[llvm] [AMDGPU][GlobalISel] Fix register count for d16 gather4 on gfx8.1/gfx9 (PR #219891)

[Matt Arsenault via llvm-commits]

================
@@ -2188,7 +2193,7 @@ bool AMDGPUInstructionSelector::selectImageIntrinsic(
       VDataTy = MRI->getType(VDataOut);
       NumVDataDwords = DMaskLanes;
 
-      if (IsD16 && !STI.hasUnpackedD16VMem())
+      if (IsD16 && NoUnpackedD16VMem && !HasGather4D16Bug)
----------------
arsenm wrote:

Should this be applied during the custom lowering rather than during selection? 

https://github.com/llvm/llvm-project/pull/219891