| blob | 0cf945202a2d4a2567adf26fcbccc7d0e48d7617 |
1 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
2 ; RUN: llc < %s -mtriple=i686-unknown-unknown -mattr=+avx512f -O3 | FileCheck %s --check-prefixes=CHECK,NODQ-32
3 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512f -O3 | FileCheck %s --check-prefixes=CHECK,NODQ-64
4 ; RUN: llc < %s -mtriple=i686-unknown-unknown -mattr=avx512f,avx512dq -O3 | FileCheck %s --check-prefixes=CHECK,DQ,DQ-32
5 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=avx512f,avx512dq -O3 | FileCheck %s --check-prefixes=CHECK,DQ,DQ-64
7 declare <16 x float> @llvm.experimental.constrained.sitofp.v16f32.v16i1(<16 x i1>, metadata, metadata)
8 declare <16 x float> @llvm.experimental.constrained.uitofp.v16f32.v16i1(<16 x i1>, metadata, metadata)
9 declare <16 x float> @llvm.experimental.constrained.sitofp.v16f32.v16i8(<16 x i8>, metadata, metadata)
10 declare <16 x float> @llvm.experimental.constrained.uitofp.v16f32.v16i8(<16 x i8>, metadata, metadata)
11 declare <16 x float> @llvm.experimental.constrained.sitofp.v16f32.v16i16(<16 x i16>, metadata, metadata)
12 declare <16 x float> @llvm.experimental.constrained.uitofp.v16f32.v16i16(<16 x i16>, metadata, metadata)
13 declare <16 x float> @llvm.experimental.constrained.sitofp.v16f32.v16i32(<16 x i32>, metadata, metadata)
14 declare <16 x float> @llvm.experimental.constrained.uitofp.v16f32.v16i32(<16 x i32>, metadata, metadata)
15 declare <8 x double> @llvm.experimental.constrained.sitofp.v8f64.v8i1(<8 x i1>, metadata, metadata)
16 declare <8 x double> @llvm.experimental.constrained.uitofp.v8f64.v8i1(<8 x i1>, metadata, metadata)
17 declare <8 x double> @llvm.experimental.constrained.sitofp.v8f64.v8i8(<8 x i8>, metadata, metadata)
18 declare <8 x double> @llvm.experimental.constrained.uitofp.v8f64.v8i8(<8 x i8>, metadata, metadata)
19 declare <8 x double> @llvm.experimental.constrained.sitofp.v8f64.v8i16(<8 x i16>, metadata, metadata)
20 declare <8 x double> @llvm.experimental.constrained.uitofp.v8f64.v8i16(<8 x i16>, metadata, metadata)
21 declare <8 x double> @llvm.experimental.constrained.sitofp.v8f64.v8i32(<8 x i32>, metadata, metadata)
22 declare <8 x double> @llvm.experimental.constrained.uitofp.v8f64.v8i32(<8 x i32>, metadata, metadata)
23 declare <8 x double> @llvm.experimental.constrained.sitofp.v8f64.v8i64(<8 x i64>, metadata, metadata)
24 declare <8 x double> @llvm.experimental.constrained.uitofp.v8f64.v8i64(<8 x i64>, metadata, metadata)
25 declare <8 x float> @llvm.experimental.constrained.sitofp.v8f32.v8i64(<8 x i64>, metadata, metadata)
26 declare <8 x float> @llvm.experimental.constrained.uitofp.v8f32.v8i64(<8 x i64>, metadata, metadata)
28 define <16 x float> @sitofp_v16i1_v16f32(<16 x i1> %x) #0 {
29 ; CHECK-LABEL: sitofp_v16i1_v16f32:
30 ; CHECK: # %bb.0:
31 ; CHECK-NEXT: vpmovzxbd {{.*#+}} zmm0 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero,xmm0[2],zero,zero,zero,xmm0[3],zero,zero,zero,xmm0[4],zero,zero,zero,xmm0[5],zero,zero,zero,xmm0[6],zero,zero,zero,xmm0[7],zero,zero,zero,xmm0[8],zero,zero,zero,xmm0[9],zero,zero,zero,xmm0[10],zero,zero,zero,xmm0[11],zero,zero,zero,xmm0[12],zero,zero,zero,xmm0[13],zero,zero,zero,xmm0[14],zero,zero,zero,xmm0[15],zero,zero,zero
32 ; CHECK-NEXT: vpslld $31, %zmm0, %zmm0
33 ; CHECK-NEXT: vpsrad $31, %zmm0, %zmm0
34 ; CHECK-NEXT: vcvtdq2ps %zmm0, %zmm0
35 ; CHECK-NEXT: ret{{[l|q]}}
36 %result = call <16 x float> @llvm.experimental.constrained.sitofp.v16f32.v16i1(<16 x i1> %x,
37 metadata !"round.dynamic",
38 metadata !"fpexcept.strict") #0
39 ret <16 x float> %result
40 }
42 define <16 x float> @uitofp_v16i1_v16f32(<16 x i1> %x) #0 {
43 ; NODQ-32-LABEL: uitofp_v16i1_v16f32:
44 ; NODQ-32: # %bb.0:
45 ; NODQ-32-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}, %xmm0, %xmm0
46 ; NODQ-32-NEXT: vpmovzxbd {{.*#+}} zmm0 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero,xmm0[2],zero,zero,zero,xmm0[3],zero,zero,zero,xmm0[4],zero,zero,zero,xmm0[5],zero,zero,zero,xmm0[6],zero,zero,zero,xmm0[7],zero,zero,zero,xmm0[8],zero,zero,zero,xmm0[9],zero,zero,zero,xmm0[10],zero,zero,zero,xmm0[11],zero,zero,zero,xmm0[12],zero,zero,zero,xmm0[13],zero,zero,zero,xmm0[14],zero,zero,zero,xmm0[15],zero,zero,zero
47 ; NODQ-32-NEXT: vcvtdq2ps %zmm0, %zmm0
48 ; NODQ-32-NEXT: retl
49 ;
50 ; NODQ-64-LABEL: uitofp_v16i1_v16f32:
51 ; NODQ-64: # %bb.0:
52 ; NODQ-64-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm0
53 ; NODQ-64-NEXT: vpmovzxbd {{.*#+}} zmm0 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero,xmm0[2],zero,zero,zero,xmm0[3],zero,zero,zero,xmm0[4],zero,zero,zero,xmm0[5],zero,zero,zero,xmm0[6],zero,zero,zero,xmm0[7],zero,zero,zero,xmm0[8],zero,zero,zero,xmm0[9],zero,zero,zero,xmm0[10],zero,zero,zero,xmm0[11],zero,zero,zero,xmm0[12],zero,zero,zero,xmm0[13],zero,zero,zero,xmm0[14],zero,zero,zero,xmm0[15],zero,zero,zero
54 ; NODQ-64-NEXT: vcvtdq2ps %zmm0, %zmm0
55 ; NODQ-64-NEXT: retq
56 ;
57 ; DQ-32-LABEL: uitofp_v16i1_v16f32:
58 ; DQ-32: # %bb.0:
59 ; DQ-32-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}, %xmm0, %xmm0
60 ; DQ-32-NEXT: vpmovzxbd {{.*#+}} zmm0 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero,xmm0[2],zero,zero,zero,xmm0[3],zero,zero,zero,xmm0[4],zero,zero,zero,xmm0[5],zero,zero,zero,xmm0[6],zero,zero,zero,xmm0[7],zero,zero,zero,xmm0[8],zero,zero,zero,xmm0[9],zero,zero,zero,xmm0[10],zero,zero,zero,xmm0[11],zero,zero,zero,xmm0[12],zero,zero,zero,xmm0[13],zero,zero,zero,xmm0[14],zero,zero,zero,xmm0[15],zero,zero,zero
61 ; DQ-32-NEXT: vcvtdq2ps %zmm0, %zmm0
62 ; DQ-32-NEXT: retl
63 ;
64 ; DQ-64-LABEL: uitofp_v16i1_v16f32:
65 ; DQ-64: # %bb.0:
66 ; DQ-64-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm0
67 ; DQ-64-NEXT: vpmovzxbd {{.*#+}} zmm0 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero,xmm0[2],zero,zero,zero,xmm0[3],zero,zero,zero,xmm0[4],zero,zero,zero,xmm0[5],zero,zero,zero,xmm0[6],zero,zero,zero,xmm0[7],zero,zero,zero,xmm0[8],zero,zero,zero,xmm0[9],zero,zero,zero,xmm0[10],zero,zero,zero,xmm0[11],zero,zero,zero,xmm0[12],zero,zero,zero,xmm0[13],zero,zero,zero,xmm0[14],zero,zero,zero,xmm0[15],zero,zero,zero
68 ; DQ-64-NEXT: vcvtdq2ps %zmm0, %zmm0
69 ; DQ-64-NEXT: retq
70 %result = call <16 x float> @llvm.experimental.constrained.uitofp.v16f32.v16i1(<16 x i1> %x,
71 metadata !"round.dynamic",
72 metadata !"fpexcept.strict") #0
73 ret <16 x float> %result
74 }
76 define <16 x float> @sitofp_v16i8_v16f32(<16 x i8> %x) #0 {
77 ; CHECK-LABEL: sitofp_v16i8_v16f32:
78 ; CHECK: # %bb.0:
79 ; CHECK-NEXT: vpmovsxbd %xmm0, %zmm0
80 ; CHECK-NEXT: vcvtdq2ps %zmm0, %zmm0
81 ; CHECK-NEXT: ret{{[l|q]}}
82 %result = call <16 x float> @llvm.experimental.constrained.sitofp.v16f32.v16i8(<16 x i8> %x,
83 metadata !"round.dynamic",
84 metadata !"fpexcept.strict") #0
85 ret <16 x float> %result
86 }
88 define <16 x float> @uitofp_v16i8_v16f32(<16 x i8> %x) #0 {
89 ; CHECK-LABEL: uitofp_v16i8_v16f32:
90 ; CHECK: # %bb.0:
91 ; CHECK-NEXT: vpmovzxbd {{.*#+}} zmm0 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero,xmm0[2],zero,zero,zero,xmm0[3],zero,zero,zero,xmm0[4],zero,zero,zero,xmm0[5],zero,zero,zero,xmm0[6],zero,zero,zero,xmm0[7],zero,zero,zero,xmm0[8],zero,zero,zero,xmm0[9],zero,zero,zero,xmm0[10],zero,zero,zero,xmm0[11],zero,zero,zero,xmm0[12],zero,zero,zero,xmm0[13],zero,zero,zero,xmm0[14],zero,zero,zero,xmm0[15],zero,zero,zero
92 ; CHECK-NEXT: vcvtdq2ps %zmm0, %zmm0
93 ; CHECK-NEXT: ret{{[l|q]}}
94 %result = call <16 x float> @llvm.experimental.constrained.uitofp.v16f32.v16i8(<16 x i8> %x,
95 metadata !"round.dynamic",
96 metadata !"fpexcept.strict") #0
97 ret <16 x float> %result
98 }
100 define <16 x float> @sitofp_v16i16_v16f32(<16 x i16> %x) #0 {
101 ; CHECK-LABEL: sitofp_v16i16_v16f32:
102 ; CHECK: # %bb.0:
103 ; CHECK-NEXT: vpmovsxwd %ymm0, %zmm0
104 ; CHECK-NEXT: vcvtdq2ps %zmm0, %zmm0
105 ; CHECK-NEXT: ret{{[l|q]}}
106 %result = call <16 x float> @llvm.experimental.constrained.sitofp.v16f32.v16i16(<16 x i16> %x,
107 metadata !"round.dynamic",
108 metadata !"fpexcept.strict") #0
109 ret <16 x float> %result
110 }
112 define <16 x float> @uitofp_v16i16_v16f32(<16 x i16> %x) #0 {
113 ; CHECK-LABEL: uitofp_v16i16_v16f32:
114 ; CHECK: # %bb.0:
115 ; CHECK-NEXT: vpmovzxwd {{.*#+}} zmm0 = ymm0[0],zero,ymm0[1],zero,ymm0[2],zero,ymm0[3],zero,ymm0[4],zero,ymm0[5],zero,ymm0[6],zero,ymm0[7],zero,ymm0[8],zero,ymm0[9],zero,ymm0[10],zero,ymm0[11],zero,ymm0[12],zero,ymm0[13],zero,ymm0[14],zero,ymm0[15],zero
116 ; CHECK-NEXT: vcvtdq2ps %zmm0, %zmm0
117 ; CHECK-NEXT: ret{{[l|q]}}
118 %result = call <16 x float> @llvm.experimental.constrained.uitofp.v16f32.v16i16(<16 x i16> %x,
119 metadata !"round.dynamic",
120 metadata !"fpexcept.strict") #0
