| blob | 6269cd4df1119f2d7177869f0d1e39f6318d6588 |
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-prefixes=SSE,SSE2
3 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+sse4.1 | FileCheck %s --check-prefixes=SSE,SSE41
4 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx | FileCheck %s --check-prefixes=AVX,AVX1
5 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx2 | FileCheck %s --check-prefixes=AVX,AVX2
6 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512f | FileCheck %s --check-prefixes=AVX512,AVX512F
7 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512f,+avx512bw | FileCheck %s --check-prefixes=AVX512,AVX512BW
8 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512f,+avx512bw,+avx512vl | FileCheck %s --check-prefixes=AVX512,AVX512VL
10 ;
11 ; Truncate
12 ;
14 define i1 @trunc_v2i64_v2i1(<2 x i64>) {
15 ; SSE-LABEL: trunc_v2i64_v2i1:
16 ; SSE: # %bb.0:
17 ; SSE-NEXT: psllq $63, %xmm0
18 ; SSE-NEXT: movmskpd %xmm0, %eax
19 ; SSE-NEXT: testl %eax, %eax
20 ; SSE-NEXT: setne %al
21 ; SSE-NEXT: retq
22 ;
23 ; AVX-LABEL: trunc_v2i64_v2i1:
24 ; AVX: # %bb.0:
25 ; AVX-NEXT: vpsllq $63, %xmm0, %xmm0
26 ; AVX-NEXT: vmovmskpd %xmm0, %eax
27 ; AVX-NEXT: testl %eax, %eax
28 ; AVX-NEXT: setne %al
29 ; AVX-NEXT: retq
30 ;
31 ; AVX512F-LABEL: trunc_v2i64_v2i1:
32 ; AVX512F: # %bb.0:
33 ; AVX512F-NEXT: vpsllq $63, %xmm0, %xmm0
34 ; AVX512F-NEXT: vptestmq %zmm0, %zmm0, %k0
35 ; AVX512F-NEXT: kmovw %k0, %eax
36 ; AVX512F-NEXT: testb $3, %al
37 ; AVX512F-NEXT: setne %al
38 ; AVX512F-NEXT: vzeroupper
39 ; AVX512F-NEXT: retq
40 ;
41 ; AVX512BW-LABEL: trunc_v2i64_v2i1:
42 ; AVX512BW: # %bb.0:
43 ; AVX512BW-NEXT: vpsllq $63, %xmm0, %xmm0
44 ; AVX512BW-NEXT: vptestmq %zmm0, %zmm0, %k0
45 ; AVX512BW-NEXT: kmovd %k0, %eax
46 ; AVX512BW-NEXT: testb $3, %al
47 ; AVX512BW-NEXT: setne %al
48 ; AVX512BW-NEXT: vzeroupper
49 ; AVX512BW-NEXT: retq
50 ;
51 ; AVX512VL-LABEL: trunc_v2i64_v2i1:
52 ; AVX512VL: # %bb.0:
53 ; AVX512VL-NEXT: vpsllq $63, %xmm0, %xmm0
54 ; AVX512VL-NEXT: vptestmq %xmm0, %xmm0, %k0
55 ; AVX512VL-NEXT: kmovd %k0, %eax
56 ; AVX512VL-NEXT: testb %al, %al
57 ; AVX512VL-NEXT: setne %al
58 ; AVX512VL-NEXT: retq
59 %a = trunc <2 x i64> %0 to <2 x i1>
60 %b = call i1 @llvm.vector.reduce.or.v2i1(<2 x i1> %a)
61 ret i1 %b
62 }
64 define i1 @trunc_v4i32_v4i1(<4 x i32>) {
65 ; SSE-LABEL: trunc_v4i32_v4i1:
66 ; SSE: # %bb.0:
67 ; SSE-NEXT: pslld $31, %xmm0
68 ; SSE-NEXT: movmskps %xmm0, %eax
69 ; SSE-NEXT: testl %eax, %eax
70 ; SSE-NEXT: setne %al
71 ; SSE-NEXT: retq
72 ;
73 ; AVX-LABEL: trunc_v4i32_v4i1:
74 ; AVX: # %bb.0:
75 ; AVX-NEXT: vpslld $31, %xmm0, %xmm0
76 ; AVX-NEXT: vmovmskps %xmm0, %eax
77 ; AVX-NEXT: testl %eax, %eax
78 ; AVX-NEXT: setne %al
79 ; AVX-NEXT: retq
80 ;
81 ; AVX512F-LABEL: trunc_v4i32_v4i1:
82 ; AVX512F: # %bb.0:
83 ; AVX512F-NEXT: vpslld $31, %xmm0, %xmm0
84 ; AVX512F-NEXT: vptestmd %zmm0, %zmm0, %k0
85 ; AVX512F-NEXT: kmovw %k0, %eax
86 ; AVX512F-NEXT: testb $15, %al
87 ; AVX512F-NEXT: setne %al
88 ; AVX512F-NEXT: vzeroupper
89 ; AVX512F-NEXT: retq
90 ;
91 ; AVX512BW-LABEL: trunc_v4i32_v4i1:
92 ; AVX512BW: # %bb.0:
93 ; AVX512BW-NEXT: vpslld $31, %xmm0, %xmm0
94 ; AVX512BW-NEXT: vptestmd %zmm0, %zmm0, %k0
95 ; AVX512BW-NEXT: kmovd %k0, %eax
96 ; AVX512BW-NEXT: testb $15, %al
97 ; AVX512BW-NEXT: setne %al
98 ; AVX512BW-NEXT: vzeroupper
99 ; AVX512BW-NEXT: retq
100 ;
101 ; AVX512VL-LABEL: trunc_v4i32_v4i1:
102 ; AVX512VL: # %bb.0:
103 ; AVX512VL-NEXT: vpslld $31, %xmm0, %xmm0
104 ; AVX512VL-NEXT: vptestmd %xmm0, %xmm0, %k0
105 ; AVX512VL-NEXT: kmovd %k0, %eax
106 ; AVX512VL-NEXT: testb %al, %al
107 ; AVX512VL-NEXT: setne %al
108 ; AVX512VL-NEXT: retq
109 %a = trunc <4 x i32> %0 to <4 x i1>
110 %b = call i1 @llvm.vector.reduce.or.v4i1(<4 x i1> %a)
111 ret i1 %b
112 }
114 define i1 @trunc_v8i16_v8i1(<8 x i8>) {
115 ; SSE2-LABEL: trunc_v8i16_v8i1:
116 ; SSE2: # %bb.0:
117 ; SSE2-NEXT: punpcklbw {{.*#+}} xmm0 = xmm0[0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7]
118 ; SSE2-NEXT: psllw $15, %xmm0
119 ; SSE2-NEXT: pmovmskb %xmm0, %eax
120 ; SSE2-NEXT: testl $43690, %eax # imm = 0xAAAA
121 ; SSE2-NEXT: setne %al
122 ; SSE2-NEXT: retq
123 ;
124 ; SSE41-LABEL: trunc_v8i16_v8i1:
125 ; SSE41: # %bb.0:
126 ; SSE41-NEXT: pmovzxbw {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero
127 ; SSE41-NEXT: psllw $15, %xmm0
128 ; SSE41-NEXT: pmovmskb %xmm0, %eax
129 ; SSE41-NEXT: testl $43690, %eax # imm = 0xAAAA
130 ; SSE41-NEXT: setne %al
131 ; SSE41-NEXT: retq
132 ;
133 ; AVX-LABEL: trunc_v8i16_v8i1:
134 ; AVX: # %bb.0:
135 ; AVX-NEXT: vpmovzxbw {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero
136 ; AVX-NEXT: vpsllw $15, %xmm0, %xmm0
137 ; AVX-NEXT: vpmovmskb %xmm0, %eax
138 ; AVX-NEXT: testl $43690, %eax # imm = 0xAAAA
139 ; AVX-NEXT: setne %al
140 ; AVX-NEXT: retq
141 ;
142 ; AVX512F-LABEL: trunc_v8i16_v8i1:
143 ; AVX512F: # %bb.0:
144 ; AVX512F-NEXT: vpmovsxbd %xmm0, %zmm0
145 ; AVX512F-NEXT: vpslld $31, %zmm0, %zmm0
146 ; AVX512F-NEXT: vptestmd %zmm0, %zmm0, %k0
147 ; AVX512F-NEXT: kmovw %k0, %eax
148 ; AVX512F-NEXT: testb %al, %al
149 ; AVX512F-NEXT: setne %al
150 ; AVX512F-NEXT: vzeroupper
151 ; AVX512F-NEXT: retq
152 ;
153 ; AVX512BW-LABEL: trunc_v8i16_v8i1:
154 ; AVX512BW: # %bb.0:
155 ; AVX512BW-NEXT: vpsllw $7, %xmm0, %xmm0
156 ; AVX512BW-NEXT: vpmovb2m %zmm0, %k0
157 ; AVX512BW-NEXT: kmovd %k0, %eax
158 ; AVX512BW-NEXT: testb %al, %al
159 ; AVX512BW-NEXT: setne %al
160 ; AVX512BW-NEXT: vzeroupper
161 ; AVX512BW-NEXT: retq
162 ;
163 ; AVX512VL-LABEL: trunc_v8i16_v8i1:
164 ; AVX512VL: # %bb.0:
165 ; AVX512VL-NEXT: vpsllw $7, %xmm0, %xmm0
166 ; AVX512VL-NEXT: vpmovb2m %xmm0, %k0
167 ; AVX512VL-NEXT: kmovd %k0, %eax
168 ; AVX512VL-NEXT: testb %al, %al
169 ; AVX512VL-NEXT: setne %al
170 ; AVX512VL-NEXT: retq
171 %a = trunc <8 x i8> %0 to <8 x i1>
172 %b = call i1 @llvm.vector.reduce.or.v8i1(<8 x i1> %a)
173 ret i1 %b
174 }
176 define i1 @trunc_v16i8_v16i1(<16 x i8>) {
177 ; SSE-LABEL: trunc_v16i8_v16i1:
178 ; SSE: # %bb.0:
179 ; SSE-NEXT: psllw $7, %xmm0
180 ; SSE-NEXT: pmovmskb %xmm0, %eax
181 ; SSE-NEXT: testl %eax, %eax
182 ; SSE-NEXT: setne %al
183 ; SSE-NEXT: retq
184 ;
185 ; AVX-LABEL: trunc_v16i8_v16i1:
186 ; AVX: # %bb.0:
187 ; AVX-NEXT: vpsllw $7, %xmm0, %xmm0
188 ; AVX-NEXT: vpmovmskb %xmm0, %eax
189 ; AVX-NEXT: testl %eax, %eax
190 ; AVX-NEXT: setne %al
191 ; AVX-NEXT: retq
192 ;
193 ; AVX512-LABEL: trunc_v16i8_v16i1:
194 ; AVX512: # %bb.0:
195 ; AVX512-NEXT: vpsllw $7, %xmm0, %xmm0
196 ; AVX512-NEXT: vpmovmskb %xmm0, %eax
197 ; AVX512-NEXT: testl %eax, %eax
198 ; AVX512-NEXT: setne %al
199 ; AVX512-NEXT: retq
200 %a = trunc <16 x i8> %0 to <16 x i1>
201 %b = call i1 @llvm.vector.reduce.or.v16i1(<16 x i1> %a)
202 ret i1 %b
203 }
205 define i1 @trunc_v4i64_v4i1(<4 x i64>) {
206 ; SSE-LABEL: trunc_v4i64_v4i1:
207 ; SSE: # %bb.0:
208 ; SSE-NEXT: shufps {{.*#+}} xmm0 = xmm0[0,2],xmm1[0,2]
209 ; SSE-NEXT: pslld $31, %xmm0
210 ; SSE-NEXT: movmskps %xmm0, %eax
211 ; SSE-NEXT: testl %eax, %eax
212 ; SSE-NEXT: setne %al
213 ; SSE-NEXT: retq
214 ;
215 ; AVX1-LABEL: trunc_v4i64_v4i1:
216 ; AVX1: # %bb.0:
217 ; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm1
218 ; AVX1-NEXT: vshufps {{.*#+}} xmm0 = xmm0[0,2],xmm1[0,2]
219 ; AVX1-NEXT: vpslld $31, %xmm0, %xmm0
220 ; AVX1-NEXT: vmovmskps %xmm0, %eax
