Lines Matching refs:CU
4 …sa -mcpu=gfx1010 -mattr=+cumode -verify-machineinstrs < %s | FileCheck --check-prefixes=GFX10-CU %s
7 …sa -mcpu=gfx1100 -mattr=+cumode -verify-machineinstrs < %s | FileCheck --check-prefixes=GFX11-CU %s
38 ; GFX10-CU-LABEL: flat_nontemporal_load_0:
39 ; GFX10-CU: ; %bb.0: ; %entry
40 ; GFX10-CU-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x0
41 ; GFX10-CU-NEXT: s_waitcnt lgkmcnt(0)
42 ; GFX10-CU-NEXT: v_mov_b32_e32 v0, s0
43 ; GFX10-CU-NEXT: v_mov_b32_e32 v1, s1
44 ; GFX10-CU-NEXT: flat_load_dword v2, v[0:1] glc dlc
45 ; GFX10-CU-NEXT: s_waitcnt vmcnt(0)
46 ; GFX10-CU-NEXT: v_mov_b32_e32 v0, s2
47 ; GFX10-CU-NEXT: v_mov_b32_e32 v1, s3
48 ; GFX10-CU-NEXT: s_waitcnt lgkmcnt(0)
49 ; GFX10-CU-NEXT: flat_store_dword v[0:1], v2
50 ; GFX10-CU-NEXT: s_endpgm
79 ; GFX11-CU-LABEL: flat_nontemporal_load_0:
80 ; GFX11-CU: ; %bb.0: ; %entry
81 ; GFX11-CU-NEXT: s_load_b128 s[0:3], s[0:1], 0x0
82 ; GFX11-CU-NEXT: s_waitcnt lgkmcnt(0)
83 ; GFX11-CU-NEXT: v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
84 ; GFX11-CU-NEXT: flat_load_b32 v2, v[0:1] glc dlc
85 ; GFX11-CU-NEXT: s_waitcnt vmcnt(0)
86 ; GFX11-CU-NEXT: v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
87 ; GFX11-CU-NEXT: s_waitcnt lgkmcnt(0)
88 ; GFX11-CU-NEXT: flat_store_b32 v[0:1], v2
89 ; GFX11-CU-NEXT: s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
90 ; GFX11-CU-NEXT: s_endpgm
130 ; GFX10-CU-LABEL: flat_nontemporal_load_1:
131 ; GFX10-CU: ; %bb.0: ; %entry
132 ; GFX10-CU-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x0
133 ; GFX10-CU-NEXT: v_lshlrev_b32_e32 v0, 2, v0
134 ; GFX10-CU-NEXT: s_waitcnt lgkmcnt(0)
135 ; GFX10-CU-NEXT: v_add_co_u32 v0, s0, s0, v0
136 ; GFX10-CU-NEXT: v_add_co_ci_u32_e64 v1, s0, s1, 0, s0
137 ; GFX10-CU-NEXT: flat_load_dword v2, v[0:1] glc dlc
138 ; GFX10-CU-NEXT: s_waitcnt vmcnt(0)
139 ; GFX10-CU-NEXT: v_mov_b32_e32 v0, s2
140 ; GFX10-CU-NEXT: v_mov_b32_e32 v1, s3
141 ; GFX10-CU-NEXT: s_waitcnt lgkmcnt(0)
142 ; GFX10-CU-NEXT: flat_store_dword v[0:1], v2
143 ; GFX10-CU-NEXT: s_endpgm
177 ; GFX11-CU-LABEL: flat_nontemporal_load_1:
178 ; GFX11-CU: ; %bb.0: ; %entry
179 ; GFX11-CU-NEXT: s_load_b128 s[0:3], s[0:1], 0x0
180 ; GFX11-CU-NEXT: v_lshlrev_b32_e32 v0, 2, v0
181 ; GFX11-CU-NEXT: s_waitcnt lgkmcnt(0)
182 ; GFX11-CU-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
183 ; GFX11-CU-NEXT: v_add_co_u32 v0, s0, s0, v0
184 ; GFX11-CU-NEXT: v_add_co_ci_u32_e64 v1, null, s1, 0, s0
185 ; GFX11-CU-NEXT: flat_load_b32 v2, v[0:1] glc dlc
186 ; GFX11-CU-NEXT: s_waitcnt vmcnt(0)
187 ; GFX11-CU-NEXT: v_dual_mov_b32 v1, s3 :: v_dual_mov_b32 v0, s2
