| blob | b9724f8c4d6db6d0f9d4ff4f20e80ac44b7f60e4 |
1 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
2 ; RUN: llc < %s -stack-symbol-ordering=0 -mtriple=x86_64-apple-darwin -mattr=+avx512f | FileCheck %s --check-prefix=CHECK --check-prefix=KNL
3 ; RUN: llc < %s -stack-symbol-ordering=0 -mtriple=x86_64-apple-darwin -mattr=+avx512f,+avx512bw,+avx512vl,+avx512dq | FileCheck %s --check-prefix=CHECK --check-prefix=SKX
4 ; RUN: llc < %s -stack-symbol-ordering=0 -mtriple=x86_64-apple-darwin -mattr=+avx512bw | FileCheck %s --check-prefix=CHECK --check-prefix=AVX512BW
5 ; RUN: llc < %s -stack-symbol-ordering=0 -mtriple=x86_64-apple-darwin -mattr=+avx512dq | FileCheck %s --check-prefix=CHECK --check-prefix=AVX512DQ
6 ; RUN: llc < %s -stack-symbol-ordering=0 -mtriple=i686-apple-darwin -mattr=+avx512f,+avx512bw,+avx512vl,+avx512dq | FileCheck %s --check-prefix=X86
9 define i16 @mask16(i16 %x) {
10 ; CHECK-LABEL: mask16:
11 ; CHECK: ## %bb.0:
12 ; CHECK-NEXT: movl %edi, %eax
13 ; CHECK-NEXT: notl %eax
14 ; CHECK-NEXT: ## kill: def $ax killed $ax killed $eax
15 ; CHECK-NEXT: retq
16 ;
17 ; X86-LABEL: mask16:
18 ; X86: ## %bb.0:
19 ; X86-NEXT: movl {{[0-9]+}}(%esp), %eax
20 ; X86-NEXT: notl %eax
21 ; X86-NEXT: ## kill: def $ax killed $ax killed $eax
22 ; X86-NEXT: retl
23 %m0 = bitcast i16 %x to <16 x i1>
24 %m1 = xor <16 x i1> %m0, <i1 -1, i1 -1, i1 -1, i1 -1, i1 -1, i1 -1, i1 -1, i1 -1, i1 -1, i1 -1, i1 -1, i1 -1, i1 -1, i1 -1, i1 -1, i1 -1>
25 %ret = bitcast <16 x i1> %m1 to i16
26 ret i16 %ret
27 }
29 define i32 @mask16_zext(i16 %x) {
30 ; CHECK-LABEL: mask16_zext:
31 ; CHECK: ## %bb.0:
32 ; CHECK-NEXT: notl %edi
33 ; CHECK-NEXT: movzwl %di, %eax
34 ; CHECK-NEXT: retq
35 ;
36 ; X86-LABEL: mask16_zext:
37 ; X86: ## %bb.0:
38 ; X86-NEXT: movzwl {{[0-9]+}}(%esp), %eax
39 ; X86-NEXT: xorl $65535, %eax ## imm = 0xFFFF
40 ; X86-NEXT: retl
41 %m0 = bitcast i16 %x to <16 x i1>
42 %m1 = xor <16 x i1> %m0, <i1 -1, i1 -1, i1 -1, i1 -1, i1 -1, i1 -1, i1 -1, i1 -1, i1 -1, i1 -1, i1 -1, i1 -1, i1 -1, i1 -1, i1 -1, i1 -1>
43 %m2 = bitcast <16 x i1> %m1 to i16
44 %ret = zext i16 %m2 to i32
45 ret i32 %ret
46 }
48 define i8 @mask8(i8 %x) {
49 ; CHECK-LABEL: mask8:
50 ; CHECK: ## %bb.0:
51 ; CHECK-NEXT: movl %edi, %eax
52 ; CHECK-NEXT: notb %al
53 ; CHECK-NEXT: ## kill: def $al killed $al killed $eax
54 ; CHECK-NEXT: retq
55 ;
56 ; X86-LABEL: mask8:
57 ; X86: ## %bb.0:
58 ; X86-NEXT: movb {{[0-9]+}}(%esp), %al
59 ; X86-NEXT: notb %al
60 ; X86-NEXT: retl
61 %m0 = bitcast i8 %x to <8 x i1>
62 %m1 = xor <8 x i1> %m0, <i1 -1, i1 -1, i1 -1, i1 -1, i1 -1, i1 -1, i1 -1, i1 -1>
63 %ret = bitcast <8 x i1> %m1 to i8
64 ret i8 %ret
65 }
67 define i32 @mask8_zext(i8 %x) {
68 ; CHECK-LABEL: mask8_zext:
69 ; CHECK: ## %bb.0:
70 ; CHECK-NEXT: notb %dil
71 ; CHECK-NEXT: movzbl %dil, %eax
72 ; CHECK-NEXT: retq
73 ;
74 ; X86-LABEL: mask8_zext:
75 ; X86: ## %bb.0:
76 ; X86-NEXT: movb {{[0-9]+}}(%esp), %al
77 ; X86-NEXT: notb %al
78 ; X86-NEXT: movzbl %al, %eax
79 ; X86-NEXT: retl
80 %m0 = bitcast i8 %x to <8 x i1>
81 %m1 = xor <8 x i1> %m0, <i1 -1, i1 -1, i1 -1, i1 -1, i1 -1, i1 -1, i1 -1, i1 -1>
82 %m2 = bitcast <8 x i1> %m1 to i8
83 %ret = zext i8 %m2 to i32
84 ret i32 %ret
85 }
87 define void @mask16_mem(i16* %ptr) {
88 ; CHECK-LABEL: mask16_mem:
89 ; CHECK: ## %bb.0:
90 ; CHECK-NEXT: kmovw (%rdi), %k0
91 ; CHECK-NEXT: knotw %k0, %k0
92 ; CHECK-NEXT: kmovw %k0, (%rdi)
93 ; CHECK-NEXT: retq
94 ;
95 ; X86-LABEL: mask16_mem:
96 ; X86: ## %bb.0:
97 ; X86-NEXT: movl {{[0-9]+}}(%esp), %eax
98 ; X86-NEXT: kmovw (%eax), %k0
99 ; X86-NEXT: knotw %k0, %k0
100 ; X86-NEXT: kmovw %k0, (%eax)
101 ; X86-NEXT: retl
102 %x = load i16, i16* %ptr, align 4
103 %m0 = bitcast i16 %x to <16 x i1>
104 %m1 = xor <16 x i1> %m0, <i1 -1, i1 -1, i1 -1, i1 -1, i1 -1, i1 -1, i1 -1, i1 -1, i1 -1, i1 -1, i1 -1, i1 -1, i1 -1, i1 -1, i1 -1, i1 -1>
105 %ret = bitcast <16 x i1> %m1 to i16
106 store i16 %ret, i16* %ptr, align 4
107 ret void
108 }
110 define void @mask8_mem(i8* %ptr) {
111 ; KNL-LABEL: mask8_mem:
112 ; KNL: ## %bb.0:
113 ; KNL-NEXT: notb (%rdi)
114 ; KNL-NEXT: retq
115 ;
116 ; SKX-LABEL: mask8_mem:
117 ; SKX: ## %bb.0:
118 ; SKX-NEXT: kmovb (%rdi), %k0
119 ; SKX-NEXT: knotb %k0, %k0
120 ; SKX-NEXT: kmovb %k0, (%rdi)
121 ; SKX-NEXT: retq
122 ;
123 ; AVX512BW-LABEL: mask8_mem:
124 ; AVX512BW: ## %bb.0:
125 ; AVX512BW-NEXT: notb (%rdi)
126 ; AVX512BW-NEXT: retq
127 ;
128 ; AVX512DQ-LABEL: mask8_mem:
129 ; AVX512DQ: ## %bb.0:
130 ; AVX512DQ-NEXT: kmovb (%rdi), %k0
131 ; AVX512DQ-NEXT: knotb %k0, %k0
132 ; AVX512DQ-NEXT: kmovb %k0, (%rdi)
133 ; AVX512DQ-NEXT: retq
134 ;
135 ; X86-LABEL: mask8_mem:
136 ; X86: ## %bb.0:
137 ; X86-NEXT: movl {{[0-9]+}}(%esp), %eax
138 ; X86-NEXT: kmovb (%eax), %k0
139 ; X86-NEXT: knotb %k0, %k0
140 ; X86-NEXT: kmovb %k0, (%eax)
141 ; X86-NEXT: retl
142 %x = load i8, i8* %ptr, align 4
143 %m0 = bitcast i8 %x to <8 x i1>
144 %m1 = xor <8 x i1> %m0, <i1 -1, i1 -1, i1 -1, i1 -1, i1 -1, i1 -1, i1 -1, i1 -1>
145 %ret = bitcast <8 x i1> %m1 to i8
146 store i8 %ret, i8* %ptr, align 4
147 ret void
148 }
150 define i16 @mand16(i16 %x, i16 %y) {
151 ; CHECK-LABEL: mand16:
152 ; CHECK: ## %bb.0:
153 ; CHECK-NEXT: movl %edi, %eax
154 ; CHECK-NEXT: movl %edi, %ecx
155 ; CHECK-NEXT: andl %esi, %ecx
156 ; CHECK-NEXT: xorl %esi, %eax
157 ; CHECK-NEXT: orl %ecx, %eax
158 ; CHECK-NEXT: ## kill: def $ax killed $ax killed $eax
159 ; CHECK-NEXT: retq
160 ;
161 ; X86-LABEL: mand16:
162 ; X86: ## %bb.0:
163 ; X86-NEXT: movl {{[0-9]+}}(%esp), %ecx
164 ; X86-NEXT: movl {{[0-9]+}}(%esp), %eax
165 ; X86-NEXT: movl %eax, %edx
166 ; X86-NEXT: andl %ecx, %edx
167 ; X86-NEXT: xorl %ecx, %eax
168 ; X86-NEXT: orl %edx, %eax
169 ; X86-NEXT: ## kill: def $ax killed $ax killed $eax
170 ; X86-NEXT: retl
171 %ma = bitcast i16 %x to <16 x i1>
172 %mb = bitcast i16 %y to <16 x i1>
173 %mc = and <16 x i1> %ma, %mb
174 %md = xor <16 x i1> %ma, %mb
175 %me = or <16 x i1> %mc, %md
176 %ret = bitcast <16 x i1> %me to i16
177 ret i16 %ret
178 }
180 define i16 @mand16_mem(<16 x i1>* %x, <16 x i1>* %y) {
181 ; KNL-LABEL: mand16_mem:
182 ; KNL: ## %bb.0:
183 ; KNL-NEXT: kmovw (%rdi), %k0
184 ; KNL-NEXT: kmovw (%rsi), %k1
185 ; KNL-NEXT: kandw %k1, %k0, %k2
186 ; KNL-NEXT: kxorw %k1, %k0, %k0
187 ; KNL-NEXT: korw %k0, %k2, %k0
188 ; KNL-NEXT: kmovw %k0, %eax
189 ; KNL-NEXT: ## kill: def $ax killed $ax killed $eax
190 ; KNL-NEXT: retq
191 ;
192 ; SKX-LABEL: mand16_mem:
193 ; SKX: ## %bb.0:
194 ; SKX-NEXT: kmovw (%rdi), %k0
195 ; SKX-NEXT: kmovw (%rsi), %k1
196 ; SKX-NEXT: kandw %k1, %k0, %k2
197 ; SKX-NEXT: kxorw %k1, %k0, %k0
198 ; SKX-NEXT: korw %k0, %k2, %k0
199 ; SKX-NEXT: kmovd %k0, %eax
200 ; SKX-NEXT: ## kill: def $ax killed $ax killed $eax
201 ; SKX-NEXT: retq
202 ;
203 ; AVX512BW-LABEL: mand16_mem:
204 ; AVX512BW: ## %bb.0:
205 ; AVX512BW-NEXT: kmovw (%rdi), %k0
206 ; AVX512BW-NEXT: kmovw (%rsi), %k1
207 ; AVX512BW-NEXT: kandw %k1, %k0, %k2
208 ; AVX512BW-NEXT: kxorw %k1, %k0, %k0
209 ; AVX512BW-NEXT: korw %k0, %k2, %k0
210 ; AVX512BW-NEXT: kmovd %k0, %eax
211 ; AVX512BW-NEXT: ## kill: def $ax killed $ax killed $eax
212 ; AVX512BW-NEXT: retq
213 ;
214 ; AVX512DQ-LABEL: mand16_mem:
215 ; AVX512DQ: ## %bb.0:
216 ; AVX512DQ-NEXT: kmovw (%rdi), %k0
217 ; AVX512DQ-NEXT: kmovw (%rsi), %k1
218 ; AVX512DQ-NEXT: kandw %k1, %k0, %k2
219 ; AVX512DQ-NEXT: kxorw %k1, %k0, %k0
220 ; AVX512DQ-NEXT: korw %k0, %k2, %k0
221 ; AVX512DQ-NEXT: kmovw %k0, %eax
222 ; AVX512DQ-NEXT: ## kill: def $ax killed $ax killed $eax
223 ; AVX512DQ-NEXT: retq
224 ;
225 ; X86-LABEL: mand16_mem:
226 ; X86: ## %bb.0:
227 ; X86-NEXT: movl {{[0-9]+}}(%esp), %eax
228 ; X86-NEXT: movl {{[0-9]+}}(%esp), %ecx
229 ; X86-NEXT: kmovw (%ecx), %k0
230 ; X86-NEXT: kmovw (%eax), %k1
231 ; X86-NEXT: kandw %k1, %k0, %k2
232 ; X86-NEXT: kxorw %k1, %k0, %k0
233 ; X86-NEXT: korw %k0, %k2, %k0
234 ; X86-NEXT: kmovd %k0, %eax
235 ; X86-NEXT: ## kill: def $ax killed $ax killed $eax
236 ; X86-NEXT: retl
237 %ma = load <16 x i1>, <16 x i1>* %x
238 %mb = load <16 x i1>, <16 x i1>* %y
239 %mc = and <16 x i1> %ma, %mb
240 %md = xor <16 x i1> %ma, %mb
241 %me = or <16 x i1> %mc, %md
242 %ret = bitcast <16 x i1> %me to i16
243 ret i16 %ret
244 }
246 define i8 @shuf_test1(i16 %v) nounwind {
247 ; KNL-LABEL: shuf_test1:
248 ; KNL: ## %bb.0:
249 ; KNL-NEXT: kmovw %edi, %k0
250 ; KNL-NEXT: kshiftrw $8, %k0, %k0
251 ; KNL-NEXT: kmovw %k0, %eax
252 ; KNL-NEXT: ## kill: def $al killed $al killed $eax
253 ; KNL-NEXT: retq
254 ;
255 ; SKX-LABEL: shuf_test1:
256 ; SKX: ## %bb.0:
257 ; SKX-NEXT: kmovd %edi, %k0
258 ; SKX-NEXT: kshiftrw $8, %k0, %k0
259 ; SKX-NEXT: kmovd %k0, %eax
260 ; SKX-NEXT: ## kill: def $al killed $al killed $eax
261 ; SKX-NEXT: retq
262 ;
263 ; AVX512BW-LABEL: shuf_test1:
264 ; AVX512BW: ## %bb.0:
265 ; AVX512BW-NEXT: kmovd %edi, %k0
266 ; AVX512BW-NEXT: kshiftrw $8, %k0, %k0
267 ; AVX512BW-NEXT: kmovd %k0, %eax
268 ; AVX512BW-NEXT: ## kill: def $al killed $al killed $eax
269 ; AVX512BW-NEXT: retq
270 ;
271 ; AVX512DQ-LABEL: shuf_test1:
272 ; AVX512DQ: ## %bb.0:
273 ; AVX512DQ-NEXT: kmovw %edi, %k0
274 ; AVX512DQ-NEXT: kshiftrw $8, %k0, %k0
275 ; AVX512DQ-NEXT: kmovw %k0, %eax
276 ; AVX512DQ-NEXT: ## kill: def $al killed $al killed $eax
277 ; AVX512DQ-NEXT: retq
278 ;
279 ; X86-LABEL: shuf_test1:
280 ; X86: ## %bb.0:
281 ; X86-NEXT: movb {{[0-9]+}}(%esp), %al
282 ; X86-NEXT: retl
283 %v1 = bitcast i16 %v to <16 x i1>
284 %mask = shufflevector <16 x i1> %v1, <16 x i1> undef, <8 x i32> <i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
285 %mask1 = bitcast <8 x i1> %mask to i8
286 ret i8 %mask1
287 }
289 define i32 @zext_test1(<16 x i32> %a, <16 x i32> %b) {
290 ; KNL-LABEL: zext_test1:
291 ; KNL: ## %bb.0:
292 ; KNL-NEXT: vpcmpnleud %zmm1, %zmm0, %k0
293 ; KNL-NEXT: kshiftrw $5, %k0, %k0
294 ; KNL-NEXT: kmovw %k0, %eax
295 ; KNL-NEXT: andl $1, %eax
296 ; KNL-NEXT: vzeroupper
297 ; KNL-NEXT: retq
298 ;
299 ; SKX-LABEL: zext_test1:
300 ; SKX: ## %bb.0:
301 ; SKX-NEXT: vpcmpnleud %zmm1, %zmm0, %k0
302 ; SKX-NEXT: kshiftrw $5, %k0, %k0
303 ; SKX-NEXT: kmovd %k0, %eax
304 ; SKX-NEXT: andl $1, %eax
305 ; SKX-NEXT: vzeroupper
306 ; SKX-NEXT: retq
307 ;
308 ; AVX512BW-LABEL: zext_test1:
309 ; AVX512BW: ## %bb.0:
310 ; AVX512BW-NEXT: vpcmpnleud %zmm1, %zmm0, %k0
311 ; AVX512BW-NEXT: kshiftrw $5, %k0, %k0
312 ; AVX512BW-NEXT: kmovd %k0, %eax
313 ; AVX512BW-NEXT: andl $1, %eax
314 ; AVX512BW-NEXT: vzeroupper
315 ; AVX512BW-NEXT: retq
316 ;
317 ; AVX512DQ-LABEL: zext_test1:
318 ; AVX512DQ: ## %bb.0:
319 ; AVX512DQ-NEXT: vpcmpnleud %zmm1, %zmm0, %k0
320 ; AVX512DQ-NEXT: kshiftrw $5, %k0, %k0
321 ; AVX512DQ-NEXT: kmovw %k0, %eax
322 ; AVX512DQ-NEXT: andl $1, %eax
323 ; AVX512DQ-NEXT: vzeroupper
324 ; AVX512DQ-NEXT: retq
325 ;
326 ; X86-LABEL: zext_test1:
327 ; X86: ## %bb.0:
328 ; X86-NEXT: vpcmpnleud %zmm1, %zmm0, %k0
329 ; X86-NEXT: kshiftrw $5, %k0, %k0
330 ; X86-NEXT: kmovd %k0, %eax
331 ; X86-NEXT: andl $1, %eax
332 ; X86-NEXT: vzeroupper
333 ; X86-NEXT: retl
334 %cmp_res = icmp ugt <16 x i32> %a, %b
335 %cmp_res.i1 = extractelement <16 x i1> %cmp_res, i32 5
336 %res = zext i1 %cmp_res.i1 to i32
337 ret i32 %res
338 }
340 define i16 @zext_test2(<16 x i32> %a, <16 x i32> %b) {
341 ; KNL-LABEL: zext_test2:
342 ; KNL: ## %bb.0:
343 ; KNL-NEXT: vpcmpnleud %zmm1, %zmm0, %k0
344 ; KNL-NEXT: kshiftrw $5, %k0, %k0
345 ; KNL-NEXT: kmovw %k0, %eax
346 ; KNL-NEXT: andl $1, %eax
347 ; KNL-NEXT: ## kill: def $ax killed $ax killed $eax
348 ; KNL-NEXT: vzeroupper
349 ; KNL-NEXT: retq
350 ;
351 ; SKX-LABEL: zext_test2:
352 ; SKX: ## %bb.0:
353 ; SKX-NEXT: vpcmpnleud %zmm1, %zmm0, %k0
354 ; SKX-NEXT: kshiftrw $5, %k0, %k0
355 ; SKX-NEXT: kmovd %k0, %eax
356 ; SKX-NEXT: andl $1, %eax
357 ; SKX-NEXT: ## kill: def $ax killed $ax killed $eax
358 ; SKX-NEXT: vzeroupper
359 ; SKX-NEXT: retq
360 ;
361 ; AVX512BW-LABEL: zext_test2:
362 ; AVX512BW: ## %bb.0:
363 ; AVX512BW-NEXT: vpcmpnleud %zmm1, %zmm0, %k0
364 ; AVX512BW-NEXT: kshiftrw $5, %k0, %k0
365 ; AVX512BW-NEXT: kmovd %k0, %eax
366 ; AVX512BW-NEXT: andl $1, %eax
367 ; AVX512BW-NEXT: ## kill: def $ax killed $ax killed $eax
368 ; AVX512BW-NEXT: vzeroupper
369 ; AVX512BW-NEXT: retq
370 ;
371 ; AVX512DQ-LABEL: zext_test2:
372 ; AVX512DQ: ## %bb.0:
373 ; AVX512DQ-NEXT: vpcmpnleud %zmm1, %zmm0, %k0
374 ; AVX512DQ-NEXT: kshiftrw $5, %k0, %k0
375 ; AVX512DQ-NEXT: kmovw %k0, %eax
376 ; AVX512DQ-NEXT: andl $1, %eax
377 ; AVX512DQ-NEXT: ## kill: def $ax killed $ax killed $eax
378 ; AVX512DQ-NEXT: vzeroupper
379 ; AVX512DQ-NEXT: retq
380 ;
381 ; X86-LABEL: zext_test2:
382 ; X86: ## %bb.0:
383 ; X86-NEXT: vpcmpnleud %zmm1, %zmm0, %k0
384 ; X86-NEXT: kshiftrw $5, %k0, %k0
385 ; X86-NEXT: kmovd %k0, %eax
386 ; X86-NEXT: andl $1, %eax
387 ; X86-NEXT: ## kill: def $ax killed $ax killed $eax
388 ; X86-NEXT: vzeroupper
389 ; X86-NEXT: retl
390 %cmp_res = icmp ugt <16 x i32> %a, %b
391 %cmp_res.i1 = extractelement <16 x i1> %cmp_res, i32 5
392 %res = zext i1 %cmp_res.i1 to i16
393 ret i16 %res
394 }
396 define i8 @zext_test3(<16 x i32> %a, <16 x i32> %b) {
397 ; KNL-LABEL: zext_test3:
398 ; KNL: ## %bb.0:
399 ; KNL-NEXT: vpcmpnleud %zmm1, %zmm0, %k0
400 ; KNL-NEXT: kshiftrw $5, %k0, %k0
401 ; KNL-NEXT: kmovw %k0, %eax
402 ; KNL-NEXT: andb $1, %al
403 ; KNL-NEXT: ## kill: def $al killed $al killed $eax
404 ; KNL-NEXT: vzeroupper
405 ; KNL-NEXT: retq
406 ;
407 ; SKX-LABEL: zext_test3:
408 ; SKX: ## %bb.0:
409 ; SKX-NEXT: vpcmpnleud %zmm1, %zmm0, %k0
410 ; SKX-NEXT: kshiftrw $5, %k0, %k0
411 ; SKX-NEXT: kmovd %k0, %eax
412 ; SKX-NEXT: andb $1, %al
413 ; SKX-NEXT: ## kill: def $al killed $al killed $eax
414 ; SKX-NEXT: vzeroupper
415 ; SKX-NEXT: retq
416 ;
417 ; AVX512BW-LABEL: zext_test3:
418 ; AVX512BW: ## %bb.0:
419 ; AVX512BW-NEXT: vpcmpnleud %zmm1, %zmm0, %k0
420 ; AVX512BW-NEXT: kshiftrw $5, %k0, %k0
421 ; AVX512BW-NEXT: kmovd %k0, %eax
422 ; AVX512BW-NEXT: andb $1, %al
423 ; AVX512BW-NEXT: ## kill: def $al killed $al killed $eax
424 ; AVX512BW-NEXT: vzeroupper
425 ; AVX512BW-NEXT: retq
426 ;
427 ; AVX512DQ-LABEL: zext_test3:
428 ; AVX512DQ: ## %bb.0:
429 ; AVX512DQ-NEXT: vpcmpnleud %zmm1, %zmm0, %k0
430 ; AVX512DQ-NEXT: kshiftrw $5, %k0, %k0
431 ; AVX512DQ-NEXT: kmovw %k0, %eax
432 ; AVX512DQ-NEXT: andb $1, %al
433 ; AVX512DQ-NEXT: ## kill: def $al killed $al killed $eax
434 ; AVX512DQ-NEXT: vzeroupper
435 ; AVX512DQ-NEXT: retq
436 ;
437 ; X86-LABEL: zext_test3:
438 ; X86: ## %bb.0:
439 ; X86-NEXT: vpcmpnleud %zmm1, %zmm0, %k0
440 ; X86-NEXT: kshiftrw $5, %k0, %k0
441 ; X86-NEXT: kmovd %k0, %eax
442 ; X86-NEXT: andb $1, %al
443 ; X86-NEXT: ## kill: def $al killed $al killed $eax
444 ; X86-NEXT: vzeroupper
445 ; X86-NEXT: retl
446 %cmp_res = icmp ugt <16 x i32> %a, %b
447 %cmp_res.i1 = extractelement <16 x i1> %cmp_res, i32 5
448 %res = zext i1 %cmp_res.i1 to i8
449 ret i8 %res
450 }
452 define i8 @conv1(<8 x i1>* %R) {
453 ; CHECK-LABEL: conv1:
454 ; CHECK: ## %bb.0: ## %entry
455 ; CHECK-NEXT: movb $-1, (%rdi)
456 ; CHECK-NEXT: movb $-2, -{{[0-9]+}}(%rsp)
457 ; CHECK-NEXT: movb $-2, %al
458 ; CHECK-NEXT: retq
459 ;
460 ; X86-LABEL: conv1:
461 ; X86: ## %bb.0: ## %entry
462 ; X86-NEXT: subl $12, %esp
463 ; X86-NEXT: .cfi_def_cfa_offset 16
464 ; X86-NEXT: movl {{[0-9]+}}(%esp), %eax
465 ; X86-NEXT: movb $-1, (%eax)
466 ; X86-NEXT: movb $-2, (%esp)
467 ; X86-NEXT: movb $-2, %al
468 ; X86-NEXT: addl $12, %esp
469 ; X86-NEXT: retl
470 entry:
471 store <8 x i1> <i1 1, i1 1, i1 1, i1 1, i1 1, i1 1, i1 1, i1 1>, <8 x i1>* %R
473 %maskPtr = alloca <8 x i1>
474 store <8 x i1> <i1 0, i1 1, i1 1, i1 1, i1 1, i1 1, i1 1, i1 1>, <8 x i1>* %maskPtr
475 %mask = load <8 x i1>, <8 x i1>* %maskPtr
476 %mask_convert = bitcast <8 x i1> %mask to i8
477 ret i8 %mask_convert
478 }
480 define <4 x i32> @test4(<4 x i64> %x, <4 x i64> %y, <4 x i64> %x1, <4 x i64> %y1) {
481 ; KNL-LABEL: test4:
482 ; KNL: ## %bb.0:
483 ; KNL-NEXT: ## kill: def $ymm3 killed $ymm3 def $zmm3
484 ; KNL-NEXT: ## kill: def $ymm2 killed $ymm2 def $zmm2
485 ; KNL-NEXT: ## kill: def $ymm1 killed $ymm1 def $zmm1
486 ; KNL-NEXT: ## kill: def $ymm0 killed $ymm0 def $zmm0
487 ; KNL-NEXT: vpcmpgtq %zmm3, %zmm2, %k1
488 ; KNL-NEXT: vpcmpleq %zmm1, %zmm0, %k1 {%k1}
489 ; KNL-NEXT: vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
490 ; KNL-NEXT: ## kill: def $xmm0 killed $xmm0 killed $zmm0
491 ; KNL-NEXT: vzeroupper
492 ; KNL-NEXT: retq
493 ;
494 ; SKX-LABEL: test4:
495 ; SKX: ## %bb.0:
496 ; SKX-NEXT: vpcmpgtq %ymm3, %ymm2, %k1
497 ; SKX-NEXT: vpcmpleq %ymm1, %ymm0, %k0 {%k1}
498 ; SKX-NEXT: vpmovm2d %k0, %xmm0
499 ; SKX-NEXT: vzeroupper
500 ; SKX-NEXT: retq
501 ;
502 ; AVX512BW-LABEL: test4:
503 ; AVX512BW: ## %bb.0:
504 ; AVX512BW-NEXT: ## kill: def $ymm3 killed $ymm3 def $zmm3
505 ; AVX512BW-NEXT: ## kill: def $ymm2 killed $ymm2 def $zmm2
506 ; AVX512BW-NEXT: ## kill: def $ymm1 killed $ymm1 def $zmm1
507 ; AVX512BW-NEXT: ## kill: def $ymm0 killed $ymm0 def $zmm0
508 ; AVX512BW-NEXT: vpcmpgtq %zmm3, %zmm2, %k1
509 ; AVX512BW-NEXT: vpcmpleq %zmm1, %zmm0, %k1 {%k1}
510 ; AVX512BW-NEXT: vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
511 ; AVX512BW-NEXT: ## kill: def $xmm0 killed $xmm0 killed $zmm0
512 ; AVX512BW-NEXT: vzeroupper
513 ; AVX512BW-NEXT: retq
514 ;
515 ; AVX512DQ-LABEL: test4:
516 ; AVX512DQ: ## %bb.0:
517 ; AVX512DQ-NEXT: ## kill: def $ymm3 killed $ymm3 def $zmm3
518 ; AVX512DQ-NEXT: ## kill: def $ymm2 killed $ymm2 def $zmm2
519 ; AVX512DQ-NEXT: ## kill: def $ymm1 killed $ymm1 def $zmm1
520 ; AVX512DQ-NEXT: ## kill: def $ymm0 killed $ymm0 def $zmm0
521 ; AVX512DQ-NEXT: vpcmpgtq %zmm3, %zmm2, %k1
522 ; AVX512DQ-NEXT: vpcmpleq %zmm1, %zmm0, %k0 {%k1}
523 ; AVX512DQ-NEXT: vpmovm2d %k0, %zmm0
524 ; AVX512DQ-NEXT: ## kill: def $xmm0 killed $xmm0 killed $zmm0
525 ; AVX512DQ-NEXT: vzeroupper
526 ; AVX512DQ-NEXT: retq
527 ;
528 ; X86-LABEL: test4:
529 ; X86: ## %bb.0:
530 ; X86-NEXT: vpcmpgtq %ymm3, %ymm2, %k1
531 ; X86-NEXT: vpcmpleq %ymm1, %ymm0, %k0 {%k1}
532 ; X86-NEXT: vpmovm2d %k0, %xmm0
533 ; X86-NEXT: vzeroupper
534 ; X86-NEXT: retl
535 %x_gt_y = icmp sgt <4 x i64> %x, %y
536 %x1_gt_y1 = icmp sgt <4 x i64> %x1, %y1
537 %res = icmp sgt <4 x i1>%x_gt_y, %x1_gt_y1
538 %resse = sext <4 x i1>%res to <4 x i32>
539 ret <4 x i32> %resse
540 }
542 define <2 x i64> @test5(<2 x i64> %x, <2 x i64> %y, <2 x i64> %x1, <2 x i64> %y1) {
543 ; KNL-LABEL: test5:
544 ; KNL: ## %bb.0:
545 ; KNL-NEXT: ## kill: def $xmm3 killed $xmm3 def $zmm3
546 ; KNL-NEXT: ## kill: def $xmm2 killed $xmm2 def $zmm2
547 ; KNL-NEXT: ## kill: def $xmm1 killed $xmm1 def $zmm1
548 ; KNL-NEXT: ## kill: def $xmm0 killed $xmm0 def $zmm0
549 ; KNL-NEXT: vpcmpgtq %zmm0, %zmm1, %k1
550 ; KNL-NEXT: vpcmpleq %zmm3, %zmm2, %k1 {%k1}
551 ; KNL-NEXT: vpternlogq $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
552 ; KNL-NEXT: ## kill: def $xmm0 killed $xmm0 killed $zmm0
553 ; KNL-NEXT: vzeroupper
554 ; KNL-NEXT: retq
555 ;
556 ; SKX-LABEL: test5:
557 ; SKX: ## %bb.0:
558 ; SKX-NEXT: vpcmpgtq %xmm0, %xmm1, %k1
559 ; SKX-NEXT: vpcmpleq %xmm3, %xmm2, %k0 {%k1}
560 ; SKX-NEXT: vpmovm2q %k0, %xmm0
561 ; SKX-NEXT: retq
562 ;
563 ; AVX512BW-LABEL: test5:
564 ; AVX512BW: ## %bb.0:
565 ; AVX512BW-NEXT: ## kill: def $xmm3 killed $xmm3 def $zmm3
566 ; AVX512BW-NEXT: ## kill: def $xmm2 killed $xmm2 def $zmm2
567 ; AVX512BW-NEXT: ## kill: def $xmm1 killed $xmm1 def $zmm1
568 ; AVX512BW-NEXT: ## kill: def $xmm0 killed $xmm0 def $zmm0
569 ; AVX512BW-NEXT: vpcmpgtq %zmm0, %zmm1, %k1
570 ; AVX512BW-NEXT: vpcmpleq %zmm3, %zmm2, %k1 {%k1}
571 ; AVX512BW-NEXT: vpternlogq $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
572 ; AVX512BW-NEXT: ## kill: def $xmm0 killed $xmm0 killed $zmm0
573 ; AVX512BW-NEXT: vzeroupper
574 ; AVX512BW-NEXT: retq
575 ;
576 ; AVX512DQ-LABEL: test5:
577 ; AVX512DQ: ## %bb.0:
578 ; AVX512DQ-NEXT: ## kill: def $xmm3 killed $xmm3 def $zmm3
579 ; AVX512DQ-NEXT: ## kill: def $xmm2 killed $xmm2 def $zmm2
580 ; AVX512DQ-NEXT: ## kill: def $xmm1 killed $xmm1 def $zmm1
581 ; AVX512DQ-NEXT: ## kill: def $xmm0 killed $xmm0 def $zmm0
582 ; AVX512DQ-NEXT: vpcmpgtq %zmm0, %zmm1, %k1
583 ; AVX512DQ-NEXT: vpcmpleq %zmm3, %zmm2, %k0 {%k1}
584 ; AVX512DQ-NEXT: vpmovm2q %k0, %zmm0
585 ; AVX512DQ-NEXT: ## kill: def $xmm0 killed $xmm0 killed $zmm0
586 ; AVX512DQ-NEXT: vzeroupper
587 ; AVX512DQ-NEXT: retq
588 ;
589 ; X86-LABEL: test5:
590 ; X86: ## %bb.0:
591 ; X86-NEXT: vpcmpgtq %xmm0, %xmm1, %k1
592 ; X86-NEXT: vpcmpleq %xmm3, %xmm2, %k0 {%k1}
593 ; X86-NEXT: vpmovm2q %k0, %xmm0
594 ; X86-NEXT: retl
595 %x_gt_y = icmp slt <2 x i64> %x, %y
596 %x1_gt_y1 = icmp sgt <2 x i64> %x1, %y1
597 %res = icmp slt <2 x i1>%x_gt_y, %x1_gt_y1
598 %resse = sext <2 x i1>%res to <2 x i64>
599 ret <2 x i64> %resse
600 }define void @test6(<16 x i1> %mask) {
601 allocas:
602 %a= and <16 x i1> %mask, <i1 true, i1 false, i1 true, i1 false, i1 true, i1 false, i1 true, i1 false, i1 true, i1 false, i1 true, i1 false, i1 true, i1 false, i1 true, i1 false>
603 %b = bitcast <16 x i1> %a to i16
604 %c = icmp eq i16 %b, 0
605 br i1 %c, label %true, label %false
607 true:
608 ret void
610 false:
611 ret void
612 }
613 define void @test7(<8 x i1> %mask) {
614 ; KNL-LABEL: test7:
615 ; KNL: ## %bb.0: ## %allocas
616 ; KNL-NEXT: vpmovsxwq %xmm0, %zmm0
617 ; KNL-NEXT: vpsllq $63, %zmm0, %zmm0
618 ; KNL-NEXT: vptestmq %zmm0, %zmm0, %k0
619 ; KNL-NEXT: kmovw %k0, %eax
620 ; KNL-NEXT: orb $85, %al
621 ; KNL-NEXT: vzeroupper
622 ; KNL-NEXT: retq
623 ;
624 ; SKX-LABEL: test7:
625 ; SKX: ## %bb.0: ## %allocas
626 ; SKX-NEXT: vpsllw $15, %xmm0, %xmm0
627 ; SKX-NEXT: vpmovw2m %xmm0, %k0
628 ; SKX-NEXT: kmovd %k0, %eax
629 ; SKX-NEXT: orb $85, %al
630 ; SKX-NEXT: retq
631 ;
632 ; AVX512BW-LABEL: test7:
633 ; AVX512BW: ## %bb.0: ## %allocas
634 ; AVX512BW-NEXT: vpsllw $15, %xmm0, %xmm0
635 ; AVX512BW-NEXT: vpmovw2m %zmm0, %k0
636 ; AVX512BW-NEXT: kmovd %k0, %eax
637 ; AVX512BW-NEXT: orb $85, %al
638 ; AVX512BW-NEXT: vzeroupper
639 ; AVX512BW-NEXT: retq
640 ;
641 ; AVX512DQ-LABEL: test7:
642 ; AVX512DQ: ## %bb.0: ## %allocas
643 ; AVX512DQ-NEXT: vpmovsxwq %xmm0, %zmm0
644 ; AVX512DQ-NEXT: vpsllq $63, %zmm0, %zmm0
645 ; AVX512DQ-NEXT: vpmovq2m %zmm0, %k0
646 ; AVX512DQ-NEXT: kmovw %k0, %eax
647 ; AVX512DQ-NEXT: orb $85, %al
648 ; AVX512DQ-NEXT: vzeroupper
649 ; AVX512DQ-NEXT: retq
650 ;
651 ; X86-LABEL: test7:
652 ; X86: ## %bb.0: ## %allocas
653 ; X86-NEXT: vpsllw $15, %xmm0, %xmm0
654 ; X86-NEXT: vpmovw2m %xmm0, %k0
655 ; X86-NEXT: kmovd %k0, %eax
656 ; X86-NEXT: orb $85, %al
657 ; X86-NEXT: retl
658 allocas:
659 %a= or <8 x i1> %mask, <i1 true, i1 false, i1 true, i1 false, i1 true, i1 false, i1 true, i1 false>
660 %b = bitcast <8 x i1> %a to i8
661 %c = icmp eq i8 %b, 0
662 br i1 %c, label %true, label %false
664 true:
665 ret void
667 false:
668 ret void
669 }
670 define <16 x i8> @test8(<16 x i32>%a, <16 x i32>%b, i32 %a1, i32 %b1) {
671 ; KNL-LABEL: test8:
672 ; KNL: ## %bb.0:
673 ; KNL-NEXT: cmpl %esi, %edi
674 ; KNL-NEXT: jg LBB17_1
675 ; KNL-NEXT: ## %bb.2:
676 ; KNL-NEXT: kxorw %k0, %k0, %k1
677 ; KNL-NEXT: jmp LBB17_3
678 ; KNL-NEXT: LBB17_1:
679 ; KNL-NEXT: vpxor %xmm1, %xmm1, %xmm1
680 ; KNL-NEXT: vpcmpgtd %zmm1, %zmm0, %k1
681 ; KNL-NEXT: LBB17_3:
682 ; KNL-NEXT: vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
683 ; KNL-NEXT: vpmovdb %zmm0, %xmm0
684 ; KNL-NEXT: vzeroupper
685 ; KNL-NEXT: retq
686 ;
687 ; SKX-LABEL: test8:
688 ; SKX: ## %bb.0:
689 ; SKX-NEXT: cmpl %esi, %edi
690 ; SKX-NEXT: jg LBB17_1
691 ; SKX-NEXT: ## %bb.2:
692 ; SKX-NEXT: kxorw %k0, %k0, %k0
693 ; SKX-NEXT: vpmovm2b %k0, %xmm0
694 ; SKX-NEXT: vzeroupper
695 ; SKX-NEXT: retq
696 ; SKX-NEXT: LBB17_1:
697 ; SKX-NEXT: vpxor %xmm1, %xmm1, %xmm1
698 ; SKX-NEXT: vpcmpgtd %zmm1, %zmm0, %k0
699 ; SKX-NEXT: vpmovm2b %k0, %xmm0
700 ; SKX-NEXT: vzeroupper
701 ; SKX-NEXT: retq
702 ;
703 ; AVX512BW-LABEL: test8:
704 ; AVX512BW: ## %bb.0:
705 ; AVX512BW-NEXT: cmpl %esi, %edi
706 ; AVX512BW-NEXT: jg LBB17_1
707 ; AVX512BW-NEXT: ## %bb.2:
708 ; AVX512BW-NEXT: kxorw %k0, %k0, %k0
709 ; AVX512BW-NEXT: vpmovm2b %k0, %zmm0
710 ; AVX512BW-NEXT: ## kill: def $xmm0 killed $xmm0 killed $zmm0
711 ; AVX512BW-NEXT: vzeroupper
712 ; AVX512BW-NEXT: retq
713 ; AVX512BW-NEXT: LBB17_1:
714 ; AVX512BW-NEXT: vpxor %xmm1, %xmm1, %xmm1
715 ; AVX512BW-NEXT: vpcmpgtd %zmm1, %zmm0, %k0
716 ; AVX512BW-NEXT: vpmovm2b %k0, %zmm0
717 ; AVX512BW-NEXT: ## kill: def $xmm0 killed $xmm0 killed $zmm0
718 ; AVX512BW-NEXT: vzeroupper
719 ; AVX512BW-NEXT: retq
720 ;
721 ; AVX512DQ-LABEL: test8:
722 ; AVX512DQ: ## %bb.0:
723 ; AVX512DQ-NEXT: cmpl %esi, %edi
724 ; AVX512DQ-NEXT: jg LBB17_1
725 ; AVX512DQ-NEXT: ## %bb.2:
726 ; AVX512DQ-NEXT: kxorw %k0, %k0, %k0
727 ; AVX512DQ-NEXT: jmp LBB17_3
728 ; AVX512DQ-NEXT: LBB17_1:
729 ; AVX512DQ-NEXT: vpxor %xmm1, %xmm1, %xmm1
730 ; AVX512DQ-NEXT: vpcmpgtd %zmm1, %zmm0, %k0
731 ; AVX512DQ-NEXT: LBB17_3:
732 ; AVX512DQ-NEXT: vpmovm2d %k0, %zmm0
733 ; AVX512DQ-NEXT: vpmovdb %zmm0, %xmm0
734 ; AVX512DQ-NEXT: vzeroupper
735 ; AVX512DQ-NEXT: retq
736 ;
737 ; X86-LABEL: test8:
738 ; X86: ## %bb.0:
739 ; X86-NEXT: movl {{[0-9]+}}(%esp), %eax
740 ; X86-NEXT: cmpl {{[0-9]+}}(%esp), %eax
741 ; X86-NEXT: jg LBB17_1
742 ; X86-NEXT: ## %bb.2:
743 ; X86-NEXT: kxorw %k0, %k0, %k0
744 ; X86-NEXT: vpmovm2b %k0, %xmm0
745 ; X86-NEXT: vzeroupper
746 ; X86-NEXT: retl
747 ; X86-NEXT: LBB17_1:
748 ; X86-NEXT: vpxor %xmm1, %xmm1, %xmm1
749 ; X86-NEXT: vpcmpgtd %zmm1, %zmm0, %k0
750 ; X86-NEXT: vpmovm2b %k0, %xmm0
751 ; X86-NEXT: vzeroupper
752 ; X86-NEXT: retl
753 %cond = icmp sgt i32 %a1, %b1
754 %cmp1 = icmp sgt <16 x i32> %a, zeroinitializer
755 %cmp2 = icmp ult <16 x i32> %b, zeroinitializer
756 %mix = select i1 %cond, <16 x i1> %cmp1, <16 x i1> %cmp2
757 %res = sext <16 x i1> %mix to <16 x i8>
758 ret <16 x i8> %res
759 }
760 define <16 x i1> @test9(<16 x i1>%a, <16 x i1>%b, i32 %a1, i32 %b1) {
761 ; KNL-LABEL: test9:
762 ; KNL: ## %bb.0:
763 ; KNL-NEXT: cmpl %esi, %edi
764 ; KNL-NEXT: jg LBB18_1
765 ; KNL-NEXT: ## %bb.2:
766 ; KNL-NEXT: vpmovsxbd %xmm1, %zmm0
767 ; KNL-NEXT: jmp LBB18_3
768 ; KNL-NEXT: LBB18_1:
769 ; KNL-NEXT: vpmovsxbd %xmm0, %zmm0
770 ; KNL-NEXT: LBB18_3:
771 ; KNL-NEXT: vpslld $31, %zmm0, %zmm0
772 ; KNL-NEXT: vptestmd %zmm0, %zmm0, %k1
773 ; KNL-NEXT: vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
774 ; KNL-NEXT: vpmovdb %zmm0, %xmm0
775 ; KNL-NEXT: vzeroupper
776 ; KNL-NEXT: retq
777 ;
778 ; SKX-LABEL: test9:
779 ; SKX: ## %bb.0:
780 ; SKX-NEXT: cmpl %esi, %edi
781 ; SKX-NEXT: jg LBB18_1
782 ; SKX-NEXT: ## %bb.2:
783 ; SKX-NEXT: vpsllw $7, %xmm1, %xmm0
784 ; SKX-NEXT: jmp LBB18_3
785 ; SKX-NEXT: LBB18_1:
786 ; SKX-NEXT: vpsllw $7, %xmm0, %xmm0
787 ; SKX-NEXT: LBB18_3:
788 ; SKX-NEXT: vpmovb2m %xmm0, %k0
789 ; SKX-NEXT: vpmovm2b %k0, %xmm0
790 ; SKX-NEXT: retq
791 ;
792 ; AVX512BW-LABEL: test9:
793 ; AVX512BW: ## %bb.0:
794 ; AVX512BW-NEXT: cmpl %esi, %edi
795 ; AVX512BW-NEXT: jg LBB18_1
796 ; AVX512BW-NEXT: ## %bb.2:
797 ; AVX512BW-NEXT: vpsllw $7, %xmm1, %xmm0
798 ; AVX512BW-NEXT: jmp LBB18_3
799 ; AVX512BW-NEXT: LBB18_1:
800 ; AVX512BW-NEXT: vpsllw $7, %xmm0, %xmm0
801 ; AVX512BW-NEXT: LBB18_3:
802 ; AVX512BW-NEXT: vpmovb2m %zmm0, %k0
803 ; AVX512BW-NEXT: vpmovm2b %k0, %zmm0
804 ; AVX512BW-NEXT: ## kill: def $xmm0 killed $xmm0 killed $zmm0
805 ; AVX512BW-NEXT: vzeroupper
806 ; AVX512BW-NEXT: retq
807 ;
808 ; AVX512DQ-LABEL: test9:
809 ; AVX512DQ: ## %bb.0:
810 ; AVX512DQ-NEXT: cmpl %esi, %edi
811 ; AVX512DQ-NEXT: jg LBB18_1
812 ; AVX512DQ-NEXT: ## %bb.2:
813 ; AVX512DQ-NEXT: vpmovsxbd %xmm1, %zmm0
814 ; AVX512DQ-NEXT: jmp LBB18_3
815 ; AVX512DQ-NEXT: LBB18_1:
816 ; AVX512DQ-NEXT: vpmovsxbd %xmm0, %zmm0
817 ; AVX512DQ-NEXT: LBB18_3:
818 ; AVX512DQ-NEXT: vpslld $31, %zmm0, %zmm0
819 ; AVX512DQ-NEXT: vpmovd2m %zmm0, %k0
820 ; AVX512DQ-NEXT: vpmovm2d %k0, %zmm0
821 ; AVX512DQ-NEXT: vpmovdb %zmm0, %xmm0
822 ; AVX512DQ-NEXT: vzeroupper
823 ; AVX512DQ-NEXT: retq
824 ;
825 ; X86-LABEL: test9:
826 ; X86: ## %bb.0:
827 ; X86-NEXT: movl {{[0-9]+}}(%esp), %eax
828 ; X86-NEXT: cmpl {{[0-9]+}}(%esp), %eax
829 ; X86-NEXT: jg LBB18_1
830 ; X86-NEXT: ## %bb.2:
831 ; X86-NEXT: vpsllw $7, %xmm1, %xmm0
832 ; X86-NEXT: jmp LBB18_3
833 ; X86-NEXT: LBB18_1:
834 ; X86-NEXT: vpsllw $7, %xmm0, %xmm0
835 ; X86-NEXT: LBB18_3:
836 ; X86-NEXT: vpmovb2m %xmm0, %k0
837 ; X86-NEXT: vpmovm2b %k0, %xmm0
838 ; X86-NEXT: retl
839 %mask = icmp sgt i32 %a1, %b1
840 %c = select i1 %mask, <16 x i1>%a, <16 x i1>%b
841 ret <16 x i1>%c
842 }define <8 x i1> @test10(<8 x i1>%a, <8 x i1>%b, i32 %a1, i32 %b1) {
843 %mask = icmp sgt i32 %a1, %b1
844 %c = select i1 %mask, <8 x i1>%a, <8 x i1>%b
845 ret <8 x i1>%c
846 }
848 define <4 x i1> @test11(<4 x i1>%a, <4 x i1>%b, i32 %a1, i32 %b1) {
849 ; KNL-LABEL: test11:
850 ; KNL: ## %bb.0:
851 ; KNL-NEXT: cmpl %esi, %edi
852 ; KNL-NEXT: jg LBB20_1
853 ; KNL-NEXT: ## %bb.2:
854 ; KNL-NEXT: vpslld $31, %xmm1, %xmm0
855 ; KNL-NEXT: jmp LBB20_3
856 ; KNL-NEXT: LBB20_1:
857 ; KNL-NEXT: vpslld $31, %xmm0, %xmm0
858 ; KNL-NEXT: LBB20_3:
859 ; KNL-NEXT: vptestmd %zmm0, %zmm0, %k1
860 ; KNL-NEXT: vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
861 ; KNL-NEXT: ## kill: def $xmm0 killed $xmm0 killed $zmm0
862 ; KNL-NEXT: vzeroupper
863 ; KNL-NEXT: retq
864 ;
865 ; SKX-LABEL: test11:
866 ; SKX: ## %bb.0:
867 ; SKX-NEXT: cmpl %esi, %edi
868 ; SKX-NEXT: jg LBB20_1
869 ; SKX-NEXT: ## %bb.2:
870 ; SKX-NEXT: vpslld $31, %xmm1, %xmm0
871 ; SKX-NEXT: jmp LBB20_3
872 ; SKX-NEXT: LBB20_1:
873 ; SKX-NEXT: vpslld $31, %xmm0, %xmm0
874 ; SKX-NEXT: LBB20_3:
875 ; SKX-NEXT: vpmovd2m %xmm0, %k0
876 ; SKX-NEXT: vpmovm2d %k0, %xmm0
877 ; SKX-NEXT: retq
878 ;
879 ; AVX512BW-LABEL: test11:
880 ; AVX512BW: ## %bb.0:
881 ; AVX512BW-NEXT: cmpl %esi, %edi
882 ; AVX512BW-NEXT: jg LBB20_1
883 ; AVX512BW-NEXT: ## %bb.2:
884 ; AVX512BW-NEXT: vpslld $31, %xmm1, %xmm0
885 ; AVX512BW-NEXT: jmp LBB20_3
886 ; AVX512BW-NEXT: LBB20_1:
887 ; AVX512BW-NEXT: vpslld $31, %xmm0, %xmm0
888 ; AVX512BW-NEXT: LBB20_3:
889 ; AVX512BW-NEXT: vptestmd %zmm0, %zmm0, %k1
890 ; AVX512BW-NEXT: vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
891 ; AVX512BW-NEXT: ## kill: def $xmm0 killed $xmm0 killed $zmm0
892 ; AVX512BW-NEXT: vzeroupper
893 ; AVX512BW-NEXT: retq
894 ;
895 ; AVX512DQ-LABEL: test11:
896 ; AVX512DQ: ## %bb.0:
897 ; AVX512DQ-NEXT: cmpl %esi, %edi
898 ; AVX512DQ-NEXT: jg LBB20_1
899 ; AVX512DQ-NEXT: ## %bb.2:
900 ; AVX512DQ-NEXT: vpslld $31, %xmm1, %xmm0
901 ; AVX512DQ-NEXT: jmp LBB20_3
902 ; AVX512DQ-NEXT: LBB20_1:
903 ; AVX512DQ-NEXT: vpslld $31, %xmm0, %xmm0
904 ; AVX512DQ-NEXT: LBB20_3:
905 ; AVX512DQ-NEXT: vpmovd2m %zmm0, %k0
906 ; AVX512DQ-NEXT: vpmovm2d %k0, %zmm0
907 ; AVX512DQ-NEXT: ## kill: def $xmm0 killed $xmm0 killed $zmm0
908 ; AVX512DQ-NEXT: vzeroupper
909 ; AVX512DQ-NEXT: retq
910 ;
911 ; X86-LABEL: test11:
912 ; X86: ## %bb.0:
913 ; X86-NEXT: movl {{[0-9]+}}(%esp), %eax
914 ; X86-NEXT: cmpl {{[0-9]+}}(%esp), %eax
915 ; X86-NEXT: jg LBB20_1
916 ; X86-NEXT: ## %bb.2:
917 ; X86-NEXT: vpslld $31, %xmm1, %xmm0
918 ; X86-NEXT: jmp LBB20_3
919 ; X86-NEXT: LBB20_1:
920 ; X86-NEXT: vpslld $31, %xmm0, %xmm0
921 ; X86-NEXT: LBB20_3:
922 ; X86-NEXT: vpmovd2m %xmm0, %k0
923 ; X86-NEXT: vpmovm2d %k0, %xmm0
924 ; X86-NEXT: retl
925 %mask = icmp sgt i32 %a1, %b1
926 %c = select i1 %mask, <4 x i1>%a, <4 x i1>%b
927 ret <4 x i1>%c
928 }
930 define i32 @test12(i32 %x, i32 %y) {
931 ; CHECK-LABEL: test12:
932 ; CHECK: ## %bb.0:
933 ; CHECK-NEXT: movl %edi, %eax
934 ; CHECK-NEXT: retq
935 ;
936 ; X86-LABEL: test12:
937 ; X86: ## %bb.0:
938 ; X86-NEXT: movl {{[0-9]+}}(%esp), %eax
939 ; X86-NEXT: retl
940 %a = bitcast i16 21845 to <16 x i1>
941 %b = extractelement <16 x i1> %a, i32 0
942 %c = select i1 %b, i32 %x, i32 %y
943 ret i32 %c
944 }
946 define i32 @test13(i32 %x, i32 %y) {
947 ; CHECK-LABEL: test13:
948 ; CHECK: ## %bb.0:
949 ; CHECK-NEXT: movl %esi, %eax
950 ; CHECK-NEXT: retq
951 ;
952 ; X86-LABEL: test13:
953 ; X86: ## %bb.0:
954 ; X86-NEXT: movl {{[0-9]+}}(%esp), %eax
955 ; X86-NEXT: retl
956 %a = bitcast i16 21845 to <16 x i1>
957 %b = extractelement <16 x i1> %a, i32 3
958 %c = select i1 %b, i32 %x, i32 %y
959 ret i32 %c
960 }
962 ; Make sure we don't crash on a large vector.
