[llvm] [NVPTX] Use PRMT instruction to lower i16 bswap (PR #168968)

[Alex MacLean via llvm-commits]

================
@@ -2570,6 +2574,45 @@ static SDValue lowerTcgen05St(SDValue Op, SelectionDAG &DAG) {
   return Tcgen05StNode;
 }
 
+static SDValue lowerBSWAP(SDValue Op, SelectionDAG &DAG) {
+  SDLoc DL(Op);
+  SDValue Src = Op.getOperand(0);
+  EVT VT = Op.getValueType();
+
+  if (VT == MVT::i16) {
+    SDValue Extended = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, Src);
+    SDValue Swapped =
+        getPRMT(Extended, DAG.getConstant(0, DL, MVT::i32), 0x7701, DL, DAG);
+    return DAG.getNode(ISD::TRUNCATE, DL, MVT::i16, Swapped);
+  }
+
+  if (VT == MVT::i32) {
+    return getPRMT(Src, DAG.getConstant(0, DL, MVT::i32), 0x0123, DL, DAG);
+  }
+
+  if (VT == MVT::v2i16) {
+    SDValue Converted = DAG.getNode(ISD::BITCAST, DL, MVT::i32, Src);
----------------
AlexMaclean wrote:

Nit: use DAG.getBitcast


https://github.com/llvm/llvm-project/pull/168968