| blob | be8674741e409e74a20b49497fe7d0af3c97f8f6 |
1 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+sse4.1 | FileCheck %s --check-prefix=SSE
2 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx2 | FileCheck %s --check-prefix=AVX
4 ;
5 ; NOTE: this is generated by utils/update_llc_test_checks.py but we can't check NAN types (PR30443),
6 ; so we need to edit it to remove the NAN constant comments
7 ;
9 ; copysign(x, c1) -> fabs(x) iff ispos(c1)
10 define <4 x float> @combine_vec_fcopysign_pos_constant0(<4 x float> %x) {
11 ; SSE-LABEL: combine_vec_fcopysign_pos_constant0:
12 ; SSE: # BB#0:
13 ; SSE-NEXT: movaps {{.*#+}} xmm1 = [2.000000e+00,2.000000e+00,2.000000e+00,2.000000e+00]
14 ; SSE-NEXT: andps {{.*}}(%rip), %xmm1
15 ; SSE-NEXT: andps {{.*}}(%rip), %xmm0
16 ; SSE-NEXT: orps %xmm1, %xmm0
17 ; SSE-NEXT: retq
18 ;
19 ; AVX-LABEL: combine_vec_fcopysign_pos_constant0:
20 ; AVX: # BB#0:
21 ; AVX-NEXT: vbroadcastss {{.*}}(%rip), %xmm1
22 ; AVX-NEXT: vbroadcastss {{.*}}(%rip), %xmm2
23 ; AVX-NEXT: vandps %xmm1, %xmm2, %xmm1
24 ; AVX-NEXT: vbroadcastss {{.*}}(%rip), %xmm2
25 ; AVX-NEXT: vandps %xmm2, %xmm0, %xmm0
26 ; AVX-NEXT: vorps %xmm1, %xmm0, %xmm0
27 ; AVX-NEXT: retq
28 %1 = call <4 x float> @llvm.copysign.v4f32(<4 x float> %x, <4 x float> <float 2.0, float 2.0, float 2.0, float 2.0>)
29 ret <4 x float> %1
30 }
32 define <4 x float> @combine_vec_fcopysign_pos_constant1(<4 x float> %x) {
33 ; SSE-LABEL: combine_vec_fcopysign_pos_constant1:
34 ; SSE: # BB#0:
35 ; SSE-NEXT: movaps {{.*#+}} xmm1 = [0.000000e+00,2.000000e+00,4.000000e+00,8.000000e+00]
36 ; SSE-NEXT: andps {{.*}}(%rip), %xmm1
37 ; SSE-NEXT: andps {{.*}}(%rip), %xmm0
38 ; SSE-NEXT: orps %xmm1, %xmm0
39 ; SSE-NEXT: retq
40 ;
41 ; AVX-LABEL: combine_vec_fcopysign_pos_constant1:
42 ; AVX: # BB#0:
43 ; AVX-NEXT: vbroadcastss {{.*}}(%rip), %xmm1
44 ; AVX-NEXT: vandps %xmm1, %xmm0, %xmm0
45 ; AVX-NEXT: vbroadcastss {{.*}}(%rip), %xmm1
46 ; AVX-NEXT: vandps {{.*}}(%rip), %xmm1, %xmm1
47 ; AVX-NEXT: vorps %xmm1, %xmm0, %xmm0
48 ; AVX-NEXT: retq
49 %1 = call <4 x float> @llvm.copysign.v4f32(<4 x float> %x, <4 x float> <float 0.0, float 2.0, float 4.0, float 8.0>)
50 ret <4 x float> %1
51 }
53 define <4 x float> @combine_vec_fcopysign_fabs_sgn(<4 x float> %x, <4 x float> %y) {
54 ; SSE-LABEL: combine_vec_fcopysign_fabs_sgn:
55 ; SSE: # BB#0:
56 ; SSE-NEXT: andps {{.*}}(%rip), %xmm0
57 ; SSE-NEXT: retq
58 ;
59 ; AVX-LABEL: combine_vec_fcopysign_fabs_sgn:
60 ; AVX: # BB#0:
61 ; AVX-NEXT: vbroadcastss {{.*}}(%rip), %xmm1
62 ; AVX-NEXT: vandps %xmm1, %xmm0, %xmm0
63 ; AVX-NEXT: retq
64 %1 = call <4 x float> @llvm.fabs.v4f32(<4 x float> %y)
