xref: /llvm-project/llvm/test/CodeGen/X86/vec-strict-cmp-sub128.ll (revision 7b3bbd83c0c24087072ec5b22a76799ab31f87d5) 
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
2; RUN: llc < %s -mtriple=i686-unknown-unknown -mattr=+sse2 -O3 | FileCheck %s --check-prefix=SSE-32
3; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+sse2 -O3 | FileCheck %s --check-prefix=SSE-64
4; RUN: llc < %s -mtriple=i686-unknown-unknown -mattr=+avx -O3 | FileCheck %s --check-prefix=AVX-32
5; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx -O3 | FileCheck %s --check-prefix=AVX-64
6; RUN: llc < %s -mtriple=i686-unknown-unknown -mattr=+avx512f -mattr=+avx512vl -O3 | FileCheck %s --check-prefix=AVX512-32
7; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512f -mattr=+avx512vl -O3 | FileCheck %s --check-prefix=AVX512-64
8; RUN: llc < %s -mtriple=i686-unknown-unknown -mattr=+avx512f -mattr=+avx512f -O3 | FileCheck %s --check-prefix=AVX512F-32
9; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512f -mattr=+avx512f -O3 | FileCheck %s --check-prefix=AVX512F-64
10
11define <2 x i32> @test_v2f32_ogt_s(<2 x i32> %a, <2 x i32> %b, <2 x float> %f1, <2 x float> %f2) #0 {
12; SSE-32-LABEL: test_v2f32_ogt_s:
13; SSE-32:       # %bb.0:
14; SSE-32-NEXT:    pushl %ebp
15; SSE-32-NEXT:    movl %esp, %ebp
16; SSE-32-NEXT:    andl $-16, %esp
17; SSE-32-NEXT:    subl $16, %esp
18; SSE-32-NEXT:    movaps 8(%ebp), %xmm4
19; SSE-32-NEXT:    xorl %eax, %eax
20; SSE-32-NEXT:    comiss %xmm4, %xmm2
21; SSE-32-NEXT:    movl $-1, %ecx
22; SSE-32-NEXT:    movl $0, %edx
23; SSE-32-NEXT:    cmoval %ecx, %edx
24; SSE-32-NEXT:    movd %edx, %xmm3
25; SSE-32-NEXT:    shufps {{.*#+}} xmm4 = xmm4[1,1,1,1]
26; SSE-32-NEXT:    shufps {{.*#+}} xmm2 = xmm2[1,1,1,1]
27; SSE-32-NEXT:    comiss %xmm4, %xmm2
28; SSE-32-NEXT:    cmoval %ecx, %eax
29; SSE-32-NEXT:    movd %eax, %xmm2
30; SSE-32-NEXT:    punpckldq {{.*#+}} xmm3 = xmm3[0],xmm2[0],xmm3[1],xmm2[1]
31; SSE-32-NEXT:    pand %xmm3, %xmm0
32; SSE-32-NEXT:    pandn %xmm1, %xmm3
33; SSE-32-NEXT:    por %xmm3, %xmm0
34; SSE-32-NEXT:    movl %ebp, %esp
35; SSE-32-NEXT:    popl %ebp
36; SSE-32-NEXT:    retl
37;
38; SSE-64-LABEL: test_v2f32_ogt_s:
39; SSE-64:       # %bb.0:
40; SSE-64-NEXT:    xorl %eax, %eax
41; SSE-64-NEXT:    comiss %xmm3, %xmm2
42; SSE-64-NEXT:    movl $-1, %ecx
43; SSE-64-NEXT:    movl $0, %edx
44; SSE-64-NEXT:    cmoval %ecx, %edx
45; SSE-64-NEXT:    movd %edx, %xmm4
