| blob | dc93d5be9b3f38e6df0ed8cf2689d20960fc854b |
1 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 2
2 ; RUN: llc -mtriple=aarch64 -verify-machineinstrs %s -o - | FileCheck %s --check-prefixes=CHECK,CHECK-SD
3 ; RUN: llc -mtriple=aarch64 -global-isel -verify-machineinstrs %s -o - | FileCheck %s --check-prefixes=CHECK,CHECK-GI
5 define double @pow_f64(double %a, double %b) {
6 ; CHECK-LABEL: pow_f64:
7 ; CHECK: // %bb.0: // %entry
8 ; CHECK-NEXT: b pow
9 entry:
10 %c = call double @llvm.pow.f64(double %a, double %b)
11 ret double %c
12 }
14 define float @pow_f32(float %a, float %b) {
15 ; CHECK-LABEL: pow_f32:
16 ; CHECK: // %bb.0: // %entry
17 ; CHECK-NEXT: b powf
18 entry:
19 %c = call float @llvm.pow.f32(float %a, float %b)
20 ret float %c
21 }
23 define half @pow_f16(half %a, half %b) {
24 ; CHECK-LABEL: pow_f16:
25 ; CHECK: // %bb.0: // %entry
26 ; CHECK-NEXT: str x30, [sp, #-16]! // 8-byte Folded Spill
27 ; CHECK-NEXT: .cfi_def_cfa_offset 16
28 ; CHECK-NEXT: .cfi_offset w30, -16
29 ; CHECK-NEXT: fcvt s0, h0
30 ; CHECK-NEXT: fcvt s1, h1
31 ; CHECK-NEXT: bl powf
32 ; CHECK-NEXT: fcvt h0, s0
33 ; CHECK-NEXT: ldr x30, [sp], #16 // 8-byte Folded Reload
34 ; CHECK-NEXT: ret
35 entry:
36 %c = call half @llvm.pow.f16(half %a, half %b)
37 ret half %c
38 }
40 define fp128 @pow_fp128(fp128 %a, fp128 %b) {
41 ; CHECK-LABEL: pow_fp128:
42 ; CHECK: // %bb.0: // %entry
43 ; CHECK-NEXT: b powl
44 entry:
45 %c = call fp128 @llvm.pow.fp128(fp128 %a, fp128 %b)
46 ret fp128 %c
47 }
49 define <1 x double> @pow_v1f64(<1 x double> %x) {
50 ; CHECK-LABEL: pow_v1f64:
51 ; CHECK: // %bb.0:
52 ; CHECK-NEXT: str x30, [sp, #-16]! // 8-byte Folded Spill
53 ; CHECK-NEXT: .cfi_def_cfa_offset 16
54 ; CHECK-NEXT: .cfi_offset w30, -16
55 ; CHECK-NEXT: adrp x8, .LCPI4_0
56 ; CHECK-NEXT: ldr d1, [x8, :lo12:.LCPI4_0]
57 ; CHECK-NEXT: bl pow
58 ; CHECK-NEXT: ldr x30, [sp], #16 // 8-byte Folded Reload
59 ; CHECK-NEXT: ret
60 %c = call <1 x double> @llvm.pow.v1f64(<1 x double> %x, <1 x double> <double 3.140000e+00>)
61 ret <1 x double> %c
62 }
64 define <2 x double> @pow_v2f64(<2 x double> %a, <2 x double> %b) {
65 ; CHECK-SD-LABEL: pow_v2f64:
66 ; CHECK-SD: // %bb.0: // %entry
67 ; CHECK-SD-NEXT: sub sp, sp, #64
68 ; CHECK-SD-NEXT: str x30, [sp, #48] // 8-byte Folded Spill
69 ; CHECK-SD-NEXT: .cfi_def_cfa_offset 64
70 ; CHECK-SD-NEXT: .cfi_offset w30, -16
71 ; CHECK-SD-NEXT: stp q0, q1, [sp] // 32-byte Folded Spill
72 ; CHECK-SD-NEXT: mov d0, v0.d[1]
73 ; CHECK-SD-NEXT: mov d1, v1.d[1]
74 ; CHECK-SD-NEXT: bl pow
75 ; CHECK-SD-NEXT: // kill: def $d0 killed $d0 def $q0
76 ; CHECK-SD-NEXT: str q0, [sp, #32] // 16-byte Folded Spill
77 ; CHECK-SD-NEXT: ldp q0, q1, [sp] // 32-byte Folded Reload
78 ; CHECK-SD-NEXT: // kill: def $d0 killed $d0 killed $q0
79 ; CHECK-SD-NEXT: // kill: def $d1 killed $d1 killed $q1
80 ; CHECK-SD-NEXT: bl pow
81 ; CHECK-SD-NEXT: ldr q1, [sp, #32] // 16-byte Folded Reload
82 ; CHECK-SD-NEXT: // kill: def $d0 killed $d0 def $q0
83 ; CHECK-SD-NEXT: ldr x30, [sp, #48] // 8-byte Folded Reload
84 ; CHECK-SD-NEXT: mov v0.d[1], v1.d[0]
85 ; CHECK-SD-NEXT: add sp, sp, #64
86 ; CHECK-SD-NEXT: ret
87 ;
88 ; CHECK-GI-LABEL: pow_v2f64:
89 ; CHECK-GI: // %bb.0: // %entry
90 ; CHECK-GI-NEXT: sub sp, sp, #48
91 ; CHECK-GI-NEXT: stp d9, d8, [sp, #16] // 16-byte Folded Spill
92 ; CHECK-GI-NEXT: str x30, [sp, #32] // 8-byte Folded Spill
93 ; CHECK-GI-NEXT: .cfi_def_cfa_offset 48
94 ; CHECK-GI-NEXT: .cfi_offset w30, -16
95 ; CHECK-GI-NEXT: .cfi_offset b8, -24
96 ; CHECK-GI-NEXT: .cfi_offset b9, -32
97 ; CHECK-GI-NEXT: mov d8, v0.d[1]
98 ; CHECK-GI-NEXT: mov d9, v1.d[1]
99 ; CHECK-GI-NEXT: // kill: def $d0 killed $d0 killed $q0
100 ; CHECK-GI-NEXT: // kill: def $d1 killed $d1 killed $q1
101 ; CHECK-GI-NEXT: bl pow
102 ; CHECK-GI-NEXT: // kill: def $d0 killed $d0 def $q0
103 ; CHECK-GI-NEXT: str q0, [sp] // 16-byte Folded Spill
104 ; CHECK-GI-NEXT: fmov d1, d9
105 ; CHECK-GI-NEXT: fmov d0, d8
106 ; CHECK-GI-NEXT: bl pow
107 ; CHECK-GI-NEXT: ldr q1, [sp] // 16-byte Folded Reload
108 ; CHECK-GI-NEXT: // kill: def $d0 killed $d0 def $q0
109 ; CHECK-GI-NEXT: ldp d9, d8, [sp, #16] // 16-byte Folded Reload
110 ; CHECK-GI-NEXT: ldr x30, [sp, #32] // 8-byte Folded Reload
111 ; CHECK-GI-NEXT: mov v1.d[1], v0.d[0]
112 ; CHECK-GI-NEXT: mov v0.16b, v1.16b
113 ; CHECK-GI-NEXT: add sp, sp, #48
114 ; CHECK-GI-NEXT: ret
115 entry:
116 %c = call <2 x double> @llvm.pow.v2f64(<2 x double> %a, <2 x double> %b)
117 ret <2 x double> %c
118 }
120 define <3 x double> @pow_v3f64(<3 x double> %a, <3 x double> %b) {
121 ; CHECK-SD-LABEL: pow_v3f64:
122 ; CHECK-SD: // %bb.0: // %entry
123 ; CHECK-SD-NEXT: str d12, [sp, #-48]! // 8-byte Folded Spill
124 ; CHECK-SD-NEXT: stp d11, d10, [sp, #8] // 16-byte Folded Spill
125 ; CHECK-SD-NEXT: stp d9, d8, [sp, #24] // 16-byte Folded Spill
126 ; CHECK-SD-NEXT: str x30, [sp, #40] // 8-byte Folded Spill
127 ; CHECK-SD-NEXT: .cfi_def_cfa_offset 48
128 ; CHECK-SD-NEXT: .cfi_offset w30, -8
129 ; CHECK-SD-NEXT: .cfi_offset b8, -16
130 ; CHECK-SD-NEXT: .cfi_offset b9, -24
131 ; CHECK-SD-NEXT: .cfi_offset b10, -32
132 ; CHECK-SD-NEXT: .cfi_offset b11, -40
133 ; CHECK-SD-NEXT: .cfi_offset b12, -48
134 ; CHECK-SD-NEXT: fmov d11, d1
135 ; CHECK-SD-NEXT: fmov d1, d3
136 ; CHECK-SD-NEXT: fmov d8, d5
137 ; CHECK-SD-NEXT: fmov d9, d4
138 ; CHECK-SD-NEXT: fmov d10, d2
139 ; CHECK-SD-NEXT: bl pow
140 ; CHECK-SD-NEXT: fmov d12, d0
141 ; CHECK-SD-NEXT: fmov d0, d11
142 ; CHECK-SD-NEXT: fmov d1, d9
143 ; CHECK-SD-NEXT: bl pow
144 ; CHECK-SD-NEXT: fmov d9, d0
145 ; CHECK-SD-NEXT: fmov d0, d10
146 ; CHECK-SD-NEXT: fmov d1, d8
147 ; CHECK-SD-NEXT: bl pow
148 ; CHECK-SD-NEXT: fmov d1, d9
149 ; CHECK-SD-NEXT: ldp d9, d8, [sp, #24] // 16-byte Folded Reload
150 ; CHECK-SD-NEXT: ldp d11, d10, [sp, #8] // 16-byte Folded Reload
151 ; CHECK-SD-NEXT: fmov d2, d0
152 ; CHECK-SD-NEXT: ldr x30, [sp, #40] // 8-byte Folded Reload
153 ; CHECK-SD-NEXT: fmov d0, d12
154 ; CHECK-SD-NEXT: ldr d12, [sp], #48 // 8-byte Folded Reload
155 ; CHECK-SD-NEXT: ret
156 ;
157 ; CHECK-GI-LABEL: pow_v3f64:
158 ; CHECK-GI: // %bb.0: // %entry
159 ; CHECK-GI-NEXT: str d12, [sp, #-48]! // 8-byte Folded Spill
160 ; CHECK-GI-NEXT: stp d11, d10, [sp, #8] // 16-byte Folded Spill
161 ; CHECK-GI-NEXT: stp d9, d8, [sp, #24] // 16-byte Folded Spill
162 ; CHECK-GI-NEXT: str x30, [sp, #40] // 8-byte Folded Spill
163 ; CHECK-GI-NEXT: .cfi_def_cfa_offset 48
164 ; CHECK-GI-NEXT: .cfi_offset w30, -8
165 ; CHECK-GI-NEXT: .cfi_offset b8, -16
166 ; CHECK-GI-NEXT: .cfi_offset b9, -24
167 ; CHECK-GI-NEXT: .cfi_offset b10, -32
168 ; CHECK-GI-NEXT: .cfi_offset b11, -40
169 ; CHECK-GI-NEXT: .cfi_offset b12, -48
170 ; CHECK-GI-NEXT: fmov d8, d1
171 ; CHECK-GI-NEXT: fmov d1, d3
172 ; CHECK-GI-NEXT: fmov d9, d2
173 ; CHECK-GI-NEXT: fmov d10, d4
174 ; CHECK-GI-NEXT: fmov d11, d5
175 ; CHECK-GI-NEXT: bl pow
176 ; CHECK-GI-NEXT: fmov d12, d0
177 ; CHECK-GI-NEXT: fmov d0, d8
178 ; CHECK-GI-NEXT: fmov d1, d10
179 ; CHECK-GI-NEXT: bl pow
180 ; CHECK-GI-NEXT: fmov d8, d0
181 ; CHECK-GI-NEXT: fmov d0, d9
182 ; CHECK-GI-NEXT: fmov d1, d11
183 ; CHECK-GI-NEXT: bl pow
184 ; CHECK-GI-NEXT: fmov d1, d8
185 ; CHECK-GI-NEXT: ldp d9, d8, [sp, #24] // 16-byte Folded Reload
186 ; CHECK-GI-NEXT: ldp d11, d10, [sp, #8] // 16-byte Folded Reload
187 ; CHECK-GI-NEXT: fmov d2, d0
188 ; CHECK-GI-NEXT: ldr x30, [sp, #40] // 8-byte Folded Reload
189 ; CHECK-GI-NEXT: fmov d0, d12
190 ; CHECK-GI-NEXT: ldr d12, [sp], #48 // 8-byte Folded Reload
191 ; CHECK-GI-NEXT: ret
192 entry:
193 %c = call <3 x double> @llvm.pow.v3f64(<3 x double> %a, <3 x double> %b)
194 ret <3 x double> %c
195 }
197 define <4 x double> @pow_v4f64(<4 x double> %a, <4 x double> %b) {
198 ; CHECK-SD-LABEL: pow_v4f64:
199 ; CHECK-SD: // %bb.0: // %entry
200 ; CHECK-SD-NEXT: sub sp, sp, #96
201 ; CHECK-SD-NEXT: str x30, [sp, #80] // 8-byte Folded Spill
202 ; CHECK-SD-NEXT: .cfi_def_cfa_offset 96
203 ; CHECK-SD-NEXT: .cfi_offset w30, -16
204 ; CHECK-SD-NEXT: stp q0, q2, [sp] // 32-byte Folded Spill
205 ; CHECK-SD-NEXT: mov d0, v0.d[1]
206 ; CHECK-SD-NEXT: stp q1, q3, [sp, #48] // 32-byte Folded Spill
207 ; CHECK-SD-NEXT: mov d1, v2.d[1]
208 ; CHECK-SD-NEXT: bl pow
209 ; CHECK-SD-NEXT: // kill: def $d0 killed $d0 def $q0
210 ; CHECK-SD-NEXT: str q0, [sp, #32] // 16-byte Folded Spill
211 ; CHECK-SD-NEXT: ldp q0, q1, [sp] // 32-byte Folded Reload
212 ; CHECK-SD-NEXT: // kill: def $d0 killed $d0 killed $q0
213 ; CHECK-SD-NEXT: // kill: def $d1 killed $d1 killed $q1
214 ; CHECK-SD-NEXT: bl pow
215 ; CHECK-SD-NEXT: ldr q1, [sp, #32] // 16-byte Folded Reload
216 ; CHECK-SD-NEXT: // kill: def $d0 killed $d0 def $q0
217 ; CHECK-SD-NEXT: mov v0.d[1], v1.d[0]
218 ; CHECK-SD-NEXT: str q0, [sp, #32] // 16-byte Folded Spill
219 ; CHECK-SD-NEXT: ldp q0, q1, [sp, #48] // 32-byte Folded Reload
220 ; CHECK-SD-NEXT: mov d0, v0.d[1]
221 ; CHECK-SD-NEXT: mov d1, v1.d[1]
222 ; CHECK-SD-NEXT: bl pow
223 ; CHECK-SD-NEXT: // kill: def $d0 killed $d0 def $q0
224 ; CHECK-SD-NEXT: str q0, [sp, #16] // 16-byte Folded Spill
225 ; CHECK-SD-NEXT: ldp q0, q1, [sp, #48] // 32-byte Folded Reload
226 ; CHECK-SD-NEXT: // kill: def $d0 killed $d0 killed $q0
227 ; CHECK-SD-NEXT: // kill: def $d1 killed $d1 killed $q1
228 ; CHECK-SD-NEXT: bl pow
229 ; CHECK-SD-NEXT: fmov d1, d0
