[llvm] [NVPTX] Lower allocas to the local address space (PR #204346)

[Alex MacLean via llvm-commits]

================
@@ -1686,10 +1686,14 @@ SDValue NVPTXTargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
                   {Chain, DAG.getZExtOrTrunc(Size, DL, LocalVT),
                    DAG.getTargetConstant(Align, DL, MVT::i32)});
 
-  SDValue ASC = DAG.getAddrSpaceCast(
-      DL, Op.getValueType(), Alloc, ADDRESS_SPACE_LOCAL, ADDRESS_SPACE_GENERIC);
-
-  return DAG.getMergeValues({ASC, SDValue(Alloc.getNode(), 1)}, DL);
+  // Keep the result in the alloca's address space: return the local pointer as
+  // is when that is requested (escapes are cast by the caller), else cast it.
+  SDValue Ptr = Alloc;
+  if (Op.getValueType() != LocalVT)
----------------
AlexMaclean wrote:

This doesn't really make sense. In SDAG value types don't hold an addrspace so we're just comparing the size here. That would work for `nvptx64` with `shortprs` enabled but isn't a general solution since LOCAL and GENERIC can have the same size. Should we just remove the cast and trust the lowering pass?

https://github.com/llvm/llvm-project/pull/204346