| blob | c0e38b92430338a1d1a4dcb17df58214141292c5 |
1 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
2 ; RUN: llc -mtriple=hexagon -hexagon-hvx-widen=32 < %s -verify-machineinstrs | FileCheck %s
4 target datalayout = "e-m:e-p:32:32:32-a:0-n16:32-i64:64:64-i32:32:32-i16:16:16-i1:8:8-f32:32:32-f64:64:64-v32:32:32-v64:64:64-v512:512:512-v1024:1024:1024-v2048:2048:2048"
5 target triple = "hexagon"
7 ; s8 -> f16
8 ; No widening
9 define void @s8f16_0(ptr %a0, ptr %a1) #0 {
10 ; CHECK-LABEL: s8f16_0:
11 ; CHECK: .cfi_startproc
12 ; CHECK-NEXT: // %bb.0:
13 ; CHECK-NEXT: {
14 ; CHECK-NEXT: v0 = vmem(r0+#0)
15 ; CHECK-NEXT: }
16 ; CHECK-NEXT: {
17 ; CHECK-NEXT: r7 = #1
18 ; CHECK-NEXT: r6 = #64
19 ; CHECK-NEXT: v1:0.h = vunpack(v0.b)
20 ; CHECK-NEXT: }
21 ; CHECK-NEXT: {
22 ; CHECK-NEXT: v2.h = vsplat(r7)
23 ; CHECK-NEXT: r3:2 = combine(#31,#5)
24 ; CHECK-NEXT: v3.h = vabs(v0.h)
25 ; CHECK-NEXT: v4.h = vabs(v1.h)
26 ; CHECK-NEXT: }
27 ; CHECK-NEXT: {
28 ; CHECK-NEXT: v8.h = vsplat(r6)
29 ; CHECK-NEXT: v7.h = vsplat(r3)
30 ; CHECK-NEXT: v9 = vxor(v9,v9)
31 ; CHECK-NEXT: }
32 ; CHECK-NEXT: {
33 ; CHECK-NEXT: r5 = ##32768
34 ; CHECK-NEXT: v5.uh = vcl0(v3.uh)
35 ; CHECK-NEXT: q0 = vcmp.gt(v9.h,v0.h)
36 ; CHECK-NEXT: }
37 ; CHECK-NEXT: {
38 ; CHECK-NEXT: v10.h = vsplat(r5)
39 ; CHECK-NEXT: r4 = #10
40 ; CHECK-NEXT: v6.uh = vcl0(v4.uh)
41 ; CHECK-NEXT: v5.h = vadd(v5.h,v2.h)
42 ; CHECK-NEXT: }
43 ; CHECK-NEXT: {
44 ; CHECK-NEXT: v27 = vmux(q0,v10,v9)
45 ; CHECK-NEXT: v6.h = vadd(v6.h,v2.h)
46 ; CHECK-NEXT: }
47 ; CHECK-NEXT: {
48 ; CHECK-NEXT: v3.h = vasl(v3.h,v5.h)
49 ; CHECK-NEXT: }
50 ; CHECK-NEXT: {
51 ; CHECK-NEXT: v4.h = vasl(v4.h,v6.h)
52 ; CHECK-NEXT: v13 = vand(v3,v8)
53 ; CHECK-NEXT: v11.h = vadd(v3.h,v7.h)
54 ; CHECK-NEXT: }
55 ; CHECK-NEXT: {
56 ; CHECK-NEXT: v14.h = vadd(v4.h,v7.h)
57 ; CHECK-NEXT: q2 = vcmp.eq(v13.h,v9.h)
58 ; CHECK-NEXT: v8 = vand(v4,v8)
59 ; CHECK-NEXT: q1 = vcmp.gt(v3.uh,v11.uh)
60 ; CHECK-NEXT: }
61 ; CHECK-NEXT: {
62 ; CHECK-NEXT: v11.uh = vlsr(v11.uh,r2)
63 ; CHECK-NEXT: v13 = vmux(q2,v9,v2)
64 ; CHECK-NEXT: q2 = vcmp.eq(v8.h,v9.h)
65 ; CHECK-NEXT: q3 = vcmp.gt(v4.uh,v14.uh)
66 ; CHECK-NEXT: }
67 ; CHECK-NEXT: {
68 ; CHECK-NEXT: v20.uh = vlsr(v14.uh,r2)
69 ; CHECK-NEXT: v22 = vmux(q2,v9,v2)
70 ; CHECK-NEXT: v21 = vmux(q1,v2,v9)
71 ; CHECK-NEXT: v2 = vmux(q3,v2,v9)
72 ; CHECK-NEXT: }
73 ; CHECK-NEXT: {
74 ; CHECK-NEXT: v19.uh = vlsr(v4.uh,r2)
75 ; CHECK-NEXT: v13.h = vadd(v11.h,v13.h)
76 ; CHECK-NEXT: v24.h = vadd(v20.h,v22.h)
77 ; CHECK-NEXT: v2.h = vadd(v2.h,v7.h)
78 ; CHECK-NEXT: }
79 ; CHECK-NEXT: {
80 ; CHECK-NEXT: v12.uh = vlsr(v3.uh,r2)
81 ; CHECK-NEXT: v23.h = vadd(v21.h,v7.h)
82 ; CHECK-NEXT: v2.h = vsub(v2.h,v6.h)
83 ; CHECK-NEXT: q3 = vcmp.gt(v9.h,v1.h)
84 ; CHECK-NEXT: }
85 ; CHECK-NEXT: {
86 ; CHECK-NEXT: v11.uh = vlsr(v11.uh,r7)
87 ; CHECK-NEXT: v3.h = vsub(v23.h,v5.h)
88 ; CHECK-NEXT: q2 = vcmp.eq(v12.h,v11.h)
89 ; CHECK-NEXT: q1 = vcmp.eq(v19.h,v20.h)
90 ; CHECK-NEXT: }
91 ; CHECK-NEXT: {
92 ; CHECK-NEXT: v25.uh = vlsr(v13.uh,r7)
93 ; CHECK-NEXT: v28 = vmux(q3,v10,v9)
94 ; CHECK-NEXT: q3 = vcmp.eq(v0.h,v9.h)
95 ; CHECK-NEXT: }
96 ; CHECK-NEXT: {
97 ; CHECK-NEXT: v26.uh = vlsr(v24.uh,r7)
98 ; CHECK-NEXT: v5 = vmux(q2,v25,v11)
99 ; CHECK-NEXT: q2 = vcmp.eq(v1.h,v9.h)
100 ; CHECK-NEXT: }
101 ; CHECK-NEXT: {
102 ; CHECK-NEXT: v4.uh = vlsr(v20.uh,r7)
103 ; CHECK-NEXT: v5 = vor(v27,v5)
104 ; CHECK-NEXT: }
105 ; CHECK-NEXT: {
106 ; CHECK-NEXT: v3.h = vasl(v3.h,r4)
107 ; CHECK-NEXT: v4 = vmux(q1,v26,v4)
108 ; CHECK-NEXT: }
109 ; CHECK-NEXT: {
110 ; CHECK-NEXT: v2.h = vasl(v2.h,r4)
111 ; CHECK-NEXT: v4 = vor(v28,v4)
112 ; CHECK-NEXT: v29 = vor(v5,v3)
113 ; CHECK-NEXT: }
114 ; CHECK-NEXT: {
115 ; CHECK-NEXT: v2 = vor(v4,v2)
116 ; CHECK-NEXT: v31 = vmux(q3,v9,v29)
117 ; CHECK-NEXT: vmem(r1+#0) = v31.new
118 ; CHECK-NEXT: }
119 ; CHECK-NEXT: {
120 ; CHECK-NEXT: v30 = vmux(q2,v9,v2)
121 ; CHECK-NEXT: jumpr r31
122 ; CHECK-NEXT: vmem(r1+#1) = v30.new
123 ; CHECK-NEXT: }
124 %v0 = load <128 x i8>, ptr %a0, align 128
125 %v1 = sitofp <128 x i8> %v0 to <128 x half>
126 store <128 x half> %v1, ptr %a1, align 128
127 ret void
128 }
130 ; Widen input
131 define void @s8f16_1(ptr %a0, ptr %a1) #0 {
132 ; CHECK-LABEL: s8f16_1:
133 ; CHECK: .cfi_startproc
134 ; CHECK-NEXT: // %bb.0:
135 ; CHECK-NEXT: {
136 ; CHECK-NEXT: v0 = vmem(r0+#0)
137 ; CHECK-NEXT: }
138 ; CHECK-NEXT: {
139 ; CHECK-NEXT: r6 = #1
140 ; CHECK-NEXT: r3:2 = combine(#64,#31)
141 ; CHECK-NEXT: v1:0.h = vunpack(v0.b)
142 ; CHECK-NEXT: }
143 ; CHECK-NEXT: {
144 ; CHECK-NEXT: v3.h = vsplat(r6)
145 ; CHECK-NEXT: v4.h = vsplat(r2)
146 ; CHECK-NEXT: v2.h = vabs(v0.h)
147 ; CHECK-NEXT: v1 = vxor(v1,v1)
148 ; CHECK-NEXT: }
149 ; CHECK-NEXT: {
150 ; CHECK-NEXT: v6.h = vsplat(r3)
151 ; CHECK-NEXT: r5:4 = combine(##32768,#5)
152 ; CHECK-NEXT: r2 = #10
153 ; CHECK-NEXT: }
154 ; CHECK-NEXT: {
155 ; CHECK-NEXT: v8.h = vsplat(r5)
156 ; CHECK-NEXT: v5.uh = vcl0(v2.uh)
157 ; CHECK-NEXT: q3 = vcmp.eq(v0.h,v1.h)
158 ; CHECK-NEXT: }
159 ; CHECK-NEXT: {
160 ; CHECK-NEXT: v5.h = vadd(v5.h,v3.h)
161 ; CHECK-NEXT: }
162 ; CHECK-NEXT: {
163 ; CHECK-NEXT: v2.h = vasl(v2.h,v5.h)
164 ; CHECK-NEXT: }
165 ; CHECK-NEXT: {
166 ; CHECK-NEXT: v7.h = vadd(v2.h,v4.h)
167 ; CHECK-NEXT: v6 = vand(v2,v6)
168 ; CHECK-NEXT: }
169 ; CHECK-NEXT: {
170 ; CHECK-NEXT: v2.uh = vlsr(v2.uh,r4)
171 ; CHECK-NEXT: q0 = vcmp.eq(v6.h,v1.h)
172 ; CHECK-NEXT: q1 = vcmp.gt(v2.uh,v7.uh)
173 ; CHECK-NEXT: }
174 ; CHECK-NEXT: {
175 ; CHECK-NEXT: v25.uh = vlsr(v7.uh,r4)
176 ; CHECK-NEXT: v26 = vmux(q0,v1,v3)
177 ; CHECK-NEXT: v3 = vmux(q1,v3,v1)
178 ; CHECK-NEXT: q1 = vcmp.gt(v1.h,v0.h)
179 ; CHECK-NEXT: }
180 ; CHECK-NEXT: {
181 ; CHECK-NEXT: v7.h = vadd(v25.h,v26.h)
182 ; CHECK-NEXT: v3.h = vadd(v3.h,v4.h)
183 ; CHECK-NEXT: q2 = vcmp.eq(v2.h,v25.h)
184 ; CHECK-NEXT: v30 = vmux(q1,v8,v1)
185 ; CHECK-NEXT: }
186 ; CHECK-NEXT: {
187 ; CHECK-NEXT: v27.uh = vlsr(v25.uh,r6)
188 ; CHECK-NEXT: v28.h = vsub(v3.h,v5.h)
189 ; CHECK-NEXT: }
190 ; CHECK-NEXT: {
191 ; CHECK-NEXT: v29.uh = vlsr(v7.uh,r6)
192 ; CHECK-NEXT: }
193 ; CHECK-NEXT: {
194 ; CHECK-NEXT: v2.h = vasl(v28.h,r2)
195 ; CHECK-NEXT: v3 = vmux(q2,v29,v27)
196 ; CHECK-NEXT: }
197 ; CHECK-NEXT: {
198 ; CHECK-NEXT: v3 = vor(v30,v3)
199 ; CHECK-NEXT: }
200 ; CHECK-NEXT: {
201 ; CHECK-NEXT: v31 = vor(v3,v2)
202 ; CHECK-NEXT: }
203 ; CHECK-NEXT: {
204 ; CHECK-NEXT: v0 = vmux(q3,v1,v31)
205 ; CHECK-NEXT: jumpr r31
206 ; CHECK-NEXT: vmem(r1+#0) = v0.new
207 ; CHECK-NEXT: }
208 %v0 = load <64 x i8>, ptr %a0, align 128
209 %v1 = sitofp <64 x i8> %v0 to <64 x half>
210 store <64 x half> %v1, ptr %a1, align 128
211 ret void
212 }
215 ; s8 -> f32
216 ; No widening
217 define void @s8f32_0(ptr %a0, ptr %a1) #0 {
218 ; CHECK-LABEL: s8f32_0:
219 ; CHECK: .cfi_startproc
220 ; CHECK-NEXT: // %bb.0:
221 ; CHECK-NEXT: {
222 ; CHECK-NEXT: r7 = #64
223 ; CHECK-NEXT: r0 = #1
224 ; CHECK-NEXT: v0 = vmem(r0+#0)
225 ; CHECK-NEXT: }
226 ; CHECK-NEXT: {
227 ; CHECK-NEXT: v2 = vsplat(r0)
228 ; CHECK-NEXT: r3:2 = combine(##255,#8)
229 ; CHECK-NEXT: v1 = valign(v0,v0,r7)
230 ; CHECK-NEXT: }
231 ; CHECK-NEXT: {
232 ; CHECK-NEXT: v3 = vsplat(r3)
233 ; CHECK-NEXT: r7 = #512
234 ; CHECK-NEXT: v9:8.h = vunpack(v0.b)
235 ; CHECK-NEXT: }
236 ; CHECK-NEXT: {
237 ; CHECK-NEXT: v4 = vsplat(r7)
238 ; CHECK-NEXT: r6 = ##-2147483648
239 ; CHECK-NEXT: r5 = #159
240 ; CHECK-NEXT: }
241 ; CHECK-NEXT: {
242 ; CHECK-NEXT: r4 = #23
243 ; CHECK-NEXT: v7:6.h = vunpack(v1.b)
244 ; CHECK-NEXT: }
245 ; CHECK-NEXT: {
246 ; CHECK-NEXT: v8 = vsplat(r6)
247 ; CHECK-NEXT: v1:0.w = vunpack(v8.h)
248 ; CHECK-NEXT: }
249 ; CHECK-NEXT: {
250 ; CHECK-NEXT: v7:6.w = vunpack(v6.h)
251 ; CHECK-NEXT: v5.w = vabs(v0.w)
252 ; CHECK-NEXT: v10.w = vabs(v1.w)
253 ; CHECK-NEXT: }
254 ; CHECK-NEXT: {
255 ; CHECK-NEXT: v26.w = vabs(v6.w)
256 ; CHECK-NEXT: v13.w = vabs(v7.w)
257 ; CHECK-NEXT: }
258 ; CHECK-NEXT: {
259 ; CHECK-NEXT: v9.uw = vcl0(v5.uw)
260 ; CHECK-NEXT: }
261 ; CHECK-NEXT: {
262 ; CHECK-NEXT: v12.uw = vcl0(v26.uw)
263 ; CHECK-NEXT: v9.w = vadd(v9.w,v2.w)
264 ; CHECK-NEXT: }
265 ; CHECK-NEXT: {
266 ; CHECK-NEXT: v14.uw = vcl0(v13.uw)
267 ; CHECK-NEXT: v15.w = vadd(v12.w,v2.w)
268 ; CHECK-NEXT: }
269 ; CHECK-NEXT: {
270 ; CHECK-NEXT: v11.uw = vcl0(v10.uw)
271 ; CHECK-NEXT: v12.w = vadd(v14.w,v2.w)
272 ; CHECK-NEXT: }
273 ; CHECK-NEXT: {
274 ; CHECK-NEXT: v27.w = vasl(v26.w,v15.w)
275 ; CHECK-NEXT: v11.w = vadd(v11.w,v2.w)
276 ; CHECK-NEXT: }
277 ; CHECK-NEXT: {
278 ; CHECK-NEXT: v13.w = vasl(v13.w,v12.w)
279 ; CHECK-NEXT: v20 = vand(v27,v4)
280 ; CHECK-NEXT: v19.w = vadd(v27.w,v3.w)
281 ; CHECK-NEXT: }
282 ; CHECK-NEXT: {
283 ; CHECK-NEXT: v16.w = vasl(v5.w,v9.w)
284 ; CHECK-NEXT: v5 = vxor(v5,v5)
285 ; CHECK-NEXT: v23.w = vadd(v13.w,v3.w)
286 ; CHECK-NEXT: v28 = vand(v13,v4)
287 ; CHECK-NEXT: }
288 ; CHECK-NEXT: {
289 ; CHECK-NEXT: v17.w = vasl(v10.w,v11.w)
290 ; CHECK-NEXT: q3 = vcmp.eq(v20.w,v5.w)
291 ; CHECK-NEXT: q2 = vcmp.gt(v27.uw,v19.uw)
292 ; CHECK-NEXT: q0 = vcmp.gt(v5.w,v6.w)
293 ; CHECK-NEXT: }
294 ; CHECK-NEXT: {
295 ; CHECK-NEXT: v21.uw = vlsr(v27.uw,r2)
296 ; CHECK-NEXT: v30 = vmux(q3,v5,v2)
297 ; CHECK-NEXT: q3 = vcmp.eq(v28.w,v5.w)
298 ; CHECK-NEXT: v22 = vand(v17,v4)
299 ; CHECK-NEXT: }
300 ; CHECK-NEXT: {
301 ; CHECK-NEXT: v14.uw = vlsr(v19.uw,r2)
302 ; CHECK-NEXT: v27 = vmux(q3,v5,v2)
303 ; CHECK-NEXT: q1 = vcmp.eq(v22.w,v5.w)
304 ; CHECK-NEXT: v24 = vmux(q2,v2,v5)
305 ; CHECK-NEXT: }
306 ; CHECK-NEXT: {
307 ; CHECK-NEXT: v31.uw = vlsr(v23.uw,r2)
308 ; CHECK-NEXT: v22.w = vadd(v14.w,v30.w)
309 ; CHECK-NEXT: v30.w = vadd(v17.w,v3.w)
310 ; CHECK-NEXT: q2 = vcmp.eq(v21.w,v14.w)
311 ; CHECK-NEXT: }
312 ; CHECK-NEXT: {
313 ; CHECK-NEXT: v29.uw = vlsr(v13.uw,r2)
314 ; CHECK-NEXT: v28.w = vadd(v31.w,v27.w)
315 ; CHECK-NEXT: v3.w = vadd(v16.w,v3.w)
316 ; CHECK-NEXT: v4 = vand(v16,v4)
317 ; CHECK-NEXT: }
318 ; CHECK-NEXT: {
319 ; CHECK-NEXT: v14.uw = vlsr(v14.uw,r0)
320 ; CHECK-NEXT: q3 = vcmp.eq(v29.w,v31.w)
321 ; CHECK-NEXT: v18 = vmux(q0,v8,v5)
322 ; CHECK-NEXT: q0 = vcmp.gt(v5.w,v7.w)
323 ; CHECK-NEXT: }
324 ; CHECK-NEXT: {
325 ; CHECK-NEXT: v19.uw = vlsr(v31.uw,r0)
326 ; CHECK-NEXT: v26 = vmux(q1,v5,v2)
327 ; CHECK-NEXT: v31 = vmux(q0,v8,v5)
328 ; CHECK-NEXT: q0 = vcmp.gt(v16.uw,v3.uw)
329 ; CHECK-NEXT: }
330 ; CHECK-NEXT: {
331 ; CHECK-NEXT: v10 = vsplat(r5)
332 ; CHECK-NEXT: v29.uw = vlsr(v22.uw,r0)
333 ; CHECK-NEXT: v15.w = vsub(v24.w,v15.w)
334 ; CHECK-NEXT: }
335 ; CHECK-NEXT: {
336 ; CHECK-NEXT: v20.uw = vlsr(v28.uw,r0)
337 ; CHECK-NEXT: v14 = vmux(q2,v29,v14)
338 ; CHECK-NEXT: q2 = vcmp.gt(v13.uw,v23.uw)