65 %2 = call <4 x float> @llvm.copysign.v4f32(<4 x float> %x, <4 x float> %1)
66 ret <4 x float> %2
67 }
69 ; copysign(x, c1) -> fneg(fabs(x)) iff isneg(c1)
70 define <4 x float> @combine_vec_fcopysign_neg_constant0(<4 x float> %x) {
71 ; SSE-LABEL: combine_vec_fcopysign_neg_constant0:
72 ; SSE: # BB#0:
73 ; SSE-NEXT: movaps {{.*#+}} xmm1 = [-2.000000e+00,-2.000000e+00,-2.000000e+00,-2.000000e+00]
74 ; SSE-NEXT: andps {{.*}}(%rip), %xmm1
75 ; SSE-NEXT: andps {{.*}}(%rip), %xmm0
76 ; SSE-NEXT: orps %xmm1, %xmm0
77 ; SSE-NEXT: retq
78 ;
79 ; AVX-LABEL: combine_vec_fcopysign_neg_constant0:
80 ; AVX: # BB#0:
81 ; AVX-NEXT: vbroadcastss {{.*}}(%rip), %xmm1
82 ; AVX-NEXT: vbroadcastss {{.*}}(%rip), %xmm2
83 ; AVX-NEXT: vandps %xmm1, %xmm2, %xmm1
84 ; AVX-NEXT: vbroadcastss {{.*}}(%rip), %xmm2
85 ; AVX-NEXT: vandps %xmm2, %xmm0, %xmm0
86 ; AVX-NEXT: vorps %xmm1, %xmm0, %xmm0
87 ; AVX-NEXT: retq
88 %1 = call <4 x float> @llvm.copysign.v4f32(<4 x float> %x, <4 x float> <float -2.0, float -2.0, float -2.0, float -2.0>)
89 ret <4 x float> %1
90 }
92 define <4 x float> @combine_vec_fcopysign_neg_constant1(<4 x float> %x) {
93 ; SSE-LABEL: combine_vec_fcopysign_neg_constant1:
94 ; SSE: # BB#0:
95 ; SSE-NEXT: movaps {{.*#+}} xmm1 = [-0.000000e+00,-2.000000e+00,-4.000000e+00,-8.000000e+00]
96 ; SSE-NEXT: andps {{.*}}(%rip), %xmm1
97 ; SSE-NEXT: andps {{.*}}(%rip), %xmm0
98 ; SSE-NEXT: orps %xmm1, %xmm0
99 ; SSE-NEXT: retq
100 ;
101 ; AVX-LABEL: combine_vec_fcopysign_neg_constant1:
102 ; AVX: # BB#0:
103 ; AVX-NEXT: vbroadcastss {{.*}}(%rip), %xmm1
104 ; AVX-NEXT: vandps %xmm1, %xmm0, %xmm0
105 ; AVX-NEXT: vbroadcastss {{.*}}(%rip), %xmm1
106 ; AVX-NEXT: vandps {{.*}}(%rip), %xmm1, %xmm1
107 ; AVX-NEXT: vorps %xmm1, %xmm0, %xmm0
108 ; AVX-NEXT: retq
109 %1 = call <4 x float> @llvm.copysign.v4f32(<4 x float> %x, <4 x float> <float -0.0, float -2.0, float -4.0, float -8.0>)
110 ret <4 x float> %1
111 }
113 define <4 x float> @combine_vec_fcopysign_fneg_fabs_sgn(<4 x float> %x, <4 x float> %y) {
114 ; SSE-LABEL: combine_vec_fcopysign_fneg_fabs_sgn:
115 ; SSE: # BB#0:
116 ; SSE-NEXT: andps {{.*}}(%rip), %xmm0
117 ; SSE-NEXT: orps {{.*}}(%rip), %xmm0
118 ; SSE-NEXT: retq
119 ;
120 ; AVX-LABEL: combine_vec_fcopysign_fneg_fabs_sgn:
121 ; AVX: # BB#0:
122 ; AVX-NEXT: vbroadcastss {{.*}}(%rip), %xmm1
123 ; AVX-NEXT: vbroadcastss {{.*}}(%rip), %xmm2
124 ; AVX-NEXT: vandps %xmm2, %xmm0, %xmm0
125 ; AVX-NEXT: vorps %xmm1, %xmm0, %xmm0
126 ; AVX-NEXT: retq
127 %1 = call <4 x float> @llvm.fabs.v4f32(<4 x float> %y)
128 %2 = fsub <4 x float> <float -0.0, float -0.0, float -0.0, float -0.0>, %1