221 ; AVX1-NEXT: testl %eax, %eax
222 ; AVX1-NEXT: setne %al
223 ; AVX1-NEXT: vzeroupper
224 ; AVX1-NEXT: retq
225 ;
226 ; AVX2-LABEL: trunc_v4i64_v4i1:
227 ; AVX2: # %bb.0:
228 ; AVX2-NEXT: vpsllq $63, %ymm0, %ymm0
229 ; AVX2-NEXT: vmovmskpd %ymm0, %eax
230 ; AVX2-NEXT: testl %eax, %eax
231 ; AVX2-NEXT: setne %al
232 ; AVX2-NEXT: vzeroupper
233 ; AVX2-NEXT: retq
234 ;
235 ; AVX512F-LABEL: trunc_v4i64_v4i1:
236 ; AVX512F: # %bb.0:
237 ; AVX512F-NEXT: vpsllq $63, %ymm0, %ymm0
238 ; AVX512F-NEXT: vptestmq %zmm0, %zmm0, %k0
239 ; AVX512F-NEXT: kmovw %k0, %eax
240 ; AVX512F-NEXT: testb $15, %al
241 ; AVX512F-NEXT: setne %al
242 ; AVX512F-NEXT: vzeroupper
243 ; AVX512F-NEXT: retq
244 ;
245 ; AVX512BW-LABEL: trunc_v4i64_v4i1:
246 ; AVX512BW: # %bb.0:
247 ; AVX512BW-NEXT: vpsllq $63, %ymm0, %ymm0
248 ; AVX512BW-NEXT: vptestmq %zmm0, %zmm0, %k0
249 ; AVX512BW-NEXT: kmovd %k0, %eax
250 ; AVX512BW-NEXT: testb $15, %al
251 ; AVX512BW-NEXT: setne %al
252 ; AVX512BW-NEXT: vzeroupper
253 ; AVX512BW-NEXT: retq
254 ;
255 ; AVX512VL-LABEL: trunc_v4i64_v4i1:
256 ; AVX512VL: # %bb.0:
257 ; AVX512VL-NEXT: vpsllq $63, %ymm0, %ymm0
258 ; AVX512VL-NEXT: vptestmq %ymm0, %ymm0, %k0
259 ; AVX512VL-NEXT: kmovd %k0, %eax
260 ; AVX512VL-NEXT: testb %al, %al
261 ; AVX512VL-NEXT: setne %al
262 ; AVX512VL-NEXT: vzeroupper
263 ; AVX512VL-NEXT: retq
264 %a = trunc <4 x i64> %0 to <4 x i1>
265 %b = call i1 @llvm.vector.reduce.or.v4i1(<4 x i1> %a)
266 ret i1 %b
267 }
269 define i1 @trunc_v8i32_v8i1(<8 x i32>) {
270 ; SSE2-LABEL: trunc_v8i32_v8i1:
271 ; SSE2: # %bb.0:
272 ; SSE2-NEXT: pslld $16, %xmm1
273 ; SSE2-NEXT: psrad $16, %xmm1
274 ; SSE2-NEXT: pslld $16, %xmm0
275 ; SSE2-NEXT: psrad $16, %xmm0
276 ; SSE2-NEXT: packssdw %xmm1, %xmm0
277 ; SSE2-NEXT: psllw $15, %xmm0
278 ; SSE2-NEXT: pmovmskb %xmm0, %eax
279 ; SSE2-NEXT: testl $43690, %eax # imm = 0xAAAA
280 ; SSE2-NEXT: setne %al
281 ; SSE2-NEXT: retq
282 ;
283 ; SSE41-LABEL: trunc_v8i32_v8i1:
284 ; SSE41: # %bb.0:
285 ; SSE41-NEXT: pxor %xmm2, %xmm2
286 ; SSE41-NEXT: pblendw {{.*#+}} xmm1 = xmm1[0],xmm2[1],xmm1[2],xmm2[3],xmm1[4],xmm2[5],xmm1[6],xmm2[7]
287 ; SSE41-NEXT: pblendw {{.*#+}} xmm0 = xmm0[0],xmm2[1],xmm0[2],xmm2[3],xmm0[4],xmm2[5],xmm0[6],xmm2[7]
288 ; SSE41-NEXT: packusdw %xmm1, %xmm0
289 ; SSE41-NEXT: psllw $15, %xmm0
290 ; SSE41-NEXT: pmovmskb %xmm0, %eax
291 ; SSE41-NEXT: testl $43690, %eax # imm = 0xAAAA
292 ; SSE41-NEXT: setne %al
293 ; SSE41-NEXT: retq
294 ;
295 ; AVX1-LABEL: trunc_v8i32_v8i1:
296 ; AVX1: # %bb.0:
297 ; AVX1-NEXT: vpslld $31, %xmm0, %xmm1
298 ; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm0
299 ; AVX1-NEXT: vpslld $31, %xmm0, %xmm0
300 ; AVX1-NEXT: vinsertf128 $1, %xmm0, %ymm1, %ymm0
301 ; AVX1-NEXT: vmovmskps %ymm0, %eax
302 ; AVX1-NEXT: testl %eax, %eax
303 ; AVX1-NEXT: setne %al
304 ; AVX1-NEXT: vzeroupper
305 ; AVX1-NEXT: retq
306 ;
307 ; AVX2-LABEL: trunc_v8i32_v8i1:
308 ; AVX2: # %bb.0:
309 ; AVX2-NEXT: vpslld $31, %ymm0, %ymm0
310 ; AVX2-NEXT: vmovmskps %ymm0, %eax
311 ; AVX2-NEXT: testl %eax, %eax
312 ; AVX2-NEXT: setne %al
313 ; AVX2-NEXT: vzeroupper
314 ; AVX2-NEXT: retq
315 ;
316 ; AVX512F-LABEL: trunc_v8i32_v8i1:
317 ; AVX512F: # %bb.0:
318 ; AVX512F-NEXT: vpslld $31, %ymm0, %ymm0
319 ; AVX512F-NEXT: vptestmd %zmm0, %zmm0, %k0
320 ; AVX512F-NEXT: kmovw %k0, %eax
321 ; AVX512F-NEXT: testb %al, %al
322 ; AVX512F-NEXT: setne %al
323 ; AVX512F-NEXT: vzeroupper
324 ; AVX512F-NEXT: retq
325 ;
326 ; AVX512BW-LABEL: trunc_v8i32_v8i1:
327 ; AVX512BW: # %bb.0:
328 ; AVX512BW-NEXT: vpslld $31, %ymm0, %ymm0
329 ; AVX512BW-NEXT: vptestmd %zmm0, %zmm0, %k0
330 ; AVX512BW-NEXT: kmovd %k0, %eax
331 ; AVX512BW-NEXT: testb %al, %al
332 ; AVX512BW-NEXT: setne %al
333 ; AVX512BW-NEXT: vzeroupper
334 ; AVX512BW-NEXT: retq
335 ;
336 ; AVX512VL-LABEL: trunc_v8i32_v8i1:
337 ; AVX512VL: # %bb.0:
338 ; AVX512VL-NEXT: vpslld $31, %ymm0, %ymm0
339 ; AVX512VL-NEXT: vptestmd %ymm0, %ymm0, %k0
340 ; AVX512VL-NEXT: kmovd %k0, %eax
341 ; AVX512VL-NEXT: testb %al, %al
342 ; AVX512VL-NEXT: setne %al
343 ; AVX512VL-NEXT: vzeroupper
344 ; AVX512VL-NEXT: retq
345 %a = trunc <8 x i32> %0 to <8 x i1>
346 %b = call i1 @llvm.vector.reduce.or.v8i1(<8 x i1> %a)
347 ret i1 %b
348 }
350 define i1 @trunc_v16i16_v16i1(<16 x i16>) {
351 ; SSE-LABEL: trunc_v16i16_v16i1:
352 ; SSE: # %bb.0:
353 ; SSE-NEXT: movdqa {{.*#+}} xmm2 = [255,0,255,0,255,0,255,0,255,0,255,0,255,0,255,0]
354 ; SSE-NEXT: pand %xmm2, %xmm1
355 ; SSE-NEXT: pand %xmm2, %xmm0
356 ; SSE-NEXT: packuswb %xmm1, %xmm0
357 ; SSE-NEXT: psllw $7, %xmm0
358 ; SSE-NEXT: pmovmskb %xmm0, %eax
359 ; SSE-NEXT: testl %eax, %eax
360 ; SSE-NEXT: setne %al
361 ; SSE-NEXT: retq
362 ;
363 ; AVX1-LABEL: trunc_v16i16_v16i1:
364 ; AVX1: # %bb.0:
365 ; AVX1-NEXT: vandps {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm0, %ymm0
366 ; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm1
367 ; AVX1-NEXT: vpackuswb %xmm1, %xmm0, %xmm0
368 ; AVX1-NEXT: vpsllw $7, %xmm0, %xmm0
369 ; AVX1-NEXT: vpmovmskb %xmm0, %eax
370 ; AVX1-NEXT: testl %eax, %eax
371 ; AVX1-NEXT: setne %al
372 ; AVX1-NEXT: vzeroupper
373 ; AVX1-NEXT: retq
374 ;
375 ; AVX2-LABEL: trunc_v16i16_v16i1:
376 ; AVX2: # %bb.0:
377 ; AVX2-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm0, %ymm0
378 ; AVX2-NEXT: vextracti128 $1, %ymm0, %xmm1
379 ; AVX2-NEXT: vpackuswb %xmm1, %xmm0, %xmm0
380 ; AVX2-NEXT: vpsllw $7, %xmm0, %xmm0
381 ; AVX2-NEXT: vpmovmskb %xmm0, %eax
382 ; AVX2-NEXT: testl %eax, %eax
383 ; AVX2-NEXT: setne %al
384 ; AVX2-NEXT: vzeroupper
385 ; AVX2-NEXT: retq
386 ;
387 ; AVX512F-LABEL: trunc_v16i16_v16i1:
388 ; AVX512F: # %bb.0:
389 ; AVX512F-NEXT: vpmovsxwd %ymm0, %zmm0
390 ; AVX512F-NEXT: vpslld $31, %zmm0, %zmm0
391 ; AVX512F-NEXT: vptestmd %zmm0, %zmm0, %k0
392 ; AVX512F-NEXT: kortestw %k0, %k0
393 ; AVX512F-NEXT: setne %al
394 ; AVX512F-NEXT: vzeroupper
395 ; AVX512F-NEXT: retq
396 ;
397 ; AVX512BW-LABEL: trunc_v16i16_v16i1:
398 ; AVX512BW: # %bb.0:
399 ; AVX512BW-NEXT: vpsllw $15, %ymm0, %ymm0
400 ; AVX512BW-NEXT: vpmovw2m %zmm0, %k0
401 ; AVX512BW-NEXT: kortestw %k0, %k0
402 ; AVX512BW-NEXT: setne %al
403 ; AVX512BW-NEXT: vzeroupper
404 ; AVX512BW-NEXT: retq
405 ;
406 ; AVX512VL-LABEL: trunc_v16i16_v16i1:
407 ; AVX512VL: # %bb.0:
408 ; AVX512VL-NEXT: vpsllw $15, %ymm0, %ymm0
409 ; AVX512VL-NEXT: vpmovw2m %ymm0, %k0
410 ; AVX512VL-NEXT: kortestw %k0, %k0
411 ; AVX512VL-NEXT: setne %al
412 ; AVX512VL-NEXT: vzeroupper
413 ; AVX512VL-NEXT: retq
414 %a = trunc <16 x i16> %0 to <16 x i1>
415 %b = call i1 @llvm.vector.reduce.or.v16i1(<16 x i1> %a)
416 ret i1 %b
417 }
419 define i1 @trunc_v32i8_v32i1(<32 x i8>) {
420 ; SSE-LABEL: trunc_v32i8_v32i1:
421 ; SSE: # %bb.0:
422 ; SSE-NEXT: por %xmm1, %xmm0
423 ; SSE-NEXT: psllw $7, %xmm0
424 ; SSE-NEXT: pmovmskb %xmm0, %eax
425 ; SSE-NEXT: testl %eax, %eax
426 ; SSE-NEXT: setne %al
427 ; SSE-NEXT: retq
428 ;
429 ; AVX1-LABEL: trunc_v32i8_v32i1:
430 ; AVX1: # %bb.0:
431 ; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm1
432 ; AVX1-NEXT: vpor %xmm1, %xmm0, %xmm0
433 ; AVX1-NEXT: vpsllw $7, %xmm0, %xmm0
434 ; AVX1-NEXT: vpmovmskb %xmm0, %eax
435 ; AVX1-NEXT: testl %eax, %eax
436 ; AVX1-NEXT: setne %al
437 ; AVX1-NEXT: vzeroupper
438 ; AVX1-NEXT: retq
439 ;
440 ; AVX2-LABEL: trunc_v32i8_v32i1:
441 ; AVX2: # %bb.0:
