/*
- * Copyright 2012 Facebook, Inc.
+ * Copyright 2017 Facebook, Inc.
*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
/**
* Various low-level, bit-manipulation routines.
*
- * findFirstSet(x)
+ * findFirstSet(x) [constexpr]
* find first (least significant) bit set in a value of an integral type,
* 1-based (like ffs()). 0 = no bits are set (x == 0)
*
- * findLastSet(x)
+ * findLastSet(x) [constexpr]
* find last (most significant) bit set in a value of an integral type,
* 1-based. 0 = no bits are set (x == 0)
* for x != 0, findLastSet(x) == 1 + floor(log2(x))
*
- * nextPowTwo(x)
+ * nextPowTwo(x) [constexpr]
* Finds the next power of two >= x.
*
+ * isPowTwo(x) [constexpr]
+ * return true iff x is a power of two
+ *
+ * popcount(x)
+ * return the number of 1 bits in x
+ *
* Endian
* convert between native, big, and little endian representation
* Endian::big(x) big <-> native
* Endian::little(x) little <-> native
* Endian::swap(x) big <-> little
*
- * BitIterator
- * Wrapper around an iterator over an integral type that iterates
- * over its underlying bits in MSb to LSb order
- *
- * findFirstSet(BitIterator begin, BitIterator end)
- * return a BitIterator pointing to the first 1 bit in [begin, end), or
- * end if all bits in [begin, end) are 0
- *
* @author Tudor Bosman (tudorb@fb.com)
*/
-#ifndef FOLLY_BITS_H_
-#define FOLLY_BITS_H_
-
-#include "folly/Portability.h"
+#pragma once
-#ifndef _GNU_SOURCE
-#define _GNU_SOURCE 1
+// MSVC does not support intrinsics constexpr
+#if defined(_MSC_VER)
+#define FOLLY_INTRINSIC_CONSTEXPR const
+#else
+#define FOLLY_INTRINSIC_CONSTEXPR constexpr
#endif
-#include "folly/detail/BitIteratorDetail.h"
-#include "folly/Likely.h"
-
-#include <byteswap.h>
#include <cassert>
#include <cinttypes>
-#include <cstring> // for ffs, ffsl, ffsll
-#include <endian.h>
-#include <iterator>
+#include <cstdint>
+#include <cstring>
#include <limits>
#include <type_traits>
-#include <boost/iterator/iterator_adaptor.hpp>
-#include <stdint.h>
+
+#include <folly/Portability.h>
+#include <folly/lang/Assume.h>
+#include <folly/portability/Builtins.h>
namespace folly {
// Generate overloads for findFirstSet as wrappers around
-// appropriate ffs, ffsl, ffsll functions from glibc.
-// We first define these overloads for signed types (because ffs, ffsl, ffsll
-// take int, long, and long long as arguments, respectively) and then
-// define an overload for unsigned that forwards to the overload for the
-// corresponding signed type.
