# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
# RUN: llc -mtriple=amdgcn -mcpu=gfx908 -verify-machineinstrs -run-pass si-insert-waitcnts -o - %s | FileCheck -check-prefixes=GFX9 %s
# RUN: llc -mtriple=amdgcn -mcpu=gfx1200 -verify-machineinstrs -run-pass si-insert-waitcnts -o - %s | FileCheck -check-prefixes=GFX12 %s
# For gfx12+, this test simply ensures that we don't crash in the face of manually
# inserted waitcnt intrinsics. They are still allowed for compatibility, but
# their effect in the HW is very conservative and code generation does not attempt
# to do anything with them. Developers who write code at such a low level should
# revisit their code for gfx12+ anyway.
---
name: test_waitcnt_preexisting_lgkmcnt_unmodified
body: |
bb.0:
liveins: $vgpr0
; GFX9-LABEL: name: test_waitcnt_preexisting_lgkmcnt_unmodified
; GFX9: liveins: $vgpr0
; GFX9-NEXT: {{ $}}
; GFX9-NEXT: S_WAITCNT 0
; GFX9-NEXT: $vgpr0_vgpr1 = DS_READ2_B32 $vgpr0, 0, 1, 0, implicit $m0, implicit $exec
; GFX9-NEXT: S_WAITCNT 49279
; GFX9-NEXT: $vgpr0 = FLAT_LOAD_DWORD $vgpr0_vgpr1, 0, 0, implicit $exec, implicit $flat_scr
; GFX9-NEXT: S_WAITCNT 112
; GFX9-NEXT: FLAT_STORE_DWORD $vgpr0_vgpr1, $vgpr0, 0, 0, implicit $exec, implicit $flat_scr
; GFX9-NEXT: S_ENDPGM 0
;
; GFX12-LABEL: name: test_waitcnt_preexisting_lgkmcnt_unmodified
; GFX12: liveins: $vgpr0
; GFX12-NEXT: {{ $}}
; GFX12-NEXT: S_WAIT_LOADCNT_DSCNT 0
; GFX12-NEXT: S_WAIT_EXPCNT 0
; GFX12-NEXT: S_WAIT_SAMPLECNT 0
; GFX12-NEXT: S_WAIT_BVHCNT 0
; GFX12-NEXT: S_WAIT_KMCNT 0
; GFX12-NEXT: $vgpr0_vgpr1 = DS_READ2_B32 $vgpr0, 0, 1, 0, implicit $m0, implicit $exec
; GFX12-NEXT: S_WAITCNT 49279
; GFX12-NEXT: S_WAIT_DSCNT 0
; GFX12-NEXT: $vgpr0 = FLAT_LOAD_DWORD $vgpr0_vgpr1, 0, 0, implicit $exec, implicit $flat_scr
; GFX12-NEXT: S_WAIT_LOADCNT_DSCNT 0
; GFX12-NEXT: FLAT_STORE_DWORD $vgpr0_vgpr1, $vgpr0, 0, 0, implicit $exec, implicit $flat_scr
; GFX12-NEXT: S_ENDPGM 0
$vgpr0_vgpr1 = DS_READ2_B32 $vgpr0, 0, 1, 0, implicit $m0, implicit $exec
S_WAITCNT 49279
$vgpr0 = FLAT_LOAD_DWORD $vgpr0_vgpr1, 0, 0, implicit $exec, implicit $flat_scr
FLAT_STORE_DWORD $vgpr0_vgpr1, $vgpr0, 0, 0, implicit $exec, implicit $flat_scr
S_ENDPGM 0
...