442 ; AVX2-NEXT: vpsllw $7, %ymm0, %ymm0
443 ; AVX2-NEXT: vpmovmskb %ymm0, %eax
444 ; AVX2-NEXT: testl %eax, %eax
445 ; AVX2-NEXT: setne %al
446 ; AVX2-NEXT: vzeroupper
447 ; AVX2-NEXT: retq
448 ;
449 ; AVX512F-LABEL: trunc_v32i8_v32i1:
450 ; AVX512F: # %bb.0:
451 ; AVX512F-NEXT: vextracti128 $1, %ymm0, %xmm1
452 ; AVX512F-NEXT: vpor %xmm1, %xmm0, %xmm0
453 ; AVX512F-NEXT: vpmovsxbd %xmm0, %zmm0
454 ; AVX512F-NEXT: vpslld $31, %zmm0, %zmm0
455 ; AVX512F-NEXT: vptestmd %zmm0, %zmm0, %k0
456 ; AVX512F-NEXT: kshiftrw $8, %k0, %k1
457 ; AVX512F-NEXT: korw %k1, %k0, %k0
458 ; AVX512F-NEXT: kshiftrw $4, %k0, %k1
459 ; AVX512F-NEXT: korw %k1, %k0, %k0
460 ; AVX512F-NEXT: kshiftrw $2, %k0, %k1
461 ; AVX512F-NEXT: korw %k1, %k0, %k0
462 ; AVX512F-NEXT: kshiftrw $1, %k0, %k1
463 ; AVX512F-NEXT: korw %k1, %k0, %k0
464 ; AVX512F-NEXT: kmovw %k0, %eax
465 ; AVX512F-NEXT: # kill: def $al killed $al killed $eax
466 ; AVX512F-NEXT: vzeroupper
467 ; AVX512F-NEXT: retq
468 ;
469 ; AVX512BW-LABEL: trunc_v32i8_v32i1:
470 ; AVX512BW: # %bb.0:
471 ; AVX512BW-NEXT: vpsllw $7, %ymm0, %ymm0
472 ; AVX512BW-NEXT: vpmovmskb %ymm0, %eax
473 ; AVX512BW-NEXT: testl %eax, %eax
474 ; AVX512BW-NEXT: setne %al
475 ; AVX512BW-NEXT: vzeroupper
476 ; AVX512BW-NEXT: retq
477 ;
478 ; AVX512VL-LABEL: trunc_v32i8_v32i1:
479 ; AVX512VL: # %bb.0:
480 ; AVX512VL-NEXT: vpsllw $7, %ymm0, %ymm0
481 ; AVX512VL-NEXT: vpmovmskb %ymm0, %eax
482 ; AVX512VL-NEXT: testl %eax, %eax
483 ; AVX512VL-NEXT: setne %al
484 ; AVX512VL-NEXT: vzeroupper
485 ; AVX512VL-NEXT: retq
486 %a = trunc <32 x i8> %0 to <32 x i1>
487 %b = call i1 @llvm.vector.reduce.or.v32i1(<32 x i1> %a)
488 ret i1 %b
489 }
491 define i1 @trunc_v8i64_v8i1(<8 x i64>) {
492 ; SSE2-LABEL: trunc_v8i64_v8i1:
493 ; SSE2: # %bb.0:
494 ; SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm1[0,2,2,3]
495 ; SSE2-NEXT: pshuflw {{.*#+}} xmm1 = xmm1[0,2,2,3,4,5,6,7]
496 ; SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,2,2,3]
497 ; SSE2-NEXT: pshuflw {{.*#+}} xmm0 = xmm0[0,2,2,3,4,5,6,7]
498 ; SSE2-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
499 ; SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm3[0,2,2,3]
500 ; SSE2-NEXT: pshuflw {{.*#+}} xmm1 = xmm1[0,1,0,2,4,5,6,7]
501 ; SSE2-NEXT: pshufd {{.*#+}} xmm2 = xmm2[0,2,2,3]
502 ; SSE2-NEXT: pshuflw {{.*#+}} xmm2 = xmm2[0,1,0,2,4,5,6,7]
503 ; SSE2-NEXT: punpckldq {{.*#+}} xmm2 = xmm2[0],xmm1[0],xmm2[1],xmm1[1]
504 ; SSE2-NEXT: movsd {{.*#+}} xmm2 = xmm0[0],xmm2[1]
505 ; SSE2-NEXT: psllw $15, %xmm2
506 ; SSE2-NEXT: pmovmskb %xmm2, %eax
507 ; SSE2-NEXT: testl $43690, %eax # imm = 0xAAAA
508 ; SSE2-NEXT: setne %al
509 ; SSE2-NEXT: retq
510 ;
511 ; SSE41-LABEL: trunc_v8i64_v8i1:
512 ; SSE41: # %bb.0:
513 ; SSE41-NEXT: pxor %xmm4, %xmm4
514 ; SSE41-NEXT: pblendw {{.*#+}} xmm3 = xmm3[0],xmm4[1,2,3],xmm3[4],xmm4[5,6,7]
515 ; SSE41-NEXT: pblendw {{.*#+}} xmm2 = xmm2[0],xmm4[1,2,3],xmm2[4],xmm4[5,6,7]
516 ; SSE41-NEXT: packusdw %xmm3, %xmm2
517 ; SSE41-NEXT: pblendw {{.*#+}} xmm1 = xmm1[0],xmm4[1,2,3],xmm1[4],xmm4[5,6,7]
518 ; SSE41-NEXT: pblendw {{.*#+}} xmm0 = xmm0[0],xmm4[1,2,3],xmm0[4],xmm4[5,6,7]
519 ; SSE41-NEXT: packusdw %xmm1, %xmm0
520 ; SSE41-NEXT: packusdw %xmm2, %xmm0
521 ; SSE41-NEXT: psllw $15, %xmm0
522 ; SSE41-NEXT: pmovmskb %xmm0, %eax
523 ; SSE41-NEXT: testl $43690, %eax # imm = 0xAAAA
524 ; SSE41-NEXT: setne %al
525 ; SSE41-NEXT: retq
526 ;
527 ; AVX1-LABEL: trunc_v8i64_v8i1:
528 ; AVX1: # %bb.0:
529 ; AVX1-NEXT: vextractf128 $1, %ymm1, %xmm2
530 ; AVX1-NEXT: vshufps {{.*#+}} xmm1 = xmm1[0,2],xmm2[0,2]
531 ; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm2
532 ; AVX1-NEXT: vshufps {{.*#+}} xmm0 = xmm0[0,2],xmm2[0,2]
533 ; AVX1-NEXT: vpslld $31, %xmm0, %xmm0
534 ; AVX1-NEXT: vpslld $31, %xmm1, %xmm1
535 ; AVX1-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm0
536 ; AVX1-NEXT: vmovmskps %ymm0, %eax
537 ; AVX1-NEXT: testl %eax, %eax
538 ; AVX1-NEXT: setne %al
539 ; AVX1-NEXT: vzeroupper
540 ; AVX1-NEXT: retq
541 ;
542 ; AVX2-LABEL: trunc_v8i64_v8i1:
543 ; AVX2: # %bb.0:
544 ; AVX2-NEXT: vperm2f128 {{.*#+}} ymm2 = ymm0[2,3],ymm1[2,3]
545 ; AVX2-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm0
546 ; AVX2-NEXT: vshufps {{.*#+}} ymm0 = ymm0[0,2],ymm2[0,2],ymm0[4,6],ymm2[4,6]
547 ; AVX2-NEXT: vpslld $31, %ymm0, %ymm0
548 ; AVX2-NEXT: vmovmskps %ymm0, %eax
549 ; AVX2-NEXT: testl %eax, %eax
550 ; AVX2-NEXT: setne %al
551 ; AVX2-NEXT: vzeroupper
552 ; AVX2-NEXT: retq
553 ;
554 ; AVX512F-LABEL: trunc_v8i64_v8i1:
555 ; AVX512F: # %bb.0:
556 ; AVX512F-NEXT: vpsllq $63, %zmm0, %zmm0
557 ; AVX512F-NEXT: vptestmq %zmm0, %zmm0, %k0
558 ; AVX512F-NEXT: kmovw %k0, %eax
559 ; AVX512F-NEXT: testb %al, %al
560 ; AVX512F-NEXT: setne %al
561 ; AVX512F-NEXT: vzeroupper
562 ; AVX512F-NEXT: retq
563 ;
564 ; AVX512BW-LABEL: trunc_v8i64_v8i1:
565 ; AVX512BW: # %bb.0:
566 ; AVX512BW-NEXT: vpsllq $63, %zmm0, %zmm0
567 ; AVX512BW-NEXT: vptestmq %zmm0, %zmm0, %k0
568 ; AVX512BW-NEXT: kmovd %k0, %eax
569 ; AVX512BW-NEXT: testb %al, %al
570 ; AVX512BW-NEXT: setne %al
571 ; AVX512BW-NEXT: vzeroupper
572 ; AVX512BW-NEXT: retq
573 ;
574 ; AVX512VL-LABEL: trunc_v8i64_v8i1:
575 ; AVX512VL: # %bb.0:
576 ; AVX512VL-NEXT: vpsllq $63, %zmm0, %zmm0
577 ; AVX512VL-NEXT: vptestmq %zmm0, %zmm0, %k0
578 ; AVX512VL-NEXT: kmovd %k0, %eax
579 ; AVX512VL-NEXT: testb %al, %al
580 ; AVX512VL-NEXT: setne %al
581 ; AVX512VL-NEXT: vzeroupper
582 ; AVX512VL-NEXT: retq
583 %a = trunc <8 x i64> %0 to <8 x i1>
584 %b = call i1 @llvm.vector.reduce.or.v8i1(<8 x i1> %a)
585 ret i1 %b
586 }
588 define i1 @trunc_v16i32_v16i1(<16 x i32>) {
589 ; SSE2-LABEL: trunc_v16i32_v16i1:
590 ; SSE2: # %bb.0:
591 ; SSE2-NEXT: movdqa {{.*#+}} xmm4 = [255,0,0,0,255,0,0,0,255,0,0,0,255,0,0,0]
592 ; SSE2-NEXT: pand %xmm4, %xmm3
593 ; SSE2-NEXT: pand %xmm4, %xmm2
594 ; SSE2-NEXT: packuswb %xmm3, %xmm2
595 ; SSE2-NEXT: pand %xmm4, %xmm1
596 ; SSE2-NEXT: pand %xmm4, %xmm0
597 ; SSE2-NEXT: packuswb %xmm1, %xmm0
598 ; SSE2-NEXT: packuswb %xmm2, %xmm0
599 ; SSE2-NEXT: psllw $7, %xmm0
600 ; SSE2-NEXT: pmovmskb %xmm0, %eax
601 ; SSE2-NEXT: testl %eax, %eax
602 ; SSE2-NEXT: setne %al
603 ; SSE2-NEXT: retq
604 ;
605 ; SSE41-LABEL: trunc_v16i32_v16i1:
606 ; SSE41: # %bb.0:
607 ; SSE41-NEXT: movdqa {{.*#+}} xmm4 = [255,0,0,0,255,0,0,0,255,0,0,0,255,0,0,0]
608 ; SSE41-NEXT: pand %xmm4, %xmm3
609 ; SSE41-NEXT: pand %xmm4, %xmm2
610 ; SSE41-NEXT: packusdw %xmm3, %xmm2
611 ; SSE41-NEXT: pand %xmm4, %xmm1
612 ; SSE41-NEXT: pand %xmm4, %xmm0
613 ; SSE41-NEXT: packusdw %xmm1, %xmm0
614 ; SSE41-NEXT: packuswb %xmm2, %xmm0
615 ; SSE41-NEXT: psllw $7, %xmm0
616 ; SSE41-NEXT: pmovmskb %xmm0, %eax
617 ; SSE41-NEXT: testl %eax, %eax
618 ; SSE41-NEXT: setne %al
619 ; SSE41-NEXT: retq
620 ;
621 ; AVX1-LABEL: trunc_v16i32_v16i1:
622 ; AVX1: # %bb.0:
623 ; AVX1-NEXT: vmovaps {{.*#+}} ymm2 = [255,255,255,255,255,255,255,255]
624 ; AVX1-NEXT: vandps %ymm2, %ymm1, %ymm1
625 ; AVX1-NEXT: vextractf128 $1, %ymm1, %xmm3
626 ; AVX1-NEXT: vpackusdw %xmm3, %xmm1, %xmm1
627 ; AVX1-NEXT: vandps %ymm2, %ymm0, %ymm0
628 ; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm2
629 ; AVX1-NEXT: vpackusdw %xmm2, %xmm0, %xmm0
630 ; AVX1-NEXT: vpackuswb %xmm1, %xmm0, %xmm0
631 ; AVX1-NEXT: vpsllw $7, %xmm0, %xmm0
632 ; AVX1-NEXT: vpmovmskb %xmm0, %eax
633 ; AVX1-NEXT: testl %eax, %eax
