| blob | efecfa47eb4abb5a0e18d1a07c344245ec421163 |
1 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
2 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512dq,+avx512cd,-avx512bw | FileCheck %s --check-prefix=AVX512CD
3 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512dq,+avx512cd,+avx512bw | FileCheck %s --check-prefix=AVX512CDBW
4 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512dq,-avx512cd,+avx512bw | FileCheck %s --check-prefix=AVX512BW
5 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512dq,-avx512cd,-avx512bw | FileCheck %s --check-prefix=AVX512DQ
7 define <8 x i64> @testv8i64(<8 x i64> %in) nounwind {
8 ; AVX512CD-LABEL: testv8i64:
9 ; AVX512CD: # %bb.0:
10 ; AVX512CD-NEXT: vplzcntq %zmm0, %zmm0
11 ; AVX512CD-NEXT: retq
12 ;
13 ; AVX512CDBW-LABEL: testv8i64:
14 ; AVX512CDBW: # %bb.0:
15 ; AVX512CDBW-NEXT: vplzcntq %zmm0, %zmm0
16 ; AVX512CDBW-NEXT: retq
17 ;
18 ; AVX512BW-LABEL: testv8i64:
19 ; AVX512BW: # %bb.0:
20 ; AVX512BW-NEXT: vpsrlq $1, %zmm0, %zmm1
21 ; AVX512BW-NEXT: vporq %zmm1, %zmm0, %zmm0
22 ; AVX512BW-NEXT: vpsrlq $2, %zmm0, %zmm1
23 ; AVX512BW-NEXT: vporq %zmm1, %zmm0, %zmm0
24 ; AVX512BW-NEXT: vpsrlq $4, %zmm0, %zmm1
25 ; AVX512BW-NEXT: vporq %zmm1, %zmm0, %zmm0
26 ; AVX512BW-NEXT: vpsrlq $8, %zmm0, %zmm1
27 ; AVX512BW-NEXT: vporq %zmm1, %zmm0, %zmm0
28 ; AVX512BW-NEXT: vpsrlq $16, %zmm0, %zmm1
29 ; AVX512BW-NEXT: vporq %zmm1, %zmm0, %zmm0
30 ; AVX512BW-NEXT: vpsrlq $32, %zmm0, %zmm1
31 ; AVX512BW-NEXT: vporq %zmm1, %zmm0, %zmm0
32 ; AVX512BW-NEXT: vpbroadcastb {{.*#+}} zmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
33 ; AVX512BW-NEXT: vpandnq %zmm1, %zmm0, %zmm2
34 ; AVX512BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
35 ; AVX512BW-NEXT: # zmm3 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
36 ; AVX512BW-NEXT: vpshufb %zmm2, %zmm3, %zmm2
37 ; AVX512BW-NEXT: vpternlogq $15, %zmm0, %zmm0, %zmm0
38 ; AVX512BW-NEXT: vpsrlw $4, %zmm0, %zmm0
39 ; AVX512BW-NEXT: vpandq %zmm1, %zmm0, %zmm0
40 ; AVX512BW-NEXT: vpshufb %zmm0, %zmm3, %zmm0
41 ; AVX512BW-NEXT: vpaddb %zmm2, %zmm0, %zmm0
42 ; AVX512BW-NEXT: vpxor %xmm1, %xmm1, %xmm1
43 ; AVX512BW-NEXT: vpsadbw %zmm1, %zmm0, %zmm0
44 ; AVX512BW-NEXT: retq
45 ;
46 ; AVX512DQ-LABEL: testv8i64:
47 ; AVX512DQ: # %bb.0:
48 ; AVX512DQ-NEXT: vpsrlq $1, %zmm0, %zmm1
49 ; AVX512DQ-NEXT: vporq %zmm1, %zmm0, %zmm0
50 ; AVX512DQ-NEXT: vpsrlq $2, %zmm0, %zmm1
51 ; AVX512DQ-NEXT: vporq %zmm1, %zmm0, %zmm0
52 ; AVX512DQ-NEXT: vpsrlq $4, %zmm0, %zmm1
53 ; AVX512DQ-NEXT: vporq %zmm1, %zmm0, %zmm0
54 ; AVX512DQ-NEXT: vpsrlq $8, %zmm0, %zmm1
55 ; AVX512DQ-NEXT: vporq %zmm1, %zmm0, %zmm0
56 ; AVX512DQ-NEXT: vpsrlq $16, %zmm0, %zmm1
57 ; AVX512DQ-NEXT: vporq %zmm1, %zmm0, %zmm0
58 ; AVX512DQ-NEXT: vpsrlq $32, %zmm0, %zmm1
59 ; AVX512DQ-NEXT: vporq %zmm1, %zmm0, %zmm0
60 ; AVX512DQ-NEXT: vpbroadcastb {{.*#+}} ymm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
61 ; AVX512DQ-NEXT: vpandn %ymm1, %ymm0, %ymm2
62 ; AVX512DQ-NEXT: vpternlogq $15, %zmm0, %zmm0, %zmm0
63 ; AVX512DQ-NEXT: vextracti64x4 $1, %zmm0, %ymm3
64 ; AVX512DQ-NEXT: vpand %ymm1, %ymm3, %ymm4
65 ; AVX512DQ-NEXT: vbroadcasti128 {{.*#+}} ymm5 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
66 ; AVX512DQ-NEXT: # ymm5 = mem[0,1,0,1]
67 ; AVX512DQ-NEXT: vpshufb %ymm4, %ymm5, %ymm4
68 ; AVX512DQ-NEXT: vpsrlw $4, %ymm3, %ymm3
69 ; AVX512DQ-NEXT: vpand %ymm1, %ymm3, %ymm3
70 ; AVX512DQ-NEXT: vpshufb %ymm3, %ymm5, %ymm3
71 ; AVX512DQ-NEXT: vpaddb %ymm4, %ymm3, %ymm3
72 ; AVX512DQ-NEXT: vpxor %xmm4, %xmm4, %xmm4
73 ; AVX512DQ-NEXT: vpsadbw %ymm4, %ymm3, %ymm3
74 ; AVX512DQ-NEXT: vpshufb %ymm2, %ymm5, %ymm2
75 ; AVX512DQ-NEXT: vpsrlw $4, %ymm0, %ymm0
76 ; AVX512DQ-NEXT: vpand %ymm1, %ymm0, %ymm0
77 ; AVX512DQ-NEXT: vpshufb %ymm0, %ymm5, %ymm0
78 ; AVX512DQ-NEXT: vpaddb %ymm2, %ymm0, %ymm0
79 ; AVX512DQ-NEXT: vpsadbw %ymm4, %ymm0, %ymm0
80 ; AVX512DQ-NEXT: vinserti64x4 $1, %ymm3, %zmm0, %zmm0
81 ; AVX512DQ-NEXT: retq
82 %out = call <8 x i64> @llvm.ctlz.v8i64(<8 x i64> %in, i1 0)
83 ret <8 x i64> %out
84 }
86 define <8 x i64> @testv8i64u(<8 x i64> %in) nounwind {
87 ; AVX512CD-LABEL: testv8i64u:
88 ; AVX512CD: # %bb.0:
89 ; AVX512CD-NEXT: vplzcntq %zmm0, %zmm0
90 ; AVX512CD-NEXT: retq
91 ;
92 ; AVX512CDBW-LABEL: testv8i64u:
93 ; AVX512CDBW: # %bb.0:
94 ; AVX512CDBW-NEXT: vplzcntq %zmm0, %zmm0
95 ; AVX512CDBW-NEXT: retq
96 ;
97 ; AVX512BW-LABEL: testv8i64u:
98 ; AVX512BW: # %bb.0:
99 ; AVX512BW-NEXT: vpsrlq $1, %zmm0, %zmm1
100 ; AVX512BW-NEXT: vporq %zmm1, %zmm0, %zmm0
101 ; AVX512BW-NEXT: vpsrlq $2, %zmm0, %zmm1
102 ; AVX512BW-NEXT: vporq %zmm1, %zmm0, %zmm0
103 ; AVX512BW-NEXT: vpsrlq $4, %zmm0, %zmm1
104 ; AVX512BW-NEXT: vporq %zmm1, %zmm0, %zmm0
105 ; AVX512BW-NEXT: vpsrlq $8, %zmm0, %zmm1
106 ; AVX512BW-NEXT: vporq %zmm1, %zmm0, %zmm0
107 ; AVX512BW-NEXT: vpsrlq $16, %zmm0, %zmm1
108 ; AVX512BW-NEXT: vporq %zmm1, %zmm0, %zmm0
109 ; AVX512BW-NEXT: vpsrlq $32, %zmm0, %zmm1
110 ; AVX512BW-NEXT: vporq %zmm1, %zmm0, %zmm0
111 ; AVX512BW-NEXT: vpbroadcastb {{.*#+}} zmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
112 ; AVX512BW-NEXT: vpandnq %zmm1, %zmm0, %zmm2
113 ; AVX512BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