46; SSE-64-NEXT:    shufps {{.*#+}} xmm3 = xmm3[1,1,1,1]
47; SSE-64-NEXT:    shufps {{.*#+}} xmm2 = xmm2[1,1,1,1]
48; SSE-64-NEXT:    comiss %xmm3, %xmm2
49; SSE-64-NEXT:    cmoval %ecx, %eax
50; SSE-64-NEXT:    movd %eax, %xmm2
51; SSE-64-NEXT:    punpckldq {{.*#+}} xmm4 = xmm4[0],xmm2[0],xmm4[1],xmm2[1]
52; SSE-64-NEXT:    pand %xmm4, %xmm0
53; SSE-64-NEXT:    pandn %xmm1, %xmm4
54; SSE-64-NEXT:    por %xmm4, %xmm0
55; SSE-64-NEXT:    retq
56;
57; AVX-32-LABEL: test_v2f32_ogt_s:
58; AVX-32:       # %bb.0:
59; AVX-32-NEXT:    pushl %ebp
60; AVX-32-NEXT:    movl %esp, %ebp
61; AVX-32-NEXT:    andl $-16, %esp
62; AVX-32-NEXT:    subl $16, %esp
63; AVX-32-NEXT:    vmovshdup {{.*#+}} xmm3 = xmm2[1,1,3,3]
64; AVX-32-NEXT:    xorl %eax, %eax
65; AVX-32-NEXT:    vcomiss 12(%ebp), %xmm3
66; AVX-32-NEXT:    movl $-1, %ecx
67; AVX-32-NEXT:    movl $0, %edx
68; AVX-32-NEXT:    cmoval %ecx, %edx
69; AVX-32-NEXT:    vcomiss 8(%ebp), %xmm2
70; AVX-32-NEXT:    cmoval %ecx, %eax
71; AVX-32-NEXT:    vmovd %eax, %xmm2
72; AVX-32-NEXT:    vpinsrd $1, %edx, %xmm2, %xmm2
73; AVX-32-NEXT:    vblendvps %xmm2, %xmm0, %xmm1, %xmm0
74; AVX-32-NEXT:    movl %ebp, %esp
75; AVX-32-NEXT:    popl %ebp
76; AVX-32-NEXT:    retl
77;
78; AVX-64-LABEL: test_v2f32_ogt_s:
79; AVX-64:       # %bb.0:
80; AVX-64-NEXT:    vmovshdup {{.*#+}} xmm4 = xmm3[1,1,3,3]
81; AVX-64-NEXT:    vmovshdup {{.*#+}} xmm5 = xmm2[1,1,3,3]
82; AVX-64-NEXT:    xorl %eax, %eax
83; AVX-64-NEXT:    vcomiss %xmm4, %xmm5
84; AVX-64-NEXT:    movl $-1, %ecx
85; AVX-64-NEXT:    movl $0, %edx
86; AVX-64-NEXT:    cmoval %ecx, %edx
87; AVX-64-NEXT:    vcomiss %xmm3, %xmm2
88; AVX-64-NEXT:    cmoval %ecx, %eax
89; AVX-64-NEXT:    vmovd %eax, %xmm2
90; AVX-64-NEXT:    vpinsrd $1, %edx, %xmm2, %xmm2
91; AVX-64-NEXT:    vblendvps %xmm2, %xmm0, %xmm1, %xmm0
92; AVX-64-NEXT:    retq
93;
94; AVX512-32-LABEL: test_v2f32_ogt_s:
95; AVX512-32:       # %bb.0:
96; AVX512-32-NEXT:    pushl %ebp
97; AVX512-32-NEXT:    movl %esp, %ebp
98; AVX512-32-NEXT:    andl $-16, %esp
99; AVX512-32-NEXT:    subl $16, %esp
100; AVX512-32-NEXT:    vcomiss 8(%ebp), %xmm2
101; AVX512-32-NEXT:    seta %al
102; AVX512-32-NEXT:    andl $1, %eax
103; AVX512-32-NEXT:    kmovw %eax, %k0
104; AVX512-32-NEXT:    vmovshdup {{.*#+}} xmm2 = xmm2[1,1,3,3]
105; AVX512-32-NEXT:    vcomiss 12(%ebp), %xmm2
106; AVX512-32-NEXT:    seta %al
107; AVX512-32-NEXT:    kmovw %eax, %k1
108; AVX512-32-NEXT:    kshiftlw $15, %k1, %k1
