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, 252 insertions, 0 deletions

diff --git a/absl/random/internal/pool_urbg.cc b/absl/random/internal/pool_urbg.cc
new file mode 100644
index 000000000000..b24eeeffeaa9
--- /dev/null
+++ b/absl/random/internal/pool_urbg.cc
@@ -0,0 +1,252 @@
+// 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.
+
+#include "absl/random/internal/pool_urbg.h"
+
+#include <algorithm>
+#include <atomic>
+#include <cstdint>
+#include <cstring>
+#include <iterator>
+
+#include "absl/base/attributes.h"
+#include "absl/base/call_once.h"
+#include "absl/base/config.h"
+#include "absl/base/internal/endian.h"
+#include "absl/base/internal/raw_logging.h"
+#include "absl/base/internal/spinlock.h"
+#include "absl/base/internal/sysinfo.h"
+#include "absl/base/internal/unaligned_access.h"
+#include "absl/base/optimization.h"
+#include "absl/random/internal/randen.h"
+#include "absl/random/internal/seed_material.h"
+#include "absl/random/seed_gen_exception.h"
+
+using absl::base_internal::SpinLock;
+using absl::base_internal::SpinLockHolder;
+
+namespace absl {
+namespace random_internal {
+namespace {
+
+// RandenPoolEntry is a thread-safe pseudorandom bit generator, implementing a
+// single generator within a RandenPool<T>. It is an internal implementation
+// detail, and does not aim to conform to [rand.req.urng].
+//
+// NOTE: There are alignment issues when used on ARM, for instance.
+// See the allocation code in PoolAlignedAlloc().
+class RandenPoolEntry {
+ public:
+  static constexpr size_t kState = RandenTraits::kStateBytes / sizeof(uint32_t);
+  static constexpr size_t kCapacity =
+      RandenTraits::kCapacityBytes / sizeof(uint32_t);
+
+  void Init(absl::Span<const uint32_t> data) {
+    SpinLockHolder l(&mu_);  // Always uncontested.
+    std::copy(data.begin(), data.end(), std::begin(state_));
+    next_ = kState;
+  }
+
+  // Copy bytes into out.
+  void Fill(uint8_t* out, size_t bytes) LOCKS_EXCLUDED(mu_);
+
+  // Returns random bits from the buffer in units of T.
+  template <typename T>
+  inline T Generate() LOCKS_EXCLUDED(mu_);
+
+  inline void MaybeRefill() EXCLUSIVE_LOCKS_REQUIRED(mu_) {
+    if (next_ >= kState) {
+      next_ = kCapacity;
+      impl_.Generate(state_);
+    }
+  }
+
+ private:
+  // Randen URBG state.
+  uint32_t state_[kState] GUARDED_BY(mu_);  // First to satisfy alignment.
+  SpinLock mu_;
+  const Randen impl_;
+  size_t next_ GUARDED_BY(mu_);
+};
+
+template <>
+inline uint8_t RandenPoolEntry::Generate<uint8_t>() {
+  SpinLockHolder l(&mu_);
+  MaybeRefill();
+  return static_cast<uint8_t>(state_[next_++]);
+}
+
+template <>
+inline uint16_t RandenPoolEntry::Generate<uint16_t>() {
+  SpinLockHolder l(&mu_);
+  MaybeRefill();
+  return static_cast<uint16_t>(state_[next_++]);
+}
+
+template <>
+inline uint32_t RandenPoolEntry::Generate<uint32_t>() {
+  SpinLockHolder l(&mu_);
+  MaybeRefill();
+  return state_[next_++];
+}
+
+template <>
+inline uint64_t RandenPoolEntry::Generate<uint64_t>() {
+  SpinLockHolder l(&mu_);
+  if (next_ >= kState - 1) {
+    next_ = kCapacity;
+    impl_.Generate(state_);
+  }
+  auto p = state_ + next_;
+  next_ += 2;
+
+  uint64_t result;
+  std::memcpy(&result, p, sizeof(result));
+  return result;
+}
+
+void RandenPoolEntry::Fill(uint8_t* out, size_t bytes) {
+  SpinLockHolder l(&mu_);
+  while (bytes > 0) {
+    MaybeRefill();
+    size_t remaining = (kState - next_) * sizeof(state_[0]);
+    size_t to_copy = std::min(bytes, remaining);
+    std::memcpy(out, &state_[next_], to_copy);
+    out += to_copy;
+    bytes -= to_copy;
+    next_ += (to_copy + sizeof(state_[0]) - 1) / sizeof(state_[0]);
+  }
+}
+
+// Number of pooled urbg entries.
+static constexpr int kPoolSize = 8;
+
+// Shared pool entries.
+static absl::once_flag pool_once;
+ABSL_CACHELINE_ALIGNED static RandenPoolEntry* shared_pools[kPoolSize];
+
+// Returns an id in the range [0 ... kPoolSize), which indexes into the
+// pool of random engines.