634 ; AVX1-NEXT: setne %al
635 ; AVX1-NEXT: vzeroupper
636 ; AVX1-NEXT: retq
637 ;
638 ; AVX2-LABEL: trunc_v16i32_v16i1:
639 ; AVX2: # %bb.0:
640 ; AVX2-NEXT: vpbroadcastd {{.*#+}} ymm2 = [255,255,255,255,255,255,255,255]
641 ; AVX2-NEXT: vpand %ymm2, %ymm1, %ymm1
642 ; AVX2-NEXT: vpand %ymm2, %ymm0, %ymm0
643 ; AVX2-NEXT: vpackusdw %ymm1, %ymm0, %ymm0
644 ; AVX2-NEXT: vextracti128 $1, %ymm0, %xmm1
645 ; AVX2-NEXT: vpackuswb %xmm1, %xmm0, %xmm0
646 ; AVX2-NEXT: vpshufd {{.*#+}} xmm0 = xmm0[0,2,1,3]
647 ; AVX2-NEXT: vpsllw $7, %xmm0, %xmm0
648 ; AVX2-NEXT: vpmovmskb %xmm0, %eax
649 ; AVX2-NEXT: testl %eax, %eax
650 ; AVX2-NEXT: setne %al
651 ; AVX2-NEXT: vzeroupper
652 ; AVX2-NEXT: retq
653 ;
654 ; AVX512-LABEL: trunc_v16i32_v16i1:
655 ; AVX512: # %bb.0:
656 ; AVX512-NEXT: vpslld $31, %zmm0, %zmm0
657 ; AVX512-NEXT: vptestmd %zmm0, %zmm0, %k0
658 ; AVX512-NEXT: kortestw %k0, %k0
659 ; AVX512-NEXT: setne %al
660 ; AVX512-NEXT: vzeroupper
661 ; AVX512-NEXT: retq
662 %a = trunc <16 x i32> %0 to <16 x i1>
663 %b = call i1 @llvm.vector.reduce.or.v16i1(<16 x i1> %a)
664 ret i1 %b
665 }
667 define i1 @trunc_v32i16_v32i1(<32 x i16>) {
668 ; SSE-LABEL: trunc_v32i16_v32i1:
669 ; SSE: # %bb.0:
670 ; SSE-NEXT: movdqa {{.*#+}} xmm4 = [255,255,255,255,255,255,255,255]
671 ; SSE-NEXT: pand %xmm4, %xmm3
672 ; SSE-NEXT: pand %xmm4, %xmm2
673 ; SSE-NEXT: packuswb %xmm3, %xmm2
674 ; SSE-NEXT: pand %xmm4, %xmm1
675 ; SSE-NEXT: pand %xmm4, %xmm0
676 ; SSE-NEXT: packuswb %xmm1, %xmm0
677 ; SSE-NEXT: por %xmm2, %xmm0
678 ; SSE-NEXT: psllw $7, %xmm0
679 ; SSE-NEXT: pmovmskb %xmm0, %eax
680 ; SSE-NEXT: testl %eax, %eax
681 ; SSE-NEXT: setne %al
682 ; SSE-NEXT: retq
683 ;
684 ; AVX1-LABEL: trunc_v32i16_v32i1:
685 ; AVX1: # %bb.0:
686 ; AVX1-NEXT: vorps %ymm1, %ymm0, %ymm0
687 ; AVX1-NEXT: vandps {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm0, %ymm0
688 ; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm1
689 ; AVX1-NEXT: vpackuswb %xmm1, %xmm0, %xmm0
690 ; AVX1-NEXT: vpsllw $7, %xmm0, %xmm0
691 ; AVX1-NEXT: vpmovmskb %xmm0, %eax
692 ; AVX1-NEXT: testl %eax, %eax
693 ; AVX1-NEXT: setne %al
694 ; AVX1-NEXT: vzeroupper
695 ; AVX1-NEXT: retq
696 ;
697 ; AVX2-LABEL: trunc_v32i16_v32i1:
698 ; AVX2: # %bb.0:
699 ; AVX2-NEXT: vpbroadcastw {{.*#+}} ymm2 = [255,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255]
700 ; AVX2-NEXT: vpand %ymm2, %ymm1, %ymm1
701 ; AVX2-NEXT: vpand %ymm2, %ymm0, %ymm0
702 ; AVX2-NEXT: vpackuswb %ymm1, %ymm0, %ymm0
703 ; AVX2-NEXT: vpermq {{.*#+}} ymm0 = ymm0[0,2,1,3]
704 ; AVX2-NEXT: vpsllw $7, %ymm0, %ymm0
705 ; AVX2-NEXT: vpmovmskb %ymm0, %eax
706 ; AVX2-NEXT: testl %eax, %eax
707 ; AVX2-NEXT: setne %al
708 ; AVX2-NEXT: vzeroupper
709 ; AVX2-NEXT: retq
710 ;
711 ; AVX512F-LABEL: trunc_v32i16_v32i1:
712 ; AVX512F: # %bb.0:
713 ; AVX512F-NEXT: vextracti64x4 $1, %zmm0, %ymm1
714 ; AVX512F-NEXT: vpor %ymm1, %ymm0, %ymm0
715 ; AVX512F-NEXT: vpmovsxwd %ymm0, %zmm0
716 ; AVX512F-NEXT: vpslld $31, %zmm0, %zmm0
717 ; AVX512F-NEXT: vptestmd %zmm0, %zmm0, %k0
718 ; AVX512F-NEXT: kshiftrw $8, %k0, %k1
719 ; AVX512F-NEXT: korw %k1, %k0, %k0
720 ; AVX512F-NEXT: kshiftrw $4, %k0, %k1
721 ; AVX512F-NEXT: korw %k1, %k0, %k0
722 ; AVX512F-NEXT: kshiftrw $2, %k0, %k1
723 ; AVX512F-NEXT: korw %k1, %k0, %k0
724 ; AVX512F-NEXT: kshiftrw $1, %k0, %k1
725 ; AVX512F-NEXT: korw %k1, %k0, %k0
726 ; AVX512F-NEXT: kmovw %k0, %eax
727 ; AVX512F-NEXT: # kill: def $al killed $al killed $eax
728 ; AVX512F-NEXT: vzeroupper
729 ; AVX512F-NEXT: retq
730 ;
731 ; AVX512BW-LABEL: trunc_v32i16_v32i1:
732 ; AVX512BW: # %bb.0:
733 ; AVX512BW-NEXT: vpsllw $15, %zmm0, %zmm0
734 ; AVX512BW-NEXT: vpmovw2m %zmm0, %k0
735 ; AVX512BW-NEXT: kortestd %k0, %k0
736 ; AVX512BW-NEXT: setne %al
737 ; AVX512BW-NEXT: vzeroupper
738 ; AVX512BW-NEXT: retq
739 ;
740 ; AVX512VL-LABEL: trunc_v32i16_v32i1:
741 ; AVX512VL: # %bb.0:
742 ; AVX512VL-NEXT: vpsllw $15, %zmm0, %zmm0
743 ; AVX512VL-NEXT: vpmovw2m %zmm0, %k0
744 ; AVX512VL-NEXT: kortestd %k0, %k0
745 ; AVX512VL-NEXT: setne %al
746 ; AVX512VL-NEXT: vzeroupper
747 ; AVX512VL-NEXT: retq
748 %a = trunc <32 x i16> %0 to <32 x i1>
749 %b = call i1 @llvm.vector.reduce.or.v32i1(<32 x i1> %a)
750 ret i1 %b
751 }
753 define i1 @trunc_v64i8_v64i1(<64 x i8>) {
754 ; SSE-LABEL: trunc_v64i8_v64i1:
755 ; SSE: # %bb.0:
756 ; SSE-NEXT: por %xmm3, %xmm1
757 ; SSE-NEXT: por %xmm2, %xmm1
758 ; SSE-NEXT: por %xmm0, %xmm1
759 ; SSE-NEXT: psllw $7, %xmm1
760 ; SSE-NEXT: pmovmskb %xmm1, %eax
761 ; SSE-NEXT: testl %eax, %eax
762 ; SSE-NEXT: setne %al
763 ; SSE-NEXT: retq
764 ;
765 ; AVX1-LABEL: trunc_v64i8_v64i1:
766 ; AVX1: # %bb.0:
767 ; AVX1-NEXT: vorps %ymm1, %ymm0, %ymm0
768 ; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm1
769 ; AVX1-NEXT: vorps %xmm1, %xmm0, %xmm0
770 ; AVX1-NEXT: vpsllw $7, %xmm0, %xmm0
771 ; AVX1-NEXT: vpmovmskb %xmm0, %eax
772 ; AVX1-NEXT: testl %eax, %eax
773 ; AVX1-NEXT: setne %al
774 ; AVX1-NEXT: vzeroupper
775 ; AVX1-NEXT: retq
776 ;
777 ; AVX2-LABEL: trunc_v64i8_v64i1:
778 ; AVX2: # %bb.0:
779 ; AVX2-NEXT: vpor %ymm1, %ymm0, %ymm0
780 ; AVX2-NEXT: vpsllw $7, %ymm0, %ymm0
781 ; AVX2-NEXT: vpmovmskb %ymm0, %eax
782 ; AVX2-NEXT: testl %eax, %eax
783 ; AVX2-NEXT: setne %al
784 ; AVX2-NEXT: vzeroupper
785 ; AVX2-NEXT: retq
786 ;
787 ; AVX512F-LABEL: trunc_v64i8_v64i1:
788 ; AVX512F: # %bb.0:
789 ; AVX512F-NEXT: vextracti64x4 $1, %zmm0, %ymm1
790 ; AVX512F-NEXT: vextracti128 $1, %ymm1, %xmm2
791 ; AVX512F-NEXT: vextracti128 $1, %ymm0, %xmm3
792 ; AVX512F-NEXT: vpor %xmm2, %xmm3, %xmm2
793 ; AVX512F-NEXT: vpor %xmm2, %xmm1, %xmm1
794 ; AVX512F-NEXT: vpor %xmm1, %xmm0, %xmm0
795 ; AVX512F-NEXT: vpmovsxbd %xmm0, %zmm0
796 ; AVX512F-NEXT: vpslld $31, %zmm0, %zmm0
797 ; AVX512F-NEXT: vptestmd %zmm0, %zmm0, %k0
798 ; AVX512F-NEXT: kshiftrw $8, %k0, %k1
799 ; AVX512F-NEXT: korw %k1, %k0, %k0
800 ; AVX512F-NEXT: kshiftrw $4, %k0, %k1
801 ; AVX512F-NEXT: korw %k1, %k0, %k0
802 ; AVX512F-NEXT: kshiftrw $2, %k0, %k1
803 ; AVX512F-NEXT: korw %k1, %k0, %k0
804 ; AVX512F-NEXT: kshiftrw $1, %k0, %k1
805 ; AVX512F-NEXT: korw %k1, %k0, %k0
806 ; AVX512F-NEXT: kmovw %k0, %eax
807 ; AVX512F-NEXT: # kill: def $al killed $al killed $eax
808 ; AVX512F-NEXT: vzeroupper
809 ; AVX512F-NEXT: retq
810 ;
811 ; AVX512BW-LABEL: trunc_v64i8_v64i1:
812 ; AVX512BW: # %bb.0:
813 ; AVX512BW-NEXT: vpsllw $7, %zmm0, %zmm0
814 ; AVX512BW-NEXT: vpmovb2m %zmm0, %k0
815 ; AVX512BW-NEXT: kortestq %k0, %k0
816 ; AVX512BW-NEXT: setne %al
817 ; AVX512BW-NEXT: vzeroupper
818 ; AVX512BW-NEXT: retq
819 ;
820 ; AVX512VL-LABEL: trunc_v64i8_v64i1:
821 ; AVX512VL: # %bb.0:
822 ; AVX512VL-NEXT: vpsllw $7, %zmm0, %zmm0
823 ; AVX512VL-NEXT: vpmovb2m %zmm0, %k0
824 ; AVX512VL-NEXT: kortestq %k0, %k0
825 ; AVX512VL-NEXT: setne %al
826 ; AVX512VL-NEXT: vzeroupper
827 ; AVX512VL-NEXT: retq
828 %a = trunc <64 x i8> %0 to <64 x i1>
829 %b = call i1 @llvm.vector.reduce.or.v64i1(<64 x i1> %a)
830 ret i1 %b
831 }
833 ;
834 ; Comparison
835 ;
837 define i1 @icmp_v2i64_v2i1(<2 x i64>) {
838 ; SSE2-LABEL: icmp_v2i64_v2i1:
839 ; SSE2: # %bb.0:
840 ; SSE2-NEXT: pxor %xmm1, %xmm1
841 ; SSE2-NEXT: pcmpeqd %xmm0, %xmm1
842 ; SSE2-NEXT: pshufd {{.*#+}} xmm0 = xmm1[1,0,3,2]
843 ; SSE2-NEXT: pand %xmm1, %xmm0
844 ; SSE2-NEXT: movmskpd %xmm0, %eax
845 ; SSE2-NEXT: testl %eax, %eax
