Export of internal Abseil changes.

--
178e7a9a76fc8fcd6df6335b59139cbe644a16b9 by Jon Cohen <cohenjon@google.com>:

Import of CCTZ from GitHub.

PiperOrigin-RevId: 220523164

--
59ef14fe7034a3148f1e9cef1f128b8ca264b444 by Jon Cohen <cohenjon@google.com>:

Don't assume how much std::vector's constructors allocate in InlinedVector's test for scoped_allocator_adaptor support.

PiperOrigin-RevId: 220464683

--
6f8351be43a44a8f10bf20612b2cc744a4a911c7 by Jon Cohen <cohenjon@google.com>:

Add VS Code and some Bazel output files to absl/.gitignore

PiperOrigin-RevId: 220464362

--
43fac22f8af6b6ed55309a784a9d25d837393d0e by Abseil Team <absl-team@google.com>:

absl: fix SpinLock::EncodeWaitCycles

If a thread has ever observed or set kSpinLockSleeper, it must
never leave 0 in kWaitTimeMask because at this point it is
expected to wake subsequent threads. Current calculations in
EncodeWaitCycles can result in 0 in kWaitTimeMask and lead to
missed wakeups. This is mostly theoretical today, because
the futex call needs to finish within 128 cycles (futex can
return immediately without waiting, but 128 cycles still
look too low for this). But this can well fire in future
if we bump granularity and/or threshold for recording contention.

Use kSpinLockSleeper instead of 0.

PiperOrigin-RevId: 220463123

--
def9b7e3d45c99d68cc52a4429256116d7f421f2 by Abseil Team <absl-team@google.com>:

absl: optimize SpinLock::SlowLock

Currently we record contention even after the first initial spin.
This leads to several performance issues:
1. If we succeed in acquiring the lock after the initial spin,
overall wait time can be within tens/hundreds of nanoseconds.
Recording such low wait time looks completely unnecessary and excessive.
From some point of view this is not even a wait, because we did not sleep.
And, for example, Mutex does not record contention in this case.
In majority of cases the lock should be acquired exactly during the initial
spin, yet we still go through full overhead of submitting contention.
2. Whenever a thread submits contention it also calls FUTEX_WAKE
(there is no way to understand if it's necessary or not when wait value
is stored in the lock). So if there are just 2 threads and a brief
contention, the second thread will still call FUTEX_WAKE which
is completely unnecessary overhead.

Don't record contention after the initial spin wait.

FWIW this also removes 2 CycleClock::Now calls and EncodeWaitCycles
from the common hot path.

PiperOrigin-RevId: 220379972

--
75b0c0cb214de904ea622f81ec3f4eabdc8695b0 by Derek Mauro <dmauro@google.com>:

Supress MSVC warnings in raw_hash_set's use of TrailingZeros and LeadingZeros.
https://github.com/abseil/abseil-cpp/issues/208

PiperOrigin-RevId: 220372204
GitOrigin-RevId: 178e7a9a76fc8fcd6df6335b59139cbe644a16b9
Change-Id: I3a66af4e050810a3274e45d4e055b2aa19ffba1b
This commit is contained in:
Abseil Team 2018-11-07 13:52:16 -08:00 committed by Jon Cohen
parent 7990fd459e
commit 070f6e47b3
13 changed files with 70 additions and 40 deletions

4
.gitignore vendored
View file

@ -1,8 +1,12 @@
# Ignore all bazel-* symlinks.
/bazel-*
# Ignore Bazel verbose explanations
--verbose_explanations
# Ignore CMake usual build directory
build
# Ignore Vim files
*.swp
# Ignore QtCreator Project file
CMakeLists.txt.user
# Ignore VS Code files
.vscode/*

