| blob | 7d1c52cd65451226c97ca64f2b4434a9f60cdaf1 |
1 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
2 ; RUN: llc < %s -mtriple=x86_64-apple-darwin -mattr=+sse2 | FileCheck %s --check-prefix=SSE2
3 ; RUN: llc < %s -mtriple=x86_64-apple-darwin -mattr=+sse4.1 | FileCheck %s --check-prefixes=SSE41
4 ; RUN: llc < %s -mtriple=x86_64-apple-darwin -mattr=+avx | FileCheck %s --check-prefixes=AVX,AVX1
5 ; RUN: llc < %s -mtriple=x86_64-apple-darwin -mattr=+avx512f,+avx512vl | FileCheck %s --check-prefixes=AVX,AVX512,AVX512F
6 ; RUN: llc < %s -mtriple=x86_64-apple-darwin -mattr=+avx512fp16,+avx512vl | FileCheck %s --check-prefixes=AVX,AVX512,AVX512FP16
8 define half @roundeven_f16(half %h) {
9 ; SSE2-LABEL: roundeven_f16:
10 ; SSE2: ## %bb.0: ## %entry
11 ; SSE2-NEXT: pushq %rax
12 ; SSE2-NEXT: .cfi_def_cfa_offset 16
13 ; SSE2-NEXT: pextrw $0, %xmm0, %eax
14 ; SSE2-NEXT: movzwl %ax, %edi
15 ; SSE2-NEXT: callq ___extendhfsf2
16 ; SSE2-NEXT: callq _roundevenf
17 ; SSE2-NEXT: callq ___truncsfhf2
18 ; SSE2-NEXT: ## kill: def $ax killed $ax def $eax
19 ; SSE2-NEXT: pinsrw $0, %eax, %xmm0
20 ; SSE2-NEXT: popq %rax
21 ; SSE2-NEXT: retq
22 ;
23 ; SSE41-LABEL: roundeven_f16:
24 ; SSE41: ## %bb.0: ## %entry
25 ; SSE41-NEXT: pushq %rax
26 ; SSE41-NEXT: .cfi_def_cfa_offset 16
27 ; SSE41-NEXT: pextrw $0, %xmm0, %eax
28 ; SSE41-NEXT: movzwl %ax, %edi
29 ; SSE41-NEXT: callq ___extendhfsf2
30 ; SSE41-NEXT: roundss $8, %xmm0, %xmm0
31 ; SSE41-NEXT: callq ___truncsfhf2
32 ; SSE41-NEXT: ## kill: def $ax killed $ax def $eax
33 ; SSE41-NEXT: pinsrw $0, %eax, %xmm0
34 ; SSE41-NEXT: popq %rax
35 ; SSE41-NEXT: retq
36 ;
37 ; AVX1-LABEL: roundeven_f16:
38 ; AVX1: ## %bb.0: ## %entry
39 ; AVX1-NEXT: pushq %rax
40 ; AVX1-NEXT: .cfi_def_cfa_offset 16
41 ; AVX1-NEXT: vpextrw $0, %xmm0, %eax
42 ; AVX1-NEXT: movzwl %ax, %edi
43 ; AVX1-NEXT: callq ___extendhfsf2
44 ; AVX1-NEXT: vroundss $8, %xmm0, %xmm0, %xmm0
45 ; AVX1-NEXT: callq ___truncsfhf2
46 ; AVX1-NEXT: ## kill: def $ax killed $ax def $eax
47 ; AVX1-NEXT: vpinsrw $0, %eax, %xmm0, %xmm0
48 ; AVX1-NEXT: popq %rax
49 ; AVX1-NEXT: retq
50 ;
51 ; AVX512F-LABEL: roundeven_f16:
52 ; AVX512F: ## %bb.0: ## %entry
53 ; AVX512F-NEXT: vcvtph2ps %xmm0, %xmm0
54 ; AVX512F-NEXT: vroundss $8, %xmm0, %xmm0, %xmm0
55 ; AVX512F-NEXT: vcvtps2ph $4, %xmm0, %xmm0
56 ; AVX512F-NEXT: vmovd %xmm0, %eax
57 ; AVX512F-NEXT: vpinsrw $0, %eax, %xmm0, %xmm0
58 ; AVX512F-NEXT: retq
59 ;
60 ; AVX512FP16-LABEL: roundeven_f16:
61 ; AVX512FP16: ## %bb.0: ## %entry
62 ; AVX512FP16-NEXT: vrndscalesh $8, %xmm0, %xmm0, %xmm0
63 ; AVX512FP16-NEXT: retq
64 entry:
65 %a = call half @llvm.roundeven.f16(half %h)
66 ret half %a
67 }
69 define float @roundeven_f32(float %x) {
70 ; SSE2-LABEL: roundeven_f32:
71 ; SSE2: ## %bb.0:
72 ; SSE2-NEXT: jmp _roundevenf ## TAILCALL
73 ;
74 ; SSE41-LABEL: roundeven_f32:
75 ; SSE41: ## %bb.0:
76 ; SSE41-NEXT: roundss $8, %xmm0, %xmm0
77 ; SSE41-NEXT: retq
78 ;
79 ; AVX-LABEL: roundeven_f32:
80 ; AVX: ## %bb.0:
81 ; AVX-NEXT: vroundss $8, %xmm0, %xmm0, %xmm0
82 ; AVX-NEXT: retq
