| blob | 656f1e031a7cbe0a81792c0fcbd6f123450355b7 |
1 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
2 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+SSE2 | FileCheck %s --check-prefix=SSE2-SSSE3
3 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+SSSE3 | FileCheck %s --check-prefix=SSE2-SSSE3
4 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx | FileCheck %s --check-prefixes=AVX12,AVX1
5 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx2 | FileCheck %s --check-prefixes=AVX12,AVX2
6 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512f,+avx512vl | FileCheck %s --check-prefixes=AVX512,AVX512F
7 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512f,+avx512vl,+avx512bw | FileCheck %s --check-prefixes=AVX512,AVX512BW
9 define i16 @v16i16(<16 x i16> %a, <16 x i16> %b) {
10 ; SSE2-SSSE3-LABEL: v16i16:
11 ; SSE2-SSSE3: # %bb.0:
12 ; SSE2-SSSE3-NEXT: pcmpgtw %xmm3, %xmm1
13 ; SSE2-SSSE3-NEXT: pcmpgtw %xmm2, %xmm0
14 ; SSE2-SSSE3-NEXT: packsswb %xmm1, %xmm0
15 ; SSE2-SSSE3-NEXT: pmovmskb %xmm0, %eax
16 ; SSE2-SSSE3-NEXT: # kill: def $ax killed $ax killed $eax
17 ; SSE2-SSSE3-NEXT: retq
18 ;
19 ; AVX1-LABEL: v16i16:
20 ; AVX1: # %bb.0:
21 ; AVX1-NEXT: vextractf128 $1, %ymm1, %xmm2
22 ; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm3
23 ; AVX1-NEXT: vpcmpgtw %xmm2, %xmm3, %xmm2
24 ; AVX1-NEXT: vpcmpgtw %xmm1, %xmm0, %xmm0
25 ; AVX1-NEXT: vpacksswb %xmm2, %xmm0, %xmm0
26 ; AVX1-NEXT: vpmovmskb %xmm0, %eax
27 ; AVX1-NEXT: # kill: def $ax killed $ax killed $eax
28 ; AVX1-NEXT: vzeroupper
29 ; AVX1-NEXT: retq
30 ;
31 ; AVX2-LABEL: v16i16:
32 ; AVX2: # %bb.0:
33 ; AVX2-NEXT: vpcmpgtw %ymm1, %ymm0, %ymm0
34 ; AVX2-NEXT: vextracti128 $1, %ymm0, %xmm1
35 ; AVX2-NEXT: vpacksswb %xmm1, %xmm0, %xmm0
36 ; AVX2-NEXT: vpmovmskb %xmm0, %eax
37 ; AVX2-NEXT: # kill: def $ax killed $ax killed $eax
38 ; AVX2-NEXT: vzeroupper
39 ; AVX2-NEXT: retq
40 ;
41 ; AVX512F-LABEL: v16i16:
42 ; AVX512F: # %bb.0:
43 ; AVX512F-NEXT: vpcmpgtw %ymm1, %ymm0, %ymm0
44 ; AVX512F-NEXT: vpmovsxwd %ymm0, %zmm0
45 ; AVX512F-NEXT: vptestmd %zmm0, %zmm0, %k0
46 ; AVX512F-NEXT: kmovw %k0, %eax
47 ; AVX512F-NEXT: # kill: def $ax killed $ax killed $eax
48 ; AVX512F-NEXT: vzeroupper
49 ; AVX512F-NEXT: retq
50 ;
51 ; AVX512BW-LABEL: v16i16:
52 ; AVX512BW: # %bb.0:
53 ; AVX512BW-NEXT: vpcmpgtw %ymm1, %ymm0, %k0
