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

[Alexey Bataev 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()) {
----------------
alexey-bataev wrote:

```suggestion
    if (match(Op, m_OneUse(m_Shuffle(m_Value(InnerOp), m_Undef(),
                                     m_Mask(InnerMask)))) &&
        InnerOp->getType() == Op->getType() &&
        all_of(InnerMask,
               [NumSrcElts](int M) { return M < (int)NumSrcElts; })) {
```

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