230 ; CHECK-SD-NEXT: ldp q2, q0, [sp, #16] // 32-byte Folded Reload
231 ; CHECK-SD-NEXT: ldr x30, [sp, #80] // 8-byte Folded Reload
232 ; CHECK-SD-NEXT: mov v1.d[1], v2.d[0]
233 ; CHECK-SD-NEXT: add sp, sp, #96
234 ; CHECK-SD-NEXT: ret
235 ;
236 ; CHECK-GI-LABEL: pow_v4f64:
237 ; CHECK-GI: // %bb.0: // %entry
238 ; CHECK-GI-NEXT: sub sp, sp, #112
239 ; CHECK-GI-NEXT: stp d11, d10, [sp, #64] // 16-byte Folded Spill
240 ; CHECK-GI-NEXT: stp d9, d8, [sp, #80] // 16-byte Folded Spill
241 ; CHECK-GI-NEXT: str x30, [sp, #96] // 8-byte Folded Spill
242 ; CHECK-GI-NEXT: .cfi_def_cfa_offset 112
243 ; CHECK-GI-NEXT: .cfi_offset w30, -16
244 ; CHECK-GI-NEXT: .cfi_offset b8, -24
245 ; CHECK-GI-NEXT: .cfi_offset b9, -32
246 ; CHECK-GI-NEXT: .cfi_offset b10, -40
247 ; CHECK-GI-NEXT: .cfi_offset b11, -48
248 ; CHECK-GI-NEXT: mov v4.16b, v1.16b
249 ; CHECK-GI-NEXT: str q1, [sp, #32] // 16-byte Folded Spill
250 ; CHECK-GI-NEXT: mov v1.16b, v2.16b
251 ; CHECK-GI-NEXT: str q3, [sp] // 16-byte Folded Spill
252 ; CHECK-GI-NEXT: mov d8, v0.d[1]
253 ; CHECK-GI-NEXT: mov d10, v2.d[1]
254 ; CHECK-GI-NEXT: mov d11, v3.d[1]
255 ; CHECK-GI-NEXT: // kill: def $d0 killed $d0 killed $q0
256 ; CHECK-GI-NEXT: // kill: def $d1 killed $d1 killed $q1
257 ; CHECK-GI-NEXT: mov d9, v4.d[1]
258 ; CHECK-GI-NEXT: bl pow
259 ; CHECK-GI-NEXT: // kill: def $d0 killed $d0 def $q0
260 ; CHECK-GI-NEXT: str q0, [sp, #48] // 16-byte Folded Spill
261 ; CHECK-GI-NEXT: fmov d1, d10
262 ; CHECK-GI-NEXT: fmov d0, d8
263 ; CHECK-GI-NEXT: bl pow
264 ; CHECK-GI-NEXT: // kill: def $d0 killed $d0 def $q0
265 ; CHECK-GI-NEXT: str q0, [sp, #16] // 16-byte Folded Spill
266 ; CHECK-GI-NEXT: ldr q1, [sp] // 16-byte Folded Reload
267 ; CHECK-GI-NEXT: ldr q0, [sp, #32] // 16-byte Folded Reload
268 ; CHECK-GI-NEXT: // kill: def $d1 killed $d1 killed $q1
269 ; CHECK-GI-NEXT: // kill: def $d0 killed $d0 killed $q0
270 ; CHECK-GI-NEXT: bl pow
271 ; CHECK-GI-NEXT: // kill: def $d0 killed $d0 def $q0
272 ; CHECK-GI-NEXT: str q0, [sp, #32] // 16-byte Folded Spill
273 ; CHECK-GI-NEXT: fmov d1, d11
274 ; CHECK-GI-NEXT: fmov d0, d9
275 ; CHECK-GI-NEXT: bl pow
276 ; CHECK-GI-NEXT: ldp q3, q1, [sp, #16] // 32-byte Folded Reload
277 ; CHECK-GI-NEXT: // kill: def $d0 killed $d0 def $q0
278 ; CHECK-GI-NEXT: ldr q2, [sp, #48] // 16-byte Folded Reload
279 ; CHECK-GI-NEXT: ldp d9, d8, [sp, #80] // 16-byte Folded Reload
280 ; CHECK-GI-NEXT: ldr x30, [sp, #96] // 8-byte Folded Reload
281 ; CHECK-GI-NEXT: ldp d11, d10, [sp, #64] // 16-byte Folded Reload
282 ; CHECK-GI-NEXT: mov v2.d[1], v3.d[0]
283 ; CHECK-GI-NEXT: mov v1.d[1], v0.d[0]
284 ; CHECK-GI-NEXT: mov v0.16b, v2.16b
285 ; CHECK-GI-NEXT: add sp, sp, #112
286 ; CHECK-GI-NEXT: ret
287 entry:
288 %c = call <4 x double> @llvm.pow.v4f64(<4 x double> %a, <4 x double> %b)
289 ret <4 x double> %c
290 }
292 define <2 x float> @pow_v2f32(<2 x float> %a, <2 x float> %b) {
293 ; CHECK-SD-LABEL: pow_v2f32:
294 ; CHECK-SD: // %bb.0: // %entry
295 ; CHECK-SD-NEXT: sub sp, sp, #64
296 ; CHECK-SD-NEXT: str x30, [sp, #48] // 8-byte Folded Spill
297 ; CHECK-SD-NEXT: .cfi_def_cfa_offset 64
298 ; CHECK-SD-NEXT: .cfi_offset w30, -16
299 ; CHECK-SD-NEXT: // kill: def $d1 killed $d1 def $q1
300 ; CHECK-SD-NEXT: // kill: def $d0 killed $d0 def $q0
301 ; CHECK-SD-NEXT: stp q0, q1, [sp] // 32-byte Folded Spill
302 ; CHECK-SD-NEXT: mov s0, v0.s[1]
303 ; CHECK-SD-NEXT: mov s1, v1.s[1]
304 ; CHECK-SD-NEXT: bl powf
305 ; CHECK-SD-NEXT: // kill: def $s0 killed $s0 def $q0
306 ; CHECK-SD-NEXT: str q0, [sp, #32] // 16-byte Folded Spill
307 ; CHECK-SD-NEXT: ldp q0, q1, [sp] // 32-byte Folded Reload
308 ; CHECK-SD-NEXT: // kill: def $s0 killed $s0 killed $q0
309 ; CHECK-SD-NEXT: // kill: def $s1 killed $s1 killed $q1
310 ; CHECK-SD-NEXT: bl powf
311 ; CHECK-SD-NEXT: ldr q1, [sp, #32] // 16-byte Folded Reload
312 ; CHECK-SD-NEXT: // kill: def $s0 killed $s0 def $q0
313 ; CHECK-SD-NEXT: ldr x30, [sp, #48] // 8-byte Folded Reload
314 ; CHECK-SD-NEXT: mov v0.s[1], v1.s[0]
315 ; CHECK-SD-NEXT: // kill: def $d0 killed $d0 killed $q0
316 ; CHECK-SD-NEXT: add sp, sp, #64
317 ; CHECK-SD-NEXT: ret
318 ;
319 ; CHECK-GI-LABEL: pow_v2f32:
320 ; CHECK-GI: // %bb.0: // %entry
321 ; CHECK-GI-NEXT: sub sp, sp, #48
322 ; CHECK-GI-NEXT: stp d9, d8, [sp, #16] // 16-byte Folded Spill
323 ; CHECK-GI-NEXT: str x30, [sp, #32] // 8-byte Folded Spill
324 ; CHECK-GI-NEXT: .cfi_def_cfa_offset 48
325 ; CHECK-GI-NEXT: .cfi_offset w30, -16
326 ; CHECK-GI-NEXT: .cfi_offset b8, -24
327 ; CHECK-GI-NEXT: .cfi_offset b9, -32
328 ; CHECK-GI-NEXT: // kill: def $d0 killed $d0 def $q0
329 ; CHECK-GI-NEXT: // kill: def $d1 killed $d1 def $q1
330 ; CHECK-GI-NEXT: mov s8, v0.s[1]
331 ; CHECK-GI-NEXT: mov s9, v1.s[1]
332 ; CHECK-GI-NEXT: // kill: def $s0 killed $s0 killed $q0
333 ; CHECK-GI-NEXT: // kill: def $s1 killed $s1 killed $q1
334 ; CHECK-GI-NEXT: bl powf
335 ; CHECK-GI-NEXT: // kill: def $s0 killed $s0 def $q0
336 ; CHECK-GI-NEXT: str q0, [sp] // 16-byte Folded Spill
337 ; CHECK-GI-NEXT: fmov s1, s9
338 ; CHECK-GI-NEXT: fmov s0, s8
339 ; CHECK-GI-NEXT: bl powf
340 ; CHECK-GI-NEXT: ldr q1, [sp] // 16-byte Folded Reload
341 ; CHECK-GI-NEXT: // kill: def $s0 killed $s0 def $q0
342 ; CHECK-GI-NEXT: ldp d9, d8, [sp, #16] // 16-byte Folded Reload
343 ; CHECK-GI-NEXT: ldr x30, [sp, #32] // 8-byte Folded Reload
344 ; CHECK-GI-NEXT: mov v1.s[1], v0.s[0]
345 ; CHECK-GI-NEXT: fmov d0, d1
346 ; CHECK-GI-NEXT: add sp, sp, #48
347 ; CHECK-GI-NEXT: ret
348 entry:
349 %c = call <2 x float> @llvm.pow.v2f32(<2 x float> %a, <2 x float> %b)
350 ret <2 x float> %c
351 }
353 define <3 x float> @pow_v3f32(<3 x float> %a, <3 x float> %b) {
354 ; CHECK-SD-LABEL: pow_v3f32:
355 ; CHECK-SD: // %bb.0: // %entry
356 ; CHECK-SD-NEXT: sub sp, sp, #64
357 ; CHECK-SD-NEXT: str x30, [sp, #48] // 8-byte Folded Spill
358 ; CHECK-SD-NEXT: .cfi_def_cfa_offset 64
359 ; CHECK-SD-NEXT: .cfi_offset w30, -16
360 ; CHECK-SD-NEXT: stp q0, q1, [sp, #16] // 32-byte Folded Spill
361 ; CHECK-SD-NEXT: mov s0, v0.s[1]
362 ; CHECK-SD-NEXT: mov s1, v1.s[1]
363 ; CHECK-SD-NEXT: bl powf
364 ; CHECK-SD-NEXT: // kill: def $s0 killed $s0 def $q0
365 ; CHECK-SD-NEXT: str q0, [sp] // 16-byte Folded Spill
366 ; CHECK-SD-NEXT: ldp q0, q1, [sp, #16] // 32-byte Folded Reload
367 ; CHECK-SD-NEXT: // kill: def $s0 killed $s0 killed $q0
368 ; CHECK-SD-NEXT: // kill: def $s1 killed $s1 killed $q1
369 ; CHECK-SD-NEXT: bl powf
370 ; CHECK-SD-NEXT: ldr q1, [sp] // 16-byte Folded Reload
371 ; CHECK-SD-NEXT: // kill: def $s0 killed $s0 def $q0
372 ; CHECK-SD-NEXT: mov v0.s[1], v1.s[0]
373 ; CHECK-SD-NEXT: str q0, [sp] // 16-byte Folded Spill
374 ; CHECK-SD-NEXT: ldp q0, q1, [sp, #16] // 32-byte Folded Reload
375 ; CHECK-SD-NEXT: mov s0, v0.s[2]
376 ; CHECK-SD-NEXT: mov s1, v1.s[2]
377 ; CHECK-SD-NEXT: bl powf
378 ; CHECK-SD-NEXT: ldr q1, [sp] // 16-byte Folded Reload
379 ; CHECK-SD-NEXT: // kill: def $s0 killed $s0 def $q0
380 ; CHECK-SD-NEXT: ldr x30, [sp, #48] // 8-byte Folded Reload
381 ; CHECK-SD-NEXT: mov v1.s[2], v0.s[0]
382 ; CHECK-SD-NEXT: mov v0.16b, v1.16b
383 ; CHECK-SD-NEXT: add sp, sp, #64
384 ; CHECK-SD-NEXT: ret
385 ;
386 ; CHECK-GI-LABEL: pow_v3f32:
387 ; CHECK-GI: // %bb.0: // %entry
388 ; CHECK-GI-NEXT: sub sp, sp, #80
389 ; CHECK-GI-NEXT: stp d11, d10, [sp, #32] // 16-byte Folded Spill
390 ; CHECK-GI-NEXT: stp d9, d8, [sp, #48] // 16-byte Folded Spill
391 ; CHECK-GI-NEXT: str x30, [sp, #64] // 8-byte Folded Spill
392 ; CHECK-GI-NEXT: .cfi_def_cfa_offset 80
393 ; CHECK-GI-NEXT: .cfi_offset w30, -16
394 ; CHECK-GI-NEXT: .cfi_offset b8, -24
395 ; CHECK-GI-NEXT: .cfi_offset b9, -32
396 ; CHECK-GI-NEXT: .cfi_offset b10, -40
397 ; CHECK-GI-NEXT: .cfi_offset b11, -48
398 ; CHECK-GI-NEXT: mov s8, v0.s[1]
399 ; CHECK-GI-NEXT: mov s9, v0.s[2]
400 ; CHECK-GI-NEXT: // kill: def $s0 killed $s0 killed $q0
401 ; CHECK-GI-NEXT: mov s10, v1.s[1]
402 ; CHECK-GI-NEXT: mov s11, v1.s[2]
403 ; CHECK-GI-NEXT: // kill: def $s1 killed $s1 killed $q1
404 ; CHECK-GI-NEXT: bl powf
405 ; CHECK-GI-NEXT: // kill: def $s0 killed $s0 def $q0
406 ; CHECK-GI-NEXT: str q0, [sp, #16] // 16-byte Folded Spill
407 ; CHECK-GI-NEXT: fmov s1, s10
408 ; CHECK-GI-NEXT: fmov s0, s8
409 ; CHECK-GI-NEXT: bl powf
410 ; CHECK-GI-NEXT: // kill: def $s0 killed $s0 def $q0
411 ; CHECK-GI-NEXT: str q0, [sp] // 16-byte Folded Spill
412 ; CHECK-GI-NEXT: fmov s1, s11
413 ; CHECK-GI-NEXT: fmov s0, s9
414 ; CHECK-GI-NEXT: bl powf
415 ; CHECK-GI-NEXT: ldp q2, q1, [sp] // 32-byte Folded Reload
416 ; CHECK-GI-NEXT: // kill: def $s0 killed $s0 def $q0
417 ; CHECK-GI-NEXT: ldr x30, [sp, #64] // 8-byte Folded Reload
418 ; CHECK-GI-NEXT: ldp d9, d8, [sp, #48] // 16-byte Folded Reload
419 ; CHECK-GI-NEXT: ldp d11, d10, [sp, #32] // 16-byte Folded Reload
420 ; CHECK-GI-NEXT: mov v1.s[1], v2.s[0]
421 ; CHECK-GI-NEXT: mov v1.s[2], v0.s[0]
422 ; CHECK-GI-NEXT: mov v0.16b, v1.16b
423 ; CHECK-GI-NEXT: add sp, sp, #80
424 ; CHECK-GI-NEXT: ret
425 entry:
426 %c = call <3 x float> @llvm.pow.v3f32(<3 x float> %a, <3 x float> %b)
427 ret <3 x float> %c
428 }
430 define <4 x float> @pow_v4f32(<4 x float> %a, <4 x float> %b) {
431 ; CHECK-SD-LABEL: pow_v4f32:
432 ; CHECK-SD: // %bb.0: // %entry
433 ; CHECK-SD-NEXT: sub sp, sp, #64
434 ; CHECK-SD-NEXT: str x30, [sp, #48] // 8-byte Folded Spill
