1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
2; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+sse2      | FileCheck %s --check-prefix=SSE --check-prefix=SSE2
3; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+sse4.2    | FileCheck %s --check-prefix=SSE --check-prefix=SSE42
4; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx       | FileCheck %s --check-prefix=AVX --check-prefix=AVX1
5; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx2      | FileCheck %s --check-prefix=AVX --check-prefix=AVX2
6; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512f   | FileCheck %s --check-prefix=AVX --check-prefix=AVX512 --check-prefix=AVX512F
7
8;
9; General cases - packing of vector comparison to legal vector result types
10;
11
12define <16 x i8> @vselect_packss_v16i16(<16 x i16> %a0, <16 x i16> %a1, <16 x i8> %a2, <16 x i8> %a3) {
13; SSE2-LABEL: vselect_packss_v16i16:
14; SSE2:       # BB#0:
15; SSE2-NEXT:    pcmpeqw %xmm3, %xmm1
16; SSE2-NEXT:    pcmpeqw %xmm2, %xmm0
17; SSE2-NEXT:    packsswb %xmm1, %xmm0
18; SSE2-NEXT:    pand %xmm0, %xmm4
19; SSE2-NEXT:    pandn %xmm5, %xmm0
20; SSE2-NEXT:    por %xmm4, %xmm0
21; SSE2-NEXT:    retq
22;
23; SSE42-LABEL: vselect_packss_v16i16:
24; SSE42:       # BB#0:
25; SSE42-NEXT:    pcmpeqw %xmm3, %xmm1
26; SSE42-NEXT:    pcmpeqw %xmm2, %xmm0
27; SSE42-NEXT:    packsswb %xmm1, %xmm0
28; SSE42-NEXT:    pblendvb %xmm0, %xmm4, %xmm5
29; SSE42-NEXT:    movdqa %xmm5, %xmm0
30; SSE42-NEXT:    retq
31;
32; AVX1-LABEL: vselect_packss_v16i16:
33; AVX1:       # BB#0:
34; AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm4
35; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm5
36; AVX1-NEXT:    vpcmpeqw %xmm4, %xmm5, %xmm4
37; AVX1-NEXT:    vpcmpeqw %xmm1, %xmm0, %xmm0
38; AVX1-NEXT:    vpacksswb %xmm4, %xmm0, %xmm0
39; AVX1-NEXT:    vpblendvb %xmm0, %xmm2, %xmm3, %xmm0
40; AVX1-NEXT:    vzeroupper
41; AVX1-NEXT:    retq
42;
43; AVX2-LABEL: vselect_packss_v16i16:
44; AVX2:       # BB#0:
45; AVX2-NEXT:    vpcmpeqw %ymm1, %ymm0, %ymm0
46; AVX2-NEXT:    vextracti128 $1, %ymm0, %xmm1
47; AVX2-NEXT:    vpacksswb %xmm1, %xmm0, %xmm0
48; AVX2-NEXT:    vpblendvb %xmm0, %xmm2, %xmm3, %xmm0
49; AVX2-NEXT:    vzeroupper
50; AVX2-NEXT:    retq
51;
52; AVX512-LABEL: vselect_packss_v16i16:
53; AVX512:       # BB#0:
54; AVX512-NEXT:    vpcmpeqw %ymm1, %ymm0, %ymm0
55; AVX512-NEXT:    vpmovsxwd %ymm0, %zmm0
56; AVX512-NEXT:    vpmovdb %zmm0, %xmm0