188 ; GFX11-CU-NEXT: s_waitcnt lgkmcnt(0)
189 ; GFX11-CU-NEXT: flat_store_b32 v[0:1], v2
190 ; GFX11-CU-NEXT: s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
191 ; GFX11-CU-NEXT: s_endpgm
230 ; GFX10-CU-LABEL: flat_nontemporal_store_0:
231 ; GFX10-CU: ; %bb.0: ; %entry
232 ; GFX10-CU-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x0
233 ; GFX10-CU-NEXT: s_waitcnt lgkmcnt(0)
234 ; GFX10-CU-NEXT: v_mov_b32_e32 v0, s0
235 ; GFX10-CU-NEXT: v_mov_b32_e32 v1, s1
236 ; GFX10-CU-NEXT: flat_load_dword v2, v[0:1]
237 ; GFX10-CU-NEXT: v_mov_b32_e32 v0, s2
238 ; GFX10-CU-NEXT: v_mov_b32_e32 v1, s3
239 ; GFX10-CU-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
240 ; GFX10-CU-NEXT: flat_store_dword v[0:1], v2
241 ; GFX10-CU-NEXT: s_waitcnt_vscnt null, 0x0
242 ; GFX10-CU-NEXT: s_endpgm
271 ; GFX11-CU-LABEL: flat_nontemporal_store_0:
272 ; GFX11-CU: ; %bb.0: ; %entry
273 ; GFX11-CU-NEXT: s_load_b128 s[0:3], s[0:1], 0x0
274 ; GFX11-CU-NEXT: s_waitcnt lgkmcnt(0)
275 ; GFX11-CU-NEXT: v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
276 ; GFX11-CU-NEXT: flat_load_b32 v2, v[0:1]
277 ; GFX11-CU-NEXT: v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
278 ; GFX11-CU-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
279 ; GFX11-CU-NEXT: flat_store_b32 v[0:1], v2 dlc
280 ; GFX11-CU-NEXT: s_waitcnt_vscnt null, 0x0
281 ; GFX11-CU-NEXT: s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
282 ; GFX11-CU-NEXT: s_endpgm
322 ; GFX10-CU-LABEL: flat_nontemporal_store_1:
323 ; GFX10-CU: ; %bb.0: ; %entry
324 ; GFX10-CU-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x0
325 ; GFX10-CU-NEXT: v_lshlrev_b32_e32 v0, 2, v0
326 ; GFX10-CU-NEXT: s_waitcnt lgkmcnt(0)
327 ; GFX10-CU-NEXT: v_mov_b32_e32 v1, s0
328 ; GFX10-CU-NEXT: v_mov_b32_e32 v2, s1
329 ; GFX10-CU-NEXT: v_add_co_u32 v0, s0, s2, v0
330 ; GFX10-CU-NEXT: flat_load_dword v2, v[1:2]
331 ; GFX10-CU-NEXT: v_add_co_ci_u32_e64 v1, s0, s3, 0, s0
332 ; GFX10-CU-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
333 ; GFX10-CU-NEXT: flat_store_dword v[0:1], v2
334 ; GFX10-CU-NEXT: s_waitcnt_vscnt null, 0x0
335 ; GFX10-CU-NEXT: s_endpgm
369 ; GFX11-CU-LABEL: flat_nontemporal_store_1:
370 ; GFX11-CU: ; %bb.0: ; %entry
371 ; GFX11-CU-NEXT: s_load_b128 s[0:3], s[0:1], 0x0
372 ; GFX11-CU-NEXT: s_waitcnt lgkmcnt(0)
373 ; GFX11-CU-NEXT: v_dual_mov_b32 v1, s0 :: v_dual_lshlrev_b32 v0, 2, v0
374 ; GFX11-CU-NEXT: v_mov_b32_e32 v2, s1
375 ; GFX11-CU-NEXT: s_delay_alu instid0(VALU_DEP_2)
376 ; GFX11-CU-NEXT: v_add_co_u32 v0, s0, s2, v0
377 ; GFX11-CU-NEXT: flat_load_b32 v2, v[1:2]
378 ; GFX11-CU-NEXT: v_add_co_ci_u32_e64 v1, null, s3, 0, s0
