|
1 | 1 | # NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
|
2 |
| -# RUN: llc -mtriple=amdgcn -mcpu=gfx1200 -verify-machineinstrs -run-pass si-insert-hard-clauses %s -o - | FileCheck %s -check-prefixes=GCN |
3 |
| -# RUN: llc -mtriple=amdgcn -mcpu=gfx1250 -run-pass si-insert-hard-clauses %s -o - | FileCheck %s -check-prefixes=GCN |
| 2 | +# RUN: llc -mtriple=amdgcn -mcpu=gfx1200 -verify-machineinstrs -run-pass si-insert-hard-clauses %s -o - | FileCheck %s -check-prefixes=GCN,GFX1200 |
| 3 | +# RUN: llc -mtriple=amdgcn -mcpu=gfx1250 -run-pass si-insert-hard-clauses %s -o - | FileCheck %s -check-prefixes=GCN,GFX1250 |
| 4 | + |
| 5 | +--- |
| 6 | +name: non_flat_vmem_loads |
| 7 | +tracksRegLiveness: true |
| 8 | +body: | |
| 9 | + bb.0: |
| 10 | + liveins: $vgpr0, $vgpr1, $vgpr2 |
| 11 | + ; GFX1200-LABEL: name: non_flat_vmem_loads |
| 12 | + ; GFX1200: liveins: $vgpr0, $vgpr1, $vgpr2 |
| 13 | + ; GFX1200-NEXT: {{ $}} |
| 14 | + ; GFX1200-NEXT: BUNDLE implicit-def $vgpr6_vgpr7, implicit-def $vgpr8_vgpr9, implicit-def $vgpr10_vgpr11, implicit $vgpr0_vgpr1, implicit $exec, implicit $m0 { |
| 15 | + ; GFX1200-NEXT: S_CLAUSE 2 |
| 16 | + ; GFX1200-NEXT: $vgpr6_vgpr7 = GLOBAL_LOAD_TR_B64_w32 $vgpr0_vgpr1, 8, 0, implicit $exec |
| 17 | + ; GFX1200-NEXT: $vgpr8_vgpr9 = CLUSTER_LOAD_B64 $vgpr0_vgpr1, 16, 0, implicit $m0, implicit $exec |
| 18 | + ; GFX1200-NEXT: $vgpr10_vgpr11 = GLOBAL_LOAD_MONITOR_B64 $vgpr0_vgpr1, 24, 0, implicit $exec |
| 19 | + ; GFX1200-NEXT: } |
| 20 | + ; GFX1200-NEXT: GLOBAL_LOAD_ASYNC_TO_LDS_B32 $vgpr2, $vgpr0_vgpr1, 0, 0, implicit-def $asynccnt, implicit $exec, implicit $asynccnt |
| 21 | + ; GFX1200-NEXT: $vgpr4 = GLOBAL_LOAD_DWORD $vgpr0_vgpr1, 32, 0, implicit $exec |
| 22 | + ; |
| 23 | + ; GFX1250-LABEL: name: non_flat_vmem_loads |
| 24 | + ; GFX1250: liveins: $vgpr0, $vgpr1, $vgpr2 |
| 25 | + ; GFX1250-NEXT: {{ $}} |
| 26 | + ; GFX1250-NEXT: BUNDLE implicit-def $vgpr6_vgpr7, implicit-def $vgpr8_vgpr9, implicit-def $vgpr10_vgpr11, implicit-def $asynccnt, implicit-def $vgpr4, implicit $vgpr0_vgpr1, implicit $exec, implicit $m0, implicit $vgpr2, implicit $asynccnt { |
| 27 | + ; GFX1250-NEXT: S_CLAUSE 4 |
| 28 | + ; GFX1250-NEXT: $vgpr6_vgpr7 = GLOBAL_LOAD_TR_B64_w32 $vgpr0_vgpr1, 8, 0, implicit $exec |
| 29 | + ; GFX1250-NEXT: $vgpr8_vgpr9 = CLUSTER_LOAD_B64 $vgpr0_vgpr1, 16, 0, implicit $m0, implicit $exec |
| 30 | + ; GFX1250-NEXT: $vgpr10_vgpr11 = GLOBAL_LOAD_MONITOR_B64 $vgpr0_vgpr1, 24, 0, implicit $exec |
| 31 | + ; GFX1250-NEXT: GLOBAL_LOAD_ASYNC_TO_LDS_B32 $vgpr2, $vgpr0_vgpr1, 0, 0, implicit-def $asynccnt, implicit $exec, implicit $asynccnt |
| 32 | + ; GFX1250-NEXT: $vgpr4 = GLOBAL_LOAD_DWORD $vgpr0_vgpr1, 32, 0, implicit $exec |
| 33 | + ; GFX1250-NEXT: } |
| 34 | + $vgpr6_vgpr7 = GLOBAL_LOAD_TR_B64_w32 $vgpr0_vgpr1, 8, 0, implicit $exec |
| 35 | + $vgpr8_vgpr9 = CLUSTER_LOAD_B64 $vgpr0_vgpr1, 16, 0, implicit $m0, implicit $exec |
| 36 | + $vgpr10_vgpr11 = GLOBAL_LOAD_MONITOR_B64 $vgpr0_vgpr1, 24, 0, implicit $exec |
| 37 | + GLOBAL_LOAD_ASYNC_TO_LDS_B32 $vgpr2, $vgpr0_vgpr1, 0, 0, implicit-def $asynccnt, implicit $exec, implicit $asynccnt |
| 38 | + $vgpr4 = GLOBAL_LOAD_DWORD $vgpr0_vgpr1, 32, 0, implicit $exec |
| 39 | +... |
4 | 40 |
|
5 | 41 | ---
|
6 | 42 | name: async_load_flat_monitor_load
|
|
0 commit comments