LibUnicode: Perform code point property lookups in constant time

We currently produce a single table for all categories of code point
properties (GeneralCategory, Script, etc.). Each row contains a field
indicating the range of code points to which that property applies. At
runtime, we then do a binary search through that table to decide if a
code point has a property.

This changes our approach to generate a 2-stage lookup table for each of
those categories. There is an in-depth explanation of these tables above
the new `create_code_point_tables` method. The end effect is that code
point property lookup is reduced from a binary search to constant-time
array lookups.

In total, this change:

    * Increases the size of libunicode.so from 2.7 MB to 2.9 MB.

    * Reduces the runtime of the new benchmark test case added here from
      3.576s to 1.020s (a 3.5x speedup).

    * In a profile of resizing a TextEditor window with a 3MB file open,
      the runtime of checking if a code point has a word break property
      reduces from ~81% to ~56%.
This commit is contained in:
Timothy Flynn 2023-07-20 06:35:59 -04:00 committed by Andreas Kling
parent 8f1d73abde
commit 0652cc48c0
Notes: sideshowbarker 2024-07-18 22:57:59 +09:00
2 changed files with 247 additions and 61 deletions

View File

@ -7,6 +7,7 @@
#include "GeneratorUtil.h"
#include <AK/AllOf.h>
#include <AK/Array.h>
#include <AK/BinarySearch.h>
#include <AK/CharacterTypes.h>
#include <AK/DeprecatedString.h>
#include <AK/Error.h>
@ -96,6 +97,18 @@ struct BlockName {
size_t name { 0 };
};
struct CodePointTables {
static constexpr auto MSB_COUNT = 16u;
static_assert(MSB_COUNT < 24u);
static constexpr auto LSB_COUNT = 24u - MSB_COUNT;
static constexpr auto LSB_MASK = NumericLimits<u32>::max() >> (NumericLimits<u32>::digits() - LSB_COUNT);
Vector<size_t> stage1;
Vector<size_t> stage2;
Vector<Vector<bool>> unique_properties;
};
struct UnicodeData {
UniqueStringStorage unique_strings;
@ -156,6 +169,14 @@ struct UnicodeData {
PropList grapheme_break_props;
PropList word_break_props;
PropList sentence_break_props;
CodePointTables general_category_tables;
CodePointTables property_tables;
CodePointTables script_tables;
CodePointTables script_extension_tables;
CodePointTables grapheme_break_tables;
CodePointTables word_break_tables;
CodePointTables sentence_break_tables;
};
static DeprecatedString sanitize_entry(DeprecatedString const& entry)
@ -1071,67 +1092,79 @@ static constexpr Array<@mapping_type@, @size@> s_@name@_mappings { {
return data.decomposition_mapping;
});
auto append_code_point_range_list = [&](DeprecatedString name, Vector<Unicode::CodePointRange> const& ranges) {
generator.set("name", name);
generator.set("size", DeprecatedString::number(ranges.size()));
generator.append(R"~~~(
static constexpr Array<CodePointRange, @size@> @name@ { {
auto append_code_point_tables = [&](StringView collection_snake, CodePointTables const& tables) -> ErrorOr<void> {
auto append_stage = [&](auto const& stage, auto name, auto type) -> ErrorOr<void> {
TRY(generator.set("name", TRY(String::formatted("{}_{}", collection_snake, name))));
TRY(generator.set("size", TRY(String::number(stage.size()))));
generator.set("type", type);
generator.append(R"~~~(
static constexpr Array<@type@, @size@> @name@ { {
)~~~");
constexpr size_t max_ranges_per_row = 20;
size_t ranges_in_current_row = 0;
static constexpr size_t max_values_per_row = 300;
size_t values_in_current_row = 0;
for (auto const& range : ranges) {
if (ranges_in_current_row++ > 0)
generator.append(" ");
for (auto value : stage) {
if (values_in_current_row++ > 0)
generator.append(", ");
generator.set("first", DeprecatedString::formatted("{:#x}", range.first));
