http://gcc.gnu.org/bugzilla/show_bug.cgi?id=55131
mawenqi <mawenqi108 at gmail dot com> changed:
What |Removed |Added
----------------------------------------------------------------------------
Status|RESOLVED |UNCONFIRMED
Component|inline-asm |c++
Version|4.7.0 |unknown
Resolution|INVALID |
Severity|normal |blocker
--- Comment #2 from mawenqi <mawenqi108 at gmail dot com> 2012-10-30 06:22:07
UTC ---
(In reply to comment #1)
> This is not a bug.
> The produced assembly looks like:
> movl 8(%ebp), %edi # %1
> movl 12(%ebp), %esi# %2
> movl 0(%esi), %eax
> movl 4(%esi), %edx
> movl (%ecx), %ebx# %3
> movl (%eax), %ecx# %4
>
> By the time the last statement happens, eax has already been clobbered. You
> never said you are clobber eax in the inline-asm so it chose the 4th operand
> as
> being eax. You were getting lucky in 3.4.6 with the inline-asm really,
>
>
> I don't see why you don't use the __sync_* (or even better the __atomic_*)
> builtins for doing the compare and swap?
Thanks a lot for your help!
This is the old legacy code. After replaced original implementation with
buildin function __atomic_compare_exchange_n, now everything is fine!
static inline bool MyAtomic_CAS64(volatile unsigned long long* tgt,
unsigned long long* old,
unsigned long long rep)
{
return __atomic_compare_exchange_n(tgt, old, rep,
false, __ATOMIC_SEQ_CST, __ATOMIC_SEQ_CST);
}
Thanks again!
