salu-to-valu.ll - OpenGrok cross reference for /external/llvm/test/CodeGen/AMDGPU/salu-to-valu.ll

Lines Matching refs:GCN
1 …-mcpu=tahiti -verify-machineinstrs < %s | FileCheck -check-prefix=GCN -check-prefix=GCN-NOHSA -che…
2 …mcpu=bonaire -verify-machineinstrs < %s | FileCheck -check-prefix=GCN -check-prefix=GCN-NOHSA -che…
3 …ire -verify-machineinstrs < %s | FileCheck -check-prefix=GCN -check-prefix=CI --check-prefix=GCN-H…
15 ; GCN-LABEL: {{^}}mubuf:
18 ; GCN-NOT: s_mov_b64 s[{{[0-9]+:[0-9]+}}], v
22 ; GCN-NOHSA: buffer_load_ubyte v{{[0-9]+}}, v[{{[0-9]+:[0-9]+}}], s[{{[0-9]+:[0-9]+}}], 0 addr64
23 ; GCN-NOHSA: buffer_load_ubyte v{{[0-9]+}}, v[{{[0-9]+:[0-9]+}}], s[{{[0-9]+:[0-9]+}}], 0 addr64
24 ; GCN-HSA: flat_load_ubyte v{{[0-9]+}}, v[{{[0-9]+:[0-9]+}}
25 ; GCN-HSA: flat_load_ubyte v{{[0-9]+}}, v[{{[0-9]+:[0-9]+}}
56 ; GCN-LABEL: {{^}}smrd_valu:
58 ; GCN: v_readfirstlane_b32 s[[PTR_LO:[0-9]+]], v{{[0-9]+}}
59 ; GCN: v_readfirstlane_b32 s[[PTR_HI:[0-9]+]], v{{[0-9]+}}
65 ; GCN: v_mov_b32_e32 [[V_OUT:v[0-9]+]], [[OUT]]
66 ; GCN-NOHSA: buffer_store_dword [[V_OUT]]
67 ; GCN-HSA: flat_store_dword {{.*}}, [[V_OUT]]
92 ; GCN-LABEL: {{^}}smrd_valu2:
93 ; GCN-NOHSA-NOT: v_add
94 ; GCN-NOHSA: buffer_load_dword v{{[0-9]+}}, v{{\[[0-9]+:[0-9]+\]}}, s[{{[0-9]+:[0-9]+}}], 0 addr64 …
95 ; GCN-HSA: flat_load_dword v{{[0-9]+}}, v[{{[0-9]+:[0-9]+}}]
107 ; GCN-LABEL: {{^}}smrd_valu_ci_offset:
108 ; GCN-NOHSA-NOT: v_add
109 ; GCN-NOHSA: s_movk_i32 [[OFFSET:s[0-9]+]], 0x4e20{{$}}
110 ; GCN-NOHSA-NOT: v_add
111 ; GCN-NOHSA: buffer_load_dword v{{[0-9]+}}, v{{\[[0-9]+:[0-9]+\]}}, s[{{[0-9]+:[0-9]+}}], [[OFFSET]…
112 ; GCN-NOHSA: v_add_i32_e32
113 ; GCN-NOHSA: buffer_store_dword
114 ; GCN-HSA: flat_load_dword v{{[0-9]+}}, v[{{[0-9]+:[0-9]+}}]
115 ; GCN-HSA: flat_store_dword v[{{[0-9]+:[0-9]+}}], v{{[0-9]+}}
127 ; GCN-LABEL: {{^}}smrd_valu_ci_offset_x2:
128 ; GCN-NOHSA-NOT: v_add
129 ; GCN-NOHSA: s_mov_b32 [[OFFSET:s[0-9]+]], 0x9c40{{$}}
130 ; GCN-NOHSA-NOT: v_add
131 ; GCN-NOHSA: buffer_load_dwordx2 v{{\[[0-9]+:[0-9]+\]}}, v{{\[[0-9]+:[0-9]+\]}}, s[{{[0-9]+:[0-9]+}…
132 ; GCN-NOHSA: v_or_b32_e32 {{v[0-9]+}}, {{s[0-9]+}}, {{v[0-9]+}}
