| blob | f3f65c54564005a5396f4e6516caa50ab4b82ff6 |
1 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
2 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -disable-peephole -mattr=+avx512f,+avx512bw,+avx512vl,+avx512dq | FileCheck %s --check-prefix=CHECK --check-prefix=VLX
3 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -disable-peephole -mattr=+avx512f | FileCheck %s --check-prefix=CHECK --check-prefix=NoVLX
5 define zeroext i32 @test_vpcmpeqb_v16i1_v32i1_mask(<2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
6 ; VLX-LABEL: test_vpcmpeqb_v16i1_v32i1_mask:
7 ; VLX: # %bb.0: # %entry
8 ; VLX-NEXT: vpcmpeqb %xmm1, %xmm0, %k0
9 ; VLX-NEXT: kmovd %k0, %eax
10 ; VLX-NEXT: retq
11 ;
12 ; NoVLX-LABEL: test_vpcmpeqb_v16i1_v32i1_mask:
13 ; NoVLX: # %bb.0: # %entry
14 ; NoVLX-NEXT: vpcmpeqb %xmm1, %xmm0, %xmm0
15 ; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0
16 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
17 ; NoVLX-NEXT: kmovw %k0, %eax
18 ; NoVLX-NEXT: vzeroupper
19 ; NoVLX-NEXT: retq
20 entry:
21 %0 = bitcast <2 x i64> %__a to <16 x i8>
22 %1 = bitcast <2 x i64> %__b to <16 x i8>
23 %2 = icmp eq <16 x i8> %0, %1
24 %3 = shufflevector <16 x i1> %2, <16 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>
25 %4 = bitcast <32 x i1> %3 to i32
26 ret i32 %4
27 }
29 define zeroext i32 @test_vpcmpeqb_v16i1_v32i1_mask_mem(<2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
30 ; VLX-LABEL: test_vpcmpeqb_v16i1_v32i1_mask_mem:
31 ; VLX: # %bb.0: # %entry
32 ; VLX-NEXT: vpcmpeqb (%rdi), %xmm0, %k0
33 ; VLX-NEXT: kmovd %k0, %eax
34 ; VLX-NEXT: retq
35 ;
36 ; NoVLX-LABEL: test_vpcmpeqb_v16i1_v32i1_mask_mem:
37 ; NoVLX: # %bb.0: # %entry
38 ; NoVLX-NEXT: vpcmpeqb (%rdi), %xmm0, %xmm0
39 ; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0
40 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
41 ; NoVLX-NEXT: kmovw %k0, %eax
42 ; NoVLX-NEXT: vzeroupper
43 ; NoVLX-NEXT: retq
44 entry:
45 %0 = bitcast <2 x i64> %__a to <16 x i8>
46 %load = load <2 x i64>, <2 x i64>* %__b
47 %1 = bitcast <2 x i64> %load to <16 x i8>
48 %2 = icmp eq <16 x i8> %0, %1
49 %3 = shufflevector <16 x i1> %2, <16 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>
50 %4 = bitcast <32 x i1> %3 to i32
51 ret i32 %4
52 }
54 define zeroext i32 @test_masked_vpcmpeqb_v16i1_v32i1_mask(i16 zeroext %__u, <2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
55 ; VLX-LABEL: test_masked_vpcmpeqb_v16i1_v32i1_mask:
56 ; VLX: # %bb.0: # %entry
57 ; VLX-NEXT: kmovd %edi, %k1
58 ; VLX-NEXT: vpcmpeqb %xmm1, %xmm0, %k0 {%k1}
59 ; VLX-NEXT: kmovd %k0, %eax
60 ; VLX-NEXT: retq
61 ;
62 ; NoVLX-LABEL: test_masked_vpcmpeqb_v16i1_v32i1_mask:
63 ; NoVLX: # %bb.0: # %entry
64 ; NoVLX-NEXT: vpcmpeqb %xmm1, %xmm0, %xmm0
65 ; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0
66 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
67 ; NoVLX-NEXT: kmovw %k0, %eax
68 ; NoVLX-NEXT: andl %edi, %eax
69 ; NoVLX-NEXT: vzeroupper
70 ; NoVLX-NEXT: retq
71 entry:
72 %0 = bitcast <2 x i64> %__a to <16 x i8>
73 %1 = bitcast <2 x i64> %__b to <16 x i8>
74 %2 = icmp eq <16 x i8> %0, %1
75 %3 = bitcast i16 %__u to <16 x i1>
76 %4 = and <16 x i1> %2, %3
77 %5 = shufflevector <16 x i1> %4, <16 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>
78 %6 = bitcast <32 x i1> %5 to i32
79 ret i32 %6
80 }
82 define zeroext i32 @test_masked_vpcmpeqb_v16i1_v32i1_mask_mem(i16 zeroext %__u, <2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
83 ; VLX-LABEL: test_masked_vpcmpeqb_v16i1_v32i1_mask_mem:
84 ; VLX: # %bb.0: # %entry
85 ; VLX-NEXT: kmovd %edi, %k1
86 ; VLX-NEXT: vpcmpeqb (%rsi), %xmm0, %k0 {%k1}
87 ; VLX-NEXT: kmovd %k0, %eax
88 ; VLX-NEXT: retq
89 ;
90 ; NoVLX-LABEL: test_masked_vpcmpeqb_v16i1_v32i1_mask_mem:
91 ; NoVLX: # %bb.0: # %entry
92 ; NoVLX-NEXT: vpcmpeqb (%rsi), %xmm0, %xmm0
93 ; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0
94 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
95 ; NoVLX-NEXT: kmovw %k0, %eax
96 ; NoVLX-NEXT: andl %edi, %eax
97 ; NoVLX-NEXT: vzeroupper
98 ; NoVLX-NEXT: retq
99 entry:
100 %0 = bitcast <2 x i64> %__a to <16 x i8>
101 %load = load <2 x i64>, <2 x i64>* %__b
102 %1 = bitcast <2 x i64> %load to <16 x i8>
103 %2 = icmp eq <16 x i8> %0, %1
104 %3 = bitcast i16 %__u to <16 x i1>
105 %4 = and <16 x i1> %2, %3
106 %5 = shufflevector <16 x i1> %4, <16 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>
107 %6 = bitcast <32 x i1> %5 to i32
108 ret i32 %6
109 }
112 define zeroext i64 @test_vpcmpeqb_v16i1_v64i1_mask(<2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
113 ; VLX-LABEL: test_vpcmpeqb_v16i1_v64i1_mask:
114 ; VLX: # %bb.0: # %entry
115 ; VLX-NEXT: vpcmpeqb %xmm1, %xmm0, %k0
116 ; VLX-NEXT: kmovq %k0, %rax
117 ; VLX-NEXT: retq
118 ;
119 ; NoVLX-LABEL: test_vpcmpeqb_v16i1_v64i1_mask:
120 ; NoVLX: # %bb.0: # %entry
121 ; NoVLX-NEXT: vpcmpeqb %xmm1, %xmm0, %xmm0
122 ; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0
123 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
124 ; NoVLX-NEXT: kmovw %k0, %eax
125 ; NoVLX-NEXT: vzeroupper
126 ; NoVLX-NEXT: retq
127 entry:
128 %0 = bitcast <2 x i64> %__a to <16 x i8>
129 %1 = bitcast <2 x i64> %__b to <16 x i8>
130 %2 = icmp eq <16 x i8> %0, %1
131 %3 = shufflevector <16 x i1> %2, <16 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>
132 %4 = bitcast <64 x i1> %3 to i64
133 ret i64 %4
134 }
136 define zeroext i64 @test_vpcmpeqb_v16i1_v64i1_mask_mem(<2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
137 ; VLX-LABEL: test_vpcmpeqb_v16i1_v64i1_mask_mem:
138 ; VLX: # %bb.0: # %entry
139 ; VLX-NEXT: vpcmpeqb (%rdi), %xmm0, %k0
140 ; VLX-NEXT: kmovq %k0, %rax
141 ; VLX-NEXT: retq
142 ;
143 ; NoVLX-LABEL: test_vpcmpeqb_v16i1_v64i1_mask_mem:
144 ; NoVLX: # %bb.0: # %entry
145 ; NoVLX-NEXT: vpcmpeqb (%rdi), %xmm0, %xmm0
146 ; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0
147 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
148 ; NoVLX-NEXT: kmovw %k0, %eax
149 ; NoVLX-NEXT: vzeroupper
150 ; NoVLX-NEXT: retq
151 entry:
152 %0 = bitcast <2 x i64> %__a to <16 x i8>
153 %load = load <2 x i64>, <2 x i64>* %__b
154 %1 = bitcast <2 x i64> %load to <16 x i8>
155 %2 = icmp eq <16 x i8> %0, %1
156 %3 = shufflevector <16 x i1> %2, <16 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>
157 %4 = bitcast <64 x i1> %3 to i64
158 ret i64 %4
159 }
161 define zeroext i64 @test_masked_vpcmpeqb_v16i1_v64i1_mask(i16 zeroext %__u, <2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
162 ; VLX-LABEL: test_masked_vpcmpeqb_v16i1_v64i1_mask:
163 ; VLX: # %bb.0: # %entry
164 ; VLX-NEXT: kmovd %edi, %k1
165 ; VLX-NEXT: vpcmpeqb %xmm1, %xmm0, %k0 {%k1}
166 ; VLX-NEXT: kmovq %k0, %rax
167 ; VLX-NEXT: retq
168 ;
169 ; NoVLX-LABEL: test_masked_vpcmpeqb_v16i1_v64i1_mask:
170 ; NoVLX: # %bb.0: # %entry
171 ; NoVLX-NEXT: vpcmpeqb %xmm1, %xmm0, %xmm0
172 ; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0
173 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
174 ; NoVLX-NEXT: kmovw %k0, %eax
175 ; NoVLX-NEXT: andl %edi, %eax
176 ; NoVLX-NEXT: vzeroupper
177 ; NoVLX-NEXT: retq
178 entry:
179 %0 = bitcast <2 x i64> %__a to <16 x i8>
180 %1 = bitcast <2 x i64> %__b to <16 x i8>
181 %2 = icmp eq <16 x i8> %0, %1
182 %3 = bitcast i16 %__u to <16 x i1>
183 %4 = and <16 x i1> %2, %3
184 %5 = shufflevector <16 x i1> %4, <16 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>
185 %6 = bitcast <64 x i1> %5 to i64
186 ret i64 %6
187 }
189 define zeroext i64 @test_masked_vpcmpeqb_v16i1_v64i1_mask_mem(i16 zeroext %__u, <2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
190 ; VLX-LABEL: test_masked_vpcmpeqb_v16i1_v64i1_mask_mem:
191 ; VLX: # %bb.0: # %entry
192 ; VLX-NEXT: kmovd %edi, %k1
193 ; VLX-NEXT: vpcmpeqb (%rsi), %xmm0, %k0 {%k1}
194 ; VLX-NEXT: kmovq %k0, %rax
195 ; VLX-NEXT: retq
196 ;
197 ; NoVLX-LABEL: test_masked_vpcmpeqb_v16i1_v64i1_mask_mem:
198 ; NoVLX: # %bb.0: # %entry
199 ; NoVLX-NEXT: vpcmpeqb (%rsi), %xmm0, %xmm0
200 ; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0
201 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
202 ; NoVLX-NEXT: kmovw %k0, %eax
203 ; NoVLX-NEXT: andl %edi, %eax
204 ; NoVLX-NEXT: vzeroupper
205 ; NoVLX-NEXT: retq
206 entry:
207 %0 = bitcast <2 x i64> %__a to <16 x i8>
208 %load = load <2 x i64>, <2 x i64>* %__b
209 %1 = bitcast <2 x i64> %load to <16 x i8>
210 %2 = icmp eq <16 x i8> %0, %1
211 %3 = bitcast i16 %__u to <16 x i1>
212 %4 = and <16 x i1> %2, %3
213 %5 = shufflevector <16 x i1> %4, <16 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>
214 %6 = bitcast <64 x i1> %5 to i64
215 ret i64 %6
216 }
219 define zeroext i64 @test_vpcmpeqb_v32i1_v64i1_mask(<4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
220 ; VLX-LABEL: test_vpcmpeqb_v32i1_v64i1_mask:
221 ; VLX: # %bb.0: # %entry
222 ; VLX-NEXT: vpcmpeqb %ymm1, %ymm0, %k0
223 ; VLX-NEXT: kmovq %k0, %rax
224 ; VLX-NEXT: vzeroupper
225 ; VLX-NEXT: retq
226 ;
227 ; NoVLX-LABEL: test_vpcmpeqb_v32i1_v64i1_mask:
228 ; NoVLX: # %bb.0: # %entry
229 ; NoVLX-NEXT: vpcmpeqb %ymm1, %ymm0, %ymm0
230 ; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm1
231 ; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
232 ; NoVLX-NEXT: kmovw %k0, %ecx
233 ; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm0
234 ; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0
235 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
236 ; NoVLX-NEXT: kmovw %k0, %eax
237 ; NoVLX-NEXT: shll $16, %eax
238 ; NoVLX-NEXT: orl %ecx, %eax
239 ; NoVLX-NEXT: vzeroupper
240 ; NoVLX-NEXT: retq
241 entry:
242 %0 = bitcast <4 x i64> %__a to <32 x i8>
243 %1 = bitcast <4 x i64> %__b to <32 x i8>
244 %2 = icmp eq <32 x i8> %0, %1
245 %3 = shufflevector <32 x i1> %2, <32 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63>
246 %4 = bitcast <64 x i1> %3 to i64
247 ret i64 %4
248 }
250 define zeroext i64 @test_vpcmpeqb_v32i1_v64i1_mask_mem(<4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
251 ; VLX-LABEL: test_vpcmpeqb_v32i1_v64i1_mask_mem:
252 ; VLX: # %bb.0: # %entry
253 ; VLX-NEXT: vpcmpeqb (%rdi), %ymm0, %k0
254 ; VLX-NEXT: kmovq %k0, %rax
255 ; VLX-NEXT: vzeroupper
256 ; VLX-NEXT: retq
257 ;
258 ; NoVLX-LABEL: test_vpcmpeqb_v32i1_v64i1_mask_mem:
259 ; NoVLX: # %bb.0: # %entry
260 ; NoVLX-NEXT: vpcmpeqb (%rdi), %ymm0, %ymm0
261 ; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm1
262 ; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
263 ; NoVLX-NEXT: kmovw %k0, %ecx
264 ; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm0
265 ; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0
266 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
267 ; NoVLX-NEXT: kmovw %k0, %eax
268 ; NoVLX-NEXT: shll $16, %eax
269 ; NoVLX-NEXT: orl %ecx, %eax
270 ; NoVLX-NEXT: vzeroupper
271 ; NoVLX-NEXT: retq
272 entry:
273 %0 = bitcast <4 x i64> %__a to <32 x i8>
274 %load = load <4 x i64>, <4 x i64>* %__b
275 %1 = bitcast <4 x i64> %load to <32 x i8>
276 %2 = icmp eq <32 x i8> %0, %1
277 %3 = shufflevector <32 x i1> %2, <32 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63>
278 %4 = bitcast <64 x i1> %3 to i64
279 ret i64 %4
280 }
282 define zeroext i64 @test_masked_vpcmpeqb_v32i1_v64i1_mask(i32 zeroext %__u, <4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
283 ; VLX-LABEL: test_masked_vpcmpeqb_v32i1_v64i1_mask:
284 ; VLX: # %bb.0: # %entry
285 ; VLX-NEXT: kmovd %edi, %k1
286 ; VLX-NEXT: vpcmpeqb %ymm1, %ymm0, %k0 {%k1}
287 ; VLX-NEXT: kmovq %k0, %rax
288 ; VLX-NEXT: vzeroupper
289 ; VLX-NEXT: retq
290 ;
291 ; NoVLX-LABEL: test_masked_vpcmpeqb_v32i1_v64i1_mask:
292 ; NoVLX: # %bb.0: # %entry
293 ; NoVLX-NEXT: vpcmpeqb %ymm1, %ymm0, %ymm0
294 ; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm1
295 ; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
296 ; NoVLX-NEXT: kmovw %k0, %eax
297 ; NoVLX-NEXT: andl %edi, %eax
298 ; NoVLX-NEXT: shrl $16, %edi
299 ; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm0
300 ; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0
301 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
302 ; NoVLX-NEXT: kmovw %k0, %ecx
303 ; NoVLX-NEXT: andl %edi, %ecx
304 ; NoVLX-NEXT: shll $16, %ecx
305 ; NoVLX-NEXT: movzwl %ax, %eax
306 ; NoVLX-NEXT: orl %ecx, %eax
307 ; NoVLX-NEXT: vzeroupper
308 ; NoVLX-NEXT: retq
309 entry:
310 %0 = bitcast <4 x i64> %__a to <32 x i8>
311 %1 = bitcast <4 x i64> %__b to <32 x i8>
312 %2 = icmp eq <32 x i8> %0, %1
313 %3 = bitcast i32 %__u to <32 x i1>
314 %4 = and <32 x i1> %2, %3
315 %5 = shufflevector <32 x i1> %4, <32 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63>
316 %6 = bitcast <64 x i1> %5 to i64
317 ret i64 %6
318 }
320 define zeroext i64 @test_masked_vpcmpeqb_v32i1_v64i1_mask_mem(i32 zeroext %__u, <4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
321 ; VLX-LABEL: test_masked_vpcmpeqb_v32i1_v64i1_mask_mem:
322 ; VLX: # %bb.0: # %entry
323 ; VLX-NEXT: kmovd %edi, %k1
324 ; VLX-NEXT: vpcmpeqb (%rsi), %ymm0, %k0 {%k1}
325 ; VLX-NEXT: kmovq %k0, %rax
326 ; VLX-NEXT: vzeroupper
327 ; VLX-NEXT: retq
328 ;
329 ; NoVLX-LABEL: test_masked_vpcmpeqb_v32i1_v64i1_mask_mem:
330 ; NoVLX: # %bb.0: # %entry
331 ; NoVLX-NEXT: vpcmpeqb (%rsi), %ymm0, %ymm0
332 ; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm1
333 ; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
334 ; NoVLX-NEXT: kmovw %k0, %eax
335 ; NoVLX-NEXT: andl %edi, %eax
336 ; NoVLX-NEXT: shrl $16, %edi
337 ; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm0
338 ; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0
339 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
340 ; NoVLX-NEXT: kmovw %k0, %ecx
341 ; NoVLX-NEXT: andl %edi, %ecx
342 ; NoVLX-NEXT: shll $16, %ecx
343 ; NoVLX-NEXT: movzwl %ax, %eax
344 ; NoVLX-NEXT: orl %ecx, %eax
345 ; NoVLX-NEXT: vzeroupper
346 ; NoVLX-NEXT: retq
347 entry:
348 %0 = bitcast <4 x i64> %__a to <32 x i8>
349 %load = load <4 x i64>, <4 x i64>* %__b
350 %1 = bitcast <4 x i64> %load to <32 x i8>
351 %2 = icmp eq <32 x i8> %0, %1
352 %3 = bitcast i32 %__u to <32 x i1>
353 %4 = and <32 x i1> %2, %3
354 %5 = shufflevector <32 x i1> %4, <32 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63>
355 %6 = bitcast <64 x i1> %5 to i64
356 ret i64 %6
357 }
360 define zeroext i16 @test_vpcmpeqw_v8i1_v16i1_mask(<2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
361 ; VLX-LABEL: test_vpcmpeqw_v8i1_v16i1_mask:
362 ; VLX: # %bb.0: # %entry
363 ; VLX-NEXT: vpcmpeqw %xmm1, %xmm0, %k0
364 ; VLX-NEXT: kmovd %k0, %eax
365 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
366 ; VLX-NEXT: retq
367 ;
368 ; NoVLX-LABEL: test_vpcmpeqw_v8i1_v16i1_mask:
369 ; NoVLX: # %bb.0: # %entry
370 ; NoVLX-NEXT: vpcmpeqw %xmm1, %xmm0, %xmm0
371 ; NoVLX-NEXT: vpmovsxwq %xmm0, %zmm0
372 ; NoVLX-NEXT: vptestmq %zmm0, %zmm0, %k0
373 ; NoVLX-NEXT: kmovw %k0, %eax
374 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
375 ; NoVLX-NEXT: vzeroupper
376 ; NoVLX-NEXT: retq
377 entry:
378 %0 = bitcast <2 x i64> %__a to <8 x i16>
379 %1 = bitcast <2 x i64> %__b to <8 x i16>
380 %2 = icmp eq <8 x i16> %0, %1
381 %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
382 %4 = bitcast <16 x i1> %3 to i16
383 ret i16 %4
384 }
386 define zeroext i16 @test_vpcmpeqw_v8i1_v16i1_mask_mem(<2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
387 ; VLX-LABEL: test_vpcmpeqw_v8i1_v16i1_mask_mem:
388 ; VLX: # %bb.0: # %entry
389 ; VLX-NEXT: vpcmpeqw (%rdi), %xmm0, %k0
390 ; VLX-NEXT: kmovd %k0, %eax
391 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
392 ; VLX-NEXT: retq
393 ;
394 ; NoVLX-LABEL: test_vpcmpeqw_v8i1_v16i1_mask_mem:
395 ; NoVLX: # %bb.0: # %entry
396 ; NoVLX-NEXT: vpcmpeqw (%rdi), %xmm0, %xmm0
397 ; NoVLX-NEXT: vpmovsxwq %xmm0, %zmm0
398 ; NoVLX-NEXT: vptestmq %zmm0, %zmm0, %k0
399 ; NoVLX-NEXT: kmovw %k0, %eax
400 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
401 ; NoVLX-NEXT: vzeroupper
402 ; NoVLX-NEXT: retq
403 entry:
404 %0 = bitcast <2 x i64> %__a to <8 x i16>
405 %load = load <2 x i64>, <2 x i64>* %__b
406 %1 = bitcast <2 x i64> %load to <8 x i16>
407 %2 = icmp eq <8 x i16> %0, %1
408 %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
409 %4 = bitcast <16 x i1> %3 to i16
410 ret i16 %4
411 }
413 define zeroext i16 @test_masked_vpcmpeqw_v8i1_v16i1_mask(i8 zeroext %__u, <2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
414 ; VLX-LABEL: test_masked_vpcmpeqw_v8i1_v16i1_mask:
415 ; VLX: # %bb.0: # %entry
416 ; VLX-NEXT: kmovd %edi, %k1
417 ; VLX-NEXT: vpcmpeqw %xmm1, %xmm0, %k0 {%k1}
418 ; VLX-NEXT: kmovd %k0, %eax
419 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
420 ; VLX-NEXT: retq
421 ;
422 ; NoVLX-LABEL: test_masked_vpcmpeqw_v8i1_v16i1_mask:
423 ; NoVLX: # %bb.0: # %entry
424 ; NoVLX-NEXT: vpcmpeqw %xmm1, %xmm0, %xmm0
425 ; NoVLX-NEXT: vpmovsxwq %xmm0, %zmm0
426 ; NoVLX-NEXT: kmovw %edi, %k1
427 ; NoVLX-NEXT: vptestmq %zmm0, %zmm0, %k0 {%k1}
428 ; NoVLX-NEXT: kmovw %k0, %eax
429 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
430 ; NoVLX-NEXT: vzeroupper
431 ; NoVLX-NEXT: retq
432 entry:
433 %0 = bitcast <2 x i64> %__a to <8 x i16>
434 %1 = bitcast <2 x i64> %__b to <8 x i16>
435 %2 = icmp eq <8 x i16> %0, %1
436 %3 = bitcast i8 %__u to <8 x i1>
437 %4 = and <8 x i1> %2, %3
438 %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
439 %6 = bitcast <16 x i1> %5 to i16
440 ret i16 %6
441 }
443 define zeroext i16 @test_masked_vpcmpeqw_v8i1_v16i1_mask_mem(i8 zeroext %__u, <2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
444 ; VLX-LABEL: test_masked_vpcmpeqw_v8i1_v16i1_mask_mem:
445 ; VLX: # %bb.0: # %entry
446 ; VLX-NEXT: kmovd %edi, %k1
447 ; VLX-NEXT: vpcmpeqw (%rsi), %xmm0, %k0 {%k1}
448 ; VLX-NEXT: kmovd %k0, %eax
449 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
450 ; VLX-NEXT: retq
451 ;
452 ; NoVLX-LABEL: test_masked_vpcmpeqw_v8i1_v16i1_mask_mem:
453 ; NoVLX: # %bb.0: # %entry
454 ; NoVLX-NEXT: vpcmpeqw (%rsi), %xmm0, %xmm0
455 ; NoVLX-NEXT: vpmovsxwq %xmm0, %zmm0
456 ; NoVLX-NEXT: kmovw %edi, %k1
457 ; NoVLX-NEXT: vptestmq %zmm0, %zmm0, %k0 {%k1}
458 ; NoVLX-NEXT: kmovw %k0, %eax
459 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
460 ; NoVLX-NEXT: vzeroupper
461 ; NoVLX-NEXT: retq
462 entry:
463 %0 = bitcast <2 x i64> %__a to <8 x i16>
464 %load = load <2 x i64>, <2 x i64>* %__b
465 %1 = bitcast <2 x i64> %load to <8 x i16>
466 %2 = icmp eq <8 x i16> %0, %1
467 %3 = bitcast i8 %__u to <8 x i1>
468 %4 = and <8 x i1> %2, %3
469 %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
470 %6 = bitcast <16 x i1> %5 to i16
471 ret i16 %6
472 }
475 define zeroext i32 @test_vpcmpeqw_v8i1_v32i1_mask(<2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
476 ; VLX-LABEL: test_vpcmpeqw_v8i1_v32i1_mask:
477 ; VLX: # %bb.0: # %entry
478 ; VLX-NEXT: vpcmpeqw %xmm1, %xmm0, %k0
479 ; VLX-NEXT: kmovd %k0, %eax
480 ; VLX-NEXT: retq
481 ;
482 ; NoVLX-LABEL: test_vpcmpeqw_v8i1_v32i1_mask:
483 ; NoVLX: # %bb.0: # %entry
484 ; NoVLX-NEXT: vpcmpeqw %xmm1, %xmm0, %xmm0
485 ; NoVLX-NEXT: vpmovsxwq %xmm0, %zmm0
486 ; NoVLX-NEXT: vptestmq %zmm0, %zmm0, %k0
487 ; NoVLX-NEXT: kmovw %k0, %eax
488 ; NoVLX-NEXT: vzeroupper
489 ; NoVLX-NEXT: retq
490 entry:
491 %0 = bitcast <2 x i64> %__a to <8 x i16>
492 %1 = bitcast <2 x i64> %__b to <8 x i16>
493 %2 = icmp eq <8 x i16> %0, %1
494 %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
495 %4 = bitcast <32 x i1> %3 to i32
496 ret i32 %4
497 }
499 define zeroext i32 @test_vpcmpeqw_v8i1_v32i1_mask_mem(<2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
500 ; VLX-LABEL: test_vpcmpeqw_v8i1_v32i1_mask_mem:
501 ; VLX: # %bb.0: # %entry
502 ; VLX-NEXT: vpcmpeqw (%rdi), %xmm0, %k0
503 ; VLX-NEXT: kmovd %k0, %eax
504 ; VLX-NEXT: retq
505 ;
506 ; NoVLX-LABEL: test_vpcmpeqw_v8i1_v32i1_mask_mem:
507 ; NoVLX: # %bb.0: # %entry
508 ; NoVLX-NEXT: vpcmpeqw (%rdi), %xmm0, %xmm0
509 ; NoVLX-NEXT: vpmovsxwq %xmm0, %zmm0
510 ; NoVLX-NEXT: vptestmq %zmm0, %zmm0, %k0
511 ; NoVLX-NEXT: kmovw %k0, %eax
512 ; NoVLX-NEXT: vzeroupper
513 ; NoVLX-NEXT: retq
514 entry:
515 %0 = bitcast <2 x i64> %__a to <8 x i16>
516 %load = load <2 x i64>, <2 x i64>* %__b
517 %1 = bitcast <2 x i64> %load to <8 x i16>
518 %2 = icmp eq <8 x i16> %0, %1
519 %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
520 %4 = bitcast <32 x i1> %3 to i32
521 ret i32 %4
522 }
524 define zeroext i32 @test_masked_vpcmpeqw_v8i1_v32i1_mask(i8 zeroext %__u, <2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
525 ; VLX-LABEL: test_masked_vpcmpeqw_v8i1_v32i1_mask:
526 ; VLX: # %bb.0: # %entry
527 ; VLX-NEXT: kmovd %edi, %k1
528 ; VLX-NEXT: vpcmpeqw %xmm1, %xmm0, %k0 {%k1}
529 ; VLX-NEXT: kmovd %k0, %eax
530 ; VLX-NEXT: retq
531 ;
532 ; NoVLX-LABEL: test_masked_vpcmpeqw_v8i1_v32i1_mask:
533 ; NoVLX: # %bb.0: # %entry
534 ; NoVLX-NEXT: vpcmpeqw %xmm1, %xmm0, %xmm0
535 ; NoVLX-NEXT: vpmovsxwq %xmm0, %zmm0
536 ; NoVLX-NEXT: kmovw %edi, %k1
537 ; NoVLX-NEXT: vptestmq %zmm0, %zmm0, %k0 {%k1}
538 ; NoVLX-NEXT: kmovw %k0, %eax
539 ; NoVLX-NEXT: vzeroupper
540 ; NoVLX-NEXT: retq
541 entry:
542 %0 = bitcast <2 x i64> %__a to <8 x i16>
543 %1 = bitcast <2 x i64> %__b to <8 x i16>
544 %2 = icmp eq <8 x i16> %0, %1
545 %3 = bitcast i8 %__u to <8 x i1>
546 %4 = and <8 x i1> %2, %3
547 %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
548 %6 = bitcast <32 x i1> %5 to i32
549 ret i32 %6
550 }
552 define zeroext i32 @test_masked_vpcmpeqw_v8i1_v32i1_mask_mem(i8 zeroext %__u, <2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
553 ; VLX-LABEL: test_masked_vpcmpeqw_v8i1_v32i1_mask_mem:
554 ; VLX: # %bb.0: # %entry
555 ; VLX-NEXT: kmovd %edi, %k1
556 ; VLX-NEXT: vpcmpeqw (%rsi), %xmm0, %k0 {%k1}
557 ; VLX-NEXT: kmovd %k0, %eax
558 ; VLX-NEXT: retq
559 ;
560 ; NoVLX-LABEL: test_masked_vpcmpeqw_v8i1_v32i1_mask_mem:
561 ; NoVLX: # %bb.0: # %entry
562 ; NoVLX-NEXT: vpcmpeqw (%rsi), %xmm0, %xmm0
563 ; NoVLX-NEXT: vpmovsxwq %xmm0, %zmm0
564 ; NoVLX-NEXT: kmovw %edi, %k1
565 ; NoVLX-NEXT: vptestmq %zmm0, %zmm0, %k0 {%k1}
566 ; NoVLX-NEXT: kmovw %k0, %eax
567 ; NoVLX-NEXT: vzeroupper
568 ; NoVLX-NEXT: retq
569 entry:
570 %0 = bitcast <2 x i64> %__a to <8 x i16>
571 %load = load <2 x i64>, <2 x i64>* %__b
572 %1 = bitcast <2 x i64> %load to <8 x i16>
573 %2 = icmp eq <8 x i16> %0, %1
574 %3 = bitcast i8 %__u to <8 x i1>
575 %4 = and <8 x i1> %2, %3
576 %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
577 %6 = bitcast <32 x i1> %5 to i32
578 ret i32 %6
579 }
582 define zeroext i64 @test_vpcmpeqw_v8i1_v64i1_mask(<2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
583 ; VLX-LABEL: test_vpcmpeqw_v8i1_v64i1_mask:
584 ; VLX: # %bb.0: # %entry
585 ; VLX-NEXT: vpcmpeqw %xmm1, %xmm0, %k0
586 ; VLX-NEXT: kmovq %k0, %rax
587 ; VLX-NEXT: retq
588 ;
589 ; NoVLX-LABEL: test_vpcmpeqw_v8i1_v64i1_mask:
590 ; NoVLX: # %bb.0: # %entry
591 ; NoVLX-NEXT: vpcmpeqw %xmm1, %xmm0, %xmm0
592 ; NoVLX-NEXT: vpmovsxwq %xmm0, %zmm0
593 ; NoVLX-NEXT: vptestmq %zmm0, %zmm0, %k0
594 ; NoVLX-NEXT: kmovw %k0, %eax
595 ; NoVLX-NEXT: vzeroupper
596 ; NoVLX-NEXT: retq
597 entry:
598 %0 = bitcast <2 x i64> %__a to <8 x i16>
599 %1 = bitcast <2 x i64> %__b to <8 x i16>
600 %2 = icmp eq <8 x i16> %0, %1
601 %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
602 %4 = bitcast <64 x i1> %3 to i64
603 ret i64 %4
604 }
606 define zeroext i64 @test_vpcmpeqw_v8i1_v64i1_mask_mem(<2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
607 ; VLX-LABEL: test_vpcmpeqw_v8i1_v64i1_mask_mem:
608 ; VLX: # %bb.0: # %entry
609 ; VLX-NEXT: vpcmpeqw (%rdi), %xmm0, %k0
610 ; VLX-NEXT: kmovq %k0, %rax
611 ; VLX-NEXT: retq
612 ;
613 ; NoVLX-LABEL: test_vpcmpeqw_v8i1_v64i1_mask_mem:
614 ; NoVLX: # %bb.0: # %entry
615 ; NoVLX-NEXT: vpcmpeqw (%rdi), %xmm0, %xmm0
616 ; NoVLX-NEXT: vpmovsxwq %xmm0, %zmm0
617 ; NoVLX-NEXT: vptestmq %zmm0, %zmm0, %k0
618 ; NoVLX-NEXT: kmovw %k0, %eax
619 ; NoVLX-NEXT: vzeroupper
620 ; NoVLX-NEXT: retq
621 entry:
622 %0 = bitcast <2 x i64> %__a to <8 x i16>
623 %load = load <2 x i64>, <2 x i64>* %__b
624 %1 = bitcast <2 x i64> %load to <8 x i16>
625 %2 = icmp eq <8 x i16> %0, %1
626 %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
627 %4 = bitcast <64 x i1> %3 to i64
628 ret i64 %4
629 }
631 define zeroext i64 @test_masked_vpcmpeqw_v8i1_v64i1_mask(i8 zeroext %__u, <2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
632 ; VLX-LABEL: test_masked_vpcmpeqw_v8i1_v64i1_mask:
633 ; VLX: # %bb.0: # %entry
634 ; VLX-NEXT: kmovd %edi, %k1
635 ; VLX-NEXT: vpcmpeqw %xmm1, %xmm0, %k0 {%k1}
636 ; VLX-NEXT: kmovq %k0, %rax
637 ; VLX-NEXT: retq
638 ;
639 ; NoVLX-LABEL: test_masked_vpcmpeqw_v8i1_v64i1_mask:
640 ; NoVLX: # %bb.0: # %entry
641 ; NoVLX-NEXT: vpcmpeqw %xmm1, %xmm0, %xmm0
642 ; NoVLX-NEXT: vpmovsxwq %xmm0, %zmm0
643 ; NoVLX-NEXT: kmovw %edi, %k1
644 ; NoVLX-NEXT: vptestmq %zmm0, %zmm0, %k0 {%k1}
645 ; NoVLX-NEXT: kmovw %k0, %eax
646 ; NoVLX-NEXT: vzeroupper
647 ; NoVLX-NEXT: retq
648 entry:
649 %0 = bitcast <2 x i64> %__a to <8 x i16>
650 %1 = bitcast <2 x i64> %__b to <8 x i16>
651 %2 = icmp eq <8 x i16> %0, %1
652 %3 = bitcast i8 %__u to <8 x i1>
653 %4 = and <8 x i1> %2, %3
654 %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
655 %6 = bitcast <64 x i1> %5 to i64
656 ret i64 %6
657 }
659 define zeroext i64 @test_masked_vpcmpeqw_v8i1_v64i1_mask_mem(i8 zeroext %__u, <2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
660 ; VLX-LABEL: test_masked_vpcmpeqw_v8i1_v64i1_mask_mem:
661 ; VLX: # %bb.0: # %entry
662 ; VLX-NEXT: kmovd %edi, %k1
663 ; VLX-NEXT: vpcmpeqw (%rsi), %xmm0, %k0 {%k1}
664 ; VLX-NEXT: kmovq %k0, %rax
665 ; VLX-NEXT: retq
666 ;
667 ; NoVLX-LABEL: test_masked_vpcmpeqw_v8i1_v64i1_mask_mem:
668 ; NoVLX: # %bb.0: # %entry
669 ; NoVLX-NEXT: vpcmpeqw (%rsi), %xmm0, %xmm0
670 ; NoVLX-NEXT: vpmovsxwq %xmm0, %zmm0
671 ; NoVLX-NEXT: kmovw %edi, %k1
672 ; NoVLX-NEXT: vptestmq %zmm0, %zmm0, %k0 {%k1}
673 ; NoVLX-NEXT: kmovw %k0, %eax
674 ; NoVLX-NEXT: vzeroupper
675 ; NoVLX-NEXT: retq
676 entry:
677 %0 = bitcast <2 x i64> %__a to <8 x i16>
678 %load = load <2 x i64>, <2 x i64>* %__b
679 %1 = bitcast <2 x i64> %load to <8 x i16>
680 %2 = icmp eq <8 x i16> %0, %1
681 %3 = bitcast i8 %__u to <8 x i1>
682 %4 = and <8 x i1> %2, %3
683 %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
684 %6 = bitcast <64 x i1> %5 to i64
685 ret i64 %6
686 }
689 define zeroext i32 @test_vpcmpeqw_v16i1_v32i1_mask(<4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
690 ; VLX-LABEL: test_vpcmpeqw_v16i1_v32i1_mask:
691 ; VLX: # %bb.0: # %entry
692 ; VLX-NEXT: vpcmpeqw %ymm1, %ymm0, %k0
693 ; VLX-NEXT: kmovd %k0, %eax
694 ; VLX-NEXT: vzeroupper
695 ; VLX-NEXT: retq
696 ;
697 ; NoVLX-LABEL: test_vpcmpeqw_v16i1_v32i1_mask:
698 ; NoVLX: # %bb.0: # %entry
699 ; NoVLX-NEXT: vpcmpeqw %ymm1, %ymm0, %ymm0
700 ; NoVLX-NEXT: vpmovsxwd %ymm0, %zmm0
701 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
702 ; NoVLX-NEXT: kmovw %k0, %eax
703 ; NoVLX-NEXT: vzeroupper
704 ; NoVLX-NEXT: retq
705 entry:
706 %0 = bitcast <4 x i64> %__a to <16 x i16>
707 %1 = bitcast <4 x i64> %__b to <16 x i16>
708 %2 = icmp eq <16 x i16> %0, %1
709 %3 = shufflevector <16 x i1> %2, <16 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>
710 %4 = bitcast <32 x i1> %3 to i32
711 ret i32 %4
712 }
714 define zeroext i32 @test_vpcmpeqw_v16i1_v32i1_mask_mem(<4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
715 ; VLX-LABEL: test_vpcmpeqw_v16i1_v32i1_mask_mem:
716 ; VLX: # %bb.0: # %entry
717 ; VLX-NEXT: vpcmpeqw (%rdi), %ymm0, %k0
718 ; VLX-NEXT: kmovd %k0, %eax
719 ; VLX-NEXT: vzeroupper
720 ; VLX-NEXT: retq
721 ;
722 ; NoVLX-LABEL: test_vpcmpeqw_v16i1_v32i1_mask_mem:
723 ; NoVLX: # %bb.0: # %entry
724 ; NoVLX-NEXT: vpcmpeqw (%rdi), %ymm0, %ymm0
725 ; NoVLX-NEXT: vpmovsxwd %ymm0, %zmm0
726 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
727 ; NoVLX-NEXT: kmovw %k0, %eax
728 ; NoVLX-NEXT: vzeroupper
729 ; NoVLX-NEXT: retq
730 entry:
731 %0 = bitcast <4 x i64> %__a to <16 x i16>
732 %load = load <4 x i64>, <4 x i64>* %__b
733 %1 = bitcast <4 x i64> %load to <16 x i16>
734 %2 = icmp eq <16 x i16> %0, %1
735 %3 = shufflevector <16 x i1> %2, <16 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>
736 %4 = bitcast <32 x i1> %3 to i32
737 ret i32 %4
738 }
740 define zeroext i32 @test_masked_vpcmpeqw_v16i1_v32i1_mask(i16 zeroext %__u, <4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
741 ; VLX-LABEL: test_masked_vpcmpeqw_v16i1_v32i1_mask:
742 ; VLX: # %bb.0: # %entry
743 ; VLX-NEXT: kmovd %edi, %k1
744 ; VLX-NEXT: vpcmpeqw %ymm1, %ymm0, %k0 {%k1}
745 ; VLX-NEXT: kmovd %k0, %eax
746 ; VLX-NEXT: vzeroupper
747 ; VLX-NEXT: retq
748 ;
749 ; NoVLX-LABEL: test_masked_vpcmpeqw_v16i1_v32i1_mask:
750 ; NoVLX: # %bb.0: # %entry
751 ; NoVLX-NEXT: vpcmpeqw %ymm1, %ymm0, %ymm0
752 ; NoVLX-NEXT: vpmovsxwd %ymm0, %zmm0
753 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
754 ; NoVLX-NEXT: kmovw %k0, %eax
755 ; NoVLX-NEXT: andl %edi, %eax
756 ; NoVLX-NEXT: vzeroupper
757 ; NoVLX-NEXT: retq
758 entry:
759 %0 = bitcast <4 x i64> %__a to <16 x i16>
760 %1 = bitcast <4 x i64> %__b to <16 x i16>
761 %2 = icmp eq <16 x i16> %0, %1
762 %3 = bitcast i16 %__u to <16 x i1>
763 %4 = and <16 x i1> %2, %3
764 %5 = shufflevector <16 x i1> %4, <16 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>
765 %6 = bitcast <32 x i1> %5 to i32
766 ret i32 %6
767 }
769 define zeroext i32 @test_masked_vpcmpeqw_v16i1_v32i1_mask_mem(i16 zeroext %__u, <4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
770 ; VLX-LABEL: test_masked_vpcmpeqw_v16i1_v32i1_mask_mem:
771 ; VLX: # %bb.0: # %entry
772 ; VLX-NEXT: kmovd %edi, %k1
773 ; VLX-NEXT: vpcmpeqw (%rsi), %ymm0, %k0 {%k1}
774 ; VLX-NEXT: kmovd %k0, %eax
775 ; VLX-NEXT: vzeroupper
776 ; VLX-NEXT: retq
777 ;
778 ; NoVLX-LABEL: test_masked_vpcmpeqw_v16i1_v32i1_mask_mem:
779 ; NoVLX: # %bb.0: # %entry
780 ; NoVLX-NEXT: vpcmpeqw (%rsi), %ymm0, %ymm0
781 ; NoVLX-NEXT: vpmovsxwd %ymm0, %zmm0
782 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
783 ; NoVLX-NEXT: kmovw %k0, %eax
784 ; NoVLX-NEXT: andl %edi, %eax
785 ; NoVLX-NEXT: vzeroupper
786 ; NoVLX-NEXT: retq
787 entry:
788 %0 = bitcast <4 x i64> %__a to <16 x i16>
789 %load = load <4 x i64>, <4 x i64>* %__b
790 %1 = bitcast <4 x i64> %load to <16 x i16>
791 %2 = icmp eq <16 x i16> %0, %1
792 %3 = bitcast i16 %__u to <16 x i1>
793 %4 = and <16 x i1> %2, %3
794 %5 = shufflevector <16 x i1> %4, <16 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>
795 %6 = bitcast <32 x i1> %5 to i32
796 ret i32 %6
797 }
800 define zeroext i64 @test_vpcmpeqw_v16i1_v64i1_mask(<4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
801 ; VLX-LABEL: test_vpcmpeqw_v16i1_v64i1_mask:
802 ; VLX: # %bb.0: # %entry
803 ; VLX-NEXT: vpcmpeqw %ymm1, %ymm0, %k0
804 ; VLX-NEXT: kmovq %k0, %rax
805 ; VLX-NEXT: vzeroupper
806 ; VLX-NEXT: retq
807 ;
808 ; NoVLX-LABEL: test_vpcmpeqw_v16i1_v64i1_mask:
809 ; NoVLX: # %bb.0: # %entry
810 ; NoVLX-NEXT: vpcmpeqw %ymm1, %ymm0, %ymm0
811 ; NoVLX-NEXT: vpmovsxwd %ymm0, %zmm0
812 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
813 ; NoVLX-NEXT: kmovw %k0, %eax
814 ; NoVLX-NEXT: vzeroupper
815 ; NoVLX-NEXT: retq
816 entry:
817 %0 = bitcast <4 x i64> %__a to <16 x i16>
818 %1 = bitcast <4 x i64> %__b to <16 x i16>
819 %2 = icmp eq <16 x i16> %0, %1
820 %3 = shufflevector <16 x i1> %2, <16 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>
821 %4 = bitcast <64 x i1> %3 to i64
822 ret i64 %4
823 }
825 define zeroext i64 @test_vpcmpeqw_v16i1_v64i1_mask_mem(<4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
826 ; VLX-LABEL: test_vpcmpeqw_v16i1_v64i1_mask_mem:
827 ; VLX: # %bb.0: # %entry
828 ; VLX-NEXT: vpcmpeqw (%rdi), %ymm0, %k0
829 ; VLX-NEXT: kmovq %k0, %rax
830 ; VLX-NEXT: vzeroupper
831 ; VLX-NEXT: retq
832 ;
833 ; NoVLX-LABEL: test_vpcmpeqw_v16i1_v64i1_mask_mem:
834 ; NoVLX: # %bb.0: # %entry
835 ; NoVLX-NEXT: vpcmpeqw (%rdi), %ymm0, %ymm0
836 ; NoVLX-NEXT: vpmovsxwd %ymm0, %zmm0
837 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
838 ; NoVLX-NEXT: kmovw %k0, %eax
839 ; NoVLX-NEXT: vzeroupper
840 ; NoVLX-NEXT: retq
841 entry:
842 %0 = bitcast <4 x i64> %__a to <16 x i16>
843 %load = load <4 x i64>, <4 x i64>* %__b
844 %1 = bitcast <4 x i64> %load to <16 x i16>
845 %2 = icmp eq <16 x i16> %0, %1
846 %3 = shufflevector <16 x i1> %2, <16 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>
847 %4 = bitcast <64 x i1> %3 to i64
848 ret i64 %4
849 }
851 define zeroext i64 @test_masked_vpcmpeqw_v16i1_v64i1_mask(i16 zeroext %__u, <4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
852 ; VLX-LABEL: test_masked_vpcmpeqw_v16i1_v64i1_mask:
853 ; VLX: # %bb.0: # %entry
854 ; VLX-NEXT: kmovd %edi, %k1
855 ; VLX-NEXT: vpcmpeqw %ymm1, %ymm0, %k0 {%k1}
856 ; VLX-NEXT: kmovq %k0, %rax
857 ; VLX-NEXT: vzeroupper
858 ; VLX-NEXT: retq
859 ;
860 ; NoVLX-LABEL: test_masked_vpcmpeqw_v16i1_v64i1_mask:
861 ; NoVLX: # %bb.0: # %entry
862 ; NoVLX-NEXT: vpcmpeqw %ymm1, %ymm0, %ymm0
863 ; NoVLX-NEXT: vpmovsxwd %ymm0, %zmm0
864 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
865 ; NoVLX-NEXT: kmovw %k0, %eax
866 ; NoVLX-NEXT: andl %edi, %eax
867 ; NoVLX-NEXT: vzeroupper
868 ; NoVLX-NEXT: retq
869 entry:
870 %0 = bitcast <4 x i64> %__a to <16 x i16>
871 %1 = bitcast <4 x i64> %__b to <16 x i16>
872 %2 = icmp eq <16 x i16> %0, %1
873 %3 = bitcast i16 %__u to <16 x i1>
874 %4 = and <16 x i1> %2, %3
875 %5 = shufflevector <16 x i1> %4, <16 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>
876 %6 = bitcast <64 x i1> %5 to i64
877 ret i64 %6
878 }
880 define zeroext i64 @test_masked_vpcmpeqw_v16i1_v64i1_mask_mem(i16 zeroext %__u, <4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
881 ; VLX-LABEL: test_masked_vpcmpeqw_v16i1_v64i1_mask_mem:
882 ; VLX: # %bb.0: # %entry
883 ; VLX-NEXT: kmovd %edi, %k1
884 ; VLX-NEXT: vpcmpeqw (%rsi), %ymm0, %k0 {%k1}
885 ; VLX-NEXT: kmovq %k0, %rax
886 ; VLX-NEXT: vzeroupper
887 ; VLX-NEXT: retq
888 ;
889 ; NoVLX-LABEL: test_masked_vpcmpeqw_v16i1_v64i1_mask_mem:
890 ; NoVLX: # %bb.0: # %entry
891 ; NoVLX-NEXT: vpcmpeqw (%rsi), %ymm0, %ymm0
892 ; NoVLX-NEXT: vpmovsxwd %ymm0, %zmm0
893 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
894 ; NoVLX-NEXT: kmovw %k0, %eax
895 ; NoVLX-NEXT: andl %edi, %eax
896 ; NoVLX-NEXT: vzeroupper
897 ; NoVLX-NEXT: retq
898 entry:
899 %0 = bitcast <4 x i64> %__a to <16 x i16>
900 %load = load <4 x i64>, <4 x i64>* %__b
901 %1 = bitcast <4 x i64> %load to <16 x i16>
902 %2 = icmp eq <16 x i16> %0, %1
903 %3 = bitcast i16 %__u to <16 x i1>
904 %4 = and <16 x i1> %2, %3
905 %5 = shufflevector <16 x i1> %4, <16 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>
906 %6 = bitcast <64 x i1> %5 to i64
907 ret i64 %6
908 }
911 define zeroext i64 @test_vpcmpeqw_v32i1_v64i1_mask(<8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
912 ; VLX-LABEL: test_vpcmpeqw_v32i1_v64i1_mask:
913 ; VLX: # %bb.0: # %entry
914 ; VLX-NEXT: vpcmpeqw %zmm1, %zmm0, %k0
915 ; VLX-NEXT: kmovq %k0, %rax
916 ; VLX-NEXT: vzeroupper
917 ; VLX-NEXT: retq
918 ;
919 ; NoVLX-LABEL: test_vpcmpeqw_v32i1_v64i1_mask:
920 ; NoVLX: # %bb.0: # %entry
921 ; NoVLX-NEXT: vextracti64x4 $1, %zmm0, %ymm2
922 ; NoVLX-NEXT: vextracti64x4 $1, %zmm1, %ymm3
923 ; NoVLX-NEXT: vpcmpeqw %ymm3, %ymm2, %ymm2
924 ; NoVLX-NEXT: vpcmpeqw %ymm1, %ymm0, %ymm0
925 ; NoVLX-NEXT: vpmovsxwd %ymm0, %zmm0
926 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
927 ; NoVLX-NEXT: kmovw %k0, %ecx
928 ; NoVLX-NEXT: vpmovsxwd %ymm2, %zmm0
929 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
930 ; NoVLX-NEXT: kmovw %k0, %eax
931 ; NoVLX-NEXT: shll $16, %eax
932 ; NoVLX-NEXT: orl %ecx, %eax
933 ; NoVLX-NEXT: vzeroupper
934 ; NoVLX-NEXT: retq
935 entry:
936 %0 = bitcast <8 x i64> %__a to <32 x i16>
937 %1 = bitcast <8 x i64> %__b to <32 x i16>
938 %2 = icmp eq <32 x i16> %0, %1
939 %3 = shufflevector <32 x i1> %2, <32 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63>
940 %4 = bitcast <64 x i1> %3 to i64
941 ret i64 %4
942 }
944 define zeroext i64 @test_vpcmpeqw_v32i1_v64i1_mask_mem(<8 x i64> %__a, <8 x i64>* %__b) local_unnamed_addr {
945 ; VLX-LABEL: test_vpcmpeqw_v32i1_v64i1_mask_mem:
946 ; VLX: # %bb.0: # %entry
947 ; VLX-NEXT: vpcmpeqw (%rdi), %zmm0, %k0
948 ; VLX-NEXT: kmovq %k0, %rax
949 ; VLX-NEXT: vzeroupper
950 ; VLX-NEXT: retq
951 ;
952 ; NoVLX-LABEL: test_vpcmpeqw_v32i1_v64i1_mask_mem:
953 ; NoVLX: # %bb.0: # %entry
954 ; NoVLX-NEXT: vextracti64x4 $1, %zmm0, %ymm1
955 ; NoVLX-NEXT: vpcmpeqw (%rdi), %ymm0, %ymm0
956 ; NoVLX-NEXT: vpmovsxwd %ymm0, %zmm0
957 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
958 ; NoVLX-NEXT: kmovw %k0, %ecx
959 ; NoVLX-NEXT: vpcmpeqw 32(%rdi), %ymm1, %ymm0
960 ; NoVLX-NEXT: vpmovsxwd %ymm0, %zmm0
961 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
962 ; NoVLX-NEXT: kmovw %k0, %eax
963 ; NoVLX-NEXT: shll $16, %eax
964 ; NoVLX-NEXT: orl %ecx, %eax
965 ; NoVLX-NEXT: vzeroupper
966 ; NoVLX-NEXT: retq
967 entry:
968 %0 = bitcast <8 x i64> %__a to <32 x i16>
969 %load = load <8 x i64>, <8 x i64>* %__b
970 %1 = bitcast <8 x i64> %load to <32 x i16>
971 %2 = icmp eq <32 x i16> %0, %1
972 %3 = shufflevector <32 x i1> %2, <32 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63>
973 %4 = bitcast <64 x i1> %3 to i64
974 ret i64 %4
975 }
977 define zeroext i64 @test_masked_vpcmpeqw_v32i1_v64i1_mask(i32 zeroext %__u, <8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
978 ; VLX-LABEL: test_masked_vpcmpeqw_v32i1_v64i1_mask:
979 ; VLX: # %bb.0: # %entry
980 ; VLX-NEXT: kmovd %edi, %k1
981 ; VLX-NEXT: vpcmpeqw %zmm1, %zmm0, %k0 {%k1}
982 ; VLX-NEXT: kmovq %k0, %rax
983 ; VLX-NEXT: vzeroupper
984 ; VLX-NEXT: retq
985 ;
986 ; NoVLX-LABEL: test_masked_vpcmpeqw_v32i1_v64i1_mask:
987 ; NoVLX: # %bb.0: # %entry
988 ; NoVLX-NEXT: vpcmpeqw %ymm1, %ymm0, %ymm2
989 ; NoVLX-NEXT: vpmovsxwd %ymm2, %zmm2
990 ; NoVLX-NEXT: vptestmd %zmm2, %zmm2, %k0
991 ; NoVLX-NEXT: kmovw %k0, %eax
992 ; NoVLX-NEXT: andl %edi, %eax
993 ; NoVLX-NEXT: shrl $16, %edi
994 ; NoVLX-NEXT: vextracti64x4 $1, %zmm0, %ymm0
995 ; NoVLX-NEXT: vextracti64x4 $1, %zmm1, %ymm1
996 ; NoVLX-NEXT: vpcmpeqw %ymm1, %ymm0, %ymm0
997 ; NoVLX-NEXT: vpmovsxwd %ymm0, %zmm0
998 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
999 ; NoVLX-NEXT: kmovw %k0, %ecx
1000 ; NoVLX-NEXT: andl %edi, %ecx
1001 ; NoVLX-NEXT: shll $16, %ecx
1002 ; NoVLX-NEXT: movzwl %ax, %eax
1003 ; NoVLX-NEXT: orl %ecx, %eax
1004 ; NoVLX-NEXT: vzeroupper
1005 ; NoVLX-NEXT: retq
1006 entry:
1007 %0 = bitcast <8 x i64> %__a to <32 x i16>
1008 %1 = bitcast <8 x i64> %__b to <32 x i16>
1009 %2 = icmp eq <32 x i16> %0, %1
1010 %3 = bitcast i32 %__u to <32 x i1>
1011 %4 = and <32 x i1> %2, %3
1012 %5 = shufflevector <32 x i1> %4, <32 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63>
1013 %6 = bitcast <64 x i1> %5 to i64
1014 ret i64 %6
1015 }
1017 define zeroext i64 @test_masked_vpcmpeqw_v32i1_v64i1_mask_mem(i32 zeroext %__u, <8 x i64> %__a, <8 x i64>* %__b) local_unnamed_addr {
1018 ; VLX-LABEL: test_masked_vpcmpeqw_v32i1_v64i1_mask_mem:
1019 ; VLX: # %bb.0: # %entry
1020 ; VLX-NEXT: kmovd %edi, %k1
1021 ; VLX-NEXT: vpcmpeqw (%rsi), %zmm0, %k0 {%k1}
1022 ; VLX-NEXT: kmovq %k0, %rax
1023 ; VLX-NEXT: vzeroupper
1024 ; VLX-NEXT: retq
1025 ;
1026 ; NoVLX-LABEL: test_masked_vpcmpeqw_v32i1_v64i1_mask_mem:
1027 ; NoVLX: # %bb.0: # %entry
1028 ; NoVLX-NEXT: vpcmpeqw (%rsi), %ymm0, %ymm1
1029 ; NoVLX-NEXT: vpmovsxwd %ymm1, %zmm1
1030 ; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
1031 ; NoVLX-NEXT: kmovw %k0, %eax
1032 ; NoVLX-NEXT: andl %edi, %eax
1033 ; NoVLX-NEXT: shrl $16, %edi
1034 ; NoVLX-NEXT: vextracti64x4 $1, %zmm0, %ymm0
1035 ; NoVLX-NEXT: vpcmpeqw 32(%rsi), %ymm0, %ymm0
1036 ; NoVLX-NEXT: vpmovsxwd %ymm0, %zmm0
1037 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
1038 ; NoVLX-NEXT: kmovw %k0, %ecx
1039 ; NoVLX-NEXT: andl %edi, %ecx
1040 ; NoVLX-NEXT: shll $16, %ecx
1041 ; NoVLX-NEXT: movzwl %ax, %eax
1042 ; NoVLX-NEXT: orl %ecx, %eax
1043 ; NoVLX-NEXT: vzeroupper
1044 ; NoVLX-NEXT: retq
1045 entry:
1046 %0 = bitcast <8 x i64> %__a to <32 x i16>
1047 %load = load <8 x i64>, <8 x i64>* %__b
1048 %1 = bitcast <8 x i64> %load to <32 x i16>
1049 %2 = icmp eq <32 x i16> %0, %1
1050 %3 = bitcast i32 %__u to <32 x i1>
1051 %4 = and <32 x i1> %2, %3
1052 %5 = shufflevector <32 x i1> %4, <32 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63>
1053 %6 = bitcast <64 x i1> %5 to i64
1054 ret i64 %6
1055 }
1058 define zeroext i8 @test_vpcmpeqd_v4i1_v8i1_mask(<2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
1059 ; VLX-LABEL: test_vpcmpeqd_v4i1_v8i1_mask:
1060 ; VLX: # %bb.0: # %entry
1061 ; VLX-NEXT: vpcmpeqd %xmm1, %xmm0, %k0
1062 ; VLX-NEXT: kmovd %k0, %eax
1063 ; VLX-NEXT: # kill: def $al killed $al killed $eax
1064 ; VLX-NEXT: retq
1065 ;
1066 ; NoVLX-LABEL: test_vpcmpeqd_v4i1_v8i1_mask:
1067 ; NoVLX: # %bb.0: # %entry
1068 ; NoVLX-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
1069 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
1070 ; NoVLX-NEXT: vpcmpeqd %zmm1, %zmm0, %k0
1071 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
1072 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
1073 ; NoVLX-NEXT: kmovw %k0, %eax
1074 ; NoVLX-NEXT: # kill: def $al killed $al killed $eax
1075 ; NoVLX-NEXT: vzeroupper
1076 ; NoVLX-NEXT: retq
1077 entry:
1078 %0 = bitcast <2 x i64> %__a to <4 x i32>
1079 %1 = bitcast <2 x i64> %__b to <4 x i32>
1080 %2 = icmp eq <4 x i32> %0, %1
1081 %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
1082 %4 = bitcast <8 x i1> %3 to i8
1083 ret i8 %4
1084 }
1086 define zeroext i8 @test_vpcmpeqd_v4i1_v8i1_mask_mem(<2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
1087 ; VLX-LABEL: test_vpcmpeqd_v4i1_v8i1_mask_mem:
1088 ; VLX: # %bb.0: # %entry
1089 ; VLX-NEXT: vpcmpeqd (%rdi), %xmm0, %k0
1090 ; VLX-NEXT: kmovd %k0, %eax
1091 ; VLX-NEXT: # kill: def $al killed $al killed $eax
1092 ; VLX-NEXT: retq
1093 ;
1094 ; NoVLX-LABEL: test_vpcmpeqd_v4i1_v8i1_mask_mem:
1095 ; NoVLX: # %bb.0: # %entry
1096 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
1097 ; NoVLX-NEXT: vmovdqa (%rdi), %xmm1
1098 ; NoVLX-NEXT: vpcmpeqd %zmm1, %zmm0, %k0
1099 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
1100 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
1101 ; NoVLX-NEXT: kmovw %k0, %eax
1102 ; NoVLX-NEXT: # kill: def $al killed $al killed $eax
1103 ; NoVLX-NEXT: vzeroupper
1104 ; NoVLX-NEXT: retq
1105 entry:
1106 %0 = bitcast <2 x i64> %__a to <4 x i32>
1107 %load = load <2 x i64>, <2 x i64>* %__b
1108 %1 = bitcast <2 x i64> %load to <4 x i32>
1109 %2 = icmp eq <4 x i32> %0, %1
1110 %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
1111 %4 = bitcast <8 x i1> %3 to i8
1112 ret i8 %4
1113 }
1115 define zeroext i8 @test_masked_vpcmpeqd_v4i1_v8i1_mask(i8 zeroext %__u, <2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
1116 ; VLX-LABEL: test_masked_vpcmpeqd_v4i1_v8i1_mask:
1117 ; VLX: # %bb.0: # %entry
1118 ; VLX-NEXT: kmovd %edi, %k1
1119 ; VLX-NEXT: vpcmpeqd %xmm1, %xmm0, %k0 {%k1}
1120 ; VLX-NEXT: kmovd %k0, %eax
1121 ; VLX-NEXT: # kill: def $al killed $al killed $eax
1122 ; VLX-NEXT: retq
1123 ;
1124 ; NoVLX-LABEL: test_masked_vpcmpeqd_v4i1_v8i1_mask:
1125 ; NoVLX: # %bb.0: # %entry
1126 ; NoVLX-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
1127 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
1128 ; NoVLX-NEXT: kmovw %edi, %k1
1129 ; NoVLX-NEXT: vpcmpeqd %zmm1, %zmm0, %k0 {%k1}
1130 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
1131 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
1132 ; NoVLX-NEXT: kmovw %k0, %eax
1133 ; NoVLX-NEXT: # kill: def $al killed $al killed $eax
1134 ; NoVLX-NEXT: vzeroupper
1135 ; NoVLX-NEXT: retq
1136 entry:
1137 %0 = bitcast <2 x i64> %__a to <4 x i32>
1138 %1 = bitcast <2 x i64> %__b to <4 x i32>
1139 %2 = icmp eq <4 x i32> %0, %1
1140 %3 = bitcast i8 %__u to <8 x i1>
1141 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
1142 %4 = and <4 x i1> %2, %extract.i
1143 %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
1144 %6 = bitcast <8 x i1> %5 to i8
1145 ret i8 %6
1146 }
1148 define zeroext i8 @test_masked_vpcmpeqd_v4i1_v8i1_mask_mem(i8 zeroext %__u, <2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
1149 ; VLX-LABEL: test_masked_vpcmpeqd_v4i1_v8i1_mask_mem:
1150 ; VLX: # %bb.0: # %entry
1151 ; VLX-NEXT: kmovd %edi, %k1
1152 ; VLX-NEXT: vpcmpeqd (%rsi), %xmm0, %k0 {%k1}
1153 ; VLX-NEXT: kmovd %k0, %eax
1154 ; VLX-NEXT: # kill: def $al killed $al killed $eax
1155 ; VLX-NEXT: retq
1156 ;
1157 ; NoVLX-LABEL: test_masked_vpcmpeqd_v4i1_v8i1_mask_mem:
1158 ; NoVLX: # %bb.0: # %entry
1159 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
1160 ; NoVLX-NEXT: vmovdqa (%rsi), %xmm1
1161 ; NoVLX-NEXT: kmovw %edi, %k1
1162 ; NoVLX-NEXT: vpcmpeqd %zmm1, %zmm0, %k0 {%k1}
1163 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
1164 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
1165 ; NoVLX-NEXT: kmovw %k0, %eax
1166 ; NoVLX-NEXT: # kill: def $al killed $al killed $eax
1167 ; NoVLX-NEXT: vzeroupper
1168 ; NoVLX-NEXT: retq
1169 entry:
1170 %0 = bitcast <2 x i64> %__a to <4 x i32>
1171 %load = load <2 x i64>, <2 x i64>* %__b
1172 %1 = bitcast <2 x i64> %load to <4 x i32>
1173 %2 = icmp eq <4 x i32> %0, %1
1174 %3 = bitcast i8 %__u to <8 x i1>
1175 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
1176 %4 = and <4 x i1> %2, %extract.i
1177 %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
1178 %6 = bitcast <8 x i1> %5 to i8
1179 ret i8 %6
1180 }
1183 define zeroext i8 @test_vpcmpeqd_v4i1_v8i1_mask_mem_b(<2 x i64> %__a, i32* %__b) local_unnamed_addr {
1184 ; VLX-LABEL: test_vpcmpeqd_v4i1_v8i1_mask_mem_b:
1185 ; VLX: # %bb.0: # %entry
1186 ; VLX-NEXT: vpcmpeqd (%rdi){1to4}, %xmm0, %k0
1187 ; VLX-NEXT: kmovd %k0, %eax
1188 ; VLX-NEXT: # kill: def $al killed $al killed $eax
1189 ; VLX-NEXT: retq
1190 ;
1191 ; NoVLX-LABEL: test_vpcmpeqd_v4i1_v8i1_mask_mem_b:
1192 ; NoVLX: # %bb.0: # %entry
1193 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
1194 ; NoVLX-NEXT: vpcmpeqd (%rdi){1to16}, %zmm0, %k0
1195 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
1196 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
1197 ; NoVLX-NEXT: kmovw %k0, %eax
1198 ; NoVLX-NEXT: # kill: def $al killed $al killed $eax
1199 ; NoVLX-NEXT: vzeroupper
1200 ; NoVLX-NEXT: retq
1201 entry:
1202 %0 = bitcast <2 x i64> %__a to <4 x i32>
1203 %load = load i32, i32* %__b
1204 %vec = insertelement <4 x i32> undef, i32 %load, i32 0
1205 %1 = shufflevector <4 x i32> %vec, <4 x i32> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
1206 %2 = icmp eq <4 x i32> %0, %1
1207 %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
1208 %4 = bitcast <8 x i1> %3 to i8
1209 ret i8 %4
1210 }
1212 define zeroext i8 @test_masked_vpcmpeqd_v4i1_v8i1_mask_mem_b(i8 zeroext %__u, <2 x i64> %__a, i32* %__b) local_unnamed_addr {
1213 ; VLX-LABEL: test_masked_vpcmpeqd_v4i1_v8i1_mask_mem_b:
1214 ; VLX: # %bb.0: # %entry
1215 ; VLX-NEXT: kmovd %edi, %k1
1216 ; VLX-NEXT: vpcmpeqd (%rsi){1to4}, %xmm0, %k0 {%k1}
1217 ; VLX-NEXT: kmovd %k0, %eax
1218 ; VLX-NEXT: # kill: def $al killed $al killed $eax
1219 ; VLX-NEXT: retq
1220 ;
1221 ; NoVLX-LABEL: test_masked_vpcmpeqd_v4i1_v8i1_mask_mem_b:
1222 ; NoVLX: # %bb.0: # %entry
1223 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
1224 ; NoVLX-NEXT: kmovw %edi, %k1
1225 ; NoVLX-NEXT: vpcmpeqd (%rsi){1to16}, %zmm0, %k0 {%k1}
1226 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
1227 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
1228 ; NoVLX-NEXT: kmovw %k0, %eax
1229 ; NoVLX-NEXT: # kill: def $al killed $al killed $eax
1230 ; NoVLX-NEXT: vzeroupper
1231 ; NoVLX-NEXT: retq
1232 entry:
1233 %0 = bitcast <2 x i64> %__a to <4 x i32>
1234 %load = load i32, i32* %__b
1235 %vec = insertelement <4 x i32> undef, i32 %load, i32 0
1236 %1 = shufflevector <4 x i32> %vec, <4 x i32> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
1237 %2 = icmp eq <4 x i32> %0, %1
1238 %3 = bitcast i8 %__u to <8 x i1>
1239 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
1240 %4 = and <4 x i1> %extract.i, %2
1241 %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
1242 %6 = bitcast <8 x i1> %5 to i8
1243 ret i8 %6
1244 }
1247 define zeroext i16 @test_vpcmpeqd_v4i1_v16i1_mask(<2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
1248 ; VLX-LABEL: test_vpcmpeqd_v4i1_v16i1_mask:
1249 ; VLX: # %bb.0: # %entry
1250 ; VLX-NEXT: vpcmpeqd %xmm1, %xmm0, %k0
1251 ; VLX-NEXT: kmovd %k0, %eax
1252 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
1253 ; VLX-NEXT: retq
1254 ;
1255 ; NoVLX-LABEL: test_vpcmpeqd_v4i1_v16i1_mask:
1256 ; NoVLX: # %bb.0: # %entry
1257 ; NoVLX-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
1258 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
1259 ; NoVLX-NEXT: vpcmpeqd %zmm1, %zmm0, %k0
1260 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
1261 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
1262 ; NoVLX-NEXT: kmovw %k0, %eax
1263 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
1264 ; NoVLX-NEXT: vzeroupper
1265 ; NoVLX-NEXT: retq
1266 entry:
1267 %0 = bitcast <2 x i64> %__a to <4 x i32>
1268 %1 = bitcast <2 x i64> %__b to <4 x i32>
1269 %2 = icmp eq <4 x i32> %0, %1
1270 %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
1271 %4 = bitcast <16 x i1> %3 to i16
1272 ret i16 %4
1273 }
1275 define zeroext i16 @test_vpcmpeqd_v4i1_v16i1_mask_mem(<2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
1276 ; VLX-LABEL: test_vpcmpeqd_v4i1_v16i1_mask_mem:
1277 ; VLX: # %bb.0: # %entry
1278 ; VLX-NEXT: vpcmpeqd (%rdi), %xmm0, %k0
1279 ; VLX-NEXT: kmovd %k0, %eax
1280 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
1281 ; VLX-NEXT: retq
1282 ;
1283 ; NoVLX-LABEL: test_vpcmpeqd_v4i1_v16i1_mask_mem:
1284 ; NoVLX: # %bb.0: # %entry
1285 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
1286 ; NoVLX-NEXT: vmovdqa (%rdi), %xmm1
1287 ; NoVLX-NEXT: vpcmpeqd %zmm1, %zmm0, %k0
1288 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
1289 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
1290 ; NoVLX-NEXT: kmovw %k0, %eax
1291 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
1292 ; NoVLX-NEXT: vzeroupper
1293 ; NoVLX-NEXT: retq
1294 entry:
1295 %0 = bitcast <2 x i64> %__a to <4 x i32>
1296 %load = load <2 x i64>, <2 x i64>* %__b
1297 %1 = bitcast <2 x i64> %load to <4 x i32>
1298 %2 = icmp eq <4 x i32> %0, %1
1299 %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
1300 %4 = bitcast <16 x i1> %3 to i16
1301 ret i16 %4
1302 }
1304 define zeroext i16 @test_masked_vpcmpeqd_v4i1_v16i1_mask(i8 zeroext %__u, <2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
1305 ; VLX-LABEL: test_masked_vpcmpeqd_v4i1_v16i1_mask:
1306 ; VLX: # %bb.0: # %entry
1307 ; VLX-NEXT: kmovd %edi, %k1
1308 ; VLX-NEXT: vpcmpeqd %xmm1, %xmm0, %k0 {%k1}
1309 ; VLX-NEXT: kmovd %k0, %eax
1310 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
1311 ; VLX-NEXT: retq
1312 ;
1313 ; NoVLX-LABEL: test_masked_vpcmpeqd_v4i1_v16i1_mask:
1314 ; NoVLX: # %bb.0: # %entry
1315 ; NoVLX-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
1316 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
1317 ; NoVLX-NEXT: kmovw %edi, %k1
1318 ; NoVLX-NEXT: vpcmpeqd %zmm1, %zmm0, %k0 {%k1}
1319 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
1320 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
1321 ; NoVLX-NEXT: kmovw %k0, %eax
1322 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
1323 ; NoVLX-NEXT: vzeroupper
1324 ; NoVLX-NEXT: retq
1325 entry:
1326 %0 = bitcast <2 x i64> %__a to <4 x i32>
1327 %1 = bitcast <2 x i64> %__b to <4 x i32>
1328 %2 = icmp eq <4 x i32> %0, %1
1329 %3 = bitcast i8 %__u to <8 x i1>
1330 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
1331 %4 = and <4 x i1> %2, %extract.i
1332 %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
1333 %6 = bitcast <16 x i1> %5 to i16
1334 ret i16 %6
1335 }
1337 define zeroext i16 @test_masked_vpcmpeqd_v4i1_v16i1_mask_mem(i8 zeroext %__u, <2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
1338 ; VLX-LABEL: test_masked_vpcmpeqd_v4i1_v16i1_mask_mem:
1339 ; VLX: # %bb.0: # %entry
1340 ; VLX-NEXT: kmovd %edi, %k1
1341 ; VLX-NEXT: vpcmpeqd (%rsi), %xmm0, %k0 {%k1}
1342 ; VLX-NEXT: kmovd %k0, %eax
1343 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
1344 ; VLX-NEXT: retq
1345 ;
1346 ; NoVLX-LABEL: test_masked_vpcmpeqd_v4i1_v16i1_mask_mem:
1347 ; NoVLX: # %bb.0: # %entry
1348 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
1349 ; NoVLX-NEXT: vmovdqa (%rsi), %xmm1
1350 ; NoVLX-NEXT: kmovw %edi, %k1
1351 ; NoVLX-NEXT: vpcmpeqd %zmm1, %zmm0, %k0 {%k1}
1352 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
1353 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
1354 ; NoVLX-NEXT: kmovw %k0, %eax
1355 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
1356 ; NoVLX-NEXT: vzeroupper
1357 ; NoVLX-NEXT: retq
1358 entry:
1359 %0 = bitcast <2 x i64> %__a to <4 x i32>
1360 %load = load <2 x i64>, <2 x i64>* %__b
1361 %1 = bitcast <2 x i64> %load to <4 x i32>
1362 %2 = icmp eq <4 x i32> %0, %1
1363 %3 = bitcast i8 %__u to <8 x i1>
1364 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
1365 %4 = and <4 x i1> %2, %extract.i
1366 %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
1367 %6 = bitcast <16 x i1> %5 to i16
1368 ret i16 %6
1369 }
1372 define zeroext i16 @test_vpcmpeqd_v4i1_v16i1_mask_mem_b(<2 x i64> %__a, i32* %__b) local_unnamed_addr {
1373 ; VLX-LABEL: test_vpcmpeqd_v4i1_v16i1_mask_mem_b:
1374 ; VLX: # %bb.0: # %entry
1375 ; VLX-NEXT: vpcmpeqd (%rdi){1to4}, %xmm0, %k0
1376 ; VLX-NEXT: kmovd %k0, %eax
1377 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
1378 ; VLX-NEXT: retq
1379 ;
1380 ; NoVLX-LABEL: test_vpcmpeqd_v4i1_v16i1_mask_mem_b:
1381 ; NoVLX: # %bb.0: # %entry
1382 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
1383 ; NoVLX-NEXT: vpcmpeqd (%rdi){1to16}, %zmm0, %k0
1384 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
1385 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
1386 ; NoVLX-NEXT: kmovw %k0, %eax
1387 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
1388 ; NoVLX-NEXT: vzeroupper
1389 ; NoVLX-NEXT: retq
1390 entry:
1391 %0 = bitcast <2 x i64> %__a to <4 x i32>
1392 %load = load i32, i32* %__b
1393 %vec = insertelement <4 x i32> undef, i32 %load, i32 0
1394 %1 = shufflevector <4 x i32> %vec, <4 x i32> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
1395 %2 = icmp eq <4 x i32> %0, %1
1396 %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
1397 %4 = bitcast <16 x i1> %3 to i16
1398 ret i16 %4
1399 }
1401 define zeroext i16 @test_masked_vpcmpeqd_v4i1_v16i1_mask_mem_b(i8 zeroext %__u, <2 x i64> %__a, i32* %__b) local_unnamed_addr {
1402 ; VLX-LABEL: test_masked_vpcmpeqd_v4i1_v16i1_mask_mem_b:
1403 ; VLX: # %bb.0: # %entry
1404 ; VLX-NEXT: kmovd %edi, %k1
1405 ; VLX-NEXT: vpcmpeqd (%rsi){1to4}, %xmm0, %k0 {%k1}
1406 ; VLX-NEXT: kmovd %k0, %eax
1407 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
1408 ; VLX-NEXT: retq
1409 ;
1410 ; NoVLX-LABEL: test_masked_vpcmpeqd_v4i1_v16i1_mask_mem_b:
1411 ; NoVLX: # %bb.0: # %entry
1412 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
1413 ; NoVLX-NEXT: kmovw %edi, %k1
1414 ; NoVLX-NEXT: vpcmpeqd (%rsi){1to16}, %zmm0, %k0 {%k1}
1415 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
1416 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
1417 ; NoVLX-NEXT: kmovw %k0, %eax
1418 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
1419 ; NoVLX-NEXT: vzeroupper
1420 ; NoVLX-NEXT: retq
1421 entry:
1422 %0 = bitcast <2 x i64> %__a to <4 x i32>
1423 %load = load i32, i32* %__b
1424 %vec = insertelement <4 x i32> undef, i32 %load, i32 0
1425 %1 = shufflevector <4 x i32> %vec, <4 x i32> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
1426 %2 = icmp eq <4 x i32> %0, %1
1427 %3 = bitcast i8 %__u to <8 x i1>
1428 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
1429 %4 = and <4 x i1> %extract.i, %2
1430 %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
1431 %6 = bitcast <16 x i1> %5 to i16
1432 ret i16 %6
1433 }
1436 define zeroext i32 @test_vpcmpeqd_v4i1_v32i1_mask(<2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
1437 ; VLX-LABEL: test_vpcmpeqd_v4i1_v32i1_mask:
1438 ; VLX: # %bb.0: # %entry
1439 ; VLX-NEXT: vpcmpeqd %xmm1, %xmm0, %k0
1440 ; VLX-NEXT: kmovd %k0, %eax
1441 ; VLX-NEXT: retq
1442 ;
1443 ; NoVLX-LABEL: test_vpcmpeqd_v4i1_v32i1_mask:
1444 ; NoVLX: # %bb.0: # %entry
1445 ; NoVLX-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
1446 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
1447 ; NoVLX-NEXT: vpcmpeqd %zmm1, %zmm0, %k0
1448 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
1449 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
1450 ; NoVLX-NEXT: kmovw %k0, %eax
1451 ; NoVLX-NEXT: vzeroupper
1452 ; NoVLX-NEXT: retq
1453 entry:
1454 %0 = bitcast <2 x i64> %__a to <4 x i32>
1455 %1 = bitcast <2 x i64> %__b to <4 x i32>
1456 %2 = icmp eq <4 x i32> %0, %1
1457 %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
1458 %4 = bitcast <32 x i1> %3 to i32
1459 ret i32 %4
1460 }
1462 define zeroext i32 @test_vpcmpeqd_v4i1_v32i1_mask_mem(<2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
1463 ; VLX-LABEL: test_vpcmpeqd_v4i1_v32i1_mask_mem:
1464 ; VLX: # %bb.0: # %entry
1465 ; VLX-NEXT: vpcmpeqd (%rdi), %xmm0, %k0
1466 ; VLX-NEXT: kmovd %k0, %eax
1467 ; VLX-NEXT: retq
1468 ;
1469 ; NoVLX-LABEL: test_vpcmpeqd_v4i1_v32i1_mask_mem:
1470 ; NoVLX: # %bb.0: # %entry
1471 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
1472 ; NoVLX-NEXT: vmovdqa (%rdi), %xmm1
1473 ; NoVLX-NEXT: vpcmpeqd %zmm1, %zmm0, %k0
1474 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
1475 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
1476 ; NoVLX-NEXT: kmovw %k0, %eax
1477 ; NoVLX-NEXT: vzeroupper
1478 ; NoVLX-NEXT: retq
1479 entry:
1480 %0 = bitcast <2 x i64> %__a to <4 x i32>
1481 %load = load <2 x i64>, <2 x i64>* %__b
1482 %1 = bitcast <2 x i64> %load to <4 x i32>
1483 %2 = icmp eq <4 x i32> %0, %1
1484 %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
1485 %4 = bitcast <32 x i1> %3 to i32
1486 ret i32 %4
1487 }
1489 define zeroext i32 @test_masked_vpcmpeqd_v4i1_v32i1_mask(i8 zeroext %__u, <2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
1490 ; VLX-LABEL: test_masked_vpcmpeqd_v4i1_v32i1_mask:
1491 ; VLX: # %bb.0: # %entry
1492 ; VLX-NEXT: kmovd %edi, %k1
1493 ; VLX-NEXT: vpcmpeqd %xmm1, %xmm0, %k0 {%k1}
1494 ; VLX-NEXT: kmovd %k0, %eax
1495 ; VLX-NEXT: retq
1496 ;
1497 ; NoVLX-LABEL: test_masked_vpcmpeqd_v4i1_v32i1_mask:
1498 ; NoVLX: # %bb.0: # %entry
1499 ; NoVLX-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
1500 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
1501 ; NoVLX-NEXT: kmovw %edi, %k1
1502 ; NoVLX-NEXT: vpcmpeqd %zmm1, %zmm0, %k0 {%k1}
1503 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
1504 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
1505 ; NoVLX-NEXT: kmovw %k0, %eax
1506 ; NoVLX-NEXT: vzeroupper
1507 ; NoVLX-NEXT: retq
1508 entry:
1509 %0 = bitcast <2 x i64> %__a to <4 x i32>
1510 %1 = bitcast <2 x i64> %__b to <4 x i32>
1511 %2 = icmp eq <4 x i32> %0, %1
1512 %3 = bitcast i8 %__u to <8 x i1>
1513 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
1514 %4 = and <4 x i1> %2, %extract.i
1515 %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
1516 %6 = bitcast <32 x i1> %5 to i32
1517 ret i32 %6
1518 }
1520 define zeroext i32 @test_masked_vpcmpeqd_v4i1_v32i1_mask_mem(i8 zeroext %__u, <2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
1521 ; VLX-LABEL: test_masked_vpcmpeqd_v4i1_v32i1_mask_mem:
1522 ; VLX: # %bb.0: # %entry
1523 ; VLX-NEXT: kmovd %edi, %k1
1524 ; VLX-NEXT: vpcmpeqd (%rsi), %xmm0, %k0 {%k1}
1525 ; VLX-NEXT: kmovd %k0, %eax
1526 ; VLX-NEXT: retq
1527 ;
1528 ; NoVLX-LABEL: test_masked_vpcmpeqd_v4i1_v32i1_mask_mem:
1529 ; NoVLX: # %bb.0: # %entry
1530 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
1531 ; NoVLX-NEXT: vmovdqa (%rsi), %xmm1
1532 ; NoVLX-NEXT: kmovw %edi, %k1
1533 ; NoVLX-NEXT: vpcmpeqd %zmm1, %zmm0, %k0 {%k1}
1534 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
1535 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
1536 ; NoVLX-NEXT: kmovw %k0, %eax
1537 ; NoVLX-NEXT: vzeroupper
1538 ; NoVLX-NEXT: retq
1539 entry:
1540 %0 = bitcast <2 x i64> %__a to <4 x i32>
1541 %load = load <2 x i64>, <2 x i64>* %__b
1542 %1 = bitcast <2 x i64> %load to <4 x i32>
1543 %2 = icmp eq <4 x i32> %0, %1
1544 %3 = bitcast i8 %__u to <8 x i1>
1545 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
1546 %4 = and <4 x i1> %2, %extract.i
1547 %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
1548 %6 = bitcast <32 x i1> %5 to i32
1549 ret i32 %6
1550 }
1553 define zeroext i32 @test_vpcmpeqd_v4i1_v32i1_mask_mem_b(<2 x i64> %__a, i32* %__b) local_unnamed_addr {
1554 ; VLX-LABEL: test_vpcmpeqd_v4i1_v32i1_mask_mem_b:
1555 ; VLX: # %bb.0: # %entry
1556 ; VLX-NEXT: vpcmpeqd (%rdi){1to4}, %xmm0, %k0
1557 ; VLX-NEXT: kmovd %k0, %eax
1558 ; VLX-NEXT: retq
1559 ;
1560 ; NoVLX-LABEL: test_vpcmpeqd_v4i1_v32i1_mask_mem_b:
1561 ; NoVLX: # %bb.0: # %entry
1562 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
1563 ; NoVLX-NEXT: vpcmpeqd (%rdi){1to16}, %zmm0, %k0
1564 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
1565 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
1566 ; NoVLX-NEXT: kmovw %k0, %eax
1567 ; NoVLX-NEXT: vzeroupper
1568 ; NoVLX-NEXT: retq
1569 entry:
1570 %0 = bitcast <2 x i64> %__a to <4 x i32>
1571 %load = load i32, i32* %__b
1572 %vec = insertelement <4 x i32> undef, i32 %load, i32 0
1573 %1 = shufflevector <4 x i32> %vec, <4 x i32> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
1574 %2 = icmp eq <4 x i32> %0, %1
1575 %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
1576 %4 = bitcast <32 x i1> %3 to i32
1577 ret i32 %4
1578 }
1580 define zeroext i32 @test_masked_vpcmpeqd_v4i1_v32i1_mask_mem_b(i8 zeroext %__u, <2 x i64> %__a, i32* %__b) local_unnamed_addr {
1581 ; VLX-LABEL: test_masked_vpcmpeqd_v4i1_v32i1_mask_mem_b:
1582 ; VLX: # %bb.0: # %entry
1583 ; VLX-NEXT: kmovd %edi, %k1
1584 ; VLX-NEXT: vpcmpeqd (%rsi){1to4}, %xmm0, %k0 {%k1}
1585 ; VLX-NEXT: kmovd %k0, %eax
1586 ; VLX-NEXT: retq
1587 ;
1588 ; NoVLX-LABEL: test_masked_vpcmpeqd_v4i1_v32i1_mask_mem_b:
1589 ; NoVLX: # %bb.0: # %entry
1590 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
1591 ; NoVLX-NEXT: kmovw %edi, %k1
1592 ; NoVLX-NEXT: vpcmpeqd (%rsi){1to16}, %zmm0, %k0 {%k1}
1593 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
1594 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
1595 ; NoVLX-NEXT: kmovw %k0, %eax
1596 ; NoVLX-NEXT: vzeroupper
1597 ; NoVLX-NEXT: retq
1598 entry:
1599 %0 = bitcast <2 x i64> %__a to <4 x i32>
1600 %load = load i32, i32* %__b
1601 %vec = insertelement <4 x i32> undef, i32 %load, i32 0
1602 %1 = shufflevector <4 x i32> %vec, <4 x i32> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
1603 %2 = icmp eq <4 x i32> %0, %1
1604 %3 = bitcast i8 %__u to <8 x i1>
1605 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
1606 %4 = and <4 x i1> %extract.i, %2
1607 %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
1608 %6 = bitcast <32 x i1> %5 to i32
1609 ret i32 %6
1610 }
1613 define zeroext i64 @test_vpcmpeqd_v4i1_v64i1_mask(<2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
1614 ; VLX-LABEL: test_vpcmpeqd_v4i1_v64i1_mask:
1615 ; VLX: # %bb.0: # %entry
1616 ; VLX-NEXT: vpcmpeqd %xmm1, %xmm0, %k0
1617 ; VLX-NEXT: kmovq %k0, %rax
1618 ; VLX-NEXT: retq
1619 ;
1620 ; NoVLX-LABEL: test_vpcmpeqd_v4i1_v64i1_mask:
1621 ; NoVLX: # %bb.0: # %entry
1622 ; NoVLX-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
1623 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
1624 ; NoVLX-NEXT: vpcmpeqd %zmm1, %zmm0, %k0
1625 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
1626 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
1627 ; NoVLX-NEXT: kmovw %k0, %eax
1628 ; NoVLX-NEXT: vzeroupper
1629 ; NoVLX-NEXT: retq
1630 entry:
1631 %0 = bitcast <2 x i64> %__a to <4 x i32>
1632 %1 = bitcast <2 x i64> %__b to <4 x i32>
1633 %2 = icmp eq <4 x i32> %0, %1
1634 %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
1635 %4 = bitcast <64 x i1> %3 to i64
1636 ret i64 %4
1637 }
1639 define zeroext i64 @test_vpcmpeqd_v4i1_v64i1_mask_mem(<2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
1640 ; VLX-LABEL: test_vpcmpeqd_v4i1_v64i1_mask_mem:
1641 ; VLX: # %bb.0: # %entry
1642 ; VLX-NEXT: vpcmpeqd (%rdi), %xmm0, %k0
1643 ; VLX-NEXT: kmovq %k0, %rax
1644 ; VLX-NEXT: retq
1645 ;
1646 ; NoVLX-LABEL: test_vpcmpeqd_v4i1_v64i1_mask_mem:
1647 ; NoVLX: # %bb.0: # %entry
1648 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
1649 ; NoVLX-NEXT: vmovdqa (%rdi), %xmm1
1650 ; NoVLX-NEXT: vpcmpeqd %zmm1, %zmm0, %k0
1651 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
1652 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
1653 ; NoVLX-NEXT: kmovw %k0, %eax
1654 ; NoVLX-NEXT: vzeroupper
1655 ; NoVLX-NEXT: retq
1656 entry:
1657 %0 = bitcast <2 x i64> %__a to <4 x i32>
1658 %load = load <2 x i64>, <2 x i64>* %__b
1659 %1 = bitcast <2 x i64> %load to <4 x i32>
1660 %2 = icmp eq <4 x i32> %0, %1
1661 %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
1662 %4 = bitcast <64 x i1> %3 to i64
1663 ret i64 %4
1664 }
1666 define zeroext i64 @test_masked_vpcmpeqd_v4i1_v64i1_mask(i8 zeroext %__u, <2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
1667 ; VLX-LABEL: test_masked_vpcmpeqd_v4i1_v64i1_mask:
1668 ; VLX: # %bb.0: # %entry
1669 ; VLX-NEXT: kmovd %edi, %k1
1670 ; VLX-NEXT: vpcmpeqd %xmm1, %xmm0, %k0 {%k1}
1671 ; VLX-NEXT: kmovq %k0, %rax
1672 ; VLX-NEXT: retq
1673 ;
1674 ; NoVLX-LABEL: test_masked_vpcmpeqd_v4i1_v64i1_mask:
1675 ; NoVLX: # %bb.0: # %entry
1676 ; NoVLX-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
1677 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
1678 ; NoVLX-NEXT: kmovw %edi, %k1
1679 ; NoVLX-NEXT: vpcmpeqd %zmm1, %zmm0, %k0 {%k1}
1680 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
1681 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
1682 ; NoVLX-NEXT: kmovw %k0, %eax
1683 ; NoVLX-NEXT: vzeroupper
1684 ; NoVLX-NEXT: retq
1685 entry:
1686 %0 = bitcast <2 x i64> %__a to <4 x i32>
1687 %1 = bitcast <2 x i64> %__b to <4 x i32>
1688 %2 = icmp eq <4 x i32> %0, %1
1689 %3 = bitcast i8 %__u to <8 x i1>
1690 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
1691 %4 = and <4 x i1> %2, %extract.i
1692 %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
1693 %6 = bitcast <64 x i1> %5 to i64
1694 ret i64 %6
1695 }
1697 define zeroext i64 @test_masked_vpcmpeqd_v4i1_v64i1_mask_mem(i8 zeroext %__u, <2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
1698 ; VLX-LABEL: test_masked_vpcmpeqd_v4i1_v64i1_mask_mem:
1699 ; VLX: # %bb.0: # %entry
1700 ; VLX-NEXT: kmovd %edi, %k1
1701 ; VLX-NEXT: vpcmpeqd (%rsi), %xmm0, %k0 {%k1}
1702 ; VLX-NEXT: kmovq %k0, %rax
1703 ; VLX-NEXT: retq
1704 ;
1705 ; NoVLX-LABEL: test_masked_vpcmpeqd_v4i1_v64i1_mask_mem:
1706 ; NoVLX: # %bb.0: # %entry
1707 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
1708 ; NoVLX-NEXT: vmovdqa (%rsi), %xmm1
1709 ; NoVLX-NEXT: kmovw %edi, %k1
1710 ; NoVLX-NEXT: vpcmpeqd %zmm1, %zmm0, %k0 {%k1}
1711 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
1712 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
1713 ; NoVLX-NEXT: kmovw %k0, %eax
1714 ; NoVLX-NEXT: vzeroupper
1715 ; NoVLX-NEXT: retq
1716 entry:
1717 %0 = bitcast <2 x i64> %__a to <4 x i32>
1718 %load = load <2 x i64>, <2 x i64>* %__b
1719 %1 = bitcast <2 x i64> %load to <4 x i32>
1720 %2 = icmp eq <4 x i32> %0, %1
1721 %3 = bitcast i8 %__u to <8 x i1>
1722 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
1723 %4 = and <4 x i1> %2, %extract.i
1724 %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
1725 %6 = bitcast <64 x i1> %5 to i64
1726 ret i64 %6
1727 }
1730 define zeroext i64 @test_vpcmpeqd_v4i1_v64i1_mask_mem_b(<2 x i64> %__a, i32* %__b) local_unnamed_addr {
1731 ; VLX-LABEL: test_vpcmpeqd_v4i1_v64i1_mask_mem_b:
1732 ; VLX: # %bb.0: # %entry
1733 ; VLX-NEXT: vpcmpeqd (%rdi){1to4}, %xmm0, %k0
1734 ; VLX-NEXT: kmovq %k0, %rax
1735 ; VLX-NEXT: retq
1736 ;
1737 ; NoVLX-LABEL: test_vpcmpeqd_v4i1_v64i1_mask_mem_b:
1738 ; NoVLX: # %bb.0: # %entry
1739 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
1740 ; NoVLX-NEXT: vpcmpeqd (%rdi){1to16}, %zmm0, %k0
1741 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
1742 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
1743 ; NoVLX-NEXT: kmovw %k0, %eax
1744 ; NoVLX-NEXT: vzeroupper
1745 ; NoVLX-NEXT: retq
1746 entry:
1747 %0 = bitcast <2 x i64> %__a to <4 x i32>
1748 %load = load i32, i32* %__b
1749 %vec = insertelement <4 x i32> undef, i32 %load, i32 0
1750 %1 = shufflevector <4 x i32> %vec, <4 x i32> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
1751 %2 = icmp eq <4 x i32> %0, %1
1752 %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
1753 %4 = bitcast <64 x i1> %3 to i64
1754 ret i64 %4
1755 }
1757 define zeroext i64 @test_masked_vpcmpeqd_v4i1_v64i1_mask_mem_b(i8 zeroext %__u, <2 x i64> %__a, i32* %__b) local_unnamed_addr {
1758 ; VLX-LABEL: test_masked_vpcmpeqd_v4i1_v64i1_mask_mem_b:
1759 ; VLX: # %bb.0: # %entry
1760 ; VLX-NEXT: kmovd %edi, %k1
1761 ; VLX-NEXT: vpcmpeqd (%rsi){1to4}, %xmm0, %k0 {%k1}
1762 ; VLX-NEXT: kmovq %k0, %rax
1763 ; VLX-NEXT: retq
1764 ;
1765 ; NoVLX-LABEL: test_masked_vpcmpeqd_v4i1_v64i1_mask_mem_b:
1766 ; NoVLX: # %bb.0: # %entry
1767 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
1768 ; NoVLX-NEXT: kmovw %edi, %k1
1769 ; NoVLX-NEXT: vpcmpeqd (%rsi){1to16}, %zmm0, %k0 {%k1}
1770 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
1771 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
1772 ; NoVLX-NEXT: kmovw %k0, %eax
1773 ; NoVLX-NEXT: vzeroupper
1774 ; NoVLX-NEXT: retq
1775 entry:
1776 %0 = bitcast <2 x i64> %__a to <4 x i32>
1777 %load = load i32, i32* %__b
1778 %vec = insertelement <4 x i32> undef, i32 %load, i32 0
1779 %1 = shufflevector <4 x i32> %vec, <4 x i32> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
1780 %2 = icmp eq <4 x i32> %0, %1
1781 %3 = bitcast i8 %__u to <8 x i1>
1782 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
1783 %4 = and <4 x i1> %extract.i, %2
1784 %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
1785 %6 = bitcast <64 x i1> %5 to i64
1786 ret i64 %6
1787 }
1790 define zeroext i16 @test_vpcmpeqd_v8i1_v16i1_mask(<4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
1791 ; VLX-LABEL: test_vpcmpeqd_v8i1_v16i1_mask:
1792 ; VLX: # %bb.0: # %entry
1793 ; VLX-NEXT: vpcmpeqd %ymm1, %ymm0, %k0
1794 ; VLX-NEXT: kmovd %k0, %eax
1795 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
1796 ; VLX-NEXT: vzeroupper
1797 ; VLX-NEXT: retq
1798 ;
1799 ; NoVLX-LABEL: test_vpcmpeqd_v8i1_v16i1_mask:
1800 ; NoVLX: # %bb.0: # %entry
1801 ; NoVLX-NEXT: # kill: def $ymm1 killed $ymm1 def $zmm1
1802 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
1803 ; NoVLX-NEXT: vpcmpeqd %zmm1, %zmm0, %k0
1804 ; NoVLX-NEXT: kshiftlw $8, %k0, %k0
1805 ; NoVLX-NEXT: kshiftrw $8, %k0, %k0
1806 ; NoVLX-NEXT: kmovw %k0, %eax
1807 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
1808 ; NoVLX-NEXT: vzeroupper
1809 ; NoVLX-NEXT: retq
1810 entry:
1811 %0 = bitcast <4 x i64> %__a to <8 x i32>
1812 %1 = bitcast <4 x i64> %__b to <8 x i32>
1813 %2 = icmp eq <8 x i32> %0, %1
1814 %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
1815 %4 = bitcast <16 x i1> %3 to i16
1816 ret i16 %4
1817 }
1819 define zeroext i16 @test_vpcmpeqd_v8i1_v16i1_mask_mem(<4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
1820 ; VLX-LABEL: test_vpcmpeqd_v8i1_v16i1_mask_mem:
1821 ; VLX: # %bb.0: # %entry
1822 ; VLX-NEXT: vpcmpeqd (%rdi), %ymm0, %k0
1823 ; VLX-NEXT: kmovd %k0, %eax
1824 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
1825 ; VLX-NEXT: vzeroupper
1826 ; VLX-NEXT: retq
1827 ;
1828 ; NoVLX-LABEL: test_vpcmpeqd_v8i1_v16i1_mask_mem:
1829 ; NoVLX: # %bb.0: # %entry
1830 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
1831 ; NoVLX-NEXT: vmovdqa (%rdi), %ymm1
1832 ; NoVLX-NEXT: vpcmpeqd %zmm1, %zmm0, %k0
1833 ; NoVLX-NEXT: kshiftlw $8, %k0, %k0
1834 ; NoVLX-NEXT: kshiftrw $8, %k0, %k0
1835 ; NoVLX-NEXT: kmovw %k0, %eax
1836 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
1837 ; NoVLX-NEXT: vzeroupper
1838 ; NoVLX-NEXT: retq
1839 entry:
1840 %0 = bitcast <4 x i64> %__a to <8 x i32>
1841 %load = load <4 x i64>, <4 x i64>* %__b
1842 %1 = bitcast <4 x i64> %load to <8 x i32>
1843 %2 = icmp eq <8 x i32> %0, %1
1844 %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
1845 %4 = bitcast <16 x i1> %3 to i16
1846 ret i16 %4
1847 }
1849 define zeroext i16 @test_masked_vpcmpeqd_v8i1_v16i1_mask(i8 zeroext %__u, <4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
1850 ; VLX-LABEL: test_masked_vpcmpeqd_v8i1_v16i1_mask:
1851 ; VLX: # %bb.0: # %entry
1852 ; VLX-NEXT: kmovd %edi, %k1
1853 ; VLX-NEXT: vpcmpeqd %ymm1, %ymm0, %k0 {%k1}
1854 ; VLX-NEXT: kmovd %k0, %eax
1855 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
1856 ; VLX-NEXT: vzeroupper
1857 ; VLX-NEXT: retq
1858 ;
1859 ; NoVLX-LABEL: test_masked_vpcmpeqd_v8i1_v16i1_mask:
1860 ; NoVLX: # %bb.0: # %entry
1861 ; NoVLX-NEXT: # kill: def $ymm1 killed $ymm1 def $zmm1
1862 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
1863 ; NoVLX-NEXT: kmovw %edi, %k1
1864 ; NoVLX-NEXT: vpcmpeqd %zmm1, %zmm0, %k0 {%k1}
1865 ; NoVLX-NEXT: kshiftlw $8, %k0, %k0
1866 ; NoVLX-NEXT: kshiftrw $8, %k0, %k0
1867 ; NoVLX-NEXT: kmovw %k0, %eax
1868 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
1869 ; NoVLX-NEXT: vzeroupper
1870 ; NoVLX-NEXT: retq
1871 entry:
1872 %0 = bitcast <4 x i64> %__a to <8 x i32>
1873 %1 = bitcast <4 x i64> %__b to <8 x i32>
1874 %2 = icmp eq <8 x i32> %0, %1
1875 %3 = bitcast i8 %__u to <8 x i1>
1876 %4 = and <8 x i1> %2, %3
1877 %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
1878 %6 = bitcast <16 x i1> %5 to i16
1879 ret i16 %6
1880 }
1882 define zeroext i16 @test_masked_vpcmpeqd_v8i1_v16i1_mask_mem(i8 zeroext %__u, <4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
1883 ; VLX-LABEL: test_masked_vpcmpeqd_v8i1_v16i1_mask_mem:
1884 ; VLX: # %bb.0: # %entry
1885 ; VLX-NEXT: kmovd %edi, %k1
1886 ; VLX-NEXT: vpcmpeqd (%rsi), %ymm0, %k0 {%k1}
1887 ; VLX-NEXT: kmovd %k0, %eax
1888 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
1889 ; VLX-NEXT: vzeroupper
1890 ; VLX-NEXT: retq
1891 ;
1892 ; NoVLX-LABEL: test_masked_vpcmpeqd_v8i1_v16i1_mask_mem:
1893 ; NoVLX: # %bb.0: # %entry
1894 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
1895 ; NoVLX-NEXT: vmovdqa (%rsi), %ymm1
1896 ; NoVLX-NEXT: kmovw %edi, %k1
1897 ; NoVLX-NEXT: vpcmpeqd %zmm1, %zmm0, %k0 {%k1}
1898 ; NoVLX-NEXT: kshiftlw $8, %k0, %k0
1899 ; NoVLX-NEXT: kshiftrw $8, %k0, %k0
1900 ; NoVLX-NEXT: kmovw %k0, %eax
1901 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
1902 ; NoVLX-NEXT: vzeroupper
1903 ; NoVLX-NEXT: retq
1904 entry:
1905 %0 = bitcast <4 x i64> %__a to <8 x i32>
1906 %load = load <4 x i64>, <4 x i64>* %__b
1907 %1 = bitcast <4 x i64> %load to <8 x i32>
1908 %2 = icmp eq <8 x i32> %0, %1
1909 %3 = bitcast i8 %__u to <8 x i1>
1910 %4 = and <8 x i1> %2, %3
1911 %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
1912 %6 = bitcast <16 x i1> %5 to i16
1913 ret i16 %6
1914 }
1917 define zeroext i16 @test_vpcmpeqd_v8i1_v16i1_mask_mem_b(<4 x i64> %__a, i32* %__b) local_unnamed_addr {
1918 ; VLX-LABEL: test_vpcmpeqd_v8i1_v16i1_mask_mem_b:
1919 ; VLX: # %bb.0: # %entry
1920 ; VLX-NEXT: vpcmpeqd (%rdi){1to8}, %ymm0, %k0
1921 ; VLX-NEXT: kmovd %k0, %eax
1922 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
1923 ; VLX-NEXT: vzeroupper
1924 ; VLX-NEXT: retq
1925 ;
1926 ; NoVLX-LABEL: test_vpcmpeqd_v8i1_v16i1_mask_mem_b:
1927 ; NoVLX: # %bb.0: # %entry
1928 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
1929 ; NoVLX-NEXT: vpcmpeqd (%rdi){1to16}, %zmm0, %k0
1930 ; NoVLX-NEXT: kshiftlw $8, %k0, %k0
1931 ; NoVLX-NEXT: kshiftrw $8, %k0, %k0
1932 ; NoVLX-NEXT: kmovw %k0, %eax
1933 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
1934 ; NoVLX-NEXT: vzeroupper
1935 ; NoVLX-NEXT: retq
1936 entry:
1937 %0 = bitcast <4 x i64> %__a to <8 x i32>
1938 %load = load i32, i32* %__b
1939 %vec = insertelement <8 x i32> undef, i32 %load, i32 0
1940 %1 = shufflevector <8 x i32> %vec, <8 x i32> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
1941 %2 = icmp eq <8 x i32> %0, %1
1942 %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
1943 %4 = bitcast <16 x i1> %3 to i16
1944 ret i16 %4
1945 }
1947 define zeroext i16 @test_masked_vpcmpeqd_v8i1_v16i1_mask_mem_b(i8 zeroext %__u, <4 x i64> %__a, i32* %__b) local_unnamed_addr {
1948 ; VLX-LABEL: test_masked_vpcmpeqd_v8i1_v16i1_mask_mem_b:
1949 ; VLX: # %bb.0: # %entry
1950 ; VLX-NEXT: kmovd %edi, %k1
1951 ; VLX-NEXT: vpcmpeqd (%rsi){1to8}, %ymm0, %k0 {%k1}
1952 ; VLX-NEXT: kmovd %k0, %eax
1953 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
1954 ; VLX-NEXT: vzeroupper
1955 ; VLX-NEXT: retq
1956 ;
1957 ; NoVLX-LABEL: test_masked_vpcmpeqd_v8i1_v16i1_mask_mem_b:
1958 ; NoVLX: # %bb.0: # %entry
1959 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
1960 ; NoVLX-NEXT: kmovw %edi, %k1
1961 ; NoVLX-NEXT: vpcmpeqd (%rsi){1to16}, %zmm0, %k0 {%k1}
1962 ; NoVLX-NEXT: kshiftlw $8, %k0, %k0
1963 ; NoVLX-NEXT: kshiftrw $8, %k0, %k0
1964 ; NoVLX-NEXT: kmovw %k0, %eax
1965 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
1966 ; NoVLX-NEXT: vzeroupper
1967 ; NoVLX-NEXT: retq
1968 entry:
1969 %0 = bitcast <4 x i64> %__a to <8 x i32>
1970 %load = load i32, i32* %__b
1971 %vec = insertelement <8 x i32> undef, i32 %load, i32 0
1972 %1 = shufflevector <8 x i32> %vec, <8 x i32> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
1973 %2 = icmp eq <8 x i32> %0, %1
1974 %3 = bitcast i8 %__u to <8 x i1>
1975 %4 = and <8 x i1> %3, %2
1976 %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
1977 %6 = bitcast <16 x i1> %5 to i16
1978 ret i16 %6
1979 }
1982 define zeroext i32 @test_vpcmpeqd_v8i1_v32i1_mask(<4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
1983 ; VLX-LABEL: test_vpcmpeqd_v8i1_v32i1_mask:
1984 ; VLX: # %bb.0: # %entry
1985 ; VLX-NEXT: vpcmpeqd %ymm1, %ymm0, %k0
1986 ; VLX-NEXT: kmovd %k0, %eax
1987 ; VLX-NEXT: vzeroupper
1988 ; VLX-NEXT: retq
1989 ;
1990 ; NoVLX-LABEL: test_vpcmpeqd_v8i1_v32i1_mask:
1991 ; NoVLX: # %bb.0: # %entry
1992 ; NoVLX-NEXT: # kill: def $ymm1 killed $ymm1 def $zmm1
1993 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
1994 ; NoVLX-NEXT: vpcmpeqd %zmm1, %zmm0, %k0
1995 ; NoVLX-NEXT: kshiftlw $8, %k0, %k0
1996 ; NoVLX-NEXT: kshiftrw $8, %k0, %k0
1997 ; NoVLX-NEXT: kmovw %k0, %eax
1998 ; NoVLX-NEXT: vzeroupper
1999 ; NoVLX-NEXT: retq
2000 entry:
2001 %0 = bitcast <4 x i64> %__a to <8 x i32>
2002 %1 = bitcast <4 x i64> %__b to <8 x i32>
2003 %2 = icmp eq <8 x i32> %0, %1
2004 %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
2005 %4 = bitcast <32 x i1> %3 to i32
2006 ret i32 %4
2007 }
2009 define zeroext i32 @test_vpcmpeqd_v8i1_v32i1_mask_mem(<4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
2010 ; VLX-LABEL: test_vpcmpeqd_v8i1_v32i1_mask_mem:
2011 ; VLX: # %bb.0: # %entry
2012 ; VLX-NEXT: vpcmpeqd (%rdi), %ymm0, %k0
2013 ; VLX-NEXT: kmovd %k0, %eax
2014 ; VLX-NEXT: vzeroupper
2015 ; VLX-NEXT: retq
2016 ;
2017 ; NoVLX-LABEL: test_vpcmpeqd_v8i1_v32i1_mask_mem:
2018 ; NoVLX: # %bb.0: # %entry
2019 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
2020 ; NoVLX-NEXT: vmovdqa (%rdi), %ymm1
2021 ; NoVLX-NEXT: vpcmpeqd %zmm1, %zmm0, %k0
2022 ; NoVLX-NEXT: kshiftlw $8, %k0, %k0
2023 ; NoVLX-NEXT: kshiftrw $8, %k0, %k0
2024 ; NoVLX-NEXT: kmovw %k0, %eax
2025 ; NoVLX-NEXT: vzeroupper
2026 ; NoVLX-NEXT: retq
2027 entry:
2028 %0 = bitcast <4 x i64> %__a to <8 x i32>
2029 %load = load <4 x i64>, <4 x i64>* %__b
2030 %1 = bitcast <4 x i64> %load to <8 x i32>
2031 %2 = icmp eq <8 x i32> %0, %1
2032 %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
2033 %4 = bitcast <32 x i1> %3 to i32
2034 ret i32 %4
2035 }
2037 define zeroext i32 @test_masked_vpcmpeqd_v8i1_v32i1_mask(i8 zeroext %__u, <4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
2038 ; VLX-LABEL: test_masked_vpcmpeqd_v8i1_v32i1_mask:
2039 ; VLX: # %bb.0: # %entry
2040 ; VLX-NEXT: kmovd %edi, %k1
2041 ; VLX-NEXT: vpcmpeqd %ymm1, %ymm0, %k0 {%k1}
2042 ; VLX-NEXT: kmovd %k0, %eax
2043 ; VLX-NEXT: vzeroupper
2044 ; VLX-NEXT: retq
2045 ;
2046 ; NoVLX-LABEL: test_masked_vpcmpeqd_v8i1_v32i1_mask:
2047 ; NoVLX: # %bb.0: # %entry
2048 ; NoVLX-NEXT: # kill: def $ymm1 killed $ymm1 def $zmm1
2049 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
2050 ; NoVLX-NEXT: kmovw %edi, %k1
2051 ; NoVLX-NEXT: vpcmpeqd %zmm1, %zmm0, %k0 {%k1}
2052 ; NoVLX-NEXT: kshiftlw $8, %k0, %k0
2053 ; NoVLX-NEXT: kshiftrw $8, %k0, %k0
2054 ; NoVLX-NEXT: kmovw %k0, %eax
2055 ; NoVLX-NEXT: vzeroupper
2056 ; NoVLX-NEXT: retq
2057 entry:
2058 %0 = bitcast <4 x i64> %__a to <8 x i32>
2059 %1 = bitcast <4 x i64> %__b to <8 x i32>
2060 %2 = icmp eq <8 x i32> %0, %1
2061 %3 = bitcast i8 %__u to <8 x i1>
2062 %4 = and <8 x i1> %2, %3
2063 %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
2064 %6 = bitcast <32 x i1> %5 to i32
2065 ret i32 %6
2066 }
2068 define zeroext i32 @test_masked_vpcmpeqd_v8i1_v32i1_mask_mem(i8 zeroext %__u, <4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
2069 ; VLX-LABEL: test_masked_vpcmpeqd_v8i1_v32i1_mask_mem:
2070 ; VLX: # %bb.0: # %entry
2071 ; VLX-NEXT: kmovd %edi, %k1
2072 ; VLX-NEXT: vpcmpeqd (%rsi), %ymm0, %k0 {%k1}
2073 ; VLX-NEXT: kmovd %k0, %eax
2074 ; VLX-NEXT: vzeroupper
2075 ; VLX-NEXT: retq
2076 ;
2077 ; NoVLX-LABEL: test_masked_vpcmpeqd_v8i1_v32i1_mask_mem:
2078 ; NoVLX: # %bb.0: # %entry
2079 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
2080 ; NoVLX-NEXT: vmovdqa (%rsi), %ymm1
2081 ; NoVLX-NEXT: kmovw %edi, %k1
2082 ; NoVLX-NEXT: vpcmpeqd %zmm1, %zmm0, %k0 {%k1}
2083 ; NoVLX-NEXT: kshiftlw $8, %k0, %k0
2084 ; NoVLX-NEXT: kshiftrw $8, %k0, %k0
2085 ; NoVLX-NEXT: kmovw %k0, %eax
2086 ; NoVLX-NEXT: vzeroupper
2087 ; NoVLX-NEXT: retq
2088 entry:
2089 %0 = bitcast <4 x i64> %__a to <8 x i32>
2090 %load = load <4 x i64>, <4 x i64>* %__b
2091 %1 = bitcast <4 x i64> %load to <8 x i32>
2092 %2 = icmp eq <8 x i32> %0, %1
2093 %3 = bitcast i8 %__u to <8 x i1>
2094 %4 = and <8 x i1> %2, %3
2095 %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
2096 %6 = bitcast <32 x i1> %5 to i32
2097 ret i32 %6
2098 }
2101 define zeroext i32 @test_vpcmpeqd_v8i1_v32i1_mask_mem_b(<4 x i64> %__a, i32* %__b) local_unnamed_addr {
2102 ; VLX-LABEL: test_vpcmpeqd_v8i1_v32i1_mask_mem_b:
2103 ; VLX: # %bb.0: # %entry
2104 ; VLX-NEXT: vpcmpeqd (%rdi){1to8}, %ymm0, %k0
2105 ; VLX-NEXT: kmovd %k0, %eax
2106 ; VLX-NEXT: vzeroupper
2107 ; VLX-NEXT: retq
2108 ;
2109 ; NoVLX-LABEL: test_vpcmpeqd_v8i1_v32i1_mask_mem_b:
2110 ; NoVLX: # %bb.0: # %entry
2111 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
2112 ; NoVLX-NEXT: vpcmpeqd (%rdi){1to16}, %zmm0, %k0
2113 ; NoVLX-NEXT: kshiftlw $8, %k0, %k0
2114 ; NoVLX-NEXT: kshiftrw $8, %k0, %k0
2115 ; NoVLX-NEXT: kmovw %k0, %eax
2116 ; NoVLX-NEXT: vzeroupper
2117 ; NoVLX-NEXT: retq
2118 entry:
2119 %0 = bitcast <4 x i64> %__a to <8 x i32>
2120 %load = load i32, i32* %__b
2121 %vec = insertelement <8 x i32> undef, i32 %load, i32 0
2122 %1 = shufflevector <8 x i32> %vec, <8 x i32> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
2123 %2 = icmp eq <8 x i32> %0, %1
2124 %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
2125 %4 = bitcast <32 x i1> %3 to i32
2126 ret i32 %4
2127 }
2129 define zeroext i32 @test_masked_vpcmpeqd_v8i1_v32i1_mask_mem_b(i8 zeroext %__u, <4 x i64> %__a, i32* %__b) local_unnamed_addr {
2130 ; VLX-LABEL: test_masked_vpcmpeqd_v8i1_v32i1_mask_mem_b:
2131 ; VLX: # %bb.0: # %entry
2132 ; VLX-NEXT: kmovd %edi, %k1
2133 ; VLX-NEXT: vpcmpeqd (%rsi){1to8}, %ymm0, %k0 {%k1}
2134 ; VLX-NEXT: kmovd %k0, %eax
2135 ; VLX-NEXT: vzeroupper
2136 ; VLX-NEXT: retq
2137 ;
2138 ; NoVLX-LABEL: test_masked_vpcmpeqd_v8i1_v32i1_mask_mem_b:
2139 ; NoVLX: # %bb.0: # %entry
2140 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
2141 ; NoVLX-NEXT: kmovw %edi, %k1
2142 ; NoVLX-NEXT: vpcmpeqd (%rsi){1to16}, %zmm0, %k0 {%k1}
2143 ; NoVLX-NEXT: kshiftlw $8, %k0, %k0
2144 ; NoVLX-NEXT: kshiftrw $8, %k0, %k0
2145 ; NoVLX-NEXT: kmovw %k0, %eax
2146 ; NoVLX-NEXT: vzeroupper
2147 ; NoVLX-NEXT: retq
2148 entry:
2149 %0 = bitcast <4 x i64> %__a to <8 x i32>
2150 %load = load i32, i32* %__b
2151 %vec = insertelement <8 x i32> undef, i32 %load, i32 0
2152 %1 = shufflevector <8 x i32> %vec, <8 x i32> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
2153 %2 = icmp eq <8 x i32> %0, %1
2154 %3 = bitcast i8 %__u to <8 x i1>
2155 %4 = and <8 x i1> %3, %2
2156 %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
2157 %6 = bitcast <32 x i1> %5 to i32
2158 ret i32 %6
2159 }
2162 define zeroext i64 @test_vpcmpeqd_v8i1_v64i1_mask(<4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
2163 ; VLX-LABEL: test_vpcmpeqd_v8i1_v64i1_mask:
2164 ; VLX: # %bb.0: # %entry
2165 ; VLX-NEXT: vpcmpeqd %ymm1, %ymm0, %k0
2166 ; VLX-NEXT: kmovq %k0, %rax
2167 ; VLX-NEXT: vzeroupper
2168 ; VLX-NEXT: retq
2169 ;
2170 ; NoVLX-LABEL: test_vpcmpeqd_v8i1_v64i1_mask:
2171 ; NoVLX: # %bb.0: # %entry
2172 ; NoVLX-NEXT: # kill: def $ymm1 killed $ymm1 def $zmm1
2173 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
2174 ; NoVLX-NEXT: vpcmpeqd %zmm1, %zmm0, %k0
2175 ; NoVLX-NEXT: kshiftlw $8, %k0, %k0
2176 ; NoVLX-NEXT: kshiftrw $8, %k0, %k0
2177 ; NoVLX-NEXT: kmovw %k0, %eax
2178 ; NoVLX-NEXT: vzeroupper
2179 ; NoVLX-NEXT: retq
2180 entry:
2181 %0 = bitcast <4 x i64> %__a to <8 x i32>
2182 %1 = bitcast <4 x i64> %__b to <8 x i32>
2183 %2 = icmp eq <8 x i32> %0, %1
2184 %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
2185 %4 = bitcast <64 x i1> %3 to i64
2186 ret i64 %4
2187 }
2189 define zeroext i64 @test_vpcmpeqd_v8i1_v64i1_mask_mem(<4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
2190 ; VLX-LABEL: test_vpcmpeqd_v8i1_v64i1_mask_mem:
2191 ; VLX: # %bb.0: # %entry
2192 ; VLX-NEXT: vpcmpeqd (%rdi), %ymm0, %k0
2193 ; VLX-NEXT: kmovq %k0, %rax
2194 ; VLX-NEXT: vzeroupper
2195 ; VLX-NEXT: retq
2196 ;
2197 ; NoVLX-LABEL: test_vpcmpeqd_v8i1_v64i1_mask_mem:
2198 ; NoVLX: # %bb.0: # %entry
2199 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
2200 ; NoVLX-NEXT: vmovdqa (%rdi), %ymm1
2201 ; NoVLX-NEXT: vpcmpeqd %zmm1, %zmm0, %k0
2202 ; NoVLX-NEXT: kshiftlw $8, %k0, %k0
2203 ; NoVLX-NEXT: kshiftrw $8, %k0, %k0
2204 ; NoVLX-NEXT: kmovw %k0, %eax
2205 ; NoVLX-NEXT: vzeroupper
2206 ; NoVLX-NEXT: retq
2207 entry:
2208 %0 = bitcast <4 x i64> %__a to <8 x i32>
2209 %load = load <4 x i64>, <4 x i64>* %__b
2210 %1 = bitcast <4 x i64> %load to <8 x i32>
2211 %2 = icmp eq <8 x i32> %0, %1
2212 %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
2213 %4 = bitcast <64 x i1> %3 to i64
2214 ret i64 %4
2215 }
2217 define zeroext i64 @test_masked_vpcmpeqd_v8i1_v64i1_mask(i8 zeroext %__u, <4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
2218 ; VLX-LABEL: test_masked_vpcmpeqd_v8i1_v64i1_mask:
2219 ; VLX: # %bb.0: # %entry
2220 ; VLX-NEXT: kmovd %edi, %k1
2221 ; VLX-NEXT: vpcmpeqd %ymm1, %ymm0, %k0 {%k1}
2222 ; VLX-NEXT: kmovq %k0, %rax
2223 ; VLX-NEXT: vzeroupper
2224 ; VLX-NEXT: retq
2225 ;
2226 ; NoVLX-LABEL: test_masked_vpcmpeqd_v8i1_v64i1_mask:
2227 ; NoVLX: # %bb.0: # %entry
2228 ; NoVLX-NEXT: # kill: def $ymm1 killed $ymm1 def $zmm1
2229 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
2230 ; NoVLX-NEXT: kmovw %edi, %k1
2231 ; NoVLX-NEXT: vpcmpeqd %zmm1, %zmm0, %k0 {%k1}
2232 ; NoVLX-NEXT: kshiftlw $8, %k0, %k0
2233 ; NoVLX-NEXT: kshiftrw $8, %k0, %k0
2234 ; NoVLX-NEXT: kmovw %k0, %eax
2235 ; NoVLX-NEXT: vzeroupper
2236 ; NoVLX-NEXT: retq
2237 entry:
2238 %0 = bitcast <4 x i64> %__a to <8 x i32>
2239 %1 = bitcast <4 x i64> %__b to <8 x i32>
2240 %2 = icmp eq <8 x i32> %0, %1
2241 %3 = bitcast i8 %__u to <8 x i1>
2242 %4 = and <8 x i1> %2, %3
2243 %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
2244 %6 = bitcast <64 x i1> %5 to i64
2245 ret i64 %6
2246 }
2248 define zeroext i64 @test_masked_vpcmpeqd_v8i1_v64i1_mask_mem(i8 zeroext %__u, <4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
2249 ; VLX-LABEL: test_masked_vpcmpeqd_v8i1_v64i1_mask_mem:
2250 ; VLX: # %bb.0: # %entry
2251 ; VLX-NEXT: kmovd %edi, %k1
2252 ; VLX-NEXT: vpcmpeqd (%rsi), %ymm0, %k0 {%k1}
2253 ; VLX-NEXT: kmovq %k0, %rax
2254 ; VLX-NEXT: vzeroupper
2255 ; VLX-NEXT: retq
2256 ;
2257 ; NoVLX-LABEL: test_masked_vpcmpeqd_v8i1_v64i1_mask_mem:
2258 ; NoVLX: # %bb.0: # %entry
2259 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
2260 ; NoVLX-NEXT: vmovdqa (%rsi), %ymm1
2261 ; NoVLX-NEXT: kmovw %edi, %k1
2262 ; NoVLX-NEXT: vpcmpeqd %zmm1, %zmm0, %k0 {%k1}
2263 ; NoVLX-NEXT: kshiftlw $8, %k0, %k0
2264 ; NoVLX-NEXT: kshiftrw $8, %k0, %k0
2265 ; NoVLX-NEXT: kmovw %k0, %eax
2266 ; NoVLX-NEXT: vzeroupper
2267 ; NoVLX-NEXT: retq
2268 entry:
2269 %0 = bitcast <4 x i64> %__a to <8 x i32>
2270 %load = load <4 x i64>, <4 x i64>* %__b
2271 %1 = bitcast <4 x i64> %load to <8 x i32>
2272 %2 = icmp eq <8 x i32> %0, %1
2273 %3 = bitcast i8 %__u to <8 x i1>
2274 %4 = and <8 x i1> %2, %3
2275 %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
2276 %6 = bitcast <64 x i1> %5 to i64
2277 ret i64 %6
2278 }
2281 define zeroext i64 @test_vpcmpeqd_v8i1_v64i1_mask_mem_b(<4 x i64> %__a, i32* %__b) local_unnamed_addr {
2282 ; VLX-LABEL: test_vpcmpeqd_v8i1_v64i1_mask_mem_b:
2283 ; VLX: # %bb.0: # %entry
2284 ; VLX-NEXT: vpcmpeqd (%rdi){1to8}, %ymm0, %k0
2285 ; VLX-NEXT: kmovq %k0, %rax
2286 ; VLX-NEXT: vzeroupper
2287 ; VLX-NEXT: retq
2288 ;
2289 ; NoVLX-LABEL: test_vpcmpeqd_v8i1_v64i1_mask_mem_b:
2290 ; NoVLX: # %bb.0: # %entry
2291 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
2292 ; NoVLX-NEXT: vpcmpeqd (%rdi){1to16}, %zmm0, %k0
2293 ; NoVLX-NEXT: kshiftlw $8, %k0, %k0
2294 ; NoVLX-NEXT: kshiftrw $8, %k0, %k0
2295 ; NoVLX-NEXT: kmovw %k0, %eax
2296 ; NoVLX-NEXT: vzeroupper
2297 ; NoVLX-NEXT: retq
2298 entry:
2299 %0 = bitcast <4 x i64> %__a to <8 x i32>
2300 %load = load i32, i32* %__b
2301 %vec = insertelement <8 x i32> undef, i32 %load, i32 0
2302 %1 = shufflevector <8 x i32> %vec, <8 x i32> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
2303 %2 = icmp eq <8 x i32> %0, %1
2304 %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
2305 %4 = bitcast <64 x i1> %3 to i64
2306 ret i64 %4
2307 }
2309 define zeroext i64 @test_masked_vpcmpeqd_v8i1_v64i1_mask_mem_b(i8 zeroext %__u, <4 x i64> %__a, i32* %__b) local_unnamed_addr {
2310 ; VLX-LABEL: test_masked_vpcmpeqd_v8i1_v64i1_mask_mem_b:
2311 ; VLX: # %bb.0: # %entry
2312 ; VLX-NEXT: kmovd %edi, %k1
2313 ; VLX-NEXT: vpcmpeqd (%rsi){1to8}, %ymm0, %k0 {%k1}
2314 ; VLX-NEXT: kmovq %k0, %rax
2315 ; VLX-NEXT: vzeroupper
2316 ; VLX-NEXT: retq
2317 ;
2318 ; NoVLX-LABEL: test_masked_vpcmpeqd_v8i1_v64i1_mask_mem_b:
2319 ; NoVLX: # %bb.0: # %entry
2320 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
2321 ; NoVLX-NEXT: kmovw %edi, %k1
2322 ; NoVLX-NEXT: vpcmpeqd (%rsi){1to16}, %zmm0, %k0 {%k1}
2323 ; NoVLX-NEXT: kshiftlw $8, %k0, %k0
2324 ; NoVLX-NEXT: kshiftrw $8, %k0, %k0
2325 ; NoVLX-NEXT: kmovw %k0, %eax
2326 ; NoVLX-NEXT: vzeroupper
2327 ; NoVLX-NEXT: retq
2328 entry:
2329 %0 = bitcast <4 x i64> %__a to <8 x i32>
2330 %load = load i32, i32* %__b
2331 %vec = insertelement <8 x i32> undef, i32 %load, i32 0
2332 %1 = shufflevector <8 x i32> %vec, <8 x i32> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
2333 %2 = icmp eq <8 x i32> %0, %1
2334 %3 = bitcast i8 %__u to <8 x i1>
2335 %4 = and <8 x i1> %3, %2
2336 %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
2337 %6 = bitcast <64 x i1> %5 to i64
2338 ret i64 %6
2339 }
2342 define zeroext i32 @test_vpcmpeqd_v16i1_v32i1_mask(<8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
2343 ; VLX-LABEL: test_vpcmpeqd_v16i1_v32i1_mask:
2344 ; VLX: # %bb.0: # %entry
2345 ; VLX-NEXT: vpcmpeqd %zmm1, %zmm0, %k0
2346 ; VLX-NEXT: kmovd %k0, %eax
2347 ; VLX-NEXT: vzeroupper
2348 ; VLX-NEXT: retq
2349 ;
2350 ; NoVLX-LABEL: test_vpcmpeqd_v16i1_v32i1_mask:
2351 ; NoVLX: # %bb.0: # %entry
2352 ; NoVLX-NEXT: vpcmpeqd %zmm1, %zmm0, %k0
2353 ; NoVLX-NEXT: kmovw %k0, %eax
2354 ; NoVLX-NEXT: vzeroupper
2355 ; NoVLX-NEXT: retq
2356 entry:
2357 %0 = bitcast <8 x i64> %__a to <16 x i32>
2358 %1 = bitcast <8 x i64> %__b to <16 x i32>
2359 %2 = icmp eq <16 x i32> %0, %1
2360 %3 = shufflevector <16 x i1> %2, <16 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>
2361 %4 = bitcast <32 x i1> %3 to i32
2362 ret i32 %4
2363 }
2365 define zeroext i32 @test_vpcmpeqd_v16i1_v32i1_mask_mem(<8 x i64> %__a, <8 x i64>* %__b) local_unnamed_addr {
2366 ; VLX-LABEL: test_vpcmpeqd_v16i1_v32i1_mask_mem:
2367 ; VLX: # %bb.0: # %entry
2368 ; VLX-NEXT: vpcmpeqd (%rdi), %zmm0, %k0
2369 ; VLX-NEXT: kmovd %k0, %eax
2370 ; VLX-NEXT: vzeroupper
2371 ; VLX-NEXT: retq
2372 ;
2373 ; NoVLX-LABEL: test_vpcmpeqd_v16i1_v32i1_mask_mem:
2374 ; NoVLX: # %bb.0: # %entry
2375 ; NoVLX-NEXT: vpcmpeqd (%rdi), %zmm0, %k0
2376 ; NoVLX-NEXT: kmovw %k0, %eax
2377 ; NoVLX-NEXT: vzeroupper
2378 ; NoVLX-NEXT: retq
2379 entry:
2380 %0 = bitcast <8 x i64> %__a to <16 x i32>
2381 %load = load <8 x i64>, <8 x i64>* %__b
2382 %1 = bitcast <8 x i64> %load to <16 x i32>
2383 %2 = icmp eq <16 x i32> %0, %1
2384 %3 = shufflevector <16 x i1> %2, <16 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>
2385 %4 = bitcast <32 x i1> %3 to i32
2386 ret i32 %4
2387 }
2389 define zeroext i32 @test_masked_vpcmpeqd_v16i1_v32i1_mask(i16 zeroext %__u, <8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
2390 ; VLX-LABEL: test_masked_vpcmpeqd_v16i1_v32i1_mask:
2391 ; VLX: # %bb.0: # %entry
2392 ; VLX-NEXT: kmovd %edi, %k1
2393 ; VLX-NEXT: vpcmpeqd %zmm1, %zmm0, %k0 {%k1}
2394 ; VLX-NEXT: kmovd %k0, %eax
2395 ; VLX-NEXT: vzeroupper
2396 ; VLX-NEXT: retq
2397 ;
2398 ; NoVLX-LABEL: test_masked_vpcmpeqd_v16i1_v32i1_mask:
2399 ; NoVLX: # %bb.0: # %entry
2400 ; NoVLX-NEXT: vpcmpeqd %zmm1, %zmm0, %k0
2401 ; NoVLX-NEXT: kmovw %k0, %eax
2402 ; NoVLX-NEXT: andl %edi, %eax
2403 ; NoVLX-NEXT: vzeroupper
2404 ; NoVLX-NEXT: retq
2405 entry:
2406 %0 = bitcast <8 x i64> %__a to <16 x i32>
2407 %1 = bitcast <8 x i64> %__b to <16 x i32>
2408 %2 = icmp eq <16 x i32> %0, %1
2409 %3 = bitcast i16 %__u to <16 x i1>
2410 %4 = and <16 x i1> %2, %3
2411 %5 = shufflevector <16 x i1> %4, <16 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>
2412 %6 = bitcast <32 x i1> %5 to i32
2413 ret i32 %6
2414 }
2416 define zeroext i32 @test_masked_vpcmpeqd_v16i1_v32i1_mask_mem(i16 zeroext %__u, <8 x i64> %__a, <8 x i64>* %__b) local_unnamed_addr {
2417 ; VLX-LABEL: test_masked_vpcmpeqd_v16i1_v32i1_mask_mem:
2418 ; VLX: # %bb.0: # %entry
2419 ; VLX-NEXT: kmovd %edi, %k1
2420 ; VLX-NEXT: vpcmpeqd (%rsi), %zmm0, %k0 {%k1}
2421 ; VLX-NEXT: kmovd %k0, %eax
2422 ; VLX-NEXT: vzeroupper
2423 ; VLX-NEXT: retq
2424 ;
2425 ; NoVLX-LABEL: test_masked_vpcmpeqd_v16i1_v32i1_mask_mem:
2426 ; NoVLX: # %bb.0: # %entry
2427 ; NoVLX-NEXT: vpcmpeqd (%rsi), %zmm0, %k0
2428 ; NoVLX-NEXT: kmovw %k0, %eax
2429 ; NoVLX-NEXT: andl %edi, %eax
2430 ; NoVLX-NEXT: vzeroupper
2431 ; NoVLX-NEXT: retq
2432 entry:
2433 %0 = bitcast <8 x i64> %__a to <16 x i32>
2434 %load = load <8 x i64>, <8 x i64>* %__b
2435 %1 = bitcast <8 x i64> %load to <16 x i32>
2436 %2 = icmp eq <16 x i32> %0, %1
2437 %3 = bitcast i16 %__u to <16 x i1>
2438 %4 = and <16 x i1> %2, %3
2439 %5 = shufflevector <16 x i1> %4, <16 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>
2440 %6 = bitcast <32 x i1> %5 to i32
2441 ret i32 %6
2442 }
2445 define zeroext i32 @test_vpcmpeqd_v16i1_v32i1_mask_mem_b(<8 x i64> %__a, i32* %__b) local_unnamed_addr {
2446 ; VLX-LABEL: test_vpcmpeqd_v16i1_v32i1_mask_mem_b:
2447 ; VLX: # %bb.0: # %entry
2448 ; VLX-NEXT: vpcmpeqd (%rdi){1to16}, %zmm0, %k0
2449 ; VLX-NEXT: kmovd %k0, %eax
2450 ; VLX-NEXT: vzeroupper
2451 ; VLX-NEXT: retq
2452 ;
2453 ; NoVLX-LABEL: test_vpcmpeqd_v16i1_v32i1_mask_mem_b:
2454 ; NoVLX: # %bb.0: # %entry
2455 ; NoVLX-NEXT: vpcmpeqd (%rdi){1to16}, %zmm0, %k0
2456 ; NoVLX-NEXT: kmovw %k0, %eax
2457 ; NoVLX-NEXT: vzeroupper
2458 ; NoVLX-NEXT: retq
2459 entry:
2460 %0 = bitcast <8 x i64> %__a to <16 x i32>
2461 %load = load i32, i32* %__b
2462 %vec = insertelement <16 x i32> undef, i32 %load, i32 0
2463 %1 = shufflevector <16 x i32> %vec, <16 x i32> undef, <16 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
2464 %2 = icmp eq <16 x i32> %0, %1
2465 %3 = shufflevector <16 x i1> %2, <16 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>
2466 %4 = bitcast <32 x i1> %3 to i32
2467 ret i32 %4
2468 }
2470 define zeroext i32 @test_masked_vpcmpeqd_v16i1_v32i1_mask_mem_b(i16 zeroext %__u, <8 x i64> %__a, i32* %__b) local_unnamed_addr {
2471 ; VLX-LABEL: test_masked_vpcmpeqd_v16i1_v32i1_mask_mem_b:
2472 ; VLX: # %bb.0: # %entry
2473 ; VLX-NEXT: kmovd %edi, %k1
2474 ; VLX-NEXT: vpcmpeqd (%rsi){1to16}, %zmm0, %k0 {%k1}
2475 ; VLX-NEXT: kmovd %k0, %eax
2476 ; VLX-NEXT: vzeroupper
2477 ; VLX-NEXT: retq
2478 ;
2479 ; NoVLX-LABEL: test_masked_vpcmpeqd_v16i1_v32i1_mask_mem_b:
2480 ; NoVLX: # %bb.0: # %entry
2481 ; NoVLX-NEXT: vpcmpeqd (%rsi){1to16}, %zmm0, %k0
2482 ; NoVLX-NEXT: kmovw %k0, %eax
2483 ; NoVLX-NEXT: andl %edi, %eax
2484 ; NoVLX-NEXT: vzeroupper
2485 ; NoVLX-NEXT: retq
2486 entry:
2487 %0 = bitcast <8 x i64> %__a to <16 x i32>
2488 %load = load i32, i32* %__b
2489 %vec = insertelement <16 x i32> undef, i32 %load, i32 0
2490 %1 = shufflevector <16 x i32> %vec, <16 x i32> undef, <16 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
2491 %2 = icmp eq <16 x i32> %0, %1
2492 %3 = bitcast i16 %__u to <16 x i1>
2493 %4 = and <16 x i1> %3, %2
2494 %5 = shufflevector <16 x i1> %4, <16 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>
2495 %6 = bitcast <32 x i1> %5 to i32
2496 ret i32 %6
2497 }
2500 define zeroext i64 @test_vpcmpeqd_v16i1_v64i1_mask(<8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
2501 ; VLX-LABEL: test_vpcmpeqd_v16i1_v64i1_mask:
2502 ; VLX: # %bb.0: # %entry
2503 ; VLX-NEXT: vpcmpeqd %zmm1, %zmm0, %k0
2504 ; VLX-NEXT: kmovq %k0, %rax
2505 ; VLX-NEXT: vzeroupper
2506 ; VLX-NEXT: retq
2507 ;
2508 ; NoVLX-LABEL: test_vpcmpeqd_v16i1_v64i1_mask:
2509 ; NoVLX: # %bb.0: # %entry
2510 ; NoVLX-NEXT: vpcmpeqd %zmm1, %zmm0, %k0
2511 ; NoVLX-NEXT: kmovw %k0, %eax
2512 ; NoVLX-NEXT: vzeroupper
2513 ; NoVLX-NEXT: retq
2514 entry:
2515 %0 = bitcast <8 x i64> %__a to <16 x i32>
2516 %1 = bitcast <8 x i64> %__b to <16 x i32>
2517 %2 = icmp eq <16 x i32> %0, %1
2518 %3 = shufflevector <16 x i1> %2, <16 x i1> zeroinitializer, <64 x i32> <i32 0,i32 1,i32 2,i32 3,i32 4,i32 5,i32 6,i32 7,i32 8,i32 9,i32 10,i32 11,i32 12,i32 13,i32 14,i32 15,i32 16,i32 17,i32 18,i32 19,i32 20,i32 21,i32 22,i32 23,i32 24,i32 25,i32 26,i32 27,i32 28,i32 29,i32 30,i32 31,i32 16,i32 17,i32 18,i32 19,i32 20,i32 21,i32 22,i32 23,i32 24,i32 25,i32 26,i32 27,i32 28,i32 29,i32 30,i32 31,i32 16,i32 17,i32 18,i32 19,i32 20,i32 21,i32 22,i32 23,i32 24,i32 25,i32 26,i32 27,i32 28,i32 29,i32 30,i32 31>
2519 %4 = bitcast <64 x i1> %3 to i64
2520 ret i64 %4
2521 }
2523 define zeroext i64 @test_vpcmpeqd_v16i1_v64i1_mask_mem(<8 x i64> %__a, <8 x i64>* %__b) local_unnamed_addr {
2524 ; VLX-LABEL: test_vpcmpeqd_v16i1_v64i1_mask_mem:
2525 ; VLX: # %bb.0: # %entry
2526 ; VLX-NEXT: vpcmpeqd (%rdi), %zmm0, %k0
2527 ; VLX-NEXT: kmovq %k0, %rax
2528 ; VLX-NEXT: vzeroupper
2529 ; VLX-NEXT: retq
2530 ;
2531 ; NoVLX-LABEL: test_vpcmpeqd_v16i1_v64i1_mask_mem:
2532 ; NoVLX: # %bb.0: # %entry
2533 ; NoVLX-NEXT: vpcmpeqd (%rdi), %zmm0, %k0
2534 ; NoVLX-NEXT: kmovw %k0, %eax
2535 ; NoVLX-NEXT: vzeroupper
2536 ; NoVLX-NEXT: retq
2537 entry:
2538 %0 = bitcast <8 x i64> %__a to <16 x i32>
2539 %load = load <8 x i64>, <8 x i64>* %__b
2540 %1 = bitcast <8 x i64> %load to <16 x i32>
2541 %2 = icmp eq <16 x i32> %0, %1
2542 %3 = shufflevector <16 x i1> %2, <16 x i1> zeroinitializer, <64 x i32> <i32 0,i32 1,i32 2,i32 3,i32 4,i32 5,i32 6,i32 7,i32 8,i32 9,i32 10,i32 11,i32 12,i32 13,i32 14,i32 15,i32 16,i32 17,i32 18,i32 19,i32 20,i32 21,i32 22,i32 23,i32 24,i32 25,i32 26,i32 27,i32 28,i32 29,i32 30,i32 31,i32 16,i32 17,i32 18,i32 19,i32 20,i32 21,i32 22,i32 23,i32 24,i32 25,i32 26,i32 27,i32 28,i32 29,i32 30,i32 31,i32 16,i32 17,i32 18,i32 19,i32 20,i32 21,i32 22,i32 23,i32 24,i32 25,i32 26,i32 27,i32 28,i32 29,i32 30,i32 31>
2543 %4 = bitcast <64 x i1> %3 to i64
2544 ret i64 %4
2545 }
2547 define zeroext i64 @test_masked_vpcmpeqd_v16i1_v64i1_mask(i16 zeroext %__u, <8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
2548 ; VLX-LABEL: test_masked_vpcmpeqd_v16i1_v64i1_mask:
2549 ; VLX: # %bb.0: # %entry
2550 ; VLX-NEXT: kmovd %edi, %k1
2551 ; VLX-NEXT: vpcmpeqd %zmm1, %zmm0, %k0 {%k1}
2552 ; VLX-NEXT: kmovq %k0, %rax
2553 ; VLX-NEXT: vzeroupper
2554 ; VLX-NEXT: retq
2555 ;
2556 ; NoVLX-LABEL: test_masked_vpcmpeqd_v16i1_v64i1_mask:
2557 ; NoVLX: # %bb.0: # %entry
2558 ; NoVLX-NEXT: vpcmpeqd %zmm1, %zmm0, %k0
2559 ; NoVLX-NEXT: kmovw %k0, %eax
2560 ; NoVLX-NEXT: andl %edi, %eax
2561 ; NoVLX-NEXT: vzeroupper
2562 ; NoVLX-NEXT: retq
2563 entry:
2564 %0 = bitcast <8 x i64> %__a to <16 x i32>
2565 %1 = bitcast <8 x i64> %__b to <16 x i32>
2566 %2 = icmp eq <16 x i32> %0, %1
2567 %3 = bitcast i16 %__u to <16 x i1>
2568 %4 = and <16 x i1> %2, %3
2569 %5 = shufflevector <16 x i1> %4, <16 x i1> zeroinitializer, <64 x i32> <i32 0,i32 1,i32 2,i32 3,i32 4,i32 5,i32 6,i32 7,i32 8,i32 9,i32 10,i32 11,i32 12,i32 13,i32 14,i32 15,i32 16,i32 17,i32 18,i32 19,i32 20,i32 21,i32 22,i32 23,i32 24,i32 25,i32 26,i32 27,i32 28,i32 29,i32 30,i32 31,i32 16,i32 17,i32 18,i32 19,i32 20,i32 21,i32 22,i32 23,i32 24,i32 25,i32 26,i32 27,i32 28,i32 29,i32 30,i32 31,i32 16,i32 17,i32 18,i32 19,i32 20,i32 21,i32 22,i32 23,i32 24,i32 25,i32 26,i32 27,i32 28,i32 29,i32 30,i32 31>
2570 %6 = bitcast <64 x i1> %5 to i64
2571 ret i64 %6
2572 }
2574 define zeroext i64 @test_masked_vpcmpeqd_v16i1_v64i1_mask_mem(i16 zeroext %__u, <8 x i64> %__a, <8 x i64>* %__b) local_unnamed_addr {
2575 ; VLX-LABEL: test_masked_vpcmpeqd_v16i1_v64i1_mask_mem:
2576 ; VLX: # %bb.0: # %entry
2577 ; VLX-NEXT: kmovd %edi, %k1
2578 ; VLX-NEXT: vpcmpeqd (%rsi), %zmm0, %k0 {%k1}
2579 ; VLX-NEXT: kmovq %k0, %rax
2580 ; VLX-NEXT: vzeroupper
2581 ; VLX-NEXT: retq
2582 ;
2583 ; NoVLX-LABEL: test_masked_vpcmpeqd_v16i1_v64i1_mask_mem:
2584 ; NoVLX: # %bb.0: # %entry
2585 ; NoVLX-NEXT: vpcmpeqd (%rsi), %zmm0, %k0
2586 ; NoVLX-NEXT: kmovw %k0, %eax
2587 ; NoVLX-NEXT: andl %edi, %eax
2588 ; NoVLX-NEXT: vzeroupper
2589 ; NoVLX-NEXT: retq
2590 entry:
2591 %0 = bitcast <8 x i64> %__a to <16 x i32>
2592 %load = load <8 x i64>, <8 x i64>* %__b
2593 %1 = bitcast <8 x i64> %load to <16 x i32>
2594 %2 = icmp eq <16 x i32> %0, %1
2595 %3 = bitcast i16 %__u to <16 x i1>
2596 %4 = and <16 x i1> %2, %3
2597 %5 = shufflevector <16 x i1> %4, <16 x i1> zeroinitializer, <64 x i32> <i32 0,i32 1,i32 2,i32 3,i32 4,i32 5,i32 6,i32 7,i32 8,i32 9,i32 10,i32 11,i32 12,i32 13,i32 14,i32 15,i32 16,i32 17,i32 18,i32 19,i32 20,i32 21,i32 22,i32 23,i32 24,i32 25,i32 26,i32 27,i32 28,i32 29,i32 30,i32 31,i32 16,i32 17,i32 18,i32 19,i32 20,i32 21,i32 22,i32 23,i32 24,i32 25,i32 26,i32 27,i32 28,i32 29,i32 30,i32 31,i32 16,i32 17,i32 18,i32 19,i32 20,i32 21,i32 22,i32 23,i32 24,i32 25,i32 26,i32 27,i32 28,i32 29,i32 30,i32 31>
2598 %6 = bitcast <64 x i1> %5 to i64
2599 ret i64 %6
2600 }
2603 define zeroext i64 @test_vpcmpeqd_v16i1_v64i1_mask_mem_b(<8 x i64> %__a, i32* %__b) local_unnamed_addr {
2604 ; VLX-LABEL: test_vpcmpeqd_v16i1_v64i1_mask_mem_b:
2605 ; VLX: # %bb.0: # %entry
2606 ; VLX-NEXT: vpcmpeqd (%rdi){1to16}, %zmm0, %k0
2607 ; VLX-NEXT: kmovq %k0, %rax
2608 ; VLX-NEXT: vzeroupper
2609 ; VLX-NEXT: retq
2610 ;
2611 ; NoVLX-LABEL: test_vpcmpeqd_v16i1_v64i1_mask_mem_b:
2612 ; NoVLX: # %bb.0: # %entry
2613 ; NoVLX-NEXT: vpcmpeqd (%rdi){1to16}, %zmm0, %k0
2614 ; NoVLX-NEXT: kmovw %k0, %eax
2615 ; NoVLX-NEXT: vzeroupper
2616 ; NoVLX-NEXT: retq
2617 entry:
2618 %0 = bitcast <8 x i64> %__a to <16 x i32>
2619 %load = load i32, i32* %__b
2620 %vec = insertelement <16 x i32> undef, i32 %load, i32 0
2621 %1 = shufflevector <16 x i32> %vec, <16 x i32> undef, <16 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
2622 %2 = icmp eq <16 x i32> %0, %1
2623 %3 = shufflevector <16 x i1> %2, <16 x i1> zeroinitializer, <64 x i32> <i32 0,i32 1,i32 2,i32 3,i32 4,i32 5,i32 6,i32 7,i32 8,i32 9,i32 10,i32 11,i32 12,i32 13,i32 14,i32 15,i32 16,i32 17,i32 18,i32 19,i32 20,i32 21,i32 22,i32 23,i32 24,i32 25,i32 26,i32 27,i32 28,i32 29,i32 30,i32 31,i32 16,i32 17,i32 18,i32 19,i32 20,i32 21,i32 22,i32 23,i32 24,i32 25,i32 26,i32 27,i32 28,i32 29,i32 30,i32 31,i32 16,i32 17,i32 18,i32 19,i32 20,i32 21,i32 22,i32 23,i32 24,i32 25,i32 26,i32 27,i32 28,i32 29,i32 30,i32 31>
2624 %4 = bitcast <64 x i1> %3 to i64
2625 ret i64 %4
2626 }
2628 define zeroext i64 @test_masked_vpcmpeqd_v16i1_v64i1_mask_mem_b(i16 zeroext %__u, <8 x i64> %__a, i32* %__b) local_unnamed_addr {
2629 ; VLX-LABEL: test_masked_vpcmpeqd_v16i1_v64i1_mask_mem_b:
2630 ; VLX: # %bb.0: # %entry
2631 ; VLX-NEXT: kmovd %edi, %k1
2632 ; VLX-NEXT: vpcmpeqd (%rsi){1to16}, %zmm0, %k0 {%k1}
2633 ; VLX-NEXT: kmovq %k0, %rax
2634 ; VLX-NEXT: vzeroupper
2635 ; VLX-NEXT: retq
2636 ;
2637 ; NoVLX-LABEL: test_masked_vpcmpeqd_v16i1_v64i1_mask_mem_b:
2638 ; NoVLX: # %bb.0: # %entry
2639 ; NoVLX-NEXT: vpcmpeqd (%rsi){1to16}, %zmm0, %k0
2640 ; NoVLX-NEXT: kmovw %k0, %eax
2641 ; NoVLX-NEXT: andl %edi, %eax
2642 ; NoVLX-NEXT: vzeroupper
2643 ; NoVLX-NEXT: retq
2644 entry:
2645 %0 = bitcast <8 x i64> %__a to <16 x i32>
2646 %load = load i32, i32* %__b
2647 %vec = insertelement <16 x i32> undef, i32 %load, i32 0
2648 %1 = shufflevector <16 x i32> %vec, <16 x i32> undef, <16 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
2649 %2 = icmp eq <16 x i32> %0, %1
2650 %3 = bitcast i16 %__u to <16 x i1>
2651 %4 = and <16 x i1> %3, %2
2652 %5 = shufflevector <16 x i1> %4, <16 x i1> zeroinitializer, <64 x i32> <i32 0,i32 1,i32 2,i32 3,i32 4,i32 5,i32 6,i32 7,i32 8,i32 9,i32 10,i32 11,i32 12,i32 13,i32 14,i32 15,i32 16,i32 17,i32 18,i32 19,i32 20,i32 21,i32 22,i32 23,i32 24,i32 25,i32 26,i32 27,i32 28,i32 29,i32 30,i32 31,i32 16,i32 17,i32 18,i32 19,i32 20,i32 21,i32 22,i32 23,i32 24,i32 25,i32 26,i32 27,i32 28,i32 29,i32 30,i32 31,i32 16,i32 17,i32 18,i32 19,i32 20,i32 21,i32 22,i32 23,i32 24,i32 25,i32 26,i32 27,i32 28,i32 29,i32 30,i32 31>
2653 %6 = bitcast <64 x i1> %5 to i64
2654 ret i64 %6
2655 }
2658 define zeroext i4 @test_vpcmpeqq_v2i1_v4i1_mask(<2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
2659 ; VLX-LABEL: test_vpcmpeqq_v2i1_v4i1_mask:
2660 ; VLX: # %bb.0: # %entry
2661 ; VLX-NEXT: vpcmpeqq %xmm1, %xmm0, %k0
2662 ; VLX-NEXT: kmovb %k0, %eax
2663 ; VLX-NEXT: retq
2664 ;
2665 ; NoVLX-LABEL: test_vpcmpeqq_v2i1_v4i1_mask:
2666 ; NoVLX: # %bb.0: # %entry
2667 ; NoVLX-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
2668 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
2669 ; NoVLX-NEXT: vpcmpeqq %zmm1, %zmm0, %k0
2670 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
2671 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
2672 ; NoVLX-NEXT: kmovw %k0, %eax
2673 ; NoVLX-NEXT: vzeroupper
2674 ; NoVLX-NEXT: retq
2675 entry:
2676 %0 = bitcast <2 x i64> %__a to <2 x i64>
2677 %1 = bitcast <2 x i64> %__b to <2 x i64>
2678 %2 = icmp eq <2 x i64> %0, %1
2679 %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
2680 %4 = bitcast <4 x i1> %3 to i4
2681 ret i4 %4
2682 }
2684 define zeroext i4 @test_vpcmpeqq_v2i1_v4i1_mask_mem(<2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
2685 ; VLX-LABEL: test_vpcmpeqq_v2i1_v4i1_mask_mem:
2686 ; VLX: # %bb.0: # %entry
2687 ; VLX-NEXT: vpcmpeqq (%rdi), %xmm0, %k0
2688 ; VLX-NEXT: kmovb %k0, %eax
2689 ; VLX-NEXT: retq
2690 ;
2691 ; NoVLX-LABEL: test_vpcmpeqq_v2i1_v4i1_mask_mem:
2692 ; NoVLX: # %bb.0: # %entry
2693 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
2694 ; NoVLX-NEXT: vmovdqa (%rdi), %xmm1
2695 ; NoVLX-NEXT: vpcmpeqq %zmm1, %zmm0, %k0
2696 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
2697 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
2698 ; NoVLX-NEXT: kmovw %k0, %eax
2699 ; NoVLX-NEXT: vzeroupper
2700 ; NoVLX-NEXT: retq
2701 entry:
2702 %0 = bitcast <2 x i64> %__a to <2 x i64>
2703 %load = load <2 x i64>, <2 x i64>* %__b
2704 %1 = bitcast <2 x i64> %load to <2 x i64>
2705 %2 = icmp eq <2 x i64> %0, %1
2706 %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
2707 %4 = bitcast <4 x i1> %3 to i4
2708 ret i4 %4
2709 }
2711 define zeroext i4 @test_masked_vpcmpeqq_v2i1_v4i1_mask(i8 zeroext %__u, <2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
2712 ; VLX-LABEL: test_masked_vpcmpeqq_v2i1_v4i1_mask:
2713 ; VLX: # %bb.0: # %entry
2714 ; VLX-NEXT: kmovd %edi, %k1
2715 ; VLX-NEXT: vpcmpeqq %xmm1, %xmm0, %k0 {%k1}
2716 ; VLX-NEXT: kmovb %k0, %eax
2717 ; VLX-NEXT: retq
2718 ;
2719 ; NoVLX-LABEL: test_masked_vpcmpeqq_v2i1_v4i1_mask:
2720 ; NoVLX: # %bb.0: # %entry
2721 ; NoVLX-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
2722 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
2723 ; NoVLX-NEXT: kmovw %edi, %k1
2724 ; NoVLX-NEXT: vpcmpeqq %zmm1, %zmm0, %k0 {%k1}
2725 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
2726 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
2727 ; NoVLX-NEXT: kmovw %k0, %eax
2728 ; NoVLX-NEXT: vzeroupper
2729 ; NoVLX-NEXT: retq
2730 entry:
2731 %0 = bitcast <2 x i64> %__a to <2 x i64>
2732 %1 = bitcast <2 x i64> %__b to <2 x i64>
2733 %2 = icmp eq <2 x i64> %0, %1
2734 %3 = bitcast i8 %__u to <8 x i1>
2735 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <2 x i32> <i32 0, i32 1>
2736 %4 = and <2 x i1> %2, %extract.i
2737 %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
2738 %6 = bitcast <4 x i1> %5 to i4
2739 ret i4 %6
2740 }
2742 define zeroext i4 @test_masked_vpcmpeqq_v2i1_v4i1_mask_mem(i8 zeroext %__u, <2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
2743 ; VLX-LABEL: test_masked_vpcmpeqq_v2i1_v4i1_mask_mem:
2744 ; VLX: # %bb.0: # %entry
2745 ; VLX-NEXT: kmovd %edi, %k1
2746 ; VLX-NEXT: vpcmpeqq (%rsi), %xmm0, %k0 {%k1}
2747 ; VLX-NEXT: kmovb %k0, %eax
2748 ; VLX-NEXT: retq
2749 ;
2750 ; NoVLX-LABEL: test_masked_vpcmpeqq_v2i1_v4i1_mask_mem:
2751 ; NoVLX: # %bb.0: # %entry
2752 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
2753 ; NoVLX-NEXT: vmovdqa (%rsi), %xmm1
2754 ; NoVLX-NEXT: kmovw %edi, %k1
2755 ; NoVLX-NEXT: vpcmpeqq %zmm1, %zmm0, %k0 {%k1}
2756 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
2757 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
2758 ; NoVLX-NEXT: kmovw %k0, %eax
2759 ; NoVLX-NEXT: vzeroupper
2760 ; NoVLX-NEXT: retq
2761 entry:
2762 %0 = bitcast <2 x i64> %__a to <2 x i64>
2763 %load = load <2 x i64>, <2 x i64>* %__b
2764 %1 = bitcast <2 x i64> %load to <2 x i64>
2765 %2 = icmp eq <2 x i64> %0, %1
2766 %3 = bitcast i8 %__u to <8 x i1>
2767 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <2 x i32> <i32 0, i32 1>
2768 %4 = and <2 x i1> %2, %extract.i
2769 %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
2770 %6 = bitcast <4 x i1> %5 to i4
2771 ret i4 %6
2772 }
2775 define zeroext i4 @test_vpcmpeqq_v2i1_v4i1_mask_mem_b(<2 x i64> %__a, i64* %__b) local_unnamed_addr {
2776 ; VLX-LABEL: test_vpcmpeqq_v2i1_v4i1_mask_mem_b:
2777 ; VLX: # %bb.0: # %entry
2778 ; VLX-NEXT: vpcmpeqq (%rdi){1to2}, %xmm0, %k0
2779 ; VLX-NEXT: kmovb %k0, %eax
2780 ; VLX-NEXT: retq
2781 ;
2782 ; NoVLX-LABEL: test_vpcmpeqq_v2i1_v4i1_mask_mem_b:
2783 ; NoVLX: # %bb.0: # %entry
2784 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
2785 ; NoVLX-NEXT: vpcmpeqq (%rdi){1to8}, %zmm0, %k0
2786 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
2787 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
2788 ; NoVLX-NEXT: kmovw %k0, %eax
2789 ; NoVLX-NEXT: vzeroupper
2790 ; NoVLX-NEXT: retq
2791 entry:
2792 %0 = bitcast <2 x i64> %__a to <2 x i64>
2793 %load = load i64, i64* %__b
2794 %vec = insertelement <2 x i64> undef, i64 %load, i32 0
2795 %1 = shufflevector <2 x i64> %vec, <2 x i64> undef, <2 x i32> <i32 0, i32 0>
2796 %2 = icmp eq <2 x i64> %0, %1
2797 %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
2798 %4 = bitcast <4 x i1> %3 to i4
2799 ret i4 %4
2800 }
2802 define zeroext i4 @test_masked_vpcmpeqq_v2i1_v4i1_mask_mem_b(i8 zeroext %__u, <2 x i64> %__a, i64* %__b) local_unnamed_addr {
2803 ; VLX-LABEL: test_masked_vpcmpeqq_v2i1_v4i1_mask_mem_b:
2804 ; VLX: # %bb.0: # %entry
2805 ; VLX-NEXT: kmovd %edi, %k1
2806 ; VLX-NEXT: vpcmpeqq (%rsi){1to2}, %xmm0, %k0 {%k1}
2807 ; VLX-NEXT: kmovb %k0, %eax
2808 ; VLX-NEXT: retq
2809 ;
2810 ; NoVLX-LABEL: test_masked_vpcmpeqq_v2i1_v4i1_mask_mem_b:
2811 ; NoVLX: # %bb.0: # %entry
2812 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
2813 ; NoVLX-NEXT: kmovw %edi, %k1
2814 ; NoVLX-NEXT: vpcmpeqq (%rsi){1to8}, %zmm0, %k0 {%k1}
2815 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
2816 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
2817 ; NoVLX-NEXT: kmovw %k0, %eax
2818 ; NoVLX-NEXT: vzeroupper
2819 ; NoVLX-NEXT: retq
2820 entry:
2821 %0 = bitcast <2 x i64> %__a to <2 x i64>
2822 %load = load i64, i64* %__b
2823 %vec = insertelement <2 x i64> undef, i64 %load, i32 0
2824 %1 = shufflevector <2 x i64> %vec, <2 x i64> undef, <2 x i32> <i32 0, i32 0>
2825 %2 = icmp eq <2 x i64> %0, %1
2826 %3 = bitcast i8 %__u to <8 x i1>
2827 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <2 x i32> <i32 0, i32 1>
2828 %4 = and <2 x i1> %extract.i, %2
2829 %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
2830 %6 = bitcast <4 x i1> %5 to i4
2831 ret i4 %6
2832 }
2835 define zeroext i8 @test_vpcmpeqq_v2i1_v8i1_mask(<2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
2836 ; VLX-LABEL: test_vpcmpeqq_v2i1_v8i1_mask:
2837 ; VLX: # %bb.0: # %entry
2838 ; VLX-NEXT: vpcmpeqq %xmm1, %xmm0, %k0
2839 ; VLX-NEXT: kmovd %k0, %eax
2840 ; VLX-NEXT: # kill: def $al killed $al killed $eax
2841 ; VLX-NEXT: retq
2842 ;
2843 ; NoVLX-LABEL: test_vpcmpeqq_v2i1_v8i1_mask:
2844 ; NoVLX: # %bb.0: # %entry
2845 ; NoVLX-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
2846 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
2847 ; NoVLX-NEXT: vpcmpeqq %zmm1, %zmm0, %k0
2848 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
2849 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
2850 ; NoVLX-NEXT: kmovw %k0, %eax
2851 ; NoVLX-NEXT: # kill: def $al killed $al killed $eax
2852 ; NoVLX-NEXT: vzeroupper
2853 ; NoVLX-NEXT: retq
2854 entry:
2855 %0 = bitcast <2 x i64> %__a to <2 x i64>
2856 %1 = bitcast <2 x i64> %__b to <2 x i64>
2857 %2 = icmp eq <2 x i64> %0, %1
2858 %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
2859 %4 = bitcast <8 x i1> %3 to i8
2860 ret i8 %4
2861 }
2863 define zeroext i8 @test_vpcmpeqq_v2i1_v8i1_mask_mem(<2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
2864 ; VLX-LABEL: test_vpcmpeqq_v2i1_v8i1_mask_mem:
2865 ; VLX: # %bb.0: # %entry
2866 ; VLX-NEXT: vpcmpeqq (%rdi), %xmm0, %k0
2867 ; VLX-NEXT: kmovd %k0, %eax
2868 ; VLX-NEXT: # kill: def $al killed $al killed $eax
2869 ; VLX-NEXT: retq
2870 ;
2871 ; NoVLX-LABEL: test_vpcmpeqq_v2i1_v8i1_mask_mem:
2872 ; NoVLX: # %bb.0: # %entry
2873 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
2874 ; NoVLX-NEXT: vmovdqa (%rdi), %xmm1
2875 ; NoVLX-NEXT: vpcmpeqq %zmm1, %zmm0, %k0
2876 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
2877 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
2878 ; NoVLX-NEXT: kmovw %k0, %eax
2879 ; NoVLX-NEXT: # kill: def $al killed $al killed $eax
2880 ; NoVLX-NEXT: vzeroupper
2881 ; NoVLX-NEXT: retq
2882 entry:
2883 %0 = bitcast <2 x i64> %__a to <2 x i64>
2884 %load = load <2 x i64>, <2 x i64>* %__b
2885 %1 = bitcast <2 x i64> %load to <2 x i64>
2886 %2 = icmp eq <2 x i64> %0, %1
2887 %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
2888 %4 = bitcast <8 x i1> %3 to i8
2889 ret i8 %4
2890 }
2892 define zeroext i8 @test_masked_vpcmpeqq_v2i1_v8i1_mask(i8 zeroext %__u, <2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
2893 ; VLX-LABEL: test_masked_vpcmpeqq_v2i1_v8i1_mask:
2894 ; VLX: # %bb.0: # %entry
2895 ; VLX-NEXT: kmovd %edi, %k1
2896 ; VLX-NEXT: vpcmpeqq %xmm1, %xmm0, %k0 {%k1}
2897 ; VLX-NEXT: kmovd %k0, %eax
2898 ; VLX-NEXT: # kill: def $al killed $al killed $eax
2899 ; VLX-NEXT: retq
2900 ;
2901 ; NoVLX-LABEL: test_masked_vpcmpeqq_v2i1_v8i1_mask:
2902 ; NoVLX: # %bb.0: # %entry
2903 ; NoVLX-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
2904 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
2905 ; NoVLX-NEXT: kmovw %edi, %k1
2906 ; NoVLX-NEXT: vpcmpeqq %zmm1, %zmm0, %k0 {%k1}
2907 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
2908 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
2909 ; NoVLX-NEXT: kmovw %k0, %eax
2910 ; NoVLX-NEXT: # kill: def $al killed $al killed $eax
2911 ; NoVLX-NEXT: vzeroupper
2912 ; NoVLX-NEXT: retq
2913 entry:
2914 %0 = bitcast <2 x i64> %__a to <2 x i64>
2915 %1 = bitcast <2 x i64> %__b to <2 x i64>
2916 %2 = icmp eq <2 x i64> %0, %1
2917 %3 = bitcast i8 %__u to <8 x i1>
2918 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <2 x i32> <i32 0, i32 1>
2919 %4 = and <2 x i1> %2, %extract.i
2920 %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
2921 %6 = bitcast <8 x i1> %5 to i8
2922 ret i8 %6
2923 }
2925 define zeroext i8 @test_masked_vpcmpeqq_v2i1_v8i1_mask_mem(i8 zeroext %__u, <2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
2926 ; VLX-LABEL: test_masked_vpcmpeqq_v2i1_v8i1_mask_mem:
2927 ; VLX: # %bb.0: # %entry
2928 ; VLX-NEXT: kmovd %edi, %k1
2929 ; VLX-NEXT: vpcmpeqq (%rsi), %xmm0, %k0 {%k1}
2930 ; VLX-NEXT: kmovd %k0, %eax
2931 ; VLX-NEXT: # kill: def $al killed $al killed $eax
2932 ; VLX-NEXT: retq
2933 ;
2934 ; NoVLX-LABEL: test_masked_vpcmpeqq_v2i1_v8i1_mask_mem:
2935 ; NoVLX: # %bb.0: # %entry
2936 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
2937 ; NoVLX-NEXT: vmovdqa (%rsi), %xmm1
2938 ; NoVLX-NEXT: kmovw %edi, %k1
2939 ; NoVLX-NEXT: vpcmpeqq %zmm1, %zmm0, %k0 {%k1}
2940 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
2941 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
2942 ; NoVLX-NEXT: kmovw %k0, %eax
2943 ; NoVLX-NEXT: # kill: def $al killed $al killed $eax
2944 ; NoVLX-NEXT: vzeroupper
2945 ; NoVLX-NEXT: retq
2946 entry:
2947 %0 = bitcast <2 x i64> %__a to <2 x i64>
2948 %load = load <2 x i64>, <2 x i64>* %__b
2949 %1 = bitcast <2 x i64> %load to <2 x i64>
2950 %2 = icmp eq <2 x i64> %0, %1
2951 %3 = bitcast i8 %__u to <8 x i1>
2952 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <2 x i32> <i32 0, i32 1>
2953 %4 = and <2 x i1> %2, %extract.i
2954 %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
2955 %6 = bitcast <8 x i1> %5 to i8
2956 ret i8 %6
2957 }
2960 define zeroext i8 @test_vpcmpeqq_v2i1_v8i1_mask_mem_b(<2 x i64> %__a, i64* %__b) local_unnamed_addr {
2961 ; VLX-LABEL: test_vpcmpeqq_v2i1_v8i1_mask_mem_b:
2962 ; VLX: # %bb.0: # %entry
2963 ; VLX-NEXT: vpcmpeqq (%rdi){1to2}, %xmm0, %k0
2964 ; VLX-NEXT: kmovd %k0, %eax
2965 ; VLX-NEXT: # kill: def $al killed $al killed $eax
2966 ; VLX-NEXT: retq
2967 ;
2968 ; NoVLX-LABEL: test_vpcmpeqq_v2i1_v8i1_mask_mem_b:
2969 ; NoVLX: # %bb.0: # %entry
2970 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
2971 ; NoVLX-NEXT: vpcmpeqq (%rdi){1to8}, %zmm0, %k0
2972 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
2973 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
2974 ; NoVLX-NEXT: kmovw %k0, %eax
2975 ; NoVLX-NEXT: # kill: def $al killed $al killed $eax
2976 ; NoVLX-NEXT: vzeroupper
2977 ; NoVLX-NEXT: retq
2978 entry:
2979 %0 = bitcast <2 x i64> %__a to <2 x i64>
2980 %load = load i64, i64* %__b
2981 %vec = insertelement <2 x i64> undef, i64 %load, i32 0
2982 %1 = shufflevector <2 x i64> %vec, <2 x i64> undef, <2 x i32> <i32 0, i32 0>