963 define i32 @test13_crash(i32 %x, i32 %y) {
964 ; CHECK-LABEL: test13_crash:
965 ; CHECK: ## %bb.0:
966 ; CHECK-NEXT: movl %edi, %eax
967 ; CHECK-NEXT: retq
968 ;
969 ; X86-LABEL: test13_crash:
970 ; X86: ## %bb.0:
971 ; X86-NEXT: movl {{[0-9]+}}(%esp), %eax
972 ; X86-NEXT: retl
973 %a = bitcast i128 2184568686868686868686868686 to <128 x i1>
974 %b = extractelement <128 x i1> %a, i32 3
975 %c = select i1 %b, i32 %x, i32 %y
976 ret i32 %c
977 }
979 define <4 x i1> @test14() {
980 ; CHECK-LABEL: test14:
981 ; CHECK: ## %bb.0:
982 ; CHECK-NEXT: vmovaps {{.*#+}} xmm0 = [1,1,0,1]
983 ; CHECK-NEXT: retq
984 ;
985 ; X86-LABEL: test14:
986 ; X86: ## %bb.0:
987 ; X86-NEXT: vmovaps {{.*#+}} xmm0 = [1,1,0,1]
988 ; X86-NEXT: retl
989 %a = bitcast i16 21845 to <16 x i1>
990 %b = extractelement <16 x i1> %a, i32 2
991 %c = insertelement <4 x i1> <i1 true, i1 false, i1 false, i1 true>, i1 %b, i32 1
992 ret <4 x i1> %c
993 }
995 define <16 x i1> @test15(i32 %x, i32 %y) {
996 ; KNL-LABEL: test15:
997 ; KNL: ## %bb.0:
998 ; KNL-NEXT: cmpl %esi, %edi
999 ; KNL-NEXT: movl $21845, %eax ## imm = 0x5555
1000 ; KNL-NEXT: movl $1, %ecx
1001 ; KNL-NEXT: cmovgl %eax, %ecx
1002 ; KNL-NEXT: kmovw %ecx, %k1
1003 ; KNL-NEXT: vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
1004 ; KNL-NEXT: vpmovdb %zmm0, %xmm0
1005 ; KNL-NEXT: vzeroupper
1006 ; KNL-NEXT: retq
1007 ;
1008 ; SKX-LABEL: test15:
1009 ; SKX: ## %bb.0:
1010 ; SKX-NEXT: cmpl %esi, %edi
1011 ; SKX-NEXT: movl $21845, %eax ## imm = 0x5555
1012 ; SKX-NEXT: movl $1, %ecx
1013 ; SKX-NEXT: cmovgl %eax, %ecx
1014 ; SKX-NEXT: kmovd %ecx, %k0
1015 ; SKX-NEXT: vpmovm2b %k0, %xmm0
1016 ; SKX-NEXT: retq
1017 ;
1018 ; AVX512BW-LABEL: test15:
1019 ; AVX512BW: ## %bb.0:
1020 ; AVX512BW-NEXT: cmpl %esi, %edi
1021 ; AVX512BW-NEXT: movl $21845, %eax ## imm = 0x5555
1022 ; AVX512BW-NEXT: movl $1, %ecx
1023 ; AVX512BW-NEXT: cmovgl %eax, %ecx
1024 ; AVX512BW-NEXT: kmovd %ecx, %k0
1025 ; AVX512BW-NEXT: vpmovm2b %k0, %zmm0
1026 ; AVX512BW-NEXT: ## kill: def $xmm0 killed $xmm0 killed $zmm0
1027 ; AVX512BW-NEXT: vzeroupper
1028 ; AVX512BW-NEXT: retq
1029 ;
1030 ; AVX512DQ-LABEL: test15:
1031 ; AVX512DQ: ## %bb.0:
1032 ; AVX512DQ-NEXT: cmpl %esi, %edi
1033 ; AVX512DQ-NEXT: movl $21845, %eax ## imm = 0x5555
1034 ; AVX512DQ-NEXT: movl $1, %ecx
1035 ; AVX512DQ-NEXT: cmovgl %eax, %ecx
1036 ; AVX512DQ-NEXT: kmovw %ecx, %k0
1037 ; AVX512DQ-NEXT: vpmovm2d %k0, %zmm0
1038 ; AVX512DQ-NEXT: vpmovdb %zmm0, %xmm0
1039 ; AVX512DQ-NEXT: vzeroupper
1040 ; AVX512DQ-NEXT: retq
1041 ;
1042 ; X86-LABEL: test15:
1043 ; X86: ## %bb.0:
1044 ; X86-NEXT: movl {{[0-9]+}}(%esp), %eax
1045 ; X86-NEXT: cmpl {{[0-9]+}}(%esp), %eax
1046 ; X86-NEXT: movl $21845, %eax ## imm = 0x5555
1047 ; X86-NEXT: movl $1, %ecx
1048 ; X86-NEXT: cmovgl %eax, %ecx
1049 ; X86-NEXT: kmovd %ecx, %k0
1050 ; X86-NEXT: vpmovm2b %k0, %xmm0
1051 ; X86-NEXT: retl
1052 %a = bitcast i16 21845 to <16 x i1>
1053 %b = bitcast i16 1 to <16 x i1>
1054 %mask = icmp sgt i32 %x, %y
1055 %c = select i1 %mask, <16 x i1> %a, <16 x i1> %b
1056 ret <16 x i1> %c
1057 }
1059 define <64 x i8> @test16(i64 %x) {
1060 ;
1061 ; KNL-LABEL: test16:
1062 ; KNL: ## %bb.0:
1063 ; KNL-NEXT: movq %rdi, %rax
1064 ; KNL-NEXT: movl %edi, %ecx
1065 ; KNL-NEXT: kmovw %edi, %k0
1066 ; KNL-NEXT: shrq $32, %rdi
1067 ; KNL-NEXT: shrq $48, %rax
1068 ; KNL-NEXT: shrl $16, %ecx
1069 ; KNL-NEXT: kmovw %ecx, %k1
1070 ; KNL-NEXT: kmovw %eax, %k2
1071 ; KNL-NEXT: kmovw %edi, %k3
1072 ; KNL-NEXT: kshiftrw $6, %k0, %k4
1073 ; KNL-NEXT: kshiftlw $6, %k4, %k4
1074 ; KNL-NEXT: kshiftlw $11, %k0, %k0
1075 ; KNL-NEXT: kshiftrw $11, %k0, %k0
1076 ; KNL-NEXT: movb $1, %al
1077 ; KNL-NEXT: kmovw %eax, %k5
1078 ; KNL-NEXT: kshiftlw $15, %k5, %k5
1079 ; KNL-NEXT: kshiftrw $10, %k5, %k5
1080 ; KNL-NEXT: korw %k5, %k4, %k4
1081 ; KNL-NEXT: korw %k4, %k0, %k4
1082 ; KNL-NEXT: vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k3} {z}
1083 ; KNL-NEXT: vpmovdb %zmm0, %xmm0
1084 ; KNL-NEXT: vpternlogd $255, %zmm1, %zmm1, %zmm1 {%k2} {z}
1085 ; KNL-NEXT: vpmovdb %zmm1, %xmm1
1086 ; KNL-NEXT: vinserti128 $1, %xmm1, %ymm0, %ymm0
1087 ; KNL-NEXT: vpternlogd $255, %zmm1, %zmm1, %zmm1 {%k4} {z}
1088 ; KNL-NEXT: vpmovdb %zmm1, %xmm1
1089 ; KNL-NEXT: vpternlogd $255, %zmm2, %zmm2, %zmm2 {%k1} {z}
1090 ; KNL-NEXT: vpmovdb %zmm2, %xmm2
1091 ; KNL-NEXT: vinserti128 $1, %xmm2, %ymm1, %ymm1
1092 ; KNL-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm0
1093 ; KNL-NEXT: retq
1094 ;
1095 ; SKX-LABEL: test16:
1096 ; SKX: ## %bb.0:
1097 ; SKX-NEXT: kmovq %rdi, %k0
1098 ; SKX-NEXT: kshiftrq $6, %k0, %k1
1099 ; SKX-NEXT: kshiftlq $6, %k1, %k1
1100 ; SKX-NEXT: kshiftlq $59, %k0, %k0
1101 ; SKX-NEXT: kshiftrq $59, %k0, %k0
1102 ; SKX-NEXT: movb $1, %al
1103 ; SKX-NEXT: kmovd %eax, %k2
1104 ; SKX-NEXT: kshiftlq $63, %k2, %k2
1105 ; SKX-NEXT: kshiftrq $58, %k2, %k2
1106 ; SKX-NEXT: korq %k2, %k1, %k1
1107 ; SKX-NEXT: korq %k1, %k0, %k0
1108 ; SKX-NEXT: vpmovm2b %k0, %zmm0
1109 ; SKX-NEXT: retq
1110 ;
1111 ; AVX512BW-LABEL: test16:
1112 ; AVX512BW: ## %bb.0:
1113 ; AVX512BW-NEXT: kmovq %rdi, %k0
1114 ; AVX512BW-NEXT: kshiftrq $6, %k0, %k1
1115 ; AVX512BW-NEXT: kshiftlq $6, %k1, %k1
1116 ; AVX512BW-NEXT: kshiftlq $59, %k0, %k0
1117 ; AVX512BW-NEXT: kshiftrq $59, %k0, %k0
1118 ; AVX512BW-NEXT: movb $1, %al
1119 ; AVX512BW-NEXT: kmovd %eax, %k2
1120 ; AVX512BW-NEXT: kshiftlq $63, %k2, %k2
1121 ; AVX512BW-NEXT: kshiftrq $58, %k2, %k2
1122 ; AVX512BW-NEXT: korq %k2, %k1, %k1
1123 ; AVX512BW-NEXT: korq %k1, %k0, %k0
1124 ; AVX512BW-NEXT: vpmovm2b %k0, %zmm0
1125 ; AVX512BW-NEXT: retq
1126 ;
1127 ; AVX512DQ-LABEL: test16:
1128 ; AVX512DQ: ## %bb.0:
1129 ; AVX512DQ-NEXT: movq %rdi, %rax
1130 ; AVX512DQ-NEXT: movl %edi, %ecx
1131 ; AVX512DQ-NEXT: kmovw %edi, %k1
1132 ; AVX512DQ-NEXT: shrq $32, %rdi
1133 ; AVX512DQ-NEXT: shrq $48, %rax
1134 ; AVX512DQ-NEXT: shrl $16, %ecx
1135 ; AVX512DQ-NEXT: kmovw %ecx, %k0
1136 ; AVX512DQ-NEXT: kmovw %eax, %k2
1137 ; AVX512DQ-NEXT: kmovw %edi, %k3
1138 ; AVX512DQ-NEXT: kshiftrw $6, %k1, %k4
1139 ; AVX512DQ-NEXT: kshiftlw $6, %k4, %k4
1140 ; AVX512DQ-NEXT: kshiftlw $11, %k1, %k1
1141 ; AVX512DQ-NEXT: kshiftrw $11, %k1, %k1
1142 ; AVX512DQ-NEXT: movb $1, %al
1143 ; AVX512DQ-NEXT: kmovw %eax, %k5
1144 ; AVX512DQ-NEXT: kshiftlw $15, %k5, %k5
1145 ; AVX512DQ-NEXT: kshiftrw $10, %k5, %k5
1146 ; AVX512DQ-NEXT: korw %k5, %k4, %k4
1147 ; AVX512DQ-NEXT: korw %k4, %k1, %k1
1148 ; AVX512DQ-NEXT: vpmovm2d %k3, %zmm0
1149 ; AVX512DQ-NEXT: vpmovdb %zmm0, %xmm0
1150 ; AVX512DQ-NEXT: vpmovm2d %k2, %zmm1
1151 ; AVX512DQ-NEXT: vpmovdb %zmm1, %xmm1
1152 ; AVX512DQ-NEXT: vinserti128 $1, %xmm1, %ymm0, %ymm0
1153 ; AVX512DQ-NEXT: vpmovm2d %k1, %zmm1
1154 ; AVX512DQ-NEXT: vpmovdb %zmm1, %xmm1
1155 ; AVX512DQ-NEXT: vpmovm2d %k0, %zmm2
1156 ; AVX512DQ-NEXT: vpmovdb %zmm2, %xmm2
1157 ; AVX512DQ-NEXT: vinserti128 $1, %xmm2, %ymm1, %ymm1
1158 ; AVX512DQ-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm0
1159 ; AVX512DQ-NEXT: retq
1160 ;
1161 ; X86-LABEL: test16:
1162 ; X86: ## %bb.0:
1163 ; X86-NEXT: kmovq {{[0-9]+}}(%esp), %k0
1164 ; X86-NEXT: kshiftrq $6, %k0, %k1
1165 ; X86-NEXT: kshiftlq $6, %k1, %k1
1166 ; X86-NEXT: kshiftlq $59, %k0, %k0
1167 ; X86-NEXT: kshiftrq $59, %k0, %k0
1168 ; X86-NEXT: movb $1, %al
1169 ; X86-NEXT: kmovd %eax, %k2
1170 ; X86-NEXT: kshiftlq $63, %k2, %k2
1171 ; X86-NEXT: kshiftrq $58, %k2, %k2
1172 ; X86-NEXT: korq %k2, %k1, %k1
1173 ; X86-NEXT: korq %k1, %k0, %k0
1174 ; X86-NEXT: vpmovm2b %k0, %zmm0
1175 ; X86-NEXT: retl
1176 %a = bitcast i64 %x to <64 x i1>
1177 %b = insertelement <64 x i1>%a, i1 true, i32 5
1178 %c = sext <64 x i1>%b to <64 x i8>
1179 ret <64 x i8>%c
1180 }
1182 define <64 x i8> @test17(i64 %x, i32 %y, i32 %z) {
1183 ;
1184 ; KNL-LABEL: test17:
1185 ; KNL: ## %bb.0:
1186 ; KNL-NEXT: movq %rdi, %rax
1187 ; KNL-NEXT: movl %edi, %ecx
1188 ; KNL-NEXT: kmovw %edi, %k0
1189 ; KNL-NEXT: shrq $32, %rdi
1190 ; KNL-NEXT: shrq $48, %rax
1191 ; KNL-NEXT: shrl $16, %ecx
1192 ; KNL-NEXT: kmovw %ecx, %k1
1193 ; KNL-NEXT: kmovw %eax, %k2
1194 ; KNL-NEXT: kmovw %edi, %k3
1195 ; KNL-NEXT: cmpl %edx, %esi
1196 ; KNL-NEXT: setg %al
1197 ; KNL-NEXT: kshiftrw $6, %k0, %k4
1198 ; KNL-NEXT: kshiftlw $6, %k4, %k4
1199 ; KNL-NEXT: kshiftlw $11, %k0, %k0
1200 ; KNL-NEXT: kshiftrw $11, %k0, %k0
1201 ; KNL-NEXT: kmovw %eax, %k5
1202 ; KNL-NEXT: kshiftlw $15, %k5, %k5
1203 ; KNL-NEXT: kshiftrw $10, %k5, %k5
1204 ; KNL-NEXT: korw %k5, %k4, %k4
1205 ; KNL-NEXT: korw %k4, %k0, %k4
1206 ; KNL-NEXT: vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k3} {z}
1207 ; KNL-NEXT: vpmovdb %zmm0, %xmm0
1208 ; KNL-NEXT: vpternlogd $255, %zmm1, %zmm1, %zmm1 {%k2} {z}
1209 ; KNL-NEXT: vpmovdb %zmm1, %xmm1
1210 ; KNL-NEXT: vinserti128 $1, %xmm1, %ymm0, %ymm0
1211 ; KNL-NEXT: vpternlogd $255, %zmm1, %zmm1, %zmm1 {%k4} {z}
1212 ; KNL-NEXT: vpmovdb %zmm1, %xmm1
1213 ; KNL-NEXT: vpternlogd $255, %zmm2, %zmm2, %zmm2 {%k1} {z}
1214 ; KNL-NEXT: vpmovdb %zmm2, %xmm2
1215 ; KNL-NEXT: vinserti128 $1, %xmm2, %ymm1, %ymm1
1216 ; KNL-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm0
1217 ; KNL-NEXT: retq
1218 ;
1219 ; SKX-LABEL: test17:
1220 ; SKX: ## %bb.0:
1221 ; SKX-NEXT: kmovq %rdi, %k0
1222 ; SKX-NEXT: cmpl %edx, %esi
1223 ; SKX-NEXT: setg %al
1224 ; SKX-NEXT: kshiftrq $6, %k0, %k1
1225 ; SKX-NEXT: kshiftlq $6, %k1, %k1
1226 ; SKX-NEXT: kshiftlq $59, %k0, %k0
1227 ; SKX-NEXT: kshiftrq $59, %k0, %k0
1228 ; SKX-NEXT: kmovd %eax, %k2
1229 ; SKX-NEXT: kshiftlq $63, %k2, %k2
1230 ; SKX-NEXT: kshiftrq $58, %k2, %k2
1231 ; SKX-NEXT: korq %k2, %k1, %k1
1232 ; SKX-NEXT: korq %k1, %k0, %k0
1233 ; SKX-NEXT: vpmovm2b %k0, %zmm0
1234 ; SKX-NEXT: retq
1235 ;
1236 ; AVX512BW-LABEL: test17:
1237 ; AVX512BW: ## %bb.0:
1238 ; AVX512BW-NEXT: kmovq %rdi, %k0
1239 ; AVX512BW-NEXT: cmpl %edx, %esi
1240 ; AVX512BW-NEXT: setg %al
1241 ; AVX512BW-NEXT: kshiftrq $6, %k0, %k1
1242 ; AVX512BW-NEXT: kshiftlq $6, %k1, %k1
1243 ; AVX512BW-NEXT: kshiftlq $59, %k0, %k0
1244 ; AVX512BW-NEXT: kshiftrq $59, %k0, %k0
1245 ; AVX512BW-NEXT: kmovd %eax, %k2
1246 ; AVX512BW-NEXT: kshiftlq $63, %k2, %k2
1247 ; AVX512BW-NEXT: kshiftrq $58, %k2, %k2
1248 ; AVX512BW-NEXT: korq %k2, %k1, %k1
1249 ; AVX512BW-NEXT: korq %k1, %k0, %k0
1250 ; AVX512BW-NEXT: vpmovm2b %k0, %zmm0
1251 ; AVX512BW-NEXT: retq
1252 ;
1253 ; AVX512DQ-LABEL: test17:
1254 ; AVX512DQ: ## %bb.0:
1255 ; AVX512DQ-NEXT: movq %rdi, %rax
1256 ; AVX512DQ-NEXT: movl %edi, %ecx
1257 ; AVX512DQ-NEXT: kmovw %edi, %k1
1258 ; AVX512DQ-NEXT: shrq $32, %rdi
1259 ; AVX512DQ-NEXT: shrq $48, %rax
1260 ; AVX512DQ-NEXT: shrl $16, %ecx
1261 ; AVX512DQ-NEXT: kmovw %ecx, %k0
1262 ; AVX512DQ-NEXT: kmovw %eax, %k2
1263 ; AVX512DQ-NEXT: kmovw %edi, %k3
1264 ; AVX512DQ-NEXT: cmpl %edx, %esi
1265 ; AVX512DQ-NEXT: setg %al
1266 ; AVX512DQ-NEXT: kshiftrw $6, %k1, %k4
1267 ; AVX512DQ-NEXT: kshiftlw $6, %k4, %k4
1268 ; AVX512DQ-NEXT: kshiftlw $11, %k1, %k1
1269 ; AVX512DQ-NEXT: kshiftrw $11, %k1, %k1
1270 ; AVX512DQ-NEXT: kmovw %eax, %k5
1271 ; AVX512DQ-NEXT: kshiftlw $15, %k5, %k5
1272 ; AVX512DQ-NEXT: kshiftrw $10, %k5, %k5
1273 ; AVX512DQ-NEXT: korw %k5, %k4, %k4
1274 ; AVX512DQ-NEXT: korw %k4, %k1, %k1
1275 ; AVX512DQ-NEXT: vpmovm2d %k3, %zmm0
1276 ; AVX512DQ-NEXT: vpmovdb %zmm0, %xmm0
1277 ; AVX512DQ-NEXT: vpmovm2d %k2, %zmm1
1278 ; AVX512DQ-NEXT: vpmovdb %zmm1, %xmm1
1279 ; AVX512DQ-NEXT: vinserti128 $1, %xmm1, %ymm0, %ymm0
1280 ; AVX512DQ-NEXT: vpmovm2d %k1, %zmm1
1281 ; AVX512DQ-NEXT: vpmovdb %zmm1, %xmm1
1282 ; AVX512DQ-NEXT: vpmovm2d %k0, %zmm2
1283 ; AVX512DQ-NEXT: vpmovdb %zmm2, %xmm2
1284 ; AVX512DQ-NEXT: vinserti128 $1, %xmm2, %ymm1, %ymm1
1285 ; AVX512DQ-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm0
1286 ; AVX512DQ-NEXT: retq
1287 ;
1288 ; X86-LABEL: test17:
1289 ; X86: ## %bb.0:
1290 ; X86-NEXT: movl {{[0-9]+}}(%esp), %eax
1291 ; X86-NEXT: kmovq {{[0-9]+}}(%esp), %k0
1292 ; X86-NEXT: cmpl {{[0-9]+}}(%esp), %eax
1293 ; X86-NEXT: setg %al
1294 ; X86-NEXT: kshiftrq $6, %k0, %k1
1295 ; X86-NEXT: kshiftlq $6, %k1, %k1
1296 ; X86-NEXT: kshiftlq $59, %k0, %k0
1297 ; X86-NEXT: kshiftrq $59, %k0, %k0
1298 ; X86-NEXT: kmovd %eax, %k2
1299 ; X86-NEXT: kshiftlq $63, %k2, %k2
1300 ; X86-NEXT: kshiftrq $58, %k2, %k2
1301 ; X86-NEXT: korq %k2, %k1, %k1
1302 ; X86-NEXT: korq %k1, %k0, %k0
1303 ; X86-NEXT: vpmovm2b %k0, %zmm0
1304 ; X86-NEXT: retl
1305 %a = bitcast i64 %x to <64 x i1>
1306 %b = icmp sgt i32 %y, %z
1307 %c = insertelement <64 x i1>%a, i1 %b, i32 5
1308 %d = sext <64 x i1>%c to <64 x i8>
1309 ret <64 x i8>%d
1310 }
1312 define <8 x i1> @test18(i8 %a, i16 %y) {
1313 ; KNL-LABEL: test18:
1314 ; KNL: ## %bb.0:
1315 ; KNL-NEXT: kmovw %edi, %k0
1316 ; KNL-NEXT: kmovw %esi, %k1
1317 ; KNL-NEXT: kshiftrw $8, %k1, %k2
1318 ; KNL-NEXT: kshiftrw $9, %k1, %k1
1319 ; KNL-NEXT: kshiftlw $10, %k0, %k0
1320 ; KNL-NEXT: kshiftrw $10, %k0, %k0
1321 ; KNL-NEXT: kshiftlw $7, %k0, %k3
1322 ; KNL-NEXT: kshiftlw $6, %k1, %k1
1323 ; KNL-NEXT: korw %k1, %k3, %k1
1324 ; KNL-NEXT: korw %k1, %k0, %k0
1325 ; KNL-NEXT: kshiftlw $9, %k0, %k0
1326 ; KNL-NEXT: kshiftrw $9, %k0, %k0
1327 ; KNL-NEXT: kshiftlw $7, %k2, %k1
1328 ; KNL-NEXT: korw %k1, %k0, %k1
1329 ; KNL-NEXT: vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
1330 ; KNL-NEXT: vpmovdw %zmm0, %ymm0
1331 ; KNL-NEXT: ## kill: def $xmm0 killed $xmm0 killed $ymm0
1332 ; KNL-NEXT: vzeroupper
1333 ; KNL-NEXT: retq
1334 ;
1335 ; SKX-LABEL: test18:
1336 ; SKX: ## %bb.0:
1337 ; SKX-NEXT: kmovd %edi, %k0
1338 ; SKX-NEXT: kmovd %esi, %k1
1339 ; SKX-NEXT: kshiftrw $8, %k1, %k2
1340 ; SKX-NEXT: kshiftrw $9, %k1, %k1
1341 ; SKX-NEXT: kshiftlb $2, %k0, %k0
1342 ; SKX-NEXT: kshiftrb $2, %k0, %k0
1343 ; SKX-NEXT: kshiftlb $7, %k0, %k3
1344 ; SKX-NEXT: kshiftlb $6, %k1, %k1
1345 ; SKX-NEXT: korb %k1, %k3, %k1
1346 ; SKX-NEXT: korb %k1, %k0, %k0
1347 ; SKX-NEXT: kshiftlb $1, %k0, %k0
1348 ; SKX-NEXT: kshiftrb $1, %k0, %k0
1349 ; SKX-NEXT: kshiftlb $7, %k2, %k1
1350 ; SKX-NEXT: korb %k1, %k0, %k0
1351 ; SKX-NEXT: vpmovm2w %k0, %xmm0
1352 ; SKX-NEXT: retq
1353 ;
1354 ; AVX512BW-LABEL: test18:
1355 ; AVX512BW: ## %bb.0:
1356 ; AVX512BW-NEXT: kmovd %edi, %k0
1357 ; AVX512BW-NEXT: kmovd %esi, %k1
1358 ; AVX512BW-NEXT: kshiftrw $8, %k1, %k2
1359 ; AVX512BW-NEXT: kshiftrw $9, %k1, %k1
1360 ; AVX512BW-NEXT: kshiftlw $10, %k0, %k0
1361 ; AVX512BW-NEXT: kshiftrw $10, %k0, %k0
1362 ; AVX512BW-NEXT: kshiftlw $7, %k0, %k3
1363 ; AVX512BW-NEXT: kshiftlw $6, %k1, %k1
1364 ; AVX512BW-NEXT: korw %k1, %k3, %k1
1365 ; AVX512BW-NEXT: korw %k1, %k0, %k0
1366 ; AVX512BW-NEXT: kshiftlw $9, %k0, %k0
1367 ; AVX512BW-NEXT: kshiftrw $9, %k0, %k0
1368 ; AVX512BW-NEXT: kshiftlw $7, %k2, %k1
1369 ; AVX512BW-NEXT: korw %k1, %k0, %k0
1370 ; AVX512BW-NEXT: vpmovm2w %k0, %zmm0
1371 ; AVX512BW-NEXT: ## kill: def $xmm0 killed $xmm0 killed $zmm0
1372 ; AVX512BW-NEXT: vzeroupper
1373 ; AVX512BW-NEXT: retq
1374 ;
1375 ; AVX512DQ-LABEL: test18:
1376 ; AVX512DQ: ## %bb.0:
1377 ; AVX512DQ-NEXT: kmovw %edi, %k0
1378 ; AVX512DQ-NEXT: kmovw %esi, %k1
1379 ; AVX512DQ-NEXT: kshiftrw $8, %k1, %k2
1380 ; AVX512DQ-NEXT: kshiftrw $9, %k1, %k1
1381 ; AVX512DQ-NEXT: kshiftlb $2, %k0, %k0
1382 ; AVX512DQ-NEXT: kshiftrb $2, %k0, %k0
1383 ; AVX512DQ-NEXT: kshiftlb $7, %k0, %k3
1384 ; AVX512DQ-NEXT: kshiftlb $6, %k1, %k1
1385 ; AVX512DQ-NEXT: korb %k1, %k3, %k1
1386 ; AVX512DQ-NEXT: korb %k1, %k0, %k0
1387 ; AVX512DQ-NEXT: kshiftlb $1, %k0, %k0
1388 ; AVX512DQ-NEXT: kshiftrb $1, %k0, %k0
1389 ; AVX512DQ-NEXT: kshiftlb $7, %k2, %k1
1390 ; AVX512DQ-NEXT: korb %k1, %k0, %k0
1391 ; AVX512DQ-NEXT: vpmovm2d %k0, %zmm0
1392 ; AVX512DQ-NEXT: vpmovdw %zmm0, %ymm0
1393 ; AVX512DQ-NEXT: ## kill: def $xmm0 killed $xmm0 killed $ymm0
1394 ; AVX512DQ-NEXT: vzeroupper
1395 ; AVX512DQ-NEXT: retq
1396 ;
1397 ; X86-LABEL: test18:
1398 ; X86: ## %bb.0:
1399 ; X86-NEXT: kmovb {{[0-9]+}}(%esp), %k0
1400 ; X86-NEXT: kmovw {{[0-9]+}}(%esp), %k1
1401 ; X86-NEXT: kshiftrw $8, %k1, %k2
1402 ; X86-NEXT: kshiftrw $9, %k1, %k1
1403 ; X86-NEXT: kshiftlb $7, %k0, %k3
1404 ; X86-NEXT: kshiftlb $2, %k0, %k0
1405 ; X86-NEXT: kshiftrb $2, %k0, %k0
1406 ; X86-NEXT: kshiftlb $6, %k1, %k1
1407 ; X86-NEXT: korb %k1, %k3, %k1
1408 ; X86-NEXT: korb %k1, %k0, %k0
1409 ; X86-NEXT: kshiftlb $1, %k0, %k0
1410 ; X86-NEXT: kshiftrb $1, %k0, %k0
1411 ; X86-NEXT: kshiftlb $7, %k2, %k1
1412 ; X86-NEXT: korb %k1, %k0, %k0
1413 ; X86-NEXT: vpmovm2w %k0, %xmm0
1414 ; X86-NEXT: retl
1415 %b = bitcast i8 %a to <8 x i1>
1416 %b1 = bitcast i16 %y to <16 x i1>
1417 %el1 = extractelement <16 x i1>%b1, i32 8
1418 %el2 = extractelement <16 x i1>%b1, i32 9
1419 %c = insertelement <8 x i1>%b, i1 %el1, i32 7
1420 %d = insertelement <8 x i1>%c, i1 %el2, i32 6
1421 ret <8 x i1>%d
1422 }
1423 define <32 x i16> @test21(<32 x i16> %x , <32 x i1> %mask) nounwind readnone {
1424 ; KNL-LABEL: test21:
1425 ; KNL: ## %bb.0:
1426 ; KNL-NEXT: vpmovzxbw {{.*#+}} ymm2 = xmm1[0],zero,xmm1[1],zero,xmm1[2],zero,xmm1[3],zero,xmm1[4],zero,xmm1[5],zero,xmm1[6],zero,xmm1[7],zero,xmm1[8],zero,xmm1[9],zero,xmm1[10],zero,xmm1[11],zero,xmm1[12],zero,xmm1[13],zero,xmm1[14],zero,xmm1[15],zero
1427 ; KNL-NEXT: vextracti64x4 $1, %zmm0, %ymm3
1428 ; KNL-NEXT: vextracti128 $1, %ymm1, %xmm1
1429 ; KNL-NEXT: vpmovzxbw {{.*#+}} ymm1 = xmm1[0],zero,xmm1[1],zero,xmm1[2],zero,xmm1[3],zero,xmm1[4],zero,xmm1[5],zero,xmm1[6],zero,xmm1[7],zero,xmm1[8],zero,xmm1[9],zero,xmm1[10],zero,xmm1[11],zero,xmm1[12],zero,xmm1[13],zero,xmm1[14],zero,xmm1[15],zero
1430 ; KNL-NEXT: vpsllw $15, %ymm1, %ymm1
1431 ; KNL-NEXT: vpsraw $15, %ymm1, %ymm1
1432 ; KNL-NEXT: vpand %ymm3, %ymm1, %ymm1
1433 ; KNL-NEXT: vpsllw $15, %ymm2, %ymm2
1434 ; KNL-NEXT: vpsraw $15, %ymm2, %ymm2
1435 ; KNL-NEXT: vpand %ymm0, %ymm2, %ymm0
1436 ; KNL-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm0
1437 ; KNL-NEXT: retq
1438 ;
1439 ; SKX-LABEL: test21:
1440 ; SKX: ## %bb.0:
1441 ; SKX-NEXT: vpsllw $7, %ymm1, %ymm1
1442 ; SKX-NEXT: vpmovb2m %ymm1, %k1
1443 ; SKX-NEXT: vmovdqu16 %zmm0, %zmm0 {%k1} {z}
1444 ; SKX-NEXT: retq
1445 ;
1446 ; AVX512BW-LABEL: test21:
1447 ; AVX512BW: ## %bb.0:
1448 ; AVX512BW-NEXT: vpsllw $7, %ymm1, %ymm1
1449 ; AVX512BW-NEXT: vpmovb2m %zmm1, %k1
1450 ; AVX512BW-NEXT: vmovdqu16 %zmm0, %zmm0 {%k1} {z}
1451 ; AVX512BW-NEXT: retq
1452 ;
1453 ; AVX512DQ-LABEL: test21:
1454 ; AVX512DQ: ## %bb.0:
1455 ; AVX512DQ-NEXT: vpmovzxbw {{.*#+}} ymm2 = xmm1[0],zero,xmm1[1],zero,xmm1[2],zero,xmm1[3],zero,xmm1[4],zero,xmm1[5],zero,xmm1[6],zero,xmm1[7],zero,xmm1[8],zero,xmm1[9],zero,xmm1[10],zero,xmm1[11],zero,xmm1[12],zero,xmm1[13],zero,xmm1[14],zero,xmm1[15],zero
1456 ; AVX512DQ-NEXT: vextracti64x4 $1, %zmm0, %ymm3
1457 ; AVX512DQ-NEXT: vextracti128 $1, %ymm1, %xmm1
1458 ; AVX512DQ-NEXT: vpmovzxbw {{.*#+}} ymm1 = xmm1[0],zero,xmm1[1],zero,xmm1[2],zero,xmm1[3],zero,xmm1[4],zero,xmm1[5],zero,xmm1[6],zero,xmm1[7],zero,xmm1[8],zero,xmm1[9],zero,xmm1[10],zero,xmm1[11],zero,xmm1[12],zero,xmm1[13],zero,xmm1[14],zero,xmm1[15],zero
1459 ; AVX512DQ-NEXT: vpsllw $15, %ymm1, %ymm1
1460 ; AVX512DQ-NEXT: vpsraw $15, %ymm1, %ymm1
1461 ; AVX512DQ-NEXT: vpand %ymm3, %ymm1, %ymm1
1462 ; AVX512DQ-NEXT: vpsllw $15, %ymm2, %ymm2
1463 ; AVX512DQ-NEXT: vpsraw $15, %ymm2, %ymm2
1464 ; AVX512DQ-NEXT: vpand %ymm0, %ymm2, %ymm0
1465 ; AVX512DQ-NEXT: vinserti64x4 $1, %ymm1, %zmm0, %zmm0
1466 ; AVX512DQ-NEXT: retq
1467 ;
1468 ; X86-LABEL: test21:
1469 ; X86: ## %bb.0:
1470 ; X86-NEXT: vpsllw $7, %ymm1, %ymm1
1471 ; X86-NEXT: vpmovb2m %ymm1, %k1
1472 ; X86-NEXT: vmovdqu16 %zmm0, %zmm0 {%k1} {z}
1473 ; X86-NEXT: retl
1474 %ret = select <32 x i1> %mask, <32 x i16> %x, <32 x i16> zeroinitializer
1475 ret <32 x i16> %ret
1476 }
1478 define void @test22(<4 x i1> %a, <4 x i1>* %addr) {
1479 ; KNL-LABEL: test22:
1480 ; KNL: ## %bb.0:
1481 ; KNL-NEXT: vpslld $31, %xmm0, %xmm0
1482 ; KNL-NEXT: vptestmd %zmm0, %zmm0, %k0
1483 ; KNL-NEXT: kmovw %k0, %eax
1484 ; KNL-NEXT: movb %al, (%rdi)
1485 ; KNL-NEXT: vzeroupper
1486 ; KNL-NEXT: retq
1487 ;
1488 ; SKX-LABEL: test22:
1489 ; SKX: ## %bb.0:
1490 ; SKX-NEXT: vpslld $31, %xmm0, %xmm0
1491 ; SKX-NEXT: vpmovd2m %xmm0, %k0
1492 ; SKX-NEXT: kmovb %k0, (%rdi)
1493 ; SKX-NEXT: retq
1494 ;
1495 ; AVX512BW-LABEL: test22:
1496 ; AVX512BW: ## %bb.0:
1497 ; AVX512BW-NEXT: vpslld $31, %xmm0, %xmm0
1498 ; AVX512BW-NEXT: vptestmd %zmm0, %zmm0, %k0
1499 ; AVX512BW-NEXT: kmovd %k0, %eax
1500 ; AVX512BW-NEXT: movb %al, (%rdi)
1501 ; AVX512BW-NEXT: vzeroupper
1502 ; AVX512BW-NEXT: retq
1503 ;
1504 ; AVX512DQ-LABEL: test22:
1505 ; AVX512DQ: ## %bb.0:
1506 ; AVX512DQ-NEXT: vpslld $31, %xmm0, %xmm0
1507 ; AVX512DQ-NEXT: vpmovd2m %zmm0, %k0
1508 ; AVX512DQ-NEXT: kmovb %k0, (%rdi)
1509 ; AVX512DQ-NEXT: vzeroupper
1510 ; AVX512DQ-NEXT: retq
1511 ;
1512 ; X86-LABEL: test22:
1513 ; X86: ## %bb.0:
1514 ; X86-NEXT: vpslld $31, %xmm0, %xmm0
1515 ; X86-NEXT: vpmovd2m %xmm0, %k0
1516 ; X86-NEXT: movl {{[0-9]+}}(%esp), %eax
1517 ; X86-NEXT: kmovb %k0, (%eax)
1518 ; X86-NEXT: retl
1519 store <4 x i1> %a, <4 x i1>* %addr
1520 ret void
1521 }
1523 define void @test23(<2 x i1> %a, <2 x i1>* %addr) {
1524 ; KNL-LABEL: test23:
1525 ; KNL: ## %bb.0:
1526 ; KNL-NEXT: vpsllq $63, %xmm0, %xmm0
1527 ; KNL-NEXT: vptestmq %zmm0, %zmm0, %k0
1528 ; KNL-NEXT: kmovw %k0, %eax
1529 ; KNL-NEXT: movb %al, (%rdi)
1530 ; KNL-NEXT: vzeroupper
1531 ; KNL-NEXT: retq
1532 ;
1533 ; SKX-LABEL: test23:
1534 ; SKX: ## %bb.0:
1535 ; SKX-NEXT: vpsllq $63, %xmm0, %xmm0
1536 ; SKX-NEXT: vpmovq2m %xmm0, %k0
1537 ; SKX-NEXT: kmovb %k0, (%rdi)
1538 ; SKX-NEXT: retq
1539 ;
1540 ; AVX512BW-LABEL: test23:
1541 ; AVX512BW: ## %bb.0:
1542 ; AVX512BW-NEXT: vpsllq $63, %xmm0, %xmm0
1543 ; AVX512BW-NEXT: vptestmq %zmm0, %zmm0, %k0
1544 ; AVX512BW-NEXT: kmovd %k0, %eax
1545 ; AVX512BW-NEXT: movb %al, (%rdi)
1546 ; AVX512BW-NEXT: vzeroupper
1547 ; AVX512BW-NEXT: retq
1548 ;
1549 ; AVX512DQ-LABEL: test23:
1550 ; AVX512DQ: ## %bb.0:
1551 ; AVX512DQ-NEXT: vpsllq $63, %xmm0, %xmm0
1552 ; AVX512DQ-NEXT: vpmovq2m %zmm0, %k0
1553 ; AVX512DQ-NEXT: kmovb %k0, (%rdi)
1554 ; AVX512DQ-NEXT: vzeroupper
1555 ; AVX512DQ-NEXT: retq
1556 ;
1557 ; X86-LABEL: test23:
1558 ; X86: ## %bb.0:
1559 ; X86-NEXT: vpsllq $63, %xmm0, %xmm0
1560 ; X86-NEXT: vpmovq2m %xmm0, %k0
1561 ; X86-NEXT: movl {{[0-9]+}}(%esp), %eax
1562 ; X86-NEXT: kmovb %k0, (%eax)
1563 ; X86-NEXT: retl
1564 store <2 x i1> %a, <2 x i1>* %addr
1565 ret void
1566 }
1568 define void @store_v1i1(<1 x i1> %c , <1 x i1>* %ptr) {
1569 ; KNL-LABEL: store_v1i1:
1570 ; KNL: ## %bb.0:
1571 ; KNL-NEXT: kmovw %edi, %k0
1572 ; KNL-NEXT: kxnorw %k0, %k0, %k1
1573 ; KNL-NEXT: kxorw %k1, %k0, %k0
1574 ; KNL-NEXT: kmovw %k0, %eax
1575 ; KNL-NEXT: movb %al, (%rsi)
1576 ; KNL-NEXT: retq
1577 ;
1578 ; SKX-LABEL: store_v1i1:
1579 ; SKX: ## %bb.0:
1580 ; SKX-NEXT: kmovd %edi, %k0
1581 ; SKX-NEXT: kxnorw %k0, %k0, %k1
1582 ; SKX-NEXT: kxorw %k1, %k0, %k0
1583 ; SKX-NEXT: kmovb %k0, (%rsi)
1584 ; SKX-NEXT: retq
1585 ;
1586 ; AVX512BW-LABEL: store_v1i1:
1587 ; AVX512BW: ## %bb.0:
1588 ; AVX512BW-NEXT: kmovd %edi, %k0
1589 ; AVX512BW-NEXT: kxnorw %k0, %k0, %k1
1590 ; AVX512BW-NEXT: kxorw %k1, %k0, %k0
1591 ; AVX512BW-NEXT: kmovd %k0, %eax
1592 ; AVX512BW-NEXT: movb %al, (%rsi)
1593 ; AVX512BW-NEXT: retq
1594 ;
1595 ; AVX512DQ-LABEL: store_v1i1:
1596 ; AVX512DQ: ## %bb.0:
1597 ; AVX512DQ-NEXT: kmovw %edi, %k0
1598 ; AVX512DQ-NEXT: kxnorw %k0, %k0, %k1
1599 ; AVX512DQ-NEXT: kxorw %k1, %k0, %k0
1600 ; AVX512DQ-NEXT: kmovb %k0, (%rsi)
1601 ; AVX512DQ-NEXT: retq
1602 ;
1603 ; X86-LABEL: store_v1i1:
1604 ; X86: ## %bb.0:
1605 ; X86-NEXT: kmovd {{[0-9]+}}(%esp), %k0
1606 ; X86-NEXT: movl {{[0-9]+}}(%esp), %eax
1607 ; X86-NEXT: kxnorw %k0, %k0, %k1
1608 ; X86-NEXT: kxorw %k1, %k0, %k0
1609 ; X86-NEXT: kmovb %k0, (%eax)
1610 ; X86-NEXT: retl
1611 %x = xor <1 x i1> %c, <i1 1>
1612 store <1 x i1> %x, <1 x i1>* %ptr, align 4
1613 ret void
1614 }
1616 define void @store_v2i1(<2 x i1> %c , <2 x i1>* %ptr) {
1617 ; KNL-LABEL: store_v2i1:
1618 ; KNL: ## %bb.0:
1619 ; KNL-NEXT: vpsllq $63, %xmm0, %xmm0
1620 ; KNL-NEXT: vptestnmq %zmm0, %zmm0, %k0
1621 ; KNL-NEXT: kmovw %k0, %eax
1622 ; KNL-NEXT: movb %al, (%rdi)
1623 ; KNL-NEXT: vzeroupper
1624 ; KNL-NEXT: retq
1625 ;
1626 ; SKX-LABEL: store_v2i1:
1627 ; SKX: ## %bb.0:
1628 ; SKX-NEXT: vpsllq $63, %xmm0, %xmm0
