[llvm] [AMDGPU] Optimize fsub and fneg when packed fp32 ops are supported (PR #195962)

[Matt Arsenault via llvm-commits]

================
@@ -1134,6 +1134,9 @@ AMDGPULegalizerInfo::AMDGPULegalizerInfo(const GCNSubtarget &ST_,
       .lowerFor({S64, S16, V2S16});
   }
 
+  if (ST.hasPackedFP32Ops())
+    FSubActions.lowerFor({V2S32}).clampMaxNumElements(0, S32, 2);
----------------
arsenm wrote:

Why don't we have the element clamp for the legal v2f16 case 

https://github.com/llvm/llvm-project/pull/195962