846 ; SSE2-NEXT: setne %al
847 ; SSE2-NEXT: retq
848 ;
849 ; SSE41-LABEL: icmp_v2i64_v2i1:
850 ; SSE41: # %bb.0:
851 ; SSE41-NEXT: pxor %xmm1, %xmm1
852 ; SSE41-NEXT: pcmpeqq %xmm0, %xmm1
853 ; SSE41-NEXT: movmskpd %xmm1, %eax
854 ; SSE41-NEXT: testl %eax, %eax
855 ; SSE41-NEXT: setne %al
856 ; SSE41-NEXT: retq
857 ;
858 ; AVX-LABEL: icmp_v2i64_v2i1:
859 ; AVX: # %bb.0:
860 ; AVX-NEXT: vpxor %xmm1, %xmm1, %xmm1
861 ; AVX-NEXT: vpcmpeqq %xmm1, %xmm0, %xmm0
862 ; AVX-NEXT: vmovmskpd %xmm0, %eax
863 ; AVX-NEXT: testl %eax, %eax
864 ; AVX-NEXT: setne %al
865 ; AVX-NEXT: retq
866 ;
867 ; AVX512F-LABEL: icmp_v2i64_v2i1:
868 ; AVX512F: # %bb.0:
869 ; AVX512F-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
870 ; AVX512F-NEXT: vptestnmq %zmm0, %zmm0, %k0
871 ; AVX512F-NEXT: kmovw %k0, %eax
872 ; AVX512F-NEXT: testb $3, %al
873 ; AVX512F-NEXT: setne %al
874 ; AVX512F-NEXT: vzeroupper
875 ; AVX512F-NEXT: retq
876 ;
877 ; AVX512BW-LABEL: icmp_v2i64_v2i1:
878 ; AVX512BW: # %bb.0:
879 ; AVX512BW-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
880 ; AVX512BW-NEXT: vptestnmq %zmm0, %zmm0, %k0
881 ; AVX512BW-NEXT: kmovd %k0, %eax
882 ; AVX512BW-NEXT: testb $3, %al
883 ; AVX512BW-NEXT: setne %al
884 ; AVX512BW-NEXT: vzeroupper
885 ; AVX512BW-NEXT: retq
886 ;
887 ; AVX512VL-LABEL: icmp_v2i64_v2i1:
888 ; AVX512VL: # %bb.0:
889 ; AVX512VL-NEXT: vptestnmq %xmm0, %xmm0, %k0
890 ; AVX512VL-NEXT: kmovd %k0, %eax
891 ; AVX512VL-NEXT: testb %al, %al
892 ; AVX512VL-NEXT: setne %al
893 ; AVX512VL-NEXT: retq
894 %a = icmp eq <2 x i64> %0, zeroinitializer
895 %b = call i1 @llvm.vector.reduce.or.v2i1(<2 x i1> %a)
896 ret i1 %b
897 }
899 define i1 @icmp_v4i32_v4i1(<4 x i32>) {
900 ; SSE-LABEL: icmp_v4i32_v4i1:
901 ; SSE: # %bb.0:
902 ; SSE-NEXT: pxor %xmm1, %xmm1
903 ; SSE-NEXT: pcmpeqd %xmm0, %xmm1
904 ; SSE-NEXT: movmskps %xmm1, %eax
905 ; SSE-NEXT: testl %eax, %eax
906 ; SSE-NEXT: setne %al
907 ; SSE-NEXT: retq
908 ;
909 ; AVX-LABEL: icmp_v4i32_v4i1:
910 ; AVX: # %bb.0:
911 ; AVX-NEXT: vpxor %xmm1, %xmm1, %xmm1
912 ; AVX-NEXT: vpcmpeqd %xmm1, %xmm0, %xmm0
913 ; AVX-NEXT: vmovmskps %xmm0, %eax
914 ; AVX-NEXT: testl %eax, %eax
915 ; AVX-NEXT: setne %al
916 ; AVX-NEXT: retq
917 ;
918 ; AVX512F-LABEL: icmp_v4i32_v4i1:
919 ; AVX512F: # %bb.0:
920 ; AVX512F-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
921 ; AVX512F-NEXT: vptestnmd %zmm0, %zmm0, %k0
922 ; AVX512F-NEXT: kmovw %k0, %eax
923 ; AVX512F-NEXT: testb $15, %al
924 ; AVX512F-NEXT: setne %al
925 ; AVX512F-NEXT: vzeroupper
926 ; AVX512F-NEXT: retq
927 ;
928 ; AVX512BW-LABEL: icmp_v4i32_v4i1:
929 ; AVX512BW: # %bb.0:
930 ; AVX512BW-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
931 ; AVX512BW-NEXT: vptestnmd %zmm0, %zmm0, %k0
932 ; AVX512BW-NEXT: kmovd %k0, %eax
933 ; AVX512BW-NEXT: testb $15, %al
934 ; AVX512BW-NEXT: setne %al
935 ; AVX512BW-NEXT: vzeroupper
936 ; AVX512BW-NEXT: retq
937 ;
938 ; AVX512VL-LABEL: icmp_v4i32_v4i1:
939 ; AVX512VL: # %bb.0:
940 ; AVX512VL-NEXT: vptestnmd %xmm0, %xmm0, %k0
941 ; AVX512VL-NEXT: kmovd %k0, %eax
942 ; AVX512VL-NEXT: testb %al, %al
943 ; AVX512VL-NEXT: setne %al
944 ; AVX512VL-NEXT: retq
945 %a = icmp eq <4 x i32> %0, zeroinitializer
946 %b = call i1 @llvm.vector.reduce.or.v4i1(<4 x i1> %a)
947 ret i1 %b
948 }
950 define i1 @icmp_v8i16_v8i1(<8 x i8>) {
951 ; SSE-LABEL: icmp_v8i16_v8i1:
952 ; SSE: # %bb.0:
953 ; SSE-NEXT: pxor %xmm1, %xmm1
954 ; SSE-NEXT: pcmpeqb %xmm0, %xmm1
955 ; SSE-NEXT: pmovmskb %xmm1, %eax
956 ; SSE-NEXT: testb %al, %al
957 ; SSE-NEXT: setne %al
958 ; SSE-NEXT: retq
959 ;
960 ; AVX-LABEL: icmp_v8i16_v8i1:
961 ; AVX: # %bb.0:
962 ; AVX-NEXT: vpxor %xmm1, %xmm1, %xmm1
963 ; AVX-NEXT: vpcmpeqb %xmm1, %xmm0, %xmm0
964 ; AVX-NEXT: vpmovmskb %xmm0, %eax
965 ; AVX-NEXT: testb %al, %al
966 ; AVX-NEXT: setne %al
967 ; AVX-NEXT: retq
968 ;
969 ; AVX512F-LABEL: icmp_v8i16_v8i1:
970 ; AVX512F: # %bb.0:
971 ; AVX512F-NEXT: vpxor %xmm1, %xmm1, %xmm1
972 ; AVX512F-NEXT: vpcmpeqb %xmm1, %xmm0, %xmm0
973 ; AVX512F-NEXT: vpmovsxbd %xmm0, %zmm0
974 ; AVX512F-NEXT: vptestmd %zmm0, %zmm0, %k0
975 ; AVX512F-NEXT: kmovw %k0, %eax
976 ; AVX512F-NEXT: testb %al, %al
977 ; AVX512F-NEXT: setne %al
978 ; AVX512F-NEXT: vzeroupper
979 ; AVX512F-NEXT: retq
980 ;
981 ; AVX512BW-LABEL: icmp_v8i16_v8i1:
982 ; AVX512BW: # %bb.0:
983 ; AVX512BW-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
984 ; AVX512BW-NEXT: vptestnmb %zmm0, %zmm0, %k0
985 ; AVX512BW-NEXT: kmovd %k0, %eax
986 ; AVX512BW-NEXT: testb %al, %al
987 ; AVX512BW-NEXT: setne %al
988 ; AVX512BW-NEXT: vzeroupper
989 ; AVX512BW-NEXT: retq
990 ;
991 ; AVX512VL-LABEL: icmp_v8i16_v8i1:
992 ; AVX512VL: # %bb.0:
993 ; AVX512VL-NEXT: vptestnmb %xmm0, %xmm0, %k0
994 ; AVX512VL-NEXT: kmovd %k0, %eax
995 ; AVX512VL-NEXT: testb %al, %al
996 ; AVX512VL-NEXT: setne %al
997 ; AVX512VL-NEXT: retq
998 %a = icmp eq <8 x i8> %0, zeroinitializer
999 %b = call i1 @llvm.vector.reduce.or.v8i1(<8 x i1> %a)
1000 ret i1 %b
1001 }
1003 define i1 @icmp_v16i8_v16i1(<16 x i8>) {
1004 ; SSE-LABEL: icmp_v16i8_v16i1:
1005 ; SSE: # %bb.0:
1006 ; SSE-NEXT: pxor %xmm1, %xmm1
1007 ; SSE-NEXT: pcmpeqb %xmm0, %xmm1
1008 ; SSE-NEXT: pmovmskb %xmm1, %eax
1009 ; SSE-NEXT: testl %eax, %eax
1010 ; SSE-NEXT: setne %al
1011 ; SSE-NEXT: retq
1012 ;
1013 ; AVX-LABEL: icmp_v16i8_v16i1:
1014 ; AVX: # %bb.0:
1015 ; AVX-NEXT: vpxor %xmm1, %xmm1, %xmm1
1016 ; AVX-NEXT: vpcmpeqb %xmm1, %xmm0, %xmm0
1017 ; AVX-NEXT: vpmovmskb %xmm0, %eax
1018 ; AVX-NEXT: testl %eax, %eax
1019 ; AVX-NEXT: setne %al
1020 ; AVX-NEXT: retq
1021 ;
1022 ; AVX512F-LABEL: icmp_v16i8_v16i1:
1023 ; AVX512F: # %bb.0:
1024 ; AVX512F-NEXT: vpxor %xmm1, %xmm1, %xmm1
1025 ; AVX512F-NEXT: vpcmpeqb %xmm1, %xmm0, %xmm0
1026 ; AVX512F-NEXT: vpmovmskb %xmm0, %eax
1027 ; AVX512F-NEXT: testl %eax, %eax
1028 ; AVX512F-NEXT: setne %al
1029 ; AVX512F-NEXT: retq
1030 ;
1031 ; AVX512BW-LABEL: icmp_v16i8_v16i1:
1032 ; AVX512BW: # %bb.0:
1033 ; AVX512BW-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
1034 ; AVX512BW-NEXT: vptestnmb %zmm0, %zmm0, %k0
1035 ; AVX512BW-NEXT: kortestw %k0, %k0
1036 ; AVX512BW-NEXT: setne %al
1037 ; AVX512BW-NEXT: vzeroupper
1038 ; AVX512BW-NEXT: retq
1039 ;
1040 ; AVX512VL-LABEL: icmp_v16i8_v16i1:
1041 ; AVX512VL: # %bb.0:
1042 ; AVX512VL-NEXT: vptestnmb %xmm0, %xmm0, %k0
1043 ; AVX512VL-NEXT: kortestw %k0, %k0
1044 ; AVX512VL-NEXT: setne %al
1045 ; AVX512VL-NEXT: retq
1046 %a = icmp eq <16 x i8> %0, zeroinitializer
1047 %b = call i1 @llvm.vector.reduce.or.v16i1(<16 x i1> %a)
1048 ret i1 %b
1049 }
1051 define i1 @icmp_v4i64_v4i1(<4 x i64>) {
1052 ; SSE2-LABEL: icmp_v4i64_v4i1:
1053 ; SSE2: # %bb.0:
1054 ; SSE2-NEXT: pxor %xmm2, %xmm2
1055 ; SSE2-NEXT: pcmpeqd %xmm2, %xmm1
1056 ; SSE2-NEXT: pshufd {{.*#+}} xmm3 = xmm1[1,0,3,2]
1057 ; SSE2-NEXT: pand %xmm1, %xmm3
1058 ; SSE2-NEXT: pcmpeqd %xmm2, %xmm0
1059 ; SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm0[1,0,3,2]
1060 ; SSE2-NEXT: pand %xmm0, %xmm1
1061 ; SSE2-NEXT: packssdw %xmm3, %xmm1
1062 ; SSE2-NEXT: movmskps %xmm1, %eax
1063 ; SSE2-NEXT: testl %eax, %eax
1064 ; SSE2-NEXT: setne %al
1065 ; SSE2-NEXT: retq
1066 ;
1067 ; SSE41-LABEL: icmp_v4i64_v4i1:
1068 ; SSE41: # %bb.0:
1069 ; SSE41-NEXT: pxor %xmm2, %xmm2
1070 ; SSE41-NEXT: pcmpeqq %xmm2, %xmm1
1071 ; SSE41-NEXT: pcmpeqq %xmm2, %xmm0
