implementing extended bracket expressions in gnulib [was Re: Dealing with character ranges in grep]

[Paolo Bonzini]

On 06/09/2011 01:12 PM, Bruno Haible wrote:
> What would it take to let distros/people use --with-included-regex and
> get understandable semantics for ranges + working equivalence classes?
>
> I would prefer that to your proposal, because it cannot be seen as a
> regression by people who care about equivalence classes.

My proposal wouldn't change defaults, which is why I believe that this 
is a separate topic.  You quoted

> > With my proposal, distros/people that use --with-included-regex would
> > get understandable semantics + no equivalence classes

but snipped this part:

> > Right now, some distros/people use --with-included-regex and get
> > broken semantics + no equivalence classes

So, I agree that understandable semantics for ranges + working 
equivalence classes would be the best, and if gnulib could provide that, 
I would champion making --with-included-regex the default.  However, 1) 
Aharon would like to release gawk 4.0 in the very near future, and 2) 
adding an extension to glibc takes time.  That's why I prefer to work in 
smaller steps.

> Can that be done through gnulib code? If not, what do we need from glibc
> to get it done in gnulib?

We'd need glibc to export two functions in both multi-byte and 
wide-character versions:

1) streqcoll(S1, S2) and wcseqcoll(S1, S2) would be the same as strcoll 
and wcscoll, but they would compare only according to primary weights. 
A slightly more formal definition is that streqcoll(S1, S2) == 0 iff S1 
matches the \`[=C1=][=C2=][=C3=]...[=Cn=]\' regular expression, where Ci 
are the characters of S2 (I'd need to double check this against POSIX 
though).  When non-zero, the result of streqcoll(S1, S2) would be the 
same as strcoll(S1, S2).  Likewise, glibc could provide streqxfrm and 
wcseqxfrm, with the definition that strcmp(streqxfrm(S1), streqxfrm(S2)) 
== streqcoll(S1, S2).

2) On top of this, [.ss.] could be implemented using an additional 
function mbelemlen(S) giving the length of the first collation element 
in S.  [.S1.] would be rejected unless mbelemlen(S1) == strlen(S1), and 
[.S1.] would match S2 if strcoll(S1, strndup(S2, mbelemlen(S2))) == 0. 
wcelemlen could be provided likewise.

These are the minimal extensions that would be required to support full 
regular expression features portably and in a manner that is compatible 
with glibc, except for ranges (which we don't care about, do we?).

Paolo