Locale Support in Qt 5

Early Qt localization support was weak with a design based on Windows APIs, but with its own code and data. Over time new features have been introduced based on CLDR data, but the current QLocale implementation is still incomplete, inconsistent, not fully integrated with the host system locale, and unnecessarily bloats the core library. It is clear that Qt needs a new approach.

Current Support

Qt provides two core options for localization:

• System Locale: Use the default system locale settings, the implementation being platform specific. This should include access to both the current user locale and the underlying system default locale.
• Custom Locale: An application can create any custom locale supported by the CLDR data embedded in QtCore which is then used by Qt's internal code routines. These may or may not match the host's available locales.

In theory, the System Locale should use the facilities of the host system, falling back to Qt's CLDR-based facilities where the host system doesn't provide the required data or API. In practice, Windows and Mac use an inconsistent mixture of the host facilities and Qt's CLDR-based facilities. On all other platforms such as Linux and Android Qt's CLDR-based facilities are always used.

Locale Changes

In theory, all Qt apps should fix the locale used at start-up and not change that locale until the entire UI can be refreshed at the same time. The app should respond to a change locale system event by refreshing the QLocale and then refreshing the GUI. Unfortunately this facility appears to be incomplete on most if not all platforms, with no obvious api available. (More research required).

• On Windows and Unix, the locale ID is fixed on creation and an internal refresh call is available.
• On Mac, CFLocaleCopyCurrent is used for every call then released, so the cached object may be reused but this is not guaranteed. No internal refresh call is available

Number Formatting

All number formatting, both system and custom locales, is done using the internal CLDR data and Qt code, no host resources are used beyond using the locale name to find the internal CLDR data to use.

• Any host data that differs from the CLDR data in Qt or any user overrides are not used.
• The Qt internal code only supports single-character number punctuation marks.
• The Qt internal code does not support proper number-grouping.
• To be able to remove the CLDR data from Qt will require obtaining the data from the host instead.
• Using the host for both data and code may affect performance
• Using the host for code may affect functionality used by other Qt code such as QValidaters, QString, and QDateTime

Future Support

A number of key requirements for future improvements were defined:

• Minimise locale data shipped with Qt
• Minimise localization code needing to be maintained by Qt
• Integrate fully with the system locale on all platforms
• Integrate fully with any user level overrides
• Add support for time zones
• Add support for calendar systems
• Add support for localized collation
• Add support for advanced formatting/parsing options such as spell-out, duration, etc.

These can be summarised as reducing the Qt code base, integrating fully with the host platform, and supporting advanced features.

In the past, Qt has emphasised consistent behaviour across platforms as being the key driver, at the expense of Qt apps looking and behaving the same as all other apps on the users platform, i.e. making life easier for the developer rather than the end user. A better balance needs to be found between the two positions, but with the emphasis on fitting in with the users expectations for the fundamentals, i.e. date and number formats must be as the user expects them to be unless the developer has good reason. This means better using the host facilities for the primary localization features.

Implementation Options

A number of options have been investigated for implementing these improvements.

The initial plan was to extend the existing QLocale code and data with new code, code from KDE, and more data from CLDR As a test case calendar system support was added using KDE code and CLDR data, but this proved to be unacceptable in terms of the library size required by the new data. It would also result in a lot of new code that Qt would struggle to code and support, especially in difficult areas like collation. It also failed the test for improving host integration.

The second plan was to utilise ICU as the localization back-end on all platforms to minimise code and data requirements within Qt. This had the advantage of a single code base and consistent behaviour and feature set across platforms. The main disadvantage was that ICU does not respect a users personal preference overrides or the host settings, so would be inconsistent. This option proved unworkable however due to resistance from Windows devs to the extra dependency and download, Apple App Store policies preventing linking to ICU, and Android only shipping the Java version.

The third plan adopted is to implement individual system back-ends for each host platform to fully utilise the system locale resources. While more code than option 2, and limited to a lowest common denominator feature set, it will at least provide a fully integrated appearance with the host system, and is the only design pattern that will work on all platforms without additional large amounts of code or data or external library dependencies.

The greatest weakness with option 3 is that the lowest common feature set is very low indeed, limited to what the Win32 API in Windows 7 supports, a very limited feature set no better than the current QLocale support. Given that Windows 7 will need to be supported for at least 5 more years (official end-of-support is 2020) then we are unlikely to be be able to add any advanced formatting support into QtCore during the 5.x series, and possibly even a future 6.x series. This will fail to meet the needs of many developers who require advanced formatting. Instead, a separate library will be required for features.

