Re: [Patch] mt_allocator: spare mem & fix alignment problems

[Loren James Rittle <rittle at latour dot rsch dot comm dot mot dot com>]

Benjamin Kosnik <bkoz@redhat.com> wrote:

>> Well, I'm happy to report (after all basic_string is my (old) baby ;) 
>> that with std::string it does *not* fail...

> 'tis as I feared.

>> Is it possible that actually rope is not as thread-safe as was supposed 
>> to be??

When I last looked at rope (to fix a PR), I noted that it contained
memory leaks in MT cases under our current configuration.  I have not
looked at the current failure yet.

> ...rope certainly seems implicated. It is strange, however, that this
> only fails with __mt_alloc and not with new_allocator or __pool_alloc.

It is possible that there were races that are more likely to appear
with the new allocator...  Even if we remove rope, it would be nice to
understand the failure (to be sure it isn't the allocator itself).

> We need to deprecate rope. We have no evidence that it's being used, and

Here is the nicest thing I can say: rope is one of those cool things
that was never standardized.  The implementation we have was designed
to work best with GC.  That is fine but we don't do GC in the library
implementation.

There was support in principle to keep it last time the issue was
raised.  All I can say is that I will commit to work on failing test
cases (as time permits ;).  It really needs a re-write (at the least
to remove the crusty configuration code) but I can't commit to that.

Regards,
Loren