[llvm] [RISCV] Canonicalize the true operand of vselect to fold more vmerge.vvm instructions (PR #206449)

[Sam Elliott via llvm-commits]

================
@@ -20430,16 +20461,24 @@ static bool matchSelectAddSub(SDValue TrueVal, SDValue FalseVal, bool &SwapCC) {
           (TrueVal.getOperand(1) == A && TrueVal.getOperand(0) == B));
 }
 
-/// Convert vselect CC, (add a, b), (sub a, b) to add a, (vselect CC, -b, b).
-/// This allows us match a vadd.vv fed by a masked vrsub, which reduces
-/// register pressure over the add followed by masked vsub sequence.
-static SDValue performVSELECTCombine(SDNode *N, SelectionDAG &DAG) {
+static SDValue performVSELECTCombine(SDNode *N, SelectionDAG &DAG,
+                                     const RISCVSubtarget &Subtarget) {
   SDLoc DL(N);
   EVT VT = N->getValueType(0);
   SDValue CC = N->getOperand(0);
   SDValue TrueVal = N->getOperand(1);
   SDValue FalseVal = N->getOperand(2);
 
+  // Convert (vselect CC, true, false) to (vselect InvertCC, false, true when)
----------------
lenary wrote:

```suggestion
  // Convert (vselect CC, true, false) to (vselect InvertCC, false, true) when
```

https://github.com/llvm/llvm-project/pull/206449