| blob | 95c8aa6be9f434d181d4cd0db7aab904d1a319d6 |
1 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
2 ; RUN: llc -mtriple=x86_64-unknown-linux-gnu -mattr=+avx512f,+avx512vl,+avx512dq,+fast-variable-crosslane-shuffle,+fast-variable-perlane-shuffle %s -o - | FileCheck %s
3 ; RUN: llc -mtriple=x86_64-unknown-linux-gnu -mattr=+avx512f,+avx512vl,+avx512dq,+fast-variable-perlane-shuffle %s -o - | FileCheck %s
5 define <4 x i32> @test_2xi32_to_4xi32(<4 x i32> %vec) {
6 ; CHECK-LABEL: test_2xi32_to_4xi32:
7 ; CHECK: # %bb.0:
8 ; CHECK-NEXT: vmovddup {{.*#+}} xmm0 = xmm0[0,0]
9 ; CHECK-NEXT: retq
10 %res = shufflevector <4 x i32> %vec, <4 x i32> undef, <4 x i32> <i32 0, i32 1, i32 0, i32 1>
11 ret <4 x i32> %res
12 }
13 define <4 x i32> @test_masked_2xi32_to_4xi32_mask0(<4 x i32> %vec, <4 x i32> %default, <4 x i32> %mask) {
14 ; CHECK-LABEL: test_masked_2xi32_to_4xi32_mask0:
15 ; CHECK: # %bb.0:
16 ; CHECK-NEXT: vptestnmd %xmm2, %xmm2, %k1
17 ; CHECK-NEXT: vbroadcasti32x2 {{.*#+}} xmm1 {%k1} = xmm0[0,1,0,1]
18 ; CHECK-NEXT: vmovdqa %xmm1, %xmm0
19 ; CHECK-NEXT: retq
20 %shuf = shufflevector <4 x i32> %vec, <4 x i32> undef, <4 x i32> <i32 0, i32 1, i32 0, i32 1>
21 %cmp = icmp eq <4 x i32> %mask, zeroinitializer
22 %res = select <4 x i1> %cmp, <4 x i32> %shuf, <4 x i32> %default
23 ret <4 x i32> %res
24 }
26 define <4 x i32> @test_masked_z_2xi32_to_4xi32_mask0(<4 x i32> %vec, <4 x i32> %mask) {
27 ; CHECK-LABEL: test_masked_z_2xi32_to_4xi32_mask0:
28 ; CHECK: # %bb.0:
29 ; CHECK-NEXT: vptestnmd %xmm1, %xmm1, %k1
30 ; CHECK-NEXT: vbroadcasti32x2 {{.*#+}} xmm0 {%k1} {z} = xmm0[0,1,0,1]
31 ; CHECK-NEXT: retq
32 %shuf = shufflevector <4 x i32> %vec, <4 x i32> undef, <4 x i32> <i32 0, i32 1, i32 0, i32 1>
33 %cmp = icmp eq <4 x i32> %mask, zeroinitializer
34 %res = select <4 x i1> %cmp, <4 x i32> %shuf, <4 x i32> zeroinitializer
35 ret <4 x i32> %res
36 }
37 define <4 x i32> @test_masked_2xi32_to_4xi32_mask1(<4 x i32> %vec, <4 x i32> %default, <4 x i32> %mask) {
38 ; CHECK-LABEL: test_masked_2xi32_to_4xi32_mask1:
39 ; CHECK: # %bb.0:
40 ; CHECK-NEXT: vptestnmd %xmm2, %xmm2, %k1
41 ; CHECK-NEXT: vbroadcasti32x2 {{.*#+}} xmm1 {%k1} = xmm0[0,1,0,1]
42 ; CHECK-NEXT: vmovdqa %xmm1, %xmm0
43 ; CHECK-NEXT: retq
44 %shuf = shufflevector <4 x i32> %vec, <4 x i32> undef, <4 x i32> <i32 0, i32 1, i32 0, i32 1>
45 %cmp = icmp eq <4 x i32> %mask, zeroinitializer
46 %res = select <4 x i1> %cmp, <4 x i32> %shuf, <4 x i32> %default
47 ret <4 x i32> %res
48 }
50 define <4 x i32> @test_masked_z_2xi32_to_4xi32_mask1(<4 x i32> %vec, <4 x i32> %mask) {
51 ; CHECK-LABEL: test_masked_z_2xi32_to_4xi32_mask1:
52 ; CHECK: # %bb.0:
53 ; CHECK-NEXT: vptestnmd %xmm1, %xmm1, %k1
54 ; CHECK-NEXT: vbroadcasti32x2 {{.*#+}} xmm0 {%k1} {z} = xmm0[0,1,0,1]
55 ; CHECK-NEXT: retq
56 %shuf = shufflevector <4 x i32> %vec, <4 x i32> undef, <4 x i32> <i32 0, i32 1, i32 0, i32 1>
57 %cmp = icmp eq <4 x i32> %mask, zeroinitializer
58 %res = select <4 x i1> %cmp, <4 x i32> %shuf, <4 x i32> zeroinitializer
59 ret <4 x i32> %res
60 }
61 define <4 x i32> @test_masked_2xi32_to_4xi32_mask2(<4 x i32> %vec, <4 x i32> %default, <4 x i32> %mask) {
62 ; CHECK-LABEL: test_masked_2xi32_to_4xi32_mask2:
63 ; CHECK: # %bb.0:
64 ; CHECK-NEXT: vptestnmd %xmm2, %xmm2, %k1
65 ; CHECK-NEXT: vbroadcasti32x2 {{.*#+}} xmm1 {%k1} = xmm0[0,1,0,1]
66 ; CHECK-NEXT: vmovdqa %xmm1, %xmm0
67 ; CHECK-NEXT: retq
68 %shuf = shufflevector <4 x i32> %vec, <4 x i32> undef, <4 x i32> <i32 0, i32 1, i32 0, i32 1>
69 %cmp = icmp eq <4 x i32> %mask, zeroinitializer
70 %res = select <4 x i1> %cmp, <4 x i32> %shuf, <4 x i32> %default
71 ret <4 x i32> %res
72 }
74 define <4 x i32> @test_masked_z_2xi32_to_4xi32_mask2(<4 x i32> %vec, <4 x i32> %mask) {
75 ; CHECK-LABEL: test_masked_z_2xi32_to_4xi32_mask2:
76 ; CHECK: # %bb.0:
77 ; CHECK-NEXT: vptestnmd %xmm1, %xmm1, %k1
78 ; CHECK-NEXT: vbroadcasti32x2 {{.*#+}} xmm0 {%k1} {z} = xmm0[0,1,0,1]
79 ; CHECK-NEXT: retq
80 %shuf = shufflevector <4 x i32> %vec, <4 x i32> undef, <4 x i32> <i32 0, i32 1, i32 0, i32 1>
81 %cmp = icmp eq <4 x i32> %mask, zeroinitializer
82 %res = select <4 x i1> %cmp, <4 x i32> %shuf, <4 x i32> zeroinitializer
83 ret <4 x i32> %res
84 }
85 define <4 x i32> @test_masked_2xi32_to_4xi32_mask3(<4 x i32> %vec, <4 x i32> %default, <4 x i32> %mask) {
86 ; CHECK-LABEL: test_masked_2xi32_to_4xi32_mask3:
87 ; CHECK: # %bb.0:
88 ; CHECK-NEXT: vptestnmd %xmm2, %xmm2, %k1
89 ; CHECK-NEXT: vbroadcasti32x2 {{.*#+}} xmm1 {%k1} = xmm0[0,1,0,1]
90 ; CHECK-NEXT: vmovdqa %xmm1, %xmm0
91 ; CHECK-NEXT: retq
92 %shuf = shufflevector <4 x i32> %vec, <4 x i32> undef, <4 x i32> <i32 0, i32 1, i32 0, i32 1>
93 %cmp = icmp eq <4 x i32> %mask, zeroinitializer
94 %res = select <4 x i1> %cmp, <4 x i32> %shuf, <4 x i32> %default
95 ret <4 x i32> %res
96 }
98 define <4 x i32> @test_masked_z_2xi32_to_4xi32_mask3(<4 x i32> %vec, <4 x i32> %mask) {
99 ; CHECK-LABEL: test_masked_z_2xi32_to_4xi32_mask3:
100 ; CHECK: # %bb.0:
101 ; CHECK-NEXT: vptestnmd %xmm1, %xmm1, %k1
102 ; CHECK-NEXT: vbroadcasti32x2 {{.*#+}} xmm0 {%k1} {z} = xmm0[0,1,0,1]
103 ; CHECK-NEXT: retq
104 %shuf = shufflevector <4 x i32> %vec, <4 x i32> undef, <4 x i32> <i32 0, i32 1, i32 0, i32 1>
105 %cmp = icmp eq <4 x i32> %mask, zeroinitializer
106 %res = select <4 x i1> %cmp, <4 x i32> %shuf, <4 x i32> zeroinitializer
107 ret <4 x i32> %res
108 }
109 define <8 x i32> @test_2xi32_to_8xi32(<8 x i32> %vec) {
110 ; CHECK-LABEL: test_2xi32_to_8xi32:
111 ; CHECK: # %bb.0:
112 ; CHECK-NEXT: vbroadcastsd %xmm0, %ymm0
113 ; CHECK-NEXT: retq
114 %res = shufflevector <8 x i32> %vec, <8 x i32> undef, <8 x i32> <i32 0, i32 1, i32 0, i32 1, i32 0, i32 1, i32 0, i32 1>
115 ret <8 x i32> %res
116 }
117 define <8 x i32> @test_masked_2xi32_to_8xi32_mask0(<8 x i32> %vec, <8 x i32> %default, <8 x i32> %mask) {
118 ; CHECK-LABEL: test_masked_2xi32_to_8xi32_mask0:
119 ; CHECK: # %bb.0:
120 ; CHECK-NEXT: vptestnmd %ymm2, %ymm2, %k1
121 ; CHECK-NEXT: vbroadcasti32x2 {{.*#+}} ymm1 {%k1} = xmm0[0,1,0,1,0,1,0,1]
122 ; CHECK-NEXT: vmovdqa %ymm1, %ymm0
123 ; CHECK-NEXT: retq
124 %shuf = shufflevector <8 x i32> %vec, <8 x i32> undef, <8 x i32> <i32 0, i32 1, i32 0, i32 1, i32 0, i32 1, i32 0, i32 1>
125 %cmp = icmp eq <8 x i32> %mask, zeroinitializer
126 %res = select <8 x i1> %cmp, <8 x i32> %shuf, <8 x i32> %default
127 ret <8 x i32> %res
128 }
130 define <8 x i32> @test_masked_z_2xi32_to_8xi32_mask0(<8 x i32> %vec, <8 x i32> %mask) {
131 ; CHECK-LABEL: test_masked_z_2xi32_to_8xi32_mask0:
132 ; CHECK: # %bb.0:
133 ; CHECK-NEXT: vptestnmd %ymm1, %ymm1, %k1
134 ; CHECK-NEXT: vbroadcasti32x2 {{.*#+}} ymm0 {%k1} {z} = xmm0[0,1,0,1,0,1,0,1]
135 ; CHECK-NEXT: retq
136 %shuf = shufflevector <8 x i32> %vec, <8 x i32> undef, <8 x i32> <i32 0, i32 1, i32 0, i32 1, i32 0, i32 1, i32 0, i32 1>
137 %cmp = icmp eq <8 x i32> %mask, zeroinitializer
138 %res = select <8 x i1> %cmp, <8 x i32> %shuf, <8 x i32> zeroinitializer
139 ret <8 x i32> %res
140 }
141 define <8 x i32> @test_masked_2xi32_to_8xi32_mask1(<8 x i32> %vec, <8 x i32> %default, <8 x i32> %mask) {
142 ; CHECK-LABEL: test_masked_2xi32_to_8xi32_mask1:
143 ; CHECK: # %bb.0:
144 ; CHECK-NEXT: vptestnmd %ymm2, %ymm2, %k1
145 ; CHECK-NEXT: vbroadcasti32x2 {{.*#+}} ymm1 {%k1} = xmm0[0,1,0,1,0,1,0,1]
146 ; CHECK-NEXT: vmovdqa %ymm1, %ymm0
147 ; CHECK-NEXT: retq
148 %shuf = shufflevector <8 x i32> %vec, <8 x i32> undef, <8 x i32> <i32 0, i32 1, i32 0, i32 1, i32 0, i32 1, i32 0, i32 1>
149 %cmp = icmp eq <8 x i32> %mask, zeroinitializer
150 %res = select <8 x i1> %cmp, <8 x i32> %shuf, <8 x i32> %default
151 ret <8 x i32> %res
152 }
154 define <8 x i32> @test_masked_z_2xi32_to_8xi32_mask1(<8 x i32> %vec, <8 x i32> %mask) {
155 ; CHECK-LABEL: test_masked_z_2xi32_to_8xi32_mask1:
156 ; CHECK: # %bb.0:
157 ; CHECK-NEXT: vptestnmd %ymm1, %ymm1, %k1
158 ; CHECK-NEXT: vbroadcasti32x2 {{.*#+}} ymm0 {%k1} {z} = xmm0[0,1,0,1,0,1,0,1]
159 ; CHECK-NEXT: retq
160 %shuf = shufflevector <8 x i32> %vec, <8 x i32> undef, <8 x i32> <i32 0, i32 1, i32 0, i32 1, i32 0, i32 1, i32 0, i32 1>
161 %cmp = icmp eq <8 x i32> %mask, zeroinitializer
162 %res = select <8 x i1> %cmp, <8 x i32> %shuf, <8 x i32> zeroinitializer
163 ret <8 x i32> %res
164 }
165 define <8 x i32> @test_masked_2xi32_to_8xi32_mask2(<8 x i32> %vec, <8 x i32> %default, <8 x i32> %mask) {
166 ; CHECK-LABEL: test_masked_2xi32_to_8xi32_mask2:
167 ; CHECK: # %bb.0:
168 ; CHECK-NEXT: vptestnmd %ymm2, %ymm2, %k1
169 ; CHECK-NEXT: vbroadcasti32x2 {{.*#+}} ymm1 {%k1} = xmm0[0,1,0,1,0,1,0,1]
