Subj : Re: A question about atomic_ptr
To   : comp.programming.threads
From : Peter Dimov
Date : Tue Apr 19 2005 03:26 am

Chris Thomasson wrote:
>> mfence fails with exception C000001E, "an attempt was made to
>> execute an invalid lock sequence". Athlon XP 2800+.
>>
>>
>
> Try replacing the inline assembly with this:
>
>
> __asm__ __volatile__ ("lock; addl $0, 0(%%esp)" : : : "memory");
>
>
> __asm__ __volatile__ ("mfence" : : : "memory");

This won't work as I don't compile with GCC, but if you are concerned about 
compiler reordering, this is not an issue. I've looked at the disassembly 
and it's correct. The mfence instruction just causes an exception.

The program that I'm running is at the end. Here's the relevant part:

0040103E  mov         eax,1DCD6500h
  {
   add dword ptr [esp], 0
00401043  add         dword ptr [esp],0
   mfence
00401047  mfence

 for( int i = 0; i < n; ++i )
 {
  __asm
0040104A  dec         eax
0040104B  jne         main+43h (401043h)
  }
 }

and here's the exception:

Unhandled exception at 0x00401047 in testbed.exe: 0xC000001E: } An attempt 
was made to execute an invalid lock sequence.

Complete program:

#include <windows.h>
#include <stdio.h>

int main()
{
    int const n = 500000000;

    timeBeginPeriod( 1 );

    DWORD t = timeGetTime();

    for( int i = 0; i < n; ++i )
    {
        __asm
        {
            lock add dword ptr [esp], 0
        }
    }

    printf( "lock prefix time: %d\n", timeGetTime() - t );

    t = timeGetTime();

    for( int i = 0; i < n; ++i )
    {
        __asm
        {
            add dword ptr [esp], 0
            mfence
        }
    }

    printf( "mfence time: %d\n", timeGetTime() - t );

    timeEndPeriod( 1 );
}

.