Mandating that the compiler cannot elide the copy to a local

variable, in case there is unlocked concurrent access to the
source tensor.  No noticeable speed changes for scatter
(it should not affect the asm output in any case I can
think of;  it's designed to prevent pathological corner cases.)
Change: 116181195

1 files changed, 13 insertions, 0 deletions

diff --git a/tensorflow/core/kernels/bounds_check.h b/tensorflow/core/kernels/bounds_check.h
index 286ef8959c..805f698f11 100644
--- a/tensorflow/core/kernels/bounds_check.h
+++ b/tensorflow/core/kernels/bounds_check.h
@@ -33,6 +33,19 @@ EIGEN_ALWAYS_INLINE bool FastBoundsCheck(Index index, Index limit) {
                          static_cast<UIndex>(limit));
 }
 
+namespace internal {
+// Ensure that the compiler cannot elide a copy into a local, for
+// bounds checking on source tensors that might be updated asynchronously.
+// This function may only be used on primitive integral types (int32, int64,
+// etc).  It does not guarantee any atomicity or barriers.
+template <typename T>
+const T SubtleMustCopy(const T &x) {
+  static_assert(std::is_integral<T>::value,
+                "must_copy can only be used on integer types.");
+  auto *to_x = reinterpret_cast<const volatile T *>(&x);
+  return *to_x;
+}
+}  // namespace tensorflow::internal
 }  // namespace tensorflow
 
 #endif  // TENSORFLOW_UTIL_BOUNDS_CHECK_H_