Binary format for hyphenation patterns

In the current state, hyphenation in all languages than Sanskrit seems
to work (case-folding edge cases). Thus, we just disable Sanskrit.
Packed tries are implemented, but not the finite state machine
(space/speed tradeoff).

This commit contains a throw-away test app, which runs on the host.
I think I want to replace it with unit tests, but I'm including it in
the CL because it's useful during development.

Bug: 21562869
Bug: 21826930
Bug: 23317038
Bug: 23317904

Change-Id: I7479a565a4a062fa319651c2c14c0fa18c5ceaea

app/Android.mk

+# Copyright (C) 2015 The Android Open Source Project
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#      http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# see how_to_run.txt for instructions on running these tests
+
+LOCAL_PATH := $(call my-dir)
+
+include $(CLEAR_VARS)
+
+LOCAL_MODULE := hyphtool
+LOCAL_MODULE_TAGS := optional
+
+LOCAL_STATIC_LIBRARIES := libminikin_host
+
+# Shared libraries which are dependencies of minikin; these are not automatically
+# pulled in by the build system (and thus sadly must be repeated).
+
+LOCAL_SHARED_LIBRARIES := \
+    liblog \
+    libicuuc-host
+
+LOCAL_SRC_FILES += \
+    HyphTool.cpp
+
+include $(BUILD_HOST_EXECUTABLE)

app/HyphTool.cpp

+#include <stdio.h>
+#include <sys/stat.h>
+#include <string.h>
+
+#include "utils/Log.h"
+
+#include <vector>
+#include <minikin/Hyphenator.h>
+
+using android::Hyphenator;
+
+Hyphenator* loadHybFile(const char* fn) {
+    struct stat statbuf;
+    int status = stat(fn, &statbuf);
+    if (status < 0) {
+        fprintf(stderr, "error opening %s\n", fn);
+        return nullptr;
+    }
+    size_t size = statbuf.st_size;
+    FILE* f = fopen(fn, "rb");
+    if (f == NULL) {
+        fprintf(stderr, "error opening %s\n", fn);
+        return nullptr;
+    }
+    uint8_t* buf = new uint8_t[size];
+    size_t read_size = fread(buf, 1, size, f);
+    if (read_size < size) {
+        fprintf(stderr, "error reading %s\n", fn);
+        delete[] buf;
+        return nullptr;
+    }
+    return Hyphenator::loadBinary(buf);
+}
+
+int main(int argc, char** argv) {
+    Hyphenator* hyph = loadHybFile("/tmp/en.hyb");  // should also be configurable
+    std::vector<uint8_t> result;
+    std::vector<uint16_t> word;
+    if (argc < 2) {
+        fprintf(stderr, "usage: hyphtool word\n");
+        return 1;
+    }
+    char* asciiword = argv[1];
+    size_t len = strlen(asciiword);
+    for (size_t i = 0; i < len; i++) {
+        uint32_t c = asciiword[i];
+        if (c == '-') {
+            c = 0x00AD;
+        }
+        // ASCII (or possibly ISO Latin 1), but kinda painful to do utf conversion :(
+        word.push_back(c);
+    }
+    hyph->hyphenate(&result, word.data(), word.size());
+    for (size_t i = 0; i < len; i++) {
+        if (result[i] != 0) {
+            printf("-");
+        }
+        printf("%c", word[i]);
+    }
+    printf("\n");
+    return 0;
+}

