Stephane Chazelas wrote:

> I don't know the details of why it's done that way, but I'm not
> sure I can see how calling pcre_exec that way can be quicker
> than calling it on each individual line/record.

It can be hundreds of times faster in common cases. See:

http://git.savannah.gnu.org/cgit/grep.git/commit/?id=f6603c4e1e04dbb87a7232c4b44acc6afdf65fef

> Note that this is still wrong:
>
> $ printf 'a\nb\0' | ./src/grep -zxP a
> a
> b

Thanks, fixed by installing the attached.

> Removing PCRE_MULTILINE (and get back to calling pcre_exec on
> every record separately) would help except in the cases where the
> user does:
>
> grep -xzP '(?m)a'

I don't think grep can address this problem, as in general that would require 
interpreting the PCRE pattern at run-time and grep should not be delving into 
PCRE internals. Uses of (?m) lead to unspecified behavior in grep, and 
applications should not rely on any particular behavior in this area. This is 
firmly in the Perl tradition, as the Perl documentation for this part of the 
regular expression syntax says "The stability of these extensions varies widely. 
Some ... are experimental and may change without warning or be completely 
removed." Also, the grep manual says that -P "is highly experimental". User 
beware, that's all.