1629 ; SKX-NEXT: vpmovq2m %xmm0, %k0
1630 ; SKX-NEXT: knotw %k0, %k0
1631 ; SKX-NEXT: kmovb %k0, (%rdi)
1632 ; SKX-NEXT: retq
1633 ;
1634 ; AVX512BW-LABEL: store_v2i1:
1635 ; AVX512BW: ## %bb.0:
1636 ; AVX512BW-NEXT: vpsllq $63, %xmm0, %xmm0
1637 ; AVX512BW-NEXT: vptestnmq %zmm0, %zmm0, %k0
1638 ; AVX512BW-NEXT: kmovd %k0, %eax
1639 ; AVX512BW-NEXT: movb %al, (%rdi)
1640 ; AVX512BW-NEXT: vzeroupper
1641 ; AVX512BW-NEXT: retq
1642 ;
1643 ; AVX512DQ-LABEL: store_v2i1:
1644 ; AVX512DQ: ## %bb.0:
1645 ; AVX512DQ-NEXT: vpsllq $63, %xmm0, %xmm0
1646 ; AVX512DQ-NEXT: vpmovq2m %zmm0, %k0
1647 ; AVX512DQ-NEXT: knotw %k0, %k0
1648 ; AVX512DQ-NEXT: kmovb %k0, (%rdi)
1649 ; AVX512DQ-NEXT: vzeroupper
1650 ; AVX512DQ-NEXT: retq
1651 ;
1652 ; X86-LABEL: store_v2i1:
1653 ; X86: ## %bb.0:
1654 ; X86-NEXT: vpsllq $63, %xmm0, %xmm0
1655 ; X86-NEXT: vpmovq2m %xmm0, %k0
1656 ; X86-NEXT: movl {{[0-9]+}}(%esp), %eax
1657 ; X86-NEXT: knotw %k0, %k0
1658 ; X86-NEXT: kmovb %k0, (%eax)
1659 ; X86-NEXT: retl
1660 %x = xor <2 x i1> %c, <i1 1, i1 1>
1661 store <2 x i1> %x, <2 x i1>* %ptr, align 4
1662 ret void
1663 }
1665 define void @store_v4i1(<4 x i1> %c , <4 x i1>* %ptr) {
1666 ; KNL-LABEL: store_v4i1:
1667 ; KNL: ## %bb.0:
1668 ; KNL-NEXT: vpslld $31, %xmm0, %xmm0
1669 ; KNL-NEXT: vptestnmd %zmm0, %zmm0, %k0
1670 ; KNL-NEXT: kmovw %k0, %eax
1671 ; KNL-NEXT: movb %al, (%rdi)
1672 ; KNL-NEXT: vzeroupper
1673 ; KNL-NEXT: retq
1674 ;
1675 ; SKX-LABEL: store_v4i1:
1676 ; SKX: ## %bb.0:
1677 ; SKX-NEXT: vpslld $31, %xmm0, %xmm0
1678 ; SKX-NEXT: vpmovd2m %xmm0, %k0
1679 ; SKX-NEXT: knotw %k0, %k0
1680 ; SKX-NEXT: kmovb %k0, (%rdi)
1681 ; SKX-NEXT: retq
1682 ;
1683 ; AVX512BW-LABEL: store_v4i1:
1684 ; AVX512BW: ## %bb.0:
1685 ; AVX512BW-NEXT: vpslld $31, %xmm0, %xmm0
1686 ; AVX512BW-NEXT: vptestnmd %zmm0, %zmm0, %k0
1687 ; AVX512BW-NEXT: kmovd %k0, %eax
1688 ; AVX512BW-NEXT: movb %al, (%rdi)
1689 ; AVX512BW-NEXT: vzeroupper
1690 ; AVX512BW-NEXT: retq
1691 ;
1692 ; AVX512DQ-LABEL: store_v4i1:
1693 ; AVX512DQ: ## %bb.0:
1694 ; AVX512DQ-NEXT: vpslld $31, %xmm0, %xmm0
1695 ; AVX512DQ-NEXT: vpmovd2m %zmm0, %k0
1696 ; AVX512DQ-NEXT: knotw %k0, %k0
1697 ; AVX512DQ-NEXT: kmovb %k0, (%rdi)
1698 ; AVX512DQ-NEXT: vzeroupper
1699 ; AVX512DQ-NEXT: retq
1700 ;
1701 ; X86-LABEL: store_v4i1:
1702 ; X86: ## %bb.0:
1703 ; X86-NEXT: vpslld $31, %xmm0, %xmm0
1704 ; X86-NEXT: vpmovd2m %xmm0, %k0
1705 ; X86-NEXT: movl {{[0-9]+}}(%esp), %eax
1706 ; X86-NEXT: knotw %k0, %k0
1707 ; X86-NEXT: kmovb %k0, (%eax)
1708 ; X86-NEXT: retl
1709 %x = xor <4 x i1> %c, <i1 1, i1 1, i1 1, i1 1>
1710 store <4 x i1> %x, <4 x i1>* %ptr, align 4
1711 ret void
1712 }
1714 define void @store_v8i1(<8 x i1> %c , <8 x i1>* %ptr) {
1715 ; KNL-LABEL: store_v8i1:
1716 ; KNL: ## %bb.0:
1717 ; KNL-NEXT: vpmovsxwq %xmm0, %zmm0
1718 ; KNL-NEXT: vpsllq $63, %zmm0, %zmm0
1719 ; KNL-NEXT: vptestnmq %zmm0, %zmm0, %k0
1720 ; KNL-NEXT: kmovw %k0, %eax
1721 ; KNL-NEXT: movb %al, (%rdi)
1722 ; KNL-NEXT: vzeroupper
1723 ; KNL-NEXT: retq
1724 ;
1725 ; SKX-LABEL: store_v8i1:
1726 ; SKX: ## %bb.0:
1727 ; SKX-NEXT: vpsllw $15, %xmm0, %xmm0
1728 ; SKX-NEXT: vpmovw2m %xmm0, %k0
1729 ; SKX-NEXT: knotb %k0, %k0
1730 ; SKX-NEXT: kmovb %k0, (%rdi)
1731 ; SKX-NEXT: retq
1732 ;
1733 ; AVX512BW-LABEL: store_v8i1:
1734 ; AVX512BW: ## %bb.0:
1735 ; AVX512BW-NEXT: vpsllw $15, %xmm0, %xmm0
1736 ; AVX512BW-NEXT: vpmovw2m %zmm0, %k0
1737 ; AVX512BW-NEXT: knotw %k0, %k0
1738 ; AVX512BW-NEXT: kmovd %k0, %eax
1739 ; AVX512BW-NEXT: movb %al, (%rdi)
1740 ; AVX512BW-NEXT: vzeroupper
1741 ; AVX512BW-NEXT: retq
1742 ;
1743 ; AVX512DQ-LABEL: store_v8i1:
1744 ; AVX512DQ: ## %bb.0:
1745 ; AVX512DQ-NEXT: vpmovsxwq %xmm0, %zmm0
1746 ; AVX512DQ-NEXT: vpsllq $63, %zmm0, %zmm0
1747 ; AVX512DQ-NEXT: vpmovq2m %zmm0, %k0
1748 ; AVX512DQ-NEXT: knotb %k0, %k0
1749 ; AVX512DQ-NEXT: kmovb %k0, (%rdi)
1750 ; AVX512DQ-NEXT: vzeroupper
1751 ; AVX512DQ-NEXT: retq
1752 ;
1753 ; X86-LABEL: store_v8i1:
1754 ; X86: ## %bb.0:
1755 ; X86-NEXT: vpsllw $15, %xmm0, %xmm0
1756 ; X86-NEXT: vpmovw2m %xmm0, %k0
1757 ; X86-NEXT: movl {{[0-9]+}}(%esp), %eax
1758 ; X86-NEXT: knotb %k0, %k0
1759 ; X86-NEXT: kmovb %k0, (%eax)
1760 ; X86-NEXT: retl
1761 %x = xor <8 x i1> %c, <i1 1, i1 1, i1 1, i1 1, i1 1, i1 1, i1 1, i1 1>
1762 store <8 x i1> %x, <8 x i1>* %ptr, align 4
1763 ret void
1764 }
1766 define void @store_v16i1(<16 x i1> %c , <16 x i1>* %ptr) {
1767 ; KNL-LABEL: store_v16i1:
1768 ; KNL: ## %bb.0:
1769 ; KNL-NEXT: vpmovsxbd %xmm0, %zmm0
1770 ; KNL-NEXT: vpslld $31, %zmm0, %zmm0
1771 ; KNL-NEXT: vptestnmd %zmm0, %zmm0, %k0
1772 ; KNL-NEXT: kmovw %k0, (%rdi)
1773 ; KNL-NEXT: vzeroupper
1774 ; KNL-NEXT: retq
1775 ;
1776 ; SKX-LABEL: store_v16i1:
1777 ; SKX: ## %bb.0:
1778 ; SKX-NEXT: vpsllw $7, %xmm0, %xmm0
1779 ; SKX-NEXT: vpmovb2m %xmm0, %k0
1780 ; SKX-NEXT: knotw %k0, %k0
1781 ; SKX-NEXT: kmovw %k0, (%rdi)
1782 ; SKX-NEXT: retq
1783 ;
1784 ; AVX512BW-LABEL: store_v16i1:
1785 ; AVX512BW: ## %bb.0:
1786 ; AVX512BW-NEXT: vpsllw $7, %xmm0, %xmm0
1787 ; AVX512BW-NEXT: vpmovb2m %zmm0, %k0
1788 ; AVX512BW-NEXT: knotw %k0, %k0
1789 ; AVX512BW-NEXT: kmovw %k0, (%rdi)
1790 ; AVX512BW-NEXT: vzeroupper
1791 ; AVX512BW-NEXT: retq
1792 ;
1793 ; AVX512DQ-LABEL: store_v16i1:
1794 ; AVX512DQ: ## %bb.0:
1795 ; AVX512DQ-NEXT: vpmovsxbd %xmm0, %zmm0
1796 ; AVX512DQ-NEXT: vpslld $31, %zmm0, %zmm0
1797 ; AVX512DQ-NEXT: vpmovd2m %zmm0, %k0
1798 ; AVX512DQ-NEXT: knotw %k0, %k0
1799 ; AVX512DQ-NEXT: kmovw %k0, (%rdi)
1800 ; AVX512DQ-NEXT: vzeroupper
1801 ; AVX512DQ-NEXT: retq
1802 ;
1803 ; X86-LABEL: store_v16i1:
1804 ; X86: ## %bb.0:
1805 ; X86-NEXT: vpsllw $7, %xmm0, %xmm0
1806 ; X86-NEXT: vpmovb2m %xmm0, %k0
1807 ; X86-NEXT: movl {{[0-9]+}}(%esp), %eax
1808 ; X86-NEXT: knotw %k0, %k0
1809 ; X86-NEXT: kmovw %k0, (%eax)
1810 ; X86-NEXT: retl
1811 %x = xor <16 x i1> %c, <i1 1, i1 1, i1 1, i1 1, i1 1, i1 1, i1 1, i1 1, i1 1, i1 1, i1 1, i1 1, i1 1, i1 1, i1 1, i1 1>
1812 store <16 x i1> %x, <16 x i1>* %ptr, align 4
1813 ret void
1814 }
1816 ;void f2(int);
1817 ;void f1(int c)
1818 ;{
1819 ; static int v = 0;
1820 ; if (v == 0)
1821 ; v = 1;
1822 ; else
1823 ; v = 0;
1824 ; f2(v);
1825 ;}
1827 @f1.v = internal unnamed_addr global i1 false, align 4
1829 define void @f1(i32 %c) {
1830 ; CHECK-LABEL: f1:
1831 ; CHECK: ## %bb.0: ## %entry
1832 ; CHECK-NEXT: movzbl {{.*}}(%rip), %edi
1833 ; CHECK-NEXT: xorl $1, %edi
1834 ; CHECK-NEXT: movb %dil, {{.*}}(%rip)
1835 ; CHECK-NEXT: jmp _f2 ## TAILCALL
1836 ;
1837 ; X86-LABEL: f1:
1838 ; X86: ## %bb.0: ## %entry
1839 ; X86-NEXT: subl $12, %esp
1840 ; X86-NEXT: .cfi_def_cfa_offset 16
1841 ; X86-NEXT: movzbl _f1.v, %eax
1842 ; X86-NEXT: xorl $1, %eax
1843 ; X86-NEXT: movb %al, _f1.v
1844 ; X86-NEXT: movl %eax, (%esp)
1845 ; X86-NEXT: calll _f2
1846 ; X86-NEXT: addl $12, %esp
1847 ; X86-NEXT: retl
1848 entry:
1849 %.b1 = load i1, i1* @f1.v, align 4
1850 %not..b1 = xor i1 %.b1, true
1851 store i1 %not..b1, i1* @f1.v, align 4
1852 %0 = zext i1 %not..b1 to i32
1853 tail call void @f2(i32 %0) #2
1854 ret void
1855 }
1857 declare void @f2(i32) #1
1859 define void @store_i16_i1(i16 %x, i1 *%y) {
1860 ; CHECK-LABEL: store_i16_i1:
1861 ; CHECK: ## %bb.0:
1862 ; CHECK-NEXT: andl $1, %edi
1863 ; CHECK-NEXT: movb %dil, (%rsi)
1864 ; CHECK-NEXT: retq
1865 ;
1866 ; X86-LABEL: store_i16_i1:
1867 ; X86: ## %bb.0:
1868 ; X86-NEXT: movl {{[0-9]+}}(%esp), %eax
1869 ; X86-NEXT: movzwl {{[0-9]+}}(%esp), %ecx
1870 ; X86-NEXT: andl $1, %ecx
1871 ; X86-NEXT: movb %cl, (%eax)
1872 ; X86-NEXT: retl
1873 %c = trunc i16 %x to i1
1874 store i1 %c, i1* %y
1875 ret void
1876 }
1878 define void @store_i8_i1(i8 %x, i1 *%y) {
1879 ; CHECK-LABEL: store_i8_i1:
1880 ; CHECK: ## %bb.0:
1881 ; CHECK-NEXT: andl $1, %edi
1882 ; CHECK-NEXT: movb %dil, (%rsi)
1883 ; CHECK-NEXT: retq
1884 ;
1885 ; X86-LABEL: store_i8_i1:
1886 ; X86: ## %bb.0:
1887 ; X86-NEXT: movl {{[0-9]+}}(%esp), %eax
1888 ; X86-NEXT: movb {{[0-9]+}}(%esp), %cl
1889 ; X86-NEXT: andb $1, %cl
1890 ; X86-NEXT: movb %cl, (%eax)
1891 ; X86-NEXT: retl
1892 %c = trunc i8 %x to i1
1893 store i1 %c, i1* %y
1894 ret void
1895 }
1897 define <32 x i16> @test_build_vec_v32i1(<32 x i16> %x) {
1898 ; KNL-LABEL: test_build_vec_v32i1:
1899 ; KNL: ## %bb.0:
1900 ; KNL-NEXT: vextractf64x4 $1, %zmm0, %ymm1
1901 ; KNL-NEXT: vandps {{.*}}(%rip), %ymm1, %ymm1
1902 ; KNL-NEXT: vandps {{.*}}(%rip), %ymm0, %ymm0
1903 ; KNL-NEXT: vinsertf64x4 $1, %ymm1, %zmm0, %zmm0
1904 ; KNL-NEXT: retq
1905 ;
1906 ; SKX-LABEL: test_build_vec_v32i1:
1907 ; SKX: ## %bb.0:
1908 ; SKX-NEXT: vandps {{.*}}(%rip), %zmm0, %zmm0
1909 ; SKX-NEXT: retq
1910 ;
1911 ; AVX512BW-LABEL: test_build_vec_v32i1:
1912 ; AVX512BW: ## %bb.0:
1913 ; AVX512BW-NEXT: vpandq {{.*}}(%rip), %zmm0, %zmm0
1914 ; AVX512BW-NEXT: retq
1915 ;
1916 ; AVX512DQ-LABEL: test_build_vec_v32i1:
1917 ; AVX512DQ: ## %bb.0:
1918 ; AVX512DQ-NEXT: vextractf64x4 $1, %zmm0, %ymm1
1919 ; AVX512DQ-NEXT: vandps {{.*}}(%rip), %ymm1, %ymm1
1920 ; AVX512DQ-NEXT: vandps {{.*}}(%rip), %ymm0, %ymm0
1921 ; AVX512DQ-NEXT: vinsertf64x4 $1, %ymm1, %zmm0, %zmm0
1922 ; AVX512DQ-NEXT: retq
1923 ;
1924 ; X86-LABEL: test_build_vec_v32i1:
1925 ; X86: ## %bb.0:
1926 ; X86-NEXT: vandps LCPI40_0, %zmm0, %zmm0
1927 ; X86-NEXT: retl
1928 %ret = select <32 x i1> <i1 true, i1 false, i1 true, i1 false, i1 true, i1 false, i1 false, i1 true, i1 false, i1 false, i1 true, i1 false, i1 true, i1 false, i1 true, i1 false, i1 true, i1 false, i1 true, i1 false, i1 false, i1 false, i1 true, i1 false, i1 true, i1 false, i1 false, i1 true, i1 true, i1 false, i1 true, i1 false>, <32 x i16> %x, <32 x i16> zeroinitializer
1929 ret <32 x i16> %ret
1930 }
1932 define <32 x i16> @test_build_vec_v32i1_optsize(<32 x i16> %x) optsize {
1933 ; KNL-LABEL: test_build_vec_v32i1_optsize:
1934 ; KNL: ## %bb.0:
1935 ; KNL-NEXT: vextractf64x4 $1, %zmm0, %ymm1
1936 ; KNL-NEXT: vandps {{.*}}(%rip), %ymm1, %ymm1
1937 ; KNL-NEXT: vandps {{.*}}(%rip), %ymm0, %ymm0
1938 ; KNL-NEXT: vinsertf64x4 $1, %ymm1, %zmm0, %zmm0
1939 ; KNL-NEXT: retq
1940 ;
1941 ; SKX-LABEL: test_build_vec_v32i1_optsize:
1942 ; SKX: ## %bb.0:
1943 ; SKX-NEXT: movl $1497715861, %eax ## imm = 0x59455495
1944 ; SKX-NEXT: kmovd %eax, %k1
1945 ; SKX-NEXT: vmovdqu16 %zmm0, %zmm0 {%k1} {z}
1946 ; SKX-NEXT: retq
1947 ;
1948 ; AVX512BW-LABEL: test_build_vec_v32i1_optsize:
1949 ; AVX512BW: ## %bb.0:
1950 ; AVX512BW-NEXT: movl $1497715861, %eax ## imm = 0x59455495
1951 ; AVX512BW-NEXT: kmovd %eax, %k1
1952 ; AVX512BW-NEXT: vmovdqu16 %zmm0, %zmm0 {%k1} {z}
1953 ; AVX512BW-NEXT: retq
1954 ;
1955 ; AVX512DQ-LABEL: test_build_vec_v32i1_optsize:
1956 ; AVX512DQ: ## %bb.0:
1957 ; AVX512DQ-NEXT: vextractf64x4 $1, %zmm0, %ymm1
1958 ; AVX512DQ-NEXT: vandps {{.*}}(%rip), %ymm1, %ymm1
1959 ; AVX512DQ-NEXT: vandps {{.*}}(%rip), %ymm0, %ymm0
1960 ; AVX512DQ-NEXT: vinsertf64x4 $1, %ymm1, %zmm0, %zmm0
1961 ; AVX512DQ-NEXT: retq
1962 ;
1963 ; X86-LABEL: test_build_vec_v32i1_optsize:
1964 ; X86: ## %bb.0:
1965 ; X86-NEXT: movl $1497715861, %eax ## imm = 0x59455495
1966 ; X86-NEXT: kmovd %eax, %k1
1967 ; X86-NEXT: vmovdqu16 %zmm0, %zmm0 {%k1} {z}
1968 ; X86-NEXT: retl
1969 %ret = select <32 x i1> <i1 true, i1 false, i1 true, i1 false, i1 true, i1 false, i1 false, i1 true, i1 false, i1 false, i1 true, i1 false, i1 true, i1 false, i1 true, i1 false, i1 true, i1 false, i1 true, i1 false, i1 false, i1 false, i1 true, i1 false, i1 true, i1 false, i1 false, i1 true, i1 true, i1 false, i1 true, i1 false>, <32 x i16> %x, <32 x i16> zeroinitializer
1970 ret <32 x i16> %ret
1971 }
1973 define <64 x i8> @test_build_vec_v64i1(<64 x i8> %x) {
1974 ; KNL-LABEL: test_build_vec_v64i1:
1975 ; KNL: ## %bb.0:
1976 ; KNL-NEXT: vextractf64x4 $1, %zmm0, %ymm1
1977 ; KNL-NEXT: vandps {{.*}}(%rip), %ymm1, %ymm1
1978 ; KNL-NEXT: vandps {{.*}}(%rip), %ymm0, %ymm0
1979 ; KNL-NEXT: vinsertf64x4 $1, %ymm1, %zmm0, %zmm0
1980 ; KNL-NEXT: retq
1981 ;
1982 ; SKX-LABEL: test_build_vec_v64i1:
1983 ; SKX: ## %bb.0:
1984 ; SKX-NEXT: vpshufb {{.*#+}} zmm0 = zero,zero,zmm0[2],zero,zero,zero,zmm0[6],zero,zmm0[8],zero,zmm0[10],zero,zmm0[12],zero,zero,zmm0[15],zero,zero,zmm0[18],zero,zmm0[20],zero,zmm0[22],zero,zmm0[24],zero,zero,zmm0[27],zero,zero,zmm0[30],zero,zmm0[32],zero,zmm0[34],zero,zero,zero,zmm0[38],zero,zmm0[40],zero,zero,zmm0[43,44],zero,zmm0[46],zero,zmm0[48],zero,zmm0[50],zero,zero,zero,zmm0[54],zero,zmm0[56],zero,zero,zmm0[59,60],zero,zmm0[62],zero
1985 ; SKX-NEXT: retq
1986 ;
1987 ; AVX512BW-LABEL: test_build_vec_v64i1:
1988 ; AVX512BW: ## %bb.0:
1989 ; AVX512BW-NEXT: vpshufb {{.*#+}} zmm0 = zero,zero,zmm0[2],zero,zero,zero,zmm0[6],zero,zmm0[8],zero,zmm0[10],zero,zmm0[12],zero,zero,zmm0[15],zero,zero,zmm0[18],zero,zmm0[20],zero,zmm0[22],zero,zmm0[24],zero,zero,zmm0[27],zero,zero,zmm0[30],zero,zmm0[32],zero,zmm0[34],zero,zero,zero,zmm0[38],zero,zmm0[40],zero,zero,zmm0[43,44],zero,zmm0[46],zero,zmm0[48],zero,zmm0[50],zero,zero,zero,zmm0[54],zero,zmm0[56],zero,zero,zmm0[59,60],zero,zmm0[62],zero
1990 ; AVX512BW-NEXT: retq
1991 ;
1992 ; AVX512DQ-LABEL: test_build_vec_v64i1:
1993 ; AVX512DQ: ## %bb.0:
1994 ; AVX512DQ-NEXT: vextractf64x4 $1, %zmm0, %ymm1
1995 ; AVX512DQ-NEXT: vandps {{.*}}(%rip), %ymm1, %ymm1
1996 ; AVX512DQ-NEXT: vandps {{.*}}(%rip), %ymm0, %ymm0
1997 ; AVX512DQ-NEXT: vinsertf64x4 $1, %ymm1, %zmm0, %zmm0
1998 ; AVX512DQ-NEXT: retq
1999 ;
2000 ; X86-LABEL: test_build_vec_v64i1:
2001 ; X86: ## %bb.0:
2002 ; X86-NEXT: vpshufb {{.*#+}} zmm0 = zero,zero,zmm0[2],zero,zero,zero,zmm0[6],zero,zmm0[8],zero,zmm0[10],zero,zmm0[12],zero,zero,zmm0[15],zero,zero,zmm0[18],zero,zmm0[20],zero,zmm0[22],zero,zmm0[24],zero,zero,zmm0[27],zero,zero,zmm0[30],zero,zmm0[32],zero,zmm0[34],zero,zero,zero,zmm0[38],zero,zmm0[40],zero,zero,zmm0[43,44],zero,zmm0[46],zero,zmm0[48],zero,zmm0[50],zero,zero,zero,zmm0[54],zero,zmm0[56],zero,zero,zmm0[59,60],zero,zmm0[62],zero
2003 ; X86-NEXT: retl
2004 %ret = select <64 x i1> <i1 false, i1 false, i1 true, i1 false, i1 false, i1 false, i1 true, i1 false, i1 true, i1 false, i1 true, i1 false, i1 true, i1 false, i1 false, i1 true, i1 false, i1 false, i1 true, i1 false, i1 true, i1 false, i1 true, i1 false, i1 true, i1 false, i1 false, i1 true, i1 false, i1 false, i1 true, i1 false, i1 true, i1 false, i1 true, i1 false, i1 false, i1 false, i1 true, i1 false, i1 true, i1 false, i1 false, i1 true, i1 true, i1 false, i1 true, i1 false, i1 true, i1 false, i1 true, i1 false, i1 false, i1 false, i1 true, i1 false, i1 true, i1 false, i1 false, i1 true, i1 true, i1 false, i1 true, i1 false>, <64 x i8> %x, <64 x i8> zeroinitializer
2005 ret <64 x i8> %ret
2006 }
2008 define void @ktest_1(<8 x double> %in, double * %base) {
2009 ; KNL-LABEL: ktest_1:
2010 ; KNL: ## %bb.0:
2011 ; KNL-NEXT: vcmpgtpd (%rdi), %zmm0, %k1
2012 ; KNL-NEXT: vmovupd 8(%rdi), %zmm1 {%k1} {z}
2013 ; KNL-NEXT: vcmpltpd %zmm1, %zmm0, %k0 {%k1}
2014 ; KNL-NEXT: kmovw %k0, %eax
2015 ; KNL-NEXT: testb %al, %al
2016 ; KNL-NEXT: je LBB43_2
2017 ; KNL-NEXT: ## %bb.1: ## %L1
2018 ; KNL-NEXT: vmovapd %zmm0, (%rdi)
2019 ; KNL-NEXT: vzeroupper
2020 ; KNL-NEXT: retq
2021 ; KNL-NEXT: LBB43_2: ## %L2
2022 ; KNL-NEXT: vmovapd %zmm0, 8(%rdi)
2023 ; KNL-NEXT: vzeroupper
2024 ; KNL-NEXT: retq
2025 ;
2026 ; SKX-LABEL: ktest_1:
2027 ; SKX: ## %bb.0:
2028 ; SKX-NEXT: vcmpgtpd (%rdi), %zmm0, %k1
2029 ; SKX-NEXT: vmovupd 8(%rdi), %zmm1 {%k1} {z}
2030 ; SKX-NEXT: vcmpltpd %zmm1, %zmm0, %k0
2031 ; SKX-NEXT: ktestb %k0, %k1
2032 ; SKX-NEXT: je LBB43_2
2033 ; SKX-NEXT: ## %bb.1: ## %L1
2034 ; SKX-NEXT: vmovapd %zmm0, (%rdi)
2035 ; SKX-NEXT: vzeroupper
2036 ; SKX-NEXT: retq
2037 ; SKX-NEXT: LBB43_2: ## %L2
2038 ; SKX-NEXT: vmovapd %zmm0, 8(%rdi)
2039 ; SKX-NEXT: vzeroupper
2040 ; SKX-NEXT: retq
2041 ;
2042 ; AVX512BW-LABEL: ktest_1:
2043 ; AVX512BW: ## %bb.0:
2044 ; AVX512BW-NEXT: vcmpgtpd (%rdi), %zmm0, %k1
2045 ; AVX512BW-NEXT: vmovupd 8(%rdi), %zmm1 {%k1} {z}
2046 ; AVX512BW-NEXT: vcmpltpd %zmm1, %zmm0, %k0 {%k1}
2047 ; AVX512BW-NEXT: kmovd %k0, %eax
2048 ; AVX512BW-NEXT: testb %al, %al
2049 ; AVX512BW-NEXT: je LBB43_2
2050 ; AVX512BW-NEXT: ## %bb.1: ## %L1
2051 ; AVX512BW-NEXT: vmovapd %zmm0, (%rdi)
2052 ; AVX512BW-NEXT: vzeroupper
2053 ; AVX512BW-NEXT: retq
2054 ; AVX512BW-NEXT: LBB43_2: ## %L2
2055 ; AVX512BW-NEXT: vmovapd %zmm0, 8(%rdi)
2056 ; AVX512BW-NEXT: vzeroupper
2057 ; AVX512BW-NEXT: retq
2058 ;
2059 ; AVX512DQ-LABEL: ktest_1:
2060 ; AVX512DQ: ## %bb.0:
2061 ; AVX512DQ-NEXT: vcmpgtpd (%rdi), %zmm0, %k1
2062 ; AVX512DQ-NEXT: vmovupd 8(%rdi), %zmm1 {%k1} {z}
2063 ; AVX512DQ-NEXT: vcmpltpd %zmm1, %zmm0, %k0
2064 ; AVX512DQ-NEXT: ktestb %k0, %k1
2065 ; AVX512DQ-NEXT: je LBB43_2
2066 ; AVX512DQ-NEXT: ## %bb.1: ## %L1
2067 ; AVX512DQ-NEXT: vmovapd %zmm0, (%rdi)
2068 ; AVX512DQ-NEXT: vzeroupper
2069 ; AVX512DQ-NEXT: retq
2070 ; AVX512DQ-NEXT: LBB43_2: ## %L2
2071 ; AVX512DQ-NEXT: vmovapd %zmm0, 8(%rdi)
2072 ; AVX512DQ-NEXT: vzeroupper
2073 ; AVX512DQ-NEXT: retq
2074 ;
2075 ; X86-LABEL: ktest_1:
2076 ; X86: ## %bb.0:
2077 ; X86-NEXT: movl {{[0-9]+}}(%esp), %eax
2078 ; X86-NEXT: vcmpgtpd (%eax), %zmm0, %k1
2079 ; X86-NEXT: vmovupd 8(%eax), %zmm1 {%k1} {z}
2080 ; X86-NEXT: vcmpltpd %zmm1, %zmm0, %k0
2081 ; X86-NEXT: ktestb %k0, %k1
2082 ; X86-NEXT: je LBB43_2
2083 ; X86-NEXT: ## %bb.1: ## %L1
2084 ; X86-NEXT: vmovapd %zmm0, (%eax)
2085 ; X86-NEXT: vzeroupper
2086 ; X86-NEXT: retl
2087 ; X86-NEXT: LBB43_2: ## %L2
2088 ; X86-NEXT: vmovapd %zmm0, 8(%eax)
2089 ; X86-NEXT: vzeroupper
2090 ; X86-NEXT: retl
2091 %addr1 = getelementptr double, double * %base, i64 0
2092 %addr2 = getelementptr double, double * %base, i64 1
2094 %vaddr1 = bitcast double* %addr1 to <8 x double>*
2095 %vaddr2 = bitcast double* %addr2 to <8 x double>*
2097 %val1 = load <8 x double>, <8 x double> *%vaddr1, align 1
2098 %val2 = load <8 x double>, <8 x double> *%vaddr2, align 1
2100 %sel1 = fcmp ogt <8 x double>%in, %val1
2101 %val3 = select <8 x i1> %sel1, <8 x double> %val2, <8 x double> zeroinitializer
2102 %sel2 = fcmp olt <8 x double> %in, %val3
2103 %sel3 = and <8 x i1> %sel1, %sel2
2105 %int_sel3 = bitcast <8 x i1> %sel3 to i8
2106 %res = icmp eq i8 %int_sel3, zeroinitializer
2107 br i1 %res, label %L2, label %L1
2108 L1:
2109 store <8 x double> %in, <8 x double>* %vaddr1
2110 br label %End
2111 L2:
2112 store <8 x double> %in, <8 x double>* %vaddr2
2113 br label %End
2114 End:
2115 ret void
2116 }
2118 define void @ktest_2(<32 x float> %in, float * %base) {
2119 ;
2120 ; KNL-LABEL: ktest_2:
2121 ; KNL: ## %bb.0:
2122 ; KNL-NEXT: vcmpgtps 64(%rdi), %zmm1, %k1
2123 ; KNL-NEXT: vcmpgtps (%rdi), %zmm0, %k2
2124 ; KNL-NEXT: vmovups 4(%rdi), %zmm2 {%k2} {z}
2125 ; KNL-NEXT: vmovups 68(%rdi), %zmm3 {%k1} {z}
2126 ; KNL-NEXT: vcmpltps %zmm3, %zmm1, %k0
2127 ; KNL-NEXT: vcmpltps %zmm2, %zmm0, %k3
2128 ; KNL-NEXT: korw %k3, %k2, %k2
2129 ; KNL-NEXT: kmovw %k2, %eax
2130 ; KNL-NEXT: korw %k0, %k1, %k0
2131 ; KNL-NEXT: kmovw %k0, %ecx
2132 ; KNL-NEXT: shll $16, %ecx
2133 ; KNL-NEXT: orl %eax, %ecx
2134 ; KNL-NEXT: je LBB44_2
2135 ; KNL-NEXT: ## %bb.1: ## %L1
2136 ; KNL-NEXT: vmovaps %zmm0, (%rdi)
2137 ; KNL-NEXT: vmovaps %zmm1, 64(%rdi)
2138 ; KNL-NEXT: vzeroupper
2139 ; KNL-NEXT: retq
2140 ; KNL-NEXT: LBB44_2: ## %L2
2141 ; KNL-NEXT: vmovaps %zmm0, 4(%rdi)
2142 ; KNL-NEXT: vmovaps %zmm1, 68(%rdi)
2143 ; KNL-NEXT: vzeroupper
2144 ; KNL-NEXT: retq
2145 ;
2146 ; SKX-LABEL: ktest_2:
2147 ; SKX: ## %bb.0:
2148 ; SKX-NEXT: vcmpgtps (%rdi), %zmm0, %k1
2149 ; SKX-NEXT: vcmpgtps 64(%rdi), %zmm1, %k2
2150 ; SKX-NEXT: kunpckwd %k1, %k2, %k0
2151 ; SKX-NEXT: vmovups 68(%rdi), %zmm2 {%k2} {z}
2152 ; SKX-NEXT: vmovups 4(%rdi), %zmm3 {%k1} {z}
2153 ; SKX-NEXT: vcmpltps %zmm3, %zmm0, %k1
2154 ; SKX-NEXT: vcmpltps %zmm2, %zmm1, %k2
2155 ; SKX-NEXT: kunpckwd %k1, %k2, %k1
2156 ; SKX-NEXT: kortestd %k1, %k0
2157 ; SKX-NEXT: je LBB44_2
2158 ; SKX-NEXT: ## %bb.1: ## %L1
2159 ; SKX-NEXT: vmovaps %zmm0, (%rdi)
2160 ; SKX-NEXT: vmovaps %zmm1, 64(%rdi)
2161 ; SKX-NEXT: vzeroupper
2162 ; SKX-NEXT: retq
2163 ; SKX-NEXT: LBB44_2: ## %L2
2164 ; SKX-NEXT: vmovaps %zmm0, 4(%rdi)
2165 ; SKX-NEXT: vmovaps %zmm1, 68(%rdi)
2166 ; SKX-NEXT: vzeroupper
2167 ; SKX-NEXT: retq
2168 ;
2169 ; AVX512BW-LABEL: ktest_2:
2170 ; AVX512BW: ## %bb.0:
2171 ; AVX512BW-NEXT: vcmpgtps (%rdi), %zmm0, %k1
2172 ; AVX512BW-NEXT: vcmpgtps 64(%rdi), %zmm1, %k2
2173 ; AVX512BW-NEXT: kunpckwd %k1, %k2, %k0
2174 ; AVX512BW-NEXT: vmovups 68(%rdi), %zmm2 {%k2} {z}
2175 ; AVX512BW-NEXT: vmovups 4(%rdi), %zmm3 {%k1} {z}
2176 ; AVX512BW-NEXT: vcmpltps %zmm3, %zmm0, %k1
2177 ; AVX512BW-NEXT: vcmpltps %zmm2, %zmm1, %k2
2178 ; AVX512BW-NEXT: kunpckwd %k1, %k2, %k1
2179 ; AVX512BW-NEXT: kortestd %k1, %k0
2180 ; AVX512BW-NEXT: je LBB44_2
2181 ; AVX512BW-NEXT: ## %bb.1: ## %L1
2182 ; AVX512BW-NEXT: vmovaps %zmm0, (%rdi)
2183 ; AVX512BW-NEXT: vmovaps %zmm1, 64(%rdi)
2184 ; AVX512BW-NEXT: vzeroupper
2185 ; AVX512BW-NEXT: retq
2186 ; AVX512BW-NEXT: LBB44_2: ## %L2
2187 ; AVX512BW-NEXT: vmovaps %zmm0, 4(%rdi)
2188 ; AVX512BW-NEXT: vmovaps %zmm1, 68(%rdi)
2189 ; AVX512BW-NEXT: vzeroupper
2190 ; AVX512BW-NEXT: retq
2191 ;
2192 ; AVX512DQ-LABEL: ktest_2:
2193 ; AVX512DQ: ## %bb.0:
2194 ; AVX512DQ-NEXT: vcmpgtps 64(%rdi), %zmm1, %k1
2195 ; AVX512DQ-NEXT: vcmpgtps (%rdi), %zmm0, %k2
2196 ; AVX512DQ-NEXT: vmovups 4(%rdi), %zmm2 {%k2} {z}
2197 ; AVX512DQ-NEXT: vmovups 68(%rdi), %zmm3 {%k1} {z}
2198 ; AVX512DQ-NEXT: vcmpltps %zmm3, %zmm1, %k0
2199 ; AVX512DQ-NEXT: vcmpltps %zmm2, %zmm0, %k3
2200 ; AVX512DQ-NEXT: korw %k3, %k2, %k2
2201 ; AVX512DQ-NEXT: kmovw %k2, %eax
2202 ; AVX512DQ-NEXT: korw %k0, %k1, %k0
2203 ; AVX512DQ-NEXT: kmovw %k0, %ecx
2204 ; AVX512DQ-NEXT: shll $16, %ecx
2205 ; AVX512DQ-NEXT: orl %eax, %ecx
2206 ; AVX512DQ-NEXT: je LBB44_2
2207 ; AVX512DQ-NEXT: ## %bb.1: ## %L1
2208 ; AVX512DQ-NEXT: vmovaps %zmm0, (%rdi)
2209 ; AVX512DQ-NEXT: vmovaps %zmm1, 64(%rdi)
2210 ; AVX512DQ-NEXT: vzeroupper
2211 ; AVX512DQ-NEXT: retq
2212 ; AVX512DQ-NEXT: LBB44_2: ## %L2
2213 ; AVX512DQ-NEXT: vmovaps %zmm0, 4(%rdi)
2214 ; AVX512DQ-NEXT: vmovaps %zmm1, 68(%rdi)
2215 ; AVX512DQ-NEXT: vzeroupper
2216 ; AVX512DQ-NEXT: retq
2217 ;
2218 ; X86-LABEL: ktest_2:
2219 ; X86: ## %bb.0:
2220 ; X86-NEXT: movl {{[0-9]+}}(%esp), %eax
2221 ; X86-NEXT: vcmpgtps (%eax), %zmm0, %k1
2222 ; X86-NEXT: vcmpgtps 64(%eax), %zmm1, %k2
2223 ; X86-NEXT: kunpckwd %k1, %k2, %k0
2224 ; X86-NEXT: vmovups 68(%eax), %zmm2 {%k2} {z}
2225 ; X86-NEXT: vmovups 4(%eax), %zmm3 {%k1} {z}
2226 ; X86-NEXT: vcmpltps %zmm3, %zmm0, %k1
2227 ; X86-NEXT: vcmpltps %zmm2, %zmm1, %k2
2228 ; X86-NEXT: kunpckwd %k1, %k2, %k1
2229 ; X86-NEXT: kortestd %k1, %k0
2230 ; X86-NEXT: je LBB44_2
2231 ; X86-NEXT: ## %bb.1: ## %L1
2232 ; X86-NEXT: vmovaps %zmm0, (%eax)
2233 ; X86-NEXT: vmovaps %zmm1, 64(%eax)
2234 ; X86-NEXT: vzeroupper
2235 ; X86-NEXT: retl
2236 ; X86-NEXT: LBB44_2: ## %L2
2237 ; X86-NEXT: vmovaps %zmm0, 4(%eax)
2238 ; X86-NEXT: vmovaps %zmm1, 68(%eax)
2239 ; X86-NEXT: vzeroupper
2240 ; X86-NEXT: retl
2241 %addr1 = getelementptr float, float * %base, i64 0
2242 %addr2 = getelementptr float, float * %base, i64 1
2244 %vaddr1 = bitcast float* %addr1 to <32 x float>*
2245 %vaddr2 = bitcast float* %addr2 to <32 x float>*
2247 %val1 = load <32 x float>, <32 x float> *%vaddr1, align 1
2248 %val2 = load <32 x float>, <32 x float> *%vaddr2, align 1
2250 %sel1 = fcmp ogt <32 x float>%in, %val1
2251 %val3 = select <32 x i1> %sel1, <32 x float> %val2, <32 x float> zeroinitializer
2252 %sel2 = fcmp olt <32 x float> %in, %val3
2253 %sel3 = or <32 x i1> %sel1, %sel2
2255 %int_sel3 = bitcast <32 x i1> %sel3 to i32
2256 %res = icmp eq i32 %int_sel3, zeroinitializer
2257 br i1 %res, label %L2, label %L1
2258 L1:
2259 store <32 x float> %in, <32 x float>* %vaddr1
2260 br label %End
2261 L2:
2262 store <32 x float> %in, <32 x float>* %vaddr2
2263 br label %End
2264 End:
2265 ret void
2266 }
2268 define <8 x i64> @load_8i1(<8 x i1>* %a) {
2269 ; KNL-LABEL: load_8i1:
2270 ; KNL: ## %bb.0:
2271 ; KNL-NEXT: kmovw (%rdi), %k1
2272 ; KNL-NEXT: vpternlogq $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
2273 ; KNL-NEXT: retq
2274 ;
2275 ; SKX-LABEL: load_8i1:
2276 ; SKX: ## %bb.0:
2277 ; SKX-NEXT: kmovb (%rdi), %k0
2278 ; SKX-NEXT: vpmovm2q %k0, %zmm0
2279 ; SKX-NEXT: retq
2280 ;
2281 ; AVX512BW-LABEL: load_8i1:
2282 ; AVX512BW: ## %bb.0:
