[Bug libstdc++/22309] mt allocator doesn't pthread_key_delete it's keys

["bkoz at gcc dot gnu dot org" <gcc-bugzilla@gcc.gnu.org>]

------- Additional Comments From bkoz at gcc dot gnu dot org  2005-09-09 09:19 -------

Hey Jakub. Yeah, I think this can be back-ported. I put in my patch, which looks
pretty good on x86/linux. We could proably do something more elaborate to not
duplicate some of the symbols but I'm feeling lazy... this is good enough.

Sorry about the confusion about the static freelist_mutex. Yes, this was what I
was getting at by "linkage clarity" ... I think we are on the same page. Yes,
this is a good idea, and I'll clean this up too as a follow-on patch.

Umm... so, no, I don't get what you are getting at WRT destructors in
libstdc++.so. I think you mean functions in testsuite_shared.so that are asm
destrutors but that use __mt_alloc to do allocations or something. You mean,
what then? Maybe we'd both be better off if you came up with an example...

I think just static ordering in this file will work. Won't that just work? Man,
I want that to work. If not, then the init priorty stuff can be re-added. The
thing is, there are no config tests for init priority...  ..... I thought
support of that was kind of dodgy, non-ELF... if this feature is going to be
used, there are other things with __ioinit that can also be done. 

Anyway. Of more immediate concern is that the testsuite dejagnu hacking is quite
weak. For instance, The testsuite file should just be run on linux, so I think
that is ok. But the shared object is unconditionally compiled, even for
--disable-shared builds.... also, compiled with a complete hack for flags,
likely to only work on linux... It is kind of funny, just to be re-creating an
autoconf way of doing this, but doing it in dejagnu. Ugh. Some of this stuff
would be easier if we could use Make to build instead of dejagnu.... 


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=22309