RE_NO_SUB thread-safety (was Re: A little more regex.h pedantry)

[Paolo Bonzini <bonzini at gnu dot org>]

On 07/31/2010 12:52 AM, Eric Blake wrote:
> > >  The code in re_compile_pattern says:
> > >
> > >     /* And GNU code determines whether or not to get register information
> > >        by passing null for the REGS argument to re_match, etc., not by
> > >        setting no_sub, unless RE_NO_SUB is set.  */
> > >     bufp->no_sub = !!(re_syntax_options & RE_NO_SUB);
> > >
> > >  It doesn't seem to be conditional on grouping.
>
> Aha.  The answer, then, is that you used to be able to supply no_sub
> prior to compilation; but nowadays you have the flag RE_NO_SUB to do the
> job instead.  So I agree with documenting that no_sub is private (by not
> listing it as public).

This is unfortunate because it is not thread safe.  I never noticed this 
problem with the GNU API, but it is quite bad.

Glibc people, would you agree with adding a 
re_compile_pattern_with_syntax function (which would likely be a good 
idea anyway)?

Paolo