Export of internal Abseil changes.

--
7a6ff16a85beb730c172d5d25cf1b5e1be885c56 by Laramie Leavitt <lar@google.com>:

Internal change.

PiperOrigin-RevId: 254454546

--
ff8f9bafaefc26d451f576ea4a06d150aed63f6f by Andy Soffer <asoffer@google.com>:

Internal changes

PiperOrigin-RevId: 254451562

--
deefc5b651b479ce36f0b4ef203e119c0c8936f2 by CJ Johnson <johnsoncj@google.com>:

Account for subtracting unsigned values from the size of InlinedVector

PiperOrigin-RevId: 254450625

--
3c677316a27bcadc17e41957c809ca472d5fef14 by Andy Soffer <asoffer@google.com>:

Add C++17's std::make_from_tuple to absl/utility/utility.h

PiperOrigin-RevId: 254411573

--
4ee3536a918830eeec402a28fc31a62c7c90b940 by CJ Johnson <johnsoncj@google.com>:

Adds benchmark for the rest of the InlinedVector public API

PiperOrigin-RevId: 254408378

--
e5a21a00700ee83498ff1efbf649169756463ee4 by CJ Johnson <johnsoncj@google.com>:

Updates the definition of InlinedVector::shrink_to_fit() to be exception safe and adds exception safety tests for it.

PiperOrigin-RevId: 254401387

--
2ea82e72b86d82d78b4e4712a63a55981b53c64b by Laramie Leavitt <lar@google.com>:

Use absl::InsecureBitGen in place of std::mt19937
in tests absl/random/...distribution_test.cc

PiperOrigin-RevId: 254289444

--
fa099e02c413a7ffda732415e8105cad26a90337 by Andy Soffer <asoffer@google.com>:

Internal changes

PiperOrigin-RevId: 254286334

--
ce34b7f36933b30cfa35b9c9a5697a792b5666e4 by Andy Soffer <asoffer@google.com>:

Internal changes

PiperOrigin-RevId: 254273059

--
6f9c473da7c2090c2e85a37c5f00622e8a912a89 by Jorg Brown <jorg@google.com>:

Change absl::container_internal::CompressedTuple to instantiate its
internal Storage class with the name of the type it's holding, rather
than the name of the Tuple.  This is not an externally-visible change,
other than less compiler memory is used and less debug information is
generated.

PiperOrigin-RevId: 254269285

--
8bd3c186bf2fc0c55d8a2dd6f28a5327502c9fba by Andy Soffer <asoffer@google.com>:

Adding short-hand IntervalClosed for IntervalClosedClosed and IntervalOpen for
IntervalOpenOpen.

PiperOrigin-RevId: 254252419

--
ea957f99b6a04fccd42aa05605605f3b44b1ecfd by Abseil Team <absl-team@google.com>:

Do not directly use __SIZEOF_INT128__.

In order to avoid linker errors when building with clang-cl (__fixunsdfti, __udivti3 and __fixunssfti are undefined), this CL uses ABSL_HAVE_INTRINSIC_INT128 which is not defined for clang-cl.

PiperOrigin-RevId: 254250739

--
89ab385cd26b34d64130bce856253aaba96d2345 by Andy Soffer <asoffer@google.com>:

Internal changes

PiperOrigin-RevId: 254242321

--
cffc793d93eca6d6bdf7de733847b6ab4a255ae9 by CJ Johnson <johnsoncj@google.com>:

Adds benchmark for InlinedVector::reserve(size_type)

PiperOrigin-RevId: 254199226

--
c90c7a9fa3c8f0c9d5114036979548b055ea2f2a by Gennadiy Rozental <rogeeff@google.com>:

Import of CCTZ from GitHub.

PiperOrigin-RevId: 254072387

--
c4c388beae016c9570ab54ffa1d52660e4a85b7b by Laramie Leavitt <lar@google.com>:

Internal cleanup.

PiperOrigin-RevId: 254062381

--
d3c992e221cc74e5372d0c8fa410170b6a43c062 by Tom Manshreck <shreck@google.com>:

Update distributions.h to Abseil standards

PiperOrigin-RevId: 254054946

--
d15ad0035c34ef11b14fadc5a4a2d3ec415f5518 by CJ Johnson <johnsoncj@google.com>:

Removes functions with only one caller from the implementation details of InlinedVector by manually inlining the definitions

PiperOrigin-RevId: 254005427

--
2f37e807efc3a8ef1f4b539bdd379917d4151520 by Andy Soffer <asoffer@google.com>:

Initial release of Abseil Random

PiperOrigin-RevId: 253999861

--
24ed1694b6430791d781ed533a8f8ccf6cac5856 by CJ Johnson <johnsoncj@google.com>:

Updates the definition of InlinedVector::assign(...)/InlinedVector::operator=(...) to new, exception-safe implementations with exception safety tests to boot

PiperOrigin-RevId: 253993691

--
5613d95f5a7e34a535cfaeadce801441e990843e by CJ Johnson <johnsoncj@google.com>:

Adds benchmarks for InlinedVector::shrink_to_fit()

PiperOrigin-RevId: 253989647

--
2a96ddfdac40bbb8cb6a7f1aeab90917067c6e63 by Abseil Team <absl-team@google.com>:

Initial release of Abseil Random

PiperOrigin-RevId: 253927497

--
bf1aff8fc9ffa921ad74643e9525ecf25b0d8dc1 by Andy Soffer <asoffer@google.com>:

Initial release of Abseil Random

PiperOrigin-RevId: 253920512

--
bfc03f4a3dcda3cf3a4b84bdb84cda24e3394f41 by Laramie Leavitt <lar@google.com>:

Internal change.

PiperOrigin-RevId: 253886486

--
05036cfcc078ca7c5f581a00dfb0daed568cbb69 by Eric Fiselier <ericwf@google.com>:

Don't include `winsock2.h` because it drags in `windows.h` and friends,
and they define awful macros like OPAQUE, ERROR, and more. This has the
potential to break abseil users.

Instead we only forward declare `timeval` and require Windows users
include `winsock2.h` themselves. This is both inconsistent and poor QoI, but so
including 'windows.h' is bad too.

PiperOrigin-RevId: 253852615
GitOrigin-RevId: 7a6ff16a85beb730c172d5d25cf1b5e1be885c56
Change-Id: Icd6aff87da26f29ec8915da856f051129987cef6