generator.set("last", DeprecatedString::formatted("{:#x}", range.last));
generator.append("{ @first@, @last@ },");
TRY(generator.set("value", TRY(String::number(value))));
generator.append("@value@");
if (ranges_in_current_row == max_ranges_per_row) {
ranges_in_current_row = 0;
generator.append("\n ");
if (values_in_current_row == max_values_per_row) {
values_in_current_row = 0;
generator.append(",\n ");
}
}
}
generator.append(R"~~~(
} };
)~~~");
};
auto append_prop_list = [&](StringView collection_name, StringView property_format, PropList const& property_list) {
for (auto const& property : property_list) {
auto name = DeprecatedString::formatted(property_format, property.key);
append_code_point_range_list(move(name), property.value);
}
auto property_names = property_list.keys();
quick_sort(property_names);
generator.set("name", collection_name);
generator.set("size", DeprecatedString::number(property_names.size()));
generator.append(R"~~~(
static constexpr Array<ReadonlySpan<CodePointRange>, @size@> @name@ { {)~~~");
for (auto const& property_name : property_names) {
generator.set("name", DeprecatedString::formatted(property_format, property_name));
generator.append(R"~~~(
@name@.span(),)~~~");
}
generator.append(R"~~~(
} };
)~~~");
return {};
};
auto append_unique_properties = [&](auto const& unique_properties) -> ErrorOr<void> {
TRY(generator.set("name", TRY(String::formatted("{}_unique_properties", collection_snake))));
TRY(generator.set("outer_size", TRY(String::number(unique_properties.size()))));
TRY(generator.set("inner_size", TRY(String::number(unique_properties[0].size()))));
generator.append(R"~~~(
static constexpr Array<Array<bool, @inner_size@>, @outer_size@> @name@ { {)~~~");
for (auto const& property_set : unique_properties) {
generator.append(R"~~~(
{ )~~~");
for (auto value : property_set) {
TRY(generator.set("value", TRY(String::formatted("{}", value))));
generator.append("@value@, ");
}
generator.append(" },");
}
generator.append(R"~~~(
} };
)~~~");
return {};
};
TRY(append_stage(tables.stage1, "stage1"sv, "u16"sv));
TRY(append_stage(tables.stage2, "stage2"sv, "u16"sv));
TRY(append_unique_properties(tables.unique_properties));
return {};
};
append_prop_list("s_general_categories"sv, "s_general_category_{}"sv, unicode_data.general_categories);
append_prop_list("s_properties"sv, "s_property_{}"sv, unicode_data.prop_list);
append_prop_list("s_scripts"sv, "s_script_{}"sv, unicode_data.script_list);
append_prop_list("s_script_extensions"sv, "s_script_extension_{}"sv, unicode_data.script_extensions);
append_prop_list("s_grapheme_break_properties"sv, "s_grapheme_break_property_{}"sv, unicode_data.grapheme_break_props);
append_prop_list("s_word_break_properties"sv, "s_word_break_property_{}"sv, unicode_data.word_break_props);
append_prop_list("s_sentence_break_properties"sv, "s_sentence_break_property_{}"sv, unicode_data.sentence_break_props);
TRY(append_code_point_tables("s_general_categories"sv, unicode_data.general_category_tables));
TRY(append_code_point_tables("s_properties"sv, unicode_data.property_tables));
TRY(append_code_point_tables("s_scripts"sv, unicode_data.script_tables));
TRY(append_code_point_tables("s_script_extensions"sv, unicode_data.script_extension_tables));
TRY(append_code_point_tables("s_grapheme_break_properties"sv, unicode_data.grapheme_break_tables));
TRY(append_code_point_tables("s_word_break_properties"sv, unicode_data.word_break_tables));
TRY(append_code_point_tables("s_sentence_break_properties"sv, unicode_data.sentence_break_tables));
auto append_code_point_display_names = [&](StringView type, StringView name, auto const& display_names) {
constexpr size_t max_values_per_row = 30;
@ -1271,20 +1304,27 @@ Optional<CodePointDecomposition const> code_point_decomposition_by_index(size_t
