xref: /llvm-project/llvm/test/CodeGen/X86/avx512-insert-extract.ll (revision e088249b74586590c9e143d85b97a175acc9465e) 
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
2; RUN: llc < %s -mtriple=x86_64-apple-darwin -mattr=+avx512f | FileCheck --check-prefixes=CHECK,KNL %s
3; RUN: llc < %s -mtriple=x86_64-apple-darwin -mattr=+avx512f,+avx512bw,+avx512vl,+avx512dq | FileCheck --check-prefixes=CHECK,SKX %s
4; RUN: llc < %s -mtriple=x86_64-apple-darwin -mattr=+avx512f,+avx512bw,+avx512vl,+avx512dq,+avx512vbmi | FileCheck --check-prefixes=CHECK,SKX %s
5
6define <16 x float> @test1(<16 x float> %x, ptr %br, float %y) nounwind {
7; KNL-LABEL: test1:
8; KNL:       ## %bb.0:
9; KNL-NEXT:    vinsertps {{.*#+}} xmm2 = xmm0[0],mem[0],xmm0[2,3]
10; KNL-NEXT:    vinsertf32x4 $0, %xmm2, %zmm0, %zmm0
11; KNL-NEXT:    movw $16384, %ax ## imm = 0x4000
12; KNL-NEXT:    kmovw %eax, %k1
13; KNL-NEXT:    vbroadcastss %xmm1, %zmm0 {%k1}
14; KNL-NEXT:    retq
15;
16; SKX-LABEL: test1:
17; SKX:       ## %bb.0:
18; SKX-NEXT:    vinsertps {{.*#+}} xmm2 = xmm0[0],mem[0],xmm0[2,3]
19; SKX-NEXT:    vinsertf32x4 $0, %xmm2, %zmm0, %zmm0
20; SKX-NEXT:    movw $16384, %ax ## imm = 0x4000
21; SKX-NEXT:    kmovd %eax, %k1
22; SKX-NEXT:    vbroadcastss %xmm1, %zmm0 {%k1}
23; SKX-NEXT:    retq
24  %rrr = load float, ptr %br
25  %rrr2 = insertelement <16 x float> %x, float %rrr, i32 1
26  %rrr3 = insertelement <16 x float> %rrr2, float %y, i32 14
27  ret <16 x float> %rrr3
28}
29
30define <8 x double> @test2(<8 x double> %x, ptr %br, double %y) nounwind {
31; KNL-LABEL: test2:
32; KNL:       ## %bb.0:
33; KNL-NEXT:    vmovhpd {{.*#+}} xmm2 = xmm0[0],mem[0]
34; KNL-NEXT:    vinsertf32x4 $0, %xmm2, %zmm0, %zmm0
35; KNL-NEXT:    movb $64, %al
36; KNL-NEXT:    kmovw %eax, %k1
37; KNL-NEXT:    vbroadcastsd %xmm1, %zmm0 {%k1}
38; KNL-NEXT:    retq
39;
40; SKX-LABEL: test2:
41; SKX:       ## %bb.0:
42; SKX-NEXT:    vmovhpd {{.*#+}} xmm2 = xmm0[0],mem[0]
43; SKX-NEXT:    vinsertf32x4 $0, %xmm2, %zmm0, %zmm0
44; SKX-NEXT:    movb $64, %al
45; SKX-NEXT:    kmovd %eax, %k1
46; SKX-NEXT:    vbroadcastsd %xmm1, %zmm0 {%k1}
47; SKX-NEXT:    retq
48  %rrr = load double, ptr %br
49  %rrr2 = insertelement <8 x double> %x, double %rrr, i32 1
50  %rrr3 = insertelement <8 x double> %rrr2, double %y, i32 6
51  ret <8 x double> %rrr3
52}
53
54define <16 x float> @test3(<16 x float> %x) nounwind {
55; CHECK-LABEL: test3:
56; CHECK:       ## %bb.0:
57; CHECK-NEXT:    vextractf128 $1, %ymm0, %xmm1
58; CHECK-NEXT:    vinsertps {{.*#+}} xmm1 = xmm0[0],xmm1[0],xmm0[2,3]
59; CHECK-NEXT:    vinsertf32x4 $0, %xmm1, %zmm0, %zmm0
60; CHECK-NEXT:    retq
61  %eee = extractelement <16 x float> %x, i32 4
62  %rrr2 = insertelement <16 x float> %x, float %eee, i32 1
63  ret <16 x float> %rrr2
64}
65
66define <8 x i64> @test4(<8 x i64> %x) nounwind {
67; CHECK-LABEL: test4:
68; CHECK:       ## %bb.0:
69; CHECK-NEXT:    vextractf32x4 $2, %zmm0, %xmm1
70; CHECK-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm0[0],xmm1[0]
71; CHECK-NEXT:    vinsertf32x4 $0, %xmm1, %zmm0, %zmm0
72; CHECK-NEXT:    retq
73  %eee = extractelement <8 x i64> %x, i32 4
74  %rrr2 = insertelement <8 x i64> %x, i64 %eee, i32 1
75  ret <8 x i64> %rrr2
76}
77
78define i32 @test5(<4 x float> %x) nounwind {
79; CHECK-LABEL: test5:
80; CHECK:       ## %bb.0:
81; CHECK-NEXT:    vextractps $3, %xmm0, %eax
82; CHECK-NEXT:    retq
83  %ef = extractelement <4 x float> %x, i32 3
84  %ei = bitcast float %ef to i32
85  ret i32 %ei
86}
87
88define void @test6(<4 x float> %x, ptr %out) nounwind {
89; CHECK-LABEL: test6:
90; CHECK:       ## %bb.0:
91; CHECK-NEXT:    vextractps $3, %xmm0, (%rdi)
92; CHECK-NEXT:    retq
93  %ef = extractelement <4 x float> %x, i32 3
94  store float %ef, ptr %out, align 4
95  ret void
96}
97
98define float @test7(<16 x float> %x, i32 %ind) nounwind {
99; CHECK-LABEL: test7:
100; CHECK:       ## %bb.0:
101; CHECK-NEXT:    pushq %rbp
102; CHECK-NEXT:    movq %rsp, %rbp
103; CHECK-NEXT:    andq $-64, %rsp
104; CHECK-NEXT:    subq $128, %rsp
105; CHECK-NEXT:    ## kill: def $edi killed $edi def $rdi
106; CHECK-NEXT:    vmovaps %zmm0, (%rsp)
107; CHECK-NEXT:    andl $15, %edi
108; CHECK-NEXT:    vmovss {{.*#+}} xmm0 = mem[0],zero,zero,zero
109; CHECK-NEXT:    movq %rbp, %rsp
110; CHECK-NEXT:    popq %rbp
111; CHECK-NEXT:    vzeroupper
112; CHECK-NEXT:    retq
113  %e = extractelement <16 x float> %x, i32 %ind
114  ret float %e
115}
116
117define double @test8(<8 x double> %x, i32 %ind) nounwind {
118; CHECK-LABEL: test8:
119; CHECK:       ## %bb.0:
120; CHECK-NEXT:    pushq %rbp
121; CHECK-NEXT:    movq %rsp, %rbp
122; CHECK-NEXT:    andq $-64, %rsp
123; CHECK-NEXT:    subq $128, %rsp
124; CHECK-NEXT:    ## kill: def $edi killed $edi def $rdi
125; CHECK-NEXT:    vmovaps %zmm0, (%rsp)
126; CHECK-NEXT:    andl $7, %edi
127; CHECK-NEXT:    vmovsd {{.*#+}} xmm0 = mem[0],zero
128; CHECK-NEXT:    movq %rbp, %rsp
129; CHECK-NEXT:    popq %rbp
130; CHECK-NEXT:    vzeroupper
131; CHECK-NEXT:    retq
132  %e = extractelement <8 x double> %x, i32 %ind
133  ret double %e
134}
135
136define float @test9(<8 x float> %x, i32 %ind) nounwind {
137; CHECK-LABEL: test9:
138; CHECK:       ## %bb.0:
139; CHECK-NEXT:    pushq %rbp
140; CHECK-NEXT:    movq %rsp, %rbp
141; CHECK-NEXT:    andq $-32, %rsp
142; CHECK-NEXT:    subq $64, %rsp
143; CHECK-NEXT:    ## kill: def $edi killed $edi def $rdi
144; CHECK-NEXT:    vmovaps %ymm0, (%rsp)
145; CHECK-NEXT:    andl $7, %edi
146; CHECK-NEXT:    vmovss {{.*#+}} xmm0 = mem[0],zero,zero,zero
147; CHECK-NEXT:    movq %rbp, %rsp
148; CHECK-NEXT:    popq %rbp
149; CHECK-NEXT:    vzeroupper
150; CHECK-NEXT:    retq
151  %e = extractelement <8 x float> %x, i32 %ind
152  ret float %e
153}
154
155define i32 @test10(<16 x i32> %x, i32 %ind) nounwind {
156; CHECK-LABEL: test10:
157; CHECK:       ## %bb.0:
158; CHECK-NEXT:    pushq %rbp
159; CHECK-NEXT:    movq %rsp, %rbp
160; CHECK-NEXT:    andq $-64, %rsp
161; CHECK-NEXT:    subq $128, %rsp
162; CHECK-NEXT:    ## kill: def $edi killed $edi def $rdi
163; CHECK-NEXT:    vmovaps %zmm0, (%rsp)
164; CHECK-NEXT:    andl $15, %edi
165; CHECK-NEXT:    movl (%rsp,%rdi,4), %eax
166; CHECK-NEXT:    movq %rbp, %rsp
167; CHECK-NEXT:    popq %rbp
168; CHECK-NEXT:    vzeroupper
169; CHECK-NEXT:    retq
170  %e = extractelement <16 x i32> %x, i32 %ind
171  ret i32 %e
172}
173
174define <16 x i32> @test11(<16 x i32>%a, <16 x i32>%b) nounwind {
175; KNL-LABEL: test11:
176; KNL:       ## %bb.0:
177; KNL-NEXT:    vpcmpltud %zmm1, %zmm0, %k0
178; KNL-NEXT:    kmovw %k0, %eax
179; KNL-NEXT:    testb $16, %al
180; KNL-NEXT:    je LBB10_2
181; KNL-NEXT:  ## %bb.1: ## %A
182; KNL-NEXT:    vmovdqa64 %zmm1, %zmm0
183; KNL-NEXT:    retq
184; KNL-NEXT:  LBB10_2: ## %B
185; KNL-NEXT:    vpaddd %zmm0, %zmm1, %zmm0
186; KNL-NEXT:    retq
187;
188; SKX-LABEL: test11:
189; SKX:       ## %bb.0:
190; SKX-NEXT:    vpcmpltud %zmm1, %zmm0, %k0
191; SKX-NEXT:    kmovd %k0, %eax
192; SKX-NEXT:    testb $16, %al
193; SKX-NEXT:    je LBB10_2
194; SKX-NEXT:  ## %bb.1: ## %A
195; SKX-NEXT:    vmovdqa64 %zmm1, %zmm0
196; SKX-NEXT:    retq
197; SKX-NEXT:  LBB10_2: ## %B
198; SKX-NEXT:    vpaddd %zmm0, %zmm1, %zmm0
199; SKX-NEXT:    retq
200  %cmp_res = icmp ult <16 x i32> %a, %b
201  %ia = extractelement <16 x i1> %cmp_res, i32 4
202  br i1 %ia, label %A, label %B
203  A:
204    ret <16 x i32>%b
205  B:
206   %c = add <16 x i32>%b, %a
207   ret <16 x i32>%c
208}
209
210define i64 @test12(<16 x i64>%a, <16 x i64>%b, i64 %a1, i64 %b1) nounwind {
211; KNL-LABEL: test12:
212; KNL:       ## %bb.0:
213; KNL-NEXT:    movq %rdi, %rax
214; KNL-NEXT:    vpcmpgtq %zmm0, %zmm2, %k0
215; KNL-NEXT:    kmovw %k0, %ecx
216; KNL-NEXT:    testb $1, %cl
217; KNL-NEXT:    cmoveq %rsi, %rax
218; KNL-NEXT:    vzeroupper
219; KNL-NEXT:    retq
220;
221; SKX-LABEL: test12:
222; SKX:       ## %bb.0:
223; SKX-NEXT:    movq %rdi, %rax
224; SKX-NEXT:    vpcmpgtq %zmm0, %zmm2, %k0
225; SKX-NEXT:    kmovd %k0, %ecx
226; SKX-NEXT:    testb $1, %cl
227; SKX-NEXT:    cmoveq %rsi, %rax
228; SKX-NEXT:    vzeroupper
229; SKX-NEXT:    retq
230  %cmpvector_func.i = icmp slt <16 x i64> %a, %b
231  %extract24vector_func.i = extractelement <16 x i1> %cmpvector_func.i, i32 0
232  %res = select i1 %extract24vector_func.i, i64 %a1, i64 %b1
233  ret i64 %res
234}
235
236define i16 @test13(i32 %a, i32 %b) nounwind {
237; KNL-LABEL: test13:
238; KNL:       ## %bb.0:
239; KNL-NEXT:    cmpl %esi, %edi
240; KNL-NEXT:    setb %al
241; KNL-NEXT:    movw $-4, %cx
242; KNL-NEXT:    kmovw %ecx, %k0
243; KNL-NEXT:    kshiftrw $1, %k0, %k0
244; KNL-NEXT:    kshiftlw $1, %k0, %k0
245; KNL-NEXT:    andl $1, %eax
246; KNL-NEXT:    kmovw %eax, %k1
247; KNL-NEXT:    korw %k1, %k0, %k0
248; KNL-NEXT:    kmovw %k0, %eax
249; KNL-NEXT:    ## kill: def $ax killed $ax killed $eax
250; KNL-NEXT:    retq
251;
252; SKX-LABEL: test13:
253; SKX:       ## %bb.0:
254; SKX-NEXT:    cmpl %esi, %edi
255; SKX-NEXT:    setb %al
256; SKX-NEXT:    movw $-4, %cx
257; SKX-NEXT:    kmovd %ecx, %k0
258; SKX-NEXT:    kshiftrw $1, %k0, %k0
259; SKX-NEXT:    kshiftlw $1, %k0, %k0
260; SKX-NEXT:    andl $1, %eax
261; SKX-NEXT:    kmovw %eax, %k1
262; SKX-NEXT:    korw %k1, %k0, %k0
263; SKX-NEXT:    kmovd %k0, %eax
264; SKX-NEXT:    ## kill: def $ax killed $ax killed $eax
265; SKX-NEXT:    retq
266  %cmp_res = icmp ult i32 %a, %b
267  %maskv = insertelement <16 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 %cmp_res, i32 0
268  %res = bitcast <16 x i1> %maskv to i16
269  ret i16 %res
270}
271
272define i64 @test14(<8 x i64>%a, <8 x i64>%b, i64 %a1, i64 %b1) nounwind {
273; KNL-LABEL: test14:
274; KNL:       ## %bb.0:
275; KNL-NEXT:    movq %rdi, %rax
276; KNL-NEXT:    vpcmpgtq %zmm0, %zmm1, %k0
277; KNL-NEXT:    kmovw %k0, %ecx
278; KNL-NEXT:    testb $16, %cl
279; KNL-NEXT:    cmoveq %rsi, %rax
280; KNL-NEXT:    vzeroupper
281; KNL-NEXT:    retq
282;
283; SKX-LABEL: test14:
284; SKX:       ## %bb.0:
285; SKX-NEXT:    movq %rdi, %rax
286; SKX-NEXT:    vpcmpgtq %zmm0, %zmm1, %k0
287; SKX-NEXT:    kmovd %k0, %ecx
288; SKX-NEXT:    testb $16, %cl
289; SKX-NEXT:    cmoveq %rsi, %rax
290; SKX-NEXT:    vzeroupper
291; SKX-NEXT:    retq
292  %cmpvector_func.i = icmp slt <8 x i64> %a, %b
293  %extract24vector_func.i = extractelement <8 x i1> %cmpvector_func.i, i32 4
294  %res = select i1 %extract24vector_func.i, i64 %a1, i64 %b1
295  ret i64 %res
296}
297
298define i16 @test15(ptr%addr) nounwind {
299; CHECK-LABEL: test15:
300; CHECK:       ## %bb.0:
301; CHECK-NEXT:    movzbl (%rdi), %eax
302; CHECK-NEXT:    negl %eax
303; CHECK-NEXT:    ## kill: def $ax killed $ax killed $eax
304; CHECK-NEXT:    retq
305  %x = load i1 , ptr %addr, align 1
306  %x1 = insertelement <16 x i1> undef, i1 %x, i32 10
307  %x2 = bitcast <16 x i1>%x1 to i16
308  ret i16 %x2
309}
310
311define i16 @test16(ptr%addr, i16 %a) nounwind {
312; KNL-LABEL: test16:
313; KNL:       ## %bb.0:
314; KNL-NEXT:    movzbl (%rdi), %eax
315; KNL-NEXT:    kmovw %esi, %k0