339 ; CHECK-NEXT: v15.w = vadd(v15.w,v10.w)
340 ; CHECK-NEXT: }
341 ; CHECK-NEXT: {
342 ; CHECK-NEXT: v25.uw = vlsr(v30.uw,r2)
343 ; CHECK-NEXT: v19 = vmux(q3,v20,v19)
344 ; CHECK-NEXT: q3 = vcmp.eq(v4.w,v5.w)
345 ; CHECK-NEXT: v27 = vmux(q2,v2,v5)
346 ; CHECK-NEXT: }
347 ; CHECK-NEXT: {
348 ; CHECK-NEXT: v3.uw = vlsr(v3.uw,r2)
349 ; CHECK-NEXT: q2 = vcmp.gt(v17.uw,v30.uw)
350 ; CHECK-NEXT: v28.w = vadd(v25.w,v26.w)
351 ; CHECK-NEXT: v29 = vmux(q3,v5,v2)
352 ; CHECK-NEXT: }
353 ; CHECK-NEXT: {
354 ; CHECK-NEXT: v17.uw = vlsr(v17.uw,r2)
355 ; CHECK-NEXT: v19 = vor(v31,v19)
356 ; CHECK-NEXT: v31 = vmux(q2,v2,v5)
357 ; CHECK-NEXT: v2 = vmux(q0,v2,v5)
358 ; CHECK-NEXT: }
359 ; CHECK-NEXT: {
360 ; CHECK-NEXT: v24.uw = vlsr(v16.uw,r2)
361 ; CHECK-NEXT: v30.w = vadd(v3.w,v29.w)
362 ; CHECK-NEXT: v2.w = vsub(v2.w,v9.w)
363 ; CHECK-NEXT: v11.w = vsub(v31.w,v11.w)
364 ; CHECK-NEXT: }
365 ; CHECK-NEXT: {
366 ; CHECK-NEXT: v16.uw = vlsr(v28.uw,r0)
367 ; CHECK-NEXT: q3 = vcmp.eq(v17.w,v25.w)
368 ; CHECK-NEXT: v4.w = vsub(v27.w,v12.w)
369 ; CHECK-NEXT: v2.w = vadd(v2.w,v10.w)
370 ; CHECK-NEXT: }
371 ; CHECK-NEXT: {
372 ; CHECK-NEXT: v13.uw = vlsr(v25.uw,r0)
373 ; CHECK-NEXT: q0 = vcmp.eq(v24.w,v3.w)
374 ; CHECK-NEXT: v21.w = vadd(v11.w,v10.w)
375 ; CHECK-NEXT: q2 = vcmp.gt(v5.w,v1.w)
376 ; CHECK-NEXT: }
377 ; CHECK-NEXT: {
378 ; CHECK-NEXT: v22.uw = vlsr(v30.uw,r0)
379 ; CHECK-NEXT: v23 = vmux(q3,v16,v13)
380 ; CHECK-NEXT: q3 = vcmp.gt(v5.w,v0.w)
381 ; CHECK-NEXT: v24 = vmux(q2,v8,v5)
382 ; CHECK-NEXT: }
383 ; CHECK-NEXT: {
384 ; CHECK-NEXT: v3.uw = vlsr(v3.uw,r0)
385 ; CHECK-NEXT: v4.w = vadd(v4.w,v10.w)
386 ; CHECK-NEXT: v8 = vmux(q3,v8,v5)
387 ; CHECK-NEXT: v10 = vor(v24,v23)
388 ; CHECK-NEXT: }
389 ; CHECK-NEXT: {
390 ; CHECK-NEXT: v9.w = vasl(v21.w,r4)
391 ; CHECK-NEXT: v3 = vmux(q0,v22,v3)
392 ; CHECK-NEXT: v14 = vor(v18,v14)
393 ; CHECK-NEXT: q2 = vcmp.eq(v1.w,v5.w)
394 ; CHECK-NEXT: }
395 ; CHECK-NEXT: {
396 ; CHECK-NEXT: v2.w = vasl(v2.w,r4)
397 ; CHECK-NEXT: v3 = vor(v8,v3)
398 ; CHECK-NEXT: v25 = vor(v10,v9)
399 ; CHECK-NEXT: q3 = vcmp.eq(v0.w,v5.w)
400 ; CHECK-NEXT: }
401 ; CHECK-NEXT: {
402 ; CHECK-NEXT: v15.w = vasl(v15.w,r4)
403 ; CHECK-NEXT: v2 = vor(v3,v2)
404 ; CHECK-NEXT: v27 = vmux(q2,v5,v25)
405 ; CHECK-NEXT: vmem(r1+#1) = v27.new
406 ; CHECK-NEXT: }
407 ; CHECK-NEXT: {
408 ; CHECK-NEXT: v26.w = vasl(v4.w,r4)
409 ; CHECK-NEXT: v29 = vmux(q3,v5,v2)
410 ; CHECK-NEXT: q2 = vcmp.eq(v7.w,v5.w)
411 ; CHECK-NEXT: vmem(r1+#0) = v29.new
412 ; CHECK-NEXT: }
413 ; CHECK-NEXT: {
414 ; CHECK-NEXT: v28 = vor(v19,v26)
415 ; CHECK-NEXT: v30 = vor(v14,v15)
416 ; CHECK-NEXT: q3 = vcmp.eq(v6.w,v5.w)
417 ; CHECK-NEXT: }
418 ; CHECK-NEXT: {
419 ; CHECK-NEXT: v0 = vmux(q2,v5,v28)
420 ; CHECK-NEXT: v31 = vmux(q3,v5,v30)
421 ; CHECK-NEXT: vmem(r1+#3) = v0.new
422 ; CHECK-NEXT: }
423 ; CHECK-NEXT: {
424 ; CHECK-NEXT: jumpr r31
425 ; CHECK-NEXT: vmem(r1+#2) = v31
426 ; CHECK-NEXT: }
427 %v0 = load <128 x i8>, ptr %a0, align 128
428 %v1 = sitofp <128 x i8> %v0 to <128 x float>
429 store <128 x float> %v1, ptr %a1, align 128
430 ret void
431 }
433 ; Widen input #1
434 define void @s8f32_1(ptr %a0, ptr %a1) #0 {
435 ; CHECK-LABEL: s8f32_1:
436 ; CHECK: .cfi_startproc
437 ; CHECK-NEXT: // %bb.0:
438 ; CHECK-NEXT: {
439 ; CHECK-NEXT: r0 = #1
440 ; CHECK-NEXT: v3:2.h = vunpack(v0.b)
441 ; CHECK-NEXT: v0.cur = vmem(r0+#0)
442 ; CHECK-NEXT: }
443 ; CHECK-NEXT: {
444 ; CHECK-NEXT: v1 = vsplat(r0)
445 ; CHECK-NEXT: r3:2 = combine(##255,#8)
446 ; CHECK-NEXT: r6 = #512
447 ; CHECK-NEXT: }
448 ; CHECK-NEXT: {
449 ; CHECK-NEXT: v7 = vsplat(r3)
450 ; CHECK-NEXT: v3:2.w = vunpack(v2.h)
451 ; CHECK-NEXT: v22 = vxor(v22,v22)
452 ; CHECK-NEXT: }
453 ; CHECK-NEXT: {
454 ; CHECK-NEXT: v10 = vsplat(r6)
455 ; CHECK-NEXT: r7 = ##-2147483648
456 ; CHECK-NEXT: r5 = #159
457 ; CHECK-NEXT: }
458 ; CHECK-NEXT: {
459 ; CHECK-NEXT: v9 = vsplat(r7)
460 ; CHECK-NEXT: v4.w = vabs(v2.w)
461 ; CHECK-NEXT: v5.w = vabs(v3.w)
462 ; CHECK-NEXT: q0 = vcmp.gt(v22.w,v2.w)
463 ; CHECK-NEXT: }
464 ; CHECK-NEXT: {
465 ; CHECK-NEXT: v12 = vsplat(r5)
466 ; CHECK-NEXT: r4 = #23
467 ; CHECK-NEXT: v11 = vmux(q0,v9,v22)
468 ; CHECK-NEXT: q0 = vcmp.gt(v22.w,v3.w)
469 ; CHECK-NEXT: }
470 ; CHECK-NEXT: {
471 ; CHECK-NEXT: v6.uw = vcl0(v4.uw)
472 ; CHECK-NEXT: v30 = vmux(q0,v9,v22)
473 ; CHECK-NEXT: }
474 ; CHECK-NEXT: {
475 ; CHECK-NEXT: v8.uw = vcl0(v5.uw)
476 ; CHECK-NEXT: v6.w = vadd(v6.w,v1.w)
477 ; CHECK-NEXT: }
478 ; CHECK-NEXT: {
479 ; CHECK-NEXT: v8.w = vadd(v8.w,v1.w)
480 ; CHECK-NEXT: }
481 ; CHECK-NEXT: {
482 ; CHECK-NEXT: v4.w = vasl(v4.w,v6.w)
483 ; CHECK-NEXT: }
484 ; CHECK-NEXT: {
485 ; CHECK-NEXT: v5.w = vasl(v5.w,v8.w)
486 ; CHECK-NEXT: v13 = vand(v4,v10)
487 ; CHECK-NEXT: v14.w = vadd(v4.w,v7.w)
488 ; CHECK-NEXT: }
489 ; CHECK-NEXT: {
490 ; CHECK-NEXT: v10 = vand(v5,v10)
491 ; CHECK-NEXT: v7.w = vadd(v5.w,v7.w)
492 ; CHECK-NEXT: q2 = vcmp.gt(v4.uw,v14.uw)
493 ; CHECK-NEXT: q1 = vcmp.eq(v13.w,v22.w)
494 ; CHECK-NEXT: }
495 ; CHECK-NEXT: {
496 ; CHECK-NEXT: v14.uw = vlsr(v14.uw,r2)
497 ; CHECK-NEXT: q3 = vcmp.eq(v10.w,v22.w)
498 ; CHECK-NEXT: v25 = vmux(q2,v1,v22)
499 ; CHECK-NEXT: q2 = vcmp.gt(v5.uw,v7.uw)
500 ; CHECK-NEXT: }
501 ; CHECK-NEXT: {
502 ; CHECK-NEXT: v7.uw = vlsr(v7.uw,r2)
503 ; CHECK-NEXT: v26 = vmux(q1,v22,v1)
504 ; CHECK-NEXT: v27 = vmux(q3,v22,v1)
505 ; CHECK-NEXT: v1 = vmux(q2,v1,v22)
506 ; CHECK-NEXT: }
507 ; CHECK-NEXT: {
508 ; CHECK-NEXT: v24.uw = vlsr(v5.uw,r2)
509 ; CHECK-NEXT: v5.w = vadd(v14.w,v26.w)
510 ; CHECK-NEXT: v29.w = vadd(v7.w,v27.w)
511 ; CHECK-NEXT: v6.w = vsub(v25.w,v6.w)
512 ; CHECK-NEXT: }
513 ; CHECK-NEXT: {
514 ; CHECK-NEXT: v23.uw = vlsr(v4.uw,r2)
515 ; CHECK-NEXT: v1.w = vsub(v1.w,v8.w)
516 ; CHECK-NEXT: v6.w = vadd(v6.w,v12.w)
517 ; CHECK-NEXT: q3 = vcmp.eq(v24.w,v7.w)
518 ; CHECK-NEXT: }
519 ; CHECK-NEXT: {
520 ; CHECK-NEXT: v28.uw = vlsr(v14.uw,r0)
521 ; CHECK-NEXT: v1.w = vadd(v1.w,v12.w)
522 ; CHECK-NEXT: q1 = vcmp.eq(v23.w,v14.w)
523 ; CHECK-NEXT: q2 = vcmp.eq(v3.w,v22.w)
524 ; CHECK-NEXT: }
525 ; CHECK-NEXT: {
526 ; CHECK-NEXT: v5.uw = vlsr(v5.uw,r0)
527 ; CHECK-NEXT: }
528 ; CHECK-NEXT: {
529 ; CHECK-NEXT: v7.uw = vlsr(v7.uw,r0)
530 ; CHECK-NEXT: v5 = vmux(q1,v5,v28)
531 ; CHECK-NEXT: }
532 ; CHECK-NEXT: {
533 ; CHECK-NEXT: v4.uw = vlsr(v29.uw,r0)
534 ; CHECK-NEXT: v5 = vor(v11,v5)
535 ; CHECK-NEXT: }
536 ; CHECK-NEXT: {
537 ; CHECK-NEXT: v6.w = vasl(v6.w,r4)
538 ; CHECK-NEXT: v4 = vmux(q3,v4,v7)
539 ; CHECK-NEXT: q3 = vcmp.eq(v2.w,v22.w)
540 ; CHECK-NEXT: }
541 ; CHECK-NEXT: {
542 ; CHECK-NEXT: v1.w = vasl(v1.w,r4)
543 ; CHECK-NEXT: v4 = vor(v30,v4)
544 ; CHECK-NEXT: v31 = vor(v5,v6)
545 ; CHECK-NEXT: }
546 ; CHECK-NEXT: {
547 ; CHECK-NEXT: v1 = vor(v4,v1)
548 ; CHECK-NEXT: v0 = vmux(q3,v22,v31)
549 ; CHECK-NEXT: vmem(r1+#0) = v0.new
550 ; CHECK-NEXT: }
551 ; CHECK-NEXT: {
552 ; CHECK-NEXT: v1 = vmux(q2,v22,v1)
553 ; CHECK-NEXT: jumpr r31
554 ; CHECK-NEXT: vmem(r1+#1) = v1.new
555 ; CHECK-NEXT: }
556 %v0 = load <64 x i8>, ptr %a0, align 128
557 %v1 = sitofp <64 x i8> %v0 to <64 x float>
558 store <64 x float> %v1, ptr %a1, align 128
559 ret void
560 }
562 ; Widen input #2
563 define void @s8f32_2(ptr %a0, ptr %a1) #0 {
564 ; CHECK-LABEL: s8f32_2:
565 ; CHECK: .cfi_startproc
566 ; CHECK-NEXT: // %bb.0:
567 ; CHECK-NEXT: {
568 ; CHECK-NEXT: v0 = vmem(r0+#0)
569 ; CHECK-NEXT: }
570 ; CHECK-NEXT: {
571 ; CHECK-NEXT: r0 = #1
572 ; CHECK-NEXT: r3 = #512
573 ; CHECK-NEXT: v1:0.h = vunpack(v0.b)
574 ; CHECK-NEXT: }
575 ; CHECK-NEXT: {
576 ; CHECK-NEXT: v2 = vsplat(r0)
577 ; CHECK-NEXT: v4 = vsplat(r3)
578 ; CHECK-NEXT: r2 = #255
579 ; CHECK-NEXT: v3 = vxor(v3,v3)
580 ; CHECK-NEXT: }
581 ; CHECK-NEXT: {
582 ; CHECK-NEXT: r7:6 = combine(##-2147483648,#8)
583 ; CHECK-NEXT: r4 = #159
584 ; CHECK-NEXT: v1:0.w = vunpack(v0.h)
585 ; CHECK-NEXT: }
586 ; CHECK-NEXT: {
587 ; CHECK-NEXT: v1 = vsplat(r2)
588 ; CHECK-NEXT: v8 = vsplat(r4)
589 ; CHECK-NEXT: v5.w = vabs(v0.w)
590 ; CHECK-NEXT: q2 = vcmp.gt(v3.w,v0.w)
591 ; CHECK-NEXT: }
592 ; CHECK-NEXT: {
593 ; CHECK-NEXT: v7 = vsplat(r7)
594 ; CHECK-NEXT: r2 = #23
595 ; CHECK-NEXT: }
596 ; CHECK-NEXT: {
597 ; CHECK-NEXT: v6.uw = vcl0(v5.uw)
598 ; CHECK-NEXT: v30 = vmux(q2,v7,v3)
599 ; CHECK-NEXT: }
600 ; CHECK-NEXT: {
601 ; CHECK-NEXT: v6.w = vadd(v6.w,v2.w)
602 ; CHECK-NEXT: }
603 ; CHECK-NEXT: {
604 ; CHECK-NEXT: v5.w = vasl(v5.w,v6.w)
605 ; CHECK-NEXT: }
606 ; CHECK-NEXT: {
607 ; CHECK-NEXT: v1.w = vadd(v5.w,v1.w)
608 ; CHECK-NEXT: v4 = vand(v5,v4)
609 ; CHECK-NEXT: }
610 ; CHECK-NEXT: {
611 ; CHECK-NEXT: v5.uw = vlsr(v5.uw,r6)
612 ; CHECK-NEXT: q0 = vcmp.eq(v4.w,v3.w)
613 ; CHECK-NEXT: q1 = vcmp.gt(v5.uw,v1.uw)
614 ; CHECK-NEXT: }
615 ; CHECK-NEXT: {
616 ; CHECK-NEXT: v1.uw = vlsr(v1.uw,r6)
617 ; CHECK-NEXT: v4 = vmux(q0,v3,v2)
618 ; CHECK-NEXT: v2 = vmux(q1,v2,v3)
619 ; CHECK-NEXT: }
620 ; CHECK-NEXT: {
621 ; CHECK-NEXT: v4.w = vadd(v1.w,v4.w)
622 ; CHECK-NEXT: v2.w = vsub(v2.w,v6.w)
623 ; CHECK-NEXT: q3 = vcmp.eq(v5.w,v1.w)
624 ; CHECK-NEXT: }
625 ; CHECK-NEXT: {
626 ; CHECK-NEXT: v28.uw = vlsr(v1.uw,r0)
627 ; CHECK-NEXT: v2.w = vadd(v2.w,v8.w)
628 ; CHECK-NEXT: }
629 ; CHECK-NEXT: {
630 ; CHECK-NEXT: v29.uw = vlsr(v4.uw,r0)
631 ; CHECK-NEXT: }
632 ; CHECK-NEXT: {
633 ; CHECK-NEXT: v2.w = vasl(v2.w,r2)
634 ; CHECK-NEXT: v1 = vmux(q3,v29,v28)
635 ; CHECK-NEXT: q3 = vcmp.eq(v0.w,v3.w)
636 ; CHECK-NEXT: }
637 ; CHECK-NEXT: {
638 ; CHECK-NEXT: v1 = vor(v30,v1)
639 ; CHECK-NEXT: }
640 ; CHECK-NEXT: {
641 ; CHECK-NEXT: v31 = vor(v1,v2)
642 ; CHECK-NEXT: }
643 ; CHECK-NEXT: {
644 ; CHECK-NEXT: v0 = vmux(q3,v3,v31)
645 ; CHECK-NEXT: jumpr r31
646 ; CHECK-NEXT: vmem(r1+#0) = v0.new
647 ; CHECK-NEXT: }
648 %v0 = load <32 x i8>, ptr %a0, align 128
649 %v1 = sitofp <32 x i8> %v0 to <32 x float>
650 store <32 x float> %v1, ptr %a1, align 128
651 ret void
652 }
655 ; s16 -> f16
656 ; No widening
657 define void @s16f16_0(ptr %a0, ptr %a1) #0 {
658 ; CHECK-LABEL: s16f16_0:
659 ; CHECK: .cfi_startproc
660 ; CHECK-NEXT: // %bb.0:
661 ; CHECK-NEXT: {
662 ; CHECK-NEXT: r6 = #1
663 ; CHECK-NEXT: r3:2 = combine(#64,#31)
664 ; CHECK-NEXT: v1.h = vabs(v0.h)
665 ; CHECK-NEXT: v0.cur = vmem(r0+#0)
666 ; CHECK-NEXT: }
667 ; CHECK-NEXT: {
668 ; CHECK-NEXT: v3.h = vsplat(r6)
669 ; CHECK-NEXT: v5.h = vsplat(r2)
670 ; CHECK-NEXT: v2 = vxor(v2,v2)
671 ; CHECK-NEXT: }
672 ; CHECK-NEXT: {
673 ; CHECK-NEXT: v6.h = vsplat(r3)
674 ; CHECK-NEXT: r5:4 = combine(##32768,#5)
675 ; CHECK-NEXT: v4.uh = vcl0(v1.uh)
676 ; CHECK-NEXT: }
677 ; CHECK-NEXT: {
678 ; CHECK-NEXT: v8.h = vsplat(r5)
679 ; CHECK-NEXT: r2 = #10
680 ; CHECK-NEXT: v4.h = vadd(v4.h,v3.h)
