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

[Yingwei Zheng via llvm-commits]

================
@@ -3519,6 +3520,97 @@ 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) {
+    if (Mask0[I] >= 0 && Mask1[I] >= 0)
+      NewMask.push_back(Mask0[I] - Mask1[I]);
+    else if (Mask0[I] == -1 && Mask1[I] == -1)
+      continue;
+    else
+      return false;
+  }
+
+  // Ensure all elements of the new mask are equal. If the difference between
+  // the incoming mask elements is the same, the two must be constant offsets
+  // of one another.
+  if (NewMask.empty() ||
+      !std::equal(NewMask.begin() + 1u, NewMask.end(), NewMask.begin()))
+    return false;
+
+  // Create new mask using difference of the two incoming masks.
+  int MaskOffset = NewMask[0u];
+  unsigned Index = (InputNumElements - MaskOffset) % InputNumElements;
+  NewMask.clear();
+
+  for (unsigned I = 0u; I < InputNumElements; ++I) {
+    NewMask.push_back(Index);
+    Index = (Index + 1u) % InputNumElements;
+  }
+
+  // Calculate costs for worst cases and compare.
+  auto const Kind = TTI::SK_PermuteSingleSrc;
+  auto OldCost = std::max(TTI.getShuffleCost(Kind, InputVT, Mask0, CostKind),
+                          TTI.getShuffleCost(Kind, InputVT, Mask1, CostKind));
+  auto NewCost = TTI.getShuffleCost(Kind, InputVT, NewMask, CostKind) +
+                 TTI.getShuffleCost(Kind, InputVT, Mask1, CostKind);
+
+  if (NewCost > OldCost)
+    return false;
+
+  // Create new shuffles and narrowed phi.
+  auto Builder = IRBuilder(&I);
+  Builder.SetInsertPoint(Shuf);
----------------
dtcxzyw wrote:

```suggestion
  auto Builder = IRBuilder(Shuf);
```

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