diff --git a/openmp/libomptarget/DeviceRTL/include/Synchronization.h b/openmp/libomptarget/DeviceRTL/include/Synchronization.h
--- a/openmp/libomptarget/DeviceRTL/include/Synchronization.h
+++ b/openmp/libomptarget/DeviceRTL/include/Synchronization.h
@@ -74,6 +74,10 @@
 /// Atomically add \p V to \p *Addr with \p Ordering semantics.
 uint64_t add(uint64_t *Addr, uint64_t V, int Ordering);
 
+/// Atomically write \p V to \p *Addr with \p Ordering semantics and return the
+/// old value of \p *Addr.
+uint32_t exchange(uint32_t *Addr, uint32_t V, int Ordering);
+
 } // namespace atomic
 
 } // namespace _OMP
diff --git a/openmp/libomptarget/DeviceRTL/src/Debug.cpp b/openmp/libomptarget/DeviceRTL/src/Debug.cpp
--- a/openmp/libomptarget/DeviceRTL/src/Debug.cpp
+++ b/openmp/libomptarget/DeviceRTL/src/Debug.cpp
@@ -13,16 +13,35 @@
 #include "Debug.h"
 #include "Configuration.h"
 #include "Mapping.h"
+#include "Synchronization.h"
 #include "Types.h"
+#include "Utils.h"
 
 using namespace _OMP;
 
 #pragma omp declare target
 
+/// Uninitialized on purpose to avoid any cost in case assertions are disabled
+/// or we don't validate any. The likelihood AssertionFlag contains the value we
+/// use to identify a broken assumption is negligible. Using the undefined value
+/// in the comparison and then branch is technically UB, however the atomic
+/// (builtin) access we use to read it does not expose the undef value to the
+/// compiler yet. The hardware will not exploit the UB and we are A-OK as long
+/// as LLVM won't look through __atomic_exchange. By the time it does we
+/// hopefully have source level `freeze` intrinsics.
+static uint32_t AssertionFlag [[clang::loader_uninitialized]];
+
 extern "C" {
 void __assert_assume(bool cond, const char *exp, const char *file, int line) {
-  if (!cond && config::isDebugMode(config::DebugKind::Assertion)) {
-    PRINTF("ASSERTION failed: %s at %s, line %d\n", exp, file, line);
+  if (config::isDebugMode(config::DebugKind::Assertion) && !cond) {
+    uint32_t exp_low, exp_high;
+    utils::unpack(uint64_t(exp), exp_low, exp_high);
+    uint32_t file_low, file_high;
+    utils::unpack(uint64_t(file), file_low, file_high);
+    uint32_t hash = (exp_low << 1) ^ (exp_high << 3) ^ (file_low << 7) ^
+                    (file_high << 11) ^ (line);
+    if (hash != atomic::exchange(&AssertionFlag, hash, __ATOMIC_SEQ_CST))
+      PRINTF("ASSERTION failed: %s at %s, line %d\n", exp, file, line);
     __builtin_trap();
   }
 
diff --git a/openmp/libomptarget/DeviceRTL/src/Synchronization.cpp b/openmp/libomptarget/DeviceRTL/src/Synchronization.cpp
--- a/openmp/libomptarget/DeviceRTL/src/Synchronization.cpp
+++ b/openmp/libomptarget/DeviceRTL/src/Synchronization.cpp
@@ -262,6 +262,10 @@
   return impl::atomicAdd(Addr, V, Ordering);
 }
 
+uint32_t atomic::exchange(uint32_t *Addr, uint32_t V, int Ordering) {
+  return impl::atomicExchange(Addr, V, Ordering);
+}
+
 extern "C" {
 void __kmpc_ordered(IdentTy *Loc, int32_t TId) {}