75 lines · plain
1# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py2# RUN: llc -mtriple=amdgcn -mcpu=gfx1200 -verify-machineinstrs -run-pass si-insert-hard-clauses %s -o - | FileCheck %s -check-prefixes=GCN,GFX12003# RUN: llc -mtriple=amdgcn -mcpu=gfx1250 -run-pass si-insert-hard-clauses %s -o - | FileCheck %s -check-prefixes=GCN,GFX12504 5---6name: non_flat_vmem_loads7tracksRegLiveness: true8body: |9 bb.0:10 liveins: $vgpr0, $vgpr1, $vgpr211 ; GFX1200-LABEL: name: non_flat_vmem_loads12 ; GFX1200: liveins: $vgpr0, $vgpr1, $vgpr213 ; 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 216 ; GFX1200-NEXT: $vgpr6_vgpr7 = GLOBAL_LOAD_TR_B64_w32 $vgpr0_vgpr1, 8, 0, implicit $exec17 ; GFX1200-NEXT: $vgpr8_vgpr9 = CLUSTER_LOAD_B64 $vgpr0_vgpr1, 16, 0, implicit $m0, implicit $exec18 ; GFX1200-NEXT: $vgpr10_vgpr11 = GLOBAL_LOAD_MONITOR_B64 $vgpr0_vgpr1, 24, 0, implicit $exec19 ; GFX1200-NEXT: }20 ; GFX1200-NEXT: GLOBAL_LOAD_ASYNC_TO_LDS_B32 $vgpr2, $vgpr0_vgpr1, 0, 0, implicit-def $asynccnt, implicit $exec, implicit $asynccnt21 ; GFX1200-NEXT: $vgpr4 = GLOBAL_LOAD_DWORD $vgpr0_vgpr1, 32, 0, implicit $exec22 ;23 ; GFX1250-LABEL: name: non_flat_vmem_loads24 ; GFX1250: liveins: $vgpr0, $vgpr1, $vgpr225 ; 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 428 ; GFX1250-NEXT: $vgpr6_vgpr7 = GLOBAL_LOAD_TR_B64_w32 $vgpr0_vgpr1, 8, 0, implicit $exec29 ; GFX1250-NEXT: $vgpr8_vgpr9 = CLUSTER_LOAD_B64 $vgpr0_vgpr1, 16, 0, implicit $m0, implicit $exec30 ; GFX1250-NEXT: $vgpr10_vgpr11 = GLOBAL_LOAD_MONITOR_B64 $vgpr0_vgpr1, 24, 0, implicit $exec31 ; GFX1250-NEXT: GLOBAL_LOAD_ASYNC_TO_LDS_B32 $vgpr2, $vgpr0_vgpr1, 0, 0, implicit-def $asynccnt, implicit $exec, implicit $asynccnt32 ; GFX1250-NEXT: $vgpr4 = GLOBAL_LOAD_DWORD $vgpr0_vgpr1, 32, 0, implicit $exec33 ; GFX1250-NEXT: }34 $vgpr6_vgpr7 = GLOBAL_LOAD_TR_B64_w32 $vgpr0_vgpr1, 8, 0, implicit $exec35 $vgpr8_vgpr9 = CLUSTER_LOAD_B64 $vgpr0_vgpr1, 16, 0, implicit $m0, implicit $exec36 $vgpr10_vgpr11 = GLOBAL_LOAD_MONITOR_B64 $vgpr0_vgpr1, 24, 0, implicit $exec37 GLOBAL_LOAD_ASYNC_TO_LDS_B32 $vgpr2, $vgpr0_vgpr1, 0, 0, implicit-def $asynccnt, implicit $exec, implicit $asynccnt38 $vgpr4 = GLOBAL_LOAD_DWORD $vgpr0_vgpr1, 32, 0, implicit $exec39...40 41---42name: async_load_flat_monitor_load43tracksRegLiveness: true44body: |45 bb.0:46 liveins: $vgpr0, $vgpr1, $vgpr247 ; GCN-LABEL: name: async_load_flat_monitor_load48 ; GCN: liveins: $vgpr0, $vgpr1, $vgpr249 ; GCN-NEXT: {{ $}}50 ; GCN-NEXT: GLOBAL_LOAD_ASYNC_TO_LDS_B32 $vgpr2, $vgpr0_vgpr1, 0, 0, implicit-def $asynccnt, implicit $exec, implicit $asynccnt51 ; GCN-NEXT: $vgpr0 = FLAT_LOAD_MONITOR_B32 $vgpr0_vgpr1, 8, 0, implicit $exec, implicit $flat_scr52 GLOBAL_LOAD_ASYNC_TO_LDS_B32 $vgpr2, $vgpr0_vgpr1, 0, 0, implicit-def $asynccnt, implicit $exec, implicit $asynccnt53 $vgpr0 = FLAT_LOAD_MONITOR_B32 $vgpr0_vgpr1, 8, 0, implicit $exec, implicit $flat_scr54...55 56---57name: flat_monitor_loads_flat_load58tracksRegLiveness: true59body: |60 bb.0:61 liveins: $vgpr0, $vgpr162 ; GCN-LABEL: name: flat_monitor_loads_flat_load63 ; GCN: liveins: $vgpr0, $vgpr164 ; GCN-NEXT: {{ $}}65 ; GCN-NEXT: BUNDLE implicit-def $vgpr4, implicit-def $vgpr3, implicit-def $vgpr0, implicit $vgpr0_vgpr1, implicit $exec, implicit $flat_scr {66 ; GCN-NEXT: S_CLAUSE 267 ; GCN-NEXT: $vgpr4 = FLAT_LOAD_MONITOR_B32 $vgpr0_vgpr1, 0, 0, implicit $exec, implicit $flat_scr68 ; GCN-NEXT: $vgpr3 = FLAT_LOAD_MONITOR_B32 $vgpr0_vgpr1, 8, 0, implicit $exec, implicit $flat_scr69 ; GCN-NEXT: $vgpr0 = FLAT_LOAD_DWORD $vgpr0_vgpr1, 0, 0, implicit $exec, implicit $flat_scr70 ; GCN-NEXT: }71 $vgpr4 = FLAT_LOAD_MONITOR_B32 $vgpr0_vgpr1, 0, 0, implicit $exec, implicit $flat_scr72 $vgpr3 = FLAT_LOAD_MONITOR_B32 $vgpr0_vgpr1, 8, 0, implicit $exec, implicit $flat_scr73 $vgpr0 = FLAT_LOAD_DWORD $vgpr0_vgpr1, 0, 0, implicit $exec, implicit $flat_scr74...75