#pragma once #include #include #include #include #include #pragma clang diagnostic push #pragma clang diagnostic ignored "-Wunneeded-internal-declaration" #include #pragma clang diagnostic pop /** Methods for reading floating point numbers from text with decimal representation. * There are "precise", "fast" and "simple" implementations. * * Neither of methods support hexadecimal numbers (0xABC), binary exponent (1p100), leading plus sign. * * Precise method always returns a number that is the closest machine representable number to the input. * * Fast method is faster (up to 3 times) and usually return the same value, * but in rare cases result may differ by lest significant bit (for Float32) * and by up to two least significant bits (for Float64) from precise method. * Also fast method may parse some garbage as some other unspecified garbage. * * Simple method is little faster for cases of parsing short (few digit) integers, but less precise and slower in other cases. * It's not recommended to use simple method and it is left only for reference. * * For performance test, look at 'read_float_perf' test. * * For precision test. * Parse all existing Float32 numbers: CREATE TABLE test.floats ENGINE = Log AS SELECT reinterpretAsFloat32(reinterpretAsString(toUInt32(number))) AS x FROM numbers(0x100000000); WITH toFloat32(toString(x)) AS y, reinterpretAsUInt32(reinterpretAsString(x)) AS bin_x, reinterpretAsUInt32(reinterpretAsString(y)) AS bin_y, abs(bin_x - bin_y) AS diff SELECT diff, count() FROM test.floats WHERE NOT isNaN(x) GROUP BY diff ORDER BY diff ASC LIMIT 100 * Here are the results: * Precise: ┌─diff─┬────count()─┐ │ 0 │ 4278190082 │ └──────┴────────────┘ (100% roundtrip property) Fast: ┌─diff─┬────count()─┐ │ 0 │ 3685260580 │ │ 1 │ 592929502 │ └──────┴────────────┘ (The difference is 1 in least significant bit in 13.8% of numbers.) Simple: ┌─diff─┬────count()─┐ │ 0 │ 2169879994 │ │ 1 │ 1807178292 │ │ 2 │ 269505944 │ │ 3 │ 28826966 │ │ 4 │ 2566488 │ │ 5 │ 212878 │ │ 6 │ 18276 │ │ 7 │ 1214 │ │ 8 │ 30 │ └──────┴────────────┘ * Parse random Float64 numbers: WITH rand64() AS bin_x, reinterpretAsFloat64(reinterpretAsString(bin_x)) AS x, toFloat64(toString(x)) AS y, reinterpretAsUInt64(reinterpretAsString(y)) AS bin_y, abs(bin_x - bin_y) AS diff SELECT diff, count() FROM numbers(100000000) WHERE NOT isNaN(x) GROUP BY diff ORDER BY diff ASC LIMIT 100 */ namespace DB { namespace ErrorCodes { extern const int CANNOT_PARSE_NUMBER; } /// Returns true, iff parsed. bool parseInfinity(ReadBuffer & buf); bool parseNaN(ReadBuffer & buf); void assertInfinity(ReadBuffer & buf); void assertNaN(ReadBuffer & buf); template bool assertOrParseInfinity(ReadBuffer & buf) { if constexpr (throw_exception) { assertInfinity(buf); return true; } else return parseInfinity(buf); } template bool assertOrParseNaN(ReadBuffer & buf) { if constexpr (throw_exception) { assertNaN(buf); return true; } else return parseNaN(buf); } template ReturnType readFloatTextPreciseImpl(T & x, ReadBuffer & buf) { static_assert(std::is_same_v || std::is_same_v, "Argument for readFloatTextPreciseImpl must be float or double"); static_assert('a' > '.' && 'A' > '.' && '\n' < '.' && '\t' < '.' && '\'' < '.' && '"' < '.', "Layout of char is not like ASCII"); static constexpr bool throw_exception = std::is_same_v; static constexpr int MAX_LENGTH = 316; /// NOLINTBEGIN(readability-else-after-return) ReadBufferFromMemory * buf_from_memory = dynamic_cast(&buf); /// Fast path (avoid copying) if the buffer have at least MAX_LENGTH bytes or buf is ReadBufferFromMemory if (likely(!buf.eof() && (buf_from_memory || buf.position() + MAX_LENGTH <= buf.buffer().end()))) { auto * initial_position = buf.position(); auto res = fast_float::from_chars(initial_position, buf.buffer().end(), x); if (unlikely(res.ec != std::errc())) { if constexpr (throw_exception) throw Exception( ErrorCodes::CANNOT_PARSE_NUMBER, "Cannot read floating point value here: {}", String(initial_position, buf.buffer().end() - initial_position)); else return ReturnType(false); } buf.position() += res.ptr - initial_position; return ReturnType(true); } else { /// Slow path. Copy characters that may be present in floating point number to temporary buffer. bool negative = false; /// We check eof here because we can parse +inf +nan while (!buf.eof()) { switch (*buf.position()) { case '+': ++buf.position(); continue; case '-': { negative = true; ++buf.position(); continue; } case 'i': [[fallthrough]]; case 'I': { if (assertOrParseInfinity(buf)) { x = std::numeric_limits::infinity(); if (negative) x = -x; return ReturnType(true); } return ReturnType(false); } case 'n': [[fallthrough]]; case 'N': { if (assertOrParseNaN(buf)) { x = std::numeric_limits::quiet_NaN(); if (negative) x = -x; return ReturnType(true); } return ReturnType(false); } default: break; } break; } char tmp_buf[MAX_LENGTH]; int num_copied_chars = 0; while (!buf.eof() && num_copied_chars < MAX_LENGTH) { char c = *buf.position(); if (!