316; KNL-NEXT:    movw $-1025, %cx ## imm = 0xFBFF
317; KNL-NEXT:    kmovw %ecx, %k1
318; KNL-NEXT:    kandw %k1, %k0, %k0
319; KNL-NEXT:    kmovw %eax, %k1
320; KNL-NEXT:    kshiftlw $15, %k1, %k1
321; KNL-NEXT:    kshiftrw $5, %k1, %k1
322; KNL-NEXT:    korw %k1, %k0, %k0
323; KNL-NEXT:    kmovw %k0, %eax
324; KNL-NEXT:    ## kill: def $ax killed $ax killed $eax
325; KNL-NEXT:    retq
326;
327; SKX-LABEL: test16:
328; SKX:       ## %bb.0:
329; SKX-NEXT:    kmovb (%rdi), %k0
330; SKX-NEXT:    kmovd %esi, %k1
331; SKX-NEXT:    movw $-1025, %ax ## imm = 0xFBFF
332; SKX-NEXT:    kmovd %eax, %k2
333; SKX-NEXT:    kandw %k2, %k1, %k1
334; SKX-NEXT:    kshiftlw $15, %k0, %k0
335; SKX-NEXT:    kshiftrw $5, %k0, %k0
336; SKX-NEXT:    korw %k0, %k1, %k0
337; SKX-NEXT:    kmovd %k0, %eax
338; SKX-NEXT:    ## kill: def $ax killed $ax killed $eax
339; SKX-NEXT:    retq
340  %x = load i1 , ptr %addr, align 128
341  %a1 = bitcast i16 %a to <16 x i1>
342  %x1 = insertelement <16 x i1> %a1, i1 %x, i32 10
343  %x2 = bitcast <16 x i1>%x1 to i16
344  ret i16 %x2
345}
346
347define i8 @test17(ptr%addr, i8 %a) nounwind {
348; KNL-LABEL: test17:
349; KNL:       ## %bb.0:
350; KNL-NEXT:    movzbl (%rdi), %eax
351; KNL-NEXT:    kmovw %esi, %k0
352; KNL-NEXT:    movw $-17, %cx
353; KNL-NEXT:    kmovw %ecx, %k1
354; KNL-NEXT:    kandw %k1, %k0, %k0
355; KNL-NEXT:    kmovw %eax, %k1
356; KNL-NEXT:    kshiftlw $15, %k1, %k1
357; KNL-NEXT:    kshiftrw $11, %k1, %k1
358; KNL-NEXT:    korw %k1, %k0, %k0
359; KNL-NEXT:    kmovw %k0, %eax
360; KNL-NEXT:    ## kill: def $al killed $al killed $eax
361; KNL-NEXT:    retq
362;
363; SKX-LABEL: test17:
364; SKX:       ## %bb.0:
365; SKX-NEXT:    kmovb (%rdi), %k0
366; SKX-NEXT:    kmovd %esi, %k1
367; SKX-NEXT:    movb $-17, %al
368; SKX-NEXT:    kmovd %eax, %k2
369; SKX-NEXT:    kandb %k2, %k1, %k1
370; SKX-NEXT:    kshiftlb $7, %k0, %k0
371; SKX-NEXT:    kshiftrb $3, %k0, %k0
372; SKX-NEXT:    korb %k0, %k1, %k0
373; SKX-NEXT:    kmovd %k0, %eax
374; SKX-NEXT:    ## kill: def $al killed $al killed $eax
375; SKX-NEXT:    retq
376  %x = load i1 , ptr %addr, align 128
377  %a1 = bitcast i8 %a to <8 x i1>
378  %x1 = insertelement <8 x i1> %a1, i1 %x, i32 4
379  %x2 = bitcast <8 x i1>%x1 to i8
380  ret i8 %x2
381}
382
383define i64 @extract_v8i64(<8 x i64> %x, ptr %dst) nounwind {
384; CHECK-LABEL: extract_v8i64:
385; CHECK:       ## %bb.0:
386; CHECK-NEXT:    vpextrq $1, %xmm0, %rax
387; CHECK-NEXT:    vextracti128 $1, %ymm0, %xmm0
388; CHECK-NEXT:    vpextrq $1, %xmm0, (%rdi)
389; CHECK-NEXT:    vzeroupper
390; CHECK-NEXT:    retq
391  %r1 = extractelement <8 x i64> %x, i32 1
392  %r2 = extractelement <8 x i64> %x, i32 3
393  store i64 %r2, ptr %dst, align 1
394  ret i64 %r1
395}
396
397define i64 @extract_v4i64(<4 x i64> %x, ptr %dst) nounwind {
398; CHECK-LABEL: extract_v4i64:
399; CHECK:       ## %bb.0:
400; CHECK-NEXT:    vpextrq $1, %xmm0, %rax
401; CHECK-NEXT:    vextracti128 $1, %ymm0, %xmm0
402; CHECK-NEXT:    vpextrq $1, %xmm0, (%rdi)
403; CHECK-NEXT:    vzeroupper
404; CHECK-NEXT:    retq
405  %r1 = extractelement <4 x i64> %x, i32 1
406  %r2 = extractelement <4 x i64> %x, i32 3
407  store i64 %r2, ptr %dst, align 1
408  ret i64 %r1
409}
410
411define i64 @extract_v2i64(<2 x i64> %x, ptr %dst) nounwind {
412; CHECK-LABEL: extract_v2i64:
413; CHECK:       ## %bb.0:
414; CHECK-NEXT:    vmovq %xmm0, %rax
415; CHECK-NEXT:    vpextrq $1, %xmm0, (%rdi)
416; CHECK-NEXT:    retq
417  %r1 = extractelement <2 x i64> %x, i32 0
418  %r2 = extractelement <2 x i64> %x, i32 1
419  store i64 %r2, ptr %dst, align 1
420  ret i64 %r1
421}
422
423define i32 @extract_v16i32(<16 x i32> %x, ptr %dst) nounwind {
424; CHECK-LABEL: extract_v16i32:
425; CHECK:       ## %bb.0:
426; CHECK-NEXT:    vextractps $1, %xmm0, %eax
427; CHECK-NEXT:    vextractf128 $1, %ymm0, %xmm0
428; CHECK-NEXT:    vextractps $1, %xmm0, (%rdi)
429; CHECK-NEXT:    vzeroupper
430; CHECK-NEXT:    retq
431  %r1 = extractelement <16 x i32> %x, i32 1
432  %r2 = extractelement <16 x i32> %x, i32 5
433  store i32 %r2, ptr %dst, align 1
434  ret i32 %r1
435}
436
437define i32 @extract_v8i32(<8 x i32> %x, ptr %dst) nounwind {
438; CHECK-LABEL: extract_v8i32:
439; CHECK:       ## %bb.0:
440; CHECK-NEXT:    vextractps $1, %xmm0, %eax
441; CHECK-NEXT:    vextractf128 $1, %ymm0, %xmm0
442; CHECK-NEXT:    vextractps $1, %xmm0, (%rdi)
443; CHECK-NEXT:    vzeroupper
444; CHECK-NEXT:    retq
445  %r1 = extractelement <8 x i32> %x, i32 1
446  %r2 = extractelement <8 x i32> %x, i32 5
447  store i32 %r2, ptr %dst, align 1
448  ret i32 %r1
449}
450
451define i32 @extract_v4i32(<4 x i32> %x, ptr %dst) nounwind {
452; CHECK-LABEL: extract_v4i32:
453; CHECK:       ## %bb.0:
454; CHECK-NEXT:    vextractps $1, %xmm0, %eax
455; CHECK-NEXT:    vextractps $3, %xmm0, (%rdi)
456; CHECK-NEXT:    retq
457  %r1 = extractelement <4 x i32> %x, i32 1
458  %r2 = extractelement <4 x i32> %x, i32 3
459  store i32 %r2, ptr %dst, align 1
460  ret i32 %r1
461}
462
463define i16 @extract_v32i16(<32 x i16> %x, ptr %dst) nounwind {
464; CHECK-LABEL: extract_v32i16:
465; CHECK:       ## %bb.0:
466; CHECK-NEXT:    vpextrw $1, %xmm0, %eax
467; CHECK-NEXT:    vextracti128 $1, %ymm0, %xmm0
468; CHECK-NEXT:    vpextrw $1, %xmm0, (%rdi)
469; CHECK-NEXT:    ## kill: def $ax killed $ax killed $eax
470; CHECK-NEXT:    vzeroupper
471; CHECK-NEXT:    retq
472  %r1 = extractelement <32 x i16> %x, i32 1
473  %r2 = extractelement <32 x i16> %x, i32 9
474  store i16 %r2, ptr %dst, align 1
475  ret i16 %r1
476}
477
478define i16 @extract_v16i16(<16 x i16> %x, ptr %dst) nounwind {
479; CHECK-LABEL: extract_v16i16:
480; CHECK:       ## %bb.0:
481; CHECK-NEXT:    vpextrw $1, %xmm0, %eax
482; CHECK-NEXT:    vextracti128 $1, %ymm0, %xmm0
483; CHECK-NEXT:    vpextrw $1, %xmm0, (%rdi)
484; CHECK-NEXT:    ## kill: def $ax killed $ax killed $eax
485; CHECK-NEXT:    vzeroupper
486; CHECK-NEXT:    retq
487  %r1 = extractelement <16 x i16> %x, i32 1
488  %r2 = extractelement <16 x i16> %x, i32 9
489  store i16 %r2, ptr %dst, align 1
490  ret i16 %r1
491}
492
493define i16 @extract_v8i16(<8 x i16> %x, ptr %dst) nounwind {
494; CHECK-LABEL: extract_v8i16:
495; CHECK:       ## %bb.0:
496; CHECK-NEXT:    vpextrw $1, %xmm0, %eax
497; CHECK-NEXT:    vpextrw $3, %xmm0, (%rdi)
498; CHECK-NEXT:    ## kill: def $ax killed $ax killed $eax
499; CHECK-NEXT:    retq
500  %r1 = extractelement <8 x i16> %x, i32 1
501  %r2 = extractelement <8 x i16> %x, i32 3
502  store i16 %r2, ptr %dst, align 1
503  ret i16 %r1
504}
505
506define i8 @extract_v64i8(<64 x i8> %x, ptr %dst) nounwind {
507; CHECK-LABEL: extract_v64i8:
508; CHECK:       ## %bb.0:
509; CHECK-NEXT:    vpextrb $1, %xmm0, %eax
510; CHECK-NEXT:    vextracti128 $1, %ymm0, %xmm0
511; CHECK-NEXT:    vpextrb $1, %xmm0, (%rdi)
512; CHECK-NEXT:    ## kill: def $al killed $al killed $eax
513; CHECK-NEXT:    vzeroupper
514; CHECK-NEXT:    retq
515  %r1 = extractelement <64 x i8> %x, i32 1
516  %r2 = extractelement <64 x i8> %x, i32 17
517  store i8 %r2, ptr %dst, align 1
518  ret i8 %r1
519}
520
521define i8 @extract_v32i8(<32 x i8> %x, ptr %dst) nounwind {
522; CHECK-LABEL: extract_v32i8:
523; CHECK:       ## %bb.0:
524; CHECK-NEXT:    vpextrb $1, %xmm0, %eax
525; CHECK-NEXT:    vextracti128 $1, %ymm0, %xmm0
526; CHECK-NEXT:    vpextrb $1, %xmm0, (%rdi)
527; CHECK-NEXT:    ## kill: def $al killed $al killed $eax
528; CHECK-NEXT:    vzeroupper
529; CHECK-NEXT:    retq
530  %r1 = extractelement <32 x i8> %x, i32 1
531  %r2 = extractelement <32 x i8> %x, i32 17
532  store i8 %r2, ptr %dst, align 1
533  ret i8 %r1
534}
535
536define i8 @extract_v16i8(<16 x i8> %x, ptr %dst) nounwind {
537; CHECK-LABEL: extract_v16i8:
538; CHECK:       ## %bb.0:
539; CHECK-NEXT:    vpextrb $1, %xmm0, %eax
540; CHECK-NEXT:    vpextrb $3, %xmm0, (%rdi)
541; CHECK-NEXT:    ## kill: def $al killed $al killed $eax
542; CHECK-NEXT:    retq
543  %r1 = extractelement <16 x i8> %x, i32 1
544  %r2 = extractelement <16 x i8> %x, i32 3
545  store i8 %r2, ptr %dst, align 1
546  ret i8 %r1
547}
548
549define <8 x i64> @insert_v8i64(<8 x i64> %x, i64 %y , ptr %ptr) nounwind {
550; KNL-LABEL: insert_v8i64:
551; KNL:       ## %bb.0:
552; KNL-NEXT:    vpinsrq $1, (%rsi), %xmm0, %xmm1
553; KNL-NEXT:    vinserti32x4 $0, %xmm1, %zmm0, %zmm0
554; KNL-NEXT:    movb $8, %al
555; KNL-NEXT:    kmovw %eax, %k1
556; KNL-NEXT:    vpbroadcastq %rdi, %zmm0 {%k1}
557; KNL-NEXT:    retq
558;
559; SKX-LABEL: insert_v8i64:
560; SKX:       ## %bb.0:
561; SKX-NEXT:    vpinsrq $1, (%rsi), %xmm0, %xmm1
562; SKX-NEXT:    vinserti32x4 $0, %xmm1, %zmm0, %zmm0
563; SKX-NEXT:    movb $8, %al
564; SKX-NEXT:    kmovd %eax, %k1
565; SKX-NEXT:    vpbroadcastq %rdi, %zmm0 {%k1}
566; SKX-NEXT:    retq
567  %val = load i64, ptr %ptr
568  %r1 = insertelement <8 x i64> %x, i64 %val, i32 1
569  %r2 = insertelement <8 x i64> %r1, i64 %y, i32 3
570  ret <8 x i64> %r2
571}
572
573define <4 x i64> @insert_v4i64(<4 x i64> %x, i64 %y , ptr %ptr) nounwind {
574; KNL-LABEL: insert_v4i64:
575; KNL:       ## %bb.0:
576; KNL-NEXT:    vpinsrq $1, (%rsi), %xmm0, %xmm1
577; KNL-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
578; KNL-NEXT:    vmovq %rdi, %xmm1
579; KNL-NEXT:    vpbroadcastq %xmm1, %ymm1
580; KNL-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
581; KNL-NEXT:    retq
582;
583; SKX-LABEL: insert_v4i64:
584; SKX:       ## %bb.0:
585; SKX-NEXT:    vpinsrq $1, (%rsi), %xmm0, %xmm1
586; SKX-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
587; SKX-NEXT:    vpbroadcastq %rdi, %ymm1
588; SKX-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5],ymm1[6,7]
589; SKX-NEXT:    retq
590  %val = load i64, ptr %ptr
591  %r1 = insertelement <4 x i64> %x, i64 %val, i32 1
592  %r2 = insertelement <4 x i64> %r1, i64 %y, i32 3
593  ret <4 x i64> %r2
594}
595
596define <2 x i64> @insert_v2i64(<2 x i64> %x, i64 %y , ptr %ptr) nounwind {
597; CHECK-LABEL: insert_v2i64:
598; CHECK:       ## %bb.0:
599; CHECK-NEXT:    vmovq {{.*#+}} xmm0 = mem[0],zero
600; CHECK-NEXT:    vmovq %rdi, %xmm1
601; CHECK-NEXT:    vpunpcklqdq {{.*#+}} xmm0 = xmm1[0],xmm0[0]
602; CHECK-NEXT:    retq
603  %val = load i64, ptr %ptr
604  %r1 = insertelement <2 x i64> %x, i64 %val, i32 1
605  %r2 = insertelement <2 x i64> %r1, i64 %y, i32 0
606  ret <2 x i64> %r2
607}
608
609define <16 x i32> @insert_v16i32(<16 x i32> %x, i32 %y, ptr %ptr) nounwind {
610; KNL-LABEL: insert_v16i32:
611; KNL:       ## %bb.0:
612; KNL-NEXT:    vpinsrd $1, (%rsi), %xmm0, %xmm1
613; KNL-NEXT:    vinserti32x4 $0, %xmm1, %zmm0, %zmm0
614; KNL-NEXT:    movw $32, %ax
615; KNL-NEXT:    kmovw %eax, %k1
616; KNL-NEXT:    vpbroadcastd %edi, %zmm0 {%k1}
617; KNL-NEXT:    retq
618;
619; SKX-LABEL: insert_v16i32:
620; SKX:       ## %bb.0:
621; SKX-NEXT:    vpinsrd $1, (%rsi), %xmm0, %xmm1
622; SKX-NEXT:    vinserti32x4 $0, %xmm1, %zmm0, %zmm0
623; SKX-NEXT:    movw $32, %ax
624; SKX-NEXT:    kmovd %eax, %k1
625; SKX-NEXT:    vpbroadcastd %edi, %zmm0 {%k1}