57; AVX512-NEXT:    vpblendvb %xmm0, %xmm2, %xmm3, %xmm0
58; AVX512-NEXT:    vzeroupper
59; AVX512-NEXT:    retq
60  %1 = icmp eq <16 x i16> %a0, %a1
61  %2 = sext <16 x i1> %1 to <16 x i8>
62  %3 = and <16 x i8> %2, %a2
63  %4 = xor <16 x i8> %2, <i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1>
64  %5 = and <16 x i8> %4, %a3
65  %6 = or <16 x i8> %3, %5
66  ret <16 x i8> %6
67}
68
69define <16 x i8> @vselect_packss_v16i32(<16 x i32> %a0, <16 x i32> %a1, <16 x i8> %a2, <16 x i8> %a3) {
70; SSE2-LABEL: vselect_packss_v16i32:
71; SSE2:       # BB#0:
72; SSE2-NEXT:    pcmpeqd %xmm7, %xmm3
73; SSE2-NEXT:    pcmpeqd %xmm6, %xmm2
74; SSE2-NEXT:    packssdw %xmm3, %xmm2
75; SSE2-NEXT:    pcmpeqd %xmm5, %xmm1
76; SSE2-NEXT:    pcmpeqd %xmm4, %xmm0
77; SSE2-NEXT:    packssdw %xmm1, %xmm0
78; SSE2-NEXT:    packsswb %xmm2, %xmm0
79; SSE2-NEXT:    movdqa {{[0-9]+}}(%rsp), %xmm1
80; SSE2-NEXT:    pand %xmm0, %xmm1
81; SSE2-NEXT:    pandn {{[0-9]+}}(%rsp), %xmm0
82; SSE2-NEXT:    por %xmm1, %xmm0
83; SSE2-NEXT:    retq
84;
85; SSE42-LABEL: vselect_packss_v16i32:
86; SSE42:       # BB#0:
87; SSE42-NEXT:    movdqa {{[0-9]+}}(%rsp), %xmm8
88; SSE42-NEXT:    pcmpeqd %xmm7, %xmm3
89; SSE42-NEXT:    pcmpeqd %xmm6, %xmm2
90; SSE42-NEXT:    packssdw %xmm3, %xmm2
91; SSE42-NEXT:    pcmpeqd %xmm5, %xmm1
92; SSE42-NEXT:    pcmpeqd %xmm4, %xmm0
93; SSE42-NEXT:    packssdw %xmm1, %xmm0
94; SSE42-NEXT:    packsswb %xmm2, %xmm0
95; SSE42-NEXT:    pblendvb %xmm0, {{[0-9]+}}(%rsp), %xmm8
96; SSE42-NEXT:    movdqa %xmm8, %xmm0
97; SSE42-NEXT:    retq
98;
99; AVX1-LABEL: vselect_packss_v16i32:
100; AVX1:       # BB#0:
101; AVX1-NEXT:    vextractf128 $1, %ymm3, %xmm6
102; AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm7
103; AVX1-NEXT:    vpcmpeqd %xmm6, %xmm7, %xmm6
104; AVX1-NEXT:    vpcmpeqd %xmm3, %xmm1, %xmm1
105; AVX1-NEXT:    vpackssdw %xmm6, %xmm1, %xmm1
106; AVX1-NEXT:    vextractf128 $1, %ymm2, %xmm3
107; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm6
108; AVX1-NEXT:    vpcmpeqd %xmm3, %xmm6, %xmm3
109; AVX1-NEXT:    vpcmpeqd %xmm2, %xmm0, %xmm0
110; AVX1-NEXT:    vpackssdw %xmm3, %xmm0, %xmm0
111; AVX1-NEXT:    vpacksswb %xmm1, %xmm0, %xmm0
112; AVX1-NEXT:    vpblendvb %xmm0, %xmm4, %xmm5, %xmm0
113; AVX1-NEXT:    vzeroupper
114; AVX1-NEXT:    retq
115;
116; AVX2-LABEL: vselect_packss_v16i32:
117; AVX2:       # BB#0:
118; AVX2-NEXT:    vpcmpeqd %ymm3, %ymm1, %ymm1
119; AVX2-NEXT:    vpcmpeqd %ymm2, %ymm0, %ymm0
