debug mode performance patch

[Jonathan Wakely]

On 16 November 2010 21:14, François Dumont wrote:
>
>    Here is a new proposition with the deadlock issue fixed. I use auto_ptr
> to acquire it even if it is deprecated because unique_ptr is only available
> in gnu++0x mode and the library is not built in this mode, is it fine ? I
> also created a test to show the problem before the fix. In fact I even had
> to add a sleep between the 2 locks in debug.cc to reproduce the deadlock
> temporarily, it shows that the problem existed and that it has been removed
> now.

Great, thanks for checking that.

I'm not keen on the use of auto_ptr here, not because of auto_ptr
itself, I just don't think it's necessary to allocated the
scoped_locks on the heap - that adds back some of the contention
you're trying to remove!  Could you not move the body of the swap
function into a separate function and call it after locking either one
or two mutexes?  e.g. the do_swap function in my mail on November 7th.