109; AVX512-32-NEXT:    kshiftrw $14, %k1, %k1
110; AVX512-32-NEXT:    korw %k1, %k0, %k1
111; AVX512-32-NEXT:    vpblendmd %xmm0, %xmm1, %xmm0 {%k1}
112; AVX512-32-NEXT:    movl %ebp, %esp
113; AVX512-32-NEXT:    popl %ebp
114; AVX512-32-NEXT:    retl
115;
116; AVX512-64-LABEL: test_v2f32_ogt_s:
117; AVX512-64:       # %bb.0:
118; AVX512-64-NEXT:    vcomiss %xmm3, %xmm2
119; AVX512-64-NEXT:    seta %al
120; AVX512-64-NEXT:    andl $1, %eax
121; AVX512-64-NEXT:    kmovw %eax, %k0
122; AVX512-64-NEXT:    vmovshdup {{.*#+}} xmm3 = xmm3[1,1,3,3]
123; AVX512-64-NEXT:    vmovshdup {{.*#+}} xmm2 = xmm2[1,1,3,3]
124; AVX512-64-NEXT:    vcomiss %xmm3, %xmm2
125; AVX512-64-NEXT:    seta %al
126; AVX512-64-NEXT:    kmovw %eax, %k1
127; AVX512-64-NEXT:    kshiftlw $15, %k1, %k1
128; AVX512-64-NEXT:    kshiftrw $14, %k1, %k1
129; AVX512-64-NEXT:    korw %k1, %k0, %k1
130; AVX512-64-NEXT:    vpblendmd %xmm0, %xmm1, %xmm0 {%k1}
131; AVX512-64-NEXT:    retq
132;
133; AVX512F-32-LABEL: test_v2f32_ogt_s:
134; AVX512F-32:       # %bb.0:
135; AVX512F-32-NEXT:    pushl %ebp
136; AVX512F-32-NEXT:    movl %esp, %ebp
137; AVX512F-32-NEXT:    andl $-16, %esp
138; AVX512F-32-NEXT:    subl $16, %esp
139; AVX512F-32-NEXT:    # kill: def $xmm1 killed $xmm1 def $zmm1
140; AVX512F-32-NEXT:    # kill: def $xmm0 killed $xmm0 def $zmm0
141; AVX512F-32-NEXT:    vcomiss 8(%ebp), %xmm2
142; AVX512F-32-NEXT:    seta %al
143; AVX512F-32-NEXT:    andl $1, %eax
144; AVX512F-32-NEXT:    kmovw %eax, %k0
145; AVX512F-32-NEXT:    vmovshdup {{.*#+}} xmm2 = xmm2[1,1,3,3]
146; AVX512F-32-NEXT:    vcomiss 12(%ebp), %xmm2
147; AVX512F-32-NEXT:    seta %al
148; AVX512F-32-NEXT:    kmovw %eax, %k1
149; AVX512F-32-NEXT:    kshiftlw $15, %k1, %k1
150; AVX512F-32-NEXT:    kshiftrw $14, %k1, %k1
151; AVX512F-32-NEXT:    korw %k1, %k0, %k1
152; AVX512F-32-NEXT:    vpblendmd %zmm0, %zmm1, %zmm0 {%k1}
153; AVX512F-32-NEXT:    # kill: def $xmm0 killed $xmm0 killed $zmm0
154; AVX512F-32-NEXT:    movl %ebp, %esp
155; AVX512F-32-NEXT:    popl %ebp
156; AVX512F-32-NEXT:    vzeroupper
157; AVX512F-32-NEXT:    retl
158;
159; AVX512F-64-LABEL: test_v2f32_ogt_s:
160; AVX512F-64:       # %bb.0:
161; AVX512F-64-NEXT:    # kill: def $xmm1 killed $xmm1 def $zmm1
162; AVX512F-64-NEXT:    # kill: def $xmm0 killed $xmm0 def $zmm0
163; AVX512F-64-NEXT:    vcomiss %xmm3, %xmm2
164; AVX512F-64-NEXT:    seta %al
165; AVX512F-64-NEXT:    andl $1, %eax
