[clang][cuda] Use the source filename for module ID

[mtrofin]

Prior to #184065 (relanded in #201849), the code introduced in https://reviews.llvm.org/D42922 was using the GUID of an internal linkage GlobalValue to create a module id, which would then be used in a few cuda-specific places (including creating a symbol name suffix). #184065 assumed the linkage of that symbol is external - which it isn't - and, thus, all module IDs computed for this would be identical.

The fix is to not rely on GlobalValue GUIDs in the first place. What is needed here is a hash that's specific to this module. So we're creating that and explicitly decoupling that calculation from GlobalValue's notion of GUIDs.

The change should be safe wrt uniqueness / hash collisions: the old behavior was computing a hash where the only distinguishing component was the module's getSourceFileName.

FWIW, there's a stronger module ID we could compute, with llvm::getUniqueModuleId (or with a trivial refactoring of it, because it prepends a ".").

[mtrofin]

• [clang][cuda] Use the source filename for module ID #209239 👈 (View in Graphite)
• main

This stack of pull requests is managed by Graphite. Learn more about stacking.

[llvmorg-github-actions]

@llvm/pr-subscribers-clang

@llvm/pr-subscribers-clang-codegen

Author: Mircea Trofin (mtrofin)

Changes

Prior to #184065 (relanded in #201849), the code introduced in https://reviews.llvm.org/D42922 was using the GUID of an internal linkage GlobalValue to create a module id, which would then be used in a few cuda-specific places (including creating a symbol name suffix). #184065 assumed the linkage of that symbol is external - which it isn't - and, thus, all module IDs computed for this would be identical.

The fix is to not rely on GlobalValue GUIDs in the first place. What is needed here is a hash that's specific to this module. So we're creating that and explicitly decoupling that calculation from GlobalValue's notion of GUIDs.

The change should be safe wrt uniqueness / hash collisions: the old behavior was computing a hash where the only distinguishing component was the module's getSourceFileName.

---

Full diff: https://github.com/llvm/llvm-project/pull/209239.diff

2 Files Affected:

• (modified) clang/lib/CodeGen/CGCUDANV.cpp (+1-1)
• (modified) clang/test/CodeGenCUDA/device-stub.cu (+13)

diff --git a/clang/lib/CodeGen/CGCUDANV.cpp b/clang/lib/CodeGen/CGCUDANV.cpp
index 733f2b7ecb46e..6467a5883b927 100644
--- a/clang/lib/CodeGen/CGCUDANV.cpp
+++ b/clang/lib/CodeGen/CGCUDANV.cpp
@@ -1046,7 +1046,7 @@ llvm::Function *CGNVCUDARuntime::makeModuleCtorFunction() {
     OS << ModuleIDPrefix
        << llvm::format("%" PRIx64,
                        llvm::GlobalValue::getGUIDAssumingExternalLinkage(
-                           FatbinWrapper->getName()));
+                           TheModule.getSourceFileName()));
     llvm::Constant *ModuleIDConstant = makeConstantArray(
         std::string(ModuleID), "", ModuleIDSectionName, 32, /*AddNull=*/true);
 
diff --git a/clang/test/CodeGenCUDA/device-stub.cu b/clang/test/CodeGenCUDA/device-stub.cu
index 38c7dc711ef1d..307ab190aa5be 100644
--- a/clang/test/CodeGenCUDA/device-stub.cu
+++ b/clang/test/CodeGenCUDA/device-stub.cu
@@ -64,6 +64,19 @@
 // RUN:     -o - -x hip\
 // RUN:   | FileCheck -allow-deprecated-dag-overlap %s --check-prefixes=ALL,WIN,HIP,HIPNEF
 
+// Verify that module IDs are distinct when the module source path is distinct.
+// RUN: rm -rf %t_distinct
+// RUN: mkdir -p %t_distinct
+// RUN: cp %s %t_distinct/filename1.cu
+// RUN: cp %s %t_distinct/filename2.cu
+// RUN: %clang_cc1 -triple x86_64-linux-gnu -emit-llvm %t_distinct/filename1.cu -I%S \
+// RUN:     -target-sdk-version=8.0 -fgpu-rdc -fcuda-include-gpubinary %t \
+// RUN:     -o - | grep __nv_module_id &> %t_module_id_1
+// RUN: %clang_cc1 -triple x86_64-linux-gnu -emit-llvm %t_distinct/filename2.cu -I%S \
+// RUN:     -target-sdk-version=8.0 -fgpu-rdc -fcuda-include-gpubinary %t \
+// RUN:     -o - | grep __nv_module_id &> %t_module_id_2
+// RUN: not diff %t_module_id_1 %t_module_id_2
+
 #include "Inputs/cuda.h"
 
 #ifndef NOGLOBALS

[alexfh]

Thank you! I've verified that it fixes the compilation problems caused by 39dcb0f.

[teresajohnson]

lgtm but with a comment suggestion

[yxsamliu]

LGTM. Matches CUDA RDC per-TU registration.

[mtrofin]

ignoring CI LLDB API failures, low chance they are related, plus apparently known issue (@boomanaiden154)