626; SKX-NEXT:    retq
627  %val = load i32, ptr %ptr
628  %r1 = insertelement <16 x i32> %x, i32 %val, i32 1
629  %r2 = insertelement <16 x i32> %r1, i32 %y, i32 5
630  ret <16 x i32> %r2
631}
632
633define <8 x i32> @insert_v8i32(<8 x i32> %x, i32 %y, ptr %ptr) nounwind {
634; KNL-LABEL: insert_v8i32:
635; KNL:       ## %bb.0:
636; KNL-NEXT:    vpinsrd $1, (%rsi), %xmm0, %xmm1
637; KNL-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
638; KNL-NEXT:    vmovd %edi, %xmm1
639; KNL-NEXT:    vpbroadcastd %xmm1, %ymm1
640; KNL-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm1[5],ymm0[6,7]
641; KNL-NEXT:    retq
642;
643; SKX-LABEL: insert_v8i32:
644; SKX:       ## %bb.0:
645; SKX-NEXT:    vpinsrd $1, (%rsi), %xmm0, %xmm1
646; SKX-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
647; SKX-NEXT:    vpbroadcastd %edi, %ymm1
648; SKX-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3,4],ymm1[5],ymm0[6,7]
649; SKX-NEXT:    retq
650  %val = load i32, ptr %ptr
651  %r1 = insertelement <8 x i32> %x, i32 %val, i32 1
652  %r2 = insertelement <8 x i32> %r1, i32 %y, i32 5
653  ret <8 x i32> %r2
654}
655
656define <4 x i32> @insert_v4i32(<4 x i32> %x, i32 %y, ptr %ptr) nounwind {
657; CHECK-LABEL: insert_v4i32:
658; CHECK:       ## %bb.0:
659; CHECK-NEXT:    vpinsrd $1, (%rsi), %xmm0, %xmm0
660; CHECK-NEXT:    vpinsrd $3, %edi, %xmm0, %xmm0
661; CHECK-NEXT:    retq
662  %val = load i32, ptr %ptr
663  %r1 = insertelement <4 x i32> %x, i32 %val, i32 1
664  %r2 = insertelement <4 x i32> %r1, i32 %y, i32 3
665  ret <4 x i32> %r2
666}
667
668define <32 x i16> @insert_v32i16(<32 x i16> %x, i16 %y, ptr %ptr) nounwind {
669; KNL-LABEL: insert_v32i16:
670; KNL:       ## %bb.0:
671; KNL-NEXT:    vpinsrw $1, (%rsi), %xmm0, %xmm1
672; KNL-NEXT:    vinserti32x4 $0, %xmm1, %zmm0, %zmm1
673; KNL-NEXT:    vmovd %edi, %xmm0
674; KNL-NEXT:    vpbroadcastw %xmm0, %ymm0
675; KNL-NEXT:    vpternlogq {{.*#+}} zmm0 = zmm0 ^ (mem & (zmm0 ^ zmm1))
676; KNL-NEXT:    retq
677;
678; SKX-LABEL: insert_v32i16:
679; SKX:       ## %bb.0:
680; SKX-NEXT:    vpinsrw $1, (%rsi), %xmm0, %xmm1
681; SKX-NEXT:    vinserti32x4 $0, %xmm1, %zmm0, %zmm0
682; SKX-NEXT:    movl $512, %eax ## imm = 0x200
683; SKX-NEXT:    kmovd %eax, %k1
684; SKX-NEXT:    vpbroadcastw %edi, %zmm0 {%k1}
685; SKX-NEXT:    retq
686  %val = load i16, ptr %ptr
687  %r1 = insertelement <32 x i16> %x, i16 %val, i32 1
688  %r2 = insertelement <32 x i16> %r1, i16 %y, i32 9
689  ret <32 x i16> %r2
690}
691
692define <16 x i16> @insert_v16i16(<16 x i16> %x, i16 %y, ptr %ptr) nounwind {
693; KNL-LABEL: insert_v16i16:
694; KNL:       ## %bb.0:
695; KNL-NEXT:    vpinsrw $1, (%rsi), %xmm0, %xmm1
696; KNL-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
697; KNL-NEXT:    vmovd %edi, %xmm1
698; KNL-NEXT:    vpbroadcastw %xmm1, %ymm1
699; KNL-NEXT:    vpblendw {{.*#+}} ymm1 = ymm0[0],ymm1[1],ymm0[2,3,4,5,6,7,8],ymm1[9],ymm0[10,11,12,13,14,15]
700; KNL-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
701; KNL-NEXT:    retq
702;
703; SKX-LABEL: insert_v16i16:
704; SKX:       ## %bb.0:
705; SKX-NEXT:    vpinsrw $1, (%rsi), %xmm0, %xmm1
706; SKX-NEXT:    vpblendd {{.*#+}} ymm0 = ymm1[0,1,2,3],ymm0[4,5,6,7]
707; SKX-NEXT:    vpbroadcastw %edi, %ymm1
708; SKX-NEXT:    vpblendw {{.*#+}} ymm1 = ymm0[0],ymm1[1],ymm0[2,3,4,5,6,7,8],ymm1[9],ymm0[10,11,12,13,14,15]
709; SKX-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
710; SKX-NEXT:    retq
711  %val = load i16, ptr %ptr
712  %r1 = insertelement <16 x i16> %x, i16 %val, i32 1
713  %r2 = insertelement <16 x i16> %r1, i16 %y, i32 9
714  ret <16 x i16> %r2
715}
716
717define <8 x i16> @insert_v8i16(<8 x i16> %x, i16 %y, ptr %ptr) nounwind {
718; CHECK-LABEL: insert_v8i16:
719; CHECK:       ## %bb.0:
720; CHECK-NEXT:    vpinsrw $1, (%rsi), %xmm0, %xmm0
721; CHECK-NEXT:    vpinsrw $5, %edi, %xmm0, %xmm0
722; CHECK-NEXT:    retq
723  %val = load i16, ptr %ptr
724  %r1 = insertelement <8 x i16> %x, i16 %val, i32 1
725  %r2 = insertelement <8 x i16> %r1, i16 %y, i32 5
726  ret <8 x i16> %r2
727}
728
729define <64 x i8> @insert_v64i8(<64 x i8> %x, i8 %y, ptr %ptr) nounwind {
730; CHECK-LABEL: insert_v64i8:
731; CHECK:       ## %bb.0:
732; CHECK-NEXT:    vpinsrb $1, (%rsi), %xmm0, %xmm1
733; CHECK-NEXT:    vinserti32x4 $0, %xmm1, %zmm0, %zmm1
734; CHECK-NEXT:    vextracti32x4 $3, %zmm0, %xmm0
735; CHECK-NEXT:    vpinsrb $2, %edi, %xmm0, %xmm0
736; CHECK-NEXT:    vinserti32x4 $3, %xmm0, %zmm1, %zmm0
737; CHECK-NEXT:    retq
738  %val = load i8, ptr %ptr
739  %r1 = insertelement <64 x i8> %x, i8 %val, i32 1
740  %r2 = insertelement <64 x i8> %r1, i8 %y, i32 50
741  ret <64 x i8> %r2
742}
743
744define <32 x i8> @insert_v32i8(<32 x i8> %x, i8 %y, ptr %ptr) nounwind {
745; CHECK-LABEL: insert_v32i8:
746; CHECK:       ## %bb.0:
747; CHECK-NEXT:    vpinsrb $1, (%rsi), %xmm0, %xmm1
748; CHECK-NEXT:    vextracti128 $1, %ymm0, %xmm0
749; CHECK-NEXT:    vpinsrb $1, %edi, %xmm0, %xmm0
750; CHECK-NEXT:    vinserti128 $1, %xmm0, %ymm1, %ymm0
751; CHECK-NEXT:    retq
752  %val = load i8, ptr %ptr
753  %r1 = insertelement <32 x i8> %x, i8 %val, i32 1
754  %r2 = insertelement <32 x i8> %r1, i8 %y, i32 17
755  ret <32 x i8> %r2
756}
757
758define <16 x i8> @insert_v16i8(<16 x i8> %x, i8 %y, ptr %ptr) nounwind {
759; CHECK-LABEL: insert_v16i8:
760; CHECK:       ## %bb.0:
761; CHECK-NEXT:    vpinsrb $3, (%rsi), %xmm0, %xmm0
762; CHECK-NEXT:    vpinsrb $10, %edi, %xmm0, %xmm0
763; CHECK-NEXT:    retq
764  %val = load i8, ptr %ptr
765  %r1 = insertelement <16 x i8> %x, i8 %val, i32 3
766  %r2 = insertelement <16 x i8> %r1, i8 %y, i32 10
767  ret <16 x i8> %r2
768}
769
770define <8 x i64> @test_insert_128_v8i64(<8 x i64> %x, i64 %y) nounwind {
771; CHECK-LABEL: test_insert_128_v8i64:
772; CHECK:       ## %bb.0:
773; CHECK-NEXT:    vpinsrq $1, %rdi, %xmm0, %xmm1
774; CHECK-NEXT:    vinserti32x4 $0, %xmm1, %zmm0, %zmm0
775; CHECK-NEXT:    retq
776  %r = insertelement <8 x i64> %x, i64 %y, i32 1
777  ret <8 x i64> %r
778}
779
780define <16 x i32> @test_insert_128_v16i32(<16 x i32> %x, i32 %y) nounwind {
781; CHECK-LABEL: test_insert_128_v16i32:
782; CHECK:       ## %bb.0:
783; CHECK-NEXT:    vpinsrd $1, %edi, %xmm0, %xmm1
784; CHECK-NEXT:    vinserti32x4 $0, %xmm1, %zmm0, %zmm0
785; CHECK-NEXT:    retq
786  %r = insertelement <16 x i32> %x, i32 %y, i32 1
787  ret <16 x i32> %r
788}
789
790define <8 x double> @test_insert_128_v8f64(<8 x double> %x, double %y) nounwind {
791; CHECK-LABEL: test_insert_128_v8f64:
792; CHECK:       ## %bb.0:
793; CHECK-NEXT:    vmovlhps {{.*#+}} xmm1 = xmm0[0],xmm1[0]
794; CHECK-NEXT:    vinsertf32x4 $0, %xmm1, %zmm0, %zmm0
795; CHECK-NEXT:    retq
796  %r = insertelement <8 x double> %x, double %y, i32 1
797  ret <8 x double> %r
798}
799
800define <16 x float> @test_insert_128_v16f32(<16 x float> %x, float %y) nounwind {
801; CHECK-LABEL: test_insert_128_v16f32:
802; CHECK:       ## %bb.0:
803; CHECK-NEXT:    vinsertps {{.*#+}} xmm1 = xmm0[0],xmm1[0],xmm0[2,3]
804; CHECK-NEXT:    vinsertf32x4 $0, %xmm1, %zmm0, %zmm0
805; CHECK-NEXT:    retq
806  %r = insertelement <16 x float> %x, float %y, i32 1
807  ret <16 x float> %r
808}
809
810define <16 x i16> @test_insert_128_v16i16(<16 x i16> %x, i16 %y) nounwind {
811; KNL-LABEL: test_insert_128_v16i16:
812; KNL:       ## %bb.0:
813; KNL-NEXT:    vmovd %edi, %xmm1
814; KNL-NEXT:    vpbroadcastw %xmm1, %ymm1
815; KNL-NEXT:    vpblendw {{.*#+}} ymm1 = ymm0[0,1],ymm1[2],ymm0[3,4,5,6,7,8,9],ymm1[10],ymm0[11,12,13,14,15]
816; KNL-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
817; KNL-NEXT:    retq
818;
819; SKX-LABEL: test_insert_128_v16i16:
820; SKX:       ## %bb.0:
821; SKX-NEXT:    vpbroadcastw %edi, %ymm1
822; SKX-NEXT:    vpblendw {{.*#+}} ymm1 = ymm0[0,1],ymm1[2],ymm0[3,4,5,6,7,8,9],ymm1[10],ymm0[11,12,13,14,15]
823; SKX-NEXT:    vpblendd {{.*#+}} ymm0 = ymm0[0,1,2,3],ymm1[4,5,6,7]
824; SKX-NEXT:    retq
825  %r = insertelement <16 x i16> %x, i16 %y, i32 10
826  ret <16 x i16> %r
827}
828
829define <32 x i8> @test_insert_128_v32i8(<32 x i8> %x, i8 %y) nounwind {
830; CHECK-LABEL: test_insert_128_v32i8:
831; CHECK:       ## %bb.0:
832; CHECK-NEXT:    vextracti128 $1, %ymm0, %xmm1
833; CHECK-NEXT:    vpinsrb $4, %edi, %xmm1, %xmm1
834; CHECK-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm0
835; CHECK-NEXT:    retq
836  %r = insertelement <32 x i8> %x, i8 %y, i32 20
837  ret <32 x i8> %r
838}
839
840define i32 @test_insertelement_v32i1(i32 %a, i32 %b, <32 x i32> %x , <32 x i32> %y) nounwind {
841; KNL-LABEL: test_insertelement_v32i1:
842; KNL:       ## %bb.0:
843; KNL-NEXT:    cmpl %esi, %edi
844; KNL-NEXT:    setb %al
845; KNL-NEXT:    vpcmpltud %zmm3, %zmm1, %k0
846; KNL-NEXT:    kmovw %k0, %ecx
847; KNL-NEXT:    shll $16, %ecx
848; KNL-NEXT:    movw $-17, %dx
849; KNL-NEXT:    kmovw %edx, %k1
850; KNL-NEXT:    vpcmpltud %zmm2, %zmm0, %k0 {%k1}
851; KNL-NEXT:    kmovw %eax, %k1
852; KNL-NEXT:    kshiftlw $15, %k1, %k1
853; KNL-NEXT:    kshiftrw $11, %k1, %k1
854; KNL-NEXT:    korw %k1, %k0, %k0
855; KNL-NEXT:    kmovw %k0, %eax
856; KNL-NEXT:    orl %ecx, %eax
857; KNL-NEXT:    vzeroupper
858; KNL-NEXT:    retq
859;
860; SKX-LABEL: test_insertelement_v32i1:
861; SKX:       ## %bb.0:
862; SKX-NEXT:    cmpl %esi, %edi
863; SKX-NEXT:    setb %al
864; SKX-NEXT:    vpcmpltud %zmm2, %zmm0, %k0
865; SKX-NEXT:    vpcmpltud %zmm3, %zmm1, %k1
866; SKX-NEXT:    kunpckwd %k0, %k1, %k0
867; SKX-NEXT:    movl $-17, %ecx
868; SKX-NEXT:    kmovd %ecx, %k1
869; SKX-NEXT:    kandd %k1, %k0, %k0
870; SKX-NEXT:    kmovd %eax, %k1
871; SKX-NEXT:    kshiftld $31, %k1, %k1
872; SKX-NEXT:    kshiftrd $27, %k1, %k1
873; SKX-NEXT:    kord %k1, %k0, %k0
874; SKX-NEXT:    kmovd %k0, %eax
875; SKX-NEXT:    vzeroupper
876; SKX-NEXT:    retq
877  %cmp_res_i1 = icmp ult i32 %a, %b
878  %cmp_cmp_vec = icmp ult <32 x i32> %x, %y
879  %maskv = insertelement <32 x i1> %cmp_cmp_vec, i1 %cmp_res_i1, i32 4
880  %res = bitcast <32 x i1> %maskv to i32
881  ret i32 %res
882}
883
884define i8 @test_iinsertelement_v4i1(i32 %a, i32 %b, <4 x i32> %x , <4 x i32> %y) nounwind {
885; KNL-LABEL: test_iinsertelement_v4i1:
886; KNL:       ## %bb.0:
887; KNL-NEXT:    ## kill: def $xmm1 killed $xmm1 def $zmm1
888; KNL-NEXT:    ## kill: def $xmm0 killed $xmm0 def $zmm0
889; KNL-NEXT:    cmpl %esi, %edi
890; KNL-NEXT:    setb %al
891; KNL-NEXT:    movw $-5, %cx
892; KNL-NEXT:    kmovw %ecx, %k1
893; KNL-NEXT:    vpcmpltud %zmm1, %zmm0, %k0 {%k1}
894; KNL-NEXT:    kmovw %eax, %k1
895; KNL-NEXT:    kshiftlw $15, %k1, %k1
896; KNL-NEXT:    kshiftrw $13, %k1, %k1
897; KNL-NEXT:    korw %k1, %k0, %k0
898; KNL-NEXT:    kmovw %k0, %eax
899; KNL-NEXT:    ## kill: def $al killed $al killed $eax
900; KNL-NEXT:    vzeroupper
901; KNL-NEXT:    retq
902;
903; SKX-LABEL: test_iinsertelement_v4i1:
904; SKX:       ## %bb.0:
905; SKX-NEXT:    cmpl %esi, %edi
906; SKX-NEXT:    setb %al
907; SKX-NEXT:    movb $-5, %cl
908; SKX-NEXT:    kmovd %ecx, %k1
909; SKX-NEXT:    vpcmpltud %xmm1, %xmm0, %k0 {%k1}
