From cd9d9018e0db8139e48cb722e9e9685d2a212c8b Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Ra=C3=BAl=20Mar=C3=ADn?= Date: Thu, 7 Mar 2024 17:15:42 +0100 Subject: [PATCH] Be able to iterate --- base/base/CMakeLists.txt | 1 + base/base/itoa.cpp | 503 +++++++++++++++++++++++++++++++++++ base/base/itoa.h | 498 +++------------------------------- src/Functions/CMakeLists.txt | 1 + 4 files changed, 540 insertions(+), 463 deletions(-) create mode 100644 base/base/itoa.cpp diff --git a/base/base/CMakeLists.txt b/base/base/CMakeLists.txt index 548ba01d86a..55d046767b8 100644 --- a/base/base/CMakeLists.txt +++ b/base/base/CMakeLists.txt @@ -19,6 +19,7 @@ set (SRCS getPageSize.cpp getThreadId.cpp int8_to_string.cpp + itoa.cpp JSON.cpp mremap.cpp phdr_cache.cpp diff --git a/base/base/itoa.cpp b/base/base/itoa.cpp new file mode 100644 index 00000000000..9fefc9f0f07 --- /dev/null +++ b/base/base/itoa.cpp @@ -0,0 +1,503 @@ +// Based on https://github.com/amdn/itoa and combined with our optimizations +// +//=== itoa.h - Fast integer to ascii conversion --*- C++ -*-// +// +// The MIT License (MIT) +// Copyright (c) 2016 Arturo Martin-de-Nicolas +// +// Permission is hereby granted, free of charge, to any person obtaining a copy +// of this software and associated documentation files (the "Software"), to deal +// in the Software without restriction, including without limitation the rights +// to use, copy, modify, merge, publish, distribute, sublicense, and/or sell +// copies of the Software, and to permit persons to whom the Software is +// furnished to do so, subject to the following conditions: +// +// The above copyright notice and this permission notice shall be included +// in all copies or substantial portions of the Software. +// +// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE +// AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, +// OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE +// SOFTWARE. +//===----------------------------------------------------------------------===// + +#include +#include +#include +#include +#include +#include +#include + + +template +int digits10(T x) +{ + if (x < T(10ULL)) + return 1; + if (x < T(100ULL)) + return 2; + if constexpr (sizeof(T) == 1) + return 3; + else + { + if (x < T(1000ULL)) + return 3; + + if (x < T(1000000000000ULL)) + { + if (x < T(100000000ULL)) + { + if (x < T(1000000ULL)) + { + if (x < T(10000ULL)) + return 4; + else + return 5 + (x >= T(100000ULL)); + } + + return 7 + (x >= T(10000000ULL)); + } + + if (x < T(10000000000ULL)) + return 9 + (x >= T(1000000000ULL)); + + return 11 + (x >= T(100000000000ULL)); + } + + return 12 + digits10(x / T(1000000000000ULL)); + } +} + + +namespace +{ + +template +static constexpr T pow10(size_t x) +{ + return x ? 10 * pow10(x - 1) : 1; +} + +// Division by a power of 10 is implemented using a multiplicative inverse. +// This strength reduction is also done by optimizing compilers, but +// presently the fastest results are produced by using the values +// for the multiplication and the shift as given by the algorithm +// described by Agner Fog in "Optimizing Subroutines in Assembly Language" +// +// http://www.agner.org/optimize/optimizing_assembly.pdf +// +// "Integer division by a constant (all processors) +// A floating point number can be divided by a constant by multiplying +// with the reciprocal. If we want to do the same with integers, we have +// to scale the reciprocal by 2n and then shift the product to the right +// by n. There are various algorithms for finding a suitable value of n +// and compensating for rounding errors. The algorithm described below +// was invented by Terje Mathisen, Norway, and not published elsewhere." + +/// Division by constant is performed by: +/// 1. Adding 1 if needed; +/// 2. Multiplying by another constant; +/// 3. Shifting right by another constant. +template +struct Division +{ + static constexpr bool add{add_}; + static constexpr UInt multiplier{multiplier_}; + static constexpr unsigned shift{shift_}; +}; + +/// Select a type with appropriate number of bytes from the list of types. +/// First parameter is the number of bytes requested. Then goes a list of types with 1, 2, 4, ... number of bytes. +/// Example: SelectType<4, uint8_t, uint16_t, uint32_t, uint64_t> will select uint32_t. +template +struct SelectType +{ + using Result = typename SelectType::Result; +}; + +template +struct SelectType<1, T, Ts...