[llvm] [RISCV] Lower VP_SELECT constant false to use vmerge.vxm/vmerge.vim (PR #144461)

[Liao Chunyu via llvm-commits]

================
@@ -13176,6 +13182,23 @@ SDValue RISCVTargetLowering::lowerVPFPIntConvOp(SDValue Op,
   return convertFromScalableVector(VT, Result, DAG, Subtarget);
 }
 
+SDValue
+RISCVTargetLowering::lowerVPSelectConstantFalse(SDValue Op,
+                                                SelectionDAG &DAG) const {
+  SDLoc DL(Op);
+  MVT VT = Op.getSimpleValueType();
+  SDValue TrueVal = Op.getOperand(1);
+  SDValue FalseVal = Op.getOperand(2);
+  SDValue VL = Op.getOperand(3);
+
+  MVT MaskVT = VT.changeVectorElementType(MVT::i1);
+  SDValue AllOneMask = DAG.getNode(RISCVISD::VMSET_VL, DL, MaskVT, VL);
+  SDValue NewMask = DAG.getNode(RISCVISD::VMXOR_VL, DL, MaskVT,
----------------
ChunyuLiao wrote:

By switching to SDNode, we do not  need additional patches, the existing optimization is reused. thanks.

https://github.com/llvm/llvm-project/pull/144461