166; AVX512F-64-NEXT:    kmovw %eax, %k0
167; AVX512F-64-NEXT:    vmovshdup {{.*#+}} xmm3 = xmm3[1,1,3,3]
168; AVX512F-64-NEXT:    vmovshdup {{.*#+}} xmm2 = xmm2[1,1,3,3]
169; AVX512F-64-NEXT:    vcomiss %xmm3, %xmm2
170; AVX512F-64-NEXT:    seta %al
171; AVX512F-64-NEXT:    kmovw %eax, %k1
172; AVX512F-64-NEXT:    kshiftlw $15, %k1, %k1
173; AVX512F-64-NEXT:    kshiftrw $14, %k1, %k1
174; AVX512F-64-NEXT:    korw %k1, %k0, %k1
175; AVX512F-64-NEXT:    vpblendmd %zmm0, %zmm1, %zmm0 {%k1}
176; AVX512F-64-NEXT:    # kill: def $xmm0 killed $xmm0 killed $zmm0
177; AVX512F-64-NEXT:    vzeroupper
178; AVX512F-64-NEXT:    retq
179  %cond = call <2 x i1> @llvm.experimental.constrained.fcmps.v2f32(
180                                               <2 x float> %f1, <2 x float> %f2, metadata !"ogt",
181                                               metadata !"fpexcept.strict") #0
182  %res = select <2 x i1> %cond, <2 x i32> %a, <2 x i32> %b
183  ret <2 x i32> %res
184}
185
186define <2 x i32> @test_v2f32_oeq_q(<2 x i32> %a, <2 x i32> %b, <2 x float> %f1, <2 x float> %f2) #0 {
187; SSE-32-LABEL: test_v2f32_oeq_q:
188; SSE-32:       # %bb.0:
189; SSE-32-NEXT:    pushl %ebp
190; SSE-32-NEXT:    movl %esp, %ebp
191; SSE-32-NEXT:    andl $-16, %esp
192; SSE-32-NEXT:    subl $16, %esp
193; SSE-32-NEXT:    movaps 8(%ebp), %xmm4
194; SSE-32-NEXT:    xorl %eax, %eax
195; SSE-32-NEXT:    ucomiss %xmm4, %xmm2
196; SSE-32-NEXT:    movl $-1, %ecx
197; SSE-32-NEXT:    movl $-1, %edx
198; SSE-32-NEXT:    cmovnel %eax, %edx
199; SSE-32-NEXT:    cmovpl %eax, %edx
200; SSE-32-NEXT:    movd %edx, %xmm3
201; SSE-32-NEXT:    shufps {{.*#+}} xmm4 = xmm4[1,1,1,1]
202; SSE-32-NEXT:    shufps {{.*#+}} xmm2 = xmm2[1,1,1,1]
203; SSE-32-NEXT:    ucomiss %xmm4, %xmm2
204; SSE-32-NEXT:    cmovnel %eax, %ecx
205; SSE-32-NEXT:    cmovpl %eax, %ecx
206; SSE-32-NEXT:    movd %ecx, %xmm2
207; SSE-32-NEXT:    punpckldq {{.*#+}} xmm3 = xmm3[0],xmm2[0],xmm3[1],xmm2[1]
208; SSE-32-NEXT:    pand %xmm3, %xmm0
209; SSE-32-NEXT:    pandn %xmm1, %xmm3
210; SSE-32-NEXT:    por %xmm3, %xmm0
211; SSE-32-NEXT:    movl %ebp, %esp
212; SSE-32-NEXT:    popl %ebp
213; SSE-32-NEXT:    retl
214;
215; SSE-64-LABEL: test_v2f32_oeq_q:
216; SSE-64:       # %bb.0:
217; SSE-64-NEXT:    xorl %eax, %eax
218; SSE-64-NEXT:    ucomiss %xmm3, %xmm2
219; SSE-64-NEXT:    movl $-1, %ecx
220; SSE-64-NEXT:    movl $-1, %edx
221; SSE-64-NEXT:    cmovnel %eax, %edx