910; SKX-NEXT:    kmovd %eax, %k1
911; SKX-NEXT:    kshiftlb $7, %k1, %k1
912; SKX-NEXT:    kshiftrb $5, %k1, %k1
913; SKX-NEXT:    korw %k1, %k0, %k0
914; SKX-NEXT:    kmovd %k0, %eax
915; SKX-NEXT:    ## kill: def $al killed $al killed $eax
916; SKX-NEXT:    retq
917  %cmp_res_i1 = icmp ult i32 %a, %b
918  %cmp_cmp_vec = icmp ult <4 x i32> %x, %y
919  %maskv = insertelement <4 x i1> %cmp_cmp_vec, i1 %cmp_res_i1, i32 2
920  %res0 = shufflevector <4 x i1> %maskv, <4 x i1> undef , <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 4, i32 4, i32 4>
921  %res = bitcast <8 x i1> %res0 to i8
922  ret i8 %res
923}
924
925define i8 @test_iinsertelement_v2i1(i32 %a, i32 %b, <2 x i64> %x , <2 x i64> %y) nounwind {
926; KNL-LABEL: test_iinsertelement_v2i1:
927; KNL:       ## %bb.0:
928; KNL-NEXT:    ## kill: def $xmm1 killed $xmm1 def $zmm1
929; KNL-NEXT:    ## kill: def $xmm0 killed $xmm0 def $zmm0
930; KNL-NEXT:    cmpl %esi, %edi
931; KNL-NEXT:    setb %al
932; KNL-NEXT:    vpcmpltuq %zmm1, %zmm0, %k0
933; KNL-NEXT:    kshiftlw $15, %k0, %k0
934; KNL-NEXT:    kshiftrw $15, %k0, %k0
935; KNL-NEXT:    kmovw %eax, %k1
936; KNL-NEXT:    kshiftlw $1, %k1, %k1
937; KNL-NEXT:    korw %k1, %k0, %k0
938; KNL-NEXT:    kmovw %k0, %eax
939; KNL-NEXT:    ## kill: def $al killed $al killed $eax
940; KNL-NEXT:    vzeroupper
941; KNL-NEXT:    retq
942;
943; SKX-LABEL: test_iinsertelement_v2i1:
944; SKX:       ## %bb.0:
945; SKX-NEXT:    cmpl %esi, %edi
946; SKX-NEXT:    setb %al
947; SKX-NEXT:    vpcmpltuq %xmm1, %xmm0, %k0
948; SKX-NEXT:    kshiftlb $7, %k0, %k0
949; SKX-NEXT:    kshiftrb $7, %k0, %k0
950; SKX-NEXT:    kmovd %eax, %k1
951; SKX-NEXT:    kshiftlb $1, %k1, %k1
952; SKX-NEXT:    korw %k1, %k0, %k0
953; SKX-NEXT:    kmovd %k0, %eax
954; SKX-NEXT:    ## kill: def $al killed $al killed $eax
955; SKX-NEXT:    retq
956  %cmp_res_i1 = icmp ult i32 %a, %b
957  %cmp_cmp_vec = icmp ult <2 x i64> %x, %y
958  %maskv = insertelement <2 x i1> %cmp_cmp_vec, i1 %cmp_res_i1, i32 1
959  %res0 = shufflevector <2 x i1> %maskv, <2 x i1> undef , <8 x i32> <i32 0, i32 1, i32 2, i32 2, i32 2, i32 2, i32 2, i32 2>
960  %res = bitcast <8 x i1> %res0 to i8
961  ret i8 %res
962}
963
964define zeroext i8 @test_extractelement_v2i1(<2 x i64> %a, <2 x i64> %b) nounwind {
965; KNL-LABEL: test_extractelement_v2i1:
966; KNL:       ## %bb.0:
967; KNL-NEXT:    ## kill: def $xmm1 killed $xmm1 def $zmm1
968; KNL-NEXT:    ## kill: def $xmm0 killed $xmm0 def $zmm0
969; KNL-NEXT:    vpcmpnleuq %zmm1, %zmm0, %k0
970; KNL-NEXT:    kmovw %k0, %ecx
971; KNL-NEXT:    andl $1, %ecx
972; KNL-NEXT:    movl $4, %eax
973; KNL-NEXT:    subl %ecx, %eax
974; KNL-NEXT:    vzeroupper
975; KNL-NEXT:    retq
976;
977; SKX-LABEL: test_extractelement_v2i1:
978; SKX:       ## %bb.0:
979; SKX-NEXT:    vpcmpnleuq %xmm1, %xmm0, %k0
980; SKX-NEXT:    kmovd %k0, %ecx
981; SKX-NEXT:    andl $1, %ecx
982; SKX-NEXT:    movl $4, %eax
983; SKX-NEXT:    subl %ecx, %eax
984; SKX-NEXT:    retq
985  %t1 = icmp ugt <2 x i64> %a, %b
986  %t2 = extractelement <2 x i1> %t1, i32 0
987  %res = select i1 %t2, i8 3, i8 4
988  ret i8 %res
989}
990
991define zeroext i8 @extractelement_v2i1_alt(<2 x i64> %a, <2 x i64> %b) nounwind {
992; KNL-LABEL: extractelement_v2i1_alt:
993; KNL:       ## %bb.0:
994; KNL-NEXT:    ## kill: def $xmm1 killed $xmm1 def $zmm1
995; KNL-NEXT:    ## kill: def $xmm0 killed $xmm0 def $zmm0
996; KNL-NEXT:    vpcmpnleuq %zmm1, %zmm0, %k0
997; KNL-NEXT:    kmovw %k0, %eax
998; KNL-NEXT:    andb $1, %al
999; KNL-NEXT:    movb $4, %cl
1000; KNL-NEXT:    subb %al, %cl
1001; KNL-NEXT:    movzbl %cl, %eax
1002; KNL-NEXT:    vzeroupper
1003; KNL-NEXT:    retq
1004;
1005; SKX-LABEL: extractelement_v2i1_alt:
1006; SKX:       ## %bb.0:
1007; SKX-NEXT:    vpcmpnleuq %xmm1, %xmm0, %k0
1008; SKX-NEXT:    kmovd %k0, %eax
1009; SKX-NEXT:    andb $1, %al
1010; SKX-NEXT:    movb $4, %cl
1011; SKX-NEXT:    subb %al, %cl
1012; SKX-NEXT:    movzbl %cl, %eax
1013; SKX-NEXT:    retq
1014  %t1 = icmp ugt <2 x i64> %a, %b
1015  %t2 = extractelement <2 x i1> %t1, i32 0
1016  %sext = sext i1 %t2 to i8
1017  %res = add i8 %sext, 4
1018  ret i8 %res
1019}
1020
1021define zeroext i8 @test_extractelement_v4i1(<4 x i32> %a, <4 x i32> %b) nounwind {
1022; KNL-LABEL: test_extractelement_v4i1:
1023; KNL:       ## %bb.0:
1024; KNL-NEXT:    ## kill: def $xmm1 killed $xmm1 def $zmm1
1025; KNL-NEXT:    ## kill: def $xmm0 killed $xmm0 def $zmm0
1026; KNL-NEXT:    vpcmpnleud %zmm1, %zmm0, %k0
1027; KNL-NEXT:    kshiftrw $3, %k0, %k0
1028; KNL-NEXT:    kmovw %k0, %eax
1029; KNL-NEXT:    andl $1, %eax
1030; KNL-NEXT:    vzeroupper
1031; KNL-NEXT:    retq
1032;
1033; SKX-LABEL: test_extractelement_v4i1:
1034; SKX:       ## %bb.0:
1035; SKX-NEXT:    vpcmpnleud %xmm1, %xmm0, %k0
1036; SKX-NEXT:    kshiftrb $3, %k0, %k0
1037; SKX-NEXT:    kmovd %k0, %eax
1038; SKX-NEXT:    andl $1, %eax
1039; SKX-NEXT:    retq
1040  %t1 = icmp ugt <4 x i32> %a, %b
1041  %t2 = extractelement <4 x i1> %t1, i32 3
1042  %res = zext i1 %t2 to i8
1043  ret i8 %res
1044}
1045
1046define zeroext i8 @test_extractelement_v32i1(<32 x i8> %a, <32 x i8> %b) nounwind {
1047; KNL-LABEL: test_extractelement_v32i1:
1048; KNL:       ## %bb.0:
1049; KNL-NEXT:    vpminub %xmm1, %xmm0, %xmm1
1050; KNL-NEXT:    vpcmpeqb %xmm1, %xmm0, %xmm0
1051; KNL-NEXT:    vpextrb $2, %xmm0, %eax
1052; KNL-NEXT:    notb %al
1053; KNL-NEXT:    movzbl %al, %eax
1054; KNL-NEXT:    andl $1, %eax
1055; KNL-NEXT:    vzeroupper
1056; KNL-NEXT:    retq
1057;
1058; SKX-LABEL: test_extractelement_v32i1:
1059; SKX:       ## %bb.0:
1060; SKX-NEXT:    vpcmpnleub %ymm1, %ymm0, %k0
1061; SKX-NEXT:    kshiftrd $2, %k0, %k0
1062; SKX-NEXT:    kmovd %k0, %eax
1063; SKX-NEXT:    andl $1, %eax
1064; SKX-NEXT:    vzeroupper
1065; SKX-NEXT:    retq
1066  %t1 = icmp ugt <32 x i8> %a, %b
1067  %t2 = extractelement <32 x i1> %t1, i32 2
1068  %res = zext i1 %t2 to i8
1069  ret i8 %res
1070}
1071
1072define zeroext i8 @test_extractelement_v64i1(<64 x i8> %a, <64 x i8> %b) nounwind {
1073; KNL-LABEL: test_extractelement_v64i1:
1074; KNL:       ## %bb.0:
1075; KNL-NEXT:    vextracti64x4 $1, %zmm1, %ymm1
1076; KNL-NEXT:    vextracti64x4 $1, %zmm0, %ymm0
1077; KNL-NEXT:    vpminub %ymm1, %ymm0, %ymm1
1078; KNL-NEXT:    vpcmpeqb %ymm1, %ymm0, %ymm0
1079; KNL-NEXT:    vextracti128 $1, %ymm0, %xmm0
1080; KNL-NEXT:    vpextrb $15, %xmm0, %eax
1081; KNL-NEXT:    notb %al
1082; KNL-NEXT:    movzbl %al, %ecx
1083; KNL-NEXT:    andl $1, %ecx
1084; KNL-NEXT:    movl $4, %eax
1085; KNL-NEXT:    subl %ecx, %eax
1086; KNL-NEXT:    vzeroupper
1087; KNL-NEXT:    retq
1088;
1089; SKX-LABEL: test_extractelement_v64i1:
1090; SKX:       ## %bb.0:
1091; SKX-NEXT:    vpcmpnleub %zmm1, %zmm0, %k0
1092; SKX-NEXT:    kshiftrq $63, %k0, %k0
1093; SKX-NEXT:    kmovd %k0, %ecx
1094; SKX-NEXT:    andl $1, %ecx
1095; SKX-NEXT:    movl $4, %eax
1096; SKX-NEXT:    subl %ecx, %eax
1097; SKX-NEXT:    vzeroupper
1098; SKX-NEXT:    retq
1099  %t1 = icmp ugt <64 x i8> %a, %b
1100  %t2 = extractelement <64 x i1> %t1, i32 63
1101  %res = select i1 %t2, i8 3, i8 4
1102  ret i8 %res
1103}
1104
1105define zeroext i8 @extractelement_v64i1_alt(<64 x i8> %a, <64 x i8> %b) nounwind {
1106; KNL-LABEL: extractelement_v64i1_alt:
1107; KNL:       ## %bb.0:
1108; KNL-NEXT:    vextracti64x4 $1, %zmm1, %ymm1
1109; KNL-NEXT:    vextracti64x4 $1, %zmm0, %ymm0
1110; KNL-NEXT:    vpminub %ymm1, %ymm0, %ymm1
1111; KNL-NEXT:    vpcmpeqb %ymm1, %ymm0, %ymm0
1112; KNL-NEXT:    vextracti128 $1, %ymm0, %xmm0
1113; KNL-NEXT:    vpextrb $15, %xmm0, %eax
1114; KNL-NEXT:    notb %al
1115; KNL-NEXT:    addb $4, %al
1116; KNL-NEXT:    movzbl %al, %eax
1117; KNL-NEXT:    vzeroupper
1118; KNL-NEXT:    retq
1119;
1120; SKX-LABEL: extractelement_v64i1_alt:
1121; SKX:       ## %bb.0:
1122; SKX-NEXT:    vpcmpnleub %zmm1, %zmm0, %k0
1123; SKX-NEXT:    kshiftrq $63, %k0, %k0
1124; SKX-NEXT:    kmovd %k0, %eax
1125; SKX-NEXT:    andb $1, %al
1126; SKX-NEXT:    movb $4, %cl
1127; SKX-NEXT:    subb %al, %cl
1128; SKX-NEXT:    movzbl %cl, %eax
1129; SKX-NEXT:    vzeroupper
1130; SKX-NEXT:    retq
1131  %t1 = icmp ugt <64 x i8> %a, %b
1132  %t2 = extractelement <64 x i1> %t1, i32 63
1133  %sext = sext i1 %t2 to i8
1134  %res = add i8 %sext, 4
1135  ret i8 %res
1136}
1137
1138define i64 @test_extractelement_variable_v2i64(<2 x i64> %t1, i32 %index) nounwind {
1139; CHECK-LABEL: test_extractelement_variable_v2i64:
1140; CHECK:       ## %bb.0:
1141; CHECK-NEXT:    ## kill: def $edi killed $edi def $rdi
1142; CHECK-NEXT:    vmovaps %xmm0, -{{[0-9]+}}(%rsp)
1143; CHECK-NEXT:    andl $1, %edi
1144; CHECK-NEXT:    movq -24(%rsp,%rdi,8), %rax
1145; CHECK-NEXT:    retq
1146  %t2 = extractelement <2 x i64> %t1, i32 %index
1147  ret i64 %t2
1148}
1149
1150define i64 @test_extractelement_variable_v4i64(<4 x i64> %t1, i32 %index) nounwind {
1151; CHECK-LABEL: test_extractelement_variable_v4i64:
1152; CHECK:       ## %bb.0:
1153; CHECK-NEXT:    pushq %rbp
1154; CHECK-NEXT:    movq %rsp, %rbp
1155; CHECK-NEXT:    andq $-32, %rsp
1156; CHECK-NEXT:    subq $64, %rsp
1157; CHECK-NEXT:    ## kill: def $edi killed $edi def $rdi
1158; CHECK-NEXT:    vmovaps %ymm0, (%rsp)
1159; CHECK-NEXT:    andl $3, %edi
1160; CHECK-NEXT:    movq (%rsp,%rdi,8), %rax
1161; CHECK-NEXT:    movq %rbp, %rsp
1162; CHECK-NEXT:    popq %rbp
1163; CHECK-NEXT:    vzeroupper
1164; CHECK-NEXT:    retq
1165  %t2 = extractelement <4 x i64> %t1, i32 %index
1166  ret i64 %t2
1167}
1168
1169define i64 @test_extractelement_variable_v8i64(<8 x i64> %t1, i32 %index) nounwind {
1170; CHECK-LABEL: test_extractelement_variable_v8i64:
1171; CHECK:       ## %bb.0:
1172; CHECK-NEXT:    pushq %rbp
1173; CHECK-NEXT:    movq %rsp, %rbp
1174; CHECK-NEXT:    andq $-64, %rsp
1175; CHECK-NEXT:    subq $128, %rsp
1176; CHECK-NEXT:    ## kill: def $edi killed $edi def $rdi
1177; CHECK-NEXT:    vmovaps %zmm0, (%rsp)
1178; CHECK-NEXT:    andl $7, %edi
1179; CHECK-NEXT:    movq (%rsp,%rdi,8), %rax
1180; CHECK-NEXT:    movq %rbp, %rsp
1181; CHECK-NEXT:    popq %rbp
1182; CHECK-NEXT:    vzeroupper
1183; CHECK-NEXT:    retq
1184  %t2 = extractelement <8 x i64> %t1, i32 %index
1185  ret i64 %t2
1186}
1187
1188define double @test_extractelement_variable_v2f64(<2 x double> %t1, i32 %index) nounwind {
1189; CHECK-LABEL: test_extractelement_variable_v2f64:
1190; CHECK:       ## %bb.0:
1191; CHECK-NEXT:    ## kill: def $edi killed $edi def $rdi
1192; CHECK-NEXT:    vmovaps %xmm0, -{{[0-9]+}}(%rsp)
1193; CHECK-NEXT:    andl $1, %edi
1194; CHECK-NEXT:    vmovsd {{.*#+}} xmm0 = mem[0],zero
1195; CHECK-NEXT:    retq
1196  %t2 = extractelement <2 x double> %t1, i32 %index
1197  ret double %t2
1198}
1199
1200define double @test_extractelement_variable_v4f64(<4 x double> %t1, i32 %index) nounwind {
1201; CHECK-LABEL: test_extractelement_variable_v4f64:
1202; CHECK:       ## %bb.0:
1203; CHECK-NEXT:    pushq %rbp
1204; CHECK-NEXT:    movq %rsp, %rbp
1205; CHECK-NEXT:    andq $-32, %rsp
1206; CHECK-NEXT:    subq $64, %rsp