681 ; CHECK-NEXT: q3 = vcmp.eq(v0.h,v2.h)
682 ; CHECK-NEXT: }
683 ; CHECK-NEXT: {
684 ; CHECK-NEXT: v1.h = vasl(v1.h,v4.h)
685 ; CHECK-NEXT: }
686 ; CHECK-NEXT: {
687 ; CHECK-NEXT: v7.h = vadd(v1.h,v5.h)
688 ; CHECK-NEXT: v6 = vand(v1,v6)
689 ; CHECK-NEXT: }
690 ; CHECK-NEXT: {
691 ; CHECK-NEXT: v1.uh = vlsr(v1.uh,r4)
692 ; CHECK-NEXT: q0 = vcmp.eq(v6.h,v2.h)
693 ; CHECK-NEXT: q1 = vcmp.gt(v1.uh,v7.uh)
694 ; CHECK-NEXT: }
695 ; CHECK-NEXT: {
696 ; CHECK-NEXT: v25.uh = vlsr(v7.uh,r4)
697 ; CHECK-NEXT: v26 = vmux(q0,v2,v3)
698 ; CHECK-NEXT: v3 = vmux(q1,v3,v2)
699 ; CHECK-NEXT: q1 = vcmp.gt(v2.h,v0.h)
700 ; CHECK-NEXT: }
701 ; CHECK-NEXT: {
702 ; CHECK-NEXT: v7.h = vadd(v25.h,v26.h)
703 ; CHECK-NEXT: v3.h = vadd(v3.h,v5.h)
704 ; CHECK-NEXT: q2 = vcmp.eq(v1.h,v25.h)
705 ; CHECK-NEXT: v30 = vmux(q1,v8,v2)
706 ; CHECK-NEXT: }
707 ; CHECK-NEXT: {
708 ; CHECK-NEXT: v27.uh = vlsr(v25.uh,r6)
709 ; CHECK-NEXT: v28.h = vsub(v3.h,v4.h)
710 ; CHECK-NEXT: }
711 ; CHECK-NEXT: {
712 ; CHECK-NEXT: v29.uh = vlsr(v7.uh,r6)
713 ; CHECK-NEXT: }
714 ; CHECK-NEXT: {
715 ; CHECK-NEXT: v1.h = vasl(v28.h,r2)
716 ; CHECK-NEXT: v3 = vmux(q2,v29,v27)
717 ; CHECK-NEXT: }
718 ; CHECK-NEXT: {
719 ; CHECK-NEXT: v3 = vor(v30,v3)
720 ; CHECK-NEXT: }
721 ; CHECK-NEXT: {
722 ; CHECK-NEXT: v31 = vor(v3,v1)
723 ; CHECK-NEXT: }
724 ; CHECK-NEXT: {
725 ; CHECK-NEXT: v0 = vmux(q3,v2,v31)
726 ; CHECK-NEXT: jumpr r31
727 ; CHECK-NEXT: vmem(r1+#0) = v0.new
728 ; CHECK-NEXT: }
729 %v0 = load <64 x i16>, ptr %a0, align 128
730 %v1 = sitofp <64 x i16> %v0 to <64 x half>
731 store <64 x half> %v1, ptr %a1, align 128
732 ret void
733 }
735 ; Widen input and result
736 define void @s16f16_1(ptr %a0, ptr %a1) #0 {
737 ; CHECK-LABEL: s16f16_1:
738 ; CHECK: .cfi_startproc
739 ; CHECK-NEXT: // %bb.0:
740 ; CHECK-NEXT: {
741 ; CHECK-NEXT: r3:2 = combine(#31,#1)
742 ; CHECK-NEXT: r7 = #64
743 ; CHECK-NEXT: v1.h = vabs(v0.h)
744 ; CHECK-NEXT: v0.cur = vmem(r0+#0)
745 ; CHECK-NEXT: }
746 ; CHECK-NEXT: {
747 ; CHECK-NEXT: v2.h = vsplat(r2)
748 ; CHECK-NEXT: v5.h = vsplat(r3)
749 ; CHECK-NEXT: r6 = #5
750 ; CHECK-NEXT: v3 = vxor(v3,v3)
751 ; CHECK-NEXT: }
752 ; CHECK-NEXT: {
753 ; CHECK-NEXT: v6.h = vsplat(r7)
754 ; CHECK-NEXT: r4 = ##32768
755 ; CHECK-NEXT: v4.uh = vcl0(v1.uh)
756 ; CHECK-NEXT: }
757 ; CHECK-NEXT: {
758 ; CHECK-NEXT: v8.h = vsplat(r4)
759 ; CHECK-NEXT: r4 = #10
760 ; CHECK-NEXT: q2 = vcmp.gt(v3.h,v0.h)
761 ; CHECK-NEXT: v4.h = vadd(v4.h,v2.h)
762 ; CHECK-NEXT: }
763 ; CHECK-NEXT: {
764 ; CHECK-NEXT: v30 = vmux(q2,v8,v3)
765 ; CHECK-NEXT: q2 = vcmp.eq(v0.h,v3.h)
766 ; CHECK-NEXT: }
767 ; CHECK-NEXT: {
768 ; CHECK-NEXT: v1.h = vasl(v1.h,v4.h)
769 ; CHECK-NEXT: }
770 ; CHECK-NEXT: {
771 ; CHECK-NEXT: v7.h = vadd(v1.h,v5.h)
772 ; CHECK-NEXT: v6 = vand(v1,v6)
773 ; CHECK-NEXT: }
774 ; CHECK-NEXT: {
775 ; CHECK-NEXT: v1.uh = vlsr(v1.uh,r6)
776 ; CHECK-NEXT: q1 = vcmp.eq(v6.h,v3.h)
777 ; CHECK-NEXT: q0 = vcmp.gt(v1.uh,v7.uh)
778 ; CHECK-NEXT: }
779 ; CHECK-NEXT: {
780 ; CHECK-NEXT: v25.uh = vlsr(v7.uh,r6)
781 ; CHECK-NEXT: v26 = vmux(q1,v3,v2)
782 ; CHECK-NEXT: v2 = vmux(q0,v2,v3)
783 ; CHECK-NEXT: }
784 ; CHECK-NEXT: {
785 ; CHECK-NEXT: v7.h = vadd(v25.h,v26.h)
786 ; CHECK-NEXT: v2.h = vadd(v2.h,v5.h)
787 ; CHECK-NEXT: q3 = vcmp.eq(v1.h,v25.h)
788 ; CHECK-NEXT: }
789 ; CHECK-NEXT: {
790 ; CHECK-NEXT: v27.uh = vlsr(v25.uh,r2)
791 ; CHECK-NEXT: v28.h = vsub(v2.h,v4.h)
792 ; CHECK-NEXT: }
793 ; CHECK-NEXT: {
794 ; CHECK-NEXT: v29.uh = vlsr(v7.uh,r2)
795 ; CHECK-NEXT: }
796 ; CHECK-NEXT: {
797 ; CHECK-NEXT: v1.h = vasl(v28.h,r4)
798 ; CHECK-NEXT: q3 = vsetq(r7)
799 ; CHECK-NEXT: v2 = vmux(q3,v29,v27)
800 ; CHECK-NEXT: }
801 ; CHECK-NEXT: {
802 ; CHECK-NEXT: v2 = vor(v30,v2)
803 ; CHECK-NEXT: }
804 ; CHECK-NEXT: {
805 ; CHECK-NEXT: v31 = vor(v2,v1)
806 ; CHECK-NEXT: }
807 ; CHECK-NEXT: {
808 ; CHECK-NEXT: v0 = vmux(q2,v3,v31)
809 ; CHECK-NEXT: }
810 ; CHECK-NEXT: {
811 ; CHECK-NEXT: jumpr r31
812 ; CHECK-NEXT: if (q3) vmem(r1+#0) = v0
813 ; CHECK-NEXT: }
814 %v0 = load <32 x i16>, ptr %a0, align 128
815 %v1 = sitofp <32 x i16> %v0 to <32 x half>
816 store <32 x half> %v1, ptr %a1, align 128
817 ret void
818 }
821 ; s16 -> f32
822 ; No widening
823 define void @s16f32_0(ptr %a0, ptr %a1) #0 {
824 ; CHECK-LABEL: s16f32_0:
825 ; CHECK: .cfi_startproc
826 ; CHECK-NEXT: // %bb.0:
827 ; CHECK-NEXT: {
828 ; CHECK-NEXT: v0 = vmem(r0+#0)
829 ; CHECK-NEXT: }
830 ; CHECK-NEXT: {
831 ; CHECK-NEXT: r0 = #1
832 ; CHECK-NEXT: r3:2 = combine(##255,#8)
833 ; CHECK-NEXT: v1:0.w = vunpack(v0.h)
834 ; CHECK-NEXT: }
835 ; CHECK-NEXT: {
836 ; CHECK-NEXT: v3 = vsplat(r0)
837 ; CHECK-NEXT: r7 = #512
838 ; CHECK-NEXT: v4.w = vabs(v0.w)
839 ; CHECK-NEXT: v6.w = vabs(v1.w)
840 ; CHECK-NEXT: }
841 ; CHECK-NEXT: {
842 ; CHECK-NEXT: v5 = vsplat(r3)
843 ; CHECK-NEXT: v9 = vsplat(r7)
844 ; CHECK-NEXT: r5 = #159
845 ; CHECK-NEXT: v2 = vxor(v2,v2)
846 ; CHECK-NEXT: }
847 ; CHECK-NEXT: {
848 ; CHECK-NEXT: v13 = vsplat(r5)
849 ; CHECK-NEXT: r6 = ##-2147483648
850 ; CHECK-NEXT: v7.uw = vcl0(v4.uw)
851 ; CHECK-NEXT: }
852 ; CHECK-NEXT: {
853 ; CHECK-NEXT: v10 = vsplat(r6)
854 ; CHECK-NEXT: v8.uw = vcl0(v6.uw)
855 ; CHECK-NEXT: q0 = vcmp.gt(v2.w,v0.w)
856 ; CHECK-NEXT: v7.w = vadd(v7.w,v3.w)
857 ; CHECK-NEXT: }
858 ; CHECK-NEXT: {
859 ; CHECK-NEXT: r4 = #23
860 ; CHECK-NEXT: v8.w = vadd(v8.w,v3.w)
861 ; CHECK-NEXT: v27 = vmux(q0,v10,v2)
862 ; CHECK-NEXT: }
863 ; CHECK-NEXT: {
864 ; CHECK-NEXT: v4.w = vasl(v4.w,v7.w)
865 ; CHECK-NEXT: }
866 ; CHECK-NEXT: {
867 ; CHECK-NEXT: v6.w = vasl(v6.w,v8.w)
868 ; CHECK-NEXT: v11.w = vadd(v4.w,v5.w)
869 ; CHECK-NEXT: v12 = vand(v4,v9)
870 ; CHECK-NEXT: }
871 ; CHECK-NEXT: {
872 ; CHECK-NEXT: v5.w = vadd(v6.w,v5.w)
873 ; CHECK-NEXT: v9 = vand(v6,v9)
874 ; CHECK-NEXT: q1 = vcmp.eq(v12.w,v2.w)
875 ; CHECK-NEXT: q2 = vcmp.gt(v4.uw,v11.uw)
876 ; CHECK-NEXT: }
877 ; CHECK-NEXT: {
878 ; CHECK-NEXT: v22.uw = vlsr(v11.uw,r2)
879 ; CHECK-NEXT: q3 = vcmp.eq(v9.w,v2.w)
880 ; CHECK-NEXT: v23 = vmux(q1,v2,v3)
881 ; CHECK-NEXT: v14 = vmux(q2,v3,v2)
882 ; CHECK-NEXT: }
883 ; CHECK-NEXT: {
884 ; CHECK-NEXT: v24.uw = vlsr(v5.uw,r2)
885 ; CHECK-NEXT: v11.w = vadd(v22.w,v23.w)
886 ; CHECK-NEXT: q2 = vcmp.gt(v6.uw,v5.uw)
887 ; CHECK-NEXT: v25 = vmux(q3,v2,v3)
888 ; CHECK-NEXT: }
889 ; CHECK-NEXT: {
890 ; CHECK-NEXT: v21.uw = vlsr(v4.uw,r2)
891 ; CHECK-NEXT: v5.w = vadd(v24.w,v25.w)
892 ; CHECK-NEXT: v3 = vmux(q2,v3,v2)
893 ; CHECK-NEXT: v7.w = vsub(v14.w,v7.w)
894 ; CHECK-NEXT: }
895 ; CHECK-NEXT: {
896 ; CHECK-NEXT: v6.uw = vlsr(v6.uw,r2)
897 ; CHECK-NEXT: v3.w = vsub(v3.w,v8.w)
898 ; CHECK-NEXT: q3 = vcmp.eq(v21.w,v22.w)
899 ; CHECK-NEXT: v7.w = vadd(v7.w,v13.w)
900 ; CHECK-NEXT: }
901 ; CHECK-NEXT: {
902 ; CHECK-NEXT: v4.uw = vlsr(v22.uw,r0)
903 ; CHECK-NEXT: v3.w = vadd(v3.w,v13.w)
904 ; CHECK-NEXT: q2 = vcmp.eq(v6.w,v24.w)
905 ; CHECK-NEXT: }
906 ; CHECK-NEXT: {
907 ; CHECK-NEXT: v11.uw = vlsr(v11.uw,r0)
908 ; CHECK-NEXT: }
909 ; CHECK-NEXT: {
910 ; CHECK-NEXT: v5.uw = vlsr(v5.uw,r0)
911 ; CHECK-NEXT: v4 = vmux(q3,v11,v4)
912 ; CHECK-NEXT: q3 = vcmp.gt(v2.w,v1.w)
913 ; CHECK-NEXT: }
914 ; CHECK-NEXT: {
915 ; CHECK-NEXT: v26.uw = vlsr(v24.uw,r0)
916 ; CHECK-NEXT: v28 = vmux(q3,v10,v2)
917 ; CHECK-NEXT: v4 = vor(v27,v4)
918 ; CHECK-NEXT: q3 = vcmp.eq(v0.w,v2.w)
919 ; CHECK-NEXT: }
920 ; CHECK-NEXT: {
921 ; CHECK-NEXT: v7.w = vasl(v7.w,r4)
922 ; CHECK-NEXT: v5 = vmux(q2,v5,v26)
923 ; CHECK-NEXT: q2 = vcmp.eq(v1.w,v2.w)
924 ; CHECK-NEXT: }
925 ; CHECK-NEXT: {
926 ; CHECK-NEXT: v3.w = vasl(v3.w,r4)
927 ; CHECK-NEXT: v5 = vor(v28,v5)
928 ; CHECK-NEXT: v29 = vor(v4,v7)
929 ; CHECK-NEXT: }
930 ; CHECK-NEXT: {
931 ; CHECK-NEXT: v3 = vor(v5,v3)
932 ; CHECK-NEXT: v31 = vmux(q3,v2,v29)
933 ; CHECK-NEXT: vmem(r1+#0) = v31.new
934 ; CHECK-NEXT: }
935 ; CHECK-NEXT: {
936 ; CHECK-NEXT: v30 = vmux(q2,v2,v3)
937 ; CHECK-NEXT: jumpr r31
938 ; CHECK-NEXT: vmem(r1+#1) = v30.new
939 ; CHECK-NEXT: }
940 %v0 = load <64 x i16>, ptr %a0, align 128
941 %v1 = sitofp <64 x i16> %v0 to <64 x float>
942 store <64 x float> %v1, ptr %a1, align 128
943 ret void
944 }
946 ; Widen input
947 define void @s16f32_1(ptr %a0, ptr %a1) #0 {
948 ; CHECK-LABEL: s16f32_1:
949 ; CHECK: .cfi_startproc
950 ; CHECK-NEXT: // %bb.0:
951 ; CHECK-NEXT: {
952 ; CHECK-NEXT: v0 = vmem(r0+#0)
953 ; CHECK-NEXT: }
954 ; CHECK-NEXT: {
955 ; CHECK-NEXT: r0 = #1
956 ; CHECK-NEXT: r2 = #255
957 ; CHECK-NEXT: v1:0.w = vunpack(v0.h)
958 ; CHECK-NEXT: }
959 ; CHECK-NEXT: {
960 ; CHECK-NEXT: v3 = vsplat(r0)
961 ; CHECK-NEXT: v4 = vsplat(r2)
962 ; CHECK-NEXT: r3 = #512
963 ; CHECK-NEXT: v2.w = vabs(v0.w)
964 ; CHECK-NEXT: }
965 ; CHECK-NEXT: {
966 ; CHECK-NEXT: v6 = vsplat(r3)
967 ; CHECK-NEXT: r7:6 = combine(##-2147483648,#8)
968 ; CHECK-NEXT: v1 = vxor(v1,v1)
969 ; CHECK-NEXT: }
970 ; CHECK-NEXT: {
971 ; CHECK-NEXT: r4 = #159
972 ; CHECK-NEXT: v5.uw = vcl0(v2.uw)
973 ; CHECK-NEXT: }
974 ; CHECK-NEXT: {
975 ; CHECK-NEXT: v7 = vsplat(r4)
976 ; CHECK-NEXT: v29 = vsplat(r7)
977 ; CHECK-NEXT: q2 = vcmp.gt(v1.w,v0.w)
978 ; CHECK-NEXT: v5.w = vadd(v5.w,v3.w)
979 ; CHECK-NEXT: }
980 ; CHECK-NEXT: {
981 ; CHECK-NEXT: r2 = #23
982 ; CHECK-NEXT: }
983 ; CHECK-NEXT: {
984 ; CHECK-NEXT: v2.w = vasl(v2.w,v5.w)
985 ; CHECK-NEXT: }
986 ; CHECK-NEXT: {
987 ; CHECK-NEXT: v4.w = vadd(v2.w,v4.w)
988 ; CHECK-NEXT: v6 = vand(v2,v6)
989 ; CHECK-NEXT: }
990 ; CHECK-NEXT: {
991 ; CHECK-NEXT: v2.uw = vlsr(v2.uw,r6)
992 ; CHECK-NEXT: q0 = vcmp.eq(v6.w,v1.w)
993 ; CHECK-NEXT: q1 = vcmp.gt(v2.uw,v4.uw)
994 ; CHECK-NEXT: }
995 ; CHECK-NEXT: {
996 ; CHECK-NEXT: v4.uw = vlsr(v4.uw,r6)
997 ; CHECK-NEXT: v6 = vmux(q0,v1,v3)
998 ; CHECK-NEXT: v3 = vmux(q1,v3,v1)
999 ; CHECK-NEXT: }
1000 ; CHECK-NEXT: {
1001 ; CHECK-NEXT: v6.w = vadd(v4.w,v6.w)
1002 ; CHECK-NEXT: v27.w = vsub(v3.w,v5.w)
1003 ; CHECK-NEXT: q3 = vcmp.eq(v2.w,v4.w)
1004 ; CHECK-NEXT: }
1005 ; CHECK-NEXT: {
1006 ; CHECK-NEXT: v28.uw = vlsr(v4.uw,r0)
1007 ; CHECK-NEXT: v2.w = vadd(v27.w,v7.w)
1008 ; CHECK-NEXT: v4 = vmux(q2,v29,v1)
1009 ; CHECK-NEXT: }
1010 ; CHECK-NEXT: {
1011 ; CHECK-NEXT: v30.uw = vlsr(v6.uw,r0)
1012 ; CHECK-NEXT: }
1013 ; CHECK-NEXT: {
1014 ; CHECK-NEXT: v2.w = vasl(v2.w,r2)
1015 ; CHECK-NEXT: v3 = vmux(q3,v30,v28)
1016 ; CHECK-NEXT: q3 = vcmp.eq(v0.w,v1.w)
1017 ; CHECK-NEXT: }
1018 ; CHECK-NEXT: {
1019 ; CHECK-NEXT: v3 = vor(v4,v3)
1020 ; CHECK-NEXT: }
1021 ; CHECK-NEXT: {
1022 ; CHECK-NEXT: v31 = vor(v3,v2)
1023 ; CHECK-NEXT: }
1024 ; CHECK-NEXT: {
1025 ; CHECK-NEXT: v0 = vmux(q3,v1,v31)
1026 ; CHECK-NEXT: jumpr r31
1027 ; CHECK-NEXT: vmem(r1+#0) = v0.new
1028 ; CHECK-NEXT: }
1029 %v0 = load <32 x i16>, ptr %a0, align 128