View file

@ -95,13 +95,9 @@ void SpinLock::InitLinkerInitializedAndCooperative() {
}
// Monitor the lock to see if its value changes within some time period
// (adaptive_spin_count loop iterations). A timestamp indicating
// when the thread initially started waiting for the lock is passed in via
// the initial_wait_timestamp value. The total wait time in cycles for the
// lock is returned in the wait_cycles parameter. The last value read
// from the lock is returned from the method.
uint32_t SpinLock::SpinLoop(int64_t initial_wait_timestamp,
uint32_t *wait_cycles) {
// (adaptive_spin_count loop iterations). The last value read from the lock
// is returned from the method.
uint32_t SpinLock::SpinLoop() {
// We are already in the slow path of SpinLock, initialize the
// adaptive_spin_count here.
ABSL_CONST_INIT static absl::once_flag init_adaptive_spin_count;
@ -115,22 +111,21 @@ uint32_t SpinLock::SpinLoop(int64_t initial_wait_timestamp,
do {
lock_value = lockword_.load(std::memory_order_relaxed);
} while ((lock_value & kSpinLockHeld) != 0 && --c > 0);
uint32_t spin_loop_wait_cycles =
EncodeWaitCycles(initial_wait_timestamp, CycleClock::Now());
*wait_cycles = spin_loop_wait_cycles;
return TryLockInternal(lock_value, spin_loop_wait_cycles);
return lock_value;
}
void SpinLock::SlowLock() {
uint32_t lock_value = SpinLoop();
lock_value = TryLockInternal(lock_value, 0);
if ((lock_value & kSpinLockHeld) == 0) {
return;
}
// The lock was not obtained initially, so this thread needs to wait for
// it. Record the current timestamp in the local variable wait_start_time
// so the total wait time can be stored in the lockword once this thread
// obtains the lock.
int64_t wait_start_time = CycleClock::Now();
uint32_t wait_cycles;
uint32_t lock_value = SpinLoop(wait_start_time, &wait_cycles);
uint32_t wait_cycles = 0;
int lock_wait_call_count = 0;
while ((lock_value & kSpinLockHeld) != 0) {
// If the lock is currently held, but not marked as having a sleeper, mark
@ -170,7 +165,9 @@ void SpinLock::SlowLock() {
ABSL_TSAN_MUTEX_POST_DIVERT(this, 0);
// Spin again after returning from the wait routine to give this thread
// some chance of obtaining the lock.
lock_value = SpinLoop(wait_start_time, &wait_cycles);
lock_value = SpinLoop();
wait_cycles = EncodeWaitCycles(wait_start_time, CycleClock::Now());
lock_value = TryLockInternal(lock_value, wait_cycles);
}
}
@ -206,14 +203,20 @@ uint32_t SpinLock::EncodeWaitCycles(int64_t wait_start_time,
(wait_end_time - wait_start_time) >> PROFILE_TIMESTAMP_SHIFT;
// Return a representation of the time spent waiting that can be stored in
// the lock word's upper bits. bit_cast is required as Atomic32 is signed.
const uint32_t clamped = static_cast<uint32_t>(
// the lock word's upper bits.
uint32_t clamped = static_cast<uint32_t>(
std::min(scaled_wait_time, kMaxWaitTime) << LOCKWORD_RESERVED_SHIFT);
// bump up value if necessary to avoid returning kSpinLockSleeper.
const uint32_t after_spinlock_sleeper =
kSpinLockSleeper + (1 << LOCKWORD_RESERVED_SHIFT);
return clamped == kSpinLockSleeper ? after_spinlock_sleeper : clamped;
if (clamped == 0) {
return kSpinLockSleeper; // Just wake waiters, but don't record contention.
}
// Bump up value if necessary to avoid returning kSpinLockSleeper.
const uint32_t kMinWaitTime =
kSpinLockSleeper + (1 << LOCKWORD_RESERVED_SHIFT);
if (clamped == kSpinLockSleeper) {
return kMinWaitTime;
}
return clamped;
}
uint64_t SpinLock::DecodeWaitCycles(uint32_t lock_value) {

View file

@ -161,7 +161,7 @@ class LOCKABLE SpinLock {
void InitLinkerInitializedAndCooperative();
void SlowLock() ABSL_ATTRIBUTE_COLD;
void SlowUnlock(uint32_t lock_value) ABSL_ATTRIBUTE_COLD;
uint32_t SpinLoop(int64_t initial_wait_timestamp, uint32_t* wait_cycles);
uint32_t SpinLoop();
inline bool TryLockImpl() {
uint32_t lock_value = lockword_.load(std::memory_order_relaxed);

View file

@ -155,7 +155,8 @@ TEST(SpinLock, WaitCyclesEncoding) {
// Test corner cases
int64_t start_time = time_distribution(generator);
EXPECT_EQ(0, SpinLockTest::EncodeWaitCycles(start_time, start_time));
EXPECT_EQ(kSpinLockSleeper,
SpinLockTest::EncodeWaitCycles(start_time, start_time));
EXPECT_EQ(0, SpinLockTest::DecodeWaitCycles(0));
EXPECT_EQ(0, SpinLockTest::DecodeWaitCycles(kLockwordReservedMask));
EXPECT_EQ(kMaxCycles & ~kProfileTimestampMask,

