the position can be changed depending on the plural, as requested by Amir in
http://lists.wikimedia.org/pipermail/mediawiki-i18n/2011-August/000338.html
Placing the text after the unit could get funny in cases where there is
more text than just the number and the unit, but it's not a problem for the
parserfunction to fix.
The transformations for ParserFunctions.i18n.php were
On pfunc-convert-unit messages: s/$2/$3/, s/$1/$2/
s/'pfunc-convert-unit-.*' *=> '/\0$1/
The translations of prefixed units in some languages may be a little tricky if the prefix is inflected etc, but you know what? You can always use ParserFunctions!! :-D
* subclass ConvertError and ExprError from MWException
* invoking the parser for error messages clears state and exposes strip markers, and the error messages should be in the content language anyway
* clean up the magic word list
* Use autoloader for ConvertParser, and ExprParser which the code was cloned from
* Fix PHP 5.1 compatibility, while there's no requirement for us to keep it, there's also no need to break it unnecessarily, and it's a really trivial fix in this case.
* Tidy whitespace
* Use ConvertDimension::getLocalisedName() for the dimensionmismatch error, as intended
[[Template:Convert]] on enwiki is a behemoth of a construction that just about manages to do this sort of conversion, taking {{convert|5|mi|km}} and outputting "5 miles (8 km)", etc. To port this to another wiki requires copying over three and a half thousand subtemplates. The additional load produced by including numerous copies of this template is measurable on large pages on enwiki, and it eats voraciously into the template limits.
This revision introduces {{#convert: 5 mi | km }}, outputting "8 km" or thereabouts. See http://www.mediawiki.org/wiki/User:Happy-melon/Convert for more details, or look at the examples in the parser tests.
In a very rough profile, comparing 50 calls to {{convert}} verses the same 50 calls to the wrapper template shown at the link above, the parser function implementation reduces page load time by 72%, preprocessor node count by 83%, post-expand include size by 86% and template argument size by 97%. More detailed profiling would probably reveal places where extra caching could improve performance further.
The primary reason for putting it in ParserFunctions instead of its own extension is availability: PFs are already available across the cluster, and it's accepted as an essential extension for any wiki wishing to emulate or mirror WMF content. One less separate extension installed on the cluster is one less extension which has to be matched by reusers.
It's still missing a lot of units, which I ran out of patience to copy from {{convert}}; I thought I'd get some feedback on the infrastructure first.
