| blob | 26a4a6743cffaeec2c7a7eef47f51f492e1f671a |
1 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
2 ; RUN: llc < %s -amdgpu-scalarize-global-loads=false -mtriple=amdgcn-- -mcpu=tahiti -verify-machineinstrs | FileCheck %s --check-prefix=SI
3 ; RUN: llc < %s -amdgpu-scalarize-global-loads=false -mtriple=amdgcn-- -mcpu=fiji -mattr=-flat-for-global -verify-machineinstrs | FileCheck %s --check-prefix=VI
4 ; RUN: llc < %s -amdgpu-scalarize-global-loads=false -mtriple=amdgcn-- -mcpu=gfx1100 -mattr=-flat-for-global -verify-machineinstrs | FileCheck %s --check-prefix=GFX11
6 define amdgpu_kernel void @select_f16(
7 ; SI-LABEL: select_f16:
8 ; SI: ; %bb.0: ; %entry
9 ; SI-NEXT: s_load_dwordx8 s[8:15], s[4:5], 0x9
10 ; SI-NEXT: s_mov_b32 s3, 0xf000
11 ; SI-NEXT: s_mov_b32 s2, -1
12 ; SI-NEXT: s_load_dwordx2 s[4:5], s[4:5], 0x11
13 ; SI-NEXT: s_mov_b32 s18, s2
14 ; SI-NEXT: s_waitcnt lgkmcnt(0)
15 ; SI-NEXT: s_mov_b32 s16, s10
16 ; SI-NEXT: s_mov_b32 s17, s11
17 ; SI-NEXT: s_mov_b32 s19, s3
18 ; SI-NEXT: s_mov_b32 s20, s12
19 ; SI-NEXT: s_mov_b32 s21, s13
20 ; SI-NEXT: s_mov_b32 s22, s2
21 ; SI-NEXT: s_mov_b32 s23, s3
22 ; SI-NEXT: s_mov_b32 s12, s14
23 ; SI-NEXT: s_mov_b32 s13, s15
24 ; SI-NEXT: s_mov_b32 s14, s2
25 ; SI-NEXT: s_mov_b32 s15, s3
26 ; SI-NEXT: s_mov_b32 s6, s2
27 ; SI-NEXT: s_mov_b32 s7, s3
28 ; SI-NEXT: buffer_load_ushort v0, off, s[16:19], 0 glc
29 ; SI-NEXT: s_waitcnt vmcnt(0)
30 ; SI-NEXT: buffer_load_ushort v1, off, s[20:23], 0 glc
31 ; SI-NEXT: s_waitcnt vmcnt(0)
32 ; SI-NEXT: buffer_load_ushort v2, off, s[12:15], 0 glc
33 ; SI-NEXT: s_waitcnt vmcnt(0)
34 ; SI-NEXT: buffer_load_ushort v3, off, s[4:7], 0 glc
35 ; SI-NEXT: s_waitcnt vmcnt(0)
36 ; SI-NEXT: s_mov_b32 s0, s8
37 ; SI-NEXT: s_mov_b32 s1, s9
38 ; SI-NEXT: v_cvt_f32_f16_e32 v0, v0
39 ; SI-NEXT: v_cvt_f32_f16_e32 v1, v1
40 ; SI-NEXT: v_cvt_f32_f16_e32 v2, v2
41 ; SI-NEXT: v_cvt_f32_f16_e32 v3, v3
42 ; SI-NEXT: v_cmp_lt_f32_e32 vcc, v0, v1
43 ; SI-NEXT: v_cndmask_b32_e32 v0, v3, v2, vcc
44 ; SI-NEXT: v_cvt_f16_f32_e32 v0, v0
45 ; SI-NEXT: buffer_store_short v0, off, s[0:3], 0
46 ; SI-NEXT: s_endpgm
47 ;
48 ; VI-LABEL: select_f16:
49 ; VI: ; %bb.0: ; %entry
50 ; VI-NEXT: s_load_dwordx8 s[8:15], s[4:5], 0x24
51 ; VI-NEXT: s_mov_b32 s3, 0xf000
52 ; VI-NEXT: s_mov_b32 s2, -1
53 ; VI-NEXT: s_load_dwordx2 s[4:5], s[4:5], 0x44
54 ; VI-NEXT: s_mov_b32 s18, s2
55 ; VI-NEXT: s_waitcnt lgkmcnt(0)
56 ; VI-NEXT: s_mov_b32 s16, s10
57 ; VI-NEXT: s_mov_b32 s17, s11
58 ; VI-NEXT: s_mov_b32 s19, s3
59 ; VI-NEXT: s_mov_b32 s20, s12
60 ; VI-NEXT: s_mov_b32 s21, s13
61 ; VI-NEXT: s_mov_b32 s22, s2
62 ; VI-NEXT: s_mov_b32 s23, s3
63 ; VI-NEXT: s_mov_b32 s12, s14
64 ; VI-NEXT: s_mov_b32 s13, s15
65 ; VI-NEXT: s_mov_b32 s14, s2
66 ; VI-NEXT: s_mov_b32 s15, s3
67 ; VI-NEXT: s_mov_b32 s6, s2
68 ; VI-NEXT: s_mov_b32 s7, s3
69 ; VI-NEXT: buffer_load_ushort v0, off, s[16:19], 0 glc
70 ; VI-NEXT: s_waitcnt vmcnt(0)
71 ; VI-NEXT: buffer_load_ushort v1, off, s[20:23], 0 glc
72 ; VI-NEXT: s_waitcnt vmcnt(0)
73 ; VI-NEXT: buffer_load_ushort v2, off, s[12:15], 0 glc
74 ; VI-NEXT: s_waitcnt vmcnt(0)
75 ; VI-NEXT: buffer_load_ushort v3, off, s[4:7], 0 glc
76 ; VI-NEXT: s_waitcnt vmcnt(0)
77 ; VI-NEXT: s_mov_b32 s0, s8
78 ; VI-NEXT: s_mov_b32 s1, s9
79 ; VI-NEXT: v_cmp_lt_f16_e32 vcc, v0, v1
80 ; VI-NEXT: v_cndmask_b32_e32 v0, v3, v2, vcc
81 ; VI-NEXT: buffer_store_short v0, off, s[0:3], 0
82 ; VI-NEXT: s_endpgm
83 ;
84 ; GFX11-LABEL: select_f16:
85 ; GFX11: ; %bb.0: ; %entry
86 ; GFX11-NEXT: s_clause 0x1
87 ; GFX11-NEXT: s_load_b256 s[8:15], s[4:5], 0x24
88 ; GFX11-NEXT: s_load_b64 s[0:1], s[4:5], 0x44
89 ; GFX11-NEXT: s_mov_b32 s6, -1
90 ; GFX11-NEXT: s_mov_b32 s7, 0x31016000
91 ; GFX11-NEXT: s_mov_b32 s18, s6
92 ; GFX11-NEXT: s_mov_b32 s19, s7
93 ; GFX11-NEXT: s_mov_b32 s22, s6
94 ; GFX11-NEXT: s_mov_b32 s23, s7
95 ; GFX11-NEXT: s_mov_b32 s26, s6
96 ; GFX11-NEXT: s_mov_b32 s27, s7
97 ; GFX11-NEXT: s_mov_b32 s2, s6
98 ; GFX11-NEXT: s_mov_b32 s3, s7
99 ; GFX11-NEXT: s_waitcnt lgkmcnt(0)
100 ; GFX11-NEXT: s_mov_b32 s16, s10
101 ; GFX11-NEXT: s_mov_b32 s17, s11
102 ; GFX11-NEXT: s_mov_b32 s20, s12
103 ; GFX11-NEXT: s_mov_b32 s21, s13
104 ; GFX11-NEXT: s_mov_b32 s24, s14
105 ; GFX11-NEXT: s_mov_b32 s25, s15
106 ; GFX11-NEXT: buffer_load_u16 v0, off, s[16:19], 0 glc dlc
107 ; GFX11-NEXT: s_waitcnt vmcnt(0)
108 ; GFX11-NEXT: buffer_load_u16 v1, off, s[20:23], 0 glc dlc
109 ; GFX11-NEXT: s_waitcnt vmcnt(0)
110 ; GFX11-NEXT: buffer_load_u16 v2, off, s[24:27], 0 glc dlc
111 ; GFX11-NEXT: s_waitcnt vmcnt(0)
112 ; GFX11-NEXT: buffer_load_u16 v3, off, s[0:3], 0 glc dlc
113 ; GFX11-NEXT: s_waitcnt vmcnt(0)
114 ; GFX11-NEXT: s_mov_b32 s4, s8
115 ; GFX11-NEXT: s_mov_b32 s5, s9
116 ; GFX11-NEXT: v_cmp_lt_f16_e32 vcc_lo, v0, v1
117 ; GFX11-NEXT: v_cndmask_b32_e32 v0, v3, v2, vcc_lo
118 ; GFX11-NEXT: buffer_store_b16 v0, off, s[4:7], 0
119 ; GFX11-NEXT: s_endpgm
121 ptr addrspace(1) %r,
122 ptr addrspace(1) %a,
123 ptr addrspace(1) %b,
124 ptr addrspace(1) %c,
125 ptr addrspace(1) %d) {
126 entry:
127 %a.val = load volatile half, ptr addrspace(1) %a
128 %b.val = load volatile half, ptr addrspace(1) %b
129 %c.val = load volatile half, ptr addrspace(1) %c
130 %d.val = load volatile half, ptr addrspace(1) %d
131 %fcmp = fcmp olt half %a.val, %b.val
132 %r.val = select i1 %fcmp, half %c.val, half %d.val
133 store half %r.val, ptr addrspace(1) %r
134 ret void
135 }
137 define amdgpu_kernel void @select_f16_imm_a(
138 ; SI-LABEL: select_f16_imm_a:
139 ; SI: ; %bb.0: ; %entry
140 ; SI-NEXT: s_load_dwordx8 s[0:7], s[4:5], 0x9
141 ; SI-NEXT: s_mov_b32 s11, 0xf000
142 ; SI-NEXT: s_mov_b32 s10, -1
143 ; SI-NEXT: s_mov_b32 s14, s10
144 ; SI-NEXT: s_mov_b32 s15, s11
145 ; SI-NEXT: s_waitcnt lgkmcnt(0)
146 ; SI-NEXT: s_mov_b32 s12, s2
147 ; SI-NEXT: s_mov_b32 s13, s3
148 ; SI-NEXT: s_mov_b32 s16, s4
149 ; SI-NEXT: s_mov_b32 s17, s5
150 ; SI-NEXT: s_mov_b32 s18, s10
151 ; SI-NEXT: s_mov_b32 s19, s11
152 ; SI-NEXT: s_mov_b32 s4, s6
153 ; SI-NEXT: s_mov_b32 s5, s7
154 ; SI-NEXT: s_mov_b32 s6, s10
155 ; SI-NEXT: s_mov_b32 s7, s11
156 ; SI-NEXT: buffer_load_ushort v0, off, s[12:15], 0 glc
157 ; SI-NEXT: s_waitcnt vmcnt(0)
158 ; SI-NEXT: buffer_load_ushort v1, off, s[16:19], 0 glc
159 ; SI-NEXT: s_waitcnt vmcnt(0)
160 ; SI-NEXT: buffer_load_ushort v2, off, s[4:7], 0 glc
161 ; SI-NEXT: s_waitcnt vmcnt(0)
162 ; SI-NEXT: s_mov_b32 s8, s0
163 ; SI-NEXT: s_mov_b32 s9, s1
164 ; SI-NEXT: v_cvt_f32_f16_e32 v0, v0
165 ; SI-NEXT: v_cvt_f32_f16_e32 v1, v1
166 ; SI-NEXT: v_cvt_f32_f16_e32 v2, v2
167 ; SI-NEXT: v_cmp_lt_f32_e32 vcc, 0.5, v0
168 ; SI-NEXT: v_cndmask_b32_e32 v0, v2, v1, vcc
169 ; SI-NEXT: v_cvt_f16_f32_e32 v0, v0
170 ; SI-NEXT: buffer_store_short v0, off, s[8:11], 0
171 ; SI-NEXT: s_endpgm
172 ;
173 ; VI-LABEL: select_f16_imm_a:
174 ; VI: ; %bb.0: ; %entry
175 ; VI-NEXT: s_load_dwordx8 s[0:7], s[4:5], 0x24
176 ; VI-NEXT: s_mov_b32 s11, 0xf000
177 ; VI-NEXT: s_mov_b32 s10, -1
178 ; VI-NEXT: s_mov_b32 s14, s10
179 ; VI-NEXT: s_mov_b32 s15, s11
180 ; VI-NEXT: s_waitcnt lgkmcnt(0)
181 ; VI-NEXT: s_mov_b32 s12, s2
182 ; VI-NEXT: s_mov_b32 s13, s3
183 ; VI-NEXT: s_mov_b32 s16, s4
184 ; VI-NEXT: s_mov_b32 s17, s5
185 ; VI-NEXT: s_mov_b32 s18, s10
186 ; VI-NEXT: s_mov_b32 s19, s11
187 ; VI-NEXT: s_mov_b32 s4, s6
188 ; VI-NEXT: s_mov_b32 s5, s7
189 ; VI-NEXT: s_mov_b32 s6, s10
190 ; VI-NEXT: s_mov_b32 s7, s11
191 ; VI-NEXT: buffer_load_ushort v0, off, s[12:15], 0 glc
192 ; VI-NEXT: s_waitcnt vmcnt(0)
193 ; VI-NEXT: buffer_load_ushort v1, off, s[16:19], 0 glc
194 ; VI-NEXT: s_waitcnt vmcnt(0)
195 ; VI-NEXT: buffer_load_ushort v2, off, s[4:7], 0 glc
196 ; VI-NEXT: s_waitcnt vmcnt(0)
197 ; VI-NEXT: s_mov_b32 s8, s0
198 ; VI-NEXT: s_mov_b32 s9, s1
199 ; VI-NEXT: v_cmp_lt_f16_e32 vcc, 0.5, v0
200 ; VI-NEXT: v_cndmask_b32_e32 v0, v2, v1, vcc
201 ; VI-NEXT: buffer_store_short v0, off, s[8:11], 0
202 ; VI-NEXT: s_endpgm
203 ;
204 ; GFX11-LABEL: select_f16_imm_a:
205 ; GFX11: ; %bb.0: ; %entry
206 ; GFX11-NEXT: s_load_b256 s[0:7], s[4:5], 0x24
207 ; GFX11-NEXT: s_mov_b32 s10, -1
208 ; GFX11-NEXT: s_mov_b32 s11, 0x31016000
209 ; GFX11-NEXT: s_mov_b32 s14, s10
210 ; GFX11-NEXT: s_mov_b32 s15, s11
211 ; GFX11-NEXT: s_mov_b32 s18, s10
212 ; GFX11-NEXT: s_mov_b32 s19, s11
213 ; GFX11-NEXT: s_mov_b32 s22, s10
214 ; GFX11-NEXT: s_mov_b32 s23, s11
215 ; GFX11-NEXT: s_waitcnt lgkmcnt(0)
216 ; GFX11-NEXT: s_mov_b32 s12, s2
217 ; GFX11-NEXT: s_mov_b32 s13, s3
218 ; GFX11-NEXT: s_mov_b32 s16, s4
219 ; GFX11-NEXT: s_mov_b32 s17, s5
220 ; GFX11-NEXT: s_mov_b32 s20, s6
221 ; GFX11-NEXT: s_mov_b32 s21, s7
222 ; GFX11-NEXT: buffer_load_u16 v0, off, s[12:15], 0 glc dlc
223 ; GFX11-NEXT: s_waitcnt vmcnt(0)
224 ; GFX11-NEXT: buffer_load_u16 v1, off, s[16:19], 0 glc dlc
225 ; GFX11-NEXT: s_waitcnt vmcnt(0)
226 ; GFX11-NEXT: buffer_load_u16 v2, off, s[20:23], 0 glc dlc
227 ; GFX11-NEXT: s_waitcnt vmcnt(0)
228 ; GFX11-NEXT: s_mov_b32 s8, s0
229 ; GFX11-NEXT: s_mov_b32 s9, s1
230 ; GFX11-NEXT: v_cmp_lt_f16_e32 vcc_lo, 0.5, v0
231 ; GFX11-NEXT: v_cndmask_b32_e32 v0, v2, v1, vcc_lo
232 ; GFX11-NEXT: buffer_store_b16 v0, off, s[8:11], 0
233 ; GFX11-NEXT: s_endpgm
234 ptr addrspace(1) %r,
235 ptr addrspace(1) %b,
236 ptr addrspace(1) %c,
237 ptr addrspace(1) %d) {
238 entry:
239 %b.val = load volatile half, ptr addrspace(1) %b
240 %c.val = load volatile half, ptr addrspace(1) %c
241 %d.val = load volatile half, ptr addrspace(1) %d
242 %fcmp = fcmp olt half 0xH3800, %b.val
243 %r.val = select i1 %fcmp, half %c.val, half %d.val
244 store half %r.val, ptr addrspace(1) %r
245 ret void
246 }
248 define amdgpu_kernel void @select_f16_imm_b(
249 ; SI-LABEL: select_f16_imm_b:
250 ; SI: ; %bb.0: ; %entry
251 ; SI-NEXT: s_load_dwordx8 s[0:7], s[4:5], 0x9
252 ; SI-NEXT: s_mov_b32 s11, 0xf000
253 ; SI-NEXT: s_mov_b32 s10, -1
254 ; SI-NEXT: s_mov_b32 s14, s10
255 ; SI-NEXT: s_mov_b32 s15, s11
256 ; SI-NEXT: s_waitcnt lgkmcnt(0)
257 ; SI-NEXT: s_mov_b32 s12, s2
258 ; SI-NEXT: s_mov_b32 s13, s3
259 ; SI-NEXT: s_mov_b32 s16, s4
260 ; SI-NEXT: s_mov_b32 s17, s5
261 ; SI-NEXT: s_mov_b32 s18, s10
262 ; SI-NEXT: s_mov_b32 s19, s11
263 ; SI-NEXT: s_mov_b32 s4, s6
264 ; SI-NEXT: s_mov_b32 s5, s7
265 ; SI-NEXT: s_mov_b32 s6, s10
266 ; SI-NEXT: s_mov_b32 s7, s11
267 ; SI-NEXT: buffer_load_ushort v0, off, s[12:15], 0 glc
268 ; SI-NEXT: s_waitcnt vmcnt(0)
269 ; SI-NEXT: buffer_load_ushort v1, off, s[16:19], 0 glc
270 ; SI-NEXT: s_waitcnt vmcnt(0)
271 ; SI-NEXT: buffer_load_ushort v2, off, s[4:7], 0 glc
272 ; SI-NEXT: s_waitcnt vmcnt(0)
273 ; SI-NEXT: s_mov_b32 s8, s0
274 ; SI-NEXT: s_mov_b32 s9, s1
275 ; SI-NEXT: v_cvt_f32_f16_e32 v0, v0
276 ; SI-NEXT: v_cvt_f32_f16_e32 v1, v1
277 ; SI-NEXT: v_cvt_f32_f16_e32 v2, v2
278 ; SI-NEXT: v_cmp_gt_f32_e32 vcc, 0.5, v0
279 ; SI-NEXT: v_cndmask_b32_e32 v0, v2, v1, vcc
280 ; SI-NEXT: v_cvt_f16_f32_e32 v0, v0
281 ; SI-NEXT: buffer_store_short v0, off, s[8:11], 0
282 ; SI-NEXT: s_endpgm
283 ;
284 ; VI-LABEL: select_f16_imm_b:
285 ; VI: ; %bb.0: ; %entry
286 ; VI-NEXT: s_load_dwordx8 s[0:7], s[4:5], 0x24
287 ; VI-NEXT: s_mov_b32 s11, 0xf000
288 ; VI-NEXT: s_mov_b32 s10, -1
289 ; VI-NEXT: s_mov_b32 s14, s10
290 ; VI-NEXT: s_mov_b32 s15, s11
291 ; VI-NEXT: s_waitcnt lgkmcnt(0)
292 ; VI-NEXT: s_mov_b32 s12, s2
293 ; VI-NEXT: s_mov_b32 s13, s3
294 ; VI-NEXT: s_mov_b32 s16, s4
295 ; VI-NEXT: s_mov_b32 s17, s5
296 ; VI-NEXT: s_mov_b32 s18, s10
297 ; VI-NEXT: s_mov_b32 s19, s11
298 ; VI-NEXT: s_mov_b32 s4, s6
299 ; VI-NEXT: s_mov_b32 s5, s7
300 ; VI-NEXT: s_mov_b32 s6, s10
301 ; VI-NEXT: s_mov_b32 s7, s11
302 ; VI-NEXT: buffer_load_ushort v0, off, s[12:15], 0 glc
303 ; VI-NEXT: s_waitcnt vmcnt(0)
304 ; VI-NEXT: buffer_load_ushort v1, off, s[16:19], 0 glc
305 ; VI-NEXT: s_waitcnt vmcnt(0)
306 ; VI-NEXT: buffer_load_ushort v2, off, s[4:7], 0 glc
307 ; VI-NEXT: s_waitcnt vmcnt(0)
308 ; VI-NEXT: s_mov_b32 s8, s0
309 ; VI-NEXT: s_mov_b32 s9, s1
310 ; VI-NEXT: v_cmp_gt_f16_e32 vcc, 0.5, v0
311 ; VI-NEXT: v_cndmask_b32_e32 v0, v2, v1, vcc
312 ; VI-NEXT: buffer_store_short v0, off, s[8:11], 0
313 ; VI-NEXT: s_endpgm
314 ;
315 ; GFX11-LABEL: select_f16_imm_b:
316 ; GFX11: ; %bb.0: ; %entry
317 ; GFX11-NEXT: s_load_b256 s[0:7], s[4:5], 0x24
318 ; GFX11-NEXT: s_mov_b32 s10, -1
319 ; GFX11-NEXT: s_mov_b32 s11, 0x31016000
320 ; GFX11-NEXT: s_mov_b32 s14, s10
321 ; GFX11-NEXT: s_mov_b32 s15, s11
322 ; GFX11-NEXT: s_mov_b32 s18, s10
323 ; GFX11-NEXT: s_mov_b32 s19, s11
324 ; GFX11-NEXT: s_mov_b32 s22, s10
325 ; GFX11-NEXT: s_mov_b32 s23, s11
326 ; GFX11-NEXT: s_waitcnt lgkmcnt(0)
327 ; GFX11-NEXT: s_mov_b32 s12, s2
328 ; GFX11-NEXT: s_mov_b32 s13, s3
329 ; GFX11-NEXT: s_mov_b32 s16, s4
330 ; GFX11-NEXT: s_mov_b32 s17, s5
331 ; GFX11-NEXT: s_mov_b32 s20, s6
332 ; GFX11-NEXT: s_mov_b32 s21, s7
333 ; GFX11-NEXT: buffer_load_u16 v0, off, s[12:15], 0 glc dlc
334 ; GFX11-NEXT: s_waitcnt vmcnt(0)
335 ; GFX11-NEXT: buffer_load_u16 v1, off, s[16:19], 0 glc dlc
336 ; GFX11-NEXT: s_waitcnt vmcnt(0)
337 ; GFX11-NEXT: buffer_load_u16 v2, off, s[20:23], 0 glc dlc
338 ; GFX11-NEXT: s_waitcnt vmcnt(0)
339 ; GFX11-NEXT: s_mov_b32 s8, s0
340 ; GFX11-NEXT: s_mov_b32 s9, s1
341 ; GFX11-NEXT: v_cmp_gt_f16_e32 vcc_lo, 0.5, v0
342 ; GFX11-NEXT: v_cndmask_b32_e32 v0, v2, v1, vcc_lo
343 ; GFX11-NEXT: buffer_store_b16 v0, off, s[8:11], 0
344 ; GFX11-NEXT: s_endpgm
345 ptr addrspace(1) %r,
346 ptr addrspace(1) %a,
347 ptr addrspace(1) %c,
348 ptr addrspace(1) %d) {
349 entry:
350 %a.val = load volatile half, ptr addrspace(1) %a
351 %c.val = load volatile half, ptr addrspace(1) %c
352 %d.val = load volatile half, ptr addrspace(1) %d
353 %fcmp = fcmp olt half %a.val, 0xH3800
354 %r.val = select i1 %fcmp, half %c.val, half %d.val
355 store half %r.val, ptr addrspace(1) %r
356 ret void
357 }
359 define amdgpu_kernel void @select_f16_imm_c(
360 ; SI-LABEL: select_f16_imm_c:
361 ; SI: ; %bb.0: ; %entry
362 ; SI-NEXT: s_load_dwordx8 s[0:7], s[4:5], 0x9
363 ; SI-NEXT: s_mov_b32 s11, 0xf000
364 ; SI-NEXT: s_mov_b32 s10, -1
365 ; SI-NEXT: s_mov_b32 s14, s10
366 ; SI-NEXT: s_mov_b32 s15, s11
367 ; SI-NEXT: s_waitcnt lgkmcnt(0)
368 ; SI-NEXT: s_mov_b32 s12, s2
369 ; SI-NEXT: s_mov_b32 s13, s3
370 ; SI-NEXT: s_mov_b32 s16, s4
371 ; SI-NEXT: s_mov_b32 s17, s5
372 ; SI-NEXT: s_mov_b32 s18, s10
373 ; SI-NEXT: s_mov_b32 s19, s11
374 ; SI-NEXT: s_mov_b32 s4, s6
375 ; SI-NEXT: s_mov_b32 s5, s7
376 ; SI-NEXT: s_mov_b32 s6, s10
377 ; SI-NEXT: s_mov_b32 s7, s11
378 ; SI-NEXT: buffer_load_ushort v0, off, s[12:15], 0 glc
379 ; SI-NEXT: s_waitcnt vmcnt(0)
380 ; SI-NEXT: buffer_load_ushort v1, off, s[16:19], 0 glc
381 ; SI-NEXT: s_waitcnt vmcnt(0)
382 ; SI-NEXT: buffer_load_ushort v2, off, s[4:7], 0 glc
383 ; SI-NEXT: s_waitcnt vmcnt(0)
384 ; SI-NEXT: s_mov_b32 s8, s0
385 ; SI-NEXT: s_mov_b32 s9, s1
386 ; SI-NEXT: v_cvt_f32_f16_e32 v0, v0
387 ; SI-NEXT: v_cvt_f32_f16_e32 v1, v1
388 ; SI-NEXT: v_cvt_f32_f16_e32 v2, v2
389 ; SI-NEXT: v_cmp_nlt_f32_e32 vcc, v0, v1
390 ; SI-NEXT: v_cndmask_b32_e32 v0, 0.5, v2, vcc
391 ; SI-NEXT: v_cvt_f16_f32_e32 v0, v0
392 ; SI-NEXT: buffer_store_short v0, off, s[8:11], 0
393 ; SI-NEXT: s_endpgm
394 ;
395 ; VI-LABEL: select_f16_imm_c:
396 ; VI: ; %bb.0: ; %entry
397 ; VI-NEXT: s_load_dwordx8 s[0:7], s[4:5], 0x24
398 ; VI-NEXT: s_mov_b32 s11, 0xf000
399 ; VI-NEXT: s_mov_b32 s10, -1
400 ; VI-NEXT: s_mov_b32 s14, s10
401 ; VI-NEXT: s_mov_b32 s15, s11
402 ; VI-NEXT: s_waitcnt lgkmcnt(0)
403 ; VI-NEXT: s_mov_b32 s12, s2
404 ; VI-NEXT: s_mov_b32 s13, s3
405 ; VI-NEXT: s_mov_b32 s16, s4
406 ; VI-NEXT: s_mov_b32 s17, s5
407 ; VI-NEXT: s_mov_b32 s18, s10
408 ; VI-NEXT: s_mov_b32 s19, s11
409 ; VI-NEXT: s_mov_b32 s4, s6
410 ; VI-NEXT: s_mov_b32 s5, s7
411 ; VI-NEXT: s_mov_b32 s6, s10
412 ; VI-NEXT: s_mov_b32 s7, s11
413 ; VI-NEXT: buffer_load_ushort v0, off, s[12:15], 0 glc
414 ; VI-NEXT: s_waitcnt vmcnt(0)
415 ; VI-NEXT: buffer_load_ushort v1, off, s[16:19], 0 glc
416 ; VI-NEXT: s_waitcnt vmcnt(0)
417 ; VI-NEXT: buffer_load_ushort v2, off, s[4:7], 0 glc
418 ; VI-NEXT: s_waitcnt vmcnt(0)
419 ; VI-NEXT: v_mov_b32_e32 v3, 0x3800
420 ; VI-NEXT: s_mov_b32 s8, s0
421 ; VI-NEXT: s_mov_b32 s9, s1
422 ; VI-NEXT: v_cmp_nlt_f16_e32 vcc, v0, v1
423 ; VI-NEXT: v_cndmask_b32_e32 v0, v3, v2, vcc
424 ; VI-NEXT: buffer_store_short v0, off, s[8:11], 0
425 ; VI-NEXT: s_endpgm
426 ;
427 ; GFX11-LABEL: select_f16_imm_c:
428 ; GFX11: ; %bb.0: ; %entry
429 ; GFX11-NEXT: s_load_b256 s[0:7], s[4:5], 0x24
430 ; GFX11-NEXT: s_mov_b32 s10, -1
431 ; GFX11-NEXT: s_mov_b32 s11, 0x31016000
432 ; GFX11-NEXT: s_mov_b32 s14, s10
433 ; GFX11-NEXT: s_mov_b32 s15, s11
434 ; GFX11-NEXT: s_mov_b32 s18, s10
435 ; GFX11-NEXT: s_mov_b32 s19, s11
436 ; GFX11-NEXT: s_mov_b32 s22, s10
437 ; GFX11-NEXT: s_mov_b32 s23, s11
438 ; GFX11-NEXT: s_waitcnt lgkmcnt(0)
439 ; GFX11-NEXT: s_mov_b32 s12, s2
440 ; GFX11-NEXT: s_mov_b32 s13, s3
441 ; GFX11-NEXT: s_mov_b32 s16, s4
442 ; GFX11-NEXT: s_mov_b32 s17, s5
443 ; GFX11-NEXT: s_mov_b32 s20, s6
444 ; GFX11-NEXT: s_mov_b32 s21, s7
445 ; GFX11-NEXT: buffer_load_u16 v0, off, s[12:15], 0 glc dlc
446 ; GFX11-NEXT: s_waitcnt vmcnt(0)
447 ; GFX11-NEXT: buffer_load_u16 v1, off, s[16:19], 0 glc dlc
448 ; GFX11-NEXT: s_waitcnt vmcnt(0)
449 ; GFX11-NEXT: buffer_load_u16 v2, off, s[20:23], 0 glc dlc
450 ; GFX11-NEXT: s_waitcnt vmcnt(0)
451 ; GFX11-NEXT: s_mov_b32 s8, s0
452 ; GFX11-NEXT: s_mov_b32 s9, s1
453 ; GFX11-NEXT: v_cmp_nlt_f16_e32 vcc_lo, v0, v1
454 ; GFX11-NEXT: v_cndmask_b32_e32 v0, 0x3800, v2, vcc_lo
455 ; GFX11-NEXT: buffer_store_b16 v0, off, s[8:11], 0
456 ; GFX11-NEXT: s_endpgm
457 ptr addrspace(1) %r,
458 ptr addrspace(1) %a,
459 ptr addrspace(1) %b,
460 ptr addrspace(1) %d) {
461 entry:
462 %a.val = load volatile half, ptr addrspace(1) %a
463 %b.val = load volatile half, ptr addrspace(1) %b
464 %d.val = load volatile half, ptr addrspace(1) %d