1030 %v1 = sitofp <32 x i16> %v0 to <32 x float>
1031 store <32 x float> %v1, ptr %a1, align 128
1032 ret void
1033 }
1036 ; s32 -> f16
1037 ; No widening
1038 define void @s32f16_0(ptr %a0, ptr %a1) #0 {
1039 ; CHECK-LABEL: s32f16_0:
1040 ; CHECK: .cfi_startproc
1041 ; CHECK-NEXT: // %bb.0:
1042 ; CHECK-NEXT: {
1043 ; CHECK-NEXT: r3:2 = combine(#8,#1)
1044 ; CHECK-NEXT: r6 = #255
1045 ; CHECK-NEXT: v6.w = vabs(v1.w)
1046 ; CHECK-NEXT: v1.cur = vmem(r0+#0)
1047 ; CHECK-NEXT: }
1048 ; CHECK-NEXT: {
1049 ; CHECK-NEXT: v2 = vsplat(r2)
1050 ; CHECK-NEXT: r4 = #512
1051 ; CHECK-NEXT: v5.w = vabs(v0.w)
1052 ; CHECK-NEXT: v0.cur = vmem(r0+#1)
1053 ; CHECK-NEXT: }
1054 ; CHECK-NEXT: {
1055 ; CHECK-NEXT: v9 = vsplat(r4)
1056 ; CHECK-NEXT: v8 = vsplat(r6)
1057 ; CHECK-NEXT: v3.uw = vcl0(v6.uw)
1058 ; CHECK-NEXT: v20 = vxor(v20,v20)
1059 ; CHECK-NEXT: }
1060 ; CHECK-NEXT: {
1061 ; CHECK-NEXT: r4 = #159
1062 ; CHECK-NEXT: v4.uw = vcl0(v5.uw)
1063 ; CHECK-NEXT: v3.w = vadd(v3.w,v2.w)
1064 ; CHECK-NEXT: }
1065 ; CHECK-NEXT: {
1066 ; CHECK-NEXT: v27 = vsplat(r4)
1067 ; CHECK-NEXT: r5 = ##-2147483648
1068 ; CHECK-NEXT: v7.w = vadd(v4.w,v2.w)
1069 ; CHECK-NEXT: }
1070 ; CHECK-NEXT: {
1071 ; CHECK-NEXT: v13 = vsplat(r5)
1072 ; CHECK-NEXT: v6.w = vasl(v6.w,v3.w)
1073 ; CHECK-NEXT: q0 = vcmp.gt(v20.w,v1.w)
1074 ; CHECK-NEXT: }
1075 ; CHECK-NEXT: {
1076 ; CHECK-NEXT: v5.w = vasl(v5.w,v7.w)
1077 ; CHECK-NEXT: v26 = vmux(q0,v13,v20)
1078 ; CHECK-NEXT: v10.w = vadd(v6.w,v8.w)
1079 ; CHECK-NEXT: v11 = vand(v6,v9)
1080 ; CHECK-NEXT: }
1081 ; CHECK-NEXT: {
1082 ; CHECK-NEXT: v9 = vand(v5,v9)
1083 ; CHECK-NEXT: q3 = vcmp.eq(v11.w,v20.w)
1084 ; CHECK-NEXT: v8.w = vadd(v5.w,v8.w)
1085 ; CHECK-NEXT: q1 = vcmp.gt(v6.uw,v10.uw)
1086 ; CHECK-NEXT: }
1087 ; CHECK-NEXT: {
1088 ; CHECK-NEXT: v21.uw = vlsr(v10.uw,r3)
1089 ; CHECK-NEXT: q2 = vcmp.eq(v9.w,v20.w)
1090 ; CHECK-NEXT: v22 = vmux(q3,v20,v2)
1091 ; CHECK-NEXT: q3 = vcmp.gt(v5.uw,v8.uw)
1092 ; CHECK-NEXT: }
1093 ; CHECK-NEXT: {
1094 ; CHECK-NEXT: v8.uw = vlsr(v8.uw,r3)
1095 ; CHECK-NEXT: v9.w = vadd(v21.w,v22.w)
1096 ; CHECK-NEXT: v24 = vmux(q2,v20,v2)
1097 ; CHECK-NEXT: v23 = vmux(q1,v2,v20)
1098 ; CHECK-NEXT: }
1099 ; CHECK-NEXT: {
1100 ; CHECK-NEXT: v12.uw = vlsr(v6.uw,r3)
1101 ; CHECK-NEXT: v2 = vmux(q3,v2,v20)
1102 ; CHECK-NEXT: v25.w = vadd(v8.w,v24.w)
1103 ; CHECK-NEXT: v3.w = vsub(v23.w,v3.w)
1104 ; CHECK-NEXT: }
1105 ; CHECK-NEXT: {
1106 ; CHECK-NEXT: v5.uw = vlsr(v5.uw,r3)
1107 ; CHECK-NEXT: v2.w = vsub(v2.w,v7.w)
1108 ; CHECK-NEXT: q3 = vcmp.eq(v12.w,v21.w)
1109 ; CHECK-NEXT: v3.w = vadd(v3.w,v27.w)
1110 ; CHECK-NEXT: }
1111 ; CHECK-NEXT: {
1112 ; CHECK-NEXT: r3 = #23
1113 ; CHECK-NEXT: v6.uw = vlsr(v21.uw,r2)
1114 ; CHECK-NEXT: q2 = vcmp.eq(v5.w,v8.w)
1115 ; CHECK-NEXT: v2.w = vadd(v2.w,v27.w)
1116 ; CHECK-NEXT: }
1117 ; CHECK-NEXT: {
1118 ; CHECK-NEXT: v9.uw = vlsr(v9.uw,r2)
1119 ; CHECK-NEXT: }
1120 ; CHECK-NEXT: {
1121 ; CHECK-NEXT: v28.uw = vlsr(v25.uw,r2)
1122 ; CHECK-NEXT: v6 = vmux(q3,v9,v6)
1123 ; CHECK-NEXT: q3 = vcmp.gt(v20.w,v0.w)
1124 ; CHECK-NEXT: }
1125 ; CHECK-NEXT: {
1126 ; CHECK-NEXT: v29.uw = vlsr(v8.uw,r2)
1127 ; CHECK-NEXT: v30 = vmux(q3,v13,v20)
1128 ; CHECK-NEXT: v6 = vor(v26,v6)
1129 ; CHECK-NEXT: q3 = vcmp.eq(v0.w,v20.w)
1130 ; CHECK-NEXT: }
1131 ; CHECK-NEXT: {
1132 ; CHECK-NEXT: v3.w = vasl(v3.w,r3)
1133 ; CHECK-NEXT: v5 = vmux(q2,v28,v29)
1134 ; CHECK-NEXT: q2 = vcmp.eq(v1.w,v20.w)
1135 ; CHECK-NEXT: }
1136 ; CHECK-NEXT: {
1137 ; CHECK-NEXT: v2.w = vasl(v2.w,r3)
1138 ; CHECK-NEXT: v31 = vor(v30,v5)
1139 ; CHECK-NEXT: v3 = vor(v6,v3)
1140 ; CHECK-NEXT: }
1141 ; CHECK-NEXT: {
1142 ; CHECK-NEXT: v1 = vor(v31,v2)
1143 ; CHECK-NEXT: v3 = vmux(q2,v20,v3)
1144 ; CHECK-NEXT: }
1145 ; CHECK-NEXT: {
1146 ; CHECK-NEXT: v0 = vmux(q3,v20,v1)
1147 ; CHECK-NEXT: }
1148 ; CHECK-NEXT: {
1149 ; CHECK-NEXT: v2.qf32 = vadd(v3.sf,v20.sf)
1150 ; CHECK-NEXT: }
1151 ; CHECK-NEXT: {
1152 ; CHECK-NEXT: v3.qf32 = vadd(v0.sf,v20.sf)
1153 ; CHECK-NEXT: }
1154 ; CHECK-NEXT: {
1155 ; CHECK-NEXT: v0.hf = v3:2.qf32
1156 ; CHECK-NEXT: }
1157 ; CHECK-NEXT: {
1158 ; CHECK-NEXT: v0.h = vdeal(v0.h)
1159 ; CHECK-NEXT: jumpr r31
1160 ; CHECK-NEXT: vmem(r1+#0) = v0.new
1161 ; CHECK-NEXT: }
1162 %v0 = load <64 x i32>, ptr %a0, align 128
1163 %v1 = sitofp <64 x i32> %v0 to <64 x half>
1164 store <64 x half> %v1, ptr %a1, align 128
1165 ret void
1166 }
1168 ; Widen result
1169 define void @s32f16_1(ptr %a0, ptr %a1) #0 {
1170 ; CHECK-LABEL: s32f16_1:
1171 ; CHECK: .cfi_startproc
1172 ; CHECK-NEXT: // %bb.0:
1173 ; CHECK-NEXT: {
1174 ; CHECK-NEXT: r6 = #1
1175 ; CHECK-NEXT: v1.w = vabs(v0.w)
1176 ; CHECK-NEXT: v0.cur = vmem(r0+#0)
1177 ; CHECK-NEXT: }
1178 ; CHECK-NEXT: {
1179 ; CHECK-NEXT: v2 = vsplat(r6)
1180 ; CHECK-NEXT: r3:2 = combine(##255,#8)
1181 ; CHECK-NEXT: r4 = #512
1182 ; CHECK-NEXT: }
1183 ; CHECK-NEXT: {
1184 ; CHECK-NEXT: v5 = vsplat(r3)
1185 ; CHECK-NEXT: v6 = vsplat(r4)
1186 ; CHECK-NEXT: v4.uw = vcl0(v1.uw)
1187 ; CHECK-NEXT: v3 = vxor(v3,v3)
1188 ; CHECK-NEXT: }
1189 ; CHECK-NEXT: {
1190 ; CHECK-NEXT: r5 = #159
1191 ; CHECK-NEXT: r4 = ##-2147483648
1192 ; CHECK-NEXT: v4.w = vadd(v4.w,v2.w)
1193 ; CHECK-NEXT: }
1194 ; CHECK-NEXT: {
1195 ; CHECK-NEXT: v28 = vsplat(r5)
1196 ; CHECK-NEXT: v29 = vsplat(r4)
1197 ; CHECK-NEXT: q3 = vcmp.gt(v3.w,v0.w)
1198 ; CHECK-NEXT: }
1199 ; CHECK-NEXT: {
1200 ; CHECK-NEXT: r3 = #23
1201 ; CHECK-NEXT: v1.w = vasl(v1.w,v4.w)
1202 ; CHECK-NEXT: v31 = vmux(q3,v29,v3)
1203 ; CHECK-NEXT: }
1204 ; CHECK-NEXT: {
1205 ; CHECK-NEXT: v5.w = vadd(v1.w,v5.w)
1206 ; CHECK-NEXT: v6 = vand(v1,v6)
1207 ; CHECK-NEXT: }
1208 ; CHECK-NEXT: {
1209 ; CHECK-NEXT: v7.uw = vlsr(v1.uw,r2)
1210 ; CHECK-NEXT: q0 = vcmp.eq(v6.w,v3.w)
1211 ; CHECK-NEXT: q1 = vcmp.gt(v1.uw,v5.uw)
1212 ; CHECK-NEXT: }
1213 ; CHECK-NEXT: {
1214 ; CHECK-NEXT: r2 = #64
1215 ; CHECK-NEXT: v1.uw = vlsr(v5.uw,r2)
1216 ; CHECK-NEXT: v27 = vmux(q0,v3,v2)
1217 ; CHECK-NEXT: v2 = vmux(q1,v2,v3)
1218 ; CHECK-NEXT: }
1219 ; CHECK-NEXT: {
1220 ; CHECK-NEXT: q3 = vsetq(r2)
1221 ; CHECK-NEXT: v5.w = vadd(v1.w,v27.w)
1222 ; CHECK-NEXT: v2.w = vsub(v2.w,v4.w)
1223 ; CHECK-NEXT: q2 = vcmp.eq(v7.w,v1.w)
1224 ; CHECK-NEXT: }
1225 ; CHECK-NEXT: {
1226 ; CHECK-NEXT: v1.uw = vlsr(v1.uw,r6)
1227 ; CHECK-NEXT: v2.w = vadd(v2.w,v28.w)
1228 ; CHECK-NEXT: }
1229 ; CHECK-NEXT: {
1230 ; CHECK-NEXT: v30.uw = vlsr(v5.uw,r6)
1231 ; CHECK-NEXT: }
1232 ; CHECK-NEXT: {
1233 ; CHECK-NEXT: v2.w = vasl(v2.w,r3)
1234 ; CHECK-NEXT: v1 = vmux(q2,v30,v1)
1235 ; CHECK-NEXT: q2 = vcmp.eq(v0.w,v3.w)
1236 ; CHECK-NEXT: }
1237 ; CHECK-NEXT: {
1238 ; CHECK-NEXT: v1 = vor(v31,v1)
1239 ; CHECK-NEXT: }
1240 ; CHECK-NEXT: {
1241 ; CHECK-NEXT: v1.qf32 = vadd(v3.sf,v3.sf)
1242 ; CHECK-NEXT: v0 = vor(v1,v2)
1243 ; CHECK-NEXT: }
1244 ; CHECK-NEXT: {
1245 ; CHECK-NEXT: v0 = vmux(q2,v3,v0)
1246 ; CHECK-NEXT: }
1247 ; CHECK-NEXT: {
1248 ; CHECK-NEXT: v0.qf32 = vadd(v0.sf,v3.sf)
1249 ; CHECK-NEXT: }
1250 ; CHECK-NEXT: {
1251 ; CHECK-NEXT: v0.hf = v1:0.qf32
1252 ; CHECK-NEXT: }
1253 ; CHECK-NEXT: {
1254 ; CHECK-NEXT: v0.h = vdeal(v0.h)
1255 ; CHECK-NEXT: }
1256 ; CHECK-NEXT: {
1257 ; CHECK-NEXT: jumpr r31
1258 ; CHECK-NEXT: if (q3) vmem(r1+#0) = v0
1259 ; CHECK-NEXT: }
1260 %v0 = load <32 x i32>, ptr %a0, align 128
1261 %v1 = sitofp <32 x i32> %v0 to <32 x half>
1262 store <32 x half> %v1, ptr %a1, align 128
1263 ret void
1264 }
1266 ; s32 -> f32
1267 ; No widening
1268 define void @s32f32_0(ptr %a0, ptr %a1) #0 {
1269 ; CHECK-LABEL: s32f32_0:
1270 ; CHECK: .cfi_startproc
1271 ; CHECK-NEXT: // %bb.0:
1272 ; CHECK-NEXT: {
1273 ; CHECK-NEXT: r0 = #1
1274 ; CHECK-NEXT: r2 = #255
1275 ; CHECK-NEXT: v1.w = vabs(v0.w)
1276 ; CHECK-NEXT: v0.cur = vmem(r0+#0)
1277 ; CHECK-NEXT: }
1278 ; CHECK-NEXT: {
1279 ; CHECK-NEXT: v3 = vsplat(r0)
1280 ; CHECK-NEXT: v5 = vsplat(r2)
1281 ; CHECK-NEXT: r3 = #512
1282 ; CHECK-NEXT: v2 = vxor(v2,v2)
1283 ; CHECK-NEXT: }
1284 ; CHECK-NEXT: {
1285 ; CHECK-NEXT: v6 = vsplat(r3)
1286 ; CHECK-NEXT: r7:6 = combine(##-2147483648,#8)
1287 ; CHECK-NEXT: v4.uw = vcl0(v1.uw)
1288 ; CHECK-NEXT: }
1289 ; CHECK-NEXT: {
1290 ; CHECK-NEXT: r4 = #159
1291 ; CHECK-NEXT: v4.w = vadd(v4.w,v3.w)
1292 ; CHECK-NEXT: }
1293 ; CHECK-NEXT: {
1294 ; CHECK-NEXT: v7 = vsplat(r4)
1295 ; CHECK-NEXT: v29 = vsplat(r7)
1296 ; CHECK-NEXT: r2 = #23
1297 ; CHECK-NEXT: q2 = vcmp.gt(v2.w,v0.w)
1298 ; CHECK-NEXT: }
1299 ; CHECK-NEXT: {
1300 ; CHECK-NEXT: v1.w = vasl(v1.w,v4.w)
1301 ; CHECK-NEXT: }
1302 ; CHECK-NEXT: {
1303 ; CHECK-NEXT: v5.w = vadd(v1.w,v5.w)
1304 ; CHECK-NEXT: v6 = vand(v1,v6)
1305 ; CHECK-NEXT: }
1306 ; CHECK-NEXT: {
1307 ; CHECK-NEXT: v1.uw = vlsr(v1.uw,r6)
1308 ; CHECK-NEXT: q0 = vcmp.eq(v6.w,v2.w)
1309 ; CHECK-NEXT: q1 = vcmp.gt(v1.uw,v5.uw)
1310 ; CHECK-NEXT: }
1311 ; CHECK-NEXT: {
1312 ; CHECK-NEXT: v5.uw = vlsr(v5.uw,r6)
1313 ; CHECK-NEXT: v6 = vmux(q0,v2,v3)
1314 ; CHECK-NEXT: v3 = vmux(q1,v3,v2)
1315 ; CHECK-NEXT: }
1316 ; CHECK-NEXT: {
1317 ; CHECK-NEXT: v6.w = vadd(v5.w,v6.w)
1318 ; CHECK-NEXT: v27.w = vsub(v3.w,v4.w)
1319 ; CHECK-NEXT: q3 = vcmp.eq(v1.w,v5.w)
1320 ; CHECK-NEXT: v4 = vmux(q2,v29,v2)
1321 ; CHECK-NEXT: }
1322 ; CHECK-NEXT: {
1323 ; CHECK-NEXT: v28.uw = vlsr(v5.uw,r0)
1324 ; CHECK-NEXT: v1.w = vadd(v27.w,v7.w)
1325 ; CHECK-NEXT: }
1326 ; CHECK-NEXT: {
1327 ; CHECK-NEXT: v30.uw = vlsr(v6.uw,r0)
1328 ; CHECK-NEXT: }
1329 ; CHECK-NEXT: {
1330 ; CHECK-NEXT: v1.w = vasl(v1.w,r2)
1331 ; CHECK-NEXT: v3 = vmux(q3,v30,v28)
1332 ; CHECK-NEXT: q3 = vcmp.eq(v0.w,v2.w)
1333 ; CHECK-NEXT: }
1334 ; CHECK-NEXT: {
1335 ; CHECK-NEXT: v3 = vor(v4,v3)
1336 ; CHECK-NEXT: }
1337 ; CHECK-NEXT: {
1338 ; CHECK-NEXT: v31 = vor(v3,v1)
1339 ; CHECK-NEXT: }
1340 ; CHECK-NEXT: {
1341 ; CHECK-NEXT: v0 = vmux(q3,v2,v31)
1342 ; CHECK-NEXT: jumpr r31
1343 ; CHECK-NEXT: vmem(r1+#0) = v0.new
1344 ; CHECK-NEXT: }
1345 %v0 = load <32 x i32>, ptr %a0, align 128
1346 %v1 = sitofp <32 x i32> %v0 to <32 x float>
1347 store <32 x float> %v1, ptr %a1, align 128
1348 ret void
1349 }
1351 ; Widen input and result
1352 define void @s32f32_1(ptr %a0, ptr %a1) #0 {
1353 ; CHECK-LABEL: s32f32_1:
1354 ; CHECK: .cfi_startproc
1355 ; CHECK-NEXT: // %bb.0:
1356 ; CHECK-NEXT: {
1357 ; CHECK-NEXT: r0 = #1
1358 ; CHECK-NEXT: r2 = #255
1359 ; CHECK-NEXT: v1.w = vabs(v0.w)
1360 ; CHECK-NEXT: v0.cur = vmem(r0+#0)
1361 ; CHECK-NEXT: }
1362 ; CHECK-NEXT: {
1363 ; CHECK-NEXT: v2 = vsplat(r0)
1364 ; CHECK-NEXT: v5 = vsplat(r2)
1365 ; CHECK-NEXT: r3 = #512
1366 ; CHECK-NEXT: v3 = vxor(v3,v3)
1367 ; CHECK-NEXT: }
1368 ; CHECK-NEXT: {
1369 ; CHECK-NEXT: v6 = vsplat(r3)
1370 ; CHECK-NEXT: r7:6 = combine(##-2147483648,#8)