83 %a = call float @llvm.roundeven.f32(float %x)
84 ret float %a
85 }
87 define double @roundeven_f64(double %x) {
88 ; SSE2-LABEL: roundeven_f64:
89 ; SSE2: ## %bb.0:
90 ; SSE2-NEXT: jmp _roundeven ## TAILCALL
91 ;
92 ; SSE41-LABEL: roundeven_f64:
93 ; SSE41: ## %bb.0:
94 ; SSE41-NEXT: roundsd $8, %xmm0, %xmm0
95 ; SSE41-NEXT: retq
96 ;
97 ; AVX-LABEL: roundeven_f64:
98 ; AVX: ## %bb.0:
99 ; AVX-NEXT: vroundsd $8, %xmm0, %xmm0, %xmm0
100 ; AVX-NEXT: retq
101 %a = call double @llvm.roundeven.f64(double %x)
102 ret double %a
103 }
105 define <4 x float> @roundeven_v4f32(<4 x float> %x) {
106 ; SSE2-LABEL: roundeven_v4f32:
107 ; SSE2: ## %bb.0:
108 ; SSE2-NEXT: subq $56, %rsp
109 ; SSE2-NEXT: .cfi_def_cfa_offset 64
110 ; SSE2-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
111 ; SSE2-NEXT: shufps {{.*#+}} xmm0 = xmm0[3,3,3,3]
112 ; SSE2-NEXT: callq _roundevenf
113 ; SSE2-NEXT: movaps %xmm0, (%rsp) ## 16-byte Spill
114 ; SSE2-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
115 ; SSE2-NEXT: movhlps {{.*#+}} xmm0 = xmm0[1,1]
116 ; SSE2-NEXT: callq _roundevenf
117 ; SSE2-NEXT: unpcklps (%rsp), %xmm0 ## 16-byte Folded Reload
118 ; SSE2-NEXT: ## xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]
119 ; SSE2-NEXT: movaps %xmm0, (%rsp) ## 16-byte Spill
120 ; SSE2-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
121 ; SSE2-NEXT: callq _roundevenf
122 ; SSE2-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
123 ; SSE2-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
124 ; SSE2-NEXT: shufps {{.*#+}} xmm0 = xmm0[1,1,1,1]
125 ; SSE2-NEXT: callq _roundevenf
126 ; SSE2-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 ## 16-byte Reload
127 ; SSE2-NEXT: unpcklps {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]
128 ; SSE2-NEXT: unpcklpd (%rsp), %xmm1 ## 16-byte Folded Reload
129 ; SSE2-NEXT: ## xmm1 = xmm1[0],mem[0]
130 ; SSE2-NEXT: movaps %xmm1, %xmm0
131 ; SSE2-NEXT: addq $56, %rsp
132 ; SSE2-NEXT: retq
133 ;
134 ; SSE41-LABEL: roundeven_v4f32:
135 ; SSE41: ## %bb.0:
136 ; SSE41-NEXT: roundps $8, %xmm0, %xmm0
137 ; SSE41-NEXT: retq
138 ;
139 ; AVX-LABEL: roundeven_v4f32:
140 ; AVX: ## %bb.0:
141 ; AVX-NEXT: vroundps $8, %xmm0, %xmm0
142 ; AVX-NEXT: retq
143 %a = call <4 x float> @llvm.roundeven.v4f32(<4 x float> %x)
144 ret <4 x float> %a
145 }
147 define <2 x double> @roundeven_v2f64(<2 x double> %x) {
148 ; SSE2-LABEL: roundeven_v2f64:
149 ; SSE2: ## %bb.0:
150 ; SSE2-NEXT: subq $40, %rsp
151 ; SSE2-NEXT: .cfi_def_cfa_offset 48
152 ; SSE2-NEXT: movaps %xmm0, (%rsp) ## 16-byte Spill
153 ; SSE2-NEXT: callq _roundeven
154 ; SSE2-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
155 ; SSE2-NEXT: movaps (%rsp), %xmm0 ## 16-byte Reload
156 ; SSE2-NEXT: movhlps {{.*#+}} xmm0 = xmm0[1,1]
157 ; SSE2-NEXT: callq _roundeven
158 ; SSE2-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 ## 16-byte Reload
159 ; SSE2-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
160 ; SSE2-NEXT: movaps %xmm1, %xmm0
161 ; SSE2-NEXT: addq $40, %rsp
162 ; SSE2-NEXT: retq
163 ;
164 ; SSE41-LABEL: roundeven_v2f64:
165 ; SSE41: ## %bb.0:
166 ; SSE41-NEXT: roundpd $8, %xmm0, %xmm0
167 ; SSE41-NEXT: retq
168 ;
169 ; AVX-LABEL: roundeven_v2f64:
170 ; AVX: ## %bb.0:
171 ; AVX-NEXT: vroundpd $8, %xmm0, %xmm0
172 ; AVX-NEXT: retq
173 %a = call <2 x double> @llvm.roundeven.v2f64(<2 x double> %x)
174 ret <2 x double> %a
175 }
177 define <8 x float> @roundeven_v8f32(<8 x float> %x) {
178 ; SSE2-LABEL: roundeven_v8f32:
179 ; SSE2: ## %bb.0:
180 ; SSE2-NEXT: subq $72, %rsp
181 ; SSE2-NEXT: .cfi_def_cfa_offset 80
182 ; SSE2-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
183 ; SSE2-NEXT: movaps %xmm0, (%rsp) ## 16-byte Spill
184 ; SSE2-NEXT: shufps {{.*#+}} xmm0 = xmm0[3,3,3,3]
185 ; SSE2-NEXT: callq _roundevenf
186 ; SSE2-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
187 ; SSE2-NEXT: movaps (%rsp), %xmm0 ## 16-byte Reload
188 ; SSE2-NEXT: movhlps {{.*#+}} xmm0 = xmm0[1,1]
189 ; SSE2-NEXT: callq _roundevenf
190 ; SSE2-NEXT: unpcklps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Folded Reload
191 ; SSE2-NEXT: ## xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]
192 ; SSE2-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
193 ; SSE2-NEXT: movaps (%rsp), %xmm0 ## 16-byte Reload
194 ; SSE2-NEXT: callq _roundevenf
195 ; SSE2-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
196 ; SSE2-NEXT: movaps (%rsp), %xmm0 ## 16-byte Reload
197 ; SSE2-NEXT: shufps {{.*#+}} xmm0 = xmm0[1,1,1,1]
198 ; SSE2-NEXT: callq _roundevenf
199 ; SSE2-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 ## 16-byte Reload
200 ; SSE2-NEXT: unpcklps {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]
201 ; SSE2-NEXT: unpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 ## 16-byte Folded Reload
202 ; SSE2-NEXT: ## xmm1 = xmm1[0],mem[0]
203 ; SSE2-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
204 ; SSE2-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
205 ; SSE2-NEXT: shufps {{.*#+}} xmm0 = xmm0[3,3,3,3]
206 ; SSE2-NEXT: callq _roundevenf
207 ; SSE2-NEXT: movaps %xmm0, (%rsp) ## 16-byte Spill
208 ; SSE2-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
209 ; SSE2-NEXT: movhlps {{.*#+}} xmm0 = xmm0[1,1]
210 ; SSE2-NEXT: callq _roundevenf
211 ; SSE2-NEXT: unpcklps (%rsp), %xmm0 ## 16-byte Folded Reload
212 ; SSE2-NEXT: ## xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]
213 ; SSE2-NEXT: movaps %xmm0, (%rsp) ## 16-byte Spill
214 ; SSE2-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
215 ; SSE2-NEXT: callq _roundevenf
216 ; SSE2-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
217 ; SSE2-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
218 ; SSE2-NEXT: shufps {{.*#+}} xmm0 = xmm0[1,1,1,1]
219 ; SSE2-NEXT: callq _roundevenf
220 ; SSE2-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 ## 16-byte Reload
221 ; SSE2-NEXT: unpcklps {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]
222 ; SSE2-NEXT: unpcklpd (%rsp), %xmm1 ## 16-byte Folded Reload
223 ; SSE2-NEXT: ## xmm1 = xmm1[0],mem[0]
224 ; SSE2-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
225 ; SSE2-NEXT: addq $72, %rsp
226 ; SSE2-NEXT: retq
