Re: [Patch (preview)] libstdc++/24469

[Paolo Carlini <pcarlini at suse dot de>]

Benjamin Kosnik wrote:

> I like this idea, you two! Thanks for updating the comments as well. 
>
> More elaborate documentation changes would of course be welcome.
>  
Yes, I will, ASAP.

> > Yes, I think it is, we are doing much worse things about alignment in 
> > the current basic_string... ;) Also, note that, for now, at least, we 
> > are not suggesting mt_allocator for all the possible targets in the 
> > world, but essentially for gnu/linux and a bit more (the are nasty 
> > issues about static ordering, things like that)
> >    
> Hmm.
>
> Another way to do this is to take v7 mt_allocator, add these patches,
> and then rename to something like __power2_allocator (suggestions
> welcome) and check it in as a new file, and then add deprecated
> warnings to __mt_alloc headers.
Yes, eventually we could even start on a new allocator, adding 
clean-ups, optimizations, and so on. Note however, that some of those 
nasty issues which led us to enable (temporarily) mt_alloc only for some 
targets (gnu/linux, essentially) I have no idea how to fix ;) A nice 
thing of the fix that we have is that it will avoid races *also* to 
existing binaries compiled vs the current mt_alloc!

> > By the way, updated version of our work forthcoming...
> >    
> Ok. I'm very hardware constrained at the moment so will only be of use
> sanity checking the single-cpu performance.
>  
Thanks for the offer, but for single-cpu, everthing is really, really 
mooth, just have a look to the files I posted yesterday. Anyway, even if 
we see a small performance impact (I don't think I seeing a measurable 
one anywhere), we are also fixing a *serious* correctness issue: a race 
on an allocator targeted to mt applications, heem, how embarassing ;)

Paolo.