120; AVX2-NEXT:    vpacksswb %ymm1, %ymm0, %ymm0
121; AVX2-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[0,2,1,3]
122; AVX2-NEXT:    vextracti128 $1, %ymm0, %xmm1
123; AVX2-NEXT:    vpacksswb %xmm1, %xmm0, %xmm0
124; AVX2-NEXT:    vpand %xmm4, %xmm0, %xmm1
125; AVX2-NEXT:    vpandn %xmm5, %xmm0, %xmm0
126; AVX2-NEXT:    vpor %xmm0, %xmm1, %xmm0
127; AVX2-NEXT:    vzeroupper
128; AVX2-NEXT:    retq
129;
130; AVX512-LABEL: vselect_packss_v16i32:
131; AVX512:       # BB#0:
132; AVX512-NEXT:    vpcmpeqd %zmm1, %zmm0, %k1
133; AVX512-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
134; AVX512-NEXT:    vpmovdb %zmm0, %xmm0
135; AVX512-NEXT:    vpblendvb %xmm0, %xmm2, %xmm3, %xmm0
136; AVX512-NEXT:    vzeroupper
137; AVX512-NEXT:    retq
138  %1 = icmp eq <16 x i32> %a0, %a1
139  %2 = sext <16 x i1> %1 to <16 x i8>
140  %3 = and <16 x i8> %2, %a2
141  %4 = xor <16 x i8> %2, <i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1>
142  %5 = and <16 x i8> %4, %a3
143  %6 = or <16 x i8> %3, %5
144  ret <16 x i8> %6
145}
146
147define <16 x i8> @vselect_packss_v16i64(<16 x i64> %a0, <16 x i64> %a1, <16 x i8> %a2, <16 x i8> %a3) {
148; SSE2-LABEL: vselect_packss_v16i64:
149; SSE2:       # BB#0:
150; SSE2-NEXT:    pcmpeqd {{[0-9]+}}(%rsp), %xmm7
151; SSE2-NEXT:    pshufd {{.*#+}} xmm8 = xmm7[1,0,3,2]
152; SSE2-NEXT:    pand %xmm7, %xmm8
153; SSE2-NEXT:    pcmpeqd {{[0-9]+}}(%rsp), %xmm6
154; SSE2-NEXT:    pshufd {{.*#+}} xmm7 = xmm6[1,0,3,2]
155; SSE2-NEXT:    pand %xmm6, %xmm7
156; SSE2-NEXT:    packssdw %xmm8, %xmm7
157; SSE2-NEXT:    pcmpeqd {{[0-9]+}}(%rsp), %xmm5
158; SSE2-NEXT:    pshufd {{.*#+}} xmm6 = xmm5[1,0,3,2]
159; SSE2-NEXT:    pand %xmm5, %xmm6
160; SSE2-NEXT:    pcmpeqd {{[0-9]+}}(%rsp), %xmm4
161; SSE2-NEXT:    pshufd {{.*#+}} xmm5 = xmm4[1,0,3,2]
162; SSE2-NEXT:    pand %xmm4, %xmm5
163; SSE2-NEXT:    packssdw %xmm6, %xmm5
164; SSE2-NEXT:    packssdw %xmm7, %xmm5
165; SSE2-NEXT:    pcmpeqd {{[0-9]+}}(%rsp), %xmm3
166; SSE2-NEXT:    pshufd {{.*#+}} xmm4 = xmm3[1,0,3,2]
167; SSE2-NEXT:    pand %xmm3, %xmm4
168; SSE2-NEXT:    pcmpeqd {{[0-9]+}}(%rsp), %xmm2
169; SSE2-NEXT:    pshufd {{.*#+}} xmm3 = xmm2[1,0,3,2]
170; SSE2-NEXT:    pand %xmm2, %xmm3
171; SSE2-NEXT:    packssdw %xmm4, %xmm3
172; SSE2-NEXT:    pcmpeqd {{[0-9]+}}(%rsp), %xmm1
173; SSE2-NEXT:    pshufd {{.*#+}} xmm2 = xmm1[1,0,3,2]
174; SSE2-NEXT:    pand %xmm1, %xmm2
175; SSE2-NEXT:    pcmpeqd {{[0-9]+}}(%rsp), %xmm0
176; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[1,0,3,2]
177; SSE2-NEXT:    pand %xmm0, %xmm1
178; SSE2-NEXT:    packssdw %xmm2, %xmm1
179; SSE2-NEXT:    packssdw %xmm3, %xmm1
180; SSE2-NEXT:    packsswb %xmm5, %xmm1
181; SSE2-NEXT:    movdqa {{[0-9]+}}(%rsp), %xmm0
182; SSE2-NEXT:    pand %xmm1, %xmm0
183; SSE2-NEXT:    pandn {{[0-9]+}}(%rsp), %xmm1
184; SSE2-NEXT:    por %xmm0, %xmm1
185; SSE2-NEXT:    movdqa %xmm1, %xmm0
186; SSE2-NEXT:    retq
187;
188; SSE42-LABEL: vselect_packss_v16i64:
189; SSE42:       # BB#0:
190; SSE42-NEXT:    pcmpeqq {{[0-9]+}}(%rsp), %xmm7
191; SSE42-NEXT:    pcmpeqq {{[0-9]+}}(%rsp), %xmm6
192; SSE42-NEXT:    packssdw %xmm7, %xmm6
193; SSE42-NEXT:    pcmpeqq {{[0-9]+}}(%rsp), %xmm5
194; SSE42-NEXT:    pcmpeqq {{[0-9]+}}(%rsp), %xmm4
195; SSE42-NEXT:    packssdw %xmm5, %xmm4
196; SSE42-NEXT:    packssdw %xmm6, %xmm4
197; SSE42-NEXT:    pcmpeqq {{[0-9]+}}(%rsp), %xmm3
198; SSE42-NEXT:    pcmpeqq {{[0-9]+}}(%rsp), %xmm2
199; SSE42-NEXT:    packssdw %xmm3, %xmm2
200; SSE42-NEXT:    pcmpeqq {{[0-9]+}}(%rsp), %xmm1
201; SSE42-NEXT:    pcmpeqq {{[0-9]+}}(%rsp), %xmm0
202; SSE42-NEXT:    packssdw %xmm1, %xmm0
203; SSE42-NEXT:    packssdw %xmm2, %xmm0
204; SSE42-NEXT:    packsswb %xmm4, %xmm0
205; SSE42-NEXT:    movdqa {{[0-9]+}}(%rsp), %xmm1
206; SSE42-NEXT:    pand %xmm0, %xmm1
207; SSE42-NEXT:    pandn {{[0-9]+}}(%rsp), %xmm0
208; SSE42-NEXT:    por %xmm1, %xmm0
209; SSE42-NEXT:    retq
210;
211; AVX1-LABEL: vselect_packss_v16i64:
212; AVX1:       # BB#0:
213; AVX1-NEXT:    vextractf128 $1, %ymm7, %xmm8
214; AVX1-NEXT:    vextractf128 $1, %ymm3, %xmm9
215; AVX1-NEXT:    vpcmpeqq %xmm8, %xmm9, %xmm8
216; AVX1-NEXT:    vpcmpeqq %xmm7, %xmm3, %xmm3
217; AVX1-NEXT:    vpackssdw %xmm8, %xmm3, %xmm8
218; AVX1-NEXT:    vextractf128 $1, %ymm6, %xmm7
219; AVX1-NEXT:    vextractf128 $1, %ymm2, %xmm3
220; AVX1-NEXT:    vpcmpeqq %xmm7, %xmm3, %xmm3
221; AVX1-NEXT:    vpcmpeqq %xmm6, %xmm2, %xmm2
222; AVX1-NEXT:    vpackssdw %xmm3, %xmm2, %xmm2
223; AVX1-NEXT:    vpackssdw %xmm8, %xmm2, %xmm2
224; AVX1-NEXT:    vextractf128 $1, %ymm5, %xmm3
225; AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm6
226; AVX1-NEXT:    vpcmpeqq %xmm3, %xmm6, %xmm3
227; AVX1-NEXT:    vpcmpeqq %xmm5, %xmm1, %xmm1
228; AVX1-NEXT:    vpackssdw %xmm3, %xmm1, %xmm1
229; AVX1-NEXT:    vextractf128 $1, %ymm4, %xmm3
230; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm5
231; AVX1-NEXT:    vpcmpeqq %xmm3, %xmm5, %xmm3