> +{ + using Result = T; +}; + + +/// Division by 10^N where N is the size of the type. +template +using DivisionBy10PowN = typename SelectType< + N, + Division, /// divide by 10 + Division, /// divide by 100 + Division, /// divide by 10000 + Division /// divide by 100000000 + >::Result; + +template +using UnsignedOfSize = typename SelectType::Result; + +/// Holds the result of dividing an unsigned N-byte variable by 10^N resulting in +template +struct QuotientAndRemainder +{ + UnsignedOfSize quotient; // quotient with fewer than 2*N decimal digits + UnsignedOfSize remainder; // remainder with at most N decimal digits +}; + +template +QuotientAndRemainder static inline split(UnsignedOfSize value) +{ + constexpr DivisionBy10PowN division; + + UnsignedOfSize quotient = (division.multiplier * (UnsignedOfSize<2 * N>(value) + division.add)) >> division.shift; + UnsignedOfSize remainder = static_cast>(value - quotient * pow10>(N)); + + return {quotient, remainder}; +} + + +static inline char * outDigit(char * p, uint8_t value) +{ + *p = '0' + value; + ++p; + return p; +} + +// Using a lookup table to convert binary numbers from 0 to 99 +// into ascii characters as described by Andrei Alexandrescu in +// https://www.facebook.com/notes/facebook-engineering/three-optimization-tips-for-c/10151361643253920/ + +static const char digits[201] = "00010203040506070809" + "10111213141516171819" + "20212223242526272829" + "30313233343536373839" + "40414243444546474849" + "50515253545556575859" + "60616263646566676869" + "70717273747576777879" + "80818283848586878889" + "90919293949596979899"; + +static inline char * outTwoDigits(char * p, uint8_t value) +{ + memcpy(p, &digits[value * 2], 2); + p += 2; + return p; +} + + +namespace convert +{ +template +static char * head(char * p, UInt u); +template +static char * tail(char * p, UInt u); + +//===----------------------------------------------------------===// +// head: find most significant digit, skip leading zeros +//===----------------------------------------------------------===// + +// "x" contains quotient and remainder after division by 10^N +// quotient is less than 10^N +template +static inline char * head(char * p, QuotientAndRemainder x) +{ + p = head(p, UnsignedOfSize(x.quotient)); + p = tail(p, x.remainder); + return p; +} + +// "u" is less than 10^2*N +template +static inline char * head(char * p, UInt u) +{ + return u < pow10>(N) ? head(p, UnsignedOfSize(u)) : head(p, split(u)); +} + +// recursion base case, selected when "u" is one byte +template <> +inline char * head, 1>(char * p, UnsignedOfSize<1> u) +{ + return u < 10 ? outDigit(p, u) : outTwoDigits(p, u); +} + +//===----------------------------------------------------------===// +// tail: produce all digits including leading zeros +//===----------------------------------------------------------===// + +// recursive step, "u" is less than 10^2*N +template +static inline char * tail(char * p, UInt u) +{ + QuotientAndRemainder x = split(u); + p = tail(p, UnsignedOfSize(x.quotient)); + p = tail(p, x.remainder); + return p; +} + +// recursion base case, selected when "u" is one byte +template <> +inline char * tail, 1>(char * p, UnsignedOfSize<1> u) +{ + return outTwoDigits(p, u); +} + +//===----------------------------------------------------------===// +// large values are >= 10^2*N +// where x contains quotient and remainder after division by 10^N +//===----------------------------------------------------------===// + +template +static inline char * large(char * p, QuotientAndRemainder x) +{ + QuotientAndRemainder y = split(x.quotient); + p = head(p, UnsignedOfSize(y.quotient)); + p = tail(p, y.remainder); + p = tail(p, x.remainder); + return p; +} + +//===----------------------------------------------------------===// +// handle values of "u" that might be >= 10^2*N +// where N is the size of "u" in bytes +//===----------------------------------------------------------===// + +template +static inline char * uitoa(char * p, UInt u) +{ + if (u < pow10>(N)) + return head(p, UnsignedOfSize(u)); + QuotientAndRemainder x = split(u); + + return u < pow10>(2 * N) ? head(p, x) : large(p, x); +} + +// selected when "u" is one byte +template <> +inline char * uitoa, 1>(char * p, UnsignedOfSize<1> u) +{ + if (u < 10) + return outDigit(p, u); + else if (u < 100) + return outTwoDigits(p, u); + else + { + p = outDigit(p, u / 100); + p = outTwoDigits(p, u % 100); + return p; + } +} + +//===----------------------------------------------------------===// +// handle unsigned and signed integral operands +//===----------------------------------------------------------===// + +// itoa: handle unsigned integral operands (selected by SFINAE) +template && std::is_integral_v> * = nullptr> +static inline char * itoa(U u, char * p) +{ + return convert::uitoa(p, u); +} + +// itoa: handle signed integral operands (selected by SFINAE) +template && std::is_integral_v> * = nullptr> +static inline char * itoa(I i, char * p) +{ + // Need "mask" to be filled with a copy of the sign bit. + // If "i" is a negative value, then the result of "operator >>" + // is implementation-defined, though usually it is an arithmetic + // right shift that replicates the sign bit. + // Use a conditional expression to be portable, + // a good optimizing compiler generates an arithmetic right shift + // and avoids the conditional branch. + UnsignedOfSize mask = i < 0 ? ~UnsignedOfSize(0) : 0; + // Now get the absolute value of "i" and cast to unsigned type UnsignedOfSize. + // Cannot use std::abs() because the result is undefined + // in 2's complement systems for the most-negative value. + // Want to avoid conditional branch for performance reasons since + // CPU branch prediction will be ineffective when negative values + // occur randomly. + // Let "u" be "i" cast to unsigned type UnsignedOfSize. + // Subtract "u" from 2*u if "i" is positive or 0 if "i" is negative. + // This yields the absolute value with the desired type without + // using a conditional branch and without invoking undefined or + // implementation defined behavior: + UnsignedOfSize u = ((2 * UnsignedOfSize(i)) & ~mask) - UnsignedOfSize(i); + // Unconditionally store a minus sign when producing digits + // in a forward direction and increment the pointer only if + // the value is in fact negative. + // This avoids a conditional branch and is safe because we will + // always produce at least one digit and it will overwrite the + // minus sign when the value is not negative. + *p = '-'; + p += (mask & 1); + p = convert::uitoa(p, u); + return p; +} +} + + +template +static NO_INLINE char * writeUIntText(T _x, char * p) +{ + static_assert(std::is_same_v || std::is_same_v); + using T_ = std::conditional_t< + std::is_same_v, + unsigned __int128, +#if defined(__x86_64__) +# pragma clang diagnostic push +# pragma clang diagnostic ignored "-Wbit-int-extension" + unsigned _BitInt(256) +# pragma clang diagnostic pop +#else + T +#endif + >; + + T_ x; + T_ hundred(100ULL); + if constexpr (std::is_same_v) + { + x = (T_(_x.items[T::_impl::little(1)]) << 64) + T_(_x.items[T::_impl::little(0)]); + } + else + { +#if defined(__x86_64__) + x = (T_(_x.items[T::_impl::little(3)]) << 192) + (T_(_x.items[T::_impl::little(2)]) << 128) + + (T_(_x.items[T::_impl::little(1)]) << 64) + T_(_x.items[T::_impl::little(0)]); +#else + x = _x; +#endif + } + + int len = digits10(x); + auto * pp = p + len; + while (x >= hundred) + { + const auto i = x % hundred; + x /= hundred; + pp -= 2; + outTwoDigits(pp, i); + } + if (x < 10) + *p = '0' + x; + else + outTwoDigits(p, x); + return p + len; +} + +static ALWAYS_INLINE inline char * writeLeadingMinus(char * pos) +{ + *pos = '-'; + return pos + 1; +} + +template +static ALWAYS_INLINE inline char * writeSIntText(T x, char * pos) +{ + static_assert(std::is_same_v || std::is_same_v); + + using UnsignedT = make_unsigned_t; + static constexpr T min_int = UnsignedT(1) << (sizeof(T) * 8 - 1); + + if (unlikely(x == min_int)) + { + if constexpr (std::is_same_v) + { + const char * res = "-170141183460469231731687303715884105728"; + memcpy(pos, res, strlen(res)); + return pos + strlen(res); + } + else if constexpr (std::is_same_v) + { + const char * res = "-57896044618658097711785492504343953926634992332820282019728792003956564819968"; + memcpy(pos, res, strlen(res)); + return pos + strlen(res); + } + } + + if (x < 0) + { + x = -x; + pos = writeLeadingMinus(pos); + } + return writeUIntText(UnsignedT(x), pos); +} +} + +template +char * itoa(T i, char * p) +{ + return convert::itoa(i, p); +} + +template <> +char * itoa(UInt8 i, char * p) +{ + return convert::itoa(uint8_t(i), p); +} + +template <> +char * itoa(Int8 i, char * p) +{ + return convert::itoa(int8_t(i), p); +} + +template <> +char * itoa(UInt128 i, char * p) +{ + return writeUIntText(i, p); +} + +template <> +char * itoa(Int128 i, char * p) +{ + return writeSIntText(i, p); +} + +template <> +char * itoa(UInt256 i, char * p) +{ + return writeUIntText(i, p); +} + +template <> +char * itoa(Int256 i, char * p) +{ + return writeSIntText(i, p); +} + +#define FOR_MISSING_INTEGER_TYPES(M) \ + M(int8_t) \ + M(uint8_t) \ + M(UInt16) \ + M(UInt32) \ + M(UInt64) \ + M(Int16) \ + M(Int32) \ + M(Int64) + +#define INSTANTIATION(T) template char * itoa(T i, char * p); +FOR_MISSING_INTEGER_TYPES(INSTANTIATION) + +#undef FOR_MISSING_INTEGER_TYPES +#undef INSTANTIATION + + +#define DIGITS_INTEGER_TYPES(M) \ + M(uint8_t) \ + M(UInt8) \ + M(UInt16) \ + M(UInt32) \ + M(UInt64) \ + M(UInt128) \ + M(UInt256) + +#define INSTANTIATION(T) template int digits10(T x); +DIGITS_INTEGER_TYPES(INSTANTIATION) + +#undef DIGITS_INTEGER_TYPES +#undef INSTANTIATION diff --git a/base/base/itoa.h b/base/base/itoa.h index a36eecaf1e5..71603cdeb88 100644 --- a/base/base/itoa.h +++ b/base/base/itoa.h @@ -1,474 +1,46 @@ #pragma once -// Based on https://github.com/amdn/itoa and combined with our optimizations -// -//=== itoa.h - Fast integer to ascii conversion --*- C++ -*-// -// -// The MIT License (MIT) -// Copyright (c) 2016 Arturo Martin-de-Nicolas -// -// Permission is hereby granted, free of charge, to any person obtaining a copy -// of this software and associated documentation files (the "Software"), to deal -// in the Software without restriction, including without limitation the rights -// to use, copy, modify, merge, publish, distribute, sublicense, and/or sell -// copies of the Software, and to permit persons to whom the Software is -// furnished to do so, subject to the following conditions: -// -// The above copyright notice and this permission notice shall be included -// in all copies or substantial portions of the Software. -// -// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR -// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, -// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE -// AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER -// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, -// OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE -// SOFTWARE. -//===----------------------------------------------------------------------===// - -#include -#include -#include -#include #include +template char * itoa(T i, char * p); -template -inline int digits10(T x) -{ - if (x < 10ULL) - return 1; - if (x < 100ULL) - return 2; - if (x < 1000ULL) - return 3; +template <> char * itoa(UInt8 i, char * p); +template <> char * itoa(Int8 i, char * p); +template <> char * itoa(UInt128 i, char * p); +template <> char * itoa(Int128 i, char * p); +template <> char * itoa(UInt256 i, char * p); +template <> char * itoa(Int256 i, char * p); - if (x < 1000000000000ULL) - { - if (x < 100000000ULL) - { - if (x < 1000000ULL) - { - if (x < 10000ULL) - return 4; - else - return 5 + (x >= 