---
name: test_waitcnt_preexisting_vmcnt_unmodified
body: |
bb.0:
liveins: $vgpr0_vgpr1
; GFX9-LABEL: name: test_waitcnt_preexisting_vmcnt_unmodified
; GFX9: liveins: $vgpr0_vgpr1
; GFX9-NEXT: {{ $}}
; GFX9-NEXT: S_WAITCNT 0
; GFX9-NEXT: $vgpr0_vgpr1 = GLOBAL_LOAD_DWORDX2 $vgpr0_vgpr1, 0, 0, implicit $exec
; GFX9-NEXT: S_WAITCNT 3952
; GFX9-NEXT: $vgpr0 = FLAT_LOAD_DWORD $vgpr0_vgpr1, 0, 0, implicit $exec, implicit $flat_scr
; GFX9-NEXT: S_WAITCNT 112
; GFX9-NEXT: FLAT_STORE_DWORD $vgpr0_vgpr1, $vgpr0, 0, 0, implicit $exec, implicit $flat_scr
; GFX9-NEXT: S_ENDPGM 0
;
; GFX12-LABEL: name: test_waitcnt_preexisting_vmcnt_unmodified
; GFX12: liveins: $vgpr0_vgpr1
; GFX12-NEXT: {{ $}}
; GFX12-NEXT: S_WAIT_LOADCNT_DSCNT 0
; GFX12-NEXT: S_WAIT_EXPCNT 0
; GFX12-NEXT: S_WAIT_SAMPLECNT 0
; GFX12-NEXT: S_WAIT_BVHCNT 0
; GFX12-NEXT: S_WAIT_KMCNT 0
; GFX12-NEXT: $vgpr0_vgpr1 = GLOBAL_LOAD_DWORDX2 $vgpr0_vgpr1, 0, 0, implicit $exec
; GFX12-NEXT: S_WAITCNT 3952
; GFX12-NEXT: S_WAIT_LOADCNT 0
; GFX12-NEXT: $vgpr0 = FLAT_LOAD_DWORD $vgpr0_vgpr1, 0, 0, implicit $exec, implicit $flat_scr
; GFX12-NEXT: S_WAIT_LOADCNT_DSCNT 0
; GFX12-NEXT: FLAT_STORE_DWORD $vgpr0_vgpr1, $vgpr0, 0, 0, implicit $exec, implicit $flat_scr
; GFX12-NEXT: S_ENDPGM 0
$vgpr0_vgpr1 = GLOBAL_LOAD_DWORDX2 $vgpr0_vgpr1, 0, 0, implicit $exec
S_WAITCNT 3952
$vgpr0 = FLAT_LOAD_DWORD $vgpr0_vgpr1, 0, 0, implicit $exec, implicit $flat_scr
FLAT_STORE_DWORD $vgpr0_vgpr1, $vgpr0, 0, 0, implicit $exec, implicit $flat_scr
S_ENDPGM 0
...
# Respect preexisting waitcnt and add required wait.
---
name: test_waitcnt_preexisting_vmcnt_needs_lgkmcnt
body: |
bb.0:
liveins: $vgpr0
; GFX9-LABEL: name: test_waitcnt_preexisting_vmcnt_needs_lgkmcnt
; GFX9: liveins: $vgpr0
; GFX9-NEXT: {{ $}}
; GFX9-NEXT: S_WAITCNT 0
; GFX9-NEXT: $vgpr0_vgpr1 = DS_READ2_B32 $vgpr0, 0, 1, 0, implicit $m0, implicit $exec
; GFX9-NEXT: S_WAITCNT 112
; GFX9-NEXT: $vgpr0 = FLAT_LOAD_DWORD $vgpr0_vgpr1, 0, 0, implicit $exec, implicit $flat_scr
; GFX9-NEXT: S_WAITCNT 112
; GFX9-NEXT: FLAT_STORE_DWORD $vgpr0_vgpr1, $vgpr0, 0, 0, implicit $exec, implicit $flat_scr
; GFX9-NEXT: S_ENDPGM 0
;
; GFX12-LABEL: name: test_waitcnt_preexisting_vmcnt_needs_lgkmcnt
; GFX12: liveins: $vgpr0
; GFX12-NEXT: {{ $}}
; GFX12-NEXT: S_WAIT_LOADCNT_DSCNT 0
; GFX12-NEXT: S_WAIT_EXPCNT 0
; GFX12-NEXT: S_WAIT_SAMPLECNT 0
; GFX12-NEXT: S_WAIT_BVHCNT 0
; GFX12-NEXT: S_WAIT_KMCNT 0
; GFX12-NEXT: $vgpr0_vgpr1 = DS_READ2_B32 $vgpr0, 0, 1, 0, implicit $m0, implicit $exec
; GFX12-NEXT: S_WAITCNT 3952
; GFX12-NEXT: S_WAIT_DSCNT 0
; GFX12-NEXT: $vgpr0 = FLAT_LOAD_DWORD $vgpr0_vgpr1, 0, 0, implicit $exec, implicit $flat_scr
; GFX12-NEXT: S_WAIT_LOADCNT_DSCNT 0
; GFX12-NEXT: FLAT_STORE_DWORD $vgpr0_vgpr1, $vgpr0, 0, 0, implicit $exec, implicit $flat_scr
; GFX12-NEXT: S_ENDPGM 0
$vgpr0_vgpr1 = DS_READ2_B32 $vgpr0, 0, 1, 0, implicit $m0, implicit $exec
S_WAITCNT 3952
$vgpr0 = FLAT_LOAD_DWORD $vgpr0_vgpr1, 0, 0, implicit $exec, implicit $flat_scr
FLAT_STORE_DWORD $vgpr0_vgpr1, $vgpr0, 0, 0, implicit $exec, implicit $flat_scr
S_ENDPGM 0
...