2283 ; AVX512BW-NEXT: kmovw (%rdi), %k1
2284 ; AVX512BW-NEXT: vpternlogq $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
2285 ; AVX512BW-NEXT: retq
2286 ;
2287 ; AVX512DQ-LABEL: load_8i1:
2288 ; AVX512DQ: ## %bb.0:
2289 ; AVX512DQ-NEXT: kmovb (%rdi), %k0
2290 ; AVX512DQ-NEXT: vpmovm2q %k0, %zmm0
2291 ; AVX512DQ-NEXT: retq
2292 ;
2293 ; X86-LABEL: load_8i1:
2294 ; X86: ## %bb.0:
2295 ; X86-NEXT: movl {{[0-9]+}}(%esp), %eax
2296 ; X86-NEXT: kmovb (%eax), %k0
2297 ; X86-NEXT: vpmovm2q %k0, %zmm0
2298 ; X86-NEXT: retl
2299 %b = load <8 x i1>, <8 x i1>* %a
2300 %c = sext <8 x i1> %b to <8 x i64>
2301 ret <8 x i64> %c
2302 }
2304 define <16 x i32> @load_16i1(<16 x i1>* %a) {
2305 ; KNL-LABEL: load_16i1:
2306 ; KNL: ## %bb.0:
2307 ; KNL-NEXT: kmovw (%rdi), %k1
2308 ; KNL-NEXT: vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
2309 ; KNL-NEXT: retq
2310 ;
2311 ; SKX-LABEL: load_16i1:
2312 ; SKX: ## %bb.0:
2313 ; SKX-NEXT: kmovw (%rdi), %k0
2314 ; SKX-NEXT: vpmovm2d %k0, %zmm0
2315 ; SKX-NEXT: retq
2316 ;
2317 ; AVX512BW-LABEL: load_16i1:
2318 ; AVX512BW: ## %bb.0:
2319 ; AVX512BW-NEXT: kmovw (%rdi), %k1
2320 ; AVX512BW-NEXT: vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
2321 ; AVX512BW-NEXT: retq
2322 ;
2323 ; AVX512DQ-LABEL: load_16i1:
2324 ; AVX512DQ: ## %bb.0:
2325 ; AVX512DQ-NEXT: kmovw (%rdi), %k0
2326 ; AVX512DQ-NEXT: vpmovm2d %k0, %zmm0
2327 ; AVX512DQ-NEXT: retq
2328 ;
2329 ; X86-LABEL: load_16i1:
2330 ; X86: ## %bb.0:
2331 ; X86-NEXT: movl {{[0-9]+}}(%esp), %eax
2332 ; X86-NEXT: kmovw (%eax), %k0
2333 ; X86-NEXT: vpmovm2d %k0, %zmm0
2334 ; X86-NEXT: retl
2335 %b = load <16 x i1>, <16 x i1>* %a
2336 %c = sext <16 x i1> %b to <16 x i32>
2337 ret <16 x i32> %c
2338 }
2340 define <2 x i16> @load_2i1(<2 x i1>* %a) {
2341 ; KNL-LABEL: load_2i1:
2342 ; KNL: ## %bb.0:
2343 ; KNL-NEXT: kmovw (%rdi), %k1
2344 ; KNL-NEXT: vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
2345 ; KNL-NEXT: vpmovdw %zmm0, %ymm0
2346 ; KNL-NEXT: ## kill: def $xmm0 killed $xmm0 killed $ymm0
2347 ; KNL-NEXT: vzeroupper
2348 ; KNL-NEXT: retq
2349 ;
2350 ; SKX-LABEL: load_2i1:
2351 ; SKX: ## %bb.0:
2352 ; SKX-NEXT: kmovb (%rdi), %k0
2353 ; SKX-NEXT: vpmovm2w %k0, %xmm0
2354 ; SKX-NEXT: retq
2355 ;
2356 ; AVX512BW-LABEL: load_2i1:
2357 ; AVX512BW: ## %bb.0:
2358 ; AVX512BW-NEXT: kmovw (%rdi), %k0
2359 ; AVX512BW-NEXT: vpmovm2w %k0, %zmm0
2360 ; AVX512BW-NEXT: ## kill: def $xmm0 killed $xmm0 killed $zmm0
2361 ; AVX512BW-NEXT: vzeroupper
2362 ; AVX512BW-NEXT: retq
2363 ;
2364 ; AVX512DQ-LABEL: load_2i1:
2365 ; AVX512DQ: ## %bb.0:
2366 ; AVX512DQ-NEXT: kmovb (%rdi), %k0
2367 ; AVX512DQ-NEXT: vpmovm2d %k0, %zmm0
2368 ; AVX512DQ-NEXT: vpmovdw %zmm0, %ymm0
2369 ; AVX512DQ-NEXT: ## kill: def $xmm0 killed $xmm0 killed $ymm0
2370 ; AVX512DQ-NEXT: vzeroupper
2371 ; AVX512DQ-NEXT: retq
2372 ;
2373 ; X86-LABEL: load_2i1:
2374 ; X86: ## %bb.0:
2375 ; X86-NEXT: movl {{[0-9]+}}(%esp), %eax
2376 ; X86-NEXT: kmovb (%eax), %k0
2377 ; X86-NEXT: vpmovm2w %k0, %xmm0
2378 ; X86-NEXT: retl
2379 %b = load <2 x i1>, <2 x i1>* %a
2380 %c = sext <2 x i1> %b to <2 x i16>
2381 ret <2 x i16> %c
2382 }
2384 define <4 x i16> @load_4i1(<4 x i1>* %a) {
2385 ; KNL-LABEL: load_4i1:
2386 ; KNL: ## %bb.0:
2387 ; KNL-NEXT: kmovw (%rdi), %k1
2388 ; KNL-NEXT: vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
2389 ; KNL-NEXT: vpmovdw %zmm0, %ymm0
2390 ; KNL-NEXT: ## kill: def $xmm0 killed $xmm0 killed $ymm0
2391 ; KNL-NEXT: vzeroupper
2392 ; KNL-NEXT: retq
2393 ;
2394 ; SKX-LABEL: load_4i1:
2395 ; SKX: ## %bb.0:
2396 ; SKX-NEXT: kmovb (%rdi), %k0
2397 ; SKX-NEXT: vpmovm2w %k0, %xmm0
2398 ; SKX-NEXT: retq
2399 ;
2400 ; AVX512BW-LABEL: load_4i1:
2401 ; AVX512BW: ## %bb.0:
2402 ; AVX512BW-NEXT: kmovw (%rdi), %k0
2403 ; AVX512BW-NEXT: vpmovm2w %k0, %zmm0
2404 ; AVX512BW-NEXT: ## kill: def $xmm0 killed $xmm0 killed $zmm0
2405 ; AVX512BW-NEXT: vzeroupper
2406 ; AVX512BW-NEXT: retq
2407 ;
2408 ; AVX512DQ-LABEL: load_4i1:
2409 ; AVX512DQ: ## %bb.0:
2410 ; AVX512DQ-NEXT: kmovb (%rdi), %k0
2411 ; AVX512DQ-NEXT: vpmovm2d %k0, %zmm0
2412 ; AVX512DQ-NEXT: vpmovdw %zmm0, %ymm0
2413 ; AVX512DQ-NEXT: ## kill: def $xmm0 killed $xmm0 killed $ymm0
2414 ; AVX512DQ-NEXT: vzeroupper
2415 ; AVX512DQ-NEXT: retq
2416 ;
2417 ; X86-LABEL: load_4i1:
2418 ; X86: ## %bb.0:
2419 ; X86-NEXT: movl {{[0-9]+}}(%esp), %eax
2420 ; X86-NEXT: kmovb (%eax), %k0
2421 ; X86-NEXT: vpmovm2w %k0, %xmm0
2422 ; X86-NEXT: retl
2423 %b = load <4 x i1>, <4 x i1>* %a
2424 %c = sext <4 x i1> %b to <4 x i16>
2425 ret <4 x i16> %c
2426 }
2428 define <32 x i16> @load_32i1(<32 x i1>* %a) {
2429 ; KNL-LABEL: load_32i1:
2430 ; KNL: ## %bb.0:
2431 ; KNL-NEXT: kmovw (%rdi), %k1
2432 ; KNL-NEXT: kmovw 2(%rdi), %k2
2433 ; KNL-NEXT: vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k2} {z}
2434 ; KNL-NEXT: vpmovdw %zmm0, %ymm0
2435 ; KNL-NEXT: vpternlogd $255, %zmm1, %zmm1, %zmm1 {%k1} {z}
2436 ; KNL-NEXT: vpmovdw %zmm1, %ymm1
2437 ; KNL-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm0
2438 ; KNL-NEXT: retq
2439 ;
2440 ; SKX-LABEL: load_32i1:
2441 ; SKX: ## %bb.0:
2442 ; SKX-NEXT: kmovd (%rdi), %k0
2443 ; SKX-NEXT: vpmovm2w %k0, %zmm0
2444 ; SKX-NEXT: retq
2445 ;
2446 ; AVX512BW-LABEL: load_32i1:
2447 ; AVX512BW: ## %bb.0:
2448 ; AVX512BW-NEXT: kmovd (%rdi), %k0
2449 ; AVX512BW-NEXT: vpmovm2w %k0, %zmm0
2450 ; AVX512BW-NEXT: retq
2451 ;
2452 ; AVX512DQ-LABEL: load_32i1:
2453 ; AVX512DQ: ## %bb.0:
2454 ; AVX512DQ-NEXT: kmovw (%rdi), %k0
2455 ; AVX512DQ-NEXT: kmovw 2(%rdi), %k1
2456 ; AVX512DQ-NEXT: vpmovm2d %k1, %zmm0
2457 ; AVX512DQ-NEXT: vpmovdw %zmm0, %ymm0
2458 ; AVX512DQ-NEXT: vpmovm2d %k0, %zmm1
2459 ; AVX512DQ-NEXT: vpmovdw %zmm1, %ymm1
2460 ; AVX512DQ-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm0
2461 ; AVX512DQ-NEXT: retq
2462 ;
2463 ; X86-LABEL: load_32i1:
2464 ; X86: ## %bb.0:
2465 ; X86-NEXT: movl {{[0-9]+}}(%esp), %eax
2466 ; X86-NEXT: kmovd (%eax), %k0
2467 ; X86-NEXT: vpmovm2w %k0, %zmm0
2468 ; X86-NEXT: retl
2469 %b = load <32 x i1>, <32 x i1>* %a
2470 %c = sext <32 x i1> %b to <32 x i16>
2471 ret <32 x i16> %c
2472 }
2474 define <64 x i8> @load_64i1(<64 x i1>* %a) {
2475 ; KNL-LABEL: load_64i1:
2476 ; KNL: ## %bb.0:
2477 ; KNL-NEXT: kmovw (%rdi), %k1
2478 ; KNL-NEXT: kmovw 2(%rdi), %k2
2479 ; KNL-NEXT: kmovw 4(%rdi), %k3
2480 ; KNL-NEXT: kmovw 6(%rdi), %k4
2481 ; KNL-NEXT: vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k3} {z}
2482 ; KNL-NEXT: vpmovdb %zmm0, %xmm0
2483 ; KNL-NEXT: vpternlogd $255, %zmm1, %zmm1, %zmm1 {%k4} {z}
2484 ; KNL-NEXT: vpmovdb %zmm1, %xmm1
2485 ; KNL-NEXT: vinserti128 $1, %xmm1, %ymm0, %ymm0
2486 ; KNL-NEXT: vpternlogd $255, %zmm1, %zmm1, %zmm1 {%k1} {z}
2487 ; KNL-NEXT: vpmovdb %zmm1, %xmm1
2488 ; KNL-NEXT: vpternlogd $255, %zmm2, %zmm2, %zmm2 {%k2} {z}
2489 ; KNL-NEXT: vpmovdb %zmm2, %xmm2
2490 ; KNL-NEXT: vinserti128 $1, %xmm2, %ymm1, %ymm1
2491 ; KNL-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm0
2492 ; KNL-NEXT: retq
2493 ;
2494 ; SKX-LABEL: load_64i1:
2495 ; SKX: ## %bb.0:
2496 ; SKX-NEXT: kmovq (%rdi), %k0
2497 ; SKX-NEXT: vpmovm2b %k0, %zmm0
2498 ; SKX-NEXT: retq
2499 ;
2500 ; AVX512BW-LABEL: load_64i1:
2501 ; AVX512BW: ## %bb.0:
2502 ; AVX512BW-NEXT: kmovq (%rdi), %k0
2503 ; AVX512BW-NEXT: vpmovm2b %k0, %zmm0
2504 ; AVX512BW-NEXT: retq
2505 ;
2506 ; AVX512DQ-LABEL: load_64i1:
2507 ; AVX512DQ: ## %bb.0:
2508 ; AVX512DQ-NEXT: kmovw (%rdi), %k0
2509 ; AVX512DQ-NEXT: kmovw 2(%rdi), %k1
2510 ; AVX512DQ-NEXT: kmovw 4(%rdi), %k2
2511 ; AVX512DQ-NEXT: kmovw 6(%rdi), %k3
2512 ; AVX512DQ-NEXT: vpmovm2d %k2, %zmm0
2513 ; AVX512DQ-NEXT: vpmovdb %zmm0, %xmm0
2514 ; AVX512DQ-NEXT: vpmovm2d %k3, %zmm1
2515 ; AVX512DQ-NEXT: vpmovdb %zmm1, %xmm1
2516 ; AVX512DQ-NEXT: vinserti128 $1, %xmm1, %ymm0, %ymm0
2517 ; AVX512DQ-NEXT: vpmovm2d %k0, %zmm1
2518 ; AVX512DQ-NEXT: vpmovdb %zmm1, %xmm1
2519 ; AVX512DQ-NEXT: vpmovm2d %k1, %zmm2
2520 ; AVX512DQ-NEXT: vpmovdb %zmm2, %xmm2
2521 ; AVX512DQ-NEXT: vinserti128 $1, %xmm2, %ymm1, %ymm1
2522 ; AVX512DQ-NEXT: vinserti64x4 $1, %ymm0, %zmm1, %zmm0
2523 ; AVX512DQ-NEXT: retq
2524 ;
2525 ; X86-LABEL: load_64i1:
2526 ; X86: ## %bb.0:
2527 ; X86-NEXT: movl {{[0-9]+}}(%esp), %eax
2528 ; X86-NEXT: kmovq (%eax), %k0
2529 ; X86-NEXT: vpmovm2b %k0, %zmm0
2530 ; X86-NEXT: retl
2531 %b = load <64 x i1>, <64 x i1>* %a
2532 %c = sext <64 x i1> %b to <64 x i8>
2533 ret <64 x i8> %c
2534 }
2536 define void @store_8i1(<8 x i1>* %a, <8 x i1> %v) {
2537 ; KNL-LABEL: store_8i1:
2538 ; KNL: ## %bb.0:
2539 ; KNL-NEXT: vpmovsxwq %xmm0, %zmm0
2540 ; KNL-NEXT: vpsllq $63, %zmm0, %zmm0
2541 ; KNL-NEXT: vptestmq %zmm0, %zmm0, %k0
2542 ; KNL-NEXT: kmovw %k0, %eax
2543 ; KNL-NEXT: movb %al, (%rdi)
2544 ; KNL-NEXT: vzeroupper
2545 ; KNL-NEXT: retq
2546 ;
2547 ; SKX-LABEL: store_8i1:
2548 ; SKX: ## %bb.0:
2549 ; SKX-NEXT: vpsllw $15, %xmm0, %xmm0
2550 ; SKX-NEXT: vpmovw2m %xmm0, %k0
2551 ; SKX-NEXT: kmovb %k0, (%rdi)
2552 ; SKX-NEXT: retq
2553 ;
2554 ; AVX512BW-LABEL: store_8i1:
2555 ; AVX512BW: ## %bb.0:
2556 ; AVX512BW-NEXT: vpsllw $15, %xmm0, %xmm0
2557 ; AVX512BW-NEXT: vpmovw2m %zmm0, %k0
2558 ; AVX512BW-NEXT: kmovd %k0, %eax
2559 ; AVX512BW-NEXT: movb %al, (%rdi)
2560 ; AVX512BW-NEXT: vzeroupper
2561 ; AVX512BW-NEXT: retq
2562 ;
2563 ; AVX512DQ-LABEL: store_8i1:
2564 ; AVX512DQ: ## %bb.0:
2565 ; AVX512DQ-NEXT: vpmovsxwq %xmm0, %zmm0
2566 ; AVX512DQ-NEXT: vpsllq $63, %zmm0, %zmm0
2567 ; AVX512DQ-NEXT: vpmovq2m %zmm0, %k0
2568 ; AVX512DQ-NEXT: kmovb %k0, (%rdi)
2569 ; AVX512DQ-NEXT: vzeroupper
2570 ; AVX512DQ-NEXT: retq
2571 ;
2572 ; X86-LABEL: store_8i1:
2573 ; X86: ## %bb.0:
2574 ; X86-NEXT: vpsllw $15, %xmm0, %xmm0
2575 ; X86-NEXT: vpmovw2m %xmm0, %k0
2576 ; X86-NEXT: movl {{[0-9]+}}(%esp), %eax
2577 ; X86-NEXT: kmovb %k0, (%eax)
2578 ; X86-NEXT: retl
2579 store <8 x i1> %v, <8 x i1>* %a
2580 ret void
2581 }
2583 define void @store_8i1_1(<8 x i1>* %a, <8 x i16> %v) {
2584 ; KNL-LABEL: store_8i1_1:
2585 ; KNL: ## %bb.0:
2586 ; KNL-NEXT: vpmovsxwq %xmm0, %zmm0
2587 ; KNL-NEXT: vpsllq $63, %zmm0, %zmm0
2588 ; KNL-NEXT: vptestmq %zmm0, %zmm0, %k0
2589 ; KNL-NEXT: kmovw %k0, %eax
2590 ; KNL-NEXT: movb %al, (%rdi)
2591 ; KNL-NEXT: vzeroupper
2592 ; KNL-NEXT: retq
2593 ;
2594 ; SKX-LABEL: store_8i1_1:
2595 ; SKX: ## %bb.0:
2596 ; SKX-NEXT: vpsllw $15, %xmm0, %xmm0
2597 ; SKX-NEXT: vpmovw2m %xmm0, %k0
2598 ; SKX-NEXT: kmovb %k0, (%rdi)
2599 ; SKX-NEXT: retq
2600 ;
2601 ; AVX512BW-LABEL: store_8i1_1:
2602 ; AVX512BW: ## %bb.0:
2603 ; AVX512BW-NEXT: vpsllw $15, %xmm0, %xmm0
2604 ; AVX512BW-NEXT: vpmovw2m %zmm0, %k0
2605 ; AVX512BW-NEXT: kmovd %k0, %eax
2606 ; AVX512BW-NEXT: movb %al, (%rdi)
2607 ; AVX512BW-NEXT: vzeroupper
2608 ; AVX512BW-NEXT: retq
2609 ;
2610 ; AVX512DQ-LABEL: store_8i1_1:
2611 ; AVX512DQ: ## %bb.0:
2612 ; AVX512DQ-NEXT: vpmovsxwq %xmm0, %zmm0
2613 ; AVX512DQ-NEXT: vpsllq $63, %zmm0, %zmm0
2614 ; AVX512DQ-NEXT: vpmovq2m %zmm0, %k0
2615 ; AVX512DQ-NEXT: kmovb %k0, (%rdi)
2616 ; AVX512DQ-NEXT: vzeroupper
2617 ; AVX512DQ-NEXT: retq
2618 ;
2619 ; X86-LABEL: store_8i1_1:
2620 ; X86: ## %bb.0:
2621 ; X86-NEXT: movl {{[0-9]+}}(%esp), %eax
2622 ; X86-NEXT: vpsllw $15, %xmm0, %xmm0
2623 ; X86-NEXT: vpmovw2m %xmm0, %k0
2624 ; X86-NEXT: kmovb %k0, (%eax)
2625 ; X86-NEXT: retl
2626 %v1 = trunc <8 x i16> %v to <8 x i1>
2627 store <8 x i1> %v1, <8 x i1>* %a
2628 ret void
2629 }
2631 define void @store_16i1(<16 x i1>* %a, <16 x i1> %v) {
2632 ; KNL-LABEL: store_16i1:
2633 ; KNL: ## %bb.0:
2634 ; KNL-NEXT: vpmovsxbd %xmm0, %zmm0
2635 ; KNL-NEXT: vpslld $31, %zmm0, %zmm0
2636 ; KNL-NEXT: vptestmd %zmm0, %zmm0, %k0
2637 ; KNL-NEXT: kmovw %k0, (%rdi)
2638 ; KNL-NEXT: vzeroupper
2639 ; KNL-NEXT: retq
2640 ;
2641 ; SKX-LABEL: store_16i1:
2642 ; SKX: ## %bb.0:
2643 ; SKX-NEXT: vpsllw $7, %xmm0, %xmm0
2644 ; SKX-NEXT: vpmovb2m %xmm0, %k0
2645 ; SKX-NEXT: kmovw %k0, (%rdi)
2646 ; SKX-NEXT: retq
2647 ;
2648 ; AVX512BW-LABEL: store_16i1:
2649 ; AVX512BW: ## %bb.0:
2650 ; AVX512BW-NEXT: vpsllw $7, %xmm0, %xmm0
2651 ; AVX512BW-NEXT: vpmovb2m %zmm0, %k0
2652 ; AVX512BW-NEXT: kmovw %k0, (%rdi)
2653 ; AVX512BW-NEXT: vzeroupper
2654 ; AVX512BW-NEXT: retq
2655 ;
2656 ; AVX512DQ-LABEL: store_16i1:
2657 ; AVX512DQ: ## %bb.0:
2658 ; AVX512DQ-NEXT: vpmovsxbd %xmm0, %zmm0
2659 ; AVX512DQ-NEXT: vpslld $31, %zmm0, %zmm0
2660 ; AVX512DQ-NEXT: vpmovd2m %zmm0, %k0
2661 ; AVX512DQ-NEXT: kmovw %k0, (%rdi)
2662 ; AVX512DQ-NEXT: vzeroupper
2663 ; AVX512DQ-NEXT: retq
2664 ;
2665 ; X86-LABEL: store_16i1:
2666 ; X86: ## %bb.0:
2667 ; X86-NEXT: vpsllw $7, %xmm0, %xmm0
2668 ; X86-NEXT: vpmovb2m %xmm0, %k0
2669 ; X86-NEXT: movl {{[0-9]+}}(%esp), %eax
2670 ; X86-NEXT: kmovw %k0, (%eax)
2671 ; X86-NEXT: retl
2672 store <16 x i1> %v, <16 x i1>* %a
2673 ret void
2674 }
2676 define void @store_32i1(<32 x i1>* %a, <32 x i1> %v) {
2677 ; KNL-LABEL: store_32i1:
2678 ; KNL: ## %bb.0:
2679 ; KNL-NEXT: vpmovsxbd %xmm0, %zmm1
2680 ; KNL-NEXT: vpslld $31, %zmm1, %zmm1
2681 ; KNL-NEXT: vptestmd %zmm1, %zmm1, %k0
2682 ; KNL-NEXT: vextracti128 $1, %ymm0, %xmm0
2683 ; KNL-NEXT: vpmovsxbd %xmm0, %zmm0
2684 ; KNL-NEXT: vpslld $31, %zmm0, %zmm0
2685 ; KNL-NEXT: vptestmd %zmm0, %zmm0, %k1
2686 ; KNL-NEXT: kmovw %k1, 2(%rdi)
2687 ; KNL-NEXT: kmovw %k0, (%rdi)
2688 ; KNL-NEXT: vzeroupper
2689 ; KNL-NEXT: retq
2690 ;
2691 ; SKX-LABEL: store_32i1:
2692 ; SKX: ## %bb.0:
2693 ; SKX-NEXT: vpsllw $7, %ymm0, %ymm0
2694 ; SKX-NEXT: vpmovb2m %ymm0, %k0
2695 ; SKX-NEXT: kmovd %k0, (%rdi)
2696 ; SKX-NEXT: vzeroupper
2697 ; SKX-NEXT: retq
2698 ;
2699 ; AVX512BW-LABEL: store_32i1:
2700 ; AVX512BW: ## %bb.0:
2701 ; AVX512BW-NEXT: vpsllw $7, %ymm0, %ymm0
2702 ; AVX512BW-NEXT: vpmovb2m %zmm0, %k0
2703 ; AVX512BW-NEXT: kmovd %k0, (%rdi)
2704 ; AVX512BW-NEXT: vzeroupper
2705 ; AVX512BW-NEXT: retq
2706 ;
2707 ; AVX512DQ-LABEL: store_32i1:
2708 ; AVX512DQ: ## %bb.0:
2709 ; AVX512DQ-NEXT: vpmovsxbd %xmm0, %zmm1
2710 ; AVX512DQ-NEXT: vpslld $31, %zmm1, %zmm1
2711 ; AVX512DQ-NEXT: vpmovd2m %zmm1, %k0
2712 ; AVX512DQ-NEXT: vextracti128 $1, %ymm0, %xmm0
2713 ; AVX512DQ-NEXT: vpmovsxbd %xmm0, %zmm0
2714 ; AVX512DQ-NEXT: vpslld $31, %zmm0, %zmm0
2715 ; AVX512DQ-NEXT: vpmovd2m %zmm0, %k1
2716 ; AVX512DQ-NEXT: kmovw %k1, 2(%rdi)
2717 ; AVX512DQ-NEXT: kmovw %k0, (%rdi)
2718 ; AVX512DQ-NEXT: vzeroupper
2719 ; AVX512DQ-NEXT: retq
2720 ;
2721 ; X86-LABEL: store_32i1:
2722 ; X86: ## %bb.0:
2723 ; X86-NEXT: vpsllw $7, %ymm0, %ymm0
2724 ; X86-NEXT: vpmovb2m %ymm0, %k0
2725 ; X86-NEXT: movl {{[0-9]+}}(%esp), %eax
2726 ; X86-NEXT: kmovd %k0, (%eax)
2727 ; X86-NEXT: vzeroupper
2728 ; X86-NEXT: retl
2729 store <32 x i1> %v, <32 x i1>* %a
2730 ret void
2731 }
2733 define void @store_32i1_1(<32 x i1>* %a, <32 x i16> %v) {
2734 ; KNL-LABEL: store_32i1_1:
2735 ; KNL: ## %bb.0:
2736 ; KNL-NEXT: vextracti64x4 $1, %zmm0, %ymm1
2737 ; KNL-NEXT: vpmovsxwd %ymm0, %zmm0
2738 ; KNL-NEXT: vpslld $31, %zmm0, %zmm0
2739 ; KNL-NEXT: vptestmd %zmm0, %zmm0, %k0
2740 ; KNL-NEXT: vpmovsxwd %ymm1, %zmm0
2741 ; KNL-NEXT: vpslld $31, %zmm0, %zmm0
2742 ; KNL-NEXT: vptestmd %zmm0, %zmm0, %k1
2743 ; KNL-NEXT: kmovw %k1, 2(%rdi)
2744 ; KNL-NEXT: kmovw %k0, (%rdi)
2745 ; KNL-NEXT: vzeroupper
2746 ; KNL-NEXT: retq
2747 ;
2748 ; SKX-LABEL: store_32i1_1:
2749 ; SKX: ## %bb.0:
2750 ; SKX-NEXT: vpsllw $15, %zmm0, %zmm0
2751 ; SKX-NEXT: vpmovw2m %zmm0, %k0
2752 ; SKX-NEXT: kmovd %k0, (%rdi)
2753 ; SKX-NEXT: vzeroupper
2754 ; SKX-NEXT: retq
2755 ;
2756 ; AVX512BW-LABEL: store_32i1_1:
2757 ; AVX512BW: ## %bb.0:
2758 ; AVX512BW-NEXT: vpsllw $15, %zmm0, %zmm0
2759 ; AVX512BW-NEXT: vpmovw2m %zmm0, %k0
2760 ; AVX512BW-NEXT: kmovd %k0, (%rdi)
2761 ; AVX512BW-NEXT: vzeroupper
2762 ; AVX512BW-NEXT: retq
2763 ;
2764 ; AVX512DQ-LABEL: store_32i1_1:
2765 ; AVX512DQ: ## %bb.0:
2766 ; AVX512DQ-NEXT: vextracti64x4 $1, %zmm0, %ymm1
2767 ; AVX512DQ-NEXT: vpmovsxwd %ymm0, %zmm0
2768 ; AVX512DQ-NEXT: vpslld $31, %zmm0, %zmm0
2769 ; AVX512DQ-NEXT: vpmovd2m %zmm0, %k0
2770 ; AVX512DQ-NEXT: vpmovsxwd %ymm1, %zmm0
2771 ; AVX512DQ-NEXT: vpslld $31, %zmm0, %zmm0
2772 ; AVX512DQ-NEXT: vpmovd2m %zmm0, %k1
2773 ; AVX512DQ-NEXT: kmovw %k1, 2(%rdi)
2774 ; AVX512DQ-NEXT: kmovw %k0, (%rdi)
2775 ; AVX512DQ-NEXT: vzeroupper
2776 ; AVX512DQ-NEXT: retq
2777 ;
2778 ; X86-LABEL: store_32i1_1:
2779 ; X86: ## %bb.0:
2780 ; X86-NEXT: movl {{[0-9]+}}(%esp), %eax
2781 ; X86-NEXT: vpsllw $15, %zmm0, %zmm0
2782 ; X86-NEXT: vpmovw2m %zmm0, %k0
2783 ; X86-NEXT: kmovd %k0, (%eax)
2784 ; X86-NEXT: vzeroupper
2785 ; X86-NEXT: retl
2786 %v1 = trunc <32 x i16> %v to <32 x i1>
2787 store <32 x i1> %v1, <32 x i1>* %a
2788 ret void
2789 }
2792 define void @store_64i1(<64 x i1>* %a, <64 x i1> %v) {
2793 ;
2794 ; KNL-LABEL: store_64i1:
2795 ; KNL: ## %bb.0:
2796 ; KNL-NEXT: kmovw %esi, %k0
2797 ; KNL-NEXT: kshiftlw $15, %k0, %k0
2798 ; KNL-NEXT: kshiftrw $15, %k0, %k0
2799 ; KNL-NEXT: kshiftlw $2, %k0, %k2
2800 ; KNL-NEXT: kmovw %k2, {{[-0-9]+}}(%r{{[sb]}}p) ## 2-byte Spill
2801 ; KNL-NEXT: kmovw %edx, %k1
2802 ; KNL-NEXT: kshiftlw $1, %k1, %k1
2803 ; KNL-NEXT: korw %k1, %k2, %k1
2804 ; KNL-NEXT: korw %k1, %k0, %k0
2805 ; KNL-NEXT: kshiftlw $14, %k0, %k0
2806 ; KNL-NEXT: kshiftrw $14, %k0, %k0
2807 ; KNL-NEXT: kshiftlw $3, %k0, %k3
2808 ; KNL-NEXT: kmovw %ecx, %k1
2809 ; KNL-NEXT: kshiftlw $2, %k1, %k1
2810 ; KNL-NEXT: korw %k1, %k3, %k1
2811 ; KNL-NEXT: korw %k1, %k0, %k0
2812 ; KNL-NEXT: kshiftlw $13, %k0, %k0
2813 ; KNL-NEXT: kshiftrw $13, %k0, %k0
2814 ; KNL-NEXT: kshiftlw $4, %k0, %k4
2815 ; KNL-NEXT: kmovw %r8d, %k1
2816 ; KNL-NEXT: kshiftlw $3, %k1, %k1
2817 ; KNL-NEXT: korw %k1, %k4, %k1
2818 ; KNL-NEXT: korw %k1, %k0, %k0
2819 ; KNL-NEXT: kshiftlw $12, %k0, %k0
2820 ; KNL-NEXT: kshiftrw $12, %k0, %k0
2821 ; KNL-NEXT: kshiftlw $5, %k0, %k5
2822 ; KNL-NEXT: kmovw %r9d, %k1
2823 ; KNL-NEXT: kshiftlw $4, %k1, %k1
2824 ; KNL-NEXT: korw %k1, %k5, %k1
2825 ; KNL-NEXT: korw %k1, %k0, %k0
2826 ; KNL-NEXT: kshiftlw $11, %k0, %k0
2827 ; KNL-NEXT: kshiftrw $11, %k0, %k0
2828 ; KNL-NEXT: kshiftlw $6, %k0, %k6
2829 ; KNL-NEXT: movb {{[0-9]+}}(%rsp), %al
2830 ; KNL-NEXT: kmovw %eax, %k1
2831 ; KNL-NEXT: kshiftlw $5, %k1, %k1
2832 ; KNL-NEXT: korw %k1, %k6, %k1
2833 ; KNL-NEXT: korw %k1, %k0, %k0
2834 ; KNL-NEXT: kshiftlw $10, %k0, %k0
2835 ; KNL-NEXT: kshiftrw $10, %k0, %k0
2836 ; KNL-NEXT: kshiftlw $7, %k0, %k7
2837 ; KNL-NEXT: movb {{[0-9]+}}(%rsp), %al
2838 ; KNL-NEXT: kmovw %eax, %k1
2839 ; KNL-NEXT: kshiftlw $6, %k1, %k1
2840 ; KNL-NEXT: korw %k1, %k7, %k1
2841 ; KNL-NEXT: korw %k1, %k0, %k0
2842 ; KNL-NEXT: kshiftlw $9, %k0, %k0
2843 ; KNL-NEXT: kshiftrw $9, %k0, %k0
2844 ; KNL-NEXT: movb {{[0-9]+}}(%rsp), %al
2845 ; KNL-NEXT: kmovw %eax, %k1
2846 ; KNL-NEXT: kshiftlw $7, %k1, %k1
2847 ; KNL-NEXT: kshiftlw $8, %k0, %k2
2848 ; KNL-NEXT: kmovw %k2, {{[-0-9]+}}(%r{{[sb]}}p) ## 2-byte Spill
2849 ; KNL-NEXT: korw %k1, %k2, %k1
2850 ; KNL-NEXT: korw %k1, %k0, %k0
2851 ; KNL-NEXT: kshiftlw $8, %k0, %k0
2852 ; KNL-NEXT: kshiftrw $8, %k0, %k0
2853 ; KNL-NEXT: movb {{[0-9]+}}(%rsp), %al
2854 ; KNL-NEXT: kmovw %eax, %k1
2855 ; KNL-NEXT: kshiftlw $8, %k1, %k1
2856 ; KNL-NEXT: kshiftlw $9, %k0, %k2
2857 ; KNL-NEXT: kmovw %k2, {{[-0-9]+}}(%r{{[sb]}}p) ## 2-byte Spill
2858 ; KNL-NEXT: korw %k1, %k2, %k1
2859 ; KNL-NEXT: korw %k1, %k0, %k0
2860 ; KNL-NEXT: kshiftlw $7, %k0, %k0
2861 ; KNL-NEXT: kshiftrw $7, %k0, %k0
2862 ; KNL-NEXT: movb {{[0-9]+}}(%rsp), %al
2863 ; KNL-NEXT: kmovw %eax, %k1
2864 ; KNL-NEXT: kshiftlw $9, %k1, %k1
2865 ; KNL-NEXT: kshiftlw $10, %k0, %k2
2866 ; KNL-NEXT: kmovw %k2, {{[-0-9]+}}(%r{{[sb]}}p) ## 2-byte Spill
2867 ; KNL-NEXT: korw %k1, %k2, %k1
2868 ; KNL-NEXT: korw %k1, %k0, %k0
2869 ; KNL-NEXT: kshiftlw $6, %k0, %k0
2870 ; KNL-NEXT: kshiftrw $6, %k0, %k0
2871 ; KNL-NEXT: movb {{[0-9]+}}(%rsp), %al
2872 ; KNL-NEXT: kmovw %eax, %k1
2873 ; KNL-NEXT: kshiftlw $10, %k1, %k1
2874 ; KNL-NEXT: kshiftlw $11, %k0, %k2
2875 ; KNL-NEXT: kmovw %k2, {{[-0-9]+}}(%r{{[sb]}}p) ## 2-byte Spill
2876 ; KNL-NEXT: korw %k1, %k2, %k1
2877 ; KNL-NEXT: korw %k1, %k0, %k0
2878 ; KNL-NEXT: kshiftlw $5, %k0, %k0
2879 ; KNL-NEXT: kshiftrw $5, %k0, %k0
2880 ; KNL-NEXT: movb {{[0-9]+}}(%rsp), %al
2881 ; KNL-NEXT: kmovw %eax, %k1
2882 ; KNL-NEXT: kshiftlw $11, %k1, %k1
2883 ; KNL-NEXT: kshiftlw $12, %k0, %k2
2884 ; KNL-NEXT: kmovw %k2, {{[-0-9]+}}(%r{{[sb]}}p) ## 2-byte Spill
2885 ; KNL-NEXT: korw %k1, %k2, %k1
2886 ; KNL-NEXT: korw %k1, %k0, %k0
2887 ; KNL-NEXT: kshiftlw $4, %k0, %k0
2888 ; KNL-NEXT: kshiftrw $4, %k0, %k0
2889 ; KNL-NEXT: movb {{[0-9]+}}(%rsp), %al
2890 ; KNL-NEXT: kmovw %eax, %k1
2891 ; KNL-NEXT: kshiftlw $12, %k1, %k1
2892 ; KNL-NEXT: kshiftlw $13, %k0, %k2
2893 ; KNL-NEXT: kmovw %k2, {{[-0-9]+}}(%r{{[sb]}}p) ## 2-byte Spill
2894 ; KNL-NEXT: korw %k1, %k2, %k1
2895 ; KNL-NEXT: korw %k1, %k0, %k0
2896 ; KNL-NEXT: kshiftlw $3, %k0, %k0
2897 ; KNL-NEXT: kshiftrw $3, %k0, %k0
2898 ; KNL-NEXT: movb {{[0-9]+}}(%rsp), %al
2899 ; KNL-NEXT: kmovw %eax, %k1
2900 ; KNL-NEXT: kshiftlw $13, %k1, %k1
2901 ; KNL-NEXT: kshiftlw $14, %k0, %k2
2902 ; KNL-NEXT: kmovw %k2, {{[-0-9]+}}(%r{{[sb]}}p) ## 2-byte Spill
2903 ; KNL-NEXT: korw %k1, %k2, %k1
2904 ; KNL-NEXT: korw %k1, %k0, %k0
2905 ; KNL-NEXT: kshiftlw $2, %k0, %k0
2906 ; KNL-NEXT: kshiftrw $2, %k0, %k2
2907 ; KNL-NEXT: movb {{[0-9]+}}(%rsp), %al
2908 ; KNL-NEXT: kmovw %eax, %k1
2909 ; KNL-NEXT: kshiftlw $14, %k1, %k0
2910 ; KNL-NEXT: kshiftlw $15, %k0, %k1
2911 ; KNL-NEXT: kmovw %k1, {{[-0-9]+}}(%r{{[sb]}}p) ## 2-byte Spill
2912 ; KNL-NEXT: korw %k0, %k1, %k0
2913 ; KNL-NEXT: korw %k0, %k2, %k0
2914 ; KNL-NEXT: kshiftlw $1, %k0, %k0
2915 ; KNL-NEXT: kshiftrw $1, %k0, %k0
2916 ; KNL-NEXT: movb {{[0-9]+}}(%rsp), %al
2917 ; KNL-NEXT: kmovw %eax, %k2
2918 ; KNL-NEXT: kshiftlw $15, %k2, %k2
2919 ; KNL-NEXT: korw %k2, %k0, %k0
2920 ; KNL-NEXT: kmovw %k0, {{[-0-9]+}}(%r{{[sb]}}p) ## 2-byte Spill
2921 ; KNL-NEXT: movb {{[0-9]+}}(%rsp), %al
2922 ; KNL-NEXT: kmovw %eax, %k0
2923 ; KNL-NEXT: kshiftlw $15, %k0, %k0
2924 ; KNL-NEXT: kshiftrw $15, %k0, %k0
2925 ; KNL-NEXT: movb {{[0-9]+}}(%rsp), %al
2926 ; KNL-NEXT: kmovw %eax, %k2
2927 ; KNL-NEXT: kshiftlw $1, %k2, %k2
2928 ; KNL-NEXT: kmovw {{[-0-9]+}}(%r{{[sb]}}p), %k1 ## 2-byte Reload
2929 ; KNL-NEXT: korw %k2, %k1, %k2
2930 ; KNL-NEXT: korw %k2, %k0, %k0
2931 ; KNL-NEXT: kshiftlw $14, %k0, %k0
2932 ; KNL-NEXT: kshiftrw $14, %k0, %k0
2933 ; KNL-NEXT: movb {{[0-9]+}}(%rsp), %al
2934 ; KNL-NEXT: kmovw %eax, %k2
2935 ; KNL-NEXT: kshiftlw $2, %k2, %k2
2936 ; KNL-NEXT: kmovw %k3, {{[-0-9]+}}(%r{{[sb]}}p) ## 2-byte Spill
2937 ; KNL-NEXT: korw %k2, %k3, %k2
2938 ; KNL-NEXT: korw %k2, %k0, %k0
2939 ; KNL-NEXT: kshiftlw $13, %k0, %k0
2940 ; KNL-NEXT: kshiftrw $13, %k0, %k0
2941 ; KNL-NEXT: movb {{[0-9]+}}(%rsp), %al
2942 ; KNL-NEXT: kmovw %eax, %k2
2943 ; KNL-NEXT: kshiftlw $3, %k2, %k2
2944 ; KNL-NEXT: kmovw %k4, {{[-0-9]+}}(%r{{[sb]}}p) ## 2-byte Spill
2945 ; KNL-NEXT: korw %k2, %k4, %k2
2946 ; KNL-NEXT: korw %k2, %k0, %k0
2947 ; KNL-NEXT: kshiftlw $12, %k0, %k0
2948 ; KNL-NEXT: kshiftrw $12, %k0, %k0
2949 ; KNL-NEXT: movb {{[0-9]+}}(%rsp), %al
2950 ; KNL-NEXT: kmovw %eax, %k2
2951 ; KNL-NEXT: kshiftlw $4, %k2, %k2
2952 ; KNL-NEXT: kmovw %k5, {{[-0-9]+}}(%r{{[sb]}}p) ## 2-byte Spill
2953 ; KNL-NEXT: korw %k2, %k5, %k2
2954 ; KNL-NEXT: korw %k2, %k0, %k0
2955 ; KNL-NEXT: kshiftlw $11, %k0, %k0
2956 ; KNL-NEXT: kshiftrw $11, %k0, %k0
2957 ; KNL-NEXT: movb {{[0-9]+}}(%rsp), %al
2958 ; KNL-NEXT: kmovw %eax, %k2
2959 ; KNL-NEXT: kshiftlw $5, %k2, %k2
2960 ; KNL-NEXT: korw %k2, %k6, %k2
2961 ; KNL-NEXT: korw %k2, %k0, %k0
2962 ; KNL-NEXT: kshiftlw $10, %k0, %k0
2963 ; KNL-NEXT: kshiftrw $10, %k0, %k0
2964 ; KNL-NEXT: movb {{[0-9]+}}(%rsp), %al
2965 ; KNL-NEXT: kmovw %eax, %k2
2966 ; KNL-NEXT: kshiftlw $6, %k2, %k2
2967 ; KNL-NEXT: kmovw %k7, {{[-0-9]+}}(%r{{[sb]}}p) ## 2-byte Spill
2968 ; KNL-NEXT: korw %k2, %k7, %k2
2969 ; KNL-NEXT: korw %k2, %k0, %k0
2970 ; KNL-NEXT: kshiftlw $9, %k0, %k0
2971 ; KNL-NEXT: kshiftrw $9, %k0, %k0
2972 ; KNL-NEXT: movb {{[0-9]+}}(%rsp), %al
2973 ; KNL-NEXT: kmovw %eax, %k2
2974 ; KNL-NEXT: kshiftlw $7, %k2, %k2
2975 ; KNL-NEXT: kmovw {{[-0-9]+}}(%r{{[sb]}}p), %k1 ## 2-byte Reload
2976 ; KNL-NEXT: korw %k2, %k1, %k2
2977 ; KNL-NEXT: korw %k2, %k0, %k0
2978 ; KNL-NEXT: kshiftlw $8, %k0, %k0
2979 ; KNL-NEXT: kshiftrw $8, %k0, %k0
2980 ; KNL-NEXT: movb {{[0-9]+}}(%rsp), %al