232; AVX1-NEXT:    vpcmpeqq %xmm4, %xmm0, %xmm0
233; AVX1-NEXT:    vpackssdw %xmm3, %xmm0, %xmm0
234; AVX1-NEXT:    vpackssdw %xmm1, %xmm0, %xmm0
235; AVX1-NEXT:    vpacksswb %xmm2, %xmm0, %xmm0
236; AVX1-NEXT:    vpand {{[0-9]+}}(%rsp), %xmm0, %xmm1
237; AVX1-NEXT:    vpandn {{[0-9]+}}(%rsp), %xmm0, %xmm0
238; AVX1-NEXT:    vpor %xmm0, %xmm1, %xmm0
239; AVX1-NEXT:    vzeroupper
240; AVX1-NEXT:    retq
241;
242; AVX2-LABEL: vselect_packss_v16i64:
243; AVX2:       # BB#0:
244; AVX2-NEXT:    vpcmpeqq %ymm7, %ymm3, %ymm3
245; AVX2-NEXT:    vpcmpeqq %ymm6, %ymm2, %ymm2
246; AVX2-NEXT:    vpackssdw %ymm3, %ymm2, %ymm2
247; AVX2-NEXT:    vpermq {{.*#+}} ymm2 = ymm2[0,2,1,3]
248; AVX2-NEXT:    vpcmpeqq %ymm5, %ymm1, %ymm1
249; AVX2-NEXT:    vpcmpeqq %ymm4, %ymm0, %ymm0
250; AVX2-NEXT:    vpackssdw %ymm1, %ymm0, %ymm0
251; AVX2-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[0,2,1,3]
252; AVX2-NEXT:    vpacksswb %ymm2, %ymm0, %ymm0
253; AVX2-NEXT:    vpermq {{.*#+}} ymm0 = ymm0[0,2,1,3]
254; AVX2-NEXT:    vextracti128 $1, %ymm0, %xmm1
255; AVX2-NEXT:    vpacksswb %xmm1, %xmm0, %xmm0
256; AVX2-NEXT:    vpand {{[0-9]+}}(%rsp), %xmm0, %xmm1
257; AVX2-NEXT:    vpandn {{[0-9]+}}(%rsp), %xmm0, %xmm0
258; AVX2-NEXT:    vpor %xmm0, %xmm1, %xmm0
259; AVX2-NEXT:    vzeroupper
260; AVX2-NEXT:    retq
261;
262; AVX512-LABEL: vselect_packss_v16i64:
263; AVX512:       # BB#0:
264; AVX512-NEXT:    vextracti32x4 $3, %zmm2, %xmm6
265; AVX512-NEXT:    vpextrq $1, %xmm6, %rcx
266; AVX512-NEXT:    vextracti32x4 $3, %zmm0, %xmm7
267; AVX512-NEXT:    vpextrq $1, %xmm7, %rdx
268; AVX512-NEXT:    xorl %eax, %eax
269; AVX512-NEXT:    cmpq %rcx, %rdx
270; AVX512-NEXT:    movq $-1, %rcx
271; AVX512-NEXT:    movl $0, %edx
272; AVX512-NEXT:    cmoveq %rcx, %rdx
273; AVX512-NEXT:    vmovq %rdx, %xmm8
274; AVX512-NEXT:    vmovq %xmm6, %rdx
275; AVX512-NEXT:    vmovq %xmm7, %rsi
276; AVX512-NEXT:    cmpq %rdx, %rsi
277; AVX512-NEXT:    movl $0, %edx
278; AVX512-NEXT:    cmoveq %rcx, %rdx
279; AVX512-NEXT:    vmovq %rdx, %xmm6
280; AVX512-NEXT:    vpunpcklqdq {{.*#+}} xmm8 = xmm6[0],xmm8[0]
281; AVX512-NEXT:    vextracti32x4 $2, %zmm2, %xmm7
282; AVX512-NEXT:    vpextrq $1, %xmm7, %rdx
283; AVX512-NEXT:    vextracti32x4 $2, %zmm0, %xmm6
284; AVX512-NEXT:    vpextrq $1, %xmm6, %rsi
285; AVX512-NEXT:    cmpq %rdx, %rsi
286; AVX512-NEXT:    movl $0, %edx
287; AVX512-NEXT:    cmoveq %rcx, %rdx
288; AVX512-NEXT:    vmovq %rdx, %xmm9
