mdebray/tvl-depot
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Export of internal Abseil changes.

Browse source 
--
cd076f55c1fa600131f6dda392533dfe61679fc0 by Abseil Team <absl-team@google.com>:

Internal change

PiperOrigin-RevId: 224008762

--
e05f62b01286d51044ff86ec6ef565749b9faf82 by Abseil Team <absl-team@google.com>:

Create a pow10() test helper function to compute guaranteed-precise double values of 10^x. Not all standard libraries ship bit-accurate pow() functions, causing tests to fail that rely on expected values generated by it.

PiperOrigin-RevId: 223883762

--
fd88e5e3f7ab80f7f5df9fd1488cd58b4573be69 by Abseil Team <absl-team@google.com>:

Remove some absl:: qualifications to work around inline namespace bugs on MSVC 2015.

PiperOrigin-RevId: 223869642

--
6276cfff969d596edd36a2bbaba65ee045808903 by Abseil Team <absl-team@google.com>:

Update absl/memory/CMakeLists.txt to use new functions
i.e. absl_cc_(library|test)

PiperOrigin-RevId: 223854224

--
359de9afc7a34c975fd3e0cbc52afd96637d97bd by Chris Kennelly <ckennelly@google.com>:

Mark spinlock_benchmark_common as alwayslink = 1.

PiperOrigin-RevId: 223844536

--
450cd8cbe2789a6d54ed1eb87170259bb334f8b9 by Abseil Team <absl-team@google.com>:

Support .* (pointer-to-member dereference) expressions in demangle.cc.

PiperOrigin-RevId: 223826797

--
772ca92179c3634f3e31a80bbc272ed8022e3572 by Abseil Team <absl-team@google.com>:

Fix misspellings in absl::variant comments and replace a ' with a `.

PiperOrigin-RevId: 223807911

--
35dcdc2fbf299d195658aac101887f6dcad1de2f by Abseil Team <absl-team@google.com>:

Bug fix in CMakeLists.txt file (SRCS --> HDRS).

The compressed_tuple header-only library is being defined
with the SRCS parameter instead of the HDRS parameter and
this has been observed to cause some builds on some platforms
to attempt to create a static library from it which fails
since there are no .cc sources.

PiperOrigin-RevId: 223805367

--
4a57a3d2045bb137c0c97958e45ce425190b8d3e by Chris Kennelly <ckennelly@google.com>:

Add test that absl::make_unique value initializes memory.

PiperOrigin-RevId: 223801819

--
dfe8289d7f4dcc6bb568a26aaf192a89e896bdfd by Chris Kennelly <ckennelly@google.com>:

SpinLock: Use exchange to avoid missing wakeups.

The default fast path for SpinLock::Unlock does not use an atomic.  If the
SpinLock becomes contended while we are unlocking between lockword_.load and
lockword_.store, we will fail to wake up the new waiter.  This can cause
unexpected latency spikes.

PiperOrigin-RevId: 223800369

--
9b9d35df786482f0016f77dd31691eff81503d23 by Abseil Team <absl-team@google.com>:

Update absl/hash/CMakeLists.txt to use new functions
i.e. absl_cc_(library|test)

PiperOrigin-RevId: 223755819

--
c2014e2704b87e7cdce2d2a0287c7e2397752296 by Abseil Team <absl-team@google.com>:

Update absl/debugging/CMakeLists.txt to use new functions
i.e. absl_cc_(library|test)

PiperOrigin-RevId: 223751986

--
d83a4e09126400e3fd80645cb03ee558f532271e by Derek Mauro <dmauro@google.com>:

Cleanup synchronization benchmarks.

PiperOrigin-RevId: 223589416

--
fad140b473586531b5b12843f942ec27dfcf5e93 by CJ Johnson <johnsoncj@google.com>:

Makes unifies the order of forward_iterator and input_iterator overloads

PiperOrigin-RevId: 223580660

--
6cd7c96faa7cc5f79f574e35a1b13837ef187d05 by Abseil Team <absl-team@google.com>:

Internal Change.