1072 ; SSE41-NEXT: packssdw %xmm1, %xmm0
1073 ; SSE41-NEXT: movmskps %xmm0, %eax
1074 ; SSE41-NEXT: testl %eax, %eax
1075 ; SSE41-NEXT: setne %al
1076 ; SSE41-NEXT: retq
1077 ;
1078 ; AVX1-LABEL: icmp_v4i64_v4i1:
1079 ; AVX1: # %bb.0:
1080 ; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm1
1081 ; AVX1-NEXT: vpxor %xmm2, %xmm2, %xmm2
1082 ; AVX1-NEXT: vpcmpeqq %xmm2, %xmm1, %xmm1
1083 ; AVX1-NEXT: vpcmpeqq %xmm2, %xmm0, %xmm0
1084 ; AVX1-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm0
1085 ; AVX1-NEXT: vmovmskpd %ymm0, %eax
1086 ; AVX1-NEXT: testl %eax, %eax
1087 ; AVX1-NEXT: setne %al
1088 ; AVX1-NEXT: vzeroupper
1089 ; AVX1-NEXT: retq
1090 ;
1091 ; AVX2-LABEL: icmp_v4i64_v4i1:
1092 ; AVX2: # %bb.0:
1093 ; AVX2-NEXT: vpxor %xmm1, %xmm1, %xmm1
1094 ; AVX2-NEXT: vpcmpeqq %ymm1, %ymm0, %ymm0
1095 ; AVX2-NEXT: vmovmskpd %ymm0, %eax
1096 ; AVX2-NEXT: testl %eax, %eax
1097 ; AVX2-NEXT: setne %al
1098 ; AVX2-NEXT: vzeroupper
1099 ; AVX2-NEXT: retq
1100 ;
1101 ; AVX512F-LABEL: icmp_v4i64_v4i1:
1102 ; AVX512F: # %bb.0:
1103 ; AVX512F-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
1104 ; AVX512F-NEXT: vptestnmq %zmm0, %zmm0, %k0
1105 ; AVX512F-NEXT: kmovw %k0, %eax
1106 ; AVX512F-NEXT: testb $15, %al
1107 ; AVX512F-NEXT: setne %al
1108 ; AVX512F-NEXT: vzeroupper
1109 ; AVX512F-NEXT: retq
1110 ;
1111 ; AVX512BW-LABEL: icmp_v4i64_v4i1:
1112 ; AVX512BW: # %bb.0:
1113 ; AVX512BW-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
1114 ; AVX512BW-NEXT: vptestnmq %zmm0, %zmm0, %k0
1115 ; AVX512BW-NEXT: kmovd %k0, %eax
1116 ; AVX512BW-NEXT: testb $15, %al
1117 ; AVX512BW-NEXT: setne %al
1118 ; AVX512BW-NEXT: vzeroupper
1119 ; AVX512BW-NEXT: retq
1120 ;
1121 ; AVX512VL-LABEL: icmp_v4i64_v4i1:
1122 ; AVX512VL: # %bb.0:
1123 ; AVX512VL-NEXT: vptestnmq %ymm0, %ymm0, %k0
1124 ; AVX512VL-NEXT: kmovd %k0, %eax
1125 ; AVX512VL-NEXT: testb %al, %al
1126 ; AVX512VL-NEXT: setne %al
1127 ; AVX512VL-NEXT: vzeroupper
1128 ; AVX512VL-NEXT: retq
1129 %a = icmp eq <4 x i64> %0, zeroinitializer
1130 %b = call i1 @llvm.vector.reduce.or.v4i1(<4 x i1> %a)
1131 ret i1 %b
1132 }
1134 define i1 @icmp_v8i32_v8i1(<8 x i32>) {
1135 ; SSE-LABEL: icmp_v8i32_v8i1:
1136 ; SSE: # %bb.0:
1137 ; SSE-NEXT: pxor %xmm2, %xmm2
1138 ; SSE-NEXT: pcmpeqd %xmm2, %xmm1
1139 ; SSE-NEXT: pcmpeqd %xmm2, %xmm0
1140 ; SSE-NEXT: packssdw %xmm1, %xmm0
1141 ; SSE-NEXT: pmovmskb %xmm0, %eax
1142 ; SSE-NEXT: testl %eax, %eax
1143 ; SSE-NEXT: setne %al
1144 ; SSE-NEXT: retq
1145 ;
1146 ; AVX1-LABEL: icmp_v8i32_v8i1:
1147 ; AVX1: # %bb.0:
1148 ; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm1
1149 ; AVX1-NEXT: vpxor %xmm2, %xmm2, %xmm2
1150 ; AVX1-NEXT: vpcmpeqd %xmm2, %xmm1, %xmm1
1151 ; AVX1-NEXT: vpcmpeqd %xmm2, %xmm0, %xmm0
1152 ; AVX1-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm0
1153 ; AVX1-NEXT: vmovmskps %ymm0, %eax
1154 ; AVX1-NEXT: testl %eax, %eax
1155 ; AVX1-NEXT: setne %al
1156 ; AVX1-NEXT: vzeroupper
1157 ; AVX1-NEXT: retq
1158 ;
1159 ; AVX2-LABEL: icmp_v8i32_v8i1:
1160 ; AVX2: # %bb.0:
1161 ; AVX2-NEXT: vpxor %xmm1, %xmm1, %xmm1
1162 ; AVX2-NEXT: vpcmpeqd %ymm1, %ymm0, %ymm0
1163 ; AVX2-NEXT: vmovmskps %ymm0, %eax
1164 ; AVX2-NEXT: testl %eax, %eax
1165 ; AVX2-NEXT: setne %al
1166 ; AVX2-NEXT: vzeroupper
1167 ; AVX2-NEXT: retq
1168 ;
1169 ; AVX512F-LABEL: icmp_v8i32_v8i1:
1170 ; AVX512F: # %bb.0:
1171 ; AVX512F-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
1172 ; AVX512F-NEXT: vptestnmd %zmm0, %zmm0, %k0
1173 ; AVX512F-NEXT: kmovw %k0, %eax
1174 ; AVX512F-NEXT: testb %al, %al
1175 ; AVX512F-NEXT: setne %al
1176 ; AVX512F-NEXT: vzeroupper
1177 ; AVX512F-NEXT: retq
1178 ;
1179 ; AVX512BW-LABEL: icmp_v8i32_v8i1:
1180 ; AVX512BW: # %bb.0:
1181 ; AVX512BW-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
1182 ; AVX512BW-NEXT: vptestnmd %zmm0, %zmm0, %k0
1183 ; AVX512BW-NEXT: kmovd %k0, %eax
1184 ; AVX512BW-NEXT: testb %al, %al
1185 ; AVX512BW-NEXT: setne %al
1186 ; AVX512BW-NEXT: vzeroupper
1187 ; AVX512BW-NEXT: retq
1188 ;
1189 ; AVX512VL-LABEL: icmp_v8i32_v8i1:
1190 ; AVX512VL: # %bb.0:
1191 ; AVX512VL-NEXT: vptestnmd %ymm0, %ymm0, %k0
1192 ; AVX512VL-NEXT: kmovd %k0, %eax
1193 ; AVX512VL-NEXT: testb %al, %al
1194 ; AVX512VL-NEXT: setne %al
1195 ; AVX512VL-NEXT: vzeroupper
1196 ; AVX512VL-NEXT: retq
1197 %a = icmp eq <8 x i32> %0, zeroinitializer
1198 %b = call i1 @llvm.vector.reduce.or.v8i1(<8 x i1> %a)
1199 ret i1 %b
1200 }
1202 define i1 @icmp_v16i16_v16i1(<16 x i16>) {
1203 ; SSE-LABEL: icmp_v16i16_v16i1:
1204 ; SSE: # %bb.0:
1205 ; SSE-NEXT: pxor %xmm2, %xmm2
1206 ; SSE-NEXT: pcmpeqw %xmm2, %xmm1
1207 ; SSE-NEXT: pcmpeqw %xmm2, %xmm0
1208 ; SSE-NEXT: packsswb %xmm1, %xmm0
1209 ; SSE-NEXT: pmovmskb %xmm0, %eax
1210 ; SSE-NEXT: testl %eax, %eax
1211 ; SSE-NEXT: setne %al
1212 ; SSE-NEXT: retq
1213 ;
1214 ; AVX1-LABEL: icmp_v16i16_v16i1:
1215 ; AVX1: # %bb.0:
1216 ; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm1
1217 ; AVX1-NEXT: vpxor %xmm2, %xmm2, %xmm2
1218 ; AVX1-NEXT: vpcmpeqw %xmm2, %xmm1, %xmm1
1219 ; AVX1-NEXT: vpcmpeqw %xmm2, %xmm0, %xmm0
1220 ; AVX1-NEXT: vpacksswb %xmm1, %xmm0, %xmm0
1221 ; AVX1-NEXT: vpmovmskb %xmm0, %eax
1222 ; AVX1-NEXT: testl %eax, %eax
1223 ; AVX1-NEXT: setne %al
1224 ; AVX1-NEXT: vzeroupper
1225 ; AVX1-NEXT: retq
1226 ;
1227 ; AVX2-LABEL: icmp_v16i16_v16i1:
1228 ; AVX2: # %bb.0:
1229 ; AVX2-NEXT: vpxor %xmm1, %xmm1, %xmm1
1230 ; AVX2-NEXT: vpcmpeqw %ymm1, %ymm0, %ymm0
1231 ; AVX2-NEXT: vpmovmskb %ymm0, %eax
1232 ; AVX2-NEXT: testl %eax, %eax
1233 ; AVX2-NEXT: setne %al
1234 ; AVX2-NEXT: vzeroupper
1235 ; AVX2-NEXT: retq
1236 ;
1237 ; AVX512F-LABEL: icmp_v16i16_v16i1:
1238 ; AVX512F: # %bb.0:
1239 ; AVX512F-NEXT: vpxor %xmm1, %xmm1, %xmm1
1240 ; AVX512F-NEXT: vpcmpeqw %ymm1, %ymm0, %ymm0
1241 ; AVX512F-NEXT: vpmovsxwd %ymm0, %zmm0
1242 ; AVX512F-NEXT: vptestmd %zmm0, %zmm0, %k0
1243 ; AVX512F-NEXT: kortestw %k0, %k0
1244 ; AVX512F-NEXT: setne %al
1245 ; AVX512F-NEXT: vzeroupper
1246 ; AVX512F-NEXT: retq
1247 ;
1248 ; AVX512BW-LABEL: icmp_v16i16_v16i1:
1249 ; AVX512BW: # %bb.0:
1250 ; AVX512BW-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
1251 ; AVX512BW-NEXT: vptestnmw %zmm0, %zmm0, %k0
1252 ; AVX512BW-NEXT: kortestw %k0, %k0
1253 ; AVX512BW-NEXT: setne %al
1254 ; AVX512BW-NEXT: vzeroupper
1255 ; AVX512BW-NEXT: retq
1256 ;
1257 ; AVX512VL-LABEL: icmp_v16i16_v16i1:
1258 ; AVX512VL: # %bb.0:
1259 ; AVX512VL-NEXT: vptestnmw %ymm0, %ymm0, %k0
1260 ; AVX512VL-NEXT: kortestw %k0, %k0
1261 ; AVX512VL-NEXT: setne %al
1262 ; AVX512VL-NEXT: vzeroupper
1263 ; AVX512VL-NEXT: retq
1264 %a = icmp eq <16 x i16> %0, zeroinitializer
1265 %b = call i1 @llvm.vector.reduce.or.v16i1(<16 x i1> %a)
1266 ret i1 %b
1267 }
1269 define i1 @icmp_v32i8_v32i1(<32 x i8>) {
1270 ; SSE-LABEL: icmp_v32i8_v32i1:
1271 ; SSE: # %bb.0:
1272 ; SSE-NEXT: pxor %xmm2, %xmm2
1273 ; SSE-NEXT: pcmpeqb %xmm2, %xmm1
1274 ; SSE-NEXT: pcmpeqb %xmm2, %xmm0
1275 ; SSE-NEXT: por %xmm1, %xmm0
1276 ; SSE-NEXT: pmovmskb %xmm0, %eax
1277 ; SSE-NEXT: testl %eax, %eax
1278 ; SSE-NEXT: setne %al
1279 ; SSE-NEXT: retq
1280 ;
1281 ; AVX1-LABEL: icmp_v32i8_v32i1:
1282 ; AVX1: # %bb.0:
1283 ; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm1
1284 ; AVX1-NEXT: vpxor %xmm2, %xmm2, %xmm2
1285 ; AVX1-NEXT: vpcmpeqb %xmm2, %xmm1, %xmm1
1286 ; AVX1-NEXT: vpcmpeqb %xmm2, %xmm0, %xmm0
