[llvm] [SimplifyCFG] Treat umul + extract pattern as cheap single instruction. (PR #124933)

[Gábor Spaits via llvm-commits]

spaits wrote:

Gentle ping @nikic .
I have checked `foldTwoEntryPHINode` and `dominatesMergePoint`.

It turns out, that the cost of
```llvm
 %mul = tail call { i64, i1 } @llvm.umul.with.overflow.i64(i64 %add, i64 %x)
```
is 1, when there is no target specified. When using the riscv64 target then this value is 5.

`dominatesMergePoint` works recursively so I think it would be more complex, to "change" the cost of the pattern there. I would personally keep things the way they are in the current version of the patch, but if you would want this to move elsewhere the I am going to do that.

https://github.com/llvm/llvm-project/pull/124933