114 ; AVX512BW-NEXT: # zmm3 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
115 ; AVX512BW-NEXT: vpshufb %zmm2, %zmm3, %zmm2
116 ; AVX512BW-NEXT: vpternlogq $15, %zmm0, %zmm0, %zmm0
117 ; AVX512BW-NEXT: vpsrlw $4, %zmm0, %zmm0
118 ; AVX512BW-NEXT: vpandq %zmm1, %zmm0, %zmm0
119 ; AVX512BW-NEXT: vpshufb %zmm0, %zmm3, %zmm0
120 ; AVX512BW-NEXT: vpaddb %zmm2, %zmm0, %zmm0
121 ; AVX512BW-NEXT: vpxor %xmm1, %xmm1, %xmm1
122 ; AVX512BW-NEXT: vpsadbw %zmm1, %zmm0, %zmm0
123 ; AVX512BW-NEXT: retq
124 ;
125 ; AVX512DQ-LABEL: testv8i64u:
126 ; AVX512DQ: # %bb.0:
127 ; AVX512DQ-NEXT: vpsrlq $1, %zmm0, %zmm1
128 ; AVX512DQ-NEXT: vporq %zmm1, %zmm0, %zmm0
129 ; AVX512DQ-NEXT: vpsrlq $2, %zmm0, %zmm1
130 ; AVX512DQ-NEXT: vporq %zmm1, %zmm0, %zmm0
131 ; AVX512DQ-NEXT: vpsrlq $4, %zmm0, %zmm1
132 ; AVX512DQ-NEXT: vporq %zmm1, %zmm0, %zmm0
133 ; AVX512DQ-NEXT: vpsrlq $8, %zmm0, %zmm1
134 ; AVX512DQ-NEXT: vporq %zmm1, %zmm0, %zmm0
135 ; AVX512DQ-NEXT: vpsrlq $16, %zmm0, %zmm1
136 ; AVX512DQ-NEXT: vporq %zmm1, %zmm0, %zmm0
137 ; AVX512DQ-NEXT: vpsrlq $32, %zmm0, %zmm1
138 ; AVX512DQ-NEXT: vporq %zmm1, %zmm0, %zmm0
139 ; AVX512DQ-NEXT: vpbroadcastb {{.*#+}} ymm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
140 ; AVX512DQ-NEXT: vpandn %ymm1, %ymm0, %ymm2
141 ; AVX512DQ-NEXT: vpternlogq $15, %zmm0, %zmm0, %zmm0
142 ; AVX512DQ-NEXT: vextracti64x4 $1, %zmm0, %ymm3
143 ; AVX512DQ-NEXT: vpand %ymm1, %ymm3, %ymm4
144 ; AVX512DQ-NEXT: vbroadcasti128 {{.*#+}} ymm5 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
145 ; AVX512DQ-NEXT: # ymm5 = mem[0,1,0,1]
146 ; AVX512DQ-NEXT: vpshufb %ymm4, %ymm5, %ymm4
147 ; AVX512DQ-NEXT: vpsrlw $4, %ymm3, %ymm3
148 ; AVX512DQ-NEXT: vpand %ymm1, %ymm3, %ymm3
149 ; AVX512DQ-NEXT: vpshufb %ymm3, %ymm5, %ymm3
150 ; AVX512DQ-NEXT: vpaddb %ymm4, %ymm3, %ymm3
151 ; AVX512DQ-NEXT: vpxor %xmm4, %xmm4, %xmm4
152 ; AVX512DQ-NEXT: vpsadbw %ymm4, %ymm3, %ymm3
153 ; AVX512DQ-NEXT: vpshufb %ymm2, %ymm5, %ymm2
154 ; AVX512DQ-NEXT: vpsrlw $4, %ymm0, %ymm0
155 ; AVX512DQ-NEXT: vpand %ymm1, %ymm0, %ymm0
156 ; AVX512DQ-NEXT: vpshufb %ymm0, %ymm5, %ymm0
157 ; AVX512DQ-NEXT: vpaddb %ymm2, %ymm0, %ymm0
158 ; AVX512DQ-NEXT: vpsadbw %ymm4, %ymm0, %ymm0
159 ; AVX512DQ-NEXT: vinserti64x4 $1, %ymm3, %zmm0, %zmm0
160 ; AVX512DQ-NEXT: retq
161 %out = call <8 x i64> @llvm.ctlz.v8i64(<8 x i64> %in, i1 -1)
162 ret <8 x i64> %out
163 }
165 define <16 x i32> @testv16i32(<16 x i32> %in) nounwind {
166 ; AVX512CD-LABEL: testv16i32:
167 ; AVX512CD: # %bb.0:
168 ; AVX512CD-NEXT: vplzcntd %zmm0, %zmm0
169 ; AVX512CD-NEXT: retq
170 ;
171 ; AVX512CDBW-LABEL: testv16i32:
172 ; AVX512CDBW: # %bb.0:
173 ; AVX512CDBW-NEXT: vplzcntd %zmm0, %zmm0
174 ; AVX512CDBW-NEXT: retq
175 ;
176 ; AVX512BW-LABEL: testv16i32:
177 ; AVX512BW: # %bb.0:
178 ; AVX512BW-NEXT: vpsrld $1, %zmm0, %zmm1
179 ; AVX512BW-NEXT: vpord %zmm1, %zmm0, %zmm0
180 ; AVX512BW-NEXT: vpsrld $2, %zmm0, %zmm1
181 ; AVX512BW-NEXT: vpord %zmm1, %zmm0, %zmm0
182 ; AVX512BW-NEXT: vpsrld $4, %zmm0, %zmm1
183 ; AVX512BW-NEXT: vpord %zmm1, %zmm0, %zmm0
184 ; AVX512BW-NEXT: vpsrld $8, %zmm0, %zmm1
185 ; AVX512BW-NEXT: vpord %zmm1, %zmm0, %zmm0
186 ; AVX512BW-NEXT: vpsrld $16, %zmm0, %zmm1
187 ; AVX512BW-NEXT: vpord %zmm1, %zmm0, %zmm0
188 ; AVX512BW-NEXT: vpbroadcastb {{.*#+}} zmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
189 ; AVX512BW-NEXT: vpandnq %zmm1, %zmm0, %zmm2
190 ; AVX512BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
191 ; AVX512BW-NEXT: # zmm3 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
192 ; AVX512BW-NEXT: vpshufb %zmm2, %zmm3, %zmm2
193 ; AVX512BW-NEXT: vpternlogq $15, %zmm0, %zmm0, %zmm0
194 ; AVX512BW-NEXT: vpsrlw $4, %zmm0, %zmm0
195 ; AVX512BW-NEXT: vpandq %zmm1, %zmm0, %zmm0
196 ; AVX512BW-NEXT: vpshufb %zmm0, %zmm3, %zmm0
197 ; AVX512BW-NEXT: vpaddb %zmm2, %zmm0, %zmm0
198 ; AVX512BW-NEXT: vpxor %xmm1, %xmm1, %xmm1
199 ; AVX512BW-NEXT: vpunpckhdq {{.*#+}} zmm2 = zmm0[2],zmm1[2],zmm0[3],zmm1[3],zmm0[6],zmm1[6],zmm0[7],zmm1[7],zmm0[10],zmm1[10],zmm0[11],zmm1[11],zmm0[14],zmm1[14],zmm0[15],zmm1[15]
200 ; AVX512BW-NEXT: vpsadbw %zmm1, %zmm2, %zmm2
201 ; AVX512BW-NEXT: vpunpckldq {{.*#+}} zmm0 = zmm0[0],zmm1[0],zmm0[1],zmm1[1],zmm0[4],zmm1[4],zmm0[5],zmm1[5],zmm0[8],zmm1[8],zmm0[9],zmm1[9],zmm0[12],zmm1[12],zmm0[13],zmm1[13]
202 ; AVX512BW-NEXT: vpsadbw %zmm1, %zmm0, %zmm0
203 ; AVX512BW-NEXT: vpackuswb %zmm2, %zmm0, %zmm0
204 ; AVX512BW-NEXT: retq
205 ;
206 ; AVX512DQ-LABEL: testv16i32:
207 ; AVX512DQ: # %bb.0:
208 ; AVX512DQ-NEXT: vpsrld $1, %zmm0, %zmm1
209 ; AVX512DQ-NEXT: vpord %zmm1, %zmm0, %zmm0
210 ; AVX512DQ-NEXT: vpsrld $2, %zmm0, %zmm1
211 ; AVX512DQ-NEXT: vpord %zmm1, %zmm0, %zmm0
212 ; AVX512DQ-NEXT: vpsrld $4, %zmm0, %zmm1
213 ; AVX512DQ-NEXT: vpord %zmm1, %zmm0, %zmm0
214 ; AVX512DQ-NEXT: vpsrld $8, %zmm0, %zmm1
215 ; AVX512DQ-NEXT: vpord %zmm1, %zmm0, %zmm0
216 ; AVX512DQ-NEXT: vpsrld $16, %zmm0, %zmm1
217 ; AVX512DQ-NEXT: vpord %zmm1, %zmm0, %zmm1
218 ; AVX512DQ-NEXT: vpbroadcastb {{.*#+}} ymm0 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
219 ; AVX512DQ-NEXT: vpandn %ymm0, %ymm1, %ymm2
220 ; AVX512DQ-NEXT: vpternlogq $15, %zmm1, %zmm1, %zmm1
221 ; AVX512DQ-NEXT: vextracti64x4 $1, %zmm1, %ymm3
222 ; AVX512DQ-NEXT: vpand %ymm0, %ymm3, %ymm4