289; AVX512-NEXT:    vmovq %xmm7, %rdx
290; AVX512-NEXT:    vmovq %xmm6, %rsi
291; AVX512-NEXT:    cmpq %rdx, %rsi
292; AVX512-NEXT:    movl $0, %edx
293; AVX512-NEXT:    cmoveq %rcx, %rdx
294; AVX512-NEXT:    vmovq %rdx, %xmm6
295; AVX512-NEXT:    vpunpcklqdq {{.*#+}} xmm6 = xmm6[0],xmm9[0]
296; AVX512-NEXT:    vinserti128 $1, %xmm8, %ymm6, %ymm8
297; AVX512-NEXT:    vextracti128 $1, %ymm2, %xmm7
298; AVX512-NEXT:    vpextrq $1, %xmm7, %rdx
299; AVX512-NEXT:    vextracti128 $1, %ymm0, %xmm6
300; AVX512-NEXT:    vpextrq $1, %xmm6, %rsi
301; AVX512-NEXT:    cmpq %rdx, %rsi
302; AVX512-NEXT:    movl $0, %edx
303; AVX512-NEXT:    cmoveq %rcx, %rdx
304; AVX512-NEXT:    vmovq %rdx, %xmm9
305; AVX512-NEXT:    vmovq %xmm7, %rdx
306; AVX512-NEXT:    vmovq %xmm6, %rsi
307; AVX512-NEXT:    cmpq %rdx, %rsi
308; AVX512-NEXT:    movl $0, %edx
309; AVX512-NEXT:    cmoveq %rcx, %rdx
310; AVX512-NEXT:    vmovq %rdx, %xmm6
311; AVX512-NEXT:    vpunpcklqdq {{.*#+}} xmm6 = xmm6[0],xmm9[0]
312; AVX512-NEXT:    vpextrq $1, %xmm2, %rdx
313; AVX512-NEXT:    vpextrq $1, %xmm0, %rsi
314; AVX512-NEXT:    cmpq %rdx, %rsi
315; AVX512-NEXT:    movl $0, %edx
316; AVX512-NEXT:    cmoveq %rcx, %rdx
317; AVX512-NEXT:    vmovq %rdx, %xmm7
318; AVX512-NEXT:    vmovq %xmm2, %rdx
319; AVX512-NEXT:    vmovq %xmm0, %rsi
320; AVX512-NEXT:    cmpq %rdx, %rsi
321; AVX512-NEXT:    movl $0, %edx
322; AVX512-NEXT:    cmoveq %rcx, %rdx
323; AVX512-NEXT:    vmovq %rdx, %xmm0
324; AVX512-NEXT:    vpunpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm7[0]
325; AVX512-NEXT:    vinserti128 $1, %xmm6, %ymm0, %ymm0
326; AVX512-NEXT:    vinserti64x4 $1, %ymm8, %zmm0, %zmm0
327; AVX512-NEXT:    vpmovqd %zmm0, %ymm8
328; AVX512-NEXT:    vextracti32x4 $3, %zmm3, %xmm2
329; AVX512-NEXT:    vpextrq $1, %xmm2, %rdx
330; AVX512-NEXT:    vextracti32x4 $3, %zmm1, %xmm6
331; AVX512-NEXT:    vpextrq $1, %xmm6, %rsi
332; AVX512-NEXT:    cmpq %rdx, %rsi
333; AVX512-NEXT:    movl $0, %edx
334; AVX512-NEXT:    cmoveq %rcx, %rdx
335; AVX512-NEXT:    vmovq %rdx, %xmm7
336; AVX512-NEXT:    vmovq %xmm2, %rdx
337; AVX512-NEXT:    vmovq %xmm6, %rsi
338; AVX512-NEXT:    cmpq %rdx, %rsi
339; AVX512-NEXT:    movl $0, %edx
340; AVX512-NEXT:    cmoveq %rcx, %rdx
341; AVX512-NEXT:    vmovq %rdx, %xmm2
342; AVX512-NEXT:    vpunpcklqdq {{.*#+}} xmm2 = xmm2[0],xmm7[0]
343; AVX512-NEXT:    vextracti32x4 $2, %zmm3, %xmm6
344; AVX512-NEXT:    vpextrq $1, %xmm6, %rdx
345; AVX512-NEXT:    vextracti32x4 $2, %zmm1, %xmm7