54 ; AVX512BW-NEXT: kmovd %k0, %eax
55 ; AVX512BW-NEXT: # kill: def $ax killed $ax killed $eax
56 ; AVX512BW-NEXT: vzeroupper
57 ; AVX512BW-NEXT: retq
58 %x = icmp sgt <16 x i16> %a, %b
59 %res = bitcast <16 x i1> %x to i16
60 ret i16 %res
61 }
63 define i8 @v8i32(<8 x i32> %a, <8 x i32> %b) {
64 ; SSE2-SSSE3-LABEL: v8i32:
65 ; SSE2-SSSE3: # %bb.0:
66 ; SSE2-SSSE3-NEXT: pcmpgtd %xmm3, %xmm1
67 ; SSE2-SSSE3-NEXT: pcmpgtd %xmm2, %xmm0
68 ; SSE2-SSSE3-NEXT: packssdw %xmm1, %xmm0
69 ; SSE2-SSSE3-NEXT: packsswb %xmm0, %xmm0
70 ; SSE2-SSSE3-NEXT: pmovmskb %xmm0, %eax
71 ; SSE2-SSSE3-NEXT: # kill: def $al killed $al killed $eax
72 ; SSE2-SSSE3-NEXT: retq
73 ;
74 ; AVX1-LABEL: v8i32:
75 ; AVX1: # %bb.0:
76 ; AVX1-NEXT: vextractf128 $1, %ymm1, %xmm2
77 ; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm3
78 ; AVX1-NEXT: vpcmpgtd %xmm2, %xmm3, %xmm2
79 ; AVX1-NEXT: vpcmpgtd %xmm1, %xmm0, %xmm0
80 ; AVX1-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm0
81 ; AVX1-NEXT: vmovmskps %ymm0, %eax
82 ; AVX1-NEXT: # kill: def $al killed $al killed $eax
83 ; AVX1-NEXT: vzeroupper
84 ; AVX1-NEXT: retq
85 ;
86 ; AVX2-LABEL: v8i32:
87 ; AVX2: # %bb.0:
88 ; AVX2-NEXT: vpcmpgtd %ymm1, %ymm0, %ymm0
89 ; AVX2-NEXT: vmovmskps %ymm0, %eax
90 ; AVX2-NEXT: # kill: def $al killed $al killed $eax
91 ; AVX2-NEXT: vzeroupper
92 ; AVX2-NEXT: retq
93 ;
94 ; AVX512F-LABEL: v8i32:
95 ; AVX512F: # %bb.0:
96 ; AVX512F-NEXT: vpcmpgtd %ymm1, %ymm0, %k0
97 ; AVX512F-NEXT: kmovw %k0, %eax
98 ; AVX512F-NEXT: # kill: def $al killed $al killed $eax
99 ; AVX512F-NEXT: vzeroupper
100 ; AVX512F-NEXT: retq
101 ;
102 ; AVX512BW-LABEL: v8i32:
103 ; AVX512BW: # %bb.0:
104 ; AVX512BW-NEXT: vpcmpgtd %ymm1, %ymm0, %k0
105 ; AVX512BW-NEXT: kmovd %k0, %eax
106 ; AVX512BW-NEXT: # kill: def $al killed $al killed $eax
107 ; AVX512BW-NEXT: vzeroupper
108 ; AVX512BW-NEXT: retq
109 %x = icmp sgt <8 x i32> %a, %b
110 %res = bitcast <8 x i1> %x to i8
111 ret i8 %res
112 }
114 define i8 @v8f32(<8 x float> %a, <8 x float> %b) {
115 ; SSE2-SSSE3-LABEL: v8f32:
116 ; SSE2-SSSE3: # %bb.0:
117 ; SSE2-SSSE3-NEXT: cmpltps %xmm1, %xmm3
118 ; SSE2-SSSE3-NEXT: cmpltps %xmm0, %xmm2
119 ; SSE2-SSSE3-NEXT: packssdw %xmm3, %xmm2
120 ; SSE2-SSSE3-NEXT: packsswb %xmm2, %xmm2
121 ; SSE2-SSSE3-NEXT: pmovmskb %xmm2, %eax
