| blob | 83606a4bdff9bd2dd4a06cf5bcf66a47200de840 |
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
11 define <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), %xmm3
19 ; SSE-32-NEXT: xorl %eax, %eax
20 ; SSE-32-NEXT: comiss %xmm3, %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, %xmm4
25 ; SSE-32-NEXT: shufps {{.*#+}} xmm3 = xmm3[1,1,1,1]
26 ; SSE-32-NEXT: shufps {{.*#+}} xmm2 = xmm2[1,1,1,1]
27 ; SSE-32-NEXT: comiss %xmm3, %xmm2
28 ; SSE-32-NEXT: cmoval %ecx, %eax
29 ; SSE-32-NEXT: movd %eax, %xmm2
30 ; SSE-32-NEXT: punpckldq {{.*#+}} xmm4 = xmm4[0],xmm2[0],xmm4[1],xmm2[1]
31 ; SSE-32-NEXT: pand %xmm4, %xmm0
32 ; SSE-32-NEXT: pandn %xmm1, %xmm4
33 ; SSE-32-NEXT: por %xmm4, %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: movw $-3, %ax
101 ; AVX512-32-NEXT: kmovw %eax, %k0
102 ; AVX512-32-NEXT: vcomiss 8(%ebp), %xmm2
103 ; AVX512-32-NEXT: seta %al
104 ; AVX512-32-NEXT: kmovw %eax, %k1
105 ; AVX512-32-NEXT: kandw %k0, %k1, %k0
106 ; AVX512-32-NEXT: vmovshdup {{.*#+}} xmm2 = xmm2[1,1,3,3]
107 ; AVX512-32-NEXT: vcomiss 12(%ebp), %xmm2
108 ; AVX512-32-NEXT: seta %al
109 ; AVX512-32-NEXT: kmovw %eax, %k1
110 ; AVX512-32-NEXT: kshiftlw $15, %k1, %k1
111 ; AVX512-32-NEXT: kshiftrw $14, %k1, %k1
112 ; AVX512-32-NEXT: korw %k1, %k0, %k1
113 ; AVX512-32-NEXT: vpblendmd %xmm0, %xmm1, %xmm0 {%k1}
114 ; AVX512-32-NEXT: movl %ebp, %esp
115 ; AVX512-32-NEXT: popl %ebp
116 ; AVX512-32-NEXT: retl
117 ;
118 ; AVX512-64-LABEL: test_v2f32_ogt_s:
119 ; AVX512-64: # %bb.0:
120 ; AVX512-64-NEXT: movw $-3, %ax
121 ; AVX512-64-NEXT: kmovw %eax, %k0
122 ; AVX512-64-NEXT: vcomiss %xmm3, %xmm2
123 ; AVX512-64-NEXT: seta %al
124 ; AVX512-64-NEXT: kmovw %eax, %k1
125 ; AVX512-64-NEXT: kandw %k0, %k1, %k0
126 ; AVX512-64-NEXT: vmovshdup {{.*#+}} xmm3 = xmm3[1,1,3,3]
127 ; AVX512-64-NEXT: vmovshdup {{.*#+}} xmm2 = xmm2[1,1,3,3]
128 ; AVX512-64-NEXT: vcomiss %xmm3, %xmm2
129 ; AVX512-64-NEXT: seta %al
130 ; AVX512-64-NEXT: kmovw %eax, %k1
131 ; AVX512-64-NEXT: kshiftlw $15, %k1, %k1
132 ; AVX512-64-NEXT: kshiftrw $14, %k1, %k1
133 ; AVX512-64-NEXT: korw %k1, %k0, %k1
134 ; AVX512-64-NEXT: vpblendmd %xmm0, %xmm1, %xmm0 {%k1}
135 ; AVX512-64-NEXT: retq
136 ;
137 ; AVX512F-32-LABEL: test_v2f32_ogt_s:
138 ; AVX512F-32: # %bb.0:
139 ; AVX512F-32-NEXT: pushl %ebp
140 ; AVX512F-32-NEXT: movl %esp, %ebp
141 ; AVX512F-32-NEXT: andl $-16, %esp
142 ; AVX512F-32-NEXT: subl $16, %esp
143 ; AVX512F-32-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
144 ; AVX512F-32-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
145 ; AVX512F-32-NEXT: movw $-3, %ax
146 ; AVX512F-32-NEXT: kmovw %eax, %k0
147 ; AVX512F-32-NEXT: vcomiss 8(%ebp), %xmm2