223 ; AVX512DQ-NEXT: vbroadcasti128 {{.*#+}} ymm5 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
224 ; AVX512DQ-NEXT: # ymm5 = mem[0,1,0,1]
225 ; AVX512DQ-NEXT: vpshufb %ymm4, %ymm5, %ymm4
226 ; AVX512DQ-NEXT: vpsrlw $4, %ymm3, %ymm3
227 ; AVX512DQ-NEXT: vpand %ymm0, %ymm3, %ymm3
228 ; AVX512DQ-NEXT: vpshufb %ymm3, %ymm5, %ymm3
229 ; AVX512DQ-NEXT: vpaddb %ymm4, %ymm3, %ymm3
230 ; AVX512DQ-NEXT: vpxor %xmm4, %xmm4, %xmm4
231 ; AVX512DQ-NEXT: vpunpckhdq {{.*#+}} ymm6 = ymm3[2],ymm4[2],ymm3[3],ymm4[3],ymm3[6],ymm4[6],ymm3[7],ymm4[7]
232 ; AVX512DQ-NEXT: vpsadbw %ymm4, %ymm6, %ymm6
233 ; AVX512DQ-NEXT: vpunpckldq {{.*#+}} ymm3 = ymm3[0],ymm4[0],ymm3[1],ymm4[1],ymm3[4],ymm4[4],ymm3[5],ymm4[5]
234 ; AVX512DQ-NEXT: vpsadbw %ymm4, %ymm3, %ymm3
235 ; AVX512DQ-NEXT: vpackuswb %ymm6, %ymm3, %ymm3
236 ; AVX512DQ-NEXT: vpshufb %ymm2, %ymm5, %ymm2
237 ; AVX512DQ-NEXT: vpsrlw $4, %ymm1, %ymm1
238 ; AVX512DQ-NEXT: vpand %ymm0, %ymm1, %ymm0
239 ; AVX512DQ-NEXT: vpshufb %ymm0, %ymm5, %ymm0
240 ; AVX512DQ-NEXT: vpaddb %ymm2, %ymm0, %ymm0
241 ; AVX512DQ-NEXT: vpunpckhdq {{.*#+}} ymm1 = ymm0[2],ymm4[2],ymm0[3],ymm4[3],ymm0[6],ymm4[6],ymm0[7],ymm4[7]
242 ; AVX512DQ-NEXT: vpsadbw %ymm4, %ymm1, %ymm1
243 ; AVX512DQ-NEXT: vpunpckldq {{.*#+}} ymm0 = ymm0[0],ymm4[0],ymm0[1],ymm4[1],ymm0[4],ymm4[4],ymm0[5],ymm4[5]
244 ; AVX512DQ-NEXT: vpsadbw %ymm4, %ymm0, %ymm0
245 ; AVX512DQ-NEXT: vpackuswb %ymm1, %ymm0, %ymm0
246 ; AVX512DQ-NEXT: vinserti64x4 $1, %ymm3, %zmm0, %zmm0
247 ; AVX512DQ-NEXT: retq
248 %out = call <16 x i32> @llvm.ctlz.v16i32(<16 x i32> %in, i1 0)
249 ret <16 x i32> %out
250 }
252 define <16 x i32> @testv16i32u(<16 x i32> %in) nounwind {
253 ; AVX512CD-LABEL: testv16i32u:
254 ; AVX512CD: # %bb.0:
255 ; AVX512CD-NEXT: vplzcntd %zmm0, %zmm0
256 ; AVX512CD-NEXT: retq
257 ;
258 ; AVX512CDBW-LABEL: testv16i32u:
259 ; AVX512CDBW: # %bb.0:
260 ; AVX512CDBW-NEXT: vplzcntd %zmm0, %zmm0
261 ; AVX512CDBW-NEXT: retq
262 ;
263 ; AVX512BW-LABEL: testv16i32u:
264 ; AVX512BW: # %bb.0:
265 ; AVX512BW-NEXT: vpsrld $1, %zmm0, %zmm1
266 ; AVX512BW-NEXT: vpord %zmm1, %zmm0, %zmm0
267 ; AVX512BW-NEXT: vpsrld $2, %zmm0, %zmm1
268 ; AVX512BW-NEXT: vpord %zmm1, %zmm0, %zmm0
269 ; AVX512BW-NEXT: vpsrld $4, %zmm0, %zmm1
270 ; AVX512BW-NEXT: vpord %zmm1, %zmm0, %zmm0
271 ; AVX512BW-NEXT: vpsrld $8, %zmm0, %zmm1
272 ; AVX512BW-NEXT: vpord %zmm1, %zmm0, %zmm0
273 ; AVX512BW-NEXT: vpsrld $16, %zmm0, %zmm1
274 ; AVX512BW-NEXT: vpord %zmm1, %zmm0, %zmm0
275 ; AVX512BW-NEXT: vpbroadcastb {{.*#+}} zmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
276 ; AVX512BW-NEXT: vpandnq %zmm1, %zmm0, %zmm2
277 ; AVX512BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm3 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
278 ; AVX512BW-NEXT: # zmm3 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
279 ; AVX512BW-NEXT: vpshufb %zmm2, %zmm3, %zmm2
280 ; AVX512BW-NEXT: vpternlogq $15, %zmm0, %zmm0, %zmm0
281 ; AVX512BW-NEXT: vpsrlw $4, %zmm0, %zmm0
282 ; AVX512BW-NEXT: vpandq %zmm1, %zmm0, %zmm0
283 ; AVX512BW-NEXT: vpshufb %zmm0, %zmm3, %zmm0
284 ; AVX512BW-NEXT: vpaddb %zmm2, %zmm0, %zmm0
285 ; AVX512BW-NEXT: vpxor %xmm1, %xmm1, %xmm1
286 ; AVX512BW-NEXT: vpunpckhdq {{.*#+}} zmm2 = zmm0[2],zmm1[2],zmm0[3],zmm1[3],zmm0[6],zmm1[6],zmm0[7],zmm1[7],zmm0[10],zmm1[10],zmm0[11],zmm1[11],zmm0[14],zmm1[14],zmm0[15],zmm1[15]
287 ; AVX512BW-NEXT: vpsadbw %zmm1, %zmm2, %zmm2
288 ; AVX512BW-NEXT: vpunpckldq {{.*#+}} zmm0 = zmm0[0],zmm1[0],zmm0[1],zmm1[1],zmm0[4],zmm1[4],zmm0[5],zmm1[5],zmm0[8],zmm1[8],zmm0[9],zmm1[9],zmm0[12],zmm1[12],zmm0[13],zmm1[13]
289 ; AVX512BW-NEXT: vpsadbw %zmm1, %zmm0, %zmm0
290 ; AVX512BW-NEXT: vpackuswb %zmm2, %zmm0, %zmm0
291 ; AVX512BW-NEXT: retq
292 ;
293 ; AVX512DQ-LABEL: testv16i32u:
294 ; AVX512DQ: # %bb.0:
295 ; AVX512DQ-NEXT: vpsrld $1, %zmm0, %zmm1
296 ; AVX512DQ-NEXT: vpord %zmm1, %zmm0, %zmm0
297 ; AVX512DQ-NEXT: vpsrld $2, %zmm0, %zmm1
298 ; AVX512DQ-NEXT: vpord %zmm1, %zmm0, %zmm0
299 ; AVX512DQ-NEXT: vpsrld $4, %zmm0, %zmm1
300 ; AVX512DQ-NEXT: vpord %zmm1, %zmm0, %zmm0
301 ; AVX512DQ-NEXT: vpsrld $8, %zmm0, %zmm1
302 ; AVX512DQ-NEXT: vpord %zmm1, %zmm0, %zmm0
303 ; AVX512DQ-NEXT: vpsrld $16, %zmm0, %zmm1
304 ; AVX512DQ-NEXT: vpord %zmm1, %zmm0, %zmm1
305 ; AVX512DQ-NEXT: vpbroadcastb {{.*#+}} ymm0 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
306 ; AVX512DQ-NEXT: vpandn %ymm0, %ymm1, %ymm2
307 ; AVX512DQ-NEXT: vpternlogq $15, %zmm1, %zmm1, %zmm1
308 ; AVX512DQ-NEXT: vextracti64x4 $1, %zmm1, %ymm3
309 ; AVX512DQ-NEXT: vpand %ymm0, %ymm3, %ymm4
310 ; AVX512DQ-NEXT: vbroadcasti128 {{.*#+}} ymm5 = [0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4,0,1,1,2,1,2,2,3,1,2,2,3,2,3,3,4]
311 ; AVX512DQ-NEXT: # ymm5 = mem[0,1,0,1]
312 ; AVX512DQ-NEXT: vpshufb %ymm4, %ymm5, %ymm4
313 ; AVX512DQ-NEXT: vpsrlw $4, %ymm3, %ymm3
314 ; AVX512DQ-NEXT: vpand %ymm0, %ymm3, %ymm3
315 ; AVX512DQ-NEXT: vpshufb %ymm3, %ymm5, %ymm3
316 ; AVX512DQ-NEXT: vpaddb %ymm4, %ymm3, %ymm3
317 ; AVX512DQ-NEXT: vpxor %xmm4, %xmm4, %xmm4
318 ; AVX512DQ-NEXT: vpunpckhdq {{.*#+}} ymm6 = ymm3[2],ymm4[2],ymm3[3],ymm4[3],ymm3[6],ymm4[6],ymm3[7],ymm4[7]
319 ; AVX512DQ-NEXT: vpsadbw %ymm4, %ymm6, %ymm6