2981 ; KNL-NEXT: kmovw %eax, %k2
2982 ; KNL-NEXT: kshiftlw $8, %k2, %k2
2983 ; KNL-NEXT: kmovw {{[-0-9]+}}(%r{{[sb]}}p), %k1 ## 2-byte Reload
2984 ; KNL-NEXT: korw %k2, %k1, %k2
2985 ; KNL-NEXT: korw %k2, %k0, %k0
2986 ; KNL-NEXT: kshiftlw $7, %k0, %k0
2987 ; KNL-NEXT: kshiftrw $7, %k0, %k0
2988 ; KNL-NEXT: movb {{[0-9]+}}(%rsp), %al
2989 ; KNL-NEXT: kmovw %eax, %k2
2990 ; KNL-NEXT: kshiftlw $9, %k2, %k2
2991 ; KNL-NEXT: kmovw {{[-0-9]+}}(%r{{[sb]}}p), %k1 ## 2-byte Reload
2992 ; KNL-NEXT: korw %k2, %k1, %k2
2993 ; KNL-NEXT: korw %k2, %k0, %k0
2994 ; KNL-NEXT: kshiftlw $6, %k0, %k0
2995 ; KNL-NEXT: kshiftrw $6, %k0, %k0
2996 ; KNL-NEXT: movb {{[0-9]+}}(%rsp), %al
2997 ; KNL-NEXT: kmovw %eax, %k2
2998 ; KNL-NEXT: kshiftlw $10, %k2, %k2
2999 ; KNL-NEXT: kmovw {{[-0-9]+}}(%r{{[sb]}}p), %k1 ## 2-byte Reload
3000 ; KNL-NEXT: korw %k2, %k1, %k2
3001 ; KNL-NEXT: korw %k2, %k0, %k0
3002 ; KNL-NEXT: kshiftlw $5, %k0, %k0
3003 ; KNL-NEXT: kshiftrw $5, %k0, %k0
3004 ; KNL-NEXT: movb {{[0-9]+}}(%rsp), %al
3005 ; KNL-NEXT: kmovw %eax, %k2
3006 ; KNL-NEXT: kshiftlw $11, %k2, %k2
3007 ; KNL-NEXT: kmovw {{[-0-9]+}}(%r{{[sb]}}p), %k1 ## 2-byte Reload
3008 ; KNL-NEXT: korw %k2, %k1, %k2
3009 ; KNL-NEXT: korw %k2, %k0, %k0
3010 ; KNL-NEXT: kshiftlw $4, %k0, %k0
3011 ; KNL-NEXT: kshiftrw $4, %k0, %k0
3012 ; KNL-NEXT: movb {{[0-9]+}}(%rsp), %al
3013 ; KNL-NEXT: kmovw %eax, %k2
3014 ; KNL-NEXT: kshiftlw $12, %k2, %k2
3015 ; KNL-NEXT: kmovw {{[-0-9]+}}(%r{{[sb]}}p), %k1 ## 2-byte Reload
3016 ; KNL-NEXT: korw %k2, %k1, %k2
3017 ; KNL-NEXT: korw %k2, %k0, %k0
3018 ; KNL-NEXT: kshiftlw $3, %k0, %k0
3019 ; KNL-NEXT: kshiftrw $3, %k0, %k0
3020 ; KNL-NEXT: movb {{[0-9]+}}(%rsp), %al
3021 ; KNL-NEXT: kmovw %eax, %k2
3022 ; KNL-NEXT: kshiftlw $13, %k2, %k2
3023 ; KNL-NEXT: kmovw {{[-0-9]+}}(%r{{[sb]}}p), %k1 ## 2-byte Reload
3024 ; KNL-NEXT: korw %k2, %k1, %k2
3025 ; KNL-NEXT: korw %k2, %k0, %k0
3026 ; KNL-NEXT: kshiftlw $2, %k0, %k0
3027 ; KNL-NEXT: kshiftrw $2, %k0, %k0
3028 ; KNL-NEXT: movb {{[0-9]+}}(%rsp), %al
3029 ; KNL-NEXT: kmovw %eax, %k2
3030 ; KNL-NEXT: kshiftlw $14, %k2, %k2
3031 ; KNL-NEXT: kmovw {{[-0-9]+}}(%r{{[sb]}}p), %k1 ## 2-byte Reload
3032 ; KNL-NEXT: korw %k2, %k1, %k2
3033 ; KNL-NEXT: korw %k2, %k0, %k0
3034 ; KNL-NEXT: kshiftlw $1, %k0, %k0
3035 ; KNL-NEXT: kshiftrw $1, %k0, %k0
3036 ; KNL-NEXT: movb {{[0-9]+}}(%rsp), %al
3037 ; KNL-NEXT: kmovw %eax, %k2
3038 ; KNL-NEXT: kshiftlw $15, %k2, %k2
3039 ; KNL-NEXT: korw %k2, %k0, %k0
3040 ; KNL-NEXT: kmovw %k0, {{[-0-9]+}}(%r{{[sb]}}p) ## 2-byte Spill
3041 ; KNL-NEXT: movb {{[0-9]+}}(%rsp), %al
3042 ; KNL-NEXT: kmovw %eax, %k0
3043 ; KNL-NEXT: kshiftlw $15, %k0, %k0
3044 ; KNL-NEXT: kshiftrw $15, %k0, %k0
3045 ; KNL-NEXT: movb {{[0-9]+}}(%rsp), %al
3046 ; KNL-NEXT: kmovw %eax, %k2
3047 ; KNL-NEXT: kshiftlw $1, %k2, %k2
3048 ; KNL-NEXT: kmovw {{[-0-9]+}}(%r{{[sb]}}p), %k1 ## 2-byte Reload
3049 ; KNL-NEXT: korw %k2, %k1, %k2
3050 ; KNL-NEXT: korw %k2, %k0, %k0
3051 ; KNL-NEXT: kshiftlw $14, %k0, %k0
3052 ; KNL-NEXT: kshiftrw $14, %k0, %k0
3053 ; KNL-NEXT: movb {{[0-9]+}}(%rsp), %al
3054 ; KNL-NEXT: kmovw %eax, %k2
3055 ; KNL-NEXT: kshiftlw $2, %k2, %k2
3056 ; KNL-NEXT: korw %k2, %k3, %k2
3057 ; KNL-NEXT: korw %k2, %k0, %k0
3058 ; KNL-NEXT: kshiftlw $13, %k0, %k0
3059 ; KNL-NEXT: kshiftrw $13, %k0, %k0
3060 ; KNL-NEXT: movb {{[0-9]+}}(%rsp), %al
3061 ; KNL-NEXT: kmovw %eax, %k2
3062 ; KNL-NEXT: kshiftlw $3, %k2, %k2
3063 ; KNL-NEXT: korw %k2, %k4, %k2
3064 ; KNL-NEXT: korw %k2, %k0, %k0
3065 ; KNL-NEXT: kshiftlw $12, %k0, %k0
3066 ; KNL-NEXT: kshiftrw $12, %k0, %k0
3067 ; KNL-NEXT: movb {{[0-9]+}}(%rsp), %al
3068 ; KNL-NEXT: kmovw %eax, %k2
3069 ; KNL-NEXT: kshiftlw $4, %k2, %k2
3070 ; KNL-NEXT: korw %k2, %k5, %k2
3071 ; KNL-NEXT: korw %k2, %k0, %k0
3072 ; KNL-NEXT: kshiftlw $11, %k0, %k0
3073 ; KNL-NEXT: kshiftrw $11, %k0, %k0
3074 ; KNL-NEXT: movb {{[0-9]+}}(%rsp), %al
3075 ; KNL-NEXT: kmovw %eax, %k2
3076 ; KNL-NEXT: kshiftlw $5, %k2, %k2
3077 ; KNL-NEXT: korw %k2, %k6, %k2
3078 ; KNL-NEXT: korw %k2, %k0, %k0
3079 ; KNL-NEXT: kshiftlw $10, %k0, %k0
3080 ; KNL-NEXT: kshiftrw $10, %k0, %k0
3081 ; KNL-NEXT: movb {{[0-9]+}}(%rsp), %al
3082 ; KNL-NEXT: kmovw %eax, %k2
3083 ; KNL-NEXT: kshiftlw $6, %k2, %k2
3084 ; KNL-NEXT: korw %k2, %k7, %k2
3085 ; KNL-NEXT: korw %k2, %k0, %k0
3086 ; KNL-NEXT: kshiftlw $9, %k0, %k0
3087 ; KNL-NEXT: kshiftrw $9, %k0, %k0
3088 ; KNL-NEXT: movb {{[0-9]+}}(%rsp), %al
3089 ; KNL-NEXT: kmovw %eax, %k2
3090 ; KNL-NEXT: kshiftlw $7, %k2, %k2
3091 ; KNL-NEXT: kmovw {{[-0-9]+}}(%r{{[sb]}}p), %k1 ## 2-byte Reload
3092 ; KNL-NEXT: korw %k2, %k1, %k2
3093 ; KNL-NEXT: korw %k2, %k0, %k0
3094 ; KNL-NEXT: kshiftlw $8, %k0, %k0
3095 ; KNL-NEXT: kshiftrw $8, %k0, %k0
3096 ; KNL-NEXT: movb {{[0-9]+}}(%rsp), %al
3097 ; KNL-NEXT: kmovw %eax, %k2
3098 ; KNL-NEXT: kshiftlw $8, %k2, %k2
3099 ; KNL-NEXT: kmovw {{[-0-9]+}}(%r{{[sb]}}p), %k3 ## 2-byte Reload
3100 ; KNL-NEXT: korw %k2, %k3, %k2
3101 ; KNL-NEXT: korw %k2, %k0, %k0
3102 ; KNL-NEXT: kshiftlw $7, %k0, %k0
3103 ; KNL-NEXT: kshiftrw $7, %k0, %k0
3104 ; KNL-NEXT: movb {{[0-9]+}}(%rsp), %al
3105 ; KNL-NEXT: kmovw %eax, %k2
3106 ; KNL-NEXT: kshiftlw $9, %k2, %k2
3107 ; KNL-NEXT: kmovw {{[-0-9]+}}(%r{{[sb]}}p), %k4 ## 2-byte Reload
3108 ; KNL-NEXT: korw %k2, %k4, %k2
3109 ; KNL-NEXT: korw %k2, %k0, %k0
3110 ; KNL-NEXT: kshiftlw $6, %k0, %k0
3111 ; KNL-NEXT: kshiftrw $6, %k0, %k0
3112 ; KNL-NEXT: movb {{[0-9]+}}(%rsp), %al
3113 ; KNL-NEXT: kmovw %eax, %k2
3114 ; KNL-NEXT: kshiftlw $10, %k2, %k2
3115 ; KNL-NEXT: kmovw {{[-0-9]+}}(%r{{[sb]}}p), %k5 ## 2-byte Reload
3116 ; KNL-NEXT: korw %k2, %k5, %k2
3117 ; KNL-NEXT: korw %k2, %k0, %k0
3118 ; KNL-NEXT: kshiftlw $5, %k0, %k0
3119 ; KNL-NEXT: kshiftrw $5, %k0, %k0
3120 ; KNL-NEXT: movb {{[0-9]+}}(%rsp), %al
3121 ; KNL-NEXT: kmovw %eax, %k2
3122 ; KNL-NEXT: kshiftlw $11, %k2, %k2
3123 ; KNL-NEXT: kmovw {{[-0-9]+}}(%r{{[sb]}}p), %k7 ## 2-byte Reload
3124 ; KNL-NEXT: korw %k2, %k7, %k2
3125 ; KNL-NEXT: korw %k2, %k0, %k0
3126 ; KNL-NEXT: kshiftlw $4, %k0, %k0
3127 ; KNL-NEXT: kshiftrw $4, %k0, %k0
3128 ; KNL-NEXT: movb {{[0-9]+}}(%rsp), %al
3129 ; KNL-NEXT: kmovw %eax, %k2
3130 ; KNL-NEXT: kshiftlw $12, %k2, %k2
3131 ; KNL-NEXT: kmovw {{[-0-9]+}}(%r{{[sb]}}p), %k7 ## 2-byte Reload
3132 ; KNL-NEXT: korw %k2, %k7, %k2
3133 ; KNL-NEXT: korw %k2, %k0, %k0
3134 ; KNL-NEXT: kshiftlw $3, %k0, %k0
3135 ; KNL-NEXT: kshiftrw $3, %k0, %k0
3136 ; KNL-NEXT: movb {{[0-9]+}}(%rsp), %al
3137 ; KNL-NEXT: kmovw %eax, %k2
3138 ; KNL-NEXT: kshiftlw $13, %k2, %k2
3139 ; KNL-NEXT: kmovw {{[-0-9]+}}(%r{{[sb]}}p), %k7 ## 2-byte Reload
3140 ; KNL-NEXT: korw %k2, %k7, %k2
3141 ; KNL-NEXT: korw %k2, %k0, %k0
3142 ; KNL-NEXT: kshiftlw $2, %k0, %k0
3143 ; KNL-NEXT: kshiftrw $2, %k0, %k0
3144 ; KNL-NEXT: movb {{[0-9]+}}(%rsp), %al
3145 ; KNL-NEXT: kmovw %eax, %k2
3146 ; KNL-NEXT: kshiftlw $14, %k2, %k2
3147 ; KNL-NEXT: kmovw {{[-0-9]+}}(%r{{[sb]}}p), %k7 ## 2-byte Reload
3148 ; KNL-NEXT: korw %k2, %k7, %k2
3149 ; KNL-NEXT: korw %k2, %k0, %k0
3150 ; KNL-NEXT: kshiftlw $1, %k0, %k0
3151 ; KNL-NEXT: kshiftrw $1, %k0, %k0
3152 ; KNL-NEXT: movb {{[0-9]+}}(%rsp), %al
3153 ; KNL-NEXT: kmovw %eax, %k2
3154 ; KNL-NEXT: kshiftlw $15, %k2, %k2
3155 ; KNL-NEXT: korw %k2, %k0, %k0
3156 ; KNL-NEXT: kmovw %k0, {{[-0-9]+}}(%r{{[sb]}}p) ## 2-byte Spill
3157 ; KNL-NEXT: movb {{[0-9]+}}(%rsp), %al
3158 ; KNL-NEXT: kmovw %eax, %k2
3159 ; KNL-NEXT: kshiftlw $1, %k2, %k2
3160 ; KNL-NEXT: kmovw {{[-0-9]+}}(%r{{[sb]}}p), %k0 ## 2-byte Reload
3161 ; KNL-NEXT: korw %k2, %k0, %k2
3162 ; KNL-NEXT: movb {{[0-9]+}}(%rsp), %al
3163 ; KNL-NEXT: kmovw %eax, %k7
3164 ; KNL-NEXT: kshiftlw $15, %k7, %k7
3165 ; KNL-NEXT: kshiftrw $15, %k7, %k7
3166 ; KNL-NEXT: korw %k2, %k7, %k2
3167 ; KNL-NEXT: movb {{[0-9]+}}(%rsp), %al
3168 ; KNL-NEXT: kmovw %eax, %k7
3169 ; KNL-NEXT: kshiftlw $2, %k7, %k7
3170 ; KNL-NEXT: kmovw {{[-0-9]+}}(%r{{[sb]}}p), %k0 ## 2-byte Reload
3171 ; KNL-NEXT: korw %k7, %k0, %k7
3172 ; KNL-NEXT: kshiftlw $14, %k2, %k2
3173 ; KNL-NEXT: kshiftrw $14, %k2, %k2
3174 ; KNL-NEXT: korw %k7, %k2, %k2
3175 ; KNL-NEXT: movb {{[0-9]+}}(%rsp), %al
3176 ; KNL-NEXT: kmovw %eax, %k7
3177 ; KNL-NEXT: kshiftlw $3, %k7, %k7
3178 ; KNL-NEXT: kmovw {{[-0-9]+}}(%r{{[sb]}}p), %k0 ## 2-byte Reload
3179 ; KNL-NEXT: korw %k7, %k0, %k7
3180 ; KNL-NEXT: kshiftlw $13, %k2, %k2
3181 ; KNL-NEXT: kshiftrw $13, %k2, %k2
3182 ; KNL-NEXT: korw %k7, %k2, %k2
3183 ; KNL-NEXT: movb {{[0-9]+}}(%rsp), %al
3184 ; KNL-NEXT: kmovw %eax, %k7
3185 ; KNL-NEXT: kshiftlw $4, %k7, %k7
3186 ; KNL-NEXT: kmovw {{[-0-9]+}}(%r{{[sb]}}p), %k0 ## 2-byte Reload
3187 ; KNL-NEXT: korw %k7, %k0, %k7
3188 ; KNL-NEXT: kshiftlw $12, %k2, %k2
3189 ; KNL-NEXT: kshiftrw $12, %k2, %k2
3190 ; KNL-NEXT: korw %k7, %k2, %k2
3191 ; KNL-NEXT: movb {{[0-9]+}}(%rsp), %al
3192 ; KNL-NEXT: kmovw %eax, %k7
3193 ; KNL-NEXT: kshiftlw $5, %k7, %k7
3194 ; KNL-NEXT: korw %k7, %k6, %k7
3195 ; KNL-NEXT: kshiftlw $11, %k2, %k2
3196 ; KNL-NEXT: kshiftrw $11, %k2, %k2
3197 ; KNL-NEXT: korw %k7, %k2, %k2
3198 ; KNL-NEXT: movb {{[0-9]+}}(%rsp), %al
3199 ; KNL-NEXT: kmovw %eax, %k7
3200 ; KNL-NEXT: kshiftlw $6, %k7, %k7
3201 ; KNL-NEXT: kmovw {{[-0-9]+}}(%r{{[sb]}}p), %k0 ## 2-byte Reload
3202 ; KNL-NEXT: korw %k7, %k0, %k7
3203 ; KNL-NEXT: kshiftlw $10, %k2, %k2
3204 ; KNL-NEXT: kshiftrw $10, %k2, %k2
3205 ; KNL-NEXT: korw %k7, %k2, %k2
3206 ; KNL-NEXT: movb {{[0-9]+}}(%rsp), %al
3207 ; KNL-NEXT: kmovw %eax, %k7
3208 ; KNL-NEXT: kshiftlw $7, %k7, %k7
3209 ; KNL-NEXT: korw %k7, %k1, %k7
3210 ; KNL-NEXT: kshiftlw $9, %k2, %k2
3211 ; KNL-NEXT: kshiftrw $9, %k2, %k2
3212 ; KNL-NEXT: korw %k7, %k2, %k2
3213 ; KNL-NEXT: movb {{[0-9]+}}(%rsp), %al
3214 ; KNL-NEXT: kmovw %eax, %k7
3215 ; KNL-NEXT: kshiftlw $8, %k7, %k7
3216 ; KNL-NEXT: korw %k7, %k3, %k7
3217 ; KNL-NEXT: kshiftlw $8, %k2, %k2
3218 ; KNL-NEXT: kshiftrw $8, %k2, %k2
3219 ; KNL-NEXT: korw %k7, %k2, %k2
3220 ; KNL-NEXT: movb {{[0-9]+}}(%rsp), %al
3221 ; KNL-NEXT: kmovw %eax, %k7
3222 ; KNL-NEXT: kshiftlw $9, %k7, %k7
3223 ; KNL-NEXT: korw %k7, %k4, %k7
3224 ; KNL-NEXT: kshiftlw $7, %k2, %k2
3225 ; KNL-NEXT: kshiftrw $7, %k2, %k2
3226 ; KNL-NEXT: korw %k7, %k2, %k2
3227 ; KNL-NEXT: movb {{[0-9]+}}(%rsp), %al
3228 ; KNL-NEXT: kmovw %eax, %k7
3229 ; KNL-NEXT: kshiftlw $10, %k7, %k7
3230 ; KNL-NEXT: korw %k7, %k5, %k6
3231 ; KNL-NEXT: kshiftlw $6, %k2, %k2
3232 ; KNL-NEXT: kshiftrw $6, %k2, %k2
3233 ; KNL-NEXT: korw %k6, %k2, %k2
3234 ; KNL-NEXT: movb {{[0-9]+}}(%rsp), %al
3235 ; KNL-NEXT: kmovw %eax, %k6
3236 ; KNL-NEXT: kshiftlw $11, %k6, %k6
3237 ; KNL-NEXT: kmovw {{[-0-9]+}}(%r{{[sb]}}p), %k1 ## 2-byte Reload
3238 ; KNL-NEXT: korw %k6, %k1, %k5
3239 ; KNL-NEXT: kshiftlw $5, %k2, %k2
3240 ; KNL-NEXT: kshiftrw $5, %k2, %k2
3241 ; KNL-NEXT: korw %k5, %k2, %k2
3242 ; KNL-NEXT: movb {{[0-9]+}}(%rsp), %al
3243 ; KNL-NEXT: kmovw %eax, %k5
3244 ; KNL-NEXT: kshiftlw $12, %k5, %k5
3245 ; KNL-NEXT: kmovw {{[-0-9]+}}(%r{{[sb]}}p), %k1 ## 2-byte Reload
3246 ; KNL-NEXT: korw %k5, %k1, %k4
3247 ; KNL-NEXT: kshiftlw $4, %k2, %k2
3248 ; KNL-NEXT: kshiftrw $4, %k2, %k2
3249 ; KNL-NEXT: korw %k4, %k2, %k2
3250 ; KNL-NEXT: movb {{[0-9]+}}(%rsp), %al
3251 ; KNL-NEXT: kmovw %eax, %k4
3252 ; KNL-NEXT: kshiftlw $13, %k4, %k4
3253 ; KNL-NEXT: kmovw {{[-0-9]+}}(%r{{[sb]}}p), %k1 ## 2-byte Reload
3254 ; KNL-NEXT: korw %k4, %k1, %k3
3255 ; KNL-NEXT: kshiftlw $3, %k2, %k2
3256 ; KNL-NEXT: kshiftrw $3, %k2, %k2
3257 ; KNL-NEXT: korw %k3, %k2, %k2
3258 ; KNL-NEXT: movb {{[0-9]+}}(%rsp), %al
3259 ; KNL-NEXT: kmovw %eax, %k3
3260 ; KNL-NEXT: kshiftlw $14, %k3, %k3
3261 ; KNL-NEXT: kmovw {{[-0-9]+}}(%r{{[sb]}}p), %k1 ## 2-byte Reload
3262 ; KNL-NEXT: korw %k3, %k1, %k1
3263 ; KNL-NEXT: kshiftlw $2, %k2, %k2
3264 ; KNL-NEXT: kshiftrw $2, %k2, %k2
3265 ; KNL-NEXT: korw %k1, %k2, %k1
3266 ; KNL-NEXT: kshiftlw $1, %k1, %k1
3267 ; KNL-NEXT: kshiftrw $1, %k1, %k1
3268 ; KNL-NEXT: movb {{[0-9]+}}(%rsp), %al
3269 ; KNL-NEXT: kmovw %eax, %k2
3270 ; KNL-NEXT: kshiftlw $15, %k2, %k2
3271 ; KNL-NEXT: korw %k2, %k1, %k1
3272 ; KNL-NEXT: kmovw %k1, 6(%rdi)
3273 ; KNL-NEXT: kmovw {{[-0-9]+}}(%r{{[sb]}}p), %k0 ## 2-byte Reload
3274 ; KNL-NEXT: kmovw %k0, 4(%rdi)
3275 ; KNL-NEXT: kmovw {{[-0-9]+}}(%r{{[sb]}}p), %k0 ## 2-byte Reload
3276 ; KNL-NEXT: kmovw %k0, 2(%rdi)
3277 ; KNL-NEXT: kmovw {{[-0-9]+}}(%r{{[sb]}}p), %k0 ## 2-byte Reload
3278 ; KNL-NEXT: kmovw %k0, (%rdi)
3279 ; KNL-NEXT: retq
3280 ;
3281 ; SKX-LABEL: store_64i1:
3282 ; SKX: ## %bb.0:
3283 ; SKX-NEXT: vpsllw $7, %zmm0, %zmm0
3284 ; SKX-NEXT: vpmovb2m %zmm0, %k0
3285 ; SKX-NEXT: kmovq %k0, (%rdi)
3286 ; SKX-NEXT: vzeroupper
3287 ; SKX-NEXT: retq
3288 ;
3289 ; AVX512BW-LABEL: store_64i1:
3290 ; AVX512BW: ## %bb.0:
3291 ; AVX512BW-NEXT: vpsllw $7, %zmm0, %zmm0
3292 ; AVX512BW-NEXT: vpmovb2m %zmm0, %k0
3293 ; AVX512BW-NEXT: kmovq %k0, (%rdi)
3294 ; AVX512BW-NEXT: vzeroupper
3295 ; AVX512BW-NEXT: retq
3296 ;
3297 ; AVX512DQ-LABEL: store_64i1:
3298 ; AVX512DQ: ## %bb.0:
3299 ; AVX512DQ-NEXT: kmovw %esi, %k0
3300 ; AVX512DQ-NEXT: kshiftlw $15, %k0, %k0
3301 ; AVX512DQ-NEXT: kshiftrw $15, %k0, %k0
3302 ; AVX512DQ-NEXT: kshiftlw $2, %k0, %k2
3303 ; AVX512DQ-NEXT: kmovw %k2, {{[-0-9]+}}(%r{{[sb]}}p) ## 2-byte Spill
3304 ; AVX512DQ-NEXT: kmovw %edx, %k1
3305 ; AVX512DQ-NEXT: kshiftlw $1, %k1, %k1
3306 ; AVX512DQ-NEXT: korw %k1, %k2, %k1
3307 ; AVX512DQ-NEXT: korw %k1, %k0, %k0
3308 ; AVX512DQ-NEXT: kshiftlw $14, %k0, %k0
3309 ; AVX512DQ-NEXT: kshiftrw $14, %k0, %k0
3310 ; AVX512DQ-NEXT: kshiftlw $3, %k0, %k3
3311 ; AVX512DQ-NEXT: kmovw %ecx, %k1
3312 ; AVX512DQ-NEXT: kshiftlw $2, %k1, %k1
3313 ; AVX512DQ-NEXT: korw %k1, %k3, %k1
3314 ; AVX512DQ-NEXT: korw %k1, %k0, %k0
3315 ; AVX512DQ-NEXT: kshiftlw $13, %k0, %k0
3316 ; AVX512DQ-NEXT: kshiftrw $13, %k0, %k0
3317 ; AVX512DQ-NEXT: kshiftlw $4, %k0, %k4
3318 ; AVX512DQ-NEXT: kmovw %r8d, %k1
3319 ; AVX512DQ-NEXT: kshiftlw $3, %k1, %k1
3320 ; AVX512DQ-NEXT: korw %k1, %k4, %k1
3321 ; AVX512DQ-NEXT: korw %k1, %k0, %k0
3322 ; AVX512DQ-NEXT: kshiftlw $12, %k0, %k0
3323 ; AVX512DQ-NEXT: kshiftrw $12, %k0, %k0
3324 ; AVX512DQ-NEXT: kshiftlw $5, %k0, %k5
3325 ; AVX512DQ-NEXT: kmovw %r9d, %k1
3326 ; AVX512DQ-NEXT: kshiftlw $4, %k1, %k1
3327 ; AVX512DQ-NEXT: korw %k1, %k5, %k1
3328 ; AVX512DQ-NEXT: korw %k1, %k0, %k0
3329 ; AVX512DQ-NEXT: kshiftlw $11, %k0, %k0
3330 ; AVX512DQ-NEXT: kshiftrw $11, %k0, %k0
3331 ; AVX512DQ-NEXT: kshiftlw $6, %k0, %k6
3332 ; AVX512DQ-NEXT: movb {{[0-9]+}}(%rsp), %al
3333 ; AVX512DQ-NEXT: kmovw %eax, %k1
3334 ; AVX512DQ-NEXT: kshiftlw $5, %k1, %k1
3335 ; AVX512DQ-NEXT: korw %k1, %k6, %k1
3336 ; AVX512DQ-NEXT: korw %k1, %k0, %k0
3337 ; AVX512DQ-NEXT: kshiftlw $10, %k0, %k0
3338 ; AVX512DQ-NEXT: kshiftrw $10, %k0, %k0
3339 ; AVX512DQ-NEXT: kshiftlw $7, %k0, %k7
3340 ; AVX512DQ-NEXT: movb {{[0-9]+}}(%rsp), %al
3341 ; AVX512DQ-NEXT: kmovw %eax, %k1
3342 ; AVX512DQ-NEXT: kshiftlw $6, %k1, %k1
3343 ; AVX512DQ-NEXT: korw %k1, %k7, %k1
3344 ; AVX512DQ-NEXT: korw %k1, %k0, %k0
3345 ; AVX512DQ-NEXT: kshiftlw $9, %k0, %k0
3346 ; AVX512DQ-NEXT: kshiftrw $9, %k0, %k0
3347 ; AVX512DQ-NEXT: movb {{[0-9]+}}(%rsp), %al
3348 ; AVX512DQ-NEXT: kmovw %eax, %k1
3349 ; AVX512DQ-NEXT: kshiftlw $7, %k1, %k1
3350 ; AVX512DQ-NEXT: kshiftlw $8, %k0, %k2
3351 ; AVX512DQ-NEXT: kmovw %k2, {{[-0-9]+}}(%r{{[sb]}}p) ## 2-byte Spill
3352 ; AVX512DQ-NEXT: korw %k1, %k2, %k1
3353 ; AVX512DQ-NEXT: korw %k1, %k0, %k0
3354 ; AVX512DQ-NEXT: kshiftlw $8, %k0, %k0
3355 ; AVX512DQ-NEXT: kshiftrw $8, %k0, %k0
3356 ; AVX512DQ-NEXT: movb {{[0-9]+}}(%rsp), %al
3357 ; AVX512DQ-NEXT: kmovw %eax, %k1
3358 ; AVX512DQ-NEXT: kshiftlw $8, %k1, %k1
3359 ; AVX512DQ-NEXT: kshiftlw $9, %k0, %k2
3360 ; AVX512DQ-NEXT: kmovw %k2, {{[-0-9]+}}(%r{{[sb]}}p) ## 2-byte Spill
3361 ; AVX512DQ-NEXT: korw %k1, %k2, %k1
3362 ; AVX512DQ-NEXT: korw %k1, %k0, %k0
3363 ; AVX512DQ-NEXT: kshiftlw $7, %k0, %k0
3364 ; AVX512DQ-NEXT: kshiftrw $7, %k0, %k0
3365 ; AVX512DQ-NEXT: movb {{[0-9]+}}(%rsp), %al
3366 ; AVX512DQ-NEXT: kmovw %eax, %k1
3367 ; AVX512DQ-NEXT: kshiftlw $9, %k1, %k1
3368 ; AVX512DQ-NEXT: kshiftlw $10, %k0, %k2
3369 ; AVX512DQ-NEXT: kmovw %k2, {{[-0-9]+}}(%r{{[sb]}}p) ## 2-byte Spill
3370 ; AVX512DQ-NEXT: korw %k1, %k2, %k1
3371 ; AVX512DQ-NEXT: korw %k1, %k0, %k0
3372 ; AVX512DQ-NEXT: kshiftlw $6, %k0, %k0
3373 ; AVX512DQ-NEXT: kshiftrw $6, %k0, %k0
3374 ; AVX512DQ-NEXT: movb {{[0-9]+}}(%rsp), %al
3375 ; AVX512DQ-NEXT: kmovw %eax, %k1
3376 ; AVX512DQ-NEXT: kshiftlw $10, %k1, %k1
3377 ; AVX512DQ-NEXT: kshiftlw $11, %k0, %k2
3378 ; AVX512DQ-NEXT: kmovw %k2, {{[-0-9]+}}(%r{{[sb]}}p) ## 2-byte Spill
3379 ; AVX512DQ-NEXT: korw %k1, %k2, %k1
3380 ; AVX512DQ-NEXT: korw %k1, %k0, %k0
3381 ; AVX512DQ-NEXT: kshiftlw $5, %k0, %k0
3382 ; AVX512DQ-NEXT: kshiftrw $5, %k0, %k0
3383 ; AVX512DQ-NEXT: movb {{[0-9]+}}(%rsp), %al
3384 ; AVX512DQ-NEXT: kmovw %eax, %k1
3385 ; AVX512DQ-NEXT: kshiftlw $11, %k1, %k1
3386 ; AVX512DQ-NEXT: kshiftlw $12, %k0, %k2
3387 ; AVX512DQ-NEXT: kmovw %k2, {{[-0-9]+}}(%r{{[sb]}}p) ## 2-byte Spill
3388 ; AVX512DQ-NEXT: korw %k1, %k2, %k1
3389 ; AVX512DQ-NEXT: korw %k1, %k0, %k0
3390 ; AVX512DQ-NEXT: kshiftlw $4, %k0, %k0
3391 ; AVX512DQ-NEXT: kshiftrw $4, %k0, %k0
3392 ; AVX512DQ-NEXT: movb {{[0-9]+}}(%rsp), %al
3393 ; AVX512DQ-NEXT: kmovw %eax, %k1
3394 ; AVX512DQ-NEXT: kshiftlw $12, %k1, %k1
3395 ; AVX512DQ-NEXT: kshiftlw $13, %k0, %k2
3396 ; AVX512DQ-NEXT: kmovw %k2, {{[-0-9]+}}(%r{{[sb]}}p) ## 2-byte Spill
3397 ; AVX512DQ-NEXT: korw %k1, %k2, %k1
3398 ; AVX512DQ-NEXT: korw %k1, %k0, %k0
3399 ; AVX512DQ-NEXT: kshiftlw $3, %k0, %k0
3400 ; AVX512DQ-NEXT: kshiftrw $3, %k0, %k0
3401 ; AVX512DQ-NEXT: movb {{[0-9]+}}(%rsp), %al
3402 ; AVX512DQ-NEXT: kmovw %eax, %k1
3403 ; AVX512DQ-NEXT: kshiftlw $13, %k1, %k1
3404 ; AVX512DQ-NEXT: kshiftlw $14, %k0, %k2
3405 ; AVX512DQ-NEXT: kmovw %k2, {{[-0-9]+}}(%r{{[sb]}}p) ## 2-byte Spill
3406 ; AVX512DQ-NEXT: korw %k1, %k2, %k1
3407 ; AVX512DQ-NEXT: korw %k1, %k0, %k0
3408 ; AVX512DQ-NEXT: kshiftlw $2, %k0, %k0
3409 ; AVX512DQ-NEXT: kshiftrw $2, %k0, %k2
3410 ; AVX512DQ-NEXT: movb {{[0-9]+}}(%rsp), %al
3411 ; AVX512DQ-NEXT: kmovw %eax, %k1
3412 ; AVX512DQ-NEXT: kshiftlw $14, %k1, %k0
3413 ; AVX512DQ-NEXT: kshiftlw $15, %k0, %k1
3414 ; AVX512DQ-NEXT: kmovw %k1, {{[-0-9]+}}(%r{{[sb]}}p) ## 2-byte Spill
3415 ; AVX512DQ-NEXT: korw %k0, %k1, %k0
3416 ; AVX512DQ-NEXT: korw %k0, %k2, %k0
3417 ; AVX512DQ-NEXT: kshiftlw $1, %k0, %k0
3418 ; AVX512DQ-NEXT: kshiftrw $1, %k0, %k0
3419 ; AVX512DQ-NEXT: movb {{[0-9]+}}(%rsp), %al
3420 ; AVX512DQ-NEXT: kmovw %eax, %k2
3421 ; AVX512DQ-NEXT: kshiftlw $15, %k2, %k2
3422 ; AVX512DQ-NEXT: korw %k2, %k0, %k0
3423 ; AVX512DQ-NEXT: kmovw %k0, {{[-0-9]+}}(%r{{[sb]}}p) ## 2-byte Spill
3424 ; AVX512DQ-NEXT: movb {{[0-9]+}}(%rsp), %al
3425 ; AVX512DQ-NEXT: kmovw %eax, %k0
3426 ; AVX512DQ-NEXT: kshiftlw $15, %k0, %k0
3427 ; AVX512DQ-NEXT: kshiftrw $15, %k0, %k0
3428 ; AVX512DQ-NEXT: movb {{[0-9]+}}(%rsp), %al
3429 ; AVX512DQ-NEXT: kmovw %eax, %k2
3430 ; AVX512DQ-NEXT: kshiftlw $1, %k2, %k2
3431 ; AVX512DQ-NEXT: kmovw {{[-0-9]+}}(%r{{[sb]}}p), %k1 ## 2-byte Reload
3432 ; AVX512DQ-NEXT: korw %k2, %k1, %k2
3433 ; AVX512DQ-NEXT: korw %k2, %k0, %k0
3434 ; AVX512DQ-NEXT: kshiftlw $14, %k0, %k0
3435 ; AVX512DQ-NEXT: kshiftrw $14, %k0, %k0
3436 ; AVX512DQ-NEXT: movb {{[0-9]+}}(%rsp), %al
3437 ; AVX512DQ-NEXT: kmovw %eax, %k2
3438 ; AVX512DQ-NEXT: kshiftlw $2, %k2, %k2
3439 ; AVX512DQ-NEXT: kmovw %k3, {{[-0-9]+}}(%r{{[sb]}}p) ## 2-byte Spill
3440 ; AVX512DQ-NEXT: korw %k2, %k3, %k2
3441 ; AVX512DQ-NEXT: korw %k2, %k0, %k0
3442 ; AVX512DQ-NEXT: kshiftlw $13, %k0, %k0
3443 ; AVX512DQ-NEXT: kshiftrw $13, %k0, %k0
3444 ; AVX512DQ-NEXT: movb {{[0-9]+}}(%rsp), %al
3445 ; AVX512DQ-NEXT: kmovw %eax, %k2
3446 ; AVX512DQ-NEXT: kshiftlw $3, %k2, %k2
3447 ; AVX512DQ-NEXT: kmovw %k4, {{[-0-9]+}}(%r{{[sb]}}p) ## 2-byte Spill
3448 ; AVX512DQ-NEXT: korw %k2, %k4, %k2
3449 ; AVX512DQ-NEXT: korw %k2, %k0, %k0
3450 ; AVX512DQ-NEXT: kshiftlw $12, %k0, %k0
3451 ; AVX512DQ-NEXT: kshiftrw $12, %k0, %k0
3452 ; AVX512DQ-NEXT: movb {{[0-9]+}}(%rsp), %al
3453 ; AVX512DQ-NEXT: kmovw %eax, %k2
3454 ; AVX512DQ-NEXT: kshiftlw $4, %k2, %k2
3455 ; AVX512DQ-NEXT: kmovw %k5, {{[-0-9]+}}(%r{{[sb]}}p) ## 2-byte Spill
3456 ; AVX512DQ-NEXT: korw %k2, %k5, %k2
3457 ; AVX512DQ-NEXT: korw %k2, %k0, %k0
3458 ; AVX512DQ-NEXT: kshiftlw $11, %k0, %k0
3459 ; AVX512DQ-NEXT: kshiftrw $11, %k0, %k0
3460 ; AVX512DQ-NEXT: movb {{[0-9]+}}(%rsp), %al
3461 ; AVX512DQ-NEXT: kmovw %eax, %k2
3462 ; AVX512DQ-NEXT: kshiftlw $5, %k2, %k2
3463 ; AVX512DQ-NEXT: korw %k2, %k6, %k2
3464 ; AVX512DQ-NEXT: korw %k2, %k0, %k0
3465 ; AVX512DQ-NEXT: kshiftlw $10, %k0, %k0
3466 ; AVX512DQ-NEXT: kshiftrw $10, %k0, %k0
3467 ; AVX512DQ-NEXT: movb {{[0-9]+}}(%rsp), %al
3468 ; AVX512DQ-NEXT: kmovw %eax, %k2
3469 ; AVX512DQ-NEXT: kshiftlw $6, %k2, %k2
3470 ; AVX512DQ-NEXT: kmovw %k7, {{[-0-9]+}}(%r{{[sb]}}p) ## 2-byte Spill
3471 ; AVX512DQ-NEXT: korw %k2, %k7, %k2
3472 ; AVX512DQ-NEXT: korw %k2, %k0, %k0
3473 ; AVX512DQ-NEXT: kshiftlw $9, %k0, %k0
3474 ; AVX512DQ-NEXT: kshiftrw $9, %k0, %k0
3475 ; AVX512DQ-NEXT: movb {{[0-9]+}}(%rsp), %al
3476 ; AVX512DQ-NEXT: kmovw %eax, %k2
3477 ; AVX512DQ-NEXT: kshiftlw $7, %k2, %k2
3478 ; AVX512DQ-NEXT: kmovw {{[-0-9]+}}(%r{{[sb]}}p), %k1 ## 2-byte Reload
3479 ; AVX512DQ-NEXT: korw %k2, %k1, %k2
3480 ; AVX512DQ-NEXT: korw %k2, %k0, %k0
3481 ; AVX512DQ-NEXT: kshiftlw $8, %k0, %k0
3482 ; AVX512DQ-NEXT: kshiftrw $8, %k0, %k0
3483 ; AVX512DQ-NEXT: movb {{[0-9]+}}(%rsp), %al
3484 ; AVX512DQ-NEXT: kmovw %eax, %k2
3485 ; AVX512DQ-NEXT: kshiftlw $8, %k2, %k2
3486 ; AVX512DQ-NEXT: kmovw {{[-0-9]+}}(%r{{[sb]}}p), %k1 ## 2-byte Reload
3487 ; AVX512DQ-NEXT: korw %k2, %k1, %k2
3488 ; AVX512DQ-NEXT: korw %k2, %k0, %k0
3489 ; AVX512DQ-NEXT: kshiftlw $7, %k0, %k0
3490 ; AVX512DQ-NEXT: kshiftrw $7, %k0, %k0
3491 ; AVX512DQ-NEXT: movb {{[0-9]+}}(%rsp), %al
3492 ; AVX512DQ-NEXT: kmovw %eax, %k2
3493 ; AVX512DQ-NEXT: kshiftlw $9, %k2, %k2
3494 ; AVX512DQ-NEXT: kmovw {{[-0-9]+}}(%r{{[sb]}}p), %k1 ## 2-byte Reload
3495 ; AVX512DQ-NEXT: korw %k2, %k1, %k2
3496 ; AVX512DQ-NEXT: korw %k2, %k0, %k0
3497 ; AVX512DQ-NEXT: kshiftlw $6, %k0, %k0
3498 ; AVX512DQ-NEXT: kshiftrw $6, %k0, %k0
3499 ; AVX512DQ-NEXT: movb {{[0-9]+}}(%rsp), %al
3500 ; AVX512DQ-NEXT: kmovw %eax, %k2
3501 ; AVX512DQ-NEXT: kshiftlw $10, %k2, %k2
3502 ; AVX512DQ-NEXT: kmovw {{[-0-9]+}}(%r{{[sb]}}p), %k1 ## 2-byte Reload
3503 ; AVX512DQ-NEXT: korw %k2, %k1, %k2
3504 ; AVX512DQ-NEXT: korw %k2, %k0, %k0
3505 ; AVX512DQ-NEXT: kshiftlw $5, %k0, %k0
3506 ; AVX512DQ-NEXT: kshiftrw $5, %k0, %k0
3507 ; AVX512DQ-NEXT: movb {{[0-9]+}}(%rsp), %al
3508 ; AVX512DQ-NEXT: kmovw %eax, %k2
3509 ; AVX512DQ-NEXT: kshiftlw $11, %k2, %k2
3510 ; AVX512DQ-NEXT: kmovw {{[-0-9]+}}(%r{{[sb]}}p), %k1 ## 2-byte Reload
3511 ; AVX512DQ-NEXT: korw %k2, %k1, %k2
3512 ; AVX512DQ-NEXT: korw %k2, %k0, %k0
3513 ; AVX512DQ-NEXT: kshiftlw $4, %k0, %k0
3514 ; AVX512DQ-NEXT: kshiftrw $4, %k0, %k0
3515 ; AVX512DQ-NEXT: movb {{[0-9]+}}(%rsp), %al
3516 ; AVX512DQ-NEXT: kmovw %eax, %k2
3517 ; AVX512DQ-NEXT: kshiftlw $12, %k2, %k2
3518 ; AVX512DQ-NEXT: kmovw {{[-0-9]+}}(%r{{[sb]}}p), %k1 ## 2-byte Reload
3519 ; AVX512DQ-NEXT: korw %k2, %k1, %k2
3520 ; AVX512DQ-NEXT: korw %k2, %k0, %k0
3521 ; AVX512DQ-NEXT: kshiftlw $3, %k0, %k0
3522 ; AVX512DQ-NEXT: kshiftrw $3, %k0, %k0
3523 ; AVX512DQ-NEXT: movb {{[0-9]+}}(%rsp), %al
3524 ; AVX512DQ-NEXT: kmovw %eax, %k2
3525 ; AVX512DQ-NEXT: kshiftlw $13, %k2, %k2
3526 ; AVX512DQ-NEXT: kmovw {{[-0-9]+}}(%r{{[sb]}}p), %k1 ## 2-byte Reload
3527 ; AVX512DQ-NEXT: korw %k2, %k1, %k2
3528 ; AVX512DQ-NEXT: korw %k2, %k0, %k0
3529 ; AVX512DQ-NEXT: kshiftlw $2, %k0, %k0
3530 ; AVX512DQ-NEXT: kshiftrw $2, %k0, %k0
3531 ; AVX512DQ-NEXT: movb {{[0-9]+}}(%rsp), %al
3532 ; AVX512DQ-NEXT: kmovw %eax, %k2
3533 ; AVX512DQ-NEXT: kshiftlw $14, %k2, %k2
3534 ; AVX512DQ-NEXT: kmovw {{[-0-9]+}}(%r{{[sb]}}p), %k1 ## 2-byte Reload
3535 ; AVX512DQ-NEXT: korw %k2, %k1, %k2
3536 ; AVX512DQ-NEXT: korw %k2, %k0, %k0
3537 ; AVX512DQ-NEXT: kshiftlw $1, %k0, %k0
3538 ; AVX512DQ-NEXT: kshiftrw $1, %k0, %k0
3539 ; AVX512DQ-NEXT: movb {{[0-9]+}}(%rsp), %al
3540 ; AVX512DQ-NEXT: kmovw %eax, %k2
3541 ; AVX512DQ-NEXT: kshiftlw $15, %k2, %k2
