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

Princeton Ferro via llvm-commits llvm-commits at lists.llvm.org
Fri Nov 21 16:13:18 PST 2025


================
@@ -2570,6 +2572,44 @@ 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();
+
+  switch (VT.getSimpleVT().SimpleTy) {
+  case MVT::i16: {
+    SDValue Extended = DAG.getNode(ISD::ANY_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);
+  }
+  case MVT::i32: {
+    return getPRMT(Src, DAG.getConstant(0, DL, MVT::i32), 0x0123, DL, DAG);
+  }
+  case MVT::v2i16: {
+    SDValue Converted = DAG.getBitcast(MVT::i32, Src);
+    SDValue Swapped =
+        getPRMT(Converted, DAG.getConstant(0, DL, MVT::i32), 0x2301, DL, DAG);
+    return DAG.getNode(ISD::BITCAST, DL, MVT::v2i16, Swapped);
+  }
+  case MVT::i64: {
----------------
Prince781 wrote:

Handle v2i32?

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


More information about the llvm-commits mailing list