346; AVX512-NEXT:    vpextrq $1, %xmm7, %rsi
347; AVX512-NEXT:    cmpq %rdx, %rsi
348; AVX512-NEXT:    movl $0, %edx
349; AVX512-NEXT:    cmoveq %rcx, %rdx
350; AVX512-NEXT:    vmovq %rdx, %xmm0
351; AVX512-NEXT:    vmovq %xmm6, %rdx
352; AVX512-NEXT:    vmovq %xmm7, %rsi
353; AVX512-NEXT:    cmpq %rdx, %rsi
354; AVX512-NEXT:    movl $0, %edx
355; AVX512-NEXT:    cmoveq %rcx, %rdx
356; AVX512-NEXT:    vmovq %rdx, %xmm6
357; AVX512-NEXT:    vpunpcklqdq {{.*#+}} xmm0 = xmm6[0],xmm0[0]
358; AVX512-NEXT:    vinserti128 $1, %xmm2, %ymm0, %ymm2
359; AVX512-NEXT:    vextracti128 $1, %ymm3, %xmm0
360; AVX512-NEXT:    vpextrq $1, %xmm0, %rdx
361; AVX512-NEXT:    vextracti128 $1, %ymm1, %xmm6
362; AVX512-NEXT:    vpextrq $1, %xmm6, %rsi
363; AVX512-NEXT:    cmpq %rdx, %rsi
364; AVX512-NEXT:    movl $0, %edx
365; AVX512-NEXT:    cmoveq %rcx, %rdx
366; AVX512-NEXT:    vmovq %rdx, %xmm7
367; AVX512-NEXT:    vmovq %xmm0, %rdx
368; AVX512-NEXT:    vmovq %xmm6, %rsi
369; AVX512-NEXT:    cmpq %rdx, %rsi
370; AVX512-NEXT:    movl $0, %edx
371; AVX512-NEXT:    cmoveq %rcx, %rdx
372; AVX512-NEXT:    vmovq %rdx, %xmm0
373; AVX512-NEXT:    vpunpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm7[0]
374; AVX512-NEXT:    vpextrq $1, %xmm3, %rdx
375; AVX512-NEXT:    vpextrq $1, %xmm1, %rsi
376; AVX512-NEXT:    cmpq %rdx, %rsi
377; AVX512-NEXT:    movl $0, %edx
378; AVX512-NEXT:    cmoveq %rcx, %rdx
379; AVX512-NEXT:    vmovq %rdx, %xmm6
380; AVX512-NEXT:    vmovq %xmm3, %rdx
381; AVX512-NEXT:    vmovq %xmm1, %rsi
382; AVX512-NEXT:    cmpq %rdx, %rsi
383; AVX512-NEXT:    cmoveq %rcx, %rax
384; AVX512-NEXT:    vmovq %rax, %xmm1
385; AVX512-NEXT:    vpunpcklqdq {{.*#+}} xmm1 = xmm1[0],xmm6[0]
386; AVX512-NEXT:    vinserti128 $1, %xmm0, %ymm1, %ymm0
387; AVX512-NEXT:    vinserti64x4 $1, %ymm2, %zmm0, %zmm0
388; AVX512-NEXT:    vpmovqd %zmm0, %ymm0
389; AVX512-NEXT:    vinserti64x4 $1, %ymm0, %zmm8, %zmm0
390; AVX512-NEXT:    vpmovdb %zmm0, %xmm0
391; AVX512-NEXT:    vpblendvb %xmm0, %xmm4, %xmm5, %xmm0
392; AVX512-NEXT:    vzeroupper
393; AVX512-NEXT:    retq
394  %1 = icmp eq <16 x i64> %a0, %a1
395  %2 = sext <16 x i1> %1 to <16 x i8>
396  %3 = and <16 x i8> %2, %a2
397  %4 = xor <16 x i8> %2, <i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1>
398  %5 = and <16 x i8> %4, %a3
399  %6 = or <16 x i8> %3, %5
400  ret <16 x i8> %6
401}
402
403;
404; PACKSS case
405;
406