148 ; AVX512F-32-NEXT: seta %al
149 ; AVX512F-32-NEXT: kmovw %eax, %k1
150 ; AVX512F-32-NEXT: kandw %k0, %k1, %k0
151 ; AVX512F-32-NEXT: vmovshdup {{.*#+}} xmm2 = xmm2[1,1,3,3]
152 ; AVX512F-32-NEXT: vcomiss 12(%ebp), %xmm2
153 ; AVX512F-32-NEXT: seta %al
154 ; AVX512F-32-NEXT: kmovw %eax, %k1
155 ; AVX512F-32-NEXT: kshiftlw $15, %k1, %k1
156 ; AVX512F-32-NEXT: kshiftrw $14, %k1, %k1
157 ; AVX512F-32-NEXT: korw %k1, %k0, %k1
158 ; AVX512F-32-NEXT: vpblendmd %zmm0, %zmm1, %zmm0 {%k1}
159 ; AVX512F-32-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
160 ; AVX512F-32-NEXT: movl %ebp, %esp
161 ; AVX512F-32-NEXT: popl %ebp
162 ; AVX512F-32-NEXT: vzeroupper
163 ; AVX512F-32-NEXT: retl
164 ;
165 ; AVX512F-64-LABEL: test_v2f32_ogt_s:
166 ; AVX512F-64: # %bb.0:
167 ; AVX512F-64-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
168 ; AVX512F-64-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
169 ; AVX512F-64-NEXT: movw $-3, %ax
170 ; AVX512F-64-NEXT: kmovw %eax, %k0
171 ; AVX512F-64-NEXT: vcomiss %xmm3, %xmm2
172 ; AVX512F-64-NEXT: seta %al
173 ; AVX512F-64-NEXT: kmovw %eax, %k1
174 ; AVX512F-64-NEXT: kandw %k0, %k1, %k0
175 ; AVX512F-64-NEXT: vmovshdup {{.*#+}} xmm3 = xmm3[1,1,3,3]
176 ; AVX512F-64-NEXT: vmovshdup {{.*#+}} xmm2 = xmm2[1,1,3,3]
177 ; AVX512F-64-NEXT: vcomiss %xmm3, %xmm2
178 ; AVX512F-64-NEXT: seta %al
179 ; AVX512F-64-NEXT: kmovw %eax, %k1
180 ; AVX512F-64-NEXT: kshiftlw $15, %k1, %k1
181 ; AVX512F-64-NEXT: kshiftrw $14, %k1, %k1
182 ; AVX512F-64-NEXT: korw %k1, %k0, %k1
183 ; AVX512F-64-NEXT: vpblendmd %zmm0, %zmm1, %zmm0 {%k1}
184 ; AVX512F-64-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
185 ; AVX512F-64-NEXT: vzeroupper
186 ; AVX512F-64-NEXT: retq
187 %cond = call <2 x i1> @llvm.experimental.constrained.fcmps.v2f32(
188 <2 x float> %f1, <2 x float> %f2, metadata !"ogt",
189 metadata !"fpexcept.strict") #0
190 %res = select <2 x i1> %cond, <2 x i32> %a, <2 x i32> %b
191 ret <2 x i32> %res
192 }
194 define <2 x i32> @test_v2f32_oeq_q(<2 x i32> %a, <2 x i32> %b, <2 x float> %f1, <2 x float> %f2) #0 {
195 ; SSE-32-LABEL: test_v2f32_oeq_q:
196 ; SSE-32: # %bb.0:
197 ; SSE-32-NEXT: pushl %ebp
198 ; SSE-32-NEXT: movl %esp, %ebp
199 ; SSE-32-NEXT: andl $-16, %esp
200 ; SSE-32-NEXT: subl $16, %esp
201 ; SSE-32-NEXT: movaps 8(%ebp), %xmm3
202 ; SSE-32-NEXT: xorl %eax, %eax
203 ; SSE-32-NEXT: ucomiss %xmm3, %xmm2
204 ; SSE-32-NEXT: movl $-1, %ecx
205 ; SSE-32-NEXT: movl $-1, %edx
206 ; SSE-32-NEXT: cmovnel %eax, %edx
207 ; SSE-32-NEXT: cmovpl %eax, %edx
208 ; SSE-32-NEXT: movd %edx, %xmm4
209 ; SSE-32-NEXT: shufps {{.*#+}} xmm3 = xmm3[1,1,1,1]
