[llvm] [AMDGPU] Use alias info to relax waitcounts for LDS DMA (PR #74537)
Matt Arsenault via llvm-commits
llvm-commits at lists.llvm.org
Thu Dec 7 03:04:31 PST 2023
================
@@ -703,8 +713,37 @@ void WaitcntBrackets::updateByEvent(const SIInstrInfo *TII,
setRegScore(RegNo, T, CurrScore);
}
}
- if (Inst.mayStore() && (TII->isDS(Inst) || mayWriteLDSThroughDMA(Inst))) {
- setRegScore(SQ_MAX_PGM_VGPRS + EXTRA_VGPR_LDS, T, CurrScore);
+ if (Inst.mayStore() &&
+ (TII->isDS(Inst) || TII->mayWriteLDSThroughDMA(Inst))) {
+ // MUBUF and FLAT LDS DMA operations need a wait on vmcnt before LDS
+ // written can be accessed. A load from LDS to VMEM does not need a wait.
+ unsigned Slot = 0;
+ for (const auto *MemOp : Inst.memoperands()) {
+ if (!MemOp->isStore() ||
+ MemOp->getAddrSpace() != AMDGPUAS::LOCAL_ADDRESS)
+ continue;
+ // Comparing just AA info does not guarantee memoperands are equal
+ // in general, but this is so for LDS DMA on practice.
+ auto AAI = MemOp->getAAInfo();
+ if (!AAI)
+ break;
+ for (unsigned I = 0, E = LDSDMAStores.size(); I != E && !Slot; ++I) {
+ for (const auto *MemOp : LDSDMAStores[I]->memoperands()) {
+ if (MemOp->isStore() && AAI == MemOp->getAAInfo()) {
----------------
arsenm wrote:
I'm not sure I understand the equality check here; you perform an actual mayAlias query later? What is the point of this filter? Don't you need to consider any possibly aliasing write as an event?
https://github.com/llvm/llvm-project/pull/74537
More information about the llvm-commits
mailing list