320 ; AVX512DQ-NEXT: vpunpckldq {{.*#+}} ymm3 = ymm3[0],ymm4[0],ymm3[1],ymm4[1],ymm3[4],ymm4[4],ymm3[5],ymm4[5]
321 ; AVX512DQ-NEXT: vpsadbw %ymm4, %ymm3, %ymm3
322 ; AVX512DQ-NEXT: vpackuswb %ymm6, %ymm3, %ymm3
323 ; AVX512DQ-NEXT: vpshufb %ymm2, %ymm5, %ymm2
324 ; AVX512DQ-NEXT: vpsrlw $4, %ymm1, %ymm1
325 ; AVX512DQ-NEXT: vpand %ymm0, %ymm1, %ymm0
326 ; AVX512DQ-NEXT: vpshufb %ymm0, %ymm5, %ymm0
327 ; AVX512DQ-NEXT: vpaddb %ymm2, %ymm0, %ymm0
328 ; AVX512DQ-NEXT: vpunpckhdq {{.*#+}} ymm1 = ymm0[2],ymm4[2],ymm0[3],ymm4[3],ymm0[6],ymm4[6],ymm0[7],ymm4[7]
329 ; AVX512DQ-NEXT: vpsadbw %ymm4, %ymm1, %ymm1
330 ; AVX512DQ-NEXT: vpunpckldq {{.*#+}} ymm0 = ymm0[0],ymm4[0],ymm0[1],ymm4[1],ymm0[4],ymm4[4],ymm0[5],ymm4[5]
331 ; AVX512DQ-NEXT: vpsadbw %ymm4, %ymm0, %ymm0
332 ; AVX512DQ-NEXT: vpackuswb %ymm1, %ymm0, %ymm0
333 ; AVX512DQ-NEXT: vinserti64x4 $1, %ymm3, %zmm0, %zmm0
334 ; AVX512DQ-NEXT: retq
335 %out = call <16 x i32> @llvm.ctlz.v16i32(<16 x i32> %in, i1 -1)
336 ret <16 x i32> %out
337 }
339 define <32 x i16> @testv32i16(<32 x i16> %in) nounwind {
340 ; AVX512CD-LABEL: testv32i16:
341 ; AVX512CD: # %bb.0:
342 ; AVX512CD-NEXT: vextracti64x4 $1, %zmm0, %ymm1
343 ; AVX512CD-NEXT: vpmovzxwd {{.*#+}} zmm1 = ymm1[0],zero,ymm1[1],zero,ymm1[2],zero,ymm1[3],zero,ymm1[4],zero,ymm1[5],zero,ymm1[6],zero,ymm1[7],zero,ymm1[8],zero,ymm1[9],zero,ymm1[10],zero,ymm1[11],zero,ymm1[12],zero,ymm1[13],zero,ymm1[14],zero,ymm1[15],zero
344 ; AVX512CD-NEXT: vplzcntd %zmm1, %zmm1
345 ; AVX512CD-NEXT: vpmovdw %zmm1, %ymm1
346 ; AVX512CD-NEXT: vpbroadcastw {{.*#+}} ymm2 = [16,16,16,16,16,16,16,16,16,16,16,16,16,16,16,16]
347 ; AVX512CD-NEXT: vpsubw %ymm2, %ymm1, %ymm1
348 ; AVX512CD-NEXT: vpmovzxwd {{.*#+}} zmm0 = ymm0[0],zero,ymm0[1],zero,ymm0[2],zero,ymm0[3],zero,ymm0[4],zero,ymm0[5],zero,ymm0[6],zero,ymm0[7],zero,ymm0[8],zero,ymm0[9],zero,ymm0[10],zero,ymm0[11],zero,ymm0[12],zero,ymm0[13],zero,ymm0[14],zero,ymm0[15],zero
349 ; AVX512CD-NEXT: vplzcntd %zmm0, %zmm0
350 ; AVX512CD-NEXT: vpmovdw %zmm0, %ymm0
351 ; AVX512CD-NEXT: vpsubw %ymm2, %ymm0, %ymm0
352 ; AVX512CD-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm0
353 ; AVX512CD-NEXT: retq
354 ;
355 ; AVX512CDBW-LABEL: testv32i16:
356 ; AVX512CDBW: # %bb.0:
357 ; AVX512CDBW-NEXT: vpmovzxwd {{.*#+}} zmm1 = ymm0[0],zero,ymm0[1],zero,ymm0[2],zero,ymm0[3],zero,ymm0[4],zero,ymm0[5],zero,ymm0[6],zero,ymm0[7],zero,ymm0[8],zero,ymm0[9],zero,ymm0[10],zero,ymm0[11],zero,ymm0[12],zero,ymm0[13],zero,ymm0[14],zero,ymm0[15],zero
358 ; AVX512CDBW-NEXT: vplzcntd %zmm1, %zmm1
359 ; AVX512CDBW-NEXT: vpmovdw %zmm1, %ymm1
360 ; AVX512CDBW-NEXT: vextracti64x4 $1, %zmm0, %ymm0
361 ; AVX512CDBW-NEXT: vpmovzxwd {{.*#+}} zmm0 = ymm0[0],zero,ymm0[1],zero,ymm0[2],zero,ymm0[3],zero,ymm0[4],zero,ymm0[5],zero,ymm0[6],zero,ymm0[7],zero,ymm0[8],zero,ymm0[9],zero,ymm0[10],zero,ymm0[11],zero,ymm0[12],zero,ymm0[13],zero,ymm0[14],zero,ymm0[15],zero
362 ; AVX512CDBW-NEXT: vplzcntd %zmm0, %zmm0
363 ; AVX512CDBW-NEXT: vpmovdw %zmm0, %ymm0
364 ; AVX512CDBW-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm0
365 ; AVX512CDBW-NEXT: vpsubw {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm0
366 ; AVX512CDBW-NEXT: retq
367 ;
368 ; AVX512BW-LABEL: testv32i16:
369 ; AVX512BW: # %bb.0:
370 ; AVX512BW-NEXT: vpsrlw $4, %zmm0, %zmm1
371 ; AVX512BW-NEXT: vpandd {{\.?LCPI[0-9]+_[0-9]+}}(%rip){1to16}, %zmm1, %zmm1
372 ; AVX512BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm2 = [4,3,2,2,1,1,1,1,0,0,0,0,0,0,0,0,4,3,2,2,1,1,1,1,0,0,0,0,0,0,0,0,4,3,2,2,1,1,1,1,0,0,0,0,0,0,0,0,4,3,2,2,1,1,1,1,0,0,0,0,0,0,0,0]
373 ; AVX512BW-NEXT: # zmm2 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
374 ; AVX512BW-NEXT: vpshufb %zmm1, %zmm2, %zmm3
375 ; AVX512BW-NEXT: vpshufb %zmm0, %zmm2, %zmm2
376 ; AVX512BW-NEXT: vptestnmb %zmm1, %zmm1, %k0
377 ; AVX512BW-NEXT: vpmovm2b %k0, %zmm1
378 ; AVX512BW-NEXT: vpandq %zmm1, %zmm2, %zmm1
379 ; AVX512BW-NEXT: vpaddb %zmm3, %zmm1, %zmm1
380 ; AVX512BW-NEXT: vptestnmb %zmm0, %zmm0, %k0
381 ; AVX512BW-NEXT: vpmovm2b %k0, %zmm0
382 ; AVX512BW-NEXT: vpsrlw $8, %zmm0, %zmm0
383 ; AVX512BW-NEXT: vpandq %zmm0, %zmm1, %zmm0
384 ; AVX512BW-NEXT: vpsrlw $8, %zmm1, %zmm1
385 ; AVX512BW-NEXT: vpaddw %zmm0, %zmm1, %zmm0
386 ; AVX512BW-NEXT: retq
387 ;
388 ; AVX512DQ-LABEL: testv32i16:
389 ; AVX512DQ: # %bb.0:
390 ; AVX512DQ-NEXT: vextracti64x4 $1, %zmm0, %ymm1
391 ; AVX512DQ-NEXT: vbroadcasti128 {{.*#+}} ymm2 = [4,3,2,2,1,1,1,1,0,0,0,0,0,0,0,0,4,3,2,2,1,1,1,1,0,0,0,0,0,0,0,0]
392 ; AVX512DQ-NEXT: # ymm2 = mem[0,1,0,1]
393 ; AVX512DQ-NEXT: vpshufb %ymm1, %ymm2, %ymm3
394 ; AVX512DQ-NEXT: vpsrlw $4, %ymm1, %ymm4
395 ; AVX512DQ-NEXT: vpbroadcastb {{.*#+}} ymm5 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
396 ; AVX512DQ-NEXT: vpand %ymm5, %ymm4, %ymm4
397 ; AVX512DQ-NEXT: vpxor %xmm6, %xmm6, %xmm6
398 ; AVX512DQ-NEXT: vpcmpeqb %ymm6, %ymm4, %ymm7
399 ; AVX512DQ-NEXT: vpand %ymm7, %ymm3, %ymm3
400 ; AVX512DQ-NEXT: vpshufb %ymm4, %ymm2, %ymm4
401 ; AVX512DQ-NEXT: vpaddb %ymm4, %ymm3, %ymm3
402 ; AVX512DQ-NEXT: vpcmpeqb %ymm6, %ymm1, %ymm1
403 ; AVX512DQ-NEXT: vpsrlw $8, %ymm1, %ymm1
404 ; AVX512DQ-NEXT: vpand %ymm1, %ymm3, %ymm1
405 ; AVX512DQ-NEXT: vpsrlw $8, %ymm3, %ymm3
406 ; AVX512DQ-NEXT: vpaddw %ymm1, %ymm3, %ymm1
