[flang-commits] [clang-tools-extra] [flang] [libc] [lldb] [compiler-rt] [lld] [llvm] [libcxx] [clang] [AMDGPU] Use alias info to relax waitcounts for LDS DMA (PR #74537)

[Stanislav Mekhanoshin via flang-commits]

================
@@ -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
----------------
rampitec wrote:

Right, there is no PSV. I have mentioned PSV because you have earlier suggested to use it. For the real IR value: it is not helpful to compare it. The IR value is a GEP, and this GEP is always different. I.e. these values never compare equal. The rest of the IR is already gone and unavailable for the analysis. Even if it would be available this GEP will address kernel module LDS variable, a single huge LDS array, and will be useless again. In this case it will tell you any LDS operation aliases any other. Now during the module LDS lowering I am creating alias scope info specifically to disambiguate aliasing after the pass has squashed all LDS variables.

https://github.com/llvm/llvm-project/pull/74537