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

[Rajveer Singh Bharadwaj via llvm-commits]

================
@@ -2988,6 +2989,241 @@ bool VectorCombine::foldShuffleFromReductions(Instruction &I) {
   return foldSelectShuffle(*Shuffle, true);
 }
 
+bool VectorCombine::foldShuffleChainsToReduce(Instruction &I) {
+  auto *EEI = dyn_cast<ExtractElementInst>(&I);
+  if (!EEI)
+    return false;
+
+  std::queue<Value *> InstWorklist;
+  InstructionCost OrigCost = 0;
+
+  Value *InitEEV = nullptr;
+
+  unsigned int CommonCallOp = 0;
+  Instruction::BinaryOps CommonBinOp = Instruction::BinaryOpsEnd;
+
+  bool IsFirstCallOrBinInst = true;
+  bool ShouldBeCallOrBinInst = true;
+
+  SmallVector<Value *, 3> PrevVecV(3, nullptr);
----------------
Rajveer100 wrote:

Yeah, documentation was the last thing left to add for this fold pattern. Since you asked, `PrevVecV` is used to store the last two instructions which are used either singly for shuffle vector or for binary operations. Specifically, `0` and `1` would be the ones that keep changing as we go up, whereas `2` is the one we get from extract operation.

These help differentiate between them in case they are mixed up.

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