435 ; CHECK-SD-NEXT: .cfi_def_cfa_offset 64
436 ; CHECK-SD-NEXT: .cfi_offset w30, -16
437 ; CHECK-SD-NEXT: stp q0, q1, [sp, #16] // 32-byte Folded Spill
438 ; CHECK-SD-NEXT: mov s0, v0.s[1]
439 ; CHECK-SD-NEXT: mov s1, v1.s[1]
440 ; CHECK-SD-NEXT: bl powf
441 ; CHECK-SD-NEXT: // kill: def $s0 killed $s0 def $q0
442 ; CHECK-SD-NEXT: str q0, [sp] // 16-byte Folded Spill
443 ; CHECK-SD-NEXT: ldp q0, q1, [sp, #16] // 32-byte Folded Reload
444 ; CHECK-SD-NEXT: // kill: def $s0 killed $s0 killed $q0
445 ; CHECK-SD-NEXT: // kill: def $s1 killed $s1 killed $q1
446 ; CHECK-SD-NEXT: bl powf
447 ; CHECK-SD-NEXT: ldr q1, [sp] // 16-byte Folded Reload
448 ; CHECK-SD-NEXT: // kill: def $s0 killed $s0 def $q0
449 ; CHECK-SD-NEXT: mov v0.s[1], v1.s[0]
450 ; CHECK-SD-NEXT: str q0, [sp] // 16-byte Folded Spill
451 ; CHECK-SD-NEXT: ldp q0, q1, [sp, #16] // 32-byte Folded Reload
452 ; CHECK-SD-NEXT: mov s0, v0.s[2]
453 ; CHECK-SD-NEXT: mov s1, v1.s[2]
454 ; CHECK-SD-NEXT: bl powf
455 ; CHECK-SD-NEXT: ldr q1, [sp] // 16-byte Folded Reload
456 ; CHECK-SD-NEXT: // kill: def $s0 killed $s0 def $q0
457 ; CHECK-SD-NEXT: mov v1.s[2], v0.s[0]
458 ; CHECK-SD-NEXT: str q1, [sp] // 16-byte Folded Spill
459 ; CHECK-SD-NEXT: ldp q0, q1, [sp, #16] // 32-byte Folded Reload
460 ; CHECK-SD-NEXT: mov s0, v0.s[3]
461 ; CHECK-SD-NEXT: mov s1, v1.s[3]
462 ; CHECK-SD-NEXT: bl powf
463 ; CHECK-SD-NEXT: ldr q1, [sp] // 16-byte Folded Reload
464 ; CHECK-SD-NEXT: // kill: def $s0 killed $s0 def $q0
465 ; CHECK-SD-NEXT: ldr x30, [sp, #48] // 8-byte Folded Reload
466 ; CHECK-SD-NEXT: mov v1.s[3], v0.s[0]
467 ; CHECK-SD-NEXT: mov v0.16b, v1.16b
468 ; CHECK-SD-NEXT: add sp, sp, #64
469 ; CHECK-SD-NEXT: ret
470 ;
471 ; CHECK-GI-LABEL: pow_v4f32:
472 ; CHECK-GI: // %bb.0: // %entry
473 ; CHECK-GI-NEXT: sub sp, sp, #112
474 ; CHECK-GI-NEXT: stp d13, d12, [sp, #48] // 16-byte Folded Spill
475 ; CHECK-GI-NEXT: stp d11, d10, [sp, #64] // 16-byte Folded Spill
476 ; CHECK-GI-NEXT: stp d9, d8, [sp, #80] // 16-byte Folded Spill
477 ; CHECK-GI-NEXT: str x30, [sp, #96] // 8-byte Folded Spill
478 ; CHECK-GI-NEXT: .cfi_def_cfa_offset 112
479 ; CHECK-GI-NEXT: .cfi_offset w30, -16
480 ; CHECK-GI-NEXT: .cfi_offset b8, -24
481 ; CHECK-GI-NEXT: .cfi_offset b9, -32
482 ; CHECK-GI-NEXT: .cfi_offset b10, -40
483 ; CHECK-GI-NEXT: .cfi_offset b11, -48
484 ; CHECK-GI-NEXT: .cfi_offset b12, -56
485 ; CHECK-GI-NEXT: .cfi_offset b13, -64
486 ; CHECK-GI-NEXT: mov s8, v0.s[1]
487 ; CHECK-GI-NEXT: mov s9, v0.s[2]
488 ; CHECK-GI-NEXT: mov s10, v0.s[3]
489 ; CHECK-GI-NEXT: mov s11, v1.s[1]
490 ; CHECK-GI-NEXT: // kill: def $s0 killed $s0 killed $q0
491 ; CHECK-GI-NEXT: mov s12, v1.s[2]
492 ; CHECK-GI-NEXT: mov s13, v1.s[3]
493 ; CHECK-GI-NEXT: // kill: def $s1 killed $s1 killed $q1
494 ; CHECK-GI-NEXT: bl powf
495 ; CHECK-GI-NEXT: // kill: def $s0 killed $s0 def $q0
496 ; CHECK-GI-NEXT: str q0, [sp, #32] // 16-byte Folded Spill
497 ; CHECK-GI-NEXT: fmov s1, s11
498 ; CHECK-GI-NEXT: fmov s0, s8
499 ; CHECK-GI-NEXT: bl powf
500 ; CHECK-GI-NEXT: // kill: def $s0 killed $s0 def $q0
501 ; CHECK-GI-NEXT: str q0, [sp, #16] // 16-byte Folded Spill
502 ; CHECK-GI-NEXT: fmov s1, s12
503 ; CHECK-GI-NEXT: fmov s0, s9
504 ; CHECK-GI-NEXT: bl powf
505 ; CHECK-GI-NEXT: // kill: def $s0 killed $s0 def $q0
506 ; CHECK-GI-NEXT: str q0, [sp] // 16-byte Folded Spill
507 ; CHECK-GI-NEXT: fmov s1, s13
508 ; CHECK-GI-NEXT: fmov s0, s10
509 ; CHECK-GI-NEXT: bl powf
510 ; CHECK-GI-NEXT: ldp q2, q1, [sp, #16] // 32-byte Folded Reload
511 ; CHECK-GI-NEXT: // kill: def $s0 killed $s0 def $q0
512 ; CHECK-GI-NEXT: ldr x30, [sp, #96] // 8-byte Folded Reload
513 ; CHECK-GI-NEXT: ldp d9, d8, [sp, #80] // 16-byte Folded Reload
514 ; CHECK-GI-NEXT: ldp d11, d10, [sp, #64] // 16-byte Folded Reload
515 ; CHECK-GI-NEXT: mov v1.s[1], v2.s[0]
516 ; CHECK-GI-NEXT: ldr q2, [sp] // 16-byte Folded Reload
517 ; CHECK-GI-NEXT: ldp d13, d12, [sp, #48] // 16-byte Folded Reload
518 ; CHECK-GI-NEXT: mov v1.s[2], v2.s[0]
519 ; CHECK-GI-NEXT: mov v1.s[3], v0.s[0]
520 ; CHECK-GI-NEXT: mov v0.16b, v1.16b
521 ; CHECK-GI-NEXT: add sp, sp, #112
522 ; CHECK-GI-NEXT: ret
523 entry:
524 %c = call <4 x float> @llvm.pow.v4f32(<4 x float> %a, <4 x float> %b)
525 ret <4 x float> %c
526 }
528 define <8 x float> @pow_v8f32(<8 x float> %a, <8 x float> %b) {
529 ; CHECK-SD-LABEL: pow_v8f32:
530 ; CHECK-SD: // %bb.0: // %entry
531 ; CHECK-SD-NEXT: sub sp, sp, #96
532 ; CHECK-SD-NEXT: str x30, [sp, #80] // 8-byte Folded Spill
533 ; CHECK-SD-NEXT: .cfi_def_cfa_offset 96
534 ; CHECK-SD-NEXT: .cfi_offset w30, -16
535 ; CHECK-SD-NEXT: stp q0, q2, [sp] // 32-byte Folded Spill
536 ; CHECK-SD-NEXT: mov s0, v0.s[1]
537 ; CHECK-SD-NEXT: stp q1, q3, [sp, #32] // 32-byte Folded Spill
538 ; CHECK-SD-NEXT: mov s1, v2.s[1]
539 ; CHECK-SD-NEXT: bl powf
540 ; CHECK-SD-NEXT: // kill: def $s0 killed $s0 def $q0
541 ; CHECK-SD-NEXT: str q0, [sp, #64] // 16-byte Folded Spill
542 ; CHECK-SD-NEXT: ldp q0, q1, [sp] // 32-byte Folded Reload
543 ; CHECK-SD-NEXT: // kill: def $s0 killed $s0 killed $q0
544 ; CHECK-SD-NEXT: // kill: def $s1 killed $s1 killed $q1
545 ; CHECK-SD-NEXT: bl powf
546 ; CHECK-SD-NEXT: ldr q1, [sp, #64] // 16-byte Folded Reload
547 ; CHECK-SD-NEXT: // kill: def $s0 killed $s0 def $q0
548 ; CHECK-SD-NEXT: mov v0.s[1], v1.s[0]
549 ; CHECK-SD-NEXT: str q0, [sp, #64] // 16-byte Folded Spill
550 ; CHECK-SD-NEXT: ldp q0, q1, [sp] // 32-byte Folded Reload
551 ; CHECK-SD-NEXT: mov s0, v0.s[2]
552 ; CHECK-SD-NEXT: mov s1, v1.s[2]
553 ; CHECK-SD-NEXT: bl powf
554 ; CHECK-SD-NEXT: ldr q1, [sp, #64] // 16-byte Folded Reload
555 ; CHECK-SD-NEXT: // kill: def $s0 killed $s0 def $q0
556 ; CHECK-SD-NEXT: mov v1.s[2], v0.s[0]
557 ; CHECK-SD-NEXT: str q1, [sp, #64] // 16-byte Folded Spill
558 ; CHECK-SD-NEXT: ldp q0, q1, [sp] // 32-byte Folded Reload
559 ; CHECK-SD-NEXT: mov s0, v0.s[3]
560 ; CHECK-SD-NEXT: mov s1, v1.s[3]
561 ; CHECK-SD-NEXT: bl powf
562 ; CHECK-SD-NEXT: ldr q1, [sp, #64] // 16-byte Folded Reload
563 ; CHECK-SD-NEXT: // kill: def $s0 killed $s0 def $q0
564 ; CHECK-SD-NEXT: mov v1.s[3], v0.s[0]
565 ; CHECK-SD-NEXT: str q1, [sp, #64] // 16-byte Folded Spill
566 ; CHECK-SD-NEXT: ldp q0, q1, [sp, #32] // 32-byte Folded Reload
567 ; CHECK-SD-NEXT: mov s0, v0.s[1]
568 ; CHECK-SD-NEXT: mov s1, v1.s[1]
569 ; CHECK-SD-NEXT: bl powf
570 ; CHECK-SD-NEXT: // kill: def $s0 killed $s0 def $q0
571 ; CHECK-SD-NEXT: str q0, [sp, #16] // 16-byte Folded Spill
572 ; CHECK-SD-NEXT: ldp q0, q1, [sp, #32] // 32-byte Folded Reload
573 ; CHECK-SD-NEXT: // kill: def $s0 killed $s0 killed $q0
574 ; CHECK-SD-NEXT: // kill: def $s1 killed $s1 killed $q1
575 ; CHECK-SD-NEXT: bl powf
576 ; CHECK-SD-NEXT: ldr q1, [sp, #16] // 16-byte Folded Reload
577 ; CHECK-SD-NEXT: // kill: def $s0 killed $s0 def $q0
578 ; CHECK-SD-NEXT: mov v0.s[1], v1.s[0]
579 ; CHECK-SD-NEXT: str q0, [sp, #16] // 16-byte Folded Spill
580 ; CHECK-SD-NEXT: ldp q0, q1, [sp, #32] // 32-byte Folded Reload
581 ; CHECK-SD-NEXT: mov s0, v0.s[2]
582 ; CHECK-SD-NEXT: mov s1, v1.s[2]
583 ; CHECK-SD-NEXT: bl powf
584 ; CHECK-SD-NEXT: ldr q1, [sp, #16] // 16-byte Folded Reload
585 ; CHECK-SD-NEXT: // kill: def $s0 killed $s0 def $q0
586 ; CHECK-SD-NEXT: mov v1.s[2], v0.s[0]
587 ; CHECK-SD-NEXT: str q1, [sp, #16] // 16-byte Folded Spill
588 ; CHECK-SD-NEXT: ldp q0, q1, [sp, #32] // 32-byte Folded Reload
589 ; CHECK-SD-NEXT: mov s0, v0.s[3]
590 ; CHECK-SD-NEXT: mov s1, v1.s[3]
591 ; CHECK-SD-NEXT: bl powf
592 ; CHECK-SD-NEXT: fmov s2, s0
593 ; CHECK-SD-NEXT: ldr q1, [sp, #16] // 16-byte Folded Reload
594 ; CHECK-SD-NEXT: ldr q0, [sp, #64] // 16-byte Folded Reload
595 ; CHECK-SD-NEXT: ldr x30, [sp, #80] // 8-byte Folded Reload
596 ; CHECK-SD-NEXT: mov v1.s[3], v2.s[0]
597 ; CHECK-SD-NEXT: add sp, sp, #96
598 ; CHECK-SD-NEXT: ret
599 ;
600 ; CHECK-GI-LABEL: pow_v8f32:
601 ; CHECK-GI: // %bb.0: // %entry
602 ; CHECK-GI-NEXT: sub sp, sp, #208
603 ; CHECK-GI-NEXT: stp d15, d14, [sp, #128] // 16-byte Folded Spill
604 ; CHECK-GI-NEXT: stp d13, d12, [sp, #144] // 16-byte Folded Spill
605 ; CHECK-GI-NEXT: stp d11, d10, [sp, #160] // 16-byte Folded Spill
606 ; CHECK-GI-NEXT: stp d9, d8, [sp, #176] // 16-byte Folded Spill
607 ; CHECK-GI-NEXT: str x30, [sp, #192] // 8-byte Folded Spill
608 ; CHECK-GI-NEXT: .cfi_def_cfa_offset 208
609 ; CHECK-GI-NEXT: .cfi_offset w30, -16
610 ; CHECK-GI-NEXT: .cfi_offset b8, -24
611 ; CHECK-GI-NEXT: .cfi_offset b9, -32
612 ; CHECK-GI-NEXT: .cfi_offset b10, -40
613 ; CHECK-GI-NEXT: .cfi_offset b11, -48
614 ; CHECK-GI-NEXT: .cfi_offset b12, -56
615 ; CHECK-GI-NEXT: .cfi_offset b13, -64
616 ; CHECK-GI-NEXT: .cfi_offset b14, -72
617 ; CHECK-GI-NEXT: .cfi_offset b15, -80
618 ; CHECK-GI-NEXT: mov v4.16b, v1.16b
619 ; CHECK-GI-NEXT: mov v1.16b, v2.16b
620 ; CHECK-GI-NEXT: mov s8, v0.s[1]
621 ; CHECK-GI-NEXT: mov s9, v0.s[2]
622 ; CHECK-GI-NEXT: mov s10, v0.s[3]
623 ; CHECK-GI-NEXT: // kill: def $s0 killed $s0 killed $q0
624 ; CHECK-GI-NEXT: mov s12, v3.s[1]
625 ; CHECK-GI-NEXT: mov s11, v3.s[2]
626 ; CHECK-GI-NEXT: mov s2, v4.s[1]
627 ; CHECK-GI-NEXT: stp q3, q4, [sp] // 32-byte Folded Spill
628 ; CHECK-GI-NEXT: mov s5, v4.s[3]
629 ; CHECK-GI-NEXT: mov s14, v1.s[1]
630 ; CHECK-GI-NEXT: mov s15, v1.s[2]
631 ; CHECK-GI-NEXT: mov s13, v1.s[3]
632 ; CHECK-GI-NEXT: // kill: def $s1 killed $s1 killed $q1