1287 ; AVX1-NEXT: vpor %xmm1, %xmm0, %xmm0
1288 ; AVX1-NEXT: vpmovmskb %xmm0, %eax
1289 ; AVX1-NEXT: testl %eax, %eax
1290 ; AVX1-NEXT: setne %al
1291 ; AVX1-NEXT: vzeroupper
1292 ; AVX1-NEXT: retq
1293 ;
1294 ; AVX2-LABEL: icmp_v32i8_v32i1:
1295 ; AVX2: # %bb.0:
1296 ; AVX2-NEXT: vpxor %xmm1, %xmm1, %xmm1
1297 ; AVX2-NEXT: vpcmpeqb %ymm1, %ymm0, %ymm0
1298 ; AVX2-NEXT: vpmovmskb %ymm0, %eax
1299 ; AVX2-NEXT: testl %eax, %eax
1300 ; AVX2-NEXT: setne %al
1301 ; AVX2-NEXT: vzeroupper
1302 ; AVX2-NEXT: retq
1303 ;
1304 ; AVX512F-LABEL: icmp_v32i8_v32i1:
1305 ; AVX512F: # %bb.0:
1306 ; AVX512F-NEXT: vpxor %xmm1, %xmm1, %xmm1
1307 ; AVX512F-NEXT: vpcmpeqb %ymm1, %ymm0, %ymm0
1308 ; AVX512F-NEXT: vextracti128 $1, %ymm0, %xmm1
1309 ; AVX512F-NEXT: vpor %xmm1, %xmm0, %xmm0
1310 ; AVX512F-NEXT: vpmovsxbd %xmm0, %zmm0
1311 ; AVX512F-NEXT: vptestmd %zmm0, %zmm0, %k0
1312 ; AVX512F-NEXT: kshiftrw $8, %k0, %k1
1313 ; AVX512F-NEXT: korw %k1, %k0, %k0
1314 ; AVX512F-NEXT: kshiftrw $4, %k0, %k1
1315 ; AVX512F-NEXT: korw %k1, %k0, %k0
1316 ; AVX512F-NEXT: kshiftrw $2, %k0, %k1
1317 ; AVX512F-NEXT: korw %k1, %k0, %k0
1318 ; AVX512F-NEXT: kshiftrw $1, %k0, %k1
1319 ; AVX512F-NEXT: korw %k1, %k0, %k0
1320 ; AVX512F-NEXT: kmovw %k0, %eax
1321 ; AVX512F-NEXT: # kill: def $al killed $al killed $eax
1322 ; AVX512F-NEXT: vzeroupper
1323 ; AVX512F-NEXT: retq
1324 ;
1325 ; AVX512BW-LABEL: icmp_v32i8_v32i1:
1326 ; AVX512BW: # %bb.0:
1327 ; AVX512BW-NEXT: # kill: def $ymm0 killed $ymm0 def $zmm0
1328 ; AVX512BW-NEXT: vptestnmb %zmm0, %zmm0, %k0
1329 ; AVX512BW-NEXT: kortestd %k0, %k0
1330 ; AVX512BW-NEXT: setne %al
1331 ; AVX512BW-NEXT: vzeroupper
1332 ; AVX512BW-NEXT: retq
1333 ;
1334 ; AVX512VL-LABEL: icmp_v32i8_v32i1:
1335 ; AVX512VL: # %bb.0:
1336 ; AVX512VL-NEXT: vptestnmb %ymm0, %ymm0, %k0
1337 ; AVX512VL-NEXT: kortestd %k0, %k0
1338 ; AVX512VL-NEXT: setne %al
1339 ; AVX512VL-NEXT: vzeroupper
1340 ; AVX512VL-NEXT: retq
1341 %a = icmp eq <32 x i8> %0, zeroinitializer
1342 %b = call i1 @llvm.vector.reduce.or.v32i1(<32 x i1> %a)
1343 ret i1 %b
1344 }
1346 define i1 @icmp_v8i64_v8i1(<8 x i64>) {
1347 ; SSE2-LABEL: icmp_v8i64_v8i1:
1348 ; SSE2: # %bb.0:
1349 ; SSE2-NEXT: pxor %xmm4, %xmm4
1350 ; SSE2-NEXT: pcmpeqd %xmm4, %xmm3
1351 ; SSE2-NEXT: pshufd {{.*#+}} xmm5 = xmm3[1,0,3,2]
1352 ; SSE2-NEXT: pand %xmm3, %xmm5
1353 ; SSE2-NEXT: pcmpeqd %xmm4, %xmm2
1354 ; SSE2-NEXT: pshufd {{.*#+}} xmm3 = xmm2[1,0,3,2]
1355 ; SSE2-NEXT: pand %xmm2, %xmm3
1356 ; SSE2-NEXT: packssdw %xmm5, %xmm3
1357 ; SSE2-NEXT: pcmpeqd %xmm4, %xmm1
1358 ; SSE2-NEXT: pshufd {{.*#+}} xmm2 = xmm1[1,0,3,2]
1359 ; SSE2-NEXT: pand %xmm1, %xmm2
1360 ; SSE2-NEXT: pcmpeqd %xmm4, %xmm0
1361 ; SSE2-NEXT: pshufd {{.*#+}} xmm1 = xmm0[1,0,3,2]
1362 ; SSE2-NEXT: pand %xmm0, %xmm1
1363 ; SSE2-NEXT: packssdw %xmm2, %xmm1
1364 ; SSE2-NEXT: packssdw %xmm3, %xmm1
1365 ; SSE2-NEXT: pmovmskb %xmm1, %eax
1366 ; SSE2-NEXT: testl $43690, %eax # imm = 0xAAAA
1367 ; SSE2-NEXT: setne %al
1368 ; SSE2-NEXT: retq
1369 ;
1370 ; SSE41-LABEL: icmp_v8i64_v8i1:
1371 ; SSE41: # %bb.0:
1372 ; SSE41-NEXT: pxor %xmm4, %xmm4
1373 ; SSE41-NEXT: pcmpeqq %xmm4, %xmm3
1374 ; SSE41-NEXT: pcmpeqq %xmm4, %xmm2
1375 ; SSE41-NEXT: packssdw %xmm3, %xmm2
1376 ; SSE41-NEXT: pcmpeqq %xmm4, %xmm1
1377 ; SSE41-NEXT: pcmpeqq %xmm4, %xmm0
1378 ; SSE41-NEXT: packssdw %xmm1, %xmm0
1379 ; SSE41-NEXT: packssdw %xmm2, %xmm0
1380 ; SSE41-NEXT: pmovmskb %xmm0, %eax
1381 ; SSE41-NEXT: testl $43690, %eax # imm = 0xAAAA
1382 ; SSE41-NEXT: setne %al
1383 ; SSE41-NEXT: retq
1384 ;
1385 ; AVX1-LABEL: icmp_v8i64_v8i1:
1386 ; AVX1: # %bb.0:
1387 ; AVX1-NEXT: vextractf128 $1, %ymm1, %xmm2
1388 ; AVX1-NEXT: vpxor %xmm3, %xmm3, %xmm3
1389 ; AVX1-NEXT: vpcmpeqq %xmm3, %xmm2, %xmm2
1390 ; AVX1-NEXT: vpcmpeqq %xmm3, %xmm1, %xmm1
1391 ; AVX1-NEXT: vpackssdw %xmm2, %xmm1, %xmm1
1392 ; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm2
1393 ; AVX1-NEXT: vpcmpeqq %xmm3, %xmm2, %xmm2
1394 ; AVX1-NEXT: vpcmpeqq %xmm3, %xmm0, %xmm0
1395 ; AVX1-NEXT: vpackssdw %xmm2, %xmm0, %xmm0
1396 ; AVX1-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm0
1397 ; AVX1-NEXT: vmovmskps %ymm0, %eax
1398 ; AVX1-NEXT: testl %eax, %eax
1399 ; AVX1-NEXT: setne %al
1400 ; AVX1-NEXT: vzeroupper
1401 ; AVX1-NEXT: retq
1402 ;
1403 ; AVX2-LABEL: icmp_v8i64_v8i1:
1404 ; AVX2: # %bb.0:
1405 ; AVX2-NEXT: vpxor %xmm2, %xmm2, %xmm2
1406 ; AVX2-NEXT: vpcmpeqq %ymm2, %ymm1, %ymm1
1407 ; AVX2-NEXT: vpcmpeqq %ymm2, %ymm0, %ymm0
1408 ; AVX2-NEXT: vpackssdw %ymm1, %ymm0, %ymm0
1409 ; AVX2-NEXT: vmovmskps %ymm0, %eax
1410 ; AVX2-NEXT: testl %eax, %eax
1411 ; AVX2-NEXT: setne %al
1412 ; AVX2-NEXT: vzeroupper
1413 ; AVX2-NEXT: retq
1414 ;
1415 ; AVX512F-LABEL: icmp_v8i64_v8i1:
1416 ; AVX512F: # %bb.0:
1417 ; AVX512F-NEXT: vptestnmq %zmm0, %zmm0, %k0
1418 ; AVX512F-NEXT: kmovw %k0, %eax
1419 ; AVX512F-NEXT: testb %al, %al
1420 ; AVX512F-NEXT: setne %al
1421 ; AVX512F-NEXT: vzeroupper
1422 ; AVX512F-NEXT: retq
1423 ;
1424 ; AVX512BW-LABEL: icmp_v8i64_v8i1:
1425 ; AVX512BW: # %bb.0:
1426 ; AVX512BW-NEXT: vptestnmq %zmm0, %zmm0, %k0
1427 ; AVX512BW-NEXT: kmovd %k0, %eax
1428 ; AVX512BW-NEXT: testb %al, %al
1429 ; AVX512BW-NEXT: setne %al
1430 ; AVX512BW-NEXT: vzeroupper
1431 ; AVX512BW-NEXT: retq
1432 ;
1433 ; AVX512VL-LABEL: icmp_v8i64_v8i1:
1434 ; AVX512VL: # %bb.0:
1435 ; AVX512VL-NEXT: vptestnmq %zmm0, %zmm0, %k0
1436 ; AVX512VL-NEXT: kmovd %k0, %eax
1437 ; AVX512VL-NEXT: testb %al, %al
1438 ; AVX512VL-NEXT: setne %al
1439 ; AVX512VL-NEXT: vzeroupper
1440 ; AVX512VL-NEXT: retq
1441 %a = icmp eq <8 x i64> %0, zeroinitializer
1442 %b = call i1 @llvm.vector.reduce.or.v8i1(<8 x i1> %a)
1443 ret i1 %b
1444 }
1446 define i1 @icmp_v16i32_v16i1(<16 x i32>) {
1447 ; SSE-LABEL: icmp_v16i32_v16i1:
1448 ; SSE: # %bb.0:
1449 ; SSE-NEXT: pxor %xmm4, %xmm4
1450 ; SSE-NEXT: pcmpeqd %xmm4, %xmm3
1451 ; SSE-NEXT: pcmpeqd %xmm4, %xmm2
1452 ; SSE-NEXT: packssdw %xmm3, %xmm2
1453 ; SSE-NEXT: pcmpeqd %xmm4, %xmm1
1454 ; SSE-NEXT: pcmpeqd %xmm4, %xmm0
1455 ; SSE-NEXT: packssdw %xmm1, %xmm0
1456 ; SSE-NEXT: packsswb %xmm2, %xmm0
1457 ; SSE-NEXT: pmovmskb %xmm0, %eax
1458 ; SSE-NEXT: testl %eax, %eax
1459 ; SSE-NEXT: setne %al
1460 ; SSE-NEXT: retq
1461 ;
1462 ; AVX1-LABEL: icmp_v16i32_v16i1:
1463 ; AVX1: # %bb.0:
1464 ; AVX1-NEXT: vextractf128 $1, %ymm1, %xmm2
1465 ; AVX1-NEXT: vpxor %xmm3, %xmm3, %xmm3
1466 ; AVX1-NEXT: vpcmpeqd %xmm3, %xmm2, %xmm2
1467 ; AVX1-NEXT: vpcmpeqd %xmm3, %xmm1, %xmm1
1468 ; AVX1-NEXT: vpackssdw %xmm2, %xmm1, %xmm1
1469 ; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm2
1470 ; AVX1-NEXT: vpcmpeqd %xmm3, %xmm2, %xmm2
1471 ; AVX1-NEXT: vpcmpeqd %xmm3, %xmm0, %xmm0
1472 ; AVX1-NEXT: vpackssdw %xmm2, %xmm0, %xmm0
1473 ; AVX1-NEXT: vpacksswb %xmm1, %xmm0, %xmm0
1474 ; AVX1-NEXT: vpmovmskb %xmm0, %eax
1475 ; AVX1-NEXT: testl %eax, %eax
1476 ; AVX1-NEXT: setne %al
1477 ; AVX1-NEXT: vzeroupper
1478 ; AVX1-NEXT: retq
1479 ;
1480 ; AVX2-LABEL: icmp_v16i32_v16i1:
1481 ; AVX2: # %bb.0:
1482 ; AVX2-NEXT: vpxor %xmm2, %xmm2, %xmm2
1483 ; AVX2-NEXT: vpcmpeqd %ymm2, %ymm1, %ymm1
1484 ; AVX2-NEXT: vpcmpeqd %ymm2, %ymm0, %ymm0
1485 ; AVX2-NEXT: vpackssdw %ymm1, %ymm0, %ymm0
1486 ; AVX2-NEXT: vpmovmskb %ymm0, %eax
1487 ; AVX2-NEXT: testl %eax, %eax
1488 ; AVX2-NEXT: setne %al
1489 ; AVX2-NEXT: vzeroupper
