On Wed, Jun 05, 2013 at 02:16:22AM -0400, Andrew Gregory wrote:
On 06/05/13 at 02:35pm, Allan McRae wrote:
On 05/06/13 02:34, Dave Reisner wrote:
On Tue, Jun 04, 2013 at 06:24:03PM +0200, Patrick Steinhardt wrote:
On Tue, Jun 04, 2013 at 11:46:50AM -0400, Dave Reisner wrote:
On Tue, Jun 04, 2013 at 05:03:47PM +0200, Patrick Steinhardt wrote:
Until now it was not easily possible to remove all files but some in a given directory. By enabling Perl regular expressions for NoExtract this is now made possible through negative lookahead.
Fixes FS#31749. ---
This patch is work in progress. I want to check if there is interest in having this feature available in pacman and how the extra dependency on libpcre is received. If it is well received I might still do some rework for the actual matching of NoExtract items, as currently I'm always recompiling the pattern for each match.
I've considered moving to PCRE, but I'm curious:
1) Why do you need this for NoExtract? Please be *specific* 2) What is insufficient about POSIX regex? (#include
) My usecase is specifically the directory /usr/share/locale. As I only need two files/directories (locale.alias and en_US) in there I'd need to specify a list of a whopping 106 entries that shouldn't be extracted.
POSIX regular expressions do not support negative lookarounds. Thus I might add entries like usr/share/locale{af,am,...} but it is not possible to remove everything but those two entries. With Perl regular expressions this is as easy as usr/share/locale/((?!en_US|locale.alias).+).
I'm sure there are other usecases where POSIX regular expressions may be insufficient but can't currently think about any other. In generl PCRE is in the core repo, widely used and very powerful, so I don't see any reason to not use it.
I think a major hurdle here is going to be that we've already standardized on fnmatch() as the matcher for not only NoExtract, but also IgnorePkg, IgnoreGroup, HoldPkg, and NoUpgrade. Moving to regex based matching flat out breaks expectations since a valid glob will *not* match the same data when interpreted as a regex, and it might not even be a valid regex (consider "*foo").
If you want to propose such a move, you either need a migration path (unlikely), or some sort of flag to determine what style of matching is performed on these config options, something like:
MatchStyle = (Glob|Regex)
Along with the necessary documentation. I'm not sure this is a road I want to go down.
I am not sure what I think about this. So here are a collection of thoughts:
1) Being able to ignore e.g. all locales apart from the one you want would be a good thing.
2) PCRE is a widely used library and I doubt many Linux distributions do not have it installed (and have grep linked to it)
3) This would need to be entirely optional at configure time, much like gpgme is.
4) I really do not want another configuration option for this. It would have to be glob matches when built without pcre and regex when built with it. But then how would the user know which pacman is using?
5) The upgrade path is particularly important given pacman is mostly (only) used on rolling release distros. But would the distribution adding a note in pacman.conf that the user will merge be enough?
So... that has me leaning towards this being a good idea.
@Dave: would suggesting the distros handle the "migration path" be fine with you? I'm not sure wildcards in any of those configuration options are widely used.
Allan
I tend to think that the particular problem at issue here would be better solved by a negation operator ala gitignore:
NoExtract = usr/share/locale/* !usr/share/locale/en_US/*
If we do add pcre support, could we include the type of pattern being used in the pattern itself? Something like:
NoExtract = glob::usr/share/locale/* NoExtract = pcre::usr/share/locale/(?!en_US).*
That way there would be no ambiguity about which type of pattern is being used and we could default to a glob if no type is specified, allowing existing configs to continue to work.
This way of handling would work well with the build time option Allan wants: if pcre support is not builtin all pattern starting with pcre:: might just be skipped. Still, it doesn't feel right to prefix every regex and it's unlikely users will mix glob and pcre. Because of this I'd prefer the MatchStyle option. If pacman wasn't built with pcre it might throw a warning/error at startup when MatchStyle is set to pcre and inform the user that it cannot handle that because of missing libpcre support.
apg