[llvm] [AMDGPU] Do not scale private alloca size when using flat-scratch (PR #201142)

Matt Arsenault via llvm-commits llvm-commits at lists.llvm.org
Wed Jun 3 03:40:21 PDT 2026


Juan Manuel Martinez =?utf-8?q?CaamaƱo?Message-ID:
In-Reply-To: <llvm.org/llvm/llvm-project/pull/201142 at github.com>


================
@@ -4694,21 +4698,29 @@ SDValue SITargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
   assert(Size.getValueType() == MVT::i32 && "Size must be 32-bit");
   SDValue NewSP;
   if (isa<ConstantSDNode>(Size)) {
-    // For constant sized alloca, scale alloca size by wave-size
-    SDValue ScaledSize = DAG.getNode(
-        ISD::SHL, dl, VT, Size,
-        DAG.getConstant(Subtarget->getWavefrontSizeLog2(), dl, MVT::i32));
+    // Increase the stack pointer by the size of the alloca.
+    // If not using flat-scratch, we have to scale the size by the wave-size.
+    SDValue ScaledSize = Size;
+    if (!HasFlatScratch) {
+      ScaledSize =
+          DAG.getNode(ISD::SHL, dl, VT, Size,
+                      DAG.getConstant(WavefrontSizeLog2, dl, MVT::i32));
+    }
----------------
arsenm wrote:

```suggestion
    SDValue ScaledSize = HasFlatScratch ? Size : 
          DAG.getNode(ISD::SHL, dl, VT, Size,
                      DAG.getConstant(WavefrontSizeLog2, dl, MVT::i32));
```

https://github.com/llvm/llvm-project/pull/201142


More information about the llvm-commits mailing list