ClickHouse/base/common/itoa.h
2020-02-14 17:48:30 +03:00

411 lines
13 KiB
C++

#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 <cstdint>
#include <cstddef>
#include <cstring>
#include <type_traits>
#include "likely.h"
using int128_t = __int128;
using uint128_t = unsigned __int128;
namespace impl
{
template <typename T>
static constexpr T pow10(size_t x)
{
return x ? 10 * pow10<T>(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 <typename UInt, bool add_, UInt multiplier_, unsigned shift_>
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 <size_t N, typename T, typename... Ts>
struct SelectType
{
using Result = typename SelectType<N / 2, Ts...>::Result;
};
template <typename T, typename... Ts>
struct SelectType<1, T, Ts...>
{
using Result = T;
};
/// Division by 10^N where N is the size of the type.
template <size_t N>
using DivisionBy10PowN = typename SelectType
<
N,
Division<uint8_t, 0, 205U, 11>, /// divide by 10
Division<uint16_t, 1, 41943U, 22>, /// divide by 100
Division<uint32_t, 0, 3518437209U, 45>, /// divide by 10000
Division<uint64_t, 0, 12379400392853802749ULL, 90> /// divide by 100000000
>::Result;
template <size_t N>
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 <size_t N>
struct QuotientAndRemainder
{
UnsignedOfSize<N> quotient; // quotient with fewer than 2*N decimal digits
UnsignedOfSize<N / 2> remainder; // remainder with at most N decimal digits
};
template <size_t N>
QuotientAndRemainder<N> static inline split(UnsignedOfSize<N> value)
{
constexpr DivisionBy10PowN<N> division;
UnsignedOfSize<N> quotient = (division.multiplier * (UnsignedOfSize<2 * N>(value) + division.add)) >> division.shift;
UnsignedOfSize<N / 2> remainder = value - quotient * pow10<UnsignedOfSize<N / 2>>(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 <typename UInt, size_t N = sizeof(UInt)> static char * head(char * p, UInt u);
template <typename UInt, size_t N = sizeof(UInt)> 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 <size_t N>
static inline char * head(char * p, QuotientAndRemainder<N> x)
{
p = head(p, UnsignedOfSize<N / 2>(x.quotient));
p = tail(p, x.remainder);
return p;
}
// "u" is less than 10^2*N
template <typename UInt, size_t N>
static inline char * head(char * p, UInt u)
{
return u < pow10<UnsignedOfSize<N>>(N)
? head(p, UnsignedOfSize<N / 2>(u))
: head<N>(p, split<N>(u));
}
// recursion base case, selected when "u" is one byte
template <>
inline char * head<UnsignedOfSize<1>, 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 <typename UInt, size_t N>
static inline char * tail(char * p, UInt u)
{
QuotientAndRemainder<N> x = split<N>(u);
p = tail(p, UnsignedOfSize<N / 2>(x.quotient));
p = tail(p, x.remainder);
return p;
}
// recursion base case, selected when "u" is one byte
template <>
inline char * tail<UnsignedOfSize<1>, 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 <size_t N>
static inline char * large(char * p, QuotientAndRemainder<N> x)
{
QuotientAndRemainder<N> y = split<N>(x.quotient);
p = head(p, UnsignedOfSize<N / 2>(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 <typename UInt, size_t N = sizeof(UInt)>
static inline char * uitoa(char * p, UInt u)
{
if (u < pow10<UnsignedOfSize<N>>(N))
return head(p, UnsignedOfSize<N / 2>(u));
QuotientAndRemainder<N> x = split<N>(u);
return u < pow10<UnsignedOfSize<N>>(2 * N)
? head<N>(p, x)
: large<N>(p, x);
}
// selected when "u" is one byte
template <>
inline char * uitoa<UnsignedOfSize<1>, 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 <typename U, std::enable_if_t<!std::is_signed_v<U> && std::is_integral_v<U>> * = nullptr>
static inline char * itoa(U u, char * p)
{
return convert::uitoa(p, u);
}
// itoa: handle signed integral operands (selected by SFINAE)
template <typename I, size_t N = sizeof(I), std::enable_if_t<std::is_signed_v<I> && std::is_integral_v<I>> * = 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<N> mask = i < 0 ? ~UnsignedOfSize<N>(0) : 0;
// Now get the absolute value of "i" and cast to unsigned type UnsignedOfSize<N>.
// 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<N>.
// 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<N> u = ((2 * UnsignedOfSize<N>(i)) & ~mask) - UnsignedOfSize<N>(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;
}
}
static inline int digits10(uint128_t x)
{
if (x < 10ULL)
return 1;
if (x < 100ULL)
return 2;
if (x < 1000ULL)
return 3;
if (x < 1000000000000ULL)
{
if (x < 100000000ULL)
{
if (x < 1000000ULL)
{
if (x < 10000ULL)
return 4;
else
return 5 + (x >= 100000ULL);
}
return 7 + (x >= 10000000ULL);
}
if (x < 10000000000ULL)
return 9 + (x >= 1000000000ULL);
return 11 + (x >= 100000000000ULL);
}
return 12 + digits10(x / 1000000000000ULL);
}
static inline char * writeUIntText(uint128_t x, char * p)
{
int len = digits10(x);
auto pp = p + len;
while (x >= 100)
{
const auto i = x % 100;
x /= 100;
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;
}
static inline char * writeSIntText(int128_t x, char * pos)
{
static const int128_t min_int128 = int128_t(0x8000000000000000ll) << 64;
if (unlikely(x == min_int128))
{
memcpy(pos, "-170141183460469231731687303715884105728", 40);
return pos + 40;
}
if (x < 0)
{
x = -x;
pos = writeLeadingMinus(pos);
}
return writeUIntText(static_cast<uint128_t>(x), pos);
}
}
template <typename I>
char * itoa(I i, char * p)
{
return impl::convert::itoa(i, p);
}
template <>
inline char * itoa<uint128_t>(uint128_t i, char * p)
{
return impl::writeUIntText(i, p);
}
template <>
inline char * itoa<int128_t>(int128_t i, char * p)
{
return impl::writeSIntText(i, p);
}