1371 ; CHECK-NEXT: v4.uw = vcl0(v1.uw)
1372 ; CHECK-NEXT: }
1373 ; CHECK-NEXT: {
1374 ; CHECK-NEXT: r4 = #159
1375 ; CHECK-NEXT: v4.w = vadd(v4.w,v2.w)
1376 ; CHECK-NEXT: }
1377 ; CHECK-NEXT: {
1378 ; CHECK-NEXT: v7 = vsplat(r4)
1379 ; CHECK-NEXT: v29 = vsplat(r7)
1380 ; CHECK-NEXT: r3 = #23
1381 ; CHECK-NEXT: q3 = vcmp.gt(v3.w,v0.w)
1382 ; CHECK-NEXT: }
1383 ; CHECK-NEXT: {
1384 ; CHECK-NEXT: r2 = #64
1385 ; CHECK-NEXT: v1.w = vasl(v1.w,v4.w)
1386 ; CHECK-NEXT: }
1387 ; CHECK-NEXT: {
1388 ; CHECK-NEXT: v5.w = vadd(v1.w,v5.w)
1389 ; CHECK-NEXT: v6 = vand(v1,v6)
1390 ; CHECK-NEXT: }
1391 ; CHECK-NEXT: {
1392 ; CHECK-NEXT: v1.uw = vlsr(v1.uw,r6)
1393 ; CHECK-NEXT: q0 = vcmp.eq(v6.w,v3.w)
1394 ; CHECK-NEXT: q1 = vcmp.gt(v1.uw,v5.uw)
1395 ; CHECK-NEXT: }
1396 ; CHECK-NEXT: {
1397 ; CHECK-NEXT: v5.uw = vlsr(v5.uw,r6)
1398 ; CHECK-NEXT: v6 = vmux(q0,v3,v2)
1399 ; CHECK-NEXT: v2 = vmux(q1,v2,v3)
1400 ; CHECK-NEXT: }
1401 ; CHECK-NEXT: {
1402 ; CHECK-NEXT: v6.w = vadd(v5.w,v6.w)
1403 ; CHECK-NEXT: v27.w = vsub(v2.w,v4.w)
1404 ; CHECK-NEXT: q2 = vcmp.eq(v1.w,v5.w)
1405 ; CHECK-NEXT: v4 = vmux(q3,v29,v3)
1406 ; CHECK-NEXT: }
1407 ; CHECK-NEXT: {
1408 ; CHECK-NEXT: v28.uw = vlsr(v5.uw,r0)
1409 ; CHECK-NEXT: q3 = vsetq(r2)
1410 ; CHECK-NEXT: v1.w = vadd(v27.w,v7.w)
1411 ; CHECK-NEXT: }
1412 ; CHECK-NEXT: {
1413 ; CHECK-NEXT: v30.uw = vlsr(v6.uw,r0)
1414 ; CHECK-NEXT: }
1415 ; CHECK-NEXT: {
1416 ; CHECK-NEXT: v1.w = vasl(v1.w,r3)
1417 ; CHECK-NEXT: v2 = vmux(q2,v30,v28)
1418 ; CHECK-NEXT: q2 = vcmp.eq(v0.w,v3.w)
1419 ; CHECK-NEXT: }
1420 ; CHECK-NEXT: {
1421 ; CHECK-NEXT: v2 = vor(v4,v2)
1422 ; CHECK-NEXT: }
1423 ; CHECK-NEXT: {
1424 ; CHECK-NEXT: v31 = vor(v2,v1)
1425 ; CHECK-NEXT: }
1426 ; CHECK-NEXT: {
1427 ; CHECK-NEXT: v0 = vmux(q2,v3,v31)
1428 ; CHECK-NEXT: }
1429 ; CHECK-NEXT: {
1430 ; CHECK-NEXT: jumpr r31
1431 ; CHECK-NEXT: if (q3) vmem(r1+#0) = v0
1432 ; CHECK-NEXT: }
1433 %v0 = load <16 x i32>, ptr %a0, align 128
1434 %v1 = sitofp <16 x i32> %v0 to <16 x float>
1435 store <16 x float> %v1, ptr %a1, align 128
1436 ret void
1437 }
1440 ; u8 -> f16
1441 ; No widening
1442 define void @u8f16_0(ptr %a0, ptr %a1) #0 {
1443 ; CHECK-LABEL: u8f16_0:
1444 ; CHECK: .cfi_startproc
1445 ; CHECK-NEXT: // %bb.0:
1446 ; CHECK-NEXT: {
1447 ; CHECK-NEXT: v0 = vmem(r0+#0)
1448 ; CHECK-NEXT: }
1449 ; CHECK-NEXT: {
1450 ; CHECK-NEXT: r6 = #1
1451 ; CHECK-NEXT: r3:2 = combine(#31,#5)
1452 ; CHECK-NEXT: v1:0.uh = vunpack(v0.ub)
1453 ; CHECK-NEXT: }
1454 ; CHECK-NEXT: {
1455 ; CHECK-NEXT: v3.h = vsplat(r6)
1456 ; CHECK-NEXT: v4.h = vsplat(r3)
1457 ; CHECK-NEXT: r5 = #64
1458 ; CHECK-NEXT: v2 = vxor(v2,v2)
1459 ; CHECK-NEXT: }
1460 ; CHECK-NEXT: {
1461 ; CHECK-NEXT: v6.h = vsplat(r5)
1462 ; CHECK-NEXT: r4 = #10
1463 ; CHECK-NEXT: v5.uh = vcl0(v0.uh)
1464 ; CHECK-NEXT: }
1465 ; CHECK-NEXT: {
1466 ; CHECK-NEXT: v7.uh = vcl0(v1.uh)
1467 ; CHECK-NEXT: v5.h = vadd(v5.h,v3.h)
1468 ; CHECK-NEXT: }
1469 ; CHECK-NEXT: {
1470 ; CHECK-NEXT: v7.h = vadd(v7.h,v3.h)
1471 ; CHECK-NEXT: }
1472 ; CHECK-NEXT: {
1473 ; CHECK-NEXT: v8.h = vasl(v0.h,v5.h)
1474 ; CHECK-NEXT: }
1475 ; CHECK-NEXT: {
1476 ; CHECK-NEXT: v11.h = vasl(v1.h,v7.h)
1477 ; CHECK-NEXT: v10 = vand(v8,v6)
1478 ; CHECK-NEXT: v9.h = vadd(v8.h,v4.h)
1479 ; CHECK-NEXT: }
1480 ; CHECK-NEXT: {
1481 ; CHECK-NEXT: v22.h = vadd(v11.h,v4.h)
1482 ; CHECK-NEXT: v6 = vand(v11,v6)
1483 ; CHECK-NEXT: q0 = vcmp.gt(v8.uh,v9.uh)
1484 ; CHECK-NEXT: q1 = vcmp.eq(v10.h,v2.h)
1485 ; CHECK-NEXT: }
1486 ; CHECK-NEXT: {
1487 ; CHECK-NEXT: v21.uh = vlsr(v8.uh,r2)
1488 ; CHECK-NEXT: q2 = vcmp.eq(v6.h,v2.h)
1489 ; CHECK-NEXT: q3 = vcmp.gt(v11.uh,v22.uh)
1490 ; CHECK-NEXT: v12 = vmux(q1,v2,v3)
1491 ; CHECK-NEXT: }
1492 ; CHECK-NEXT: {
1493 ; CHECK-NEXT: v9.uh = vlsr(v9.uh,r2)
1494 ; CHECK-NEXT: v13 = vmux(q2,v2,v3)
1495 ; CHECK-NEXT: v25 = vmux(q0,v3,v2)
1496 ; CHECK-NEXT: v3 = vmux(q3,v3,v2)
1497 ; CHECK-NEXT: }
1498 ; CHECK-NEXT: {
1499 ; CHECK-NEXT: v8.uh = vlsr(v22.uh,r2)
1500 ; CHECK-NEXT: v24.h = vadd(v9.h,v12.h)
1501 ; CHECK-NEXT: v3.h = vadd(v3.h,v4.h)
1502 ; CHECK-NEXT: v12.h = vadd(v25.h,v4.h)
1503 ; CHECK-NEXT: }
1504 ; CHECK-NEXT: {
1505 ; CHECK-NEXT: v23.uh = vlsr(v11.uh,r2)
1506 ; CHECK-NEXT: v13.h = vadd(v8.h,v13.h)
1507 ; CHECK-NEXT: v5.h = vsub(v12.h,v5.h)
1508 ; CHECK-NEXT: v3.h = vsub(v3.h,v7.h)
1509 ; CHECK-NEXT: }
1510 ; CHECK-NEXT: {
1511 ; CHECK-NEXT: v14.uh = vlsr(v9.uh,r6)
1512 ; CHECK-NEXT: q2 = vcmp.eq(v21.h,v9.h)
1513 ; CHECK-NEXT: q3 = vcmp.eq(v23.h,v8.h)
1514 ; CHECK-NEXT: }
1515 ; CHECK-NEXT: {
1516 ; CHECK-NEXT: v26.uh = vlsr(v24.uh,r6)
1517 ; CHECK-NEXT: }
1518 ; CHECK-NEXT: {
1519 ; CHECK-NEXT: v27.uh = vlsr(v13.uh,r6)
1520 ; CHECK-NEXT: v4 = vmux(q2,v26,v14)
1521 ; CHECK-NEXT: q2 = vcmp.eq(v1.h,v2.h)
1522 ; CHECK-NEXT: }
1523 ; CHECK-NEXT: {
1524 ; CHECK-NEXT: v28.uh = vlsr(v8.uh,r6)
1525 ; CHECK-NEXT: }
1526 ; CHECK-NEXT: {
1527 ; CHECK-NEXT: v5.h = vasl(v5.h,r4)
1528 ; CHECK-NEXT: v6 = vmux(q3,v27,v28)
1529 ; CHECK-NEXT: q3 = vcmp.eq(v0.h,v2.h)
1530 ; CHECK-NEXT: }
1531 ; CHECK-NEXT: {
1532 ; CHECK-NEXT: v3.h = vasl(v3.h,r4)
1533 ; CHECK-NEXT: v29 = vor(v4,v5)
1534 ; CHECK-NEXT: }
1535 ; CHECK-NEXT: {
1536 ; CHECK-NEXT: v3 = vor(v6,v3)
1537 ; CHECK-NEXT: v31 = vmux(q3,v2,v29)
1538 ; CHECK-NEXT: vmem(r1+#0) = v31.new
1539 ; CHECK-NEXT: }
1540 ; CHECK-NEXT: {
1541 ; CHECK-NEXT: v30 = vmux(q2,v2,v3)
1542 ; CHECK-NEXT: jumpr r31
1543 ; CHECK-NEXT: vmem(r1+#1) = v30.new
1544 ; CHECK-NEXT: }
1545 %v0 = load <128 x i8>, ptr %a0, align 128
1546 %v1 = uitofp <128 x i8> %v0 to <128 x half>
1547 store <128 x half> %v1, ptr %a1, align 128
1548 ret void
1549 }
1551 ; Widen input
1552 define void @u8f16_1(ptr %a0, ptr %a1) #0 {
1553 ; CHECK-LABEL: u8f16_1:
1554 ; CHECK: .cfi_startproc
1555 ; CHECK-NEXT: // %bb.0:
1556 ; CHECK-NEXT: {
1557 ; CHECK-NEXT: v0 = vmem(r0+#0)
1558 ; CHECK-NEXT: }
1559 ; CHECK-NEXT: {
1560 ; CHECK-NEXT: r6 = #1
1561 ; CHECK-NEXT: r3:2 = combine(#64,#31)
1562 ; CHECK-NEXT: v1:0.uh = vunpack(v0.ub)
1563 ; CHECK-NEXT: }
1564 ; CHECK-NEXT: {
1565 ; CHECK-NEXT: v1.h = vsplat(r6)
1566 ; CHECK-NEXT: v4.h = vsplat(r2)
1567 ; CHECK-NEXT: r5 = #5
1568 ; CHECK-NEXT: v2 = vxor(v2,v2)
1569 ; CHECK-NEXT: }
1570 ; CHECK-NEXT: {
1571 ; CHECK-NEXT: v5.h = vsplat(r3)
1572 ; CHECK-NEXT: r4 = #10
1573 ; CHECK-NEXT: v3.uh = vcl0(v0.uh)
1574 ; CHECK-NEXT: q3 = vcmp.eq(v0.h,v2.h)
1575 ; CHECK-NEXT: }
1576 ; CHECK-NEXT: {
1577 ; CHECK-NEXT: v3.h = vadd(v3.h,v1.h)
1578 ; CHECK-NEXT: }
1579 ; CHECK-NEXT: {
1580 ; CHECK-NEXT: v6.h = vasl(v0.h,v3.h)
1581 ; CHECK-NEXT: }
1582 ; CHECK-NEXT: {
1583 ; CHECK-NEXT: v7.h = vadd(v6.h,v4.h)
1584 ; CHECK-NEXT: v5 = vand(v6,v5)
1585 ; CHECK-NEXT: }
1586 ; CHECK-NEXT: {
1587 ; CHECK-NEXT: v6.uh = vlsr(v6.uh,r5)
1588 ; CHECK-NEXT: q0 = vcmp.gt(v6.uh,v7.uh)
1589 ; CHECK-NEXT: q1 = vcmp.eq(v5.h,v2.h)
1590 ; CHECK-NEXT: }
1591 ; CHECK-NEXT: {
1592 ; CHECK-NEXT: v26.uh = vlsr(v7.uh,r5)
1593 ; CHECK-NEXT: v27 = vmux(q1,v2,v1)
1594 ; CHECK-NEXT: v1 = vmux(q0,v1,v2)
1595 ; CHECK-NEXT: }
1596 ; CHECK-NEXT: {
1597 ; CHECK-NEXT: v1.h = vadd(v1.h,v4.h)
1598 ; CHECK-NEXT: v28.h = vadd(v26.h,v27.h)
1599 ; CHECK-NEXT: q2 = vcmp.eq(v6.h,v26.h)
1600 ; CHECK-NEXT: }
1601 ; CHECK-NEXT: {
1602 ; CHECK-NEXT: v29.uh = vlsr(v26.uh,r6)
1603 ; CHECK-NEXT: v1.h = vsub(v1.h,v3.h)
1604 ; CHECK-NEXT: }
1605 ; CHECK-NEXT: {
1606 ; CHECK-NEXT: v30.uh = vlsr(v28.uh,r6)
1607 ; CHECK-NEXT: }
1608 ; CHECK-NEXT: {
1609 ; CHECK-NEXT: v1.h = vasl(v1.h,r4)
1610 ; CHECK-NEXT: v3 = vmux(q2,v30,v29)
1611 ; CHECK-NEXT: }
1612 ; CHECK-NEXT: {
1613 ; CHECK-NEXT: v31 = vor(v3,v1)
1614 ; CHECK-NEXT: }
1615 ; CHECK-NEXT: {
1616 ; CHECK-NEXT: v0 = vmux(q3,v2,v31)
1617 ; CHECK-NEXT: jumpr r31
1618 ; CHECK-NEXT: vmem(r1+#0) = v0.new
1619 ; CHECK-NEXT: }
1620 %v0 = load <64 x i8>, ptr %a0, align 128
1621 %v1 = uitofp <64 x i8> %v0 to <64 x half>
1622 store <64 x half> %v1, ptr %a1, align 128
1623 ret void
1624 }
1627 ; u8 -> f32
1628 ; No widening
1629 define void @u8f32_0(ptr %a0, ptr %a1) #0 {
1630 ; CHECK-LABEL: u8f32_0:
1631 ; CHECK: .cfi_startproc
1632 ; CHECK-NEXT: // %bb.0:
1633 ; CHECK-NEXT: {
1634 ; CHECK-NEXT: r7 = #64
1635 ; CHECK-NEXT: r0 = #1
1636 ; CHECK-NEXT: r6 = #512
1637 ; CHECK-NEXT: v0 = vmem(r0+#0)
1638 ; CHECK-NEXT: }
1639 ; CHECK-NEXT: {
1640 ; CHECK-NEXT: v4 = vsplat(r0)
1641 ; CHECK-NEXT: r3:2 = combine(##255,#8)
1642 ; CHECK-NEXT: v1 = valign(v0,v0,r7)
1643 ; CHECK-NEXT: }
1644 ; CHECK-NEXT: {
1645 ; CHECK-NEXT: v15 = vsplat(r6)
1646 ; CHECK-NEXT: v6 = vsplat(r3)
1647 ; CHECK-NEXT: r5 = #159
1648 ; CHECK-NEXT: v3:2.uh = vunpack(v0.ub)
1649 ; CHECK-NEXT: }
1650 ; CHECK-NEXT: {
1651 ; CHECK-NEXT: r4 = #23
1652 ; CHECK-NEXT: v31:30.uh = vunpack(v1.ub)
1653 ; CHECK-NEXT: }
1654 ; CHECK-NEXT: {
1655 ; CHECK-NEXT: v3:2.uw = vunpack(v2.uh)
1656 ; CHECK-NEXT: }
1657 ; CHECK-NEXT: {
1658 ; CHECK-NEXT: v1:0.uw = vunpack(v30.uh)
1659 ; CHECK-NEXT: }
1660 ; CHECK-NEXT: {
1661 ; CHECK-NEXT: v5.uw = vcl0(v2.uw)
1662 ; CHECK-NEXT: }
1663 ; CHECK-NEXT: {
1664 ; CHECK-NEXT: v7.uw = vcl0(v0.uw)
1665 ; CHECK-NEXT: v5.w = vadd(v5.w,v4.w)
1666 ; CHECK-NEXT: }
1667 ; CHECK-NEXT: {
1668 ; CHECK-NEXT: v8.uw = vcl0(v3.uw)
1669 ; CHECK-NEXT: v11.w = vadd(v7.w,v4.w)
1670 ; CHECK-NEXT: v7 = vxor(v7,v7)
1671 ; CHECK-NEXT: }
1672 ; CHECK-NEXT: {
1673 ; CHECK-NEXT: v9.uw = vcl0(v1.uw)
1674 ; CHECK-NEXT: v10.w = vadd(v8.w,v4.w)
1675 ; CHECK-NEXT: }
1676 ; CHECK-NEXT: {
1677 ; CHECK-NEXT: v9 = vsplat(r5)
1678 ; CHECK-NEXT: v14.w = vasl(v0.w,v11.w)
1679 ; CHECK-NEXT: v8.w = vadd(v9.w,v4.w)
1680 ; CHECK-NEXT: }
1681 ; CHECK-NEXT: {
1682 ; CHECK-NEXT: v12.w = vasl(v2.w,v5.w)
1683 ; CHECK-NEXT: v24 = vand(v14,v15)
1684 ; CHECK-NEXT: v20.w = vadd(v14.w,v6.w)
1685 ; CHECK-NEXT: }
1686 ; CHECK-NEXT: {
1687 ; CHECK-NEXT: v13.w = vasl(v3.w,v10.w)
1688 ; CHECK-NEXT: v19 = vand(v12,v15)
1689 ; CHECK-NEXT: q3 = vcmp.eq(v24.w,v7.w)
1690 ; CHECK-NEXT: v18.w = vadd(v12.w,v6.w)
1691 ; CHECK-NEXT: }
1692 ; CHECK-NEXT: {
1693 ; CHECK-NEXT: v16.w = vasl(v1.w,v8.w)
1694 ; CHECK-NEXT: v23 = vand(v13,v15)
1695 ; CHECK-NEXT: v22.w = vadd(v13.w,v6.w)
1696 ; CHECK-NEXT: q0 = vcmp.gt(v14.uw,v20.uw)
1697 ; CHECK-NEXT: }
1698 ; CHECK-NEXT: {
1699 ; CHECK-NEXT: v6.w = vadd(v16.w,v6.w)
1700 ; CHECK-NEXT: v15 = vand(v16,v15)
1701 ; CHECK-NEXT: v30 = vmux(q3,v7,v4)
1702 ; CHECK-NEXT: q2 = vcmp.eq(v19.w,v7.w)
1703 ; CHECK-NEXT: }
1704 ; CHECK-NEXT: {
1705 ; CHECK-NEXT: v21.uw = vlsr(v14.uw,r2)
1706 ; CHECK-NEXT: q3 = vcmp.eq(v15.w,v7.w)
1707 ; CHECK-NEXT: v28 = vmux(q0,v4,v7)
1708 ; CHECK-NEXT: q1 = vcmp.eq(v23.w,v7.w)
1709 ; CHECK-NEXT: }
1710 ; CHECK-NEXT: {
