[llvm] [AMDGPU] Update base addr of dyn alloca considering GrowingUp stack (PR #119822)

[Sergei Barannikov via llvm-commits]

================
@@ -4023,10 +4024,20 @@ SDValue SITargetLowering::lowerDYNAMIC_STACKALLOCImpl(SDValue Op,
   Chain = DAG.getCALLSEQ_START(Chain, 0, 0, dl);
 
   SDValue Size = Tmp2.getOperand(1);
-  SDValue SP = DAG.getCopyFromReg(Chain, dl, SPReg, VT);
-  Chain = SP.getValue(1);
+  SDValue SPOld = DAG.getCopyFromReg(Chain, dl, SPReg, VT);
+  Chain = SPOld.getValue(1);
   MaybeAlign Alignment = cast<ConstantSDNode>(Tmp3)->getMaybeAlignValue();
   const TargetFrameLowering *TFL = Subtarget->getFrameLowering();
+  Align StackAlign = TFL->getStackAlign();
+  if (Alignment && *Alignment > StackAlign) {
----------------
s-barannikov wrote:

With the above suggestion:
```suggestion
  if (Alignment > StackAlign) {
```

https://github.com/llvm/llvm-project/pull/119822