This is the mail archive of the
gcc-bugs@gcc.gnu.org
mailing list for the GCC project.
[Bug libgcc/71744] Concurrently throwing exceptions is not scalable
- From: "torvald at gcc dot gnu.org" <gcc-bugzilla at gcc dot gnu dot org>
- To: gcc-bugs at gcc dot gnu dot org
- Date: Wed, 14 Sep 2016 10:19:50 +0000
- Subject: [Bug libgcc/71744] Concurrently throwing exceptions is not scalable
- Auto-submitted: auto-generated
- References: <bug-71744-4@http.gcc.gnu.org/bugzilla/>
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=71744
--- Comment #15 from torvald at gcc dot gnu.org ---
(In reply to Jakub Jelinek from comment #14)
> Looking at that patchset, I just want to say that the use of TLS in libgcc
> is very much undesirable, it affects every program even when not using
> exceptions (significant portion of them), especially when it uses that much
> (6 * sizeof(void *) * 8 + sizeof(void *) + 2 * 8 = 408 bytes per thread is
> significant, plus it uses the slow local dynamic TLS model.
I agree.
> Similarly, the 64 recursive locks in libc, again, significant amount of
> memory
> per process that doesn't care about exceptions,
That might be reducable if we build custom locks and don't use pthread ones,
but we'll at least have 64b per lock.
> and especially signficiant
> slowdown on dlopen/dlclose. While some programs abuse exceptions for normal
> control flow, others abuse dlopen/dlclose.
>
> A rwlock ought to be better, but I don't think we have recursive rwlock in
> libc
rdlock can be recursive, rwlock cannot. It is something we could add with not
much effort in a custom rwlock, I think.
> yet (all we need is being able to recursively lock it for writing, for
> reading
> we might not need to allow recursion in this case).
Not having to deal with recursive rdlock can make things easier for the rwlock
implementation.
> Plus the current rwlock
> implementation still takes uses a per-process lock to guard the inner
> operations
> on the rwlock, so while rwlock helps with longer critical sections, for very
> short ones it likely doesn't scale all that well.
This will change as soon as my new rwlock is upstream. Especially short rdlock
sections will be much faster.