2983 %2 = icmp eq <2 x i64> %0, %1
2984 %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
2985 %4 = bitcast <8 x i1> %3 to i8
2986 ret i8 %4
2987 }
2989 define zeroext i8 @test_masked_vpcmpeqq_v2i1_v8i1_mask_mem_b(i8 zeroext %__u, <2 x i64> %__a, i64* %__b) local_unnamed_addr {
2990 ; VLX-LABEL: test_masked_vpcmpeqq_v2i1_v8i1_mask_mem_b:
2991 ; VLX: # %bb.0: # %entry
2992 ; VLX-NEXT: kmovd %edi, %k1
2993 ; VLX-NEXT: vpcmpeqq (%rsi){1to2}, %xmm0, %k0 {%k1}
2994 ; VLX-NEXT: kmovd %k0, %eax
2995 ; VLX-NEXT: # kill: def $al killed $al killed $eax
2996 ; VLX-NEXT: retq
2997 ;
2998 ; NoVLX-LABEL: test_masked_vpcmpeqq_v2i1_v8i1_mask_mem_b:
2999 ; NoVLX: # %bb.0: # %entry
3000 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
3001 ; NoVLX-NEXT: kmovw %edi, %k1
3002 ; NoVLX-NEXT: vpcmpeqq (%rsi){1to8}, %zmm0, %k0 {%k1}
3003 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
3004 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
3005 ; NoVLX-NEXT: kmovw %k0, %eax
3006 ; NoVLX-NEXT: # kill: def $al killed $al killed $eax
3007 ; NoVLX-NEXT: vzeroupper
3008 ; NoVLX-NEXT: retq
3009 entry:
3010 %0 = bitcast <2 x i64> %__a to <2 x i64>
3011 %load = load i64, i64* %__b
3012 %vec = insertelement <2 x i64> undef, i64 %load, i32 0
3013 %1 = shufflevector <2 x i64> %vec, <2 x i64> undef, <2 x i32> <i32 0, i32 0>
3014 %2 = icmp eq <2 x i64> %0, %1
3015 %3 = bitcast i8 %__u to <8 x i1>
3016 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <2 x i32> <i32 0, i32 1>
3017 %4 = and <2 x i1> %extract.i, %2
3018 %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
3019 %6 = bitcast <8 x i1> %5 to i8
3020 ret i8 %6
3021 }
3024 define zeroext i16 @test_vpcmpeqq_v2i1_v16i1_mask(<2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
3025 ; VLX-LABEL: test_vpcmpeqq_v2i1_v16i1_mask:
3026 ; VLX: # %bb.0: # %entry
3027 ; VLX-NEXT: vpcmpeqq %xmm1, %xmm0, %k0
3028 ; VLX-NEXT: kmovd %k0, %eax
3029 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
3030 ; VLX-NEXT: retq
3031 ;
3032 ; NoVLX-LABEL: test_vpcmpeqq_v2i1_v16i1_mask:
3033 ; NoVLX: # %bb.0: # %entry
3034 ; NoVLX-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
3035 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
3036 ; NoVLX-NEXT: vpcmpeqq %zmm1, %zmm0, %k0
3037 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
3038 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
3039 ; NoVLX-NEXT: kmovw %k0, %eax
3040 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
3041 ; NoVLX-NEXT: vzeroupper
3042 ; NoVLX-NEXT: retq
3043 entry:
3044 %0 = bitcast <2 x i64> %__a to <2 x i64>
3045 %1 = bitcast <2 x i64> %__b to <2 x i64>
3046 %2 = icmp eq <2 x i64> %0, %1
3047 %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
3048 %4 = bitcast <16 x i1> %3 to i16
3049 ret i16 %4
3050 }
3052 define zeroext i16 @test_vpcmpeqq_v2i1_v16i1_mask_mem(<2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
3053 ; VLX-LABEL: test_vpcmpeqq_v2i1_v16i1_mask_mem:
3054 ; VLX: # %bb.0: # %entry
3055 ; VLX-NEXT: vpcmpeqq (%rdi), %xmm0, %k0
3056 ; VLX-NEXT: kmovd %k0, %eax
3057 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
3058 ; VLX-NEXT: retq
3059 ;
3060 ; NoVLX-LABEL: test_vpcmpeqq_v2i1_v16i1_mask_mem:
3061 ; NoVLX: # %bb.0: # %entry
3062 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
3063 ; NoVLX-NEXT: vmovdqa (%rdi), %xmm1
3064 ; NoVLX-NEXT: vpcmpeqq %zmm1, %zmm0, %k0
3065 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
3066 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
3067 ; NoVLX-NEXT: kmovw %k0, %eax
3068 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
3069 ; NoVLX-NEXT: vzeroupper
3070 ; NoVLX-NEXT: retq
3071 entry:
3072 %0 = bitcast <2 x i64> %__a to <2 x i64>
3073 %load = load <2 x i64>, <2 x i64>* %__b
3074 %1 = bitcast <2 x i64> %load to <2 x i64>
3075 %2 = icmp eq <2 x i64> %0, %1
3076 %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
3077 %4 = bitcast <16 x i1> %3 to i16
3078 ret i16 %4
3079 }
3081 define zeroext i16 @test_masked_vpcmpeqq_v2i1_v16i1_mask(i8 zeroext %__u, <2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
3082 ; VLX-LABEL: test_masked_vpcmpeqq_v2i1_v16i1_mask:
3083 ; VLX: # %bb.0: # %entry
3084 ; VLX-NEXT: kmovd %edi, %k1
3085 ; VLX-NEXT: vpcmpeqq %xmm1, %xmm0, %k0 {%k1}
3086 ; VLX-NEXT: kmovd %k0, %eax
3087 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
3088 ; VLX-NEXT: retq
3089 ;
3090 ; NoVLX-LABEL: test_masked_vpcmpeqq_v2i1_v16i1_mask:
3091 ; NoVLX: # %bb.0: # %entry
3092 ; NoVLX-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
3093 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
3094 ; NoVLX-NEXT: kmovw %edi, %k1
3095 ; NoVLX-NEXT: vpcmpeqq %zmm1, %zmm0, %k0 {%k1}
3096 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
3097 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
3098 ; NoVLX-NEXT: kmovw %k0, %eax
3099 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
3100 ; NoVLX-NEXT: vzeroupper
3101 ; NoVLX-NEXT: retq
3102 entry:
3103 %0 = bitcast <2 x i64> %__a to <2 x i64>
3104 %1 = bitcast <2 x i64> %__b to <2 x i64>
3105 %2 = icmp eq <2 x i64> %0, %1
3106 %3 = bitcast i8 %__u to <8 x i1>
3107 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <2 x i32> <i32 0, i32 1>
3108 %4 = and <2 x i1> %2, %extract.i
3109 %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
3110 %6 = bitcast <16 x i1> %5 to i16
3111 ret i16 %6
3112 }
3114 define zeroext i16 @test_masked_vpcmpeqq_v2i1_v16i1_mask_mem(i8 zeroext %__u, <2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
3115 ; VLX-LABEL: test_masked_vpcmpeqq_v2i1_v16i1_mask_mem:
3116 ; VLX: # %bb.0: # %entry
3117 ; VLX-NEXT: kmovd %edi, %k1
3118 ; VLX-NEXT: vpcmpeqq (%rsi), %xmm0, %k0 {%k1}
3119 ; VLX-NEXT: kmovd %k0, %eax
3120 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
3121 ; VLX-NEXT: retq
3122 ;
3123 ; NoVLX-LABEL: test_masked_vpcmpeqq_v2i1_v16i1_mask_mem:
3124 ; NoVLX: # %bb.0: # %entry
3125 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
3126 ; NoVLX-NEXT: vmovdqa (%rsi), %xmm1
3127 ; NoVLX-NEXT: kmovw %edi, %k1
3128 ; NoVLX-NEXT: vpcmpeqq %zmm1, %zmm0, %k0 {%k1}
3129 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
3130 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
3131 ; NoVLX-NEXT: kmovw %k0, %eax
3132 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
3133 ; NoVLX-NEXT: vzeroupper
3134 ; NoVLX-NEXT: retq
3135 entry:
3136 %0 = bitcast <2 x i64> %__a to <2 x i64>
3137 %load = load <2 x i64>, <2 x i64>* %__b
3138 %1 = bitcast <2 x i64> %load to <2 x i64>
3139 %2 = icmp eq <2 x i64> %0, %1
3140 %3 = bitcast i8 %__u to <8 x i1>
3141 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <2 x i32> <i32 0, i32 1>
3142 %4 = and <2 x i1> %2, %extract.i
3143 %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
3144 %6 = bitcast <16 x i1> %5 to i16
3145 ret i16 %6
3146 }
3149 define zeroext i16 @test_vpcmpeqq_v2i1_v16i1_mask_mem_b(<2 x i64> %__a, i64* %__b) local_unnamed_addr {
3150 ; VLX-LABEL: test_vpcmpeqq_v2i1_v16i1_mask_mem_b:
3151 ; VLX: # %bb.0: # %entry
3152 ; VLX-NEXT: vpcmpeqq (%rdi){1to2}, %xmm0, %k0
3153 ; VLX-NEXT: kmovd %k0, %eax
3154 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
3155 ; VLX-NEXT: retq
3156 ;
3157 ; NoVLX-LABEL: test_vpcmpeqq_v2i1_v16i1_mask_mem_b:
3158 ; NoVLX: # %bb.0: # %entry
3159 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
3160 ; NoVLX-NEXT: vpcmpeqq (%rdi){1to8}, %zmm0, %k0
3161 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
3162 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
3163 ; NoVLX-NEXT: kmovw %k0, %eax
3164 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
3165 ; NoVLX-NEXT: vzeroupper
3166 ; NoVLX-NEXT: retq
3167 entry:
3168 %0 = bitcast <2 x i64> %__a to <2 x i64>
3169 %load = load i64, i64* %__b
3170 %vec = insertelement <2 x i64> undef, i64 %load, i32 0
3171 %1 = shufflevector <2 x i64> %vec, <2 x i64> undef, <2 x i32> <i32 0, i32 0>
3172 %2 = icmp eq <2 x i64> %0, %1
3173 %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
3174 %4 = bitcast <16 x i1> %3 to i16
3175 ret i16 %4
3176 }
3178 define zeroext i16 @test_masked_vpcmpeqq_v2i1_v16i1_mask_mem_b(i8 zeroext %__u, <2 x i64> %__a, i64* %__b) local_unnamed_addr {
3179 ; VLX-LABEL: test_masked_vpcmpeqq_v2i1_v16i1_mask_mem_b:
3180 ; VLX: # %bb.0: # %entry
3181 ; VLX-NEXT: kmovd %edi, %k1
3182 ; VLX-NEXT: vpcmpeqq (%rsi){1to2}, %xmm0, %k0 {%k1}
3183 ; VLX-NEXT: kmovd %k0, %eax
3184 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
3185 ; VLX-NEXT: retq
3186 ;
3187 ; NoVLX-LABEL: test_masked_vpcmpeqq_v2i1_v16i1_mask_mem_b:
3188 ; NoVLX: # %bb.0: # %entry
3189 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
3190 ; NoVLX-NEXT: kmovw %edi, %k1
3191 ; NoVLX-NEXT: vpcmpeqq (%rsi){1to8}, %zmm0, %k0 {%k1}
3192 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
3193 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
3194 ; NoVLX-NEXT: kmovw %k0, %eax
3195 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
3196 ; NoVLX-NEXT: vzeroupper
3197 ; NoVLX-NEXT: retq
3198 entry:
3199 %0 = bitcast <2 x i64> %__a to <2 x i64>
3200 %load = load i64, i64* %__b
3201 %vec = insertelement <2 x i64> undef, i64 %load, i32 0
3202 %1 = shufflevector <2 x i64> %vec, <2 x i64> undef, <2 x i32> <i32 0, i32 0>
3203 %2 = icmp eq <2 x i64> %0, %1
3204 %3 = bitcast i8 %__u to <8 x i1>
3205 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <2 x i32> <i32 0, i32 1>
3206 %4 = and <2 x i1> %extract.i, %2
3207 %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
3208 %6 = bitcast <16 x i1> %5 to i16
3209 ret i16 %6
3210 }
3213 define zeroext i32 @test_vpcmpeqq_v2i1_v32i1_mask(<2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
3214 ; VLX-LABEL: test_vpcmpeqq_v2i1_v32i1_mask:
3215 ; VLX: # %bb.0: # %entry
3216 ; VLX-NEXT: vpcmpeqq %xmm1, %xmm0, %k0
3217 ; VLX-NEXT: kmovd %k0, %eax
3218 ; VLX-NEXT: retq
3219 ;
3220 ; NoVLX-LABEL: test_vpcmpeqq_v2i1_v32i1_mask:
3221 ; NoVLX: # %bb.0: # %entry
3222 ; NoVLX-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
3223 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
3224 ; NoVLX-NEXT: vpcmpeqq %zmm1, %zmm0, %k0
3225 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
3226 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
3227 ; NoVLX-NEXT: kmovw %k0, %eax
3228 ; NoVLX-NEXT: vzeroupper
3229 ; NoVLX-NEXT: retq
3230 entry:
3231 %0 = bitcast <2 x i64> %__a to <2 x i64>
3232 %1 = bitcast <2 x i64> %__b to <2 x i64>
3233 %2 = icmp eq <2 x i64> %0, %1
3234 %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
3235 %4 = bitcast <32 x i1> %3 to i32
3236 ret i32 %4
3237 }
3239 define zeroext i32 @test_vpcmpeqq_v2i1_v32i1_mask_mem(<2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
3240 ; VLX-LABEL: test_vpcmpeqq_v2i1_v32i1_mask_mem:
3241 ; VLX: # %bb.0: # %entry
3242 ; VLX-NEXT: vpcmpeqq (%rdi), %xmm0, %k0
3243 ; VLX-NEXT: kmovd %k0, %eax
3244 ; VLX-NEXT: retq
3245 ;
3246 ; NoVLX-LABEL: test_vpcmpeqq_v2i1_v32i1_mask_mem:
3247 ; NoVLX: # %bb.0: # %entry
3248 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
3249 ; NoVLX-NEXT: vmovdqa (%rdi), %xmm1
3250 ; NoVLX-NEXT: vpcmpeqq %zmm1, %zmm0, %k0
3251 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
3252 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
3253 ; NoVLX-NEXT: kmovw %k0, %eax
3254 ; NoVLX-NEXT: vzeroupper
3255 ; NoVLX-NEXT: retq
3256 entry:
3257 %0 = bitcast <2 x i64> %__a to <2 x i64>
3258 %load = load <2 x i64>, <2 x i64>* %__b
3259 %1 = bitcast <2 x i64> %load to <2 x i64>
3260 %2 = icmp eq <2 x i64> %0, %1
3261 %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
3262 %4 = bitcast <32 x i1> %3 to i32
3263 ret i32 %4
3264 }
3266 define zeroext i32 @test_masked_vpcmpeqq_v2i1_v32i1_mask(i8 zeroext %__u, <2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
3267 ; VLX-LABEL: test_masked_vpcmpeqq_v2i1_v32i1_mask:
3268 ; VLX: # %bb.0: # %entry
3269 ; VLX-NEXT: kmovd %edi, %k1
3270 ; VLX-NEXT: vpcmpeqq %xmm1, %xmm0, %k0 {%k1}
3271 ; VLX-NEXT: kmovd %k0, %eax
3272 ; VLX-NEXT: retq
3273 ;
3274 ; NoVLX-LABEL: test_masked_vpcmpeqq_v2i1_v32i1_mask:
3275 ; NoVLX: # %bb.0: # %entry
3276 ; NoVLX-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
3277 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
3278 ; NoVLX-NEXT: kmovw %edi, %k1
3279 ; NoVLX-NEXT: vpcmpeqq %zmm1, %zmm0, %k0 {%k1}
3280 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
3281 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
3282 ; NoVLX-NEXT: kmovw %k0, %eax
3283 ; NoVLX-NEXT: vzeroupper
3284 ; NoVLX-NEXT: retq
3285 entry:
3286 %0 = bitcast <2 x i64> %__a to <2 x i64>
3287 %1 = bitcast <2 x i64> %__b to <2 x i64>
3288 %2 = icmp eq <2 x i64> %0, %1
3289 %3 = bitcast i8 %__u to <8 x i1>
3290 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <2 x i32> <i32 0, i32 1>
3291 %4 = and <2 x i1> %2, %extract.i
3292 %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
3293 %6 = bitcast <32 x i1> %5 to i32
3294 ret i32 %6
3295 }
3297 define zeroext i32 @test_masked_vpcmpeqq_v2i1_v32i1_mask_mem(i8 zeroext %__u, <2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
3298 ; VLX-LABEL: test_masked_vpcmpeqq_v2i1_v32i1_mask_mem:
3299 ; VLX: # %bb.0: # %entry
3300 ; VLX-NEXT: kmovd %edi, %k1
3301 ; VLX-NEXT: vpcmpeqq (%rsi), %xmm0, %k0 {%k1}
3302 ; VLX-NEXT: kmovd %k0, %eax
3303 ; VLX-NEXT: retq
3304 ;
3305 ; NoVLX-LABEL: test_masked_vpcmpeqq_v2i1_v32i1_mask_mem:
3306 ; NoVLX: # %bb.0: # %entry
3307 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
3308 ; NoVLX-NEXT: vmovdqa (%rsi), %xmm1
3309 ; NoVLX-NEXT: kmovw %edi, %k1
3310 ; NoVLX-NEXT: vpcmpeqq %zmm1, %zmm0, %k0 {%k1}
3311 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
3312 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
3313 ; NoVLX-NEXT: kmovw %k0, %eax
3314 ; NoVLX-NEXT: vzeroupper
3315 ; NoVLX-NEXT: retq
3316 entry:
3317 %0 = bitcast <2 x i64> %__a to <2 x i64>
3318 %load = load <2 x i64>, <2 x i64>* %__b
3319 %1 = bitcast <2 x i64> %load to <2 x i64>
3320 %2 = icmp eq <2 x i64> %0, %1
3321 %3 = bitcast i8 %__u to <8 x i1>
3322 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <2 x i32> <i32 0, i32 1>
3323 %4 = and <2 x i1> %2, %extract.i
3324 %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
3325 %6 = bitcast <32 x i1> %5 to i32
3326 ret i32 %6
3327 }
3330 define zeroext i32 @test_vpcmpeqq_v2i1_v32i1_mask_mem_b(<2 x i64> %__a, i64* %__b) local_unnamed_addr {
3331 ; VLX-LABEL: test_vpcmpeqq_v2i1_v32i1_mask_mem_b:
3332 ; VLX: # %bb.0: # %entry
3333 ; VLX-NEXT: vpcmpeqq (%rdi){1to2}, %xmm0, %k0
3334 ; VLX-NEXT: kmovd %k0, %eax
3335 ; VLX-NEXT: retq
3336 ;
3337 ; NoVLX-LABEL: test_vpcmpeqq_v2i1_v32i1_mask_mem_b:
3338 ; NoVLX: # %bb.0: # %entry
3339 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
3340 ; NoVLX-NEXT: vpcmpeqq (%rdi){1to8}, %zmm0, %k0
3341 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
3342 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
3343 ; NoVLX-NEXT: kmovw %k0, %eax
3344 ; NoVLX-NEXT: vzeroupper
3345 ; NoVLX-NEXT: retq
3346 entry:
3347 %0 = bitcast <2 x i64> %__a to <2 x i64>
3348 %load = load i64, i64* %__b
3349 %vec = insertelement <2 x i64> undef, i64 %load, i32 0
3350 %1 = shufflevector <2 x i64> %vec, <2 x i64> undef, <2 x i32> <i32 0, i32 0>
3351 %2 = icmp eq <2 x i64> %0, %1
3352 %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
3353 %4 = bitcast <32 x i1> %3 to i32
3354 ret i32 %4
3355 }
3357 define zeroext i32 @test_masked_vpcmpeqq_v2i1_v32i1_mask_mem_b(i8 zeroext %__u, <2 x i64> %__a, i64* %__b) local_unnamed_addr {
3358 ; VLX-LABEL: test_masked_vpcmpeqq_v2i1_v32i1_mask_mem_b:
3359 ; VLX: # %bb.0: # %entry
3360 ; VLX-NEXT: kmovd %edi, %k1
3361 ; VLX-NEXT: vpcmpeqq (%rsi){1to2}, %xmm0, %k0 {%k1}
3362 ; VLX-NEXT: kmovd %k0, %eax
3363 ; VLX-NEXT: retq
3364 ;
3365 ; NoVLX-LABEL: test_masked_vpcmpeqq_v2i1_v32i1_mask_mem_b:
3366 ; NoVLX: # %bb.0: # %entry
3367 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
3368 ; NoVLX-NEXT: kmovw %edi, %k1
3369 ; NoVLX-NEXT: vpcmpeqq (%rsi){1to8}, %zmm0, %k0 {%k1}
3370 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
3371 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
3372 ; NoVLX-NEXT: kmovw %k0, %eax
3373 ; NoVLX-NEXT: vzeroupper
3374 ; NoVLX-NEXT: retq
3375 entry:
3376 %0 = bitcast <2 x i64> %__a to <2 x i64>
3377 %load = load i64, i64* %__b
3378 %vec = insertelement <2 x i64> undef, i64 %load, i32 0
3379 %1 = shufflevector <2 x i64> %vec, <2 x i64> undef, <2 x i32> <i32 0, i32 0>
3380 %2 = icmp eq <2 x i64> %0, %1
3381 %3 = bitcast i8 %__u to <8 x i1>
3382 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <2 x i32> <i32 0, i32 1>
3383 %4 = and <2 x i1> %extract.i, %2
3384 %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
3385 %6 = bitcast <32 x i1> %5 to i32
3386 ret i32 %6
3387 }
3390 define zeroext i64 @test_vpcmpeqq_v2i1_v64i1_mask(<2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
3391 ; VLX-LABEL: test_vpcmpeqq_v2i1_v64i1_mask:
3392 ; VLX: # %bb.0: # %entry
3393 ; VLX-NEXT: vpcmpeqq %xmm1, %xmm0, %k0
3394 ; VLX-NEXT: kmovq %k0, %rax
3395 ; VLX-NEXT: retq
3396 ;
3397 ; NoVLX-LABEL: test_vpcmpeqq_v2i1_v64i1_mask:
3398 ; NoVLX: # %bb.0: # %entry
3399 ; NoVLX-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
3400 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
3401 ; NoVLX-NEXT: vpcmpeqq %zmm1, %zmm0, %k0
3402 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
3403 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
3404 ; NoVLX-NEXT: kmovw %k0, %eax
3405 ; NoVLX-NEXT: vzeroupper
3406 ; NoVLX-NEXT: retq
3407 entry:
3408 %0 = bitcast <2 x i64> %__a to <2 x i64>
3409 %1 = bitcast <2 x i64> %__b to <2 x i64>
3410 %2 = icmp eq <2 x i64> %0, %1
3411 %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
3412 %4 = bitcast <64 x i1> %3 to i64
3413 ret i64 %4
3414 }
3416 define zeroext i64 @test_vpcmpeqq_v2i1_v64i1_mask_mem(<2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
3417 ; VLX-LABEL: test_vpcmpeqq_v2i1_v64i1_mask_mem:
3418 ; VLX: # %bb.0: # %entry
3419 ; VLX-NEXT: vpcmpeqq (%rdi), %xmm0, %k0
3420 ; VLX-NEXT: kmovq %k0, %rax
3421 ; VLX-NEXT: retq
3422 ;
3423 ; NoVLX-LABEL: test_vpcmpeqq_v2i1_v64i1_mask_mem:
3424 ; NoVLX: # %bb.0: # %entry
3425 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
3426 ; NoVLX-NEXT: vmovdqa (%rdi), %xmm1
3427 ; NoVLX-NEXT: vpcmpeqq %zmm1, %zmm0, %k0
3428 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
3429 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
3430 ; NoVLX-NEXT: kmovw %k0, %eax
3431 ; NoVLX-NEXT: vzeroupper
3432 ; NoVLX-NEXT: retq
3433 entry:
3434 %0 = bitcast <2 x i64> %__a to <2 x i64>
3435 %load = load <2 x i64>, <2 x i64>* %__b
3436 %1 = bitcast <2 x i64> %load to <2 x i64>
3437 %2 = icmp eq <2 x i64> %0, %1
3438 %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
3439 %4 = bitcast <64 x i1> %3 to i64
3440 ret i64 %4
3441 }
3443 define zeroext i64 @test_masked_vpcmpeqq_v2i1_v64i1_mask(i8 zeroext %__u, <2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
3444 ; VLX-LABEL: test_masked_vpcmpeqq_v2i1_v64i1_mask:
3445 ; VLX: # %bb.0: # %entry
3446 ; VLX-NEXT: kmovd %edi, %k1
3447 ; VLX-NEXT: vpcmpeqq %xmm1, %xmm0, %k0 {%k1}
3448 ; VLX-NEXT: kmovq %k0, %rax
3449 ; VLX-NEXT: retq
3450 ;
3451 ; NoVLX-LABEL: test_masked_vpcmpeqq_v2i1_v64i1_mask:
3452 ; NoVLX: # %bb.0: # %entry
3453 ; NoVLX-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
3454 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
3455 ; NoVLX-NEXT: kmovw %edi, %k1
3456 ; NoVLX-NEXT: vpcmpeqq %zmm1, %zmm0, %k0 {%k1}
3457 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
3458 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
3459 ; NoVLX-NEXT: kmovw %k0, %eax
3460 ; NoVLX-NEXT: vzeroupper
3461 ; NoVLX-NEXT: retq
3462 entry:
3463 %0 = bitcast <2 x i64> %__a to <2 x i64>
3464 %1 = bitcast <2 x i64> %__b to <2 x i64>
3465 %2 = icmp eq <2 x i64> %0, %1
3466 %3 = bitcast i8 %__u to <8 x i1>
3467 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <2 x i32> <i32 0, i32 1>
3468 %4 = and <2 x i1> %2, %extract.i
3469 %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
3470 %6 = bitcast <64 x i1> %5 to i64
3471 ret i64 %6
3472 }
3474 define zeroext i64 @test_masked_vpcmpeqq_v2i1_v64i1_mask_mem(i8 zeroext %__u, <2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
3475 ; VLX-LABEL: test_masked_vpcmpeqq_v2i1_v64i1_mask_mem:
3476 ; VLX: # %bb.0: # %entry
3477 ; VLX-NEXT: kmovd %edi, %k1
3478 ; VLX-NEXT: vpcmpeqq (%rsi), %xmm0, %k0 {%k1}
3479 ; VLX-NEXT: kmovq %k0, %rax
3480 ; VLX-NEXT: retq
3481 ;
3482 ; NoVLX-LABEL: test_masked_vpcmpeqq_v2i1_v64i1_mask_mem:
3483 ; NoVLX: # %bb.0: # %entry
3484 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
3485 ; NoVLX-NEXT: vmovdqa (%rsi), %xmm1
3486 ; NoVLX-NEXT: kmovw %edi, %k1
3487 ; NoVLX-NEXT: vpcmpeqq %zmm1, %zmm0, %k0 {%k1}
3488 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
3489 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
3490 ; NoVLX-NEXT: kmovw %k0, %eax
3491 ; NoVLX-NEXT: vzeroupper
3492 ; NoVLX-NEXT: retq
3493 entry:
3494 %0 = bitcast <2 x i64> %__a to <2 x i64>
3495 %load = load <2 x i64>, <2 x i64>* %__b
3496 %1 = bitcast <2 x i64> %load to <2 x i64>
3497 %2 = icmp eq <2 x i64> %0, %1
3498 %3 = bitcast i8 %__u to <8 x i1>
3499 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <2 x i32> <i32 0, i32 1>
3500 %4 = and <2 x i1> %2, %extract.i
3501 %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
3502 %6 = bitcast <64 x i1> %5 to i64
3503 ret i64 %6
3504 }
3507 define zeroext i64 @test_vpcmpeqq_v2i1_v64i1_mask_mem_b(<2 x i64> %__a, i64* %__b) local_unnamed_addr {
3508 ; VLX-LABEL: test_vpcmpeqq_v2i1_v64i1_mask_mem_b:
3509 ; VLX: # %bb.0: # %entry
3510 ; VLX-NEXT: vpcmpeqq (%rdi){1to2}, %xmm0, %k0
3511 ; VLX-NEXT: kmovq %k0, %rax
3512 ; VLX-NEXT: retq
3513 ;
3514 ; NoVLX-LABEL: test_vpcmpeqq_v2i1_v64i1_mask_mem_b:
3515 ; NoVLX: # %bb.0: # %entry
3516 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
3517 ; NoVLX-NEXT: vpcmpeqq (%rdi){1to8}, %zmm0, %k0
3518 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
3519 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
3520 ; NoVLX-NEXT: kmovw %k0, %eax
3521 ; NoVLX-NEXT: vzeroupper
3522 ; NoVLX-NEXT: retq
3523 entry:
3524 %0 = bitcast <2 x i64> %__a to <2 x i64>
3525 %load = load i64, i64* %__b
3526 %vec = insertelement <2 x i64> undef, i64 %load, i32 0
3527 %1 = shufflevector <2 x i64> %vec, <2 x i64> undef, <2 x i32> <i32 0, i32 0>
3528 %2 = icmp eq <2 x i64> %0, %1
3529 %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
3530 %4 = bitcast <64 x i1> %3 to i64
3531 ret i64 %4
3532 }
3534 define zeroext i64 @test_masked_vpcmpeqq_v2i1_v64i1_mask_mem_b(i8 zeroext %__u, <2 x i64> %__a, i64* %__b) local_unnamed_addr {
3535 ; VLX-LABEL: test_masked_vpcmpeqq_v2i1_v64i1_mask_mem_b:
3536 ; VLX: # %bb.0: # %entry
3537 ; VLX-NEXT: kmovd %edi, %k1
3538 ; VLX-NEXT: vpcmpeqq (%rsi){1to2}, %xmm0, %k0 {%k1}
3539 ; VLX-NEXT: kmovq %k0, %rax
3540 ; VLX-NEXT: retq
3541 ;
3542 ; NoVLX-LABEL: test_masked_vpcmpeqq_v2i1_v64i1_mask_mem_b:
3543 ; NoVLX: # %bb.0: # %entry
3544 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
3545 ; NoVLX-NEXT: kmovw %edi, %k1
3546 ; NoVLX-NEXT: vpcmpeqq (%rsi){1to8}, %zmm0, %k0 {%k1}
3547 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
3548 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
3549 ; NoVLX-NEXT: kmovw %k0, %eax
3550 ; NoVLX-NEXT: vzeroupper
3551 ; NoVLX-NEXT: retq
3552 entry:
3553 %0 = bitcast <2 x i64> %__a to <2 x i64>
3554 %load = load i64, i64* %__b
3555 %vec = insertelement <2 x i64> undef, i64 %load, i32 0
3556 %1 = shufflevector <2 x i64> %vec, <2 x i64> undef, <2 x i32> <i32 0, i32 0>
3557 %2 = icmp eq <2 x i64> %0, %1
3558 %3 = bitcast i8 %__u to <8 x i1>
3559 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <2 x i32> <i32 0, i32 1>
3560 %4 = and <2 x i1> %extract.i, %2
3561 %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
3562 %6 = bitcast <64 x i1> %5 to i64
3563 ret i64 %6
3564 }
3567 define zeroext i8 @test_vpcmpeqq_v4i1_v8i1_mask(<4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
3568 ; VLX-LABEL: test_vpcmpeqq_v4i1_v8i1_mask:
3569 ; VLX: # %bb.0: # %entry
3570 ; VLX-NEXT: vpcmpeqq %ymm1, %ymm0, %k0
3571 ; VLX-NEXT: kmovd %k0, %eax
3572 ; VLX-NEXT: # kill: def $al killed $al killed $eax
3573 ; VLX-NEXT: vzeroupper
3574 ; VLX-NEXT: retq
3575 ;
3576 ; NoVLX-LABEL: test_vpcmpeqq_v4i1_v8i1_mask:
3577 ; NoVLX: # %bb.0: # %entry
3578 ; NoVLX-NEXT: # kill: def $ymm1 killed $ymm1 def $zmm1
3579 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
3580 ; NoVLX-NEXT: vpcmpeqq %zmm1, %zmm0, %k0
3581 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
3582 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
3583 ; NoVLX-NEXT: kmovw %k0, %eax
3584 ; NoVLX-NEXT: # kill: def $al killed $al killed $eax
3585 ; NoVLX-NEXT: vzeroupper
3586 ; NoVLX-NEXT: retq
3587 entry:
3588 %0 = bitcast <4 x i64> %__a to <4 x i64>
3589 %1 = bitcast <4 x i64> %__b to <4 x i64>
3590 %2 = icmp eq <4 x i64> %0, %1
3591 %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
3592 %4 = bitcast <8 x i1> %3 to i8
3593 ret i8 %4
3594 }
3596 define zeroext i8 @test_vpcmpeqq_v4i1_v8i1_mask_mem(<4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
3597 ; VLX-LABEL: test_vpcmpeqq_v4i1_v8i1_mask_mem:
3598 ; VLX: # %bb.0: # %entry
3599 ; VLX-NEXT: vpcmpeqq (%rdi), %ymm0, %k0
3600 ; VLX-NEXT: kmovd %k0, %eax
3601 ; VLX-NEXT: # kill: def $al killed $al killed $eax
3602 ; VLX-NEXT: vzeroupper
3603 ; VLX-NEXT: retq
3604 ;
3605 ; NoVLX-LABEL: test_vpcmpeqq_v4i1_v8i1_mask_mem:
3606 ; NoVLX: # %bb.0: # %entry
3607 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
3608 ; NoVLX-NEXT: vmovdqa (%rdi), %ymm1
3609 ; NoVLX-NEXT: vpcmpeqq %zmm1, %zmm0, %k0
3610 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
3611 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
3612 ; NoVLX-NEXT: kmovw %k0, %eax
3613 ; NoVLX-NEXT: # kill: def $al killed $al killed $eax
3614 ; NoVLX-NEXT: vzeroupper
3615 ; NoVLX-NEXT: retq
3616 entry:
3617 %0 = bitcast <4 x i64> %__a to <4 x i64>
3618 %load = load <4 x i64>, <4 x i64>* %__b
3619 %1 = bitcast <4 x i64> %load to <4 x i64>
3620 %2 = icmp eq <4 x i64> %0, %1
3621 %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
3622 %4 = bitcast <8 x i1> %3 to i8
3623 ret i8 %4
3624 }
3626 define zeroext i8 @test_masked_vpcmpeqq_v4i1_v8i1_mask(i8 zeroext %__u, <4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
3627 ; VLX-LABEL: test_masked_vpcmpeqq_v4i1_v8i1_mask:
3628 ; VLX: # %bb.0: # %entry
3629 ; VLX-NEXT: kmovd %edi, %k1
3630 ; VLX-NEXT: vpcmpeqq %ymm1, %ymm0, %k0 {%k1}
3631 ; VLX-NEXT: kmovd %k0, %eax
3632 ; VLX-NEXT: # kill: def $al killed $al killed $eax
3633 ; VLX-NEXT: vzeroupper
3634 ; VLX-NEXT: retq
3635 ;
3636 ; NoVLX-LABEL: test_masked_vpcmpeqq_v4i1_v8i1_mask:
3637 ; NoVLX: # %bb.0: # %entry
3638 ; NoVLX-NEXT: # kill: def $ymm1 killed $ymm1 def $zmm1
3639 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
3640 ; NoVLX-NEXT: kmovw %edi, %k1
3641 ; NoVLX-NEXT: vpcmpeqq %zmm1, %zmm0, %k0 {%k1}
3642 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
3643 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
3644 ; NoVLX-NEXT: kmovw %k0, %eax
3645 ; NoVLX-NEXT: # kill: def $al killed $al killed $eax
3646 ; NoVLX-NEXT: vzeroupper
3647 ; NoVLX-NEXT: retq
3648 entry:
3649 %0 = bitcast <4 x i64> %__a to <4 x i64>
3650 %1 = bitcast <4 x i64> %__b to <4 x i64>
3651 %2 = icmp eq <4 x i64> %0, %1
3652 %3 = bitcast i8 %__u to <8 x i1>
3653 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
3654 %4 = and <4 x i1> %2, %extract.i
3655 %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
3656 %6 = bitcast <8 x i1> %5 to i8
3657 ret i8 %6
3658 }
3660 define zeroext i8 @test_masked_vpcmpeqq_v4i1_v8i1_mask_mem(i8 zeroext %__u, <4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
3661 ; VLX-LABEL: test_masked_vpcmpeqq_v4i1_v8i1_mask_mem:
3662 ; VLX: # %bb.0: # %entry
3663 ; VLX-NEXT: kmovd %edi, %k1
3664 ; VLX-NEXT: vpcmpeqq (%rsi), %ymm0, %k0 {%k1}
3665 ; VLX-NEXT: kmovd %k0, %eax
3666 ; VLX-NEXT: # kill: def $al killed $al killed $eax
3667 ; VLX-NEXT: vzeroupper
3668 ; VLX-NEXT: retq
3669 ;
3670 ; NoVLX-LABEL: test_masked_vpcmpeqq_v4i1_v8i1_mask_mem:
3671 ; NoVLX: # %bb.0: # %entry
3672 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
3673 ; NoVLX-NEXT: vmovdqa (%rsi), %ymm1
3674 ; NoVLX-NEXT: kmovw %edi, %k1
3675 ; NoVLX-NEXT: vpcmpeqq %zmm1, %zmm0, %k0 {%k1}
3676 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
3677 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
3678 ; NoVLX-NEXT: kmovw %k0, %eax
3679 ; NoVLX-NEXT: # kill: def $al killed $al killed $eax
3680 ; NoVLX-NEXT: vzeroupper
3681 ; NoVLX-NEXT: retq
3682 entry:
3683 %0 = bitcast <4 x i64> %__a to <4 x i64>
3684 %load = load <4 x i64>, <4 x i64>* %__b
3685 %1 = bitcast <4 x i64> %load to <4 x i64>
3686 %2 = icmp eq <4 x i64> %0, %1
3687 %3 = bitcast i8 %__u to <8 x i1>
3688 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
3689 %4 = and <4 x i1> %2, %extract.i
3690 %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
3691 %6 = bitcast <8 x i1> %5 to i8
3692 ret i8 %6
3693 }
3696 define zeroext i8 @test_vpcmpeqq_v4i1_v8i1_mask_mem_b(<4 x i64> %__a, i64* %__b) local_unnamed_addr {
3697 ; VLX-LABEL: test_vpcmpeqq_v4i1_v8i1_mask_mem_b:
3698 ; VLX: # %bb.0: # %entry
3699 ; VLX-NEXT: vpcmpeqq (%rdi){1to4}, %ymm0, %k0
3700 ; VLX-NEXT: kmovd %k0, %eax
3701 ; VLX-NEXT: # kill: def $al killed $al killed $eax
3702 ; VLX-NEXT: vzeroupper
3703 ; VLX-NEXT: retq
3704 ;
3705 ; NoVLX-LABEL: test_vpcmpeqq_v4i1_v8i1_mask_mem_b:
3706 ; NoVLX: # %bb.0: # %entry
3707 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
3708 ; NoVLX-NEXT: vpcmpeqq (%rdi){1to8}, %zmm0, %k0
3709 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
3710 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
3711 ; NoVLX-NEXT: kmovw %k0, %eax
3712 ; NoVLX-NEXT: # kill: def $al killed $al killed $eax
3713 ; NoVLX-NEXT: vzeroupper
3714 ; NoVLX-NEXT: retq
3715 entry:
3716 %0 = bitcast <4 x i64> %__a to <4 x i64>
3717 %load = load i64, i64* %__b
3718 %vec = insertelement <4 x i64> undef, i64 %load, i32 0
3719 %1 = shufflevector <4 x i64> %vec, <4 x i64> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
3720 %2 = icmp eq <4 x i64> %0, %1
3721 %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
3722 %4 = bitcast <8 x i1> %3 to i8
3723 ret i8 %4
3724 }
3726 define zeroext i8 @test_masked_vpcmpeqq_v4i1_v8i1_mask_mem_b(i8 zeroext %__u, <4 x i64> %__a, i64* %__b) local_unnamed_addr {
3727 ; VLX-LABEL: test_masked_vpcmpeqq_v4i1_v8i1_mask_mem_b:
3728 ; VLX: # %bb.0: # %entry
3729 ; VLX-NEXT: kmovd %edi, %k1
3730 ; VLX-NEXT: vpcmpeqq (%rsi){1to4}, %ymm0, %k0 {%k1}
3731 ; VLX-NEXT: kmovd %k0, %eax
3732 ; VLX-NEXT: # kill: def $al killed $al killed $eax
3733 ; VLX-NEXT: vzeroupper
3734 ; VLX-NEXT: retq
3735 ;
3736 ; NoVLX-LABEL: test_masked_vpcmpeqq_v4i1_v8i1_mask_mem_b:
3737 ; NoVLX: # %bb.0: # %entry
3738 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
3739 ; NoVLX-NEXT: kmovw %edi, %k1
3740 ; NoVLX-NEXT: vpcmpeqq (%rsi){1to8}, %zmm0, %k0 {%k1}
3741 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
3742 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
3743 ; NoVLX-NEXT: kmovw %k0, %eax
3744 ; NoVLX-NEXT: # kill: def $al killed $al killed $eax
3745 ; NoVLX-NEXT: vzeroupper
3746 ; NoVLX-NEXT: retq
3747 entry:
3748 %0 = bitcast <4 x i64> %__a to <4 x i64>
3749 %load = load i64, i64* %__b
3750 %vec = insertelement <4 x i64> undef, i64 %load, i32 0
3751 %1 = shufflevector <4 x i64> %vec, <4 x i64> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
3752 %2 = icmp eq <4 x i64> %0, %1
3753 %3 = bitcast i8 %__u to <8 x i1>
3754 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
3755 %4 = and <4 x i1> %extract.i, %2
3756 %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
3757 %6 = bitcast <8 x i1> %5 to i8
3758 ret i8 %6
3759 }
3762 define zeroext i16 @test_vpcmpeqq_v4i1_v16i1_mask(<4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
3763 ; VLX-LABEL: test_vpcmpeqq_v4i1_v16i1_mask:
3764 ; VLX: # %bb.0: # %entry
3765 ; VLX-NEXT: vpcmpeqq %ymm1, %ymm0, %k0
3766 ; VLX-NEXT: kmovd %k0, %eax
3767 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
3768 ; VLX-NEXT: vzeroupper
3769 ; VLX-NEXT: retq
3770 ;
3771 ; NoVLX-LABEL: test_vpcmpeqq_v4i1_v16i1_mask:
3772 ; NoVLX: # %bb.0: # %entry
3773 ; NoVLX-NEXT: # kill: def $ymm1 killed $ymm1 def $zmm1
3774 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
3775 ; NoVLX-NEXT: vpcmpeqq %zmm1, %zmm0, %k0
3776 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
3777 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
3778 ; NoVLX-NEXT: kmovw %k0, %eax
3779 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
3780 ; NoVLX-NEXT: vzeroupper
3781 ; NoVLX-NEXT: retq
3782 entry:
3783 %0 = bitcast <4 x i64> %__a to <4 x i64>
3784 %1 = bitcast <4 x i64> %__b to <4 x i64>
3785 %2 = icmp eq <4 x i64> %0, %1
3786 %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
3787 %4 = bitcast <16 x i1> %3 to i16
3788 ret i16 %4
3789 }
3791 define zeroext i16 @test_vpcmpeqq_v4i1_v16i1_mask_mem(<4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
3792 ; VLX-LABEL: test_vpcmpeqq_v4i1_v16i1_mask_mem:
3793 ; VLX: # %bb.0: # %entry
3794 ; VLX-NEXT: vpcmpeqq (%rdi), %ymm0, %k0
3795 ; VLX-NEXT: kmovd %k0, %eax
3796 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
3797 ; VLX-NEXT: vzeroupper
3798 ; VLX-NEXT: retq
3799 ;
3800 ; NoVLX-LABEL: test_vpcmpeqq_v4i1_v16i1_mask_mem:
3801 ; NoVLX: # %bb.0: # %entry
3802 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
3803 ; NoVLX-NEXT: vmovdqa (%rdi), %ymm1
3804 ; NoVLX-NEXT: vpcmpeqq %zmm1, %zmm0, %k0
3805 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
3806 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
3807 ; NoVLX-NEXT: kmovw %k0, %eax
3808 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
3809 ; NoVLX-NEXT: vzeroupper
3810 ; NoVLX-NEXT: retq
3811 entry:
3812 %0 = bitcast <4 x i64> %__a to <4 x i64>
3813 %load = load <4 x i64>, <4 x i64>* %__b
3814 %1 = bitcast <4 x i64> %load to <4 x i64>
3815 %2 = icmp eq <4 x i64> %0, %1
3816 %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
3817 %4 = bitcast <16 x i1> %3 to i16
3818 ret i16 %4
3819 }
3821 define zeroext i16 @test_masked_vpcmpeqq_v4i1_v16i1_mask(i8 zeroext %__u, <4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
3822 ; VLX-LABEL: test_masked_vpcmpeqq_v4i1_v16i1_mask:
3823 ; VLX: # %bb.0: # %entry
3824 ; VLX-NEXT: kmovd %edi, %k1
3825 ; VLX-NEXT: vpcmpeqq %ymm1, %ymm0, %k0 {%k1}
3826 ; VLX-NEXT: kmovd %k0, %eax
3827 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
3828 ; VLX-NEXT: vzeroupper
3829 ; VLX-NEXT: retq
3830 ;
3831 ; NoVLX-LABEL: test_masked_vpcmpeqq_v4i1_v16i1_mask:
3832 ; NoVLX: # %bb.0: # %entry
3833 ; NoVLX-NEXT: # kill: def $ymm1 killed $ymm1 def $zmm1
3834 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
3835 ; NoVLX-NEXT: kmovw %edi, %k1
3836 ; NoVLX-NEXT: vpcmpeqq %zmm1, %zmm0, %k0 {%k1}
3837 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
3838 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
3839 ; NoVLX-NEXT: kmovw %k0, %eax
3840 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
3841 ; NoVLX-NEXT: vzeroupper
3842 ; NoVLX-NEXT: retq
3843 entry:
3844 %0 = bitcast <4 x i64> %__a to <4 x i64>
3845 %1 = bitcast <4 x i64> %__b to <4 x i64>
3846 %2 = icmp eq <4 x i64> %0, %1
3847 %3 = bitcast i8 %__u to <8 x i1>
3848 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
3849 %4 = and <4 x i1> %2, %extract.i
3850 %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
3851 %6 = bitcast <16 x i1> %5 to i16
3852 ret i16 %6
3853 }
3855 define zeroext i16 @test_masked_vpcmpeqq_v4i1_v16i1_mask_mem(i8 zeroext %__u, <4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
3856 ; VLX-LABEL: test_masked_vpcmpeqq_v4i1_v16i1_mask_mem:
3857 ; VLX: # %bb.0: # %entry
3858 ; VLX-NEXT: kmovd %edi, %k1
3859 ; VLX-NEXT: vpcmpeqq (%rsi), %ymm0, %k0 {%k1}
3860 ; VLX-NEXT: kmovd %k0, %eax
3861 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
3862 ; VLX-NEXT: vzeroupper
3863 ; VLX-NEXT: retq
3864 ;
3865 ; NoVLX-LABEL: test_masked_vpcmpeqq_v4i1_v16i1_mask_mem:
3866 ; NoVLX: # %bb.0: # %entry
3867 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
3868 ; NoVLX-NEXT: vmovdqa (%rsi), %ymm1
3869 ; NoVLX-NEXT: kmovw %edi, %k1
3870 ; NoVLX-NEXT: vpcmpeqq %zmm1, %zmm0, %k0 {%k1}
3871 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
3872 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
3873 ; NoVLX-NEXT: kmovw %k0, %eax
3874 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
3875 ; NoVLX-NEXT: vzeroupper
3876 ; NoVLX-NEXT: retq
3877 entry:
3878 %0 = bitcast <4 x i64> %__a to <4 x i64>
3879 %load = load <4 x i64>, <4 x i64>* %__b
3880 %1 = bitcast <4 x i64> %load to <4 x i64>
3881 %2 = icmp eq <4 x i64> %0, %1
3882 %3 = bitcast i8 %__u to <8 x i1>
3883 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
3884 %4 = and <4 x i1> %2, %extract.i
3885 %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
3886 %6 = bitcast <16 x i1> %5 to i16
3887 ret i16 %6
3888 }
3891 define zeroext i16 @test_vpcmpeqq_v4i1_v16i1_mask_mem_b(<4 x i64> %__a, i64* %__b) local_unnamed_addr {
3892 ; VLX-LABEL: test_vpcmpeqq_v4i1_v16i1_mask_mem_b:
3893 ; VLX: # %bb.0: # %entry
3894 ; VLX-NEXT: vpcmpeqq (%rdi){1to4}, %ymm0, %k0
3895 ; VLX-NEXT: kmovd %k0, %eax
3896 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
3897 ; VLX-NEXT: vzeroupper
3898 ; VLX-NEXT: retq
3899 ;
3900 ; NoVLX-LABEL: test_vpcmpeqq_v4i1_v16i1_mask_mem_b:
3901 ; NoVLX: # %bb.0: # %entry
3902 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
3903 ; NoVLX-NEXT: vpcmpeqq (%rdi){1to8}, %zmm0, %k0
3904 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
3905 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
3906 ; NoVLX-NEXT: kmovw %k0, %eax
3907 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
3908 ; NoVLX-NEXT: vzeroupper
3909 ; NoVLX-NEXT: retq
3910 entry:
3911 %0 = bitcast <4 x i64> %__a to <4 x i64>
3912 %load = load i64, i64* %__b
3913 %vec = insertelement <4 x i64> undef, i64 %load, i32 0
3914 %1 = shufflevector <4 x i64> %vec, <4 x i64> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
3915 %2 = icmp eq <4 x i64> %0, %1
3916 %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
3917 %4 = bitcast <16 x i1> %3 to i16
3918 ret i16 %4
3919 }
3921 define zeroext i16 @test_masked_vpcmpeqq_v4i1_v16i1_mask_mem_b(i8 zeroext %__u, <4 x i64> %__a, i64* %__b) local_unnamed_addr {
3922 ; VLX-LABEL: test_masked_vpcmpeqq_v4i1_v16i1_mask_mem_b:
3923 ; VLX: # %bb.0: # %entry
3924 ; VLX-NEXT: kmovd %edi, %k1
3925 ; VLX-NEXT: vpcmpeqq (%rsi){1to4}, %ymm0, %k0 {%k1}
3926 ; VLX-NEXT: kmovd %k0, %eax
3927 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
3928 ; VLX-NEXT: vzeroupper
3929 ; VLX-NEXT: retq
3930 ;
3931 ; NoVLX-LABEL: test_masked_vpcmpeqq_v4i1_v16i1_mask_mem_b:
3932 ; NoVLX: # %bb.0: # %entry
3933 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
3934 ; NoVLX-NEXT: kmovw %edi, %k1
3935 ; NoVLX-NEXT: vpcmpeqq (%rsi){1to8}, %zmm0, %k0 {%k1}
3936 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
3937 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
3938 ; NoVLX-NEXT: kmovw %k0, %eax
3939 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
3940 ; NoVLX-NEXT: vzeroupper
3941 ; NoVLX-NEXT: retq
3942 entry:
3943 %0 = bitcast <4 x i64> %__a to <4 x i64>
3944 %load = load i64, i64* %__b
3945 %vec = insertelement <4 x i64> undef, i64 %load, i32 0
3946 %1 = shufflevector <4 x i64> %vec, <4 x i64> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
3947 %2 = icmp eq <4 x i64> %0, %1
3948 %3 = bitcast i8 %__u to <8 x i1>
3949 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
3950 %4 = and <4 x i1> %extract.i, %2
3951 %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
3952 %6 = bitcast <16 x i1> %5 to i16
3953 ret i16 %6
3954 }
3957 define zeroext i32 @test_vpcmpeqq_v4i1_v32i1_mask(<4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
3958 ; VLX-LABEL: test_vpcmpeqq_v4i1_v32i1_mask:
3959 ; VLX: # %bb.0: # %entry
3960 ; VLX-NEXT: vpcmpeqq %ymm1, %ymm0, %k0
3961 ; VLX-NEXT: kmovd %k0, %eax
3962 ; VLX-NEXT: vzeroupper
3963 ; VLX-NEXT: retq
3964 ;
3965 ; NoVLX-LABEL: test_vpcmpeqq_v4i1_v32i1_mask:
3966 ; NoVLX: # %bb.0: # %entry
3967 ; NoVLX-NEXT: # kill: def $ymm1 killed $ymm1 def $zmm1
3968 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
3969 ; NoVLX-NEXT: vpcmpeqq %zmm1, %zmm0, %k0
3970 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
3971 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
3972 ; NoVLX-NEXT: kmovw %k0, %eax
3973 ; NoVLX-NEXT: vzeroupper
3974 ; NoVLX-NEXT: retq
3975 entry:
3976 %0 = bitcast <4 x i64> %__a to <4 x i64>
3977 %1 = bitcast <4 x i64> %__b to <4 x i64>
3978 %2 = icmp eq <4 x i64> %0, %1
3979 %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
3980 %4 = bitcast <32 x i1> %3 to i32
3981 ret i32 %4
3982 }
3984 define zeroext i32 @test_vpcmpeqq_v4i1_v32i1_mask_mem(<4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
3985 ; VLX-LABEL: test_vpcmpeqq_v4i1_v32i1_mask_mem:
3986 ; VLX: # %bb.0: # %entry
3987 ; VLX-NEXT: vpcmpeqq (%rdi), %ymm0, %k0
3988 ; VLX-NEXT: kmovd %k0, %eax
3989 ; VLX-NEXT: vzeroupper
3990 ; VLX-NEXT: retq
3991 ;
3992 ; NoVLX-LABEL: test_vpcmpeqq_v4i1_v32i1_mask_mem:
3993 ; NoVLX: # %bb.0: # %entry
3994 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
3995 ; NoVLX-NEXT: vmovdqa (%rdi), %ymm1
3996 ; NoVLX-NEXT: vpcmpeqq %zmm1, %zmm0, %k0
3997 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
3998 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
3999 ; NoVLX-NEXT: kmovw %k0, %eax
4000 ; NoVLX-NEXT: vzeroupper
4001 ; NoVLX-NEXT: retq
4002 entry:
4003 %0 = bitcast <4 x i64> %__a to <4 x i64>
4004 %load = load <4 x i64>, <4 x i64>* %__b
4005 %1 = bitcast <4 x i64> %load to <4 x i64>
4006 %2 = icmp eq <4 x i64> %0, %1
4007 %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
4008 %4 = bitcast <32 x i1> %3 to i32
4009 ret i32 %4
4010 }
4012 define zeroext i32 @test_masked_vpcmpeqq_v4i1_v32i1_mask(i8 zeroext %__u, <4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
4013 ; VLX-LABEL: test_masked_vpcmpeqq_v4i1_v32i1_mask:
4014 ; VLX: # %bb.0: # %entry
4015 ; VLX-NEXT: kmovd %edi, %k1
4016 ; VLX-NEXT: vpcmpeqq %ymm1, %ymm0, %k0 {%k1}
4017 ; VLX-NEXT: kmovd %k0, %eax
4018 ; VLX-NEXT: vzeroupper
4019 ; VLX-NEXT: retq
4020 ;
4021 ; NoVLX-LABEL: test_masked_vpcmpeqq_v4i1_v32i1_mask:
4022 ; NoVLX: # %bb.0: # %entry
4023 ; NoVLX-NEXT: # kill: def $ymm1 killed $ymm1 def $zmm1
4024 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
4025 ; NoVLX-NEXT: kmovw %edi, %k1
4026 ; NoVLX-NEXT: vpcmpeqq %zmm1, %zmm0, %k0 {%k1}
4027 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
4028 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
4029 ; NoVLX-NEXT: kmovw %k0, %eax
4030 ; NoVLX-NEXT: vzeroupper
4031 ; NoVLX-NEXT: retq
4032 entry:
4033 %0 = bitcast <4 x i64> %__a to <4 x i64>
4034 %1 = bitcast <4 x i64> %__b to <4 x i64>
4035 %2 = icmp eq <4 x i64> %0, %1
4036 %3 = bitcast i8 %__u to <8 x i1>
4037 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
4038 %4 = and <4 x i1> %2, %extract.i
4039 %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
4040 %6 = bitcast <32 x i1> %5 to i32
4041 ret i32 %6
4042 }
4044 define zeroext i32 @test_masked_vpcmpeqq_v4i1_v32i1_mask_mem(i8 zeroext %__u, <4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
4045 ; VLX-LABEL: test_masked_vpcmpeqq_v4i1_v32i1_mask_mem:
4046 ; VLX: # %bb.0: # %entry
4047 ; VLX-NEXT: kmovd %edi, %k1
4048 ; VLX-NEXT: vpcmpeqq (%rsi), %ymm0, %k0 {%k1}
4049 ; VLX-NEXT: kmovd %k0, %eax
4050 ; VLX-NEXT: vzeroupper
4051 ; VLX-NEXT: retq
4052 ;
4053 ; NoVLX-LABEL: test_masked_vpcmpeqq_v4i1_v32i1_mask_mem:
4054 ; NoVLX: # %bb.0: # %entry
4055 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
4056 ; NoVLX-NEXT: vmovdqa (%rsi), %ymm1
4057 ; NoVLX-NEXT: kmovw %edi, %k1
4058 ; NoVLX-NEXT: vpcmpeqq %zmm1, %zmm0, %k0 {%k1}
4059 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
4060 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
4061 ; NoVLX-NEXT: kmovw %k0, %eax
4062 ; NoVLX-NEXT: vzeroupper
4063 ; NoVLX-NEXT: retq
4064 entry:
4065 %0 = bitcast <4 x i64> %__a to <4 x i64>
4066 %load = load <4 x i64>, <4 x i64>* %__b
4067 %1 = bitcast <4 x i64> %load to <4 x i64>
4068 %2 = icmp eq <4 x i64> %0, %1
4069 %3 = bitcast i8 %__u to <8 x i1>
4070 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
4071 %4 = and <4 x i1> %2, %extract.i
4072 %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
4073 %6 = bitcast <32 x i1> %5 to i32
4074 ret i32 %6
4075 }
4078 define zeroext i32 @test_vpcmpeqq_v4i1_v32i1_mask_mem_b(<4 x i64> %__a, i64* %__b) local_unnamed_addr {
4079 ; VLX-LABEL: test_vpcmpeqq_v4i1_v32i1_mask_mem_b:
4080 ; VLX: # %bb.0: # %entry
4081 ; VLX-NEXT: vpcmpeqq (%rdi){1to4}, %ymm0, %k0
4082 ; VLX-NEXT: kmovd %k0, %eax
4083 ; VLX-NEXT: vzeroupper
4084 ; VLX-NEXT: retq
4085 ;
4086 ; NoVLX-LABEL: test_vpcmpeqq_v4i1_v32i1_mask_mem_b:
4087 ; NoVLX: # %bb.0: # %entry
4088 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
4089 ; NoVLX-NEXT: vpcmpeqq (%rdi){1to8}, %zmm0, %k0
4090 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
4091 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
4092 ; NoVLX-NEXT: kmovw %k0, %eax
4093 ; NoVLX-NEXT: vzeroupper
4094 ; NoVLX-NEXT: retq
4095 entry:
4096 %0 = bitcast <4 x i64> %__a to <4 x i64>
4097 %load = load i64, i64* %__b
4098 %vec = insertelement <4 x i64> undef, i64 %load, i32 0
4099 %1 = shufflevector <4 x i64> %vec, <4 x i64> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
4100 %2 = icmp eq <4 x i64> %0, %1
4101 %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
4102 %4 = bitcast <32 x i1> %3 to i32
4103 ret i32 %4
4104 }
4106 define zeroext i32 @test_masked_vpcmpeqq_v4i1_v32i1_mask_mem_b(i8 zeroext %__u, <4 x i64> %__a, i64* %__b) local_unnamed_addr {
4107 ; VLX-LABEL: test_masked_vpcmpeqq_v4i1_v32i1_mask_mem_b:
4108 ; VLX: # %bb.0: # %entry
4109 ; VLX-NEXT: kmovd %edi, %k1
4110 ; VLX-NEXT: vpcmpeqq (%rsi){1to4}, %ymm0, %k0 {%k1}
4111 ; VLX-NEXT: kmovd %k0, %eax
4112 ; VLX-NEXT: vzeroupper
4113 ; VLX-NEXT: retq
4114 ;
4115 ; NoVLX-LABEL: test_masked_vpcmpeqq_v4i1_v32i1_mask_mem_b:
4116 ; NoVLX: # %bb.0: # %entry
4117 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
4118 ; NoVLX-NEXT: kmovw %edi, %k1
4119 ; NoVLX-NEXT: vpcmpeqq (%rsi){1to8}, %zmm0, %k0 {%k1}
4120 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
4121 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
4122 ; NoVLX-NEXT: kmovw %k0, %eax
4123 ; NoVLX-NEXT: vzeroupper
4124 ; NoVLX-NEXT: retq
4125 entry:
4126 %0 = bitcast <4 x i64> %__a to <4 x i64>
4127 %load = load i64, i64* %__b
4128 %vec = insertelement <4 x i64> undef, i64 %load, i32 0
4129 %1 = shufflevector <4 x i64> %vec, <4 x i64> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
4130 %2 = icmp eq <4 x i64> %0, %1
4131 %3 = bitcast i8 %__u to <8 x i1>
4132 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
4133 %4 = and <4 x i1> %extract.i, %2
4134 %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
4135 %6 = bitcast <32 x i1> %5 to i32
4136 ret i32 %6
4137 }
4140 define zeroext i64 @test_vpcmpeqq_v4i1_v64i1_mask(<4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
4141 ; VLX-LABEL: test_vpcmpeqq_v4i1_v64i1_mask:
4142 ; VLX: # %bb.0: # %entry
4143 ; VLX-NEXT: vpcmpeqq %ymm1, %ymm0, %k0
4144 ; VLX-NEXT: kmovq %k0, %rax
4145 ; VLX-NEXT: vzeroupper
4146 ; VLX-NEXT: retq
4147 ;
4148 ; NoVLX-LABEL: test_vpcmpeqq_v4i1_v64i1_mask:
4149 ; NoVLX: # %bb.0: # %entry
4150 ; NoVLX-NEXT: # kill: def $ymm1 killed $ymm1 def $zmm1
4151 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
4152 ; NoVLX-NEXT: vpcmpeqq %zmm1, %zmm0, %k0
4153 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
4154 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
4155 ; NoVLX-NEXT: kmovw %k0, %eax
4156 ; NoVLX-NEXT: vzeroupper
4157 ; NoVLX-NEXT: retq
4158 entry:
4159 %0 = bitcast <4 x i64> %__a to <4 x i64>
4160 %1 = bitcast <4 x i64> %__b to <4 x i64>
4161 %2 = icmp eq <4 x i64> %0, %1
4162 %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
4163 %4 = bitcast <64 x i1> %3 to i64
4164 ret i64 %4
4165 }
4167 define zeroext i64 @test_vpcmpeqq_v4i1_v64i1_mask_mem(<4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
4168 ; VLX-LABEL: test_vpcmpeqq_v4i1_v64i1_mask_mem:
4169 ; VLX: # %bb.0: # %entry
4170 ; VLX-NEXT: vpcmpeqq (%rdi), %ymm0, %k0
4171 ; VLX-NEXT: kmovq %k0, %rax
4172 ; VLX-NEXT: vzeroupper
4173 ; VLX-NEXT: retq
4174 ;
4175 ; NoVLX-LABEL: test_vpcmpeqq_v4i1_v64i1_mask_mem:
4176 ; NoVLX: # %bb.0: # %entry
4177 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
4178 ; NoVLX-NEXT: vmovdqa (%rdi), %ymm1
4179 ; NoVLX-NEXT: vpcmpeqq %zmm1, %zmm0, %k0
4180 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
4181 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
4182 ; NoVLX-NEXT: kmovw %k0, %eax
4183 ; NoVLX-NEXT: vzeroupper
4184 ; NoVLX-NEXT: retq
4185 entry:
4186 %0 = bitcast <4 x i64> %__a to <4 x i64>
4187 %load = load <4 x i64>, <4 x i64>* %__b
4188 %1 = bitcast <4 x i64> %load to <4 x i64>
4189 %2 = icmp eq <4 x i64> %0, %1
4190 %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
4191 %4 = bitcast <64 x i1> %3 to i64
4192 ret i64 %4
4193 }
4195 define zeroext i64 @test_masked_vpcmpeqq_v4i1_v64i1_mask(i8 zeroext %__u, <4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
4196 ; VLX-LABEL: test_masked_vpcmpeqq_v4i1_v64i1_mask:
4197 ; VLX: # %bb.0: # %entry
4198 ; VLX-NEXT: kmovd %edi, %k1
4199 ; VLX-NEXT: vpcmpeqq %ymm1, %ymm0, %k0 {%k1}
4200 ; VLX-NEXT: kmovq %k0, %rax
4201 ; VLX-NEXT: vzeroupper
4202 ; VLX-NEXT: retq
4203 ;
4204 ; NoVLX-LABEL: test_masked_vpcmpeqq_v4i1_v64i1_mask:
4205 ; NoVLX: # %bb.0: # %entry
4206 ; NoVLX-NEXT: # kill: def $ymm1 killed $ymm1 def $zmm1
4207 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
4208 ; NoVLX-NEXT: kmovw %edi, %k1
4209 ; NoVLX-NEXT: vpcmpeqq %zmm1, %zmm0, %k0 {%k1}
4210 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
4211 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
4212 ; NoVLX-NEXT: kmovw %k0, %eax
4213 ; NoVLX-NEXT: vzeroupper
4214 ; NoVLX-NEXT: retq
4215 entry:
4216 %0 = bitcast <4 x i64> %__a to <4 x i64>
4217 %1 = bitcast <4 x i64> %__b to <4 x i64>
4218 %2 = icmp eq <4 x i64> %0, %1
4219 %3 = bitcast i8 %__u to <8 x i1>
4220 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
4221 %4 = and <4 x i1> %2, %extract.i
4222 %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
4223 %6 = bitcast <64 x i1> %5 to i64
4224 ret i64 %6
4225 }
4227 define zeroext i64 @test_masked_vpcmpeqq_v4i1_v64i1_mask_mem(i8 zeroext %__u, <4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
4228 ; VLX-LABEL: test_masked_vpcmpeqq_v4i1_v64i1_mask_mem:
4229 ; VLX: # %bb.0: # %entry
4230 ; VLX-NEXT: kmovd %edi, %k1
4231 ; VLX-NEXT: vpcmpeqq (%rsi), %ymm0, %k0 {%k1}
4232 ; VLX-NEXT: kmovq %k0, %rax
4233 ; VLX-NEXT: vzeroupper
4234 ; VLX-NEXT: retq
4235 ;
4236 ; NoVLX-LABEL: test_masked_vpcmpeqq_v4i1_v64i1_mask_mem:
4237 ; NoVLX: # %bb.0: # %entry
4238 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
4239 ; NoVLX-NEXT: vmovdqa (%rsi), %ymm1
4240 ; NoVLX-NEXT: kmovw %edi, %k1
4241 ; NoVLX-NEXT: vpcmpeqq %zmm1, %zmm0, %k0 {%k1}
4242 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
4243 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
4244 ; NoVLX-NEXT: kmovw %k0, %eax
4245 ; NoVLX-NEXT: vzeroupper
4246 ; NoVLX-NEXT: retq
4247 entry:
4248 %0 = bitcast <4 x i64> %__a to <4 x i64>
4249 %load = load <4 x i64>, <4 x i64>* %__b
4250 %1 = bitcast <4 x i64> %load to <4 x i64>
4251 %2 = icmp eq <4 x i64> %0, %1
4252 %3 = bitcast i8 %__u to <8 x i1>
4253 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
4254 %4 = and <4 x i1> %2, %extract.i
4255 %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
4256 %6 = bitcast <64 x i1> %5 to i64
4257 ret i64 %6
4258 }
4261 define zeroext i64 @test_vpcmpeqq_v4i1_v64i1_mask_mem_b(<4 x i64> %__a, i64* %__b) local_unnamed_addr {
4262 ; VLX-LABEL: test_vpcmpeqq_v4i1_v64i1_mask_mem_b:
4263 ; VLX: # %bb.0: # %entry
4264 ; VLX-NEXT: vpcmpeqq (%rdi){1to4}, %ymm0, %k0
4265 ; VLX-NEXT: kmovq %k0, %rax
4266 ; VLX-NEXT: vzeroupper
4267 ; VLX-NEXT: retq
4268 ;
4269 ; NoVLX-LABEL: test_vpcmpeqq_v4i1_v64i1_mask_mem_b:
4270 ; NoVLX: # %bb.0: # %entry
4271 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
4272 ; NoVLX-NEXT: vpcmpeqq (%rdi){1to8}, %zmm0, %k0
4273 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
4274 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
4275 ; NoVLX-NEXT: kmovw %k0, %eax
4276 ; NoVLX-NEXT: vzeroupper
4277 ; NoVLX-NEXT: retq
4278 entry:
4279 %0 = bitcast <4 x i64> %__a to <4 x i64>
4280 %load = load i64, i64* %__b
4281 %vec = insertelement <4 x i64> undef, i64 %load, i32 0
4282 %1 = shufflevector <4 x i64> %vec, <4 x i64> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
4283 %2 = icmp eq <4 x i64> %0, %1
4284 %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
4285 %4 = bitcast <64 x i1> %3 to i64
4286 ret i64 %4
4287 }
4289 define zeroext i64 @test_masked_vpcmpeqq_v4i1_v64i1_mask_mem_b(i8 zeroext %__u, <4 x i64> %__a, i64* %__b) local_unnamed_addr {
4290 ; VLX-LABEL: test_masked_vpcmpeqq_v4i1_v64i1_mask_mem_b:
4291 ; VLX: # %bb.0: # %entry
4292 ; VLX-NEXT: kmovd %edi, %k1
4293 ; VLX-NEXT: vpcmpeqq (%rsi){1to4}, %ymm0, %k0 {%k1}
4294 ; VLX-NEXT: kmovq %k0, %rax
4295 ; VLX-NEXT: vzeroupper
4296 ; VLX-NEXT: retq
4297 ;
4298 ; NoVLX-LABEL: test_masked_vpcmpeqq_v4i1_v64i1_mask_mem_b:
4299 ; NoVLX: # %bb.0: # %entry
4300 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
4301 ; NoVLX-NEXT: kmovw %edi, %k1
4302 ; NoVLX-NEXT: vpcmpeqq (%rsi){1to8}, %zmm0, %k0 {%k1}
4303 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
4304 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
4305 ; NoVLX-NEXT: kmovw %k0, %eax
4306 ; NoVLX-NEXT: vzeroupper
4307 ; NoVLX-NEXT: retq
4308 entry:
4309 %0 = bitcast <4 x i64> %__a to <4 x i64>
4310 %load = load i64, i64* %__b
4311 %vec = insertelement <4 x i64> undef, i64 %load, i32 0
4312 %1 = shufflevector <4 x i64> %vec, <4 x i64> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
4313 %2 = icmp eq <4 x i64> %0, %1
4314 %3 = bitcast i8 %__u to <8 x i1>
4315 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
4316 %4 = and <4 x i1> %extract.i, %2
4317 %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
4318 %6 = bitcast <64 x i1> %5 to i64
4319 ret i64 %6
4320 }
4323 define zeroext i16 @test_vpcmpeqq_v8i1_v16i1_mask(<8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
4324 ; VLX-LABEL: test_vpcmpeqq_v8i1_v16i1_mask:
4325 ; VLX: # %bb.0: # %entry
4326 ; VLX-NEXT: vpcmpeqq %zmm1, %zmm0, %k0
4327 ; VLX-NEXT: kmovd %k0, %eax
4328 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
4329 ; VLX-NEXT: vzeroupper
4330 ; VLX-NEXT: retq
4331 ;
4332 ; NoVLX-LABEL: test_vpcmpeqq_v8i1_v16i1_mask:
4333 ; NoVLX: # %bb.0: # %entry
4334 ; NoVLX-NEXT: vpcmpeqq %zmm1, %zmm0, %k0
4335 ; NoVLX-NEXT: kmovw %k0, %eax
4336 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
4337 ; NoVLX-NEXT: vzeroupper
4338 ; NoVLX-NEXT: retq
4339 entry:
4340 %0 = bitcast <8 x i64> %__a to <8 x i64>
4341 %1 = bitcast <8 x i64> %__b to <8 x i64>
4342 %2 = icmp eq <8 x i64> %0, %1
4343 %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
4344 %4 = bitcast <16 x i1> %3 to i16
4345 ret i16 %4
4346 }
4348 define zeroext i16 @test_vpcmpeqq_v8i1_v16i1_mask_mem(<8 x i64> %__a, <8 x i64>* %__b) local_unnamed_addr {
4349 ; VLX-LABEL: test_vpcmpeqq_v8i1_v16i1_mask_mem:
4350 ; VLX: # %bb.0: # %entry
4351 ; VLX-NEXT: vpcmpeqq (%rdi), %zmm0, %k0
4352 ; VLX-NEXT: kmovd %k0, %eax
4353 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
4354 ; VLX-NEXT: vzeroupper
4355 ; VLX-NEXT: retq
4356 ;
4357 ; NoVLX-LABEL: test_vpcmpeqq_v8i1_v16i1_mask_mem:
4358 ; NoVLX: # %bb.0: # %entry
4359 ; NoVLX-NEXT: vpcmpeqq (%rdi), %zmm0, %k0
4360 ; NoVLX-NEXT: kmovw %k0, %eax
4361 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
4362 ; NoVLX-NEXT: vzeroupper
4363 ; NoVLX-NEXT: retq
4364 entry:
4365 %0 = bitcast <8 x i64> %__a to <8 x i64>
4366 %load = load <8 x i64>, <8 x i64>* %__b
4367 %1 = bitcast <8 x i64> %load to <8 x i64>
4368 %2 = icmp eq <8 x i64> %0, %1
4369 %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
4370 %4 = bitcast <16 x i1> %3 to i16
4371 ret i16 %4
4372 }
4374 define zeroext i16 @test_masked_vpcmpeqq_v8i1_v16i1_mask(i8 zeroext %__u, <8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
4375 ; VLX-LABEL: test_masked_vpcmpeqq_v8i1_v16i1_mask:
4376 ; VLX: # %bb.0: # %entry
4377 ; VLX-NEXT: kmovd %edi, %k1
4378 ; VLX-NEXT: vpcmpeqq %zmm1, %zmm0, %k0 {%k1}
4379 ; VLX-NEXT: kmovd %k0, %eax
4380 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
4381 ; VLX-NEXT: vzeroupper
4382 ; VLX-NEXT: retq
4383 ;
4384 ; NoVLX-LABEL: test_masked_vpcmpeqq_v8i1_v16i1_mask:
4385 ; NoVLX: # %bb.0: # %entry
4386 ; NoVLX-NEXT: kmovw %edi, %k1
4387 ; NoVLX-NEXT: vpcmpeqq %zmm1, %zmm0, %k0 {%k1}
4388 ; NoVLX-NEXT: kmovw %k0, %eax
4389 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
4390 ; NoVLX-NEXT: vzeroupper
4391 ; NoVLX-NEXT: retq
4392 entry:
4393 %0 = bitcast <8 x i64> %__a to <8 x i64>
4394 %1 = bitcast <8 x i64> %__b to <8 x i64>
4395 %2 = icmp eq <8 x i64> %0, %1
4396 %3 = bitcast i8 %__u to <8 x i1>
4397 %4 = and <8 x i1> %2, %3
4398 %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
4399 %6 = bitcast <16 x i1> %5 to i16
4400 ret i16 %6
4401 }
4403 define zeroext i16 @test_masked_vpcmpeqq_v8i1_v16i1_mask_mem(i8 zeroext %__u, <8 x i64> %__a, <8 x i64>* %__b) local_unnamed_addr {
4404 ; VLX-LABEL: test_masked_vpcmpeqq_v8i1_v16i1_mask_mem:
4405 ; VLX: # %bb.0: # %entry
4406 ; VLX-NEXT: kmovd %edi, %k1
4407 ; VLX-NEXT: vpcmpeqq (%rsi), %zmm0, %k0 {%k1}
4408 ; VLX-NEXT: kmovd %k0, %eax
4409 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
4410 ; VLX-NEXT: vzeroupper
4411 ; VLX-NEXT: retq
4412 ;
4413 ; NoVLX-LABEL: test_masked_vpcmpeqq_v8i1_v16i1_mask_mem:
4414 ; NoVLX: # %bb.0: # %entry
4415 ; NoVLX-NEXT: kmovw %edi, %k1
4416 ; NoVLX-NEXT: vpcmpeqq (%rsi), %zmm0, %k0 {%k1}
4417 ; NoVLX-NEXT: kmovw %k0, %eax
4418 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
4419 ; NoVLX-NEXT: vzeroupper
4420 ; NoVLX-NEXT: retq
4421 entry:
4422 %0 = bitcast <8 x i64> %__a to <8 x i64>
4423 %load = load <8 x i64>, <8 x i64>* %__b
4424 %1 = bitcast <8 x i64> %load to <8 x i64>
4425 %2 = icmp eq <8 x i64> %0, %1
4426 %3 = bitcast i8 %__u to <8 x i1>
4427 %4 = and <8 x i1> %2, %3
4428 %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
4429 %6 = bitcast <16 x i1> %5 to i16
4430 ret i16 %6
4431 }
4434 define zeroext i16 @test_vpcmpeqq_v8i1_v16i1_mask_mem_b(<8 x i64> %__a, i64* %__b) local_unnamed_addr {
4435 ; VLX-LABEL: test_vpcmpeqq_v8i1_v16i1_mask_mem_b:
4436 ; VLX: # %bb.0: # %entry
4437 ; VLX-NEXT: vpcmpeqq (%rdi){1to8}, %zmm0, %k0
4438 ; VLX-NEXT: kmovd %k0, %eax
4439 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
4440 ; VLX-NEXT: vzeroupper
4441 ; VLX-NEXT: retq
4442 ;
4443 ; NoVLX-LABEL: test_vpcmpeqq_v8i1_v16i1_mask_mem_b:
4444 ; NoVLX: # %bb.0: # %entry
4445 ; NoVLX-NEXT: vpcmpeqq (%rdi){1to8}, %zmm0, %k0
4446 ; NoVLX-NEXT: kmovw %k0, %eax
4447 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
4448 ; NoVLX-NEXT: vzeroupper
4449 ; NoVLX-NEXT: retq
4450 entry:
4451 %0 = bitcast <8 x i64> %__a to <8 x i64>
4452 %load = load i64, i64* %__b
4453 %vec = insertelement <8 x i64> undef, i64 %load, i32 0
4454 %1 = shufflevector <8 x i64> %vec, <8 x i64> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
4455 %2 = icmp eq <8 x i64> %0, %1
4456 %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
4457 %4 = bitcast <16 x i1> %3 to i16
4458 ret i16 %4
4459 }
4461 define zeroext i16 @test_masked_vpcmpeqq_v8i1_v16i1_mask_mem_b(i8 zeroext %__u, <8 x i64> %__a, i64* %__b) local_unnamed_addr {
4462 ; VLX-LABEL: test_masked_vpcmpeqq_v8i1_v16i1_mask_mem_b:
4463 ; VLX: # %bb.0: # %entry
4464 ; VLX-NEXT: kmovd %edi, %k1
4465 ; VLX-NEXT: vpcmpeqq (%rsi){1to8}, %zmm0, %k0 {%k1}
4466 ; VLX-NEXT: kmovd %k0, %eax
4467 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
4468 ; VLX-NEXT: vzeroupper
4469 ; VLX-NEXT: retq
4470 ;
4471 ; NoVLX-LABEL: test_masked_vpcmpeqq_v8i1_v16i1_mask_mem_b:
4472 ; NoVLX: # %bb.0: # %entry
4473 ; NoVLX-NEXT: kmovw %edi, %k1
4474 ; NoVLX-NEXT: vpcmpeqq (%rsi){1to8}, %zmm0, %k0 {%k1}
4475 ; NoVLX-NEXT: kmovw %k0, %eax
4476 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
4477 ; NoVLX-NEXT: vzeroupper
4478 ; NoVLX-NEXT: retq
4479 entry:
4480 %0 = bitcast <8 x i64> %__a to <8 x i64>
4481 %load = load i64, i64* %__b
4482 %vec = insertelement <8 x i64> undef, i64 %load, i32 0
4483 %1 = shufflevector <8 x i64> %vec, <8 x i64> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
4484 %2 = icmp eq <8 x i64> %0, %1
4485 %3 = bitcast i8 %__u to <8 x i1>
4486 %4 = and <8 x i1> %3, %2
4487 %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
4488 %6 = bitcast <16 x i1> %5 to i16
4489 ret i16 %6
4490 }
4493 define zeroext i32 @test_vpcmpeqq_v8i1_v32i1_mask(<8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
4494 ; VLX-LABEL: test_vpcmpeqq_v8i1_v32i1_mask:
4495 ; VLX: # %bb.0: # %entry
4496 ; VLX-NEXT: vpcmpeqq %zmm1, %zmm0, %k0
4497 ; VLX-NEXT: kmovd %k0, %eax
4498 ; VLX-NEXT: vzeroupper
4499 ; VLX-NEXT: retq
4500 ;
4501 ; NoVLX-LABEL: test_vpcmpeqq_v8i1_v32i1_mask:
4502 ; NoVLX: # %bb.0: # %entry
4503 ; NoVLX-NEXT: vpcmpeqq %zmm1, %zmm0, %k0
4504 ; NoVLX-NEXT: kmovw %k0, %eax
4505 ; NoVLX-NEXT: vzeroupper
4506 ; NoVLX-NEXT: retq
4507 entry:
4508 %0 = bitcast <8 x i64> %__a to <8 x i64>
4509 %1 = bitcast <8 x i64> %__b to <8 x i64>
4510 %2 = icmp eq <8 x i64> %0, %1
4511 %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
4512 %4 = bitcast <32 x i1> %3 to i32
4513 ret i32 %4
4514 }
4516 define zeroext i32 @test_vpcmpeqq_v8i1_v32i1_mask_mem(<8 x i64> %__a, <8 x i64>* %__b) local_unnamed_addr {
4517 ; VLX-LABEL: test_vpcmpeqq_v8i1_v32i1_mask_mem:
4518 ; VLX: # %bb.0: # %entry
4519 ; VLX-NEXT: vpcmpeqq (%rdi), %zmm0, %k0
4520 ; VLX-NEXT: kmovd %k0, %eax
4521 ; VLX-NEXT: vzeroupper
4522 ; VLX-NEXT: retq
4523 ;
4524 ; NoVLX-LABEL: test_vpcmpeqq_v8i1_v32i1_mask_mem:
4525 ; NoVLX: # %bb.0: # %entry
4526 ; NoVLX-NEXT: vpcmpeqq (%rdi), %zmm0, %k0
4527 ; NoVLX-NEXT: kmovw %k0, %eax
4528 ; NoVLX-NEXT: vzeroupper
4529 ; NoVLX-NEXT: retq
4530 entry:
4531 %0 = bitcast <8 x i64> %__a to <8 x i64>
4532 %load = load <8 x i64>, <8 x i64>* %__b
4533 %1 = bitcast <8 x i64> %load to <8 x i64>
4534 %2 = icmp eq <8 x i64> %0, %1
4535 %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
4536 %4 = bitcast <32 x i1> %3 to i32
4537 ret i32 %4
4538 }
4540 define zeroext i32 @test_masked_vpcmpeqq_v8i1_v32i1_mask(i8 zeroext %__u, <8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
4541 ; VLX-LABEL: test_masked_vpcmpeqq_v8i1_v32i1_mask:
4542 ; VLX: # %bb.0: # %entry
4543 ; VLX-NEXT: kmovd %edi, %k1
4544 ; VLX-NEXT: vpcmpeqq %zmm1, %zmm0, %k0 {%k1}
4545 ; VLX-NEXT: kmovd %k0, %eax
4546 ; VLX-NEXT: vzeroupper
4547 ; VLX-NEXT: retq
4548 ;
4549 ; NoVLX-LABEL: test_masked_vpcmpeqq_v8i1_v32i1_mask:
4550 ; NoVLX: # %bb.0: # %entry
4551 ; NoVLX-NEXT: kmovw %edi, %k1
4552 ; NoVLX-NEXT: vpcmpeqq %zmm1, %zmm0, %k0 {%k1}
4553 ; NoVLX-NEXT: kmovw %k0, %eax
4554 ; NoVLX-NEXT: vzeroupper
4555 ; NoVLX-NEXT: retq
4556 entry:
4557 %0 = bitcast <8 x i64> %__a to <8 x i64>
4558 %1 = bitcast <8 x i64> %__b to <8 x i64>
4559 %2 = icmp eq <8 x i64> %0, %1
4560 %3 = bitcast i8 %__u to <8 x i1>
4561 %4 = and <8 x i1> %2, %3
4562 %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
4563 %6 = bitcast <32 x i1> %5 to i32
4564 ret i32 %6
4565 }
4567 define zeroext i32 @test_masked_vpcmpeqq_v8i1_v32i1_mask_mem(i8 zeroext %__u, <8 x i64> %__a, <8 x i64>* %__b) local_unnamed_addr {
4568 ; VLX-LABEL: test_masked_vpcmpeqq_v8i1_v32i1_mask_mem:
4569 ; VLX: # %bb.0: # %entry
4570 ; VLX-NEXT: kmovd %edi, %k1
4571 ; VLX-NEXT: vpcmpeqq (%rsi), %zmm0, %k0 {%k1}
4572 ; VLX-NEXT: kmovd %k0, %eax
4573 ; VLX-NEXT: vzeroupper
4574 ; VLX-NEXT: retq
4575 ;
4576 ; NoVLX-LABEL: test_masked_vpcmpeqq_v8i1_v32i1_mask_mem:
4577 ; NoVLX: # %bb.0: # %entry
4578 ; NoVLX-NEXT: kmovw %edi, %k1
4579 ; NoVLX-NEXT: vpcmpeqq (%rsi), %zmm0, %k0 {%k1}
4580 ; NoVLX-NEXT: kmovw %k0, %eax
4581 ; NoVLX-NEXT: vzeroupper
4582 ; NoVLX-NEXT: retq
4583 entry:
4584 %0 = bitcast <8 x i64> %__a to <8 x i64>
4585 %load = load <8 x i64>, <8 x i64>* %__b
4586 %1 = bitcast <8 x i64> %load to <8 x i64>
4587 %2 = icmp eq <8 x i64> %0, %1
4588 %3 = bitcast i8 %__u to <8 x i1>
4589 %4 = and <8 x i1> %2, %3
4590 %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
4591 %6 = bitcast <32 x i1> %5 to i32
4592 ret i32 %6
4593 }
4596 define zeroext i32 @test_vpcmpeqq_v8i1_v32i1_mask_mem_b(<8 x i64> %__a, i64* %__b) local_unnamed_addr {
4597 ; VLX-LABEL: test_vpcmpeqq_v8i1_v32i1_mask_mem_b:
4598 ; VLX: # %bb.0: # %entry
4599 ; VLX-NEXT: vpcmpeqq (%rdi){1to8}, %zmm0, %k0
4600 ; VLX-NEXT: kmovd %k0, %eax
4601 ; VLX-NEXT: vzeroupper
4602 ; VLX-NEXT: retq
4603 ;
4604 ; NoVLX-LABEL: test_vpcmpeqq_v8i1_v32i1_mask_mem_b:
4605 ; NoVLX: # %bb.0: # %entry
4606 ; NoVLX-NEXT: vpcmpeqq (%rdi){1to8}, %zmm0, %k0
4607 ; NoVLX-NEXT: kmovw %k0, %eax
4608 ; NoVLX-NEXT: vzeroupper
4609 ; NoVLX-NEXT: retq
4610 entry:
4611 %0 = bitcast <8 x i64> %__a to <8 x i64>
4612 %load = load i64, i64* %__b
4613 %vec = insertelement <8 x i64> undef, i64 %load, i32 0
4614 %1 = shufflevector <8 x i64> %vec, <8 x i64> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
4615 %2 = icmp eq <8 x i64> %0, %1
4616 %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
4617 %4 = bitcast <32 x i1> %3 to i32
4618 ret i32 %4
4619 }
4621 define zeroext i32 @test_masked_vpcmpeqq_v8i1_v32i1_mask_mem_b(i8 zeroext %__u, <8 x i64> %__a, i64* %__b) local_unnamed_addr {
4622 ; VLX-LABEL: test_masked_vpcmpeqq_v8i1_v32i1_mask_mem_b:
4623 ; VLX: # %bb.0: # %entry
4624 ; VLX-NEXT: kmovd %edi, %k1
4625 ; VLX-NEXT: vpcmpeqq (%rsi){1to8}, %zmm0, %k0 {%k1}
4626 ; VLX-NEXT: kmovd %k0, %eax
4627 ; VLX-NEXT: vzeroupper
4628 ; VLX-NEXT: retq
4629 ;
4630 ; NoVLX-LABEL: test_masked_vpcmpeqq_v8i1_v32i1_mask_mem_b:
4631 ; NoVLX: # %bb.0: # %entry
4632 ; NoVLX-NEXT: kmovw %edi, %k1
4633 ; NoVLX-NEXT: vpcmpeqq (%rsi){1to8}, %zmm0, %k0 {%k1}
4634 ; NoVLX-NEXT: kmovw %k0, %eax
4635 ; NoVLX-NEXT: vzeroupper
4636 ; NoVLX-NEXT: retq
4637 entry:
4638 %0 = bitcast <8 x i64> %__a to <8 x i64>
4639 %load = load i64, i64* %__b
4640 %vec = insertelement <8 x i64> undef, i64 %load, i32 0
4641 %1 = shufflevector <8 x i64> %vec, <8 x i64> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
4642 %2 = icmp eq <8 x i64> %0, %1
4643 %3 = bitcast i8 %__u to <8 x i1>
4644 %4 = and <8 x i1> %3, %2
4645 %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
4646 %6 = bitcast <32 x i1> %5 to i32
4647 ret i32 %6
4648 }
4651 define zeroext i64 @test_vpcmpeqq_v8i1_v64i1_mask(<8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
4652 ; VLX-LABEL: test_vpcmpeqq_v8i1_v64i1_mask:
4653 ; VLX: # %bb.0: # %entry
4654 ; VLX-NEXT: vpcmpeqq %zmm1, %zmm0, %k0
4655 ; VLX-NEXT: kmovq %k0, %rax
4656 ; VLX-NEXT: vzeroupper
4657 ; VLX-NEXT: retq
4658 ;
4659 ; NoVLX-LABEL: test_vpcmpeqq_v8i1_v64i1_mask:
4660 ; NoVLX: # %bb.0: # %entry
4661 ; NoVLX-NEXT: vpcmpeqq %zmm1, %zmm0, %k0
4662 ; NoVLX-NEXT: kmovw %k0, %eax
4663 ; NoVLX-NEXT: vzeroupper
4664 ; NoVLX-NEXT: retq
4665 entry:
4666 %0 = bitcast <8 x i64> %__a to <8 x i64>
4667 %1 = bitcast <8 x i64> %__b to <8 x i64>
4668 %2 = icmp eq <8 x i64> %0, %1
4669 %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
4670 %4 = bitcast <64 x i1> %3 to i64
4671 ret i64 %4
4672 }
4674 define zeroext i64 @test_vpcmpeqq_v8i1_v64i1_mask_mem(<8 x i64> %__a, <8 x i64>* %__b) local_unnamed_addr {
4675 ; VLX-LABEL: test_vpcmpeqq_v8i1_v64i1_mask_mem:
4676 ; VLX: # %bb.0: # %entry
4677 ; VLX-NEXT: vpcmpeqq (%rdi), %zmm0, %k0
4678 ; VLX-NEXT: kmovq %k0, %rax
4679 ; VLX-NEXT: vzeroupper
4680 ; VLX-NEXT: retq
4681 ;
4682 ; NoVLX-LABEL: test_vpcmpeqq_v8i1_v64i1_mask_mem:
4683 ; NoVLX: # %bb.0: # %entry
4684 ; NoVLX-NEXT: vpcmpeqq (%rdi), %zmm0, %k0
4685 ; NoVLX-NEXT: kmovw %k0, %eax
4686 ; NoVLX-NEXT: vzeroupper
4687 ; NoVLX-NEXT: retq
4688 entry:
4689 %0 = bitcast <8 x i64> %__a to <8 x i64>
4690 %load = load <8 x i64>, <8 x i64>* %__b
4691 %1 = bitcast <8 x i64> %load to <8 x i64>
4692 %2 = icmp eq <8 x i64> %0, %1
4693 %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
4694 %4 = bitcast <64 x i1> %3 to i64
4695 ret i64 %4
4696 }
4698 define zeroext i64 @test_masked_vpcmpeqq_v8i1_v64i1_mask(i8 zeroext %__u, <8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
4699 ; VLX-LABEL: test_masked_vpcmpeqq_v8i1_v64i1_mask:
4700 ; VLX: # %bb.0: # %entry
4701 ; VLX-NEXT: kmovd %edi, %k1
4702 ; VLX-NEXT: vpcmpeqq %zmm1, %zmm0, %k0 {%k1}
4703 ; VLX-NEXT: kmovq %k0, %rax
4704 ; VLX-NEXT: vzeroupper
4705 ; VLX-NEXT: retq
4706 ;
4707 ; NoVLX-LABEL: test_masked_vpcmpeqq_v8i1_v64i1_mask:
4708 ; NoVLX: # %bb.0: # %entry
4709 ; NoVLX-NEXT: kmovw %edi, %k1
4710 ; NoVLX-NEXT: vpcmpeqq %zmm1, %zmm0, %k0 {%k1}
4711 ; NoVLX-NEXT: kmovw %k0, %eax
4712 ; NoVLX-NEXT: vzeroupper
4713 ; NoVLX-NEXT: retq
4714 entry:
4715 %0 = bitcast <8 x i64> %__a to <8 x i64>
4716 %1 = bitcast <8 x i64> %__b to <8 x i64>
4717 %2 = icmp eq <8 x i64> %0, %1
4718 %3 = bitcast i8 %__u to <8 x i1>
4719 %4 = and <8 x i1> %2, %3
4720 %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
4721 %6 = bitcast <64 x i1> %5 to i64
4722 ret i64 %6
4723 }
4725 define zeroext i64 @test_masked_vpcmpeqq_v8i1_v64i1_mask_mem(i8 zeroext %__u, <8 x i64> %__a, <8 x i64>* %__b) local_unnamed_addr {
4726 ; VLX-LABEL: test_masked_vpcmpeqq_v8i1_v64i1_mask_mem:
4727 ; VLX: # %bb.0: # %entry
4728 ; VLX-NEXT: kmovd %edi, %k1
4729 ; VLX-NEXT: vpcmpeqq (%rsi), %zmm0, %k0 {%k1}
4730 ; VLX-NEXT: kmovq %k0, %rax
4731 ; VLX-NEXT: vzeroupper
4732 ; VLX-NEXT: retq
4733 ;
4734 ; NoVLX-LABEL: test_masked_vpcmpeqq_v8i1_v64i1_mask_mem:
4735 ; NoVLX: # %bb.0: # %entry
4736 ; NoVLX-NEXT: kmovw %edi, %k1
4737 ; NoVLX-NEXT: vpcmpeqq (%rsi), %zmm0, %k0 {%k1}
4738 ; NoVLX-NEXT: kmovw %k0, %eax
4739 ; NoVLX-NEXT: vzeroupper
4740 ; NoVLX-NEXT: retq
4741 entry:
4742 %0 = bitcast <8 x i64> %__a to <8 x i64>
4743 %load = load <8 x i64>, <8 x i64>* %__b
4744 %1 = bitcast <8 x i64> %load to <8 x i64>
4745 %2 = icmp eq <8 x i64> %0, %1
4746 %3 = bitcast i8 %__u to <8 x i1>
4747 %4 = and <8 x i1> %2, %3
4748 %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
4749 %6 = bitcast <64 x i1> %5 to i64
4750 ret i64 %6
4751 }
4754 define zeroext i64 @test_vpcmpeqq_v8i1_v64i1_mask_mem_b(<8 x i64> %__a, i64* %__b) local_unnamed_addr {
4755 ; VLX-LABEL: test_vpcmpeqq_v8i1_v64i1_mask_mem_b:
4756 ; VLX: # %bb.0: # %entry
4757 ; VLX-NEXT: vpcmpeqq (%rdi){1to8}, %zmm0, %k0
4758 ; VLX-NEXT: kmovq %k0, %rax
4759 ; VLX-NEXT: vzeroupper
4760 ; VLX-NEXT: retq
4761 ;
4762 ; NoVLX-LABEL: test_vpcmpeqq_v8i1_v64i1_mask_mem_b:
4763 ; NoVLX: # %bb.0: # %entry
4764 ; NoVLX-NEXT: vpcmpeqq (%rdi){1to8}, %zmm0, %k0
4765 ; NoVLX-NEXT: kmovw %k0, %eax
4766 ; NoVLX-NEXT: vzeroupper
4767 ; NoVLX-NEXT: retq
4768 entry:
4769 %0 = bitcast <8 x i64> %__a to <8 x i64>
4770 %load = load i64, i64* %__b
4771 %vec = insertelement <8 x i64> undef, i64 %load, i32 0
4772 %1 = shufflevector <8 x i64> %vec, <8 x i64> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
4773 %2 = icmp eq <8 x i64> %0, %1
4774 %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
4775 %4 = bitcast <64 x i1> %3 to i64
4776 ret i64 %4
4777 }
4779 define zeroext i64 @test_masked_vpcmpeqq_v8i1_v64i1_mask_mem_b(i8 zeroext %__u, <8 x i64> %__a, i64* %__b) local_unnamed_addr {
4780 ; VLX-LABEL: test_masked_vpcmpeqq_v8i1_v64i1_mask_mem_b:
4781 ; VLX: # %bb.0: # %entry
4782 ; VLX-NEXT: kmovd %edi, %k1
4783 ; VLX-NEXT: vpcmpeqq (%rsi){1to8}, %zmm0, %k0 {%k1}
4784 ; VLX-NEXT: kmovq %k0, %rax
4785 ; VLX-NEXT: vzeroupper
4786 ; VLX-NEXT: retq
4787 ;
4788 ; NoVLX-LABEL: test_masked_vpcmpeqq_v8i1_v64i1_mask_mem_b:
4789 ; NoVLX: # %bb.0: # %entry
4790 ; NoVLX-NEXT: kmovw %edi, %k1
4791 ; NoVLX-NEXT: vpcmpeqq (%rsi){1to8}, %zmm0, %k0 {%k1}
4792 ; NoVLX-NEXT: kmovw %k0, %eax
4793 ; NoVLX-NEXT: vzeroupper
4794 ; NoVLX-NEXT: retq
4795 entry:
4796 %0 = bitcast <8 x i64> %__a to <8 x i64>
4797 %load = load i64, i64* %__b
4798 %vec = insertelement <8 x i64> undef, i64 %load, i32 0
4799 %1 = shufflevector <8 x i64> %vec, <8 x i64> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
4800 %2 = icmp eq <8 x i64> %0, %1
4801 %3 = bitcast i8 %__u to <8 x i1>
4802 %4 = and <8 x i1> %3, %2
4803 %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
4804 %6 = bitcast <64 x i1> %5 to i64
4805 ret i64 %6
4806 }
4809 define zeroext i32 @test_vpcmpsgtb_v16i1_v32i1_mask(<2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
4810 ; VLX-LABEL: test_vpcmpsgtb_v16i1_v32i1_mask:
4811 ; VLX: # %bb.0: # %entry
4812 ; VLX-NEXT: vpcmpgtb %xmm1, %xmm0, %k0
4813 ; VLX-NEXT: kmovd %k0, %eax
4814 ; VLX-NEXT: retq
4815 ;
4816 ; NoVLX-LABEL: test_vpcmpsgtb_v16i1_v32i1_mask:
4817 ; NoVLX: # %bb.0: # %entry
4818 ; NoVLX-NEXT: vpcmpgtb %xmm1, %xmm0, %xmm0
4819 ; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0
4820 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
4821 ; NoVLX-NEXT: kmovw %k0, %eax
4822 ; NoVLX-NEXT: vzeroupper
4823 ; NoVLX-NEXT: retq
4824 entry:
4825 %0 = bitcast <2 x i64> %__a to <16 x i8>
4826 %1 = bitcast <2 x i64> %__b to <16 x i8>
4827 %2 = icmp sgt <16 x i8> %0, %1
4828 %3 = shufflevector <16 x i1> %2, <16 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>
4829 %4 = bitcast <32 x i1> %3 to i32
4830 ret i32 %4
4831 }
4833 define zeroext i32 @test_vpcmpsgtb_v16i1_v32i1_mask_mem(<2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
4834 ; VLX-LABEL: test_vpcmpsgtb_v16i1_v32i1_mask_mem:
4835 ; VLX: # %bb.0: # %entry
4836 ; VLX-NEXT: vpcmpgtb (%rdi), %xmm0, %k0
4837 ; VLX-NEXT: kmovd %k0, %eax
4838 ; VLX-NEXT: retq
4839 ;
4840 ; NoVLX-LABEL: test_vpcmpsgtb_v16i1_v32i1_mask_mem:
4841 ; NoVLX: # %bb.0: # %entry
4842 ; NoVLX-NEXT: vpcmpgtb (%rdi), %xmm0, %xmm0
4843 ; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0
4844 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
4845 ; NoVLX-NEXT: kmovw %k0, %eax
4846 ; NoVLX-NEXT: vzeroupper
4847 ; NoVLX-NEXT: retq
4848 entry:
4849 %0 = bitcast <2 x i64> %__a to <16 x i8>
4850 %load = load <2 x i64>, <2 x i64>* %__b
4851 %1 = bitcast <2 x i64> %load to <16 x i8>
4852 %2 = icmp sgt <16 x i8> %0, %1
4853 %3 = shufflevector <16 x i1> %2, <16 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>
4854 %4 = bitcast <32 x i1> %3 to i32
4855 ret i32 %4
4856 }
4858 define zeroext i32 @test_masked_vpcmpsgtb_v16i1_v32i1_mask(i16 zeroext %__u, <2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
4859 ; VLX-LABEL: test_masked_vpcmpsgtb_v16i1_v32i1_mask:
4860 ; VLX: # %bb.0: # %entry
4861 ; VLX-NEXT: kmovd %edi, %k1
4862 ; VLX-NEXT: vpcmpgtb %xmm1, %xmm0, %k0 {%k1}
4863 ; VLX-NEXT: kmovd %k0, %eax
4864 ; VLX-NEXT: retq
4865 ;
4866 ; NoVLX-LABEL: test_masked_vpcmpsgtb_v16i1_v32i1_mask:
4867 ; NoVLX: # %bb.0: # %entry
4868 ; NoVLX-NEXT: vpcmpgtb %xmm1, %xmm0, %xmm0
4869 ; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0
4870 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
4871 ; NoVLX-NEXT: kmovw %k0, %eax
4872 ; NoVLX-NEXT: andl %edi, %eax
4873 ; NoVLX-NEXT: vzeroupper
4874 ; NoVLX-NEXT: retq
4875 entry:
4876 %0 = bitcast <2 x i64> %__a to <16 x i8>
4877 %1 = bitcast <2 x i64> %__b to <16 x i8>
4878 %2 = icmp sgt <16 x i8> %0, %1
4879 %3 = bitcast i16 %__u to <16 x i1>
4880 %4 = and <16 x i1> %2, %3
4881 %5 = shufflevector <16 x i1> %4, <16 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>
4882 %6 = bitcast <32 x i1> %5 to i32
4883 ret i32 %6
4884 }
4886 define zeroext i32 @test_masked_vpcmpsgtb_v16i1_v32i1_mask_mem(i16 zeroext %__u, <2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
4887 ; VLX-LABEL: test_masked_vpcmpsgtb_v16i1_v32i1_mask_mem:
4888 ; VLX: # %bb.0: # %entry
4889 ; VLX-NEXT: kmovd %edi, %k1
4890 ; VLX-NEXT: vpcmpgtb (%rsi), %xmm0, %k0 {%k1}
4891 ; VLX-NEXT: kmovd %k0, %eax
4892 ; VLX-NEXT: retq
4893 ;
4894 ; NoVLX-LABEL: test_masked_vpcmpsgtb_v16i1_v32i1_mask_mem:
4895 ; NoVLX: # %bb.0: # %entry
4896 ; NoVLX-NEXT: vpcmpgtb (%rsi), %xmm0, %xmm0
4897 ; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0
4898 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
4899 ; NoVLX-NEXT: kmovw %k0, %eax
4900 ; NoVLX-NEXT: andl %edi, %eax
4901 ; NoVLX-NEXT: vzeroupper
4902 ; NoVLX-NEXT: retq
4903 entry:
4904 %0 = bitcast <2 x i64> %__a to <16 x i8>
4905 %load = load <2 x i64>, <2 x i64>* %__b
4906 %1 = bitcast <2 x i64> %load to <16 x i8>
4907 %2 = icmp sgt <16 x i8> %0, %1
4908 %3 = bitcast i16 %__u to <16 x i1>
4909 %4 = and <16 x i1> %2, %3
4910 %5 = shufflevector <16 x i1> %4, <16 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>
4911 %6 = bitcast <32 x i1> %5 to i32
4912 ret i32 %6
4913 }
4916 define zeroext i64 @test_vpcmpsgtb_v16i1_v64i1_mask(<2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
4917 ; VLX-LABEL: test_vpcmpsgtb_v16i1_v64i1_mask:
4918 ; VLX: # %bb.0: # %entry
4919 ; VLX-NEXT: vpcmpgtb %xmm1, %xmm0, %k0
4920 ; VLX-NEXT: kmovq %k0, %rax
4921 ; VLX-NEXT: retq
4922 ;
4923 ; NoVLX-LABEL: test_vpcmpsgtb_v16i1_v64i1_mask:
4924 ; NoVLX: # %bb.0: # %entry
4925 ; NoVLX-NEXT: vpcmpgtb %xmm1, %xmm0, %xmm0
4926 ; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0
4927 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
4928 ; NoVLX-NEXT: kmovw %k0, %eax
4929 ; NoVLX-NEXT: vzeroupper
4930 ; NoVLX-NEXT: retq
4931 entry:
4932 %0 = bitcast <2 x i64> %__a to <16 x i8>
4933 %1 = bitcast <2 x i64> %__b to <16 x i8>
4934 %2 = icmp sgt <16 x i8> %0, %1
4935 %3 = shufflevector <16 x i1> %2, <16 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>
4936 %4 = bitcast <64 x i1> %3 to i64
4937 ret i64 %4
4938 }
4940 define zeroext i64 @test_vpcmpsgtb_v16i1_v64i1_mask_mem(<2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
4941 ; VLX-LABEL: test_vpcmpsgtb_v16i1_v64i1_mask_mem:
4942 ; VLX: # %bb.0: # %entry
4943 ; VLX-NEXT: vpcmpgtb (%rdi), %xmm0, %k0
4944 ; VLX-NEXT: kmovq %k0, %rax
4945 ; VLX-NEXT: retq
4946 ;
4947 ; NoVLX-LABEL: test_vpcmpsgtb_v16i1_v64i1_mask_mem:
4948 ; NoVLX: # %bb.0: # %entry
4949 ; NoVLX-NEXT: vpcmpgtb (%rdi), %xmm0, %xmm0
4950 ; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0
4951 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
4952 ; NoVLX-NEXT: kmovw %k0, %eax
4953 ; NoVLX-NEXT: vzeroupper
4954 ; NoVLX-NEXT: retq
4955 entry:
4956 %0 = bitcast <2 x i64> %__a to <16 x i8>
4957 %load = load <2 x i64>, <2 x i64>* %__b
4958 %1 = bitcast <2 x i64> %load to <16 x i8>
4959 %2 = icmp sgt <16 x i8> %0, %1
4960 %3 = shufflevector <16 x i1> %2, <16 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>
4961 %4 = bitcast <64 x i1> %3 to i64
4962 ret i64 %4
4963 }
4965 define zeroext i64 @test_masked_vpcmpsgtb_v16i1_v64i1_mask(i16 zeroext %__u, <2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
4966 ; VLX-LABEL: test_masked_vpcmpsgtb_v16i1_v64i1_mask:
4967 ; VLX: # %bb.0: # %entry
4968 ; VLX-NEXT: kmovd %edi, %k1
4969 ; VLX-NEXT: vpcmpgtb %xmm1, %xmm0, %k0 {%k1}
4970 ; VLX-NEXT: kmovq %k0, %rax
4971 ; VLX-NEXT: retq
4972 ;
4973 ; NoVLX-LABEL: test_masked_vpcmpsgtb_v16i1_v64i1_mask:
4974 ; NoVLX: # %bb.0: # %entry
4975 ; NoVLX-NEXT: vpcmpgtb %xmm1, %xmm0, %xmm0
4976 ; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0
4977 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
4978 ; NoVLX-NEXT: kmovw %k0, %eax
4979 ; NoVLX-NEXT: andl %edi, %eax
4980 ; NoVLX-NEXT: vzeroupper
4981 ; NoVLX-NEXT: retq
4982 entry:
4983 %0 = bitcast <2 x i64> %__a to <16 x i8>
4984 %1 = bitcast <2 x i64> %__b to <16 x i8>
4985 %2 = icmp sgt <16 x i8> %0, %1
4986 %3 = bitcast i16 %__u to <16 x i1>
4987 %4 = and <16 x i1> %2, %3
4988 %5 = shufflevector <16 x i1> %4, <16 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>
4989 %6 = bitcast <64 x i1> %5 to i64
4990 ret i64 %6
4991 }
4993 define zeroext i64 @test_masked_vpcmpsgtb_v16i1_v64i1_mask_mem(i16 zeroext %__u, <2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
4994 ; VLX-LABEL: test_masked_vpcmpsgtb_v16i1_v64i1_mask_mem:
4995 ; VLX: # %bb.0: # %entry
4996 ; VLX-NEXT: kmovd %edi, %k1
4997 ; VLX-NEXT: vpcmpgtb (%rsi), %xmm0, %k0 {%k1}
4998 ; VLX-NEXT: kmovq %k0, %rax
4999 ; VLX-NEXT: retq
5000 ;
5001 ; NoVLX-LABEL: test_masked_vpcmpsgtb_v16i1_v64i1_mask_mem:
5002 ; NoVLX: # %bb.0: # %entry
5003 ; NoVLX-NEXT: vpcmpgtb (%rsi), %xmm0, %xmm0
5004 ; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0
5005 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
5006 ; NoVLX-NEXT: kmovw %k0, %eax
5007 ; NoVLX-NEXT: andl %edi, %eax
5008 ; NoVLX-NEXT: vzeroupper
5009 ; NoVLX-NEXT: retq
5010 entry:
5011 %0 = bitcast <2 x i64> %__a to <16 x i8>
5012 %load = load <2 x i64>, <2 x i64>* %__b
5013 %1 = bitcast <2 x i64> %load to <16 x i8>
5014 %2 = icmp sgt <16 x i8> %0, %1
5015 %3 = bitcast i16 %__u to <16 x i1>
5016 %4 = and <16 x i1> %2, %3
5017 %5 = shufflevector <16 x i1> %4, <16 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>
5018 %6 = bitcast <64 x i1> %5 to i64
5019 ret i64 %6
5020 }
5023 define zeroext i64 @test_vpcmpsgtb_v32i1_v64i1_mask(<4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
5024 ; VLX-LABEL: test_vpcmpsgtb_v32i1_v64i1_mask:
5025 ; VLX: # %bb.0: # %entry
5026 ; VLX-NEXT: vpcmpgtb %ymm1, %ymm0, %k0
5027 ; VLX-NEXT: kmovq %k0, %rax
5028 ; VLX-NEXT: vzeroupper
5029 ; VLX-NEXT: retq
5030 ;
5031 ; NoVLX-LABEL: test_vpcmpsgtb_v32i1_v64i1_mask:
5032 ; NoVLX: # %bb.0: # %entry
5033 ; NoVLX-NEXT: vpcmpgtb %ymm1, %ymm0, %ymm0
5034 ; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm1
5035 ; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
5036 ; NoVLX-NEXT: kmovw %k0, %ecx
5037 ; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm0
5038 ; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0
5039 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
5040 ; NoVLX-NEXT: kmovw %k0, %eax
5041 ; NoVLX-NEXT: shll $16, %eax
5042 ; NoVLX-NEXT: orl %ecx, %eax
5043 ; NoVLX-NEXT: vzeroupper
5044 ; NoVLX-NEXT: retq
5045 entry:
5046 %0 = bitcast <4 x i64> %__a to <32 x i8>
5047 %1 = bitcast <4 x i64> %__b to <32 x i8>
5048 %2 = icmp sgt <32 x i8> %0, %1
5049 %3 = shufflevector <32 x i1> %2, <32 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63>
5050 %4 = bitcast <64 x i1> %3 to i64
5051 ret i64 %4
5052 }
5054 define zeroext i64 @test_vpcmpsgtb_v32i1_v64i1_mask_mem(<4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
5055 ; VLX-LABEL: test_vpcmpsgtb_v32i1_v64i1_mask_mem:
5056 ; VLX: # %bb.0: # %entry
5057 ; VLX-NEXT: vpcmpgtb (%rdi), %ymm0, %k0
5058 ; VLX-NEXT: kmovq %k0, %rax
5059 ; VLX-NEXT: vzeroupper
5060 ; VLX-NEXT: retq
5061 ;
5062 ; NoVLX-LABEL: test_vpcmpsgtb_v32i1_v64i1_mask_mem:
5063 ; NoVLX: # %bb.0: # %entry
5064 ; NoVLX-NEXT: vpcmpgtb (%rdi), %ymm0, %ymm0
5065 ; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm1
5066 ; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
5067 ; NoVLX-NEXT: kmovw %k0, %ecx
5068 ; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm0
5069 ; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0
5070 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
5071 ; NoVLX-NEXT: kmovw %k0, %eax
5072 ; NoVLX-NEXT: shll $16, %eax
5073 ; NoVLX-NEXT: orl %ecx, %eax
5074 ; NoVLX-NEXT: vzeroupper
5075 ; NoVLX-NEXT: retq
5076 entry:
5077 %0 = bitcast <4 x i64> %__a to <32 x i8>
5078 %load = load <4 x i64>, <4 x i64>* %__b
5079 %1 = bitcast <4 x i64> %load to <32 x i8>
5080 %2 = icmp sgt <32 x i8> %0, %1
5081 %3 = shufflevector <32 x i1> %2, <32 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63>
5082 %4 = bitcast <64 x i1> %3 to i64
5083 ret i64 %4
5084 }
5086 define zeroext i64 @test_masked_vpcmpsgtb_v32i1_v64i1_mask(i32 zeroext %__u, <4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
5087 ; VLX-LABEL: test_masked_vpcmpsgtb_v32i1_v64i1_mask:
5088 ; VLX: # %bb.0: # %entry
5089 ; VLX-NEXT: kmovd %edi, %k1
5090 ; VLX-NEXT: vpcmpgtb %ymm1, %ymm0, %k0 {%k1}
5091 ; VLX-NEXT: kmovq %k0, %rax
5092 ; VLX-NEXT: vzeroupper
5093 ; VLX-NEXT: retq
5094 ;
5095 ; NoVLX-LABEL: test_masked_vpcmpsgtb_v32i1_v64i1_mask:
5096 ; NoVLX: # %bb.0: # %entry
5097 ; NoVLX-NEXT: vpcmpgtb %ymm1, %ymm0, %ymm0
5098 ; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm1
5099 ; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
5100 ; NoVLX-NEXT: kmovw %k0, %eax
5101 ; NoVLX-NEXT: andl %edi, %eax
5102 ; NoVLX-NEXT: shrl $16, %edi
5103 ; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm0
5104 ; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0
5105 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
5106 ; NoVLX-NEXT: kmovw %k0, %ecx
5107 ; NoVLX-NEXT: andl %edi, %ecx
5108 ; NoVLX-NEXT: shll $16, %ecx
5109 ; NoVLX-NEXT: movzwl %ax, %eax
5110 ; NoVLX-NEXT: orl %ecx, %eax
5111 ; NoVLX-NEXT: vzeroupper
5112 ; NoVLX-NEXT: retq
5113 entry:
5114 %0 = bitcast <4 x i64> %__a to <32 x i8>
5115 %1 = bitcast <4 x i64> %__b to <32 x i8>
5116 %2 = icmp sgt <32 x i8> %0, %1
5117 %3 = bitcast i32 %__u to <32 x i1>
5118 %4 = and <32 x i1> %2, %3
5119 %5 = shufflevector <32 x i1> %4, <32 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63>
5120 %6 = bitcast <64 x i1> %5 to i64
5121 ret i64 %6
5122 }
5124 define zeroext i64 @test_masked_vpcmpsgtb_v32i1_v64i1_mask_mem(i32 zeroext %__u, <4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
5125 ; VLX-LABEL: test_masked_vpcmpsgtb_v32i1_v64i1_mask_mem:
5126 ; VLX: # %bb.0: # %entry
5127 ; VLX-NEXT: kmovd %edi, %k1
5128 ; VLX-NEXT: vpcmpgtb (%rsi), %ymm0, %k0 {%k1}
5129 ; VLX-NEXT: kmovq %k0, %rax
5130 ; VLX-NEXT: vzeroupper
5131 ; VLX-NEXT: retq
5132 ;
5133 ; NoVLX-LABEL: test_masked_vpcmpsgtb_v32i1_v64i1_mask_mem:
5134 ; NoVLX: # %bb.0: # %entry
5135 ; NoVLX-NEXT: vpcmpgtb (%rsi), %ymm0, %ymm0
5136 ; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm1
5137 ; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
5138 ; NoVLX-NEXT: kmovw %k0, %eax
5139 ; NoVLX-NEXT: andl %edi, %eax
5140 ; NoVLX-NEXT: shrl $16, %edi
5141 ; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm0
5142 ; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0
5143 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
5144 ; NoVLX-NEXT: kmovw %k0, %ecx
5145 ; NoVLX-NEXT: andl %edi, %ecx
5146 ; NoVLX-NEXT: shll $16, %ecx
5147 ; NoVLX-NEXT: movzwl %ax, %eax
5148 ; NoVLX-NEXT: orl %ecx, %eax
5149 ; NoVLX-NEXT: vzeroupper
5150 ; NoVLX-NEXT: retq
5151 entry:
5152 %0 = bitcast <4 x i64> %__a to <32 x i8>
5153 %load = load <4 x i64>, <4 x i64>* %__b
5154 %1 = bitcast <4 x i64> %load to <32 x i8>
5155 %2 = icmp sgt <32 x i8> %0, %1
5156 %3 = bitcast i32 %__u to <32 x i1>
5157 %4 = and <32 x i1> %2, %3
5158 %5 = shufflevector <32 x i1> %4, <32 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63>
5159 %6 = bitcast <64 x i1> %5 to i64
5160 ret i64 %6
5161 }
5164 define zeroext i16 @test_vpcmpsgtw_v8i1_v16i1_mask(<2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
5165 ; VLX-LABEL: test_vpcmpsgtw_v8i1_v16i1_mask:
5166 ; VLX: # %bb.0: # %entry
5167 ; VLX-NEXT: vpcmpgtw %xmm1, %xmm0, %k0
5168 ; VLX-NEXT: kmovd %k0, %eax
5169 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
5170 ; VLX-NEXT: retq
5171 ;
5172 ; NoVLX-LABEL: test_vpcmpsgtw_v8i1_v16i1_mask:
5173 ; NoVLX: # %bb.0: # %entry
5174 ; NoVLX-NEXT: vpcmpgtw %xmm1, %xmm0, %xmm0
5175 ; NoVLX-NEXT: vpmovsxwq %xmm0, %zmm0
5176 ; NoVLX-NEXT: vptestmq %zmm0, %zmm0, %k0
5177 ; NoVLX-NEXT: kmovw %k0, %eax
5178 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
5179 ; NoVLX-NEXT: vzeroupper
5180 ; NoVLX-NEXT: retq
5181 entry:
5182 %0 = bitcast <2 x i64> %__a to <8 x i16>
5183 %1 = bitcast <2 x i64> %__b to <8 x i16>
5184 %2 = icmp sgt <8 x i16> %0, %1
5185 %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
5186 %4 = bitcast <16 x i1> %3 to i16
5187 ret i16 %4
5188 }
5190 define zeroext i16 @test_vpcmpsgtw_v8i1_v16i1_mask_mem(<2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
5191 ; VLX-LABEL: test_vpcmpsgtw_v8i1_v16i1_mask_mem:
5192 ; VLX: # %bb.0: # %entry
5193 ; VLX-NEXT: vpcmpgtw (%rdi), %xmm0, %k0
5194 ; VLX-NEXT: kmovd %k0, %eax
5195 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
5196 ; VLX-NEXT: retq
5197 ;
5198 ; NoVLX-LABEL: test_vpcmpsgtw_v8i1_v16i1_mask_mem:
5199 ; NoVLX: # %bb.0: # %entry
5200 ; NoVLX-NEXT: vpcmpgtw (%rdi), %xmm0, %xmm0
5201 ; NoVLX-NEXT: vpmovsxwq %xmm0, %zmm0
5202 ; NoVLX-NEXT: vptestmq %zmm0, %zmm0, %k0
5203 ; NoVLX-NEXT: kmovw %k0, %eax
5204 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
5205 ; NoVLX-NEXT: vzeroupper
5206 ; NoVLX-NEXT: retq
5207 entry:
5208 %0 = bitcast <2 x i64> %__a to <8 x i16>
5209 %load = load <2 x i64>, <2 x i64>* %__b
5210 %1 = bitcast <2 x i64> %load to <8 x i16>
5211 %2 = icmp sgt <8 x i16> %0, %1
5212 %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
5213 %4 = bitcast <16 x i1> %3 to i16
5214 ret i16 %4
5215 }
5217 define zeroext i16 @test_masked_vpcmpsgtw_v8i1_v16i1_mask(i8 zeroext %__u, <2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
5218 ; VLX-LABEL: test_masked_vpcmpsgtw_v8i1_v16i1_mask:
5219 ; VLX: # %bb.0: # %entry
5220 ; VLX-NEXT: kmovd %edi, %k1
5221 ; VLX-NEXT: vpcmpgtw %xmm1, %xmm0, %k0 {%k1}
5222 ; VLX-NEXT: kmovd %k0, %eax
5223 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
5224 ; VLX-NEXT: retq
5225 ;
5226 ; NoVLX-LABEL: test_masked_vpcmpsgtw_v8i1_v16i1_mask:
5227 ; NoVLX: # %bb.0: # %entry
5228 ; NoVLX-NEXT: vpcmpgtw %xmm1, %xmm0, %xmm0
5229 ; NoVLX-NEXT: vpmovsxwq %xmm0, %zmm0
5230 ; NoVLX-NEXT: kmovw %edi, %k1
5231 ; NoVLX-NEXT: vptestmq %zmm0, %zmm0, %k0 {%k1}
5232 ; NoVLX-NEXT: kmovw %k0, %eax
5233 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
5234 ; NoVLX-NEXT: vzeroupper
5235 ; NoVLX-NEXT: retq
5236 entry:
5237 %0 = bitcast <2 x i64> %__a to <8 x i16>
5238 %1 = bitcast <2 x i64> %__b to <8 x i16>
5239 %2 = icmp sgt <8 x i16> %0, %1
5240 %3 = bitcast i8 %__u to <8 x i1>
5241 %4 = and <8 x i1> %2, %3
5242 %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
5243 %6 = bitcast <16 x i1> %5 to i16
5244 ret i16 %6
5245 }
5247 define zeroext i16 @test_masked_vpcmpsgtw_v8i1_v16i1_mask_mem(i8 zeroext %__u, <2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
5248 ; VLX-LABEL: test_masked_vpcmpsgtw_v8i1_v16i1_mask_mem:
5249 ; VLX: # %bb.0: # %entry
5250 ; VLX-NEXT: kmovd %edi, %k1
5251 ; VLX-NEXT: vpcmpgtw (%rsi), %xmm0, %k0 {%k1}
5252 ; VLX-NEXT: kmovd %k0, %eax
5253 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
5254 ; VLX-NEXT: retq
5255 ;
5256 ; NoVLX-LABEL: test_masked_vpcmpsgtw_v8i1_v16i1_mask_mem:
5257 ; NoVLX: # %bb.0: # %entry
5258 ; NoVLX-NEXT: vpcmpgtw (%rsi), %xmm0, %xmm0
5259 ; NoVLX-NEXT: vpmovsxwq %xmm0, %zmm0
5260 ; NoVLX-NEXT: kmovw %edi, %k1
5261 ; NoVLX-NEXT: vptestmq %zmm0, %zmm0, %k0 {%k1}
5262 ; NoVLX-NEXT: kmovw %k0, %eax
5263 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
5264 ; NoVLX-NEXT: vzeroupper
5265 ; NoVLX-NEXT: retq
5266 entry:
5267 %0 = bitcast <2 x i64> %__a to <8 x i16>
5268 %load = load <2 x i64>, <2 x i64>* %__b
5269 %1 = bitcast <2 x i64> %load to <8 x i16>
5270 %2 = icmp sgt <8 x i16> %0, %1
5271 %3 = bitcast i8 %__u to <8 x i1>
5272 %4 = and <8 x i1> %2, %3
5273 %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
5274 %6 = bitcast <16 x i1> %5 to i16
5275 ret i16 %6
5276 }
5279 define zeroext i32 @test_vpcmpsgtw_v8i1_v32i1_mask(<2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
5280 ; VLX-LABEL: test_vpcmpsgtw_v8i1_v32i1_mask:
5281 ; VLX: # %bb.0: # %entry
5282 ; VLX-NEXT: vpcmpgtw %xmm1, %xmm0, %k0
5283 ; VLX-NEXT: kmovd %k0, %eax
5284 ; VLX-NEXT: retq
5285 ;
5286 ; NoVLX-LABEL: test_vpcmpsgtw_v8i1_v32i1_mask:
5287 ; NoVLX: # %bb.0: # %entry
5288 ; NoVLX-NEXT: vpcmpgtw %xmm1, %xmm0, %xmm0
5289 ; NoVLX-NEXT: vpmovsxwq %xmm0, %zmm0
5290 ; NoVLX-NEXT: vptestmq %zmm0, %zmm0, %k0
5291 ; NoVLX-NEXT: kmovw %k0, %eax
5292 ; NoVLX-NEXT: vzeroupper
5293 ; NoVLX-NEXT: retq
5294 entry:
5295 %0 = bitcast <2 x i64> %__a to <8 x i16>
5296 %1 = bitcast <2 x i64> %__b to <8 x i16>
5297 %2 = icmp sgt <8 x i16> %0, %1
5298 %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
5299 %4 = bitcast <32 x i1> %3 to i32
5300 ret i32 %4
5301 }
5303 define zeroext i32 @test_vpcmpsgtw_v8i1_v32i1_mask_mem(<2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
5304 ; VLX-LABEL: test_vpcmpsgtw_v8i1_v32i1_mask_mem:
5305 ; VLX: # %bb.0: # %entry
5306 ; VLX-NEXT: vpcmpgtw (%rdi), %xmm0, %k0
5307 ; VLX-NEXT: kmovd %k0, %eax
5308 ; VLX-NEXT: retq
5309 ;
5310 ; NoVLX-LABEL: test_vpcmpsgtw_v8i1_v32i1_mask_mem:
5311 ; NoVLX: # %bb.0: # %entry
5312 ; NoVLX-NEXT: vpcmpgtw (%rdi), %xmm0, %xmm0
5313 ; NoVLX-NEXT: vpmovsxwq %xmm0, %zmm0
5314 ; NoVLX-NEXT: vptestmq %zmm0, %zmm0, %k0
5315 ; NoVLX-NEXT: kmovw %k0, %eax
5316 ; NoVLX-NEXT: vzeroupper
5317 ; NoVLX-NEXT: retq
5318 entry:
5319 %0 = bitcast <2 x i64> %__a to <8 x i16>
5320 %load = load <2 x i64>, <2 x i64>* %__b
5321 %1 = bitcast <2 x i64> %load to <8 x i16>
5322 %2 = icmp sgt <8 x i16> %0, %1
5323 %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
5324 %4 = bitcast <32 x i1> %3 to i32
5325 ret i32 %4
5326 }
5328 define zeroext i32 @test_masked_vpcmpsgtw_v8i1_v32i1_mask(i8 zeroext %__u, <2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
5329 ; VLX-LABEL: test_masked_vpcmpsgtw_v8i1_v32i1_mask:
5330 ; VLX: # %bb.0: # %entry
5331 ; VLX-NEXT: kmovd %edi, %k1
5332 ; VLX-NEXT: vpcmpgtw %xmm1, %xmm0, %k0 {%k1}
5333 ; VLX-NEXT: kmovd %k0, %eax
5334 ; VLX-NEXT: retq
5335 ;
5336 ; NoVLX-LABEL: test_masked_vpcmpsgtw_v8i1_v32i1_mask:
5337 ; NoVLX: # %bb.0: # %entry
5338 ; NoVLX-NEXT: vpcmpgtw %xmm1, %xmm0, %xmm0
5339 ; NoVLX-NEXT: vpmovsxwq %xmm0, %zmm0
5340 ; NoVLX-NEXT: kmovw %edi, %k1
5341 ; NoVLX-NEXT: vptestmq %zmm0, %zmm0, %k0 {%k1}
5342 ; NoVLX-NEXT: kmovw %k0, %eax
5343 ; NoVLX-NEXT: vzeroupper
5344 ; NoVLX-NEXT: retq
5345 entry:
5346 %0 = bitcast <2 x i64> %__a to <8 x i16>
5347 %1 = bitcast <2 x i64> %__b to <8 x i16>
5348 %2 = icmp sgt <8 x i16> %0, %1
5349 %3 = bitcast i8 %__u to <8 x i1>
5350 %4 = and <8 x i1> %2, %3
5351 %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
5352 %6 = bitcast <32 x i1> %5 to i32
5353 ret i32 %6
5354 }
5356 define zeroext i32 @test_masked_vpcmpsgtw_v8i1_v32i1_mask_mem(i8 zeroext %__u, <2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
5357 ; VLX-LABEL: test_masked_vpcmpsgtw_v8i1_v32i1_mask_mem:
5358 ; VLX: # %bb.0: # %entry
5359 ; VLX-NEXT: kmovd %edi, %k1
5360 ; VLX-NEXT: vpcmpgtw (%rsi), %xmm0, %k0 {%k1}
5361 ; VLX-NEXT: kmovd %k0, %eax
5362 ; VLX-NEXT: retq
5363 ;
5364 ; NoVLX-LABEL: test_masked_vpcmpsgtw_v8i1_v32i1_mask_mem:
5365 ; NoVLX: # %bb.0: # %entry
5366 ; NoVLX-NEXT: vpcmpgtw (%rsi), %xmm0, %xmm0
5367 ; NoVLX-NEXT: vpmovsxwq %xmm0, %zmm0
5368 ; NoVLX-NEXT: kmovw %edi, %k1
5369 ; NoVLX-NEXT: vptestmq %zmm0, %zmm0, %k0 {%k1}
5370 ; NoVLX-NEXT: kmovw %k0, %eax
5371 ; NoVLX-NEXT: vzeroupper
5372 ; NoVLX-NEXT: retq
5373 entry:
5374 %0 = bitcast <2 x i64> %__a to <8 x i16>
5375 %load = load <2 x i64>, <2 x i64>* %__b
5376 %1 = bitcast <2 x i64> %load to <8 x i16>
5377 %2 = icmp sgt <8 x i16> %0, %1
5378 %3 = bitcast i8 %__u to <8 x i1>
5379 %4 = and <8 x i1> %2, %3
5380 %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
5381 %6 = bitcast <32 x i1> %5 to i32
5382 ret i32 %6
5383 }
5386 define zeroext i64 @test_vpcmpsgtw_v8i1_v64i1_mask(<2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
5387 ; VLX-LABEL: test_vpcmpsgtw_v8i1_v64i1_mask:
5388 ; VLX: # %bb.0: # %entry
5389 ; VLX-NEXT: vpcmpgtw %xmm1, %xmm0, %k0
5390 ; VLX-NEXT: kmovq %k0, %rax
5391 ; VLX-NEXT: retq
5392 ;
5393 ; NoVLX-LABEL: test_vpcmpsgtw_v8i1_v64i1_mask:
5394 ; NoVLX: # %bb.0: # %entry
5395 ; NoVLX-NEXT: vpcmpgtw %xmm1, %xmm0, %xmm0
5396 ; NoVLX-NEXT: vpmovsxwq %xmm0, %zmm0
5397 ; NoVLX-NEXT: vptestmq %zmm0, %zmm0, %k0
5398 ; NoVLX-NEXT: kmovw %k0, %eax
5399 ; NoVLX-NEXT: vzeroupper
5400 ; NoVLX-NEXT: retq
5401 entry:
5402 %0 = bitcast <2 x i64> %__a to <8 x i16>
5403 %1 = bitcast <2 x i64> %__b to <8 x i16>
5404 %2 = icmp sgt <8 x i16> %0, %1
5405 %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
5406 %4 = bitcast <64 x i1> %3 to i64
5407 ret i64 %4
5408 }
5410 define zeroext i64 @test_vpcmpsgtw_v8i1_v64i1_mask_mem(<2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
5411 ; VLX-LABEL: test_vpcmpsgtw_v8i1_v64i1_mask_mem:
5412 ; VLX: # %bb.0: # %entry
5413 ; VLX-NEXT: vpcmpgtw (%rdi), %xmm0, %k0
5414 ; VLX-NEXT: kmovq %k0, %rax
5415 ; VLX-NEXT: retq
5416 ;
5417 ; NoVLX-LABEL: test_vpcmpsgtw_v8i1_v64i1_mask_mem:
5418 ; NoVLX: # %bb.0: # %entry
5419 ; NoVLX-NEXT: vpcmpgtw (%rdi), %xmm0, %xmm0
5420 ; NoVLX-NEXT: vpmovsxwq %xmm0, %zmm0
5421 ; NoVLX-NEXT: vptestmq %zmm0, %zmm0, %k0
5422 ; NoVLX-NEXT: kmovw %k0, %eax
5423 ; NoVLX-NEXT: vzeroupper
5424 ; NoVLX-NEXT: retq
5425 entry:
5426 %0 = bitcast <2 x i64> %__a to <8 x i16>
5427 %load = load <2 x i64>, <2 x i64>* %__b
5428 %1 = bitcast <2 x i64> %load to <8 x i16>
5429 %2 = icmp sgt <8 x i16> %0, %1
5430 %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
5431 %4 = bitcast <64 x i1> %3 to i64
5432 ret i64 %4
5433 }
5435 define zeroext i64 @test_masked_vpcmpsgtw_v8i1_v64i1_mask(i8 zeroext %__u, <2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
5436 ; VLX-LABEL: test_masked_vpcmpsgtw_v8i1_v64i1_mask:
5437 ; VLX: # %bb.0: # %entry
5438 ; VLX-NEXT: kmovd %edi, %k1
5439 ; VLX-NEXT: vpcmpgtw %xmm1, %xmm0, %k0 {%k1}
5440 ; VLX-NEXT: kmovq %k0, %rax
5441 ; VLX-NEXT: retq
5442 ;
5443 ; NoVLX-LABEL: test_masked_vpcmpsgtw_v8i1_v64i1_mask:
5444 ; NoVLX: # %bb.0: # %entry
5445 ; NoVLX-NEXT: vpcmpgtw %xmm1, %xmm0, %xmm0
5446 ; NoVLX-NEXT: vpmovsxwq %xmm0, %zmm0
5447 ; NoVLX-NEXT: kmovw %edi, %k1
5448 ; NoVLX-NEXT: vptestmq %zmm0, %zmm0, %k0 {%k1}
5449 ; NoVLX-NEXT: kmovw %k0, %eax
5450 ; NoVLX-NEXT: vzeroupper
5451 ; NoVLX-NEXT: retq
5452 entry:
5453 %0 = bitcast <2 x i64> %__a to <8 x i16>
5454 %1 = bitcast <2 x i64> %__b to <8 x i16>
5455 %2 = icmp sgt <8 x i16> %0, %1
5456 %3 = bitcast i8 %__u to <8 x i1>
5457 %4 = and <8 x i1> %2, %3
5458 %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
5459 %6 = bitcast <64 x i1> %5 to i64
5460 ret i64 %6
5461 }
5463 define zeroext i64 @test_masked_vpcmpsgtw_v8i1_v64i1_mask_mem(i8 zeroext %__u, <2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
5464 ; VLX-LABEL: test_masked_vpcmpsgtw_v8i1_v64i1_mask_mem:
5465 ; VLX: # %bb.0: # %entry
5466 ; VLX-NEXT: kmovd %edi, %k1
5467 ; VLX-NEXT: vpcmpgtw (%rsi), %xmm0, %k0 {%k1}
5468 ; VLX-NEXT: kmovq %k0, %rax
5469 ; VLX-NEXT: retq
5470 ;
5471 ; NoVLX-LABEL: test_masked_vpcmpsgtw_v8i1_v64i1_mask_mem:
5472 ; NoVLX: # %bb.0: # %entry
5473 ; NoVLX-NEXT: vpcmpgtw (%rsi), %xmm0, %xmm0
5474 ; NoVLX-NEXT: vpmovsxwq %xmm0, %zmm0
5475 ; NoVLX-NEXT: kmovw %edi, %k1
5476 ; NoVLX-NEXT: vptestmq %zmm0, %zmm0, %k0 {%k1}
5477 ; NoVLX-NEXT: kmovw %k0, %eax
5478 ; NoVLX-NEXT: vzeroupper
5479 ; NoVLX-NEXT: retq
5480 entry:
5481 %0 = bitcast <2 x i64> %__a to <8 x i16>
5482 %load = load <2 x i64>, <2 x i64>* %__b
5483 %1 = bitcast <2 x i64> %load to <8 x i16>
5484 %2 = icmp sgt <8 x i16> %0, %1
5485 %3 = bitcast i8 %__u to <8 x i1>
5486 %4 = and <8 x i1> %2, %3
5487 %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
5488 %6 = bitcast <64 x i1> %5 to i64
5489 ret i64 %6
5490 }
5493 define zeroext i32 @test_vpcmpsgtw_v16i1_v32i1_mask(<4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
5494 ; VLX-LABEL: test_vpcmpsgtw_v16i1_v32i1_mask:
5495 ; VLX: # %bb.0: # %entry
5496 ; VLX-NEXT: vpcmpgtw %ymm1, %ymm0, %k0
5497 ; VLX-NEXT: kmovd %k0, %eax
5498 ; VLX-NEXT: vzeroupper
5499 ; VLX-NEXT: retq
5500 ;
5501 ; NoVLX-LABEL: test_vpcmpsgtw_v16i1_v32i1_mask:
5502 ; NoVLX: # %bb.0: # %entry
5503 ; NoVLX-NEXT: vpcmpgtw %ymm1, %ymm0, %ymm0
5504 ; NoVLX-NEXT: vpmovsxwd %ymm0, %zmm0
5505 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
5506 ; NoVLX-NEXT: kmovw %k0, %eax
5507 ; NoVLX-NEXT: vzeroupper
5508 ; NoVLX-NEXT: retq
5509 entry:
5510 %0 = bitcast <4 x i64> %__a to <16 x i16>
5511 %1 = bitcast <4 x i64> %__b to <16 x i16>
5512 %2 = icmp sgt <16 x i16> %0, %1
5513 %3 = shufflevector <16 x i1> %2, <16 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>
5514 %4 = bitcast <32 x i1> %3 to i32
5515 ret i32 %4
5516 }
5518 define zeroext i32 @test_vpcmpsgtw_v16i1_v32i1_mask_mem(<4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
5519 ; VLX-LABEL: test_vpcmpsgtw_v16i1_v32i1_mask_mem:
5520 ; VLX: # %bb.0: # %entry
5521 ; VLX-NEXT: vpcmpgtw (%rdi), %ymm0, %k0
5522 ; VLX-NEXT: kmovd %k0, %eax
5523 ; VLX-NEXT: vzeroupper
5524 ; VLX-NEXT: retq
5525 ;
5526 ; NoVLX-LABEL: test_vpcmpsgtw_v16i1_v32i1_mask_mem:
5527 ; NoVLX: # %bb.0: # %entry
5528 ; NoVLX-NEXT: vpcmpgtw (%rdi), %ymm0, %ymm0
5529 ; NoVLX-NEXT: vpmovsxwd %ymm0, %zmm0
5530 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
5531 ; NoVLX-NEXT: kmovw %k0, %eax
5532 ; NoVLX-NEXT: vzeroupper
5533 ; NoVLX-NEXT: retq
5534 entry:
5535 %0 = bitcast <4 x i64> %__a to <16 x i16>
5536 %load = load <4 x i64>, <4 x i64>* %__b
5537 %1 = bitcast <4 x i64> %load to <16 x i16>
5538 %2 = icmp sgt <16 x i16> %0, %1
5539 %3 = shufflevector <16 x i1> %2, <16 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>
5540 %4 = bitcast <32 x i1> %3 to i32
5541 ret i32 %4
5542 }
5544 define zeroext i32 @test_masked_vpcmpsgtw_v16i1_v32i1_mask(i16 zeroext %__u, <4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
5545 ; VLX-LABEL: test_masked_vpcmpsgtw_v16i1_v32i1_mask:
5546 ; VLX: # %bb.0: # %entry
5547 ; VLX-NEXT: kmovd %edi, %k1
5548 ; VLX-NEXT: vpcmpgtw %ymm1, %ymm0, %k0 {%k1}
5549 ; VLX-NEXT: kmovd %k0, %eax
5550 ; VLX-NEXT: vzeroupper
5551 ; VLX-NEXT: retq
5552 ;
5553 ; NoVLX-LABEL: test_masked_vpcmpsgtw_v16i1_v32i1_mask:
5554 ; NoVLX: # %bb.0: # %entry
5555 ; NoVLX-NEXT: vpcmpgtw %ymm1, %ymm0, %ymm0
5556 ; NoVLX-NEXT: vpmovsxwd %ymm0, %zmm0
5557 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
5558 ; NoVLX-NEXT: kmovw %k0, %eax
5559 ; NoVLX-NEXT: andl %edi, %eax
5560 ; NoVLX-NEXT: vzeroupper
5561 ; NoVLX-NEXT: retq
5562 entry:
5563 %0 = bitcast <4 x i64> %__a to <16 x i16>
5564 %1 = bitcast <4 x i64> %__b to <16 x i16>
5565 %2 = icmp sgt <16 x i16> %0, %1
5566 %3 = bitcast i16 %__u to <16 x i1>
5567 %4 = and <16 x i1> %2, %3
5568 %5 = shufflevector <16 x i1> %4, <16 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>
5569 %6 = bitcast <32 x i1> %5 to i32
5570 ret i32 %6
5571 }
5573 define zeroext i32 @test_masked_vpcmpsgtw_v16i1_v32i1_mask_mem(i16 zeroext %__u, <4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
5574 ; VLX-LABEL: test_masked_vpcmpsgtw_v16i1_v32i1_mask_mem:
5575 ; VLX: # %bb.0: # %entry
5576 ; VLX-NEXT: kmovd %edi, %k1
5577 ; VLX-NEXT: vpcmpgtw (%rsi), %ymm0, %k0 {%k1}
5578 ; VLX-NEXT: kmovd %k0, %eax
5579 ; VLX-NEXT: vzeroupper
5580 ; VLX-NEXT: retq
5581 ;
5582 ; NoVLX-LABEL: test_masked_vpcmpsgtw_v16i1_v32i1_mask_mem:
5583 ; NoVLX: # %bb.0: # %entry
5584 ; NoVLX-NEXT: vpcmpgtw (%rsi), %ymm0, %ymm0
5585 ; NoVLX-NEXT: vpmovsxwd %ymm0, %zmm0
5586 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
5587 ; NoVLX-NEXT: kmovw %k0, %eax
5588 ; NoVLX-NEXT: andl %edi, %eax
5589 ; NoVLX-NEXT: vzeroupper
5590 ; NoVLX-NEXT: retq
5591 entry:
5592 %0 = bitcast <4 x i64> %__a to <16 x i16>
5593 %load = load <4 x i64>, <4 x i64>* %__b
5594 %1 = bitcast <4 x i64> %load to <16 x i16>
5595 %2 = icmp sgt <16 x i16> %0, %1
5596 %3 = bitcast i16 %__u to <16 x i1>
5597 %4 = and <16 x i1> %2, %3
5598 %5 = shufflevector <16 x i1> %4, <16 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>
5599 %6 = bitcast <32 x i1> %5 to i32
5600 ret i32 %6
5601 }
5604 define zeroext i64 @test_vpcmpsgtw_v16i1_v64i1_mask(<4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
5605 ; VLX-LABEL: test_vpcmpsgtw_v16i1_v64i1_mask:
5606 ; VLX: # %bb.0: # %entry
5607 ; VLX-NEXT: vpcmpgtw %ymm1, %ymm0, %k0
5608 ; VLX-NEXT: kmovq %k0, %rax
5609 ; VLX-NEXT: vzeroupper
5610 ; VLX-NEXT: retq
5611 ;
5612 ; NoVLX-LABEL: test_vpcmpsgtw_v16i1_v64i1_mask:
5613 ; NoVLX: # %bb.0: # %entry
5614 ; NoVLX-NEXT: vpcmpgtw %ymm1, %ymm0, %ymm0
5615 ; NoVLX-NEXT: vpmovsxwd %ymm0, %zmm0
5616 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
5617 ; NoVLX-NEXT: kmovw %k0, %eax
5618 ; NoVLX-NEXT: vzeroupper
5619 ; NoVLX-NEXT: retq
5620 entry:
5621 %0 = bitcast <4 x i64> %__a to <16 x i16>
5622 %1 = bitcast <4 x i64> %__b to <16 x i16>
5623 %2 = icmp sgt <16 x i16> %0, %1
5624 %3 = shufflevector <16 x i1> %2, <16 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>
5625 %4 = bitcast <64 x i1> %3 to i64
5626 ret i64 %4
5627 }
5629 define zeroext i64 @test_vpcmpsgtw_v16i1_v64i1_mask_mem(<4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
5630 ; VLX-LABEL: test_vpcmpsgtw_v16i1_v64i1_mask_mem:
5631 ; VLX: # %bb.0: # %entry
5632 ; VLX-NEXT: vpcmpgtw (%rdi), %ymm0, %k0
5633 ; VLX-NEXT: kmovq %k0, %rax
5634 ; VLX-NEXT: vzeroupper
5635 ; VLX-NEXT: retq
5636 ;
5637 ; NoVLX-LABEL: test_vpcmpsgtw_v16i1_v64i1_mask_mem:
5638 ; NoVLX: # %bb.0: # %entry
5639 ; NoVLX-NEXT: vpcmpgtw (%rdi), %ymm0, %ymm0
5640 ; NoVLX-NEXT: vpmovsxwd %ymm0, %zmm0
5641 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
5642 ; NoVLX-NEXT: kmovw %k0, %eax
5643 ; NoVLX-NEXT: vzeroupper
5644 ; NoVLX-NEXT: retq
5645 entry:
5646 %0 = bitcast <4 x i64> %__a to <16 x i16>
5647 %load = load <4 x i64>, <4 x i64>* %__b
5648 %1 = bitcast <4 x i64> %load to <16 x i16>
5649 %2 = icmp sgt <16 x i16> %0, %1
5650 %3 = shufflevector <16 x i1> %2, <16 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>
5651 %4 = bitcast <64 x i1> %3 to i64
5652 ret i64 %4
5653 }
5655 define zeroext i64 @test_masked_vpcmpsgtw_v16i1_v64i1_mask(i16 zeroext %__u, <4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
5656 ; VLX-LABEL: test_masked_vpcmpsgtw_v16i1_v64i1_mask:
5657 ; VLX: # %bb.0: # %entry
5658 ; VLX-NEXT: kmovd %edi, %k1
5659 ; VLX-NEXT: vpcmpgtw %ymm1, %ymm0, %k0 {%k1}
5660 ; VLX-NEXT: kmovq %k0, %rax
5661 ; VLX-NEXT: vzeroupper
5662 ; VLX-NEXT: retq
5663 ;
5664 ; NoVLX-LABEL: test_masked_vpcmpsgtw_v16i1_v64i1_mask:
5665 ; NoVLX: # %bb.0: # %entry
5666 ; NoVLX-NEXT: vpcmpgtw %ymm1, %ymm0, %ymm0
5667 ; NoVLX-NEXT: vpmovsxwd %ymm0, %zmm0
5668 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
5669 ; NoVLX-NEXT: kmovw %k0, %eax
5670 ; NoVLX-NEXT: andl %edi, %eax
5671 ; NoVLX-NEXT: vzeroupper
5672 ; NoVLX-NEXT: retq
5673 entry:
5674 %0 = bitcast <4 x i64> %__a to <16 x i16>
5675 %1 = bitcast <4 x i64> %__b to <16 x i16>
5676 %2 = icmp sgt <16 x i16> %0, %1
5677 %3 = bitcast i16 %__u to <16 x i1>
5678 %4 = and <16 x i1> %2, %3
5679 %5 = shufflevector <16 x i1> %4, <16 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>
5680 %6 = bitcast <64 x i1> %5 to i64
5681 ret i64 %6
5682 }
5684 define zeroext i64 @test_masked_vpcmpsgtw_v16i1_v64i1_mask_mem(i16 zeroext %__u, <4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
5685 ; VLX-LABEL: test_masked_vpcmpsgtw_v16i1_v64i1_mask_mem:
5686 ; VLX: # %bb.0: # %entry
5687 ; VLX-NEXT: kmovd %edi, %k1
5688 ; VLX-NEXT: vpcmpgtw (%rsi), %ymm0, %k0 {%k1}
5689 ; VLX-NEXT: kmovq %k0, %rax
5690 ; VLX-NEXT: vzeroupper
5691 ; VLX-NEXT: retq
5692 ;
5693 ; NoVLX-LABEL: test_masked_vpcmpsgtw_v16i1_v64i1_mask_mem:
5694 ; NoVLX: # %bb.0: # %entry
5695 ; NoVLX-NEXT: vpcmpgtw (%rsi), %ymm0, %ymm0
5696 ; NoVLX-NEXT: vpmovsxwd %ymm0, %zmm0
5697 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
5698 ; NoVLX-NEXT: kmovw %k0, %eax
5699 ; NoVLX-NEXT: andl %edi, %eax
5700 ; NoVLX-NEXT: vzeroupper
5701 ; NoVLX-NEXT: retq
5702 entry:
5703 %0 = bitcast <4 x i64> %__a to <16 x i16>
5704 %load = load <4 x i64>, <4 x i64>* %__b
5705 %1 = bitcast <4 x i64> %load to <16 x i16>
5706 %2 = icmp sgt <16 x i16> %0, %1
5707 %3 = bitcast i16 %__u to <16 x i1>
5708 %4 = and <16 x i1> %2, %3
5709 %5 = shufflevector <16 x i1> %4, <16 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>
5710 %6 = bitcast <64 x i1> %5 to i64
5711 ret i64 %6
5712 }
5715 define zeroext i64 @test_vpcmpsgtw_v32i1_v64i1_mask(<8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
5716 ; VLX-LABEL: test_vpcmpsgtw_v32i1_v64i1_mask:
5717 ; VLX: # %bb.0: # %entry
5718 ; VLX-NEXT: vpcmpgtw %zmm1, %zmm0, %k0
5719 ; VLX-NEXT: kmovq %k0, %rax
5720 ; VLX-NEXT: vzeroupper
5721 ; VLX-NEXT: retq
5722 ;
5723 ; NoVLX-LABEL: test_vpcmpsgtw_v32i1_v64i1_mask:
5724 ; NoVLX: # %bb.0: # %entry
5725 ; NoVLX-NEXT: vextracti64x4 $1, %zmm0, %ymm2
5726 ; NoVLX-NEXT: vextracti64x4 $1, %zmm1, %ymm3
5727 ; NoVLX-NEXT: vpcmpgtw %ymm3, %ymm2, %ymm2
5728 ; NoVLX-NEXT: vpcmpgtw %ymm1, %ymm0, %ymm0
5729 ; NoVLX-NEXT: vpmovsxwd %ymm0, %zmm0
5730 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
5731 ; NoVLX-NEXT: kmovw %k0, %ecx
5732 ; NoVLX-NEXT: vpmovsxwd %ymm2, %zmm0
5733 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
5734 ; NoVLX-NEXT: kmovw %k0, %eax
5735 ; NoVLX-NEXT: shll $16, %eax
5736 ; NoVLX-NEXT: orl %ecx, %eax
5737 ; NoVLX-NEXT: vzeroupper
5738 ; NoVLX-NEXT: retq
5739 entry:
5740 %0 = bitcast <8 x i64> %__a to <32 x i16>
5741 %1 = bitcast <8 x i64> %__b to <32 x i16>
5742 %2 = icmp sgt <32 x i16> %0, %1
5743 %3 = shufflevector <32 x i1> %2, <32 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63>
5744 %4 = bitcast <64 x i1> %3 to i64
5745 ret i64 %4
5746 }
5748 define zeroext i64 @test_vpcmpsgtw_v32i1_v64i1_mask_mem(<8 x i64> %__a, <8 x i64>* %__b) local_unnamed_addr {
5749 ; VLX-LABEL: test_vpcmpsgtw_v32i1_v64i1_mask_mem:
5750 ; VLX: # %bb.0: # %entry
5751 ; VLX-NEXT: vpcmpgtw (%rdi), %zmm0, %k0
5752 ; VLX-NEXT: kmovq %k0, %rax
5753 ; VLX-NEXT: vzeroupper
5754 ; VLX-NEXT: retq
5755 ;
5756 ; NoVLX-LABEL: test_vpcmpsgtw_v32i1_v64i1_mask_mem:
5757 ; NoVLX: # %bb.0: # %entry
5758 ; NoVLX-NEXT: vextracti64x4 $1, %zmm0, %ymm1
5759 ; NoVLX-NEXT: vpcmpgtw (%rdi), %ymm0, %ymm0
5760 ; NoVLX-NEXT: vpmovsxwd %ymm0, %zmm0
5761 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
5762 ; NoVLX-NEXT: kmovw %k0, %ecx
5763 ; NoVLX-NEXT: vpcmpgtw 32(%rdi), %ymm1, %ymm0
5764 ; NoVLX-NEXT: vpmovsxwd %ymm0, %zmm0
5765 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
5766 ; NoVLX-NEXT: kmovw %k0, %eax
5767 ; NoVLX-NEXT: shll $16, %eax
5768 ; NoVLX-NEXT: orl %ecx, %eax
5769 ; NoVLX-NEXT: vzeroupper
5770 ; NoVLX-NEXT: retq
5771 entry:
5772 %0 = bitcast <8 x i64> %__a to <32 x i16>
5773 %load = load <8 x i64>, <8 x i64>* %__b
5774 %1 = bitcast <8 x i64> %load to <32 x i16>
5775 %2 = icmp sgt <32 x i16> %0, %1
5776 %3 = shufflevector <32 x i1> %2, <32 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63>
5777 %4 = bitcast <64 x i1> %3 to i64
5778 ret i64 %4
5779 }
5781 define zeroext i64 @test_masked_vpcmpsgtw_v32i1_v64i1_mask(i32 zeroext %__u, <8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
5782 ; VLX-LABEL: test_masked_vpcmpsgtw_v32i1_v64i1_mask:
5783 ; VLX: # %bb.0: # %entry
5784 ; VLX-NEXT: kmovd %edi, %k1
5785 ; VLX-NEXT: vpcmpgtw %zmm1, %zmm0, %k0 {%k1}
5786 ; VLX-NEXT: kmovq %k0, %rax
5787 ; VLX-NEXT: vzeroupper
5788 ; VLX-NEXT: retq
5789 ;
5790 ; NoVLX-LABEL: test_masked_vpcmpsgtw_v32i1_v64i1_mask:
5791 ; NoVLX: # %bb.0: # %entry
5792 ; NoVLX-NEXT: vpcmpgtw %ymm1, %ymm0, %ymm2
5793 ; NoVLX-NEXT: vpmovsxwd %ymm2, %zmm2
5794 ; NoVLX-NEXT: vptestmd %zmm2, %zmm2, %k0
5795 ; NoVLX-NEXT: kmovw %k0, %eax
5796 ; NoVLX-NEXT: andl %edi, %eax
5797 ; NoVLX-NEXT: shrl $16, %edi
5798 ; NoVLX-NEXT: vextracti64x4 $1, %zmm0, %ymm0
5799 ; NoVLX-NEXT: vextracti64x4 $1, %zmm1, %ymm1
5800 ; NoVLX-NEXT: vpcmpgtw %ymm1, %ymm0, %ymm0
5801 ; NoVLX-NEXT: vpmovsxwd %ymm0, %zmm0
5802 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
5803 ; NoVLX-NEXT: kmovw %k0, %ecx
5804 ; NoVLX-NEXT: andl %edi, %ecx
5805 ; NoVLX-NEXT: shll $16, %ecx
5806 ; NoVLX-NEXT: movzwl %ax, %eax
5807 ; NoVLX-NEXT: orl %ecx, %eax
5808 ; NoVLX-NEXT: vzeroupper
5809 ; NoVLX-NEXT: retq
5810 entry:
5811 %0 = bitcast <8 x i64> %__a to <32 x i16>
5812 %1 = bitcast <8 x i64> %__b to <32 x i16>
5813 %2 = icmp sgt <32 x i16> %0, %1
5814 %3 = bitcast i32 %__u to <32 x i1>
5815 %4 = and <32 x i1> %2, %3
5816 %5 = shufflevector <32 x i1> %4, <32 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63>
5817 %6 = bitcast <64 x i1> %5 to i64
5818 ret i64 %6
5819 }
5821 define zeroext i64 @test_masked_vpcmpsgtw_v32i1_v64i1_mask_mem(i32 zeroext %__u, <8 x i64> %__a, <8 x i64>* %__b) local_unnamed_addr {
5822 ; VLX-LABEL: test_masked_vpcmpsgtw_v32i1_v64i1_mask_mem:
5823 ; VLX: # %bb.0: # %entry
5824 ; VLX-NEXT: kmovd %edi, %k1
5825 ; VLX-NEXT: vpcmpgtw (%rsi), %zmm0, %k0 {%k1}
5826 ; VLX-NEXT: kmovq %k0, %rax
5827 ; VLX-NEXT: vzeroupper
5828 ; VLX-NEXT: retq
5829 ;
5830 ; NoVLX-LABEL: test_masked_vpcmpsgtw_v32i1_v64i1_mask_mem:
5831 ; NoVLX: # %bb.0: # %entry
5832 ; NoVLX-NEXT: vpcmpgtw (%rsi), %ymm0, %ymm1
5833 ; NoVLX-NEXT: vpmovsxwd %ymm1, %zmm1
5834 ; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
5835 ; NoVLX-NEXT: kmovw %k0, %eax
5836 ; NoVLX-NEXT: andl %edi, %eax
5837 ; NoVLX-NEXT: shrl $16, %edi
5838 ; NoVLX-NEXT: vextracti64x4 $1, %zmm0, %ymm0
5839 ; NoVLX-NEXT: vpcmpgtw 32(%rsi), %ymm0, %ymm0
5840 ; NoVLX-NEXT: vpmovsxwd %ymm0, %zmm0
5841 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
5842 ; NoVLX-NEXT: kmovw %k0, %ecx
5843 ; NoVLX-NEXT: andl %edi, %ecx
5844 ; NoVLX-NEXT: shll $16, %ecx
5845 ; NoVLX-NEXT: movzwl %ax, %eax
5846 ; NoVLX-NEXT: orl %ecx, %eax
5847 ; NoVLX-NEXT: vzeroupper
5848 ; NoVLX-NEXT: retq
5849 entry:
5850 %0 = bitcast <8 x i64> %__a to <32 x i16>
5851 %load = load <8 x i64>, <8 x i64>* %__b
5852 %1 = bitcast <8 x i64> %load to <32 x i16>
5853 %2 = icmp sgt <32 x i16> %0, %1
5854 %3 = bitcast i32 %__u to <32 x i1>
5855 %4 = and <32 x i1> %2, %3
5856 %5 = shufflevector <32 x i1> %4, <32 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63>
5857 %6 = bitcast <64 x i1> %5 to i64
5858 ret i64 %6
5859 }
5862 define zeroext i8 @test_vpcmpsgtd_v4i1_v8i1_mask(<2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
5863 ; VLX-LABEL: test_vpcmpsgtd_v4i1_v8i1_mask:
5864 ; VLX: # %bb.0: # %entry
5865 ; VLX-NEXT: vpcmpgtd %xmm1, %xmm0, %k0
5866 ; VLX-NEXT: kmovd %k0, %eax
5867 ; VLX-NEXT: # kill: def $al killed $al killed $eax
5868 ; VLX-NEXT: retq
5869 ;
5870 ; NoVLX-LABEL: test_vpcmpsgtd_v4i1_v8i1_mask:
5871 ; NoVLX: # %bb.0: # %entry
5872 ; NoVLX-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
5873 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
5874 ; NoVLX-NEXT: vpcmpgtd %zmm1, %zmm0, %k0
5875 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
5876 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
5877 ; NoVLX-NEXT: kmovw %k0, %eax
5878 ; NoVLX-NEXT: # kill: def $al killed $al killed $eax
5879 ; NoVLX-NEXT: vzeroupper
5880 ; NoVLX-NEXT: retq
5881 entry:
5882 %0 = bitcast <2 x i64> %__a to <4 x i32>
5883 %1 = bitcast <2 x i64> %__b to <4 x i32>
5884 %2 = icmp sgt <4 x i32> %0, %1
5885 %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
5886 %4 = bitcast <8 x i1> %3 to i8
5887 ret i8 %4
5888 }
5890 define zeroext i8 @test_vpcmpsgtd_v4i1_v8i1_mask_mem(<2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
5891 ; VLX-LABEL: test_vpcmpsgtd_v4i1_v8i1_mask_mem:
5892 ; VLX: # %bb.0: # %entry
5893 ; VLX-NEXT: vpcmpgtd (%rdi), %xmm0, %k0
5894 ; VLX-NEXT: kmovd %k0, %eax
5895 ; VLX-NEXT: # kill: def $al killed $al killed $eax
5896 ; VLX-NEXT: retq
5897 ;
5898 ; NoVLX-LABEL: test_vpcmpsgtd_v4i1_v8i1_mask_mem:
5899 ; NoVLX: # %bb.0: # %entry
5900 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
5901 ; NoVLX-NEXT: vmovdqa (%rdi), %xmm1
5902 ; NoVLX-NEXT: vpcmpgtd %zmm1, %zmm0, %k0
5903 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
5904 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
5905 ; NoVLX-NEXT: kmovw %k0, %eax
5906 ; NoVLX-NEXT: # kill: def $al killed $al killed $eax
5907 ; NoVLX-NEXT: vzeroupper
5908 ; NoVLX-NEXT: retq
5909 entry:
5910 %0 = bitcast <2 x i64> %__a to <4 x i32>
5911 %load = load <2 x i64>, <2 x i64>* %__b
5912 %1 = bitcast <2 x i64> %load to <4 x i32>
5913 %2 = icmp sgt <4 x i32> %0, %1
5914 %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
5915 %4 = bitcast <8 x i1> %3 to i8
5916 ret i8 %4
5917 }
5919 define zeroext i8 @test_masked_vpcmpsgtd_v4i1_v8i1_mask(i8 zeroext %__u, <2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
5920 ; VLX-LABEL: test_masked_vpcmpsgtd_v4i1_v8i1_mask:
5921 ; VLX: # %bb.0: # %entry
5922 ; VLX-NEXT: kmovd %edi, %k1
5923 ; VLX-NEXT: vpcmpgtd %xmm1, %xmm0, %k0 {%k1}
5924 ; VLX-NEXT: kmovd %k0, %eax
5925 ; VLX-NEXT: # kill: def $al killed $al killed $eax
5926 ; VLX-NEXT: retq
5927 ;
5928 ; NoVLX-LABEL: test_masked_vpcmpsgtd_v4i1_v8i1_mask:
5929 ; NoVLX: # %bb.0: # %entry
5930 ; NoVLX-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
5931 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
5932 ; NoVLX-NEXT: kmovw %edi, %k1
5933 ; NoVLX-NEXT: vpcmpgtd %zmm1, %zmm0, %k0 {%k1}
5934 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
5935 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
5936 ; NoVLX-NEXT: kmovw %k0, %eax
5937 ; NoVLX-NEXT: # kill: def $al killed $al killed $eax
5938 ; NoVLX-NEXT: vzeroupper
5939 ; NoVLX-NEXT: retq
5940 entry:
5941 %0 = bitcast <2 x i64> %__a to <4 x i32>
5942 %1 = bitcast <2 x i64> %__b to <4 x i32>
5943 %2 = icmp sgt <4 x i32> %0, %1
5944 %3 = bitcast i8 %__u to <8 x i1>
5945 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
5946 %4 = and <4 x i1> %2, %extract.i
5947 %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
5948 %6 = bitcast <8 x i1> %5 to i8
5949 ret i8 %6
5950 }
5952 define zeroext i8 @test_masked_vpcmpsgtd_v4i1_v8i1_mask_mem(i8 zeroext %__u, <2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
5953 ; VLX-LABEL: test_masked_vpcmpsgtd_v4i1_v8i1_mask_mem:
5954 ; VLX: # %bb.0: # %entry
5955 ; VLX-NEXT: kmovd %edi, %k1
5956 ; VLX-NEXT: vpcmpgtd (%rsi), %xmm0, %k0 {%k1}
5957 ; VLX-NEXT: kmovd %k0, %eax
5958 ; VLX-NEXT: # kill: def $al killed $al killed $eax
5959 ; VLX-NEXT: retq
5960 ;
5961 ; NoVLX-LABEL: test_masked_vpcmpsgtd_v4i1_v8i1_mask_mem:
5962 ; NoVLX: # %bb.0: # %entry
5963 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
5964 ; NoVLX-NEXT: vmovdqa (%rsi), %xmm1
5965 ; NoVLX-NEXT: kmovw %edi, %k1
5966 ; NoVLX-NEXT: vpcmpgtd %zmm1, %zmm0, %k0 {%k1}
5967 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
5968 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
5969 ; NoVLX-NEXT: kmovw %k0, %eax
5970 ; NoVLX-NEXT: # kill: def $al killed $al killed $eax
5971 ; NoVLX-NEXT: vzeroupper
5972 ; NoVLX-NEXT: retq
5973 entry:
5974 %0 = bitcast <2 x i64> %__a to <4 x i32>
5975 %load = load <2 x i64>, <2 x i64>* %__b
5976 %1 = bitcast <2 x i64> %load to <4 x i32>
5977 %2 = icmp sgt <4 x i32> %0, %1
5978 %3 = bitcast i8 %__u to <8 x i1>
5979 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
5980 %4 = and <4 x i1> %2, %extract.i
5981 %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
5982 %6 = bitcast <8 x i1> %5 to i8
5983 ret i8 %6
5984 }
5987 define zeroext i8 @test_vpcmpsgtd_v4i1_v8i1_mask_mem_b(<2 x i64> %__a, i32* %__b) local_unnamed_addr {
5988 ; VLX-LABEL: test_vpcmpsgtd_v4i1_v8i1_mask_mem_b:
5989 ; VLX: # %bb.0: # %entry
5990 ; VLX-NEXT: vpcmpgtd (%rdi){1to4}, %xmm0, %k0
5991 ; VLX-NEXT: kmovd %k0, %eax
5992 ; VLX-NEXT: # kill: def $al killed $al killed $eax
5993 ; VLX-NEXT: retq
5994 ;
5995 ; NoVLX-LABEL: test_vpcmpsgtd_v4i1_v8i1_mask_mem_b:
5996 ; NoVLX: # %bb.0: # %entry
5997 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
5998 ; NoVLX-NEXT: vpcmpgtd (%rdi){1to16}, %zmm0, %k0
5999 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
6000 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
6001 ; NoVLX-NEXT: kmovw %k0, %eax
6002 ; NoVLX-NEXT: # kill: def $al killed $al killed $eax
6003 ; NoVLX-NEXT: vzeroupper
6004 ; NoVLX-NEXT: retq