121 ret <16 x float> %result
122 }
124 define <16 x float> @sitofp_v16i32_v16f32(<16 x i32> %x) #0 {
125 ; CHECK-LABEL: sitofp_v16i32_v16f32:
126 ; CHECK: # %bb.0:
127 ; CHECK-NEXT: vcvtdq2ps %zmm0, %zmm0
128 ; CHECK-NEXT: ret{{[l|q]}}
129 %result = call <16 x float> @llvm.experimental.constrained.sitofp.v16f32.v16i32(<16 x i32> %x,
130 metadata !"round.dynamic",
131 metadata !"fpexcept.strict") #0
132 ret <16 x float> %result
133 }
135 define <16 x float> @uitofp_v16i32_v16f32(<16 x i32> %x) #0 {
136 ; CHECK-LABEL: uitofp_v16i32_v16f32:
137 ; CHECK: # %bb.0:
138 ; CHECK-NEXT: vcvtudq2ps %zmm0, %zmm0
139 ; CHECK-NEXT: ret{{[l|q]}}
140 %result = call <16 x float> @llvm.experimental.constrained.uitofp.v16f32.v16i32(<16 x i32> %x,
141 metadata !"round.dynamic",
142 metadata !"fpexcept.strict") #0
143 ret <16 x float> %result
144 }
146 define <8 x double> @sitofp_v8i1_v8f64(<8 x i1> %x) #0 {
147 ; CHECK-LABEL: sitofp_v8i1_v8f64:
148 ; CHECK: # %bb.0:
149 ; CHECK-NEXT: vpmovzxwd {{.*#+}} ymm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero
150 ; CHECK-NEXT: vpslld $31, %ymm0, %ymm0
151 ; CHECK-NEXT: vpsrad $31, %ymm0, %ymm0
152 ; CHECK-NEXT: vcvtdq2pd %ymm0, %zmm0
153 ; CHECK-NEXT: ret{{[l|q]}}
154 %result = call <8 x double> @llvm.experimental.constrained.sitofp.v8f64.v8i1(<8 x i1> %x,
155 metadata !"round.dynamic",
156 metadata !"fpexcept.strict") #0
157 ret <8 x double> %result
158 }
160 define <8 x double> @uitofp_v8i1_v8f64(<8 x i1> %x) #0 {
161 ; NODQ-32-LABEL: uitofp_v8i1_v8f64:
162 ; NODQ-32: # %bb.0:
163 ; NODQ-32-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}, %xmm0, %xmm0
164 ; NODQ-32-NEXT: vpmovzxwd {{.*#+}} ymm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero
165 ; NODQ-32-NEXT: vcvtdq2pd %ymm0, %zmm0
166 ; NODQ-32-NEXT: retl
167 ;
168 ; NODQ-64-LABEL: uitofp_v8i1_v8f64:
169 ; NODQ-64: # %bb.0:
170 ; NODQ-64-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm0
171 ; NODQ-64-NEXT: vpmovzxwd {{.*#+}} ymm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero
172 ; NODQ-64-NEXT: vcvtdq2pd %ymm0, %zmm0
173 ; NODQ-64-NEXT: retq
174 ;
175 ; DQ-32-LABEL: uitofp_v8i1_v8f64:
176 ; DQ-32: # %bb.0:
177 ; DQ-32-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}, %xmm0, %xmm0
178 ; DQ-32-NEXT: vpmovzxwd {{.*#+}} ymm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero
179 ; DQ-32-NEXT: vcvtdq2pd %ymm0, %zmm0
180 ; DQ-32-NEXT: retl
181 ;
182 ; DQ-64-LABEL: uitofp_v8i1_v8f64:
183 ; DQ-64: # %bb.0:
184 ; DQ-64-NEXT: vpand {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm0
185 ; DQ-64-NEXT: vpmovzxwd {{.*#+}} ymm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero
186 ; DQ-64-NEXT: vcvtdq2pd %ymm0, %zmm0
187 ; DQ-64-NEXT: retq
188 %result = call <8 x double> @llvm.experimental.constrained.uitofp.v8f64.v8i1(<8 x i1> %x,
189 metadata !"round.dynamic",
190 metadata !"fpexcept.strict") #0
191 ret <8 x double> %result
192 }
194 define <8 x double> @sitofp_v8i8_v8f64(<8 x i8> %x) #0 {
195 ; CHECK-LABEL: sitofp_v8i8_v8f64:
196 ; CHECK: # %bb.0:
197 ; CHECK-NEXT: vpmovsxbd %xmm0, %ymm0
198 ; CHECK-NEXT: vcvtdq2pd %ymm0, %zmm0
199 ; CHECK-NEXT: ret{{[l|q]}}
200 %result = call <8 x double> @llvm.experimental.constrained.sitofp.v8f64.v8i8(<8 x i8> %x,
201 metadata !"round.dynamic",
202 metadata !"fpexcept.strict") #0
203 ret <8 x double> %result
204 }
206 define <8 x double> @uitofp_v8i8_v8f64(<8 x i8> %x) #0 {
207 ; CHECK-LABEL: uitofp_v8i8_v8f64:
208 ; CHECK: # %bb.0:
209 ; CHECK-NEXT: vpmovzxbd {{.*#+}} ymm0 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero,xmm0[2],zero,zero,zero,xmm0[3],zero,zero,zero,xmm0[4],zero,zero,zero,xmm0[5],zero,zero,zero,xmm0[6],zero,zero,zero,xmm0[7],zero,zero,zero
210 ; CHECK-NEXT: vcvtdq2pd %ymm0, %zmm0
211 ; CHECK-NEXT: ret{{[l|q]}}
212 %result = call <8 x double> @llvm.experimental.constrained.uitofp.v8f64.v8i8(<8 x i8> %x,
213 metadata !"round.dynamic",
214 metadata !"fpexcept.strict") #0
215 ret <8 x double> %result
216 }
218 define <8 x double> @sitofp_v8i16_v8f64(<8 x i16> %x) #0 {
219 ; CHECK-LABEL: sitofp_v8i16_v8f64:
220 ; CHECK: # %bb.0:
221 ; CHECK-NEXT: vpmovsxwd %xmm0, %ymm0
222 ; CHECK-NEXT: vcvtdq2pd %ymm0, %zmm0
223 ; CHECK-NEXT: ret{{[l|q]}}
224 %result = call <8 x double> @llvm.experimental.constrained.sitofp.v8f64.v8i16(<8 x i16> %x,
225 metadata !"round.dynamic",
226 metadata !"fpexcept.strict") #0
