Bazaar

bzr.dev does not support pyrex 0.9.4.1

Bug #276868 reported by GuilhemBichot on 2008-10-01

Affects		Status	Importance	Assigned to	Milestone
	Bazaar	Fix Released	Medium	John A Meinel	Bazaar 1.8

Bug Description

Linux 32 bit gcc 4.1.2 (opensuse 10.2).
Must have been introduced in the past weeks. When I do a fresh branch of bzr.dev and execute "make", I see many compiler warnings in _btree_serializer_c.c bzrlib/_knit_load_data_c.c bzrlib/_dirstate_helpers_c.c bzrlib/_readdir_pyx.c (attached). The generated bzr segfaults on simple operations like:
bzr init bzr1 # initial branch
cd bzr1
cat >foo <<EOF
1
2
3
4
5
6
EOF
bzr add foo
bzr commit -m "A"
bzr branch . ../bzr2 # segfault here
After upgrading pyrex (I was using opensuse 10.2's version) to the latest one, warnings and segfault go away.

Revision history for this message

GuilhemBichot (guilhem-bichot) wrote on 2008-10-01:

build log (you can see compiler warnings) Edit (19.4 KiB, text/plain)

Revision history for this message

John A Meinel (jameinel) wrote on 2008-10-01:

We tracked it down to revision 3739 on bzr.dev which happens to be:

3739 Canonical.com Patch Queue Manager 2008-09-26 [merge]
(robertc) Create a pyrex optimised iter_changes for dirstate trees.
(Robert Collins)

So it seems something in the new iter_changes code causes pyrex 0.9.4.1 to generate code that will cause a segfault.

Revision history for this message

John A Meinel (jameinel) wrote on 2008-10-01:

Trying to debug the pyrex output, it appears that it sometimes does:

PY_XDECREF(pointer); pointer = 0;

and then later does

PY_DECREF(pointer).

So it seems we are just triggering a bug in pyrex 0.9.4.1. Further, it seems to only happen because the functions are long and complex enough that it re-uses some of its temporary variables.

There doesn't seem to be an easy way to work around this, so instead we'll just blacklist that version of the compiler. It won't effect released versions, because we always bundle the .c files anyway.