1711 ; CHECK-NEXT: v14.uw = vlsr(v20.uw,r2)
1712 ; CHECK-NEXT: v26 = vmux(q3,v7,v4)
1713 ; CHECK-NEXT: v11.w = vsub(v28.w,v11.w)
1714 ; CHECK-NEXT: q3 = vcmp.gt(v13.uw,v22.uw)
1715 ; CHECK-NEXT: }
1716 ; CHECK-NEXT: {
1717 ; CHECK-NEXT: v15.uw = vlsr(v6.uw,r2)
1718 ; CHECK-NEXT: v20.w = vadd(v14.w,v30.w)
1719 ; CHECK-NEXT: v30 = vmux(q1,v7,v4)
1720 ; CHECK-NEXT: v31 = vmux(q2,v7,v4)
1721 ; CHECK-NEXT: }
1722 ; CHECK-NEXT: {
1723 ; CHECK-NEXT: v19.uw = vlsr(v18.uw,r2)
1724 ; CHECK-NEXT: v29.w = vadd(v15.w,v26.w)
1725 ; CHECK-NEXT: q1 = vcmp.gt(v12.uw,v18.uw)
1726 ; CHECK-NEXT: v11.w = vadd(v11.w,v9.w)
1727 ; CHECK-NEXT: }
1728 ; CHECK-NEXT: {
1729 ; CHECK-NEXT: v28.uw = vlsr(v22.uw,r2)
1730 ; CHECK-NEXT: v23.w = vadd(v19.w,v31.w)
1731 ; CHECK-NEXT: v22 = vmux(q3,v4,v7)
1732 ; CHECK-NEXT: q3 = vcmp.gt(v16.uw,v6.uw)
1733 ; CHECK-NEXT: }
1734 ; CHECK-NEXT: {
1735 ; CHECK-NEXT: v24.uw = vlsr(v29.uw,r0)
1736 ; CHECK-NEXT: v31.w = vadd(v28.w,v30.w)
1737 ; CHECK-NEXT: v30 = vmux(q1,v4,v7)
1738 ; CHECK-NEXT: v4 = vmux(q3,v4,v7)
1739 ; CHECK-NEXT: }
1740 ; CHECK-NEXT: {
1741 ; CHECK-NEXT: v17.uw = vlsr(v12.uw,r2)
1742 ; CHECK-NEXT: v5.w = vsub(v30.w,v5.w)
1743 ; CHECK-NEXT: v29.w = vsub(v22.w,v10.w)
1744 ; CHECK-NEXT: v4.w = vsub(v4.w,v8.w)
1745 ; CHECK-NEXT: }
1746 ; CHECK-NEXT: {
1747 ; CHECK-NEXT: v13.uw = vlsr(v13.uw,r2)
1748 ; CHECK-NEXT: v6.w = vadd(v29.w,v9.w)
1749 ; CHECK-NEXT: v5.w = vadd(v5.w,v9.w)
1750 ; CHECK-NEXT: q0 = vcmp.eq(v21.w,v14.w)
1751 ; CHECK-NEXT: }
1752 ; CHECK-NEXT: {
1753 ; CHECK-NEXT: v25.uw = vlsr(v16.uw,r2)
1754 ; CHECK-NEXT: q2 = vcmp.eq(v17.w,v19.w)
1755 ; CHECK-NEXT: q3 = vcmp.eq(v13.w,v28.w)
1756 ; CHECK-NEXT: v4.w = vadd(v4.w,v9.w)
1757 ; CHECK-NEXT: }
1758 ; CHECK-NEXT: {
1759 ; CHECK-NEXT: v21.uw = vlsr(v23.uw,r0)
1760 ; CHECK-NEXT: q1 = vcmp.eq(v25.w,v15.w)
1761 ; CHECK-NEXT: }
1762 ; CHECK-NEXT: {
1763 ; CHECK-NEXT: v23.uw = vlsr(v19.uw,r0)
1764 ; CHECK-NEXT: }
1765 ; CHECK-NEXT: {
1766 ; CHECK-NEXT: v31.uw = vlsr(v31.uw,r0)
1767 ; CHECK-NEXT: v23 = vmux(q2,v21,v23)
1768 ; CHECK-NEXT: q2 = vcmp.eq(v3.w,v7.w)
1769 ; CHECK-NEXT: }
1770 ; CHECK-NEXT: {
1771 ; CHECK-NEXT: v16.uw = vlsr(v28.uw,r0)
1772 ; CHECK-NEXT: }
1773 ; CHECK-NEXT: {
1774 ; CHECK-NEXT: v26.uw = vlsr(v15.uw,r0)
1775 ; CHECK-NEXT: v8 = vmux(q3,v31,v16)
1776 ; CHECK-NEXT: q3 = vcmp.eq(v2.w,v7.w)
1777 ; CHECK-NEXT: }
1778 ; CHECK-NEXT: {
1779 ; CHECK-NEXT: v6.w = vasl(v6.w,r4)
1780 ; CHECK-NEXT: v22 = vmux(q1,v24,v26)
1781 ; CHECK-NEXT: }
1782 ; CHECK-NEXT: {
1783 ; CHECK-NEXT: v5.w = vasl(v5.w,r4)
1784 ; CHECK-NEXT: v6 = vor(v8,v6)
1785 ; CHECK-NEXT: }
1786 ; CHECK-NEXT: {
1787 ; CHECK-NEXT: v27.uw = vlsr(v14.uw,r0)
1788 ; CHECK-NEXT: v25 = vor(v23,v5)
1789 ; CHECK-NEXT: v26 = vmux(q2,v7,v6)
1790 ; CHECK-NEXT: vmem(r1+#1) = v26.new
1791 ; CHECK-NEXT: }
1792 ; CHECK-NEXT: {
1793 ; CHECK-NEXT: v20.uw = vlsr(v20.uw,r0)
1794 ; CHECK-NEXT: v28 = vmux(q3,v7,v25)
1795 ; CHECK-NEXT: q2 = vcmp.eq(v1.w,v7.w)
1796 ; CHECK-NEXT: vmem(r1+#0) = v28.new
1797 ; CHECK-NEXT: }
1798 ; CHECK-NEXT: {
1799 ; CHECK-NEXT: v11.w = vasl(v11.w,r4)
1800 ; CHECK-NEXT: v20 = vmux(q0,v20,v27)
1801 ; CHECK-NEXT: q3 = vcmp.eq(v0.w,v7.w)
1802 ; CHECK-NEXT: }
1803 ; CHECK-NEXT: {
1804 ; CHECK-NEXT: v24.w = vasl(v4.w,r4)
1805 ; CHECK-NEXT: v29 = vor(v20,v11)
1806 ; CHECK-NEXT: }
1807 ; CHECK-NEXT: {
1808 ; CHECK-NEXT: v27 = vor(v22,v24)
1809 ; CHECK-NEXT: v31 = vmux(q3,v7,v29)
1810 ; CHECK-NEXT: vmem(r1+#2) = v31.new
1811 ; CHECK-NEXT: }
1812 ; CHECK-NEXT: {
1813 ; CHECK-NEXT: v30 = vmux(q2,v7,v27)
1814 ; CHECK-NEXT: jumpr r31
1815 ; CHECK-NEXT: vmem(r1+#3) = v30.new
1816 ; CHECK-NEXT: }
1817 %v0 = load <128 x i8>, ptr %a0, align 128
1818 %v1 = uitofp <128 x i8> %v0 to <128 x float>
1819 store <128 x float> %v1, ptr %a1, align 128
1820 ret void
1821 }
1823 ; Widen input #1
1824 define void @u8f32_1(ptr %a0, ptr %a1) #0 {
1825 ; CHECK-LABEL: u8f32_1:
1826 ; CHECK: .cfi_startproc
1827 ; CHECK-NEXT: // %bb.0:
1828 ; CHECK-NEXT: {
1829 ; CHECK-NEXT: r7 = #1
1830 ; CHECK-NEXT: r6 = #512
1831 ; CHECK-NEXT: v3:2.uh = vunpack(v0.ub)
1832 ; CHECK-NEXT: v0.cur = vmem(r0+#0)
1833 ; CHECK-NEXT: }
1834 ; CHECK-NEXT: {
1835 ; CHECK-NEXT: v1 = vsplat(r7)
1836 ; CHECK-NEXT: v8 = vsplat(r6)
1837 ; CHECK-NEXT: r3:2 = combine(##255,#8)
1838 ; CHECK-NEXT: }
1839 ; CHECK-NEXT: {
1840 ; CHECK-NEXT: v6 = vsplat(r3)
1841 ; CHECK-NEXT: r5 = #159
1842 ; CHECK-NEXT: v3:2.uw = vunpack(v2.uh)
1843 ; CHECK-NEXT: v21 = vxor(v21,v21)
1844 ; CHECK-NEXT: }
1845 ; CHECK-NEXT: {
1846 ; CHECK-NEXT: v13 = vsplat(r5)
1847 ; CHECK-NEXT: r4 = #23
1848 ; CHECK-NEXT: }
1849 ; CHECK-NEXT: {
1850 ; CHECK-NEXT: v4.uw = vcl0(v2.uw)
1851 ; CHECK-NEXT: }
1852 ; CHECK-NEXT: {
1853 ; CHECK-NEXT: v5.uw = vcl0(v3.uw)
1854 ; CHECK-NEXT: v4.w = vadd(v4.w,v1.w)
1855 ; CHECK-NEXT: }
1856 ; CHECK-NEXT: {
1857 ; CHECK-NEXT: v5.w = vadd(v5.w,v1.w)
1858 ; CHECK-NEXT: }
1859 ; CHECK-NEXT: {
1860 ; CHECK-NEXT: v7.w = vasl(v2.w,v4.w)
1861 ; CHECK-NEXT: }
1862 ; CHECK-NEXT: {
1863 ; CHECK-NEXT: v9.w = vasl(v3.w,v5.w)
1864 ; CHECK-NEXT: v11 = vand(v7,v8)
1865 ; CHECK-NEXT: v10.w = vadd(v7.w,v6.w)
1866 ; CHECK-NEXT: }
1867 ; CHECK-NEXT: {
1868 ; CHECK-NEXT: v6.w = vadd(v9.w,v6.w)
1869 ; CHECK-NEXT: q1 = vcmp.eq(v11.w,v21.w)
1870 ; CHECK-NEXT: v8 = vand(v9,v8)
1871 ; CHECK-NEXT: q0 = vcmp.gt(v7.uw,v10.uw)
1872 ; CHECK-NEXT: }
1873 ; CHECK-NEXT: {
1874 ; CHECK-NEXT: v22.uw = vlsr(v10.uw,r2)
1875 ; CHECK-NEXT: v24 = vmux(q1,v21,v1)
1876 ; CHECK-NEXT: q3 = vcmp.eq(v8.w,v21.w)
1877 ; CHECK-NEXT: q1 = vcmp.gt(v9.uw,v6.uw)
1878 ; CHECK-NEXT: }
1879 ; CHECK-NEXT: {
1880 ; CHECK-NEXT: v23.uw = vlsr(v6.uw,r2)
1881 ; CHECK-NEXT: v25 = vmux(q0,v1,v21)
1882 ; CHECK-NEXT: v27 = vmux(q3,v21,v1)
1883 ; CHECK-NEXT: v1 = vmux(q1,v1,v21)
1884 ; CHECK-NEXT: }
1885 ; CHECK-NEXT: {
1886 ; CHECK-NEXT: v4.w = vsub(v25.w,v4.w)
1887 ; CHECK-NEXT: v1.w = vsub(v1.w,v5.w)
1888 ; CHECK-NEXT: v10.w = vadd(v22.w,v24.w)
1889 ; CHECK-NEXT: v28.w = vadd(v23.w,v27.w)
1890 ; CHECK-NEXT: }
1891 ; CHECK-NEXT: {
1892 ; CHECK-NEXT: v12.uw = vlsr(v7.uw,r2)
1893 ; CHECK-NEXT: v4.w = vadd(v4.w,v13.w)
1894 ; CHECK-NEXT: v1.w = vadd(v1.w,v13.w)
1895 ; CHECK-NEXT: }
1896 ; CHECK-NEXT: {
1897 ; CHECK-NEXT: v26.uw = vlsr(v9.uw,r2)
1898 ; CHECK-NEXT: q2 = vcmp.eq(v12.w,v22.w)
1899 ; CHECK-NEXT: }
1900 ; CHECK-NEXT: {
1901 ; CHECK-NEXT: v11.uw = vlsr(v22.uw,r7)
1902 ; CHECK-NEXT: q3 = vcmp.eq(v26.w,v23.w)
1903 ; CHECK-NEXT: }
1904 ; CHECK-NEXT: {
1905 ; CHECK-NEXT: v30.uw = vlsr(v10.uw,r7)
1906 ; CHECK-NEXT: }
1907 ; CHECK-NEXT: {
1908 ; CHECK-NEXT: v29.uw = vlsr(v23.uw,r7)
1909 ; CHECK-NEXT: v5 = vmux(q2,v30,v11)
1910 ; CHECK-NEXT: q2 = vcmp.eq(v3.w,v21.w)
1911 ; CHECK-NEXT: }
1912 ; CHECK-NEXT: {
1913 ; CHECK-NEXT: v6.uw = vlsr(v28.uw,r7)
1914 ; CHECK-NEXT: }
1915 ; CHECK-NEXT: {
1916 ; CHECK-NEXT: v4.w = vasl(v4.w,r4)
1917 ; CHECK-NEXT: v6 = vmux(q3,v6,v29)
1918 ; CHECK-NEXT: q3 = vcmp.eq(v2.w,v21.w)
1919 ; CHECK-NEXT: }
1920 ; CHECK-NEXT: {
1921 ; CHECK-NEXT: v1.w = vasl(v1.w,r4)
1922 ; CHECK-NEXT: v31 = vor(v5,v4)
1923 ; CHECK-NEXT: }
1924 ; CHECK-NEXT: {
1925 ; CHECK-NEXT: v1 = vor(v6,v1)
1926 ; CHECK-NEXT: v0 = vmux(q3,v21,v31)
1927 ; CHECK-NEXT: vmem(r1+#0) = v0.new
1928 ; CHECK-NEXT: }
1929 ; CHECK-NEXT: {
1930 ; CHECK-NEXT: v1 = vmux(q2,v21,v1)
1931 ; CHECK-NEXT: jumpr r31
1932 ; CHECK-NEXT: vmem(r1+#1) = v1.new
1933 ; CHECK-NEXT: }
1934 %v0 = load <64 x i8>, ptr %a0, align 128
1935 %v1 = uitofp <64 x i8> %v0 to <64 x float>
1936 store <64 x float> %v1, ptr %a1, align 128
1937 ret void
1938 }
1940 ; Widen input #2
1941 define void @u8f32_2(ptr %a0, ptr %a1) #0 {
1942 ; CHECK-LABEL: u8f32_2:
1943 ; CHECK: .cfi_startproc
1944 ; CHECK-NEXT: // %bb.0:
1945 ; CHECK-NEXT: {
1946 ; CHECK-NEXT: v0 = vmem(r0+#0)
1947 ; CHECK-NEXT: }
1948 ; CHECK-NEXT: {
1949 ; CHECK-NEXT: r6 = #1
1950 ; CHECK-NEXT: r3 = #512
1951 ; CHECK-NEXT: v1:0.uh = vunpack(v0.ub)
1952 ; CHECK-NEXT: }
1953 ; CHECK-NEXT: {
1954 ; CHECK-NEXT: v2 = vsplat(r6)
1955 ; CHECK-NEXT: v4 = vsplat(r3)
1956 ; CHECK-NEXT: r2 = #255
1957 ; CHECK-NEXT: v3 = vxor(v3,v3)
1958 ; CHECK-NEXT: }
1959 ; CHECK-NEXT: {
1960 ; CHECK-NEXT: r5:4 = combine(##159,#8)
1961 ; CHECK-NEXT: v1:0.uw = vunpack(v0.uh)
1962 ; CHECK-NEXT: }
1963 ; CHECK-NEXT: {
1964 ; CHECK-NEXT: v1 = vsplat(r2)
1965 ; CHECK-NEXT: v7 = vsplat(r5)
1966 ; CHECK-NEXT: q3 = vcmp.eq(v0.w,v3.w)
1967 ; CHECK-NEXT: }
1968 ; CHECK-NEXT: {
1969 ; CHECK-NEXT: v5.uw = vcl0(v0.uw)
1970 ; CHECK-NEXT: }
1971 ; CHECK-NEXT: {
1972 ; CHECK-NEXT: v5.w = vadd(v5.w,v2.w)
1973 ; CHECK-NEXT: }
1974 ; CHECK-NEXT: {
1975 ; CHECK-NEXT: v6.w = vasl(v0.w,v5.w)
1976 ; CHECK-NEXT: }
1977 ; CHECK-NEXT: {
1978 ; CHECK-NEXT: v1.w = vadd(v6.w,v1.w)
1979 ; CHECK-NEXT: v4 = vand(v6,v4)
1980 ; CHECK-NEXT: }
1981 ; CHECK-NEXT: {
1982 ; CHECK-NEXT: v6.uw = vlsr(v6.uw,r4)
1983 ; CHECK-NEXT: q0 = vcmp.gt(v6.uw,v1.uw)
1984 ; CHECK-NEXT: q1 = vcmp.eq(v4.w,v3.w)
1985 ; CHECK-NEXT: }
1986 ; CHECK-NEXT: {
1987 ; CHECK-NEXT: r4 = #23
1988 ; CHECK-NEXT: v1.uw = vlsr(v1.uw,r4)
1989 ; CHECK-NEXT: v4 = vmux(q1,v3,v2)
1990 ; CHECK-NEXT: v2 = vmux(q0,v2,v3)
1991 ; CHECK-NEXT: }
1992 ; CHECK-NEXT: {
1993 ; CHECK-NEXT: v2.w = vsub(v2.w,v5.w)
1994 ; CHECK-NEXT: v4.w = vadd(v1.w,v4.w)
1995 ; CHECK-NEXT: q2 = vcmp.eq(v6.w,v1.w)
1996 ; CHECK-NEXT: }
1997 ; CHECK-NEXT: {
1998 ; CHECK-NEXT: v29.uw = vlsr(v1.uw,r6)
1999 ; CHECK-NEXT: v2.w = vadd(v2.w,v7.w)
2000 ; CHECK-NEXT: }
2001 ; CHECK-NEXT: {
2002 ; CHECK-NEXT: v30.uw = vlsr(v4.uw,r6)
2003 ; CHECK-NEXT: }
2004 ; CHECK-NEXT: {
2005 ; CHECK-NEXT: v2.w = vasl(v2.w,r4)
2006 ; CHECK-NEXT: v1 = vmux(q2,v30,v29)
2007 ; CHECK-NEXT: }
2008 ; CHECK-NEXT: {
2009 ; CHECK-NEXT: v31 = vor(v1,v2)
2010 ; CHECK-NEXT: }
2011 ; CHECK-NEXT: {
2012 ; CHECK-NEXT: v0 = vmux(q3,v3,v31)
2013 ; CHECK-NEXT: jumpr r31
2014 ; CHECK-NEXT: vmem(r1+#0) = v0.new
2015 ; CHECK-NEXT: }
2016 %v0 = load <32 x i8>, ptr %a0, align 128
2017 %v1 = uitofp <32 x i8> %v0 to <32 x float>
2018 store <32 x float> %v1, ptr %a1, align 128
2019 ret void
2020 }
2023 ; u16 -> f16
2024 ; No widening
2025 define void @u16f16_0(ptr %a0, ptr %a1) #0 {
2026 ; CHECK-LABEL: u16f16_0:
2027 ; CHECK: .cfi_startproc
2028 ; CHECK-NEXT: // %bb.0:
2029 ; CHECK-NEXT: {
2030 ; CHECK-NEXT: r3:2 = combine(#64,#1)
2031 ; CHECK-NEXT: r5 = #31
2032 ; CHECK-NEXT: v1.uh = vcl0(v0.uh)
2033 ; CHECK-NEXT: v0.cur = vmem(r0+#0)
2034 ; CHECK-NEXT: }
2035 ; CHECK-NEXT: {
2036 ; CHECK-NEXT: v2.h = vsplat(r2)
2037 ; CHECK-NEXT: v5.h = vsplat(r3)
2038 ; CHECK-NEXT: r4 = #5
2039 ; CHECK-NEXT: v3 = vxor(v3,v3)
2040 ; CHECK-NEXT: }
2041 ; CHECK-NEXT: {
