[AMDGPU] Allow any linkage for dynlds (#84742)

Solves SWDEV-449592

2 files changed, 40 insertions, 7 deletions

diff --git a/llvm/lib/Target/AMDGPU/Utils/AMDGPUMemoryUtils.cpp b/llvm/lib/Target/AMDGPU/Utils/AMDGPUMemoryUtils.cpp
index 25e628e..79c359a 100644
--- a/llvm/lib/Target/AMDGPU/Utils/AMDGPUMemoryUtils.cpp
+++ b/llvm/lib/Target/AMDGPU/Utils/AMDGPUMemoryUtils.cpp
@@ -32,16 +32,12 @@ Align getAlign(DataLayout const &DL, const GlobalVariable *GV) {
 }
 
 bool isDynamicLDS(const GlobalVariable &GV) {
-  // external zero size addrspace(3) without initializer implies cuda/hip extern
-  // __shared__ the semantics for such a variable appears to be that all extern
-  // __shared__ variables alias one another. This hits different handling.
+  // external zero size addrspace(3) without initializer is dynlds.
   const Module *M = GV.getParent();
   const DataLayout &DL = M->getDataLayout();
-  if (GV.getType()->getPointerAddressSpace() != AMDGPUAS::LOCAL_ADDRESS) {
+  if (GV.getType()->getPointerAddressSpace() != AMDGPUAS::LOCAL_ADDRESS)
     return false;
-  }
-  uint64_t AllocSize = DL.getTypeAllocSize(GV.getValueType());
-  return GV.hasExternalLinkage() && AllocSize == 0;
+  return DL.getTypeAllocSize(GV.getValueType()) == 0;
 }
 
 bool isLDSVariableToLower(const GlobalVariable &GV) {
diff --git a/llvm/test/CodeGen/AMDGPU/lower-module-lds-zero-size-arr.ll b/llvm/test/CodeGen/AMDGPU/lower-module-lds-zero-size-arr.ll
new file mode 100644
index 0000000..da1d23f
--- /dev/null
+++ b/llvm/test/CodeGen/AMDGPU/lower-module-lds-zero-size-arr.ll
@@ -0,0 +1,37 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 4
+; RUN: opt -S -mtriple=amdgcn-- -passes=amdgpu-lower-module-lds < %s | FileCheck %s
+
+; This is an extension and should be rejected by the front-end in most cases.
+; If it goes through, lower it as dynlds.
+
+@Var0 = linkonce_odr hidden local_unnamed_addr addrspace(3) global [0 x float] poison
+
+define void @fn(float %val, i32 %idx) {
+; CHECK-LABEL: define void @fn(
+; CHECK-SAME: float [[VAL:%.*]], i32 [[IDX:%.*]]) {
+; CHECK-NEXT:    [[TMP1:%.*]] = call i32 @llvm.amdgcn.lds.kernel.id()
+; CHECK-NEXT:    [[VAR0:%.*]] = getelementptr inbounds [1 x i32], ptr addrspace(4) @llvm.amdgcn.dynlds.offset.table, i32 0, i32 [[TMP1]]
+; CHECK-NEXT:    [[TMP2:%.*]] = load i32, ptr addrspace(4) [[VAR0]], align 4
+; CHECK-NEXT:    [[VAR01:%.*]] = inttoptr i32 [[TMP2]] to ptr addrspace(3)
+; CHECK-NEXT:    [[PTR:%.*]] = getelementptr i32, ptr addrspace(3) [[VAR01]], i32 [[IDX]]
+; CHECK-NEXT:    store float [[VAL]], ptr addrspace(3) [[PTR]], align 4
+; CHECK-NEXT:    ret void
+;
+  %ptr = getelementptr i32, ptr addrspace(3) @Var0, i32 %idx
+  store float %val, ptr addrspace(3) %ptr
+  ret void
+}
+
+define amdgpu_kernel void @kernelA(float %val, i32 %idx) {
+; CHECK-LABEL: define amdgpu_kernel void @kernelA(
+; CHECK-SAME: float [[VAL:%.*]], i32 [[IDX:%.*]]) !llvm.amdgcn.lds.kernel.id [[META1:![0-9]+]] {
+; CHECK-NEXT:    call void @llvm.donothing() [ "ExplicitUse"(ptr addrspace(3) @llvm.amdgcn.kernelA.dynlds) ]
+; CHECK-NEXT:    tail call void @fn(float [[VAL]], i32 [[IDX]])
+; CHECK-NEXT:    ret void
+;
+  tail call void @fn(float %val, i32 %idx)
+  ret void
+}
+;.
+; CHECK: [[META1]] = !{i32 0}
+;.