633 ; CHECK-GI-NEXT: str s2, [sp, #48] // 4-byte Folded Spill
634 ; CHECK-GI-NEXT: mov s2, v4.s[2]
635 ; CHECK-GI-NEXT: str s2, [sp, #112] // 4-byte Folded Spill
636 ; CHECK-GI-NEXT: mov s2, v3.s[3]
637 ; CHECK-GI-NEXT: stp s2, s5, [sp, #200] // 8-byte Folded Spill
638 ; CHECK-GI-NEXT: bl powf
639 ; CHECK-GI-NEXT: // kill: def $s0 killed $s0 def $q0
640 ; CHECK-GI-NEXT: str q0, [sp, #96] // 16-byte Folded Spill
641 ; CHECK-GI-NEXT: fmov s1, s14
642 ; CHECK-GI-NEXT: fmov s0, s8
643 ; CHECK-GI-NEXT: bl powf
644 ; CHECK-GI-NEXT: // kill: def $s0 killed $s0 def $q0
645 ; CHECK-GI-NEXT: str q0, [sp, #32] // 16-byte Folded Spill
646 ; CHECK-GI-NEXT: fmov s1, s15
647 ; CHECK-GI-NEXT: fmov s0, s9
648 ; CHECK-GI-NEXT: bl powf
649 ; CHECK-GI-NEXT: // kill: def $s0 killed $s0 def $q0
650 ; CHECK-GI-NEXT: str q0, [sp, #80] // 16-byte Folded Spill
651 ; CHECK-GI-NEXT: fmov s1, s13
652 ; CHECK-GI-NEXT: fmov s0, s10
653 ; CHECK-GI-NEXT: bl powf
654 ; CHECK-GI-NEXT: // kill: def $s0 killed $s0 def $q0
655 ; CHECK-GI-NEXT: str q0, [sp, #64] // 16-byte Folded Spill
656 ; CHECK-GI-NEXT: ldp q1, q0, [sp] // 32-byte Folded Reload
657 ; CHECK-GI-NEXT: // kill: def $s0 killed $s0 killed $q0
658 ; CHECK-GI-NEXT: // kill: def $s1 killed $s1 killed $q1
659 ; CHECK-GI-NEXT: bl powf
660 ; CHECK-GI-NEXT: fmov s1, s12
661 ; CHECK-GI-NEXT: // kill: def $s0 killed $s0 def $q0
662 ; CHECK-GI-NEXT: str q0, [sp, #16] // 16-byte Folded Spill
663 ; CHECK-GI-NEXT: ldr s0, [sp, #48] // 4-byte Folded Reload
664 ; CHECK-GI-NEXT: bl powf
665 ; CHECK-GI-NEXT: fmov s1, s11
666 ; CHECK-GI-NEXT: // kill: def $s0 killed $s0 def $q0
667 ; CHECK-GI-NEXT: str q0, [sp, #48] // 16-byte Folded Spill
668 ; CHECK-GI-NEXT: ldr s0, [sp, #112] // 4-byte Folded Reload
669 ; CHECK-GI-NEXT: bl powf
670 ; CHECK-GI-NEXT: // kill: def $s0 killed $s0 def $q0
671 ; CHECK-GI-NEXT: str q0, [sp, #112] // 16-byte Folded Spill
672 ; CHECK-GI-NEXT: ldp s1, s0, [sp, #200] // 8-byte Folded Reload
673 ; CHECK-GI-NEXT: bl powf
674 ; CHECK-GI-NEXT: ldp q3, q2, [sp, #16] // 32-byte Folded Reload
675 ; CHECK-GI-NEXT: // kill: def $s0 killed $s0 def $q0
676 ; CHECK-GI-NEXT: ldr q1, [sp, #96] // 16-byte Folded Reload
677 ; CHECK-GI-NEXT: ldp d9, d8, [sp, #176] // 16-byte Folded Reload
678 ; CHECK-GI-NEXT: ldr x30, [sp, #192] // 8-byte Folded Reload
679 ; CHECK-GI-NEXT: ldp d11, d10, [sp, #160] // 16-byte Folded Reload
680 ; CHECK-GI-NEXT: mov v1.s[1], v2.s[0]
681 ; CHECK-GI-NEXT: ldr q2, [sp, #48] // 16-byte Folded Reload
682 ; CHECK-GI-NEXT: ldp d13, d12, [sp, #144] // 16-byte Folded Reload
683 ; CHECK-GI-NEXT: mov v3.s[1], v2.s[0]
684 ; CHECK-GI-NEXT: ldr q2, [sp, #80] // 16-byte Folded Reload
685 ; CHECK-GI-NEXT: ldp d15, d14, [sp, #128] // 16-byte Folded Reload
686 ; CHECK-GI-NEXT: mov v1.s[2], v2.s[0]
687 ; CHECK-GI-NEXT: ldr q2, [sp, #112] // 16-byte Folded Reload
688 ; CHECK-GI-NEXT: mov v3.s[2], v2.s[0]
689 ; CHECK-GI-NEXT: ldr q2, [sp, #64] // 16-byte Folded Reload
690 ; CHECK-GI-NEXT: mov v1.s[3], v2.s[0]
691 ; CHECK-GI-NEXT: mov v3.s[3], v0.s[0]
692 ; CHECK-GI-NEXT: mov v2.16b, v1.16b
693 ; CHECK-GI-NEXT: mov v1.16b, v3.16b
694 ; CHECK-GI-NEXT: mov v0.16b, v2.16b
695 ; CHECK-GI-NEXT: add sp, sp, #208
696 ; CHECK-GI-NEXT: ret
697 entry:
698 %c = call <8 x float> @llvm.pow.v8f32(<8 x float> %a, <8 x float> %b)
699 ret <8 x float> %c
700 }
702 define <7 x half> @pow_v7f16(<7 x half> %a, <7 x half> %b) {
703 ; CHECK-SD-LABEL: pow_v7f16:
704 ; CHECK-SD: // %bb.0: // %entry
705 ; CHECK-SD-NEXT: sub sp, sp, #64
706 ; CHECK-SD-NEXT: str x30, [sp, #48] // 8-byte Folded Spill
707 ; CHECK-SD-NEXT: .cfi_def_cfa_offset 64
708 ; CHECK-SD-NEXT: .cfi_offset w30, -16
709 ; CHECK-SD-NEXT: mov h3, v0.h[1]
710 ; CHECK-SD-NEXT: mov h2, v1.h[1]
711 ; CHECK-SD-NEXT: stp q0, q1, [sp] // 32-byte Folded Spill
712 ; CHECK-SD-NEXT: fcvt s0, h3
713 ; CHECK-SD-NEXT: fcvt s1, h2
714 ; CHECK-SD-NEXT: bl powf
715 ; CHECK-SD-NEXT: fcvt h0, s0
716 ; CHECK-SD-NEXT: str q0, [sp, #32] // 16-byte Folded Spill
717 ; CHECK-SD-NEXT: ldp q0, q1, [sp] // 32-byte Folded Reload
718 ; CHECK-SD-NEXT: fcvt s0, h0
719 ; CHECK-SD-NEXT: fcvt s1, h1
720 ; CHECK-SD-NEXT: bl powf
721 ; CHECK-SD-NEXT: fcvt h0, s0
722 ; CHECK-SD-NEXT: ldp q1, q2, [sp, #16] // 32-byte Folded Reload
723 ; CHECK-SD-NEXT: mov h1, v1.h[2]
724 ; CHECK-SD-NEXT: mov v0.h[1], v2.h[0]
725 ; CHECK-SD-NEXT: fcvt s1, h1
726 ; CHECK-SD-NEXT: str q0, [sp, #32] // 16-byte Folded Spill
727 ; CHECK-SD-NEXT: ldr q0, [sp] // 16-byte Folded Reload
728 ; CHECK-SD-NEXT: mov h0, v0.h[2]
729 ; CHECK-SD-NEXT: fcvt s0, h0
730 ; CHECK-SD-NEXT: bl powf
731 ; CHECK-SD-NEXT: fcvt h0, s0
732 ; CHECK-SD-NEXT: ldr q1, [sp, #32] // 16-byte Folded Reload
733 ; CHECK-SD-NEXT: mov v1.h[2], v0.h[0]
734 ; CHECK-SD-NEXT: str q1, [sp, #32] // 16-byte Folded Spill
735 ; CHECK-SD-NEXT: ldp q0, q1, [sp] // 32-byte Folded Reload
736 ; CHECK-SD-NEXT: mov h0, v0.h[3]
737 ; CHECK-SD-NEXT: mov h1, v1.h[3]
738 ; CHECK-SD-NEXT: fcvt s0, h0
739 ; CHECK-SD-NEXT: fcvt s1, h1
740 ; CHECK-SD-NEXT: bl powf
741 ; CHECK-SD-NEXT: fcvt h0, s0
742 ; CHECK-SD-NEXT: ldr q1, [sp, #32] // 16-byte Folded Reload
743 ; CHECK-SD-NEXT: mov v1.h[3], v0.h[0]
744 ; CHECK-SD-NEXT: str q1, [sp, #32] // 16-byte Folded Spill
745 ; CHECK-SD-NEXT: ldp q0, q1, [sp] // 32-byte Folded Reload
746 ; CHECK-SD-NEXT: mov h0, v0.h[4]
747 ; CHECK-SD-NEXT: mov h1, v1.h[4]
748 ; CHECK-SD-NEXT: fcvt s0, h0
749 ; CHECK-SD-NEXT: fcvt s1, h1
750 ; CHECK-SD-NEXT: bl powf
751 ; CHECK-SD-NEXT: fcvt h0, s0
752 ; CHECK-SD-NEXT: ldr q1, [sp, #32] // 16-byte Folded Reload
753 ; CHECK-SD-NEXT: mov v1.h[4], v0.h[0]
754 ; CHECK-SD-NEXT: str q1, [sp, #32] // 16-byte Folded Spill
755 ; CHECK-SD-NEXT: ldp q0, q1, [sp] // 32-byte Folded Reload
756 ; CHECK-SD-NEXT: mov h0, v0.h[5]
757 ; CHECK-SD-NEXT: mov h1, v1.h[5]
758 ; CHECK-SD-NEXT: fcvt s0, h0
759 ; CHECK-SD-NEXT: fcvt s1, h1
760 ; CHECK-SD-NEXT: bl powf
761 ; CHECK-SD-NEXT: fcvt h0, s0
762 ; CHECK-SD-NEXT: ldr q1, [sp, #32] // 16-byte Folded Reload
763 ; CHECK-SD-NEXT: mov v1.h[5], v0.h[0]
764 ; CHECK-SD-NEXT: str q1, [sp, #32] // 16-byte Folded Spill
765 ; CHECK-SD-NEXT: ldp q0, q1, [sp] // 32-byte Folded Reload
766 ; CHECK-SD-NEXT: mov h0, v0.h[6]
767 ; CHECK-SD-NEXT: mov h1, v1.h[6]
768 ; CHECK-SD-NEXT: fcvt s0, h0
769 ; CHECK-SD-NEXT: fcvt s1, h1
770 ; CHECK-SD-NEXT: bl powf
771 ; CHECK-SD-NEXT: fcvt h0, s0
772 ; CHECK-SD-NEXT: ldr q1, [sp, #32] // 16-byte Folded Reload
773 ; CHECK-SD-NEXT: mov v1.h[6], v0.h[0]
774 ; CHECK-SD-NEXT: str q1, [sp, #32] // 16-byte Folded Spill
775 ; CHECK-SD-NEXT: ldp q0, q1, [sp] // 32-byte Folded Reload
776 ; CHECK-SD-NEXT: mov h0, v0.h[7]
777 ; CHECK-SD-NEXT: mov h1, v1.h[7]
778 ; CHECK-SD-NEXT: fcvt s0, h0
779 ; CHECK-SD-NEXT: fcvt s1, h1
780 ; CHECK-SD-NEXT: bl powf
781 ; CHECK-SD-NEXT: fcvt h1, s0
782 ; CHECK-SD-NEXT: ldr q0, [sp, #32] // 16-byte Folded Reload
783 ; CHECK-SD-NEXT: ldr x30, [sp, #48] // 8-byte Folded Reload
784 ; CHECK-SD-NEXT: mov v0.h[7], v1.h[0]
785 ; CHECK-SD-NEXT: add sp, sp, #64
786 ; CHECK-SD-NEXT: ret
787 ;
788 ; CHECK-GI-LABEL: pow_v7f16:
789 ; CHECK-GI: // %bb.0: // %entry
790 ; CHECK-GI-NEXT: sub sp, sp, #176
791 ; CHECK-GI-NEXT: stp d15, d14, [sp, #96] // 16-byte Folded Spill
792 ; CHECK-GI-NEXT: stp d13, d12, [sp, #112] // 16-byte Folded Spill
793 ; CHECK-GI-NEXT: stp d11, d10, [sp, #128] // 16-byte Folded Spill
794 ; CHECK-GI-NEXT: stp d9, d8, [sp, #144] // 16-byte Folded Spill
795 ; CHECK-GI-NEXT: str x30, [sp, #160] // 8-byte Folded Spill
796 ; CHECK-GI-NEXT: .cfi_def_cfa_offset 176
797 ; CHECK-GI-NEXT: .cfi_offset w30, -16
798 ; CHECK-GI-NEXT: .cfi_offset b8, -24
799 ; CHECK-GI-NEXT: .cfi_offset b9, -32
800 ; CHECK-GI-NEXT: .cfi_offset b10, -40
801 ; CHECK-GI-NEXT: .cfi_offset b11, -48
802 ; CHECK-GI-NEXT: .cfi_offset b12, -56
803 ; CHECK-GI-NEXT: .cfi_offset b13, -64
804 ; CHECK-GI-NEXT: .cfi_offset b14, -72
805 ; CHECK-GI-NEXT: .cfi_offset b15, -80
806 ; CHECK-GI-NEXT: mov h2, v0.h[5]
807 ; CHECK-GI-NEXT: mov h9, v0.h[1]
808 ; CHECK-GI-NEXT: mov h10, v0.h[2]
809 ; CHECK-GI-NEXT: mov h11, v0.h[3]
810 ; CHECK-GI-NEXT: mov h12, v0.h[4]
811 ; CHECK-GI-NEXT: mov h14, v1.h[1]
812 ; CHECK-GI-NEXT: mov h15, v1.h[2]
813 ; CHECK-GI-NEXT: mov h8, v1.h[3]
814 ; CHECK-GI-NEXT: mov h13, v1.h[4]
815 ; CHECK-GI-NEXT: str h2, [sp, #48] // 2-byte Folded Spill
816 ; CHECK-GI-NEXT: mov h2, v0.h[6]
817 ; CHECK-GI-NEXT: fcvt s0, h0
818 ; CHECK-GI-NEXT: str h2, [sp, #80] // 2-byte Folded Spill
819 ; CHECK-GI-NEXT: mov h2, v1.h[5]
820 ; CHECK-GI-NEXT: str h2, [sp, #172] // 2-byte Folded Spill
821 ; CHECK-GI-NEXT: mov h2, v1.h[6]
822 ; CHECK-GI-NEXT: fcvt s1, h1
823 ; CHECK-GI-NEXT: str h2, [sp, #174] // 2-byte Folded Spill
824 ; CHECK-GI-NEXT: bl powf
825 ; CHECK-GI-NEXT: fcvt s2, h9
826 ; CHECK-GI-NEXT: fcvt h0, s0
827 ; CHECK-GI-NEXT: fcvt s1, h14
828 ; CHECK-GI-NEXT: str q0, [sp, #64] // 16-byte Folded Spill
829 ; CHECK-GI-NEXT: fmov s0, s2
830 ; CHECK-GI-NEXT: bl powf
831 ; CHECK-GI-NEXT: fcvt s2, h10
832 ; CHECK-GI-NEXT: fcvt h0, s0
833 ; CHECK-GI-NEXT: fcvt s1, h15
834 ; CHECK-GI-NEXT: str q0, [sp, #32] // 16-byte Folded Spill
835 ; CHECK-GI-NEXT: fmov s0, s2
836 ; CHECK-GI-NEXT: bl powf
837 ; CHECK-GI-NEXT: fcvt s2, h11
838 ; CHECK-GI-NEXT: fcvt h0, s0
839 ; CHECK-GI-NEXT: fcvt s1, h8