227 ;
228 ; SSE41-LABEL: roundeven_v8f32:
229 ; SSE41: ## %bb.0:
230 ; SSE41-NEXT: roundps $8, %xmm0, %xmm0
231 ; SSE41-NEXT: roundps $8, %xmm1, %xmm1
232 ; SSE41-NEXT: retq
233 ;
234 ; AVX-LABEL: roundeven_v8f32:
235 ; AVX: ## %bb.0:
236 ; AVX-NEXT: vroundps $8, %ymm0, %ymm0
237 ; AVX-NEXT: retq
238 %a = call <8 x float> @llvm.roundeven.v8f32(<8 x float> %x)
239 ret <8 x float> %a
240 }
242 define <4 x double> @roundeven_v4f64(<4 x double> %x) {
243 ; SSE2-LABEL: roundeven_v4f64:
244 ; SSE2: ## %bb.0:
245 ; SSE2-NEXT: subq $56, %rsp
246 ; SSE2-NEXT: .cfi_def_cfa_offset 64
247 ; SSE2-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
248 ; SSE2-NEXT: movaps %xmm0, (%rsp) ## 16-byte Spill
249 ; SSE2-NEXT: callq _roundeven
250 ; SSE2-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
251 ; SSE2-NEXT: movaps (%rsp), %xmm0 ## 16-byte Reload
252 ; SSE2-NEXT: movhlps {{.*#+}} xmm0 = xmm0[1,1]
253 ; SSE2-NEXT: callq _roundeven
254 ; SSE2-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 ## 16-byte Reload
255 ; SSE2-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
256 ; SSE2-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
257 ; SSE2-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
258 ; SSE2-NEXT: callq _roundeven
259 ; SSE2-NEXT: movaps %xmm0, (%rsp) ## 16-byte Spill
260 ; SSE2-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
261 ; SSE2-NEXT: movhlps {{.*#+}} xmm0 = xmm0[1,1]
262 ; SSE2-NEXT: callq _roundeven
263 ; SSE2-NEXT: movaps (%rsp), %xmm1 ## 16-byte Reload
264 ; SSE2-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
265 ; SSE2-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
266 ; SSE2-NEXT: addq $56, %rsp
267 ; SSE2-NEXT: retq
268 ;
269 ; SSE41-LABEL: roundeven_v4f64:
270 ; SSE41: ## %bb.0:
271 ; SSE41-NEXT: roundpd $8, %xmm0, %xmm0
272 ; SSE41-NEXT: roundpd $8, %xmm1, %xmm1
273 ; SSE41-NEXT: retq
274 ;
275 ; AVX-LABEL: roundeven_v4f64:
276 ; AVX: ## %bb.0:
277 ; AVX-NEXT: vroundpd $8, %ymm0, %ymm0
278 ; AVX-NEXT: retq
279 %a = call <4 x double> @llvm.roundeven.v4f64(<4 x double> %x)
280 ret <4 x double> %a
281 }
283 define <16 x float> @roundeven_v16f32(<16 x float> %x) {
284 ; SSE2-LABEL: roundeven_v16f32:
285 ; SSE2: ## %bb.0:
286 ; SSE2-NEXT: subq $104, %rsp
287 ; SSE2-NEXT: .cfi_def_cfa_offset 112
288 ; SSE2-NEXT: movaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
289 ; SSE2-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
290 ; SSE2-NEXT: movaps %xmm1, (%rsp) ## 16-byte Spill
291 ; SSE2-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
292 ; SSE2-NEXT: shufps {{.*#+}} xmm0 = xmm0[3,3,3,3]
293 ; SSE2-NEXT: callq _roundevenf
294 ; SSE2-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
295 ; SSE2-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
296 ; SSE2-NEXT: movhlps {{.*#+}} xmm0 = xmm0[1,1]
297 ; SSE2-NEXT: callq _roundevenf
298 ; SSE2-NEXT: unpcklps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Folded Reload
299 ; SSE2-NEXT: ## xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]
300 ; SSE2-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