122 ; SSE2-SSSE3-NEXT: # kill: def $al killed $al killed $eax
123 ; SSE2-SSSE3-NEXT: retq
124 ;
125 ; AVX12-LABEL: v8f32:
126 ; AVX12: # %bb.0:
127 ; AVX12-NEXT: vcmpltps %ymm0, %ymm1, %ymm0
128 ; AVX12-NEXT: vmovmskps %ymm0, %eax
129 ; AVX12-NEXT: # kill: def $al killed $al killed $eax
130 ; AVX12-NEXT: vzeroupper
131 ; AVX12-NEXT: retq
132 ;
133 ; AVX512F-LABEL: v8f32:
134 ; AVX512F: # %bb.0:
135 ; AVX512F-NEXT: vcmpltps %ymm0, %ymm1, %k0
136 ; AVX512F-NEXT: kmovw %k0, %eax
137 ; AVX512F-NEXT: # kill: def $al killed $al killed $eax
138 ; AVX512F-NEXT: vzeroupper
139 ; AVX512F-NEXT: retq
140 ;
141 ; AVX512BW-LABEL: v8f32:
142 ; AVX512BW: # %bb.0:
143 ; AVX512BW-NEXT: vcmpltps %ymm0, %ymm1, %k0
144 ; AVX512BW-NEXT: kmovd %k0, %eax
145 ; AVX512BW-NEXT: # kill: def $al killed $al killed $eax
146 ; AVX512BW-NEXT: vzeroupper
147 ; AVX512BW-NEXT: retq
148 %x = fcmp ogt <8 x float> %a, %b
149 %res = bitcast <8 x i1> %x to i8
150 ret i8 %res
151 }
153 define i32 @v32i8(<32 x i8> %a, <32 x i8> %b) {
154 ; SSE2-SSSE3-LABEL: v32i8:
155 ; SSE2-SSSE3: # %bb.0:
156 ; SSE2-SSSE3-NEXT: pcmpgtb %xmm2, %xmm0
157 ; SSE2-SSSE3-NEXT: pmovmskb %xmm0, %ecx
158 ; SSE2-SSSE3-NEXT: pcmpgtb %xmm3, %xmm1
159 ; SSE2-SSSE3-NEXT: pmovmskb %xmm1, %eax
160 ; SSE2-SSSE3-NEXT: shll $16, %eax
161 ; SSE2-SSSE3-NEXT: orl %ecx, %eax
162 ; SSE2-SSSE3-NEXT: retq
163 ;
164 ; AVX1-LABEL: v32i8:
165 ; AVX1: # %bb.0:
166 ; AVX1-NEXT: vpcmpgtb %xmm1, %xmm0, %xmm2
167 ; AVX1-NEXT: vpmovmskb %xmm2, %ecx
168 ; AVX1-NEXT: vextractf128 $1, %ymm1, %xmm1
169 ; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm0
170 ; AVX1-NEXT: vpcmpgtb %xmm1, %xmm0, %xmm0
171 ; AVX1-NEXT: vpmovmskb %xmm0, %eax
172 ; AVX1-NEXT: shll $16, %eax
173 ; AVX1-NEXT: orl %ecx, %eax
174 ; AVX1-NEXT: vzeroupper
175 ; AVX1-NEXT: retq
176 ;
177 ; AVX2-LABEL: v32i8:
178 ; AVX2: # %bb.0:
179 ; AVX2-NEXT: vpcmpgtb %ymm1, %ymm0, %ymm0
180 ; AVX2-NEXT: vpmovmskb %ymm0, %eax
181 ; AVX2-NEXT: vzeroupper
182 ; AVX2-NEXT: retq
183 ;
184 ; AVX512F-LABEL: v32i8:
185 ; AVX512F: # %bb.0:
186 ; AVX512F-NEXT: vpcmpgtb %ymm1, %ymm0, %ymm0
187 ; AVX512F-NEXT: vpmovmskb %ymm0, %eax
188 ; AVX512F-NEXT: vzeroupper
189 ; AVX512F-NEXT: retq
190 ;
191 ; AVX512BW-LABEL: v32i8:
192 ; AVX512BW: # %bb.0:
193 ; AVX512BW-NEXT: vpcmpgtb %ymm1, %ymm0, %k0
194 ; AVX512BW-NEXT: kmovd %k0, %eax
