Memory barriers vs lock/unlock

[Paolo Carlini]

Peter Dimov wrote:

> Have you tried calling _S_initialize before spawning the threads? If
> the problem disappears, then _S_initialize isn't thread safe. If it
> persists... I'm out of ideas :-)

Thanks Peter. I'll try to return to the issue before the end of the week
(right now hard at work both on the unequal allocators thing and deque).

Actually, I'm not positive we can fix that bug at the library level: the
library in 4_0-branch is very, very similar in the sensitive areas
(locale, iostreams, string, everywhere we pay attention to MT) and
*never* fails on all the multiprocessor ia64 machines I tried. On the
other hand, the mainline compiler has got the the new ia64-style atomic
builtins machinery and exactly around that date (mid April), consistenly
2, often 4, library testcases stressing MT started consistently failing...

Paolo.