[llvm] [AMDGPU] Use `v_cvt_pk_*` instructions for saturated conversions (PR #202680)

[Matt Arsenault via llvm-commits]

================
@@ -3949,6 +3949,10 @@ SDValue AMDGPUTargetLowering::LowerFP_TO_INT_SAT(const SDValue Op,
   if (DstVT == MVT::i16 && SatWidth == DstWidth && SrcVT == MVT::f16)
     return Op;
 
+  if (Subtarget->hasGFX11Insts() && DstVT == MVT::i16 && SrcVT == MVT::f32 &&
----------------
arsenm wrote:

Avoid generation checks 

https://github.com/llvm/llvm-project/pull/202680