Re: [RFC/PATCH] ignore memcmp() overreading in bsearch() callback

From: Jeff King <peff@peff.net>
Date: Jan 15, 2013, 17:18 UTC
Message-ID: <20130115171844.GA32227@sigill.intra.peff.net>
In-Reply-To: <7vmwwa4c8r.fsf@alter.siamese.dyndns.org>

On Tue, Jan 15, 2013 at 08:55:32AM -0800, Junio C Hamano wrote:

Show 17 quoted lines

> Jeff King <peff@peff.net> writes:
> 
> > On Mon, Jan 14, 2013 at 03:36:21PM -0800, Junio C Hamano wrote:
> >
> >> It appears that memcmp() uses the usual "one word at a time"
> >> comparison and triggers valgrind in a callback of bsearch() used in
> >> the refname search.  I can easily trigger problems in any script
> >> with test_commit (e.g. "sh t0101-at-syntax.sh --valgrind -i -v")
> >> without this suppression.
> >
> > Out of curiosity, what platform do you see this on? I can't reproduce on
> > glibc.
> 
>     Debian GNU/Linux 6.0.6 (squeeze), on Linux 2.6.32-5-amd64.
>     libc-bin              2.11.3-4
>     valgrind-3.6.0.SVN-Debian
>     gcc                   4:4.4.5-1

Interesting. I can reproduce easily on my squeeze machine, but not my wheezy. So presumably it is a false positive fixed either in libc (I have 2.13-38 on the "good" box) or valgrind (1:3.8.1-1).

However, the error that valgrind reports is on the call to "strlen(ent->name)", not memcmp (but it has suffered from the same SSE issues in the past).

So I feel pretty confident that it really is a false positive; you may want to double-check the offending call for the commit message (though I would not be surprised if it is triggerable from both). I think it also means that René's suggestion to use strncmp cannot be relied on to silence the warning (though I am not opposed to doing it anyway if we think it is more clear).

-Peff