}
)~~~");
auto append_prop_search = [&](StringView enum_title, StringView enum_snake, StringView collection_name) {
auto append_prop_search = [&](StringView enum_title, StringView enum_snake, StringView collection_name) -> ErrorOr<void> {
generator.set("enum_title", enum_title);
generator.set("enum_snake", enum_snake);
generator.set("collection_name", collection_name);
TRY(generator.set("LSB_COUNT", TRY(String::number(CodePointTables::LSB_COUNT))));
TRY(generator.set("LSB_MASK", TRY(String::formatted("{:#x}", CodePointTables::LSB_MASK))));
generator.append(R"~~~(
bool code_point_has_@enum_snake@(u32 code_point, @enum_title@ @enum_snake@)
{
auto index = static_cast<@enum_title@UnderlyingType>(@enum_snake@);
auto const& ranges = @collection_name@.at(index);
auto stage1_index = code_point >> @LSB_COUNT@;
auto stage2_index = @collection_name@_stage1[stage1_index] + (code_point & @LSB_MASK@);
auto unique_properties_index = @collection_name@_stage2[stage2_index];
auto const* range = binary_search(ranges, code_point, nullptr, CodePointRangeComparator {});
return range != nullptr;
auto const& property_set = @collection_name@_unique_properties[unique_properties_index];
return property_set[to_underlying(@enum_snake@)];
}
)~~~");
return {};
};
auto append_from_string = [&](StringView enum_title, StringView enum_snake, auto const& prop_list, Vector<Alias> const& aliases) -> ErrorOr<void> {
@ -1312,19 +1352,19 @@ bool code_point_has_@enum_snake@(u32 code_point, @enum_title@ @enum_snake@)
TRY(append_from_string("Locale"sv, "locale"sv, unicode_data.locales, {}));
append_prop_search("GeneralCategory"sv, "general_category"sv, "s_general_categories"sv);
TRY(append_prop_search("GeneralCategory"sv, "general_category"sv, "s_general_categories"sv));
TRY(append_from_string("GeneralCategory"sv, "general_category"sv, unicode_data.general_categories, unicode_data.general_category_aliases));
append_prop_search("Property"sv, "property"sv, "s_properties"sv);
TRY(append_prop_search("Property"sv, "property"sv, "s_properties"sv));
TRY(append_from_string("Property"sv, "property"sv, unicode_data.prop_list, unicode_data.prop_aliases));
append_prop_search("Script"sv, "script"sv, "s_scripts"sv);
append_prop_search("Script"sv, "script_extension"sv, "s_script_extensions"sv);
TRY(append_prop_search("Script"sv, "script"sv, "s_scripts"sv));
TRY(append_prop_search("Script"sv, "script_extension"sv, "s_script_extensions"sv));
TRY(append_from_string("Script"sv, "script"sv, unicode_data.script_list, unicode_data.script_aliases));
append_prop_search("GraphemeBreakProperty"sv, "grapheme_break_property"sv, "s_grapheme_break_properties"sv);
append_prop_search("WordBreakProperty"sv, "word_break_property"sv, "s_word_break_properties"sv);
append_prop_search("SentenceBreakProperty"sv, "sentence_break_property"sv, "s_sentence_break_properties"sv);
TRY(append_prop_search("GraphemeBreakProperty"sv, "grapheme_break_property"sv, "s_grapheme_break_properties"sv));
TRY(append_prop_search("WordBreakProperty"sv, "word_break_property"sv, "s_word_break_properties"sv));
TRY(append_prop_search("SentenceBreakProperty"sv, "sentence_break_property"sv, "s_sentence_break_properties"sv));
generator.append(R"~~~(
}
@ -1460,6 +1500,128 @@ static ErrorOr<void> normalize_script_extensions(PropList& script_extensions, Pr
return {};
}
// The goal here is to produce a set of tables that represent a category of code point properties for every code point.
// The most naive method would be to generate a single table per category, each with one entry per code point. Each of
// those tables would have a size of 0x10ffff though, which is a non-starter. Instead, we create a set of 2-stage lookup
// tables per category.
//
