[llvm] [RISCV] Allow folding vmerge with implicit merge operand when true has tied dest (PR #78565)

[Luke Lau via llvm-commits]

lukel97 wrote:

> Could this PR replace #78403 ?

Not on its own, because the issue in #78403 stems from the fact that `DAGCombiner::foldSelectWithIdentityConstant` doesn't kick in with the sext/zext in between the binary op, so we don't transform

```
add a, (zext (vselect Cond, 0, b))
```

to

```
vselect Cond, a, (add a, (zext FVal))
```

So the vselect/vmerge won't be pulled outside the add which is needed for the vmerge peephole.

With that said, I returned to this PR because I was attempting to generalize #78403 by teaching `DAGCombiner::foldSelectWithIdentityConstant` to look through extends. That way we would have the vmerge pulled outside the add/sub, and the vmerge peephole would automatically take care of folding the mask and we wouldn't need combineVWADDWSelect. And I think we would also be able to handle muls and floating point binary ops for free too. But this PR was needed to preserve some of the test cases in #78403



https://github.com/llvm/llvm-project/pull/78565