465 %fcmp = fcmp olt half %a.val, %b.val
466 %r.val = select i1 %fcmp, half 0xH3800, half %d.val
467 store half %r.val, ptr addrspace(1) %r
468 ret void
469 }
471 define amdgpu_kernel void @select_f16_imm_d(
472 ; SI-LABEL: select_f16_imm_d:
473 ; SI: ; %bb.0: ; %entry
474 ; SI-NEXT: s_load_dwordx8 s[0:7], s[4:5], 0x9
475 ; SI-NEXT: s_mov_b32 s11, 0xf000
476 ; SI-NEXT: s_mov_b32 s10, -1
477 ; SI-NEXT: s_mov_b32 s14, s10
478 ; SI-NEXT: s_mov_b32 s15, s11
479 ; SI-NEXT: s_waitcnt lgkmcnt(0)
480 ; SI-NEXT: s_mov_b32 s12, s2
481 ; SI-NEXT: s_mov_b32 s13, s3
482 ; SI-NEXT: s_mov_b32 s16, s4
483 ; SI-NEXT: s_mov_b32 s17, s5
484 ; SI-NEXT: s_mov_b32 s18, s10
485 ; SI-NEXT: s_mov_b32 s19, s11
486 ; SI-NEXT: s_mov_b32 s4, s6
487 ; SI-NEXT: s_mov_b32 s5, s7
488 ; SI-NEXT: s_mov_b32 s6, s10
489 ; SI-NEXT: s_mov_b32 s7, s11
490 ; SI-NEXT: buffer_load_ushort v0, off, s[12:15], 0 glc
491 ; SI-NEXT: s_waitcnt vmcnt(0)
492 ; SI-NEXT: buffer_load_ushort v1, off, s[16:19], 0 glc
493 ; SI-NEXT: s_waitcnt vmcnt(0)
494 ; SI-NEXT: buffer_load_ushort v2, off, s[4:7], 0 glc
495 ; SI-NEXT: s_waitcnt vmcnt(0)
496 ; SI-NEXT: s_mov_b32 s8, s0
497 ; SI-NEXT: s_mov_b32 s9, s1
498 ; SI-NEXT: v_cvt_f32_f16_e32 v0, v0
499 ; SI-NEXT: v_cvt_f32_f16_e32 v1, v1
500 ; SI-NEXT: v_cvt_f32_f16_e32 v2, v2
501 ; SI-NEXT: v_cmp_lt_f32_e32 vcc, v0, v1
502 ; SI-NEXT: v_cndmask_b32_e32 v0, 0.5, v2, vcc
503 ; SI-NEXT: v_cvt_f16_f32_e32 v0, v0
504 ; SI-NEXT: buffer_store_short v0, off, s[8:11], 0
505 ; SI-NEXT: s_endpgm
506 ;
507 ; VI-LABEL: select_f16_imm_d:
508 ; VI: ; %bb.0: ; %entry
509 ; VI-NEXT: s_load_dwordx8 s[0:7], s[4:5], 0x24
510 ; VI-NEXT: s_mov_b32 s11, 0xf000
511 ; VI-NEXT: s_mov_b32 s10, -1
512 ; VI-NEXT: s_mov_b32 s14, s10
513 ; VI-NEXT: s_mov_b32 s15, s11
514 ; VI-NEXT: s_waitcnt lgkmcnt(0)
515 ; VI-NEXT: s_mov_b32 s12, s2
516 ; VI-NEXT: s_mov_b32 s13, s3
517 ; VI-NEXT: s_mov_b32 s16, s4
518 ; VI-NEXT: s_mov_b32 s17, s5
519 ; VI-NEXT: s_mov_b32 s18, s10
520 ; VI-NEXT: s_mov_b32 s19, s11
521 ; VI-NEXT: s_mov_b32 s4, s6
522 ; VI-NEXT: s_mov_b32 s5, s7
523 ; VI-NEXT: s_mov_b32 s6, s10
524 ; VI-NEXT: s_mov_b32 s7, s11
525 ; VI-NEXT: buffer_load_ushort v0, off, s[12:15], 0 glc
526 ; VI-NEXT: s_waitcnt vmcnt(0)
527 ; VI-NEXT: buffer_load_ushort v1, off, s[16:19], 0 glc
528 ; VI-NEXT: s_waitcnt vmcnt(0)
529 ; VI-NEXT: buffer_load_ushort v2, off, s[4:7], 0 glc
530 ; VI-NEXT: s_waitcnt vmcnt(0)
531 ; VI-NEXT: v_mov_b32_e32 v3, 0x3800
532 ; VI-NEXT: s_mov_b32 s8, s0
533 ; VI-NEXT: s_mov_b32 s9, s1
534 ; VI-NEXT: v_cmp_lt_f16_e32 vcc, v0, v1
535 ; VI-NEXT: v_cndmask_b32_e32 v0, v3, v2, vcc
536 ; VI-NEXT: buffer_store_short v0, off, s[8:11], 0
537 ; VI-NEXT: s_endpgm
538 ;
539 ; GFX11-LABEL: select_f16_imm_d:
540 ; GFX11: ; %bb.0: ; %entry
541 ; GFX11-NEXT: s_load_b256 s[0:7], s[4:5], 0x24
542 ; GFX11-NEXT: s_mov_b32 s10, -1
543 ; GFX11-NEXT: s_mov_b32 s11, 0x31016000
544 ; GFX11-NEXT: s_mov_b32 s14, s10
545 ; GFX11-NEXT: s_mov_b32 s15, s11
546 ; GFX11-NEXT: s_mov_b32 s18, s10
547 ; GFX11-NEXT: s_mov_b32 s19, s11
548 ; GFX11-NEXT: s_mov_b32 s22, s10
549 ; GFX11-NEXT: s_mov_b32 s23, s11
550 ; GFX11-NEXT: s_waitcnt lgkmcnt(0)
551 ; GFX11-NEXT: s_mov_b32 s12, s2
552 ; GFX11-NEXT: s_mov_b32 s13, s3
553 ; GFX11-NEXT: s_mov_b32 s16, s4
554 ; GFX11-NEXT: s_mov_b32 s17, s5
555 ; GFX11-NEXT: s_mov_b32 s20, s6
556 ; GFX11-NEXT: s_mov_b32 s21, s7
557 ; GFX11-NEXT: buffer_load_u16 v0, off, s[12:15], 0 glc dlc
558 ; GFX11-NEXT: s_waitcnt vmcnt(0)
559 ; GFX11-NEXT: buffer_load_u16 v1, off, s[16:19], 0 glc dlc
560 ; GFX11-NEXT: s_waitcnt vmcnt(0)
561 ; GFX11-NEXT: buffer_load_u16 v2, off, s[20:23], 0 glc dlc
562 ; GFX11-NEXT: s_waitcnt vmcnt(0)
563 ; GFX11-NEXT: s_mov_b32 s8, s0
564 ; GFX11-NEXT: s_mov_b32 s9, s1
565 ; GFX11-NEXT: v_cmp_lt_f16_e32 vcc_lo, v0, v1
566 ; GFX11-NEXT: v_cndmask_b32_e32 v0, 0x3800, v2, vcc_lo
567 ; GFX11-NEXT: buffer_store_b16 v0, off, s[8:11], 0
568 ; GFX11-NEXT: s_endpgm
569 ptr addrspace(1) %r,
570 ptr addrspace(1) %a,
571 ptr addrspace(1) %b,
572 ptr addrspace(1) %c) {
573 entry:
574 %a.val = load volatile half, ptr addrspace(1) %a
575 %b.val = load volatile half, ptr addrspace(1) %b
576 %c.val = load volatile half, ptr addrspace(1) %c
577 %fcmp = fcmp olt half %a.val, %b.val
578 %r.val = select i1 %fcmp, half %c.val, half 0xH3800
579 store half %r.val, ptr addrspace(1) %r
580 ret void
581 }
583 define amdgpu_kernel void @select_v2f16(
584 ; SI-LABEL: select_v2f16:
585 ; SI: ; %bb.0: ; %entry
586 ; SI-NEXT: s_load_dwordx8 s[8:15], s[4:5], 0x9
587 ; SI-NEXT: s_load_dwordx2 s[4:5], s[4:5], 0x11
588 ; SI-NEXT: s_mov_b32 s3, 0xf000
589 ; SI-NEXT: s_mov_b32 s2, -1
590 ; SI-NEXT: s_mov_b32 s18, s2
591 ; SI-NEXT: s_waitcnt lgkmcnt(0)
592 ; SI-NEXT: s_mov_b32 s16, s10
593 ; SI-NEXT: s_mov_b32 s17, s11
594 ; SI-NEXT: s_mov_b32 s19, s3
595 ; SI-NEXT: s_mov_b32 s20, s12
596 ; SI-NEXT: s_mov_b32 s21, s13
597 ; SI-NEXT: s_mov_b32 s22, s2
598 ; SI-NEXT: s_mov_b32 s23, s3
599 ; SI-NEXT: s_mov_b32 s6, s2
600 ; SI-NEXT: s_mov_b32 s7, s3
601 ; SI-NEXT: buffer_load_dword v1, off, s[16:19], 0
602 ; SI-NEXT: s_mov_b32 s12, s14
603 ; SI-NEXT: s_mov_b32 s13, s15
604 ; SI-NEXT: buffer_load_dword v0, off, s[4:7], 0
605 ; SI-NEXT: s_mov_b32 s14, s2
606 ; SI-NEXT: s_mov_b32 s15, s3
607 ; SI-NEXT: buffer_load_dword v2, off, s[20:23], 0
608 ; SI-NEXT: buffer_load_dword v3, off, s[12:15], 0
609 ; SI-NEXT: s_mov_b32 s0, s8
610 ; SI-NEXT: s_mov_b32 s1, s9
611 ; SI-NEXT: s_waitcnt vmcnt(3)
612 ; SI-NEXT: v_lshrrev_b32_e32 v5, 16, v1
613 ; SI-NEXT: v_cvt_f32_f16_e32 v5, v5
614 ; SI-NEXT: v_cvt_f32_f16_e32 v1, v1
615 ; SI-NEXT: s_waitcnt vmcnt(2)
616 ; SI-NEXT: v_cvt_f32_f16_e32 v4, v0
617 ; SI-NEXT: v_lshrrev_b32_e32 v0, 16, v0
618 ; SI-NEXT: s_waitcnt vmcnt(1)
619 ; SI-NEXT: v_lshrrev_b32_e32 v6, 16, v2
620 ; SI-NEXT: s_waitcnt vmcnt(0)
621 ; SI-NEXT: v_lshrrev_b32_e32 v7, 16, v3
622 ; SI-NEXT: v_cvt_f32_f16_e32 v6, v6
623 ; SI-NEXT: v_cvt_f32_f16_e32 v0, v0
624 ; SI-NEXT: v_cvt_f32_f16_e32 v7, v7
625 ; SI-NEXT: v_cvt_f32_f16_e32 v2, v2
626 ; SI-NEXT: v_cvt_f32_f16_e32 v3, v3
627 ; SI-NEXT: v_cmp_lt_f32_e32 vcc, v5, v6
628 ; SI-NEXT: v_cndmask_b32_e32 v0, v0, v7, vcc
629 ; SI-NEXT: v_cmp_lt_f32_e32 vcc, v1, v2
630 ; SI-NEXT: v_cvt_f16_f32_e32 v0, v0
631 ; SI-NEXT: v_cndmask_b32_e32 v1, v4, v3, vcc
632 ; SI-NEXT: v_cvt_f16_f32_e32 v1, v1
633 ; SI-NEXT: v_lshlrev_b32_e32 v0, 16, v0
634 ; SI-NEXT: v_or_b32_e32 v0, v1, v0
635 ; SI-NEXT: buffer_store_dword v0, off, s[0:3], 0
636 ; SI-NEXT: s_endpgm
637 ;
638 ; VI-LABEL: select_v2f16:
639 ; VI: ; %bb.0: ; %entry
640 ; VI-NEXT: s_load_dwordx8 s[8:15], s[4:5], 0x24
641 ; VI-NEXT: s_load_dwordx2 s[4:5], s[4:5], 0x44
642 ; VI-NEXT: s_mov_b32 s3, 0xf000
643 ; VI-NEXT: s_mov_b32 s2, -1
644 ; VI-NEXT: s_mov_b32 s6, s2
645 ; VI-NEXT: s_mov_b32 s7, s3
646 ; VI-NEXT: s_waitcnt lgkmcnt(0)
647 ; VI-NEXT: s_mov_b32 s16, s10
648 ; VI-NEXT: s_mov_b32 s17, s11
649 ; VI-NEXT: s_mov_b32 s18, s2
650 ; VI-NEXT: s_mov_b32 s19, s3
651 ; VI-NEXT: s_mov_b32 s20, s12
652 ; VI-NEXT: s_mov_b32 s21, s13
653 ; VI-NEXT: s_mov_b32 s22, s2
654 ; VI-NEXT: s_mov_b32 s23, s3
655 ; VI-NEXT: s_mov_b32 s12, s14
656 ; VI-NEXT: s_mov_b32 s13, s15
657 ; VI-NEXT: s_mov_b32 s14, s2
658 ; VI-NEXT: buffer_load_dword v0, off, s[4:7], 0
659 ; VI-NEXT: buffer_load_dword v1, off, s[20:23], 0
660 ; VI-NEXT: buffer_load_dword v2, off, s[16:19], 0
661 ; VI-NEXT: s_mov_b32 s15, s3
662 ; VI-NEXT: buffer_load_dword v3, off, s[12:15], 0
663 ; VI-NEXT: s_mov_b32 s0, s8
664 ; VI-NEXT: s_mov_b32 s1, s9
665 ; VI-NEXT: s_waitcnt vmcnt(3)
666 ; VI-NEXT: v_lshrrev_b32_e32 v4, 16, v0
667 ; VI-NEXT: s_waitcnt vmcnt(2)
668 ; VI-NEXT: v_lshrrev_b32_e32 v5, 16, v1
669 ; VI-NEXT: s_waitcnt vmcnt(1)
670 ; VI-NEXT: v_lshrrev_b32_e32 v6, 16, v2
671 ; VI-NEXT: v_cmp_lt_f16_e32 vcc, v2, v1
672 ; VI-NEXT: s_waitcnt vmcnt(0)
673 ; VI-NEXT: v_cndmask_b32_e32 v0, v0, v3, vcc
674 ; VI-NEXT: v_lshrrev_b32_e32 v1, 16, v3
675 ; VI-NEXT: v_cmp_lt_f16_e32 vcc, v6, v5
676 ; VI-NEXT: v_cndmask_b32_e32 v1, v4, v1, vcc
677 ; VI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
678 ; VI-NEXT: v_or_b32_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
679 ; VI-NEXT: buffer_store_dword v0, off, s[0:3], 0
680 ; VI-NEXT: s_endpgm
681 ;
682 ; GFX11-LABEL: select_v2f16:
683 ; GFX11: ; %bb.0: ; %entry
684 ; GFX11-NEXT: s_clause 0x1
685 ; GFX11-NEXT: s_load_b256 s[8:15], s[4:5], 0x24
686 ; GFX11-NEXT: s_load_b64 s[4:5], s[4:5], 0x44
687 ; GFX11-NEXT: s_mov_b32 s2, -1
688 ; GFX11-NEXT: s_mov_b32 s3, 0x31016000
689 ; GFX11-NEXT: s_mov_b32 s6, s2
690 ; GFX11-NEXT: s_mov_b32 s7, s3
691 ; GFX11-NEXT: s_mov_b32 s22, s2
692 ; GFX11-NEXT: s_mov_b32 s23, s3
693 ; GFX11-NEXT: s_mov_b32 s18, s2
694 ; GFX11-NEXT: s_mov_b32 s19, s3
695 ; GFX11-NEXT: s_mov_b32 s26, s2
696 ; GFX11-NEXT: s_mov_b32 s27, s3
697 ; GFX11-NEXT: s_waitcnt lgkmcnt(0)
698 ; GFX11-NEXT: s_mov_b32 s20, s12
699 ; GFX11-NEXT: s_mov_b32 s21, s13
700 ; GFX11-NEXT: s_mov_b32 s16, s10
701 ; GFX11-NEXT: s_mov_b32 s17, s11
702 ; GFX11-NEXT: s_mov_b32 s24, s14
703 ; GFX11-NEXT: s_mov_b32 s25, s15
704 ; GFX11-NEXT: buffer_load_b32 v0, off, s[4:7], 0
705 ; GFX11-NEXT: buffer_load_b32 v1, off, s[20:23], 0
706 ; GFX11-NEXT: buffer_load_b32 v2, off, s[16:19], 0
707 ; GFX11-NEXT: buffer_load_b32 v3, off, s[24:27], 0
708 ; GFX11-NEXT: s_mov_b32 s0, s8
709 ; GFX11-NEXT: s_mov_b32 s1, s9
710 ; GFX11-NEXT: s_waitcnt vmcnt(3)
711 ; GFX11-NEXT: v_lshrrev_b32_e32 v4, 16, v0
712 ; GFX11-NEXT: s_waitcnt vmcnt(2)
713 ; GFX11-NEXT: v_lshrrev_b32_e32 v5, 16, v1
714 ; GFX11-NEXT: s_waitcnt vmcnt(1)
715 ; GFX11-NEXT: v_lshrrev_b32_e32 v6, 16, v2
716 ; GFX11-NEXT: v_cmp_lt_f16_e32 vcc_lo, v2, v1
717 ; GFX11-NEXT: s_waitcnt vmcnt(0)
718 ; GFX11-NEXT: v_lshrrev_b32_e32 v7, 16, v3
719 ; GFX11-NEXT: v_cndmask_b32_e32 v0, v0, v3, vcc_lo
720 ; GFX11-NEXT: v_cmp_lt_f16_e32 vcc_lo, v6, v5
721 ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
722 ; GFX11-NEXT: v_dual_cndmask_b32 v1, v4, v7 :: v_dual_and_b32 v0, 0xffff, v0
723 ; GFX11-NEXT: v_lshl_or_b32 v0, v1, 16, v0
724 ; GFX11-NEXT: buffer_store_b32 v0, off, s[0:3], 0
725 ; GFX11-NEXT: s_endpgm
726 ptr addrspace(1) %r,
727 ptr addrspace(1) %a,
728 ptr addrspace(1) %b,
729 ptr addrspace(1) %c,
730 ptr addrspace(1) %d) {
731 entry:
732 %a.val = load <2 x half>, ptr addrspace(1) %a
733 %b.val = load <2 x half>, ptr addrspace(1) %b
734 %c.val = load <2 x half>, ptr addrspace(1) %c
735 %d.val = load <2 x half>, ptr addrspace(1) %d
736 %fcmp = fcmp olt <2 x half> %a.val, %b.val
737 %r.val = select <2 x i1> %fcmp, <2 x half> %c.val, <2 x half> %d.val
738 store <2 x half> %r.val, ptr addrspace(1) %r
739 ret void
740 }
742 define amdgpu_kernel void @select_v2f16_imm_a(
743 ; SI-LABEL: select_v2f16_imm_a:
744 ; SI: ; %bb.0: ; %entry
745 ; SI-NEXT: s_load_dwordx8 s[0:7], s[4:5], 0x9
746 ; SI-NEXT: s_mov_b32 s11, 0xf000
747 ; SI-NEXT: s_mov_b32 s10, -1
748 ; SI-NEXT: s_mov_b32 s14, s10
749 ; SI-NEXT: s_mov_b32 s15, s11
750 ; SI-NEXT: s_waitcnt lgkmcnt(0)
751 ; SI-NEXT: s_mov_b32 s12, s2
752 ; SI-NEXT: s_mov_b32 s13, s3
753 ; SI-NEXT: s_mov_b32 s16, s4
754 ; SI-NEXT: s_mov_b32 s17, s5
755 ; SI-NEXT: s_mov_b32 s18, s10
756 ; SI-NEXT: s_mov_b32 s19, s11
757 ; SI-NEXT: buffer_load_dword v0, off, s[12:15], 0
758 ; SI-NEXT: s_mov_b32 s4, s6
759 ; SI-NEXT: s_mov_b32 s5, s7
760 ; SI-NEXT: s_mov_b32 s6, s10
761 ; SI-NEXT: s_mov_b32 s7, s11
762 ; SI-NEXT: buffer_load_dword v1, off, s[16:19], 0
763 ; SI-NEXT: buffer_load_dword v2, off, s[4:7], 0
764 ; SI-NEXT: s_mov_b32 s2, 0x3f200000
765 ; SI-NEXT: s_mov_b32 s8, s0
766 ; SI-NEXT: s_mov_b32 s9, s1
767 ; SI-NEXT: s_waitcnt vmcnt(2)
768 ; SI-NEXT: v_lshrrev_b32_e32 v3, 16, v0
769 ; SI-NEXT: v_cvt_f32_f16_e32 v3, v3
770 ; SI-NEXT: v_cvt_f32_f16_e32 v0, v0
771 ; SI-NEXT: s_waitcnt vmcnt(1)
772 ; SI-NEXT: v_lshrrev_b32_e32 v4, 16, v1
773 ; SI-NEXT: s_waitcnt vmcnt(0)
774 ; SI-NEXT: v_lshrrev_b32_e32 v5, 16, v2
775 ; SI-NEXT: v_cvt_f32_f16_e32 v4, v4
776 ; SI-NEXT: v_cvt_f32_f16_e32 v5, v5
777 ; SI-NEXT: v_cvt_f32_f16_e32 v1, v1
778 ; SI-NEXT: v_cvt_f32_f16_e32 v2, v2
779 ; SI-NEXT: v_cmp_lt_f32_e32 vcc, s2, v3
780 ; SI-NEXT: v_cndmask_b32_e32 v3, v5, v4, vcc
781 ; SI-NEXT: v_cmp_lt_f32_e32 vcc, 0.5, v0
782 ; SI-NEXT: v_cvt_f16_f32_e32 v3, v3
783 ; SI-NEXT: v_cndmask_b32_e32 v0, v2, v1, vcc
784 ; SI-NEXT: v_cvt_f16_f32_e32 v0, v0
785 ; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v3
786 ; SI-NEXT: v_or_b32_e32 v0, v0, v1
787 ; SI-NEXT: buffer_store_dword v0, off, s[8:11], 0
788 ; SI-NEXT: s_endpgm
789 ;
790 ; VI-LABEL: select_v2f16_imm_a:
791 ; VI: ; %bb.0: ; %entry
792 ; VI-NEXT: s_load_dwordx8 s[0:7], s[4:5], 0x24
793 ; VI-NEXT: s_mov_b32 s11, 0xf000
794 ; VI-NEXT: s_mov_b32 s10, -1
795 ; VI-NEXT: s_mov_b32 s14, s10
796 ; VI-NEXT: s_mov_b32 s15, s11
797 ; VI-NEXT: s_waitcnt lgkmcnt(0)
798 ; VI-NEXT: s_mov_b32 s12, s2
799 ; VI-NEXT: s_mov_b32 s13, s3
800 ; VI-NEXT: s_mov_b32 s16, s4
801 ; VI-NEXT: s_mov_b32 s17, s5
802 ; VI-NEXT: s_mov_b32 s18, s10
803 ; VI-NEXT: s_mov_b32 s19, s11
804 ; VI-NEXT: s_mov_b32 s4, s6
805 ; VI-NEXT: s_mov_b32 s5, s7
806 ; VI-NEXT: s_mov_b32 s6, s10
807 ; VI-NEXT: s_mov_b32 s7, s11
808 ; VI-NEXT: buffer_load_dword v0, off, s[12:15], 0
809 ; VI-NEXT: buffer_load_dword v1, off, s[16:19], 0
810 ; VI-NEXT: buffer_load_dword v2, off, s[4:7], 0
811 ; VI-NEXT: s_movk_i32 s2, 0x3900
812 ; VI-NEXT: s_mov_b32 s8, s0
813 ; VI-NEXT: s_mov_b32 s9, s1
814 ; VI-NEXT: s_waitcnt vmcnt(2)
815 ; VI-NEXT: v_lshrrev_b32_e32 v3, 16, v0
816 ; VI-NEXT: v_cmp_lt_f16_e32 vcc, 0.5, v0
817 ; VI-NEXT: s_waitcnt vmcnt(0)
818 ; VI-NEXT: v_cndmask_b32_e32 v0, v2, v1, vcc
819 ; VI-NEXT: v_lshrrev_b32_e32 v1, 16, v1
820 ; VI-NEXT: v_lshrrev_b32_e32 v2, 16, v2
821 ; VI-NEXT: v_cmp_lt_f16_e32 vcc, s2, v3
822 ; VI-NEXT: v_cndmask_b32_e32 v1, v2, v1, vcc
823 ; VI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
824 ; VI-NEXT: v_or_b32_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
825 ; VI-NEXT: buffer_store_dword v0, off, s[8:11], 0
826 ; VI-NEXT: s_endpgm
827 ;
828 ; GFX11-LABEL: select_v2f16_imm_a:
829 ; GFX11: ; %bb.0: ; %entry
830 ; GFX11-NEXT: s_load_b256 s[0:7], s[4:5], 0x24
831 ; GFX11-NEXT: s_mov_b32 s10, -1
832 ; GFX11-NEXT: s_mov_b32 s11, 0x31016000
833 ; GFX11-NEXT: s_mov_b32 s14, s10
834 ; GFX11-NEXT: s_mov_b32 s15, s11
835 ; GFX11-NEXT: s_mov_b32 s18, s10
836 ; GFX11-NEXT: s_mov_b32 s19, s11
837 ; GFX11-NEXT: s_mov_b32 s22, s10
838 ; GFX11-NEXT: s_mov_b32 s23, s11
839 ; GFX11-NEXT: s_waitcnt lgkmcnt(0)
840 ; GFX11-NEXT: s_mov_b32 s12, s2
841 ; GFX11-NEXT: s_mov_b32 s13, s3
842 ; GFX11-NEXT: s_mov_b32 s16, s4
843 ; GFX11-NEXT: s_mov_b32 s17, s5
844 ; GFX11-NEXT: s_mov_b32 s20, s6
845 ; GFX11-NEXT: s_mov_b32 s21, s7
846 ; GFX11-NEXT: buffer_load_b32 v0, off, s[12:15], 0
847 ; GFX11-NEXT: buffer_load_b32 v1, off, s[16:19], 0
848 ; GFX11-NEXT: buffer_load_b32 v2, off, s[20:23], 0
849 ; GFX11-NEXT: s_mov_b32 s8, s0
850 ; GFX11-NEXT: s_mov_b32 s9, s1
851 ; GFX11-NEXT: s_waitcnt vmcnt(2)
852 ; GFX11-NEXT: v_lshrrev_b32_e32 v3, 16, v0
853 ; GFX11-NEXT: v_cmp_lt_f16_e32 vcc_lo, 0.5, v0
854 ; GFX11-NEXT: s_waitcnt vmcnt(1)
855 ; GFX11-NEXT: v_lshrrev_b32_e32 v4, 16, v1
856 ; GFX11-NEXT: s_waitcnt vmcnt(0)
857 ; GFX11-NEXT: v_lshrrev_b32_e32 v5, 16, v2
858 ; GFX11-NEXT: v_cndmask_b32_e32 v0, v2, v1, vcc_lo
859 ; GFX11-NEXT: v_cmp_lt_f16_e32 vcc_lo, 0x3900, v3
860 ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
861 ; GFX11-NEXT: v_cndmask_b32_e32 v1, v5, v4, vcc_lo
862 ; GFX11-NEXT: v_and_b32_e32 v0, 0xffff, v0
863 ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
864 ; GFX11-NEXT: v_lshl_or_b32 v0, v1, 16, v0
865 ; GFX11-NEXT: buffer_store_b32 v0, off, s[8:11], 0
866 ; GFX11-NEXT: s_endpgm
867 ptr addrspace(1) %r,
868 ptr addrspace(1) %b,
869 ptr addrspace(1) %c,
870 ptr addrspace(1) %d) {
871 entry:
872 %b.val = load <2 x half>, ptr addrspace(1) %b
873 %c.val = load <2 x half>, ptr addrspace(1) %c
874 %d.val = load <2 x half>, ptr addrspace(1) %d
875 %fcmp = fcmp olt <2 x half> <half 0xH3800, half 0xH3900>, %b.val
876 %r.val = select <2 x i1> %fcmp, <2 x half> %c.val, <2 x half> %d.val
877 store <2 x half> %r.val, ptr addrspace(1) %r
878 ret void
879 }
881 define amdgpu_kernel void @select_v2f16_imm_b(
882 ; SI-LABEL: select_v2f16_imm_b:
883 ; SI: ; %bb.0: ; %entry
884 ; SI-NEXT: s_load_dwordx8 s[0:7], s[4:5], 0x9
885 ; SI-NEXT: s_mov_b32 s11, 0xf000
886 ; SI-NEXT: s_mov_b32 s10, -1
887 ; SI-NEXT: s_mov_b32 s14, s10
888 ; SI-NEXT: s_mov_b32 s15, s11
889 ; SI-NEXT: s_waitcnt lgkmcnt(0)
890 ; SI-NEXT: s_mov_b32 s12, s2
891 ; SI-NEXT: s_mov_b32 s13, s3
892 ; SI-NEXT: s_mov_b32 s16, s4
893 ; SI-NEXT: s_mov_b32 s17, s5
894 ; SI-NEXT: s_mov_b32 s18, s10
895 ; SI-NEXT: s_mov_b32 s19, s11
896 ; SI-NEXT: buffer_load_dword v0, off, s[12:15], 0
897 ; SI-NEXT: s_mov_b32 s4, s6
898 ; SI-NEXT: s_mov_b32 s5, s7
899 ; SI-NEXT: s_mov_b32 s6, s10
900 ; SI-NEXT: s_mov_b32 s7, s11
901 ; SI-NEXT: buffer_load_dword v1, off, s[16:19], 0
902 ; SI-NEXT: buffer_load_dword v2, off, s[4:7], 0
903 ; SI-NEXT: s_mov_b32 s2, 0x3f200000
904 ; SI-NEXT: s_mov_b32 s8, s0
905 ; SI-NEXT: s_mov_b32 s9, s1
906 ; SI-NEXT: s_waitcnt vmcnt(2)
907 ; SI-NEXT: v_lshrrev_b32_e32 v3, 16, v0
908 ; SI-NEXT: v_cvt_f32_f16_e32 v3, v3
909 ; SI-NEXT: v_cvt_f32_f16_e32 v0, v0
910 ; SI-NEXT: s_waitcnt vmcnt(1)
911 ; SI-NEXT: v_lshrrev_b32_e32 v4, 16, v1
912 ; SI-NEXT: s_waitcnt vmcnt(0)
913 ; SI-NEXT: v_lshrrev_b32_e32 v5, 16, v2
914 ; SI-NEXT: v_cvt_f32_f16_e32 v4, v4
915 ; SI-NEXT: v_cvt_f32_f16_e32 v5, v5
916 ; SI-NEXT: v_cvt_f32_f16_e32 v1, v1
917 ; SI-NEXT: v_cvt_f32_f16_e32 v2, v2
918 ; SI-NEXT: v_cmp_gt_f32_e32 vcc, s2, v3
919 ; SI-NEXT: v_cndmask_b32_e32 v3, v5, v4, vcc
920 ; SI-NEXT: v_cmp_gt_f32_e32 vcc, 0.5, v0
921 ; SI-NEXT: v_cvt_f16_f32_e32 v3, v3