---
name: test_waitcnt_preexisting_lgkmcnt_needs_vmcnt
body: |
bb.0:
liveins: $vgpr0_vgpr1
; GFX9-LABEL: name: test_waitcnt_preexisting_lgkmcnt_needs_vmcnt
; GFX9: liveins: $vgpr0_vgpr1
; GFX9-NEXT: {{ $}}
; GFX9-NEXT: S_WAITCNT 0
; GFX9-NEXT: $vgpr0_vgpr1 = GLOBAL_LOAD_DWORDX2 $vgpr0_vgpr1, 0, 0, implicit $exec
; GFX9-NEXT: S_WAITCNT 112
; GFX9-NEXT: $vgpr0 = FLAT_LOAD_DWORD $vgpr0_vgpr1, 0, 0, implicit $exec, implicit $flat_scr
; GFX9-NEXT: S_WAITCNT 112
; GFX9-NEXT: FLAT_STORE_DWORD $vgpr0_vgpr1, $vgpr0, 0, 0, implicit $exec, implicit $flat_scr
; GFX9-NEXT: S_ENDPGM 0
;
; GFX12-LABEL: name: test_waitcnt_preexisting_lgkmcnt_needs_vmcnt
; GFX12: liveins: $vgpr0_vgpr1
; GFX12-NEXT: {{ $}}
; GFX12-NEXT: S_WAIT_LOADCNT_DSCNT 0
; GFX12-NEXT: S_WAIT_EXPCNT 0
; GFX12-NEXT: S_WAIT_SAMPLECNT 0
; GFX12-NEXT: S_WAIT_BVHCNT 0
; GFX12-NEXT: S_WAIT_KMCNT 0
; GFX12-NEXT: $vgpr0_vgpr1 = GLOBAL_LOAD_DWORDX2 $vgpr0_vgpr1, 0, 0, implicit $exec
; GFX12-NEXT: S_WAITCNT 49279
; GFX12-NEXT: S_WAIT_LOADCNT 0
; GFX12-NEXT: $vgpr0 = FLAT_LOAD_DWORD $vgpr0_vgpr1, 0, 0, implicit $exec, implicit $flat_scr
; GFX12-NEXT: S_WAIT_LOADCNT_DSCNT 0
; GFX12-NEXT: FLAT_STORE_DWORD $vgpr0_vgpr1, $vgpr0, 0, 0, implicit $exec, implicit $flat_scr
; GFX12-NEXT: S_ENDPGM 0
$vgpr0_vgpr1 = GLOBAL_LOAD_DWORDX2 $vgpr0_vgpr1, 0, 0, implicit $exec
S_WAITCNT 49279
$vgpr0 = FLAT_LOAD_DWORD $vgpr0_vgpr1, 0, 0, implicit $exec, implicit $flat_scr
FLAT_STORE_DWORD $vgpr0_vgpr1, $vgpr0, 0, 0, implicit $exec, implicit $flat_scr
S_ENDPGM 0
...
