diff --git a/runtime/hsa-runtime/hsacore.so.def b/runtime/hsa-runtime/hsacore.so.def
index c415bb811a..ea90b3363e 100644
--- a/runtime/hsa-runtime/hsacore.so.def
+++ b/runtime/hsa-runtime/hsacore.so.def
@@ -220,6 +220,7 @@ global:
 	hsa_amd_queue_set_priority;
 	hsa_amd_register_deallocation_callback;
 	hsa_amd_deregister_deallocation_callback;
+	_amdgpu_r_debug;
 
 local:
     *;
diff --git a/runtime/hsa-runtime/loader/executable.cpp b/runtime/hsa-runtime/loader/executable.cpp
index 0ec868cbc9..5d65a32b97 100644
--- a/runtime/hsa-runtime/loader/executable.cpp
+++ b/runtime/hsa-runtime/loader/executable.cpp
@@ -62,18 +62,17 @@
 using namespace amd::hsa;
 using namespace amd::hsa::common;
 
-#if defined __clang__
-#define NONOPTIMIZE __attribute__((noinline, optnone))
-#else
-#define NONOPTIMIZE __attribute__((noinline, optimize(0)))
-#endif
-
-NONOPTIMIZE static void _loader_debug_state() {};
-r_debug _amdgpu_r_debug __attribute__((visibility("default"))) = {1,
-                                                                  nullptr,
-                                                                  reinterpret_cast<uintptr_t>(&_loader_debug_state),
-                                                                  r_debug::RT_CONSISTENT,
-                                                                  0};
+// Having a side effect prevents call site optimization that allows removal of a noinline function call
+// with no side effect.
+__attribute__((noinline)) static void _loader_debug_state() {
+  static volatile int function_needs_a_side_effect = 0;
+  function_needs_a_side_effect ^= 1;
+}
+HSA_API r_debug _amdgpu_r_debug = {1,
+                           nullptr,
+                           reinterpret_cast<uintptr_t>(&_loader_debug_state),
+                           r_debug::RT_CONSISTENT,
+                           0};
 static link_map* r_debug_tail = nullptr;
 
 namespace amd {