[llvm] [AMDGPU] Add custom lowering of llvm.convert.from.arbitrary.fp for FP8 (PR #194144)

[Dmitry Sidorov via llvm-commits]

================
@@ -10569,6 +10584,123 @@ SDValue SITargetLowering::lowerWorkitemID(SelectionDAG &DAG, SDValue Op,
                      DAG.getValueType(SmallVT));
 }
 
+// Pack vector inputs as hw instruction is packed.
+// FIXME: note, packing loses lane-wise poison. Should we do something about
+// it?
+SDValue SITargetLowering::packBytesToI32(SelectionDAG &DAG, const SDLoc &SL,
+                                         SDValue Src, unsigned NumBytes,
+                                         unsigned FirstLane) {
+  EVT SrcEltVT = Src.getValueType().getVectorElementType();
+  SDValue PackedI32;
+  for (unsigned I = 0; I != NumBytes; ++I) {
----------------
MrSidims wrote:

Applied, thanks!

https://github.com/llvm/llvm-project/pull/194144