[llvm] AMDGPU: Delay value replacement in PromoteAlloca (PR #186944)

[Juan Manuel Martinez Caamaño via llvm-commits]

================
@@ -420,14 +460,23 @@ bool AMDGPUPromoteAllocaImpl::run(Function &F, bool PromoteToLDS) {
 
   bool Changed = false;
   SetVector<IntrinsicInst *> DeferredIntrs;
+  ValueReplacer VR;
+  SmallVector<Instruction *> ToBeErased;
+
   for (AllocaAnalysis &AA : Allocas) {
     if (AA.Vector.Ty) {
       std::optional<TypeSize> Size = AA.Alloca->getAllocationSize(*DL);
       assert(Size); // Expected to succeed on non-array alloca.
       const unsigned AllocaCost = Size->getFixedValue() * 8;
       // First, check if we have enough budget to vectorize this alloca.
       if (AllocaCost <= VectorizationBudget) {
-        promoteAllocaToVector(AA);
+        promoteAllocaToVector(AA, VR);
+
+        VR.postErase(AA.Vector.Worklist);
+        // Append in reverse order so that further users would be erased first.
+        VR.postErase(reverse(AA.Vector.UsersToRemove));
----------------
jmmartinez wrote:

Why do we have to call `reverse` in here? We're already calling `dropDropableUsers`, so we could erase the elements in any order we want since the users should be empty.

---

If we find out that we can remove `reverse`, we would be able to have a single `postErase` method taking an `ArrayRef` (a single element can convert to `ArrayRef` if I recall correctly).

https://github.com/llvm/llvm-project/pull/186944