922 ; SI-NEXT: v_cndmask_b32_e32 v0, v2, v1, vcc
923 ; SI-NEXT: v_cvt_f16_f32_e32 v0, v0
924 ; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v3
925 ; SI-NEXT: v_or_b32_e32 v0, v0, v1
926 ; SI-NEXT: buffer_store_dword v0, off, s[8:11], 0
927 ; SI-NEXT: s_endpgm
928 ;
929 ; VI-LABEL: select_v2f16_imm_b:
930 ; VI: ; %bb.0: ; %entry
931 ; VI-NEXT: s_load_dwordx8 s[0:7], s[4:5], 0x24
932 ; VI-NEXT: s_mov_b32 s11, 0xf000
933 ; VI-NEXT: s_mov_b32 s10, -1
934 ; VI-NEXT: s_mov_b32 s14, s10
935 ; VI-NEXT: s_mov_b32 s15, s11
936 ; VI-NEXT: s_waitcnt lgkmcnt(0)
937 ; VI-NEXT: s_mov_b32 s12, s2
938 ; VI-NEXT: s_mov_b32 s13, s3
939 ; VI-NEXT: s_mov_b32 s16, s4
940 ; VI-NEXT: s_mov_b32 s17, s5
941 ; VI-NEXT: s_mov_b32 s18, s10
942 ; VI-NEXT: s_mov_b32 s19, s11
943 ; VI-NEXT: s_mov_b32 s4, s6
944 ; VI-NEXT: s_mov_b32 s5, s7
945 ; VI-NEXT: s_mov_b32 s6, s10
946 ; VI-NEXT: s_mov_b32 s7, s11
947 ; VI-NEXT: buffer_load_dword v0, off, s[12:15], 0
948 ; VI-NEXT: buffer_load_dword v1, off, s[16:19], 0
949 ; VI-NEXT: buffer_load_dword v2, off, s[4:7], 0
950 ; VI-NEXT: s_movk_i32 s2, 0x3900
951 ; VI-NEXT: s_mov_b32 s8, s0
952 ; VI-NEXT: s_mov_b32 s9, s1
953 ; VI-NEXT: s_waitcnt vmcnt(2)
954 ; VI-NEXT: v_lshrrev_b32_e32 v3, 16, v0
955 ; VI-NEXT: v_cmp_gt_f16_e32 vcc, 0.5, v0
956 ; VI-NEXT: s_waitcnt vmcnt(0)
957 ; VI-NEXT: v_cndmask_b32_e32 v0, v2, v1, vcc
958 ; VI-NEXT: v_lshrrev_b32_e32 v1, 16, v1
959 ; VI-NEXT: v_lshrrev_b32_e32 v2, 16, v2
960 ; VI-NEXT: v_cmp_gt_f16_e32 vcc, s2, v3
961 ; VI-NEXT: v_cndmask_b32_e32 v1, v2, v1, vcc
962 ; VI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
963 ; VI-NEXT: v_or_b32_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
964 ; VI-NEXT: buffer_store_dword v0, off, s[8:11], 0
965 ; VI-NEXT: s_endpgm
966 ;
967 ; GFX11-LABEL: select_v2f16_imm_b:
968 ; GFX11: ; %bb.0: ; %entry
969 ; GFX11-NEXT: s_load_b256 s[0:7], s[4:5], 0x24
970 ; GFX11-NEXT: s_mov_b32 s10, -1
971 ; GFX11-NEXT: s_mov_b32 s11, 0x31016000
972 ; GFX11-NEXT: s_mov_b32 s14, s10
973 ; GFX11-NEXT: s_mov_b32 s15, s11
974 ; GFX11-NEXT: s_mov_b32 s18, s10
975 ; GFX11-NEXT: s_mov_b32 s19, s11
976 ; GFX11-NEXT: s_mov_b32 s22, s10
977 ; GFX11-NEXT: s_mov_b32 s23, s11
978 ; GFX11-NEXT: s_waitcnt lgkmcnt(0)
979 ; GFX11-NEXT: s_mov_b32 s12, s2
980 ; GFX11-NEXT: s_mov_b32 s13, s3
981 ; GFX11-NEXT: s_mov_b32 s16, s4
982 ; GFX11-NEXT: s_mov_b32 s17, s5
983 ; GFX11-NEXT: s_mov_b32 s20, s6
984 ; GFX11-NEXT: s_mov_b32 s21, s7
985 ; GFX11-NEXT: buffer_load_b32 v0, off, s[12:15], 0
986 ; GFX11-NEXT: buffer_load_b32 v1, off, s[16:19], 0
987 ; GFX11-NEXT: buffer_load_b32 v2, off, s[20:23], 0
988 ; GFX11-NEXT: s_mov_b32 s8, s0
989 ; GFX11-NEXT: s_mov_b32 s9, s1
990 ; GFX11-NEXT: s_waitcnt vmcnt(2)
991 ; GFX11-NEXT: v_lshrrev_b32_e32 v3, 16, v0
992 ; GFX11-NEXT: v_cmp_gt_f16_e32 vcc_lo, 0.5, v0
993 ; GFX11-NEXT: s_waitcnt vmcnt(1)
994 ; GFX11-NEXT: v_lshrrev_b32_e32 v4, 16, v1
995 ; GFX11-NEXT: s_waitcnt vmcnt(0)
996 ; GFX11-NEXT: v_lshrrev_b32_e32 v5, 16, v2
997 ; GFX11-NEXT: v_cndmask_b32_e32 v0, v2, v1, vcc_lo
998 ; GFX11-NEXT: v_cmp_gt_f16_e32 vcc_lo, 0x3900, v3
999 ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
1000 ; GFX11-NEXT: v_cndmask_b32_e32 v1, v5, v4, vcc_lo
1001 ; GFX11-NEXT: v_and_b32_e32 v0, 0xffff, v0
1002 ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
1003 ; GFX11-NEXT: v_lshl_or_b32 v0, v1, 16, v0
1004 ; GFX11-NEXT: buffer_store_b32 v0, off, s[8:11], 0
1005 ; GFX11-NEXT: s_endpgm
1006 ptr addrspace(1) %r,
1007 ptr addrspace(1) %a,
1008 ptr addrspace(1) %c,
1009 ptr addrspace(1) %d) {
1010 entry:
1011 %a.val = load <2 x half>, ptr addrspace(1) %a
1012 %c.val = load <2 x half>, ptr addrspace(1) %c
1013 %d.val = load <2 x half>, ptr addrspace(1) %d
1014 %fcmp = fcmp olt <2 x half> %a.val, <half 0xH3800, half 0xH3900>
1015 %r.val = select <2 x i1> %fcmp, <2 x half> %c.val, <2 x half> %d.val
1016 store <2 x half> %r.val, ptr addrspace(1) %r
1017 ret void
1018 }
1020 define amdgpu_kernel void @select_v2f16_imm_c(
1021 ; SI-LABEL: select_v2f16_imm_c:
1022 ; SI: ; %bb.0: ; %entry
1023 ; SI-NEXT: s_load_dwordx8 s[0:7], s[4:5], 0x9
1024 ; SI-NEXT: s_mov_b32 s11, 0xf000
1025 ; SI-NEXT: s_mov_b32 s10, -1
1026 ; SI-NEXT: s_mov_b32 s14, s10
1027 ; SI-NEXT: s_mov_b32 s15, s11
1028 ; SI-NEXT: s_waitcnt lgkmcnt(0)
1029 ; SI-NEXT: s_mov_b32 s12, s2
1030 ; SI-NEXT: s_mov_b32 s13, s3
1031 ; SI-NEXT: s_mov_b32 s16, s4
1032 ; SI-NEXT: s_mov_b32 s17, s5
1033 ; SI-NEXT: s_mov_b32 s18, s10
1034 ; SI-NEXT: s_mov_b32 s19, s11
1035 ; SI-NEXT: buffer_load_dword v0, off, s[12:15], 0
1036 ; SI-NEXT: s_mov_b32 s4, s6
1037 ; SI-NEXT: s_mov_b32 s5, s7
1038 ; SI-NEXT: s_mov_b32 s6, s10
1039 ; SI-NEXT: s_mov_b32 s7, s11
1040 ; SI-NEXT: buffer_load_dword v1, off, s[16:19], 0
1041 ; SI-NEXT: buffer_load_dword v2, off, s[4:7], 0
1042 ; SI-NEXT: v_mov_b32_e32 v3, 0x3f200000
1043 ; SI-NEXT: s_mov_b32 s8, s0
1044 ; SI-NEXT: s_mov_b32 s9, s1
1045 ; SI-NEXT: s_waitcnt vmcnt(2)
1046 ; SI-NEXT: v_cvt_f32_f16_e32 v4, v0
1047 ; SI-NEXT: v_lshrrev_b32_e32 v0, 16, v0
1048 ; SI-NEXT: v_cvt_f32_f16_e32 v0, v0
1049 ; SI-NEXT: s_waitcnt vmcnt(1)
1050 ; SI-NEXT: v_lshrrev_b32_e32 v5, 16, v1
1051 ; SI-NEXT: s_waitcnt vmcnt(0)
1052 ; SI-NEXT: v_lshrrev_b32_e32 v6, 16, v2
1053 ; SI-NEXT: v_cvt_f32_f16_e32 v5, v5
1054 ; SI-NEXT: v_cvt_f32_f16_e32 v6, v6
1055 ; SI-NEXT: v_cvt_f32_f16_e32 v1, v1
1056 ; SI-NEXT: v_cvt_f32_f16_e32 v2, v2
1057 ; SI-NEXT: v_cmp_nlt_f32_e32 vcc, v0, v5
1058 ; SI-NEXT: v_cndmask_b32_e32 v0, v3, v6, vcc
1059 ; SI-NEXT: v_cmp_nlt_f32_e32 vcc, v4, v1
1060 ; SI-NEXT: v_cvt_f16_f32_e32 v0, v0
1061 ; SI-NEXT: v_cndmask_b32_e32 v1, 0.5, v2, vcc
1062 ; SI-NEXT: v_cvt_f16_f32_e32 v1, v1
1063 ; SI-NEXT: v_lshlrev_b32_e32 v0, 16, v0
1064 ; SI-NEXT: v_or_b32_e32 v0, v1, v0
1065 ; SI-NEXT: buffer_store_dword v0, off, s[8:11], 0
1066 ; SI-NEXT: s_endpgm
1067 ;
1068 ; VI-LABEL: select_v2f16_imm_c:
1069 ; VI: ; %bb.0: ; %entry
1070 ; VI-NEXT: s_load_dwordx8 s[0:7], s[4:5], 0x24
1071 ; VI-NEXT: s_mov_b32 s11, 0xf000
1072 ; VI-NEXT: s_mov_b32 s10, -1
1073 ; VI-NEXT: s_mov_b32 s18, s10
1074 ; VI-NEXT: s_mov_b32 s19, s11
1075 ; VI-NEXT: s_waitcnt lgkmcnt(0)
1076 ; VI-NEXT: s_mov_b32 s16, s4
1077 ; VI-NEXT: s_mov_b32 s17, s5
1078 ; VI-NEXT: s_mov_b32 s14, s10
1079 ; VI-NEXT: s_mov_b32 s12, s2
1080 ; VI-NEXT: s_mov_b32 s13, s3
1081 ; VI-NEXT: s_mov_b32 s15, s11
1082 ; VI-NEXT: s_mov_b32 s4, s6
1083 ; VI-NEXT: s_mov_b32 s5, s7
1084 ; VI-NEXT: s_mov_b32 s6, s10
1085 ; VI-NEXT: buffer_load_dword v0, off, s[16:19], 0
1086 ; VI-NEXT: buffer_load_dword v1, off, s[12:15], 0
1087 ; VI-NEXT: s_mov_b32 s7, s11
1088 ; VI-NEXT: buffer_load_dword v2, off, s[4:7], 0
1089 ; VI-NEXT: v_mov_b32_e32 v3, 0x3800
1090 ; VI-NEXT: v_mov_b32_e32 v4, 0x3900
1091 ; VI-NEXT: s_mov_b32 s8, s0
1092 ; VI-NEXT: s_mov_b32 s9, s1
1093 ; VI-NEXT: s_waitcnt vmcnt(2)
1094 ; VI-NEXT: v_lshrrev_b32_e32 v5, 16, v0
1095 ; VI-NEXT: s_waitcnt vmcnt(1)
1096 ; VI-NEXT: v_lshrrev_b32_e32 v6, 16, v1
1097 ; VI-NEXT: v_cmp_nlt_f16_e32 vcc, v1, v0
1098 ; VI-NEXT: s_waitcnt vmcnt(0)
1099 ; VI-NEXT: v_cndmask_b32_e32 v0, v3, v2, vcc
1100 ; VI-NEXT: v_lshrrev_b32_e32 v1, 16, v2
1101 ; VI-NEXT: v_cmp_nlt_f16_e32 vcc, v6, v5
1102 ; VI-NEXT: v_cndmask_b32_e32 v1, v4, v1, vcc
1103 ; VI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
1104 ; VI-NEXT: v_or_b32_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
1105 ; VI-NEXT: buffer_store_dword v0, off, s[8:11], 0
1106 ; VI-NEXT: s_endpgm
1107 ;
1108 ; GFX11-LABEL: select_v2f16_imm_c:
1109 ; GFX11: ; %bb.0: ; %entry
1110 ; GFX11-NEXT: s_load_b256 s[0:7], s[4:5], 0x24
1111 ; GFX11-NEXT: s_mov_b32 s10, -1
1112 ; GFX11-NEXT: s_mov_b32 s11, 0x31016000
1113 ; GFX11-NEXT: s_mov_b32 s18, s10
1114 ; GFX11-NEXT: s_mov_b32 s19, s11
1115 ; GFX11-NEXT: s_mov_b32 s14, s10
1116 ; GFX11-NEXT: s_mov_b32 s15, s11
1117 ; GFX11-NEXT: s_mov_b32 s22, s10
1118 ; GFX11-NEXT: s_mov_b32 s23, s11
1119 ; GFX11-NEXT: s_waitcnt lgkmcnt(0)
1120 ; GFX11-NEXT: s_mov_b32 s16, s4
1121 ; GFX11-NEXT: s_mov_b32 s17, s5
1122 ; GFX11-NEXT: s_mov_b32 s12, s2
1123 ; GFX11-NEXT: s_mov_b32 s13, s3
1124 ; GFX11-NEXT: s_mov_b32 s20, s6
1125 ; GFX11-NEXT: s_mov_b32 s21, s7
1126 ; GFX11-NEXT: buffer_load_b32 v0, off, s[16:19], 0
1127 ; GFX11-NEXT: buffer_load_b32 v1, off, s[12:15], 0
1128 ; GFX11-NEXT: buffer_load_b32 v2, off, s[20:23], 0
1129 ; GFX11-NEXT: s_mov_b32 s8, s0
1130 ; GFX11-NEXT: s_mov_b32 s9, s1
1131 ; GFX11-NEXT: s_waitcnt vmcnt(2)
1132 ; GFX11-NEXT: v_lshrrev_b32_e32 v3, 16, v0
1133 ; GFX11-NEXT: s_waitcnt vmcnt(1)
1134 ; GFX11-NEXT: v_lshrrev_b32_e32 v4, 16, v1
1135 ; GFX11-NEXT: v_cmp_nlt_f16_e32 vcc_lo, v1, v0
1136 ; GFX11-NEXT: s_waitcnt vmcnt(0)
1137 ; GFX11-NEXT: v_lshrrev_b32_e32 v5, 16, v2
1138 ; GFX11-NEXT: v_cndmask_b32_e32 v0, 0x3800, v2, vcc_lo
1139 ; GFX11-NEXT: v_cmp_nlt_f16_e32 vcc_lo, v4, v3
1140 ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_4)
1141 ; GFX11-NEXT: v_and_b32_e32 v0, 0xffff, v0
1142 ; GFX11-NEXT: v_cndmask_b32_e32 v1, 0x3900, v5, vcc_lo
1143 ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
1144 ; GFX11-NEXT: v_lshl_or_b32 v0, v1, 16, v0
1145 ; GFX11-NEXT: buffer_store_b32 v0, off, s[8:11], 0
1146 ; GFX11-NEXT: s_endpgm
1147 ptr addrspace(1) %r,
1148 ptr addrspace(1) %a,
1149 ptr addrspace(1) %b,
1150 ptr addrspace(1) %d) {
1151 entry:
1152 %a.val = load <2 x half>, ptr addrspace(1) %a
1153 %b.val = load <2 x half>, ptr addrspace(1) %b
1154 %d.val = load <2 x half>, ptr addrspace(1) %d
1155 %fcmp = fcmp olt <2 x half> %a.val, %b.val
1156 %r.val = select <2 x i1> %fcmp, <2 x half> <half 0xH3800, half 0xH3900>, <2 x half> %d.val
1157 store <2 x half> %r.val, ptr addrspace(1) %r
1158 ret void
1159 }
1161 define amdgpu_kernel void @select_v2f16_imm_d(
1162 ; SI-LABEL: select_v2f16_imm_d:
1163 ; SI: ; %bb.0: ; %entry
1164 ; SI-NEXT: s_load_dwordx8 s[0:7], s[4:5], 0x9
1165 ; SI-NEXT: s_mov_b32 s11, 0xf000
1166 ; SI-NEXT: s_mov_b32 s10, -1
1167 ; SI-NEXT: s_mov_b32 s14, s10
1168 ; SI-NEXT: s_mov_b32 s15, s11
1169 ; SI-NEXT: s_waitcnt lgkmcnt(0)
1170 ; SI-NEXT: s_mov_b32 s12, s2
1171 ; SI-NEXT: s_mov_b32 s13, s3
1172 ; SI-NEXT: s_mov_b32 s16, s4
1173 ; SI-NEXT: s_mov_b32 s17, s5
1174 ; SI-NEXT: s_mov_b32 s18, s10
1175 ; SI-NEXT: s_mov_b32 s19, s11
1176 ; SI-NEXT: buffer_load_dword v0, off, s[12:15], 0
1177 ; SI-NEXT: s_mov_b32 s4, s6
1178 ; SI-NEXT: s_mov_b32 s5, s7
1179 ; SI-NEXT: s_mov_b32 s6, s10
1180 ; SI-NEXT: s_mov_b32 s7, s11
1181 ; SI-NEXT: buffer_load_dword v1, off, s[16:19], 0
1182 ; SI-NEXT: buffer_load_dword v2, off, s[4:7], 0
1183 ; SI-NEXT: v_mov_b32_e32 v3, 0x3f200000
1184 ; SI-NEXT: s_mov_b32 s8, s0
1185 ; SI-NEXT: s_mov_b32 s9, s1
1186 ; SI-NEXT: s_waitcnt vmcnt(2)
1187 ; SI-NEXT: v_lshrrev_b32_e32 v4, 16, v0
1188 ; SI-NEXT: v_cvt_f32_f16_e32 v4, v4
1189 ; SI-NEXT: v_cvt_f32_f16_e32 v0, v0
1190 ; SI-NEXT: s_waitcnt vmcnt(1)
1191 ; SI-NEXT: v_lshrrev_b32_e32 v5, 16, v1
1192 ; SI-NEXT: s_waitcnt vmcnt(0)
1193 ; SI-NEXT: v_lshrrev_b32_e32 v6, 16, v2
1194 ; SI-NEXT: v_cvt_f32_f16_e32 v5, v5
1195 ; SI-NEXT: v_cvt_f32_f16_e32 v6, v6
1196 ; SI-NEXT: v_cvt_f32_f16_e32 v1, v1
1197 ; SI-NEXT: v_cvt_f32_f16_e32 v2, v2
1198 ; SI-NEXT: v_cmp_lt_f32_e32 vcc, v4, v5
1199 ; SI-NEXT: v_cndmask_b32_e32 v3, v3, v6, vcc
1200 ; SI-NEXT: v_cmp_lt_f32_e32 vcc, v0, v1
1201 ; SI-NEXT: v_cvt_f16_f32_e32 v3, v3
1202 ; SI-NEXT: v_cndmask_b32_e32 v0, 0.5, v2, vcc
1203 ; SI-NEXT: v_cvt_f16_f32_e32 v0, v0
1204 ; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v3
1205 ; SI-NEXT: v_or_b32_e32 v0, v0, v1
1206 ; SI-NEXT: buffer_store_dword v0, off, s[8:11], 0
1207 ; SI-NEXT: s_endpgm
1208 ;
1209 ; VI-LABEL: select_v2f16_imm_d:
1210 ; VI: ; %bb.0: ; %entry
1211 ; VI-NEXT: s_load_dwordx8 s[0:7], s[4:5], 0x24
1212 ; VI-NEXT: s_mov_b32 s11, 0xf000
1213 ; VI-NEXT: s_mov_b32 s10, -1
1214 ; VI-NEXT: s_mov_b32 s18, s10
1215 ; VI-NEXT: s_mov_b32 s19, s11
1216 ; VI-NEXT: s_waitcnt lgkmcnt(0)
1217 ; VI-NEXT: s_mov_b32 s16, s4
1218 ; VI-NEXT: s_mov_b32 s17, s5
1219 ; VI-NEXT: s_mov_b32 s14, s10
1220 ; VI-NEXT: s_mov_b32 s12, s2
1221 ; VI-NEXT: s_mov_b32 s13, s3
1222 ; VI-NEXT: s_mov_b32 s15, s11
1223 ; VI-NEXT: s_mov_b32 s4, s6
1224 ; VI-NEXT: s_mov_b32 s5, s7
1225 ; VI-NEXT: s_mov_b32 s6, s10
1226 ; VI-NEXT: buffer_load_dword v0, off, s[16:19], 0
1227 ; VI-NEXT: buffer_load_dword v1, off, s[12:15], 0
1228 ; VI-NEXT: s_mov_b32 s7, s11
1229 ; VI-NEXT: buffer_load_dword v2, off, s[4:7], 0
1230 ; VI-NEXT: v_mov_b32_e32 v3, 0x3800
1231 ; VI-NEXT: v_mov_b32_e32 v4, 0x3900
1232 ; VI-NEXT: s_mov_b32 s8, s0
1233 ; VI-NEXT: s_mov_b32 s9, s1
1234 ; VI-NEXT: s_waitcnt vmcnt(2)
1235 ; VI-NEXT: v_lshrrev_b32_e32 v5, 16, v0
1236 ; VI-NEXT: s_waitcnt vmcnt(1)
1237 ; VI-NEXT: v_lshrrev_b32_e32 v6, 16, v1
1238 ; VI-NEXT: v_cmp_lt_f16_e32 vcc, v1, v0
1239 ; VI-NEXT: s_waitcnt vmcnt(0)
1240 ; VI-NEXT: v_cndmask_b32_e32 v0, v3, v2, vcc
1241 ; VI-NEXT: v_lshrrev_b32_e32 v1, 16, v2
1242 ; VI-NEXT: v_cmp_lt_f16_e32 vcc, v6, v5
1243 ; VI-NEXT: v_cndmask_b32_e32 v1, v4, v1, vcc
1244 ; VI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
1245 ; VI-NEXT: v_or_b32_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
1246 ; VI-NEXT: buffer_store_dword v0, off, s[8:11], 0
1247 ; VI-NEXT: s_endpgm
1248 ;
1249 ; GFX11-LABEL: select_v2f16_imm_d:
1250 ; GFX11: ; %bb.0: ; %entry
1251 ; GFX11-NEXT: s_load_b256 s[0:7], s[4:5], 0x24
1252 ; GFX11-NEXT: s_mov_b32 s10, -1
1253 ; GFX11-NEXT: s_mov_b32 s11, 0x31016000
1254 ; GFX11-NEXT: s_mov_b32 s18, s10
1255 ; GFX11-NEXT: s_mov_b32 s19, s11
1256 ; GFX11-NEXT: s_mov_b32 s14, s10
1257 ; GFX11-NEXT: s_mov_b32 s15, s11
1258 ; GFX11-NEXT: s_mov_b32 s22, s10
1259 ; GFX11-NEXT: s_mov_b32 s23, s11
1260 ; GFX11-NEXT: s_waitcnt lgkmcnt(0)
1261 ; GFX11-NEXT: s_mov_b32 s16, s4
1262 ; GFX11-NEXT: s_mov_b32 s17, s5
1263 ; GFX11-NEXT: s_mov_b32 s12, s2
1264 ; GFX11-NEXT: s_mov_b32 s13, s3
1265 ; GFX11-NEXT: s_mov_b32 s20, s6
1266 ; GFX11-NEXT: s_mov_b32 s21, s7
1267 ; GFX11-NEXT: buffer_load_b32 v0, off, s[16:19], 0
1268 ; GFX11-NEXT: buffer_load_b32 v1, off, s[12:15], 0
1269 ; GFX11-NEXT: buffer_load_b32 v2, off, s[20:23], 0
1270 ; GFX11-NEXT: s_mov_b32 s8, s0
1271 ; GFX11-NEXT: s_mov_b32 s9, s1
1272 ; GFX11-NEXT: s_waitcnt vmcnt(2)
1273 ; GFX11-NEXT: v_lshrrev_b32_e32 v3, 16, v0
1274 ; GFX11-NEXT: s_waitcnt vmcnt(1)
1275 ; GFX11-NEXT: v_lshrrev_b32_e32 v4, 16, v1
1276 ; GFX11-NEXT: v_cmp_lt_f16_e32 vcc_lo, v1, v0
1277 ; GFX11-NEXT: s_waitcnt vmcnt(0)
1278 ; GFX11-NEXT: v_lshrrev_b32_e32 v5, 16, v2
1279 ; GFX11-NEXT: v_cndmask_b32_e32 v0, 0x3800, v2, vcc_lo
1280 ; GFX11-NEXT: v_cmp_lt_f16_e32 vcc_lo, v4, v3
1281 ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_4)
1282 ; GFX11-NEXT: v_and_b32_e32 v0, 0xffff, v0
1283 ; GFX11-NEXT: v_cndmask_b32_e32 v1, 0x3900, v5, vcc_lo
1284 ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
1285 ; GFX11-NEXT: v_lshl_or_b32 v0, v1, 16, v0
1286 ; GFX11-NEXT: buffer_store_b32 v0, off, s[8:11], 0
1287 ; GFX11-NEXT: s_endpgm
1288 ptr addrspace(1) %r,
1289 ptr addrspace(1) %a,
1290 ptr addrspace(1) %b,
1291 ptr addrspace(1) %c) {
1292 entry:
1293 %a.val = load <2 x half>, ptr addrspace(1) %a
1294 %b.val = load <2 x half>, ptr addrspace(1) %b
1295 %c.val = load <2 x half>, ptr addrspace(1) %c
1296 %fcmp = fcmp olt <2 x half> %a.val, %b.val
1297 %r.val = select <2 x i1> %fcmp, <2 x half> %c.val, <2 x half> <half 0xH3800, half 0xH3900>
1298 store <2 x half> %r.val, ptr addrspace(1) %r
1299 ret void
1300 }
1302 define <4 x half> @v_select_v4f16(<4 x half> %a, <4 x half> %b, i32 %cond) {
1303 ; SI-LABEL: v_select_v4f16:
1304 ; SI: ; %bb.0:
1305 ; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1306 ; SI-NEXT: v_cvt_f16_f32_e32 v3, v3
1307 ; SI-NEXT: v_cvt_f16_f32_e32 v1, v1
1308 ; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
1309 ; SI-NEXT: v_cvt_f16_f32_e32 v7, v7
1310 ; SI-NEXT: v_cvt_f16_f32_e32 v0, v0
1311 ; SI-NEXT: v_cvt_f16_f32_e32 v5, v5
1312 ; SI-NEXT: v_cvt_f16_f32_e32 v6, v6
1313 ; SI-NEXT: v_cvt_f16_f32_e32 v4, v4
1314 ; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v3
1315 ; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
1316 ; SI-NEXT: v_or_b32_e32 v2, v2, v3
1317 ; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v7
1318 ; SI-NEXT: v_or_b32_e32 v0, v0, v1
1319 ; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v5
1320 ; SI-NEXT: v_or_b32_e32 v3, v6, v3
1321 ; SI-NEXT: v_or_b32_e32 v1, v4, v1
1322 ; SI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v8
1323 ; SI-NEXT: v_cndmask_b32_e32 v1, v1, v0, vcc
1324 ; SI-NEXT: v_cndmask_b32_e32 v3, v3, v2, vcc
1325 ; SI-NEXT: v_cvt_f32_f16_e32 v0, v1
1326 ; SI-NEXT: v_cvt_f32_f16_e32 v2, v3
1327 ; SI-NEXT: v_lshrrev_b32_e32 v1, 16, v1
1328 ; SI-NEXT: v_lshrrev_b32_e32 v3, 16, v3
1329 ; SI-NEXT: v_cvt_f32_f16_e32 v1, v1
1330 ; SI-NEXT: v_cvt_f32_f16_e32 v3, v3
1331 ; SI-NEXT: s_setpc_b64 s[30:31]
1332 ;
1333 ; VI-LABEL: v_select_v4f16:
1334 ; VI: ; %bb.0:
1335 ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1336 ; VI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v4
1337 ; VI-NEXT: v_cndmask_b32_e32 v0, v2, v0, vcc
1338 ; VI-NEXT: v_cndmask_b32_e32 v1, v3, v1, vcc
1339 ; VI-NEXT: s_setpc_b64 s[30:31]
1340 ;
1341 ; GFX11-LABEL: v_select_v4f16:
1342 ; GFX11: ; %bb.0:
1343 ; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1344 ; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v4
1345 ; GFX11-NEXT: v_dual_cndmask_b32 v0, v2, v0 :: v_dual_cndmask_b32 v1, v3, v1
1346 ; GFX11-NEXT: s_setpc_b64 s[30:31]
1347 %cmp = icmp eq i32 %cond, 0
1348 %select = select i1 %cmp, <4 x half> %a, <4 x half> %b
1349 ret <4 x half> %select
1350 }
1352 define <4 x half> @v_vselect_v4f16(<4 x half> %a, <4 x half> %b, <4 x i32> %cond) {
1353 ; SI-LABEL: v_vselect_v4f16:
1354 ; SI: ; %bb.0:
1355 ; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1356 ; SI-NEXT: v_cvt_f16_f32_e32 v0, v0