195 ; AVX512BW-NEXT: vzeroupper
196 ; AVX512BW-NEXT: retq
197 %x = icmp sgt <32 x i8> %a, %b
198 %res = bitcast <32 x i1> %x to i32
199 ret i32 %res
200 }
202 define i4 @v4i64(<4 x i64> %a, <4 x i64> %b) {
203 ; SSE2-SSSE3-LABEL: v4i64:
204 ; SSE2-SSSE3: # %bb.0:
205 ; SSE2-SSSE3-NEXT: movdqa {{.*#+}} xmm4 = [2147483648,2147483648]
206 ; SSE2-SSSE3-NEXT: pxor %xmm4, %xmm3
207 ; SSE2-SSSE3-NEXT: pxor %xmm4, %xmm1
208 ; SSE2-SSSE3-NEXT: movdqa %xmm1, %xmm5
209 ; SSE2-SSSE3-NEXT: pcmpgtd %xmm3, %xmm5
210 ; SSE2-SSSE3-NEXT: pxor %xmm4, %xmm2
211 ; SSE2-SSSE3-NEXT: pxor %xmm4, %xmm0
212 ; SSE2-SSSE3-NEXT: movdqa %xmm0, %xmm4
213 ; SSE2-SSSE3-NEXT: pcmpgtd %xmm2, %xmm4
214 ; SSE2-SSSE3-NEXT: movdqa %xmm4, %xmm6
215 ; SSE2-SSSE3-NEXT: shufps {{.*#+}} xmm6 = xmm6[0,2],xmm5[0,2]
216 ; SSE2-SSSE3-NEXT: pcmpeqd %xmm3, %xmm1
217 ; SSE2-SSSE3-NEXT: pcmpeqd %xmm2, %xmm0
218 ; SSE2-SSSE3-NEXT: shufps {{.*#+}} xmm0 = xmm0[1,3],xmm1[1,3]
219 ; SSE2-SSSE3-NEXT: andps %xmm6, %xmm0
220 ; SSE2-SSSE3-NEXT: shufps {{.*#+}} xmm4 = xmm4[1,3],xmm5[1,3]
221 ; SSE2-SSSE3-NEXT: orps %xmm0, %xmm4
222 ; SSE2-SSSE3-NEXT: movmskps %xmm4, %eax
223 ; SSE2-SSSE3-NEXT: # kill: def $al killed $al killed $eax
224 ; SSE2-SSSE3-NEXT: retq
225 ;
226 ; AVX1-LABEL: v4i64:
227 ; AVX1: # %bb.0:
228 ; AVX1-NEXT: vextractf128 $1, %ymm1, %xmm2
229 ; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm3
230 ; AVX1-NEXT: vpcmpgtq %xmm2, %xmm3, %xmm2
231 ; AVX1-NEXT: vpcmpgtq %xmm1, %xmm0, %xmm0
232 ; AVX1-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm0
233 ; AVX1-NEXT: vmovmskpd %ymm0, %eax
234 ; AVX1-NEXT: # kill: def $al killed $al killed $eax
235 ; AVX1-NEXT: vzeroupper
236 ; AVX1-NEXT: retq
237 ;
238 ; AVX2-LABEL: v4i64:
239 ; AVX2: # %bb.0:
240 ; AVX2-NEXT: vpcmpgtq %ymm1, %ymm0, %ymm0
241 ; AVX2-NEXT: vmovmskpd %ymm0, %eax
242 ; AVX2-NEXT: # kill: def $al killed $al killed $eax
243 ; AVX2-NEXT: vzeroupper
244 ; AVX2-NEXT: retq
245 ;
246 ; AVX512F-LABEL: v4i64:
247 ; AVX512F: # %bb.0:
248 ; AVX512F-NEXT: vpcmpgtq %ymm1, %ymm0, %k0
249 ; AVX512F-NEXT: kmovw %k0, %eax
250 ; AVX512F-NEXT: # kill: def $al killed $al killed $eax
251 ; AVX512F-NEXT: vzeroupper
252 ; AVX512F-NEXT: retq
253 ;
254 ; AVX512BW-LABEL: v4i64:
255 ; AVX512BW: # %bb.0:
256 ; AVX512BW-NEXT: vpcmpgtq %ymm1, %ymm0, %k0
