[AMDGPU][NFC] Use getScoreUB for XCNT insertion.

[easyonaadit]

No description provided.

[llvmbot]

@llvm/pr-subscribers-backend-amdgpu

Author: Aaditya (easyonaadit)

Changes

---

Full diff: https://github.com/llvm/llvm-project/pull/162448.diff

1 Files Affected:

• (modified) llvm/lib/Target/AMDGPU/SIInsertWaitcnts.cpp (+1-1)

diff --git a/llvm/lib/Target/AMDGPU/SIInsertWaitcnts.cpp b/llvm/lib/Target/AMDGPU/SIInsertWaitcnts.cpp
index 5e27b37809c7c..6dcbced010a5a 100644
--- a/llvm/lib/Target/AMDGPU/SIInsertWaitcnts.cpp
+++ b/llvm/lib/Target/AMDGPU/SIInsertWaitcnts.cpp
@@ -1019,7 +1019,7 @@ void WaitcntBrackets::updateByEvent(WaitEventType E, MachineInstr &Inst) {
       // SMEM and VMEM operations. So there will never be
       // outstanding address translations for both SMEM and
       // VMEM at the same time.
-      setScoreLB(T, CurrScore - 1);
+      setScoreLB(T, getScoreUB(T) - 1);
       PendingEvents &= ~(1 << OtherEvent);
     }
     for (const MachineOperand &Op : Inst.all_uses())

[shiltian]

I'm not sure if we need a test here?

[cdevadas]

I'm not sure if we need a test here?

This is a better way of achieving the code change that was introduced in #160812.

[jayfoad]

Why do you still subtract 1? Is that required for some reason?

[easyonaadit]

I want to keep exactly one pending Xcnt event.
UBScore gets updated at the start of this function, if I use that as-is, it'll mark all xcnt events as completed.

[jayfoad]

I see. I think this is a hint that you're doing this too late, and it might work out cleaner to implement it at the end of generateWaitcntInstBefore near where we handle S_BARRIER.

[jayfoad]

• @Pierre-vh FYI.

[jayfoad]

Current patch seems fine anyway, since it's a tiny improvement on the status quo.