2042 ; CHECK-NEXT: v4.h = vsplat(r5)
2043 ; CHECK-NEXT: r3 = #10
2044 ; CHECK-NEXT: v1.h = vadd(v1.h,v2.h)
2045 ; CHECK-NEXT: q3 = vcmp.eq(v0.h,v3.h)
2046 ; CHECK-NEXT: }
2047 ; CHECK-NEXT: {
2048 ; CHECK-NEXT: v6.h = vasl(v0.h,v1.h)
2049 ; CHECK-NEXT: }
2050 ; CHECK-NEXT: {
2051 ; CHECK-NEXT: v7.h = vadd(v6.h,v4.h)
2052 ; CHECK-NEXT: v5 = vand(v6,v5)
2053 ; CHECK-NEXT: }
2054 ; CHECK-NEXT: {
2055 ; CHECK-NEXT: v6.uh = vlsr(v6.uh,r4)
2056 ; CHECK-NEXT: q0 = vcmp.eq(v5.h,v3.h)
2057 ; CHECK-NEXT: q1 = vcmp.gt(v6.uh,v7.uh)
2058 ; CHECK-NEXT: }
2059 ; CHECK-NEXT: {
2060 ; CHECK-NEXT: v26.uh = vlsr(v7.uh,r4)
2061 ; CHECK-NEXT: v27 = vmux(q0,v3,v2)
2062 ; CHECK-NEXT: v2 = vmux(q1,v2,v3)
2063 ; CHECK-NEXT: }
2064 ; CHECK-NEXT: {
2065 ; CHECK-NEXT: v2.h = vadd(v2.h,v4.h)
2066 ; CHECK-NEXT: v28.h = vadd(v26.h,v27.h)
2067 ; CHECK-NEXT: q2 = vcmp.eq(v6.h,v26.h)
2068 ; CHECK-NEXT: }
2069 ; CHECK-NEXT: {
2070 ; CHECK-NEXT: v29.uh = vlsr(v26.uh,r2)
2071 ; CHECK-NEXT: v1.h = vsub(v2.h,v1.h)
2072 ; CHECK-NEXT: }
2073 ; CHECK-NEXT: {
2074 ; CHECK-NEXT: v30.uh = vlsr(v28.uh,r2)
2075 ; CHECK-NEXT: }
2076 ; CHECK-NEXT: {
2077 ; CHECK-NEXT: v1.h = vasl(v1.h,r3)
2078 ; CHECK-NEXT: v2 = vmux(q2,v30,v29)
2079 ; CHECK-NEXT: }
2080 ; CHECK-NEXT: {
2081 ; CHECK-NEXT: v31 = vor(v2,v1)
2082 ; CHECK-NEXT: }
2083 ; CHECK-NEXT: {
2084 ; CHECK-NEXT: v0 = vmux(q3,v3,v31)
2085 ; CHECK-NEXT: jumpr r31
2086 ; CHECK-NEXT: vmem(r1+#0) = v0.new
2087 ; CHECK-NEXT: }
2088 %v0 = load <64 x i16>, ptr %a0, align 128
2089 %v1 = uitofp <64 x i16> %v0 to <64 x half>
2090 store <64 x half> %v1, ptr %a1, align 128
2091 ret void
2092 }
2094 ; Widen input and result
2095 define void @u16f16_1(ptr %a0, ptr %a1) #0 {
2096 ; CHECK-LABEL: u16f16_1:
2097 ; CHECK: .cfi_startproc
2098 ; CHECK-NEXT: // %bb.0:
2099 ; CHECK-NEXT: {
2100 ; CHECK-NEXT: r3:2 = combine(#31,#1)
2101 ; CHECK-NEXT: r6 = #64
2102 ; CHECK-NEXT: v1.uh = vcl0(v0.uh)
2103 ; CHECK-NEXT: v0.cur = vmem(r0+#0)
2104 ; CHECK-NEXT: }
2105 ; CHECK-NEXT: {
2106 ; CHECK-NEXT: v2.h = vsplat(r2)
2107 ; CHECK-NEXT: v4.h = vsplat(r3)
2108 ; CHECK-NEXT: r5 = #5
2109 ; CHECK-NEXT: v3 = vxor(v3,v3)
2110 ; CHECK-NEXT: }
2111 ; CHECK-NEXT: {
2112 ; CHECK-NEXT: v5.h = vsplat(r6)
2113 ; CHECK-NEXT: r4 = #10
2114 ; CHECK-NEXT: v1.h = vadd(v1.h,v2.h)
2115 ; CHECK-NEXT: q2 = vcmp.eq(v0.h,v3.h)
2116 ; CHECK-NEXT: }
2117 ; CHECK-NEXT: {
2118 ; CHECK-NEXT: q3 = vsetq(r6)
2119 ; CHECK-NEXT: }
2120 ; CHECK-NEXT: {
2121 ; CHECK-NEXT: v6.h = vasl(v0.h,v1.h)
2122 ; CHECK-NEXT: }
2123 ; CHECK-NEXT: {
2124 ; CHECK-NEXT: v7.h = vadd(v6.h,v4.h)
2125 ; CHECK-NEXT: v5 = vand(v6,v5)
2126 ; CHECK-NEXT: }
2127 ; CHECK-NEXT: {
2128 ; CHECK-NEXT: v6.uh = vlsr(v6.uh,r5)
2129 ; CHECK-NEXT: q1 = vcmp.eq(v5.h,v3.h)
2130 ; CHECK-NEXT: q0 = vcmp.gt(v6.uh,v7.uh)
2131 ; CHECK-NEXT: }
2132 ; CHECK-NEXT: {
2133 ; CHECK-NEXT: v7.uh = vlsr(v7.uh,r5)
2134 ; CHECK-NEXT: v5 = vmux(q1,v3,v2)
2135 ; CHECK-NEXT: v2 = vmux(q0,v2,v3)
2136 ; CHECK-NEXT: }
2137 ; CHECK-NEXT: {
2138 ; CHECK-NEXT: v2.h = vadd(v2.h,v4.h)
2139 ; CHECK-NEXT: v28.h = vadd(v7.h,v5.h)
2140 ; CHECK-NEXT: q1 = vcmp.eq(v6.h,v7.h)
2141 ; CHECK-NEXT: }
2142 ; CHECK-NEXT: {
2143 ; CHECK-NEXT: v29.uh = vlsr(v7.uh,r2)
2144 ; CHECK-NEXT: v1.h = vsub(v2.h,v1.h)
2145 ; CHECK-NEXT: }
2146 ; CHECK-NEXT: {
2147 ; CHECK-NEXT: v30.uh = vlsr(v28.uh,r2)
2148 ; CHECK-NEXT: }
2149 ; CHECK-NEXT: {
2150 ; CHECK-NEXT: v1.h = vasl(v1.h,r4)
2151 ; CHECK-NEXT: v2 = vmux(q1,v30,v29)
2152 ; CHECK-NEXT: }
2153 ; CHECK-NEXT: {
2154 ; CHECK-NEXT: v31 = vor(v2,v1)
2155 ; CHECK-NEXT: }
2156 ; CHECK-NEXT: {
2157 ; CHECK-NEXT: v0 = vmux(q2,v3,v31)
2158 ; CHECK-NEXT: }
2159 ; CHECK-NEXT: {
2160 ; CHECK-NEXT: jumpr r31
2161 ; CHECK-NEXT: if (q3) vmem(r1+#0) = v0
2162 ; CHECK-NEXT: }
2163 %v0 = load <32 x i16>, ptr %a0, align 128
2164 %v1 = uitofp <32 x i16> %v0 to <32 x half>
2165 store <32 x half> %v1, ptr %a1, align 128
2166 ret void
2167 }
2170 ; u16 -> f32
2171 ; No widening
2172 define void @u16f32_0(ptr %a0, ptr %a1) #0 {
2173 ; CHECK-LABEL: u16f32_0:
2174 ; CHECK: .cfi_startproc
2175 ; CHECK-NEXT: // %bb.0:
2176 ; CHECK-NEXT: {
2177 ; CHECK-NEXT: v0 = vmem(r0+#0)
2178 ; CHECK-NEXT: }
2179 ; CHECK-NEXT: {
2180 ; CHECK-NEXT: r7 = #1
2181 ; CHECK-NEXT: r3:2 = combine(##255,#8)
2182 ; CHECK-NEXT: v1:0.uw = vunpack(v0.uh)
2183 ; CHECK-NEXT: }
2184 ; CHECK-NEXT: {
2185 ; CHECK-NEXT: v3 = vsplat(r7)
2186 ; CHECK-NEXT: v6 = vsplat(r3)
2187 ; CHECK-NEXT: r6 = #512
2188 ; CHECK-NEXT: v2 = vxor(v2,v2)
2189 ; CHECK-NEXT: }
2190 ; CHECK-NEXT: {
2191 ; CHECK-NEXT: v8 = vsplat(r6)
2192 ; CHECK-NEXT: r5 = #159
2193 ; CHECK-NEXT: r4 = #23
2194 ; CHECK-NEXT: v4.uw = vcl0(v0.uw)
2195 ; CHECK-NEXT: }
2196 ; CHECK-NEXT: {
2197 ; CHECK-NEXT: v14 = vsplat(r5)
2198 ; CHECK-NEXT: v5.uw = vcl0(v1.uw)
2199 ; CHECK-NEXT: v4.w = vadd(v4.w,v3.w)
2200 ; CHECK-NEXT: }
2201 ; CHECK-NEXT: {
2202 ; CHECK-NEXT: v5.w = vadd(v5.w,v3.w)
2203 ; CHECK-NEXT: }
2204 ; CHECK-NEXT: {
2205 ; CHECK-NEXT: v7.w = vasl(v0.w,v4.w)
2206 ; CHECK-NEXT: }
2207 ; CHECK-NEXT: {
2208 ; CHECK-NEXT: v9.w = vasl(v1.w,v5.w)
2209 ; CHECK-NEXT: v10.w = vadd(v7.w,v6.w)
2210 ; CHECK-NEXT: v11 = vand(v7,v8)
2211 ; CHECK-NEXT: }
2212 ; CHECK-NEXT: {
2213 ; CHECK-NEXT: v6.w = vadd(v9.w,v6.w)
2214 ; CHECK-NEXT: v8 = vand(v9,v8)
2215 ; CHECK-NEXT: q1 = vcmp.eq(v11.w,v2.w)
2216 ; CHECK-NEXT: q0 = vcmp.gt(v7.uw,v10.uw)
2217 ; CHECK-NEXT: }
2218 ; CHECK-NEXT: {
2219 ; CHECK-NEXT: v19.uw = vlsr(v10.uw,r2)
2220 ; CHECK-NEXT: q2 = vcmp.eq(v8.w,v2.w)
2221 ; CHECK-NEXT: q3 = vcmp.gt(v9.uw,v6.uw)
2222 ; CHECK-NEXT: v20 = vmux(q1,v2,v3)
2223 ; CHECK-NEXT: }
2224 ; CHECK-NEXT: {
2225 ; CHECK-NEXT: v21.uw = vlsr(v6.uw,r2)
2226 ; CHECK-NEXT: v22 = vmux(q2,v2,v3)
2227 ; CHECK-NEXT: v25 = vmux(q0,v3,v2)
2228 ; CHECK-NEXT: v3 = vmux(q3,v3,v2)
2229 ; CHECK-NEXT: }
2230 ; CHECK-NEXT: {
2231 ; CHECK-NEXT: v4.w = vsub(v25.w,v4.w)
2232 ; CHECK-NEXT: v3.w = vsub(v3.w,v5.w)
2233 ; CHECK-NEXT: v23.w = vadd(v19.w,v20.w)
2234 ; CHECK-NEXT: v10.w = vadd(v21.w,v22.w)
2235 ; CHECK-NEXT: }
2236 ; CHECK-NEXT: {
2237 ; CHECK-NEXT: v12.uw = vlsr(v7.uw,r2)
2238 ; CHECK-NEXT: v4.w = vadd(v4.w,v14.w)
2239 ; CHECK-NEXT: v3.w = vadd(v3.w,v14.w)
2240 ; CHECK-NEXT: }
2241 ; CHECK-NEXT: {
2242 ; CHECK-NEXT: v24.uw = vlsr(v9.uw,r2)
2243 ; CHECK-NEXT: q2 = vcmp.eq(v12.w,v19.w)
2244 ; CHECK-NEXT: }
2245 ; CHECK-NEXT: {
2246 ; CHECK-NEXT: v13.uw = vlsr(v19.uw,r7)
2247 ; CHECK-NEXT: q3 = vcmp.eq(v24.w,v21.w)
2248 ; CHECK-NEXT: }
2249 ; CHECK-NEXT: {
2250 ; CHECK-NEXT: v26.uw = vlsr(v23.uw,r7)
2251 ; CHECK-NEXT: }
2252 ; CHECK-NEXT: {
2253 ; CHECK-NEXT: v27.uw = vlsr(v10.uw,r7)
2254 ; CHECK-NEXT: v5 = vmux(q2,v26,v13)
2255 ; CHECK-NEXT: q2 = vcmp.eq(v1.w,v2.w)
2256 ; CHECK-NEXT: }
2257 ; CHECK-NEXT: {
2258 ; CHECK-NEXT: v28.uw = vlsr(v21.uw,r7)
2259 ; CHECK-NEXT: }
2260 ; CHECK-NEXT: {
2261 ; CHECK-NEXT: v4.w = vasl(v4.w,r4)
2262 ; CHECK-NEXT: v6 = vmux(q3,v27,v28)
2263 ; CHECK-NEXT: q3 = vcmp.eq(v0.w,v2.w)
2264 ; CHECK-NEXT: }
2265 ; CHECK-NEXT: {
2266 ; CHECK-NEXT: v3.w = vasl(v3.w,r4)
2267 ; CHECK-NEXT: v29 = vor(v5,v4)
2268 ; CHECK-NEXT: }
2269 ; CHECK-NEXT: {
2270 ; CHECK-NEXT: v3 = vor(v6,v3)
2271 ; CHECK-NEXT: v31 = vmux(q3,v2,v29)
2272 ; CHECK-NEXT: vmem(r1+#0) = v31.new
2273 ; CHECK-NEXT: }
2274 ; CHECK-NEXT: {
2275 ; CHECK-NEXT: v30 = vmux(q2,v2,v3)
2276 ; CHECK-NEXT: jumpr r31
2277 ; CHECK-NEXT: vmem(r1+#1) = v30.new
2278 ; CHECK-NEXT: }
2279 %v0 = load <64 x i16>, ptr %a0, align 128
2280 %v1 = uitofp <64 x i16> %v0 to <64 x float>
2281 store <64 x float> %v1, ptr %a1, align 128
2282 ret void
2283 }
2285 ; Widen input
2286 define void @u16f32_1(ptr %a0, ptr %a1) #0 {
2287 ; CHECK-LABEL: u16f32_1:
2288 ; CHECK: .cfi_startproc
2289 ; CHECK-NEXT: // %bb.0:
2290 ; CHECK-NEXT: {
2291 ; CHECK-NEXT: v0 = vmem(r0+#0)
2292 ; CHECK-NEXT: }
2293 ; CHECK-NEXT: {
2294 ; CHECK-NEXT: r6 = #1
2295 ; CHECK-NEXT: r2 = #255
2296 ; CHECK-NEXT: v1:0.uw = vunpack(v0.uh)
2297 ; CHECK-NEXT: }
2298 ; CHECK-NEXT: {
2299 ; CHECK-NEXT: v1 = vsplat(r6)
2300 ; CHECK-NEXT: v4 = vsplat(r2)
2301 ; CHECK-NEXT: r3 = #512
2302 ; CHECK-NEXT: v2 = vxor(v2,v2)
2303 ; CHECK-NEXT: }
2304 ; CHECK-NEXT: {
2305 ; CHECK-NEXT: v5 = vsplat(r3)
2306 ; CHECK-NEXT: r5:4 = combine(##159,#8)
2307 ; CHECK-NEXT: v3.uw = vcl0(v0.uw)
2308 ; CHECK-NEXT: }
2309 ; CHECK-NEXT: {
2310 ; CHECK-NEXT: v7 = vsplat(r5)
2311 ; CHECK-NEXT: q3 = vcmp.eq(v0.w,v2.w)
2312 ; CHECK-NEXT: v3.w = vadd(v3.w,v1.w)
2313 ; CHECK-NEXT: }
2314 ; CHECK-NEXT: {
2315 ; CHECK-NEXT: v6.w = vasl(v0.w,v3.w)
2316 ; CHECK-NEXT: }
2317 ; CHECK-NEXT: {
2318 ; CHECK-NEXT: v4.w = vadd(v6.w,v4.w)
2319 ; CHECK-NEXT: v5 = vand(v6,v5)
2320 ; CHECK-NEXT: }
2321 ; CHECK-NEXT: {
2322 ; CHECK-NEXT: v6.uw = vlsr(v6.uw,r4)
2323 ; CHECK-NEXT: q0 = vcmp.gt(v6.uw,v4.uw)
2324 ; CHECK-NEXT: q1 = vcmp.eq(v5.w,v2.w)
2325 ; CHECK-NEXT: }
2326 ; CHECK-NEXT: {
2327 ; CHECK-NEXT: r4 = #23
2328 ; CHECK-NEXT: v4.uw = vlsr(v4.uw,r4)
2329 ; CHECK-NEXT: v5 = vmux(q1,v2,v1)
2330 ; CHECK-NEXT: v1 = vmux(q0,v1,v2)
2331 ; CHECK-NEXT: }
2332 ; CHECK-NEXT: {
2333 ; CHECK-NEXT: v1.w = vsub(v1.w,v3.w)
2334 ; CHECK-NEXT: v29.w = vadd(v4.w,v5.w)
2335 ; CHECK-NEXT: q2 = vcmp.eq(v6.w,v4.w)
2336 ; CHECK-NEXT: }
2337 ; CHECK-NEXT: {
2338 ; CHECK-NEXT: v30.uw = vlsr(v4.uw,r6)
2339 ; CHECK-NEXT: v1.w = vadd(v1.w,v7.w)
2340 ; CHECK-NEXT: }
2341 ; CHECK-NEXT: {
2342 ; CHECK-NEXT: v3.uw = vlsr(v29.uw,r6)
2343 ; CHECK-NEXT: }
2344 ; CHECK-NEXT: {
2345 ; CHECK-NEXT: v1.w = vasl(v1.w,r4)
2346 ; CHECK-NEXT: v3 = vmux(q2,v3,v30)
2347 ; CHECK-NEXT: }
2348 ; CHECK-NEXT: {
2349 ; CHECK-NEXT: v31 = vor(v3,v1)
2350 ; CHECK-NEXT: }
2351 ; CHECK-NEXT: {
2352 ; CHECK-NEXT: v0 = vmux(q3,v2,v31)
2353 ; CHECK-NEXT: jumpr r31
2354 ; CHECK-NEXT: vmem(r1+#0) = v0.new
2355 ; CHECK-NEXT: }
2356 %v0 = load <32 x i16>, ptr %a0, align 128
2357 %v1 = uitofp <32 x i16> %v0 to <32 x float>
2358 store <32 x float> %v1, ptr %a1, align 128
2359 ret void
2360 }
2363 ; u32 -> f16
2364 ; No widening
2365 define void @u32f16_0(ptr %a0, ptr %a1) #0 {
2366 ; CHECK-LABEL: u32f16_0:
2367 ; CHECK: .cfi_startproc
2368 ; CHECK-NEXT: // %bb.0:
2369 ; CHECK-NEXT: {
2370 ; CHECK-NEXT: r3:2 = combine(#8,#1)
2371 ; CHECK-NEXT: r6 = #255
2372 ; CHECK-NEXT: v3.uw = vcl0(v0.uw)
2373 ; CHECK-NEXT: v0.cur = vmem(r0+#1)
2374 ; CHECK-NEXT: }
2375 ; CHECK-NEXT: {
2376 ; CHECK-NEXT: v2 = vsplat(r2)
2377 ; CHECK-NEXT: r4 = #512
2378 ; CHECK-NEXT: v4.uw = vcl0(v1.uw)
2379 ; CHECK-NEXT: v1.cur = vmem(r0+#0)
2380 ; CHECK-NEXT: }
2381 ; CHECK-NEXT: {
2382 ; CHECK-NEXT: v7 = vsplat(r4)
2383 ; CHECK-NEXT: v6 = vsplat(r6)
2384 ; CHECK-NEXT: v4.w = vadd(v4.w,v2.w)