129 %3 = call <4 x float> @llvm.copysign.v4f32(<4 x float> %x, <4 x float> %2)
130 ret <4 x float> %3
131 }
133 ; copysign(fabs(x), y) -> copysign(x, y)
134 define <4 x float> @combine_vec_fcopysign_fabs_mag(<4 x float> %x, <4 x float> %y) {
135 ; SSE-LABEL: combine_vec_fcopysign_fabs_mag:
136 ; SSE: # BB#0:
137 ; SSE-NEXT: andps {{.*}}(%rip), %xmm1
138 ; SSE-NEXT: andps {{.*}}(%rip), %xmm0
139 ; SSE-NEXT: orps %xmm1, %xmm0
140 ; SSE-NEXT: retq
141 ;
142 ; AVX-LABEL: combine_vec_fcopysign_fabs_mag:
143 ; AVX: # BB#0:
144 ; AVX-NEXT: vbroadcastss {{.*}}(%rip), %xmm2
145 ; AVX-NEXT: vandps %xmm2, %xmm1, %xmm1
146 ; AVX-NEXT: vbroadcastss {{.*}}(%rip), %xmm2
147 ; AVX-NEXT: vandps %xmm2, %xmm0, %xmm0
148 ; AVX-NEXT: vorps %xmm1, %xmm0, %xmm0
149 ; AVX-NEXT: retq
150 %1 = call <4 x float> @llvm.fabs.v4f32(<4 x float> %x)
151 %2 = call <4 x float> @llvm.copysign.v4f32(<4 x float> %1, <4 x float> %y)
152 ret <4 x float> %2
153 }
155 ; copysign(fneg(x), y) -> copysign(x, y)
156 define <4 x float> @combine_vec_fcopysign_fneg_mag(<4 x float> %x, <4 x float> %y) {
157 ; SSE-LABEL: combine_vec_fcopysign_fneg_mag:
158 ; SSE: # BB#0:
159 ; SSE-NEXT: andps {{.*}}(%rip), %xmm1
160 ; SSE-NEXT: andps {{.*}}(%rip), %xmm0
161 ; SSE-NEXT: orps %xmm1, %xmm0
162 ; SSE-NEXT: retq
163 ;
164 ; AVX-LABEL: combine_vec_fcopysign_fneg_mag:
165 ; AVX: # BB#0:
166 ; AVX-NEXT: vbroadcastss {{.*}}(%rip), %xmm2
167 ; AVX-NEXT: vandps %xmm2, %xmm1, %xmm1
168 ; AVX-NEXT: vbroadcastss {{.*}}(%rip), %xmm2
169 ; AVX-NEXT: vandps %xmm2, %xmm0, %xmm0
170 ; AVX-NEXT: vorps %xmm1, %xmm0, %xmm0
171 ; AVX-NEXT: retq
172 %1 = fsub <4 x float> <float -0.0, float -0.0, float -0.0, float -0.0>, %x
173 %2 = call <4 x float> @llvm.copysign.v4f32(<4 x float> %1, <4 x float> %y)
174 ret <4 x float> %2
175 }
177 ; copysign(copysign(x,z), y) -> copysign(x, y)
178 define <4 x float> @combine_vec_fcopysign_fcopysign_mag(<4 x float> %x, <4 x float> %y, <4 x float> %z) {
179 ; SSE-LABEL: combine_vec_fcopysign_fcopysign_mag:
180 ; SSE: # BB#0:
181 ; SSE-NEXT: andps {{.*}}(%rip), %xmm1
182 ; SSE-NEXT: andps {{.*}}(%rip), %xmm0
183 ; SSE-NEXT: orps %xmm1, %xmm0
184 ; SSE-NEXT: retq
185 ;
186 ; AVX-LABEL: combine_vec_fcopysign_fcopysign_mag:
187 ; AVX: # BB#0:
188 ; AVX-NEXT: vbroadcastss {{.*}}(%rip), %xmm2
189 ; AVX-NEXT: vandps %xmm2, %xmm1, %xmm1
190 ; AVX-NEXT: vbroadcastss {{.*}}(%rip), %xmm2
191 ; AVX-NEXT: vandps %xmm2, %xmm0, %xmm0
192 ; AVX-NEXT: vorps %xmm1, %xmm0, %xmm0
193 ; AVX-NEXT: retq
194 %1 = call <4 x float> @llvm.copysign.v4f32(<4 x float> %x, <4 x float> %z)
195 %2 = call <4 x float> @llvm.copysign.v4f32(<4 x float> %1, <4 x float> %y)