840 ; CHECK-GI-NEXT: str q0, [sp, #16] // 16-byte Folded Spill
841 ; CHECK-GI-NEXT: fmov s0, s2
842 ; CHECK-GI-NEXT: bl powf
843 ; CHECK-GI-NEXT: fcvt s2, h12
844 ; CHECK-GI-NEXT: fcvt h0, s0
845 ; CHECK-GI-NEXT: fcvt s1, h13
846 ; CHECK-GI-NEXT: str q0, [sp] // 16-byte Folded Spill
847 ; CHECK-GI-NEXT: fmov s0, s2
848 ; CHECK-GI-NEXT: bl powf
849 ; CHECK-GI-NEXT: ldr h1, [sp, #48] // 2-byte Folded Reload
850 ; CHECK-GI-NEXT: fcvt h0, s0
851 ; CHECK-GI-NEXT: fcvt s2, h1
852 ; CHECK-GI-NEXT: str q0, [sp, #48] // 16-byte Folded Spill
853 ; CHECK-GI-NEXT: ldr h0, [sp, #172] // 2-byte Folded Reload
854 ; CHECK-GI-NEXT: fcvt s1, h0
855 ; CHECK-GI-NEXT: fmov s0, s2
856 ; CHECK-GI-NEXT: bl powf
857 ; CHECK-GI-NEXT: ldr h1, [sp, #80] // 2-byte Folded Reload
858 ; CHECK-GI-NEXT: fcvt h0, s0
859 ; CHECK-GI-NEXT: fcvt s2, h1
860 ; CHECK-GI-NEXT: str q0, [sp, #80] // 16-byte Folded Spill
861 ; CHECK-GI-NEXT: ldr h0, [sp, #174] // 2-byte Folded Reload
862 ; CHECK-GI-NEXT: fcvt s1, h0
863 ; CHECK-GI-NEXT: fmov s0, s2
864 ; CHECK-GI-NEXT: bl powf
865 ; CHECK-GI-NEXT: ldp q2, q3, [sp, #16] // 32-byte Folded Reload
866 ; CHECK-GI-NEXT: fcvt h0, s0
867 ; CHECK-GI-NEXT: ldr q1, [sp, #64] // 16-byte Folded Reload
868 ; CHECK-GI-NEXT: ldp d9, d8, [sp, #144] // 16-byte Folded Reload
869 ; CHECK-GI-NEXT: ldp d11, d10, [sp, #128] // 16-byte Folded Reload
870 ; CHECK-GI-NEXT: ldr x30, [sp, #160] // 8-byte Folded Reload
871 ; CHECK-GI-NEXT: mov v1.h[1], v3.h[0]
872 ; CHECK-GI-NEXT: ldp d13, d12, [sp, #112] // 16-byte Folded Reload
873 ; CHECK-GI-NEXT: ldp d15, d14, [sp, #96] // 16-byte Folded Reload
874 ; CHECK-GI-NEXT: mov v1.h[2], v2.h[0]
875 ; CHECK-GI-NEXT: ldr q2, [sp] // 16-byte Folded Reload
876 ; CHECK-GI-NEXT: mov v1.h[3], v2.h[0]
877 ; CHECK-GI-NEXT: ldr q2, [sp, #48] // 16-byte Folded Reload
878 ; CHECK-GI-NEXT: mov v1.h[4], v2.h[0]
879 ; CHECK-GI-NEXT: ldr q2, [sp, #80] // 16-byte Folded Reload
880 ; CHECK-GI-NEXT: mov v1.h[5], v2.h[0]
881 ; CHECK-GI-NEXT: mov v1.h[6], v0.h[0]
882 ; CHECK-GI-NEXT: mov v0.16b, v1.16b
883 ; CHECK-GI-NEXT: add sp, sp, #176
884 ; CHECK-GI-NEXT: ret
885 entry:
886 %c = call <7 x half> @llvm.pow.v7f16(<7 x half> %a, <7 x half> %b)
887 ret <7 x half> %c
888 }
890 define <4 x half> @pow_v4f16(<4 x half> %a, <4 x half> %b) {
891 ; CHECK-SD-LABEL: pow_v4f16:
892 ; CHECK-SD: // %bb.0: // %entry
893 ; CHECK-SD-NEXT: sub sp, sp, #64
894 ; CHECK-SD-NEXT: str x30, [sp, #48] // 8-byte Folded Spill
895 ; CHECK-SD-NEXT: .cfi_def_cfa_offset 64
896 ; CHECK-SD-NEXT: .cfi_offset w30, -16
897 ; CHECK-SD-NEXT: // kill: def $d1 killed $d1 def $q1
898 ; CHECK-SD-NEXT: // kill: def $d0 killed $d0 def $q0
899 ; CHECK-SD-NEXT: mov h3, v0.h[1]
900 ; CHECK-SD-NEXT: mov h2, v1.h[1]
901 ; CHECK-SD-NEXT: stp q0, q1, [sp, #16] // 32-byte Folded Spill
902 ; CHECK-SD-NEXT: fcvt s0, h3
903 ; CHECK-SD-NEXT: fcvt s1, h2
904 ; CHECK-SD-NEXT: bl powf
905 ; CHECK-SD-NEXT: ldr q1, [sp, #16] // 16-byte Folded Reload
906 ; CHECK-SD-NEXT: fcvt h0, s0
907 ; CHECK-SD-NEXT: fcvt s2, h1
908 ; CHECK-SD-NEXT: str q0, [sp] // 16-byte Folded Spill
909 ; CHECK-SD-NEXT: ldr q0, [sp, #32] // 16-byte Folded Reload
910 ; CHECK-SD-NEXT: fcvt s1, h0
911 ; CHECK-SD-NEXT: fmov s0, s2
912 ; CHECK-SD-NEXT: bl powf
913 ; CHECK-SD-NEXT: ldr q1, [sp, #16] // 16-byte Folded Reload
914 ; CHECK-SD-NEXT: fcvt h3, s0
915 ; CHECK-SD-NEXT: ldr q0, [sp, #32] // 16-byte Folded Reload
916 ; CHECK-SD-NEXT: mov h1, v1.h[2]
917 ; CHECK-SD-NEXT: mov h2, v0.h[2]
918 ; CHECK-SD-NEXT: ldr q0, [sp] // 16-byte Folded Reload
919 ; CHECK-SD-NEXT: mov v3.h[1], v0.h[0]
920 ; CHECK-SD-NEXT: fcvt s0, h1
921 ; CHECK-SD-NEXT: fcvt s1, h2
922 ; CHECK-SD-NEXT: str q3, [sp] // 16-byte Folded Spill
923 ; CHECK-SD-NEXT: bl powf
924 ; CHECK-SD-NEXT: ldp q1, q2, [sp, #16] // 32-byte Folded Reload
925 ; CHECK-SD-NEXT: fcvt h0, s0
926 ; CHECK-SD-NEXT: ldr q3, [sp] // 16-byte Folded Reload
927 ; CHECK-SD-NEXT: mov h1, v1.h[3]
928 ; CHECK-SD-NEXT: mov h2, v2.h[3]
929 ; CHECK-SD-NEXT: mov v3.h[2], v0.h[0]
930 ; CHECK-SD-NEXT: fcvt s0, h1
931 ; CHECK-SD-NEXT: fcvt s1, h2
932 ; CHECK-SD-NEXT: str q3, [sp] // 16-byte Folded Spill
933 ; CHECK-SD-NEXT: bl powf
934 ; CHECK-SD-NEXT: fcvt h1, s0
935 ; CHECK-SD-NEXT: ldr q0, [sp] // 16-byte Folded Reload
936 ; CHECK-SD-NEXT: ldr x30, [sp, #48] // 8-byte Folded Reload
937 ; CHECK-SD-NEXT: mov v0.h[3], v1.h[0]
938 ; CHECK-SD-NEXT: // kill: def $d0 killed $d0 killed $q0
939 ; CHECK-SD-NEXT: add sp, sp, #64
940 ; CHECK-SD-NEXT: ret
941 ;
942 ; CHECK-GI-LABEL: pow_v4f16:
943 ; CHECK-GI: // %bb.0: // %entry
944 ; CHECK-GI-NEXT: sub sp, sp, #112
945 ; CHECK-GI-NEXT: stp d13, d12, [sp, #48] // 16-byte Folded Spill
946 ; CHECK-GI-NEXT: stp d11, d10, [sp, #64] // 16-byte Folded Spill
947 ; CHECK-GI-NEXT: stp d9, d8, [sp, #80] // 16-byte Folded Spill
948 ; CHECK-GI-NEXT: str x30, [sp, #96] // 8-byte Folded Spill
949 ; CHECK-GI-NEXT: .cfi_def_cfa_offset 112
950 ; CHECK-GI-NEXT: .cfi_offset w30, -16
951 ; CHECK-GI-NEXT: .cfi_offset b8, -24
952 ; CHECK-GI-NEXT: .cfi_offset b9, -32
953 ; CHECK-GI-NEXT: .cfi_offset b10, -40
954 ; CHECK-GI-NEXT: .cfi_offset b11, -48
955 ; CHECK-GI-NEXT: .cfi_offset b12, -56
956 ; CHECK-GI-NEXT: .cfi_offset b13, -64
957 ; CHECK-GI-NEXT: // kill: def $d0 killed $d0 def $q0
958 ; CHECK-GI-NEXT: // kill: def $d1 killed $d1 def $q1
959 ; CHECK-GI-NEXT: mov h8, v0.h[1]
960 ; CHECK-GI-NEXT: mov h9, v0.h[2]
961 ; CHECK-GI-NEXT: mov h10, v0.h[3]
962 ; CHECK-GI-NEXT: mov h11, v1.h[1]
963 ; CHECK-GI-NEXT: mov h12, v1.h[2]
964 ; CHECK-GI-NEXT: mov h13, v1.h[3]
965 ; CHECK-GI-NEXT: fcvt s0, h0
966 ; CHECK-GI-NEXT: fcvt s1, h1
967 ; CHECK-GI-NEXT: bl powf
968 ; CHECK-GI-NEXT: fcvt s2, h8
969 ; CHECK-GI-NEXT: fcvt h0, s0
970 ; CHECK-GI-NEXT: fcvt s1, h11
971 ; CHECK-GI-NEXT: str q0, [sp, #32] // 16-byte Folded Spill
972 ; CHECK-GI-NEXT: fmov s0, s2
973 ; CHECK-GI-NEXT: bl powf
974 ; CHECK-GI-NEXT: fcvt s2, h9
975 ; CHECK-GI-NEXT: fcvt h0, s0
976 ; CHECK-GI-NEXT: fcvt s1, h12
977 ; CHECK-GI-NEXT: str q0, [sp] // 16-byte Folded Spill
978 ; CHECK-GI-NEXT: fmov s0, s2
979 ; CHECK-GI-NEXT: bl powf
980 ; CHECK-GI-NEXT: fcvt s2, h10
981 ; CHECK-GI-NEXT: fcvt h0, s0
982 ; CHECK-GI-NEXT: fcvt s1, h13
983 ; CHECK-GI-NEXT: str q0, [sp, #16] // 16-byte Folded Spill
984 ; CHECK-GI-NEXT: fmov s0, s2
985 ; CHECK-GI-NEXT: bl powf
986 ; CHECK-GI-NEXT: ldp q3, q2, [sp] // 32-byte Folded Reload
987 ; CHECK-GI-NEXT: fcvt h0, s0
988 ; CHECK-GI-NEXT: ldr q1, [sp, #32] // 16-byte Folded Reload
989 ; CHECK-GI-NEXT: ldp d9, d8, [sp, #80] // 16-byte Folded Reload
990 ; CHECK-GI-NEXT: ldp d11, d10, [sp, #64] // 16-byte Folded Reload
991 ; CHECK-GI-NEXT: ldr x30, [sp, #96] // 8-byte Folded Reload
992 ; CHECK-GI-NEXT: mov v1.h[1], v3.h[0]
993 ; CHECK-GI-NEXT: ldp d13, d12, [sp, #48] // 16-byte Folded Reload
994 ; CHECK-GI-NEXT: mov v1.h[2], v2.h[0]
995 ; CHECK-GI-NEXT: mov v1.h[3], v0.h[0]
996 ; CHECK-GI-NEXT: mov v0.16b, v1.16b
997 ; CHECK-GI-NEXT: // kill: def $d0 killed $d0 killed $q0
998 ; CHECK-GI-NEXT: add sp, sp, #112
999 ; CHECK-GI-NEXT: ret
1000 entry:
1001 %c = call <4 x half> @llvm.pow.v4f16(<4 x half> %a, <4 x half> %b)
1002 ret <4 x half> %c
1003 }
1005 define <8 x half> @pow_v8f16(<8 x half> %a, <8 x half> %b) {
1006 ; CHECK-SD-LABEL: pow_v8f16:
1007 ; CHECK-SD: // %bb.0: // %entry
1008 ; CHECK-SD-NEXT: sub sp, sp, #64
1009 ; CHECK-SD-NEXT: str x30, [sp, #48] // 8-byte Folded Spill
1010 ; CHECK-SD-NEXT: .cfi_def_cfa_offset 64
1011 ; CHECK-SD-NEXT: .cfi_offset w30, -16
1012 ; CHECK-SD-NEXT: mov h3, v0.h[1]
1013 ; CHECK-SD-NEXT: mov h2, v1.h[1]
1014 ; CHECK-SD-NEXT: stp q0, q1, [sp] // 32-byte Folded Spill
1015 ; CHECK-SD-NEXT: fcvt s0, h3
1016 ; CHECK-SD-NEXT: fcvt s1, h2
1017 ; CHECK-SD-NEXT: bl powf
1018 ; CHECK-SD-NEXT: fcvt h0, s0
1019 ; CHECK-SD-NEXT: str q0, [sp, #32] // 16-byte Folded Spill
1020 ; CHECK-SD-NEXT: ldp q0, q1, [sp] // 32-byte Folded Reload
1021 ; CHECK-SD-NEXT: fcvt s0, h0
1022 ; CHECK-SD-NEXT: fcvt s1, h1
1023 ; CHECK-SD-NEXT: bl powf
1024 ; CHECK-SD-NEXT: fcvt h0, s0
1025 ; CHECK-SD-NEXT: ldp q1, q2, [sp, #16] // 32-byte Folded Reload
1026 ; CHECK-SD-NEXT: mov h1, v1.h[2]
1027 ; CHECK-SD-NEXT: mov v0.h[1], v2.h[0]
1028 ; CHECK-SD-NEXT: fcvt s1, h1
1029 ; CHECK-SD-NEXT: str q0, [sp, #32] // 16-byte Folded Spill
1030 ; CHECK-SD-NEXT: ldr q0, [sp] // 16-byte Folded Reload
1031 ; CHECK-SD-NEXT: mov h0, v0.h[2]
1032 ; CHECK-SD-NEXT: fcvt s0, h0
1033 ; CHECK-SD-NEXT: bl powf
1034 ; CHECK-SD-NEXT: fcvt h0, s0
1035 ; CHECK-SD-NEXT: ldr q1, [sp, #32] // 16-byte Folded Reload
1036 ; CHECK-SD-NEXT: mov v1.h[2], v0.h[0]
1037 ; CHECK-SD-NEXT: str q1, [sp, #32] // 16-byte Folded Spill
1038 ; CHECK-SD-NEXT: ldp q0, q1, [sp] // 32-byte Folded Reload
1039 ; CHECK-SD-NEXT: mov h0, v0.h[3]
1040 ; CHECK-SD-NEXT: mov h1, v1.h[3]
1041 ; CHECK-SD-NEXT: fcvt s0, h0
1042 ; CHECK-SD-NEXT: fcvt s1, h1
1043 ; CHECK-SD-NEXT: bl powf
1044 ; CHECK-SD-NEXT: fcvt h0, s0
1045 ; CHECK-SD-NEXT: ldr q1, [sp, #32] // 16-byte Folded Reload
1046 ; CHECK-SD-NEXT: mov v1.h[3], v0.h[0]
1047 ; CHECK-SD-NEXT: str q1, [sp, #32] // 16-byte Folded Spill
1048 ; CHECK-SD-NEXT: ldp q0, q1, [sp] // 32-byte Folded Reload
1049 ; CHECK-SD-NEXT: mov h0, v0.h[4]
1050 ; CHECK-SD-NEXT: mov h1, v1.h[4]
1051 ; CHECK-SD-NEXT: fcvt s0, h0
1052 ; CHECK-SD-NEXT: fcvt s1, h1
1053 ; CHECK-SD-NEXT: bl powf