+//
+// Each thread to access the pool is assigned a sequential ID (without reuse)
+// from the pool-id space; the id is cached in a thread_local variable.
+// This id is assigned based on the arrival-order of the thread to the
+// GetPoolID call; this has no binary, CL, or runtime stability because
+// on subsequent runs the order within the same program may be significantly
+// different. However, as other thread IDs are not assigned sequentially,
+// this is not expected to matter.
+int GetPoolID() {
+  static_assert(kPoolSize >= 1,
+                "At least one urbg instance is required for PoolURBG");
+
+  ABSL_CONST_INIT static std::atomic<int64_t> sequence{0};
+
+#ifdef ABSL_HAVE_THREAD_LOCAL
+  static thread_local int my_pool_id = -1;
+  if (ABSL_PREDICT_FALSE(my_pool_id < 0)) {
+    my_pool_id = (sequence++ % kPoolSize);
+  }
+  return my_pool_id;
+#else
+  static pthread_key_t tid_key = [] {
+    pthread_key_t tmp_key;
+    int err = pthread_key_create(&tmp_key, nullptr);
+    if (err) {
+      ABSL_RAW_LOG(FATAL, "pthread_key_create failed with %d", err);
+    }
+    return tmp_key;
+  }();
+
+  // Store the value in the pthread_{get/set}specific. However an uninitialized
+  // value is 0, so add +1 to distinguish from the null value.
+  intptr_t my_pool_id =
+      reinterpret_cast<intptr_t>(pthread_getspecific(tid_key));
+  if (ABSL_PREDICT_FALSE(my_pool_id == 0)) {
+    // No allocated ID, allocate the next value, cache it, and return.
+    my_pool_id = (sequence++ % kPoolSize) + 1;
+    int err = pthread_setspecific(tid_key, reinterpret_cast<void*>(my_pool_id));
+    if (err) {
+      ABSL_RAW_LOG(FATAL, "pthread_setspecific failed with %d", err);
+    }
+  }
+  return my_pool_id - 1;
+#endif
+}
+
+// Allocate a RandenPoolEntry with at least 32-byte alignment, which is required
+// by ARM platform code.
+RandenPoolEntry* PoolAlignedAlloc() {
+  constexpr size_t kAlignment =
+      ABSL_CACHELINE_SIZE > 32 ? ABSL_CACHELINE_SIZE : 32;
+
+  // Not all the platforms that we build for have std::aligned_alloc, however
+  // since we never free these objects, we can over allocate and munge the
+  // pointers to the correct alignment.
+  void* memory = std::malloc(sizeof(RandenPoolEntry) + kAlignment);
+  auto x = reinterpret_cast<intptr_t>(memory);
+  auto y = x % kAlignment;
+  void* aligned =
+      (y == 0) ? memory : reinterpret_cast<void*>(x + kAlignment - y);
+  return new (aligned) RandenPoolEntry();
+}
+
+// Allocate and initialize kPoolSize objects of type RandenPoolEntry.
+//
+// The initialization strategy is to initialize one object directly from
+// OS entropy, then to use that object to seed all of the individual
+// pool instances.
+void InitPoolURBG() {
+  static constexpr size_t kSeedSize =
+      RandenTraits::kStateBytes / sizeof(uint32_t);
+  // Read the seed data from OS entropy once.
+  uint32_t seed_material[kPoolSize * kSeedSize];
+  if (!random_internal::ReadSeedMaterialFromOSEntropy(
+          absl::MakeSpan(seed_material))) {
+    random_internal::ThrowSeedGenException();
+  }
+  for (int i = 0; i < kPoolSize; i++) {
+    shared_pools[i] = PoolAlignedAlloc();
+    shared_pools[i]->Init(
+        absl::MakeSpan(&seed_material[i * kSeedSize], kSeedSize));
+  }
+}
+
+// Returns the pool entry for the current thread.
+RandenPoolEntry* GetPoolForCurrentThread() {
+  absl::call_once(pool_once, InitPoolURBG);
+  return shared_pools[GetPoolID()];
+}
+
+}  // namespace
+
+template <typename T>
+typename RandenPool<T>::result_type RandenPool<T>::Generate() {
+  auto* pool = GetPoolForCurrentThread();
+  return pool->Generate<T>();
+}
+
+template <typename T>
+void RandenPool<T>::Fill(absl::Span<result_type> data) {
+  auto* pool = GetPoolForCurrentThread();
+  pool->Fill(reinterpret_cast<uint8_t*>(data.data()),
+             data.size() * sizeof(result_type));
+}
+
+template class RandenPool<uint8_t>;
+template class RandenPool<uint16_t>;
+template class RandenPool<uint32_t>;
+template class RandenPool<uint64_t>;
+
+}  // namespace random_internal
+}  // namespace absl