[llvm] [NVPTX] support dynamic allocas with PTX alloca instruction (PR #84585)

[Artem Belevich via llvm-commits]

================
@@ -2211,14 +2210,24 @@ SDValue NVPTXTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
 
 SDValue NVPTXTargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
                                                      SelectionDAG &DAG) const {
-  const Function &Fn = DAG.getMachineFunction().getFunction();
-
-  DiagnosticInfoUnsupported NoDynamicAlloca(
-      Fn, "dynamic alloca unsupported by NVPTX backend",
-      SDLoc(Op).getDebugLoc());
-  DAG.getContext()->diagnose(NoDynamicAlloca);
-  auto Ops = {DAG.getConstant(0, SDLoc(), Op.getValueType()), Op.getOperand(0)};
-  return DAG.getMergeValues(Ops, SDLoc());
+  SDValue Chain = Op.getOperand(0);
+  SDValue Size = Op.getOperand(1);
+  uint64_t Align = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue();
+  SDLoc DL(Op.getNode());
+
+  // The size for ptx alloca instruction is 64-bit for m64 and 32-bit for m32.
+  if (nvTM->is64Bit())
+    Size = DAG.getZExtOrTrunc(Size, DL, MVT::i64);
+  else
+    Size = DAG.getZExtOrTrunc(Size, DL, MVT::i32);
+
+  SDValue AllocOps[] = {Chain, Size,
+                        DAG.getTargetConstant(Align, DL, MVT::i32)};
+  SDValue Alloca = DAG.getNode(NVPTXISD::DYNAMIC_STACKALLOC, DL,
+                               nvTM->is64Bit() ? MVT::i64 : MVT::i32, AllocOps);
+
+  SDValue MergeOps[] = {Alloca, Chain};
+  return DAG.getMergeValues(MergeOps, DL);
----------------
Artem-B wrote:

We still need to diagnose it in some cases, as `alloca` instruction is only available on sm_52+ and PTX7.3+

https://github.com/llvm/llvm-project/pull/84585