210 ; SSE-32-NEXT: shufps {{.*#+}} xmm2 = xmm2[1,1,1,1]
211 ; SSE-32-NEXT: ucomiss %xmm3, %xmm2
212 ; SSE-32-NEXT: cmovnel %eax, %ecx
213 ; SSE-32-NEXT: cmovpl %eax, %ecx
214 ; SSE-32-NEXT: movd %ecx, %xmm2
215 ; SSE-32-NEXT: punpckldq {{.*#+}} xmm4 = xmm4[0],xmm2[0],xmm4[1],xmm2[1]
216 ; SSE-32-NEXT: pand %xmm4, %xmm0
217 ; SSE-32-NEXT: pandn %xmm1, %xmm4
218 ; SSE-32-NEXT: por %xmm4, %xmm0
219 ; SSE-32-NEXT: movl %ebp, %esp
220 ; SSE-32-NEXT: popl %ebp
221 ; SSE-32-NEXT: retl
222 ;
223 ; SSE-64-LABEL: test_v2f32_oeq_q:
224 ; SSE-64: # %bb.0:
225 ; SSE-64-NEXT: xorl %eax, %eax
226 ; SSE-64-NEXT: ucomiss %xmm3, %xmm2
227 ; SSE-64-NEXT: movl $-1, %ecx
228 ; SSE-64-NEXT: movl $-1, %edx
229 ; SSE-64-NEXT: cmovnel %eax, %edx
230 ; SSE-64-NEXT: cmovpl %eax, %edx
231 ; SSE-64-NEXT: movd %edx, %xmm4
232 ; SSE-64-NEXT: shufps {{.*#+}} xmm3 = xmm3[1,1,1,1]
233 ; SSE-64-NEXT: shufps {{.*#+}} xmm2 = xmm2[1,1,1,1]
234 ; SSE-64-NEXT: ucomiss %xmm3, %xmm2
235 ; SSE-64-NEXT: cmovnel %eax, %ecx
236 ; SSE-64-NEXT: cmovpl %eax, %ecx
237 ; SSE-64-NEXT: movd %ecx, %xmm2
238 ; SSE-64-NEXT: punpckldq {{.*#+}} xmm4 = xmm4[0],xmm2[0],xmm4[1],xmm2[1]
239 ; SSE-64-NEXT: pand %xmm4, %xmm0
240 ; SSE-64-NEXT: pandn %xmm1, %xmm4
241 ; SSE-64-NEXT: por %xmm4, %xmm0
242 ; SSE-64-NEXT: retq
243 ;
244 ; AVX-32-LABEL: test_v2f32_oeq_q:
245 ; AVX-32: # %bb.0:
246 ; AVX-32-NEXT: pushl %ebp
247 ; AVX-32-NEXT: movl %esp, %ebp
248 ; AVX-32-NEXT: andl $-16, %esp
249 ; AVX-32-NEXT: subl $16, %esp
250 ; AVX-32-NEXT: vmovshdup {{.*#+}} xmm3 = xmm2[1,1,3,3]
251 ; AVX-32-NEXT: xorl %eax, %eax
252 ; AVX-32-NEXT: vucomiss 12(%ebp), %xmm3
253 ; AVX-32-NEXT: movl $-1, %ecx
254 ; AVX-32-NEXT: movl $-1, %edx
255 ; AVX-32-NEXT: cmovnel %eax, %edx
256 ; AVX-32-NEXT: cmovpl %eax, %edx
257 ; AVX-32-NEXT: vucomiss 8(%ebp), %xmm2
258 ; AVX-32-NEXT: cmovnel %eax, %ecx
259 ; AVX-32-NEXT: cmovpl %eax, %ecx
260 ; AVX-32-NEXT: vmovd %ecx, %xmm2
261 ; AVX-32-NEXT: vpinsrd $1, %edx, %xmm2, %xmm2
262 ; AVX-32-NEXT: vblendvps %xmm2, %xmm0, %xmm1, %xmm0
263 ; AVX-32-NEXT: movl %ebp, %esp
264 ; AVX-32-NEXT: popl %ebp
265 ; AVX-32-NEXT: retl
266 ;
267 ; AVX-64-LABEL: test_v2f32_oeq_q:
268 ; AVX-64: # %bb.0:
269 ; AVX-64-NEXT: vmovshdup {{.*#+}} xmm4 = xmm3[1,1,3,3]
270 ; AVX-64-NEXT: vmovshdup {{.*#+}} xmm5 = xmm2[1,1,3,3]
271 ; AVX-64-NEXT: xorl %eax, %eax
272 ; AVX-64-NEXT: vucomiss %xmm4, %xmm5
273 ; AVX-64-NEXT: movl $-1, %ecx
274 ; AVX-64-NEXT: movl $-1, %edx
275 ; AVX-64-NEXT: cmovnel %eax, %edx
276 ; AVX-64-NEXT: cmovpl %eax, %edx
277 ; AVX-64-NEXT: vucomiss %xmm3, %xmm2
278 ; AVX-64-NEXT: cmovnel %eax, %ecx
279 ; AVX-64-NEXT: cmovpl %eax, %ecx
280 ; AVX-64-NEXT: vmovd %ecx, %xmm2
