AMDGPU: Fix handling of negative scratch offset

[macurtis-amd]

No description provided.

[llvmbot]

@llvm/pr-subscribers-backend-amdgpu

Author: None (macurtis-amd)

Changes

---

Full diff: https://github.com/llvm/llvm-project/pull/122135.diff

2 Files Affected:

• (modified) llvm/lib/Target/AMDGPU/AMDGPUISelDAGToDAG.cpp (+1-1)
• (added) llvm/test/CodeGen/AMDGPU/bug-neg-scratch-offset.ll (+36)

diff --git a/llvm/lib/Target/AMDGPU/AMDGPUISelDAGToDAG.cpp b/llvm/lib/Target/AMDGPU/AMDGPUISelDAGToDAG.cpp
index d9eaf82c521404..27e9018d68a03e 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUISelDAGToDAG.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPUISelDAGToDAG.cpp
@@ -1997,7 +1997,7 @@ bool AMDGPUDAGToDAGISel::SelectScratchSVAddr(SDNode *N, SDValue Addr,
   if (checkFlatScratchSVSSwizzleBug(VAddr, SAddr, ImmOffset))
     return false;
   SAddr = SelectSAddrFI(CurDAG, SAddr);
-  Offset = CurDAG->getTargetConstant(ImmOffset, SDLoc(), MVT::i32);
+  Offset = CurDAG->getSignedTargetConstant(ImmOffset, SDLoc(), MVT::i32);
   return true;
 }
 
diff --git a/llvm/test/CodeGen/AMDGPU/bug-neg-scratch-offset.ll b/llvm/test/CodeGen/AMDGPU/bug-neg-scratch-offset.ll
new file mode 100644
index 00000000000000..9141b4fd188ce1
--- /dev/null
+++ b/llvm/test/CodeGen/AMDGPU/bug-neg-scratch-offset.ll
@@ -0,0 +1,36 @@
+; RUN: llc < %s -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1200 -stop-after=amdgpu-isel -verify-machineinstrs | FileCheck %s
+
+target datalayout = "e-p:64:64-p1:64:64-p2:32:32-p3:32:32-p4:64:64-p5:32:32-p6:32:32-p7:160:256:256:32-p8:128:128-p9:192:256:256:32-i64:64-v16:16-v24:32-v32:32-v48:64-v96:128-v192:256-v256:256-v512:512-v1024:1024-v2048:2048-n32:64-S32-A5-G1-ni:7:8:9"
+target triple = "amdgcn-amd-amdhsa"
+
+%union.anon.41 = type { [4 x i64] }
+%union.anon.2 = type { i8 }
+
+define fastcc void @_ZN10PrimitivesI12rccl_bfloat810FuncMinMaxIS0_E13FanAsymmetricILi1ELi1EELi1E10ProtoLL128Li0EE9localCopyEPS0_S7_i(i32 %0, i64 %idx.ext62.i.i) {
+entry:
+  %1 = alloca %union.anon.41, i32 0, align 8, addrspace(5)
+  %add.ptr63.i.i3 = getelementptr %union.anon.2, ptr null, i64 %idx.ext62.i.i
+  br label %for.body69.i.i.epil3
+
+for.body69.i.i.epil3:                             ; preds = %for.body69.i.i.epil3, %entry
+  %i.0117.i.i.epil4 = phi i32 [ %inc.i.i.7.epil, %for.body69.i.i.epil3 ], [ %0, %entry ]
+  %conv65.i.i.epil5 = zext i32 %i.0117.i.i.epil4 to i64
+  %arrayidx73.i.i.epil6 = getelementptr [32 x i8], ptr addrspace(5) %1, i32 0, i32 %i.0117.i.i.epil4
+  %add.ptr75.i.i.epil7 = getelementptr i8, ptr %add.ptr63.i.i3, i64 %conv65.i.i.epil5
+  %2 = load <4 x i8>, ptr addrspace(5) %arrayidx73.i.i.epil6, align 8
+  store <4 x i8> %2, ptr %add.ptr75.i.i.epil7, align 1
+  %inc.i.i.3.epil = or disjoint i32 %i.0117.i.i.epil4, 1
+  %conv65.i.i.4.epil = zext i32 %inc.i.i.3.epil to i64
+  %arrayidx73.i.i.4.epil = getelementptr [32 x i8], ptr addrspace(5) %1, i32 0, i32 %inc.i.i.3.epil
+  %add.ptr75.i.i.4.epil = getelementptr i8, ptr %add.ptr63.i.i3, i64 %conv65.i.i.4.epil
+  %3 = load <4 x i8>, ptr addrspace(5) %arrayidx73.i.i.4.epil, align 4
+  store <4 x i8> %3, ptr %add.ptr75.i.i.4.epil, align 1
+  %inc.i.i.7.epil = add nuw i32 %i.0117.i.i.epil4, 1
+  br label %for.body69.i.i.epil3
+
+for.body69.i.i.epil3.for.cond.cleanup68.loopexit.i.i.unr-lcssa_crit_edge: ; No predecessors!
+  %conv65.i.i.epil = zext i32 %inc.i.i.7.epil to i64
+  ret void
+}
+
+; CHECK: SCRATCH_LOAD_DWORD_SVS %{{[0-9]+}}, %{{[0-9]+}}, -1

[shiltian]

this two lines are not needed

[shiltian]

Did I travel in time? I thought I saw a very similar PR yesterday.
I'm not. #121978

[macurtis-amd]

Did I travel in time? I thought I saw a very similar PR yesterday. I'm not. #1219

I'm too slow ... closing.