133 ; GCN-NOHSA: v_or_b32_e32 {{v[0-9]+}}, {{s[0-9]+}}, {{v[0-9]+}}
134 ; GCN-NOHSA: buffer_store_dwordx2
135 ; GCN-HSA: flat_load_dwordx2 v[{{[0-9]+:[0-9]+}}], v[{{[0-9]+:[0-9]+}}]
147 ; GCN-LABEL: {{^}}smrd_valu_ci_offset_x4:
148 ; GCN-NOHSA-NOT: v_add
149 ; GCN-NOHSA: s_movk_i32 [[OFFSET:s[0-9]+]], 0x4d20{{$}}
150 ; GCN-NOHSA-NOT: v_add
151 ; GCN-NOHSA: buffer_load_dwordx4 v{{\[[0-9]+:[0-9]+\]}}, v{{\[[0-9]+:[0-9]+\]}}, s[{{[0-9]+:[0-9]+}…
152 ; GCN-NOHSA: v_or_b32_e32 {{v[0-9]+}}, {{s[0-9]+}}, {{v[0-9]+}}
153 ; GCN-NOHSA: v_or_b32_e32 {{v[0-9]+}}, {{s[0-9]+}}, {{v[0-9]+}}
154 ; GCN-NOHSA: v_or_b32_e32 {{v[0-9]+}}, {{s[0-9]+}}, {{v[0-9]+}}
155 ; GCN-NOHSA: v_or_b32_e32 {{v[0-9]+}}, {{s[0-9]+}}, {{v[0-9]+}}
156 ; GCN-NOHSA: buffer_store_dwordx4
157 ; GCN-HSA: flat_load_dwordx4 v[{{[0-9]+:[0-9]+}}], v[{{[0-9]+:[0-9]+}}]
172 ; GCN-LABEL: {{^}}smrd_valu_ci_offset_x8:
173 ; GCN-NOHSA-NOT: v_add
174 ; GCN-NOHSA: s_mov_b32 [[OFFSET0:s[0-9]+]], 0x9a40{{$}}
175 ; GCN-NOHSA-NOT: v_add
176 ; GCN-NOHSA: buffer_load_dwordx4 v{{\[[0-9]+:[0-9]+\]}}, v{{\[[0-9]+:[0-9]+\]}}, s[{{[0-9]+:[0-9]+}…
177 ; GCN-NOHSA-NOT: v_add
178 ; GCN-NOHSA: s_mov_b32 [[OFFSET1:s[0-9]+]], 0x9a50{{$}}
179 ; GCN-NOHSA-NOT: v_add
180 ; GCN-NOHSA: buffer_load_dwordx4 v{{\[[0-9]+:[0-9]+\]}}, v{{\[[0-9]+:[0-9]+\]}}, s[{{[0-9]+:[0-9]+}…
182 ; GCN-NOHSA: v_or_b32_e32 {{v[0-9]+}}, {{s[0-9]+}}, {{v[0-9]+}}
183 ; GCN-NOHSA: v_or_b32_e32 {{v[0-9]+}}, {{s[0-9]+}}, {{v[0-9]+}}
184 ; GCN-NOHSA: v_or_b32_e32 {{v[0-9]+}}, {{s[0-9]+}}, {{v[0-9]+}}
185 ; GCN-NOHSA: v_or_b32_e32 {{v[0-9]+}}, {{s[0-9]+}}, {{v[0-9]+}}
186 ; GCN-NOHSA: v_or_b32_e32 {{v[0-9]+}}, {{s[0-9]+}}, {{v[0-9]+}}
187 ; GCN-NOHSA: v_or_b32_e32 {{v[0-9]+}}, {{s[0-9]+}}, {{v[0-9]+}}
188 ; GCN-NOHSA: v_or_b32_e32 {{v[0-9]+}}, {{s[0-9]+}}, {{v[0-9]+}}
189 ; GCN-NOHSA: v_or_b32_e32 {{v[0-9]+}}, {{s[0-9]+}}, {{v[0-9]+}}
190 ; GCN-NOHSA: buffer_store_dwordx4
191 ; GCN-NOHSA: buffer_store_dwordx4
192 ; GCN-HSA: flat_load_dwordx4
193 ; GCN-HSA: flat_load_dwordx4
205 ; GCN-LABEL: {{^}}smrd_valu_ci_offset_x16:
207 ; GCN-NOHSA-DAG: s_mov_b32 [[OFFSET0:s[0-9]+]], 0x13480{{$}}
208 ; GCN-NOHSA-DAG: buffer_load_dwordx4 v{{\[[0-9]+:[0-9]+\]}}, v{{\[[0-9]+:[0-9]+\]}}, s[{{[0-9]+:[0-…