301 ; SSE2-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
302 ; SSE2-NEXT: callq _roundevenf
303 ; SSE2-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
304 ; SSE2-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
305 ; SSE2-NEXT: shufps {{.*#+}} xmm0 = xmm0[1,1,1,1]
306 ; SSE2-NEXT: callq _roundevenf
307 ; SSE2-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 ## 16-byte Reload
308 ; SSE2-NEXT: unpcklps {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]
309 ; SSE2-NEXT: unpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 ## 16-byte Folded Reload
310 ; SSE2-NEXT: ## xmm1 = xmm1[0],mem[0]
311 ; SSE2-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
312 ; SSE2-NEXT: movaps (%rsp), %xmm0 ## 16-byte Reload
313 ; SSE2-NEXT: shufps {{.*#+}} xmm0 = xmm0[3,3,3,3]
314 ; SSE2-NEXT: callq _roundevenf
315 ; SSE2-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
316 ; SSE2-NEXT: movaps (%rsp), %xmm0 ## 16-byte Reload
317 ; SSE2-NEXT: movhlps {{.*#+}} xmm0 = xmm0[1,1]
318 ; SSE2-NEXT: callq _roundevenf
319 ; SSE2-NEXT: unpcklps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Folded Reload
320 ; SSE2-NEXT: ## xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]
321 ; SSE2-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
322 ; SSE2-NEXT: movaps (%rsp), %xmm0 ## 16-byte Reload
323 ; SSE2-NEXT: callq _roundevenf
324 ; SSE2-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
325 ; SSE2-NEXT: movaps (%rsp), %xmm0 ## 16-byte Reload
326 ; SSE2-NEXT: shufps {{.*#+}} xmm0 = xmm0[1,1,1,1]
327 ; SSE2-NEXT: callq _roundevenf
328 ; SSE2-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 ## 16-byte Reload
329 ; SSE2-NEXT: unpcklps {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]
330 ; SSE2-NEXT: unpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 ## 16-byte Folded Reload
331 ; SSE2-NEXT: ## xmm1 = xmm1[0],mem[0]
332 ; SSE2-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
333 ; SSE2-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
334 ; SSE2-NEXT: shufps {{.*#+}} xmm0 = xmm0[3,3,3,3]
335 ; SSE2-NEXT: callq _roundevenf
336 ; SSE2-NEXT: movaps %xmm0, (%rsp) ## 16-byte Spill
337 ; SSE2-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
338 ; SSE2-NEXT: movhlps {{.*#+}} xmm0 = xmm0[1,1]
339 ; SSE2-NEXT: callq _roundevenf
340 ; SSE2-NEXT: unpcklps (%rsp), %xmm0 ## 16-byte Folded Reload
341 ; SSE2-NEXT: ## xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]
342 ; SSE2-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
343 ; SSE2-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
344 ; SSE2-NEXT: callq _roundevenf
345 ; SSE2-NEXT: movaps %xmm0, (%rsp) ## 16-byte Spill
346 ; SSE2-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
347 ; SSE2-NEXT: shufps {{.*#+}} xmm0 = xmm0[1,1,1,1]
348 ; SSE2-NEXT: callq _roundevenf
349 ; SSE2-NEXT: movaps (%rsp), %xmm1 ## 16-byte Reload
350 ; SSE2-NEXT: unpcklps {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]
351 ; SSE2-NEXT: unpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 ## 16-byte Folded Reload