1207; CHECK-NEXT:    ## kill: def $edi killed $edi def $rdi
1208; CHECK-NEXT:    vmovaps %ymm0, (%rsp)
1209; CHECK-NEXT:    andl $3, %edi
1210; CHECK-NEXT:    vmovsd {{.*#+}} xmm0 = mem[0],zero
1211; CHECK-NEXT:    movq %rbp, %rsp
1212; CHECK-NEXT:    popq %rbp
1213; CHECK-NEXT:    vzeroupper
1214; CHECK-NEXT:    retq
1215  %t2 = extractelement <4 x double> %t1, i32 %index
1216  ret double %t2
1217}
1218
1219define double @test_extractelement_variable_v8f64(<8 x double> %t1, i32 %index) nounwind {
1220; CHECK-LABEL: test_extractelement_variable_v8f64:
1221; CHECK:       ## %bb.0:
1222; CHECK-NEXT:    pushq %rbp
1223; CHECK-NEXT:    movq %rsp, %rbp
1224; CHECK-NEXT:    andq $-64, %rsp
1225; CHECK-NEXT:    subq $128, %rsp
1226; CHECK-NEXT:    ## kill: def $edi killed $edi def $rdi
1227; CHECK-NEXT:    vmovaps %zmm0, (%rsp)
1228; CHECK-NEXT:    andl $7, %edi
1229; CHECK-NEXT:    vmovsd {{.*#+}} xmm0 = mem[0],zero
1230; CHECK-NEXT:    movq %rbp, %rsp
1231; CHECK-NEXT:    popq %rbp
1232; CHECK-NEXT:    vzeroupper
1233; CHECK-NEXT:    retq
1234  %t2 = extractelement <8 x double> %t1, i32 %index
1235  ret double %t2
1236}
1237
1238define i32 @test_extractelement_variable_v4i32(<4 x i32> %t1, i32 %index) nounwind {
1239; CHECK-LABEL: test_extractelement_variable_v4i32:
1240; CHECK:       ## %bb.0:
1241; CHECK-NEXT:    ## kill: def $edi killed $edi def $rdi
1242; CHECK-NEXT:    vmovaps %xmm0, -{{[0-9]+}}(%rsp)
1243; CHECK-NEXT:    andl $3, %edi
1244; CHECK-NEXT:    movl -24(%rsp,%rdi,4), %eax
1245; CHECK-NEXT:    retq
1246  %t2 = extractelement <4 x i32> %t1, i32 %index
1247  ret i32 %t2
1248}
1249
1250define i32 @test_extractelement_variable_v8i32(<8 x i32> %t1, i32 %index) nounwind {
1251; CHECK-LABEL: test_extractelement_variable_v8i32:
1252; CHECK:       ## %bb.0:
1253; CHECK-NEXT:    pushq %rbp
1254; CHECK-NEXT:    movq %rsp, %rbp
1255; CHECK-NEXT:    andq $-32, %rsp
1256; CHECK-NEXT:    subq $64, %rsp
1257; CHECK-NEXT:    ## kill: def $edi killed $edi def $rdi
1258; CHECK-NEXT:    vmovaps %ymm0, (%rsp)
1259; CHECK-NEXT:    andl $7, %edi
1260; CHECK-NEXT:    movl (%rsp,%rdi,4), %eax
1261; CHECK-NEXT:    movq %rbp, %rsp
1262; CHECK-NEXT:    popq %rbp
1263; CHECK-NEXT:    vzeroupper
1264; CHECK-NEXT:    retq
1265  %t2 = extractelement <8 x i32> %t1, i32 %index
1266  ret i32 %t2
1267}
1268
1269define i32 @test_extractelement_variable_v16i32(<16 x i32> %t1, i32 %index) nounwind {
1270; CHECK-LABEL: test_extractelement_variable_v16i32:
1271; CHECK:       ## %bb.0:
1272; CHECK-NEXT:    pushq %rbp
1273; CHECK-NEXT:    movq %rsp, %rbp
1274; CHECK-NEXT:    andq $-64, %rsp
1275; CHECK-NEXT:    subq $128, %rsp
1276; CHECK-NEXT:    ## kill: def $edi killed $edi def $rdi
1277; CHECK-NEXT:    vmovaps %zmm0, (%rsp)
1278; CHECK-NEXT:    andl $15, %edi
1279; CHECK-NEXT:    movl (%rsp,%rdi,4), %eax
1280; CHECK-NEXT:    movq %rbp, %rsp
1281; CHECK-NEXT:    popq %rbp
1282; CHECK-NEXT:    vzeroupper
1283; CHECK-NEXT:    retq
1284  %t2 = extractelement <16 x i32> %t1, i32 %index
1285  ret i32 %t2
1286}
1287
1288define float @test_extractelement_variable_v4f32(<4 x float> %t1, i32 %index) nounwind {
1289; CHECK-LABEL: test_extractelement_variable_v4f32:
1290; CHECK:       ## %bb.0:
1291; CHECK-NEXT:    ## kill: def $edi killed $edi def $rdi
1292; CHECK-NEXT:    vmovaps %xmm0, -{{[0-9]+}}(%rsp)
1293; CHECK-NEXT:    andl $3, %edi
1294; CHECK-NEXT:    vmovss {{.*#+}} xmm0 = mem[0],zero,zero,zero
1295; CHECK-NEXT:    retq
1296  %t2 = extractelement <4 x float> %t1, i32 %index
1297  ret float %t2
1298}
1299
1300define float @test_extractelement_variable_v8f32(<8 x float> %t1, i32 %index) nounwind {
1301; CHECK-LABEL: test_extractelement_variable_v8f32:
1302; CHECK:       ## %bb.0:
1303; CHECK-NEXT:    pushq %rbp
1304; CHECK-NEXT:    movq %rsp, %rbp
1305; CHECK-NEXT:    andq $-32, %rsp
1306; CHECK-NEXT:    subq $64, %rsp
1307; CHECK-NEXT:    ## kill: def $edi killed $edi def $rdi
1308; CHECK-NEXT:    vmovaps %ymm0, (%rsp)
1309; CHECK-NEXT:    andl $7, %edi
1310; CHECK-NEXT:    vmovss {{.*#+}} xmm0 = mem[0],zero,zero,zero
1311; CHECK-NEXT:    movq %rbp, %rsp
1312; CHECK-NEXT:    popq %rbp
1313; CHECK-NEXT:    vzeroupper
1314; CHECK-NEXT:    retq
1315  %t2 = extractelement <8 x float> %t1, i32 %index
1316  ret float %t2
1317}
1318
1319define float @test_extractelement_variable_v16f32(<16 x float> %t1, i32 %index) nounwind {
1320; CHECK-LABEL: test_extractelement_variable_v16f32:
1321; CHECK:       ## %bb.0:
1322; CHECK-NEXT:    pushq %rbp
1323; CHECK-NEXT:    movq %rsp, %rbp
1324; CHECK-NEXT:    andq $-64, %rsp
1325; CHECK-NEXT:    subq $128, %rsp
1326; CHECK-NEXT:    ## kill: def $edi killed $edi def $rdi
1327; CHECK-NEXT:    vmovaps %zmm0, (%rsp)
1328; CHECK-NEXT:    andl $15, %edi
1329; CHECK-NEXT:    vmovss {{.*#+}} xmm0 = mem[0],zero,zero,zero
1330; CHECK-NEXT:    movq %rbp, %rsp
1331; CHECK-NEXT:    popq %rbp
1332; CHECK-NEXT:    vzeroupper
1333; CHECK-NEXT:    retq
1334  %t2 = extractelement <16 x float> %t1, i32 %index
1335  ret float %t2
1336}
1337
1338define i16 @test_extractelement_variable_v8i16(<8 x i16> %t1, i32 %index) nounwind {
1339; CHECK-LABEL: test_extractelement_variable_v8i16:
1340; CHECK:       ## %bb.0:
1341; CHECK-NEXT:    ## kill: def $edi killed $edi def $rdi
1342; CHECK-NEXT:    vmovaps %xmm0, -{{[0-9]+}}(%rsp)
1343; CHECK-NEXT:    andl $7, %edi
1344; CHECK-NEXT:    movzwl -24(%rsp,%rdi,2), %eax
1345; CHECK-NEXT:    retq
1346  %t2 = extractelement <8 x i16> %t1, i32 %index
1347  ret i16 %t2
1348}
1349
1350define i16 @test_extractelement_variable_v16i16(<16 x i16> %t1, i32 %index) nounwind {
1351; CHECK-LABEL: test_extractelement_variable_v16i16:
1352; CHECK:       ## %bb.0:
1353; CHECK-NEXT:    pushq %rbp
1354; CHECK-NEXT:    movq %rsp, %rbp
1355; CHECK-NEXT:    andq $-32, %rsp
1356; CHECK-NEXT:    subq $64, %rsp
1357; CHECK-NEXT:    ## kill: def $edi killed $edi def $rdi
1358; CHECK-NEXT:    vmovaps %ymm0, (%rsp)
1359; CHECK-NEXT:    andl $15, %edi
1360; CHECK-NEXT:    movzwl (%rsp,%rdi,2), %eax
1361; CHECK-NEXT:    movq %rbp, %rsp
1362; CHECK-NEXT:    popq %rbp
1363; CHECK-NEXT:    vzeroupper
1364; CHECK-NEXT:    retq
1365  %t2 = extractelement <16 x i16> %t1, i32 %index
1366  ret i16 %t2
1367}
1368
1369define i16 @test_extractelement_variable_v32i16(<32 x i16> %t1, i32 %index) nounwind {
1370; CHECK-LABEL: test_extractelement_variable_v32i16:
1371; CHECK:       ## %bb.0:
1372; CHECK-NEXT:    pushq %rbp
1373; CHECK-NEXT:    movq %rsp, %rbp
1374; CHECK-NEXT:    andq $-64, %rsp
1375; CHECK-NEXT:    subq $128, %rsp
1376; CHECK-NEXT:    ## kill: def $edi killed $edi def $rdi
1377; CHECK-NEXT:    vmovaps %zmm0, (%rsp)
1378; CHECK-NEXT:    andl $31, %edi
1379; CHECK-NEXT:    movzwl (%rsp,%rdi,2), %eax
1380; CHECK-NEXT:    movq %rbp, %rsp
1381; CHECK-NEXT:    popq %rbp
1382; CHECK-NEXT:    vzeroupper
1383; CHECK-NEXT:    retq
1384  %t2 = extractelement <32 x i16> %t1, i32 %index
1385  ret i16 %t2
1386}
1387
1388define i8 @test_extractelement_variable_v16i8(<16 x i8> %t1, i32 %index) nounwind {
1389; CHECK-LABEL: test_extractelement_variable_v16i8:
1390; CHECK:       ## %bb.0:
1391; CHECK-NEXT:    ## kill: def $edi killed $edi def $rdi
1392; CHECK-NEXT:    vmovaps %xmm0, -{{[0-9]+}}(%rsp)
1393; CHECK-NEXT:    andl $15, %edi
1394; CHECK-NEXT:    movzbl -24(%rsp,%rdi), %eax
1395; CHECK-NEXT:    retq
1396  %t2 = extractelement <16 x i8> %t1, i32 %index
1397  ret i8 %t2
1398}
1399
1400define i8 @test_extractelement_variable_v32i8(<32 x i8> %t1, i32 %index) nounwind {
1401; CHECK-LABEL: test_extractelement_variable_v32i8:
1402; CHECK:       ## %bb.0:
1403; CHECK-NEXT:    pushq %rbp
1404; CHECK-NEXT:    movq %rsp, %rbp
1405; CHECK-NEXT:    andq $-32, %rsp
1406; CHECK-NEXT:    subq $64, %rsp
1407; CHECK-NEXT:    ## kill: def $edi killed $edi def $rdi
1408; CHECK-NEXT:    vmovaps %ymm0, (%rsp)
1409; CHECK-NEXT:    andl $31, %edi
1410; CHECK-NEXT:    movzbl (%rsp,%rdi), %eax
1411; CHECK-NEXT:    movq %rbp, %rsp
1412; CHECK-NEXT:    popq %rbp
1413; CHECK-NEXT:    vzeroupper
1414; CHECK-NEXT:    retq
1415
1416  %t2 = extractelement <32 x i8> %t1, i32 %index
1417  ret i8 %t2
1418}
1419
1420define i8 @test_extractelement_variable_v64i8(<64 x i8> %t1, i32 %index) nounwind {
1421; CHECK-LABEL: test_extractelement_variable_v64i8:
1422; CHECK:       ## %bb.0:
1423; CHECK-NEXT:    pushq %rbp
1424; CHECK-NEXT:    movq %rsp, %rbp
1425; CHECK-NEXT:    andq $-64, %rsp
1426; CHECK-NEXT:    subq $128, %rsp
1427; CHECK-NEXT:    ## kill: def $edi killed $edi def $rdi
1428; CHECK-NEXT:    vmovaps %zmm0, (%rsp)
1429; CHECK-NEXT:    andl $63, %edi
1430; CHECK-NEXT:    movzbl (%rsp,%rdi), %eax
1431; CHECK-NEXT:    movq %rbp, %rsp
1432; CHECK-NEXT:    popq %rbp
1433; CHECK-NEXT:    vzeroupper
1434; CHECK-NEXT:    retq
1435
1436  %t2 = extractelement <64 x i8> %t1, i32 %index
1437  ret i8 %t2
1438}
1439
1440define i8 @test_extractelement_variable_v64i8_indexi8(<64 x i8> %t1, i8 %index) nounwind {
1441; CHECK-LABEL: test_extractelement_variable_v64i8_indexi8:
1442; CHECK:       ## %bb.0:
1443; CHECK-NEXT:    pushq %rbp
1444; CHECK-NEXT:    movq %rsp, %rbp
1445; CHECK-NEXT:    andq $-64, %rsp
1446; CHECK-NEXT:    subq $128, %rsp
1447; CHECK-NEXT:    addb %dil, %dil
1448; CHECK-NEXT:    vmovaps %zmm0, (%rsp)
1449; CHECK-NEXT:    movzbl %dil, %eax
1450; CHECK-NEXT:    andl $63, %eax
1451; CHECK-NEXT:    movzbl (%rsp,%rax), %eax
1452; CHECK-NEXT:    movq %rbp, %rsp
1453; CHECK-NEXT:    popq %rbp
1454; CHECK-NEXT:    vzeroupper
1455; CHECK-NEXT:    retq
1456
1457  %i  = add i8 %index, %index
1458  %t2 = extractelement <64 x i8> %t1, i8 %i
1459  ret i8 %t2
1460}
1461
1462define zeroext i8 @test_extractelement_varible_v2i1(<2 x i64> %a, <2 x i64> %b, i32 %index) nounwind {
1463; KNL-LABEL: test_extractelement_varible_v2i1:
1464; KNL:       ## %bb.0:
1465; KNL-NEXT:    ## kill: def $edi killed $edi def $rdi
1466; KNL-NEXT:    ## kill: def $xmm1 killed $xmm1 def $zmm1
1467; KNL-NEXT:    ## kill: def $xmm0 killed $xmm0 def $zmm0
1468; KNL-NEXT:    vpcmpnleuq %zmm1, %zmm0, %k1
1469; KNL-NEXT:    vpternlogq {{.*#+}} zmm0 {%k1} {z} = -1
1470; KNL-NEXT:    vmovdqa %xmm0, -{{[0-9]+}}(%rsp)
1471; KNL-NEXT:    andl $1, %edi
1472; KNL-NEXT:    movzbl -24(%rsp,%rdi,8), %eax
1473; KNL-NEXT:    andl $1, %eax
1474; KNL-NEXT:    vzeroupper
1475; KNL-NEXT:    retq
1476;
1477; SKX-LABEL: test_extractelement_varible_v2i1:
1478; SKX:       ## %bb.0:
1479; SKX-NEXT:    ## kill: def $edi killed $edi def $rdi
1480; SKX-NEXT:    vpcmpnleuq %xmm1, %xmm0, %k0
1481; SKX-NEXT:    vpmovm2q %k0, %xmm0
1482; SKX-NEXT:    vmovdqa %xmm0, -{{[0-9]+}}(%rsp)
1483; SKX-NEXT:    andl $1, %edi
1484; SKX-NEXT:    movzbl -24(%rsp,%rdi,8), %eax
1485; SKX-NEXT:    andl $1, %eax
1486; SKX-NEXT:    retq
1487  %t1 = icmp ugt <2 x i64> %a, %b
1488  %t2 = extractelement <2 x i1> %t1, i32 %index
1489  %res = zext i1 %t2 to i8
1490  ret i8 %res
1491}
1492
1493define zeroext i8 @test_extractelement_varible_v4i1(<4 x i32> %a, <4 x i32> %b, i32 %index) nounwind {