170 ; CHECK-NEXT: vmovdqa %ymm1, %ymm0
171 ; CHECK-NEXT: retq
172 %shuf = shufflevector <8 x i32> %vec, <8 x i32> undef, <8 x i32> <i32 0, i32 1, i32 0, i32 1, i32 0, i32 1, i32 0, i32 1>
173 %cmp = icmp eq <8 x i32> %mask, zeroinitializer
174 %res = select <8 x i1> %cmp, <8 x i32> %shuf, <8 x i32> %default
175 ret <8 x i32> %res
176 }
178 define <8 x i32> @test_masked_z_2xi32_to_8xi32_mask2(<8 x i32> %vec, <8 x i32> %mask) {
179 ; CHECK-LABEL: test_masked_z_2xi32_to_8xi32_mask2:
180 ; CHECK: # %bb.0:
181 ; CHECK-NEXT: vptestnmd %ymm1, %ymm1, %k1
182 ; CHECK-NEXT: vbroadcasti32x2 {{.*#+}} ymm0 {%k1} {z} = xmm0[0,1,0,1,0,1,0,1]
183 ; CHECK-NEXT: retq
184 %shuf = shufflevector <8 x i32> %vec, <8 x i32> undef, <8 x i32> <i32 0, i32 1, i32 0, i32 1, i32 0, i32 1, i32 0, i32 1>
185 %cmp = icmp eq <8 x i32> %mask, zeroinitializer
186 %res = select <8 x i1> %cmp, <8 x i32> %shuf, <8 x i32> zeroinitializer
187 ret <8 x i32> %res
188 }
189 define <8 x i32> @test_masked_2xi32_to_8xi32_mask3(<8 x i32> %vec, <8 x i32> %default, <8 x i32> %mask) {
190 ; CHECK-LABEL: test_masked_2xi32_to_8xi32_mask3:
191 ; CHECK: # %bb.0:
192 ; CHECK-NEXT: vptestnmd %ymm2, %ymm2, %k1
193 ; CHECK-NEXT: vbroadcasti32x2 {{.*#+}} ymm1 {%k1} = xmm0[0,1,0,1,0,1,0,1]
194 ; CHECK-NEXT: vmovdqa %ymm1, %ymm0
195 ; CHECK-NEXT: retq
196 %shuf = shufflevector <8 x i32> %vec, <8 x i32> undef, <8 x i32> <i32 0, i32 1, i32 0, i32 1, i32 0, i32 1, i32 0, i32 1>
197 %cmp = icmp eq <8 x i32> %mask, zeroinitializer
198 %res = select <8 x i1> %cmp, <8 x i32> %shuf, <8 x i32> %default
199 ret <8 x i32> %res
200 }
202 define <8 x i32> @test_masked_z_2xi32_to_8xi32_mask3(<8 x i32> %vec, <8 x i32> %mask) {
203 ; CHECK-LABEL: test_masked_z_2xi32_to_8xi32_mask3:
204 ; CHECK: # %bb.0:
205 ; CHECK-NEXT: vptestnmd %ymm1, %ymm1, %k1
206 ; CHECK-NEXT: vbroadcasti32x2 {{.*#+}} ymm0 {%k1} {z} = xmm0[0,1,0,1,0,1,0,1]
207 ; CHECK-NEXT: retq
208 %shuf = shufflevector <8 x i32> %vec, <8 x i32> undef, <8 x i32> <i32 0, i32 1, i32 0, i32 1, i32 0, i32 1, i32 0, i32 1>
209 %cmp = icmp eq <8 x i32> %mask, zeroinitializer
210 %res = select <8 x i1> %cmp, <8 x i32> %shuf, <8 x i32> zeroinitializer
211 ret <8 x i32> %res
212 }
213 define <16 x i32> @test_2xi32_to_16xi32(<16 x i32> %vec) {
214 ; CHECK-LABEL: test_2xi32_to_16xi32:
215 ; CHECK: # %bb.0:
216 ; CHECK-NEXT: vbroadcastsd %xmm0, %zmm0
217 ; CHECK-NEXT: retq
218 %res = shufflevector <16 x i32> %vec, <16 x i32> undef, <16 x i32> <i32 0, i32 1, i32 0, i32 1, i32 0, i32 1, i32 0, i32 1, i32 0, i32 1, i32 0, i32 1, i32 0, i32 1, i32 0, i32 1>
219 ret <16 x i32> %res
220 }
221 define <16 x i32> @test_masked_2xi32_to_16xi32_mask0(<16 x i32> %vec, <16 x i32> %default, <16 x i32> %mask) {
222 ; CHECK-LABEL: test_masked_2xi32_to_16xi32_mask0:
223 ; CHECK: # %bb.0:
224 ; CHECK-NEXT: vptestnmd %zmm2, %zmm2, %k1
225 ; CHECK-NEXT: vbroadcasti32x2 {{.*#+}} zmm1 {%k1} = xmm0[0,1,0,1,0,1,0,1,0,1,0,1,0,1,0,1]
226 ; CHECK-NEXT: vmovdqa64 %zmm1, %zmm0
227 ; CHECK-NEXT: retq
228 %shuf = shufflevector <16 x i32> %vec, <16 x i32> undef, <16 x i32> <i32 0, i32 1, i32 0, i32 1, i32 0, i32 1, i32 0, i32 1, i32 0, i32 1, i32 0, i32 1, i32 0, i32 1, i32 0, i32 1>
229 %cmp = icmp eq <16 x i32> %mask, zeroinitializer
230 %res = select <16 x i1> %cmp, <16 x i32> %shuf, <16 x i32> %default
231 ret <16 x i32> %res
232 }
234 define <16 x i32> @test_masked_z_2xi32_to_16xi32_mask0(<16 x i32> %vec, <16 x i32> %mask) {
235 ; CHECK-LABEL: test_masked_z_2xi32_to_16xi32_mask0:
236 ; CHECK: # %bb.0:
237 ; CHECK-NEXT: vptestnmd %zmm1, %zmm1, %k1
238 ; CHECK-NEXT: vbroadcasti32x2 {{.*#+}} zmm0 {%k1} {z} = xmm0[0,1,0,1,0,1,0,1,0,1,0,1,0,1,0,1]
239 ; CHECK-NEXT: retq
240 %shuf = shufflevector <16 x i32> %vec, <16 x i32> undef, <16 x i32> <i32 0, i32 1, i32 0, i32 1, i32 0, i32 1, i32 0, i32 1, i32 0, i32 1, i32 0, i32 1, i32 0, i32 1, i32 0, i32 1>
241 %cmp = icmp eq <16 x i32> %mask, zeroinitializer
242 %res = select <16 x i1> %cmp, <16 x i32> %shuf, <16 x i32> zeroinitializer
243 ret <16 x i32> %res
244 }
245 define <16 x i32> @test_masked_2xi32_to_16xi32_mask1(<16 x i32> %vec, <16 x i32> %default, <16 x i32> %mask) {
246 ; CHECK-LABEL: test_masked_2xi32_to_16xi32_mask1:
247 ; CHECK: # %bb.0:
248 ; CHECK-NEXT: vptestnmd %zmm2, %zmm2, %k1
249 ; CHECK-NEXT: vbroadcasti32x2 {{.*#+}} zmm1 {%k1} = xmm0[0,1,0,1,0,1,0,1,0,1,0,1,0,1,0,1]
250 ; CHECK-NEXT: vmovdqa64 %zmm1, %zmm0
251 ; CHECK-NEXT: retq
252 %shuf = shufflevector <16 x i32> %vec, <16 x i32> undef, <16 x i32> <i32 0, i32 1, i32 0, i32 1, i32 0, i32 1, i32 0, i32 1, i32 0, i32 1, i32 0, i32 1, i32 0, i32 1, i32 0, i32 1>
253 %cmp = icmp eq <16 x i32> %mask, zeroinitializer
254 %res = select <16 x i1> %cmp, <16 x i32> %shuf, <16 x i32> %default
255 ret <16 x i32> %res
256 }
258 define <16 x i32> @test_masked_z_2xi32_to_16xi32_mask1(<16 x i32> %vec, <16 x i32> %mask) {
259 ; CHECK-LABEL: test_masked_z_2xi32_to_16xi32_mask1:
260 ; CHECK: # %bb.0:
261 ; CHECK-NEXT: vptestnmd %zmm1, %zmm1, %k1
262 ; CHECK-NEXT: vbroadcasti32x2 {{.*#+}} zmm0 {%k1} {z} = xmm0[0,1,0,1,0,1,0,1,0,1,0,1,0,1,0,1]
263 ; CHECK-NEXT: retq
264 %shuf = shufflevector <16 x i32> %vec, <16 x i32> undef, <16 x i32> <i32 0, i32 1, i32 0, i32 1, i32 0, i32 1, i32 0, i32 1, i32 0, i32 1, i32 0, i32 1, i32 0, i32 1, i32 0, i32 1>
265 %cmp = icmp eq <16 x i32> %mask, zeroinitializer
266 %res = select <16 x i1> %cmp, <16 x i32> %shuf, <16 x i32> zeroinitializer
267 ret <16 x i32> %res
268 }
269 define <16 x i32> @test_masked_2xi32_to_16xi32_mask2(<16 x i32> %vec, <16 x i32> %default, <16 x i32> %mask) {
270 ; CHECK-LABEL: test_masked_2xi32_to_16xi32_mask2:
271 ; CHECK: # %bb.0:
272 ; CHECK-NEXT: vptestnmd %zmm2, %zmm2, %k1
273 ; CHECK-NEXT: vbroadcasti32x2 {{.*#+}} zmm1 {%k1} = xmm0[0,1,0,1,0,1,0,1,0,1,0,1,0,1,0,1]
274 ; CHECK-NEXT: vmovdqa64 %zmm1, %zmm0
275 ; CHECK-NEXT: retq
276 %shuf = shufflevector <16 x i32> %vec, <16 x i32> undef, <16 x i32> <i32 0, i32 1, i32 0, i32 1, i32 0, i32 1, i32 0, i32 1, i32 0, i32 1, i32 0, i32 1, i32 0, i32 1, i32 0, i32 1>
277 %cmp = icmp eq <16 x i32> %mask, zeroinitializer
278 %res = select <16 x i1> %cmp, <16 x i32> %shuf, <16 x i32> %default
279 ret <16 x i32> %res
280 }
282 define <16 x i32> @test_masked_z_2xi32_to_16xi32_mask2(<16 x i32> %vec, <16 x i32> %mask) {
283 ; CHECK-LABEL: test_masked_z_2xi32_to_16xi32_mask2:
284 ; CHECK: # %bb.0:
285 ; CHECK-NEXT: vptestnmd %zmm1, %zmm1, %k1
286 ; CHECK-NEXT: vbroadcasti32x2 {{.*#+}} zmm0 {%k1} {z} = xmm0[0,1,0,1,0,1,0,1,0,1,0,1,0,1,0,1]
287 ; CHECK-NEXT: retq
288 %shuf = shufflevector <16 x i32> %vec, <16 x i32> undef, <16 x i32> <i32 0, i32 1, i32 0, i32 1, i32 0, i32 1, i32 0, i32 1, i32 0, i32 1, i32 0, i32 1, i32 0, i32 1, i32 0, i32 1>
289 %cmp = icmp eq <16 x i32> %mask, zeroinitializer
290 %res = select <16 x i1> %cmp, <16 x i32> %shuf, <16 x i32> zeroinitializer
291 ret <16 x i32> %res
292 }
293 define <16 x i32> @test_masked_2xi32_to_16xi32_mask3(<16 x i32> %vec, <16 x i32> %default, <16 x i32> %mask) {
294 ; CHECK-LABEL: test_masked_2xi32_to_16xi32_mask3:
295 ; CHECK: # %bb.0:
296 ; CHECK-NEXT: vptestnmd %zmm2, %zmm2, %k1
297 ; CHECK-NEXT: vbroadcasti32x2 {{.*#+}} zmm1 {%k1} = xmm0[0,1,0,1,0,1,0,1,0,1,0,1,0,1,0,1]
298 ; CHECK-NEXT: vmovdqa64 %zmm1, %zmm0
299 ; CHECK-NEXT: retq
300 %shuf = shufflevector <16 x i32> %vec, <16 x i32> undef, <16 x i32> <i32 0, i32 1, i32 0, i32 1, i32 0, i32 1, i32 0, i32 1, i32 0, i32 1, i32 0, i32 1, i32 0, i32 1, i32 0, i32 1>
301 %cmp = icmp eq <16 x i32> %mask, zeroinitializer
302 %res = select <16 x i1> %cmp, <16 x i32> %shuf, <16 x i32> %default
303 ret <16 x i32> %res
304 }
306 define <16 x i32> @test_masked_z_2xi32_to_16xi32_mask3(<16 x i32> %vec, <16 x i32> %mask) {
307 ; CHECK-LABEL: test_masked_z_2xi32_to_16xi32_mask3:
308 ; CHECK: # %bb.0:
309 ; CHECK-NEXT: vptestnmd %zmm1, %zmm1, %k1
310 ; CHECK-NEXT: vbroadcasti32x2 {{.*#+}} zmm0 {%k1} {z} = xmm0[0,1,0,1,0,1,0,1,0,1,0,1,0,1,0,1]
311 ; CHECK-NEXT: retq
312 %shuf = shufflevector <16 x i32> %vec, <16 x i32> undef, <16 x i32> <i32 0, i32 1, i32 0, i32 1, i32 0, i32 1, i32 0, i32 1, i32 0, i32 1, i32 0, i32 1, i32 0, i32 1, i32 0, i32 1>
313 %cmp = icmp eq <16 x i32> %mask, zeroinitializer
314 %res = select <16 x i1> %cmp, <16 x i32> %shuf, <16 x i32> zeroinitializer
315 ret <16 x i32> %res
316 }
317 define <4 x i32> @test_2xi32_to_4xi32_mem(ptr %vp) {
318 ; CHECK-LABEL: test_2xi32_to_4xi32_mem:
319 ; CHECK: # %bb.0:
320 ; CHECK-NEXT: vmovddup {{.*#+}} xmm0 = mem[0,0]
321 ; CHECK-NEXT: retq
322 %vec = load <2 x i32>, ptr %vp
