Export of internal Abseil changes

--
e54b9c7bbb0c58475676c268e2e19c69f4bce48a by Jorg Brown <jorg@google.com>:

Tweak ABSL_PREDICT_TRUE slightly, for better code on some platforms and/or
optimization levels.  "false || (x)" is more verbose than "!!(x)", but
ultimately more efficient.

For example, given this code:

void InitIfNecessary() {
  if (ABSL_PREDICT_TRUE(NeedsInit())) {
    SlowInitIfNecessary();
  }
}

Clang with default optimization level will produce:

Before this CL              After this CL
InitIfNecessary:            InitIfNecessary:
  push rbp                    push rbp
  mov  rbp, rsp               mov  rbp, rsp
  call NeedsInit              call NeedsInit
  xor  al, -1
  xor  al, -1
  test al, 1                  test al, 1
  jne  .LBB2_1                jne  .LBB3_1
  jmp  .LBB2_2                jmp  .LBB3_2
.LBB2_1:                    .LBB3_1:
  call SlowInitIfNecessary    call SlowInitIfNecessary
.LBB2_2:                    .LBB3_2:
  pop  rbp                    pop  rbp
  ret                         ret
PiperOrigin-RevId: 276401386

--
0a3c4dfd8342bf2b1b11a87f1c662c883f73cab7 by Abseil Team <absl-team@google.com>:

Fix comment nit: sem_open => sem_init.

The code calls sem_init, not sem_open, to initialize an unnamed semaphore.
(sem_open creates or opens a named semaphore.)

PiperOrigin-RevId: 276344072

--
b36a664e9459057509a90e83d3482e1d3a4c44c7 by Abseil Team <absl-team@google.com>:

Fix typo in flat_hash_map.h: exchaged -> exchanged

PiperOrigin-RevId: 276295792

--
7bbd8d18276eb110c8335743e35fceb662ddf3d6 by Samuel Benzaquen <sbenza@google.com>:

Add assertions to verify use of iterators.

PiperOrigin-RevId: 276283300

--
677398a8ffcb1f59182cffe57a4fe7ff147a0404 by Laramie Leavitt <lar@google.com>:

Migrate distribution_impl.h/cc to generate_real.h/cc.

Combine the methods RandU64To<Float,Double> into a single method:
GenerateRealFromBits().

Remove rejection sampling from absl::uniform_real_distribution.

PiperOrigin-RevId: 276158675

--
c60c9d11d24b0c546329d998e78e15a84b3153f5 by Abseil Team <absl-team@google.com>:

Internal change

PiperOrigin-RevId: 276126962

--
4c840cab6a8d86efa29b397cafaf7520eece68cc by Andy Soffer <asoffer@google.com>:

Update CMakeLists.txt to address https://github.com/abseil/abseil-cpp/issues/365.
This does not cover every platform, but it does at least address the
first-order issue of assuming gcc implies x86.

PiperOrigin-RevId: 276116253

--
98da366e6b5d51afe5d7ac6722126aca23d85ee6 by Abseil Team <absl-team@google.com>:

Internal change

PiperOrigin-RevId: 276097452
GitOrigin-RevId: e54b9c7bbb0c58475676c268e2e19c69f4bce48a
Change-Id: I02d84454bb71ab21ad3d39650acf6cc6e36f58d7