This is the option now documented below. Detailed planning for the ICU option 2 can be found at Qt-5-ICU. Many of these details will still apply to the new separate library, host-based plan.

Solution Design

The solution proposed is split into two sections, changes to the current QLocale code, and the new QtGlobal library.

Changes to QtCore::QLocale:

• Remove all custom locale data and only support what locales the host platform does. If an app requires more locales they can choose to use QtGlobal for this purpose.
• Remove all internal formatting code
• Create a new private QPlatformLocale class to implement backends on each platform. The api will match QLocale to improve performance. Each platform will receive a clean implementation to ensure only latest api is cleanly and correctly used.
• Windows will convert from LCID to Locale Names and will use full Win32 api available, supplemented by fallback code in limited circumstances
• WinRT will use the new WinRT api instead for improved support
• Mac will solely use Mac api
• Linux will use ICU api
• Embedded will have a minimal fallback C implementation for when platform devs do not require localization
• A number of key missing features could be added to QLocale to reduce teh need to use QtGlobal, such as 24-hour time formatting

Other QtCore changes:

• Add QCalendarSystem support across all platforms
• Update QTimeZone support to the new minimum platform features

New QtGlobal module:

• A Qt Add-on that ships as part of the main Qt release so it may be relied upon
• The api implemented will be teh ICU/CLDR style api as implemented on all main platforms
• The module will implement api support for as advanced a feature set as possible, with devs checking at runtime which features may or may not be available in a local host. Where a dev does call an unavailable feature thsi will degrade gracefully by default
• On all platforms except Windows 7 uses the full host api available
• On Windows 7 will be built against ICU and will require ICU to be shipped with the app
• Any dev may choose to build using ICU for any advanced features not available on their platform.

Long term, once Windows 7 support is dropped, QLocale would be dropped and QtGlobal would be merged into QtCore.

A major issue with this design is the heavy overlap of shared code between QtCore and QtGlobal, i.e. the Linux backend in QtCore will implement ICU support, which will be replicated and enhanced in QtGlobal. This would be a necessary evil however to allow for QtGlobal to be built on Windows 7 using ICU. This would not be possible if QtGlobal was to use a shared private api exported by QtCore. It may still be possible for Mac and WinRT to share their implementations in this way. In this event, the resulting QtGlobal library could be very small as a result.

An alternative is to make the new API part of QtCore but with Windows 7 devs advised that if they don't use ICU then the new API will not actually work. This would be controversial. In any event, teh new code will be easiest developed outside QtCore and migrated later if required.

Platform Support

While we cannot use ICU directly on all platforms, all platforms except Windows 7 (Win32) use ICU or CLDR data as a base and so have a consistent feature set, object model and api that we can design to. Careful API design using new classes for the new features, an API that can be queried for supported features, and sensible fallback options where a feature is unsupported on minor platforms will allow for different levels of support on different platform versions while allowing the latest versions to make most new features available.

Choosing the supported lowest common set of advanced features is slightly tricky depending on whether we choose the lowest deployment version of each platform supported by the Community (say OSX 10.7 supporting ICU 4.6) or the lowest Reference or Officially supported platform (say OSX 10.8 supporting ICU 49). Currently the absolute minimum version would be RHEL using ICU 4.2 but this is very restricting.

As at Qt5.6 the minimum supported platforms are:

Platform	Reference	Official	Community	ICU
Windows	7	7	7	Win32 Only
Windows Embedded			7.0?	Own API
Windows Phone	8.1	8.1?		Own API
Windows RT	8.1	8.1?		Own API
Mac OSX	10.8	10.8	10.7?	4.6 or 4.9 via Own API
Mac iOS	8.0	5.0?		? via Own API
Linux Ubuntu	14.04	14.04	11.10?	4.4 or 52
Linux RHEL	6.6	6.6	?	4.2
Linux OpenSuse	13.1	13.1	?	51
Android	4.1 (API Level 16)	4.1 (API Level 16)		4.8 via Own API
QNX			SDP 6.6	?

ICU

ICU ships standard on all modern Linux and BSD distros, and is shipped standard with QNX. With POSIX locale functions being clearly deficient (i.e. no calendar system support, poor collation, etc) then this is the preferred back-end for use on all Unix-like platforms, including embedded and QNX. One limiting factor is that the C++ ABI is not stable between versions so the C API must be used instead, and this lacks many of the features of the C++ API.

