Re: regexp-replace and ^, word boundary or look-behind operators

[Stephane Chazelas <stephane.chazelas@xxxxxxxxx>]

2019-12-16 21:10:13 +0000, Stephane Chazelas:
> The way regexp-replace works means that these things:
> 
> $ a='aaab'; regexp-replace a '^a' x; echo "$a"
> xxxb
> $ a='abab'; regexp-replace a '\<ab' '<$MATCH>'; echo $a
> <ab><ab>
> $ set -o rematchpcre
> $ a=xxx; regexp-replace a '(?<!x)x' y; echo $a
> yyy
[...]

FWIW, looks like some sed implementations (like that of the
heirloom toolchest or busybox) or ksh93 have the same problem:

$ echo xxx | busybox sed 's/\<x/y/g'
yyy
$ a=xxx ksh -c 'echo ${a//~(E:^x)/y}'
yyy
$ a=xxx ksh -c 'echo ${a//[[:<:]]x/y}'
yyy

It may be that the POSIX regex API doesn't have a way to fix
that (REG_NOTBOL addresses the ^ case, but there's nothing about
\< / \b / [[:<]] which are non-POSIX extensions anyway).

PCRE should be OK, so it could be just a matter of
exposing it via the pcre_match builtin and document the
limitation otherwise for EREs (PCRE is the new de-facto standard
anyway).

-- 
Stephane