Fixed-width strings

Design Goals

Currently, Tamarin-Tracing strings are UTF-8. As much as these are desirable in terms of memory consumption, index access and other functions are unbearably slow.

The design goals for the new StringObject class are

1. Fixed-width strings with fixed widths of 8, 16 and 32 bits. Widths are either automatic (based on UTF-8 input), or manual (as created).
   • 8-bit strings contain the first 256 Unicode characters. UTF-8 is not supported in strings.
   • 16-bit strings are UTF-16, including surrogate pairs.
2. String contents may be based upon static data.
3. String concatenation creates a tree structure, which is flattened when needed.
4. String APIs should be as close to SpiderMonkey string APIs as possible to make ActionMonkey implementation easy.
5. Preferred access to characters is via a charAt() method that returns 32-bit characters regardless of the underlying implementation. Low level API strXX() are available for direct read-only buffer access if really needed.
6. Integrators control which width is being used internally. If they only feed 16-bit data, or UTF-8 data with the desired width set to 16 bits, only 16-bit data will be used. There will be a compile-time flag that comments out code that deals with other widths that 16 bits.

Strings are immutable, but their contents may change, invisible to the owner of the instance. String data can be 8, 16, or 32 bits wide; A string may contain a pointer to a string buffer, which may need to be deleted or not; or the string data may immediately follow the String instance in memory (by doing a raw allocate followed by an in-place constructor call); or it can hold two String references as the result of a concat operation. Finally, a string can be the result of a substring operation. The contents may change dynamically during a flatten operation. Therefore, a String instance contains a union with a tag.

Out-of-memory conditions will be handled by the allocator in a future version. The String class will use checks for NULL, and return NULL for new strings. For internal operations, the result of out-of-memory conditions still need to be defined.

UTF-8, UTF-16 and UTF-32

The core String class will ignore all encoding related issues. A string is just an array of characters. Widening a string from 16 to 32 bits will, for example, not combine a surrogate pair to a new character, as narrowing a string will not crack a 32-bit character with a value outside the Basic Multilingual Pane (0x0000-0xFFFF) into two surrogate pairs. The latter limitation prohibits automatic narrowing of 32-bit strings into 16-bit strings if the 32-bit strings contain characters with a value > 0xFFFF.

For these conversions, and for UTF-8 encoding and decoding, separate layers will be provided that return error codes. These APIs include:

• Widening and narrowing with surrogate pair processing for 16-bit strings
• Creating a string out of UTF-8 data
• Creating an UTF-8 data buffer out of a string

These layers will catch the following:

• Invalid UTF-8 character sequences
• Single 16-bit and 32-bit characters that have a value between 0xD800 and 0xDFFF
• 32-bit characters with a value > 0x10FFFF
• Any other conditions that Unicode 5 considers to be ill-formed.

Creation

Strings may either be created with 8, 16, or 32 bit data. In addition, strings may be created with UTF-8 data, which results in the smallest width that can hold the data, or a desired width that may cause the creation method to return NULL if the UTF-8 string contains characters that cannot be represented in the desired width. This is the case for 8-bit strings, and for 16-bit strings, if the character exceeds the value 0x10FFFF.

Strings are never zero-terminated. Zero-characters are legal as part of a string.

Strings are created using static creator functions. This allows the implementation to use raw memory allocation and in-place constructor calls to avoid having to do two memory allocations, one for the instance, and the other for the data. Strings created that way contain the data right behind the instance data.

The maximum string width determines the way strings are created. It is an optional argument to the string constructors.

1. 8 bits: If the source data contains 16 or 32 bit data, the return value is null.
2. 16 bits: If the source data contains 32 bit values, surrogate pairs are created. If a character is > 0x10FFFF, NULL is returned.

This allows implementers to define the maximum width of strings; they can choose to use 8, 16 or 32 bits only, or they can choose to go with whatever width that fits best. If they choose best-fit widths, string creation methods do not create UTF-16 surrogate pairs. If a script creates surrogate pairs, these will remain in strings, though, although a flattening operation could detect surrogate pairs and widen the flattened string to 32 bits.

Question: How are out-of-memory conditions handled? The current implementation often just assumes success. There should be some sort of exception, and the same mechanism should be used to report strings that cannot be created.

Concatenation, substrings, and flattening

It would not be a good idea to create a new, flat string every time two strings are concatenated. Consider this loop:

var s = "";
for (var i = 32; i <= 1024; i++)
  s += String.fromCharCode (i);

If a new, flat string would be created on every iteration, this would lead to a almost 1000 copy operations, with a growing string buffer. Instead, the resulting string contains two String pointers that point to the two source strings.

The above example would create a deep tree, which is also undesirable. Therefore, a String instance contains a treeDepth field that contains the deepest depth of both subtrees plus one. The concat operation will contain a threshold where a string will be flattened before it is used for concatenation. This value should be determined using various benchmarks for optimal memory/performance ration. Also, the field is limited in size (10 bits?), so at some point automatic flattening is forced.

