Skip to content

[AMDGPU] Fix missing S_WAIT_XCNT with multiple pending VMEMs #166779

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
wants to merge 5 commits into
base: main
Choose a base branch
from
Open

[AMDGPU] Fix missing S_WAIT_XCNT with multiple pending VMEMs #166779

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
5806f11
Tweak test case
jayfoad Nov 6, 2025
5b828b0
Simplify and inline applyPendingXcntGroup. NFC.
jayfoad Nov 6, 2025
8042095
Refactor with hasMixedPendingEvents. NFC.
jayfoad Nov 6, 2025
fb6364c
Refactor. NFC.
jayfoad Nov 6, 2025
e3e274f
[AMDGPU] Fix missing S_WAIT_XCNT with multiple pending VMEMs
jayfoad Nov 6, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
18 changes: 6 additions & 12 deletions llvm/lib/Target/AMDGPU/SIInsertWaitcnts.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1291,21 +1291,15 @@ void WaitcntBrackets::applyXcnt(const AMDGPU::Waitcnt &Wait) {
// On entry to a block with multiple predescessors, there may
// be pending SMEM and VMEM events active at the same time.
// In such cases, only clear one active event at a time.
auto applyPendingXcntGroup = [this](unsigned E) {
unsigned LowerBound = getScoreLB(X_CNT);
applyWaitcnt(X_CNT, 0);
PendingEvents |= (1 << E);
setScoreLB(X_CNT, LowerBound);
};

// Wait on XCNT is redundant if we are already waiting for a load to complete.
// SMEM can return out of order, so only omit XCNT wait if we are waiting till
// zero.
if (Wait.KmCnt == 0 && hasPendingEvent(SMEM_GROUP)) {
if (hasPendingEvent(VMEM_GROUP))
applyPendingXcntGroup(VMEM_GROUP);
else
if (!hasMixedPendingEvents(X_CNT))
applyWaitcnt(X_CNT, 0);
else
PendingEvents &= ~(1 << SMEM_GROUP);
return;
}

Expand All @@ -1314,10 +1308,10 @@ void WaitcntBrackets::applyXcnt(const AMDGPU::Waitcnt &Wait) {
// decremented to the same number as LOADCnt.
if (Wait.LoadCnt != ~0u && hasPendingEvent(VMEM_GROUP) &&
!hasPendingEvent(STORE_CNT)) {
if (hasPendingEvent(SMEM_GROUP))
applyPendingXcntGroup(SMEM_GROUP);
else
if (!hasMixedPendingEvents(X_CNT))
applyWaitcnt(X_CNT, std::min(Wait.XCnt, Wait.LoadCnt));
else if (Wait.LoadCnt == 0)
PendingEvents &= ~(1 << VMEM_GROUP);
return;
}

Expand Down
10 changes: 5 additions & 5 deletions llvm/test/CodeGen/AMDGPU/wait-xcnt.mir
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1069,7 +1069,6 @@ body: |
$sgpr0 = S_MOV_B32 $sgpr0
...

# FIXME: Missing S_WAIT_XCNT before overwriting vgpr0.
---
name: mixed_pending_events
tracksRegLiveness: true
Expand All @@ -1088,8 +1087,8 @@ body: |
; GCN-NEXT: successors: %bb.2(0x80000000)
; GCN-NEXT: liveins: $vgpr0_vgpr1, $sgpr2
; GCN-NEXT: {{ $}}
; GCN-NEXT: $vgpr2 = GLOBAL_LOAD_DWORD $vgpr0_vgpr1, 0, 0, implicit $exec
; GCN-NEXT: $vgpr3 = GLOBAL_LOAD_DWORD $vgpr0_vgpr1, 0, 0, implicit $exec
; GCN-NEXT: $vgpr2 = GLOBAL_LOAD_DWORD $vgpr0_vgpr1, 100, 0, implicit $exec
; GCN-NEXT: $vgpr3 = GLOBAL_LOAD_DWORD $vgpr0_vgpr1, 200, 0, implicit $exec
; GCN-NEXT: {{ $}}
; GCN-NEXT: bb.2:
; GCN-NEXT: liveins: $sgpr2, $vgpr2
Expand All @@ -1098,15 +1097,16 @@ body: |
; GCN-NEXT: $vgpr2 = V_MOV_B32_e32 $vgpr2, implicit $exec
; GCN-NEXT: S_WAIT_KMCNT 0
; GCN-NEXT: $sgpr2 = S_MOV_B32 $sgpr2
; GCN-NEXT: S_WAIT_XCNT 0
; GCN-NEXT: $vgpr0 = V_MOV_B32_e32 0, implicit $exec
bb.0:
liveins: $vgpr0_vgpr1, $sgpr0_sgpr1, $scc
$sgpr2 = S_LOAD_DWORD_IMM $sgpr0_sgpr1, 0, 0
S_CBRANCH_SCC1 %bb.2, implicit $scc
bb.1:
liveins: $vgpr0_vgpr1, $sgpr2
$vgpr2 = GLOBAL_LOAD_DWORD $vgpr0_vgpr1, 0, 0, implicit $exec
$vgpr3 = GLOBAL_LOAD_DWORD $vgpr0_vgpr1, 0, 0, implicit $exec
$vgpr2 = GLOBAL_LOAD_DWORD $vgpr0_vgpr1, 100, 0, implicit $exec
$vgpr3 = GLOBAL_LOAD_DWORD $vgpr0_vgpr1, 200, 0, implicit $exec
bb.2:
liveins: $sgpr2, $vgpr2
$vgpr2 = V_MOV_B32_e32 $vgpr2, implicit $exec
Expand Down