1357 ; SI-NEXT: v_cvt_f16_f32_e32 v4, v4
1358 ; SI-NEXT: v_cvt_f16_f32_e32 v1, v1
1359 ; SI-NEXT: v_cvt_f16_f32_e32 v5, v5
1360 ; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
1361 ; SI-NEXT: v_cvt_f16_f32_e32 v6, v6
1362 ; SI-NEXT: v_cvt_f16_f32_e32 v3, v3
1363 ; SI-NEXT: v_cvt_f16_f32_e32 v7, v7
1364 ; SI-NEXT: v_cvt_f32_f16_e32 v0, v0
1365 ; SI-NEXT: v_cvt_f32_f16_e32 v4, v4
1366 ; SI-NEXT: v_cvt_f32_f16_e32 v1, v1
1367 ; SI-NEXT: v_cvt_f32_f16_e32 v5, v5
1368 ; SI-NEXT: v_cvt_f32_f16_e32 v2, v2
1369 ; SI-NEXT: v_cvt_f32_f16_e32 v6, v6
1370 ; SI-NEXT: v_cvt_f32_f16_e32 v3, v3
1371 ; SI-NEXT: v_cvt_f32_f16_e32 v7, v7
1372 ; SI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v8
1373 ; SI-NEXT: v_cndmask_b32_e32 v0, v4, v0, vcc
1374 ; SI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v9
1375 ; SI-NEXT: v_cndmask_b32_e32 v1, v5, v1, vcc
1376 ; SI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v10
1377 ; SI-NEXT: v_cndmask_b32_e32 v2, v6, v2, vcc
1378 ; SI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v11
1379 ; SI-NEXT: v_cndmask_b32_e32 v3, v7, v3, vcc
1380 ; SI-NEXT: s_setpc_b64 s[30:31]
1381 ;
1382 ; VI-LABEL: v_vselect_v4f16:
1383 ; VI: ; %bb.0:
1384 ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1385 ; VI-NEXT: v_lshrrev_b32_e32 v8, 16, v1
1386 ; VI-NEXT: v_lshrrev_b32_e32 v9, 16, v3
1387 ; VI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v7
1388 ; VI-NEXT: v_cndmask_b32_e32 v7, v9, v8, vcc
1389 ; VI-NEXT: v_lshrrev_b32_e32 v8, 16, v0
1390 ; VI-NEXT: v_lshrrev_b32_e32 v9, 16, v2
1391 ; VI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v5
1392 ; VI-NEXT: v_cndmask_b32_e32 v5, v9, v8, vcc
1393 ; VI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v6
1394 ; VI-NEXT: v_cndmask_b32_e32 v1, v3, v1, vcc
1395 ; VI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v4
1396 ; VI-NEXT: v_cndmask_b32_e32 v0, v2, v0, vcc
1397 ; VI-NEXT: v_lshlrev_b32_e32 v2, 16, v5
1398 ; VI-NEXT: v_or_b32_sdwa v0, v0, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
1399 ; VI-NEXT: v_lshlrev_b32_e32 v2, 16, v7
1400 ; VI-NEXT: v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
1401 ; VI-NEXT: s_setpc_b64 s[30:31]
1402 ;
1403 ; GFX11-LABEL: v_vselect_v4f16:
1404 ; GFX11: ; %bb.0:
1405 ; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1406 ; GFX11-NEXT: v_lshrrev_b32_e32 v8, 16, v1
1407 ; GFX11-NEXT: v_lshrrev_b32_e32 v9, 16, v3
1408 ; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v7
1409 ; GFX11-NEXT: v_lshrrev_b32_e32 v10, 16, v0
1410 ; GFX11-NEXT: v_lshrrev_b32_e32 v11, 16, v2
1411 ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3)
1412 ; GFX11-NEXT: v_cndmask_b32_e32 v7, v9, v8, vcc_lo
1413 ; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v5
1414 ; GFX11-NEXT: v_cndmask_b32_e32 v5, v11, v10, vcc_lo
1415 ; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v4
1416 ; GFX11-NEXT: v_cndmask_b32_e32 v0, v2, v0, vcc_lo
1417 ; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v6
1418 ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_1)
1419 ; GFX11-NEXT: v_perm_b32 v0, v5, v0, 0x5040100
1420 ; GFX11-NEXT: v_cndmask_b32_e32 v1, v3, v1, vcc_lo
1421 ; GFX11-NEXT: v_perm_b32 v1, v7, v1, 0x5040100
1422 ; GFX11-NEXT: s_setpc_b64 s[30:31]
1423 %cmp = icmp eq <4 x i32> %cond, zeroinitializer
1424 %select = select <4 x i1> %cmp, <4 x half> %a, <4 x half> %b
1425 ret <4 x half> %select
1426 }
1428 define <8 x half> @v_select_v8f16(<8 x half> %a, <8 x half> %b, i32 %cond) {
1429 ; SI-LABEL: v_select_v8f16:
1430 ; SI: ; %bb.0:
1431 ; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1432 ; SI-NEXT: v_cvt_f16_f32_e32 v7, v7
1433 ; SI-NEXT: v_cvt_f16_f32_e32 v5, v5
1434 ; SI-NEXT: v_cvt_f16_f32_e32 v3, v3
1435 ; SI-NEXT: v_cvt_f16_f32_e32 v1, v1
1436 ; SI-NEXT: v_cvt_f16_f32_e32 v6, v6
1437 ; SI-NEXT: v_cvt_f16_f32_e32 v15, v15
1438 ; SI-NEXT: v_cvt_f16_f32_e32 v4, v4
1439 ; SI-NEXT: v_cvt_f16_f32_e32 v13, v13
1440 ; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
1441 ; SI-NEXT: v_cvt_f16_f32_e32 v11, v11
1442 ; SI-NEXT: v_cvt_f16_f32_e32 v0, v0
1443 ; SI-NEXT: v_cvt_f16_f32_e32 v9, v9
1444 ; SI-NEXT: v_cvt_f16_f32_e32 v14, v14
1445 ; SI-NEXT: v_cvt_f16_f32_e32 v12, v12
1446 ; SI-NEXT: v_cvt_f16_f32_e32 v10, v10
1447 ; SI-NEXT: v_cvt_f16_f32_e32 v8, v8
1448 ; SI-NEXT: v_lshlrev_b32_e32 v7, 16, v7
1449 ; SI-NEXT: v_lshlrev_b32_e32 v5, 16, v5
1450 ; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v3
1451 ; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
1452 ; SI-NEXT: v_or_b32_e32 v6, v6, v7
1453 ; SI-NEXT: v_lshlrev_b32_e32 v7, 16, v15
1454 ; SI-NEXT: v_or_b32_e32 v4, v4, v5
1455 ; SI-NEXT: v_lshlrev_b32_e32 v5, 16, v13
1456 ; SI-NEXT: v_or_b32_e32 v2, v2, v3
1457 ; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v11
1458 ; SI-NEXT: v_or_b32_e32 v0, v0, v1
1459 ; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v9
1460 ; SI-NEXT: v_or_b32_e32 v7, v14, v7
1461 ; SI-NEXT: v_or_b32_e32 v5, v12, v5
1462 ; SI-NEXT: v_or_b32_e32 v3, v10, v3
1463 ; SI-NEXT: v_or_b32_e32 v1, v8, v1
1464 ; SI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v16
1465 ; SI-NEXT: v_cndmask_b32_e32 v1, v1, v0, vcc
1466 ; SI-NEXT: v_cndmask_b32_e32 v3, v3, v2, vcc
1467 ; SI-NEXT: v_cndmask_b32_e32 v5, v5, v4, vcc
1468 ; SI-NEXT: v_cndmask_b32_e32 v7, v7, v6, vcc
1469 ; SI-NEXT: v_cvt_f32_f16_e32 v0, v1
1470 ; SI-NEXT: v_cvt_f32_f16_e32 v2, v3
1471 ; SI-NEXT: v_cvt_f32_f16_e32 v4, v5
1472 ; SI-NEXT: v_cvt_f32_f16_e32 v6, v7
1473 ; SI-NEXT: v_lshrrev_b32_e32 v1, 16, v1
1474 ; SI-NEXT: v_lshrrev_b32_e32 v3, 16, v3
1475 ; SI-NEXT: v_lshrrev_b32_e32 v5, 16, v5
1476 ; SI-NEXT: v_lshrrev_b32_e32 v7, 16, v7
1477 ; SI-NEXT: v_cvt_f32_f16_e32 v1, v1
1478 ; SI-NEXT: v_cvt_f32_f16_e32 v3, v3
1479 ; SI-NEXT: v_cvt_f32_f16_e32 v5, v5
1480 ; SI-NEXT: v_cvt_f32_f16_e32 v7, v7
1481 ; SI-NEXT: s_setpc_b64 s[30:31]
1482 ;
1483 ; VI-LABEL: v_select_v8f16:
1484 ; VI: ; %bb.0:
1485 ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1486 ; VI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v8
1487 ; VI-NEXT: v_cndmask_b32_e32 v0, v4, v0, vcc
1488 ; VI-NEXT: v_cndmask_b32_e32 v1, v5, v1, vcc
1489 ; VI-NEXT: v_cndmask_b32_e32 v2, v6, v2, vcc
1490 ; VI-NEXT: v_cndmask_b32_e32 v3, v7, v3, vcc
1491 ; VI-NEXT: s_setpc_b64 s[30:31]
1492 ;
1493 ; GFX11-LABEL: v_select_v8f16:
1494 ; GFX11: ; %bb.0:
1495 ; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1496 ; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v8
1497 ; GFX11-NEXT: v_dual_cndmask_b32 v0, v4, v0 :: v_dual_cndmask_b32 v1, v5, v1
1498 ; GFX11-NEXT: v_dual_cndmask_b32 v2, v6, v2 :: v_dual_cndmask_b32 v3, v7, v3
1499 ; GFX11-NEXT: s_setpc_b64 s[30:31]
1500 %cmp = icmp eq i32 %cond, 0
1501 %select = select i1 %cmp, <8 x half> %a, <8 x half> %b
1502 ret <8 x half> %select
1503 }
1505 define <8 x half> @v_vselect_v8f16(<8 x half> %a, <8 x half> %b, <8 x i32> %cond) {
1506 ; SI-LABEL: v_vselect_v8f16:
1507 ; SI: ; %bb.0:
1508 ; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1509 ; SI-NEXT: v_cvt_f16_f32_e32 v0, v0
1510 ; SI-NEXT: v_cvt_f16_f32_e32 v8, v8
1511 ; SI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v16
1512 ; SI-NEXT: v_cvt_f16_f32_e32 v1, v1
1513 ; SI-NEXT: v_cvt_f32_f16_e32 v0, v0
1514 ; SI-NEXT: v_cvt_f32_f16_e32 v8, v8
1515 ; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
1516 ; SI-NEXT: v_cvt_f16_f32_e32 v10, v10
1517 ; SI-NEXT: v_cvt_f32_f16_e32 v1, v1
1518 ; SI-NEXT: v_cndmask_b32_e32 v0, v8, v0, vcc
1519 ; SI-NEXT: v_cvt_f16_f32_e32 v8, v9
1520 ; SI-NEXT: v_cvt_f16_f32_e32 v3, v3
1521 ; SI-NEXT: v_cvt_f16_f32_e32 v11, v11
1522 ; SI-NEXT: v_cvt_f16_f32_e32 v4, v4
1523 ; SI-NEXT: v_cvt_f32_f16_e32 v8, v8
1524 ; SI-NEXT: v_cvt_f16_f32_e32 v12, v12
1525 ; SI-NEXT: v_cvt_f16_f32_e32 v5, v5
1526 ; SI-NEXT: v_cvt_f16_f32_e32 v13, v13
1527 ; SI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v17
1528 ; SI-NEXT: v_cvt_f16_f32_e32 v6, v6
1529 ; SI-NEXT: v_cvt_f16_f32_e32 v9, v14
1530 ; SI-NEXT: v_cvt_f32_f16_e32 v2, v2
1531 ; SI-NEXT: v_cvt_f32_f16_e32 v10, v10
1532 ; SI-NEXT: v_cndmask_b32_e32 v1, v8, v1, vcc
1533 ; SI-NEXT: v_cvt_f16_f32_e32 v7, v7
1534 ; SI-NEXT: v_cvt_f16_f32_e32 v8, v15
1535 ; SI-NEXT: v_cvt_f32_f16_e32 v3, v3
1536 ; SI-NEXT: v_cvt_f32_f16_e32 v11, v11
1537 ; SI-NEXT: v_cvt_f32_f16_e32 v4, v4
1538 ; SI-NEXT: v_cvt_f32_f16_e32 v12, v12
1539 ; SI-NEXT: v_cvt_f32_f16_e32 v5, v5
1540 ; SI-NEXT: v_cvt_f32_f16_e32 v13, v13
1541 ; SI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v18
1542 ; SI-NEXT: v_cvt_f32_f16_e32 v6, v6
1543 ; SI-NEXT: v_cvt_f32_f16_e32 v9, v9
1544 ; SI-NEXT: v_cndmask_b32_e32 v2, v10, v2, vcc
1545 ; SI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v19
1546 ; SI-NEXT: v_cvt_f32_f16_e32 v7, v7
1547 ; SI-NEXT: v_cvt_f32_f16_e32 v8, v8
1548 ; SI-NEXT: v_cndmask_b32_e32 v3, v11, v3, vcc
1549 ; SI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v20
1550 ; SI-NEXT: v_cndmask_b32_e32 v4, v12, v4, vcc
1551 ; SI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v21
1552 ; SI-NEXT: v_cndmask_b32_e32 v5, v13, v5, vcc
1553 ; SI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v22
1554 ; SI-NEXT: v_cndmask_b32_e32 v6, v9, v6, vcc
1555 ; SI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v23
1556 ; SI-NEXT: v_cndmask_b32_e32 v7, v8, v7, vcc
1557 ; SI-NEXT: s_setpc_b64 s[30:31]
1558 ;
1559 ; VI-LABEL: v_vselect_v8f16:
1560 ; VI: ; %bb.0:
1561 ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1562 ; VI-NEXT: v_lshrrev_b32_e32 v16, 16, v3
1563 ; VI-NEXT: v_lshrrev_b32_e32 v17, 16, v7
1564 ; VI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v15
1565 ; VI-NEXT: v_cndmask_b32_e32 v15, v17, v16, vcc
1566 ; VI-NEXT: v_lshrrev_b32_e32 v16, 16, v2
1567 ; VI-NEXT: v_lshrrev_b32_e32 v17, 16, v6
1568 ; VI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v13
1569 ; VI-NEXT: v_cndmask_b32_e32 v13, v17, v16, vcc
1570 ; VI-NEXT: v_lshrrev_b32_e32 v16, 16, v1
1571 ; VI-NEXT: v_lshrrev_b32_e32 v17, 16, v5
1572 ; VI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v11
1573 ; VI-NEXT: v_cndmask_b32_e32 v11, v17, v16, vcc
1574 ; VI-NEXT: v_lshrrev_b32_e32 v16, 16, v0
1575 ; VI-NEXT: v_lshrrev_b32_e32 v17, 16, v4
1576 ; VI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v9
1577 ; VI-NEXT: v_cndmask_b32_e32 v9, v17, v16, vcc
1578 ; VI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v14
1579 ; VI-NEXT: v_cndmask_b32_e32 v3, v7, v3, vcc
1580 ; VI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v12
1581 ; VI-NEXT: v_cndmask_b32_e32 v2, v6, v2, vcc
1582 ; VI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v10
1583 ; VI-NEXT: v_cndmask_b32_e32 v1, v5, v1, vcc
1584 ; VI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v8
1585 ; VI-NEXT: v_cndmask_b32_e32 v0, v4, v0, vcc
1586 ; VI-NEXT: v_lshlrev_b32_e32 v4, 16, v9
1587 ; VI-NEXT: v_or_b32_sdwa v0, v0, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
1588 ; VI-NEXT: v_lshlrev_b32_e32 v4, 16, v11
1589 ; VI-NEXT: v_or_b32_sdwa v1, v1, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
1590 ; VI-NEXT: v_lshlrev_b32_e32 v4, 16, v13
1591 ; VI-NEXT: v_or_b32_sdwa v2, v2, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
1592 ; VI-NEXT: v_lshlrev_b32_e32 v4, 16, v15
1593 ; VI-NEXT: v_or_b32_sdwa v3, v3, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
1594 ; VI-NEXT: s_setpc_b64 s[30:31]
1595 ;
1596 ; GFX11-LABEL: v_vselect_v8f16:
1597 ; GFX11: ; %bb.0:
1598 ; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1599 ; GFX11-NEXT: v_lshrrev_b32_e32 v16, 16, v3
1600 ; GFX11-NEXT: v_lshrrev_b32_e32 v17, 16, v7
1601 ; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v15
1602 ; GFX11-NEXT: v_lshrrev_b32_e32 v18, 16, v1
1603 ; GFX11-NEXT: v_lshrrev_b32_e32 v19, 16, v5
1604 ; GFX11-NEXT: v_lshrrev_b32_e32 v20, 16, v0
1605 ; GFX11-NEXT: v_lshrrev_b32_e32 v21, 16, v4
1606 ; GFX11-NEXT: v_cndmask_b32_e32 v15, v17, v16, vcc_lo
1607 ; GFX11-NEXT: v_lshrrev_b32_e32 v16, 16, v2
1608 ; GFX11-NEXT: v_lshrrev_b32_e32 v17, 16, v6
1609 ; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v13
1610 ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2)
1611 ; GFX11-NEXT: v_cndmask_b32_e32 v13, v17, v16, vcc_lo
1612 ; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v11
1613 ; GFX11-NEXT: v_cndmask_b32_e32 v11, v19, v18, vcc_lo
1614 ; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v9
1615 ; GFX11-NEXT: v_cndmask_b32_e32 v9, v21, v20, vcc_lo
1616 ; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v12
1617 ; GFX11-NEXT: v_cndmask_b32_e32 v2, v6, v2, vcc_lo
1618 ; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v8
1619 ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_2)
1620 ; GFX11-NEXT: v_perm_b32 v2, v13, v2, 0x5040100
1621 ; GFX11-NEXT: v_cndmask_b32_e32 v0, v4, v0, vcc_lo
1622 ; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v10
1623 ; GFX11-NEXT: v_perm_b32 v0, v9, v0, 0x5040100
1624 ; GFX11-NEXT: v_cndmask_b32_e32 v1, v5, v1, vcc_lo
1625 ; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v14
1626 ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_1)
1627 ; GFX11-NEXT: v_perm_b32 v1, v11, v1, 0x5040100
1628 ; GFX11-NEXT: v_cndmask_b32_e32 v3, v7, v3, vcc_lo
1629 ; GFX11-NEXT: v_perm_b32 v3, v15, v3, 0x5040100
1630 ; GFX11-NEXT: s_setpc_b64 s[30:31]
1631 %cmp = icmp eq <8 x i32> %cond, zeroinitializer
1632 %select = select <8 x i1> %cmp, <8 x half> %a, <8 x half> %b
1633 ret <8 x half> %select
1634 }
1636 define <16 x half> @v_select_v16f16(<16 x half> %a, <16 x half> %b, i32 %cond) {
1637 ; SI-LABEL: v_select_v16f16:
1638 ; SI: ; %bb.0:
1639 ; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1640 ; SI-NEXT: v_cvt_f16_f32_e32 v13, v13
1641 ; SI-NEXT: v_cvt_f16_f32_e32 v15, v15
1642 ; SI-NEXT: v_cvt_f16_f32_e32 v12, v12
1643 ; SI-NEXT: v_cvt_f16_f32_e32 v14, v14
1644 ; SI-NEXT: v_lshlrev_b32_e32 v13, 16, v13
1645 ; SI-NEXT: v_cvt_f16_f32_e32 v11, v11
1646 ; SI-NEXT: v_lshlrev_b32_e32 v15, 16, v15
1647 ; SI-NEXT: v_or_b32_e32 v12, v12, v13
1648 ; SI-NEXT: v_cvt_f16_f32_e32 v13, v29
1649 ; SI-NEXT: v_cvt_f16_f32_e32 v10, v10
1650 ; SI-NEXT: v_or_b32_e32 v14, v14, v15
1651 ; SI-NEXT: v_cvt_f16_f32_e32 v15, v28
1652 ; SI-NEXT: v_lshlrev_b32_e32 v11, 16, v11
1653 ; SI-NEXT: v_lshlrev_b32_e32 v13, 16, v13
1654 ; SI-NEXT: v_or_b32_e32 v10, v10, v11
1655 ; SI-NEXT: buffer_load_dword v11, off, s[0:3], s32
1656 ; SI-NEXT: v_or_b32_e32 v13, v15, v13
1657 ; SI-NEXT: v_cvt_f16_f32_e32 v15, v27
1658 ; SI-NEXT: v_cvt_f16_f32_e32 v26, v26
1659 ; SI-NEXT: v_cvt_f16_f32_e32 v3, v3
1660 ; SI-NEXT: v_cvt_f16_f32_e32 v9, v9
1661 ; SI-NEXT: v_lshlrev_b32_e32 v15, 16, v15
1662 ; SI-NEXT: v_or_b32_e32 v15, v26, v15
1663 ; SI-NEXT: buffer_load_dword v26, off, s[0:3], s32 offset:4
1664 ; SI-NEXT: v_cvt_f16_f32_e32 v7, v7
1665 ; SI-NEXT: v_cvt_f16_f32_e32 v5, v5
1666 ; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
1667 ; SI-NEXT: v_cvt_f16_f32_e32 v8, v8
1668 ; SI-NEXT: v_cvt_f16_f32_e32 v6, v6
1669 ; SI-NEXT: v_cvt_f16_f32_e32 v4, v4
1670 ; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v3
1671 ; SI-NEXT: v_lshlrev_b32_e32 v9, 16, v9
1672 ; SI-NEXT: v_lshlrev_b32_e32 v7, 16, v7
1673 ; SI-NEXT: v_lshlrev_b32_e32 v5, 16, v5
1674 ; SI-NEXT: v_or_b32_e32 v2, v2, v3
1675 ; SI-NEXT: v_cvt_f16_f32_e32 v1, v1
1676 ; SI-NEXT: v_or_b32_e32 v8, v8, v9
1677 ; SI-NEXT: v_cvt_f16_f32_e32 v9, v25
1678 ; SI-NEXT: v_or_b32_e32 v6, v6, v7
1679 ; SI-NEXT: v_cvt_f16_f32_e32 v7, v23
1680 ; SI-NEXT: v_or_b32_e32 v4, v4, v5
1681 ; SI-NEXT: v_cvt_f16_f32_e32 v5, v21
1682 ; SI-NEXT: v_cvt_f16_f32_e32 v19, v19
1683 ; SI-NEXT: v_cvt_f16_f32_e32 v0, v0
1684 ; SI-NEXT: v_cvt_f16_f32_e32 v17, v17
1685 ; SI-NEXT: v_cvt_f16_f32_e32 v24, v24
1686 ; SI-NEXT: v_cvt_f16_f32_e32 v22, v22
1687 ; SI-NEXT: v_cvt_f16_f32_e32 v20, v20
1688 ; SI-NEXT: v_cvt_f16_f32_e32 v18, v18
1689 ; SI-NEXT: v_cvt_f16_f32_e32 v16, v16
1690 ; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
1691 ; SI-NEXT: v_lshlrev_b32_e32 v9, 16, v9
1692 ; SI-NEXT: v_lshlrev_b32_e32 v7, 16, v7
1693 ; SI-NEXT: v_lshlrev_b32_e32 v5, 16, v5
1694 ; SI-NEXT: v_or_b32_e32 v0, v0, v1
1695 ; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v17
1696 ; SI-NEXT: v_or_b32_e32 v9, v24, v9
1697 ; SI-NEXT: v_or_b32_e32 v7, v22, v7
1698 ; SI-NEXT: v_or_b32_e32 v5, v20, v5
1699 ; SI-NEXT: v_or_b32_e32 v1, v16, v1
1700 ; SI-NEXT: s_waitcnt vmcnt(1)
1701 ; SI-NEXT: v_cvt_f16_f32_e32 v3, v11
1702 ; SI-NEXT: v_cvt_f16_f32_e32 v11, v30
1703 ; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v3
1704 ; SI-NEXT: v_or_b32_e32 v3, v11, v3
1705 ; SI-NEXT: v_lshlrev_b32_e32 v11, 16, v19
1706 ; SI-NEXT: v_or_b32_e32 v11, v18, v11
1707 ; SI-NEXT: s_waitcnt vmcnt(0)
1708 ; SI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v26
1709 ; SI-NEXT: v_cndmask_b32_e32 v1, v1, v0, vcc
1710 ; SI-NEXT: v_cndmask_b32_e32 v11, v11, v2, vcc
1711 ; SI-NEXT: v_cndmask_b32_e32 v5, v5, v4, vcc
1712 ; SI-NEXT: v_cndmask_b32_e32 v7, v7, v6, vcc
1713 ; SI-NEXT: v_cndmask_b32_e32 v9, v9, v8, vcc
1714 ; SI-NEXT: v_cndmask_b32_e32 v15, v15, v10, vcc
1715 ; SI-NEXT: v_cndmask_b32_e32 v13, v13, v12, vcc
1716 ; SI-NEXT: v_cndmask_b32_e32 v16, v3, v14, vcc
1717 ; SI-NEXT: v_cvt_f32_f16_e32 v0, v1
1718 ; SI-NEXT: v_cvt_f32_f16_e32 v2, v11
1719 ; SI-NEXT: v_cvt_f32_f16_e32 v4, v5
1720 ; SI-NEXT: v_cvt_f32_f16_e32 v6, v7
1721 ; SI-NEXT: v_cvt_f32_f16_e32 v8, v9
1722 ; SI-NEXT: v_cvt_f32_f16_e32 v10, v15
1723 ; SI-NEXT: v_cvt_f32_f16_e32 v12, v13
1724 ; SI-NEXT: v_lshrrev_b32_e32 v1, 16, v1
1725 ; SI-NEXT: v_lshrrev_b32_e32 v3, 16, v11
1726 ; SI-NEXT: v_lshrrev_b32_e32 v5, 16, v5
1727 ; SI-NEXT: v_lshrrev_b32_e32 v7, 16, v7
1728 ; SI-NEXT: v_lshrrev_b32_e32 v9, 16, v9
1729 ; SI-NEXT: v_lshrrev_b32_e32 v11, 16, v15
