[llvm] [RISCV] Don't lose elements from False in vmerge -> vmv peephole (PR #149720)

[Luke Lau via llvm-commits]

================
@@ -78,12 +78,12 @@ body: |
     ; CHECK-NEXT: %false:vrnov0 = COPY $v9
     ; CHECK-NEXT: %mask:vmv0 = COPY $v0
     ; CHECK-NEXT: %true:vrnov0 = PseudoVADD_VV_M1_MASK %false, $noreg, $noreg, %mask, 4, 5 /* e32 */, 0 /* tu, mu */
-    ; CHECK-NEXT: %x:vr = PseudoVMV_V_V_M1 %pt, %true, 8, 5 /* e32 */, 0 /* tu, mu */
+    ; CHECK-NEXT: %x:vr = PseudoVMV_V_V_M1 %pt, %true, 4, 5 /* e32 */, 0 /* tu, mu */
     %pt:vrnov0 = COPY $v8
     %false:vrnov0 = COPY $v9
     %mask:vmv0 = COPY $v0
-    %true:vrnov0 = PseudoVADD_VV_M1_MASK %false, $noreg, $noreg, %mask, 4, 5 /* e32 */, 0 /* tu, mu */
-    %x:vrnov0 = PseudoVMERGE_VVM_M1 %pt, %false, %true, %mask, 8, 5 /* e32 */
+    %true:vrnov0 = PseudoVADD_VV_M1_MASK %false, $noreg, $noreg, %mask, 8, 5 /* e32 */, 0 /* tu, mu */
+    %x:vrnov0 = PseudoVMERGE_VVM_M1 %pt, %false, %true, %mask, 4, 5 /* e32 */
----------------
lukel97 wrote:

I guess we could still fold this case where true's vl < vmerge's vl if we check that true's policy operand is undisturbed. But that would require checking that true has a policy op. I'm not sure if this case is that common in practice.

https://github.com/llvm/llvm-project/pull/149720