# Apply wait for all counters from preexisting waitcnt regardless of the wait
# required by the next instruction.
---
name: test_waitcnt_preexisting_apply_all_counters
body: |
bb.0:
liveins: $vgpr0_vgpr1, $vgpr2
; GFX9-LABEL: name: test_waitcnt_preexisting_apply_all_counters
; GFX9: liveins: $vgpr0_vgpr1, $vgpr2
; GFX9-NEXT: {{ $}}
; GFX9-NEXT: S_WAITCNT 0
; GFX9-NEXT: $vgpr4_vgpr5 = GLOBAL_LOAD_DWORDX2 $vgpr0_vgpr1, 0, 0, implicit $exec
; GFX9-NEXT: $vgpr6_vgpr7 = DS_READ2_B32 $vgpr2, 0, 1, 0, implicit $m0, implicit $exec
; GFX9-NEXT: S_WAITCNT 0
; GFX9-NEXT: $vgpr6 = V_OR_B32_e32 1, killed $vgpr6, implicit $exec
; GFX9-NEXT: $vgpr0 = FLAT_LOAD_DWORD $vgpr4_vgpr5, 0, 0, implicit $exec, implicit $flat_scr
; GFX9-NEXT: S_WAITCNT 112
; GFX9-NEXT: FLAT_STORE_DWORD $vgpr0_vgpr1, $vgpr0, 0, 0, implicit $exec, implicit $flat_scr
;
; GFX12-LABEL: name: test_waitcnt_preexisting_apply_all_counters
; GFX12: liveins: $vgpr0_vgpr1, $vgpr2
; GFX12-NEXT: {{ $}}
; GFX12-NEXT: S_WAIT_LOADCNT_DSCNT 0
; GFX12-NEXT: S_WAIT_EXPCNT 0
; GFX12-NEXT: S_WAIT_SAMPLECNT 0
; GFX12-NEXT: S_WAIT_BVHCNT 0
; GFX12-NEXT: S_WAIT_KMCNT 0
; GFX12-NEXT: $vgpr4_vgpr5 = GLOBAL_LOAD_DWORDX2 $vgpr0_vgpr1, 0, 0, implicit $exec
; GFX12-NEXT: $vgpr6_vgpr7 = DS_READ2_B32 $vgpr2, 0, 1, 0, implicit $m0, implicit $exec
; GFX12-NEXT: S_WAITCNT 0
; GFX12-NEXT: S_WAIT_DSCNT 0
; GFX12-NEXT: $vgpr6 = V_OR_B32_e32 1, killed $vgpr6, implicit $exec
; GFX12-NEXT: S_WAIT_LOADCNT 0
; GFX12-NEXT: $vgpr0 = FLAT_LOAD_DWORD $vgpr4_vgpr5, 0, 0, implicit $exec, implicit $flat_scr
; GFX12-NEXT: S_WAIT_LOADCNT_DSCNT 0
; GFX12-NEXT: FLAT_STORE_DWORD $vgpr0_vgpr1, $vgpr0, 0, 0, implicit $exec, implicit $flat_scr
$vgpr4_vgpr5 = GLOBAL_LOAD_DWORDX2 $vgpr0_vgpr1, 0, 0, implicit $exec
$vgpr6_vgpr7 = DS_READ2_B32 $vgpr2, 0, 1, 0, implicit $m0, implicit $exec
S_WAITCNT 0
$vgpr6 = V_OR_B32_e32 1, killed $vgpr6, implicit $exec
$vgpr0 = FLAT_LOAD_DWORD $vgpr4_vgpr5, 0, 0, implicit $exec, implicit $flat_scr
FLAT_STORE_DWORD $vgpr0_vgpr1, $vgpr0, 0, 0, implicit $exec, implicit $flat_scr
...