doc/hyb_file_format.md

+# Hyb (hyphenation pattern binary) file format
+
+The hyb file format is how hyphenation patterns are stored in the system image.
+
+Goals include:
+
+* Concise (system image space is at a premium)
+* Usable when mmap'ed, so it doesn't take significant physical RAM
+* Fast to compute
+* Simple
+
+It is _not_ intended as an interchange format, so there is no attempt to make the format
+extensible or facilitate backward and forward compatibility.
+
+Further, at some point we will probably pack patterns for multiple languages into a single
+physical file, to reduce number of open mmap'ed files. This document doesn't cover that.
+
+## Theoretical basis
+
+At heart, the file contains packed tries with suffix compression, actually quite similar
+to the implementation in TeX.
+
+The file contains three sections. The first section represents the "alphabet," including
+case folding. It is effectively a map from Unicode code point to a small integer.
+
+The second section contains the trie in packed form. It is an array of 3-tuples, packed
+into a 32 bit integer. Each (suffix-compressed) trie node has a unique index within this
+array, and the pattern field in the tuple is the pattern for that node. Further, each edge
+in the trie has an entry in the array, and the character and link fields in the tuple
+represent the label and destination of the edge. The packing strategy is as in
+[Word Hy-phen-a-tion by Com-put-er](http://www.tug.org/docs/liang/liang-thesis.pdf) by
+Franklin Mark Liang.
+
+The trie representation is similar but not identical to the "double-array trie".
+The fundamental operation of lookup of the edge from `s` to `t` with label `c` is
+to compare `c == character[s + c]`, and if so, `t = link[s + c]`.
+
+The third section contains the pattern strings. This section is in two parts: first,
+an array with a 3-tuple for each pattern (length, number of trailing 0's, and offset
+into the string pool); and second, the string pool. Each pattern is encoded as a byte
+(packing 2 per byte would be possible but the space savings would not be signficant).
+
+As much as possible of the file is represented as 32 bit integers, as that is especially
+efficent to access. All are little-endian (this could be revised if the code ever needs
+to be ported to big-endian systems).
+
+## Header
+
+```
+uint32_t magic == 0x62ad7968
+uint32_t version = 0
+uint32_t alphabet_offset (in bytes)
+uint32_t trie_offset (in bytes)
+uint32_t pattern_offset (in bytes)
+uint32_t file_size (in bytes)
+```
+
+Offsets are from the front of the file, and in bytes.
+
+## Alphabet
+
+The alphabet table comes in two versions. The first is well suited to dense Unicode
+ranges and is limited to 256. The second is more general, but lookups will be slower.
+
+### Alphabet, direct version
+
+```
+uint32_t version = 0
+uint32_t min_codepoint
+uint32_t max_codepoint (exclusive)
+uint8_t[] data
+```
+
+The size of the data array is max_codepoint - min_codepoint. 0 represents an unmapped
+character. Note that, in the current implementation, automatic hyphenation is disabled
+for any word containing an unmapped character.
+
+In general, pad bytes follow this table, aligning the next table to a 4-byte boundary.
+
+### Alphabet, general version
+
+```
+uint32_t version = 1
+uint32_t n_entries
+uint32_t[n_entries] data
+```
+
+Each element in the data table is `(codepoint << 11) | value`. Note that this is
+restricted to 11 bits (2048 possible values). The largest known current value is 483
+(for Sanskrit).
+
+The entries are sorted by codepoint, to facilitate binary search. Another reasonable
+implementation for consumers of the data would be to build a hash table at load time.
+
+## Trie
+
+```
+uint32_t version = 0
+uint32_t char_mask
+uint32_t link_shift
+uint32_t link_mask
+uint32_t pattern_shift
+uint32_t n_entries
+uint32_t[n_entries] data
+```
+
+Each element in the data table is `(pattern << pattern_shift) | (link << link_shift) | char`.
+
+All known pattern tables fit in 32 bits total. If this is exceeded, there is a fairly
+straightforward tweak, where each node occupies a slot by itself (as opposed to sharing
+it with edge slots), which would require very minimal changes to the implementation (TODO
+present in more detail).
+
+## Pattern
+
+```
+uint32_t version = 0
+uint32_t n_entries
+uint32_t pattern_offset (in bytes)
+uint32_t pattern_size (in bytes)
+uint32_t[n_entries] data
+uint8_t[] pattern_buf
+```
+
+Each element in data table is `(len << 26) | (shift << 20) | offset`, where an offset of 0
+points to the first byte of pattern_buf.
+
+Generally pattern_offset is `16 + 4 * n_entries`.
+
+For example, 'a4m5ato' would be represented as `[4, 5, 0, 0, 0]`, then len = 2, shift = 3, and
+offset points to [4, 5] in the pattern buffer.
+
+Future extension: additional data representing nonstandard hyphenation. See
+[Automatic non-standard hyphenation in OpenOffice.org](https://www.tug.org/TUGboat/tb27-1/tb86nemeth.pdf)
+for more information about that issue.

include/minikin/Hyphenator.h

@@ -26,11 +26,8 @@
 
 namespace android {
 
-class Trie {
-public:
-    std::vector<uint8_t> result;
-    std::unordered_map<uint16_t, Trie> succ;
-};
+// hyb file header; implementation details are in the .cpp file
+struct Header;
 
 class Hyphenator {
 public:
@@ -44,19 +41,43 @@ public:
     // Example: word is "hyphen", result is [0 0 1 0 0 0], corresponding to "hy-phen".
     void hyphenate(std::vector<uint8_t>* result, const uint16_t* word, size_t len);
 
+    // pattern data is in binary format, as described in doc/hyb_file_format.md. Note:
+    // the caller is responsible for ensuring that the lifetime of the pattern data is
+    // at least as long as the Hyphenator object.
+
+    // Note: nullptr is valid input, in which case the hyphenator only processes soft hyphens
+    static Hyphenator* loadBinary(const uint8_t* patternData);
+
 private:
-    void addPattern(const uint16_t* pattern, size_t size);
+    // apply soft hyphens only, ignoring patterns
+    void hyphenateSoft(uint8_t* result, const uint16_t* word, size_t len);
 
-    void hyphenateSoft(std::vector<uint8_t>* result, const uint16_t* word, size_t len);
+    // try looking up word in alphabet table, return false if any code units fail to map
+    // Note that this methor writes len+2 entries into alpha_codes (including start and stop)
+    bool alphabetLookup(uint16_t* alpha_codes, const uint16_t* word, size_t len);
+
+    // calculate hyphenation from patterns, assuming alphabet lookup has already been done
+    void hyphenateFromCodes(uint8_t* result, const uint16_t* codes, size_t len);
 
     // TODO: these should become parameters, as they might vary by locale, screen size, and
     // possibly explicit user control.
     static const int MIN_PREFIX = 2;
     static const int MIN_SUFFIX = 3;
 
-    Trie root;
+    // See also LONGEST_HYPHENATED_WORD in LineBreaker.cpp. Here the constant is used so
+    // that temporary buffers can be stack-allocated without waste, which is a slightly
+    // different use case. It measures UTF-16 code units.
+    static const size_t MAX_HYPHENATED_SIZE = 64;
+
+    const uint8_t* patternData;
+
+    // accessors for binary data
+    const Header* getHeader() const {
+        return reinterpret_cast<const Header*>(patternData);
+    }
+
 };
 
