[llvm] [VectorCombine][AMDGPU] Narrow Phi of Shuffles. (PR #140188)

[Simon Pilgrim via llvm-commits]

================
@@ -3519,6 +3520,95 @@ bool VectorCombine::foldInterleaveIntrinsics(Instruction &I) {
   return true;
 }
 
+/// Attempt to narrow a phi of shufflevector instructions where the two incoming
+/// values have the same operands but different masks. If the two shuffle masks
+/// are offsets of one another we can use one branch to rotate the incoming
+/// vector and perform one larger shuffle after the phi.
+bool VectorCombine::shrinkPhiOfShuffles(Instruction &I) {
+  auto *Phi = dyn_cast<PHINode>(&I);
+  if (!Phi || Phi->getNumIncomingValues() != 2u)
+    return false;
+
+  Value *Op = nullptr;
+  ArrayRef<int> Mask0;
+  ArrayRef<int> Mask1;
+
+  if (!match(Phi->getOperand(0u),
+             m_OneUse(m_Shuffle(m_Value(Op), m_Poison(), m_Mask(Mask0)))) ||
+      !match(Phi->getOperand(1u),
+             m_OneUse(m_Shuffle(m_Specific(Op), m_Poison(), m_Mask(Mask1)))))
+    return false;
+
+  auto *Shuf = cast<ShuffleVectorInst>(Phi->getOperand(0u));
+
+  // Ensure result vectors are wider than the argument vector.
+  auto *InputVT = cast<FixedVectorType>(Op->getType());
+  auto *ResultVT = cast<FixedVectorType>(Shuf->getType());
+  auto const InputNumElements = InputVT->getNumElements();
+
+  if (InputNumElements >= ResultVT->getNumElements())
+    return false;
+
+  // Take the difference of the two shuffle masks at each index. Ignore poison
+  // values at the same index in both masks.
+  SmallVector<int, 16> NewMask;
+  NewMask.reserve(Mask0.size());
+
+  for (auto I = 0u; I < Mask0.size(); ++I) {
----------------
RKSimon wrote:

```
for (auto [M0,M1] : zip(Mask0, Mask1)) {
  if (M0 >= 0 && M1 >= 0)
    NewMask.push_back(M0 - M1);
  else if (M0 == -1 && M1 == -1)
    continue;
  else
    return false;
}
```

https://github.com/llvm/llvm-project/pull/140188