[llvm] [VectorCombine] foldShuffleOfBinops - fold shuffle(binop(shuffle(x),shuffle(z)),binop(shuffle(y),shuffle(w)) -> binop(shuffle(x,z),shuffle(y,w)) (PR #120984)

[Simon Pilgrim via llvm-commits]

================
@@ -1723,6 +1730,36 @@ bool VectorCombine::foldShuffleOfBinops(Instruction &I) {
       TTI.getShuffleCost(TargetTransformInfo::SK_PermuteTwoSrc, BinResTy,
                          OldMask, CostKind, 0, nullptr, {LHS, RHS}, &I);
 
+  // Handle shuffle(binop(shuffle(x),y),binop(z,shuffle(w))) style patterns
+  // where one use shuffles have gotten split across the binop/cmp. These
+  // often allow a major reduction in total cost that wouldn't happen as
+  // individual folds.
+  auto MergeInner = [&](Value *&Op, int Offset, MutableArrayRef<int> Mask,
+                        TTI::TargetCostKind CostKind) -> bool {
+    Value *InnerOp;
+    ArrayRef<int> InnerMask;
+    if (match(Op, m_OneUse(m_Shuffle(m_Value(InnerOp), m_Undef(),
+                                     m_Mask(InnerMask)))) &&
+        all_of(InnerMask,
+               [NumSrcElts](int M) { return M < (int)NumSrcElts; }) &&
+        InnerOp->getType() == Op->getType()) {
+      for (int &M : Mask)
+        if (Offset <= M && M < (int)(Offset + NumSrcElts)) {
+          M = InnerMask[M - Offset];
+          M = 0 <= M ? M + Offset : M;
----------------
RKSimon wrote:

Yes, we've replaced the original M with the InnerMask value to fold the shuffles together - and InnerMask could contain PoisonMaskElem

https://github.com/llvm/llvm-project/pull/120984