379 ; GFX11-CU-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
380 ; GFX11-CU-NEXT: flat_store_b32 v[0:1], v2 dlc
381 ; GFX11-CU-NEXT: s_waitcnt_vscnt null, 0x0
382 ; GFX11-CU-NEXT: s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
383 ; GFX11-CU-NEXT: s_endpgm
422 ; GFX10-CU-LABEL: flat_volatile_workgroup_acquire_load:
423 ; GFX10-CU: ; %bb.0: ; %entry
424 ; GFX10-CU-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x0
425 ; GFX10-CU-NEXT: s_waitcnt lgkmcnt(0)
426 ; GFX10-CU-NEXT: v_mov_b32_e32 v0, s0
427 ; GFX10-CU-NEXT: v_mov_b32_e32 v1, s1
428 ; GFX10-CU-NEXT: flat_load_dword v2, v[0:1]
429 ; GFX10-CU-NEXT: s_waitcnt lgkmcnt(0)
430 ; GFX10-CU-NEXT: v_mov_b32_e32 v0, s2
431 ; GFX10-CU-NEXT: v_mov_b32_e32 v1, s3
432 ; GFX10-CU-NEXT: s_waitcnt vmcnt(0)
433 ; GFX10-CU-NEXT: flat_store_dword v[0:1], v2
434 ; GFX10-CU-NEXT: s_endpgm
463 ; GFX11-CU-LABEL: flat_volatile_workgroup_acquire_load:
464 ; GFX11-CU: ; %bb.0: ; %entry
465 ; GFX11-CU-NEXT: s_load_b128 s[0:3], s[0:1], 0x0
466 ; GFX11-CU-NEXT: s_waitcnt lgkmcnt(0)
467 ; GFX11-CU-NEXT: v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
468 ; GFX11-CU-NEXT: flat_load_b32 v2, v[0:1]
469 ; GFX11-CU-NEXT: s_waitcnt lgkmcnt(0)
470 ; GFX11-CU-NEXT: v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
471 ; GFX11-CU-NEXT: s_waitcnt vmcnt(0)
472 ; GFX11-CU-NEXT: flat_store_b32 v[0:1], v2
473 ; GFX11-CU-NEXT: s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
474 ; GFX11-CU-NEXT: s_endpgm
509 ; GFX10-CU-LABEL: flat_volatile_workgroup_release_store:
510 ; GFX10-CU: ; %bb.0: ; %entry
511 ; GFX10-CU-NEXT: s_clause 0x1
512 ; GFX10-CU-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x8
513 ; GFX10-CU-NEXT: s_load_dword s2, s[4:5], 0x0
514 ; GFX10-CU-NEXT: s_waitcnt lgkmcnt(0)
515 ; GFX10-CU-NEXT: v_mov_b32_e32 v0, s0
516 ; GFX10-CU-NEXT: v_mov_b32_e32 v1, s1
517 ; GFX10-CU-NEXT: v_mov_b32_e32 v2, s2
518 ; GFX10-CU-NEXT: s_waitcnt lgkmcnt(0)
519 ; GFX10-CU-NEXT: flat_store_dword v[0:1], v2
520 ; GFX10-CU-NEXT: s_endpgm
548 ; GFX11-CU-LABEL: flat_volatile_workgroup_release_store:
549 ; GFX11-CU: ; %bb.0: ; %entry
550 ; GFX11-CU-NEXT: s_clause 0x1
551 ; GFX11-CU-NEXT: s_load_b64 s[2:3], s[0:1], 0x8
552 ; GFX11-CU-NEXT: s_load_b32 s0, s[0:1], 0x0
553 ; GFX11-CU-NEXT: s_waitcnt lgkmcnt(0)
554 ; GFX11-CU-NEXT: v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
555 ; GFX11-CU-NEXT: v_mov_b32_e32 v2, s0
556 ; GFX11-CU-NEXT: s_waitcnt lgkmcnt(0)
557 ; GFX11-CU-NEXT: flat_store_b32 v[0:1], v2
558 ; GFX11-CU-NEXT: s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
559 ; GFX11-CU-NEXT: s_endpgm