196 ret <4 x float> %2
197 }
199 ; copysign(x, copysign(y,z)) -> copysign(x, z)
200 define <4 x float> @combine_vec_fcopysign_fcopysign_sgn(<4 x float> %x, <4 x float> %y, <4 x float> %z) {
201 ; SSE-LABEL: combine_vec_fcopysign_fcopysign_sgn:
202 ; SSE: # BB#0:
203 ; SSE-NEXT: andps {{.*}}(%rip), %xmm2
204 ; SSE-NEXT: andps {{.*}}(%rip), %xmm0
205 ; SSE-NEXT: orps %xmm2, %xmm0
206 ; SSE-NEXT: retq
207 ;
208 ; AVX-LABEL: combine_vec_fcopysign_fcopysign_sgn:
209 ; AVX: # BB#0:
210 ; AVX-NEXT: vbroadcastss {{.*}}(%rip), %xmm1
211 ; AVX-NEXT: vandps %xmm1, %xmm2, %xmm1
212 ; AVX-NEXT: vbroadcastss {{.*}}(%rip), %xmm2
213 ; AVX-NEXT: vandps %xmm2, %xmm0, %xmm0
214 ; AVX-NEXT: vorps %xmm1, %xmm0, %xmm0
215 ; AVX-NEXT: retq
216 %1 = call <4 x float> @llvm.copysign.v4f32(<4 x float> %y, <4 x float> %z)
217 %2 = call <4 x float> @llvm.copysign.v4f32(<4 x float> %x, <4 x float> %1)
218 ret <4 x float> %2
219 }
221 ; copysign(x, fp_extend(y)) -> copysign(x, y)
222 define <4 x double> @combine_vec_fcopysign_fpext_sgn(<4 x double> %x, <4 x float> %y) {
223 ; SSE-LABEL: combine_vec_fcopysign_fpext_sgn:
224 ; SSE: # BB#0:
225 ; SSE-NEXT: movaps %xmm2, %xmm3
226 ; SSE-NEXT: cvtss2sd %xmm2, %xmm4
227 ; SSE-NEXT: movshdup {{.*#+}} xmm5 = xmm2[1,1,3,3]
228 ; SSE-NEXT: movaps %xmm2, %xmm6
229 ; SSE-NEXT: movhlps {{.*#+}} xmm6 = xmm6[1,1]
230 ; SSE-NEXT: shufps {{.*#+}} xmm3 = xmm3[3,1,2,3]
231 ; SSE-NEXT: movaps {{.*#+}} xmm7
232 ; SSE-NEXT: movaps %xmm0, %xmm2
233 ; SSE-NEXT: andps %xmm7, %xmm2
234 ; SSE-NEXT: movaps {{.*#+}} xmm8 = [-0.000000e+00,-0.000000e+00]
235 ; SSE-NEXT: andps %xmm8, %xmm4
236 ; SSE-NEXT: orps %xmm4, %xmm2
237 ; SSE-NEXT: movhlps {{.*#+}} xmm0 = xmm0[1,1]
238 ; SSE-NEXT: andps %xmm7, %xmm0
239 ; SSE-NEXT: xorps %xmm4, %xmm4
240 ; SSE-NEXT: cvtss2sd %xmm5, %xmm4
241 ; SSE-NEXT: andps %xmm8, %xmm4
242 ; SSE-NEXT: orps %xmm0, %xmm4
243 ; SSE-NEXT: movlhps {{.*#+}} xmm2 = xmm2[0],xmm4[0]
244 ; SSE-NEXT: movaps %xmm1, %xmm0
245 ; SSE-NEXT: movhlps {{.*#+}} xmm0 = xmm0[1,1]
246 ; SSE-NEXT: andps %xmm7, %xmm0
247 ; SSE-NEXT: cvtss2sd %xmm3, %xmm3
248 ; SSE-NEXT: andps %xmm8, %xmm3
249 ; SSE-NEXT: orps %xmm0, %xmm3
250 ; SSE-NEXT: andps %xmm7, %xmm1
251 ; SSE-NEXT: xorps %xmm0, %xmm0
252 ; SSE-NEXT: cvtss2sd %xmm6, %xmm0
253 ; SSE-NEXT: andps %xmm8, %xmm0
254 ; SSE-NEXT: orps %xmm0, %xmm1
255 ; SSE-NEXT: movlhps {{.*#+}} xmm1 = xmm1[0],xmm3[0]
256 ; SSE-NEXT: movaps %xmm2, %xmm0
257 ; SSE-NEXT: retq
258 ;
259 ; AVX-LABEL: combine_vec_fcopysign_fpext_sgn:
260 ; AVX: # BB#0:
261 ; AVX-NEXT: vbroadcastsd {{.*}}(%rip), %ymm2
262 ; AVX-NEXT: vandps %ymm2, %ymm0, %ymm0
263 ; AVX-NEXT: vcvtps2pd %xmm1, %ymm1