6005 entry:
6006 %0 = bitcast <2 x i64> %__a to <4 x i32>
6007 %load = load i32, i32* %__b
6008 %vec = insertelement <4 x i32> undef, i32 %load, i32 0
6009 %1 = shufflevector <4 x i32> %vec, <4 x i32> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
6010 %2 = icmp sgt <4 x i32> %0, %1
6011 %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
6012 %4 = bitcast <8 x i1> %3 to i8
6013 ret i8 %4
6014 }
6016 define zeroext i8 @test_masked_vpcmpsgtd_v4i1_v8i1_mask_mem_b(i8 zeroext %__u, <2 x i64> %__a, i32* %__b) local_unnamed_addr {
6017 ; VLX-LABEL: test_masked_vpcmpsgtd_v4i1_v8i1_mask_mem_b:
6018 ; VLX: # %bb.0: # %entry
6019 ; VLX-NEXT: kmovd %edi, %k1
6020 ; VLX-NEXT: vpcmpgtd (%rsi){1to4}, %xmm0, %k0 {%k1}
6021 ; VLX-NEXT: kmovd %k0, %eax
6022 ; VLX-NEXT: # kill: def $al killed $al killed $eax
6023 ; VLX-NEXT: retq
6024 ;
6025 ; NoVLX-LABEL: test_masked_vpcmpsgtd_v4i1_v8i1_mask_mem_b:
6026 ; NoVLX: # %bb.0: # %entry
6027 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
6028 ; NoVLX-NEXT: kmovw %edi, %k1
6029 ; NoVLX-NEXT: vpcmpgtd (%rsi){1to16}, %zmm0, %k0 {%k1}
6030 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
6031 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
6032 ; NoVLX-NEXT: kmovw %k0, %eax
6033 ; NoVLX-NEXT: # kill: def $al killed $al killed $eax
6034 ; NoVLX-NEXT: vzeroupper
6035 ; NoVLX-NEXT: retq
6036 entry:
6037 %0 = bitcast <2 x i64> %__a to <4 x i32>
6038 %load = load i32, i32* %__b
6039 %vec = insertelement <4 x i32> undef, i32 %load, i32 0
6040 %1 = shufflevector <4 x i32> %vec, <4 x i32> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
6041 %2 = icmp sgt <4 x i32> %0, %1
6042 %3 = bitcast i8 %__u to <8 x i1>
6043 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
6044 %4 = and <4 x i1> %extract.i, %2
6045 %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
6046 %6 = bitcast <8 x i1> %5 to i8
6047 ret i8 %6
6048 }
6051 define zeroext i16 @test_vpcmpsgtd_v4i1_v16i1_mask(<2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
6052 ; VLX-LABEL: test_vpcmpsgtd_v4i1_v16i1_mask:
6053 ; VLX: # %bb.0: # %entry
6054 ; VLX-NEXT: vpcmpgtd %xmm1, %xmm0, %k0
6055 ; VLX-NEXT: kmovd %k0, %eax
6056 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
6057 ; VLX-NEXT: retq
6058 ;
6059 ; NoVLX-LABEL: test_vpcmpsgtd_v4i1_v16i1_mask:
6060 ; NoVLX: # %bb.0: # %entry
6061 ; NoVLX-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
6062 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
6063 ; NoVLX-NEXT: vpcmpgtd %zmm1, %zmm0, %k0
6064 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
6065 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
6066 ; NoVLX-NEXT: kmovw %k0, %eax
6067 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
6068 ; NoVLX-NEXT: vzeroupper
6069 ; NoVLX-NEXT: retq
6070 entry:
6071 %0 = bitcast <2 x i64> %__a to <4 x i32>
6072 %1 = bitcast <2 x i64> %__b to <4 x i32>
6073 %2 = icmp sgt <4 x i32> %0, %1
6074 %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
6075 %4 = bitcast <16 x i1> %3 to i16
6076 ret i16 %4
6077 }
6079 define zeroext i16 @test_vpcmpsgtd_v4i1_v16i1_mask_mem(<2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
6080 ; VLX-LABEL: test_vpcmpsgtd_v4i1_v16i1_mask_mem:
6081 ; VLX: # %bb.0: # %entry
6082 ; VLX-NEXT: vpcmpgtd (%rdi), %xmm0, %k0
6083 ; VLX-NEXT: kmovd %k0, %eax
6084 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
6085 ; VLX-NEXT: retq
6086 ;
6087 ; NoVLX-LABEL: test_vpcmpsgtd_v4i1_v16i1_mask_mem:
6088 ; NoVLX: # %bb.0: # %entry
6089 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
6090 ; NoVLX-NEXT: vmovdqa (%rdi), %xmm1
6091 ; NoVLX-NEXT: vpcmpgtd %zmm1, %zmm0, %k0
6092 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
6093 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
6094 ; NoVLX-NEXT: kmovw %k0, %eax
6095 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
6096 ; NoVLX-NEXT: vzeroupper
6097 ; NoVLX-NEXT: retq
6098 entry:
6099 %0 = bitcast <2 x i64> %__a to <4 x i32>
6100 %load = load <2 x i64>, <2 x i64>* %__b
6101 %1 = bitcast <2 x i64> %load to <4 x i32>
6102 %2 = icmp sgt <4 x i32> %0, %1
6103 %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
6104 %4 = bitcast <16 x i1> %3 to i16
6105 ret i16 %4
6106 }
6108 define zeroext i16 @test_masked_vpcmpsgtd_v4i1_v16i1_mask(i8 zeroext %__u, <2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
6109 ; VLX-LABEL: test_masked_vpcmpsgtd_v4i1_v16i1_mask:
6110 ; VLX: # %bb.0: # %entry
6111 ; VLX-NEXT: kmovd %edi, %k1
6112 ; VLX-NEXT: vpcmpgtd %xmm1, %xmm0, %k0 {%k1}
6113 ; VLX-NEXT: kmovd %k0, %eax
6114 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
6115 ; VLX-NEXT: retq
6116 ;
6117 ; NoVLX-LABEL: test_masked_vpcmpsgtd_v4i1_v16i1_mask:
6118 ; NoVLX: # %bb.0: # %entry
6119 ; NoVLX-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
6120 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
6121 ; NoVLX-NEXT: kmovw %edi, %k1
6122 ; NoVLX-NEXT: vpcmpgtd %zmm1, %zmm0, %k0 {%k1}
6123 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
6124 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
6125 ; NoVLX-NEXT: kmovw %k0, %eax
6126 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
6127 ; NoVLX-NEXT: vzeroupper
6128 ; NoVLX-NEXT: retq
6129 entry:
6130 %0 = bitcast <2 x i64> %__a to <4 x i32>
6131 %1 = bitcast <2 x i64> %__b to <4 x i32>
6132 %2 = icmp sgt <4 x i32> %0, %1
6133 %3 = bitcast i8 %__u to <8 x i1>
6134 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
6135 %4 = and <4 x i1> %2, %extract.i
6136 %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
6137 %6 = bitcast <16 x i1> %5 to i16
6138 ret i16 %6
6139 }
6141 define zeroext i16 @test_masked_vpcmpsgtd_v4i1_v16i1_mask_mem(i8 zeroext %__u, <2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
6142 ; VLX-LABEL: test_masked_vpcmpsgtd_v4i1_v16i1_mask_mem:
6143 ; VLX: # %bb.0: # %entry
6144 ; VLX-NEXT: kmovd %edi, %k1
6145 ; VLX-NEXT: vpcmpgtd (%rsi), %xmm0, %k0 {%k1}
6146 ; VLX-NEXT: kmovd %k0, %eax
6147 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
6148 ; VLX-NEXT: retq
6149 ;
6150 ; NoVLX-LABEL: test_masked_vpcmpsgtd_v4i1_v16i1_mask_mem:
6151 ; NoVLX: # %bb.0: # %entry
6152 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
6153 ; NoVLX-NEXT: vmovdqa (%rsi), %xmm1
6154 ; NoVLX-NEXT: kmovw %edi, %k1
6155 ; NoVLX-NEXT: vpcmpgtd %zmm1, %zmm0, %k0 {%k1}
6156 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
6157 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
6158 ; NoVLX-NEXT: kmovw %k0, %eax
6159 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
6160 ; NoVLX-NEXT: vzeroupper
6161 ; NoVLX-NEXT: retq
6162 entry:
6163 %0 = bitcast <2 x i64> %__a to <4 x i32>
6164 %load = load <2 x i64>, <2 x i64>* %__b
6165 %1 = bitcast <2 x i64> %load to <4 x i32>
6166 %2 = icmp sgt <4 x i32> %0, %1
6167 %3 = bitcast i8 %__u to <8 x i1>
6168 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
6169 %4 = and <4 x i1> %2, %extract.i
6170 %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
6171 %6 = bitcast <16 x i1> %5 to i16
6172 ret i16 %6
6173 }
6176 define zeroext i16 @test_vpcmpsgtd_v4i1_v16i1_mask_mem_b(<2 x i64> %__a, i32* %__b) local_unnamed_addr {
6177 ; VLX-LABEL: test_vpcmpsgtd_v4i1_v16i1_mask_mem_b:
6178 ; VLX: # %bb.0: # %entry
6179 ; VLX-NEXT: vpcmpgtd (%rdi){1to4}, %xmm0, %k0
6180 ; VLX-NEXT: kmovd %k0, %eax
6181 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
6182 ; VLX-NEXT: retq
6183 ;
6184 ; NoVLX-LABEL: test_vpcmpsgtd_v4i1_v16i1_mask_mem_b:
6185 ; NoVLX: # %bb.0: # %entry
6186 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
6187 ; NoVLX-NEXT: vpcmpgtd (%rdi){1to16}, %zmm0, %k0
6188 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
6189 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
6190 ; NoVLX-NEXT: kmovw %k0, %eax
6191 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
6192 ; NoVLX-NEXT: vzeroupper
6193 ; NoVLX-NEXT: retq
6194 entry:
6195 %0 = bitcast <2 x i64> %__a to <4 x i32>
6196 %load = load i32, i32* %__b
6197 %vec = insertelement <4 x i32> undef, i32 %load, i32 0
6198 %1 = shufflevector <4 x i32> %vec, <4 x i32> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
6199 %2 = icmp sgt <4 x i32> %0, %1
6200 %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
6201 %4 = bitcast <16 x i1> %3 to i16
6202 ret i16 %4
6203 }
6205 define zeroext i16 @test_masked_vpcmpsgtd_v4i1_v16i1_mask_mem_b(i8 zeroext %__u, <2 x i64> %__a, i32* %__b) local_unnamed_addr {
6206 ; VLX-LABEL: test_masked_vpcmpsgtd_v4i1_v16i1_mask_mem_b:
6207 ; VLX: # %bb.0: # %entry
6208 ; VLX-NEXT: kmovd %edi, %k1
6209 ; VLX-NEXT: vpcmpgtd (%rsi){1to4}, %xmm0, %k0 {%k1}
6210 ; VLX-NEXT: kmovd %k0, %eax
6211 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
6212 ; VLX-NEXT: retq
6213 ;
6214 ; NoVLX-LABEL: test_masked_vpcmpsgtd_v4i1_v16i1_mask_mem_b:
6215 ; NoVLX: # %bb.0: # %entry
6216 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
6217 ; NoVLX-NEXT: kmovw %edi, %k1
6218 ; NoVLX-NEXT: vpcmpgtd (%rsi){1to16}, %zmm0, %k0 {%k1}
6219 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
6220 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
6221 ; NoVLX-NEXT: kmovw %k0, %eax
6222 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
6223 ; NoVLX-NEXT: vzeroupper
6224 ; NoVLX-NEXT: retq
6225 entry:
6226 %0 = bitcast <2 x i64> %__a to <4 x i32>
6227 %load = load i32, i32* %__b
6228 %vec = insertelement <4 x i32> undef, i32 %load, i32 0
6229 %1 = shufflevector <4 x i32> %vec, <4 x i32> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
6230 %2 = icmp sgt <4 x i32> %0, %1
6231 %3 = bitcast i8 %__u to <8 x i1>
6232 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
6233 %4 = and <4 x i1> %extract.i, %2
6234 %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
6235 %6 = bitcast <16 x i1> %5 to i16
6236 ret i16 %6
6237 }
6240 define zeroext i32 @test_vpcmpsgtd_v4i1_v32i1_mask(<2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
6241 ; VLX-LABEL: test_vpcmpsgtd_v4i1_v32i1_mask:
6242 ; VLX: # %bb.0: # %entry
6243 ; VLX-NEXT: vpcmpgtd %xmm1, %xmm0, %k0
6244 ; VLX-NEXT: kmovd %k0, %eax
6245 ; VLX-NEXT: retq
6246 ;
6247 ; NoVLX-LABEL: test_vpcmpsgtd_v4i1_v32i1_mask:
6248 ; NoVLX: # %bb.0: # %entry
6249 ; NoVLX-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
6250 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
6251 ; NoVLX-NEXT: vpcmpgtd %zmm1, %zmm0, %k0
6252 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
6253 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
6254 ; NoVLX-NEXT: kmovw %k0, %eax
6255 ; NoVLX-NEXT: vzeroupper
6256 ; NoVLX-NEXT: retq
6257 entry:
6258 %0 = bitcast <2 x i64> %__a to <4 x i32>
6259 %1 = bitcast <2 x i64> %__b to <4 x i32>
6260 %2 = icmp sgt <4 x i32> %0, %1
6261 %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
6262 %4 = bitcast <32 x i1> %3 to i32
6263 ret i32 %4
6264 }
6266 define zeroext i32 @test_vpcmpsgtd_v4i1_v32i1_mask_mem(<2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
6267 ; VLX-LABEL: test_vpcmpsgtd_v4i1_v32i1_mask_mem:
6268 ; VLX: # %bb.0: # %entry
6269 ; VLX-NEXT: vpcmpgtd (%rdi), %xmm0, %k0
6270 ; VLX-NEXT: kmovd %k0, %eax
6271 ; VLX-NEXT: retq
6272 ;
6273 ; NoVLX-LABEL: test_vpcmpsgtd_v4i1_v32i1_mask_mem:
6274 ; NoVLX: # %bb.0: # %entry
6275 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
6276 ; NoVLX-NEXT: vmovdqa (%rdi), %xmm1
6277 ; NoVLX-NEXT: vpcmpgtd %zmm1, %zmm0, %k0
6278 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
6279 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
6280 ; NoVLX-NEXT: kmovw %k0, %eax
6281 ; NoVLX-NEXT: vzeroupper
6282 ; NoVLX-NEXT: retq
6283 entry:
6284 %0 = bitcast <2 x i64> %__a to <4 x i32>
6285 %load = load <2 x i64>, <2 x i64>* %__b
6286 %1 = bitcast <2 x i64> %load to <4 x i32>
6287 %2 = icmp sgt <4 x i32> %0, %1
6288 %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
6289 %4 = bitcast <32 x i1> %3 to i32
6290 ret i32 %4
6291 }
6293 define zeroext i32 @test_masked_vpcmpsgtd_v4i1_v32i1_mask(i8 zeroext %__u, <2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
6294 ; VLX-LABEL: test_masked_vpcmpsgtd_v4i1_v32i1_mask:
6295 ; VLX: # %bb.0: # %entry
6296 ; VLX-NEXT: kmovd %edi, %k1
6297 ; VLX-NEXT: vpcmpgtd %xmm1, %xmm0, %k0 {%k1}
6298 ; VLX-NEXT: kmovd %k0, %eax
6299 ; VLX-NEXT: retq
6300 ;
6301 ; NoVLX-LABEL: test_masked_vpcmpsgtd_v4i1_v32i1_mask:
6302 ; NoVLX: # %bb.0: # %entry
6303 ; NoVLX-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
6304 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
6305 ; NoVLX-NEXT: kmovw %edi, %k1
6306 ; NoVLX-NEXT: vpcmpgtd %zmm1, %zmm0, %k0 {%k1}
6307 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
6308 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
6309 ; NoVLX-NEXT: kmovw %k0, %eax
6310 ; NoVLX-NEXT: vzeroupper
6311 ; NoVLX-NEXT: retq
6312 entry:
6313 %0 = bitcast <2 x i64> %__a to <4 x i32>
6314 %1 = bitcast <2 x i64> %__b to <4 x i32>
6315 %2 = icmp sgt <4 x i32> %0, %1
6316 %3 = bitcast i8 %__u to <8 x i1>
6317 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
6318 %4 = and <4 x i1> %2, %extract.i
6319 %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
6320 %6 = bitcast <32 x i1> %5 to i32
6321 ret i32 %6
6322 }
6324 define zeroext i32 @test_masked_vpcmpsgtd_v4i1_v32i1_mask_mem(i8 zeroext %__u, <2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
6325 ; VLX-LABEL: test_masked_vpcmpsgtd_v4i1_v32i1_mask_mem:
6326 ; VLX: # %bb.0: # %entry
6327 ; VLX-NEXT: kmovd %edi, %k1
6328 ; VLX-NEXT: vpcmpgtd (%rsi), %xmm0, %k0 {%k1}
6329 ; VLX-NEXT: kmovd %k0, %eax
6330 ; VLX-NEXT: retq
6331 ;
6332 ; NoVLX-LABEL: test_masked_vpcmpsgtd_v4i1_v32i1_mask_mem:
6333 ; NoVLX: # %bb.0: # %entry
6334 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
6335 ; NoVLX-NEXT: vmovdqa (%rsi), %xmm1
6336 ; NoVLX-NEXT: kmovw %edi, %k1
6337 ; NoVLX-NEXT: vpcmpgtd %zmm1, %zmm0, %k0 {%k1}
6338 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
6339 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
6340 ; NoVLX-NEXT: kmovw %k0, %eax
6341 ; NoVLX-NEXT: vzeroupper
6342 ; NoVLX-NEXT: retq
6343 entry:
6344 %0 = bitcast <2 x i64> %__a to <4 x i32>
6345 %load = load <2 x i64>, <2 x i64>* %__b
6346 %1 = bitcast <2 x i64> %load to <4 x i32>
6347 %2 = icmp sgt <4 x i32> %0, %1
6348 %3 = bitcast i8 %__u to <8 x i1>
6349 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
6350 %4 = and <4 x i1> %2, %extract.i
6351 %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
6352 %6 = bitcast <32 x i1> %5 to i32
6353 ret i32 %6
6354 }
6357 define zeroext i32 @test_vpcmpsgtd_v4i1_v32i1_mask_mem_b(<2 x i64> %__a, i32* %__b) local_unnamed_addr {
6358 ; VLX-LABEL: test_vpcmpsgtd_v4i1_v32i1_mask_mem_b:
6359 ; VLX: # %bb.0: # %entry
6360 ; VLX-NEXT: vpcmpgtd (%rdi){1to4}, %xmm0, %k0
6361 ; VLX-NEXT: kmovd %k0, %eax
6362 ; VLX-NEXT: retq
6363 ;
6364 ; NoVLX-LABEL: test_vpcmpsgtd_v4i1_v32i1_mask_mem_b:
6365 ; NoVLX: # %bb.0: # %entry
6366 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
6367 ; NoVLX-NEXT: vpcmpgtd (%rdi){1to16}, %zmm0, %k0
6368 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
6369 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
6370 ; NoVLX-NEXT: kmovw %k0, %eax
6371 ; NoVLX-NEXT: vzeroupper
6372 ; NoVLX-NEXT: retq
6373 entry:
6374 %0 = bitcast <2 x i64> %__a to <4 x i32>
6375 %load = load i32, i32* %__b
6376 %vec = insertelement <4 x i32> undef, i32 %load, i32 0
6377 %1 = shufflevector <4 x i32> %vec, <4 x i32> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
6378 %2 = icmp sgt <4 x i32> %0, %1
6379 %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
6380 %4 = bitcast <32 x i1> %3 to i32
6381 ret i32 %4
6382 }
6384 define zeroext i32 @test_masked_vpcmpsgtd_v4i1_v32i1_mask_mem_b(i8 zeroext %__u, <2 x i64> %__a, i32* %__b) local_unnamed_addr {
6385 ; VLX-LABEL: test_masked_vpcmpsgtd_v4i1_v32i1_mask_mem_b:
6386 ; VLX: # %bb.0: # %entry
6387 ; VLX-NEXT: kmovd %edi, %k1
6388 ; VLX-NEXT: vpcmpgtd (%rsi){1to4}, %xmm0, %k0 {%k1}
6389 ; VLX-NEXT: kmovd %k0, %eax
6390 ; VLX-NEXT: retq
6391 ;
6392 ; NoVLX-LABEL: test_masked_vpcmpsgtd_v4i1_v32i1_mask_mem_b:
6393 ; NoVLX: # %bb.0: # %entry
6394 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
6395 ; NoVLX-NEXT: kmovw %edi, %k1
6396 ; NoVLX-NEXT: vpcmpgtd (%rsi){1to16}, %zmm0, %k0 {%k1}
6397 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
6398 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
6399 ; NoVLX-NEXT: kmovw %k0, %eax
6400 ; NoVLX-NEXT: vzeroupper
6401 ; NoVLX-NEXT: retq
6402 entry:
6403 %0 = bitcast <2 x i64> %__a to <4 x i32>
6404 %load = load i32, i32* %__b
6405 %vec = insertelement <4 x i32> undef, i32 %load, i32 0
6406 %1 = shufflevector <4 x i32> %vec, <4 x i32> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
6407 %2 = icmp sgt <4 x i32> %0, %1
6408 %3 = bitcast i8 %__u to <8 x i1>
6409 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
6410 %4 = and <4 x i1> %extract.i, %2
6411 %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
6412 %6 = bitcast <32 x i1> %5 to i32
6413 ret i32 %6
6414 }
6417 define zeroext i64 @test_vpcmpsgtd_v4i1_v64i1_mask(<2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
6418 ; VLX-LABEL: test_vpcmpsgtd_v4i1_v64i1_mask:
6419 ; VLX: # %bb.0: # %entry
6420 ; VLX-NEXT: vpcmpgtd %xmm1, %xmm0, %k0
6421 ; VLX-NEXT: kmovq %k0, %rax
6422 ; VLX-NEXT: retq
6423 ;
6424 ; NoVLX-LABEL: test_vpcmpsgtd_v4i1_v64i1_mask:
6425 ; NoVLX: # %bb.0: # %entry
6426 ; NoVLX-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
6427 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
6428 ; NoVLX-NEXT: vpcmpgtd %zmm1, %zmm0, %k0
6429 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
6430 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
6431 ; NoVLX-NEXT: kmovw %k0, %eax
6432 ; NoVLX-NEXT: vzeroupper
6433 ; NoVLX-NEXT: retq
6434 entry:
6435 %0 = bitcast <2 x i64> %__a to <4 x i32>
6436 %1 = bitcast <2 x i64> %__b to <4 x i32>
6437 %2 = icmp sgt <4 x i32> %0, %1
6438 %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
6439 %4 = bitcast <64 x i1> %3 to i64
6440 ret i64 %4
6441 }
6443 define zeroext i64 @test_vpcmpsgtd_v4i1_v64i1_mask_mem(<2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
6444 ; VLX-LABEL: test_vpcmpsgtd_v4i1_v64i1_mask_mem:
6445 ; VLX: # %bb.0: # %entry
6446 ; VLX-NEXT: vpcmpgtd (%rdi), %xmm0, %k0
6447 ; VLX-NEXT: kmovq %k0, %rax
6448 ; VLX-NEXT: retq
6449 ;
6450 ; NoVLX-LABEL: test_vpcmpsgtd_v4i1_v64i1_mask_mem:
6451 ; NoVLX: # %bb.0: # %entry
6452 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
6453 ; NoVLX-NEXT: vmovdqa (%rdi), %xmm1
6454 ; NoVLX-NEXT: vpcmpgtd %zmm1, %zmm0, %k0
6455 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
6456 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
6457 ; NoVLX-NEXT: kmovw %k0, %eax
6458 ; NoVLX-NEXT: vzeroupper
6459 ; NoVLX-NEXT: retq
6460 entry:
6461 %0 = bitcast <2 x i64> %__a to <4 x i32>
6462 %load = load <2 x i64>, <2 x i64>* %__b
6463 %1 = bitcast <2 x i64> %load to <4 x i32>
6464 %2 = icmp sgt <4 x i32> %0, %1
6465 %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
6466 %4 = bitcast <64 x i1> %3 to i64
6467 ret i64 %4
6468 }
6470 define zeroext i64 @test_masked_vpcmpsgtd_v4i1_v64i1_mask(i8 zeroext %__u, <2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
6471 ; VLX-LABEL: test_masked_vpcmpsgtd_v4i1_v64i1_mask:
6472 ; VLX: # %bb.0: # %entry
6473 ; VLX-NEXT: kmovd %edi, %k1
6474 ; VLX-NEXT: vpcmpgtd %xmm1, %xmm0, %k0 {%k1}
6475 ; VLX-NEXT: kmovq %k0, %rax
6476 ; VLX-NEXT: retq
6477 ;
6478 ; NoVLX-LABEL: test_masked_vpcmpsgtd_v4i1_v64i1_mask:
6479 ; NoVLX: # %bb.0: # %entry
6480 ; NoVLX-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
6481 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
6482 ; NoVLX-NEXT: kmovw %edi, %k1
6483 ; NoVLX-NEXT: vpcmpgtd %zmm1, %zmm0, %k0 {%k1}
6484 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
6485 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
6486 ; NoVLX-NEXT: kmovw %k0, %eax
6487 ; NoVLX-NEXT: vzeroupper
6488 ; NoVLX-NEXT: retq
6489 entry:
6490 %0 = bitcast <2 x i64> %__a to <4 x i32>
6491 %1 = bitcast <2 x i64> %__b to <4 x i32>
6492 %2 = icmp sgt <4 x i32> %0, %1
6493 %3 = bitcast i8 %__u to <8 x i1>
6494 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
6495 %4 = and <4 x i1> %2, %extract.i
6496 %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
6497 %6 = bitcast <64 x i1> %5 to i64
6498 ret i64 %6
6499 }
6501 define zeroext i64 @test_masked_vpcmpsgtd_v4i1_v64i1_mask_mem(i8 zeroext %__u, <2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
6502 ; VLX-LABEL: test_masked_vpcmpsgtd_v4i1_v64i1_mask_mem:
6503 ; VLX: # %bb.0: # %entry
6504 ; VLX-NEXT: kmovd %edi, %k1
6505 ; VLX-NEXT: vpcmpgtd (%rsi), %xmm0, %k0 {%k1}
6506 ; VLX-NEXT: kmovq %k0, %rax
6507 ; VLX-NEXT: retq
6508 ;
6509 ; NoVLX-LABEL: test_masked_vpcmpsgtd_v4i1_v64i1_mask_mem:
6510 ; NoVLX: # %bb.0: # %entry
6511 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
6512 ; NoVLX-NEXT: vmovdqa (%rsi), %xmm1
6513 ; NoVLX-NEXT: kmovw %edi, %k1
6514 ; NoVLX-NEXT: vpcmpgtd %zmm1, %zmm0, %k0 {%k1}
6515 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
6516 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
6517 ; NoVLX-NEXT: kmovw %k0, %eax
6518 ; NoVLX-NEXT: vzeroupper
6519 ; NoVLX-NEXT: retq
6520 entry:
6521 %0 = bitcast <2 x i64> %__a to <4 x i32>
6522 %load = load <2 x i64>, <2 x i64>* %__b
6523 %1 = bitcast <2 x i64> %load to <4 x i32>
6524 %2 = icmp sgt <4 x i32> %0, %1
6525 %3 = bitcast i8 %__u to <8 x i1>
6526 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
6527 %4 = and <4 x i1> %2, %extract.i
6528 %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
6529 %6 = bitcast <64 x i1> %5 to i64
6530 ret i64 %6
6531 }
6534 define zeroext i64 @test_vpcmpsgtd_v4i1_v64i1_mask_mem_b(<2 x i64> %__a, i32* %__b) local_unnamed_addr {
6535 ; VLX-LABEL: test_vpcmpsgtd_v4i1_v64i1_mask_mem_b:
6536 ; VLX: # %bb.0: # %entry
6537 ; VLX-NEXT: vpcmpgtd (%rdi){1to4}, %xmm0, %k0
6538 ; VLX-NEXT: kmovq %k0, %rax
6539 ; VLX-NEXT: retq
6540 ;
6541 ; NoVLX-LABEL: test_vpcmpsgtd_v4i1_v64i1_mask_mem_b:
6542 ; NoVLX: # %bb.0: # %entry
6543 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
6544 ; NoVLX-NEXT: vpcmpgtd (%rdi){1to16}, %zmm0, %k0
6545 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
6546 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
6547 ; NoVLX-NEXT: kmovw %k0, %eax
6548 ; NoVLX-NEXT: vzeroupper
6549 ; NoVLX-NEXT: retq
6550 entry:
6551 %0 = bitcast <2 x i64> %__a to <4 x i32>
6552 %load = load i32, i32* %__b
6553 %vec = insertelement <4 x i32> undef, i32 %load, i32 0
6554 %1 = shufflevector <4 x i32> %vec, <4 x i32> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
6555 %2 = icmp sgt <4 x i32> %0, %1
6556 %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
6557 %4 = bitcast <64 x i1> %3 to i64
6558 ret i64 %4
6559 }
6561 define zeroext i64 @test_masked_vpcmpsgtd_v4i1_v64i1_mask_mem_b(i8 zeroext %__u, <2 x i64> %__a, i32* %__b) local_unnamed_addr {
6562 ; VLX-LABEL: test_masked_vpcmpsgtd_v4i1_v64i1_mask_mem_b:
6563 ; VLX: # %bb.0: # %entry
6564 ; VLX-NEXT: kmovd %edi, %k1
6565 ; VLX-NEXT: vpcmpgtd (%rsi){1to4}, %xmm0, %k0 {%k1}
6566 ; VLX-NEXT: kmovq %k0, %rax
6567 ; VLX-NEXT: retq
6568 ;
6569 ; NoVLX-LABEL: test_masked_vpcmpsgtd_v4i1_v64i1_mask_mem_b:
6570 ; NoVLX: # %bb.0: # %entry
6571 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
6572 ; NoVLX-NEXT: kmovw %edi, %k1
6573 ; NoVLX-NEXT: vpcmpgtd (%rsi){1to16}, %zmm0, %k0 {%k1}
6574 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
6575 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
6576 ; NoVLX-NEXT: kmovw %k0, %eax
6577 ; NoVLX-NEXT: vzeroupper
6578 ; NoVLX-NEXT: retq
6579 entry:
6580 %0 = bitcast <2 x i64> %__a to <4 x i32>
6581 %load = load i32, i32* %__b
6582 %vec = insertelement <4 x i32> undef, i32 %load, i32 0
6583 %1 = shufflevector <4 x i32> %vec, <4 x i32> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
6584 %2 = icmp sgt <4 x i32> %0, %1
6585 %3 = bitcast i8 %__u to <8 x i1>
6586 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
6587 %4 = and <4 x i1> %extract.i, %2
6588 %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
6589 %6 = bitcast <64 x i1> %5 to i64
6590 ret i64 %6
6591 }
6594 define zeroext i16 @test_vpcmpsgtd_v8i1_v16i1_mask(<4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
6595 ; VLX-LABEL: test_vpcmpsgtd_v8i1_v16i1_mask:
6596 ; VLX: # %bb.0: # %entry
6597 ; VLX-NEXT: vpcmpgtd %ymm1, %ymm0, %k0
6598 ; VLX-NEXT: kmovd %k0, %eax
6599 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
6600 ; VLX-NEXT: vzeroupper
6601 ; VLX-NEXT: retq
6602 ;
6603 ; NoVLX-LABEL: test_vpcmpsgtd_v8i1_v16i1_mask:
6604 ; NoVLX: # %bb.0: # %entry
6605 ; NoVLX-NEXT: # kill: def $ymm1 killed $ymm1 def $zmm1
6606 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
6607 ; NoVLX-NEXT: vpcmpgtd %zmm1, %zmm0, %k0
6608 ; NoVLX-NEXT: kshiftlw $8, %k0, %k0
6609 ; NoVLX-NEXT: kshiftrw $8, %k0, %k0
6610 ; NoVLX-NEXT: kmovw %k0, %eax
6611 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
6612 ; NoVLX-NEXT: vzeroupper
6613 ; NoVLX-NEXT: retq
6614 entry:
6615 %0 = bitcast <4 x i64> %__a to <8 x i32>
6616 %1 = bitcast <4 x i64> %__b to <8 x i32>
6617 %2 = icmp sgt <8 x i32> %0, %1
6618 %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
6619 %4 = bitcast <16 x i1> %3 to i16
6620 ret i16 %4
6621 }
6623 define zeroext i16 @test_vpcmpsgtd_v8i1_v16i1_mask_mem(<4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
6624 ; VLX-LABEL: test_vpcmpsgtd_v8i1_v16i1_mask_mem:
6625 ; VLX: # %bb.0: # %entry
6626 ; VLX-NEXT: vpcmpgtd (%rdi), %ymm0, %k0
6627 ; VLX-NEXT: kmovd %k0, %eax
6628 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
6629 ; VLX-NEXT: vzeroupper
6630 ; VLX-NEXT: retq
6631 ;
6632 ; NoVLX-LABEL: test_vpcmpsgtd_v8i1_v16i1_mask_mem:
6633 ; NoVLX: # %bb.0: # %entry
6634 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
6635 ; NoVLX-NEXT: vmovdqa (%rdi), %ymm1
6636 ; NoVLX-NEXT: vpcmpgtd %zmm1, %zmm0, %k0
6637 ; NoVLX-NEXT: kshiftlw $8, %k0, %k0
6638 ; NoVLX-NEXT: kshiftrw $8, %k0, %k0
6639 ; NoVLX-NEXT: kmovw %k0, %eax
6640 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
6641 ; NoVLX-NEXT: vzeroupper
6642 ; NoVLX-NEXT: retq
6643 entry:
6644 %0 = bitcast <4 x i64> %__a to <8 x i32>
6645 %load = load <4 x i64>, <4 x i64>* %__b
6646 %1 = bitcast <4 x i64> %load to <8 x i32>
6647 %2 = icmp sgt <8 x i32> %0, %1
6648 %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
6649 %4 = bitcast <16 x i1> %3 to i16
6650 ret i16 %4
6651 }
6653 define zeroext i16 @test_masked_vpcmpsgtd_v8i1_v16i1_mask(i8 zeroext %__u, <4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
6654 ; VLX-LABEL: test_masked_vpcmpsgtd_v8i1_v16i1_mask:
6655 ; VLX: # %bb.0: # %entry
6656 ; VLX-NEXT: kmovd %edi, %k1
6657 ; VLX-NEXT: vpcmpgtd %ymm1, %ymm0, %k0 {%k1}
6658 ; VLX-NEXT: kmovd %k0, %eax
6659 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
6660 ; VLX-NEXT: vzeroupper
6661 ; VLX-NEXT: retq
6662 ;
6663 ; NoVLX-LABEL: test_masked_vpcmpsgtd_v8i1_v16i1_mask:
6664 ; NoVLX: # %bb.0: # %entry
6665 ; NoVLX-NEXT: # kill: def $ymm1 killed $ymm1 def $zmm1
6666 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
6667 ; NoVLX-NEXT: kmovw %edi, %k1
6668 ; NoVLX-NEXT: vpcmpgtd %zmm1, %zmm0, %k0 {%k1}
6669 ; NoVLX-NEXT: kshiftlw $8, %k0, %k0
6670 ; NoVLX-NEXT: kshiftrw $8, %k0, %k0
6671 ; NoVLX-NEXT: kmovw %k0, %eax
6672 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
6673 ; NoVLX-NEXT: vzeroupper
6674 ; NoVLX-NEXT: retq
6675 entry:
6676 %0 = bitcast <4 x i64> %__a to <8 x i32>
6677 %1 = bitcast <4 x i64> %__b to <8 x i32>
6678 %2 = icmp sgt <8 x i32> %0, %1
6679 %3 = bitcast i8 %__u to <8 x i1>
6680 %4 = and <8 x i1> %2, %3
6681 %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
6682 %6 = bitcast <16 x i1> %5 to i16
6683 ret i16 %6
6684 }
6686 define zeroext i16 @test_masked_vpcmpsgtd_v8i1_v16i1_mask_mem(i8 zeroext %__u, <4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
6687 ; VLX-LABEL: test_masked_vpcmpsgtd_v8i1_v16i1_mask_mem:
6688 ; VLX: # %bb.0: # %entry
6689 ; VLX-NEXT: kmovd %edi, %k1
6690 ; VLX-NEXT: vpcmpgtd (%rsi), %ymm0, %k0 {%k1}
6691 ; VLX-NEXT: kmovd %k0, %eax
6692 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
6693 ; VLX-NEXT: vzeroupper
6694 ; VLX-NEXT: retq
6695 ;
6696 ; NoVLX-LABEL: test_masked_vpcmpsgtd_v8i1_v16i1_mask_mem:
6697 ; NoVLX: # %bb.0: # %entry
6698 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
6699 ; NoVLX-NEXT: vmovdqa (%rsi), %ymm1
6700 ; NoVLX-NEXT: kmovw %edi, %k1
6701 ; NoVLX-NEXT: vpcmpgtd %zmm1, %zmm0, %k0 {%k1}
6702 ; NoVLX-NEXT: kshiftlw $8, %k0, %k0
6703 ; NoVLX-NEXT: kshiftrw $8, %k0, %k0
6704 ; NoVLX-NEXT: kmovw %k0, %eax
6705 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
6706 ; NoVLX-NEXT: vzeroupper
6707 ; NoVLX-NEXT: retq
6708 entry:
6709 %0 = bitcast <4 x i64> %__a to <8 x i32>
6710 %load = load <4 x i64>, <4 x i64>* %__b
6711 %1 = bitcast <4 x i64> %load to <8 x i32>
6712 %2 = icmp sgt <8 x i32> %0, %1
6713 %3 = bitcast i8 %__u to <8 x i1>
6714 %4 = and <8 x i1> %2, %3
6715 %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
6716 %6 = bitcast <16 x i1> %5 to i16
6717 ret i16 %6
6718 }
6721 define zeroext i16 @test_vpcmpsgtd_v8i1_v16i1_mask_mem_b(<4 x i64> %__a, i32* %__b) local_unnamed_addr {
6722 ; VLX-LABEL: test_vpcmpsgtd_v8i1_v16i1_mask_mem_b:
6723 ; VLX: # %bb.0: # %entry
6724 ; VLX-NEXT: vpcmpgtd (%rdi){1to8}, %ymm0, %k0
6725 ; VLX-NEXT: kmovd %k0, %eax
6726 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
6727 ; VLX-NEXT: vzeroupper
6728 ; VLX-NEXT: retq
6729 ;
6730 ; NoVLX-LABEL: test_vpcmpsgtd_v8i1_v16i1_mask_mem_b:
6731 ; NoVLX: # %bb.0: # %entry
6732 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
6733 ; NoVLX-NEXT: vpcmpgtd (%rdi){1to16}, %zmm0, %k0
6734 ; NoVLX-NEXT: kshiftlw $8, %k0, %k0
6735 ; NoVLX-NEXT: kshiftrw $8, %k0, %k0
6736 ; NoVLX-NEXT: kmovw %k0, %eax
6737 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
6738 ; NoVLX-NEXT: vzeroupper
6739 ; NoVLX-NEXT: retq
6740 entry:
6741 %0 = bitcast <4 x i64> %__a to <8 x i32>
6742 %load = load i32, i32* %__b
6743 %vec = insertelement <8 x i32> undef, i32 %load, i32 0
6744 %1 = shufflevector <8 x i32> %vec, <8 x i32> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
6745 %2 = icmp sgt <8 x i32> %0, %1
6746 %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
6747 %4 = bitcast <16 x i1> %3 to i16
6748 ret i16 %4
6749 }
6751 define zeroext i16 @test_masked_vpcmpsgtd_v8i1_v16i1_mask_mem_b(i8 zeroext %__u, <4 x i64> %__a, i32* %__b) local_unnamed_addr {
6752 ; VLX-LABEL: test_masked_vpcmpsgtd_v8i1_v16i1_mask_mem_b:
6753 ; VLX: # %bb.0: # %entry
6754 ; VLX-NEXT: kmovd %edi, %k1
6755 ; VLX-NEXT: vpcmpgtd (%rsi){1to8}, %ymm0, %k0 {%k1}
6756 ; VLX-NEXT: kmovd %k0, %eax
6757 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
6758 ; VLX-NEXT: vzeroupper
6759 ; VLX-NEXT: retq
6760 ;
6761 ; NoVLX-LABEL: test_masked_vpcmpsgtd_v8i1_v16i1_mask_mem_b:
6762 ; NoVLX: # %bb.0: # %entry
6763 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
6764 ; NoVLX-NEXT: kmovw %edi, %k1
6765 ; NoVLX-NEXT: vpcmpgtd (%rsi){1to16}, %zmm0, %k0 {%k1}
6766 ; NoVLX-NEXT: kshiftlw $8, %k0, %k0
6767 ; NoVLX-NEXT: kshiftrw $8, %k0, %k0
6768 ; NoVLX-NEXT: kmovw %k0, %eax
6769 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
6770 ; NoVLX-NEXT: vzeroupper
6771 ; NoVLX-NEXT: retq
6772 entry:
6773 %0 = bitcast <4 x i64> %__a to <8 x i32>
6774 %load = load i32, i32* %__b
6775 %vec = insertelement <8 x i32> undef, i32 %load, i32 0
6776 %1 = shufflevector <8 x i32> %vec, <8 x i32> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
6777 %2 = icmp sgt <8 x i32> %0, %1
6778 %3 = bitcast i8 %__u to <8 x i1>
6779 %4 = and <8 x i1> %3, %2
6780 %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
6781 %6 = bitcast <16 x i1> %5 to i16
6782 ret i16 %6
6783 }
6786 define zeroext i32 @test_vpcmpsgtd_v8i1_v32i1_mask(<4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
6787 ; VLX-LABEL: test_vpcmpsgtd_v8i1_v32i1_mask:
6788 ; VLX: # %bb.0: # %entry
6789 ; VLX-NEXT: vpcmpgtd %ymm1, %ymm0, %k0
6790 ; VLX-NEXT: kmovd %k0, %eax
6791 ; VLX-NEXT: vzeroupper
6792 ; VLX-NEXT: retq
6793 ;
6794 ; NoVLX-LABEL: test_vpcmpsgtd_v8i1_v32i1_mask:
6795 ; NoVLX: # %bb.0: # %entry
6796 ; NoVLX-NEXT: # kill: def $ymm1 killed $ymm1 def $zmm1
6797 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
6798 ; NoVLX-NEXT: vpcmpgtd %zmm1, %zmm0, %k0
6799 ; NoVLX-NEXT: kshiftlw $8, %k0, %k0
6800 ; NoVLX-NEXT: kshiftrw $8, %k0, %k0
6801 ; NoVLX-NEXT: kmovw %k0, %eax
6802 ; NoVLX-NEXT: vzeroupper
6803 ; NoVLX-NEXT: retq
6804 entry:
6805 %0 = bitcast <4 x i64> %__a to <8 x i32>
6806 %1 = bitcast <4 x i64> %__b to <8 x i32>
6807 %2 = icmp sgt <8 x i32> %0, %1
6808 %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
6809 %4 = bitcast <32 x i1> %3 to i32
6810 ret i32 %4
6811 }
6813 define zeroext i32 @test_vpcmpsgtd_v8i1_v32i1_mask_mem(<4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
6814 ; VLX-LABEL: test_vpcmpsgtd_v8i1_v32i1_mask_mem:
6815 ; VLX: # %bb.0: # %entry
6816 ; VLX-NEXT: vpcmpgtd (%rdi), %ymm0, %k0
6817 ; VLX-NEXT: kmovd %k0, %eax
6818 ; VLX-NEXT: vzeroupper
6819 ; VLX-NEXT: retq
6820 ;
6821 ; NoVLX-LABEL: test_vpcmpsgtd_v8i1_v32i1_mask_mem:
6822 ; NoVLX: # %bb.0: # %entry
6823 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
6824 ; NoVLX-NEXT: vmovdqa (%rdi), %ymm1
6825 ; NoVLX-NEXT: vpcmpgtd %zmm1, %zmm0, %k0
6826 ; NoVLX-NEXT: kshiftlw $8, %k0, %k0
6827 ; NoVLX-NEXT: kshiftrw $8, %k0, %k0
6828 ; NoVLX-NEXT: kmovw %k0, %eax
6829 ; NoVLX-NEXT: vzeroupper
6830 ; NoVLX-NEXT: retq
6831 entry:
6832 %0 = bitcast <4 x i64> %__a to <8 x i32>
6833 %load = load <4 x i64>, <4 x i64>* %__b
6834 %1 = bitcast <4 x i64> %load to <8 x i32>
6835 %2 = icmp sgt <8 x i32> %0, %1
6836 %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
6837 %4 = bitcast <32 x i1> %3 to i32
6838 ret i32 %4
6839 }
6841 define zeroext i32 @test_masked_vpcmpsgtd_v8i1_v32i1_mask(i8 zeroext %__u, <4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
6842 ; VLX-LABEL: test_masked_vpcmpsgtd_v8i1_v32i1_mask:
6843 ; VLX: # %bb.0: # %entry
6844 ; VLX-NEXT: kmovd %edi, %k1
6845 ; VLX-NEXT: vpcmpgtd %ymm1, %ymm0, %k0 {%k1}
6846 ; VLX-NEXT: kmovd %k0, %eax
6847 ; VLX-NEXT: vzeroupper
6848 ; VLX-NEXT: retq
6849 ;
6850 ; NoVLX-LABEL: test_masked_vpcmpsgtd_v8i1_v32i1_mask:
6851 ; NoVLX: # %bb.0: # %entry
6852 ; NoVLX-NEXT: # kill: def $ymm1 killed $ymm1 def $zmm1
6853 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
6854 ; NoVLX-NEXT: kmovw %edi, %k1
6855 ; NoVLX-NEXT: vpcmpgtd %zmm1, %zmm0, %k0 {%k1}
6856 ; NoVLX-NEXT: kshiftlw $8, %k0, %k0
6857 ; NoVLX-NEXT: kshiftrw $8, %k0, %k0
6858 ; NoVLX-NEXT: kmovw %k0, %eax
6859 ; NoVLX-NEXT: vzeroupper
6860 ; NoVLX-NEXT: retq
6861 entry:
6862 %0 = bitcast <4 x i64> %__a to <8 x i32>
6863 %1 = bitcast <4 x i64> %__b to <8 x i32>
6864 %2 = icmp sgt <8 x i32> %0, %1
6865 %3 = bitcast i8 %__u to <8 x i1>
6866 %4 = and <8 x i1> %2, %3
6867 %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
6868 %6 = bitcast <32 x i1> %5 to i32
6869 ret i32 %6
6870 }
6872 define zeroext i32 @test_masked_vpcmpsgtd_v8i1_v32i1_mask_mem(i8 zeroext %__u, <4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
6873 ; VLX-LABEL: test_masked_vpcmpsgtd_v8i1_v32i1_mask_mem:
6874 ; VLX: # %bb.0: # %entry
6875 ; VLX-NEXT: kmovd %edi, %k1
6876 ; VLX-NEXT: vpcmpgtd (%rsi), %ymm0, %k0 {%k1}
6877 ; VLX-NEXT: kmovd %k0, %eax
6878 ; VLX-NEXT: vzeroupper
6879 ; VLX-NEXT: retq
6880 ;
6881 ; NoVLX-LABEL: test_masked_vpcmpsgtd_v8i1_v32i1_mask_mem:
6882 ; NoVLX: # %bb.0: # %entry
6883 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
6884 ; NoVLX-NEXT: vmovdqa (%rsi), %ymm1
6885 ; NoVLX-NEXT: kmovw %edi, %k1
6886 ; NoVLX-NEXT: vpcmpgtd %zmm1, %zmm0, %k0 {%k1}
6887 ; NoVLX-NEXT: kshiftlw $8, %k0, %k0
6888 ; NoVLX-NEXT: kshiftrw $8, %k0, %k0
6889 ; NoVLX-NEXT: kmovw %k0, %eax
6890 ; NoVLX-NEXT: vzeroupper
6891 ; NoVLX-NEXT: retq
6892 entry:
6893 %0 = bitcast <4 x i64> %__a to <8 x i32>
6894 %load = load <4 x i64>, <4 x i64>* %__b
6895 %1 = bitcast <4 x i64> %load to <8 x i32>
6896 %2 = icmp sgt <8 x i32> %0, %1
6897 %3 = bitcast i8 %__u to <8 x i1>
6898 %4 = and <8 x i1> %2, %3
6899 %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
6900 %6 = bitcast <32 x i1> %5 to i32
6901 ret i32 %6
6902 }
6905 define zeroext i32 @test_vpcmpsgtd_v8i1_v32i1_mask_mem_b(<4 x i64> %__a, i32* %__b) local_unnamed_addr {
6906 ; VLX-LABEL: test_vpcmpsgtd_v8i1_v32i1_mask_mem_b:
6907 ; VLX: # %bb.0: # %entry
6908 ; VLX-NEXT: vpcmpgtd (%rdi){1to8}, %ymm0, %k0
6909 ; VLX-NEXT: kmovd %k0, %eax
6910 ; VLX-NEXT: vzeroupper
6911 ; VLX-NEXT: retq
6912 ;
6913 ; NoVLX-LABEL: test_vpcmpsgtd_v8i1_v32i1_mask_mem_b:
6914 ; NoVLX: # %bb.0: # %entry
6915 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
6916 ; NoVLX-NEXT: vpcmpgtd (%rdi){1to16}, %zmm0, %k0
6917 ; NoVLX-NEXT: kshiftlw $8, %k0, %k0
6918 ; NoVLX-NEXT: kshiftrw $8, %k0, %k0
6919 ; NoVLX-NEXT: kmovw %k0, %eax
6920 ; NoVLX-NEXT: vzeroupper
6921 ; NoVLX-NEXT: retq
6922 entry:
6923 %0 = bitcast <4 x i64> %__a to <8 x i32>
6924 %load = load i32, i32* %__b
6925 %vec = insertelement <8 x i32> undef, i32 %load, i32 0
6926 %1 = shufflevector <8 x i32> %vec, <8 x i32> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
6927 %2 = icmp sgt <8 x i32> %0, %1
6928 %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
6929 %4 = bitcast <32 x i1> %3 to i32
6930 ret i32 %4
6931 }
6933 define zeroext i32 @test_masked_vpcmpsgtd_v8i1_v32i1_mask_mem_b(i8 zeroext %__u, <4 x i64> %__a, i32* %__b) local_unnamed_addr {
6934 ; VLX-LABEL: test_masked_vpcmpsgtd_v8i1_v32i1_mask_mem_b:
6935 ; VLX: # %bb.0: # %entry
6936 ; VLX-NEXT: kmovd %edi, %k1
6937 ; VLX-NEXT: vpcmpgtd (%rsi){1to8}, %ymm0, %k0 {%k1}
6938 ; VLX-NEXT: kmovd %k0, %eax
6939 ; VLX-NEXT: vzeroupper
6940 ; VLX-NEXT: retq
6941 ;
6942 ; NoVLX-LABEL: test_masked_vpcmpsgtd_v8i1_v32i1_mask_mem_b:
6943 ; NoVLX: # %bb.0: # %entry
6944 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
6945 ; NoVLX-NEXT: kmovw %edi, %k1
6946 ; NoVLX-NEXT: vpcmpgtd (%rsi){1to16}, %zmm0, %k0 {%k1}
6947 ; NoVLX-NEXT: kshiftlw $8, %k0, %k0
6948 ; NoVLX-NEXT: kshiftrw $8, %k0, %k0
6949 ; NoVLX-NEXT: kmovw %k0, %eax
6950 ; NoVLX-NEXT: vzeroupper
6951 ; NoVLX-NEXT: retq
6952 entry:
6953 %0 = bitcast <4 x i64> %__a to <8 x i32>
6954 %load = load i32, i32* %__b
6955 %vec = insertelement <8 x i32> undef, i32 %load, i32 0
6956 %1 = shufflevector <8 x i32> %vec, <8 x i32> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
6957 %2 = icmp sgt <8 x i32> %0, %1
6958 %3 = bitcast i8 %__u to <8 x i1>
6959 %4 = and <8 x i1> %3, %2
6960 %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
6961 %6 = bitcast <32 x i1> %5 to i32
6962 ret i32 %6
6963 }
6966 define zeroext i64 @test_vpcmpsgtd_v8i1_v64i1_mask(<4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
6967 ; VLX-LABEL: test_vpcmpsgtd_v8i1_v64i1_mask:
6968 ; VLX: # %bb.0: # %entry
6969 ; VLX-NEXT: vpcmpgtd %ymm1, %ymm0, %k0
6970 ; VLX-NEXT: kmovq %k0, %rax
6971 ; VLX-NEXT: vzeroupper
6972 ; VLX-NEXT: retq
6973 ;
6974 ; NoVLX-LABEL: test_vpcmpsgtd_v8i1_v64i1_mask:
6975 ; NoVLX: # %bb.0: # %entry
6976 ; NoVLX-NEXT: # kill: def $ymm1 killed $ymm1 def $zmm1
6977 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
6978 ; NoVLX-NEXT: vpcmpgtd %zmm1, %zmm0, %k0
6979 ; NoVLX-NEXT: kshiftlw $8, %k0, %k0
6980 ; NoVLX-NEXT: kshiftrw $8, %k0, %k0
6981 ; NoVLX-NEXT: kmovw %k0, %eax
6982 ; NoVLX-NEXT: vzeroupper
6983 ; NoVLX-NEXT: retq
6984 entry:
6985 %0 = bitcast <4 x i64> %__a to <8 x i32>
6986 %1 = bitcast <4 x i64> %__b to <8 x i32>
6987 %2 = icmp sgt <8 x i32> %0, %1
6988 %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
6989 %4 = bitcast <64 x i1> %3 to i64
6990 ret i64 %4
6991 }
6993 define zeroext i64 @test_vpcmpsgtd_v8i1_v64i1_mask_mem(<4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
6994 ; VLX-LABEL: test_vpcmpsgtd_v8i1_v64i1_mask_mem:
6995 ; VLX: # %bb.0: # %entry
6996 ; VLX-NEXT: vpcmpgtd (%rdi), %ymm0, %k0
6997 ; VLX-NEXT: kmovq %k0, %rax
6998 ; VLX-NEXT: vzeroupper
6999 ; VLX-NEXT: retq
7000 ;
7001 ; NoVLX-LABEL: test_vpcmpsgtd_v8i1_v64i1_mask_mem:
7002 ; NoVLX: # %bb.0: # %entry
7003 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
7004 ; NoVLX-NEXT: vmovdqa (%rdi), %ymm1
7005 ; NoVLX-NEXT: vpcmpgtd %zmm1, %zmm0, %k0
7006 ; NoVLX-NEXT: kshiftlw $8, %k0, %k0
7007 ; NoVLX-NEXT: kshiftrw $8, %k0, %k0
7008 ; NoVLX-NEXT: kmovw %k0, %eax
7009 ; NoVLX-NEXT: vzeroupper
7010 ; NoVLX-NEXT: retq
7011 entry:
7012 %0 = bitcast <4 x i64> %__a to <8 x i32>
7013 %load = load <4 x i64>, <4 x i64>* %__b
7014 %1 = bitcast <4 x i64> %load to <8 x i32>
7015 %2 = icmp sgt <8 x i32> %0, %1
7016 %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
7017 %4 = bitcast <64 x i1> %3 to i64
7018 ret i64 %4
7019 }
7021 define zeroext i64 @test_masked_vpcmpsgtd_v8i1_v64i1_mask(i8 zeroext %__u, <4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
7022 ; VLX-LABEL: test_masked_vpcmpsgtd_v8i1_v64i1_mask:
7023 ; VLX: # %bb.0: # %entry
7024 ; VLX-NEXT: kmovd %edi, %k1
7025 ; VLX-NEXT: vpcmpgtd %ymm1, %ymm0, %k0 {%k1}
7026 ; VLX-NEXT: kmovq %k0, %rax
7027 ; VLX-NEXT: vzeroupper
7028 ; VLX-NEXT: retq
7029 ;
7030 ; NoVLX-LABEL: test_masked_vpcmpsgtd_v8i1_v64i1_mask:
7031 ; NoVLX: # %bb.0: # %entry
7032 ; NoVLX-NEXT: # kill: def $ymm1 killed $ymm1 def $zmm1
7033 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
7034 ; NoVLX-NEXT: kmovw %edi, %k1
7035 ; NoVLX-NEXT: vpcmpgtd %zmm1, %zmm0, %k0 {%k1}
7036 ; NoVLX-NEXT: kshiftlw $8, %k0, %k0
7037 ; NoVLX-NEXT: kshiftrw $8, %k0, %k0
7038 ; NoVLX-NEXT: kmovw %k0, %eax
7039 ; NoVLX-NEXT: vzeroupper
7040 ; NoVLX-NEXT: retq
7041 entry:
7042 %0 = bitcast <4 x i64> %__a to <8 x i32>
7043 %1 = bitcast <4 x i64> %__b to <8 x i32>
7044 %2 = icmp sgt <8 x i32> %0, %1
7045 %3 = bitcast i8 %__u to <8 x i1>
7046 %4 = and <8 x i1> %2, %3
7047 %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
7048 %6 = bitcast <64 x i1> %5 to i64
7049 ret i64 %6
7050 }
7052 define zeroext i64 @test_masked_vpcmpsgtd_v8i1_v64i1_mask_mem(i8 zeroext %__u, <4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
7053 ; VLX-LABEL: test_masked_vpcmpsgtd_v8i1_v64i1_mask_mem:
7054 ; VLX: # %bb.0: # %entry
7055 ; VLX-NEXT: kmovd %edi, %k1
7056 ; VLX-NEXT: vpcmpgtd (%rsi), %ymm0, %k0 {%k1}
7057 ; VLX-NEXT: kmovq %k0, %rax
7058 ; VLX-NEXT: vzeroupper
7059 ; VLX-NEXT: retq
7060 ;
7061 ; NoVLX-LABEL: test_masked_vpcmpsgtd_v8i1_v64i1_mask_mem:
7062 ; NoVLX: # %bb.0: # %entry
7063 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
7064 ; NoVLX-NEXT: vmovdqa (%rsi), %ymm1
7065 ; NoVLX-NEXT: kmovw %edi, %k1
7066 ; NoVLX-NEXT: vpcmpgtd %zmm1, %zmm0, %k0 {%k1}
7067 ; NoVLX-NEXT: kshiftlw $8, %k0, %k0
7068 ; NoVLX-NEXT: kshiftrw $8, %k0, %k0
7069 ; NoVLX-NEXT: kmovw %k0, %eax
7070 ; NoVLX-NEXT: vzeroupper
7071 ; NoVLX-NEXT: retq
7072 entry:
7073 %0 = bitcast <4 x i64> %__a to <8 x i32>
7074 %load = load <4 x i64>, <4 x i64>* %__b
7075 %1 = bitcast <4 x i64> %load to <8 x i32>
7076 %2 = icmp sgt <8 x i32> %0, %1
7077 %3 = bitcast i8 %__u to <8 x i1>
7078 %4 = and <8 x i1> %2, %3
7079 %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
7080 %6 = bitcast <64 x i1> %5 to i64
7081 ret i64 %6
7082 }
7085 define zeroext i64 @test_vpcmpsgtd_v8i1_v64i1_mask_mem_b(<4 x i64> %__a, i32* %__b) local_unnamed_addr {
7086 ; VLX-LABEL: test_vpcmpsgtd_v8i1_v64i1_mask_mem_b:
7087 ; VLX: # %bb.0: # %entry
7088 ; VLX-NEXT: vpcmpgtd (%rdi){1to8}, %ymm0, %k0
7089 ; VLX-NEXT: kmovq %k0, %rax
7090 ; VLX-NEXT: vzeroupper
7091 ; VLX-NEXT: retq
7092 ;
7093 ; NoVLX-LABEL: test_vpcmpsgtd_v8i1_v64i1_mask_mem_b:
7094 ; NoVLX: # %bb.0: # %entry
7095 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
7096 ; NoVLX-NEXT: vpcmpgtd (%rdi){1to16}, %zmm0, %k0
7097 ; NoVLX-NEXT: kshiftlw $8, %k0, %k0
7098 ; NoVLX-NEXT: kshiftrw $8, %k0, %k0
7099 ; NoVLX-NEXT: kmovw %k0, %eax
7100 ; NoVLX-NEXT: vzeroupper
7101 ; NoVLX-NEXT: retq
7102 entry:
7103 %0 = bitcast <4 x i64> %__a to <8 x i32>
7104 %load = load i32, i32* %__b
7105 %vec = insertelement <8 x i32> undef, i32 %load, i32 0
7106 %1 = shufflevector <8 x i32> %vec, <8 x i32> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
7107 %2 = icmp sgt <8 x i32> %0, %1
7108 %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
7109 %4 = bitcast <64 x i1> %3 to i64
7110 ret i64 %4
7111 }
7113 define zeroext i64 @test_masked_vpcmpsgtd_v8i1_v64i1_mask_mem_b(i8 zeroext %__u, <4 x i64> %__a, i32* %__b) local_unnamed_addr {
7114 ; VLX-LABEL: test_masked_vpcmpsgtd_v8i1_v64i1_mask_mem_b:
7115 ; VLX: # %bb.0: # %entry
7116 ; VLX-NEXT: kmovd %edi, %k1
7117 ; VLX-NEXT: vpcmpgtd (%rsi){1to8}, %ymm0, %k0 {%k1}
7118 ; VLX-NEXT: kmovq %k0, %rax
7119 ; VLX-NEXT: vzeroupper
7120 ; VLX-NEXT: retq
7121 ;
7122 ; NoVLX-LABEL: test_masked_vpcmpsgtd_v8i1_v64i1_mask_mem_b:
7123 ; NoVLX: # %bb.0: # %entry
7124 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
7125 ; NoVLX-NEXT: kmovw %edi, %k1
7126 ; NoVLX-NEXT: vpcmpgtd (%rsi){1to16}, %zmm0, %k0 {%k1}
7127 ; NoVLX-NEXT: kshiftlw $8, %k0, %k0
7128 ; NoVLX-NEXT: kshiftrw $8, %k0, %k0
7129 ; NoVLX-NEXT: kmovw %k0, %eax
7130 ; NoVLX-NEXT: vzeroupper
7131 ; NoVLX-NEXT: retq
7132 entry:
7133 %0 = bitcast <4 x i64> %__a to <8 x i32>
7134 %load = load i32, i32* %__b
7135 %vec = insertelement <8 x i32> undef, i32 %load, i32 0
7136 %1 = shufflevector <8 x i32> %vec, <8 x i32> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
7137 %2 = icmp sgt <8 x i32> %0, %1
7138 %3 = bitcast i8 %__u to <8 x i1>
7139 %4 = and <8 x i1> %3, %2
7140 %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
7141 %6 = bitcast <64 x i1> %5 to i64
7142 ret i64 %6
7143 }
7146 define zeroext i32 @test_vpcmpsgtd_v16i1_v32i1_mask(<8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
7147 ; VLX-LABEL: test_vpcmpsgtd_v16i1_v32i1_mask:
7148 ; VLX: # %bb.0: # %entry
7149 ; VLX-NEXT: vpcmpgtd %zmm1, %zmm0, %k0
7150 ; VLX-NEXT: kmovd %k0, %eax
7151 ; VLX-NEXT: vzeroupper
7152 ; VLX-NEXT: retq
7153 ;
7154 ; NoVLX-LABEL: test_vpcmpsgtd_v16i1_v32i1_mask:
7155 ; NoVLX: # %bb.0: # %entry
7156 ; NoVLX-NEXT: vpcmpgtd %zmm1, %zmm0, %k0
7157 ; NoVLX-NEXT: kmovw %k0, %eax
7158 ; NoVLX-NEXT: vzeroupper
7159 ; NoVLX-NEXT: retq
7160 entry:
7161 %0 = bitcast <8 x i64> %__a to <16 x i32>
7162 %1 = bitcast <8 x i64> %__b to <16 x i32>
7163 %2 = icmp sgt <16 x i32> %0, %1
7164 %3 = shufflevector <16 x i1> %2, <16 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>
7165 %4 = bitcast <32 x i1> %3 to i32
7166 ret i32 %4
7167 }
7169 define zeroext i32 @test_vpcmpsgtd_v16i1_v32i1_mask_mem(<8 x i64> %__a, <8 x i64>* %__b) local_unnamed_addr {
7170 ; VLX-LABEL: test_vpcmpsgtd_v16i1_v32i1_mask_mem:
7171 ; VLX: # %bb.0: # %entry
7172 ; VLX-NEXT: vpcmpgtd (%rdi), %zmm0, %k0
7173 ; VLX-NEXT: kmovd %k0, %eax
7174 ; VLX-NEXT: vzeroupper
7175 ; VLX-NEXT: retq
7176 ;
7177 ; NoVLX-LABEL: test_vpcmpsgtd_v16i1_v32i1_mask_mem:
7178 ; NoVLX: # %bb.0: # %entry
7179 ; NoVLX-NEXT: vpcmpgtd (%rdi), %zmm0, %k0
7180 ; NoVLX-NEXT: kmovw %k0, %eax
7181 ; NoVLX-NEXT: vzeroupper
7182 ; NoVLX-NEXT: retq
7183 entry:
7184 %0 = bitcast <8 x i64> %__a to <16 x i32>
7185 %load = load <8 x i64>, <8 x i64>* %__b
7186 %1 = bitcast <8 x i64> %load to <16 x i32>
7187 %2 = icmp sgt <16 x i32> %0, %1
7188 %3 = shufflevector <16 x i1> %2, <16 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>
7189 %4 = bitcast <32 x i1> %3 to i32
7190 ret i32 %4
7191 }
7193 define zeroext i32 @test_masked_vpcmpsgtd_v16i1_v32i1_mask(i16 zeroext %__u, <8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
7194 ; VLX-LABEL: test_masked_vpcmpsgtd_v16i1_v32i1_mask:
7195 ; VLX: # %bb.0: # %entry
7196 ; VLX-NEXT: kmovd %edi, %k1
7197 ; VLX-NEXT: vpcmpgtd %zmm1, %zmm0, %k0 {%k1}
7198 ; VLX-NEXT: kmovd %k0, %eax
7199 ; VLX-NEXT: vzeroupper
7200 ; VLX-NEXT: retq
7201 ;
7202 ; NoVLX-LABEL: test_masked_vpcmpsgtd_v16i1_v32i1_mask:
7203 ; NoVLX: # %bb.0: # %entry
7204 ; NoVLX-NEXT: vpcmpgtd %zmm1, %zmm0, %k0
7205 ; NoVLX-NEXT: kmovw %k0, %eax
7206 ; NoVLX-NEXT: andl %edi, %eax
7207 ; NoVLX-NEXT: vzeroupper
7208 ; NoVLX-NEXT: retq
7209 entry:
7210 %0 = bitcast <8 x i64> %__a to <16 x i32>
7211 %1 = bitcast <8 x i64> %__b to <16 x i32>
7212 %2 = icmp sgt <16 x i32> %0, %1
7213 %3 = bitcast i16 %__u to <16 x i1>
7214 %4 = and <16 x i1> %2, %3
7215 %5 = shufflevector <16 x i1> %4, <16 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>
7216 %6 = bitcast <32 x i1> %5 to i32
7217 ret i32 %6
7218 }
7220 define zeroext i32 @test_masked_vpcmpsgtd_v16i1_v32i1_mask_mem(i16 zeroext %__u, <8 x i64> %__a, <8 x i64>* %__b) local_unnamed_addr {
7221 ; VLX-LABEL: test_masked_vpcmpsgtd_v16i1_v32i1_mask_mem:
7222 ; VLX: # %bb.0: # %entry
7223 ; VLX-NEXT: kmovd %edi, %k1
7224 ; VLX-NEXT: vpcmpgtd (%rsi), %zmm0, %k0 {%k1}
7225 ; VLX-NEXT: kmovd %k0, %eax
7226 ; VLX-NEXT: vzeroupper
7227 ; VLX-NEXT: retq
7228 ;
7229 ; NoVLX-LABEL: test_masked_vpcmpsgtd_v16i1_v32i1_mask_mem:
7230 ; NoVLX: # %bb.0: # %entry
7231 ; NoVLX-NEXT: vpcmpgtd (%rsi), %zmm0, %k0
7232 ; NoVLX-NEXT: kmovw %k0, %eax
7233 ; NoVLX-NEXT: andl %edi, %eax
7234 ; NoVLX-NEXT: vzeroupper
7235 ; NoVLX-NEXT: retq
7236 entry:
7237 %0 = bitcast <8 x i64> %__a to <16 x i32>
7238 %load = load <8 x i64>, <8 x i64>* %__b
7239 %1 = bitcast <8 x i64> %load to <16 x i32>
7240 %2 = icmp sgt <16 x i32> %0, %1
7241 %3 = bitcast i16 %__u to <16 x i1>
7242 %4 = and <16 x i1> %2, %3
7243 %5 = shufflevector <16 x i1> %4, <16 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>
7244 %6 = bitcast <32 x i1> %5 to i32
7245 ret i32 %6
7246 }
7249 define zeroext i32 @test_vpcmpsgtd_v16i1_v32i1_mask_mem_b(<8 x i64> %__a, i32* %__b) local_unnamed_addr {
7250 ; VLX-LABEL: test_vpcmpsgtd_v16i1_v32i1_mask_mem_b:
7251 ; VLX: # %bb.0: # %entry
7252 ; VLX-NEXT: vpcmpgtd (%rdi){1to16}, %zmm0, %k0
7253 ; VLX-NEXT: kmovd %k0, %eax
7254 ; VLX-NEXT: vzeroupper
7255 ; VLX-NEXT: retq
7256 ;
7257 ; NoVLX-LABEL: test_vpcmpsgtd_v16i1_v32i1_mask_mem_b:
7258 ; NoVLX: # %bb.0: # %entry
7259 ; NoVLX-NEXT: vpcmpgtd (%rdi){1to16}, %zmm0, %k0
7260 ; NoVLX-NEXT: kmovw %k0, %eax
7261 ; NoVLX-NEXT: vzeroupper
7262 ; NoVLX-NEXT: retq
7263 entry:
7264 %0 = bitcast <8 x i64> %__a to <16 x i32>
7265 %load = load i32, i32* %__b
7266 %vec = insertelement <16 x i32> undef, i32 %load, i32 0
7267 %1 = shufflevector <16 x i32> %vec, <16 x i32> undef, <16 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
7268 %2 = icmp sgt <16 x i32> %0, %1
7269 %3 = shufflevector <16 x i1> %2, <16 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>
7270 %4 = bitcast <32 x i1> %3 to i32
7271 ret i32 %4
7272 }
7274 define zeroext i32 @test_masked_vpcmpsgtd_v16i1_v32i1_mask_mem_b(i16 zeroext %__u, <8 x i64> %__a, i32* %__b) local_unnamed_addr {
7275 ; VLX-LABEL: test_masked_vpcmpsgtd_v16i1_v32i1_mask_mem_b:
7276 ; VLX: # %bb.0: # %entry
7277 ; VLX-NEXT: kmovd %edi, %k1
7278 ; VLX-NEXT: vpcmpgtd (%rsi){1to16}, %zmm0, %k0 {%k1}
7279 ; VLX-NEXT: kmovd %k0, %eax
7280 ; VLX-NEXT: vzeroupper
7281 ; VLX-NEXT: retq
7282 ;
7283 ; NoVLX-LABEL: test_masked_vpcmpsgtd_v16i1_v32i1_mask_mem_b:
7284 ; NoVLX: # %bb.0: # %entry
7285 ; NoVLX-NEXT: vpcmpgtd (%rsi){1to16}, %zmm0, %k0
7286 ; NoVLX-NEXT: kmovw %k0, %eax
7287 ; NoVLX-NEXT: andl %edi, %eax
7288 ; NoVLX-NEXT: vzeroupper
7289 ; NoVLX-NEXT: retq
7290 entry:
7291 %0 = bitcast <8 x i64> %__a to <16 x i32>
7292 %load = load i32, i32* %__b
7293 %vec = insertelement <16 x i32> undef, i32 %load, i32 0
7294 %1 = shufflevector <16 x i32> %vec, <16 x i32> undef, <16 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
7295 %2 = icmp sgt <16 x i32> %0, %1
7296 %3 = bitcast i16 %__u to <16 x i1>
7297 %4 = and <16 x i1> %3, %2
7298 %5 = shufflevector <16 x i1> %4, <16 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>
7299 %6 = bitcast <32 x i1> %5 to i32
7300 ret i32 %6
7301 }
7304 define zeroext i64 @test_vpcmpsgtd_v16i1_v64i1_mask(<8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
7305 ; VLX-LABEL: test_vpcmpsgtd_v16i1_v64i1_mask:
7306 ; VLX: # %bb.0: # %entry
7307 ; VLX-NEXT: vpcmpgtd %zmm1, %zmm0, %k0
7308 ; VLX-NEXT: kmovq %k0, %rax
7309 ; VLX-NEXT: vzeroupper
7310 ; VLX-NEXT: retq
7311 ;
7312 ; NoVLX-LABEL: test_vpcmpsgtd_v16i1_v64i1_mask:
7313 ; NoVLX: # %bb.0: # %entry
7314 ; NoVLX-NEXT: vpcmpgtd %zmm1, %zmm0, %k0
7315 ; NoVLX-NEXT: kmovw %k0, %eax
7316 ; NoVLX-NEXT: vzeroupper
7317 ; NoVLX-NEXT: retq
7318 entry:
7319 %0 = bitcast <8 x i64> %__a to <16 x i32>
7320 %1 = bitcast <8 x i64> %__b to <16 x i32>
7321 %2 = icmp sgt <16 x i32> %0, %1
7322 %3 = shufflevector <16 x i1> %2, <16 x i1> zeroinitializer, <64 x i32> <i32 0,i32 1,i32 2,i32 3,i32 4,i32 5,i32 6,i32 7,i32 8,i32 9,i32 10,i32 11,i32 12,i32 13,i32 14,i32 15,i32 16,i32 17,i32 18,i32 19,i32 20,i32 21,i32 22,i32 23,i32 24,i32 25,i32 26,i32 27,i32 28,i32 29,i32 30,i32 31,i32 16,i32 17,i32 18,i32 19,i32 20,i32 21,i32 22,i32 23,i32 24,i32 25,i32 26,i32 27,i32 28,i32 29,i32 30,i32 31,i32 16,i32 17,i32 18,i32 19,i32 20,i32 21,i32 22,i32 23,i32 24,i32 25,i32 26,i32 27,i32 28,i32 29,i32 30,i32 31>
7323 %4 = bitcast <64 x i1> %3 to i64
7324 ret i64 %4
7325 }
7327 define zeroext i64 @test_vpcmpsgtd_v16i1_v64i1_mask_mem(<8 x i64> %__a, <8 x i64>* %__b) local_unnamed_addr {
7328 ; VLX-LABEL: test_vpcmpsgtd_v16i1_v64i1_mask_mem:
7329 ; VLX: # %bb.0: # %entry
7330 ; VLX-NEXT: vpcmpgtd (%rdi), %zmm0, %k0
7331 ; VLX-NEXT: kmovq %k0, %rax
7332 ; VLX-NEXT: vzeroupper
7333 ; VLX-NEXT: retq
7334 ;
7335 ; NoVLX-LABEL: test_vpcmpsgtd_v16i1_v64i1_mask_mem:
7336 ; NoVLX: # %bb.0: # %entry
7337 ; NoVLX-NEXT: vpcmpgtd (%rdi), %zmm0, %k0
7338 ; NoVLX-NEXT: kmovw %k0, %eax
7339 ; NoVLX-NEXT: vzeroupper
7340 ; NoVLX-NEXT: retq
7341 entry:
7342 %0 = bitcast <8 x i64> %__a to <16 x i32>
7343 %load = load <8 x i64>, <8 x i64>* %__b
7344 %1 = bitcast <8 x i64> %load to <16 x i32>
7345 %2 = icmp sgt <16 x i32> %0, %1
7346 %3 = shufflevector <16 x i1> %2, <16 x i1> zeroinitializer, <64 x i32> <i32 0,i32 1,i32 2,i32 3,i32 4,i32 5,i32 6,i32 7,i32 8,i32 9,i32 10,i32 11,i32 12,i32 13,i32 14,i32 15,i32 16,i32 17,i32 18,i32 19,i32 20,i32 21,i32 22,i32 23,i32 24,i32 25,i32 26,i32 27,i32 28,i32 29,i32 30,i32 31,i32 16,i32 17,i32 18,i32 19,i32 20,i32 21,i32 22,i32 23,i32 24,i32 25,i32 26,i32 27,i32 28,i32 29,i32 30,i32 31,i32 16,i32 17,i32 18,i32 19,i32 20,i32 21,i32 22,i32 23,i32 24,i32 25,i32 26,i32 27,i32 28,i32 29,i32 30,i32 31>
7347 %4 = bitcast <64 x i1> %3 to i64
7348 ret i64 %4
7349 }
7351 define zeroext i64 @test_masked_vpcmpsgtd_v16i1_v64i1_mask(i16 zeroext %__u, <8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
7352 ; VLX-LABEL: test_masked_vpcmpsgtd_v16i1_v64i1_mask:
7353 ; VLX: # %bb.0: # %entry
7354 ; VLX-NEXT: kmovd %edi, %k1
7355 ; VLX-NEXT: vpcmpgtd %zmm1, %zmm0, %k0 {%k1}
7356 ; VLX-NEXT: kmovq %k0, %rax
7357 ; VLX-NEXT: vzeroupper
7358 ; VLX-NEXT: retq
7359 ;
7360 ; NoVLX-LABEL: test_masked_vpcmpsgtd_v16i1_v64i1_mask:
7361 ; NoVLX: # %bb.0: # %entry
7362 ; NoVLX-NEXT: vpcmpgtd %zmm1, %zmm0, %k0
7363 ; NoVLX-NEXT: kmovw %k0, %eax
7364 ; NoVLX-NEXT: andl %edi, %eax
7365 ; NoVLX-NEXT: vzeroupper
7366 ; NoVLX-NEXT: retq
7367 entry:
7368 %0 = bitcast <8 x i64> %__a to <16 x i32>
7369 %1 = bitcast <8 x i64> %__b to <16 x i32>
7370 %2 = icmp sgt <16 x i32> %0, %1
7371 %3 = bitcast i16 %__u to <16 x i1>
7372 %4 = and <16 x i1> %2, %3
7373 %5 = shufflevector <16 x i1> %4, <16 x i1> zeroinitializer, <64 x i32> <i32 0,i32 1,i32 2,i32 3,i32 4,i32 5,i32 6,i32 7,i32 8,i32 9,i32 10,i32 11,i32 12,i32 13,i32 14,i32 15,i32 16,i32 17,i32 18,i32 19,i32 20,i32 21,i32 22,i32 23,i32 24,i32 25,i32 26,i32 27,i32 28,i32 29,i32 30,i32 31,i32 16,i32 17,i32 18,i32 19,i32 20,i32 21,i32 22,i32 23,i32 24,i32 25,i32 26,i32 27,i32 28,i32 29,i32 30,i32 31,i32 16,i32 17,i32 18,i32 19,i32 20,i32 21,i32 22,i32 23,i32 24,i32 25,i32 26,i32 27,i32 28,i32 29,i32 30,i32 31>
7374 %6 = bitcast <64 x i1> %5 to i64
7375 ret i64 %6
7376 }
7378 define zeroext i64 @test_masked_vpcmpsgtd_v16i1_v64i1_mask_mem(i16 zeroext %__u, <8 x i64> %__a, <8 x i64>* %__b) local_unnamed_addr {
7379 ; VLX-LABEL: test_masked_vpcmpsgtd_v16i1_v64i1_mask_mem:
7380 ; VLX: # %bb.0: # %entry
7381 ; VLX-NEXT: kmovd %edi, %k1
7382 ; VLX-NEXT: vpcmpgtd (%rsi), %zmm0, %k0 {%k1}
7383 ; VLX-NEXT: kmovq %k0, %rax
7384 ; VLX-NEXT: vzeroupper
7385 ; VLX-NEXT: retq
7386 ;
7387 ; NoVLX-LABEL: test_masked_vpcmpsgtd_v16i1_v64i1_mask_mem:
7388 ; NoVLX: # %bb.0: # %entry
7389 ; NoVLX-NEXT: vpcmpgtd (%rsi), %zmm0, %k0
7390 ; NoVLX-NEXT: kmovw %k0, %eax
7391 ; NoVLX-NEXT: andl %edi, %eax
7392 ; NoVLX-NEXT: vzeroupper
7393 ; NoVLX-NEXT: retq
7394 entry:
7395 %0 = bitcast <8 x i64> %__a to <16 x i32>
7396 %load = load <8 x i64>, <8 x i64>* %__b
7397 %1 = bitcast <8 x i64> %load to <16 x i32>
7398 %2 = icmp sgt <16 x i32> %0, %1
7399 %3 = bitcast i16 %__u to <16 x i1>
7400 %4 = and <16 x i1> %2, %3
7401 %5 = shufflevector <16 x i1> %4, <16 x i1> zeroinitializer, <64 x i32> <i32 0,i32 1,i32 2,i32 3,i32 4,i32 5,i32 6,i32 7,i32 8,i32 9,i32 10,i32 11,i32 12,i32 13,i32 14,i32 15,i32 16,i32 17,i32 18,i32 19,i32 20,i32 21,i32 22,i32 23,i32 24,i32 25,i32 26,i32 27,i32 28,i32 29,i32 30,i32 31,i32 16,i32 17,i32 18,i32 19,i32 20,i32 21,i32 22,i32 23,i32 24,i32 25,i32 26,i32 27,i32 28,i32 29,i32 30,i32 31,i32 16,i32 17,i32 18,i32 19,i32 20,i32 21,i32 22,i32 23,i32 24,i32 25,i32 26,i32 27,i32 28,i32 29,i32 30,i32 31>
7402 %6 = bitcast <64 x i1> %5 to i64
7403 ret i64 %6
7404 }
7407 define zeroext i64 @test_vpcmpsgtd_v16i1_v64i1_mask_mem_b(<8 x i64> %__a, i32* %__b) local_unnamed_addr {
7408 ; VLX-LABEL: test_vpcmpsgtd_v16i1_v64i1_mask_mem_b:
7409 ; VLX: # %bb.0: # %entry
7410 ; VLX-NEXT: vpcmpgtd (%rdi){1to16}, %zmm0, %k0
7411 ; VLX-NEXT: kmovq %k0, %rax
7412 ; VLX-NEXT: vzeroupper
7413 ; VLX-NEXT: retq
7414 ;
7415 ; NoVLX-LABEL: test_vpcmpsgtd_v16i1_v64i1_mask_mem_b:
7416 ; NoVLX: # %bb.0: # %entry
7417 ; NoVLX-NEXT: vpcmpgtd (%rdi){1to16}, %zmm0, %k0
7418 ; NoVLX-NEXT: kmovw %k0, %eax
7419 ; NoVLX-NEXT: vzeroupper
7420 ; NoVLX-NEXT: retq
7421 entry:
7422 %0 = bitcast <8 x i64> %__a to <16 x i32>
7423 %load = load i32, i32* %__b
7424 %vec = insertelement <16 x i32> undef, i32 %load, i32 0
7425 %1 = shufflevector <16 x i32> %vec, <16 x i32> undef, <16 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
7426 %2 = icmp sgt <16 x i32> %0, %1
7427 %3 = shufflevector <16 x i1> %2, <16 x i1> zeroinitializer, <64 x i32> <i32 0,i32 1,i32 2,i32 3,i32 4,i32 5,i32 6,i32 7,i32 8,i32 9,i32 10,i32 11,i32 12,i32 13,i32 14,i32 15,i32 16,i32 17,i32 18,i32 19,i32 20,i32 21,i32 22,i32 23,i32 24,i32 25,i32 26,i32 27,i32 28,i32 29,i32 30,i32 31,i32 16,i32 17,i32 18,i32 19,i32 20,i32 21,i32 22,i32 23,i32 24,i32 25,i32 26,i32 27,i32 28,i32 29,i32 30,i32 31,i32 16,i32 17,i32 18,i32 19,i32 20,i32 21,i32 22,i32 23,i32 24,i32 25,i32 26,i32 27,i32 28,i32 29,i32 30,i32 31>
7428 %4 = bitcast <64 x i1> %3 to i64
7429 ret i64 %4
7430 }
7432 define zeroext i64 @test_masked_vpcmpsgtd_v16i1_v64i1_mask_mem_b(i16 zeroext %__u, <8 x i64> %__a, i32* %__b) local_unnamed_addr {
7433 ; VLX-LABEL: test_masked_vpcmpsgtd_v16i1_v64i1_mask_mem_b:
7434 ; VLX: # %bb.0: # %entry
7435 ; VLX-NEXT: kmovd %edi, %k1
7436 ; VLX-NEXT: vpcmpgtd (%rsi){1to16}, %zmm0, %k0 {%k1}
7437 ; VLX-NEXT: kmovq %k0, %rax
7438 ; VLX-NEXT: vzeroupper
7439 ; VLX-NEXT: retq
7440 ;
7441 ; NoVLX-LABEL: test_masked_vpcmpsgtd_v16i1_v64i1_mask_mem_b:
7442 ; NoVLX: # %bb.0: # %entry
7443 ; NoVLX-NEXT: vpcmpgtd (%rsi){1to16}, %zmm0, %k0
7444 ; NoVLX-NEXT: kmovw %k0, %eax
7445 ; NoVLX-NEXT: andl %edi, %eax
7446 ; NoVLX-NEXT: vzeroupper
7447 ; NoVLX-NEXT: retq
7448 entry:
7449 %0 = bitcast <8 x i64> %__a to <16 x i32>
7450 %load = load i32, i32* %__b
7451 %vec = insertelement <16 x i32> undef, i32 %load, i32 0
7452 %1 = shufflevector <16 x i32> %vec, <16 x i32> undef, <16 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
7453 %2 = icmp sgt <16 x i32> %0, %1
7454 %3 = bitcast i16 %__u to <16 x i1>
7455 %4 = and <16 x i1> %3, %2
7456 %5 = shufflevector <16 x i1> %4, <16 x i1> zeroinitializer, <64 x i32> <i32 0,i32 1,i32 2,i32 3,i32 4,i32 5,i32 6,i32 7,i32 8,i32 9,i32 10,i32 11,i32 12,i32 13,i32 14,i32 15,i32 16,i32 17,i32 18,i32 19,i32 20,i32 21,i32 22,i32 23,i32 24,i32 25,i32 26,i32 27,i32 28,i32 29,i32 30,i32 31,i32 16,i32 17,i32 18,i32 19,i32 20,i32 21,i32 22,i32 23,i32 24,i32 25,i32 26,i32 27,i32 28,i32 29,i32 30,i32 31,i32 16,i32 17,i32 18,i32 19,i32 20,i32 21,i32 22,i32 23,i32 24,i32 25,i32 26,i32 27,i32 28,i32 29,i32 30,i32 31>
7457 %6 = bitcast <64 x i1> %5 to i64
7458 ret i64 %6
7459 }
7462 define zeroext i4 @test_vpcmpsgtq_v2i1_v4i1_mask(<2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
7463 ; VLX-LABEL: test_vpcmpsgtq_v2i1_v4i1_mask:
7464 ; VLX: # %bb.0: # %entry
7465 ; VLX-NEXT: vpcmpgtq %xmm1, %xmm0, %k0
7466 ; VLX-NEXT: kmovb %k0, %eax
7467 ; VLX-NEXT: retq
7468 ;
7469 ; NoVLX-LABEL: test_vpcmpsgtq_v2i1_v4i1_mask:
7470 ; NoVLX: # %bb.0: # %entry
7471 ; NoVLX-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
7472 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
7473 ; NoVLX-NEXT: vpcmpgtq %zmm1, %zmm0, %k0
7474 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
7475 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
7476 ; NoVLX-NEXT: kmovw %k0, %eax
7477 ; NoVLX-NEXT: vzeroupper
7478 ; NoVLX-NEXT: retq
7479 entry:
7480 %0 = bitcast <2 x i64> %__a to <2 x i64>
7481 %1 = bitcast <2 x i64> %__b to <2 x i64>
7482 %2 = icmp sgt <2 x i64> %0, %1
7483 %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
7484 %4 = bitcast <4 x i1> %3 to i4
7485 ret i4 %4
7486 }
7488 define zeroext i4 @test_vpcmpsgtq_v2i1_v4i1_mask_mem(<2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
7489 ; VLX-LABEL: test_vpcmpsgtq_v2i1_v4i1_mask_mem:
7490 ; VLX: # %bb.0: # %entry
7491 ; VLX-NEXT: vpcmpgtq (%rdi), %xmm0, %k0
7492 ; VLX-NEXT: kmovb %k0, %eax
7493 ; VLX-NEXT: retq
7494 ;
7495 ; NoVLX-LABEL: test_vpcmpsgtq_v2i1_v4i1_mask_mem:
7496 ; NoVLX: # %bb.0: # %entry
7497 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
7498 ; NoVLX-NEXT: vmovdqa (%rdi), %xmm1
7499 ; NoVLX-NEXT: vpcmpgtq %zmm1, %zmm0, %k0
7500 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
7501 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
7502 ; NoVLX-NEXT: kmovw %k0, %eax
7503 ; NoVLX-NEXT: vzeroupper
7504 ; NoVLX-NEXT: retq
7505 entry:
7506 %0 = bitcast <2 x i64> %__a to <2 x i64>
7507 %load = load <2 x i64>, <2 x i64>* %__b
7508 %1 = bitcast <2 x i64> %load to <2 x i64>
7509 %2 = icmp sgt <2 x i64> %0, %1
7510 %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
7511 %4 = bitcast <4 x i1> %3 to i4
7512 ret i4 %4
7513 }
7515 define zeroext i4 @test_masked_vpcmpsgtq_v2i1_v4i1_mask(i8 zeroext %__u, <2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
7516 ; VLX-LABEL: test_masked_vpcmpsgtq_v2i1_v4i1_mask:
7517 ; VLX: # %bb.0: # %entry
7518 ; VLX-NEXT: kmovd %edi, %k1
7519 ; VLX-NEXT: vpcmpgtq %xmm1, %xmm0, %k0 {%k1}
7520 ; VLX-NEXT: kmovb %k0, %eax
7521 ; VLX-NEXT: retq
7522 ;
7523 ; NoVLX-LABEL: test_masked_vpcmpsgtq_v2i1_v4i1_mask:
7524 ; NoVLX: # %bb.0: # %entry
7525 ; NoVLX-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
7526 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
7527 ; NoVLX-NEXT: kmovw %edi, %k1
7528 ; NoVLX-NEXT: vpcmpgtq %zmm1, %zmm0, %k0 {%k1}
7529 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
7530 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
7531 ; NoVLX-NEXT: kmovw %k0, %eax
7532 ; NoVLX-NEXT: vzeroupper
7533 ; NoVLX-NEXT: retq
7534 entry:
7535 %0 = bitcast <2 x i64> %__a to <2 x i64>
7536 %1 = bitcast <2 x i64> %__b to <2 x i64>
7537 %2 = icmp sgt <2 x i64> %0, %1
7538 %3 = bitcast i8 %__u to <8 x i1>
7539 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <2 x i32> <i32 0, i32 1>
7540 %4 = and <2 x i1> %2, %extract.i
7541 %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
7542 %6 = bitcast <4 x i1> %5 to i4
7543 ret i4 %6
7544 }
7546 define zeroext i4 @test_masked_vpcmpsgtq_v2i1_v4i1_mask_mem(i8 zeroext %__u, <2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
7547 ; VLX-LABEL: test_masked_vpcmpsgtq_v2i1_v4i1_mask_mem:
7548 ; VLX: # %bb.0: # %entry
7549 ; VLX-NEXT: kmovd %edi, %k1
7550 ; VLX-NEXT: vpcmpgtq (%rsi), %xmm0, %k0 {%k1}
7551 ; VLX-NEXT: kmovb %k0, %eax
7552 ; VLX-NEXT: retq
7553 ;
7554 ; NoVLX-LABEL: test_masked_vpcmpsgtq_v2i1_v4i1_mask_mem:
7555 ; NoVLX: # %bb.0: # %entry
7556 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
7557 ; NoVLX-NEXT: vmovdqa (%rsi), %xmm1
7558 ; NoVLX-NEXT: kmovw %edi, %k1
7559 ; NoVLX-NEXT: vpcmpgtq %zmm1, %zmm0, %k0 {%k1}
7560 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
7561 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
7562 ; NoVLX-NEXT: kmovw %k0, %eax
7563 ; NoVLX-NEXT: vzeroupper
7564 ; NoVLX-NEXT: retq
7565 entry:
7566 %0 = bitcast <2 x i64> %__a to <2 x i64>
7567 %load = load <2 x i64>, <2 x i64>* %__b
7568 %1 = bitcast <2 x i64> %load to <2 x i64>
7569 %2 = icmp sgt <2 x i64> %0, %1
7570 %3 = bitcast i8 %__u to <8 x i1>
7571 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <2 x i32> <i32 0, i32 1>
7572 %4 = and <2 x i1> %2, %extract.i
7573 %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
7574 %6 = bitcast <4 x i1> %5 to i4
7575 ret i4 %6
7576 }
7579 define zeroext i4 @test_vpcmpsgtq_v2i1_v4i1_mask_mem_b(<2 x i64> %__a, i64* %__b) local_unnamed_addr {
7580 ; VLX-LABEL: test_vpcmpsgtq_v2i1_v4i1_mask_mem_b:
7581 ; VLX: # %bb.0: # %entry
7582 ; VLX-NEXT: vpcmpgtq (%rdi){1to2}, %xmm0, %k0
7583 ; VLX-NEXT: kmovb %k0, %eax
7584 ; VLX-NEXT: retq
7585 ;
7586 ; NoVLX-LABEL: test_vpcmpsgtq_v2i1_v4i1_mask_mem_b:
7587 ; NoVLX: # %bb.0: # %entry
7588 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
7589 ; NoVLX-NEXT: vpcmpgtq (%rdi){1to8}, %zmm0, %k0
7590 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
7591 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
7592 ; NoVLX-NEXT: kmovw %k0, %eax
7593 ; NoVLX-NEXT: vzeroupper
7594 ; NoVLX-NEXT: retq
7595 entry:
7596 %0 = bitcast <2 x i64> %__a to <2 x i64>
7597 %load = load i64, i64* %__b
7598 %vec = insertelement <2 x i64> undef, i64 %load, i32 0
7599 %1 = shufflevector <2 x i64> %vec, <2 x i64> undef, <2 x i32> <i32 0, i32 0>
7600 %2 = icmp sgt <2 x i64> %0, %1
7601 %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
7602 %4 = bitcast <4 x i1> %3 to i4
7603 ret i4 %4
7604 }
7606 define zeroext i4 @test_masked_vpcmpsgtq_v2i1_v4i1_mask_mem_b(i8 zeroext %__u, <2 x i64> %__a, i64* %__b) local_unnamed_addr {
7607 ; VLX-LABEL: test_masked_vpcmpsgtq_v2i1_v4i1_mask_mem_b:
7608 ; VLX: # %bb.0: # %entry
7609 ; VLX-NEXT: kmovd %edi, %k1
7610 ; VLX-NEXT: vpcmpgtq (%rsi){1to2}, %xmm0, %k0 {%k1}
7611 ; VLX-NEXT: kmovb %k0, %eax
7612 ; VLX-NEXT: retq
7613 ;
7614 ; NoVLX-LABEL: test_masked_vpcmpsgtq_v2i1_v4i1_mask_mem_b:
7615 ; NoVLX: # %bb.0: # %entry
7616 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
7617 ; NoVLX-NEXT: kmovw %edi, %k1
7618 ; NoVLX-NEXT: vpcmpgtq (%rsi){1to8}, %zmm0, %k0 {%k1}
7619 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
7620 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
7621 ; NoVLX-NEXT: kmovw %k0, %eax
7622 ; NoVLX-NEXT: vzeroupper
7623 ; NoVLX-NEXT: retq
7624 entry:
7625 %0 = bitcast <2 x i64> %__a to <2 x i64>
7626 %load = load i64, i64* %__b
7627 %vec = insertelement <2 x i64> undef, i64 %load, i32 0
7628 %1 = shufflevector <2 x i64> %vec, <2 x i64> undef, <2 x i32> <i32 0, i32 0>
7629 %2 = icmp sgt <2 x i64> %0, %1
7630 %3 = bitcast i8 %__u to <8 x i1>
7631 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <2 x i32> <i32 0, i32 1>
7632 %4 = and <2 x i1> %extract.i, %2
7633 %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
7634 %6 = bitcast <4 x i1> %5 to i4
7635 ret i4 %6
7636 }
7639 define zeroext i8 @test_vpcmpsgtq_v2i1_v8i1_mask(<2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
7640 ; VLX-LABEL: test_vpcmpsgtq_v2i1_v8i1_mask:
7641 ; VLX: # %bb.0: # %entry
7642 ; VLX-NEXT: vpcmpgtq %xmm1, %xmm0, %k0
7643 ; VLX-NEXT: kmovd %k0, %eax
7644 ; VLX-NEXT: # kill: def $al killed $al killed $eax
7645 ; VLX-NEXT: retq
7646 ;
7647 ; NoVLX-LABEL: test_vpcmpsgtq_v2i1_v8i1_mask:
7648 ; NoVLX: # %bb.0: # %entry
7649 ; NoVLX-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
7650 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
7651 ; NoVLX-NEXT: vpcmpgtq %zmm1, %zmm0, %k0
7652 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
7653 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
7654 ; NoVLX-NEXT: kmovw %k0, %eax
7655 ; NoVLX-NEXT: # kill: def $al killed $al killed $eax
7656 ; NoVLX-NEXT: vzeroupper
7657 ; NoVLX-NEXT: retq
7658 entry:
7659 %0 = bitcast <2 x i64> %__a to <2 x i64>
7660 %1 = bitcast <2 x i64> %__b to <2 x i64>
7661 %2 = icmp sgt <2 x i64> %0, %1
7662 %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
7663 %4 = bitcast <8 x i1> %3 to i8
7664 ret i8 %4
7665 }
7667 define zeroext i8 @test_vpcmpsgtq_v2i1_v8i1_mask_mem(<2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
7668 ; VLX-LABEL: test_vpcmpsgtq_v2i1_v8i1_mask_mem:
7669 ; VLX: # %bb.0: # %entry
7670 ; VLX-NEXT: vpcmpgtq (%rdi), %xmm0, %k0
7671 ; VLX-NEXT: kmovd %k0, %eax
7672 ; VLX-NEXT: # kill: def $al killed $al killed $eax
7673 ; VLX-NEXT: retq
7674 ;
7675 ; NoVLX-LABEL: test_vpcmpsgtq_v2i1_v8i1_mask_mem:
7676 ; NoVLX: # %bb.0: # %entry
7677 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
7678 ; NoVLX-NEXT: vmovdqa (%rdi), %xmm1
7679 ; NoVLX-NEXT: vpcmpgtq %zmm1, %zmm0, %k0
7680 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
7681 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
7682 ; NoVLX-NEXT: kmovw %k0, %eax
7683 ; NoVLX-NEXT: # kill: def $al killed $al killed $eax
7684 ; NoVLX-NEXT: vzeroupper
7685 ; NoVLX-NEXT: retq
7686 entry:
7687 %0 = bitcast <2 x i64> %__a to <2 x i64>
7688 %load = load <2 x i64>, <2 x i64>* %__b
7689 %1 = bitcast <2 x i64> %load to <2 x i64>
7690 %2 = icmp sgt <2 x i64> %0, %1
7691 %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
7692 %4 = bitcast <8 x i1> %3 to i8
7693 ret i8 %4
7694 }
7696 define zeroext i8 @test_masked_vpcmpsgtq_v2i1_v8i1_mask(i8 zeroext %__u, <2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
7697 ; VLX-LABEL: test_masked_vpcmpsgtq_v2i1_v8i1_mask:
7698 ; VLX: # %bb.0: # %entry
7699 ; VLX-NEXT: kmovd %edi, %k1
7700 ; VLX-NEXT: vpcmpgtq %xmm1, %xmm0, %k0 {%k1}
7701 ; VLX-NEXT: kmovd %k0, %eax
7702 ; VLX-NEXT: # kill: def $al killed $al killed $eax
7703 ; VLX-NEXT: retq
7704 ;
7705 ; NoVLX-LABEL: test_masked_vpcmpsgtq_v2i1_v8i1_mask:
7706 ; NoVLX: # %bb.0: # %entry
7707 ; NoVLX-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
7708 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
7709 ; NoVLX-NEXT: kmovw %edi, %k1
7710 ; NoVLX-NEXT: vpcmpgtq %zmm1, %zmm0, %k0 {%k1}
7711 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
7712 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
7713 ; NoVLX-NEXT: kmovw %k0, %eax
7714 ; NoVLX-NEXT: # kill: def $al killed $al killed $eax
7715 ; NoVLX-NEXT: vzeroupper
7716 ; NoVLX-NEXT: retq
7717 entry:
7718 %0 = bitcast <2 x i64> %__a to <2 x i64>
7719 %1 = bitcast <2 x i64> %__b to <2 x i64>
7720 %2 = icmp sgt <2 x i64> %0, %1
7721 %3 = bitcast i8 %__u to <8 x i1>
7722 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <2 x i32> <i32 0, i32 1>
7723 %4 = and <2 x i1> %2, %extract.i
7724 %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
7725 %6 = bitcast <8 x i1> %5 to i8
7726 ret i8 %6
7727 }
7729 define zeroext i8 @test_masked_vpcmpsgtq_v2i1_v8i1_mask_mem(i8 zeroext %__u, <2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
7730 ; VLX-LABEL: test_masked_vpcmpsgtq_v2i1_v8i1_mask_mem:
7731 ; VLX: # %bb.0: # %entry
7732 ; VLX-NEXT: kmovd %edi, %k1
7733 ; VLX-NEXT: vpcmpgtq (%rsi), %xmm0, %k0 {%k1}
7734 ; VLX-NEXT: kmovd %k0, %eax
7735 ; VLX-NEXT: # kill: def $al killed $al killed $eax
7736 ; VLX-NEXT: retq
7737 ;
7738 ; NoVLX-LABEL: test_masked_vpcmpsgtq_v2i1_v8i1_mask_mem:
7739 ; NoVLX: # %bb.0: # %entry
7740 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
7741 ; NoVLX-NEXT: vmovdqa (%rsi), %xmm1
7742 ; NoVLX-NEXT: kmovw %edi, %k1
7743 ; NoVLX-NEXT: vpcmpgtq %zmm1, %zmm0, %k0 {%k1}
7744 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
7745 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
7746 ; NoVLX-NEXT: kmovw %k0, %eax
7747 ; NoVLX-NEXT: # kill: def $al killed $al killed $eax
7748 ; NoVLX-NEXT: vzeroupper
7749 ; NoVLX-NEXT: retq
7750 entry:
7751 %0 = bitcast <2 x i64> %__a to <2 x i64>
7752 %load = load <2 x i64>, <2 x i64>* %__b
7753 %1 = bitcast <2 x i64> %load to <2 x i64>
7754 %2 = icmp sgt <2 x i64> %0, %1
7755 %3 = bitcast i8 %__u to <8 x i1>
7756 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <2 x i32> <i32 0, i32 1>
7757 %4 = and <2 x i1> %2, %extract.i
7758 %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
7759 %6 = bitcast <8 x i1> %5 to i8
7760 ret i8 %6
7761 }
7764 define zeroext i8 @test_vpcmpsgtq_v2i1_v8i1_mask_mem_b(<2 x i64> %__a, i64* %__b) local_unnamed_addr {
7765 ; VLX-LABEL: test_vpcmpsgtq_v2i1_v8i1_mask_mem_b:
7766 ; VLX: # %bb.0: # %entry
7767 ; VLX-NEXT: vpcmpgtq (%rdi){1to2}, %xmm0, %k0
7768 ; VLX-NEXT: kmovd %k0, %eax
7769 ; VLX-NEXT: # kill: def $al killed $al killed $eax
7770 ; VLX-NEXT: retq
7771 ;
7772 ; NoVLX-LABEL: test_vpcmpsgtq_v2i1_v8i1_mask_mem_b:
7773 ; NoVLX: # %bb.0: # %entry
7774 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
7775 ; NoVLX-NEXT: vpcmpgtq (%rdi){1to8}, %zmm0, %k0
7776 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
7777 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
7778 ; NoVLX-NEXT: kmovw %k0, %eax
7779 ; NoVLX-NEXT: # kill: def $al killed $al killed $eax
7780 ; NoVLX-NEXT: vzeroupper
7781 ; NoVLX-NEXT: retq
7782 entry:
7783 %0 = bitcast <2 x i64> %__a to <2 x i64>
7784 %load = load i64, i64* %__b
7785 %vec = insertelement <2 x i64> undef, i64 %load, i32 0
7786 %1 = shufflevector <2 x i64> %vec, <2 x i64> undef, <2 x i32> <i32 0, i32 0>
7787 %2 = icmp sgt <2 x i64> %0, %1
7788 %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
7789 %4 = bitcast <8 x i1> %3 to i8
7790 ret i8 %4
7791 }
7793 define zeroext i8 @test_masked_vpcmpsgtq_v2i1_v8i1_mask_mem_b(i8 zeroext %__u, <2 x i64> %__a, i64* %__b) local_unnamed_addr {
7794 ; VLX-LABEL: test_masked_vpcmpsgtq_v2i1_v8i1_mask_mem_b:
7795 ; VLX: # %bb.0: # %entry
7796 ; VLX-NEXT: kmovd %edi, %k1
7797 ; VLX-NEXT: vpcmpgtq (%rsi){1to2}, %xmm0, %k0 {%k1}
7798 ; VLX-NEXT: kmovd %k0, %eax
7799 ; VLX-NEXT: # kill: def $al killed $al killed $eax
7800 ; VLX-NEXT: retq
7801 ;
7802 ; NoVLX-LABEL: test_masked_vpcmpsgtq_v2i1_v8i1_mask_mem_b:
7803 ; NoVLX: # %bb.0: # %entry
7804 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
7805 ; NoVLX-NEXT: kmovw %edi, %k1
7806 ; NoVLX-NEXT: vpcmpgtq (%rsi){1to8}, %zmm0, %k0 {%k1}
7807 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
7808 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
7809 ; NoVLX-NEXT: kmovw %k0, %eax
7810 ; NoVLX-NEXT: # kill: def $al killed $al killed $eax
7811 ; NoVLX-NEXT: vzeroupper
7812 ; NoVLX-NEXT: retq
7813 entry:
7814 %0 = bitcast <2 x i64> %__a to <2 x i64>
7815 %load = load i64, i64* %__b
7816 %vec = insertelement <2 x i64> undef, i64 %load, i32 0
7817 %1 = shufflevector <2 x i64> %vec, <2 x i64> undef, <2 x i32> <i32 0, i32 0>
7818 %2 = icmp sgt <2 x i64> %0, %1
7819 %3 = bitcast i8 %__u to <8 x i1>
7820 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <2 x i32> <i32 0, i32 1>
7821 %4 = and <2 x i1> %extract.i, %2
7822 %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
7823 %6 = bitcast <8 x i1> %5 to i8
7824 ret i8 %6
7825 }
7828 define zeroext i16 @test_vpcmpsgtq_v2i1_v16i1_mask(<2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
7829 ; VLX-LABEL: test_vpcmpsgtq_v2i1_v16i1_mask:
7830 ; VLX: # %bb.0: # %entry
7831 ; VLX-NEXT: vpcmpgtq %xmm1, %xmm0, %k0
7832 ; VLX-NEXT: kmovd %k0, %eax
7833 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
7834 ; VLX-NEXT: retq
7835 ;
7836 ; NoVLX-LABEL: test_vpcmpsgtq_v2i1_v16i1_mask:
7837 ; NoVLX: # %bb.0: # %entry
7838 ; NoVLX-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
7839 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
7840 ; NoVLX-NEXT: vpcmpgtq %zmm1, %zmm0, %k0
7841 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
7842 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
7843 ; NoVLX-NEXT: kmovw %k0, %eax
7844 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
7845 ; NoVLX-NEXT: vzeroupper
7846 ; NoVLX-NEXT: retq
7847 entry:
7848 %0 = bitcast <2 x i64> %__a to <2 x i64>
7849 %1 = bitcast <2 x i64> %__b to <2 x i64>
7850 %2 = icmp sgt <2 x i64> %0, %1
7851 %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
7852 %4 = bitcast <16 x i1> %3 to i16
7853 ret i16 %4
7854 }
7856 define zeroext i16 @test_vpcmpsgtq_v2i1_v16i1_mask_mem(<2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
7857 ; VLX-LABEL: test_vpcmpsgtq_v2i1_v16i1_mask_mem:
7858 ; VLX: # %bb.0: # %entry
7859 ; VLX-NEXT: vpcmpgtq (%rdi), %xmm0, %k0
7860 ; VLX-NEXT: kmovd %k0, %eax
7861 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
7862 ; VLX-NEXT: retq
7863 ;
7864 ; NoVLX-LABEL: test_vpcmpsgtq_v2i1_v16i1_mask_mem:
7865 ; NoVLX: # %bb.0: # %entry
7866 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
7867 ; NoVLX-NEXT: vmovdqa (%rdi), %xmm1
7868 ; NoVLX-NEXT: vpcmpgtq %zmm1, %zmm0, %k0
7869 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
7870 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
7871 ; NoVLX-NEXT: kmovw %k0, %eax
7872 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
7873 ; NoVLX-NEXT: vzeroupper
7874 ; NoVLX-NEXT: retq
7875 entry:
7876 %0 = bitcast <2 x i64> %__a to <2 x i64>
7877 %load = load <2 x i64>, <2 x i64>* %__b
7878 %1 = bitcast <2 x i64> %load to <2 x i64>
7879 %2 = icmp sgt <2 x i64> %0, %1
7880 %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
7881 %4 = bitcast <16 x i1> %3 to i16
7882 ret i16 %4
7883 }
7885 define zeroext i16 @test_masked_vpcmpsgtq_v2i1_v16i1_mask(i8 zeroext %__u, <2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
7886 ; VLX-LABEL: test_masked_vpcmpsgtq_v2i1_v16i1_mask:
7887 ; VLX: # %bb.0: # %entry
7888 ; VLX-NEXT: kmovd %edi, %k1
7889 ; VLX-NEXT: vpcmpgtq %xmm1, %xmm0, %k0 {%k1}
7890 ; VLX-NEXT: kmovd %k0, %eax
7891 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
7892 ; VLX-NEXT: retq
7893 ;
7894 ; NoVLX-LABEL: test_masked_vpcmpsgtq_v2i1_v16i1_mask:
7895 ; NoVLX: # %bb.0: # %entry
7896 ; NoVLX-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
7897 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
7898 ; NoVLX-NEXT: kmovw %edi, %k1
7899 ; NoVLX-NEXT: vpcmpgtq %zmm1, %zmm0, %k0 {%k1}
7900 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
7901 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
7902 ; NoVLX-NEXT: kmovw %k0, %eax
7903 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
7904 ; NoVLX-NEXT: vzeroupper
7905 ; NoVLX-NEXT: retq
7906 entry:
7907 %0 = bitcast <2 x i64> %__a to <2 x i64>
7908 %1 = bitcast <2 x i64> %__b to <2 x i64>
7909 %2 = icmp sgt <2 x i64> %0, %1
7910 %3 = bitcast i8 %__u to <8 x i1>
7911 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <2 x i32> <i32 0, i32 1>
7912 %4 = and <2 x i1> %2, %extract.i
7913 %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
7914 %6 = bitcast <16 x i1> %5 to i16
7915 ret i16 %6
7916 }
7918 define zeroext i16 @test_masked_vpcmpsgtq_v2i1_v16i1_mask_mem(i8 zeroext %__u, <2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
7919 ; VLX-LABEL: test_masked_vpcmpsgtq_v2i1_v16i1_mask_mem:
7920 ; VLX: # %bb.0: # %entry
7921 ; VLX-NEXT: kmovd %edi, %k1
7922 ; VLX-NEXT: vpcmpgtq (%rsi), %xmm0, %k0 {%k1}
7923 ; VLX-NEXT: kmovd %k0, %eax
7924 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
7925 ; VLX-NEXT: retq
7926 ;
7927 ; NoVLX-LABEL: test_masked_vpcmpsgtq_v2i1_v16i1_mask_mem:
7928 ; NoVLX: # %bb.0: # %entry
7929 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
7930 ; NoVLX-NEXT: vmovdqa (%rsi), %xmm1
7931 ; NoVLX-NEXT: kmovw %edi, %k1
7932 ; NoVLX-NEXT: vpcmpgtq %zmm1, %zmm0, %k0 {%k1}
7933 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
7934 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
7935 ; NoVLX-NEXT: kmovw %k0, %eax
7936 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
7937 ; NoVLX-NEXT: vzeroupper
7938 ; NoVLX-NEXT: retq
7939 entry:
7940 %0 = bitcast <2 x i64> %__a to <2 x i64>
7941 %load = load <2 x i64>, <2 x i64>* %__b
7942 %1 = bitcast <2 x i64> %load to <2 x i64>
7943 %2 = icmp sgt <2 x i64> %0, %1
7944 %3 = bitcast i8 %__u to <8 x i1>
7945 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <2 x i32> <i32 0, i32 1>
7946 %4 = and <2 x i1> %2, %extract.i
7947 %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
7948 %6 = bitcast <16 x i1> %5 to i16
7949 ret i16 %6
7950 }
7953 define zeroext i16 @test_vpcmpsgtq_v2i1_v16i1_mask_mem_b(<2 x i64> %__a, i64* %__b) local_unnamed_addr {
7954 ; VLX-LABEL: test_vpcmpsgtq_v2i1_v16i1_mask_mem_b:
7955 ; VLX: # %bb.0: # %entry
7956 ; VLX-NEXT: vpcmpgtq (%rdi){1to2}, %xmm0, %k0
7957 ; VLX-NEXT: kmovd %k0, %eax
7958 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
7959 ; VLX-NEXT: retq
7960 ;
7961 ; NoVLX-LABEL: test_vpcmpsgtq_v2i1_v16i1_mask_mem_b:
7962 ; NoVLX: # %bb.0: # %entry
7963 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
7964 ; NoVLX-NEXT: vpcmpgtq (%rdi){1to8}, %zmm0, %k0
7965 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
7966 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
7967 ; NoVLX-NEXT: kmovw %k0, %eax
7968 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
7969 ; NoVLX-NEXT: vzeroupper
7970 ; NoVLX-NEXT: retq
7971 entry:
7972 %0 = bitcast <2 x i64> %__a to <2 x i64>
7973 %load = load i64, i64* %__b
7974 %vec = insertelement <2 x i64> undef, i64 %load, i32 0
7975 %1 = shufflevector <2 x i64> %vec, <2 x i64> undef, <2 x i32> <i32 0, i32 0>
7976 %2 = icmp sgt <2 x i64> %0, %1
7977 %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
7978 %4 = bitcast <16 x i1> %3 to i16
7979 ret i16 %4
7980 }
7982 define zeroext i16 @test_masked_vpcmpsgtq_v2i1_v16i1_mask_mem_b(i8 zeroext %__u, <2 x i64> %__a, i64* %__b) local_unnamed_addr {
7983 ; VLX-LABEL: test_masked_vpcmpsgtq_v2i1_v16i1_mask_mem_b:
7984 ; VLX: # %bb.0: # %entry
7985 ; VLX-NEXT: kmovd %edi, %k1
7986 ; VLX-NEXT: vpcmpgtq (%rsi){1to2}, %xmm0, %k0 {%k1}
7987 ; VLX-NEXT: kmovd %k0, %eax
7988 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
7989 ; VLX-NEXT: retq
7990 ;
7991 ; NoVLX-LABEL: test_masked_vpcmpsgtq_v2i1_v16i1_mask_mem_b:
7992 ; NoVLX: # %bb.0: # %entry
7993 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
7994 ; NoVLX-NEXT: kmovw %edi, %k1
7995 ; NoVLX-NEXT: vpcmpgtq (%rsi){1to8}, %zmm0, %k0 {%k1}
7996 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
7997 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
7998 ; NoVLX-NEXT: kmovw %k0, %eax
7999 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
8000 ; NoVLX-NEXT: vzeroupper
8001 ; NoVLX-NEXT: retq
8002 entry:
8003 %0 = bitcast <2 x i64> %__a to <2 x i64>
8004 %load = load i64, i64* %__b
8005 %vec = insertelement <2 x i64> undef, i64 %load, i32 0
8006 %1 = shufflevector <2 x i64> %vec, <2 x i64> undef, <2 x i32> <i32 0, i32 0>
8007 %2 = icmp sgt <2 x i64> %0, %1
8008 %3 = bitcast i8 %__u to <8 x i1>
8009 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <2 x i32> <i32 0, i32 1>
8010 %4 = and <2 x i1> %extract.i, %2
8011 %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
8012 %6 = bitcast <16 x i1> %5 to i16
8013 ret i16 %6
8014 }
8017 define zeroext i32 @test_vpcmpsgtq_v2i1_v32i1_mask(<2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
8018 ; VLX-LABEL: test_vpcmpsgtq_v2i1_v32i1_mask:
8019 ; VLX: # %bb.0: # %entry
8020 ; VLX-NEXT: vpcmpgtq %xmm1, %xmm0, %k0
8021 ; VLX-NEXT: kmovd %k0, %eax
8022 ; VLX-NEXT: retq
8023 ;
8024 ; NoVLX-LABEL: test_vpcmpsgtq_v2i1_v32i1_mask:
8025 ; NoVLX: # %bb.0: # %entry
8026 ; NoVLX-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
8027 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
8028 ; NoVLX-NEXT: vpcmpgtq %zmm1, %zmm0, %k0
8029 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
8030 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
8031 ; NoVLX-NEXT: kmovw %k0, %eax
8032 ; NoVLX-NEXT: vzeroupper
8033 ; NoVLX-NEXT: retq
8034 entry:
8035 %0 = bitcast <2 x i64> %__a to <2 x i64>
8036 %1 = bitcast <2 x i64> %__b to <2 x i64>
8037 %2 = icmp sgt <2 x i64> %0, %1
8038 %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
8039 %4 = bitcast <32 x i1> %3 to i32
8040 ret i32 %4
8041 }
8043 define zeroext i32 @test_vpcmpsgtq_v2i1_v32i1_mask_mem(<2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
8044 ; VLX-LABEL: test_vpcmpsgtq_v2i1_v32i1_mask_mem:
8045 ; VLX: # %bb.0: # %entry
8046 ; VLX-NEXT: vpcmpgtq (%rdi), %xmm0, %k0
8047 ; VLX-NEXT: kmovd %k0, %eax
8048 ; VLX-NEXT: retq
8049 ;
8050 ; NoVLX-LABEL: test_vpcmpsgtq_v2i1_v32i1_mask_mem:
8051 ; NoVLX: # %bb.0: # %entry
8052 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
8053 ; NoVLX-NEXT: vmovdqa (%rdi), %xmm1
8054 ; NoVLX-NEXT: vpcmpgtq %zmm1, %zmm0, %k0
8055 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
8056 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
8057 ; NoVLX-NEXT: kmovw %k0, %eax
8058 ; NoVLX-NEXT: vzeroupper
8059 ; NoVLX-NEXT: retq
8060 entry:
8061 %0 = bitcast <2 x i64> %__a to <2 x i64>
8062 %load = load <2 x i64>, <2 x i64>* %__b
8063 %1 = bitcast <2 x i64> %load to <2 x i64>
8064 %2 = icmp sgt <2 x i64> %0, %1
8065 %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
8066 %4 = bitcast <32 x i1> %3 to i32
8067 ret i32 %4
8068 }
8070 define zeroext i32 @test_masked_vpcmpsgtq_v2i1_v32i1_mask(i8 zeroext %__u, <2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
8071 ; VLX-LABEL: test_masked_vpcmpsgtq_v2i1_v32i1_mask:
8072 ; VLX: # %bb.0: # %entry
8073 ; VLX-NEXT: kmovd %edi, %k1
8074 ; VLX-NEXT: vpcmpgtq %xmm1, %xmm0, %k0 {%k1}
8075 ; VLX-NEXT: kmovd %k0, %eax
8076 ; VLX-NEXT: retq
8077 ;
8078 ; NoVLX-LABEL: test_masked_vpcmpsgtq_v2i1_v32i1_mask:
8079 ; NoVLX: # %bb.0: # %entry
8080 ; NoVLX-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
8081 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
8082 ; NoVLX-NEXT: kmovw %edi, %k1
8083 ; NoVLX-NEXT: vpcmpgtq %zmm1, %zmm0, %k0 {%k1}
8084 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
8085 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
8086 ; NoVLX-NEXT: kmovw %k0, %eax
8087 ; NoVLX-NEXT: vzeroupper
8088 ; NoVLX-NEXT: retq
8089 entry:
8090 %0 = bitcast <2 x i64> %__a to <2 x i64>
8091 %1 = bitcast <2 x i64> %__b to <2 x i64>
8092 %2 = icmp sgt <2 x i64> %0, %1
8093 %3 = bitcast i8 %__u to <8 x i1>
8094 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <2 x i32> <i32 0, i32 1>
8095 %4 = and <2 x i1> %2, %extract.i
8096 %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
8097 %6 = bitcast <32 x i1> %5 to i32
8098 ret i32 %6
8099 }
8101 define zeroext i32 @test_masked_vpcmpsgtq_v2i1_v32i1_mask_mem(i8 zeroext %__u, <2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
8102 ; VLX-LABEL: test_masked_vpcmpsgtq_v2i1_v32i1_mask_mem:
8103 ; VLX: # %bb.0: # %entry
8104 ; VLX-NEXT: kmovd %edi, %k1
8105 ; VLX-NEXT: vpcmpgtq (%rsi), %xmm0, %k0 {%k1}
8106 ; VLX-NEXT: kmovd %k0, %eax
8107 ; VLX-NEXT: retq
8108 ;
8109 ; NoVLX-LABEL: test_masked_vpcmpsgtq_v2i1_v32i1_mask_mem:
8110 ; NoVLX: # %bb.0: # %entry
8111 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
8112 ; NoVLX-NEXT: vmovdqa (%rsi), %xmm1
8113 ; NoVLX-NEXT: kmovw %edi, %k1
8114 ; NoVLX-NEXT: vpcmpgtq %zmm1, %zmm0, %k0 {%k1}
8115 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
8116 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
8117 ; NoVLX-NEXT: kmovw %k0, %eax
8118 ; NoVLX-NEXT: vzeroupper
8119 ; NoVLX-NEXT: retq
8120 entry:
8121 %0 = bitcast <2 x i64> %__a to <2 x i64>
8122 %load = load <2 x i64>, <2 x i64>* %__b
8123 %1 = bitcast <2 x i64> %load to <2 x i64>
8124 %2 = icmp sgt <2 x i64> %0, %1
8125 %3 = bitcast i8 %__u to <8 x i1>
8126 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <2 x i32> <i32 0, i32 1>
8127 %4 = and <2 x i1> %2, %extract.i
8128 %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
8129 %6 = bitcast <32 x i1> %5 to i32
8130 ret i32 %6
8131 }
8134 define zeroext i32 @test_vpcmpsgtq_v2i1_v32i1_mask_mem_b(<2 x i64> %__a, i64* %__b) local_unnamed_addr {
8135 ; VLX-LABEL: test_vpcmpsgtq_v2i1_v32i1_mask_mem_b:
8136 ; VLX: # %bb.0: # %entry
8137 ; VLX-NEXT: vpcmpgtq (%rdi){1to2}, %xmm0, %k0
8138 ; VLX-NEXT: kmovd %k0, %eax
8139 ; VLX-NEXT: retq
8140 ;
8141 ; NoVLX-LABEL: test_vpcmpsgtq_v2i1_v32i1_mask_mem_b:
8142 ; NoVLX: # %bb.0: # %entry
8143 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
8144 ; NoVLX-NEXT: vpcmpgtq (%rdi){1to8}, %zmm0, %k0
8145 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
8146 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
8147 ; NoVLX-NEXT: kmovw %k0, %eax
8148 ; NoVLX-NEXT: vzeroupper
8149 ; NoVLX-NEXT: retq
8150 entry:
8151 %0 = bitcast <2 x i64> %__a to <2 x i64>
8152 %load = load i64, i64* %__b
8153 %vec = insertelement <2 x i64> undef, i64 %load, i32 0
8154 %1 = shufflevector <2 x i64> %vec, <2 x i64> undef, <2 x i32> <i32 0, i32 0>
8155 %2 = icmp sgt <2 x i64> %0, %1
8156 %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
8157 %4 = bitcast <32 x i1> %3 to i32
8158 ret i32 %4
8159 }
8161 define zeroext i32 @test_masked_vpcmpsgtq_v2i1_v32i1_mask_mem_b(i8 zeroext %__u, <2 x i64> %__a, i64* %__b) local_unnamed_addr {
8162 ; VLX-LABEL: test_masked_vpcmpsgtq_v2i1_v32i1_mask_mem_b:
8163 ; VLX: # %bb.0: # %entry
8164 ; VLX-NEXT: kmovd %edi, %k1
8165 ; VLX-NEXT: vpcmpgtq (%rsi){1to2}, %xmm0, %k0 {%k1}
8166 ; VLX-NEXT: kmovd %k0, %eax
8167 ; VLX-NEXT: retq
8168 ;
8169 ; NoVLX-LABEL: test_masked_vpcmpsgtq_v2i1_v32i1_mask_mem_b:
8170 ; NoVLX: # %bb.0: # %entry
8171 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
8172 ; NoVLX-NEXT: kmovw %edi, %k1
8173 ; NoVLX-NEXT: vpcmpgtq (%rsi){1to8}, %zmm0, %k0 {%k1}
8174 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
8175 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
8176 ; NoVLX-NEXT: kmovw %k0, %eax
8177 ; NoVLX-NEXT: vzeroupper
8178 ; NoVLX-NEXT: retq
8179 entry:
8180 %0 = bitcast <2 x i64> %__a to <2 x i64>
8181 %load = load i64, i64* %__b
8182 %vec = insertelement <2 x i64> undef, i64 %load, i32 0
8183 %1 = shufflevector <2 x i64> %vec, <2 x i64> undef, <2 x i32> <i32 0, i32 0>
8184 %2 = icmp sgt <2 x i64> %0, %1
8185 %3 = bitcast i8 %__u to <8 x i1>
8186 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <2 x i32> <i32 0, i32 1>
8187 %4 = and <2 x i1> %extract.i, %2
8188 %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
8189 %6 = bitcast <32 x i1> %5 to i32
8190 ret i32 %6
8191 }
8194 define zeroext i64 @test_vpcmpsgtq_v2i1_v64i1_mask(<2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
8195 ; VLX-LABEL: test_vpcmpsgtq_v2i1_v64i1_mask:
8196 ; VLX: # %bb.0: # %entry
8197 ; VLX-NEXT: vpcmpgtq %xmm1, %xmm0, %k0
8198 ; VLX-NEXT: kmovq %k0, %rax
8199 ; VLX-NEXT: retq
8200 ;
8201 ; NoVLX-LABEL: test_vpcmpsgtq_v2i1_v64i1_mask:
8202 ; NoVLX: # %bb.0: # %entry
8203 ; NoVLX-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
8204 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
8205 ; NoVLX-NEXT: vpcmpgtq %zmm1, %zmm0, %k0
8206 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
8207 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
8208 ; NoVLX-NEXT: kmovw %k0, %eax
8209 ; NoVLX-NEXT: vzeroupper
8210 ; NoVLX-NEXT: retq
8211 entry:
8212 %0 = bitcast <2 x i64> %__a to <2 x i64>
8213 %1 = bitcast <2 x i64> %__b to <2 x i64>
8214 %2 = icmp sgt <2 x i64> %0, %1
8215 %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
8216 %4 = bitcast <64 x i1> %3 to i64
8217 ret i64 %4
8218 }
8220 define zeroext i64 @test_vpcmpsgtq_v2i1_v64i1_mask_mem(<2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
8221 ; VLX-LABEL: test_vpcmpsgtq_v2i1_v64i1_mask_mem:
8222 ; VLX: # %bb.0: # %entry
8223 ; VLX-NEXT: vpcmpgtq (%rdi), %xmm0, %k0
8224 ; VLX-NEXT: kmovq %k0, %rax
8225 ; VLX-NEXT: retq
8226 ;
8227 ; NoVLX-LABEL: test_vpcmpsgtq_v2i1_v64i1_mask_mem:
8228 ; NoVLX: # %bb.0: # %entry
8229 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
8230 ; NoVLX-NEXT: vmovdqa (%rdi), %xmm1
8231 ; NoVLX-NEXT: vpcmpgtq %zmm1, %zmm0, %k0
8232 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
8233 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
8234 ; NoVLX-NEXT: kmovw %k0, %eax
8235 ; NoVLX-NEXT: vzeroupper
8236 ; NoVLX-NEXT: retq
8237 entry:
8238 %0 = bitcast <2 x i64> %__a to <2 x i64>
8239 %load = load <2 x i64>, <2 x i64>* %__b
8240 %1 = bitcast <2 x i64> %load to <2 x i64>
8241 %2 = icmp sgt <2 x i64> %0, %1
8242 %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
8243 %4 = bitcast <64 x i1> %3 to i64
8244 ret i64 %4
8245 }
8247 define zeroext i64 @test_masked_vpcmpsgtq_v2i1_v64i1_mask(i8 zeroext %__u, <2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
8248 ; VLX-LABEL: test_masked_vpcmpsgtq_v2i1_v64i1_mask:
8249 ; VLX: # %bb.0: # %entry
8250 ; VLX-NEXT: kmovd %edi, %k1
8251 ; VLX-NEXT: vpcmpgtq %xmm1, %xmm0, %k0 {%k1}
8252 ; VLX-NEXT: kmovq %k0, %rax
8253 ; VLX-NEXT: retq
8254 ;
8255 ; NoVLX-LABEL: test_masked_vpcmpsgtq_v2i1_v64i1_mask:
8256 ; NoVLX: # %bb.0: # %entry
8257 ; NoVLX-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
8258 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
8259 ; NoVLX-NEXT: kmovw %edi, %k1
8260 ; NoVLX-NEXT: vpcmpgtq %zmm1, %zmm0, %k0 {%k1}
8261 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
8262 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
8263 ; NoVLX-NEXT: kmovw %k0, %eax
8264 ; NoVLX-NEXT: vzeroupper
8265 ; NoVLX-NEXT: retq
8266 entry:
8267 %0 = bitcast <2 x i64> %__a to <2 x i64>
8268 %1 = bitcast <2 x i64> %__b to <2 x i64>
8269 %2 = icmp sgt <2 x i64> %0, %1
8270 %3 = bitcast i8 %__u to <8 x i1>
8271 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <2 x i32> <i32 0, i32 1>
8272 %4 = and <2 x i1> %2, %extract.i
8273 %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
8274 %6 = bitcast <64 x i1> %5 to i64
8275 ret i64 %6
8276 }
8278 define zeroext i64 @test_masked_vpcmpsgtq_v2i1_v64i1_mask_mem(i8 zeroext %__u, <2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
8279 ; VLX-LABEL: test_masked_vpcmpsgtq_v2i1_v64i1_mask_mem:
8280 ; VLX: # %bb.0: # %entry
8281 ; VLX-NEXT: kmovd %edi, %k1
8282 ; VLX-NEXT: vpcmpgtq (%rsi), %xmm0, %k0 {%k1}
8283 ; VLX-NEXT: kmovq %k0, %rax
8284 ; VLX-NEXT: retq
8285 ;
8286 ; NoVLX-LABEL: test_masked_vpcmpsgtq_v2i1_v64i1_mask_mem:
8287 ; NoVLX: # %bb.0: # %entry
8288 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
8289 ; NoVLX-NEXT: vmovdqa (%rsi), %xmm1
8290 ; NoVLX-NEXT: kmovw %edi, %k1
8291 ; NoVLX-NEXT: vpcmpgtq %zmm1, %zmm0, %k0 {%k1}
8292 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
8293 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
8294 ; NoVLX-NEXT: kmovw %k0, %eax
8295 ; NoVLX-NEXT: vzeroupper
8296 ; NoVLX-NEXT: retq
8297 entry:
8298 %0 = bitcast <2 x i64> %__a to <2 x i64>
8299 %load = load <2 x i64>, <2 x i64>* %__b
8300 %1 = bitcast <2 x i64> %load to <2 x i64>
8301 %2 = icmp sgt <2 x i64> %0, %1
8302 %3 = bitcast i8 %__u to <8 x i1>
8303 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <2 x i32> <i32 0, i32 1>
8304 %4 = and <2 x i1> %2, %extract.i
8305 %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
8306 %6 = bitcast <64 x i1> %5 to i64
8307 ret i64 %6
8308 }
8311 define zeroext i64 @test_vpcmpsgtq_v2i1_v64i1_mask_mem_b(<2 x i64> %__a, i64* %__b) local_unnamed_addr {
8312 ; VLX-LABEL: test_vpcmpsgtq_v2i1_v64i1_mask_mem_b:
8313 ; VLX: # %bb.0: # %entry
8314 ; VLX-NEXT: vpcmpgtq (%rdi){1to2}, %xmm0, %k0
8315 ; VLX-NEXT: kmovq %k0, %rax
8316 ; VLX-NEXT: retq
8317 ;
8318 ; NoVLX-LABEL: test_vpcmpsgtq_v2i1_v64i1_mask_mem_b:
8319 ; NoVLX: # %bb.0: # %entry
8320 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
8321 ; NoVLX-NEXT: vpcmpgtq (%rdi){1to8}, %zmm0, %k0
8322 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
8323 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
8324 ; NoVLX-NEXT: kmovw %k0, %eax
8325 ; NoVLX-NEXT: vzeroupper
8326 ; NoVLX-NEXT: retq
8327 entry:
8328 %0 = bitcast <2 x i64> %__a to <2 x i64>
8329 %load = load i64, i64* %__b
8330 %vec = insertelement <2 x i64> undef, i64 %load, i32 0
8331 %1 = shufflevector <2 x i64> %vec, <2 x i64> undef, <2 x i32> <i32 0, i32 0>
8332 %2 = icmp sgt <2 x i64> %0, %1
8333 %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
8334 %4 = bitcast <64 x i1> %3 to i64
8335 ret i64 %4
8336 }
8338 define zeroext i64 @test_masked_vpcmpsgtq_v2i1_v64i1_mask_mem_b(i8 zeroext %__u, <2 x i64> %__a, i64* %__b) local_unnamed_addr {
8339 ; VLX-LABEL: test_masked_vpcmpsgtq_v2i1_v64i1_mask_mem_b:
8340 ; VLX: # %bb.0: # %entry
8341 ; VLX-NEXT: kmovd %edi, %k1
8342 ; VLX-NEXT: vpcmpgtq (%rsi){1to2}, %xmm0, %k0 {%k1}
8343 ; VLX-NEXT: kmovq %k0, %rax
8344 ; VLX-NEXT: retq
8345 ;
8346 ; NoVLX-LABEL: test_masked_vpcmpsgtq_v2i1_v64i1_mask_mem_b:
8347 ; NoVLX: # %bb.0: # %entry
8348 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
8349 ; NoVLX-NEXT: kmovw %edi, %k1
8350 ; NoVLX-NEXT: vpcmpgtq (%rsi){1to8}, %zmm0, %k0 {%k1}
8351 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
8352 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
8353 ; NoVLX-NEXT: kmovw %k0, %eax
8354 ; NoVLX-NEXT: vzeroupper
8355 ; NoVLX-NEXT: retq
8356 entry:
8357 %0 = bitcast <2 x i64> %__a to <2 x i64>
8358 %load = load i64, i64* %__b
8359 %vec = insertelement <2 x i64> undef, i64 %load, i32 0
8360 %1 = shufflevector <2 x i64> %vec, <2 x i64> undef, <2 x i32> <i32 0, i32 0>
8361 %2 = icmp sgt <2 x i64> %0, %1
8362 %3 = bitcast i8 %__u to <8 x i1>
8363 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <2 x i32> <i32 0, i32 1>
8364 %4 = and <2 x i1> %extract.i, %2
8365 %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
8366 %6 = bitcast <64 x i1> %5 to i64
8367 ret i64 %6
8368 }
8371 define zeroext i8 @test_vpcmpsgtq_v4i1_v8i1_mask(<4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
8372 ; VLX-LABEL: test_vpcmpsgtq_v4i1_v8i1_mask:
8373 ; VLX: # %bb.0: # %entry
8374 ; VLX-NEXT: vpcmpgtq %ymm1, %ymm0, %k0
8375 ; VLX-NEXT: kmovd %k0, %eax
8376 ; VLX-NEXT: # kill: def $al killed $al killed $eax
8377 ; VLX-NEXT: vzeroupper
8378 ; VLX-NEXT: retq
8379 ;
8380 ; NoVLX-LABEL: test_vpcmpsgtq_v4i1_v8i1_mask:
8381 ; NoVLX: # %bb.0: # %entry
8382 ; NoVLX-NEXT: # kill: def $ymm1 killed $ymm1 def $zmm1
8383 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
8384 ; NoVLX-NEXT: vpcmpgtq %zmm1, %zmm0, %k0
8385 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
8386 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
8387 ; NoVLX-NEXT: kmovw %k0, %eax
8388 ; NoVLX-NEXT: # kill: def $al killed $al killed $eax
8389 ; NoVLX-NEXT: vzeroupper
8390 ; NoVLX-NEXT: retq
8391 entry:
8392 %0 = bitcast <4 x i64> %__a to <4 x i64>
8393 %1 = bitcast <4 x i64> %__b to <4 x i64>
8394 %2 = icmp sgt <4 x i64> %0, %1
8395 %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
8396 %4 = bitcast <8 x i1> %3 to i8
8397 ret i8 %4
8398 }
8400 define zeroext i8 @test_vpcmpsgtq_v4i1_v8i1_mask_mem(<4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
8401 ; VLX-LABEL: test_vpcmpsgtq_v4i1_v8i1_mask_mem:
8402 ; VLX: # %bb.0: # %entry
8403 ; VLX-NEXT: vpcmpgtq (%rdi), %ymm0, %k0
8404 ; VLX-NEXT: kmovd %k0, %eax
8405 ; VLX-NEXT: # kill: def $al killed $al killed $eax
8406 ; VLX-NEXT: vzeroupper
8407 ; VLX-NEXT: retq
8408 ;
8409 ; NoVLX-LABEL: test_vpcmpsgtq_v4i1_v8i1_mask_mem:
8410 ; NoVLX: # %bb.0: # %entry
8411 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
8412 ; NoVLX-NEXT: vmovdqa (%rdi), %ymm1
8413 ; NoVLX-NEXT: vpcmpgtq %zmm1, %zmm0, %k0
8414 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
8415 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
8416 ; NoVLX-NEXT: kmovw %k0, %eax
8417 ; NoVLX-NEXT: # kill: def $al killed $al killed $eax
8418 ; NoVLX-NEXT: vzeroupper
8419 ; NoVLX-NEXT: retq
8420 entry:
8421 %0 = bitcast <4 x i64> %__a to <4 x i64>
8422 %load = load <4 x i64>, <4 x i64>* %__b
8423 %1 = bitcast <4 x i64> %load to <4 x i64>
8424 %2 = icmp sgt <4 x i64> %0, %1
8425 %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
8426 %4 = bitcast <8 x i1> %3 to i8
8427 ret i8 %4
8428 }
8430 define zeroext i8 @test_masked_vpcmpsgtq_v4i1_v8i1_mask(i8 zeroext %__u, <4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
8431 ; VLX-LABEL: test_masked_vpcmpsgtq_v4i1_v8i1_mask:
8432 ; VLX: # %bb.0: # %entry
8433 ; VLX-NEXT: kmovd %edi, %k1
8434 ; VLX-NEXT: vpcmpgtq %ymm1, %ymm0, %k0 {%k1}
8435 ; VLX-NEXT: kmovd %k0, %eax
8436 ; VLX-NEXT: # kill: def $al killed $al killed $eax
8437 ; VLX-NEXT: vzeroupper
8438 ; VLX-NEXT: retq
8439 ;
8440 ; NoVLX-LABEL: test_masked_vpcmpsgtq_v4i1_v8i1_mask:
8441 ; NoVLX: # %bb.0: # %entry
8442 ; NoVLX-NEXT: # kill: def $ymm1 killed $ymm1 def $zmm1
8443 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
8444 ; NoVLX-NEXT: kmovw %edi, %k1
8445 ; NoVLX-NEXT: vpcmpgtq %zmm1, %zmm0, %k0 {%k1}
8446 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
8447 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
8448 ; NoVLX-NEXT: kmovw %k0, %eax
8449 ; NoVLX-NEXT: # kill: def $al killed $al killed $eax
8450 ; NoVLX-NEXT: vzeroupper
8451 ; NoVLX-NEXT: retq
8452 entry:
8453 %0 = bitcast <4 x i64> %__a to <4 x i64>
8454 %1 = bitcast <4 x i64> %__b to <4 x i64>
8455 %2 = icmp sgt <4 x i64> %0, %1
8456 %3 = bitcast i8 %__u to <8 x i1>
8457 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
8458 %4 = and <4 x i1> %2, %extract.i
8459 %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
8460 %6 = bitcast <8 x i1> %5 to i8
8461 ret i8 %6
8462 }
8464 define zeroext i8 @test_masked_vpcmpsgtq_v4i1_v8i1_mask_mem(i8 zeroext %__u, <4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
8465 ; VLX-LABEL: test_masked_vpcmpsgtq_v4i1_v8i1_mask_mem:
8466 ; VLX: # %bb.0: # %entry
8467 ; VLX-NEXT: kmovd %edi, %k1
8468 ; VLX-NEXT: vpcmpgtq (%rsi), %ymm0, %k0 {%k1}
8469 ; VLX-NEXT: kmovd %k0, %eax
8470 ; VLX-NEXT: # kill: def $al killed $al killed $eax
8471 ; VLX-NEXT: vzeroupper
8472 ; VLX-NEXT: retq
8473 ;
8474 ; NoVLX-LABEL: test_masked_vpcmpsgtq_v4i1_v8i1_mask_mem:
8475 ; NoVLX: # %bb.0: # %entry
8476 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
8477 ; NoVLX-NEXT: vmovdqa (%rsi), %ymm1
8478 ; NoVLX-NEXT: kmovw %edi, %k1
8479 ; NoVLX-NEXT: vpcmpgtq %zmm1, %zmm0, %k0 {%k1}
8480 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
8481 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
8482 ; NoVLX-NEXT: kmovw %k0, %eax
8483 ; NoVLX-NEXT: # kill: def $al killed $al killed $eax
8484 ; NoVLX-NEXT: vzeroupper
8485 ; NoVLX-NEXT: retq
8486 entry:
8487 %0 = bitcast <4 x i64> %__a to <4 x i64>
8488 %load = load <4 x i64>, <4 x i64>* %__b
8489 %1 = bitcast <4 x i64> %load to <4 x i64>
8490 %2 = icmp sgt <4 x i64> %0, %1
8491 %3 = bitcast i8 %__u to <8 x i1>
8492 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
8493 %4 = and <4 x i1> %2, %extract.i
8494 %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
8495 %6 = bitcast <8 x i1> %5 to i8
8496 ret i8 %6
8497 }
8500 define zeroext i8 @test_vpcmpsgtq_v4i1_v8i1_mask_mem_b(<4 x i64> %__a, i64* %__b) local_unnamed_addr {
8501 ; VLX-LABEL: test_vpcmpsgtq_v4i1_v8i1_mask_mem_b:
8502 ; VLX: # %bb.0: # %entry
8503 ; VLX-NEXT: vpcmpgtq (%rdi){1to4}, %ymm0, %k0
8504 ; VLX-NEXT: kmovd %k0, %eax
8505 ; VLX-NEXT: # kill: def $al killed $al killed $eax
8506 ; VLX-NEXT: vzeroupper
8507 ; VLX-NEXT: retq
8508 ;
8509 ; NoVLX-LABEL: test_vpcmpsgtq_v4i1_v8i1_mask_mem_b:
8510 ; NoVLX: # %bb.0: # %entry
8511 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
8512 ; NoVLX-NEXT: vpcmpgtq (%rdi){1to8}, %zmm0, %k0
8513 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
8514 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
8515 ; NoVLX-NEXT: kmovw %k0, %eax
8516 ; NoVLX-NEXT: # kill: def $al killed $al killed $eax
8517 ; NoVLX-NEXT: vzeroupper
8518 ; NoVLX-NEXT: retq
8519 entry:
8520 %0 = bitcast <4 x i64> %__a to <4 x i64>
8521 %load = load i64, i64* %__b
8522 %vec = insertelement <4 x i64> undef, i64 %load, i32 0
8523 %1 = shufflevector <4 x i64> %vec, <4 x i64> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
8524 %2 = icmp sgt <4 x i64> %0, %1
8525 %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
8526 %4 = bitcast <8 x i1> %3 to i8
8527 ret i8 %4
8528 }
8530 define zeroext i8 @test_masked_vpcmpsgtq_v4i1_v8i1_mask_mem_b(i8 zeroext %__u, <4 x i64> %__a, i64* %__b) local_unnamed_addr {
8531 ; VLX-LABEL: test_masked_vpcmpsgtq_v4i1_v8i1_mask_mem_b:
8532 ; VLX: # %bb.0: # %entry
8533 ; VLX-NEXT: kmovd %edi, %k1
8534 ; VLX-NEXT: vpcmpgtq (%rsi){1to4}, %ymm0, %k0 {%k1}
8535 ; VLX-NEXT: kmovd %k0, %eax
8536 ; VLX-NEXT: # kill: def $al killed $al killed $eax
8537 ; VLX-NEXT: vzeroupper
8538 ; VLX-NEXT: retq
8539 ;
8540 ; NoVLX-LABEL: test_masked_vpcmpsgtq_v4i1_v8i1_mask_mem_b:
8541 ; NoVLX: # %bb.0: # %entry
8542 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
8543 ; NoVLX-NEXT: kmovw %edi, %k1
8544 ; NoVLX-NEXT: vpcmpgtq (%rsi){1to8}, %zmm0, %k0 {%k1}
8545 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
8546 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
8547 ; NoVLX-NEXT: kmovw %k0, %eax
8548 ; NoVLX-NEXT: # kill: def $al killed $al killed $eax
8549 ; NoVLX-NEXT: vzeroupper
8550 ; NoVLX-NEXT: retq
8551 entry:
8552 %0 = bitcast <4 x i64> %__a to <4 x i64>
8553 %load = load i64, i64* %__b
8554 %vec = insertelement <4 x i64> undef, i64 %load, i32 0
8555 %1 = shufflevector <4 x i64> %vec, <4 x i64> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
8556 %2 = icmp sgt <4 x i64> %0, %1
8557 %3 = bitcast i8 %__u to <8 x i1>
8558 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
8559 %4 = and <4 x i1> %extract.i, %2
8560 %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
8561 %6 = bitcast <8 x i1> %5 to i8
8562 ret i8 %6
8563 }
8566 define zeroext i16 @test_vpcmpsgtq_v4i1_v16i1_mask(<4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
8567 ; VLX-LABEL: test_vpcmpsgtq_v4i1_v16i1_mask:
8568 ; VLX: # %bb.0: # %entry
8569 ; VLX-NEXT: vpcmpgtq %ymm1, %ymm0, %k0
8570 ; VLX-NEXT: kmovd %k0, %eax
8571 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
8572 ; VLX-NEXT: vzeroupper
8573 ; VLX-NEXT: retq
8574 ;
8575 ; NoVLX-LABEL: test_vpcmpsgtq_v4i1_v16i1_mask:
8576 ; NoVLX: # %bb.0: # %entry
8577 ; NoVLX-NEXT: # kill: def $ymm1 killed $ymm1 def $zmm1
8578 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
8579 ; NoVLX-NEXT: vpcmpgtq %zmm1, %zmm0, %k0
8580 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
8581 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
8582 ; NoVLX-NEXT: kmovw %k0, %eax
8583 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
8584 ; NoVLX-NEXT: vzeroupper
8585 ; NoVLX-NEXT: retq
8586 entry:
8587 %0 = bitcast <4 x i64> %__a to <4 x i64>
8588 %1 = bitcast <4 x i64> %__b to <4 x i64>
8589 %2 = icmp sgt <4 x i64> %0, %1
8590 %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
8591 %4 = bitcast <16 x i1> %3 to i16
8592 ret i16 %4
8593 }
8595 define zeroext i16 @test_vpcmpsgtq_v4i1_v16i1_mask_mem(<4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
8596 ; VLX-LABEL: test_vpcmpsgtq_v4i1_v16i1_mask_mem:
8597 ; VLX: # %bb.0: # %entry
8598 ; VLX-NEXT: vpcmpgtq (%rdi), %ymm0, %k0
8599 ; VLX-NEXT: kmovd %k0, %eax
8600 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
8601 ; VLX-NEXT: vzeroupper
8602 ; VLX-NEXT: retq
8603 ;
8604 ; NoVLX-LABEL: test_vpcmpsgtq_v4i1_v16i1_mask_mem:
8605 ; NoVLX: # %bb.0: # %entry
8606 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
8607 ; NoVLX-NEXT: vmovdqa (%rdi), %ymm1
8608 ; NoVLX-NEXT: vpcmpgtq %zmm1, %zmm0, %k0
8609 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
8610 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
8611 ; NoVLX-NEXT: kmovw %k0, %eax
8612 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
8613 ; NoVLX-NEXT: vzeroupper
8614 ; NoVLX-NEXT: retq
8615 entry:
8616 %0 = bitcast <4 x i64> %__a to <4 x i64>
8617 %load = load <4 x i64>, <4 x i64>* %__b
8618 %1 = bitcast <4 x i64> %load to <4 x i64>
8619 %2 = icmp sgt <4 x i64> %0, %1
8620 %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
8621 %4 = bitcast <16 x i1> %3 to i16
8622 ret i16 %4
8623 }
8625 define zeroext i16 @test_masked_vpcmpsgtq_v4i1_v16i1_mask(i8 zeroext %__u, <4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
8626 ; VLX-LABEL: test_masked_vpcmpsgtq_v4i1_v16i1_mask:
8627 ; VLX: # %bb.0: # %entry
8628 ; VLX-NEXT: kmovd %edi, %k1
8629 ; VLX-NEXT: vpcmpgtq %ymm1, %ymm0, %k0 {%k1}
8630 ; VLX-NEXT: kmovd %k0, %eax
8631 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
8632 ; VLX-NEXT: vzeroupper
8633 ; VLX-NEXT: retq
8634 ;
8635 ; NoVLX-LABEL: test_masked_vpcmpsgtq_v4i1_v16i1_mask:
8636 ; NoVLX: # %bb.0: # %entry
8637 ; NoVLX-NEXT: # kill: def $ymm1 killed $ymm1 def $zmm1
8638 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
8639 ; NoVLX-NEXT: kmovw %edi, %k1
8640 ; NoVLX-NEXT: vpcmpgtq %zmm1, %zmm0, %k0 {%k1}
8641 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
8642 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
8643 ; NoVLX-NEXT: kmovw %k0, %eax
8644 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
8645 ; NoVLX-NEXT: vzeroupper
8646 ; NoVLX-NEXT: retq
8647 entry:
8648 %0 = bitcast <4 x i64> %__a to <4 x i64>
8649 %1 = bitcast <4 x i64> %__b to <4 x i64>
8650 %2 = icmp sgt <4 x i64> %0, %1
8651 %3 = bitcast i8 %__u to <8 x i1>
8652 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
8653 %4 = and <4 x i1> %2, %extract.i
8654 %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
8655 %6 = bitcast <16 x i1> %5 to i16
8656 ret i16 %6
8657 }
8659 define zeroext i16 @test_masked_vpcmpsgtq_v4i1_v16i1_mask_mem(i8 zeroext %__u, <4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
8660 ; VLX-LABEL: test_masked_vpcmpsgtq_v4i1_v16i1_mask_mem:
8661 ; VLX: # %bb.0: # %entry
8662 ; VLX-NEXT: kmovd %edi, %k1
8663 ; VLX-NEXT: vpcmpgtq (%rsi), %ymm0, %k0 {%k1}
8664 ; VLX-NEXT: kmovd %k0, %eax
8665 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
8666 ; VLX-NEXT: vzeroupper
8667 ; VLX-NEXT: retq
8668 ;
8669 ; NoVLX-LABEL: test_masked_vpcmpsgtq_v4i1_v16i1_mask_mem:
8670 ; NoVLX: # %bb.0: # %entry
8671 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
8672 ; NoVLX-NEXT: vmovdqa (%rsi), %ymm1
8673 ; NoVLX-NEXT: kmovw %edi, %k1
8674 ; NoVLX-NEXT: vpcmpgtq %zmm1, %zmm0, %k0 {%k1}
8675 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
8676 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
8677 ; NoVLX-NEXT: kmovw %k0, %eax
8678 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
8679 ; NoVLX-NEXT: vzeroupper
8680 ; NoVLX-NEXT: retq
8681 entry:
8682 %0 = bitcast <4 x i64> %__a to <4 x i64>
8683 %load = load <4 x i64>, <4 x i64>* %__b
8684 %1 = bitcast <4 x i64> %load to <4 x i64>
8685 %2 = icmp sgt <4 x i64> %0, %1
8686 %3 = bitcast i8 %__u to <8 x i1>
8687 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
8688 %4 = and <4 x i1> %2, %extract.i
8689 %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
8690 %6 = bitcast <16 x i1> %5 to i16
8691 ret i16 %6
8692 }
8695 define zeroext i16 @test_vpcmpsgtq_v4i1_v16i1_mask_mem_b(<4 x i64> %__a, i64* %__b) local_unnamed_addr {
8696 ; VLX-LABEL: test_vpcmpsgtq_v4i1_v16i1_mask_mem_b:
8697 ; VLX: # %bb.0: # %entry
8698 ; VLX-NEXT: vpcmpgtq (%rdi){1to4}, %ymm0, %k0
8699 ; VLX-NEXT: kmovd %k0, %eax
8700 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
8701 ; VLX-NEXT: vzeroupper
8702 ; VLX-NEXT: retq
8703 ;
8704 ; NoVLX-LABEL: test_vpcmpsgtq_v4i1_v16i1_mask_mem_b:
8705 ; NoVLX: # %bb.0: # %entry
8706 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
8707 ; NoVLX-NEXT: vpcmpgtq (%rdi){1to8}, %zmm0, %k0
8708 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
8709 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
8710 ; NoVLX-NEXT: kmovw %k0, %eax
8711 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
8712 ; NoVLX-NEXT: vzeroupper
8713 ; NoVLX-NEXT: retq
8714 entry:
8715 %0 = bitcast <4 x i64> %__a to <4 x i64>
8716 %load = load i64, i64* %__b
8717 %vec = insertelement <4 x i64> undef, i64 %load, i32 0
8718 %1 = shufflevector <4 x i64> %vec, <4 x i64> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
8719 %2 = icmp sgt <4 x i64> %0, %1
8720 %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
8721 %4 = bitcast <16 x i1> %3 to i16
8722 ret i16 %4
8723 }
8725 define zeroext i16 @test_masked_vpcmpsgtq_v4i1_v16i1_mask_mem_b(i8 zeroext %__u, <4 x i64> %__a, i64* %__b) local_unnamed_addr {
8726 ; VLX-LABEL: test_masked_vpcmpsgtq_v4i1_v16i1_mask_mem_b:
8727 ; VLX: # %bb.0: # %entry
8728 ; VLX-NEXT: kmovd %edi, %k1
8729 ; VLX-NEXT: vpcmpgtq (%rsi){1to4}, %ymm0, %k0 {%k1}
8730 ; VLX-NEXT: kmovd %k0, %eax
8731 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
8732 ; VLX-NEXT: vzeroupper
8733 ; VLX-NEXT: retq
8734 ;
8735 ; NoVLX-LABEL: test_masked_vpcmpsgtq_v4i1_v16i1_mask_mem_b:
8736 ; NoVLX: # %bb.0: # %entry
8737 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
8738 ; NoVLX-NEXT: kmovw %edi, %k1
8739 ; NoVLX-NEXT: vpcmpgtq (%rsi){1to8}, %zmm0, %k0 {%k1}
8740 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
8741 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
8742 ; NoVLX-NEXT: kmovw %k0, %eax
8743 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
8744 ; NoVLX-NEXT: vzeroupper
8745 ; NoVLX-NEXT: retq
8746 entry:
8747 %0 = bitcast <4 x i64> %__a to <4 x i64>
8748 %load = load i64, i64* %__b
8749 %vec = insertelement <4 x i64> undef, i64 %load, i32 0
8750 %1 = shufflevector <4 x i64> %vec, <4 x i64> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
8751 %2 = icmp sgt <4 x i64> %0, %1
8752 %3 = bitcast i8 %__u to <8 x i1>
8753 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
8754 %4 = and <4 x i1> %extract.i, %2
8755 %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
8756 %6 = bitcast <16 x i1> %5 to i16
8757 ret i16 %6
8758 }
8761 define zeroext i32 @test_vpcmpsgtq_v4i1_v32i1_mask(<4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
8762 ; VLX-LABEL: test_vpcmpsgtq_v4i1_v32i1_mask:
8763 ; VLX: # %bb.0: # %entry
8764 ; VLX-NEXT: vpcmpgtq %ymm1, %ymm0, %k0
8765 ; VLX-NEXT: kmovd %k0, %eax
8766 ; VLX-NEXT: vzeroupper
8767 ; VLX-NEXT: retq
8768 ;
8769 ; NoVLX-LABEL: test_vpcmpsgtq_v4i1_v32i1_mask:
8770 ; NoVLX: # %bb.0: # %entry
8771 ; NoVLX-NEXT: # kill: def $ymm1 killed $ymm1 def $zmm1
8772 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
8773 ; NoVLX-NEXT: vpcmpgtq %zmm1, %zmm0, %k0
8774 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
8775 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
8776 ; NoVLX-NEXT: kmovw %k0, %eax
8777 ; NoVLX-NEXT: vzeroupper
8778 ; NoVLX-NEXT: retq
8779 entry:
8780 %0 = bitcast <4 x i64> %__a to <4 x i64>
8781 %1 = bitcast <4 x i64> %__b to <4 x i64>
8782 %2 = icmp sgt <4 x i64> %0, %1
8783 %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
8784 %4 = bitcast <32 x i1> %3 to i32
8785 ret i32 %4
8786 }
8788 define zeroext i32 @test_vpcmpsgtq_v4i1_v32i1_mask_mem(<4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
8789 ; VLX-LABEL: test_vpcmpsgtq_v4i1_v32i1_mask_mem:
8790 ; VLX: # %bb.0: # %entry
8791 ; VLX-NEXT: vpcmpgtq (%rdi), %ymm0, %k0
8792 ; VLX-NEXT: kmovd %k0, %eax
8793 ; VLX-NEXT: vzeroupper
8794 ; VLX-NEXT: retq
8795 ;
8796 ; NoVLX-LABEL: test_vpcmpsgtq_v4i1_v32i1_mask_mem:
8797 ; NoVLX: # %bb.0: # %entry
8798 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
8799 ; NoVLX-NEXT: vmovdqa (%rdi), %ymm1
8800 ; NoVLX-NEXT: vpcmpgtq %zmm1, %zmm0, %k0
8801 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
8802 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
8803 ; NoVLX-NEXT: kmovw %k0, %eax
8804 ; NoVLX-NEXT: vzeroupper
8805 ; NoVLX-NEXT: retq
8806 entry:
8807 %0 = bitcast <4 x i64> %__a to <4 x i64>
8808 %load = load <4 x i64>, <4 x i64>* %__b
8809 %1 = bitcast <4 x i64> %load to <4 x i64>
8810 %2 = icmp sgt <4 x i64> %0, %1
8811 %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
8812 %4 = bitcast <32 x i1> %3 to i32
8813 ret i32 %4
8814 }
8816 define zeroext i32 @test_masked_vpcmpsgtq_v4i1_v32i1_mask(i8 zeroext %__u, <4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
8817 ; VLX-LABEL: test_masked_vpcmpsgtq_v4i1_v32i1_mask:
8818 ; VLX: # %bb.0: # %entry
8819 ; VLX-NEXT: kmovd %edi, %k1
8820 ; VLX-NEXT: vpcmpgtq %ymm1, %ymm0, %k0 {%k1}
8821 ; VLX-NEXT: kmovd %k0, %eax
8822 ; VLX-NEXT: vzeroupper
8823 ; VLX-NEXT: retq
8824 ;
8825 ; NoVLX-LABEL: test_masked_vpcmpsgtq_v4i1_v32i1_mask:
8826 ; NoVLX: # %bb.0: # %entry
8827 ; NoVLX-NEXT: # kill: def $ymm1 killed $ymm1 def $zmm1
8828 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
8829 ; NoVLX-NEXT: kmovw %edi, %k1
8830 ; NoVLX-NEXT: vpcmpgtq %zmm1, %zmm0, %k0 {%k1}
8831 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
8832 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
8833 ; NoVLX-NEXT: kmovw %k0, %eax
8834 ; NoVLX-NEXT: vzeroupper
8835 ; NoVLX-NEXT: retq
8836 entry:
8837 %0 = bitcast <4 x i64> %__a to <4 x i64>
8838 %1 = bitcast <4 x i64> %__b to <4 x i64>
8839 %2 = icmp sgt <4 x i64> %0, %1
8840 %3 = bitcast i8 %__u to <8 x i1>
8841 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
8842 %4 = and <4 x i1> %2, %extract.i
8843 %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
8844 %6 = bitcast <32 x i1> %5 to i32
8845 ret i32 %6
8846 }
8848 define zeroext i32 @test_masked_vpcmpsgtq_v4i1_v32i1_mask_mem(i8 zeroext %__u, <4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
8849 ; VLX-LABEL: test_masked_vpcmpsgtq_v4i1_v32i1_mask_mem:
8850 ; VLX: # %bb.0: # %entry
8851 ; VLX-NEXT: kmovd %edi, %k1
8852 ; VLX-NEXT: vpcmpgtq (%rsi), %ymm0, %k0 {%k1}
8853 ; VLX-NEXT: kmovd %k0, %eax
8854 ; VLX-NEXT: vzeroupper
8855 ; VLX-NEXT: retq
8856 ;
8857 ; NoVLX-LABEL: test_masked_vpcmpsgtq_v4i1_v32i1_mask_mem:
8858 ; NoVLX: # %bb.0: # %entry
8859 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
8860 ; NoVLX-NEXT: vmovdqa (%rsi), %ymm1
8861 ; NoVLX-NEXT: kmovw %edi, %k1
8862 ; NoVLX-NEXT: vpcmpgtq %zmm1, %zmm0, %k0 {%k1}
8863 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
8864 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
8865 ; NoVLX-NEXT: kmovw %k0, %eax
8866 ; NoVLX-NEXT: vzeroupper
8867 ; NoVLX-NEXT: retq
8868 entry:
8869 %0 = bitcast <4 x i64> %__a to <4 x i64>
8870 %load = load <4 x i64>, <4 x i64>* %__b
8871 %1 = bitcast <4 x i64> %load to <4 x i64>
8872 %2 = icmp sgt <4 x i64> %0, %1
8873 %3 = bitcast i8 %__u to <8 x i1>
8874 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
8875 %4 = and <4 x i1> %2, %extract.i
8876 %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
8877 %6 = bitcast <32 x i1> %5 to i32
8878 ret i32 %6
8879 }
8882 define zeroext i32 @test_vpcmpsgtq_v4i1_v32i1_mask_mem_b(<4 x i64> %__a, i64* %__b) local_unnamed_addr {
8883 ; VLX-LABEL: test_vpcmpsgtq_v4i1_v32i1_mask_mem_b:
8884 ; VLX: # %bb.0: # %entry
8885 ; VLX-NEXT: vpcmpgtq (%rdi){1to4}, %ymm0, %k0
8886 ; VLX-NEXT: kmovd %k0, %eax
8887 ; VLX-NEXT: vzeroupper
8888 ; VLX-NEXT: retq
8889 ;
8890 ; NoVLX-LABEL: test_vpcmpsgtq_v4i1_v32i1_mask_mem_b:
8891 ; NoVLX: # %bb.0: # %entry
8892 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
8893 ; NoVLX-NEXT: vpcmpgtq (%rdi){1to8}, %zmm0, %k0
8894 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
8895 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
8896 ; NoVLX-NEXT: kmovw %k0, %eax
8897 ; NoVLX-NEXT: vzeroupper
8898 ; NoVLX-NEXT: retq
8899 entry:
8900 %0 = bitcast <4 x i64> %__a to <4 x i64>
8901 %load = load i64, i64* %__b
8902 %vec = insertelement <4 x i64> undef, i64 %load, i32 0
8903 %1 = shufflevector <4 x i64> %vec, <4 x i64> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
8904 %2 = icmp sgt <4 x i64> %0, %1
8905 %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
8906 %4 = bitcast <32 x i1> %3 to i32
8907 ret i32 %4