+// appropriate ffs, ffsl, ffsll gcc builtins
template <class T>
+inline FOLLY_INTRINSIC_CONSTEXPR
typename std::enable_if<
(std::is_integral<T>::value &&
- std::is_signed<T>::value &&
- (std::numeric_limits<T>::digits <= std::numeric_limits<int>::digits)),
+ std::is_unsigned<T>::value &&
+ sizeof(T) <= sizeof(unsigned int)),
unsigned int>::type
findFirstSet(T x) {
- return ::ffs(static_cast<int>(x));
+ return static_cast<unsigned int>(__builtin_ffs(static_cast<int>(x)));
}
template <class T>
+inline FOLLY_INTRINSIC_CONSTEXPR
typename std::enable_if<
(std::is_integral<T>::value &&
- std::is_signed<T>::value &&
- (std::numeric_limits<T>::digits > std::numeric_limits<int>::digits) &&
- (std::numeric_limits<T>::digits <= std::numeric_limits<long>::digits)),
+ std::is_unsigned<T>::value &&
+ sizeof(T) > sizeof(unsigned int) &&
+ sizeof(T) <= sizeof(unsigned long)),
unsigned int>::type
findFirstSet(T x) {
- return ::ffsl(static_cast<long>(x));
+ return static_cast<unsigned int>(__builtin_ffsl(static_cast<long>(x)));
}
-#ifdef FOLLY_HAVE_FFSLL
-
template <class T>
+inline FOLLY_INTRINSIC_CONSTEXPR
typename std::enable_if<
(std::is_integral<T>::value &&
- std::is_signed<T>::value &&
- (std::numeric_limits<T>::digits > std::numeric_limits<long>::digits) &&
- (std::numeric_limits<T>::digits <= std::numeric_limits<long long>::digits)),
+ std::is_unsigned<T>::value &&
+ sizeof(T) > sizeof(unsigned long) &&
+ sizeof(T) <= sizeof(unsigned long long)),
unsigned int>::type
findFirstSet(T x) {
- return ::ffsll(static_cast<long long>(x));
+ return static_cast<unsigned int>(__builtin_ffsll(static_cast<long long>(x)));
}
-#endif
-
template <class T>
+inline FOLLY_INTRINSIC_CONSTEXPR
typename std::enable_if<
- (std::is_integral<T>::value &&
- !std::is_signed<T>::value),
+ (std::is_integral<T>::value && std::is_signed<T>::value),
unsigned int>::type
findFirstSet(T x) {
- // Note that conversion from an unsigned type to the corresponding signed
+ // Note that conversion from a signed type to the corresponding unsigned
// type is technically implementation-defined, but will likely work
// on any impementation that uses two's complement.
- return findFirstSet(static_cast<typename std::make_signed<T>::type>(x));
-}
-
-namespace detail {
-
-// Portable, but likely slow...
-inline unsigned int findLastSetPortable(uint64_t x) {
- unsigned int r = (x != 0); // 1-based index, except for x==0
- while (x >>= 1) {
- ++r;
- }
- return r;
+ return findFirstSet(static_cast<typename std::make_unsigned<T>::type>(x));
}
-} // namespace detail
-
-#ifdef __GNUC__
-
// findLastSet: return the 1-based index of the highest bit set
// for x > 0, findLastSet(x) == 1 + floor(log2(x))
template <class T>
+inline FOLLY_INTRINSIC_CONSTEXPR
typename std::enable_if<
(std::is_integral<T>::value &&
std::is_unsigned<T>::value &&
- (std::numeric_limits<T>::digits <=
- std::numeric_limits<unsigned int>::digits)),
+ sizeof(T) <= sizeof(unsigned int)),
unsigned int>::type
findLastSet(T x) {
- return x ? 8 * sizeof(unsigned int) - __builtin_clz(x) : 0;
+ // If X is a power of two X - Y = ((X - 1) ^ Y) + 1. Doing this transformation
+ // allows GCC to remove its own xor that it adds to implement clz using bsr
+ return x ? ((8 * sizeof(unsigned int) - 1) ^ __builtin_clz(x)) + 1 : 0;
}
template <class T>
+inline FOLLY_INTRINSIC_CONSTEXPR
typename std::enable_if<
(std::is_integral<T>::value &&
std::is_unsigned<T>::value &&
- (std::numeric_limits<T>::digits >