---
name: test_waitcnt_preexisting_combine_waitcnt
body: |
bb.0:
liveins: $vgpr0_vgpr1
; GFX9-LABEL: name: test_waitcnt_preexisting_combine_waitcnt
; GFX9: liveins: $vgpr0_vgpr1
; GFX9-NEXT: {{ $}}
; GFX9-NEXT: S_WAITCNT 0
; GFX9-NEXT: $vgpr0 = FLAT_LOAD_DWORD $vgpr0_vgpr1, 0, 0, implicit $exec, implicit $flat_scr
; GFX9-NEXT: S_WAITCNT 0
; GFX9-NEXT: FLAT_STORE_DWORD $vgpr0_vgpr1, $vgpr0, 0, 0, implicit $exec, implicit $flat_scr
;
; GFX12-LABEL: name: test_waitcnt_preexisting_combine_waitcnt
; GFX12: liveins: $vgpr0_vgpr1
; GFX12-NEXT: {{ $}}
; GFX12-NEXT: S_WAIT_LOADCNT_DSCNT 0
; GFX12-NEXT: S_WAIT_EXPCNT 0
; GFX12-NEXT: S_WAIT_SAMPLECNT 0
; GFX12-NEXT: S_WAIT_BVHCNT 0
; GFX12-NEXT: S_WAIT_KMCNT 0
; GFX12-NEXT: $vgpr0 = FLAT_LOAD_DWORD $vgpr0_vgpr1, 0, 0, implicit $exec, implicit $flat_scr
; GFX12-NEXT: S_WAITCNT 0
; GFX12-NEXT: S_WAITCNT 0
; GFX12-NEXT: S_WAITCNT 0
; GFX12-NEXT: S_WAITCNT 0
; GFX12-NEXT: S_WAITCNT 0
; GFX12-NEXT: S_WAITCNT 0
; GFX12-NEXT: S_WAIT_LOADCNT_DSCNT 0
; GFX12-NEXT: FLAT_STORE_DWORD $vgpr0_vgpr1, $vgpr0, 0, 0, implicit $exec, implicit $flat_scr
$vgpr0 = FLAT_LOAD_DWORD $vgpr0_vgpr1, 0, 0, implicit $exec, implicit $flat_scr
S_WAITCNT 0
S_WAITCNT 0
S_WAITCNT 0
S_WAITCNT 0
S_WAITCNT 0
S_WAITCNT 0
FLAT_STORE_DWORD $vgpr0_vgpr1, $vgpr0, 0, 0, implicit $exec, implicit $flat_scr
...
---
name: test_waitcnt_preexisting_combine_waitcnt_diff_counters
body: |
bb.0:
liveins: $vgpr0_vgpr1
; GFX9-LABEL: name: test_waitcnt_preexisting_combine_waitcnt_diff_counters
; GFX9: liveins: $vgpr0_vgpr1
; GFX9-NEXT: {{ $}}
; GFX9-NEXT: S_WAITCNT 0
; GFX9-NEXT: $vgpr0 = FLAT_LOAD_DWORD $vgpr0_vgpr1, 0, 0, implicit $exec, implicit $flat_scr
; GFX9-NEXT: S_WAITCNT 112
; GFX9-NEXT: FLAT_STORE_DWORD $vgpr0_vgpr1, $vgpr0, 0, 0, implicit $exec, implicit $flat_scr
;
; GFX12-LABEL: name: test_waitcnt_preexisting_combine_waitcnt_diff_counters
; GFX12: liveins: $vgpr0_vgpr1
; GFX12-NEXT: {{ $}}
; GFX12-NEXT: S_WAIT_LOADCNT_DSCNT 0
; GFX12-NEXT: S_WAIT_EXPCNT 0
; GFX12-NEXT: S_WAIT_SAMPLECNT 0
; GFX12-NEXT: S_WAIT_BVHCNT 0
; GFX12-NEXT: S_WAIT_KMCNT 0
; GFX12-NEXT: $vgpr0 = FLAT_LOAD_DWORD $vgpr0_vgpr1, 0, 0, implicit $exec, implicit $flat_scr
; GFX12-NEXT: S_WAITCNT 49279
; GFX12-NEXT: S_WAITCNT 3952
; GFX12-NEXT: S_WAIT_LOADCNT_DSCNT 0
; GFX12-NEXT: FLAT_STORE_DWORD $vgpr0_vgpr1, $vgpr0, 0, 0, implicit $exec, implicit $flat_scr
$vgpr0 = FLAT_LOAD_DWORD $vgpr0_vgpr1, 0, 0, implicit $exec, implicit $flat_scr
S_WAITCNT 49279
S_WAITCNT 3952
FLAT_STORE_DWORD $vgpr0_vgpr1, $vgpr0, 0, 0, implicit $exec, implicit $flat_scr
...