 }  // namespace android
 
-#endif   // MINIKIN_HYPHENATOR_H
\ No newline at end of file
+#endif   // MINIKIN_HYPHENATOR_H

libs/minikin/Android.mk

@@ -63,3 +63,17 @@ LOCAL_C_INCLUDES := $(minikin_c_includes)
 LOCAL_SHARED_LIBRARIES := $(minikin_shared_libraries)
 
 include $(BUILD_STATIC_LIBRARY)
+
+include $(CLEAR_VARS)
+
+# Reduced library (currently just hyphenation) for host
+
+LOCAL_MODULE := libminikin_host
+LOCAL_MODULE_TAGS := optional
+LOCAL_EXPORT_C_INCLUDE_DIRS := frameworks/minikin/include
+LOCAL_C_INCLUDES := $(minikin_c_includes)
+LOCAL_SHARED_LIBRARIES := liblog libicuuc-host
+
+LOCAL_SRC_FILES := Hyphenator.cpp
+
+include $(BUILD_HOST_STATIC_LIBRARY)

libs/minikin/Hyphenator.cpp

@@ -34,130 +34,202 @@ namespace android {
 
 static const uint16_t CHAR_SOFT_HYPHEN = 0x00AD;
 
-void Hyphenator::addPattern(const uint16_t* pattern, size_t size) {
-    vector<uint16_t> word;
-    vector<uint8_t> result;
-
-    // start by parsing the Liang-format pattern into a word and a result vector, the
-    // vector right-aligned but without leading zeros. Examples:
-    // a1bc2d -> abcd [1, 0, 2, 0]
-    // abc1 -> abc [1]
-    // 1a2b3c4d5 -> abcd [1, 2, 3, 4, 5]
-    bool lastWasLetter = false;
-    bool haveSeenNumber = false;
-    for (size_t i = 0; i < size; i++) {
-        uint16_t c = pattern[i];
-        if (isdigit(c)) {
-            result.push_back(c - '0');
-            lastWasLetter = false;
-            haveSeenNumber = true;
-        } else {
-            word.push_back(c);
-            if (lastWasLetter && haveSeenNumber) {
-                result.push_back(0);
-            }
-            lastWasLetter = true;
-        }
+// The following are structs that correspond to tables inside the hyb file format
+
+struct AlphabetTable0 {
+    uint32_t version;
+    uint32_t min_codepoint;
+    uint32_t max_codepoint;
+    uint8_t data[1];  // actually flexible array, size is known at runtime
+};
+
+struct AlphabetTable1 {
+    uint32_t version;
+    uint32_t n_entries;
+    uint32_t data[1]; // actually flexible array, size is known at runtime
+
+    static uint32_t codepoint(uint32_t entry) { return entry >> 11; }
+    static uint32_t value(uint32_t entry) { return entry & 0x7ff; }
+};
+
+struct Trie {
+    uint32_t version;
+    uint32_t char_mask;
+    uint32_t link_shift;
+    uint32_t link_mask;
+    uint32_t pattern_shift;
+    uint32_t n_entries;
+    uint32_t data[1];  // actually flexible array, size is known at runtime
+};
+
+struct Pattern {
+    uint32_t version;
+    uint32_t n_entries;
+    uint32_t pattern_offset;
+    uint32_t pattern_size;
+    uint32_t data[1];  // actually flexible array, size is known at runtime
+
+    // accessors
+    static uint32_t len(uint32_t entry) { return entry >> 26; }
+    static uint32_t shift(uint32_t entry) { return (entry >> 20) & 0x3f; }
+    const uint8_t* buf(uint32_t entry) const {
+        return reinterpret_cast<const uint8_t*>(this) + pattern_offset + (entry & 0xfffff);
+    }
+};
+
+struct Header {
+    uint32_t magic;
+    uint32_t version;
+    uint32_t alphabet_offset;
+    uint32_t trie_offset;
+    uint32_t pattern_offset;
+    uint32_t file_size;
+
+    // accessors
+    const uint8_t* bytes() const { return reinterpret_cast<const uint8_t*>(this); }
+    uint32_t alphabetVersion() const {
+        return *reinterpret_cast<const uint32_t*>(bytes() + alphabet_offset);
     }
-    if (lastWasLetter) {
-        result.push_back(0);
+    const AlphabetTable0* alphabetTable0() const {
+        return reinterpret_cast<const AlphabetTable0*>(bytes() + alphabet_offset);
     }
-    Trie* t = &root;
-    for (size_t i = 0; i < word.size(); i++) {
-        t = &t->succ[word[i]];
+    const AlphabetTable1* alphabetTable1() const {
+        return reinterpret_cast<const AlphabetTable1*>(bytes() + alphabet_offset);
+    }
+    const Trie* trieTable() const {
+        return reinterpret_cast<const Trie*>(bytes() + trie_offset);
+    }
+    const Pattern* patternTable() const {
+        return reinterpret_cast<const Pattern*>(bytes() + pattern_offset);
+    }
+};
+
+Hyphenator* Hyphenator::loadBinary(const uint8_t* patternData) {
+    Hyphenator* result = new Hyphenator;
+    result->patternData = patternData;
+    return result;
+}
+
+void Hyphenator::hyphenate(vector<uint8_t>* result, const uint16_t* word, size_t len) {
+    result->clear();
+    result->resize(len);
+    const size_t paddedLen = len + 2;  // start and stop code each count for 1
+    if (patternData != nullptr &&
+            (int)len >= MIN_PREFIX + MIN_SUFFIX && paddedLen <= MAX_HYPHENATED_SIZE) {
+        uint16_t alpha_codes[MAX_HYPHENATED_SIZE];
+        if (alphabetLookup(alpha_codes, word, len)) {
+            hyphenateFromCodes(result->data(), alpha_codes, paddedLen);
+            return;
+        }
+        // TODO: try NFC normalization
+        // TODO: handle non-BMP Unicode (requires remapping of offsets)
     }
-    t->result = result;
+    hyphenateSoft(result->data(), word, len);
 }
 
