X-Git-Url: http://plrg.eecs.uci.edu/git/?p=folly.git;a=blobdiff_plain;f=folly%2FBits.h;h=ade5bbeeb7b39572f47e73e6405a54184d72984b;hp=9795f8f5650f1337d0509c48769dceb1b37fe8c0;hb=f975d3c54800b7065985e5e548ce1d91fef4926c;hpb=8ff6fe9e5593c1a5a0de24577f1e423fb6eae503

diff --git a/folly/Bits.h b/folly/Bits.h
index 9795f8f5..ade5bbee 100644
--- a/folly/Bits.h
+++ b/folly/Bits.h
@@ -1,5 +1,5 @@
 /*
- * 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.
@@ -17,18 +17,24 @@
 /**
  * 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, findFirstSet(x) == 1 + floor(log2(x))
+ *    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
@@ -46,23 +52,26 @@
  * @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
+#if !defined(__clang__) && !(defined(_MSC_VER) && (_MSC_VER < 1900))
+#define FOLLY_INTRINSIC_CONSTEXPR constexpr
+#else
+// GCC and MSVC 2015+ are the only compilers with
+// intrinsics constexpr.
+#define FOLLY_INTRINSIC_CONSTEXPR const
 #endif
 
-#include "folly/detail/BitIteratorDetail.h"
-#include "folly/Likely.h"
+#include <folly/Portability.h>
+#include <folly/portability/Builtins.h>
+
+#include <folly/Assume.h>
+#include <folly/detail/BitIteratorDetail.h>
+#include <folly/Likely.h>
 
-#include <byteswap.h>
 #include <cassert>
+#include <cstring>
 #include <cinttypes>
-#include <cstring>  // for ffs, ffsl, ffsll
-#include <endian.h>
 #include <iterator>
 #include <limits>
 #include <type_traits>
@@ -72,127 +81,96 @@
 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));
+  return findFirstSet(static_cast<typename std::make_unsigned<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;
-}
-
-}  // 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)),
-  unsigned int>::type
-  findLastSet(T x) {
-  return x ? 8 * sizeof(unsigned long long) - __builtin_clzll(x) : 0;
-}
-
-#else  /* !__GNUC__ */
-
-template <class T>
-typename std::enable_if<
-  (std::is_integral<T>::value &&
-   std::is_unsigned<T>::value),
+   sizeof(T) > sizeof(unsigned long) &&
+   sizeof(T) <= sizeof(unsigned long long)),
   unsigned int>::type
   findLastSet(T x) {
-  return detail:findLastSetPortable(x);
+  return x ? ((8 * sizeof(unsigned long long) - 1) ^ __builtin_clzll(x)) + 1
+           : 0;
 }
 
-#endif
-
 template <class T>
+inline FOLLY_INTRINSIC_CONSTEXPR
 typename std::enable_if<
   (std::is_integral<T>::value &&
    std::is_signed<T>::value),
@@ -201,112 +179,109 @@ typename std::enable_if<
   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);
-};
-
-#define FB_GEN(t, fn) \
-template<> inline t EndianIntBase<t>::swap(t x) { return fn(x); }
-
-// 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)
+template <size_t Size>
+struct uint_types_by_size;
+
+#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;                              \
+  };
+
+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
 
 // big* convert between native and big-endian representations
@@ -329,20 +304,29 @@ struct EndianInt : public detail::EndianIntBase<T> {
 
 class Endian {
  public:
+  enum class Order : uint8_t {
+    LITTLE,
+    BIG
+  };
+
+  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
@@ -371,7 +355,7 @@ class BitIterator
   /**
    * Return the number of bits in an element of the underlying iterator.
    */
-  static size_t bitsPerBlock() {
+  static unsigned int bitsPerBlock() {
     return std::numeric_limits<
       typename std::make_unsigned<
         typename std::iterator_traits<BaseIter>::value_type
@@ -383,9 +367,9 @@ class BitIterator
    * Construct a BitIterator that points at a given bit offset (default 0)
    * in iter.
    */
-  explicit BitIterator(const BaseIter& iter, size_t bitOffset=0)
+  explicit BitIterator(const BaseIter& iter, size_t bitOff=0)
     : bititerator_detail::BitIteratorBase<BaseIter>::type(iter),
-      bitOffset_(bitOffset) {
+      bitOffset_(bitOff) {
     assert(bitOffset_ < bitsPerBlock());
   }
 
@@ -424,7 +408,7 @@ class BitIterator
 
   void advance(ssize_t n) {
     size_t bpb = bitsPerBlock();
-    ssize_t blocks = n / bpb;
+    ssize_t blocks = n / ssize_t(bpb);
     bitOffset_ += n % bpb;
     if (bitOffset_ >= bpb) {
       bitOffset_ -= bpb;
@@ -452,12 +436,12 @@ class BitIterator
   }
 
   ssize_t distance_to(const BitIterator& other) const {
-    return
-      (other.base_reference() - this->base_reference()) * bitsPerBlock() +
-      (other.bitOffset_ - bitOffset_);
+    return ssize_t(
+        (other.base_reference() - this->base_reference()) * bitsPerBlock() +
+        other.bitOffset_ - bitOffset_);
   }
 
-  ssize_t bitOffset_;
+  size_t bitOffset_;
 };
 
 /**
@@ -514,26 +498,36 @@ BitIterator<BaseIter> findFirstSet(BitIterator<BaseIter> begin,
 }
 
 
-namespace detail {
-
 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> {
+    T,
+    typename std::enable_if<std::is_pod<T>::value>::type> {
+  Unaligned() = default;  // uninitialized
+  /* implicit */ Unaligned(T v) : value(v) { }
   T value;
-} __attribute__((packed));
-
-}  // namespace detail
+} FOLLY_PACK_ATTR;
+FOLLY_PACK_POP
 
 /**
  * Read an unaligned value of type T and return it.
  */
 template <class T>
 inline T loadUnaligned(const void* p) {
-  static_assert(alignof(detail::Unaligned<T>) == 1, "Invalid alignment");
-  return static_cast<const detail::Unaligned<T>*>(p)->value;
+  static_assert(sizeof(Unaligned<T>) == sizeof(T), "Invalid unaligned size");
+  static_assert(alignof(Unaligned<T>) == 1, "Invalid alignment");
+  if (kHasUnalignedAccess) {
+    return static_cast<const Unaligned<T>*>(p)->value;
+  } else {
+    T value;
+    memcpy(&value, p, sizeof(T));
+    return value;
+  }
 }
 
 /**
@@ -541,11 +535,19 @@ inline T loadUnaligned(const void* p) {
  */
 template <class T>
 inline void storeUnaligned(void* p, T value) {
-  static_assert(alignof(detail::Unaligned<T>) == 1, "Invalid alignment");
-  static_cast<detail::Unaligned<T>*>(p)->value = value;
+  static_assert(sizeof(Unaligned<T>) == sizeof(T), "Invalid unaligned size");
+  static_assert(alignof(Unaligned<T>) == 1, "Invalid alignment");
+  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_ */
-