227 ret <8 x double> %result
228 }
230 define <8 x double> @uitofp_v8i16_v8f64(<8 x i16> %x) #0 {
231 ; CHECK-LABEL: uitofp_v8i16_v8f64:
232 ; CHECK: # %bb.0:
233 ; CHECK-NEXT: vpmovzxwd {{.*#+}} ymm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero
234 ; CHECK-NEXT: vcvtdq2pd %ymm0, %zmm0
235 ; CHECK-NEXT: ret{{[l|q]}}
236 %result = call <8 x double> @llvm.experimental.constrained.uitofp.v8f64.v8i16(<8 x i16> %x,
237 metadata !"round.dynamic",
238 metadata !"fpexcept.strict") #0
239 ret <8 x double> %result
240 }
242 define <8 x double> @sitofp_v8i32_v8f64(<8 x i32> %x) #0 {
243 ; CHECK-LABEL: sitofp_v8i32_v8f64:
244 ; CHECK: # %bb.0:
245 ; CHECK-NEXT: vcvtdq2pd %ymm0, %zmm0
246 ; CHECK-NEXT: ret{{[l|q]}}
247 %result = call <8 x double> @llvm.experimental.constrained.sitofp.v8f64.v8i32(<8 x i32> %x,
248 metadata !"round.dynamic",
249 metadata !"fpexcept.strict") #0
250 ret <8 x double> %result
251 }
253 define <8 x double> @uitofp_v8i32_v8f64(<8 x i32> %x) #0 {
254 ; CHECK-LABEL: uitofp_v8i32_v8f64:
255 ; CHECK: # %bb.0:
256 ; CHECK-NEXT: vcvtudq2pd %ymm0, %zmm0
257 ; CHECK-NEXT: ret{{[l|q]}}
258 %result = call <8 x double> @llvm.experimental.constrained.uitofp.v8f64.v8i32(<8 x i32> %x,
259 metadata !"round.dynamic",
260 metadata !"fpexcept.strict") #0
261 ret <8 x double> %result
262 }
264 define <8 x double> @sitofp_v8i64_v8f64(<8 x i64> %x) #0 {
265 ; NODQ-32-LABEL: sitofp_v8i64_v8f64:
266 ; NODQ-32: # %bb.0:
267 ; NODQ-32-NEXT: pushl %ebp
268 ; NODQ-32-NEXT: .cfi_def_cfa_offset 8
269 ; NODQ-32-NEXT: .cfi_offset %ebp, -8
270 ; NODQ-32-NEXT: movl %esp, %ebp
271 ; NODQ-32-NEXT: .cfi_def_cfa_register %ebp
272 ; NODQ-32-NEXT: andl $-8, %esp
273 ; NODQ-32-NEXT: subl $128, %esp
274 ; NODQ-32-NEXT: vextractf32x4 $2, %zmm0, %xmm1
275 ; NODQ-32-NEXT: vmovlps %xmm1, {{[0-9]+}}(%esp)
276 ; NODQ-32-NEXT: vshufps {{.*#+}} xmm1 = xmm1[2,3,2,3]
277 ; NODQ-32-NEXT: vmovlps %xmm1, {{[0-9]+}}(%esp)
278 ; NODQ-32-NEXT: vextractf32x4 $3, %zmm0, %xmm1
279 ; NODQ-32-NEXT: vmovlps %xmm1, {{[0-9]+}}(%esp)
280 ; NODQ-32-NEXT: vshufps {{.*#+}} xmm1 = xmm1[2,3,2,3]
281 ; NODQ-32-NEXT: vmovlps %xmm1, {{[0-9]+}}(%esp)
282 ; NODQ-32-NEXT: vmovlps %xmm0, {{[0-9]+}}(%esp)
283 ; NODQ-32-NEXT: vshufps {{.*#+}} xmm1 = xmm0[2,3,2,3]
284 ; NODQ-32-NEXT: vmovlps %xmm1, {{[0-9]+}}(%esp)
285 ; NODQ-32-NEXT: vextractf128 $1, %ymm0, %xmm0
286 ; NODQ-32-NEXT: vmovlps %xmm0, {{[0-9]+}}(%esp)
287 ; NODQ-32-NEXT: vshufps {{.*#+}} xmm0 = xmm0[2,3,2,3]
288 ; NODQ-32-NEXT: vmovlps %xmm0, {{[0-9]+}}(%esp)
289 ; NODQ-32-NEXT: fildll {{[0-9]+}}(%esp)
290 ; NODQ-32-NEXT: fstpl {{[0-9]+}}(%esp)
291 ; NODQ-32-NEXT: fildll {{[0-9]+}}(%esp)
292 ; NODQ-32-NEXT: fstpl {{[0-9]+}}(%esp)
293 ; NODQ-32-NEXT: fildll {{[0-9]+}}(%esp)
294 ; NODQ-32-NEXT: fstpl {{[0-9]+}}(%esp)
295 ; NODQ-32-NEXT: fildll {{[0-9]+}}(%esp)
296 ; NODQ-32-NEXT: fstpl (%esp)
297 ; NODQ-32-NEXT: fildll {{[0-9]+}}(%esp)
298 ; NODQ-32-NEXT: fstpl {{[0-9]+}}(%esp)
299 ; NODQ-32-NEXT: fildll {{[0-9]+}}(%esp)
300 ; NODQ-32-NEXT: fstpl {{[0-9]+}}(%esp)
301 ; NODQ-32-NEXT: fildll {{[0-9]+}}(%esp)
302 ; NODQ-32-NEXT: fstpl {{[0-9]+}}(%esp)
303 ; NODQ-32-NEXT: fildll {{[0-9]+}}(%esp)
304 ; NODQ-32-NEXT: fstpl {{[0-9]+}}(%esp)
305 ; NODQ-32-NEXT: wait
306 ; NODQ-32-NEXT: vmovsd {{.*#+}} xmm0 = mem[0],zero
307 ; NODQ-32-NEXT: vmovhps {{.*#+}} xmm0 = xmm0[0,1],mem[0,1]
308 ; NODQ-32-NEXT: vmovsd {{.*#+}} xmm1 = mem[0],zero
309 ; NODQ-32-NEXT: vmovhps {{.*#+}} xmm1 = xmm1[0,1],mem[0,1]
310 ; NODQ-32-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm0
311 ; NODQ-32-NEXT: vmovsd {{.*#+}} xmm1 = mem[0],zero
312 ; NODQ-32-NEXT: vmovhps {{.*#+}} xmm1 = xmm1[0,1],mem[0,1]
313 ; NODQ-32-NEXT: vmovsd {{.*#+}} xmm2 = mem[0],zero
314 ; NODQ-32-NEXT: vmovhps {{.*#+}} xmm2 = xmm2[0,1],mem[0,1]
315 ; NODQ-32-NEXT: vinsertf128 $1, %xmm2, %ymm1, %ymm1
316 ; NODQ-32-NEXT: vinsertf64x4 $1, %ymm0, %zmm1, %zmm0
317 ; NODQ-32-NEXT: movl %ebp, %esp
318 ; NODQ-32-NEXT: popl %ebp
319 ; NODQ-32-NEXT: .cfi_def_cfa %esp, 4
320 ; NODQ-32-NEXT: retl
321 ;
322 ; NODQ-64-LABEL: sitofp_v8i64_v8f64:
323 ; NODQ-64: # %bb.0:
324 ; NODQ-64-NEXT: vextracti32x4 $3, %zmm0, %xmm1
325 ; NODQ-64-NEXT: vpextrq $1, %xmm1, %rax
326 ; NODQ-64-NEXT: vcvtsi2sd %rax, %xmm2, %xmm2
327 ; NODQ-64-NEXT: vmovq %xmm1, %rax
328 ; NODQ-64-NEXT: vcvtsi2sd %rax, %xmm3, %xmm1