1494; KNL-LABEL: test_extractelement_varible_v4i1:
1495; KNL:       ## %bb.0:
1496; KNL-NEXT:    ## kill: def $edi killed $edi def $rdi
1497; KNL-NEXT:    ## kill: def $xmm1 killed $xmm1 def $zmm1
1498; KNL-NEXT:    ## kill: def $xmm0 killed $xmm0 def $zmm0
1499; KNL-NEXT:    vpcmpnleud %zmm1, %zmm0, %k1
1500; KNL-NEXT:    vpternlogd {{.*#+}} zmm0 {%k1} {z} = -1
1501; KNL-NEXT:    vmovdqa %xmm0, -{{[0-9]+}}(%rsp)
1502; KNL-NEXT:    andl $3, %edi
1503; KNL-NEXT:    movzbl -24(%rsp,%rdi,4), %eax
1504; KNL-NEXT:    andl $1, %eax
1505; KNL-NEXT:    vzeroupper
1506; KNL-NEXT:    retq
1507;
1508; SKX-LABEL: test_extractelement_varible_v4i1:
1509; SKX:       ## %bb.0:
1510; SKX-NEXT:    ## kill: def $edi killed $edi def $rdi
1511; SKX-NEXT:    vpcmpnleud %xmm1, %xmm0, %k0
1512; SKX-NEXT:    vpmovm2d %k0, %xmm0
1513; SKX-NEXT:    vmovdqa %xmm0, -{{[0-9]+}}(%rsp)
1514; SKX-NEXT:    andl $3, %edi
1515; SKX-NEXT:    movzbl -24(%rsp,%rdi,4), %eax
1516; SKX-NEXT:    andl $1, %eax
1517; SKX-NEXT:    retq
1518  %t1 = icmp ugt <4 x i32> %a, %b
1519  %t2 = extractelement <4 x i1> %t1, i32 %index
1520  %res = zext i1 %t2 to i8
1521  ret i8 %res
1522}
1523
1524define zeroext i8 @test_extractelement_varible_v8i1(<8 x i32> %a, <8 x i32> %b, i32 %index) nounwind {
1525; KNL-LABEL: test_extractelement_varible_v8i1:
1526; KNL:       ## %bb.0:
1527; KNL-NEXT:    ## kill: def $edi killed $edi def $rdi
1528; KNL-NEXT:    ## kill: def $ymm1 killed $ymm1 def $zmm1
1529; KNL-NEXT:    ## kill: def $ymm0 killed $ymm0 def $zmm0
1530; KNL-NEXT:    vpcmpnleud %zmm1, %zmm0, %k1
1531; KNL-NEXT:    vpternlogd {{.*#+}} zmm0 {%k1} {z} = -1
1532; KNL-NEXT:    vpmovdw %zmm0, %ymm0
1533; KNL-NEXT:    vmovdqa %xmm0, -{{[0-9]+}}(%rsp)
1534; KNL-NEXT:    andl $7, %edi
1535; KNL-NEXT:    movzbl -24(%rsp,%rdi,2), %eax
1536; KNL-NEXT:    andl $1, %eax
1537; KNL-NEXT:    vzeroupper
1538; KNL-NEXT:    retq
1539;
1540; SKX-LABEL: test_extractelement_varible_v8i1:
1541; SKX:       ## %bb.0:
1542; SKX-NEXT:    ## kill: def $edi killed $edi def $rdi
1543; SKX-NEXT:    vpcmpnleud %ymm1, %ymm0, %k0
1544; SKX-NEXT:    vpmovm2w %k0, %xmm0
1545; SKX-NEXT:    vmovdqa %xmm0, -{{[0-9]+}}(%rsp)
1546; SKX-NEXT:    andl $7, %edi
1547; SKX-NEXT:    movzbl -24(%rsp,%rdi,2), %eax
1548; SKX-NEXT:    andl $1, %eax
1549; SKX-NEXT:    vzeroupper
1550; SKX-NEXT:    retq
1551  %t1 = icmp ugt <8 x i32> %a, %b
1552  %t2 = extractelement <8 x i1> %t1, i32 %index
1553  %res = zext i1 %t2 to i8
1554  ret i8 %res
1555}
1556
1557define zeroext i8 @test_extractelement_varible_v16i1(<16 x i32> %a, <16 x i32> %b, i32 %index) nounwind {
1558; KNL-LABEL: test_extractelement_varible_v16i1:
1559; KNL:       ## %bb.0:
1560; KNL-NEXT:    ## kill: def $edi killed $edi def $rdi
1561; KNL-NEXT:    vpcmpnleud %zmm1, %zmm0, %k1
1562; KNL-NEXT:    vpternlogd {{.*#+}} zmm0 {%k1} {z} = -1
1563; KNL-NEXT:    vpmovdb %zmm0, -{{[0-9]+}}(%rsp)
1564; KNL-NEXT:    andl $15, %edi
1565; KNL-NEXT:    movzbl -24(%rsp,%rdi), %eax
1566; KNL-NEXT:    andl $1, %eax
1567; KNL-NEXT:    vzeroupper
1568; KNL-NEXT:    retq
1569;
1570; SKX-LABEL: test_extractelement_varible_v16i1:
1571; SKX:       ## %bb.0:
1572; SKX-NEXT:    ## kill: def $edi killed $edi def $rdi
1573; SKX-NEXT:    vpcmpnleud %zmm1, %zmm0, %k0
1574; SKX-NEXT:    vpmovm2b %k0, %xmm0
1575; SKX-NEXT:    vmovdqa %xmm0, -{{[0-9]+}}(%rsp)
1576; SKX-NEXT:    andl $15, %edi
1577; SKX-NEXT:    movzbl -24(%rsp,%rdi), %eax
1578; SKX-NEXT:    andl $1, %eax
1579; SKX-NEXT:    vzeroupper
1580; SKX-NEXT:    retq
1581  %t1 = icmp ugt <16 x i32> %a, %b
1582  %t2 = extractelement <16 x i1> %t1, i32 %index
1583  %res = zext i1 %t2 to i8
1584  ret i8 %res
1585}
1586
1587define zeroext i8 @test_extractelement_varible_v32i1(<32 x i8> %a, <32 x i8> %b, i32 %index) nounwind {
1588; KNL-LABEL: test_extractelement_varible_v32i1:
1589; KNL:       ## %bb.0:
1590; KNL-NEXT:    vpminub %ymm1, %ymm0, %ymm1
1591; KNL-NEXT:    vpcmpeqb %ymm1, %ymm0, %ymm0
1592; KNL-NEXT:    vpmovmskb %ymm0, %ecx
1593; KNL-NEXT:    xorl %eax, %eax
1594; KNL-NEXT:    btl %edi, %ecx
1595; KNL-NEXT:    setae %al
1596; KNL-NEXT:    vzeroupper
1597; KNL-NEXT:    retq
1598;
1599; SKX-LABEL: test_extractelement_varible_v32i1:
1600; SKX:       ## %bb.0:
1601; SKX-NEXT:    pushq %rbp
1602; SKX-NEXT:    movq %rsp, %rbp
1603; SKX-NEXT:    andq $-32, %rsp
1604; SKX-NEXT:    subq $64, %rsp
1605; SKX-NEXT:    ## kill: def $edi killed $edi def $rdi
1606; SKX-NEXT:    vpcmpnleub %ymm1, %ymm0, %k0
1607; SKX-NEXT:    vpmovm2b %k0, %ymm0
1608; SKX-NEXT:    vmovdqa %ymm0, (%rsp)
1609; SKX-NEXT:    andl $31, %edi
1610; SKX-NEXT:    movzbl (%rsp,%rdi), %eax
1611; SKX-NEXT:    andl $1, %eax
1612; SKX-NEXT:    movq %rbp, %rsp
1613; SKX-NEXT:    popq %rbp
1614; SKX-NEXT:    vzeroupper
1615; SKX-NEXT:    retq
1616  %t1 = icmp ugt <32 x i8> %a, %b
1617  %t2 = extractelement <32 x i1> %t1, i32 %index
1618  %res = zext i1 %t2 to i8
1619  ret i8 %res
1620}
1621
1622define <8 x i64> @insert_double_zero(<2 x i64> %a) nounwind {
1623; CHECK-LABEL: insert_double_zero:
1624; CHECK:       ## %bb.0:
1625; CHECK-NEXT:    vxorps %xmm1, %xmm1, %xmm1
1626; CHECK-NEXT:    vinsertf32x4 $2, %xmm0, %zmm1, %zmm0
1627; CHECK-NEXT:    retq
1628  %b = shufflevector <2 x i64> %a, <2 x i64> zeroinitializer, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
1629  %d = shufflevector <4 x i64> %b, <4 x i64> undef, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
1630  %e = shufflevector <8 x i64> %d, <8 x i64> zeroinitializer, <8 x i32> <i32 8, i32 9, i32 10, i32 11, i32 0, i32 1, i32 2, i32 3>
1631  ret <8 x i64> %e
1632}
1633
1634define i32 @test_insertelement_variable_v32i1(<32 x i8> %a, i8 %b, i32 %index) nounwind {
1635; KNL-LABEL: test_insertelement_variable_v32i1:
1636; KNL:       ## %bb.0:
1637; KNL-NEXT:    pushq %rbp
1638; KNL-NEXT:    movq %rsp, %rbp
1639; KNL-NEXT:    andq $-32, %rsp
1640; KNL-NEXT:    subq $64, %rsp
1641; KNL-NEXT:    ## kill: def $esi killed $esi def $rsi
1642; KNL-NEXT:    vpxor %xmm1, %xmm1, %xmm1
1643; KNL-NEXT:    vpcmpeqb %ymm1, %ymm0, %ymm0
1644; KNL-NEXT:    vpternlogq {{.*#+}} zmm0 = ~zmm0
1645; KNL-NEXT:    andl $31, %esi
1646; KNL-NEXT:    testb %dil, %dil
1647; KNL-NEXT:    vmovdqa %ymm0, (%rsp)
1648; KNL-NEXT:    setne (%rsp,%rsi)
1649; KNL-NEXT:    vpmovsxbd (%rsp), %zmm0
1650; KNL-NEXT:    vpslld $31, %zmm0, %zmm0
1651; KNL-NEXT:    vptestmd %zmm0, %zmm0, %k0
1652; KNL-NEXT:    kmovw %k0, %ecx
1653; KNL-NEXT:    vpmovsxbd {{[0-9]+}}(%rsp), %zmm0
1654; KNL-NEXT:    vpslld $31, %zmm0, %zmm0
1655; KNL-NEXT:    vptestmd %zmm0, %zmm0, %k0
1656; KNL-NEXT:    kmovw %k0, %eax
1657; KNL-NEXT:    shll $16, %eax
1658; KNL-NEXT:    orl %ecx, %eax
1659; KNL-NEXT:    movq %rbp, %rsp
1660; KNL-NEXT:    popq %rbp
1661; KNL-NEXT:    vzeroupper
1662; KNL-NEXT:    retq
1663;
1664; SKX-LABEL: test_insertelement_variable_v32i1:
1665; SKX:       ## %bb.0:
1666; SKX-NEXT:    vptestmb %ymm0, %ymm0, %k0
1667; SKX-NEXT:    testb %dil, %dil
1668; SKX-NEXT:    setne %al
1669; SKX-NEXT:    vpbroadcastb %esi, %ymm0
1670; SKX-NEXT:    vpcmpeqb {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ymm0, %k1
1671; SKX-NEXT:    vpmovm2b %k0, %ymm0
1672; SKX-NEXT:    vpbroadcastb %eax, %ymm0 {%k1}
1673; SKX-NEXT:    vpsllw $7, %ymm0, %ymm0
1674; SKX-NEXT:    vpmovb2m %ymm0, %k0
1675; SKX-NEXT:    kmovd %k0, %eax
1676; SKX-NEXT:    vzeroupper
1677; SKX-NEXT:    retq
1678  %t1 = icmp ugt <32 x i8> %a, zeroinitializer
1679  %t2 = icmp ugt i8 %b, 0
1680  %t3 = insertelement <32 x i1> %t1, i1 %t2, i32 %index
1681  %t4 = bitcast <32 x i1> %t3 to i32
1682  ret i32 %t4
1683}
1684
1685define i64 @test_insertelement_variable_v64i1(<64 x i8> %a, i8 %b, i32 %index) nounwind {
1686; KNL-LABEL: test_insertelement_variable_v64i1:
1687; KNL:       ## %bb.0:
1688; KNL-NEXT:    pushq %rbp
1689; KNL-NEXT:    movq %rsp, %rbp
1690; KNL-NEXT:    andq $-64, %rsp
1691; KNL-NEXT:    subq $128, %rsp
1692; KNL-NEXT:    ## kill: def $esi killed $esi def $rsi
1693; KNL-NEXT:    vextracti64x4 $1, %zmm0, %ymm1
1694; KNL-NEXT:    vpxor %xmm2, %xmm2, %xmm2
1695; KNL-NEXT:    vpcmpeqb %ymm2, %ymm1, %ymm1
1696; KNL-NEXT:    vpcmpeqb %ymm2, %ymm0, %ymm0
1697; KNL-NEXT:    vinserti64x4 $1, %ymm1, %zmm0, %zmm0
1698; KNL-NEXT:    vpternlogq {{.*#+}} zmm0 = ~zmm0
1699; KNL-NEXT:    andl $63, %esi
1700; KNL-NEXT:    testb %dil, %dil
1701; KNL-NEXT:    vmovdqa64 %zmm0, (%rsp)
1702; KNL-NEXT:    setne (%rsp,%rsi)
1703; KNL-NEXT:    vpmovsxbd (%rsp), %zmm0
1704; KNL-NEXT:    vpslld $31, %zmm0, %zmm0
1705; KNL-NEXT:    vptestmd %zmm0, %zmm0, %k0
1706; KNL-NEXT:    kmovw %k0, %eax
1707; KNL-NEXT:    vpmovsxbd {{[0-9]+}}(%rsp), %zmm0
1708; KNL-NEXT:    vpslld $31, %zmm0, %zmm0
1709; KNL-NEXT:    vptestmd %zmm0, %zmm0, %k0
1710; KNL-NEXT:    kmovw %k0, %ecx
1711; KNL-NEXT:    shll $16, %ecx
1712; KNL-NEXT:    orl %eax, %ecx
1713; KNL-NEXT:    vpmovsxbd {{[0-9]+}}(%rsp), %zmm0
1714; KNL-NEXT:    vpslld $31, %zmm0, %zmm0
1715; KNL-NEXT:    vptestmd %zmm0, %zmm0, %k0
1716; KNL-NEXT:    kmovw %k0, %edx
1717; KNL-NEXT:    vpmovsxbd {{[0-9]+}}(%rsp), %zmm0
1718; KNL-NEXT:    vpslld $31, %zmm0, %zmm0
1719; KNL-NEXT:    vptestmd %zmm0, %zmm0, %k0
1720; KNL-NEXT:    kmovw %k0, %eax
1721; KNL-NEXT:    shll $16, %eax
1722; KNL-NEXT:    orl %edx, %eax
1723; KNL-NEXT:    shlq $32, %rax
1724; KNL-NEXT:    orq %rcx, %rax
1725; KNL-NEXT:    movq %rbp, %rsp
1726; KNL-NEXT:    popq %rbp
1727; KNL-NEXT:    vzeroupper
1728; KNL-NEXT:    retq
1729;
1730; SKX-LABEL: test_insertelement_variable_v64i1:
1731; SKX:       ## %bb.0:
1732; SKX-NEXT:    vptestmb %zmm0, %zmm0, %k0
1733; SKX-NEXT:    testb %dil, %dil
1734; SKX-NEXT:    setne %al
1735; SKX-NEXT:    vpbroadcastb %esi, %zmm0
1736; SKX-NEXT:    vpcmpeqb {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %zmm0, %k1
1737; SKX-NEXT:    vpmovm2b %k0, %zmm0
1738; SKX-NEXT:    vpbroadcastb %eax, %zmm0 {%k1}
1739; SKX-NEXT:    vpsllw $7, %zmm0, %zmm0
1740; SKX-NEXT:    vpmovb2m %zmm0, %k0
1741; SKX-NEXT:    kmovq %k0, %rax
1742; SKX-NEXT:    vzeroupper
1743; SKX-NEXT:    retq
1744  %t1 = icmp ugt <64 x i8> %a, zeroinitializer
1745  %t2 = icmp ugt i8 %b, 0
1746  %t3 = insertelement <64 x i1> %t1, i1 %t2, i32 %index
1747  %t4 = bitcast <64 x i1> %t3 to i64
1748  ret i64 %t4
1749}
1750
1751define i96 @test_insertelement_variable_v96i1(<96 x i8> %a, i8 %b, i32 %index) nounwind {
1752; KNL-LABEL: test_insertelement_variable_v96i1:
1753; KNL:       ## %bb.0:
1754; KNL-NEXT:    pushq %rbp
1755; KNL-NEXT:    movq %rsp, %rbp
1756; KNL-NEXT:    andq $-64, %rsp
1757; KNL-NEXT:    subq $192, %rsp
1758; KNL-NEXT:    movl 744(%rbp), %eax
1759; KNL-NEXT:    andl $127, %eax
1760; KNL-NEXT:    vmovd {{.*#+}} xmm0 = mem[0],zero,zero,zero
1761; KNL-NEXT:    vpinsrb $1, 232(%rbp), %xmm0, %xmm0
1762; KNL-NEXT:    vpinsrb $2, 240(%rbp), %xmm0, %xmm0
1763; KNL-NEXT:    vpinsrb $3, 248(%rbp), %xmm0, %xmm0
1764; KNL-NEXT:    vpinsrb $4, 256(%rbp), %xmm0, %xmm0
1765; KNL-NEXT:    vpinsrb $5, 264(%rbp), %xmm0, %xmm0