- std::numeric_limits<unsigned int>::digits) &&
- (std::numeric_limits<T>::digits <=
- std::numeric_limits<unsigned long>::digits)),
+ sizeof(T) > sizeof(unsigned int) &&
+ sizeof(T) <= sizeof(unsigned long)),
unsigned int>::type
findLastSet(T x) {
- return x ? 8 * sizeof(unsigned long) - __builtin_clzl(x) : 0;
+ return x ? ((8 * sizeof(unsigned long) - 1) ^ __builtin_clzl(x)) + 1 : 0;
}
template <class T>
+inline FOLLY_INTRINSIC_CONSTEXPR
typename std::enable_if<
(std::is_integral<T>::value &&
std::is_unsigned<T>::value &&
- (std::numeric_limits<T>::digits >
- std::numeric_limits<unsigned long>::digits) &&
- (std::numeric_limits<T>::digits <=
- std::numeric_limits<unsigned long long>::digits)),
+ sizeof(T) > sizeof(unsigned long) &&
+ sizeof(T) <= sizeof(unsigned long long)),
unsigned int>::type
findLastSet(T x) {
- return x ? 8 * sizeof(unsigned long long) - __builtin_clzll(x) : 0;
+ return x ? ((8 * sizeof(unsigned long long) - 1) ^ __builtin_clzll(x)) + 1
+ : 0;
}
-#else /* !__GNUC__ */
-
-template <class T>
-typename std::enable_if<
- (std::is_integral<T>::value &&
- std::is_unsigned<T>::value),
- unsigned int>::type
- findLastSet(T x) {
- return detail:findLastSetPortable(x);
-}
-
-#endif
-
template <class T>
+inline FOLLY_INTRINSIC_CONSTEXPR
typename std::enable_if<
(std::is_integral<T>::value &&
std::is_signed<T>::value),
return findLastSet(static_cast<typename std::make_unsigned<T>::type>(x));
}
-namespace detail {
-
template <class T>
-inline
+inline FOLLY_INTRINSIC_CONSTEXPR
typename std::enable_if<
std::is_integral<T>::value && std::is_unsigned<T>::value,
T>::type
-nextPowTwoPortable(T v) {
- if (UNLIKELY(v == 0)) {
- return 1;
- }
-
- --v;
- for (uint32_t i = 1; i < sizeof(T) * 8; i <<= 8) {
- v |= (v >> i);
- v |= (v >> (i << 1));
- v |= (v >> (i << 2));
- v |= (v >> (i << 3));
- v |= (v >> (i << 4));
- v |= (v >> (i << 5));
- v |= (v >> (i << 6));
- v |= (v >> (i << 7));
- }
- return v + 1;
+nextPowTwo(T v) {
+ return v ? (T(1) << findLastSet(v - 1)) : 1;
}
-} // namespace detail
-
-#ifdef __GNUC__
-
template <class T>
-inline
-typename std::enable_if<
- std::is_integral<T>::value && std::is_unsigned<T>::value,
- T>::type
-nextPowTwo(T v) {
- if (UNLIKELY(v == 0)) {
- return 1;
- }
- return 1ul << findLastSet(v - 1);
+inline FOLLY_INTRINSIC_CONSTEXPR typename std::
+ enable_if<std::is_integral<T>::value && std::is_unsigned<T>::value, T>::type
+ prevPowTwo(T v) {
+ return v ? (T(1) << (findLastSet(v) - 1)) : 0;
}
-#else /* __GNUC__ */
-
template <class T>
-inline
-typename std::enable_if<
- std::is_integral<T>::value && std::is_unsigned<T>::value,
- T>::type
-nextPowTwo(T v) {
- return detail::nextPowTwoPortable(v);
+inline constexpr typename std::enable_if<
+ std::is_integral<T>::value && std::is_unsigned<T>::value,
+ bool>::type
+isPowTwo(T v) {
+ return (v != 0) && !(v & (v - 1));
}
-#endif /* __GNUC__ */
-
+/**
+ * Population count
+ */
+template <class T>
+inline typename std::enable_if<
+ (std::is_integral<T>::value &&
+ std::is_unsigned<T>::value &&
+ sizeof(T) <= sizeof(unsigned int)),
+ size_t>::type
+ popcount(T x) {
+ return size_t(__builtin_popcount(x));
+}
+template <class T>
+inline typename std::enable_if<
+ (std::is_integral<T>::value &&
+ std::is_unsigned<T>::value &&
+ sizeof(T) > sizeof(unsigned int) &&
+ sizeof(T) <= sizeof(unsigned long long)),
+ size_t>::type
+ popcount(T x) {
+ return size_t(__builtin_popcountll(x));
+}
/**
* Endianness detection and manipulation primitives.
