1 # Hyb (hyphenation pattern binary) file format 2 3 The hyb file format is how hyphenation patterns are stored in the system image. 4 5 Goals include: 6 7 * Concise (system image space is at a premium) 8 * Usable when mmap'ed, so it doesn't take significant physical RAM 9 * Fast to compute 10 * Simple 11 12 It is _not_ intended as an interchange format, so there is no attempt to make the format 13 extensible or facilitate backward and forward compatibility. 14 15 Further, at some point we will probably pack patterns for multiple languages into a single 16 physical file, to reduce number of open mmap'ed files. This document doesn't cover that. 17 18 ## Theoretical basis 19 20 At heart, the file contains packed tries with suffix compression, actually quite similar 21 to the implementation in TeX. 22 23 The file contains three sections. The first section represents the "alphabet," including 24 case folding. It is effectively a map from Unicode code point to a small integer. 25 26 The second section contains the trie in packed form. It is an array of 3-tuples, packed 27 into a 32 bit integer. Each (suffix-compressed) trie node has a unique index within this 28 array, and the pattern field in the tuple is the pattern for that node. Further, each edge 29 in the trie has an entry in the array, and the character and link fields in the tuple 30 represent the label and destination of the edge. The packing strategy is as in 31 [Word Hy-phen-a-tion by Com-put-er](http://www.tug.org/docs/liang/liang-thesis.pdf) by 32 Franklin Mark Liang. 33 34 The trie representation is similar but not identical to the "double-array trie". 35 The fundamental operation of lookup of the edge from `s` to `t` with label `c` is 36 to compare `c == character[s + c]`, and if so, `t = link[s + c]`. 37 38 The third section contains the pattern strings. This section is in two parts: first, 39 an array with a 3-tuple for each pattern (length, number of trailing 0's, and offset 40 into the string pool); and second, the string pool. Each pattern is encoded as a byte 41 (packing 2 per byte would be possible but the space savings would not be signficant). 42 43 As much as possible of the file is represented as 32 bit integers, as that is especially 44 efficent to access. All are little-endian (this could be revised if the code ever needs 45 to be ported to big-endian systems). 46 47 ## Header 48 49 ``` 50 uint32_t magic == 0x62ad7968 51 uint32_t version = 0 52 uint32_t alphabet_offset (in bytes) 53 uint32_t trie_offset (in bytes) 54 uint32_t pattern_offset (in bytes) 55 uint32_t file_size (in bytes) 56 ``` 57 58 Offsets are from the front of the file, and in bytes. 59 60 ## Alphabet 61 62 The alphabet table comes in two versions. The first is well suited to dense Unicode 63 ranges and is limited to 256. The second is more general, but lookups will be slower. 64 65 ### Alphabet, direct version 66 67 ``` 68 uint32_t version = 0 69 uint32_t min_codepoint 70 uint32_t max_codepoint (exclusive) 71 uint8_t[] data 72 ``` 73 74 The size of the data array is max_codepoint - min_codepoint. 0 represents an unmapped 75 character. Note that, in the current implementation, automatic hyphenation is disabled 76 for any word containing an unmapped character. 77 78 In general, pad bytes follow this table, aligning the next table to a 4-byte boundary. 79 80 ### Alphabet, general version 81 82 ``` 83 uint32_t version = 1 84 uint32_t n_entries 85 uint32_t[n_entries] data 86 ``` 87 88 Each element in the data table is `(codepoint << 11) | value`. Note that this is 89 restricted to 11 bits (2048 possible values). The largest known current value is 483 90 (for Sanskrit). 91 92 The entries are sorted by codepoint, to facilitate binary search. Another reasonable 93 implementation for consumers of the data would be to build a hash table at load time. 94 95 ## Trie 96 97 ``` 98 uint32_t version = 0 99 uint32_t char_mask 100 uint32_t link_shift 101 uint32_t link_mask 102 uint32_t pattern_shift 103 uint32_t n_entries 104 uint32_t[n_entries] data 105 ``` 106 107 Each element in the data table is `(pattern << pattern_shift) | (link << link_shift) | char`. 108 109 All known pattern tables fit in 32 bits total. If this is exceeded, there is a fairly 110 straightforward tweak, where each node occupies a slot by itself (as opposed to sharing 111 it with edge slots), which would require very minimal changes to the implementation (TODO 112 present in more detail). 113 114 ## Pattern 115 116 ``` 117 uint32_t version = 0 118 uint32_t n_entries 119 uint32_t pattern_offset (in bytes) 120 uint32_t pattern_size (in bytes) 121 uint32_t[n_entries] data 122 uint8_t[] pattern_buf 123 ``` 124 125 Each element in data table is `(len << 26) | (shift << 20) | offset`, where an offset of 0 126 points to the first byte of pattern_buf. 127 128 Generally pattern_offset is `16 + 4 * n_entries`. 129 130 For example, 'a4m5ato' would be represented as `[4, 5, 0, 0, 0]`, then len = 2, shift = 3, and 131 offset points to [4, 5] in the pattern buffer. 132 133 Future extension: additional data representing nonstandard hyphenation. See 134 [Automatic non-standard hyphenation in OpenOffice.org](https://www.tug.org/TUGboat/tb27-1/tb86nemeth.pdf) 135 for more information about that issue. 136