1054 ; CHECK-SD-NEXT: fcvt h0, s0
1055 ; CHECK-SD-NEXT: ldr q1, [sp, #32] // 16-byte Folded Reload
1056 ; CHECK-SD-NEXT: mov v1.h[4], v0.h[0]
1057 ; CHECK-SD-NEXT: str q1, [sp, #32] // 16-byte Folded Spill
1058 ; CHECK-SD-NEXT: ldp q0, q1, [sp] // 32-byte Folded Reload
1059 ; CHECK-SD-NEXT: mov h0, v0.h[5]
1060 ; CHECK-SD-NEXT: mov h1, v1.h[5]
1061 ; CHECK-SD-NEXT: fcvt s0, h0
1062 ; CHECK-SD-NEXT: fcvt s1, h1
1063 ; CHECK-SD-NEXT: bl powf
1064 ; CHECK-SD-NEXT: fcvt h0, s0
1065 ; CHECK-SD-NEXT: ldr q1, [sp, #32] // 16-byte Folded Reload
1066 ; CHECK-SD-NEXT: mov v1.h[5], v0.h[0]
1067 ; CHECK-SD-NEXT: str q1, [sp, #32] // 16-byte Folded Spill
1068 ; CHECK-SD-NEXT: ldp q0, q1, [sp] // 32-byte Folded Reload
1069 ; CHECK-SD-NEXT: mov h0, v0.h[6]
1070 ; CHECK-SD-NEXT: mov h1, v1.h[6]
1071 ; CHECK-SD-NEXT: fcvt s0, h0
1072 ; CHECK-SD-NEXT: fcvt s1, h1
1073 ; CHECK-SD-NEXT: bl powf
1074 ; CHECK-SD-NEXT: fcvt h0, s0
1075 ; CHECK-SD-NEXT: ldr q1, [sp, #32] // 16-byte Folded Reload
1076 ; CHECK-SD-NEXT: mov v1.h[6], v0.h[0]
1077 ; CHECK-SD-NEXT: str q1, [sp, #32] // 16-byte Folded Spill
1078 ; CHECK-SD-NEXT: ldp q0, q1, [sp] // 32-byte Folded Reload
1079 ; CHECK-SD-NEXT: mov h0, v0.h[7]
1080 ; CHECK-SD-NEXT: mov h1, v1.h[7]
1081 ; CHECK-SD-NEXT: fcvt s0, h0
1082 ; CHECK-SD-NEXT: fcvt s1, h1
1083 ; CHECK-SD-NEXT: bl powf
1084 ; CHECK-SD-NEXT: fcvt h1, s0
1085 ; CHECK-SD-NEXT: ldr q0, [sp, #32] // 16-byte Folded Reload
1086 ; CHECK-SD-NEXT: ldr x30, [sp, #48] // 8-byte Folded Reload
1087 ; CHECK-SD-NEXT: mov v0.h[7], v1.h[0]
1088 ; CHECK-SD-NEXT: add sp, sp, #64
1089 ; CHECK-SD-NEXT: ret
1090 ;
1091 ; CHECK-GI-LABEL: pow_v8f16:
1092 ; CHECK-GI: // %bb.0: // %entry
1093 ; CHECK-GI-NEXT: sub sp, sp, #192
1094 ; CHECK-GI-NEXT: stp d15, d14, [sp, #112] // 16-byte Folded Spill
1095 ; CHECK-GI-NEXT: stp d13, d12, [sp, #128] // 16-byte Folded Spill
1096 ; CHECK-GI-NEXT: stp d11, d10, [sp, #144] // 16-byte Folded Spill
1097 ; CHECK-GI-NEXT: stp d9, d8, [sp, #160] // 16-byte Folded Spill
1098 ; CHECK-GI-NEXT: str x30, [sp, #176] // 8-byte Folded Spill
1099 ; CHECK-GI-NEXT: .cfi_def_cfa_offset 192
1100 ; CHECK-GI-NEXT: .cfi_offset w30, -16
1101 ; CHECK-GI-NEXT: .cfi_offset b8, -24
1102 ; CHECK-GI-NEXT: .cfi_offset b9, -32
1103 ; CHECK-GI-NEXT: .cfi_offset b10, -40
1104 ; CHECK-GI-NEXT: .cfi_offset b11, -48
1105 ; CHECK-GI-NEXT: .cfi_offset b12, -56
1106 ; CHECK-GI-NEXT: .cfi_offset b13, -64
1107 ; CHECK-GI-NEXT: .cfi_offset b14, -72
1108 ; CHECK-GI-NEXT: .cfi_offset b15, -80
1109 ; CHECK-GI-NEXT: mov h2, v0.h[5]
1110 ; CHECK-GI-NEXT: mov h11, v0.h[1]
1111 ; CHECK-GI-NEXT: mov h12, v0.h[2]
1112 ; CHECK-GI-NEXT: mov h13, v0.h[3]
1113 ; CHECK-GI-NEXT: mov h14, v0.h[4]
1114 ; CHECK-GI-NEXT: mov h8, v1.h[1]
1115 ; CHECK-GI-NEXT: mov h9, v1.h[2]
1116 ; CHECK-GI-NEXT: mov h10, v1.h[3]
1117 ; CHECK-GI-NEXT: mov h15, v1.h[4]
1118 ; CHECK-GI-NEXT: str h2, [sp, #48] // 2-byte Folded Spill
1119 ; CHECK-GI-NEXT: mov h2, v0.h[6]
1120 ; CHECK-GI-NEXT: str h2, [sp, #64] // 2-byte Folded Spill
1121 ; CHECK-GI-NEXT: mov h2, v0.h[7]
1122 ; CHECK-GI-NEXT: fcvt s0, h0
1123 ; CHECK-GI-NEXT: str h2, [sp, #96] // 2-byte Folded Spill
1124 ; CHECK-GI-NEXT: mov h2, v1.h[5]
1125 ; CHECK-GI-NEXT: str h2, [sp, #186] // 2-byte Folded Spill
1126 ; CHECK-GI-NEXT: mov h2, v1.h[6]
1127 ; CHECK-GI-NEXT: str h2, [sp, #188] // 2-byte Folded Spill
1128 ; CHECK-GI-NEXT: mov h2, v1.h[7]
1129 ; CHECK-GI-NEXT: fcvt s1, h1
1130 ; CHECK-GI-NEXT: str h2, [sp, #190] // 2-byte Folded Spill
1131 ; CHECK-GI-NEXT: bl powf
1132 ; CHECK-GI-NEXT: fcvt s2, h11
1133 ; CHECK-GI-NEXT: fcvt h0, s0
1134 ; CHECK-GI-NEXT: fcvt s1, h8
1135 ; CHECK-GI-NEXT: str q0, [sp, #80] // 16-byte Folded Spill
1136 ; CHECK-GI-NEXT: fmov s0, s2
1137 ; CHECK-GI-NEXT: bl powf
1138 ; CHECK-GI-NEXT: fcvt s2, h12
1139 ; CHECK-GI-NEXT: fcvt h0, s0
1140 ; CHECK-GI-NEXT: fcvt s1, h9
1141 ; CHECK-GI-NEXT: str q0, [sp, #32] // 16-byte Folded Spill
1142 ; CHECK-GI-NEXT: fmov s0, s2
1143 ; CHECK-GI-NEXT: bl powf
1144 ; CHECK-GI-NEXT: fcvt s2, h13
1145 ; CHECK-GI-NEXT: fcvt h0, s0
1146 ; CHECK-GI-NEXT: fcvt s1, h10
1147 ; CHECK-GI-NEXT: str q0, [sp, #16] // 16-byte Folded Spill
1148 ; CHECK-GI-NEXT: fmov s0, s2
1149 ; CHECK-GI-NEXT: bl powf
1150 ; CHECK-GI-NEXT: fcvt s2, h14
1151 ; CHECK-GI-NEXT: fcvt h0, s0
1152 ; CHECK-GI-NEXT: fcvt s1, h15
1153 ; CHECK-GI-NEXT: str q0, [sp] // 16-byte Folded Spill
1154 ; CHECK-GI-NEXT: fmov s0, s2
1155 ; CHECK-GI-NEXT: bl powf
1156 ; CHECK-GI-NEXT: ldr h1, [sp, #48] // 2-byte Folded Reload
1157 ; CHECK-GI-NEXT: fcvt h0, s0
1158 ; CHECK-GI-NEXT: fcvt s2, h1
1159 ; CHECK-GI-NEXT: str q0, [sp, #48] // 16-byte Folded Spill
1160 ; CHECK-GI-NEXT: ldr h0, [sp, #186] // 2-byte Folded Reload
1161 ; CHECK-GI-NEXT: fcvt s1, h0
1162 ; CHECK-GI-NEXT: fmov s0, s2
1163 ; CHECK-GI-NEXT: bl powf
1164 ; CHECK-GI-NEXT: ldr h1, [sp, #64] // 2-byte Folded Reload
1165 ; CHECK-GI-NEXT: fcvt h0, s0
1166 ; CHECK-GI-NEXT: fcvt s2, h1
1167 ; CHECK-GI-NEXT: str q0, [sp, #64] // 16-byte Folded Spill
1168 ; CHECK-GI-NEXT: ldr h0, [sp, #188] // 2-byte Folded Reload
1169 ; CHECK-GI-NEXT: fcvt s1, h0
1170 ; CHECK-GI-NEXT: fmov s0, s2
1171 ; CHECK-GI-NEXT: bl powf
1172 ; CHECK-GI-NEXT: ldr h1, [sp, #96] // 2-byte Folded Reload
1173 ; CHECK-GI-NEXT: fcvt h0, s0
1174 ; CHECK-GI-NEXT: fcvt s2, h1
1175 ; CHECK-GI-NEXT: str q0, [sp, #96] // 16-byte Folded Spill
1176 ; CHECK-GI-NEXT: ldr h0, [sp, #190] // 2-byte Folded Reload
1177 ; CHECK-GI-NEXT: fcvt s1, h0
1178 ; CHECK-GI-NEXT: fmov s0, s2
1179 ; CHECK-GI-NEXT: bl powf
1180 ; CHECK-GI-NEXT: ldp q2, q3, [sp, #16] // 32-byte Folded Reload
1181 ; CHECK-GI-NEXT: fcvt h0, s0
1182 ; CHECK-GI-NEXT: ldr q1, [sp, #80] // 16-byte Folded Reload
1183 ; CHECK-GI-NEXT: ldp d9, d8, [sp, #160] // 16-byte Folded Reload
1184 ; CHECK-GI-NEXT: ldp d11, d10, [sp, #144] // 16-byte Folded Reload
1185 ; CHECK-GI-NEXT: ldr x30, [sp, #176] // 8-byte Folded Reload
1186 ; CHECK-GI-NEXT: mov v1.h[1], v3.h[0]
1187 ; CHECK-GI-NEXT: ldp d13, d12, [sp, #128] // 16-byte Folded Reload
1188 ; CHECK-GI-NEXT: ldp d15, d14, [sp, #112] // 16-byte Folded Reload
1189 ; CHECK-GI-NEXT: mov v1.h[2], v2.h[0]
1190 ; CHECK-GI-NEXT: ldr q2, [sp] // 16-byte Folded Reload
1191 ; CHECK-GI-NEXT: mov v1.h[3], v2.h[0]
1192 ; CHECK-GI-NEXT: ldp q3, q2, [sp, #48] // 32-byte Folded Reload
1193 ; CHECK-GI-NEXT: mov v1.h[4], v3.h[0]
1194 ; CHECK-GI-NEXT: mov v1.h[5], v2.h[0]
1195 ; CHECK-GI-NEXT: ldr q2, [sp, #96] // 16-byte Folded Reload
1196 ; CHECK-GI-NEXT: mov v1.h[6], v2.h[0]
1197 ; CHECK-GI-NEXT: mov v1.h[7], v0.h[0]
1198 ; CHECK-GI-NEXT: mov v0.16b, v1.16b
1199 ; CHECK-GI-NEXT: add sp, sp, #192
1200 ; CHECK-GI-NEXT: ret
1201 entry:
1202 %c = call <8 x half> @llvm.pow.v8f16(<8 x half> %a, <8 x half> %b)
1203 ret <8 x half> %c
1204 }
1206 define <16 x half> @pow_v16f16(<16 x half> %a, <16 x half> %b) {
1207 ; CHECK-SD-LABEL: pow_v16f16:
1208 ; CHECK-SD: // %bb.0: // %entry
1209 ; CHECK-SD-NEXT: sub sp, sp, #96
1210 ; CHECK-SD-NEXT: str x30, [sp, #80] // 8-byte Folded Spill
1211 ; CHECK-SD-NEXT: .cfi_def_cfa_offset 96
1212 ; CHECK-SD-NEXT: .cfi_offset w30, -16
1213 ; CHECK-SD-NEXT: stp q0, q1, [sp] // 32-byte Folded Spill
1214 ; CHECK-SD-NEXT: mov h1, v2.h[1]
1215 ; CHECK-SD-NEXT: stp q3, q2, [sp, #32] // 32-byte Folded Spill
1216 ; CHECK-SD-NEXT: mov h3, v0.h[1]
1217 ; CHECK-SD-NEXT: fcvt s1, h1
1218 ; CHECK-SD-NEXT: fcvt s0, h3
1219 ; CHECK-SD-NEXT: bl powf
1220 ; CHECK-SD-NEXT: fcvt h0, s0
1221 ; CHECK-SD-NEXT: ldr q1, [sp, #48] // 16-byte Folded Reload
1222 ; CHECK-SD-NEXT: fcvt s1, h1
1223 ; CHECK-SD-NEXT: str q0, [sp, #64] // 16-byte Folded Spill
1224 ; CHECK-SD-NEXT: ldr q0, [sp] // 16-byte Folded Reload
1225 ; CHECK-SD-NEXT: fcvt s0, h0
1226 ; CHECK-SD-NEXT: bl powf
1227 ; CHECK-SD-NEXT: fcvt h0, s0
1228 ; CHECK-SD-NEXT: ldp q1, q2, [sp, #48] // 32-byte Folded Reload
1229 ; CHECK-SD-NEXT: mov h1, v1.h[2]
1230 ; CHECK-SD-NEXT: mov v0.h[1], v2.h[0]
1231 ; CHECK-SD-NEXT: fcvt s1, h1
1232 ; CHECK-SD-NEXT: str q0, [sp, #64] // 16-byte Folded Spill
1233 ; CHECK-SD-NEXT: ldr q0, [sp] // 16-byte Folded Reload
1234 ; CHECK-SD-NEXT: mov h0, v0.h[2]
1235 ; CHECK-SD-NEXT: fcvt s0, h0
1236 ; CHECK-SD-NEXT: bl powf
1237 ; CHECK-SD-NEXT: fcvt h0, s0
1238 ; CHECK-SD-NEXT: ldr q1, [sp, #64] // 16-byte Folded Reload
1239 ; CHECK-SD-NEXT: mov v1.h[2], v0.h[0]
1240 ; CHECK-SD-NEXT: ldr q0, [sp] // 16-byte Folded Reload
1241 ; CHECK-SD-NEXT: mov h0, v0.h[3]
1242 ; CHECK-SD-NEXT: str q1, [sp, #64] // 16-byte Folded Spill
1243 ; CHECK-SD-NEXT: ldr q1, [sp, #48] // 16-byte Folded Reload
1244 ; CHECK-SD-NEXT: mov h1, v1.h[3]
1245 ; CHECK-SD-NEXT: fcvt s0, h0
1246 ; CHECK-SD-NEXT: fcvt s1, h1
1247 ; CHECK-SD-NEXT: bl powf
1248 ; CHECK-SD-NEXT: fcvt h0, s0
1249 ; CHECK-SD-NEXT: ldr q1, [sp, #64] // 16-byte Folded Reload
1250 ; CHECK-SD-NEXT: mov v1.h[3], v0.h[0]
1251 ; CHECK-SD-NEXT: ldr q0, [sp] // 16-byte Folded Reload
1252 ; CHECK-SD-NEXT: mov h0, v0.h[4]
1253 ; CHECK-SD-NEXT: str q1, [sp, #64] // 16-byte Folded Spill
1254 ; CHECK-SD-NEXT: ldr q1, [sp, #48] // 16-byte Folded Reload
1255 ; CHECK-SD-NEXT: mov h1, v1.h[4]
1256 ; CHECK-SD-NEXT: fcvt s0, h0
1257 ; CHECK-SD-NEXT: fcvt s1, h1
1258 ; CHECK-SD-NEXT: bl powf
1259 ; CHECK-SD-NEXT: fcvt h0, s0
1260 ; CHECK-SD-NEXT: ldr q1, [sp, #64] // 16-byte Folded Reload
