[llvm] [SelectionDAG] Fold constant min/max vector reductions (PR #209190)

[via llvm-commits]

================
@@ -7802,19 +7806,26 @@ SDValue SelectionDAG::FoldConstantArithmetic(unsigned Opcode, const SDLoc &DL,
     if (Opcode == ISD::BITCAST)
       return SDValue();
 
-    // Constant fold VECREDUCE_ADD with a BUILD_VECTOR of integer constants.
-    if (Opcode == ISD::VECREDUCE_ADD && ISD::isBuildVectorOfConstantSDNodes(N1.getNode())) {
+    // Constant fold integer vector reductions with constant BUILD_VECTORs.
+    if ((Opcode == ISD::VECREDUCE_ADD || Opcode == ISD::VECREDUCE_SMAX ||
+         Opcode == ISD::VECREDUCE_SMIN || Opcode == ISD::VECREDUCE_UMAX ||
+         Opcode == ISD::VECREDUCE_UMIN) &&
+        ISD::isBuildVectorOfConstantSDNodes(N1.getNode())) {
       unsigned EltBits = N1.getValueType().getScalarSizeInBits();
-      APInt Acc = APInt::getZero(EltBits);
+      std::optional<APInt> Acc;
----------------
mygitljf wrote:

Thanks. I removed the optional accumulator and now initialize the APInt accumulator with the identity value for BaseOpcode.
I also pulled the integer identity construction into a local getIntegerIdentity helper and reused it in both FoldConstantArithmetic and getIdentityElement. This avoids creating an SDValue just to extract its APInt, while keeping the existing floating-point handling and type-width behavior unchanged. 

https://github.com/llvm/llvm-project/pull/209190