[llvm] AMDGPU: Custom lower fptrunc vectors for f32 -> f16 (PR #141883)

[Matt Arsenault via llvm-commits]

================
@@ -6900,14 +6902,44 @@ SDValue SITargetLowering::getFPExtOrFPRound(SelectionDAG &DAG, SDValue Op,
                            DAG.getTargetConstant(0, DL, MVT::i32));
 }
 
+SDValue SITargetLowering::SplitFP_ROUNDVectorToPacks(SDValue Op,
+                                                     SelectionDAG &DAG) const {
+  unsigned Opc = Op.getOpcode();
+  EVT DstVT = Op.getValueType();
+  unsigned NumElts = DstVT.getVectorNumElements();
+  assert(NumElts % 2 == 0 && "Only handle vectors of even number of elements");
+  if (NumElts == 2) // already packed.
+    return Op;
+
+  SDValue Src = Op.getOperand(0);
+  EVT SrcVT = Src.getValueType();
+  LLVMContext &Context = *DAG.getContext();
+  EVT SrcPkVT = EVT::getVectorVT(Context, SrcVT.getScalarType(), 2);
+  EVT DstPkVT = EVT::getVectorVT(Context, DstVT.getScalarType(), 2);
+
+  SDLoc DL(Op);
+  SmallVector<SDValue, 16> Packs;
+  for (unsigned Index = 0; Index < NumElts; Index += 2) {
+    SDValue PkSrc = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SrcPkVT, Src,
+                                DAG.getConstant(Index, DL, MVT::i32));
+    SDValue PkDst = DAG.getNode(Opc, DL, DstPkVT, PkSrc,
+                                DAG.getTargetConstant(0, DL, MVT::i32));
----------------
arsenm wrote:

This should preserve the original value for the round flag 

https://github.com/llvm/llvm-project/pull/141883