[llvm] [AMDGPU] Use `v_cvt_pk_*` instructions for i16_f32 saturated conversions (PR #202680)

[Jay Foad via llvm-commits]

================
@@ -3941,12 +3941,21 @@ SDValue AMDGPUTargetLowering::LowerFP_TO_INT_SAT(const SDValue Op,
   uint64_t SatWidth = SatVT.getScalarSizeInBits();
   assert(SatWidth <= DstWidth && "Saturation width cannot exceed result width");
 
+  // Select v2f32 -> v2i16 natively to v_cvt_pk_[iu]16_f32.
+  if (DstVT.isVector()) {
+    if (DstVT == MVT::v2i16 && SatWidth == 16 && SrcVT == MVT::v2f32)
+      return Op;
+
+    return SDValue();
----------------
jayfoad wrote:

OK. I guess this is fine for now but it looks like it could maybe be cleaned up. For example there is no need for your `if`s to be nested here, you could check `DstVT == MVT::v2i16 && ...` first and then `DstVT.isVector()`. Also all of the cases that get selected directly could be inside one big `if (DstWidth == SatWidth) { ... }`.

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