[llvm] [AMDGPU] Move SGPR Count Queries into TargetParser (PR #209848)

[Matt Arsenault via llvm-commits]

================
@@ -1491,8 +1465,9 @@ unsigned getOccupancyWithNumSGPRs(const MCSubtargetInfo &STI, unsigned SGPRs) {
   if (!isSGPROccupancyLimited(STI))
     return MaxWaves;
 
-  return getOccupancyWithNumSGPRs(SGPRs, MaxWaves, getTotalNumSGPRs(STI),
-                                  getSGPRAllocGranule(STI),
+  GPUKind Kind = parseArchAMDGCN(STI.getCPU());
----------------
arsenm wrote:

Ideally we would have had independent subtarget features for the register counts, but that's another annoying change 

https://github.com/llvm/llvm-project/pull/209848