329 ; NODQ-64-NEXT: vunpcklpd {{.*#+}} xmm1 = xmm1[0],xmm2[0]
330 ; NODQ-64-NEXT: vextracti32x4 $2, %zmm0, %xmm2
331 ; NODQ-64-NEXT: vpextrq $1, %xmm2, %rax
332 ; NODQ-64-NEXT: vcvtsi2sd %rax, %xmm3, %xmm3
333 ; NODQ-64-NEXT: vmovq %xmm2, %rax
334 ; NODQ-64-NEXT: vcvtsi2sd %rax, %xmm4, %xmm2
335 ; NODQ-64-NEXT: vunpcklpd {{.*#+}} xmm2 = xmm2[0],xmm3[0]
336 ; NODQ-64-NEXT: vinsertf128 $1, %xmm1, %ymm2, %ymm1
337 ; NODQ-64-NEXT: vextracti128 $1, %ymm0, %xmm2
338 ; NODQ-64-NEXT: vpextrq $1, %xmm2, %rax
339 ; NODQ-64-NEXT: vcvtsi2sd %rax, %xmm4, %xmm3
340 ; NODQ-64-NEXT: vmovq %xmm2, %rax
341 ; NODQ-64-NEXT: vcvtsi2sd %rax, %xmm4, %xmm2
342 ; NODQ-64-NEXT: vunpcklpd {{.*#+}} xmm2 = xmm2[0],xmm3[0]
343 ; NODQ-64-NEXT: vpextrq $1, %xmm0, %rax
344 ; NODQ-64-NEXT: vcvtsi2sd %rax, %xmm4, %xmm3
345 ; NODQ-64-NEXT: vmovq %xmm0, %rax
346 ; NODQ-64-NEXT: vcvtsi2sd %rax, %xmm4, %xmm0
347 ; NODQ-64-NEXT: vunpcklpd {{.*#+}} xmm0 = xmm0[0],xmm3[0]
348 ; NODQ-64-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm0
349 ; NODQ-64-NEXT: vinsertf64x4 $1, %ymm1, %zmm0, %zmm0
350 ; NODQ-64-NEXT: retq
351 ;
352 ; DQ-LABEL: sitofp_v8i64_v8f64:
353 ; DQ: # %bb.0:
354 ; DQ-NEXT: vcvtqq2pd %zmm0, %zmm0
355 ; DQ-NEXT: ret{{[l|q]}}
356 %result = call <8 x double> @llvm.experimental.constrained.sitofp.v8f64.v8i64(<8 x i64> %x,
357 metadata !"round.dynamic",
358 metadata !"fpexcept.strict") #0
359 ret <8 x double> %result
360 }
362 define <8 x double> @uitofp_v8i64_v8f64(<8 x i64> %x) #0 {
363 ; NODQ-32-LABEL: uitofp_v8i64_v8f64:
364 ; NODQ-32: # %bb.0:
365 ; NODQ-32-NEXT: pushl %ebp
366 ; NODQ-32-NEXT: .cfi_def_cfa_offset 8
367 ; NODQ-32-NEXT: .cfi_offset %ebp, -8
368 ; NODQ-32-NEXT: movl %esp, %ebp
369 ; NODQ-32-NEXT: .cfi_def_cfa_register %ebp
370 ; NODQ-32-NEXT: andl $-8, %esp
371 ; NODQ-32-NEXT: subl $128, %esp
372 ; NODQ-32-NEXT: vextractf32x4 $2, %zmm0, %xmm3
373 ; NODQ-32-NEXT: vmovlps %xmm3, {{[0-9]+}}(%esp)
374 ; NODQ-32-NEXT: vshufps {{.*#+}} xmm1 = xmm3[2,3,2,3]
375 ; NODQ-32-NEXT: vmovlps %xmm1, {{[0-9]+}}(%esp)
376 ; NODQ-32-NEXT: vextractf32x4 $3, %zmm0, %xmm2
377 ; NODQ-32-NEXT: vmovlps %xmm2, {{[0-9]+}}(%esp)
378 ; NODQ-32-NEXT: vshufps {{.*#+}} xmm1 = xmm2[2,3,2,3]
379 ; NODQ-32-NEXT: vmovlps %xmm1, {{[0-9]+}}(%esp)
380 ; NODQ-32-NEXT: vmovlps %xmm0, {{[0-9]+}}(%esp)
381 ; NODQ-32-NEXT: vshufps {{.*#+}} xmm1 = xmm0[2,3,2,3]
382 ; NODQ-32-NEXT: vmovlps %xmm1, {{[0-9]+}}(%esp)
383 ; NODQ-32-NEXT: vextractf128 $1, %ymm0, %xmm1
384 ; NODQ-32-NEXT: vmovlps %xmm1, {{[0-9]+}}(%esp)
385 ; NODQ-32-NEXT: vshufps {{.*#+}} xmm4 = xmm1[2,3,2,3]
386 ; NODQ-32-NEXT: vmovlps %xmm4, {{[0-9]+}}(%esp)
387 ; NODQ-32-NEXT: vextractps $1, %xmm3, %eax
388 ; NODQ-32-NEXT: shrl $31, %eax
389 ; NODQ-32-NEXT: fildll {{[0-9]+}}(%esp)
390 ; NODQ-32-NEXT: fadds {{\.?LCPI[0-9]+_[0-9]+}}(,%eax,4)
391 ; NODQ-32-NEXT: fstpl {{[0-9]+}}(%esp)
392 ; NODQ-32-NEXT: wait
393 ; NODQ-32-NEXT: vextractps $3, %xmm3, %eax
394 ; NODQ-32-NEXT: shrl $31, %eax
395 ; NODQ-32-NEXT: fildll {{[0-9]+}}(%esp)
396 ; NODQ-32-NEXT: fadds {{\.?LCPI[0-9]+_[0-9]+}}(,%eax,4)
397 ; NODQ-32-NEXT: fstpl {{[0-9]+}}(%esp)
398 ; NODQ-32-NEXT: wait
399 ; NODQ-32-NEXT: vextractps $1, %xmm2, %eax
400 ; NODQ-32-NEXT: shrl $31, %eax
401 ; NODQ-32-NEXT: fildll {{[0-9]+}}(%esp)
402 ; NODQ-32-NEXT: fadds {{\.?LCPI[0-9]+_[0-9]+}}(,%eax,4)
403 ; NODQ-32-NEXT: fstpl {{[0-9]+}}(%esp)
404 ; NODQ-32-NEXT: wait
405 ; NODQ-32-NEXT: vextractps $3, %xmm2, %eax
406 ; NODQ-32-NEXT: shrl $31, %eax
407 ; NODQ-32-NEXT: fildll {{[0-9]+}}(%esp)
408 ; NODQ-32-NEXT: fadds {{\.?LCPI[0-9]+_[0-9]+}}(,%eax,4)
409 ; NODQ-32-NEXT: fstpl {{[0-9]+}}(%esp)
410 ; NODQ-32-NEXT: wait
411 ; NODQ-32-NEXT: vextractps $1, %xmm0, %eax
412 ; NODQ-32-NEXT: shrl $31, %eax
413 ; NODQ-32-NEXT: fildll {{[0-9]+}}(%esp)
414 ; NODQ-32-NEXT: fadds {{\.?LCPI[0-9]+_[0-9]+}}(,%eax,4)
415 ; NODQ-32-NEXT: fstpl (%esp)
416 ; NODQ-32-NEXT: wait
417 ; NODQ-32-NEXT: vextractps $3, %xmm0, %eax
418 ; NODQ-32-NEXT: shrl $31, %eax
419 ; NODQ-32-NEXT: fildll {{[0-9]+}}(%esp)
420 ; NODQ-32-NEXT: fadds {{\.?LCPI[0-9]+_[0-9]+}}(,%eax,4)
421 ; NODQ-32-NEXT: fstpl {{[0-9]+}}(%esp)
422 ; NODQ-32-NEXT: wait
423 ; NODQ-32-NEXT: vextractps $1, %xmm1, %eax
424 ; NODQ-32-NEXT: shrl $31, %eax
425 ; NODQ-32-NEXT: fildll {{[0-9]+}}(%esp)
426 ; NODQ-32-NEXT: fadds {{\.?LCPI[0-9]+_[0-9]+}}(,%eax,4)
427 ; NODQ-32-NEXT: fstpl {{[0-9]+}}(%esp)
