[llvm] [RISCV] Allow folding vmerge into masked ops when mask is the same (PR #97989)

[Luke Lau via llvm-commits]

================
@@ -3755,12 +3766,16 @@ bool RISCVDAGToDAGISel::performCombineVMergeAndVOps(SDNode *N) {
       return false;
   }
 
-  // If True is masked then the vmerge must have an all 1s mask, since we're
-  // going to keep the mask from True.
+  // If True is masked then the vmerge must have either the same mask or an all
+  // 1s mask, since we're going to keep the mask from True.
   if (IsMasked && Mask) {
     // FIXME: Support mask agnostic True instruction which would have an
     // undef merge operand.
-    if (!usesAllOnesMask(Mask, Glue))
+    SDValue TrueMask =
+        getMaskSetter(True->getOperand(Info->MaskOpIdx),
+                      True->getOperand(True->getNumOperands() - 1));
----------------
lukel97 wrote:

This is a bit gnarly but I still hope to eventually move all this into RISCVFoldMasks.cpp. Sorry for dropping the ball on that

https://github.com/llvm/llvm-project/pull/97989