2385 ; CHECK-NEXT: v3.w = vadd(v3.w,v2.w)
2386 ; CHECK-NEXT: }
2387 ; CHECK-NEXT: {
2388 ; CHECK-NEXT: r4 = #159
2389 ; CHECK-NEXT: v9 = vxor(v9,v9)
2390 ; CHECK-NEXT: }
2391 ; CHECK-NEXT: {
2392 ; CHECK-NEXT: v10 = vsplat(r4)
2393 ; CHECK-NEXT: v5.w = vasl(v1.w,v4.w)
2394 ; CHECK-NEXT: }
2395 ; CHECK-NEXT: {
2396 ; CHECK-NEXT: v8.w = vasl(v0.w,v3.w)
2397 ; CHECK-NEXT: v11.w = vadd(v5.w,v6.w)
2398 ; CHECK-NEXT: v13 = vand(v5,v7)
2399 ; CHECK-NEXT: }
2400 ; CHECK-NEXT: {
2401 ; CHECK-NEXT: v6.w = vadd(v8.w,v6.w)
2402 ; CHECK-NEXT: v7 = vand(v8,v7)
2403 ; CHECK-NEXT: q1 = vcmp.gt(v5.uw,v11.uw)
2404 ; CHECK-NEXT: q2 = vcmp.eq(v13.w,v9.w)
2405 ; CHECK-NEXT: }
2406 ; CHECK-NEXT: {
2407 ; CHECK-NEXT: v27.uw = vlsr(v11.uw,r3)
2408 ; CHECK-NEXT: q3 = vcmp.gt(v8.uw,v6.uw)
2409 ; CHECK-NEXT: q0 = vcmp.eq(v7.w,v9.w)
2410 ; CHECK-NEXT: v28 = vmux(q2,v9,v2)
2411 ; CHECK-NEXT: }
2412 ; CHECK-NEXT: {
2413 ; CHECK-NEXT: v6.uw = vlsr(v6.uw,r3)
2414 ; CHECK-NEXT: v29 = vmux(q1,v2,v9)
2415 ; CHECK-NEXT: v30 = vmux(q3,v2,v9)
2416 ; CHECK-NEXT: v2 = vmux(q0,v9,v2)
2417 ; CHECK-NEXT: }
2418 ; CHECK-NEXT: {
2419 ; CHECK-NEXT: v4.w = vsub(v29.w,v4.w)
2420 ; CHECK-NEXT: v7.w = vadd(v27.w,v28.w)
2421 ; CHECK-NEXT: v3.w = vsub(v30.w,v3.w)
2422 ; CHECK-NEXT: v2.w = vadd(v6.w,v2.w)
2423 ; CHECK-NEXT: }
2424 ; CHECK-NEXT: {
2425 ; CHECK-NEXT: v12.uw = vlsr(v5.uw,r3)
2426 ; CHECK-NEXT: v4.w = vadd(v4.w,v10.w)
2427 ; CHECK-NEXT: v3.w = vadd(v3.w,v10.w)
2428 ; CHECK-NEXT: q2 = vcmp.eq(v1.w,v9.w)
2429 ; CHECK-NEXT: }
2430 ; CHECK-NEXT: {
2431 ; CHECK-NEXT: r3 = #23
2432 ; CHECK-NEXT: v14.uw = vlsr(v8.uw,r3)
2433 ; CHECK-NEXT: q3 = vcmp.eq(v12.w,v27.w)
2434 ; CHECK-NEXT: }
2435 ; CHECK-NEXT: {
2436 ; CHECK-NEXT: v5.uw = vlsr(v27.uw,r2)
2437 ; CHECK-NEXT: q1 = vcmp.eq(v14.w,v6.w)
2438 ; CHECK-NEXT: }
2439 ; CHECK-NEXT: {
2440 ; CHECK-NEXT: v7.uw = vlsr(v7.uw,r2)
2441 ; CHECK-NEXT: }
2442 ; CHECK-NEXT: {
2443 ; CHECK-NEXT: v2.uw = vlsr(v2.uw,r2)
2444 ; CHECK-NEXT: v5 = vmux(q3,v7,v5)
2445 ; CHECK-NEXT: q3 = vcmp.eq(v0.w,v9.w)
2446 ; CHECK-NEXT: }
2447 ; CHECK-NEXT: {
2448 ; CHECK-NEXT: v6.uw = vlsr(v6.uw,r2)
2449 ; CHECK-NEXT: }
2450 ; CHECK-NEXT: {
2451 ; CHECK-NEXT: v4.w = vasl(v4.w,r3)
2452 ; CHECK-NEXT: v31 = vmux(q1,v2,v6)
2453 ; CHECK-NEXT: }
2454 ; CHECK-NEXT: {
2455 ; CHECK-NEXT: v2.w = vasl(v3.w,r3)
2456 ; CHECK-NEXT: v4 = vor(v5,v4)
2457 ; CHECK-NEXT: }
2458 ; CHECK-NEXT: {
2459 ; CHECK-NEXT: v1 = vor(v31,v2)
2460 ; CHECK-NEXT: v3 = vmux(q2,v9,v4)
2461 ; CHECK-NEXT: }
2462 ; CHECK-NEXT: {
2463 ; CHECK-NEXT: v0 = vmux(q3,v9,v1)
2464 ; CHECK-NEXT: }
2465 ; CHECK-NEXT: {
2466 ; CHECK-NEXT: v2.qf32 = vadd(v3.sf,v9.sf)
2467 ; CHECK-NEXT: }
2468 ; CHECK-NEXT: {
2469 ; CHECK-NEXT: v3.qf32 = vadd(v0.sf,v9.sf)
2470 ; CHECK-NEXT: }
2471 ; CHECK-NEXT: {
2472 ; CHECK-NEXT: v0.hf = v3:2.qf32
2473 ; CHECK-NEXT: }
2474 ; CHECK-NEXT: {
2475 ; CHECK-NEXT: v0.h = vdeal(v0.h)
2476 ; CHECK-NEXT: jumpr r31
2477 ; CHECK-NEXT: vmem(r1+#0) = v0.new
2478 ; CHECK-NEXT: }
2479 %v0 = load <64 x i32>, ptr %a0, align 128
2480 %v1 = uitofp <64 x i32> %v0 to <64 x half>
2481 store <64 x half> %v1, ptr %a1, align 128
2482 ret void
2483 }
2485 ; Widen result
2486 define void @u32f16_1(ptr %a0, ptr %a1) #0 {
2487 ; CHECK-LABEL: u32f16_1:
2488 ; CHECK: .cfi_startproc
2489 ; CHECK-NEXT: // %bb.0:
2490 ; CHECK-NEXT: {
2491 ; CHECK-NEXT: r3:2 = combine(##512,#1)
2492 ; CHECK-NEXT: v1.uw = vcl0(v0.uw)
2493 ; CHECK-NEXT: v0.cur = vmem(r0+#0)
2494 ; CHECK-NEXT: }
2495 ; CHECK-NEXT: {
2496 ; CHECK-NEXT: v3 = vsplat(r2)
2497 ; CHECK-NEXT: v5 = vsplat(r3)
2498 ; CHECK-NEXT: r6 = #255
2499 ; CHECK-NEXT: v2 = vxor(v2,v2)
2500 ; CHECK-NEXT: }
2501 ; CHECK-NEXT: {
2502 ; CHECK-NEXT: v4 = vsplat(r6)
2503 ; CHECK-NEXT: r5 = #8
2504 ; CHECK-NEXT: r4 = #159
2505 ; CHECK-NEXT: v1.w = vadd(v1.w,v3.w)
2506 ; CHECK-NEXT: }
2507 ; CHECK-NEXT: {
2508 ; CHECK-NEXT: v7 = vsplat(r4)
2509 ; CHECK-NEXT: r3 = #23
2510 ; CHECK-NEXT: q2 = vcmp.eq(v0.w,v2.w)
2511 ; CHECK-NEXT: }
2512 ; CHECK-NEXT: {
2513 ; CHECK-NEXT: v6.w = vasl(v0.w,v1.w)
2514 ; CHECK-NEXT: }
2515 ; CHECK-NEXT: {
2516 ; CHECK-NEXT: v4.w = vadd(v6.w,v4.w)
2517 ; CHECK-NEXT: v5 = vand(v6,v5)
2518 ; CHECK-NEXT: }
2519 ; CHECK-NEXT: {
2520 ; CHECK-NEXT: v6.uw = vlsr(v6.uw,r5)
2521 ; CHECK-NEXT: q0 = vcmp.eq(v5.w,v2.w)
2522 ; CHECK-NEXT: q1 = vcmp.gt(v6.uw,v4.uw)
2523 ; CHECK-NEXT: }
2524 ; CHECK-NEXT: {
2525 ; CHECK-NEXT: v4.uw = vlsr(v4.uw,r5)
2526 ; CHECK-NEXT: v5 = vmux(q0,v2,v3)
2527 ; CHECK-NEXT: v3 = vmux(q1,v3,v2)
2528 ; CHECK-NEXT: }
2529 ; CHECK-NEXT: {
2530 ; CHECK-NEXT: v1.w = vsub(v3.w,v1.w)
2531 ; CHECK-NEXT: v30.w = vadd(v4.w,v5.w)
2532 ; CHECK-NEXT: q1 = vcmp.eq(v6.w,v4.w)
2533 ; CHECK-NEXT: }
2534 ; CHECK-NEXT: {
2535 ; CHECK-NEXT: v31.uw = vlsr(v4.uw,r2)
2536 ; CHECK-NEXT: v1.w = vadd(v1.w,v7.w)
2537 ; CHECK-NEXT: }
2538 ; CHECK-NEXT: {
2539 ; CHECK-NEXT: r2 = #64
2540 ; CHECK-NEXT: v3.uw = vlsr(v30.uw,r2)
2541 ; CHECK-NEXT: }
2542 ; CHECK-NEXT: {
2543 ; CHECK-NEXT: v1.w = vasl(v1.w,r3)
2544 ; CHECK-NEXT: q3 = vsetq(r2)
2545 ; CHECK-NEXT: v3 = vmux(q1,v3,v31)
2546 ; CHECK-NEXT: }
2547 ; CHECK-NEXT: {
2548 ; CHECK-NEXT: v1.qf32 = vadd(v2.sf,v2.sf)
2549 ; CHECK-NEXT: v0 = vor(v3,v1)
2550 ; CHECK-NEXT: }
2551 ; CHECK-NEXT: {
2552 ; CHECK-NEXT: v0 = vmux(q2,v2,v0)
2553 ; CHECK-NEXT: }
2554 ; CHECK-NEXT: {
2555 ; CHECK-NEXT: v0.qf32 = vadd(v0.sf,v2.sf)
2556 ; CHECK-NEXT: }
2557 ; CHECK-NEXT: {
2558 ; CHECK-NEXT: v0.hf = v1:0.qf32
2559 ; CHECK-NEXT: }
2560 ; CHECK-NEXT: {
2561 ; CHECK-NEXT: v0.h = vdeal(v0.h)
2562 ; CHECK-NEXT: }
2563 ; CHECK-NEXT: {
2564 ; CHECK-NEXT: jumpr r31
2565 ; CHECK-NEXT: if (q3) vmem(r1+#0) = v0
2566 ; CHECK-NEXT: }
2567 %v0 = load <32 x i32>, ptr %a0, align 128
2568 %v1 = uitofp <32 x i32> %v0 to <32 x half>
2569 store <32 x half> %v1, ptr %a1, align 128
2570 ret void
2571 }
2573 ; u32 -> f32
2574 ; No widening
2575 define void @u32f32_0(ptr %a0, ptr %a1) #0 {
2576 ; CHECK-LABEL: u32f32_0:
2577 ; CHECK: .cfi_startproc
2578 ; CHECK-NEXT: // %bb.0:
2579 ; CHECK-NEXT: {
2580 ; CHECK-NEXT: r3:2 = combine(##512,#1)
2581 ; CHECK-NEXT: v1.uw = vcl0(v0.uw)
2582 ; CHECK-NEXT: v0.cur = vmem(r0+#0)
2583 ; CHECK-NEXT: }
2584 ; CHECK-NEXT: {
2585 ; CHECK-NEXT: v2 = vsplat(r2)
2586 ; CHECK-NEXT: v5 = vsplat(r3)
2587 ; CHECK-NEXT: r6 = #255
2588 ; CHECK-NEXT: v3 = vxor(v3,v3)
2589 ; CHECK-NEXT: }
2590 ; CHECK-NEXT: {
2591 ; CHECK-NEXT: v4 = vsplat(r6)
2592 ; CHECK-NEXT: r5 = #8
2593 ; CHECK-NEXT: r4 = #159
2594 ; CHECK-NEXT: v1.w = vadd(v1.w,v2.w)
2595 ; CHECK-NEXT: }
2596 ; CHECK-NEXT: {
2597 ; CHECK-NEXT: v7 = vsplat(r4)
2598 ; CHECK-NEXT: r3 = #23
2599 ; CHECK-NEXT: q3 = vcmp.eq(v0.w,v3.w)
2600 ; CHECK-NEXT: }
2601 ; CHECK-NEXT: {
2602 ; CHECK-NEXT: v6.w = vasl(v0.w,v1.w)
2603 ; CHECK-NEXT: }
2604 ; CHECK-NEXT: {
2605 ; CHECK-NEXT: v4.w = vadd(v6.w,v4.w)
2606 ; CHECK-NEXT: v5 = vand(v6,v5)
2607 ; CHECK-NEXT: }
2608 ; CHECK-NEXT: {
2609 ; CHECK-NEXT: v6.uw = vlsr(v6.uw,r5)
2610 ; CHECK-NEXT: q0 = vcmp.eq(v5.w,v3.w)
2611 ; CHECK-NEXT: q1 = vcmp.gt(v6.uw,v4.uw)
2612 ; CHECK-NEXT: }
2613 ; CHECK-NEXT: {
2614 ; CHECK-NEXT: v4.uw = vlsr(v4.uw,r5)
2615 ; CHECK-NEXT: v5 = vmux(q0,v3,v2)
2616 ; CHECK-NEXT: v2 = vmux(q1,v2,v3)
2617 ; CHECK-NEXT: }
2618 ; CHECK-NEXT: {
2619 ; CHECK-NEXT: v1.w = vsub(v2.w,v1.w)
2620 ; CHECK-NEXT: v29.w = vadd(v4.w,v5.w)
2621 ; CHECK-NEXT: q2 = vcmp.eq(v6.w,v4.w)
2622 ; CHECK-NEXT: }
2623 ; CHECK-NEXT: {
2624 ; CHECK-NEXT: v30.uw = vlsr(v4.uw,r2)
2625 ; CHECK-NEXT: v1.w = vadd(v1.w,v7.w)
2626 ; CHECK-NEXT: }
2627 ; CHECK-NEXT: {
2628 ; CHECK-NEXT: v2.uw = vlsr(v29.uw,r2)
2629 ; CHECK-NEXT: }
2630 ; CHECK-NEXT: {
2631 ; CHECK-NEXT: v1.w = vasl(v1.w,r3)
2632 ; CHECK-NEXT: v2 = vmux(q2,v2,v30)
2633 ; CHECK-NEXT: }
2634 ; CHECK-NEXT: {
2635 ; CHECK-NEXT: v31 = vor(v2,v1)
2636 ; CHECK-NEXT: }
2637 ; CHECK-NEXT: {
2638 ; CHECK-NEXT: v0 = vmux(q3,v3,v31)
2639 ; CHECK-NEXT: jumpr r31
2640 ; CHECK-NEXT: vmem(r1+#0) = v0.new
2641 ; CHECK-NEXT: }
2642 %v0 = load <32 x i32>, ptr %a0, align 128
2643 %v1 = uitofp <32 x i32> %v0 to <32 x float>
2644 store <32 x float> %v1, ptr %a1, align 128
2645 ret void
2646 }
2648 ; Widen input and result
2649 define void @u32f32_1(ptr %a0, ptr %a1) #0 {
2650 ; CHECK-LABEL: u32f32_1:
2651 ; CHECK: .cfi_startproc
2652 ; CHECK-NEXT: // %bb.0:
2653 ; CHECK-NEXT: {
2654 ; CHECK-NEXT: r3:2 = combine(##512,#1)
2655 ; CHECK-NEXT: v1.uw = vcl0(v0.uw)
2656 ; CHECK-NEXT: v0.cur = vmem(r0+#0)
2657 ; CHECK-NEXT: }
2658 ; CHECK-NEXT: {
2659 ; CHECK-NEXT: v2 = vsplat(r2)
2660 ; CHECK-NEXT: v5 = vsplat(r3)
2661 ; CHECK-NEXT: r6 = #255
2662 ; CHECK-NEXT: v3 = vxor(v3,v3)
2663 ; CHECK-NEXT: }
2664 ; CHECK-NEXT: {
2665 ; CHECK-NEXT: v4 = vsplat(r6)
2666 ; CHECK-NEXT: r5 = #8
2667 ; CHECK-NEXT: r4 = #159
2668 ; CHECK-NEXT: v1.w = vadd(v1.w,v2.w)
2669 ; CHECK-NEXT: }
2670 ; CHECK-NEXT: {
2671 ; CHECK-NEXT: v7 = vsplat(r4)
2672 ; CHECK-NEXT: r3 = #23
2673 ; CHECK-NEXT: q2 = vcmp.eq(v0.w,v3.w)
2674 ; CHECK-NEXT: }
2675 ; CHECK-NEXT: {
2676 ; CHECK-NEXT: v6.w = vasl(v0.w,v1.w)
2677 ; CHECK-NEXT: }
2678 ; CHECK-NEXT: {
2679 ; CHECK-NEXT: v4.w = vadd(v6.w,v4.w)
2680 ; CHECK-NEXT: v5 = vand(v6,v5)
2681 ; CHECK-NEXT: }
2682 ; CHECK-NEXT: {
2683 ; CHECK-NEXT: v6.uw = vlsr(v6.uw,r5)
2684 ; CHECK-NEXT: q0 = vcmp.eq(v5.w,v3.w)
2685 ; CHECK-NEXT: q1 = vcmp.gt(v6.uw,v4.uw)
2686 ; CHECK-NEXT: }
2687 ; CHECK-NEXT: {
2688 ; CHECK-NEXT: v4.uw = vlsr(v4.uw,r5)
2689 ; CHECK-NEXT: v5 = vmux(q0,v3,v2)
2690 ; CHECK-NEXT: v2 = vmux(q1,v2,v3)
2691 ; CHECK-NEXT: }
2692 ; CHECK-NEXT: {
2693 ; CHECK-NEXT: v1.w = vsub(v2.w,v1.w)
2694 ; CHECK-NEXT: v29.w = vadd(v4.w,v5.w)
2695 ; CHECK-NEXT: q1 = vcmp.eq(v6.w,v4.w)
2696 ; CHECK-NEXT: }
2697 ; CHECK-NEXT: {
2698 ; CHECK-NEXT: v30.uw = vlsr(v4.uw,r2)
2699 ; CHECK-NEXT: v1.w = vadd(v1.w,v7.w)
2700 ; CHECK-NEXT: }
2701 ; CHECK-NEXT: {
2702 ; CHECK-NEXT: r2 = #64
2703 ; CHECK-NEXT: v2.uw = vlsr(v29.uw,r2)
2704 ; CHECK-NEXT: }
2705 ; CHECK-NEXT: {
2706 ; CHECK-NEXT: v1.w = vasl(v1.w,r3)
2707 ; CHECK-NEXT: q3 = vsetq(r2)
2708 ; CHECK-NEXT: v2 = vmux(q1,v2,v30)
2709 ; CHECK-NEXT: }
2710 ; CHECK-NEXT: {
2711 ; CHECK-NEXT: v31 = vor(v2,v1)
2712 ; CHECK-NEXT: }
2713 ; CHECK-NEXT: {
2714 ; CHECK-NEXT: v0 = vmux(q2,v3,v31)
2715 ; CHECK-NEXT: }
2716 ; CHECK-NEXT: {
2717 ; CHECK-NEXT: jumpr r31
2718 ; CHECK-NEXT: if (q3) vmem(r1+#0) = v0
2719 ; CHECK-NEXT: }
2720 %v0 = load <16 x i32>, ptr %a0, align 128
2721 %v1 = uitofp <16 x i32> %v0 to <16 x float>
2722 store <16 x float> %v1, ptr %a1, align 128
2723 ret void
2724 }
2727 attributes #0 = { "target-features"="+v68,+hvxv68,+hvx-length128b,+hvx-qfloat" }