8908 }
8910 define zeroext i32 @test_masked_vpcmpsgtq_v4i1_v32i1_mask_mem_b(i8 zeroext %__u, <4 x i64> %__a, i64* %__b) local_unnamed_addr {
8911 ; VLX-LABEL: test_masked_vpcmpsgtq_v4i1_v32i1_mask_mem_b:
8912 ; VLX: # %bb.0: # %entry
8913 ; VLX-NEXT: kmovd %edi, %k1
8914 ; VLX-NEXT: vpcmpgtq (%rsi){1to4}, %ymm0, %k0 {%k1}
8915 ; VLX-NEXT: kmovd %k0, %eax
8916 ; VLX-NEXT: vzeroupper
8917 ; VLX-NEXT: retq
8918 ;
8919 ; NoVLX-LABEL: test_masked_vpcmpsgtq_v4i1_v32i1_mask_mem_b:
8920 ; NoVLX: # %bb.0: # %entry
8921 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
8922 ; NoVLX-NEXT: kmovw %edi, %k1
8923 ; NoVLX-NEXT: vpcmpgtq (%rsi){1to8}, %zmm0, %k0 {%k1}
8924 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
8925 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
8926 ; NoVLX-NEXT: kmovw %k0, %eax
8927 ; NoVLX-NEXT: vzeroupper
8928 ; NoVLX-NEXT: retq
8929 entry:
8930 %0 = bitcast <4 x i64> %__a to <4 x i64>
8931 %load = load i64, i64* %__b
8932 %vec = insertelement <4 x i64> undef, i64 %load, i32 0
8933 %1 = shufflevector <4 x i64> %vec, <4 x i64> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
8934 %2 = icmp sgt <4 x i64> %0, %1
8935 %3 = bitcast i8 %__u to <8 x i1>
8936 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
8937 %4 = and <4 x i1> %extract.i, %2
8938 %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
8939 %6 = bitcast <32 x i1> %5 to i32
8940 ret i32 %6
8941 }
8944 define zeroext i64 @test_vpcmpsgtq_v4i1_v64i1_mask(<4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
8945 ; VLX-LABEL: test_vpcmpsgtq_v4i1_v64i1_mask:
8946 ; VLX: # %bb.0: # %entry
8947 ; VLX-NEXT: vpcmpgtq %ymm1, %ymm0, %k0
8948 ; VLX-NEXT: kmovq %k0, %rax
8949 ; VLX-NEXT: vzeroupper
8950 ; VLX-NEXT: retq
8951 ;
8952 ; NoVLX-LABEL: test_vpcmpsgtq_v4i1_v64i1_mask:
8953 ; NoVLX: # %bb.0: # %entry
8954 ; NoVLX-NEXT: # kill: def $ymm1 killed $ymm1 def $zmm1
8955 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
8956 ; NoVLX-NEXT: vpcmpgtq %zmm1, %zmm0, %k0
8957 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
8958 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
8959 ; NoVLX-NEXT: kmovw %k0, %eax
8960 ; NoVLX-NEXT: vzeroupper
8961 ; NoVLX-NEXT: retq
8962 entry:
8963 %0 = bitcast <4 x i64> %__a to <4 x i64>
8964 %1 = bitcast <4 x i64> %__b to <4 x i64>
8965 %2 = icmp sgt <4 x i64> %0, %1
8966 %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
8967 %4 = bitcast <64 x i1> %3 to i64
8968 ret i64 %4
8969 }
8971 define zeroext i64 @test_vpcmpsgtq_v4i1_v64i1_mask_mem(<4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
8972 ; VLX-LABEL: test_vpcmpsgtq_v4i1_v64i1_mask_mem:
8973 ; VLX: # %bb.0: # %entry
8974 ; VLX-NEXT: vpcmpgtq (%rdi), %ymm0, %k0
8975 ; VLX-NEXT: kmovq %k0, %rax
8976 ; VLX-NEXT: vzeroupper
8977 ; VLX-NEXT: retq
8978 ;
8979 ; NoVLX-LABEL: test_vpcmpsgtq_v4i1_v64i1_mask_mem:
8980 ; NoVLX: # %bb.0: # %entry
8981 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
8982 ; NoVLX-NEXT: vmovdqa (%rdi), %ymm1
8983 ; NoVLX-NEXT: vpcmpgtq %zmm1, %zmm0, %k0
8984 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
8985 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
8986 ; NoVLX-NEXT: kmovw %k0, %eax
8987 ; NoVLX-NEXT: vzeroupper
8988 ; NoVLX-NEXT: retq
8989 entry:
8990 %0 = bitcast <4 x i64> %__a to <4 x i64>
8991 %load = load <4 x i64>, <4 x i64>* %__b
8992 %1 = bitcast <4 x i64> %load to <4 x i64>
8993 %2 = icmp sgt <4 x i64> %0, %1
8994 %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
8995 %4 = bitcast <64 x i1> %3 to i64
8996 ret i64 %4
8997 }
8999 define zeroext i64 @test_masked_vpcmpsgtq_v4i1_v64i1_mask(i8 zeroext %__u, <4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
9000 ; VLX-LABEL: test_masked_vpcmpsgtq_v4i1_v64i1_mask:
9001 ; VLX: # %bb.0: # %entry
9002 ; VLX-NEXT: kmovd %edi, %k1
9003 ; VLX-NEXT: vpcmpgtq %ymm1, %ymm0, %k0 {%k1}
9004 ; VLX-NEXT: kmovq %k0, %rax
9005 ; VLX-NEXT: vzeroupper
9006 ; VLX-NEXT: retq
9007 ;
9008 ; NoVLX-LABEL: test_masked_vpcmpsgtq_v4i1_v64i1_mask:
9009 ; NoVLX: # %bb.0: # %entry
9010 ; NoVLX-NEXT: # kill: def $ymm1 killed $ymm1 def $zmm1
9011 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
9012 ; NoVLX-NEXT: kmovw %edi, %k1
9013 ; NoVLX-NEXT: vpcmpgtq %zmm1, %zmm0, %k0 {%k1}
9014 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
9015 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
9016 ; NoVLX-NEXT: kmovw %k0, %eax
9017 ; NoVLX-NEXT: vzeroupper
9018 ; NoVLX-NEXT: retq
9019 entry:
9020 %0 = bitcast <4 x i64> %__a to <4 x i64>
9021 %1 = bitcast <4 x i64> %__b to <4 x i64>
9022 %2 = icmp sgt <4 x i64> %0, %1
9023 %3 = bitcast i8 %__u to <8 x i1>
9024 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
9025 %4 = and <4 x i1> %2, %extract.i
9026 %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
9027 %6 = bitcast <64 x i1> %5 to i64
9028 ret i64 %6
9029 }
9031 define zeroext i64 @test_masked_vpcmpsgtq_v4i1_v64i1_mask_mem(i8 zeroext %__u, <4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
9032 ; VLX-LABEL: test_masked_vpcmpsgtq_v4i1_v64i1_mask_mem:
9033 ; VLX: # %bb.0: # %entry
9034 ; VLX-NEXT: kmovd %edi, %k1
9035 ; VLX-NEXT: vpcmpgtq (%rsi), %ymm0, %k0 {%k1}
9036 ; VLX-NEXT: kmovq %k0, %rax
9037 ; VLX-NEXT: vzeroupper
9038 ; VLX-NEXT: retq
9039 ;
9040 ; NoVLX-LABEL: test_masked_vpcmpsgtq_v4i1_v64i1_mask_mem:
9041 ; NoVLX: # %bb.0: # %entry
9042 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
9043 ; NoVLX-NEXT: vmovdqa (%rsi), %ymm1
9044 ; NoVLX-NEXT: kmovw %edi, %k1
9045 ; NoVLX-NEXT: vpcmpgtq %zmm1, %zmm0, %k0 {%k1}
9046 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
9047 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
9048 ; NoVLX-NEXT: kmovw %k0, %eax
9049 ; NoVLX-NEXT: vzeroupper
9050 ; NoVLX-NEXT: retq
9051 entry:
9052 %0 = bitcast <4 x i64> %__a to <4 x i64>
9053 %load = load <4 x i64>, <4 x i64>* %__b
9054 %1 = bitcast <4 x i64> %load to <4 x i64>
9055 %2 = icmp sgt <4 x i64> %0, %1
9056 %3 = bitcast i8 %__u to <8 x i1>
9057 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
9058 %4 = and <4 x i1> %2, %extract.i
9059 %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
9060 %6 = bitcast <64 x i1> %5 to i64
9061 ret i64 %6
9062 }
9065 define zeroext i64 @test_vpcmpsgtq_v4i1_v64i1_mask_mem_b(<4 x i64> %__a, i64* %__b) local_unnamed_addr {
9066 ; VLX-LABEL: test_vpcmpsgtq_v4i1_v64i1_mask_mem_b:
9067 ; VLX: # %bb.0: # %entry
9068 ; VLX-NEXT: vpcmpgtq (%rdi){1to4}, %ymm0, %k0
9069 ; VLX-NEXT: kmovq %k0, %rax
9070 ; VLX-NEXT: vzeroupper
9071 ; VLX-NEXT: retq
9072 ;
9073 ; NoVLX-LABEL: test_vpcmpsgtq_v4i1_v64i1_mask_mem_b:
9074 ; NoVLX: # %bb.0: # %entry
9075 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
9076 ; NoVLX-NEXT: vpcmpgtq (%rdi){1to8}, %zmm0, %k0
9077 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
9078 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
9079 ; NoVLX-NEXT: kmovw %k0, %eax
9080 ; NoVLX-NEXT: vzeroupper
9081 ; NoVLX-NEXT: retq
9082 entry:
9083 %0 = bitcast <4 x i64> %__a to <4 x i64>
9084 %load = load i64, i64* %__b
9085 %vec = insertelement <4 x i64> undef, i64 %load, i32 0
9086 %1 = shufflevector <4 x i64> %vec, <4 x i64> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
9087 %2 = icmp sgt <4 x i64> %0, %1
9088 %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
9089 %4 = bitcast <64 x i1> %3 to i64
9090 ret i64 %4
9091 }
9093 define zeroext i64 @test_masked_vpcmpsgtq_v4i1_v64i1_mask_mem_b(i8 zeroext %__u, <4 x i64> %__a, i64* %__b) local_unnamed_addr {
9094 ; VLX-LABEL: test_masked_vpcmpsgtq_v4i1_v64i1_mask_mem_b:
9095 ; VLX: # %bb.0: # %entry
9096 ; VLX-NEXT: kmovd %edi, %k1
9097 ; VLX-NEXT: vpcmpgtq (%rsi){1to4}, %ymm0, %k0 {%k1}
9098 ; VLX-NEXT: kmovq %k0, %rax
9099 ; VLX-NEXT: vzeroupper
9100 ; VLX-NEXT: retq
9101 ;
9102 ; NoVLX-LABEL: test_masked_vpcmpsgtq_v4i1_v64i1_mask_mem_b:
9103 ; NoVLX: # %bb.0: # %entry
9104 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
9105 ; NoVLX-NEXT: kmovw %edi, %k1
9106 ; NoVLX-NEXT: vpcmpgtq (%rsi){1to8}, %zmm0, %k0 {%k1}
9107 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
9108 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
9109 ; NoVLX-NEXT: kmovw %k0, %eax
9110 ; NoVLX-NEXT: vzeroupper
9111 ; NoVLX-NEXT: retq
9112 entry:
9113 %0 = bitcast <4 x i64> %__a to <4 x i64>
9114 %load = load i64, i64* %__b
9115 %vec = insertelement <4 x i64> undef, i64 %load, i32 0
9116 %1 = shufflevector <4 x i64> %vec, <4 x i64> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
9117 %2 = icmp sgt <4 x i64> %0, %1
9118 %3 = bitcast i8 %__u to <8 x i1>
9119 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
9120 %4 = and <4 x i1> %extract.i, %2
9121 %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
9122 %6 = bitcast <64 x i1> %5 to i64
9123 ret i64 %6
9124 }
9127 define zeroext i16 @test_vpcmpsgtq_v8i1_v16i1_mask(<8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
9128 ; VLX-LABEL: test_vpcmpsgtq_v8i1_v16i1_mask:
9129 ; VLX: # %bb.0: # %entry
9130 ; VLX-NEXT: vpcmpgtq %zmm1, %zmm0, %k0
9131 ; VLX-NEXT: kmovd %k0, %eax
9132 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
9133 ; VLX-NEXT: vzeroupper
9134 ; VLX-NEXT: retq
9135 ;
9136 ; NoVLX-LABEL: test_vpcmpsgtq_v8i1_v16i1_mask:
9137 ; NoVLX: # %bb.0: # %entry
9138 ; NoVLX-NEXT: vpcmpgtq %zmm1, %zmm0, %k0
9139 ; NoVLX-NEXT: kmovw %k0, %eax
9140 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
9141 ; NoVLX-NEXT: vzeroupper
9142 ; NoVLX-NEXT: retq
9143 entry:
9144 %0 = bitcast <8 x i64> %__a to <8 x i64>
9145 %1 = bitcast <8 x i64> %__b to <8 x i64>
9146 %2 = icmp sgt <8 x i64> %0, %1
9147 %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
9148 %4 = bitcast <16 x i1> %3 to i16
9149 ret i16 %4
9150 }
9152 define zeroext i16 @test_vpcmpsgtq_v8i1_v16i1_mask_mem(<8 x i64> %__a, <8 x i64>* %__b) local_unnamed_addr {
9153 ; VLX-LABEL: test_vpcmpsgtq_v8i1_v16i1_mask_mem:
9154 ; VLX: # %bb.0: # %entry
9155 ; VLX-NEXT: vpcmpgtq (%rdi), %zmm0, %k0
9156 ; VLX-NEXT: kmovd %k0, %eax
9157 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
9158 ; VLX-NEXT: vzeroupper
9159 ; VLX-NEXT: retq
9160 ;
9161 ; NoVLX-LABEL: test_vpcmpsgtq_v8i1_v16i1_mask_mem:
9162 ; NoVLX: # %bb.0: # %entry
9163 ; NoVLX-NEXT: vpcmpgtq (%rdi), %zmm0, %k0
9164 ; NoVLX-NEXT: kmovw %k0, %eax
9165 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
9166 ; NoVLX-NEXT: vzeroupper
9167 ; NoVLX-NEXT: retq
9168 entry:
9169 %0 = bitcast <8 x i64> %__a to <8 x i64>
9170 %load = load <8 x i64>, <8 x i64>* %__b
9171 %1 = bitcast <8 x i64> %load to <8 x i64>
9172 %2 = icmp sgt <8 x i64> %0, %1
9173 %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
9174 %4 = bitcast <16 x i1> %3 to i16
9175 ret i16 %4
9176 }
9178 define zeroext i16 @test_masked_vpcmpsgtq_v8i1_v16i1_mask(i8 zeroext %__u, <8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
9179 ; VLX-LABEL: test_masked_vpcmpsgtq_v8i1_v16i1_mask:
9180 ; VLX: # %bb.0: # %entry
9181 ; VLX-NEXT: kmovd %edi, %k1
9182 ; VLX-NEXT: vpcmpgtq %zmm1, %zmm0, %k0 {%k1}
9183 ; VLX-NEXT: kmovd %k0, %eax
9184 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
9185 ; VLX-NEXT: vzeroupper
9186 ; VLX-NEXT: retq
9187 ;
9188 ; NoVLX-LABEL: test_masked_vpcmpsgtq_v8i1_v16i1_mask:
9189 ; NoVLX: # %bb.0: # %entry
9190 ; NoVLX-NEXT: kmovw %edi, %k1
9191 ; NoVLX-NEXT: vpcmpgtq %zmm1, %zmm0, %k0 {%k1}
9192 ; NoVLX-NEXT: kmovw %k0, %eax
9193 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
9194 ; NoVLX-NEXT: vzeroupper
9195 ; NoVLX-NEXT: retq
9196 entry:
9197 %0 = bitcast <8 x i64> %__a to <8 x i64>
9198 %1 = bitcast <8 x i64> %__b to <8 x i64>
9199 %2 = icmp sgt <8 x i64> %0, %1
9200 %3 = bitcast i8 %__u to <8 x i1>
9201 %4 = and <8 x i1> %2, %3
9202 %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
9203 %6 = bitcast <16 x i1> %5 to i16
9204 ret i16 %6
9205 }
9207 define zeroext i16 @test_masked_vpcmpsgtq_v8i1_v16i1_mask_mem(i8 zeroext %__u, <8 x i64> %__a, <8 x i64>* %__b) local_unnamed_addr {
9208 ; VLX-LABEL: test_masked_vpcmpsgtq_v8i1_v16i1_mask_mem:
9209 ; VLX: # %bb.0: # %entry
9210 ; VLX-NEXT: kmovd %edi, %k1
9211 ; VLX-NEXT: vpcmpgtq (%rsi), %zmm0, %k0 {%k1}
9212 ; VLX-NEXT: kmovd %k0, %eax
9213 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
9214 ; VLX-NEXT: vzeroupper
9215 ; VLX-NEXT: retq
9216 ;
9217 ; NoVLX-LABEL: test_masked_vpcmpsgtq_v8i1_v16i1_mask_mem:
9218 ; NoVLX: # %bb.0: # %entry
9219 ; NoVLX-NEXT: kmovw %edi, %k1
9220 ; NoVLX-NEXT: vpcmpgtq (%rsi), %zmm0, %k0 {%k1}
9221 ; NoVLX-NEXT: kmovw %k0, %eax
9222 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
9223 ; NoVLX-NEXT: vzeroupper
9224 ; NoVLX-NEXT: retq
9225 entry:
9226 %0 = bitcast <8 x i64> %__a to <8 x i64>
9227 %load = load <8 x i64>, <8 x i64>* %__b
9228 %1 = bitcast <8 x i64> %load to <8 x i64>
9229 %2 = icmp sgt <8 x i64> %0, %1
9230 %3 = bitcast i8 %__u to <8 x i1>
9231 %4 = and <8 x i1> %2, %3
9232 %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
9233 %6 = bitcast <16 x i1> %5 to i16
9234 ret i16 %6
9235 }
9238 define zeroext i16 @test_vpcmpsgtq_v8i1_v16i1_mask_mem_b(<8 x i64> %__a, i64* %__b) local_unnamed_addr {
9239 ; VLX-LABEL: test_vpcmpsgtq_v8i1_v16i1_mask_mem_b:
9240 ; VLX: # %bb.0: # %entry
9241 ; VLX-NEXT: vpcmpgtq (%rdi){1to8}, %zmm0, %k0
9242 ; VLX-NEXT: kmovd %k0, %eax
9243 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
9244 ; VLX-NEXT: vzeroupper
9245 ; VLX-NEXT: retq
9246 ;
9247 ; NoVLX-LABEL: test_vpcmpsgtq_v8i1_v16i1_mask_mem_b:
9248 ; NoVLX: # %bb.0: # %entry
9249 ; NoVLX-NEXT: vpcmpgtq (%rdi){1to8}, %zmm0, %k0
9250 ; NoVLX-NEXT: kmovw %k0, %eax
9251 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
9252 ; NoVLX-NEXT: vzeroupper
9253 ; NoVLX-NEXT: retq
9254 entry:
9255 %0 = bitcast <8 x i64> %__a to <8 x i64>
9256 %load = load i64, i64* %__b
9257 %vec = insertelement <8 x i64> undef, i64 %load, i32 0
9258 %1 = shufflevector <8 x i64> %vec, <8 x i64> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
9259 %2 = icmp sgt <8 x i64> %0, %1
9260 %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
9261 %4 = bitcast <16 x i1> %3 to i16
9262 ret i16 %4
9263 }
9265 define zeroext i16 @test_masked_vpcmpsgtq_v8i1_v16i1_mask_mem_b(i8 zeroext %__u, <8 x i64> %__a, i64* %__b) local_unnamed_addr {
9266 ; VLX-LABEL: test_masked_vpcmpsgtq_v8i1_v16i1_mask_mem_b:
9267 ; VLX: # %bb.0: # %entry
9268 ; VLX-NEXT: kmovd %edi, %k1
9269 ; VLX-NEXT: vpcmpgtq (%rsi){1to8}, %zmm0, %k0 {%k1}
9270 ; VLX-NEXT: kmovd %k0, %eax
9271 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
9272 ; VLX-NEXT: vzeroupper
9273 ; VLX-NEXT: retq
9274 ;
9275 ; NoVLX-LABEL: test_masked_vpcmpsgtq_v8i1_v16i1_mask_mem_b:
9276 ; NoVLX: # %bb.0: # %entry
9277 ; NoVLX-NEXT: kmovw %edi, %k1
9278 ; NoVLX-NEXT: vpcmpgtq (%rsi){1to8}, %zmm0, %k0 {%k1}
9279 ; NoVLX-NEXT: kmovw %k0, %eax
9280 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
9281 ; NoVLX-NEXT: vzeroupper
9282 ; NoVLX-NEXT: retq
9283 entry:
9284 %0 = bitcast <8 x i64> %__a to <8 x i64>
9285 %load = load i64, i64* %__b
9286 %vec = insertelement <8 x i64> undef, i64 %load, i32 0
9287 %1 = shufflevector <8 x i64> %vec, <8 x i64> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
9288 %2 = icmp sgt <8 x i64> %0, %1
9289 %3 = bitcast i8 %__u to <8 x i1>
9290 %4 = and <8 x i1> %3, %2
9291 %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
9292 %6 = bitcast <16 x i1> %5 to i16
9293 ret i16 %6
9294 }
9297 define zeroext i32 @test_vpcmpsgtq_v8i1_v32i1_mask(<8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
9298 ; VLX-LABEL: test_vpcmpsgtq_v8i1_v32i1_mask:
9299 ; VLX: # %bb.0: # %entry
9300 ; VLX-NEXT: vpcmpgtq %zmm1, %zmm0, %k0
9301 ; VLX-NEXT: kmovd %k0, %eax
9302 ; VLX-NEXT: vzeroupper
9303 ; VLX-NEXT: retq
9304 ;
9305 ; NoVLX-LABEL: test_vpcmpsgtq_v8i1_v32i1_mask:
9306 ; NoVLX: # %bb.0: # %entry
9307 ; NoVLX-NEXT: vpcmpgtq %zmm1, %zmm0, %k0
9308 ; NoVLX-NEXT: kmovw %k0, %eax
9309 ; NoVLX-NEXT: vzeroupper
9310 ; NoVLX-NEXT: retq
9311 entry:
9312 %0 = bitcast <8 x i64> %__a to <8 x i64>
9313 %1 = bitcast <8 x i64> %__b to <8 x i64>
9314 %2 = icmp sgt <8 x i64> %0, %1
9315 %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
9316 %4 = bitcast <32 x i1> %3 to i32
9317 ret i32 %4
9318 }
9320 define zeroext i32 @test_vpcmpsgtq_v8i1_v32i1_mask_mem(<8 x i64> %__a, <8 x i64>* %__b) local_unnamed_addr {
9321 ; VLX-LABEL: test_vpcmpsgtq_v8i1_v32i1_mask_mem:
9322 ; VLX: # %bb.0: # %entry
9323 ; VLX-NEXT: vpcmpgtq (%rdi), %zmm0, %k0
9324 ; VLX-NEXT: kmovd %k0, %eax
9325 ; VLX-NEXT: vzeroupper
9326 ; VLX-NEXT: retq
9327 ;
9328 ; NoVLX-LABEL: test_vpcmpsgtq_v8i1_v32i1_mask_mem:
9329 ; NoVLX: # %bb.0: # %entry
9330 ; NoVLX-NEXT: vpcmpgtq (%rdi), %zmm0, %k0
9331 ; NoVLX-NEXT: kmovw %k0, %eax
9332 ; NoVLX-NEXT: vzeroupper
9333 ; NoVLX-NEXT: retq
9334 entry:
9335 %0 = bitcast <8 x i64> %__a to <8 x i64>
9336 %load = load <8 x i64>, <8 x i64>* %__b
9337 %1 = bitcast <8 x i64> %load to <8 x i64>
9338 %2 = icmp sgt <8 x i64> %0, %1
9339 %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
9340 %4 = bitcast <32 x i1> %3 to i32
9341 ret i32 %4
9342 }
9344 define zeroext i32 @test_masked_vpcmpsgtq_v8i1_v32i1_mask(i8 zeroext %__u, <8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
9345 ; VLX-LABEL: test_masked_vpcmpsgtq_v8i1_v32i1_mask:
9346 ; VLX: # %bb.0: # %entry
9347 ; VLX-NEXT: kmovd %edi, %k1
9348 ; VLX-NEXT: vpcmpgtq %zmm1, %zmm0, %k0 {%k1}
9349 ; VLX-NEXT: kmovd %k0, %eax
9350 ; VLX-NEXT: vzeroupper
9351 ; VLX-NEXT: retq
9352 ;
9353 ; NoVLX-LABEL: test_masked_vpcmpsgtq_v8i1_v32i1_mask:
9354 ; NoVLX: # %bb.0: # %entry
9355 ; NoVLX-NEXT: kmovw %edi, %k1
9356 ; NoVLX-NEXT: vpcmpgtq %zmm1, %zmm0, %k0 {%k1}
9357 ; NoVLX-NEXT: kmovw %k0, %eax
9358 ; NoVLX-NEXT: vzeroupper
9359 ; NoVLX-NEXT: retq
9360 entry:
9361 %0 = bitcast <8 x i64> %__a to <8 x i64>
9362 %1 = bitcast <8 x i64> %__b to <8 x i64>
9363 %2 = icmp sgt <8 x i64> %0, %1
9364 %3 = bitcast i8 %__u to <8 x i1>
9365 %4 = and <8 x i1> %2, %3
9366 %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
9367 %6 = bitcast <32 x i1> %5 to i32
9368 ret i32 %6
9369 }
9371 define zeroext i32 @test_masked_vpcmpsgtq_v8i1_v32i1_mask_mem(i8 zeroext %__u, <8 x i64> %__a, <8 x i64>* %__b) local_unnamed_addr {
9372 ; VLX-LABEL: test_masked_vpcmpsgtq_v8i1_v32i1_mask_mem:
9373 ; VLX: # %bb.0: # %entry
9374 ; VLX-NEXT: kmovd %edi, %k1
9375 ; VLX-NEXT: vpcmpgtq (%rsi), %zmm0, %k0 {%k1}
9376 ; VLX-NEXT: kmovd %k0, %eax
9377 ; VLX-NEXT: vzeroupper
9378 ; VLX-NEXT: retq
9379 ;
9380 ; NoVLX-LABEL: test_masked_vpcmpsgtq_v8i1_v32i1_mask_mem:
9381 ; NoVLX: # %bb.0: # %entry
9382 ; NoVLX-NEXT: kmovw %edi, %k1
9383 ; NoVLX-NEXT: vpcmpgtq (%rsi), %zmm0, %k0 {%k1}
9384 ; NoVLX-NEXT: kmovw %k0, %eax
9385 ; NoVLX-NEXT: vzeroupper
9386 ; NoVLX-NEXT: retq
9387 entry:
9388 %0 = bitcast <8 x i64> %__a to <8 x i64>
9389 %load = load <8 x i64>, <8 x i64>* %__b
9390 %1 = bitcast <8 x i64> %load to <8 x i64>
9391 %2 = icmp sgt <8 x i64> %0, %1
9392 %3 = bitcast i8 %__u to <8 x i1>
9393 %4 = and <8 x i1> %2, %3
9394 %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
9395 %6 = bitcast <32 x i1> %5 to i32
9396 ret i32 %6
9397 }
9400 define zeroext i32 @test_vpcmpsgtq_v8i1_v32i1_mask_mem_b(<8 x i64> %__a, i64* %__b) local_unnamed_addr {
9401 ; VLX-LABEL: test_vpcmpsgtq_v8i1_v32i1_mask_mem_b:
9402 ; VLX: # %bb.0: # %entry
9403 ; VLX-NEXT: vpcmpgtq (%rdi){1to8}, %zmm0, %k0
9404 ; VLX-NEXT: kmovd %k0, %eax
9405 ; VLX-NEXT: vzeroupper
9406 ; VLX-NEXT: retq
9407 ;
9408 ; NoVLX-LABEL: test_vpcmpsgtq_v8i1_v32i1_mask_mem_b:
9409 ; NoVLX: # %bb.0: # %entry
9410 ; NoVLX-NEXT: vpcmpgtq (%rdi){1to8}, %zmm0, %k0
9411 ; NoVLX-NEXT: kmovw %k0, %eax
9412 ; NoVLX-NEXT: vzeroupper
9413 ; NoVLX-NEXT: retq
9414 entry:
9415 %0 = bitcast <8 x i64> %__a to <8 x i64>
9416 %load = load i64, i64* %__b
9417 %vec = insertelement <8 x i64> undef, i64 %load, i32 0
9418 %1 = shufflevector <8 x i64> %vec, <8 x i64> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
9419 %2 = icmp sgt <8 x i64> %0, %1
9420 %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
9421 %4 = bitcast <32 x i1> %3 to i32
9422 ret i32 %4
9423 }
9425 define zeroext i32 @test_masked_vpcmpsgtq_v8i1_v32i1_mask_mem_b(i8 zeroext %__u, <8 x i64> %__a, i64* %__b) local_unnamed_addr {
9426 ; VLX-LABEL: test_masked_vpcmpsgtq_v8i1_v32i1_mask_mem_b:
9427 ; VLX: # %bb.0: # %entry
9428 ; VLX-NEXT: kmovd %edi, %k1
9429 ; VLX-NEXT: vpcmpgtq (%rsi){1to8}, %zmm0, %k0 {%k1}
9430 ; VLX-NEXT: kmovd %k0, %eax
9431 ; VLX-NEXT: vzeroupper
9432 ; VLX-NEXT: retq
9433 ;
9434 ; NoVLX-LABEL: test_masked_vpcmpsgtq_v8i1_v32i1_mask_mem_b:
9435 ; NoVLX: # %bb.0: # %entry
9436 ; NoVLX-NEXT: kmovw %edi, %k1
9437 ; NoVLX-NEXT: vpcmpgtq (%rsi){1to8}, %zmm0, %k0 {%k1}
9438 ; NoVLX-NEXT: kmovw %k0, %eax
9439 ; NoVLX-NEXT: vzeroupper
9440 ; NoVLX-NEXT: retq
9441 entry:
9442 %0 = bitcast <8 x i64> %__a to <8 x i64>
9443 %load = load i64, i64* %__b
9444 %vec = insertelement <8 x i64> undef, i64 %load, i32 0
9445 %1 = shufflevector <8 x i64> %vec, <8 x i64> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
9446 %2 = icmp sgt <8 x i64> %0, %1
9447 %3 = bitcast i8 %__u to <8 x i1>
9448 %4 = and <8 x i1> %3, %2
9449 %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
9450 %6 = bitcast <32 x i1> %5 to i32
9451 ret i32 %6
9452 }
9455 define zeroext i64 @test_vpcmpsgtq_v8i1_v64i1_mask(<8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
9456 ; VLX-LABEL: test_vpcmpsgtq_v8i1_v64i1_mask:
9457 ; VLX: # %bb.0: # %entry
9458 ; VLX-NEXT: vpcmpgtq %zmm1, %zmm0, %k0
9459 ; VLX-NEXT: kmovq %k0, %rax
9460 ; VLX-NEXT: vzeroupper
9461 ; VLX-NEXT: retq
9462 ;
9463 ; NoVLX-LABEL: test_vpcmpsgtq_v8i1_v64i1_mask:
9464 ; NoVLX: # %bb.0: # %entry
9465 ; NoVLX-NEXT: vpcmpgtq %zmm1, %zmm0, %k0
9466 ; NoVLX-NEXT: kmovw %k0, %eax
9467 ; NoVLX-NEXT: vzeroupper
9468 ; NoVLX-NEXT: retq
9469 entry:
9470 %0 = bitcast <8 x i64> %__a to <8 x i64>
9471 %1 = bitcast <8 x i64> %__b to <8 x i64>
9472 %2 = icmp sgt <8 x i64> %0, %1
9473 %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
9474 %4 = bitcast <64 x i1> %3 to i64
9475 ret i64 %4
9476 }
9478 define zeroext i64 @test_vpcmpsgtq_v8i1_v64i1_mask_mem(<8 x i64> %__a, <8 x i64>* %__b) local_unnamed_addr {
9479 ; VLX-LABEL: test_vpcmpsgtq_v8i1_v64i1_mask_mem:
9480 ; VLX: # %bb.0: # %entry
9481 ; VLX-NEXT: vpcmpgtq (%rdi), %zmm0, %k0
9482 ; VLX-NEXT: kmovq %k0, %rax
9483 ; VLX-NEXT: vzeroupper
9484 ; VLX-NEXT: retq
9485 ;
9486 ; NoVLX-LABEL: test_vpcmpsgtq_v8i1_v64i1_mask_mem:
9487 ; NoVLX: # %bb.0: # %entry
9488 ; NoVLX-NEXT: vpcmpgtq (%rdi), %zmm0, %k0
9489 ; NoVLX-NEXT: kmovw %k0, %eax
9490 ; NoVLX-NEXT: vzeroupper
9491 ; NoVLX-NEXT: retq
9492 entry:
9493 %0 = bitcast <8 x i64> %__a to <8 x i64>
9494 %load = load <8 x i64>, <8 x i64>* %__b
9495 %1 = bitcast <8 x i64> %load to <8 x i64>
9496 %2 = icmp sgt <8 x i64> %0, %1
9497 %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
9498 %4 = bitcast <64 x i1> %3 to i64
9499 ret i64 %4
9500 }
9502 define zeroext i64 @test_masked_vpcmpsgtq_v8i1_v64i1_mask(i8 zeroext %__u, <8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
9503 ; VLX-LABEL: test_masked_vpcmpsgtq_v8i1_v64i1_mask:
9504 ; VLX: # %bb.0: # %entry
9505 ; VLX-NEXT: kmovd %edi, %k1
9506 ; VLX-NEXT: vpcmpgtq %zmm1, %zmm0, %k0 {%k1}
9507 ; VLX-NEXT: kmovq %k0, %rax
9508 ; VLX-NEXT: vzeroupper
9509 ; VLX-NEXT: retq
9510 ;
9511 ; NoVLX-LABEL: test_masked_vpcmpsgtq_v8i1_v64i1_mask:
9512 ; NoVLX: # %bb.0: # %entry
9513 ; NoVLX-NEXT: kmovw %edi, %k1
9514 ; NoVLX-NEXT: vpcmpgtq %zmm1, %zmm0, %k0 {%k1}
9515 ; NoVLX-NEXT: kmovw %k0, %eax
9516 ; NoVLX-NEXT: vzeroupper
9517 ; NoVLX-NEXT: retq
9518 entry:
9519 %0 = bitcast <8 x i64> %__a to <8 x i64>
9520 %1 = bitcast <8 x i64> %__b to <8 x i64>
9521 %2 = icmp sgt <8 x i64> %0, %1
9522 %3 = bitcast i8 %__u to <8 x i1>
9523 %4 = and <8 x i1> %2, %3
9524 %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
9525 %6 = bitcast <64 x i1> %5 to i64
9526 ret i64 %6
9527 }
9529 define zeroext i64 @test_masked_vpcmpsgtq_v8i1_v64i1_mask_mem(i8 zeroext %__u, <8 x i64> %__a, <8 x i64>* %__b) local_unnamed_addr {
9530 ; VLX-LABEL: test_masked_vpcmpsgtq_v8i1_v64i1_mask_mem:
9531 ; VLX: # %bb.0: # %entry
9532 ; VLX-NEXT: kmovd %edi, %k1
9533 ; VLX-NEXT: vpcmpgtq (%rsi), %zmm0, %k0 {%k1}
9534 ; VLX-NEXT: kmovq %k0, %rax
9535 ; VLX-NEXT: vzeroupper
9536 ; VLX-NEXT: retq
9537 ;
9538 ; NoVLX-LABEL: test_masked_vpcmpsgtq_v8i1_v64i1_mask_mem:
9539 ; NoVLX: # %bb.0: # %entry
9540 ; NoVLX-NEXT: kmovw %edi, %k1
9541 ; NoVLX-NEXT: vpcmpgtq (%rsi), %zmm0, %k0 {%k1}
9542 ; NoVLX-NEXT: kmovw %k0, %eax
9543 ; NoVLX-NEXT: vzeroupper
9544 ; NoVLX-NEXT: retq
9545 entry:
9546 %0 = bitcast <8 x i64> %__a to <8 x i64>
9547 %load = load <8 x i64>, <8 x i64>* %__b
9548 %1 = bitcast <8 x i64> %load to <8 x i64>
9549 %2 = icmp sgt <8 x i64> %0, %1
9550 %3 = bitcast i8 %__u to <8 x i1>
9551 %4 = and <8 x i1> %2, %3
9552 %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
9553 %6 = bitcast <64 x i1> %5 to i64
9554 ret i64 %6
9555 }
9558 define zeroext i64 @test_vpcmpsgtq_v8i1_v64i1_mask_mem_b(<8 x i64> %__a, i64* %__b) local_unnamed_addr {
9559 ; VLX-LABEL: test_vpcmpsgtq_v8i1_v64i1_mask_mem_b:
9560 ; VLX: # %bb.0: # %entry
9561 ; VLX-NEXT: vpcmpgtq (%rdi){1to8}, %zmm0, %k0
9562 ; VLX-NEXT: kmovq %k0, %rax
9563 ; VLX-NEXT: vzeroupper
9564 ; VLX-NEXT: retq
9565 ;
9566 ; NoVLX-LABEL: test_vpcmpsgtq_v8i1_v64i1_mask_mem_b:
9567 ; NoVLX: # %bb.0: # %entry
9568 ; NoVLX-NEXT: vpcmpgtq (%rdi){1to8}, %zmm0, %k0
9569 ; NoVLX-NEXT: kmovw %k0, %eax
9570 ; NoVLX-NEXT: vzeroupper
9571 ; NoVLX-NEXT: retq
9572 entry:
9573 %0 = bitcast <8 x i64> %__a to <8 x i64>
9574 %load = load i64, i64* %__b
9575 %vec = insertelement <8 x i64> undef, i64 %load, i32 0
9576 %1 = shufflevector <8 x i64> %vec, <8 x i64> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
9577 %2 = icmp sgt <8 x i64> %0, %1
9578 %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
9579 %4 = bitcast <64 x i1> %3 to i64
9580 ret i64 %4
9581 }
9583 define zeroext i64 @test_masked_vpcmpsgtq_v8i1_v64i1_mask_mem_b(i8 zeroext %__u, <8 x i64> %__a, i64* %__b) local_unnamed_addr {
9584 ; VLX-LABEL: test_masked_vpcmpsgtq_v8i1_v64i1_mask_mem_b:
9585 ; VLX: # %bb.0: # %entry
9586 ; VLX-NEXT: kmovd %edi, %k1
9587 ; VLX-NEXT: vpcmpgtq (%rsi){1to8}, %zmm0, %k0 {%k1}
9588 ; VLX-NEXT: kmovq %k0, %rax
9589 ; VLX-NEXT: vzeroupper
9590 ; VLX-NEXT: retq
9591 ;
9592 ; NoVLX-LABEL: test_masked_vpcmpsgtq_v8i1_v64i1_mask_mem_b:
9593 ; NoVLX: # %bb.0: # %entry
9594 ; NoVLX-NEXT: kmovw %edi, %k1
9595 ; NoVLX-NEXT: vpcmpgtq (%rsi){1to8}, %zmm0, %k0 {%k1}
9596 ; NoVLX-NEXT: kmovw %k0, %eax
9597 ; NoVLX-NEXT: vzeroupper
9598 ; NoVLX-NEXT: retq
9599 entry:
9600 %0 = bitcast <8 x i64> %__a to <8 x i64>
9601 %load = load i64, i64* %__b
9602 %vec = insertelement <8 x i64> undef, i64 %load, i32 0
9603 %1 = shufflevector <8 x i64> %vec, <8 x i64> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
9604 %2 = icmp sgt <8 x i64> %0, %1
9605 %3 = bitcast i8 %__u to <8 x i1>
9606 %4 = and <8 x i1> %3, %2
9607 %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
9608 %6 = bitcast <64 x i1> %5 to i64
9609 ret i64 %6
9610 }
9613 define zeroext i32 @test_vpcmpsgeb_v16i1_v32i1_mask(<2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
9614 ; VLX-LABEL: test_vpcmpsgeb_v16i1_v32i1_mask:
9615 ; VLX: # %bb.0: # %entry
9616 ; VLX-NEXT: vpcmpnltb %xmm1, %xmm0, %k0
9617 ; VLX-NEXT: kmovd %k0, %eax
9618 ; VLX-NEXT: retq
9619 ;
9620 ; NoVLX-LABEL: test_vpcmpsgeb_v16i1_v32i1_mask:
9621 ; NoVLX: # %bb.0: # %entry
9622 ; NoVLX-NEXT: vpcmpgtb %xmm0, %xmm1, %xmm0
9623 ; NoVLX-NEXT: vpternlogq $15, %zmm0, %zmm0, %zmm0
9624 ; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0
9625 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
9626 ; NoVLX-NEXT: kmovw %k0, %eax
9627 ; NoVLX-NEXT: vzeroupper
9628 ; NoVLX-NEXT: retq
9629 entry:
9630 %0 = bitcast <2 x i64> %__a to <16 x i8>
9631 %1 = bitcast <2 x i64> %__b to <16 x i8>
9632 %2 = icmp sge <16 x i8> %0, %1
9633 %3 = shufflevector <16 x i1> %2, <16 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>
9634 %4 = bitcast <32 x i1> %3 to i32
9635 ret i32 %4
9636 }
9638 define zeroext i32 @test_vpcmpsgeb_v16i1_v32i1_mask_mem(<2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
9639 ; VLX-LABEL: test_vpcmpsgeb_v16i1_v32i1_mask_mem:
9640 ; VLX: # %bb.0: # %entry
9641 ; VLX-NEXT: vpcmpnltb (%rdi), %xmm0, %k0
9642 ; VLX-NEXT: kmovd %k0, %eax
9643 ; VLX-NEXT: retq
9644 ;
9645 ; NoVLX-LABEL: test_vpcmpsgeb_v16i1_v32i1_mask_mem:
9646 ; NoVLX: # %bb.0: # %entry
9647 ; NoVLX-NEXT: vmovdqa (%rdi), %xmm1
9648 ; NoVLX-NEXT: vpcmpgtb %xmm0, %xmm1, %xmm0
9649 ; NoVLX-NEXT: vpternlogq $15, %zmm0, %zmm0, %zmm0
9650 ; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0
9651 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
9652 ; NoVLX-NEXT: kmovw %k0, %eax
9653 ; NoVLX-NEXT: vzeroupper
9654 ; NoVLX-NEXT: retq
9655 entry:
9656 %0 = bitcast <2 x i64> %__a to <16 x i8>
9657 %load = load <2 x i64>, <2 x i64>* %__b
9658 %1 = bitcast <2 x i64> %load to <16 x i8>
9659 %2 = icmp sge <16 x i8> %0, %1
9660 %3 = shufflevector <16 x i1> %2, <16 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>
9661 %4 = bitcast <32 x i1> %3 to i32
9662 ret i32 %4
9663 }
9665 define zeroext i32 @test_masked_vpcmpsgeb_v16i1_v32i1_mask(i16 zeroext %__u, <2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
9666 ; VLX-LABEL: test_masked_vpcmpsgeb_v16i1_v32i1_mask:
9667 ; VLX: # %bb.0: # %entry
9668 ; VLX-NEXT: kmovd %edi, %k1
9669 ; VLX-NEXT: vpcmpnltb %xmm1, %xmm0, %k0 {%k1}
9670 ; VLX-NEXT: kmovd %k0, %eax
9671 ; VLX-NEXT: retq
9672 ;
9673 ; NoVLX-LABEL: test_masked_vpcmpsgeb_v16i1_v32i1_mask:
9674 ; NoVLX: # %bb.0: # %entry
9675 ; NoVLX-NEXT: vpcmpgtb %xmm0, %xmm1, %xmm0
9676 ; NoVLX-NEXT: vpternlogq $15, %zmm0, %zmm0, %zmm0
9677 ; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0
9678 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
9679 ; NoVLX-NEXT: kmovw %k0, %eax
9680 ; NoVLX-NEXT: andl %edi, %eax
9681 ; NoVLX-NEXT: vzeroupper
9682 ; NoVLX-NEXT: retq
9683 entry:
9684 %0 = bitcast <2 x i64> %__a to <16 x i8>
9685 %1 = bitcast <2 x i64> %__b to <16 x i8>
9686 %2 = icmp sge <16 x i8> %0, %1
9687 %3 = bitcast i16 %__u to <16 x i1>
9688 %4 = and <16 x i1> %2, %3
9689 %5 = shufflevector <16 x i1> %4, <16 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>
9690 %6 = bitcast <32 x i1> %5 to i32
9691 ret i32 %6
9692 }
9694 define zeroext i32 @test_masked_vpcmpsgeb_v16i1_v32i1_mask_mem(i16 zeroext %__u, <2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
9695 ; VLX-LABEL: test_masked_vpcmpsgeb_v16i1_v32i1_mask_mem:
9696 ; VLX: # %bb.0: # %entry
9697 ; VLX-NEXT: kmovd %edi, %k1
9698 ; VLX-NEXT: vpcmpnltb (%rsi), %xmm0, %k0 {%k1}
9699 ; VLX-NEXT: kmovd %k0, %eax
9700 ; VLX-NEXT: retq
9701 ;
9702 ; NoVLX-LABEL: test_masked_vpcmpsgeb_v16i1_v32i1_mask_mem:
9703 ; NoVLX: # %bb.0: # %entry
9704 ; NoVLX-NEXT: vmovdqa (%rsi), %xmm1
9705 ; NoVLX-NEXT: vpcmpgtb %xmm0, %xmm1, %xmm0
9706 ; NoVLX-NEXT: vpternlogq $15, %zmm0, %zmm0, %zmm0
9707 ; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0
9708 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
9709 ; NoVLX-NEXT: kmovw %k0, %eax
9710 ; NoVLX-NEXT: andl %edi, %eax
9711 ; NoVLX-NEXT: vzeroupper
9712 ; NoVLX-NEXT: retq
9713 entry:
9714 %0 = bitcast <2 x i64> %__a to <16 x i8>
9715 %load = load <2 x i64>, <2 x i64>* %__b
9716 %1 = bitcast <2 x i64> %load to <16 x i8>
9717 %2 = icmp sge <16 x i8> %0, %1
9718 %3 = bitcast i16 %__u to <16 x i1>
9719 %4 = and <16 x i1> %2, %3
9720 %5 = shufflevector <16 x i1> %4, <16 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>
9721 %6 = bitcast <32 x i1> %5 to i32
9722 ret i32 %6
9723 }
9726 define zeroext i64 @test_vpcmpsgeb_v16i1_v64i1_mask(<2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
9727 ; VLX-LABEL: test_vpcmpsgeb_v16i1_v64i1_mask:
9728 ; VLX: # %bb.0: # %entry
9729 ; VLX-NEXT: vpcmpnltb %xmm1, %xmm0, %k0
9730 ; VLX-NEXT: kmovq %k0, %rax
9731 ; VLX-NEXT: retq
9732 ;
9733 ; NoVLX-LABEL: test_vpcmpsgeb_v16i1_v64i1_mask:
9734 ; NoVLX: # %bb.0: # %entry
9735 ; NoVLX-NEXT: vpcmpgtb %xmm0, %xmm1, %xmm0
9736 ; NoVLX-NEXT: vpternlogq $15, %zmm0, %zmm0, %zmm0
9737 ; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0
9738 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
9739 ; NoVLX-NEXT: kmovw %k0, %eax
9740 ; NoVLX-NEXT: vzeroupper
9741 ; NoVLX-NEXT: retq
9742 entry:
9743 %0 = bitcast <2 x i64> %__a to <16 x i8>
9744 %1 = bitcast <2 x i64> %__b to <16 x i8>
9745 %2 = icmp sge <16 x i8> %0, %1
9746 %3 = shufflevector <16 x i1> %2, <16 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>
9747 %4 = bitcast <64 x i1> %3 to i64
9748 ret i64 %4
9749 }
9751 define zeroext i64 @test_vpcmpsgeb_v16i1_v64i1_mask_mem(<2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
9752 ; VLX-LABEL: test_vpcmpsgeb_v16i1_v64i1_mask_mem:
9753 ; VLX: # %bb.0: # %entry
9754 ; VLX-NEXT: vpcmpnltb (%rdi), %xmm0, %k0
9755 ; VLX-NEXT: kmovq %k0, %rax
9756 ; VLX-NEXT: retq
9757 ;
9758 ; NoVLX-LABEL: test_vpcmpsgeb_v16i1_v64i1_mask_mem:
9759 ; NoVLX: # %bb.0: # %entry
9760 ; NoVLX-NEXT: vmovdqa (%rdi), %xmm1
9761 ; NoVLX-NEXT: vpcmpgtb %xmm0, %xmm1, %xmm0
9762 ; NoVLX-NEXT: vpternlogq $15, %zmm0, %zmm0, %zmm0
9763 ; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0
9764 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
9765 ; NoVLX-NEXT: kmovw %k0, %eax
9766 ; NoVLX-NEXT: vzeroupper
9767 ; NoVLX-NEXT: retq
9768 entry:
9769 %0 = bitcast <2 x i64> %__a to <16 x i8>
9770 %load = load <2 x i64>, <2 x i64>* %__b
9771 %1 = bitcast <2 x i64> %load to <16 x i8>
9772 %2 = icmp sge <16 x i8> %0, %1
9773 %3 = shufflevector <16 x i1> %2, <16 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>
9774 %4 = bitcast <64 x i1> %3 to i64
9775 ret i64 %4
9776 }
9778 define zeroext i64 @test_masked_vpcmpsgeb_v16i1_v64i1_mask(i16 zeroext %__u, <2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
9779 ; VLX-LABEL: test_masked_vpcmpsgeb_v16i1_v64i1_mask:
9780 ; VLX: # %bb.0: # %entry
9781 ; VLX-NEXT: kmovd %edi, %k1
9782 ; VLX-NEXT: vpcmpnltb %xmm1, %xmm0, %k0 {%k1}
9783 ; VLX-NEXT: kmovq %k0, %rax
9784 ; VLX-NEXT: retq
9785 ;
9786 ; NoVLX-LABEL: test_masked_vpcmpsgeb_v16i1_v64i1_mask:
9787 ; NoVLX: # %bb.0: # %entry
9788 ; NoVLX-NEXT: vpcmpgtb %xmm0, %xmm1, %xmm0
9789 ; NoVLX-NEXT: vpternlogq $15, %zmm0, %zmm0, %zmm0
9790 ; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0
9791 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
9792 ; NoVLX-NEXT: kmovw %k0, %eax
9793 ; NoVLX-NEXT: andl %edi, %eax
9794 ; NoVLX-NEXT: vzeroupper
9795 ; NoVLX-NEXT: retq
9796 entry:
9797 %0 = bitcast <2 x i64> %__a to <16 x i8>
9798 %1 = bitcast <2 x i64> %__b to <16 x i8>
9799 %2 = icmp sge <16 x i8> %0, %1
9800 %3 = bitcast i16 %__u to <16 x i1>
9801 %4 = and <16 x i1> %2, %3
9802 %5 = shufflevector <16 x i1> %4, <16 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>
9803 %6 = bitcast <64 x i1> %5 to i64
9804 ret i64 %6
9805 }
9807 define zeroext i64 @test_masked_vpcmpsgeb_v16i1_v64i1_mask_mem(i16 zeroext %__u, <2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
9808 ; VLX-LABEL: test_masked_vpcmpsgeb_v16i1_v64i1_mask_mem:
9809 ; VLX: # %bb.0: # %entry
9810 ; VLX-NEXT: kmovd %edi, %k1
9811 ; VLX-NEXT: vpcmpnltb (%rsi), %xmm0, %k0 {%k1}
9812 ; VLX-NEXT: kmovq %k0, %rax
9813 ; VLX-NEXT: retq
9814 ;
9815 ; NoVLX-LABEL: test_masked_vpcmpsgeb_v16i1_v64i1_mask_mem:
9816 ; NoVLX: # %bb.0: # %entry
9817 ; NoVLX-NEXT: vmovdqa (%rsi), %xmm1
9818 ; NoVLX-NEXT: vpcmpgtb %xmm0, %xmm1, %xmm0
9819 ; NoVLX-NEXT: vpternlogq $15, %zmm0, %zmm0, %zmm0
9820 ; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0
9821 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
9822 ; NoVLX-NEXT: kmovw %k0, %eax
9823 ; NoVLX-NEXT: andl %edi, %eax
9824 ; NoVLX-NEXT: vzeroupper
9825 ; NoVLX-NEXT: retq
9826 entry:
9827 %0 = bitcast <2 x i64> %__a to <16 x i8>
9828 %load = load <2 x i64>, <2 x i64>* %__b
9829 %1 = bitcast <2 x i64> %load to <16 x i8>
9830 %2 = icmp sge <16 x i8> %0, %1
9831 %3 = bitcast i16 %__u to <16 x i1>
9832 %4 = and <16 x i1> %2, %3
9833 %5 = shufflevector <16 x i1> %4, <16 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>
9834 %6 = bitcast <64 x i1> %5 to i64
9835 ret i64 %6
9836 }
9839 define zeroext i64 @test_vpcmpsgeb_v32i1_v64i1_mask(<4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
9840 ; VLX-LABEL: test_vpcmpsgeb_v32i1_v64i1_mask:
9841 ; VLX: # %bb.0: # %entry
9842 ; VLX-NEXT: vpcmpnltb %ymm1, %ymm0, %k0
9843 ; VLX-NEXT: kmovq %k0, %rax
9844 ; VLX-NEXT: vzeroupper
9845 ; VLX-NEXT: retq
9846 ;
9847 ; NoVLX-LABEL: test_vpcmpsgeb_v32i1_v64i1_mask:
9848 ; NoVLX: # %bb.0: # %entry
9849 ; NoVLX-NEXT: vpcmpgtb %ymm0, %ymm1, %ymm0
9850 ; NoVLX-NEXT: vpternlogq $15, %zmm0, %zmm0, %zmm0
9851 ; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm1
9852 ; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
9853 ; NoVLX-NEXT: kmovw %k0, %ecx
9854 ; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm0
9855 ; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0
9856 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
9857 ; NoVLX-NEXT: kmovw %k0, %eax
9858 ; NoVLX-NEXT: shll $16, %eax
9859 ; NoVLX-NEXT: orl %ecx, %eax
9860 ; NoVLX-NEXT: vzeroupper
9861 ; NoVLX-NEXT: retq
9862 entry:
9863 %0 = bitcast <4 x i64> %__a to <32 x i8>
9864 %1 = bitcast <4 x i64> %__b to <32 x i8>
9865 %2 = icmp sge <32 x i8> %0, %1
9866 %3 = shufflevector <32 x i1> %2, <32 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63>
9867 %4 = bitcast <64 x i1> %3 to i64
9868 ret i64 %4
9869 }
9871 define zeroext i64 @test_vpcmpsgeb_v32i1_v64i1_mask_mem(<4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
9872 ; VLX-LABEL: test_vpcmpsgeb_v32i1_v64i1_mask_mem:
9873 ; VLX: # %bb.0: # %entry
9874 ; VLX-NEXT: vpcmpnltb (%rdi), %ymm0, %k0
9875 ; VLX-NEXT: kmovq %k0, %rax
9876 ; VLX-NEXT: vzeroupper
9877 ; VLX-NEXT: retq
9878 ;
9879 ; NoVLX-LABEL: test_vpcmpsgeb_v32i1_v64i1_mask_mem:
9880 ; NoVLX: # %bb.0: # %entry
9881 ; NoVLX-NEXT: vmovdqa (%rdi), %ymm1
9882 ; NoVLX-NEXT: vpcmpgtb %ymm0, %ymm1, %ymm0
9883 ; NoVLX-NEXT: vpternlogq $15, %zmm0, %zmm0, %zmm0
9884 ; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm1
9885 ; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
9886 ; NoVLX-NEXT: kmovw %k0, %ecx
9887 ; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm0
9888 ; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0
9889 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
9890 ; NoVLX-NEXT: kmovw %k0, %eax
9891 ; NoVLX-NEXT: shll $16, %eax
9892 ; NoVLX-NEXT: orl %ecx, %eax
9893 ; NoVLX-NEXT: vzeroupper
9894 ; NoVLX-NEXT: retq
9895 entry:
9896 %0 = bitcast <4 x i64> %__a to <32 x i8>
9897 %load = load <4 x i64>, <4 x i64>* %__b
9898 %1 = bitcast <4 x i64> %load to <32 x i8>
9899 %2 = icmp sge <32 x i8> %0, %1
9900 %3 = shufflevector <32 x i1> %2, <32 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63>
9901 %4 = bitcast <64 x i1> %3 to i64
9902 ret i64 %4
9903 }
9905 define zeroext i64 @test_masked_vpcmpsgeb_v32i1_v64i1_mask(i32 zeroext %__u, <4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
9906 ; VLX-LABEL: test_masked_vpcmpsgeb_v32i1_v64i1_mask:
9907 ; VLX: # %bb.0: # %entry
9908 ; VLX-NEXT: kmovd %edi, %k1
9909 ; VLX-NEXT: vpcmpnltb %ymm1, %ymm0, %k0 {%k1}
9910 ; VLX-NEXT: kmovq %k0, %rax
9911 ; VLX-NEXT: vzeroupper
9912 ; VLX-NEXT: retq
9913 ;
9914 ; NoVLX-LABEL: test_masked_vpcmpsgeb_v32i1_v64i1_mask:
9915 ; NoVLX: # %bb.0: # %entry
9916 ; NoVLX-NEXT: vpcmpgtb %ymm0, %ymm1, %ymm0
9917 ; NoVLX-NEXT: vpternlogq $15, %zmm0, %zmm0, %zmm0
9918 ; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm1
9919 ; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
9920 ; NoVLX-NEXT: kmovw %k0, %eax
9921 ; NoVLX-NEXT: andl %edi, %eax
9922 ; NoVLX-NEXT: shrl $16, %edi
9923 ; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm0
9924 ; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0
9925 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
9926 ; NoVLX-NEXT: kmovw %k0, %ecx
9927 ; NoVLX-NEXT: andl %edi, %ecx
9928 ; NoVLX-NEXT: shll $16, %ecx
9929 ; NoVLX-NEXT: movzwl %ax, %eax
9930 ; NoVLX-NEXT: orl %ecx, %eax
9931 ; NoVLX-NEXT: vzeroupper
9932 ; NoVLX-NEXT: retq
9933 entry:
9934 %0 = bitcast <4 x i64> %__a to <32 x i8>
9935 %1 = bitcast <4 x i64> %__b to <32 x i8>
9936 %2 = icmp sge <32 x i8> %0, %1
9937 %3 = bitcast i32 %__u to <32 x i1>
9938 %4 = and <32 x i1> %2, %3
9939 %5 = shufflevector <32 x i1> %4, <32 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63>
9940 %6 = bitcast <64 x i1> %5 to i64
9941 ret i64 %6
9942 }
9944 define zeroext i64 @test_masked_vpcmpsgeb_v32i1_v64i1_mask_mem(i32 zeroext %__u, <4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
9945 ; VLX-LABEL: test_masked_vpcmpsgeb_v32i1_v64i1_mask_mem:
9946 ; VLX: # %bb.0: # %entry
9947 ; VLX-NEXT: kmovd %edi, %k1
9948 ; VLX-NEXT: vpcmpnltb (%rsi), %ymm0, %k0 {%k1}
9949 ; VLX-NEXT: kmovq %k0, %rax
9950 ; VLX-NEXT: vzeroupper
9951 ; VLX-NEXT: retq
9952 ;
9953 ; NoVLX-LABEL: test_masked_vpcmpsgeb_v32i1_v64i1_mask_mem:
9954 ; NoVLX: # %bb.0: # %entry
9955 ; NoVLX-NEXT: vmovdqa (%rsi), %ymm1
9956 ; NoVLX-NEXT: vpcmpgtb %ymm0, %ymm1, %ymm0
9957 ; NoVLX-NEXT: vpternlogq $15, %zmm0, %zmm0, %zmm0
9958 ; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm1
9959 ; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
9960 ; NoVLX-NEXT: kmovw %k0, %eax
9961 ; NoVLX-NEXT: andl %edi, %eax
9962 ; NoVLX-NEXT: shrl $16, %edi
9963 ; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm0
9964 ; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0
9965 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
9966 ; NoVLX-NEXT: kmovw %k0, %ecx
9967 ; NoVLX-NEXT: andl %edi, %ecx
9968 ; NoVLX-NEXT: shll $16, %ecx
9969 ; NoVLX-NEXT: movzwl %ax, %eax
9970 ; NoVLX-NEXT: orl %ecx, %eax
9971 ; NoVLX-NEXT: vzeroupper
9972 ; NoVLX-NEXT: retq
9973 entry:
9974 %0 = bitcast <4 x i64> %__a to <32 x i8>
9975 %load = load <4 x i64>, <4 x i64>* %__b
9976 %1 = bitcast <4 x i64> %load to <32 x i8>
9977 %2 = icmp sge <32 x i8> %0, %1
9978 %3 = bitcast i32 %__u to <32 x i1>
9979 %4 = and <32 x i1> %2, %3
9980 %5 = shufflevector <32 x i1> %4, <32 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63>
9981 %6 = bitcast <64 x i1> %5 to i64
9982 ret i64 %6
9983 }
9986 define zeroext i16 @test_vpcmpsgew_v8i1_v16i1_mask(<2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
9987 ; VLX-LABEL: test_vpcmpsgew_v8i1_v16i1_mask:
9988 ; VLX: # %bb.0: # %entry
9989 ; VLX-NEXT: vpcmpnltw %xmm1, %xmm0, %k0
9990 ; VLX-NEXT: kmovd %k0, %eax
9991 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
9992 ; VLX-NEXT: retq
9993 ;
9994 ; NoVLX-LABEL: test_vpcmpsgew_v8i1_v16i1_mask:
9995 ; NoVLX: # %bb.0: # %entry
9996 ; NoVLX-NEXT: vpcmpgtw %xmm0, %xmm1, %xmm0
9997 ; NoVLX-NEXT: vpternlogq $15, %zmm0, %zmm0, %zmm0
9998 ; NoVLX-NEXT: vpmovsxwq %xmm0, %zmm0
9999 ; NoVLX-NEXT: vptestmq %zmm0, %zmm0, %k0
10000 ; NoVLX-NEXT: kmovw %k0, %eax
10001 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
10002 ; NoVLX-NEXT: vzeroupper
10003 ; NoVLX-NEXT: retq
10004 entry:
10005 %0 = bitcast <2 x i64> %__a to <8 x i16>
10006 %1 = bitcast <2 x i64> %__b to <8 x i16>
10007 %2 = icmp sge <8 x i16> %0, %1
10008 %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
10009 %4 = bitcast <16 x i1> %3 to i16
10010 ret i16 %4
10011 }
10013 define zeroext i16 @test_vpcmpsgew_v8i1_v16i1_mask_mem(<2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
10014 ; VLX-LABEL: test_vpcmpsgew_v8i1_v16i1_mask_mem:
10015 ; VLX: # %bb.0: # %entry
10016 ; VLX-NEXT: vpcmpnltw (%rdi), %xmm0, %k0
10017 ; VLX-NEXT: kmovd %k0, %eax
10018 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
10019 ; VLX-NEXT: retq
10020 ;
10021 ; NoVLX-LABEL: test_vpcmpsgew_v8i1_v16i1_mask_mem:
10022 ; NoVLX: # %bb.0: # %entry
10023 ; NoVLX-NEXT: vmovdqa (%rdi), %xmm1
10024 ; NoVLX-NEXT: vpcmpgtw %xmm0, %xmm1, %xmm0
10025 ; NoVLX-NEXT: vpternlogq $15, %zmm0, %zmm0, %zmm0
10026 ; NoVLX-NEXT: vpmovsxwq %xmm0, %zmm0
10027 ; NoVLX-NEXT: vptestmq %zmm0, %zmm0, %k0
10028 ; NoVLX-NEXT: kmovw %k0, %eax
10029 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
10030 ; NoVLX-NEXT: vzeroupper
10031 ; NoVLX-NEXT: retq
10032 entry:
10033 %0 = bitcast <2 x i64> %__a to <8 x i16>
10034 %load = load <2 x i64>, <2 x i64>* %__b
10035 %1 = bitcast <2 x i64> %load to <8 x i16>
10036 %2 = icmp sge <8 x i16> %0, %1
10037 %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
10038 %4 = bitcast <16 x i1> %3 to i16
10039 ret i16 %4
10040 }
10042 define zeroext i16 @test_masked_vpcmpsgew_v8i1_v16i1_mask(i8 zeroext %__u, <2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
10043 ; VLX-LABEL: test_masked_vpcmpsgew_v8i1_v16i1_mask:
10044 ; VLX: # %bb.0: # %entry
10045 ; VLX-NEXT: kmovd %edi, %k1
10046 ; VLX-NEXT: vpcmpnltw %xmm1, %xmm0, %k0 {%k1}
10047 ; VLX-NEXT: kmovd %k0, %eax
10048 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
10049 ; VLX-NEXT: retq
10050 ;
10051 ; NoVLX-LABEL: test_masked_vpcmpsgew_v8i1_v16i1_mask:
10052 ; NoVLX: # %bb.0: # %entry
10053 ; NoVLX-NEXT: vpcmpgtw %xmm0, %xmm1, %xmm0
10054 ; NoVLX-NEXT: vpternlogq $15, %zmm0, %zmm0, %zmm0
10055 ; NoVLX-NEXT: vpmovsxwq %xmm0, %zmm0
10056 ; NoVLX-NEXT: kmovw %edi, %k1
10057 ; NoVLX-NEXT: vptestmq %zmm0, %zmm0, %k0 {%k1}
10058 ; NoVLX-NEXT: kmovw %k0, %eax
10059 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
10060 ; NoVLX-NEXT: vzeroupper
10061 ; NoVLX-NEXT: retq
10062 entry:
10063 %0 = bitcast <2 x i64> %__a to <8 x i16>
10064 %1 = bitcast <2 x i64> %__b to <8 x i16>
10065 %2 = icmp sge <8 x i16> %0, %1
10066 %3 = bitcast i8 %__u to <8 x i1>
10067 %4 = and <8 x i1> %2, %3
10068 %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
10069 %6 = bitcast <16 x i1> %5 to i16
10070 ret i16 %6
10071 }
10073 define zeroext i16 @test_masked_vpcmpsgew_v8i1_v16i1_mask_mem(i8 zeroext %__u, <2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
10074 ; VLX-LABEL: test_masked_vpcmpsgew_v8i1_v16i1_mask_mem:
10075 ; VLX: # %bb.0: # %entry
10076 ; VLX-NEXT: kmovd %edi, %k1
10077 ; VLX-NEXT: vpcmpnltw (%rsi), %xmm0, %k0 {%k1}
10078 ; VLX-NEXT: kmovd %k0, %eax
10079 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
10080 ; VLX-NEXT: retq
10081 ;
10082 ; NoVLX-LABEL: test_masked_vpcmpsgew_v8i1_v16i1_mask_mem:
10083 ; NoVLX: # %bb.0: # %entry
10084 ; NoVLX-NEXT: vmovdqa (%rsi), %xmm1
10085 ; NoVLX-NEXT: vpcmpgtw %xmm0, %xmm1, %xmm0
10086 ; NoVLX-NEXT: vpternlogq $15, %zmm0, %zmm0, %zmm0
10087 ; NoVLX-NEXT: vpmovsxwq %xmm0, %zmm0
10088 ; NoVLX-NEXT: kmovw %edi, %k1
10089 ; NoVLX-NEXT: vptestmq %zmm0, %zmm0, %k0 {%k1}
10090 ; NoVLX-NEXT: kmovw %k0, %eax
10091 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
10092 ; NoVLX-NEXT: vzeroupper
10093 ; NoVLX-NEXT: retq
10094 entry:
10095 %0 = bitcast <2 x i64> %__a to <8 x i16>
10096 %load = load <2 x i64>, <2 x i64>* %__b
10097 %1 = bitcast <2 x i64> %load to <8 x i16>
10098 %2 = icmp sge <8 x i16> %0, %1
10099 %3 = bitcast i8 %__u to <8 x i1>
10100 %4 = and <8 x i1> %2, %3
10101 %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
10102 %6 = bitcast <16 x i1> %5 to i16
10103 ret i16 %6
10104 }
10107 define zeroext i32 @test_vpcmpsgew_v8i1_v32i1_mask(<2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
10108 ; VLX-LABEL: test_vpcmpsgew_v8i1_v32i1_mask:
10109 ; VLX: # %bb.0: # %entry
10110 ; VLX-NEXT: vpcmpnltw %xmm1, %xmm0, %k0
10111 ; VLX-NEXT: kmovd %k0, %eax
10112 ; VLX-NEXT: retq
10113 ;
10114 ; NoVLX-LABEL: test_vpcmpsgew_v8i1_v32i1_mask:
10115 ; NoVLX: # %bb.0: # %entry
10116 ; NoVLX-NEXT: vpcmpgtw %xmm0, %xmm1, %xmm0
10117 ; NoVLX-NEXT: vpternlogq $15, %zmm0, %zmm0, %zmm0
10118 ; NoVLX-NEXT: vpmovsxwq %xmm0, %zmm0
10119 ; NoVLX-NEXT: vptestmq %zmm0, %zmm0, %k0
10120 ; NoVLX-NEXT: kmovw %k0, %eax
10121 ; NoVLX-NEXT: vzeroupper
10122 ; NoVLX-NEXT: retq
10123 entry:
10124 %0 = bitcast <2 x i64> %__a to <8 x i16>
10125 %1 = bitcast <2 x i64> %__b to <8 x i16>
10126 %2 = icmp sge <8 x i16> %0, %1
10127 %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
10128 %4 = bitcast <32 x i1> %3 to i32
10129 ret i32 %4
10130 }
10132 define zeroext i32 @test_vpcmpsgew_v8i1_v32i1_mask_mem(<2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
10133 ; VLX-LABEL: test_vpcmpsgew_v8i1_v32i1_mask_mem:
10134 ; VLX: # %bb.0: # %entry
10135 ; VLX-NEXT: vpcmpnltw (%rdi), %xmm0, %k0
10136 ; VLX-NEXT: kmovd %k0, %eax
10137 ; VLX-NEXT: retq
10138 ;
10139 ; NoVLX-LABEL: test_vpcmpsgew_v8i1_v32i1_mask_mem:
10140 ; NoVLX: # %bb.0: # %entry
10141 ; NoVLX-NEXT: vmovdqa (%rdi), %xmm1
10142 ; NoVLX-NEXT: vpcmpgtw %xmm0, %xmm1, %xmm0
10143 ; NoVLX-NEXT: vpternlogq $15, %zmm0, %zmm0, %zmm0
10144 ; NoVLX-NEXT: vpmovsxwq %xmm0, %zmm0
10145 ; NoVLX-NEXT: vptestmq %zmm0, %zmm0, %k0
10146 ; NoVLX-NEXT: kmovw %k0, %eax
10147 ; NoVLX-NEXT: vzeroupper
10148 ; NoVLX-NEXT: retq
10149 entry:
10150 %0 = bitcast <2 x i64> %__a to <8 x i16>
10151 %load = load <2 x i64>, <2 x i64>* %__b
10152 %1 = bitcast <2 x i64> %load to <8 x i16>
10153 %2 = icmp sge <8 x i16> %0, %1
10154 %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
10155 %4 = bitcast <32 x i1> %3 to i32
10156 ret i32 %4
10157 }
10159 define zeroext i32 @test_masked_vpcmpsgew_v8i1_v32i1_mask(i8 zeroext %__u, <2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
10160 ; VLX-LABEL: test_masked_vpcmpsgew_v8i1_v32i1_mask:
10161 ; VLX: # %bb.0: # %entry
10162 ; VLX-NEXT: kmovd %edi, %k1
10163 ; VLX-NEXT: vpcmpnltw %xmm1, %xmm0, %k0 {%k1}
10164 ; VLX-NEXT: kmovd %k0, %eax
10165 ; VLX-NEXT: retq
10166 ;
10167 ; NoVLX-LABEL: test_masked_vpcmpsgew_v8i1_v32i1_mask:
10168 ; NoVLX: # %bb.0: # %entry
10169 ; NoVLX-NEXT: vpcmpgtw %xmm0, %xmm1, %xmm0
10170 ; NoVLX-NEXT: vpternlogq $15, %zmm0, %zmm0, %zmm0
10171 ; NoVLX-NEXT: vpmovsxwq %xmm0, %zmm0
10172 ; NoVLX-NEXT: kmovw %edi, %k1
10173 ; NoVLX-NEXT: vptestmq %zmm0, %zmm0, %k0 {%k1}
10174 ; NoVLX-NEXT: kmovw %k0, %eax
10175 ; NoVLX-NEXT: vzeroupper
10176 ; NoVLX-NEXT: retq
10177 entry:
10178 %0 = bitcast <2 x i64> %__a to <8 x i16>
10179 %1 = bitcast <2 x i64> %__b to <8 x i16>
10180 %2 = icmp sge <8 x i16> %0, %1
10181 %3 = bitcast i8 %__u to <8 x i1>
10182 %4 = and <8 x i1> %2, %3
10183 %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
10184 %6 = bitcast <32 x i1> %5 to i32
10185 ret i32 %6
10186 }
10188 define zeroext i32 @test_masked_vpcmpsgew_v8i1_v32i1_mask_mem(i8 zeroext %__u, <2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
10189 ; VLX-LABEL: test_masked_vpcmpsgew_v8i1_v32i1_mask_mem:
10190 ; VLX: # %bb.0: # %entry
10191 ; VLX-NEXT: kmovd %edi, %k1
10192 ; VLX-NEXT: vpcmpnltw (%rsi), %xmm0, %k0 {%k1}
10193 ; VLX-NEXT: kmovd %k0, %eax
10194 ; VLX-NEXT: retq
10195 ;
10196 ; NoVLX-LABEL: test_masked_vpcmpsgew_v8i1_v32i1_mask_mem:
10197 ; NoVLX: # %bb.0: # %entry
10198 ; NoVLX-NEXT: vmovdqa (%rsi), %xmm1
10199 ; NoVLX-NEXT: vpcmpgtw %xmm0, %xmm1, %xmm0
10200 ; NoVLX-NEXT: vpternlogq $15, %zmm0, %zmm0, %zmm0
10201 ; NoVLX-NEXT: vpmovsxwq %xmm0, %zmm0
10202 ; NoVLX-NEXT: kmovw %edi, %k1
10203 ; NoVLX-NEXT: vptestmq %zmm0, %zmm0, %k0 {%k1}
10204 ; NoVLX-NEXT: kmovw %k0, %eax
10205 ; NoVLX-NEXT: vzeroupper
10206 ; NoVLX-NEXT: retq
10207 entry:
10208 %0 = bitcast <2 x i64> %__a to <8 x i16>
10209 %load = load <2 x i64>, <2 x i64>* %__b
10210 %1 = bitcast <2 x i64> %load to <8 x i16>
10211 %2 = icmp sge <8 x i16> %0, %1
10212 %3 = bitcast i8 %__u to <8 x i1>
10213 %4 = and <8 x i1> %2, %3
10214 %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
10215 %6 = bitcast <32 x i1> %5 to i32
10216 ret i32 %6
10217 }
10220 define zeroext i64 @test_vpcmpsgew_v8i1_v64i1_mask(<2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
10221 ; VLX-LABEL: test_vpcmpsgew_v8i1_v64i1_mask:
10222 ; VLX: # %bb.0: # %entry
10223 ; VLX-NEXT: vpcmpnltw %xmm1, %xmm0, %k0
10224 ; VLX-NEXT: kmovq %k0, %rax
10225 ; VLX-NEXT: retq
10226 ;
10227 ; NoVLX-LABEL: test_vpcmpsgew_v8i1_v64i1_mask:
10228 ; NoVLX: # %bb.0: # %entry
10229 ; NoVLX-NEXT: vpcmpgtw %xmm0, %xmm1, %xmm0
10230 ; NoVLX-NEXT: vpternlogq $15, %zmm0, %zmm0, %zmm0
10231 ; NoVLX-NEXT: vpmovsxwq %xmm0, %zmm0
10232 ; NoVLX-NEXT: vptestmq %zmm0, %zmm0, %k0
10233 ; NoVLX-NEXT: kmovw %k0, %eax
10234 ; NoVLX-NEXT: vzeroupper
10235 ; NoVLX-NEXT: retq
10236 entry:
10237 %0 = bitcast <2 x i64> %__a to <8 x i16>
10238 %1 = bitcast <2 x i64> %__b to <8 x i16>
10239 %2 = icmp sge <8 x i16> %0, %1
10240 %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
10241 %4 = bitcast <64 x i1> %3 to i64
10242 ret i64 %4
10243 }
10245 define zeroext i64 @test_vpcmpsgew_v8i1_v64i1_mask_mem(<2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
10246 ; VLX-LABEL: test_vpcmpsgew_v8i1_v64i1_mask_mem:
10247 ; VLX: # %bb.0: # %entry
10248 ; VLX-NEXT: vpcmpnltw (%rdi), %xmm0, %k0
10249 ; VLX-NEXT: kmovq %k0, %rax
10250 ; VLX-NEXT: retq
10251 ;
10252 ; NoVLX-LABEL: test_vpcmpsgew_v8i1_v64i1_mask_mem:
10253 ; NoVLX: # %bb.0: # %entry
10254 ; NoVLX-NEXT: vmovdqa (%rdi), %xmm1
10255 ; NoVLX-NEXT: vpcmpgtw %xmm0, %xmm1, %xmm0
10256 ; NoVLX-NEXT: vpternlogq $15, %zmm0, %zmm0, %zmm0
10257 ; NoVLX-NEXT: vpmovsxwq %xmm0, %zmm0
10258 ; NoVLX-NEXT: vptestmq %zmm0, %zmm0, %k0
10259 ; NoVLX-NEXT: kmovw %k0, %eax
10260 ; NoVLX-NEXT: vzeroupper
10261 ; NoVLX-NEXT: retq
10262 entry:
10263 %0 = bitcast <2 x i64> %__a to <8 x i16>
10264 %load = load <2 x i64>, <2 x i64>* %__b
10265 %1 = bitcast <2 x i64> %load to <8 x i16>
10266 %2 = icmp sge <8 x i16> %0, %1
10267 %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
10268 %4 = bitcast <64 x i1> %3 to i64
10269 ret i64 %4
10270 }
10272 define zeroext i64 @test_masked_vpcmpsgew_v8i1_v64i1_mask(i8 zeroext %__u, <2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
10273 ; VLX-LABEL: test_masked_vpcmpsgew_v8i1_v64i1_mask:
10274 ; VLX: # %bb.0: # %entry
10275 ; VLX-NEXT: kmovd %edi, %k1
10276 ; VLX-NEXT: vpcmpnltw %xmm1, %xmm0, %k0 {%k1}
10277 ; VLX-NEXT: kmovq %k0, %rax
10278 ; VLX-NEXT: retq
10279 ;
10280 ; NoVLX-LABEL: test_masked_vpcmpsgew_v8i1_v64i1_mask:
10281 ; NoVLX: # %bb.0: # %entry
10282 ; NoVLX-NEXT: vpcmpgtw %xmm0, %xmm1, %xmm0
10283 ; NoVLX-NEXT: vpternlogq $15, %zmm0, %zmm0, %zmm0
10284 ; NoVLX-NEXT: vpmovsxwq %xmm0, %zmm0
10285 ; NoVLX-NEXT: kmovw %edi, %k1
10286 ; NoVLX-NEXT: vptestmq %zmm0, %zmm0, %k0 {%k1}
10287 ; NoVLX-NEXT: kmovw %k0, %eax
10288 ; NoVLX-NEXT: vzeroupper
10289 ; NoVLX-NEXT: retq
10290 entry:
10291 %0 = bitcast <2 x i64> %__a to <8 x i16>
10292 %1 = bitcast <2 x i64> %__b to <8 x i16>
10293 %2 = icmp sge <8 x i16> %0, %1
10294 %3 = bitcast i8 %__u to <8 x i1>
10295 %4 = and <8 x i1> %2, %3
10296 %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
10297 %6 = bitcast <64 x i1> %5 to i64
10298 ret i64 %6
10299 }
10301 define zeroext i64 @test_masked_vpcmpsgew_v8i1_v64i1_mask_mem(i8 zeroext %__u, <2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
10302 ; VLX-LABEL: test_masked_vpcmpsgew_v8i1_v64i1_mask_mem:
10303 ; VLX: # %bb.0: # %entry
10304 ; VLX-NEXT: kmovd %edi, %k1
10305 ; VLX-NEXT: vpcmpnltw (%rsi), %xmm0, %k0 {%k1}
10306 ; VLX-NEXT: kmovq %k0, %rax
10307 ; VLX-NEXT: retq
10308 ;
10309 ; NoVLX-LABEL: test_masked_vpcmpsgew_v8i1_v64i1_mask_mem:
10310 ; NoVLX: # %bb.0: # %entry
10311 ; NoVLX-NEXT: vmovdqa (%rsi), %xmm1
10312 ; NoVLX-NEXT: vpcmpgtw %xmm0, %xmm1, %xmm0
10313 ; NoVLX-NEXT: vpternlogq $15, %zmm0, %zmm0, %zmm0
10314 ; NoVLX-NEXT: vpmovsxwq %xmm0, %zmm0
10315 ; NoVLX-NEXT: kmovw %edi, %k1
10316 ; NoVLX-NEXT: vptestmq %zmm0, %zmm0, %k0 {%k1}
10317 ; NoVLX-NEXT: kmovw %k0, %eax
10318 ; NoVLX-NEXT: vzeroupper
10319 ; NoVLX-NEXT: retq
10320 entry:
10321 %0 = bitcast <2 x i64> %__a to <8 x i16>
10322 %load = load <2 x i64>, <2 x i64>* %__b
10323 %1 = bitcast <2 x i64> %load to <8 x i16>
10324 %2 = icmp sge <8 x i16> %0, %1
10325 %3 = bitcast i8 %__u to <8 x i1>
10326 %4 = and <8 x i1> %2, %3
10327 %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
10328 %6 = bitcast <64 x i1> %5 to i64
10329 ret i64 %6
10330 }
10333 define zeroext i32 @test_vpcmpsgew_v16i1_v32i1_mask(<4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
10334 ; VLX-LABEL: test_vpcmpsgew_v16i1_v32i1_mask:
10335 ; VLX: # %bb.0: # %entry
10336 ; VLX-NEXT: vpcmpnltw %ymm1, %ymm0, %k0
10337 ; VLX-NEXT: kmovd %k0, %eax
10338 ; VLX-NEXT: vzeroupper
10339 ; VLX-NEXT: retq
10340 ;
10341 ; NoVLX-LABEL: test_vpcmpsgew_v16i1_v32i1_mask:
10342 ; NoVLX: # %bb.0: # %entry
10343 ; NoVLX-NEXT: vpcmpgtw %ymm0, %ymm1, %ymm0
10344 ; NoVLX-NEXT: vpternlogq $15, %zmm0, %zmm0, %zmm0
10345 ; NoVLX-NEXT: vpmovsxwd %ymm0, %zmm0
10346 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
10347 ; NoVLX-NEXT: kmovw %k0, %eax
10348 ; NoVLX-NEXT: vzeroupper
10349 ; NoVLX-NEXT: retq
10350 entry:
10351 %0 = bitcast <4 x i64> %__a to <16 x i16>
10352 %1 = bitcast <4 x i64> %__b to <16 x i16>
10353 %2 = icmp sge <16 x i16> %0, %1
10354 %3 = shufflevector <16 x i1> %2, <16 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>
10355 %4 = bitcast <32 x i1> %3 to i32
10356 ret i32 %4
10357 }
10359 define zeroext i32 @test_vpcmpsgew_v16i1_v32i1_mask_mem(<4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
10360 ; VLX-LABEL: test_vpcmpsgew_v16i1_v32i1_mask_mem:
10361 ; VLX: # %bb.0: # %entry
10362 ; VLX-NEXT: vpcmpnltw (%rdi), %ymm0, %k0
10363 ; VLX-NEXT: kmovd %k0, %eax
10364 ; VLX-NEXT: vzeroupper
10365 ; VLX-NEXT: retq
10366 ;
10367 ; NoVLX-LABEL: test_vpcmpsgew_v16i1_v32i1_mask_mem:
10368 ; NoVLX: # %bb.0: # %entry
10369 ; NoVLX-NEXT: vmovdqa (%rdi), %ymm1
10370 ; NoVLX-NEXT: vpcmpgtw %ymm0, %ymm1, %ymm0
10371 ; NoVLX-NEXT: vpternlogq $15, %zmm0, %zmm0, %zmm0
10372 ; NoVLX-NEXT: vpmovsxwd %ymm0, %zmm0
10373 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
10374 ; NoVLX-NEXT: kmovw %k0, %eax
10375 ; NoVLX-NEXT: vzeroupper
10376 ; NoVLX-NEXT: retq
10377 entry:
10378 %0 = bitcast <4 x i64> %__a to <16 x i16>
10379 %load = load <4 x i64>, <4 x i64>* %__b
10380 %1 = bitcast <4 x i64> %load to <16 x i16>
10381 %2 = icmp sge <16 x i16> %0, %1
10382 %3 = shufflevector <16 x i1> %2, <16 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>
10383 %4 = bitcast <32 x i1> %3 to i32
10384 ret i32 %4
10385 }
10387 define zeroext i32 @test_masked_vpcmpsgew_v16i1_v32i1_mask(i16 zeroext %__u, <4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
10388 ; VLX-LABEL: test_masked_vpcmpsgew_v16i1_v32i1_mask:
10389 ; VLX: # %bb.0: # %entry
10390 ; VLX-NEXT: kmovd %edi, %k1
10391 ; VLX-NEXT: vpcmpnltw %ymm1, %ymm0, %k0 {%k1}
10392 ; VLX-NEXT: kmovd %k0, %eax
10393 ; VLX-NEXT: vzeroupper
10394 ; VLX-NEXT: retq
10395 ;
10396 ; NoVLX-LABEL: test_masked_vpcmpsgew_v16i1_v32i1_mask:
10397 ; NoVLX: # %bb.0: # %entry
10398 ; NoVLX-NEXT: vpcmpgtw %ymm0, %ymm1, %ymm0
10399 ; NoVLX-NEXT: vpternlogq $15, %zmm0, %zmm0, %zmm0
10400 ; NoVLX-NEXT: vpmovsxwd %ymm0, %zmm0
10401 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
10402 ; NoVLX-NEXT: kmovw %k0, %eax
10403 ; NoVLX-NEXT: andl %edi, %eax
10404 ; NoVLX-NEXT: vzeroupper
10405 ; NoVLX-NEXT: retq
10406 entry:
10407 %0 = bitcast <4 x i64> %__a to <16 x i16>
10408 %1 = bitcast <4 x i64> %__b to <16 x i16>
10409 %2 = icmp sge <16 x i16> %0, %1
10410 %3 = bitcast i16 %__u to <16 x i1>
10411 %4 = and <16 x i1> %2, %3
10412 %5 = shufflevector <16 x i1> %4, <16 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>
10413 %6 = bitcast <32 x i1> %5 to i32
10414 ret i32 %6
10415 }
10417 define zeroext i32 @test_masked_vpcmpsgew_v16i1_v32i1_mask_mem(i16 zeroext %__u, <4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
10418 ; VLX-LABEL: test_masked_vpcmpsgew_v16i1_v32i1_mask_mem:
10419 ; VLX: # %bb.0: # %entry
10420 ; VLX-NEXT: kmovd %edi, %k1
10421 ; VLX-NEXT: vpcmpnltw (%rsi), %ymm0, %k0 {%k1}
10422 ; VLX-NEXT: kmovd %k0, %eax
10423 ; VLX-NEXT: vzeroupper
10424 ; VLX-NEXT: retq
10425 ;
10426 ; NoVLX-LABEL: test_masked_vpcmpsgew_v16i1_v32i1_mask_mem:
10427 ; NoVLX: # %bb.0: # %entry
10428 ; NoVLX-NEXT: vmovdqa (%rsi), %ymm1
10429 ; NoVLX-NEXT: vpcmpgtw %ymm0, %ymm1, %ymm0
10430 ; NoVLX-NEXT: vpternlogq $15, %zmm0, %zmm0, %zmm0
10431 ; NoVLX-NEXT: vpmovsxwd %ymm0, %zmm0
10432 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
10433 ; NoVLX-NEXT: kmovw %k0, %eax
10434 ; NoVLX-NEXT: andl %edi, %eax
10435 ; NoVLX-NEXT: vzeroupper
10436 ; NoVLX-NEXT: retq
10437 entry:
10438 %0 = bitcast <4 x i64> %__a to <16 x i16>
10439 %load = load <4 x i64>, <4 x i64>* %__b
10440 %1 = bitcast <4 x i64> %load to <16 x i16>
10441 %2 = icmp sge <16 x i16> %0, %1
10442 %3 = bitcast i16 %__u to <16 x i1>
10443 %4 = and <16 x i1> %2, %3
10444 %5 = shufflevector <16 x i1> %4, <16 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>
10445 %6 = bitcast <32 x i1> %5 to i32
10446 ret i32 %6
10447 }
10450 define zeroext i64 @test_vpcmpsgew_v16i1_v64i1_mask(<4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
10451 ; VLX-LABEL: test_vpcmpsgew_v16i1_v64i1_mask:
10452 ; VLX: # %bb.0: # %entry
10453 ; VLX-NEXT: vpcmpnltw %ymm1, %ymm0, %k0
10454 ; VLX-NEXT: kmovq %k0, %rax
10455 ; VLX-NEXT: vzeroupper
10456 ; VLX-NEXT: retq
10457 ;
10458 ; NoVLX-LABEL: test_vpcmpsgew_v16i1_v64i1_mask:
10459 ; NoVLX: # %bb.0: # %entry
10460 ; NoVLX-NEXT: vpcmpgtw %ymm0, %ymm1, %ymm0
10461 ; NoVLX-NEXT: vpternlogq $15, %zmm0, %zmm0, %zmm0
10462 ; NoVLX-NEXT: vpmovsxwd %ymm0, %zmm0
10463 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
10464 ; NoVLX-NEXT: kmovw %k0, %eax
10465 ; NoVLX-NEXT: vzeroupper
10466 ; NoVLX-NEXT: retq
10467 entry:
10468 %0 = bitcast <4 x i64> %__a to <16 x i16>
10469 %1 = bitcast <4 x i64> %__b to <16 x i16>
10470 %2 = icmp sge <16 x i16> %0, %1
10471 %3 = shufflevector <16 x i1> %2, <16 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>
10472 %4 = bitcast <64 x i1> %3 to i64
10473 ret i64 %4
10474 }
10476 define zeroext i64 @test_vpcmpsgew_v16i1_v64i1_mask_mem(<4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
10477 ; VLX-LABEL: test_vpcmpsgew_v16i1_v64i1_mask_mem:
10478 ; VLX: # %bb.0: # %entry
10479 ; VLX-NEXT: vpcmpnltw (%rdi), %ymm0, %k0
10480 ; VLX-NEXT: kmovq %k0, %rax
10481 ; VLX-NEXT: vzeroupper
10482 ; VLX-NEXT: retq
10483 ;
10484 ; NoVLX-LABEL: test_vpcmpsgew_v16i1_v64i1_mask_mem:
10485 ; NoVLX: # %bb.0: # %entry
10486 ; NoVLX-NEXT: vmovdqa (%rdi), %ymm1
10487 ; NoVLX-NEXT: vpcmpgtw %ymm0, %ymm1, %ymm0
10488 ; NoVLX-NEXT: vpternlogq $15, %zmm0, %zmm0, %zmm0
10489 ; NoVLX-NEXT: vpmovsxwd %ymm0, %zmm0
10490 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
10491 ; NoVLX-NEXT: kmovw %k0, %eax
10492 ; NoVLX-NEXT: vzeroupper
10493 ; NoVLX-NEXT: retq
10494 entry:
10495 %0 = bitcast <4 x i64> %__a to <16 x i16>
10496 %load = load <4 x i64>, <4 x i64>* %__b
10497 %1 = bitcast <4 x i64> %load to <16 x i16>
10498 %2 = icmp sge <16 x i16> %0, %1
10499 %3 = shufflevector <16 x i1> %2, <16 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>
10500 %4 = bitcast <64 x i1> %3 to i64
10501 ret i64 %4
10502 }
10504 define zeroext i64 @test_masked_vpcmpsgew_v16i1_v64i1_mask(i16 zeroext %__u, <4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
10505 ; VLX-LABEL: test_masked_vpcmpsgew_v16i1_v64i1_mask:
10506 ; VLX: # %bb.0: # %entry
10507 ; VLX-NEXT: kmovd %edi, %k1
10508 ; VLX-NEXT: vpcmpnltw %ymm1, %ymm0, %k0 {%k1}
10509 ; VLX-NEXT: kmovq %k0, %rax
10510 ; VLX-NEXT: vzeroupper
10511 ; VLX-NEXT: retq
10512 ;
10513 ; NoVLX-LABEL: test_masked_vpcmpsgew_v16i1_v64i1_mask:
10514 ; NoVLX: # %bb.0: # %entry
10515 ; NoVLX-NEXT: vpcmpgtw %ymm0, %ymm1, %ymm0
10516 ; NoVLX-NEXT: vpternlogq $15, %zmm0, %zmm0, %zmm0
10517 ; NoVLX-NEXT: vpmovsxwd %ymm0, %zmm0
10518 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
10519 ; NoVLX-NEXT: kmovw %k0, %eax
10520 ; NoVLX-NEXT: andl %edi, %eax
10521 ; NoVLX-NEXT: vzeroupper
10522 ; NoVLX-NEXT: retq
10523 entry:
10524 %0 = bitcast <4 x i64> %__a to <16 x i16>
10525 %1 = bitcast <4 x i64> %__b to <16 x i16>
10526 %2 = icmp sge <16 x i16> %0, %1
10527 %3 = bitcast i16 %__u to <16 x i1>
10528 %4 = and <16 x i1> %2, %3
10529 %5 = shufflevector <16 x i1> %4, <16 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>
10530 %6 = bitcast <64 x i1> %5 to i64
10531 ret i64 %6
10532 }
10534 define zeroext i64 @test_masked_vpcmpsgew_v16i1_v64i1_mask_mem(i16 zeroext %__u, <4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
10535 ; VLX-LABEL: test_masked_vpcmpsgew_v16i1_v64i1_mask_mem:
10536 ; VLX: # %bb.0: # %entry
10537 ; VLX-NEXT: kmovd %edi, %k1
10538 ; VLX-NEXT: vpcmpnltw (%rsi), %ymm0, %k0 {%k1}
10539 ; VLX-NEXT: kmovq %k0, %rax
10540 ; VLX-NEXT: vzeroupper
10541 ; VLX-NEXT: retq
10542 ;
10543 ; NoVLX-LABEL: test_masked_vpcmpsgew_v16i1_v64i1_mask_mem:
10544 ; NoVLX: # %bb.0: # %entry
10545 ; NoVLX-NEXT: vmovdqa (%rsi), %ymm1
10546 ; NoVLX-NEXT: vpcmpgtw %ymm0, %ymm1, %ymm0
10547 ; NoVLX-NEXT: vpternlogq $15, %zmm0, %zmm0, %zmm0
10548 ; NoVLX-NEXT: vpmovsxwd %ymm0, %zmm0
10549 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
10550 ; NoVLX-NEXT: kmovw %k0, %eax
10551 ; NoVLX-NEXT: andl %edi, %eax
10552 ; NoVLX-NEXT: vzeroupper
10553 ; NoVLX-NEXT: retq
10554 entry:
10555 %0 = bitcast <4 x i64> %__a to <16 x i16>
10556 %load = load <4 x i64>, <4 x i64>* %__b
10557 %1 = bitcast <4 x i64> %load to <16 x i16>
10558 %2 = icmp sge <16 x i16> %0, %1
10559 %3 = bitcast i16 %__u to <16 x i1>
10560 %4 = and <16 x i1> %2, %3
10561 %5 = shufflevector <16 x i1> %4, <16 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>
10562 %6 = bitcast <64 x i1> %5 to i64
10563 ret i64 %6
10564 }
10567 define zeroext i64 @test_vpcmpsgew_v32i1_v64i1_mask(<8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
10568 ; VLX-LABEL: test_vpcmpsgew_v32i1_v64i1_mask:
10569 ; VLX: # %bb.0: # %entry
10570 ; VLX-NEXT: vpcmpnltw %zmm1, %zmm0, %k0
10571 ; VLX-NEXT: kmovq %k0, %rax
10572 ; VLX-NEXT: vzeroupper
10573 ; VLX-NEXT: retq
10574 ;
10575 ; NoVLX-LABEL: test_vpcmpsgew_v32i1_v64i1_mask:
10576 ; NoVLX: # %bb.0: # %entry
10577 ; NoVLX-NEXT: vextracti64x4 $1, %zmm0, %ymm2
10578 ; NoVLX-NEXT: vextracti64x4 $1, %zmm1, %ymm3
10579 ; NoVLX-NEXT: vpcmpgtw %ymm2, %ymm3, %ymm2
10580 ; NoVLX-NEXT: vpcmpgtw %ymm0, %ymm1, %ymm0
10581 ; NoVLX-NEXT: vpternlogq $15, %zmm0, %zmm0, %zmm0
10582 ; NoVLX-NEXT: vpmovsxwd %ymm0, %zmm0
10583 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
10584 ; NoVLX-NEXT: kmovw %k0, %ecx
10585 ; NoVLX-NEXT: vpternlogq $15, %zmm2, %zmm2, %zmm2
10586 ; NoVLX-NEXT: vpmovsxwd %ymm2, %zmm0
10587 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
10588 ; NoVLX-NEXT: kmovw %k0, %eax
10589 ; NoVLX-NEXT: shll $16, %eax
10590 ; NoVLX-NEXT: orl %ecx, %eax
10591 ; NoVLX-NEXT: vzeroupper
10592 ; NoVLX-NEXT: retq
10593 entry:
10594 %0 = bitcast <8 x i64> %__a to <32 x i16>
10595 %1 = bitcast <8 x i64> %__b to <32 x i16>
10596 %2 = icmp sge <32 x i16> %0, %1
10597 %3 = shufflevector <32 x i1> %2, <32 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63>
10598 %4 = bitcast <64 x i1> %3 to i64
10599 ret i64 %4
10600 }
10602 define zeroext i64 @test_vpcmpsgew_v32i1_v64i1_mask_mem(<8 x i64> %__a, <8 x i64>* %__b) local_unnamed_addr {
10603 ; VLX-LABEL: test_vpcmpsgew_v32i1_v64i1_mask_mem:
10604 ; VLX: # %bb.0: # %entry
10605 ; VLX-NEXT: vpcmpnltw (%rdi), %zmm0, %k0
10606 ; VLX-NEXT: kmovq %k0, %rax
10607 ; VLX-NEXT: vzeroupper
10608 ; VLX-NEXT: retq
10609 ;
10610 ; NoVLX-LABEL: test_vpcmpsgew_v32i1_v64i1_mask_mem:
10611 ; NoVLX: # %bb.0: # %entry
10612 ; NoVLX-NEXT: vextracti64x4 $1, %zmm0, %ymm1
10613 ; NoVLX-NEXT: vmovdqa (%rdi), %ymm2
10614 ; NoVLX-NEXT: vmovdqa 32(%rdi), %ymm3
10615 ; NoVLX-NEXT: vpcmpgtw %ymm1, %ymm3, %ymm1
10616 ; NoVLX-NEXT: vpcmpgtw %ymm0, %ymm2, %ymm0
10617 ; NoVLX-NEXT: vpternlogq $15, %zmm0, %zmm0, %zmm0
10618 ; NoVLX-NEXT: vpmovsxwd %ymm0, %zmm0
10619 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
10620 ; NoVLX-NEXT: kmovw %k0, %ecx
10621 ; NoVLX-NEXT: vpternlogq $15, %zmm1, %zmm1, %zmm1
10622 ; NoVLX-NEXT: vpmovsxwd %ymm1, %zmm0
10623 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
10624 ; NoVLX-NEXT: kmovw %k0, %eax
10625 ; NoVLX-NEXT: shll $16, %eax
10626 ; NoVLX-NEXT: orl %ecx, %eax
10627 ; NoVLX-NEXT: vzeroupper
10628 ; NoVLX-NEXT: retq
10629 entry:
10630 %0 = bitcast <8 x i64> %__a to <32 x i16>
10631 %load = load <8 x i64>, <8 x i64>* %__b
10632 %1 = bitcast <8 x i64> %load to <32 x i16>
10633 %2 = icmp sge <32 x i16> %0, %1
10634 %3 = shufflevector <32 x i1> %2, <32 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63>
10635 %4 = bitcast <64 x i1> %3 to i64
10636 ret i64 %4
10637 }
10639 define zeroext i64 @test_masked_vpcmpsgew_v32i1_v64i1_mask(i32 zeroext %__u, <8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
10640 ; VLX-LABEL: test_masked_vpcmpsgew_v32i1_v64i1_mask:
10641 ; VLX: # %bb.0: # %entry
10642 ; VLX-NEXT: kmovd %edi, %k1
10643 ; VLX-NEXT: vpcmpnltw %zmm1, %zmm0, %k0 {%k1}
10644 ; VLX-NEXT: kmovq %k0, %rax
10645 ; VLX-NEXT: vzeroupper
10646 ; VLX-NEXT: retq
10647 ;
10648 ; NoVLX-LABEL: test_masked_vpcmpsgew_v32i1_v64i1_mask:
10649 ; NoVLX: # %bb.0: # %entry
10650 ; NoVLX-NEXT: vpcmpgtw %ymm0, %ymm1, %ymm2
10651 ; NoVLX-NEXT: vpternlogq $15, %zmm2, %zmm2, %zmm2
10652 ; NoVLX-NEXT: vpmovsxwd %ymm2, %zmm2
10653 ; NoVLX-NEXT: vptestmd %zmm2, %zmm2, %k0
10654 ; NoVLX-NEXT: kmovw %k0, %eax
10655 ; NoVLX-NEXT: andl %edi, %eax
10656 ; NoVLX-NEXT: shrl $16, %edi
10657 ; NoVLX-NEXT: vextracti64x4 $1, %zmm0, %ymm0
10658 ; NoVLX-NEXT: vextracti64x4 $1, %zmm1, %ymm1
10659 ; NoVLX-NEXT: vpcmpgtw %ymm0, %ymm1, %ymm0
10660 ; NoVLX-NEXT: vpternlogq $15, %zmm0, %zmm0, %zmm0
10661 ; NoVLX-NEXT: vpmovsxwd %ymm0, %zmm0
10662 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
10663 ; NoVLX-NEXT: kmovw %k0, %ecx
10664 ; NoVLX-NEXT: andl %edi, %ecx
10665 ; NoVLX-NEXT: shll $16, %ecx
10666 ; NoVLX-NEXT: movzwl %ax, %eax
10667 ; NoVLX-NEXT: orl %ecx, %eax
10668 ; NoVLX-NEXT: vzeroupper
10669 ; NoVLX-NEXT: retq
10670 entry:
10671 %0 = bitcast <8 x i64> %__a to <32 x i16>
10672 %1 = bitcast <8 x i64> %__b to <32 x i16>
10673 %2 = icmp sge <32 x i16> %0, %1
10674 %3 = bitcast i32 %__u to <32 x i1>
10675 %4 = and <32 x i1> %2, %3
10676 %5 = shufflevector <32 x i1> %4, <32 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63>
10677 %6 = bitcast <64 x i1> %5 to i64
10678 ret i64 %6
10679 }
10681 define zeroext i64 @test_masked_vpcmpsgew_v32i1_v64i1_mask_mem(i32 zeroext %__u, <8 x i64> %__a, <8 x i64>* %__b) local_unnamed_addr {
10682 ; VLX-LABEL: test_masked_vpcmpsgew_v32i1_v64i1_mask_mem:
10683 ; VLX: # %bb.0: # %entry
10684 ; VLX-NEXT: kmovd %edi, %k1
10685 ; VLX-NEXT: vpcmpnltw (%rsi), %zmm0, %k0 {%k1}
10686 ; VLX-NEXT: kmovq %k0, %rax
10687 ; VLX-NEXT: vzeroupper
10688 ; VLX-NEXT: retq
10689 ;
10690 ; NoVLX-LABEL: test_masked_vpcmpsgew_v32i1_v64i1_mask_mem:
10691 ; NoVLX: # %bb.0: # %entry
10692 ; NoVLX-NEXT: vmovdqa (%rsi), %ymm1
10693 ; NoVLX-NEXT: vpcmpgtw %ymm0, %ymm1, %ymm1
10694 ; NoVLX-NEXT: vpternlogq $15, %zmm1, %zmm1, %zmm1
10695 ; NoVLX-NEXT: vpmovsxwd %ymm1, %zmm1
10696 ; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
10697 ; NoVLX-NEXT: kmovw %k0, %eax
10698 ; NoVLX-NEXT: andl %edi, %eax
10699 ; NoVLX-NEXT: shrl $16, %edi
10700 ; NoVLX-NEXT: vextracti64x4 $1, %zmm0, %ymm0
10701 ; NoVLX-NEXT: vmovdqa 32(%rsi), %ymm1
10702 ; NoVLX-NEXT: vpcmpgtw %ymm0, %ymm1, %ymm0
10703 ; NoVLX-NEXT: vpternlogq $15, %zmm0, %zmm0, %zmm0
10704 ; NoVLX-NEXT: vpmovsxwd %ymm0, %zmm0
10705 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
10706 ; NoVLX-NEXT: kmovw %k0, %ecx
10707 ; NoVLX-NEXT: andl %edi, %ecx
10708 ; NoVLX-NEXT: shll $16, %ecx
10709 ; NoVLX-NEXT: movzwl %ax, %eax
10710 ; NoVLX-NEXT: orl %ecx, %eax
10711 ; NoVLX-NEXT: vzeroupper
10712 ; NoVLX-NEXT: retq
10713 entry:
10714 %0 = bitcast <8 x i64> %__a to <32 x i16>
10715 %load = load <8 x i64>, <8 x i64>* %__b
10716 %1 = bitcast <8 x i64> %load to <32 x i16>
10717 %2 = icmp sge <32 x i16> %0, %1
10718 %3 = bitcast i32 %__u to <32 x i1>
10719 %4 = and <32 x i1> %2, %3
10720 %5 = shufflevector <32 x i1> %4, <32 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63>
10721 %6 = bitcast <64 x i1> %5 to i64
10722 ret i64 %6
10723 }
10726 define zeroext i8 @test_vpcmpsged_v4i1_v8i1_mask(<2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
10727 ; VLX-LABEL: test_vpcmpsged_v4i1_v8i1_mask:
10728 ; VLX: # %bb.0: # %entry
10729 ; VLX-NEXT: vpcmpnltd %xmm1, %xmm0, %k0
10730 ; VLX-NEXT: kmovd %k0, %eax
10731 ; VLX-NEXT: # kill: def $al killed $al killed $eax
10732 ; VLX-NEXT: retq
10733 ;
10734 ; NoVLX-LABEL: test_vpcmpsged_v4i1_v8i1_mask:
10735 ; NoVLX: # %bb.0: # %entry
10736 ; NoVLX-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
10737 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
10738 ; NoVLX-NEXT: vpcmpnltd %zmm1, %zmm0, %k0
10739 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
10740 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
10741 ; NoVLX-NEXT: kmovw %k0, %eax
10742 ; NoVLX-NEXT: # kill: def $al killed $al killed $eax
10743 ; NoVLX-NEXT: vzeroupper
10744 ; NoVLX-NEXT: retq
10745 entry:
10746 %0 = bitcast <2 x i64> %__a to <4 x i32>
10747 %1 = bitcast <2 x i64> %__b to <4 x i32>
10748 %2 = icmp sge <4 x i32> %0, %1
10749 %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
10750 %4 = bitcast <8 x i1> %3 to i8
10751 ret i8 %4
10752 }
10754 define zeroext i8 @test_vpcmpsged_v4i1_v8i1_mask_mem(<2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
10755 ; VLX-LABEL: test_vpcmpsged_v4i1_v8i1_mask_mem:
10756 ; VLX: # %bb.0: # %entry
10757 ; VLX-NEXT: vpcmpnltd (%rdi), %xmm0, %k0
10758 ; VLX-NEXT: kmovd %k0, %eax
10759 ; VLX-NEXT: # kill: def $al killed $al killed $eax
10760 ; VLX-NEXT: retq
10761 ;
10762 ; NoVLX-LABEL: test_vpcmpsged_v4i1_v8i1_mask_mem:
10763 ; NoVLX: # %bb.0: # %entry
10764 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
10765 ; NoVLX-NEXT: vmovdqa (%rdi), %xmm1
10766 ; NoVLX-NEXT: vpcmpnltd %zmm1, %zmm0, %k0
10767 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
10768 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
10769 ; NoVLX-NEXT: kmovw %k0, %eax
10770 ; NoVLX-NEXT: # kill: def $al killed $al killed $eax
10771 ; NoVLX-NEXT: vzeroupper
10772 ; NoVLX-NEXT: retq
10773 entry:
10774 %0 = bitcast <2 x i64> %__a to <4 x i32>
10775 %load = load <2 x i64>, <2 x i64>* %__b
10776 %1 = bitcast <2 x i64> %load to <4 x i32>
10777 %2 = icmp sge <4 x i32> %0, %1
10778 %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
10779 %4 = bitcast <8 x i1> %3 to i8
10780 ret i8 %4
10781 }
10783 define zeroext i8 @test_masked_vpcmpsged_v4i1_v8i1_mask(i8 zeroext %__u, <2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
10784 ; VLX-LABEL: test_masked_vpcmpsged_v4i1_v8i1_mask:
10785 ; VLX: # %bb.0: # %entry
10786 ; VLX-NEXT: kmovd %edi, %k1
10787 ; VLX-NEXT: vpcmpnltd %xmm1, %xmm0, %k0 {%k1}
10788 ; VLX-NEXT: kmovd %k0, %eax
10789 ; VLX-NEXT: # kill: def $al killed $al killed $eax
10790 ; VLX-NEXT: retq
10791 ;
10792 ; NoVLX-LABEL: test_masked_vpcmpsged_v4i1_v8i1_mask:
10793 ; NoVLX: # %bb.0: # %entry
10794 ; NoVLX-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
10795 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
10796 ; NoVLX-NEXT: kmovw %edi, %k1
10797 ; NoVLX-NEXT: vpcmpnltd %zmm1, %zmm0, %k0 {%k1}
10798 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
10799 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
10800 ; NoVLX-NEXT: kmovw %k0, %eax
10801 ; NoVLX-NEXT: # kill: def $al killed $al killed $eax
10802 ; NoVLX-NEXT: vzeroupper
10803 ; NoVLX-NEXT: retq
10804 entry:
10805 %0 = bitcast <2 x i64> %__a to <4 x i32>
10806 %1 = bitcast <2 x i64> %__b to <4 x i32>
10807 %2 = icmp sge <4 x i32> %0, %1
10808 %3 = bitcast i8 %__u to <8 x i1>
10809 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
10810 %4 = and <4 x i1> %2, %extract.i
10811 %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
10812 %6 = bitcast <8 x i1> %5 to i8
10813 ret i8 %6
10814 }
10816 define zeroext i8 @test_masked_vpcmpsged_v4i1_v8i1_mask_mem(i8 zeroext %__u, <2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
10817 ; VLX-LABEL: test_masked_vpcmpsged_v4i1_v8i1_mask_mem:
10818 ; VLX: # %bb.0: # %entry
10819 ; VLX-NEXT: kmovd %edi, %k1
10820 ; VLX-NEXT: vpcmpnltd (%rsi), %xmm0, %k0 {%k1}
10821 ; VLX-NEXT: kmovd %k0, %eax
10822 ; VLX-NEXT: # kill: def $al killed $al killed $eax
10823 ; VLX-NEXT: retq
10824 ;
10825 ; NoVLX-LABEL: test_masked_vpcmpsged_v4i1_v8i1_mask_mem:
10826 ; NoVLX: # %bb.0: # %entry
10827 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
10828 ; NoVLX-NEXT: vmovdqa (%rsi), %xmm1
10829 ; NoVLX-NEXT: kmovw %edi, %k1
10830 ; NoVLX-NEXT: vpcmpnltd %zmm1, %zmm0, %k0 {%k1}
10831 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
10832 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
10833 ; NoVLX-NEXT: kmovw %k0, %eax
10834 ; NoVLX-NEXT: # kill: def $al killed $al killed $eax
10835 ; NoVLX-NEXT: vzeroupper
10836 ; NoVLX-NEXT: retq
10837 entry:
10838 %0 = bitcast <2 x i64> %__a to <4 x i32>
10839 %load = load <2 x i64>, <2 x i64>* %__b
10840 %1 = bitcast <2 x i64> %load to <4 x i32>
10841 %2 = icmp sge <4 x i32> %0, %1
10842 %3 = bitcast i8 %__u to <8 x i1>
10843 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
10844 %4 = and <4 x i1> %2, %extract.i
10845 %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
10846 %6 = bitcast <8 x i1> %5 to i8
10847 ret i8 %6
10848 }
10851 define zeroext i8 @test_vpcmpsged_v4i1_v8i1_mask_mem_b(<2 x i64> %__a, i32* %__b) local_unnamed_addr {
10852 ; VLX-LABEL: test_vpcmpsged_v4i1_v8i1_mask_mem_b:
10853 ; VLX: # %bb.0: # %entry
10854 ; VLX-NEXT: vpcmpnltd (%rdi){1to4}, %xmm0, %k0
10855 ; VLX-NEXT: kmovd %k0, %eax
10856 ; VLX-NEXT: # kill: def $al killed $al killed $eax
10857 ; VLX-NEXT: retq
10858 ;
10859 ; NoVLX-LABEL: test_vpcmpsged_v4i1_v8i1_mask_mem_b:
10860 ; NoVLX: # %bb.0: # %entry
10861 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
10862 ; NoVLX-NEXT: vpcmpnltd (%rdi){1to16}, %zmm0, %k0
10863 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
10864 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
10865 ; NoVLX-NEXT: kmovw %k0, %eax
10866 ; NoVLX-NEXT: # kill: def $al killed $al killed $eax
10867 ; NoVLX-NEXT: vzeroupper
10868 ; NoVLX-NEXT: retq
10869 entry:
10870 %0 = bitcast <2 x i64> %__a to <4 x i32>
10871 %load = load i32, i32* %__b
10872 %vec = insertelement <4 x i32> undef, i32 %load, i32 0
10873 %1 = shufflevector <4 x i32> %vec, <4 x i32> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
10874 %2 = icmp sge <4 x i32> %0, %1
10875 %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
10876 %4 = bitcast <8 x i1> %3 to i8
10877 ret i8 %4
10878 }
10880 define zeroext i8 @test_masked_vpcmpsged_v4i1_v8i1_mask_mem_b(i8 zeroext %__u, <2 x i64> %__a, i32* %__b) local_unnamed_addr {
10881 ; VLX-LABEL: test_masked_vpcmpsged_v4i1_v8i1_mask_mem_b:
10882 ; VLX: # %bb.0: # %entry
10883 ; VLX-NEXT: kmovd %edi, %k1
10884 ; VLX-NEXT: vpcmpnltd (%rsi){1to4}, %xmm0, %k0 {%k1}
10885 ; VLX-NEXT: kmovd %k0, %eax
10886 ; VLX-NEXT: # kill: def $al killed $al killed $eax
10887 ; VLX-NEXT: retq
10888 ;
10889 ; NoVLX-LABEL: test_masked_vpcmpsged_v4i1_v8i1_mask_mem_b:
10890 ; NoVLX: # %bb.0: # %entry
10891 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
10892 ; NoVLX-NEXT: kmovw %edi, %k1
10893 ; NoVLX-NEXT: vpcmpnltd (%rsi){1to16}, %zmm0, %k0 {%k1}
10894 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
10895 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
10896 ; NoVLX-NEXT: kmovw %k0, %eax
10897 ; NoVLX-NEXT: # kill: def $al killed $al killed $eax
10898 ; NoVLX-NEXT: vzeroupper
10899 ; NoVLX-NEXT: retq
10900 entry:
10901 %0 = bitcast <2 x i64> %__a to <4 x i32>
10902 %load = load i32, i32* %__b
10903 %vec = insertelement <4 x i32> undef, i32 %load, i32 0
10904 %1 = shufflevector <4 x i32> %vec, <4 x i32> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
10905 %2 = icmp sge <4 x i32> %0, %1
10906 %3 = bitcast i8 %__u to <8 x i1>
10907 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
10908 %4 = and <4 x i1> %extract.i, %2
10909 %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
10910 %6 = bitcast <8 x i1> %5 to i8
10911 ret i8 %6
10912 }
10915 define zeroext i16 @test_vpcmpsged_v4i1_v16i1_mask(<2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
10916 ; VLX-LABEL: test_vpcmpsged_v4i1_v16i1_mask:
10917 ; VLX: # %bb.0: # %entry
10918 ; VLX-NEXT: vpcmpnltd %xmm1, %xmm0, %k0
10919 ; VLX-NEXT: kmovd %k0, %eax
10920 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
10921 ; VLX-NEXT: retq
10922 ;
10923 ; NoVLX-LABEL: test_vpcmpsged_v4i1_v16i1_mask:
10924 ; NoVLX: # %bb.0: # %entry
10925 ; NoVLX-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
10926 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
10927 ; NoVLX-NEXT: vpcmpnltd %zmm1, %zmm0, %k0
10928 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
10929 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
10930 ; NoVLX-NEXT: kmovw %k0, %eax
10931 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
10932 ; NoVLX-NEXT: vzeroupper
10933 ; NoVLX-NEXT: retq
10934 entry:
10935 %0 = bitcast <2 x i64> %__a to <4 x i32>
10936 %1 = bitcast <2 x i64> %__b to <4 x i32>
10937 %2 = icmp sge <4 x i32> %0, %1
10938 %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
10939 %4 = bitcast <16 x i1> %3 to i16
10940 ret i16 %4
10941 }
10943 define zeroext i16 @test_vpcmpsged_v4i1_v16i1_mask_mem(<2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
10944 ; VLX-LABEL: test_vpcmpsged_v4i1_v16i1_mask_mem:
10945 ; VLX: # %bb.0: # %entry
10946 ; VLX-NEXT: vpcmpnltd (%rdi), %xmm0, %k0
10947 ; VLX-NEXT: kmovd %k0, %eax
10948 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
10949 ; VLX-NEXT: retq
10950 ;
10951 ; NoVLX-LABEL: test_vpcmpsged_v4i1_v16i1_mask_mem:
10952 ; NoVLX: # %bb.0: # %entry
10953 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
10954 ; NoVLX-NEXT: vmovdqa (%rdi), %xmm1
10955 ; NoVLX-NEXT: vpcmpnltd %zmm1, %zmm0, %k0
10956 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
10957 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
10958 ; NoVLX-NEXT: kmovw %k0, %eax
10959 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
10960 ; NoVLX-NEXT: vzeroupper
10961 ; NoVLX-NEXT: retq
10962 entry:
10963 %0 = bitcast <2 x i64> %__a to <4 x i32>
10964 %load = load <2 x i64>, <2 x i64>* %__b
10965 %1 = bitcast <2 x i64> %load to <4 x i32>
10966 %2 = icmp sge <4 x i32> %0, %1
10967 %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
10968 %4 = bitcast <16 x i1> %3 to i16
10969 ret i16 %4
10970 }
10972 define zeroext i16 @test_masked_vpcmpsged_v4i1_v16i1_mask(i8 zeroext %__u, <2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
10973 ; VLX-LABEL: test_masked_vpcmpsged_v4i1_v16i1_mask:
10974 ; VLX: # %bb.0: # %entry
10975 ; VLX-NEXT: kmovd %edi, %k1
10976 ; VLX-NEXT: vpcmpnltd %xmm1, %xmm0, %k0 {%k1}
10977 ; VLX-NEXT: kmovd %k0, %eax
10978 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
10979 ; VLX-NEXT: retq
10980 ;
10981 ; NoVLX-LABEL: test_masked_vpcmpsged_v4i1_v16i1_mask:
10982 ; NoVLX: # %bb.0: # %entry
10983 ; NoVLX-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
10984 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
10985 ; NoVLX-NEXT: kmovw %edi, %k1
10986 ; NoVLX-NEXT: vpcmpnltd %zmm1, %zmm0, %k0 {%k1}
10987 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
10988 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
10989 ; NoVLX-NEXT: kmovw %k0, %eax
10990 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
10991 ; NoVLX-NEXT: vzeroupper
10992 ; NoVLX-NEXT: retq
10993 entry:
10994 %0 = bitcast <2 x i64> %__a to <4 x i32>
10995 %1 = bitcast <2 x i64> %__b to <4 x i32>
10996 %2 = icmp sge <4 x i32> %0, %1
10997 %3 = bitcast i8 %__u to <8 x i1>
10998 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
10999 %4 = and <4 x i1> %2, %extract.i
11000 %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
11001 %6 = bitcast <16 x i1> %5 to i16
11002 ret i16 %6
11003 }
11005 define zeroext i16 @test_masked_vpcmpsged_v4i1_v16i1_mask_mem(i8 zeroext %__u, <2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
11006 ; VLX-LABEL: test_masked_vpcmpsged_v4i1_v16i1_mask_mem:
11007 ; VLX: # %bb.0: # %entry
11008 ; VLX-NEXT: kmovd %edi, %k1
11009 ; VLX-NEXT: vpcmpnltd (%rsi), %xmm0, %k0 {%k1}
11010 ; VLX-NEXT: kmovd %k0, %eax
11011 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
11012 ; VLX-NEXT: retq
11013 ;
11014 ; NoVLX-LABEL: test_masked_vpcmpsged_v4i1_v16i1_mask_mem:
11015 ; NoVLX: # %bb.0: # %entry
11016 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
11017 ; NoVLX-NEXT: vmovdqa (%rsi), %xmm1
11018 ; NoVLX-NEXT: kmovw %edi, %k1
11019 ; NoVLX-NEXT: vpcmpnltd %zmm1, %zmm0, %k0 {%k1}
11020 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
11021 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
11022 ; NoVLX-NEXT: kmovw %k0, %eax
11023 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
11024 ; NoVLX-NEXT: vzeroupper
11025 ; NoVLX-NEXT: retq
11026 entry:
11027 %0 = bitcast <2 x i64> %__a to <4 x i32>
11028 %load = load <2 x i64>, <2 x i64>* %__b
11029 %1 = bitcast <2 x i64> %load to <4 x i32>
11030 %2 = icmp sge <4 x i32> %0, %1
11031 %3 = bitcast i8 %__u to <8 x i1>
11032 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
11033 %4 = and <4 x i1> %2, %extract.i
11034 %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
11035 %6 = bitcast <16 x i1> %5 to i16
11036 ret i16 %6
11037 }
11040 define zeroext i16 @test_vpcmpsged_v4i1_v16i1_mask_mem_b(<2 x i64> %__a, i32* %__b) local_unnamed_addr {
11041 ; VLX-LABEL: test_vpcmpsged_v4i1_v16i1_mask_mem_b:
11042 ; VLX: # %bb.0: # %entry
11043 ; VLX-NEXT: vpcmpnltd (%rdi){1to4}, %xmm0, %k0
11044 ; VLX-NEXT: kmovd %k0, %eax
11045 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
11046 ; VLX-NEXT: retq
11047 ;
11048 ; NoVLX-LABEL: test_vpcmpsged_v4i1_v16i1_mask_mem_b:
11049 ; NoVLX: # %bb.0: # %entry
11050 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
11051 ; NoVLX-NEXT: vpcmpnltd (%rdi){1to16}, %zmm0, %k0
11052 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
11053 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
11054 ; NoVLX-NEXT: kmovw %k0, %eax
11055 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
11056 ; NoVLX-NEXT: vzeroupper
11057 ; NoVLX-NEXT: retq
11058 entry:
11059 %0 = bitcast <2 x i64> %__a to <4 x i32>
11060 %load = load i32, i32* %__b
11061 %vec = insertelement <4 x i32> undef, i32 %load, i32 0
11062 %1 = shufflevector <4 x i32> %vec, <4 x i32> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
11063 %2 = icmp sge <4 x i32> %0, %1
11064 %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
11065 %4 = bitcast <16 x i1> %3 to i16
11066 ret i16 %4
11067 }
11069 define zeroext i16 @test_masked_vpcmpsged_v4i1_v16i1_mask_mem_b(i8 zeroext %__u, <2 x i64> %__a, i32* %__b) local_unnamed_addr {
11070 ; VLX-LABEL: test_masked_vpcmpsged_v4i1_v16i1_mask_mem_b:
11071 ; VLX: # %bb.0: # %entry
11072 ; VLX-NEXT: kmovd %edi, %k1
11073 ; VLX-NEXT: vpcmpnltd (%rsi){1to4}, %xmm0, %k0 {%k1}
11074 ; VLX-NEXT: kmovd %k0, %eax
11075 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
11076 ; VLX-NEXT: retq
11077 ;
11078 ; NoVLX-LABEL: test_masked_vpcmpsged_v4i1_v16i1_mask_mem_b:
11079 ; NoVLX: # %bb.0: # %entry
11080 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
11081 ; NoVLX-NEXT: kmovw %edi, %k1
11082 ; NoVLX-NEXT: vpcmpnltd (%rsi){1to16}, %zmm0, %k0 {%k1}
11083 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
11084 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
11085 ; NoVLX-NEXT: kmovw %k0, %eax
11086 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
11087 ; NoVLX-NEXT: vzeroupper
11088 ; NoVLX-NEXT: retq
11089 entry:
11090 %0 = bitcast <2 x i64> %__a to <4 x i32>
11091 %load = load i32, i32* %__b
11092 %vec = insertelement <4 x i32> undef, i32 %load, i32 0
11093 %1 = shufflevector <4 x i32> %vec, <4 x i32> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
11094 %2 = icmp sge <4 x i32> %0, %1
11095 %3 = bitcast i8 %__u to <8 x i1>
11096 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
11097 %4 = and <4 x i1> %extract.i, %2
11098 %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
11099 %6 = bitcast <16 x i1> %5 to i16
11100 ret i16 %6
11101 }
11104 define zeroext i32 @test_vpcmpsged_v4i1_v32i1_mask(<2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
11105 ; VLX-LABEL: test_vpcmpsged_v4i1_v32i1_mask:
11106 ; VLX: # %bb.0: # %entry
11107 ; VLX-NEXT: vpcmpnltd %xmm1, %xmm0, %k0
11108 ; VLX-NEXT: kmovd %k0, %eax
11109 ; VLX-NEXT: retq
11110 ;
11111 ; NoVLX-LABEL: test_vpcmpsged_v4i1_v32i1_mask:
11112 ; NoVLX: # %bb.0: # %entry
11113 ; NoVLX-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
11114 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
11115 ; NoVLX-NEXT: vpcmpnltd %zmm1, %zmm0, %k0
11116 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
11117 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
11118 ; NoVLX-NEXT: kmovw %k0, %eax
11119 ; NoVLX-NEXT: vzeroupper
11120 ; NoVLX-NEXT: retq
11121 entry:
11122 %0 = bitcast <2 x i64> %__a to <4 x i32>
11123 %1 = bitcast <2 x i64> %__b to <4 x i32>
11124 %2 = icmp sge <4 x i32> %0, %1
11125 %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
11126 %4 = bitcast <32 x i1> %3 to i32
11127 ret i32 %4
11128 }
11130 define zeroext i32 @test_vpcmpsged_v4i1_v32i1_mask_mem(<2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
11131 ; VLX-LABEL: test_vpcmpsged_v4i1_v32i1_mask_mem:
11132 ; VLX: # %bb.0: # %entry
11133 ; VLX-NEXT: vpcmpnltd (%rdi), %xmm0, %k0
11134 ; VLX-NEXT: kmovd %k0, %eax
11135 ; VLX-NEXT: retq
11136 ;
11137 ; NoVLX-LABEL: test_vpcmpsged_v4i1_v32i1_mask_mem:
11138 ; NoVLX: # %bb.0: # %entry
11139 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
11140 ; NoVLX-NEXT: vmovdqa (%rdi), %xmm1
11141 ; NoVLX-NEXT: vpcmpnltd %zmm1, %zmm0, %k0
11142 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
11143 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
11144 ; NoVLX-NEXT: kmovw %k0, %eax
11145 ; NoVLX-NEXT: vzeroupper
11146 ; NoVLX-NEXT: retq
11147 entry:
11148 %0 = bitcast <2 x i64> %__a to <4 x i32>
11149 %load = load <2 x i64>, <2 x i64>* %__b
11150 %1 = bitcast <2 x i64> %load to <4 x i32>
11151 %2 = icmp sge <4 x i32> %0, %1
11152 %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
11153 %4 = bitcast <32 x i1> %3 to i32
11154 ret i32 %4
11155 }
11157 define zeroext i32 @test_masked_vpcmpsged_v4i1_v32i1_mask(i8 zeroext %__u, <2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
11158 ; VLX-LABEL: test_masked_vpcmpsged_v4i1_v32i1_mask:
11159 ; VLX: # %bb.0: # %entry
11160 ; VLX-NEXT: kmovd %edi, %k1
11161 ; VLX-NEXT: vpcmpnltd %xmm1, %xmm0, %k0 {%k1}
11162 ; VLX-NEXT: kmovd %k0, %eax
11163 ; VLX-NEXT: retq
11164 ;
11165 ; NoVLX-LABEL: test_masked_vpcmpsged_v4i1_v32i1_mask:
11166 ; NoVLX: # %bb.0: # %entry
11167 ; NoVLX-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
11168 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
11169 ; NoVLX-NEXT: kmovw %edi, %k1
11170 ; NoVLX-NEXT: vpcmpnltd %zmm1, %zmm0, %k0 {%k1}
11171 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
11172 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
11173 ; NoVLX-NEXT: kmovw %k0, %eax
11174 ; NoVLX-NEXT: vzeroupper
11175 ; NoVLX-NEXT: retq
11176 entry:
11177 %0 = bitcast <2 x i64> %__a to <4 x i32>
11178 %1 = bitcast <2 x i64> %__b to <4 x i32>
11179 %2 = icmp sge <4 x i32> %0, %1
11180 %3 = bitcast i8 %__u to <8 x i1>
11181 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
11182 %4 = and <4 x i1> %2, %extract.i
11183 %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
11184 %6 = bitcast <32 x i1> %5 to i32
11185 ret i32 %6
11186 }
11188 define zeroext i32 @test_masked_vpcmpsged_v4i1_v32i1_mask_mem(i8 zeroext %__u, <2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
11189 ; VLX-LABEL: test_masked_vpcmpsged_v4i1_v32i1_mask_mem:
11190 ; VLX: # %bb.0: # %entry
11191 ; VLX-NEXT: kmovd %edi, %k1
11192 ; VLX-NEXT: vpcmpnltd (%rsi), %xmm0, %k0 {%k1}
11193 ; VLX-NEXT: kmovd %k0, %eax
11194 ; VLX-NEXT: retq
11195 ;
11196 ; NoVLX-LABEL: test_masked_vpcmpsged_v4i1_v32i1_mask_mem:
11197 ; NoVLX: # %bb.0: # %entry
11198 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
11199 ; NoVLX-NEXT: vmovdqa (%rsi), %xmm1
11200 ; NoVLX-NEXT: kmovw %edi, %k1
11201 ; NoVLX-NEXT: vpcmpnltd %zmm1, %zmm0, %k0 {%k1}
11202 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
11203 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
11204 ; NoVLX-NEXT: kmovw %k0, %eax
11205 ; NoVLX-NEXT: vzeroupper
11206 ; NoVLX-NEXT: retq
11207 entry:
11208 %0 = bitcast <2 x i64> %__a to <4 x i32>
11209 %load = load <2 x i64>, <2 x i64>* %__b
11210 %1 = bitcast <2 x i64> %load to <4 x i32>
11211 %2 = icmp sge <4 x i32> %0, %1
11212 %3 = bitcast i8 %__u to <8 x i1>
11213 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
11214 %4 = and <4 x i1> %2, %extract.i
11215 %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
11216 %6 = bitcast <32 x i1> %5 to i32
11217 ret i32 %6
11218 }
11221 define zeroext i32 @test_vpcmpsged_v4i1_v32i1_mask_mem_b(<2 x i64> %__a, i32* %__b) local_unnamed_addr {
11222 ; VLX-LABEL: test_vpcmpsged_v4i1_v32i1_mask_mem_b:
11223 ; VLX: # %bb.0: # %entry
11224 ; VLX-NEXT: vpcmpnltd (%rdi){1to4}, %xmm0, %k0
11225 ; VLX-NEXT: kmovd %k0, %eax
11226 ; VLX-NEXT: retq
11227 ;
11228 ; NoVLX-LABEL: test_vpcmpsged_v4i1_v32i1_mask_mem_b:
11229 ; NoVLX: # %bb.0: # %entry
11230 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
11231 ; NoVLX-NEXT: vpcmpnltd (%rdi){1to16}, %zmm0, %k0
11232 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
11233 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
11234 ; NoVLX-NEXT: kmovw %k0, %eax
11235 ; NoVLX-NEXT: vzeroupper
11236 ; NoVLX-NEXT: retq
11237 entry:
11238 %0 = bitcast <2 x i64> %__a to <4 x i32>
11239 %load = load i32, i32* %__b
11240 %vec = insertelement <4 x i32> undef, i32 %load, i32 0
11241 %1 = shufflevector <4 x i32> %vec, <4 x i32> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
11242 %2 = icmp sge <4 x i32> %0, %1
11243 %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
11244 %4 = bitcast <32 x i1> %3 to i32
11245 ret i32 %4
11246 }
11248 define zeroext i32 @test_masked_vpcmpsged_v4i1_v32i1_mask_mem_b(i8 zeroext %__u, <2 x i64> %__a, i32* %__b) local_unnamed_addr {
11249 ; VLX-LABEL: test_masked_vpcmpsged_v4i1_v32i1_mask_mem_b:
11250 ; VLX: # %bb.0: # %entry
11251 ; VLX-NEXT: kmovd %edi, %k1
11252 ; VLX-NEXT: vpcmpnltd (%rsi){1to4}, %xmm0, %k0 {%k1}
11253 ; VLX-NEXT: kmovd %k0, %eax
11254 ; VLX-NEXT: retq
11255 ;
11256 ; NoVLX-LABEL: test_masked_vpcmpsged_v4i1_v32i1_mask_mem_b:
11257 ; NoVLX: # %bb.0: # %entry
11258 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
11259 ; NoVLX-NEXT: kmovw %edi, %k1
11260 ; NoVLX-NEXT: vpcmpnltd (%rsi){1to16}, %zmm0, %k0 {%k1}
11261 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
11262 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
11263 ; NoVLX-NEXT: kmovw %k0, %eax
11264 ; NoVLX-NEXT: vzeroupper
11265 ; NoVLX-NEXT: retq
11266 entry:
11267 %0 = bitcast <2 x i64> %__a to <4 x i32>
11268 %load = load i32, i32* %__b
11269 %vec = insertelement <4 x i32> undef, i32 %load, i32 0
11270 %1 = shufflevector <4 x i32> %vec, <4 x i32> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
11271 %2 = icmp sge <4 x i32> %0, %1
11272 %3 = bitcast i8 %__u to <8 x i1>
11273 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
11274 %4 = and <4 x i1> %extract.i, %2
11275 %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
11276 %6 = bitcast <32 x i1> %5 to i32
11277 ret i32 %6
11278 }
11281 define zeroext i64 @test_vpcmpsged_v4i1_v64i1_mask(<2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
11282 ; VLX-LABEL: test_vpcmpsged_v4i1_v64i1_mask:
11283 ; VLX: # %bb.0: # %entry
11284 ; VLX-NEXT: vpcmpnltd %xmm1, %xmm0, %k0
11285 ; VLX-NEXT: kmovq %k0, %rax
11286 ; VLX-NEXT: retq
11287 ;
11288 ; NoVLX-LABEL: test_vpcmpsged_v4i1_v64i1_mask:
11289 ; NoVLX: # %bb.0: # %entry
11290 ; NoVLX-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
11291 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
11292 ; NoVLX-NEXT: vpcmpnltd %zmm1, %zmm0, %k0
11293 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
11294 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
11295 ; NoVLX-NEXT: kmovw %k0, %eax
11296 ; NoVLX-NEXT: vzeroupper
11297 ; NoVLX-NEXT: retq
11298 entry:
11299 %0 = bitcast <2 x i64> %__a to <4 x i32>
11300 %1 = bitcast <2 x i64> %__b to <4 x i32>
11301 %2 = icmp sge <4 x i32> %0, %1
11302 %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
11303 %4 = bitcast <64 x i1> %3 to i64
11304 ret i64 %4
11305 }
11307 define zeroext i64 @test_vpcmpsged_v4i1_v64i1_mask_mem(<2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
11308 ; VLX-LABEL: test_vpcmpsged_v4i1_v64i1_mask_mem:
11309 ; VLX: # %bb.0: # %entry
11310 ; VLX-NEXT: vpcmpnltd (%rdi), %xmm0, %k0
11311 ; VLX-NEXT: kmovq %k0, %rax
11312 ; VLX-NEXT: retq
11313 ;
11314 ; NoVLX-LABEL: test_vpcmpsged_v4i1_v64i1_mask_mem:
11315 ; NoVLX: # %bb.0: # %entry
11316 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
11317 ; NoVLX-NEXT: vmovdqa (%rdi), %xmm1
11318 ; NoVLX-NEXT: vpcmpnltd %zmm1, %zmm0, %k0
11319 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
11320 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
11321 ; NoVLX-NEXT: kmovw %k0, %eax
11322 ; NoVLX-NEXT: vzeroupper
11323 ; NoVLX-NEXT: retq
11324 entry:
11325 %0 = bitcast <2 x i64> %__a to <4 x i32>
11326 %load = load <2 x i64>, <2 x i64>* %__b
11327 %1 = bitcast <2 x i64> %load to <4 x i32>
11328 %2 = icmp sge <4 x i32> %0, %1
11329 %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
11330 %4 = bitcast <64 x i1> %3 to i64
11331 ret i64 %4
11332 }
11334 define zeroext i64 @test_masked_vpcmpsged_v4i1_v64i1_mask(i8 zeroext %__u, <2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
11335 ; VLX-LABEL: test_masked_vpcmpsged_v4i1_v64i1_mask:
11336 ; VLX: # %bb.0: # %entry
11337 ; VLX-NEXT: kmovd %edi, %k1
11338 ; VLX-NEXT: vpcmpnltd %xmm1, %xmm0, %k0 {%k1}
11339 ; VLX-NEXT: kmovq %k0, %rax
11340 ; VLX-NEXT: retq
11341 ;
11342 ; NoVLX-LABEL: test_masked_vpcmpsged_v4i1_v64i1_mask:
11343 ; NoVLX: # %bb.0: # %entry
11344 ; NoVLX-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
11345 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
11346 ; NoVLX-NEXT: kmovw %edi, %k1
11347 ; NoVLX-NEXT: vpcmpnltd %zmm1, %zmm0, %k0 {%k1}
11348 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
11349 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
11350 ; NoVLX-NEXT: kmovw %k0, %eax
11351 ; NoVLX-NEXT: vzeroupper
11352 ; NoVLX-NEXT: retq
11353 entry:
11354 %0 = bitcast <2 x i64> %__a to <4 x i32>
11355 %1 = bitcast <2 x i64> %__b to <4 x i32>
11356 %2 = icmp sge <4 x i32> %0, %1
11357 %3 = bitcast i8 %__u to <8 x i1>
11358 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
11359 %4 = and <4 x i1> %2, %extract.i
11360 %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
11361 %6 = bitcast <64 x i1> %5 to i64
11362 ret i64 %6
11363 }
11365 define zeroext i64 @test_masked_vpcmpsged_v4i1_v64i1_mask_mem(i8 zeroext %__u, <2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
11366 ; VLX-LABEL: test_masked_vpcmpsged_v4i1_v64i1_mask_mem:
11367 ; VLX: # %bb.0: # %entry
11368 ; VLX-NEXT: kmovd %edi, %k1
11369 ; VLX-NEXT: vpcmpnltd (%rsi), %xmm0, %k0 {%k1}
11370 ; VLX-NEXT: kmovq %k0, %rax
11371 ; VLX-NEXT: retq
11372 ;
11373 ; NoVLX-LABEL: test_masked_vpcmpsged_v4i1_v64i1_mask_mem:
11374 ; NoVLX: # %bb.0: # %entry
11375 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
11376 ; NoVLX-NEXT: vmovdqa (%rsi), %xmm1
11377 ; NoVLX-NEXT: kmovw %edi, %k1
11378 ; NoVLX-NEXT: vpcmpnltd %zmm1, %zmm0, %k0 {%k1}
11379 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
11380 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
11381 ; NoVLX-NEXT: kmovw %k0, %eax
11382 ; NoVLX-NEXT: vzeroupper
11383 ; NoVLX-NEXT: retq
11384 entry:
11385 %0 = bitcast <2 x i64> %__a to <4 x i32>
11386 %load = load <2 x i64>, <2 x i64>* %__b
11387 %1 = bitcast <2 x i64> %load to <4 x i32>
11388 %2 = icmp sge <4 x i32> %0, %1
11389 %3 = bitcast i8 %__u to <8 x i1>
11390 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
11391 %4 = and <4 x i1> %2, %extract.i
11392 %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
11393 %6 = bitcast <64 x i1> %5 to i64
11394 ret i64 %6
11395 }
11398 define zeroext i64 @test_vpcmpsged_v4i1_v64i1_mask_mem_b(<2 x i64> %__a, i32* %__b) local_unnamed_addr {
11399 ; VLX-LABEL: test_vpcmpsged_v4i1_v64i1_mask_mem_b:
11400 ; VLX: # %bb.0: # %entry
11401 ; VLX-NEXT: vpcmpnltd (%rdi){1to4}, %xmm0, %k0
11402 ; VLX-NEXT: kmovq %k0, %rax
11403 ; VLX-NEXT: retq
11404 ;
11405 ; NoVLX-LABEL: test_vpcmpsged_v4i1_v64i1_mask_mem_b:
11406 ; NoVLX: # %bb.0: # %entry
11407 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
11408 ; NoVLX-NEXT: vpcmpnltd (%rdi){1to16}, %zmm0, %k0
11409 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
11410 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
11411 ; NoVLX-NEXT: kmovw %k0, %eax
11412 ; NoVLX-NEXT: vzeroupper
11413 ; NoVLX-NEXT: retq
11414 entry:
11415 %0 = bitcast <2 x i64> %__a to <4 x i32>
11416 %load = load i32, i32* %__b
11417 %vec = insertelement <4 x i32> undef, i32 %load, i32 0
11418 %1 = shufflevector <4 x i32> %vec, <4 x i32> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
11419 %2 = icmp sge <4 x i32> %0, %1
11420 %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
11421 %4 = bitcast <64 x i1> %3 to i64
11422 ret i64 %4
11423 }
11425 define zeroext i64 @test_masked_vpcmpsged_v4i1_v64i1_mask_mem_b(i8 zeroext %__u, <2 x i64> %__a, i32* %__b) local_unnamed_addr {
11426 ; VLX-LABEL: test_masked_vpcmpsged_v4i1_v64i1_mask_mem_b:
11427 ; VLX: # %bb.0: # %entry
11428 ; VLX-NEXT: kmovd %edi, %k1
11429 ; VLX-NEXT: vpcmpnltd (%rsi){1to4}, %xmm0, %k0 {%k1}
11430 ; VLX-NEXT: kmovq %k0, %rax
11431 ; VLX-NEXT: retq
11432 ;
11433 ; NoVLX-LABEL: test_masked_vpcmpsged_v4i1_v64i1_mask_mem_b:
11434 ; NoVLX: # %bb.0: # %entry
11435 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
11436 ; NoVLX-NEXT: kmovw %edi, %k1
11437 ; NoVLX-NEXT: vpcmpnltd (%rsi){1to16}, %zmm0, %k0 {%k1}
11438 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
11439 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
11440 ; NoVLX-NEXT: kmovw %k0, %eax
11441 ; NoVLX-NEXT: vzeroupper
11442 ; NoVLX-NEXT: retq
11443 entry:
11444 %0 = bitcast <2 x i64> %__a to <4 x i32>
11445 %load = load i32, i32* %__b
11446 %vec = insertelement <4 x i32> undef, i32 %load, i32 0
11447 %1 = shufflevector <4 x i32> %vec, <4 x i32> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
11448 %2 = icmp sge <4 x i32> %0, %1
11449 %3 = bitcast i8 %__u to <8 x i1>
11450 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
11451 %4 = and <4 x i1> %extract.i, %2
11452 %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
11453 %6 = bitcast <64 x i1> %5 to i64
11454 ret i64 %6
11455 }
11458 define zeroext i16 @test_vpcmpsged_v8i1_v16i1_mask(<4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
11459 ; VLX-LABEL: test_vpcmpsged_v8i1_v16i1_mask:
11460 ; VLX: # %bb.0: # %entry
11461 ; VLX-NEXT: vpcmpnltd %ymm1, %ymm0, %k0
11462 ; VLX-NEXT: kmovd %k0, %eax
11463 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
11464 ; VLX-NEXT: vzeroupper
11465 ; VLX-NEXT: retq
11466 ;
11467 ; NoVLX-LABEL: test_vpcmpsged_v8i1_v16i1_mask:
11468 ; NoVLX: # %bb.0: # %entry
11469 ; NoVLX-NEXT: # kill: def $ymm1 killed $ymm1 def $zmm1
11470 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
11471 ; NoVLX-NEXT: vpcmpnltd %zmm1, %zmm0, %k0
11472 ; NoVLX-NEXT: kshiftlw $8, %k0, %k0
11473 ; NoVLX-NEXT: kshiftrw $8, %k0, %k0
11474 ; NoVLX-NEXT: kmovw %k0, %eax
11475 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
11476 ; NoVLX-NEXT: vzeroupper
11477 ; NoVLX-NEXT: retq
11478 entry:
11479 %0 = bitcast <4 x i64> %__a to <8 x i32>
11480 %1 = bitcast <4 x i64> %__b to <8 x i32>
11481 %2 = icmp sge <8 x i32> %0, %1
11482 %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
11483 %4 = bitcast <16 x i1> %3 to i16
11484 ret i16 %4
11485 }
11487 define zeroext i16 @test_vpcmpsged_v8i1_v16i1_mask_mem(<4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
11488 ; VLX-LABEL: test_vpcmpsged_v8i1_v16i1_mask_mem:
11489 ; VLX: # %bb.0: # %entry
11490 ; VLX-NEXT: vpcmpnltd (%rdi), %ymm0, %k0
11491 ; VLX-NEXT: kmovd %k0, %eax
11492 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
11493 ; VLX-NEXT: vzeroupper
11494 ; VLX-NEXT: retq
11495 ;
11496 ; NoVLX-LABEL: test_vpcmpsged_v8i1_v16i1_mask_mem:
11497 ; NoVLX: # %bb.0: # %entry
11498 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
11499 ; NoVLX-NEXT: vmovdqa (%rdi), %ymm1
11500 ; NoVLX-NEXT: vpcmpnltd %zmm1, %zmm0, %k0
11501 ; NoVLX-NEXT: kshiftlw $8, %k0, %k0
11502 ; NoVLX-NEXT: kshiftrw $8, %k0, %k0
11503 ; NoVLX-NEXT: kmovw %k0, %eax
11504 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
11505 ; NoVLX-NEXT: vzeroupper
11506 ; NoVLX-NEXT: retq
11507 entry:
11508 %0 = bitcast <4 x i64> %__a to <8 x i32>
11509 %load = load <4 x i64>, <4 x i64>* %__b
11510 %1 = bitcast <4 x i64> %load to <8 x i32>
11511 %2 = icmp sge <8 x i32> %0, %1
11512 %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
11513 %4 = bitcast <16 x i1> %3 to i16
11514 ret i16 %4
11515 }
11517 define zeroext i16 @test_masked_vpcmpsged_v8i1_v16i1_mask(i8 zeroext %__u, <4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
11518 ; VLX-LABEL: test_masked_vpcmpsged_v8i1_v16i1_mask:
11519 ; VLX: # %bb.0: # %entry
11520 ; VLX-NEXT: kmovd %edi, %k1
11521 ; VLX-NEXT: vpcmpnltd %ymm1, %ymm0, %k0 {%k1}
11522 ; VLX-NEXT: kmovd %k0, %eax
11523 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
11524 ; VLX-NEXT: vzeroupper
11525 ; VLX-NEXT: retq
11526 ;
11527 ; NoVLX-LABEL: test_masked_vpcmpsged_v8i1_v16i1_mask:
11528 ; NoVLX: # %bb.0: # %entry
11529 ; NoVLX-NEXT: # kill: def $ymm1 killed $ymm1 def $zmm1
11530 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
11531 ; NoVLX-NEXT: kmovw %edi, %k1
11532 ; NoVLX-NEXT: vpcmpnltd %zmm1, %zmm0, %k0 {%k1}
11533 ; NoVLX-NEXT: kshiftlw $8, %k0, %k0
11534 ; NoVLX-NEXT: kshiftrw $8, %k0, %k0
11535 ; NoVLX-NEXT: kmovw %k0, %eax
11536 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
11537 ; NoVLX-NEXT: vzeroupper
11538 ; NoVLX-NEXT: retq
11539 entry:
11540 %0 = bitcast <4 x i64> %__a to <8 x i32>
11541 %1 = bitcast <4 x i64> %__b to <8 x i32>
11542 %2 = icmp sge <8 x i32> %0, %1
11543 %3 = bitcast i8 %__u to <8 x i1>
11544 %4 = and <8 x i1> %2, %3
11545 %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
11546 %6 = bitcast <16 x i1> %5 to i16
11547 ret i16 %6
11548 }
11550 define zeroext i16 @test_masked_vpcmpsged_v8i1_v16i1_mask_mem(i8 zeroext %__u, <4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
11551 ; VLX-LABEL: test_masked_vpcmpsged_v8i1_v16i1_mask_mem:
11552 ; VLX: # %bb.0: # %entry
11553 ; VLX-NEXT: kmovd %edi, %k1
11554 ; VLX-NEXT: vpcmpnltd (%rsi), %ymm0, %k0 {%k1}
11555 ; VLX-NEXT: kmovd %k0, %eax
11556 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
11557 ; VLX-NEXT: vzeroupper
11558 ; VLX-NEXT: retq
11559 ;
11560 ; NoVLX-LABEL: test_masked_vpcmpsged_v8i1_v16i1_mask_mem:
11561 ; NoVLX: # %bb.0: # %entry
11562 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
11563 ; NoVLX-NEXT: vmovdqa (%rsi), %ymm1
11564 ; NoVLX-NEXT: kmovw %edi, %k1
11565 ; NoVLX-NEXT: vpcmpnltd %zmm1, %zmm0, %k0 {%k1}
11566 ; NoVLX-NEXT: kshiftlw $8, %k0, %k0
11567 ; NoVLX-NEXT: kshiftrw $8, %k0, %k0
11568 ; NoVLX-NEXT: kmovw %k0, %eax
11569 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
11570 ; NoVLX-NEXT: vzeroupper
11571 ; NoVLX-NEXT: retq
11572 entry:
11573 %0 = bitcast <4 x i64> %__a to <8 x i32>
11574 %load = load <4 x i64>, <4 x i64>* %__b
11575 %1 = bitcast <4 x i64> %load to <8 x i32>
11576 %2 = icmp sge <8 x i32> %0, %1
11577 %3 = bitcast i8 %__u to <8 x i1>
11578 %4 = and <8 x i1> %2, %3
11579 %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
11580 %6 = bitcast <16 x i1> %5 to i16
11581 ret i16 %6
11582 }
11585 define zeroext i16 @test_vpcmpsged_v8i1_v16i1_mask_mem_b(<4 x i64> %__a, i32* %__b) local_unnamed_addr {
11586 ; VLX-LABEL: test_vpcmpsged_v8i1_v16i1_mask_mem_b:
11587 ; VLX: # %bb.0: # %entry
11588 ; VLX-NEXT: vpcmpnltd (%rdi){1to8}, %ymm0, %k0
11589 ; VLX-NEXT: kmovd %k0, %eax
11590 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
11591 ; VLX-NEXT: vzeroupper
11592 ; VLX-NEXT: retq
11593 ;
11594 ; NoVLX-LABEL: test_vpcmpsged_v8i1_v16i1_mask_mem_b:
11595 ; NoVLX: # %bb.0: # %entry
11596 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
11597 ; NoVLX-NEXT: vpcmpnltd (%rdi){1to16}, %zmm0, %k0
11598 ; NoVLX-NEXT: kshiftlw $8, %k0, %k0
11599 ; NoVLX-NEXT: kshiftrw $8, %k0, %k0
11600 ; NoVLX-NEXT: kmovw %k0, %eax
11601 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
11602 ; NoVLX-NEXT: vzeroupper
11603 ; NoVLX-NEXT: retq
11604 entry:
11605 %0 = bitcast <4 x i64> %__a to <8 x i32>
11606 %load = load i32, i32* %__b
11607 %vec = insertelement <8 x i32> undef, i32 %load, i32 0
11608 %1 = shufflevector <8 x i32> %vec, <8 x i32> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
11609 %2 = icmp sge <8 x i32> %0, %1
11610 %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
11611 %4 = bitcast <16 x i1> %3 to i16
11612 ret i16 %4
11613 }
11615 define zeroext i16 @test_masked_vpcmpsged_v8i1_v16i1_mask_mem_b(i8 zeroext %__u, <4 x i64> %__a, i32* %__b) local_unnamed_addr {
11616 ; VLX-LABEL: test_masked_vpcmpsged_v8i1_v16i1_mask_mem_b:
11617 ; VLX: # %bb.0: # %entry
11618 ; VLX-NEXT: kmovd %edi, %k1
11619 ; VLX-NEXT: vpcmpnltd (%rsi){1to8}, %ymm0, %k0 {%k1}
11620 ; VLX-NEXT: kmovd %k0, %eax
11621 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
11622 ; VLX-NEXT: vzeroupper
11623 ; VLX-NEXT: retq
11624 ;
11625 ; NoVLX-LABEL: test_masked_vpcmpsged_v8i1_v16i1_mask_mem_b:
11626 ; NoVLX: # %bb.0: # %entry
11627 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
11628 ; NoVLX-NEXT: kmovw %edi, %k1
11629 ; NoVLX-NEXT: vpcmpnltd (%rsi){1to16}, %zmm0, %k0 {%k1}
11630 ; NoVLX-NEXT: kshiftlw $8, %k0, %k0
11631 ; NoVLX-NEXT: kshiftrw $8, %k0, %k0
11632 ; NoVLX-NEXT: kmovw %k0, %eax
11633 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
11634 ; NoVLX-NEXT: vzeroupper
11635 ; NoVLX-NEXT: retq
11636 entry:
11637 %0 = bitcast <4 x i64> %__a to <8 x i32>
11638 %load = load i32, i32* %__b
11639 %vec = insertelement <8 x i32> undef, i32 %load, i32 0
11640 %1 = shufflevector <8 x i32> %vec, <8 x i32> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
11641 %2 = icmp sge <8 x i32> %0, %1
11642 %3 = bitcast i8 %__u to <8 x i1>
11643 %4 = and <8 x i1> %3, %2
11644 %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
11645 %6 = bitcast <16 x i1> %5 to i16
11646 ret i16 %6
11647 }
11650 define zeroext i32 @test_vpcmpsged_v8i1_v32i1_mask(<4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
11651 ; VLX-LABEL: test_vpcmpsged_v8i1_v32i1_mask:
11652 ; VLX: # %bb.0: # %entry
11653 ; VLX-NEXT: vpcmpnltd %ymm1, %ymm0, %k0
11654 ; VLX-NEXT: kmovd %k0, %eax
11655 ; VLX-NEXT: vzeroupper
11656 ; VLX-NEXT: retq
11657 ;
11658 ; NoVLX-LABEL: test_vpcmpsged_v8i1_v32i1_mask:
11659 ; NoVLX: # %bb.0: # %entry
11660 ; NoVLX-NEXT: # kill: def $ymm1 killed $ymm1 def $zmm1
11661 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
11662 ; NoVLX-NEXT: vpcmpnltd %zmm1, %zmm0, %k0
11663 ; NoVLX-NEXT: kshiftlw $8, %k0, %k0
11664 ; NoVLX-NEXT: kshiftrw $8, %k0, %k0
11665 ; NoVLX-NEXT: kmovw %k0, %eax
11666 ; NoVLX-NEXT: vzeroupper
11667 ; NoVLX-NEXT: retq
11668 entry:
11669 %0 = bitcast <4 x i64> %__a to <8 x i32>
11670 %1 = bitcast <4 x i64> %__b to <8 x i32>
11671 %2 = icmp sge <8 x i32> %0, %1
11672 %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
11673 %4 = bitcast <32 x i1> %3 to i32
11674 ret i32 %4
11675 }
11677 define zeroext i32 @test_vpcmpsged_v8i1_v32i1_mask_mem(<4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
11678 ; VLX-LABEL: test_vpcmpsged_v8i1_v32i1_mask_mem:
11679 ; VLX: # %bb.0: # %entry
11680 ; VLX-NEXT: vpcmpnltd (%rdi), %ymm0, %k0
11681 ; VLX-NEXT: kmovd %k0, %eax
11682 ; VLX-NEXT: vzeroupper
11683 ; VLX-NEXT: retq
11684 ;
11685 ; NoVLX-LABEL: test_vpcmpsged_v8i1_v32i1_mask_mem:
11686 ; NoVLX: # %bb.0: # %entry
11687 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
11688 ; NoVLX-NEXT: vmovdqa (%rdi), %ymm1
11689 ; NoVLX-NEXT: vpcmpnltd %zmm1, %zmm0, %k0
11690 ; NoVLX-NEXT: kshiftlw $8, %k0, %k0
11691 ; NoVLX-NEXT: kshiftrw $8, %k0, %k0
11692 ; NoVLX-NEXT: kmovw %k0, %eax
11693 ; NoVLX-NEXT: vzeroupper
11694 ; NoVLX-NEXT: retq
11695 entry:
11696 %0 = bitcast <4 x i64> %__a to <8 x i32>
11697 %load = load <4 x i64>, <4 x i64>* %__b
11698 %1 = bitcast <4 x i64> %load to <8 x i32>
11699 %2 = icmp sge <8 x i32> %0, %1
11700 %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
11701 %4 = bitcast <32 x i1> %3 to i32
11702 ret i32 %4
11703 }
11705 define zeroext i32 @test_masked_vpcmpsged_v8i1_v32i1_mask(i8 zeroext %__u, <4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
11706 ; VLX-LABEL: test_masked_vpcmpsged_v8i1_v32i1_mask:
11707 ; VLX: # %bb.0: # %entry
11708 ; VLX-NEXT: kmovd %edi, %k1
11709 ; VLX-NEXT: vpcmpnltd %ymm1, %ymm0, %k0 {%k1}
11710 ; VLX-NEXT: kmovd %k0, %eax
11711 ; VLX-NEXT: vzeroupper
11712 ; VLX-NEXT: retq
11713 ;
11714 ; NoVLX-LABEL: test_masked_vpcmpsged_v8i1_v32i1_mask:
11715 ; NoVLX: # %bb.0: # %entry
11716 ; NoVLX-NEXT: # kill: def $ymm1 killed $ymm1 def $zmm1
11717 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
11718 ; NoVLX-NEXT: kmovw %edi, %k1
11719 ; NoVLX-NEXT: vpcmpnltd %zmm1, %zmm0, %k0 {%k1}
11720 ; NoVLX-NEXT: kshiftlw $8, %k0, %k0
11721 ; NoVLX-NEXT: kshiftrw $8, %k0, %k0
11722 ; NoVLX-NEXT: kmovw %k0, %eax
11723 ; NoVLX-NEXT: vzeroupper
11724 ; NoVLX-NEXT: retq
11725 entry:
11726 %0 = bitcast <4 x i64> %__a to <8 x i32>
11727 %1 = bitcast <4 x i64> %__b to <8 x i32>
11728 %2 = icmp sge <8 x i32> %0, %1
11729 %3 = bitcast i8 %__u to <8 x i1>
11730 %4 = and <8 x i1> %2, %3
11731 %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
11732 %6 = bitcast <32 x i1> %5 to i32
11733 ret i32 %6
11734 }
11736 define zeroext i32 @test_masked_vpcmpsged_v8i1_v32i1_mask_mem(i8 zeroext %__u, <4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
11737 ; VLX-LABEL: test_masked_vpcmpsged_v8i1_v32i1_mask_mem:
11738 ; VLX: # %bb.0: # %entry
11739 ; VLX-NEXT: kmovd %edi, %k1
11740 ; VLX-NEXT: vpcmpnltd (%rsi), %ymm0, %k0 {%k1}
11741 ; VLX-NEXT: kmovd %k0, %eax
11742 ; VLX-NEXT: vzeroupper
11743 ; VLX-NEXT: retq
11744 ;
11745 ; NoVLX-LABEL: test_masked_vpcmpsged_v8i1_v32i1_mask_mem:
11746 ; NoVLX: # %bb.0: # %entry
11747 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
11748 ; NoVLX-NEXT: vmovdqa (%rsi), %ymm1
11749 ; NoVLX-NEXT: kmovw %edi, %k1
11750 ; NoVLX-NEXT: vpcmpnltd %zmm1, %zmm0, %k0 {%k1}
11751 ; NoVLX-NEXT: kshiftlw $8, %k0, %k0
11752 ; NoVLX-NEXT: kshiftrw $8, %k0, %k0
11753 ; NoVLX-NEXT: kmovw %k0, %eax
11754 ; NoVLX-NEXT: vzeroupper
11755 ; NoVLX-NEXT: retq
11756 entry:
11757 %0 = bitcast <4 x i64> %__a to <8 x i32>
11758 %load = load <4 x i64>, <4 x i64>* %__b
11759 %1 = bitcast <4 x i64> %load to <8 x i32>
11760 %2 = icmp sge <8 x i32> %0, %1
11761 %3 = bitcast i8 %__u to <8 x i1>
11762 %4 = and <8 x i1> %2, %3
11763 %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
11764 %6 = bitcast <32 x i1> %5 to i32
11765 ret i32 %6
11766 }
11769 define zeroext i32 @test_vpcmpsged_v8i1_v32i1_mask_mem_b(<4 x i64> %__a, i32* %__b) local_unnamed_addr {
11770 ; VLX-LABEL: test_vpcmpsged_v8i1_v32i1_mask_mem_b:
11771 ; VLX: # %bb.0: # %entry
11772 ; VLX-NEXT: vpcmpnltd (%rdi){1to8}, %ymm0, %k0
11773 ; VLX-NEXT: kmovd %k0, %eax
11774 ; VLX-NEXT: vzeroupper
11775 ; VLX-NEXT: retq
11776 ;
11777 ; NoVLX-LABEL: test_vpcmpsged_v8i1_v32i1_mask_mem_b:
11778 ; NoVLX: # %bb.0: # %entry
11779 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
11780 ; NoVLX-NEXT: vpcmpnltd (%rdi){1to16}, %zmm0, %k0
11781 ; NoVLX-NEXT: kshiftlw $8, %k0, %k0
11782 ; NoVLX-NEXT: kshiftrw $8, %k0, %k0
11783 ; NoVLX-NEXT: kmovw %k0, %eax
11784 ; NoVLX-NEXT: vzeroupper
11785 ; NoVLX-NEXT: retq
11786 entry:
11787 %0 = bitcast <4 x i64> %__a to <8 x i32>
11788 %load = load i32, i32* %__b
11789 %vec = insertelement <8 x i32> undef, i32 %load, i32 0
11790 %1 = shufflevector <8 x i32> %vec, <8 x i32> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
11791 %2 = icmp sge <8 x i32> %0, %1
11792 %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
11793 %4 = bitcast <32 x i1> %3 to i32
11794 ret i32 %4
11795 }
11797 define zeroext i32 @test_masked_vpcmpsged_v8i1_v32i1_mask_mem_b(i8 zeroext %__u, <4 x i64> %__a, i32* %__b) local_unnamed_addr {
11798 ; VLX-LABEL: test_masked_vpcmpsged_v8i1_v32i1_mask_mem_b:
11799 ; VLX: # %bb.0: # %entry
11800 ; VLX-NEXT: kmovd %edi, %k1
11801 ; VLX-NEXT: vpcmpnltd (%rsi){1to8}, %ymm0, %k0 {%k1}
11802 ; VLX-NEXT: kmovd %k0, %eax
11803 ; VLX-NEXT: vzeroupper
11804 ; VLX-NEXT: retq
11805 ;
11806 ; NoVLX-LABEL: test_masked_vpcmpsged_v8i1_v32i1_mask_mem_b:
11807 ; NoVLX: # %bb.0: # %entry
11808 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
11809 ; NoVLX-NEXT: kmovw %edi, %k1
11810 ; NoVLX-NEXT: vpcmpnltd (%rsi){1to16}, %zmm0, %k0 {%k1}
11811 ; NoVLX-NEXT: kshiftlw $8, %k0, %k0
11812 ; NoVLX-NEXT: kshiftrw $8, %k0, %k0
11813 ; NoVLX-NEXT: kmovw %k0, %eax
11814 ; NoVLX-NEXT: vzeroupper
11815 ; NoVLX-NEXT: retq
11816 entry:
11817 %0 = bitcast <4 x i64> %__a to <8 x i32>
11818 %load = load i32, i32* %__b
11819 %vec = insertelement <8 x i32> undef, i32 %load, i32 0
11820 %1 = shufflevector <8 x i32> %vec, <8 x i32> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
11821 %2 = icmp sge <8 x i32> %0, %1
11822 %3 = bitcast i8 %__u to <8 x i1>
11823 %4 = and <8 x i1> %3, %2
11824 %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
11825 %6 = bitcast <32 x i1> %5 to i32
11826 ret i32 %6
11827 }
11830 define zeroext i64 @test_vpcmpsged_v8i1_v64i1_mask(<4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
11831 ; VLX-LABEL: test_vpcmpsged_v8i1_v64i1_mask:
11832 ; VLX: # %bb.0: # %entry
11833 ; VLX-NEXT: vpcmpnltd %ymm1, %ymm0, %k0
11834 ; VLX-NEXT: kmovq %k0, %rax
11835 ; VLX-NEXT: vzeroupper
11836 ; VLX-NEXT: retq
11837 ;
11838 ; NoVLX-LABEL: test_vpcmpsged_v8i1_v64i1_mask:
11839 ; NoVLX: # %bb.0: # %entry
11840 ; NoVLX-NEXT: # kill: def $ymm1 killed $ymm1 def $zmm1
11841 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
11842 ; NoVLX-NEXT: vpcmpnltd %zmm1, %zmm0, %k0
11843 ; NoVLX-NEXT: kshiftlw $8, %k0, %k0
11844 ; NoVLX-NEXT: kshiftrw $8, %k0, %k0
11845 ; NoVLX-NEXT: kmovw %k0, %eax
11846 ; NoVLX-NEXT: vzeroupper
11847 ; NoVLX-NEXT: retq
11848 entry:
11849 %0 = bitcast <4 x i64> %__a to <8 x i32>
11850 %1 = bitcast <4 x i64> %__b to <8 x i32>
11851 %2 = icmp sge <8 x i32> %0, %1
11852 %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
11853 %4 = bitcast <64 x i1> %3 to i64
11854 ret i64 %4
11855 }
11857 define zeroext i64 @test_vpcmpsged_v8i1_v64i1_mask_mem(<4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
11858 ; VLX-LABEL: test_vpcmpsged_v8i1_v64i1_mask_mem:
11859 ; VLX: # %bb.0: # %entry
11860 ; VLX-NEXT: vpcmpnltd (%rdi), %ymm0, %k0
11861 ; VLX-NEXT: kmovq %k0, %rax
11862 ; VLX-NEXT: vzeroupper
11863 ; VLX-NEXT: retq
11864 ;
11865 ; NoVLX-LABEL: test_vpcmpsged_v8i1_v64i1_mask_mem:
11866 ; NoVLX: # %bb.0: # %entry
11867 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
11868 ; NoVLX-NEXT: vmovdqa (%rdi), %ymm1
11869 ; NoVLX-NEXT: vpcmpnltd %zmm1, %zmm0, %k0
11870 ; NoVLX-NEXT: kshiftlw $8, %k0, %k0
11871 ; NoVLX-NEXT: kshiftrw $8, %k0, %k0
11872 ; NoVLX-NEXT: kmovw %k0, %eax
11873 ; NoVLX-NEXT: vzeroupper
11874 ; NoVLX-NEXT: retq
11875 entry:
11876 %0 = bitcast <4 x i64> %__a to <8 x i32>
11877 %load = load <4 x i64>, <4 x i64>* %__b
11878 %1 = bitcast <4 x i64> %load to <8 x i32>
11879 %2 = icmp sge <8 x i32> %0, %1
11880 %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
11881 %4 = bitcast <64 x i1> %3 to i64
11882 ret i64 %4
11883 }