222; SSE-64-NEXT:    cmovpl %eax, %edx
223; SSE-64-NEXT:    movd %edx, %xmm4
224; SSE-64-NEXT:    shufps {{.*#+}} xmm3 = xmm3[1,1,1,1]
225; SSE-64-NEXT:    shufps {{.*#+}} xmm2 = xmm2[1,1,1,1]
226; SSE-64-NEXT:    ucomiss %xmm3, %xmm2
227; SSE-64-NEXT:    cmovnel %eax, %ecx
228; SSE-64-NEXT:    cmovpl %eax, %ecx
229; SSE-64-NEXT:    movd %ecx, %xmm2
230; SSE-64-NEXT:    punpckldq {{.*#+}} xmm4 = xmm4[0],xmm2[0],xmm4[1],xmm2[1]
231; SSE-64-NEXT:    pand %xmm4, %xmm0
232; SSE-64-NEXT:    pandn %xmm1, %xmm4
233; SSE-64-NEXT:    por %xmm4, %xmm0
234; SSE-64-NEXT:    retq
235;
236; AVX-32-LABEL: test_v2f32_oeq_q:
237; AVX-32:       # %bb.0:
238; AVX-32-NEXT:    pushl %ebp
239; AVX-32-NEXT:    movl %esp, %ebp
240; AVX-32-NEXT:    andl $-16, %esp
241; AVX-32-NEXT:    subl $16, %esp
242; AVX-32-NEXT:    vmovshdup {{.*#+}} xmm3 = xmm2[1,1,3,3]
243; AVX-32-NEXT:    xorl %eax, %eax
244; AVX-32-NEXT:    vucomiss 12(%ebp), %xmm3
245; AVX-32-NEXT:    movl $-1, %ecx
246; AVX-32-NEXT:    movl $-1, %edx
247; AVX-32-NEXT:    cmovnel %eax, %edx
248; AVX-32-NEXT:    cmovpl %eax, %edx
249; AVX-32-NEXT:    vucomiss 8(%ebp), %xmm2
250; AVX-32-NEXT:    cmovnel %eax, %ecx
251; AVX-32-NEXT:    cmovpl %eax, %ecx
252; AVX-32-NEXT:    vmovd %ecx, %xmm2
253; AVX-32-NEXT:    vpinsrd $1, %edx, %xmm2, %xmm2
254; AVX-32-NEXT:    vblendvps %xmm2, %xmm0, %xmm1, %xmm0
255; AVX-32-NEXT:    movl %ebp, %esp
256; AVX-32-NEXT:    popl %ebp
257; AVX-32-NEXT:    retl
258;
259; AVX-64-LABEL: test_v2f32_oeq_q:
260; AVX-64:       # %bb.0:
261; AVX-64-NEXT:    vmovshdup {{.*#+}} xmm4 = xmm3[1,1,3,3]
262; AVX-64-NEXT:    vmovshdup {{.*#+}} xmm5 = xmm2[1,1,3,3]
263; AVX-64-NEXT:    xorl %eax, %eax
264; AVX-64-NEXT:    vucomiss %xmm4, %xmm5
265; AVX-64-NEXT:    movl $-1, %ecx
266; AVX-64-NEXT:    movl $-1, %edx
267; AVX-64-NEXT:    cmovnel %eax, %edx
268; AVX-64-NEXT:    cmovpl %eax, %edx
269; AVX-64-NEXT:    vucomiss %xmm3, %xmm2
270; AVX-64-NEXT:    cmovnel %eax, %ecx
271; AVX-64-NEXT:    cmovpl %eax, %ecx
272; AVX-64-NEXT:    vmovd %ecx, %xmm2
273; AVX-64-NEXT:    vpinsrd $1, %edx, %xmm2, %xmm2
274; AVX-64-NEXT:    vblendvps %xmm2, %xmm0, %xmm1, %xmm0
275; AVX-64-NEXT:    retq
276;
277; AVX512-32-LABEL: test_v2f32_oeq_q:
278; AVX512-32:       # %bb.0:
279; AVX512-32-NEXT:    pushl %ebp
280; AVX512-32-NEXT:    movl %esp, %ebp
281; AVX512-32-NEXT:    andl $-16, %esp
282; AVX512-32-NEXT:    subl $16, %esp
283; AVX512-32-NEXT:    vucomiss 8(%ebp), %xmm2
