[llvm] [LV][VPlan] Add initial support for CSA vectorization (PR #106560)

[Michael Maitland via llvm-commits]

================
@@ -2932,6 +2938,25 @@ LoopVectorizationCostModel::getVectorIntrinsicCost(CallInst *CI,
                                    TargetTransformInfo::TCK_RecipThroughput);
 }
 
+void InnerLoopVectorizer::fixCSALiveOuts(VPTransformState &State, VPlan &Plan) {
+  for (const auto &CSA : Plan.getCSAStates()) {
+    VPCSADataUpdateRecipe *VPDataUpdate = CSA.second->getDataUpdate();
+    assert(VPDataUpdate &&
+           "VPDataUpdate must have been introduced prior to fixing live outs");
+    Value *V = VPDataUpdate->getUnderlyingValue();
+    Value *ExtractedScalar = State.get(CSA.second->getExtractScalarRecipe(), 0,
+                                       /*NeedsScalar=*/true);
+    // Fix LCSSAPhis
+    llvm::SmallPtrSet<PHINode *, 2> ToFix;
+    for (User *U : V->users())
+      if (auto *Phi = dyn_cast<PHINode>(U);
+          Phi && Phi->getParent() == LoopExitBlock)
+        ToFix.insert(Phi);
+    for (PHINode *Phi : ToFix)
----------------
michaelmaitland wrote:

It causes a crash if we do it in place, likely due to modifying `V->users()` while iterating over it. I'll take a closer look on Monday.

https://github.com/llvm/llvm-project/pull/106560