 // If any soft hyphen is present in the word, use soft hyphens to decide hyphenation,
 // as recommended in UAX #14 (Use of Soft Hyphen)
-void Hyphenator::hyphenateSoft(vector<uint8_t>* result, const uint16_t* word, size_t len) {
-    (*result)[0] = 0;
+void Hyphenator::hyphenateSoft(uint8_t* result, const uint16_t* word, size_t len) {
+    result[0] = 0;
     for (size_t i = 1; i < len; i++) {
-        (*result)[i] = word[i - 1] == CHAR_SOFT_HYPHEN;
-    }
+        result[i] = word[i - 1] == CHAR_SOFT_HYPHEN;
+     }
 }
 
-void Hyphenator::hyphenate(vector<uint8_t>* result, const uint16_t* word, size_t len) {
-    result->clear();
-    result->resize(len);
-    if (len < MIN_PREFIX + MIN_SUFFIX) return;
-    size_t maxOffset = len - MIN_SUFFIX + 1;
-    for (size_t i = 0; i < len + 1; i++) {
-        const Trie* node = &root;
-        for (size_t j = i; j < len + 2; j++) {
-            uint16_t c;
-            if (j == 0 || j == len + 1) {
-                c = '.';  // word boundary character in pattern data files
-            } else {
-                c = word[j - 1];
-                if (c == CHAR_SOFT_HYPHEN) {
-                    hyphenateSoft(result, word, len);
-                    return;
-                }
-                // TODO: This uses ICU's simple character to character lowercasing, which ignores
-                // the locale, and ignores cases when lowercasing a character results in more than
-                // one character. It should be fixed to consider the locale (in order for it to work
-                // correctly for Turkish and Azerbaijani), as well as support one-to-many, and
-                // many-to-many case conversions (including non-BMP cases).
-                if (c < 0x00C0) { // U+00C0 is the lowest uppercase non-ASCII character
-                    // Convert uppercase ASCII to lowercase ASCII, but keep other characters as-is
-                    if (0x0041 <= c && c <= 0x005A) {
-                        c += 0x0020;
-                    }
-                } else {
-                    c = u_tolower(c);
-                }
+bool Hyphenator::alphabetLookup(uint16_t* alpha_codes, const uint16_t* word, size_t len) {
+    const Header* header = getHeader();
+    // TODO: check header magic
+    uint32_t alphabetVersion = header->alphabetVersion();
+    if (alphabetVersion == 0) {
+        const AlphabetTable0* alphabet = header->alphabetTable0();
+        uint32_t min_codepoint = alphabet->min_codepoint;
+        uint32_t max_codepoint = alphabet->max_codepoint;
+        alpha_codes[0] = 0;  // word start
+        for (size_t i = 0; i < len; i++) {
+            uint16_t c = word[i];
+            if (c < min_codepoint || c >= max_codepoint) {
+                return false;
+            }
+            uint8_t code = alphabet->data[c - min_codepoint];
+            if (code == 0) {
+                return false;
+            }
+            alpha_codes[i + 1] = code;
+        }
+        alpha_codes[len + 1] = 0;  // word termination
+        return true;
+    } else if (alphabetVersion == 1) {
+        const AlphabetTable1* alphabet = header->alphabetTable1();
+        size_t n_entries = alphabet->n_entries;
+        const uint32_t* begin = alphabet->data;
+        const uint32_t* end = begin + n_entries;
+        alpha_codes[0] = 0;
+        for (size_t i = 0; i < len; i++) {
+            uint16_t c = word[i];
+            auto p = std::lower_bound(begin, end, c << 11);
+            if (p == end) {
+                return false;
             }
-            auto search = node->succ.find(c);
-            if (search != node->succ.end()) {
-                node = &search->second;
+            uint32_t entry = *p;
+            if (AlphabetTable1::codepoint(entry) != c) {
+                return false;
+            }
+            alpha_codes[i + 1] = AlphabetTable1::value(entry);
+        }
+        alpha_codes[len + 1] = 0;
+        return true;
+    }
+    return false;
+}
+
+/**
+ * Internal implementation, after conversion to codes. All case folding and normalization
+ * has been done by now, and all characters have been found in the alphabet.
+ * Note: len here is the padded length including 0 codes at start and end.
+ **/
+void Hyphenator::hyphenateFromCodes(uint8_t* result, const uint16_t* codes, size_t len) {
+    const Header* header = getHeader();
+    const Trie* trie = header->trieTable();
+    const Pattern* pattern = header->patternTable();
+    uint32_t char_mask = trie->char_mask;
+    uint32_t link_shift = trie->link_shift;
+    uint32_t link_mask = trie->link_mask;
+    uint32_t pattern_shift = trie->pattern_shift;
+    size_t maxOffset = len - MIN_SUFFIX - 1;
+    for (size_t i = 0; i < len - 1; i++) {
+        uint32_t node = 0;  // index into Trie table
+        for (size_t j = i; j < len; j++) {
+            uint16_t c = codes[j];
+            uint32_t entry = trie->data[node + c];
+            if ((entry & char_mask) == c) {
+                node = (entry & link_mask) >> link_shift;
             } else {
                 break;
             }
-            if (!node->result.empty()) {
-                int resultLen = node->result.size();
-                int offset = j + 1 - resultLen;
+            uint32_t pat_ix = trie->data[node] >> pattern_shift;
+            // pat_ix contains a 3-tuple of length, shift (number of trailing zeros), and an offset
+            // into the buf pool. This is the pattern for the substring (i..j) we just matched,
+            // which we combine (via point-wise max) into the result vector.
+            if (pat_ix != 0) {
+                uint32_t pat_entry = pattern->data[pat_ix];
+                int pat_len = Pattern::len(pat_entry);
+                int pat_shift = Pattern::shift(pat_entry);
+                const uint8_t* pat_buf = pattern->buf(pat_entry);
+                int offset = j + 1 - (pat_len + pat_shift);
+                // offset is the index within result that lines up with the start of pat_buf
                 int start = std::max(MIN_PREFIX - offset, 0);
-                int end = std::min(resultLen, (int)maxOffset - offset);
-                // TODO performance: this inner loop can profitably be optimized
+                int end = std::min(pat_len, (int)maxOffset - offset);
                 for (int k = start; k < end; k++) {
-                    (*result)[offset + k] = std::max((*result)[offset + k], node->result[k]);
-                }
-#if 0
-                // debug printing of matched patterns
-                std::string dbg;
-                for (size_t k = i; k <= j + 1; k++) {
-                    int off = k - j - 2 + resultLen;
-                    if (off >= 0 && node->result[off] != 0) {
-                        dbg.push_back((char)('0' + node->result[off]));
-                    }
-                    if (k < j + 1) {
-                        uint16_t c = (k == 0 || k == len + 1) ? '.' : word[k - 1];
-                        dbg.push_back((char)c);
-                    }
+                    result[offset + k] = std::max(result[offset + k], pat_buf[k]);
                 }
-                ALOGD("%d:%d %s", i, j, dbg.c_str());
-#endif
             }
         }
     }
     // Since the above calculation does not modify values outside
     // [MIN_PREFIX, len - MIN_SUFFIX], they are left as 0.
     for (size_t i = MIN_PREFIX; i < maxOffset; i++) {
-        (*result)[i] &= 1;
+        result[i] &= 1;
     }
 }
 