(isNumericASCII(c) || c == '-' || c == '+' || c == '.' || c == 'e' || c == 'E')) break; tmp_buf[num_copied_chars] = c; ++buf.position(); ++num_copied_chars; } fast_float::from_chars_result res; if constexpr (std::endian::native == std::endian::little) res = fast_float::from_chars(tmp_buf, tmp_buf + num_copied_chars, x); else { Float64 x64 = 0.0; res = fast_float::from_chars(tmp_buf, tmp_buf + num_copied_chars, x64); x = static_cast(x64); } if (unlikely(res.ec != std::errc() || res.ptr - tmp_buf != num_copied_chars)) { if constexpr (throw_exception) throw Exception( ErrorCodes::CANNOT_PARSE_NUMBER, "Cannot read floating point value here: {}", String(tmp_buf, num_copied_chars)); else return ReturnType(false); } if (negative) x = -x; return ReturnType(true); } /// NOLINTEND(readability-else-after-return) } // credit: https://johnnylee-sde.github.io/Fast-numeric-string-to-int/ static inline bool is_made_of_eight_digits_fast(uint64_t val) noexcept { return (((val & 0xF0F0F0F0F0F0F0F0) | (((val + 0x0606060606060606) & 0xF0F0F0F0F0F0F0F0) >> 4)) == 0x3333333333333333); } static inline bool is_made_of_eight_digits_fast(const char * chars) noexcept { uint64_t val; ::memcpy(&val, chars, 8); return is_made_of_eight_digits_fast(val); } template static inline void readUIntTextUpToNSignificantDigits(T & x, ReadBuffer & buf) { /// In optimistic case we can skip bound checking for first loop. if (buf.position() + N <= buf.buffer().end()) { for (size_t i = 0; i < N; ++i) { if (isNumericASCII(*buf.position())) { x *= 10; x += *buf.position() & 0x0F; ++buf.position(); } else return; } } else { for (size_t i = 0; i < N; ++i) { if (!buf.eof() && isNumericASCII(*buf.position())) { x *= 10; x += *buf.position() & 0x0F; ++buf.position(); } else return; } } while (!buf.eof() && (buf.position() + 8 <= buf.buffer().end()) && is_made_of_eight_digits_fast(buf.position())) { buf.position() += 8; } while (!buf.eof() && isNumericASCII(*buf.position())) ++buf.position(); } template ReturnType readFloatTextFastImpl(T & x, ReadBuffer & in, bool & has_fractional) { static_assert(std::is_same_v || std::is_same_v, "Argument for readFloatTextImpl must be float or double"); static_assert('a' > '.' && 'A' > '.' && '\n' < '.' && '\t' < '.' && '\'' < '.' && '"' < '.', "Layout of char is not like ASCII"); has_fractional = false; static constexpr bool throw_exception = std::is_same_v; bool negative = false; x = 0; UInt64 before_point = 0; UInt64 after_point = 0; int after_point_exponent = 0; int exponent = 0; if (in.eof()) { if constexpr (throw_exception) throw Exception(ErrorCodes::CANNOT_PARSE_NUMBER, "Cannot read floating point value"); else return false; } if (*in.position() == '-') { negative = true; ++in.position(); } else if (*in.position() == '+') ++in.position(); auto count_after_sign = in.count(); constexpr int significant_digits = std::numeric_limits::digits10; readUIntTextUpToNSignificantDigits(before_point, in); size_t read_digits = in.count() - count_after_sign; if (unlikely(read_digits > significant_digits)) { int before_point_additional_exponent = static_cast(read_digits) - significant_digits; x = static_cast(shift10(before_point, before_point_additional_exponent)); } else { x = before_point; /// Shortcut for the common case when there is an integer that fit in Int64. if (read_digits && (in.eof() || *in.position() < '.')) { if (negative) x = -x; return ReturnType(true); } } if (checkChar('.', in)) { has_fractional = true; auto after_point_count = in.count(); while (!in.eof() && *in.position() == '0') ++in.position(); auto after_leading_zeros_count = in.count(); int after_point_num_leading_zeros = static_cast(after_leading_zeros_count - after_point_count); readUIntTextUpToNSignificantDigits(after_point, in); read_digits = in.count() - after_leading_zeros_count; after_point_exponent = (read_digits > significant_digits ? -significant_digits : static_cast(-read_digits)) - after_point_num_leading_zeros; } if constexpr (allow_exponent) { if (checkChar('e', in) || checkChar('E', in)) { has_fractional = true; if (in.eof()) { if constexpr (throw_exception) throw Exception(ErrorCodes::CANNOT_PARSE_NUMBER, "Cannot