209 ; GCN-NOHSA-DAG: s_mov_b32 [[OFFSET1:s[0-9]+]], 0x13490{{$}}
210 ; GCN-NOHSA-DAG: buffer_load_dwordx4 v{{\[[0-9]+:[0-9]+\]}}, v{{\[[0-9]+:[0-9]+\]}}, s[{{[0-9]+:[0-…
211 ; GCN-NOHSA-DAG: s_mov_b32 [[OFFSET2:s[0-9]+]], 0x134a0{{$}}
212 ; GCN-NOHSA-DAG: buffer_load_dwordx4 v{{\[[0-9]+:[0-9]+\]}}, v{{\[[0-9]+:[0-9]+\]}}, s[{{[0-9]+:[0-…
213 ; GCN-NOHSA-DAG: s_mov_b32 [[OFFSET3:s[0-9]+]], 0x134b0{{$}}
214 ; GCN-NOHSA-DAG: buffer_load_dwordx4 v{{\[[0-9]+:[0-9]+\]}}, v{{\[[0-9]+:[0-9]+\]}}, s[{{[0-9]+:[0-…
216 ; GCN-NOHSA: v_or_b32_e32 {{v[0-9]+}}, {{s[0-9]+}}, {{v[0-9]+}}
217 ; GCN-NOHSA: v_or_b32_e32 {{v[0-9]+}}, {{s[0-9]+}}, {{v[0-9]+}}
218 ; GCN-NOHSA: v_or_b32_e32 {{v[0-9]+}}, {{s[0-9]+}}, {{v[0-9]+}}
219 ; GCN-NOHSA: v_or_b32_e32 {{v[0-9]+}}, {{s[0-9]+}}, {{v[0-9]+}}
220 ; GCN-NOHSA: v_or_b32_e32 {{v[0-9]+}}, {{s[0-9]+}}, {{v[0-9]+}}
221 ; GCN-NOHSA: v_or_b32_e32 {{v[0-9]+}}, {{s[0-9]+}}, {{v[0-9]+}}
222 ; GCN-NOHSA: v_or_b32_e32 {{v[0-9]+}}, {{s[0-9]+}}, {{v[0-9]+}}
223 ; GCN-NOHSA: v_or_b32_e32 {{v[0-9]+}}, {{s[0-9]+}}, {{v[0-9]+}}
224 ; GCN-NOHSA: buffer_store_dwordx4
225 ; GCN-NOHSA: buffer_store_dwordx4
226 ; GCN-NOHSA: buffer_store_dwordx4
227 ; GCN-NOHSA: buffer_store_dwordx4
229 ; GCN-HSA: flat_load_dwordx4
230 ; GCN-HSA: flat_load_dwordx4
231 ; GCN-HSA: flat_load_dwordx4
232 ; GCN-HSA: flat_load_dwordx4
234 ; GCN: s_endpgm
246 ; GCN-LABEL: {{^}}smrd_valu2_salu_user:
247 ; GCN-NOHSA: buffer_load_dword [[MOVED:v[0-9]+]], v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0…
248 ; GCN-HSA: flat_load_dword [[MOVED:v[0-9]+]], v[{{[0-9+:[0-9]+}}]
249 ; GCN: v_add_i32_e32 [[ADD:v[0-9]+]], vcc, s{{[0-9]+}}, [[MOVED]]
250 ; GCN-NOHSA: buffer_store_dword [[ADD]]
251 ; GCN-HSA: flat_store_dword {{.*}}, [[ADD]]
263 ; GCN-LABEL: {{^}}smrd_valu2_max_smrd_offset:
264 ; GCN-NOHSA: buffer_load_dword v{{[0-9]+}}, v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0 addr6…
265 ; GCN-HSA flat_load_dword v{{[0-9]}}, v{{[0-9]+:[0-9]+}}
276 ; GCN-LABEL: {{^}}smrd_valu2_mubuf_offset:
277 ; GCN-NOHSA-NOT: v_add
278 ; GCN-NOHSA: buffer_load_dword v{{[0-9]+}}, v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, 0 addr6…
