[llvm] [NVPTX] Don't use stack memory when bitcasting to/from v2i8 (PR #113928)

[Justin Fargnoli via llvm-commits]

================
@@ -2311,6 +2315,47 @@ NVPTXTargetLowering::LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) const {
   return DAG.getBuildVector(Node->getValueType(0), dl, Ops);
 }
 
+SDValue NVPTXTargetLowering::LowerBITCAST(SDValue Op, SelectionDAG &DAG) const {
+  // Handle bitcasting to/from v2i8 without hitting the default promotion
+  // strategy which goes through stack memory.
+  SDNode *Node = Op.getNode();
+  SDLoc dl(Node);
+
+  auto maybeBitcast = [&](EVT vt, SDValue val) {
+    if (val->getValueType(0) == vt) {
+      return val;
+    }
+    return DAG.getNode(ISD::BITCAST, dl, vt, val);
+  };
+
+  EVT VT = Op->getValueType(0);
+  EVT fromVT = Op->getOperand(0)->getValueType(0);
+
+  if (VT == MVT::v2i8) {
+    // Bitcast to i16 and unpack elements into a vector
+    SDValue reg = maybeBitcast(MVT::i16, Op->getOperand(0));
+    SDValue v0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, reg);
+    SDValue C8 = DAG.getConstant(8, dl, MVT::i16);
+    SDValue v1 = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8,
----------------
justinfargnoli wrote:

I think it would be useful to have some more descriptive variable names here and in the `else if` below. 

https://github.com/llvm/llvm-project/pull/113928