*/
namespace detail {
-template <class T>
-struct EndianIntBase {
- public:
- static T swap(T x);
-};
+template <size_t Size>
+struct uint_types_by_size;
-#define FB_GEN(t, fn) \
-template<> inline t EndianIntBase<t>::swap(t x) { return fn(x); }
+#define FB_GEN(sz, fn) \
+ static inline uint##sz##_t byteswap_gen(uint##sz##_t v) { \
+ return fn(v); \
+ } \
+ template <> \
+ struct uint_types_by_size<sz / 8> { \
+ using type = uint##sz##_t; \
+ };
-// fn(x) expands to (x) if the second argument is empty, which is exactly
-// what we want for [u]int8_t
-FB_GEN( int8_t,)
-FB_GEN(uint8_t,)
-FB_GEN( int64_t, bswap_64)
-FB_GEN(uint64_t, bswap_64)
-FB_GEN( int32_t, bswap_32)
-FB_GEN(uint32_t, bswap_32)
-FB_GEN( int16_t, bswap_16)
-FB_GEN(uint16_t, bswap_16)
+FB_GEN(8, uint8_t)
+#ifdef _MSC_VER
+FB_GEN(64, _byteswap_uint64)
+FB_GEN(32, _byteswap_ulong)
+FB_GEN(16, _byteswap_ushort)
+#else
+FB_GEN(64, __builtin_bswap64)
+FB_GEN(32, __builtin_bswap32)
+FB_GEN(16, __builtin_bswap16)
+#endif
#undef FB_GEN
-#if __BYTE_ORDER == __LITTLE_ENDIAN
-
template <class T>
-struct EndianInt : public detail::EndianIntBase<T> {
- public:
- static T big(T x) { return EndianInt::swap(x); }
- static T little(T x) { return x; }
-};
-
-#elif __BYTE_ORDER == __BIG_ENDIAN
-
-template <class T>
-struct EndianInt : public detail::EndianIntBase<T> {
- public:
- static T big(T x) { return x; }
- static T little(T x) { return EndianInt::swap(x); }
+struct EndianInt {
+ static_assert(
+ (std::is_integral<T>::value && !std::is_same<T, bool>::value) ||
+ std::is_floating_point<T>::value,
+ "template type parameter must be non-bool integral or floating point");
+ static T swap(T x) {
+ // we implement this with memcpy because that is defined behavior in C++
+ // we rely on compilers to optimize away the memcpy calls
+ constexpr auto s = sizeof(T);
+ using B = typename uint_types_by_size<s>::type;
+ B b;
+ std::memcpy(&b, &x, s);
+ b = byteswap_gen(b);
+ std::memcpy(&x, &b, s);
+ return x;
+ }
+ static T big(T x) {
+ return kIsLittleEndian ? EndianInt::swap(x) : x;
+ }
+ static T little(T x) {
+ return kIsBigEndian ? EndianInt::swap(x) : x;
+ }
};
-#else
-# error Your machine uses a weird endianness!
-#endif /* __BYTE_ORDER */
-
-} // namespace detail
+} // namespace detail
// big* convert between native and big-endian representations
// little* convert between native and little-endian representations
BIG
};
- static constexpr Order order =
-#if __BYTE_ORDER == __LITTLE_ENDIAN
- Order::LITTLE;
-#elif __BYTE_ORDER == __BIG_ENDIAN
- Order::BIG;
-#else
-# error Your machine uses a weird endianness!