3542 ; AVX512DQ-NEXT: korw %k2, %k0, %k0
3543 ; AVX512DQ-NEXT: kmovw %k0, {{[-0-9]+}}(%r{{[sb]}}p) ## 2-byte Spill
3544 ; AVX512DQ-NEXT: movb {{[0-9]+}}(%rsp), %al
3545 ; AVX512DQ-NEXT: kmovw %eax, %k0
3546 ; AVX512DQ-NEXT: kshiftlw $15, %k0, %k0
3547 ; AVX512DQ-NEXT: kshiftrw $15, %k0, %k0
3548 ; AVX512DQ-NEXT: movb {{[0-9]+}}(%rsp), %al
3549 ; AVX512DQ-NEXT: kmovw %eax, %k2
3550 ; AVX512DQ-NEXT: kshiftlw $1, %k2, %k2
3551 ; AVX512DQ-NEXT: kmovw {{[-0-9]+}}(%r{{[sb]}}p), %k1 ## 2-byte Reload
3552 ; AVX512DQ-NEXT: korw %k2, %k1, %k2
3553 ; AVX512DQ-NEXT: korw %k2, %k0, %k0
3554 ; AVX512DQ-NEXT: kshiftlw $14, %k0, %k0
3555 ; AVX512DQ-NEXT: kshiftrw $14, %k0, %k0
3556 ; AVX512DQ-NEXT: movb {{[0-9]+}}(%rsp), %al
3557 ; AVX512DQ-NEXT: kmovw %eax, %k2
3558 ; AVX512DQ-NEXT: kshiftlw $2, %k2, %k2
3559 ; AVX512DQ-NEXT: korw %k2, %k3, %k2
3560 ; AVX512DQ-NEXT: korw %k2, %k0, %k0
3561 ; AVX512DQ-NEXT: kshiftlw $13, %k0, %k0
3562 ; AVX512DQ-NEXT: kshiftrw $13, %k0, %k0
3563 ; AVX512DQ-NEXT: movb {{[0-9]+}}(%rsp), %al
3564 ; AVX512DQ-NEXT: kmovw %eax, %k2
3565 ; AVX512DQ-NEXT: kshiftlw $3, %k2, %k2
3566 ; AVX512DQ-NEXT: korw %k2, %k4, %k2
3567 ; AVX512DQ-NEXT: korw %k2, %k0, %k0
3568 ; AVX512DQ-NEXT: kshiftlw $12, %k0, %k0
3569 ; AVX512DQ-NEXT: kshiftrw $12, %k0, %k0
3570 ; AVX512DQ-NEXT: movb {{[0-9]+}}(%rsp), %al
3571 ; AVX512DQ-NEXT: kmovw %eax, %k2
3572 ; AVX512DQ-NEXT: kshiftlw $4, %k2, %k2
3573 ; AVX512DQ-NEXT: korw %k2, %k5, %k2
3574 ; AVX512DQ-NEXT: korw %k2, %k0, %k0
3575 ; AVX512DQ-NEXT: kshiftlw $11, %k0, %k0
3576 ; AVX512DQ-NEXT: kshiftrw $11, %k0, %k0
3577 ; AVX512DQ-NEXT: movb {{[0-9]+}}(%rsp), %al
3578 ; AVX512DQ-NEXT: kmovw %eax, %k2
3579 ; AVX512DQ-NEXT: kshiftlw $5, %k2, %k2
3580 ; AVX512DQ-NEXT: korw %k2, %k6, %k2
3581 ; AVX512DQ-NEXT: korw %k2, %k0, %k0
3582 ; AVX512DQ-NEXT: kshiftlw $10, %k0, %k0
3583 ; AVX512DQ-NEXT: kshiftrw $10, %k0, %k0
3584 ; AVX512DQ-NEXT: movb {{[0-9]+}}(%rsp), %al
3585 ; AVX512DQ-NEXT: kmovw %eax, %k2
3586 ; AVX512DQ-NEXT: kshiftlw $6, %k2, %k2
3587 ; AVX512DQ-NEXT: korw %k2, %k7, %k2
3588 ; AVX512DQ-NEXT: korw %k2, %k0, %k0
3589 ; AVX512DQ-NEXT: kshiftlw $9, %k0, %k0
3590 ; AVX512DQ-NEXT: kshiftrw $9, %k0, %k0
3591 ; AVX512DQ-NEXT: movb {{[0-9]+}}(%rsp), %al
3592 ; AVX512DQ-NEXT: kmovw %eax, %k2
3593 ; AVX512DQ-NEXT: kshiftlw $7, %k2, %k2
3594 ; AVX512DQ-NEXT: kmovw {{[-0-9]+}}(%r{{[sb]}}p), %k1 ## 2-byte Reload
3595 ; AVX512DQ-NEXT: korw %k2, %k1, %k2
3596 ; AVX512DQ-NEXT: korw %k2, %k0, %k0
3597 ; AVX512DQ-NEXT: kshiftlw $8, %k0, %k0
3598 ; AVX512DQ-NEXT: kshiftrw $8, %k0, %k0
3599 ; AVX512DQ-NEXT: movb {{[0-9]+}}(%rsp), %al
3600 ; AVX512DQ-NEXT: kmovw %eax, %k2
3601 ; AVX512DQ-NEXT: kshiftlw $8, %k2, %k2
3602 ; AVX512DQ-NEXT: kmovw {{[-0-9]+}}(%r{{[sb]}}p), %k3 ## 2-byte Reload
3603 ; AVX512DQ-NEXT: korw %k2, %k3, %k2
3604 ; AVX512DQ-NEXT: korw %k2, %k0, %k0
3605 ; AVX512DQ-NEXT: kshiftlw $7, %k0, %k0
3606 ; AVX512DQ-NEXT: kshiftrw $7, %k0, %k0
3607 ; AVX512DQ-NEXT: movb {{[0-9]+}}(%rsp), %al
3608 ; AVX512DQ-NEXT: kmovw %eax, %k2
3609 ; AVX512DQ-NEXT: kshiftlw $9, %k2, %k2
3610 ; AVX512DQ-NEXT: kmovw {{[-0-9]+}}(%r{{[sb]}}p), %k4 ## 2-byte Reload
3611 ; AVX512DQ-NEXT: korw %k2, %k4, %k2
3612 ; AVX512DQ-NEXT: korw %k2, %k0, %k0
3613 ; AVX512DQ-NEXT: kshiftlw $6, %k0, %k0
3614 ; AVX512DQ-NEXT: kshiftrw $6, %k0, %k0
3615 ; AVX512DQ-NEXT: movb {{[0-9]+}}(%rsp), %al
3616 ; AVX512DQ-NEXT: kmovw %eax, %k2
3617 ; AVX512DQ-NEXT: kshiftlw $10, %k2, %k2
3618 ; AVX512DQ-NEXT: kmovw {{[-0-9]+}}(%r{{[sb]}}p), %k5 ## 2-byte Reload
3619 ; AVX512DQ-NEXT: korw %k2, %k5, %k2
3620 ; AVX512DQ-NEXT: korw %k2, %k0, %k0
3621 ; AVX512DQ-NEXT: kshiftlw $5, %k0, %k0
3622 ; AVX512DQ-NEXT: kshiftrw $5, %k0, %k0
3623 ; AVX512DQ-NEXT: movb {{[0-9]+}}(%rsp), %al
3624 ; AVX512DQ-NEXT: kmovw %eax, %k2
3625 ; AVX512DQ-NEXT: kshiftlw $11, %k2, %k2
3626 ; AVX512DQ-NEXT: kmovw {{[-0-9]+}}(%r{{[sb]}}p), %k7 ## 2-byte Reload
3627 ; AVX512DQ-NEXT: korw %k2, %k7, %k2
3628 ; AVX512DQ-NEXT: korw %k2, %k0, %k0
3629 ; AVX512DQ-NEXT: kshiftlw $4, %k0, %k0
3630 ; AVX512DQ-NEXT: kshiftrw $4, %k0, %k0
3631 ; AVX512DQ-NEXT: movb {{[0-9]+}}(%rsp), %al
3632 ; AVX512DQ-NEXT: kmovw %eax, %k2
3633 ; AVX512DQ-NEXT: kshiftlw $12, %k2, %k2
3634 ; AVX512DQ-NEXT: kmovw {{[-0-9]+}}(%r{{[sb]}}p), %k7 ## 2-byte Reload
3635 ; AVX512DQ-NEXT: korw %k2, %k7, %k2
3636 ; AVX512DQ-NEXT: korw %k2, %k0, %k0
3637 ; AVX512DQ-NEXT: kshiftlw $3, %k0, %k0
3638 ; AVX512DQ-NEXT: kshiftrw $3, %k0, %k0
3639 ; AVX512DQ-NEXT: movb {{[0-9]+}}(%rsp), %al
3640 ; AVX512DQ-NEXT: kmovw %eax, %k2
3641 ; AVX512DQ-NEXT: kshiftlw $13, %k2, %k2
3642 ; AVX512DQ-NEXT: kmovw {{[-0-9]+}}(%r{{[sb]}}p), %k7 ## 2-byte Reload
3643 ; AVX512DQ-NEXT: korw %k2, %k7, %k2
3644 ; AVX512DQ-NEXT: korw %k2, %k0, %k0
3645 ; AVX512DQ-NEXT: kshiftlw $2, %k0, %k0
3646 ; AVX512DQ-NEXT: kshiftrw $2, %k0, %k0
3647 ; AVX512DQ-NEXT: movb {{[0-9]+}}(%rsp), %al
3648 ; AVX512DQ-NEXT: kmovw %eax, %k2
3649 ; AVX512DQ-NEXT: kshiftlw $14, %k2, %k2
3650 ; AVX512DQ-NEXT: kmovw {{[-0-9]+}}(%r{{[sb]}}p), %k7 ## 2-byte Reload
3651 ; AVX512DQ-NEXT: korw %k2, %k7, %k2
3652 ; AVX512DQ-NEXT: korw %k2, %k0, %k0
3653 ; AVX512DQ-NEXT: kshiftlw $1, %k0, %k0
3654 ; AVX512DQ-NEXT: kshiftrw $1, %k0, %k0
3655 ; AVX512DQ-NEXT: movb {{[0-9]+}}(%rsp), %al
3656 ; AVX512DQ-NEXT: kmovw %eax, %k2
3657 ; AVX512DQ-NEXT: kshiftlw $15, %k2, %k2
3658 ; AVX512DQ-NEXT: korw %k2, %k0, %k0
3659 ; AVX512DQ-NEXT: kmovw %k0, {{[-0-9]+}}(%r{{[sb]}}p) ## 2-byte Spill
3660 ; AVX512DQ-NEXT: movb {{[0-9]+}}(%rsp), %al
3661 ; AVX512DQ-NEXT: kmovw %eax, %k2
3662 ; AVX512DQ-NEXT: kshiftlw $1, %k2, %k2
3663 ; AVX512DQ-NEXT: kmovw {{[-0-9]+}}(%r{{[sb]}}p), %k0 ## 2-byte Reload
3664 ; AVX512DQ-NEXT: korw %k2, %k0, %k2
3665 ; AVX512DQ-NEXT: movb {{[0-9]+}}(%rsp), %al
3666 ; AVX512DQ-NEXT: kmovw %eax, %k7
3667 ; AVX512DQ-NEXT: kshiftlw $15, %k7, %k7
3668 ; AVX512DQ-NEXT: kshiftrw $15, %k7, %k7
3669 ; AVX512DQ-NEXT: korw %k2, %k7, %k2
3670 ; AVX512DQ-NEXT: movb {{[0-9]+}}(%rsp), %al
3671 ; AVX512DQ-NEXT: kmovw %eax, %k7
3672 ; AVX512DQ-NEXT: kshiftlw $2, %k7, %k7
3673 ; AVX512DQ-NEXT: kmovw {{[-0-9]+}}(%r{{[sb]}}p), %k0 ## 2-byte Reload
3674 ; AVX512DQ-NEXT: korw %k7, %k0, %k7
3675 ; AVX512DQ-NEXT: kshiftlw $14, %k2, %k2
3676 ; AVX512DQ-NEXT: kshiftrw $14, %k2, %k2
3677 ; AVX512DQ-NEXT: korw %k7, %k2, %k2
3678 ; AVX512DQ-NEXT: movb {{[0-9]+}}(%rsp), %al
3679 ; AVX512DQ-NEXT: kmovw %eax, %k7
3680 ; AVX512DQ-NEXT: kshiftlw $3, %k7, %k7
3681 ; AVX512DQ-NEXT: kmovw {{[-0-9]+}}(%r{{[sb]}}p), %k0 ## 2-byte Reload
3682 ; AVX512DQ-NEXT: korw %k7, %k0, %k7
3683 ; AVX512DQ-NEXT: kshiftlw $13, %k2, %k2
3684 ; AVX512DQ-NEXT: kshiftrw $13, %k2, %k2
3685 ; AVX512DQ-NEXT: korw %k7, %k2, %k2
3686 ; AVX512DQ-NEXT: movb {{[0-9]+}}(%rsp), %al
3687 ; AVX512DQ-NEXT: kmovw %eax, %k7
3688 ; AVX512DQ-NEXT: kshiftlw $4, %k7, %k7
3689 ; AVX512DQ-NEXT: kmovw {{[-0-9]+}}(%r{{[sb]}}p), %k0 ## 2-byte Reload
3690 ; AVX512DQ-NEXT: korw %k7, %k0, %k7
3691 ; AVX512DQ-NEXT: kshiftlw $12, %k2, %k2
3692 ; AVX512DQ-NEXT: kshiftrw $12, %k2, %k2
3693 ; AVX512DQ-NEXT: korw %k7, %k2, %k2
3694 ; AVX512DQ-NEXT: movb {{[0-9]+}}(%rsp), %al
3695 ; AVX512DQ-NEXT: kmovw %eax, %k7
3696 ; AVX512DQ-NEXT: kshiftlw $5, %k7, %k7
3697 ; AVX512DQ-NEXT: korw %k7, %k6, %k7
3698 ; AVX512DQ-NEXT: kshiftlw $11, %k2, %k2
3699 ; AVX512DQ-NEXT: kshiftrw $11, %k2, %k2
3700 ; AVX512DQ-NEXT: korw %k7, %k2, %k2
3701 ; AVX512DQ-NEXT: movb {{[0-9]+}}(%rsp), %al
3702 ; AVX512DQ-NEXT: kmovw %eax, %k7
3703 ; AVX512DQ-NEXT: kshiftlw $6, %k7, %k7
3704 ; AVX512DQ-NEXT: kmovw {{[-0-9]+}}(%r{{[sb]}}p), %k0 ## 2-byte Reload
3705 ; AVX512DQ-NEXT: korw %k7, %k0, %k7
3706 ; AVX512DQ-NEXT: kshiftlw $10, %k2, %k2
3707 ; AVX512DQ-NEXT: kshiftrw $10, %k2, %k2
3708 ; AVX512DQ-NEXT: korw %k7, %k2, %k2
3709 ; AVX512DQ-NEXT: movb {{[0-9]+}}(%rsp), %al
3710 ; AVX512DQ-NEXT: kmovw %eax, %k7
3711 ; AVX512DQ-NEXT: kshiftlw $7, %k7, %k7
3712 ; AVX512DQ-NEXT: korw %k7, %k1, %k7
3713 ; AVX512DQ-NEXT: kshiftlw $9, %k2, %k2
3714 ; AVX512DQ-NEXT: kshiftrw $9, %k2, %k2
3715 ; AVX512DQ-NEXT: korw %k7, %k2, %k2
3716 ; AVX512DQ-NEXT: movb {{[0-9]+}}(%rsp), %al
3717 ; AVX512DQ-NEXT: kmovw %eax, %k7
3718 ; AVX512DQ-NEXT: kshiftlw $8, %k7, %k7
3719 ; AVX512DQ-NEXT: korw %k7, %k3, %k7
3720 ; AVX512DQ-NEXT: kshiftlw $8, %k2, %k2
3721 ; AVX512DQ-NEXT: kshiftrw $8, %k2, %k2
3722 ; AVX512DQ-NEXT: korw %k7, %k2, %k2
3723 ; AVX512DQ-NEXT: movb {{[0-9]+}}(%rsp), %al
3724 ; AVX512DQ-NEXT: kmovw %eax, %k7
3725 ; AVX512DQ-NEXT: kshiftlw $9, %k7, %k7
3726 ; AVX512DQ-NEXT: korw %k7, %k4, %k7
3727 ; AVX512DQ-NEXT: kshiftlw $7, %k2, %k2
3728 ; AVX512DQ-NEXT: kshiftrw $7, %k2, %k2
3729 ; AVX512DQ-NEXT: korw %k7, %k2, %k2
3730 ; AVX512DQ-NEXT: movb {{[0-9]+}}(%rsp), %al
3731 ; AVX512DQ-NEXT: kmovw %eax, %k7
3732 ; AVX512DQ-NEXT: kshiftlw $10, %k7, %k7
3733 ; AVX512DQ-NEXT: korw %k7, %k5, %k6
3734 ; AVX512DQ-NEXT: kshiftlw $6, %k2, %k2
3735 ; AVX512DQ-NEXT: kshiftrw $6, %k2, %k2
3736 ; AVX512DQ-NEXT: korw %k6, %k2, %k2
3737 ; AVX512DQ-NEXT: movb {{[0-9]+}}(%rsp), %al
3738 ; AVX512DQ-NEXT: kmovw %eax, %k6
3739 ; AVX512DQ-NEXT: kshiftlw $11, %k6, %k6
3740 ; AVX512DQ-NEXT: kmovw {{[-0-9]+}}(%r{{[sb]}}p), %k1 ## 2-byte Reload
3741 ; AVX512DQ-NEXT: korw %k6, %k1, %k5
3742 ; AVX512DQ-NEXT: kshiftlw $5, %k2, %k2
3743 ; AVX512DQ-NEXT: kshiftrw $5, %k2, %k2
3744 ; AVX512DQ-NEXT: korw %k5, %k2, %k2
3745 ; AVX512DQ-NEXT: movb {{[0-9]+}}(%rsp), %al
3746 ; AVX512DQ-NEXT: kmovw %eax, %k5
3747 ; AVX512DQ-NEXT: kshiftlw $12, %k5, %k5
3748 ; AVX512DQ-NEXT: kmovw {{[-0-9]+}}(%r{{[sb]}}p), %k1 ## 2-byte Reload
3749 ; AVX512DQ-NEXT: korw %k5, %k1, %k4
3750 ; AVX512DQ-NEXT: kshiftlw $4, %k2, %k2
3751 ; AVX512DQ-NEXT: kshiftrw $4, %k2, %k2
3752 ; AVX512DQ-NEXT: korw %k4, %k2, %k2
3753 ; AVX512DQ-NEXT: movb {{[0-9]+}}(%rsp), %al
3754 ; AVX512DQ-NEXT: kmovw %eax, %k4
3755 ; AVX512DQ-NEXT: kshiftlw $13, %k4, %k4
3756 ; AVX512DQ-NEXT: kmovw {{[-0-9]+}}(%r{{[sb]}}p), %k1 ## 2-byte Reload
3757 ; AVX512DQ-NEXT: korw %k4, %k1, %k3
3758 ; AVX512DQ-NEXT: kshiftlw $3, %k2, %k2
3759 ; AVX512DQ-NEXT: kshiftrw $3, %k2, %k2
3760 ; AVX512DQ-NEXT: korw %k3, %k2, %k2
3761 ; AVX512DQ-NEXT: movb {{[0-9]+}}(%rsp), %al
3762 ; AVX512DQ-NEXT: kmovw %eax, %k3
3763 ; AVX512DQ-NEXT: kshiftlw $14, %k3, %k3
3764 ; AVX512DQ-NEXT: kmovw {{[-0-9]+}}(%r{{[sb]}}p), %k1 ## 2-byte Reload
3765 ; AVX512DQ-NEXT: korw %k3, %k1, %k1
3766 ; AVX512DQ-NEXT: kshiftlw $2, %k2, %k2
3767 ; AVX512DQ-NEXT: kshiftrw $2, %k2, %k2
3768 ; AVX512DQ-NEXT: korw %k1, %k2, %k1
3769 ; AVX512DQ-NEXT: kshiftlw $1, %k1, %k1
3770 ; AVX512DQ-NEXT: kshiftrw $1, %k1, %k1
3771 ; AVX512DQ-NEXT: movb {{[0-9]+}}(%rsp), %al
3772 ; AVX512DQ-NEXT: kmovw %eax, %k2
3773 ; AVX512DQ-NEXT: kshiftlw $15, %k2, %k2
3774 ; AVX512DQ-NEXT: korw %k2, %k1, %k1
3775 ; AVX512DQ-NEXT: kmovw %k1, 6(%rdi)
3776 ; AVX512DQ-NEXT: kmovw {{[-0-9]+}}(%r{{[sb]}}p), %k0 ## 2-byte Reload
3777 ; AVX512DQ-NEXT: kmovw %k0, 4(%rdi)
3778 ; AVX512DQ-NEXT: kmovw {{[-0-9]+}}(%r{{[sb]}}p), %k0 ## 2-byte Reload
3779 ; AVX512DQ-NEXT: kmovw %k0, 2(%rdi)
3780 ; AVX512DQ-NEXT: kmovw {{[-0-9]+}}(%r{{[sb]}}p), %k0 ## 2-byte Reload
3781 ; AVX512DQ-NEXT: kmovw %k0, (%rdi)
3782 ; AVX512DQ-NEXT: retq
3783 ;
3784 ; X86-LABEL: store_64i1:
3785 ; X86: ## %bb.0:
3786 ; X86-NEXT: vpsllw $7, %zmm0, %zmm0
3787 ; X86-NEXT: vpmovb2m %zmm0, %k0
3788 ; X86-NEXT: movl {{[0-9]+}}(%esp), %eax
3789 ; X86-NEXT: kmovq %k0, (%eax)
3790 ; X86-NEXT: vzeroupper
3791 ; X86-NEXT: retl
3792 store <64 x i1> %v, <64 x i1>* %a
3793 ret void
3794 }
3796 define i32 @test_bitcast_v8i1_zext(<16 x i32> %a) {
3797 ; KNL-LABEL: test_bitcast_v8i1_zext:
3798 ; KNL: ## %bb.0:
3799 ; KNL-NEXT: vptestnmd %zmm0, %zmm0, %k0
3800 ; KNL-NEXT: kmovw %k0, %eax
3801 ; KNL-NEXT: movzbl %al, %eax
3802 ; KNL-NEXT: addl %eax, %eax
3803 ; KNL-NEXT: vzeroupper
3804 ; KNL-NEXT: retq
3805 ;
3806 ; SKX-LABEL: test_bitcast_v8i1_zext:
3807 ; SKX: ## %bb.0:
3808 ; SKX-NEXT: vptestnmd %zmm0, %zmm0, %k0
3809 ; SKX-NEXT: kmovb %k0, %eax
3810 ; SKX-NEXT: addl %eax, %eax
3811 ; SKX-NEXT: vzeroupper
3812 ; SKX-NEXT: retq
3813 ;
3814 ; AVX512BW-LABEL: test_bitcast_v8i1_zext:
3815 ; AVX512BW: ## %bb.0:
3816 ; AVX512BW-NEXT: vptestnmd %zmm0, %zmm0, %k0
3817 ; AVX512BW-NEXT: kmovd %k0, %eax
3818 ; AVX512BW-NEXT: movzbl %al, %eax
3819 ; AVX512BW-NEXT: addl %eax, %eax
3820 ; AVX512BW-NEXT: vzeroupper
3821 ; AVX512BW-NEXT: retq
3822 ;
3823 ; AVX512DQ-LABEL: test_bitcast_v8i1_zext:
3824 ; AVX512DQ: ## %bb.0:
3825 ; AVX512DQ-NEXT: vptestnmd %zmm0, %zmm0, %k0
3826 ; AVX512DQ-NEXT: kmovb %k0, %eax
3827 ; AVX512DQ-NEXT: addl %eax, %eax
3828 ; AVX512DQ-NEXT: vzeroupper
3829 ; AVX512DQ-NEXT: retq
3830 ;
3831 ; X86-LABEL: test_bitcast_v8i1_zext:
3832 ; X86: ## %bb.0:
3833 ; X86-NEXT: vptestnmd %zmm0, %zmm0, %k0
3834 ; X86-NEXT: kmovb %k0, %eax
3835 ; X86-NEXT: addl %eax, %eax
3836 ; X86-NEXT: vzeroupper
3837 ; X86-NEXT: retl
3838 %v1 = icmp eq <16 x i32> %a, zeroinitializer
3839 %mask = shufflevector <16 x i1> %v1, <16 x i1> undef, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
3840 %mask1 = bitcast <8 x i1> %mask to i8
3841 %val = zext i8 %mask1 to i32
3842 %val1 = add i32 %val, %val
3843 ret i32 %val1
3844 }
3846 define i32 @test_bitcast_v16i1_zext(<16 x i32> %a) {
3847 ; CHECK-LABEL: test_bitcast_v16i1_zext:
3848 ; CHECK: ## %bb.0:
3849 ; CHECK-NEXT: vptestnmd %zmm0, %zmm0, %k0
3850 ; CHECK-NEXT: kmovw %k0, %eax
3851 ; CHECK-NEXT: addl %eax, %eax
3852 ; CHECK-NEXT: vzeroupper
3853 ; CHECK-NEXT: retq
3854 ;
3855 ; X86-LABEL: test_bitcast_v16i1_zext:
3856 ; X86: ## %bb.0:
3857 ; X86-NEXT: vptestnmd %zmm0, %zmm0, %k0
3858 ; X86-NEXT: kmovw %k0, %eax
3859 ; X86-NEXT: addl %eax, %eax
3860 ; X86-NEXT: vzeroupper
3861 ; X86-NEXT: retl
3862 %v1 = icmp eq <16 x i32> %a, zeroinitializer
3863 %mask1 = bitcast <16 x i1> %v1 to i16
3864 %val = zext i16 %mask1 to i32
3865 %val1 = add i32 %val, %val
3866 ret i32 %val1
3867 }
3869 define i16 @test_v16i1_add(i16 %x, i16 %y) {
3870 ; KNL-LABEL: test_v16i1_add:
3871 ; KNL: ## %bb.0:
3872 ; KNL-NEXT: kmovw %edi, %k0
3873 ; KNL-NEXT: kmovw %esi, %k1
3874 ; KNL-NEXT: kxorw %k1, %k0, %k0
3875 ; KNL-NEXT: kmovw %k0, %eax
3876 ; KNL-NEXT: ## kill: def $ax killed $ax killed $eax
3877 ; KNL-NEXT: retq
3878 ;
3879 ; SKX-LABEL: test_v16i1_add:
3880 ; SKX: ## %bb.0:
3881 ; SKX-NEXT: kmovd %edi, %k0
3882 ; SKX-NEXT: kmovd %esi, %k1
3883 ; SKX-NEXT: kxorw %k1, %k0, %k0
3884 ; SKX-NEXT: kmovd %k0, %eax
3885 ; SKX-NEXT: ## kill: def $ax killed $ax killed $eax
3886 ; SKX-NEXT: retq
3887 ;
3888 ; AVX512BW-LABEL: test_v16i1_add:
3889 ; AVX512BW: ## %bb.0:
3890 ; AVX512BW-NEXT: kmovd %edi, %k0
3891 ; AVX512BW-NEXT: kmovd %esi, %k1
3892 ; AVX512BW-NEXT: kxorw %k1, %k0, %k0
3893 ; AVX512BW-NEXT: kmovd %k0, %eax
3894 ; AVX512BW-NEXT: ## kill: def $ax killed $ax killed $eax
3895 ; AVX512BW-NEXT: retq
3896 ;
3897 ; AVX512DQ-LABEL: test_v16i1_add:
3898 ; AVX512DQ: ## %bb.0:
3899 ; AVX512DQ-NEXT: kmovw %edi, %k0
3900 ; AVX512DQ-NEXT: kmovw %esi, %k1
3901 ; AVX512DQ-NEXT: kxorw %k1, %k0, %k0
3902 ; AVX512DQ-NEXT: kmovw %k0, %eax
3903 ; AVX512DQ-NEXT: ## kill: def $ax killed $ax killed $eax
3904 ; AVX512DQ-NEXT: retq
3905 ;
3906 ; X86-LABEL: test_v16i1_add:
3907 ; X86: ## %bb.0:
3908 ; X86-NEXT: kmovw {{[0-9]+}}(%esp), %k0
3909 ; X86-NEXT: kmovw {{[0-9]+}}(%esp), %k1
3910 ; X86-NEXT: kxorw %k1, %k0, %k0
3911 ; X86-NEXT: kmovd %k0, %eax
3912 ; X86-NEXT: ## kill: def $ax killed $ax killed $eax
3913 ; X86-NEXT: retl
3914 %m0 = bitcast i16 %x to <16 x i1>
3915 %m1 = bitcast i16 %y to <16 x i1>
3916 %m2 = add <16 x i1> %m0, %m1
3917 %ret = bitcast <16 x i1> %m2 to i16
3918 ret i16 %ret
3919 }
3921 define i16 @test_v16i1_sub(i16 %x, i16 %y) {
3922 ; KNL-LABEL: test_v16i1_sub:
3923 ; KNL: ## %bb.0:
3924 ; KNL-NEXT: kmovw %edi, %k0
3925 ; KNL-NEXT: kmovw %esi, %k1
3926 ; KNL-NEXT: kxorw %k1, %k0, %k0
3927 ; KNL-NEXT: kmovw %k0, %eax
3928 ; KNL-NEXT: ## kill: def $ax killed $ax killed $eax
3929 ; KNL-NEXT: retq
3930 ;
3931 ; SKX-LABEL: test_v16i1_sub:
3932 ; SKX: ## %bb.0:
3933 ; SKX-NEXT: kmovd %edi, %k0
3934 ; SKX-NEXT: kmovd %esi, %k1
3935 ; SKX-NEXT: kxorw %k1, %k0, %k0
3936 ; SKX-NEXT: kmovd %k0, %eax
3937 ; SKX-NEXT: ## kill: def $ax killed $ax killed $eax
3938 ; SKX-NEXT: retq
3939 ;
3940 ; AVX512BW-LABEL: test_v16i1_sub:
3941 ; AVX512BW: ## %bb.0:
3942 ; AVX512BW-NEXT: kmovd %edi, %k0
3943 ; AVX512BW-NEXT: kmovd %esi, %k1
3944 ; AVX512BW-NEXT: kxorw %k1, %k0, %k0
3945 ; AVX512BW-NEXT: kmovd %k0, %eax
3946 ; AVX512BW-NEXT: ## kill: def $ax killed $ax killed $eax
3947 ; AVX512BW-NEXT: retq
3948 ;
3949 ; AVX512DQ-LABEL: test_v16i1_sub:
3950 ; AVX512DQ: ## %bb.0:
3951 ; AVX512DQ-NEXT: kmovw %edi, %k0
3952 ; AVX512DQ-NEXT: kmovw %esi, %k1
3953 ; AVX512DQ-NEXT: kxorw %k1, %k0, %k0
3954 ; AVX512DQ-NEXT: kmovw %k0, %eax
3955 ; AVX512DQ-NEXT: ## kill: def $ax killed $ax killed $eax
3956 ; AVX512DQ-NEXT: retq
3957 ;
3958 ; X86-LABEL: test_v16i1_sub:
3959 ; X86: ## %bb.0:
3960 ; X86-NEXT: kmovw {{[0-9]+}}(%esp), %k0
3961 ; X86-NEXT: kmovw {{[0-9]+}}(%esp), %k1
3962 ; X86-NEXT: kxorw %k1, %k0, %k0
3963 ; X86-NEXT: kmovd %k0, %eax
3964 ; X86-NEXT: ## kill: def $ax killed $ax killed $eax
3965 ; X86-NEXT: retl
3966 %m0 = bitcast i16 %x to <16 x i1>
3967 %m1 = bitcast i16 %y to <16 x i1>
3968 %m2 = sub <16 x i1> %m0, %m1
3969 %ret = bitcast <16 x i1> %m2 to i16
3970 ret i16 %ret
3971 }
3973 define i16 @test_v16i1_mul(i16 %x, i16 %y) {
3974 ; KNL-LABEL: test_v16i1_mul:
3975 ; KNL: ## %bb.0:
3976 ; KNL-NEXT: kmovw %edi, %k0
3977 ; KNL-NEXT: kmovw %esi, %k1
3978 ; KNL-NEXT: kandw %k1, %k0, %k0
3979 ; KNL-NEXT: kmovw %k0, %eax
3980 ; KNL-NEXT: ## kill: def $ax killed $ax killed $eax
3981 ; KNL-NEXT: retq
3982 ;
3983 ; SKX-LABEL: test_v16i1_mul:
3984 ; SKX: ## %bb.0:
3985 ; SKX-NEXT: kmovd %edi, %k0
3986 ; SKX-NEXT: kmovd %esi, %k1
3987 ; SKX-NEXT: kandw %k1, %k0, %k0
3988 ; SKX-NEXT: kmovd %k0, %eax
3989 ; SKX-NEXT: ## kill: def $ax killed $ax killed $eax
3990 ; SKX-NEXT: retq
3991 ;
3992 ; AVX512BW-LABEL: test_v16i1_mul:
3993 ; AVX512BW: ## %bb.0:
3994 ; AVX512BW-NEXT: kmovd %edi, %k0
3995 ; AVX512BW-NEXT: kmovd %esi, %k1
3996 ; AVX512BW-NEXT: kandw %k1, %k0, %k0
3997 ; AVX512BW-NEXT: kmovd %k0, %eax
3998 ; AVX512BW-NEXT: ## kill: def $ax killed $ax killed $eax
3999 ; AVX512BW-NEXT: retq
4000 ;
4001 ; AVX512DQ-LABEL: test_v16i1_mul:
4002 ; AVX512DQ: ## %bb.0:
4003 ; AVX512DQ-NEXT: kmovw %edi, %k0
4004 ; AVX512DQ-NEXT: kmovw %esi, %k1
4005 ; AVX512DQ-NEXT: kandw %k1, %k0, %k0
4006 ; AVX512DQ-NEXT: kmovw %k0, %eax
4007 ; AVX512DQ-NEXT: ## kill: def $ax killed $ax killed $eax
4008 ; AVX512DQ-NEXT: retq
4009 ;
4010 ; X86-LABEL: test_v16i1_mul:
4011 ; X86: ## %bb.0:
4012 ; X86-NEXT: kmovw {{[0-9]+}}(%esp), %k0
4013 ; X86-NEXT: kmovw {{[0-9]+}}(%esp), %k1
4014 ; X86-NEXT: kandw %k1, %k0, %k0
4015 ; X86-NEXT: kmovd %k0, %eax
4016 ; X86-NEXT: ## kill: def $ax killed $ax killed $eax
4017 ; X86-NEXT: retl
4018 %m0 = bitcast i16 %x to <16 x i1>
4019 %m1 = bitcast i16 %y to <16 x i1>
4020 %m2 = mul <16 x i1> %m0, %m1
4021 %ret = bitcast <16 x i1> %m2 to i16
4022 ret i16 %ret
4023 }
4025 define i8 @test_v8i1_add(i8 %x, i8 %y) {
4026 ; KNL-LABEL: test_v8i1_add:
4027 ; KNL: ## %bb.0:
4028 ; KNL-NEXT: kmovw %edi, %k0
4029 ; KNL-NEXT: kmovw %esi, %k1
4030 ; KNL-NEXT: kxorw %k1, %k0, %k0
4031 ; KNL-NEXT: kmovw %k0, %eax
4032 ; KNL-NEXT: ## kill: def $al killed $al killed $eax
4033 ; KNL-NEXT: retq
4034 ;
4035 ; SKX-LABEL: test_v8i1_add:
4036 ; SKX: ## %bb.0:
4037 ; SKX-NEXT: kmovd %edi, %k0
4038 ; SKX-NEXT: kmovd %esi, %k1
4039 ; SKX-NEXT: kxorb %k1, %k0, %k0
4040 ; SKX-NEXT: kmovd %k0, %eax
4041 ; SKX-NEXT: ## kill: def $al killed $al killed $eax
4042 ; SKX-NEXT: retq
4043 ;
4044 ; AVX512BW-LABEL: test_v8i1_add:
4045 ; AVX512BW: ## %bb.0:
4046 ; AVX512BW-NEXT: kmovd %edi, %k0
4047 ; AVX512BW-NEXT: kmovd %esi, %k1
4048 ; AVX512BW-NEXT: kxorw %k1, %k0, %k0
4049 ; AVX512BW-NEXT: kmovd %k0, %eax
4050 ; AVX512BW-NEXT: ## kill: def $al killed $al killed $eax
4051 ; AVX512BW-NEXT: retq
4052 ;
4053 ; AVX512DQ-LABEL: test_v8i1_add:
4054 ; AVX512DQ: ## %bb.0:
4055 ; AVX512DQ-NEXT: kmovw %edi, %k0
4056 ; AVX512DQ-NEXT: kmovw %esi, %k1
4057 ; AVX512DQ-NEXT: kxorb %k1, %k0, %k0
4058 ; AVX512DQ-NEXT: kmovw %k0, %eax
4059 ; AVX512DQ-NEXT: ## kill: def $al killed $al killed $eax
4060 ; AVX512DQ-NEXT: retq
4061 ;
4062 ; X86-LABEL: test_v8i1_add:
4063 ; X86: ## %bb.0:
4064 ; X86-NEXT: kmovb {{[0-9]+}}(%esp), %k0
4065 ; X86-NEXT: kmovb {{[0-9]+}}(%esp), %k1
4066 ; X86-NEXT: kxorb %k1, %k0, %k0
4067 ; X86-NEXT: kmovd %k0, %eax
4068 ; X86-NEXT: ## kill: def $al killed $al killed $eax
4069 ; X86-NEXT: retl
4070 %m0 = bitcast i8 %x to <8 x i1>
4071 %m1 = bitcast i8 %y to <8 x i1>
4072 %m2 = add <8 x i1> %m0, %m1
4073 %ret = bitcast <8 x i1> %m2 to i8
4074 ret i8 %ret
4075 }
4077 define i8 @test_v8i1_sub(i8 %x, i8 %y) {
4078 ; KNL-LABEL: test_v8i1_sub:
4079 ; KNL: ## %bb.0:
4080 ; KNL-NEXT: kmovw %edi, %k0
4081 ; KNL-NEXT: kmovw %esi, %k1
4082 ; KNL-NEXT: kxorw %k1, %k0, %k0
4083 ; KNL-NEXT: kmovw %k0, %eax
4084 ; KNL-NEXT: ## kill: def $al killed $al killed $eax
4085 ; KNL-NEXT: retq
4086 ;
4087 ; SKX-LABEL: test_v8i1_sub:
4088 ; SKX: ## %bb.0:
4089 ; SKX-NEXT: kmovd %edi, %k0
4090 ; SKX-NEXT: kmovd %esi, %k1
4091 ; SKX-NEXT: kxorb %k1, %k0, %k0
4092 ; SKX-NEXT: kmovd %k0, %eax
4093 ; SKX-NEXT: ## kill: def $al killed $al killed $eax
4094 ; SKX-NEXT: retq
4095 ;
4096 ; AVX512BW-LABEL: test_v8i1_sub:
4097 ; AVX512BW: ## %bb.0:
4098 ; AVX512BW-NEXT: kmovd %edi, %k0
4099 ; AVX512BW-NEXT: kmovd %esi, %k1
4100 ; AVX512BW-NEXT: kxorw %k1, %k0, %k0
4101 ; AVX512BW-NEXT: kmovd %k0, %eax
4102 ; AVX512BW-NEXT: ## kill: def $al killed $al killed $eax
4103 ; AVX512BW-NEXT: retq
4104 ;
4105 ; AVX512DQ-LABEL: test_v8i1_sub:
4106 ; AVX512DQ: ## %bb.0:
4107 ; AVX512DQ-NEXT: kmovw %edi, %k0
4108 ; AVX512DQ-NEXT: kmovw %esi, %k1
4109 ; AVX512DQ-NEXT: kxorb %k1, %k0, %k0
4110 ; AVX512DQ-NEXT: kmovw %k0, %eax
4111 ; AVX512DQ-NEXT: ## kill: def $al killed $al killed $eax
4112 ; AVX512DQ-NEXT: retq
4113 ;
4114 ; X86-LABEL: test_v8i1_sub:
4115 ; X86: ## %bb.0:
4116 ; X86-NEXT: kmovb {{[0-9]+}}(%esp), %k0
4117 ; X86-NEXT: kmovb {{[0-9]+}}(%esp), %k1
4118 ; X86-NEXT: kxorb %k1, %k0, %k0
4119 ; X86-NEXT: kmovd %k0, %eax
4120 ; X86-NEXT: ## kill: def $al killed $al killed $eax
4121 ; X86-NEXT: retl
4122 %m0 = bitcast i8 %x to <8 x i1>
4123 %m1 = bitcast i8 %y to <8 x i1>
4124 %m2 = sub <8 x i1> %m0, %m1
4125 %ret = bitcast <8 x i1> %m2 to i8
4126 ret i8 %ret
4127 }
4129 define i8 @test_v8i1_mul(i8 %x, i8 %y) {
4130 ; KNL-LABEL: test_v8i1_mul:
4131 ; KNL: ## %bb.0:
4132 ; KNL-NEXT: kmovw %edi, %k0
4133 ; KNL-NEXT: kmovw %esi, %k1
4134 ; KNL-NEXT: kandw %k1, %k0, %k0
4135 ; KNL-NEXT: kmovw %k0, %eax
4136 ; KNL-NEXT: ## kill: def $al killed $al killed $eax
4137 ; KNL-NEXT: retq
4138 ;
4139 ; SKX-LABEL: test_v8i1_mul:
4140 ; SKX: ## %bb.0:
4141 ; SKX-NEXT: kmovd %edi, %k0
4142 ; SKX-NEXT: kmovd %esi, %k1
4143 ; SKX-NEXT: kandb %k1, %k0, %k0
4144 ; SKX-NEXT: kmovd %k0, %eax
4145 ; SKX-NEXT: ## kill: def $al killed $al killed $eax
4146 ; SKX-NEXT: retq
4147 ;
4148 ; AVX512BW-LABEL: test_v8i1_mul:
4149 ; AVX512BW: ## %bb.0:
4150 ; AVX512BW-NEXT: kmovd %edi, %k0
4151 ; AVX512BW-NEXT: kmovd %esi, %k1
4152 ; AVX512BW-NEXT: kandw %k1, %k0, %k0
4153 ; AVX512BW-NEXT: kmovd %k0, %eax
4154 ; AVX512BW-NEXT: ## kill: def $al killed $al killed $eax
4155 ; AVX512BW-NEXT: retq
4156 ;
4157 ; AVX512DQ-LABEL: test_v8i1_mul:
4158 ; AVX512DQ: ## %bb.0:
4159 ; AVX512DQ-NEXT: kmovw %edi, %k0
4160 ; AVX512DQ-NEXT: kmovw %esi, %k1
4161 ; AVX512DQ-NEXT: kandb %k1, %k0, %k0
4162 ; AVX512DQ-NEXT: kmovw %k0, %eax
4163 ; AVX512DQ-NEXT: ## kill: def $al killed $al killed $eax
4164 ; AVX512DQ-NEXT: retq
4165 ;
4166 ; X86-LABEL: test_v8i1_mul:
4167 ; X86: ## %bb.0:
4168 ; X86-NEXT: kmovb {{[0-9]+}}(%esp), %k0
4169 ; X86-NEXT: kmovb {{[0-9]+}}(%esp), %k1
4170 ; X86-NEXT: kandb %k1, %k0, %k0
4171 ; X86-NEXT: kmovd %k0, %eax
4172 ; X86-NEXT: ## kill: def $al killed $al killed $eax
4173 ; X86-NEXT: retl
4174 %m0 = bitcast i8 %x to <8 x i1>
4175 %m1 = bitcast i8 %y to <8 x i1>
4176 %m2 = mul <8 x i1> %m0, %m1
4177 %ret = bitcast <8 x i1> %m2 to i8
4178 ret i8 %ret
4179 }
4181 ; Make sure we don't emit a ktest for signed comparisons.
