[llvm] [AMDGPU] Lower uniform uaddsat to SALU instructions (PR #210156)

[Jay Foad via llvm-commits]

jayfoad wrote:

Title should mention that this is the 16-bit case.

As discussed in #203155, having explicit ANDs with 0xFFFF in the selection pattern is bad because then it is too late for any DAG combines to optimize them away. Instead of this, please try adding UADDSAT to promoteUniformOpToI32, just like you did for USUBSAT in #203155.

https://github.com/llvm/llvm-project/pull/210156