1766; KNL-NEXT:    vpinsrb $6, 272(%rbp), %xmm0, %xmm0
1767; KNL-NEXT:    vpinsrb $7, 280(%rbp), %xmm0, %xmm0
1768; KNL-NEXT:    vpinsrb $8, 288(%rbp), %xmm0, %xmm0
1769; KNL-NEXT:    vpinsrb $9, 296(%rbp), %xmm0, %xmm0
1770; KNL-NEXT:    vpinsrb $10, 304(%rbp), %xmm0, %xmm0
1771; KNL-NEXT:    vpinsrb $11, 312(%rbp), %xmm0, %xmm0
1772; KNL-NEXT:    vpinsrb $12, 320(%rbp), %xmm0, %xmm0
1773; KNL-NEXT:    vpinsrb $13, 328(%rbp), %xmm0, %xmm0
1774; KNL-NEXT:    vpinsrb $14, 336(%rbp), %xmm0, %xmm0
1775; KNL-NEXT:    vpinsrb $15, 344(%rbp), %xmm0, %xmm0
1776; KNL-NEXT:    vmovd {{.*#+}} xmm1 = mem[0],zero,zero,zero
1777; KNL-NEXT:    vpinsrb $1, 360(%rbp), %xmm1, %xmm1
1778; KNL-NEXT:    vpinsrb $2, 368(%rbp), %xmm1, %xmm1
1779; KNL-NEXT:    vpinsrb $3, 376(%rbp), %xmm1, %xmm1
1780; KNL-NEXT:    vpinsrb $4, 384(%rbp), %xmm1, %xmm1
1781; KNL-NEXT:    vpinsrb $5, 392(%rbp), %xmm1, %xmm1
1782; KNL-NEXT:    vpinsrb $6, 400(%rbp), %xmm1, %xmm1
1783; KNL-NEXT:    vpinsrb $7, 408(%rbp), %xmm1, %xmm1
1784; KNL-NEXT:    vpinsrb $8, 416(%rbp), %xmm1, %xmm1
1785; KNL-NEXT:    vpinsrb $9, 424(%rbp), %xmm1, %xmm1
1786; KNL-NEXT:    vpinsrb $10, 432(%rbp), %xmm1, %xmm1
1787; KNL-NEXT:    vpinsrb $11, 440(%rbp), %xmm1, %xmm1
1788; KNL-NEXT:    vpinsrb $12, 448(%rbp), %xmm1, %xmm1
1789; KNL-NEXT:    vpinsrb $13, 456(%rbp), %xmm1, %xmm1
1790; KNL-NEXT:    vpinsrb $14, 464(%rbp), %xmm1, %xmm1
1791; KNL-NEXT:    vpinsrb $15, 472(%rbp), %xmm1, %xmm1
1792; KNL-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm1
1793; KNL-NEXT:    vpxor %xmm0, %xmm0, %xmm0
1794; KNL-NEXT:    vpcmpeqb %ymm0, %ymm1, %ymm1
1795; KNL-NEXT:    vmovd %edi, %xmm2
1796; KNL-NEXT:    vpinsrb $1, %esi, %xmm2, %xmm2
1797; KNL-NEXT:    vpinsrb $2, %edx, %xmm2, %xmm2
1798; KNL-NEXT:    vpinsrb $3, %ecx, %xmm2, %xmm2
1799; KNL-NEXT:    vpinsrb $4, %r8d, %xmm2, %xmm2
1800; KNL-NEXT:    vpinsrb $5, %r9d, %xmm2, %xmm2
1801; KNL-NEXT:    vpinsrb $6, 16(%rbp), %xmm2, %xmm2
1802; KNL-NEXT:    vpinsrb $7, 24(%rbp), %xmm2, %xmm2
1803; KNL-NEXT:    vpinsrb $8, 32(%rbp), %xmm2, %xmm2
1804; KNL-NEXT:    vpinsrb $9, 40(%rbp), %xmm2, %xmm2
1805; KNL-NEXT:    vpinsrb $10, 48(%rbp), %xmm2, %xmm2
1806; KNL-NEXT:    vpinsrb $11, 56(%rbp), %xmm2, %xmm2
1807; KNL-NEXT:    vpinsrb $12, 64(%rbp), %xmm2, %xmm2
1808; KNL-NEXT:    vpinsrb $13, 72(%rbp), %xmm2, %xmm2
1809; KNL-NEXT:    vpinsrb $14, 80(%rbp), %xmm2, %xmm2
1810; KNL-NEXT:    vpinsrb $15, 88(%rbp), %xmm2, %xmm2
1811; KNL-NEXT:    vmovd {{.*#+}} xmm3 = mem[0],zero,zero,zero
1812; KNL-NEXT:    vpinsrb $1, 104(%rbp), %xmm3, %xmm3
1813; KNL-NEXT:    vpinsrb $2, 112(%rbp), %xmm3, %xmm3
1814; KNL-NEXT:    vpinsrb $3, 120(%rbp), %xmm3, %xmm3
1815; KNL-NEXT:    vpinsrb $4, 128(%rbp), %xmm3, %xmm3
1816; KNL-NEXT:    vpinsrb $5, 136(%rbp), %xmm3, %xmm3
1817; KNL-NEXT:    vpinsrb $6, 144(%rbp), %xmm3, %xmm3
1818; KNL-NEXT:    vpinsrb $7, 152(%rbp), %xmm3, %xmm3
1819; KNL-NEXT:    vpinsrb $8, 160(%rbp), %xmm3, %xmm3
1820; KNL-NEXT:    vpinsrb $9, 168(%rbp), %xmm3, %xmm3
1821; KNL-NEXT:    vpinsrb $10, 176(%rbp), %xmm3, %xmm3
1822; KNL-NEXT:    vpinsrb $11, 184(%rbp), %xmm3, %xmm3
1823; KNL-NEXT:    vpinsrb $12, 192(%rbp), %xmm3, %xmm3
1824; KNL-NEXT:    vpinsrb $13, 200(%rbp), %xmm3, %xmm3
1825; KNL-NEXT:    vpinsrb $14, 208(%rbp), %xmm3, %xmm3
1826; KNL-NEXT:    vpinsrb $15, 216(%rbp), %xmm3, %xmm3
1827; KNL-NEXT:    vinserti128 $1, %xmm3, %ymm2, %ymm2
1828; KNL-NEXT:    vpcmpeqb %ymm0, %ymm2, %ymm2
1829; KNL-NEXT:    vinserti64x4 $1, %ymm1, %zmm2, %zmm1
1830; KNL-NEXT:    vpternlogq {{.*#+}} zmm1 = ~zmm1
1831; KNL-NEXT:    vmovd {{.*#+}} xmm2 = mem[0],zero,zero,zero
1832; KNL-NEXT:    vpinsrb $1, 488(%rbp), %xmm2, %xmm2
1833; KNL-NEXT:    vpinsrb $2, 496(%rbp), %xmm2, %xmm2
1834; KNL-NEXT:    vpinsrb $3, 504(%rbp), %xmm2, %xmm2
1835; KNL-NEXT:    vpinsrb $4, 512(%rbp), %xmm2, %xmm2
1836; KNL-NEXT:    vpinsrb $5, 520(%rbp), %xmm2, %xmm2
1837; KNL-NEXT:    vpinsrb $6, 528(%rbp), %xmm2, %xmm2
1838; KNL-NEXT:    vpinsrb $7, 536(%rbp), %xmm2, %xmm2
1839; KNL-NEXT:    vpinsrb $8, 544(%rbp), %xmm2, %xmm2
1840; KNL-NEXT:    vpinsrb $9, 552(%rbp), %xmm2, %xmm2
1841; KNL-NEXT:    vpinsrb $10, 560(%rbp), %xmm2, %xmm2
1842; KNL-NEXT:    vpinsrb $11, 568(%rbp), %xmm2, %xmm2
1843; KNL-NEXT:    vpinsrb $12, 576(%rbp), %xmm2, %xmm2
1844; KNL-NEXT:    vpinsrb $13, 584(%rbp), %xmm2, %xmm2
1845; KNL-NEXT:    vpinsrb $14, 592(%rbp), %xmm2, %xmm2
1846; KNL-NEXT:    vpinsrb $15, 600(%rbp), %xmm2, %xmm2
1847; KNL-NEXT:    vmovd {{.*#+}} xmm3 = mem[0],zero,zero,zero
1848; KNL-NEXT:    vpinsrb $1, 616(%rbp), %xmm3, %xmm3
1849; KNL-NEXT:    vpinsrb $2, 624(%rbp), %xmm3, %xmm3
1850; KNL-NEXT:    vpinsrb $3, 632(%rbp), %xmm3, %xmm3
1851; KNL-NEXT:    vpinsrb $4, 640(%rbp), %xmm3, %xmm3
1852; KNL-NEXT:    vpinsrb $5, 648(%rbp), %xmm3, %xmm3
1853; KNL-NEXT:    vpinsrb $6, 656(%rbp), %xmm3, %xmm3
1854; KNL-NEXT:    vpinsrb $7, 664(%rbp), %xmm3, %xmm3
1855; KNL-NEXT:    vpinsrb $8, 672(%rbp), %xmm3, %xmm3
1856; KNL-NEXT:    vpinsrb $9, 680(%rbp), %xmm3, %xmm3
1857; KNL-NEXT:    vpinsrb $10, 688(%rbp), %xmm3, %xmm3
1858; KNL-NEXT:    vpinsrb $11, 696(%rbp), %xmm3, %xmm3
1859; KNL-NEXT:    vpinsrb $12, 704(%rbp), %xmm3, %xmm3
1860; KNL-NEXT:    vpinsrb $13, 712(%rbp), %xmm3, %xmm3
1861; KNL-NEXT:    vpinsrb $14, 720(%rbp), %xmm3, %xmm3
1862; KNL-NEXT:    vpinsrb $15, 728(%rbp), %xmm3, %xmm3
1863; KNL-NEXT:    vinserti128 $1, %xmm3, %ymm2, %ymm2
1864; KNL-NEXT:    vpcmpeqb %ymm0, %ymm2, %ymm2
1865; KNL-NEXT:    vpternlogq {{.*#+}} zmm2 = ~zmm2
1866; KNL-NEXT:    cmpb $0, 736(%rbp)
1867; KNL-NEXT:    vmovdqa %ymm0, {{[0-9]+}}(%rsp)
1868; KNL-NEXT:    vmovdqa %ymm2, {{[0-9]+}}(%rsp)
1869; KNL-NEXT:    vmovdqa64 %zmm1, (%rsp)
1870; KNL-NEXT:    setne (%rsp,%rax)
1871; KNL-NEXT:    vpmovsxbd (%rsp), %zmm0
1872; KNL-NEXT:    vpslld $31, %zmm0, %zmm0
1873; KNL-NEXT:    vptestmd %zmm0, %zmm0, %k0
1874; KNL-NEXT:    kmovw %k0, %eax
1875; KNL-NEXT:    vpmovsxbd {{[0-9]+}}(%rsp), %zmm0
1876; KNL-NEXT:    vpslld $31, %zmm0, %zmm0
1877; KNL-NEXT:    vptestmd %zmm0, %zmm0, %k0
1878; KNL-NEXT:    kmovw %k0, %ecx
1879; KNL-NEXT:    shll $16, %ecx
1880; KNL-NEXT:    orl %eax, %ecx
1881; KNL-NEXT:    vpmovsxbd {{[0-9]+}}(%rsp), %zmm0
1882; KNL-NEXT:    vpslld $31, %zmm0, %zmm0
1883; KNL-NEXT:    vptestmd %zmm0, %zmm0, %k0
1884; KNL-NEXT:    kmovw %k0, %edx
1885; KNL-NEXT:    vpmovsxbd {{[0-9]+}}(%rsp), %zmm0
1886; KNL-NEXT:    vpslld $31, %zmm0, %zmm0
1887; KNL-NEXT:    vptestmd %zmm0, %zmm0, %k0
1888; KNL-NEXT:    kmovw %k0, %eax
1889; KNL-NEXT:    shll $16, %eax
1890; KNL-NEXT:    orl %edx, %eax
1891; KNL-NEXT:    shlq $32, %rax
1892; KNL-NEXT:    orq %rcx, %rax
1893; KNL-NEXT:    vpmovsxbd {{[0-9]+}}(%rsp), %zmm0
1894; KNL-NEXT:    vpslld $31, %zmm0, %zmm0
1895; KNL-NEXT:    vptestmd %zmm0, %zmm0, %k0
1896; KNL-NEXT:    kmovw %k0, %ecx
1897; KNL-NEXT:    vpmovsxbd {{[0-9]+}}(%rsp), %zmm0
1898; KNL-NEXT:    vpslld $31, %zmm0, %zmm0
1899; KNL-NEXT:    vptestmd %zmm0, %zmm0, %k0
1900; KNL-NEXT:    kmovw %k0, %esi
1901; KNL-NEXT:    shll $16, %esi
1902; KNL-NEXT:    orl %ecx, %esi
1903; KNL-NEXT:    vpmovsxbd {{[0-9]+}}(%rsp), %zmm0
1904; KNL-NEXT:    vpslld $31, %zmm0, %zmm0
1905; KNL-NEXT:    vptestmd %zmm0, %zmm0, %k0
1906; KNL-NEXT:    kmovw %k0, %ecx
1907; KNL-NEXT:    vpmovsxbd {{[0-9]+}}(%rsp), %zmm0
1908; KNL-NEXT:    vpslld $31, %zmm0, %zmm0
1909; KNL-NEXT:    vptestmd %zmm0, %zmm0, %k0
1910; KNL-NEXT:    kmovw %k0, %edx
1911; KNL-NEXT:    shll $16, %edx
1912; KNL-NEXT:    orl %ecx, %edx
1913; KNL-NEXT:    shlq $32, %rdx
1914; KNL-NEXT:    orq %rsi, %rdx
1915; KNL-NEXT:    movq %rbp, %rsp
1916; KNL-NEXT:    popq %rbp
1917; KNL-NEXT:    vzeroupper
1918; KNL-NEXT:    retq
1919;
1920; SKX-LABEL: test_insertelement_variable_v96i1:
1921; SKX:       ## %bb.0:
1922; SKX-NEXT:    pushq %rbp
1923; SKX-NEXT:    movq %rsp, %rbp
1924; SKX-NEXT:    andq $-64, %rsp
1925; SKX-NEXT:    subq $192, %rsp
1926; SKX-NEXT:    vmovd {{.*#+}} xmm0 = mem[0],zero,zero,zero
1927; SKX-NEXT:    vpinsrb $1, 232(%rbp), %xmm0, %xmm0
1928; SKX-NEXT:    vpinsrb $2, 240(%rbp), %xmm0, %xmm0
1929; SKX-NEXT:    vpinsrb $3, 248(%rbp), %xmm0, %xmm0
1930; SKX-NEXT:    vpinsrb $4, 256(%rbp), %xmm0, %xmm0
1931; SKX-NEXT:    vpinsrb $5, 264(%rbp), %xmm0, %xmm0
1932; SKX-NEXT:    vpinsrb $6, 272(%rbp), %xmm0, %xmm0
1933; SKX-NEXT:    vpinsrb $7, 280(%rbp), %xmm0, %xmm0
1934; SKX-NEXT:    vpinsrb $8, 288(%rbp), %xmm0, %xmm0
1935; SKX-NEXT:    vpinsrb $9, 296(%rbp), %xmm0, %xmm0
1936; SKX-NEXT:    vpinsrb $10, 304(%rbp), %xmm0, %xmm0
1937; SKX-NEXT:    vpinsrb $11, 312(%rbp), %xmm0, %xmm0
1938; SKX-NEXT:    vpinsrb $12, 320(%rbp), %xmm0, %xmm0
1939; SKX-NEXT:    vpinsrb $13, 328(%rbp), %xmm0, %xmm0
1940; SKX-NEXT:    vpinsrb $14, 336(%rbp), %xmm0, %xmm0
1941; SKX-NEXT:    vpinsrb $15, 344(%rbp), %xmm0, %xmm0
1942; SKX-NEXT:    vmovd {{.*#+}} xmm1 = mem[0],zero,zero,zero
1943; SKX-NEXT:    vpinsrb $1, 360(%rbp), %xmm1, %xmm1
1944; SKX-NEXT:    vpinsrb $2, 368(%rbp), %xmm1, %xmm1
1945; SKX-NEXT:    vpinsrb $3, 376(%rbp), %xmm1, %xmm1
1946; SKX-NEXT:    vpinsrb $4, 384(%rbp), %xmm1, %xmm1
1947; SKX-NEXT:    vpinsrb $5, 392(%rbp), %xmm1, %xmm1
1948; SKX-NEXT:    vpinsrb $6, 400(%rbp), %xmm1, %xmm1
1949; SKX-NEXT:    vpinsrb $7, 408(%rbp), %xmm1, %xmm1
1950; SKX-NEXT:    vpinsrb $8, 416(%rbp), %xmm1, %xmm1
1951; SKX-NEXT:    vpinsrb $9, 424(%rbp), %xmm1, %xmm1
1952; SKX-NEXT:    vpinsrb $10, 432(%rbp), %xmm1, %xmm1
1953; SKX-NEXT:    vpinsrb $11, 440(%rbp), %xmm1, %xmm1
1954; SKX-NEXT:    vpinsrb $12, 448(%rbp), %xmm1, %xmm1
1955; SKX-NEXT:    vpinsrb $13, 456(%rbp), %xmm1, %xmm1
1956; SKX-NEXT:    vpinsrb $14, 464(%rbp), %xmm1, %xmm1
1957; SKX-NEXT:    vpinsrb $15, 472(%rbp), %xmm1, %xmm1
1958; SKX-NEXT:    vinserti128 $1, %xmm1, %ymm0, %ymm0
1959; SKX-NEXT:    vmovd %edi, %xmm1
1960; SKX-NEXT:    vpinsrb $1, %esi, %xmm1, %xmm1
1961; SKX-NEXT:    vpinsrb $2, %edx, %xmm1, %xmm1
1962; SKX-NEXT:    vpinsrb $3, %ecx, %xmm1, %xmm1
1963; SKX-NEXT:    vpinsrb $4, %r8d, %xmm1, %xmm1
1964; SKX-NEXT:    vpinsrb $5, %r9d, %xmm1, %xmm1
1965; SKX-NEXT:    vpinsrb $6, 16(%rbp), %xmm1, %xmm1
1966; SKX-NEXT:    vpinsrb $7, 24(%rbp), %xmm1, %xmm1
1967; SKX-NEXT:    vpinsrb $8, 32(%rbp), %xmm1, %xmm1
1968; SKX-NEXT:    vpinsrb $9, 40(%rbp), %xmm1, %xmm1
1969; SKX-NEXT:    vpinsrb $10, 48(%rbp), %xmm1, %xmm1
1970; SKX-NEXT:    vpinsrb $11, 56(%rbp), %xmm1, %xmm1
1971; SKX-NEXT:    vpinsrb $12, 64(%rbp), %xmm1, %xmm1
1972; SKX-NEXT:    vpinsrb $13, 72(%rbp), %xmm1, %xmm1
1973; SKX-NEXT:    vpinsrb $14, 80(%rbp), %xmm1, %xmm1
1974; SKX-NEXT:    vpinsrb $15, 88(%rbp), %xmm1, %xmm1
1975; SKX-NEXT:    vmovd {{.*#+}} xmm2 = mem[0],zero,zero,zero
1976; SKX-NEXT:    vpinsrb $1, 104(%rbp), %xmm2, %xmm2
1977; SKX-NEXT:    vpinsrb $2, 112(%rbp), %xmm2, %xmm2
1978; SKX-NEXT:    vpinsrb $3, 120(%rbp), %xmm2, %xmm2
1979; SKX-NEXT:    vpinsrb $4, 128(%rbp), %xmm2, %xmm2
1980; SKX-NEXT:    vpinsrb $5, 136(%rbp), %xmm2, %xmm2
1981; SKX-NEXT:    vpinsrb $6, 144(%rbp), %xmm2, %xmm2
1982; SKX-NEXT:    vpinsrb $7, 152(%rbp), %xmm2, %xmm2
