[llvm] [InstCombine] Fold uitofp of masked truncations (PR #222592)

[Harrison Hao via llvm-commits]

harrisonGPU wrote:

> this seems not beneficial for: f32, f64, v2f32, v2f64 on amdgpu12.5
> 
> https://godbolt.org/z/8E9M6eM3G

Thanks, really appreciate. AMDGPU uses `n32:64` as its native integer widths, so I added an `isLegalInteger` check on the original narrow type.
This prevents the fold from widening an already legal type, avoiding the regressions you observed.

https://github.com/llvm/llvm-project/pull/222592