257 ; AVX512BW-NEXT: kmovd %k0, %eax
258 ; AVX512BW-NEXT: # kill: def $al killed $al killed $eax
259 ; AVX512BW-NEXT: vzeroupper
260 ; AVX512BW-NEXT: retq
261 %x = icmp sgt <4 x i64> %a, %b
262 %res = bitcast <4 x i1> %x to i4
263 ret i4 %res
264 }
266 define i4 @v4f64(<4 x double> %a, <4 x double> %b) {
267 ; SSE2-SSSE3-LABEL: v4f64:
268 ; SSE2-SSSE3: # %bb.0:
269 ; SSE2-SSSE3-NEXT: cmpltpd %xmm1, %xmm3
270 ; SSE2-SSSE3-NEXT: cmpltpd %xmm0, %xmm2
271 ; SSE2-SSSE3-NEXT: shufps {{.*#+}} xmm2 = xmm2[0,2],xmm3[0,2]
272 ; SSE2-SSSE3-NEXT: movmskps %xmm2, %eax
273 ; SSE2-SSSE3-NEXT: # kill: def $al killed $al killed $eax
274 ; SSE2-SSSE3-NEXT: retq
275 ;
276 ; AVX12-LABEL: v4f64:
277 ; AVX12: # %bb.0:
278 ; AVX12-NEXT: vcmpltpd %ymm0, %ymm1, %ymm0
279 ; AVX12-NEXT: vmovmskpd %ymm0, %eax
280 ; AVX12-NEXT: # kill: def $al killed $al killed $eax
281 ; AVX12-NEXT: vzeroupper
282 ; AVX12-NEXT: retq
283 ;
284 ; AVX512F-LABEL: v4f64:
285 ; AVX512F: # %bb.0:
286 ; AVX512F-NEXT: vcmpltpd %ymm0, %ymm1, %k0
287 ; AVX512F-NEXT: kmovw %k0, %eax
288 ; AVX512F-NEXT: # kill: def $al killed $al killed $eax
289 ; AVX512F-NEXT: vzeroupper
290 ; AVX512F-NEXT: retq
291 ;
292 ; AVX512BW-LABEL: v4f64:
293 ; AVX512BW: # %bb.0:
294 ; AVX512BW-NEXT: vcmpltpd %ymm0, %ymm1, %k0
295 ; AVX512BW-NEXT: kmovd %k0, %eax
296 ; AVX512BW-NEXT: # kill: def $al killed $al killed $eax
297 ; AVX512BW-NEXT: vzeroupper
298 ; AVX512BW-NEXT: retq
299 %x = fcmp ogt <4 x double> %a, %b
300 %res = bitcast <4 x i1> %x to i4
301 ret i4 %res
302 }
304 define void @bitcast_32i8_store(ptr %p, <32 x i8> %a0) {
305 ; SSE2-SSSE3-LABEL: bitcast_32i8_store:
306 ; SSE2-SSSE3: # %bb.0:
307 ; SSE2-SSSE3-NEXT: pmovmskb %xmm0, %eax
308 ; SSE2-SSSE3-NEXT: pmovmskb %xmm1, %ecx
309 ; SSE2-SSSE3-NEXT: shll $16, %ecx
310 ; SSE2-SSSE3-NEXT: orl %eax, %ecx
311 ; SSE2-SSSE3-NEXT: movl %ecx, (%rdi)
312 ; SSE2-SSSE3-NEXT: retq
313 ;
314 ; AVX1-LABEL: bitcast_32i8_store:
315 ; AVX1: # %bb.0:
316 ; AVX1-NEXT: vpmovmskb %xmm0, %eax
317 ; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm0
318 ; AVX1-NEXT: vpmovmskb %xmm0, %ecx
319 ; AVX1-NEXT: shll $16, %ecx
320 ; AVX1-NEXT: orl %eax, %ecx
321 ; AVX1-NEXT: movl %ecx, (%rdi)
322 ; AVX1-NEXT: vzeroupper
323 ; AVX1-NEXT: retq
324 ;
325 ; AVX2-LABEL: bitcast_32i8_store:
326 ; AVX2: # %bb.0:
327 ; AVX2-NEXT: vpmovmskb %ymm0, %eax