407 ; AVX512DQ-NEXT: vpshufb %ymm0, %ymm2, %ymm3
408 ; AVX512DQ-NEXT: vpsrlw $4, %ymm0, %ymm4
409 ; AVX512DQ-NEXT: vpand %ymm5, %ymm4, %ymm4
410 ; AVX512DQ-NEXT: vpcmpeqb %ymm6, %ymm4, %ymm5
411 ; AVX512DQ-NEXT: vpand %ymm5, %ymm3, %ymm3
412 ; AVX512DQ-NEXT: vpshufb %ymm4, %ymm2, %ymm2
413 ; AVX512DQ-NEXT: vpaddb %ymm2, %ymm3, %ymm2
414 ; AVX512DQ-NEXT: vpcmpeqb %ymm6, %ymm0, %ymm0
415 ; AVX512DQ-NEXT: vpsrlw $8, %ymm0, %ymm0
416 ; AVX512DQ-NEXT: vpand %ymm0, %ymm2, %ymm0
417 ; AVX512DQ-NEXT: vpsrlw $8, %ymm2, %ymm2
418 ; AVX512DQ-NEXT: vpaddw %ymm0, %ymm2, %ymm0
419 ; AVX512DQ-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm0
420 ; AVX512DQ-NEXT: retq
421 %out = call <32 x i16> @llvm.ctlz.v32i16(<32 x i16> %in, i1 0)
422 ret <32 x i16> %out
423 }
425 define <32 x i16> @testv32i16u(<32 x i16> %in) nounwind {
426 ; AVX512CD-LABEL: testv32i16u:
427 ; AVX512CD: # %bb.0:
428 ; AVX512CD-NEXT: vextracti64x4 $1, %zmm0, %ymm1
429 ; AVX512CD-NEXT: vpmovzxwd {{.*#+}} zmm1 = ymm1[0],zero,ymm1[1],zero,ymm1[2],zero,ymm1[3],zero,ymm1[4],zero,ymm1[5],zero,ymm1[6],zero,ymm1[7],zero,ymm1[8],zero,ymm1[9],zero,ymm1[10],zero,ymm1[11],zero,ymm1[12],zero,ymm1[13],zero,ymm1[14],zero,ymm1[15],zero
430 ; AVX512CD-NEXT: vplzcntd %zmm1, %zmm1
431 ; AVX512CD-NEXT: vpmovdw %zmm1, %ymm1
432 ; AVX512CD-NEXT: vpbroadcastw {{.*#+}} ymm2 = [16,16,16,16,16,16,16,16,16,16,16,16,16,16,16,16]
433 ; AVX512CD-NEXT: vpsubw %ymm2, %ymm1, %ymm1
434 ; AVX512CD-NEXT: vpmovzxwd {{.*#+}} zmm0 = ymm0[0],zero,ymm0[1],zero,ymm0[2],zero,ymm0[3],zero,ymm0[4],zero,ymm0[5],zero,ymm0[6],zero,ymm0[7],zero,ymm0[8],zero,ymm0[9],zero,ymm0[10],zero,ymm0[11],zero,ymm0[12],zero,ymm0[13],zero,ymm0[14],zero,ymm0[15],zero
435 ; AVX512CD-NEXT: vplzcntd %zmm0, %zmm0
436 ; AVX512CD-NEXT: vpmovdw %zmm0, %ymm0
437 ; AVX512CD-NEXT: vpsubw %ymm2, %ymm0, %ymm0
438 ; AVX512CD-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm0
439 ; AVX512CD-NEXT: retq
440 ;
441 ; AVX512CDBW-LABEL: testv32i16u:
442 ; AVX512CDBW: # %bb.0:
443 ; AVX512CDBW-NEXT: vpmovzxwd {{.*#+}} zmm1 = ymm0[0],zero,ymm0[1],zero,ymm0[2],zero,ymm0[3],zero,ymm0[4],zero,ymm0[5],zero,ymm0[6],zero,ymm0[7],zero,ymm0[8],zero,ymm0[9],zero,ymm0[10],zero,ymm0[11],zero,ymm0[12],zero,ymm0[13],zero,ymm0[14],zero,ymm0[15],zero
444 ; AVX512CDBW-NEXT: vplzcntd %zmm1, %zmm1
445 ; AVX512CDBW-NEXT: vpmovdw %zmm1, %ymm1
446 ; AVX512CDBW-NEXT: vextracti64x4 $1, %zmm0, %ymm0
447 ; AVX512CDBW-NEXT: vpmovzxwd {{.*#+}} zmm0 = ymm0[0],zero,ymm0[1],zero,ymm0[2],zero,ymm0[3],zero,ymm0[4],zero,ymm0[5],zero,ymm0[6],zero,ymm0[7],zero,ymm0[8],zero,ymm0[9],zero,ymm0[10],zero,ymm0[11],zero,ymm0[12],zero,ymm0[13],zero,ymm0[14],zero,ymm0[15],zero
448 ; AVX512CDBW-NEXT: vplzcntd %zmm0, %zmm0
449 ; AVX512CDBW-NEXT: vpmovdw %zmm0, %ymm0
450 ; AVX512CDBW-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm0
451 ; AVX512CDBW-NEXT: vpsubw {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm0
452 ; AVX512CDBW-NEXT: retq
453 ;
454 ; AVX512BW-LABEL: testv32i16u:
455 ; AVX512BW: # %bb.0:
456 ; AVX512BW-NEXT: vpsrlw $4, %zmm0, %zmm1
457 ; AVX512BW-NEXT: vpandd {{\.?LCPI[0-9]+_[0-9]+}}(%rip){1to16}, %zmm1, %zmm1
458 ; AVX512BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm2 = [4,3,2,2,1,1,1,1,0,0,0,0,0,0,0,0,4,3,2,2,1,1,1,1,0,0,0,0,0,0,0,0,4,3,2,2,1,1,1,1,0,0,0,0,0,0,0,0,4,3,2,2,1,1,1,1,0,0,0,0,0,0,0,0]
459 ; AVX512BW-NEXT: # zmm2 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
460 ; AVX512BW-NEXT: vpshufb %zmm1, %zmm2, %zmm3
461 ; AVX512BW-NEXT: vpshufb %zmm0, %zmm2, %zmm2
462 ; AVX512BW-NEXT: vptestnmb %zmm1, %zmm1, %k0
463 ; AVX512BW-NEXT: vpmovm2b %k0, %zmm1
464 ; AVX512BW-NEXT: vpandq %zmm1, %zmm2, %zmm1
465 ; AVX512BW-NEXT: vpaddb %zmm3, %zmm1, %zmm1
466 ; AVX512BW-NEXT: vptestnmb %zmm0, %zmm0, %k0
467 ; AVX512BW-NEXT: vpmovm2b %k0, %zmm0
468 ; AVX512BW-NEXT: vpsrlw $8, %zmm0, %zmm0
469 ; AVX512BW-NEXT: vpandq %zmm0, %zmm1, %zmm0
470 ; AVX512BW-NEXT: vpsrlw $8, %zmm1, %zmm1
471 ; AVX512BW-NEXT: vpaddw %zmm0, %zmm1, %zmm0
472 ; AVX512BW-NEXT: retq
473 ;
474 ; AVX512DQ-LABEL: testv32i16u:
475 ; AVX512DQ: # %bb.0:
476 ; AVX512DQ-NEXT: vextracti64x4 $1, %zmm0, %ymm1
477 ; AVX512DQ-NEXT: vbroadcasti128 {{.*#+}} ymm2 = [4,3,2,2,1,1,1,1,0,0,0,0,0,0,0,0,4,3,2,2,1,1,1,1,0,0,0,0,0,0,0,0]
478 ; AVX512DQ-NEXT: # ymm2 = mem[0,1,0,1]
479 ; AVX512DQ-NEXT: vpshufb %ymm1, %ymm2, %ymm3
480 ; AVX512DQ-NEXT: vpsrlw $4, %ymm1, %ymm4
481 ; AVX512DQ-NEXT: vpbroadcastb {{.*#+}} ymm5 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
482 ; AVX512DQ-NEXT: vpand %ymm5, %ymm4, %ymm4
483 ; AVX512DQ-NEXT: vpxor %xmm6, %xmm6, %xmm6
484 ; AVX512DQ-NEXT: vpcmpeqb %ymm6, %ymm4, %ymm7
485 ; AVX512DQ-NEXT: vpand %ymm7, %ymm3, %ymm3
486 ; AVX512DQ-NEXT: vpshufb %ymm4, %ymm2, %ymm4
487 ; AVX512DQ-NEXT: vpaddb %ymm4, %ymm3, %ymm3
488 ; AVX512DQ-NEXT: vpcmpeqb %ymm6, %ymm1, %ymm1
489 ; AVX512DQ-NEXT: vpsrlw $8, %ymm1, %ymm1
490 ; AVX512DQ-NEXT: vpand %ymm1, %ymm3, %ymm1
491 ; AVX512DQ-NEXT: vpsrlw $8, %ymm3, %ymm3
492 ; AVX512DQ-NEXT: vpaddw %ymm1, %ymm3, %ymm1
493 ; AVX512DQ-NEXT: vpshufb %ymm0, %ymm2, %ymm3
494 ; AVX512DQ-NEXT: vpsrlw $4, %ymm0, %ymm4
495 ; AVX512DQ-NEXT: vpand %ymm5, %ymm4, %ymm4
496 ; AVX512DQ-NEXT: vpcmpeqb %ymm6, %ymm4, %ymm5
497 ; AVX512DQ-NEXT: vpand %ymm5, %ymm3, %ymm3
