the 'messages' are already part of 'preload', so there is no need for
that. Last changes there: r101507/r102304
Change-Id: I6c9daa3c46bd7f1db851cc4620ecf59f81e6f9aa
Fallbacks didn't work at all for the new plural rule system. I fixed it
by moving the getPluralRules() and getCompiledPluralRules() calls near
to the readPHPFile() calls, before merging is done. Then I factored out
the resulting code to readSourceFilesAndRegisterDeps().
* Removed pluralRules from mergeableMapKeys, it isn't one
* Added compiledPluralRules to allKeys so that it will be merged
* When a language is not present in the CLDR XML files, return null from
getPluralRules() and getCompiledPluralRules() so that the fallback
rules won't be overridden with an empty array. Normalised it back to
an empty array in the unlikely event that there is no plural data in
the fallback sequence at all, even in English.
* Fixed private function, "protected" is the way to say private here.
Change-Id: I3a008ef7f6ed7aaa15ad25ad796f7a2b8f827fa2
Wrote a CLDR plural rule parser to replace the eval()-based one from
I58a9cdfe. It converts the infix notation of the XML files to a
sanitized RPN notation, referred to in external interfaces as the
"compiled" form. The RPN notation is cached and then executed by a
fast non-validating evaluator.
Timings for the largest rule in the XML file are ~1.2ms for
compilation and ~200us for execution.
Also:
* Lazy-load the plural rules when recache() requests them, instead of
loading them for every request.
* Language::convertPlural() needs integer keys, and CLDR only gives
string keys. The previous code was not mapping them so it didn't work
at all. I just mapped them in the order they appear in the XML file,
i.e. the first rule becomes MediaWiki's $pluralForm=0, the second
becomes $pluralForm=1, etc. Not sure if there is a more rigorous way
to do it.
Change-Id: I65ee788c1a8e5ee2ede2091990d86eb722749dd3
* Use the plurals.xml of CLDR for the plural rules of languages
* Use plurals-mediawiki.xml to override or extend the rules inside MW
* Remove the convertPlural method in each LanguageXX.php
* Parse and load the xml files in LocalisationCache
* Use the CLDRPluralRuleEvaluator.php for parsing the cldr plural rules
(This is taken from Translate extension and might require a replacement
parser without using eval)
* Add getPluralRules() to make the CLDR plural rules available to JS.
PS3: More method documentation, cleanup
Change-Id: I58a9cdfe60c7b9027bf031c91370472054f04ae2
As Brion points out in <https://bugzilla.wikimedia.org/show_bug.cgi?id=29747#c7>, the shared memory functions are very buggy. In fact, eAccelerator has disabled the shared memory functions by default for awhile now. I didn't touch APC, XCache or WinCache, but I imagine they suffer from the same problems.
Writing to cache is now handled in a new method. That makes recache() a bit
shorter and easier to read. Since there is no point in writing data to
/dev/null , return early in such a case (ie if store is LCStore_Null).
.... I've found another bug with the implementation - it also
relies on the presence of key '...:preload' and doesn't check it
when checking for the expired cache, which can lead to "Invalid or
missing localisation cache" exceptions.
* 'editfont' => 'sans-serif'; -> trivial (can be set in CSS)
* 'underline' => 0; by languages written in scripts that are hard to read with underlines -> now set in shared.css, which makes it work also when they are used as interface language (note that [lang="xx"] is not yet present on wikitext content by default but will be soon as part of my work on RTL improvements)
* 'quickbar' => 2; by RTL languages. For this, I introduced a new option (5) that sets left/right according to the directionality of your interface language (and automatic is better than setting it in MessagesXx files!). Note that the broken feature was in this case "corrected" by CSS flipping, causing preferences to say "left" while it was actually "right". This commit fixes that bug as well.
Third, it actually would have made more sense to have it them load for wgLang instead of wgContLang, but that's not possible because wgLang is itself dependent on preferences
* MW_MATH_* constants are now defined in Math extension
* Language::getMathNames() is removed
* mathNames section in message files is removed
* A hardcoded preference override in refreshLinks moved to MaintenanceRefreshLinksInit hook
* Allow any language code which consists entirely of valid title characters, and does not contain any path-traversal characters, to be customised via the uselang parameter. Language::isValidCode() represents this concept.
* Add some shortcuts preventing Language and LocalisationCache from looking for localisation files for a language code which does not follow the usual form of language codes in MediaWiki, i.e. /[a-z-]*/. This concept is represented by Language::isValidBuiltInCode().
* Do not allow colon characters in file names, per Platonides' suggestion on CR.
* In MagicWord::initRegex(), put the longest synonyms first in the alternation. This means that when one synonym is an initial substring of another, the longest one will match in precedence, which is generally in line with user expectations.
This patch automatically includes synonyms from the fallback localisation, and prevents the case flag from being overridden, thus making extensions resistant to translator error.
* MessageCache::singleton() calls wfGetMessageCacheStorage() directly instead of using $messageMemc, just in case this would be called before that variable is set
* Per TimStarling: also removed deprecated methods in MessageCache class: addMessages() and related, [get|set|enable|disable]Transform(), loadAllMessages(), loadMessageFile() and some others. Same for the legacyData stuff in LocalisationCache that was only used by MessageCache::addMessages() and related.
* Converted remaining extensions