323 %res = shufflevector <2 x i32> %vec, <2 x i32> undef, <4 x i32> <i32 0, i32 1, i32 0, i32 1>
324 ret <4 x i32> %res
325 }
326 define <4 x i32> @test_masked_2xi32_to_4xi32_mem_mask0(ptr %vp, <4 x i32> %default, <4 x i32> %mask) {
327 ; CHECK-LABEL: test_masked_2xi32_to_4xi32_mem_mask0:
328 ; CHECK: # %bb.0:
329 ; CHECK-NEXT: vptestnmd %xmm1, %xmm1, %k1
330 ; CHECK-NEXT: vbroadcasti32x2 {{.*#+}} xmm0 {%k1} = mem[0,1,0,1]
331 ; CHECK-NEXT: retq
332 %vec = load <2 x i32>, ptr %vp
333 %shuf = shufflevector <2 x i32> %vec, <2 x i32> undef, <4 x i32> <i32 0, i32 1, i32 0, i32 1>
334 %cmp = icmp eq <4 x i32> %mask, zeroinitializer
335 %res = select <4 x i1> %cmp, <4 x i32> %shuf, <4 x i32> %default
336 ret <4 x i32> %res
337 }
339 define <4 x i32> @test_masked_z_2xi32_to_4xi32_mem_mask0(ptr %vp, <4 x i32> %mask) {
340 ; CHECK-LABEL: test_masked_z_2xi32_to_4xi32_mem_mask0:
341 ; CHECK: # %bb.0:
342 ; CHECK-NEXT: vptestnmd %xmm0, %xmm0, %k1
343 ; CHECK-NEXT: vbroadcasti32x2 {{.*#+}} xmm0 {%k1} {z} = mem[0,1,0,1]
344 ; CHECK-NEXT: retq
345 %vec = load <2 x i32>, ptr %vp
346 %shuf = shufflevector <2 x i32> %vec, <2 x i32> undef, <4 x i32> <i32 0, i32 1, i32 0, i32 1>
347 %cmp = icmp eq <4 x i32> %mask, zeroinitializer
348 %res = select <4 x i1> %cmp, <4 x i32> %shuf, <4 x i32> zeroinitializer
349 ret <4 x i32> %res
350 }
351 define <4 x i32> @test_masked_2xi32_to_4xi32_mem_mask1(ptr %vp, <4 x i32> %default, <4 x i32> %mask) {
352 ; CHECK-LABEL: test_masked_2xi32_to_4xi32_mem_mask1:
353 ; CHECK: # %bb.0:
354 ; CHECK-NEXT: vptestnmd %xmm1, %xmm1, %k1
355 ; CHECK-NEXT: vbroadcasti32x2 {{.*#+}} xmm0 {%k1} = mem[0,1,0,1]
356 ; CHECK-NEXT: retq
357 %vec = load <2 x i32>, ptr %vp
358 %shuf = shufflevector <2 x i32> %vec, <2 x i32> undef, <4 x i32> <i32 0, i32 1, i32 0, i32 1>
359 %cmp = icmp eq <4 x i32> %mask, zeroinitializer
360 %res = select <4 x i1> %cmp, <4 x i32> %shuf, <4 x i32> %default
361 ret <4 x i32> %res
362 }
364 define <4 x i32> @test_masked_z_2xi32_to_4xi32_mem_mask1(ptr %vp, <4 x i32> %mask) {
365 ; CHECK-LABEL: test_masked_z_2xi32_to_4xi32_mem_mask1:
366 ; CHECK: # %bb.0:
367 ; CHECK-NEXT: vptestnmd %xmm0, %xmm0, %k1
368 ; CHECK-NEXT: vbroadcasti32x2 {{.*#+}} xmm0 {%k1} {z} = mem[0,1,0,1]
369 ; CHECK-NEXT: retq
370 %vec = load <2 x i32>, ptr %vp
371 %shuf = shufflevector <2 x i32> %vec, <2 x i32> undef, <4 x i32> <i32 0, i32 1, i32 0, i32 1>
372 %cmp = icmp eq <4 x i32> %mask, zeroinitializer
373 %res = select <4 x i1> %cmp, <4 x i32> %shuf, <4 x i32> zeroinitializer
374 ret <4 x i32> %res
375 }
376 define <4 x i32> @test_masked_2xi32_to_4xi32_mem_mask2(ptr %vp, <4 x i32> %default, <4 x i32> %mask) {
377 ; CHECK-LABEL: test_masked_2xi32_to_4xi32_mem_mask2:
378 ; CHECK: # %bb.0:
379 ; CHECK-NEXT: vptestnmd %xmm1, %xmm1, %k1
380 ; CHECK-NEXT: vbroadcasti32x2 {{.*#+}} xmm0 {%k1} = mem[0,1,0,1]
381 ; CHECK-NEXT: retq
382 %vec = load <2 x i32>, ptr %vp
383 %shuf = shufflevector <2 x i32> %vec, <2 x i32> undef, <4 x i32> <i32 0, i32 1, i32 0, i32 1>
384 %cmp = icmp eq <4 x i32> %mask, zeroinitializer
385 %res = select <4 x i1> %cmp, <4 x i32> %shuf, <4 x i32> %default
386 ret <4 x i32> %res
387 }
389 define <4 x i32> @test_masked_z_2xi32_to_4xi32_mem_mask2(ptr %vp, <4 x i32> %mask) {
390 ; CHECK-LABEL: test_masked_z_2xi32_to_4xi32_mem_mask2:
391 ; CHECK: # %bb.0:
392 ; CHECK-NEXT: vptestnmd %xmm0, %xmm0, %k1
393 ; CHECK-NEXT: vbroadcasti32x2 {{.*#+}} xmm0 {%k1} {z} = mem[0,1,0,1]
394 ; CHECK-NEXT: retq
395 %vec = load <2 x i32>, ptr %vp
396 %shuf = shufflevector <2 x i32> %vec, <2 x i32> undef, <4 x i32> <i32 0, i32 1, i32 0, i32 1>
397 %cmp = icmp eq <4 x i32> %mask, zeroinitializer
398 %res = select <4 x i1> %cmp, <4 x i32> %shuf, <4 x i32> zeroinitializer
399 ret <4 x i32> %res
400 }
401 define <4 x i32> @test_masked_2xi32_to_4xi32_mem_mask3(ptr %vp, <4 x i32> %default, <4 x i32> %mask) {
402 ; CHECK-LABEL: test_masked_2xi32_to_4xi32_mem_mask3:
403 ; CHECK: # %bb.0:
404 ; CHECK-NEXT: vptestnmd %xmm1, %xmm1, %k1
405 ; CHECK-NEXT: vbroadcasti32x2 {{.*#+}} xmm0 {%k1} = mem[0,1,0,1]
406 ; CHECK-NEXT: retq
407 %vec = load <2 x i32>, ptr %vp
408 %shuf = shufflevector <2 x i32> %vec, <2 x i32> undef, <4 x i32> <i32 0, i32 1, i32 0, i32 1>
409 %cmp = icmp eq <4 x i32> %mask, zeroinitializer
410 %res = select <4 x i1> %cmp, <4 x i32> %shuf, <4 x i32> %default
411 ret <4 x i32> %res
412 }
414 define <4 x i32> @test_masked_z_2xi32_to_4xi32_mem_mask3(ptr %vp, <4 x i32> %mask) {
415 ; CHECK-LABEL: test_masked_z_2xi32_to_4xi32_mem_mask3:
416 ; CHECK: # %bb.0:
417 ; CHECK-NEXT: vptestnmd %xmm0, %xmm0, %k1
418 ; CHECK-NEXT: vbroadcasti32x2 {{.*#+}} xmm0 {%k1} {z} = mem[0,1,0,1]
419 ; CHECK-NEXT: retq
420 %vec = load <2 x i32>, ptr %vp
421 %shuf = shufflevector <2 x i32> %vec, <2 x i32> undef, <4 x i32> <i32 0, i32 1, i32 0, i32 1>
422 %cmp = icmp eq <4 x i32> %mask, zeroinitializer
423 %res = select <4 x i1> %cmp, <4 x i32> %shuf, <4 x i32> zeroinitializer
424 ret <4 x i32> %res
425 }
426 define <8 x i32> @test_2xi32_to_8xi32_mem(ptr %vp) {
427 ; CHECK-LABEL: test_2xi32_to_8xi32_mem:
428 ; CHECK: # %bb.0:
429 ; CHECK-NEXT: vbroadcastsd (%rdi), %ymm0
430 ; CHECK-NEXT: retq
431 %vec = load <2 x i32>, ptr %vp
432 %res = shufflevector <2 x i32> %vec, <2 x i32> undef, <8 x i32> <i32 0, i32 1, i32 0, i32 1, i32 0, i32 1, i32 0, i32 1>
433 ret <8 x i32> %res
434 }
435 define <8 x i32> @test_masked_2xi32_to_8xi32_mem_mask0(ptr %vp, <8 x i32> %default, <8 x i32> %mask) {
436 ; CHECK-LABEL: test_masked_2xi32_to_8xi32_mem_mask0:
437 ; CHECK: # %bb.0:
438 ; CHECK-NEXT: vptestnmd %ymm1, %ymm1, %k1
439 ; CHECK-NEXT: vbroadcasti32x2 {{.*#+}} ymm0 {%k1} = mem[0,1,0,1,0,1,0,1]
440 ; CHECK-NEXT: retq
441 %vec = load <2 x i32>, ptr %vp
442 %shuf = shufflevector <2 x i32> %vec, <2 x i32> undef, <8 x i32> <i32 0, i32 1, i32 0, i32 1, i32 0, i32 1, i32 0, i32 1>
443 %cmp = icmp eq <8 x i32> %mask, zeroinitializer
444 %res = select <8 x i1> %cmp, <8 x i32> %shuf, <8 x i32> %default
445 ret <8 x i32> %res
446 }
448 define <8 x i32> @test_masked_z_2xi32_to_8xi32_mem_mask0(ptr %vp, <8 x i32> %mask) {
449 ; CHECK-LABEL: test_masked_z_2xi32_to_8xi32_mem_mask0:
450 ; CHECK: # %bb.0:
451 ; CHECK-NEXT: vptestnmd %ymm0, %ymm0, %k1
452 ; CHECK-NEXT: vbroadcasti32x2 {{.*#+}} ymm0 {%k1} {z} = mem[0,1,0,1,0,1,0,1]
453 ; CHECK-NEXT: retq
454 %vec = load <2 x i32>, ptr %vp
455 %shuf = shufflevector <2 x i32> %vec, <2 x i32> undef, <8 x i32> <i32 0, i32 1, i32 0, i32 1, i32 0, i32 1, i32 0, i32 1>
456 %cmp = icmp eq <8 x i32> %mask, zeroinitializer
457 %res = select <8 x i1> %cmp, <8 x i32> %shuf, <8 x i32> zeroinitializer
458 ret <8 x i32> %res
459 }
460 define <8 x i32> @test_masked_2xi32_to_8xi32_mem_mask1(ptr %vp, <8 x i32> %default, <8 x i32> %mask) {
461 ; CHECK-LABEL: test_masked_2xi32_to_8xi32_mem_mask1:
462 ; CHECK: # %bb.0:
463 ; CHECK-NEXT: vptestnmd %ymm1, %ymm1, %k1
464 ; CHECK-NEXT: vbroadcasti32x2 {{.*#+}} ymm0 {%k1} = mem[0,1,0,1,0,1,0,1]
465 ; CHECK-NEXT: retq
466 %vec = load <2 x i32>, ptr %vp
467 %shuf = shufflevector <2 x i32> %vec, <2 x i32> undef, <8 x i32> <i32 0, i32 1, i32 0, i32 1, i32 0, i32 1, i32 0, i32 1>
468 %cmp = icmp eq <8 x i32> %mask, zeroinitializer
469 %res = select <8 x i1> %cmp, <8 x i32> %shuf, <8 x i32> %default
470 ret <8 x i32> %res
471 }
473 define <8 x i32> @test_masked_z_2xi32_to_8xi32_mem_mask1(ptr %vp, <8 x i32> %mask) {
474 ; CHECK-LABEL: test_masked_z_2xi32_to_8xi32_mem_mask1:
475 ; CHECK: # %bb.0:
476 ; CHECK-NEXT: vptestnmd %ymm0, %ymm0, %k1
477 ; CHECK-NEXT: vbroadcasti32x2 {{.*#+}} ymm0 {%k1} {z} = mem[0,1,0,1,0,1,0,1]
478 ; CHECK-NEXT: retq
479 %vec = load <2 x i32>, ptr %vp
480 %shuf = shufflevector <2 x i32> %vec, <2 x i32> undef, <8 x i32> <i32 0, i32 1, i32 0, i32 1, i32 0, i32 1, i32 0, i32 1>
481 %cmp = icmp eq <8 x i32> %mask, zeroinitializer
482 %res = select <8 x i1> %cmp, <8 x i32> %shuf, <8 x i32> zeroinitializer
483 ret <8 x i32> %res
484 }
485 define <8 x i32> @test_masked_2xi32_to_8xi32_mem_mask2(ptr %vp, <8 x i32> %default, <8 x i32> %mask) {
486 ; CHECK-LABEL: test_masked_2xi32_to_8xi32_mem_mask2:
487 ; CHECK: # %bb.0:
488 ; CHECK-NEXT: vptestnmd %ymm1, %ymm1, %k1
489 ; CHECK-NEXT: vbroadcasti32x2 {{.*#+}} ymm0 {%k1} = mem[0,1,0,1,0,1,0,1]
490 ; CHECK-NEXT: retq
491 %vec = load <2 x i32>, ptr %vp
492 %shuf = shufflevector <2 x i32> %vec, <2 x i32> undef, <8 x i32> <i32 0, i32 1, i32 0, i32 1, i32 0, i32 1, i32 0, i32 1>
493 %cmp = icmp eq <8 x i32> %mask, zeroinitializer
494 %res = select <8 x i1> %cmp, <8 x i32> %shuf, <8 x i32> %default