264 ; AVX-NEXT: vbroadcastsd {{.*}}(%rip), %ymm2
265 ; AVX-NEXT: vandps %ymm2, %ymm1, %ymm1
266 ; AVX-NEXT: vorps %ymm1, %ymm0, %ymm0
267 ; AVX-NEXT: retq
268 %1 = fpext <4 x float> %y to <4 x double>
269 %2 = call <4 x double> @llvm.copysign.v4f64(<4 x double> %x, <4 x double> %1)
270 ret <4 x double> %2
271 }
273 ; copysign(x, fp_round(y)) -> copysign(x, y)
274 define <4 x float> @combine_vec_fcopysign_fptrunc_sgn(<4 x float> %x, <4 x double> %y) {
275 ; SSE-LABEL: combine_vec_fcopysign_fptrunc_sgn:
276 ; SSE: # BB#0:
277 ; SSE-NEXT: movaps %xmm0, %xmm3
278 ; SSE-NEXT: movaps {{.*#+}} xmm5
279 ; SSE-NEXT: andps %xmm5, %xmm0
280 ; SSE-NEXT: cvtsd2ss %xmm1, %xmm6
281 ; SSE-NEXT: movaps {{.*#+}} xmm4 = [-0.000000e+00,-0.000000e+00,-0.000000e+00,-0.000000e+00]
282 ; SSE-NEXT: andps %xmm4, %xmm6
283 ; SSE-NEXT: orps %xmm6, %xmm0
284 ; SSE-NEXT: movshdup {{.*#+}} xmm6 = xmm3[1,1,3,3]
285 ; SSE-NEXT: andps %xmm5, %xmm6
286 ; SSE-NEXT: movhlps {{.*#+}} xmm1 = xmm1[1,1]
287 ; SSE-NEXT: cvtsd2ss %xmm1, %xmm1
288 ; SSE-NEXT: andps %xmm4, %xmm1
289 ; SSE-NEXT: orps %xmm6, %xmm1
290 ; SSE-NEXT: unpcklps {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
291 ; SSE-NEXT: movaps %xmm3, %xmm1
292 ; SSE-NEXT: movhlps {{.*#+}} xmm1 = xmm1[1,1]
293 ; SSE-NEXT: andps %xmm5, %xmm1
294 ; SSE-NEXT: xorps %xmm6, %xmm6
295 ; SSE-NEXT: cvtsd2ss %xmm2, %xmm6
296 ; SSE-NEXT: andps %xmm4, %xmm6
297 ; SSE-NEXT: orps %xmm1, %xmm6
298 ; SSE-NEXT: insertps {{.*#+}} xmm0 = xmm0[0,1],xmm6[0],xmm0[3]
299 ; SSE-NEXT: shufps {{.*#+}} xmm3 = xmm3[3,1,2,3]
300 ; SSE-NEXT: andps %xmm5, %xmm3
301 ; SSE-NEXT: movhlps {{.*#+}} xmm2 = xmm2[1,1]
302 ; SSE-NEXT: xorps %xmm1, %xmm1
303 ; SSE-NEXT: cvtsd2ss %xmm2, %xmm1
304 ; SSE-NEXT: andps %xmm4, %xmm1
305 ; SSE-NEXT: orps %xmm3, %xmm1
306 ; SSE-NEXT: insertps {{.*#+}} xmm0 = xmm0[0,1,2],xmm1[0]
307 ; SSE-NEXT: retq
308 ;
309 ; AVX-LABEL: combine_vec_fcopysign_fptrunc_sgn:
310 ; AVX: # BB#0:
311 ; AVX-NEXT: vbroadcastss {{.*}}(%rip), %xmm2
312 ; AVX-NEXT: vandpd %xmm2, %xmm0, %xmm0
313 ; AVX-NEXT: vcvtpd2ps %ymm1, %xmm1
314 ; AVX-NEXT: vbroadcastss {{.*}}(%rip), %xmm2
315 ; AVX-NEXT: vandpd %xmm2, %xmm1, %xmm1
316 ; AVX-NEXT: vorpd %xmm1, %xmm0, %xmm0
317 ; AVX-NEXT: vzeroupper
318 ; AVX-NEXT: retq
319 %1 = fptrunc <4 x double> %y to <4 x float>
320 %2 = call <4 x float> @llvm.copysign.v4f32(<4 x float> %x, <4 x float> %1)
321 ret <4 x float> %2
322 }
324 declare <4 x float> @llvm.fabs.v4f32(<4 x float> %p)
325 declare <4 x float> @llvm.copysign.v4f32(<4 x float> %Mag, <4 x float> %Sgn)
326 declare <4 x double> @llvm.copysign.v4f64(<4 x double> %Mag, <4 x double> %Sgn)