[llvm] [VectorCombine] Fold reduce(trunc(x)) -> trunc(reduce(x)) iff cost effective (PR #81852)

[Nikita Popov via llvm-commits]

================
@@ -1526,6 +1527,67 @@ bool VectorCombine::foldShuffleFromReductions(Instruction &I) {
   return foldSelectShuffle(*Shuffle, true);
 }
 
+/// Determine if its more efficient to fold:
+///   reduce(trunc(x)) -> trunc(reduce(x)).
+bool VectorCombine::foldTruncFromReductions(Instruction &I) {
+  auto *II = dyn_cast<IntrinsicInst>(&I);
+  if (!II)
+    return false;
+
+  unsigned ReductionOpc = 0;
+  switch (II->getIntrinsicID()) {
+  case Intrinsic::vector_reduce_add:
+    ReductionOpc = Instruction::Add;
+    break;
+  case Intrinsic::vector_reduce_mul:
+    ReductionOpc = Instruction::Mul;
+    break;
+  case Intrinsic::vector_reduce_and:
+    ReductionOpc = Instruction::And;
+    break;
+  case Intrinsic::vector_reduce_or:
+    ReductionOpc = Instruction::Or;
+    break;
+  case Intrinsic::vector_reduce_xor:
+    ReductionOpc = Instruction::Xor;
+    break;
+  default:
+    return false;
+  }
+  Value *ReductionSrc = I.getOperand(0);
+
+  Value *TruncSrc;
+  if (!match(ReductionSrc, m_Trunc(m_OneUse(m_Value(TruncSrc)))))
----------------
nikic wrote:

Why is there a one-use requirement on the trunc *source*, rather than the trunc? Please add a multi-use test.

https://github.com/llvm/llvm-project/pull/81852