-Hyphenator* Hyphenator::load(const uint16_t *patternData, size_t size) {
-    Hyphenator* result = new Hyphenator;
-    for (size_t i = 0; i < size; i++) {
-        size_t end = i;
-        while (patternData[end] != '\n') end++;
-        result->addPattern(patternData + i, end - i);
-        i = end;
-    }
-    return result;
-}
-
 }  // namespace android

tools/mk_hyb_file.py

+#!/usr/bin/env python
+
+# Copyright (C) 2015 The Android Open Source Project
+#
+# Licensed under the Apache License, Version 2.0 (the 'License');
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#      http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an 'AS IS' BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+Convert hyphen files in standard TeX format (a trio of pat, chr, and hyp)
+into binary format. See doc/hyb_file_format.md for more information.
+
+Usage: mk_hyb_file.py [-v] hyph-foo.pat.txt hyph-foo.hyb
+
+Optional -v parameter turns on verbose debugging.
+
+"""
+
+from __future__ import print_function
+
+import io
+import sys
+import struct
+import math
+import getopt
+
+
+VERBOSE = False
+
+
+if sys.version_info[0] >= 3:
+    def unichr(x):
+        return chr(x)
+
+
+# number of bits required to represent numbers up to n inclusive
+def num_bits(n):
+    return 1 + int(math.log(n, 2)) if n > 0 else 0
+
+
+class Node:
+
+    def __init__(self):
+        self.succ = {}
+        self.res = None
+        self.fsm_pat = None
+        self.fail = None
+
+
+# List of free slots, implemented as doubly linked list
+class Freelist:
+
+    def __init__(self):
+        self.first = None
+        self.last = None
+        self.pred = []
+        self.succ = []
+
+    def grow(self):
+        this = len(self.pred)
+        self.pred.append(self.last)
+        self.succ.append(None)
+        if self.last is None:
+            self.first = this
+        else:
+            self.succ[self.last] = this
+        self.last = this
+
+    def next(self, cursor):
+        if cursor == 0:
+            cursor = self.first
+        if cursor is None:
+            self.grow()
+            result = self.last
+        else:
+            result = cursor
+        return result, self.succ[result]
+
+    def is_free(self, ix):
+        while ix >= len(self.pred):
+            self.grow()
+        return self.pred[ix] != -1
+
+    def use(self, ix):
+        if self.pred[ix] is None:
+            self.first = self.succ[ix]
+        else:
+            self.succ[self.pred[ix]] = self.succ[ix]
+        if self.succ[ix] is None:
+            self.last = self.pred[ix]
+        else:
+            self.pred[self.succ[ix]] = self.pred[ix]
+        if self.pred[ix] == -1:
+            assert self.pred[ix] != -1, 'double free!'
+        self.pred[ix] = -1
+
+
+def combine(a, b):
+    if a is None: return b
+    if b is None: return a
+    if len(b) < len(a): a, b = b, a
+    res = b[:len(b) - len(a)]
+    for i in range(len(a)):
+        res.append(max(a[i], b[i + len(b) - len(a)]))
+    return res
+
+
+def trim(pattern):
+    for ix in range(len(pattern)):
+        if pattern[ix] != 0:
+            return pattern[ix:]
+
+
+def pat_to_binary(pattern):
+    return b''.join(struct.pack('B', x) for x in pattern)
+
+
+class Hyph:
+
+    def __init__(self):
+        self.root = Node()
+        self.root.str = '<root>'
+        self.node_list = [self.root]
+
+    # Add a pattern (word fragment with numeric codes, such as ".ad4der")
+    def add_pat(self, pat):
+        lastWasLetter = False
+        haveSeenNumber = False
+        result = []
+        word = ''
+        for c in pat:
+            if c.isdigit():
+                result.append(int(c))
+                lastWasLetter = False
+                haveSeenNumber = True
+            else:
+                word += c
+                if lastWasLetter and haveSeenNumber:
+                    result.append(0)
+                lastWasLetter = True
+        if lastWasLetter:
+            result.append(0)
+
+        self.add_word_res(word, result)
+
+    # Add an exception (word with hyphens, such as "ta-ble")
+    def add_exception(self, hyph_word):
+        res = []
+        word = ['.']
+        need_10 = False
+        for c in hyph_word:
+            if c == '-':
+                res.append(11)
+                need_10 = False
+            else:
+                if need_10:
+                    res.append(10)
+                word.append(c)
+                need_10 = True
+        word.append('.')
+        res.append(0)
+        res.append(0)
+        if VERBOSE:
+            print(word, res)
+        self.add_word_res(''.join(word), res)
+
+    def add_word_res(self, word, result):
+        if VERBOSE:
+            print(word, result)
+
+        t = self.root
+        s = ''
+        for c in word:
+            s += c
