[llvm] [DAGCombiner] Narrow the integer source of uint_to_fp (PR #222899)

[Harrison Hao via llvm-commits]

harrisonGPU wrote:

> @harrisonGPU @MrSidims it looks like this patch #222899 and #223107 are basically doing the same thing but there are a bunch of differences in the implementations, e.g.
> 
> * this patch handles vector types
> * the other patch handles signed conversions
> 
> Can you work together to come up with one patch which is the best of both worlds?

Thanks, Jay! I really appreciate your comments. My patch affects other targets, so I need to wait for their reviews.
I have already implemented the same optimization for `SINT_TO_FP` locally, using the same approach. However, it also affects other targets, so I am waiting for the respective target owners to confirm that it is correct.

https://github.com/llvm/llvm-project/pull/222899