495 ret <8 x i32> %res
496 }
498 define <8 x i32> @test_masked_z_2xi32_to_8xi32_mem_mask2(ptr %vp, <8 x i32> %mask) {
499 ; CHECK-LABEL: test_masked_z_2xi32_to_8xi32_mem_mask2:
500 ; CHECK: # %bb.0:
501 ; CHECK-NEXT: vptestnmd %ymm0, %ymm0, %k1
502 ; CHECK-NEXT: vbroadcasti32x2 {{.*#+}} ymm0 {%k1} {z} = mem[0,1,0,1,0,1,0,1]
503 ; CHECK-NEXT: retq
504 %vec = load <2 x i32>, ptr %vp
505 %shuf = shufflevector <2 x i32> %vec, <2 x i32> undef, <8 x i32> <i32 0, i32 1, i32 0, i32 1, i32 0, i32 1, i32 0, i32 1>
506 %cmp = icmp eq <8 x i32> %mask, zeroinitializer
507 %res = select <8 x i1> %cmp, <8 x i32> %shuf, <8 x i32> zeroinitializer
508 ret <8 x i32> %res
509 }
510 define <8 x i32> @test_masked_2xi32_to_8xi32_mem_mask3(ptr %vp, <8 x i32> %default, <8 x i32> %mask) {
511 ; CHECK-LABEL: test_masked_2xi32_to_8xi32_mem_mask3:
512 ; CHECK: # %bb.0:
513 ; CHECK-NEXT: vptestnmd %ymm1, %ymm1, %k1
514 ; CHECK-NEXT: vbroadcasti32x2 {{.*#+}} ymm0 {%k1} = mem[0,1,0,1,0,1,0,1]
515 ; CHECK-NEXT: retq
516 %vec = load <2 x i32>, ptr %vp
517 %shuf = shufflevector <2 x i32> %vec, <2 x i32> undef, <8 x i32> <i32 0, i32 1, i32 0, i32 1, i32 0, i32 1, i32 0, i32 1>
518 %cmp = icmp eq <8 x i32> %mask, zeroinitializer
519 %res = select <8 x i1> %cmp, <8 x i32> %shuf, <8 x i32> %default
520 ret <8 x i32> %res
521 }
523 define <8 x i32> @test_masked_z_2xi32_to_8xi32_mem_mask3(ptr %vp, <8 x i32> %mask) {
524 ; CHECK-LABEL: test_masked_z_2xi32_to_8xi32_mem_mask3:
525 ; CHECK: # %bb.0:
526 ; CHECK-NEXT: vptestnmd %ymm0, %ymm0, %k1
527 ; CHECK-NEXT: vbroadcasti32x2 {{.*#+}} ymm0 {%k1} {z} = mem[0,1,0,1,0,1,0,1]
528 ; CHECK-NEXT: retq
529 %vec = load <2 x i32>, ptr %vp
530 %shuf = shufflevector <2 x i32> %vec, <2 x i32> undef, <8 x i32> <i32 0, i32 1, i32 0, i32 1, i32 0, i32 1, i32 0, i32 1>
531 %cmp = icmp eq <8 x i32> %mask, zeroinitializer
532 %res = select <8 x i1> %cmp, <8 x i32> %shuf, <8 x i32> zeroinitializer
533 ret <8 x i32> %res
534 }
535 define <16 x i32> @test_2xi32_to_16xi32_mem(ptr %vp) {
536 ; CHECK-LABEL: test_2xi32_to_16xi32_mem:
537 ; CHECK: # %bb.0:
538 ; CHECK-NEXT: vbroadcastsd (%rdi), %zmm0
539 ; CHECK-NEXT: retq
540 %vec = load <2 x i32>, ptr %vp
541 %res = shufflevector <2 x i32> %vec, <2 x i32> undef, <16 x i32> <i32 0, i32 1, i32 0, i32 1, i32 0, i32 1, i32 0, i32 1, i32 0, i32 1, i32 0, i32 1, i32 0, i32 1, i32 0, i32 1>
542 ret <16 x i32> %res
543 }
544 define <16 x i32> @test_masked_2xi32_to_16xi32_mem_mask0(ptr %vp, <16 x i32> %default, <16 x i32> %mask) {
545 ; CHECK-LABEL: test_masked_2xi32_to_16xi32_mem_mask0:
546 ; CHECK: # %bb.0:
547 ; CHECK-NEXT: vptestnmd %zmm1, %zmm1, %k1
548 ; CHECK-NEXT: vbroadcasti32x2 {{.*#+}} zmm0 {%k1} = mem[0,1,0,1,0,1,0,1,0,1,0,1,0,1,0,1]
549 ; CHECK-NEXT: retq
550 %vec = load <2 x i32>, ptr %vp
551 %shuf = shufflevector <2 x i32> %vec, <2 x i32> undef, <16 x i32> <i32 0, i32 1, i32 0, i32 1, i32 0, i32 1, i32 0, i32 1, i32 0, i32 1, i32 0, i32 1, i32 0, i32 1, i32 0, i32 1>
552 %cmp = icmp eq <16 x i32> %mask, zeroinitializer
553 %res = select <16 x i1> %cmp, <16 x i32> %shuf, <16 x i32> %default
554 ret <16 x i32> %res
555 }
557 define <16 x i32> @test_masked_z_2xi32_to_16xi32_mem_mask0(ptr %vp, <16 x i32> %mask) {
558 ; CHECK-LABEL: test_masked_z_2xi32_to_16xi32_mem_mask0:
559 ; CHECK: # %bb.0:
560 ; CHECK-NEXT: vptestnmd %zmm0, %zmm0, %k1
561 ; CHECK-NEXT: vbroadcasti32x2 {{.*#+}} zmm0 {%k1} {z} = mem[0,1,0,1,0,1,0,1,0,1,0,1,0,1,0,1]
562 ; CHECK-NEXT: retq
563 %vec = load <2 x i32>, ptr %vp
564 %shuf = shufflevector <2 x i32> %vec, <2 x i32> undef, <16 x i32> <i32 0, i32 1, i32 0, i32 1, i32 0, i32 1, i32 0, i32 1, i32 0, i32 1, i32 0, i32 1, i32 0, i32 1, i32 0, i32 1>
565 %cmp = icmp eq <16 x i32> %mask, zeroinitializer
566 %res = select <16 x i1> %cmp, <16 x i32> %shuf, <16 x i32> zeroinitializer
567 ret <16 x i32> %res
568 }
569 define <16 x i32> @test_masked_2xi32_to_16xi32_mem_mask1(ptr %vp, <16 x i32> %default, <16 x i32> %mask) {
570 ; CHECK-LABEL: test_masked_2xi32_to_16xi32_mem_mask1:
571 ; CHECK: # %bb.0:
572 ; CHECK-NEXT: vptestnmd %zmm1, %zmm1, %k1
573 ; CHECK-NEXT: vbroadcasti32x2 {{.*#+}} zmm0 {%k1} = mem[0,1,0,1,0,1,0,1,0,1,0,1,0,1,0,1]
574 ; CHECK-NEXT: retq
575 %vec = load <2 x i32>, ptr %vp
576 %shuf = shufflevector <2 x i32> %vec, <2 x i32> undef, <16 x i32> <i32 0, i32 1, i32 0, i32 1, i32 0, i32 1, i32 0, i32 1, i32 0, i32 1, i32 0, i32 1, i32 0, i32 1, i32 0, i32 1>
577 %cmp = icmp eq <16 x i32> %mask, zeroinitializer
578 %res = select <16 x i1> %cmp, <16 x i32> %shuf, <16 x i32> %default
579 ret <16 x i32> %res
580 }
582 define <16 x i32> @test_masked_z_2xi32_to_16xi32_mem_mask1(ptr %vp, <16 x i32> %mask) {
583 ; CHECK-LABEL: test_masked_z_2xi32_to_16xi32_mem_mask1:
584 ; CHECK: # %bb.0:
585 ; CHECK-NEXT: vptestnmd %zmm0, %zmm0, %k1
586 ; CHECK-NEXT: vbroadcasti32x2 {{.*#+}} zmm0 {%k1} {z} = mem[0,1,0,1,0,1,0,1,0,1,0,1,0,1,0,1]
587 ; CHECK-NEXT: retq
588 %vec = load <2 x i32>, ptr %vp
589 %shuf = shufflevector <2 x i32> %vec, <2 x i32> undef, <16 x i32> <i32 0, i32 1, i32 0, i32 1, i32 0, i32 1, i32 0, i32 1, i32 0, i32 1, i32 0, i32 1, i32 0, i32 1, i32 0, i32 1>
590 %cmp = icmp eq <16 x i32> %mask, zeroinitializer
591 %res = select <16 x i1> %cmp, <16 x i32> %shuf, <16 x i32> zeroinitializer
592 ret <16 x i32> %res
593 }
594 define <16 x i32> @test_masked_2xi32_to_16xi32_mem_mask2(ptr %vp, <16 x i32> %default, <16 x i32> %mask) {
595 ; CHECK-LABEL: test_masked_2xi32_to_16xi32_mem_mask2:
596 ; CHECK: # %bb.0:
597 ; CHECK-NEXT: vptestnmd %zmm1, %zmm1, %k1
598 ; CHECK-NEXT: vbroadcasti32x2 {{.*#+}} zmm0 {%k1} = mem[0,1,0,1,0,1,0,1,0,1,0,1,0,1,0,1]
599 ; CHECK-NEXT: retq
600 %vec = load <2 x i32>, ptr %vp
601 %shuf = shufflevector <2 x i32> %vec, <2 x i32> undef, <16 x i32> <i32 0, i32 1, i32 0, i32 1, i32 0, i32 1, i32 0, i32 1, i32 0, i32 1, i32 0, i32 1, i32 0, i32 1, i32 0, i32 1>
602 %cmp = icmp eq <16 x i32> %mask, zeroinitializer
603 %res = select <16 x i1> %cmp, <16 x i32> %shuf, <16 x i32> %default
604 ret <16 x i32> %res
605 }
607 define <16 x i32> @test_masked_z_2xi32_to_16xi32_mem_mask2(ptr %vp, <16 x i32> %mask) {
608 ; CHECK-LABEL: test_masked_z_2xi32_to_16xi32_mem_mask2:
609 ; CHECK: # %bb.0:
610 ; CHECK-NEXT: vptestnmd %zmm0, %zmm0, %k1
611 ; CHECK-NEXT: vbroadcasti32x2 {{.*#+}} zmm0 {%k1} {z} = mem[0,1,0,1,0,1,0,1,0,1,0,1,0,1,0,1]
612 ; CHECK-NEXT: retq
613 %vec = load <2 x i32>, ptr %vp
614 %shuf = shufflevector <2 x i32> %vec, <2 x i32> undef, <16 x i32> <i32 0, i32 1, i32 0, i32 1, i32 0, i32 1, i32 0, i32 1, i32 0, i32 1, i32 0, i32 1, i32 0, i32 1, i32 0, i32 1>
615 %cmp = icmp eq <16 x i32> %mask, zeroinitializer
616 %res = select <16 x i1> %cmp, <16 x i32> %shuf, <16 x i32> zeroinitializer
617 ret <16 x i32> %res
618 }
619 define <16 x i32> @test_masked_2xi32_to_16xi32_mem_mask3(ptr %vp, <16 x i32> %default, <16 x i32> %mask) {
620 ; CHECK-LABEL: test_masked_2xi32_to_16xi32_mem_mask3:
621 ; CHECK: # %bb.0:
622 ; CHECK-NEXT: vptestnmd %zmm1, %zmm1, %k1
623 ; CHECK-NEXT: vbroadcasti32x2 {{.*#+}} zmm0 {%k1} = mem[0,1,0,1,0,1,0,1,0,1,0,1,0,1,0,1]
624 ; CHECK-NEXT: retq
625 %vec = load <2 x i32>, ptr %vp
626 %shuf = shufflevector <2 x i32> %vec, <2 x i32> undef, <16 x i32> <i32 0, i32 1, i32 0, i32 1, i32 0, i32 1, i32 0, i32 1, i32 0, i32 1, i32 0, i32 1, i32 0, i32 1, i32 0, i32 1>
627 %cmp = icmp eq <16 x i32> %mask, zeroinitializer
628 %res = select <16 x i1> %cmp, <16 x i32> %shuf, <16 x i32> %default
629 ret <16 x i32> %res
630 }
632 define <16 x i32> @test_masked_z_2xi32_to_16xi32_mem_mask3(ptr %vp, <16 x i32> %mask) {
633 ; CHECK-LABEL: test_masked_z_2xi32_to_16xi32_mem_mask3:
634 ; CHECK: # %bb.0:
635 ; CHECK-NEXT: vptestnmd %zmm0, %zmm0, %k1
636 ; CHECK-NEXT: vbroadcasti32x2 {{.*#+}} zmm0 {%k1} {z} = mem[0,1,0,1,0,1,0,1,0,1,0,1,0,1,0,1]
637 ; CHECK-NEXT: retq
638 %vec = load <2 x i32>, ptr %vp
639 %shuf = shufflevector <2 x i32> %vec, <2 x i32> undef, <16 x i32> <i32 0, i32 1, i32 0, i32 1, i32 0, i32 1, i32 0, i32 1, i32 0, i32 1, i32 0, i32 1, i32 0, i32 1, i32 0, i32 1>
640 %cmp = icmp eq <16 x i32> %mask, zeroinitializer
641 %res = select <16 x i1> %cmp, <16 x i32> %shuf, <16 x i32> zeroinitializer
642 ret <16 x i32> %res
643 }
644 define <8 x i32> @test_4xi32_to_8xi32_mem(ptr %vp) {
645 ; CHECK-LABEL: test_4xi32_to_8xi32_mem:
646 ; CHECK: # %bb.0:
647 ; CHECK-NEXT: vbroadcasti128 {{.*#+}} ymm0 = mem[0,1,0,1]
648 ; CHECK-NEXT: retq
649 %vec = load <4 x i32>, ptr %vp
650 %res = shufflevector <4 x i32> %vec, <4 x i32> undef, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 0, i32 1, i32 2, i32 3>