1730 ; SI-NEXT: v_lshrrev_b32_e32 v13, 16, v13
1731 ; SI-NEXT: v_lshrrev_b32_e32 v15, 16, v16
1732 ; SI-NEXT: v_cvt_f32_f16_e32 v14, v16
1733 ; SI-NEXT: v_cvt_f32_f16_e32 v1, v1
1734 ; SI-NEXT: v_cvt_f32_f16_e32 v3, v3
1735 ; SI-NEXT: v_cvt_f32_f16_e32 v5, v5
1736 ; SI-NEXT: v_cvt_f32_f16_e32 v7, v7
1737 ; SI-NEXT: v_cvt_f32_f16_e32 v9, v9
1738 ; SI-NEXT: v_cvt_f32_f16_e32 v11, v11
1739 ; SI-NEXT: v_cvt_f32_f16_e32 v13, v13
1740 ; SI-NEXT: v_cvt_f32_f16_e32 v15, v15
1741 ; SI-NEXT: s_setpc_b64 s[30:31]
1742 ;
1743 ; VI-LABEL: v_select_v16f16:
1744 ; VI: ; %bb.0:
1745 ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1746 ; VI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v16
1747 ; VI-NEXT: v_cndmask_b32_e32 v0, v8, v0, vcc
1748 ; VI-NEXT: v_cndmask_b32_e32 v1, v9, v1, vcc
1749 ; VI-NEXT: v_cndmask_b32_e32 v2, v10, v2, vcc
1750 ; VI-NEXT: v_cndmask_b32_e32 v3, v11, v3, vcc
1751 ; VI-NEXT: v_cndmask_b32_e32 v4, v12, v4, vcc
1752 ; VI-NEXT: v_cndmask_b32_e32 v5, v13, v5, vcc
1753 ; VI-NEXT: v_cndmask_b32_e32 v6, v14, v6, vcc
1754 ; VI-NEXT: v_cndmask_b32_e32 v7, v15, v7, vcc
1755 ; VI-NEXT: s_setpc_b64 s[30:31]
1756 ;
1757 ; GFX11-LABEL: v_select_v16f16:
1758 ; GFX11: ; %bb.0:
1759 ; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1760 ; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v16
1761 ; GFX11-NEXT: v_dual_cndmask_b32 v0, v8, v0 :: v_dual_cndmask_b32 v1, v9, v1
1762 ; GFX11-NEXT: v_dual_cndmask_b32 v2, v10, v2 :: v_dual_cndmask_b32 v3, v11, v3
1763 ; GFX11-NEXT: v_dual_cndmask_b32 v4, v12, v4 :: v_dual_cndmask_b32 v5, v13, v5
1764 ; GFX11-NEXT: v_dual_cndmask_b32 v6, v14, v6 :: v_dual_cndmask_b32 v7, v15, v7
1765 ; GFX11-NEXT: s_setpc_b64 s[30:31]
1766 %cmp = icmp eq i32 %cond, 0
1767 %select = select i1 %cmp, <16 x half> %a, <16 x half> %b
1768 ret <16 x half> %select
1769 }
1771 define <16 x half> @v_vselect_v16f16(<16 x half> %a, <16 x half> %b, <16 x i32> %cond) {
1772 ; SI-LABEL: v_vselect_v16f16:
1773 ; SI: ; %bb.0:
1774 ; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1775 ; SI-NEXT: buffer_load_dword v36, off, s[0:3], s32 offset:4
1776 ; SI-NEXT: v_cvt_f16_f32_e32 v16, v16
1777 ; SI-NEXT: v_cvt_f16_f32_e32 v0, v0
1778 ; SI-NEXT: v_cvt_f16_f32_e32 v1, v1
1779 ; SI-NEXT: v_cvt_f16_f32_e32 v17, v17
1780 ; SI-NEXT: v_cvt_f32_f16_e32 v37, v16
1781 ; SI-NEXT: buffer_load_dword v38, off, s[0:3], s32 offset:8
1782 ; SI-NEXT: buffer_load_dword v35, off, s[0:3], s32 offset:12
1783 ; SI-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:16
1784 ; SI-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:20
1785 ; SI-NEXT: buffer_load_dword v34, off, s[0:3], s32 offset:24
1786 ; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:28
1787 ; SI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:32
1788 ; SI-NEXT: v_cvt_f32_f16_e32 v0, v0
1789 ; SI-NEXT: v_cvt_f32_f16_e32 v1, v1
1790 ; SI-NEXT: v_cvt_f32_f16_e32 v17, v17
1791 ; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
1792 ; SI-NEXT: v_cvt_f16_f32_e32 v3, v3
1793 ; SI-NEXT: v_cvt_f16_f32_e32 v4, v4
1794 ; SI-NEXT: v_cvt_f16_f32_e32 v5, v5
1795 ; SI-NEXT: v_cvt_f32_f16_e32 v2, v2
1796 ; SI-NEXT: v_cvt_f32_f16_e32 v3, v3
1797 ; SI-NEXT: v_cvt_f32_f16_e32 v4, v4
1798 ; SI-NEXT: v_cvt_f16_f32_e32 v6, v6
1799 ; SI-NEXT: v_cvt_f32_f16_e32 v5, v5
1800 ; SI-NEXT: v_cvt_f16_f32_e32 v7, v7
1801 ; SI-NEXT: v_cvt_f16_f32_e32 v8, v8
1802 ; SI-NEXT: v_cvt_f32_f16_e32 v6, v6
1803 ; SI-NEXT: v_cvt_f16_f32_e32 v9, v9
1804 ; SI-NEXT: v_cvt_f32_f16_e32 v7, v7
1805 ; SI-NEXT: v_cvt_f32_f16_e32 v8, v8
1806 ; SI-NEXT: v_cvt_f16_f32_e32 v10, v10
1807 ; SI-NEXT: v_cvt_f32_f16_e32 v9, v9
1808 ; SI-NEXT: v_cvt_f16_f32_e32 v11, v11
1809 ; SI-NEXT: v_cvt_f16_f32_e32 v12, v12
1810 ; SI-NEXT: v_cvt_f32_f16_e32 v10, v10
1811 ; SI-NEXT: v_cvt_f16_f32_e32 v13, v13
1812 ; SI-NEXT: v_cvt_f32_f16_e32 v11, v11
1813 ; SI-NEXT: v_cvt_f32_f16_e32 v12, v12
1814 ; SI-NEXT: v_cvt_f16_f32_e32 v14, v14
1815 ; SI-NEXT: v_cvt_f16_f32_e32 v15, v15
1816 ; SI-NEXT: v_cvt_f32_f16_e32 v13, v13
1817 ; SI-NEXT: v_cvt_f32_f16_e32 v14, v14
1818 ; SI-NEXT: v_cvt_f32_f16_e32 v15, v15
1819 ; SI-NEXT: s_waitcnt vmcnt(7)
1820 ; SI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v36
1821 ; SI-NEXT: v_cndmask_b32_e32 v0, v37, v0, vcc
1822 ; SI-NEXT: s_waitcnt vmcnt(6)
1823 ; SI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v38
1824 ; SI-NEXT: v_cndmask_b32_e32 v1, v17, v1, vcc
1825 ; SI-NEXT: v_cvt_f16_f32_e32 v17, v18
1826 ; SI-NEXT: s_waitcnt vmcnt(5)
1827 ; SI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v35
1828 ; SI-NEXT: v_cvt_f16_f32_e32 v18, v20
1829 ; SI-NEXT: buffer_load_dword v20, off, s[0:3], s32 offset:48
1830 ; SI-NEXT: v_cvt_f32_f16_e32 v17, v17
1831 ; SI-NEXT: v_cvt_f32_f16_e32 v18, v18
1832 ; SI-NEXT: v_cndmask_b32_e32 v2, v17, v2, vcc
1833 ; SI-NEXT: v_cvt_f16_f32_e32 v17, v19
1834 ; SI-NEXT: s_waitcnt vmcnt(5)
1835 ; SI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v32
1836 ; SI-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:36
1837 ; SI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:44
1838 ; SI-NEXT: v_cvt_f32_f16_e32 v17, v17
1839 ; SI-NEXT: v_cndmask_b32_e32 v3, v17, v3, vcc
1840 ; SI-NEXT: s_waitcnt vmcnt(6)
1841 ; SI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v33
1842 ; SI-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:40
1843 ; SI-NEXT: v_cvt_f16_f32_e32 v17, v21
1844 ; SI-NEXT: v_cndmask_b32_e32 v4, v18, v4, vcc
1845 ; SI-NEXT: v_cvt_f16_f32_e32 v18, v22
1846 ; SI-NEXT: s_waitcnt vmcnt(6)
1847 ; SI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v34
1848 ; SI-NEXT: v_cvt_f32_f16_e32 v17, v17
1849 ; SI-NEXT: v_cvt_f16_f32_e32 v22, v23
1850 ; SI-NEXT: v_cvt_f32_f16_e32 v21, v18
1851 ; SI-NEXT: buffer_load_dword v18, off, s[0:3], s32 offset:56
1852 ; SI-NEXT: v_cndmask_b32_e32 v5, v17, v5, vcc
1853 ; SI-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:52
1854 ; SI-NEXT: s_waitcnt vmcnt(7)
1855 ; SI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v31
1856 ; SI-NEXT: v_cndmask_b32_e32 v6, v21, v6, vcc
1857 ; SI-NEXT: s_waitcnt vmcnt(6)
1858 ; SI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v16
1859 ; SI-NEXT: buffer_load_dword v16, off, s[0:3], s32 offset:60
1860 ; SI-NEXT: buffer_load_dword v21, off, s[0:3], s32
1861 ; SI-NEXT: v_cvt_f32_f16_e32 v22, v22
1862 ; SI-NEXT: v_cvt_f16_f32_e32 v23, v24
1863 ; SI-NEXT: v_cvt_f16_f32_e32 v24, v25
1864 ; SI-NEXT: v_cndmask_b32_e32 v7, v22, v7, vcc
1865 ; SI-NEXT: buffer_load_dword v22, off, s[0:3], s32 offset:64
1866 ; SI-NEXT: v_cvt_f32_f16_e32 v23, v23
1867 ; SI-NEXT: v_cvt_f32_f16_e32 v24, v24
1868 ; SI-NEXT: s_waitcnt vmcnt(7)
1869 ; SI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v32
1870 ; SI-NEXT: v_cndmask_b32_e32 v8, v23, v8, vcc
1871 ; SI-NEXT: v_cvt_f16_f32_e32 v23, v26
1872 ; SI-NEXT: v_cvt_f32_f16_e32 v23, v23
1873 ; SI-NEXT: s_waitcnt vmcnt(5)
1874 ; SI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v33
1875 ; SI-NEXT: v_cndmask_b32_e32 v9, v24, v9, vcc
1876 ; SI-NEXT: v_cvt_f16_f32_e32 v24, v27
1877 ; SI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v19
1878 ; SI-NEXT: v_cvt_f16_f32_e32 v19, v28
1879 ; SI-NEXT: v_cndmask_b32_e32 v10, v23, v10, vcc
1880 ; SI-NEXT: v_cvt_f32_f16_e32 v24, v24
1881 ; SI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v20
1882 ; SI-NEXT: v_cvt_f32_f16_e32 v19, v19
1883 ; SI-NEXT: v_cvt_f16_f32_e32 v20, v29
1884 ; SI-NEXT: v_cndmask_b32_e32 v11, v24, v11, vcc
1885 ; SI-NEXT: s_waitcnt vmcnt(3)
1886 ; SI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v17
1887 ; SI-NEXT: v_cvt_f16_f32_e32 v17, v30
1888 ; SI-NEXT: v_cndmask_b32_e32 v12, v19, v12, vcc
1889 ; SI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v18
1890 ; SI-NEXT: s_waitcnt vmcnt(1)
1891 ; SI-NEXT: v_cvt_f16_f32_e32 v18, v21
1892 ; SI-NEXT: v_cvt_f32_f16_e32 v20, v20
1893 ; SI-NEXT: v_cvt_f32_f16_e32 v17, v17
1894 ; SI-NEXT: v_cvt_f32_f16_e32 v18, v18
1895 ; SI-NEXT: v_cndmask_b32_e32 v13, v20, v13, vcc
1896 ; SI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v16
1897 ; SI-NEXT: v_cndmask_b32_e32 v14, v17, v14, vcc
1898 ; SI-NEXT: s_waitcnt vmcnt(0)
1899 ; SI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v22
1900 ; SI-NEXT: v_cndmask_b32_e32 v15, v18, v15, vcc
1901 ; SI-NEXT: s_setpc_b64 s[30:31]
1902 ;
1903 ; VI-LABEL: v_vselect_v16f16:
1904 ; VI: ; %bb.0:
1905 ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1906 ; VI-NEXT: s_xor_saveexec_b64 s[4:5], -1
1907 ; VI-NEXT: buffer_store_dword v31, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
1908 ; VI-NEXT: s_mov_b64 exec, s[4:5]
1909 ; VI-NEXT: v_writelane_b32 v31, s30, 0
1910 ; VI-NEXT: v_writelane_b32 v31, s31, 1
1911 ; VI-NEXT: v_cmp_eq_u32_e64 s[4:5], 0, v16
1912 ; VI-NEXT: v_cmp_eq_u32_e64 s[18:19], 0, v17
1913 ; VI-NEXT: v_cmp_eq_u32_e64 s[30:31], 0, v29
1914 ; VI-NEXT: v_lshrrev_b32_e32 v16, 16, v6
1915 ; VI-NEXT: v_lshrrev_b32_e32 v17, 16, v14
1916 ; VI-NEXT: v_cmp_eq_u32_e64 s[6:7], 0, v18
1917 ; VI-NEXT: v_cmp_eq_u32_e64 s[28:29], 0, v27
1918 ; VI-NEXT: v_cndmask_b32_e64 v16, v17, v16, s[30:31]
1919 ; VI-NEXT: v_lshrrev_b32_e32 v17, 16, v5
1920 ; VI-NEXT: v_lshrrev_b32_e32 v18, 16, v13
1921 ; VI-NEXT: v_cmp_eq_u32_e64 s[20:21], 0, v19
1922 ; VI-NEXT: v_cmp_eq_u32_e64 s[26:27], 0, v25
1923 ; VI-NEXT: v_cndmask_b32_e64 v17, v18, v17, s[28:29]
1924 ; VI-NEXT: v_lshrrev_b32_e32 v18, 16, v4
1925 ; VI-NEXT: v_lshrrev_b32_e32 v19, 16, v12
1926 ; VI-NEXT: v_cmp_eq_u32_e64 s[8:9], 0, v20
1927 ; VI-NEXT: v_cmp_eq_u32_e64 s[24:25], 0, v23
1928 ; VI-NEXT: v_cndmask_b32_e64 v18, v19, v18, s[26:27]
1929 ; VI-NEXT: v_lshrrev_b32_e32 v19, 16, v3
1930 ; VI-NEXT: v_lshrrev_b32_e32 v20, 16, v11
1931 ; VI-NEXT: v_cmp_eq_u32_e64 s[22:23], 0, v21
1932 ; VI-NEXT: v_cndmask_b32_e64 v19, v20, v19, s[24:25]
1933 ; VI-NEXT: v_lshrrev_b32_e32 v20, 16, v2
1934 ; VI-NEXT: v_lshrrev_b32_e32 v21, 16, v10
1935 ; VI-NEXT: v_cmp_eq_u32_e64 s[10:11], 0, v22
1936 ; VI-NEXT: v_cndmask_b32_e64 v20, v21, v20, s[22:23]
1937 ; VI-NEXT: v_lshrrev_b32_e32 v21, 16, v1
1938 ; VI-NEXT: v_lshrrev_b32_e32 v22, 16, v9
1939 ; VI-NEXT: v_cndmask_b32_e64 v21, v22, v21, s[20:21]
1940 ; VI-NEXT: v_lshrrev_b32_e32 v22, 16, v0
1941 ; VI-NEXT: v_lshrrev_b32_e32 v23, 16, v8
1942 ; VI-NEXT: v_cndmask_b32_e64 v0, v8, v0, s[4:5]
1943 ; VI-NEXT: buffer_load_dword v8, off, s[0:3], s32
1944 ; VI-NEXT: v_cndmask_b32_e64 v22, v23, v22, s[18:19]
1945 ; VI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v30
1946 ; VI-NEXT: v_cndmask_b32_e64 v1, v9, v1, s[6:7]
1947 ; VI-NEXT: v_lshlrev_b32_e32 v9, 16, v22
1948 ; VI-NEXT: v_cndmask_b32_e64 v2, v10, v2, s[8:9]
1949 ; VI-NEXT: v_or_b32_sdwa v0, v0, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
1950 ; VI-NEXT: v_lshrrev_b32_e32 v9, 16, v7
1951 ; VI-NEXT: v_cndmask_b32_e32 v7, v15, v7, vcc
1952 ; VI-NEXT: v_lshrrev_b32_e32 v10, 16, v15
1953 ; VI-NEXT: v_cmp_eq_u32_e64 s[12:13], 0, v24
1954 ; VI-NEXT: v_cndmask_b32_e64 v3, v11, v3, s[10:11]
1955 ; VI-NEXT: v_cmp_eq_u32_e64 s[14:15], 0, v26
1956 ; VI-NEXT: v_cndmask_b32_e64 v4, v12, v4, s[12:13]
1957 ; VI-NEXT: v_cmp_eq_u32_e64 s[16:17], 0, v28
1958 ; VI-NEXT: v_cndmask_b32_e64 v5, v13, v5, s[14:15]
1959 ; VI-NEXT: v_cndmask_b32_e64 v6, v14, v6, s[16:17]
1960 ; VI-NEXT: v_readlane_b32 s31, v31, 1
1961 ; VI-NEXT: v_readlane_b32 s30, v31, 0
1962 ; VI-NEXT: s_waitcnt vmcnt(0)
1963 ; VI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v8
1964 ; VI-NEXT: v_cndmask_b32_e32 v8, v10, v9, vcc
1965 ; VI-NEXT: v_lshlrev_b32_e32 v9, 16, v21
1966 ; VI-NEXT: v_or_b32_sdwa v1, v1, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
1967 ; VI-NEXT: v_lshlrev_b32_e32 v9, 16, v20
1968 ; VI-NEXT: v_or_b32_sdwa v2, v2, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
1969 ; VI-NEXT: v_lshlrev_b32_e32 v9, 16, v19
1970 ; VI-NEXT: v_or_b32_sdwa v3, v3, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
1971 ; VI-NEXT: v_lshlrev_b32_e32 v9, 16, v18
1972 ; VI-NEXT: v_or_b32_sdwa v4, v4, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
1973 ; VI-NEXT: v_lshlrev_b32_e32 v9, 16, v17
1974 ; VI-NEXT: v_or_b32_sdwa v5, v5, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
1975 ; VI-NEXT: v_lshlrev_b32_e32 v9, 16, v16
1976 ; VI-NEXT: v_lshlrev_b32_e32 v8, 16, v8
1977 ; VI-NEXT: v_or_b32_sdwa v6, v6, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
1978 ; VI-NEXT: v_or_b32_sdwa v7, v7, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
1979 ; VI-NEXT: s_xor_saveexec_b64 s[4:5], -1
1980 ; VI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
1981 ; VI-NEXT: s_mov_b64 exec, s[4:5]
1982 ; VI-NEXT: s_waitcnt vmcnt(0)
1983 ; VI-NEXT: s_setpc_b64 s[30:31]
1984 ;
1985 ; GFX11-LABEL: v_vselect_v16f16:
1986 ; GFX11: ; %bb.0:
1987 ; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1988 ; GFX11-NEXT: scratch_load_b32 v31, off, s32
1989 ; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v30
1990 ; GFX11-NEXT: v_lshrrev_b32_e32 v32, 16, v7
1991 ; GFX11-NEXT: v_lshrrev_b32_e32 v34, 16, v6
1992 ; GFX11-NEXT: v_lshrrev_b32_e32 v36, 16, v5
1993 ; GFX11-NEXT: v_lshrrev_b32_e32 v38, 16, v4
1994 ; GFX11-NEXT: v_cndmask_b32_e32 v7, v15, v7, vcc_lo
1995 ; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v28
1996 ; GFX11-NEXT: v_lshrrev_b32_e32 v48, 16, v3
1997 ; GFX11-NEXT: v_lshrrev_b32_e32 v50, 16, v2
1998 ; GFX11-NEXT: v_lshrrev_b32_e32 v52, 16, v1
1999 ; GFX11-NEXT: v_lshrrev_b32_e32 v35, 16, v14
2000 ; GFX11-NEXT: v_cndmask_b32_e32 v6, v14, v6, vcc_lo
2001 ; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v26
2002 ; GFX11-NEXT: v_lshrrev_b32_e32 v54, 16, v0
2003 ; GFX11-NEXT: v_lshrrev_b32_e32 v37, 16, v13
2004 ; GFX11-NEXT: v_lshrrev_b32_e32 v55, 16, v8
2005 ; GFX11-NEXT: v_lshrrev_b32_e32 v39, 16, v12
2006 ; GFX11-NEXT: v_cndmask_b32_e32 v5, v13, v5, vcc_lo
2007 ; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v24
2008 ; GFX11-NEXT: v_lshrrev_b32_e32 v53, 16, v9
2009 ; GFX11-NEXT: v_lshrrev_b32_e32 v51, 16, v10
2010 ; GFX11-NEXT: v_lshrrev_b32_e32 v49, 16, v11
2011 ; GFX11-NEXT: v_lshrrev_b32_e32 v33, 16, v15
2012 ; GFX11-NEXT: v_cndmask_b32_e32 v4, v12, v4, vcc_lo
2013 ; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v22
2014 ; GFX11-NEXT: v_cndmask_b32_e32 v3, v11, v3, vcc_lo
2015 ; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v20
2016 ; GFX11-NEXT: v_cndmask_b32_e32 v2, v10, v2, vcc_lo
2017 ; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v18
2018 ; GFX11-NEXT: v_cndmask_b32_e32 v1, v9, v1, vcc_lo
2019 ; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v16
2020 ; GFX11-NEXT: v_cndmask_b32_e32 v0, v8, v0, vcc_lo
2021 ; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v29
2022 ; GFX11-NEXT: v_cndmask_b32_e32 v8, v35, v34, vcc_lo
2023 ; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v27
2024 ; GFX11-NEXT: v_cndmask_b32_e32 v9, v37, v36, vcc_lo
2025 ; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v25
2026 ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3)
2027 ; GFX11-NEXT: v_perm_b32 v6, v8, v6, 0x5040100
2028 ; GFX11-NEXT: v_perm_b32 v5, v9, v5, 0x5040100
2029 ; GFX11-NEXT: v_cndmask_b32_e32 v10, v39, v38, vcc_lo
2030 ; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v19
2031 ; GFX11-NEXT: v_cndmask_b32_e32 v11, v53, v52, vcc_lo
2032 ; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v17
2033 ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3)
2034 ; GFX11-NEXT: v_perm_b32 v4, v10, v4, 0x5040100
2035 ; GFX11-NEXT: v_perm_b32 v1, v11, v1, 0x5040100
2036 ; GFX11-NEXT: v_cndmask_b32_e32 v12, v55, v54, vcc_lo
2037 ; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v21
2038 ; GFX11-NEXT: v_cndmask_b32_e32 v13, v51, v50, vcc_lo
2039 ; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v23
2040 ; GFX11-NEXT: v_cndmask_b32_e32 v14, v49, v48, vcc_lo
2041 ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
2042 ; GFX11-NEXT: v_perm_b32 v2, v13, v2, 0x5040100
2043 ; GFX11-NEXT: v_perm_b32 v3, v14, v3, 0x5040100
2044 ; GFX11-NEXT: s_waitcnt vmcnt(0)
2045 ; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v31
2046 ; GFX11-NEXT: v_cndmask_b32_e32 v11, v33, v32, vcc_lo
2047 ; GFX11-NEXT: v_perm_b32 v0, v12, v0, 0x5040100
2048 ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2)
2049 ; GFX11-NEXT: v_perm_b32 v7, v11, v7, 0x5040100
2050 ; GFX11-NEXT: s_setpc_b64 s[30:31]
2051 %cmp = icmp eq <16 x i32> %cond, zeroinitializer
2052 %select = select <16 x i1> %cmp, <16 x half> %a, <16 x half> %b
2053 ret <16 x half> %select
2054 }
2056 define <32 x half> @v_select_v32f16(<32 x half> %a, <32 x half> %b, i32 %cond) {
2057 ; SI-LABEL: v_select_v32f16:
2058 ; SI: ; %bb.0:
2059 ; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
2060 ; SI-NEXT: v_cvt_f16_f32_e32 v21, v21
2061 ; SI-NEXT: v_cvt_f16_f32_e32 v20, v20
2062 ; SI-NEXT: v_cvt_f16_f32_e32 v13, v13
2063 ; SI-NEXT: v_cvt_f16_f32_e32 v11, v11
2064 ; SI-NEXT: v_lshlrev_b32_e32 v21, 16, v21
2065 ; SI-NEXT: v_or_b32_e32 v20, v20, v21
2066 ; SI-NEXT: buffer_load_dword v21, off, s[0:3], s32
2067 ; SI-NEXT: v_cvt_f16_f32_e32 v9, v9
2068 ; SI-NEXT: v_cvt_f16_f32_e32 v7, v7
2069 ; SI-NEXT: v_cvt_f16_f32_e32 v5, v5
2070 ; SI-NEXT: v_cvt_f16_f32_e32 v3, v3
2071 ; SI-NEXT: v_cvt_f16_f32_e32 v1, v1
2072 ; SI-NEXT: v_cvt_f16_f32_e32 v23, v23
2073 ; SI-NEXT: v_cvt_f16_f32_e32 v12, v12
2074 ; SI-NEXT: v_cvt_f16_f32_e32 v10, v10
2075 ; SI-NEXT: v_cvt_f16_f32_e32 v8, v8
2076 ; SI-NEXT: v_cvt_f16_f32_e32 v6, v6
2077 ; SI-NEXT: v_cvt_f16_f32_e32 v4, v4
2078 ; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
2079 ; SI-NEXT: v_cvt_f16_f32_e32 v0, v0
2080 ; SI-NEXT: v_cvt_f16_f32_e32 v22, v22
2081 ; SI-NEXT: v_lshlrev_b32_e32 v13, 16, v13
2082 ; SI-NEXT: v_lshlrev_b32_e32 v11, 16, v11
2083 ; SI-NEXT: v_lshlrev_b32_e32 v9, 16, v9
2084 ; SI-NEXT: v_lshlrev_b32_e32 v7, 16, v7