4182 define void @ktest_signed(<16 x i32> %x, <16 x i32> %y) {
4183 ; KNL-LABEL: ktest_signed:
4184 ; KNL: ## %bb.0:
4185 ; KNL-NEXT: pushq %rax
4186 ; KNL-NEXT: .cfi_def_cfa_offset 16
4187 ; KNL-NEXT: vpord %zmm1, %zmm0, %zmm0
4188 ; KNL-NEXT: vptestnmd %zmm0, %zmm0, %k0
4189 ; KNL-NEXT: kmovw %k0, %eax
4190 ; KNL-NEXT: testw %ax, %ax
4191 ; KNL-NEXT: jle LBB65_1
4192 ; KNL-NEXT: ## %bb.2: ## %bb.2
4193 ; KNL-NEXT: popq %rax
4194 ; KNL-NEXT: vzeroupper
4195 ; KNL-NEXT: retq
4196 ; KNL-NEXT: LBB65_1: ## %bb.1
4197 ; KNL-NEXT: vzeroupper
4198 ; KNL-NEXT: callq _foo
4199 ; KNL-NEXT: popq %rax
4200 ; KNL-NEXT: retq
4201 ;
4202 ; SKX-LABEL: ktest_signed:
4203 ; SKX: ## %bb.0:
4204 ; SKX-NEXT: pushq %rax
4205 ; SKX-NEXT: .cfi_def_cfa_offset 16
4206 ; SKX-NEXT: vpord %zmm1, %zmm0, %zmm0
4207 ; SKX-NEXT: vptestnmd %zmm0, %zmm0, %k0
4208 ; SKX-NEXT: kmovd %k0, %eax
4209 ; SKX-NEXT: testw %ax, %ax
4210 ; SKX-NEXT: jle LBB65_1
4211 ; SKX-NEXT: ## %bb.2: ## %bb.2
4212 ; SKX-NEXT: popq %rax
4213 ; SKX-NEXT: vzeroupper
4214 ; SKX-NEXT: retq
4215 ; SKX-NEXT: LBB65_1: ## %bb.1
4216 ; SKX-NEXT: vzeroupper
4217 ; SKX-NEXT: callq _foo
4218 ; SKX-NEXT: popq %rax
4219 ; SKX-NEXT: retq
4220 ;
4221 ; AVX512BW-LABEL: ktest_signed:
4222 ; AVX512BW: ## %bb.0:
4223 ; AVX512BW-NEXT: pushq %rax
4224 ; AVX512BW-NEXT: .cfi_def_cfa_offset 16
4225 ; AVX512BW-NEXT: vpord %zmm1, %zmm0, %zmm0
4226 ; AVX512BW-NEXT: vptestnmd %zmm0, %zmm0, %k0
4227 ; AVX512BW-NEXT: kmovd %k0, %eax
4228 ; AVX512BW-NEXT: testw %ax, %ax
4229 ; AVX512BW-NEXT: jle LBB65_1
4230 ; AVX512BW-NEXT: ## %bb.2: ## %bb.2
4231 ; AVX512BW-NEXT: popq %rax
4232 ; AVX512BW-NEXT: vzeroupper
4233 ; AVX512BW-NEXT: retq
4234 ; AVX512BW-NEXT: LBB65_1: ## %bb.1
4235 ; AVX512BW-NEXT: vzeroupper
4236 ; AVX512BW-NEXT: callq _foo
4237 ; AVX512BW-NEXT: popq %rax
4238 ; AVX512BW-NEXT: retq
4239 ;
4240 ; AVX512DQ-LABEL: ktest_signed:
4241 ; AVX512DQ: ## %bb.0:
4242 ; AVX512DQ-NEXT: pushq %rax
4243 ; AVX512DQ-NEXT: .cfi_def_cfa_offset 16
4244 ; AVX512DQ-NEXT: vpord %zmm1, %zmm0, %zmm0
4245 ; AVX512DQ-NEXT: vptestnmd %zmm0, %zmm0, %k0
4246 ; AVX512DQ-NEXT: kmovw %k0, %eax
4247 ; AVX512DQ-NEXT: testw %ax, %ax
4248 ; AVX512DQ-NEXT: jle LBB65_1
4249 ; AVX512DQ-NEXT: ## %bb.2: ## %bb.2
4250 ; AVX512DQ-NEXT: popq %rax
4251 ; AVX512DQ-NEXT: vzeroupper
4252 ; AVX512DQ-NEXT: retq
4253 ; AVX512DQ-NEXT: LBB65_1: ## %bb.1
4254 ; AVX512DQ-NEXT: vzeroupper
4255 ; AVX512DQ-NEXT: callq _foo
4256 ; AVX512DQ-NEXT: popq %rax
4257 ; AVX512DQ-NEXT: retq
4258 ;
4259 ; X86-LABEL: ktest_signed:
4260 ; X86: ## %bb.0:
4261 ; X86-NEXT: subl $12, %esp
4262 ; X86-NEXT: .cfi_def_cfa_offset 16
4263 ; X86-NEXT: vpord %zmm1, %zmm0, %zmm0
4264 ; X86-NEXT: vptestnmd %zmm0, %zmm0, %k0
4265 ; X86-NEXT: kmovd %k0, %eax
4266 ; X86-NEXT: testw %ax, %ax
4267 ; X86-NEXT: jle LBB65_1
4268 ; X86-NEXT: ## %bb.2: ## %bb.2
4269 ; X86-NEXT: addl $12, %esp
4270 ; X86-NEXT: vzeroupper
4271 ; X86-NEXT: retl
4272 ; X86-NEXT: LBB65_1: ## %bb.1
4273 ; X86-NEXT: vzeroupper
4274 ; X86-NEXT: calll _foo
4275 ; X86-NEXT: addl $12, %esp
4276 ; X86-NEXT: retl
4277 %a = icmp eq <16 x i32> %x, zeroinitializer
4278 %b = icmp eq <16 x i32> %y, zeroinitializer
4279 %c = and <16 x i1> %a, %b
4280 %d = bitcast <16 x i1> %c to i16
4281 %e = icmp sgt i16 %d, 0
4282 br i1 %e, label %bb.2, label %bb.1
4283 bb.1:
4284 call void @foo()
4285 br label %bb.2
4286 bb.2:
4287 ret void
4288 }
4289 declare void @foo()
4291 ; Make sure we can use the C flag from kortest to check for all ones.
4292 define void @ktest_allones(<16 x i32> %x, <16 x i32> %y) {
4293 ; CHECK-LABEL: ktest_allones:
4294 ; CHECK: ## %bb.0:
4295 ; CHECK-NEXT: pushq %rax
4296 ; CHECK-NEXT: .cfi_def_cfa_offset 16
4297 ; CHECK-NEXT: vpord %zmm1, %zmm0, %zmm0
4298 ; CHECK-NEXT: vptestnmd %zmm0, %zmm0, %k0
4299 ; CHECK-NEXT: kortestw %k0, %k0
4300 ; CHECK-NEXT: jb LBB66_2
4301 ; CHECK-NEXT: ## %bb.1: ## %bb.1
4302 ; CHECK-NEXT: vzeroupper
4303 ; CHECK-NEXT: callq _foo
4304 ; CHECK-NEXT: LBB66_2: ## %bb.2
4305 ; CHECK-NEXT: popq %rax
4306 ; CHECK-NEXT: vzeroupper
4307 ; CHECK-NEXT: retq
4308 ;
4309 ; X86-LABEL: ktest_allones:
4310 ; X86: ## %bb.0:
4311 ; X86-NEXT: subl $12, %esp
4312 ; X86-NEXT: .cfi_def_cfa_offset 16
4313 ; X86-NEXT: vpord %zmm1, %zmm0, %zmm0
4314 ; X86-NEXT: vptestnmd %zmm0, %zmm0, %k0
4315 ; X86-NEXT: kortestw %k0, %k0
4316 ; X86-NEXT: jb LBB66_2
4317 ; X86-NEXT: ## %bb.1: ## %bb.1
4318 ; X86-NEXT: vzeroupper
4319 ; X86-NEXT: calll _foo
4320 ; X86-NEXT: LBB66_2: ## %bb.2
4321 ; X86-NEXT: addl $12, %esp
4322 ; X86-NEXT: vzeroupper
4323 ; X86-NEXT: retl
4324 %a = icmp eq <16 x i32> %x, zeroinitializer
4325 %b = icmp eq <16 x i32> %y, zeroinitializer
4326 %c = and <16 x i1> %a, %b
4327 %d = bitcast <16 x i1> %c to i16
4328 %e = icmp eq i16 %d, -1
4329 br i1 %e, label %bb.2, label %bb.1
4330 bb.1:
4331 call void @foo()
4332 br label %bb.2
4333 bb.2:
4334 ret void
4335 }
4337 ; This is derived from an intrinsic test where v4i1 mask was created by _mm_cmp_epi32_mask, then it was passed to _mm512_mask_blend_epi32 which uses a v16i1 mask.
4338 ; The widening happens in the scalar domain between the intrinsics. The middle end optmized it to this.
4339 define <8 x i64> @mask_widening(<2 x i64> %a, <2 x i64> %b, <2 x i64> %c, <2 x i64> %d, <8 x i64> %e, <8 x i64> %f) {
4340 ; KNL-LABEL: mask_widening:
4341 ; KNL: ## %bb.0: ## %entry
4342 ; KNL-NEXT: ## kill: def $xmm1 killed $xmm1 def $zmm1
4343 ; KNL-NEXT: ## kill: def $xmm0 killed $xmm0 def $zmm0
4344 ; KNL-NEXT: vpcmpeqd %zmm1, %zmm0, %k0
4345 ; KNL-NEXT: kshiftlw $12, %k0, %k0
4346 ; KNL-NEXT: kshiftrw $12, %k0, %k1
4347 ; KNL-NEXT: vpblendmd %zmm5, %zmm4, %zmm0 {%k1}
4348 ; KNL-NEXT: retq
4349 ;
4350 ; SKX-LABEL: mask_widening:
4351 ; SKX: ## %bb.0: ## %entry
4352 ; SKX-NEXT: vpcmpeqd %xmm1, %xmm0, %k1
4353 ; SKX-NEXT: vpblendmd %zmm5, %zmm4, %zmm0 {%k1}
4354 ; SKX-NEXT: retq
4355 ;
4356 ; AVX512BW-LABEL: mask_widening:
4357 ; AVX512BW: ## %bb.0: ## %entry
4358 ; AVX512BW-NEXT: ## kill: def $xmm1 killed $xmm1 def $zmm1
4359 ; AVX512BW-NEXT: ## kill: def $xmm0 killed $xmm0 def $zmm0
4360 ; AVX512BW-NEXT: vpcmpeqd %zmm1, %zmm0, %k0
4361 ; AVX512BW-NEXT: kshiftlw $12, %k0, %k0
4362 ; AVX512BW-NEXT: kshiftrw $12, %k0, %k1
4363 ; AVX512BW-NEXT: vpblendmd %zmm5, %zmm4, %zmm0 {%k1}
4364 ; AVX512BW-NEXT: retq
4365 ;
4366 ; AVX512DQ-LABEL: mask_widening:
4367 ; AVX512DQ: ## %bb.0: ## %entry
4368 ; AVX512DQ-NEXT: ## kill: def $xmm1 killed $xmm1 def $zmm1
4369 ; AVX512DQ-NEXT: ## kill: def $xmm0 killed $xmm0 def $zmm0
4370 ; AVX512DQ-NEXT: vpcmpeqd %zmm1, %zmm0, %k0
4371 ; AVX512DQ-NEXT: kshiftlw $12, %k0, %k0
4372 ; AVX512DQ-NEXT: kshiftrw $12, %k0, %k1
4373 ; AVX512DQ-NEXT: vpblendmd %zmm5, %zmm4, %zmm0 {%k1}
4374 ; AVX512DQ-NEXT: retq
4375 ;
4376 ; X86-LABEL: mask_widening:
4377 ; X86: ## %bb.0: ## %entry
4378 ; X86-NEXT: pushl %ebp
4379 ; X86-NEXT: .cfi_def_cfa_offset 8
4380 ; X86-NEXT: .cfi_offset %ebp, -8
4381 ; X86-NEXT: movl %esp, %ebp
4382 ; X86-NEXT: .cfi_def_cfa_register %ebp
4383 ; X86-NEXT: andl $-64, %esp
4384 ; X86-NEXT: subl $64, %esp
4385 ; X86-NEXT: vpcmpeqd %xmm1, %xmm0, %k1
4386 ; X86-NEXT: vmovdqa64 8(%ebp), %zmm0
4387 ; X86-NEXT: vmovdqa32 72(%ebp), %zmm0 {%k1}
4388 ; X86-NEXT: movl %ebp, %esp
4389 ; X86-NEXT: popl %ebp
4390 ; X86-NEXT: retl
4391 entry:
4392 %0 = bitcast <2 x i64> %a to <4 x i32>
4393 %1 = bitcast <2 x i64> %b to <4 x i32>
4394 %2 = icmp eq <4 x i32> %0, %1
4395 %3 = shufflevector <4 x i1> %2, <4 x i1> zeroinitializer, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
4396 %4 = bitcast <8 x i64> %f to <16 x i32>
4397 %5 = bitcast <8 x i64> %e to <16 x i32>
4398 %6 = shufflevector <8 x i1> %3, <8 x i1> <i1 false, i1 undef, i1 false, i1 false, i1 false, i1 false, i1 false, i1 false>, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 8, i32 8, i32 8, i32 8, i32 8, i32 8, i32 8>
4399 %7 = select <16 x i1> %6, <16 x i32> %4, <16 x i32> %5
4400 %8 = bitcast <16 x i32> %7 to <8 x i64>
4401 ret <8 x i64> %8
4402 }
4404 define void @store_v64i1_constant(<64 x i1>* %R) {
4405 ; CHECK-LABEL: store_v64i1_constant:
4406 ; CHECK: ## %bb.0: ## %entry
4407 ; CHECK-NEXT: movabsq $-2305843576149381123, %rax ## imm = 0xDFFFFF7BFFFFEFFD
4408 ; CHECK-NEXT: movq %rax, (%rdi)
4409 ; CHECK-NEXT: retq
4410 ;
4411 ; X86-LABEL: store_v64i1_constant:
4412 ; X86: ## %bb.0: ## %entry
4413 ; X86-NEXT: movl {{[0-9]+}}(%esp), %eax
4414 ; X86-NEXT: movl $-536871045, 4(%eax) ## imm = 0xDFFFFF7B
4415 ; X86-NEXT: movl $-4099, (%eax) ## imm = 0xEFFD
4416 ; X86-NEXT: retl
4417 entry:
4418 store <64 x i1> <i1 1, i1 0, i1 1, i1 1, i1 1, i1 1, i1 1, i1 1, i1 1, i1 1, i1 1, i1 1, i1 0, i1 1, i1 1, i1 1, i1 1, i1 1, i1 1, i1 1, i1 1, i1 1, i1 1, i1 1, i1 1, i1 1, i1 1, i1 1, i1 1, i1 1, i1 1, i1 1, i1 1, i1 1, i1 0, i1 1, i1 1, i1 1, i1 1, i1 0, i1 1, i1 1, i1 1, i1 1, i1 1, i1 1, i1 1, i1 1, i1 1, i1 1, i1 1, i1 1, i1 1, i1 1, i1 1, i1 1, i1 1, i1 1, i1 1, i1 1, i1 1, i1 0, i1 1, i1 1>, <64 x i1>* %R
4419 ret void
4420 }
4422 define void @store_v2i1_constant(<2 x i1>* %R) {
4423 ; CHECK-LABEL: store_v2i1_constant:
4424 ; CHECK: ## %bb.0: ## %entry
4425 ; CHECK-NEXT: movb $1, (%rdi)
4426 ; CHECK-NEXT: retq
4427 ;
4428 ; X86-LABEL: store_v2i1_constant:
4429 ; X86: ## %bb.0: ## %entry
4430 ; X86-NEXT: movl {{[0-9]+}}(%esp), %eax
4431 ; X86-NEXT: movb $1, (%eax)
4432 ; X86-NEXT: retl
4433 entry:
4434 store <2 x i1> <i1 1, i1 0>, <2 x i1>* %R
4435 ret void
4436 }
4438 define void @store_v4i1_constant(<4 x i1>* %R) {
4439 ; CHECK-LABEL: store_v4i1_constant:
4440 ; CHECK: ## %bb.0: ## %entry
4441 ; CHECK-NEXT: movb $5, (%rdi)
4442 ; CHECK-NEXT: retq
4443 ;
4444 ; X86-LABEL: store_v4i1_constant:
4445 ; X86: ## %bb.0: ## %entry
4446 ; X86-NEXT: movl {{[0-9]+}}(%esp), %eax
4447 ; X86-NEXT: movb $5, (%eax)
4448 ; X86-NEXT: retl
4449 entry:
4450 store <4 x i1> <i1 1, i1 0, i1 1, i1 0>, <4 x i1>* %R
4451 ret void
4452 }
4454 ; Make sure we bring the -1 constant into the mask domain.
4455 define void @mask_not_cast(i8*, <8 x i64>, <8 x i64>, <8 x i64>, <8 x i64>) {
4456 ; CHECK-LABEL: mask_not_cast:
4457 ; CHECK: ## %bb.0:
4458 ; CHECK-NEXT: vpcmpnleud %zmm3, %zmm2, %k1
4459 ; CHECK-NEXT: vptestmd %zmm0, %zmm1, %k1 {%k1}
4460 ; CHECK-NEXT: vmovdqu32 %zmm0, (%rdi) {%k1}
4461 ; CHECK-NEXT: vzeroupper
4462 ; CHECK-NEXT: retq
4463 ;
4464 ; X86-LABEL: mask_not_cast:
4465 ; X86: ## %bb.0:
4466 ; X86-NEXT: movl {{[0-9]+}}(%esp), %eax
4467 ; X86-NEXT: vpcmpnleud %zmm3, %zmm2, %k1
4468 ; X86-NEXT: vptestmd %zmm0, %zmm1, %k1 {%k1}
4469 ; X86-NEXT: vmovdqu32 %zmm0, (%eax) {%k1}
4470 ; X86-NEXT: vzeroupper
4471 ; X86-NEXT: retl
4472 %6 = and <8 x i64> %2, %1
4473 %7 = bitcast <8 x i64> %6 to <16 x i32>
4474 %8 = icmp ne <16 x i32> %7, zeroinitializer
4475 %9 = bitcast <16 x i1> %8 to i16
4476 %10 = bitcast <8 x i64> %3 to <16 x i32>
4477 %11 = bitcast <8 x i64> %4 to <16 x i32>
4478 %12 = icmp ule <16 x i32> %10, %11
4479 %13 = bitcast <16 x i1> %12 to i16
4480 %14 = xor i16 %13, -1
4481 %15 = and i16 %14, %9
4482 %16 = bitcast <8 x i64> %1 to <16 x i32>
4483 %17 = bitcast i8* %0 to <16 x i32>*
4484 %18 = bitcast i16 %15 to <16 x i1>
4485 tail call void @llvm.masked.store.v16i32.p0v16i32(<16 x i32> %16, <16 x i32>* %17, i32 1, <16 x i1> %18) #2
4486 ret void
4487 }
4488 declare void @llvm.masked.store.v16i32.p0v16i32(<16 x i32>, <16 x i32>*, i32, <16 x i1>)
4490 define void @ktest_3(<8 x i32> %w, <8 x i32> %x, <8 x i32> %y, <8 x i32> %z) {
4491 ; KNL-LABEL: ktest_3:
4492 ; KNL: ## %bb.0:
4493 ; KNL-NEXT: pushq %rax
4494 ; KNL-NEXT: .cfi_def_cfa_offset 16
4495 ; KNL-NEXT: ## kill: def $ymm3 killed $ymm3 def $zmm3
4496 ; KNL-NEXT: ## kill: def $ymm2 killed $ymm2 def $zmm2
4497 ; KNL-NEXT: ## kill: def $ymm1 killed $ymm1 def $zmm1
4498 ; KNL-NEXT: ## kill: def $ymm0 killed $ymm0 def $zmm0
4499 ; KNL-NEXT: vptestnmd %zmm0, %zmm0, %k0
4500 ; KNL-NEXT: vptestnmd %zmm1, %zmm1, %k1
4501 ; KNL-NEXT: vptestnmd %zmm2, %zmm2, %k2
4502 ; KNL-NEXT: vptestnmd %zmm3, %zmm3, %k3
4503 ; KNL-NEXT: korw %k1, %k0, %k0
4504 ; KNL-NEXT: korw %k3, %k2, %k1
4505 ; KNL-NEXT: kandw %k1, %k0, %k0
4506 ; KNL-NEXT: kmovw %k0, %eax
4507 ; KNL-NEXT: testb %al, %al
4508 ; KNL-NEXT: je LBB72_1
4509 ; KNL-NEXT: ## %bb.2: ## %exit
4510 ; KNL-NEXT: popq %rax
4511 ; KNL-NEXT: vzeroupper
4512 ; KNL-NEXT: retq
4513 ; KNL-NEXT: LBB72_1: ## %bar
4514 ; KNL-NEXT: vzeroupper
4515 ; KNL-NEXT: callq _foo
4516 ; KNL-NEXT: popq %rax
4517 ; KNL-NEXT: retq
4518 ;
4519 ; SKX-LABEL: ktest_3:
4520 ; SKX: ## %bb.0:
4521 ; SKX-NEXT: pushq %rax
4522 ; SKX-NEXT: .cfi_def_cfa_offset 16
4523 ; SKX-NEXT: vptestnmd %ymm0, %ymm0, %k0
4524 ; SKX-NEXT: vptestnmd %ymm1, %ymm1, %k1
4525 ; SKX-NEXT: korb %k1, %k0, %k0
4526 ; SKX-NEXT: vptestnmd %ymm2, %ymm2, %k1
4527 ; SKX-NEXT: vptestnmd %ymm3, %ymm3, %k2
4528 ; SKX-NEXT: korb %k2, %k1, %k1
4529 ; SKX-NEXT: ktestb %k1, %k0
4530 ; SKX-NEXT: je LBB72_1
4531 ; SKX-NEXT: ## %bb.2: ## %exit
4532 ; SKX-NEXT: popq %rax
4533 ; SKX-NEXT: vzeroupper
4534 ; SKX-NEXT: retq
4535 ; SKX-NEXT: LBB72_1: ## %bar
4536 ; SKX-NEXT: vzeroupper
4537 ; SKX-NEXT: callq _foo
4538 ; SKX-NEXT: popq %rax
4539 ; SKX-NEXT: retq
4540 ;
4541 ; AVX512BW-LABEL: ktest_3:
4542 ; AVX512BW: ## %bb.0:
4543 ; AVX512BW-NEXT: pushq %rax
4544 ; AVX512BW-NEXT: .cfi_def_cfa_offset 16
4545 ; AVX512BW-NEXT: ## kill: def $ymm3 killed $ymm3 def $zmm3
4546 ; AVX512BW-NEXT: ## kill: def $ymm2 killed $ymm2 def $zmm2
4547 ; AVX512BW-NEXT: ## kill: def $ymm1 killed $ymm1 def $zmm1
4548 ; AVX512BW-NEXT: ## kill: def $ymm0 killed $ymm0 def $zmm0
4549 ; AVX512BW-NEXT: vptestnmd %zmm0, %zmm0, %k0
4550 ; AVX512BW-NEXT: vptestnmd %zmm1, %zmm1, %k1
4551 ; AVX512BW-NEXT: vptestnmd %zmm2, %zmm2, %k2
4552 ; AVX512BW-NEXT: vptestnmd %zmm3, %zmm3, %k3
4553 ; AVX512BW-NEXT: korw %k1, %k0, %k0
4554 ; AVX512BW-NEXT: korw %k3, %k2, %k1
4555 ; AVX512BW-NEXT: kandw %k1, %k0, %k0
4556 ; AVX512BW-NEXT: kmovd %k0, %eax
4557 ; AVX512BW-NEXT: testb %al, %al
4558 ; AVX512BW-NEXT: je LBB72_1
4559 ; AVX512BW-NEXT: ## %bb.2: ## %exit
4560 ; AVX512BW-NEXT: popq %rax
4561 ; AVX512BW-NEXT: vzeroupper
4562 ; AVX512BW-NEXT: retq
4563 ; AVX512BW-NEXT: LBB72_1: ## %bar
4564 ; AVX512BW-NEXT: vzeroupper
4565 ; AVX512BW-NEXT: callq _foo
4566 ; AVX512BW-NEXT: popq %rax
4567 ; AVX512BW-NEXT: retq
4568 ;
4569 ; AVX512DQ-LABEL: ktest_3:
4570 ; AVX512DQ: ## %bb.0:
4571 ; AVX512DQ-NEXT: pushq %rax
4572 ; AVX512DQ-NEXT: .cfi_def_cfa_offset 16
4573 ; AVX512DQ-NEXT: ## kill: def $ymm3 killed $ymm3 def $zmm3
4574 ; AVX512DQ-NEXT: ## kill: def $ymm2 killed $ymm2 def $zmm2
4575 ; AVX512DQ-NEXT: ## kill: def $ymm1 killed $ymm1 def $zmm1
4576 ; AVX512DQ-NEXT: ## kill: def $ymm0 killed $ymm0 def $zmm0
4577 ; AVX512DQ-NEXT: vptestnmd %zmm0, %zmm0, %k0
4578 ; AVX512DQ-NEXT: vptestnmd %zmm1, %zmm1, %k1
4579 ; AVX512DQ-NEXT: vptestnmd %zmm2, %zmm2, %k2
4580 ; AVX512DQ-NEXT: vptestnmd %zmm3, %zmm3, %k3
4581 ; AVX512DQ-NEXT: korb %k1, %k0, %k0
4582 ; AVX512DQ-NEXT: korb %k3, %k2, %k1
4583 ; AVX512DQ-NEXT: ktestb %k1, %k0
4584 ; AVX512DQ-NEXT: je LBB72_1
4585 ; AVX512DQ-NEXT: ## %bb.2: ## %exit
4586 ; AVX512DQ-NEXT: popq %rax
4587 ; AVX512DQ-NEXT: vzeroupper
4588 ; AVX512DQ-NEXT: retq
4589 ; AVX512DQ-NEXT: LBB72_1: ## %bar
4590 ; AVX512DQ-NEXT: vzeroupper
4591 ; AVX512DQ-NEXT: callq _foo
4592 ; AVX512DQ-NEXT: popq %rax
4593 ; AVX512DQ-NEXT: retq
4594 ;
4595 ; X86-LABEL: ktest_3:
4596 ; X86: ## %bb.0:
4597 ; X86-NEXT: subl $12, %esp
4598 ; X86-NEXT: .cfi_def_cfa_offset 16
4599 ; X86-NEXT: vptestnmd %ymm0, %ymm0, %k0
4600 ; X86-NEXT: vptestnmd %ymm1, %ymm1, %k1
4601 ; X86-NEXT: korb %k1, %k0, %k0
4602 ; X86-NEXT: vptestnmd %ymm2, %ymm2, %k1
4603 ; X86-NEXT: vptestnmd %ymm3, %ymm3, %k2
4604 ; X86-NEXT: korb %k2, %k1, %k1
4605 ; X86-NEXT: ktestb %k1, %k0
4606 ; X86-NEXT: je LBB72_1
4607 ; X86-NEXT: ## %bb.2: ## %exit
4608 ; X86-NEXT: addl $12, %esp
4609 ; X86-NEXT: vzeroupper
4610 ; X86-NEXT: retl
4611 ; X86-NEXT: LBB72_1: ## %bar
4612 ; X86-NEXT: vzeroupper
4613 ; X86-NEXT: calll _foo
4614 ; X86-NEXT: addl $12, %esp
4615 ; X86-NEXT: retl
4616 %a = icmp eq <8 x i32> %w, zeroinitializer
4617 %b = icmp eq <8 x i32> %x, zeroinitializer
4618 %c = icmp eq <8 x i32> %y, zeroinitializer
4619 %d = icmp eq <8 x i32> %z, zeroinitializer
4620 %e = or <8 x i1> %a, %b
4621 %f = or <8 x i1> %c, %d
4622 %g = and <8 x i1> %e, %f
4623 %h = bitcast <8 x i1> %g to i8
4624 %i = icmp eq i8 %h, 0
4625 br i1 %i, label %bar, label %exit
4627 bar:
4628 call void @foo()
4629 br label %exit
4631 exit:
4632 ret void
4633 }
4635 define void @ktest_4(<8 x i64> %w, <8 x i64> %x, <8 x i64> %y, <8 x i64> %z) {
4636 ; KNL-LABEL: ktest_4:
4637 ; KNL: ## %bb.0:
4638 ; KNL-NEXT: pushq %rax
4639 ; KNL-NEXT: .cfi_def_cfa_offset 16
4640 ; KNL-NEXT: vptestnmq %zmm0, %zmm0, %k0
4641 ; KNL-NEXT: vptestnmq %zmm1, %zmm1, %k1
4642 ; KNL-NEXT: vptestnmq %zmm2, %zmm2, %k2
4643 ; KNL-NEXT: vptestnmq %zmm3, %zmm3, %k3
4644 ; KNL-NEXT: korw %k1, %k0, %k0
4645 ; KNL-NEXT: korw %k3, %k2, %k1
4646 ; KNL-NEXT: kandw %k1, %k0, %k0
4647 ; KNL-NEXT: kmovw %k0, %eax
4648 ; KNL-NEXT: testb %al, %al
4649 ; KNL-NEXT: je LBB73_1
4650 ; KNL-NEXT: ## %bb.2: ## %exit
4651 ; KNL-NEXT: popq %rax
4652 ; KNL-NEXT: vzeroupper
4653 ; KNL-NEXT: retq
4654 ; KNL-NEXT: LBB73_1: ## %bar
4655 ; KNL-NEXT: vzeroupper
4656 ; KNL-NEXT: callq _foo
4657 ; KNL-NEXT: popq %rax
4658 ; KNL-NEXT: retq
4659 ;
4660 ; SKX-LABEL: ktest_4:
4661 ; SKX: ## %bb.0:
4662 ; SKX-NEXT: pushq %rax
4663 ; SKX-NEXT: .cfi_def_cfa_offset 16
4664 ; SKX-NEXT: vptestnmq %zmm0, %zmm0, %k0
4665 ; SKX-NEXT: vptestnmq %zmm1, %zmm1, %k1
4666 ; SKX-NEXT: korb %k1, %k0, %k0
4667 ; SKX-NEXT: vptestnmq %zmm2, %zmm2, %k1
4668 ; SKX-NEXT: vptestnmq %zmm3, %zmm3, %k2
4669 ; SKX-NEXT: korb %k2, %k1, %k1
4670 ; SKX-NEXT: ktestb %k1, %k0
4671 ; SKX-NEXT: je LBB73_1
4672 ; SKX-NEXT: ## %bb.2: ## %exit
4673 ; SKX-NEXT: popq %rax
4674 ; SKX-NEXT: vzeroupper
4675 ; SKX-NEXT: retq
4676 ; SKX-NEXT: LBB73_1: ## %bar
4677 ; SKX-NEXT: vzeroupper
4678 ; SKX-NEXT: callq _foo
4679 ; SKX-NEXT: popq %rax
4680 ; SKX-NEXT: retq
4681 ;
4682 ; AVX512BW-LABEL: ktest_4:
4683 ; AVX512BW: ## %bb.0:
4684 ; AVX512BW-NEXT: pushq %rax
4685 ; AVX512BW-NEXT: .cfi_def_cfa_offset 16
4686 ; AVX512BW-NEXT: vptestnmq %zmm0, %zmm0, %k0
4687 ; AVX512BW-NEXT: vptestnmq %zmm1, %zmm1, %k1
4688 ; AVX512BW-NEXT: vptestnmq %zmm2, %zmm2, %k2
4689 ; AVX512BW-NEXT: vptestnmq %zmm3, %zmm3, %k3
4690 ; AVX512BW-NEXT: korw %k1, %k0, %k0
4691 ; AVX512BW-NEXT: korw %k3, %k2, %k1
4692 ; AVX512BW-NEXT: kandw %k1, %k0, %k0
4693 ; AVX512BW-NEXT: kmovd %k0, %eax
4694 ; AVX512BW-NEXT: testb %al, %al
4695 ; AVX512BW-NEXT: je LBB73_1
4696 ; AVX512BW-NEXT: ## %bb.2: ## %exit
4697 ; AVX512BW-NEXT: popq %rax
4698 ; AVX512BW-NEXT: vzeroupper
4699 ; AVX512BW-NEXT: retq
4700 ; AVX512BW-NEXT: LBB73_1: ## %bar
4701 ; AVX512BW-NEXT: vzeroupper
4702 ; AVX512BW-NEXT: callq _foo
4703 ; AVX512BW-NEXT: popq %rax
4704 ; AVX512BW-NEXT: retq
4705 ;
4706 ; AVX512DQ-LABEL: ktest_4:
4707 ; AVX512DQ: ## %bb.0:
4708 ; AVX512DQ-NEXT: pushq %rax
4709 ; AVX512DQ-NEXT: .cfi_def_cfa_offset 16
4710 ; AVX512DQ-NEXT: vptestnmq %zmm0, %zmm0, %k0
4711 ; AVX512DQ-NEXT: vptestnmq %zmm1, %zmm1, %k1
4712 ; AVX512DQ-NEXT: korb %k1, %k0, %k0
4713 ; AVX512DQ-NEXT: vptestnmq %zmm2, %zmm2, %k1
4714 ; AVX512DQ-NEXT: vptestnmq %zmm3, %zmm3, %k2
4715 ; AVX512DQ-NEXT: korb %k2, %k1, %k1
4716 ; AVX512DQ-NEXT: ktestb %k1, %k0
4717 ; AVX512DQ-NEXT: je LBB73_1
4718 ; AVX512DQ-NEXT: ## %bb.2: ## %exit
4719 ; AVX512DQ-NEXT: popq %rax
4720 ; AVX512DQ-NEXT: vzeroupper
4721 ; AVX512DQ-NEXT: retq
4722 ; AVX512DQ-NEXT: LBB73_1: ## %bar
4723 ; AVX512DQ-NEXT: vzeroupper
4724 ; AVX512DQ-NEXT: callq _foo
4725 ; AVX512DQ-NEXT: popq %rax
4726 ; AVX512DQ-NEXT: retq
4727 ;
4728 ; X86-LABEL: ktest_4:
4729 ; X86: ## %bb.0:
4730 ; X86-NEXT: subl $12, %esp
4731 ; X86-NEXT: .cfi_def_cfa_offset 16
4732 ; X86-NEXT: vptestnmq %zmm0, %zmm0, %k0
4733 ; X86-NEXT: vptestnmq %zmm1, %zmm1, %k1
4734 ; X86-NEXT: korb %k1, %k0, %k0
4735 ; X86-NEXT: vptestnmq %zmm2, %zmm2, %k1
4736 ; X86-NEXT: vptestnmq %zmm3, %zmm3, %k2
4737 ; X86-NEXT: korb %k2, %k1, %k1
4738 ; X86-NEXT: ktestb %k1, %k0
4739 ; X86-NEXT: je LBB73_1
4740 ; X86-NEXT: ## %bb.2: ## %exit
4741 ; X86-NEXT: addl $12, %esp
4742 ; X86-NEXT: vzeroupper
4743 ; X86-NEXT: retl
4744 ; X86-NEXT: LBB73_1: ## %bar
4745 ; X86-NEXT: vzeroupper
4746 ; X86-NEXT: calll _foo
4747 ; X86-NEXT: addl $12, %esp
4748 ; X86-NEXT: retl
4749 %a = icmp eq <8 x i64> %w, zeroinitializer
4750 %b = icmp eq <8 x i64> %x, zeroinitializer
4751 %c = icmp eq <8 x i64> %y, zeroinitializer
4752 %d = icmp eq <8 x i64> %z, zeroinitializer
4753 %e = or <8 x i1> %a, %b
4754 %f = or <8 x i1> %c, %d
4755 %g = and <8 x i1> %e, %f
4756 %h = bitcast <8 x i1> %g to i8
4757 %i = icmp eq i8 %h, 0
4758 br i1 %i, label %bar, label %exit
4760 bar:
4761 call void @foo()
4762 br label %exit
4764 exit:
4765 ret void
4766 }
4768 define void @ktest_5(<16 x i32> %w, <16 x i32> %x, <16 x i32> %y, <16 x i32> %z) {
4769 ; KNL-LABEL: ktest_5:
4770 ; KNL: ## %bb.0:
4771 ; KNL-NEXT: pushq %rax
4772 ; KNL-NEXT: .cfi_def_cfa_offset 16
4773 ; KNL-NEXT: vptestnmd %zmm0, %zmm0, %k0
4774 ; KNL-NEXT: vptestnmd %zmm1, %zmm1, %k1
4775 ; KNL-NEXT: korw %k1, %k0, %k0
4776 ; KNL-NEXT: vptestnmd %zmm2, %zmm2, %k1
4777 ; KNL-NEXT: vptestnmd %zmm3, %zmm3, %k2
4778 ; KNL-NEXT: korw %k2, %k1, %k1
4779 ; KNL-NEXT: kandw %k1, %k0, %k0
4780 ; KNL-NEXT: kortestw %k0, %k0
4781 ; KNL-NEXT: je LBB74_1
4782 ; KNL-NEXT: ## %bb.2: ## %exit