+            if c not in t.succ:
+                new_node = Node()
+                new_node.str = s
+                self.node_list.append(new_node)
+                t.succ[c] = new_node
+            t = t.succ[c]
+        t.res = result
+
+    def pack(self, node_list, ch_map, use_node=False):
+        size = 0
+        self.node_map = {}
+        nodes = Freelist()
+        edges = Freelist()
+        edge_start = 1 if use_node else 0
+        for node in node_list:
+            succ = sorted([ch_map[c] + edge_start for c in node.succ.keys()])
+            if len(succ):
+                cursor = 0
+                while True:
+                    edge_ix, cursor = edges.next(cursor)
+                    ix = edge_ix - succ[0]
+                    if (ix >= 0 and nodes.is_free(ix) and
+                            all(edges.is_free(ix + s) for s in succ) and
+                            ((not use_node) or edges.is_free(ix))):
+                        break
+            elif use_node:
+                ix, _ = edges.next(0)
+                nodes.is_free(ix)  # actually don't need nodes at all when use_node,
+                # but keep it happy
+            else:
+                ix, _ = nodes.next(0)
+            node.ix = ix
+            self.node_map[ix] = node
+            nodes.use(ix)
+            size = max(size, ix)
+            if use_node:
+                edges.use(ix)
+            for s in succ:
+                edges.use(ix + s)
+        size += max(ch_map.values()) + 1
+        return size
+
+    # return list of nodes in bfs order
+    def bfs(self, ch_map):
+        result = [self.root]
+        ix = 0
+        while ix < len(result):
+            node = result[ix]
+            node.bfs_ix = ix
+            mapped = {}
+            for c, next in node.succ.items():
+                assert ch_map[c] not in mapped, 'duplicate edge ' + node.str + ' ' + hex(ord(c))
+                mapped[ch_map[c]] = next
+            for i in sorted(mapped.keys()):
+                result.append(mapped[i])
+            ix += 1
+        self.bfs_order = result
+        return result
+
+    # suffix compression - convert the trie into an acyclic digraph, merging nodes when
+    # the subtries are identical
+    def dedup(self):
+        uniques = []
+        dupmap = {}
+        dedup_ix = [0] * len(self.bfs_order)
+        for ix in reversed(range(len(self.bfs_order))):
+            # construct string representation of node
+            node = self.bfs_order[ix]
+            if node.res is None:
+                s = ''
+            else:
+                s = ''.join(str(c) for c in node.res)
+            for c in sorted(node.succ.keys()):
+                succ = node.succ[c]
+                s += ' ' + c + str(dedup_ix[succ.bfs_ix])
+            if s in dupmap:
+                dedup_ix[ix] = dupmap[s]
+            else:
+                uniques.append(node)
+                dedup_ix[ix] = ix
+            dupmap[s] = dedup_ix[ix]
+        uniques.reverse()
+        print(len(uniques), 'unique nodes,', len(self.bfs_order), 'total')
+        return dedup_ix, uniques
+
+
+# load the ".pat" file, which contains patterns such as a1b2c3
+def load(fn):
+    hyph = Hyph()
+    with io.open(fn, encoding='UTF-8') as f:
+        for l in f:
+            pat = l.strip()
+            hyph.add_pat(pat)
+    return hyph
+
+
+# load the ".chr" file, which contains the alphabet and case pairs, eg "aA", "bB" etc.
+def load_chr(fn):
+    ch_map = {'.': 0}
+    with io.open(fn, encoding='UTF-8') as f:
+        for i, l in enumerate(f):
+            l = l.strip()
+            if len(l) > 2:
+                # lowercase maps to multi-character uppercase sequence, ignore uppercase for now
+                l = l[:1]
+            else:
+                assert len(l) == 2, 'expected 2 chars in chr'
+            for c in l:
+                ch_map[c] = i + 1
+    return ch_map
+
+
+# load exceptions with explicit hyphens
+def load_hyp(hyph, fn):
+    with io.open(fn, encoding='UTF-8') as f:
+        for l in f:
+            hyph.add_exception(l.strip())
+
+
+def generate_header(alphabet, trie, pattern):
+    alphabet_off = 6 * 4
+    trie_off = alphabet_off + len(alphabet)
+    pattern_off = trie_off + len(trie)
+    file_size = pattern_off + len(pattern)
+    data = [0x62ad7968, 0, alphabet_off, trie_off, pattern_off, file_size]
+    return struct.pack('<6I', *data)
+
+
+def generate_alphabet(ch_map):
+    ch_map = ch_map.copy()
+    del ch_map['.']
+    min_ch = ord(min(ch_map))
+    max_ch = ord(max(ch_map))
+    if max_ch - min_ch < 1024 and max(ch_map.values()) < 256:
+        # generate format 0
+        data = [0] * (max_ch - min_ch + 1)
+        for c, val in ch_map.items():
+            data[ord(c) - min_ch] = val
+        result = [struct.pack('<3I', 0, min_ch, max_ch + 1)]