-#endif /* __BYTE_ORDER */
+ static constexpr Order order = kIsLittleEndian ? Order::LITTLE : Order::BIG;
template <class T> static T swap(T x) {
- return detail::EndianInt<T>::swap(x);
+ return folly::detail::EndianInt<T>::swap(x);
}
template <class T> static T big(T x) {
- return detail::EndianInt<T>::big(x);
+ return folly::detail::EndianInt<T>::big(x);
}
template <class T> static T little(T x) {
- return detail::EndianInt<T>::little(x);
+ return folly::detail::EndianInt<T>::little(x);
}
+#if !defined(__ANDROID__)
FB_GEN(64)
FB_GEN(32)
FB_GEN(16)
FB_GEN(8)
+#endif
};
#undef FB_GEN
#undef FB_GEN2
#undef FB_GEN1
-/**
- * Fast bit iteration facility.
- */
-
-
-template <class BaseIter> class BitIterator;
-template <class BaseIter>
-BitIterator<BaseIter> findFirstSet(BitIterator<BaseIter>,
- BitIterator<BaseIter>);
-/**
- * Wrapper around an iterator over an integer type that iterates
- * over its underlying bits in LSb to MSb order.
- *
- * BitIterator models the same iterator concepts as the base iterator.
- */
-template <class BaseIter>
-class BitIterator
- : public bititerator_detail::BitIteratorBase<BaseIter>::type {
- public:
- /**
- * Return the number of bits in an element of the underlying iterator.
- */
- static size_t bitsPerBlock() {
- return std::numeric_limits<
- typename std::make_unsigned<
- typename std::iterator_traits<BaseIter>::value_type
- >::type
- >::digits;
- }
-
- /**
- * Construct a BitIterator that points at a given bit offset (default 0)
- * in iter.
- */
- explicit BitIterator(const BaseIter& iter, size_t bitOffset=0)
- : bititerator_detail::BitIteratorBase<BaseIter>::type(iter),
- bitOffset_(bitOffset) {
- assert(bitOffset_ < bitsPerBlock());
- }
-
- size_t bitOffset() const {
- return bitOffset_;
- }
-
- void advanceToNextBlock() {
- bitOffset_ = 0;
- ++this->base_reference();
- }
-
- BitIterator& operator=(const BaseIter& other) {
- this->~BitIterator();
- new (this) BitIterator(other);
- return *this;
- }
-
- private:
- friend class boost::iterator_core_access;
- friend BitIterator findFirstSet<>(BitIterator, BitIterator);
-
- typedef bititerator_detail::BitReference<
- typename std::iterator_traits<BaseIter>::reference,
- typename std::iterator_traits<BaseIter>::value_type
- > BitRef;
-
- void advanceInBlock(size_t n) {
- bitOffset_ += n;
- assert(bitOffset_ < bitsPerBlock());
- }
-
- BitRef dereference() const {
- return BitRef(*this->base_reference(), bitOffset_);
- }
-
- void advance(ssize_t n) {
- size_t bpb = bitsPerBlock();
- ssize_t blocks = n / bpb;
- bitOffset_ += n % bpb;
- if (bitOffset_ >= bpb) {
- bitOffset_ -= bpb;
- ++blocks;
- }
- this->base_reference() += blocks;
- }
-
- void increment() {
- if (++bitOffset_ == bitsPerBlock()) {
- advanceToNextBlock();
- }
- }
-
- void decrement() {
- if (bitOffset_-- == 0) {
- bitOffset_ = bitsPerBlock() - 1;
- --this->base_reference();
- }
- }
-
- bool equal(const BitIterator& other) const {
- return (bitOffset_ == other.bitOffset_ &&
- this->base_reference() == other.base_reference());
- }
-
- ssize_t distance_to(const BitIterator& other) const {
- return
- (other.base_reference() - this->base_reference()) * bitsPerBlock() +
- (other.bitOffset_ - bitOffset_);
- }
-
- ssize_t bitOffset_;
-};
-
-/**
- * Helper function, so you can write
- * auto bi = makeBitIterator(container.begin());
- */
-template <class BaseIter>
-BitIterator<BaseIter> makeBitIterator(const BaseIter& iter) {
- return BitIterator<BaseIter>(iter);
-}
-
-
-/**
- * Find first bit set in a range of bit iterators.