498 ; AVX512DQ-NEXT: vpshufb %ymm4, %ymm2, %ymm2
499 ; AVX512DQ-NEXT: vpaddb %ymm2, %ymm3, %ymm2
500 ; AVX512DQ-NEXT: vpcmpeqb %ymm6, %ymm0, %ymm0
501 ; AVX512DQ-NEXT: vpsrlw $8, %ymm0, %ymm0
502 ; AVX512DQ-NEXT: vpand %ymm0, %ymm2, %ymm0
503 ; AVX512DQ-NEXT: vpsrlw $8, %ymm2, %ymm2
504 ; AVX512DQ-NEXT: vpaddw %ymm0, %ymm2, %ymm0
505 ; AVX512DQ-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm0
506 ; AVX512DQ-NEXT: retq
507 %out = call <32 x i16> @llvm.ctlz.v32i16(<32 x i16> %in, i1 -1)
508 ret <32 x i16> %out
509 }
511 define <64 x i8> @testv64i8(<64 x i8> %in) nounwind {
512 ; AVX512CD-LABEL: testv64i8:
513 ; AVX512CD: # %bb.0:
514 ; AVX512CD-NEXT: vextracti64x4 $1, %zmm0, %ymm1
515 ; AVX512CD-NEXT: vpmovzxbd {{.*#+}} zmm2 = xmm1[0],zero,zero,zero,xmm1[1],zero,zero,zero,xmm1[2],zero,zero,zero,xmm1[3],zero,zero,zero,xmm1[4],zero,zero,zero,xmm1[5],zero,zero,zero,xmm1[6],zero,zero,zero,xmm1[7],zero,zero,zero,xmm1[8],zero,zero,zero,xmm1[9],zero,zero,zero,xmm1[10],zero,zero,zero,xmm1[11],zero,zero,zero,xmm1[12],zero,zero,zero,xmm1[13],zero,zero,zero,xmm1[14],zero,zero,zero,xmm1[15],zero,zero,zero
516 ; AVX512CD-NEXT: vplzcntd %zmm2, %zmm2
517 ; AVX512CD-NEXT: vpmovdb %zmm2, %xmm2
518 ; AVX512CD-NEXT: vextracti128 $1, %ymm1, %xmm1
519 ; AVX512CD-NEXT: vpmovzxbd {{.*#+}} zmm1 = xmm1[0],zero,zero,zero,xmm1[1],zero,zero,zero,xmm1[2],zero,zero,zero,xmm1[3],zero,zero,zero,xmm1[4],zero,zero,zero,xmm1[5],zero,zero,zero,xmm1[6],zero,zero,zero,xmm1[7],zero,zero,zero,xmm1[8],zero,zero,zero,xmm1[9],zero,zero,zero,xmm1[10],zero,zero,zero,xmm1[11],zero,zero,zero,xmm1[12],zero,zero,zero,xmm1[13],zero,zero,zero,xmm1[14],zero,zero,zero,xmm1[15],zero,zero,zero
520 ; AVX512CD-NEXT: vplzcntd %zmm1, %zmm1
521 ; AVX512CD-NEXT: vpmovdb %zmm1, %xmm1
522 ; AVX512CD-NEXT: vinserti128 $1, %xmm1, %ymm2, %ymm1
523 ; AVX512CD-NEXT: vpbroadcastb {{.*#+}} ymm2 = [24,24,24,24,24,24,24,24,24,24,24,24,24,24,24,24,24,24,24,24,24,24,24,24,24,24,24,24,24,24,24,24]
524 ; AVX512CD-NEXT: vpsubb %ymm2, %ymm1, %ymm1
525 ; AVX512CD-NEXT: vpmovzxbd {{.*#+}} zmm3 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero,xmm0[2],zero,zero,zero,xmm0[3],zero,zero,zero,xmm0[4],zero,zero,zero,xmm0[5],zero,zero,zero,xmm0[6],zero,zero,zero,xmm0[7],zero,zero,zero,xmm0[8],zero,zero,zero,xmm0[9],zero,zero,zero,xmm0[10],zero,zero,zero,xmm0[11],zero,zero,zero,xmm0[12],zero,zero,zero,xmm0[13],zero,zero,zero,xmm0[14],zero,zero,zero,xmm0[15],zero,zero,zero
526 ; AVX512CD-NEXT: vplzcntd %zmm3, %zmm3
527 ; AVX512CD-NEXT: vpmovdb %zmm3, %xmm3
528 ; AVX512CD-NEXT: vextracti128 $1, %ymm0, %xmm0
529 ; AVX512CD-NEXT: vpmovzxbd {{.*#+}} zmm0 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero,xmm0[2],zero,zero,zero,xmm0[3],zero,zero,zero,xmm0[4],zero,zero,zero,xmm0[5],zero,zero,zero,xmm0[6],zero,zero,zero,xmm0[7],zero,zero,zero,xmm0[8],zero,zero,zero,xmm0[9],zero,zero,zero,xmm0[10],zero,zero,zero,xmm0[11],zero,zero,zero,xmm0[12],zero,zero,zero,xmm0[13],zero,zero,zero,xmm0[14],zero,zero,zero,xmm0[15],zero,zero,zero
530 ; AVX512CD-NEXT: vplzcntd %zmm0, %zmm0
531 ; AVX512CD-NEXT: vpmovdb %zmm0, %xmm0
532 ; AVX512CD-NEXT: vinserti128 $1, %xmm0, %ymm3, %ymm0
533 ; AVX512CD-NEXT: vpsubb %ymm2, %ymm0, %ymm0
534 ; AVX512CD-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm0
535 ; AVX512CD-NEXT: retq
536 ;
537 ; AVX512CDBW-LABEL: testv64i8:
538 ; AVX512CDBW: # %bb.0:
539 ; AVX512CDBW-NEXT: vextracti64x4 $1, %zmm0, %ymm1
540 ; AVX512CDBW-NEXT: vpmovzxbd {{.*#+}} zmm2 = xmm1[0],zero,zero,zero,xmm1[1],zero,zero,zero,xmm1[2],zero,zero,zero,xmm1[3],zero,zero,zero,xmm1[4],zero,zero,zero,xmm1[5],zero,zero,zero,xmm1[6],zero,zero,zero,xmm1[7],zero,zero,zero,xmm1[8],zero,zero,zero,xmm1[9],zero,zero,zero,xmm1[10],zero,zero,zero,xmm1[11],zero,zero,zero,xmm1[12],zero,zero,zero,xmm1[13],zero,zero,zero,xmm1[14],zero,zero,zero,xmm1[15],zero,zero,zero
541 ; AVX512CDBW-NEXT: vplzcntd %zmm2, %zmm2
542 ; AVX512CDBW-NEXT: vpmovdb %zmm2, %xmm2
543 ; AVX512CDBW-NEXT: vextracti128 $1, %ymm1, %xmm1
544 ; AVX512CDBW-NEXT: vpmovzxbd {{.*#+}} zmm1 = xmm1[0],zero,zero,zero,xmm1[1],zero,zero,zero,xmm1[2],zero,zero,zero,xmm1[3],zero,zero,zero,xmm1[4],zero,zero,zero,xmm1[5],zero,zero,zero,xmm1[6],zero,zero,zero,xmm1[7],zero,zero,zero,xmm1[8],zero,zero,zero,xmm1[9],zero,zero,zero,xmm1[10],zero,zero,zero,xmm1[11],zero,zero,zero,xmm1[12],zero,zero,zero,xmm1[13],zero,zero,zero,xmm1[14],zero,zero,zero,xmm1[15],zero,zero,zero
545 ; AVX512CDBW-NEXT: vplzcntd %zmm1, %zmm1
546 ; AVX512CDBW-NEXT: vpmovdb %zmm1, %xmm1
547 ; AVX512CDBW-NEXT: vinserti128 $1, %xmm1, %ymm2, %ymm1
548 ; AVX512CDBW-NEXT: vpmovzxbd {{.*#+}} zmm2 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero,xmm0[2],zero,zero,zero,xmm0[3],zero,zero,zero,xmm0[4],zero,zero,zero,xmm0[5],zero,zero,zero,xmm0[6],zero,zero,zero,xmm0[7],zero,zero,zero,xmm0[8],zero,zero,zero,xmm0[9],zero,zero,zero,xmm0[10],zero,zero,zero,xmm0[11],zero,zero,zero,xmm0[12],zero,zero,zero,xmm0[13],zero,zero,zero,xmm0[14],zero,zero,zero,xmm0[15],zero,zero,zero
549 ; AVX512CDBW-NEXT: vplzcntd %zmm2, %zmm2
550 ; AVX512CDBW-NEXT: vpmovdb %zmm2, %xmm2
551 ; AVX512CDBW-NEXT: vextracti128 $1, %ymm0, %xmm0