328 ; AVX2-NEXT: movl %eax, (%rdi)
329 ; AVX2-NEXT: vzeroupper
330 ; AVX2-NEXT: retq
331 ;
332 ; AVX512F-LABEL: bitcast_32i8_store:
333 ; AVX512F: # %bb.0:
334 ; AVX512F-NEXT: vpxor %xmm1, %xmm1, %xmm1
335 ; AVX512F-NEXT: vpcmpgtb %ymm0, %ymm1, %ymm0
336 ; AVX512F-NEXT: vpmovsxbd %xmm0, %zmm1
337 ; AVX512F-NEXT: vptestmd %zmm1, %zmm1, %k0
338 ; AVX512F-NEXT: vextracti128 $1, %ymm0, %xmm0
339 ; AVX512F-NEXT: vpmovsxbd %xmm0, %zmm0
340 ; AVX512F-NEXT: vptestmd %zmm0, %zmm0, %k1
341 ; AVX512F-NEXT: kmovw %k1, 2(%rdi)
342 ; AVX512F-NEXT: kmovw %k0, (%rdi)
343 ; AVX512F-NEXT: vzeroupper
344 ; AVX512F-NEXT: retq
345 ;
346 ; AVX512BW-LABEL: bitcast_32i8_store:
347 ; AVX512BW: # %bb.0:
348 ; AVX512BW-NEXT: vpmovb2m %ymm0, %k0
349 ; AVX512BW-NEXT: kmovd %k0, (%rdi)
350 ; AVX512BW-NEXT: vzeroupper
351 ; AVX512BW-NEXT: retq
352 %a1 = icmp slt <32 x i8> %a0, zeroinitializer
353 %a2 = bitcast <32 x i1> %a1 to i32
354 store i32 %a2, ptr %p
355 ret void
356 }
358 define void @bitcast_16i16_store(ptr %p, <16 x i16> %a0) {
359 ; SSE2-SSSE3-LABEL: bitcast_16i16_store:
360 ; SSE2-SSSE3: # %bb.0:
361 ; SSE2-SSSE3-NEXT: packsswb %xmm1, %xmm0
362 ; SSE2-SSSE3-NEXT: pmovmskb %xmm0, %eax
363 ; SSE2-SSSE3-NEXT: movw %ax, (%rdi)
364 ; SSE2-SSSE3-NEXT: retq
365 ;
366 ; AVX1-LABEL: bitcast_16i16_store:
367 ; AVX1: # %bb.0:
368 ; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm1
369 ; AVX1-NEXT: vpacksswb %xmm1, %xmm0, %xmm0
370 ; AVX1-NEXT: vpmovmskb %xmm0, %eax
371 ; AVX1-NEXT: movw %ax, (%rdi)
372 ; AVX1-NEXT: vzeroupper
373 ; AVX1-NEXT: retq
374 ;
375 ; AVX2-LABEL: bitcast_16i16_store:
376 ; AVX2: # %bb.0:
377 ; AVX2-NEXT: vextracti128 $1, %ymm0, %xmm1
378 ; AVX2-NEXT: vpacksswb %xmm1, %xmm0, %xmm0
379 ; AVX2-NEXT: vpmovmskb %xmm0, %eax
380 ; AVX2-NEXT: movw %ax, (%rdi)
381 ; AVX2-NEXT: vzeroupper
382 ; AVX2-NEXT: retq
383 ;
384 ; AVX512F-LABEL: bitcast_16i16_store:
385 ; AVX512F: # %bb.0:
386 ; AVX512F-NEXT: vpxor %xmm1, %xmm1, %xmm1
387 ; AVX512F-NEXT: vpcmpgtw %ymm0, %ymm1, %ymm0
388 ; AVX512F-NEXT: vpmovsxwd %ymm0, %zmm0
389 ; AVX512F-NEXT: vptestmd %zmm0, %zmm0, %k0
390 ; AVX512F-NEXT: kmovw %k0, (%rdi)
391 ; AVX512F-NEXT: vzeroupper
392 ; AVX512F-NEXT: retq
393 ;
394 ; AVX512BW-LABEL: bitcast_16i16_store:
395 ; AVX512BW: # %bb.0:
396 ; AVX512BW-NEXT: vpmovw2m %ymm0, %k0