284; AVX512-32-NEXT:    setnp %al
285; AVX512-32-NEXT:    sete %cl
286; AVX512-32-NEXT:    testb %al, %cl
287; AVX512-32-NEXT:    setne %al
288; AVX512-32-NEXT:    andl $1, %eax
289; AVX512-32-NEXT:    kmovw %eax, %k0
290; AVX512-32-NEXT:    vmovshdup {{.*#+}} xmm2 = xmm2[1,1,3,3]
291; AVX512-32-NEXT:    vucomiss 12(%ebp), %xmm2
292; AVX512-32-NEXT:    setnp %al
293; AVX512-32-NEXT:    sete %cl
294; AVX512-32-NEXT:    testb %al, %cl
295; AVX512-32-NEXT:    setne %al
296; AVX512-32-NEXT:    kmovw %eax, %k1
297; AVX512-32-NEXT:    kshiftlw $15, %k1, %k1
298; AVX512-32-NEXT:    kshiftrw $14, %k1, %k1
299; AVX512-32-NEXT:    korw %k1, %k0, %k1
300; AVX512-32-NEXT:    vpblendmd %xmm0, %xmm1, %xmm0 {%k1}
301; AVX512-32-NEXT:    movl %ebp, %esp
302; AVX512-32-NEXT:    popl %ebp
303; AVX512-32-NEXT:    retl
304;
305; AVX512-64-LABEL: test_v2f32_oeq_q:
306; AVX512-64:       # %bb.0:
307; AVX512-64-NEXT:    vucomiss %xmm3, %xmm2
308; AVX512-64-NEXT:    setnp %al
309; AVX512-64-NEXT:    sete %cl
310; AVX512-64-NEXT:    testb %al, %cl
311; AVX512-64-NEXT:    setne %al
312; AVX512-64-NEXT:    andl $1, %eax
313; AVX512-64-NEXT:    kmovw %eax, %k0
314; AVX512-64-NEXT:    vmovshdup {{.*#+}} xmm3 = xmm3[1,1,3,3]
315; AVX512-64-NEXT:    vmovshdup {{.*#+}} xmm2 = xmm2[1,1,3,3]
316; AVX512-64-NEXT:    vucomiss %xmm3, %xmm2
317; AVX512-64-NEXT:    setnp %al
318; AVX512-64-NEXT:    sete %cl
319; AVX512-64-NEXT:    testb %al, %cl
320; AVX512-64-NEXT:    setne %al
321; AVX512-64-NEXT:    kmovw %eax, %k1
322; AVX512-64-NEXT:    kshiftlw $15, %k1, %k1
323; AVX512-64-NEXT:    kshiftrw $14, %k1, %k1
324; AVX512-64-NEXT:    korw %k1, %k0, %k1
325; AVX512-64-NEXT:    vpblendmd %xmm0, %xmm1, %xmm0 {%k1}
326; AVX512-64-NEXT:    retq
327;
328; AVX512F-32-LABEL: test_v2f32_oeq_q:
329; AVX512F-32:       # %bb.0:
330; AVX512F-32-NEXT:    pushl %ebp
331; AVX512F-32-NEXT:    movl %esp, %ebp
332; AVX512F-32-NEXT:    andl $-16, %esp
333; AVX512F-32-NEXT:    subl $16, %esp
334; AVX512F-32-NEXT:    # kill: def $xmm1 killed $xmm1 def $zmm1
335; AVX512F-32-NEXT:    # kill: def $xmm0 killed $xmm0 def $zmm0
336; AVX512F-32-NEXT:    vucomiss 8(%ebp), %xmm2
337; AVX512F-32-NEXT:    setnp %al
338; AVX512F-32-NEXT:    sete %cl
339; AVX512F-32-NEXT:    testb %al, %cl
340; AVX512F-32-NEXT:    setne %al
341; AVX512F-32-NEXT:    andl $1, %eax
342; AVX512F-32-NEXT:    kmovw %eax, %k0
343; AVX512F-32-NEXT:    vmovshdup {{.*#+}} xmm2 = xmm2[1,1,3,3]