279 ; GCN-HSA: flat_load_dword v{{[0-9]}}, v[{{[0-9]+:[0-9]+}}]
290 ; GCN-LABEL: {{^}}s_load_imm_v8i32:
291 ; GCN-NOHSA: buffer_load_dwordx4
292 ; GCN-NOHSA: buffer_load_dwordx4
293 ; GCN-HSA: flat_load_dwordx4
294 ; GCN-HSA: flat_load_dwordx4
305 ; GCN-LABEL: {{^}}s_load_imm_v8i32_salu_user:
306 ; GCN-NOHSA: buffer_load_dwordx4
307 ; GCN-NOHSA: buffer_load_dwordx4
308 ; GCN-NOHSA: v_add_i32_e32
309 ; GCN-NOHSA: v_add_i32_e32
310 ; GCN-NOHSA: v_add_i32_e32
311 ; GCN-NOHSA: v_add_i32_e32
312 ; GCN-NOHSA: v_add_i32_e32
313 ; GCN-NOHSA: v_add_i32_e32
314 ; GCN-NOHSA: v_add_i32_e32
315 ; GCN-NOHSA: buffer_store_dword
316 ; GCN-HSA: flat_load_dwordx4
317 ; GCN-HSA: flat_load_dwordx4
346 ; GCN-LABEL: {{^}}s_load_imm_v16i32:
347 ; GCN-NOHSA: buffer_load_dwordx4
348 ; GCN-NOHSA: buffer_load_dwordx4
349 ; GCN-NOHSA: buffer_load_dwordx4
350 ; GCN-NOHSA: buffer_load_dwordx4
351 ; GCN-HSA: flat_load_dwordx4
352 ; GCN-HSA: flat_load_dwordx4
353 ; GCN-HSA: flat_load_dwordx4
354 ; GCN-HSA: flat_load_dwordx4
365 ; GCN-LABEL: {{^}}s_load_imm_v16i32_salu_user:
366 ; GCN-NOHSA: buffer_load_dwordx4
367 ; GCN-NOHSA: buffer_load_dwordx4
368 ; GCN-NOHSA: buffer_load_dwordx4
369 ; GCN-NOHSA: buffer_load_dwordx4
370 ; GCN-NOHSA: v_add_i32_e32
371 ; GCN-NOHSA: v_add_i32_e32
372 ; GCN-NOHSA: v_add_i32_e32
373 ; GCN-NOHSA: v_add_i32_e32
374 ; GCN-NOHSA: v_add_i32_e32
375 ; GCN-NOHSA: v_add_i32_e32
376 ; GCN-NOHSA: v_add_i32_e32
377 ; GCN-NOHSA: v_add_i32_e32
378 ; GCN-NOHSA: v_add_i32_e32
379 ; GCN-NOHSA: v_add_i32_e32
380 ; GCN-NOHSA: v_add_i32_e32
381 ; GCN-NOHSA: v_add_i32_e32
382 ; GCN-NOHSA: v_add_i32_e32
383 ; GCN-NOHSA: v_add_i32_e32
384 ; GCN-NOHSA: v_add_i32_e32
385 ; GCN-NOHSA: buffer_store_dword
386 ; GCN-HSA: flat_load_dwordx4
387 ; GCN-HSA: flat_load_dwordx4
388 ; GCN-HSA: flat_load_dwordx4
389 ; GCN-HSA: flat_load_dwordx4
437 ; GCN: s_load_dword [[SGPR:s[0-9]+]]
438 ; GCN: v_cmp_le_u32_e32 vcc, [[SGPR]], v{{[0-9]+}}
439 ; GCN: s_and_b64 vcc, exec, vcc
440 ; GCN: s_cbranch_vccnz [[EXIT:[A-Z0-9_]+]]
441 ; GCN: v_mov_b32_e32 [[ONE:v[0-9]+]], 1
442 ; GCN-NOHSA: buffer_store_dword [[ONE]]
443 ; GCN-HSA: flat_store_dword v[{{[0-9]+:[0-9]+}}], [[ONE]]
444 ; GCN; {{^}}[[EXIT]]:
445 ; GCN: s_endpgm