100000ULL); - } +#define FOR_MISSING_INTEGER_TYPES(M) \ + M(int8_t) \ + M(uint8_t) \ + M(UInt16) \ + M(UInt32) \ + M(UInt64) \ + M(Int16) \ + M(Int32) \ + M(Int64) - return 7 + (x >= 10000000ULL); - } +#define INSTANTIATION(T) \ + extern template char * itoa(T i, char * p); +FOR_MISSING_INTEGER_TYPES(INSTANTIATION) - if (x < 10000000000ULL) - return 9 + (x >= 1000000000ULL); - - return 11 + (x >= 100000000000ULL); - } - - return 12 + digits10(x / 1000000000000ULL); -} +#undef FOR_MISSING_INTEGER_TYPES +#undef INSTANTIATION -namespace impl -{ +template int digits10(T x); -template -static constexpr T pow10(size_t x) -{ - return x ? 10 * pow10(x - 1) : 1; -} - -// Division by a power of 10 is implemented using a multiplicative inverse. -// This strength reduction is also done by optimizing compilers, but -// presently the fastest results are produced by using the values -// for the multiplication and the shift as given by the algorithm -// described by Agner Fog in "Optimizing Subroutines in Assembly Language" -// -// http://www.agner.org/optimize/optimizing_assembly.pdf -// -// "Integer division by a constant (all processors) -// A floating point number can be divided by a constant by multiplying -// with the reciprocal. If we want to do the same with integers, we have -// to scale the reciprocal by 2n and then shift the product to the right -// by n. There are various algorithms for finding a suitable value of n -// and compensating for rounding errors. The algorithm described below -// was invented by Terje Mathisen, Norway, and not published elsewhere." - -/// Division by constant is performed by: -/// 1. Adding 1 if needed; -/// 2. Multiplying by another constant; -/// 3. Shifting right by another constant. -template -struct Division -{ - static constexpr bool add{add_}; - static constexpr UInt multiplier{multiplier_}; - static constexpr unsigned shift{shift_}; -}; - -/// Select a type with appropriate number of bytes from the list of types. -/// First parameter is the number of bytes requested. Then goes a list of types with 1, 2, 4, ... number of bytes. -/// Example: SelectType<4, uint8_t, uint16_t, uint32_t, uint64_t> will select uint32_t. -template -struct SelectType -{ - using Result = typename SelectType::Result; -}; - -template -struct SelectType<1, T, Ts...> -{ - using Result = T; -}; - - -/// Division by 10^N where N is the size of the type. -template -using DivisionBy10PowN = typename SelectType -< - N, - Division, /// divide by 10 - Division, /// divide by 100 - Division, /// divide by 10000 - Division /// divide by 100000000 ->::Result; - -template -using UnsignedOfSize = typename SelectType -< - N, - uint8_t, - uint16_t, - uint32_t, - uint64_t, - __uint128_t ->::Result; - -/// Holds the result of dividing an unsigned N-byte variable by 10^N resulting in -template -struct QuotientAndRemainder -{ - UnsignedOfSize quotient; // quotient with fewer than 2*N decimal digits - UnsignedOfSize remainder; // remainder with at most N decimal digits -}; - -template -QuotientAndRemainder static inline split(UnsignedOfSize value) -{ - constexpr DivisionBy10PowN division; - - UnsignedOfSize quotient = (division.multiplier * (UnsignedOfSize<2 * N>(value) + division.add)) >> division.shift; - UnsignedOfSize remainder = static_cast>(value - quotient * pow10>(N)); - - return {quotient, remainder}; -} - - -static inline char * outDigit(char * p, uint8_t value) -{ - *p = '0' + value; - ++p; - return p; -} - -// Using a lookup table to convert binary numbers from 0 to 99 -// into ascii characters as described by Andrei Alexandrescu in -// https://www.facebook.com/notes/facebook-engineering/three-optimization-tips-for-c/10151361643253920/ - -static const char digits[201] = "00010203040506070809" - "10111213141516171819" - "20212223242526272829" - "30313233343536373839" - "40414243444546474849" - "50515253545556575859" - "60616263646566676869" - "70717273747576777879" - "80818283848586878889" - "90919293949596979899"; - -static inline char * outTwoDigits(char * p, uint8_t value) -{ - memcpy(p, &digits[value * 2], 2); - p += 2; - return p; -} - - -namespace convert -{ - template static char * head(char * p, UInt u); - template static char * tail(char * p, UInt u); - - //===----------------------------------------------------------===// - // head: find most significant digit, skip leading zeros - //===----------------------------------------------------------===// - - // "x" contains quotient and remainder after division by 10^N - // quotient is less than 10^N - template - static inline char * head(char * p, QuotientAndRemainder x) - { - p = head(p, UnsignedOfSize(x.quotient)); - p = tail(p, x.remainder); - return p; - } - - // "u" is less than 10^2*N - template - static inline char * head(char * p, UInt u) - { - return u < pow10>(N) - ? head(p, UnsignedOfSize(u)) - : head(p, split(u)); - } - - // recursion base case, selected when "u" is one byte - template <> - inline char * head, 1>(char * p, UnsignedOfSize<1> u) - { - return u < 10 - ? outDigit(p, u) - : outTwoDigits(p, u); - } - - //===----------------------------------------------------------===// - // tail: produce all digits including leading zeros - //===----------------------------------------------------------===// - - // recursive step, "u" is less than 10^2*N - template - static inline char * tail(char * p, UInt u) - { - QuotientAndRemainder x = split(u); - p = tail(p, UnsignedOfSize(x.quotient)); - p = tail(p, x.remainder); - return p; - } - - // recursion base case, selected when "u" is one byte - template <> - inline char * tail, 1>(char * p, UnsignedOfSize<1> u) - { - return outTwoDigits(p, u); - } - - //===----------------------------------------------------------===// - // large values are >= 10^2*N - // where x contains quotient and remainder after division by 10^N - //===----------------------------------------------------------===// - - template - static inline char * large(char * p, QuotientAndRemainder x) - { - QuotientAndRemainder y = split(x.quotient); - p = head(p, UnsignedOfSize(y.quotient)); - p = tail(p, y.remainder); - p = tail(p, x.remainder); - return p; - } - - //===----------------------------------------------------------===// - // handle values of "u" that might be >= 10^2*N - // where N is the size of "u" in bytes - //===----------------------------------------------------------===// - - template - static inline char * uitoa(char * p, UInt u) - { - if (u < pow10>(N)) - return head(p, UnsignedOfSize(u)); - QuotientAndRemainder x = split(u); - - return u < pow10>(2 * N) - ? head(p, x) - : large(p, x); - } - - // selected when "u" is one byte - template <> - inline char * uitoa, 1>(char * p, UnsignedOfSize<1> u) - { - if (u < 10) - return outDigit(p, u); - else if (u < 100) - return outTwoDigits(p, u); - else - { - p = outDigit(p, u / 100); - p = outTwoDigits(p, u % 100); - return p; - } - } - - //===----------------------------------------------------------===// - // handle unsigned and signed integral operands - //===----------------------------------------------------------===// - - // itoa: handle unsigned integral operands (selected by SFINAE) - template && std::is_integral_v> * = nullptr> - static inline char * itoa(U u, char * p) - { - return convert::uitoa(p, u); - } - - // itoa: handle signed integral operands (selected by SFINAE) - template && std::is_integral_v> * = nullptr> - static inline char * itoa(I i, char * p) - { - // Need "mask" to be filled with a copy of the sign bit. - // If "i" is a negative value, then the result of "operator >>" - // is implementation-defined, though usually it is an arithmetic - // right shift that replicates the sign bit. - // Use a conditional expression to be portable, - // a good optimizing compiler generates an arithmetic right shift - // and avoids the conditional branch. - UnsignedOfSize mask = i < 0 ? ~UnsignedOfSize(0) : 0; - // Now get the absolute value of "i" and cast to unsigned type UnsignedOfSize. - // Cannot use std::abs() because the result is undefined - // in 2's complement systems for the most-negative