11885 define zeroext i64 @test_masked_vpcmpsged_v8i1_v64i1_mask(i8 zeroext %__u, <4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
11886 ; VLX-LABEL: test_masked_vpcmpsged_v8i1_v64i1_mask:
11887 ; VLX: # %bb.0: # %entry
11888 ; VLX-NEXT: kmovd %edi, %k1
11889 ; VLX-NEXT: vpcmpnltd %ymm1, %ymm0, %k0 {%k1}
11890 ; VLX-NEXT: kmovq %k0, %rax
11891 ; VLX-NEXT: vzeroupper
11892 ; VLX-NEXT: retq
11893 ;
11894 ; NoVLX-LABEL: test_masked_vpcmpsged_v8i1_v64i1_mask:
11895 ; NoVLX: # %bb.0: # %entry
11896 ; NoVLX-NEXT: # kill: def $ymm1 killed $ymm1 def $zmm1
11897 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
11898 ; NoVLX-NEXT: kmovw %edi, %k1
11899 ; NoVLX-NEXT: vpcmpnltd %zmm1, %zmm0, %k0 {%k1}
11900 ; NoVLX-NEXT: kshiftlw $8, %k0, %k0
11901 ; NoVLX-NEXT: kshiftrw $8, %k0, %k0
11902 ; NoVLX-NEXT: kmovw %k0, %eax
11903 ; NoVLX-NEXT: vzeroupper
11904 ; NoVLX-NEXT: retq
11905 entry:
11906 %0 = bitcast <4 x i64> %__a to <8 x i32>
11907 %1 = bitcast <4 x i64> %__b to <8 x i32>
11908 %2 = icmp sge <8 x i32> %0, %1
11909 %3 = bitcast i8 %__u to <8 x i1>
11910 %4 = and <8 x i1> %2, %3
11911 %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
11912 %6 = bitcast <64 x i1> %5 to i64
11913 ret i64 %6
11914 }
11916 define zeroext i64 @test_masked_vpcmpsged_v8i1_v64i1_mask_mem(i8 zeroext %__u, <4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
11917 ; VLX-LABEL: test_masked_vpcmpsged_v8i1_v64i1_mask_mem:
11918 ; VLX: # %bb.0: # %entry
11919 ; VLX-NEXT: kmovd %edi, %k1
11920 ; VLX-NEXT: vpcmpnltd (%rsi), %ymm0, %k0 {%k1}
11921 ; VLX-NEXT: kmovq %k0, %rax
11922 ; VLX-NEXT: vzeroupper
11923 ; VLX-NEXT: retq
11924 ;
11925 ; NoVLX-LABEL: test_masked_vpcmpsged_v8i1_v64i1_mask_mem:
11926 ; NoVLX: # %bb.0: # %entry
11927 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
11928 ; NoVLX-NEXT: vmovdqa (%rsi), %ymm1
11929 ; NoVLX-NEXT: kmovw %edi, %k1
11930 ; NoVLX-NEXT: vpcmpnltd %zmm1, %zmm0, %k0 {%k1}
11931 ; NoVLX-NEXT: kshiftlw $8, %k0, %k0
11932 ; NoVLX-NEXT: kshiftrw $8, %k0, %k0
11933 ; NoVLX-NEXT: kmovw %k0, %eax
11934 ; NoVLX-NEXT: vzeroupper
11935 ; NoVLX-NEXT: retq
11936 entry:
11937 %0 = bitcast <4 x i64> %__a to <8 x i32>
11938 %load = load <4 x i64>, <4 x i64>* %__b
11939 %1 = bitcast <4 x i64> %load to <8 x i32>
11940 %2 = icmp sge <8 x i32> %0, %1
11941 %3 = bitcast i8 %__u to <8 x i1>
11942 %4 = and <8 x i1> %2, %3
11943 %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
11944 %6 = bitcast <64 x i1> %5 to i64
11945 ret i64 %6
11946 }
11949 define zeroext i64 @test_vpcmpsged_v8i1_v64i1_mask_mem_b(<4 x i64> %__a, i32* %__b) local_unnamed_addr {
11950 ; VLX-LABEL: test_vpcmpsged_v8i1_v64i1_mask_mem_b:
11951 ; VLX: # %bb.0: # %entry
11952 ; VLX-NEXT: vpcmpnltd (%rdi){1to8}, %ymm0, %k0
11953 ; VLX-NEXT: kmovq %k0, %rax
11954 ; VLX-NEXT: vzeroupper
11955 ; VLX-NEXT: retq
11956 ;
11957 ; NoVLX-LABEL: test_vpcmpsged_v8i1_v64i1_mask_mem_b:
11958 ; NoVLX: # %bb.0: # %entry
11959 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
11960 ; NoVLX-NEXT: vpcmpnltd (%rdi){1to16}, %zmm0, %k0
11961 ; NoVLX-NEXT: kshiftlw $8, %k0, %k0
11962 ; NoVLX-NEXT: kshiftrw $8, %k0, %k0
11963 ; NoVLX-NEXT: kmovw %k0, %eax
11964 ; NoVLX-NEXT: vzeroupper
11965 ; NoVLX-NEXT: retq
11966 entry:
11967 %0 = bitcast <4 x i64> %__a to <8 x i32>
11968 %load = load i32, i32* %__b
11969 %vec = insertelement <8 x i32> undef, i32 %load, i32 0
11970 %1 = shufflevector <8 x i32> %vec, <8 x i32> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
11971 %2 = icmp sge <8 x i32> %0, %1
11972 %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
11973 %4 = bitcast <64 x i1> %3 to i64
11974 ret i64 %4
11975 }
11977 define zeroext i64 @test_masked_vpcmpsged_v8i1_v64i1_mask_mem_b(i8 zeroext %__u, <4 x i64> %__a, i32* %__b) local_unnamed_addr {
11978 ; VLX-LABEL: test_masked_vpcmpsged_v8i1_v64i1_mask_mem_b:
11979 ; VLX: # %bb.0: # %entry
11980 ; VLX-NEXT: kmovd %edi, %k1
11981 ; VLX-NEXT: vpcmpnltd (%rsi){1to8}, %ymm0, %k0 {%k1}
11982 ; VLX-NEXT: kmovq %k0, %rax
11983 ; VLX-NEXT: vzeroupper
11984 ; VLX-NEXT: retq
11985 ;
11986 ; NoVLX-LABEL: test_masked_vpcmpsged_v8i1_v64i1_mask_mem_b:
11987 ; NoVLX: # %bb.0: # %entry
11988 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
11989 ; NoVLX-NEXT: kmovw %edi, %k1
11990 ; NoVLX-NEXT: vpcmpnltd (%rsi){1to16}, %zmm0, %k0 {%k1}
11991 ; NoVLX-NEXT: kshiftlw $8, %k0, %k0
11992 ; NoVLX-NEXT: kshiftrw $8, %k0, %k0
11993 ; NoVLX-NEXT: kmovw %k0, %eax
11994 ; NoVLX-NEXT: vzeroupper
11995 ; NoVLX-NEXT: retq
11996 entry:
11997 %0 = bitcast <4 x i64> %__a to <8 x i32>
11998 %load = load i32, i32* %__b
11999 %vec = insertelement <8 x i32> undef, i32 %load, i32 0
12000 %1 = shufflevector <8 x i32> %vec, <8 x i32> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
12001 %2 = icmp sge <8 x i32> %0, %1
12002 %3 = bitcast i8 %__u to <8 x i1>
12003 %4 = and <8 x i1> %3, %2
12004 %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
12005 %6 = bitcast <64 x i1> %5 to i64
12006 ret i64 %6
12007 }
12010 define zeroext i32 @test_vpcmpsged_v16i1_v32i1_mask(<8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
12011 ; VLX-LABEL: test_vpcmpsged_v16i1_v32i1_mask:
12012 ; VLX: # %bb.0: # %entry
12013 ; VLX-NEXT: vpcmpnltd %zmm1, %zmm0, %k0
12014 ; VLX-NEXT: kmovd %k0, %eax
12015 ; VLX-NEXT: vzeroupper
12016 ; VLX-NEXT: retq
12017 ;
12018 ; NoVLX-LABEL: test_vpcmpsged_v16i1_v32i1_mask:
12019 ; NoVLX: # %bb.0: # %entry
12020 ; NoVLX-NEXT: vpcmpnltd %zmm1, %zmm0, %k0
12021 ; NoVLX-NEXT: kmovw %k0, %eax
12022 ; NoVLX-NEXT: vzeroupper
12023 ; NoVLX-NEXT: retq
12024 entry:
12025 %0 = bitcast <8 x i64> %__a to <16 x i32>
12026 %1 = bitcast <8 x i64> %__b to <16 x i32>
12027 %2 = icmp sge <16 x i32> %0, %1
12028 %3 = shufflevector <16 x i1> %2, <16 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>
12029 %4 = bitcast <32 x i1> %3 to i32
12030 ret i32 %4
12031 }
12033 define zeroext i32 @test_vpcmpsged_v16i1_v32i1_mask_mem(<8 x i64> %__a, <8 x i64>* %__b) local_unnamed_addr {
12034 ; VLX-LABEL: test_vpcmpsged_v16i1_v32i1_mask_mem:
12035 ; VLX: # %bb.0: # %entry
12036 ; VLX-NEXT: vpcmpnltd (%rdi), %zmm0, %k0
12037 ; VLX-NEXT: kmovd %k0, %eax
12038 ; VLX-NEXT: vzeroupper
12039 ; VLX-NEXT: retq
12040 ;
12041 ; NoVLX-LABEL: test_vpcmpsged_v16i1_v32i1_mask_mem:
12042 ; NoVLX: # %bb.0: # %entry
12043 ; NoVLX-NEXT: vpcmpnltd (%rdi), %zmm0, %k0
12044 ; NoVLX-NEXT: kmovw %k0, %eax
12045 ; NoVLX-NEXT: vzeroupper
12046 ; NoVLX-NEXT: retq
12047 entry:
12048 %0 = bitcast <8 x i64> %__a to <16 x i32>
12049 %load = load <8 x i64>, <8 x i64>* %__b
12050 %1 = bitcast <8 x i64> %load to <16 x i32>
12051 %2 = icmp sge <16 x i32> %0, %1
12052 %3 = shufflevector <16 x i1> %2, <16 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>
12053 %4 = bitcast <32 x i1> %3 to i32
12054 ret i32 %4
12055 }
12057 define zeroext i32 @test_masked_vpcmpsged_v16i1_v32i1_mask(i16 zeroext %__u, <8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
12058 ; VLX-LABEL: test_masked_vpcmpsged_v16i1_v32i1_mask:
12059 ; VLX: # %bb.0: # %entry
12060 ; VLX-NEXT: kmovd %edi, %k1
12061 ; VLX-NEXT: vpcmpnltd %zmm1, %zmm0, %k0 {%k1}
12062 ; VLX-NEXT: kmovd %k0, %eax
12063 ; VLX-NEXT: vzeroupper
12064 ; VLX-NEXT: retq
12065 ;
12066 ; NoVLX-LABEL: test_masked_vpcmpsged_v16i1_v32i1_mask:
12067 ; NoVLX: # %bb.0: # %entry
12068 ; NoVLX-NEXT: vpcmpnltd %zmm1, %zmm0, %k0
12069 ; NoVLX-NEXT: kmovw %k0, %eax
12070 ; NoVLX-NEXT: andl %edi, %eax
12071 ; NoVLX-NEXT: vzeroupper
12072 ; NoVLX-NEXT: retq
12073 entry:
12074 %0 = bitcast <8 x i64> %__a to <16 x i32>
12075 %1 = bitcast <8 x i64> %__b to <16 x i32>
12076 %2 = icmp sge <16 x i32> %0, %1
12077 %3 = bitcast i16 %__u to <16 x i1>
12078 %4 = and <16 x i1> %2, %3
12079 %5 = shufflevector <16 x i1> %4, <16 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>
12080 %6 = bitcast <32 x i1> %5 to i32
12081 ret i32 %6
12082 }
12084 define zeroext i32 @test_masked_vpcmpsged_v16i1_v32i1_mask_mem(i16 zeroext %__u, <8 x i64> %__a, <8 x i64>* %__b) local_unnamed_addr {
12085 ; VLX-LABEL: test_masked_vpcmpsged_v16i1_v32i1_mask_mem:
12086 ; VLX: # %bb.0: # %entry
12087 ; VLX-NEXT: kmovd %edi, %k1
12088 ; VLX-NEXT: vpcmpnltd (%rsi), %zmm0, %k0 {%k1}
12089 ; VLX-NEXT: kmovd %k0, %eax
12090 ; VLX-NEXT: vzeroupper
12091 ; VLX-NEXT: retq
12092 ;
12093 ; NoVLX-LABEL: test_masked_vpcmpsged_v16i1_v32i1_mask_mem:
12094 ; NoVLX: # %bb.0: # %entry
12095 ; NoVLX-NEXT: vpcmpnltd (%rsi), %zmm0, %k0
12096 ; NoVLX-NEXT: kmovw %k0, %eax
12097 ; NoVLX-NEXT: andl %edi, %eax
12098 ; NoVLX-NEXT: vzeroupper
12099 ; NoVLX-NEXT: retq
12100 entry:
12101 %0 = bitcast <8 x i64> %__a to <16 x i32>
12102 %load = load <8 x i64>, <8 x i64>* %__b
12103 %1 = bitcast <8 x i64> %load to <16 x i32>
12104 %2 = icmp sge <16 x i32> %0, %1
12105 %3 = bitcast i16 %__u to <16 x i1>
12106 %4 = and <16 x i1> %2, %3
12107 %5 = shufflevector <16 x i1> %4, <16 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>
12108 %6 = bitcast <32 x i1> %5 to i32
12109 ret i32 %6
12110 }
12113 define zeroext i32 @test_vpcmpsged_v16i1_v32i1_mask_mem_b(<8 x i64> %__a, i32* %__b) local_unnamed_addr {
12114 ; VLX-LABEL: test_vpcmpsged_v16i1_v32i1_mask_mem_b:
12115 ; VLX: # %bb.0: # %entry
12116 ; VLX-NEXT: vpcmpnltd (%rdi){1to16}, %zmm0, %k0
12117 ; VLX-NEXT: kmovd %k0, %eax
12118 ; VLX-NEXT: vzeroupper
12119 ; VLX-NEXT: retq
12120 ;
12121 ; NoVLX-LABEL: test_vpcmpsged_v16i1_v32i1_mask_mem_b:
12122 ; NoVLX: # %bb.0: # %entry
12123 ; NoVLX-NEXT: vpcmpnltd (%rdi){1to16}, %zmm0, %k0
12124 ; NoVLX-NEXT: kmovw %k0, %eax
12125 ; NoVLX-NEXT: vzeroupper
12126 ; NoVLX-NEXT: retq
12127 entry:
12128 %0 = bitcast <8 x i64> %__a to <16 x i32>
12129 %load = load i32, i32* %__b
12130 %vec = insertelement <16 x i32> undef, i32 %load, i32 0
12131 %1 = shufflevector <16 x i32> %vec, <16 x i32> undef, <16 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
12132 %2 = icmp sge <16 x i32> %0, %1
12133 %3 = shufflevector <16 x i1> %2, <16 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>
12134 %4 = bitcast <32 x i1> %3 to i32
12135 ret i32 %4
12136 }
12138 define zeroext i32 @test_masked_vpcmpsged_v16i1_v32i1_mask_mem_b(i16 zeroext %__u, <8 x i64> %__a, i32* %__b) local_unnamed_addr {
12139 ; VLX-LABEL: test_masked_vpcmpsged_v16i1_v32i1_mask_mem_b:
12140 ; VLX: # %bb.0: # %entry
12141 ; VLX-NEXT: kmovd %edi, %k1
12142 ; VLX-NEXT: vpcmpnltd (%rsi){1to16}, %zmm0, %k0 {%k1}
12143 ; VLX-NEXT: kmovd %k0, %eax
12144 ; VLX-NEXT: vzeroupper
12145 ; VLX-NEXT: retq
12146 ;
12147 ; NoVLX-LABEL: test_masked_vpcmpsged_v16i1_v32i1_mask_mem_b:
12148 ; NoVLX: # %bb.0: # %entry
12149 ; NoVLX-NEXT: vpcmpnltd (%rsi){1to16}, %zmm0, %k0
12150 ; NoVLX-NEXT: kmovw %k0, %eax
12151 ; NoVLX-NEXT: andl %edi, %eax
12152 ; NoVLX-NEXT: vzeroupper
12153 ; NoVLX-NEXT: retq
12154 entry:
12155 %0 = bitcast <8 x i64> %__a to <16 x i32>
12156 %load = load i32, i32* %__b
12157 %vec = insertelement <16 x i32> undef, i32 %load, i32 0
12158 %1 = shufflevector <16 x i32> %vec, <16 x i32> undef, <16 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
12159 %2 = icmp sge <16 x i32> %0, %1
12160 %3 = bitcast i16 %__u to <16 x i1>
12161 %4 = and <16 x i1> %3, %2
12162 %5 = shufflevector <16 x i1> %4, <16 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>
12163 %6 = bitcast <32 x i1> %5 to i32
12164 ret i32 %6
12165 }
12168 define zeroext i64 @test_vpcmpsged_v16i1_v64i1_mask(<8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
12169 ; VLX-LABEL: test_vpcmpsged_v16i1_v64i1_mask:
12170 ; VLX: # %bb.0: # %entry
12171 ; VLX-NEXT: vpcmpnltd %zmm1, %zmm0, %k0
12172 ; VLX-NEXT: kmovq %k0, %rax
12173 ; VLX-NEXT: vzeroupper
12174 ; VLX-NEXT: retq
12175 ;
12176 ; NoVLX-LABEL: test_vpcmpsged_v16i1_v64i1_mask:
12177 ; NoVLX: # %bb.0: # %entry
12178 ; NoVLX-NEXT: vpcmpnltd %zmm1, %zmm0, %k0
12179 ; NoVLX-NEXT: kmovw %k0, %eax
12180 ; NoVLX-NEXT: vzeroupper
12181 ; NoVLX-NEXT: retq
12182 entry:
12183 %0 = bitcast <8 x i64> %__a to <16 x i32>
12184 %1 = bitcast <8 x i64> %__b to <16 x i32>
12185 %2 = icmp sge <16 x i32> %0, %1
12186 %3 = shufflevector <16 x i1> %2, <16 x i1> zeroinitializer, <64 x i32> <i32 0,i32 1,i32 2,i32 3,i32 4,i32 5,i32 6,i32 7,i32 8,i32 9,i32 10,i32 11,i32 12,i32 13,i32 14,i32 15,i32 16,i32 17,i32 18,i32 19,i32 20,i32 21,i32 22,i32 23,i32 24,i32 25,i32 26,i32 27,i32 28,i32 29,i32 30,i32 31,i32 16,i32 17,i32 18,i32 19,i32 20,i32 21,i32 22,i32 23,i32 24,i32 25,i32 26,i32 27,i32 28,i32 29,i32 30,i32 31,i32 16,i32 17,i32 18,i32 19,i32 20,i32 21,i32 22,i32 23,i32 24,i32 25,i32 26,i32 27,i32 28,i32 29,i32 30,i32 31>
12187 %4 = bitcast <64 x i1> %3 to i64
12188 ret i64 %4
12189 }
12191 define zeroext i64 @test_vpcmpsged_v16i1_v64i1_mask_mem(<8 x i64> %__a, <8 x i64>* %__b) local_unnamed_addr {
12192 ; VLX-LABEL: test_vpcmpsged_v16i1_v64i1_mask_mem:
12193 ; VLX: # %bb.0: # %entry
12194 ; VLX-NEXT: vpcmpnltd (%rdi), %zmm0, %k0
12195 ; VLX-NEXT: kmovq %k0, %rax
12196 ; VLX-NEXT: vzeroupper
12197 ; VLX-NEXT: retq
12198 ;
12199 ; NoVLX-LABEL: test_vpcmpsged_v16i1_v64i1_mask_mem:
12200 ; NoVLX: # %bb.0: # %entry
12201 ; NoVLX-NEXT: vpcmpnltd (%rdi), %zmm0, %k0
12202 ; NoVLX-NEXT: kmovw %k0, %eax
12203 ; NoVLX-NEXT: vzeroupper
12204 ; NoVLX-NEXT: retq
12205 entry:
12206 %0 = bitcast <8 x i64> %__a to <16 x i32>
12207 %load = load <8 x i64>, <8 x i64>* %__b
12208 %1 = bitcast <8 x i64> %load to <16 x i32>
12209 %2 = icmp sge <16 x i32> %0, %1
12210 %3 = shufflevector <16 x i1> %2, <16 x i1> zeroinitializer, <64 x i32> <i32 0,i32 1,i32 2,i32 3,i32 4,i32 5,i32 6,i32 7,i32 8,i32 9,i32 10,i32 11,i32 12,i32 13,i32 14,i32 15,i32 16,i32 17,i32 18,i32 19,i32 20,i32 21,i32 22,i32 23,i32 24,i32 25,i32 26,i32 27,i32 28,i32 29,i32 30,i32 31,i32 16,i32 17,i32 18,i32 19,i32 20,i32 21,i32 22,i32 23,i32 24,i32 25,i32 26,i32 27,i32 28,i32 29,i32 30,i32 31,i32 16,i32 17,i32 18,i32 19,i32 20,i32 21,i32 22,i32 23,i32 24,i32 25,i32 26,i32 27,i32 28,i32 29,i32 30,i32 31>
12211 %4 = bitcast <64 x i1> %3 to i64
12212 ret i64 %4
12213 }
12215 define zeroext i64 @test_masked_vpcmpsged_v16i1_v64i1_mask(i16 zeroext %__u, <8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
12216 ; VLX-LABEL: test_masked_vpcmpsged_v16i1_v64i1_mask:
12217 ; VLX: # %bb.0: # %entry
12218 ; VLX-NEXT: kmovd %edi, %k1
12219 ; VLX-NEXT: vpcmpnltd %zmm1, %zmm0, %k0 {%k1}
12220 ; VLX-NEXT: kmovq %k0, %rax
12221 ; VLX-NEXT: vzeroupper
12222 ; VLX-NEXT: retq
12223 ;
12224 ; NoVLX-LABEL: test_masked_vpcmpsged_v16i1_v64i1_mask:
12225 ; NoVLX: # %bb.0: # %entry
12226 ; NoVLX-NEXT: vpcmpnltd %zmm1, %zmm0, %k0
12227 ; NoVLX-NEXT: kmovw %k0, %eax
12228 ; NoVLX-NEXT: andl %edi, %eax
12229 ; NoVLX-NEXT: vzeroupper
12230 ; NoVLX-NEXT: retq
12231 entry:
12232 %0 = bitcast <8 x i64> %__a to <16 x i32>
12233 %1 = bitcast <8 x i64> %__b to <16 x i32>
12234 %2 = icmp sge <16 x i32> %0, %1
12235 %3 = bitcast i16 %__u to <16 x i1>
12236 %4 = and <16 x i1> %2, %3
12237 %5 = shufflevector <16 x i1> %4, <16 x i1> zeroinitializer, <64 x i32> <i32 0,i32 1,i32 2,i32 3,i32 4,i32 5,i32 6,i32 7,i32 8,i32 9,i32 10,i32 11,i32 12,i32 13,i32 14,i32 15,i32 16,i32 17,i32 18,i32 19,i32 20,i32 21,i32 22,i32 23,i32 24,i32 25,i32 26,i32 27,i32 28,i32 29,i32 30,i32 31,i32 16,i32 17,i32 18,i32 19,i32 20,i32 21,i32 22,i32 23,i32 24,i32 25,i32 26,i32 27,i32 28,i32 29,i32 30,i32 31,i32 16,i32 17,i32 18,i32 19,i32 20,i32 21,i32 22,i32 23,i32 24,i32 25,i32 26,i32 27,i32 28,i32 29,i32 30,i32 31>
12238 %6 = bitcast <64 x i1> %5 to i64
12239 ret i64 %6
12240 }
12242 define zeroext i64 @test_masked_vpcmpsged_v16i1_v64i1_mask_mem(i16 zeroext %__u, <8 x i64> %__a, <8 x i64>* %__b) local_unnamed_addr {
12243 ; VLX-LABEL: test_masked_vpcmpsged_v16i1_v64i1_mask_mem:
12244 ; VLX: # %bb.0: # %entry
12245 ; VLX-NEXT: kmovd %edi, %k1
12246 ; VLX-NEXT: vpcmpnltd (%rsi), %zmm0, %k0 {%k1}
12247 ; VLX-NEXT: kmovq %k0, %rax
12248 ; VLX-NEXT: vzeroupper
12249 ; VLX-NEXT: retq
12250 ;
12251 ; NoVLX-LABEL: test_masked_vpcmpsged_v16i1_v64i1_mask_mem:
12252 ; NoVLX: # %bb.0: # %entry
12253 ; NoVLX-NEXT: vpcmpnltd (%rsi), %zmm0, %k0
12254 ; NoVLX-NEXT: kmovw %k0, %eax
12255 ; NoVLX-NEXT: andl %edi, %eax
12256 ; NoVLX-NEXT: vzeroupper
12257 ; NoVLX-NEXT: retq
12258 entry:
12259 %0 = bitcast <8 x i64> %__a to <16 x i32>
12260 %load = load <8 x i64>, <8 x i64>* %__b
12261 %1 = bitcast <8 x i64> %load to <16 x i32>
12262 %2 = icmp sge <16 x i32> %0, %1
12263 %3 = bitcast i16 %__u to <16 x i1>
12264 %4 = and <16 x i1> %2, %3
12265 %5 = shufflevector <16 x i1> %4, <16 x i1> zeroinitializer, <64 x i32> <i32 0,i32 1,i32 2,i32 3,i32 4,i32 5,i32 6,i32 7,i32 8,i32 9,i32 10,i32 11,i32 12,i32 13,i32 14,i32 15,i32 16,i32 17,i32 18,i32 19,i32 20,i32 21,i32 22,i32 23,i32 24,i32 25,i32 26,i32 27,i32 28,i32 29,i32 30,i32 31,i32 16,i32 17,i32 18,i32 19,i32 20,i32 21,i32 22,i32 23,i32 24,i32 25,i32 26,i32 27,i32 28,i32 29,i32 30,i32 31,i32 16,i32 17,i32 18,i32 19,i32 20,i32 21,i32 22,i32 23,i32 24,i32 25,i32 26,i32 27,i32 28,i32 29,i32 30,i32 31>
12266 %6 = bitcast <64 x i1> %5 to i64
12267 ret i64 %6
12268 }
12271 define zeroext i64 @test_vpcmpsged_v16i1_v64i1_mask_mem_b(<8 x i64> %__a, i32* %__b) local_unnamed_addr {
12272 ; VLX-LABEL: test_vpcmpsged_v16i1_v64i1_mask_mem_b:
12273 ; VLX: # %bb.0: # %entry
12274 ; VLX-NEXT: vpcmpnltd (%rdi){1to16}, %zmm0, %k0
12275 ; VLX-NEXT: kmovq %k0, %rax
12276 ; VLX-NEXT: vzeroupper
12277 ; VLX-NEXT: retq
12278 ;
12279 ; NoVLX-LABEL: test_vpcmpsged_v16i1_v64i1_mask_mem_b:
12280 ; NoVLX: # %bb.0: # %entry
12281 ; NoVLX-NEXT: vpcmpnltd (%rdi){1to16}, %zmm0, %k0
12282 ; NoVLX-NEXT: kmovw %k0, %eax
12283 ; NoVLX-NEXT: vzeroupper
12284 ; NoVLX-NEXT: retq
12285 entry:
12286 %0 = bitcast <8 x i64> %__a to <16 x i32>
12287 %load = load i32, i32* %__b
12288 %vec = insertelement <16 x i32> undef, i32 %load, i32 0
12289 %1 = shufflevector <16 x i32> %vec, <16 x i32> undef, <16 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
12290 %2 = icmp sge <16 x i32> %0, %1
12291 %3 = shufflevector <16 x i1> %2, <16 x i1> zeroinitializer, <64 x i32> <i32 0,i32 1,i32 2,i32 3,i32 4,i32 5,i32 6,i32 7,i32 8,i32 9,i32 10,i32 11,i32 12,i32 13,i32 14,i32 15,i32 16,i32 17,i32 18,i32 19,i32 20,i32 21,i32 22,i32 23,i32 24,i32 25,i32 26,i32 27,i32 28,i32 29,i32 30,i32 31,i32 16,i32 17,i32 18,i32 19,i32 20,i32 21,i32 22,i32 23,i32 24,i32 25,i32 26,i32 27,i32 28,i32 29,i32 30,i32 31,i32 16,i32 17,i32 18,i32 19,i32 20,i32 21,i32 22,i32 23,i32 24,i32 25,i32 26,i32 27,i32 28,i32 29,i32 30,i32 31>
12292 %4 = bitcast <64 x i1> %3 to i64
12293 ret i64 %4
12294 }
12296 define zeroext i64 @test_masked_vpcmpsged_v16i1_v64i1_mask_mem_b(i16 zeroext %__u, <8 x i64> %__a, i32* %__b) local_unnamed_addr {
12297 ; VLX-LABEL: test_masked_vpcmpsged_v16i1_v64i1_mask_mem_b:
12298 ; VLX: # %bb.0: # %entry
12299 ; VLX-NEXT: kmovd %edi, %k1
12300 ; VLX-NEXT: vpcmpnltd (%rsi){1to16}, %zmm0, %k0 {%k1}
12301 ; VLX-NEXT: kmovq %k0, %rax
12302 ; VLX-NEXT: vzeroupper
12303 ; VLX-NEXT: retq
12304 ;
12305 ; NoVLX-LABEL: test_masked_vpcmpsged_v16i1_v64i1_mask_mem_b:
12306 ; NoVLX: # %bb.0: # %entry
12307 ; NoVLX-NEXT: vpcmpnltd (%rsi){1to16}, %zmm0, %k0
12308 ; NoVLX-NEXT: kmovw %k0, %eax
12309 ; NoVLX-NEXT: andl %edi, %eax
12310 ; NoVLX-NEXT: vzeroupper
12311 ; NoVLX-NEXT: retq
12312 entry:
12313 %0 = bitcast <8 x i64> %__a to <16 x i32>
12314 %load = load i32, i32* %__b
12315 %vec = insertelement <16 x i32> undef, i32 %load, i32 0
12316 %1 = shufflevector <16 x i32> %vec, <16 x i32> undef, <16 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
12317 %2 = icmp sge <16 x i32> %0, %1
12318 %3 = bitcast i16 %__u to <16 x i1>
12319 %4 = and <16 x i1> %3, %2
12320 %5 = shufflevector <16 x i1> %4, <16 x i1> zeroinitializer, <64 x i32> <i32 0,i32 1,i32 2,i32 3,i32 4,i32 5,i32 6,i32 7,i32 8,i32 9,i32 10,i32 11,i32 12,i32 13,i32 14,i32 15,i32 16,i32 17,i32 18,i32 19,i32 20,i32 21,i32 22,i32 23,i32 24,i32 25,i32 26,i32 27,i32 28,i32 29,i32 30,i32 31,i32 16,i32 17,i32 18,i32 19,i32 20,i32 21,i32 22,i32 23,i32 24,i32 25,i32 26,i32 27,i32 28,i32 29,i32 30,i32 31,i32 16,i32 17,i32 18,i32 19,i32 20,i32 21,i32 22,i32 23,i32 24,i32 25,i32 26,i32 27,i32 28,i32 29,i32 30,i32 31>
12321 %6 = bitcast <64 x i1> %5 to i64
12322 ret i64 %6
12323 }
12326 define zeroext i4 @test_vpcmpsgeq_v2i1_v4i1_mask(<2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
12327 ; VLX-LABEL: test_vpcmpsgeq_v2i1_v4i1_mask:
12328 ; VLX: # %bb.0: # %entry
12329 ; VLX-NEXT: vpcmpnltq %xmm1, %xmm0, %k0
12330 ; VLX-NEXT: kmovb %k0, %eax
12331 ; VLX-NEXT: retq
12332 ;
12333 ; NoVLX-LABEL: test_vpcmpsgeq_v2i1_v4i1_mask:
12334 ; NoVLX: # %bb.0: # %entry
12335 ; NoVLX-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
12336 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
12337 ; NoVLX-NEXT: vpcmpnltq %zmm1, %zmm0, %k0
12338 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
12339 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
12340 ; NoVLX-NEXT: kmovw %k0, %eax
12341 ; NoVLX-NEXT: vzeroupper
12342 ; NoVLX-NEXT: retq
12343 entry:
12344 %0 = bitcast <2 x i64> %__a to <2 x i64>
12345 %1 = bitcast <2 x i64> %__b to <2 x i64>
12346 %2 = icmp sge <2 x i64> %0, %1
12347 %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
12348 %4 = bitcast <4 x i1> %3 to i4
12349 ret i4 %4
12350 }
12352 define zeroext i4 @test_vpcmpsgeq_v2i1_v4i1_mask_mem(<2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
12353 ; VLX-LABEL: test_vpcmpsgeq_v2i1_v4i1_mask_mem:
12354 ; VLX: # %bb.0: # %entry
12355 ; VLX-NEXT: vpcmpnltq (%rdi), %xmm0, %k0
12356 ; VLX-NEXT: kmovb %k0, %eax
12357 ; VLX-NEXT: retq
12358 ;
12359 ; NoVLX-LABEL: test_vpcmpsgeq_v2i1_v4i1_mask_mem:
12360 ; NoVLX: # %bb.0: # %entry
12361 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
12362 ; NoVLX-NEXT: vmovdqa (%rdi), %xmm1
12363 ; NoVLX-NEXT: vpcmpnltq %zmm1, %zmm0, %k0
12364 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
12365 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
12366 ; NoVLX-NEXT: kmovw %k0, %eax
12367 ; NoVLX-NEXT: vzeroupper
12368 ; NoVLX-NEXT: retq
12369 entry:
12370 %0 = bitcast <2 x i64> %__a to <2 x i64>
12371 %load = load <2 x i64>, <2 x i64>* %__b
12372 %1 = bitcast <2 x i64> %load to <2 x i64>
12373 %2 = icmp sge <2 x i64> %0, %1
12374 %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
12375 %4 = bitcast <4 x i1> %3 to i4
12376 ret i4 %4
12377 }
12379 define zeroext i4 @test_masked_vpcmpsgeq_v2i1_v4i1_mask(i8 zeroext %__u, <2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
12380 ; VLX-LABEL: test_masked_vpcmpsgeq_v2i1_v4i1_mask:
12381 ; VLX: # %bb.0: # %entry
12382 ; VLX-NEXT: kmovd %edi, %k1
12383 ; VLX-NEXT: vpcmpnltq %xmm1, %xmm0, %k0 {%k1}
12384 ; VLX-NEXT: kmovb %k0, %eax
12385 ; VLX-NEXT: retq
12386 ;
12387 ; NoVLX-LABEL: test_masked_vpcmpsgeq_v2i1_v4i1_mask:
12388 ; NoVLX: # %bb.0: # %entry
12389 ; NoVLX-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
12390 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
12391 ; NoVLX-NEXT: kmovw %edi, %k1
12392 ; NoVLX-NEXT: vpcmpnltq %zmm1, %zmm0, %k0 {%k1}
12393 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
12394 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
12395 ; NoVLX-NEXT: kmovw %k0, %eax
12396 ; NoVLX-NEXT: vzeroupper
12397 ; NoVLX-NEXT: retq
12398 entry:
12399 %0 = bitcast <2 x i64> %__a to <2 x i64>
12400 %1 = bitcast <2 x i64> %__b to <2 x i64>
12401 %2 = icmp sge <2 x i64> %0, %1
12402 %3 = bitcast i8 %__u to <8 x i1>
12403 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <2 x i32> <i32 0, i32 1>
12404 %4 = and <2 x i1> %2, %extract.i
12405 %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
12406 %6 = bitcast <4 x i1> %5 to i4
12407 ret i4 %6
12408 }
12410 define zeroext i4 @test_masked_vpcmpsgeq_v2i1_v4i1_mask_mem(i8 zeroext %__u, <2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
12411 ; VLX-LABEL: test_masked_vpcmpsgeq_v2i1_v4i1_mask_mem:
12412 ; VLX: # %bb.0: # %entry
12413 ; VLX-NEXT: kmovd %edi, %k1
12414 ; VLX-NEXT: vpcmpnltq (%rsi), %xmm0, %k0 {%k1}
12415 ; VLX-NEXT: kmovb %k0, %eax
12416 ; VLX-NEXT: retq
12417 ;
12418 ; NoVLX-LABEL: test_masked_vpcmpsgeq_v2i1_v4i1_mask_mem:
12419 ; NoVLX: # %bb.0: # %entry
12420 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
12421 ; NoVLX-NEXT: vmovdqa (%rsi), %xmm1
12422 ; NoVLX-NEXT: kmovw %edi, %k1
12423 ; NoVLX-NEXT: vpcmpnltq %zmm1, %zmm0, %k0 {%k1}
12424 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
12425 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
12426 ; NoVLX-NEXT: kmovw %k0, %eax
12427 ; NoVLX-NEXT: vzeroupper
12428 ; NoVLX-NEXT: retq
12429 entry:
12430 %0 = bitcast <2 x i64> %__a to <2 x i64>
12431 %load = load <2 x i64>, <2 x i64>* %__b
12432 %1 = bitcast <2 x i64> %load to <2 x i64>
12433 %2 = icmp sge <2 x i64> %0, %1
12434 %3 = bitcast i8 %__u to <8 x i1>
12435 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <2 x i32> <i32 0, i32 1>
12436 %4 = and <2 x i1> %2, %extract.i
12437 %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
12438 %6 = bitcast <4 x i1> %5 to i4
12439 ret i4 %6
12440 }
12443 define zeroext i4 @test_vpcmpsgeq_v2i1_v4i1_mask_mem_b(<2 x i64> %__a, i64* %__b) local_unnamed_addr {
12444 ; VLX-LABEL: test_vpcmpsgeq_v2i1_v4i1_mask_mem_b:
12445 ; VLX: # %bb.0: # %entry
12446 ; VLX-NEXT: vpcmpnltq (%rdi){1to2}, %xmm0, %k0
12447 ; VLX-NEXT: kmovb %k0, %eax
12448 ; VLX-NEXT: retq
12449 ;
12450 ; NoVLX-LABEL: test_vpcmpsgeq_v2i1_v4i1_mask_mem_b:
12451 ; NoVLX: # %bb.0: # %entry
12452 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
12453 ; NoVLX-NEXT: vpcmpnltq (%rdi){1to8}, %zmm0, %k0
12454 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
12455 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
12456 ; NoVLX-NEXT: kmovw %k0, %eax
12457 ; NoVLX-NEXT: vzeroupper
12458 ; NoVLX-NEXT: retq
12459 entry:
12460 %0 = bitcast <2 x i64> %__a to <2 x i64>
12461 %load = load i64, i64* %__b
12462 %vec = insertelement <2 x i64> undef, i64 %load, i32 0
12463 %1 = shufflevector <2 x i64> %vec, <2 x i64> undef, <2 x i32> <i32 0, i32 0>
12464 %2 = icmp sge <2 x i64> %0, %1
12465 %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
12466 %4 = bitcast <4 x i1> %3 to i4
12467 ret i4 %4
12468 }
12470 define zeroext i4 @test_masked_vpcmpsgeq_v2i1_v4i1_mask_mem_b(i8 zeroext %__u, <2 x i64> %__a, i64* %__b) local_unnamed_addr {
12471 ; VLX-LABEL: test_masked_vpcmpsgeq_v2i1_v4i1_mask_mem_b:
12472 ; VLX: # %bb.0: # %entry
12473 ; VLX-NEXT: kmovd %edi, %k1
12474 ; VLX-NEXT: vpcmpnltq (%rsi){1to2}, %xmm0, %k0 {%k1}
12475 ; VLX-NEXT: kmovb %k0, %eax
12476 ; VLX-NEXT: retq
12477 ;
12478 ; NoVLX-LABEL: test_masked_vpcmpsgeq_v2i1_v4i1_mask_mem_b:
12479 ; NoVLX: # %bb.0: # %entry
12480 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
12481 ; NoVLX-NEXT: kmovw %edi, %k1
12482 ; NoVLX-NEXT: vpcmpnltq (%rsi){1to8}, %zmm0, %k0 {%k1}
12483 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
12484 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
12485 ; NoVLX-NEXT: kmovw %k0, %eax
12486 ; NoVLX-NEXT: vzeroupper
12487 ; NoVLX-NEXT: retq
12488 entry:
12489 %0 = bitcast <2 x i64> %__a to <2 x i64>
12490 %load = load i64, i64* %__b
12491 %vec = insertelement <2 x i64> undef, i64 %load, i32 0
12492 %1 = shufflevector <2 x i64> %vec, <2 x i64> undef, <2 x i32> <i32 0, i32 0>
12493 %2 = icmp sge <2 x i64> %0, %1
12494 %3 = bitcast i8 %__u to <8 x i1>
12495 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <2 x i32> <i32 0, i32 1>
12496 %4 = and <2 x i1> %extract.i, %2
12497 %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
12498 %6 = bitcast <4 x i1> %5 to i4
12499 ret i4 %6
12500 }
12503 define zeroext i8 @test_vpcmpsgeq_v2i1_v8i1_mask(<2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
12504 ; VLX-LABEL: test_vpcmpsgeq_v2i1_v8i1_mask:
12505 ; VLX: # %bb.0: # %entry
12506 ; VLX-NEXT: vpcmpnltq %xmm1, %xmm0, %k0
12507 ; VLX-NEXT: kmovd %k0, %eax
12508 ; VLX-NEXT: # kill: def $al killed $al killed $eax
12509 ; VLX-NEXT: retq
12510 ;
12511 ; NoVLX-LABEL: test_vpcmpsgeq_v2i1_v8i1_mask:
12512 ; NoVLX: # %bb.0: # %entry
12513 ; NoVLX-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
12514 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
12515 ; NoVLX-NEXT: vpcmpnltq %zmm1, %zmm0, %k0
12516 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
12517 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
12518 ; NoVLX-NEXT: kmovw %k0, %eax
12519 ; NoVLX-NEXT: # kill: def $al killed $al killed $eax
12520 ; NoVLX-NEXT: vzeroupper
12521 ; NoVLX-NEXT: retq
12522 entry:
12523 %0 = bitcast <2 x i64> %__a to <2 x i64>
12524 %1 = bitcast <2 x i64> %__b to <2 x i64>
12525 %2 = icmp sge <2 x i64> %0, %1
12526 %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
12527 %4 = bitcast <8 x i1> %3 to i8
12528 ret i8 %4
12529 }
12531 define zeroext i8 @test_vpcmpsgeq_v2i1_v8i1_mask_mem(<2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
12532 ; VLX-LABEL: test_vpcmpsgeq_v2i1_v8i1_mask_mem:
12533 ; VLX: # %bb.0: # %entry
12534 ; VLX-NEXT: vpcmpnltq (%rdi), %xmm0, %k0
12535 ; VLX-NEXT: kmovd %k0, %eax
12536 ; VLX-NEXT: # kill: def $al killed $al killed $eax
12537 ; VLX-NEXT: retq
12538 ;
12539 ; NoVLX-LABEL: test_vpcmpsgeq_v2i1_v8i1_mask_mem:
12540 ; NoVLX: # %bb.0: # %entry
12541 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
12542 ; NoVLX-NEXT: vmovdqa (%rdi), %xmm1
12543 ; NoVLX-NEXT: vpcmpnltq %zmm1, %zmm0, %k0
12544 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
12545 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
12546 ; NoVLX-NEXT: kmovw %k0, %eax
12547 ; NoVLX-NEXT: # kill: def $al killed $al killed $eax
12548 ; NoVLX-NEXT: vzeroupper
12549 ; NoVLX-NEXT: retq
12550 entry:
12551 %0 = bitcast <2 x i64> %__a to <2 x i64>
12552 %load = load <2 x i64>, <2 x i64>* %__b
12553 %1 = bitcast <2 x i64> %load to <2 x i64>
12554 %2 = icmp sge <2 x i64> %0, %1
12555 %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
12556 %4 = bitcast <8 x i1> %3 to i8
12557 ret i8 %4
12558 }
12560 define zeroext i8 @test_masked_vpcmpsgeq_v2i1_v8i1_mask(i8 zeroext %__u, <2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
12561 ; VLX-LABEL: test_masked_vpcmpsgeq_v2i1_v8i1_mask:
12562 ; VLX: # %bb.0: # %entry
12563 ; VLX-NEXT: kmovd %edi, %k1
12564 ; VLX-NEXT: vpcmpnltq %xmm1, %xmm0, %k0 {%k1}
12565 ; VLX-NEXT: kmovd %k0, %eax
12566 ; VLX-NEXT: # kill: def $al killed $al killed $eax
12567 ; VLX-NEXT: retq
12568 ;
12569 ; NoVLX-LABEL: test_masked_vpcmpsgeq_v2i1_v8i1_mask:
12570 ; NoVLX: # %bb.0: # %entry
12571 ; NoVLX-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
12572 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
12573 ; NoVLX-NEXT: kmovw %edi, %k1
12574 ; NoVLX-NEXT: vpcmpnltq %zmm1, %zmm0, %k0 {%k1}
12575 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
12576 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
12577 ; NoVLX-NEXT: kmovw %k0, %eax
12578 ; NoVLX-NEXT: # kill: def $al killed $al killed $eax
12579 ; NoVLX-NEXT: vzeroupper
12580 ; NoVLX-NEXT: retq
12581 entry:
12582 %0 = bitcast <2 x i64> %__a to <2 x i64>
12583 %1 = bitcast <2 x i64> %__b to <2 x i64>
12584 %2 = icmp sge <2 x i64> %0, %1
12585 %3 = bitcast i8 %__u to <8 x i1>
12586 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <2 x i32> <i32 0, i32 1>
12587 %4 = and <2 x i1> %2, %extract.i
12588 %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
12589 %6 = bitcast <8 x i1> %5 to i8
12590 ret i8 %6
12591 }
12593 define zeroext i8 @test_masked_vpcmpsgeq_v2i1_v8i1_mask_mem(i8 zeroext %__u, <2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
12594 ; VLX-LABEL: test_masked_vpcmpsgeq_v2i1_v8i1_mask_mem:
12595 ; VLX: # %bb.0: # %entry
12596 ; VLX-NEXT: kmovd %edi, %k1
12597 ; VLX-NEXT: vpcmpnltq (%rsi), %xmm0, %k0 {%k1}
12598 ; VLX-NEXT: kmovd %k0, %eax
12599 ; VLX-NEXT: # kill: def $al killed $al killed $eax
12600 ; VLX-NEXT: retq
12601 ;
12602 ; NoVLX-LABEL: test_masked_vpcmpsgeq_v2i1_v8i1_mask_mem:
12603 ; NoVLX: # %bb.0: # %entry
12604 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
12605 ; NoVLX-NEXT: vmovdqa (%rsi), %xmm1
12606 ; NoVLX-NEXT: kmovw %edi, %k1
12607 ; NoVLX-NEXT: vpcmpnltq %zmm1, %zmm0, %k0 {%k1}
12608 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
12609 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
12610 ; NoVLX-NEXT: kmovw %k0, %eax
12611 ; NoVLX-NEXT: # kill: def $al killed $al killed $eax
12612 ; NoVLX-NEXT: vzeroupper
12613 ; NoVLX-NEXT: retq
12614 entry:
12615 %0 = bitcast <2 x i64> %__a to <2 x i64>
12616 %load = load <2 x i64>, <2 x i64>* %__b
12617 %1 = bitcast <2 x i64> %load to <2 x i64>
12618 %2 = icmp sge <2 x i64> %0, %1
12619 %3 = bitcast i8 %__u to <8 x i1>
12620 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <2 x i32> <i32 0, i32 1>
12621 %4 = and <2 x i1> %2, %extract.i
12622 %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
12623 %6 = bitcast <8 x i1> %5 to i8
12624 ret i8 %6
12625 }
12628 define zeroext i8 @test_vpcmpsgeq_v2i1_v8i1_mask_mem_b(<2 x i64> %__a, i64* %__b) local_unnamed_addr {
12629 ; VLX-LABEL: test_vpcmpsgeq_v2i1_v8i1_mask_mem_b:
12630 ; VLX: # %bb.0: # %entry
12631 ; VLX-NEXT: vpcmpnltq (%rdi){1to2}, %xmm0, %k0
12632 ; VLX-NEXT: kmovd %k0, %eax
12633 ; VLX-NEXT: # kill: def $al killed $al killed $eax
12634 ; VLX-NEXT: retq
12635 ;
12636 ; NoVLX-LABEL: test_vpcmpsgeq_v2i1_v8i1_mask_mem_b:
12637 ; NoVLX: # %bb.0: # %entry
12638 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
12639 ; NoVLX-NEXT: vpcmpnltq (%rdi){1to8}, %zmm0, %k0
12640 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
12641 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
12642 ; NoVLX-NEXT: kmovw %k0, %eax
12643 ; NoVLX-NEXT: # kill: def $al killed $al killed $eax
12644 ; NoVLX-NEXT: vzeroupper
12645 ; NoVLX-NEXT: retq
12646 entry:
12647 %0 = bitcast <2 x i64> %__a to <2 x i64>
12648 %load = load i64, i64* %__b
12649 %vec = insertelement <2 x i64> undef, i64 %load, i32 0
12650 %1 = shufflevector <2 x i64> %vec, <2 x i64> undef, <2 x i32> <i32 0, i32 0>
12651 %2 = icmp sge <2 x i64> %0, %1
12652 %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
12653 %4 = bitcast <8 x i1> %3 to i8
12654 ret i8 %4
12655 }
12657 define zeroext i8 @test_masked_vpcmpsgeq_v2i1_v8i1_mask_mem_b(i8 zeroext %__u, <2 x i64> %__a, i64* %__b) local_unnamed_addr {
12658 ; VLX-LABEL: test_masked_vpcmpsgeq_v2i1_v8i1_mask_mem_b:
12659 ; VLX: # %bb.0: # %entry
12660 ; VLX-NEXT: kmovd %edi, %k1
12661 ; VLX-NEXT: vpcmpnltq (%rsi){1to2}, %xmm0, %k0 {%k1}
12662 ; VLX-NEXT: kmovd %k0, %eax
12663 ; VLX-NEXT: # kill: def $al killed $al killed $eax
12664 ; VLX-NEXT: retq
12665 ;
12666 ; NoVLX-LABEL: test_masked_vpcmpsgeq_v2i1_v8i1_mask_mem_b:
12667 ; NoVLX: # %bb.0: # %entry
12668 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
12669 ; NoVLX-NEXT: kmovw %edi, %k1
12670 ; NoVLX-NEXT: vpcmpnltq (%rsi){1to8}, %zmm0, %k0 {%k1}
12671 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
12672 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
12673 ; NoVLX-NEXT: kmovw %k0, %eax
12674 ; NoVLX-NEXT: # kill: def $al killed $al killed $eax
12675 ; NoVLX-NEXT: vzeroupper
12676 ; NoVLX-NEXT: retq
12677 entry:
12678 %0 = bitcast <2 x i64> %__a to <2 x i64>
12679 %load = load i64, i64* %__b
12680 %vec = insertelement <2 x i64> undef, i64 %load, i32 0
12681 %1 = shufflevector <2 x i64> %vec, <2 x i64> undef, <2 x i32> <i32 0, i32 0>
12682 %2 = icmp sge <2 x i64> %0, %1
12683 %3 = bitcast i8 %__u to <8 x i1>
12684 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <2 x i32> <i32 0, i32 1>
12685 %4 = and <2 x i1> %extract.i, %2
12686 %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
12687 %6 = bitcast <8 x i1> %5 to i8
12688 ret i8 %6
12689 }
12692 define zeroext i16 @test_vpcmpsgeq_v2i1_v16i1_mask(<2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
12693 ; VLX-LABEL: test_vpcmpsgeq_v2i1_v16i1_mask:
12694 ; VLX: # %bb.0: # %entry
12695 ; VLX-NEXT: vpcmpnltq %xmm1, %xmm0, %k0
12696 ; VLX-NEXT: kmovd %k0, %eax
12697 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
12698 ; VLX-NEXT: retq
12699 ;
12700 ; NoVLX-LABEL: test_vpcmpsgeq_v2i1_v16i1_mask:
12701 ; NoVLX: # %bb.0: # %entry
12702 ; NoVLX-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
12703 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
12704 ; NoVLX-NEXT: vpcmpnltq %zmm1, %zmm0, %k0
12705 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
12706 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
12707 ; NoVLX-NEXT: kmovw %k0, %eax
12708 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
12709 ; NoVLX-NEXT: vzeroupper
12710 ; NoVLX-NEXT: retq
12711 entry:
12712 %0 = bitcast <2 x i64> %__a to <2 x i64>
12713 %1 = bitcast <2 x i64> %__b to <2 x i64>
12714 %2 = icmp sge <2 x i64> %0, %1
12715 %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
12716 %4 = bitcast <16 x i1> %3 to i16
12717 ret i16 %4
12718 }
12720 define zeroext i16 @test_vpcmpsgeq_v2i1_v16i1_mask_mem(<2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
12721 ; VLX-LABEL: test_vpcmpsgeq_v2i1_v16i1_mask_mem:
12722 ; VLX: # %bb.0: # %entry
12723 ; VLX-NEXT: vpcmpnltq (%rdi), %xmm0, %k0
12724 ; VLX-NEXT: kmovd %k0, %eax
12725 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
12726 ; VLX-NEXT: retq
12727 ;
12728 ; NoVLX-LABEL: test_vpcmpsgeq_v2i1_v16i1_mask_mem:
12729 ; NoVLX: # %bb.0: # %entry
12730 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
12731 ; NoVLX-NEXT: vmovdqa (%rdi), %xmm1
12732 ; NoVLX-NEXT: vpcmpnltq %zmm1, %zmm0, %k0
12733 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
12734 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
12735 ; NoVLX-NEXT: kmovw %k0, %eax
12736 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
12737 ; NoVLX-NEXT: vzeroupper
12738 ; NoVLX-NEXT: retq
12739 entry:
12740 %0 = bitcast <2 x i64> %__a to <2 x i64>
12741 %load = load <2 x i64>, <2 x i64>* %__b
12742 %1 = bitcast <2 x i64> %load to <2 x i64>
12743 %2 = icmp sge <2 x i64> %0, %1
12744 %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
12745 %4 = bitcast <16 x i1> %3 to i16
12746 ret i16 %4
12747 }
12749 define zeroext i16 @test_masked_vpcmpsgeq_v2i1_v16i1_mask(i8 zeroext %__u, <2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
12750 ; VLX-LABEL: test_masked_vpcmpsgeq_v2i1_v16i1_mask:
12751 ; VLX: # %bb.0: # %entry
12752 ; VLX-NEXT: kmovd %edi, %k1
12753 ; VLX-NEXT: vpcmpnltq %xmm1, %xmm0, %k0 {%k1}
12754 ; VLX-NEXT: kmovd %k0, %eax
12755 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
12756 ; VLX-NEXT: retq
12757 ;
12758 ; NoVLX-LABEL: test_masked_vpcmpsgeq_v2i1_v16i1_mask:
12759 ; NoVLX: # %bb.0: # %entry
12760 ; NoVLX-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
12761 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
12762 ; NoVLX-NEXT: kmovw %edi, %k1
12763 ; NoVLX-NEXT: vpcmpnltq %zmm1, %zmm0, %k0 {%k1}
12764 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
12765 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
12766 ; NoVLX-NEXT: kmovw %k0, %eax
12767 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
12768 ; NoVLX-NEXT: vzeroupper
12769 ; NoVLX-NEXT: retq
12770 entry:
12771 %0 = bitcast <2 x i64> %__a to <2 x i64>
12772 %1 = bitcast <2 x i64> %__b to <2 x i64>
12773 %2 = icmp sge <2 x i64> %0, %1
12774 %3 = bitcast i8 %__u to <8 x i1>
12775 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <2 x i32> <i32 0, i32 1>
12776 %4 = and <2 x i1> %2, %extract.i
12777 %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
12778 %6 = bitcast <16 x i1> %5 to i16
12779 ret i16 %6
12780 }
12782 define zeroext i16 @test_masked_vpcmpsgeq_v2i1_v16i1_mask_mem(i8 zeroext %__u, <2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
12783 ; VLX-LABEL: test_masked_vpcmpsgeq_v2i1_v16i1_mask_mem:
12784 ; VLX: # %bb.0: # %entry
12785 ; VLX-NEXT: kmovd %edi, %k1
12786 ; VLX-NEXT: vpcmpnltq (%rsi), %xmm0, %k0 {%k1}
12787 ; VLX-NEXT: kmovd %k0, %eax
12788 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
12789 ; VLX-NEXT: retq
12790 ;
12791 ; NoVLX-LABEL: test_masked_vpcmpsgeq_v2i1_v16i1_mask_mem:
12792 ; NoVLX: # %bb.0: # %entry
12793 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
12794 ; NoVLX-NEXT: vmovdqa (%rsi), %xmm1
12795 ; NoVLX-NEXT: kmovw %edi, %k1
12796 ; NoVLX-NEXT: vpcmpnltq %zmm1, %zmm0, %k0 {%k1}
12797 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
12798 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
12799 ; NoVLX-NEXT: kmovw %k0, %eax
12800 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
12801 ; NoVLX-NEXT: vzeroupper
12802 ; NoVLX-NEXT: retq
12803 entry:
12804 %0 = bitcast <2 x i64> %__a to <2 x i64>
12805 %load = load <2 x i64>, <2 x i64>* %__b
12806 %1 = bitcast <2 x i64> %load to <2 x i64>
12807 %2 = icmp sge <2 x i64> %0, %1
12808 %3 = bitcast i8 %__u to <8 x i1>
12809 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <2 x i32> <i32 0, i32 1>
12810 %4 = and <2 x i1> %2, %extract.i
12811 %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
12812 %6 = bitcast <16 x i1> %5 to i16
12813 ret i16 %6
12814 }
12817 define zeroext i16 @test_vpcmpsgeq_v2i1_v16i1_mask_mem_b(<2 x i64> %__a, i64* %__b) local_unnamed_addr {
12818 ; VLX-LABEL: test_vpcmpsgeq_v2i1_v16i1_mask_mem_b:
12819 ; VLX: # %bb.0: # %entry
12820 ; VLX-NEXT: vpcmpnltq (%rdi){1to2}, %xmm0, %k0
12821 ; VLX-NEXT: kmovd %k0, %eax
12822 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
12823 ; VLX-NEXT: retq
12824 ;
12825 ; NoVLX-LABEL: test_vpcmpsgeq_v2i1_v16i1_mask_mem_b:
12826 ; NoVLX: # %bb.0: # %entry
12827 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
12828 ; NoVLX-NEXT: vpcmpnltq (%rdi){1to8}, %zmm0, %k0
12829 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
12830 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
12831 ; NoVLX-NEXT: kmovw %k0, %eax
12832 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
12833 ; NoVLX-NEXT: vzeroupper
12834 ; NoVLX-NEXT: retq
12835 entry:
12836 %0 = bitcast <2 x i64> %__a to <2 x i64>
12837 %load = load i64, i64* %__b
12838 %vec = insertelement <2 x i64> undef, i64 %load, i32 0
12839 %1 = shufflevector <2 x i64> %vec, <2 x i64> undef, <2 x i32> <i32 0, i32 0>
12840 %2 = icmp sge <2 x i64> %0, %1
12841 %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
12842 %4 = bitcast <16 x i1> %3 to i16
12843 ret i16 %4
12844 }
12846 define zeroext i16 @test_masked_vpcmpsgeq_v2i1_v16i1_mask_mem_b(i8 zeroext %__u, <2 x i64> %__a, i64* %__b) local_unnamed_addr {
12847 ; VLX-LABEL: test_masked_vpcmpsgeq_v2i1_v16i1_mask_mem_b:
12848 ; VLX: # %bb.0: # %entry
12849 ; VLX-NEXT: kmovd %edi, %k1
12850 ; VLX-NEXT: vpcmpnltq (%rsi){1to2}, %xmm0, %k0 {%k1}
12851 ; VLX-NEXT: kmovd %k0, %eax
12852 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
12853 ; VLX-NEXT: retq
12854 ;
12855 ; NoVLX-LABEL: test_masked_vpcmpsgeq_v2i1_v16i1_mask_mem_b:
12856 ; NoVLX: # %bb.0: # %entry
12857 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
12858 ; NoVLX-NEXT: kmovw %edi, %k1
12859 ; NoVLX-NEXT: vpcmpnltq (%rsi){1to8}, %zmm0, %k0 {%k1}
12860 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
12861 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
12862 ; NoVLX-NEXT: kmovw %k0, %eax
12863 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
12864 ; NoVLX-NEXT: vzeroupper
12865 ; NoVLX-NEXT: retq
12866 entry:
12867 %0 = bitcast <2 x i64> %__a to <2 x i64>
12868 %load = load i64, i64* %__b
12869 %vec = insertelement <2 x i64> undef, i64 %load, i32 0
12870 %1 = shufflevector <2 x i64> %vec, <2 x i64> undef, <2 x i32> <i32 0, i32 0>
12871 %2 = icmp sge <2 x i64> %0, %1
12872 %3 = bitcast i8 %__u to <8 x i1>
12873 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <2 x i32> <i32 0, i32 1>
12874 %4 = and <2 x i1> %extract.i, %2
12875 %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
12876 %6 = bitcast <16 x i1> %5 to i16
12877 ret i16 %6
12878 }
12881 define zeroext i32 @test_vpcmpsgeq_v2i1_v32i1_mask(<2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
12882 ; VLX-LABEL: test_vpcmpsgeq_v2i1_v32i1_mask:
12883 ; VLX: # %bb.0: # %entry
12884 ; VLX-NEXT: vpcmpnltq %xmm1, %xmm0, %k0
12885 ; VLX-NEXT: kmovd %k0, %eax
12886 ; VLX-NEXT: retq
12887 ;
12888 ; NoVLX-LABEL: test_vpcmpsgeq_v2i1_v32i1_mask:
12889 ; NoVLX: # %bb.0: # %entry
12890 ; NoVLX-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
12891 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
12892 ; NoVLX-NEXT: vpcmpnltq %zmm1, %zmm0, %k0
12893 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
12894 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
12895 ; NoVLX-NEXT: kmovw %k0, %eax
12896 ; NoVLX-NEXT: vzeroupper
12897 ; NoVLX-NEXT: retq
12898 entry:
12899 %0 = bitcast <2 x i64> %__a to <2 x i64>
12900 %1 = bitcast <2 x i64> %__b to <2 x i64>
12901 %2 = icmp sge <2 x i64> %0, %1
12902 %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
12903 %4 = bitcast <32 x i1> %3 to i32
12904 ret i32 %4
12905 }
12907 define zeroext i32 @test_vpcmpsgeq_v2i1_v32i1_mask_mem(<2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
12908 ; VLX-LABEL: test_vpcmpsgeq_v2i1_v32i1_mask_mem:
12909 ; VLX: # %bb.0: # %entry
12910 ; VLX-NEXT: vpcmpnltq (%rdi), %xmm0, %k0
12911 ; VLX-NEXT: kmovd %k0, %eax
12912 ; VLX-NEXT: retq
12913 ;
12914 ; NoVLX-LABEL: test_vpcmpsgeq_v2i1_v32i1_mask_mem:
12915 ; NoVLX: # %bb.0: # %entry
12916 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
12917 ; NoVLX-NEXT: vmovdqa (%rdi), %xmm1
12918 ; NoVLX-NEXT: vpcmpnltq %zmm1, %zmm0, %k0
12919 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
12920 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
12921 ; NoVLX-NEXT: kmovw %k0, %eax
12922 ; NoVLX-NEXT: vzeroupper
12923 ; NoVLX-NEXT: retq
12924 entry:
12925 %0 = bitcast <2 x i64> %__a to <2 x i64>
12926 %load = load <2 x i64>, <2 x i64>* %__b
12927 %1 = bitcast <2 x i64> %load to <2 x i64>
12928 %2 = icmp sge <2 x i64> %0, %1
12929 %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
12930 %4 = bitcast <32 x i1> %3 to i32
12931 ret i32 %4
12932 }
12934 define zeroext i32 @test_masked_vpcmpsgeq_v2i1_v32i1_mask(i8 zeroext %__u, <2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
12935 ; VLX-LABEL: test_masked_vpcmpsgeq_v2i1_v32i1_mask:
12936 ; VLX: # %bb.0: # %entry
12937 ; VLX-NEXT: kmovd %edi, %k1
12938 ; VLX-NEXT: vpcmpnltq %xmm1, %xmm0, %k0 {%k1}
12939 ; VLX-NEXT: kmovd %k0, %eax
12940 ; VLX-NEXT: retq
12941 ;
12942 ; NoVLX-LABEL: test_masked_vpcmpsgeq_v2i1_v32i1_mask:
12943 ; NoVLX: # %bb.0: # %entry
12944 ; NoVLX-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
12945 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
12946 ; NoVLX-NEXT: kmovw %edi, %k1
12947 ; NoVLX-NEXT: vpcmpnltq %zmm1, %zmm0, %k0 {%k1}
12948 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
12949 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
12950 ; NoVLX-NEXT: kmovw %k0, %eax
12951 ; NoVLX-NEXT: vzeroupper
12952 ; NoVLX-NEXT: retq
12953 entry:
12954 %0 = bitcast <2 x i64> %__a to <2 x i64>
12955 %1 = bitcast <2 x i64> %__b to <2 x i64>
12956 %2 = icmp sge <2 x i64> %0, %1
12957 %3 = bitcast i8 %__u to <8 x i1>
12958 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <2 x i32> <i32 0, i32 1>
12959 %4 = and <2 x i1> %2, %extract.i
12960 %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
12961 %6 = bitcast <32 x i1> %5 to i32
12962 ret i32 %6
12963 }
12965 define zeroext i32 @test_masked_vpcmpsgeq_v2i1_v32i1_mask_mem(i8 zeroext %__u, <2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
12966 ; VLX-LABEL: test_masked_vpcmpsgeq_v2i1_v32i1_mask_mem:
12967 ; VLX: # %bb.0: # %entry
12968 ; VLX-NEXT: kmovd %edi, %k1
12969 ; VLX-NEXT: vpcmpnltq (%rsi), %xmm0, %k0 {%k1}
12970 ; VLX-NEXT: kmovd %k0, %eax
12971 ; VLX-NEXT: retq
12972 ;
12973 ; NoVLX-LABEL: test_masked_vpcmpsgeq_v2i1_v32i1_mask_mem:
12974 ; NoVLX: # %bb.0: # %entry
12975 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
12976 ; NoVLX-NEXT: vmovdqa (%rsi), %xmm1
12977 ; NoVLX-NEXT: kmovw %edi, %k1
12978 ; NoVLX-NEXT: vpcmpnltq %zmm1, %zmm0, %k0 {%k1}
12979 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
12980 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
12981 ; NoVLX-NEXT: kmovw %k0, %eax
12982 ; NoVLX-NEXT: vzeroupper
12983 ; NoVLX-NEXT: retq
12984 entry:
12985 %0 = bitcast <2 x i64> %__a to <2 x i64>
12986 %load = load <2 x i64>, <2 x i64>* %__b
12987 %1 = bitcast <2 x i64> %load to <2 x i64>
12988 %2 = icmp sge <2 x i64> %0, %1
12989 %3 = bitcast i8 %__u to <8 x i1>
12990 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <2 x i32> <i32 0, i32 1>
12991 %4 = and <2 x i1> %2, %extract.i
12992 %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
12993 %6 = bitcast <32 x i1> %5 to i32
12994 ret i32 %6
12995 }
12998 define zeroext i32 @test_vpcmpsgeq_v2i1_v32i1_mask_mem_b(<2 x i64> %__a, i64* %__b) local_unnamed_addr {
12999 ; VLX-LABEL: test_vpcmpsgeq_v2i1_v32i1_mask_mem_b:
13000 ; VLX: # %bb.0: # %entry
13001 ; VLX-NEXT: vpcmpnltq (%rdi){1to2}, %xmm0, %k0
13002 ; VLX-NEXT: kmovd %k0, %eax
13003 ; VLX-NEXT: retq
13004 ;
13005 ; NoVLX-LABEL: test_vpcmpsgeq_v2i1_v32i1_mask_mem_b:
13006 ; NoVLX: # %bb.0: # %entry
13007 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
13008 ; NoVLX-NEXT: vpcmpnltq (%rdi){1to8}, %zmm0, %k0
13009 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
13010 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
13011 ; NoVLX-NEXT: kmovw %k0, %eax
13012 ; NoVLX-NEXT: vzeroupper
13013 ; NoVLX-NEXT: retq
13014 entry:
13015 %0 = bitcast <2 x i64> %__a to <2 x i64>
13016 %load = load i64, i64* %__b
13017 %vec = insertelement <2 x i64> undef, i64 %load, i32 0
13018 %1 = shufflevector <2 x i64> %vec, <2 x i64> undef, <2 x i32> <i32 0, i32 0>
13019 %2 = icmp sge <2 x i64> %0, %1
13020 %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
13021 %4 = bitcast <32 x i1> %3 to i32
13022 ret i32 %4
13023 }
13025 define zeroext i32 @test_masked_vpcmpsgeq_v2i1_v32i1_mask_mem_b(i8 zeroext %__u, <2 x i64> %__a, i64* %__b) local_unnamed_addr {
13026 ; VLX-LABEL: test_masked_vpcmpsgeq_v2i1_v32i1_mask_mem_b:
13027 ; VLX: # %bb.0: # %entry
13028 ; VLX-NEXT: kmovd %edi, %k1
13029 ; VLX-NEXT: vpcmpnltq (%rsi){1to2}, %xmm0, %k0 {%k1}
13030 ; VLX-NEXT: kmovd %k0, %eax
13031 ; VLX-NEXT: retq
13032 ;
13033 ; NoVLX-LABEL: test_masked_vpcmpsgeq_v2i1_v32i1_mask_mem_b:
13034 ; NoVLX: # %bb.0: # %entry
13035 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
13036 ; NoVLX-NEXT: kmovw %edi, %k1
13037 ; NoVLX-NEXT: vpcmpnltq (%rsi){1to8}, %zmm0, %k0 {%k1}
13038 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
13039 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
13040 ; NoVLX-NEXT: kmovw %k0, %eax
13041 ; NoVLX-NEXT: vzeroupper
13042 ; NoVLX-NEXT: retq
13043 entry:
13044 %0 = bitcast <2 x i64> %__a to <2 x i64>
13045 %load = load i64, i64* %__b
13046 %vec = insertelement <2 x i64> undef, i64 %load, i32 0
13047 %1 = shufflevector <2 x i64> %vec, <2 x i64> undef, <2 x i32> <i32 0, i32 0>
13048 %2 = icmp sge <2 x i64> %0, %1
13049 %3 = bitcast i8 %__u to <8 x i1>
13050 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <2 x i32> <i32 0, i32 1>
13051 %4 = and <2 x i1> %extract.i, %2
13052 %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
13053 %6 = bitcast <32 x i1> %5 to i32
13054 ret i32 %6
13055 }
13058 define zeroext i64 @test_vpcmpsgeq_v2i1_v64i1_mask(<2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
13059 ; VLX-LABEL: test_vpcmpsgeq_v2i1_v64i1_mask:
13060 ; VLX: # %bb.0: # %entry
13061 ; VLX-NEXT: vpcmpnltq %xmm1, %xmm0, %k0
13062 ; VLX-NEXT: kmovq %k0, %rax
13063 ; VLX-NEXT: retq
13064 ;
13065 ; NoVLX-LABEL: test_vpcmpsgeq_v2i1_v64i1_mask:
13066 ; NoVLX: # %bb.0: # %entry
13067 ; NoVLX-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
13068 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
13069 ; NoVLX-NEXT: vpcmpnltq %zmm1, %zmm0, %k0
13070 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
13071 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
13072 ; NoVLX-NEXT: kmovw %k0, %eax
13073 ; NoVLX-NEXT: vzeroupper
13074 ; NoVLX-NEXT: retq
13075 entry:
13076 %0 = bitcast <2 x i64> %__a to <2 x i64>
13077 %1 = bitcast <2 x i64> %__b to <2 x i64>
13078 %2 = icmp sge <2 x i64> %0, %1
13079 %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
13080 %4 = bitcast <64 x i1> %3 to i64
13081 ret i64 %4
13082 }
13084 define zeroext i64 @test_vpcmpsgeq_v2i1_v64i1_mask_mem(<2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
13085 ; VLX-LABEL: test_vpcmpsgeq_v2i1_v64i1_mask_mem:
13086 ; VLX: # %bb.0: # %entry
13087 ; VLX-NEXT: vpcmpnltq (%rdi), %xmm0, %k0
13088 ; VLX-NEXT: kmovq %k0, %rax
13089 ; VLX-NEXT: retq
13090 ;
13091 ; NoVLX-LABEL: test_vpcmpsgeq_v2i1_v64i1_mask_mem:
13092 ; NoVLX: # %bb.0: # %entry
13093 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
13094 ; NoVLX-NEXT: vmovdqa (%rdi), %xmm1
13095 ; NoVLX-NEXT: vpcmpnltq %zmm1, %zmm0, %k0
13096 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
13097 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
13098 ; NoVLX-NEXT: kmovw %k0, %eax
13099 ; NoVLX-NEXT: vzeroupper
13100 ; NoVLX-NEXT: retq
13101 entry:
13102 %0 = bitcast <2 x i64> %__a to <2 x i64>
13103 %load = load <2 x i64>, <2 x i64>* %__b
13104 %1 = bitcast <2 x i64> %load to <2 x i64>
13105 %2 = icmp sge <2 x i64> %0, %1
13106 %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
13107 %4 = bitcast <64 x i1> %3 to i64
13108 ret i64 %4
13109 }
13111 define zeroext i64 @test_masked_vpcmpsgeq_v2i1_v64i1_mask(i8 zeroext %__u, <2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
13112 ; VLX-LABEL: test_masked_vpcmpsgeq_v2i1_v64i1_mask:
13113 ; VLX: # %bb.0: # %entry
13114 ; VLX-NEXT: kmovd %edi, %k1
13115 ; VLX-NEXT: vpcmpnltq %xmm1, %xmm0, %k0 {%k1}
13116 ; VLX-NEXT: kmovq %k0, %rax
13117 ; VLX-NEXT: retq
13118 ;
13119 ; NoVLX-LABEL: test_masked_vpcmpsgeq_v2i1_v64i1_mask:
13120 ; NoVLX: # %bb.0: # %entry
13121 ; NoVLX-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
13122 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
13123 ; NoVLX-NEXT: kmovw %edi, %k1
13124 ; NoVLX-NEXT: vpcmpnltq %zmm1, %zmm0, %k0 {%k1}
13125 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
13126 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
13127 ; NoVLX-NEXT: kmovw %k0, %eax
13128 ; NoVLX-NEXT: vzeroupper
13129 ; NoVLX-NEXT: retq
13130 entry:
13131 %0 = bitcast <2 x i64> %__a to <2 x i64>
13132 %1 = bitcast <2 x i64> %__b to <2 x i64>
13133 %2 = icmp sge <2 x i64> %0, %1
13134 %3 = bitcast i8 %__u to <8 x i1>
13135 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <2 x i32> <i32 0, i32 1>
13136 %4 = and <2 x i1> %2, %extract.i
13137 %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
13138 %6 = bitcast <64 x i1> %5 to i64
13139 ret i64 %6
13140 }
13142 define zeroext i64 @test_masked_vpcmpsgeq_v2i1_v64i1_mask_mem(i8 zeroext %__u, <2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
13143 ; VLX-LABEL: test_masked_vpcmpsgeq_v2i1_v64i1_mask_mem:
13144 ; VLX: # %bb.0: # %entry
13145 ; VLX-NEXT: kmovd %edi, %k1
13146 ; VLX-NEXT: vpcmpnltq (%rsi), %xmm0, %k0 {%k1}
13147 ; VLX-NEXT: kmovq %k0, %rax
13148 ; VLX-NEXT: retq
13149 ;
13150 ; NoVLX-LABEL: test_masked_vpcmpsgeq_v2i1_v64i1_mask_mem:
13151 ; NoVLX: # %bb.0: # %entry
13152 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
13153 ; NoVLX-NEXT: vmovdqa (%rsi), %xmm1
13154 ; NoVLX-NEXT: kmovw %edi, %k1
13155 ; NoVLX-NEXT: vpcmpnltq %zmm1, %zmm0, %k0 {%k1}
13156 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
13157 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
13158 ; NoVLX-NEXT: kmovw %k0, %eax
13159 ; NoVLX-NEXT: vzeroupper
13160 ; NoVLX-NEXT: retq
13161 entry:
13162 %0 = bitcast <2 x i64> %__a to <2 x i64>
13163 %load = load <2 x i64>, <2 x i64>* %__b
13164 %1 = bitcast <2 x i64> %load to <2 x i64>
13165 %2 = icmp sge <2 x i64> %0, %1
13166 %3 = bitcast i8 %__u to <8 x i1>
13167 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <2 x i32> <i32 0, i32 1>
13168 %4 = and <2 x i1> %2, %extract.i
13169 %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
13170 %6 = bitcast <64 x i1> %5 to i64
13171 ret i64 %6
13172 }
13175 define zeroext i64 @test_vpcmpsgeq_v2i1_v64i1_mask_mem_b(<2 x i64> %__a, i64* %__b) local_unnamed_addr {
13176 ; VLX-LABEL: test_vpcmpsgeq_v2i1_v64i1_mask_mem_b:
13177 ; VLX: # %bb.0: # %entry
13178 ; VLX-NEXT: vpcmpnltq (%rdi){1to2}, %xmm0, %k0
13179 ; VLX-NEXT: kmovq %k0, %rax
13180 ; VLX-NEXT: retq
13181 ;
13182 ; NoVLX-LABEL: test_vpcmpsgeq_v2i1_v64i1_mask_mem_b:
13183 ; NoVLX: # %bb.0: # %entry
13184 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
13185 ; NoVLX-NEXT: vpcmpnltq (%rdi){1to8}, %zmm0, %k0
13186 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
13187 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
13188 ; NoVLX-NEXT: kmovw %k0, %eax
13189 ; NoVLX-NEXT: vzeroupper
13190 ; NoVLX-NEXT: retq
13191 entry:
13192 %0 = bitcast <2 x i64> %__a to <2 x i64>
13193 %load = load i64, i64* %__b
13194 %vec = insertelement <2 x i64> undef, i64 %load, i32 0
13195 %1 = shufflevector <2 x i64> %vec, <2 x i64> undef, <2 x i32> <i32 0, i32 0>
13196 %2 = icmp sge <2 x i64> %0, %1
13197 %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
13198 %4 = bitcast <64 x i1> %3 to i64
13199 ret i64 %4
13200 }
13202 define zeroext i64 @test_masked_vpcmpsgeq_v2i1_v64i1_mask_mem_b(i8 zeroext %__u, <2 x i64> %__a, i64* %__b) local_unnamed_addr {
13203 ; VLX-LABEL: test_masked_vpcmpsgeq_v2i1_v64i1_mask_mem_b:
13204 ; VLX: # %bb.0: # %entry
13205 ; VLX-NEXT: kmovd %edi, %k1
13206 ; VLX-NEXT: vpcmpnltq (%rsi){1to2}, %xmm0, %k0 {%k1}
13207 ; VLX-NEXT: kmovq %k0, %rax
13208 ; VLX-NEXT: retq
13209 ;
13210 ; NoVLX-LABEL: test_masked_vpcmpsgeq_v2i1_v64i1_mask_mem_b:
13211 ; NoVLX: # %bb.0: # %entry
13212 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
13213 ; NoVLX-NEXT: kmovw %edi, %k1
13214 ; NoVLX-NEXT: vpcmpnltq (%rsi){1to8}, %zmm0, %k0 {%k1}
13215 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
13216 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
13217 ; NoVLX-NEXT: kmovw %k0, %eax
13218 ; NoVLX-NEXT: vzeroupper
13219 ; NoVLX-NEXT: retq
13220 entry:
13221 %0 = bitcast <2 x i64> %__a to <2 x i64>
13222 %load = load i64, i64* %__b
13223 %vec = insertelement <2 x i64> undef, i64 %load, i32 0
13224 %1 = shufflevector <2 x i64> %vec, <2 x i64> undef, <2 x i32> <i32 0, i32 0>
13225 %2 = icmp sge <2 x i64> %0, %1
13226 %3 = bitcast i8 %__u to <8 x i1>
13227 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <2 x i32> <i32 0, i32 1>
13228 %4 = and <2 x i1> %extract.i, %2
13229 %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
13230 %6 = bitcast <64 x i1> %5 to i64
13231 ret i64 %6
13232 }
13235 define zeroext i8 @test_vpcmpsgeq_v4i1_v8i1_mask(<4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
13236 ; VLX-LABEL: test_vpcmpsgeq_v4i1_v8i1_mask:
13237 ; VLX: # %bb.0: # %entry
13238 ; VLX-NEXT: vpcmpnltq %ymm1, %ymm0, %k0
13239 ; VLX-NEXT: kmovd %k0, %eax
13240 ; VLX-NEXT: # kill: def $al killed $al killed $eax
13241 ; VLX-NEXT: vzeroupper
13242 ; VLX-NEXT: retq
13243 ;
13244 ; NoVLX-LABEL: test_vpcmpsgeq_v4i1_v8i1_mask:
13245 ; NoVLX: # %bb.0: # %entry
13246 ; NoVLX-NEXT: # kill: def $ymm1 killed $ymm1 def $zmm1
13247 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
13248 ; NoVLX-NEXT: vpcmpnltq %zmm1, %zmm0, %k0
13249 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
13250 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
13251 ; NoVLX-NEXT: kmovw %k0, %eax
13252 ; NoVLX-NEXT: # kill: def $al killed $al killed $eax
13253 ; NoVLX-NEXT: vzeroupper
13254 ; NoVLX-NEXT: retq
13255 entry:
13256 %0 = bitcast <4 x i64> %__a to <4 x i64>
13257 %1 = bitcast <4 x i64> %__b to <4 x i64>
13258 %2 = icmp sge <4 x i64> %0, %1
13259 %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
13260 %4 = bitcast <8 x i1> %3 to i8
13261 ret i8 %4
13262 }
13264 define zeroext i8 @test_vpcmpsgeq_v4i1_v8i1_mask_mem(<4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
13265 ; VLX-LABEL: test_vpcmpsgeq_v4i1_v8i1_mask_mem:
13266 ; VLX: # %bb.0: # %entry
13267 ; VLX-NEXT: vpcmpnltq (%rdi), %ymm0, %k0
13268 ; VLX-NEXT: kmovd %k0, %eax
13269 ; VLX-NEXT: # kill: def $al killed $al killed $eax
13270 ; VLX-NEXT: vzeroupper
13271 ; VLX-NEXT: retq
13272 ;
13273 ; NoVLX-LABEL: test_vpcmpsgeq_v4i1_v8i1_mask_mem:
13274 ; NoVLX: # %bb.0: # %entry
13275 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
13276 ; NoVLX-NEXT: vmovdqa (%rdi), %ymm1
13277 ; NoVLX-NEXT: vpcmpnltq %zmm1, %zmm0, %k0
13278 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
13279 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
13280 ; NoVLX-NEXT: kmovw %k0, %eax
13281 ; NoVLX-NEXT: # kill: def $al killed $al killed $eax
13282 ; NoVLX-NEXT: vzeroupper
13283 ; NoVLX-NEXT: retq
13284 entry:
13285 %0 = bitcast <4 x i64> %__a to <4 x i64>
13286 %load = load <4 x i64>, <4 x i64>* %__b
13287 %1 = bitcast <4 x i64> %load to <4 x i64>
13288 %2 = icmp sge <4 x i64> %0, %1
13289 %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
13290 %4 = bitcast <8 x i1> %3 to i8
13291 ret i8 %4
13292 }
13294 define zeroext i8 @test_masked_vpcmpsgeq_v4i1_v8i1_mask(i8 zeroext %__u, <4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
13295 ; VLX-LABEL: test_masked_vpcmpsgeq_v4i1_v8i1_mask:
13296 ; VLX: # %bb.0: # %entry
13297 ; VLX-NEXT: kmovd %edi, %k1
13298 ; VLX-NEXT: vpcmpnltq %ymm1, %ymm0, %k0 {%k1}
13299 ; VLX-NEXT: kmovd %k0, %eax
13300 ; VLX-NEXT: # kill: def $al killed $al killed $eax
13301 ; VLX-NEXT: vzeroupper
13302 ; VLX-NEXT: retq
13303 ;
13304 ; NoVLX-LABEL: test_masked_vpcmpsgeq_v4i1_v8i1_mask:
13305 ; NoVLX: # %bb.0: # %entry
13306 ; NoVLX-NEXT: # kill: def $ymm1 killed $ymm1 def $zmm1
13307 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
13308 ; NoVLX-NEXT: kmovw %edi, %k1
13309 ; NoVLX-NEXT: vpcmpnltq %zmm1, %zmm0, %k0 {%k1}
13310 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
13311 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
13312 ; NoVLX-NEXT: kmovw %k0, %eax
13313 ; NoVLX-NEXT: # kill: def $al killed $al killed $eax
13314 ; NoVLX-NEXT: vzeroupper
13315 ; NoVLX-NEXT: retq
13316 entry:
13317 %0 = bitcast <4 x i64> %__a to <4 x i64>
13318 %1 = bitcast <4 x i64> %__b to <4 x i64>
13319 %2 = icmp sge <4 x i64> %0, %1
13320 %3 = bitcast i8 %__u to <8 x i1>
13321 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
13322 %4 = and <4 x i1> %2, %extract.i
13323 %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
13324 %6 = bitcast <8 x i1> %5 to i8
13325 ret i8 %6
13326 }
13328 define zeroext i8 @test_masked_vpcmpsgeq_v4i1_v8i1_mask_mem(i8 zeroext %__u, <4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
13329 ; VLX-LABEL: test_masked_vpcmpsgeq_v4i1_v8i1_mask_mem:
13330 ; VLX: # %bb.0: # %entry
13331 ; VLX-NEXT: kmovd %edi, %k1
13332 ; VLX-NEXT: vpcmpnltq (%rsi), %ymm0, %k0 {%k1}
13333 ; VLX-NEXT: kmovd %k0, %eax
13334 ; VLX-NEXT: # kill: def $al killed $al killed $eax
13335 ; VLX-NEXT: vzeroupper
13336 ; VLX-NEXT: retq
13337 ;
13338 ; NoVLX-LABEL: test_masked_vpcmpsgeq_v4i1_v8i1_mask_mem:
13339 ; NoVLX: # %bb.0: # %entry
13340 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
13341 ; NoVLX-NEXT: vmovdqa (%rsi), %ymm1
13342 ; NoVLX-NEXT: kmovw %edi, %k1
13343 ; NoVLX-NEXT: vpcmpnltq %zmm1, %zmm0, %k0 {%k1}
13344 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
13345 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
13346 ; NoVLX-NEXT: kmovw %k0, %eax
13347 ; NoVLX-NEXT: # kill: def $al killed $al killed $eax
13348 ; NoVLX-NEXT: vzeroupper
13349 ; NoVLX-NEXT: retq
13350 entry:
13351 %0 = bitcast <4 x i64> %__a to <4 x i64>
13352 %load = load <4 x i64>, <4 x i64>* %__b
13353 %1 = bitcast <4 x i64> %load to <4 x i64>
13354 %2 = icmp sge <4 x i64> %0, %1
13355 %3 = bitcast i8 %__u to <8 x i1>
13356 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
13357 %4 = and <4 x i1> %2, %extract.i
13358 %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
13359 %6 = bitcast <8 x i1> %5 to i8
13360 ret i8 %6
13361 }
13364 define zeroext i8 @test_vpcmpsgeq_v4i1_v8i1_mask_mem_b(<4 x i64> %__a, i64* %__b) local_unnamed_addr {
13365 ; VLX-LABEL: test_vpcmpsgeq_v4i1_v8i1_mask_mem_b:
13366 ; VLX: # %bb.0: # %entry
13367 ; VLX-NEXT: vpcmpnltq (%rdi){1to4}, %ymm0, %k0
13368 ; VLX-NEXT: kmovd %k0, %eax
13369 ; VLX-NEXT: # kill: def $al killed $al killed $eax
13370 ; VLX-NEXT: vzeroupper
13371 ; VLX-NEXT: retq
13372 ;
13373 ; NoVLX-LABEL: test_vpcmpsgeq_v4i1_v8i1_mask_mem_b:
13374 ; NoVLX: # %bb.0: # %entry
13375 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
13376 ; NoVLX-NEXT: vpcmpnltq (%rdi){1to8}, %zmm0, %k0
13377 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
13378 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
13379 ; NoVLX-NEXT: kmovw %k0, %eax
13380 ; NoVLX-NEXT: # kill: def $al killed $al killed $eax
13381 ; NoVLX-NEXT: vzeroupper
13382 ; NoVLX-NEXT: retq
13383 entry:
13384 %0 = bitcast <4 x i64> %__a to <4 x i64>
13385 %load = load i64, i64* %__b
13386 %vec = insertelement <4 x i64> undef, i64 %load, i32 0
13387 %1 = shufflevector <4 x i64> %vec, <4 x i64> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
13388 %2 = icmp sge <4 x i64> %0, %1
13389 %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
13390 %4 = bitcast <8 x i1> %3 to i8
13391 ret i8 %4
13392 }
13394 define zeroext i8 @test_masked_vpcmpsgeq_v4i1_v8i1_mask_mem_b(i8 zeroext %__u, <4 x i64> %__a, i64* %__b) local_unnamed_addr {
13395 ; VLX-LABEL: test_masked_vpcmpsgeq_v4i1_v8i1_mask_mem_b:
13396 ; VLX: # %bb.0: # %entry
13397 ; VLX-NEXT: kmovd %edi, %k1
13398 ; VLX-NEXT: vpcmpnltq (%rsi){1to4}, %ymm0, %k0 {%k1}
13399 ; VLX-NEXT: kmovd %k0, %eax
13400 ; VLX-NEXT: # kill: def $al killed $al killed $eax
13401 ; VLX-NEXT: vzeroupper
13402 ; VLX-NEXT: retq
13403 ;
13404 ; NoVLX-LABEL: test_masked_vpcmpsgeq_v4i1_v8i1_mask_mem_b:
13405 ; NoVLX: # %bb.0: # %entry
13406 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
13407 ; NoVLX-NEXT: kmovw %edi, %k1
13408 ; NoVLX-NEXT: vpcmpnltq (%rsi){1to8}, %zmm0, %k0 {%k1}
13409 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
13410 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
13411 ; NoVLX-NEXT: kmovw %k0, %eax
13412 ; NoVLX-NEXT: # kill: def $al killed $al killed $eax
13413 ; NoVLX-NEXT: vzeroupper
13414 ; NoVLX-NEXT: retq
13415 entry:
13416 %0 = bitcast <4 x i64> %__a to <4 x i64>
13417 %load = load i64, i64* %__b
13418 %vec = insertelement <4 x i64> undef, i64 %load, i32 0
13419 %1 = shufflevector <4 x i64> %vec, <4 x i64> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
13420 %2 = icmp sge <4 x i64> %0, %1
13421 %3 = bitcast i8 %__u to <8 x i1>
13422 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
13423 %4 = and <4 x i1> %extract.i, %2
13424 %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
13425 %6 = bitcast <8 x i1> %5 to i8
13426 ret i8 %6
13427 }
13430 define zeroext i16 @test_vpcmpsgeq_v4i1_v16i1_mask(<4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
13431 ; VLX-LABEL: test_vpcmpsgeq_v4i1_v16i1_mask:
13432 ; VLX: # %bb.0: # %entry
13433 ; VLX-NEXT: vpcmpnltq %ymm1, %ymm0, %k0
13434 ; VLX-NEXT: kmovd %k0, %eax
13435 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
13436 ; VLX-NEXT: vzeroupper
13437 ; VLX-NEXT: retq
13438 ;
13439 ; NoVLX-LABEL: test_vpcmpsgeq_v4i1_v16i1_mask:
13440 ; NoVLX: # %bb.0: # %entry
13441 ; NoVLX-NEXT: # kill: def $ymm1 killed $ymm1 def $zmm1
13442 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
13443 ; NoVLX-NEXT: vpcmpnltq %zmm1, %zmm0, %k0
13444 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
13445 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
13446 ; NoVLX-NEXT: kmovw %k0, %eax
13447 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
13448 ; NoVLX-NEXT: vzeroupper
13449 ; NoVLX-NEXT: retq
13450 entry:
13451 %0 = bitcast <4 x i64> %__a to <4 x i64>
13452 %1 = bitcast <4 x i64> %__b to <4 x i64>
13453 %2 = icmp sge <4 x i64> %0, %1
13454 %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
13455 %4 = bitcast <16 x i1> %3 to i16
13456 ret i16 %4
13457 }
13459 define zeroext i16 @test_vpcmpsgeq_v4i1_v16i1_mask_mem(<4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
13460 ; VLX-LABEL: test_vpcmpsgeq_v4i1_v16i1_mask_mem:
13461 ; VLX: # %bb.0: # %entry
13462 ; VLX-NEXT: vpcmpnltq (%rdi), %ymm0, %k0
13463 ; VLX-NEXT: kmovd %k0, %eax
13464 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
13465 ; VLX-NEXT: vzeroupper
13466 ; VLX-NEXT: retq
13467 ;
13468 ; NoVLX-LABEL: test_vpcmpsgeq_v4i1_v16i1_mask_mem:
13469 ; NoVLX: # %bb.0: # %entry
13470 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
13471 ; NoVLX-NEXT: vmovdqa (%rdi), %ymm1
13472 ; NoVLX-NEXT: vpcmpnltq %zmm1, %zmm0, %k0
13473 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
13474 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
13475 ; NoVLX-NEXT: kmovw %k0, %eax
13476 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
13477 ; NoVLX-NEXT: vzeroupper
13478 ; NoVLX-NEXT: retq
13479 entry:
13480 %0 = bitcast <4 x i64> %__a to <4 x i64>
13481 %load = load <4 x i64>, <4 x i64>* %__b
13482 %1 = bitcast <4 x i64> %load to <4 x i64>
13483 %2 = icmp sge <4 x i64> %0, %1
13484 %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
13485 %4 = bitcast <16 x i1> %3 to i16
13486 ret i16 %4
13487 }
13489 define zeroext i16 @test_masked_vpcmpsgeq_v4i1_v16i1_mask(i8 zeroext %__u, <4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
13490 ; VLX-LABEL: test_masked_vpcmpsgeq_v4i1_v16i1_mask:
13491 ; VLX: # %bb.0: # %entry
13492 ; VLX-NEXT: kmovd %edi, %k1
13493 ; VLX-NEXT: vpcmpnltq %ymm1, %ymm0, %k0 {%k1}
13494 ; VLX-NEXT: kmovd %k0, %eax
13495 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
13496 ; VLX-NEXT: vzeroupper
13497 ; VLX-NEXT: retq
13498 ;
13499 ; NoVLX-LABEL: test_masked_vpcmpsgeq_v4i1_v16i1_mask:
13500 ; NoVLX: # %bb.0: # %entry
13501 ; NoVLX-NEXT: # kill: def $ymm1 killed $ymm1 def $zmm1
13502 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
13503 ; NoVLX-NEXT: kmovw %edi, %k1
13504 ; NoVLX-NEXT: vpcmpnltq %zmm1, %zmm0, %k0 {%k1}
13505 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
13506 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
13507 ; NoVLX-NEXT: kmovw %k0, %eax
13508 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
13509 ; NoVLX-NEXT: vzeroupper
13510 ; NoVLX-NEXT: retq
13511 entry:
13512 %0 = bitcast <4 x i64> %__a to <4 x i64>
13513 %1 = bitcast <4 x i64> %__b to <4 x i64>
13514 %2 = icmp sge <4 x i64> %0, %1
13515 %3 = bitcast i8 %__u to <8 x i1>
13516 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
13517 %4 = and <4 x i1> %2, %extract.i
13518 %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
13519 %6 = bitcast <16 x i1> %5 to i16
13520 ret i16 %6
13521 }
13523 define zeroext i16 @test_masked_vpcmpsgeq_v4i1_v16i1_mask_mem(i8 zeroext %__u, <4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
13524 ; VLX-LABEL: test_masked_vpcmpsgeq_v4i1_v16i1_mask_mem:
13525 ; VLX: # %bb.0: # %entry
13526 ; VLX-NEXT: kmovd %edi, %k1
13527 ; VLX-NEXT: vpcmpnltq (%rsi), %ymm0, %k0 {%k1}
13528 ; VLX-NEXT: kmovd %k0, %eax
13529 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
13530 ; VLX-NEXT: vzeroupper
13531 ; VLX-NEXT: retq
13532 ;
13533 ; NoVLX-LABEL: test_masked_vpcmpsgeq_v4i1_v16i1_mask_mem:
13534 ; NoVLX: # %bb.0: # %entry
13535 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
13536 ; NoVLX-NEXT: vmovdqa (%rsi), %ymm1
13537 ; NoVLX-NEXT: kmovw %edi, %k1
13538 ; NoVLX-NEXT: vpcmpnltq %zmm1, %zmm0, %k0 {%k1}
13539 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
13540 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
13541 ; NoVLX-NEXT: kmovw %k0, %eax
13542 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
13543 ; NoVLX-NEXT: vzeroupper
13544 ; NoVLX-NEXT: retq
13545 entry:
13546 %0 = bitcast <4 x i64> %__a to <4 x i64>
13547 %load = load <4 x i64>, <4 x i64>* %__b
13548 %1 = bitcast <4 x i64> %load to <4 x i64>
13549 %2 = icmp sge <4 x i64> %0, %1
13550 %3 = bitcast i8 %__u to <8 x i1>
13551 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
13552 %4 = and <4 x i1> %2, %extract.i
13553 %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
13554 %6 = bitcast <16 x i1> %5 to i16
13555 ret i16 %6
13556 }
13559 define zeroext i16 @test_vpcmpsgeq_v4i1_v16i1_mask_mem_b(<4 x i64> %__a, i64* %__b) local_unnamed_addr {
13560 ; VLX-LABEL: test_vpcmpsgeq_v4i1_v16i1_mask_mem_b:
13561 ; VLX: # %bb.0: # %entry
13562 ; VLX-NEXT: vpcmpnltq (%rdi){1to4}, %ymm0, %k0
13563 ; VLX-NEXT: kmovd %k0, %eax
13564 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
13565 ; VLX-NEXT: vzeroupper
13566 ; VLX-NEXT: retq
13567 ;
13568 ; NoVLX-LABEL: test_vpcmpsgeq_v4i1_v16i1_mask_mem_b:
13569 ; NoVLX: # %bb.0: # %entry
13570 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
13571 ; NoVLX-NEXT: vpcmpnltq (%rdi){1to8}, %zmm0, %k0
13572 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
13573 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
13574 ; NoVLX-NEXT: kmovw %k0, %eax
13575 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
13576 ; NoVLX-NEXT: vzeroupper
13577 ; NoVLX-NEXT: retq
13578 entry:
13579 %0 = bitcast <4 x i64> %__a to <4 x i64>
13580 %load = load i64, i64* %__b
13581 %vec = insertelement <4 x i64> undef, i64 %load, i32 0
13582 %1 = shufflevector <4 x i64> %vec, <4 x i64> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
13583 %2 = icmp sge <4 x i64> %0, %1
13584 %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
13585 %4 = bitcast <16 x i1> %3 to i16
13586 ret i16 %4
13587 }
13589 define zeroext i16 @test_masked_vpcmpsgeq_v4i1_v16i1_mask_mem_b(i8 zeroext %__u, <4 x i64> %__a, i64* %__b) local_unnamed_addr {
13590 ; VLX-LABEL: test_masked_vpcmpsgeq_v4i1_v16i1_mask_mem_b:
13591 ; VLX: # %bb.0: # %entry
13592 ; VLX-NEXT: kmovd %edi, %k1
13593 ; VLX-NEXT: vpcmpnltq (%rsi){1to4}, %ymm0, %k0 {%k1}
13594 ; VLX-NEXT: kmovd %k0, %eax
13595 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
13596 ; VLX-NEXT: vzeroupper
13597 ; VLX-NEXT: retq
13598 ;
13599 ; NoVLX-LABEL: test_masked_vpcmpsgeq_v4i1_v16i1_mask_mem_b:
13600 ; NoVLX: # %bb.0: # %entry
13601 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
13602 ; NoVLX-NEXT: kmovw %edi, %k1
13603 ; NoVLX-NEXT: vpcmpnltq (%rsi){1to8}, %zmm0, %k0 {%k1}
13604 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
13605 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
13606 ; NoVLX-NEXT: kmovw %k0, %eax
13607 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
13608 ; NoVLX-NEXT: vzeroupper
13609 ; NoVLX-NEXT: retq
13610 entry:
13611 %0 = bitcast <4 x i64> %__a to <4 x i64>
13612 %load = load i64, i64* %__b
13613 %vec = insertelement <4 x i64> undef, i64 %load, i32 0
13614 %1 = shufflevector <4 x i64> %vec, <4 x i64> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
13615 %2 = icmp sge <4 x i64> %0, %1
13616 %3 = bitcast i8 %__u to <8 x i1>
13617 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
13618 %4 = and <4 x i1> %extract.i, %2
13619 %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
13620 %6 = bitcast <16 x i1> %5 to i16
13621 ret i16 %6
13622 }
13625 define zeroext i32 @test_vpcmpsgeq_v4i1_v32i1_mask(<4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
13626 ; VLX-LABEL: test_vpcmpsgeq_v4i1_v32i1_mask:
13627 ; VLX: # %bb.0: # %entry
13628 ; VLX-NEXT: vpcmpnltq %ymm1, %ymm0, %k0
13629 ; VLX-NEXT: kmovd %k0, %eax
13630 ; VLX-NEXT: vzeroupper
13631 ; VLX-NEXT: retq
13632 ;
13633 ; NoVLX-LABEL: test_vpcmpsgeq_v4i1_v32i1_mask:
13634 ; NoVLX: # %bb.0: # %entry
13635 ; NoVLX-NEXT: # kill: def $ymm1 killed $ymm1 def $zmm1
13636 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
13637 ; NoVLX-NEXT: vpcmpnltq %zmm1, %zmm0, %k0
13638 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
13639 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
13640 ; NoVLX-NEXT: kmovw %k0, %eax
13641 ; NoVLX-NEXT: vzeroupper
13642 ; NoVLX-NEXT: retq
13643 entry:
13644 %0 = bitcast <4 x i64> %__a to <4 x i64>
13645 %1 = bitcast <4 x i64> %__b to <4 x i64>
13646 %2 = icmp sge <4 x i64> %0, %1
13647 %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
13648 %4 = bitcast <32 x i1> %3 to i32
13649 ret i32 %4
13650 }
13652 define zeroext i32 @test_vpcmpsgeq_v4i1_v32i1_mask_mem(<4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
13653 ; VLX-LABEL: test_vpcmpsgeq_v4i1_v32i1_mask_mem:
13654 ; VLX: # %bb.0: # %entry
13655 ; VLX-NEXT: vpcmpnltq (%rdi), %ymm0, %k0
13656 ; VLX-NEXT: kmovd %k0, %eax
13657 ; VLX-NEXT: vzeroupper
13658 ; VLX-NEXT: retq
13659 ;
13660 ; NoVLX-LABEL: test_vpcmpsgeq_v4i1_v32i1_mask_mem:
13661 ; NoVLX: # %bb.0: # %entry
13662 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
13663 ; NoVLX-NEXT: vmovdqa (%rdi), %ymm1
13664 ; NoVLX-NEXT: vpcmpnltq %zmm1, %zmm0, %k0
13665 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
13666 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
13667 ; NoVLX-NEXT: kmovw %k0, %eax
13668 ; NoVLX-NEXT: vzeroupper
13669 ; NoVLX-NEXT: retq
13670 entry:
13671 %0 = bitcast <4 x i64> %__a to <4 x i64>
13672 %load = load <4 x i64>, <4 x i64>* %__b
13673 %1 = bitcast <4 x i64> %load to <4 x i64>
13674 %2 = icmp sge <4 x i64> %0, %1
13675 %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
13676 %4 = bitcast <32 x i1> %3 to i32
13677 ret i32 %4
13678 }
13680 define zeroext i32 @test_masked_vpcmpsgeq_v4i1_v32i1_mask(i8 zeroext %__u, <4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
13681 ; VLX-LABEL: test_masked_vpcmpsgeq_v4i1_v32i1_mask:
13682 ; VLX: # %bb.0: # %entry
13683 ; VLX-NEXT: kmovd %edi, %k1
13684 ; VLX-NEXT: vpcmpnltq %ymm1, %ymm0, %k0 {%k1}
13685 ; VLX-NEXT: kmovd %k0, %eax
13686 ; VLX-NEXT: vzeroupper
13687 ; VLX-NEXT: retq
13688 ;
13689 ; NoVLX-LABEL: test_masked_vpcmpsgeq_v4i1_v32i1_mask:
13690 ; NoVLX: # %bb.0: # %entry
13691 ; NoVLX-NEXT: # kill: def $ymm1 killed $ymm1 def $zmm1
13692 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
13693 ; NoVLX-NEXT: kmovw %edi, %k1
13694 ; NoVLX-NEXT: vpcmpnltq %zmm1, %zmm0, %k0 {%k1}
13695 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
13696 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
13697 ; NoVLX-NEXT: kmovw %k0, %eax
13698 ; NoVLX-NEXT: vzeroupper
13699 ; NoVLX-NEXT: retq
13700 entry:
13701 %0 = bitcast <4 x i64> %__a to <4 x i64>
13702 %1 = bitcast <4 x i64> %__b to <4 x i64>
13703 %2 = icmp sge <4 x i64> %0, %1
13704 %3 = bitcast i8 %__u to <8 x i1>
13705 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
13706 %4 = and <4 x i1> %2, %extract.i
13707 %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
13708 %6 = bitcast <32 x i1> %5 to i32
13709 ret i32 %6
13710 }
13712 define zeroext i32 @test_masked_vpcmpsgeq_v4i1_v32i1_mask_mem(i8 zeroext %__u, <4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
13713 ; VLX-LABEL: test_masked_vpcmpsgeq_v4i1_v32i1_mask_mem:
13714 ; VLX: # %bb.0: # %entry
13715 ; VLX-NEXT: kmovd %edi, %k1
13716 ; VLX-NEXT: vpcmpnltq (%rsi), %ymm0, %k0 {%k1}
13717 ; VLX-NEXT: kmovd %k0, %eax
13718 ; VLX-NEXT: vzeroupper
13719 ; VLX-NEXT: retq
13720 ;
13721 ; NoVLX-LABEL: test_masked_vpcmpsgeq_v4i1_v32i1_mask_mem:
13722 ; NoVLX: # %bb.0: # %entry
13723 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
13724 ; NoVLX-NEXT: vmovdqa (%rsi), %ymm1
13725 ; NoVLX-NEXT: kmovw %edi, %k1
13726 ; NoVLX-NEXT: vpcmpnltq %zmm1, %zmm0, %k0 {%k1}
13727 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
13728 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
13729 ; NoVLX-NEXT: kmovw %k0, %eax
13730 ; NoVLX-NEXT: vzeroupper
13731 ; NoVLX-NEXT: retq
13732 entry:
13733 %0 = bitcast <4 x i64> %__a to <4 x i64>
13734 %load = load <4 x i64>, <4 x i64>* %__b
13735 %1 = bitcast <4 x i64> %load to <4 x i64>
13736 %2 = icmp sge <4 x i64> %0, %1
13737 %3 = bitcast i8 %__u to <8 x i1>
13738 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
13739 %4 = and <4 x i1> %2, %extract.i
13740 %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
13741 %6 = bitcast <32 x i1> %5 to i32
13742 ret i32 %6
13743 }
13746 define zeroext i32 @test_vpcmpsgeq_v4i1_v32i1_mask_mem_b(<4 x i64> %__a, i64* %__b) local_unnamed_addr {
13747 ; VLX-LABEL: test_vpcmpsgeq_v4i1_v32i1_mask_mem_b:
13748 ; VLX: # %bb.0: # %entry
13749 ; VLX-NEXT: vpcmpnltq (%rdi){1to4}, %ymm0, %k0
13750 ; VLX-NEXT: kmovd %k0, %eax
13751 ; VLX-NEXT: vzeroupper
13752 ; VLX-NEXT: retq
13753 ;
13754 ; NoVLX-LABEL: test_vpcmpsgeq_v4i1_v32i1_mask_mem_b:
13755 ; NoVLX: # %bb.0: # %entry
13756 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
13757 ; NoVLX-NEXT: vpcmpnltq (%rdi){1to8}, %zmm0, %k0
13758 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
13759 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
13760 ; NoVLX-NEXT: kmovw %k0, %eax
13761 ; NoVLX-NEXT: vzeroupper
13762 ; NoVLX-NEXT: retq
13763 entry:
13764 %0 = bitcast <4 x i64> %__a to <4 x i64>
13765 %load = load i64, i64* %__b
13766 %vec = insertelement <4 x i64> undef, i64 %load, i32 0
13767 %1 = shufflevector <4 x i64> %vec, <4 x i64> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
13768 %2 = icmp sge <4 x i64> %0, %1
13769 %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
13770 %4 = bitcast <32 x i1> %3 to i32
13771 ret i32 %4
13772 }
13774 define zeroext i32 @test_masked_vpcmpsgeq_v4i1_v32i1_mask_mem_b(i8 zeroext %__u, <4 x i64> %__a, i64* %__b) local_unnamed_addr {
13775 ; VLX-LABEL: test_masked_vpcmpsgeq_v4i1_v32i1_mask_mem_b:
13776 ; VLX: # %bb.0: # %entry
13777 ; VLX-NEXT: kmovd %edi, %k1
13778 ; VLX-NEXT: vpcmpnltq (%rsi){1to4}, %ymm0, %k0 {%k1}
13779 ; VLX-NEXT: kmovd %k0, %eax
13780 ; VLX-NEXT: vzeroupper
13781 ; VLX-NEXT: retq
13782 ;
13783 ; NoVLX-LABEL: test_masked_vpcmpsgeq_v4i1_v32i1_mask_mem_b:
13784 ; NoVLX: # %bb.0: # %entry
13785 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
13786 ; NoVLX-NEXT: kmovw %edi, %k1
13787 ; NoVLX-NEXT: vpcmpnltq (%rsi){1to8}, %zmm0, %k0 {%k1}
13788 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
13789 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
13790 ; NoVLX-NEXT: kmovw %k0, %eax
13791 ; NoVLX-NEXT: vzeroupper
13792 ; NoVLX-NEXT: retq
13793 entry:
13794 %0 = bitcast <4 x i64> %__a to <4 x i64>
13795 %load = load i64, i64* %__b
13796 %vec = insertelement <4 x i64> undef, i64 %load, i32 0
13797 %1 = shufflevector <4 x i64> %vec, <4 x i64> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
13798 %2 = icmp sge <4 x i64> %0, %1
13799 %3 = bitcast i8 %__u to <8 x i1>
13800 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
13801 %4 = and <4 x i1> %extract.i, %2
13802 %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
13803 %6 = bitcast <32 x i1> %5 to i32
13804 ret i32 %6
13805 }
13808 define zeroext i64 @test_vpcmpsgeq_v4i1_v64i1_mask(<4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
13809 ; VLX-LABEL: test_vpcmpsgeq_v4i1_v64i1_mask:
13810 ; VLX: # %bb.0: # %entry
13811 ; VLX-NEXT: vpcmpnltq %ymm1, %ymm0, %k0
13812 ; VLX-NEXT: kmovq %k0, %rax
13813 ; VLX-NEXT: vzeroupper
13814 ; VLX-NEXT: retq
13815 ;
13816 ; NoVLX-LABEL: test_vpcmpsgeq_v4i1_v64i1_mask:
13817 ; NoVLX: # %bb.0: # %entry
13818 ; NoVLX-NEXT: # kill: def $ymm1 killed $ymm1 def $zmm1
13819 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
13820 ; NoVLX-NEXT: vpcmpnltq %zmm1, %zmm0, %k0
13821 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
13822 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
13823 ; NoVLX-NEXT: kmovw %k0, %eax
13824 ; NoVLX-NEXT: vzeroupper
13825 ; NoVLX-NEXT: retq
13826 entry:
13827 %0 = bitcast <4 x i64> %__a to <4 x i64>
13828 %1 = bitcast <4 x i64> %__b to <4 x i64>
13829 %2 = icmp sge <4 x i64> %0, %1
13830 %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
13831 %4 = bitcast <64 x i1> %3 to i64
13832 ret i64 %4
13833 }
13835 define zeroext i64 @test_vpcmpsgeq_v4i1_v64i1_mask_mem(<4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
13836 ; VLX-LABEL: test_vpcmpsgeq_v4i1_v64i1_mask_mem:
13837 ; VLX: # %bb.0: # %entry
13838 ; VLX-NEXT: vpcmpnltq (%rdi), %ymm0, %k0
13839 ; VLX-NEXT: kmovq %k0, %rax
13840 ; VLX-NEXT: vzeroupper
13841 ; VLX-NEXT: retq
13842 ;
13843 ; NoVLX-LABEL: test_vpcmpsgeq_v4i1_v64i1_mask_mem:
13844 ; NoVLX: # %bb.0: # %entry
13845 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
13846 ; NoVLX-NEXT: vmovdqa (%rdi), %ymm1
13847 ; NoVLX-NEXT: vpcmpnltq %zmm1, %zmm0, %k0
13848 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
13849 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
13850 ; NoVLX-NEXT: kmovw %k0, %eax
13851 ; NoVLX-NEXT: vzeroupper
13852 ; NoVLX-NEXT: retq
13853 entry:
13854 %0 = bitcast <4 x i64> %__a to <4 x i64>
13855 %load = load <4 x i64>, <4 x i64>* %__b
13856 %1 = bitcast <4 x i64> %load to <4 x i64>
13857 %2 = icmp sge <4 x i64> %0, %1
13858 %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
13859 %4 = bitcast <64 x i1> %3 to i64
13860 ret i64 %4
13861 }
13863 define zeroext i64 @test_masked_vpcmpsgeq_v4i1_v64i1_mask(i8 zeroext %__u, <4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
13864 ; VLX-LABEL: test_masked_vpcmpsgeq_v4i1_v64i1_mask:
13865 ; VLX: # %bb.0: # %entry
13866 ; VLX-NEXT: kmovd %edi, %k1
13867 ; VLX-NEXT: vpcmpnltq %ymm1, %ymm0, %k0 {%k1}
13868 ; VLX-NEXT: kmovq %k0, %rax
13869 ; VLX-NEXT: vzeroupper
13870 ; VLX-NEXT: retq
13871 ;
13872 ; NoVLX-LABEL: test_masked_vpcmpsgeq_v4i1_v64i1_mask:
13873 ; NoVLX: # %bb.0: # %entry
13874 ; NoVLX-NEXT: # kill: def $ymm1 killed $ymm1 def $zmm1
13875 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
13876 ; NoVLX-NEXT: kmovw %edi, %k1
13877 ; NoVLX-NEXT: vpcmpnltq %zmm1, %zmm0, %k0 {%k1}
13878 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
13879 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
13880 ; NoVLX-NEXT: kmovw %k0, %eax
13881 ; NoVLX-NEXT: vzeroupper
13882 ; NoVLX-NEXT: retq
13883 entry:
13884 %0 = bitcast <4 x i64> %__a to <4 x i64>
13885 %1 = bitcast <4 x i64> %__b to <4 x i64>
13886 %2 = icmp sge <4 x i64> %0, %1
13887 %3 = bitcast i8 %__u to <8 x i1>
13888 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
13889 %4 = and <4 x i1> %2, %extract.i
13890 %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
13891 %6 = bitcast <64 x i1> %5 to i64
13892 ret i64 %6
13893 }
13895 define zeroext i64 @test_masked_vpcmpsgeq_v4i1_v64i1_mask_mem(i8 zeroext %__u, <4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
13896 ; VLX-LABEL: test_masked_vpcmpsgeq_v4i1_v64i1_mask_mem:
13897 ; VLX: # %bb.0: # %entry
13898 ; VLX-NEXT: kmovd %edi, %k1
13899 ; VLX-NEXT: vpcmpnltq (%rsi), %ymm0, %k0 {%k1}
13900 ; VLX-NEXT: kmovq %k0, %rax
13901 ; VLX-NEXT: vzeroupper
13902 ; VLX-NEXT: retq
13903 ;
13904 ; NoVLX-LABEL: test_masked_vpcmpsgeq_v4i1_v64i1_mask_mem:
13905 ; NoVLX: # %bb.0: # %entry
13906 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
13907 ; NoVLX-NEXT: vmovdqa (%rsi), %ymm1
13908 ; NoVLX-NEXT: kmovw %edi, %k1
13909 ; NoVLX-NEXT: vpcmpnltq %zmm1, %zmm0, %k0 {%k1}
13910 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
13911 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
13912 ; NoVLX-NEXT: kmovw %k0, %eax
13913 ; NoVLX-NEXT: vzeroupper
13914 ; NoVLX-NEXT: retq
13915 entry:
13916 %0 = bitcast <4 x i64> %__a to <4 x i64>
13917 %load = load <4 x i64>, <4 x i64>* %__b
13918 %1 = bitcast <4 x i64> %load to <4 x i64>
13919 %2 = icmp sge <4 x i64> %0, %1
13920 %3 = bitcast i8 %__u to <8 x i1>
13921 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
13922 %4 = and <4 x i1> %2, %extract.i
13923 %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
13924 %6 = bitcast <64 x i1> %5 to i64
13925 ret i64 %6
13926 }
13929 define zeroext i64 @test_vpcmpsgeq_v4i1_v64i1_mask_mem_b(<4 x i64> %__a, i64* %__b) local_unnamed_addr {
13930 ; VLX-LABEL: test_vpcmpsgeq_v4i1_v64i1_mask_mem_b:
13931 ; VLX: # %bb.0: # %entry
13932 ; VLX-NEXT: vpcmpnltq (%rdi){1to4}, %ymm0, %k0
13933 ; VLX-NEXT: kmovq %k0, %rax
13934 ; VLX-NEXT: vzeroupper
13935 ; VLX-NEXT: retq
13936 ;
13937 ; NoVLX-LABEL: test_vpcmpsgeq_v4i1_v64i1_mask_mem_b:
13938 ; NoVLX: # %bb.0: # %entry
13939 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
13940 ; NoVLX-NEXT: vpcmpnltq (%rdi){1to8}, %zmm0, %k0
13941 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
13942 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
13943 ; NoVLX-NEXT: kmovw %k0, %eax
13944 ; NoVLX-NEXT: vzeroupper
13945 ; NoVLX-NEXT: retq
13946 entry:
13947 %0 = bitcast <4 x i64> %__a to <4 x i64>
13948 %load = load i64, i64* %__b
13949 %vec = insertelement <4 x i64> undef, i64 %load, i32 0
13950 %1 = shufflevector <4 x i64> %vec, <4 x i64> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
13951 %2 = icmp sge <4 x i64> %0, %1
13952 %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
13953 %4 = bitcast <64 x i1> %3 to i64
13954 ret i64 %4
13955 }
13957 define zeroext i64 @test_masked_vpcmpsgeq_v4i1_v64i1_mask_mem_b(i8 zeroext %__u, <4 x i64> %__a, i64* %__b) local_unnamed_addr {
13958 ; VLX-LABEL: test_masked_vpcmpsgeq_v4i1_v64i1_mask_mem_b:
13959 ; VLX: # %bb.0: # %entry
13960 ; VLX-NEXT: kmovd %edi, %k1
13961 ; VLX-NEXT: vpcmpnltq (%rsi){1to4}, %ymm0, %k0 {%k1}
13962 ; VLX-NEXT: kmovq %k0, %rax
13963 ; VLX-NEXT: vzeroupper
13964 ; VLX-NEXT: retq
13965 ;
13966 ; NoVLX-LABEL: test_masked_vpcmpsgeq_v4i1_v64i1_mask_mem_b:
13967 ; NoVLX: # %bb.0: # %entry
13968 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
13969 ; NoVLX-NEXT: kmovw %edi, %k1
13970 ; NoVLX-NEXT: vpcmpnltq (%rsi){1to8}, %zmm0, %k0 {%k1}
13971 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
13972 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
13973 ; NoVLX-NEXT: kmovw %k0, %eax
13974 ; NoVLX-NEXT: vzeroupper
13975 ; NoVLX-NEXT: retq
13976 entry:
13977 %0 = bitcast <4 x i64> %__a to <4 x i64>
13978 %load = load i64, i64* %__b
13979 %vec = insertelement <4 x i64> undef, i64 %load, i32 0
13980 %1 = shufflevector <4 x i64> %vec, <4 x i64> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
13981 %2 = icmp sge <4 x i64> %0, %1
13982 %3 = bitcast i8 %__u to <8 x i1>
13983 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
13984 %4 = and <4 x i1> %extract.i, %2
13985 %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
13986 %6 = bitcast <64 x i1> %5 to i64
13987 ret i64 %6
13988 }
13991 define zeroext i16 @test_vpcmpsgeq_v8i1_v16i1_mask(<8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
13992 ; VLX-LABEL: test_vpcmpsgeq_v8i1_v16i1_mask:
13993 ; VLX: # %bb.0: # %entry
13994 ; VLX-NEXT: vpcmpnltq %zmm1, %zmm0, %k0
13995 ; VLX-NEXT: kmovd %k0, %eax
13996 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
13997 ; VLX-NEXT: vzeroupper
13998 ; VLX-NEXT: retq
13999 ;
14000 ; NoVLX-LABEL: test_vpcmpsgeq_v8i1_v16i1_mask:
14001 ; NoVLX: # %bb.0: # %entry
14002 ; NoVLX-NEXT: vpcmpnltq %zmm1, %zmm0, %k0
14003 ; NoVLX-NEXT: kmovw %k0, %eax
14004 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
14005 ; NoVLX-NEXT: vzeroupper
14006 ; NoVLX-NEXT: retq
14007 entry:
14008 %0 = bitcast <8 x i64> %__a to <8 x i64>
14009 %1 = bitcast <8 x i64> %__b to <8 x i64>
14010 %2 = icmp sge <8 x i64> %0, %1
14011 %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
14012 %4 = bitcast <16 x i1> %3 to i16
14013 ret i16 %4
14014 }
14016 define zeroext i16 @test_vpcmpsgeq_v8i1_v16i1_mask_mem(<8 x i64> %__a, <8 x i64>* %__b) local_unnamed_addr {
14017 ; VLX-LABEL: test_vpcmpsgeq_v8i1_v16i1_mask_mem:
14018 ; VLX: # %bb.0: # %entry
14019 ; VLX-NEXT: vpcmpnltq (%rdi), %zmm0, %k0
14020 ; VLX-NEXT: kmovd %k0, %eax
14021 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
14022 ; VLX-NEXT: vzeroupper
14023 ; VLX-NEXT: retq
14024 ;
14025 ; NoVLX-LABEL: test_vpcmpsgeq_v8i1_v16i1_mask_mem:
14026 ; NoVLX: # %bb.0: # %entry
14027 ; NoVLX-NEXT: vpcmpnltq (%rdi), %zmm0, %k0
14028 ; NoVLX-NEXT: kmovw %k0, %eax
14029 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
14030 ; NoVLX-NEXT: vzeroupper
14031 ; NoVLX-NEXT: retq
14032 entry:
14033 %0 = bitcast <8 x i64> %__a to <8 x i64>
14034 %load = load <8 x i64>, <8 x i64>* %__b
14035 %1 = bitcast <8 x i64> %load to <8 x i64>
14036 %2 = icmp sge <8 x i64> %0, %1
14037 %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
14038 %4 = bitcast <16 x i1> %3 to i16
14039 ret i16 %4
14040 }
14042 define zeroext i16 @test_masked_vpcmpsgeq_v8i1_v16i1_mask(i8 zeroext %__u, <8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
14043 ; VLX-LABEL: test_masked_vpcmpsgeq_v8i1_v16i1_mask:
14044 ; VLX: # %bb.0: # %entry
14045 ; VLX-NEXT: kmovd %edi, %k1
14046 ; VLX-NEXT: vpcmpnltq %zmm1, %zmm0, %k0 {%k1}
14047 ; VLX-NEXT: kmovd %k0, %eax
14048 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
14049 ; VLX-NEXT: vzeroupper
14050 ; VLX-NEXT: retq
14051 ;
14052 ; NoVLX-LABEL: test_masked_vpcmpsgeq_v8i1_v16i1_mask:
14053 ; NoVLX: # %bb.0: # %entry
14054 ; NoVLX-NEXT: kmovw %edi, %k1
14055 ; NoVLX-NEXT: vpcmpnltq %zmm1, %zmm0, %k0 {%k1}
14056 ; NoVLX-NEXT: kmovw %k0, %eax
14057 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
14058 ; NoVLX-NEXT: vzeroupper
14059 ; NoVLX-NEXT: retq
14060 entry:
14061 %0 = bitcast <8 x i64> %__a to <8 x i64>
14062 %1 = bitcast <8 x i64> %__b to <8 x i64>
14063 %2 = icmp sge <8 x i64> %0, %1
14064 %3 = bitcast i8 %__u to <8 x i1>
14065 %4 = and <8 x i1> %2, %3
14066 %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
14067 %6 = bitcast <16 x i1> %5 to i16
14068 ret i16 %6
14069 }
14071 define zeroext i16 @test_masked_vpcmpsgeq_v8i1_v16i1_mask_mem(i8 zeroext %__u, <8 x i64> %__a, <8 x i64>* %__b) local_unnamed_addr {
14072 ; VLX-LABEL: test_masked_vpcmpsgeq_v8i1_v16i1_mask_mem:
14073 ; VLX: # %bb.0: # %entry
14074 ; VLX-NEXT: kmovd %edi, %k1
14075 ; VLX-NEXT: vpcmpnltq (%rsi), %zmm0, %k0 {%k1}
14076 ; VLX-NEXT: kmovd %k0, %eax
14077 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
14078 ; VLX-NEXT: vzeroupper
14079 ; VLX-NEXT: retq
14080 ;
14081 ; NoVLX-LABEL: test_masked_vpcmpsgeq_v8i1_v16i1_mask_mem:
14082 ; NoVLX: # %bb.0: # %entry
14083 ; NoVLX-NEXT: kmovw %edi, %k1
14084 ; NoVLX-NEXT: vpcmpnltq (%rsi), %zmm0, %k0 {%k1}
14085 ; NoVLX-NEXT: kmovw %k0, %eax
14086 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
14087 ; NoVLX-NEXT: vzeroupper
14088 ; NoVLX-NEXT: retq
14089 entry:
14090 %0 = bitcast <8 x i64> %__a to <8 x i64>
14091 %load = load <8 x i64>, <8 x i64>* %__b
14092 %1 = bitcast <8 x i64> %load to <8 x i64>
14093 %2 = icmp sge <8 x i64> %0, %1
14094 %3 = bitcast i8 %__u to <8 x i1>
14095 %4 = and <8 x i1> %2, %3
14096 %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
14097 %6 = bitcast <16 x i1> %5 to i16
14098 ret i16 %6
14099 }
14102 define zeroext i16 @test_vpcmpsgeq_v8i1_v16i1_mask_mem_b(<8 x i64> %__a, i64* %__b) local_unnamed_addr {
14103 ; VLX-LABEL: test_vpcmpsgeq_v8i1_v16i1_mask_mem_b:
14104 ; VLX: # %bb.0: # %entry
14105 ; VLX-NEXT: vpcmpnltq (%rdi){1to8}, %zmm0, %k0
14106 ; VLX-NEXT: kmovd %k0, %eax
14107 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
14108 ; VLX-NEXT: vzeroupper
14109 ; VLX-NEXT: retq
14110 ;
14111 ; NoVLX-LABEL: test_vpcmpsgeq_v8i1_v16i1_mask_mem_b:
14112 ; NoVLX: # %bb.0: # %entry
14113 ; NoVLX-NEXT: vpcmpnltq (%rdi){1to8}, %zmm0, %k0
14114 ; NoVLX-NEXT: kmovw %k0, %eax
14115 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
14116 ; NoVLX-NEXT: vzeroupper
14117 ; NoVLX-NEXT: retq
14118 entry:
14119 %0 = bitcast <8 x i64> %__a to <8 x i64>
14120 %load = load i64, i64* %__b
14121 %vec = insertelement <8 x i64> undef, i64 %load, i32 0
14122 %1 = shufflevector <8 x i64> %vec, <8 x i64> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
14123 %2 = icmp sge <8 x i64> %0, %1
14124 %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
14125 %4 = bitcast <16 x i1> %3 to i16
14126 ret i16 %4
14127 }
14129 define zeroext i16 @test_masked_vpcmpsgeq_v8i1_v16i1_mask_mem_b(i8 zeroext %__u, <8 x i64> %__a, i64* %__b) local_unnamed_addr {
14130 ; VLX-LABEL: test_masked_vpcmpsgeq_v8i1_v16i1_mask_mem_b:
14131 ; VLX: # %bb.0: # %entry
14132 ; VLX-NEXT: kmovd %edi, %k1
14133 ; VLX-NEXT: vpcmpnltq (%rsi){1to8}, %zmm0, %k0 {%k1}
14134 ; VLX-NEXT: kmovd %k0, %eax
14135 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
14136 ; VLX-NEXT: vzeroupper
14137 ; VLX-NEXT: retq
14138 ;
14139 ; NoVLX-LABEL: test_masked_vpcmpsgeq_v8i1_v16i1_mask_mem_b:
14140 ; NoVLX: # %bb.0: # %entry
14141 ; NoVLX-NEXT: kmovw %edi, %k1
14142 ; NoVLX-NEXT: vpcmpnltq (%rsi){1to8}, %zmm0, %k0 {%k1}
14143 ; NoVLX-NEXT: kmovw %k0, %eax
14144 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
14145 ; NoVLX-NEXT: vzeroupper
14146 ; NoVLX-NEXT: retq
14147 entry:
14148 %0 = bitcast <8 x i64> %__a to <8 x i64>
14149 %load = load i64, i64* %__b
14150 %vec = insertelement <8 x i64> undef, i64 %load, i32 0
14151 %1 = shufflevector <8 x i64> %vec, <8 x i64> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
14152 %2 = icmp sge <8 x i64> %0, %1
14153 %3 = bitcast i8 %__u to <8 x i1>
14154 %4 = and <8 x i1> %3, %2
14155 %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
14156 %6 = bitcast <16 x i1> %5 to i16
14157 ret i16 %6
14158 }
14161 define zeroext i32 @test_vpcmpsgeq_v8i1_v32i1_mask(<8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
14162 ; VLX-LABEL: test_vpcmpsgeq_v8i1_v32i1_mask:
14163 ; VLX: # %bb.0: # %entry
14164 ; VLX-NEXT: vpcmpnltq %zmm1, %zmm0, %k0
14165 ; VLX-NEXT: kmovd %k0, %eax
14166 ; VLX-NEXT: vzeroupper
14167 ; VLX-NEXT: retq
14168 ;
14169 ; NoVLX-LABEL: test_vpcmpsgeq_v8i1_v32i1_mask:
14170 ; NoVLX: # %bb.0: # %entry
14171 ; NoVLX-NEXT: vpcmpnltq %zmm1, %zmm0, %k0
14172 ; NoVLX-NEXT: kmovw %k0, %eax
14173 ; NoVLX-NEXT: vzeroupper
14174 ; NoVLX-NEXT: retq
14175 entry:
14176 %0 = bitcast <8 x i64> %__a to <8 x i64>
14177 %1 = bitcast <8 x i64> %__b to <8 x i64>
14178 %2 = icmp sge <8 x i64> %0, %1
14179 %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
14180 %4 = bitcast <32 x i1> %3 to i32
14181 ret i32 %4
14182 }
14184 define zeroext i32 @test_vpcmpsgeq_v8i1_v32i1_mask_mem(<8 x i64> %__a, <8 x i64>* %__b) local_unnamed_addr {
14185 ; VLX-LABEL: test_vpcmpsgeq_v8i1_v32i1_mask_mem:
14186 ; VLX: # %bb.0: # %entry
14187 ; VLX-NEXT: vpcmpnltq (%rdi), %zmm0, %k0
14188 ; VLX-NEXT: kmovd %k0, %eax
14189 ; VLX-NEXT: vzeroupper
14190 ; VLX-NEXT: retq
14191 ;
14192 ; NoVLX-LABEL: test_vpcmpsgeq_v8i1_v32i1_mask_mem:
14193 ; NoVLX: # %bb.0: # %entry
14194 ; NoVLX-NEXT: vpcmpnltq (%rdi), %zmm0, %k0
14195 ; NoVLX-NEXT: kmovw %k0, %eax
14196 ; NoVLX-NEXT: vzeroupper
14197 ; NoVLX-NEXT: retq
14198 entry:
14199 %0 = bitcast <8 x i64> %__a to <8 x i64>
14200 %load = load <8 x i64>, <8 x i64>* %__b
14201 %1 = bitcast <8 x i64> %load to <8 x i64>
14202 %2 = icmp sge <8 x i64> %0, %1
14203 %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
14204 %4 = bitcast <32 x i1> %3 to i32
14205 ret i32 %4
14206 }
14208 define zeroext i32 @test_masked_vpcmpsgeq_v8i1_v32i1_mask(i8 zeroext %__u, <8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
14209 ; VLX-LABEL: test_masked_vpcmpsgeq_v8i1_v32i1_mask:
14210 ; VLX: # %bb.0: # %entry
14211 ; VLX-NEXT: kmovd %edi, %k1
14212 ; VLX-NEXT: vpcmpnltq %zmm1, %zmm0, %k0 {%k1}
14213 ; VLX-NEXT: kmovd %k0, %eax
14214 ; VLX-NEXT: vzeroupper
14215 ; VLX-NEXT: retq
14216 ;
14217 ; NoVLX-LABEL: test_masked_vpcmpsgeq_v8i1_v32i1_mask:
14218 ; NoVLX: # %bb.0: # %entry
14219 ; NoVLX-NEXT: kmovw %edi, %k1
14220 ; NoVLX-NEXT: vpcmpnltq %zmm1, %zmm0, %k0 {%k1}
14221 ; NoVLX-NEXT: kmovw %k0, %eax
14222 ; NoVLX-NEXT: vzeroupper
14223 ; NoVLX-NEXT: retq
14224 entry:
14225 %0 = bitcast <8 x i64> %__a to <8 x i64>
14226 %1 = bitcast <8 x i64> %__b to <8 x i64>
14227 %2 = icmp sge <8 x i64> %0, %1
14228 %3 = bitcast i8 %__u to <8 x i1>
14229 %4 = and <8 x i1> %2, %3
14230 %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
14231 %6 = bitcast <32 x i1> %5 to i32
14232 ret i32 %6
14233 }
14235 define zeroext i32 @test_masked_vpcmpsgeq_v8i1_v32i1_mask_mem(i8 zeroext %__u, <8 x i64> %__a, <8 x i64>* %__b) local_unnamed_addr {
14236 ; VLX-LABEL: test_masked_vpcmpsgeq_v8i1_v32i1_mask_mem:
14237 ; VLX: # %bb.0: # %entry
14238 ; VLX-NEXT: kmovd %edi, %k1
14239 ; VLX-NEXT: vpcmpnltq (%rsi), %zmm0, %k0 {%k1}
14240 ; VLX-NEXT: kmovd %k0, %eax
14241 ; VLX-NEXT: vzeroupper
14242 ; VLX-NEXT: retq
14243 ;
14244 ; NoVLX-LABEL: test_masked_vpcmpsgeq_v8i1_v32i1_mask_mem:
14245 ; NoVLX: # %bb.0: # %entry
14246 ; NoVLX-NEXT: kmovw %edi, %k1
14247 ; NoVLX-NEXT: vpcmpnltq (%rsi), %zmm0, %k0 {%k1}
14248 ; NoVLX-NEXT: kmovw %k0, %eax
14249 ; NoVLX-NEXT: vzeroupper
14250 ; NoVLX-NEXT: retq
14251 entry:
14252 %0 = bitcast <8 x i64> %__a to <8 x i64>
14253 %load = load <8 x i64>, <8 x i64>* %__b
14254 %1 = bitcast <8 x i64> %load to <8 x i64>
14255 %2 = icmp sge <8 x i64> %0, %1
14256 %3 = bitcast i8 %__u to <8 x i1>
14257 %4 = and <8 x i1> %2, %3
14258 %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
14259 %6 = bitcast <32 x i1> %5 to i32
14260 ret i32 %6
14261 }
14264 define zeroext i32 @test_vpcmpsgeq_v8i1_v32i1_mask_mem_b(<8 x i64> %__a, i64* %__b) local_unnamed_addr {
14265 ; VLX-LABEL: test_vpcmpsgeq_v8i1_v32i1_mask_mem_b:
14266 ; VLX: # %bb.0: # %entry
14267 ; VLX-NEXT: vpcmpnltq (%rdi){1to8}, %zmm0, %k0
14268 ; VLX-NEXT: kmovd %k0, %eax
14269 ; VLX-NEXT: vzeroupper
14270 ; VLX-NEXT: retq
14271 ;
14272 ; NoVLX-LABEL: test_vpcmpsgeq_v8i1_v32i1_mask_mem_b:
14273 ; NoVLX: # %bb.0: # %entry
14274 ; NoVLX-NEXT: vpcmpnltq (%rdi){1to8}, %zmm0, %k0
14275 ; NoVLX-NEXT: kmovw %k0, %eax
14276 ; NoVLX-NEXT: vzeroupper
14277 ; NoVLX-NEXT: retq
14278 entry:
14279 %0 = bitcast <8 x i64> %__a to <8 x i64>
14280 %load = load i64, i64* %__b
14281 %vec = insertelement <8 x i64> undef, i64 %load, i32 0
14282 %1 = shufflevector <8 x i64> %vec, <8 x i64> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
14283 %2 = icmp sge <8 x i64> %0, %1
14284 %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
14285 %4 = bitcast <32 x i1> %3 to i32
14286 ret i32 %4
14287 }
14289 define zeroext i32 @test_masked_vpcmpsgeq_v8i1_v32i1_mask_mem_b(i8 zeroext %__u, <8 x i64> %__a, i64* %__b) local_unnamed_addr {
14290 ; VLX-LABEL: test_masked_vpcmpsgeq_v8i1_v32i1_mask_mem_b:
14291 ; VLX: # %bb.0: # %entry
14292 ; VLX-NEXT: kmovd %edi, %k1
14293 ; VLX-NEXT: vpcmpnltq (%rsi){1to8}, %zmm0, %k0 {%k1}
14294 ; VLX-NEXT: kmovd %k0, %eax
14295 ; VLX-NEXT: vzeroupper
14296 ; VLX-NEXT: retq
14297 ;
14298 ; NoVLX-LABEL: test_masked_vpcmpsgeq_v8i1_v32i1_mask_mem_b:
14299 ; NoVLX: # %bb.0: # %entry
14300 ; NoVLX-NEXT: kmovw %edi, %k1
14301 ; NoVLX-NEXT: vpcmpnltq (%rsi){1to8}, %zmm0, %k0 {%k1}
14302 ; NoVLX-NEXT: kmovw %k0, %eax
14303 ; NoVLX-NEXT: vzeroupper
14304 ; NoVLX-NEXT: retq
14305 entry:
14306 %0 = bitcast <8 x i64> %__a to <8 x i64>
14307 %load = load i64, i64* %__b
14308 %vec = insertelement <8 x i64> undef, i64 %load, i32 0
14309 %1 = shufflevector <8 x i64> %vec, <8 x i64> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
14310 %2 = icmp sge <8 x i64> %0, %1
14311 %3 = bitcast i8 %__u to <8 x i1>
14312 %4 = and <8 x i1> %3, %2
14313 %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
14314 %6 = bitcast <32 x i1> %5 to i32
14315 ret i32 %6
14316 }
14319 define zeroext i64 @test_vpcmpsgeq_v8i1_v64i1_mask(<8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
14320 ; VLX-LABEL: test_vpcmpsgeq_v8i1_v64i1_mask:
14321 ; VLX: # %bb.0: # %entry
14322 ; VLX-NEXT: vpcmpnltq %zmm1, %zmm0, %k0
14323 ; VLX-NEXT: kmovq %k0, %rax
14324 ; VLX-NEXT: vzeroupper
14325 ; VLX-NEXT: retq
14326 ;
14327 ; NoVLX-LABEL: test_vpcmpsgeq_v8i1_v64i1_mask:
14328 ; NoVLX: # %bb.0: # %entry
14329 ; NoVLX-NEXT: vpcmpnltq %zmm1, %zmm0, %k0
14330 ; NoVLX-NEXT: kmovw %k0, %eax
14331 ; NoVLX-NEXT: vzeroupper
14332 ; NoVLX-NEXT: retq
14333 entry:
14334 %0 = bitcast <8 x i64> %__a to <8 x i64>
14335 %1 = bitcast <8 x i64> %__b to <8 x i64>
14336 %2 = icmp sge <8 x i64> %0, %1
14337 %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
14338 %4 = bitcast <64 x i1> %3 to i64
14339 ret i64 %4
14340 }
14342 define zeroext i64 @test_vpcmpsgeq_v8i1_v64i1_mask_mem(<8 x i64> %__a, <8 x i64>* %__b) local_unnamed_addr {
14343 ; VLX-LABEL: test_vpcmpsgeq_v8i1_v64i1_mask_mem:
14344 ; VLX: # %bb.0: # %entry
14345 ; VLX-NEXT: vpcmpnltq (%rdi), %zmm0, %k0
14346 ; VLX-NEXT: kmovq %k0, %rax
14347 ; VLX-NEXT: vzeroupper
14348 ; VLX-NEXT: retq
14349 ;
14350 ; NoVLX-LABEL: test_vpcmpsgeq_v8i1_v64i1_mask_mem:
14351 ; NoVLX: # %bb.0: # %entry
14352 ; NoVLX-NEXT: vpcmpnltq (%rdi), %zmm0, %k0
14353 ; NoVLX-NEXT: kmovw %k0, %eax
14354 ; NoVLX-NEXT: vzeroupper
14355 ; NoVLX-NEXT: retq
14356 entry:
14357 %0 = bitcast <8 x i64> %__a to <8 x i64>
14358 %load = load <8 x i64>, <8 x i64>* %__b
14359 %1 = bitcast <8 x i64> %load to <8 x i64>
14360 %2 = icmp sge <8 x i64> %0, %1
14361 %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
14362 %4 = bitcast <64 x i1> %3 to i64
14363 ret i64 %4
14364 }
14366 define zeroext i64 @test_masked_vpcmpsgeq_v8i1_v64i1_mask(i8 zeroext %__u, <8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
14367 ; VLX-LABEL: test_masked_vpcmpsgeq_v8i1_v64i1_mask:
14368 ; VLX: # %bb.0: # %entry
14369 ; VLX-NEXT: kmovd %edi, %k1
14370 ; VLX-NEXT: vpcmpnltq %zmm1, %zmm0, %k0 {%k1}
14371 ; VLX-NEXT: kmovq %k0, %rax
14372 ; VLX-NEXT: vzeroupper
14373 ; VLX-NEXT: retq
14374 ;
14375 ; NoVLX-LABEL: test_masked_vpcmpsgeq_v8i1_v64i1_mask:
14376 ; NoVLX: # %bb.0: # %entry
14377 ; NoVLX-NEXT: kmovw %edi, %k1
14378 ; NoVLX-NEXT: vpcmpnltq %zmm1, %zmm0, %k0 {%k1}
14379 ; NoVLX-NEXT: kmovw %k0, %eax
14380 ; NoVLX-NEXT: vzeroupper
14381 ; NoVLX-NEXT: retq
14382 entry:
14383 %0 = bitcast <8 x i64> %__a to <8 x i64>
14384 %1 = bitcast <8 x i64> %__b to <8 x i64>
14385 %2 = icmp sge <8 x i64> %0, %1
14386 %3 = bitcast i8 %__u to <8 x i1>
14387 %4 = and <8 x i1> %2, %3
14388 %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
14389 %6 = bitcast <64 x i1> %5 to i64
14390 ret i64 %6
14391 }
14393 define zeroext i64 @test_masked_vpcmpsgeq_v8i1_v64i1_mask_mem(i8 zeroext %__u, <8 x i64> %__a, <8 x i64>* %__b) local_unnamed_addr {
14394 ; VLX-LABEL: test_masked_vpcmpsgeq_v8i1_v64i1_mask_mem:
14395 ; VLX: # %bb.0: # %entry
14396 ; VLX-NEXT: kmovd %edi, %k1
14397 ; VLX-NEXT: vpcmpnltq (%rsi), %zmm0, %k0 {%k1}
14398 ; VLX-NEXT: kmovq %k0, %rax
14399 ; VLX-NEXT: vzeroupper
14400 ; VLX-NEXT: retq
14401 ;
14402 ; NoVLX-LABEL: test_masked_vpcmpsgeq_v8i1_v64i1_mask_mem:
14403 ; NoVLX: # %bb.0: # %entry
14404 ; NoVLX-NEXT: kmovw %edi, %k1
14405 ; NoVLX-NEXT: vpcmpnltq (%rsi), %zmm0, %k0 {%k1}
14406 ; NoVLX-NEXT: kmovw %k0, %eax
14407 ; NoVLX-NEXT: vzeroupper
14408 ; NoVLX-NEXT: retq
14409 entry:
14410 %0 = bitcast <8 x i64> %__a to <8 x i64>
14411 %load = load <8 x i64>, <8 x i64>* %__b
14412 %1 = bitcast <8 x i64> %load to <8 x i64>
14413 %2 = icmp sge <8 x i64> %0, %1
14414 %3 = bitcast i8 %__u to <8 x i1>
14415 %4 = and <8 x i1> %2, %3
14416 %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
14417 %6 = bitcast <64 x i1> %5 to i64
14418 ret i64 %6
14419 }
14422 define zeroext i64 @test_vpcmpsgeq_v8i1_v64i1_mask_mem_b(<8 x i64> %__a, i64* %__b) local_unnamed_addr {
14423 ; VLX-LABEL: test_vpcmpsgeq_v8i1_v64i1_mask_mem_b:
14424 ; VLX: # %bb.0: # %entry
14425 ; VLX-NEXT: vpcmpnltq (%rdi){1to8}, %zmm0, %k0
14426 ; VLX-NEXT: kmovq %k0, %rax
14427 ; VLX-NEXT: vzeroupper
14428 ; VLX-NEXT: retq
14429 ;
14430 ; NoVLX-LABEL: test_vpcmpsgeq_v8i1_v64i1_mask_mem_b:
14431 ; NoVLX: # %bb.0: # %entry
14432 ; NoVLX-NEXT: vpcmpnltq (%rdi){1to8}, %zmm0, %k0
14433 ; NoVLX-NEXT: kmovw %k0, %eax
14434 ; NoVLX-NEXT: vzeroupper
14435 ; NoVLX-NEXT: retq
14436 entry:
14437 %0 = bitcast <8 x i64> %__a to <8 x i64>
14438 %load = load i64, i64* %__b
14439 %vec = insertelement <8 x i64> undef, i64 %load, i32 0
14440 %1 = shufflevector <8 x i64> %vec, <8 x i64> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
14441 %2 = icmp sge <8 x i64> %0, %1
14442 %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
14443 %4 = bitcast <64 x i1> %3 to i64
14444 ret i64 %4
14445 }
14447 define zeroext i64 @test_masked_vpcmpsgeq_v8i1_v64i1_mask_mem_b(i8 zeroext %__u, <8 x i64> %__a, i64* %__b) local_unnamed_addr {
14448 ; VLX-LABEL: test_masked_vpcmpsgeq_v8i1_v64i1_mask_mem_b:
14449 ; VLX: # %bb.0: # %entry
14450 ; VLX-NEXT: kmovd %edi, %k1
14451 ; VLX-NEXT: vpcmpnltq (%rsi){1to8}, %zmm0, %k0 {%k1}
14452 ; VLX-NEXT: kmovq %k0, %rax
14453 ; VLX-NEXT: vzeroupper
14454 ; VLX-NEXT: retq
14455 ;
14456 ; NoVLX-LABEL: test_masked_vpcmpsgeq_v8i1_v64i1_mask_mem_b:
14457 ; NoVLX: # %bb.0: # %entry
14458 ; NoVLX-NEXT: kmovw %edi, %k1
14459 ; NoVLX-NEXT: vpcmpnltq (%rsi){1to8}, %zmm0, %k0 {%k1}
14460 ; NoVLX-NEXT: kmovw %k0, %eax
14461 ; NoVLX-NEXT: vzeroupper
14462 ; NoVLX-NEXT: retq
14463 entry:
14464 %0 = bitcast <8 x i64> %__a to <8 x i64>
14465 %load = load i64, i64* %__b
14466 %vec = insertelement <8 x i64> undef, i64 %load, i32 0
14467 %1 = shufflevector <8 x i64> %vec, <8 x i64> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
14468 %2 = icmp sge <8 x i64> %0, %1
14469 %3 = bitcast i8 %__u to <8 x i1>
14470 %4 = and <8 x i1> %3, %2
14471 %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
14472 %6 = bitcast <64 x i1> %5 to i64
14473 ret i64 %6
14474 }
14477 define zeroext i32 @test_vpcmpultb_v16i1_v32i1_mask(<2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
14478 ; VLX-LABEL: test_vpcmpultb_v16i1_v32i1_mask:
14479 ; VLX: # %bb.0: # %entry
14480 ; VLX-NEXT: vpcmpltub %xmm1, %xmm0, %k0
14481 ; VLX-NEXT: kmovd %k0, %eax
14482 ; VLX-NEXT: retq
14483 ;
14484 ; NoVLX-LABEL: test_vpcmpultb_v16i1_v32i1_mask:
14485 ; NoVLX: # %bb.0: # %entry
14486 ; NoVLX-NEXT: vpmaxub %xmm1, %xmm0, %xmm1
14487 ; NoVLX-NEXT: vpcmpeqb %xmm1, %xmm0, %xmm0
14488 ; NoVLX-NEXT: vpternlogq $15, %zmm0, %zmm0, %zmm0
14489 ; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0
14490 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
14491 ; NoVLX-NEXT: kmovw %k0, %eax
14492 ; NoVLX-NEXT: vzeroupper
14493 ; NoVLX-NEXT: retq
14494 entry:
14495 %0 = bitcast <2 x i64> %__a to <16 x i8>
14496 %1 = bitcast <2 x i64> %__b to <16 x i8>
14497 %2 = icmp ult <16 x i8> %0, %1
14498 %3 = shufflevector <16 x i1> %2, <16 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>
14499 %4 = bitcast <32 x i1> %3 to i32
14500 ret i32 %4
14501 }
14503 define zeroext i32 @test_vpcmpultb_v16i1_v32i1_mask_mem(<2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
14504 ; VLX-LABEL: test_vpcmpultb_v16i1_v32i1_mask_mem:
14505 ; VLX: # %bb.0: # %entry
14506 ; VLX-NEXT: vpcmpltub (%rdi), %xmm0, %k0
14507 ; VLX-NEXT: kmovd %k0, %eax
14508 ; VLX-NEXT: retq
14509 ;
14510 ; NoVLX-LABEL: test_vpcmpultb_v16i1_v32i1_mask_mem:
14511 ; NoVLX: # %bb.0: # %entry
14512 ; NoVLX-NEXT: vpmaxub (%rdi), %xmm0, %xmm1
14513 ; NoVLX-NEXT: vpcmpeqb %xmm1, %xmm0, %xmm0
14514 ; NoVLX-NEXT: vpternlogq $15, %zmm0, %zmm0, %zmm0
14515 ; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0
14516 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
14517 ; NoVLX-NEXT: kmovw %k0, %eax
14518 ; NoVLX-NEXT: vzeroupper
14519 ; NoVLX-NEXT: retq
14520 entry:
14521 %0 = bitcast <2 x i64> %__a to <16 x i8>
14522 %load = load <2 x i64>, <2 x i64>* %__b
14523 %1 = bitcast <2 x i64> %load to <16 x i8>
14524 %2 = icmp ult <16 x i8> %0, %1
14525 %3 = shufflevector <16 x i1> %2, <16 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>
14526 %4 = bitcast <32 x i1> %3 to i32
14527 ret i32 %4
14528 }
14530 define zeroext i32 @test_masked_vpcmpultb_v16i1_v32i1_mask(i16 zeroext %__u, <2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
14531 ; VLX-LABEL: test_masked_vpcmpultb_v16i1_v32i1_mask:
14532 ; VLX: # %bb.0: # %entry
14533 ; VLX-NEXT: kmovd %edi, %k1
14534 ; VLX-NEXT: vpcmpltub %xmm1, %xmm0, %k0 {%k1}
14535 ; VLX-NEXT: kmovd %k0, %eax
14536 ; VLX-NEXT: retq
14537 ;
14538 ; NoVLX-LABEL: test_masked_vpcmpultb_v16i1_v32i1_mask:
14539 ; NoVLX: # %bb.0: # %entry
14540 ; NoVLX-NEXT: vpmaxub %xmm1, %xmm0, %xmm1
14541 ; NoVLX-NEXT: vpcmpeqb %xmm1, %xmm0, %xmm0
14542 ; NoVLX-NEXT: vpternlogq $15, %zmm0, %zmm0, %zmm0
14543 ; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0
14544 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
14545 ; NoVLX-NEXT: kmovw %k0, %eax
14546 ; NoVLX-NEXT: andl %edi, %eax
14547 ; NoVLX-NEXT: vzeroupper
14548 ; NoVLX-NEXT: retq
14549 entry:
14550 %0 = bitcast <2 x i64> %__a to <16 x i8>
14551 %1 = bitcast <2 x i64> %__b to <16 x i8>
14552 %2 = icmp ult <16 x i8> %0, %1
14553 %3 = bitcast i16 %__u to <16 x i1>
14554 %4 = and <16 x i1> %2, %3
14555 %5 = shufflevector <16 x i1> %4, <16 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>
14556 %6 = bitcast <32 x i1> %5 to i32
14557 ret i32 %6
14558 }
14560 define zeroext i32 @test_masked_vpcmpultb_v16i1_v32i1_mask_mem(i16 zeroext %__u, <2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
14561 ; VLX-LABEL: test_masked_vpcmpultb_v16i1_v32i1_mask_mem:
14562 ; VLX: # %bb.0: # %entry
14563 ; VLX-NEXT: kmovd %edi, %k1
14564 ; VLX-NEXT: vpcmpltub (%rsi), %xmm0, %k0 {%k1}
14565 ; VLX-NEXT: kmovd %k0, %eax
14566 ; VLX-NEXT: retq
14567 ;
14568 ; NoVLX-LABEL: test_masked_vpcmpultb_v16i1_v32i1_mask_mem:
14569 ; NoVLX: # %bb.0: # %entry
14570 ; NoVLX-NEXT: vpmaxub (%rsi), %xmm0, %xmm1
14571 ; NoVLX-NEXT: vpcmpeqb %xmm1, %xmm0, %xmm0
14572 ; NoVLX-NEXT: vpternlogq $15, %zmm0, %zmm0, %zmm0
14573 ; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0
14574 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
14575 ; NoVLX-NEXT: kmovw %k0, %eax
14576 ; NoVLX-NEXT: andl %edi, %eax
14577 ; NoVLX-NEXT: vzeroupper
14578 ; NoVLX-NEXT: retq
14579 entry:
14580 %0 = bitcast <2 x i64> %__a to <16 x i8>
14581 %load = load <2 x i64>, <2 x i64>* %__b
14582 %1 = bitcast <2 x i64> %load to <16 x i8>
14583 %2 = icmp ult <16 x i8> %0, %1
14584 %3 = bitcast i16 %__u to <16 x i1>
14585 %4 = and <16 x i1> %2, %3
14586 %5 = shufflevector <16 x i1> %4, <16 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>
14587 %6 = bitcast <32 x i1> %5 to i32
14588 ret i32 %6
14589 }
14592 define zeroext i64 @test_vpcmpultb_v16i1_v64i1_mask(<2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
14593 ; VLX-LABEL: test_vpcmpultb_v16i1_v64i1_mask:
14594 ; VLX: # %bb.0: # %entry
14595 ; VLX-NEXT: vpcmpltub %xmm1, %xmm0, %k0
14596 ; VLX-NEXT: kmovq %k0, %rax
14597 ; VLX-NEXT: retq
14598 ;
14599 ; NoVLX-LABEL: test_vpcmpultb_v16i1_v64i1_mask:
14600 ; NoVLX: # %bb.0: # %entry
14601 ; NoVLX-NEXT: vpmaxub %xmm1, %xmm0, %xmm1
14602 ; NoVLX-NEXT: vpcmpeqb %xmm1, %xmm0, %xmm0
14603 ; NoVLX-NEXT: vpternlogq $15, %zmm0, %zmm0, %zmm0
14604 ; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0
14605 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
14606 ; NoVLX-NEXT: kmovw %k0, %eax
14607 ; NoVLX-NEXT: vzeroupper
14608 ; NoVLX-NEXT: retq
14609 entry:
14610 %0 = bitcast <2 x i64> %__a to <16 x i8>
14611 %1 = bitcast <2 x i64> %__b to <16 x i8>
14612 %2 = icmp ult <16 x i8> %0, %1
14613 %3 = shufflevector <16 x i1> %2, <16 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>
14614 %4 = bitcast <64 x i1> %3 to i64
14615 ret i64 %4
14616 }
14618 define zeroext i64 @test_vpcmpultb_v16i1_v64i1_mask_mem(<2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
14619 ; VLX-LABEL: test_vpcmpultb_v16i1_v64i1_mask_mem:
14620 ; VLX: # %bb.0: # %entry
14621 ; VLX-NEXT: vpcmpltub (%rdi), %xmm0, %k0
14622 ; VLX-NEXT: kmovq %k0, %rax
14623 ; VLX-NEXT: retq
14624 ;
14625 ; NoVLX-LABEL: test_vpcmpultb_v16i1_v64i1_mask_mem:
14626 ; NoVLX: # %bb.0: # %entry
14627 ; NoVLX-NEXT: vpmaxub (%rdi), %xmm0, %xmm1
14628 ; NoVLX-NEXT: vpcmpeqb %xmm1, %xmm0, %xmm0
14629 ; NoVLX-NEXT: vpternlogq $15, %zmm0, %zmm0, %zmm0
14630 ; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0
14631 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
14632 ; NoVLX-NEXT: kmovw %k0, %eax
14633 ; NoVLX-NEXT: vzeroupper
14634 ; NoVLX-NEXT: retq
14635 entry:
14636 %0 = bitcast <2 x i64> %__a to <16 x i8>
14637 %load = load <2 x i64>, <2 x i64>* %__b
14638 %1 = bitcast <2 x i64> %load to <16 x i8>
14639 %2 = icmp ult <16 x i8> %0, %1
14640 %3 = shufflevector <16 x i1> %2, <16 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>
14641 %4 = bitcast <64 x i1> %3 to i64
14642 ret i64 %4
14643 }
14645 define zeroext i64 @test_masked_vpcmpultb_v16i1_v64i1_mask(i16 zeroext %__u, <2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
14646 ; VLX-LABEL: test_masked_vpcmpultb_v16i1_v64i1_mask:
14647 ; VLX: # %bb.0: # %entry
14648 ; VLX-NEXT: kmovd %edi, %k1
14649 ; VLX-NEXT: vpcmpltub %xmm1, %xmm0, %k0 {%k1}
14650 ; VLX-NEXT: kmovq %k0, %rax
14651 ; VLX-NEXT: retq
14652 ;
14653 ; NoVLX-LABEL: test_masked_vpcmpultb_v16i1_v64i1_mask:
14654 ; NoVLX: # %bb.0: # %entry
14655 ; NoVLX-NEXT: vpmaxub %xmm1, %xmm0, %xmm1
14656 ; NoVLX-NEXT: vpcmpeqb %xmm1, %xmm0, %xmm0
14657 ; NoVLX-NEXT: vpternlogq $15, %zmm0, %zmm0, %zmm0
14658 ; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0
14659 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
14660 ; NoVLX-NEXT: kmovw %k0, %eax
14661 ; NoVLX-NEXT: andl %edi, %eax
14662 ; NoVLX-NEXT: vzeroupper
14663 ; NoVLX-NEXT: retq
14664 entry:
14665 %0 = bitcast <2 x i64> %__a to <16 x i8>
14666 %1 = bitcast <2 x i64> %__b to <16 x i8>
14667 %2 = icmp ult <16 x i8> %0, %1
14668 %3 = bitcast i16 %__u to <16 x i1>
14669 %4 = and <16 x i1> %2, %3
14670 %5 = shufflevector <16 x i1> %4, <16 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>
14671 %6 = bitcast <64 x i1> %5 to i64
14672 ret i64 %6
14673 }
14675 define zeroext i64 @test_masked_vpcmpultb_v16i1_v64i1_mask_mem(i16 zeroext %__u, <2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
14676 ; VLX-LABEL: test_masked_vpcmpultb_v16i1_v64i1_mask_mem:
14677 ; VLX: # %bb.0: # %entry
14678 ; VLX-NEXT: kmovd %edi, %k1
14679 ; VLX-NEXT: vpcmpltub (%rsi), %xmm0, %k0 {%k1}
14680 ; VLX-NEXT: kmovq %k0, %rax
14681 ; VLX-NEXT: retq
14682 ;
14683 ; NoVLX-LABEL: test_masked_vpcmpultb_v16i1_v64i1_mask_mem:
14684 ; NoVLX: # %bb.0: # %entry
14685 ; NoVLX-NEXT: vpmaxub (%rsi), %xmm0, %xmm1
14686 ; NoVLX-NEXT: vpcmpeqb %xmm1, %xmm0, %xmm0
14687 ; NoVLX-NEXT: vpternlogq $15, %zmm0, %zmm0, %zmm0
14688 ; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0
14689 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
14690 ; NoVLX-NEXT: kmovw %k0, %eax
14691 ; NoVLX-NEXT: andl %edi, %eax
14692 ; NoVLX-NEXT: vzeroupper
14693 ; NoVLX-NEXT: retq
14694 entry:
14695 %0 = bitcast <2 x i64> %__a to <16 x i8>
14696 %load = load <2 x i64>, <2 x i64>* %__b
14697 %1 = bitcast <2 x i64> %load to <16 x i8>
14698 %2 = icmp ult <16 x i8> %0, %1
14699 %3 = bitcast i16 %__u to <16 x i1>
14700 %4 = and <16 x i1> %2, %3
14701 %5 = shufflevector <16 x i1> %4, <16 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>
14702 %6 = bitcast <64 x i1> %5 to i64
14703 ret i64 %6
14704 }
14707 define zeroext i64 @test_vpcmpultb_v32i1_v64i1_mask(<4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
14708 ; VLX-LABEL: test_vpcmpultb_v32i1_v64i1_mask:
14709 ; VLX: # %bb.0: # %entry
14710 ; VLX-NEXT: vpcmpltub %ymm1, %ymm0, %k0
14711 ; VLX-NEXT: kmovq %k0, %rax
14712 ; VLX-NEXT: vzeroupper
14713 ; VLX-NEXT: retq
14714 ;
14715 ; NoVLX-LABEL: test_vpcmpultb_v32i1_v64i1_mask:
14716 ; NoVLX: # %bb.0: # %entry
14717 ; NoVLX-NEXT: vpmaxub %ymm1, %ymm0, %ymm1
14718 ; NoVLX-NEXT: vpcmpeqb %ymm1, %ymm0, %ymm0
14719 ; NoVLX-NEXT: vpternlogq $15, %zmm0, %zmm0, %zmm0
14720 ; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm1
14721 ; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
14722 ; NoVLX-NEXT: kmovw %k0, %ecx
14723 ; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm0
14724 ; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0
14725 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
14726 ; NoVLX-NEXT: kmovw %k0, %eax
14727 ; NoVLX-NEXT: shll $16, %eax
14728 ; NoVLX-NEXT: orl %ecx, %eax
14729 ; NoVLX-NEXT: vzeroupper
14730 ; NoVLX-NEXT: retq
14731 entry:
14732 %0 = bitcast <4 x i64> %__a to <32 x i8>
14733 %1 = bitcast <4 x i64> %__b to <32 x i8>
14734 %2 = icmp ult <32 x i8> %0, %1
14735 %3 = shufflevector <32 x i1> %2, <32 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63>
14736 %4 = bitcast <64 x i1> %3 to i64
14737 ret i64 %4
14738 }
14740 define zeroext i64 @test_vpcmpultb_v32i1_v64i1_mask_mem(<4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