651 ret <8 x i32> %res
652 }
653 define <8 x i32> @test_masked_4xi32_to_8xi32_mem_mask0(ptr %vp, <8 x i32> %default, <8 x i32> %mask) {
654 ; CHECK-LABEL: test_masked_4xi32_to_8xi32_mem_mask0:
655 ; CHECK: # %bb.0:
656 ; CHECK-NEXT: vptestnmd %ymm1, %ymm1, %k1
657 ; CHECK-NEXT: vbroadcasti32x4 {{.*#+}} ymm0 {%k1} = mem[0,1,2,3,0,1,2,3]
658 ; CHECK-NEXT: retq
659 %vec = load <4 x i32>, ptr %vp
660 %shuf = shufflevector <4 x i32> %vec, <4 x i32> undef, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 0, i32 1, i32 2, i32 3>
661 %cmp = icmp eq <8 x i32> %mask, zeroinitializer
662 %res = select <8 x i1> %cmp, <8 x i32> %shuf, <8 x i32> %default
663 ret <8 x i32> %res
664 }
666 define <8 x i32> @test_masked_z_4xi32_to_8xi32_mem_mask0(ptr %vp, <8 x i32> %mask) {
667 ; CHECK-LABEL: test_masked_z_4xi32_to_8xi32_mem_mask0:
668 ; CHECK: # %bb.0:
669 ; CHECK-NEXT: vptestnmd %ymm0, %ymm0, %k1
670 ; CHECK-NEXT: vbroadcasti32x4 {{.*#+}} ymm0 {%k1} {z} = mem[0,1,2,3,0,1,2,3]
671 ; CHECK-NEXT: retq
672 %vec = load <4 x i32>, ptr %vp
673 %shuf = shufflevector <4 x i32> %vec, <4 x i32> undef, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 0, i32 1, i32 2, i32 3>
674 %cmp = icmp eq <8 x i32> %mask, zeroinitializer
675 %res = select <8 x i1> %cmp, <8 x i32> %shuf, <8 x i32> zeroinitializer
676 ret <8 x i32> %res
677 }
678 define <8 x i32> @test_masked_4xi32_to_8xi32_mem_mask1(ptr %vp, <8 x i32> %default, <8 x i32> %mask) {
679 ; CHECK-LABEL: test_masked_4xi32_to_8xi32_mem_mask1:
680 ; CHECK: # %bb.0:
681 ; CHECK-NEXT: vptestnmd %ymm1, %ymm1, %k1
682 ; CHECK-NEXT: vbroadcasti32x4 {{.*#+}} ymm0 {%k1} = mem[0,1,2,3,0,1,2,3]
683 ; CHECK-NEXT: retq
684 %vec = load <4 x i32>, ptr %vp
685 %shuf = shufflevector <4 x i32> %vec, <4 x i32> undef, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 0, i32 1, i32 2, i32 3>
686 %cmp = icmp eq <8 x i32> %mask, zeroinitializer
687 %res = select <8 x i1> %cmp, <8 x i32> %shuf, <8 x i32> %default
688 ret <8 x i32> %res
689 }
691 define <8 x i32> @test_masked_z_4xi32_to_8xi32_mem_mask1(ptr %vp, <8 x i32> %mask) {
692 ; CHECK-LABEL: test_masked_z_4xi32_to_8xi32_mem_mask1:
693 ; CHECK: # %bb.0:
694 ; CHECK-NEXT: vptestnmd %ymm0, %ymm0, %k1
695 ; CHECK-NEXT: vbroadcasti32x4 {{.*#+}} ymm0 {%k1} {z} = mem[0,1,2,3,0,1,2,3]
696 ; CHECK-NEXT: retq
697 %vec = load <4 x i32>, ptr %vp
698 %shuf = shufflevector <4 x i32> %vec, <4 x i32> undef, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 0, i32 1, i32 2, i32 3>
699 %cmp = icmp eq <8 x i32> %mask, zeroinitializer
700 %res = select <8 x i1> %cmp, <8 x i32> %shuf, <8 x i32> zeroinitializer
701 ret <8 x i32> %res
702 }
703 define <8 x i32> @test_masked_4xi32_to_8xi32_mem_mask2(ptr %vp, <8 x i32> %default, <8 x i32> %mask) {
704 ; CHECK-LABEL: test_masked_4xi32_to_8xi32_mem_mask2:
705 ; CHECK: # %bb.0:
706 ; CHECK-NEXT: vptestnmd %ymm1, %ymm1, %k1
707 ; CHECK-NEXT: vbroadcasti32x4 {{.*#+}} ymm0 {%k1} = mem[0,1,2,3,0,1,2,3]
708 ; CHECK-NEXT: retq
709 %vec = load <4 x i32>, ptr %vp
710 %shuf = shufflevector <4 x i32> %vec, <4 x i32> undef, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 0, i32 1, i32 2, i32 3>
711 %cmp = icmp eq <8 x i32> %mask, zeroinitializer
712 %res = select <8 x i1> %cmp, <8 x i32> %shuf, <8 x i32> %default
713 ret <8 x i32> %res
714 }
716 define <8 x i32> @test_masked_z_4xi32_to_8xi32_mem_mask2(ptr %vp, <8 x i32> %mask) {
717 ; CHECK-LABEL: test_masked_z_4xi32_to_8xi32_mem_mask2:
718 ; CHECK: # %bb.0:
719 ; CHECK-NEXT: vptestnmd %ymm0, %ymm0, %k1
720 ; CHECK-NEXT: vbroadcasti32x4 {{.*#+}} ymm0 {%k1} {z} = mem[0,1,2,3,0,1,2,3]
721 ; CHECK-NEXT: retq
722 %vec = load <4 x i32>, ptr %vp
723 %shuf = shufflevector <4 x i32> %vec, <4 x i32> undef, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 0, i32 1, i32 2, i32 3>
724 %cmp = icmp eq <8 x i32> %mask, zeroinitializer
725 %res = select <8 x i1> %cmp, <8 x i32> %shuf, <8 x i32> zeroinitializer
726 ret <8 x i32> %res
727 }
728 define <8 x i32> @test_masked_4xi32_to_8xi32_mem_mask3(ptr %vp, <8 x i32> %default, <8 x i32> %mask) {
729 ; CHECK-LABEL: test_masked_4xi32_to_8xi32_mem_mask3:
730 ; CHECK: # %bb.0:
731 ; CHECK-NEXT: vptestnmd %ymm1, %ymm1, %k1
732 ; CHECK-NEXT: vbroadcasti32x4 {{.*#+}} ymm0 {%k1} = mem[0,1,2,3,0,1,2,3]
733 ; CHECK-NEXT: retq
734 %vec = load <4 x i32>, ptr %vp
735 %shuf = shufflevector <4 x i32> %vec, <4 x i32> undef, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 0, i32 1, i32 2, i32 3>
736 %cmp = icmp eq <8 x i32> %mask, zeroinitializer
737 %res = select <8 x i1> %cmp, <8 x i32> %shuf, <8 x i32> %default
738 ret <8 x i32> %res
739 }
741 define <8 x i32> @test_masked_z_4xi32_to_8xi32_mem_mask3(ptr %vp, <8 x i32> %mask) {
742 ; CHECK-LABEL: test_masked_z_4xi32_to_8xi32_mem_mask3:
743 ; CHECK: # %bb.0:
744 ; CHECK-NEXT: vptestnmd %ymm0, %ymm0, %k1
745 ; CHECK-NEXT: vbroadcasti32x4 {{.*#+}} ymm0 {%k1} {z} = mem[0,1,2,3,0,1,2,3]
746 ; CHECK-NEXT: retq
747 %vec = load <4 x i32>, ptr %vp
748 %shuf = shufflevector <4 x i32> %vec, <4 x i32> undef, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 0, i32 1, i32 2, i32 3>
749 %cmp = icmp eq <8 x i32> %mask, zeroinitializer
750 %res = select <8 x i1> %cmp, <8 x i32> %shuf, <8 x i32> zeroinitializer
751 ret <8 x i32> %res
752 }
753 define <16 x i32> @test_4xi32_to_16xi32_mem(ptr %vp) {
754 ; CHECK-LABEL: test_4xi32_to_16xi32_mem:
755 ; CHECK: # %bb.0:
756 ; CHECK-NEXT: vbroadcasti32x4 {{.*#+}} zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
757 ; CHECK-NEXT: retq
758 %vec = load <4 x i32>, ptr %vp
759 %res = shufflevector <4 x i32> %vec, <4 x i32> undef, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 0, i32 1, i32 2, i32 3, i32 0, i32 1, i32 2, i32 3, i32 0, i32 1, i32 2, i32 3>
760 ret <16 x i32> %res
761 }
762 define <16 x i32> @test_masked_4xi32_to_16xi32_mem_mask0(ptr %vp, <16 x i32> %default, <16 x i32> %mask) {
763 ; CHECK-LABEL: test_masked_4xi32_to_16xi32_mem_mask0:
764 ; CHECK: # %bb.0:
765 ; CHECK-NEXT: vptestnmd %zmm1, %zmm1, %k1
766 ; CHECK-NEXT: vbroadcasti32x4 {{.*#+}} zmm0 {%k1} = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
767 ; CHECK-NEXT: retq
768 %vec = load <4 x i32>, ptr %vp
769 %shuf = shufflevector <4 x i32> %vec, <4 x i32> undef, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 0, i32 1, i32 2, i32 3, i32 0, i32 1, i32 2, i32 3, i32 0, i32 1, i32 2, i32 3>
770 %cmp = icmp eq <16 x i32> %mask, zeroinitializer
771 %res = select <16 x i1> %cmp, <16 x i32> %shuf, <16 x i32> %default
772 ret <16 x i32> %res
773 }
775 define <16 x i32> @test_masked_z_4xi32_to_16xi32_mem_mask0(ptr %vp, <16 x i32> %mask) {
776 ; CHECK-LABEL: test_masked_z_4xi32_to_16xi32_mem_mask0:
777 ; CHECK: # %bb.0:
778 ; CHECK-NEXT: vptestnmd %zmm0, %zmm0, %k1
779 ; CHECK-NEXT: vbroadcasti32x4 {{.*#+}} zmm0 {%k1} {z} = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
780 ; CHECK-NEXT: retq
781 %vec = load <4 x i32>, ptr %vp
782 %shuf = shufflevector <4 x i32> %vec, <4 x i32> undef, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 0, i32 1, i32 2, i32 3, i32 0, i32 1, i32 2, i32 3, i32 0, i32 1, i32 2, i32 3>
783 %cmp = icmp eq <16 x i32> %mask, zeroinitializer
784 %res = select <16 x i1> %cmp, <16 x i32> %shuf, <16 x i32> zeroinitializer
785 ret <16 x i32> %res
786 }
787 define <16 x i32> @test_masked_4xi32_to_16xi32_mem_mask1(ptr %vp, <16 x i32> %default, <16 x i32> %mask) {
788 ; CHECK-LABEL: test_masked_4xi32_to_16xi32_mem_mask1:
789 ; CHECK: # %bb.0:
790 ; CHECK-NEXT: vptestnmd %zmm1, %zmm1, %k1
791 ; CHECK-NEXT: vbroadcasti32x4 {{.*#+}} zmm0 {%k1} = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
792 ; CHECK-NEXT: retq
793 %vec = load <4 x i32>, ptr %vp
794 %shuf = shufflevector <4 x i32> %vec, <4 x i32> undef, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 0, i32 1, i32 2, i32 3, i32 0, i32 1, i32 2, i32 3, i32 0, i32 1, i32 2, i32 3>
795 %cmp = icmp eq <16 x i32> %mask, zeroinitializer
796 %res = select <16 x i1> %cmp, <16 x i32> %shuf, <16 x i32> %default
797 ret <16 x i32> %res
798 }
800 define <16 x i32> @test_masked_z_4xi32_to_16xi32_mem_mask1(ptr %vp, <16 x i32> %mask) {
801 ; CHECK-LABEL: test_masked_z_4xi32_to_16xi32_mem_mask1:
802 ; CHECK: # %bb.0:
803 ; CHECK-NEXT: vptestnmd %zmm0, %zmm0, %k1
804 ; CHECK-NEXT: vbroadcasti32x4 {{.*#+}} zmm0 {%k1} {z} = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
805 ; CHECK-NEXT: retq
806 %vec = load <4 x i32>, ptr %vp