352 ; SSE2-NEXT: ## xmm1 = xmm1[0],mem[0]
353 ; SSE2-NEXT: movaps %xmm1, (%rsp) ## 16-byte Spill
354 ; SSE2-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
355 ; SSE2-NEXT: shufps {{.*#+}} xmm0 = xmm0[3,3,3,3]
356 ; SSE2-NEXT: callq _roundevenf
357 ; SSE2-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
358 ; SSE2-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
359 ; SSE2-NEXT: movhlps {{.*#+}} xmm0 = xmm0[1,1]
360 ; SSE2-NEXT: callq _roundevenf
361 ; SSE2-NEXT: unpcklps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Folded Reload
362 ; SSE2-NEXT: ## xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]
363 ; SSE2-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
364 ; SSE2-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
365 ; SSE2-NEXT: callq _roundevenf
366 ; SSE2-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
367 ; SSE2-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
368 ; SSE2-NEXT: shufps {{.*#+}} xmm0 = xmm0[1,1,1,1]
369 ; SSE2-NEXT: callq _roundevenf
370 ; SSE2-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 ## 16-byte Reload
371 ; SSE2-NEXT: unpcklps {{.*#+}} xmm3 = xmm3[0],xmm0[0],xmm3[1],xmm0[1]
372 ; SSE2-NEXT: unpcklpd {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 ## 16-byte Folded Reload
373 ; SSE2-NEXT: ## xmm3 = xmm3[0],mem[0]
374 ; SSE2-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
375 ; SSE2-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 ## 16-byte Reload
376 ; SSE2-NEXT: movaps (%rsp), %xmm2 ## 16-byte Reload
377 ; SSE2-NEXT: addq $104, %rsp
378 ; SSE2-NEXT: retq
379 ;
380 ; SSE41-LABEL: roundeven_v16f32:
381 ; SSE41: ## %bb.0:
382 ; SSE41-NEXT: roundps $8, %xmm0, %xmm0
383 ; SSE41-NEXT: roundps $8, %xmm1, %xmm1
384 ; SSE41-NEXT: roundps $8, %xmm2, %xmm2
385 ; SSE41-NEXT: roundps $8, %xmm3, %xmm3
386 ; SSE41-NEXT: retq
387 ;
388 ; AVX1-LABEL: roundeven_v16f32:
389 ; AVX1: ## %bb.0:
390 ; AVX1-NEXT: vroundps $8, %ymm0, %ymm0
391 ; AVX1-NEXT: vroundps $8, %ymm1, %ymm1
392 ; AVX1-NEXT: retq
393 ;
394 ; AVX512-LABEL: roundeven_v16f32:
395 ; AVX512: ## %bb.0:
396 ; AVX512-NEXT: vrndscaleps $8, %zmm0, %zmm0
397 ; AVX512-NEXT: retq
398 %a = call <16 x float> @llvm.roundeven.v16f32(<16 x float> %x)
399 ret <16 x float> %a
400 }
402 define <8 x double> @roundeven_v8f64(<8 x double> %x) {
403 ; SSE2-LABEL: roundeven_v8f64:
404 ; SSE2: ## %bb.0:
405 ; SSE2-NEXT: subq $88, %rsp
406 ; SSE2-NEXT: .cfi_def_cfa_offset 96
407 ; SSE2-NEXT: movaps %xmm3, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
408 ; SSE2-NEXT: movaps %xmm2, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
409 ; SSE2-NEXT: movaps %xmm1, (%rsp) ## 16-byte Spill
410 ; SSE2-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
411 ; SSE2-NEXT: callq _roundeven
412 ; SSE2-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
413 ; SSE2-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
414 ; SSE2-NEXT: movhlps {{.*#+}} xmm0 = xmm0[1,1]
415 ; SSE2-NEXT: callq _roundeven
416 ; SSE2-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 ## 16-byte Reload