14741 ; VLX-LABEL: test_vpcmpultb_v32i1_v64i1_mask_mem:
14742 ; VLX: # %bb.0: # %entry
14743 ; VLX-NEXT: vpcmpltub (%rdi), %ymm0, %k0
14744 ; VLX-NEXT: kmovq %k0, %rax
14745 ; VLX-NEXT: vzeroupper
14746 ; VLX-NEXT: retq
14747 ;
14748 ; NoVLX-LABEL: test_vpcmpultb_v32i1_v64i1_mask_mem:
14749 ; NoVLX: # %bb.0: # %entry
14750 ; NoVLX-NEXT: vpmaxub (%rdi), %ymm0, %ymm1
14751 ; NoVLX-NEXT: vpcmpeqb %ymm1, %ymm0, %ymm0
14752 ; NoVLX-NEXT: vpternlogq $15, %zmm0, %zmm0, %zmm0
14753 ; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm1
14754 ; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
14755 ; NoVLX-NEXT: kmovw %k0, %ecx
14756 ; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm0
14757 ; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0
14758 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
14759 ; NoVLX-NEXT: kmovw %k0, %eax
14760 ; NoVLX-NEXT: shll $16, %eax
14761 ; NoVLX-NEXT: orl %ecx, %eax
14762 ; NoVLX-NEXT: vzeroupper
14763 ; NoVLX-NEXT: retq
14764 entry:
14765 %0 = bitcast <4 x i64> %__a to <32 x i8>
14766 %load = load <4 x i64>, <4 x i64>* %__b
14767 %1 = bitcast <4 x i64> %load to <32 x i8>
14768 %2 = icmp ult <32 x i8> %0, %1
14769 %3 = shufflevector <32 x i1> %2, <32 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63>
14770 %4 = bitcast <64 x i1> %3 to i64
14771 ret i64 %4
14772 }
14774 define zeroext i64 @test_masked_vpcmpultb_v32i1_v64i1_mask(i32 zeroext %__u, <4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
14775 ; VLX-LABEL: test_masked_vpcmpultb_v32i1_v64i1_mask:
14776 ; VLX: # %bb.0: # %entry
14777 ; VLX-NEXT: kmovd %edi, %k1
14778 ; VLX-NEXT: vpcmpltub %ymm1, %ymm0, %k0 {%k1}
14779 ; VLX-NEXT: kmovq %k0, %rax
14780 ; VLX-NEXT: vzeroupper
14781 ; VLX-NEXT: retq
14782 ;
14783 ; NoVLX-LABEL: test_masked_vpcmpultb_v32i1_v64i1_mask:
14784 ; NoVLX: # %bb.0: # %entry
14785 ; NoVLX-NEXT: vpmaxub %ymm1, %ymm0, %ymm1
14786 ; NoVLX-NEXT: vpcmpeqb %ymm1, %ymm0, %ymm0
14787 ; NoVLX-NEXT: vpternlogq $15, %zmm0, %zmm0, %zmm0
14788 ; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm1
14789 ; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
14790 ; NoVLX-NEXT: kmovw %k0, %eax
14791 ; NoVLX-NEXT: andl %edi, %eax
14792 ; NoVLX-NEXT: shrl $16, %edi
14793 ; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm0
14794 ; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0
14795 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
14796 ; NoVLX-NEXT: kmovw %k0, %ecx
14797 ; NoVLX-NEXT: andl %edi, %ecx
14798 ; NoVLX-NEXT: shll $16, %ecx
14799 ; NoVLX-NEXT: movzwl %ax, %eax
14800 ; NoVLX-NEXT: orl %ecx, %eax
14801 ; NoVLX-NEXT: vzeroupper
14802 ; NoVLX-NEXT: retq
14803 entry:
14804 %0 = bitcast <4 x i64> %__a to <32 x i8>
14805 %1 = bitcast <4 x i64> %__b to <32 x i8>
14806 %2 = icmp ult <32 x i8> %0, %1
14807 %3 = bitcast i32 %__u to <32 x i1>
14808 %4 = and <32 x i1> %2, %3
14809 %5 = shufflevector <32 x i1> %4, <32 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63>
14810 %6 = bitcast <64 x i1> %5 to i64
14811 ret i64 %6
14812 }
14814 define zeroext i64 @test_masked_vpcmpultb_v32i1_v64i1_mask_mem(i32 zeroext %__u, <4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
14815 ; VLX-LABEL: test_masked_vpcmpultb_v32i1_v64i1_mask_mem:
14816 ; VLX: # %bb.0: # %entry
14817 ; VLX-NEXT: kmovd %edi, %k1
14818 ; VLX-NEXT: vpcmpltub (%rsi), %ymm0, %k0 {%k1}
14819 ; VLX-NEXT: kmovq %k0, %rax
14820 ; VLX-NEXT: vzeroupper
14821 ; VLX-NEXT: retq
14822 ;
14823 ; NoVLX-LABEL: test_masked_vpcmpultb_v32i1_v64i1_mask_mem:
14824 ; NoVLX: # %bb.0: # %entry
14825 ; NoVLX-NEXT: vpmaxub (%rsi), %ymm0, %ymm1
14826 ; NoVLX-NEXT: vpcmpeqb %ymm1, %ymm0, %ymm0
14827 ; NoVLX-NEXT: vpternlogq $15, %zmm0, %zmm0, %zmm0
14828 ; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm1
14829 ; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
14830 ; NoVLX-NEXT: kmovw %k0, %eax
14831 ; NoVLX-NEXT: andl %edi, %eax
14832 ; NoVLX-NEXT: shrl $16, %edi
14833 ; NoVLX-NEXT: vextracti128 $1, %ymm0, %xmm0
14834 ; NoVLX-NEXT: vpmovsxbd %xmm0, %zmm0
14835 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
14836 ; NoVLX-NEXT: kmovw %k0, %ecx
14837 ; NoVLX-NEXT: andl %edi, %ecx
14838 ; NoVLX-NEXT: shll $16, %ecx
14839 ; NoVLX-NEXT: movzwl %ax, %eax
14840 ; NoVLX-NEXT: orl %ecx, %eax
14841 ; NoVLX-NEXT: vzeroupper
14842 ; NoVLX-NEXT: retq
14843 entry:
14844 %0 = bitcast <4 x i64> %__a to <32 x i8>
14845 %load = load <4 x i64>, <4 x i64>* %__b
14846 %1 = bitcast <4 x i64> %load to <32 x i8>
14847 %2 = icmp ult <32 x i8> %0, %1
14848 %3 = bitcast i32 %__u to <32 x i1>
14849 %4 = and <32 x i1> %2, %3
14850 %5 = shufflevector <32 x i1> %4, <32 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63>
14851 %6 = bitcast <64 x i1> %5 to i64
14852 ret i64 %6
14853 }
14856 define zeroext i16 @test_vpcmpultw_v8i1_v16i1_mask(<2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
14857 ; VLX-LABEL: test_vpcmpultw_v8i1_v16i1_mask:
14858 ; VLX: # %bb.0: # %entry
14859 ; VLX-NEXT: vpcmpltuw %xmm1, %xmm0, %k0
14860 ; VLX-NEXT: kmovd %k0, %eax
14861 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
14862 ; VLX-NEXT: retq
14863 ;
14864 ; NoVLX-LABEL: test_vpcmpultw_v8i1_v16i1_mask:
14865 ; NoVLX: # %bb.0: # %entry
14866 ; NoVLX-NEXT: vpmaxuw %xmm1, %xmm0, %xmm1
14867 ; NoVLX-NEXT: vpcmpeqw %xmm1, %xmm0, %xmm0
14868 ; NoVLX-NEXT: vpternlogq $15, %zmm0, %zmm0, %zmm0
14869 ; NoVLX-NEXT: vpmovsxwq %xmm0, %zmm0
14870 ; NoVLX-NEXT: vptestmq %zmm0, %zmm0, %k0
14871 ; NoVLX-NEXT: kmovw %k0, %eax
14872 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
14873 ; NoVLX-NEXT: vzeroupper
14874 ; NoVLX-NEXT: retq
14875 entry:
14876 %0 = bitcast <2 x i64> %__a to <8 x i16>
14877 %1 = bitcast <2 x i64> %__b to <8 x i16>
14878 %2 = icmp ult <8 x i16> %0, %1
14879 %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
14880 %4 = bitcast <16 x i1> %3 to i16
14881 ret i16 %4
14882 }
14884 define zeroext i16 @test_vpcmpultw_v8i1_v16i1_mask_mem(<2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
14885 ; VLX-LABEL: test_vpcmpultw_v8i1_v16i1_mask_mem:
14886 ; VLX: # %bb.0: # %entry
14887 ; VLX-NEXT: vpcmpltuw (%rdi), %xmm0, %k0
14888 ; VLX-NEXT: kmovd %k0, %eax
14889 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
14890 ; VLX-NEXT: retq
14891 ;
14892 ; NoVLX-LABEL: test_vpcmpultw_v8i1_v16i1_mask_mem:
14893 ; NoVLX: # %bb.0: # %entry
14894 ; NoVLX-NEXT: vpmaxuw (%rdi), %xmm0, %xmm1
14895 ; NoVLX-NEXT: vpcmpeqw %xmm1, %xmm0, %xmm0
14896 ; NoVLX-NEXT: vpternlogq $15, %zmm0, %zmm0, %zmm0
14897 ; NoVLX-NEXT: vpmovsxwq %xmm0, %zmm0
14898 ; NoVLX-NEXT: vptestmq %zmm0, %zmm0, %k0
14899 ; NoVLX-NEXT: kmovw %k0, %eax
14900 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
14901 ; NoVLX-NEXT: vzeroupper
14902 ; NoVLX-NEXT: retq
14903 entry:
14904 %0 = bitcast <2 x i64> %__a to <8 x i16>
14905 %load = load <2 x i64>, <2 x i64>* %__b
14906 %1 = bitcast <2 x i64> %load to <8 x i16>
14907 %2 = icmp ult <8 x i16> %0, %1
14908 %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
14909 %4 = bitcast <16 x i1> %3 to i16
14910 ret i16 %4
14911 }
14913 define zeroext i16 @test_masked_vpcmpultw_v8i1_v16i1_mask(i8 zeroext %__u, <2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
14914 ; VLX-LABEL: test_masked_vpcmpultw_v8i1_v16i1_mask:
14915 ; VLX: # %bb.0: # %entry
14916 ; VLX-NEXT: kmovd %edi, %k1
14917 ; VLX-NEXT: vpcmpltuw %xmm1, %xmm0, %k0 {%k1}
14918 ; VLX-NEXT: kmovd %k0, %eax
14919 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
14920 ; VLX-NEXT: retq
14921 ;
14922 ; NoVLX-LABEL: test_masked_vpcmpultw_v8i1_v16i1_mask:
14923 ; NoVLX: # %bb.0: # %entry
14924 ; NoVLX-NEXT: vpmaxuw %xmm1, %xmm0, %xmm1
14925 ; NoVLX-NEXT: vpcmpeqw %xmm1, %xmm0, %xmm0
14926 ; NoVLX-NEXT: vpternlogq $15, %zmm0, %zmm0, %zmm0
14927 ; NoVLX-NEXT: vpmovsxwq %xmm0, %zmm0
14928 ; NoVLX-NEXT: kmovw %edi, %k1
14929 ; NoVLX-NEXT: vptestmq %zmm0, %zmm0, %k0 {%k1}
14930 ; NoVLX-NEXT: kmovw %k0, %eax
14931 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
14932 ; NoVLX-NEXT: vzeroupper
14933 ; NoVLX-NEXT: retq
14934 entry:
14935 %0 = bitcast <2 x i64> %__a to <8 x i16>
14936 %1 = bitcast <2 x i64> %__b to <8 x i16>
14937 %2 = icmp ult <8 x i16> %0, %1
14938 %3 = bitcast i8 %__u to <8 x i1>
14939 %4 = and <8 x i1> %2, %3
14940 %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
14941 %6 = bitcast <16 x i1> %5 to i16
14942 ret i16 %6
14943 }
14945 define zeroext i16 @test_masked_vpcmpultw_v8i1_v16i1_mask_mem(i8 zeroext %__u, <2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
14946 ; VLX-LABEL: test_masked_vpcmpultw_v8i1_v16i1_mask_mem:
14947 ; VLX: # %bb.0: # %entry
14948 ; VLX-NEXT: kmovd %edi, %k1
14949 ; VLX-NEXT: vpcmpltuw (%rsi), %xmm0, %k0 {%k1}
14950 ; VLX-NEXT: kmovd %k0, %eax
14951 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
14952 ; VLX-NEXT: retq
14953 ;
14954 ; NoVLX-LABEL: test_masked_vpcmpultw_v8i1_v16i1_mask_mem:
14955 ; NoVLX: # %bb.0: # %entry
14956 ; NoVLX-NEXT: vpmaxuw (%rsi), %xmm0, %xmm1
14957 ; NoVLX-NEXT: vpcmpeqw %xmm1, %xmm0, %xmm0
14958 ; NoVLX-NEXT: vpternlogq $15, %zmm0, %zmm0, %zmm0
14959 ; NoVLX-NEXT: vpmovsxwq %xmm0, %zmm0
14960 ; NoVLX-NEXT: kmovw %edi, %k1
14961 ; NoVLX-NEXT: vptestmq %zmm0, %zmm0, %k0 {%k1}
14962 ; NoVLX-NEXT: kmovw %k0, %eax
14963 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
14964 ; NoVLX-NEXT: vzeroupper
14965 ; NoVLX-NEXT: retq
14966 entry:
14967 %0 = bitcast <2 x i64> %__a to <8 x i16>
14968 %load = load <2 x i64>, <2 x i64>* %__b
14969 %1 = bitcast <2 x i64> %load to <8 x i16>
14970 %2 = icmp ult <8 x i16> %0, %1
14971 %3 = bitcast i8 %__u to <8 x i1>
14972 %4 = and <8 x i1> %2, %3
14973 %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
14974 %6 = bitcast <16 x i1> %5 to i16
14975 ret i16 %6
14976 }
14979 define zeroext i32 @test_vpcmpultw_v8i1_v32i1_mask(<2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
14980 ; VLX-LABEL: test_vpcmpultw_v8i1_v32i1_mask:
14981 ; VLX: # %bb.0: # %entry
14982 ; VLX-NEXT: vpcmpltuw %xmm1, %xmm0, %k0
14983 ; VLX-NEXT: kmovd %k0, %eax
14984 ; VLX-NEXT: retq
14985 ;
14986 ; NoVLX-LABEL: test_vpcmpultw_v8i1_v32i1_mask:
14987 ; NoVLX: # %bb.0: # %entry
14988 ; NoVLX-NEXT: vpmaxuw %xmm1, %xmm0, %xmm1
14989 ; NoVLX-NEXT: vpcmpeqw %xmm1, %xmm0, %xmm0
14990 ; NoVLX-NEXT: vpternlogq $15, %zmm0, %zmm0, %zmm0
14991 ; NoVLX-NEXT: vpmovsxwq %xmm0, %zmm0
14992 ; NoVLX-NEXT: vptestmq %zmm0, %zmm0, %k0
14993 ; NoVLX-NEXT: kmovw %k0, %eax
14994 ; NoVLX-NEXT: vzeroupper
14995 ; NoVLX-NEXT: retq
14996 entry:
14997 %0 = bitcast <2 x i64> %__a to <8 x i16>
14998 %1 = bitcast <2 x i64> %__b to <8 x i16>
14999 %2 = icmp ult <8 x i16> %0, %1
15000 %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
15001 %4 = bitcast <32 x i1> %3 to i32
15002 ret i32 %4
15003 }
15005 define zeroext i32 @test_vpcmpultw_v8i1_v32i1_mask_mem(<2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
15006 ; VLX-LABEL: test_vpcmpultw_v8i1_v32i1_mask_mem:
15007 ; VLX: # %bb.0: # %entry
15008 ; VLX-NEXT: vpcmpltuw (%rdi), %xmm0, %k0
15009 ; VLX-NEXT: kmovd %k0, %eax
15010 ; VLX-NEXT: retq
15011 ;
15012 ; NoVLX-LABEL: test_vpcmpultw_v8i1_v32i1_mask_mem:
15013 ; NoVLX: # %bb.0: # %entry
15014 ; NoVLX-NEXT: vpmaxuw (%rdi), %xmm0, %xmm1
15015 ; NoVLX-NEXT: vpcmpeqw %xmm1, %xmm0, %xmm0
15016 ; NoVLX-NEXT: vpternlogq $15, %zmm0, %zmm0, %zmm0
15017 ; NoVLX-NEXT: vpmovsxwq %xmm0, %zmm0
15018 ; NoVLX-NEXT: vptestmq %zmm0, %zmm0, %k0
15019 ; NoVLX-NEXT: kmovw %k0, %eax
15020 ; NoVLX-NEXT: vzeroupper
15021 ; NoVLX-NEXT: retq
15022 entry:
15023 %0 = bitcast <2 x i64> %__a to <8 x i16>
15024 %load = load <2 x i64>, <2 x i64>* %__b
15025 %1 = bitcast <2 x i64> %load to <8 x i16>
15026 %2 = icmp ult <8 x i16> %0, %1
15027 %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
15028 %4 = bitcast <32 x i1> %3 to i32
15029 ret i32 %4
15030 }
15032 define zeroext i32 @test_masked_vpcmpultw_v8i1_v32i1_mask(i8 zeroext %__u, <2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
15033 ; VLX-LABEL: test_masked_vpcmpultw_v8i1_v32i1_mask:
15034 ; VLX: # %bb.0: # %entry
15035 ; VLX-NEXT: kmovd %edi, %k1
15036 ; VLX-NEXT: vpcmpltuw %xmm1, %xmm0, %k0 {%k1}
15037 ; VLX-NEXT: kmovd %k0, %eax
15038 ; VLX-NEXT: retq
15039 ;
15040 ; NoVLX-LABEL: test_masked_vpcmpultw_v8i1_v32i1_mask:
15041 ; NoVLX: # %bb.0: # %entry
15042 ; NoVLX-NEXT: vpmaxuw %xmm1, %xmm0, %xmm1
15043 ; NoVLX-NEXT: vpcmpeqw %xmm1, %xmm0, %xmm0
15044 ; NoVLX-NEXT: vpternlogq $15, %zmm0, %zmm0, %zmm0
15045 ; NoVLX-NEXT: vpmovsxwq %xmm0, %zmm0
15046 ; NoVLX-NEXT: kmovw %edi, %k1
15047 ; NoVLX-NEXT: vptestmq %zmm0, %zmm0, %k0 {%k1}
15048 ; NoVLX-NEXT: kmovw %k0, %eax
15049 ; NoVLX-NEXT: vzeroupper
15050 ; NoVLX-NEXT: retq
15051 entry:
15052 %0 = bitcast <2 x i64> %__a to <8 x i16>
15053 %1 = bitcast <2 x i64> %__b to <8 x i16>
15054 %2 = icmp ult <8 x i16> %0, %1
15055 %3 = bitcast i8 %__u to <8 x i1>
15056 %4 = and <8 x i1> %2, %3
15057 %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
15058 %6 = bitcast <32 x i1> %5 to i32
15059 ret i32 %6
15060 }
15062 define zeroext i32 @test_masked_vpcmpultw_v8i1_v32i1_mask_mem(i8 zeroext %__u, <2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
15063 ; VLX-LABEL: test_masked_vpcmpultw_v8i1_v32i1_mask_mem:
15064 ; VLX: # %bb.0: # %entry
15065 ; VLX-NEXT: kmovd %edi, %k1
15066 ; VLX-NEXT: vpcmpltuw (%rsi), %xmm0, %k0 {%k1}
15067 ; VLX-NEXT: kmovd %k0, %eax
15068 ; VLX-NEXT: retq
15069 ;
15070 ; NoVLX-LABEL: test_masked_vpcmpultw_v8i1_v32i1_mask_mem:
15071 ; NoVLX: # %bb.0: # %entry
15072 ; NoVLX-NEXT: vpmaxuw (%rsi), %xmm0, %xmm1
15073 ; NoVLX-NEXT: vpcmpeqw %xmm1, %xmm0, %xmm0
15074 ; NoVLX-NEXT: vpternlogq $15, %zmm0, %zmm0, %zmm0
15075 ; NoVLX-NEXT: vpmovsxwq %xmm0, %zmm0
15076 ; NoVLX-NEXT: kmovw %edi, %k1
15077 ; NoVLX-NEXT: vptestmq %zmm0, %zmm0, %k0 {%k1}
15078 ; NoVLX-NEXT: kmovw %k0, %eax
15079 ; NoVLX-NEXT: vzeroupper
15080 ; NoVLX-NEXT: retq
15081 entry:
15082 %0 = bitcast <2 x i64> %__a to <8 x i16>
15083 %load = load <2 x i64>, <2 x i64>* %__b
15084 %1 = bitcast <2 x i64> %load to <8 x i16>
15085 %2 = icmp ult <8 x i16> %0, %1
15086 %3 = bitcast i8 %__u to <8 x i1>
15087 %4 = and <8 x i1> %2, %3
15088 %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
15089 %6 = bitcast <32 x i1> %5 to i32
15090 ret i32 %6
15091 }
15094 define zeroext i64 @test_vpcmpultw_v8i1_v64i1_mask(<2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
15095 ; VLX-LABEL: test_vpcmpultw_v8i1_v64i1_mask:
15096 ; VLX: # %bb.0: # %entry
15097 ; VLX-NEXT: vpcmpltuw %xmm1, %xmm0, %k0
15098 ; VLX-NEXT: kmovq %k0, %rax
15099 ; VLX-NEXT: retq
15100 ;
15101 ; NoVLX-LABEL: test_vpcmpultw_v8i1_v64i1_mask:
15102 ; NoVLX: # %bb.0: # %entry
15103 ; NoVLX-NEXT: vpmaxuw %xmm1, %xmm0, %xmm1
15104 ; NoVLX-NEXT: vpcmpeqw %xmm1, %xmm0, %xmm0
15105 ; NoVLX-NEXT: vpternlogq $15, %zmm0, %zmm0, %zmm0
15106 ; NoVLX-NEXT: vpmovsxwq %xmm0, %zmm0
15107 ; NoVLX-NEXT: vptestmq %zmm0, %zmm0, %k0
15108 ; NoVLX-NEXT: kmovw %k0, %eax
15109 ; NoVLX-NEXT: vzeroupper
15110 ; NoVLX-NEXT: retq
15111 entry:
15112 %0 = bitcast <2 x i64> %__a to <8 x i16>
15113 %1 = bitcast <2 x i64> %__b to <8 x i16>
15114 %2 = icmp ult <8 x i16> %0, %1
15115 %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
15116 %4 = bitcast <64 x i1> %3 to i64
15117 ret i64 %4
15118 }
15120 define zeroext i64 @test_vpcmpultw_v8i1_v64i1_mask_mem(<2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
15121 ; VLX-LABEL: test_vpcmpultw_v8i1_v64i1_mask_mem:
15122 ; VLX: # %bb.0: # %entry
15123 ; VLX-NEXT: vpcmpltuw (%rdi), %xmm0, %k0
15124 ; VLX-NEXT: kmovq %k0, %rax
15125 ; VLX-NEXT: retq
15126 ;
15127 ; NoVLX-LABEL: test_vpcmpultw_v8i1_v64i1_mask_mem:
15128 ; NoVLX: # %bb.0: # %entry
15129 ; NoVLX-NEXT: vpmaxuw (%rdi), %xmm0, %xmm1
15130 ; NoVLX-NEXT: vpcmpeqw %xmm1, %xmm0, %xmm0
15131 ; NoVLX-NEXT: vpternlogq $15, %zmm0, %zmm0, %zmm0
15132 ; NoVLX-NEXT: vpmovsxwq %xmm0, %zmm0
15133 ; NoVLX-NEXT: vptestmq %zmm0, %zmm0, %k0
15134 ; NoVLX-NEXT: kmovw %k0, %eax
15135 ; NoVLX-NEXT: vzeroupper
15136 ; NoVLX-NEXT: retq
15137 entry:
15138 %0 = bitcast <2 x i64> %__a to <8 x i16>
15139 %load = load <2 x i64>, <2 x i64>* %__b
15140 %1 = bitcast <2 x i64> %load to <8 x i16>
15141 %2 = icmp ult <8 x i16> %0, %1
15142 %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
15143 %4 = bitcast <64 x i1> %3 to i64
15144 ret i64 %4
15145 }
15147 define zeroext i64 @test_masked_vpcmpultw_v8i1_v64i1_mask(i8 zeroext %__u, <2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
15148 ; VLX-LABEL: test_masked_vpcmpultw_v8i1_v64i1_mask:
15149 ; VLX: # %bb.0: # %entry
15150 ; VLX-NEXT: kmovd %edi, %k1
15151 ; VLX-NEXT: vpcmpltuw %xmm1, %xmm0, %k0 {%k1}
15152 ; VLX-NEXT: kmovq %k0, %rax
15153 ; VLX-NEXT: retq
15154 ;
15155 ; NoVLX-LABEL: test_masked_vpcmpultw_v8i1_v64i1_mask:
15156 ; NoVLX: # %bb.0: # %entry
15157 ; NoVLX-NEXT: vpmaxuw %xmm1, %xmm0, %xmm1
15158 ; NoVLX-NEXT: vpcmpeqw %xmm1, %xmm0, %xmm0
15159 ; NoVLX-NEXT: vpternlogq $15, %zmm0, %zmm0, %zmm0
15160 ; NoVLX-NEXT: vpmovsxwq %xmm0, %zmm0
15161 ; NoVLX-NEXT: kmovw %edi, %k1
15162 ; NoVLX-NEXT: vptestmq %zmm0, %zmm0, %k0 {%k1}
15163 ; NoVLX-NEXT: kmovw %k0, %eax
15164 ; NoVLX-NEXT: vzeroupper
15165 ; NoVLX-NEXT: retq
15166 entry:
15167 %0 = bitcast <2 x i64> %__a to <8 x i16>
15168 %1 = bitcast <2 x i64> %__b to <8 x i16>
15169 %2 = icmp ult <8 x i16> %0, %1
15170 %3 = bitcast i8 %__u to <8 x i1>
15171 %4 = and <8 x i1> %2, %3
15172 %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
15173 %6 = bitcast <64 x i1> %5 to i64
15174 ret i64 %6
15175 }
15177 define zeroext i64 @test_masked_vpcmpultw_v8i1_v64i1_mask_mem(i8 zeroext %__u, <2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
15178 ; VLX-LABEL: test_masked_vpcmpultw_v8i1_v64i1_mask_mem:
15179 ; VLX: # %bb.0: # %entry
15180 ; VLX-NEXT: kmovd %edi, %k1
15181 ; VLX-NEXT: vpcmpltuw (%rsi), %xmm0, %k0 {%k1}
15182 ; VLX-NEXT: kmovq %k0, %rax
15183 ; VLX-NEXT: retq
15184 ;
15185 ; NoVLX-LABEL: test_masked_vpcmpultw_v8i1_v64i1_mask_mem:
15186 ; NoVLX: # %bb.0: # %entry
15187 ; NoVLX-NEXT: vpmaxuw (%rsi), %xmm0, %xmm1
15188 ; NoVLX-NEXT: vpcmpeqw %xmm1, %xmm0, %xmm0
15189 ; NoVLX-NEXT: vpternlogq $15, %zmm0, %zmm0, %zmm0
15190 ; NoVLX-NEXT: vpmovsxwq %xmm0, %zmm0
15191 ; NoVLX-NEXT: kmovw %edi, %k1
15192 ; NoVLX-NEXT: vptestmq %zmm0, %zmm0, %k0 {%k1}
15193 ; NoVLX-NEXT: kmovw %k0, %eax
15194 ; NoVLX-NEXT: vzeroupper
15195 ; NoVLX-NEXT: retq
15196 entry:
15197 %0 = bitcast <2 x i64> %__a to <8 x i16>
15198 %load = load <2 x i64>, <2 x i64>* %__b
15199 %1 = bitcast <2 x i64> %load to <8 x i16>
15200 %2 = icmp ult <8 x i16> %0, %1
15201 %3 = bitcast i8 %__u to <8 x i1>
15202 %4 = and <8 x i1> %2, %3
15203 %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
15204 %6 = bitcast <64 x i1> %5 to i64
15205 ret i64 %6
15206 }
15209 define zeroext i32 @test_vpcmpultw_v16i1_v32i1_mask(<4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
15210 ; VLX-LABEL: test_vpcmpultw_v16i1_v32i1_mask:
15211 ; VLX: # %bb.0: # %entry
15212 ; VLX-NEXT: vpcmpltuw %ymm1, %ymm0, %k0
15213 ; VLX-NEXT: kmovd %k0, %eax
15214 ; VLX-NEXT: vzeroupper
15215 ; VLX-NEXT: retq
15216 ;
15217 ; NoVLX-LABEL: test_vpcmpultw_v16i1_v32i1_mask:
15218 ; NoVLX: # %bb.0: # %entry
15219 ; NoVLX-NEXT: vpmaxuw %ymm1, %ymm0, %ymm1
15220 ; NoVLX-NEXT: vpcmpeqw %ymm1, %ymm0, %ymm0
15221 ; NoVLX-NEXT: vpternlogq $15, %zmm0, %zmm0, %zmm0
15222 ; NoVLX-NEXT: vpmovsxwd %ymm0, %zmm0
15223 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
15224 ; NoVLX-NEXT: kmovw %k0, %eax
15225 ; NoVLX-NEXT: vzeroupper
15226 ; NoVLX-NEXT: retq
15227 entry:
15228 %0 = bitcast <4 x i64> %__a to <16 x i16>
15229 %1 = bitcast <4 x i64> %__b to <16 x i16>
15230 %2 = icmp ult <16 x i16> %0, %1
15231 %3 = shufflevector <16 x i1> %2, <16 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>
15232 %4 = bitcast <32 x i1> %3 to i32
15233 ret i32 %4
15234 }
15236 define zeroext i32 @test_vpcmpultw_v16i1_v32i1_mask_mem(<4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
15237 ; VLX-LABEL: test_vpcmpultw_v16i1_v32i1_mask_mem:
15238 ; VLX: # %bb.0: # %entry
15239 ; VLX-NEXT: vpcmpltuw (%rdi), %ymm0, %k0
15240 ; VLX-NEXT: kmovd %k0, %eax
15241 ; VLX-NEXT: vzeroupper
15242 ; VLX-NEXT: retq
15243 ;
15244 ; NoVLX-LABEL: test_vpcmpultw_v16i1_v32i1_mask_mem:
15245 ; NoVLX: # %bb.0: # %entry
15246 ; NoVLX-NEXT: vpmaxuw (%rdi), %ymm0, %ymm1
15247 ; NoVLX-NEXT: vpcmpeqw %ymm1, %ymm0, %ymm0
15248 ; NoVLX-NEXT: vpternlogq $15, %zmm0, %zmm0, %zmm0
15249 ; NoVLX-NEXT: vpmovsxwd %ymm0, %zmm0
15250 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
15251 ; NoVLX-NEXT: kmovw %k0, %eax
15252 ; NoVLX-NEXT: vzeroupper
15253 ; NoVLX-NEXT: retq
15254 entry:
15255 %0 = bitcast <4 x i64> %__a to <16 x i16>
15256 %load = load <4 x i64>, <4 x i64>* %__b
15257 %1 = bitcast <4 x i64> %load to <16 x i16>
15258 %2 = icmp ult <16 x i16> %0, %1
15259 %3 = shufflevector <16 x i1> %2, <16 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>
15260 %4 = bitcast <32 x i1> %3 to i32
15261 ret i32 %4
15262 }
15264 define zeroext i32 @test_masked_vpcmpultw_v16i1_v32i1_mask(i16 zeroext %__u, <4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
15265 ; VLX-LABEL: test_masked_vpcmpultw_v16i1_v32i1_mask:
15266 ; VLX: # %bb.0: # %entry
15267 ; VLX-NEXT: kmovd %edi, %k1
15268 ; VLX-NEXT: vpcmpltuw %ymm1, %ymm0, %k0 {%k1}
15269 ; VLX-NEXT: kmovd %k0, %eax
15270 ; VLX-NEXT: vzeroupper
15271 ; VLX-NEXT: retq
15272 ;
15273 ; NoVLX-LABEL: test_masked_vpcmpultw_v16i1_v32i1_mask:
15274 ; NoVLX: # %bb.0: # %entry
15275 ; NoVLX-NEXT: vpmaxuw %ymm1, %ymm0, %ymm1
15276 ; NoVLX-NEXT: vpcmpeqw %ymm1, %ymm0, %ymm0
15277 ; NoVLX-NEXT: vpternlogq $15, %zmm0, %zmm0, %zmm0
15278 ; NoVLX-NEXT: vpmovsxwd %ymm0, %zmm0
15279 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
15280 ; NoVLX-NEXT: kmovw %k0, %eax
15281 ; NoVLX-NEXT: andl %edi, %eax
15282 ; NoVLX-NEXT: vzeroupper
15283 ; NoVLX-NEXT: retq
15284 entry:
15285 %0 = bitcast <4 x i64> %__a to <16 x i16>
15286 %1 = bitcast <4 x i64> %__b to <16 x i16>
15287 %2 = icmp ult <16 x i16> %0, %1
15288 %3 = bitcast i16 %__u to <16 x i1>
15289 %4 = and <16 x i1> %2, %3
15290 %5 = shufflevector <16 x i1> %4, <16 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>
15291 %6 = bitcast <32 x i1> %5 to i32
15292 ret i32 %6
15293 }
15295 define zeroext i32 @test_masked_vpcmpultw_v16i1_v32i1_mask_mem(i16 zeroext %__u, <4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
15296 ; VLX-LABEL: test_masked_vpcmpultw_v16i1_v32i1_mask_mem:
15297 ; VLX: # %bb.0: # %entry
15298 ; VLX-NEXT: kmovd %edi, %k1
15299 ; VLX-NEXT: vpcmpltuw (%rsi), %ymm0, %k0 {%k1}
15300 ; VLX-NEXT: kmovd %k0, %eax
15301 ; VLX-NEXT: vzeroupper
15302 ; VLX-NEXT: retq
15303 ;
15304 ; NoVLX-LABEL: test_masked_vpcmpultw_v16i1_v32i1_mask_mem:
15305 ; NoVLX: # %bb.0: # %entry
15306 ; NoVLX-NEXT: vpmaxuw (%rsi), %ymm0, %ymm1
15307 ; NoVLX-NEXT: vpcmpeqw %ymm1, %ymm0, %ymm0
15308 ; NoVLX-NEXT: vpternlogq $15, %zmm0, %zmm0, %zmm0
15309 ; NoVLX-NEXT: vpmovsxwd %ymm0, %zmm0
15310 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
15311 ; NoVLX-NEXT: kmovw %k0, %eax
15312 ; NoVLX-NEXT: andl %edi, %eax
15313 ; NoVLX-NEXT: vzeroupper
15314 ; NoVLX-NEXT: retq
15315 entry:
15316 %0 = bitcast <4 x i64> %__a to <16 x i16>
15317 %load = load <4 x i64>, <4 x i64>* %__b
15318 %1 = bitcast <4 x i64> %load to <16 x i16>
15319 %2 = icmp ult <16 x i16> %0, %1
15320 %3 = bitcast i16 %__u to <16 x i1>
15321 %4 = and <16 x i1> %2, %3
15322 %5 = shufflevector <16 x i1> %4, <16 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>
15323 %6 = bitcast <32 x i1> %5 to i32
15324 ret i32 %6
15325 }
15328 define zeroext i64 @test_vpcmpultw_v16i1_v64i1_mask(<4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
15329 ; VLX-LABEL: test_vpcmpultw_v16i1_v64i1_mask:
15330 ; VLX: # %bb.0: # %entry
15331 ; VLX-NEXT: vpcmpltuw %ymm1, %ymm0, %k0
15332 ; VLX-NEXT: kmovq %k0, %rax
15333 ; VLX-NEXT: vzeroupper
15334 ; VLX-NEXT: retq
15335 ;
15336 ; NoVLX-LABEL: test_vpcmpultw_v16i1_v64i1_mask:
15337 ; NoVLX: # %bb.0: # %entry
15338 ; NoVLX-NEXT: vpmaxuw %ymm1, %ymm0, %ymm1
15339 ; NoVLX-NEXT: vpcmpeqw %ymm1, %ymm0, %ymm0
15340 ; NoVLX-NEXT: vpternlogq $15, %zmm0, %zmm0, %zmm0
15341 ; NoVLX-NEXT: vpmovsxwd %ymm0, %zmm0
15342 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
15343 ; NoVLX-NEXT: kmovw %k0, %eax
15344 ; NoVLX-NEXT: vzeroupper
15345 ; NoVLX-NEXT: retq
15346 entry:
15347 %0 = bitcast <4 x i64> %__a to <16 x i16>
15348 %1 = bitcast <4 x i64> %__b to <16 x i16>
15349 %2 = icmp ult <16 x i16> %0, %1
15350 %3 = shufflevector <16 x i1> %2, <16 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>
15351 %4 = bitcast <64 x i1> %3 to i64
15352 ret i64 %4
15353 }
15355 define zeroext i64 @test_vpcmpultw_v16i1_v64i1_mask_mem(<4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
15356 ; VLX-LABEL: test_vpcmpultw_v16i1_v64i1_mask_mem:
15357 ; VLX: # %bb.0: # %entry
15358 ; VLX-NEXT: vpcmpltuw (%rdi), %ymm0, %k0
15359 ; VLX-NEXT: kmovq %k0, %rax
15360 ; VLX-NEXT: vzeroupper
15361 ; VLX-NEXT: retq
15362 ;
15363 ; NoVLX-LABEL: test_vpcmpultw_v16i1_v64i1_mask_mem:
15364 ; NoVLX: # %bb.0: # %entry
15365 ; NoVLX-NEXT: vpmaxuw (%rdi), %ymm0, %ymm1
15366 ; NoVLX-NEXT: vpcmpeqw %ymm1, %ymm0, %ymm0
15367 ; NoVLX-NEXT: vpternlogq $15, %zmm0, %zmm0, %zmm0
15368 ; NoVLX-NEXT: vpmovsxwd %ymm0, %zmm0
15369 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
15370 ; NoVLX-NEXT: kmovw %k0, %eax
15371 ; NoVLX-NEXT: vzeroupper
15372 ; NoVLX-NEXT: retq
15373 entry:
15374 %0 = bitcast <4 x i64> %__a to <16 x i16>
15375 %load = load <4 x i64>, <4 x i64>* %__b
15376 %1 = bitcast <4 x i64> %load to <16 x i16>
15377 %2 = icmp ult <16 x i16> %0, %1
15378 %3 = shufflevector <16 x i1> %2, <16 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>
15379 %4 = bitcast <64 x i1> %3 to i64
15380 ret i64 %4
15381 }
15383 define zeroext i64 @test_masked_vpcmpultw_v16i1_v64i1_mask(i16 zeroext %__u, <4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
15384 ; VLX-LABEL: test_masked_vpcmpultw_v16i1_v64i1_mask:
15385 ; VLX: # %bb.0: # %entry
15386 ; VLX-NEXT: kmovd %edi, %k1
15387 ; VLX-NEXT: vpcmpltuw %ymm1, %ymm0, %k0 {%k1}
15388 ; VLX-NEXT: kmovq %k0, %rax
15389 ; VLX-NEXT: vzeroupper
15390 ; VLX-NEXT: retq
15391 ;
15392 ; NoVLX-LABEL: test_masked_vpcmpultw_v16i1_v64i1_mask:
15393 ; NoVLX: # %bb.0: # %entry
15394 ; NoVLX-NEXT: vpmaxuw %ymm1, %ymm0, %ymm1
15395 ; NoVLX-NEXT: vpcmpeqw %ymm1, %ymm0, %ymm0
15396 ; NoVLX-NEXT: vpternlogq $15, %zmm0, %zmm0, %zmm0
15397 ; NoVLX-NEXT: vpmovsxwd %ymm0, %zmm0
15398 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
15399 ; NoVLX-NEXT: kmovw %k0, %eax
15400 ; NoVLX-NEXT: andl %edi, %eax
15401 ; NoVLX-NEXT: vzeroupper
15402 ; NoVLX-NEXT: retq
15403 entry:
15404 %0 = bitcast <4 x i64> %__a to <16 x i16>
15405 %1 = bitcast <4 x i64> %__b to <16 x i16>
15406 %2 = icmp ult <16 x i16> %0, %1
15407 %3 = bitcast i16 %__u to <16 x i1>
15408 %4 = and <16 x i1> %2, %3
15409 %5 = shufflevector <16 x i1> %4, <16 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>
15410 %6 = bitcast <64 x i1> %5 to i64
15411 ret i64 %6
15412 }
15414 define zeroext i64 @test_masked_vpcmpultw_v16i1_v64i1_mask_mem(i16 zeroext %__u, <4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
15415 ; VLX-LABEL: test_masked_vpcmpultw_v16i1_v64i1_mask_mem:
15416 ; VLX: # %bb.0: # %entry
15417 ; VLX-NEXT: kmovd %edi, %k1
15418 ; VLX-NEXT: vpcmpltuw (%rsi), %ymm0, %k0 {%k1}
15419 ; VLX-NEXT: kmovq %k0, %rax
15420 ; VLX-NEXT: vzeroupper
15421 ; VLX-NEXT: retq
15422 ;
15423 ; NoVLX-LABEL: test_masked_vpcmpultw_v16i1_v64i1_mask_mem:
15424 ; NoVLX: # %bb.0: # %entry
15425 ; NoVLX-NEXT: vpmaxuw (%rsi), %ymm0, %ymm1
15426 ; NoVLX-NEXT: vpcmpeqw %ymm1, %ymm0, %ymm0
15427 ; NoVLX-NEXT: vpternlogq $15, %zmm0, %zmm0, %zmm0
15428 ; NoVLX-NEXT: vpmovsxwd %ymm0, %zmm0
15429 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
15430 ; NoVLX-NEXT: kmovw %k0, %eax
15431 ; NoVLX-NEXT: andl %edi, %eax
15432 ; NoVLX-NEXT: vzeroupper
15433 ; NoVLX-NEXT: retq
15434 entry:
15435 %0 = bitcast <4 x i64> %__a to <16 x i16>
15436 %load = load <4 x i64>, <4 x i64>* %__b
15437 %1 = bitcast <4 x i64> %load to <16 x i16>
15438 %2 = icmp ult <16 x i16> %0, %1
15439 %3 = bitcast i16 %__u to <16 x i1>
15440 %4 = and <16 x i1> %2, %3
15441 %5 = shufflevector <16 x i1> %4, <16 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>
15442 %6 = bitcast <64 x i1> %5 to i64
15443 ret i64 %6
15444 }
15447 define zeroext i64 @test_vpcmpultw_v32i1_v64i1_mask(<8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
15448 ; VLX-LABEL: test_vpcmpultw_v32i1_v64i1_mask:
15449 ; VLX: # %bb.0: # %entry
15450 ; VLX-NEXT: vpcmpltuw %zmm1, %zmm0, %k0
15451 ; VLX-NEXT: kmovq %k0, %rax
15452 ; VLX-NEXT: vzeroupper
15453 ; VLX-NEXT: retq
15454 ;
15455 ; NoVLX-LABEL: test_vpcmpultw_v32i1_v64i1_mask:
15456 ; NoVLX: # %bb.0: # %entry
15457 ; NoVLX-NEXT: vextracti64x4 $1, %zmm0, %ymm2
15458 ; NoVLX-NEXT: vextracti64x4 $1, %zmm1, %ymm3
15459 ; NoVLX-NEXT: vpmaxuw %ymm1, %ymm0, %ymm1
15460 ; NoVLX-NEXT: vpcmpeqw %ymm1, %ymm0, %ymm0
15461 ; NoVLX-NEXT: vpternlogq $15, %zmm0, %zmm0, %zmm0
15462 ; NoVLX-NEXT: vpmovsxwd %ymm0, %zmm0
15463 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
15464 ; NoVLX-NEXT: kmovw %k0, %ecx
15465 ; NoVLX-NEXT: vpmaxuw %ymm3, %ymm2, %ymm0
15466 ; NoVLX-NEXT: vpcmpeqw %ymm0, %ymm2, %ymm0
15467 ; NoVLX-NEXT: vpternlogq $15, %zmm0, %zmm0, %zmm0
15468 ; NoVLX-NEXT: vpmovsxwd %ymm0, %zmm0
15469 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
15470 ; NoVLX-NEXT: kmovw %k0, %eax
15471 ; NoVLX-NEXT: shll $16, %eax
15472 ; NoVLX-NEXT: orl %ecx, %eax
15473 ; NoVLX-NEXT: vzeroupper
15474 ; NoVLX-NEXT: retq
15475 entry:
15476 %0 = bitcast <8 x i64> %__a to <32 x i16>
15477 %1 = bitcast <8 x i64> %__b to <32 x i16>
15478 %2 = icmp ult <32 x i16> %0, %1
15479 %3 = shufflevector <32 x i1> %2, <32 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63>
15480 %4 = bitcast <64 x i1> %3 to i64
15481 ret i64 %4
15482 }
15484 define zeroext i64 @test_vpcmpultw_v32i1_v64i1_mask_mem(<8 x i64> %__a, <8 x i64>* %__b) local_unnamed_addr {
15485 ; VLX-LABEL: test_vpcmpultw_v32i1_v64i1_mask_mem:
15486 ; VLX: # %bb.0: # %entry
15487 ; VLX-NEXT: vpcmpltuw (%rdi), %zmm0, %k0
15488 ; VLX-NEXT: kmovq %k0, %rax
15489 ; VLX-NEXT: vzeroupper
15490 ; VLX-NEXT: retq
15491 ;
15492 ; NoVLX-LABEL: test_vpcmpultw_v32i1_v64i1_mask_mem:
15493 ; NoVLX: # %bb.0: # %entry
15494 ; NoVLX-NEXT: vextracti64x4 $1, %zmm0, %ymm1
15495 ; NoVLX-NEXT: vpmaxuw (%rdi), %ymm0, %ymm2
15496 ; NoVLX-NEXT: vpcmpeqw %ymm2, %ymm0, %ymm0
15497 ; NoVLX-NEXT: vpternlogq $15, %zmm0, %zmm0, %zmm0
15498 ; NoVLX-NEXT: vpmovsxwd %ymm0, %zmm0
15499 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
15500 ; NoVLX-NEXT: kmovw %k0, %ecx
15501 ; NoVLX-NEXT: vpmaxuw 32(%rdi), %ymm1, %ymm0
15502 ; NoVLX-NEXT: vpcmpeqw %ymm0, %ymm1, %ymm0
15503 ; NoVLX-NEXT: vpternlogq $15, %zmm0, %zmm0, %zmm0
15504 ; NoVLX-NEXT: vpmovsxwd %ymm0, %zmm0
15505 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
15506 ; NoVLX-NEXT: kmovw %k0, %eax
15507 ; NoVLX-NEXT: shll $16, %eax
15508 ; NoVLX-NEXT: orl %ecx, %eax
15509 ; NoVLX-NEXT: vzeroupper
15510 ; NoVLX-NEXT: retq
15511 entry:
15512 %0 = bitcast <8 x i64> %__a to <32 x i16>
15513 %load = load <8 x i64>, <8 x i64>* %__b
15514 %1 = bitcast <8 x i64> %load to <32 x i16>
15515 %2 = icmp ult <32 x i16> %0, %1
15516 %3 = shufflevector <32 x i1> %2, <32 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63>
15517 %4 = bitcast <64 x i1> %3 to i64
15518 ret i64 %4
15519 }
15521 define zeroext i64 @test_masked_vpcmpultw_v32i1_v64i1_mask(i32 zeroext %__u, <8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
15522 ; VLX-LABEL: test_masked_vpcmpultw_v32i1_v64i1_mask:
15523 ; VLX: # %bb.0: # %entry
15524 ; VLX-NEXT: kmovd %edi, %k1
15525 ; VLX-NEXT: vpcmpltuw %zmm1, %zmm0, %k0 {%k1}
15526 ; VLX-NEXT: kmovq %k0, %rax
15527 ; VLX-NEXT: vzeroupper
15528 ; VLX-NEXT: retq
15529 ;
15530 ; NoVLX-LABEL: test_masked_vpcmpultw_v32i1_v64i1_mask:
15531 ; NoVLX: # %bb.0: # %entry
15532 ; NoVLX-NEXT: vpmaxuw %ymm1, %ymm0, %ymm2
15533 ; NoVLX-NEXT: vpcmpeqw %ymm2, %ymm0, %ymm2
15534 ; NoVLX-NEXT: vpternlogq $15, %zmm2, %zmm2, %zmm2
15535 ; NoVLX-NEXT: vpmovsxwd %ymm2, %zmm2
15536 ; NoVLX-NEXT: vptestmd %zmm2, %zmm2, %k0
15537 ; NoVLX-NEXT: kmovw %k0, %eax
15538 ; NoVLX-NEXT: andl %edi, %eax
15539 ; NoVLX-NEXT: shrl $16, %edi
15540 ; NoVLX-NEXT: vextracti64x4 $1, %zmm0, %ymm0
15541 ; NoVLX-NEXT: vextracti64x4 $1, %zmm1, %ymm1
15542 ; NoVLX-NEXT: vpmaxuw %ymm1, %ymm0, %ymm1
15543 ; NoVLX-NEXT: vpcmpeqw %ymm1, %ymm0, %ymm0
15544 ; NoVLX-NEXT: vpternlogq $15, %zmm0, %zmm0, %zmm0
15545 ; NoVLX-NEXT: vpmovsxwd %ymm0, %zmm0
15546 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
15547 ; NoVLX-NEXT: kmovw %k0, %ecx
15548 ; NoVLX-NEXT: andl %edi, %ecx
15549 ; NoVLX-NEXT: shll $16, %ecx
15550 ; NoVLX-NEXT: movzwl %ax, %eax
15551 ; NoVLX-NEXT: orl %ecx, %eax
15552 ; NoVLX-NEXT: vzeroupper
15553 ; NoVLX-NEXT: retq
15554 entry:
15555 %0 = bitcast <8 x i64> %__a to <32 x i16>
15556 %1 = bitcast <8 x i64> %__b to <32 x i16>
15557 %2 = icmp ult <32 x i16> %0, %1
15558 %3 = bitcast i32 %__u to <32 x i1>
15559 %4 = and <32 x i1> %2, %3
15560 %5 = shufflevector <32 x i1> %4, <32 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63>
15561 %6 = bitcast <64 x i1> %5 to i64
15562 ret i64 %6
15563 }
15565 define zeroext i64 @test_masked_vpcmpultw_v32i1_v64i1_mask_mem(i32 zeroext %__u, <8 x i64> %__a, <8 x i64>* %__b) local_unnamed_addr {
15566 ; VLX-LABEL: test_masked_vpcmpultw_v32i1_v64i1_mask_mem:
15567 ; VLX: # %bb.0: # %entry
15568 ; VLX-NEXT: kmovd %edi, %k1
15569 ; VLX-NEXT: vpcmpltuw (%rsi), %zmm0, %k0 {%k1}
15570 ; VLX-NEXT: kmovq %k0, %rax
15571 ; VLX-NEXT: vzeroupper
15572 ; VLX-NEXT: retq
15573 ;
15574 ; NoVLX-LABEL: test_masked_vpcmpultw_v32i1_v64i1_mask_mem:
15575 ; NoVLX: # %bb.0: # %entry
15576 ; NoVLX-NEXT: vpmaxuw (%rsi), %ymm0, %ymm1
15577 ; NoVLX-NEXT: vpcmpeqw %ymm1, %ymm0, %ymm1
15578 ; NoVLX-NEXT: vpternlogq $15, %zmm1, %zmm1, %zmm1
15579 ; NoVLX-NEXT: vpmovsxwd %ymm1, %zmm1
15580 ; NoVLX-NEXT: vptestmd %zmm1, %zmm1, %k0
15581 ; NoVLX-NEXT: kmovw %k0, %eax
15582 ; NoVLX-NEXT: andl %edi, %eax
15583 ; NoVLX-NEXT: shrl $16, %edi
15584 ; NoVLX-NEXT: vextracti64x4 $1, %zmm0, %ymm0
15585 ; NoVLX-NEXT: vpmaxuw 32(%rsi), %ymm0, %ymm1
15586 ; NoVLX-NEXT: vpcmpeqw %ymm1, %ymm0, %ymm0
15587 ; NoVLX-NEXT: vpternlogq $15, %zmm0, %zmm0, %zmm0
15588 ; NoVLX-NEXT: vpmovsxwd %ymm0, %zmm0
15589 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
15590 ; NoVLX-NEXT: kmovw %k0, %ecx
15591 ; NoVLX-NEXT: andl %edi, %ecx
15592 ; NoVLX-NEXT: shll $16, %ecx
15593 ; NoVLX-NEXT: movzwl %ax, %eax
15594 ; NoVLX-NEXT: orl %ecx, %eax
15595 ; NoVLX-NEXT: vzeroupper
15596 ; NoVLX-NEXT: retq
15597 entry:
15598 %0 = bitcast <8 x i64> %__a to <32 x i16>
15599 %load = load <8 x i64>, <8 x i64>* %__b
15600 %1 = bitcast <8 x i64> %load to <32 x i16>
15601 %2 = icmp ult <32 x i16> %0, %1
15602 %3 = bitcast i32 %__u to <32 x i1>
15603 %4 = and <32 x i1> %2, %3
15604 %5 = shufflevector <32 x i1> %4, <32 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31, i32 32, i32 33, i32 34, i32 35, i32 36, i32 37, i32 38, i32 39, i32 40, i32 41, i32 42, i32 43, i32 44, i32 45, i32 46, i32 47, i32 48, i32 49, i32 50, i32 51, i32 52, i32 53, i32 54, i32 55, i32 56, i32 57, i32 58, i32 59, i32 60, i32 61, i32 62, i32 63>
15605 %6 = bitcast <64 x i1> %5 to i64
15606 ret i64 %6
15607 }
15610 define zeroext i8 @test_vpcmpultd_v4i1_v8i1_mask(<2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
15611 ; VLX-LABEL: test_vpcmpultd_v4i1_v8i1_mask:
15612 ; VLX: # %bb.0: # %entry
15613 ; VLX-NEXT: vpcmpltud %xmm1, %xmm0, %k0
15614 ; VLX-NEXT: kmovd %k0, %eax
15615 ; VLX-NEXT: # kill: def $al killed $al killed $eax
15616 ; VLX-NEXT: retq
15617 ;
15618 ; NoVLX-LABEL: test_vpcmpultd_v4i1_v8i1_mask:
15619 ; NoVLX: # %bb.0: # %entry
15620 ; NoVLX-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
15621 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
15622 ; NoVLX-NEXT: vpcmpltud %zmm1, %zmm0, %k0
15623 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
15624 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
15625 ; NoVLX-NEXT: kmovw %k0, %eax
15626 ; NoVLX-NEXT: # kill: def $al killed $al killed $eax
15627 ; NoVLX-NEXT: vzeroupper
15628 ; NoVLX-NEXT: retq
15629 entry:
15630 %0 = bitcast <2 x i64> %__a to <4 x i32>
15631 %1 = bitcast <2 x i64> %__b to <4 x i32>
15632 %2 = icmp ult <4 x i32> %0, %1
15633 %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
15634 %4 = bitcast <8 x i1> %3 to i8
15635 ret i8 %4
15636 }
15638 define zeroext i8 @test_vpcmpultd_v4i1_v8i1_mask_mem(<2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
15639 ; VLX-LABEL: test_vpcmpultd_v4i1_v8i1_mask_mem:
15640 ; VLX: # %bb.0: # %entry
15641 ; VLX-NEXT: vpcmpltud (%rdi), %xmm0, %k0
15642 ; VLX-NEXT: kmovd %k0, %eax
15643 ; VLX-NEXT: # kill: def $al killed $al killed $eax
15644 ; VLX-NEXT: retq
15645 ;
15646 ; NoVLX-LABEL: test_vpcmpultd_v4i1_v8i1_mask_mem:
15647 ; NoVLX: # %bb.0: # %entry
15648 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
15649 ; NoVLX-NEXT: vmovdqa (%rdi), %xmm1
15650 ; NoVLX-NEXT: vpcmpltud %zmm1, %zmm0, %k0
15651 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
15652 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
15653 ; NoVLX-NEXT: kmovw %k0, %eax
15654 ; NoVLX-NEXT: # kill: def $al killed $al killed $eax
15655 ; NoVLX-NEXT: vzeroupper
15656 ; NoVLX-NEXT: retq
15657 entry:
15658 %0 = bitcast <2 x i64> %__a to <4 x i32>
15659 %load = load <2 x i64>, <2 x i64>* %__b
15660 %1 = bitcast <2 x i64> %load to <4 x i32>
15661 %2 = icmp ult <4 x i32> %0, %1
15662 %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
15663 %4 = bitcast <8 x i1> %3 to i8
15664 ret i8 %4
15665 }
15667 define zeroext i8 @test_masked_vpcmpultd_v4i1_v8i1_mask(i8 zeroext %__u, <2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
15668 ; VLX-LABEL: test_masked_vpcmpultd_v4i1_v8i1_mask:
15669 ; VLX: # %bb.0: # %entry
15670 ; VLX-NEXT: kmovd %edi, %k1
15671 ; VLX-NEXT: vpcmpltud %xmm1, %xmm0, %k0 {%k1}
15672 ; VLX-NEXT: kmovd %k0, %eax
15673 ; VLX-NEXT: # kill: def $al killed $al killed $eax
15674 ; VLX-NEXT: retq
15675 ;
15676 ; NoVLX-LABEL: test_masked_vpcmpultd_v4i1_v8i1_mask:
15677 ; NoVLX: # %bb.0: # %entry
15678 ; NoVLX-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
15679 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
15680 ; NoVLX-NEXT: kmovw %edi, %k1
15681 ; NoVLX-NEXT: vpcmpltud %zmm1, %zmm0, %k0 {%k1}
15682 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
15683 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
15684 ; NoVLX-NEXT: kmovw %k0, %eax
15685 ; NoVLX-NEXT: # kill: def $al killed $al killed $eax
15686 ; NoVLX-NEXT: vzeroupper
15687 ; NoVLX-NEXT: retq
15688 entry:
15689 %0 = bitcast <2 x i64> %__a to <4 x i32>
15690 %1 = bitcast <2 x i64> %__b to <4 x i32>
15691 %2 = icmp ult <4 x i32> %0, %1
15692 %3 = bitcast i8 %__u to <8 x i1>
15693 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
15694 %4 = and <4 x i1> %2, %extract.i
15695 %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
15696 %6 = bitcast <8 x i1> %5 to i8
15697 ret i8 %6
15698 }
15700 define zeroext i8 @test_masked_vpcmpultd_v4i1_v8i1_mask_mem(i8 zeroext %__u, <2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
15701 ; VLX-LABEL: test_masked_vpcmpultd_v4i1_v8i1_mask_mem:
15702 ; VLX: # %bb.0: # %entry
15703 ; VLX-NEXT: kmovd %edi, %k1
15704 ; VLX-NEXT: vpcmpltud (%rsi), %xmm0, %k0 {%k1}
15705 ; VLX-NEXT: kmovd %k0, %eax
15706 ; VLX-NEXT: # kill: def $al killed $al killed $eax
15707 ; VLX-NEXT: retq
15708 ;
15709 ; NoVLX-LABEL: test_masked_vpcmpultd_v4i1_v8i1_mask_mem:
15710 ; NoVLX: # %bb.0: # %entry
15711 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
15712 ; NoVLX-NEXT: vmovdqa (%rsi), %xmm1
15713 ; NoVLX-NEXT: kmovw %edi, %k1
15714 ; NoVLX-NEXT: vpcmpltud %zmm1, %zmm0, %k0 {%k1}
15715 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
15716 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
15717 ; NoVLX-NEXT: kmovw %k0, %eax
15718 ; NoVLX-NEXT: # kill: def $al killed $al killed $eax
15719 ; NoVLX-NEXT: vzeroupper
15720 ; NoVLX-NEXT: retq
15721 entry:
15722 %0 = bitcast <2 x i64> %__a to <4 x i32>
15723 %load = load <2 x i64>, <2 x i64>* %__b
15724 %1 = bitcast <2 x i64> %load to <4 x i32>
15725 %2 = icmp ult <4 x i32> %0, %1
15726 %3 = bitcast i8 %__u to <8 x i1>
15727 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
15728 %4 = and <4 x i1> %2, %extract.i
15729 %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
15730 %6 = bitcast <8 x i1> %5 to i8
15731 ret i8 %6
15732 }
15735 define zeroext i8 @test_vpcmpultd_v4i1_v8i1_mask_mem_b(<2 x i64> %__a, i32* %__b) local_unnamed_addr {
15736 ; VLX-LABEL: test_vpcmpultd_v4i1_v8i1_mask_mem_b:
15737 ; VLX: # %bb.0: # %entry
15738 ; VLX-NEXT: vpcmpltud (%rdi){1to4}, %xmm0, %k0
15739 ; VLX-NEXT: kmovd %k0, %eax
15740 ; VLX-NEXT: # kill: def $al killed $al killed $eax
15741 ; VLX-NEXT: retq
15742 ;
15743 ; NoVLX-LABEL: test_vpcmpultd_v4i1_v8i1_mask_mem_b:
15744 ; NoVLX: # %bb.0: # %entry
15745 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
15746 ; NoVLX-NEXT: vpcmpltud (%rdi){1to16}, %zmm0, %k0
15747 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
15748 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
15749 ; NoVLX-NEXT: kmovw %k0, %eax
15750 ; NoVLX-NEXT: # kill: def $al killed $al killed $eax
15751 ; NoVLX-NEXT: vzeroupper
15752 ; NoVLX-NEXT: retq
15753 entry:
15754 %0 = bitcast <2 x i64> %__a to <4 x i32>
15755 %load = load i32, i32* %__b
15756 %vec = insertelement <4 x i32> undef, i32 %load, i32 0
15757 %1 = shufflevector <4 x i32> %vec, <4 x i32> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
15758 %2 = icmp ult <4 x i32> %0, %1
15759 %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
15760 %4 = bitcast <8 x i1> %3 to i8
15761 ret i8 %4
15762 }
15764 define zeroext i8 @test_masked_vpcmpultd_v4i1_v8i1_mask_mem_b(i8 zeroext %__u, <2 x i64> %__a, i32* %__b) local_unnamed_addr {
15765 ; VLX-LABEL: test_masked_vpcmpultd_v4i1_v8i1_mask_mem_b:
15766 ; VLX: # %bb.0: # %entry
15767 ; VLX-NEXT: kmovd %edi, %k1
15768 ; VLX-NEXT: vpcmpltud (%rsi){1to4}, %xmm0, %k0 {%k1}
15769 ; VLX-NEXT: kmovd %k0, %eax
15770 ; VLX-NEXT: # kill: def $al killed $al killed $eax
15771 ; VLX-NEXT: retq
15772 ;
15773 ; NoVLX-LABEL: test_masked_vpcmpultd_v4i1_v8i1_mask_mem_b:
15774 ; NoVLX: # %bb.0: # %entry
15775 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
15776 ; NoVLX-NEXT: kmovw %edi, %k1
15777 ; NoVLX-NEXT: vpcmpltud (%rsi){1to16}, %zmm0, %k0 {%k1}
15778 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
15779 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
15780 ; NoVLX-NEXT: kmovw %k0, %eax
15781 ; NoVLX-NEXT: # kill: def $al killed $al killed $eax
15782 ; NoVLX-NEXT: vzeroupper
15783 ; NoVLX-NEXT: retq
15784 entry:
15785 %0 = bitcast <2 x i64> %__a to <4 x i32>
15786 %load = load i32, i32* %__b
15787 %vec = insertelement <4 x i32> undef, i32 %load, i32 0
15788 %1 = shufflevector <4 x i32> %vec, <4 x i32> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
15789 %2 = icmp ult <4 x i32> %0, %1
15790 %3 = bitcast i8 %__u to <8 x i1>
15791 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
15792 %4 = and <4 x i1> %extract.i, %2
15793 %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
15794 %6 = bitcast <8 x i1> %5 to i8
15795 ret i8 %6
15796 }
15799 define zeroext i16 @test_vpcmpultd_v4i1_v16i1_mask(<2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
15800 ; VLX-LABEL: test_vpcmpultd_v4i1_v16i1_mask:
15801 ; VLX: # %bb.0: # %entry
15802 ; VLX-NEXT: vpcmpltud %xmm1, %xmm0, %k0
15803 ; VLX-NEXT: kmovd %k0, %eax
15804 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
15805 ; VLX-NEXT: retq
15806 ;
15807 ; NoVLX-LABEL: test_vpcmpultd_v4i1_v16i1_mask:
15808 ; NoVLX: # %bb.0: # %entry
15809 ; NoVLX-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
15810 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
15811 ; NoVLX-NEXT: vpcmpltud %zmm1, %zmm0, %k0
15812 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
15813 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
15814 ; NoVLX-NEXT: kmovw %k0, %eax
15815 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
15816 ; NoVLX-NEXT: vzeroupper
15817 ; NoVLX-NEXT: retq
15818 entry:
15819 %0 = bitcast <2 x i64> %__a to <4 x i32>
15820 %1 = bitcast <2 x i64> %__b to <4 x i32>
15821 %2 = icmp ult <4 x i32> %0, %1
15822 %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
15823 %4 = bitcast <16 x i1> %3 to i16
15824 ret i16 %4
15825 }
15827 define zeroext i16 @test_vpcmpultd_v4i1_v16i1_mask_mem(<2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
15828 ; VLX-LABEL: test_vpcmpultd_v4i1_v16i1_mask_mem:
15829 ; VLX: # %bb.0: # %entry
15830 ; VLX-NEXT: vpcmpltud (%rdi), %xmm0, %k0
15831 ; VLX-NEXT: kmovd %k0, %eax
15832 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
15833 ; VLX-NEXT: retq
15834 ;
15835 ; NoVLX-LABEL: test_vpcmpultd_v4i1_v16i1_mask_mem:
15836 ; NoVLX: # %bb.0: # %entry
15837 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
15838 ; NoVLX-NEXT: vmovdqa (%rdi), %xmm1
15839 ; NoVLX-NEXT: vpcmpltud %zmm1, %zmm0, %k0
15840 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
15841 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
15842 ; NoVLX-NEXT: kmovw %k0, %eax
15843 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
15844 ; NoVLX-NEXT: vzeroupper
15845 ; NoVLX-NEXT: retq
15846 entry:
15847 %0 = bitcast <2 x i64> %__a to <4 x i32>
15848 %load = load <2 x i64>, <2 x i64>* %__b
15849 %1 = bitcast <2 x i64> %load to <4 x i32>
15850 %2 = icmp ult <4 x i32> %0, %1
15851 %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
15852 %4 = bitcast <16 x i1> %3 to i16
15853 ret i16 %4
15854 }
15856 define zeroext i16 @test_masked_vpcmpultd_v4i1_v16i1_mask(i8 zeroext %__u, <2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
15857 ; VLX-LABEL: test_masked_vpcmpultd_v4i1_v16i1_mask:
15858 ; VLX: # %bb.0: # %entry
15859 ; VLX-NEXT: kmovd %edi, %k1
15860 ; VLX-NEXT: vpcmpltud %xmm1, %xmm0, %k0 {%k1}
15861 ; VLX-NEXT: kmovd %k0, %eax
15862 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
15863 ; VLX-NEXT: retq
15864 ;
15865 ; NoVLX-LABEL: test_masked_vpcmpultd_v4i1_v16i1_mask:
15866 ; NoVLX: # %bb.0: # %entry
15867 ; NoVLX-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
15868 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
15869 ; NoVLX-NEXT: kmovw %edi, %k1
15870 ; NoVLX-NEXT: vpcmpltud %zmm1, %zmm0, %k0 {%k1}
15871 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
15872 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
15873 ; NoVLX-NEXT: kmovw %k0, %eax
15874 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
15875 ; NoVLX-NEXT: vzeroupper
15876 ; NoVLX-NEXT: retq
15877 entry:
15878 %0 = bitcast <2 x i64> %__a to <4 x i32>
15879 %1 = bitcast <2 x i64> %__b to <4 x i32>
15880 %2 = icmp ult <4 x i32> %0, %1
15881 %3 = bitcast i8 %__u to <8 x i1>
15882 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
15883 %4 = and <4 x i1> %2, %extract.i
15884 %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
15885 %6 = bitcast <16 x i1> %5 to i16
15886 ret i16 %6
15887 }
15889 define zeroext i16 @test_masked_vpcmpultd_v4i1_v16i1_mask_mem(i8 zeroext %__u, <2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
15890 ; VLX-LABEL: test_masked_vpcmpultd_v4i1_v16i1_mask_mem:
15891 ; VLX: # %bb.0: # %entry
15892 ; VLX-NEXT: kmovd %edi, %k1
15893 ; VLX-NEXT: vpcmpltud (%rsi), %xmm0, %k0 {%k1}
15894 ; VLX-NEXT: kmovd %k0, %eax
15895 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
15896 ; VLX-NEXT: retq
15897 ;
15898 ; NoVLX-LABEL: test_masked_vpcmpultd_v4i1_v16i1_mask_mem:
15899 ; NoVLX: # %bb.0: # %entry
15900 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
15901 ; NoVLX-NEXT: vmovdqa (%rsi), %xmm1
15902 ; NoVLX-NEXT: kmovw %edi, %k1
15903 ; NoVLX-NEXT: vpcmpltud %zmm1, %zmm0, %k0 {%k1}
15904 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
15905 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
15906 ; NoVLX-NEXT: kmovw %k0, %eax
15907 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
15908 ; NoVLX-NEXT: vzeroupper
15909 ; NoVLX-NEXT: retq
15910 entry:
15911 %0 = bitcast <2 x i64> %__a to <4 x i32>
15912 %load = load <2 x i64>, <2 x i64>* %__b
15913 %1 = bitcast <2 x i64> %load to <4 x i32>
15914 %2 = icmp ult <4 x i32> %0, %1
15915 %3 = bitcast i8 %__u to <8 x i1>
15916 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
15917 %4 = and <4 x i1> %2, %extract.i
15918 %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
15919 %6 = bitcast <16 x i1> %5 to i16
15920 ret i16 %6
15921 }
15924 define zeroext i16 @test_vpcmpultd_v4i1_v16i1_mask_mem_b(<2 x i64> %__a, i32* %__b) local_unnamed_addr {
15925 ; VLX-LABEL: test_vpcmpultd_v4i1_v16i1_mask_mem_b:
15926 ; VLX: # %bb.0: # %entry
15927 ; VLX-NEXT: vpcmpltud (%rdi){1to4}, %xmm0, %k0
15928 ; VLX-NEXT: kmovd %k0, %eax
15929 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
15930 ; VLX-NEXT: retq
15931 ;
15932 ; NoVLX-LABEL: test_vpcmpultd_v4i1_v16i1_mask_mem_b:
15933 ; NoVLX: # %bb.0: # %entry
15934 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
15935 ; NoVLX-NEXT: vpcmpltud (%rdi){1to16}, %zmm0, %k0
15936 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
15937 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
15938 ; NoVLX-NEXT: kmovw %k0, %eax
15939 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
15940 ; NoVLX-NEXT: vzeroupper
15941 ; NoVLX-NEXT: retq
15942 entry:
15943 %0 = bitcast <2 x i64> %__a to <4 x i32>
15944 %load = load i32, i32* %__b
15945 %vec = insertelement <4 x i32> undef, i32 %load, i32 0
15946 %1 = shufflevector <4 x i32> %vec, <4 x i32> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
15947 %2 = icmp ult <4 x i32> %0, %1
15948 %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
15949 %4 = bitcast <16 x i1> %3 to i16
15950 ret i16 %4
15951 }
15953 define zeroext i16 @test_masked_vpcmpultd_v4i1_v16i1_mask_mem_b(i8 zeroext %__u, <2 x i64> %__a, i32* %__b) local_unnamed_addr {
15954 ; VLX-LABEL: test_masked_vpcmpultd_v4i1_v16i1_mask_mem_b:
15955 ; VLX: # %bb.0: # %entry
15956 ; VLX-NEXT: kmovd %edi, %k1
15957 ; VLX-NEXT: vpcmpltud (%rsi){1to4}, %xmm0, %k0 {%k1}
15958 ; VLX-NEXT: kmovd %k0, %eax
15959 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
15960 ; VLX-NEXT: retq
15961 ;
15962 ; NoVLX-LABEL: test_masked_vpcmpultd_v4i1_v16i1_mask_mem_b:
15963 ; NoVLX: # %bb.0: # %entry
15964 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
15965 ; NoVLX-NEXT: kmovw %edi, %k1
15966 ; NoVLX-NEXT: vpcmpltud (%rsi){1to16}, %zmm0, %k0 {%k1}
15967 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
15968 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
15969 ; NoVLX-NEXT: kmovw %k0, %eax
15970 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
15971 ; NoVLX-NEXT: vzeroupper
15972 ; NoVLX-NEXT: retq
15973 entry:
15974 %0 = bitcast <2 x i64> %__a to <4 x i32>
15975 %load = load i32, i32* %__b
15976 %vec = insertelement <4 x i32> undef, i32 %load, i32 0
15977 %1 = shufflevector <4 x i32> %vec, <4 x i32> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
15978 %2 = icmp ult <4 x i32> %0, %1
15979 %3 = bitcast i8 %__u to <8 x i1>
15980 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
15981 %4 = and <4 x i1> %extract.i, %2
15982 %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
15983 %6 = bitcast <16 x i1> %5 to i16
15984 ret i16 %6
15985 }
15988 define zeroext i32 @test_vpcmpultd_v4i1_v32i1_mask(<2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
15989 ; VLX-LABEL: test_vpcmpultd_v4i1_v32i1_mask:
15990 ; VLX: # %bb.0: # %entry
15991 ; VLX-NEXT: vpcmpltud %xmm1, %xmm0, %k0
15992 ; VLX-NEXT: kmovd %k0, %eax
15993 ; VLX-NEXT: retq
15994 ;
15995 ; NoVLX-LABEL: test_vpcmpultd_v4i1_v32i1_mask:
15996 ; NoVLX: # %bb.0: # %entry
15997 ; NoVLX-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
15998 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
15999 ; NoVLX-NEXT: vpcmpltud %zmm1, %zmm0, %k0
16000 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
16001 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
16002 ; NoVLX-NEXT: kmovw %k0, %eax
16003 ; NoVLX-NEXT: vzeroupper
16004 ; NoVLX-NEXT: retq
16005 entry:
16006 %0 = bitcast <2 x i64> %__a to <4 x i32>
16007 %1 = bitcast <2 x i64> %__b to <4 x i32>
16008 %2 = icmp ult <4 x i32> %0, %1
16009 %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
16010 %4 = bitcast <32 x i1> %3 to i32
16011 ret i32 %4
16012 }
16014 define zeroext i32 @test_vpcmpultd_v4i1_v32i1_mask_mem(<2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
16015 ; VLX-LABEL: test_vpcmpultd_v4i1_v32i1_mask_mem:
16016 ; VLX: # %bb.0: # %entry
16017 ; VLX-NEXT: vpcmpltud (%rdi), %xmm0, %k0
16018 ; VLX-NEXT: kmovd %k0, %eax
16019 ; VLX-NEXT: retq
16020 ;
16021 ; NoVLX-LABEL: test_vpcmpultd_v4i1_v32i1_mask_mem:
16022 ; NoVLX: # %bb.0: # %entry
16023 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
16024 ; NoVLX-NEXT: vmovdqa (%rdi), %xmm1
16025 ; NoVLX-NEXT: vpcmpltud %zmm1, %zmm0, %k0
16026 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
16027 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
16028 ; NoVLX-NEXT: kmovw %k0, %eax
16029 ; NoVLX-NEXT: vzeroupper
16030 ; NoVLX-NEXT: retq
16031 entry:
16032 %0 = bitcast <2 x i64> %__a to <4 x i32>
16033 %load = load <2 x i64>, <2 x i64>* %__b
16034 %1 = bitcast <2 x i64> %load to <4 x i32>
16035 %2 = icmp ult <4 x i32> %0, %1
16036 %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
16037 %4 = bitcast <32 x i1> %3 to i32
16038 ret i32 %4
16039 }
16041 define zeroext i32 @test_masked_vpcmpultd_v4i1_v32i1_mask(i8 zeroext %__u, <2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
16042 ; VLX-LABEL: test_masked_vpcmpultd_v4i1_v32i1_mask:
16043 ; VLX: # %bb.0: # %entry
16044 ; VLX-NEXT: kmovd %edi, %k1
16045 ; VLX-NEXT: vpcmpltud %xmm1, %xmm0, %k0 {%k1}
16046 ; VLX-NEXT: kmovd %k0, %eax
16047 ; VLX-NEXT: retq
16048 ;
16049 ; NoVLX-LABEL: test_masked_vpcmpultd_v4i1_v32i1_mask:
16050 ; NoVLX: # %bb.0: # %entry
16051 ; NoVLX-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
16052 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
16053 ; NoVLX-NEXT: kmovw %edi, %k1
16054 ; NoVLX-NEXT: vpcmpltud %zmm1, %zmm0, %k0 {%k1}
16055 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
16056 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
16057 ; NoVLX-NEXT: kmovw %k0, %eax
16058 ; NoVLX-NEXT: vzeroupper
16059 ; NoVLX-NEXT: retq
16060 entry:
16061 %0 = bitcast <2 x i64> %__a to <4 x i32>
16062 %1 = bitcast <2 x i64> %__b to <4 x i32>
16063 %2 = icmp ult <4 x i32> %0, %1
16064 %3 = bitcast i8 %__u to <8 x i1>
16065 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
16066 %4 = and <4 x i1> %2, %extract.i
16067 %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
16068 %6 = bitcast <32 x i1> %5 to i32
16069 ret i32 %6
16070 }
16072 define zeroext i32 @test_masked_vpcmpultd_v4i1_v32i1_mask_mem(i8 zeroext %__u, <2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
16073 ; VLX-LABEL: test_masked_vpcmpultd_v4i1_v32i1_mask_mem:
16074 ; VLX: # %bb.0: # %entry
16075 ; VLX-NEXT: kmovd %edi, %k1
16076 ; VLX-NEXT: vpcmpltud (%rsi), %xmm0, %k0 {%k1}
16077 ; VLX-NEXT: kmovd %k0, %eax
16078 ; VLX-NEXT: retq
16079 ;
16080 ; NoVLX-LABEL: test_masked_vpcmpultd_v4i1_v32i1_mask_mem:
16081 ; NoVLX: # %bb.0: # %entry
16082 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
16083 ; NoVLX-NEXT: vmovdqa (%rsi), %xmm1
16084 ; NoVLX-NEXT: kmovw %edi, %k1
16085 ; NoVLX-NEXT: vpcmpltud %zmm1, %zmm0, %k0 {%k1}
16086 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
16087 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
16088 ; NoVLX-NEXT: kmovw %k0, %eax
16089 ; NoVLX-NEXT: vzeroupper
16090 ; NoVLX-NEXT: retq
16091 entry:
16092 %0 = bitcast <2 x i64> %__a to <4 x i32>
16093 %load = load <2 x i64>, <2 x i64>* %__b
16094 %1 = bitcast <2 x i64> %load to <4 x i32>
16095 %2 = icmp ult <4 x i32> %0, %1
16096 %3 = bitcast i8 %__u to <8 x i1>
16097 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
16098 %4 = and <4 x i1> %2, %extract.i
16099 %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
16100 %6 = bitcast <32 x i1> %5 to i32
16101 ret i32 %6
16102 }
16105 define zeroext i32 @test_vpcmpultd_v4i1_v32i1_mask_mem_b(<2 x i64> %__a, i32* %__b) local_unnamed_addr {
16106 ; VLX-LABEL: test_vpcmpultd_v4i1_v32i1_mask_mem_b:
16107 ; VLX: # %bb.0: # %entry
16108 ; VLX-NEXT: vpcmpltud (%rdi){1to4}, %xmm0, %k0
16109 ; VLX-NEXT: kmovd %k0, %eax
16110 ; VLX-NEXT: retq
16111 ;
16112 ; NoVLX-LABEL: test_vpcmpultd_v4i1_v32i1_mask_mem_b:
16113 ; NoVLX: # %bb.0: # %entry
16114 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
16115 ; NoVLX-NEXT: vpcmpltud (%rdi){1to16}, %zmm0, %k0
16116 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
16117 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
16118 ; NoVLX-NEXT: kmovw %k0, %eax
16119 ; NoVLX-NEXT: vzeroupper
16120 ; NoVLX-NEXT: retq
16121 entry:
16122 %0 = bitcast <2 x i64> %__a to <4 x i32>
16123 %load = load i32, i32* %__b
16124 %vec = insertelement <4 x i32> undef, i32 %load, i32 0
16125 %1 = shufflevector <4 x i32> %vec, <4 x i32> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
16126 %2 = icmp ult <4 x i32> %0, %1
16127 %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
16128 %4 = bitcast <32 x i1> %3 to i32
16129 ret i32 %4
16130 }
16132 define zeroext i32 @test_masked_vpcmpultd_v4i1_v32i1_mask_mem_b(i8 zeroext %__u, <2 x i64> %__a, i32* %__b) local_unnamed_addr {
16133 ; VLX-LABEL: test_masked_vpcmpultd_v4i1_v32i1_mask_mem_b:
16134 ; VLX: # %bb.0: # %entry
16135 ; VLX-NEXT: kmovd %edi, %k1
16136 ; VLX-NEXT: vpcmpltud (%rsi){1to4}, %xmm0, %k0 {%k1}
16137 ; VLX-NEXT: kmovd %k0, %eax
16138 ; VLX-NEXT: retq
16139 ;
16140 ; NoVLX-LABEL: test_masked_vpcmpultd_v4i1_v32i1_mask_mem_b:
16141 ; NoVLX: # %bb.0: # %entry
16142 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
16143 ; NoVLX-NEXT: kmovw %edi, %k1
16144 ; NoVLX-NEXT: vpcmpltud (%rsi){1to16}, %zmm0, %k0 {%k1}
16145 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
16146 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
16147 ; NoVLX-NEXT: kmovw %k0, %eax
16148 ; NoVLX-NEXT: vzeroupper
16149 ; NoVLX-NEXT: retq
16150 entry:
16151 %0 = bitcast <2 x i64> %__a to <4 x i32>
16152 %load = load i32, i32* %__b
16153 %vec = insertelement <4 x i32> undef, i32 %load, i32 0
16154 %1 = shufflevector <4 x i32> %vec, <4 x i32> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
16155 %2 = icmp ult <4 x i32> %0, %1
16156 %3 = bitcast i8 %__u to <8 x i1>
16157 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
16158 %4 = and <4 x i1> %extract.i, %2
16159 %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
16160 %6 = bitcast <32 x i1> %5 to i32
16161 ret i32 %6
16162 }
16165 define zeroext i64 @test_vpcmpultd_v4i1_v64i1_mask(<2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
16166 ; VLX-LABEL: test_vpcmpultd_v4i1_v64i1_mask:
16167 ; VLX: # %bb.0: # %entry
16168 ; VLX-NEXT: vpcmpltud %xmm1, %xmm0, %k0
16169 ; VLX-NEXT: kmovq %k0, %rax
16170 ; VLX-NEXT: retq
16171 ;
16172 ; NoVLX-LABEL: test_vpcmpultd_v4i1_v64i1_mask:
16173 ; NoVLX: # %bb.0: # %entry
16174 ; NoVLX-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
16175 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
16176 ; NoVLX-NEXT: vpcmpltud %zmm1, %zmm0, %k0
16177 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
16178 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
16179 ; NoVLX-NEXT: kmovw %k0, %eax
16180 ; NoVLX-NEXT: vzeroupper
16181 ; NoVLX-NEXT: retq
16182 entry:
16183 %0 = bitcast <2 x i64> %__a to <4 x i32>
16184 %1 = bitcast <2 x i64> %__b to <4 x i32>
16185 %2 = icmp ult <4 x i32> %0, %1
16186 %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
16187 %4 = bitcast <64 x i1> %3 to i64
16188 ret i64 %4
16189 }
16191 define zeroext i64 @test_vpcmpultd_v4i1_v64i1_mask_mem(<2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
16192 ; VLX-LABEL: test_vpcmpultd_v4i1_v64i1_mask_mem:
16193 ; VLX: # %bb.0: # %entry
16194 ; VLX-NEXT: vpcmpltud (%rdi), %xmm0, %k0
16195 ; VLX-NEXT: kmovq %k0, %rax
16196 ; VLX-NEXT: retq
16197 ;
16198 ; NoVLX-LABEL: test_vpcmpultd_v4i1_v64i1_mask_mem:
16199 ; NoVLX: # %bb.0: # %entry
16200 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
16201 ; NoVLX-NEXT: vmovdqa (%rdi), %xmm1
16202 ; NoVLX-NEXT: vpcmpltud %zmm1, %zmm0, %k0
16203 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
16204 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
16205 ; NoVLX-NEXT: kmovw %k0, %eax
16206 ; NoVLX-NEXT: vzeroupper
16207 ; NoVLX-NEXT: retq
16208 entry:
16209 %0 = bitcast <2 x i64> %__a to <4 x i32>
16210 %load = load <2 x i64>, <2 x i64>* %__b
16211 %1 = bitcast <2 x i64> %load to <4 x i32>
16212 %2 = icmp ult <4 x i32> %0, %1
16213 %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
16214 %4 = bitcast <64 x i1> %3 to i64
16215 ret i64 %4
16216 }
16218 define zeroext i64 @test_masked_vpcmpultd_v4i1_v64i1_mask(i8 zeroext %__u, <2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
16219 ; VLX-LABEL: test_masked_vpcmpultd_v4i1_v64i1_mask:
16220 ; VLX: # %bb.0: # %entry
16221 ; VLX-NEXT: kmovd %edi, %k1
16222 ; VLX-NEXT: vpcmpltud %xmm1, %xmm0, %k0 {%k1}
16223 ; VLX-NEXT: kmovq %k0, %rax
16224 ; VLX-NEXT: retq
16225 ;
16226 ; NoVLX-LABEL: test_masked_vpcmpultd_v4i1_v64i1_mask:
16227 ; NoVLX: # %bb.0: # %entry
16228 ; NoVLX-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
16229 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
16230 ; NoVLX-NEXT: kmovw %edi, %k1
16231 ; NoVLX-NEXT: vpcmpltud %zmm1, %zmm0, %k0 {%k1}
16232 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
16233 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
16234 ; NoVLX-NEXT: kmovw %k0, %eax
16235 ; NoVLX-NEXT: vzeroupper
16236 ; NoVLX-NEXT: retq
16237 entry:
16238 %0 = bitcast <2 x i64> %__a to <4 x i32>
16239 %1 = bitcast <2 x i64> %__b to <4 x i32>
16240 %2 = icmp ult <4 x i32> %0, %1
16241 %3 = bitcast i8 %__u to <8 x i1>
16242 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
16243 %4 = and <4 x i1> %2, %extract.i
16244 %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
16245 %6 = bitcast <64 x i1> %5 to i64
16246 ret i64 %6
16247 }
16249 define zeroext i64 @test_masked_vpcmpultd_v4i1_v64i1_mask_mem(i8 zeroext %__u, <2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
16250 ; VLX-LABEL: test_masked_vpcmpultd_v4i1_v64i1_mask_mem:
16251 ; VLX: # %bb.0: # %entry
16252 ; VLX-NEXT: kmovd %edi, %k1
16253 ; VLX-NEXT: vpcmpltud (%rsi), %xmm0, %k0 {%k1}
16254 ; VLX-NEXT: kmovq %k0, %rax
16255 ; VLX-NEXT: retq
16256 ;
16257 ; NoVLX-LABEL: test_masked_vpcmpultd_v4i1_v64i1_mask_mem:
16258 ; NoVLX: # %bb.0: # %entry
16259 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
16260 ; NoVLX-NEXT: vmovdqa (%rsi), %xmm1
16261 ; NoVLX-NEXT: kmovw %edi, %k1
16262 ; NoVLX-NEXT: vpcmpltud %zmm1, %zmm0, %k0 {%k1}
16263 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
16264 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
16265 ; NoVLX-NEXT: kmovw %k0, %eax
16266 ; NoVLX-NEXT: vzeroupper
16267 ; NoVLX-NEXT: retq
16268 entry:
16269 %0 = bitcast <2 x i64> %__a to <4 x i32>
16270 %load = load <2 x i64>, <2 x i64>* %__b
16271 %1 = bitcast <2 x i64> %load to <4 x i32>
16272 %2 = icmp ult <4 x i32> %0, %1
16273 %3 = bitcast i8 %__u to <8 x i1>
16274 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
16275 %4 = and <4 x i1> %2, %extract.i
16276 %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
16277 %6 = bitcast <64 x i1> %5 to i64
16278 ret i64 %6
16279 }
16282 define zeroext i64 @test_vpcmpultd_v4i1_v64i1_mask_mem_b(<2 x i64> %__a, i32* %__b) local_unnamed_addr {
16283 ; VLX-LABEL: test_vpcmpultd_v4i1_v64i1_mask_mem_b:
16284 ; VLX: # %bb.0: # %entry
16285 ; VLX-NEXT: vpcmpltud (%rdi){1to4}, %xmm0, %k0
16286 ; VLX-NEXT: kmovq %k0, %rax
16287 ; VLX-NEXT: retq
16288 ;
16289 ; NoVLX-LABEL: test_vpcmpultd_v4i1_v64i1_mask_mem_b:
16290 ; NoVLX: # %bb.0: # %entry
16291 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
16292 ; NoVLX-NEXT: vpcmpltud (%rdi){1to16}, %zmm0, %k0
16293 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
16294 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
16295 ; NoVLX-NEXT: kmovw %k0, %eax
16296 ; NoVLX-NEXT: vzeroupper
16297 ; NoVLX-NEXT: retq
16298 entry:
16299 %0 = bitcast <2 x i64> %__a to <4 x i32>
16300 %load = load i32, i32* %__b
16301 %vec = insertelement <4 x i32> undef, i32 %load, i32 0
16302 %1 = shufflevector <4 x i32> %vec, <4 x i32> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
16303 %2 = icmp ult <4 x i32> %0, %1
16304 %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
16305 %4 = bitcast <64 x i1> %3 to i64
16306 ret i64 %4
16307 }
16309 define zeroext i64 @test_masked_vpcmpultd_v4i1_v64i1_mask_mem_b(i8 zeroext %__u, <2 x i64> %__a, i32* %__b) local_unnamed_addr {
16310 ; VLX-LABEL: test_masked_vpcmpultd_v4i1_v64i1_mask_mem_b:
16311 ; VLX: # %bb.0: # %entry
16312 ; VLX-NEXT: kmovd %edi, %k1
16313 ; VLX-NEXT: vpcmpltud (%rsi){1to4}, %xmm0, %k0 {%k1}
16314 ; VLX-NEXT: kmovq %k0, %rax
16315 ; VLX-NEXT: retq
16316 ;
16317 ; NoVLX-LABEL: test_masked_vpcmpultd_v4i1_v64i1_mask_mem_b:
16318 ; NoVLX: # %bb.0: # %entry
16319 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
16320 ; NoVLX-NEXT: kmovw %edi, %k1
16321 ; NoVLX-NEXT: vpcmpltud (%rsi){1to16}, %zmm0, %k0 {%k1}
16322 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
16323 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
16324 ; NoVLX-NEXT: kmovw %k0, %eax
16325 ; NoVLX-NEXT: vzeroupper
16326 ; NoVLX-NEXT: retq
16327 entry:
16328 %0 = bitcast <2 x i64> %__a to <4 x i32>
16329 %load = load i32, i32* %__b
16330 %vec = insertelement <4 x i32> undef, i32 %load, i32 0
16331 %1 = shufflevector <4 x i32> %vec, <4 x i32> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
16332 %2 = icmp ult <4 x i32> %0, %1
16333 %3 = bitcast i8 %__u to <8 x i1>
16334 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
16335 %4 = and <4 x i1> %extract.i, %2
16336 %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
16337 %6 = bitcast <64 x i1> %5 to i64
16338 ret i64 %6
16339 }
16342 define zeroext i16 @test_vpcmpultd_v8i1_v16i1_mask(<4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
16343 ; VLX-LABEL: test_vpcmpultd_v8i1_v16i1_mask:
16344 ; VLX: # %bb.0: # %entry
16345 ; VLX-NEXT: vpcmpltud %ymm1, %ymm0, %k0
16346 ; VLX-NEXT: kmovd %k0, %eax
16347 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
16348 ; VLX-NEXT: vzeroupper
16349 ; VLX-NEXT: retq
16350 ;
16351 ; NoVLX-LABEL: test_vpcmpultd_v8i1_v16i1_mask:
16352 ; NoVLX: # %bb.0: # %entry
16353 ; NoVLX-NEXT: # kill: def $ymm1 killed $ymm1 def $zmm1
16354 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
16355 ; NoVLX-NEXT: vpcmpltud %zmm1, %zmm0, %k0
16356 ; NoVLX-NEXT: kshiftlw $8, %k0, %k0
16357 ; NoVLX-NEXT: kshiftrw $8, %k0, %k0
16358 ; NoVLX-NEXT: kmovw %k0, %eax
16359 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
16360 ; NoVLX-NEXT: vzeroupper
16361 ; NoVLX-NEXT: retq
16362 entry:
16363 %0 = bitcast <4 x i64> %__a to <8 x i32>
16364 %1 = bitcast <4 x i64> %__b to <8 x i32>
16365 %2 = icmp ult <8 x i32> %0, %1
16366 %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
16367 %4 = bitcast <16 x i1> %3 to i16
16368 ret i16 %4
16369 }
16371 define zeroext i16 @test_vpcmpultd_v8i1_v16i1_mask_mem(<4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
16372 ; VLX-LABEL: test_vpcmpultd_v8i1_v16i1_mask_mem:
16373 ; VLX: # %bb.0: # %entry
16374 ; VLX-NEXT: vpcmpltud (%rdi), %ymm0, %k0
16375 ; VLX-NEXT: kmovd %k0, %eax
16376 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
16377 ; VLX-NEXT: vzeroupper
16378 ; VLX-NEXT: retq
16379 ;
16380 ; NoVLX-LABEL: test_vpcmpultd_v8i1_v16i1_mask_mem:
16381 ; NoVLX: # %bb.0: # %entry
16382 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
16383 ; NoVLX-NEXT: vmovdqa (%rdi), %ymm1
16384 ; NoVLX-NEXT: vpcmpltud %zmm1, %zmm0, %k0
16385 ; NoVLX-NEXT: kshiftlw $8, %k0, %k0
16386 ; NoVLX-NEXT: kshiftrw $8, %k0, %k0
16387 ; NoVLX-NEXT: kmovw %k0, %eax
16388 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
16389 ; NoVLX-NEXT: vzeroupper
16390 ; NoVLX-NEXT: retq
16391 entry:
16392 %0 = bitcast <4 x i64> %__a to <8 x i32>
16393 %load = load <4 x i64>, <4 x i64>* %__b
16394 %1 = bitcast <4 x i64> %load to <8 x i32>
16395 %2 = icmp ult <8 x i32> %0, %1
16396 %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
16397 %4 = bitcast <16 x i1> %3 to i16
16398 ret i16 %4
16399 }
16401 define zeroext i16 @test_masked_vpcmpultd_v8i1_v16i1_mask(i8 zeroext %__u, <4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
16402 ; VLX-LABEL: test_masked_vpcmpultd_v8i1_v16i1_mask:
16403 ; VLX: # %bb.0: # %entry
16404 ; VLX-NEXT: kmovd %edi, %k1
16405 ; VLX-NEXT: vpcmpltud %ymm1, %ymm0, %k0 {%k1}
16406 ; VLX-NEXT: kmovd %k0, %eax
16407 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
16408 ; VLX-NEXT: vzeroupper
16409 ; VLX-NEXT: retq
16410 ;
16411 ; NoVLX-LABEL: test_masked_vpcmpultd_v8i1_v16i1_mask:
16412 ; NoVLX: # %bb.0: # %entry
16413 ; NoVLX-NEXT: # kill: def $ymm1 killed $ymm1 def $zmm1
16414 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
16415 ; NoVLX-NEXT: kmovw %edi, %k1
16416 ; NoVLX-NEXT: vpcmpltud %zmm1, %zmm0, %k0 {%k1}
16417 ; NoVLX-NEXT: kshiftlw $8, %k0, %k0
16418 ; NoVLX-NEXT: kshiftrw $8, %k0, %k0
16419 ; NoVLX-NEXT: kmovw %k0, %eax
16420 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
16421 ; NoVLX-NEXT: vzeroupper
16422 ; NoVLX-NEXT: retq
16423 entry:
16424 %0 = bitcast <4 x i64> %__a to <8 x i32>
16425 %1 = bitcast <4 x i64> %__b to <8 x i32>
16426 %2 = icmp ult <8 x i32> %0, %1
16427 %3 = bitcast i8 %__u to <8 x i1>
16428 %4 = and <8 x i1> %2, %3
16429 %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
16430 %6 = bitcast <16 x i1> %5 to i16
16431 ret i16 %6
16432 }
16434 define zeroext i16 @test_masked_vpcmpultd_v8i1_v16i1_mask_mem(i8 zeroext %__u, <4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
16435 ; VLX-LABEL: test_masked_vpcmpultd_v8i1_v16i1_mask_mem:
16436 ; VLX: # %bb.0: # %entry
16437 ; VLX-NEXT: kmovd %edi, %k1
16438 ; VLX-NEXT: vpcmpltud (%rsi), %ymm0, %k0 {%k1}
16439 ; VLX-NEXT: kmovd %k0, %eax
16440 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
16441 ; VLX-NEXT: vzeroupper
16442 ; VLX-NEXT: retq
16443 ;
16444 ; NoVLX-LABEL: test_masked_vpcmpultd_v8i1_v16i1_mask_mem:
16445 ; NoVLX: # %bb.0: # %entry
16446 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
16447 ; NoVLX-NEXT: vmovdqa (%rsi), %ymm1
16448 ; NoVLX-NEXT: kmovw %edi, %k1
16449 ; NoVLX-NEXT: vpcmpltud %zmm1, %zmm0, %k0 {%k1}
16450 ; NoVLX-NEXT: kshiftlw $8, %k0, %k0
16451 ; NoVLX-NEXT: kshiftrw $8, %k0, %k0
16452 ; NoVLX-NEXT: kmovw %k0, %eax
16453 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
16454 ; NoVLX-NEXT: vzeroupper
16455 ; NoVLX-NEXT: retq
16456 entry:
16457 %0 = bitcast <4 x i64> %__a to <8 x i32>
16458 %load = load <4 x i64>, <4 x i64>* %__b
16459 %1 = bitcast <4 x i64> %load to <8 x i32>
16460 %2 = icmp ult <8 x i32> %0, %1
16461 %3 = bitcast i8 %__u to <8 x i1>
16462 %4 = and <8 x i1> %2, %3
16463 %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
16464 %6 = bitcast <16 x i1> %5 to i16
16465 ret i16 %6
16466 }
16469 define zeroext i16 @test_vpcmpultd_v8i1_v16i1_mask_mem_b(<4 x i64> %__a, i32* %__b) local_unnamed_addr {
16470 ; VLX-LABEL: test_vpcmpultd_v8i1_v16i1_mask_mem_b:
16471 ; VLX: # %bb.0: # %entry
16472 ; VLX-NEXT: vpcmpltud (%rdi){1to8}, %ymm0, %k0
16473 ; VLX-NEXT: kmovd %k0, %eax
16474 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
16475 ; VLX-NEXT: vzeroupper
16476 ; VLX-NEXT: retq
16477 ;
16478 ; NoVLX-LABEL: test_vpcmpultd_v8i1_v16i1_mask_mem_b:
16479 ; NoVLX: # %bb.0: # %entry
16480 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
16481 ; NoVLX-NEXT: vpcmpltud (%rdi){1to16}, %zmm0, %k0
16482 ; NoVLX-NEXT: kshiftlw $8, %k0, %k0
16483 ; NoVLX-NEXT: kshiftrw $8, %k0, %k0
16484 ; NoVLX-NEXT: kmovw %k0, %eax
16485 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
16486 ; NoVLX-NEXT: vzeroupper
16487 ; NoVLX-NEXT: retq
16488 entry:
16489 %0 = bitcast <4 x i64> %__a to <8 x i32>
16490 %load = load i32, i32* %__b
16491 %vec = insertelement <8 x i32> undef, i32 %load, i32 0
16492 %1 = shufflevector <8 x i32> %vec, <8 x i32> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
16493 %2 = icmp ult <8 x i32> %0, %1
16494 %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
16495 %4 = bitcast <16 x i1> %3 to i16
16496 ret i16 %4
16497 }
16499 define zeroext i16 @test_masked_vpcmpultd_v8i1_v16i1_mask_mem_b(i8 zeroext %__u, <4 x i64> %__a, i32* %__b) local_unnamed_addr {
16500 ; VLX-LABEL: test_masked_vpcmpultd_v8i1_v16i1_mask_mem_b:
16501 ; VLX: # %bb.0: # %entry
16502 ; VLX-NEXT: kmovd %edi, %k1
16503 ; VLX-NEXT: vpcmpltud (%rsi){1to8}, %ymm0, %k0 {%k1}
16504 ; VLX-NEXT: kmovd %k0, %eax
16505 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
16506 ; VLX-NEXT: vzeroupper
16507 ; VLX-NEXT: retq
16508 ;
16509 ; NoVLX-LABEL: test_masked_vpcmpultd_v8i1_v16i1_mask_mem_b:
16510 ; NoVLX: # %bb.0: # %entry
16511 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
16512 ; NoVLX-NEXT: kmovw %edi, %k1
16513 ; NoVLX-NEXT: vpcmpltud (%rsi){1to16}, %zmm0, %k0 {%k1}
16514 ; NoVLX-NEXT: kshiftlw $8, %k0, %k0
16515 ; NoVLX-NEXT: kshiftrw $8, %k0, %k0
16516 ; NoVLX-NEXT: kmovw %k0, %eax
16517 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
16518 ; NoVLX-NEXT: vzeroupper
16519 ; NoVLX-NEXT: retq
16520 entry:
16521 %0 = bitcast <4 x i64> %__a to <8 x i32>
16522 %load = load i32, i32* %__b
16523 %vec = insertelement <8 x i32> undef, i32 %load, i32 0
16524 %1 = shufflevector <8 x i32> %vec, <8 x i32> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
16525 %2 = icmp ult <8 x i32> %0, %1
16526 %3 = bitcast i8 %__u to <8 x i1>
16527 %4 = and <8 x i1> %3, %2
16528 %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
16529 %6 = bitcast <16 x i1> %5 to i16
16530 ret i16 %6
16531 }
16534 define zeroext i32 @test_vpcmpultd_v8i1_v32i1_mask(<4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
16535 ; VLX-LABEL: test_vpcmpultd_v8i1_v32i1_mask:
16536 ; VLX: # %bb.0: # %entry
16537 ; VLX-NEXT: vpcmpltud %ymm1, %ymm0, %k0
16538 ; VLX-NEXT: kmovd %k0, %eax
16539 ; VLX-NEXT: vzeroupper
16540 ; VLX-NEXT: retq
16541 ;
16542 ; NoVLX-LABEL: test_vpcmpultd_v8i1_v32i1_mask:
16543 ; NoVLX: # %bb.0: # %entry
16544 ; NoVLX-NEXT: # kill: def $ymm1 killed $ymm1 def $zmm1
16545 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
16546 ; NoVLX-NEXT: vpcmpltud %zmm1, %zmm0, %k0
16547 ; NoVLX-NEXT: kshiftlw $8, %k0, %k0
16548 ; NoVLX-NEXT: kshiftrw $8, %k0, %k0
16549 ; NoVLX-NEXT: kmovw %k0, %eax
16550 ; NoVLX-NEXT: vzeroupper
16551 ; NoVLX-NEXT: retq
16552 entry:
16553 %0 = bitcast <4 x i64> %__a to <8 x i32>
16554 %1 = bitcast <4 x i64> %__b to <8 x i32>
16555 %2 = icmp ult <8 x i32> %0, %1
16556 %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
16557 %4 = bitcast <32 x i1> %3 to i32
16558 ret i32 %4
16559 }
16561 define zeroext i32 @test_vpcmpultd_v8i1_v32i1_mask_mem(<4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
16562 ; VLX-LABEL: test_vpcmpultd_v8i1_v32i1_mask_mem:
16563 ; VLX: # %bb.0: # %entry
16564 ; VLX-NEXT: vpcmpltud (%rdi), %ymm0, %k0
16565 ; VLX-NEXT: kmovd %k0, %eax
16566 ; VLX-NEXT: vzeroupper
16567 ; VLX-NEXT: retq
16568 ;
16569 ; NoVLX-LABEL: test_vpcmpultd_v8i1_v32i1_mask_mem:
16570 ; NoVLX: # %bb.0: # %entry
16571 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
16572 ; NoVLX-NEXT: vmovdqa (%rdi), %ymm1
16573 ; NoVLX-NEXT: vpcmpltud %zmm1, %zmm0, %k0
16574 ; NoVLX-NEXT: kshiftlw $8, %k0, %k0
16575 ; NoVLX-NEXT: kshiftrw $8, %k0, %k0
16576 ; NoVLX-NEXT: kmovw %k0, %eax
16577 ; NoVLX-NEXT: vzeroupper
16578 ; NoVLX-NEXT: retq
16579 entry:
16580 %0 = bitcast <4 x i64> %__a to <8 x i32>
16581 %load = load <4 x i64>, <4 x i64>* %__b
16582 %1 = bitcast <4 x i64> %load to <8 x i32>
16583 %2 = icmp ult <8 x i32> %0, %1
16584 %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
16585 %4 = bitcast <32 x i1> %3 to i32
16586 ret i32 %4
16587 }
16589 define zeroext i32 @test_masked_vpcmpultd_v8i1_v32i1_mask(i8 zeroext %__u, <4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
16590 ; VLX-LABEL: test_masked_vpcmpultd_v8i1_v32i1_mask:
16591 ; VLX: # %bb.0: # %entry
16592 ; VLX-NEXT: kmovd %edi, %k1
16593 ; VLX-NEXT: vpcmpltud %ymm1, %ymm0, %k0 {%k1}
16594 ; VLX-NEXT: kmovd %k0, %eax
16595 ; VLX-NEXT: vzeroupper
16596 ; VLX-NEXT: retq
16597 ;
16598 ; NoVLX-LABEL: test_masked_vpcmpultd_v8i1_v32i1_mask:
16599 ; NoVLX: # %bb.0: # %entry
16600 ; NoVLX-NEXT: # kill: def $ymm1 killed $ymm1 def $zmm1
16601 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
16602 ; NoVLX-NEXT: kmovw %edi, %k1
16603 ; NoVLX-NEXT: vpcmpltud %zmm1, %zmm0, %k0 {%k1}
16604 ; NoVLX-NEXT: kshiftlw $8, %k0, %k0
16605 ; NoVLX-NEXT: kshiftrw $8, %k0, %k0
16606 ; NoVLX-NEXT: kmovw %k0, %eax
16607 ; NoVLX-NEXT: vzeroupper
16608 ; NoVLX-NEXT: retq
16609 entry:
16610 %0 = bitcast <4 x i64> %__a to <8 x i32>
16611 %1 = bitcast <4 x i64> %__b to <8 x i32>
16612 %2 = icmp ult <8 x i32> %0, %1
16613 %3 = bitcast i8 %__u to <8 x i1>
16614 %4 = and <8 x i1> %2, %3
16615 %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
16616 %6 = bitcast <32 x i1> %5 to i32
16617 ret i32 %6
16618 }
16620 define zeroext i32 @test_masked_vpcmpultd_v8i1_v32i1_mask_mem(i8 zeroext %__u, <4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
16621 ; VLX-LABEL: test_masked_vpcmpultd_v8i1_v32i1_mask_mem:
16622 ; VLX: # %bb.0: # %entry
16623 ; VLX-NEXT: kmovd %edi, %k1
16624 ; VLX-NEXT: vpcmpltud (%rsi), %ymm0, %k0 {%k1}
16625 ; VLX-NEXT: kmovd %k0, %eax
16626 ; VLX-NEXT: vzeroupper
16627 ; VLX-NEXT: retq
16628 ;
16629 ; NoVLX-LABEL: test_masked_vpcmpultd_v8i1_v32i1_mask_mem:
16630 ; NoVLX: # %bb.0: # %entry
16631 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
16632 ; NoVLX-NEXT: vmovdqa (%rsi), %ymm1
16633 ; NoVLX-NEXT: kmovw %edi, %k1
16634 ; NoVLX-NEXT: vpcmpltud %zmm1, %zmm0, %k0 {%k1}
16635 ; NoVLX-NEXT: kshiftlw $8, %k0, %k0
16636 ; NoVLX-NEXT: kshiftrw $8, %k0, %k0
16637 ; NoVLX-NEXT: kmovw %k0, %eax
16638 ; NoVLX-NEXT: vzeroupper
16639 ; NoVLX-NEXT: retq
16640 entry:
16641 %0 = bitcast <4 x i64> %__a to <8 x i32>
16642 %load = load <4 x i64>, <4 x i64>* %__b
16643 %1 = bitcast <4 x i64> %load to <8 x i32>
16644 %2 = icmp ult <8 x i32> %0, %1
16645 %3 = bitcast i8 %__u to <8 x i1>
16646 %4 = and <8 x i1> %2, %3
16647 %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
16648 %6 = bitcast <32 x i1> %5 to i32
16649 ret i32 %6
16650 }
16653 define zeroext i32 @test_vpcmpultd_v8i1_v32i1_mask_mem_b(<4 x i64> %__a, i32* %__b) local_unnamed_addr {
16654 ; VLX-LABEL: test_vpcmpultd_v8i1_v32i1_mask_mem_b:
16655 ; VLX: # %bb.0: # %entry
16656 ; VLX-NEXT: vpcmpltud (%rdi){1to8}, %ymm0, %k0
16657 ; VLX-NEXT: kmovd %k0, %eax
16658 ; VLX-NEXT: vzeroupper
16659 ; VLX-NEXT: retq
16660 ;
16661 ; NoVLX-LABEL: test_vpcmpultd_v8i1_v32i1_mask_mem_b:
16662 ; NoVLX: # %bb.0: # %entry
16663 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
16664 ; NoVLX-NEXT: vpcmpltud (%rdi){1to16}, %zmm0, %k0
16665 ; NoVLX-NEXT: kshiftlw $8, %k0, %k0
16666 ; NoVLX-NEXT: kshiftrw $8, %k0, %k0
16667 ; NoVLX-NEXT: kmovw %k0, %eax
16668 ; NoVLX-NEXT: vzeroupper
16669 ; NoVLX-NEXT: retq
16670 entry:
16671 %0 = bitcast <4 x i64> %__a to <8 x i32>
16672 %load = load i32, i32* %__b
16673 %vec = insertelement <8 x i32> undef, i32 %load, i32 0
16674 %1 = shufflevector <8 x i32> %vec, <8 x i32> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
16675 %2 = icmp ult <8 x i32> %0, %1
16676 %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
16677 %4 = bitcast <32 x i1> %3 to i32
16678 ret i32 %4
16679 }
16681 define zeroext i32 @test_masked_vpcmpultd_v8i1_v32i1_mask_mem_b(i8 zeroext %__u, <4 x i64> %__a, i32* %__b) local_unnamed_addr {
16682 ; VLX-LABEL: test_masked_vpcmpultd_v8i1_v32i1_mask_mem_b:
16683 ; VLX: # %bb.0: # %entry
16684 ; VLX-NEXT: kmovd %edi, %k1
16685 ; VLX-NEXT: vpcmpltud (%rsi){1to8}, %ymm0, %k0 {%k1}
16686 ; VLX-NEXT: kmovd %k0, %eax
16687 ; VLX-NEXT: vzeroupper
16688 ; VLX-NEXT: retq
16689 ;
16690 ; NoVLX-LABEL: test_masked_vpcmpultd_v8i1_v32i1_mask_mem_b:
16691 ; NoVLX: # %bb.0: # %entry
16692 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
16693 ; NoVLX-NEXT: kmovw %edi, %k1
16694 ; NoVLX-NEXT: vpcmpltud (%rsi){1to16}, %zmm0, %k0 {%k1}
16695 ; NoVLX-NEXT: kshiftlw $8, %k0, %k0
16696 ; NoVLX-NEXT: kshiftrw $8, %k0, %k0
16697 ; NoVLX-NEXT: kmovw %k0, %eax
16698 ; NoVLX-NEXT: vzeroupper
16699 ; NoVLX-NEXT: retq
16700 entry:
16701 %0 = bitcast <4 x i64> %__a to <8 x i32>
16702 %load = load i32, i32* %__b
16703 %vec = insertelement <8 x i32> undef, i32 %load, i32 0
16704 %1 = shufflevector <8 x i32> %vec, <8 x i32> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
16705 %2 = icmp ult <8 x i32> %0, %1
16706 %3 = bitcast i8 %__u to <8 x i1>
16707 %4 = and <8 x i1> %3, %2
16708 %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
16709 %6 = bitcast <32 x i1> %5 to i32
16710 ret i32 %6
16711 }
16714 define zeroext i64 @test_vpcmpultd_v8i1_v64i1_mask(<4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
16715 ; VLX-LABEL: test_vpcmpultd_v8i1_v64i1_mask:
16716 ; VLX: # %bb.0: # %entry
16717 ; VLX-NEXT: vpcmpltud %ymm1, %ymm0, %k0
16718 ; VLX-NEXT: kmovq %k0, %rax
16719 ; VLX-NEXT: vzeroupper
16720 ; VLX-NEXT: retq
16721 ;
16722 ; NoVLX-LABEL: test_vpcmpultd_v8i1_v64i1_mask:
16723 ; NoVLX: # %bb.0: # %entry
16724 ; NoVLX-NEXT: # kill: def $ymm1 killed $ymm1 def $zmm1
16725 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
16726 ; NoVLX-NEXT: vpcmpltud %zmm1, %zmm0, %k0
16727 ; NoVLX-NEXT: kshiftlw $8, %k0, %k0
16728 ; NoVLX-NEXT: kshiftrw $8, %k0, %k0
16729 ; NoVLX-NEXT: kmovw %k0, %eax
16730 ; NoVLX-NEXT: vzeroupper
16731 ; NoVLX-NEXT: retq
16732 entry:
16733 %0 = bitcast <4 x i64> %__a to <8 x i32>
16734 %1 = bitcast <4 x i64> %__b to <8 x i32>
16735 %2 = icmp ult <8 x i32> %0, %1
16736 %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
16737 %4 = bitcast <64 x i1> %3 to i64
16738 ret i64 %4
16739 }
16741 define zeroext i64 @test_vpcmpultd_v8i1_v64i1_mask_mem(<4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
16742 ; VLX-LABEL: test_vpcmpultd_v8i1_v64i1_mask_mem:
16743 ; VLX: # %bb.0: # %entry
16744 ; VLX-NEXT: vpcmpltud (%rdi), %ymm0, %k0
16745 ; VLX-NEXT: kmovq %k0, %rax
16746 ; VLX-NEXT: vzeroupper
16747 ; VLX-NEXT: retq
16748 ;
16749 ; NoVLX-LABEL: test_vpcmpultd_v8i1_v64i1_mask_mem:
16750 ; NoVLX: # %bb.0: # %entry
16751 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
16752 ; NoVLX-NEXT: vmovdqa (%rdi), %ymm1
16753 ; NoVLX-NEXT: vpcmpltud %zmm1, %zmm0, %k0
16754 ; NoVLX-NEXT: kshiftlw $8, %k0, %k0
16755 ; NoVLX-NEXT: kshiftrw $8, %k0, %k0
16756 ; NoVLX-NEXT: kmovw %k0, %eax
16757 ; NoVLX-NEXT: vzeroupper
16758 ; NoVLX-NEXT: retq
16759 entry:
16760 %0 = bitcast <4 x i64> %__a to <8 x i32>
16761 %load = load <4 x i64>, <4 x i64>* %__b
16762 %1 = bitcast <4 x i64> %load to <8 x i32>
16763 %2 = icmp ult <8 x i32> %0, %1
16764 %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
16765 %4 = bitcast <64 x i1> %3 to i64
16766 ret i64 %4
16767 }
16769 define zeroext i64 @test_masked_vpcmpultd_v8i1_v64i1_mask(i8 zeroext %__u, <4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
16770 ; VLX-LABEL: test_masked_vpcmpultd_v8i1_v64i1_mask:
16771 ; VLX: # %bb.0: # %entry
16772 ; VLX-NEXT: kmovd %edi, %k1
16773 ; VLX-NEXT: vpcmpltud %ymm1, %ymm0, %k0 {%k1}
16774 ; VLX-NEXT: kmovq %k0, %rax
16775 ; VLX-NEXT: vzeroupper
16776 ; VLX-NEXT: retq
16777 ;
16778 ; NoVLX-LABEL: test_masked_vpcmpultd_v8i1_v64i1_mask:
16779 ; NoVLX: # %bb.0: # %entry
16780 ; NoVLX-NEXT: # kill: def $ymm1 killed $ymm1 def $zmm1
16781 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
16782 ; NoVLX-NEXT: kmovw %edi, %k1
16783 ; NoVLX-NEXT: vpcmpltud %zmm1, %zmm0, %k0 {%k1}
16784 ; NoVLX-NEXT: kshiftlw $8, %k0, %k0
16785 ; NoVLX-NEXT: kshiftrw $8, %k0, %k0
16786 ; NoVLX-NEXT: kmovw %k0, %eax
16787 ; NoVLX-NEXT: vzeroupper
16788 ; NoVLX-NEXT: retq
16789 entry:
16790 %0 = bitcast <4 x i64> %__a to <8 x i32>
16791 %1 = bitcast <4 x i64> %__b to <8 x i32>
16792 %2 = icmp ult <8 x i32> %0, %1
16793 %3 = bitcast i8 %__u to <8 x i1>
16794 %4 = and <8 x i1> %2, %3
16795 %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
16796 %6 = bitcast <64 x i1> %5 to i64
16797 ret i64 %6
16798 }
16800 define zeroext i64 @test_masked_vpcmpultd_v8i1_v64i1_mask_mem(i8 zeroext %__u, <4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
16801 ; VLX-LABEL: test_masked_vpcmpultd_v8i1_v64i1_mask_mem:
16802 ; VLX: # %bb.0: # %entry
16803 ; VLX-NEXT: kmovd %edi, %k1
16804 ; VLX-NEXT: vpcmpltud (%rsi), %ymm0, %k0 {%k1}
16805 ; VLX-NEXT: kmovq %k0, %rax
16806 ; VLX-NEXT: vzeroupper
16807 ; VLX-NEXT: retq
16808 ;
16809 ; NoVLX-LABEL: test_masked_vpcmpultd_v8i1_v64i1_mask_mem:
16810 ; NoVLX: # %bb.0: # %entry
16811 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
16812 ; NoVLX-NEXT: vmovdqa (%rsi), %ymm1
16813 ; NoVLX-NEXT: kmovw %edi, %k1
16814 ; NoVLX-NEXT: vpcmpltud %zmm1, %zmm0, %k0 {%k1}
16815 ; NoVLX-NEXT: kshiftlw $8, %k0, %k0
16816 ; NoVLX-NEXT: kshiftrw $8, %k0, %k0
16817 ; NoVLX-NEXT: kmovw %k0, %eax
16818 ; NoVLX-NEXT: vzeroupper
16819 ; NoVLX-NEXT: retq
16820 entry:
16821 %0 = bitcast <4 x i64> %__a to <8 x i32>
16822 %load = load <4 x i64>, <4 x i64>* %__b
16823 %1 = bitcast <4 x i64> %load to <8 x i32>
16824 %2 = icmp ult <8 x i32> %0, %1
16825 %3 = bitcast i8 %__u to <8 x i1>
16826 %4 = and <8 x i1> %2, %3
16827 %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
16828 %6 = bitcast <64 x i1> %5 to i64
16829 ret i64 %6
16830 }
16833 define zeroext i64 @test_vpcmpultd_v8i1_v64i1_mask_mem_b(<4 x i64> %__a, i32* %__b) local_unnamed_addr {
16834 ; VLX-LABEL: test_vpcmpultd_v8i1_v64i1_mask_mem_b:
16835 ; VLX: # %bb.0: # %entry
16836 ; VLX-NEXT: vpcmpltud (%rdi){1to8}, %ymm0, %k0
16837 ; VLX-NEXT: kmovq %k0, %rax
16838 ; VLX-NEXT: vzeroupper
16839 ; VLX-NEXT: retq
16840 ;
16841 ; NoVLX-LABEL: test_vpcmpultd_v8i1_v64i1_mask_mem_b:
16842 ; NoVLX: # %bb.0: # %entry
16843 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
16844 ; NoVLX-NEXT: vpcmpltud (%rdi){1to16}, %zmm0, %k0
16845 ; NoVLX-NEXT: kshiftlw $8, %k0, %k0
16846 ; NoVLX-NEXT: kshiftrw $8, %k0, %k0
16847 ; NoVLX-NEXT: kmovw %k0, %eax
16848 ; NoVLX-NEXT: vzeroupper
16849 ; NoVLX-NEXT: retq
16850 entry:
16851 %0 = bitcast <4 x i64> %__a to <8 x i32>
16852 %load = load i32, i32* %__b
16853 %vec = insertelement <8 x i32> undef, i32 %load, i32 0
16854 %1 = shufflevector <8 x i32> %vec, <8 x i32> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
16855 %2 = icmp ult <8 x i32> %0, %1
16856 %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
16857 %4 = bitcast <64 x i1> %3 to i64
16858 ret i64 %4
16859 }
16861 define zeroext i64 @test_masked_vpcmpultd_v8i1_v64i1_mask_mem_b(i8 zeroext %__u, <4 x i64> %__a, i32* %__b) local_unnamed_addr {
16862 ; VLX-LABEL: test_masked_vpcmpultd_v8i1_v64i1_mask_mem_b:
16863 ; VLX: # %bb.0: # %entry
16864 ; VLX-NEXT: kmovd %edi, %k1
16865 ; VLX-NEXT: vpcmpltud (%rsi){1to8}, %ymm0, %k0 {%k1}
16866 ; VLX-NEXT: kmovq %k0, %rax
16867 ; VLX-NEXT: vzeroupper
16868 ; VLX-NEXT: retq
16869 ;
16870 ; NoVLX-LABEL: test_masked_vpcmpultd_v8i1_v64i1_mask_mem_b:
16871 ; NoVLX: # %bb.0: # %entry
16872 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
16873 ; NoVLX-NEXT: kmovw %edi, %k1
16874 ; NoVLX-NEXT: vpcmpltud (%rsi){1to16}, %zmm0, %k0 {%k1}
16875 ; NoVLX-NEXT: kshiftlw $8, %k0, %k0
16876 ; NoVLX-NEXT: kshiftrw $8, %k0, %k0
16877 ; NoVLX-NEXT: kmovw %k0, %eax
16878 ; NoVLX-NEXT: vzeroupper
16879 ; NoVLX-NEXT: retq
16880 entry:
16881 %0 = bitcast <4 x i64> %__a to <8 x i32>
16882 %load = load i32, i32* %__b
16883 %vec = insertelement <8 x i32> undef, i32 %load, i32 0
16884 %1 = shufflevector <8 x i32> %vec, <8 x i32> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
16885 %2 = icmp ult <8 x i32> %0, %1
16886 %3 = bitcast i8 %__u to <8 x i1>
16887 %4 = and <8 x i1> %3, %2
16888 %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
16889 %6 = bitcast <64 x i1> %5 to i64
16890 ret i64 %6
16891 }
16894 define zeroext i32 @test_vpcmpultd_v16i1_v32i1_mask(<8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
16895 ; VLX-LABEL: test_vpcmpultd_v16i1_v32i1_mask:
16896 ; VLX: # %bb.0: # %entry
16897 ; VLX-NEXT: vpcmpltud %zmm1, %zmm0, %k0
16898 ; VLX-NEXT: kmovd %k0, %eax
16899 ; VLX-NEXT: vzeroupper
16900 ; VLX-NEXT: retq
16901 ;
16902 ; NoVLX-LABEL: test_vpcmpultd_v16i1_v32i1_mask:
16903 ; NoVLX: # %bb.0: # %entry
16904 ; NoVLX-NEXT: vpcmpltud %zmm1, %zmm0, %k0
16905 ; NoVLX-NEXT: kmovw %k0, %eax
16906 ; NoVLX-NEXT: vzeroupper
16907 ; NoVLX-NEXT: retq
16908 entry:
16909 %0 = bitcast <8 x i64> %__a to <16 x i32>
16910 %1 = bitcast <8 x i64> %__b to <16 x i32>
16911 %2 = icmp ult <16 x i32> %0, %1
16912 %3 = shufflevector <16 x i1> %2, <16 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>
16913 %4 = bitcast <32 x i1> %3 to i32
16914 ret i32 %4
16915 }
16917 define zeroext i32 @test_vpcmpultd_v16i1_v32i1_mask_mem(<8 x i64> %__a, <8 x i64>* %__b) local_unnamed_addr {
16918 ; VLX-LABEL: test_vpcmpultd_v16i1_v32i1_mask_mem:
16919 ; VLX: # %bb.0: # %entry
16920 ; VLX-NEXT: vpcmpltud (%rdi), %zmm0, %k0
16921 ; VLX-NEXT: kmovd %k0, %eax
16922 ; VLX-NEXT: vzeroupper
16923 ; VLX-NEXT: retq
16924 ;
16925 ; NoVLX-LABEL: test_vpcmpultd_v16i1_v32i1_mask_mem:
16926 ; NoVLX: # %bb.0: # %entry
16927 ; NoVLX-NEXT: vpcmpltud (%rdi), %zmm0, %k0
16928 ; NoVLX-NEXT: kmovw %k0, %eax
16929 ; NoVLX-NEXT: vzeroupper
16930 ; NoVLX-NEXT: retq
16931 entry:
16932 %0 = bitcast <8 x i64> %__a to <16 x i32>
16933 %load = load <8 x i64>, <8 x i64>* %__b
16934 %1 = bitcast <8 x i64> %load to <16 x i32>
16935 %2 = icmp ult <16 x i32> %0, %1
16936 %3 = shufflevector <16 x i1> %2, <16 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>
16937 %4 = bitcast <32 x i1> %3 to i32
16938 ret i32 %4
16939 }
16941 define zeroext i32 @test_masked_vpcmpultd_v16i1_v32i1_mask(i16 zeroext %__u, <8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
16942 ; VLX-LABEL: test_masked_vpcmpultd_v16i1_v32i1_mask:
16943 ; VLX: # %bb.0: # %entry
16944 ; VLX-NEXT: kmovd %edi, %k1
16945 ; VLX-NEXT: vpcmpltud %zmm1, %zmm0, %k0 {%k1}
16946 ; VLX-NEXT: kmovd %k0, %eax
16947 ; VLX-NEXT: vzeroupper
16948 ; VLX-NEXT: retq
16949 ;
16950 ; NoVLX-LABEL: test_masked_vpcmpultd_v16i1_v32i1_mask:
16951 ; NoVLX: # %bb.0: # %entry
16952 ; NoVLX-NEXT: vpcmpltud %zmm1, %zmm0, %k0
16953 ; NoVLX-NEXT: kmovw %k0, %eax
16954 ; NoVLX-NEXT: andl %edi, %eax
16955 ; NoVLX-NEXT: vzeroupper
16956 ; NoVLX-NEXT: retq
16957 entry:
16958 %0 = bitcast <8 x i64> %__a to <16 x i32>
16959 %1 = bitcast <8 x i64> %__b to <16 x i32>
16960 %2 = icmp ult <16 x i32> %0, %1
16961 %3 = bitcast i16 %__u to <16 x i1>
16962 %4 = and <16 x i1> %2, %3
16963 %5 = shufflevector <16 x i1> %4, <16 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>
16964 %6 = bitcast <32 x i1> %5 to i32
16965 ret i32 %6
16966 }
16968 define zeroext i32 @test_masked_vpcmpultd_v16i1_v32i1_mask_mem(i16 zeroext %__u, <8 x i64> %__a, <8 x i64>* %__b) local_unnamed_addr {
16969 ; VLX-LABEL: test_masked_vpcmpultd_v16i1_v32i1_mask_mem:
16970 ; VLX: # %bb.0: # %entry
16971 ; VLX-NEXT: kmovd %edi, %k1
16972 ; VLX-NEXT: vpcmpltud (%rsi), %zmm0, %k0 {%k1}
16973 ; VLX-NEXT: kmovd %k0, %eax
16974 ; VLX-NEXT: vzeroupper
16975 ; VLX-NEXT: retq
16976 ;
16977 ; NoVLX-LABEL: test_masked_vpcmpultd_v16i1_v32i1_mask_mem:
16978 ; NoVLX: # %bb.0: # %entry
16979 ; NoVLX-NEXT: vpcmpltud (%rsi), %zmm0, %k0
16980 ; NoVLX-NEXT: kmovw %k0, %eax
16981 ; NoVLX-NEXT: andl %edi, %eax
16982 ; NoVLX-NEXT: vzeroupper
16983 ; NoVLX-NEXT: retq
16984 entry:
16985 %0 = bitcast <8 x i64> %__a to <16 x i32>
16986 %load = load <8 x i64>, <8 x i64>* %__b
16987 %1 = bitcast <8 x i64> %load to <16 x i32>
16988 %2 = icmp ult <16 x i32> %0, %1
16989 %3 = bitcast i16 %__u to <16 x i1>
16990 %4 = and <16 x i1> %2, %3
16991 %5 = shufflevector <16 x i1> %4, <16 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>
16992 %6 = bitcast <32 x i1> %5 to i32
16993 ret i32 %6
16994 }
16997 define zeroext i32 @test_vpcmpultd_v16i1_v32i1_mask_mem_b(<8 x i64> %__a, i32* %__b) local_unnamed_addr {
16998 ; VLX-LABEL: test_vpcmpultd_v16i1_v32i1_mask_mem_b:
16999 ; VLX: # %bb.0: # %entry
17000 ; VLX-NEXT: vpcmpltud (%rdi){1to16}, %zmm0, %k0
17001 ; VLX-NEXT: kmovd %k0, %eax
17002 ; VLX-NEXT: vzeroupper
17003 ; VLX-NEXT: retq
17004 ;
17005 ; NoVLX-LABEL: test_vpcmpultd_v16i1_v32i1_mask_mem_b:
17006 ; NoVLX: # %bb.0: # %entry
17007 ; NoVLX-NEXT: vpcmpltud (%rdi){1to16}, %zmm0, %k0
17008 ; NoVLX-NEXT: kmovw %k0, %eax
17009 ; NoVLX-NEXT: vzeroupper
17010 ; NoVLX-NEXT: retq
17011 entry:
17012 %0 = bitcast <8 x i64> %__a to <16 x i32>
17013 %load = load i32, i32* %__b
17014 %vec = insertelement <16 x i32> undef, i32 %load, i32 0
17015 %1 = shufflevector <16 x i32> %vec, <16 x i32> undef, <16 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
17016 %2 = icmp ult <16 x i32> %0, %1
17017 %3 = shufflevector <16 x i1> %2, <16 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>
17018 %4 = bitcast <32 x i1> %3 to i32
17019 ret i32 %4
17020 }
17022 define zeroext i32 @test_masked_vpcmpultd_v16i1_v32i1_mask_mem_b(i16 zeroext %__u, <8 x i64> %__a, i32* %__b) local_unnamed_addr {
17023 ; VLX-LABEL: test_masked_vpcmpultd_v16i1_v32i1_mask_mem_b:
17024 ; VLX: # %bb.0: # %entry
17025 ; VLX-NEXT: kmovd %edi, %k1
17026 ; VLX-NEXT: vpcmpltud (%rsi){1to16}, %zmm0, %k0 {%k1}
17027 ; VLX-NEXT: kmovd %k0, %eax
17028 ; VLX-NEXT: vzeroupper
17029 ; VLX-NEXT: retq
17030 ;
17031 ; NoVLX-LABEL: test_masked_vpcmpultd_v16i1_v32i1_mask_mem_b:
17032 ; NoVLX: # %bb.0: # %entry
17033 ; NoVLX-NEXT: vpcmpltud (%rsi){1to16}, %zmm0, %k0
17034 ; NoVLX-NEXT: kmovw %k0, %eax
17035 ; NoVLX-NEXT: andl %edi, %eax
17036 ; NoVLX-NEXT: vzeroupper
17037 ; NoVLX-NEXT: retq
17038 entry:
17039 %0 = bitcast <8 x i64> %__a to <16 x i32>
17040 %load = load i32, i32* %__b
17041 %vec = insertelement <16 x i32> undef, i32 %load, i32 0
17042 %1 = shufflevector <16 x i32> %vec, <16 x i32> undef, <16 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
17043 %2 = icmp ult <16 x i32> %0, %1
17044 %3 = bitcast i16 %__u to <16 x i1>
17045 %4 = and <16 x i1> %3, %2
17046 %5 = shufflevector <16 x i1> %4, <16 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>
17047 %6 = bitcast <32 x i1> %5 to i32
17048 ret i32 %6
17049 }
17052 define zeroext i64 @test_vpcmpultd_v16i1_v64i1_mask(<8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
17053 ; VLX-LABEL: test_vpcmpultd_v16i1_v64i1_mask:
17054 ; VLX: # %bb.0: # %entry
17055 ; VLX-NEXT: vpcmpltud %zmm1, %zmm0, %k0
17056 ; VLX-NEXT: kmovq %k0, %rax
17057 ; VLX-NEXT: vzeroupper
17058 ; VLX-NEXT: retq
17059 ;
17060 ; NoVLX-LABEL: test_vpcmpultd_v16i1_v64i1_mask:
17061 ; NoVLX: # %bb.0: # %entry
17062 ; NoVLX-NEXT: vpcmpltud %zmm1, %zmm0, %k0
17063 ; NoVLX-NEXT: kmovw %k0, %eax
17064 ; NoVLX-NEXT: vzeroupper
17065 ; NoVLX-NEXT: retq
17066 entry:
17067 %0 = bitcast <8 x i64> %__a to <16 x i32>
17068 %1 = bitcast <8 x i64> %__b to <16 x i32>
17069 %2 = icmp ult <16 x i32> %0, %1
17070 %3 = shufflevector <16 x i1> %2, <16 x i1> zeroinitializer, <64 x i32> <i32 0,i32 1,i32 2,i32 3,i32 4,i32 5,i32 6,i32 7,i32 8,i32 9,i32 10,i32 11,i32 12,i32 13,i32 14,i32 15,i32 16,i32 17,i32 18,i32 19,i32 20,i32 21,i32 22,i32 23,i32 24,i32 25,i32 26,i32 27,i32 28,i32 29,i32 30,i32 31,i32 16,i32 17,i32 18,i32 19,i32 20,i32 21,i32 22,i32 23,i32 24,i32 25,i32 26,i32 27,i32 28,i32 29,i32 30,i32 31,i32 16,i32 17,i32 18,i32 19,i32 20,i32 21,i32 22,i32 23,i32 24,i32 25,i32 26,i32 27,i32 28,i32 29,i32 30,i32 31>
17071 %4 = bitcast <64 x i1> %3 to i64
17072 ret i64 %4
17073 }
17075 define zeroext i64 @test_vpcmpultd_v16i1_v64i1_mask_mem(<8 x i64> %__a, <8 x i64>* %__b) local_unnamed_addr {
17076 ; VLX-LABEL: test_vpcmpultd_v16i1_v64i1_mask_mem:
17077 ; VLX: # %bb.0: # %entry
17078 ; VLX-NEXT: vpcmpltud (%rdi), %zmm0, %k0
17079 ; VLX-NEXT: kmovq %k0, %rax
17080 ; VLX-NEXT: vzeroupper
17081 ; VLX-NEXT: retq
17082 ;
17083 ; NoVLX-LABEL: test_vpcmpultd_v16i1_v64i1_mask_mem:
17084 ; NoVLX: # %bb.0: # %entry
17085 ; NoVLX-NEXT: vpcmpltud (%rdi), %zmm0, %k0
17086 ; NoVLX-NEXT: kmovw %k0, %eax
17087 ; NoVLX-NEXT: vzeroupper
17088 ; NoVLX-NEXT: retq
17089 entry:
17090 %0 = bitcast <8 x i64> %__a to <16 x i32>
17091 %load = load <8 x i64>, <8 x i64>* %__b
17092 %1 = bitcast <8 x i64> %load to <16 x i32>
17093 %2 = icmp ult <16 x i32> %0, %1
17094 %3 = shufflevector <16 x i1> %2, <16 x i1> zeroinitializer, <64 x i32> <i32 0,i32 1,i32 2,i32 3,i32 4,i32 5,i32 6,i32 7,i32 8,i32 9,i32 10,i32 11,i32 12,i32 13,i32 14,i32 15,i32 16,i32 17,i32 18,i32 19,i32 20,i32 21,i32 22,i32 23,i32 24,i32 25,i32 26,i32 27,i32 28,i32 29,i32 30,i32 31,i32 16,i32 17,i32 18,i32 19,i32 20,i32 21,i32 22,i32 23,i32 24,i32 25,i32 26,i32 27,i32 28,i32 29,i32 30,i32 31,i32 16,i32 17,i32 18,i32 19,i32 20,i32 21,i32 22,i32 23,i32 24,i32 25,i32 26,i32 27,i32 28,i32 29,i32 30,i32 31>
17095 %4 = bitcast <64 x i1> %3 to i64
17096 ret i64 %4
17097 }
17099 define zeroext i64 @test_masked_vpcmpultd_v16i1_v64i1_mask(i16 zeroext %__u, <8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
17100 ; VLX-LABEL: test_masked_vpcmpultd_v16i1_v64i1_mask:
17101 ; VLX: # %bb.0: # %entry
17102 ; VLX-NEXT: kmovd %edi, %k1
17103 ; VLX-NEXT: vpcmpltud %zmm1, %zmm0, %k0 {%k1}
17104 ; VLX-NEXT: kmovq %k0, %rax
17105 ; VLX-NEXT: vzeroupper
17106 ; VLX-NEXT: retq
17107 ;
17108 ; NoVLX-LABEL: test_masked_vpcmpultd_v16i1_v64i1_mask:
17109 ; NoVLX: # %bb.0: # %entry
17110 ; NoVLX-NEXT: vpcmpltud %zmm1, %zmm0, %k0
17111 ; NoVLX-NEXT: kmovw %k0, %eax
17112 ; NoVLX-NEXT: andl %edi, %eax
17113 ; NoVLX-NEXT: vzeroupper
17114 ; NoVLX-NEXT: retq
17115 entry:
17116 %0 = bitcast <8 x i64> %__a to <16 x i32>
17117 %1 = bitcast <8 x i64> %__b to <16 x i32>
17118 %2 = icmp ult <16 x i32> %0, %1
17119 %3 = bitcast i16 %__u to <16 x i1>
17120 %4 = and <16 x i1> %2, %3
17121 %5 = shufflevector <16 x i1> %4, <16 x i1> zeroinitializer, <64 x i32> <i32 0,i32 1,i32 2,i32 3,i32 4,i32 5,i32 6,i32 7,i32 8,i32 9,i32 10,i32 11,i32 12,i32 13,i32 14,i32 15,i32 16,i32 17,i32 18,i32 19,i32 20,i32 21,i32 22,i32 23,i32 24,i32 25,i32 26,i32 27,i32 28,i32 29,i32 30,i32 31,i32 16,i32 17,i32 18,i32 19,i32 20,i32 21,i32 22,i32 23,i32 24,i32 25,i32 26,i32 27,i32 28,i32 29,i32 30,i32 31,i32 16,i32 17,i32 18,i32 19,i32 20,i32 21,i32 22,i32 23,i32 24,i32 25,i32 26,i32 27,i32 28,i32 29,i32 30,i32 31>
17122 %6 = bitcast <64 x i1> %5 to i64
17123 ret i64 %6
17124 }
17126 define zeroext i64 @test_masked_vpcmpultd_v16i1_v64i1_mask_mem(i16 zeroext %__u, <8 x i64> %__a, <8 x i64>* %__b) local_unnamed_addr {
17127 ; VLX-LABEL: test_masked_vpcmpultd_v16i1_v64i1_mask_mem:
17128 ; VLX: # %bb.0: # %entry
17129 ; VLX-NEXT: kmovd %edi, %k1
17130 ; VLX-NEXT: vpcmpltud (%rsi), %zmm0, %k0 {%k1}
17131 ; VLX-NEXT: kmovq %k0, %rax
17132 ; VLX-NEXT: vzeroupper
17133 ; VLX-NEXT: retq
17134 ;
17135 ; NoVLX-LABEL: test_masked_vpcmpultd_v16i1_v64i1_mask_mem:
17136 ; NoVLX: # %bb.0: # %entry
17137 ; NoVLX-NEXT: vpcmpltud (%rsi), %zmm0, %k0
17138 ; NoVLX-NEXT: kmovw %k0, %eax
17139 ; NoVLX-NEXT: andl %edi, %eax
17140 ; NoVLX-NEXT: vzeroupper
17141 ; NoVLX-NEXT: retq
17142 entry:
17143 %0 = bitcast <8 x i64> %__a to <16 x i32>
17144 %load = load <8 x i64>, <8 x i64>* %__b
17145 %1 = bitcast <8 x i64> %load to <16 x i32>
17146 %2 = icmp ult <16 x i32> %0, %1
17147 %3 = bitcast i16 %__u to <16 x i1>
17148 %4 = and <16 x i1> %2, %3
17149 %5 = shufflevector <16 x i1> %4, <16 x i1> zeroinitializer, <64 x i32> <i32 0,i32 1,i32 2,i32 3,i32 4,i32 5,i32 6,i32 7,i32 8,i32 9,i32 10,i32 11,i32 12,i32 13,i32 14,i32 15,i32 16,i32 17,i32 18,i32 19,i32 20,i32 21,i32 22,i32 23,i32 24,i32 25,i32 26,i32 27,i32 28,i32 29,i32 30,i32 31,i32 16,i32 17,i32 18,i32 19,i32 20,i32 21,i32 22,i32 23,i32 24,i32 25,i32 26,i32 27,i32 28,i32 29,i32 30,i32 31,i32 16,i32 17,i32 18,i32 19,i32 20,i32 21,i32 22,i32 23,i32 24,i32 25,i32 26,i32 27,i32 28,i32 29,i32 30,i32 31>
17150 %6 = bitcast <64 x i1> %5 to i64
17151 ret i64 %6
17152 }
17155 define zeroext i64 @test_vpcmpultd_v16i1_v64i1_mask_mem_b(<8 x i64> %__a, i32* %__b) local_unnamed_addr {
17156 ; VLX-LABEL: test_vpcmpultd_v16i1_v64i1_mask_mem_b:
17157 ; VLX: # %bb.0: # %entry
17158 ; VLX-NEXT: vpcmpltud (%rdi){1to16}, %zmm0, %k0
17159 ; VLX-NEXT: kmovq %k0, %rax
17160 ; VLX-NEXT: vzeroupper
17161 ; VLX-NEXT: retq
17162 ;
17163 ; NoVLX-LABEL: test_vpcmpultd_v16i1_v64i1_mask_mem_b:
17164 ; NoVLX: # %bb.0: # %entry
17165 ; NoVLX-NEXT: vpcmpltud (%rdi){1to16}, %zmm0, %k0
17166 ; NoVLX-NEXT: kmovw %k0, %eax
17167 ; NoVLX-NEXT: vzeroupper
17168 ; NoVLX-NEXT: retq
17169 entry:
17170 %0 = bitcast <8 x i64> %__a to <16 x i32>
17171 %load = load i32, i32* %__b
17172 %vec = insertelement <16 x i32> undef, i32 %load, i32 0
17173 %1 = shufflevector <16 x i32> %vec, <16 x i32> undef, <16 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
17174 %2 = icmp ult <16 x i32> %0, %1
17175 %3 = shufflevector <16 x i1> %2, <16 x i1> zeroinitializer, <64 x i32> <i32 0,i32 1,i32 2,i32 3,i32 4,i32 5,i32 6,i32 7,i32 8,i32 9,i32 10,i32 11,i32 12,i32 13,i32 14,i32 15,i32 16,i32 17,i32 18,i32 19,i32 20,i32 21,i32 22,i32 23,i32 24,i32 25,i32 26,i32 27,i32 28,i32 29,i32 30,i32 31,i32 16,i32 17,i32 18,i32 19,i32 20,i32 21,i32 22,i32 23,i32 24,i32 25,i32 26,i32 27,i32 28,i32 29,i32 30,i32 31,i32 16,i32 17,i32 18,i32 19,i32 20,i32 21,i32 22,i32 23,i32 24,i32 25,i32 26,i32 27,i32 28,i32 29,i32 30,i32 31>
17176 %4 = bitcast <64 x i1> %3 to i64
17177 ret i64 %4
17178 }
17180 define zeroext i64 @test_masked_vpcmpultd_v16i1_v64i1_mask_mem_b(i16 zeroext %__u, <8 x i64> %__a, i32* %__b) local_unnamed_addr {
17181 ; VLX-LABEL: test_masked_vpcmpultd_v16i1_v64i1_mask_mem_b:
17182 ; VLX: # %bb.0: # %entry
17183 ; VLX-NEXT: kmovd %edi, %k1
17184 ; VLX-NEXT: vpcmpltud (%rsi){1to16}, %zmm0, %k0 {%k1}
17185 ; VLX-NEXT: kmovq %k0, %rax
17186 ; VLX-NEXT: vzeroupper
17187 ; VLX-NEXT: retq
17188 ;
17189 ; NoVLX-LABEL: test_masked_vpcmpultd_v16i1_v64i1_mask_mem_b:
17190 ; NoVLX: # %bb.0: # %entry
17191 ; NoVLX-NEXT: vpcmpltud (%rsi){1to16}, %zmm0, %k0
17192 ; NoVLX-NEXT: kmovw %k0, %eax
17193 ; NoVLX-NEXT: andl %edi, %eax
17194 ; NoVLX-NEXT: vzeroupper
17195 ; NoVLX-NEXT: retq
17196 entry:
17197 %0 = bitcast <8 x i64> %__a to <16 x i32>
17198 %load = load i32, i32* %__b
17199 %vec = insertelement <16 x i32> undef, i32 %load, i32 0
17200 %1 = shufflevector <16 x i32> %vec, <16 x i32> undef, <16 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
17201 %2 = icmp ult <16 x i32> %0, %1
17202 %3 = bitcast i16 %__u to <16 x i1>
17203 %4 = and <16 x i1> %3, %2
17204 %5 = shufflevector <16 x i1> %4, <16 x i1> zeroinitializer, <64 x i32> <i32 0,i32 1,i32 2,i32 3,i32 4,i32 5,i32 6,i32 7,i32 8,i32 9,i32 10,i32 11,i32 12,i32 13,i32 14,i32 15,i32 16,i32 17,i32 18,i32 19,i32 20,i32 21,i32 22,i32 23,i32 24,i32 25,i32 26,i32 27,i32 28,i32 29,i32 30,i32 31,i32 16,i32 17,i32 18,i32 19,i32 20,i32 21,i32 22,i32 23,i32 24,i32 25,i32 26,i32 27,i32 28,i32 29,i32 30,i32 31,i32 16,i32 17,i32 18,i32 19,i32 20,i32 21,i32 22,i32 23,i32 24,i32 25,i32 26,i32 27,i32 28,i32 29,i32 30,i32 31>
17205 %6 = bitcast <64 x i1> %5 to i64
17206 ret i64 %6
17207 }
17210 define zeroext i4 @test_vpcmpultq_v2i1_v4i1_mask(<2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
17211 ; VLX-LABEL: test_vpcmpultq_v2i1_v4i1_mask:
17212 ; VLX: # %bb.0: # %entry
17213 ; VLX-NEXT: vpcmpltuq %xmm1, %xmm0, %k0
17214 ; VLX-NEXT: kmovb %k0, %eax
17215 ; VLX-NEXT: retq
17216 ;
17217 ; NoVLX-LABEL: test_vpcmpultq_v2i1_v4i1_mask:
17218 ; NoVLX: # %bb.0: # %entry
17219 ; NoVLX-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
17220 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
17221 ; NoVLX-NEXT: vpcmpltuq %zmm1, %zmm0, %k0
17222 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
17223 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
17224 ; NoVLX-NEXT: kmovw %k0, %eax
17225 ; NoVLX-NEXT: vzeroupper
17226 ; NoVLX-NEXT: retq
17227 entry:
17228 %0 = bitcast <2 x i64> %__a to <2 x i64>
17229 %1 = bitcast <2 x i64> %__b to <2 x i64>
17230 %2 = icmp ult <2 x i64> %0, %1
17231 %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
17232 %4 = bitcast <4 x i1> %3 to i4
17233 ret i4 %4
17234 }
17236 define zeroext i4 @test_vpcmpultq_v2i1_v4i1_mask_mem(<2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
17237 ; VLX-LABEL: test_vpcmpultq_v2i1_v4i1_mask_mem:
17238 ; VLX: # %bb.0: # %entry
17239 ; VLX-NEXT: vpcmpltuq (%rdi), %xmm0, %k0
17240 ; VLX-NEXT: kmovb %k0, %eax
17241 ; VLX-NEXT: retq
17242 ;
17243 ; NoVLX-LABEL: test_vpcmpultq_v2i1_v4i1_mask_mem:
17244 ; NoVLX: # %bb.0: # %entry
17245 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
17246 ; NoVLX-NEXT: vmovdqa (%rdi), %xmm1
17247 ; NoVLX-NEXT: vpcmpltuq %zmm1, %zmm0, %k0
17248 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
17249 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
17250 ; NoVLX-NEXT: kmovw %k0, %eax
17251 ; NoVLX-NEXT: vzeroupper
17252 ; NoVLX-NEXT: retq
17253 entry:
17254 %0 = bitcast <2 x i64> %__a to <2 x i64>
17255 %load = load <2 x i64>, <2 x i64>* %__b
17256 %1 = bitcast <2 x i64> %load to <2 x i64>
17257 %2 = icmp ult <2 x i64> %0, %1
17258 %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
17259 %4 = bitcast <4 x i1> %3 to i4
17260 ret i4 %4
17261 }
17263 define zeroext i4 @test_masked_vpcmpultq_v2i1_v4i1_mask(i8 zeroext %__u, <2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
17264 ; VLX-LABEL: test_masked_vpcmpultq_v2i1_v4i1_mask:
17265 ; VLX: # %bb.0: # %entry
17266 ; VLX-NEXT: kmovd %edi, %k1
17267 ; VLX-NEXT: vpcmpltuq %xmm1, %xmm0, %k0 {%k1}
17268 ; VLX-NEXT: kmovb %k0, %eax
17269 ; VLX-NEXT: retq
17270 ;
17271 ; NoVLX-LABEL: test_masked_vpcmpultq_v2i1_v4i1_mask:
17272 ; NoVLX: # %bb.0: # %entry
17273 ; NoVLX-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
17274 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
17275 ; NoVLX-NEXT: kmovw %edi, %k1
17276 ; NoVLX-NEXT: vpcmpltuq %zmm1, %zmm0, %k0 {%k1}
17277 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
17278 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
17279 ; NoVLX-NEXT: kmovw %k0, %eax
17280 ; NoVLX-NEXT: vzeroupper
17281 ; NoVLX-NEXT: retq
17282 entry:
17283 %0 = bitcast <2 x i64> %__a to <2 x i64>
17284 %1 = bitcast <2 x i64> %__b to <2 x i64>
17285 %2 = icmp ult <2 x i64> %0, %1
17286 %3 = bitcast i8 %__u to <8 x i1>
17287 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <2 x i32> <i32 0, i32 1>
17288 %4 = and <2 x i1> %2, %extract.i
17289 %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
17290 %6 = bitcast <4 x i1> %5 to i4
17291 ret i4 %6
17292 }
17294 define zeroext i4 @test_masked_vpcmpultq_v2i1_v4i1_mask_mem(i8 zeroext %__u, <2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
17295 ; VLX-LABEL: test_masked_vpcmpultq_v2i1_v4i1_mask_mem:
17296 ; VLX: # %bb.0: # %entry
17297 ; VLX-NEXT: kmovd %edi, %k1
17298 ; VLX-NEXT: vpcmpltuq (%rsi), %xmm0, %k0 {%k1}
17299 ; VLX-NEXT: kmovb %k0, %eax
17300 ; VLX-NEXT: retq
17301 ;
17302 ; NoVLX-LABEL: test_masked_vpcmpultq_v2i1_v4i1_mask_mem:
17303 ; NoVLX: # %bb.0: # %entry
17304 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
17305 ; NoVLX-NEXT: vmovdqa (%rsi), %xmm1
17306 ; NoVLX-NEXT: kmovw %edi, %k1
17307 ; NoVLX-NEXT: vpcmpltuq %zmm1, %zmm0, %k0 {%k1}
17308 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
17309 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
17310 ; NoVLX-NEXT: kmovw %k0, %eax
17311 ; NoVLX-NEXT: vzeroupper
17312 ; NoVLX-NEXT: retq
17313 entry:
17314 %0 = bitcast <2 x i64> %__a to <2 x i64>
17315 %load = load <2 x i64>, <2 x i64>* %__b
17316 %1 = bitcast <2 x i64> %load to <2 x i64>
17317 %2 = icmp ult <2 x i64> %0, %1
17318 %3 = bitcast i8 %__u to <8 x i1>
17319 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <2 x i32> <i32 0, i32 1>
17320 %4 = and <2 x i1> %2, %extract.i
17321 %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
17322 %6 = bitcast <4 x i1> %5 to i4
17323 ret i4 %6
17324 }
17327 define zeroext i4 @test_vpcmpultq_v2i1_v4i1_mask_mem_b(<2 x i64> %__a, i64* %__b) local_unnamed_addr {
17328 ; VLX-LABEL: test_vpcmpultq_v2i1_v4i1_mask_mem_b:
17329 ; VLX: # %bb.0: # %entry
17330 ; VLX-NEXT: vpcmpltuq (%rdi){1to2}, %xmm0, %k0
17331 ; VLX-NEXT: kmovb %k0, %eax
17332 ; VLX-NEXT: retq
17333 ;
17334 ; NoVLX-LABEL: test_vpcmpultq_v2i1_v4i1_mask_mem_b:
17335 ; NoVLX: # %bb.0: # %entry
17336 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
17337 ; NoVLX-NEXT: vpcmpltuq (%rdi){1to8}, %zmm0, %k0
17338 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
17339 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
17340 ; NoVLX-NEXT: kmovw %k0, %eax
17341 ; NoVLX-NEXT: vzeroupper
17342 ; NoVLX-NEXT: retq
17343 entry:
17344 %0 = bitcast <2 x i64> %__a to <2 x i64>
17345 %load = load i64, i64* %__b
17346 %vec = insertelement <2 x i64> undef, i64 %load, i32 0
17347 %1 = shufflevector <2 x i64> %vec, <2 x i64> undef, <2 x i32> <i32 0, i32 0>
17348 %2 = icmp ult <2 x i64> %0, %1
17349 %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
17350 %4 = bitcast <4 x i1> %3 to i4
17351 ret i4 %4
17352 }
17354 define zeroext i4 @test_masked_vpcmpultq_v2i1_v4i1_mask_mem_b(i8 zeroext %__u, <2 x i64> %__a, i64* %__b) local_unnamed_addr {
17355 ; VLX-LABEL: test_masked_vpcmpultq_v2i1_v4i1_mask_mem_b:
17356 ; VLX: # %bb.0: # %entry
17357 ; VLX-NEXT: kmovd %edi, %k1
17358 ; VLX-NEXT: vpcmpltuq (%rsi){1to2}, %xmm0, %k0 {%k1}
17359 ; VLX-NEXT: kmovb %k0, %eax
17360 ; VLX-NEXT: retq
17361 ;
17362 ; NoVLX-LABEL: test_masked_vpcmpultq_v2i1_v4i1_mask_mem_b:
17363 ; NoVLX: # %bb.0: # %entry
17364 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
17365 ; NoVLX-NEXT: kmovw %edi, %k1
17366 ; NoVLX-NEXT: vpcmpltuq (%rsi){1to8}, %zmm0, %k0 {%k1}
17367 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
17368 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
17369 ; NoVLX-NEXT: kmovw %k0, %eax
17370 ; NoVLX-NEXT: vzeroupper
17371 ; NoVLX-NEXT: retq
17372 entry:
17373 %0 = bitcast <2 x i64> %__a to <2 x i64>
17374 %load = load i64, i64* %__b
17375 %vec = insertelement <2 x i64> undef, i64 %load, i32 0
17376 %1 = shufflevector <2 x i64> %vec, <2 x i64> undef, <2 x i32> <i32 0, i32 0>
17377 %2 = icmp ult <2 x i64> %0, %1
17378 %3 = bitcast i8 %__u to <8 x i1>
17379 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <2 x i32> <i32 0, i32 1>
17380 %4 = and <2 x i1> %extract.i, %2
17381 %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
17382 %6 = bitcast <4 x i1> %5 to i4
17383 ret i4 %6
17384 }
17387 define zeroext i8 @test_vpcmpultq_v2i1_v8i1_mask(<2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
17388 ; VLX-LABEL: test_vpcmpultq_v2i1_v8i1_mask:
17389 ; VLX: # %bb.0: # %entry
17390 ; VLX-NEXT: vpcmpltuq %xmm1, %xmm0, %k0
17391 ; VLX-NEXT: kmovd %k0, %eax
17392 ; VLX-NEXT: # kill: def $al killed $al killed $eax
17393 ; VLX-NEXT: retq
17394 ;
17395 ; NoVLX-LABEL: test_vpcmpultq_v2i1_v8i1_mask:
17396 ; NoVLX: # %bb.0: # %entry
17397 ; NoVLX-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
17398 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
17399 ; NoVLX-NEXT: vpcmpltuq %zmm1, %zmm0, %k0
17400 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
17401 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
17402 ; NoVLX-NEXT: kmovw %k0, %eax
17403 ; NoVLX-NEXT: # kill: def $al killed $al killed $eax
17404 ; NoVLX-NEXT: vzeroupper
17405 ; NoVLX-NEXT: retq
17406 entry:
17407 %0 = bitcast <2 x i64> %__a to <2 x i64>
17408 %1 = bitcast <2 x i64> %__b to <2 x i64>
17409 %2 = icmp ult <2 x i64> %0, %1
17410 %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
17411 %4 = bitcast <8 x i1> %3 to i8
17412 ret i8 %4
17413 }
17415 define zeroext i8 @test_vpcmpultq_v2i1_v8i1_mask_mem(<2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
17416 ; VLX-LABEL: test_vpcmpultq_v2i1_v8i1_mask_mem:
17417 ; VLX: # %bb.0: # %entry
17418 ; VLX-NEXT: vpcmpltuq (%rdi), %xmm0, %k0
17419 ; VLX-NEXT: kmovd %k0, %eax
17420 ; VLX-NEXT: # kill: def $al killed $al killed $eax
17421 ; VLX-NEXT: retq
17422 ;
17423 ; NoVLX-LABEL: test_vpcmpultq_v2i1_v8i1_mask_mem:
17424 ; NoVLX: # %bb.0: # %entry
17425 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
17426 ; NoVLX-NEXT: vmovdqa (%rdi), %xmm1
17427 ; NoVLX-NEXT: vpcmpltuq %zmm1, %zmm0, %k0
17428 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
17429 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
17430 ; NoVLX-NEXT: kmovw %k0, %eax
17431 ; NoVLX-NEXT: # kill: def $al killed $al killed $eax
17432 ; NoVLX-NEXT: vzeroupper
17433 ; NoVLX-NEXT: retq
17434 entry:
17435 %0 = bitcast <2 x i64> %__a to <2 x i64>
17436 %load = load <2 x i64>, <2 x i64>* %__b
17437 %1 = bitcast <2 x i64> %load to <2 x i64>
17438 %2 = icmp ult <2 x i64> %0, %1
17439 %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
17440 %4 = bitcast <8 x i1> %3 to i8
17441 ret i8 %4
17442 }
17444 define zeroext i8 @test_masked_vpcmpultq_v2i1_v8i1_mask(i8 zeroext %__u, <2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
17445 ; VLX-LABEL: test_masked_vpcmpultq_v2i1_v8i1_mask:
17446 ; VLX: # %bb.0: # %entry
17447 ; VLX-NEXT: kmovd %edi, %k1
17448 ; VLX-NEXT: vpcmpltuq %xmm1, %xmm0, %k0 {%k1}
17449 ; VLX-NEXT: kmovd %k0, %eax
17450 ; VLX-NEXT: # kill: def $al killed $al killed $eax
17451 ; VLX-NEXT: retq
17452 ;
17453 ; NoVLX-LABEL: test_masked_vpcmpultq_v2i1_v8i1_mask:
17454 ; NoVLX: # %bb.0: # %entry
17455 ; NoVLX-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
17456 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
17457 ; NoVLX-NEXT: kmovw %edi, %k1
17458 ; NoVLX-NEXT: vpcmpltuq %zmm1, %zmm0, %k0 {%k1}
17459 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
17460 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
17461 ; NoVLX-NEXT: kmovw %k0, %eax
17462 ; NoVLX-NEXT: # kill: def $al killed $al killed $eax
17463 ; NoVLX-NEXT: vzeroupper
17464 ; NoVLX-NEXT: retq
17465 entry:
17466 %0 = bitcast <2 x i64> %__a to <2 x i64>
17467 %1 = bitcast <2 x i64> %__b to <2 x i64>
17468 %2 = icmp ult <2 x i64> %0, %1
17469 %3 = bitcast i8 %__u to <8 x i1>
17470 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <2 x i32> <i32 0, i32 1>
17471 %4 = and <2 x i1> %2, %extract.i
17472 %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
17473 %6 = bitcast <8 x i1> %5 to i8
17474 ret i8 %6
17475 }
17477 define zeroext i8 @test_masked_vpcmpultq_v2i1_v8i1_mask_mem(i8 zeroext %__u, <2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
17478 ; VLX-LABEL: test_masked_vpcmpultq_v2i1_v8i1_mask_mem:
17479 ; VLX: # %bb.0: # %entry
17480 ; VLX-NEXT: kmovd %edi, %k1
17481 ; VLX-NEXT: vpcmpltuq (%rsi), %xmm0, %k0 {%k1}
17482 ; VLX-NEXT: kmovd %k0, %eax
17483 ; VLX-NEXT: # kill: def $al killed $al killed $eax
17484 ; VLX-NEXT: retq
17485 ;
17486 ; NoVLX-LABEL: test_masked_vpcmpultq_v2i1_v8i1_mask_mem:
17487 ; NoVLX: # %bb.0: # %entry
17488 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
17489 ; NoVLX-NEXT: vmovdqa (%rsi), %xmm1
17490 ; NoVLX-NEXT: kmovw %edi, %k1
17491 ; NoVLX-NEXT: vpcmpltuq %zmm1, %zmm0, %k0 {%k1}
17492 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
17493 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
17494 ; NoVLX-NEXT: kmovw %k0, %eax
17495 ; NoVLX-NEXT: # kill: def $al killed $al killed $eax
17496 ; NoVLX-NEXT: vzeroupper
17497 ; NoVLX-NEXT: retq
17498 entry:
17499 %0 = bitcast <2 x i64> %__a to <2 x i64>
17500 %load = load <2 x i64>, <2 x i64>* %__b
17501 %1 = bitcast <2 x i64> %load to <2 x i64>
17502 %2 = icmp ult <2 x i64> %0, %1
17503 %3 = bitcast i8 %__u to <8 x i1>
17504 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <2 x i32> <i32 0, i32 1>
17505 %4 = and <2 x i1> %2, %extract.i
17506 %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
17507 %6 = bitcast <8 x i1> %5 to i8
17508 ret i8 %6
17509 }
17512 define zeroext i8 @test_vpcmpultq_v2i1_v8i1_mask_mem_b(<2 x i64> %__a, i64* %__b) local_unnamed_addr {
17513 ; VLX-LABEL: test_vpcmpultq_v2i1_v8i1_mask_mem_b:
17514 ; VLX: # %bb.0: # %entry
17515 ; VLX-NEXT: vpcmpltuq (%rdi){1to2}, %xmm0, %k0
17516 ; VLX-NEXT: kmovd %k0, %eax
17517 ; VLX-NEXT: # kill: def $al killed $al killed $eax
17518 ; VLX-NEXT: retq
17519 ;
17520 ; NoVLX-LABEL: test_vpcmpultq_v2i1_v8i1_mask_mem_b:
17521 ; NoVLX: # %bb.0: # %entry
17522 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
17523 ; NoVLX-NEXT: vpcmpltuq (%rdi){1to8}, %zmm0, %k0
17524 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
17525 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
17526 ; NoVLX-NEXT: kmovw %k0, %eax
17527 ; NoVLX-NEXT: # kill: def $al killed $al killed $eax
17528 ; NoVLX-NEXT: vzeroupper
17529 ; NoVLX-NEXT: retq
17530 entry:
17531 %0 = bitcast <2 x i64> %__a to <2 x i64>
17532 %load = load i64, i64* %__b
17533 %vec = insertelement <2 x i64> undef, i64 %load, i32 0
17534 %1 = shufflevector <2 x i64> %vec, <2 x i64> undef, <2 x i32> <i32 0, i32 0>
17535 %2 = icmp ult <2 x i64> %0, %1
17536 %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
17537 %4 = bitcast <8 x i1> %3 to i8
17538 ret i8 %4
17539 }
17541 define zeroext i8 @test_masked_vpcmpultq_v2i1_v8i1_mask_mem_b(i8 zeroext %__u, <2 x i64> %__a, i64* %__b) local_unnamed_addr {
17542 ; VLX-LABEL: test_masked_vpcmpultq_v2i1_v8i1_mask_mem_b:
17543 ; VLX: # %bb.0: # %entry
17544 ; VLX-NEXT: kmovd %edi, %k1
17545 ; VLX-NEXT: vpcmpltuq (%rsi){1to2}, %xmm0, %k0 {%k1}
17546 ; VLX-NEXT: kmovd %k0, %eax
17547 ; VLX-NEXT: # kill: def $al killed $al killed $eax
17548 ; VLX-NEXT: retq
17549 ;
17550 ; NoVLX-LABEL: test_masked_vpcmpultq_v2i1_v8i1_mask_mem_b:
17551 ; NoVLX: # %bb.0: # %entry
17552 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
17553 ; NoVLX-NEXT: kmovw %edi, %k1
17554 ; NoVLX-NEXT: vpcmpltuq (%rsi){1to8}, %zmm0, %k0 {%k1}
17555 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
17556 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
17557 ; NoVLX-NEXT: kmovw %k0, %eax
17558 ; NoVLX-NEXT: # kill: def $al killed $al killed $eax
17559 ; NoVLX-NEXT: vzeroupper
17560 ; NoVLX-NEXT: retq
17561 entry:
17562 %0 = bitcast <2 x i64> %__a to <2 x i64>
17563 %load = load i64, i64* %__b
17564 %vec = insertelement <2 x i64> undef, i64 %load, i32 0
17565 %1 = shufflevector <2 x i64> %vec, <2 x i64> undef, <2 x i32> <i32 0, i32 0>
17566 %2 = icmp ult <2 x i64> %0, %1
17567 %3 = bitcast i8 %__u to <8 x i1>
17568 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <2 x i32> <i32 0, i32 1>
17569 %4 = and <2 x i1> %extract.i, %2
17570 %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
17571 %6 = bitcast <8 x i1> %5 to i8
17572 ret i8 %6
17573 }
17576 define zeroext i16 @test_vpcmpultq_v2i1_v16i1_mask(<2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
17577 ; VLX-LABEL: test_vpcmpultq_v2i1_v16i1_mask:
17578 ; VLX: # %bb.0: # %entry
17579 ; VLX-NEXT: vpcmpltuq %xmm1, %xmm0, %k0
17580 ; VLX-NEXT: kmovd %k0, %eax
17581 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
17582 ; VLX-NEXT: retq
17583 ;
17584 ; NoVLX-LABEL: test_vpcmpultq_v2i1_v16i1_mask:
17585 ; NoVLX: # %bb.0: # %entry
17586 ; NoVLX-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
17587 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
17588 ; NoVLX-NEXT: vpcmpltuq %zmm1, %zmm0, %k0
17589 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
17590 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
17591 ; NoVLX-NEXT: kmovw %k0, %eax
17592 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
17593 ; NoVLX-NEXT: vzeroupper
17594 ; NoVLX-NEXT: retq
17595 entry:
