Troubles with nptl on Geode

[Pirmin Walthert]

Hello

I'm the developer of a system we use as CPE for our triple-play
customers. It's based on an ALIX board and since 2 years we are using
uClibc instead of glibc. After some troubles with callweaver (an
asterisk fork) and linuxthreads.old our distribution is almost
"uncrasheable" since 2 years (using the 2.6.27 kernel series until
december and 2.6.32 since then) after we switched to the new
linuxthreads implementation. (We also made a stable arm-be port).

In May I started playing with the git-branch and started using nptl. I
was very surprised that the system was working stable from the first
moment on (congratulations to the developers!!!). Yesterday I tried to
update to the newest version and callweaver started to crash on every
startup!

I've compiled the whole distribution from scratch with several versions
now and it seems like version 8b48f745be1e086d6e486bbb8167e770f3d1fbc5
(commit: Choose the correct version for setting up errno) still works
perfectly while the troubles start to occur with
c06619d7ab2ebbbc93427ca6a49dcc8d479db549: "nptl: simplify buildsys".

Note that I only looked at the nptl-directory-revisions and didn't
checkout every version between these commits (however
c06619d7ab2ebbbc93427ca6a49dcc8d479db549 ist the first commit after
8b48f745be1e086d6e486bbb8167e770f3d1fbc5 in the nptl directory).

I'm cross-compiling from a x86-64 system to i586 (TARGET_ARCH="i386" and
CONFIG_586MMX=y), what should work with AMD Geode LX (and always did in
the past).

My build-process is: linux-headers, gmp, mpfr, binutils, static gcc
(because uclibc-nptl needs a minimal cross-compiler for generating the
headers), uclibc-headers, static gcc agin (don't know if this is
necessary), uclibc, gcc, busybox, ...

Versions: Kernel 2.6.32.17, gcc-4.3.3, binutils 2.19.1, busybox 1.17.1

Do you have an idea what I'm doing wrong or where the bug could be? I'm
pretty sure that the build-process is clean and is not using wrong
headers or similar as even the arm-be cross-compilation succeeded (and
worked) one year ago, using the same method.

Best regards,

Pirmin Walthert