1261 ; CHECK-SD-NEXT: mov v1.h[4], v0.h[0]
1262 ; CHECK-SD-NEXT: ldr q0, [sp] // 16-byte Folded Reload
1263 ; CHECK-SD-NEXT: mov h0, v0.h[5]
1264 ; CHECK-SD-NEXT: str q1, [sp, #64] // 16-byte Folded Spill
1265 ; CHECK-SD-NEXT: ldr q1, [sp, #48] // 16-byte Folded Reload
1266 ; CHECK-SD-NEXT: mov h1, v1.h[5]
1267 ; CHECK-SD-NEXT: fcvt s0, h0
1268 ; CHECK-SD-NEXT: fcvt s1, h1
1269 ; CHECK-SD-NEXT: bl powf
1270 ; CHECK-SD-NEXT: fcvt h0, s0
1271 ; CHECK-SD-NEXT: ldr q1, [sp, #64] // 16-byte Folded Reload
1272 ; CHECK-SD-NEXT: mov v1.h[5], v0.h[0]
1273 ; CHECK-SD-NEXT: ldr q0, [sp] // 16-byte Folded Reload
1274 ; CHECK-SD-NEXT: mov h0, v0.h[6]
1275 ; CHECK-SD-NEXT: str q1, [sp, #64] // 16-byte Folded Spill
1276 ; CHECK-SD-NEXT: ldr q1, [sp, #48] // 16-byte Folded Reload
1277 ; CHECK-SD-NEXT: mov h1, v1.h[6]
1278 ; CHECK-SD-NEXT: fcvt s0, h0
1279 ; CHECK-SD-NEXT: fcvt s1, h1
1280 ; CHECK-SD-NEXT: bl powf
1281 ; CHECK-SD-NEXT: fcvt h0, s0
1282 ; CHECK-SD-NEXT: ldr q1, [sp, #64] // 16-byte Folded Reload
1283 ; CHECK-SD-NEXT: mov v1.h[6], v0.h[0]
1284 ; CHECK-SD-NEXT: ldr q0, [sp] // 16-byte Folded Reload
1285 ; CHECK-SD-NEXT: mov h0, v0.h[7]
1286 ; CHECK-SD-NEXT: str q1, [sp, #64] // 16-byte Folded Spill
1287 ; CHECK-SD-NEXT: ldr q1, [sp, #48] // 16-byte Folded Reload
1288 ; CHECK-SD-NEXT: mov h1, v1.h[7]
1289 ; CHECK-SD-NEXT: fcvt s0, h0
1290 ; CHECK-SD-NEXT: fcvt s1, h1
1291 ; CHECK-SD-NEXT: bl powf
1292 ; CHECK-SD-NEXT: fcvt h0, s0
1293 ; CHECK-SD-NEXT: ldr q1, [sp, #64] // 16-byte Folded Reload
1294 ; CHECK-SD-NEXT: mov v1.h[7], v0.h[0]
1295 ; CHECK-SD-NEXT: str q1, [sp, #64] // 16-byte Folded Spill
1296 ; CHECK-SD-NEXT: ldp q0, q1, [sp, #16] // 32-byte Folded Reload
1297 ; CHECK-SD-NEXT: mov h0, v0.h[1]
1298 ; CHECK-SD-NEXT: mov h1, v1.h[1]
1299 ; CHECK-SD-NEXT: fcvt s0, h0
1300 ; CHECK-SD-NEXT: fcvt s1, h1
1301 ; CHECK-SD-NEXT: bl powf
1302 ; CHECK-SD-NEXT: fcvt h0, s0
1303 ; CHECK-SD-NEXT: str q0, [sp, #48] // 16-byte Folded Spill
1304 ; CHECK-SD-NEXT: ldp q0, q1, [sp, #16] // 32-byte Folded Reload
1305 ; CHECK-SD-NEXT: fcvt s0, h0
1306 ; CHECK-SD-NEXT: fcvt s1, h1
1307 ; CHECK-SD-NEXT: bl powf
1308 ; CHECK-SD-NEXT: fcvt h0, s0
1309 ; CHECK-SD-NEXT: ldp q1, q2, [sp, #32] // 32-byte Folded Reload
1310 ; CHECK-SD-NEXT: mov h1, v1.h[2]
1311 ; CHECK-SD-NEXT: mov v0.h[1], v2.h[0]
1312 ; CHECK-SD-NEXT: fcvt s1, h1
1313 ; CHECK-SD-NEXT: str q0, [sp, #48] // 16-byte Folded Spill
1314 ; CHECK-SD-NEXT: ldr q0, [sp, #16] // 16-byte Folded Reload
1315 ; CHECK-SD-NEXT: mov h0, v0.h[2]
1316 ; CHECK-SD-NEXT: fcvt s0, h0
1317 ; CHECK-SD-NEXT: bl powf
1318 ; CHECK-SD-NEXT: fcvt h0, s0
1319 ; CHECK-SD-NEXT: ldr q1, [sp, #48] // 16-byte Folded Reload
1320 ; CHECK-SD-NEXT: mov v1.h[2], v0.h[0]
1321 ; CHECK-SD-NEXT: str q1, [sp, #48] // 16-byte Folded Spill
1322 ; CHECK-SD-NEXT: ldp q0, q1, [sp, #16] // 32-byte Folded Reload
1323 ; CHECK-SD-NEXT: mov h0, v0.h[3]
1324 ; CHECK-SD-NEXT: mov h1, v1.h[3]
1325 ; CHECK-SD-NEXT: fcvt s0, h0
1326 ; CHECK-SD-NEXT: fcvt s1, h1
1327 ; CHECK-SD-NEXT: bl powf
1328 ; CHECK-SD-NEXT: fcvt h0, s0
1329 ; CHECK-SD-NEXT: ldr q1, [sp, #48] // 16-byte Folded Reload
1330 ; CHECK-SD-NEXT: mov v1.h[3], v0.h[0]
1331 ; CHECK-SD-NEXT: str q1, [sp, #48] // 16-byte Folded Spill
1332 ; CHECK-SD-NEXT: ldp q0, q1, [sp, #16] // 32-byte Folded Reload
1333 ; CHECK-SD-NEXT: mov h0, v0.h[4]
1334 ; CHECK-SD-NEXT: mov h1, v1.h[4]
1335 ; CHECK-SD-NEXT: fcvt s0, h0
1336 ; CHECK-SD-NEXT: fcvt s1, h1
1337 ; CHECK-SD-NEXT: bl powf
1338 ; CHECK-SD-NEXT: fcvt h0, s0
1339 ; CHECK-SD-NEXT: ldr q1, [sp, #48] // 16-byte Folded Reload
1340 ; CHECK-SD-NEXT: mov v1.h[4], v0.h[0]
1341 ; CHECK-SD-NEXT: str q1, [sp, #48] // 16-byte Folded Spill
1342 ; CHECK-SD-NEXT: ldp q0, q1, [sp, #16] // 32-byte Folded Reload
1343 ; CHECK-SD-NEXT: mov h0, v0.h[5]
1344 ; CHECK-SD-NEXT: mov h1, v1.h[5]
1345 ; CHECK-SD-NEXT: fcvt s0, h0
1346 ; CHECK-SD-NEXT: fcvt s1, h1
1347 ; CHECK-SD-NEXT: bl powf
1348 ; CHECK-SD-NEXT: fcvt h0, s0
1349 ; CHECK-SD-NEXT: ldr q1, [sp, #48] // 16-byte Folded Reload
1350 ; CHECK-SD-NEXT: mov v1.h[5], v0.h[0]
1351 ; CHECK-SD-NEXT: str q1, [sp, #48] // 16-byte Folded Spill
1352 ; CHECK-SD-NEXT: ldp q0, q1, [sp, #16] // 32-byte Folded Reload
1353 ; CHECK-SD-NEXT: mov h0, v0.h[6]
1354 ; CHECK-SD-NEXT: mov h1, v1.h[6]
1355 ; CHECK-SD-NEXT: fcvt s0, h0
1356 ; CHECK-SD-NEXT: fcvt s1, h1
1357 ; CHECK-SD-NEXT: bl powf
1358 ; CHECK-SD-NEXT: fcvt h0, s0
1359 ; CHECK-SD-NEXT: ldr q1, [sp, #48] // 16-byte Folded Reload
1360 ; CHECK-SD-NEXT: mov v1.h[6], v0.h[0]
1361 ; CHECK-SD-NEXT: str q1, [sp, #48] // 16-byte Folded Spill
1362 ; CHECK-SD-NEXT: ldp q0, q1, [sp, #16] // 32-byte Folded Reload
1363 ; CHECK-SD-NEXT: mov h0, v0.h[7]
1364 ; CHECK-SD-NEXT: mov h1, v1.h[7]
1365 ; CHECK-SD-NEXT: fcvt s0, h0
1366 ; CHECK-SD-NEXT: fcvt s1, h1
1367 ; CHECK-SD-NEXT: bl powf
1368 ; CHECK-SD-NEXT: fmov s1, s0
1369 ; CHECK-SD-NEXT: ldr x30, [sp, #80] // 8-byte Folded Reload
1370 ; CHECK-SD-NEXT: fcvt h2, s1
1371 ; CHECK-SD-NEXT: ldp q1, q0, [sp, #48] // 32-byte Folded Reload
1372 ; CHECK-SD-NEXT: mov v1.h[7], v2.h[0]
1373 ; CHECK-SD-NEXT: add sp, sp, #96
1374 ; CHECK-SD-NEXT: ret
1375 ;
1376 ; CHECK-GI-LABEL: pow_v16f16:
1377 ; CHECK-GI: // %bb.0: // %entry
1378 ; CHECK-GI-NEXT: sub sp, sp, #448
1379 ; CHECK-GI-NEXT: stp d15, d14, [sp, #368] // 16-byte Folded Spill
1380 ; CHECK-GI-NEXT: stp d13, d12, [sp, #384] // 16-byte Folded Spill
1381 ; CHECK-GI-NEXT: stp d11, d10, [sp, #400] // 16-byte Folded Spill
1382 ; CHECK-GI-NEXT: stp d9, d8, [sp, #416] // 16-byte Folded Spill
1383 ; CHECK-GI-NEXT: stp x29, x30, [sp, #432] // 16-byte Folded Spill
1384 ; CHECK-GI-NEXT: .cfi_def_cfa_offset 448
1385 ; CHECK-GI-NEXT: .cfi_offset w30, -8
1386 ; CHECK-GI-NEXT: .cfi_offset w29, -16
1387 ; CHECK-GI-NEXT: .cfi_offset b8, -24
1388 ; CHECK-GI-NEXT: .cfi_offset b9, -32
1389 ; CHECK-GI-NEXT: .cfi_offset b10, -40
1390 ; CHECK-GI-NEXT: .cfi_offset b11, -48
1391 ; CHECK-GI-NEXT: .cfi_offset b12, -56
1392 ; CHECK-GI-NEXT: .cfi_offset b13, -64
1393 ; CHECK-GI-NEXT: .cfi_offset b14, -72
1394 ; CHECK-GI-NEXT: .cfi_offset b15, -80
1395 ; CHECK-GI-NEXT: mov v4.16b, v1.16b
1396 ; CHECK-GI-NEXT: str q1, [sp, #112] // 16-byte Folded Spill
1397 ; CHECK-GI-NEXT: mov h1, v0.h[4]
1398 ; CHECK-GI-NEXT: mov h12, v0.h[1]
1399 ; CHECK-GI-NEXT: mov h13, v0.h[2]
1400 ; CHECK-GI-NEXT: str q3, [sp, #16] // 16-byte Folded Spill
1401 ; CHECK-GI-NEXT: mov h14, v0.h[3]
1402 ; CHECK-GI-NEXT: mov h15, v2.h[1]
1403 ; CHECK-GI-NEXT: mov h8, v2.h[2]
1404 ; CHECK-GI-NEXT: mov h9, v2.h[3]
1405 ; CHECK-GI-NEXT: mov h10, v2.h[4]
1406 ; CHECK-GI-NEXT: mov h11, v2.h[5]
1407 ; CHECK-GI-NEXT: str h1, [sp, #272] // 2-byte Folded Spill
1408 ; CHECK-GI-NEXT: mov h1, v0.h[5]
1409 ; CHECK-GI-NEXT: str h1, [sp, #240] // 2-byte Folded Spill
1410 ; CHECK-GI-NEXT: mov h1, v0.h[6]
1411 ; CHECK-GI-NEXT: str h1, [sp, #176] // 2-byte Folded Spill
1412 ; CHECK-GI-NEXT: mov h1, v0.h[7]
1413 ; CHECK-GI-NEXT: fcvt s0, h0
1414 ; CHECK-GI-NEXT: str h1, [sp, #144] // 2-byte Folded Spill
1415 ; CHECK-GI-NEXT: mov h1, v4.h[1]
1416 ; CHECK-GI-NEXT: str h1, [sp, #48] // 2-byte Folded Spill
1417 ; CHECK-GI-NEXT: mov h1, v4.h[2]
1418 ; CHECK-GI-NEXT: str h1, [sp, #80] // 2-byte Folded Spill
1419 ; CHECK-GI-NEXT: mov h1, v4.h[3]
1420 ; CHECK-GI-NEXT: str h1, [sp, #128] // 2-byte Folded Spill
1421 ; CHECK-GI-NEXT: mov h1, v4.h[4]
1422 ; CHECK-GI-NEXT: str h1, [sp, #192] // 2-byte Folded Spill
1423 ; CHECK-GI-NEXT: mov h1, v4.h[5]
1424 ; CHECK-GI-NEXT: str h1, [sp, #256] // 2-byte Folded Spill
1425 ; CHECK-GI-NEXT: mov h1, v4.h[6]
1426 ; CHECK-GI-NEXT: str h1, [sp, #336] // 2-byte Folded Spill
1427 ; CHECK-GI-NEXT: mov h1, v4.h[7]
1428 ; CHECK-GI-NEXT: str h1, [sp, #352] // 2-byte Folded Spill
1429 ; CHECK-GI-NEXT: mov h1, v2.h[6]
1430 ; CHECK-GI-NEXT: str h1, [sp, #12] // 2-byte Folded Spill
1431 ; CHECK-GI-NEXT: mov h1, v2.h[7]
1432 ; CHECK-GI-NEXT: str h1, [sp, #14] // 2-byte Folded Spill
1433 ; CHECK-GI-NEXT: mov h1, v3.h[1]
1434 ; CHECK-GI-NEXT: str h1, [sp, #44] // 2-byte Folded Spill
1435 ; CHECK-GI-NEXT: mov h1, v3.h[2]
1436 ; CHECK-GI-NEXT: str h1, [sp, #46] // 2-byte Folded Spill
1437 ; CHECK-GI-NEXT: mov h1, v3.h[3]
1438 ; CHECK-GI-NEXT: str h1, [sp, #78] // 2-byte Folded Spill
1439 ; CHECK-GI-NEXT: mov h1, v3.h[4]
1440 ; CHECK-GI-NEXT: str h1, [sp, #110] // 2-byte Folded Spill
1441 ; CHECK-GI-NEXT: mov h1, v3.h[5]
1442 ; CHECK-GI-NEXT: str h1, [sp, #174] // 2-byte Folded Spill
1443 ; CHECK-GI-NEXT: mov h1, v3.h[6]
1444 ; CHECK-GI-NEXT: str h1, [sp, #238] // 2-byte Folded Spill
1445 ; CHECK-GI-NEXT: mov h1, v3.h[7]
1446 ; CHECK-GI-NEXT: str h1, [sp, #302] // 2-byte Folded Spill
1447 ; CHECK-GI-NEXT: fcvt s1, h2
1448 ; CHECK-GI-NEXT: bl powf
1449 ; CHECK-GI-NEXT: fcvt s2, h12
1450 ; CHECK-GI-NEXT: fcvt h0, s0
1451 ; CHECK-GI-NEXT: fcvt s1, h15
1452 ; CHECK-GI-NEXT: str q0, [sp, #304] // 16-byte Folded Spill
1453 ; CHECK-GI-NEXT: fmov s0, s2
1454 ; CHECK-GI-NEXT: bl powf
1455 ; CHECK-GI-NEXT: fcvt s2, h13
1456 ; CHECK-GI-NEXT: fcvt h0, s0
1457 ; CHECK-GI-NEXT: fcvt s1, h8
1458 ; CHECK-GI-NEXT: str q0, [sp, #208] // 16-byte Folded Spill
1459 ; CHECK-GI-NEXT: fmov s0, s2
1460 ; CHECK-GI-NEXT: bl powf
1461 ; CHECK-GI-NEXT: fcvt s2, h14
1462 ; CHECK-GI-NEXT: fcvt h0, s0
1463 ; CHECK-GI-NEXT: fcvt s1, h9