17596 %0 = bitcast <2 x i64> %__a to <2 x i64>
17597 %1 = bitcast <2 x i64> %__b to <2 x i64>
17598 %2 = icmp ult <2 x i64> %0, %1
17599 %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
17600 %4 = bitcast <16 x i1> %3 to i16
17601 ret i16 %4
17602 }
17604 define zeroext i16 @test_vpcmpultq_v2i1_v16i1_mask_mem(<2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
17605 ; VLX-LABEL: test_vpcmpultq_v2i1_v16i1_mask_mem:
17606 ; VLX: # %bb.0: # %entry
17607 ; VLX-NEXT: vpcmpltuq (%rdi), %xmm0, %k0
17608 ; VLX-NEXT: kmovd %k0, %eax
17609 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
17610 ; VLX-NEXT: retq
17611 ;
17612 ; NoVLX-LABEL: test_vpcmpultq_v2i1_v16i1_mask_mem:
17613 ; NoVLX: # %bb.0: # %entry
17614 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
17615 ; NoVLX-NEXT: vmovdqa (%rdi), %xmm1
17616 ; NoVLX-NEXT: vpcmpltuq %zmm1, %zmm0, %k0
17617 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
17618 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
17619 ; NoVLX-NEXT: kmovw %k0, %eax
17620 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
17621 ; NoVLX-NEXT: vzeroupper
17622 ; NoVLX-NEXT: retq
17623 entry:
17624 %0 = bitcast <2 x i64> %__a to <2 x i64>
17625 %load = load <2 x i64>, <2 x i64>* %__b
17626 %1 = bitcast <2 x i64> %load to <2 x i64>
17627 %2 = icmp ult <2 x i64> %0, %1
17628 %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
17629 %4 = bitcast <16 x i1> %3 to i16
17630 ret i16 %4
17631 }
17633 define zeroext i16 @test_masked_vpcmpultq_v2i1_v16i1_mask(i8 zeroext %__u, <2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
17634 ; VLX-LABEL: test_masked_vpcmpultq_v2i1_v16i1_mask:
17635 ; VLX: # %bb.0: # %entry
17636 ; VLX-NEXT: kmovd %edi, %k1
17637 ; VLX-NEXT: vpcmpltuq %xmm1, %xmm0, %k0 {%k1}
17638 ; VLX-NEXT: kmovd %k0, %eax
17639 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
17640 ; VLX-NEXT: retq
17641 ;
17642 ; NoVLX-LABEL: test_masked_vpcmpultq_v2i1_v16i1_mask:
17643 ; NoVLX: # %bb.0: # %entry
17644 ; NoVLX-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
17645 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
17646 ; NoVLX-NEXT: kmovw %edi, %k1
17647 ; NoVLX-NEXT: vpcmpltuq %zmm1, %zmm0, %k0 {%k1}
17648 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
17649 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
17650 ; NoVLX-NEXT: kmovw %k0, %eax
17651 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
17652 ; NoVLX-NEXT: vzeroupper
17653 ; NoVLX-NEXT: retq
17654 entry:
17655 %0 = bitcast <2 x i64> %__a to <2 x i64>
17656 %1 = bitcast <2 x i64> %__b to <2 x i64>
17657 %2 = icmp ult <2 x i64> %0, %1
17658 %3 = bitcast i8 %__u to <8 x i1>
17659 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <2 x i32> <i32 0, i32 1>
17660 %4 = and <2 x i1> %2, %extract.i
17661 %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
17662 %6 = bitcast <16 x i1> %5 to i16
17663 ret i16 %6
17664 }
17666 define zeroext i16 @test_masked_vpcmpultq_v2i1_v16i1_mask_mem(i8 zeroext %__u, <2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
17667 ; VLX-LABEL: test_masked_vpcmpultq_v2i1_v16i1_mask_mem:
17668 ; VLX: # %bb.0: # %entry
17669 ; VLX-NEXT: kmovd %edi, %k1
17670 ; VLX-NEXT: vpcmpltuq (%rsi), %xmm0, %k0 {%k1}
17671 ; VLX-NEXT: kmovd %k0, %eax
17672 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
17673 ; VLX-NEXT: retq
17674 ;
17675 ; NoVLX-LABEL: test_masked_vpcmpultq_v2i1_v16i1_mask_mem:
17676 ; NoVLX: # %bb.0: # %entry
17677 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
17678 ; NoVLX-NEXT: vmovdqa (%rsi), %xmm1
17679 ; NoVLX-NEXT: kmovw %edi, %k1
17680 ; NoVLX-NEXT: vpcmpltuq %zmm1, %zmm0, %k0 {%k1}
17681 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
17682 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
17683 ; NoVLX-NEXT: kmovw %k0, %eax
17684 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
17685 ; NoVLX-NEXT: vzeroupper
17686 ; NoVLX-NEXT: retq
17687 entry:
17688 %0 = bitcast <2 x i64> %__a to <2 x i64>
17689 %load = load <2 x i64>, <2 x i64>* %__b
17690 %1 = bitcast <2 x i64> %load to <2 x i64>
17691 %2 = icmp ult <2 x i64> %0, %1
17692 %3 = bitcast i8 %__u to <8 x i1>
17693 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <2 x i32> <i32 0, i32 1>
17694 %4 = and <2 x i1> %2, %extract.i
17695 %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
17696 %6 = bitcast <16 x i1> %5 to i16
17697 ret i16 %6
17698 }
17701 define zeroext i16 @test_vpcmpultq_v2i1_v16i1_mask_mem_b(<2 x i64> %__a, i64* %__b) local_unnamed_addr {
17702 ; VLX-LABEL: test_vpcmpultq_v2i1_v16i1_mask_mem_b:
17703 ; VLX: # %bb.0: # %entry
17704 ; VLX-NEXT: vpcmpltuq (%rdi){1to2}, %xmm0, %k0
17705 ; VLX-NEXT: kmovd %k0, %eax
17706 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
17707 ; VLX-NEXT: retq
17708 ;
17709 ; NoVLX-LABEL: test_vpcmpultq_v2i1_v16i1_mask_mem_b:
17710 ; NoVLX: # %bb.0: # %entry
17711 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
17712 ; NoVLX-NEXT: vpcmpltuq (%rdi){1to8}, %zmm0, %k0
17713 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
17714 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
17715 ; NoVLX-NEXT: kmovw %k0, %eax
17716 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
17717 ; NoVLX-NEXT: vzeroupper
17718 ; NoVLX-NEXT: retq
17719 entry:
17720 %0 = bitcast <2 x i64> %__a to <2 x i64>
17721 %load = load i64, i64* %__b
17722 %vec = insertelement <2 x i64> undef, i64 %load, i32 0
17723 %1 = shufflevector <2 x i64> %vec, <2 x i64> undef, <2 x i32> <i32 0, i32 0>
17724 %2 = icmp ult <2 x i64> %0, %1
17725 %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
17726 %4 = bitcast <16 x i1> %3 to i16
17727 ret i16 %4
17728 }
17730 define zeroext i16 @test_masked_vpcmpultq_v2i1_v16i1_mask_mem_b(i8 zeroext %__u, <2 x i64> %__a, i64* %__b) local_unnamed_addr {
17731 ; VLX-LABEL: test_masked_vpcmpultq_v2i1_v16i1_mask_mem_b:
17732 ; VLX: # %bb.0: # %entry
17733 ; VLX-NEXT: kmovd %edi, %k1
17734 ; VLX-NEXT: vpcmpltuq (%rsi){1to2}, %xmm0, %k0 {%k1}
17735 ; VLX-NEXT: kmovd %k0, %eax
17736 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
17737 ; VLX-NEXT: retq
17738 ;
17739 ; NoVLX-LABEL: test_masked_vpcmpultq_v2i1_v16i1_mask_mem_b:
17740 ; NoVLX: # %bb.0: # %entry
17741 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
17742 ; NoVLX-NEXT: kmovw %edi, %k1
17743 ; NoVLX-NEXT: vpcmpltuq (%rsi){1to8}, %zmm0, %k0 {%k1}
17744 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
17745 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
17746 ; NoVLX-NEXT: kmovw %k0, %eax
17747 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
17748 ; NoVLX-NEXT: vzeroupper
17749 ; NoVLX-NEXT: retq
17750 entry:
17751 %0 = bitcast <2 x i64> %__a to <2 x i64>
17752 %load = load i64, i64* %__b
17753 %vec = insertelement <2 x i64> undef, i64 %load, i32 0
17754 %1 = shufflevector <2 x i64> %vec, <2 x i64> undef, <2 x i32> <i32 0, i32 0>
17755 %2 = icmp ult <2 x i64> %0, %1
17756 %3 = bitcast i8 %__u to <8 x i1>
17757 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <2 x i32> <i32 0, i32 1>
17758 %4 = and <2 x i1> %extract.i, %2
17759 %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
17760 %6 = bitcast <16 x i1> %5 to i16
17761 ret i16 %6
17762 }
17765 define zeroext i32 @test_vpcmpultq_v2i1_v32i1_mask(<2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
17766 ; VLX-LABEL: test_vpcmpultq_v2i1_v32i1_mask:
17767 ; VLX: # %bb.0: # %entry
17768 ; VLX-NEXT: vpcmpltuq %xmm1, %xmm0, %k0
17769 ; VLX-NEXT: kmovd %k0, %eax
17770 ; VLX-NEXT: retq
17771 ;
17772 ; NoVLX-LABEL: test_vpcmpultq_v2i1_v32i1_mask:
17773 ; NoVLX: # %bb.0: # %entry
17774 ; NoVLX-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
17775 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
17776 ; NoVLX-NEXT: vpcmpltuq %zmm1, %zmm0, %k0
17777 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
17778 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
17779 ; NoVLX-NEXT: kmovw %k0, %eax
17780 ; NoVLX-NEXT: vzeroupper
17781 ; NoVLX-NEXT: retq
17782 entry:
17783 %0 = bitcast <2 x i64> %__a to <2 x i64>
17784 %1 = bitcast <2 x i64> %__b to <2 x i64>
17785 %2 = icmp ult <2 x i64> %0, %1
17786 %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
17787 %4 = bitcast <32 x i1> %3 to i32
17788 ret i32 %4
17789 }
17791 define zeroext i32 @test_vpcmpultq_v2i1_v32i1_mask_mem(<2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
17792 ; VLX-LABEL: test_vpcmpultq_v2i1_v32i1_mask_mem:
17793 ; VLX: # %bb.0: # %entry
17794 ; VLX-NEXT: vpcmpltuq (%rdi), %xmm0, %k0
17795 ; VLX-NEXT: kmovd %k0, %eax
17796 ; VLX-NEXT: retq
17797 ;
17798 ; NoVLX-LABEL: test_vpcmpultq_v2i1_v32i1_mask_mem:
17799 ; NoVLX: # %bb.0: # %entry
17800 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
17801 ; NoVLX-NEXT: vmovdqa (%rdi), %xmm1
17802 ; NoVLX-NEXT: vpcmpltuq %zmm1, %zmm0, %k0
17803 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
17804 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
17805 ; NoVLX-NEXT: kmovw %k0, %eax
17806 ; NoVLX-NEXT: vzeroupper
17807 ; NoVLX-NEXT: retq
17808 entry:
17809 %0 = bitcast <2 x i64> %__a to <2 x i64>
17810 %load = load <2 x i64>, <2 x i64>* %__b
17811 %1 = bitcast <2 x i64> %load to <2 x i64>
17812 %2 = icmp ult <2 x i64> %0, %1
17813 %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
17814 %4 = bitcast <32 x i1> %3 to i32
17815 ret i32 %4
17816 }
17818 define zeroext i32 @test_masked_vpcmpultq_v2i1_v32i1_mask(i8 zeroext %__u, <2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
17819 ; VLX-LABEL: test_masked_vpcmpultq_v2i1_v32i1_mask:
17820 ; VLX: # %bb.0: # %entry
17821 ; VLX-NEXT: kmovd %edi, %k1
17822 ; VLX-NEXT: vpcmpltuq %xmm1, %xmm0, %k0 {%k1}
17823 ; VLX-NEXT: kmovd %k0, %eax
17824 ; VLX-NEXT: retq
17825 ;
17826 ; NoVLX-LABEL: test_masked_vpcmpultq_v2i1_v32i1_mask:
17827 ; NoVLX: # %bb.0: # %entry
17828 ; NoVLX-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
17829 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
17830 ; NoVLX-NEXT: kmovw %edi, %k1
17831 ; NoVLX-NEXT: vpcmpltuq %zmm1, %zmm0, %k0 {%k1}
17832 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
17833 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
17834 ; NoVLX-NEXT: kmovw %k0, %eax
17835 ; NoVLX-NEXT: vzeroupper
17836 ; NoVLX-NEXT: retq
17837 entry:
17838 %0 = bitcast <2 x i64> %__a to <2 x i64>
17839 %1 = bitcast <2 x i64> %__b to <2 x i64>
17840 %2 = icmp ult <2 x i64> %0, %1
17841 %3 = bitcast i8 %__u to <8 x i1>
17842 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <2 x i32> <i32 0, i32 1>
17843 %4 = and <2 x i1> %2, %extract.i
17844 %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
17845 %6 = bitcast <32 x i1> %5 to i32
17846 ret i32 %6
17847 }
17849 define zeroext i32 @test_masked_vpcmpultq_v2i1_v32i1_mask_mem(i8 zeroext %__u, <2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
17850 ; VLX-LABEL: test_masked_vpcmpultq_v2i1_v32i1_mask_mem:
17851 ; VLX: # %bb.0: # %entry
17852 ; VLX-NEXT: kmovd %edi, %k1
17853 ; VLX-NEXT: vpcmpltuq (%rsi), %xmm0, %k0 {%k1}
17854 ; VLX-NEXT: kmovd %k0, %eax
17855 ; VLX-NEXT: retq
17856 ;
17857 ; NoVLX-LABEL: test_masked_vpcmpultq_v2i1_v32i1_mask_mem:
17858 ; NoVLX: # %bb.0: # %entry
17859 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
17860 ; NoVLX-NEXT: vmovdqa (%rsi), %xmm1
17861 ; NoVLX-NEXT: kmovw %edi, %k1
17862 ; NoVLX-NEXT: vpcmpltuq %zmm1, %zmm0, %k0 {%k1}
17863 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
17864 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
17865 ; NoVLX-NEXT: kmovw %k0, %eax
17866 ; NoVLX-NEXT: vzeroupper
17867 ; NoVLX-NEXT: retq
17868 entry:
17869 %0 = bitcast <2 x i64> %__a to <2 x i64>
17870 %load = load <2 x i64>, <2 x i64>* %__b
17871 %1 = bitcast <2 x i64> %load to <2 x i64>
17872 %2 = icmp ult <2 x i64> %0, %1
17873 %3 = bitcast i8 %__u to <8 x i1>
17874 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <2 x i32> <i32 0, i32 1>
17875 %4 = and <2 x i1> %2, %extract.i
17876 %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
17877 %6 = bitcast <32 x i1> %5 to i32
17878 ret i32 %6
17879 }
17882 define zeroext i32 @test_vpcmpultq_v2i1_v32i1_mask_mem_b(<2 x i64> %__a, i64* %__b) local_unnamed_addr {
17883 ; VLX-LABEL: test_vpcmpultq_v2i1_v32i1_mask_mem_b:
17884 ; VLX: # %bb.0: # %entry
17885 ; VLX-NEXT: vpcmpltuq (%rdi){1to2}, %xmm0, %k0
17886 ; VLX-NEXT: kmovd %k0, %eax
17887 ; VLX-NEXT: retq
17888 ;
17889 ; NoVLX-LABEL: test_vpcmpultq_v2i1_v32i1_mask_mem_b:
17890 ; NoVLX: # %bb.0: # %entry
17891 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
17892 ; NoVLX-NEXT: vpcmpltuq (%rdi){1to8}, %zmm0, %k0
17893 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
17894 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
17895 ; NoVLX-NEXT: kmovw %k0, %eax
17896 ; NoVLX-NEXT: vzeroupper
17897 ; NoVLX-NEXT: retq
17898 entry:
17899 %0 = bitcast <2 x i64> %__a to <2 x i64>
17900 %load = load i64, i64* %__b
17901 %vec = insertelement <2 x i64> undef, i64 %load, i32 0
17902 %1 = shufflevector <2 x i64> %vec, <2 x i64> undef, <2 x i32> <i32 0, i32 0>
17903 %2 = icmp ult <2 x i64> %0, %1
17904 %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
17905 %4 = bitcast <32 x i1> %3 to i32
17906 ret i32 %4
17907 }
17909 define zeroext i32 @test_masked_vpcmpultq_v2i1_v32i1_mask_mem_b(i8 zeroext %__u, <2 x i64> %__a, i64* %__b) local_unnamed_addr {
17910 ; VLX-LABEL: test_masked_vpcmpultq_v2i1_v32i1_mask_mem_b:
17911 ; VLX: # %bb.0: # %entry
17912 ; VLX-NEXT: kmovd %edi, %k1
17913 ; VLX-NEXT: vpcmpltuq (%rsi){1to2}, %xmm0, %k0 {%k1}
17914 ; VLX-NEXT: kmovd %k0, %eax
17915 ; VLX-NEXT: retq
17916 ;
17917 ; NoVLX-LABEL: test_masked_vpcmpultq_v2i1_v32i1_mask_mem_b:
17918 ; NoVLX: # %bb.0: # %entry
17919 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
17920 ; NoVLX-NEXT: kmovw %edi, %k1
17921 ; NoVLX-NEXT: vpcmpltuq (%rsi){1to8}, %zmm0, %k0 {%k1}
17922 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
17923 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
17924 ; NoVLX-NEXT: kmovw %k0, %eax
17925 ; NoVLX-NEXT: vzeroupper
17926 ; NoVLX-NEXT: retq
17927 entry:
17928 %0 = bitcast <2 x i64> %__a to <2 x i64>
17929 %load = load i64, i64* %__b
17930 %vec = insertelement <2 x i64> undef, i64 %load, i32 0
17931 %1 = shufflevector <2 x i64> %vec, <2 x i64> undef, <2 x i32> <i32 0, i32 0>
17932 %2 = icmp ult <2 x i64> %0, %1
17933 %3 = bitcast i8 %__u to <8 x i1>
17934 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <2 x i32> <i32 0, i32 1>
17935 %4 = and <2 x i1> %extract.i, %2
17936 %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
17937 %6 = bitcast <32 x i1> %5 to i32
17938 ret i32 %6
17939 }
17942 define zeroext i64 @test_vpcmpultq_v2i1_v64i1_mask(<2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
17943 ; VLX-LABEL: test_vpcmpultq_v2i1_v64i1_mask:
17944 ; VLX: # %bb.0: # %entry
17945 ; VLX-NEXT: vpcmpltuq %xmm1, %xmm0, %k0
17946 ; VLX-NEXT: kmovq %k0, %rax
17947 ; VLX-NEXT: retq
17948 ;
17949 ; NoVLX-LABEL: test_vpcmpultq_v2i1_v64i1_mask:
17950 ; NoVLX: # %bb.0: # %entry
17951 ; NoVLX-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
17952 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
17953 ; NoVLX-NEXT: vpcmpltuq %zmm1, %zmm0, %k0
17954 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
17955 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
17956 ; NoVLX-NEXT: kmovw %k0, %eax
17957 ; NoVLX-NEXT: vzeroupper
17958 ; NoVLX-NEXT: retq
17959 entry:
17960 %0 = bitcast <2 x i64> %__a to <2 x i64>
17961 %1 = bitcast <2 x i64> %__b to <2 x i64>
17962 %2 = icmp ult <2 x i64> %0, %1
17963 %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
17964 %4 = bitcast <64 x i1> %3 to i64
17965 ret i64 %4
17966 }
17968 define zeroext i64 @test_vpcmpultq_v2i1_v64i1_mask_mem(<2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
17969 ; VLX-LABEL: test_vpcmpultq_v2i1_v64i1_mask_mem:
17970 ; VLX: # %bb.0: # %entry
17971 ; VLX-NEXT: vpcmpltuq (%rdi), %xmm0, %k0
17972 ; VLX-NEXT: kmovq %k0, %rax
17973 ; VLX-NEXT: retq
17974 ;
17975 ; NoVLX-LABEL: test_vpcmpultq_v2i1_v64i1_mask_mem:
17976 ; NoVLX: # %bb.0: # %entry
17977 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
17978 ; NoVLX-NEXT: vmovdqa (%rdi), %xmm1
17979 ; NoVLX-NEXT: vpcmpltuq %zmm1, %zmm0, %k0
17980 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
17981 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
17982 ; NoVLX-NEXT: kmovw %k0, %eax
17983 ; NoVLX-NEXT: vzeroupper
17984 ; NoVLX-NEXT: retq
17985 entry:
17986 %0 = bitcast <2 x i64> %__a to <2 x i64>
17987 %load = load <2 x i64>, <2 x i64>* %__b
17988 %1 = bitcast <2 x i64> %load to <2 x i64>
17989 %2 = icmp ult <2 x i64> %0, %1
17990 %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
17991 %4 = bitcast <64 x i1> %3 to i64
17992 ret i64 %4
17993 }
17995 define zeroext i64 @test_masked_vpcmpultq_v2i1_v64i1_mask(i8 zeroext %__u, <2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
17996 ; VLX-LABEL: test_masked_vpcmpultq_v2i1_v64i1_mask:
17997 ; VLX: # %bb.0: # %entry
17998 ; VLX-NEXT: kmovd %edi, %k1
17999 ; VLX-NEXT: vpcmpltuq %xmm1, %xmm0, %k0 {%k1}
18000 ; VLX-NEXT: kmovq %k0, %rax
18001 ; VLX-NEXT: retq
18002 ;
18003 ; NoVLX-LABEL: test_masked_vpcmpultq_v2i1_v64i1_mask:
18004 ; NoVLX: # %bb.0: # %entry
18005 ; NoVLX-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
18006 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
18007 ; NoVLX-NEXT: kmovw %edi, %k1
18008 ; NoVLX-NEXT: vpcmpltuq %zmm1, %zmm0, %k0 {%k1}
18009 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
18010 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
18011 ; NoVLX-NEXT: kmovw %k0, %eax
18012 ; NoVLX-NEXT: vzeroupper
18013 ; NoVLX-NEXT: retq
18014 entry:
18015 %0 = bitcast <2 x i64> %__a to <2 x i64>
18016 %1 = bitcast <2 x i64> %__b to <2 x i64>
18017 %2 = icmp ult <2 x i64> %0, %1
18018 %3 = bitcast i8 %__u to <8 x i1>
18019 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <2 x i32> <i32 0, i32 1>
18020 %4 = and <2 x i1> %2, %extract.i
18021 %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
18022 %6 = bitcast <64 x i1> %5 to i64
18023 ret i64 %6
18024 }
18026 define zeroext i64 @test_masked_vpcmpultq_v2i1_v64i1_mask_mem(i8 zeroext %__u, <2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
18027 ; VLX-LABEL: test_masked_vpcmpultq_v2i1_v64i1_mask_mem:
18028 ; VLX: # %bb.0: # %entry
18029 ; VLX-NEXT: kmovd %edi, %k1
18030 ; VLX-NEXT: vpcmpltuq (%rsi), %xmm0, %k0 {%k1}
18031 ; VLX-NEXT: kmovq %k0, %rax
18032 ; VLX-NEXT: retq
18033 ;
18034 ; NoVLX-LABEL: test_masked_vpcmpultq_v2i1_v64i1_mask_mem:
18035 ; NoVLX: # %bb.0: # %entry
18036 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
18037 ; NoVLX-NEXT: vmovdqa (%rsi), %xmm1
18038 ; NoVLX-NEXT: kmovw %edi, %k1
18039 ; NoVLX-NEXT: vpcmpltuq %zmm1, %zmm0, %k0 {%k1}
18040 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
18041 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
18042 ; NoVLX-NEXT: kmovw %k0, %eax
18043 ; NoVLX-NEXT: vzeroupper
18044 ; NoVLX-NEXT: retq
18045 entry:
18046 %0 = bitcast <2 x i64> %__a to <2 x i64>
18047 %load = load <2 x i64>, <2 x i64>* %__b
18048 %1 = bitcast <2 x i64> %load to <2 x i64>
18049 %2 = icmp ult <2 x i64> %0, %1
18050 %3 = bitcast i8 %__u to <8 x i1>
18051 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <2 x i32> <i32 0, i32 1>
18052 %4 = and <2 x i1> %2, %extract.i
18053 %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
18054 %6 = bitcast <64 x i1> %5 to i64
18055 ret i64 %6
18056 }
18059 define zeroext i64 @test_vpcmpultq_v2i1_v64i1_mask_mem_b(<2 x i64> %__a, i64* %__b) local_unnamed_addr {
18060 ; VLX-LABEL: test_vpcmpultq_v2i1_v64i1_mask_mem_b:
18061 ; VLX: # %bb.0: # %entry
18062 ; VLX-NEXT: vpcmpltuq (%rdi){1to2}, %xmm0, %k0
18063 ; VLX-NEXT: kmovq %k0, %rax
18064 ; VLX-NEXT: retq
18065 ;
18066 ; NoVLX-LABEL: test_vpcmpultq_v2i1_v64i1_mask_mem_b:
18067 ; NoVLX: # %bb.0: # %entry
18068 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
18069 ; NoVLX-NEXT: vpcmpltuq (%rdi){1to8}, %zmm0, %k0
18070 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
18071 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
18072 ; NoVLX-NEXT: kmovw %k0, %eax
18073 ; NoVLX-NEXT: vzeroupper
18074 ; NoVLX-NEXT: retq
18075 entry:
18076 %0 = bitcast <2 x i64> %__a to <2 x i64>
18077 %load = load i64, i64* %__b
18078 %vec = insertelement <2 x i64> undef, i64 %load, i32 0
18079 %1 = shufflevector <2 x i64> %vec, <2 x i64> undef, <2 x i32> <i32 0, i32 0>
18080 %2 = icmp ult <2 x i64> %0, %1
18081 %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
18082 %4 = bitcast <64 x i1> %3 to i64
18083 ret i64 %4
18084 }
18086 define zeroext i64 @test_masked_vpcmpultq_v2i1_v64i1_mask_mem_b(i8 zeroext %__u, <2 x i64> %__a, i64* %__b) local_unnamed_addr {
18087 ; VLX-LABEL: test_masked_vpcmpultq_v2i1_v64i1_mask_mem_b:
18088 ; VLX: # %bb.0: # %entry
18089 ; VLX-NEXT: kmovd %edi, %k1
18090 ; VLX-NEXT: vpcmpltuq (%rsi){1to2}, %xmm0, %k0 {%k1}
18091 ; VLX-NEXT: kmovq %k0, %rax
18092 ; VLX-NEXT: retq
18093 ;
18094 ; NoVLX-LABEL: test_masked_vpcmpultq_v2i1_v64i1_mask_mem_b:
18095 ; NoVLX: # %bb.0: # %entry
18096 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
18097 ; NoVLX-NEXT: kmovw %edi, %k1
18098 ; NoVLX-NEXT: vpcmpltuq (%rsi){1to8}, %zmm0, %k0 {%k1}
18099 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
18100 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
18101 ; NoVLX-NEXT: kmovw %k0, %eax
18102 ; NoVLX-NEXT: vzeroupper
18103 ; NoVLX-NEXT: retq
18104 entry:
18105 %0 = bitcast <2 x i64> %__a to <2 x i64>
18106 %load = load i64, i64* %__b
18107 %vec = insertelement <2 x i64> undef, i64 %load, i32 0
18108 %1 = shufflevector <2 x i64> %vec, <2 x i64> undef, <2 x i32> <i32 0, i32 0>
18109 %2 = icmp ult <2 x i64> %0, %1
18110 %3 = bitcast i8 %__u to <8 x i1>
18111 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <2 x i32> <i32 0, i32 1>
18112 %4 = and <2 x i1> %extract.i, %2
18113 %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
18114 %6 = bitcast <64 x i1> %5 to i64
18115 ret i64 %6
18116 }
18119 define zeroext i8 @test_vpcmpultq_v4i1_v8i1_mask(<4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
18120 ; VLX-LABEL: test_vpcmpultq_v4i1_v8i1_mask:
18121 ; VLX: # %bb.0: # %entry
18122 ; VLX-NEXT: vpcmpltuq %ymm1, %ymm0, %k0
18123 ; VLX-NEXT: kmovd %k0, %eax
18124 ; VLX-NEXT: # kill: def $al killed $al killed $eax
18125 ; VLX-NEXT: vzeroupper
18126 ; VLX-NEXT: retq
18127 ;
18128 ; NoVLX-LABEL: test_vpcmpultq_v4i1_v8i1_mask:
18129 ; NoVLX: # %bb.0: # %entry
18130 ; NoVLX-NEXT: # kill: def $ymm1 killed $ymm1 def $zmm1
18131 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
18132 ; NoVLX-NEXT: vpcmpltuq %zmm1, %zmm0, %k0
18133 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
18134 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
18135 ; NoVLX-NEXT: kmovw %k0, %eax
18136 ; NoVLX-NEXT: # kill: def $al killed $al killed $eax
18137 ; NoVLX-NEXT: vzeroupper
18138 ; NoVLX-NEXT: retq
18139 entry:
18140 %0 = bitcast <4 x i64> %__a to <4 x i64>
18141 %1 = bitcast <4 x i64> %__b to <4 x i64>
18142 %2 = icmp ult <4 x i64> %0, %1
18143 %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
18144 %4 = bitcast <8 x i1> %3 to i8
18145 ret i8 %4
18146 }
18148 define zeroext i8 @test_vpcmpultq_v4i1_v8i1_mask_mem(<4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
18149 ; VLX-LABEL: test_vpcmpultq_v4i1_v8i1_mask_mem:
18150 ; VLX: # %bb.0: # %entry
18151 ; VLX-NEXT: vpcmpltuq (%rdi), %ymm0, %k0
18152 ; VLX-NEXT: kmovd %k0, %eax
18153 ; VLX-NEXT: # kill: def $al killed $al killed $eax
18154 ; VLX-NEXT: vzeroupper
18155 ; VLX-NEXT: retq
18156 ;
18157 ; NoVLX-LABEL: test_vpcmpultq_v4i1_v8i1_mask_mem:
18158 ; NoVLX: # %bb.0: # %entry
18159 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
18160 ; NoVLX-NEXT: vmovdqa (%rdi), %ymm1
18161 ; NoVLX-NEXT: vpcmpltuq %zmm1, %zmm0, %k0
18162 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
18163 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
18164 ; NoVLX-NEXT: kmovw %k0, %eax
18165 ; NoVLX-NEXT: # kill: def $al killed $al killed $eax
18166 ; NoVLX-NEXT: vzeroupper
18167 ; NoVLX-NEXT: retq
18168 entry:
18169 %0 = bitcast <4 x i64> %__a to <4 x i64>
18170 %load = load <4 x i64>, <4 x i64>* %__b
18171 %1 = bitcast <4 x i64> %load to <4 x i64>
18172 %2 = icmp ult <4 x i64> %0, %1
18173 %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
18174 %4 = bitcast <8 x i1> %3 to i8
18175 ret i8 %4
18176 }
18178 define zeroext i8 @test_masked_vpcmpultq_v4i1_v8i1_mask(i8 zeroext %__u, <4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
18179 ; VLX-LABEL: test_masked_vpcmpultq_v4i1_v8i1_mask:
18180 ; VLX: # %bb.0: # %entry
18181 ; VLX-NEXT: kmovd %edi, %k1
18182 ; VLX-NEXT: vpcmpltuq %ymm1, %ymm0, %k0 {%k1}
18183 ; VLX-NEXT: kmovd %k0, %eax
18184 ; VLX-NEXT: # kill: def $al killed $al killed $eax
18185 ; VLX-NEXT: vzeroupper
18186 ; VLX-NEXT: retq
18187 ;
18188 ; NoVLX-LABEL: test_masked_vpcmpultq_v4i1_v8i1_mask:
18189 ; NoVLX: # %bb.0: # %entry
18190 ; NoVLX-NEXT: # kill: def $ymm1 killed $ymm1 def $zmm1
18191 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
18192 ; NoVLX-NEXT: kmovw %edi, %k1
18193 ; NoVLX-NEXT: vpcmpltuq %zmm1, %zmm0, %k0 {%k1}
18194 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
18195 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
18196 ; NoVLX-NEXT: kmovw %k0, %eax
18197 ; NoVLX-NEXT: # kill: def $al killed $al killed $eax
18198 ; NoVLX-NEXT: vzeroupper
18199 ; NoVLX-NEXT: retq
18200 entry:
18201 %0 = bitcast <4 x i64> %__a to <4 x i64>
18202 %1 = bitcast <4 x i64> %__b to <4 x i64>
18203 %2 = icmp ult <4 x i64> %0, %1
18204 %3 = bitcast i8 %__u to <8 x i1>
18205 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
18206 %4 = and <4 x i1> %2, %extract.i
18207 %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
18208 %6 = bitcast <8 x i1> %5 to i8
18209 ret i8 %6
18210 }
18212 define zeroext i8 @test_masked_vpcmpultq_v4i1_v8i1_mask_mem(i8 zeroext %__u, <4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
18213 ; VLX-LABEL: test_masked_vpcmpultq_v4i1_v8i1_mask_mem:
18214 ; VLX: # %bb.0: # %entry
18215 ; VLX-NEXT: kmovd %edi, %k1
18216 ; VLX-NEXT: vpcmpltuq (%rsi), %ymm0, %k0 {%k1}
18217 ; VLX-NEXT: kmovd %k0, %eax
18218 ; VLX-NEXT: # kill: def $al killed $al killed $eax
18219 ; VLX-NEXT: vzeroupper
18220 ; VLX-NEXT: retq
18221 ;
18222 ; NoVLX-LABEL: test_masked_vpcmpultq_v4i1_v8i1_mask_mem:
18223 ; NoVLX: # %bb.0: # %entry
18224 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
18225 ; NoVLX-NEXT: vmovdqa (%rsi), %ymm1
18226 ; NoVLX-NEXT: kmovw %edi, %k1
18227 ; NoVLX-NEXT: vpcmpltuq %zmm1, %zmm0, %k0 {%k1}
18228 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
18229 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
18230 ; NoVLX-NEXT: kmovw %k0, %eax
18231 ; NoVLX-NEXT: # kill: def $al killed $al killed $eax
18232 ; NoVLX-NEXT: vzeroupper
18233 ; NoVLX-NEXT: retq
18234 entry:
18235 %0 = bitcast <4 x i64> %__a to <4 x i64>
18236 %load = load <4 x i64>, <4 x i64>* %__b
18237 %1 = bitcast <4 x i64> %load to <4 x i64>
18238 %2 = icmp ult <4 x i64> %0, %1
18239 %3 = bitcast i8 %__u to <8 x i1>
18240 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
18241 %4 = and <4 x i1> %2, %extract.i
18242 %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
18243 %6 = bitcast <8 x i1> %5 to i8
18244 ret i8 %6
18245 }
18248 define zeroext i8 @test_vpcmpultq_v4i1_v8i1_mask_mem_b(<4 x i64> %__a, i64* %__b) local_unnamed_addr {
18249 ; VLX-LABEL: test_vpcmpultq_v4i1_v8i1_mask_mem_b:
18250 ; VLX: # %bb.0: # %entry
18251 ; VLX-NEXT: vpcmpltuq (%rdi){1to4}, %ymm0, %k0
18252 ; VLX-NEXT: kmovd %k0, %eax
18253 ; VLX-NEXT: # kill: def $al killed $al killed $eax
18254 ; VLX-NEXT: vzeroupper
18255 ; VLX-NEXT: retq
18256 ;
18257 ; NoVLX-LABEL: test_vpcmpultq_v4i1_v8i1_mask_mem_b:
18258 ; NoVLX: # %bb.0: # %entry
18259 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
18260 ; NoVLX-NEXT: vpcmpltuq (%rdi){1to8}, %zmm0, %k0
18261 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
18262 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
18263 ; NoVLX-NEXT: kmovw %k0, %eax
18264 ; NoVLX-NEXT: # kill: def $al killed $al killed $eax
18265 ; NoVLX-NEXT: vzeroupper
18266 ; NoVLX-NEXT: retq
18267 entry:
18268 %0 = bitcast <4 x i64> %__a to <4 x i64>
18269 %load = load i64, i64* %__b
18270 %vec = insertelement <4 x i64> undef, i64 %load, i32 0
18271 %1 = shufflevector <4 x i64> %vec, <4 x i64> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
18272 %2 = icmp ult <4 x i64> %0, %1
18273 %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
18274 %4 = bitcast <8 x i1> %3 to i8
18275 ret i8 %4
18276 }
18278 define zeroext i8 @test_masked_vpcmpultq_v4i1_v8i1_mask_mem_b(i8 zeroext %__u, <4 x i64> %__a, i64* %__b) local_unnamed_addr {
18279 ; VLX-LABEL: test_masked_vpcmpultq_v4i1_v8i1_mask_mem_b:
18280 ; VLX: # %bb.0: # %entry
18281 ; VLX-NEXT: kmovd %edi, %k1
18282 ; VLX-NEXT: vpcmpltuq (%rsi){1to4}, %ymm0, %k0 {%k1}
18283 ; VLX-NEXT: kmovd %k0, %eax
18284 ; VLX-NEXT: # kill: def $al killed $al killed $eax
18285 ; VLX-NEXT: vzeroupper
18286 ; VLX-NEXT: retq
18287 ;
18288 ; NoVLX-LABEL: test_masked_vpcmpultq_v4i1_v8i1_mask_mem_b:
18289 ; NoVLX: # %bb.0: # %entry
18290 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
18291 ; NoVLX-NEXT: kmovw %edi, %k1
18292 ; NoVLX-NEXT: vpcmpltuq (%rsi){1to8}, %zmm0, %k0 {%k1}
18293 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
18294 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
18295 ; NoVLX-NEXT: kmovw %k0, %eax
18296 ; NoVLX-NEXT: # kill: def $al killed $al killed $eax
18297 ; NoVLX-NEXT: vzeroupper
18298 ; NoVLX-NEXT: retq
18299 entry:
18300 %0 = bitcast <4 x i64> %__a to <4 x i64>
18301 %load = load i64, i64* %__b
18302 %vec = insertelement <4 x i64> undef, i64 %load, i32 0
18303 %1 = shufflevector <4 x i64> %vec, <4 x i64> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
18304 %2 = icmp ult <4 x i64> %0, %1
18305 %3 = bitcast i8 %__u to <8 x i1>
18306 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
18307 %4 = and <4 x i1> %extract.i, %2
18308 %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
18309 %6 = bitcast <8 x i1> %5 to i8
18310 ret i8 %6
18311 }
18314 define zeroext i16 @test_vpcmpultq_v4i1_v16i1_mask(<4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
18315 ; VLX-LABEL: test_vpcmpultq_v4i1_v16i1_mask:
18316 ; VLX: # %bb.0: # %entry
18317 ; VLX-NEXT: vpcmpltuq %ymm1, %ymm0, %k0
18318 ; VLX-NEXT: kmovd %k0, %eax
18319 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
18320 ; VLX-NEXT: vzeroupper
18321 ; VLX-NEXT: retq
18322 ;
18323 ; NoVLX-LABEL: test_vpcmpultq_v4i1_v16i1_mask:
18324 ; NoVLX: # %bb.0: # %entry
18325 ; NoVLX-NEXT: # kill: def $ymm1 killed $ymm1 def $zmm1
18326 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
18327 ; NoVLX-NEXT: vpcmpltuq %zmm1, %zmm0, %k0
18328 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
18329 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
18330 ; NoVLX-NEXT: kmovw %k0, %eax
18331 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
18332 ; NoVLX-NEXT: vzeroupper
18333 ; NoVLX-NEXT: retq
18334 entry:
18335 %0 = bitcast <4 x i64> %__a to <4 x i64>
18336 %1 = bitcast <4 x i64> %__b to <4 x i64>
18337 %2 = icmp ult <4 x i64> %0, %1
18338 %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
18339 %4 = bitcast <16 x i1> %3 to i16
18340 ret i16 %4
18341 }
18343 define zeroext i16 @test_vpcmpultq_v4i1_v16i1_mask_mem(<4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
18344 ; VLX-LABEL: test_vpcmpultq_v4i1_v16i1_mask_mem:
18345 ; VLX: # %bb.0: # %entry
18346 ; VLX-NEXT: vpcmpltuq (%rdi), %ymm0, %k0
18347 ; VLX-NEXT: kmovd %k0, %eax
18348 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
18349 ; VLX-NEXT: vzeroupper
18350 ; VLX-NEXT: retq
18351 ;
18352 ; NoVLX-LABEL: test_vpcmpultq_v4i1_v16i1_mask_mem:
18353 ; NoVLX: # %bb.0: # %entry
18354 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
18355 ; NoVLX-NEXT: vmovdqa (%rdi), %ymm1
18356 ; NoVLX-NEXT: vpcmpltuq %zmm1, %zmm0, %k0
18357 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
18358 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
18359 ; NoVLX-NEXT: kmovw %k0, %eax
18360 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
18361 ; NoVLX-NEXT: vzeroupper
18362 ; NoVLX-NEXT: retq
18363 entry:
18364 %0 = bitcast <4 x i64> %__a to <4 x i64>
18365 %load = load <4 x i64>, <4 x i64>* %__b
18366 %1 = bitcast <4 x i64> %load to <4 x i64>
18367 %2 = icmp ult <4 x i64> %0, %1
18368 %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
18369 %4 = bitcast <16 x i1> %3 to i16
18370 ret i16 %4
18371 }
18373 define zeroext i16 @test_masked_vpcmpultq_v4i1_v16i1_mask(i8 zeroext %__u, <4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
18374 ; VLX-LABEL: test_masked_vpcmpultq_v4i1_v16i1_mask:
18375 ; VLX: # %bb.0: # %entry
18376 ; VLX-NEXT: kmovd %edi, %k1
18377 ; VLX-NEXT: vpcmpltuq %ymm1, %ymm0, %k0 {%k1}
18378 ; VLX-NEXT: kmovd %k0, %eax
18379 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
18380 ; VLX-NEXT: vzeroupper
18381 ; VLX-NEXT: retq
18382 ;
18383 ; NoVLX-LABEL: test_masked_vpcmpultq_v4i1_v16i1_mask:
18384 ; NoVLX: # %bb.0: # %entry
18385 ; NoVLX-NEXT: # kill: def $ymm1 killed $ymm1 def $zmm1
18386 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
18387 ; NoVLX-NEXT: kmovw %edi, %k1
18388 ; NoVLX-NEXT: vpcmpltuq %zmm1, %zmm0, %k0 {%k1}
18389 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
18390 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
18391 ; NoVLX-NEXT: kmovw %k0, %eax
18392 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
18393 ; NoVLX-NEXT: vzeroupper
18394 ; NoVLX-NEXT: retq
18395 entry:
18396 %0 = bitcast <4 x i64> %__a to <4 x i64>
18397 %1 = bitcast <4 x i64> %__b to <4 x i64>
18398 %2 = icmp ult <4 x i64> %0, %1
18399 %3 = bitcast i8 %__u to <8 x i1>
18400 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
18401 %4 = and <4 x i1> %2, %extract.i
18402 %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
18403 %6 = bitcast <16 x i1> %5 to i16
18404 ret i16 %6
18405 }
18407 define zeroext i16 @test_masked_vpcmpultq_v4i1_v16i1_mask_mem(i8 zeroext %__u, <4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
18408 ; VLX-LABEL: test_masked_vpcmpultq_v4i1_v16i1_mask_mem:
18409 ; VLX: # %bb.0: # %entry
18410 ; VLX-NEXT: kmovd %edi, %k1
18411 ; VLX-NEXT: vpcmpltuq (%rsi), %ymm0, %k0 {%k1}
18412 ; VLX-NEXT: kmovd %k0, %eax
18413 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
18414 ; VLX-NEXT: vzeroupper
18415 ; VLX-NEXT: retq
18416 ;
18417 ; NoVLX-LABEL: test_masked_vpcmpultq_v4i1_v16i1_mask_mem:
18418 ; NoVLX: # %bb.0: # %entry
18419 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
18420 ; NoVLX-NEXT: vmovdqa (%rsi), %ymm1
18421 ; NoVLX-NEXT: kmovw %edi, %k1
18422 ; NoVLX-NEXT: vpcmpltuq %zmm1, %zmm0, %k0 {%k1}
18423 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
18424 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
18425 ; NoVLX-NEXT: kmovw %k0, %eax
18426 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
18427 ; NoVLX-NEXT: vzeroupper
18428 ; NoVLX-NEXT: retq
18429 entry:
18430 %0 = bitcast <4 x i64> %__a to <4 x i64>
18431 %load = load <4 x i64>, <4 x i64>* %__b
18432 %1 = bitcast <4 x i64> %load to <4 x i64>
18433 %2 = icmp ult <4 x i64> %0, %1
18434 %3 = bitcast i8 %__u to <8 x i1>
18435 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
18436 %4 = and <4 x i1> %2, %extract.i
18437 %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
18438 %6 = bitcast <16 x i1> %5 to i16
18439 ret i16 %6
18440 }
18443 define zeroext i16 @test_vpcmpultq_v4i1_v16i1_mask_mem_b(<4 x i64> %__a, i64* %__b) local_unnamed_addr {
18444 ; VLX-LABEL: test_vpcmpultq_v4i1_v16i1_mask_mem_b:
18445 ; VLX: # %bb.0: # %entry
18446 ; VLX-NEXT: vpcmpltuq (%rdi){1to4}, %ymm0, %k0
18447 ; VLX-NEXT: kmovd %k0, %eax
18448 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
18449 ; VLX-NEXT: vzeroupper
18450 ; VLX-NEXT: retq
18451 ;
18452 ; NoVLX-LABEL: test_vpcmpultq_v4i1_v16i1_mask_mem_b:
18453 ; NoVLX: # %bb.0: # %entry
18454 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
18455 ; NoVLX-NEXT: vpcmpltuq (%rdi){1to8}, %zmm0, %k0
18456 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
18457 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
18458 ; NoVLX-NEXT: kmovw %k0, %eax
18459 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
18460 ; NoVLX-NEXT: vzeroupper
18461 ; NoVLX-NEXT: retq
18462 entry:
18463 %0 = bitcast <4 x i64> %__a to <4 x i64>
18464 %load = load i64, i64* %__b
18465 %vec = insertelement <4 x i64> undef, i64 %load, i32 0
18466 %1 = shufflevector <4 x i64> %vec, <4 x i64> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
18467 %2 = icmp ult <4 x i64> %0, %1
18468 %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
18469 %4 = bitcast <16 x i1> %3 to i16
18470 ret i16 %4
18471 }
18473 define zeroext i16 @test_masked_vpcmpultq_v4i1_v16i1_mask_mem_b(i8 zeroext %__u, <4 x i64> %__a, i64* %__b) local_unnamed_addr {
18474 ; VLX-LABEL: test_masked_vpcmpultq_v4i1_v16i1_mask_mem_b:
18475 ; VLX: # %bb.0: # %entry
18476 ; VLX-NEXT: kmovd %edi, %k1
18477 ; VLX-NEXT: vpcmpltuq (%rsi){1to4}, %ymm0, %k0 {%k1}
18478 ; VLX-NEXT: kmovd %k0, %eax
18479 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
18480 ; VLX-NEXT: vzeroupper
18481 ; VLX-NEXT: retq
18482 ;
18483 ; NoVLX-LABEL: test_masked_vpcmpultq_v4i1_v16i1_mask_mem_b:
18484 ; NoVLX: # %bb.0: # %entry
18485 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
18486 ; NoVLX-NEXT: kmovw %edi, %k1
18487 ; NoVLX-NEXT: vpcmpltuq (%rsi){1to8}, %zmm0, %k0 {%k1}
18488 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
18489 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
18490 ; NoVLX-NEXT: kmovw %k0, %eax
18491 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
18492 ; NoVLX-NEXT: vzeroupper
18493 ; NoVLX-NEXT: retq
18494 entry:
18495 %0 = bitcast <4 x i64> %__a to <4 x i64>
18496 %load = load i64, i64* %__b
18497 %vec = insertelement <4 x i64> undef, i64 %load, i32 0
18498 %1 = shufflevector <4 x i64> %vec, <4 x i64> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
18499 %2 = icmp ult <4 x i64> %0, %1
18500 %3 = bitcast i8 %__u to <8 x i1>
18501 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
18502 %4 = and <4 x i1> %extract.i, %2
18503 %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
18504 %6 = bitcast <16 x i1> %5 to i16
18505 ret i16 %6
18506 }
18509 define zeroext i32 @test_vpcmpultq_v4i1_v32i1_mask(<4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
18510 ; VLX-LABEL: test_vpcmpultq_v4i1_v32i1_mask:
18511 ; VLX: # %bb.0: # %entry
18512 ; VLX-NEXT: vpcmpltuq %ymm1, %ymm0, %k0
18513 ; VLX-NEXT: kmovd %k0, %eax
18514 ; VLX-NEXT: vzeroupper
18515 ; VLX-NEXT: retq
18516 ;
18517 ; NoVLX-LABEL: test_vpcmpultq_v4i1_v32i1_mask:
18518 ; NoVLX: # %bb.0: # %entry
18519 ; NoVLX-NEXT: # kill: def $ymm1 killed $ymm1 def $zmm1
18520 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
18521 ; NoVLX-NEXT: vpcmpltuq %zmm1, %zmm0, %k0
18522 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
18523 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
18524 ; NoVLX-NEXT: kmovw %k0, %eax
18525 ; NoVLX-NEXT: vzeroupper
18526 ; NoVLX-NEXT: retq
18527 entry:
18528 %0 = bitcast <4 x i64> %__a to <4 x i64>
18529 %1 = bitcast <4 x i64> %__b to <4 x i64>
18530 %2 = icmp ult <4 x i64> %0, %1
18531 %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
18532 %4 = bitcast <32 x i1> %3 to i32
18533 ret i32 %4
18534 }
18536 define zeroext i32 @test_vpcmpultq_v4i1_v32i1_mask_mem(<4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
18537 ; VLX-LABEL: test_vpcmpultq_v4i1_v32i1_mask_mem:
18538 ; VLX: # %bb.0: # %entry
18539 ; VLX-NEXT: vpcmpltuq (%rdi), %ymm0, %k0
18540 ; VLX-NEXT: kmovd %k0, %eax
18541 ; VLX-NEXT: vzeroupper
18542 ; VLX-NEXT: retq
18543 ;
18544 ; NoVLX-LABEL: test_vpcmpultq_v4i1_v32i1_mask_mem:
18545 ; NoVLX: # %bb.0: # %entry
18546 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
18547 ; NoVLX-NEXT: vmovdqa (%rdi), %ymm1
18548 ; NoVLX-NEXT: vpcmpltuq %zmm1, %zmm0, %k0
18549 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
18550 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
18551 ; NoVLX-NEXT: kmovw %k0, %eax
18552 ; NoVLX-NEXT: vzeroupper
18553 ; NoVLX-NEXT: retq
18554 entry:
18555 %0 = bitcast <4 x i64> %__a to <4 x i64>
18556 %load = load <4 x i64>, <4 x i64>* %__b
18557 %1 = bitcast <4 x i64> %load to <4 x i64>
18558 %2 = icmp ult <4 x i64> %0, %1
18559 %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
18560 %4 = bitcast <32 x i1> %3 to i32
18561 ret i32 %4
18562 }
18564 define zeroext i32 @test_masked_vpcmpultq_v4i1_v32i1_mask(i8 zeroext %__u, <4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
18565 ; VLX-LABEL: test_masked_vpcmpultq_v4i1_v32i1_mask:
18566 ; VLX: # %bb.0: # %entry
18567 ; VLX-NEXT: kmovd %edi, %k1
18568 ; VLX-NEXT: vpcmpltuq %ymm1, %ymm0, %k0 {%k1}
18569 ; VLX-NEXT: kmovd %k0, %eax
18570 ; VLX-NEXT: vzeroupper
18571 ; VLX-NEXT: retq
18572 ;
18573 ; NoVLX-LABEL: test_masked_vpcmpultq_v4i1_v32i1_mask:
18574 ; NoVLX: # %bb.0: # %entry
18575 ; NoVLX-NEXT: # kill: def $ymm1 killed $ymm1 def $zmm1
18576 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
18577 ; NoVLX-NEXT: kmovw %edi, %k1
18578 ; NoVLX-NEXT: vpcmpltuq %zmm1, %zmm0, %k0 {%k1}
18579 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
18580 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
18581 ; NoVLX-NEXT: kmovw %k0, %eax
18582 ; NoVLX-NEXT: vzeroupper
18583 ; NoVLX-NEXT: retq
18584 entry:
18585 %0 = bitcast <4 x i64> %__a to <4 x i64>
18586 %1 = bitcast <4 x i64> %__b to <4 x i64>
18587 %2 = icmp ult <4 x i64> %0, %1
18588 %3 = bitcast i8 %__u to <8 x i1>
18589 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
18590 %4 = and <4 x i1> %2, %extract.i
18591 %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
18592 %6 = bitcast <32 x i1> %5 to i32
18593 ret i32 %6
18594 }
18596 define zeroext i32 @test_masked_vpcmpultq_v4i1_v32i1_mask_mem(i8 zeroext %__u, <4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
18597 ; VLX-LABEL: test_masked_vpcmpultq_v4i1_v32i1_mask_mem:
18598 ; VLX: # %bb.0: # %entry
18599 ; VLX-NEXT: kmovd %edi, %k1
18600 ; VLX-NEXT: vpcmpltuq (%rsi), %ymm0, %k0 {%k1}
18601 ; VLX-NEXT: kmovd %k0, %eax
18602 ; VLX-NEXT: vzeroupper
18603 ; VLX-NEXT: retq
18604 ;
18605 ; NoVLX-LABEL: test_masked_vpcmpultq_v4i1_v32i1_mask_mem:
18606 ; NoVLX: # %bb.0: # %entry
18607 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
18608 ; NoVLX-NEXT: vmovdqa (%rsi), %ymm1
18609 ; NoVLX-NEXT: kmovw %edi, %k1
18610 ; NoVLX-NEXT: vpcmpltuq %zmm1, %zmm0, %k0 {%k1}
18611 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
18612 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
18613 ; NoVLX-NEXT: kmovw %k0, %eax
18614 ; NoVLX-NEXT: vzeroupper
18615 ; NoVLX-NEXT: retq
18616 entry:
18617 %0 = bitcast <4 x i64> %__a to <4 x i64>
18618 %load = load <4 x i64>, <4 x i64>* %__b
18619 %1 = bitcast <4 x i64> %load to <4 x i64>
18620 %2 = icmp ult <4 x i64> %0, %1
18621 %3 = bitcast i8 %__u to <8 x i1>
18622 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
18623 %4 = and <4 x i1> %2, %extract.i
18624 %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
18625 %6 = bitcast <32 x i1> %5 to i32
18626 ret i32 %6
18627 }
18630 define zeroext i32 @test_vpcmpultq_v4i1_v32i1_mask_mem_b(<4 x i64> %__a, i64* %__b) local_unnamed_addr {
18631 ; VLX-LABEL: test_vpcmpultq_v4i1_v32i1_mask_mem_b:
18632 ; VLX: # %bb.0: # %entry
18633 ; VLX-NEXT: vpcmpltuq (%rdi){1to4}, %ymm0, %k0
18634 ; VLX-NEXT: kmovd %k0, %eax
18635 ; VLX-NEXT: vzeroupper
18636 ; VLX-NEXT: retq
18637 ;
18638 ; NoVLX-LABEL: test_vpcmpultq_v4i1_v32i1_mask_mem_b:
18639 ; NoVLX: # %bb.0: # %entry
18640 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
18641 ; NoVLX-NEXT: vpcmpltuq (%rdi){1to8}, %zmm0, %k0
18642 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
18643 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
18644 ; NoVLX-NEXT: kmovw %k0, %eax
18645 ; NoVLX-NEXT: vzeroupper
18646 ; NoVLX-NEXT: retq
18647 entry:
18648 %0 = bitcast <4 x i64> %__a to <4 x i64>
18649 %load = load i64, i64* %__b
18650 %vec = insertelement <4 x i64> undef, i64 %load, i32 0
18651 %1 = shufflevector <4 x i64> %vec, <4 x i64> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
18652 %2 = icmp ult <4 x i64> %0, %1
18653 %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
18654 %4 = bitcast <32 x i1> %3 to i32
18655 ret i32 %4
18656 }
18658 define zeroext i32 @test_masked_vpcmpultq_v4i1_v32i1_mask_mem_b(i8 zeroext %__u, <4 x i64> %__a, i64* %__b) local_unnamed_addr {
18659 ; VLX-LABEL: test_masked_vpcmpultq_v4i1_v32i1_mask_mem_b:
18660 ; VLX: # %bb.0: # %entry
18661 ; VLX-NEXT: kmovd %edi, %k1
18662 ; VLX-NEXT: vpcmpltuq (%rsi){1to4}, %ymm0, %k0 {%k1}
18663 ; VLX-NEXT: kmovd %k0, %eax
18664 ; VLX-NEXT: vzeroupper
18665 ; VLX-NEXT: retq
18666 ;
18667 ; NoVLX-LABEL: test_masked_vpcmpultq_v4i1_v32i1_mask_mem_b:
18668 ; NoVLX: # %bb.0: # %entry
18669 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
18670 ; NoVLX-NEXT: kmovw %edi, %k1
18671 ; NoVLX-NEXT: vpcmpltuq (%rsi){1to8}, %zmm0, %k0 {%k1}
18672 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
18673 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
18674 ; NoVLX-NEXT: kmovw %k0, %eax
18675 ; NoVLX-NEXT: vzeroupper
18676 ; NoVLX-NEXT: retq
18677 entry:
18678 %0 = bitcast <4 x i64> %__a to <4 x i64>
18679 %load = load i64, i64* %__b
18680 %vec = insertelement <4 x i64> undef, i64 %load, i32 0
18681 %1 = shufflevector <4 x i64> %vec, <4 x i64> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
18682 %2 = icmp ult <4 x i64> %0, %1
18683 %3 = bitcast i8 %__u to <8 x i1>
18684 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
18685 %4 = and <4 x i1> %extract.i, %2
18686 %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
18687 %6 = bitcast <32 x i1> %5 to i32
18688 ret i32 %6
18689 }
18692 define zeroext i64 @test_vpcmpultq_v4i1_v64i1_mask(<4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
18693 ; VLX-LABEL: test_vpcmpultq_v4i1_v64i1_mask:
18694 ; VLX: # %bb.0: # %entry
18695 ; VLX-NEXT: vpcmpltuq %ymm1, %ymm0, %k0
18696 ; VLX-NEXT: kmovq %k0, %rax
18697 ; VLX-NEXT: vzeroupper
18698 ; VLX-NEXT: retq
18699 ;
18700 ; NoVLX-LABEL: test_vpcmpultq_v4i1_v64i1_mask:
18701 ; NoVLX: # %bb.0: # %entry
18702 ; NoVLX-NEXT: # kill: def $ymm1 killed $ymm1 def $zmm1
18703 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
18704 ; NoVLX-NEXT: vpcmpltuq %zmm1, %zmm0, %k0
18705 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
18706 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
18707 ; NoVLX-NEXT: kmovw %k0, %eax
18708 ; NoVLX-NEXT: vzeroupper
18709 ; NoVLX-NEXT: retq
18710 entry:
18711 %0 = bitcast <4 x i64> %__a to <4 x i64>
18712 %1 = bitcast <4 x i64> %__b to <4 x i64>
18713 %2 = icmp ult <4 x i64> %0, %1
18714 %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
18715 %4 = bitcast <64 x i1> %3 to i64
18716 ret i64 %4
18717 }
18719 define zeroext i64 @test_vpcmpultq_v4i1_v64i1_mask_mem(<4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
18720 ; VLX-LABEL: test_vpcmpultq_v4i1_v64i1_mask_mem:
18721 ; VLX: # %bb.0: # %entry
18722 ; VLX-NEXT: vpcmpltuq (%rdi), %ymm0, %k0
18723 ; VLX-NEXT: kmovq %k0, %rax
18724 ; VLX-NEXT: vzeroupper
18725 ; VLX-NEXT: retq
18726 ;
18727 ; NoVLX-LABEL: test_vpcmpultq_v4i1_v64i1_mask_mem:
18728 ; NoVLX: # %bb.0: # %entry
18729 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
18730 ; NoVLX-NEXT: vmovdqa (%rdi), %ymm1
18731 ; NoVLX-NEXT: vpcmpltuq %zmm1, %zmm0, %k0
18732 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
18733 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
18734 ; NoVLX-NEXT: kmovw %k0, %eax
18735 ; NoVLX-NEXT: vzeroupper
18736 ; NoVLX-NEXT: retq
18737 entry:
18738 %0 = bitcast <4 x i64> %__a to <4 x i64>
18739 %load = load <4 x i64>, <4 x i64>* %__b
18740 %1 = bitcast <4 x i64> %load to <4 x i64>
18741 %2 = icmp ult <4 x i64> %0, %1
18742 %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
18743 %4 = bitcast <64 x i1> %3 to i64
18744 ret i64 %4
18745 }
18747 define zeroext i64 @test_masked_vpcmpultq_v4i1_v64i1_mask(i8 zeroext %__u, <4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
18748 ; VLX-LABEL: test_masked_vpcmpultq_v4i1_v64i1_mask:
18749 ; VLX: # %bb.0: # %entry
18750 ; VLX-NEXT: kmovd %edi, %k1
18751 ; VLX-NEXT: vpcmpltuq %ymm1, %ymm0, %k0 {%k1}
18752 ; VLX-NEXT: kmovq %k0, %rax
18753 ; VLX-NEXT: vzeroupper
18754 ; VLX-NEXT: retq
18755 ;
18756 ; NoVLX-LABEL: test_masked_vpcmpultq_v4i1_v64i1_mask:
18757 ; NoVLX: # %bb.0: # %entry
18758 ; NoVLX-NEXT: # kill: def $ymm1 killed $ymm1 def $zmm1
18759 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
18760 ; NoVLX-NEXT: kmovw %edi, %k1
18761 ; NoVLX-NEXT: vpcmpltuq %zmm1, %zmm0, %k0 {%k1}
18762 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
18763 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
18764 ; NoVLX-NEXT: kmovw %k0, %eax
18765 ; NoVLX-NEXT: vzeroupper
18766 ; NoVLX-NEXT: retq
18767 entry:
18768 %0 = bitcast <4 x i64> %__a to <4 x i64>
18769 %1 = bitcast <4 x i64> %__b to <4 x i64>
18770 %2 = icmp ult <4 x i64> %0, %1
18771 %3 = bitcast i8 %__u to <8 x i1>
18772 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
18773 %4 = and <4 x i1> %2, %extract.i
18774 %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
18775 %6 = bitcast <64 x i1> %5 to i64
18776 ret i64 %6
18777 }
18779 define zeroext i64 @test_masked_vpcmpultq_v4i1_v64i1_mask_mem(i8 zeroext %__u, <4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
18780 ; VLX-LABEL: test_masked_vpcmpultq_v4i1_v64i1_mask_mem:
18781 ; VLX: # %bb.0: # %entry
18782 ; VLX-NEXT: kmovd %edi, %k1
18783 ; VLX-NEXT: vpcmpltuq (%rsi), %ymm0, %k0 {%k1}
18784 ; VLX-NEXT: kmovq %k0, %rax
18785 ; VLX-NEXT: vzeroupper
18786 ; VLX-NEXT: retq
18787 ;
18788 ; NoVLX-LABEL: test_masked_vpcmpultq_v4i1_v64i1_mask_mem:
18789 ; NoVLX: # %bb.0: # %entry
18790 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
18791 ; NoVLX-NEXT: vmovdqa (%rsi), %ymm1
18792 ; NoVLX-NEXT: kmovw %edi, %k1
18793 ; NoVLX-NEXT: vpcmpltuq %zmm1, %zmm0, %k0 {%k1}
18794 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
18795 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
18796 ; NoVLX-NEXT: kmovw %k0, %eax
18797 ; NoVLX-NEXT: vzeroupper
18798 ; NoVLX-NEXT: retq
18799 entry:
18800 %0 = bitcast <4 x i64> %__a to <4 x i64>
18801 %load = load <4 x i64>, <4 x i64>* %__b
18802 %1 = bitcast <4 x i64> %load to <4 x i64>
18803 %2 = icmp ult <4 x i64> %0, %1
18804 %3 = bitcast i8 %__u to <8 x i1>
18805 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
18806 %4 = and <4 x i1> %2, %extract.i
18807 %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
18808 %6 = bitcast <64 x i1> %5 to i64
18809 ret i64 %6
18810 }
18813 define zeroext i64 @test_vpcmpultq_v4i1_v64i1_mask_mem_b(<4 x i64> %__a, i64* %__b) local_unnamed_addr {
18814 ; VLX-LABEL: test_vpcmpultq_v4i1_v64i1_mask_mem_b:
18815 ; VLX: # %bb.0: # %entry
18816 ; VLX-NEXT: vpcmpltuq (%rdi){1to4}, %ymm0, %k0
18817 ; VLX-NEXT: kmovq %k0, %rax
18818 ; VLX-NEXT: vzeroupper
18819 ; VLX-NEXT: retq
18820 ;
18821 ; NoVLX-LABEL: test_vpcmpultq_v4i1_v64i1_mask_mem_b:
18822 ; NoVLX: # %bb.0: # %entry
18823 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
18824 ; NoVLX-NEXT: vpcmpltuq (%rdi){1to8}, %zmm0, %k0
18825 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
18826 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
18827 ; NoVLX-NEXT: kmovw %k0, %eax
18828 ; NoVLX-NEXT: vzeroupper
18829 ; NoVLX-NEXT: retq
18830 entry:
18831 %0 = bitcast <4 x i64> %__a to <4 x i64>
18832 %load = load i64, i64* %__b
18833 %vec = insertelement <4 x i64> undef, i64 %load, i32 0
18834 %1 = shufflevector <4 x i64> %vec, <4 x i64> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
18835 %2 = icmp ult <4 x i64> %0, %1
18836 %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
18837 %4 = bitcast <64 x i1> %3 to i64
18838 ret i64 %4
18839 }
18841 define zeroext i64 @test_masked_vpcmpultq_v4i1_v64i1_mask_mem_b(i8 zeroext %__u, <4 x i64> %__a, i64* %__b) local_unnamed_addr {
18842 ; VLX-LABEL: test_masked_vpcmpultq_v4i1_v64i1_mask_mem_b:
18843 ; VLX: # %bb.0: # %entry
18844 ; VLX-NEXT: kmovd %edi, %k1
18845 ; VLX-NEXT: vpcmpltuq (%rsi){1to4}, %ymm0, %k0 {%k1}
18846 ; VLX-NEXT: kmovq %k0, %rax
18847 ; VLX-NEXT: vzeroupper
18848 ; VLX-NEXT: retq
18849 ;
18850 ; NoVLX-LABEL: test_masked_vpcmpultq_v4i1_v64i1_mask_mem_b:
18851 ; NoVLX: # %bb.0: # %entry
18852 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
18853 ; NoVLX-NEXT: kmovw %edi, %k1
18854 ; NoVLX-NEXT: vpcmpltuq (%rsi){1to8}, %zmm0, %k0 {%k1}
18855 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
18856 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
18857 ; NoVLX-NEXT: kmovw %k0, %eax
18858 ; NoVLX-NEXT: vzeroupper
18859 ; NoVLX-NEXT: retq
18860 entry:
18861 %0 = bitcast <4 x i64> %__a to <4 x i64>
18862 %load = load i64, i64* %__b
18863 %vec = insertelement <4 x i64> undef, i64 %load, i32 0
18864 %1 = shufflevector <4 x i64> %vec, <4 x i64> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
18865 %2 = icmp ult <4 x i64> %0, %1
18866 %3 = bitcast i8 %__u to <8 x i1>
18867 %extract.i = shufflevector <8 x i1> %3, <8 x i1> undef, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
18868 %4 = and <4 x i1> %extract.i, %2
18869 %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
18870 %6 = bitcast <64 x i1> %5 to i64
18871 ret i64 %6
18872 }
18875 define zeroext i16 @test_vpcmpultq_v8i1_v16i1_mask(<8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
18876 ; VLX-LABEL: test_vpcmpultq_v8i1_v16i1_mask:
18877 ; VLX: # %bb.0: # %entry
18878 ; VLX-NEXT: vpcmpltuq %zmm1, %zmm0, %k0
18879 ; VLX-NEXT: kmovd %k0, %eax
18880 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
18881 ; VLX-NEXT: vzeroupper
18882 ; VLX-NEXT: retq
18883 ;
18884 ; NoVLX-LABEL: test_vpcmpultq_v8i1_v16i1_mask:
18885 ; NoVLX: # %bb.0: # %entry
18886 ; NoVLX-NEXT: vpcmpltuq %zmm1, %zmm0, %k0
18887 ; NoVLX-NEXT: kmovw %k0, %eax
18888 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
18889 ; NoVLX-NEXT: vzeroupper
18890 ; NoVLX-NEXT: retq
18891 entry:
18892 %0 = bitcast <8 x i64> %__a to <8 x i64>
18893 %1 = bitcast <8 x i64> %__b to <8 x i64>
18894 %2 = icmp ult <8 x i64> %0, %1
18895 %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
18896 %4 = bitcast <16 x i1> %3 to i16
18897 ret i16 %4
18898 }
18900 define zeroext i16 @test_vpcmpultq_v8i1_v16i1_mask_mem(<8 x i64> %__a, <8 x i64>* %__b) local_unnamed_addr {
18901 ; VLX-LABEL: test_vpcmpultq_v8i1_v16i1_mask_mem:
18902 ; VLX: # %bb.0: # %entry
18903 ; VLX-NEXT: vpcmpltuq (%rdi), %zmm0, %k0
18904 ; VLX-NEXT: kmovd %k0, %eax
18905 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
18906 ; VLX-NEXT: vzeroupper
18907 ; VLX-NEXT: retq
18908 ;
18909 ; NoVLX-LABEL: test_vpcmpultq_v8i1_v16i1_mask_mem:
18910 ; NoVLX: # %bb.0: # %entry
18911 ; NoVLX-NEXT: vpcmpltuq (%rdi), %zmm0, %k0
18912 ; NoVLX-NEXT: kmovw %k0, %eax
18913 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
18914 ; NoVLX-NEXT: vzeroupper
18915 ; NoVLX-NEXT: retq
18916 entry:
18917 %0 = bitcast <8 x i64> %__a to <8 x i64>
18918 %load = load <8 x i64>, <8 x i64>* %__b
18919 %1 = bitcast <8 x i64> %load to <8 x i64>
18920 %2 = icmp ult <8 x i64> %0, %1
18921 %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
18922 %4 = bitcast <16 x i1> %3 to i16
18923 ret i16 %4
18924 }
18926 define zeroext i16 @test_masked_vpcmpultq_v8i1_v16i1_mask(i8 zeroext %__u, <8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
18927 ; VLX-LABEL: test_masked_vpcmpultq_v8i1_v16i1_mask:
18928 ; VLX: # %bb.0: # %entry
18929 ; VLX-NEXT: kmovd %edi, %k1
18930 ; VLX-NEXT: vpcmpltuq %zmm1, %zmm0, %k0 {%k1}
18931 ; VLX-NEXT: kmovd %k0, %eax
18932 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
18933 ; VLX-NEXT: vzeroupper
18934 ; VLX-NEXT: retq
18935 ;
18936 ; NoVLX-LABEL: test_masked_vpcmpultq_v8i1_v16i1_mask:
18937 ; NoVLX: # %bb.0: # %entry
18938 ; NoVLX-NEXT: kmovw %edi, %k1
18939 ; NoVLX-NEXT: vpcmpltuq %zmm1, %zmm0, %k0 {%k1}
18940 ; NoVLX-NEXT: kmovw %k0, %eax
18941 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
18942 ; NoVLX-NEXT: vzeroupper
18943 ; NoVLX-NEXT: retq
18944 entry:
18945 %0 = bitcast <8 x i64> %__a to <8 x i64>
18946 %1 = bitcast <8 x i64> %__b to <8 x i64>
18947 %2 = icmp ult <8 x i64> %0, %1
18948 %3 = bitcast i8 %__u to <8 x i1>
18949 %4 = and <8 x i1> %2, %3
18950 %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
18951 %6 = bitcast <16 x i1> %5 to i16
18952 ret i16 %6
18953 }
18955 define zeroext i16 @test_masked_vpcmpultq_v8i1_v16i1_mask_mem(i8 zeroext %__u, <8 x i64> %__a, <8 x i64>* %__b) local_unnamed_addr {
18956 ; VLX-LABEL: test_masked_vpcmpultq_v8i1_v16i1_mask_mem:
18957 ; VLX: # %bb.0: # %entry
18958 ; VLX-NEXT: kmovd %edi, %k1
18959 ; VLX-NEXT: vpcmpltuq (%rsi), %zmm0, %k0 {%k1}
18960 ; VLX-NEXT: kmovd %k0, %eax
18961 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
18962 ; VLX-NEXT: vzeroupper
18963 ; VLX-NEXT: retq
18964 ;
18965 ; NoVLX-LABEL: test_masked_vpcmpultq_v8i1_v16i1_mask_mem:
18966 ; NoVLX: # %bb.0: # %entry
18967 ; NoVLX-NEXT: kmovw %edi, %k1
18968 ; NoVLX-NEXT: vpcmpltuq (%rsi), %zmm0, %k0 {%k1}
18969 ; NoVLX-NEXT: kmovw %k0, %eax
18970 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
18971 ; NoVLX-NEXT: vzeroupper
18972 ; NoVLX-NEXT: retq
18973 entry:
18974 %0 = bitcast <8 x i64> %__a to <8 x i64>
18975 %load = load <8 x i64>, <8 x i64>* %__b
18976 %1 = bitcast <8 x i64> %load to <8 x i64>
18977 %2 = icmp ult <8 x i64> %0, %1
18978 %3 = bitcast i8 %__u to <8 x i1>
18979 %4 = and <8 x i1> %2, %3
18980 %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
18981 %6 = bitcast <16 x i1> %5 to i16
18982 ret i16 %6
18983 }
18986 define zeroext i16 @test_vpcmpultq_v8i1_v16i1_mask_mem_b(<8 x i64> %__a, i64* %__b) local_unnamed_addr {
18987 ; VLX-LABEL: test_vpcmpultq_v8i1_v16i1_mask_mem_b:
18988 ; VLX: # %bb.0: # %entry
18989 ; VLX-NEXT: vpcmpltuq (%rdi){1to8}, %zmm0, %k0
18990 ; VLX-NEXT: kmovd %k0, %eax
18991 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
18992 ; VLX-NEXT: vzeroupper
18993 ; VLX-NEXT: retq
18994 ;
18995 ; NoVLX-LABEL: test_vpcmpultq_v8i1_v16i1_mask_mem_b:
18996 ; NoVLX: # %bb.0: # %entry
18997 ; NoVLX-NEXT: vpcmpltuq (%rdi){1to8}, %zmm0, %k0
18998 ; NoVLX-NEXT: kmovw %k0, %eax
18999 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
19000 ; NoVLX-NEXT: vzeroupper
19001 ; NoVLX-NEXT: retq
19002 entry:
19003 %0 = bitcast <8 x i64> %__a to <8 x i64>
19004 %load = load i64, i64* %__b
19005 %vec = insertelement <8 x i64> undef, i64 %load, i32 0
19006 %1 = shufflevector <8 x i64> %vec, <8 x i64> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
19007 %2 = icmp ult <8 x i64> %0, %1
19008 %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
19009 %4 = bitcast <16 x i1> %3 to i16
19010 ret i16 %4
19011 }
19013 define zeroext i16 @test_masked_vpcmpultq_v8i1_v16i1_mask_mem_b(i8 zeroext %__u, <8 x i64> %__a, i64* %__b) local_unnamed_addr {
19014 ; VLX-LABEL: test_masked_vpcmpultq_v8i1_v16i1_mask_mem_b:
19015 ; VLX: # %bb.0: # %entry
19016 ; VLX-NEXT: kmovd %edi, %k1
19017 ; VLX-NEXT: vpcmpltuq (%rsi){1to8}, %zmm0, %k0 {%k1}
19018 ; VLX-NEXT: kmovd %k0, %eax
19019 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
19020 ; VLX-NEXT: vzeroupper
19021 ; VLX-NEXT: retq
19022 ;
19023 ; NoVLX-LABEL: test_masked_vpcmpultq_v8i1_v16i1_mask_mem_b:
19024 ; NoVLX: # %bb.0: # %entry
19025 ; NoVLX-NEXT: kmovw %edi, %k1
19026 ; NoVLX-NEXT: vpcmpltuq (%rsi){1to8}, %zmm0, %k0 {%k1}
19027 ; NoVLX-NEXT: kmovw %k0, %eax
19028 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
19029 ; NoVLX-NEXT: vzeroupper
19030 ; NoVLX-NEXT: retq
19031 entry:
19032 %0 = bitcast <8 x i64> %__a to <8 x i64>
19033 %load = load i64, i64* %__b
19034 %vec = insertelement <8 x i64> undef, i64 %load, i32 0
19035 %1 = shufflevector <8 x i64> %vec, <8 x i64> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
19036 %2 = icmp ult <8 x i64> %0, %1
19037 %3 = bitcast i8 %__u to <8 x i1>
19038 %4 = and <8 x i1> %3, %2
19039 %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
19040 %6 = bitcast <16 x i1> %5 to i16
19041 ret i16 %6
19042 }
19045 define zeroext i32 @test_vpcmpultq_v8i1_v32i1_mask(<8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
19046 ; VLX-LABEL: test_vpcmpultq_v8i1_v32i1_mask:
19047 ; VLX: # %bb.0: # %entry
19048 ; VLX-NEXT: vpcmpltuq %zmm1, %zmm0, %k0
19049 ; VLX-NEXT: kmovd %k0, %eax
19050 ; VLX-NEXT: vzeroupper
19051 ; VLX-NEXT: retq
19052 ;
19053 ; NoVLX-LABEL: test_vpcmpultq_v8i1_v32i1_mask:
19054 ; NoVLX: # %bb.0: # %entry
19055 ; NoVLX-NEXT: vpcmpltuq %zmm1, %zmm0, %k0
19056 ; NoVLX-NEXT: kmovw %k0, %eax
19057 ; NoVLX-NEXT: vzeroupper
19058 ; NoVLX-NEXT: retq
19059 entry:
19060 %0 = bitcast <8 x i64> %__a to <8 x i64>
19061 %1 = bitcast <8 x i64> %__b to <8 x i64>
19062 %2 = icmp ult <8 x i64> %0, %1
19063 %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
19064 %4 = bitcast <32 x i1> %3 to i32
19065 ret i32 %4
19066 }
19068 define zeroext i32 @test_vpcmpultq_v8i1_v32i1_mask_mem(<8 x i64> %__a, <8 x i64>* %__b) local_unnamed_addr {
19069 ; VLX-LABEL: test_vpcmpultq_v8i1_v32i1_mask_mem:
19070 ; VLX: # %bb.0: # %entry
19071 ; VLX-NEXT: vpcmpltuq (%rdi), %zmm0, %k0
19072 ; VLX-NEXT: kmovd %k0, %eax
19073 ; VLX-NEXT: vzeroupper
19074 ; VLX-NEXT: retq
19075 ;
19076 ; NoVLX-LABEL: test_vpcmpultq_v8i1_v32i1_mask_mem:
19077 ; NoVLX: # %bb.0: # %entry
19078 ; NoVLX-NEXT: vpcmpltuq (%rdi), %zmm0, %k0
19079 ; NoVLX-NEXT: kmovw %k0, %eax
19080 ; NoVLX-NEXT: vzeroupper
19081 ; NoVLX-NEXT: retq
19082 entry:
19083 %0 = bitcast <8 x i64> %__a to <8 x i64>
19084 %load = load <8 x i64>, <8 x i64>* %__b
19085 %1 = bitcast <8 x i64> %load to <8 x i64>
19086 %2 = icmp ult <8 x i64> %0, %1
19087 %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
19088 %4 = bitcast <32 x i1> %3 to i32
19089 ret i32 %4
19090 }
19092 define zeroext i32 @test_masked_vpcmpultq_v8i1_v32i1_mask(i8 zeroext %__u, <8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
19093 ; VLX-LABEL: test_masked_vpcmpultq_v8i1_v32i1_mask:
19094 ; VLX: # %bb.0: # %entry
19095 ; VLX-NEXT: kmovd %edi, %k1
19096 ; VLX-NEXT: vpcmpltuq %zmm1, %zmm0, %k0 {%k1}
19097 ; VLX-NEXT: kmovd %k0, %eax
19098 ; VLX-NEXT: vzeroupper
19099 ; VLX-NEXT: retq
19100 ;
19101 ; NoVLX-LABEL: test_masked_vpcmpultq_v8i1_v32i1_mask:
19102 ; NoVLX: # %bb.0: # %entry
19103 ; NoVLX-NEXT: kmovw %edi, %k1
19104 ; NoVLX-NEXT: vpcmpltuq %zmm1, %zmm0, %k0 {%k1}
19105 ; NoVLX-NEXT: kmovw %k0, %eax
19106 ; NoVLX-NEXT: vzeroupper
19107 ; NoVLX-NEXT: retq
19108 entry:
19109 %0 = bitcast <8 x i64> %__a to <8 x i64>
19110 %1 = bitcast <8 x i64> %__b to <8 x i64>
19111 %2 = icmp ult <8 x i64> %0, %1
19112 %3 = bitcast i8 %__u to <8 x i1>
19113 %4 = and <8 x i1> %2, %3
19114 %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
19115 %6 = bitcast <32 x i1> %5 to i32
19116 ret i32 %6
19117 }
19119 define zeroext i32 @test_masked_vpcmpultq_v8i1_v32i1_mask_mem(i8 zeroext %__u, <8 x i64> %__a, <8 x i64>* %__b) local_unnamed_addr {
19120 ; VLX-LABEL: test_masked_vpcmpultq_v8i1_v32i1_mask_mem:
19121 ; VLX: # %bb.0: # %entry
19122 ; VLX-NEXT: kmovd %edi, %k1
19123 ; VLX-NEXT: vpcmpltuq (%rsi), %zmm0, %k0 {%k1}
19124 ; VLX-NEXT: kmovd %k0, %eax
19125 ; VLX-NEXT: vzeroupper
19126 ; VLX-NEXT: retq
19127 ;
19128 ; NoVLX-LABEL: test_masked_vpcmpultq_v8i1_v32i1_mask_mem:
19129 ; NoVLX: # %bb.0: # %entry
19130 ; NoVLX-NEXT: kmovw %edi, %k1
19131 ; NoVLX-NEXT: vpcmpltuq (%rsi), %zmm0, %k0 {%k1}
19132 ; NoVLX-NEXT: kmovw %k0, %eax
19133 ; NoVLX-NEXT: vzeroupper
19134 ; NoVLX-NEXT: retq
19135 entry:
19136 %0 = bitcast <8 x i64> %__a to <8 x i64>
19137 %load = load <8 x i64>, <8 x i64>* %__b
19138 %1 = bitcast <8 x i64> %load to <8 x i64>
19139 %2 = icmp ult <8 x i64> %0, %1
19140 %3 = bitcast i8 %__u to <8 x i1>
19141 %4 = and <8 x i1> %2, %3
19142 %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
19143 %6 = bitcast <32 x i1> %5 to i32
19144 ret i32 %6
19145 }
19148 define zeroext i32 @test_vpcmpultq_v8i1_v32i1_mask_mem_b(<8 x i64> %__a, i64* %__b) local_unnamed_addr {
19149 ; VLX-LABEL: test_vpcmpultq_v8i1_v32i1_mask_mem_b:
19150 ; VLX: # %bb.0: # %entry
19151 ; VLX-NEXT: vpcmpltuq (%rdi){1to8}, %zmm0, %k0
19152 ; VLX-NEXT: kmovd %k0, %eax
19153 ; VLX-NEXT: vzeroupper
19154 ; VLX-NEXT: retq
19155 ;
19156 ; NoVLX-LABEL: test_vpcmpultq_v8i1_v32i1_mask_mem_b:
19157 ; NoVLX: # %bb.0: # %entry
19158 ; NoVLX-NEXT: vpcmpltuq (%rdi){1to8}, %zmm0, %k0
19159 ; NoVLX-NEXT: kmovw %k0, %eax
19160 ; NoVLX-NEXT: vzeroupper
19161 ; NoVLX-NEXT: retq
19162 entry:
19163 %0 = bitcast <8 x i64> %__a to <8 x i64>
19164 %load = load i64, i64* %__b
19165 %vec = insertelement <8 x i64> undef, i64 %load, i32 0
19166 %1 = shufflevector <8 x i64> %vec, <8 x i64> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
19167 %2 = icmp ult <8 x i64> %0, %1
19168 %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
19169 %4 = bitcast <32 x i1> %3 to i32
19170 ret i32 %4
19171 }
19173 define zeroext i32 @test_masked_vpcmpultq_v8i1_v32i1_mask_mem_b(i8 zeroext %__u, <8 x i64> %__a, i64* %__b) local_unnamed_addr {
19174 ; VLX-LABEL: test_masked_vpcmpultq_v8i1_v32i1_mask_mem_b:
19175 ; VLX: # %bb.0: # %entry
19176 ; VLX-NEXT: kmovd %edi, %k1
19177 ; VLX-NEXT: vpcmpltuq (%rsi){1to8}, %zmm0, %k0 {%k1}
19178 ; VLX-NEXT: kmovd %k0, %eax
19179 ; VLX-NEXT: vzeroupper
19180 ; VLX-NEXT: retq
19181 ;
19182 ; NoVLX-LABEL: test_masked_vpcmpultq_v8i1_v32i1_mask_mem_b:
19183 ; NoVLX: # %bb.0: # %entry
19184 ; NoVLX-NEXT: kmovw %edi, %k1
19185 ; NoVLX-NEXT: vpcmpltuq (%rsi){1to8}, %zmm0, %k0 {%k1}
19186 ; NoVLX-NEXT: kmovw %k0, %eax
19187 ; NoVLX-NEXT: vzeroupper
19188 ; NoVLX-NEXT: retq
19189 entry:
19190 %0 = bitcast <8 x i64> %__a to <8 x i64>
19191 %load = load i64, i64* %__b
19192 %vec = insertelement <8 x i64> undef, i64 %load, i32 0
19193 %1 = shufflevector <8 x i64> %vec, <8 x i64> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
19194 %2 = icmp ult <8 x i64> %0, %1
19195 %3 = bitcast i8 %__u to <8 x i1>
19196 %4 = and <8 x i1> %3, %2
19197 %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
19198 %6 = bitcast <32 x i1> %5 to i32
19199 ret i32 %6
19200 }
19203 define zeroext i64 @test_vpcmpultq_v8i1_v64i1_mask(<8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
19204 ; VLX-LABEL: test_vpcmpultq_v8i1_v64i1_mask:
19205 ; VLX: # %bb.0: # %entry
19206 ; VLX-NEXT: vpcmpltuq %zmm1, %zmm0, %k0
19207 ; VLX-NEXT: kmovq %k0, %rax
19208 ; VLX-NEXT: vzeroupper
19209 ; VLX-NEXT: retq
19210 ;
19211 ; NoVLX-LABEL: test_vpcmpultq_v8i1_v64i1_mask:
19212 ; NoVLX: # %bb.0: # %entry
19213 ; NoVLX-NEXT: vpcmpltuq %zmm1, %zmm0, %k0
19214 ; NoVLX-NEXT: kmovw %k0, %eax
19215 ; NoVLX-NEXT: vzeroupper
19216 ; NoVLX-NEXT: retq
19217 entry:
19218 %0 = bitcast <8 x i64> %__a to <8 x i64>
19219 %1 = bitcast <8 x i64> %__b to <8 x i64>
19220 %2 = icmp ult <8 x i64> %0, %1
19221 %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
19222 %4 = bitcast <64 x i1> %3 to i64
19223 ret i64 %4
19224 }
19226 define zeroext i64 @test_vpcmpultq_v8i1_v64i1_mask_mem(<8 x i64> %__a, <8 x i64>* %__b) local_unnamed_addr {
19227 ; VLX-LABEL: test_vpcmpultq_v8i1_v64i1_mask_mem:
19228 ; VLX: # %bb.0: # %entry
19229 ; VLX-NEXT: vpcmpltuq (%rdi), %zmm0, %k0
19230 ; VLX-NEXT: kmovq %k0, %rax
19231 ; VLX-NEXT: vzeroupper
19232 ; VLX-NEXT: retq
19233 ;
19234 ; NoVLX-LABEL: test_vpcmpultq_v8i1_v64i1_mask_mem:
19235 ; NoVLX: # %bb.0: # %entry
19236 ; NoVLX-NEXT: vpcmpltuq (%rdi), %zmm0, %k0
19237 ; NoVLX-NEXT: kmovw %k0, %eax
19238 ; NoVLX-NEXT: vzeroupper
19239 ; NoVLX-NEXT: retq
19240 entry:
19241 %0 = bitcast <8 x i64> %__a to <8 x i64>
19242 %load = load <8 x i64>, <8 x i64>* %__b
19243 %1 = bitcast <8 x i64> %load to <8 x i64>
19244 %2 = icmp ult <8 x i64> %0, %1
19245 %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
19246 %4 = bitcast <64 x i1> %3 to i64
19247 ret i64 %4
19248 }
19250 define zeroext i64 @test_masked_vpcmpultq_v8i1_v64i1_mask(i8 zeroext %__u, <8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
19251 ; VLX-LABEL: test_masked_vpcmpultq_v8i1_v64i1_mask:
19252 ; VLX: # %bb.0: # %entry
19253 ; VLX-NEXT: kmovd %edi, %k1
19254 ; VLX-NEXT: vpcmpltuq %zmm1, %zmm0, %k0 {%k1}
19255 ; VLX-NEXT: kmovq %k0, %rax
19256 ; VLX-NEXT: vzeroupper
19257 ; VLX-NEXT: retq
19258 ;
19259 ; NoVLX-LABEL: test_masked_vpcmpultq_v8i1_v64i1_mask:
19260 ; NoVLX: # %bb.0: # %entry
19261 ; NoVLX-NEXT: kmovw %edi, %k1
19262 ; NoVLX-NEXT: vpcmpltuq %zmm1, %zmm0, %k0 {%k1}
19263 ; NoVLX-NEXT: kmovw %k0, %eax
19264 ; NoVLX-NEXT: vzeroupper
19265 ; NoVLX-NEXT: retq
19266 entry:
19267 %0 = bitcast <8 x i64> %__a to <8 x i64>
19268 %1 = bitcast <8 x i64> %__b to <8 x i64>
19269 %2 = icmp ult <8 x i64> %0, %1
19270 %3 = bitcast i8 %__u to <8 x i1>
19271 %4 = and <8 x i1> %2, %3
19272 %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
19273 %6 = bitcast <64 x i1> %5 to i64
19274 ret i64 %6
19275 }
19277 define zeroext i64 @test_masked_vpcmpultq_v8i1_v64i1_mask_mem(i8 zeroext %__u, <8 x i64> %__a, <8 x i64>* %__b) local_unnamed_addr {
19278 ; VLX-LABEL: test_masked_vpcmpultq_v8i1_v64i1_mask_mem:
19279 ; VLX: # %bb.0: # %entry
19280 ; VLX-NEXT: kmovd %edi, %k1
19281 ; VLX-NEXT: vpcmpltuq (%rsi), %zmm0, %k0 {%k1}
19282 ; VLX-NEXT: kmovq %k0, %rax
19283 ; VLX-NEXT: vzeroupper
19284 ; VLX-NEXT: retq
19285 ;
19286 ; NoVLX-LABEL: test_masked_vpcmpultq_v8i1_v64i1_mask_mem:
19287 ; NoVLX: # %bb.0: # %entry
19288 ; NoVLX-NEXT: kmovw %edi, %k1
19289 ; NoVLX-NEXT: vpcmpltuq (%rsi), %zmm0, %k0 {%k1}
19290 ; NoVLX-NEXT: kmovw %k0, %eax
19291 ; NoVLX-NEXT: vzeroupper
19292 ; NoVLX-NEXT: retq
19293 entry:
19294 %0 = bitcast <8 x i64> %__a to <8 x i64>
19295 %load = load <8 x i64>, <8 x i64>* %__b
19296 %1 = bitcast <8 x i64> %load to <8 x i64>
19297 %2 = icmp ult <8 x i64> %0, %1
19298 %3 = bitcast i8 %__u to <8 x i1>
19299 %4 = and <8 x i1> %2, %3
19300 %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
19301 %6 = bitcast <64 x i1> %5 to i64
19302 ret i64 %6
19303 }
19306 define zeroext i64 @test_vpcmpultq_v8i1_v64i1_mask_mem_b(<8 x i64> %__a, i64* %__b) local_unnamed_addr {
19307 ; VLX-LABEL: test_vpcmpultq_v8i1_v64i1_mask_mem_b:
19308 ; VLX: # %bb.0: # %entry
19309 ; VLX-NEXT: vpcmpltuq (%rdi){1to8}, %zmm0, %k0
19310 ; VLX-NEXT: kmovq %k0, %rax
19311 ; VLX-NEXT: vzeroupper
19312 ; VLX-NEXT: retq
19313 ;
19314 ; NoVLX-LABEL: test_vpcmpultq_v8i1_v64i1_mask_mem_b:
19315 ; NoVLX: # %bb.0: # %entry
19316 ; NoVLX-NEXT: vpcmpltuq (%rdi){1to8}, %zmm0, %k0
19317 ; NoVLX-NEXT: kmovw %k0, %eax
19318 ; NoVLX-NEXT: vzeroupper
19319 ; NoVLX-NEXT: retq
19320 entry:
19321 %0 = bitcast <8 x i64> %__a to <8 x i64>
19322 %load = load i64, i64* %__b
19323 %vec = insertelement <8 x i64> undef, i64 %load, i32 0
19324 %1 = shufflevector <8 x i64> %vec, <8 x i64> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
19325 %2 = icmp ult <8 x i64> %0, %1
19326 %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
19327 %4 = bitcast <64 x i1> %3 to i64
19328 ret i64 %4
19329 }
19331 define zeroext i64 @test_masked_vpcmpultq_v8i1_v64i1_mask_mem_b(i8 zeroext %__u, <8 x i64> %__a, i64* %__b) local_unnamed_addr {
19332 ; VLX-LABEL: test_masked_vpcmpultq_v8i1_v64i1_mask_mem_b:
19333 ; VLX: # %bb.0: # %entry
19334 ; VLX-NEXT: kmovd %edi, %k1
19335 ; VLX-NEXT: vpcmpltuq (%rsi){1to8}, %zmm0, %k0 {%k1}
19336 ; VLX-NEXT: kmovq %k0, %rax
19337 ; VLX-NEXT: vzeroupper
19338 ; VLX-NEXT: retq
19339 ;
19340 ; NoVLX-LABEL: test_masked_vpcmpultq_v8i1_v64i1_mask_mem_b:
19341 ; NoVLX: # %bb.0: # %entry
19342 ; NoVLX-NEXT: kmovw %edi, %k1
19343 ; NoVLX-NEXT: vpcmpltuq (%rsi){1to8}, %zmm0, %k0 {%k1}
19344 ; NoVLX-NEXT: kmovw %k0, %eax
19345 ; NoVLX-NEXT: vzeroupper
19346 ; NoVLX-NEXT: retq
19347 entry:
19348 %0 = bitcast <8 x i64> %__a to <8 x i64>
19349 %load = load i64, i64* %__b
19350 %vec = insertelement <8 x i64> undef, i64 %load, i32 0
19351 %1 = shufflevector <8 x i64> %vec, <8 x i64> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
19352 %2 = icmp ult <8 x i64> %0, %1
19353 %3 = bitcast i8 %__u to <8 x i1>
19354 %4 = and <8 x i1> %3, %2
19355 %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
19356 %6 = bitcast <64 x i1> %5 to i64
19357 ret i64 %6
19358 }
19361 declare <16 x i1> @llvm.x86.avx512.cmp.ps.512(<16 x float>, <16 x float>, i32, i32)
19362 define zeroext i8 @test_vcmpoeqps_v4i1_v8i1_mask(<2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
19363 ; VLX-LABEL: test_vcmpoeqps_v4i1_v8i1_mask:
19364 ; VLX: # %bb.0: # %entry
19365 ; VLX-NEXT: vcmpeqps %xmm1, %xmm0, %k0
19366 ; VLX-NEXT: kmovd %k0, %eax
19367 ; VLX-NEXT: # kill: def $al killed $al killed $eax
19368 ; VLX-NEXT: retq
19369 ;
19370 ; NoVLX-LABEL: test_vcmpoeqps_v4i1_v8i1_mask:
19371 ; NoVLX: # %bb.0: # %entry
19372 ; NoVLX-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
19373 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
19374 ; NoVLX-NEXT: vcmpeqps %zmm1, %zmm0, %k0
19375 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
19376 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
19377 ; NoVLX-NEXT: kmovw %k0, %eax
19378 ; NoVLX-NEXT: # kill: def $al killed $al killed $eax
19379 ; NoVLX-NEXT: vzeroupper
19380 ; NoVLX-NEXT: retq
19381 entry:
19382 %0 = bitcast <2 x i64> %__a to <4 x float>
19383 %1 = bitcast <2 x i64> %__b to <4 x float>
19384 %2 = fcmp oeq <4 x float> %0, %1
19385 %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
19386 %4 = bitcast <8 x i1> %3 to i8
19387 ret i8 %4
19388 }
19390 define zeroext i8 @test_vcmpoeqps_v4i1_v8i1_mask_mem(<2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
19391 ; VLX-LABEL: test_vcmpoeqps_v4i1_v8i1_mask_mem:
19392 ; VLX: # %bb.0: # %entry
19393 ; VLX-NEXT: vcmpeqps (%rdi), %xmm0, %k0
19394 ; VLX-NEXT: kmovd %k0, %eax
19395 ; VLX-NEXT: # kill: def $al killed $al killed $eax
19396 ; VLX-NEXT: retq
19397 ;
19398 ; NoVLX-LABEL: test_vcmpoeqps_v4i1_v8i1_mask_mem:
19399 ; NoVLX: # %bb.0: # %entry
19400 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
19401 ; NoVLX-NEXT: vmovaps (%rdi), %xmm1
19402 ; NoVLX-NEXT: vcmpeqps %zmm1, %zmm0, %k0
19403 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
19404 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
19405 ; NoVLX-NEXT: kmovw %k0, %eax
19406 ; NoVLX-NEXT: # kill: def $al killed $al killed $eax
19407 ; NoVLX-NEXT: vzeroupper
19408 ; NoVLX-NEXT: retq
19409 entry:
19410 %0 = bitcast <2 x i64> %__a to <4 x float>
19411 %load = load <2 x i64>, <2 x i64>* %__b
19412 %1 = bitcast <2 x i64> %load to <4 x float>
19413 %2 = fcmp oeq <4 x float> %0, %1
19414 %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
19415 %4 = bitcast <8 x i1> %3 to i8
19416 ret i8 %4
19417 }
19419 define zeroext i8 @test_vcmpoeqps_v4i1_v8i1_mask_mem_b(<2 x i64> %__a, float* %__b) local_unnamed_addr {
19420 ; VLX-LABEL: test_vcmpoeqps_v4i1_v8i1_mask_mem_b:
19421 ; VLX: # %bb.0: # %entry
19422 ; VLX-NEXT: vcmpeqps (%rdi){1to4}, %xmm0, %k0
19423 ; VLX-NEXT: kmovd %k0, %eax
19424 ; VLX-NEXT: # kill: def $al killed $al killed $eax
19425 ; VLX-NEXT: retq
19426 ;
19427 ; NoVLX-LABEL: test_vcmpoeqps_v4i1_v8i1_mask_mem_b:
19428 ; NoVLX: # %bb.0: # %entry
19429 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
19430 ; NoVLX-NEXT: vcmpeqps (%rdi){1to16}, %zmm0, %k0
19431 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
19432 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
19433 ; NoVLX-NEXT: kmovw %k0, %eax
19434 ; NoVLX-NEXT: # kill: def $al killed $al killed $eax
19435 ; NoVLX-NEXT: vzeroupper
19436 ; NoVLX-NEXT: retq
19437 entry:
19438 %0 = bitcast <2 x i64> %__a to <4 x float>
19439 %load = load float, float* %__b
19440 %vec = insertelement <4 x float> undef, float %load, i32 0
19441 %1 = shufflevector <4 x float> %vec, <4 x float> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
19442 %2 = fcmp oeq <4 x float> %0, %1
19443 %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
19444 %4 = bitcast <8 x i1> %3 to i8
19445 ret i8 %4
19446 }
19448 define zeroext i8 @test_masked_vcmpoeqps_v4i1_v8i1_mask(i4 zeroext %__u, <2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
19449 ; VLX-LABEL: test_masked_vcmpoeqps_v4i1_v8i1_mask:
19450 ; VLX: # %bb.0: # %entry
19451 ; VLX-NEXT: kmovd %edi, %k1
19452 ; VLX-NEXT: vcmpeqps %xmm1, %xmm0, %k0 {%k1}
19453 ; VLX-NEXT: kmovd %k0, %eax
19454 ; VLX-NEXT: # kill: def $al killed $al killed $eax
19455 ; VLX-NEXT: retq
19456 ;
19457 ; NoVLX-LABEL: test_masked_vcmpoeqps_v4i1_v8i1_mask:
19458 ; NoVLX: # %bb.0: # %entry
19459 ; NoVLX-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
19460 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
19461 ; NoVLX-NEXT: kmovw %edi, %k1
19462 ; NoVLX-NEXT: vcmpeqps %zmm1, %zmm0, %k0 {%k1}
19463 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
19464 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
19465 ; NoVLX-NEXT: kmovw %k0, %eax
19466 ; NoVLX-NEXT: # kill: def $al killed $al killed $eax
19467 ; NoVLX-NEXT: vzeroupper
19468 ; NoVLX-NEXT: retq
19469 entry:
19470 %0 = bitcast <2 x i64> %__a to <4 x float>
19471 %1 = bitcast <2 x i64> %__b to <4 x float>
19472 %2 = fcmp oeq <4 x float> %0, %1
19473 %3 = bitcast i4 %__u to <4 x i1>
19474 %4 = and <4 x i1> %2, %3
19475 %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
19476 %6 = bitcast <8 x i1> %5 to i8
19477 ret i8 %6
19478 }
19480 define zeroext i8 @test_masked_vcmpoeqps_v4i1_v8i1_mask_mem(i4 zeroext %__u, <2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
19481 ; VLX-LABEL: test_masked_vcmpoeqps_v4i1_v8i1_mask_mem:
19482 ; VLX: # %bb.0: # %entry
19483 ; VLX-NEXT: kmovd %edi, %k1
19484 ; VLX-NEXT: vcmpeqps (%rsi), %xmm0, %k0 {%k1}
19485 ; VLX-NEXT: kmovd %k0, %eax
19486 ; VLX-NEXT: # kill: def $al killed $al killed $eax
19487 ; VLX-NEXT: retq
19488 ;
19489 ; NoVLX-LABEL: test_masked_vcmpoeqps_v4i1_v8i1_mask_mem:
19490 ; NoVLX: # %bb.0: # %entry
19491 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
19492 ; NoVLX-NEXT: kmovw %edi, %k1
19493 ; NoVLX-NEXT: vmovaps (%rsi), %xmm1
19494 ; NoVLX-NEXT: vcmpeqps %zmm1, %zmm0, %k0 {%k1}
19495 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
19496 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
19497 ; NoVLX-NEXT: kmovw %k0, %eax
19498 ; NoVLX-NEXT: # kill: def $al killed $al killed $eax
19499 ; NoVLX-NEXT: vzeroupper
19500 ; NoVLX-NEXT: retq
19501 entry:
19502 %0 = bitcast <2 x i64> %__a to <4 x float>
19503 %load = load <2 x i64>, <2 x i64>* %__b
19504 %1 = bitcast <2 x i64> %load to <4 x float>
19505 %2 = fcmp oeq <4 x float> %0, %1
19506 %3 = bitcast i4 %__u to <4 x i1>
19507 %4 = and <4 x i1> %2, %3
19508 %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
19509 %6 = bitcast <8 x i1> %5 to i8
19510 ret i8 %6
19511 }
19513 define zeroext i8 @test_masked_vcmpoeqps_v4i1_v8i1_mask_mem_b(i4 zeroext %__u, <2 x i64> %__a, float* %__b) local_unnamed_addr {
19514 ; VLX-LABEL: test_masked_vcmpoeqps_v4i1_v8i1_mask_mem_b:
19515 ; VLX: # %bb.0: # %entry
19516 ; VLX-NEXT: kmovd %edi, %k1
19517 ; VLX-NEXT: vcmpeqps (%rsi){1to4}, %xmm0, %k0 {%k1}
19518 ; VLX-NEXT: kmovd %k0, %eax
19519 ; VLX-NEXT: # kill: def $al killed $al killed $eax
19520 ; VLX-NEXT: retq
19521 ;
19522 ; NoVLX-LABEL: test_masked_vcmpoeqps_v4i1_v8i1_mask_mem_b:
19523 ; NoVLX: # %bb.0: # %entry
19524 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
19525 ; NoVLX-NEXT: kmovw %edi, %k1
19526 ; NoVLX-NEXT: vcmpeqps (%rsi){1to16}, %zmm0, %k0 {%k1}
19527 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
19528 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
19529 ; NoVLX-NEXT: kmovw %k0, %eax
19530 ; NoVLX-NEXT: # kill: def $al killed $al killed $eax
19531 ; NoVLX-NEXT: vzeroupper
19532 ; NoVLX-NEXT: retq
19533 entry:
19534 %0 = bitcast <2 x i64> %__a to <4 x float>
19535 %load = load float, float* %__b
19536 %vec = insertelement <4 x float> undef, float %load, i32 0
19537 %1 = shufflevector <4 x float> %vec, <4 x float> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
19538 %2 = fcmp oeq <4 x float> %0, %1
19539 %3 = bitcast i4 %__u to <4 x i1>
19540 %4 = and <4 x i1> %2, %3
19541 %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
19542 %6 = bitcast <8 x i1> %5 to i8
19543 ret i8 %6
19544 }
19548 define zeroext i16 @test_vcmpoeqps_v4i1_v16i1_mask(<2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
19549 ; VLX-LABEL: test_vcmpoeqps_v4i1_v16i1_mask:
19550 ; VLX: # %bb.0: # %entry
19551 ; VLX-NEXT: vcmpeqps %xmm1, %xmm0, %k0
19552 ; VLX-NEXT: kmovd %k0, %eax
19553 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
19554 ; VLX-NEXT: retq
19555 ;
19556 ; NoVLX-LABEL: test_vcmpoeqps_v4i1_v16i1_mask:
19557 ; NoVLX: # %bb.0: # %entry
19558 ; NoVLX-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
19559 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
19560 ; NoVLX-NEXT: vcmpeqps %zmm1, %zmm0, %k0
19561 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
19562 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
19563 ; NoVLX-NEXT: kmovw %k0, %eax
19564 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
19565 ; NoVLX-NEXT: vzeroupper
19566 ; NoVLX-NEXT: retq
19567 entry:
19568 %0 = bitcast <2 x i64> %__a to <4 x float>
19569 %1 = bitcast <2 x i64> %__b to <4 x float>
19570 %2 = fcmp oeq <4 x float> %0, %1
19571 %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
19572 %4 = bitcast <16 x i1> %3 to i16
19573 ret i16 %4
19574 }
19576 define zeroext i16 @test_vcmpoeqps_v4i1_v16i1_mask_mem(<2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
19577 ; VLX-LABEL: test_vcmpoeqps_v4i1_v16i1_mask_mem:
19578 ; VLX: # %bb.0: # %entry
19579 ; VLX-NEXT: vcmpeqps (%rdi), %xmm0, %k0
19580 ; VLX-NEXT: kmovd %k0, %eax
19581 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
19582 ; VLX-NEXT: retq
19583 ;
19584 ; NoVLX-LABEL: test_vcmpoeqps_v4i1_v16i1_mask_mem:
19585 ; NoVLX: # %bb.0: # %entry
19586 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
19587 ; NoVLX-NEXT: vmovaps (%rdi), %xmm1
19588 ; NoVLX-NEXT: vcmpeqps %zmm1, %zmm0, %k0
19589 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
19590 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
19591 ; NoVLX-NEXT: kmovw %k0, %eax
19592 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
19593 ; NoVLX-NEXT: vzeroupper
19594 ; NoVLX-NEXT: retq
19595 entry:
19596 %0 = bitcast <2 x i64> %__a to <4 x float>
19597 %load = load <2 x i64>, <2 x i64>* %__b
19598 %1 = bitcast <2 x i64> %load to <4 x float>
19599 %2 = fcmp oeq <4 x float> %0, %1
19600 %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
19601 %4 = bitcast <16 x i1> %3 to i16
19602 ret i16 %4
19603 }
19605 define zeroext i16 @test_vcmpoeqps_v4i1_v16i1_mask_mem_b(<2 x i64> %__a, float* %__b) local_unnamed_addr {
19606 ; VLX-LABEL: test_vcmpoeqps_v4i1_v16i1_mask_mem_b:
19607 ; VLX: # %bb.0: # %entry
19608 ; VLX-NEXT: vcmpeqps (%rdi){1to4}, %xmm0, %k0
19609 ; VLX-NEXT: kmovd %k0, %eax
19610 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
19611 ; VLX-NEXT: retq
19612 ;
19613 ; NoVLX-LABEL: test_vcmpoeqps_v4i1_v16i1_mask_mem_b:
19614 ; NoVLX: # %bb.0: # %entry
19615 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
19616 ; NoVLX-NEXT: vcmpeqps (%rdi){1to16}, %zmm0, %k0
19617 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
19618 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
19619 ; NoVLX-NEXT: kmovw %k0, %eax
19620 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
19621 ; NoVLX-NEXT: vzeroupper
19622 ; NoVLX-NEXT: retq
19623 entry:
19624 %0 = bitcast <2 x i64> %__a to <4 x float>
19625 %load = load float, float* %__b
19626 %vec = insertelement <4 x float> undef, float %load, i32 0
19627 %1 = shufflevector <4 x float> %vec, <4 x float> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
19628 %2 = fcmp oeq <4 x float> %0, %1
19629 %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
19630 %4 = bitcast <16 x i1> %3 to i16
19631 ret i16 %4
19632 }
19634 define zeroext i16 @test_masked_vcmpoeqps_v4i1_v16i1_mask(i4 zeroext %__u, <2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
19635 ; VLX-LABEL: test_masked_vcmpoeqps_v4i1_v16i1_mask:
19636 ; VLX: # %bb.0: # %entry
19637 ; VLX-NEXT: kmovd %edi, %k1
19638 ; VLX-NEXT: vcmpeqps %xmm1, %xmm0, %k0 {%k1}
19639 ; VLX-NEXT: kmovd %k0, %eax
19640 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
19641 ; VLX-NEXT: retq
19642 ;
19643 ; NoVLX-LABEL: test_masked_vcmpoeqps_v4i1_v16i1_mask:
19644 ; NoVLX: # %bb.0: # %entry
19645 ; NoVLX-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
19646 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
19647 ; NoVLX-NEXT: kmovw %edi, %k1
19648 ; NoVLX-NEXT: vcmpeqps %zmm1, %zmm0, %k0 {%k1}
19649 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
19650 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
19651 ; NoVLX-NEXT: kmovw %k0, %eax
19652 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
19653 ; NoVLX-NEXT: vzeroupper
19654 ; NoVLX-NEXT: retq
19655 entry:
19656 %0 = bitcast <2 x i64> %__a to <4 x float>
19657 %1 = bitcast <2 x i64> %__b to <4 x float>
19658 %2 = fcmp oeq <4 x float> %0, %1
19659 %3 = bitcast i4 %__u to <4 x i1>
19660 %4 = and <4 x i1> %2, %3
19661 %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
19662 %6 = bitcast <16 x i1> %5 to i16
19663 ret i16 %6
19664 }
19666 define zeroext i16 @test_masked_vcmpoeqps_v4i1_v16i1_mask_mem(i4 zeroext %__u, <2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
19667 ; VLX-LABEL: test_masked_vcmpoeqps_v4i1_v16i1_mask_mem:
19668 ; VLX: # %bb.0: # %entry
19669 ; VLX-NEXT: kmovd %edi, %k1
19670 ; VLX-NEXT: vcmpeqps (%rsi), %xmm0, %k0 {%k1}
19671 ; VLX-NEXT: kmovd %k0, %eax
19672 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
19673 ; VLX-NEXT: retq
19674 ;
19675 ; NoVLX-LABEL: test_masked_vcmpoeqps_v4i1_v16i1_mask_mem:
19676 ; NoVLX: # %bb.0: # %entry
19677 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
19678 ; NoVLX-NEXT: kmovw %edi, %k1
19679 ; NoVLX-NEXT: vmovaps (%rsi), %xmm1
19680 ; NoVLX-NEXT: vcmpeqps %zmm1, %zmm0, %k0 {%k1}
19681 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
19682 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
19683 ; NoVLX-NEXT: kmovw %k0, %eax
19684 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
19685 ; NoVLX-NEXT: vzeroupper
19686 ; NoVLX-NEXT: retq
19687 entry:
19688 %0 = bitcast <2 x i64> %__a to <4 x float>
19689 %load = load <2 x i64>, <2 x i64>* %__b
19690 %1 = bitcast <2 x i64> %load to <4 x float>
19691 %2 = fcmp oeq <4 x float> %0, %1
19692 %3 = bitcast i4 %__u to <4 x i1>
19693 %4 = and <4 x i1> %2, %3
19694 %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
19695 %6 = bitcast <16 x i1> %5 to i16
19696 ret i16 %6
19697 }
19699 define zeroext i16 @test_masked_vcmpoeqps_v4i1_v16i1_mask_mem_b(i4 zeroext %__u, <2 x i64> %__a, float* %__b) local_unnamed_addr {
19700 ; VLX-LABEL: test_masked_vcmpoeqps_v4i1_v16i1_mask_mem_b:
19701 ; VLX: # %bb.0: # %entry
19702 ; VLX-NEXT: kmovd %edi, %k1
19703 ; VLX-NEXT: vcmpeqps (%rsi){1to4}, %xmm0, %k0 {%k1}
19704 ; VLX-NEXT: kmovd %k0, %eax
19705 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
19706 ; VLX-NEXT: retq
19707 ;
19708 ; NoVLX-LABEL: test_masked_vcmpoeqps_v4i1_v16i1_mask_mem_b:
19709 ; NoVLX: # %bb.0: # %entry
19710 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
19711 ; NoVLX-NEXT: kmovw %edi, %k1
19712 ; NoVLX-NEXT: vcmpeqps (%rsi){1to16}, %zmm0, %k0 {%k1}
19713 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
19714 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
19715 ; NoVLX-NEXT: kmovw %k0, %eax
19716 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
19717 ; NoVLX-NEXT: vzeroupper
19718 ; NoVLX-NEXT: retq
19719 entry:
19720 %0 = bitcast <2 x i64> %__a to <4 x float>
19721 %load = load float, float* %__b
19722 %vec = insertelement <4 x float> undef, float %load, i32 0
19723 %1 = shufflevector <4 x float> %vec, <4 x float> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
19724 %2 = fcmp oeq <4 x float> %0, %1
19725 %3 = bitcast i4 %__u to <4 x i1>
19726 %4 = and <4 x i1> %2, %3
19727 %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
19728 %6 = bitcast <16 x i1> %5 to i16
19729 ret i16 %6
19730 }
19734 define zeroext i32 @test_vcmpoeqps_v4i1_v32i1_mask(<2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
19735 ; VLX-LABEL: test_vcmpoeqps_v4i1_v32i1_mask:
19736 ; VLX: # %bb.0: # %entry
19737 ; VLX-NEXT: vcmpeqps %xmm1, %xmm0, %k0
19738 ; VLX-NEXT: kmovd %k0, %eax
19739 ; VLX-NEXT: retq
19740 ;
19741 ; NoVLX-LABEL: test_vcmpoeqps_v4i1_v32i1_mask:
19742 ; NoVLX: # %bb.0: # %entry
19743 ; NoVLX-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
19744 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
19745 ; NoVLX-NEXT: vcmpeqps %zmm1, %zmm0, %k0
19746 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
19747 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
19748 ; NoVLX-NEXT: kmovw %k0, %eax
19749 ; NoVLX-NEXT: vzeroupper
19750 ; NoVLX-NEXT: retq
19751 entry:
19752 %0 = bitcast <2 x i64> %__a to <4 x float>
19753 %1 = bitcast <2 x i64> %__b to <4 x float>
19754 %2 = fcmp oeq <4 x float> %0, %1
19755 %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
19756 %4 = bitcast <32 x i1> %3 to i32
19757 ret i32 %4
19758 }
19760 define zeroext i32 @test_vcmpoeqps_v4i1_v32i1_mask_mem(<2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
19761 ; VLX-LABEL: test_vcmpoeqps_v4i1_v32i1_mask_mem:
19762 ; VLX: # %bb.0: # %entry
19763 ; VLX-NEXT: vcmpeqps (%rdi), %xmm0, %k0
19764 ; VLX-NEXT: kmovd %k0, %eax
19765 ; VLX-NEXT: retq
19766 ;
19767 ; NoVLX-LABEL: test_vcmpoeqps_v4i1_v32i1_mask_mem:
19768 ; NoVLX: # %bb.0: # %entry
19769 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
19770 ; NoVLX-NEXT: vmovaps (%rdi), %xmm1
19771 ; NoVLX-NEXT: vcmpeqps %zmm1, %zmm0, %k0
19772 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
19773 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
19774 ; NoVLX-NEXT: kmovw %k0, %eax
19775 ; NoVLX-NEXT: vzeroupper
19776 ; NoVLX-NEXT: retq
19777 entry:
19778 %0 = bitcast <2 x i64> %__a to <4 x float>
19779 %load = load <2 x i64>, <2 x i64>* %__b
19780 %1 = bitcast <2 x i64> %load to <4 x float>
19781 %2 = fcmp oeq <4 x float> %0, %1
19782 %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
19783 %4 = bitcast <32 x i1> %3 to i32
19784 ret i32 %4
19785 }
19787 define zeroext i32 @test_vcmpoeqps_v4i1_v32i1_mask_mem_b(<2 x i64> %__a, float* %__b) local_unnamed_addr {
19788 ; VLX-LABEL: test_vcmpoeqps_v4i1_v32i1_mask_mem_b:
19789 ; VLX: # %bb.0: # %entry
19790 ; VLX-NEXT: vcmpeqps (%rdi){1to4}, %xmm0, %k0
19791 ; VLX-NEXT: kmovd %k0, %eax
19792 ; VLX-NEXT: retq
19793 ;
19794 ; NoVLX-LABEL: test_vcmpoeqps_v4i1_v32i1_mask_mem_b:
19795 ; NoVLX: # %bb.0: # %entry
19796 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
19797 ; NoVLX-NEXT: vcmpeqps (%rdi){1to16}, %zmm0, %k0
19798 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
19799 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
19800 ; NoVLX-NEXT: kmovw %k0, %eax
19801 ; NoVLX-NEXT: vzeroupper
19802 ; NoVLX-NEXT: retq
19803 entry:
19804 %0 = bitcast <2 x i64> %__a to <4 x float>
19805 %load = load float, float* %__b
19806 %vec = insertelement <4 x float> undef, float %load, i32 0
19807 %1 = shufflevector <4 x float> %vec, <4 x float> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
19808 %2 = fcmp oeq <4 x float> %0, %1
19809 %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
19810 %4 = bitcast <32 x i1> %3 to i32
19811 ret i32 %4
19812 }
19814 define zeroext i32 @test_masked_vcmpoeqps_v4i1_v32i1_mask(i4 zeroext %__u, <2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
19815 ; VLX-LABEL: test_masked_vcmpoeqps_v4i1_v32i1_mask:
19816 ; VLX: # %bb.0: # %entry
19817 ; VLX-NEXT: kmovd %edi, %k1
19818 ; VLX-NEXT: vcmpeqps %xmm1, %xmm0, %k0 {%k1}
19819 ; VLX-NEXT: kmovd %k0, %eax
19820 ; VLX-NEXT: retq
19821 ;
19822 ; NoVLX-LABEL: test_masked_vcmpoeqps_v4i1_v32i1_mask:
19823 ; NoVLX: # %bb.0: # %entry
19824 ; NoVLX-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
19825 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
19826 ; NoVLX-NEXT: kmovw %edi, %k1
19827 ; NoVLX-NEXT: vcmpeqps %zmm1, %zmm0, %k0 {%k1}
19828 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
19829 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
19830 ; NoVLX-NEXT: kmovw %k0, %eax
19831 ; NoVLX-NEXT: vzeroupper
19832 ; NoVLX-NEXT: retq
19833 entry:
19834 %0 = bitcast <2 x i64> %__a to <4 x float>
19835 %1 = bitcast <2 x i64> %__b to <4 x float>
19836 %2 = fcmp oeq <4 x float> %0, %1
19837 %3 = bitcast i4 %__u to <4 x i1>
19838 %4 = and <4 x i1> %2, %3
19839 %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
19840 %6 = bitcast <32 x i1> %5 to i32
19841 ret i32 %6
19842 }
19844 define zeroext i32 @test_masked_vcmpoeqps_v4i1_v32i1_mask_mem(i4 zeroext %__u, <2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
19845 ; VLX-LABEL: test_masked_vcmpoeqps_v4i1_v32i1_mask_mem:
19846 ; VLX: # %bb.0: # %entry
19847 ; VLX-NEXT: kmovd %edi, %k1
19848 ; VLX-NEXT: vcmpeqps (%rsi), %xmm0, %k0 {%k1}
19849 ; VLX-NEXT: kmovd %k0, %eax
19850 ; VLX-NEXT: retq
19851 ;
19852 ; NoVLX-LABEL: test_masked_vcmpoeqps_v4i1_v32i1_mask_mem:
19853 ; NoVLX: # %bb.0: # %entry
19854 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
19855 ; NoVLX-NEXT: kmovw %edi, %k1
19856 ; NoVLX-NEXT: vmovaps (%rsi), %xmm1
19857 ; NoVLX-NEXT: vcmpeqps %zmm1, %zmm0, %k0 {%k1}
19858 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
19859 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
19860 ; NoVLX-NEXT: kmovw %k0, %eax
19861 ; NoVLX-NEXT: vzeroupper
19862 ; NoVLX-NEXT: retq
19863 entry:
19864 %0 = bitcast <2 x i64> %__a to <4 x float>
19865 %load = load <2 x i64>, <2 x i64>* %__b
19866 %1 = bitcast <2 x i64> %load to <4 x float>
19867 %2 = fcmp oeq <4 x float> %0, %1
19868 %3 = bitcast i4 %__u to <4 x i1>
19869 %4 = and <4 x i1> %2, %3
19870 %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
19871 %6 = bitcast <32 x i1> %5 to i32
19872 ret i32 %6
19873 }
19875 define zeroext i32 @test_masked_vcmpoeqps_v4i1_v32i1_mask_mem_b(i4 zeroext %__u, <2 x i64> %__a, float* %__b) local_unnamed_addr {
19876 ; VLX-LABEL: test_masked_vcmpoeqps_v4i1_v32i1_mask_mem_b:
19877 ; VLX: # %bb.0: # %entry
19878 ; VLX-NEXT: kmovd %edi, %k1
19879 ; VLX-NEXT: vcmpeqps (%rsi){1to4}, %xmm0, %k0 {%k1}
19880 ; VLX-NEXT: kmovd %k0, %eax
19881 ; VLX-NEXT: retq
19882 ;
19883 ; NoVLX-LABEL: test_masked_vcmpoeqps_v4i1_v32i1_mask_mem_b:
19884 ; NoVLX: # %bb.0: # %entry
19885 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
19886 ; NoVLX-NEXT: kmovw %edi, %k1
19887 ; NoVLX-NEXT: vcmpeqps (%rsi){1to16}, %zmm0, %k0 {%k1}
19888 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
19889 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
19890 ; NoVLX-NEXT: kmovw %k0, %eax
19891 ; NoVLX-NEXT: vzeroupper
19892 ; NoVLX-NEXT: retq
19893 entry:
19894 %0 = bitcast <2 x i64> %__a to <4 x float>
19895 %load = load float, float* %__b
19896 %vec = insertelement <4 x float> undef, float %load, i32 0
19897 %1 = shufflevector <4 x float> %vec, <4 x float> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
19898 %2 = fcmp oeq <4 x float> %0, %1
19899 %3 = bitcast i4 %__u to <4 x i1>
19900 %4 = and <4 x i1> %2, %3
19901 %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
19902 %6 = bitcast <32 x i1> %5 to i32
19903 ret i32 %6
19904 }
19908 define zeroext i64 @test_vcmpoeqps_v4i1_v64i1_mask(<2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
19909 ; VLX-LABEL: test_vcmpoeqps_v4i1_v64i1_mask:
19910 ; VLX: # %bb.0: # %entry
19911 ; VLX-NEXT: vcmpeqps %xmm1, %xmm0, %k0
19912 ; VLX-NEXT: kmovq %k0, %rax
19913 ; VLX-NEXT: retq
19914 ;
19915 ; NoVLX-LABEL: test_vcmpoeqps_v4i1_v64i1_mask:
19916 ; NoVLX: # %bb.0: # %entry
19917 ; NoVLX-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
19918 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
19919 ; NoVLX-NEXT: vcmpeqps %zmm1, %zmm0, %k0
19920 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
19921 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
19922 ; NoVLX-NEXT: kmovw %k0, %eax
19923 ; NoVLX-NEXT: vzeroupper
19924 ; NoVLX-NEXT: retq
19925 entry:
19926 %0 = bitcast <2 x i64> %__a to <4 x float>
19927 %1 = bitcast <2 x i64> %__b to <4 x float>
19928 %2 = fcmp oeq <4 x float> %0, %1
19929 %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
19930 %4 = bitcast <64 x i1> %3 to i64
19931 ret i64 %4
19932 }
19934 define zeroext i64 @test_vcmpoeqps_v4i1_v64i1_mask_mem(<2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
19935 ; VLX-LABEL: test_vcmpoeqps_v4i1_v64i1_mask_mem:
19936 ; VLX: # %bb.0: # %entry
19937 ; VLX-NEXT: vcmpeqps (%rdi), %xmm0, %k0
19938 ; VLX-NEXT: kmovq %k0, %rax
19939 ; VLX-NEXT: retq
19940 ;
19941 ; NoVLX-LABEL: test_vcmpoeqps_v4i1_v64i1_mask_mem:
19942 ; NoVLX: # %bb.0: # %entry
19943 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
19944 ; NoVLX-NEXT: vmovaps (%rdi), %xmm1
19945 ; NoVLX-NEXT: vcmpeqps %zmm1, %zmm0, %k0
19946 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
19947 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
19948 ; NoVLX-NEXT: kmovw %k0, %eax
19949 ; NoVLX-NEXT: vzeroupper
19950 ; NoVLX-NEXT: retq
19951 entry:
19952 %0 = bitcast <2 x i64> %__a to <4 x float>
19953 %load = load <2 x i64>, <2 x i64>* %__b
19954 %1 = bitcast <2 x i64> %load to <4 x float>
19955 %2 = fcmp oeq <4 x float> %0, %1
19956 %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
19957 %4 = bitcast <64 x i1> %3 to i64
19958 ret i64 %4
19959 }
19961 define zeroext i64 @test_vcmpoeqps_v4i1_v64i1_mask_mem_b(<2 x i64> %__a, float* %__b) local_unnamed_addr {
19962 ; VLX-LABEL: test_vcmpoeqps_v4i1_v64i1_mask_mem_b:
19963 ; VLX: # %bb.0: # %entry
19964 ; VLX-NEXT: vcmpeqps (%rdi){1to4}, %xmm0, %k0
19965 ; VLX-NEXT: kmovq %k0, %rax
19966 ; VLX-NEXT: retq
19967 ;
19968 ; NoVLX-LABEL: test_vcmpoeqps_v4i1_v64i1_mask_mem_b:
19969 ; NoVLX: # %bb.0: # %entry
19970 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
19971 ; NoVLX-NEXT: vcmpeqps (%rdi){1to16}, %zmm0, %k0
19972 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
19973 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
19974 ; NoVLX-NEXT: kmovw %k0, %eax
19975 ; NoVLX-NEXT: vzeroupper
19976 ; NoVLX-NEXT: retq
19977 entry:
19978 %0 = bitcast <2 x i64> %__a to <4 x float>
19979 %load = load float, float* %__b
19980 %vec = insertelement <4 x float> undef, float %load, i32 0
19981 %1 = shufflevector <4 x float> %vec, <4 x float> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
19982 %2 = fcmp oeq <4 x float> %0, %1
19983 %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
19984 %4 = bitcast <64 x i1> %3 to i64
19985 ret i64 %4
19986 }
19988 define zeroext i64 @test_masked_vcmpoeqps_v4i1_v64i1_mask(i4 zeroext %__u, <2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
19989 ; VLX-LABEL: test_masked_vcmpoeqps_v4i1_v64i1_mask:
19990 ; VLX: # %bb.0: # %entry
19991 ; VLX-NEXT: kmovd %edi, %k1
19992 ; VLX-NEXT: vcmpeqps %xmm1, %xmm0, %k0 {%k1}
19993 ; VLX-NEXT: kmovq %k0, %rax
19994 ; VLX-NEXT: retq
19995 ;
19996 ; NoVLX-LABEL: test_masked_vcmpoeqps_v4i1_v64i1_mask:
19997 ; NoVLX: # %bb.0: # %entry
19998 ; NoVLX-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
19999 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
20000 ; NoVLX-NEXT: kmovw %edi, %k1
20001 ; NoVLX-NEXT: vcmpeqps %zmm1, %zmm0, %k0 {%k1}
20002 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
20003 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
20004 ; NoVLX-NEXT: kmovw %k0, %eax
20005 ; NoVLX-NEXT: vzeroupper
20006 ; NoVLX-NEXT: retq
20007 entry:
20008 %0 = bitcast <2 x i64> %__a to <4 x float>
20009 %1 = bitcast <2 x i64> %__b to <4 x float>
20010 %2 = fcmp oeq <4 x float> %0, %1
20011 %3 = bitcast i4 %__u to <4 x i1>
20012 %4 = and <4 x i1> %2, %3
20013 %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
20014 %6 = bitcast <64 x i1> %5 to i64
20015 ret i64 %6
20016 }
20018 define zeroext i64 @test_masked_vcmpoeqps_v4i1_v64i1_mask_mem(i4 zeroext %__u, <2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
20019 ; VLX-LABEL: test_masked_vcmpoeqps_v4i1_v64i1_mask_mem:
20020 ; VLX: # %bb.0: # %entry
20021 ; VLX-NEXT: kmovd %edi, %k1
20022 ; VLX-NEXT: vcmpeqps (%rsi), %xmm0, %k0 {%k1}
20023 ; VLX-NEXT: kmovq %k0, %rax
20024 ; VLX-NEXT: retq
20025 ;
20026 ; NoVLX-LABEL: test_masked_vcmpoeqps_v4i1_v64i1_mask_mem:
20027 ; NoVLX: # %bb.0: # %entry
20028 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
20029 ; NoVLX-NEXT: kmovw %edi, %k1
20030 ; NoVLX-NEXT: vmovaps (%rsi), %xmm1
20031 ; NoVLX-NEXT: vcmpeqps %zmm1, %zmm0, %k0 {%k1}
20032 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
20033 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
20034 ; NoVLX-NEXT: kmovw %k0, %eax
20035 ; NoVLX-NEXT: vzeroupper
20036 ; NoVLX-NEXT: retq
20037 entry:
20038 %0 = bitcast <2 x i64> %__a to <4 x float>
20039 %load = load <2 x i64>, <2 x i64>* %__b
20040 %1 = bitcast <2 x i64> %load to <4 x float>
20041 %2 = fcmp oeq <4 x float> %0, %1
20042 %3 = bitcast i4 %__u to <4 x i1>
20043 %4 = and <4 x i1> %2, %3
20044 %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
20045 %6 = bitcast <64 x i1> %5 to i64
20046 ret i64 %6
20047 }
20049 define zeroext i64 @test_masked_vcmpoeqps_v4i1_v64i1_mask_mem_b(i4 zeroext %__u, <2 x i64> %__a, float* %__b) local_unnamed_addr {
20050 ; VLX-LABEL: test_masked_vcmpoeqps_v4i1_v64i1_mask_mem_b:
20051 ; VLX: # %bb.0: # %entry
20052 ; VLX-NEXT: kmovd %edi, %k1
20053 ; VLX-NEXT: vcmpeqps (%rsi){1to4}, %xmm0, %k0 {%k1}
20054 ; VLX-NEXT: kmovq %k0, %rax
20055 ; VLX-NEXT: retq
20056 ;
20057 ; NoVLX-LABEL: test_masked_vcmpoeqps_v4i1_v64i1_mask_mem_b:
20058 ; NoVLX: # %bb.0: # %entry
20059 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
20060 ; NoVLX-NEXT: kmovw %edi, %k1
20061 ; NoVLX-NEXT: vcmpeqps (%rsi){1to16}, %zmm0, %k0 {%k1}
20062 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
20063 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
20064 ; NoVLX-NEXT: kmovw %k0, %eax
20065 ; NoVLX-NEXT: vzeroupper
20066 ; NoVLX-NEXT: retq
20067 entry:
20068 %0 = bitcast <2 x i64> %__a to <4 x float>
20069 %load = load float, float* %__b
20070 %vec = insertelement <4 x float> undef, float %load, i32 0
20071 %1 = shufflevector <4 x float> %vec, <4 x float> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
20072 %2 = fcmp oeq <4 x float> %0, %1
20073 %3 = bitcast i4 %__u to <4 x i1>
20074 %4 = and <4 x i1> %2, %3
20075 %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
20076 %6 = bitcast <64 x i1> %5 to i64
20077 ret i64 %6
20078 }
20082 define zeroext i16 @test_vcmpoeqps_v8i1_v16i1_mask(<4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
20083 ; VLX-LABEL: test_vcmpoeqps_v8i1_v16i1_mask:
20084 ; VLX: # %bb.0: # %entry
20085 ; VLX-NEXT: vcmpeqps %ymm1, %ymm0, %k0
20086 ; VLX-NEXT: kmovd %k0, %eax
20087 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
20088 ; VLX-NEXT: vzeroupper
20089 ; VLX-NEXT: retq
20090 ;
20091 ; NoVLX-LABEL: test_vcmpoeqps_v8i1_v16i1_mask:
20092 ; NoVLX: # %bb.0: # %entry
20093 ; NoVLX-NEXT: # kill: def $ymm1 killed $ymm1 def $zmm1
20094 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
20095 ; NoVLX-NEXT: vcmpeqps %zmm1, %zmm0, %k0
20096 ; NoVLX-NEXT: kshiftlw $8, %k0, %k0
20097 ; NoVLX-NEXT: kshiftrw $8, %k0, %k0
20098 ; NoVLX-NEXT: kmovw %k0, %eax
20099 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
20100 ; NoVLX-NEXT: vzeroupper
20101 ; NoVLX-NEXT: retq
20102 entry:
20103 %0 = bitcast <4 x i64> %__a to <8 x float>
20104 %1 = bitcast <4 x i64> %__b to <8 x float>
20105 %2 = fcmp oeq <8 x float> %0, %1
20106 %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
20107 %4 = bitcast <16 x i1> %3 to i16
20108 ret i16 %4
20109 }
20111 define zeroext i16 @test_vcmpoeqps_v8i1_v16i1_mask_mem(<4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
20112 ; VLX-LABEL: test_vcmpoeqps_v8i1_v16i1_mask_mem:
20113 ; VLX: # %bb.0: # %entry
20114 ; VLX-NEXT: vcmpeqps (%rdi), %ymm0, %k0
20115 ; VLX-NEXT: kmovd %k0, %eax
20116 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
20117 ; VLX-NEXT: vzeroupper
20118 ; VLX-NEXT: retq
20119 ;
20120 ; NoVLX-LABEL: test_vcmpoeqps_v8i1_v16i1_mask_mem:
20121 ; NoVLX: # %bb.0: # %entry
20122 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
20123 ; NoVLX-NEXT: vmovaps (%rdi), %ymm1
20124 ; NoVLX-NEXT: vcmpeqps %zmm1, %zmm0, %k0
20125 ; NoVLX-NEXT: kshiftlw $8, %k0, %k0
20126 ; NoVLX-NEXT: kshiftrw $8, %k0, %k0
20127 ; NoVLX-NEXT: kmovw %k0, %eax
20128 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
20129 ; NoVLX-NEXT: vzeroupper
20130 ; NoVLX-NEXT: retq
20131 entry:
20132 %0 = bitcast <4 x i64> %__a to <8 x float>
20133 %load = load <4 x i64>, <4 x i64>* %__b
20134 %1 = bitcast <4 x i64> %load to <8 x float>
20135 %2 = fcmp oeq <8 x float> %0, %1
20136 %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
20137 %4 = bitcast <16 x i1> %3 to i16
20138 ret i16 %4
20139 }
20141 define zeroext i16 @test_vcmpoeqps_v8i1_v16i1_mask_mem_b(<4 x i64> %__a, float* %__b) local_unnamed_addr {
20142 ; VLX-LABEL: test_vcmpoeqps_v8i1_v16i1_mask_mem_b:
20143 ; VLX: # %bb.0: # %entry
20144 ; VLX-NEXT: vcmpeqps (%rdi){1to8}, %ymm0, %k0
20145 ; VLX-NEXT: kmovd %k0, %eax
20146 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
20147 ; VLX-NEXT: vzeroupper
20148 ; VLX-NEXT: retq
20149 ;
20150 ; NoVLX-LABEL: test_vcmpoeqps_v8i1_v16i1_mask_mem_b:
20151 ; NoVLX: # %bb.0: # %entry
20152 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
20153 ; NoVLX-NEXT: vcmpeqps (%rdi){1to16}, %zmm0, %k0
20154 ; NoVLX-NEXT: kshiftlw $8, %k0, %k0
20155 ; NoVLX-NEXT: kshiftrw $8, %k0, %k0
20156 ; NoVLX-NEXT: kmovw %k0, %eax
20157 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
20158 ; NoVLX-NEXT: vzeroupper
20159 ; NoVLX-NEXT: retq
20160 entry:
20161 %0 = bitcast <4 x i64> %__a to <8 x float>
20162 %load = load float, float* %__b
20163 %vec = insertelement <8 x float> undef, float %load, i32 0
20164 %1 = shufflevector <8 x float> %vec, <8 x float> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
20165 %2 = fcmp oeq <8 x float> %0, %1
20166 %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
20167 %4 = bitcast <16 x i1> %3 to i16
20168 ret i16 %4
20169 }
20171 define zeroext i16 @test_masked_vcmpoeqps_v8i1_v16i1_mask(i8 zeroext %__u, <4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
20172 ; VLX-LABEL: test_masked_vcmpoeqps_v8i1_v16i1_mask:
20173 ; VLX: # %bb.0: # %entry
20174 ; VLX-NEXT: kmovd %edi, %k1
20175 ; VLX-NEXT: vcmpeqps %ymm1, %ymm0, %k0 {%k1}
20176 ; VLX-NEXT: kmovd %k0, %eax
20177 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
20178 ; VLX-NEXT: vzeroupper
20179 ; VLX-NEXT: retq
20180 ;
20181 ; NoVLX-LABEL: test_masked_vcmpoeqps_v8i1_v16i1_mask:
20182 ; NoVLX: # %bb.0: # %entry
20183 ; NoVLX-NEXT: # kill: def $ymm1 killed $ymm1 def $zmm1
20184 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
20185 ; NoVLX-NEXT: kmovw %edi, %k1
20186 ; NoVLX-NEXT: vcmpeqps %zmm1, %zmm0, %k0 {%k1}
20187 ; NoVLX-NEXT: kshiftlw $8, %k0, %k0
20188 ; NoVLX-NEXT: kshiftrw $8, %k0, %k0
20189 ; NoVLX-NEXT: kmovw %k0, %eax
20190 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
20191 ; NoVLX-NEXT: vzeroupper
20192 ; NoVLX-NEXT: retq
20193 entry:
20194 %0 = bitcast <4 x i64> %__a to <8 x float>
20195 %1 = bitcast <4 x i64> %__b to <8 x float>
20196 %2 = fcmp oeq <8 x float> %0, %1
20197 %3 = bitcast i8 %__u to <8 x i1>
20198 %4 = and <8 x i1> %2, %3
20199 %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
20200 %6 = bitcast <16 x i1> %5 to i16
20201 ret i16 %6
20202 }
20204 define zeroext i16 @test_masked_vcmpoeqps_v8i1_v16i1_mask_mem(i8 zeroext %__u, <4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
20205 ; VLX-LABEL: test_masked_vcmpoeqps_v8i1_v16i1_mask_mem:
20206 ; VLX: # %bb.0: # %entry
20207 ; VLX-NEXT: kmovd %edi, %k1
20208 ; VLX-NEXT: vcmpeqps (%rsi), %ymm0, %k0 {%k1}
20209 ; VLX-NEXT: kmovd %k0, %eax
20210 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
20211 ; VLX-NEXT: vzeroupper
20212 ; VLX-NEXT: retq
20213 ;
20214 ; NoVLX-LABEL: test_masked_vcmpoeqps_v8i1_v16i1_mask_mem:
20215 ; NoVLX: # %bb.0: # %entry
20216 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
20217 ; NoVLX-NEXT: vmovaps (%rsi), %ymm1
20218 ; NoVLX-NEXT: kmovw %edi, %k1
20219 ; NoVLX-NEXT: vcmpeqps %zmm1, %zmm0, %k0 {%k1}
20220 ; NoVLX-NEXT: kshiftlw $8, %k0, %k0
20221 ; NoVLX-NEXT: kshiftrw $8, %k0, %k0
20222 ; NoVLX-NEXT: kmovw %k0, %eax
20223 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
20224 ; NoVLX-NEXT: vzeroupper
20225 ; NoVLX-NEXT: retq
20226 entry:
20227 %0 = bitcast <4 x i64> %__a to <8 x float>
20228 %load = load <4 x i64>, <4 x i64>* %__b
20229 %1 = bitcast <4 x i64> %load to <8 x float>
20230 %2 = fcmp oeq <8 x float> %0, %1
20231 %3 = bitcast i8 %__u to <8 x i1>
20232 %4 = and <8 x i1> %2, %3
20233 %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
20234 %6 = bitcast <16 x i1> %5 to i16
20235 ret i16 %6
20236 }
20238 define zeroext i16 @test_masked_vcmpoeqps_v8i1_v16i1_mask_mem_b(i8 zeroext %__u, <4 x i64> %__a, float* %__b) local_unnamed_addr {
20239 ; VLX-LABEL: test_masked_vcmpoeqps_v8i1_v16i1_mask_mem_b:
20240 ; VLX: # %bb.0: # %entry
20241 ; VLX-NEXT: kmovd %edi, %k1
20242 ; VLX-NEXT: vcmpeqps (%rsi){1to8}, %ymm0, %k0 {%k1}
20243 ; VLX-NEXT: kmovd %k0, %eax
20244 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
20245 ; VLX-NEXT: vzeroupper
20246 ; VLX-NEXT: retq
20247 ;
20248 ; NoVLX-LABEL: test_masked_vcmpoeqps_v8i1_v16i1_mask_mem_b:
20249 ; NoVLX: # %bb.0: # %entry
20250 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
20251 ; NoVLX-NEXT: kmovw %edi, %k1
20252 ; NoVLX-NEXT: vcmpeqps (%rsi){1to16}, %zmm0, %k0 {%k1}
20253 ; NoVLX-NEXT: kshiftlw $8, %k0, %k0
20254 ; NoVLX-NEXT: kshiftrw $8, %k0, %k0
20255 ; NoVLX-NEXT: kmovw %k0, %eax
20256 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
20257 ; NoVLX-NEXT: vzeroupper
20258 ; NoVLX-NEXT: retq
20259 entry:
20260 %0 = bitcast <4 x i64> %__a to <8 x float>
20261 %load = load float, float* %__b
20262 %vec = insertelement <8 x float> undef, float %load, i32 0
20263 %1 = shufflevector <8 x float> %vec, <8 x float> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
20264 %2 = fcmp oeq <8 x float> %0, %1
20265 %3 = bitcast i8 %__u to <8 x i1>
20266 %4 = and <8 x i1> %2, %3
20267 %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
20268 %6 = bitcast <16 x i1> %5 to i16
20269 ret i16 %6
20270 }
20274 define zeroext i32 @test_vcmpoeqps_v8i1_v32i1_mask(<4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
20275 ; VLX-LABEL: test_vcmpoeqps_v8i1_v32i1_mask:
20276 ; VLX: # %bb.0: # %entry
20277 ; VLX-NEXT: vcmpeqps %ymm1, %ymm0, %k0
20278 ; VLX-NEXT: kmovd %k0, %eax
20279 ; VLX-NEXT: vzeroupper
20280 ; VLX-NEXT: retq
20281 ;
20282 ; NoVLX-LABEL: test_vcmpoeqps_v8i1_v32i1_mask:
20283 ; NoVLX: # %bb.0: # %entry
20284 ; NoVLX-NEXT: # kill: def $ymm1 killed $ymm1 def $zmm1
20285 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
20286 ; NoVLX-NEXT: vcmpeqps %zmm1, %zmm0, %k0
20287 ; NoVLX-NEXT: kshiftlw $8, %k0, %k0
20288 ; NoVLX-NEXT: kshiftrw $8, %k0, %k0
20289 ; NoVLX-NEXT: kmovw %k0, %eax
20290 ; NoVLX-NEXT: vzeroupper
20291 ; NoVLX-NEXT: retq
20292 entry:
20293 %0 = bitcast <4 x i64> %__a to <8 x float>
20294 %1 = bitcast <4 x i64> %__b to <8 x float>
20295 %2 = fcmp oeq <8 x float> %0, %1
20296 %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
20297 %4 = bitcast <32 x i1> %3 to i32
20298 ret i32 %4
20299 }
20301 define zeroext i32 @test_vcmpoeqps_v8i1_v32i1_mask_mem(<4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
20302 ; VLX-LABEL: test_vcmpoeqps_v8i1_v32i1_mask_mem:
20303 ; VLX: # %bb.0: # %entry
20304 ; VLX-NEXT: vcmpeqps (%rdi), %ymm0, %k0
20305 ; VLX-NEXT: kmovd %k0, %eax
20306 ; VLX-NEXT: vzeroupper
20307 ; VLX-NEXT: retq
20308 ;
20309 ; NoVLX-LABEL: test_vcmpoeqps_v8i1_v32i1_mask_mem:
20310 ; NoVLX: # %bb.0: # %entry
20311 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
20312 ; NoVLX-NEXT: vmovaps (%rdi), %ymm1
20313 ; NoVLX-NEXT: vcmpeqps %zmm1, %zmm0, %k0
20314 ; NoVLX-NEXT: kshiftlw $8, %k0, %k0
20315 ; NoVLX-NEXT: kshiftrw $8, %k0, %k0
20316 ; NoVLX-NEXT: kmovw %k0, %eax
20317 ; NoVLX-NEXT: vzeroupper
20318 ; NoVLX-NEXT: retq
20319 entry:
20320 %0 = bitcast <4 x i64> %__a to <8 x float>
20321 %load = load <4 x i64>, <4 x i64>* %__b
20322 %1 = bitcast <4 x i64> %load to <8 x float>
20323 %2 = fcmp oeq <8 x float> %0, %1
20324 %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
20325 %4 = bitcast <32 x i1> %3 to i32
20326 ret i32 %4
20327 }
20329 define zeroext i32 @test_vcmpoeqps_v8i1_v32i1_mask_mem_b(<4 x i64> %__a, float* %__b) local_unnamed_addr {
20330 ; VLX-LABEL: test_vcmpoeqps_v8i1_v32i1_mask_mem_b:
20331 ; VLX: # %bb.0: # %entry
20332 ; VLX-NEXT: vcmpeqps (%rdi){1to8}, %ymm0, %k0
20333 ; VLX-NEXT: kmovd %k0, %eax
20334 ; VLX-NEXT: vzeroupper
20335 ; VLX-NEXT: retq
20336 ;
20337 ; NoVLX-LABEL: test_vcmpoeqps_v8i1_v32i1_mask_mem_b:
20338 ; NoVLX: # %bb.0: # %entry
20339 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
20340 ; NoVLX-NEXT: vcmpeqps (%rdi){1to16}, %zmm0, %k0
20341 ; NoVLX-NEXT: kshiftlw $8, %k0, %k0
20342 ; NoVLX-NEXT: kshiftrw $8, %k0, %k0
20343 ; NoVLX-NEXT: kmovw %k0, %eax
20344 ; NoVLX-NEXT: vzeroupper
20345 ; NoVLX-NEXT: retq
20346 entry:
20347 %0 = bitcast <4 x i64> %__a to <8 x float>
20348 %load = load float, float* %__b
20349 %vec = insertelement <8 x float> undef, float %load, i32 0
20350 %1 = shufflevector <8 x float> %vec, <8 x float> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
20351 %2 = fcmp oeq <8 x float> %0, %1
20352 %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
20353 %4 = bitcast <32 x i1> %3 to i32
20354 ret i32 %4
20355 }
20357 define zeroext i32 @test_masked_vcmpoeqps_v8i1_v32i1_mask(i8 zeroext %__u, <4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
20358 ; VLX-LABEL: test_masked_vcmpoeqps_v8i1_v32i1_mask:
20359 ; VLX: # %bb.0: # %entry
20360 ; VLX-NEXT: kmovd %edi, %k1
20361 ; VLX-NEXT: vcmpeqps %ymm1, %ymm0, %k0 {%k1}
20362 ; VLX-NEXT: kmovd %k0, %eax
20363 ; VLX-NEXT: vzeroupper
20364 ; VLX-NEXT: retq
20365 ;
20366 ; NoVLX-LABEL: test_masked_vcmpoeqps_v8i1_v32i1_mask:
20367 ; NoVLX: # %bb.0: # %entry
20368 ; NoVLX-NEXT: # kill: def $ymm1 killed $ymm1 def $zmm1
20369 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
20370 ; NoVLX-NEXT: kmovw %edi, %k1
20371 ; NoVLX-NEXT: vcmpeqps %zmm1, %zmm0, %k0 {%k1}
20372 ; NoVLX-NEXT: kshiftlw $8, %k0, %k0
20373 ; NoVLX-NEXT: kshiftrw $8, %k0, %k0
20374 ; NoVLX-NEXT: kmovw %k0, %eax
20375 ; NoVLX-NEXT: vzeroupper
20376 ; NoVLX-NEXT: retq
20377 entry:
20378 %0 = bitcast <4 x i64> %__a to <8 x float>
20379 %1 = bitcast <4 x i64> %__b to <8 x float>
20380 %2 = fcmp oeq <8 x float> %0, %1
20381 %3 = bitcast i8 %__u to <8 x i1>
20382 %4 = and <8 x i1> %2, %3
20383 %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
20384 %6 = bitcast <32 x i1> %5 to i32
20385 ret i32 %6
20386 }
20388 define zeroext i32 @test_masked_vcmpoeqps_v8i1_v32i1_mask_mem(i8 zeroext %__u, <4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
20389 ; VLX-LABEL: test_masked_vcmpoeqps_v8i1_v32i1_mask_mem:
20390 ; VLX: # %bb.0: # %entry
20391 ; VLX-NEXT: kmovd %edi, %k1
20392 ; VLX-NEXT: vcmpeqps (%rsi), %ymm0, %k0 {%k1}
20393 ; VLX-NEXT: kmovd %k0, %eax
20394 ; VLX-NEXT: vzeroupper
20395 ; VLX-NEXT: retq
20396 ;
20397 ; NoVLX-LABEL: test_masked_vcmpoeqps_v8i1_v32i1_mask_mem:
20398 ; NoVLX: # %bb.0: # %entry
20399 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
20400 ; NoVLX-NEXT: vmovaps (%rsi), %ymm1
20401 ; NoVLX-NEXT: kmovw %edi, %k1
20402 ; NoVLX-NEXT: vcmpeqps %zmm1, %zmm0, %k0 {%k1}
20403 ; NoVLX-NEXT: kshiftlw $8, %k0, %k0
20404 ; NoVLX-NEXT: kshiftrw $8, %k0, %k0
20405 ; NoVLX-NEXT: kmovw %k0, %eax
20406 ; NoVLX-NEXT: vzeroupper
20407 ; NoVLX-NEXT: retq
20408 entry:
20409 %0 = bitcast <4 x i64> %__a to <8 x float>
20410 %load = load <4 x i64>, <4 x i64>* %__b
20411 %1 = bitcast <4 x i64> %load to <8 x float>
20412 %2 = fcmp oeq <8 x float> %0, %1
20413 %3 = bitcast i8 %__u to <8 x i1>
20414 %4 = and <8 x i1> %2, %3
20415 %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
20416 %6 = bitcast <32 x i1> %5 to i32
20417 ret i32 %6
20418 }
20420 define zeroext i32 @test_masked_vcmpoeqps_v8i1_v32i1_mask_mem_b(i8 zeroext %__u, <4 x i64> %__a, float* %__b) local_unnamed_addr {
20421 ; VLX-LABEL: test_masked_vcmpoeqps_v8i1_v32i1_mask_mem_b:
20422 ; VLX: # %bb.0: # %entry
20423 ; VLX-NEXT: kmovd %edi, %k1
20424 ; VLX-NEXT: vcmpeqps (%rsi){1to8}, %ymm0, %k0 {%k1}
20425 ; VLX-NEXT: kmovd %k0, %eax
20426 ; VLX-NEXT: vzeroupper
20427 ; VLX-NEXT: retq
20428 ;
20429 ; NoVLX-LABEL: test_masked_vcmpoeqps_v8i1_v32i1_mask_mem_b:
20430 ; NoVLX: # %bb.0: # %entry
20431 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
20432 ; NoVLX-NEXT: kmovw %edi, %k1
20433 ; NoVLX-NEXT: vcmpeqps (%rsi){1to16}, %zmm0, %k0 {%k1}
20434 ; NoVLX-NEXT: kshiftlw $8, %k0, %k0
20435 ; NoVLX-NEXT: kshiftrw $8, %k0, %k0
20436 ; NoVLX-NEXT: kmovw %k0, %eax
20437 ; NoVLX-NEXT: vzeroupper
20438 ; NoVLX-NEXT: retq
20439 entry:
20440 %0 = bitcast <4 x i64> %__a to <8 x float>
20441 %load = load float, float* %__b
20442 %vec = insertelement <8 x float> undef, float %load, i32 0
20443 %1 = shufflevector <8 x float> %vec, <8 x float> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
20444 %2 = fcmp oeq <8 x float> %0, %1
20445 %3 = bitcast i8 %__u to <8 x i1>
20446 %4 = and <8 x i1> %2, %3
20447 %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
20448 %6 = bitcast <32 x i1> %5 to i32
20449 ret i32 %6
20450 }
20454 define zeroext i64 @test_vcmpoeqps_v8i1_v64i1_mask(<4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
20455 ; VLX-LABEL: test_vcmpoeqps_v8i1_v64i1_mask:
20456 ; VLX: # %bb.0: # %entry
20457 ; VLX-NEXT: vcmpeqps %ymm1, %ymm0, %k0
20458 ; VLX-NEXT: kmovq %k0, %rax
20459 ; VLX-NEXT: vzeroupper
20460 ; VLX-NEXT: retq
20461 ;
20462 ; NoVLX-LABEL: test_vcmpoeqps_v8i1_v64i1_mask:
20463 ; NoVLX: # %bb.0: # %entry
20464 ; NoVLX-NEXT: # kill: def $ymm1 killed $ymm1 def $zmm1
20465 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
20466 ; NoVLX-NEXT: vcmpeqps %zmm1, %zmm0, %k0
20467 ; NoVLX-NEXT: kshiftlw $8, %k0, %k0
20468 ; NoVLX-NEXT: kshiftrw $8, %k0, %k0
20469 ; NoVLX-NEXT: kmovw %k0, %eax
20470 ; NoVLX-NEXT: vzeroupper
20471 ; NoVLX-NEXT: retq
20472 entry:
20473 %0 = bitcast <4 x i64> %__a to <8 x float>
20474 %1 = bitcast <4 x i64> %__b to <8 x float>
20475 %2 = fcmp oeq <8 x float> %0, %1
20476 %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
20477 %4 = bitcast <64 x i1> %3 to i64
20478 ret i64 %4
20479 }
20481 define zeroext i64 @test_vcmpoeqps_v8i1_v64i1_mask_mem(<4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
20482 ; VLX-LABEL: test_vcmpoeqps_v8i1_v64i1_mask_mem:
20483 ; VLX: # %bb.0: # %entry
20484 ; VLX-NEXT: vcmpeqps (%rdi), %ymm0, %k0
20485 ; VLX-NEXT: kmovq %k0, %rax
20486 ; VLX-NEXT: vzeroupper
20487 ; VLX-NEXT: retq
20488 ;
20489 ; NoVLX-LABEL: test_vcmpoeqps_v8i1_v64i1_mask_mem:
20490 ; NoVLX: # %bb.0: # %entry
20491 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
20492 ; NoVLX-NEXT: vmovaps (%rdi), %ymm1
20493 ; NoVLX-NEXT: vcmpeqps %zmm1, %zmm0, %k0
20494 ; NoVLX-NEXT: kshiftlw $8, %k0, %k0
20495 ; NoVLX-NEXT: kshiftrw $8, %k0, %k0
20496 ; NoVLX-NEXT: kmovw %k0, %eax
20497 ; NoVLX-NEXT: vzeroupper
20498 ; NoVLX-NEXT: retq
20499 entry:
20500 %0 = bitcast <4 x i64> %__a to <8 x float>
20501 %load = load <4 x i64>, <4 x i64>* %__b
20502 %1 = bitcast <4 x i64> %load to <8 x float>
20503 %2 = fcmp oeq <8 x float> %0, %1
20504 %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
20505 %4 = bitcast <64 x i1> %3 to i64
20506 ret i64 %4
20507 }
