[llvm] [VPlan] Process simplifyRecipes via a worklist (PR #133977)

[Luke Lau via llvm-commits]

================
@@ -1075,16 +1003,110 @@ static void simplifyRecipe(VPRecipeBase &R, VPTypeAnalysis &TypeInfo) {
              m_DerivedIV(m_SpecificInt(0), m_SpecificInt(0), m_VPValue()))) &&
       TypeInfo.inferScalarType(R.getOperand(1)) ==
           TypeInfo.inferScalarType(R.getVPSingleValue()))
-    return R.getVPSingleValue()->replaceAllUsesWith(R.getOperand(1));
+    return R.getOperand(1);
+
+  return nullptr;
 }
 
 void VPlanTransforms::simplifyRecipes(VPlan &Plan, Type &CanonicalIVTy) {
   ReversePostOrderTraversal<VPBlockDeepTraversalWrapper<VPBlockBase *>> RPOT(
       Plan.getEntry());
   VPTypeAnalysis TypeInfo(&CanonicalIVTy);
+  SetVector<VPRecipeBase *> Worklist;
+  for (VPBasicBlock *VPBB : VPBlockUtils::blocksOnly<VPBasicBlock>(RPOT))
+    for (VPRecipeBase &R : make_early_inc_range(*VPBB))
+      Worklist.insert(&R);
+
+  while (!Worklist.empty()) {
+    VPRecipeBase *R = Worklist.pop_back_val();
+    if (VPValue *Result = simplifyRecipe(*R, TypeInfo)) {
+      R->getVPSingleValue()->replaceAllUsesWith(Result);
+      R->eraseFromParent();
----------------
lukel97 wrote:

@artagnon erasing the old instruction decreases the number of uses which can affect what transforms kick in, e.g. the blend simplifications are sensitive to the number of uses.

Also if the old recipe uses the new result, it will get added back to the worklist which will loop

https://github.com/llvm/llvm-project/pull/133977