The defining platform here is RHEL 6.6 which only uses ICU 4.2, but it could be argued that RHEL is mostly a server distro and so localization is perhaps not as high a profile requirement there, so a later target of ICU 51 (OpenSUSE 13.1) could be used with a degraded feature-set on RHEL.

Handling of user-level overrides could occur at app level, but a platform wide solution should be sought (e.g. so KDE can define desktop-wide overrides).

Linux	ICU
RHEL 6.6	4.2
Ubuntu 11.10 LTS (Oneiric)	4.4
Ubuntu 12.04 LTS (Precise)	4.8
Ubuntu 14.04 LTS (Trusty)	52
OpenSuse 13.1	51
OpenSuse 13.2	53
QNX 6.6.0	?

Mac OS X / iOS

Mac uses ICU for localization, but does not ship the ICU headers and prohibits apps that link to ICU from the App Store. Instead we must use the native Mac API which in reality is a thin wrapper around ICU with a simplified API and some Mac convenience methods added. The Mac API also uses any user level overrides which using ICU directly would not. One problem is that the wrapper API's may not have been updated to provide access to new features in each ICU release used.

Qt 5.4 supports 10.6 at community level, and 10.7 at Official and Reference level. Qt 5.5. drops 10.7 as an Official and Reference platform, making OSX 10.8 and ICU 49 the lowest supported version.

OSX	ICU
10.6	4.0
10.7	4.6
10.8	49
10.9	51
10.10	53
10.11	55

Reference: http://opensource.apple.com/

Android

Android uses ICU for localization, but only ships the Java version of the library and data files. We will use the Java Android api via JNI.

The ICU version changes in Android are as follows:

Version	Codename	API Level	ICU
2.3.3	Gingerbread	10	4.4
3.0	Honeycomb	11	4.4
4.0	Ice Cream Sandwich	14	4.6
4.1	Jelly Bean	16	4.8
4.3	Jelly Bean	18	50
4.4	Kit Kat	19	51
5.0	Lollipop	21	53

Reference: http://developer.android.com/reference/java/util/Locale.html

Windows 7

Windows 7 continues to use the Win32 API for localization functions, supplemented with some new functions using Locale Names in place of LCID's and improved calendar data access. Unfortunately, the core number and date formatter calls have not changed and so remain the lowest common denominator and preventing and new advanced formatting API. Initial analysis does indicate that the calendar api is sufficient to implement the minimal required cross-platform support.

Windows Runtime

WinRT and Windows Phone provide advanced localization functions clearly based on CLDR data and broadly comparable to ICU. An initial review of the WinRT api indicates all required features are exposed, but this needs to be fully documented.

Embedded / Fallback

Some embedded platforms may prefer not to ship ICU, or may only have very simple localization requirements. It will be required to provide a fall-back implementation for these platforms. This could be the existing QLocale code and database, but that would be a substantial maintenance burden. The alternative is to ship a simple C-locale back-end or a pass-through to POSIX. One option is to provide good documentation on building the minimal ICU required to support the embedded platform's locales.

API Design

The new API will be implemented as a set of new classes completely separate to the existing QLocale class:

• QLocaleCode
• QNumberFormatter
• QDateTimeFormatter
• QCalendar
• QTimeZone

There are a number of very strong reasons for this:

• This is the design pattern used by ICU, OSX, Windows, and Java that all devs are already familiar with
• It is more efficient as ICU splits the locale resource files that are loaded by the Number and DateTime formatters, so a monolithic class would take longer to load
• It represents a clear break with the old API and format codes, making it clearer to devs that behaviour and codes have changed
• It prevents API bloat by having a single formatter api for different formatter types rather than multiple calls with prefixes or extra enums
• It allows different levels of feature support for different formatter types on different platforms or versions while keeping the same simple api across them all

The old QLocale backend code will be removed and replaced with calls to the new code. It is expected this compatibility layer will take considerable effort and testing to ensure backwards compatible behaviour.

ICU Analysis

This section analyses the available ICU API for a common feature set. This analysis is based on the ICU 4.2 C api as used in RHEL 6.6, but will also analyse later versions for new features.

Calendar Support

Note ICU Calendar C API merges Calendar and Time Zone into one 'object'.

Supported range in ICU 56: Julian day numbers of -0x7F000000 to +0x7F000000. This corresponds to years from ~5,800,000 BCE to ~5,800,000 CE. Was previously wider range.