read floating point value: nothing after exponent"); else return false; } bool exponent_negative = false; if (*in.position() == '-') { exponent_negative = true; ++in.position(); } else if (*in.position() == '+') { ++in.position(); } readUIntTextUpToNSignificantDigits<4>(exponent, in); if (exponent_negative) exponent = -exponent; } } if (after_point) { x += static_cast(shift10(after_point, after_point_exponent)); } if (exponent) { x = static_cast(shift10(x, exponent)); } if (negative) x = -x; auto num_characters_without_sign = in.count() - count_after_sign; /// Denormals. At most one character is read before denormal and it is '-'. if (num_characters_without_sign == 0) { if (in.eof()) { if constexpr (throw_exception) throw Exception(ErrorCodes::CANNOT_PARSE_NUMBER, "Cannot read floating point value: no digits read"); else return false; } if (*in.position() == '+') { ++in.position(); if (in.eof()) { if constexpr (throw_exception) throw Exception(ErrorCodes::CANNOT_PARSE_NUMBER, "Cannot read floating point value: nothing after plus sign"); else return false; } else if (negative) { if constexpr (throw_exception) throw Exception(ErrorCodes::CANNOT_PARSE_NUMBER, "Cannot read floating point value: plus after minus sign"); else return false; } } if (*in.position() == 'i' || *in.position() == 'I') { if (assertOrParseInfinity(in)) { x = std::numeric_limits::infinity(); if (negative) x = -x; return ReturnType(true); } return ReturnType(false); } if (*in.position() == 'n' || *in.position() == 'N') { if (assertOrParseNaN(in)) { x = std::numeric_limits::quiet_NaN(); if (negative) x = -x; return ReturnType(true); } return ReturnType(false); } } return ReturnType(true); } template ReturnType readFloatTextSimpleImpl(T & x, ReadBuffer & buf) { static constexpr bool throw_exception = std::is_same_v; bool negative = false; x = 0; bool after_point = false; T power_of_ten = 1; if (buf.eof()) throwReadAfterEOF(); while (!buf.eof()) { switch (*buf.position()) { case '+': break; case '-': negative = true; break; case '.': after_point = true; break; case '0': [[fallthrough]]; case '1': [[fallthrough]]; case '2': [[fallthrough]]; case '3': [[fallthrough]]; case '4': [[fallthrough]]; case '5': [[fallthrough]]; case '6': [[fallthrough]]; case '7': [[fallthrough]]; case '8': [[fallthrough]]; case '9': if (after_point) { power_of_ten /= 10; x += (*buf.position() - '0') * power_of_ten; } else { x *= 10; x += *buf.position() - '0'; } break; case 'e': [[fallthrough]]; case 'E': { ++buf.position(); Int32 exponent = 0; readIntText(exponent, buf); x = shift10(x, exponent); if (negative) x = -x; return ReturnType(true); } case 'i': [[fallthrough]]; case 'I': { if (assertOrParseInfinity(buf)) { x = std::numeric_limits::infinity(); if (negative) x = -x; return ReturnType(true); } return ReturnType(false); } case 'n': [[fallthrough]]; case 'N': { if (assertOrParseNaN(buf)) { x = std::numeric_limits::quiet_NaN(); if (negative) x = -x; return ReturnType(true); } return ReturnType(false); } default: { if (negative) x = -x; return ReturnType(true); } } ++buf.position(); } if (negative) x = -x; return ReturnType(true); } template void readFloatTextPrecise(T & x, ReadBuffer & in) { readFloatTextPreciseImpl(x, in); } template bool tryReadFloatTextPrecise(T & x, ReadBuffer & in) { return readFloatTextPreciseImpl(x, in); } template void readFloatTextFast(T & x, ReadBuffer & in) { bool has_fractional; readFloatTextFastImpl(x, in, has_fractional); } template bool tryReadFloatTextFast(T & x, ReadBuffer & in) { bool has_fractional; return readFloatTextFastImpl(x, in, has_fractional); } template void readFloatTextSimple(T & x, ReadBuffer & in) { readFloatTextSimpleImpl(x, in); } template bool tryReadFloatTextSimple(T & x, ReadBuffer & in) { return readFloatTextSimpleImpl(x, in); } /// Implementation that is selected as default. template void readFloatText(T & x, ReadBuffer & in) { readFloatTextFast(x, in); } template bool tryReadFloatText(T & x, ReadBuffer & in) { return tryReadFloatTextFast(x, in); } /// Don't read exponent part of the number. template bool tryReadFloatTextNoExponent(T & x, ReadBuffer & in) { bool has_fractional; return readFloatTextFastImpl(x, in, has_fractional); } /// With a @has_fractional flag /// Used for input_format_try_infer_integers template bool tryReadFloatTextExt(T & x, ReadBuffer & in, bool & has_fractional) { return readFloatTextFastImpl(x, in, has_fractional); } template bool tryReadFloatTextExtNoExponent(T & x, ReadBuffer & in, bool & has_fractional) { return readFloatTextFastImpl(x, in, has_fractional); } }