Use the intrinsic version of the Interlocked* functions.

base/atomicops_internals_x86_msvc.h

Index: base/atomicops_internals_x86_msvc.h
diff --git a/base/atomicops_internals_x86_msvc.h b/base/atomicops_internals_x86_msvc.h
index 016744c144416f869606881e87ee7d8072f75cdd..1d326d1fbcf2364ba23302ae7809ec6cd95d788f 100644
--- a/base/atomicops_internals_x86_msvc.h
+++ b/base/atomicops_internals_x86_msvc.h
@@ -9,6 +9,8 @@
 
 #include <windows.h>
 
+#include <intrin.h>
+
 #include "base/macros.h"
 
 #if defined(ARCH_CPU_64_BITS)
@@ -20,13 +22,17 @@
 #undef MemoryBarrier
 #endif
 
+#pragma intrinsic(_InterlockedCompareExchange)

[Sigurður Ásgeirsson, 2014/06/02 19:43:08]

I'm concerned that you're setting this for all clients of this file. Maybe
that's benign, but I just don't know. Another thing that bugs me is that
winbase.h seems to define these functions to their intrinsic counterparts. Why
isn't that hitting for these invocations?

[Sébastien Marchand, 2014/06/03 14:39:48]

winbase.h seems to only define these functions to their intrinsic counterpart in
some cases:
- if defined(_M_IA64) && !defined(RC_INVOKED)
- if defined(_M_ARM) && !defined(RC_INVOKED)
- if defined(_M_AMD64) && !defined(RC_INVOKED)

+#pragma intrinsic(_InterlockedExchange)
+#pragma intrinsic(_InterlockedExchangeAdd)
+
 namespace base {
 namespace subtle {
 
 inline Atomic32 NoBarrier_CompareAndSwap(volatile Atomic32* ptr,
                                          Atomic32 old_value,
                                          Atomic32 new_value) {
-  LONG result = InterlockedCompareExchange(
+  LONG result = _InterlockedCompareExchange(
       reinterpret_cast<volatile LONG*>(ptr),
       static_cast<LONG>(new_value),
       static_cast<LONG>(old_value));
@@ -35,7 +41,7 @@ inline Atomic32 NoBarrier_CompareAndSwap(volatile Atomic32* ptr,
 
 inline Atomic32 NoBarrier_AtomicExchange(volatile Atomic32* ptr,
                                          Atomic32 new_value) {
-  LONG result = InterlockedExchange(
+  LONG result = _InterlockedExchange(
       reinterpret_cast<volatile LONG*>(ptr),
       static_cast<LONG>(new_value));
   return static_cast<Atomic32>(result);
@@ -43,7 +49,7 @@ inline Atomic32 NoBarrier_AtomicExchange(volatile Atomic32* ptr,
 
 inline Atomic32 Barrier_AtomicIncrement(volatile Atomic32* ptr,
                                         Atomic32 increment) {
-  return InterlockedExchangeAdd(
+  return _InterlockedExchangeAdd(
       reinterpret_cast<volatile LONG*>(ptr),
       static_cast<LONG>(increment)) + increment;
 }