[PGO][HIP] Stop pulling ROCm.o into every PGO host link

[yxsamliu]

PR #177665 added an unconditional extern reference to __llvm_profile_hip_collect_device_data from InstrProfilingFile.c, which forces InstrProfilingPlatformROCm.o (and its sanitizer_common / interception dependencies) out of libclang_rt.profile.a in every PGO binary. That breaks bots without -lpthread and races dlsym/PLT state in non-HIP programs via the interceptor constructor.

Fix:

• Declare the hook COMPILER_RT_WEAK and gate the call on its address. No COMPILER_RT_VISIBILITY: a hidden weak-undef function would be non-preemptible and the address test would fold to true.
• Gate installHipModuleInterceptors on dlsym(hipModuleLoad) so the constructor is a no-op if ROCm.o is still pulled in.

Fixes:

• https://lab.llvm.org/buildbot/#/builders/66/builds/31311
• https://lab.llvm.org/buildbot/#/builders/174/builds/36180

Verified:

• check-profile 134/134 pass.
• nm on a non-HIP clang -fprofile-generate binary: zero installHip/ROCm/sanitizer/hip_collect symbols.
• HIP offload PGO end-to-end on gfx1101 (compile → run → llvm-profdata merge → llvm-cov) still works; interceptor installs, device profile collected via shared API.

[llvmorg-github-actions]

@llvm/pr-subscribers-pgo

Author: Yaxun (Sam) Liu (yxsamliu)

Changes

PR #177665 added an unconditional extern reference to __llvm_profile_hip_collect_device_data from InstrProfilingFile.c, which forces InstrProfilingPlatformROCm.o (and its sanitizer_common / interception dependencies) out of libclang_rt.profile.a in every PGO binary. That breaks bots without -lpthread and races dlsym/PLT state in non-HIP programs via the interceptor constructor.

Fix:

• Declare the hook COMPILER_RT_WEAK and gate the call on its address. No COMPILER_RT_VISIBILITY: a hidden weak-undef function would be non-preemptible and the address test would fold to true.
• Gate installHipModuleInterceptors on dlsym(hipModuleLoad) so the constructor is a no-op if ROCm.o is still pulled in.

Fixes:

• https://lab.llvm.org/buildbot/#/builders/66/builds/31311
• https://lab.llvm.org/buildbot/#/builders/174/builds/36180

Verified:

• check-profile 134/134 pass.
• nm on a non-HIP clang -fprofile-generate binary: zero installHip/ROCm/sanitizer/hip_collect symbols.
• HIP offload PGO end-to-end on gfx1101 (compile → run → llvm-profdata merge → llvm-cov) still works; interceptor installs, device profile collected via shared API.

---

Full diff: https://github.com/llvm/llvm-project/pull/200101.diff

2 Files Affected:

• (modified) compiler-rt/lib/profile/InstrProfilingFile.c (+13-7)
• (modified) compiler-rt/lib/profile/InstrProfilingPlatformROCm.cpp (+7)

diff --git a/compiler-rt/lib/profile/InstrProfilingFile.c b/compiler-rt/lib/profile/InstrProfilingFile.c
index 78e1edb54768f..8d068c103574f 100644
--- a/compiler-rt/lib/profile/InstrProfilingFile.c
+++ b/compiler-rt/lib/profile/InstrProfilingFile.c
@@ -41,9 +41,14 @@
 #include "InstrProfilingPort.h"
 #include "InstrProfilingUtil.h"
 
-/* HIP / offload collection hook implemented in InstrProfilingPlatformROCm.c.
- * It is a no-op when no offload profile data was registered. */
-extern int __llvm_profile_hip_collect_device_data(void);
+/* Weak so non-HIP programs do not force InstrProfilingPlatformROCm.o (and its
+ * transitive sanitizer_common / interception dependencies) into the host link
+ * out of libclang_rt.profile.a. HIP programs emit strong references to other
+ * ROCm-runtime symbols (e.g. __llvm_profile_offload_register_shadow_variable)
+ * that pull in the strong definition.
+ * No COMPILER_RT_VISIBILITY: a hidden weak-undefined symbol is non-preemptible
+ * and the address test at the call site would fold to true. */
+COMPILER_RT_WEAK int __llvm_profile_hip_collect_device_data(void);
 
 /* From where is profile name specified.
  * The order the enumerators define their
@@ -1202,10 +1207,11 @@ int __llvm_profile_write_file(void) {
   if (rc)
     PROF_ERR("Failed to write file \"%s\": %s\n", Filename, strerror(errno));
 
-  /* No-op when no HIP shadow variables or dynamic modules are registered,
-   * or when the HIP runtime is not loaded. Warning on failure is handled
-   * inside the callee so non-HIP programs do not see spurious noise. */
-  (void)__llvm_profile_hip_collect_device_data();
+  /* Weak: only invoked when InstrProfilingPlatformROCm.o is in the link,
+   * which happens when the program references other ROCm-runtime symbols
+   * (HIP-with-PGO). Warning on failure is handled inside the callee. */
+  if (__llvm_profile_hip_collect_device_data)
+    (void)__llvm_profile_hip_collect_device_data();
 
   // Restore SIGKILL.
   if (PDeathSig == 1)
diff --git a/compiler-rt/lib/profile/InstrProfilingPlatformROCm.cpp b/compiler-rt/lib/profile/InstrProfilingPlatformROCm.cpp
index 9e234da5b5cc4..ee00c572e3a42 100644
--- a/compiler-rt/lib/profile/InstrProfilingPlatformROCm.cpp
+++ b/compiler-rt/lib/profile/InstrProfilingPlatformROCm.cpp
@@ -24,6 +24,7 @@ extern "C" {
 #define WIN32_LEAN_AND_MEAN
 #include <windows.h>
 #else
+#include <dlfcn.h>
 #include <pthread.h>
 #endif
 
@@ -878,6 +879,12 @@ INTERCEPTOR(int, hipModuleUnload, void *module) {
 }
 
 __attribute__((constructor)) static void installHipModuleInterceptors() {
+  /* Skip when the HIP runtime is not loaded. INTERCEPT_FUNCTION uses the
+   * sanitizer interception framework, which can perturb dlsym/PLT state for
+   * the rest of the process even when the target symbol is absent; non-HIP
+   * programs linked with libclang_rt.profile.a must see zero side effects. */
+  if (!dlsym(RTLD_DEFAULT, "hipModuleLoad"))
+    return;
   if (!INTERCEPT_FUNCTION(hipModuleLoad))
     return;
   if (isVerboseMode())

[github-actions]

🪟 Windows x64 Test Results

• 1631 tests passed
• 1342 tests skipped

✅ The build succeeded and all tests passed.