20509 define zeroext i64 @test_vcmpoeqps_v8i1_v64i1_mask_mem_b(<4 x i64> %__a, float* %__b) local_unnamed_addr {
20510 ; VLX-LABEL: test_vcmpoeqps_v8i1_v64i1_mask_mem_b:
20511 ; VLX: # %bb.0: # %entry
20512 ; VLX-NEXT: vcmpeqps (%rdi){1to8}, %ymm0, %k0
20513 ; VLX-NEXT: kmovq %k0, %rax
20514 ; VLX-NEXT: vzeroupper
20515 ; VLX-NEXT: retq
20516 ;
20517 ; NoVLX-LABEL: test_vcmpoeqps_v8i1_v64i1_mask_mem_b:
20518 ; NoVLX: # %bb.0: # %entry
20519 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
20520 ; NoVLX-NEXT: vcmpeqps (%rdi){1to16}, %zmm0, %k0
20521 ; NoVLX-NEXT: kshiftlw $8, %k0, %k0
20522 ; NoVLX-NEXT: kshiftrw $8, %k0, %k0
20523 ; NoVLX-NEXT: kmovw %k0, %eax
20524 ; NoVLX-NEXT: vzeroupper
20525 ; NoVLX-NEXT: retq
20526 entry:
20527 %0 = bitcast <4 x i64> %__a to <8 x float>
20528 %load = load float, float* %__b
20529 %vec = insertelement <8 x float> undef, float %load, i32 0
20530 %1 = shufflevector <8 x float> %vec, <8 x float> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
20531 %2 = fcmp oeq <8 x float> %0, %1
20532 %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
20533 %4 = bitcast <64 x i1> %3 to i64
20534 ret i64 %4
20535 }
20537 define zeroext i64 @test_masked_vcmpoeqps_v8i1_v64i1_mask(i8 zeroext %__u, <4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
20538 ; VLX-LABEL: test_masked_vcmpoeqps_v8i1_v64i1_mask:
20539 ; VLX: # %bb.0: # %entry
20540 ; VLX-NEXT: kmovd %edi, %k1
20541 ; VLX-NEXT: vcmpeqps %ymm1, %ymm0, %k0 {%k1}
20542 ; VLX-NEXT: kmovq %k0, %rax
20543 ; VLX-NEXT: vzeroupper
20544 ; VLX-NEXT: retq
20545 ;
20546 ; NoVLX-LABEL: test_masked_vcmpoeqps_v8i1_v64i1_mask:
20547 ; NoVLX: # %bb.0: # %entry
20548 ; NoVLX-NEXT: # kill: def $ymm1 killed $ymm1 def $zmm1
20549 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
20550 ; NoVLX-NEXT: kmovw %edi, %k1
20551 ; NoVLX-NEXT: vcmpeqps %zmm1, %zmm0, %k0 {%k1}
20552 ; NoVLX-NEXT: kshiftlw $8, %k0, %k0
20553 ; NoVLX-NEXT: kshiftrw $8, %k0, %k0
20554 ; NoVLX-NEXT: kmovw %k0, %eax
20555 ; NoVLX-NEXT: vzeroupper
20556 ; NoVLX-NEXT: retq
20557 entry:
20558 %0 = bitcast <4 x i64> %__a to <8 x float>
20559 %1 = bitcast <4 x i64> %__b to <8 x float>
20560 %2 = fcmp oeq <8 x float> %0, %1
20561 %3 = bitcast i8 %__u to <8 x i1>
20562 %4 = and <8 x i1> %2, %3
20563 %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
20564 %6 = bitcast <64 x i1> %5 to i64
20565 ret i64 %6
20566 }
20568 define zeroext i64 @test_masked_vcmpoeqps_v8i1_v64i1_mask_mem(i8 zeroext %__u, <4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
20569 ; VLX-LABEL: test_masked_vcmpoeqps_v8i1_v64i1_mask_mem:
20570 ; VLX: # %bb.0: # %entry
20571 ; VLX-NEXT: kmovd %edi, %k1
20572 ; VLX-NEXT: vcmpeqps (%rsi), %ymm0, %k0 {%k1}
20573 ; VLX-NEXT: kmovq %k0, %rax
20574 ; VLX-NEXT: vzeroupper
20575 ; VLX-NEXT: retq
20576 ;
20577 ; NoVLX-LABEL: test_masked_vcmpoeqps_v8i1_v64i1_mask_mem:
20578 ; NoVLX: # %bb.0: # %entry
20579 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
20580 ; NoVLX-NEXT: vmovaps (%rsi), %ymm1
20581 ; NoVLX-NEXT: kmovw %edi, %k1
20582 ; NoVLX-NEXT: vcmpeqps %zmm1, %zmm0, %k0 {%k1}
20583 ; NoVLX-NEXT: kshiftlw $8, %k0, %k0
20584 ; NoVLX-NEXT: kshiftrw $8, %k0, %k0
20585 ; NoVLX-NEXT: kmovw %k0, %eax
20586 ; NoVLX-NEXT: vzeroupper
20587 ; NoVLX-NEXT: retq
20588 entry:
20589 %0 = bitcast <4 x i64> %__a to <8 x float>
20590 %load = load <4 x i64>, <4 x i64>* %__b
20591 %1 = bitcast <4 x i64> %load to <8 x float>
20592 %2 = fcmp oeq <8 x float> %0, %1
20593 %3 = bitcast i8 %__u to <8 x i1>
20594 %4 = and <8 x i1> %2, %3
20595 %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
20596 %6 = bitcast <64 x i1> %5 to i64
20597 ret i64 %6
20598 }
20600 define zeroext i64 @test_masked_vcmpoeqps_v8i1_v64i1_mask_mem_b(i8 zeroext %__u, <4 x i64> %__a, float* %__b) local_unnamed_addr {
20601 ; VLX-LABEL: test_masked_vcmpoeqps_v8i1_v64i1_mask_mem_b:
20602 ; VLX: # %bb.0: # %entry
20603 ; VLX-NEXT: kmovd %edi, %k1
20604 ; VLX-NEXT: vcmpeqps (%rsi){1to8}, %ymm0, %k0 {%k1}
20605 ; VLX-NEXT: kmovq %k0, %rax
20606 ; VLX-NEXT: vzeroupper
20607 ; VLX-NEXT: retq
20608 ;
20609 ; NoVLX-LABEL: test_masked_vcmpoeqps_v8i1_v64i1_mask_mem_b:
20610 ; NoVLX: # %bb.0: # %entry
20611 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
20612 ; NoVLX-NEXT: kmovw %edi, %k1
20613 ; NoVLX-NEXT: vcmpeqps (%rsi){1to16}, %zmm0, %k0 {%k1}
20614 ; NoVLX-NEXT: kshiftlw $8, %k0, %k0
20615 ; NoVLX-NEXT: kshiftrw $8, %k0, %k0
20616 ; NoVLX-NEXT: kmovw %k0, %eax
20617 ; NoVLX-NEXT: vzeroupper
20618 ; NoVLX-NEXT: retq
20619 entry:
20620 %0 = bitcast <4 x i64> %__a to <8 x float>
20621 %load = load float, float* %__b
20622 %vec = insertelement <8 x float> undef, float %load, i32 0
20623 %1 = shufflevector <8 x float> %vec, <8 x float> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
20624 %2 = fcmp oeq <8 x float> %0, %1
20625 %3 = bitcast i8 %__u to <8 x i1>
20626 %4 = and <8 x i1> %2, %3
20627 %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
20628 %6 = bitcast <64 x i1> %5 to i64
20629 ret i64 %6
20630 }
20634 define zeroext i32 @test_vcmpoeqps_v16i1_v32i1_mask(<8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
20635 ; VLX-LABEL: test_vcmpoeqps_v16i1_v32i1_mask:
20636 ; VLX: # %bb.0: # %entry
20637 ; VLX-NEXT: vcmpeqps %zmm1, %zmm0, %k0
20638 ; VLX-NEXT: kmovd %k0, %eax
20639 ; VLX-NEXT: vzeroupper
20640 ; VLX-NEXT: retq
20641 ;
20642 ; NoVLX-LABEL: test_vcmpoeqps_v16i1_v32i1_mask:
20643 ; NoVLX: # %bb.0: # %entry
20644 ; NoVLX-NEXT: vcmpeqps %zmm1, %zmm0, %k0
20645 ; NoVLX-NEXT: kmovw %k0, %eax
20646 ; NoVLX-NEXT: vzeroupper
20647 ; NoVLX-NEXT: retq
20648 entry:
20649 %0 = bitcast <8 x i64> %__a to <16 x float>
20650 %1 = bitcast <8 x i64> %__b to <16 x float>
20651 %2 = fcmp oeq <16 x float> %0, %1
20652 %3 = shufflevector <16 x i1> %2, <16 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>
20653 %4 = bitcast <32 x i1> %3 to i32
20654 ret i32 %4
20655 }
20657 define zeroext i32 @test_vcmpoeqps_v16i1_v32i1_mask_mem(<8 x i64> %__a, <8 x i64>* %__b) local_unnamed_addr {
20658 ; VLX-LABEL: test_vcmpoeqps_v16i1_v32i1_mask_mem:
20659 ; VLX: # %bb.0: # %entry
20660 ; VLX-NEXT: vcmpeqps (%rdi), %zmm0, %k0
20661 ; VLX-NEXT: kmovd %k0, %eax
20662 ; VLX-NEXT: vzeroupper
20663 ; VLX-NEXT: retq
20664 ;
20665 ; NoVLX-LABEL: test_vcmpoeqps_v16i1_v32i1_mask_mem:
20666 ; NoVLX: # %bb.0: # %entry
20667 ; NoVLX-NEXT: vcmpeqps (%rdi), %zmm0, %k0
20668 ; NoVLX-NEXT: kmovw %k0, %eax
20669 ; NoVLX-NEXT: vzeroupper
20670 ; NoVLX-NEXT: retq
20671 entry:
20672 %0 = bitcast <8 x i64> %__a to <16 x float>
20673 %load = load <8 x i64>, <8 x i64>* %__b
20674 %1 = bitcast <8 x i64> %load to <16 x float>
20675 %2 = fcmp oeq <16 x float> %0, %1
20676 %3 = shufflevector <16 x i1> %2, <16 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>
20677 %4 = bitcast <32 x i1> %3 to i32
20678 ret i32 %4
20679 }
20681 define zeroext i32 @test_vcmpoeqps_v16i1_v32i1_mask_mem_b(<8 x i64> %__a, float* %__b) local_unnamed_addr {
20682 ; VLX-LABEL: test_vcmpoeqps_v16i1_v32i1_mask_mem_b:
20683 ; VLX: # %bb.0: # %entry
20684 ; VLX-NEXT: vcmpeqps (%rdi){1to16}, %zmm0, %k0
20685 ; VLX-NEXT: kmovd %k0, %eax
20686 ; VLX-NEXT: vzeroupper
20687 ; VLX-NEXT: retq
20688 ;
20689 ; NoVLX-LABEL: test_vcmpoeqps_v16i1_v32i1_mask_mem_b:
20690 ; NoVLX: # %bb.0: # %entry
20691 ; NoVLX-NEXT: vcmpeqps (%rdi){1to16}, %zmm0, %k0
20692 ; NoVLX-NEXT: kmovw %k0, %eax
20693 ; NoVLX-NEXT: vzeroupper
20694 ; NoVLX-NEXT: retq
20695 entry:
20696 %0 = bitcast <8 x i64> %__a to <16 x float>
20697 %load = load float, float* %__b
20698 %vec = insertelement <16 x float> undef, float %load, i32 0
20699 %1 = shufflevector <16 x float> %vec, <16 x float> undef, <16 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
20700 %2 = fcmp oeq <16 x float> %0, %1
20701 %3 = shufflevector <16 x i1> %2, <16 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>
20702 %4 = bitcast <32 x i1> %3 to i32
20703 ret i32 %4
20704 }
20706 define zeroext i32 @test_masked_vcmpoeqps_v16i1_v32i1_mask(i16 zeroext %__u, <8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
20707 ; VLX-LABEL: test_masked_vcmpoeqps_v16i1_v32i1_mask:
20708 ; VLX: # %bb.0: # %entry
20709 ; VLX-NEXT: kmovd %edi, %k1
20710 ; VLX-NEXT: vcmpeqps %zmm1, %zmm0, %k0 {%k1}
20711 ; VLX-NEXT: kmovd %k0, %eax
20712 ; VLX-NEXT: vzeroupper
20713 ; VLX-NEXT: retq
20714 ;
20715 ; NoVLX-LABEL: test_masked_vcmpoeqps_v16i1_v32i1_mask:
20716 ; NoVLX: # %bb.0: # %entry
20717 ; NoVLX-NEXT: vcmpeqps %zmm1, %zmm0, %k0
20718 ; NoVLX-NEXT: kmovw %k0, %eax
20719 ; NoVLX-NEXT: andl %edi, %eax
20720 ; NoVLX-NEXT: vzeroupper
20721 ; NoVLX-NEXT: retq
20722 entry:
20723 %0 = bitcast <8 x i64> %__a to <16 x float>
20724 %1 = bitcast <8 x i64> %__b to <16 x float>
20725 %2 = fcmp oeq <16 x float> %0, %1
20726 %3 = bitcast i16 %__u to <16 x i1>
20727 %4 = and <16 x i1> %2, %3
20728 %5 = shufflevector <16 x i1> %4, <16 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>
20729 %6 = bitcast <32 x i1> %5 to i32
20730 ret i32 %6
20731 }
20733 define zeroext i32 @test_masked_vcmpoeqps_v16i1_v32i1_mask_mem(i16 zeroext %__u, <8 x i64> %__a, <8 x i64>* %__b) local_unnamed_addr {
20734 ; VLX-LABEL: test_masked_vcmpoeqps_v16i1_v32i1_mask_mem:
20735 ; VLX: # %bb.0: # %entry
20736 ; VLX-NEXT: kmovd %edi, %k1
20737 ; VLX-NEXT: vcmpeqps (%rsi), %zmm0, %k0 {%k1}
20738 ; VLX-NEXT: kmovd %k0, %eax
20739 ; VLX-NEXT: vzeroupper
20740 ; VLX-NEXT: retq
20741 ;
20742 ; NoVLX-LABEL: test_masked_vcmpoeqps_v16i1_v32i1_mask_mem:
20743 ; NoVLX: # %bb.0: # %entry
20744 ; NoVLX-NEXT: vcmpeqps (%rsi), %zmm0, %k0
20745 ; NoVLX-NEXT: kmovw %k0, %eax
20746 ; NoVLX-NEXT: andl %edi, %eax
20747 ; NoVLX-NEXT: vzeroupper
20748 ; NoVLX-NEXT: retq
20749 entry:
20750 %0 = bitcast <8 x i64> %__a to <16 x float>
20751 %load = load <8 x i64>, <8 x i64>* %__b
20752 %1 = bitcast <8 x i64> %load to <16 x float>
20753 %2 = fcmp oeq <16 x float> %0, %1
20754 %3 = bitcast i16 %__u to <16 x i1>
20755 %4 = and <16 x i1> %2, %3
20756 %5 = shufflevector <16 x i1> %4, <16 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>
20757 %6 = bitcast <32 x i1> %5 to i32
20758 ret i32 %6
20759 }
20761 define zeroext i32 @test_masked_vcmpoeqps_v16i1_v32i1_mask_mem_b(i16 zeroext %__u, <8 x i64> %__a, float* %__b) local_unnamed_addr {
20762 ; VLX-LABEL: test_masked_vcmpoeqps_v16i1_v32i1_mask_mem_b:
20763 ; VLX: # %bb.0: # %entry
20764 ; VLX-NEXT: kmovd %edi, %k1
20765 ; VLX-NEXT: vcmpeqps (%rsi){1to16}, %zmm0, %k0 {%k1}
20766 ; VLX-NEXT: kmovd %k0, %eax
20767 ; VLX-NEXT: vzeroupper
20768 ; VLX-NEXT: retq
20769 ;
20770 ; NoVLX-LABEL: test_masked_vcmpoeqps_v16i1_v32i1_mask_mem_b:
20771 ; NoVLX: # %bb.0: # %entry
20772 ; NoVLX-NEXT: vcmpeqps (%rsi){1to16}, %zmm0, %k0
20773 ; NoVLX-NEXT: kmovw %k0, %eax
20774 ; NoVLX-NEXT: andl %edi, %eax
20775 ; NoVLX-NEXT: vzeroupper
20776 ; NoVLX-NEXT: retq
20777 entry:
20778 %0 = bitcast <8 x i64> %__a to <16 x float>
20779 %load = load float, float* %__b
20780 %vec = insertelement <16 x float> undef, float %load, i32 0
20781 %1 = shufflevector <16 x float> %vec, <16 x float> undef, <16 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
20782 %2 = fcmp oeq <16 x float> %0, %1
20783 %3 = bitcast i16 %__u to <16 x i1>
20784 %4 = and <16 x i1> %2, %3
20785 %5 = shufflevector <16 x i1> %4, <16 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>
20786 %6 = bitcast <32 x i1> %5 to i32
20787 ret i32 %6
20788 }
20792 define zeroext i32 @test_vcmpoeqps_v16i1_v32i1_sae_mask(<8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
20793 ; CHECK-LABEL: test_vcmpoeqps_v16i1_v32i1_sae_mask:
20794 ; CHECK: # %bb.0: # %entry
20795 ; CHECK-NEXT: vcmpleps {sae}, %zmm1, %zmm0, %k0
20796 ; CHECK-NEXT: kmovw %k0, %eax
20797 ; CHECK-NEXT: vzeroupper
20798 ; CHECK-NEXT: retq
20799 entry:
20800 %0 = bitcast <8 x i64> %__a to <16 x float>
20801 %1 = bitcast <8 x i64> %__b to <16 x float>
20802 %2 = call <16 x i1> @llvm.x86.avx512.cmp.ps.512(<16 x float> %0, <16 x float> %1, i32 2, i32 8)
20803 %3 = bitcast <16 x i1> %2 to i16
20804 %4 = zext i16 %3 to i32
20805 ret i32 %4
20806 }
20808 define zeroext i32 @test_masked_vcmpoeqps_v16i1_v32i1_sae_mask(i16 zeroext %__u, <8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
20809 ; VLX-LABEL: test_masked_vcmpoeqps_v16i1_v32i1_sae_mask:
20810 ; VLX: # %bb.0: # %entry
20811 ; VLX-NEXT: vcmpleps {sae}, %zmm1, %zmm0, %k0
20812 ; VLX-NEXT: kmovd %k0, %eax
20813 ; VLX-NEXT: andl %edi, %eax
20814 ; VLX-NEXT: vzeroupper
20815 ; VLX-NEXT: retq
20816 ;
20817 ; NoVLX-LABEL: test_masked_vcmpoeqps_v16i1_v32i1_sae_mask:
20818 ; NoVLX: # %bb.0: # %entry
20819 ; NoVLX-NEXT: vcmpleps {sae}, %zmm1, %zmm0, %k0
20820 ; NoVLX-NEXT: kmovw %k0, %eax
20821 ; NoVLX-NEXT: andl %edi, %eax
20822 ; NoVLX-NEXT: vzeroupper
20823 ; NoVLX-NEXT: retq
20824 entry:
20825 %0 = bitcast <8 x i64> %__a to <16 x float>
20826 %1 = bitcast <8 x i64> %__b to <16 x float>
20827 %2 = call <16 x i1> @llvm.x86.avx512.cmp.ps.512(<16 x float> %0, <16 x float> %1, i32 2, i32 8)
20828 %3 = bitcast i16 %__u to <16 x i1>
20829 %4 = and <16 x i1> %2, %3
20830 %5 = bitcast <16 x i1> %4 to i16
20831 %6 = zext i16 %5 to i32
20832 ret i32 %6
20833 }
20837 define zeroext i64 @test_vcmpoeqps_v16i1_v64i1_mask(<8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
20838 ; VLX-LABEL: test_vcmpoeqps_v16i1_v64i1_mask:
20839 ; VLX: # %bb.0: # %entry
20840 ; VLX-NEXT: vcmpeqps %zmm1, %zmm0, %k0
20841 ; VLX-NEXT: kmovq %k0, %rax
20842 ; VLX-NEXT: vzeroupper
20843 ; VLX-NEXT: retq
20844 ;
20845 ; NoVLX-LABEL: test_vcmpoeqps_v16i1_v64i1_mask:
20846 ; NoVLX: # %bb.0: # %entry
20847 ; NoVLX-NEXT: vcmpeqps %zmm1, %zmm0, %k0
20848 ; NoVLX-NEXT: kmovw %k0, %eax
20849 ; NoVLX-NEXT: vzeroupper
20850 ; NoVLX-NEXT: retq
20851 entry:
20852 %0 = bitcast <8 x i64> %__a to <16 x float>
20853 %1 = bitcast <8 x i64> %__b to <16 x float>
20854 %2 = fcmp oeq <16 x float> %0, %1
20855 %3 = shufflevector <16 x i1> %2, <16 x i1> zeroinitializer, <64 x i32> <i32 0,i32 1,i32 2,i32 3,i32 4,i32 5,i32 6,i32 7,i32 8,i32 9,i32 10,i32 11,i32 12,i32 13,i32 14,i32 15,i32 16,i32 17,i32 18,i32 19,i32 20,i32 21,i32 22,i32 23,i32 24,i32 25,i32 26,i32 27,i32 28,i32 29,i32 30,i32 31,i32 16,i32 17,i32 18,i32 19,i32 20,i32 21,i32 22,i32 23,i32 24,i32 25,i32 26,i32 27,i32 28,i32 29,i32 30,i32 31,i32 16,i32 17,i32 18,i32 19,i32 20,i32 21,i32 22,i32 23,i32 24,i32 25,i32 26,i32 27,i32 28,i32 29,i32 30,i32 31>
20856 %4 = bitcast <64 x i1> %3 to i64
20857 ret i64 %4
20858 }
20860 define zeroext i64 @test_vcmpoeqps_v16i1_v64i1_mask_mem(<8 x i64> %__a, <8 x i64>* %__b) local_unnamed_addr {
20861 ; VLX-LABEL: test_vcmpoeqps_v16i1_v64i1_mask_mem:
20862 ; VLX: # %bb.0: # %entry
20863 ; VLX-NEXT: vcmpeqps (%rdi), %zmm0, %k0
20864 ; VLX-NEXT: kmovq %k0, %rax
20865 ; VLX-NEXT: vzeroupper
20866 ; VLX-NEXT: retq
20867 ;
20868 ; NoVLX-LABEL: test_vcmpoeqps_v16i1_v64i1_mask_mem:
20869 ; NoVLX: # %bb.0: # %entry
20870 ; NoVLX-NEXT: vcmpeqps (%rdi), %zmm0, %k0
20871 ; NoVLX-NEXT: kmovw %k0, %eax
20872 ; NoVLX-NEXT: vzeroupper
20873 ; NoVLX-NEXT: retq
20874 entry:
20875 %0 = bitcast <8 x i64> %__a to <16 x float>
20876 %load = load <8 x i64>, <8 x i64>* %__b
20877 %1 = bitcast <8 x i64> %load to <16 x float>
20878 %2 = fcmp oeq <16 x float> %0, %1
20879 %3 = shufflevector <16 x i1> %2, <16 x i1> zeroinitializer, <64 x i32> <i32 0,i32 1,i32 2,i32 3,i32 4,i32 5,i32 6,i32 7,i32 8,i32 9,i32 10,i32 11,i32 12,i32 13,i32 14,i32 15,i32 16,i32 17,i32 18,i32 19,i32 20,i32 21,i32 22,i32 23,i32 24,i32 25,i32 26,i32 27,i32 28,i32 29,i32 30,i32 31,i32 16,i32 17,i32 18,i32 19,i32 20,i32 21,i32 22,i32 23,i32 24,i32 25,i32 26,i32 27,i32 28,i32 29,i32 30,i32 31,i32 16,i32 17,i32 18,i32 19,i32 20,i32 21,i32 22,i32 23,i32 24,i32 25,i32 26,i32 27,i32 28,i32 29,i32 30,i32 31>
20880 %4 = bitcast <64 x i1> %3 to i64
20881 ret i64 %4
20882 }
20884 define zeroext i64 @test_vcmpoeqps_v16i1_v64i1_mask_mem_b(<8 x i64> %__a, float* %__b) local_unnamed_addr {
20885 ; VLX-LABEL: test_vcmpoeqps_v16i1_v64i1_mask_mem_b:
20886 ; VLX: # %bb.0: # %entry
20887 ; VLX-NEXT: vcmpeqps (%rdi){1to16}, %zmm0, %k0
20888 ; VLX-NEXT: kmovq %k0, %rax
20889 ; VLX-NEXT: vzeroupper
20890 ; VLX-NEXT: retq
20891 ;
20892 ; NoVLX-LABEL: test_vcmpoeqps_v16i1_v64i1_mask_mem_b:
20893 ; NoVLX: # %bb.0: # %entry
20894 ; NoVLX-NEXT: vcmpeqps (%rdi){1to16}, %zmm0, %k0
20895 ; NoVLX-NEXT: kmovw %k0, %eax
20896 ; NoVLX-NEXT: vzeroupper
20897 ; NoVLX-NEXT: retq
20898 entry:
20899 %0 = bitcast <8 x i64> %__a to <16 x float>
20900 %load = load float, float* %__b
20901 %vec = insertelement <16 x float> undef, float %load, i32 0
20902 %1 = shufflevector <16 x float> %vec, <16 x float> undef, <16 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
20903 %2 = fcmp oeq <16 x float> %0, %1
20904 %3 = shufflevector <16 x i1> %2, <16 x i1> zeroinitializer, <64 x i32> <i32 0,i32 1,i32 2,i32 3,i32 4,i32 5,i32 6,i32 7,i32 8,i32 9,i32 10,i32 11,i32 12,i32 13,i32 14,i32 15,i32 16,i32 17,i32 18,i32 19,i32 20,i32 21,i32 22,i32 23,i32 24,i32 25,i32 26,i32 27,i32 28,i32 29,i32 30,i32 31,i32 16,i32 17,i32 18,i32 19,i32 20,i32 21,i32 22,i32 23,i32 24,i32 25,i32 26,i32 27,i32 28,i32 29,i32 30,i32 31,i32 16,i32 17,i32 18,i32 19,i32 20,i32 21,i32 22,i32 23,i32 24,i32 25,i32 26,i32 27,i32 28,i32 29,i32 30,i32 31>
20905 %4 = bitcast <64 x i1> %3 to i64
20906 ret i64 %4
20907 }
20909 define zeroext i64 @test_masked_vcmpoeqps_v16i1_v64i1_mask(i16 zeroext %__u, <8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
20910 ; VLX-LABEL: test_masked_vcmpoeqps_v16i1_v64i1_mask:
20911 ; VLX: # %bb.0: # %entry
20912 ; VLX-NEXT: kmovd %edi, %k1
20913 ; VLX-NEXT: vcmpeqps %zmm1, %zmm0, %k0 {%k1}
20914 ; VLX-NEXT: kmovq %k0, %rax
20915 ; VLX-NEXT: vzeroupper
20916 ; VLX-NEXT: retq
20917 ;
20918 ; NoVLX-LABEL: test_masked_vcmpoeqps_v16i1_v64i1_mask:
20919 ; NoVLX: # %bb.0: # %entry
20920 ; NoVLX-NEXT: vcmpeqps %zmm1, %zmm0, %k0
20921 ; NoVLX-NEXT: kmovw %k0, %eax
20922 ; NoVLX-NEXT: andl %edi, %eax
20923 ; NoVLX-NEXT: vzeroupper
20924 ; NoVLX-NEXT: retq
20925 entry:
20926 %0 = bitcast <8 x i64> %__a to <16 x float>
20927 %1 = bitcast <8 x i64> %__b to <16 x float>
20928 %2 = fcmp oeq <16 x float> %0, %1
20929 %3 = bitcast i16 %__u to <16 x i1>
20930 %4 = and <16 x i1> %2, %3
20931 %5 = shufflevector <16 x i1> %4, <16 x i1> zeroinitializer, <64 x i32> <i32 0,i32 1,i32 2,i32 3,i32 4,i32 5,i32 6,i32 7,i32 8,i32 9,i32 10,i32 11,i32 12,i32 13,i32 14,i32 15,i32 16,i32 17,i32 18,i32 19,i32 20,i32 21,i32 22,i32 23,i32 24,i32 25,i32 26,i32 27,i32 28,i32 29,i32 30,i32 31,i32 16,i32 17,i32 18,i32 19,i32 20,i32 21,i32 22,i32 23,i32 24,i32 25,i32 26,i32 27,i32 28,i32 29,i32 30,i32 31,i32 16,i32 17,i32 18,i32 19,i32 20,i32 21,i32 22,i32 23,i32 24,i32 25,i32 26,i32 27,i32 28,i32 29,i32 30,i32 31>
20932 %6 = bitcast <64 x i1> %5 to i64
20933 ret i64 %6
20934 }
20936 define zeroext i64 @test_masked_vcmpoeqps_v16i1_v64i1_mask_mem(i16 zeroext %__u, <8 x i64> %__a, <8 x i64>* %__b) local_unnamed_addr {
20937 ; VLX-LABEL: test_masked_vcmpoeqps_v16i1_v64i1_mask_mem:
20938 ; VLX: # %bb.0: # %entry
20939 ; VLX-NEXT: kmovd %edi, %k1
20940 ; VLX-NEXT: vcmpeqps (%rsi), %zmm0, %k0 {%k1}
20941 ; VLX-NEXT: kmovq %k0, %rax
20942 ; VLX-NEXT: vzeroupper
20943 ; VLX-NEXT: retq
20944 ;
20945 ; NoVLX-LABEL: test_masked_vcmpoeqps_v16i1_v64i1_mask_mem:
20946 ; NoVLX: # %bb.0: # %entry
20947 ; NoVLX-NEXT: vcmpeqps (%rsi), %zmm0, %k0
20948 ; NoVLX-NEXT: kmovw %k0, %eax
20949 ; NoVLX-NEXT: andl %edi, %eax
20950 ; NoVLX-NEXT: vzeroupper
20951 ; NoVLX-NEXT: retq
20952 entry:
20953 %0 = bitcast <8 x i64> %__a to <16 x float>
20954 %load = load <8 x i64>, <8 x i64>* %__b
20955 %1 = bitcast <8 x i64> %load to <16 x float>
20956 %2 = fcmp oeq <16 x float> %0, %1
20957 %3 = bitcast i16 %__u to <16 x i1>
20958 %4 = and <16 x i1> %2, %3
20959 %5 = shufflevector <16 x i1> %4, <16 x i1> zeroinitializer, <64 x i32> <i32 0,i32 1,i32 2,i32 3,i32 4,i32 5,i32 6,i32 7,i32 8,i32 9,i32 10,i32 11,i32 12,i32 13,i32 14,i32 15,i32 16,i32 17,i32 18,i32 19,i32 20,i32 21,i32 22,i32 23,i32 24,i32 25,i32 26,i32 27,i32 28,i32 29,i32 30,i32 31,i32 16,i32 17,i32 18,i32 19,i32 20,i32 21,i32 22,i32 23,i32 24,i32 25,i32 26,i32 27,i32 28,i32 29,i32 30,i32 31,i32 16,i32 17,i32 18,i32 19,i32 20,i32 21,i32 22,i32 23,i32 24,i32 25,i32 26,i32 27,i32 28,i32 29,i32 30,i32 31>
20960 %6 = bitcast <64 x i1> %5 to i64
20961 ret i64 %6
20962 }
20964 define zeroext i64 @test_masked_vcmpoeqps_v16i1_v64i1_mask_mem_b(i16 zeroext %__u, <8 x i64> %__a, float* %__b) local_unnamed_addr {
20965 ; VLX-LABEL: test_masked_vcmpoeqps_v16i1_v64i1_mask_mem_b:
20966 ; VLX: # %bb.0: # %entry
20967 ; VLX-NEXT: kmovd %edi, %k1
20968 ; VLX-NEXT: vcmpeqps (%rsi){1to16}, %zmm0, %k0 {%k1}
20969 ; VLX-NEXT: kmovq %k0, %rax
20970 ; VLX-NEXT: vzeroupper
20971 ; VLX-NEXT: retq
20972 ;
20973 ; NoVLX-LABEL: test_masked_vcmpoeqps_v16i1_v64i1_mask_mem_b:
20974 ; NoVLX: # %bb.0: # %entry
20975 ; NoVLX-NEXT: vcmpeqps (%rsi){1to16}, %zmm0, %k0
20976 ; NoVLX-NEXT: kmovw %k0, %eax
20977 ; NoVLX-NEXT: andl %edi, %eax
20978 ; NoVLX-NEXT: vzeroupper
20979 ; NoVLX-NEXT: retq
20980 entry:
20981 %0 = bitcast <8 x i64> %__a to <16 x float>
20982 %load = load float, float* %__b
20983 %vec = insertelement <16 x float> undef, float %load, i32 0
20984 %1 = shufflevector <16 x float> %vec, <16 x float> undef, <16 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
20985 %2 = fcmp oeq <16 x float> %0, %1
20986 %3 = bitcast i16 %__u to <16 x i1>
20987 %4 = and <16 x i1> %2, %3
20988 %5 = shufflevector <16 x i1> %4, <16 x i1> zeroinitializer, <64 x i32> <i32 0,i32 1,i32 2,i32 3,i32 4,i32 5,i32 6,i32 7,i32 8,i32 9,i32 10,i32 11,i32 12,i32 13,i32 14,i32 15,i32 16,i32 17,i32 18,i32 19,i32 20,i32 21,i32 22,i32 23,i32 24,i32 25,i32 26,i32 27,i32 28,i32 29,i32 30,i32 31,i32 16,i32 17,i32 18,i32 19,i32 20,i32 21,i32 22,i32 23,i32 24,i32 25,i32 26,i32 27,i32 28,i32 29,i32 30,i32 31,i32 16,i32 17,i32 18,i32 19,i32 20,i32 21,i32 22,i32 23,i32 24,i32 25,i32 26,i32 27,i32 28,i32 29,i32 30,i32 31>
20989 %6 = bitcast <64 x i1> %5 to i64
20990 ret i64 %6
20991 }
20995 define zeroext i64 @test_vcmpoeqps_v16i1_v64i1_sae_mask(<8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
20996 ; CHECK-LABEL: test_vcmpoeqps_v16i1_v64i1_sae_mask:
20997 ; CHECK: # %bb.0: # %entry
20998 ; CHECK-NEXT: vcmpleps {sae}, %zmm1, %zmm0, %k0
20999 ; CHECK-NEXT: kmovw %k0, %eax
21000 ; CHECK-NEXT: vzeroupper
21001 ; CHECK-NEXT: retq
21002 entry:
21003 %0 = bitcast <8 x i64> %__a to <16 x float>
21004 %1 = bitcast <8 x i64> %__b to <16 x float>
21005 %2 = call <16 x i1> @llvm.x86.avx512.cmp.ps.512(<16 x float> %0, <16 x float> %1, i32 2, i32 8)
21006 %3 = bitcast <16 x i1> %2 to i16
21007 %4 = zext i16 %3 to i64
21008 ret i64 %4
21009 }
21011 define zeroext i64 @test_masked_vcmpoeqps_v16i1_v64i1_sae_mask(i16 zeroext %__u, <8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
21012 ; VLX-LABEL: test_masked_vcmpoeqps_v16i1_v64i1_sae_mask:
21013 ; VLX: # %bb.0: # %entry
21014 ; VLX-NEXT: vcmpleps {sae}, %zmm1, %zmm0, %k0
21015 ; VLX-NEXT: kmovd %k0, %eax
21016 ; VLX-NEXT: andl %edi, %eax
21017 ; VLX-NEXT: vzeroupper
21018 ; VLX-NEXT: retq
21019 ;
21020 ; NoVLX-LABEL: test_masked_vcmpoeqps_v16i1_v64i1_sae_mask:
21021 ; NoVLX: # %bb.0: # %entry
21022 ; NoVLX-NEXT: vcmpleps {sae}, %zmm1, %zmm0, %k0
21023 ; NoVLX-NEXT: kmovw %k0, %eax
21024 ; NoVLX-NEXT: andl %edi, %eax
21025 ; NoVLX-NEXT: vzeroupper
21026 ; NoVLX-NEXT: retq
21027 entry:
21028 %0 = bitcast <8 x i64> %__a to <16 x float>
21029 %1 = bitcast <8 x i64> %__b to <16 x float>
21030 %2 = call <16 x i1> @llvm.x86.avx512.cmp.ps.512(<16 x float> %0, <16 x float> %1, i32 2, i32 8)
21031 %3 = bitcast i16 %__u to <16 x i1>
21032 %4 = and <16 x i1> %2, %3
21033 %5 = bitcast <16 x i1> %4 to i16
21034 %6 = zext i16 %5 to i64
21035 ret i64 %6
21036 }
21040 declare <8 x i1> @llvm.x86.avx512.cmp.pd.512(<8 x double>, <8 x double>, i32, i32)
21041 define zeroext i4 @test_vcmpoeqpd_v2i1_v4i1_mask(<2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
21042 ; VLX-LABEL: test_vcmpoeqpd_v2i1_v4i1_mask:
21043 ; VLX: # %bb.0: # %entry
21044 ; VLX-NEXT: vcmpeqpd %xmm1, %xmm0, %k0
21045 ; VLX-NEXT: kmovb %k0, %eax
21046 ; VLX-NEXT: retq
21047 ;
21048 ; NoVLX-LABEL: test_vcmpoeqpd_v2i1_v4i1_mask:
21049 ; NoVLX: # %bb.0: # %entry
21050 ; NoVLX-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
21051 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
21052 ; NoVLX-NEXT: vcmpeqpd %zmm1, %zmm0, %k0
21053 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
21054 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
21055 ; NoVLX-NEXT: kmovw %k0, %eax
21056 ; NoVLX-NEXT: vzeroupper
21057 ; NoVLX-NEXT: retq
21058 entry:
21059 %0 = bitcast <2 x i64> %__a to <2 x double>
21060 %1 = bitcast <2 x i64> %__b to <2 x double>
21061 %2 = fcmp oeq <2 x double> %0, %1
21062 %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
21063 %4 = bitcast <4 x i1> %3 to i4
21064 ret i4 %4
21065 }
21067 define zeroext i4 @test_vcmpoeqpd_v2i1_v4i1_mask_mem(<2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
21068 ; VLX-LABEL: test_vcmpoeqpd_v2i1_v4i1_mask_mem:
21069 ; VLX: # %bb.0: # %entry
21070 ; VLX-NEXT: vcmpeqpd (%rdi), %xmm0, %k0
21071 ; VLX-NEXT: kmovb %k0, %eax
21072 ; VLX-NEXT: retq
21073 ;
21074 ; NoVLX-LABEL: test_vcmpoeqpd_v2i1_v4i1_mask_mem:
21075 ; NoVLX: # %bb.0: # %entry
21076 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
21077 ; NoVLX-NEXT: vmovapd (%rdi), %xmm1
21078 ; NoVLX-NEXT: vcmpeqpd %zmm1, %zmm0, %k0
21079 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
21080 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
21081 ; NoVLX-NEXT: kmovw %k0, %eax
21082 ; NoVLX-NEXT: vzeroupper
21083 ; NoVLX-NEXT: retq
21084 entry:
21085 %0 = bitcast <2 x i64> %__a to <2 x double>
21086 %load = load <2 x i64>, <2 x i64>* %__b
21087 %1 = bitcast <2 x i64> %load to <2 x double>
21088 %2 = fcmp oeq <2 x double> %0, %1
21089 %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
21090 %4 = bitcast <4 x i1> %3 to i4
21091 ret i4 %4
21092 }
21094 define zeroext i4 @test_vcmpoeqpd_v2i1_v4i1_mask_mem_b(<2 x i64> %__a, double* %__b) local_unnamed_addr {
21095 ; VLX-LABEL: test_vcmpoeqpd_v2i1_v4i1_mask_mem_b:
21096 ; VLX: # %bb.0: # %entry
21097 ; VLX-NEXT: vcmpeqpd (%rdi){1to2}, %xmm0, %k0
21098 ; VLX-NEXT: kmovb %k0, %eax
21099 ; VLX-NEXT: retq
21100 ;
21101 ; NoVLX-LABEL: test_vcmpoeqpd_v2i1_v4i1_mask_mem_b:
21102 ; NoVLX: # %bb.0: # %entry
21103 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
21104 ; NoVLX-NEXT: vcmpeqpd (%rdi){1to8}, %zmm0, %k0
21105 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
21106 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
21107 ; NoVLX-NEXT: kmovw %k0, %eax
21108 ; NoVLX-NEXT: vzeroupper
21109 ; NoVLX-NEXT: retq
21110 entry:
21111 %0 = bitcast <2 x i64> %__a to <2 x double>
21112 %load = load double, double* %__b
21113 %vec = insertelement <2 x double> undef, double %load, i32 0
21114 %1 = shufflevector <2 x double> %vec, <2 x double> undef, <2 x i32> <i32 0, i32 0>
21115 %2 = fcmp oeq <2 x double> %0, %1
21116 %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
21117 %4 = bitcast <4 x i1> %3 to i4
21118 ret i4 %4
21119 }
21121 define zeroext i4 @test_masked_vcmpoeqpd_v2i1_v4i1_mask(i2 zeroext %__u, <2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
21122 ; VLX-LABEL: test_masked_vcmpoeqpd_v2i1_v4i1_mask:
21123 ; VLX: # %bb.0: # %entry
21124 ; VLX-NEXT: kmovd %edi, %k1
21125 ; VLX-NEXT: vcmpeqpd %xmm1, %xmm0, %k0 {%k1}
21126 ; VLX-NEXT: kmovb %k0, %eax
21127 ; VLX-NEXT: retq
21128 ;
21129 ; NoVLX-LABEL: test_masked_vcmpoeqpd_v2i1_v4i1_mask:
21130 ; NoVLX: # %bb.0: # %entry
21131 ; NoVLX-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
21132 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
21133 ; NoVLX-NEXT: kmovw %edi, %k1
21134 ; NoVLX-NEXT: vcmpeqpd %zmm1, %zmm0, %k0 {%k1}
21135 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
21136 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
21137 ; NoVLX-NEXT: kmovw %k0, %eax
21138 ; NoVLX-NEXT: vzeroupper
21139 ; NoVLX-NEXT: retq
21140 entry:
21141 %0 = bitcast <2 x i64> %__a to <2 x double>
21142 %1 = bitcast <2 x i64> %__b to <2 x double>
21143 %2 = fcmp oeq <2 x double> %0, %1
21144 %3 = bitcast i2 %__u to <2 x i1>
21145 %4 = and <2 x i1> %2, %3
21146 %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
21147 %6 = bitcast <4 x i1> %5 to i4
21148 ret i4 %6
21149 }
21151 define zeroext i4 @test_masked_vcmpoeqpd_v2i1_v4i1_mask_mem(i2 zeroext %__u, <2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
21152 ; VLX-LABEL: test_masked_vcmpoeqpd_v2i1_v4i1_mask_mem:
21153 ; VLX: # %bb.0: # %entry
21154 ; VLX-NEXT: kmovd %edi, %k1
21155 ; VLX-NEXT: vcmpeqpd (%rsi), %xmm0, %k0 {%k1}
21156 ; VLX-NEXT: kmovb %k0, %eax
21157 ; VLX-NEXT: retq
21158 ;
21159 ; NoVLX-LABEL: test_masked_vcmpoeqpd_v2i1_v4i1_mask_mem:
21160 ; NoVLX: # %bb.0: # %entry
21161 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
21162 ; NoVLX-NEXT: kmovw %edi, %k1
21163 ; NoVLX-NEXT: vmovapd (%rsi), %xmm1
21164 ; NoVLX-NEXT: vcmpeqpd %zmm1, %zmm0, %k0 {%k1}
21165 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
21166 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
21167 ; NoVLX-NEXT: kmovw %k0, %eax
21168 ; NoVLX-NEXT: vzeroupper
21169 ; NoVLX-NEXT: retq
21170 entry:
21171 %0 = bitcast <2 x i64> %__a to <2 x double>
21172 %load = load <2 x i64>, <2 x i64>* %__b
21173 %1 = bitcast <2 x i64> %load to <2 x double>
21174 %2 = fcmp oeq <2 x double> %0, %1
21175 %3 = bitcast i2 %__u to <2 x i1>
21176 %4 = and <2 x i1> %2, %3
21177 %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
21178 %6 = bitcast <4 x i1> %5 to i4
21179 ret i4 %6
21180 }
21182 define zeroext i4 @test_masked_vcmpoeqpd_v2i1_v4i1_mask_mem_b(i2 zeroext %__u, <2 x i64> %__a, double* %__b) local_unnamed_addr {
21183 ; VLX-LABEL: test_masked_vcmpoeqpd_v2i1_v4i1_mask_mem_b:
21184 ; VLX: # %bb.0: # %entry
21185 ; VLX-NEXT: kmovd %edi, %k1
21186 ; VLX-NEXT: vcmpeqpd (%rsi){1to2}, %xmm0, %k0 {%k1}
21187 ; VLX-NEXT: kmovb %k0, %eax
21188 ; VLX-NEXT: retq
21189 ;
21190 ; NoVLX-LABEL: test_masked_vcmpoeqpd_v2i1_v4i1_mask_mem_b:
21191 ; NoVLX: # %bb.0: # %entry
21192 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
21193 ; NoVLX-NEXT: kmovw %edi, %k1
21194 ; NoVLX-NEXT: vcmpeqpd (%rsi){1to8}, %zmm0, %k0 {%k1}
21195 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
21196 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
21197 ; NoVLX-NEXT: kmovw %k0, %eax
21198 ; NoVLX-NEXT: vzeroupper
21199 ; NoVLX-NEXT: retq
21200 entry:
21201 %0 = bitcast <2 x i64> %__a to <2 x double>
21202 %load = load double, double* %__b
21203 %vec = insertelement <2 x double> undef, double %load, i32 0
21204 %1 = shufflevector <2 x double> %vec, <2 x double> undef, <2 x i32> <i32 0, i32 0>
21205 %2 = fcmp oeq <2 x double> %0, %1
21206 %3 = bitcast i2 %__u to <2 x i1>
21207 %4 = and <2 x i1> %2, %3
21208 %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
21209 %6 = bitcast <4 x i1> %5 to i4
21210 ret i4 %6
21211 }
21215 define zeroext i8 @test_vcmpoeqpd_v2i1_v8i1_mask(<2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
21216 ; VLX-LABEL: test_vcmpoeqpd_v2i1_v8i1_mask:
21217 ; VLX: # %bb.0: # %entry
21218 ; VLX-NEXT: vcmpeqpd %xmm1, %xmm0, %k0
21219 ; VLX-NEXT: kmovd %k0, %eax
21220 ; VLX-NEXT: # kill: def $al killed $al killed $eax
21221 ; VLX-NEXT: retq
21222 ;
21223 ; NoVLX-LABEL: test_vcmpoeqpd_v2i1_v8i1_mask:
21224 ; NoVLX: # %bb.0: # %entry
21225 ; NoVLX-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
21226 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
21227 ; NoVLX-NEXT: vcmpeqpd %zmm1, %zmm0, %k0
21228 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
21229 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
21230 ; NoVLX-NEXT: kmovw %k0, %eax
21231 ; NoVLX-NEXT: # kill: def $al killed $al killed $eax
21232 ; NoVLX-NEXT: vzeroupper
21233 ; NoVLX-NEXT: retq
21234 entry:
21235 %0 = bitcast <2 x i64> %__a to <2 x double>
21236 %1 = bitcast <2 x i64> %__b to <2 x double>
21237 %2 = fcmp oeq <2 x double> %0, %1
21238 %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
21239 %4 = bitcast <8 x i1> %3 to i8
21240 ret i8 %4
21241 }
21243 define zeroext i8 @test_vcmpoeqpd_v2i1_v8i1_mask_mem(<2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
21244 ; VLX-LABEL: test_vcmpoeqpd_v2i1_v8i1_mask_mem:
21245 ; VLX: # %bb.0: # %entry
21246 ; VLX-NEXT: vcmpeqpd (%rdi), %xmm0, %k0
21247 ; VLX-NEXT: kmovd %k0, %eax
21248 ; VLX-NEXT: # kill: def $al killed $al killed $eax
21249 ; VLX-NEXT: retq
21250 ;
21251 ; NoVLX-LABEL: test_vcmpoeqpd_v2i1_v8i1_mask_mem:
21252 ; NoVLX: # %bb.0: # %entry
21253 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
21254 ; NoVLX-NEXT: vmovapd (%rdi), %xmm1
21255 ; NoVLX-NEXT: vcmpeqpd %zmm1, %zmm0, %k0
21256 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
21257 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
21258 ; NoVLX-NEXT: kmovw %k0, %eax
21259 ; NoVLX-NEXT: # kill: def $al killed $al killed $eax
21260 ; NoVLX-NEXT: vzeroupper
21261 ; NoVLX-NEXT: retq
21262 entry:
21263 %0 = bitcast <2 x i64> %__a to <2 x double>
21264 %load = load <2 x i64>, <2 x i64>* %__b
21265 %1 = bitcast <2 x i64> %load to <2 x double>
21266 %2 = fcmp oeq <2 x double> %0, %1
21267 %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
21268 %4 = bitcast <8 x i1> %3 to i8
21269 ret i8 %4
21270 }
21272 define zeroext i8 @test_vcmpoeqpd_v2i1_v8i1_mask_mem_b(<2 x i64> %__a, double* %__b) local_unnamed_addr {
21273 ; VLX-LABEL: test_vcmpoeqpd_v2i1_v8i1_mask_mem_b:
21274 ; VLX: # %bb.0: # %entry
21275 ; VLX-NEXT: vcmpeqpd (%rdi){1to2}, %xmm0, %k0
21276 ; VLX-NEXT: kmovd %k0, %eax
21277 ; VLX-NEXT: # kill: def $al killed $al killed $eax
21278 ; VLX-NEXT: retq
21279 ;
21280 ; NoVLX-LABEL: test_vcmpoeqpd_v2i1_v8i1_mask_mem_b:
21281 ; NoVLX: # %bb.0: # %entry
21282 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
21283 ; NoVLX-NEXT: vcmpeqpd (%rdi){1to8}, %zmm0, %k0
21284 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
21285 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
21286 ; NoVLX-NEXT: kmovw %k0, %eax
21287 ; NoVLX-NEXT: # kill: def $al killed $al killed $eax
21288 ; NoVLX-NEXT: vzeroupper
21289 ; NoVLX-NEXT: retq
21290 entry:
21291 %0 = bitcast <2 x i64> %__a to <2 x double>
21292 %load = load double, double* %__b
21293 %vec = insertelement <2 x double> undef, double %load, i32 0
21294 %1 = shufflevector <2 x double> %vec, <2 x double> undef, <2 x i32> <i32 0, i32 0>
21295 %2 = fcmp oeq <2 x double> %0, %1
21296 %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
21297 %4 = bitcast <8 x i1> %3 to i8
21298 ret i8 %4
21299 }
21301 define zeroext i8 @test_masked_vcmpoeqpd_v2i1_v8i1_mask(i2 zeroext %__u, <2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
21302 ; VLX-LABEL: test_masked_vcmpoeqpd_v2i1_v8i1_mask:
21303 ; VLX: # %bb.0: # %entry
21304 ; VLX-NEXT: kmovd %edi, %k1
21305 ; VLX-NEXT: vcmpeqpd %xmm1, %xmm0, %k0 {%k1}
21306 ; VLX-NEXT: kmovd %k0, %eax
21307 ; VLX-NEXT: # kill: def $al killed $al killed $eax
21308 ; VLX-NEXT: retq
21309 ;
21310 ; NoVLX-LABEL: test_masked_vcmpoeqpd_v2i1_v8i1_mask:
21311 ; NoVLX: # %bb.0: # %entry
21312 ; NoVLX-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
21313 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
21314 ; NoVLX-NEXT: kmovw %edi, %k1
21315 ; NoVLX-NEXT: vcmpeqpd %zmm1, %zmm0, %k0 {%k1}
21316 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
21317 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
21318 ; NoVLX-NEXT: kmovw %k0, %eax
21319 ; NoVLX-NEXT: # kill: def $al killed $al killed $eax
21320 ; NoVLX-NEXT: vzeroupper
21321 ; NoVLX-NEXT: retq
21322 entry:
21323 %0 = bitcast <2 x i64> %__a to <2 x double>
21324 %1 = bitcast <2 x i64> %__b to <2 x double>
21325 %2 = fcmp oeq <2 x double> %0, %1
21326 %3 = bitcast i2 %__u to <2 x i1>
21327 %4 = and <2 x i1> %2, %3
21328 %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
21329 %6 = bitcast <8 x i1> %5 to i8
21330 ret i8 %6
21331 }
21333 define zeroext i8 @test_masked_vcmpoeqpd_v2i1_v8i1_mask_mem(i2 zeroext %__u, <2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
21334 ; VLX-LABEL: test_masked_vcmpoeqpd_v2i1_v8i1_mask_mem:
21335 ; VLX: # %bb.0: # %entry
21336 ; VLX-NEXT: kmovd %edi, %k1
21337 ; VLX-NEXT: vcmpeqpd (%rsi), %xmm0, %k0 {%k1}
21338 ; VLX-NEXT: kmovd %k0, %eax
21339 ; VLX-NEXT: # kill: def $al killed $al killed $eax
21340 ; VLX-NEXT: retq
21341 ;
21342 ; NoVLX-LABEL: test_masked_vcmpoeqpd_v2i1_v8i1_mask_mem:
21343 ; NoVLX: # %bb.0: # %entry
21344 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
21345 ; NoVLX-NEXT: kmovw %edi, %k1
21346 ; NoVLX-NEXT: vmovapd (%rsi), %xmm1
21347 ; NoVLX-NEXT: vcmpeqpd %zmm1, %zmm0, %k0 {%k1}
21348 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
21349 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
21350 ; NoVLX-NEXT: kmovw %k0, %eax
21351 ; NoVLX-NEXT: # kill: def $al killed $al killed $eax
21352 ; NoVLX-NEXT: vzeroupper
21353 ; NoVLX-NEXT: retq
21354 entry:
21355 %0 = bitcast <2 x i64> %__a to <2 x double>
21356 %load = load <2 x i64>, <2 x i64>* %__b
21357 %1 = bitcast <2 x i64> %load to <2 x double>
21358 %2 = fcmp oeq <2 x double> %0, %1
21359 %3 = bitcast i2 %__u to <2 x i1>
21360 %4 = and <2 x i1> %2, %3
21361 %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
21362 %6 = bitcast <8 x i1> %5 to i8
21363 ret i8 %6
21364 }
21366 define zeroext i8 @test_masked_vcmpoeqpd_v2i1_v8i1_mask_mem_b(i2 zeroext %__u, <2 x i64> %__a, double* %__b) local_unnamed_addr {
21367 ; VLX-LABEL: test_masked_vcmpoeqpd_v2i1_v8i1_mask_mem_b:
21368 ; VLX: # %bb.0: # %entry
21369 ; VLX-NEXT: kmovd %edi, %k1
21370 ; VLX-NEXT: vcmpeqpd (%rsi){1to2}, %xmm0, %k0 {%k1}
21371 ; VLX-NEXT: kmovd %k0, %eax
21372 ; VLX-NEXT: # kill: def $al killed $al killed $eax
21373 ; VLX-NEXT: retq
21374 ;
21375 ; NoVLX-LABEL: test_masked_vcmpoeqpd_v2i1_v8i1_mask_mem_b:
21376 ; NoVLX: # %bb.0: # %entry
21377 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
21378 ; NoVLX-NEXT: kmovw %edi, %k1
21379 ; NoVLX-NEXT: vcmpeqpd (%rsi){1to8}, %zmm0, %k0 {%k1}
21380 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
21381 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
21382 ; NoVLX-NEXT: kmovw %k0, %eax
21383 ; NoVLX-NEXT: # kill: def $al killed $al killed $eax
21384 ; NoVLX-NEXT: vzeroupper
21385 ; NoVLX-NEXT: retq
21386 entry:
21387 %0 = bitcast <2 x i64> %__a to <2 x double>
21388 %load = load double, double* %__b
21389 %vec = insertelement <2 x double> undef, double %load, i32 0
21390 %1 = shufflevector <2 x double> %vec, <2 x double> undef, <2 x i32> <i32 0, i32 0>
21391 %2 = fcmp oeq <2 x double> %0, %1
21392 %3 = bitcast i2 %__u to <2 x i1>
21393 %4 = and <2 x i1> %2, %3
21394 %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
21395 %6 = bitcast <8 x i1> %5 to i8
21396 ret i8 %6
21397 }
21401 define zeroext i16 @test_vcmpoeqpd_v2i1_v16i1_mask(<2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
21402 ; VLX-LABEL: test_vcmpoeqpd_v2i1_v16i1_mask:
21403 ; VLX: # %bb.0: # %entry
21404 ; VLX-NEXT: vcmpeqpd %xmm1, %xmm0, %k0
21405 ; VLX-NEXT: kmovd %k0, %eax
21406 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
21407 ; VLX-NEXT: retq
21408 ;
21409 ; NoVLX-LABEL: test_vcmpoeqpd_v2i1_v16i1_mask:
21410 ; NoVLX: # %bb.0: # %entry
21411 ; NoVLX-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
21412 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
21413 ; NoVLX-NEXT: vcmpeqpd %zmm1, %zmm0, %k0
21414 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
21415 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
21416 ; NoVLX-NEXT: kmovw %k0, %eax
21417 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
21418 ; NoVLX-NEXT: vzeroupper
21419 ; NoVLX-NEXT: retq
21420 entry:
21421 %0 = bitcast <2 x i64> %__a to <2 x double>
21422 %1 = bitcast <2 x i64> %__b to <2 x double>
21423 %2 = fcmp oeq <2 x double> %0, %1
21424 %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
21425 %4 = bitcast <16 x i1> %3 to i16
21426 ret i16 %4
21427 }
21429 define zeroext i16 @test_vcmpoeqpd_v2i1_v16i1_mask_mem(<2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
21430 ; VLX-LABEL: test_vcmpoeqpd_v2i1_v16i1_mask_mem:
21431 ; VLX: # %bb.0: # %entry
21432 ; VLX-NEXT: vcmpeqpd (%rdi), %xmm0, %k0
21433 ; VLX-NEXT: kmovd %k0, %eax
21434 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
21435 ; VLX-NEXT: retq
21436 ;
21437 ; NoVLX-LABEL: test_vcmpoeqpd_v2i1_v16i1_mask_mem:
21438 ; NoVLX: # %bb.0: # %entry
21439 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
21440 ; NoVLX-NEXT: vmovapd (%rdi), %xmm1
21441 ; NoVLX-NEXT: vcmpeqpd %zmm1, %zmm0, %k0
21442 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
21443 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
21444 ; NoVLX-NEXT: kmovw %k0, %eax
21445 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
21446 ; NoVLX-NEXT: vzeroupper
21447 ; NoVLX-NEXT: retq
21448 entry:
21449 %0 = bitcast <2 x i64> %__a to <2 x double>
21450 %load = load <2 x i64>, <2 x i64>* %__b
21451 %1 = bitcast <2 x i64> %load to <2 x double>
21452 %2 = fcmp oeq <2 x double> %0, %1
21453 %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
21454 %4 = bitcast <16 x i1> %3 to i16
21455 ret i16 %4
21456 }
21458 define zeroext i16 @test_vcmpoeqpd_v2i1_v16i1_mask_mem_b(<2 x i64> %__a, double* %__b) local_unnamed_addr {
21459 ; VLX-LABEL: test_vcmpoeqpd_v2i1_v16i1_mask_mem_b:
21460 ; VLX: # %bb.0: # %entry
21461 ; VLX-NEXT: vcmpeqpd (%rdi){1to2}, %xmm0, %k0
21462 ; VLX-NEXT: kmovd %k0, %eax
21463 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
21464 ; VLX-NEXT: retq
21465 ;
21466 ; NoVLX-LABEL: test_vcmpoeqpd_v2i1_v16i1_mask_mem_b:
21467 ; NoVLX: # %bb.0: # %entry
21468 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
21469 ; NoVLX-NEXT: vcmpeqpd (%rdi){1to8}, %zmm0, %k0
21470 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
21471 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
21472 ; NoVLX-NEXT: kmovw %k0, %eax
21473 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
21474 ; NoVLX-NEXT: vzeroupper
21475 ; NoVLX-NEXT: retq
21476 entry:
21477 %0 = bitcast <2 x i64> %__a to <2 x double>
21478 %load = load double, double* %__b
21479 %vec = insertelement <2 x double> undef, double %load, i32 0
21480 %1 = shufflevector <2 x double> %vec, <2 x double> undef, <2 x i32> <i32 0, i32 0>
21481 %2 = fcmp oeq <2 x double> %0, %1
21482 %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
21483 %4 = bitcast <16 x i1> %3 to i16
21484 ret i16 %4
21485 }
21487 define zeroext i16 @test_masked_vcmpoeqpd_v2i1_v16i1_mask(i2 zeroext %__u, <2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
21488 ; VLX-LABEL: test_masked_vcmpoeqpd_v2i1_v16i1_mask:
21489 ; VLX: # %bb.0: # %entry
21490 ; VLX-NEXT: kmovd %edi, %k1
21491 ; VLX-NEXT: vcmpeqpd %xmm1, %xmm0, %k0 {%k1}
21492 ; VLX-NEXT: kmovd %k0, %eax
21493 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
21494 ; VLX-NEXT: retq
21495 ;
21496 ; NoVLX-LABEL: test_masked_vcmpoeqpd_v2i1_v16i1_mask:
21497 ; NoVLX: # %bb.0: # %entry
21498 ; NoVLX-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
21499 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
21500 ; NoVLX-NEXT: kmovw %edi, %k1
21501 ; NoVLX-NEXT: vcmpeqpd %zmm1, %zmm0, %k0 {%k1}
21502 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
21503 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
21504 ; NoVLX-NEXT: kmovw %k0, %eax
21505 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
21506 ; NoVLX-NEXT: vzeroupper
21507 ; NoVLX-NEXT: retq
21508 entry:
21509 %0 = bitcast <2 x i64> %__a to <2 x double>
21510 %1 = bitcast <2 x i64> %__b to <2 x double>
21511 %2 = fcmp oeq <2 x double> %0, %1
21512 %3 = bitcast i2 %__u to <2 x i1>
21513 %4 = and <2 x i1> %2, %3
21514 %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
21515 %6 = bitcast <16 x i1> %5 to i16
21516 ret i16 %6
21517 }
21519 define zeroext i16 @test_masked_vcmpoeqpd_v2i1_v16i1_mask_mem(i2 zeroext %__u, <2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
21520 ; VLX-LABEL: test_masked_vcmpoeqpd_v2i1_v16i1_mask_mem:
21521 ; VLX: # %bb.0: # %entry
21522 ; VLX-NEXT: kmovd %edi, %k1
21523 ; VLX-NEXT: vcmpeqpd (%rsi), %xmm0, %k0 {%k1}
21524 ; VLX-NEXT: kmovd %k0, %eax
21525 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
21526 ; VLX-NEXT: retq
21527 ;
21528 ; NoVLX-LABEL: test_masked_vcmpoeqpd_v2i1_v16i1_mask_mem:
21529 ; NoVLX: # %bb.0: # %entry
21530 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
21531 ; NoVLX-NEXT: kmovw %edi, %k1
21532 ; NoVLX-NEXT: vmovapd (%rsi), %xmm1
21533 ; NoVLX-NEXT: vcmpeqpd %zmm1, %zmm0, %k0 {%k1}
21534 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
21535 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
21536 ; NoVLX-NEXT: kmovw %k0, %eax
21537 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
21538 ; NoVLX-NEXT: vzeroupper
21539 ; NoVLX-NEXT: retq
21540 entry:
21541 %0 = bitcast <2 x i64> %__a to <2 x double>
21542 %load = load <2 x i64>, <2 x i64>* %__b
21543 %1 = bitcast <2 x i64> %load to <2 x double>
21544 %2 = fcmp oeq <2 x double> %0, %1
21545 %3 = bitcast i2 %__u to <2 x i1>
21546 %4 = and <2 x i1> %2, %3
21547 %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
21548 %6 = bitcast <16 x i1> %5 to i16
21549 ret i16 %6
21550 }
21552 define zeroext i16 @test_masked_vcmpoeqpd_v2i1_v16i1_mask_mem_b(i2 zeroext %__u, <2 x i64> %__a, double* %__b) local_unnamed_addr {
21553 ; VLX-LABEL: test_masked_vcmpoeqpd_v2i1_v16i1_mask_mem_b:
21554 ; VLX: # %bb.0: # %entry
21555 ; VLX-NEXT: kmovd %edi, %k1
21556 ; VLX-NEXT: vcmpeqpd (%rsi){1to2}, %xmm0, %k0 {%k1}
21557 ; VLX-NEXT: kmovd %k0, %eax
21558 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
21559 ; VLX-NEXT: retq
21560 ;
21561 ; NoVLX-LABEL: test_masked_vcmpoeqpd_v2i1_v16i1_mask_mem_b:
21562 ; NoVLX: # %bb.0: # %entry
21563 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
21564 ; NoVLX-NEXT: kmovw %edi, %k1
21565 ; NoVLX-NEXT: vcmpeqpd (%rsi){1to8}, %zmm0, %k0 {%k1}
21566 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
21567 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
21568 ; NoVLX-NEXT: kmovw %k0, %eax
21569 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
21570 ; NoVLX-NEXT: vzeroupper
21571 ; NoVLX-NEXT: retq
21572 entry:
21573 %0 = bitcast <2 x i64> %__a to <2 x double>
21574 %load = load double, double* %__b
21575 %vec = insertelement <2 x double> undef, double %load, i32 0
21576 %1 = shufflevector <2 x double> %vec, <2 x double> undef, <2 x i32> <i32 0, i32 0>
21577 %2 = fcmp oeq <2 x double> %0, %1
21578 %3 = bitcast i2 %__u to <2 x i1>
21579 %4 = and <2 x i1> %2, %3
21580 %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
21581 %6 = bitcast <16 x i1> %5 to i16
21582 ret i16 %6
21583 }
21587 define zeroext i32 @test_vcmpoeqpd_v2i1_v32i1_mask(<2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
21588 ; VLX-LABEL: test_vcmpoeqpd_v2i1_v32i1_mask:
21589 ; VLX: # %bb.0: # %entry
21590 ; VLX-NEXT: vcmpeqpd %xmm1, %xmm0, %k0
21591 ; VLX-NEXT: kmovd %k0, %eax
21592 ; VLX-NEXT: retq
21593 ;
21594 ; NoVLX-LABEL: test_vcmpoeqpd_v2i1_v32i1_mask:
21595 ; NoVLX: # %bb.0: # %entry
21596 ; NoVLX-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
21597 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
21598 ; NoVLX-NEXT: vcmpeqpd %zmm1, %zmm0, %k0
21599 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
21600 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
21601 ; NoVLX-NEXT: kmovw %k0, %eax
21602 ; NoVLX-NEXT: vzeroupper
21603 ; NoVLX-NEXT: retq
21604 entry:
21605 %0 = bitcast <2 x i64> %__a to <2 x double>
21606 %1 = bitcast <2 x i64> %__b to <2 x double>
21607 %2 = fcmp oeq <2 x double> %0, %1
21608 %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
21609 %4 = bitcast <32 x i1> %3 to i32
21610 ret i32 %4
21611 }
21613 define zeroext i32 @test_vcmpoeqpd_v2i1_v32i1_mask_mem(<2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
21614 ; VLX-LABEL: test_vcmpoeqpd_v2i1_v32i1_mask_mem:
21615 ; VLX: # %bb.0: # %entry
21616 ; VLX-NEXT: vcmpeqpd (%rdi), %xmm0, %k0
21617 ; VLX-NEXT: kmovd %k0, %eax
21618 ; VLX-NEXT: retq
21619 ;
21620 ; NoVLX-LABEL: test_vcmpoeqpd_v2i1_v32i1_mask_mem:
21621 ; NoVLX: # %bb.0: # %entry
21622 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
21623 ; NoVLX-NEXT: vmovapd (%rdi), %xmm1
21624 ; NoVLX-NEXT: vcmpeqpd %zmm1, %zmm0, %k0
21625 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
21626 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
21627 ; NoVLX-NEXT: kmovw %k0, %eax
21628 ; NoVLX-NEXT: vzeroupper
21629 ; NoVLX-NEXT: retq
21630 entry:
21631 %0 = bitcast <2 x i64> %__a to <2 x double>
21632 %load = load <2 x i64>, <2 x i64>* %__b
21633 %1 = bitcast <2 x i64> %load to <2 x double>
21634 %2 = fcmp oeq <2 x double> %0, %1
21635 %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
21636 %4 = bitcast <32 x i1> %3 to i32
21637 ret i32 %4
21638 }
21640 define zeroext i32 @test_vcmpoeqpd_v2i1_v32i1_mask_mem_b(<2 x i64> %__a, double* %__b) local_unnamed_addr {
21641 ; VLX-LABEL: test_vcmpoeqpd_v2i1_v32i1_mask_mem_b:
21642 ; VLX: # %bb.0: # %entry
21643 ; VLX-NEXT: vcmpeqpd (%rdi){1to2}, %xmm0, %k0
21644 ; VLX-NEXT: kmovd %k0, %eax
21645 ; VLX-NEXT: retq
21646 ;
21647 ; NoVLX-LABEL: test_vcmpoeqpd_v2i1_v32i1_mask_mem_b:
21648 ; NoVLX: # %bb.0: # %entry
21649 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
21650 ; NoVLX-NEXT: vcmpeqpd (%rdi){1to8}, %zmm0, %k0
21651 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
21652 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
21653 ; NoVLX-NEXT: kmovw %k0, %eax
21654 ; NoVLX-NEXT: vzeroupper
21655 ; NoVLX-NEXT: retq
21656 entry:
21657 %0 = bitcast <2 x i64> %__a to <2 x double>
21658 %load = load double, double* %__b
21659 %vec = insertelement <2 x double> undef, double %load, i32 0
21660 %1 = shufflevector <2 x double> %vec, <2 x double> undef, <2 x i32> <i32 0, i32 0>
21661 %2 = fcmp oeq <2 x double> %0, %1
21662 %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
21663 %4 = bitcast <32 x i1> %3 to i32
21664 ret i32 %4
21665 }
21667 define zeroext i32 @test_masked_vcmpoeqpd_v2i1_v32i1_mask(i2 zeroext %__u, <2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
21668 ; VLX-LABEL: test_masked_vcmpoeqpd_v2i1_v32i1_mask:
21669 ; VLX: # %bb.0: # %entry
21670 ; VLX-NEXT: kmovd %edi, %k1
21671 ; VLX-NEXT: vcmpeqpd %xmm1, %xmm0, %k0 {%k1}
21672 ; VLX-NEXT: kmovd %k0, %eax
21673 ; VLX-NEXT: retq
21674 ;
21675 ; NoVLX-LABEL: test_masked_vcmpoeqpd_v2i1_v32i1_mask:
21676 ; NoVLX: # %bb.0: # %entry
21677 ; NoVLX-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
21678 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
21679 ; NoVLX-NEXT: kmovw %edi, %k1
21680 ; NoVLX-NEXT: vcmpeqpd %zmm1, %zmm0, %k0 {%k1}
21681 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
21682 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
21683 ; NoVLX-NEXT: kmovw %k0, %eax
21684 ; NoVLX-NEXT: vzeroupper
21685 ; NoVLX-NEXT: retq
21686 entry:
21687 %0 = bitcast <2 x i64> %__a to <2 x double>
21688 %1 = bitcast <2 x i64> %__b to <2 x double>
21689 %2 = fcmp oeq <2 x double> %0, %1
21690 %3 = bitcast i2 %__u to <2 x i1>
21691 %4 = and <2 x i1> %2, %3
21692 %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
21693 %6 = bitcast <32 x i1> %5 to i32
21694 ret i32 %6
21695 }
21697 define zeroext i32 @test_masked_vcmpoeqpd_v2i1_v32i1_mask_mem(i2 zeroext %__u, <2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
21698 ; VLX-LABEL: test_masked_vcmpoeqpd_v2i1_v32i1_mask_mem:
21699 ; VLX: # %bb.0: # %entry
21700 ; VLX-NEXT: kmovd %edi, %k1
21701 ; VLX-NEXT: vcmpeqpd (%rsi), %xmm0, %k0 {%k1}
21702 ; VLX-NEXT: kmovd %k0, %eax
21703 ; VLX-NEXT: retq
21704 ;
21705 ; NoVLX-LABEL: test_masked_vcmpoeqpd_v2i1_v32i1_mask_mem:
21706 ; NoVLX: # %bb.0: # %entry
21707 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
21708 ; NoVLX-NEXT: kmovw %edi, %k1
21709 ; NoVLX-NEXT: vmovapd (%rsi), %xmm1
21710 ; NoVLX-NEXT: vcmpeqpd %zmm1, %zmm0, %k0 {%k1}
21711 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
21712 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
21713 ; NoVLX-NEXT: kmovw %k0, %eax
21714 ; NoVLX-NEXT: vzeroupper
21715 ; NoVLX-NEXT: retq
21716 entry:
21717 %0 = bitcast <2 x i64> %__a to <2 x double>
21718 %load = load <2 x i64>, <2 x i64>* %__b
21719 %1 = bitcast <2 x i64> %load to <2 x double>
21720 %2 = fcmp oeq <2 x double> %0, %1
21721 %3 = bitcast i2 %__u to <2 x i1>
21722 %4 = and <2 x i1> %2, %3
21723 %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
21724 %6 = bitcast <32 x i1> %5 to i32
21725 ret i32 %6
21726 }
21728 define zeroext i32 @test_masked_vcmpoeqpd_v2i1_v32i1_mask_mem_b(i2 zeroext %__u, <2 x i64> %__a, double* %__b) local_unnamed_addr {
21729 ; VLX-LABEL: test_masked_vcmpoeqpd_v2i1_v32i1_mask_mem_b:
21730 ; VLX: # %bb.0: # %entry
21731 ; VLX-NEXT: kmovd %edi, %k1
21732 ; VLX-NEXT: vcmpeqpd (%rsi){1to2}, %xmm0, %k0 {%k1}
21733 ; VLX-NEXT: kmovd %k0, %eax
21734 ; VLX-NEXT: retq
21735 ;
21736 ; NoVLX-LABEL: test_masked_vcmpoeqpd_v2i1_v32i1_mask_mem_b:
21737 ; NoVLX: # %bb.0: # %entry
21738 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
21739 ; NoVLX-NEXT: kmovw %edi, %k1
21740 ; NoVLX-NEXT: vcmpeqpd (%rsi){1to8}, %zmm0, %k0 {%k1}
21741 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
21742 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
21743 ; NoVLX-NEXT: kmovw %k0, %eax
21744 ; NoVLX-NEXT: vzeroupper
21745 ; NoVLX-NEXT: retq
21746 entry:
21747 %0 = bitcast <2 x i64> %__a to <2 x double>
21748 %load = load double, double* %__b
21749 %vec = insertelement <2 x double> undef, double %load, i32 0
21750 %1 = shufflevector <2 x double> %vec, <2 x double> undef, <2 x i32> <i32 0, i32 0>
21751 %2 = fcmp oeq <2 x double> %0, %1
21752 %3 = bitcast i2 %__u to <2 x i1>
21753 %4 = and <2 x i1> %2, %3
21754 %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
21755 %6 = bitcast <32 x i1> %5 to i32
21756 ret i32 %6
21757 }
21761 define zeroext i64 @test_vcmpoeqpd_v2i1_v64i1_mask(<2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
21762 ; VLX-LABEL: test_vcmpoeqpd_v2i1_v64i1_mask:
21763 ; VLX: # %bb.0: # %entry
21764 ; VLX-NEXT: vcmpeqpd %xmm1, %xmm0, %k0
21765 ; VLX-NEXT: kmovq %k0, %rax
21766 ; VLX-NEXT: retq
21767 ;
21768 ; NoVLX-LABEL: test_vcmpoeqpd_v2i1_v64i1_mask:
21769 ; NoVLX: # %bb.0: # %entry
21770 ; NoVLX-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
21771 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
21772 ; NoVLX-NEXT: vcmpeqpd %zmm1, %zmm0, %k0
21773 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
21774 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
21775 ; NoVLX-NEXT: kmovw %k0, %eax
21776 ; NoVLX-NEXT: vzeroupper
21777 ; NoVLX-NEXT: retq
21778 entry:
21779 %0 = bitcast <2 x i64> %__a to <2 x double>
21780 %1 = bitcast <2 x i64> %__b to <2 x double>
21781 %2 = fcmp oeq <2 x double> %0, %1
21782 %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
21783 %4 = bitcast <64 x i1> %3 to i64
21784 ret i64 %4
21785 }
21787 define zeroext i64 @test_vcmpoeqpd_v2i1_v64i1_mask_mem(<2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
21788 ; VLX-LABEL: test_vcmpoeqpd_v2i1_v64i1_mask_mem:
21789 ; VLX: # %bb.0: # %entry
21790 ; VLX-NEXT: vcmpeqpd (%rdi), %xmm0, %k0
21791 ; VLX-NEXT: kmovq %k0, %rax
21792 ; VLX-NEXT: retq
21793 ;
21794 ; NoVLX-LABEL: test_vcmpoeqpd_v2i1_v64i1_mask_mem:
21795 ; NoVLX: # %bb.0: # %entry
21796 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
21797 ; NoVLX-NEXT: vmovapd (%rdi), %xmm1
21798 ; NoVLX-NEXT: vcmpeqpd %zmm1, %zmm0, %k0
21799 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
21800 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
21801 ; NoVLX-NEXT: kmovw %k0, %eax
21802 ; NoVLX-NEXT: vzeroupper
21803 ; NoVLX-NEXT: retq
21804 entry:
21805 %0 = bitcast <2 x i64> %__a to <2 x double>
21806 %load = load <2 x i64>, <2 x i64>* %__b
21807 %1 = bitcast <2 x i64> %load to <2 x double>
21808 %2 = fcmp oeq <2 x double> %0, %1
21809 %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
21810 %4 = bitcast <64 x i1> %3 to i64
21811 ret i64 %4
21812 }
21814 define zeroext i64 @test_vcmpoeqpd_v2i1_v64i1_mask_mem_b(<2 x i64> %__a, double* %__b) local_unnamed_addr {
21815 ; VLX-LABEL: test_vcmpoeqpd_v2i1_v64i1_mask_mem_b:
21816 ; VLX: # %bb.0: # %entry
21817 ; VLX-NEXT: vcmpeqpd (%rdi){1to2}, %xmm0, %k0
21818 ; VLX-NEXT: kmovq %k0, %rax
21819 ; VLX-NEXT: retq
21820 ;
21821 ; NoVLX-LABEL: test_vcmpoeqpd_v2i1_v64i1_mask_mem_b:
21822 ; NoVLX: # %bb.0: # %entry
21823 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
21824 ; NoVLX-NEXT: vcmpeqpd (%rdi){1to8}, %zmm0, %k0
21825 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
21826 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
21827 ; NoVLX-NEXT: kmovw %k0, %eax
21828 ; NoVLX-NEXT: vzeroupper
21829 ; NoVLX-NEXT: retq
21830 entry:
21831 %0 = bitcast <2 x i64> %__a to <2 x double>
21832 %load = load double, double* %__b
21833 %vec = insertelement <2 x double> undef, double %load, i32 0
21834 %1 = shufflevector <2 x double> %vec, <2 x double> undef, <2 x i32> <i32 0, i32 0>
21835 %2 = fcmp oeq <2 x double> %0, %1
21836 %3 = shufflevector <2 x i1> %2, <2 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
21837 %4 = bitcast <64 x i1> %3 to i64
21838 ret i64 %4
21839 }
21841 define zeroext i64 @test_masked_vcmpoeqpd_v2i1_v64i1_mask(i2 zeroext %__u, <2 x i64> %__a, <2 x i64> %__b) local_unnamed_addr {
21842 ; VLX-LABEL: test_masked_vcmpoeqpd_v2i1_v64i1_mask:
21843 ; VLX: # %bb.0: # %entry
21844 ; VLX-NEXT: kmovd %edi, %k1
21845 ; VLX-NEXT: vcmpeqpd %xmm1, %xmm0, %k0 {%k1}
21846 ; VLX-NEXT: kmovq %k0, %rax
21847 ; VLX-NEXT: retq
21848 ;
21849 ; NoVLX-LABEL: test_masked_vcmpoeqpd_v2i1_v64i1_mask:
21850 ; NoVLX: # %bb.0: # %entry
21851 ; NoVLX-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
21852 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
21853 ; NoVLX-NEXT: kmovw %edi, %k1
21854 ; NoVLX-NEXT: vcmpeqpd %zmm1, %zmm0, %k0 {%k1}
21855 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
21856 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
21857 ; NoVLX-NEXT: kmovw %k0, %eax
21858 ; NoVLX-NEXT: vzeroupper
21859 ; NoVLX-NEXT: retq
21860 entry:
21861 %0 = bitcast <2 x i64> %__a to <2 x double>
21862 %1 = bitcast <2 x i64> %__b to <2 x double>
21863 %2 = fcmp oeq <2 x double> %0, %1
21864 %3 = bitcast i2 %__u to <2 x i1>
21865 %4 = and <2 x i1> %2, %3
21866 %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
21867 %6 = bitcast <64 x i1> %5 to i64
21868 ret i64 %6
21869 }
21871 define zeroext i64 @test_masked_vcmpoeqpd_v2i1_v64i1_mask_mem(i2 zeroext %__u, <2 x i64> %__a, <2 x i64>* %__b) local_unnamed_addr {
21872 ; VLX-LABEL: test_masked_vcmpoeqpd_v2i1_v64i1_mask_mem:
21873 ; VLX: # %bb.0: # %entry
21874 ; VLX-NEXT: kmovd %edi, %k1
21875 ; VLX-NEXT: vcmpeqpd (%rsi), %xmm0, %k0 {%k1}
21876 ; VLX-NEXT: kmovq %k0, %rax
21877 ; VLX-NEXT: retq
21878 ;
21879 ; NoVLX-LABEL: test_masked_vcmpoeqpd_v2i1_v64i1_mask_mem:
21880 ; NoVLX: # %bb.0: # %entry
21881 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
21882 ; NoVLX-NEXT: kmovw %edi, %k1
21883 ; NoVLX-NEXT: vmovapd (%rsi), %xmm1
21884 ; NoVLX-NEXT: vcmpeqpd %zmm1, %zmm0, %k0 {%k1}
21885 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
21886 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
21887 ; NoVLX-NEXT: kmovw %k0, %eax
21888 ; NoVLX-NEXT: vzeroupper
21889 ; NoVLX-NEXT: retq
21890 entry:
21891 %0 = bitcast <2 x i64> %__a to <2 x double>
21892 %load = load <2 x i64>, <2 x i64>* %__b
21893 %1 = bitcast <2 x i64> %load to <2 x double>
21894 %2 = fcmp oeq <2 x double> %0, %1
21895 %3 = bitcast i2 %__u to <2 x i1>
21896 %4 = and <2 x i1> %2, %3
21897 %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
21898 %6 = bitcast <64 x i1> %5 to i64
21899 ret i64 %6
21900 }
21902 define zeroext i64 @test_masked_vcmpoeqpd_v2i1_v64i1_mask_mem_b(i2 zeroext %__u, <2 x i64> %__a, double* %__b) local_unnamed_addr {
21903 ; VLX-LABEL: test_masked_vcmpoeqpd_v2i1_v64i1_mask_mem_b:
21904 ; VLX: # %bb.0: # %entry
21905 ; VLX-NEXT: kmovd %edi, %k1
21906 ; VLX-NEXT: vcmpeqpd (%rsi){1to2}, %xmm0, %k0 {%k1}
21907 ; VLX-NEXT: kmovq %k0, %rax
21908 ; VLX-NEXT: retq
21909 ;
21910 ; NoVLX-LABEL: test_masked_vcmpoeqpd_v2i1_v64i1_mask_mem_b:
21911 ; NoVLX: # %bb.0: # %entry
21912 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
21913 ; NoVLX-NEXT: kmovw %edi, %k1
21914 ; NoVLX-NEXT: vcmpeqpd (%rsi){1to8}, %zmm0, %k0 {%k1}
21915 ; NoVLX-NEXT: kshiftlw $14, %k0, %k0
21916 ; NoVLX-NEXT: kshiftrw $14, %k0, %k0
21917 ; NoVLX-NEXT: kmovw %k0, %eax
21918 ; NoVLX-NEXT: vzeroupper
21919 ; NoVLX-NEXT: retq
21920 entry:
21921 %0 = bitcast <2 x i64> %__a to <2 x double>
21922 %load = load double, double* %__b
21923 %vec = insertelement <2 x double> undef, double %load, i32 0
21924 %1 = shufflevector <2 x double> %vec, <2 x double> undef, <2 x i32> <i32 0, i32 0>
21925 %2 = fcmp oeq <2 x double> %0, %1
21926 %3 = bitcast i2 %__u to <2 x i1>
21927 %4 = and <2 x i1> %2, %3
21928 %5 = shufflevector <2 x i1> %4, <2 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3, i32 2, i32 3>
21929 %6 = bitcast <64 x i1> %5 to i64
21930 ret i64 %6
21931 }
21935 define zeroext i8 @test_vcmpoeqpd_v4i1_v8i1_mask(<4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
21936 ; VLX-LABEL: test_vcmpoeqpd_v4i1_v8i1_mask:
21937 ; VLX: # %bb.0: # %entry
21938 ; VLX-NEXT: vcmpeqpd %ymm1, %ymm0, %k0
21939 ; VLX-NEXT: kmovd %k0, %eax
21940 ; VLX-NEXT: # kill: def $al killed $al killed $eax
21941 ; VLX-NEXT: vzeroupper
21942 ; VLX-NEXT: retq
21943 ;
21944 ; NoVLX-LABEL: test_vcmpoeqpd_v4i1_v8i1_mask:
21945 ; NoVLX: # %bb.0: # %entry
21946 ; NoVLX-NEXT: # kill: def $ymm1 killed $ymm1 def $zmm1
21947 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
21948 ; NoVLX-NEXT: vcmpeqpd %zmm1, %zmm0, %k0
21949 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
21950 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
21951 ; NoVLX-NEXT: kmovw %k0, %eax
21952 ; NoVLX-NEXT: # kill: def $al killed $al killed $eax
21953 ; NoVLX-NEXT: vzeroupper
21954 ; NoVLX-NEXT: retq
21955 entry:
21956 %0 = bitcast <4 x i64> %__a to <4 x double>
21957 %1 = bitcast <4 x i64> %__b to <4 x double>
21958 %2 = fcmp oeq <4 x double> %0, %1
21959 %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
21960 %4 = bitcast <8 x i1> %3 to i8
21961 ret i8 %4
21962 }
21964 define zeroext i8 @test_vcmpoeqpd_v4i1_v8i1_mask_mem(<4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
21965 ; VLX-LABEL: test_vcmpoeqpd_v4i1_v8i1_mask_mem:
21966 ; VLX: # %bb.0: # %entry
21967 ; VLX-NEXT: vcmpeqpd (%rdi), %ymm0, %k0
21968 ; VLX-NEXT: kmovd %k0, %eax
21969 ; VLX-NEXT: # kill: def $al killed $al killed $eax
21970 ; VLX-NEXT: vzeroupper
21971 ; VLX-NEXT: retq
21972 ;
21973 ; NoVLX-LABEL: test_vcmpoeqpd_v4i1_v8i1_mask_mem:
21974 ; NoVLX: # %bb.0: # %entry
21975 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
21976 ; NoVLX-NEXT: vmovapd (%rdi), %ymm1
21977 ; NoVLX-NEXT: vcmpeqpd %zmm1, %zmm0, %k0
21978 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
21979 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
21980 ; NoVLX-NEXT: kmovw %k0, %eax
21981 ; NoVLX-NEXT: # kill: def $al killed $al killed $eax
21982 ; NoVLX-NEXT: vzeroupper
21983 ; NoVLX-NEXT: retq
21984 entry:
21985 %0 = bitcast <4 x i64> %__a to <4 x double>
21986 %load = load <4 x i64>, <4 x i64>* %__b
21987 %1 = bitcast <4 x i64> %load to <4 x double>
21988 %2 = fcmp oeq <4 x double> %0, %1
21989 %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
21990 %4 = bitcast <8 x i1> %3 to i8
21991 ret i8 %4
21992 }
21994 define zeroext i8 @test_vcmpoeqpd_v4i1_v8i1_mask_mem_b(<4 x i64> %__a, double* %__b) local_unnamed_addr {
21995 ; VLX-LABEL: test_vcmpoeqpd_v4i1_v8i1_mask_mem_b:
21996 ; VLX: # %bb.0: # %entry
21997 ; VLX-NEXT: vcmpeqpd (%rdi){1to4}, %ymm0, %k0
21998 ; VLX-NEXT: kmovd %k0, %eax
21999 ; VLX-NEXT: # kill: def $al killed $al killed $eax
22000 ; VLX-NEXT: vzeroupper
22001 ; VLX-NEXT: retq
22002 ;
22003 ; NoVLX-LABEL: test_vcmpoeqpd_v4i1_v8i1_mask_mem_b:
22004 ; NoVLX: # %bb.0: # %entry
22005 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
22006 ; NoVLX-NEXT: vcmpeqpd (%rdi){1to8}, %zmm0, %k0
22007 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
22008 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
22009 ; NoVLX-NEXT: kmovw %k0, %eax
22010 ; NoVLX-NEXT: # kill: def $al killed $al killed $eax
22011 ; NoVLX-NEXT: vzeroupper
22012 ; NoVLX-NEXT: retq
22013 entry:
22014 %0 = bitcast <4 x i64> %__a to <4 x double>
22015 %load = load double, double* %__b
22016 %vec = insertelement <4 x double> undef, double %load, i32 0
22017 %1 = shufflevector <4 x double> %vec, <4 x double> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
22018 %2 = fcmp oeq <4 x double> %0, %1
22019 %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
22020 %4 = bitcast <8 x i1> %3 to i8
22021 ret i8 %4
22022 }
22024 define zeroext i8 @test_masked_vcmpoeqpd_v4i1_v8i1_mask(i4 zeroext %__u, <4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
22025 ; VLX-LABEL: test_masked_vcmpoeqpd_v4i1_v8i1_mask:
22026 ; VLX: # %bb.0: # %entry
22027 ; VLX-NEXT: kmovd %edi, %k1
22028 ; VLX-NEXT: vcmpeqpd %ymm1, %ymm0, %k0 {%k1}
22029 ; VLX-NEXT: kmovd %k0, %eax
22030 ; VLX-NEXT: # kill: def $al killed $al killed $eax
22031 ; VLX-NEXT: vzeroupper
22032 ; VLX-NEXT: retq
22033 ;
22034 ; NoVLX-LABEL: test_masked_vcmpoeqpd_v4i1_v8i1_mask:
22035 ; NoVLX: # %bb.0: # %entry
22036 ; NoVLX-NEXT: # kill: def $ymm1 killed $ymm1 def $zmm1
22037 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
22038 ; NoVLX-NEXT: kmovw %edi, %k1
22039 ; NoVLX-NEXT: vcmpeqpd %zmm1, %zmm0, %k0 {%k1}
22040 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
22041 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
22042 ; NoVLX-NEXT: kmovw %k0, %eax
22043 ; NoVLX-NEXT: # kill: def $al killed $al killed $eax
22044 ; NoVLX-NEXT: vzeroupper
22045 ; NoVLX-NEXT: retq
22046 entry:
22047 %0 = bitcast <4 x i64> %__a to <4 x double>
22048 %1 = bitcast <4 x i64> %__b to <4 x double>
22049 %2 = fcmp oeq <4 x double> %0, %1
22050 %3 = bitcast i4 %__u to <4 x i1>
22051 %4 = and <4 x i1> %2, %3
22052 %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
22053 %6 = bitcast <8 x i1> %5 to i8
22054 ret i8 %6
22055 }
22057 define zeroext i8 @test_masked_vcmpoeqpd_v4i1_v8i1_mask_mem(i4 zeroext %__u, <4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
22058 ; VLX-LABEL: test_masked_vcmpoeqpd_v4i1_v8i1_mask_mem:
22059 ; VLX: # %bb.0: # %entry
22060 ; VLX-NEXT: kmovd %edi, %k1
22061 ; VLX-NEXT: vcmpeqpd (%rsi), %ymm0, %k0 {%k1}
22062 ; VLX-NEXT: kmovd %k0, %eax
22063 ; VLX-NEXT: # kill: def $al killed $al killed $eax
22064 ; VLX-NEXT: vzeroupper
22065 ; VLX-NEXT: retq
22066 ;
22067 ; NoVLX-LABEL: test_masked_vcmpoeqpd_v4i1_v8i1_mask_mem:
22068 ; NoVLX: # %bb.0: # %entry
22069 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
22070 ; NoVLX-NEXT: kmovw %edi, %k1
22071 ; NoVLX-NEXT: vmovapd (%rsi), %ymm1
22072 ; NoVLX-NEXT: vcmpeqpd %zmm1, %zmm0, %k0 {%k1}
22073 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
22074 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
22075 ; NoVLX-NEXT: kmovw %k0, %eax
22076 ; NoVLX-NEXT: # kill: def $al killed $al killed $eax
22077 ; NoVLX-NEXT: vzeroupper
22078 ; NoVLX-NEXT: retq
22079 entry:
22080 %0 = bitcast <4 x i64> %__a to <4 x double>
22081 %load = load <4 x i64>, <4 x i64>* %__b
22082 %1 = bitcast <4 x i64> %load to <4 x double>
22083 %2 = fcmp oeq <4 x double> %0, %1
22084 %3 = bitcast i4 %__u to <4 x i1>
22085 %4 = and <4 x i1> %2, %3
22086 %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
22087 %6 = bitcast <8 x i1> %5 to i8
22088 ret i8 %6
22089 }
22091 define zeroext i8 @test_masked_vcmpoeqpd_v4i1_v8i1_mask_mem_b(i4 zeroext %__u, <4 x i64> %__a, double* %__b) local_unnamed_addr {
22092 ; VLX-LABEL: test_masked_vcmpoeqpd_v4i1_v8i1_mask_mem_b:
22093 ; VLX: # %bb.0: # %entry
22094 ; VLX-NEXT: kmovd %edi, %k1
22095 ; VLX-NEXT: vcmpeqpd (%rsi){1to4}, %ymm0, %k0 {%k1}
22096 ; VLX-NEXT: kmovd %k0, %eax
22097 ; VLX-NEXT: # kill: def $al killed $al killed $eax
22098 ; VLX-NEXT: vzeroupper
22099 ; VLX-NEXT: retq
22100 ;
22101 ; NoVLX-LABEL: test_masked_vcmpoeqpd_v4i1_v8i1_mask_mem_b:
22102 ; NoVLX: # %bb.0: # %entry
22103 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
22104 ; NoVLX-NEXT: kmovw %edi, %k1
22105 ; NoVLX-NEXT: vcmpeqpd (%rsi){1to8}, %zmm0, %k0 {%k1}
22106 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
22107 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
22108 ; NoVLX-NEXT: kmovw %k0, %eax
22109 ; NoVLX-NEXT: # kill: def $al killed $al killed $eax
22110 ; NoVLX-NEXT: vzeroupper
22111 ; NoVLX-NEXT: retq
22112 entry:
22113 %0 = bitcast <4 x i64> %__a to <4 x double>
22114 %load = load double, double* %__b
22115 %vec = insertelement <4 x double> undef, double %load, i32 0
22116 %1 = shufflevector <4 x double> %vec, <4 x double> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
22117 %2 = fcmp oeq <4 x double> %0, %1
22118 %3 = bitcast i4 %__u to <4 x i1>
22119 %4 = and <4 x i1> %2, %3
22120 %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
22121 %6 = bitcast <8 x i1> %5 to i8
22122 ret i8 %6
22123 }
22127 define zeroext i16 @test_vcmpoeqpd_v4i1_v16i1_mask(<4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
22128 ; VLX-LABEL: test_vcmpoeqpd_v4i1_v16i1_mask:
22129 ; VLX: # %bb.0: # %entry
22130 ; VLX-NEXT: vcmpeqpd %ymm1, %ymm0, %k0
22131 ; VLX-NEXT: kmovd %k0, %eax
22132 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
22133 ; VLX-NEXT: vzeroupper
22134 ; VLX-NEXT: retq
22135 ;
22136 ; NoVLX-LABEL: test_vcmpoeqpd_v4i1_v16i1_mask:
22137 ; NoVLX: # %bb.0: # %entry
22138 ; NoVLX-NEXT: # kill: def $ymm1 killed $ymm1 def $zmm1
22139 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
22140 ; NoVLX-NEXT: vcmpeqpd %zmm1, %zmm0, %k0
22141 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
22142 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
22143 ; NoVLX-NEXT: kmovw %k0, %eax
22144 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
22145 ; NoVLX-NEXT: vzeroupper
22146 ; NoVLX-NEXT: retq
22147 entry:
22148 %0 = bitcast <4 x i64> %__a to <4 x double>
22149 %1 = bitcast <4 x i64> %__b to <4 x double>
22150 %2 = fcmp oeq <4 x double> %0, %1
22151 %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
22152 %4 = bitcast <16 x i1> %3 to i16
22153 ret i16 %4
22154 }
22156 define zeroext i16 @test_vcmpoeqpd_v4i1_v16i1_mask_mem(<4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
22157 ; VLX-LABEL: test_vcmpoeqpd_v4i1_v16i1_mask_mem:
22158 ; VLX: # %bb.0: # %entry
22159 ; VLX-NEXT: vcmpeqpd (%rdi), %ymm0, %k0
22160 ; VLX-NEXT: kmovd %k0, %eax
22161 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
22162 ; VLX-NEXT: vzeroupper
22163 ; VLX-NEXT: retq
22164 ;
22165 ; NoVLX-LABEL: test_vcmpoeqpd_v4i1_v16i1_mask_mem:
22166 ; NoVLX: # %bb.0: # %entry
22167 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
22168 ; NoVLX-NEXT: vmovapd (%rdi), %ymm1
22169 ; NoVLX-NEXT: vcmpeqpd %zmm1, %zmm0, %k0
22170 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
22171 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
22172 ; NoVLX-NEXT: kmovw %k0, %eax
22173 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
22174 ; NoVLX-NEXT: vzeroupper
22175 ; NoVLX-NEXT: retq
22176 entry:
22177 %0 = bitcast <4 x i64> %__a to <4 x double>
22178 %load = load <4 x i64>, <4 x i64>* %__b
22179 %1 = bitcast <4 x i64> %load to <4 x double>
22180 %2 = fcmp oeq <4 x double> %0, %1
22181 %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
22182 %4 = bitcast <16 x i1> %3 to i16
22183 ret i16 %4
22184 }
22186 define zeroext i16 @test_vcmpoeqpd_v4i1_v16i1_mask_mem_b(<4 x i64> %__a, double* %__b) local_unnamed_addr {
22187 ; VLX-LABEL: test_vcmpoeqpd_v4i1_v16i1_mask_mem_b:
22188 ; VLX: # %bb.0: # %entry
22189 ; VLX-NEXT: vcmpeqpd (%rdi){1to4}, %ymm0, %k0
22190 ; VLX-NEXT: kmovd %k0, %eax
22191 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
22192 ; VLX-NEXT: vzeroupper
22193 ; VLX-NEXT: retq
22194 ;
22195 ; NoVLX-LABEL: test_vcmpoeqpd_v4i1_v16i1_mask_mem_b:
22196 ; NoVLX: # %bb.0: # %entry
22197 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
22198 ; NoVLX-NEXT: vcmpeqpd (%rdi){1to8}, %zmm0, %k0
22199 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
22200 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
22201 ; NoVLX-NEXT: kmovw %k0, %eax
22202 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
22203 ; NoVLX-NEXT: vzeroupper
22204 ; NoVLX-NEXT: retq
22205 entry:
22206 %0 = bitcast <4 x i64> %__a to <4 x double>
22207 %load = load double, double* %__b
22208 %vec = insertelement <4 x double> undef, double %load, i32 0
22209 %1 = shufflevector <4 x double> %vec, <4 x double> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
22210 %2 = fcmp oeq <4 x double> %0, %1
22211 %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
22212 %4 = bitcast <16 x i1> %3 to i16
22213 ret i16 %4
22214 }
22216 define zeroext i16 @test_masked_vcmpoeqpd_v4i1_v16i1_mask(i4 zeroext %__u, <4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
22217 ; VLX-LABEL: test_masked_vcmpoeqpd_v4i1_v16i1_mask:
22218 ; VLX: # %bb.0: # %entry
22219 ; VLX-NEXT: kmovd %edi, %k1
22220 ; VLX-NEXT: vcmpeqpd %ymm1, %ymm0, %k0 {%k1}
22221 ; VLX-NEXT: kmovd %k0, %eax
22222 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
22223 ; VLX-NEXT: vzeroupper
22224 ; VLX-NEXT: retq
22225 ;
22226 ; NoVLX-LABEL: test_masked_vcmpoeqpd_v4i1_v16i1_mask:
22227 ; NoVLX: # %bb.0: # %entry
22228 ; NoVLX-NEXT: # kill: def $ymm1 killed $ymm1 def $zmm1
22229 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
22230 ; NoVLX-NEXT: kmovw %edi, %k1
22231 ; NoVLX-NEXT: vcmpeqpd %zmm1, %zmm0, %k0 {%k1}
22232 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
22233 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
22234 ; NoVLX-NEXT: kmovw %k0, %eax
22235 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
22236 ; NoVLX-NEXT: vzeroupper
22237 ; NoVLX-NEXT: retq
22238 entry:
22239 %0 = bitcast <4 x i64> %__a to <4 x double>
22240 %1 = bitcast <4 x i64> %__b to <4 x double>
22241 %2 = fcmp oeq <4 x double> %0, %1
22242 %3 = bitcast i4 %__u to <4 x i1>
22243 %4 = and <4 x i1> %2, %3
22244 %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
22245 %6 = bitcast <16 x i1> %5 to i16
22246 ret i16 %6
22247 }
22249 define zeroext i16 @test_masked_vcmpoeqpd_v4i1_v16i1_mask_mem(i4 zeroext %__u, <4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
22250 ; VLX-LABEL: test_masked_vcmpoeqpd_v4i1_v16i1_mask_mem:
22251 ; VLX: # %bb.0: # %entry
22252 ; VLX-NEXT: kmovd %edi, %k1
22253 ; VLX-NEXT: vcmpeqpd (%rsi), %ymm0, %k0 {%k1}
22254 ; VLX-NEXT: kmovd %k0, %eax
22255 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
22256 ; VLX-NEXT: vzeroupper
22257 ; VLX-NEXT: retq
22258 ;
22259 ; NoVLX-LABEL: test_masked_vcmpoeqpd_v4i1_v16i1_mask_mem:
22260 ; NoVLX: # %bb.0: # %entry
22261 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
22262 ; NoVLX-NEXT: kmovw %edi, %k1
22263 ; NoVLX-NEXT: vmovapd (%rsi), %ymm1
22264 ; NoVLX-NEXT: vcmpeqpd %zmm1, %zmm0, %k0 {%k1}
22265 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
22266 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
22267 ; NoVLX-NEXT: kmovw %k0, %eax
22268 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
22269 ; NoVLX-NEXT: vzeroupper
22270 ; NoVLX-NEXT: retq
22271 entry:
22272 %0 = bitcast <4 x i64> %__a to <4 x double>
22273 %load = load <4 x i64>, <4 x i64>* %__b
22274 %1 = bitcast <4 x i64> %load to <4 x double>
22275 %2 = fcmp oeq <4 x double> %0, %1
22276 %3 = bitcast i4 %__u to <4 x i1>
22277 %4 = and <4 x i1> %2, %3
22278 %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
22279 %6 = bitcast <16 x i1> %5 to i16
22280 ret i16 %6
22281 }
22283 define zeroext i16 @test_masked_vcmpoeqpd_v4i1_v16i1_mask_mem_b(i4 zeroext %__u, <4 x i64> %__a, double* %__b) local_unnamed_addr {
22284 ; VLX-LABEL: test_masked_vcmpoeqpd_v4i1_v16i1_mask_mem_b:
22285 ; VLX: # %bb.0: # %entry
22286 ; VLX-NEXT: kmovd %edi, %k1
22287 ; VLX-NEXT: vcmpeqpd (%rsi){1to4}, %ymm0, %k0 {%k1}
22288 ; VLX-NEXT: kmovd %k0, %eax
22289 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
22290 ; VLX-NEXT: vzeroupper
22291 ; VLX-NEXT: retq
22292 ;
22293 ; NoVLX-LABEL: test_masked_vcmpoeqpd_v4i1_v16i1_mask_mem_b:
22294 ; NoVLX: # %bb.0: # %entry
22295 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
22296 ; NoVLX-NEXT: kmovw %edi, %k1
22297 ; NoVLX-NEXT: vcmpeqpd (%rsi){1to8}, %zmm0, %k0 {%k1}
22298 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
22299 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
22300 ; NoVLX-NEXT: kmovw %k0, %eax
22301 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
22302 ; NoVLX-NEXT: vzeroupper
22303 ; NoVLX-NEXT: retq
22304 entry:
22305 %0 = bitcast <4 x i64> %__a to <4 x double>
22306 %load = load double, double* %__b
22307 %vec = insertelement <4 x double> undef, double %load, i32 0
22308 %1 = shufflevector <4 x double> %vec, <4 x double> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
22309 %2 = fcmp oeq <4 x double> %0, %1
22310 %3 = bitcast i4 %__u to <4 x i1>
22311 %4 = and <4 x i1> %2, %3
22312 %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
22313 %6 = bitcast <16 x i1> %5 to i16
22314 ret i16 %6
22315 }
22319 define zeroext i32 @test_vcmpoeqpd_v4i1_v32i1_mask(<4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
22320 ; VLX-LABEL: test_vcmpoeqpd_v4i1_v32i1_mask:
22321 ; VLX: # %bb.0: # %entry
22322 ; VLX-NEXT: vcmpeqpd %ymm1, %ymm0, %k0
22323 ; VLX-NEXT: kmovd %k0, %eax
22324 ; VLX-NEXT: vzeroupper
22325 ; VLX-NEXT: retq
22326 ;
22327 ; NoVLX-LABEL: test_vcmpoeqpd_v4i1_v32i1_mask:
22328 ; NoVLX: # %bb.0: # %entry
22329 ; NoVLX-NEXT: # kill: def $ymm1 killed $ymm1 def $zmm1
22330 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
22331 ; NoVLX-NEXT: vcmpeqpd %zmm1, %zmm0, %k0
22332 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
22333 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
22334 ; NoVLX-NEXT: kmovw %k0, %eax
22335 ; NoVLX-NEXT: vzeroupper
22336 ; NoVLX-NEXT: retq
22337 entry:
22338 %0 = bitcast <4 x i64> %__a to <4 x double>
22339 %1 = bitcast <4 x i64> %__b to <4 x double>
22340 %2 = fcmp oeq <4 x double> %0, %1
22341 %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
22342 %4 = bitcast <32 x i1> %3 to i32
22343 ret i32 %4
22344 }
22346 define zeroext i32 @test_vcmpoeqpd_v4i1_v32i1_mask_mem(<4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
22347 ; VLX-LABEL: test_vcmpoeqpd_v4i1_v32i1_mask_mem:
22348 ; VLX: # %bb.0: # %entry
22349 ; VLX-NEXT: vcmpeqpd (%rdi), %ymm0, %k0
22350 ; VLX-NEXT: kmovd %k0, %eax
22351 ; VLX-NEXT: vzeroupper
22352 ; VLX-NEXT: retq
22353 ;
22354 ; NoVLX-LABEL: test_vcmpoeqpd_v4i1_v32i1_mask_mem:
22355 ; NoVLX: # %bb.0: # %entry
22356 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
22357 ; NoVLX-NEXT: vmovapd (%rdi), %ymm1
22358 ; NoVLX-NEXT: vcmpeqpd %zmm1, %zmm0, %k0
22359 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
22360 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
22361 ; NoVLX-NEXT: kmovw %k0, %eax
22362 ; NoVLX-NEXT: vzeroupper
22363 ; NoVLX-NEXT: retq
22364 entry:
22365 %0 = bitcast <4 x i64> %__a to <4 x double>
22366 %load = load <4 x i64>, <4 x i64>* %__b
22367 %1 = bitcast <4 x i64> %load to <4 x double>
22368 %2 = fcmp oeq <4 x double> %0, %1
22369 %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
22370 %4 = bitcast <32 x i1> %3 to i32
22371 ret i32 %4
22372 }
22374 define zeroext i32 @test_vcmpoeqpd_v4i1_v32i1_mask_mem_b(<4 x i64> %__a, double* %__b) local_unnamed_addr {
22375 ; VLX-LABEL: test_vcmpoeqpd_v4i1_v32i1_mask_mem_b:
22376 ; VLX: # %bb.0: # %entry
22377 ; VLX-NEXT: vcmpeqpd (%rdi){1to4}, %ymm0, %k0
22378 ; VLX-NEXT: kmovd %k0, %eax
22379 ; VLX-NEXT: vzeroupper
22380 ; VLX-NEXT: retq
22381 ;
22382 ; NoVLX-LABEL: test_vcmpoeqpd_v4i1_v32i1_mask_mem_b:
22383 ; NoVLX: # %bb.0: # %entry
22384 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
22385 ; NoVLX-NEXT: vcmpeqpd (%rdi){1to8}, %zmm0, %k0
22386 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
22387 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
22388 ; NoVLX-NEXT: kmovw %k0, %eax
22389 ; NoVLX-NEXT: vzeroupper
22390 ; NoVLX-NEXT: retq
22391 entry:
22392 %0 = bitcast <4 x i64> %__a to <4 x double>
22393 %load = load double, double* %__b
22394 %vec = insertelement <4 x double> undef, double %load, i32 0
22395 %1 = shufflevector <4 x double> %vec, <4 x double> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
22396 %2 = fcmp oeq <4 x double> %0, %1
22397 %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
22398 %4 = bitcast <32 x i1> %3 to i32
22399 ret i32 %4
22400 }
22402 define zeroext i32 @test_masked_vcmpoeqpd_v4i1_v32i1_mask(i4 zeroext %__u, <4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
22403 ; VLX-LABEL: test_masked_vcmpoeqpd_v4i1_v32i1_mask:
22404 ; VLX: # %bb.0: # %entry
22405 ; VLX-NEXT: kmovd %edi, %k1
22406 ; VLX-NEXT: vcmpeqpd %ymm1, %ymm0, %k0 {%k1}
22407 ; VLX-NEXT: kmovd %k0, %eax
22408 ; VLX-NEXT: vzeroupper
22409 ; VLX-NEXT: retq
22410 ;
22411 ; NoVLX-LABEL: test_masked_vcmpoeqpd_v4i1_v32i1_mask:
22412 ; NoVLX: # %bb.0: # %entry
22413 ; NoVLX-NEXT: # kill: def $ymm1 killed $ymm1 def $zmm1
22414 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
22415 ; NoVLX-NEXT: kmovw %edi, %k1
22416 ; NoVLX-NEXT: vcmpeqpd %zmm1, %zmm0, %k0 {%k1}
22417 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
22418 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
22419 ; NoVLX-NEXT: kmovw %k0, %eax
22420 ; NoVLX-NEXT: vzeroupper
22421 ; NoVLX-NEXT: retq
22422 entry:
22423 %0 = bitcast <4 x i64> %__a to <4 x double>
22424 %1 = bitcast <4 x i64> %__b to <4 x double>
22425 %2 = fcmp oeq <4 x double> %0, %1
22426 %3 = bitcast i4 %__u to <4 x i1>
22427 %4 = and <4 x i1> %2, %3
22428 %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
22429 %6 = bitcast <32 x i1> %5 to i32
22430 ret i32 %6
22431 }
22433 define zeroext i32 @test_masked_vcmpoeqpd_v4i1_v32i1_mask_mem(i4 zeroext %__u, <4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
22434 ; VLX-LABEL: test_masked_vcmpoeqpd_v4i1_v32i1_mask_mem:
22435 ; VLX: # %bb.0: # %entry
22436 ; VLX-NEXT: kmovd %edi, %k1
22437 ; VLX-NEXT: vcmpeqpd (%rsi), %ymm0, %k0 {%k1}
22438 ; VLX-NEXT: kmovd %k0, %eax
22439 ; VLX-NEXT: vzeroupper
22440 ; VLX-NEXT: retq
22441 ;
22442 ; NoVLX-LABEL: test_masked_vcmpoeqpd_v4i1_v32i1_mask_mem:
22443 ; NoVLX: # %bb.0: # %entry
22444 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
22445 ; NoVLX-NEXT: kmovw %edi, %k1
22446 ; NoVLX-NEXT: vmovapd (%rsi), %ymm1
22447 ; NoVLX-NEXT: vcmpeqpd %zmm1, %zmm0, %k0 {%k1}
22448 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
22449 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
22450 ; NoVLX-NEXT: kmovw %k0, %eax
22451 ; NoVLX-NEXT: vzeroupper
22452 ; NoVLX-NEXT: retq
22453 entry:
22454 %0 = bitcast <4 x i64> %__a to <4 x double>
22455 %load = load <4 x i64>, <4 x i64>* %__b
22456 %1 = bitcast <4 x i64> %load to <4 x double>
22457 %2 = fcmp oeq <4 x double> %0, %1
22458 %3 = bitcast i4 %__u to <4 x i1>
22459 %4 = and <4 x i1> %2, %3
22460 %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
22461 %6 = bitcast <32 x i1> %5 to i32
22462 ret i32 %6
22463 }
22465 define zeroext i32 @test_masked_vcmpoeqpd_v4i1_v32i1_mask_mem_b(i4 zeroext %__u, <4 x i64> %__a, double* %__b) local_unnamed_addr {
22466 ; VLX-LABEL: test_masked_vcmpoeqpd_v4i1_v32i1_mask_mem_b:
22467 ; VLX: # %bb.0: # %entry
22468 ; VLX-NEXT: kmovd %edi, %k1
22469 ; VLX-NEXT: vcmpeqpd (%rsi){1to4}, %ymm0, %k0 {%k1}
22470 ; VLX-NEXT: kmovd %k0, %eax
22471 ; VLX-NEXT: vzeroupper
22472 ; VLX-NEXT: retq
22473 ;
22474 ; NoVLX-LABEL: test_masked_vcmpoeqpd_v4i1_v32i1_mask_mem_b:
22475 ; NoVLX: # %bb.0: # %entry
22476 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
22477 ; NoVLX-NEXT: kmovw %edi, %k1
22478 ; NoVLX-NEXT: vcmpeqpd (%rsi){1to8}, %zmm0, %k0 {%k1}
22479 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
22480 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
22481 ; NoVLX-NEXT: kmovw %k0, %eax
22482 ; NoVLX-NEXT: vzeroupper
22483 ; NoVLX-NEXT: retq
22484 entry:
22485 %0 = bitcast <4 x i64> %__a to <4 x double>
22486 %load = load double, double* %__b
22487 %vec = insertelement <4 x double> undef, double %load, i32 0
22488 %1 = shufflevector <4 x double> %vec, <4 x double> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
22489 %2 = fcmp oeq <4 x double> %0, %1
22490 %3 = bitcast i4 %__u to <4 x i1>
22491 %4 = and <4 x i1> %2, %3
22492 %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
22493 %6 = bitcast <32 x i1> %5 to i32
22494 ret i32 %6
22495 }
22499 define zeroext i64 @test_vcmpoeqpd_v4i1_v64i1_mask(<4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
22500 ; VLX-LABEL: test_vcmpoeqpd_v4i1_v64i1_mask:
22501 ; VLX: # %bb.0: # %entry
22502 ; VLX-NEXT: vcmpeqpd %ymm1, %ymm0, %k0
22503 ; VLX-NEXT: kmovq %k0, %rax
22504 ; VLX-NEXT: vzeroupper
22505 ; VLX-NEXT: retq
22506 ;
22507 ; NoVLX-LABEL: test_vcmpoeqpd_v4i1_v64i1_mask:
22508 ; NoVLX: # %bb.0: # %entry
22509 ; NoVLX-NEXT: # kill: def $ymm1 killed $ymm1 def $zmm1
22510 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
22511 ; NoVLX-NEXT: vcmpeqpd %zmm1, %zmm0, %k0
22512 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
22513 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
22514 ; NoVLX-NEXT: kmovw %k0, %eax
22515 ; NoVLX-NEXT: vzeroupper
22516 ; NoVLX-NEXT: retq
22517 entry:
22518 %0 = bitcast <4 x i64> %__a to <4 x double>
22519 %1 = bitcast <4 x i64> %__b to <4 x double>
22520 %2 = fcmp oeq <4 x double> %0, %1
22521 %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
22522 %4 = bitcast <64 x i1> %3 to i64
22523 ret i64 %4
22524 }
22526 define zeroext i64 @test_vcmpoeqpd_v4i1_v64i1_mask_mem(<4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
22527 ; VLX-LABEL: test_vcmpoeqpd_v4i1_v64i1_mask_mem:
22528 ; VLX: # %bb.0: # %entry
22529 ; VLX-NEXT: vcmpeqpd (%rdi), %ymm0, %k0
22530 ; VLX-NEXT: kmovq %k0, %rax
22531 ; VLX-NEXT: vzeroupper
22532 ; VLX-NEXT: retq
22533 ;
22534 ; NoVLX-LABEL: test_vcmpoeqpd_v4i1_v64i1_mask_mem:
22535 ; NoVLX: # %bb.0: # %entry
22536 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
22537 ; NoVLX-NEXT: vmovapd (%rdi), %ymm1
22538 ; NoVLX-NEXT: vcmpeqpd %zmm1, %zmm0, %k0
22539 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
22540 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
22541 ; NoVLX-NEXT: kmovw %k0, %eax
22542 ; NoVLX-NEXT: vzeroupper
22543 ; NoVLX-NEXT: retq
22544 entry:
22545 %0 = bitcast <4 x i64> %__a to <4 x double>
22546 %load = load <4 x i64>, <4 x i64>* %__b
22547 %1 = bitcast <4 x i64> %load to <4 x double>
22548 %2 = fcmp oeq <4 x double> %0, %1
22549 %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
22550 %4 = bitcast <64 x i1> %3 to i64
22551 ret i64 %4
22552 }
22554 define zeroext i64 @test_vcmpoeqpd_v4i1_v64i1_mask_mem_b(<4 x i64> %__a, double* %__b) local_unnamed_addr {
22555 ; VLX-LABEL: test_vcmpoeqpd_v4i1_v64i1_mask_mem_b:
22556 ; VLX: # %bb.0: # %entry
22557 ; VLX-NEXT: vcmpeqpd (%rdi){1to4}, %ymm0, %k0
22558 ; VLX-NEXT: kmovq %k0, %rax
22559 ; VLX-NEXT: vzeroupper
22560 ; VLX-NEXT: retq
22561 ;
22562 ; NoVLX-LABEL: test_vcmpoeqpd_v4i1_v64i1_mask_mem_b:
22563 ; NoVLX: # %bb.0: # %entry
22564 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
22565 ; NoVLX-NEXT: vcmpeqpd (%rdi){1to8}, %zmm0, %k0
22566 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
22567 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
22568 ; NoVLX-NEXT: kmovw %k0, %eax
22569 ; NoVLX-NEXT: vzeroupper
22570 ; NoVLX-NEXT: retq
22571 entry:
22572 %0 = bitcast <4 x i64> %__a to <4 x double>
22573 %load = load double, double* %__b
22574 %vec = insertelement <4 x double> undef, double %load, i32 0
22575 %1 = shufflevector <4 x double> %vec, <4 x double> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
22576 %2 = fcmp oeq <4 x double> %0, %1
22577 %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
22578 %4 = bitcast <64 x i1> %3 to i64
22579 ret i64 %4
22580 }
22582 define zeroext i64 @test_masked_vcmpoeqpd_v4i1_v64i1_mask(i4 zeroext %__u, <4 x i64> %__a, <4 x i64> %__b) local_unnamed_addr {
22583 ; VLX-LABEL: test_masked_vcmpoeqpd_v4i1_v64i1_mask:
22584 ; VLX: # %bb.0: # %entry
22585 ; VLX-NEXT: kmovd %edi, %k1
22586 ; VLX-NEXT: vcmpeqpd %ymm1, %ymm0, %k0 {%k1}
22587 ; VLX-NEXT: kmovq %k0, %rax
22588 ; VLX-NEXT: vzeroupper
22589 ; VLX-NEXT: retq
22590 ;
22591 ; NoVLX-LABEL: test_masked_vcmpoeqpd_v4i1_v64i1_mask:
22592 ; NoVLX: # %bb.0: # %entry
22593 ; NoVLX-NEXT: # kill: def $ymm1 killed $ymm1 def $zmm1
22594 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
22595 ; NoVLX-NEXT: kmovw %edi, %k1
22596 ; NoVLX-NEXT: vcmpeqpd %zmm1, %zmm0, %k0 {%k1}
22597 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
22598 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
22599 ; NoVLX-NEXT: kmovw %k0, %eax
22600 ; NoVLX-NEXT: vzeroupper
22601 ; NoVLX-NEXT: retq
22602 entry:
22603 %0 = bitcast <4 x i64> %__a to <4 x double>
22604 %1 = bitcast <4 x i64> %__b to <4 x double>
22605 %2 = fcmp oeq <4 x double> %0, %1
22606 %3 = bitcast i4 %__u to <4 x i1>
22607 %4 = and <4 x i1> %2, %3
22608 %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
22609 %6 = bitcast <64 x i1> %5 to i64
22610 ret i64 %6
22611 }
22613 define zeroext i64 @test_masked_vcmpoeqpd_v4i1_v64i1_mask_mem(i4 zeroext %__u, <4 x i64> %__a, <4 x i64>* %__b) local_unnamed_addr {
22614 ; VLX-LABEL: test_masked_vcmpoeqpd_v4i1_v64i1_mask_mem:
22615 ; VLX: # %bb.0: # %entry
22616 ; VLX-NEXT: kmovd %edi, %k1
22617 ; VLX-NEXT: vcmpeqpd (%rsi), %ymm0, %k0 {%k1}
22618 ; VLX-NEXT: kmovq %k0, %rax
22619 ; VLX-NEXT: vzeroupper
22620 ; VLX-NEXT: retq
22621 ;
22622 ; NoVLX-LABEL: test_masked_vcmpoeqpd_v4i1_v64i1_mask_mem:
22623 ; NoVLX: # %bb.0: # %entry
22624 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
22625 ; NoVLX-NEXT: kmovw %edi, %k1
22626 ; NoVLX-NEXT: vmovapd (%rsi), %ymm1
22627 ; NoVLX-NEXT: vcmpeqpd %zmm1, %zmm0, %k0 {%k1}
22628 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
22629 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
22630 ; NoVLX-NEXT: kmovw %k0, %eax
22631 ; NoVLX-NEXT: vzeroupper
22632 ; NoVLX-NEXT: retq
22633 entry:
22634 %0 = bitcast <4 x i64> %__a to <4 x double>
22635 %load = load <4 x i64>, <4 x i64>* %__b
22636 %1 = bitcast <4 x i64> %load to <4 x double>
22637 %2 = fcmp oeq <4 x double> %0, %1
22638 %3 = bitcast i4 %__u to <4 x i1>
22639 %4 = and <4 x i1> %2, %3
22640 %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
22641 %6 = bitcast <64 x i1> %5 to i64
22642 ret i64 %6
22643 }
22645 define zeroext i64 @test_masked_vcmpoeqpd_v4i1_v64i1_mask_mem_b(i4 zeroext %__u, <4 x i64> %__a, double* %__b) local_unnamed_addr {
22646 ; VLX-LABEL: test_masked_vcmpoeqpd_v4i1_v64i1_mask_mem_b:
22647 ; VLX: # %bb.0: # %entry
22648 ; VLX-NEXT: kmovd %edi, %k1
22649 ; VLX-NEXT: vcmpeqpd (%rsi){1to4}, %ymm0, %k0 {%k1}
22650 ; VLX-NEXT: kmovq %k0, %rax
22651 ; VLX-NEXT: vzeroupper
22652 ; VLX-NEXT: retq
22653 ;
22654 ; NoVLX-LABEL: test_masked_vcmpoeqpd_v4i1_v64i1_mask_mem_b:
22655 ; NoVLX: # %bb.0: # %entry
22656 ; NoVLX-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
22657 ; NoVLX-NEXT: kmovw %edi, %k1
22658 ; NoVLX-NEXT: vcmpeqpd (%rsi){1to8}, %zmm0, %k0 {%k1}
22659 ; NoVLX-NEXT: kshiftlw $12, %k0, %k0
22660 ; NoVLX-NEXT: kshiftrw $12, %k0, %k0
22661 ; NoVLX-NEXT: kmovw %k0, %eax
22662 ; NoVLX-NEXT: vzeroupper
22663 ; NoVLX-NEXT: retq
22664 entry:
22665 %0 = bitcast <4 x i64> %__a to <4 x double>
22666 %load = load double, double* %__b
22667 %vec = insertelement <4 x double> undef, double %load, i32 0
22668 %1 = shufflevector <4 x double> %vec, <4 x double> undef, <4 x i32> <i32 0, i32 0, i32 0, i32 0>
22669 %2 = fcmp oeq <4 x double> %0, %1
22670 %3 = bitcast i4 %__u to <4 x i1>
22671 %4 = and <4 x i1> %2, %3
22672 %5 = shufflevector <4 x i1> %4, <4 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7, i32 4, i32 5, i32 6, i32 7>
22673 %6 = bitcast <64 x i1> %5 to i64
22674 ret i64 %6
22675 }
22679 define zeroext i16 @test_vcmpoeqpd_v8i1_v16i1_mask(<8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
22680 ; VLX-LABEL: test_vcmpoeqpd_v8i1_v16i1_mask:
22681 ; VLX: # %bb.0: # %entry
22682 ; VLX-NEXT: vcmpeqpd %zmm1, %zmm0, %k0
22683 ; VLX-NEXT: kmovd %k0, %eax
22684 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
22685 ; VLX-NEXT: vzeroupper
22686 ; VLX-NEXT: retq
22687 ;
22688 ; NoVLX-LABEL: test_vcmpoeqpd_v8i1_v16i1_mask:
22689 ; NoVLX: # %bb.0: # %entry
22690 ; NoVLX-NEXT: vcmpeqpd %zmm1, %zmm0, %k0
22691 ; NoVLX-NEXT: kmovw %k0, %eax
22692 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
22693 ; NoVLX-NEXT: vzeroupper
22694 ; NoVLX-NEXT: retq
22695 entry:
22696 %0 = bitcast <8 x i64> %__a to <8 x double>
22697 %1 = bitcast <8 x i64> %__b to <8 x double>
22698 %2 = fcmp oeq <8 x double> %0, %1
22699 %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
22700 %4 = bitcast <16 x i1> %3 to i16
22701 ret i16 %4
22702 }
22704 define zeroext i16 @test_vcmpoeqpd_v8i1_v16i1_mask_mem(<8 x i64> %__a, <8 x i64>* %__b) local_unnamed_addr {
22705 ; VLX-LABEL: test_vcmpoeqpd_v8i1_v16i1_mask_mem:
22706 ; VLX: # %bb.0: # %entry
22707 ; VLX-NEXT: vcmpeqpd (%rdi), %zmm0, %k0
22708 ; VLX-NEXT: kmovd %k0, %eax
22709 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
22710 ; VLX-NEXT: vzeroupper
22711 ; VLX-NEXT: retq
22712 ;
22713 ; NoVLX-LABEL: test_vcmpoeqpd_v8i1_v16i1_mask_mem:
22714 ; NoVLX: # %bb.0: # %entry
22715 ; NoVLX-NEXT: vcmpeqpd (%rdi), %zmm0, %k0
22716 ; NoVLX-NEXT: kmovw %k0, %eax
22717 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
22718 ; NoVLX-NEXT: vzeroupper
22719 ; NoVLX-NEXT: retq
22720 entry:
22721 %0 = bitcast <8 x i64> %__a to <8 x double>
22722 %load = load <8 x i64>, <8 x i64>* %__b
22723 %1 = bitcast <8 x i64> %load to <8 x double>
22724 %2 = fcmp oeq <8 x double> %0, %1
22725 %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
22726 %4 = bitcast <16 x i1> %3 to i16
22727 ret i16 %4
22728 }
22730 define zeroext i16 @test_vcmpoeqpd_v8i1_v16i1_mask_mem_b(<8 x i64> %__a, double* %__b) local_unnamed_addr {
22731 ; VLX-LABEL: test_vcmpoeqpd_v8i1_v16i1_mask_mem_b:
22732 ; VLX: # %bb.0: # %entry
22733 ; VLX-NEXT: vcmpeqpd (%rdi){1to8}, %zmm0, %k0
22734 ; VLX-NEXT: kmovd %k0, %eax
22735 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
22736 ; VLX-NEXT: vzeroupper
22737 ; VLX-NEXT: retq
22738 ;
22739 ; NoVLX-LABEL: test_vcmpoeqpd_v8i1_v16i1_mask_mem_b:
22740 ; NoVLX: # %bb.0: # %entry
22741 ; NoVLX-NEXT: vcmpeqpd (%rdi){1to8}, %zmm0, %k0
22742 ; NoVLX-NEXT: kmovw %k0, %eax
22743 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
22744 ; NoVLX-NEXT: vzeroupper
22745 ; NoVLX-NEXT: retq
22746 entry:
22747 %0 = bitcast <8 x i64> %__a to <8 x double>
22748 %load = load double, double* %__b
22749 %vec = insertelement <8 x double> undef, double %load, i32 0
22750 %1 = shufflevector <8 x double> %vec, <8 x double> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
22751 %2 = fcmp oeq <8 x double> %0, %1
22752 %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
22753 %4 = bitcast <16 x i1> %3 to i16
22754 ret i16 %4
22755 }
22757 define zeroext i16 @test_masked_vcmpoeqpd_v8i1_v16i1_mask(i8 zeroext %__u, <8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
22758 ; VLX-LABEL: test_masked_vcmpoeqpd_v8i1_v16i1_mask:
22759 ; VLX: # %bb.0: # %entry
22760 ; VLX-NEXT: kmovd %edi, %k1
22761 ; VLX-NEXT: vcmpeqpd %zmm1, %zmm0, %k0 {%k1}
22762 ; VLX-NEXT: kmovd %k0, %eax
22763 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
22764 ; VLX-NEXT: vzeroupper
22765 ; VLX-NEXT: retq
22766 ;
22767 ; NoVLX-LABEL: test_masked_vcmpoeqpd_v8i1_v16i1_mask:
22768 ; NoVLX: # %bb.0: # %entry
22769 ; NoVLX-NEXT: kmovw %edi, %k1
22770 ; NoVLX-NEXT: vcmpeqpd %zmm1, %zmm0, %k0 {%k1}
22771 ; NoVLX-NEXT: kmovw %k0, %eax
22772 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
22773 ; NoVLX-NEXT: vzeroupper
22774 ; NoVLX-NEXT: retq
22775 entry:
22776 %0 = bitcast <8 x i64> %__a to <8 x double>
22777 %1 = bitcast <8 x i64> %__b to <8 x double>
22778 %2 = fcmp oeq <8 x double> %0, %1
22779 %3 = bitcast i8 %__u to <8 x i1>
22780 %4 = and <8 x i1> %2, %3
22781 %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
22782 %6 = bitcast <16 x i1> %5 to i16
22783 ret i16 %6
22784 }
22786 define zeroext i16 @test_masked_vcmpoeqpd_v8i1_v16i1_mask_mem(i8 zeroext %__u, <8 x i64> %__a, <8 x i64>* %__b) local_unnamed_addr {
22787 ; VLX-LABEL: test_masked_vcmpoeqpd_v8i1_v16i1_mask_mem:
22788 ; VLX: # %bb.0: # %entry
22789 ; VLX-NEXT: kmovd %edi, %k1
22790 ; VLX-NEXT: vcmpeqpd (%rsi), %zmm0, %k0 {%k1}
22791 ; VLX-NEXT: kmovd %k0, %eax
22792 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
22793 ; VLX-NEXT: vzeroupper
22794 ; VLX-NEXT: retq
22795 ;
22796 ; NoVLX-LABEL: test_masked_vcmpoeqpd_v8i1_v16i1_mask_mem:
22797 ; NoVLX: # %bb.0: # %entry
22798 ; NoVLX-NEXT: kmovw %edi, %k1
22799 ; NoVLX-NEXT: vcmpeqpd (%rsi), %zmm0, %k0 {%k1}
22800 ; NoVLX-NEXT: kmovw %k0, %eax
22801 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
22802 ; NoVLX-NEXT: vzeroupper
22803 ; NoVLX-NEXT: retq
22804 entry:
22805 %0 = bitcast <8 x i64> %__a to <8 x double>
22806 %load = load <8 x i64>, <8 x i64>* %__b
22807 %1 = bitcast <8 x i64> %load to <8 x double>
22808 %2 = fcmp oeq <8 x double> %0, %1
22809 %3 = bitcast i8 %__u to <8 x i1>
22810 %4 = and <8 x i1> %2, %3
22811 %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
22812 %6 = bitcast <16 x i1> %5 to i16
22813 ret i16 %6
22814 }
22816 define zeroext i16 @test_masked_vcmpoeqpd_v8i1_v16i1_mask_mem_b(i8 zeroext %__u, <8 x i64> %__a, double* %__b) local_unnamed_addr {
22817 ; VLX-LABEL: test_masked_vcmpoeqpd_v8i1_v16i1_mask_mem_b:
22818 ; VLX: # %bb.0: # %entry
22819 ; VLX-NEXT: kmovd %edi, %k1
22820 ; VLX-NEXT: vcmpeqpd (%rsi){1to8}, %zmm0, %k0 {%k1}
22821 ; VLX-NEXT: kmovd %k0, %eax
22822 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
22823 ; VLX-NEXT: vzeroupper
22824 ; VLX-NEXT: retq
22825 ;
22826 ; NoVLX-LABEL: test_masked_vcmpoeqpd_v8i1_v16i1_mask_mem_b:
22827 ; NoVLX: # %bb.0: # %entry
22828 ; NoVLX-NEXT: kmovw %edi, %k1
22829 ; NoVLX-NEXT: vcmpeqpd (%rsi){1to8}, %zmm0, %k0 {%k1}
22830 ; NoVLX-NEXT: kmovw %k0, %eax
22831 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
22832 ; NoVLX-NEXT: vzeroupper
22833 ; NoVLX-NEXT: retq
22834 entry:
22835 %0 = bitcast <8 x i64> %__a to <8 x double>
22836 %load = load double, double* %__b
22837 %vec = insertelement <8 x double> undef, double %load, i32 0
22838 %1 = shufflevector <8 x double> %vec, <8 x double> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
22839 %2 = fcmp oeq <8 x double> %0, %1
22840 %3 = bitcast i8 %__u to <8 x i1>
22841 %4 = and <8 x i1> %2, %3
22842 %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
22843 %6 = bitcast <16 x i1> %5 to i16
22844 ret i16 %6
22845 }
22849 define zeroext i16 @test_vcmpoeqpd_v8i1_v16i1_sae_mask(<8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
22850 ; VLX-LABEL: test_vcmpoeqpd_v8i1_v16i1_sae_mask:
22851 ; VLX: # %bb.0: # %entry
22852 ; VLX-NEXT: vcmplepd {sae}, %zmm1, %zmm0, %k0
22853 ; VLX-NEXT: kmovd %k0, %eax
22854 ; VLX-NEXT: movzbl %al, %eax
22855 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
22856 ; VLX-NEXT: vzeroupper
22857 ; VLX-NEXT: retq
22858 ;
22859 ; NoVLX-LABEL: test_vcmpoeqpd_v8i1_v16i1_sae_mask:
22860 ; NoVLX: # %bb.0: # %entry
22861 ; NoVLX-NEXT: vcmplepd {sae}, %zmm1, %zmm0, %k0
22862 ; NoVLX-NEXT: kmovw %k0, %eax
22863 ; NoVLX-NEXT: movzbl %al, %eax
22864 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
22865 ; NoVLX-NEXT: vzeroupper
22866 ; NoVLX-NEXT: retq
22867 entry:
22868 %0 = bitcast <8 x i64> %__a to <8 x double>
22869 %1 = bitcast <8 x i64> %__b to <8 x double>
22870 %2 = call <8 x i1> @llvm.x86.avx512.cmp.pd.512(<8 x double> %0, <8 x double> %1, i32 2, i32 8)
22871 %3 = bitcast <8 x i1> %2 to i8
22872 %4 = zext i8 %3 to i16
22873 ret i16 %4
22874 }
22876 define zeroext i16 @test_masked_vcmpoeqpd_v8i1_v16i1_sae_mask(i8 zeroext %__u, <8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
22877 ; VLX-LABEL: test_masked_vcmpoeqpd_v8i1_v16i1_sae_mask:
22878 ; VLX: # %bb.0: # %entry
22879 ; VLX-NEXT: vcmplepd {sae}, %zmm1, %zmm0, %k0
22880 ; VLX-NEXT: kmovd %k0, %eax
22881 ; VLX-NEXT: andb %dil, %al
22882 ; VLX-NEXT: movzbl %al, %eax
22883 ; VLX-NEXT: # kill: def $ax killed $ax killed $eax
22884 ; VLX-NEXT: vzeroupper
22885 ; VLX-NEXT: retq
22886 ;
22887 ; NoVLX-LABEL: test_masked_vcmpoeqpd_v8i1_v16i1_sae_mask:
22888 ; NoVLX: # %bb.0: # %entry
22889 ; NoVLX-NEXT: vcmplepd {sae}, %zmm1, %zmm0, %k0
22890 ; NoVLX-NEXT: kmovw %k0, %eax
22891 ; NoVLX-NEXT: andb %dil, %al
22892 ; NoVLX-NEXT: movzbl %al, %eax
22893 ; NoVLX-NEXT: # kill: def $ax killed $ax killed $eax
22894 ; NoVLX-NEXT: vzeroupper
22895 ; NoVLX-NEXT: retq
22896 entry:
22897 %0 = bitcast <8 x i64> %__a to <8 x double>
22898 %1 = bitcast <8 x i64> %__b to <8 x double>
22899 %2 = call <8 x i1> @llvm.x86.avx512.cmp.pd.512(<8 x double> %0, <8 x double> %1, i32 2, i32 8)
22900 %3 = bitcast i8 %__u to <8 x i1>
22901 %4 = and <8 x i1> %2, %3
22902 %5 = bitcast <8 x i1> %4 to i8
22903 %6 = zext i8 %5 to i16
22904 ret i16 %6
22905 }
22909 define zeroext i32 @test_vcmpoeqpd_v8i1_v32i1_mask(<8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
22910 ; VLX-LABEL: test_vcmpoeqpd_v8i1_v32i1_mask:
22911 ; VLX: # %bb.0: # %entry
22912 ; VLX-NEXT: vcmpeqpd %zmm1, %zmm0, %k0
22913 ; VLX-NEXT: kmovd %k0, %eax
22914 ; VLX-NEXT: vzeroupper
22915 ; VLX-NEXT: retq
22916 ;
22917 ; NoVLX-LABEL: test_vcmpoeqpd_v8i1_v32i1_mask:
22918 ; NoVLX: # %bb.0: # %entry
22919 ; NoVLX-NEXT: vcmpeqpd %zmm1, %zmm0, %k0
22920 ; NoVLX-NEXT: kmovw %k0, %eax
22921 ; NoVLX-NEXT: vzeroupper
22922 ; NoVLX-NEXT: retq
22923 entry:
22924 %0 = bitcast <8 x i64> %__a to <8 x double>
22925 %1 = bitcast <8 x i64> %__b to <8 x double>
22926 %2 = fcmp oeq <8 x double> %0, %1
22927 %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
22928 %4 = bitcast <32 x i1> %3 to i32
22929 ret i32 %4
22930 }
22932 define zeroext i32 @test_vcmpoeqpd_v8i1_v32i1_mask_mem(<8 x i64> %__a, <8 x i64>* %__b) local_unnamed_addr {
22933 ; VLX-LABEL: test_vcmpoeqpd_v8i1_v32i1_mask_mem:
22934 ; VLX: # %bb.0: # %entry
22935 ; VLX-NEXT: vcmpeqpd (%rdi), %zmm0, %k0
22936 ; VLX-NEXT: kmovd %k0, %eax
22937 ; VLX-NEXT: vzeroupper
22938 ; VLX-NEXT: retq
22939 ;
22940 ; NoVLX-LABEL: test_vcmpoeqpd_v8i1_v32i1_mask_mem:
22941 ; NoVLX: # %bb.0: # %entry
22942 ; NoVLX-NEXT: vcmpeqpd (%rdi), %zmm0, %k0
22943 ; NoVLX-NEXT: kmovw %k0, %eax
22944 ; NoVLX-NEXT: vzeroupper
22945 ; NoVLX-NEXT: retq
22946 entry:
22947 %0 = bitcast <8 x i64> %__a to <8 x double>
22948 %load = load <8 x i64>, <8 x i64>* %__b
22949 %1 = bitcast <8 x i64> %load to <8 x double>
22950 %2 = fcmp oeq <8 x double> %0, %1
22951 %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
22952 %4 = bitcast <32 x i1> %3 to i32
22953 ret i32 %4
22954 }
22956 define zeroext i32 @test_vcmpoeqpd_v8i1_v32i1_mask_mem_b(<8 x i64> %__a, double* %__b) local_unnamed_addr {
22957 ; VLX-LABEL: test_vcmpoeqpd_v8i1_v32i1_mask_mem_b:
22958 ; VLX: # %bb.0: # %entry
22959 ; VLX-NEXT: vcmpeqpd (%rdi){1to8}, %zmm0, %k0
22960 ; VLX-NEXT: kmovd %k0, %eax
22961 ; VLX-NEXT: vzeroupper
22962 ; VLX-NEXT: retq
22963 ;
22964 ; NoVLX-LABEL: test_vcmpoeqpd_v8i1_v32i1_mask_mem_b:
22965 ; NoVLX: # %bb.0: # %entry
22966 ; NoVLX-NEXT: vcmpeqpd (%rdi){1to8}, %zmm0, %k0
22967 ; NoVLX-NEXT: kmovw %k0, %eax
22968 ; NoVLX-NEXT: vzeroupper
22969 ; NoVLX-NEXT: retq
22970 entry:
22971 %0 = bitcast <8 x i64> %__a to <8 x double>
22972 %load = load double, double* %__b
22973 %vec = insertelement <8 x double> undef, double %load, i32 0
22974 %1 = shufflevector <8 x double> %vec, <8 x double> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
22975 %2 = fcmp oeq <8 x double> %0, %1
22976 %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
22977 %4 = bitcast <32 x i1> %3 to i32
22978 ret i32 %4
22979 }
22981 define zeroext i32 @test_masked_vcmpoeqpd_v8i1_v32i1_mask(i8 zeroext %__u, <8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
22982 ; VLX-LABEL: test_masked_vcmpoeqpd_v8i1_v32i1_mask:
22983 ; VLX: # %bb.0: # %entry
22984 ; VLX-NEXT: kmovd %edi, %k1
22985 ; VLX-NEXT: vcmpeqpd %zmm1, %zmm0, %k0 {%k1}
22986 ; VLX-NEXT: kmovd %k0, %eax
22987 ; VLX-NEXT: vzeroupper
22988 ; VLX-NEXT: retq
22989 ;
22990 ; NoVLX-LABEL: test_masked_vcmpoeqpd_v8i1_v32i1_mask:
22991 ; NoVLX: # %bb.0: # %entry
22992 ; NoVLX-NEXT: kmovw %edi, %k1
22993 ; NoVLX-NEXT: vcmpeqpd %zmm1, %zmm0, %k0 {%k1}
22994 ; NoVLX-NEXT: kmovw %k0, %eax
22995 ; NoVLX-NEXT: vzeroupper
22996 ; NoVLX-NEXT: retq
22997 entry:
22998 %0 = bitcast <8 x i64> %__a to <8 x double>
22999 %1 = bitcast <8 x i64> %__b to <8 x double>
23000 %2 = fcmp oeq <8 x double> %0, %1
23001 %3 = bitcast i8 %__u to <8 x i1>
23002 %4 = and <8 x i1> %2, %3
23003 %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
23004 %6 = bitcast <32 x i1> %5 to i32
23005 ret i32 %6
23006 }
23008 define zeroext i32 @test_masked_vcmpoeqpd_v8i1_v32i1_mask_mem(i8 zeroext %__u, <8 x i64> %__a, <8 x i64>* %__b) local_unnamed_addr {
23009 ; VLX-LABEL: test_masked_vcmpoeqpd_v8i1_v32i1_mask_mem:
23010 ; VLX: # %bb.0: # %entry
23011 ; VLX-NEXT: kmovd %edi, %k1
23012 ; VLX-NEXT: vcmpeqpd (%rsi), %zmm0, %k0 {%k1}
23013 ; VLX-NEXT: kmovd %k0, %eax
23014 ; VLX-NEXT: vzeroupper
23015 ; VLX-NEXT: retq
23016 ;
23017 ; NoVLX-LABEL: test_masked_vcmpoeqpd_v8i1_v32i1_mask_mem:
23018 ; NoVLX: # %bb.0: # %entry
23019 ; NoVLX-NEXT: kmovw %edi, %k1
23020 ; NoVLX-NEXT: vcmpeqpd (%rsi), %zmm0, %k0 {%k1}
23021 ; NoVLX-NEXT: kmovw %k0, %eax
23022 ; NoVLX-NEXT: vzeroupper
23023 ; NoVLX-NEXT: retq
23024 entry:
23025 %0 = bitcast <8 x i64> %__a to <8 x double>
23026 %load = load <8 x i64>, <8 x i64>* %__b
23027 %1 = bitcast <8 x i64> %load to <8 x double>
23028 %2 = fcmp oeq <8 x double> %0, %1
23029 %3 = bitcast i8 %__u to <8 x i1>
23030 %4 = and <8 x i1> %2, %3
23031 %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
23032 %6 = bitcast <32 x i1> %5 to i32
23033 ret i32 %6
23034 }
23036 define zeroext i32 @test_masked_vcmpoeqpd_v8i1_v32i1_mask_mem_b(i8 zeroext %__u, <8 x i64> %__a, double* %__b) local_unnamed_addr {
23037 ; VLX-LABEL: test_masked_vcmpoeqpd_v8i1_v32i1_mask_mem_b:
23038 ; VLX: # %bb.0: # %entry
23039 ; VLX-NEXT: kmovd %edi, %k1
23040 ; VLX-NEXT: vcmpeqpd (%rsi){1to8}, %zmm0, %k0 {%k1}
23041 ; VLX-NEXT: kmovd %k0, %eax
23042 ; VLX-NEXT: vzeroupper
23043 ; VLX-NEXT: retq
23044 ;
23045 ; NoVLX-LABEL: test_masked_vcmpoeqpd_v8i1_v32i1_mask_mem_b:
23046 ; NoVLX: # %bb.0: # %entry
23047 ; NoVLX-NEXT: kmovw %edi, %k1
23048 ; NoVLX-NEXT: vcmpeqpd (%rsi){1to8}, %zmm0, %k0 {%k1}
23049 ; NoVLX-NEXT: kmovw %k0, %eax
23050 ; NoVLX-NEXT: vzeroupper
23051 ; NoVLX-NEXT: retq
23052 entry:
23053 %0 = bitcast <8 x i64> %__a to <8 x double>
23054 %load = load double, double* %__b
23055 %vec = insertelement <8 x double> undef, double %load, i32 0
23056 %1 = shufflevector <8 x double> %vec, <8 x double> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
23057 %2 = fcmp oeq <8 x double> %0, %1
23058 %3 = bitcast i8 %__u to <8 x i1>
23059 %4 = and <8 x i1> %2, %3
23060 %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
23061 %6 = bitcast <32 x i1> %5 to i32
23062 ret i32 %6
23063 }
23067 define zeroext i32 @test_vcmpoeqpd_v8i1_v32i1_sae_mask(<8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
23068 ; VLX-LABEL: test_vcmpoeqpd_v8i1_v32i1_sae_mask:
23069 ; VLX: # %bb.0: # %entry
23070 ; VLX-NEXT: vcmplepd {sae}, %zmm1, %zmm0, %k0
23071 ; VLX-NEXT: kmovb %k0, %eax
23072 ; VLX-NEXT: vzeroupper
23073 ; VLX-NEXT: retq
23074 ;
23075 ; NoVLX-LABEL: test_vcmpoeqpd_v8i1_v32i1_sae_mask:
23076 ; NoVLX: # %bb.0: # %entry
23077 ; NoVLX-NEXT: vcmplepd {sae}, %zmm1, %zmm0, %k0
23078 ; NoVLX-NEXT: kmovw %k0, %eax
23079 ; NoVLX-NEXT: movzbl %al, %eax
23080 ; NoVLX-NEXT: vzeroupper
23081 ; NoVLX-NEXT: retq
23082 entry:
23083 %0 = bitcast <8 x i64> %__a to <8 x double>
23084 %1 = bitcast <8 x i64> %__b to <8 x double>
23085 %2 = call <8 x i1> @llvm.x86.avx512.cmp.pd.512(<8 x double> %0, <8 x double> %1, i32 2, i32 8)
23086 %3 = bitcast <8 x i1> %2 to i8
23087 %4 = zext i8 %3 to i32
23088 ret i32 %4
23089 }
23091 define zeroext i32 @test_masked_vcmpoeqpd_v8i1_v32i1_sae_mask(i8 zeroext %__u, <8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
23092 ; VLX-LABEL: test_masked_vcmpoeqpd_v8i1_v32i1_sae_mask:
23093 ; VLX: # %bb.0: # %entry
23094 ; VLX-NEXT: vcmplepd {sae}, %zmm1, %zmm0, %k0
23095 ; VLX-NEXT: kmovd %k0, %eax
23096 ; VLX-NEXT: andb %dil, %al
23097 ; VLX-NEXT: movzbl %al, %eax
23098 ; VLX-NEXT: vzeroupper
23099 ; VLX-NEXT: retq
23100 ;
23101 ; NoVLX-LABEL: test_masked_vcmpoeqpd_v8i1_v32i1_sae_mask:
23102 ; NoVLX: # %bb.0: # %entry
23103 ; NoVLX-NEXT: vcmplepd {sae}, %zmm1, %zmm0, %k0
23104 ; NoVLX-NEXT: kmovw %k0, %eax
23105 ; NoVLX-NEXT: andb %dil, %al
23106 ; NoVLX-NEXT: movzbl %al, %eax
23107 ; NoVLX-NEXT: vzeroupper
23108 ; NoVLX-NEXT: retq
23109 entry:
23110 %0 = bitcast <8 x i64> %__a to <8 x double>
23111 %1 = bitcast <8 x i64> %__b to <8 x double>
23112 %2 = call <8 x i1> @llvm.x86.avx512.cmp.pd.512(<8 x double> %0, <8 x double> %1, i32 2, i32 8)
23113 %3 = bitcast i8 %__u to <8 x i1>
23114 %4 = and <8 x i1> %2, %3
23115 %5 = bitcast <8 x i1> %4 to i8
23116 %6 = zext i8 %5 to i32
23117 ret i32 %6
23118 }
23122 define zeroext i64 @test_vcmpoeqpd_v8i1_v64i1_mask(<8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
23123 ; VLX-LABEL: test_vcmpoeqpd_v8i1_v64i1_mask:
23124 ; VLX: # %bb.0: # %entry
23125 ; VLX-NEXT: vcmpeqpd %zmm1, %zmm0, %k0
23126 ; VLX-NEXT: kmovq %k0, %rax
23127 ; VLX-NEXT: vzeroupper
23128 ; VLX-NEXT: retq
23129 ;
23130 ; NoVLX-LABEL: test_vcmpoeqpd_v8i1_v64i1_mask:
23131 ; NoVLX: # %bb.0: # %entry
23132 ; NoVLX-NEXT: vcmpeqpd %zmm1, %zmm0, %k0
23133 ; NoVLX-NEXT: kmovw %k0, %eax
23134 ; NoVLX-NEXT: vzeroupper
23135 ; NoVLX-NEXT: retq
23136 entry:
23137 %0 = bitcast <8 x i64> %__a to <8 x double>
23138 %1 = bitcast <8 x i64> %__b to <8 x double>
23139 %2 = fcmp oeq <8 x double> %0, %1
23140 %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
23141 %4 = bitcast <64 x i1> %3 to i64
23142 ret i64 %4
23143 }
23145 define zeroext i64 @test_vcmpoeqpd_v8i1_v64i1_mask_mem(<8 x i64> %__a, <8 x i64>* %__b) local_unnamed_addr {
23146 ; VLX-LABEL: test_vcmpoeqpd_v8i1_v64i1_mask_mem:
23147 ; VLX: # %bb.0: # %entry
23148 ; VLX-NEXT: vcmpeqpd (%rdi), %zmm0, %k0
23149 ; VLX-NEXT: kmovq %k0, %rax
23150 ; VLX-NEXT: vzeroupper
23151 ; VLX-NEXT: retq
23152 ;
23153 ; NoVLX-LABEL: test_vcmpoeqpd_v8i1_v64i1_mask_mem:
23154 ; NoVLX: # %bb.0: # %entry
23155 ; NoVLX-NEXT: vcmpeqpd (%rdi), %zmm0, %k0
23156 ; NoVLX-NEXT: kmovw %k0, %eax
23157 ; NoVLX-NEXT: vzeroupper
23158 ; NoVLX-NEXT: retq
23159 entry:
23160 %0 = bitcast <8 x i64> %__a to <8 x double>
23161 %load = load <8 x i64>, <8 x i64>* %__b
23162 %1 = bitcast <8 x i64> %load to <8 x double>
23163 %2 = fcmp oeq <8 x double> %0, %1
23164 %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
23165 %4 = bitcast <64 x i1> %3 to i64
23166 ret i64 %4
23167 }
23169 define zeroext i64 @test_vcmpoeqpd_v8i1_v64i1_mask_mem_b(<8 x i64> %__a, double* %__b) local_unnamed_addr {
23170 ; VLX-LABEL: test_vcmpoeqpd_v8i1_v64i1_mask_mem_b:
23171 ; VLX: # %bb.0: # %entry
23172 ; VLX-NEXT: vcmpeqpd (%rdi){1to8}, %zmm0, %k0
23173 ; VLX-NEXT: kmovq %k0, %rax
23174 ; VLX-NEXT: vzeroupper
23175 ; VLX-NEXT: retq
23176 ;
23177 ; NoVLX-LABEL: test_vcmpoeqpd_v8i1_v64i1_mask_mem_b:
23178 ; NoVLX: # %bb.0: # %entry
23179 ; NoVLX-NEXT: vcmpeqpd (%rdi){1to8}, %zmm0, %k0
23180 ; NoVLX-NEXT: kmovw %k0, %eax
23181 ; NoVLX-NEXT: vzeroupper
23182 ; NoVLX-NEXT: retq
23183 entry:
23184 %0 = bitcast <8 x i64> %__a to <8 x double>
23185 %load = load double, double* %__b
23186 %vec = insertelement <8 x double> undef, double %load, i32 0
23187 %1 = shufflevector <8 x double> %vec, <8 x double> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
23188 %2 = fcmp oeq <8 x double> %0, %1
23189 %3 = shufflevector <8 x i1> %2, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
23190 %4 = bitcast <64 x i1> %3 to i64
23191 ret i64 %4
23192 }
23194 define zeroext i64 @test_masked_vcmpoeqpd_v8i1_v64i1_mask(i8 zeroext %__u, <8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
23195 ; VLX-LABEL: test_masked_vcmpoeqpd_v8i1_v64i1_mask:
23196 ; VLX: # %bb.0: # %entry
23197 ; VLX-NEXT: kmovd %edi, %k1
23198 ; VLX-NEXT: vcmpeqpd %zmm1, %zmm0, %k0 {%k1}
23199 ; VLX-NEXT: kmovq %k0, %rax
23200 ; VLX-NEXT: vzeroupper
23201 ; VLX-NEXT: retq
23202 ;
23203 ; NoVLX-LABEL: test_masked_vcmpoeqpd_v8i1_v64i1_mask:
23204 ; NoVLX: # %bb.0: # %entry
23205 ; NoVLX-NEXT: kmovw %edi, %k1
23206 ; NoVLX-NEXT: vcmpeqpd %zmm1, %zmm0, %k0 {%k1}
23207 ; NoVLX-NEXT: kmovw %k0, %eax
23208 ; NoVLX-NEXT: vzeroupper
23209 ; NoVLX-NEXT: retq
23210 entry:
23211 %0 = bitcast <8 x i64> %__a to <8 x double>
23212 %1 = bitcast <8 x i64> %__b to <8 x double>
23213 %2 = fcmp oeq <8 x double> %0, %1
23214 %3 = bitcast i8 %__u to <8 x i1>
23215 %4 = and <8 x i1> %2, %3
23216 %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
23217 %6 = bitcast <64 x i1> %5 to i64
23218 ret i64 %6
23219 }
23221 define zeroext i64 @test_masked_vcmpoeqpd_v8i1_v64i1_mask_mem(i8 zeroext %__u, <8 x i64> %__a, <8 x i64>* %__b) local_unnamed_addr {
23222 ; VLX-LABEL: test_masked_vcmpoeqpd_v8i1_v64i1_mask_mem:
23223 ; VLX: # %bb.0: # %entry
23224 ; VLX-NEXT: kmovd %edi, %k1
23225 ; VLX-NEXT: vcmpeqpd (%rsi), %zmm0, %k0 {%k1}
23226 ; VLX-NEXT: kmovq %k0, %rax
23227 ; VLX-NEXT: vzeroupper
23228 ; VLX-NEXT: retq
23229 ;
23230 ; NoVLX-LABEL: test_masked_vcmpoeqpd_v8i1_v64i1_mask_mem:
23231 ; NoVLX: # %bb.0: # %entry
23232 ; NoVLX-NEXT: kmovw %edi, %k1
23233 ; NoVLX-NEXT: vcmpeqpd (%rsi), %zmm0, %k0 {%k1}
23234 ; NoVLX-NEXT: kmovw %k0, %eax
23235 ; NoVLX-NEXT: vzeroupper
23236 ; NoVLX-NEXT: retq
23237 entry:
23238 %0 = bitcast <8 x i64> %__a to <8 x double>
23239 %load = load <8 x i64>, <8 x i64>* %__b
23240 %1 = bitcast <8 x i64> %load to <8 x double>
23241 %2 = fcmp oeq <8 x double> %0, %1
23242 %3 = bitcast i8 %__u to <8 x i1>
23243 %4 = and <8 x i1> %2, %3
23244 %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
23245 %6 = bitcast <64 x i1> %5 to i64
23246 ret i64 %6
23247 }
23249 define zeroext i64 @test_masked_vcmpoeqpd_v8i1_v64i1_mask_mem_b(i8 zeroext %__u, <8 x i64> %__a, double* %__b) local_unnamed_addr {
23250 ; VLX-LABEL: test_masked_vcmpoeqpd_v8i1_v64i1_mask_mem_b:
23251 ; VLX: # %bb.0: # %entry
23252 ; VLX-NEXT: kmovd %edi, %k1
23253 ; VLX-NEXT: vcmpeqpd (%rsi){1to8}, %zmm0, %k0 {%k1}
23254 ; VLX-NEXT: kmovq %k0, %rax
23255 ; VLX-NEXT: vzeroupper
23256 ; VLX-NEXT: retq
23257 ;
23258 ; NoVLX-LABEL: test_masked_vcmpoeqpd_v8i1_v64i1_mask_mem_b:
23259 ; NoVLX: # %bb.0: # %entry
23260 ; NoVLX-NEXT: kmovw %edi, %k1
23261 ; NoVLX-NEXT: vcmpeqpd (%rsi){1to8}, %zmm0, %k0 {%k1}
23262 ; NoVLX-NEXT: kmovw %k0, %eax
23263 ; NoVLX-NEXT: vzeroupper
23264 ; NoVLX-NEXT: retq
23265 entry:
23266 %0 = bitcast <8 x i64> %__a to <8 x double>
23267 %load = load double, double* %__b
23268 %vec = insertelement <8 x double> undef, double %load, i32 0
23269 %1 = shufflevector <8 x double> %vec, <8 x double> undef, <8 x i32> <i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0>
23270 %2 = fcmp oeq <8 x double> %0, %1
23271 %3 = bitcast i8 %__u to <8 x i1>
23272 %4 = and <8 x i1> %2, %3
23273 %5 = shufflevector <8 x i1> %4, <8 x i1> zeroinitializer, <64 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
23274 %6 = bitcast <64 x i1> %5 to i64
23275 ret i64 %6
23276 }
23280 define zeroext i64 @test_vcmpoeqpd_v8i1_v64i1_sae_mask(<8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
23281 ; VLX-LABEL: test_vcmpoeqpd_v8i1_v64i1_sae_mask:
23282 ; VLX: # %bb.0: # %entry
23283 ; VLX-NEXT: vcmplepd {sae}, %zmm1, %zmm0, %k0
23284 ; VLX-NEXT: kmovb %k0, %eax
23285 ; VLX-NEXT: vzeroupper
23286 ; VLX-NEXT: retq
23287 ;
23288 ; NoVLX-LABEL: test_vcmpoeqpd_v8i1_v64i1_sae_mask:
23289 ; NoVLX: # %bb.0: # %entry
23290 ; NoVLX-NEXT: vcmplepd {sae}, %zmm1, %zmm0, %k0
23291 ; NoVLX-NEXT: kmovw %k0, %eax
23292 ; NoVLX-NEXT: movzbl %al, %eax
23293 ; NoVLX-NEXT: vzeroupper
23294 ; NoVLX-NEXT: retq
23295 entry:
23296 %0 = bitcast <8 x i64> %__a to <8 x double>
23297 %1 = bitcast <8 x i64> %__b to <8 x double>
23298 %2 = call <8 x i1> @llvm.x86.avx512.cmp.pd.512(<8 x double> %0, <8 x double> %1, i32 2, i32 8)
23299 %3 = bitcast <8 x i1> %2 to i8
23300 %4 = zext i8 %3 to i64
23301 ret i64 %4
23302 }
23304 define zeroext i64 @test_masked_vcmpoeqpd_v8i1_v64i1_sae_mask(i8 zeroext %__u, <8 x i64> %__a, <8 x i64> %__b) local_unnamed_addr {
23305 ; VLX-LABEL: test_masked_vcmpoeqpd_v8i1_v64i1_sae_mask:
23306 ; VLX: # %bb.0: # %entry
23307 ; VLX-NEXT: vcmplepd {sae}, %zmm1, %zmm0, %k0
23308 ; VLX-NEXT: kmovd %k0, %eax
23309 ; VLX-NEXT: andb %dil, %al
23310 ; VLX-NEXT: movzbl %al, %eax
23311 ; VLX-NEXT: vzeroupper
23312 ; VLX-NEXT: retq
23313 ;
23314 ; NoVLX-LABEL: test_masked_vcmpoeqpd_v8i1_v64i1_sae_mask:
23315 ; NoVLX: # %bb.0: # %entry
23316 ; NoVLX-NEXT: vcmplepd {sae}, %zmm1, %zmm0, %k0
23317 ; NoVLX-NEXT: kmovw %k0, %eax
23318 ; NoVLX-NEXT: andb %dil, %al
23319 ; NoVLX-NEXT: movzbl %al, %eax
23320 ; NoVLX-NEXT: vzeroupper
23321 ; NoVLX-NEXT: retq
23322 entry:
23323 %0 = bitcast <8 x i64> %__a to <8 x double>
23324 %1 = bitcast <8 x i64> %__b to <8 x double>
23325 %2 = call <8 x i1> @llvm.x86.avx512.cmp.pd.512(<8 x double> %0, <8 x double> %1, i32 2, i32 8)
23326 %3 = bitcast i8 %__u to <8 x i1>
23327 %4 = and <8 x i1> %2, %3
23328 %5 = bitcast <8 x i1> %4 to i8
23329 %6 = zext i8 %5 to i64
23330 ret i64 %6
23331 }
23333 ; Test that we understand that cmpps with rounding zeros the upper bits of the mask register.
23334 define i32 @test_cmpm_rnd_zero(<16 x float> %a, <16 x float> %b) {
23335 ; VLX-LABEL: test_cmpm_rnd_zero:
23336 ; VLX: # %bb.0:
23337 ; VLX-NEXT: vcmpleps {sae}, %zmm1, %zmm0, %k0
23338 ; VLX-NEXT: kmovd %k0, %eax
23339 ; VLX-NEXT: vzeroupper
23340 ; VLX-NEXT: retq
23341 ;
23342 ; NoVLX-LABEL: test_cmpm_rnd_zero:
23343 ; NoVLX: # %bb.0:
23344 ; NoVLX-NEXT: vcmpleps {sae}, %zmm1, %zmm0, %k0
23345 ; NoVLX-NEXT: kmovw %k0, %eax
23346 ; NoVLX-NEXT: vzeroupper
23347 ; NoVLX-NEXT: retq
23348 %res = call <16 x i1> @llvm.x86.avx512.cmp.ps.512(<16 x float> %a, <16 x float> %b, i32 2, i32 8)
23349 %1 = bitcast <16 x i1> %res to i16
23350 %cast = bitcast i16 %1 to <16 x i1>
23351 %shuffle = shufflevector <16 x i1> %cast, <16 x i1> zeroinitializer, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 31>
23352 %cast2 = bitcast <32 x i1> %shuffle to i32
23353 ret i32 %cast2
23354 }
23356 define i8 @mask_zero_lower(<4 x i32> %a) {
23357 ; VLX-LABEL: mask_zero_lower:
23358 ; VLX: # %bb.0:
23359 ; VLX-NEXT: vptestmd %xmm0, %xmm0, %k0
23360 ; VLX-NEXT: kshiftlb $4, %k0, %k0
23361 ; VLX-NEXT: kmovd %k0, %eax
23362 ; VLX-NEXT: # kill: def $al killed $al killed $eax
23363 ; VLX-NEXT: retq
23364 ;
23365 ; NoVLX-LABEL: mask_zero_lower:
23366 ; NoVLX: # %bb.0:
23367 ; NoVLX-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
23368 ; NoVLX-NEXT: vptestmd %zmm0, %zmm0, %k0
23369 ; NoVLX-NEXT: kshiftlw $4, %k0, %k0
23370 ; NoVLX-NEXT: kmovw %k0, %eax
23371 ; NoVLX-NEXT: # kill: def $al killed $al killed $eax
23372 ; NoVLX-NEXT: vzeroupper
23373 ; NoVLX-NEXT: retq
23374 %cmp = icmp ne <4 x i32> %a, zeroinitializer
23375 %concat = shufflevector <4 x i1> %cmp, <4 x i1> zeroinitializer, <8 x i32> <i32 4, i32 5, i32 6, i32 7, i32 0, i32 1, i32 2, i32 3>
23376 %cast = bitcast <8 x i1> %concat to i8
23377 ret i8 %cast
23378 }