# Apply preexisting waitcnt when no wait is immediately needed.
# FIXME: Move waitcnt as late as possible.
---
name: test_waitcnt_preexisting_early_wait
body: |
bb.0:
liveins: $vgpr0_vgpr1
; GFX9-LABEL: name: test_waitcnt_preexisting_early_wait
; GFX9: liveins: $vgpr0_vgpr1
; GFX9-NEXT: {{ $}}
; GFX9-NEXT: S_WAITCNT 0
; GFX9-NEXT: $vgpr0 = FLAT_LOAD_DWORD $vgpr0_vgpr1, 0, 0, implicit $exec, implicit $flat_scr
; GFX9-NEXT: S_WAITCNT 0
; GFX9-NEXT: S_NOP 0
; GFX9-NEXT: S_NOP 0
; GFX9-NEXT: S_NOP 0
; GFX9-NEXT: FLAT_STORE_DWORD $vgpr0_vgpr1, $vgpr0, 0, 0, implicit $exec, implicit $flat_scr
; GFX9-NEXT: S_ENDPGM 0
;
; GFX12-LABEL: name: test_waitcnt_preexisting_early_wait
; GFX12: liveins: $vgpr0_vgpr1
; GFX12-NEXT: {{ $}}
; GFX12-NEXT: S_WAIT_LOADCNT_DSCNT 0
; GFX12-NEXT: S_WAIT_EXPCNT 0
; GFX12-NEXT: S_WAIT_SAMPLECNT 0
; GFX12-NEXT: S_WAIT_BVHCNT 0
; GFX12-NEXT: S_WAIT_KMCNT 0
; GFX12-NEXT: $vgpr0 = FLAT_LOAD_DWORD $vgpr0_vgpr1, 0, 0, implicit $exec, implicit $flat_scr
; GFX12-NEXT: S_WAITCNT 0
; GFX12-NEXT: S_NOP 0
; GFX12-NEXT: S_NOP 0
; GFX12-NEXT: S_NOP 0
; GFX12-NEXT: S_WAIT_LOADCNT_DSCNT 0
; GFX12-NEXT: FLAT_STORE_DWORD $vgpr0_vgpr1, $vgpr0, 0, 0, implicit $exec, implicit $flat_scr
; GFX12-NEXT: S_ENDPGM 0
$vgpr0 = FLAT_LOAD_DWORD $vgpr0_vgpr1, 0, 0, implicit $exec, implicit $flat_scr
S_WAITCNT 0
S_NOP 0
S_NOP 0
S_NOP 0
FLAT_STORE_DWORD $vgpr0_vgpr1, $vgpr0, 0, 0, implicit $exec, implicit $flat_scr
S_ENDPGM 0
...
---
name: test_waitcnt_preexisting_ignore_kill
body: |
bb.0:
liveins: $vgpr0_vgpr1
; GFX9-LABEL: name: test_waitcnt_preexisting_ignore_kill
; GFX9: liveins: $vgpr0_vgpr1
; GFX9-NEXT: {{ $}}
; GFX9-NEXT: S_WAITCNT 0
; GFX9-NEXT: $vgpr0 = FLAT_LOAD_DWORD $vgpr0_vgpr1, 0, 0, implicit $exec, implicit $flat_scr
; GFX9-NEXT: S_WAITCNT 3952
; GFX9-NEXT: KILL $vgpr0
;
; GFX12-LABEL: name: test_waitcnt_preexisting_ignore_kill
; GFX12: liveins: $vgpr0_vgpr1
; GFX12-NEXT: {{ $}}
; GFX12-NEXT: S_WAIT_LOADCNT_DSCNT 0
; GFX12-NEXT: S_WAIT_EXPCNT 0
; GFX12-NEXT: S_WAIT_SAMPLECNT 0
; GFX12-NEXT: S_WAIT_BVHCNT 0
; GFX12-NEXT: S_WAIT_KMCNT 0
; GFX12-NEXT: $vgpr0 = FLAT_LOAD_DWORD $vgpr0_vgpr1, 0, 0, implicit $exec, implicit $flat_scr
; GFX12-NEXT: S_WAITCNT 3952
; GFX12-NEXT: KILL $vgpr0
$vgpr0 = FLAT_LOAD_DWORD $vgpr0_vgpr1, 0, 0, implicit $exec, implicit $flat_scr
S_WAITCNT 3952
KILL $vgpr0
...
