[HIP] fix host-used external kernel

[yxsamliu]

In -fgpu-rdc mode, when an external kernel is used by a host function with weak_odr linkage (e.g. explicitly instantiated template function), the kernel should not be marked as host-used external kernel, since the host function may be dropped by the linker. Mark the external kernel as host-used external kernel will force a reference to the external kernel, which the user may not define in other TU.

Fixes: #83771

[llvmbot]

@llvm/pr-subscribers-clang

Author: Yaxun (Sam) Liu (yxsamliu)

Changes

In -fgpu-rdc mode, when an external kernel is used by a host function with weak_odr linkage (e.g. explicitly instantiated template function), the kernel should not be marked as host-used external kernel, since the host function may be dropped by the linker. Mark the external kernel as host-used external kernel will force a reference to the external kernel, which the user may not define in other TU.

Fixes: #83771

---

Full diff: https://github.com/llvm/llvm-project/pull/83870.diff

3 Files Affected:

• (modified) clang/lib/Sema/SemaCUDA.cpp (+2-1)
• (modified) clang/lib/Sema/SemaExpr.cpp (+3-1)
• (modified) clang/test/CodeGenCUDA/host-used-extern.cu (+13)

diff --git a/clang/lib/Sema/SemaCUDA.cpp b/clang/lib/Sema/SemaCUDA.cpp
index 6a66ecf6f94c17..40a2d47e4da127 100644
--- a/clang/lib/Sema/SemaCUDA.cpp
+++ b/clang/lib/Sema/SemaCUDA.cpp
@@ -895,7 +895,8 @@ bool Sema::CheckCUDACall(SourceLocation Loc, FunctionDecl *Callee) {
   if (DiagKind == SemaDiagnosticBuilder::K_Nop) {
     // For -fgpu-rdc, keep track of external kernels used by host functions.
     if (LangOpts.CUDAIsDevice && LangOpts.GPURelocatableDeviceCode &&
-        Callee->hasAttr<CUDAGlobalAttr>() && !Callee->isDefined())
+        Callee->hasAttr<CUDAGlobalAttr>() && !Callee->isDefined() &&
+        getASTContext().GetGVALinkageForFunction(Caller) == GVA_StrongExternal)
       getASTContext().CUDAExternalDeviceDeclODRUsedByHost.insert(Callee);
     return true;
   }
diff --git a/clang/lib/Sema/SemaExpr.cpp b/clang/lib/Sema/SemaExpr.cpp
index 0a449fc1082bd4..dd16fc4eeffb72 100644
--- a/clang/lib/Sema/SemaExpr.cpp
+++ b/clang/lib/Sema/SemaExpr.cpp
@@ -19211,7 +19211,9 @@ MarkVarDeclODRUsed(ValueDecl *V, SourceLocation Loc, Sema &SemaRef,
     } else if (VarTarget == Sema::CVT_Device &&
                !Var->hasAttr<CUDASharedAttr>() &&
                (UserTarget == Sema::CFT_Host ||
-                UserTarget == Sema::CFT_HostDevice)) {
+                UserTarget == Sema::CFT_HostDevice) &&
+               SemaRef.getASTContext().GetGVALinkageForFunction(FD) ==
+                   GVA_StrongExternal) {
       // Record a CUDA/HIP device side variable if it is ODR-used
       // by host code. This is done conservatively, when the variable is
       // referenced in any of the following contexts:
diff --git a/clang/test/CodeGenCUDA/host-used-extern.cu b/clang/test/CodeGenCUDA/host-used-extern.cu
index e8f8e12aad47d1..1db6123916c700 100644
--- a/clang/test/CodeGenCUDA/host-used-extern.cu
+++ b/clang/test/CodeGenCUDA/host-used-extern.cu
@@ -24,6 +24,7 @@
 
 // NEG-NOT: @__clang_gpu_used_external = {{.*}} @_Z7kernel2v
 // NEG-NOT: @__clang_gpu_used_external = {{.*}} @_Z7kernel3v
+// XEG-NOT: @__clang_gpu_used_external = {{.*}} @_Z7kernel5v
 // NEG-NOT: @__clang_gpu_used_external = {{.*}} @var2
 // NEG-NOT: @__clang_gpu_used_external = {{.*}} @var3
 // NEG-NOT: @__clang_gpu_used_external = {{.*}} @ext_shvar
@@ -44,6 +45,10 @@ __global__ void kernel3();
 // kernel4 is marked as used even though it is not called.
 __global__ void kernel4();
 
+// kernel5 is not marked as used since it is called by host function
+// with linkonce_odr linkage, which may be dropped by linker.
+__global__ void kernel5();
+
 extern __device__ int var1;
 
 __device__ int var2;
@@ -67,3 +72,11 @@ __global__ void test_lambda_using_extern_shared() {
   };
   lambda();
 }
+
+template<class T>
+void template_caller() {
+  kernel5<<<1, 1>>>();
+  var1 = 1;
+}
+
+template void template_caller<int>();

[Artem-B]

LGTM in principle, but I'd run it by someone with more familiarity with linking quirks.

@MaskRay PTAL, when you get a chance.

[Artem-B]

Did you mean NEG-NOT ?

[yxsamliu]

fixed