[llvm] [VectorCombine] New folding pattern for extract/binop/shuffle chains (PR #145232)

[Yingwei Zheng via llvm-commits]

================
@@ -2910,6 +2911,133 @@ bool VectorCombine::foldShuffleFromReductions(Instruction &I) {
   return foldSelectShuffle(*Shuffle, true);
 }
 
+bool VectorCombine::foldShuffleChainsToReduce(Instruction &I) {
+  auto *SVI = dyn_cast<ShuffleVectorInst>(&I);
----------------
dtcxzyw wrote:

This transform should start with the root instruction `ExtractElementInst` instead of the first instruction `ShuffleVectorInst`, then visit binop/shufflevector instructions through the def-use chain. `CI->getNextNode()` is expensive and unreliable. You can imagine that there are many unrelated instructions inserted into the middle of this pattern...



https://github.com/llvm/llvm-project/pull/145232