1 files changed, 0 insertions, 194 deletions

diff --git a/absl/random/internal/distribution_impl.h b/absl/random/internal/distribution_impl.h
deleted file mode 100644
index 49b3e1a6e33d..000000000000
--- a/absl/random/internal/distribution_impl.h
+++ /dev/null
@@ -1,194 +0,0 @@
-// Copyright 2017 The Abseil Authors.
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//      https://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-#ifndef ABSL_RANDOM_INTERNAL_DISTRIBUTION_IMPL_H_
-#define ABSL_RANDOM_INTERNAL_DISTRIBUTION_IMPL_H_
-
-// This file contains some implementation details which are used by one or more
-// of the absl random number distributions.
-
-#include <cfloat>
-#include <cstddef>
-#include <cstdint>
-#include <cstring>
-#include <limits>
-#include <type_traits>
-
-#if (defined(_WIN32) || defined(_WIN64)) && defined(_M_IA64)
-#include <intrin.h>  // NOLINT(build/include_order)
-#pragma intrinsic(_umul128)
-#define ABSL_INTERNAL_USE_UMUL128 1
-#endif
-
-#include "absl/base/config.h"
-#include "absl/base/internal/bits.h"
-#include "absl/numeric/int128.h"
-#include "absl/random/internal/fastmath.h"
-#include "absl/random/internal/traits.h"
-
-namespace absl {
-namespace random_internal {
-
-// Creates a double from `bits`, with the template fields controlling the
-// output.
-//
-// RandU64To is both more efficient and generates more unique values in the
-// result interval than known implementations of std::generate_canonical().
-//
-// The `Signed` parameter controls whether positive, negative, or both are
-// returned (thus affecting the output interval).
-//   When Signed == SignedValueT, range is U(-1, 1)
-//   When Signed == NegativeValueT, range is U(-1, 0)
-//   When Signed == PositiveValueT, range is U(0, 1)
-//
-// When the `IncludeZero` parameter is true, the function may return 0 for some
-// inputs, otherwise it never returns 0.
-//
-// The `ExponentBias` parameter determines the scale of the output range by
-// adjusting the exponent.
-//
-// When a value in U(0,1) is required, use:
-//   RandU64ToDouble<PositiveValueT, true, 0>();
-//
-// When a value in U(-1,1) is required, use:
-//   RandU64ToDouble<SignedValueT, false, 0>() => U(-1, 1)
-// This generates more distinct values than the mathematically equivalent
-// expression `U(0, 1) * 2.0 - 1.0`, and is preferable.
-//
-// Scaling the result by powers of 2 (and avoiding a multiply) is also possible:
-//   RandU64ToDouble<PositiveValueT, false, 1>();  => U(0, 2)
-//   RandU64ToDouble<PositiveValueT, false, -1>();  => U(0, 0.5)
-//
-
-// Tristate types controlling the output.
-struct PositiveValueT {};
-struct NegativeValueT {};
-struct SignedValueT {};
-
-// RandU64ToDouble is the double-result variant of RandU64To, described above.
-template <typename Signed, bool IncludeZero, int ExponentBias = 0>
-inline double RandU64ToDouble(uint64_t bits) {
-  static_assert(std::is_same<Signed, PositiveValueT>::value ||
-                    std::is_same<Signed, NegativeValueT>::value ||
-                    std::is_same<Signed, SignedValueT>::value,
-                "");
-
-  // Maybe use the left-most bit for a sign bit.
-  uint64_t sign = std::is_same<Signed, NegativeValueT>::value
-                      ? 0x8000000000000000ull
-                      : 0;  // Sign bits.
-
-  if (std::is_same<Signed, SignedValueT>::value) {
-    sign = bits & 0x8000000000000000ull;
-    bits = bits & 0x7FFFFFFFFFFFFFFFull;
-  }
-  if (IncludeZero) {
-    if (bits == 0u) return 0;
-  }
-
-  // Number of leading zeros is mapped to the exponent: 2^-clz
-  int clz = base_internal::CountLeadingZeros64(bits);
-  // Shift number left to erase leading zeros.
-  bits <<= IncludeZero ? clz : (clz & 63);
-
-  // Shift number right to remove bits that overflow double mantissa.  The
-  // direction of the shift depends on `clz`.
-  bits >>= (64 - DBL_MANT_DIG);
-
-  // Compute IEEE 754 double exponent.
-  // In the Signed case, bits is a 63-bit number with a 0 msb.  Adjust the
-  // exponent to account for that.
-  const uint64_t exp =
-      (std::is_same<Signed, SignedValueT>::value ? 1023U : 1022U) +
-      static_cast<uint64_t>(ExponentBias - clz);
-  constexpr int kExp = DBL_MANT_DIG - 1;
-  // Construct IEEE 754 double from exponent and mantissa.
-  const uint64_t val = sign | (exp << kExp) | (bits & ((1ULL << kExp) - 1U));
-
-  double res;
-  static_assert(sizeof(res) == sizeof(val), "double is not 64 bit");
-  // Memcpy value from "val" to "res" to avoid aliasing problems.  Assumes that
-  // endian-ness is same for double and uint64_t.
-  std::memcpy(&res, &val, sizeof(res));
-
-  return res;
-}
-
-// RandU64ToFloat is the float-result variant of RandU64To, described above.
-template <typename Signed, bool IncludeZero, int ExponentBias = 0>
-inline float RandU64ToFloat(uint64_t bits) {
-  static_assert(std::is_same<Signed, PositiveValueT>::value ||
-                    std::is_same<Signed, NegativeValueT>::value ||
-                    std::is_same<Signed, SignedValueT>::value,
-                "");
-
-  // Maybe use the left-most bit for a sign bit.
-  uint64_t sign = std::is_same<Signed, NegativeValueT>::value
-                      ? 0x80000000ul
-                      : 0;  // Sign bits.
-
-  if (std::is_same<Signed, SignedValueT>::value) {
-    uint64_t a = bits & 0x8000000000000000ull;
-    sign = static_cast<uint32_t>(a >> 32);
-    bits = bits & 0x7FFFFFFFFFFFFFFFull;
-  }
-  if (IncludeZero) {
-    if (bits == 0u) return 0;
-  }
-
-  // Number of leading zeros is mapped to the exponent: 2^-clz
-  int clz = base_internal::CountLeadingZeros64(bits);
-  // Shift number left to erase leading zeros.
-  bits <<= IncludeZero ? clz : (clz & 63);
-  // Shift number right to remove bits that overflow double mantissa.  The
-  // direction of the shift depends on `clz`.
-  bits >>= (64 - FLT_MANT_DIG);
-
-  // Construct IEEE 754 float exponent.
-  // In the Signed case, bits is a 63-bit number with a 0 msb.  Adjust the
-  // exponent to account for that.
-  const uint32_t exp =
-      (std::is_same<Signed, SignedValueT>::value ? 127U : 126U) +
-      static_cast<uint32_t>(ExponentBias - clz);
-  constexpr int kExp = FLT_MANT_DIG - 1;
-  const uint32_t val = sign | (exp << kExp) | (bits & ((1U << kExp) - 1U));
-
-  float res;
-  static_assert(sizeof(res) == sizeof(val), "float is not 32 bit");
-  // Assumes that endian-ness is same for float and uint32_t.
-  std::memcpy(&res, &val, sizeof(res));
-
-  return res;
-}
-
-template <typename Result>
-struct RandU64ToReal {
-  template <typename Signed, bool IncludeZero, int ExponentBias = 0>
-  static inline Result Value(uint64_t bits) {
-    return RandU64ToDouble<Signed, IncludeZero, ExponentBias>(bits);
-  }
-};
-
-template <>
-struct RandU64ToReal<float> {
-  template <typename Signed, bool IncludeZero, int ExponentBias = 0>
-  static inline float Value(uint64_t bits) {
-    return RandU64ToFloat<Signed, IncludeZero, ExponentBias>(bits);
-  }
-};
-
-}  // namespace random_internal
-}  // namespace absl
-
-#endif  // ABSL_RANDOM_INTERNAL_DISTRIBUTION_IMPL_H_