// To do so, it's important to note that Unicode tends to organize code points with similar properties together. This
// leads to long series of code points with identical properties. Therefore, if we divide the 0x10ffff code points into
// fixed-size blocks, many of those blocks will also be identical.
//
// So we iterate over every code point, classifying each one for the category of interest. We represent a classification
// as a list of booleans. We store the classification in the CodePointTables::unique_properties list for this category.
// As the name implies, this list is de-duplicated; we store the index into this list in a separate list, which we call
// a "block".
//
// As we iterate, we "pause" every BLOCK_SIZE code points to examine the block. If the block is unique so far, we extend
// CodePointTables::stage2 with the entries of that block (so CodePointTables::stage2 is also a list of indices into
// CodePointTables::unique_properties). We then append the index of the start of that block in CodePointTables::stage2
// to CodePointTables::stage1.
//
// The value of BLOCK_SIZE is determined by CodePointTables::MSB_COUNT and CodePointTables::LSB_COUNT. We need 24 bits
// to describe all code points; the blocks we create are based on splitting these bits into 2 segments. We currently use
// a 16:8 bit split. So when perform a runtime lookup of a code point in the 2-stage tables, we:
//
// 1. Use most-significant 16 bits of the code point as the index into CodePointTables::stage1. That value is the
// index into CodePointTables::stage2 of the start of the block that contains properties for this code point.
//
// 2. Add the least-significant 8 bits of the code point to that value, to use as the index into
// CodePointTables::stage2. As described above, that value is the index into CodePointTables::unique_properties,
// which contains the classification for this code point.
//
// Using the code point GeneralCategory as an example, we end up with a CodePointTables::stage1 with a size of ~4000,
// a CodePointTables::stage2 with a size of ~40,000, and a CodePointTables::unique_properties with a size of ~30. So
// this process reduces over 1 million entries (0x10ffff) to ~44,030.
//
// For much more in-depth reading, see: https://icu.unicode.org/design/struct/utrie
static ErrorOr<void> create_code_point_tables(UnicodeData& unicode_data)
{
static constexpr auto MAX_CODE_POINT = 0x10ffffu;
struct TableMetadata {
static ErrorOr<TableMetadata> create(PropList const& property_list)
{
TableMetadata data;
TRY(data.property_values.try_ensure_capacity(property_list.size()));
TRY(data.property_set.try_ensure_capacity(property_list.size()));
auto property_names = property_list.keys();
quick_sort(property_names);
for (auto const& property_name : property_names) {
auto const& code_point_rages = property_list.get(property_name).value();
data.property_values.unchecked_append(code_point_rages);
}
return data;
}
Vector<typename PropList::ValueType const&> property_values;
Vector<bool> property_set;
Vector<size_t> current_block;
HashMap<decltype(current_block), size_t> unique_blocks;
};
auto update_tables = [](auto code_point, auto& tables, auto& metadata) -> ErrorOr<void> {
static constexpr auto BLOCK_SIZE = CodePointTables::LSB_MASK + 1;
for (auto const& property_values : metadata.property_values) {
auto has_property = binary_search(property_values, code_point, nullptr, Unicode::CodePointRangeComparator {}) != nullptr;
metadata.property_set.unchecked_append(has_property);
}
size_t unique_properties_index = 0;
if (auto block_index = tables.unique_properties.find_first_index(metadata.property_set); block_index.has_value()) {
unique_properties_index = *block_index;
} else {
unique_properties_index = tables.unique_properties.size();