2085 ; SI-NEXT: v_lshlrev_b32_e32 v5, 16, v5
2086 ; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v3
2087 ; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
2088 ; SI-NEXT: v_lshlrev_b32_e32 v23, 16, v23
2089 ; SI-NEXT: v_or_b32_e32 v12, v12, v13
2090 ; SI-NEXT: buffer_load_dword v13, off, s[0:3], s32 offset:56
2091 ; SI-NEXT: v_or_b32_e32 v10, v10, v11
2092 ; SI-NEXT: buffer_load_dword v11, off, s[0:3], s32 offset:48
2093 ; SI-NEXT: v_or_b32_e32 v8, v8, v9
2094 ; SI-NEXT: buffer_load_dword v9, off, s[0:3], s32 offset:40
2095 ; SI-NEXT: v_or_b32_e32 v6, v6, v7
2096 ; SI-NEXT: buffer_load_dword v7, off, s[0:3], s32 offset:32
2097 ; SI-NEXT: v_or_b32_e32 v4, v4, v5
2098 ; SI-NEXT: buffer_load_dword v5, off, s[0:3], s32 offset:24
2099 ; SI-NEXT: v_or_b32_e32 v2, v2, v3
2100 ; SI-NEXT: buffer_load_dword v3, off, s[0:3], s32 offset:16
2101 ; SI-NEXT: v_or_b32_e32 v0, v0, v1
2102 ; SI-NEXT: buffer_load_dword v1, off, s[0:3], s32 offset:8
2103 ; SI-NEXT: v_cvt_f16_f32_e32 v25, v25
2104 ; SI-NEXT: v_or_b32_e32 v22, v22, v23
2105 ; SI-NEXT: v_cvt_f16_f32_e32 v23, v30
2106 ; SI-NEXT: v_cvt_f16_f32_e32 v24, v24
2107 ; SI-NEXT: v_lshlrev_b32_e32 v25, 16, v25
2108 ; SI-NEXT: v_cvt_f16_f32_e32 v27, v27
2109 ; SI-NEXT: v_cvt_f16_f32_e32 v26, v26
2110 ; SI-NEXT: v_or_b32_e32 v24, v24, v25
2111 ; SI-NEXT: v_cvt_f16_f32_e32 v29, v29
2112 ; SI-NEXT: v_lshlrev_b32_e32 v27, 16, v27
2113 ; SI-NEXT: v_or_b32_e32 v26, v26, v27
2114 ; SI-NEXT: v_cvt_f16_f32_e32 v28, v28
2115 ; SI-NEXT: v_lshlrev_b32_e32 v29, 16, v29
2116 ; SI-NEXT: v_cvt_f16_f32_e32 v19, v19
2117 ; SI-NEXT: v_cvt_f16_f32_e32 v18, v18
2118 ; SI-NEXT: v_or_b32_e32 v28, v28, v29
2119 ; SI-NEXT: v_cvt_f16_f32_e32 v17, v17
2120 ; SI-NEXT: v_lshlrev_b32_e32 v19, 16, v19
2121 ; SI-NEXT: v_or_b32_e32 v18, v18, v19
2122 ; SI-NEXT: v_cvt_f16_f32_e32 v16, v16
2123 ; SI-NEXT: v_lshlrev_b32_e32 v17, 16, v17
2124 ; SI-NEXT: v_cvt_f16_f32_e32 v15, v15
2125 ; SI-NEXT: v_cvt_f16_f32_e32 v14, v14
2126 ; SI-NEXT: v_or_b32_e32 v16, v16, v17
2127 ; SI-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:124
2128 ; SI-NEXT: v_lshlrev_b32_e32 v15, 16, v15
2129 ; SI-NEXT: v_or_b32_e32 v14, v14, v15
2130 ; SI-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:116
2131 ; SI-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:108
2132 ; SI-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:100
2133 ; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:92
2134 ; SI-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:84
2135 ; SI-NEXT: buffer_load_dword v19, off, s[0:3], s32 offset:80
2136 ; SI-NEXT: buffer_load_dword v17, off, s[0:3], s32 offset:72
2137 ; SI-NEXT: buffer_load_dword v15, off, s[0:3], s32 offset:64
2138 ; SI-NEXT: s_waitcnt vmcnt(14)
2139 ; SI-NEXT: v_cvt_f16_f32_e32 v21, v21
2140 ; SI-NEXT: v_lshlrev_b32_e32 v21, 16, v21
2141 ; SI-NEXT: v_or_b32_e32 v21, v23, v21
2142 ; SI-NEXT: buffer_load_dword v23, off, s[0:3], s32 offset:128
2143 ; SI-NEXT: v_cvt_f16_f32_e32 v13, v13
2144 ; SI-NEXT: v_cvt_f16_f32_e32 v11, v11
2145 ; SI-NEXT: v_lshlrev_b32_e32 v13, 16, v13
2146 ; SI-NEXT: s_waitcnt vmcnt(14)
2147 ; SI-NEXT: v_cvt_f16_f32_e32 v9, v9
2148 ; SI-NEXT: v_lshlrev_b32_e32 v11, 16, v11
2149 ; SI-NEXT: s_waitcnt vmcnt(13)
2150 ; SI-NEXT: v_cvt_f16_f32_e32 v7, v7
2151 ; SI-NEXT: s_waitcnt vmcnt(12)
2152 ; SI-NEXT: v_cvt_f16_f32_e32 v5, v5
2153 ; SI-NEXT: v_lshlrev_b32_e32 v9, 16, v9
2154 ; SI-NEXT: s_waitcnt vmcnt(11)
2155 ; SI-NEXT: v_cvt_f16_f32_e32 v3, v3
2156 ; SI-NEXT: v_lshlrev_b32_e32 v7, 16, v7
2157 ; SI-NEXT: v_lshlrev_b32_e32 v5, 16, v5
2158 ; SI-NEXT: s_waitcnt vmcnt(10)
2159 ; SI-NEXT: v_cvt_f16_f32_e32 v1, v1
2160 ; SI-NEXT: v_lshlrev_b32_e32 v3, 16, v3
2161 ; SI-NEXT: v_lshlrev_b32_e32 v1, 16, v1
2162 ; SI-NEXT: s_waitcnt vmcnt(9)
2163 ; SI-NEXT: v_cvt_f16_f32_e32 v25, v25
2164 ; SI-NEXT: s_waitcnt vmcnt(8)
2165 ; SI-NEXT: v_cvt_f16_f32_e32 v27, v27
2166 ; SI-NEXT: s_waitcnt vmcnt(7)
2167 ; SI-NEXT: v_cvt_f16_f32_e32 v29, v29
2168 ; SI-NEXT: s_waitcnt vmcnt(6)
2169 ; SI-NEXT: v_cvt_f16_f32_e32 v30, v30
2170 ; SI-NEXT: s_waitcnt vmcnt(5)
2171 ; SI-NEXT: v_cvt_f16_f32_e32 v31, v31
2172 ; SI-NEXT: s_waitcnt vmcnt(4)
2173 ; SI-NEXT: v_cvt_f16_f32_e32 v32, v32
2174 ; SI-NEXT: s_waitcnt vmcnt(0)
2175 ; SI-NEXT: v_cvt_f16_f32_e32 v23, v23
2176 ; SI-NEXT: v_lshlrev_b32_e32 v23, 16, v23
2177 ; SI-NEXT: v_or_b32_e32 v23, v25, v23
2178 ; SI-NEXT: buffer_load_dword v25, off, s[0:3], s32 offset:120
2179 ; SI-NEXT: s_waitcnt vmcnt(0)
2180 ; SI-NEXT: v_cvt_f16_f32_e32 v25, v25
2181 ; SI-NEXT: v_lshlrev_b32_e32 v25, 16, v25
2182 ; SI-NEXT: v_or_b32_e32 v25, v27, v25
2183 ; SI-NEXT: buffer_load_dword v27, off, s[0:3], s32 offset:112
2184 ; SI-NEXT: s_waitcnt vmcnt(0)
2185 ; SI-NEXT: v_cvt_f16_f32_e32 v27, v27
2186 ; SI-NEXT: v_lshlrev_b32_e32 v27, 16, v27
2187 ; SI-NEXT: v_or_b32_e32 v27, v29, v27
2188 ; SI-NEXT: buffer_load_dword v29, off, s[0:3], s32 offset:104
2189 ; SI-NEXT: s_waitcnt vmcnt(0)
2190 ; SI-NEXT: v_cvt_f16_f32_e32 v29, v29
2191 ; SI-NEXT: v_lshlrev_b32_e32 v29, 16, v29
2192 ; SI-NEXT: v_or_b32_e32 v29, v30, v29
2193 ; SI-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:96
2194 ; SI-NEXT: s_waitcnt vmcnt(0)
2195 ; SI-NEXT: v_cvt_f16_f32_e32 v30, v30
2196 ; SI-NEXT: v_lshlrev_b32_e32 v30, 16, v30
2197 ; SI-NEXT: v_or_b32_e32 v30, v31, v30
2198 ; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:88
2199 ; SI-NEXT: s_waitcnt vmcnt(0)
2200 ; SI-NEXT: v_cvt_f16_f32_e32 v31, v31
2201 ; SI-NEXT: v_lshlrev_b32_e32 v31, 16, v31
2202 ; SI-NEXT: v_or_b32_e32 v31, v32, v31
2203 ; SI-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:76
2204 ; SI-NEXT: v_cvt_f16_f32_e32 v19, v19
2205 ; SI-NEXT: v_lshlrev_b32_e32 v19, 16, v19
2206 ; SI-NEXT: s_waitcnt vmcnt(0)
2207 ; SI-NEXT: v_cvt_f16_f32_e32 v32, v32
2208 ; SI-NEXT: v_or_b32_e32 v19, v32, v19
2209 ; SI-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:68
2210 ; SI-NEXT: v_cvt_f16_f32_e32 v17, v17
2211 ; SI-NEXT: v_lshlrev_b32_e32 v17, 16, v17
2212 ; SI-NEXT: s_waitcnt vmcnt(0)
2213 ; SI-NEXT: v_cvt_f16_f32_e32 v32, v32
2214 ; SI-NEXT: v_or_b32_e32 v17, v32, v17
2215 ; SI-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:60
2216 ; SI-NEXT: v_cvt_f16_f32_e32 v15, v15
2217 ; SI-NEXT: v_lshlrev_b32_e32 v15, 16, v15
2218 ; SI-NEXT: s_waitcnt vmcnt(0)
2219 ; SI-NEXT: v_cvt_f16_f32_e32 v32, v32
2220 ; SI-NEXT: v_or_b32_e32 v15, v32, v15
2221 ; SI-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:52
2222 ; SI-NEXT: s_waitcnt vmcnt(0)
2223 ; SI-NEXT: v_cvt_f16_f32_e32 v32, v32
2224 ; SI-NEXT: v_or_b32_e32 v13, v32, v13
2225 ; SI-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:44
2226 ; SI-NEXT: s_waitcnt vmcnt(0)
2227 ; SI-NEXT: v_cvt_f16_f32_e32 v32, v32
2228 ; SI-NEXT: v_or_b32_e32 v11, v32, v11
2229 ; SI-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:36
2230 ; SI-NEXT: s_waitcnt vmcnt(0)
2231 ; SI-NEXT: v_cvt_f16_f32_e32 v32, v32
2232 ; SI-NEXT: v_or_b32_e32 v9, v32, v9
2233 ; SI-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:28
2234 ; SI-NEXT: s_waitcnt vmcnt(0)
2235 ; SI-NEXT: v_cvt_f16_f32_e32 v32, v32
2236 ; SI-NEXT: v_or_b32_e32 v7, v32, v7
2237 ; SI-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:20
2238 ; SI-NEXT: s_waitcnt vmcnt(0)
2239 ; SI-NEXT: v_cvt_f16_f32_e32 v32, v32
2240 ; SI-NEXT: v_or_b32_e32 v5, v32, v5
2241 ; SI-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:12
2242 ; SI-NEXT: s_waitcnt vmcnt(0)
2243 ; SI-NEXT: v_cvt_f16_f32_e32 v32, v32
2244 ; SI-NEXT: v_or_b32_e32 v3, v32, v3
2245 ; SI-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:4
2246 ; SI-NEXT: s_waitcnt vmcnt(0)
2247 ; SI-NEXT: v_cvt_f16_f32_e32 v32, v32
2248 ; SI-NEXT: v_or_b32_e32 v1, v32, v1
2249 ; SI-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:132
2250 ; SI-NEXT: s_waitcnt vmcnt(0)
2251 ; SI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v32
2252 ; SI-NEXT: v_cndmask_b32_e32 v1, v1, v0, vcc
2253 ; SI-NEXT: v_cndmask_b32_e32 v3, v3, v2, vcc
2254 ; SI-NEXT: v_cndmask_b32_e32 v5, v5, v4, vcc
2255 ; SI-NEXT: v_cndmask_b32_e32 v7, v7, v6, vcc
2256 ; SI-NEXT: v_cndmask_b32_e32 v9, v9, v8, vcc
2257 ; SI-NEXT: v_cndmask_b32_e32 v11, v11, v10, vcc
2258 ; SI-NEXT: v_cndmask_b32_e32 v13, v13, v12, vcc
2259 ; SI-NEXT: v_cndmask_b32_e32 v15, v15, v14, vcc
2260 ; SI-NEXT: v_cndmask_b32_e32 v17, v17, v16, vcc
2261 ; SI-NEXT: v_cndmask_b32_e32 v19, v19, v18, vcc
2262 ; SI-NEXT: v_cndmask_b32_e32 v31, v31, v20, vcc
2263 ; SI-NEXT: v_cndmask_b32_e32 v30, v30, v22, vcc
2264 ; SI-NEXT: v_cndmask_b32_e32 v29, v29, v24, vcc
2265 ; SI-NEXT: v_cndmask_b32_e32 v27, v27, v26, vcc
2266 ; SI-NEXT: v_cndmask_b32_e32 v32, v25, v28, vcc
2267 ; SI-NEXT: v_cndmask_b32_e32 v33, v23, v21, vcc
2268 ; SI-NEXT: v_cvt_f32_f16_e32 v0, v1
2269 ; SI-NEXT: v_cvt_f32_f16_e32 v2, v3
2270 ; SI-NEXT: v_cvt_f32_f16_e32 v4, v5
2271 ; SI-NEXT: v_cvt_f32_f16_e32 v6, v7
2272 ; SI-NEXT: v_cvt_f32_f16_e32 v8, v9
2273 ; SI-NEXT: v_lshrrev_b32_e32 v1, 16, v1
2274 ; SI-NEXT: v_lshrrev_b32_e32 v3, 16, v3
2275 ; SI-NEXT: v_lshrrev_b32_e32 v5, 16, v5
2276 ; SI-NEXT: v_lshrrev_b32_e32 v7, 16, v7
2277 ; SI-NEXT: v_lshrrev_b32_e32 v9, 16, v9
2278 ; SI-NEXT: v_cvt_f32_f16_e32 v10, v11
2279 ; SI-NEXT: v_lshrrev_b32_e32 v11, 16, v11
2280 ; SI-NEXT: v_cvt_f32_f16_e32 v12, v13
2281 ; SI-NEXT: v_lshrrev_b32_e32 v13, 16, v13
2282 ; SI-NEXT: v_cvt_f32_f16_e32 v14, v15
2283 ; SI-NEXT: v_lshrrev_b32_e32 v15, 16, v15
2284 ; SI-NEXT: v_cvt_f32_f16_e32 v16, v17
2285 ; SI-NEXT: v_lshrrev_b32_e32 v17, 16, v17
2286 ; SI-NEXT: v_cvt_f32_f16_e32 v18, v19
2287 ; SI-NEXT: v_lshrrev_b32_e32 v19, 16, v19
2288 ; SI-NEXT: v_cvt_f32_f16_e32 v20, v31
2289 ; SI-NEXT: v_lshrrev_b32_e32 v21, 16, v31
2290 ; SI-NEXT: v_lshrrev_b32_e32 v23, 16, v30
2291 ; SI-NEXT: v_cvt_f32_f16_e32 v24, v29
2292 ; SI-NEXT: v_lshrrev_b32_e32 v25, 16, v29
2293 ; SI-NEXT: v_cvt_f32_f16_e32 v26, v27
2294 ; SI-NEXT: v_lshrrev_b32_e32 v27, 16, v27
2295 ; SI-NEXT: v_lshrrev_b32_e32 v29, 16, v32
2296 ; SI-NEXT: v_lshrrev_b32_e32 v31, 16, v33
2297 ; SI-NEXT: v_cvt_f32_f16_e32 v1, v1
2298 ; SI-NEXT: v_cvt_f32_f16_e32 v3, v3
2299 ; SI-NEXT: v_cvt_f32_f16_e32 v5, v5
2300 ; SI-NEXT: v_cvt_f32_f16_e32 v7, v7
2301 ; SI-NEXT: v_cvt_f32_f16_e32 v9, v9
2302 ; SI-NEXT: v_cvt_f32_f16_e32 v11, v11
2303 ; SI-NEXT: v_cvt_f32_f16_e32 v13, v13
2304 ; SI-NEXT: v_cvt_f32_f16_e32 v15, v15
2305 ; SI-NEXT: v_cvt_f32_f16_e32 v17, v17
2306 ; SI-NEXT: v_cvt_f32_f16_e32 v19, v19
2307 ; SI-NEXT: v_cvt_f32_f16_e32 v21, v21
2308 ; SI-NEXT: v_cvt_f32_f16_e32 v22, v30
2309 ; SI-NEXT: v_cvt_f32_f16_e32 v23, v23
2310 ; SI-NEXT: v_cvt_f32_f16_e32 v25, v25
2311 ; SI-NEXT: v_cvt_f32_f16_e32 v27, v27
2312 ; SI-NEXT: v_cvt_f32_f16_e32 v28, v32
2313 ; SI-NEXT: v_cvt_f32_f16_e32 v29, v29
2314 ; SI-NEXT: v_cvt_f32_f16_e32 v30, v33
2315 ; SI-NEXT: v_cvt_f32_f16_e32 v31, v31
2316 ; SI-NEXT: s_setpc_b64 s[30:31]
2317 ;
2318 ; VI-LABEL: v_select_v32f16:
2319 ; VI: ; %bb.0:
2320 ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
2321 ; VI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:4
2322 ; VI-NEXT: s_waitcnt vmcnt(0)
2323 ; VI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v31
2324 ; VI-NEXT: v_cndmask_b32_e32 v0, v16, v0, vcc
2325 ; VI-NEXT: buffer_load_dword v16, off, s[0:3], s32
2326 ; VI-NEXT: v_cndmask_b32_e32 v1, v17, v1, vcc
2327 ; VI-NEXT: v_cndmask_b32_e32 v2, v18, v2, vcc
2328 ; VI-NEXT: v_cndmask_b32_e32 v3, v19, v3, vcc
2329 ; VI-NEXT: v_cndmask_b32_e32 v4, v20, v4, vcc
2330 ; VI-NEXT: v_cndmask_b32_e32 v5, v21, v5, vcc
2331 ; VI-NEXT: v_cndmask_b32_e32 v6, v22, v6, vcc
2332 ; VI-NEXT: v_cndmask_b32_e32 v7, v23, v7, vcc
2333 ; VI-NEXT: v_cndmask_b32_e32 v8, v24, v8, vcc
2334 ; VI-NEXT: v_cndmask_b32_e32 v9, v25, v9, vcc
2335 ; VI-NEXT: v_cndmask_b32_e32 v10, v26, v10, vcc
2336 ; VI-NEXT: v_cndmask_b32_e32 v11, v27, v11, vcc
2337 ; VI-NEXT: v_cndmask_b32_e32 v12, v28, v12, vcc
2338 ; VI-NEXT: v_cndmask_b32_e32 v13, v29, v13, vcc
2339 ; VI-NEXT: v_cndmask_b32_e32 v14, v30, v14, vcc
2340 ; VI-NEXT: s_waitcnt vmcnt(0)
2341 ; VI-NEXT: v_cndmask_b32_e32 v15, v16, v15, vcc
2342 ; VI-NEXT: s_setpc_b64 s[30:31]
2343 ;
2344 ; GFX11-LABEL: v_select_v32f16:
2345 ; GFX11: ; %bb.0:
2346 ; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
2347 ; GFX11-NEXT: s_clause 0x1
2348 ; GFX11-NEXT: scratch_load_b32 v31, off, s32 offset:4
2349 ; GFX11-NEXT: scratch_load_b32 v32, off, s32
2350 ; GFX11-NEXT: s_waitcnt vmcnt(1)
2351 ; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v31
2352 ; GFX11-NEXT: v_dual_cndmask_b32 v0, v16, v0 :: v_dual_cndmask_b32 v1, v17, v1
2353 ; GFX11-NEXT: v_dual_cndmask_b32 v2, v18, v2 :: v_dual_cndmask_b32 v3, v19, v3
2354 ; GFX11-NEXT: v_dual_cndmask_b32 v4, v20, v4 :: v_dual_cndmask_b32 v5, v21, v5
2355 ; GFX11-NEXT: v_dual_cndmask_b32 v6, v22, v6 :: v_dual_cndmask_b32 v7, v23, v7
2356 ; GFX11-NEXT: v_dual_cndmask_b32 v8, v24, v8 :: v_dual_cndmask_b32 v9, v25, v9
2357 ; GFX11-NEXT: v_dual_cndmask_b32 v10, v26, v10 :: v_dual_cndmask_b32 v11, v27, v11
2358 ; GFX11-NEXT: v_dual_cndmask_b32 v12, v28, v12 :: v_dual_cndmask_b32 v13, v29, v13
2359 ; GFX11-NEXT: s_waitcnt vmcnt(0)
2360 ; GFX11-NEXT: v_dual_cndmask_b32 v14, v30, v14 :: v_dual_cndmask_b32 v15, v32, v15
2361 ; GFX11-NEXT: s_setpc_b64 s[30:31]
2362 %cmp = icmp eq i32 %cond, 0
2363 %select = select i1 %cmp, <32 x half> %a, <32 x half> %b
2364 ret <32 x half> %select
2365 }
2367 define <32 x half> @v_vselect_v32f16(<32 x half> %a, <32 x half> %b, <32 x i32> %cond) {
2368 ; SI-LABEL: v_vselect_v32f16:
2369 ; SI: ; %bb.0:
2370 ; SI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
2371 ; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:132
2372 ; SI-NEXT: v_cvt_f16_f32_e32 v0, v0
2373 ; SI-NEXT: v_cvt_f16_f32_e32 v1, v1
2374 ; SI-NEXT: v_cvt_f16_f32_e32 v2, v2
2375 ; SI-NEXT: v_cvt_f16_f32_e32 v3, v3
2376 ; SI-NEXT: v_cvt_f32_f16_e32 v0, v0
2377 ; SI-NEXT: v_cvt_f32_f16_e32 v1, v1
2378 ; SI-NEXT: v_cvt_f32_f16_e32 v2, v2
2379 ; SI-NEXT: v_cvt_f32_f16_e32 v3, v3
2380 ; SI-NEXT: v_cvt_f16_f32_e32 v4, v4
2381 ; SI-NEXT: v_cvt_f16_f32_e32 v5, v5
2382 ; SI-NEXT: v_cvt_f16_f32_e32 v6, v6
2383 ; SI-NEXT: v_cvt_f16_f32_e32 v7, v7
2384 ; SI-NEXT: v_cvt_f32_f16_e32 v4, v4
2385 ; SI-NEXT: v_cvt_f32_f16_e32 v5, v5
2386 ; SI-NEXT: v_cvt_f32_f16_e32 v6, v6
2387 ; SI-NEXT: v_cvt_f32_f16_e32 v7, v7
2388 ; SI-NEXT: v_cvt_f16_f32_e32 v8, v8
2389 ; SI-NEXT: v_cvt_f16_f32_e32 v9, v9
2390 ; SI-NEXT: v_cvt_f16_f32_e32 v10, v10
2391 ; SI-NEXT: v_cvt_f16_f32_e32 v11, v11
2392 ; SI-NEXT: v_cvt_f32_f16_e32 v8, v8
2393 ; SI-NEXT: v_cvt_f32_f16_e32 v9, v9
2394 ; SI-NEXT: v_cvt_f32_f16_e32 v10, v10
2395 ; SI-NEXT: v_cvt_f32_f16_e32 v11, v11
2396 ; SI-NEXT: v_cvt_f16_f32_e32 v12, v12
2397 ; SI-NEXT: v_cvt_f16_f32_e32 v13, v13
2398 ; SI-NEXT: v_cvt_f16_f32_e32 v14, v14
2399 ; SI-NEXT: v_cvt_f16_f32_e32 v15, v15
2400 ; SI-NEXT: v_cvt_f32_f16_e32 v12, v12
2401 ; SI-NEXT: v_cvt_f32_f16_e32 v13, v13
2402 ; SI-NEXT: v_cvt_f32_f16_e32 v14, v14
2403 ; SI-NEXT: v_cvt_f32_f16_e32 v15, v15
2404 ; SI-NEXT: v_cvt_f16_f32_e32 v16, v16
2405 ; SI-NEXT: v_cvt_f16_f32_e32 v17, v17
2406 ; SI-NEXT: v_cvt_f16_f32_e32 v18, v18
2407 ; SI-NEXT: v_cvt_f16_f32_e32 v19, v19
2408 ; SI-NEXT: v_cvt_f32_f16_e32 v16, v16
2409 ; SI-NEXT: v_cvt_f32_f16_e32 v17, v17
2410 ; SI-NEXT: v_cvt_f32_f16_e32 v18, v18
2411 ; SI-NEXT: v_cvt_f32_f16_e32 v19, v19
2412 ; SI-NEXT: v_cvt_f16_f32_e32 v20, v20
2413 ; SI-NEXT: v_cvt_f16_f32_e32 v21, v21
2414 ; SI-NEXT: v_cvt_f16_f32_e32 v22, v22
2415 ; SI-NEXT: v_cvt_f16_f32_e32 v23, v23
2416 ; SI-NEXT: v_cvt_f32_f16_e32 v20, v20
2417 ; SI-NEXT: v_cvt_f32_f16_e32 v21, v21
2418 ; SI-NEXT: v_cvt_f32_f16_e32 v22, v22
2419 ; SI-NEXT: v_cvt_f32_f16_e32 v23, v23
2420 ; SI-NEXT: v_cvt_f16_f32_e32 v24, v24
2421 ; SI-NEXT: v_cvt_f16_f32_e32 v25, v25
2422 ; SI-NEXT: v_cvt_f16_f32_e32 v26, v26
2423 ; SI-NEXT: v_cvt_f16_f32_e32 v27, v27
2424 ; SI-NEXT: v_cvt_f32_f16_e32 v24, v24
2425 ; SI-NEXT: v_cvt_f32_f16_e32 v25, v25
2426 ; SI-NEXT: v_cvt_f32_f16_e32 v26, v26
2427 ; SI-NEXT: v_cvt_f32_f16_e32 v27, v27
2428 ; SI-NEXT: v_cvt_f16_f32_e32 v28, v28
2429 ; SI-NEXT: v_cvt_f16_f32_e32 v29, v29
2430 ; SI-NEXT: v_cvt_f16_f32_e32 v30, v30
2431 ; SI-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:128
2432 ; SI-NEXT: v_cvt_f32_f16_e32 v28, v28
2433 ; SI-NEXT: v_cvt_f32_f16_e32 v29, v29
2434 ; SI-NEXT: v_cvt_f32_f16_e32 v30, v30
2435 ; SI-NEXT: s_waitcnt vmcnt(1)
2436 ; SI-NEXT: v_cmp_eq_u32_e64 s[16:17], 0, v31
2437 ; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:136
2438 ; SI-NEXT: s_waitcnt vmcnt(1)
2439 ; SI-NEXT: v_cvt_f16_f32_e32 v32, v32
2440 ; SI-NEXT: v_cvt_f32_f16_e32 v32, v32
2441 ; SI-NEXT: s_waitcnt vmcnt(0)
2442 ; SI-NEXT: v_cmp_eq_u32_e64 s[14:15], 0, v31
2443 ; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:140
2444 ; SI-NEXT: s_waitcnt vmcnt(0)
2445 ; SI-NEXT: v_cmp_eq_u32_e64 s[12:13], 0, v31
2446 ; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:144
2447 ; SI-NEXT: s_waitcnt vmcnt(0)
2448 ; SI-NEXT: v_cmp_eq_u32_e64 s[10:11], 0, v31
2449 ; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:148
2450 ; SI-NEXT: s_waitcnt vmcnt(0)
2451 ; SI-NEXT: v_cmp_eq_u32_e64 s[8:9], 0, v31
2452 ; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:152
2453 ; SI-NEXT: s_waitcnt vmcnt(0)
2454 ; SI-NEXT: v_cmp_eq_u32_e64 s[6:7], 0, v31
2455 ; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:156
2456 ; SI-NEXT: s_waitcnt vmcnt(0)
2457 ; SI-NEXT: v_cmp_eq_u32_e64 s[4:5], 0, v31
2458 ; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:160
2459 ; SI-NEXT: s_waitcnt vmcnt(0)
2460 ; SI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v31
2461 ; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:4
2462 ; SI-NEXT: s_waitcnt vmcnt(0)
2463 ; SI-NEXT: v_cvt_f16_f32_e32 v31, v31
2464 ; SI-NEXT: v_cvt_f32_f16_e32 v31, v31
