[llvm] [AMDGPU] Promote uniform ops to I32 in ISel (PR #106383)

[Matt Arsenault via llvm-commits]

================
@@ -145,6 +145,31 @@ def expand_promoted_fmed3 : GICombineRule<
 
 } // End Predicates = [NotHasMed3_16]
 
+def promote_i16_uniform_binops_frag : GICombinePatFrag<
----------------
arsenm wrote:

I'd first implement all the DAG parts in one patch and leave the IR pass as it is. Then the next 2 parts can be done in either order, either do the globalisel parts or remove the IR promotion and let the regression happen 

https://github.com/llvm/llvm-project/pull/106383