552 ; AVX512CDBW-NEXT: vpmovzxbd {{.*#+}} zmm0 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero,xmm0[2],zero,zero,zero,xmm0[3],zero,zero,zero,xmm0[4],zero,zero,zero,xmm0[5],zero,zero,zero,xmm0[6],zero,zero,zero,xmm0[7],zero,zero,zero,xmm0[8],zero,zero,zero,xmm0[9],zero,zero,zero,xmm0[10],zero,zero,zero,xmm0[11],zero,zero,zero,xmm0[12],zero,zero,zero,xmm0[13],zero,zero,zero,xmm0[14],zero,zero,zero,xmm0[15],zero,zero,zero
553 ; AVX512CDBW-NEXT: vplzcntd %zmm0, %zmm0
554 ; AVX512CDBW-NEXT: vpmovdb %zmm0, %xmm0
555 ; AVX512CDBW-NEXT: vinserti128 $1, %xmm0, %ymm2, %ymm0
556 ; AVX512CDBW-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm0
557 ; AVX512CDBW-NEXT: vpsubb {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm0
558 ; AVX512CDBW-NEXT: retq
559 ;
560 ; AVX512BW-LABEL: testv64i8:
561 ; AVX512BW: # %bb.0:
562 ; AVX512BW-NEXT: vpsrlw $4, %zmm0, %zmm1
563 ; AVX512BW-NEXT: vpandd {{\.?LCPI[0-9]+_[0-9]+}}(%rip){1to16}, %zmm1, %zmm1
564 ; AVX512BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm2 = [4,3,2,2,1,1,1,1,0,0,0,0,0,0,0,0,4,3,2,2,1,1,1,1,0,0,0,0,0,0,0,0,4,3,2,2,1,1,1,1,0,0,0,0,0,0,0,0,4,3,2,2,1,1,1,1,0,0,0,0,0,0,0,0]
565 ; AVX512BW-NEXT: # zmm2 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
566 ; AVX512BW-NEXT: vpshufb %zmm1, %zmm2, %zmm3
567 ; AVX512BW-NEXT: vpshufb %zmm0, %zmm2, %zmm0
568 ; AVX512BW-NEXT: vptestnmb %zmm1, %zmm1, %k0
569 ; AVX512BW-NEXT: vpmovm2b %k0, %zmm1
570 ; AVX512BW-NEXT: vpandq %zmm1, %zmm0, %zmm0
571 ; AVX512BW-NEXT: vpaddb %zmm3, %zmm0, %zmm0
572 ; AVX512BW-NEXT: retq
573 ;
574 ; AVX512DQ-LABEL: testv64i8:
575 ; AVX512DQ: # %bb.0:
576 ; AVX512DQ-NEXT: vextracti64x4 $1, %zmm0, %ymm1
577 ; AVX512DQ-NEXT: vbroadcasti128 {{.*#+}} ymm2 = [4,3,2,2,1,1,1,1,0,0,0,0,0,0,0,0,4,3,2,2,1,1,1,1,0,0,0,0,0,0,0,0]
578 ; AVX512DQ-NEXT: # ymm2 = mem[0,1,0,1]
579 ; AVX512DQ-NEXT: vpshufb %ymm1, %ymm2, %ymm3
580 ; AVX512DQ-NEXT: vpsrlw $4, %ymm1, %ymm1
581 ; AVX512DQ-NEXT: vpbroadcastb {{.*#+}} ymm4 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
582 ; AVX512DQ-NEXT: vpand %ymm4, %ymm1, %ymm1
583 ; AVX512DQ-NEXT: vpxor %xmm5, %xmm5, %xmm5
584 ; AVX512DQ-NEXT: vpcmpeqb %ymm5, %ymm1, %ymm6
585 ; AVX512DQ-NEXT: vpand %ymm6, %ymm3, %ymm3
586 ; AVX512DQ-NEXT: vpshufb %ymm1, %ymm2, %ymm1
587 ; AVX512DQ-NEXT: vpaddb %ymm1, %ymm3, %ymm1
588 ; AVX512DQ-NEXT: vpshufb %ymm0, %ymm2, %ymm3
589 ; AVX512DQ-NEXT: vpsrlw $4, %ymm0, %ymm0
590 ; AVX512DQ-NEXT: vpand %ymm4, %ymm0, %ymm0
591 ; AVX512DQ-NEXT: vpcmpeqb %ymm5, %ymm0, %ymm4
592 ; AVX512DQ-NEXT: vpand %ymm4, %ymm3, %ymm3
593 ; AVX512DQ-NEXT: vpshufb %ymm0, %ymm2, %ymm0
594 ; AVX512DQ-NEXT: vpaddb %ymm0, %ymm3, %ymm0
595 ; AVX512DQ-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm0
596 ; AVX512DQ-NEXT: retq
597 %out = call <64 x i8> @llvm.ctlz.v64i8(<64 x i8> %in, i1 0)
598 ret <64 x i8> %out
599 }
601 define <64 x i8> @testv64i8u(<64 x i8> %in) nounwind {
602 ; AVX512CD-LABEL: testv64i8u:
603 ; AVX512CD: # %bb.0:
604 ; AVX512CD-NEXT: vextracti64x4 $1, %zmm0, %ymm1
605 ; AVX512CD-NEXT: vpmovzxbd {{.*#+}} zmm2 = xmm1[0],zero,zero,zero,xmm1[1],zero,zero,zero,xmm1[2],zero,zero,zero,xmm1[3],zero,zero,zero,xmm1[4],zero,zero,zero,xmm1[5],zero,zero,zero,xmm1[6],zero,zero,zero,xmm1[7],zero,zero,zero,xmm1[8],zero,zero,zero,xmm1[9],zero,zero,zero,xmm1[10],zero,zero,zero,xmm1[11],zero,zero,zero,xmm1[12],zero,zero,zero,xmm1[13],zero,zero,zero,xmm1[14],zero,zero,zero,xmm1[15],zero,zero,zero
606 ; AVX512CD-NEXT: vplzcntd %zmm2, %zmm2
607 ; AVX512CD-NEXT: vpmovdb %zmm2, %xmm2
608 ; AVX512CD-NEXT: vextracti128 $1, %ymm1, %xmm1
609 ; AVX512CD-NEXT: vpmovzxbd {{.*#+}} zmm1 = xmm1[0],zero,zero,zero,xmm1[1],zero,zero,zero,xmm1[2],zero,zero,zero,xmm1[3],zero,zero,zero,xmm1[4],zero,zero,zero,xmm1[5],zero,zero,zero,xmm1[6],zero,zero,zero,xmm1[7],zero,zero,zero,xmm1[8],zero,zero,zero,xmm1[9],zero,zero,zero,xmm1[10],zero,zero,zero,xmm1[11],zero,zero,zero,xmm1[12],zero,zero,zero,xmm1[13],zero,zero,zero,xmm1[14],zero,zero,zero,xmm1[15],zero,zero,zero
610 ; AVX512CD-NEXT: vplzcntd %zmm1, %zmm1
611 ; AVX512CD-NEXT: vpmovdb %zmm1, %xmm1
612 ; AVX512CD-NEXT: vinserti128 $1, %xmm1, %ymm2, %ymm1
613 ; AVX512CD-NEXT: vpbroadcastb {{.*#+}} ymm2 = [24,24,24,24,24,24,24,24,24,24,24,24,24,24,24,24,24,24,24,24,24,24,24,24,24,24,24,24,24,24,24,24]
614 ; AVX512CD-NEXT: vpsubb %ymm2, %ymm1, %ymm1
615 ; AVX512CD-NEXT: vpmovzxbd {{.*#+}} zmm3 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero,xmm0[2],zero,zero,zero,xmm0[3],zero,zero,zero,xmm0[4],zero,zero,zero,xmm0[5],zero,zero,zero,xmm0[6],zero,zero,zero,xmm0[7],zero,zero,zero,xmm0[8],zero,zero,zero,xmm0[9],zero,zero,zero,xmm0[10],zero,zero,zero,xmm0[11],zero,zero,zero,xmm0[12],zero,zero,zero,xmm0[13],zero,zero,zero,xmm0[14],zero,zero,zero,xmm0[15],zero,zero,zero
616 ; AVX512CD-NEXT: vplzcntd %zmm3, %zmm3
617 ; AVX512CD-NEXT: vpmovdb %zmm3, %xmm3
618 ; AVX512CD-NEXT: vextracti128 $1, %ymm0, %xmm0
619 ; AVX512CD-NEXT: vpmovzxbd {{.*#+}} zmm0 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero,xmm0[2],zero,zero,zero,xmm0[3],zero,zero,zero,xmm0[4],zero,zero,zero,xmm0[5],zero,zero,zero,xmm0[6],zero,zero,zero,xmm0[7],zero,zero,zero,xmm0[8],zero,zero,zero,xmm0[9],zero,zero,zero,xmm0[10],zero,zero,zero,xmm0[11],zero,zero,zero,xmm0[12],zero,zero,zero,xmm0[13],zero,zero,zero,xmm0[14],zero,zero,zero,xmm0[15],zero,zero,zero