+        for b in data:
+            result.append(struct.pack('<B', b))
+    else:
+        # generate format 1
+        assert max(ch_map.values()) < 2048, 'max number of unique characters exceeded'
+        result = [struct.pack('<2I', 1, len(ch_map))]
+        for c, val in sorted(ch_map.items()):
+            data = (ord(c) << 11) | val
+            result.append(struct.pack('<I', data))
+    binary = b''.join(result)
+    if len(binary) % 4 != 0:
+        binary += b'\x00' * (4 - len(binary) % 4)
+    return binary
+
+
+# assumes hyph structure has been packed, ie node.ix values have been set
+def generate_trie(hyph, ch_map, n_trie, dedup_ix, dedup_nodes, patmap):
+    ch_array = [0] * n_trie
+    link_array = [0] * n_trie
+    pat_array = [0] * n_trie
+    link_shift = num_bits(max(ch_map.values()))
+    char_mask = (1 << link_shift) - 1
+    pattern_shift = link_shift + num_bits(n_trie - 1)
+    link_mask = (1 << pattern_shift) - (1 << link_shift)
+    result = [struct.pack('<6I', 0, char_mask, link_shift, link_mask, pattern_shift, n_trie)]
+
+    for node in dedup_nodes:
+        ix = node.ix
+        if node.res is not None:
+            pat_array[ix] = patmap[pat_to_binary(node.res)]
+        for c, next in node.succ.items():
+            c_num = ch_map[c]
+            link_ix = ix + c_num
+            ch_array[link_ix] = c_num
+            if dedup_ix is None:
+                dedup_next = next
+            else:
+                dedup_next = hyph.bfs_order[dedup_ix[next.bfs_ix]]
+            link_array[link_ix] = dedup_next.ix
+
+    for i in range(n_trie):
+        #print((pat_array[i], link_array[i], ch_array[i]))
+        packed = (pat_array[i] << pattern_shift) | (link_array[i] << link_shift) | ch_array[i]
+        result.append(struct.pack('<I', packed))
+    return b''.join(result)
+
+
+def generate_pattern(pats):
+    pat_array = [0]
+    patmap = {b'': 0}
+
+    raw_pat_array = []
+    raw_pat_size = 0
+    raw_patmap = {}
+
+    for pat in pats:
+        if pat is None:
+            continue
+        pat_str = pat_to_binary(pat)
+        if pat_str not in patmap:
+            shift = 0
+            while shift < len(pat) and pat[len(pat) - shift - 1] == 0:
+                shift += 1
+            rawpat = pat_str[:len(pat) - shift]
+            if rawpat not in raw_patmap:
+                raw_patmap[rawpat] = raw_pat_size
+                raw_pat_array.append(rawpat)
+                raw_pat_size += len(rawpat)
+            data = (len(rawpat) << 26) | (shift << 20) | raw_patmap[rawpat]
+            patmap[pat_str] = len(pat_array)
+            pat_array.append(data)
+    data = [0, len(pat_array), 16 + 4 * len(pat_array), raw_pat_size]
+    result = [struct.pack('<4I', *data)]
+    for x in pat_array:
+        result.append(struct.pack('<I', x))
+    result.extend(raw_pat_array)
+    return patmap, b''.join(result)
+
+
+def generate_hyb_file(hyph, ch_map, hyb_fn):
+    bfs = hyph.bfs(ch_map)
+    dedup_ix, dedup_nodes = hyph.dedup()
+    n_trie = hyph.pack(dedup_nodes, ch_map)
+    alphabet = generate_alphabet(ch_map)
+    patmap, pattern = generate_pattern([n.res for n in hyph.node_list])
+    trie = generate_trie(hyph, ch_map, n_trie, dedup_ix, dedup_nodes, patmap)
+    header = generate_header(alphabet, trie, pattern)
+
+    with open(hyb_fn, 'wb') as f:
+        f.write(header)
+        f.write(alphabet)
+        f.write(trie)
+        f.write(pattern)
+
+
+# Verify that the file contains the same lines as the lines argument, in arbitrary order
+def verify_file_sorted(lines, fn):
+    file_lines = [l.strip() for l in io.open(fn)]
+    line_set = set(lines)
+    file_set = set(file_lines)
+    if line_set == file_set:
+        return True
+    for line in line_set - file_set:
+        print(repr(line) + ' in reconstruction, not in file')
+    for line in file_set - line_set:
+        print(repr(line) + ' in file, not in reconstruction')
+    return False
+
+
+def map_to_chr(alphabet_map):
+    result = []
+    ch_map = {}
+    for val in alphabet_map.values():
+        chs = [ch for ch in alphabet_map if alphabet_map[ch] == val]
+        # non-cased characters (like Ethopic) are in both, matching chr file
+        lowercase = [ch for ch in chs if not ch.isupper()]
+        uppercase = [ch for ch in chs if not ch.islower()]
+        # print(val, `lowercase`, `uppercase`)
+        assert len(lowercase) == 1, 'expected 1 lowercase character'
+        assert 0 <= len(uppercase) <= 1, 'expected 0 or 1 uppercase character'
+        ch_map[val] = lowercase[0]
+        result.append(''.join(lowercase + uppercase))
+    ch_map[0] = '.'
+    return (ch_map, result)
+
+
+def get_pattern(pattern_data, ix):
+    pattern_offset = struct.unpack('<I', pattern_data[8:12])[0]
+    entry = struct.unpack('<I', pattern_data[16 + ix * 4: 16 + ix * 4 + 4])[0]