428 ; NODQ-32-NEXT: wait
429 ; NODQ-32-NEXT: vextractps $3, %xmm1, %eax
430 ; NODQ-32-NEXT: shrl $31, %eax
431 ; NODQ-32-NEXT: fildll {{[0-9]+}}(%esp)
432 ; NODQ-32-NEXT: fadds {{\.?LCPI[0-9]+_[0-9]+}}(,%eax,4)
433 ; NODQ-32-NEXT: fstpl {{[0-9]+}}(%esp)
434 ; NODQ-32-NEXT: wait
435 ; NODQ-32-NEXT: vmovsd {{.*#+}} xmm0 = mem[0],zero
436 ; NODQ-32-NEXT: vmovhps {{.*#+}} xmm0 = xmm0[0,1],mem[0,1]
437 ; NODQ-32-NEXT: vmovsd {{.*#+}} xmm1 = mem[0],zero
438 ; NODQ-32-NEXT: vmovhps {{.*#+}} xmm1 = xmm1[0,1],mem[0,1]
439 ; NODQ-32-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm0
440 ; NODQ-32-NEXT: vmovsd {{.*#+}} xmm1 = mem[0],zero
441 ; NODQ-32-NEXT: vmovhps {{.*#+}} xmm1 = xmm1[0,1],mem[0,1]
442 ; NODQ-32-NEXT: vmovsd {{.*#+}} xmm2 = mem[0],zero
443 ; NODQ-32-NEXT: vmovhps {{.*#+}} xmm2 = xmm2[0,1],mem[0,1]
444 ; NODQ-32-NEXT: vinsertf128 $1, %xmm2, %ymm1, %ymm1
445 ; NODQ-32-NEXT: vinsertf64x4 $1, %ymm0, %zmm1, %zmm0
446 ; NODQ-32-NEXT: movl %ebp, %esp
447 ; NODQ-32-NEXT: popl %ebp
448 ; NODQ-32-NEXT: .cfi_def_cfa %esp, 4
449 ; NODQ-32-NEXT: retl
450 ;
451 ; NODQ-64-LABEL: uitofp_v8i64_v8f64:
452 ; NODQ-64: # %bb.0:
453 ; NODQ-64-NEXT: vextracti32x4 $3, %zmm0, %xmm1
454 ; NODQ-64-NEXT: vpextrq $1, %xmm1, %rax
455 ; NODQ-64-NEXT: vcvtusi2sd %rax, %xmm2, %xmm2
456 ; NODQ-64-NEXT: vmovq %xmm1, %rax
457 ; NODQ-64-NEXT: vcvtusi2sd %rax, %xmm3, %xmm1
458 ; NODQ-64-NEXT: vunpcklpd {{.*#+}} xmm1 = xmm1[0],xmm2[0]
459 ; NODQ-64-NEXT: vextracti32x4 $2, %zmm0, %xmm2
460 ; NODQ-64-NEXT: vpextrq $1, %xmm2, %rax
461 ; NODQ-64-NEXT: vcvtusi2sd %rax, %xmm3, %xmm3
462 ; NODQ-64-NEXT: vmovq %xmm2, %rax
463 ; NODQ-64-NEXT: vcvtusi2sd %rax, %xmm4, %xmm2
464 ; NODQ-64-NEXT: vunpcklpd {{.*#+}} xmm2 = xmm2[0],xmm3[0]
465 ; NODQ-64-NEXT: vinsertf128 $1, %xmm1, %ymm2, %ymm1
466 ; NODQ-64-NEXT: vextracti128 $1, %ymm0, %xmm2
467 ; NODQ-64-NEXT: vpextrq $1, %xmm2, %rax
468 ; NODQ-64-NEXT: vcvtusi2sd %rax, %xmm4, %xmm3
469 ; NODQ-64-NEXT: vmovq %xmm2, %rax
470 ; NODQ-64-NEXT: vcvtusi2sd %rax, %xmm4, %xmm2
471 ; NODQ-64-NEXT: vunpcklpd {{.*#+}} xmm2 = xmm2[0],xmm3[0]
472 ; NODQ-64-NEXT: vpextrq $1, %xmm0, %rax
473 ; NODQ-64-NEXT: vcvtusi2sd %rax, %xmm4, %xmm3
474 ; NODQ-64-NEXT: vmovq %xmm0, %rax
475 ; NODQ-64-NEXT: vcvtusi2sd %rax, %xmm4, %xmm0
476 ; NODQ-64-NEXT: vunpcklpd {{.*#+}} xmm0 = xmm0[0],xmm3[0]
477 ; NODQ-64-NEXT: vinsertf128 $1, %xmm2, %ymm0, %ymm0
478 ; NODQ-64-NEXT: vinsertf64x4 $1, %ymm1, %zmm0, %zmm0
479 ; NODQ-64-NEXT: retq
480 ;
481 ; DQ-LABEL: uitofp_v8i64_v8f64:
482 ; DQ: # %bb.0:
483 ; DQ-NEXT: vcvtuqq2pd %zmm0, %zmm0
484 ; DQ-NEXT: ret{{[l|q]}}
485 %result = call <8 x double> @llvm.experimental.constrained.uitofp.v8f64.v8i64(<8 x i64> %x,
486 metadata !"round.dynamic",
487 metadata !"fpexcept.strict") #0
488 ret <8 x double> %result
489 }
491 define <8 x float> @sitofp_v8i64_v8f32(<8 x i64> %x) #0 {
492 ; NODQ-32-LABEL: sitofp_v8i64_v8f32:
493 ; NODQ-32: # %bb.0:
494 ; NODQ-32-NEXT: pushl %ebp
495 ; NODQ-32-NEXT: .cfi_def_cfa_offset 8
496 ; NODQ-32-NEXT: .cfi_offset %ebp, -8
497 ; NODQ-32-NEXT: movl %esp, %ebp
498 ; NODQ-32-NEXT: .cfi_def_cfa_register %ebp
499 ; NODQ-32-NEXT: andl $-8, %esp
500 ; NODQ-32-NEXT: subl $96, %esp
501 ; NODQ-32-NEXT: vmovlps %xmm0, {{[0-9]+}}(%esp)
502 ; NODQ-32-NEXT: vshufps {{.*#+}} xmm1 = xmm0[2,3,2,3]
503 ; NODQ-32-NEXT: vmovlps %xmm1, {{[0-9]+}}(%esp)
504 ; NODQ-32-NEXT: vextractf128 $1, %ymm0, %xmm1
505 ; NODQ-32-NEXT: vmovlps %xmm1, {{[0-9]+}}(%esp)
506 ; NODQ-32-NEXT: vshufps {{.*#+}} xmm1 = xmm1[2,3,2,3]
507 ; NODQ-32-NEXT: vmovlps %xmm1, {{[0-9]+}}(%esp)
508 ; NODQ-32-NEXT: vextractf32x4 $2, %zmm0, %xmm1
509 ; NODQ-32-NEXT: vmovlps %xmm1, {{[0-9]+}}(%esp)
510 ; NODQ-32-NEXT: vshufps {{.*#+}} xmm1 = xmm1[2,3,2,3]
511 ; NODQ-32-NEXT: vmovlps %xmm1, {{[0-9]+}}(%esp)
512 ; NODQ-32-NEXT: vextractf32x4 $3, %zmm0, %xmm0
513 ; NODQ-32-NEXT: vmovlps %xmm0, {{[0-9]+}}(%esp)
514 ; NODQ-32-NEXT: vshufps {{.*#+}} xmm0 = xmm0[2,3,2,3]
515 ; NODQ-32-NEXT: vmovlps %xmm0, {{[0-9]+}}(%esp)
516 ; NODQ-32-NEXT: fildll {{[0-9]+}}(%esp)
517 ; NODQ-32-NEXT: fstps {{[0-9]+}}(%esp)
518 ; NODQ-32-NEXT: fildll {{[0-9]+}}(%esp)
519 ; NODQ-32-NEXT: fstps {{[0-9]+}}(%esp)
520 ; NODQ-32-NEXT: fildll {{[0-9]+}}(%esp)
521 ; NODQ-32-NEXT: fstps {{[0-9]+}}(%esp)
522 ; NODQ-32-NEXT: fildll {{[0-9]+}}(%esp)
523 ; NODQ-32-NEXT: fstps {{[0-9]+}}(%esp)