807 %shuf = shufflevector <4 x i32> %vec, <4 x i32> undef, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 0, i32 1, i32 2, i32 3, i32 0, i32 1, i32 2, i32 3, i32 0, i32 1, i32 2, i32 3>
808 %cmp = icmp eq <16 x i32> %mask, zeroinitializer
809 %res = select <16 x i1> %cmp, <16 x i32> %shuf, <16 x i32> zeroinitializer
810 ret <16 x i32> %res
811 }
812 define <16 x i32> @test_masked_4xi32_to_16xi32_mem_mask2(ptr %vp, <16 x i32> %default, <16 x i32> %mask) {
813 ; CHECK-LABEL: test_masked_4xi32_to_16xi32_mem_mask2:
814 ; CHECK: # %bb.0:
815 ; CHECK-NEXT: vptestnmd %zmm1, %zmm1, %k1
816 ; CHECK-NEXT: vbroadcasti32x4 {{.*#+}} zmm0 {%k1} = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
817 ; CHECK-NEXT: retq
818 %vec = load <4 x i32>, ptr %vp
819 %shuf = shufflevector <4 x i32> %vec, <4 x i32> undef, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 0, i32 1, i32 2, i32 3, i32 0, i32 1, i32 2, i32 3, i32 0, i32 1, i32 2, i32 3>
820 %cmp = icmp eq <16 x i32> %mask, zeroinitializer
821 %res = select <16 x i1> %cmp, <16 x i32> %shuf, <16 x i32> %default
822 ret <16 x i32> %res
823 }
825 define <16 x i32> @test_masked_z_4xi32_to_16xi32_mem_mask2(ptr %vp, <16 x i32> %mask) {
826 ; CHECK-LABEL: test_masked_z_4xi32_to_16xi32_mem_mask2:
827 ; CHECK: # %bb.0:
828 ; CHECK-NEXT: vptestnmd %zmm0, %zmm0, %k1
829 ; CHECK-NEXT: vbroadcasti32x4 {{.*#+}} zmm0 {%k1} {z} = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
830 ; CHECK-NEXT: retq
831 %vec = load <4 x i32>, ptr %vp
832 %shuf = shufflevector <4 x i32> %vec, <4 x i32> undef, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 0, i32 1, i32 2, i32 3, i32 0, i32 1, i32 2, i32 3, i32 0, i32 1, i32 2, i32 3>
833 %cmp = icmp eq <16 x i32> %mask, zeroinitializer
834 %res = select <16 x i1> %cmp, <16 x i32> %shuf, <16 x i32> zeroinitializer
835 ret <16 x i32> %res
836 }
837 define <16 x i32> @test_masked_4xi32_to_16xi32_mem_mask3(ptr %vp, <16 x i32> %default, <16 x i32> %mask) {
838 ; CHECK-LABEL: test_masked_4xi32_to_16xi32_mem_mask3:
839 ; CHECK: # %bb.0:
840 ; CHECK-NEXT: vptestnmd %zmm1, %zmm1, %k1
841 ; CHECK-NEXT: vbroadcasti32x4 {{.*#+}} zmm0 {%k1} = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
842 ; CHECK-NEXT: retq
843 %vec = load <4 x i32>, ptr %vp
844 %shuf = shufflevector <4 x i32> %vec, <4 x i32> undef, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 0, i32 1, i32 2, i32 3, i32 0, i32 1, i32 2, i32 3, i32 0, i32 1, i32 2, i32 3>
845 %cmp = icmp eq <16 x i32> %mask, zeroinitializer
846 %res = select <16 x i1> %cmp, <16 x i32> %shuf, <16 x i32> %default
847 ret <16 x i32> %res
848 }
850 define <16 x i32> @test_masked_z_4xi32_to_16xi32_mem_mask3(ptr %vp, <16 x i32> %mask) {
851 ; CHECK-LABEL: test_masked_z_4xi32_to_16xi32_mem_mask3:
852 ; CHECK: # %bb.0:
853 ; CHECK-NEXT: vptestnmd %zmm0, %zmm0, %k1
854 ; CHECK-NEXT: vbroadcasti32x4 {{.*#+}} zmm0 {%k1} {z} = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
855 ; CHECK-NEXT: retq
856 %vec = load <4 x i32>, ptr %vp
857 %shuf = shufflevector <4 x i32> %vec, <4 x i32> undef, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 0, i32 1, i32 2, i32 3, i32 0, i32 1, i32 2, i32 3, i32 0, i32 1, i32 2, i32 3>
858 %cmp = icmp eq <16 x i32> %mask, zeroinitializer
859 %res = select <16 x i1> %cmp, <16 x i32> %shuf, <16 x i32> zeroinitializer
860 ret <16 x i32> %res
861 }
862 define <4 x i64> @test_2xi64_to_4xi64_mem(ptr %vp) {
863 ; CHECK-LABEL: test_2xi64_to_4xi64_mem:
864 ; CHECK: # %bb.0:
865 ; CHECK-NEXT: vbroadcasti128 {{.*#+}} ymm0 = mem[0,1,0,1]
866 ; CHECK-NEXT: retq
867 %vec = load <2 x i64>, ptr %vp
868 %res = shufflevector <2 x i64> %vec, <2 x i64> undef, <4 x i32> <i32 0, i32 1, i32 0, i32 1>
869 ret <4 x i64> %res
870 }
871 define <4 x i64> @test_masked_2xi64_to_4xi64_mem_mask0(ptr %vp, <4 x i64> %default, <4 x i64> %mask) {
872 ; CHECK-LABEL: test_masked_2xi64_to_4xi64_mem_mask0:
873 ; CHECK: # %bb.0:
874 ; CHECK-NEXT: vptestnmq %ymm1, %ymm1, %k1
875 ; CHECK-NEXT: vbroadcasti64x2 {{.*#+}} ymm0 {%k1} = mem[0,1,0,1]
876 ; CHECK-NEXT: retq
877 %vec = load <2 x i64>, ptr %vp
878 %shuf = shufflevector <2 x i64> %vec, <2 x i64> undef, <4 x i32> <i32 0, i32 1, i32 0, i32 1>
879 %cmp = icmp eq <4 x i64> %mask, zeroinitializer
880 %res = select <4 x i1> %cmp, <4 x i64> %shuf, <4 x i64> %default
881 ret <4 x i64> %res
882 }
884 define <4 x i64> @test_masked_z_2xi64_to_4xi64_mem_mask0(ptr %vp, <4 x i64> %mask) {
885 ; CHECK-LABEL: test_masked_z_2xi64_to_4xi64_mem_mask0:
886 ; CHECK: # %bb.0:
887 ; CHECK-NEXT: vptestnmq %ymm0, %ymm0, %k1
888 ; CHECK-NEXT: vbroadcasti64x2 {{.*#+}} ymm0 {%k1} {z} = mem[0,1,0,1]
889 ; CHECK-NEXT: retq
890 %vec = load <2 x i64>, ptr %vp
891 %shuf = shufflevector <2 x i64> %vec, <2 x i64> undef, <4 x i32> <i32 0, i32 1, i32 0, i32 1>
892 %cmp = icmp eq <4 x i64> %mask, zeroinitializer
893 %res = select <4 x i1> %cmp, <4 x i64> %shuf, <4 x i64> zeroinitializer
894 ret <4 x i64> %res
895 }
896 define <4 x i64> @test_masked_2xi64_to_4xi64_mem_mask1(ptr %vp, <4 x i64> %default, <4 x i64> %mask) {
897 ; CHECK-LABEL: test_masked_2xi64_to_4xi64_mem_mask1:
898 ; CHECK: # %bb.0:
899 ; CHECK-NEXT: vptestnmq %ymm1, %ymm1, %k1
900 ; CHECK-NEXT: vbroadcasti64x2 {{.*#+}} ymm0 {%k1} = mem[0,1,0,1]
901 ; CHECK-NEXT: retq
902 %vec = load <2 x i64>, ptr %vp
903 %shuf = shufflevector <2 x i64> %vec, <2 x i64> undef, <4 x i32> <i32 0, i32 1, i32 0, i32 1>
904 %cmp = icmp eq <4 x i64> %mask, zeroinitializer
905 %res = select <4 x i1> %cmp, <4 x i64> %shuf, <4 x i64> %default
906 ret <4 x i64> %res
907 }
909 define <4 x i64> @test_masked_z_2xi64_to_4xi64_mem_mask1(ptr %vp, <4 x i64> %mask) {
910 ; CHECK-LABEL: test_masked_z_2xi64_to_4xi64_mem_mask1:
911 ; CHECK: # %bb.0:
912 ; CHECK-NEXT: vptestnmq %ymm0, %ymm0, %k1
913 ; CHECK-NEXT: vbroadcasti64x2 {{.*#+}} ymm0 {%k1} {z} = mem[0,1,0,1]
914 ; CHECK-NEXT: retq
915 %vec = load <2 x i64>, ptr %vp
916 %shuf = shufflevector <2 x i64> %vec, <2 x i64> undef, <4 x i32> <i32 0, i32 1, i32 0, i32 1>
917 %cmp = icmp eq <4 x i64> %mask, zeroinitializer
918 %res = select <4 x i1> %cmp, <4 x i64> %shuf, <4 x i64> zeroinitializer
919 ret <4 x i64> %res
920 }
921 define <4 x i64> @test_masked_2xi64_to_4xi64_mem_mask2(ptr %vp, <4 x i64> %default, <4 x i64> %mask) {
922 ; CHECK-LABEL: test_masked_2xi64_to_4xi64_mem_mask2:
923 ; CHECK: # %bb.0:
924 ; CHECK-NEXT: vptestnmq %ymm1, %ymm1, %k1
925 ; CHECK-NEXT: vbroadcasti64x2 {{.*#+}} ymm0 {%k1} = mem[0,1,0,1]
926 ; CHECK-NEXT: retq
927 %vec = load <2 x i64>, ptr %vp
928 %shuf = shufflevector <2 x i64> %vec, <2 x i64> undef, <4 x i32> <i32 0, i32 1, i32 0, i32 1>
929 %cmp = icmp eq <4 x i64> %mask, zeroinitializer
930 %res = select <4 x i1> %cmp, <4 x i64> %shuf, <4 x i64> %default
931 ret <4 x i64> %res
932 }
934 define <4 x i64> @test_masked_z_2xi64_to_4xi64_mem_mask2(ptr %vp, <4 x i64> %mask) {
935 ; CHECK-LABEL: test_masked_z_2xi64_to_4xi64_mem_mask2:
936 ; CHECK: # %bb.0:
937 ; CHECK-NEXT: vptestnmq %ymm0, %ymm0, %k1
938 ; CHECK-NEXT: vbroadcasti64x2 {{.*#+}} ymm0 {%k1} {z} = mem[0,1,0,1]
939 ; CHECK-NEXT: retq
940 %vec = load <2 x i64>, ptr %vp
941 %shuf = shufflevector <2 x i64> %vec, <2 x i64> undef, <4 x i32> <i32 0, i32 1, i32 0, i32 1>
942 %cmp = icmp eq <4 x i64> %mask, zeroinitializer
943 %res = select <4 x i1> %cmp, <4 x i64> %shuf, <4 x i64> zeroinitializer
944 ret <4 x i64> %res
945 }
946 define <4 x i64> @test_masked_2xi64_to_4xi64_mem_mask3(ptr %vp, <4 x i64> %default, <4 x i64> %mask) {
947 ; CHECK-LABEL: test_masked_2xi64_to_4xi64_mem_mask3:
948 ; CHECK: # %bb.0:
949 ; CHECK-NEXT: vptestnmq %ymm1, %ymm1, %k1
950 ; CHECK-NEXT: vbroadcasti64x2 {{.*#+}} ymm0 {%k1} = mem[0,1,0,1]
951 ; CHECK-NEXT: retq
952 %vec = load <2 x i64>, ptr %vp
953 %shuf = shufflevector <2 x i64> %vec, <2 x i64> undef, <4 x i32> <i32 0, i32 1, i32 0, i32 1>
954 %cmp = icmp eq <4 x i64> %mask, zeroinitializer
955 %res = select <4 x i1> %cmp, <4 x i64> %shuf, <4 x i64> %default
956 ret <4 x i64> %res
957 }
959 define <4 x i64> @test_masked_z_2xi64_to_4xi64_mem_mask3(ptr %vp, <4 x i64> %mask) {
960 ; CHECK-LABEL: test_masked_z_2xi64_to_4xi64_mem_mask3:
961 ; CHECK: # %bb.0:
962 ; CHECK-NEXT: vptestnmq %ymm0, %ymm0, %k1
963 ; CHECK-NEXT: vbroadcasti64x2 {{.*#+}} ymm0 {%k1} {z} = mem[0,1,0,1]
964 ; CHECK-NEXT: retq
965 %vec = load <2 x i64>, ptr %vp
966 %shuf = shufflevector <2 x i64> %vec, <2 x i64> undef, <4 x i32> <i32 0, i32 1, i32 0, i32 1>
967 %cmp = icmp eq <4 x i64> %mask, zeroinitializer
968 %res = select <4 x i1> %cmp, <4 x i64> %shuf, <4 x i64> zeroinitializer
969 ret <4 x i64> %res
970 }
971 define <8 x i64> @test_2xi64_to_8xi64_mem(ptr %vp) {
