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

[Pierre van Houtryve via llvm-commits]

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

This will switch the CGP i16 promotion off, so we'd let GISel regress in the meantime?

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