[llvm] [AMDGPU] Fix mul combine for MUL24 (PR #79110)

[Jay Foad via llvm-commits]

jayfoad wrote:

> > "mul24 x, (add y, 1) -> add (mul24 x, y), x" does not seem safe at all. You know that y+1 is a 24-bit value, but that does not guarantee that y is a 24-bit value.
> 
> Ah indeed, good catch. Should the combine just be removed?

I guess it's safe for a regular mul, so maybe move the "Skip if already mul24" up a bit?

https://github.com/llvm/llvm-project/pull/79110