281 ; AVX-64-NEXT: vpinsrd $1, %edx, %xmm2, %xmm2
282 ; AVX-64-NEXT: vblendvps %xmm2, %xmm0, %xmm1, %xmm0
283 ; AVX-64-NEXT: retq
284 ;
285 ; AVX512-32-LABEL: test_v2f32_oeq_q:
286 ; AVX512-32: # %bb.0:
287 ; AVX512-32-NEXT: pushl %ebp
288 ; AVX512-32-NEXT: movl %esp, %ebp
289 ; AVX512-32-NEXT: andl $-16, %esp
290 ; AVX512-32-NEXT: subl $16, %esp
291 ; AVX512-32-NEXT: movw $-3, %ax
292 ; AVX512-32-NEXT: kmovw %eax, %k0
293 ; AVX512-32-NEXT: vucomiss 8(%ebp), %xmm2
294 ; AVX512-32-NEXT: setnp %al
295 ; AVX512-32-NEXT: sete %cl
296 ; AVX512-32-NEXT: testb %al, %cl
297 ; AVX512-32-NEXT: setne %al
298 ; AVX512-32-NEXT: kmovw %eax, %k1
299 ; AVX512-32-NEXT: kandw %k0, %k1, %k0
300 ; AVX512-32-NEXT: vmovshdup {{.*#+}} xmm2 = xmm2[1,1,3,3]
301 ; AVX512-32-NEXT: vucomiss 12(%ebp), %xmm2
302 ; AVX512-32-NEXT: setnp %al
303 ; AVX512-32-NEXT: sete %cl
304 ; AVX512-32-NEXT: testb %al, %cl
305 ; AVX512-32-NEXT: setne %al
306 ; AVX512-32-NEXT: kmovw %eax, %k1
307 ; AVX512-32-NEXT: kshiftlw $15, %k1, %k1
308 ; AVX512-32-NEXT: kshiftrw $14, %k1, %k1
309 ; AVX512-32-NEXT: korw %k1, %k0, %k1
310 ; AVX512-32-NEXT: vpblendmd %xmm0, %xmm1, %xmm0 {%k1}
311 ; AVX512-32-NEXT: movl %ebp, %esp
312 ; AVX512-32-NEXT: popl %ebp
313 ; AVX512-32-NEXT: retl
314 ;
315 ; AVX512-64-LABEL: test_v2f32_oeq_q:
316 ; AVX512-64: # %bb.0:
317 ; AVX512-64-NEXT: vucomiss %xmm3, %xmm2
318 ; AVX512-64-NEXT: setnp %al
319 ; AVX512-64-NEXT: sete %cl
320 ; AVX512-64-NEXT: testb %al, %cl
321 ; AVX512-64-NEXT: setne %al
322 ; AVX512-64-NEXT: kmovw %eax, %k0
323 ; AVX512-64-NEXT: movw $-3, %ax
324 ; AVX512-64-NEXT: kmovw %eax, %k1
325 ; AVX512-64-NEXT: kandw %k1, %k0, %k0
326 ; AVX512-64-NEXT: vmovshdup {{.*#+}} xmm3 = xmm3[1,1,3,3]
327 ; AVX512-64-NEXT: vmovshdup {{.*#+}} xmm2 = xmm2[1,1,3,3]
328 ; AVX512-64-NEXT: vucomiss %xmm3, %xmm2
329 ; AVX512-64-NEXT: setnp %al
330 ; AVX512-64-NEXT: sete %cl
331 ; AVX512-64-NEXT: testb %al, %cl
332 ; AVX512-64-NEXT: setne %al
333 ; AVX512-64-NEXT: kmovw %eax, %k1
334 ; AVX512-64-NEXT: kshiftlw $15, %k1, %k1
335 ; AVX512-64-NEXT: kshiftrw $14, %k1, %k1
336 ; AVX512-64-NEXT: korw %k1, %k0, %k1
337 ; AVX512-64-NEXT: vpblendmd %xmm0, %xmm1, %xmm0 {%k1}
338 ; AVX512-64-NEXT: retq
339 ;
340 ; AVX512F-32-LABEL: test_v2f32_oeq_q:
341 ; AVX512F-32: # %bb.0:
342 ; AVX512F-32-NEXT: pushl %ebp
343 ; AVX512F-32-NEXT: movl %esp, %ebp
344 ; AVX512F-32-NEXT: andl $-16, %esp
345 ; AVX512F-32-NEXT: subl $16, %esp
346 ; AVX512F-32-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
347 ; AVX512F-32-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
348 ; AVX512F-32-NEXT: movw $-3, %ax
349 ; AVX512F-32-NEXT: kmovw %eax, %k0
350 ; AVX512F-32-NEXT: vucomiss 8(%ebp), %xmm2