TRY(tables.unique_properties.try_append(metadata.property_set));
}
TRY(metadata.current_block.try_append(unique_properties_index));
metadata.property_set.clear_with_capacity();
if (metadata.current_block.size() == BLOCK_SIZE || code_point == MAX_CODE_POINT) {
size_t stage2_index = 0;
if (auto block_index = metadata.unique_blocks.get(metadata.current_block); block_index.has_value()) {
stage2_index = *block_index;
} else {
stage2_index = tables.stage2.size();
TRY(tables.stage2.try_extend(metadata.current_block));
TRY(metadata.unique_blocks.try_set(metadata.current_block, stage2_index));
}
TRY(tables.stage1.try_append(stage2_index));
metadata.current_block.clear_with_capacity();
}
return {};
};
auto general_category_metadata = TRY(TableMetadata::create(unicode_data.general_categories));
auto property_metadata = TRY(TableMetadata::create(unicode_data.prop_list));
auto script_metadata = TRY(TableMetadata::create(unicode_data.script_list));
auto script_extension_metadata = TRY(TableMetadata::create(unicode_data.script_extensions));
auto grapheme_break_metadata = TRY(TableMetadata::create(unicode_data.grapheme_break_props));
auto word_break_metadata = TRY(TableMetadata::create(unicode_data.word_break_props));
auto sentence_break_metadata = TRY(TableMetadata::create(unicode_data.sentence_break_props));
for (u32 code_point = 0; code_point <= MAX_CODE_POINT; ++code_point) {
TRY(update_tables(code_point, unicode_data.general_category_tables, general_category_metadata));
TRY(update_tables(code_point, unicode_data.property_tables, property_metadata));
TRY(update_tables(code_point, unicode_data.script_tables, script_metadata));
TRY(update_tables(code_point, unicode_data.script_extension_tables, script_extension_metadata));
TRY(update_tables(code_point, unicode_data.grapheme_break_tables, grapheme_break_metadata));
TRY(update_tables(code_point, unicode_data.word_break_tables, word_break_metadata));
TRY(update_tables(code_point, unicode_data.sentence_break_tables, sentence_break_metadata));
}
return {};
}
ErrorOr<int> serenity_main(Main::Arguments arguments)
{
StringView generated_header_path;
@ -1551,6 +1713,8 @@ ErrorOr<int> serenity_main(Main::Arguments arguments)
TRY(parse_value_alias_list(*prop_value_alias_file, "sc"sv, unicode_data.script_list.keys(), unicode_data.script_aliases, false));
TRY(normalize_script_extensions(unicode_data.script_extensions, unicode_data.script_list, unicode_data.script_aliases));
TRY(create_code_point_tables(unicode_data));
TRY(generate_unicode_data_header(*generated_header_file, unicode_data));
TRY(generate_unicode_data_implementation(*generated_implementation_file, unicode_data));

View File

@ -594,6 +594,28 @@ TEST_CASE(general_category)
}
}
BENCHMARK_CASE(general_category_performance)
{
auto general_category_cased_letter = Unicode::general_category_from_string("Cased_Letter"sv).value();
for (size_t i = 0; i < 1'000'000; ++i) {
for (u32 code_point = 0; code_point <= 0x1f; ++code_point)
EXPECT(!Unicode::code_point_has_general_category(code_point, general_category_cased_letter));
for (u32 code_point = 0x41; code_point <= 0x5a; ++code_point)
EXPECT(Unicode::code_point_has_general_category(code_point, general_category_cased_letter));
for (u32 code_point = 0x61; code_point <= 0x7a; ++code_point)
EXPECT(Unicode::code_point_has_general_category(code_point, general_category_cased_letter));
for (u32 code_point = 0xe000; code_point <= 0xe100; ++code_point)
EXPECT(!Unicode::code_point_has_general_category(code_point, general_category_cased_letter));
for (u32 code_point = 0x101fe; code_point <= 0x1027f; ++code_point)
EXPECT(!Unicode::code_point_has_general_category(code_point, general_category_cased_letter));
}
}
TEST_CASE(property)
{
auto property = [](StringView name) {