344; AVX512F-32-NEXT:    vucomiss 12(%ebp), %xmm2
345; AVX512F-32-NEXT:    setnp %al
346; AVX512F-32-NEXT:    sete %cl
347; AVX512F-32-NEXT:    testb %al, %cl
348; AVX512F-32-NEXT:    setne %al
349; AVX512F-32-NEXT:    kmovw %eax, %k1
350; AVX512F-32-NEXT:    kshiftlw $15, %k1, %k1
351; AVX512F-32-NEXT:    kshiftrw $14, %k1, %k1
352; AVX512F-32-NEXT:    korw %k1, %k0, %k1
353; AVX512F-32-NEXT:    vpblendmd %zmm0, %zmm1, %zmm0 {%k1}
354; AVX512F-32-NEXT:    # kill: def $xmm0 killed $xmm0 killed $zmm0
355; AVX512F-32-NEXT:    movl %ebp, %esp
356; AVX512F-32-NEXT:    popl %ebp
357; AVX512F-32-NEXT:    vzeroupper
358; AVX512F-32-NEXT:    retl
359;
360; AVX512F-64-LABEL: test_v2f32_oeq_q:
361; AVX512F-64:       # %bb.0:
362; AVX512F-64-NEXT:    # kill: def $xmm1 killed $xmm1 def $zmm1
363; AVX512F-64-NEXT:    # kill: def $xmm0 killed $xmm0 def $zmm0
364; AVX512F-64-NEXT:    vucomiss %xmm3, %xmm2
365; AVX512F-64-NEXT:    setnp %al
366; AVX512F-64-NEXT:    sete %cl
367; AVX512F-64-NEXT:    testb %al, %cl
368; AVX512F-64-NEXT:    setne %al
369; AVX512F-64-NEXT:    andl $1, %eax
370; AVX512F-64-NEXT:    kmovw %eax, %k0
371; AVX512F-64-NEXT:    vmovshdup {{.*#+}} xmm3 = xmm3[1,1,3,3]
372; AVX512F-64-NEXT:    vmovshdup {{.*#+}} xmm2 = xmm2[1,1,3,3]
373; AVX512F-64-NEXT:    vucomiss %xmm3, %xmm2
374; AVX512F-64-NEXT:    setnp %al
375; AVX512F-64-NEXT:    sete %cl
376; AVX512F-64-NEXT:    testb %al, %cl
377; AVX512F-64-NEXT:    setne %al
378; AVX512F-64-NEXT:    kmovw %eax, %k1
379; AVX512F-64-NEXT:    kshiftlw $15, %k1, %k1
380; AVX512F-64-NEXT:    kshiftrw $14, %k1, %k1
381; AVX512F-64-NEXT:    korw %k1, %k0, %k1
382; AVX512F-64-NEXT:    vpblendmd %zmm0, %zmm1, %zmm0 {%k1}
383; AVX512F-64-NEXT:    # kill: def $xmm0 killed $xmm0 killed $zmm0
384; AVX512F-64-NEXT:    vzeroupper
385; AVX512F-64-NEXT:    retq
386  %cond = call <2 x i1> @llvm.experimental.constrained.fcmp.v2f32(
387                                               <2 x float> %f1, <2 x float> %f2, metadata !"oeq",
388                                               metadata !"fpexcept.strict") #0
389  %res = select <2 x i1> %cond, <2 x i32> %a, <2 x i32> %b
390  ret <2 x i32> %res
391}
392
393attributes #0 = { strictfp nounwind }
394
395declare <2 x i1> @llvm.experimental.constrained.fcmp.v2f32(<2 x float>, <2 x float>, metadata, metadata)
396declare <2 x i1> @llvm.experimental.constrained.fcmps.v2f32(<2 x float>, <2 x float>, metadata, metadata)
397