diff options
Diffstat (limited to 'include/unicode/optimization.h')
| -rw-r--r-- | include/unicode/optimization.h | 325 |
1 files changed, 325 insertions, 0 deletions
diff --git a/include/unicode/optimization.h b/include/unicode/optimization.h new file mode 100644 index 0000000..d7b054d --- /dev/null +++ b/include/unicode/optimization.h @@ -0,0 +1,325 @@ +// +// Reichwein.IT Unicode Library +// +// Optimized conversion functions for UTF input and output +// + +#pragma once + +#include "unicode/endian.h" +#include "unicode/iso.h" +#include "unicode/predicate.h" +#include "unicode/types.h" +#include "unicode/type_traits.h" +#include "unicode/utf.h" + +#include <algorithm> +#include <array> +#include <cstdint> +#include <iterator> +#include <memory> +#include <stdexcept> +#include <string> +#include <type_traits> +#include <utility> + +namespace unicode { + + // Helper function: Item distance of specified iterators + // std::distance doesn't work here: it is based on "output" distance of iterators + template<class Iterator> + inline size_t input_distance(const Iterator& it1, const Iterator& it2) + { + return it2 - it1; + } + + // Helper function: Distance of specified iterator content data in bytes + template<class Iterator> + inline size_t input_distance_bytes(const Iterator& it1, const Iterator& it2) + { + return input_distance(it1, it2) * sizeof(typename Iterator::value_type); + } + + // Optimizations following: + static const size_t accu_size {sizeof(size_t)}; + + template<int value_size> + struct ConvertInputOptimizer {}; + + template<> struct ConvertInputOptimizer<1> + { + static const uint32_t ascii_mask { 0x80808080 }; + }; + + template<> struct ConvertInputOptimizer<2> + { + static const uint32_t ascii_mask { 0xFF80FF80 }; + }; + + template<> struct ConvertInputOptimizer<4> + { + static const uint32_t ascii_mask { 0xFFFFFF80 }; + }; + + template<int AccuSize, class ConvertInputOptimizer> + struct ArchitectureOptimizer {}; + + // On 32 bit architecture, calculate with 32 bit accumulator value + // (hoping the compiler will put it into a 32 bit register) + template<class ConvertInputOptimizer> + struct ArchitectureOptimizer<4, ConvertInputOptimizer> + { + typedef ConvertInputOptimizer input_optimizer; + typedef uint32_t accu_type; + static const accu_type addr_mask {accu_size - 1}; + static const accu_type ascii_mask { (accu_type)input_optimizer::ascii_mask }; + static const accu_type ascii_value { 0ULL }; + + template<typename input_value_type, class output_string_type> + inline static void append(const input_value_type* addr, output_string_type& s) + { + if constexpr(sizeof(input_value_type) == sizeof(typename output_string_type::value_type)) { + s.append(reinterpret_cast<const typename output_string_type::value_type*>(addr), accu_size / sizeof(input_value_type)); + } else if constexpr(is_utf_8_v<input_value_type>) { + s.append({static_cast<typename output_string_type::value_type>(addr[0]), + static_cast<typename output_string_type::value_type>(addr[1]), + static_cast<typename output_string_type::value_type>(addr[2]), + static_cast<typename output_string_type::value_type>(addr[3])}); + } else if constexpr(is_utf_16_v<input_value_type>) { + s.append({static_cast<typename output_string_type::value_type>(addr[0]), + static_cast<typename output_string_type::value_type>(addr[1])}); + } else if constexpr(is_utf_32_v<input_value_type>) { + s.append({static_cast<typename output_string_type::value_type>(addr[0])}); + } + } + }; + + // On 64 bit architecture, calculate with 64 bit accumulator value + // (hoping the compiler will put it into a 64 bit register) + template<class ConvertInputOptimizer> + struct ArchitectureOptimizer<8, ConvertInputOptimizer> + { + typedef ConvertInputOptimizer input_optimizer; + typedef uint64_t accu_type; + static const accu_type addr_mask {accu_size - 1}; + static const accu_type ascii_mask { ((accu_type)input_optimizer::ascii_mask) << 32 | (accu_type)input_optimizer::ascii_mask }; + static const accu_type ascii_value { 0ULL }; + + template<typename input_value_type, class output_string_type> + inline static void append(const input_value_type* addr, output_string_type& s) + { + if constexpr(sizeof(input_value_type) == sizeof(typename output_string_type::value_type)) { + s.append(reinterpret_cast<const typename output_string_type::value_type*>(addr), accu_size / sizeof(input_value_type)); + } else if constexpr(is_utf_8_v<input_value_type>) { + s.append({static_cast<typename output_string_type::value_type>(addr[0]), + static_cast<typename output_string_type::value_type>(addr[1]), + static_cast<typename output_string_type::value_type>(addr[2]), + static_cast<typename output_string_type::value_type>(addr[3]), + static_cast<typename output_string_type::value_type>(addr[4]), + static_cast<typename output_string_type::value_type>(addr[5]), + static_cast<typename output_string_type::value_type>(addr[6]), + static_cast<typename output_string_type::value_type>(addr[7])}); + } else if constexpr(is_utf_16_v<input_value_type>) { + s.append({static_cast<typename output_string_type::value_type>(addr[0]), + static_cast<typename output_string_type::value_type>(addr[1]), + static_cast<typename output_string_type::value_type>(addr[2]), + static_cast<typename output_string_type::value_type>(addr[3])}); + } else if constexpr(is_utf_32_v<input_value_type>) { + s.append({static_cast<typename output_string_type::value_type>(addr[0]), + static_cast<typename output_string_type::value_type>(addr[1])}); + } + } + + }; // class ArchitectureOptimizer + + // Optimize for the case of all ASCII (7-bit) data in a accu size row + // From and To are Encodings + template<typename From, typename To, std::enable_if_t<is_encoding_v<From> && is_encoding_v<To>, bool> = true> + typename To::string_type convert_optimized(const typename From::string_type& s) + { + typename To::string_type result; + typedef ConvertInputOptimizer<sizeof(typename From::value_type)> input_optimizer; + typedef ArchitectureOptimizer<accu_size, input_optimizer> arch_optimizer; + + auto begin{From::begin(s)}; + auto end{From::end(s)}; + auto back_inserter{To::back_inserter(result)}; + auto addr{reinterpret_cast<const typename arch_optimizer::accu_type*>(&s.data()[s.size() - input_distance(begin, end)])}; + while (input_distance_bytes(begin, end) >= accu_size) { + if (((uintptr_t)(void*)addr & arch_optimizer::addr_mask) == 0) { + while (input_distance_bytes(begin, end) >= accu_size) { + typename arch_optimizer::accu_type data{*addr}; + if ((data & arch_optimizer::ascii_mask) == arch_optimizer::ascii_value) +#if __cplusplus >= 202002L + [[likely]] +#endif + { + arch_optimizer::template append(reinterpret_cast<const typename From::value_type*>(addr), result); + begin += accu_size / sizeof(typename From::value_type); + ++addr; + } else { + // just advance one code unit for now and break to trigger unoptimized + // version until next accu boundary + back_inserter = *begin; + ++begin; + break; + } + } + } + + // keep up after unaligned Non-ASCII code points + while (begin != end && (uintptr_t)(void*)(addr = reinterpret_cast<const typename arch_optimizer::accu_type*>(&s.data()[s.size() - input_distance(begin, end)])) & arch_optimizer::addr_mask) { + back_inserter = *begin; + ++begin; + } + } + + // remainder < 8 bytes + while (begin != end) { + back_inserter = *begin; + ++begin; + } + + return result; + } + + template<size_t bits_to_compare = 32, typename To, typename std::enable_if_t<is_utf_8_v<To>, bool> = true> + inline void append_utf(std::basic_string<To>& result, const char32_t& value) + { + using From = char32_t; + if (bits_to_compare <= 7 || value < 0x80) { // 1 byte + result.push_back(static_cast<To>(value)); + } else if (bits_to_compare <= 11 || value < 0x800) { // 2 bytes + result.append({utf8_byte_n_of_m<0,2,From,To>(value), utf8_byte_n_of_m<1,2,From,To>(value)}); + } else if (bits_to_compare <= 16 || value < 0x10000) { // 3 bytes + result.append({utf8_byte_n_of_m<0,3,From,To>(value), utf8_byte_n_of_m<1,3,From,To>(value), utf8_byte_n_of_m<2,3,From,To>(value)}); + } else { // 4 bytes + // expect value to be already valid Unicode values + result.append({utf8_byte_n_of_m<0,4,From,To>(value), utf8_byte_n_of_m<1,4,From,To>(value), utf8_byte_n_of_m<2,4,From,To>(value), utf8_byte_n_of_m<3,4,From,To>(value)}); + } + } + + template<size_t bits_to_compare = 32, typename To, typename std::enable_if_t<is_utf_16_v<To>, bool> = true> + inline void append_utf(std::basic_string<To>& result, const char32_t& value) + { + if (bits_to_compare <= 16 || value <= 0xFFFF) { // expect value to be already valid Unicode values + result.push_back(static_cast<To>(value)); + } else { + char32_t