Calendar Systems:

• "gregorian"
• "chinese"
• "coptic"
• "ethiopic"
• "taiwan"
• "indian"
• "persian"
• "japanese"
• "buddist"
• "islamic"
• "hebrew"
• "dangi" ICU 50 Java ICU 51 C
• "islamic-umalqura" ICU 52
• "islamic-tbla" ICU 52

Date/Time Components:

• UCAL_ERA
• UCAL_YEAR
• UCAL_MONTH
• UCAL_WEEK_OF_YEAR
• UCAL_WEEK_OF_MONTH
• UCAL_DAY_OF_YEAR
• UCAL_DAY_OF_WEEK
• UCAL_DAY_OF_WEEK_IN_MONTH
• UCAL_AM_PM
• UCAL_HOUR
• UCAL_HOUR_OF_DAY
• UCAL_MINUTE
• UCAL_SECOND
• UCAL_MILLISECOND
• UCAL_ZONE_OFFSET
• UCAL_DST_OFFSET
• UCAL_YEAR_WOY
• UCAL_DOW_LOCAL
• UCAL_EXTENDED_YEAR
• UCAL_JULIAN_DAY
• UCAL_MILLISECONDS_IN_DAY
• UCAL_IS_LEAP_MONTH
• UCAL_FIELD_COUNT
• UCAL_DAY_OF_MONTH

Other:

• Min/Max values
• Lenient
• Gregorian change date
• First day of week
• Min days in first week
• ICU 4.4 Weekday/Weekend
• ICU 49 Ambiguous Wall Time
• ICU 50 TZ Transitions
• ICU 51 Get TZ ID
• ICU 52 Windows TZ ID

Mac Analysis

This section analyses the available Mac API for a common feature set. This analysis is based on a minimum OSX 10.8 but will also analyse later versions for new features.

Calendar Support

Calendar Systems

• API Docs say enum deprecated since 10.10?
• NSGregorianCalendar
• NSBuddhistCalendar
• NSChineseCalendar
• NSHebrewCalendar
• NSIslamicCalendar
• NSIslamicCivilCalendar
• NSJapaneseCalendar
• NSRepublicOfChinaCalendar
• NSPersianCalendar
• NSIndianCalendar
• NSISO8601Calendar - Doc says not implemented?
• OSX 10.11 System Preferences lists all above as available, plus Amete Alem, Umm al-Qura, Islamic Tabular

Win Analysis

This section analyses the available Windows API for a common feature set. This analysis is based on Windows 7, but will also analyse later versions for new features.

Components

QTimeZone

QTimeZone has been successfully implemented in Qt 5.2 using separate back-ends for each system but a common API. The design of this class will be copied for much of the new QLocale implementation, especially QCalendar.

In 5.3 a number of new features are still required, including a QEvent for TimeZoneChanged and a QTimeZoneDatabase class to load TZ databases on any platform.

Old design details can be found at http://wiki.qt.io/Qt-5-QTimeZone

QCalendar

QCalendar will follow the design of QTimeZone to wrap the system provided calendar calculators.

Not all platforms equally support the same set of calendar systems, although this is slowly converging thanks to the increasing use of CLDR and ICU. While QCalendar will have to define an enum for all possible calendar systems, it will also have to provide an availableCalendarSystems() api to describe what systems are available to be set in the formatter.

QCalendar will implement baseline support for as many calendars as possible so that even where a calendar system is not available on a host system it may still be used as an optional calculation class, but not in the formatter.

ICU C++: http://icu-project.org/apiref/icu4c/classicu_1_1Calendar.html

• ICU C: http://icu-project.org/apiref/icu4c/ucal_8h.html
• Android / ICU4J: http://developer.android.com/reference/java/util/Calendar.html
• Mac Cocoa: https://developer.apple.com/library/mac/documentation/Cocoa/Reference/Foundation/Classes/NSCalendar_Class/Reference/NSCalendar.html
• WinRT: http://msdn.microsoft.com/en-US/library/windows/apps/windows.globalization.calendar.aspx
• Win32: http://msdn.microsoft.com/en-us/library/windows/desktop/dd319114(v=vs.85).aspx

Calendar	ICU	Mac	Android	WinRT	Win32	KDE
Gregorian	gregorian
Chinese Lunar	chinese
Coptic	coptic
Ethiopic	ethiopic
Ethiopic Al Amate	???
Indian National	indian
Jalali	persian
Hebrew	hebrew
Islamic Civil	islamic
Islamic Um-al-Qura	slamic-umalqura (ICU 52)
Islamic Tabular	islamic-tbla (ICU 52)
Japanese (Gregorian)	japanese
Dangi (Korean Lunar)	dangi (ICU 51)
Taiwan (Gregorian)	taiwan
Thai (Gregorian)	buddist