Re: [PATCH] fall back to glibc matcher if a MBCSET is found

[Paolo Bonzini]

On 09/08/2010 11:05 AM, Jim Meyering wrote:
> Thank you for the patch.
>
> If this change really does fix a correctness bug,
> then it deserves a NEWS entry with enough detail to confirm that,
> and, if at all possible, a test suite addition.

It fixes equivalence classes (e.g. matching [[=a=]] against à), but only 
--without-included-regex.  See attached patches.

The presence of this check in regex.m4

             if (sizeof (regoff_t) < sizeof (ptrdiff_t)
                 || sizeof (regoff_t) < sizeof (ssize_t))

unfortunately means that all existing systems will use the inferior 
gnulib regex rather than glibc regex.  In turn, this means that grep 
will nowhere support equivalence classes out-of-the-box.

> Similarly, if it works around a performance problem,
> it would help me evaluate it if you were to provide evidence.

yes 1234567890123456789012345678901234567890123456789012567890 | \
  sed 100000q | time ./grep '[a-z]'

shows 0.91s with the patch and 1.21s without.  Since this is not an 
asymptotic improvement, it is hard to test it reliably, and is secondary 
anyway compared to the correctness problem above.

Paolo

fallback.patch
Description: Text document