Re: [PATCH 07/10] regex: fix longstanding backref match bug

[Dmitry V. Levin]

On Fri, Feb 05, 2021 at 05:25:59PM -0800, Paul Eggert wrote:
> This fixes a longstanding glibc bug concerning backreferences
> <https://sourceware.org/11053> (2009-12-04).

Nit: this short URL doesn't work, while the longer one does:
https://sourceware.org/bugzilla/show_bug.cgi?id=11053

> * lib/regexec.c (proceed_next_node, push_fail_stack)
> (pop_fail_stack): Push and pop the previous registers
> as well as the current ones.  All callers changed.
> (set_regs): Also pop if CUR_NODE has already been checked,
> so that it does not get added as a duplicate set entry.
> (update_regs): Fix comment location.
> * tests/test-regex.c (tests): New constant.
> (bug_regex11): New test function.
> (main): Bump alarm value.  Call new test function.
> ---
>  ChangeLog          |  13 ++++++
>  lib/regexec.c      |  26 +++++++----
>  tests/test-regex.c | 113 ++++++++++++++++++++++++++++++++++++++++++++-
>  3 files changed, 141 insertions(+), 11 deletions(-)

Apparently, this commit changes behaviour of GNU sed in a way that looks
like regression.

When GNU sed is built with this gnulib commit (v0.1-4426-g70b673eb7):
$ echo 123 |sed -r 's/^1*+(.)/\1/'
3
$ echo 123 |sed -r 's/^(1*)*(.)/\2/'
3

When GNU sed is built with the previous gnulib commit (v0.1-4425-gf5596242f):
$ echo 123 |sed -r 's/^1*+(.)/\1/'
23
$ echo 123 |sed -r 's/^(1*)*(.)/\2/'
23

There are regexps in the wild containing "*+" that used to work before,
see e.g.
https://lore.kernel.org/lkml/20210414182723.1670663-1-vt@altlinux.org/T/#u


-- 
ldv