4783 ; KNL-NEXT: popq %rax
4784 ; KNL-NEXT: vzeroupper
4785 ; KNL-NEXT: retq
4786 ; KNL-NEXT: LBB74_1: ## %bar
4787 ; KNL-NEXT: vzeroupper
4788 ; KNL-NEXT: callq _foo
4789 ; KNL-NEXT: popq %rax
4790 ; KNL-NEXT: retq
4791 ;
4792 ; SKX-LABEL: ktest_5:
4793 ; SKX: ## %bb.0:
4794 ; SKX-NEXT: pushq %rax
4795 ; SKX-NEXT: .cfi_def_cfa_offset 16
4796 ; SKX-NEXT: vptestnmd %zmm0, %zmm0, %k0
4797 ; SKX-NEXT: vptestnmd %zmm1, %zmm1, %k1
4798 ; SKX-NEXT: korw %k1, %k0, %k0
4799 ; SKX-NEXT: vptestnmd %zmm2, %zmm2, %k1
4800 ; SKX-NEXT: vptestnmd %zmm3, %zmm3, %k2
4801 ; SKX-NEXT: korw %k2, %k1, %k1
4802 ; SKX-NEXT: ktestw %k1, %k0
4803 ; SKX-NEXT: je LBB74_1
4804 ; SKX-NEXT: ## %bb.2: ## %exit
4805 ; SKX-NEXT: popq %rax
4806 ; SKX-NEXT: vzeroupper
4807 ; SKX-NEXT: retq
4808 ; SKX-NEXT: LBB74_1: ## %bar
4809 ; SKX-NEXT: vzeroupper
4810 ; SKX-NEXT: callq _foo
4811 ; SKX-NEXT: popq %rax
4812 ; SKX-NEXT: retq
4813 ;
4814 ; AVX512BW-LABEL: ktest_5:
4815 ; AVX512BW: ## %bb.0:
4816 ; AVX512BW-NEXT: pushq %rax
4817 ; AVX512BW-NEXT: .cfi_def_cfa_offset 16
4818 ; AVX512BW-NEXT: vptestnmd %zmm0, %zmm0, %k0
4819 ; AVX512BW-NEXT: vptestnmd %zmm1, %zmm1, %k1
4820 ; AVX512BW-NEXT: korw %k1, %k0, %k0
4821 ; AVX512BW-NEXT: vptestnmd %zmm2, %zmm2, %k1
4822 ; AVX512BW-NEXT: vptestnmd %zmm3, %zmm3, %k2
4823 ; AVX512BW-NEXT: korw %k2, %k1, %k1
4824 ; AVX512BW-NEXT: kandw %k1, %k0, %k0
4825 ; AVX512BW-NEXT: kortestw %k0, %k0
4826 ; AVX512BW-NEXT: je LBB74_1
4827 ; AVX512BW-NEXT: ## %bb.2: ## %exit
4828 ; AVX512BW-NEXT: popq %rax
4829 ; AVX512BW-NEXT: vzeroupper
4830 ; AVX512BW-NEXT: retq
4831 ; AVX512BW-NEXT: LBB74_1: ## %bar
4832 ; AVX512BW-NEXT: vzeroupper
4833 ; AVX512BW-NEXT: callq _foo
4834 ; AVX512BW-NEXT: popq %rax
4835 ; AVX512BW-NEXT: retq
4836 ;
4837 ; AVX512DQ-LABEL: ktest_5:
4838 ; AVX512DQ: ## %bb.0:
4839 ; AVX512DQ-NEXT: pushq %rax
4840 ; AVX512DQ-NEXT: .cfi_def_cfa_offset 16
4841 ; AVX512DQ-NEXT: vptestnmd %zmm0, %zmm0, %k0
4842 ; AVX512DQ-NEXT: vptestnmd %zmm1, %zmm1, %k1
4843 ; AVX512DQ-NEXT: korw %k1, %k0, %k0
4844 ; AVX512DQ-NEXT: vptestnmd %zmm2, %zmm2, %k1
4845 ; AVX512DQ-NEXT: vptestnmd %zmm3, %zmm3, %k2
4846 ; AVX512DQ-NEXT: korw %k2, %k1, %k1
4847 ; AVX512DQ-NEXT: ktestw %k1, %k0
4848 ; AVX512DQ-NEXT: je LBB74_1
4849 ; AVX512DQ-NEXT: ## %bb.2: ## %exit
4850 ; AVX512DQ-NEXT: popq %rax
4851 ; AVX512DQ-NEXT: vzeroupper
4852 ; AVX512DQ-NEXT: retq
4853 ; AVX512DQ-NEXT: LBB74_1: ## %bar
4854 ; AVX512DQ-NEXT: vzeroupper
4855 ; AVX512DQ-NEXT: callq _foo
4856 ; AVX512DQ-NEXT: popq %rax
4857 ; AVX512DQ-NEXT: retq
4858 ;
4859 ; X86-LABEL: ktest_5:
4860 ; X86: ## %bb.0:
4861 ; X86-NEXT: subl $12, %esp
4862 ; X86-NEXT: .cfi_def_cfa_offset 16
4863 ; X86-NEXT: vptestnmd %zmm0, %zmm0, %k0
4864 ; X86-NEXT: vptestnmd %zmm1, %zmm1, %k1
4865 ; X86-NEXT: korw %k1, %k0, %k0
4866 ; X86-NEXT: vptestnmd %zmm2, %zmm2, %k1
4867 ; X86-NEXT: vptestnmd %zmm3, %zmm3, %k2
4868 ; X86-NEXT: korw %k2, %k1, %k1
4869 ; X86-NEXT: ktestw %k1, %k0
4870 ; X86-NEXT: je LBB74_1
4871 ; X86-NEXT: ## %bb.2: ## %exit
4872 ; X86-NEXT: addl $12, %esp
4873 ; X86-NEXT: vzeroupper
4874 ; X86-NEXT: retl
4875 ; X86-NEXT: LBB74_1: ## %bar
4876 ; X86-NEXT: vzeroupper
4877 ; X86-NEXT: calll _foo
4878 ; X86-NEXT: addl $12, %esp
4879 ; X86-NEXT: retl
4880 %a = icmp eq <16 x i32> %w, zeroinitializer
4881 %b = icmp eq <16 x i32> %x, zeroinitializer
4882 %c = icmp eq <16 x i32> %y, zeroinitializer
4883 %d = icmp eq <16 x i32> %z, zeroinitializer
4884 %e = or <16 x i1> %a, %b
4885 %f = or <16 x i1> %c, %d
4886 %g = and <16 x i1> %e, %f
4887 %h = bitcast <16 x i1> %g to i16
4888 %i = icmp eq i16 %h, 0
4889 br i1 %i, label %bar, label %exit
4891 bar:
4892 call void @foo()
4893 br label %exit
4895 exit:
4896 ret void
4897 }
4899 define void @ktest_6(<32 x i16> %w, <32 x i16> %x, <32 x i16> %y, <32 x i16> %z) {
4900 ; KNL-LABEL: ktest_6:
4901 ; KNL: ## %bb.0:
4902 ; KNL-NEXT: pushq %rax
4903 ; KNL-NEXT: .cfi_def_cfa_offset 16
4904 ; KNL-NEXT: vextracti64x4 $1, %zmm3, %ymm4
4905 ; KNL-NEXT: vextracti64x4 $1, %zmm2, %ymm5
4906 ; KNL-NEXT: vextracti64x4 $1, %zmm1, %ymm6
4907 ; KNL-NEXT: vextracti64x4 $1, %zmm0, %ymm7
4908 ; KNL-NEXT: vpxor %xmm8, %xmm8, %xmm8
4909 ; KNL-NEXT: vpcmpeqw %ymm8, %ymm0, %ymm0
4910 ; KNL-NEXT: vpcmpeqw %ymm8, %ymm7, %ymm7
4911 ; KNL-NEXT: vpcmpeqw %ymm8, %ymm1, %ymm1
4912 ; KNL-NEXT: vpor %ymm1, %ymm0, %ymm0
4913 ; KNL-NEXT: vpcmpeqw %ymm8, %ymm6, %ymm1
4914 ; KNL-NEXT: vpor %ymm1, %ymm7, %ymm1
4915 ; KNL-NEXT: vpcmpeqw %ymm8, %ymm2, %ymm2
4916 ; KNL-NEXT: vpcmpeqw %ymm8, %ymm5, %ymm5
4917 ; KNL-NEXT: vpcmpeqw %ymm8, %ymm3, %ymm3
4918 ; KNL-NEXT: vpor %ymm3, %ymm2, %ymm2
4919 ; KNL-NEXT: vpand %ymm2, %ymm0, %ymm0
4920 ; KNL-NEXT: vpcmpeqw %ymm8, %ymm4, %ymm2
4921 ; KNL-NEXT: vpor %ymm2, %ymm5, %ymm2
4922 ; KNL-NEXT: vpand %ymm2, %ymm1, %ymm1
4923 ; KNL-NEXT: vpmovsxwd %ymm0, %zmm0
4924 ; KNL-NEXT: vptestmd %zmm0, %zmm0, %k0
4925 ; KNL-NEXT: kmovw %k0, %eax
4926 ; KNL-NEXT: vpmovsxwd %ymm1, %zmm0
4927 ; KNL-NEXT: vptestmd %zmm0, %zmm0, %k0
4928 ; KNL-NEXT: kmovw %k0, %ecx
4929 ; KNL-NEXT: shll $16, %ecx
4930 ; KNL-NEXT: orl %eax, %ecx
4931 ; KNL-NEXT: je LBB75_1
4932 ; KNL-NEXT: ## %bb.2: ## %exit
4933 ; KNL-NEXT: popq %rax
4934 ; KNL-NEXT: vzeroupper
4935 ; KNL-NEXT: retq
4936 ; KNL-NEXT: LBB75_1: ## %bar
4937 ; KNL-NEXT: vzeroupper
4938 ; KNL-NEXT: callq _foo
4939 ; KNL-NEXT: popq %rax
4940 ; KNL-NEXT: retq
4941 ;
4942 ; SKX-LABEL: ktest_6:
4943 ; SKX: ## %bb.0:
4944 ; SKX-NEXT: pushq %rax
4945 ; SKX-NEXT: .cfi_def_cfa_offset 16
4946 ; SKX-NEXT: vptestnmw %zmm0, %zmm0, %k0
4947 ; SKX-NEXT: vptestnmw %zmm1, %zmm1, %k1
4948 ; SKX-NEXT: kord %k1, %k0, %k0
4949 ; SKX-NEXT: vptestnmw %zmm2, %zmm2, %k1
4950 ; SKX-NEXT: vptestnmw %zmm3, %zmm3, %k2
4951 ; SKX-NEXT: kord %k2, %k1, %k1
4952 ; SKX-NEXT: ktestd %k1, %k0
4953 ; SKX-NEXT: je LBB75_1
4954 ; SKX-NEXT: ## %bb.2: ## %exit
4955 ; SKX-NEXT: popq %rax
4956 ; SKX-NEXT: vzeroupper
4957 ; SKX-NEXT: retq
4958 ; SKX-NEXT: LBB75_1: ## %bar
4959 ; SKX-NEXT: vzeroupper
4960 ; SKX-NEXT: callq _foo
4961 ; SKX-NEXT: popq %rax
4962 ; SKX-NEXT: retq
4963 ;
4964 ; AVX512BW-LABEL: ktest_6:
4965 ; AVX512BW: ## %bb.0:
4966 ; AVX512BW-NEXT: pushq %rax
4967 ; AVX512BW-NEXT: .cfi_def_cfa_offset 16
4968 ; AVX512BW-NEXT: vptestnmw %zmm0, %zmm0, %k0
4969 ; AVX512BW-NEXT: vptestnmw %zmm1, %zmm1, %k1
4970 ; AVX512BW-NEXT: kord %k1, %k0, %k0
4971 ; AVX512BW-NEXT: vptestnmw %zmm2, %zmm2, %k1
4972 ; AVX512BW-NEXT: vptestnmw %zmm3, %zmm3, %k2
4973 ; AVX512BW-NEXT: kord %k2, %k1, %k1
4974 ; AVX512BW-NEXT: ktestd %k1, %k0
4975 ; AVX512BW-NEXT: je LBB75_1
4976 ; AVX512BW-NEXT: ## %bb.2: ## %exit
4977 ; AVX512BW-NEXT: popq %rax
4978 ; AVX512BW-NEXT: vzeroupper
4979 ; AVX512BW-NEXT: retq
4980 ; AVX512BW-NEXT: LBB75_1: ## %bar
4981 ; AVX512BW-NEXT: vzeroupper
4982 ; AVX512BW-NEXT: callq _foo
4983 ; AVX512BW-NEXT: popq %rax
4984 ; AVX512BW-NEXT: retq
4985 ;
4986 ; AVX512DQ-LABEL: ktest_6:
4987 ; AVX512DQ: ## %bb.0:
4988 ; AVX512DQ-NEXT: pushq %rax
4989 ; AVX512DQ-NEXT: .cfi_def_cfa_offset 16
4990 ; AVX512DQ-NEXT: vextracti64x4 $1, %zmm3, %ymm4
4991 ; AVX512DQ-NEXT: vextracti64x4 $1, %zmm2, %ymm5
4992 ; AVX512DQ-NEXT: vextracti64x4 $1, %zmm1, %ymm6
4993 ; AVX512DQ-NEXT: vextracti64x4 $1, %zmm0, %ymm7
4994 ; AVX512DQ-NEXT: vpxor %xmm8, %xmm8, %xmm8
4995 ; AVX512DQ-NEXT: vpcmpeqw %ymm8, %ymm0, %ymm0
4996 ; AVX512DQ-NEXT: vpcmpeqw %ymm8, %ymm7, %ymm7
4997 ; AVX512DQ-NEXT: vpcmpeqw %ymm8, %ymm1, %ymm1
4998 ; AVX512DQ-NEXT: vpor %ymm1, %ymm0, %ymm0
4999 ; AVX512DQ-NEXT: vpcmpeqw %ymm8, %ymm6, %ymm1
5000 ; AVX512DQ-NEXT: vpor %ymm1, %ymm7, %ymm1
5001 ; AVX512DQ-NEXT: vpcmpeqw %ymm8, %ymm2, %ymm2
5002 ; AVX512DQ-NEXT: vpcmpeqw %ymm8, %ymm5, %ymm5
5003 ; AVX512DQ-NEXT: vpcmpeqw %ymm8, %ymm3, %ymm3
5004 ; AVX512DQ-NEXT: vpor %ymm3, %ymm2, %ymm2
5005 ; AVX512DQ-NEXT: vpand %ymm2, %ymm0, %ymm0
5006 ; AVX512DQ-NEXT: vpcmpeqw %ymm8, %ymm4, %ymm2
5007 ; AVX512DQ-NEXT: vpor %ymm2, %ymm5, %ymm2
5008 ; AVX512DQ-NEXT: vpand %ymm2, %ymm1, %ymm1
5009 ; AVX512DQ-NEXT: vpmovsxwd %ymm0, %zmm0
5010 ; AVX512DQ-NEXT: vpmovd2m %zmm0, %k0
5011 ; AVX512DQ-NEXT: kmovw %k0, %eax
5012 ; AVX512DQ-NEXT: vpmovsxwd %ymm1, %zmm0
5013 ; AVX512DQ-NEXT: vpmovd2m %zmm0, %k0
5014 ; AVX512DQ-NEXT: kmovw %k0, %ecx
5015 ; AVX512DQ-NEXT: shll $16, %ecx
5016 ; AVX512DQ-NEXT: orl %eax, %ecx
5017 ; AVX512DQ-NEXT: je LBB75_1
5018 ; AVX512DQ-NEXT: ## %bb.2: ## %exit
5019 ; AVX512DQ-NEXT: popq %rax
5020 ; AVX512DQ-NEXT: vzeroupper
5021 ; AVX512DQ-NEXT: retq
5022 ; AVX512DQ-NEXT: LBB75_1: ## %bar
5023 ; AVX512DQ-NEXT: vzeroupper
5024 ; AVX512DQ-NEXT: callq _foo
5025 ; AVX512DQ-NEXT: popq %rax
5026 ; AVX512DQ-NEXT: retq
5027 ;
5028 ; X86-LABEL: ktest_6:
5029 ; X86: ## %bb.0:
5030 ; X86-NEXT: subl $12, %esp
5031 ; X86-NEXT: .cfi_def_cfa_offset 16
5032 ; X86-NEXT: vptestnmw %zmm0, %zmm0, %k0
5033 ; X86-NEXT: vptestnmw %zmm1, %zmm1, %k1
5034 ; X86-NEXT: kord %k1, %k0, %k0
5035 ; X86-NEXT: vptestnmw %zmm2, %zmm2, %k1
5036 ; X86-NEXT: vptestnmw %zmm3, %zmm3, %k2
5037 ; X86-NEXT: kord %k2, %k1, %k1
5038 ; X86-NEXT: ktestd %k1, %k0
5039 ; X86-NEXT: je LBB75_1
5040 ; X86-NEXT: ## %bb.2: ## %exit
5041 ; X86-NEXT: addl $12, %esp
5042 ; X86-NEXT: vzeroupper
5043 ; X86-NEXT: retl
5044 ; X86-NEXT: LBB75_1: ## %bar
5045 ; X86-NEXT: vzeroupper
5046 ; X86-NEXT: calll _foo
5047 ; X86-NEXT: addl $12, %esp
5048 ; X86-NEXT: retl
5049 %a = icmp eq <32 x i16> %w, zeroinitializer
5050 %b = icmp eq <32 x i16> %x, zeroinitializer
5051 %c = icmp eq <32 x i16> %y, zeroinitializer
5052 %d = icmp eq <32 x i16> %z, zeroinitializer
5053 %e = or <32 x i1> %a, %b
5054 %f = or <32 x i1> %c, %d
5055 %g = and <32 x i1> %e, %f
5056 %h = bitcast <32 x i1> %g to i32
5057 %i = icmp eq i32 %h, 0
5058 br i1 %i, label %bar, label %exit
5060 bar:
5061 call void @foo()
5062 br label %exit
5064 exit:
5065 ret void
5066 }
5068 define void @ktest_7(<64 x i8> %w, <64 x i8> %x, <64 x i8> %y, <64 x i8> %z) {
5069 ; KNL-LABEL: ktest_7:
5070 ; KNL: ## %bb.0:
5071 ; KNL-NEXT: pushq %rax
5072 ; KNL-NEXT: .cfi_def_cfa_offset 16
5073 ; KNL-NEXT: vextracti64x4 $1, %zmm3, %ymm9
5074 ; KNL-NEXT: vextracti64x4 $1, %zmm2, %ymm10
5075 ; KNL-NEXT: vextracti64x4 $1, %zmm1, %ymm11
5076 ; KNL-NEXT: vextracti64x4 $1, %zmm0, %ymm7
5077 ; KNL-NEXT: vpxor %xmm8, %xmm8, %xmm8
5078 ; KNL-NEXT: vpcmpeqb %ymm8, %ymm0, %ymm13
5079 ; KNL-NEXT: vextracti128 $1, %ymm13, %xmm4
5080 ; KNL-NEXT: vpcmpeqb %ymm8, %ymm7, %ymm7
5081 ; KNL-NEXT: vextracti128 $1, %ymm7, %xmm5
5082 ; KNL-NEXT: vpcmpeqb %ymm8, %ymm1, %ymm1
5083 ; KNL-NEXT: vextracti128 $1, %ymm1, %xmm6
5084 ; KNL-NEXT: vpor %xmm6, %xmm4, %xmm12
5085 ; KNL-NEXT: vpcmpeqb %ymm8, %ymm11, %ymm6
5086 ; KNL-NEXT: vextracti128 $1, %ymm6, %xmm4
5087 ; KNL-NEXT: vpor %xmm4, %xmm5, %xmm11
5088 ; KNL-NEXT: vpcmpeqb %ymm8, %ymm2, %ymm2
5089 ; KNL-NEXT: vextracti128 $1, %ymm2, %xmm5
5090 ; KNL-NEXT: vpcmpeqb %ymm8, %ymm10, %ymm10
5091 ; KNL-NEXT: vextracti128 $1, %ymm10, %xmm4
5092 ; KNL-NEXT: vpcmpeqb %ymm8, %ymm3, %ymm3
5093 ; KNL-NEXT: vextracti128 $1, %ymm3, %xmm0
5094 ; KNL-NEXT: vpor %xmm0, %xmm5, %xmm0
5095 ; KNL-NEXT: vpand %xmm0, %xmm12, %xmm12
5096 ; KNL-NEXT: vpcmpeqb %ymm8, %ymm9, %ymm5
5097 ; KNL-NEXT: vextracti128 $1, %ymm5, %xmm0
5098 ; KNL-NEXT: vpor %xmm0, %xmm4, %xmm0
5099 ; KNL-NEXT: vpand %xmm0, %xmm11, %xmm0
5100 ; KNL-NEXT: vpor %xmm6, %xmm7, %xmm4
5101 ; KNL-NEXT: vpor %xmm1, %xmm13, %xmm1
5102 ; KNL-NEXT: vpor %xmm5, %xmm10, %xmm5
5103 ; KNL-NEXT: vpand %xmm5, %xmm4, %xmm4
5104 ; KNL-NEXT: vpor %xmm3, %xmm2, %xmm2
5105 ; KNL-NEXT: vpand %xmm2, %xmm1, %xmm1
5106 ; KNL-NEXT: vpmovsxbd %xmm1, %zmm1
5107 ; KNL-NEXT: vptestmd %zmm1, %zmm1, %k0
5108 ; KNL-NEXT: kmovw %k0, %eax
5109 ; KNL-NEXT: vpmovsxbd %xmm12, %zmm1
5110 ; KNL-NEXT: vptestmd %zmm1, %zmm1, %k0
5111 ; KNL-NEXT: kmovw %k0, %ecx
5112 ; KNL-NEXT: shll $16, %ecx
5113 ; KNL-NEXT: orl %eax, %ecx
5114 ; KNL-NEXT: vpmovsxbd %xmm4, %zmm1
5115 ; KNL-NEXT: vptestmd %zmm1, %zmm1, %k0
5116 ; KNL-NEXT: kmovw %k0, %eax
5117 ; KNL-NEXT: vpmovsxbd %xmm0, %zmm0
5118 ; KNL-NEXT: vptestmd %zmm0, %zmm0, %k0
5119 ; KNL-NEXT: kmovw %k0, %edx
5120 ; KNL-NEXT: shll $16, %edx
5121 ; KNL-NEXT: orl %eax, %edx
5122 ; KNL-NEXT: shlq $32, %rdx
5123 ; KNL-NEXT: orq %rcx, %rdx
5124 ; KNL-NEXT: je LBB76_1
5125 ; KNL-NEXT: ## %bb.2: ## %exit
5126 ; KNL-NEXT: popq %rax
5127 ; KNL-NEXT: vzeroupper
5128 ; KNL-NEXT: retq
5129 ; KNL-NEXT: LBB76_1: ## %bar
5130 ; KNL-NEXT: vzeroupper
5131 ; KNL-NEXT: callq _foo
5132 ; KNL-NEXT: popq %rax
5133 ; KNL-NEXT: retq
5134 ;
5135 ; SKX-LABEL: ktest_7:
5136 ; SKX: ## %bb.0:
5137 ; SKX-NEXT: pushq %rax
5138 ; SKX-NEXT: .cfi_def_cfa_offset 16
5139 ; SKX-NEXT: vptestnmb %zmm0, %zmm0, %k0
5140 ; SKX-NEXT: vptestnmb %zmm1, %zmm1, %k1
5141 ; SKX-NEXT: korq %k1, %k0, %k0
5142 ; SKX-NEXT: vptestnmb %zmm2, %zmm2, %k1
5143 ; SKX-NEXT: vptestnmb %zmm3, %zmm3, %k2
5144 ; SKX-NEXT: korq %k2, %k1, %k1
5145 ; SKX-NEXT: ktestq %k1, %k0
5146 ; SKX-NEXT: je LBB76_1
5147 ; SKX-NEXT: ## %bb.2: ## %exit
5148 ; SKX-NEXT: popq %rax
5149 ; SKX-NEXT: vzeroupper
5150 ; SKX-NEXT: retq
5151 ; SKX-NEXT: LBB76_1: ## %bar
5152 ; SKX-NEXT: vzeroupper
5153 ; SKX-NEXT: callq _foo
5154 ; SKX-NEXT: popq %rax
5155 ; SKX-NEXT: retq
5156 ;
5157 ; AVX512BW-LABEL: ktest_7:
5158 ; AVX512BW: ## %bb.0:
5159 ; AVX512BW-NEXT: pushq %rax
5160 ; AVX512BW-NEXT: .cfi_def_cfa_offset 16
5161 ; AVX512BW-NEXT: vptestnmb %zmm0, %zmm0, %k0
5162 ; AVX512BW-NEXT: vptestnmb %zmm1, %zmm1, %k1
5163 ; AVX512BW-NEXT: korq %k1, %k0, %k0
5164 ; AVX512BW-NEXT: vptestnmb %zmm2, %zmm2, %k1
5165 ; AVX512BW-NEXT: vptestnmb %zmm3, %zmm3, %k2
5166 ; AVX512BW-NEXT: korq %k2, %k1, %k1
5167 ; AVX512BW-NEXT: ktestq %k1, %k0
5168 ; AVX512BW-NEXT: je LBB76_1
5169 ; AVX512BW-NEXT: ## %bb.2: ## %exit
5170 ; AVX512BW-NEXT: popq %rax
5171 ; AVX512BW-NEXT: vzeroupper
5172 ; AVX512BW-NEXT: retq
5173 ; AVX512BW-NEXT: LBB76_1: ## %bar
5174 ; AVX512BW-NEXT: vzeroupper
5175 ; AVX512BW-NEXT: callq _foo
5176 ; AVX512BW-NEXT: popq %rax
5177 ; AVX512BW-NEXT: retq
5178 ;
5179 ; AVX512DQ-LABEL: ktest_7:
5180 ; AVX512DQ: ## %bb.0:
5181 ; AVX512DQ-NEXT: pushq %rax
5182 ; AVX512DQ-NEXT: .cfi_def_cfa_offset 16
5183 ; AVX512DQ-NEXT: vextracti64x4 $1, %zmm3, %ymm9
5184 ; AVX512DQ-NEXT: vextracti64x4 $1, %zmm2, %ymm10
5185 ; AVX512DQ-NEXT: vextracti64x4 $1, %zmm1, %ymm11
5186 ; AVX512DQ-NEXT: vextracti64x4 $1, %zmm0, %ymm7
5187 ; AVX512DQ-NEXT: vpxor %xmm8, %xmm8, %xmm8
5188 ; AVX512DQ-NEXT: vpcmpeqb %ymm8, %ymm0, %ymm13
5189 ; AVX512DQ-NEXT: vextracti128 $1, %ymm13, %xmm4
5190 ; AVX512DQ-NEXT: vpcmpeqb %ymm8, %ymm7, %ymm7
5191 ; AVX512DQ-NEXT: vextracti128 $1, %ymm7, %xmm5
5192 ; AVX512DQ-NEXT: vpcmpeqb %ymm8, %ymm1, %ymm1
5193 ; AVX512DQ-NEXT: vextracti128 $1, %ymm1, %xmm6
5194 ; AVX512DQ-NEXT: vpor %xmm6, %xmm4, %xmm12
5195 ; AVX512DQ-NEXT: vpcmpeqb %ymm8, %ymm11, %ymm6
5196 ; AVX512DQ-NEXT: vextracti128 $1, %ymm6, %xmm4
5197 ; AVX512DQ-NEXT: vpor %xmm4, %xmm5, %xmm11
5198 ; AVX512DQ-NEXT: vpcmpeqb %ymm8, %ymm2, %ymm2
5199 ; AVX512DQ-NEXT: vextracti128 $1, %ymm2, %xmm5
5200 ; AVX512DQ-NEXT: vpcmpeqb %ymm8, %ymm10, %ymm10
5201 ; AVX512DQ-NEXT: vextracti128 $1, %ymm10, %xmm4
5202 ; AVX512DQ-NEXT: vpcmpeqb %ymm8, %ymm3, %ymm3
5203 ; AVX512DQ-NEXT: vextracti128 $1, %ymm3, %xmm0
5204 ; AVX512DQ-NEXT: vpor %xmm0, %xmm5, %xmm0
5205 ; AVX512DQ-NEXT: vpand %xmm0, %xmm12, %xmm12
5206 ; AVX512DQ-NEXT: vpcmpeqb %ymm8, %ymm9, %ymm5
5207 ; AVX512DQ-NEXT: vextracti128 $1, %ymm5, %xmm0
5208 ; AVX512DQ-NEXT: vpor %xmm0, %xmm4, %xmm0
5209 ; AVX512DQ-NEXT: vpand %xmm0, %xmm11, %xmm0
5210 ; AVX512DQ-NEXT: vpor %xmm6, %xmm7, %xmm4
5211 ; AVX512DQ-NEXT: vpor %xmm1, %xmm13, %xmm1
5212 ; AVX512DQ-NEXT: vpor %xmm5, %xmm10, %xmm5
5213 ; AVX512DQ-NEXT: vpand %xmm5, %xmm4, %xmm4
5214 ; AVX512DQ-NEXT: vpor %xmm3, %xmm2, %xmm2
5215 ; AVX512DQ-NEXT: vpand %xmm2, %xmm1, %xmm1
5216 ; AVX512DQ-NEXT: vpmovsxbd %xmm1, %zmm1
5217 ; AVX512DQ-NEXT: vpmovd2m %zmm1, %k0
5218 ; AVX512DQ-NEXT: kmovw %k0, %eax
5219 ; AVX512DQ-NEXT: vpmovsxbd %xmm12, %zmm1
5220 ; AVX512DQ-NEXT: vpmovd2m %zmm1, %k0
5221 ; AVX512DQ-NEXT: kmovw %k0, %ecx
5222 ; AVX512DQ-NEXT: shll $16, %ecx
5223 ; AVX512DQ-NEXT: orl %eax, %ecx
5224 ; AVX512DQ-NEXT: vpmovsxbd %xmm4, %zmm1
5225 ; AVX512DQ-NEXT: vpmovd2m %zmm1, %k0
5226 ; AVX512DQ-NEXT: kmovw %k0, %eax
5227 ; AVX512DQ-NEXT: vpmovsxbd %xmm0, %zmm0
5228 ; AVX512DQ-NEXT: vpmovd2m %zmm0, %k0
5229 ; AVX512DQ-NEXT: kmovw %k0, %edx
5230 ; AVX512DQ-NEXT: shll $16, %edx
5231 ; AVX512DQ-NEXT: orl %eax, %edx
5232 ; AVX512DQ-NEXT: shlq $32, %rdx
5233 ; AVX512DQ-NEXT: orq %rcx, %rdx
5234 ; AVX512DQ-NEXT: je LBB76_1
5235 ; AVX512DQ-NEXT: ## %bb.2: ## %exit
5236 ; AVX512DQ-NEXT: popq %rax
5237 ; AVX512DQ-NEXT: vzeroupper
5238 ; AVX512DQ-NEXT: retq
5239 ; AVX512DQ-NEXT: LBB76_1: ## %bar
5240 ; AVX512DQ-NEXT: vzeroupper
5241 ; AVX512DQ-NEXT: callq _foo
5242 ; AVX512DQ-NEXT: popq %rax
5243 ; AVX512DQ-NEXT: retq
5244 ;
5245 ; X86-LABEL: ktest_7:
5246 ; X86: ## %bb.0:
5247 ; X86-NEXT: subl $12, %esp
5248 ; X86-NEXT: .cfi_def_cfa_offset 16
5249 ; X86-NEXT: vptestnmb %zmm0, %zmm0, %k0
5250 ; X86-NEXT: vptestnmb %zmm1, %zmm1, %k1
5251 ; X86-NEXT: korq %k1, %k0, %k0
5252 ; X86-NEXT: vptestnmb %zmm2, %zmm2, %k1
5253 ; X86-NEXT: vptestnmb %zmm3, %zmm3, %k2
5254 ; X86-NEXT: korq %k2, %k1, %k1
5255 ; X86-NEXT: kandq %k1, %k0, %k0
5256 ; X86-NEXT: kshiftrq $32, %k0, %k1
5257 ; X86-NEXT: kortestd %k1, %k0
5258 ; X86-NEXT: je LBB76_1
5259 ; X86-NEXT: ## %bb.2: ## %exit
5260 ; X86-NEXT: addl $12, %esp
5261 ; X86-NEXT: vzeroupper
5262 ; X86-NEXT: retl
5263 ; X86-NEXT: LBB76_1: ## %bar
5264 ; X86-NEXT: vzeroupper
5265 ; X86-NEXT: calll _foo
5266 ; X86-NEXT: addl $12, %esp
5267 ; X86-NEXT: retl
5268 %a = icmp eq <64 x i8> %w, zeroinitializer
5269 %b = icmp eq <64 x i8> %x, zeroinitializer
5270 %c = icmp eq <64 x i8> %y, zeroinitializer
5271 %d = icmp eq <64 x i8> %z, zeroinitializer
5272 %e = or <64 x i1> %a, %b
5273 %f = or <64 x i1> %c, %d
5274 %g = and <64 x i1> %e, %f
5275 %h = bitcast <64 x i1> %g to i64
5276 %i = icmp eq i64 %h, 0
5277 br i1 %i, label %bar, label %exit
5279 bar:
5280 call void @foo()
5281 br label %exit
5283 exit:
5284 ret void
5285 }
5287 define <64 x i1> @mask64_insert(i32 %a) {
5288 ; KNL-LABEL: mask64_insert:
5289 ; KNL: ## %bb.0:
5290 ; KNL-NEXT: movq %rdi, %rax
5291 ; KNL-NEXT: movw $-4, %cx
5292 ; KNL-NEXT: kmovw %ecx, %k0
5293 ; KNL-NEXT: kshiftrw $1, %k0, %k0
5294 ; KNL-NEXT: kshiftlw $1, %k0, %k0
5295 ; KNL-NEXT: andl $1, %esi
5296 ; KNL-NEXT: kmovw %esi, %k1
5297 ; KNL-NEXT: korw %k1, %k0, %k0
5298 ; KNL-NEXT: kmovw %k0, (%rdi)
5299 ; KNL-NEXT: movw $-3, 6(%rdi)
5300 ; KNL-NEXT: movl $-131075, 2(%rdi) ## imm = 0xFFFDFFFD
5301 ; KNL-NEXT: retq
5302 ;
5303 ; SKX-LABEL: mask64_insert:
5304 ; SKX: ## %bb.0:
5305 ; SKX-NEXT: kmovd %edi, %k0
5306 ; SKX-NEXT: kshiftlq $63, %k0, %k0
5307 ; SKX-NEXT: kshiftrq $63, %k0, %k0
5308 ; SKX-NEXT: movabsq $-562958543486980, %rax ## imm = 0xFFFDFFFDFFFDFFFC
5309 ; SKX-NEXT: kmovq %rax, %k1
5310 ; SKX-NEXT: kshiftrq $1, %k1, %k1
5311 ; SKX-NEXT: kshiftlq $1, %k1, %k1
5312 ; SKX-NEXT: korq %k0, %k1, %k0
5313 ; SKX-NEXT: vpmovm2b %k0, %zmm0
5314 ; SKX-NEXT: retq
5315 ;
5316 ; AVX512BW-LABEL: mask64_insert:
5317 ; AVX512BW: ## %bb.0:
5318 ; AVX512BW-NEXT: kmovd %edi, %k0
5319 ; AVX512BW-NEXT: kshiftlq $63, %k0, %k0
5320 ; AVX512BW-NEXT: kshiftrq $63, %k0, %k0
5321 ; AVX512BW-NEXT: movabsq $-562958543486980, %rax ## imm = 0xFFFDFFFDFFFDFFFC
5322 ; AVX512BW-NEXT: kmovq %rax, %k1
5323 ; AVX512BW-NEXT: kshiftrq $1, %k1, %k1
5324 ; AVX512BW-NEXT: kshiftlq $1, %k1, %k1
5325 ; AVX512BW-NEXT: korq %k0, %k1, %k0
5326 ; AVX512BW-NEXT: vpmovm2b %k0, %zmm0
5327 ; AVX512BW-NEXT: retq
5328 ;
5329 ; AVX512DQ-LABEL: mask64_insert:
5330 ; AVX512DQ: ## %bb.0:
5331 ; AVX512DQ-NEXT: movq %rdi, %rax
5332 ; AVX512DQ-NEXT: movw $-4, %cx
5333 ; AVX512DQ-NEXT: kmovw %ecx, %k0
5334 ; AVX512DQ-NEXT: kshiftrw $1, %k0, %k0
5335 ; AVX512DQ-NEXT: kshiftlw $1, %k0, %k0
5336 ; AVX512DQ-NEXT: andl $1, %esi
5337 ; AVX512DQ-NEXT: kmovw %esi, %k1
5338 ; AVX512DQ-NEXT: korw %k1, %k0, %k0
5339 ; AVX512DQ-NEXT: kmovw %k0, (%rdi)
5340 ; AVX512DQ-NEXT: movw $-3, 6(%rdi)
5341 ; AVX512DQ-NEXT: movl $-131075, 2(%rdi) ## imm = 0xFFFDFFFD
5342 ; AVX512DQ-NEXT: retq
5343 ;
5344 ; X86-LABEL: mask64_insert:
5345 ; X86: ## %bb.0:
5346 ; X86-NEXT: kmovb {{[0-9]+}}(%esp), %k0
5347 ; X86-NEXT: movl $-131076, %eax ## imm = 0xFFFDFFFC
5348 ; X86-NEXT: kmovd %eax, %k1
5349 ; X86-NEXT: movl $-131075, %eax ## imm = 0xFFFDFFFD
5350 ; X86-NEXT: kmovd %eax, %k2
5351 ; X86-NEXT: kunpckdq %k1, %k2, %k1
5352 ; X86-NEXT: kshiftrq $1, %k1, %k1
5353 ; X86-NEXT: kshiftlq $1, %k1, %k1
5354 ; X86-NEXT: kshiftlq $63, %k0, %k0
5355 ; X86-NEXT: kshiftrq $63, %k0, %k0
5356 ; X86-NEXT: korq %k0, %k1, %k0
5357 ; X86-NEXT: vpmovm2b %k0, %zmm0
5358 ; X86-NEXT: retl
5359 %a_i = trunc i32 %a to i1
5360 %maskv = insertelement <64 x i1> <i1 true, i1 false, i1 true, i1 true, i1 true, i1 true, i1 true, i1 true, i1 true, i1 true, i1 true, i1 true, i1 true, i1 true, i1 true, i1 true, i1 true, i1 false, i1 true, i1 true, i1 true, i1 true, i1 true, i1 true, i1 true, i1 true, i1 true, i1 true, i1 true, i1 true, i1 true, i1 true, i1 true, i1 false, i1 true, i1 true, i1 true, i1 true, i1 true, i1 true, i1 true, i1 true, i1 true, i1 true, i1 true, i1 true, i1 true, i1 true, i1 true, i1 false, i1 true, i1 true, i1 true, i1 true, i1 true, i1 true, i1 true, i1 true, i1 true, i1 true, i1 true, i1 true, i1 true, i1 true>, i1 %a_i, i32 0
5361 ret <64 x i1> %maskv
5362 }