2465 ; SI-NEXT: v_cndmask_b32_e64 v0, v31, v0, s[16:17]
2466 ; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:8
2467 ; SI-NEXT: s_waitcnt vmcnt(0)
2468 ; SI-NEXT: v_cvt_f16_f32_e32 v31, v31
2469 ; SI-NEXT: v_cvt_f32_f16_e32 v31, v31
2470 ; SI-NEXT: v_cndmask_b32_e64 v1, v31, v1, s[14:15]
2471 ; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:12
2472 ; SI-NEXT: s_waitcnt vmcnt(0)
2473 ; SI-NEXT: v_cvt_f16_f32_e32 v31, v31
2474 ; SI-NEXT: v_cvt_f32_f16_e32 v31, v31
2475 ; SI-NEXT: v_cndmask_b32_e64 v2, v31, v2, s[12:13]
2476 ; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:16
2477 ; SI-NEXT: s_waitcnt vmcnt(0)
2478 ; SI-NEXT: v_cvt_f16_f32_e32 v31, v31
2479 ; SI-NEXT: v_cvt_f32_f16_e32 v31, v31
2480 ; SI-NEXT: v_cndmask_b32_e64 v3, v31, v3, s[10:11]
2481 ; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:20
2482 ; SI-NEXT: s_waitcnt vmcnt(0)
2483 ; SI-NEXT: v_cvt_f16_f32_e32 v31, v31
2484 ; SI-NEXT: v_cvt_f32_f16_e32 v31, v31
2485 ; SI-NEXT: v_cndmask_b32_e64 v4, v31, v4, s[8:9]
2486 ; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:24
2487 ; SI-NEXT: s_waitcnt vmcnt(0)
2488 ; SI-NEXT: v_cvt_f16_f32_e32 v31, v31
2489 ; SI-NEXT: v_cvt_f32_f16_e32 v31, v31
2490 ; SI-NEXT: v_cndmask_b32_e64 v5, v31, v5, s[6:7]
2491 ; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:28
2492 ; SI-NEXT: s_waitcnt vmcnt(0)
2493 ; SI-NEXT: v_cvt_f16_f32_e32 v31, v31
2494 ; SI-NEXT: v_cvt_f32_f16_e32 v31, v31
2495 ; SI-NEXT: v_cndmask_b32_e64 v6, v31, v6, s[4:5]
2496 ; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:32
2497 ; SI-NEXT: s_waitcnt vmcnt(0)
2498 ; SI-NEXT: v_cvt_f16_f32_e32 v31, v31
2499 ; SI-NEXT: v_cvt_f32_f16_e32 v31, v31
2500 ; SI-NEXT: v_cndmask_b32_e32 v7, v31, v7, vcc
2501 ; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:164
2502 ; SI-NEXT: s_waitcnt vmcnt(0)
2503 ; SI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v31
2504 ; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:36
2505 ; SI-NEXT: s_waitcnt vmcnt(0)
2506 ; SI-NEXT: v_cvt_f16_f32_e32 v31, v31
2507 ; SI-NEXT: v_cvt_f32_f16_e32 v31, v31
2508 ; SI-NEXT: v_cndmask_b32_e32 v8, v31, v8, vcc
2509 ; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:168
2510 ; SI-NEXT: s_waitcnt vmcnt(0)
2511 ; SI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v31
2512 ; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:40
2513 ; SI-NEXT: s_waitcnt vmcnt(0)
2514 ; SI-NEXT: v_cvt_f16_f32_e32 v31, v31
2515 ; SI-NEXT: v_cvt_f32_f16_e32 v31, v31
2516 ; SI-NEXT: v_cndmask_b32_e32 v9, v31, v9, vcc
2517 ; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:172
2518 ; SI-NEXT: s_waitcnt vmcnt(0)
2519 ; SI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v31
2520 ; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:44
2521 ; SI-NEXT: s_waitcnt vmcnt(0)
2522 ; SI-NEXT: v_cvt_f16_f32_e32 v31, v31
2523 ; SI-NEXT: v_cvt_f32_f16_e32 v31, v31
2524 ; SI-NEXT: v_cndmask_b32_e32 v10, v31, v10, vcc
2525 ; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:176
2526 ; SI-NEXT: s_waitcnt vmcnt(0)
2527 ; SI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v31
2528 ; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:48
2529 ; SI-NEXT: s_waitcnt vmcnt(0)
2530 ; SI-NEXT: v_cvt_f16_f32_e32 v31, v31
2531 ; SI-NEXT: v_cvt_f32_f16_e32 v31, v31
2532 ; SI-NEXT: v_cndmask_b32_e32 v11, v31, v11, vcc
2533 ; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:180
2534 ; SI-NEXT: s_waitcnt vmcnt(0)
2535 ; SI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v31
2536 ; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:52
2537 ; SI-NEXT: s_waitcnt vmcnt(0)
2538 ; SI-NEXT: v_cvt_f16_f32_e32 v31, v31
2539 ; SI-NEXT: v_cvt_f32_f16_e32 v31, v31
2540 ; SI-NEXT: v_cndmask_b32_e32 v12, v31, v12, vcc
2541 ; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:184
2542 ; SI-NEXT: s_waitcnt vmcnt(0)
2543 ; SI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v31
2544 ; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:56
2545 ; SI-NEXT: s_waitcnt vmcnt(0)
2546 ; SI-NEXT: v_cvt_f16_f32_e32 v31, v31
2547 ; SI-NEXT: v_cvt_f32_f16_e32 v31, v31
2548 ; SI-NEXT: v_cndmask_b32_e32 v13, v31, v13, vcc
2549 ; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:188
2550 ; SI-NEXT: s_waitcnt vmcnt(0)
2551 ; SI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v31
2552 ; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:60
2553 ; SI-NEXT: s_waitcnt vmcnt(0)
2554 ; SI-NEXT: v_cvt_f16_f32_e32 v31, v31
2555 ; SI-NEXT: v_cvt_f32_f16_e32 v31, v31
2556 ; SI-NEXT: v_cndmask_b32_e32 v14, v31, v14, vcc
2557 ; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:192
2558 ; SI-NEXT: s_waitcnt vmcnt(0)
2559 ; SI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v31
2560 ; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:64
2561 ; SI-NEXT: s_waitcnt vmcnt(0)
2562 ; SI-NEXT: v_cvt_f16_f32_e32 v31, v31
2563 ; SI-NEXT: v_cvt_f32_f16_e32 v31, v31
2564 ; SI-NEXT: v_cndmask_b32_e32 v15, v31, v15, vcc
2565 ; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:196
2566 ; SI-NEXT: s_waitcnt vmcnt(0)
2567 ; SI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v31
2568 ; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:68
2569 ; SI-NEXT: s_waitcnt vmcnt(0)
2570 ; SI-NEXT: v_cvt_f16_f32_e32 v31, v31
2571 ; SI-NEXT: v_cvt_f32_f16_e32 v31, v31
2572 ; SI-NEXT: v_cndmask_b32_e32 v16, v31, v16, vcc
2573 ; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:200
2574 ; SI-NEXT: s_waitcnt vmcnt(0)
2575 ; SI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v31
2576 ; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:72
2577 ; SI-NEXT: s_waitcnt vmcnt(0)
2578 ; SI-NEXT: v_cvt_f16_f32_e32 v31, v31
2579 ; SI-NEXT: v_cvt_f32_f16_e32 v31, v31
2580 ; SI-NEXT: v_cndmask_b32_e32 v17, v31, v17, vcc
2581 ; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:204
2582 ; SI-NEXT: s_waitcnt vmcnt(0)
2583 ; SI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v31
2584 ; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:76
2585 ; SI-NEXT: s_waitcnt vmcnt(0)
2586 ; SI-NEXT: v_cvt_f16_f32_e32 v31, v31
2587 ; SI-NEXT: v_cvt_f32_f16_e32 v31, v31
2588 ; SI-NEXT: v_cndmask_b32_e32 v18, v31, v18, vcc
2589 ; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:208
2590 ; SI-NEXT: s_waitcnt vmcnt(0)
2591 ; SI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v31
2592 ; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:80
2593 ; SI-NEXT: s_waitcnt vmcnt(0)
2594 ; SI-NEXT: v_cvt_f16_f32_e32 v31, v31
2595 ; SI-NEXT: v_cvt_f32_f16_e32 v31, v31
2596 ; SI-NEXT: v_cndmask_b32_e32 v19, v31, v19, vcc
2597 ; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:212
2598 ; SI-NEXT: s_waitcnt vmcnt(0)
2599 ; SI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v31
2600 ; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:84
2601 ; SI-NEXT: s_waitcnt vmcnt(0)
2602 ; SI-NEXT: v_cvt_f16_f32_e32 v31, v31
2603 ; SI-NEXT: v_cvt_f32_f16_e32 v31, v31
2604 ; SI-NEXT: v_cndmask_b32_e32 v20, v31, v20, vcc
2605 ; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:216
2606 ; SI-NEXT: s_waitcnt vmcnt(0)
2607 ; SI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v31
2608 ; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:88
2609 ; SI-NEXT: s_waitcnt vmcnt(0)
2610 ; SI-NEXT: v_cvt_f16_f32_e32 v31, v31
2611 ; SI-NEXT: v_cvt_f32_f16_e32 v31, v31
2612 ; SI-NEXT: v_cndmask_b32_e32 v21, v31, v21, vcc
2613 ; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:220
2614 ; SI-NEXT: s_waitcnt vmcnt(0)
2615 ; SI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v31
2616 ; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:92
2617 ; SI-NEXT: s_waitcnt vmcnt(0)
2618 ; SI-NEXT: v_cvt_f16_f32_e32 v31, v31
2619 ; SI-NEXT: v_cvt_f32_f16_e32 v31, v31
2620 ; SI-NEXT: v_cndmask_b32_e32 v22, v31, v22, vcc
2621 ; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:224
2622 ; SI-NEXT: s_waitcnt vmcnt(0)
2623 ; SI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v31
2624 ; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:96
2625 ; SI-NEXT: s_waitcnt vmcnt(0)
2626 ; SI-NEXT: v_cvt_f16_f32_e32 v31, v31
2627 ; SI-NEXT: v_cvt_f32_f16_e32 v31, v31
2628 ; SI-NEXT: v_cndmask_b32_e32 v23, v31, v23, vcc
2629 ; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:228
2630 ; SI-NEXT: s_waitcnt vmcnt(0)
2631 ; SI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v31
2632 ; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:100
2633 ; SI-NEXT: s_waitcnt vmcnt(0)
2634 ; SI-NEXT: v_cvt_f16_f32_e32 v31, v31
2635 ; SI-NEXT: v_cvt_f32_f16_e32 v31, v31
2636 ; SI-NEXT: v_cndmask_b32_e32 v24, v31, v24, vcc
2637 ; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:232
2638 ; SI-NEXT: s_waitcnt vmcnt(0)
2639 ; SI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v31
2640 ; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:104
2641 ; SI-NEXT: s_waitcnt vmcnt(0)
2642 ; SI-NEXT: v_cvt_f16_f32_e32 v31, v31
2643 ; SI-NEXT: v_cvt_f32_f16_e32 v31, v31
2644 ; SI-NEXT: v_cndmask_b32_e32 v25, v31, v25, vcc
2645 ; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:236
2646 ; SI-NEXT: s_waitcnt vmcnt(0)
2647 ; SI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v31
2648 ; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:108
2649 ; SI-NEXT: s_waitcnt vmcnt(0)
2650 ; SI-NEXT: v_cvt_f16_f32_e32 v31, v31
2651 ; SI-NEXT: v_cvt_f32_f16_e32 v31, v31
2652 ; SI-NEXT: v_cndmask_b32_e32 v26, v31, v26, vcc
2653 ; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:240
2654 ; SI-NEXT: s_waitcnt vmcnt(0)
2655 ; SI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v31
2656 ; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:112
2657 ; SI-NEXT: s_waitcnt vmcnt(0)
2658 ; SI-NEXT: v_cvt_f16_f32_e32 v31, v31
2659 ; SI-NEXT: v_cvt_f32_f16_e32 v31, v31
2660 ; SI-NEXT: v_cndmask_b32_e32 v27, v31, v27, vcc
2661 ; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:244
2662 ; SI-NEXT: s_waitcnt vmcnt(0)
2663 ; SI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v31
2664 ; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:116
2665 ; SI-NEXT: s_waitcnt vmcnt(0)
2666 ; SI-NEXT: v_cvt_f16_f32_e32 v31, v31
2667 ; SI-NEXT: v_cvt_f32_f16_e32 v31, v31
2668 ; SI-NEXT: v_cndmask_b32_e32 v28, v31, v28, vcc
2669 ; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:248
2670 ; SI-NEXT: s_waitcnt vmcnt(0)
2671 ; SI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v31
2672 ; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:120
2673 ; SI-NEXT: s_waitcnt vmcnt(0)
2674 ; SI-NEXT: v_cvt_f16_f32_e32 v31, v31
2675 ; SI-NEXT: v_cvt_f32_f16_e32 v31, v31
2676 ; SI-NEXT: v_cndmask_b32_e32 v29, v31, v29, vcc
2677 ; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:252
2678 ; SI-NEXT: s_waitcnt vmcnt(0)
2679 ; SI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v31
2680 ; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:124
2681 ; SI-NEXT: s_waitcnt vmcnt(0)
2682 ; SI-NEXT: v_cvt_f16_f32_e32 v31, v31
2683 ; SI-NEXT: v_cvt_f32_f16_e32 v31, v31
2684 ; SI-NEXT: v_cndmask_b32_e32 v30, v31, v30, vcc
2685 ; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:256
2686 ; SI-NEXT: s_waitcnt vmcnt(0)
2687 ; SI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v31
2688 ; SI-NEXT: buffer_load_dword v31, off, s[0:3], s32
2689 ; SI-NEXT: s_waitcnt vmcnt(0)
2690 ; SI-NEXT: v_cvt_f16_f32_e32 v31, v31
2691 ; SI-NEXT: v_cvt_f32_f16_e32 v31, v31
2692 ; SI-NEXT: v_cndmask_b32_e32 v31, v32, v31, vcc
2693 ; SI-NEXT: s_setpc_b64 s[30:31]
2694 ;
2695 ; VI-LABEL: v_vselect_v32f16:
2696 ; VI: ; %bb.0:
2697 ; VI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
2698 ; VI-NEXT: buffer_store_dword v40, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
2699 ; VI-NEXT: buffer_store_dword v41, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
2700 ; VI-NEXT: buffer_store_dword v42, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
2701 ; VI-NEXT: buffer_store_dword v43, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
2702 ; VI-NEXT: buffer_store_dword v44, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
2703 ; VI-NEXT: buffer_store_dword v45, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
2704 ; VI-NEXT: buffer_store_dword v46, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
2705 ; VI-NEXT: buffer_store_dword v47, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
2706 ; VI-NEXT: buffer_store_dword v56, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
2707 ; VI-NEXT: buffer_store_dword v57, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
2708 ; VI-NEXT: buffer_store_dword v58, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
2709 ; VI-NEXT: buffer_store_dword v59, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
2710 ; VI-NEXT: buffer_store_dword v60, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
2711 ; VI-NEXT: buffer_load_dword v36, off, s[0:3], s32 offset:120
2712 ; VI-NEXT: buffer_load_dword v35, off, s[0:3], s32 offset:112
2713 ; VI-NEXT: buffer_load_dword v34, off, s[0:3], s32 offset:104
2714 ; VI-NEXT: buffer_load_dword v33, off, s[0:3], s32 offset:96
2715 ; VI-NEXT: buffer_load_dword v32, off, s[0:3], s32 offset:88
2716 ; VI-NEXT: buffer_load_dword v31, off, s[0:3], s32 offset:80
2717 ; VI-NEXT: buffer_load_dword v39, off, s[0:3], s32 offset:72
2718 ; VI-NEXT: buffer_load_dword v37, off, s[0:3], s32
2719 ; VI-NEXT: buffer_load_dword v55, off, s[0:3], s32 offset:128
2720 ; VI-NEXT: buffer_load_dword v50, off, s[0:3], s32 offset:64
2721 ; VI-NEXT: buffer_load_dword v53, off, s[0:3], s32 offset:56
2722 ; VI-NEXT: buffer_load_dword v54, off, s[0:3], s32 offset:48
2723 ; VI-NEXT: buffer_load_dword v52, off, s[0:3], s32 offset:40
2724 ; VI-NEXT: buffer_load_dword v51, off, s[0:3], s32 offset:32
2725 ; VI-NEXT: buffer_load_dword v49, off, s[0:3], s32 offset:24
2726 ; VI-NEXT: buffer_load_dword v48, off, s[0:3], s32 offset:16
2727 ; VI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:8
2728 ; VI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:124
2729 ; VI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:116
2730 ; VI-NEXT: v_lshrrev_b32_e32 v38, 16, v14
2731 ; VI-NEXT: v_lshrrev_b32_e32 v43, 16, v30
2732 ; VI-NEXT: v_lshrrev_b32_e32 v44, 16, v13
2733 ; VI-NEXT: v_lshrrev_b32_e32 v45, 16, v29
2734 ; VI-NEXT: v_lshrrev_b32_e32 v46, 16, v12
2735 ; VI-NEXT: v_lshrrev_b32_e32 v47, 16, v28
2736 ; VI-NEXT: v_lshrrev_b32_e32 v56, 16, v11
2737 ; VI-NEXT: v_lshrrev_b32_e32 v57, 16, v27
2738 ; VI-NEXT: v_lshrrev_b32_e32 v58, 16, v10
2739 ; VI-NEXT: v_lshrrev_b32_e32 v59, 16, v26
2740 ; VI-NEXT: v_lshrrev_b32_e32 v60, 16, v9
2741 ; VI-NEXT: s_waitcnt vmcnt(14)
2742 ; VI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v36
2743 ; VI-NEXT: v_cndmask_b32_e32 v36, v43, v38, vcc
2744 ; VI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v35
2745 ; VI-NEXT: v_cndmask_b32_e32 v35, v45, v44, vcc
2746 ; VI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v34
2747 ; VI-NEXT: v_cndmask_b32_e32 v34, v47, v46, vcc
2748 ; VI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v33
2749 ; VI-NEXT: v_cndmask_b32_e32 v33, v57, v56, vcc
2750 ; VI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v32
2751 ; VI-NEXT: v_lshrrev_b32_e32 v38, 16, v25
2752 ; VI-NEXT: v_cndmask_b32_e32 v32, v59, v58, vcc
2753 ; VI-NEXT: s_waitcnt vmcnt(13)
2754 ; VI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v31
2755 ; VI-NEXT: v_lshrrev_b32_e32 v43, 16, v8
2756 ; VI-NEXT: v_lshrrev_b32_e32 v44, 16, v24
2757 ; VI-NEXT: v_cndmask_b32_e32 v38, v38, v60, vcc
2758 ; VI-NEXT: s_waitcnt vmcnt(12)
2759 ; VI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v39
2760 ; VI-NEXT: v_lshrrev_b32_e32 v45, 16, v15
2761 ; VI-NEXT: v_cndmask_b32_e32 v39, v44, v43, vcc
2762 ; VI-NEXT: s_waitcnt vmcnt(11)
2763 ; VI-NEXT: v_lshrrev_b32_e32 v31, 16, v37
2764 ; VI-NEXT: s_waitcnt vmcnt(10)
2765 ; VI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v55
2766 ; VI-NEXT: v_cndmask_b32_e32 v31, v31, v45, vcc
2767 ; VI-NEXT: v_lshrrev_b32_e32 v55, 16, v7
2768 ; VI-NEXT: v_lshrrev_b32_e32 v43, 16, v23
2769 ; VI-NEXT: s_waitcnt vmcnt(9)
2770 ; VI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v50
2771 ; VI-NEXT: v_cndmask_b32_e32 v50, v43, v55, vcc
2772 ; VI-NEXT: v_lshrrev_b32_e32 v55, 16, v6
2773 ; VI-NEXT: v_lshrrev_b32_e32 v43, 16, v22
2774 ; VI-NEXT: s_waitcnt vmcnt(8)
2775 ; VI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v53
2776 ; VI-NEXT: v_cndmask_b32_e32 v53, v43, v55, vcc
2777 ; VI-NEXT: v_lshrrev_b32_e32 v55, 16, v5
2778 ; VI-NEXT: v_lshrrev_b32_e32 v43, 16, v21
2779 ; VI-NEXT: s_waitcnt vmcnt(7)
2780 ; VI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v54
2781 ; VI-NEXT: v_cndmask_b32_e32 v54, v43, v55, vcc
2782 ; VI-NEXT: v_lshrrev_b32_e32 v55, 16, v4
2783 ; VI-NEXT: v_lshrrev_b32_e32 v43, 16, v20
2784 ; VI-NEXT: s_waitcnt vmcnt(6)
2785 ; VI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v52
2786 ; VI-NEXT: v_cndmask_b32_e32 v52, v43, v55, vcc
2787 ; VI-NEXT: v_lshrrev_b32_e32 v55, 16, v3
2788 ; VI-NEXT: v_lshrrev_b32_e32 v43, 16, v19
2789 ; VI-NEXT: s_waitcnt vmcnt(5)
2790 ; VI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v51
2791 ; VI-NEXT: v_cndmask_b32_e32 v51, v43, v55, vcc
2792 ; VI-NEXT: v_lshrrev_b32_e32 v55, 16, v2
2793 ; VI-NEXT: v_lshrrev_b32_e32 v43, 16, v18
2794 ; VI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:108
2795 ; VI-NEXT: s_waitcnt vmcnt(5)
2796 ; VI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v49
2797 ; VI-NEXT: v_cndmask_b32_e32 v49, v43, v55, vcc
2798 ; VI-NEXT: buffer_load_dword v55, off, s[0:3], s32 offset:100
2799 ; VI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:92
2800 ; VI-NEXT: v_lshrrev_b32_e32 v43, 16, v1
2801 ; VI-NEXT: v_lshrrev_b32_e32 v46, 16, v17
2802 ; VI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:84
