[llvm] [AMDGPU] Sink uniform buffer address offsets into soffset (PR #160939)

[Prasoon Mishra via llvm-commits]

================
@@ -328,9 +328,6 @@ bool AMDGPUDAGToDAGISel::matchLoadD16FromBuildVector(SDNode *N) const {
 }
 
 void AMDGPUDAGToDAGISel::PreprocessISelDAG() {
-  if (!Subtarget->d16PreservesUnusedBits())
----------------
PrasoonMishra wrote:

The buffer optimization implemented doesn't require d16 support, so I moved the gate to only apply to the D16-specific optimization. Also, this optimization does not require a GlobalISel version. DAGtoDAG combines run as part of the SelectionDAG pipelines.

https://github.com/llvm/llvm-project/pull/160939