351 ; AVX512F-32-NEXT: setnp %al
352 ; AVX512F-32-NEXT: sete %cl
353 ; AVX512F-32-NEXT: testb %al, %cl
354 ; AVX512F-32-NEXT: setne %al
355 ; AVX512F-32-NEXT: kmovw %eax, %k1
356 ; AVX512F-32-NEXT: kandw %k0, %k1, %k0
357 ; AVX512F-32-NEXT: vmovshdup {{.*#+}} xmm2 = xmm2[1,1,3,3]
358 ; AVX512F-32-NEXT: vucomiss 12(%ebp), %xmm2
359 ; AVX512F-32-NEXT: setnp %al
360 ; AVX512F-32-NEXT: sete %cl
361 ; AVX512F-32-NEXT: testb %al, %cl
362 ; AVX512F-32-NEXT: setne %al
363 ; AVX512F-32-NEXT: kmovw %eax, %k1
364 ; AVX512F-32-NEXT: kshiftlw $15, %k1, %k1
365 ; AVX512F-32-NEXT: kshiftrw $14, %k1, %k1
366 ; AVX512F-32-NEXT: korw %k1, %k0, %k1
367 ; AVX512F-32-NEXT: vpblendmd %zmm0, %zmm1, %zmm0 {%k1}
368 ; AVX512F-32-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
369 ; AVX512F-32-NEXT: movl %ebp, %esp
370 ; AVX512F-32-NEXT: popl %ebp
371 ; AVX512F-32-NEXT: vzeroupper
372 ; AVX512F-32-NEXT: retl
373 ;
374 ; AVX512F-64-LABEL: test_v2f32_oeq_q:
375 ; AVX512F-64: # %bb.0:
376 ; AVX512F-64-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
377 ; AVX512F-64-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
378 ; AVX512F-64-NEXT: vucomiss %xmm3, %xmm2
379 ; AVX512F-64-NEXT: setnp %al
380 ; AVX512F-64-NEXT: sete %cl
381 ; AVX512F-64-NEXT: testb %al, %cl
382 ; AVX512F-64-NEXT: setne %al
383 ; AVX512F-64-NEXT: kmovw %eax, %k0
384 ; AVX512F-64-NEXT: movw $-3, %ax
385 ; AVX512F-64-NEXT: kmovw %eax, %k1
386 ; AVX512F-64-NEXT: kandw %k1, %k0, %k0
387 ; AVX512F-64-NEXT: vmovshdup {{.*#+}} xmm3 = xmm3[1,1,3,3]
388 ; AVX512F-64-NEXT: vmovshdup {{.*#+}} xmm2 = xmm2[1,1,3,3]
389 ; AVX512F-64-NEXT: vucomiss %xmm3, %xmm2
390 ; AVX512F-64-NEXT: setnp %al
391 ; AVX512F-64-NEXT: sete %cl
392 ; AVX512F-64-NEXT: testb %al, %cl
393 ; AVX512F-64-NEXT: setne %al
394 ; AVX512F-64-NEXT: kmovw %eax, %k1
395 ; AVX512F-64-NEXT: kshiftlw $15, %k1, %k1
396 ; AVX512F-64-NEXT: kshiftrw $14, %k1, %k1
397 ; AVX512F-64-NEXT: korw %k1, %k0, %k1
398 ; AVX512F-64-NEXT: vpblendmd %zmm0, %zmm1, %zmm0 {%k1}
399 ; AVX512F-64-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
400 ; AVX512F-64-NEXT: vzeroupper
401 ; AVX512F-64-NEXT: retq
402 %cond = call <2 x i1> @llvm.experimental.constrained.fcmp.v2f32(
403 <2 x float> %f1, <2 x float> %f2, metadata !"oeq",
404 metadata !"fpexcept.strict") #0
405 %res = select <2 x i1> %cond, <2 x i32> %a, <2 x i32> %b
406 ret <2 x i32> %res
407 }
409 attributes #0 = { strictfp nounwind }
411 declare <2 x i1> @llvm.experimental.constrained.fcmp.v2f32(<2 x float>, <2 x float>, metadata, metadata)
412 declare <2 x i1> @llvm.experimental.constrained.fcmps.v2f32(<2 x float>, <2 x float>, metadata, metadata)