[llvm] [AMDGPU] Optimize rotate/funnel shift pattern matching in instruction selection (PR #149817)

[Matt Arsenault via llvm-commits]

================
@@ -4105,6 +4112,168 @@ void AMDGPUDAGToDAGISel::PostprocessISelDAG() {
   } while (IsModified);
 }
 
+// Pattern matching for rotate/funnel shift operations
+// and converts them to v_alignbit_b32 instructions
----------------
arsenm wrote:

Why does this need manual selection? Several of these cases look like they belong in the funnel shift combines or could be moved to patterns 

https://github.com/llvm/llvm-project/pull/149817