[OMPIRBuilder] Use AS 0 for internal variables for AMDGPU

[sarnex]

We see some libomptarget test failures if we use the default global AS.

See #166459 for more info.

[llvmbot]

@llvm/pr-subscribers-clang

@llvm/pr-subscribers-flang-openmp

Author: Nick Sarnie (sarnex)

Changes

We see some libomptarget test failures if we use the default global as.

See #166459 for more info.

---

Full diff: https://github.com/llvm/llvm-project/pull/167377.diff

2 Files Affected:

• (modified) clang/test/OpenMP/force-usm.c (+1-1)
• (modified) llvm/lib/Frontend/OpenMP/OMPIRBuilder.cpp (+9-2)

diff --git a/clang/test/OpenMP/force-usm.c b/clang/test/OpenMP/force-usm.c
index 45c0e28b525da..5c63a9a5e7004 100644
--- a/clang/test/OpenMP/force-usm.c
+++ b/clang/test/OpenMP/force-usm.c
@@ -46,7 +46,7 @@ int main(void) {
 // CHECK-USM-NEXT:    br i1 [[EXEC_USER_CODE]], label [[USER_CODE_ENTRY:%.*]], label [[WORKER_EXIT:%.*]]
 // CHECK-USM:       user_code.entry:
 // CHECK-USM-NEXT:    store i32 1, ptr [[TMP0]], align 4
-// CHECK-USM-NEXT:    [[TMP2:%.*]] = load ptr, ptr addrspace(1) @pGI_decl_tgt_ref_ptr, align 8
+// CHECK-USM-NEXT:    [[TMP2:%.*]] = load ptr, ptr @pGI_decl_tgt_ref_ptr, align 8
 // CHECK-USM-NEXT:    [[TMP3:%.*]] = load ptr, ptr [[TMP2]], align 8
 // CHECK-USM-NEXT:    store i32 2, ptr [[TMP3]], align 4
 // CHECK-USM-NEXT:    call void @__kmpc_target_deinit()
diff --git a/llvm/lib/Frontend/OpenMP/OMPIRBuilder.cpp b/llvm/lib/Frontend/OpenMP/OMPIRBuilder.cpp
index 7dc32fda0eed6..44b4195982c49 100644
--- a/llvm/lib/Frontend/OpenMP/OMPIRBuilder.cpp
+++ b/llvm/lib/Frontend/OpenMP/OMPIRBuilder.cpp
@@ -8472,8 +8472,15 @@ GlobalVariable *OpenMPIRBuilder::getOrCreateInternalVariable(
     // create different versions of the function for different OMP internal
     // variables.
     const DataLayout &DL = M.getDataLayout();
-    unsigned AddressSpaceVal =
-        AddressSpace ? *AddressSpace : DL.getDefaultGlobalsAddressSpace();
+    // TODO: Investigate why AMDGPU expects AS 0 for globals even though the
+    // default global AS is 1.
+    // See double-target-call-with-declare-target.f90 and
+    // declare-target-vars-in-target-region.f90 libomptarget
+    // tests.
+    unsigned AddressSpaceVal = AddressSpace ? *AddressSpace
+                               : M.getTargetTriple().isAMDGPU()
+                                   ? 0
+                                   : DL.getDefaultGlobalsAddressSpace();
     auto Linkage = this->M.getTargetTriple().getArch() == Triple::wasm32
                        ? GlobalValue::InternalLinkage
                        : GlobalValue::CommonLinkage;

[Kewen12]

Hi! I tested this PR and the following fortran tests are no longer failed:

libomptarget :: amdgcn-amd-amdhsa :: offloading/fortran/declare-target-vars-in-target-region.f90
libomptarget :: amdgcn-amd-amdhsa :: offloading/fortran/double-target-call-with-declare-target.f90
libomptarget :: amdgcn-amd-amdhsa :: offloading/fortran/target-map-declare-target-link-common-block.f90

I am going to land this PR to unblock our bot. https://lab.llvm.org/staging/#/builders/105

[sarnex]

Thanks @Kewen12!