value. - // Want to avoid conditional branch for performance reasons since - // CPU branch prediction will be ineffective when negative values - // occur randomly. - // Let "u" be "i" cast to unsigned type UnsignedOfSize. - // Subtract "u" from 2*u if "i" is positive or 0 if "i" is negative. - // This yields the absolute value with the desired type without - // using a conditional branch and without invoking undefined or - // implementation defined behavior: - UnsignedOfSize u = ((2 * UnsignedOfSize(i)) & ~mask) - UnsignedOfSize(i); - // Unconditionally store a minus sign when producing digits - // in a forward direction and increment the pointer only if - // the value is in fact negative. - // This avoids a conditional branch and is safe because we will - // always produce at least one digit and it will overwrite the - // minus sign when the value is not negative. - *p = '-'; - p += (mask & 1); - p = convert::uitoa(p, u); - return p; - } -} - - -template -static inline char * writeUIntText(T _x, char * p) -{ - int len = digits10(_x); - static_assert(std::is_same_v || std::is_same_v); - using T_ = std::conditional_t< - std::is_same_v, - unsigned __int128, -#if defined(__x86_64__) -#pragma clang diagnostic push -#pragma clang diagnostic ignored "-Wbit-int-extension" - unsigned _BitInt(256) -#pragma clang diagnostic pop -#else - T -#endif - >; - - T_ x; - T_ hundred(100ULL); - if constexpr (std::is_same_v) - { - x = (T_(_x.items[T::_impl::little(1)]) << 64) + T_(_x.items[T::_impl::little(0)]); - } - else - { -#if defined(__x86_64__) - x = (T_(_x.items[T::_impl::little(3)]) << 192) + (T_(_x.items[T::_impl::little(2)]) << 128) + - (T_(_x.items[T::_impl::little(1)]) << 64) + T_(_x.items[T::_impl::little(0)]); -#else - x = _x; -#endif - } - - auto * pp = p + len; - while (x >= hundred) - { - const auto i = x % hundred; - x /= hundred; - pp -= 2; - outTwoDigits(pp, i); - } - if (x < 10) - *p = '0' + x; - else - outTwoDigits(p, x); - return p + len; -} - -static inline char * writeLeadingMinus(char * pos) -{ - *pos = '-'; - return pos + 1; -} - -template -static inline char * writeSIntText(T x, char * pos) -{ - static_assert(std::is_same_v || std::is_same_v); - - using UnsignedT = make_unsigned_t; - static constexpr T min_int = UnsignedT(1) << (sizeof(T) * 8 - 1); - - if (unlikely(x == min_int)) - { - if constexpr (std::is_same_v) - { - const char * res = "-170141183460469231731687303715884105728"; - memcpy(pos, res, strlen(res)); - return pos + strlen(res); - } - else if constexpr (std::is_same_v) - { - const char * res = "-57896044618658097711785492504343953926634992332820282019728792003956564819968"; - memcpy(pos, res, strlen(res)); - return pos + strlen(res); - } - } - - if (x < 0) - { - x = -x; - pos = writeLeadingMinus(pos); - } - return writeUIntText(UnsignedT(x), pos); -} - -} - -template -char * itoa(I i, char * p) -{ - return impl::convert::itoa(i, p); -} - -template <> -inline char * itoa(char8_t i, char * p) -{ - return impl::convert::itoa(uint8_t(i), p); -} - -template <> -inline char * itoa(UInt128 i, char * p) -{ - return impl::writeUIntText(i, p); -} - -template <> -inline char * itoa(Int128 i, char * p) -{ - return impl::writeSIntText(i, p); -} - -template <> -inline char * itoa(UInt256 i, char * p) -{ - return impl::writeUIntText(i, p); -} - -template <> -inline char * itoa(Int256 i, char * p) -{ - return impl::writeSIntText(i, p); -} +#define DIGITS_INTEGER_TYPES(M) \ + M(uint8_t) \ + M(UInt8) \ + M(UInt16) \ + M(UInt32) \ + M(UInt64) \ + M(UInt128) \ + M(UInt256) +#define INSTANTIATION(T) \ + extern template int digits10(T x); +DIGITS_INTEGER_TYPES(INSTANTIATION) +#undef DIGITS_INTEGER_TYPES +#undef INSTANTIATION diff --git a/src/Functions/CMakeLists.txt b/src/Functions/CMakeLists.txt index ac3e3671ae0..dea369a508a 100644 --- a/src/Functions/CMakeLists.txt +++ b/src/Functions/CMakeLists.txt @@ -37,6 +37,7 @@ list (APPEND PUBLIC_LIBS clickhouse_dictionaries_embedded clickhouse_parsers ch_contrib::consistent_hashing + common dbms ch_contrib::metrohash ch_contrib::murmurhash