The problem was that the x86 assembly based atomic CAS
(compare and swap) implementation could copy the wrong
value to the ebx register, where the cmpxchg8b expects
to see part of the "comparand" value. Since the original
value in the ebx register is saved in the stack (that is,
the push instruction causes the stack pointer to change),
a wrong offset could be used if the compiler decides to
put the source of the comparand value in the stack.

The solution is to copy the comparand value directly from
memory. Since the comparand value is 64-bits wide, it is
copied in two steps over to the ebx and ecx registers.

include/atomic/x86-gcc.h:
  For reference, an excerpt from a faulty binary follows.
  
  It is a disassembly of my_atomic-t, compiled at -O3 with
  ICC 11.0. Most of the code deals with preparations for
  a atomic cmpxchg8b operation. This instruction compares
  the value in edx:eax with the destination operand. If the
  values are equal, the value in ecx:ebx is stored in the
  destination, otherwise the value in the destination operand
  is copied into edx:eax.
  
  In this case, my_atomic_add64 is implemented as a compare
  and exchange. The addition is done over temporary storage
  and loaded into the destination if the original term value
  is still valid.
  
    volatile int64 a64;
    int64 b=0x1000200030004000LL;
    a64=0;
        mov    0xfffffda8(%ebx),%eax
        xor    %ebp,%ebp
        mov    %ebp,(%eax)
        mov    %ebp,0x4(%eax)
    my_atomic_add64(&a64, b);
        mov    0xfffffda8(%ebx),%ebp      # Load address of a64
        mov    0x0(%ebp),%edx             # Copy value
        mov    0x4(%ebp),%ecx
        mov    %edx,0xc(%esp)             # Assign to tmp var in the stack
        mov    %ecx,0x10(%esp)
        add    $0x30004000,%edx           # Sum values
        adc    $0x10002000,%ecx
        mov    %edx,0x8(%esp)             # Save part of result for later
        mov    0x0(%ebp),%esi             # Copy value of a64 again
        mov    0x4(%ebp),%edi
        mov    0xc(%esp),%eax             # Load the value of a64 used
        mov    0x10(%esp),%edx            # for comparison
        mov    %esi,(%esp)
        mov    %edi,0x4(%esp)
        push   %ebx                       # Push %ebx into stack. Changes esp.
        mov    0x8(%esp),%ebx             # Wrong restore of the result.
        lock cmpxchg8b 0x0(%ebp)
        sete   %cl
        pop    %ebx