1464 ; CHECK-GI-NEXT: str q0, [sp, #320] // 16-byte Folded Spill
1465 ; CHECK-GI-NEXT: fmov s0, s2
1466 ; CHECK-GI-NEXT: bl powf
1467 ; CHECK-GI-NEXT: ldr h1, [sp, #272] // 2-byte Folded Reload
1468 ; CHECK-GI-NEXT: fcvt h0, s0
1469 ; CHECK-GI-NEXT: fcvt s2, h1
1470 ; CHECK-GI-NEXT: fcvt s1, h10
1471 ; CHECK-GI-NEXT: str q0, [sp, #272] // 16-byte Folded Spill
1472 ; CHECK-GI-NEXT: fmov s0, s2
1473 ; CHECK-GI-NEXT: bl powf
1474 ; CHECK-GI-NEXT: ldr h1, [sp, #240] // 2-byte Folded Reload
1475 ; CHECK-GI-NEXT: fcvt h0, s0
1476 ; CHECK-GI-NEXT: fcvt s2, h1
1477 ; CHECK-GI-NEXT: fcvt s1, h11
1478 ; CHECK-GI-NEXT: str q0, [sp, #240] // 16-byte Folded Spill
1479 ; CHECK-GI-NEXT: fmov s0, s2
1480 ; CHECK-GI-NEXT: bl powf
1481 ; CHECK-GI-NEXT: ldr h1, [sp, #176] // 2-byte Folded Reload
1482 ; CHECK-GI-NEXT: fcvt h0, s0
1483 ; CHECK-GI-NEXT: fcvt s2, h1
1484 ; CHECK-GI-NEXT: str q0, [sp, #176] // 16-byte Folded Spill
1485 ; CHECK-GI-NEXT: ldr h0, [sp, #12] // 2-byte Folded Reload
1486 ; CHECK-GI-NEXT: fcvt s1, h0
1487 ; CHECK-GI-NEXT: fmov s0, s2
1488 ; CHECK-GI-NEXT: bl powf
1489 ; CHECK-GI-NEXT: ldr h1, [sp, #144] // 2-byte Folded Reload
1490 ; CHECK-GI-NEXT: fcvt h0, s0
1491 ; CHECK-GI-NEXT: fcvt s2, h1
1492 ; CHECK-GI-NEXT: str q0, [sp, #144] // 16-byte Folded Spill
1493 ; CHECK-GI-NEXT: ldr h0, [sp, #14] // 2-byte Folded Reload
1494 ; CHECK-GI-NEXT: fcvt s1, h0
1495 ; CHECK-GI-NEXT: fmov s0, s2
1496 ; CHECK-GI-NEXT: bl powf
1497 ; CHECK-GI-NEXT: ldr q1, [sp, #112] // 16-byte Folded Reload
1498 ; CHECK-GI-NEXT: fcvt h0, s0
1499 ; CHECK-GI-NEXT: fcvt s2, h1
1500 ; CHECK-GI-NEXT: str q0, [sp, #112] // 16-byte Folded Spill
1501 ; CHECK-GI-NEXT: ldr q0, [sp, #16] // 16-byte Folded Reload
1502 ; CHECK-GI-NEXT: fcvt s1, h0
1503 ; CHECK-GI-NEXT: fmov s0, s2
1504 ; CHECK-GI-NEXT: bl powf
1505 ; CHECK-GI-NEXT: ldr h1, [sp, #48] // 2-byte Folded Reload
1506 ; CHECK-GI-NEXT: fcvt h0, s0
1507 ; CHECK-GI-NEXT: fcvt s2, h1
1508 ; CHECK-GI-NEXT: str q0, [sp, #48] // 16-byte Folded Spill
1509 ; CHECK-GI-NEXT: ldr h0, [sp, #44] // 2-byte Folded Reload
1510 ; CHECK-GI-NEXT: fcvt s1, h0
1511 ; CHECK-GI-NEXT: fmov s0, s2
1512 ; CHECK-GI-NEXT: bl powf
1513 ; CHECK-GI-NEXT: ldr h1, [sp, #80] // 2-byte Folded Reload
1514 ; CHECK-GI-NEXT: fcvt h0, s0
1515 ; CHECK-GI-NEXT: fcvt s2, h1
1516 ; CHECK-GI-NEXT: str q0, [sp, #80] // 16-byte Folded Spill
1517 ; CHECK-GI-NEXT: ldr h0, [sp, #46] // 2-byte Folded Reload
1518 ; CHECK-GI-NEXT: fcvt s1, h0
1519 ; CHECK-GI-NEXT: fmov s0, s2
1520 ; CHECK-GI-NEXT: bl powf
1521 ; CHECK-GI-NEXT: ldr h1, [sp, #128] // 2-byte Folded Reload
1522 ; CHECK-GI-NEXT: fcvt h0, s0
1523 ; CHECK-GI-NEXT: fcvt s2, h1
1524 ; CHECK-GI-NEXT: str q0, [sp, #128] // 16-byte Folded Spill
1525 ; CHECK-GI-NEXT: ldr h0, [sp, #78] // 2-byte Folded Reload
1526 ; CHECK-GI-NEXT: fcvt s1, h0
1527 ; CHECK-GI-NEXT: fmov s0, s2
1528 ; CHECK-GI-NEXT: bl powf
1529 ; CHECK-GI-NEXT: ldr h1, [sp, #192] // 2-byte Folded Reload
1530 ; CHECK-GI-NEXT: fcvt h0, s0
1531 ; CHECK-GI-NEXT: fcvt s2, h1
1532 ; CHECK-GI-NEXT: str q0, [sp, #192] // 16-byte Folded Spill
1533 ; CHECK-GI-NEXT: ldr h0, [sp, #110] // 2-byte Folded Reload
1534 ; CHECK-GI-NEXT: fcvt s1, h0
1535 ; CHECK-GI-NEXT: fmov s0, s2
1536 ; CHECK-GI-NEXT: bl powf
1537 ; CHECK-GI-NEXT: ldr h1, [sp, #256] // 2-byte Folded Reload
1538 ; CHECK-GI-NEXT: fcvt h0, s0
1539 ; CHECK-GI-NEXT: fcvt s2, h1
1540 ; CHECK-GI-NEXT: str q0, [sp, #256] // 16-byte Folded Spill
1541 ; CHECK-GI-NEXT: ldr h0, [sp, #174] // 2-byte Folded Reload
1542 ; CHECK-GI-NEXT: fcvt s1, h0
1543 ; CHECK-GI-NEXT: fmov s0, s2
1544 ; CHECK-GI-NEXT: bl powf
1545 ; CHECK-GI-NEXT: ldr h1, [sp, #336] // 2-byte Folded Reload
1546 ; CHECK-GI-NEXT: fcvt h0, s0
1547 ; CHECK-GI-NEXT: fcvt s2, h1
1548 ; CHECK-GI-NEXT: str q0, [sp, #336] // 16-byte Folded Spill
1549 ; CHECK-GI-NEXT: ldr h0, [sp, #238] // 2-byte Folded Reload
1550 ; CHECK-GI-NEXT: fcvt s1, h0
1551 ; CHECK-GI-NEXT: fmov s0, s2
1552 ; CHECK-GI-NEXT: bl powf
1553 ; CHECK-GI-NEXT: ldr h1, [sp, #352] // 2-byte Folded Reload
1554 ; CHECK-GI-NEXT: fcvt h0, s0
1555 ; CHECK-GI-NEXT: fcvt s2, h1
1556 ; CHECK-GI-NEXT: str q0, [sp, #352] // 16-byte Folded Spill
1557 ; CHECK-GI-NEXT: ldr h0, [sp, #302] // 2-byte Folded Reload
1558 ; CHECK-GI-NEXT: fcvt s1, h0
1559 ; CHECK-GI-NEXT: fmov s0, s2
1560 ; CHECK-GI-NEXT: bl powf
1561 ; CHECK-GI-NEXT: ldr q3, [sp, #304] // 16-byte Folded Reload
1562 ; CHECK-GI-NEXT: ldr q2, [sp, #208] // 16-byte Folded Reload
1563 ; CHECK-GI-NEXT: ldr q1, [sp, #48] // 16-byte Folded Reload
1564 ; CHECK-GI-NEXT: ldp x29, x30, [sp, #432] // 16-byte Folded Reload
1565 ; CHECK-GI-NEXT: mov v3.h[1], v2.h[0]
1566 ; CHECK-GI-NEXT: ldr q2, [sp, #80] // 16-byte Folded Reload
1567 ; CHECK-GI-NEXT: ldp d9, d8, [sp, #416] // 16-byte Folded Reload
1568 ; CHECK-GI-NEXT: mov v1.h[1], v2.h[0]
1569 ; CHECK-GI-NEXT: ldr q2, [sp, #320] // 16-byte Folded Reload
1570 ; CHECK-GI-NEXT: ldp d11, d10, [sp, #400] // 16-byte Folded Reload
1571 ; CHECK-GI-NEXT: mov v3.h[2], v2.h[0]
1572 ; CHECK-GI-NEXT: ldr q2, [sp, #128] // 16-byte Folded Reload
1573 ; CHECK-GI-NEXT: ldp d13, d12, [sp, #384] // 16-byte Folded Reload
1574 ; CHECK-GI-NEXT: mov v1.h[2], v2.h[0]
1575 ; CHECK-GI-NEXT: ldr q2, [sp, #272] // 16-byte Folded Reload
1576 ; CHECK-GI-NEXT: ldp d15, d14, [sp, #368] // 16-byte Folded Reload
1577 ; CHECK-GI-NEXT: mov v3.h[3], v2.h[0]
1578 ; CHECK-GI-NEXT: ldr q2, [sp, #192] // 16-byte Folded Reload
1579 ; CHECK-GI-NEXT: mov v1.h[3], v2.h[0]
1580 ; CHECK-GI-NEXT: ldp q4, q2, [sp, #240] // 32-byte Folded Reload
1581 ; CHECK-GI-NEXT: mov v3.h[4], v4.h[0]
1582 ; CHECK-GI-NEXT: mov v1.h[4], v2.h[0]
1583 ; CHECK-GI-NEXT: ldr q2, [sp, #176] // 16-byte Folded Reload
1584 ; CHECK-GI-NEXT: mov v3.h[5], v2.h[0]
1585 ; CHECK-GI-NEXT: ldr q2, [sp, #336] // 16-byte Folded Reload
1586 ; CHECK-GI-NEXT: mov v1.h[5], v2.h[0]
1587 ; CHECK-GI-NEXT: fcvt h2, s0
1588 ; CHECK-GI-NEXT: ldr q0, [sp, #144] // 16-byte Folded Reload
1589 ; CHECK-GI-NEXT: mov v3.h[6], v0.h[0]
1590 ; CHECK-GI-NEXT: ldr q0, [sp, #352] // 16-byte Folded Reload
1591 ; CHECK-GI-NEXT: mov v1.h[6], v0.h[0]
1592 ; CHECK-GI-NEXT: ldr q0, [sp, #112] // 16-byte Folded Reload
1593 ; CHECK-GI-NEXT: mov v3.h[7], v0.h[0]
1594 ; CHECK-GI-NEXT: mov v1.h[7], v2.h[0]
1595 ; CHECK-GI-NEXT: mov v0.16b, v3.16b
1596 ; CHECK-GI-NEXT: add sp, sp, #448
1597 ; CHECK-GI-NEXT: ret
1598 entry:
1599 %c = call <16 x half> @llvm.pow.v16f16(<16 x half> %a, <16 x half> %b)
1600 ret <16 x half> %c
1601 }
1603 define <2 x fp128> @pow_v2fp128(<2 x fp128> %a, <2 x fp128> %b) {
1604 ; CHECK-SD-LABEL: pow_v2fp128:
1605 ; CHECK-SD: // %bb.0: // %entry
1606 ; CHECK-SD-NEXT: sub sp, sp, #64
1607 ; CHECK-SD-NEXT: str x30, [sp, #48] // 8-byte Folded Spill
1608 ; CHECK-SD-NEXT: .cfi_def_cfa_offset 64
1609 ; CHECK-SD-NEXT: .cfi_offset w30, -16
1610 ; CHECK-SD-NEXT: stp q1, q3, [sp, #16] // 32-byte Folded Spill
1611 ; CHECK-SD-NEXT: mov v1.16b, v2.16b
1612 ; CHECK-SD-NEXT: bl powl
1613 ; CHECK-SD-NEXT: str q0, [sp] // 16-byte Folded Spill
1614 ; CHECK-SD-NEXT: ldp q0, q1, [sp, #16] // 32-byte Folded Reload
1615 ; CHECK-SD-NEXT: bl powl
1616 ; CHECK-SD-NEXT: mov v1.16b, v0.16b
1617 ; CHECK-SD-NEXT: ldr q0, [sp] // 16-byte Folded Reload
1618 ; CHECK-SD-NEXT: ldr x30, [sp, #48] // 8-byte Folded Reload
1619 ; CHECK-SD-NEXT: add sp, sp, #64
1620 ; CHECK-SD-NEXT: ret
1621 ;
1622 ; CHECK-GI-LABEL: pow_v2fp128:
1623 ; CHECK-GI: // %bb.0: // %entry
1624 ; CHECK-GI-NEXT: sub sp, sp, #64
1625 ; CHECK-GI-NEXT: str x30, [sp, #48] // 8-byte Folded Spill
1626 ; CHECK-GI-NEXT: .cfi_def_cfa_offset 64
1627 ; CHECK-GI-NEXT: .cfi_offset w30, -16
1628 ; CHECK-GI-NEXT: stp q3, q1, [sp, #16] // 32-byte Folded Spill
1629 ; CHECK-GI-NEXT: mov v1.16b, v2.16b
1630 ; CHECK-GI-NEXT: bl powl
1631 ; CHECK-GI-NEXT: str q0, [sp] // 16-byte Folded Spill
1632 ; CHECK-GI-NEXT: ldp q1, q0, [sp, #16] // 32-byte Folded Reload
1633 ; CHECK-GI-NEXT: bl powl
1634 ; CHECK-GI-NEXT: mov v1.16b, v0.16b
1635 ; CHECK-GI-NEXT: ldr q0, [sp] // 16-byte Folded Reload
1636 ; CHECK-GI-NEXT: ldr x30, [sp, #48] // 8-byte Folded Reload
1637 ; CHECK-GI-NEXT: add sp, sp, #64
1638 ; CHECK-GI-NEXT: ret
1639 entry:
1640 %c = call <2 x fp128> @llvm.pow.v2fp128(<2 x fp128> %a, <2 x fp128> %b)
1641 ret <2 x fp128> %c
1642 }
1644 declare <16 x half> @llvm.pow.v16f16(<16 x half>, <16 x half>)
1645 declare <2 x double> @llvm.pow.v2f64(<2 x double>, <2 x double>)
1646 declare <2 x float> @llvm.pow.v2f32(<2 x float>, <2 x float>)
1647 declare <2 x fp128> @llvm.pow.v2fp128(<2 x fp128>, <2 x fp128>)
1648 declare <3 x double> @llvm.pow.v3f64(<3 x double>, <3 x double>)
1649 declare <3 x float> @llvm.pow.v3f32(<3 x float>, <3 x float>)
1650 declare <4 x double> @llvm.pow.v4f64(<4 x double>, <4 x double>)
1651 declare <4 x float> @llvm.pow.v4f32(<4 x float>, <4 x float>)
1652 declare <4 x half> @llvm.pow.v4f16(<4 x half>, <4 x half>)
1653 declare <7 x half> @llvm.pow.v7f16(<7 x half>, <7 x half>)
1654 declare <8 x float> @llvm.pow.v8f32(<8 x float>, <8 x float>)
1655 declare <8 x half> @llvm.pow.v8f16(<8 x half>, <8 x half>)
1656 declare double @llvm.pow.f64(double, double)
1657 declare float @llvm.pow.f32(float, float)
1658 declare fp128 @llvm.pow.fp128(fp128, fp128)
1659 declare half @llvm.pow.f16(half, half)