# Combine preexisting waitcnt with wait added to the start of a non-entry function.
---
name: test_waitcnt_preexisting_func_start
body: |
bb.0:
; GFX9-LABEL: name: test_waitcnt_preexisting_func_start
; GFX9: S_WAITCNT 0
; GFX9-NEXT: S_ENDPGM 0
;
; GFX12-LABEL: name: test_waitcnt_preexisting_func_start
; GFX12: S_WAIT_LOADCNT_DSCNT 0
; GFX12-NEXT: S_WAIT_EXPCNT 0
; GFX12-NEXT: S_WAIT_SAMPLECNT 0
; GFX12-NEXT: S_WAIT_BVHCNT 0
; GFX12-NEXT: S_WAIT_KMCNT 0
; GFX12-NEXT: S_WAITCNT 0
; GFX12-NEXT: S_ENDPGM 0
S_WAITCNT 0
S_ENDPGM 0
...
# Verify that extra waitcnt are not added after buffer invalidate instructions.
---
name: test_waitcnt_preexisting_buffer_inv
body: |
bb.0:
; GFX9-LABEL: name: test_waitcnt_preexisting_buffer_inv
; GFX9: S_WAITCNT 0
; GFX9-NEXT: $vgpr0_vgpr1 = GLOBAL_LOAD_DWORDX2 $vgpr0_vgpr1, 0, 0, implicit $exec
; GFX9-NEXT: S_WAITCNT 3952
; GFX9-NEXT: BUFFER_INVL2 implicit $exec
; GFX9-NEXT: BUFFER_WBINVL1_VOL implicit $exec
; GFX9-NEXT: $vgpr0 = FLAT_LOAD_DWORD $vgpr0_vgpr1, 0, 0, implicit $exec, implicit $flat_scr
; GFX9-NEXT: S_WAITCNT 112
; GFX9-NEXT: FLAT_STORE_DWORD $vgpr0_vgpr1, $vgpr0, 0, 0, implicit $exec, implicit $flat_scr
; GFX9-NEXT: S_ENDPGM 0
;
; GFX12-LABEL: name: test_waitcnt_preexisting_buffer_inv
; GFX12: S_WAIT_LOADCNT_DSCNT 0
; GFX12-NEXT: S_WAIT_EXPCNT 0
; GFX12-NEXT: S_WAIT_SAMPLECNT 0
; GFX12-NEXT: S_WAIT_BVHCNT 0
; GFX12-NEXT: S_WAIT_KMCNT 0
; GFX12-NEXT: $vgpr0_vgpr1 = GLOBAL_LOAD_DWORDX2 $vgpr0_vgpr1, 0, 0, implicit $exec
; GFX12-NEXT: S_WAITCNT 3952
; GFX12-NEXT: BUFFER_INVL2 implicit $exec
; GFX12-NEXT: S_WAIT_LOADCNT 0
; GFX12-NEXT: BUFFER_WBINVL1_VOL implicit $exec
; GFX12-NEXT: $vgpr0 = FLAT_LOAD_DWORD $vgpr0_vgpr1, 0, 0, implicit $exec, implicit $flat_scr
; GFX12-NEXT: S_WAIT_LOADCNT_DSCNT 0
; GFX12-NEXT: FLAT_STORE_DWORD $vgpr0_vgpr1, $vgpr0, 0, 0, implicit $exec, implicit $flat_scr
; GFX12-NEXT: S_ENDPGM 0
$vgpr0_vgpr1 = GLOBAL_LOAD_DWORDX2 $vgpr0_vgpr1, 0, 0, implicit $exec
S_WAITCNT 3952
BUFFER_INVL2 implicit $exec
BUFFER_WBINVL1_VOL implicit $exec
$vgpr0 = FLAT_LOAD_DWORD $vgpr0_vgpr1, 0, 0, implicit $exec, implicit $flat_scr
FLAT_STORE_DWORD $vgpr0_vgpr1, $vgpr0, 0, 0, implicit $exec, implicit $flat_scr
S_ENDPGM 0
...
Codebase Browser
llvm