524 ; NODQ-32-NEXT: fildll {{[0-9]+}}(%esp)
525 ; NODQ-32-NEXT: fstps {{[0-9]+}}(%esp)
526 ; NODQ-32-NEXT: fildll {{[0-9]+}}(%esp)
527 ; NODQ-32-NEXT: fstps {{[0-9]+}}(%esp)
528 ; NODQ-32-NEXT: fildll {{[0-9]+}}(%esp)
529 ; NODQ-32-NEXT: fstps {{[0-9]+}}(%esp)
530 ; NODQ-32-NEXT: fildll {{[0-9]+}}(%esp)
531 ; NODQ-32-NEXT: fstps (%esp)
532 ; NODQ-32-NEXT: wait
533 ; NODQ-32-NEXT: vmovss {{.*#+}} xmm0 = mem[0],zero,zero,zero
534 ; NODQ-32-NEXT: vinsertps {{.*#+}} xmm0 = xmm0[0],mem[0],xmm0[2,3]
535 ; NODQ-32-NEXT: vinsertps {{.*#+}} xmm0 = xmm0[0,1],mem[0],xmm0[3]
536 ; NODQ-32-NEXT: vinsertps {{.*#+}} xmm0 = xmm0[0,1,2],mem[0]
537 ; NODQ-32-NEXT: vmovss {{.*#+}} xmm1 = mem[0],zero,zero,zero
538 ; NODQ-32-NEXT: vinsertps {{.*#+}} xmm1 = xmm1[0],mem[0],xmm1[2,3]
539 ; NODQ-32-NEXT: vinsertps {{.*#+}} xmm1 = xmm1[0,1],mem[0],xmm1[3]
540 ; NODQ-32-NEXT: vinsertps {{.*#+}} xmm1 = xmm1[0,1,2],mem[0]
541 ; NODQ-32-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm0
542 ; NODQ-32-NEXT: movl %ebp, %esp
543 ; NODQ-32-NEXT: popl %ebp
544 ; NODQ-32-NEXT: .cfi_def_cfa %esp, 4
545 ; NODQ-32-NEXT: retl
546 ;
547 ; NODQ-64-LABEL: sitofp_v8i64_v8f32:
548 ; NODQ-64: # %bb.0:
549 ; NODQ-64-NEXT: vextracti32x4 $2, %zmm0, %xmm1
550 ; NODQ-64-NEXT: vpextrq $1, %xmm1, %rax
551 ; NODQ-64-NEXT: vcvtsi2ss %rax, %xmm2, %xmm2
552 ; NODQ-64-NEXT: vmovq %xmm1, %rax
553 ; NODQ-64-NEXT: vcvtsi2ss %rax, %xmm3, %xmm1
554 ; NODQ-64-NEXT: vinsertps {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[2,3]
555 ; NODQ-64-NEXT: vextracti32x4 $3, %zmm0, %xmm2
556 ; NODQ-64-NEXT: vmovq %xmm2, %rax
557 ; NODQ-64-NEXT: vcvtsi2ss %rax, %xmm3, %xmm3
558 ; NODQ-64-NEXT: vinsertps {{.*#+}} xmm1 = xmm1[0,1],xmm3[0],xmm1[3]
559 ; NODQ-64-NEXT: vpextrq $1, %xmm2, %rax
560 ; NODQ-64-NEXT: vcvtsi2ss %rax, %xmm4, %xmm2
561 ; NODQ-64-NEXT: vinsertps {{.*#+}} xmm1 = xmm1[0,1,2],xmm2[0]
562 ; NODQ-64-NEXT: vpextrq $1, %xmm0, %rax
563 ; NODQ-64-NEXT: vcvtsi2ss %rax, %xmm4, %xmm2
564 ; NODQ-64-NEXT: vmovq %xmm0, %rax
565 ; NODQ-64-NEXT: vcvtsi2ss %rax, %xmm4, %xmm3
566 ; NODQ-64-NEXT: vinsertps {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[2,3]
567 ; NODQ-64-NEXT: vextracti128 $1, %ymm0, %xmm0
568 ; NODQ-64-NEXT: vmovq %xmm0, %rax
569 ; NODQ-64-NEXT: vcvtsi2ss %rax, %xmm4, %xmm3
570 ; NODQ-64-NEXT: vinsertps {{.*#+}} xmm2 = xmm2[0,1],xmm3[0],xmm2[3]
571 ; NODQ-64-NEXT: vpextrq $1, %xmm0, %rax
572 ; NODQ-64-NEXT: vcvtsi2ss %rax, %xmm4, %xmm0
573 ; NODQ-64-NEXT: vinsertps {{.*#+}} xmm0 = xmm2[0,1,2],xmm0[0]
574 ; NODQ-64-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm0
575 ; NODQ-64-NEXT: retq
576 ;
577 ; DQ-LABEL: sitofp_v8i64_v8f32:
578 ; DQ: # %bb.0:
579 ; DQ-NEXT: vcvtqq2ps %zmm0, %ymm0
580 ; DQ-NEXT: ret{{[l|q]}}
581 %result = call <8 x float> @llvm.experimental.constrained.sitofp.v8f32.v8i64(<8 x i64> %x,
582 metadata !"round.dynamic",
583 metadata !"fpexcept.strict") #0
584 ret <8 x float> %result
585 }
587 define <8 x float> @uitofp_v8i64_v8f32(<8 x i64> %x) #0 {
588 ; NODQ-32-LABEL: uitofp_v8i64_v8f32:
589 ; NODQ-32: # %bb.0:
590 ; NODQ-32-NEXT: pushl %ebp
591 ; NODQ-32-NEXT: .cfi_def_cfa_offset 8
592 ; NODQ-32-NEXT: .cfi_offset %ebp, -8
593 ; NODQ-32-NEXT: movl %esp, %ebp
594 ; NODQ-32-NEXT: .cfi_def_cfa_register %ebp
595 ; NODQ-32-NEXT: andl $-8, %esp
596 ; NODQ-32-NEXT: subl $96, %esp
597 ; NODQ-32-NEXT: vmovlps %xmm0, {{[0-9]+}}(%esp)
598 ; NODQ-32-NEXT: vshufps {{.*#+}} xmm1 = xmm0[2,3,2,3]
599 ; NODQ-32-NEXT: vmovlps %xmm1, {{[0-9]+}}(%esp)
600 ; NODQ-32-NEXT: vextractf128 $1, %ymm0, %xmm3
601 ; NODQ-32-NEXT: vmovlps %xmm3, {{[0-9]+}}(%esp)
602 ; NODQ-32-NEXT: vshufps {{.*#+}} xmm1 = xmm3[2,3,2,3]
603 ; NODQ-32-NEXT: vmovlps %xmm1, {{[0-9]+}}(%esp)
604 ; NODQ-32-NEXT: vextractf32x4 $2, %zmm0, %xmm2
605 ; NODQ-32-NEXT: vmovlps %xmm2, {{[0-9]+}}(%esp)
606 ; NODQ-32-NEXT: vshufps {{.*#+}} xmm1 = xmm2[2,3,2,3]
607 ; NODQ-32-NEXT: vmovlps %xmm1, {{[0-9]+}}(%esp)
608 ; NODQ-32-NEXT: vextractf32x4 $3, %zmm0, %xmm1
609 ; NODQ-32-NEXT: vmovlps %xmm1, {{[0-9]+}}(%esp)
610 ; NODQ-32-NEXT: vshufps {{.*#+}} xmm4 = xmm1[2,3,2,3]
611 ; NODQ-32-NEXT: vmovlps %xmm4, {{[0-9]+}}(%esp)
612 ; NODQ-32-NEXT: vextractps $1, %xmm0, %eax
613 ; NODQ-32-NEXT: shrl $31, %eax
614 ; NODQ-32-NEXT: fildll {{[0-9]+}}(%esp)
615 ; NODQ-32-NEXT: fadds {{\.?LCPI[0-9]+_[0-9]+}}(,%eax,4)
616 ; NODQ-32-NEXT: fstps (%esp)
617 ; NODQ-32-NEXT: wait