972 ; CHECK-LABEL: test_2xi64_to_8xi64_mem:
973 ; CHECK: # %bb.0:
974 ; CHECK-NEXT: vbroadcasti32x4 {{.*#+}} zmm0 = mem[0,1,2,3,0,1,2,3,0,1,2,3,0,1,2,3]
975 ; CHECK-NEXT: retq
976 %vec = load <2 x i64>, ptr %vp
977 %res = shufflevector <2 x i64> %vec, <2 x i64> undef, <8 x i32> <i32 0, i32 1, i32 0, i32 1, i32 0, i32 1, i32 0, i32 1>
978 ret <8 x i64> %res
979 }
980 define <8 x i64> @test_masked_2xi64_to_8xi64_mem_mask0(ptr %vp, <8 x i64> %default, <8 x i64> %mask) {
981 ; CHECK-LABEL: test_masked_2xi64_to_8xi64_mem_mask0:
982 ; CHECK: # %bb.0:
983 ; CHECK-NEXT: vptestnmq %zmm1, %zmm1, %k1
984 ; CHECK-NEXT: vbroadcasti64x2 {{.*#+}} zmm0 {%k1} = mem[0,1,0,1,0,1,0,1]
985 ; CHECK-NEXT: retq
986 %vec = load <2 x i64>, ptr %vp
987 %shuf = shufflevector <2 x i64> %vec, <2 x i64> undef, <8 x i32> <i32 0, i32 1, i32 0, i32 1, i32 0, i32 1, i32 0, i32 1>
988 %cmp = icmp eq <8 x i64> %mask, zeroinitializer
989 %res = select <8 x i1> %cmp, <8 x i64> %shuf, <8 x i64> %default
990 ret <8 x i64> %res
991 }
993 define <8 x i64> @test_masked_z_2xi64_to_8xi64_mem_mask0(ptr %vp, <8 x i64> %mask) {
994 ; CHECK-LABEL: test_masked_z_2xi64_to_8xi64_mem_mask0:
995 ; CHECK: # %bb.0:
996 ; CHECK-NEXT: vptestnmq %zmm0, %zmm0, %k1
997 ; CHECK-NEXT: vbroadcasti64x2 {{.*#+}} zmm0 {%k1} {z} = mem[0,1,0,1,0,1,0,1]
998 ; CHECK-NEXT: retq
999 %vec = load <2 x i64>, ptr %vp
1000 %shuf = shufflevector <2 x i64> %vec, <2 x i64> undef, <8 x i32> <i32 0, i32 1, i32 0, i32 1, i32 0, i32 1, i32 0, i32 1>
1001 %cmp = icmp eq <8 x i64> %mask, zeroinitializer
1002 %res = select <8 x i1> %cmp, <8 x i64> %shuf, <8 x i64> zeroinitializer
1003 ret <8 x i64> %res
1004 }
1005 define <8 x i64> @test_masked_2xi64_to_8xi64_mem_mask1(ptr %vp, <8 x i64> %default, <8 x i64> %mask) {
1006 ; CHECK-LABEL: test_masked_2xi64_to_8xi64_mem_mask1:
1007 ; CHECK: # %bb.0:
1008 ; CHECK-NEXT: vptestnmq %zmm1, %zmm1, %k1
1009 ; CHECK-NEXT: vbroadcasti64x2 {{.*#+}} zmm0 {%k1} = mem[0,1,0,1,0,1,0,1]
1010 ; CHECK-NEXT: retq
1011 %vec = load <2 x i64>, ptr %vp
1012 %shuf = shufflevector <2 x i64> %vec, <2 x i64> undef, <8 x i32> <i32 0, i32 1, i32 0, i32 1, i32 0, i32 1, i32 0, i32 1>
1013 %cmp = icmp eq <8 x i64> %mask, zeroinitializer
1014 %res = select <8 x i1> %cmp, <8 x i64> %shuf, <8 x i64> %default
1015 ret <8 x i64> %res
1016 }
1018 define <8 x i64> @test_masked_z_2xi64_to_8xi64_mem_mask1(ptr %vp, <8 x i64> %mask) {
1019 ; CHECK-LABEL: test_masked_z_2xi64_to_8xi64_mem_mask1:
1020 ; CHECK: # %bb.0:
1021 ; CHECK-NEXT: vptestnmq %zmm0, %zmm0, %k1
1022 ; CHECK-NEXT: vbroadcasti64x2 {{.*#+}} zmm0 {%k1} {z} = mem[0,1,0,1,0,1,0,1]
1023 ; CHECK-NEXT: retq
1024 %vec = load <2 x i64>, ptr %vp
1025 %shuf = shufflevector <2 x i64> %vec, <2 x i64> undef, <8 x i32> <i32 0, i32 1, i32 0, i32 1, i32 0, i32 1, i32 0, i32 1>
1026 %cmp = icmp eq <8 x i64> %mask, zeroinitializer
1027 %res = select <8 x i1> %cmp, <8 x i64> %shuf, <8 x i64> zeroinitializer
1028 ret <8 x i64> %res
1029 }
1030 define <8 x i64> @test_masked_2xi64_to_8xi64_mem_mask2(ptr %vp, <8 x i64> %default, <8 x i64> %mask) {
1031 ; CHECK-LABEL: test_masked_2xi64_to_8xi64_mem_mask2:
1032 ; CHECK: # %bb.0:
1033 ; CHECK-NEXT: vptestnmq %zmm1, %zmm1, %k1
1034 ; CHECK-NEXT: vbroadcasti64x2 {{.*#+}} zmm0 {%k1} = mem[0,1,0,1,0,1,0,1]
1035 ; CHECK-NEXT: retq
1036 %vec = load <2 x i64>, ptr %vp
1037 %shuf = shufflevector <2 x i64> %vec, <2 x i64> undef, <8 x i32> <i32 0, i32 1, i32 0, i32 1, i32 0, i32 1, i32 0, i32 1>
1038 %cmp = icmp eq <8 x i64> %mask, zeroinitializer
1039 %res = select <8 x i1> %cmp, <8 x i64> %shuf, <8 x i64> %default
1040 ret <8 x i64> %res
1041 }
1043 define <8 x i64> @test_masked_z_2xi64_to_8xi64_mem_mask2(ptr %vp, <8 x i64> %mask) {
1044 ; CHECK-LABEL: test_masked_z_2xi64_to_8xi64_mem_mask2:
1045 ; CHECK: # %bb.0:
1046 ; CHECK-NEXT: vptestnmq %zmm0, %zmm0, %k1
1047 ; CHECK-NEXT: vbroadcasti64x2 {{.*#+}} zmm0 {%k1} {z} = mem[0,1,0,1,0,1,0,1]
1048 ; CHECK-NEXT: retq
1049 %vec = load <2 x i64>, ptr %vp
1050 %shuf = shufflevector <2 x i64> %vec, <2 x i64> undef, <8 x i32> <i32 0, i32 1, i32 0, i32 1, i32 0, i32 1, i32 0, i32 1>
1051 %cmp = icmp eq <8 x i64> %mask, zeroinitializer
1052 %res = select <8 x i1> %cmp, <8 x i64> %shuf, <8 x i64> zeroinitializer
1053 ret <8 x i64> %res
1054 }
1055 define <8 x i64> @test_masked_2xi64_to_8xi64_mem_mask3(ptr %vp, <8 x i64> %default, <8 x i64> %mask) {
1056 ; CHECK-LABEL: test_masked_2xi64_to_8xi64_mem_mask3:
1057 ; CHECK: # %bb.0:
1058 ; CHECK-NEXT: vptestnmq %zmm1, %zmm1, %k1
1059 ; CHECK-NEXT: vbroadcasti64x2 {{.*#+}} zmm0 {%k1} = mem[0,1,0,1,0,1,0,1]
1060 ; CHECK-NEXT: retq
1061 %vec = load <2 x i64>, ptr %vp
1062 %shuf = shufflevector <2 x i64> %vec, <2 x i64> undef, <8 x i32> <i32 0, i32 1, i32 0, i32 1, i32 0, i32 1, i32 0, i32 1>
1063 %cmp = icmp eq <8 x i64> %mask, zeroinitializer
1064 %res = select <8 x i1> %cmp, <8 x i64> %shuf, <8 x i64> %default
1065 ret <8 x i64> %res
1066 }
1068 define <8 x i64> @test_masked_z_2xi64_to_8xi64_mem_mask3(ptr %vp, <8 x i64> %mask) {
1069 ; CHECK-LABEL: test_masked_z_2xi64_to_8xi64_mem_mask3:
1070 ; CHECK: # %bb.0:
1071 ; CHECK-NEXT: vptestnmq %zmm0, %zmm0, %k1
1072 ; CHECK-NEXT: vbroadcasti64x2 {{.*#+}} zmm0 {%k1} {z} = mem[0,1,0,1,0,1,0,1]
1073 ; CHECK-NEXT: retq
1074 %vec = load <2 x i64>, ptr %vp
1075 %shuf = shufflevector <2 x i64> %vec, <2 x i64> undef, <8 x i32> <i32 0, i32 1, i32 0, i32 1, i32 0, i32 1, i32 0, i32 1>
1076 %cmp = icmp eq <8 x i64> %mask, zeroinitializer
1077 %res = select <8 x i1> %cmp, <8 x i64> %shuf, <8 x i64> zeroinitializer
1078 ret <8 x i64> %res
1079 }
1080 define <16 x i32> @test_8xi32_to_16xi32_mem(ptr %vp) {
1081 ; CHECK-LABEL: test_8xi32_to_16xi32_mem:
1082 ; CHECK: # %bb.0:
1083 ; CHECK-NEXT: vbroadcasti64x4 {{.*#+}} zmm0 = mem[0,1,2,3,0,1,2,3]
1084 ; CHECK-NEXT: retq
1085 %vec = load <8 x i32>, ptr %vp
1086 %res = shufflevector <8 x i32> %vec, <8 x i32> undef, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
1087 ret <16 x i32> %res
1088 }
1089 define <16 x i32> @test_masked_8xi32_to_16xi32_mem_mask0(ptr %vp, <16 x i32> %default, <16 x i32> %mask) {
1090 ; CHECK-LABEL: test_masked_8xi32_to_16xi32_mem_mask0:
1091 ; CHECK: # %bb.0:
1092 ; CHECK-NEXT: vptestnmd %zmm1, %zmm1, %k1
1093 ; CHECK-NEXT: vbroadcasti32x8 {{.*#+}} zmm0 {%k1} = mem[0,1,2,3,4,5,6,7,0,1,2,3,4,5,6,7]
1094 ; CHECK-NEXT: retq
1095 %vec = load <8 x i32>, ptr %vp
1096 %shuf = shufflevector <8 x i32> %vec, <8 x i32> undef, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
1097 %cmp = icmp eq <16 x i32> %mask, zeroinitializer
1098 %res = select <16 x i1> %cmp, <16 x i32> %shuf, <16 x i32> %default
1099 ret <16 x i32> %res
1100 }
1102 define <16 x i32> @test_masked_z_8xi32_to_16xi32_mem_mask0(ptr %vp, <16 x i32> %mask) {
1103 ; CHECK-LABEL: test_masked_z_8xi32_to_16xi32_mem_mask0:
1104 ; CHECK: # %bb.0:
1105 ; CHECK-NEXT: vptestnmd %zmm0, %zmm0, %k1
1106 ; CHECK-NEXT: vbroadcasti32x8 {{.*#+}} zmm0 {%k1} {z} = mem[0,1,2,3,4,5,6,7,0,1,2,3,4,5,6,7]
1107 ; CHECK-NEXT: retq
1108 %vec = load <8 x i32>, ptr %vp
1109 %shuf = shufflevector <8 x i32> %vec, <8 x i32> undef, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
1110 %cmp = icmp eq <16 x i32> %mask, zeroinitializer
1111 %res = select <16 x i1> %cmp, <16 x i32> %shuf, <16 x i32> zeroinitializer
1112 ret <16 x i32> %res
1113 }
1114 define <16 x i32> @test_masked_8xi32_to_16xi32_mem_mask1(ptr %vp, <16 x i32> %default, <16 x i32> %mask) {
1115 ; CHECK-LABEL: test_masked_8xi32_to_16xi32_mem_mask1:
1116 ; CHECK: # %bb.0:
1117 ; CHECK-NEXT: vptestnmd %zmm1, %zmm1, %k1
1118 ; CHECK-NEXT: vbroadcasti32x8 {{.*#+}} zmm0 {%k1} = mem[0,1,2,3,4,5,6,7,0,1,2,3,4,5,6,7]
1119 ; CHECK-NEXT: retq
1120 %vec = load <8 x i32>, ptr %vp
1121 %shuf = shufflevector <8 x i32> %vec, <8 x i32> undef, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
1122 %cmp = icmp eq <16 x i32> %mask, zeroinitializer
1123 %res = select <16 x i1> %cmp, <16 x i32> %shuf, <16 x i32> %default
1124 ret <16 x i32> %res
1125 }
1127 define <16 x i32> @test_masked_z_8xi32_to_16xi32_mem_mask1(ptr %vp, <16 x i32> %mask) {
1128 ; CHECK-LABEL: test_masked_z_8xi32_to_16xi32_mem_mask1:
1129 ; CHECK: # %bb.0:
1130 ; CHECK-NEXT: vptestnmd %zmm0, %zmm0, %k1
1131 ; CHECK-NEXT: vbroadcasti32x8 {{.*#+}} zmm0 {%k1} {z} = mem[0,1,2,3,4,5,6,7,0,1,2,3,4,5,6,7]
1132 ; CHECK-NEXT: retq
1133 %vec = load <8 x i32>, ptr %vp
1134 %shuf = shufflevector <8 x i32> %vec, <8 x i32> undef, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
1135 %cmp = icmp eq <16 x i32> %mask, zeroinitializer