397 ; AVX512BW-NEXT: kmovw %k0, (%rdi)
398 ; AVX512BW-NEXT: vzeroupper
399 ; AVX512BW-NEXT: retq
400 %a1 = icmp slt <16 x i16> %a0, zeroinitializer
401 %a2 = bitcast <16 x i1> %a1 to i16
402 store i16 %a2, ptr %p
403 ret void
404 }
406 define void @bitcast_8i32_store(ptr %p, <8 x i32> %a0) {
407 ; SSE2-SSSE3-LABEL: bitcast_8i32_store:
408 ; SSE2-SSSE3: # %bb.0:
409 ; SSE2-SSSE3-NEXT: packssdw %xmm1, %xmm0
410 ; SSE2-SSSE3-NEXT: packsswb %xmm0, %xmm0
411 ; SSE2-SSSE3-NEXT: pmovmskb %xmm0, %eax
412 ; SSE2-SSSE3-NEXT: movb %al, (%rdi)
413 ; SSE2-SSSE3-NEXT: retq
414 ;
415 ; AVX12-LABEL: bitcast_8i32_store:
416 ; AVX12: # %bb.0:
417 ; AVX12-NEXT: vmovmskps %ymm0, %eax
418 ; AVX12-NEXT: movb %al, (%rdi)
419 ; AVX12-NEXT: vzeroupper
420 ; AVX12-NEXT: retq
421 ;
422 ; AVX512-LABEL: bitcast_8i32_store:
423 ; AVX512: # %bb.0:
424 ; AVX512-NEXT: vmovmskps %ymm0, %eax
425 ; AVX512-NEXT: movb %al, (%rdi)
426 ; AVX512-NEXT: vzeroupper
427 ; AVX512-NEXT: retq
428 %a1 = icmp slt <8 x i32> %a0, zeroinitializer
429 %a2 = bitcast <8 x i1> %a1 to i8
430 store i8 %a2, ptr %p
431 ret void
432 }
434 define void @bitcast_4i64_store(ptr %p, <4 x i64> %a0) {
435 ; SSE2-SSSE3-LABEL: bitcast_4i64_store:
436 ; SSE2-SSSE3: # %bb.0:
437 ; SSE2-SSSE3-NEXT: shufps {{.*#+}} xmm0 = xmm0[1,3],xmm1[1,3]
438 ; SSE2-SSSE3-NEXT: movmskps %xmm0, %eax
439 ; SSE2-SSSE3-NEXT: movb %al, (%rdi)
440 ; SSE2-SSSE3-NEXT: retq
441 ;
442 ; AVX12-LABEL: bitcast_4i64_store:
443 ; AVX12: # %bb.0:
444 ; AVX12-NEXT: vmovmskpd %ymm0, %eax
445 ; AVX12-NEXT: movb %al, (%rdi)
446 ; AVX12-NEXT: vzeroupper
447 ; AVX12-NEXT: retq
448 ;
449 ; AVX512F-LABEL: bitcast_4i64_store:
450 ; AVX512F: # %bb.0:
451 ; AVX512F-NEXT: vpxor %xmm1, %xmm1, %xmm1
452 ; AVX512F-NEXT: vpcmpgtq %ymm0, %ymm1, %k0
453 ; AVX512F-NEXT: kmovw %k0, %eax
454 ; AVX512F-NEXT: movb %al, (%rdi)
455 ; AVX512F-NEXT: vzeroupper
456 ; AVX512F-NEXT: retq
457 ;
458 ; AVX512BW-LABEL: bitcast_4i64_store:
459 ; AVX512BW: # %bb.0:
460 ; AVX512BW-NEXT: vpxor %xmm1, %xmm1, %xmm1
461 ; AVX512BW-NEXT: vpcmpgtq %ymm0, %ymm1, %k0
462 ; AVX512BW-NEXT: kmovd %k0, %eax
463 ; AVX512BW-NEXT: movb %al, (%rdi)
464 ; AVX512BW-NEXT: vzeroupper
465 ; AVX512BW-NEXT: retq
466 %a1 = icmp slt <4 x i64> %a0, zeroinitializer
467 %a2 = bitcast <4 x i1> %a1 to i4
468 store i4 %a2, ptr %p
469 ret void
470 }