+    pat_len = entry >> 26
+    pat_shift = (entry >> 20) & 0x1f
+    offset = pattern_offset + (entry & 0xfffff)
+    return pattern_data[offset: offset + pat_len] + b'\0' * pat_shift
+
+
+def traverse_trie(ix, s, trie_data, ch_map, pattern_data, patterns, exceptions):
+    (char_mask, link_shift, link_mask, pattern_shift) = struct.unpack('<4I', trie_data[4:20])
+    node_entry = struct.unpack('<I', trie_data[24 + ix * 4: 24 + ix * 4 + 4])[0]
+    pattern = node_entry >> pattern_shift
+    if pattern:
+        result = []
+        is_exception = False
+        pat = get_pattern(pattern_data, pattern)
+        for i in range(len(s) + 1):
+            pat_off = i - 1 + len(pat) - len(s)
+            if pat_off < 0:
+                code = 0
+            else:
+                code = struct.unpack('B', pat[pat_off : pat_off + 1])[0]
+            if 1 <= code <= 9:
+                result.append('%d' % code)
+            elif code == 10:
+                is_exception = True
+            elif code == 11:
+                result.append('-')
+                is_exception = True
+            else:
+                assert code == 0, 'unexpected code'
+            if i < len(s):
+                result.append(s[i])
+        pat_str = ''.join(result)
+        #print(`pat_str`, `pat`)
+        if is_exception:
+            assert pat_str[0] == '.', "expected leading '.'"
+            assert pat_str[-1] == '.', "expected trailing '.'"
+            exceptions.append(pat_str[1:-1])  # strip leading and trailing '.'
+        else:
+            patterns.append(pat_str)
+    for ch in ch_map:
+        edge_entry = struct.unpack('<I', trie_data[24 + (ix + ch) * 4: 24 + (ix + ch) * 4 + 4])[0]
+        link = (edge_entry & link_mask) >> link_shift
+        if link != 0 and ch == (edge_entry & char_mask):
+            sch = s + ch_map[ch]
+            traverse_trie(link, sch, trie_data, ch_map, pattern_data, patterns, exceptions)
+
+
+# Verify the generated binary file by reconstructing the textual representations
+# from the binary hyb file, then checking that they're identical (mod the order of
+# lines within the file, which is irrelevant). This function makes assumptions that
+# are stronger than absolutely necessary (in particular, that the patterns are in
+# lowercase as defined by python islower).
+def verify_hyb_file(hyb_fn, pat_fn, chr_fn, hyp_fn):
+    with open(hyb_fn, 'rb') as f:
+        hyb_data = f.read()
+    header = hyb_data[0: 6 * 4]
+    (magic, version, alphabet_off, trie_off, pattern_off, file_size) = struct.unpack('<6I', header)
+    alphabet_data = hyb_data[alphabet_off:trie_off]
+    trie_data = hyb_data[trie_off:pattern_off]
+    pattern_data = hyb_data[pattern_off:file_size]
+
+    # reconstruct alphabet table
+    alphabet_version = struct.unpack('<I', alphabet_data[:4])[0]
+    alphabet_map = {}
+    if alphabet_version == 0:
+        (min_ch, max_ch) = struct.unpack('<2I', alphabet_data[4:12])
+        for ch in range(min_ch, max_ch):
+            offset = 12 + ch - min_ch
+            b = struct.unpack('B', alphabet_data[offset : offset + 1])[0]
+            if b != 0:
+                alphabet_map[unichr(ch)] = b
+    else:
+        assert alphabet_version == 1
+        n_entries = struct.unpack('<I', alphabet_data[4:8])[0]
+        for i in range(n_entries):
+            entry = struct.unpack('<I', alphabet_data[8 + 4 * i: 8 + 4 * i + 4])[0]
+            alphabet_map[unichr(entry >> 11)] = entry & 0x7ff
+
+    ch_map, reconstructed_chr = map_to_chr(alphabet_map)
+
+    # EXCEPTION for Armenian (hy), we don't really deal with the uppercase form of U+0587
+    if u'\u0587' in reconstructed_chr:
+        reconstructed_chr.remove(u'\u0587')
+        reconstructed_chr.append(u'\u0587\u0535\u0552')
+
+    assert verify_file_sorted(reconstructed_chr, chr_fn), 'alphabet table not verified'
+
+    # reconstruct trie
+    patterns = []
+    exceptions = []
+    traverse_trie(0, '', trie_data, ch_map, pattern_data, patterns, exceptions)
+    assert verify_file_sorted(patterns, pat_fn), 'pattern table not verified'
+    assert verify_file_sorted(exceptions, hyp_fn), 'exception table not verified'
+
+
+def main():
+    global VERBOSE
+    try:
+        opts, args = getopt.getopt(sys.argv[1:], 'v')
+    except getopt.GetoptError as err:
+        print(str(err))
+        sys.exit(1)
+    for o, _ in opts:
+        if o == '-v':
+            VERBOSE = True
+    pat_fn, out_fn = args
+    hyph = load(pat_fn)
+    if pat_fn.endswith('.pat.txt'):
+        chr_fn = pat_fn[:-8] + '.chr.txt'
+        ch_map = load_chr(chr_fn)
+        hyp_fn = pat_fn[:-8] + '.hyp.txt'
+        load_hyp(hyph, hyp_fn)
+        generate_hyb_file(hyph, ch_map, out_fn)
+        verify_hyb_file(out_fn, pat_fn, chr_fn, hyp_fn)
+
+if __name__ == '__main__':
+    main()