618 ; NODQ-32-NEXT: vextractps $3, %xmm0, %eax
619 ; NODQ-32-NEXT: shrl $31, %eax
620 ; NODQ-32-NEXT: fildll {{[0-9]+}}(%esp)
621 ; NODQ-32-NEXT: fadds {{\.?LCPI[0-9]+_[0-9]+}}(,%eax,4)
622 ; NODQ-32-NEXT: fstps {{[0-9]+}}(%esp)
623 ; NODQ-32-NEXT: wait
624 ; NODQ-32-NEXT: vextractps $1, %xmm3, %eax
625 ; NODQ-32-NEXT: shrl $31, %eax
626 ; NODQ-32-NEXT: fildll {{[0-9]+}}(%esp)
627 ; NODQ-32-NEXT: fadds {{\.?LCPI[0-9]+_[0-9]+}}(,%eax,4)
628 ; NODQ-32-NEXT: fstps {{[0-9]+}}(%esp)
629 ; NODQ-32-NEXT: wait
630 ; NODQ-32-NEXT: vextractps $3, %xmm3, %eax
631 ; NODQ-32-NEXT: shrl $31, %eax
632 ; NODQ-32-NEXT: fildll {{[0-9]+}}(%esp)
633 ; NODQ-32-NEXT: fadds {{\.?LCPI[0-9]+_[0-9]+}}(,%eax,4)
634 ; NODQ-32-NEXT: fstps {{[0-9]+}}(%esp)
635 ; NODQ-32-NEXT: wait
636 ; NODQ-32-NEXT: vextractps $1, %xmm2, %eax
637 ; NODQ-32-NEXT: shrl $31, %eax
638 ; NODQ-32-NEXT: fildll {{[0-9]+}}(%esp)
639 ; NODQ-32-NEXT: fadds {{\.?LCPI[0-9]+_[0-9]+}}(,%eax,4)
640 ; NODQ-32-NEXT: fstps {{[0-9]+}}(%esp)
641 ; NODQ-32-NEXT: wait
642 ; NODQ-32-NEXT: vextractps $3, %xmm2, %eax
643 ; NODQ-32-NEXT: shrl $31, %eax
644 ; NODQ-32-NEXT: fildll {{[0-9]+}}(%esp)
645 ; NODQ-32-NEXT: fadds {{\.?LCPI[0-9]+_[0-9]+}}(,%eax,4)
646 ; NODQ-32-NEXT: fstps {{[0-9]+}}(%esp)
647 ; NODQ-32-NEXT: wait
648 ; NODQ-32-NEXT: vextractps $1, %xmm1, %eax
649 ; NODQ-32-NEXT: shrl $31, %eax
650 ; NODQ-32-NEXT: fildll {{[0-9]+}}(%esp)
651 ; NODQ-32-NEXT: fadds {{\.?LCPI[0-9]+_[0-9]+}}(,%eax,4)
652 ; NODQ-32-NEXT: fstps {{[0-9]+}}(%esp)
653 ; NODQ-32-NEXT: wait
654 ; NODQ-32-NEXT: vextractps $3, %xmm1, %eax
655 ; NODQ-32-NEXT: shrl $31, %eax
656 ; NODQ-32-NEXT: fildll {{[0-9]+}}(%esp)
657 ; NODQ-32-NEXT: fadds {{\.?LCPI[0-9]+_[0-9]+}}(,%eax,4)
658 ; NODQ-32-NEXT: fstps {{[0-9]+}}(%esp)
659 ; NODQ-32-NEXT: wait
660 ; NODQ-32-NEXT: vmovss {{.*#+}} xmm0 = mem[0],zero,zero,zero
661 ; NODQ-32-NEXT: vinsertps {{.*#+}} xmm0 = xmm0[0],mem[0],xmm0[2,3]
662 ; NODQ-32-NEXT: vinsertps {{.*#+}} xmm0 = xmm0[0,1],mem[0],xmm0[3]
663 ; NODQ-32-NEXT: vinsertps {{.*#+}} xmm0 = xmm0[0,1,2],mem[0]
664 ; NODQ-32-NEXT: vmovss {{.*#+}} xmm1 = mem[0],zero,zero,zero
665 ; NODQ-32-NEXT: vinsertps {{.*#+}} xmm1 = xmm1[0],mem[0],xmm1[2,3]
666 ; NODQ-32-NEXT: vinsertps {{.*#+}} xmm1 = xmm1[0,1],mem[0],xmm1[3]
667 ; NODQ-32-NEXT: vinsertps {{.*#+}} xmm1 = xmm1[0,1,2],mem[0]
668 ; NODQ-32-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm0
669 ; NODQ-32-NEXT: movl %ebp, %esp
670 ; NODQ-32-NEXT: popl %ebp
671 ; NODQ-32-NEXT: .cfi_def_cfa %esp, 4
672 ; NODQ-32-NEXT: retl
673 ;
674 ; NODQ-64-LABEL: uitofp_v8i64_v8f32:
675 ; NODQ-64: # %bb.0:
676 ; NODQ-64-NEXT: vextracti32x4 $2, %zmm0, %xmm1
677 ; NODQ-64-NEXT: vpextrq $1, %xmm1, %rax
678 ; NODQ-64-NEXT: vcvtusi2ss %rax, %xmm2, %xmm2
679 ; NODQ-64-NEXT: vmovq %xmm1, %rax
680 ; NODQ-64-NEXT: vcvtusi2ss %rax, %xmm3, %xmm1
681 ; NODQ-64-NEXT: vinsertps {{.*#+}} xmm1 = xmm1[0],xmm2[0],xmm1[2,3]
682 ; NODQ-64-NEXT: vextracti32x4 $3, %zmm0, %xmm2
683 ; NODQ-64-NEXT: vmovq %xmm2, %rax
684 ; NODQ-64-NEXT: vcvtusi2ss %rax, %xmm3, %xmm3
685 ; NODQ-64-NEXT: vinsertps {{.*#+}} xmm1 = xmm1[0,1],xmm3[0],xmm1[3]
686 ; NODQ-64-NEXT: vpextrq $1, %xmm2, %rax
687 ; NODQ-64-NEXT: vcvtusi2ss %rax, %xmm4, %xmm2
688 ; NODQ-64-NEXT: vinsertps {{.*#+}} xmm1 = xmm1[0,1,2],xmm2[0]
689 ; NODQ-64-NEXT: vpextrq $1, %xmm0, %rax
690 ; NODQ-64-NEXT: vcvtusi2ss %rax, %xmm4, %xmm2
691 ; NODQ-64-NEXT: vmovq %xmm0, %rax
692 ; NODQ-64-NEXT: vcvtusi2ss %rax, %xmm4, %xmm3
693 ; NODQ-64-NEXT: vinsertps {{.*#+}} xmm2 = xmm3[0],xmm2[0],xmm3[2,3]
694 ; NODQ-64-NEXT: vextracti128 $1, %ymm0, %xmm0
695 ; NODQ-64-NEXT: vmovq %xmm0, %rax
696 ; NODQ-64-NEXT: vcvtusi2ss %rax, %xmm4, %xmm3
697 ; NODQ-64-NEXT: vinsertps {{.*#+}} xmm2 = xmm2[0,1],xmm3[0],xmm2[3]
698 ; NODQ-64-NEXT: vpextrq $1, %xmm0, %rax
699 ; NODQ-64-NEXT: vcvtusi2ss %rax, %xmm4, %xmm0
700 ; NODQ-64-NEXT: vinsertps {{.*#+}} xmm0 = xmm2[0,1,2],xmm0[0]
701 ; NODQ-64-NEXT: vinsertf128 $1, %xmm1, %ymm0, %ymm0
702 ; NODQ-64-NEXT: retq
703 ;
704 ; DQ-LABEL: uitofp_v8i64_v8f32:
705 ; DQ: # %bb.0:
706 ; DQ-NEXT: vcvtuqq2ps %zmm0, %ymm0
707 ; DQ-NEXT: ret{{[l|q]}}
708 %result = call <8 x float> @llvm.experimental.constrained.uitofp.v8f32.v8i64(<8 x i64> %x,
709 metadata !"round.dynamic",
710 metadata !"fpexcept.strict") #0
711 ret <8 x float> %result
712 }
714 attributes #0 = { strictfp }