2803 ; VI-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:76
2804 ; VI-NEXT: s_waitcnt vmcnt(8)
2805 ; VI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v48
2806 ; VI-NEXT: v_cndmask_b32_e32 v48, v46, v43, vcc
2807 ; VI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:68
2808 ; VI-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:60
2809 ; VI-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:52
2810 ; VI-NEXT: v_lshrrev_b32_e32 v46, 16, v0
2811 ; VI-NEXT: v_lshrrev_b32_e32 v58, 16, v16
2812 ; VI-NEXT: s_waitcnt vmcnt(10)
2813 ; VI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v40
2814 ; VI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:44
2815 ; VI-NEXT: v_cndmask_b32_e32 v46, v58, v46, vcc
2816 ; VI-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:36
2817 ; VI-NEXT: s_waitcnt vmcnt(11)
2818 ; VI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v41
2819 ; VI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:28
2820 ; VI-NEXT: v_cndmask_b32_e32 v15, v37, v15, vcc
2821 ; VI-NEXT: buffer_load_dword v37, off, s[0:3], s32 offset:20
2822 ; VI-NEXT: s_waitcnt vmcnt(12)
2823 ; VI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v42
2824 ; VI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:12
2825 ; VI-NEXT: v_cndmask_b32_e32 v14, v30, v14, vcc
2826 ; VI-NEXT: buffer_load_dword v30, off, s[0:3], s32 offset:4
2827 ; VI-NEXT: s_waitcnt vmcnt(13)
2828 ; VI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v44
2829 ; VI-NEXT: v_cndmask_b32_e32 v13, v29, v13, vcc
2830 ; VI-NEXT: s_waitcnt vmcnt(12)
2831 ; VI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v55
2832 ; VI-NEXT: v_cndmask_b32_e32 v12, v28, v12, vcc
2833 ; VI-NEXT: s_waitcnt vmcnt(11)
2834 ; VI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v45
2835 ; VI-NEXT: v_cndmask_b32_e32 v11, v27, v11, vcc
2836 ; VI-NEXT: s_waitcnt vmcnt(10)
2837 ; VI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v47
2838 ; VI-NEXT: v_cndmask_b32_e32 v10, v26, v10, vcc
2839 ; VI-NEXT: s_waitcnt vmcnt(9)
2840 ; VI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v56
2841 ; VI-NEXT: v_cndmask_b32_e32 v9, v25, v9, vcc
2842 ; VI-NEXT: s_waitcnt vmcnt(8)
2843 ; VI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v43
2844 ; VI-NEXT: v_cndmask_b32_e32 v8, v24, v8, vcc
2845 ; VI-NEXT: s_waitcnt vmcnt(7)
2846 ; VI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v57
2847 ; VI-NEXT: v_cndmask_b32_e32 v7, v23, v7, vcc
2848 ; VI-NEXT: s_waitcnt vmcnt(6)
2849 ; VI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v59
2850 ; VI-NEXT: v_cndmask_b32_e32 v6, v22, v6, vcc
2851 ; VI-NEXT: s_waitcnt vmcnt(5)
2852 ; VI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v40
2853 ; VI-NEXT: v_cndmask_b32_e32 v5, v21, v5, vcc
2854 ; VI-NEXT: s_waitcnt vmcnt(4)
2855 ; VI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v58
2856 ; VI-NEXT: v_cndmask_b32_e32 v4, v20, v4, vcc
2857 ; VI-NEXT: s_waitcnt vmcnt(3)
2858 ; VI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v41
2859 ; VI-NEXT: v_cndmask_b32_e32 v3, v19, v3, vcc
2860 ; VI-NEXT: s_waitcnt vmcnt(2)
2861 ; VI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v37
2862 ; VI-NEXT: v_cndmask_b32_e32 v2, v18, v2, vcc
2863 ; VI-NEXT: s_waitcnt vmcnt(1)
2864 ; VI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v42
2865 ; VI-NEXT: v_cndmask_b32_e32 v1, v17, v1, vcc
2866 ; VI-NEXT: s_waitcnt vmcnt(0)
2867 ; VI-NEXT: v_cmp_eq_u32_e32 vcc, 0, v30
2868 ; VI-NEXT: v_cndmask_b32_e32 v0, v16, v0, vcc
2869 ; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v46
2870 ; VI-NEXT: buffer_load_dword v60, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
2871 ; VI-NEXT: buffer_load_dword v59, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
2872 ; VI-NEXT: buffer_load_dword v58, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
2873 ; VI-NEXT: buffer_load_dword v57, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
2874 ; VI-NEXT: buffer_load_dword v56, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
2875 ; VI-NEXT: buffer_load_dword v47, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
2876 ; VI-NEXT: buffer_load_dword v46, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
2877 ; VI-NEXT: buffer_load_dword v45, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
2878 ; VI-NEXT: buffer_load_dword v44, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
2879 ; VI-NEXT: buffer_load_dword v43, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
2880 ; VI-NEXT: buffer_load_dword v42, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
2881 ; VI-NEXT: buffer_load_dword v41, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
2882 ; VI-NEXT: buffer_load_dword v40, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
2883 ; VI-NEXT: v_or_b32_sdwa v0, v0, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
2884 ; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v48
2885 ; VI-NEXT: v_or_b32_sdwa v1, v1, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
2886 ; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v49
2887 ; VI-NEXT: v_or_b32_sdwa v2, v2, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
2888 ; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v51
2889 ; VI-NEXT: v_or_b32_sdwa v3, v3, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
2890 ; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v52
2891 ; VI-NEXT: v_or_b32_sdwa v4, v4, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
2892 ; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v54
2893 ; VI-NEXT: v_or_b32_sdwa v5, v5, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
2894 ; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v53
2895 ; VI-NEXT: v_or_b32_sdwa v6, v6, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
2896 ; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v50
2897 ; VI-NEXT: v_or_b32_sdwa v7, v7, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
2898 ; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v39
2899 ; VI-NEXT: v_or_b32_sdwa v8, v8, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
2900 ; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v38
2901 ; VI-NEXT: v_or_b32_sdwa v9, v9, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
2902 ; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v32
2903 ; VI-NEXT: v_or_b32_sdwa v10, v10, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
2904 ; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v33
2905 ; VI-NEXT: v_or_b32_sdwa v11, v11, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
2906 ; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v34
2907 ; VI-NEXT: v_or_b32_sdwa v12, v12, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
2908 ; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v35
2909 ; VI-NEXT: v_or_b32_sdwa v13, v13, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
2910 ; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v36
2911 ; VI-NEXT: v_or_b32_sdwa v14, v14, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
2912 ; VI-NEXT: v_lshlrev_b32_e32 v16, 16, v31
2913 ; VI-NEXT: v_or_b32_sdwa v15, v15, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
2914 ; VI-NEXT: s_waitcnt vmcnt(0)
2915 ; VI-NEXT: s_setpc_b64 s[30:31]
2916 ;
2917 ; GFX11-LABEL: v_vselect_v32f16:
2918 ; GFX11: ; %bb.0:
2919 ; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
2920 ; GFX11-NEXT: s_clause 0x1f
2921 ; GFX11-NEXT: scratch_load_b32 v31, off, s32 offset:120
2922 ; GFX11-NEXT: scratch_load_b32 v32, off, s32 offset:112
2923 ; GFX11-NEXT: scratch_load_b32 v33, off, s32
2924 ; GFX11-NEXT: scratch_load_b32 v34, off, s32 offset:104
2925 ; GFX11-NEXT: scratch_load_b32 v35, off, s32 offset:96
2926 ; GFX11-NEXT: scratch_load_b32 v36, off, s32 offset:88
2927 ; GFX11-NEXT: scratch_load_b32 v37, off, s32 offset:80
2928 ; GFX11-NEXT: scratch_load_b32 v38, off, s32 offset:72
2929 ; GFX11-NEXT: scratch_load_b32 v39, off, s32 offset:64
2930 ; GFX11-NEXT: scratch_load_b32 v48, off, s32 offset:56
2931 ; GFX11-NEXT: scratch_load_b32 v49, off, s32 offset:48
2932 ; GFX11-NEXT: scratch_load_b32 v50, off, s32 offset:40
2933 ; GFX11-NEXT: scratch_load_b32 v51, off, s32 offset:32
2934 ; GFX11-NEXT: scratch_load_b32 v52, off, s32 offset:24
2935 ; GFX11-NEXT: scratch_load_b32 v53, off, s32 offset:16
2936 ; GFX11-NEXT: scratch_load_b32 v54, off, s32 offset:8
2937 ; GFX11-NEXT: scratch_load_b32 v55, off, s32 offset:124
2938 ; GFX11-NEXT: scratch_load_b32 v64, off, s32 offset:116
2939 ; GFX11-NEXT: scratch_load_b32 v65, off, s32 offset:108
2940 ; GFX11-NEXT: scratch_load_b32 v66, off, s32 offset:100
2941 ; GFX11-NEXT: scratch_load_b32 v67, off, s32 offset:92
2942 ; GFX11-NEXT: scratch_load_b32 v68, off, s32 offset:84
2943 ; GFX11-NEXT: scratch_load_b32 v69, off, s32 offset:76
2944 ; GFX11-NEXT: scratch_load_b32 v70, off, s32 offset:68
2945 ; GFX11-NEXT: scratch_load_b32 v71, off, s32 offset:60
2946 ; GFX11-NEXT: scratch_load_b32 v80, off, s32 offset:52
2947 ; GFX11-NEXT: scratch_load_b32 v81, off, s32 offset:44
2948 ; GFX11-NEXT: scratch_load_b32 v82, off, s32 offset:36
2949 ; GFX11-NEXT: scratch_load_b32 v83, off, s32 offset:28
2950 ; GFX11-NEXT: scratch_load_b32 v84, off, s32 offset:12
2951 ; GFX11-NEXT: scratch_load_b32 v85, off, s32 offset:4
2952 ; GFX11-NEXT: scratch_load_b32 v86, off, s32 offset:20
2953 ; GFX11-NEXT: scratch_load_b32 v87, off, s32 offset:128
2954 ; GFX11-NEXT: v_lshrrev_b32_e32 v97, 16, v14
2955 ; GFX11-NEXT: v_lshrrev_b32_e32 v98, 16, v30
2956 ; GFX11-NEXT: v_lshrrev_b32_e32 v99, 16, v13
2957 ; GFX11-NEXT: v_lshrrev_b32_e32 v100, 16, v29
2958 ; GFX11-NEXT: v_lshrrev_b32_e32 v101, 16, v12
2959 ; GFX11-NEXT: v_lshrrev_b32_e32 v102, 16, v28
2960 ; GFX11-NEXT: v_lshrrev_b32_e32 v103, 16, v11
2961 ; GFX11-NEXT: v_lshrrev_b32_e32 v112, 16, v27
2962 ; GFX11-NEXT: v_lshrrev_b32_e32 v113, 16, v10
2963 ; GFX11-NEXT: v_lshrrev_b32_e32 v114, 16, v26
2964 ; GFX11-NEXT: v_lshrrev_b32_e32 v115, 16, v9
2965 ; GFX11-NEXT: v_lshrrev_b32_e32 v116, 16, v25
2966 ; GFX11-NEXT: v_lshrrev_b32_e32 v117, 16, v8
2967 ; GFX11-NEXT: v_lshrrev_b32_e32 v118, 16, v24
2968 ; GFX11-NEXT: v_lshrrev_b32_e32 v119, 16, v7
2969 ; GFX11-NEXT: v_lshrrev_b32_e32 v128, 16, v23
2970 ; GFX11-NEXT: v_lshrrev_b32_e32 v129, 16, v6
2971 ; GFX11-NEXT: v_lshrrev_b32_e32 v130, 16, v22
2972 ; GFX11-NEXT: v_lshrrev_b32_e32 v131, 16, v5
2973 ; GFX11-NEXT: v_lshrrev_b32_e32 v132, 16, v21
2974 ; GFX11-NEXT: v_lshrrev_b32_e32 v133, 16, v4
2975 ; GFX11-NEXT: v_lshrrev_b32_e32 v134, 16, v20
2976 ; GFX11-NEXT: v_lshrrev_b32_e32 v135, 16, v3
2977 ; GFX11-NEXT: v_lshrrev_b32_e32 v144, 16, v19
2978 ; GFX11-NEXT: v_lshrrev_b32_e32 v145, 16, v2
2979 ; GFX11-NEXT: v_lshrrev_b32_e32 v146, 16, v18
2980 ; GFX11-NEXT: v_lshrrev_b32_e32 v147, 16, v1
2981 ; GFX11-NEXT: v_lshrrev_b32_e32 v96, 16, v15
2982 ; GFX11-NEXT: s_waitcnt vmcnt(32)
2983 ; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v31
2984 ; GFX11-NEXT: v_lshrrev_b32_e32 v31, 16, v17
2985 ; GFX11-NEXT: v_cndmask_b32_e32 v97, v98, v97, vcc_lo
2986 ; GFX11-NEXT: s_waitcnt vmcnt(31)
2987 ; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v32
2988 ; GFX11-NEXT: v_lshrrev_b32_e32 v98, 16, v0
2989 ; GFX11-NEXT: v_lshrrev_b32_e32 v32, 16, v16
2990 ; GFX11-NEXT: v_cndmask_b32_e32 v99, v100, v99, vcc_lo
2991 ; GFX11-NEXT: s_waitcnt vmcnt(29)
2992 ; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v34
2993 ; GFX11-NEXT: v_lshrrev_b32_e32 v100, 16, v33
2994 ; GFX11-NEXT: v_cndmask_b32_e32 v34, v102, v101, vcc_lo
2995 ; GFX11-NEXT: s_waitcnt vmcnt(28)
2996 ; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v35
2997 ; GFX11-NEXT: v_cndmask_b32_e32 v35, v112, v103, vcc_lo
2998 ; GFX11-NEXT: s_waitcnt vmcnt(27)
2999 ; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v36
3000 ; GFX11-NEXT: v_cndmask_b32_e32 v36, v114, v113, vcc_lo
3001 ; GFX11-NEXT: s_waitcnt vmcnt(26)
3002 ; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v37
3003 ; GFX11-NEXT: v_cndmask_b32_e32 v37, v116, v115, vcc_lo
3004 ; GFX11-NEXT: s_waitcnt vmcnt(25)
3005 ; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v38
3006 ; GFX11-NEXT: v_cndmask_b32_e32 v38, v118, v117, vcc_lo
3007 ; GFX11-NEXT: s_waitcnt vmcnt(24)
3008 ; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v39
3009 ; GFX11-NEXT: v_cndmask_b32_e32 v39, v128, v119, vcc_lo
3010 ; GFX11-NEXT: s_waitcnt vmcnt(23)
3011 ; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v48
3012 ; GFX11-NEXT: v_cndmask_b32_e32 v48, v130, v129, vcc_lo
3013 ; GFX11-NEXT: s_waitcnt vmcnt(22)
3014 ; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v49
3015 ; GFX11-NEXT: v_cndmask_b32_e32 v49, v132, v131, vcc_lo
3016 ; GFX11-NEXT: s_waitcnt vmcnt(21)
3017 ; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v50
3018 ; GFX11-NEXT: v_cndmask_b32_e32 v50, v134, v133, vcc_lo
3019 ; GFX11-NEXT: s_waitcnt vmcnt(20)
3020 ; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v51
3021 ; GFX11-NEXT: v_cndmask_b32_e32 v51, v144, v135, vcc_lo
3022 ; GFX11-NEXT: s_waitcnt vmcnt(19)
3023 ; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v52
3024 ; GFX11-NEXT: v_cndmask_b32_e32 v52, v146, v145, vcc_lo
3025 ; GFX11-NEXT: s_waitcnt vmcnt(18)
3026 ; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v53
3027 ; GFX11-NEXT: v_cndmask_b32_e32 v31, v31, v147, vcc_lo
3028 ; GFX11-NEXT: s_waitcnt vmcnt(17)
3029 ; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v54
3030 ; GFX11-NEXT: v_cndmask_b32_e32 v32, v32, v98, vcc_lo
3031 ; GFX11-NEXT: s_waitcnt vmcnt(16)
3032 ; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v55
3033 ; GFX11-NEXT: v_cndmask_b32_e32 v15, v33, v15, vcc_lo
3034 ; GFX11-NEXT: s_waitcnt vmcnt(15)
3035 ; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v64
3036 ; GFX11-NEXT: v_cndmask_b32_e32 v14, v30, v14, vcc_lo
3037 ; GFX11-NEXT: s_waitcnt vmcnt(14)
3038 ; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v65
3039 ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2)
3040 ; GFX11-NEXT: v_perm_b32 v14, v97, v14, 0x5040100
3041 ; GFX11-NEXT: v_cndmask_b32_e32 v13, v29, v13, vcc_lo
3042 ; GFX11-NEXT: s_waitcnt vmcnt(13)
3043 ; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v66
3044 ; GFX11-NEXT: v_cndmask_b32_e32 v12, v28, v12, vcc_lo
3045 ; GFX11-NEXT: s_waitcnt vmcnt(12)
3046 ; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v67
3047 ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2)
3048 ; GFX11-NEXT: v_perm_b32 v12, v34, v12, 0x5040100
3049 ; GFX11-NEXT: v_cndmask_b32_e32 v11, v27, v11, vcc_lo
3050 ; GFX11-NEXT: s_waitcnt vmcnt(11)
3051 ; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v68
3052 ; GFX11-NEXT: v_cndmask_b32_e32 v10, v26, v10, vcc_lo
3053 ; GFX11-NEXT: s_waitcnt vmcnt(10)
3054 ; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v69
3055 ; GFX11-NEXT: v_perm_b32 v13, v99, v13, 0x5040100
3056 ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3)
3057 ; GFX11-NEXT: v_perm_b32 v10, v36, v10, 0x5040100
3058 ; GFX11-NEXT: v_cndmask_b32_e32 v9, v25, v9, vcc_lo
3059 ; GFX11-NEXT: s_waitcnt vmcnt(9)
3060 ; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v70
3061 ; GFX11-NEXT: v_cndmask_b32_e32 v8, v24, v8, vcc_lo
3062 ; GFX11-NEXT: s_waitcnt vmcnt(8)
3063 ; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v71
3064 ; GFX11-NEXT: v_perm_b32 v11, v35, v11, 0x5040100
3065 ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3)
3066 ; GFX11-NEXT: v_perm_b32 v8, v38, v8, 0x5040100
3067 ; GFX11-NEXT: v_cndmask_b32_e32 v7, v23, v7, vcc_lo
3068 ; GFX11-NEXT: s_waitcnt vmcnt(7)
3069 ; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v80
3070 ; GFX11-NEXT: v_cndmask_b32_e32 v6, v22, v6, vcc_lo
3071 ; GFX11-NEXT: s_waitcnt vmcnt(6)
3072 ; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v81
3073 ; GFX11-NEXT: v_perm_b32 v9, v37, v9, 0x5040100
3074 ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3)
3075 ; GFX11-NEXT: v_perm_b32 v6, v48, v6, 0x5040100
3076 ; GFX11-NEXT: v_cndmask_b32_e32 v5, v21, v5, vcc_lo
3077 ; GFX11-NEXT: s_waitcnt vmcnt(5)
3078 ; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v82
3079 ; GFX11-NEXT: v_cndmask_b32_e32 v4, v20, v4, vcc_lo
3080 ; GFX11-NEXT: s_waitcnt vmcnt(4)
3081 ; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v83
3082 ; GFX11-NEXT: v_perm_b32 v7, v39, v7, 0x5040100
3083 ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3)
3084 ; GFX11-NEXT: v_perm_b32 v4, v50, v4, 0x5040100
3085 ; GFX11-NEXT: v_cndmask_b32_e32 v3, v19, v3, vcc_lo
3086 ; GFX11-NEXT: s_waitcnt vmcnt(3)
3087 ; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v84
3088 ; GFX11-NEXT: v_cndmask_b32_e32 v1, v17, v1, vcc_lo
3089 ; GFX11-NEXT: s_waitcnt vmcnt(2)
3090 ; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v85
3091 ; GFX11-NEXT: v_cndmask_b32_e32 v0, v16, v0, vcc_lo
3092 ; GFX11-NEXT: s_waitcnt vmcnt(1)
3093 ; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v86
3094 ; GFX11-NEXT: v_perm_b32 v5, v49, v5, 0x5040100
3095 ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_3)
3096 ; GFX11-NEXT: v_perm_b32 v0, v32, v0, 0x5040100
3097 ; GFX11-NEXT: v_cndmask_b32_e32 v2, v18, v2, vcc_lo
3098 ; GFX11-NEXT: s_waitcnt vmcnt(0)
3099 ; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, 0, v87
3100 ; GFX11-NEXT: v_perm_b32 v3, v51, v3, 0x5040100
3101 ; GFX11-NEXT: v_perm_b32 v2, v52, v2, 0x5040100
3102 ; GFX11-NEXT: v_cndmask_b32_e32 v16, v100, v96, vcc_lo
3103 ; GFX11-NEXT: v_perm_b32 v1, v31, v1, 0x5040100
3104 ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2)
3105 ; GFX11-NEXT: v_perm_b32 v15, v16, v15, 0x5040100
3106 ; GFX11-NEXT: s_setpc_b64 s[30:31]
3107 %cmp = icmp eq <32 x i32> %cond, zeroinitializer
3108 %select = select <32 x i1> %cmp, <32 x half> %a, <32 x half> %b
3109 ret <32 x half> %select
3110 }