1136 %res = select <16 x i1> %cmp, <16 x i32> %shuf, <16 x i32> zeroinitializer
1137 ret <16 x i32> %res
1138 }
1139 define <16 x i32> @test_masked_8xi32_to_16xi32_mem_mask2(ptr %vp, <16 x i32> %default, <16 x i32> %mask) {
1140 ; CHECK-LABEL: test_masked_8xi32_to_16xi32_mem_mask2:
1141 ; CHECK: # %bb.0:
1142 ; CHECK-NEXT: vptestnmd %zmm1, %zmm1, %k1
1143 ; CHECK-NEXT: vbroadcasti32x8 {{.*#+}} zmm0 {%k1} = mem[0,1,2,3,4,5,6,7,0,1,2,3,4,5,6,7]
1144 ; CHECK-NEXT: retq
1145 %vec = load <8 x i32>, ptr %vp
1146 %shuf = shufflevector <8 x i32> %vec, <8 x i32> undef, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
1147 %cmp = icmp eq <16 x i32> %mask, zeroinitializer
1148 %res = select <16 x i1> %cmp, <16 x i32> %shuf, <16 x i32> %default
1149 ret <16 x i32> %res
1150 }
1152 define <16 x i32> @test_masked_z_8xi32_to_16xi32_mem_mask2(ptr %vp, <16 x i32> %mask) {
1153 ; CHECK-LABEL: test_masked_z_8xi32_to_16xi32_mem_mask2:
1154 ; CHECK: # %bb.0:
1155 ; CHECK-NEXT: vptestnmd %zmm0, %zmm0, %k1
1156 ; CHECK-NEXT: vbroadcasti32x8 {{.*#+}} zmm0 {%k1} {z} = mem[0,1,2,3,4,5,6,7,0,1,2,3,4,5,6,7]
1157 ; CHECK-NEXT: retq
1158 %vec = load <8 x i32>, ptr %vp
1159 %shuf = shufflevector <8 x i32> %vec, <8 x i32> undef, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
1160 %cmp = icmp eq <16 x i32> %mask, zeroinitializer
1161 %res = select <16 x i1> %cmp, <16 x i32> %shuf, <16 x i32> zeroinitializer
1162 ret <16 x i32> %res
1163 }
1164 define <16 x i32> @test_masked_8xi32_to_16xi32_mem_mask3(ptr %vp, <16 x i32> %default, <16 x i32> %mask) {
1165 ; CHECK-LABEL: test_masked_8xi32_to_16xi32_mem_mask3:
1166 ; CHECK: # %bb.0:
1167 ; CHECK-NEXT: vptestnmd %zmm1, %zmm1, %k1
1168 ; CHECK-NEXT: vbroadcasti32x8 {{.*#+}} zmm0 {%k1} = mem[0,1,2,3,4,5,6,7,0,1,2,3,4,5,6,7]
1169 ; CHECK-NEXT: retq
1170 %vec = load <8 x i32>, ptr %vp
1171 %shuf = shufflevector <8 x i32> %vec, <8 x i32> undef, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
1172 %cmp = icmp eq <16 x i32> %mask, zeroinitializer
1173 %res = select <16 x i1> %cmp, <16 x i32> %shuf, <16 x i32> %default
1174 ret <16 x i32> %res
1175 }
1177 define <16 x i32> @test_masked_z_8xi32_to_16xi32_mem_mask3(ptr %vp, <16 x i32> %mask) {
1178 ; CHECK-LABEL: test_masked_z_8xi32_to_16xi32_mem_mask3:
1179 ; CHECK: # %bb.0:
1180 ; CHECK-NEXT: vptestnmd %zmm0, %zmm0, %k1
1181 ; CHECK-NEXT: vbroadcasti32x8 {{.*#+}} zmm0 {%k1} {z} = mem[0,1,2,3,4,5,6,7,0,1,2,3,4,5,6,7]
1182 ; CHECK-NEXT: retq
1183 %vec = load <8 x i32>, ptr %vp
1184 %shuf = shufflevector <8 x i32> %vec, <8 x i32> undef, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
1185 %cmp = icmp eq <16 x i32> %mask, zeroinitializer
1186 %res = select <16 x i1> %cmp, <16 x i32> %shuf, <16 x i32> zeroinitializer
1187 ret <16 x i32> %res
1188 }
1189 define <8 x i64> @test_4xi64_to_8xi64_mem(ptr %vp) {
1190 ; CHECK-LABEL: test_4xi64_to_8xi64_mem:
1191 ; CHECK: # %bb.0:
1192 ; CHECK-NEXT: vbroadcasti64x4 {{.*#+}} zmm0 = mem[0,1,2,3,0,1,2,3]
1193 ; CHECK-NEXT: retq
1194 %vec = load <4 x i64>, ptr %vp
1195 %res = shufflevector <4 x i64> %vec, <4 x i64> undef, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 0, i32 1, i32 2, i32 3>
1196 ret <8 x i64> %res
1197 }
1198 define <8 x i64> @test_masked_4xi64_to_8xi64_mem_mask0(ptr %vp, <8 x i64> %default, <8 x i64> %mask) {
1199 ; CHECK-LABEL: test_masked_4xi64_to_8xi64_mem_mask0:
1200 ; CHECK: # %bb.0:
1201 ; CHECK-NEXT: vptestnmq %zmm1, %zmm1, %k1
1202 ; CHECK-NEXT: vbroadcasti64x4 {{.*#+}} zmm0 {%k1} = mem[0,1,2,3,0,1,2,3]
1203 ; CHECK-NEXT: retq
1204 %vec = load <4 x i64>, ptr %vp
1205 %shuf = shufflevector <4 x i64> %vec, <4 x i64> undef, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 0, i32 1, i32 2, i32 3>
1206 %cmp = icmp eq <8 x i64> %mask, zeroinitializer
1207 %res = select <8 x i1> %cmp, <8 x i64> %shuf, <8 x i64> %default
1208 ret <8 x i64> %res
1209 }
1211 define <8 x i64> @test_masked_z_4xi64_to_8xi64_mem_mask0(ptr %vp, <8 x i64> %mask) {
1212 ; CHECK-LABEL: test_masked_z_4xi64_to_8xi64_mem_mask0:
1213 ; CHECK: # %bb.0:
1214 ; CHECK-NEXT: vptestnmq %zmm0, %zmm0, %k1
1215 ; CHECK-NEXT: vbroadcasti64x4 {{.*#+}} zmm0 {%k1} {z} = mem[0,1,2,3,0,1,2,3]
1216 ; CHECK-NEXT: retq
1217 %vec = load <4 x i64>, ptr %vp
1218 %shuf = shufflevector <4 x i64> %vec, <4 x i64> undef, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 0, i32 1, i32 2, i32 3>
1219 %cmp = icmp eq <8 x i64> %mask, zeroinitializer
1220 %res = select <8 x i1> %cmp, <8 x i64> %shuf, <8 x i64> zeroinitializer
1221 ret <8 x i64> %res
1222 }
1223 define <8 x i64> @test_masked_4xi64_to_8xi64_mem_mask1(ptr %vp, <8 x i64> %default, <8 x i64> %mask) {
1224 ; CHECK-LABEL: test_masked_4xi64_to_8xi64_mem_mask1:
1225 ; CHECK: # %bb.0:
1226 ; CHECK-NEXT: vptestnmq %zmm1, %zmm1, %k1
1227 ; CHECK-NEXT: vbroadcasti64x4 {{.*#+}} zmm0 {%k1} = mem[0,1,2,3,0,1,2,3]
1228 ; CHECK-NEXT: retq
1229 %vec = load <4 x i64>, ptr %vp
1230 %shuf = shufflevector <4 x i64> %vec, <4 x i64> undef, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 0, i32 1, i32 2, i32 3>
1231 %cmp = icmp eq <8 x i64> %mask, zeroinitializer
1232 %res = select <8 x i1> %cmp, <8 x i64> %shuf, <8 x i64> %default
1233 ret <8 x i64> %res
1234 }
1236 define <8 x i64> @test_masked_z_4xi64_to_8xi64_mem_mask1(ptr %vp, <8 x i64> %mask) {
1237 ; CHECK-LABEL: test_masked_z_4xi64_to_8xi64_mem_mask1:
1238 ; CHECK: # %bb.0:
1239 ; CHECK-NEXT: vptestnmq %zmm0, %zmm0, %k1
1240 ; CHECK-NEXT: vbroadcasti64x4 {{.*#+}} zmm0 {%k1} {z} = mem[0,1,2,3,0,1,2,3]
1241 ; CHECK-NEXT: retq
1242 %vec = load <4 x i64>, ptr %vp
1243 %shuf = shufflevector <4 x i64> %vec, <4 x i64> undef, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 0, i32 1, i32 2, i32 3>
1244 %cmp = icmp eq <8 x i64> %mask, zeroinitializer
1245 %res = select <8 x i1> %cmp, <8 x i64> %shuf, <8 x i64> zeroinitializer
1246 ret <8 x i64> %res
1247 }
1248 define <8 x i64> @test_masked_4xi64_to_8xi64_mem_mask2(ptr %vp, <8 x i64> %default, <8 x i64> %mask) {
1249 ; CHECK-LABEL: test_masked_4xi64_to_8xi64_mem_mask2:
1250 ; CHECK: # %bb.0:
1251 ; CHECK-NEXT: vptestnmq %zmm1, %zmm1, %k1
1252 ; CHECK-NEXT: vbroadcasti64x4 {{.*#+}} zmm0 {%k1} = mem[0,1,2,3,0,1,2,3]
1253 ; CHECK-NEXT: retq
1254 %vec = load <4 x i64>, ptr %vp
1255 %shuf = shufflevector <4 x i64> %vec, <4 x i64> undef, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 0, i32 1, i32 2, i32 3>
1256 %cmp = icmp eq <8 x i64> %mask, zeroinitializer
1257 %res = select <8 x i1> %cmp, <8 x i64> %shuf, <8 x i64> %default
1258 ret <8 x i64> %res
1259 }
1261 define <8 x i64> @test_masked_z_4xi64_to_8xi64_mem_mask2(ptr %vp, <8 x i64> %mask) {
1262 ; CHECK-LABEL: test_masked_z_4xi64_to_8xi64_mem_mask2:
1263 ; CHECK: # %bb.0:
1264 ; CHECK-NEXT: vptestnmq %zmm0, %zmm0, %k1
1265 ; CHECK-NEXT: vbroadcasti64x4 {{.*#+}} zmm0 {%k1} {z} = mem[0,1,2,3,0,1,2,3]
1266 ; CHECK-NEXT: retq
1267 %vec = load <4 x i64>, ptr %vp
1268 %shuf = shufflevector <4 x i64> %vec, <4 x i64> undef, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 0, i32 1, i32 2, i32 3>
1269 %cmp = icmp eq <8 x i64> %mask, zeroinitializer
1270 %res = select <8 x i1> %cmp, <8 x i64> %shuf, <8 x i64> zeroinitializer
1271 ret <8 x i64> %res
1272 }
1273 define <8 x i64> @test_masked_4xi64_to_8xi64_mem_mask3(ptr %vp, <8 x i64> %default, <8 x i64> %mask) {
1274 ; CHECK-LABEL: test_masked_4xi64_to_8xi64_mem_mask3:
1275 ; CHECK: # %bb.0:
1276 ; CHECK-NEXT: vptestnmq %zmm1, %zmm1, %k1
1277 ; CHECK-NEXT: vbroadcasti64x4 {{.*#+}} zmm0 {%k1} = mem[0,1,2,3,0,1,2,3]
1278 ; CHECK-NEXT: retq
1279 %vec = load <4 x i64>, ptr %vp
1280 %shuf = shufflevector <4 x i64> %vec, <4 x i64> undef, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 0, i32 1, i32 2, i32 3>
1281 %cmp = icmp eq <8 x i64> %mask, zeroinitializer
1282 %res = select <8 x i1> %cmp, <8 x i64> %shuf, <8 x i64> %default
1283 ret <8 x i64> %res
1284 }
1286 define <8 x i64> @test_masked_z_4xi64_to_8xi64_mem_mask3(ptr %vp, <8 x i64> %mask) {
1287 ; CHECK-LABEL: test_masked_z_4xi64_to_8xi64_mem_mask3:
1288 ; CHECK: # %bb.0:
1289 ; CHECK-NEXT: vptestnmq %zmm0, %zmm0, %k1
1290 ; CHECK-NEXT: vbroadcasti64x4 {{.*#+}} zmm0 {%k1} {z} = mem[0,1,2,3,0,1,2,3]
1291 ; CHECK-NEXT: retq
1292 %vec = load <4 x i64>, ptr %vp
1293 %shuf = shufflevector <4 x i64> %vec, <4 x i64> undef, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 0, i32 1, i32 2, i32 3>
1294 %cmp = icmp eq <8 x i64> %mask, zeroinitializer
1295 %res = select <8 x i1> %cmp, <8 x i64> %shuf, <8 x i64> zeroinitializer
1296 ret <8 x i64> %res
1297 }