407define <16 x i8> @vselect_packss(<16 x i16> %a0, <16 x i16> %a1, <16 x i8> %a2, <16 x i8> %a3) {
408; SSE2-LABEL: vselect_packss:
409; SSE2:       # BB#0:
410; SSE2-NEXT:    pcmpeqw %xmm3, %xmm1
411; SSE2-NEXT:    pcmpeqw %xmm2, %xmm0
412; SSE2-NEXT:    packsswb %xmm1, %xmm0
413; SSE2-NEXT:    pand %xmm0, %xmm4
414; SSE2-NEXT:    pandn %xmm5, %xmm0
415; SSE2-NEXT:    por %xmm4, %xmm0
416; SSE2-NEXT:    retq
417;
418; SSE42-LABEL: vselect_packss:
419; SSE42:       # BB#0:
420; SSE42-NEXT:    pcmpeqw %xmm3, %xmm1
421; SSE42-NEXT:    pcmpeqw %xmm2, %xmm0
422; SSE42-NEXT:    packsswb %xmm1, %xmm0
423; SSE42-NEXT:    pblendvb %xmm0, %xmm4, %xmm5
424; SSE42-NEXT:    movdqa %xmm5, %xmm0
425; SSE42-NEXT:    retq
426;
427; AVX1-LABEL: vselect_packss:
428; AVX1:       # BB#0:
429; AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm4
430; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm5
431; AVX1-NEXT:    vpcmpeqw %xmm4, %xmm5, %xmm4
432; AVX1-NEXT:    vpcmpeqw %xmm1, %xmm0, %xmm0
433; AVX1-NEXT:    vpacksswb %xmm4, %xmm0, %xmm0
434; AVX1-NEXT:    vpblendvb %xmm0, %xmm2, %xmm3, %xmm0
435; AVX1-NEXT:    vzeroupper
436; AVX1-NEXT:    retq
437;
438; AVX2-LABEL: vselect_packss:
439; AVX2:       # BB#0:
440; AVX2-NEXT:    vpcmpeqw %ymm1, %ymm0, %ymm0
441; AVX2-NEXT:    vextracti128 $1, %ymm0, %xmm1
442; AVX2-NEXT:    vpacksswb %xmm1, %xmm0, %xmm0
443; AVX2-NEXT:    vpblendvb %xmm0, %xmm2, %xmm3, %xmm0
444; AVX2-NEXT:    vzeroupper
445; AVX2-NEXT:    retq
446;
447; AVX512-LABEL: vselect_packss:
448; AVX512:       # BB#0:
449; AVX512-NEXT:    vpcmpeqw %ymm1, %ymm0, %ymm0
450; AVX512-NEXT:    vextracti128 $1, %ymm0, %xmm1
451; AVX512-NEXT:    vpacksswb %xmm1, %xmm0, %xmm0
452; AVX512-NEXT:    vpblendvb %xmm0, %xmm2, %xmm3, %xmm0
453; AVX512-NEXT:    vzeroupper
454; AVX512-NEXT:    retq
455  %1 = icmp eq <16 x i16> %a0, %a1
456  %2 = sext <16 x i1> %1 to <16 x i16>
457  %3 = shufflevector <16 x i16> %2, <16 x i16> undef, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
458  %4 = shufflevector <16 x i16> %2, <16 x i16> undef, <8 x i32> <i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
459  %5 = tail call <16 x i8> @llvm.x86.sse2.packsswb.128(<8 x i16> %3, <8 x i16> %4)
460  %6 = and <16 x i8> %5, %a2
461  %7 = xor <16 x i8> %5, <i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1, i8 -1>
462  %8 = and <16 x i8> %7, %a3
463  %9 = or <16 x i8> %6, %8
464  ret <16 x i8> %9
465}
466declare <16 x i8> @llvm.x86.sse2.packsswb.128(<8 x i16>, <8 x i16>)
467