620 ; AVX512CD-NEXT: vplzcntd %zmm0, %zmm0
621 ; AVX512CD-NEXT: vpmovdb %zmm0, %xmm0
622 ; AVX512CD-NEXT: vinserti128 $1, %xmm0, %ymm3, %ymm0
623 ; AVX512CD-NEXT: vpsubb %ymm2, %ymm0, %ymm0
624 ; AVX512CD-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm0
625 ; AVX512CD-NEXT: retq
626 ;
627 ; AVX512CDBW-LABEL: testv64i8u:
628 ; AVX512CDBW: # %bb.0:
629 ; AVX512CDBW-NEXT: vextracti64x4 $1, %zmm0, %ymm1
630 ; AVX512CDBW-NEXT: vpmovzxbd {{.*#+}} zmm2 = xmm1[0],zero,zero,zero,xmm1[1],zero,zero,zero,xmm1[2],zero,zero,zero,xmm1[3],zero,zero,zero,xmm1[4],zero,zero,zero,xmm1[5],zero,zero,zero,xmm1[6],zero,zero,zero,xmm1[7],zero,zero,zero,xmm1[8],zero,zero,zero,xmm1[9],zero,zero,zero,xmm1[10],zero,zero,zero,xmm1[11],zero,zero,zero,xmm1[12],zero,zero,zero,xmm1[13],zero,zero,zero,xmm1[14],zero,zero,zero,xmm1[15],zero,zero,zero
631 ; AVX512CDBW-NEXT: vplzcntd %zmm2, %zmm2
632 ; AVX512CDBW-NEXT: vpmovdb %zmm2, %xmm2
633 ; AVX512CDBW-NEXT: vextracti128 $1, %ymm1, %xmm1
634 ; AVX512CDBW-NEXT: vpmovzxbd {{.*#+}} zmm1 = xmm1[0],zero,zero,zero,xmm1[1],zero,zero,zero,xmm1[2],zero,zero,zero,xmm1[3],zero,zero,zero,xmm1[4],zero,zero,zero,xmm1[5],zero,zero,zero,xmm1[6],zero,zero,zero,xmm1[7],zero,zero,zero,xmm1[8],zero,zero,zero,xmm1[9],zero,zero,zero,xmm1[10],zero,zero,zero,xmm1[11],zero,zero,zero,xmm1[12],zero,zero,zero,xmm1[13],zero,zero,zero,xmm1[14],zero,zero,zero,xmm1[15],zero,zero,zero
635 ; AVX512CDBW-NEXT: vplzcntd %zmm1, %zmm1
636 ; AVX512CDBW-NEXT: vpmovdb %zmm1, %xmm1
637 ; AVX512CDBW-NEXT: vinserti128 $1, %xmm1, %ymm2, %ymm1
638 ; AVX512CDBW-NEXT: vpmovzxbd {{.*#+}} zmm2 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero,xmm0[2],zero,zero,zero,xmm0[3],zero,zero,zero,xmm0[4],zero,zero,zero,xmm0[5],zero,zero,zero,xmm0[6],zero,zero,zero,xmm0[7],zero,zero,zero,xmm0[8],zero,zero,zero,xmm0[9],zero,zero,zero,xmm0[10],zero,zero,zero,xmm0[11],zero,zero,zero,xmm0[12],zero,zero,zero,xmm0[13],zero,zero,zero,xmm0[14],zero,zero,zero,xmm0[15],zero,zero,zero
639 ; AVX512CDBW-NEXT: vplzcntd %zmm2, %zmm2
640 ; AVX512CDBW-NEXT: vpmovdb %zmm2, %xmm2
641 ; AVX512CDBW-NEXT: vextracti128 $1, %ymm0, %xmm0
642 ; AVX512CDBW-NEXT: vpmovzxbd {{.*#+}} zmm0 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero,xmm0[2],zero,zero,zero,xmm0[3],zero,zero,zero,xmm0[4],zero,zero,zero,xmm0[5],zero,zero,zero,xmm0[6],zero,zero,zero,xmm0[7],zero,zero,zero,xmm0[8],zero,zero,zero,xmm0[9],zero,zero,zero,xmm0[10],zero,zero,zero,xmm0[11],zero,zero,zero,xmm0[12],zero,zero,zero,xmm0[13],zero,zero,zero,xmm0[14],zero,zero,zero,xmm0[15],zero,zero,zero
643 ; AVX512CDBW-NEXT: vplzcntd %zmm0, %zmm0
644 ; AVX512CDBW-NEXT: vpmovdb %zmm0, %xmm0
645 ; AVX512CDBW-NEXT: vinserti128 $1, %xmm0, %ymm2, %ymm0
646 ; AVX512CDBW-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm0
647 ; AVX512CDBW-NEXT: vpsubb {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %zmm0
648 ; AVX512CDBW-NEXT: retq
649 ;
650 ; AVX512BW-LABEL: testv64i8u:
651 ; AVX512BW: # %bb.0:
652 ; AVX512BW-NEXT: vpsrlw $4, %zmm0, %zmm1
653 ; AVX512BW-NEXT: vpandd {{\.?LCPI[0-9]+_[0-9]+}}(%rip){1to16}, %zmm1, %zmm1
654 ; AVX512BW-NEXT: vbroadcasti32x4 {{.*#+}} zmm2 = [4,3,2,2,1,1,1,1,0,0,0,0,0,0,0,0,4,3,2,2,1,1,1,1,0,0,0,0,0,0,0,0,4,3,2,2,1,1,1,1,0,0,0,0,0,0,0,0,4,3,2,2,1,1,1,1,0,0,0,0,0,0,0,0]
655 ; AVX512BW-NEXT: # zmm2 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
656 ; AVX512BW-NEXT: vpshufb %zmm1, %zmm2, %zmm3
657 ; AVX512BW-NEXT: vpshufb %zmm0, %zmm2, %zmm0
658 ; AVX512BW-NEXT: vptestnmb %zmm1, %zmm1, %k0
659 ; AVX512BW-NEXT: vpmovm2b %k0, %zmm1
660 ; AVX512BW-NEXT: vpandq %zmm1, %zmm0, %zmm0
661 ; AVX512BW-NEXT: vpaddb %zmm3, %zmm0, %zmm0
662 ; AVX512BW-NEXT: retq
663 ;
664 ; AVX512DQ-LABEL: testv64i8u:
665 ; AVX512DQ: # %bb.0:
666 ; AVX512DQ-NEXT: vextracti64x4 $1, %zmm0, %ymm1
667 ; AVX512DQ-NEXT: vbroadcasti128 {{.*#+}} ymm2 = [4,3,2,2,1,1,1,1,0,0,0,0,0,0,0,0,4,3,2,2,1,1,1,1,0,0,0,0,0,0,0,0]
668 ; AVX512DQ-NEXT: # ymm2 = mem[0,1,0,1]
669 ; AVX512DQ-NEXT: vpshufb %ymm1, %ymm2, %ymm3
670 ; AVX512DQ-NEXT: vpsrlw $4, %ymm1, %ymm1
671 ; AVX512DQ-NEXT: vpbroadcastb {{.*#+}} ymm4 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
672 ; AVX512DQ-NEXT: vpand %ymm4, %ymm1, %ymm1
673 ; AVX512DQ-NEXT: vpxor %xmm5, %xmm5, %xmm5
674 ; AVX512DQ-NEXT: vpcmpeqb %ymm5, %ymm1, %ymm6
675 ; AVX512DQ-NEXT: vpand %ymm6, %ymm3, %ymm3
676 ; AVX512DQ-NEXT: vpshufb %ymm1, %ymm2, %ymm1
677 ; AVX512DQ-NEXT: vpaddb %ymm1, %ymm3, %ymm1
678 ; AVX512DQ-NEXT: vpshufb %ymm0, %ymm2, %ymm3
679 ; AVX512DQ-NEXT: vpsrlw $4, %ymm0, %ymm0
680 ; AVX512DQ-NEXT: vpand %ymm4, %ymm0, %ymm0
681 ; AVX512DQ-NEXT: vpcmpeqb %ymm5, %ymm0, %ymm4
682 ; AVX512DQ-NEXT: vpand %ymm4, %ymm3, %ymm3
683 ; AVX512DQ-NEXT: vpshufb %ymm0, %ymm2, %ymm0
684 ; AVX512DQ-NEXT: vpaddb %ymm0, %ymm3, %ymm0
685 ; AVX512DQ-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm0
686 ; AVX512DQ-NEXT: retq
687 %out = call <64 x i8> @llvm.ctlz.v64i8(<64 x i8> %in, i1 -1)
688 ret <64 x i8> %out
689 }
691 declare <8 x i64> @llvm.ctlz.v8i64(<8 x i64>, i1)
692 declare <16 x i32> @llvm.ctlz.v16i32(<16 x i32>, i1)
693 declare <32 x i16> @llvm.ctlz.v32i16(<32 x i16>, i1)
694 declare <64 x i8> @llvm.ctlz.v64i8(<64 x i8>, i1)