1983; SKX-NEXT:    vpinsrb $8, 160(%rbp), %xmm2, %xmm2
1984; SKX-NEXT:    vpinsrb $9, 168(%rbp), %xmm2, %xmm2
1985; SKX-NEXT:    vpinsrb $10, 176(%rbp), %xmm2, %xmm2
1986; SKX-NEXT:    vpinsrb $11, 184(%rbp), %xmm2, %xmm2
1987; SKX-NEXT:    vpinsrb $12, 192(%rbp), %xmm2, %xmm2
1988; SKX-NEXT:    vpinsrb $13, 200(%rbp), %xmm2, %xmm2
1989; SKX-NEXT:    vpinsrb $14, 208(%rbp), %xmm2, %xmm2
1990; SKX-NEXT:    vpinsrb $15, 216(%rbp), %xmm2, %xmm2
1991; SKX-NEXT:    vinserti128 $1, %xmm2, %ymm1, %ymm1
1992; SKX-NEXT:    vinserti64x4 $1, %ymm0, %zmm1, %zmm0
1993; SKX-NEXT:    vmovd {{.*#+}} xmm1 = mem[0],zero,zero,zero
1994; SKX-NEXT:    vpinsrb $1, 488(%rbp), %xmm1, %xmm1
1995; SKX-NEXT:    vpinsrb $2, 496(%rbp), %xmm1, %xmm1
1996; SKX-NEXT:    vpinsrb $3, 504(%rbp), %xmm1, %xmm1
1997; SKX-NEXT:    vpinsrb $4, 512(%rbp), %xmm1, %xmm1
1998; SKX-NEXT:    vpinsrb $5, 520(%rbp), %xmm1, %xmm1
1999; SKX-NEXT:    vpinsrb $6, 528(%rbp), %xmm1, %xmm1
2000; SKX-NEXT:    vpinsrb $7, 536(%rbp), %xmm1, %xmm1
2001; SKX-NEXT:    vpinsrb $8, 544(%rbp), %xmm1, %xmm1
2002; SKX-NEXT:    vpinsrb $9, 552(%rbp), %xmm1, %xmm1
2003; SKX-NEXT:    vpinsrb $10, 560(%rbp), %xmm1, %xmm1
2004; SKX-NEXT:    vpinsrb $11, 568(%rbp), %xmm1, %xmm1
2005; SKX-NEXT:    vpinsrb $12, 576(%rbp), %xmm1, %xmm1
2006; SKX-NEXT:    vpinsrb $13, 584(%rbp), %xmm1, %xmm1
2007; SKX-NEXT:    vpinsrb $14, 592(%rbp), %xmm1, %xmm1
2008; SKX-NEXT:    vpinsrb $15, 600(%rbp), %xmm1, %xmm1
2009; SKX-NEXT:    vmovd {{.*#+}} xmm2 = mem[0],zero,zero,zero
2010; SKX-NEXT:    vpinsrb $1, 616(%rbp), %xmm2, %xmm2
2011; SKX-NEXT:    vpinsrb $2, 624(%rbp), %xmm2, %xmm2
2012; SKX-NEXT:    vpinsrb $3, 632(%rbp), %xmm2, %xmm2
2013; SKX-NEXT:    vpinsrb $4, 640(%rbp), %xmm2, %xmm2
2014; SKX-NEXT:    vpinsrb $5, 648(%rbp), %xmm2, %xmm2
2015; SKX-NEXT:    vpinsrb $6, 656(%rbp), %xmm2, %xmm2
2016; SKX-NEXT:    vpinsrb $7, 664(%rbp), %xmm2, %xmm2
2017; SKX-NEXT:    vpinsrb $8, 672(%rbp), %xmm2, %xmm2
2018; SKX-NEXT:    vpinsrb $9, 680(%rbp), %xmm2, %xmm2
2019; SKX-NEXT:    vpinsrb $10, 688(%rbp), %xmm2, %xmm2
2020; SKX-NEXT:    vpinsrb $11, 696(%rbp), %xmm2, %xmm2
2021; SKX-NEXT:    vpinsrb $12, 704(%rbp), %xmm2, %xmm2
2022; SKX-NEXT:    vpinsrb $13, 712(%rbp), %xmm2, %xmm2
2023; SKX-NEXT:    vpinsrb $14, 720(%rbp), %xmm2, %xmm2
2024; SKX-NEXT:    vpinsrb $15, 728(%rbp), %xmm2, %xmm2
2025; SKX-NEXT:    vinserti128 $1, %xmm2, %ymm1, %ymm1
2026; SKX-NEXT:    movl 744(%rbp), %eax
2027; SKX-NEXT:    andl $127, %eax
2028; SKX-NEXT:    vptestmb %zmm0, %zmm0, %k0
2029; SKX-NEXT:    vptestmb %zmm1, %zmm1, %k1
2030; SKX-NEXT:    cmpb $0, 736(%rbp)
2031; SKX-NEXT:    vpmovm2b %k1, %zmm0
2032; SKX-NEXT:    vmovdqa64 %zmm0, {{[0-9]+}}(%rsp)
2033; SKX-NEXT:    vpmovm2b %k0, %zmm0
2034; SKX-NEXT:    vmovdqa64 %zmm0, (%rsp)
2035; SKX-NEXT:    setne (%rsp,%rax)
2036; SKX-NEXT:    vpsllw $7, {{[0-9]+}}(%rsp), %zmm0
2037; SKX-NEXT:    vpmovb2m %zmm0, %k0
2038; SKX-NEXT:    vpsllw $7, (%rsp), %zmm0
2039; SKX-NEXT:    vpmovb2m %zmm0, %k1
2040; SKX-NEXT:    kmovq %k1, %rax
2041; SKX-NEXT:    kmovq %k0, %rdx
2042; SKX-NEXT:    movq %rbp, %rsp
2043; SKX-NEXT:    popq %rbp
2044; SKX-NEXT:    vzeroupper
2045; SKX-NEXT:    retq
2046  %t1 = icmp ugt <96 x i8> %a, zeroinitializer
2047  %t2 = icmp ugt i8 %b, 0
2048  %t3 = insertelement <96 x i1> %t1, i1 %t2, i32 %index
2049  %t4 = bitcast <96 x i1> %t3 to i96
2050  ret i96 %t4
2051}
2052
2053define i128 @test_insertelement_variable_v128i1(<128 x i8> %a, i8 %b, i32 %index) nounwind {
2054; KNL-LABEL: test_insertelement_variable_v128i1:
2055; KNL:       ## %bb.0:
2056; KNL-NEXT:    pushq %rbp
2057; KNL-NEXT:    movq %rsp, %rbp
2058; KNL-NEXT:    andq $-64, %rsp
2059; KNL-NEXT:    subq $192, %rsp
2060; KNL-NEXT:    ## kill: def $esi killed $esi def $rsi
2061; KNL-NEXT:    vextracti64x4 $1, %zmm0, %ymm2
2062; KNL-NEXT:    vpxor %xmm3, %xmm3, %xmm3
2063; KNL-NEXT:    vpcmpeqb %ymm3, %ymm2, %ymm2
2064; KNL-NEXT:    vpcmpeqb %ymm3, %ymm0, %ymm0
2065; KNL-NEXT:    vinserti64x4 $1, %ymm2, %zmm0, %zmm0
2066; KNL-NEXT:    vpternlogq {{.*#+}} zmm0 = ~zmm0
2067; KNL-NEXT:    vextracti64x4 $1, %zmm1, %ymm2
2068; KNL-NEXT:    vpcmpeqb %ymm3, %ymm2, %ymm2
2069; KNL-NEXT:    vpcmpeqb %ymm3, %ymm1, %ymm1
2070; KNL-NEXT:    vinserti64x4 $1, %ymm2, %zmm1, %zmm1
2071; KNL-NEXT:    vpternlogq {{.*#+}} zmm1 = ~zmm1
2072; KNL-NEXT:    andl $127, %esi
2073; KNL-NEXT:    testb %dil, %dil
2074; KNL-NEXT:    vmovdqa64 %zmm1, {{[0-9]+}}(%rsp)
2075; KNL-NEXT:    vmovdqa64 %zmm0, (%rsp)
2076; KNL-NEXT:    setne (%rsp,%rsi)
2077; KNL-NEXT:    vpmovsxbd (%rsp), %zmm0
2078; KNL-NEXT:    vpslld $31, %zmm0, %zmm0
2079; KNL-NEXT:    vptestmd %zmm0, %zmm0, %k0
2080; KNL-NEXT:    kmovw %k0, %eax
2081; KNL-NEXT:    vpmovsxbd {{[0-9]+}}(%rsp), %zmm0
2082; KNL-NEXT:    vpslld $31, %zmm0, %zmm0
2083; KNL-NEXT:    vptestmd %zmm0, %zmm0, %k0
2084; KNL-NEXT:    kmovw %k0, %ecx
2085; KNL-NEXT:    shll $16, %ecx
2086; KNL-NEXT:    orl %eax, %ecx
2087; KNL-NEXT:    vpmovsxbd {{[0-9]+}}(%rsp), %zmm0
2088; KNL-NEXT:    vpslld $31, %zmm0, %zmm0
2089; KNL-NEXT:    vptestmd %zmm0, %zmm0, %k0
2090; KNL-NEXT:    kmovw %k0, %edx
2091; KNL-NEXT:    vpmovsxbd {{[0-9]+}}(%rsp), %zmm0
2092; KNL-NEXT:    vpslld $31, %zmm0, %zmm0
2093; KNL-NEXT:    vptestmd %zmm0, %zmm0, %k0
2094; KNL-NEXT:    kmovw %k0, %eax
2095; KNL-NEXT:    shll $16, %eax
2096; KNL-NEXT:    orl %edx, %eax
2097; KNL-NEXT:    shlq $32, %rax
2098; KNL-NEXT:    orq %rcx, %rax
2099; KNL-NEXT:    vpmovsxbd {{[0-9]+}}(%rsp), %zmm0
2100; KNL-NEXT:    vpslld $31, %zmm0, %zmm0
2101; KNL-NEXT:    vptestmd %zmm0, %zmm0, %k0
2102; KNL-NEXT:    kmovw %k0, %ecx
2103; KNL-NEXT:    vpmovsxbd {{[0-9]+}}(%rsp), %zmm0
2104; KNL-NEXT:    vpslld $31, %zmm0, %zmm0
2105; KNL-NEXT:    vptestmd %zmm0, %zmm0, %k0
2106; KNL-NEXT:    kmovw %k0, %esi
2107; KNL-NEXT:    shll $16, %esi
2108; KNL-NEXT:    orl %ecx, %esi
2109; KNL-NEXT:    vpmovsxbd {{[0-9]+}}(%rsp), %zmm0
2110; KNL-NEXT:    vpslld $31, %zmm0, %zmm0
2111; KNL-NEXT:    vptestmd %zmm0, %zmm0, %k0
2112; KNL-NEXT:    kmovw %k0, %ecx
2113; KNL-NEXT:    vpmovsxbd {{[0-9]+}}(%rsp), %zmm0
2114; KNL-NEXT:    vpslld $31, %zmm0, %zmm0
2115; KNL-NEXT:    vptestmd %zmm0, %zmm0, %k0
2116; KNL-NEXT:    kmovw %k0, %edx
2117; KNL-NEXT:    shll $16, %edx
2118; KNL-NEXT:    orl %ecx, %edx
2119; KNL-NEXT:    shlq $32, %rdx
2120; KNL-NEXT:    orq %rsi, %rdx
2121; KNL-NEXT:    movq %rbp, %rsp
2122; KNL-NEXT:    popq %rbp
2123; KNL-NEXT:    vzeroupper
2124; KNL-NEXT:    retq
2125;
2126; SKX-LABEL: test_insertelement_variable_v128i1:
2127; SKX:       ## %bb.0:
2128; SKX-NEXT:    pushq %rbp
2129; SKX-NEXT:    movq %rsp, %rbp
2130; SKX-NEXT:    andq $-64, %rsp
2131; SKX-NEXT:    subq $192, %rsp
2132; SKX-NEXT:    ## kill: def $esi killed $esi def $rsi
2133; SKX-NEXT:    vptestmb %zmm0, %zmm0, %k0
2134; SKX-NEXT:    vptestmb %zmm1, %zmm1, %k1
2135; SKX-NEXT:    andl $127, %esi
2136; SKX-NEXT:    testb %dil, %dil
2137; SKX-NEXT:    vpmovm2b %k1, %zmm0
2138; SKX-NEXT:    vmovdqa64 %zmm0, {{[0-9]+}}(%rsp)
2139; SKX-NEXT:    vpmovm2b %k0, %zmm0
2140; SKX-NEXT:    vmovdqa64 %zmm0, (%rsp)
2141; SKX-NEXT:    setne (%rsp,%rsi)
2142; SKX-NEXT:    vpsllw $7, {{[0-9]+}}(%rsp), %zmm0
2143; SKX-NEXT:    vpmovb2m %zmm0, %k0
2144; SKX-NEXT:    vpsllw $7, (%rsp), %zmm0
2145; SKX-NEXT:    vpmovb2m %zmm0, %k1
2146; SKX-NEXT:    kmovq %k1, %rax
2147; SKX-NEXT:    kmovq %k0, %rdx
2148; SKX-NEXT:    movq %rbp, %rsp
2149; SKX-NEXT:    popq %rbp
2150; SKX-NEXT:    vzeroupper
2151; SKX-NEXT:    retq
2152  %t1 = icmp ugt <128 x i8> %a, zeroinitializer
2153  %t2 = icmp ugt i8 %b, 0
2154  %t3 = insertelement <128 x i1> %t1, i1 %t2, i32 %index
2155  %t4 = bitcast <128 x i1> %t3 to i128
2156  ret i128 %t4
2157}
2158
2159define void @test_concat_v2i1(ptr %arg, ptr %arg1, ptr %arg2) nounwind {
2160; KNL-LABEL: test_concat_v2i1:
2161; KNL:       ## %bb.0:
2162; KNL-NEXT:    vmovq {{.*#+}} xmm0 = mem[0],zero
2163; KNL-NEXT:    vcvtph2ps %xmm0, %xmm1
2164; KNL-NEXT:    vmovss {{.*#+}} xmm2 = [6.0E+0,0.0E+0,0.0E+0,0.0E+0]
2165; KNL-NEXT:    vucomiss %xmm2, %xmm1
2166; KNL-NEXT:    setb %al
2167; KNL-NEXT:    andl $1, %eax
2168; KNL-NEXT:    kmovw %eax, %k0
2169; KNL-NEXT:    vpshuflw {{.*#+}} xmm0 = xmm0[1,1,1,1,4,5,6,7]
2170; KNL-NEXT:    vcvtph2ps %xmm0, %xmm0
2171; KNL-NEXT:    vucomiss %xmm2, %xmm0
2172; KNL-NEXT:    setb %al
2173; KNL-NEXT:    kmovw %eax, %k1
2174; KNL-NEXT:    kshiftlw $1, %k1, %k1
2175; KNL-NEXT:    korw %k1, %k0, %k0
2176; KNL-NEXT:    vxorps %xmm2, %xmm2, %xmm2
2177; KNL-NEXT:    vucomiss %xmm2, %xmm1
2178; KNL-NEXT:    seta %al
2179; KNL-NEXT:    andl $1, %eax
2180; KNL-NEXT:    kmovw %eax, %k1
2181; KNL-NEXT:    vucomiss %xmm2, %xmm0
2182; KNL-NEXT:    seta %al
2183; KNL-NEXT:    kmovw %eax, %k2
2184; KNL-NEXT:    kshiftlw $1, %k2, %k2
2185; KNL-NEXT:    korw %k2, %k1, %k1
2186; KNL-NEXT:    kandw %k1, %k0, %k1
2187; KNL-NEXT:    vmovq {{.*#+}} xmm0 = mem[0],zero
2188; KNL-NEXT:    vpternlogd {{.*#+}} zmm1 {%k1} {z} = -1
2189; KNL-NEXT:    vpmovdw %zmm1, %ymm1
2190; KNL-NEXT:    vpand %xmm0, %xmm1, %xmm0
2191; KNL-NEXT:    vmovd %xmm0, (%rdx)
2192; KNL-NEXT:    vzeroupper
2193; KNL-NEXT:    retq
2194;
2195; SKX-LABEL: test_concat_v2i1:
2196; SKX:       ## %bb.0:
2197; SKX-NEXT:    vmovsd {{.*#+}} xmm0 = mem[0],zero
2198; SKX-NEXT:    vcvtph2ps %xmm0, %ymm0
2199; SKX-NEXT:    vcmpltps {{\.?LCPI[0-9]+_[0-9]+}}(%rip){1to8}, %ymm0, %k0
2200; SKX-NEXT:    vxorps %xmm1, %xmm1, %xmm1
2201; SKX-NEXT:    vcmpltps %ymm0, %ymm1, %k1
2202; SKX-NEXT:    kandw %k1, %k0, %k1
2203; SKX-NEXT:    vmovq {{.*#+}} xmm0 = mem[0],zero
2204; SKX-NEXT:    vmovdqu16 %xmm0, %xmm0 {%k1} {z}
2205; SKX-NEXT:    vmovd %xmm0, (%rdx)
2206; SKX-NEXT:    vzeroupper
2207; SKX-NEXT:    retq
2208  %tmp = load <2 x half>, ptr %arg, align 8
2209  %tmp3 = fcmp fast olt <2 x half> %tmp, <half 0xH4600, half 0xH4600>
2210  %tmp4 = fcmp fast ogt <2 x half> %tmp, zeroinitializer
2211  %tmp5 = and <2 x i1> %tmp3, %tmp4
2212  %tmp6 = load <2 x half>, ptr %arg1, align 8
2213  %tmp7 = select <2 x i1> %tmp5, <2 x half> %tmp6, <2 x half> zeroinitializer
2214  store <2 x half> %tmp7, ptr %arg2, align 8
2215  ret void
2216}
2217
2218; Check that an extracted bool element comparison is correctly extended when
2219; the bool vector is bitcasted instead.
2220define void @test_extractelement_v64i1_cmp0(<64 x i1> %0) nounwind {
2221; KNL-LABEL: test_extractelement_v64i1_cmp0:
2222; KNL:       ## %bb.0: ## %entry
2223; KNL-NEXT:    retq
2224;
2225; SKX-LABEL: test_extractelement_v64i1_cmp0:
2226; SKX:       ## %bb.0: ## %entry
2227; SKX-NEXT:    vzeroupper
2228; SKX-NEXT:    retq
2229entry:
2230  %1 = extractelement <64 x i1> %0, i64 0
2231  br i1 %1, label %common.ret, label %2
2232
2233common.ret:                                       ; preds = %2, %entry
2234  ret void
2235
22362:                                                ; preds = %entry
2237  %3 = extractelement <64 x i1> zeroinitializer, i64 0
2238  br label %common.ret
2239}
2240