1 files changed, 198 insertions, 0 deletions

diff --git a/absl/random/bernoulli_distribution.h b/absl/random/bernoulli_distribution.h
new file mode 100644
index 000000000000..326fcb6e012d
--- /dev/null
+++ b/absl/random/bernoulli_distribution.h
@@ -0,0 +1,198 @@
+// 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_BERNOULLI_DISTRIBUTION_H_
+#define ABSL_RANDOM_BERNOULLI_DISTRIBUTION_H_
+
+#include <cstdint>
+#include <istream>
+#include <limits>
+
+#include "absl/base/optimization.h"
+#include "absl/random/internal/fast_uniform_bits.h"
+#include "absl/random/internal/iostream_state_saver.h"
+
+namespace absl {
+
+// absl::bernoulli_distribution is a drop in replacement for
+// std::bernoulli_distribution. It guarantees that (given a perfect
+// UniformRandomBitGenerator) the acceptance probability is *exactly* equal to
+// the given double.
+//
+// The implementation assumes that double is IEEE754
+class bernoulli_distribution {
+ public:
+  using result_type = bool;
+
+  class param_type {
+   public:
+    using distribution_type = bernoulli_distribution;
+
+    explicit param_type(double p = 0.5) : prob_(p) {
+      assert(p >= 0.0 && p <= 1.0);
+    }
+
+    double p() const { return prob_; }
+
+    friend bool operator==(const param_type& p1, const param_type& p2) {
+      return p1.p() == p2.p();
+    }
+    friend bool operator!=(const param_type& p1, const param_type& p2) {
+      return p1.p() != p2.p();
+    }
+
+   private:
+    double prob_;
+  };
+
+  bernoulli_distribution() : bernoulli_distribution(0.5) {}
+
+  explicit bernoulli_distribution(double p) : param_(p) {}
+
+  explicit bernoulli_distribution(param_type p) : param_(p) {}
+
+  // no-op
+  void reset() {}
+
+  template <typename URBG>
+  bool operator()(URBG& g) {  // NOLINT(runtime/references)
+    return Generate(param_.p(), g);
+  }
+
+  template <typename URBG>
+  bool operator()(URBG& g,  // NOLINT(runtime/references)
+                  const param_type& param) {
+    return Generate(param.p(), g);
+  }
+
+  param_type param() const { return param_; }
+  void param(const param_type& param) { param_ = param; }
+
+  double p() const { return param_.p(); }
+
+  result_type(min)() const { return false; }
+  result_type(max)() const { return true; }
+
+  friend bool operator==(const bernoulli_distribution& d1,
+                         const bernoulli_distribution& d2) {
+    return d1.param_ == d2.param_;
+  }
+
+  friend bool operator!=(const bernoulli_distribution& d1,
+                         const bernoulli_distribution& d2) {
+    return d1.param_ != d2.param_;
+  }
+
+ private:
+  static constexpr uint64_t kP32 = static_cast<uint64_t>(1) << 32;
+
+  template <typename URBG>
+  static bool Generate(double p, URBG& g);  // NOLINT(runtime/references)
+
+  param_type param_;
+};
+
+template <typename CharT, typename Traits>
+std::basic_ostream<CharT, Traits>& operator<<(
+    std::basic_ostream<CharT, Traits>& os,  // NOLINT(runtime/references)
+    const bernoulli_distribution& x) {
+  auto saver = random_internal::make_ostream_state_saver(os);
+  os.precision(random_internal::stream_precision_helper<double>::kPrecision);
+  os << x.p();
+  return os;
+}
+
+template <typename CharT, typename Traits>
+std::basic_istream<CharT, Traits>& operator>>(
+    std::basic_istream<CharT, Traits>& is,  // NOLINT(runtime/references)
+    bernoulli_distribution& x) {            // NOLINT(runtime/references)
+  auto saver = random_internal::make_istream_state_saver(is);
+  auto p = random_internal::read_floating_point<double>(is);
+  if (!is.fail()) {
+    x.param(bernoulli_distribution::param_type(p));
+  }
+  return is;
+}
+
+template <typename URBG>
+bool bernoulli_distribution::Generate(double p,
+                                      URBG& g) {  // NOLINT(runtime/references)
+  random_internal::FastUniformBits<uint32_t> fast_u32;
+
+  while (true) {
+    // There are two aspects of the definition of `c` below that are worth
+    // commenting on.  First, because `p` is in the range [0, 1], `c` is in the
+    // range [0, 2^32] which does not fit in a uint32_t and therefore requires
+    // 64 bits.
+    //
+    // Second, `c` is constructed by first casting explicitly to a signed
+    // integer and then converting implicitly to an unsigned integer of the same
+    // size.  This is done because the hardware conversion instructions produce
+    // signed integers from double; if taken as a uint64_t the conversion would
+    // be wrong for doubles greater than 2^63 (not relevant in this use-case).
+    // If converted directly to an unsigned integer, the compiler would end up
+    // emitting code to handle such large values that are not relevant due to
+    // the known bounds on `c`.  To avoid these extra instructions this
+    // implementation converts first to the signed type and then use the
+    // implicit conversion to unsigned (which is a no-op).
+    const uint64_t c = static_cast<int64_t>(p * kP32);
+    const uint32_t v = fast_u32(g);
+    // FAST PATH: this path fails with probability 1/2^32.  Note that simply
+    // returning v <= c would approximate P very well (up to an absolute error
+    // of 1/2^32); the slow path (taken in that range of possible error, in the
+    // case of equality) eliminates the remaining error.
+    if (ABSL_PREDICT_TRUE(v != c)) return v < c;
+
+    // It is guaranteed that `q` is strictly less than 1, because if `q` were
+    // greater than or equal to 1, the same would be true for `p`. Certainly `p`
+    // cannot be greater than 1, and if `p == 1`, then the fast path would
+    // necessary have been taken already.
+    const double q = static_cast<double>(c) / kP32;
+
+    // The probability of acceptance on the fast path is `q` and so the
+    // probability of acceptance here should be `p - q`.
+    //
+    // Note that `q` is obtained from `p` via some shifts and conversions, the
+    // upshot of which is that `q` is simply `p` with some of the
+    // least-significant bits of its mantissa set to zero. This means that the
+    // difference `p - q` will not have any rounding errors. To see why, pretend
+    // that double has 10 bits of resolution and q is obtained from `p` in such
+    // a way that the 4 least-significant bits of its mantissa are set to zero.
+    // For example:
+    //   p   = 1.1100111011 * 2^-1
+    //   q   = 1.1100110000 * 2^-1
+    // p - q = 1.011        * 2^-8
+    // The difference `p - q` has exactly the nonzero mantissa bits that were
+    // "lost" in `q` producing a number which is certainly representable in a
+    // double.
+    const double left = p - q;
+
+    // By construction, the probability of being on this slow path is 1/2^32, so
+    // P(accept in slow path) = P(accept| in slow path) * P(slow path),
+    // which means the probability of acceptance here is `1 / (left * kP32)`:
+    const double here = left * kP32;
+
+    // The simplest way to compute the result of this trial is to repeat the
+    // whole algorithm with the new probability. This terminates because even
+    // given  arbitrarily unfriendly "random" bits, each iteration either
+    // multiplies a tiny probability by 2^32 (if c == 0) or strips off some
+    // number of nonzero mantissa bits. That process is bounded.
+    if (here == 0) return false;
+    p = here;
+  }
+}
+
+}  // namespace absl
+
+#endif  // ABSL_RANDOM_BERNOULLI_DISTRIBUTION_H_