Hi Siddhesh,

It seems the glibc manual doesn't cover REG_STARTEND.  We're having
some discussion about it, while we're improving the manual page, and
we have some slight disagreement about what are the constraints for
nmatch and pmatch when REG_STARTEND is specified.

Since this is a BSD extension, not in POSIX, we can't refer to the
standard.  And since glibc doesn't document it, I can't know for
sure your intention.

I'm not sure how much we should rely on implementation details, and
if those will hold forever, or _FORTIFY_SOURCE and the like are
allowed to constrain some current details.

The discussion is in this thread, if you want to read it entirely:
<https://lore.kernel.org/linux-man/09bda4f5-ab6f-a369-ad36-0779dc3749b4@gmail.com/T/#m7dd7771072df3be97bc831f581dc7bfea684197d>

The question is:

If REG_STARTEND was specified _and_ REG_NOSUB was also specified,
regexec(3) will read from pmatch[0] and will effectively ignore
nmatch (the number of elements in pmatch[]), since it won't write
to pmatch[] at all.

Is the following call valid, or is it UB?

    regmatch_t  pmatch = {
        .rm_so = string,
        .rm_eo = string + 42,  // Assume this offset is valid
    };

    regexec(preg, string, 0, pmatch, REG_NOSUB | REG_STARTEND);

How about this?

    regexec(preg, string, 999, pmatch, REG_NOSUB | REG_STARTEND);


Current implementations will work, because nmatch is effectively
ignored.  But is it intended to be this way, or just an implementation
detail?

Thanks!

Alex

-- 
<http://www.alejandro-colomar.es/>
GPG key fingerprint: A9348594CE31283A826FBDD8D57633D441E25BB5