Re: have you ever mistyped [[:lower:]] as [:lower:] ?

[Paolo Bonzini]

On 09/01/2010 10:11 AM, Jim Meyering wrote:
> No, that's not the same at all.  grep must not do that.
> It is one thing to fail for an obviously erroneous construct.
> It would be worse to silently transform it into the "intended"
> one, since then GNU grep would silently work as intended, but that
> same erroneous command would produce different results with non-GNU grep.
>
> Here, we're making it clear that this is a serious error.

It's not, as it is syntactically correct.

> > Second, if this was done, it should operate in the same way in sed,
> > expr, awk, and all other GNU programs that deal with regexes.  (And
> > possibly in glibc too).
>
> It would be nice to make other GNU programs provide the same new
> feature.  However, if they don't (or don't right away), it's not
> a big deal.

I know that sed won't as long as I'm maintainer.

> > If you want to add --warn=error (which is a "superset" of
> > --warn=always behavior), that's fine and I actually like the idea.
> > But I think making it the default non-POSIXLY_CORRECT behavior is
> > wrong.  Honestly, if this happened I would regret having introduced
> > the feature in the first place.
>
> I hope you don't regret it.
> Sometimes you just have to admit that
> POSIX-is-clear-and-POSIX-can-be-improved.

POSIX can be improved in many ways, and GNU is a testimony to this.  In 
fact, I should have participated to POSIX more and made sure some of my 
sed extensions went into POSIX.2-2008.

But this is not a case in which POSIX can be improved.  POSIX provides a 
nice grammar for regex, and our warning is a hack on top of that 
grammar.  POSIX provides the clean thing as a standard should, and we 
build on top of it a useful hack.

What could have a place in a standard, is a mechanism for grep to warn 
about doubtful regular expressions.  Mandating "what is" a doubtful 
regular expression (which is a prerequisite if you want grep to exit 
with status 2) does not have its place in a standard.  The C standard 
does not say what algorithms to use in order to find uninitialized 
variables or dead stores.

> We should not let standards get in the way of improving our software.
>
> That we are making this the default behavior
> is a tribute to the usefulness of your new feature.

It's not.  It's a huge mistake, because making it an error means 
changing the regex grammar (and making it unnecessarily complicated and 
contrived).

> The following change-set implements what Paul and I have been advocating.
> I'll push it later today or tomorrow.

I still think this is wrong, and doubly wrong because I cannot disable 
it on my system without breaking it with POSIXLY_CORRECT.  Please, 
please implement --warn=error instead.  Only then we can discuss about:

- making it the default

- making --warn=auto choose between "error" (instead of "yes") and "no"

- leaving things as is, releasing 2.6.4 or 2.7 now, and revisiting the 
topic again later

Paolo

ps: maybe you can look for another opinion and ask glibc to implement 
this in regex.  Then watch Ulrich closing it as nonsense in 2 nanoseconds.