417 ; SSE2-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
418 ; SSE2-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
419 ; SSE2-NEXT: movaps (%rsp), %xmm0 ## 16-byte Reload
420 ; SSE2-NEXT: callq _roundeven
421 ; SSE2-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
422 ; SSE2-NEXT: movaps (%rsp), %xmm0 ## 16-byte Reload
423 ; SSE2-NEXT: movhlps {{.*#+}} xmm0 = xmm0[1,1]
424 ; SSE2-NEXT: callq _roundeven
425 ; SSE2-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 ## 16-byte Reload
426 ; SSE2-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
427 ; SSE2-NEXT: movaps %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
428 ; SSE2-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
429 ; SSE2-NEXT: callq _roundeven
430 ; SSE2-NEXT: movaps %xmm0, (%rsp) ## 16-byte Spill
431 ; SSE2-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
432 ; SSE2-NEXT: movhlps {{.*#+}} xmm0 = xmm0[1,1]
433 ; SSE2-NEXT: callq _roundeven
434 ; SSE2-NEXT: movaps (%rsp), %xmm1 ## 16-byte Reload
435 ; SSE2-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm0[0]
436 ; SSE2-NEXT: movaps %xmm1, (%rsp) ## 16-byte Spill
437 ; SSE2-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
438 ; SSE2-NEXT: callq _roundeven
439 ; SSE2-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) ## 16-byte Spill
440 ; SSE2-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
441 ; SSE2-NEXT: movhlps {{.*#+}} xmm0 = xmm0[1,1]
442 ; SSE2-NEXT: callq _roundeven
443 ; SSE2-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 ## 16-byte Reload
444 ; SSE2-NEXT: movlhps {{.*#+}} xmm3 = xmm3[0],xmm0[0]
445 ; SSE2-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 ## 16-byte Reload
446 ; SSE2-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 ## 16-byte Reload
447 ; SSE2-NEXT: movaps (%rsp), %xmm2 ## 16-byte Reload
448 ; SSE2-NEXT: addq $88, %rsp
449 ; SSE2-NEXT: retq
450 ;
451 ; SSE41-LABEL: roundeven_v8f64:
452 ; SSE41: ## %bb.0:
453 ; SSE41-NEXT: roundpd $8, %xmm0, %xmm0
454 ; SSE41-NEXT: roundpd $8, %xmm1, %xmm1
455 ; SSE41-NEXT: roundpd $8, %xmm2, %xmm2
456 ; SSE41-NEXT: roundpd $8, %xmm3, %xmm3
457 ; SSE41-NEXT: retq
458 ;
459 ; AVX1-LABEL: roundeven_v8f64:
460 ; AVX1: ## %bb.0:
461 ; AVX1-NEXT: vroundpd $8, %ymm0, %ymm0
462 ; AVX1-NEXT: vroundpd $8, %ymm1, %ymm1
463 ; AVX1-NEXT: retq
464 ;
465 ; AVX512-LABEL: roundeven_v8f64:
466 ; AVX512: ## %bb.0:
467 ; AVX512-NEXT: vrndscalepd $8, %zmm0, %zmm0
468 ; AVX512-NEXT: retq
469 %a = call <8 x double> @llvm.roundeven.v8f64(<8 x double> %x)
470 ret <8 x double> %a
471 }
473 declare half @llvm.roundeven.f16(half)
474 declare float @llvm.roundeven.f32(float)
475 declare double @llvm.roundeven.f64(double)
476 declare <4 x float> @llvm.roundeven.v4f32(<4 x float>)
477 declare <2 x double> @llvm.roundeven.v2f64(<2 x double>)
478 declare <8 x float> @llvm.roundeven.v8f32(<8 x float>)
479 declare <4 x double> @llvm.roundeven.v4f64(<4 x double>)
480 declare <16 x float> @llvm.roundeven.v16f32(<16 x float>)
481 declare <8 x double> @llvm.roundeven.v8f64(<8 x double>)