The __tls_get_addr() routine is called by the compiler for __thread
accesses (barring optimizations, IE mode, etc). But I'm confused by how
the "generation" code flow is supposed to work.
The dlopen() and dlclose() routines, for example, both increment the
global _dl_tls_generation counter after
On 12/09/2009 11:09 AM, Chris Metcalf wrote:
My proposed fix is something like the following, in dl-tls.c. (I'm
running code patched up from the old SVN NPTL branch, but looking at the
latest GIT NPTL it seems pretty similar here.) Obviously a real patch
would re-indent all the code contained w
-BEGIN PGP SIGNED MESSAGE-
Hash: SHA1
Austin Foxley wrote:
> On 12/09/2009 11:09 AM, Chris Metcalf wrote:
>> My proposed fix is something like the following, in dl-tls.c. (I'm
>> running code patched up from the old SVN NPTL branch, but looking at the
>> latest GIT NPTL it seems pretty si
On 12/15/2009 9:07 AM, Carmelo AMOROSO wrote:
> Austin Foxley wrote:
>
>> On 12/09/2009 11:09 AM, Chris Metcalf wrote:
>>
>>> My proposed fix is something like the following, in dl-tls.c. (I'm
>>> running code patched up from the old SVN NPTL branch, but looking at the
>>> latest GIT NPTL
In my email quoted below, I found that __tls_get_addr() under the uClibc
version we were running with was always running the slow path. We were
running an old version from the nptl branch, back at change 22990 on svn.
Since then, I've ported glibc-2.11.2 to tile. I looked at the performance
of
