Message Formatting Redesign

Goals

- Modularization: We want to be able to offer formatting without static dependencies on the majority of ICU.
  - We should have simple formatting support in the ICU4C common library, with just string substitution and "select" formatting, as a basic formatting API and for Locale Display Names and similar which use CLDR patterns like "{0} {1}" for the ordering of items.
  - We should have formatting with numeric support, adding support for "number", "choice" and "plural" argument types.
  - Finally, full support.
- Threading: Significant objects and methods should be thread-safe, so that objects can be cached and used concurrently.
- Performance: It would be nice if it was fairly fast to format a string with arguments without remembering context. It should be possible to improve speed with caching, but one-shot usage should be ok.
  - For example, applying choice/plural/select formats should not require re-parsing their selected message fragments.
- Extensibility: Currently, the interaction between MessageFormat and Choice/Plural/SelectFormat is hardcoded and relies on RTTI. It should be possible to add another such argument formatter without modifying the core formatting class.

Ideas

Message class

Superseded by icu-design 2011-02-08 00:59: ICU4J API *pre*-proposal: MessagePattern class

A parser of the basic MessageFormat syntax, and container for the parsed message representation. It would parse the whole message string, remove the syntax it recognizes (quotes and braces, some whitespace) and build a simple data structure for efficient usage.

- setAutoQuoteApostrophe(boolean mode)
- parse(CharSequence msg[, UErrorCode])
- is-a Freezable -- cache-friendly and thread-safe once frozen
- String getParsedString() -- syntax removed, usable only with indexes
- int countIndexes()
- int getIndex(int i)
- (possibly helpers like int findEndOfArgument(int startIndex))

Design for indexes: (designed for easy processing)

- The indexes would be bit fields, with bits 24..0 for the actual index into the parsed string. (This "limits" parsed message strings to a length of 0x1ffffff.) The indexes are stored in ascending order.
- Bits 26..25:
  - 0: Start of argument name/number
  - 1: Start of argument type
  - 2: Start of argument style (pattern)
  - 3: End of argument
  - Issue: We need to allow nested {fragments} that do not use argument syntax, and distinguish them from arguments.
  - How does the parser distinguish a {fragment} like in plural and select formats from an {argument}, without the parser knowing details of the subformat style syntax? Would the parser have to be limited to just strip '{} and note where the {} were with nesting levels, but without preparsing argument numbers/names/types?
- Bits 30..27: Nesting level 0..15, increases at each { and decreases at each }.
- Bit 31: Sign bit, not set.

Parsed message string:

- Quotes (apostrophes) removed except where they were quoted. Text is always literal.
- Syntactic (unquoted) braces removed. Arguments are found via the indexes.
- Argument numbers are preparsed and stored as a UChar 0..0x3f. (This "limits" the number of numerically indexed arguments to 64.)
- Argument names are trimmed and checked for [:ID_START:][:ID_CONTINUE:]* syntax.
- Argument types are trimmed.

Example:

Original message string: "The {0} '{}' chased {1, select, female {her} other {his}} tail."

Parsed message string: "The \u0000 {} chased \u0001select female her other his tail."

Indexes: 1+start+4, 1+end+5, 1+start+16, 1+type+17, 1+style+23, 2+start+31, 2+end+34, 2+start+41, 2+end+44, 1+end+44

Formatter class

A formatter has-a Message and formats it, invoking argument formatters on the way. Recursion eliminates the MessageFormat-Choice/Plural/Select RTTI magic.

Design choice:

1. We could provide Formatter subclasses that know about the levels (simple/with numbers/all) of argument formatters. It would be customizable by writing additional subclasses.
2. Pro: This is easy to implement.
3. Con: You have to use a formatter that supports the argument types in your message.
4. Con: (minor) Requires writing a Formatter subclass for supporting new argument types. (Probably very few users would provide argument formatters beyond what ICU provides.)
5. We could use registration: Register an argument type with an argument formatter factory. The factory would take locale, type and style and return an argument formatter.
6. Pro: After registration, the same Formatter class would be able to handle all registered argument types.
7. Con: When would we do the registration in C++? Would we require users to call one of several functions to register some bundle of argument formatters?
8. Con: Argument formatter access would be through a synchronized registry cache getter.

Public Formatter API: (not thread-safe; making this immutable seems limiting, or seems to require a worker object)

- Formatter(Locale locale[, UErrorCode])
- aliasMessage(Message msg) -- uses the Message but does not take ownership of it
- ?parseMessageString(CharSequence[, UErrorCode]) -- convenience?
- (format() methods would be non-const, setting the args into the Formatter object, which will probably also have a cache of specific argument formatters like MessageFormat does)
- format(Appendable dest, Object... args)
- format(Appendable dest, Map namedArgs)

Protected Formatter API (if we use subclasses):

- Object getArgument(int argNumber)
- Object getArgument(String argName)
- format(Appendable dest, Message msg, int styleStart /* use msg.getIndex(styleStart) */, int argNumber, String argType)
- format(Appendable dest, Message msg, int styleStart /* use msg.getIndex(styleStart) */, String argName, String argType)
- format(Appendable dest, Message msg, int msgStringStart, int msgStringLimit) -- for recursive calls from choice, and called with (0, msg.getParsedString().length()) from the public format() methods
- format(Appendable dest, Message msg, int nestedIndex) -- for recursive calls from plural/select

NumberFormat

NumberFormat should have an optimization for small integers. It could cache a limit value such that formatting numbers below the limit is simple and fast. For example, if the NumberFormat uses decimal digits from somewhere on the BMP, the limit could be 1000 for standard grouping (for -1000<intValue<1000), or much larger if the grouping separator is not used. Formatting a below-limit integer should use very fast code.

Formattable

In C++, we will have to move Formattable to the common library. (In Java, the arguments are passed in as Object.) Check if we can do so without dragging formatting code with it.