View file

@ -1728,39 +1728,58 @@ TEST(AllocatorSupportTest, ScopedAllocatorWorks) {
std::scoped_allocator_adaptor<CountingAllocator<StdVector>>;
using AllocVec = absl::InlinedVector<StdVector, 4, MyAlloc>;
// MSVC 2017's std::vector allocates different amounts of memory in debug
// versus opt mode.
int64_t test_allocated = 0;
StdVector v(CountingAllocator<int>{&test_allocated});
// The amount of memory allocated by a default constructed vector<int>
auto default_std_vec_allocated = test_allocated;
v.push_back(1);
// The amound of memory allocated by a copy-constructed vector<int> with one
// element.
int64_t one_element_std_vec_copy_allocated = test_allocated;
int64_t allocated = 0;
AllocVec vec(MyAlloc{CountingAllocator<StdVector>{&allocated}});
EXPECT_EQ(allocated, 0);
// This default constructs a vector<int>, but the allocator should pass itself
// into the vector<int>.
// into the vector<int>, so check allocation compared to that.
// The absl::InlinedVector does not allocate any memory.
// The vector<int> does not allocate any memory.
// The vector<int> may allocate any memory.
auto expected = default_std_vec_allocated;
vec.resize(1);
EXPECT_EQ(allocated, 0);
EXPECT_EQ(allocated, expected);
// We make vector<int> allocate memory.
// It must go through the allocator even though we didn't construct the
// vector directly.
// vector directly. This assumes that vec[0] doesn't need to grow its
// allocation.
expected += sizeof(int);
vec[0].push_back(1);
EXPECT_EQ(allocated, sizeof(int) * 1);
EXPECT_EQ(allocated, expected);
// Another allocating vector.
expected += one_element_std_vec_copy_allocated;
vec.push_back(vec[0]);
EXPECT_EQ(allocated, sizeof(int) * 2);
EXPECT_EQ(allocated, expected);
// Overflow the inlined memory.
// The absl::InlinedVector will now allocate.
expected += sizeof(StdVector) * 8 + default_std_vec_allocated * 3;
vec.resize(5);
EXPECT_EQ(allocated, sizeof(int) * 2 + sizeof(StdVector) * 8);
EXPECT_EQ(allocated, expected);
// Adding one more in external mode should also work.
expected += one_element_std_vec_copy_allocated;
vec.push_back(vec[0]);
EXPECT_EQ(allocated, sizeof(int) * 3 + sizeof(StdVector) * 8);
EXPECT_EQ(allocated, expected);
// And extending these should still work.
// And extending these should still work. This assumes that vec[0] does not
// need to grow its allocation.
expected += sizeof(int);
vec[0].push_back(1);
EXPECT_EQ(allocated, sizeof(int) * 4 + sizeof(StdVector) * 8);
EXPECT_EQ(allocated, expected);
vec.clear();
EXPECT_EQ(allocated, 0);

View file

@ -208,14 +208,17 @@ constexpr bool IsNoThrowSwappable() {
template <typename T>
int TrailingZeros(T x) {
return sizeof(T) == 8 ? base_internal::CountTrailingZerosNonZero64(x)
: base_internal::CountTrailingZerosNonZero32(x);
return sizeof(T) == 8 ? base_internal::CountTrailingZerosNonZero64(
static_cast<uint64_t>(x))
: base_internal::CountTrailingZerosNonZero32(
static_cast<uint32_t>(x));
}
template <typename T>
int LeadingZeros(T x) {
return sizeof(T) == 8 ? base_internal::CountLeadingZeros64(x)
: base_internal::CountLeadingZeros32(x);
return sizeof(T) == 8
? base_internal::CountLeadingZeros64(static_cast<uint64_t>(x))
: base_internal::CountLeadingZeros32(static_cast<uint32_t>(x));
}
// An abstraction over a bitmask. It provides an easy way to iterate through the

View file

@ -1 +1 @@
2018f-1-g401c42d
2018g-9-gf0d2759

View file

@ -289,9 +289,9 @@ RS,BA,HR,ME,MK,SI +4450+02030 Europe/Belgrade
RU +5443+02030 Europe/Kaliningrad MSK-01 - Kaliningrad
RU +554521+0373704 Europe/Moscow MSK+00 - Moscow area
RU +4457+03406 Europe/Simferopol MSK+00 - Crimea
RU +4844+04425 Europe/Volgograd MSK+00 - Volgograd
RU +5836+04939 Europe/Kirov MSK+00 - Kirov
RU +4621+04803 Europe/Astrakhan MSK+01 - Astrakhan
RU +4844+04425 Europe/Volgograd MSK+01 - Volgograd
RU +5134+04602 Europe/Saratov MSK+01 - Saratov
RU +5420+04824 Europe/Ulyanovsk MSK+01 - Ulyanovsk
RU +5312+05009 Europe/Samara MSK+01 - Samara, Udmurtia