value_reduced{value - 0x10000}; + result.append({static_cast<To>((value_reduced >> 10) + 0xD800), static_cast<To>((value_reduced & 0x3FF) + 0xDC00)}); + } + } + + template<size_t bits_to_compare = 32, typename To, typename std::enable_if_t<is_utf_32_v<To>, bool> = true> + inline void append_utf(std::basic_string<To>& result, const char32_t& value) + { + // expect value to be already valid Unicode values (checked in input iterator) + result.push_back(static_cast<To>(value)); + } + + // Little Endian optimized version for UTF-8 + // In block_mode, at least 4 bytes are in accu. On first call, even 8. + // otherwise, at least one code unit is in accu + template<typename From, typename To, bool block_mode = true, typename std::enable_if_t<is_utf_8_v<From>, bool> = true> + inline static void append_accu(std::basic_string<To>& result, uint64_t& accu, int& bytes_in_accu) + { + if (block_mode && bytes_in_accu == 8 && (accu & 0x8080808080808080) == 0) +#if __cplusplus >= 202002L + [[likely]] +#endif + { + result.append({ + static_cast<To>(accu & 0x7F), + static_cast<To>((accu >> 8) & 0x7F), + static_cast<To>((accu >> 16) & 0x7F), + static_cast<To>((accu >> 24) & 0x7F), + static_cast<To>((accu >> 32) & 0x7F), + static_cast<To>((accu >> 40) & 0x7F), + static_cast<To>((accu >> 48) & 0x7F), + static_cast<To>((accu >> 56) & 0x7F), + }); + accu = 0; + bytes_in_accu = 0; + } else if ((accu & 0x80) == 0) { // 1 byte sequence + append_utf<7>(result, static_cast<char32_t>(accu & 0x7F)); + accu >>= 8; + bytes_in_accu -= 1; + } else if ((block_mode || bytes_in_accu >= 2) && (accu & 0xC0E0) == 0x80C0) { // 2 byte sequence + char32_t value {static_cast<char32_t>(((accu & 0x1F) << 6) | ((accu >> 8) & 0x3f))}; + accu >>= 16; + bytes_in_accu -= 2; + if (is_valid_unicode<11>(value)) + append_utf<11>(result, value); + else +#if __cplusplus >= 202002L + [[unlikely]] +#endif + throw std::invalid_argument("Invalid Unicode character in 2 byte UTF-8 sequence"); + } else if ((block_mode || bytes_in_accu >= 3) && (accu & 0xC0C0F0) == 0x8080E0) { // 3 byte sequence + char32_t value {static_cast<char32_t>(((accu & 0x0F) << 12) | ((accu >> 2) & 0x0FC0) | ((accu >> 16) & 0x3f))}; + accu >>= 24; + bytes_in_accu -= 3; + if (is_valid_unicode<16>(value)) + append_utf<16>(result, value); + else +#if __cplusplus >= 202002L + [[unlikely]] +#endif + throw std::invalid_argument("Invalid Unicode character in 3 byte UTF-8 sequence"); + } else if ((block_mode || bytes_in_accu >= 4) && (accu & 0xC0C0C0F8) == 0x808080F0) { // 4 byte sequence + char32_t value {static_cast<char32_t>(((accu & 0x07) << 18) | ((accu << 4) & 0x3f000) | ((accu >> 10) & 0xFC0) | ((accu >> 24) & 0x3f))}; + accu >>= 32; + bytes_in_accu -= 4; + if (is_valid_unicode<21>(value)) + append_utf(result, value); + else +#if __cplusplus >= 202002L + [[unlikely]] +#endif + throw std::invalid_argument("Invalid Unicode character in 4 byte UTF-8 sequence"); + } else +#if __cplusplus >= 202002L + [[unlikely]] +#endif + throw std::invalid_argument("Invalid UTF-8 byte sequence"); + } + + // Little Endian optimized version + template<typename From, typename To, std::enable_if_t<is_encoding_v<From> && is_encoding_v<To>, bool> = true> + typename To::string_type convert_optimized_utf(const typename From::string_type& s) + { + typename To::string_type result; + uint64_t accu{}; + int bytes_in_accu{}; + + size_t s_index{}; + size_t s_size{s.size()}; + while (s_index + 8 / sizeof(typename From::value_type) <= s_size) { + // read input + // assume: bytes_in_accu < 8 + accu |= (*reinterpret_cast<const uint64_t*>(&(s.data()[s_index]))) << (bytes_in_accu * 8); + s_index += (8 - bytes_in_accu) / sizeof(typename From::value_type); + bytes_in_accu = 8; + + while (bytes_in_accu >= 4) { + append_accu<typename From::value_type, typename To::value_type, true>(result, accu, bytes_in_accu); + } + } + + // 0..3 bytes left in accu + // 0..7 bytes left in s + + while (s_index < s_size || bytes_in_accu > 0) { + while (s_index < s_size && bytes_in_accu < 8) { + accu |= static_cast<uint64_t>(*reinterpret_cast<const uint8_t*>(&(s.data()[s_index]))) << (bytes_in_accu * 8); + ++s_index; + bytes_in_accu += sizeof(typename From::value_type); + } + + append_accu<typename From::value_type, typename To::value_type, false>(result, accu, bytes_in_accu); + } + return result; + } + +} // namespace unicode + |