In-place concatenation is supported. This method allocates string data buffers at multiples of bytes (16 or 32 bytes). Concatenating to such a string would fill the buffer until available space is exhausted, and create a new String instance that would point into the original buffer, but with a larger length. This way, the original string seems unchanged because its length and buffer did not change, and the new string shares the same buffer, with a larger length. This technique reduces memory allocation and provides flattened strings most of the time. This is especially desirable if the loop is an concat /access loop, where a simple concat would create a tree that is flattened in every loop iteration.

In the future, the tracing optimizer may detect concatenation in a loop and inform the String implementation about such a loop, which then would e.g. allocate larger data buffers to speed up these loops.

The .abc image contains UTF-8 strings that are not 0-terminated. A future abc format could deliver 8/16/32 bit fixed with strings, so it is desirable to support a string object that can point to external data. If the UTF-8 string contains Latin-1 characters only, the original buffer is used to create the String instance avoiding the copying of data.

Getting a substring also flattens the source string. The substring is an instance that contains a pointer to the source string, and pointer to the start of the source string buffer. The length field contains the string length. This string is already flat, although it contains a reference to another string. It may be desirable to have a separate flattening function for this case, and for the the case of a "super-string" that an in-place concatenating operation created.

When a string is flattened, its two String pointers are replaced with a flat data buffer. The resulting width of the string is determined by the widths of the strings in the tree. Usually, the resulting string width is the widest of all substrings found. If desired (with an #ifdef), substrings could also be analyzed if they are wider than the containing data, if e.g. a 16-bit strings only contains 8-bit characters. This is, of course, a performance hit, but may be desired if memory footprint is important, because flattening the concatenation of a, say, 8-bit string and a 16-bit string may result in a 8-bit string if the flattener is allowed to check the contents of the 16-bit string.

16-bit strings containing surrogate pairs will never automatically be widened to 32 bits, losing surrogate pairs, because this would change the string length and the location of its characters. There will be a separate API that allows for this conversion.

Thread safety

Since strings are immutable, they are by definition thread safe. The only unsafe operation if the flattening operation, and in-place concatenation. Currently, TT is not thread safe, so the sensible code should be clearly marked with a TODO comment until a global threading solution for TT is available.

SpiderMonkey compatibility

This is just the start of this section...

• The SM API offers the registration of string finalizers. Strings can be created with custom buffers that the finalizer takes care of deallocating. There are some predefined internal finalizers. All finalizers are stored in a global table with a fixed maximum size. The registration of a finalizer returns an index value into this table. Finalizer indexes correspond to string type enumerators (see last section). The size of that array is determined with a #define to allow ActionMonkey to compile the desired maximum size.

• JS_GetStringChars() returns a pointer to UTF-16 characters, and JS_GetStringBytes() returns a pointer to UTF-8 characters. Both buffers are guaranteed to live as long as the string instance lives. SM maintains a separate cache for this purpose, where string buffers are garbage-collected. Other encodings may be requested as well.

• What about growable strings? They are contradictory to the immutability of Tamarin strings. They could be implemented, but there should be some safety measure so they cannot be passed in to the engine. SRJ: I don't know if growable strings are a requirement for SM compatibility or not, but if they aren't, I think we're better off keeping strings immutable. It fits the ES model well and allows for useful simplification of the code.

• SM (and probably other integrations as well) would like to use only UTF-16. The String implementation should, if possible, use macros that remove unnecessary code if only UTF-16 is to be supported.

Sample instance data

This is a sample representation of instance data, just to illustrate the concept. In this sample, a String instance would occupy between 8+n (direct data) and 16 bytes, assuming 4-byte pointers. The first 64 bits would be the lengths and the bits-and-flags field, followed by aligned pointers.

The listing uses bit fields, but I think that we should use masks and shifts to ensure the correct width and alignment of data, along with inline methods to access the fields.

The string type is a tag that describes which union to use. In addition, it it an index into the table of finalizers. Negative values are internal finalizers, while 0 and positive values are external finalizers.

class String ... {
  ...
private:
  // common data
  uint32_t length;
  struct {
    unsigned int width:2;       // 0:1, 1:2, 2:not used, 3:4
      signed int type:9;        // string type, same as finalizer index
    unsigned int dynamic:1;     // if nonzero, buffer must be deleted
    unsigned int treeDepth:x;   // to be defined (more fields may follow)
    unsigned int padding:y;     // padding to 32 bits
  } data;
  // variable data according to tag
  union {

    struct {                    // data follows directly (normal case)
      union {
        unsigned char c8 [100]; // big number for debug display
        utf16_t c16 [100];      // actual size varies
        utf32_t c32 [100];
      }
    } direct;

    struct {                    // string with buffer (also flattened)
      union {
        unsigned char* c8;
        utf16_t* c16;
        utf32_t* c32;
      }
    } buffer;

    struct {                    // concatenated
      String* left, *right;
    } concat;

    struct {                    // substring and "superstring"
      String* source;           // flattened string source
      union {
        unsigned char* c8;      // buffer into source's data
        utf16_t* c16;
        utf32_t* c32;
      }
    } dependent;
  }
}