[tkreuzer] 53151: [INTRIN] memory barriers - implement a fixed version of __invlpg for MSVC, since the intrinsic is broken and can generate wrong instructions, when optimization is enabled - add MS...

Author: tkreuzer
Date: Tue Aug  9 14:22:29 2011
New Revision: 53151

URL: http://svn.reactos.org/svn/reactos?rev=53151&view=rev
Log:
[INTRIN]
memory barriers
- implement a fixed version of __invlpg for MSVC, since the intrinsic is broken and can
generate wrong instructions, when optimization is enabled
- add MSVC versions of _mm_mfence, _mm_lfence, _mm_sfence and __faststorefence
- add _mm_mfence() for gcc
- give _mm_lfence, _mm_sfence, __invlpg and __wbinvd proper 

Modified:
    trunk/reactos/include/crt/mingw32/intrin_x86.h
    trunk/reactos/include/crt/msc/intrin.h

Modified: trunk/reactos/include/crt/mingw32/intrin_x86.h
URL:
http://svn.reactos.org/svn/reactos/trunk/reactos/include/crt/mingw32/intrin…
==============================================================================
--- trunk/reactos/include/crt/mingw32/intrin_x86.h [iso-8859-1] (original)
+++ trunk/reactos/include/crt/mingw32/intrin_x86.h [iso-8859-1] Tue Aug  9 14:22:29 2011
@@ -81,24 +81,6 @@
 
 /*** Memory barriers ***/
 
-#ifdef _x86_64
-__INTRIN_INLINE void __faststorefence(void)
-{
-    long local;
-	__asm__ __volatile__("lock; orl $0, %0;" : : "m"(local));
-}
-#endif
-
-__INTRIN_INLINE void _mm_lfence(void)
-{
-	__asm__ __volatile__("lfence");
-}
-
-__INTRIN_INLINE void _mm_sfence(void)
-{
-	__asm__ __volatile__("sfence");
-}
-
 __INTRIN_INLINE void _ReadWriteBarrier(void)
 {
 	__asm__ __volatile__("" : : : "memory");
@@ -107,6 +89,34 @@
 /* GCC only supports full barriers */
 #define _ReadBarrier _ReadWriteBarrier
 #define _WriteBarrier _ReadWriteBarrier
+
+__INTRIN_INLINE void _mm_mfence(void)
+{
+	__asm__ __volatile__("mfence" : : : "memory");
+}
+
+__INTRIN_INLINE void _mm_lfence(void)
+{
+	_ReadBarrier();
+	__asm__ __volatile__("lfence");
+	_ReadBarrier();
+}
+
+__INTRIN_INLINE void _mm_sfence(void)
+{
+	_WriteBarrier();
+	__asm__ __volatile__("sfence");
+	_WriteBarrier();
+}
+
+#ifdef _x86_64
+__INTRIN_INLINE void __faststorefence(void)
+{
+    long local;
+	__asm__ __volatile__("lock; orl $0, %0;" : : "m"(local));
+}
+#endif
+
 
 /*** Atomic operations ***/
 
@@ -1438,7 +1448,7 @@
 
 __INTRIN_INLINE void __invlpg(void * const Address)
 {
-	__asm__("invlpg %[Address]" : : [Address] "m" (*((unsigned char
*)(Address))));
+	__asm__("invlpg %[Address]" : : [Address] "m" (*((unsigned char
*)(Address))) : "memory");
 }
 
 
@@ -1482,7 +1492,7 @@
 
 __INTRIN_INLINE void __wbinvd(void)
 {
-	__asm__ __volatile__("wbinvd");
+	__asm__ __volatile__("wbinvd" : : : "memory");
 }
 
 __INTRIN_INLINE void __lidt(void *Source)

Modified: trunk/reactos/include/crt/msc/intrin.h
URL:
http://svn.reactos.org/svn/reactos/trunk/reactos/include/crt/msc/intrin.h?r…
==============================================================================
--- trunk/reactos/include/crt/msc/intrin.h [iso-8859-1] (original)
+++ trunk/reactos/include/crt/msc/intrin.h [iso-8859-1] Tue Aug  9 14:22:29 2011
@@ -13,14 +13,25 @@
 unsigned int __getcallerseflags(void);
 #pragma intrinsic(__getcallerseflags)
 
-/*** Atomic operations ***/
+/*** Memory barriers ***/
 void _ReadWriteBarrier(void);
 #pragma intrinsic(_ReadWriteBarrier)
 void _ReadBarrier(void);
 #pragma intrinsic(_ReadBarrier)
 void _WriteBarrier(void);
 #pragma intrinsic(_WriteBarrier)
-
+void _mm_mfence(void);
+#pragma intrinsic(_mm_mfence)
+void _mm_lfence(void);
+#pragma intrinsic(_mm_lfence)
+void _mm_sfence(void);
+#pragma intrinsic(_mm_sfence)
+#ifdef _M_AMD64
+void __faststorefence(void);
+#pragma intrinsic(__faststorefence)
+#endif
+
+/*** Atomic operations ***/
 long _InterlockedCompareExchange(volatile long * const Destination, const long Exchange,
const long Comperand);
 #pragma intrinsic(_InterlockedCompareExchange)
 long _InterlockedExchange(volatile long * const Target, const long Value);
@@ -297,6 +308,23 @@
 void __invlpg(void * const Address);
 #pragma intrinsic(__invlpg)
 
+// This intrinsic is broken and generates wrong opcodes,
+// when optimization is enabled!
+#pragma warning(push)
+#pragma warning(disable:4711)
+void  __forceinline __invlpg_fixed(void * const Address)
+{
+    _ReadWriteBarrier();
+   __asm
+   {
+       mov eax, Address
+       invlpg [eax]
+   }
+    _ReadWriteBarrier();
+}
+#pragma warning(pop)
+#define __invlpg __invlpg_fixed
+
 /*** System operations ***/
 unsigned __int64 __readmsr(const int reg);
 #pragma intrinsic(__readmsr)