- * 4.5x faster than the obvious std::find(begin, end, true);
- */
-template <class BaseIter>
-BitIterator<BaseIter> findFirstSet(BitIterator<BaseIter> begin,
- BitIterator<BaseIter> end) {
- // shortcut to avoid ugly static_cast<>
- static const typename BaseIter::value_type one = 1;
-
- while (begin.base() != end.base()) {
- typename BaseIter::value_type v = *begin.base();
- // mask out the bits that don't matter (< begin.bitOffset)
- v &= ~((one << begin.bitOffset()) - 1);
- size_t firstSet = findFirstSet(v);
- if (firstSet) {
- --firstSet; // now it's 0-based
- assert(firstSet >= begin.bitOffset());
- begin.advanceInBlock(firstSet - begin.bitOffset());
- return begin;
- }
- begin.advanceToNextBlock();
- }
-
- // now begin points to the same block as end
- if (end.bitOffset() != 0) { // assume end is dereferenceable
- typename BaseIter::value_type v = *begin.base();
- // mask out the bits that don't matter (< begin.bitOffset)
- v &= ~((one << begin.bitOffset()) - 1);
- // mask out the bits that don't matter (>= end.bitOffset)
- v &= (one << end.bitOffset()) - 1;
- size_t firstSet = findFirstSet(v);
- if (firstSet) {
- --firstSet; // now it's 0-based
- assert(firstSet >= begin.bitOffset());
- begin.advanceInBlock(firstSet - begin.bitOffset());
- return begin;
- }
- }
-
- return end;
-}
-
template <class T, class Enable=void> struct Unaligned;
/**
* Representation of an unaligned value of a POD type.
*/
+FOLLY_PACK_PUSH
template <class T>
struct Unaligned<
T,
typename std::enable_if<std::is_pod<T>::value>::type> {
- Unaligned() { } // uninitialized
+ Unaligned() = default; // uninitialized
/* implicit */ Unaligned(T v) : value(v) { }
T value;
-} __attribute__((packed));
+} FOLLY_PACK_ATTR;
+FOLLY_PACK_POP
/**
* Read an unaligned value of type T and return it.
inline T loadUnaligned(const void* p) {
static_assert(sizeof(Unaligned<T>) == sizeof(T), "Invalid unaligned size");
static_assert(alignof(Unaligned<T>) == 1, "Invalid alignment");
- return static_cast<const Unaligned<T>*>(p)->value;
+ if (kHasUnalignedAccess) {
+ return static_cast<const Unaligned<T>*>(p)->value;
+ } else {
+ T value;
+ memcpy(&value, p, sizeof(T));
+ return value;
+ }
}
/**
inline void storeUnaligned(void* p, T value) {
static_assert(sizeof(Unaligned<T>) == sizeof(T), "Invalid unaligned size");
static_assert(alignof(Unaligned<T>) == 1, "Invalid alignment");
- new (p) Unaligned<T>(value);
+ if (kHasUnalignedAccess) {
+ // Prior to C++14, the spec says that a placement new like this
+ // is required to check that p is not nullptr, and to do nothing
+ // if p is a nullptr. By assuming it's not a nullptr, we get a
+ // nice loud segfault in optimized builds if p is nullptr, rather
+ // than just silently doing nothing.
+ folly::assume(p != nullptr);
+ new (p) Unaligned<T>(value);
+ } else {
+ memcpy(p, &value, sizeof(T));
+ }
}
-} // namespace folly
-
-#endif /* FOLLY_BITS_H_ */
-
+} // namespace folly
projects / folly.git / blobdiff