1490 ; AVX2-NEXT: retq
1491 ;
1492 ; AVX512-LABEL: icmp_v16i32_v16i1:
1493 ; AVX512: # %bb.0:
1494 ; AVX512-NEXT: vptestnmd %zmm0, %zmm0, %k0
1495 ; AVX512-NEXT: kortestw %k0, %k0
1496 ; AVX512-NEXT: setne %al
1497 ; AVX512-NEXT: vzeroupper
1498 ; AVX512-NEXT: retq
1499 %a = icmp eq <16 x i32> %0, zeroinitializer
1500 %b = call i1 @llvm.vector.reduce.or.v16i1(<16 x i1> %a)
1501 ret i1 %b
1502 }
1504 define i1 @icmp_v32i16_v32i1(<32 x i16>) {
1505 ; SSE-LABEL: icmp_v32i16_v32i1:
1506 ; SSE: # %bb.0:
1507 ; SSE-NEXT: pxor %xmm4, %xmm4
1508 ; SSE-NEXT: pcmpeqw %xmm4, %xmm1
1509 ; SSE-NEXT: pcmpeqw %xmm4, %xmm0
1510 ; SSE-NEXT: packsswb %xmm1, %xmm0
1511 ; SSE-NEXT: pcmpeqw %xmm4, %xmm3
1512 ; SSE-NEXT: pcmpeqw %xmm4, %xmm2
1513 ; SSE-NEXT: packsswb %xmm3, %xmm2
1514 ; SSE-NEXT: por %xmm0, %xmm2
1515 ; SSE-NEXT: pmovmskb %xmm2, %eax
1516 ; SSE-NEXT: testl %eax, %eax
1517 ; SSE-NEXT: setne %al
1518 ; SSE-NEXT: retq
1519 ;
1520 ; AVX1-LABEL: icmp_v32i16_v32i1:
1521 ; AVX1: # %bb.0:
1522 ; AVX1-NEXT: vextractf128 $1, %ymm1, %xmm2
1523 ; AVX1-NEXT: vpxor %xmm3, %xmm3, %xmm3
1524 ; AVX1-NEXT: vpcmpeqw %xmm3, %xmm2, %xmm2
1525 ; AVX1-NEXT: vpcmpeqw %xmm3, %xmm1, %xmm1
1526 ; AVX1-NEXT: vpacksswb %xmm2, %xmm1, %xmm1
1527 ; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm2
1528 ; AVX1-NEXT: vpcmpeqw %xmm3, %xmm2, %xmm2
1529 ; AVX1-NEXT: vpcmpeqw %xmm3, %xmm0, %xmm0
1530 ; AVX1-NEXT: vpacksswb %xmm2, %xmm0, %xmm0
1531 ; AVX1-NEXT: vpor %xmm1, %xmm0, %xmm0
1532 ; AVX1-NEXT: vpmovmskb %xmm0, %eax
1533 ; AVX1-NEXT: testl %eax, %eax
1534 ; AVX1-NEXT: setne %al
1535 ; AVX1-NEXT: vzeroupper
1536 ; AVX1-NEXT: retq
1537 ;
1538 ; AVX2-LABEL: icmp_v32i16_v32i1:
1539 ; AVX2: # %bb.0:
1540 ; AVX2-NEXT: vpxor %xmm2, %xmm2, %xmm2
1541 ; AVX2-NEXT: vpcmpeqw %ymm2, %ymm1, %ymm1
1542 ; AVX2-NEXT: vpcmpeqw %ymm2, %ymm0, %ymm0
1543 ; AVX2-NEXT: vpacksswb %ymm1, %ymm0, %ymm0
1544 ; AVX2-NEXT: vpmovmskb %ymm0, %eax
1545 ; AVX2-NEXT: testl %eax, %eax
1546 ; AVX2-NEXT: setne %al
1547 ; AVX2-NEXT: vzeroupper
1548 ; AVX2-NEXT: retq
1549 ;
1550 ; AVX512F-LABEL: icmp_v32i16_v32i1:
1551 ; AVX512F: # %bb.0:
1552 ; AVX512F-NEXT: vextracti64x4 $1, %zmm0, %ymm1
1553 ; AVX512F-NEXT: vpxor %xmm2, %xmm2, %xmm2
1554 ; AVX512F-NEXT: vpcmpeqw %ymm2, %ymm1, %ymm1
1555 ; AVX512F-NEXT: vpcmpeqw %ymm2, %ymm0, %ymm0
1556 ; AVX512F-NEXT: vpor %ymm1, %ymm0, %ymm0
1557 ; AVX512F-NEXT: vpmovsxwd %ymm0, %zmm0
1558 ; AVX512F-NEXT: vptestmd %zmm0, %zmm0, %k0
1559 ; AVX512F-NEXT: kshiftrw $8, %k0, %k1
1560 ; AVX512F-NEXT: korw %k1, %k0, %k0
1561 ; AVX512F-NEXT: kshiftrw $4, %k0, %k1
1562 ; AVX512F-NEXT: korw %k1, %k0, %k0
1563 ; AVX512F-NEXT: kshiftrw $2, %k0, %k1
1564 ; AVX512F-NEXT: korw %k1, %k0, %k0
1565 ; AVX512F-NEXT: kshiftrw $1, %k0, %k1
1566 ; AVX512F-NEXT: korw %k1, %k0, %k0
1567 ; AVX512F-NEXT: kmovw %k0, %eax
1568 ; AVX512F-NEXT: # kill: def $al killed $al killed $eax
1569 ; AVX512F-NEXT: vzeroupper
1570 ; AVX512F-NEXT: retq
1571 ;
1572 ; AVX512BW-LABEL: icmp_v32i16_v32i1:
1573 ; AVX512BW: # %bb.0:
1574 ; AVX512BW-NEXT: vptestnmw %zmm0, %zmm0, %k0
1575 ; AVX512BW-NEXT: kortestd %k0, %k0
1576 ; AVX512BW-NEXT: setne %al
1577 ; AVX512BW-NEXT: vzeroupper
1578 ; AVX512BW-NEXT: retq
1579 ;
1580 ; AVX512VL-LABEL: icmp_v32i16_v32i1:
1581 ; AVX512VL: # %bb.0:
1582 ; AVX512VL-NEXT: vptestnmw %zmm0, %zmm0, %k0
1583 ; AVX512VL-NEXT: kortestd %k0, %k0
1584 ; AVX512VL-NEXT: setne %al
1585 ; AVX512VL-NEXT: vzeroupper
1586 ; AVX512VL-NEXT: retq
1587 %a = icmp eq <32 x i16> %0, zeroinitializer
1588 %b = call i1 @llvm.vector.reduce.or.v32i1(<32 x i1> %a)
1589 ret i1 %b
1590 }
1592 define i1 @icmp_v64i8_v64i1(<64 x i8>) {
1593 ; SSE-LABEL: icmp_v64i8_v64i1:
1594 ; SSE: # %bb.0:
1595 ; SSE-NEXT: pxor %xmm4, %xmm4
1596 ; SSE-NEXT: pcmpeqb %xmm4, %xmm2
1597 ; SSE-NEXT: pcmpeqb %xmm4, %xmm0
1598 ; SSE-NEXT: pcmpeqb %xmm4, %xmm3
1599 ; SSE-NEXT: pcmpeqb %xmm4, %xmm1
1600 ; SSE-NEXT: por %xmm3, %xmm1
1601 ; SSE-NEXT: por %xmm2, %xmm1
1602 ; SSE-NEXT: por %xmm0, %xmm1
1603 ; SSE-NEXT: pmovmskb %xmm1, %eax
1604 ; SSE-NEXT: testl %eax, %eax
1605 ; SSE-NEXT: setne %al
1606 ; SSE-NEXT: retq
1607 ;
1608 ; AVX1-LABEL: icmp_v64i8_v64i1:
1609 ; AVX1: # %bb.0:
1610 ; AVX1-NEXT: vpxor %xmm2, %xmm2, %xmm2
1611 ; AVX1-NEXT: vpcmpeqb %xmm2, %xmm1, %xmm3
1612 ; AVX1-NEXT: vpcmpeqb %xmm2, %xmm0, %xmm4
1613 ; AVX1-NEXT: vextractf128 $1, %ymm1, %xmm1
1614 ; AVX1-NEXT: vpcmpeqb %xmm2, %xmm1, %xmm1
1615 ; AVX1-NEXT: vextractf128 $1, %ymm0, %xmm0
1616 ; AVX1-NEXT: vpcmpeqb %xmm2, %xmm0, %xmm0
1617 ; AVX1-NEXT: vpor %xmm1, %xmm0, %xmm0
1618 ; AVX1-NEXT: vpor %xmm0, %xmm3, %xmm0
1619 ; AVX1-NEXT: vpor %xmm0, %xmm4, %xmm0
1620 ; AVX1-NEXT: vpmovmskb %xmm0, %eax
1621 ; AVX1-NEXT: testl %eax, %eax
1622 ; AVX1-NEXT: setne %al
1623 ; AVX1-NEXT: vzeroupper
1624 ; AVX1-NEXT: retq
1625 ;
1626 ; AVX2-LABEL: icmp_v64i8_v64i1:
1627 ; AVX2: # %bb.0:
1628 ; AVX2-NEXT: vpxor %xmm2, %xmm2, %xmm2
1629 ; AVX2-NEXT: vpcmpeqb %ymm2, %ymm1, %ymm1
1630 ; AVX2-NEXT: vpcmpeqb %ymm2, %ymm0, %ymm0
1631 ; AVX2-NEXT: vpor %ymm1, %ymm0, %ymm0
1632 ; AVX2-NEXT: vpmovmskb %ymm0, %eax
1633 ; AVX2-NEXT: testl %eax, %eax
1634 ; AVX2-NEXT: setne %al
1635 ; AVX2-NEXT: vzeroupper
1636 ; AVX2-NEXT: retq
1637 ;
1638 ; AVX512F-LABEL: icmp_v64i8_v64i1:
1639 ; AVX512F: # %bb.0:
1640 ; AVX512F-NEXT: vextracti64x4 $1, %zmm0, %ymm1
1641 ; AVX512F-NEXT: vpxor %xmm2, %xmm2, %xmm2
1642 ; AVX512F-NEXT: vpcmpeqb %ymm2, %ymm1, %ymm1
1643 ; AVX512F-NEXT: vpcmpeqb %ymm2, %ymm0, %ymm0
1644 ; AVX512F-NEXT: vextracti128 $1, %ymm1, %xmm2
1645 ; AVX512F-NEXT: vextracti128 $1, %ymm0, %xmm3
1646 ; AVX512F-NEXT: vpor %xmm2, %xmm3, %xmm2
1647 ; AVX512F-NEXT: vpor %xmm2, %xmm1, %xmm1
1648 ; AVX512F-NEXT: vpor %xmm1, %xmm0, %xmm0
1649 ; AVX512F-NEXT: vpmovsxbd %xmm0, %zmm0
1650 ; AVX512F-NEXT: vpslld $31, %zmm0, %zmm0
1651 ; AVX512F-NEXT: vptestmd %zmm0, %zmm0, %k0
1652 ; AVX512F-NEXT: kshiftrw $8, %k0, %k1
1653 ; AVX512F-NEXT: korw %k1, %k0, %k0
1654 ; AVX512F-NEXT: kshiftrw $4, %k0, %k1
1655 ; AVX512F-NEXT: korw %k1, %k0, %k0
1656 ; AVX512F-NEXT: kshiftrw $2, %k0, %k1
1657 ; AVX512F-NEXT: korw %k1, %k0, %k0
1658 ; AVX512F-NEXT: kshiftrw $1, %k0, %k1
1659 ; AVX512F-NEXT: korw %k1, %k0, %k0
1660 ; AVX512F-NEXT: kmovw %k0, %eax
1661 ; AVX512F-NEXT: # kill: def $al killed $al killed $eax
1662 ; AVX512F-NEXT: vzeroupper
1663 ; AVX512F-NEXT: retq
1664 ;
1665 ; AVX512BW-LABEL: icmp_v64i8_v64i1:
1666 ; AVX512BW: # %bb.0:
1667 ; AVX512BW-NEXT: vptestnmb %zmm0, %zmm0, %k0
1668 ; AVX512BW-NEXT: kortestq %k0, %k0
1669 ; AVX512BW-NEXT: setne %al
1670 ; AVX512BW-NEXT: vzeroupper
1671 ; AVX512BW-NEXT: retq
1672 ;
1673 ; AVX512VL-LABEL: icmp_v64i8_v64i1:
1674 ; AVX512VL: # %bb.0:
1675 ; AVX512VL-NEXT: vptestnmb %zmm0, %zmm0, %k0
1676 ; AVX512VL-NEXT: kortestq %k0, %k0
1677 ; AVX512VL-NEXT: setne %al
1678 ; AVX512VL-NEXT: vzeroupper
1679 ; AVX512VL-NEXT: retq
1680 %a = icmp eq <64 x i8> %0, zeroinitializer
1681 %b = call i1 @llvm.vector.reduce.or.v64i1(<64 x i1> %a)
1682 ret i1 %b
1683 }
1685 declare i1 @llvm.vector.reduce.or.v2i1(<2 x i1>)
1686 declare i1 @llvm.vector.reduce.or.v4i1(<4 x i1>)
1687 declare i1 @llvm.vector.reduce.or.v8i1(<8 x i1>)
1688 declare i1 @llvm.vector.reduce.or.v16i1(<16 x i1>)
1689 declare i1 @llvm.vector.reduce.or.v32i1(<32 x i1>)
1690 declare i1 @llvm.vector.reduce.or.v64i1(<64 x i1>)