PiperOrigin-RevId: 223561629

--
bd2e545356b0f548af0e3c14bb2f7f0e712e49d0 by Shaindel Schwartz <shaindel@google.com>:

Remove misleading comments. try_emplace() does not exist for the hash_set containers.

PiperOrigin-RevId: 223543089

--
0cd380a53b587eb7aacc4003a4a3bbb6c78d7c10 by Derek Mauro <dmauro@google.com>:

Internal change

PiperOrigin-RevId: 223512551

--
7156dfee599cb72e9adddfe0e6ae07a95ddf10bb by Greg Miller <jgm@google.com>:

Fixes UB that would result from constructing, multiplying, or dividing a
Duration with a double "NaN" value. This CL changes the absl::Duration
*implementation* to return an InfiniteDuration value that has the same sign as
the given NaN.

PiperOrigin-RevId: 223407499

--
196b7d18609958267951882baf7f9429e49bcafa by CJ Johnson <johnsoncj@google.com>:

Addresses NVCC+MSVC compilation bug where `inlined_capacity()` was not considered valid in constexpr

PiperOrigin-RevId: 223397718
GitOrigin-RevId: cd076f55c1fa600131f6dda392533dfe61679fc0
Change-Id: I5423ca6470f661a7c6f73aa8fee49990446f157f
This commit is contained in:
Abseil Team 2018-12-04 11:01:12 -08:00 • committed by CJ Johnson
parent 926bfeb9ff
commit 44b0fafc62
30 changed files with 1126 additions and 383 deletions
Download patch file Download diff file Expand all files Collapse all files

24
absl/synchronization/BUILD.bazel
View file

@ -170,18 +170,32 @@ cc_test(
],
)
cc_test(
name = "mutex_benchmark",
cc_library(
name = "mutex_benchmark_common",
testonly = 1,
srcs = ["mutex_benchmark.cc"],
copts = ABSL_TEST_COPTS,
tags = ["benchmark"],
visibility = ["//visibility:private"],
copts = ABSL_DEFAULT_COPTS,
visibility = [
"//absl/synchronization:__pkg__",
],
deps = [
":synchronization",
":thread_pool",
"//absl/base",
"//absl/base:base_internal",
"@com_github_google_benchmark//:benchmark_main",
],
alwayslink = 1,
)
cc_binary(
name = "mutex_benchmark",
testonly = 1,
copts = ABSL_DEFAULT_COPTS,
visibility = ["//visibility:private"],
deps = [
":mutex_benchmark_common",
],
)
cc_test(

151
absl/synchronization/mutex_benchmark.cc
View file

@ -12,16 +12,154 @@
// See the License for the specific language governing permissions and
// limitations under the License.
#include <cstdint>
#include <mutex> // NOLINT(build/c++11)
#include <vector>
#include "benchmark/benchmark.h"
#include "absl/base/internal/sysinfo.h"
#include "absl/base/internal/cycleclock.h"
#include "absl/base/internal/spinlock.h"
#include "absl/synchronization/blocking_counter.h"
#include "absl/synchronization/internal/thread_pool.h"
#include "absl/synchronization/mutex.h"
#include "benchmark/benchmark.h"
namespace {
void BM_Mutex(benchmark::State& state) {
static absl::Mutex* mu = new absl::Mutex;
for (auto _ : state) {
absl::MutexLock lock(mu);
}
}
BENCHMARK(BM_Mutex)->UseRealTime()->Threads(1)->ThreadPerCpu();
static void DelayNs(int64_t ns, int* data) {
int64_t end = absl::base_internal::CycleClock::Now() +
ns * absl::base_internal::CycleClock::Frequency() / 1e9;
while (absl::base_internal::CycleClock::Now() < end) {
++(*data);
benchmark::DoNotOptimize(*data);
}
}
template <typename MutexType>
class RaiiLocker {
public:
explicit RaiiLocker(MutexType* mu) : mu_(mu) { mu_->Lock(); }
~RaiiLocker() { mu_->Unlock(); }
private:
MutexType* mu_;
};
template <>
class RaiiLocker<std::mutex> {
public:
explicit RaiiLocker(std::mutex* mu) : mu_(mu) { mu_->lock(); }
~RaiiLocker() { mu_->unlock(); }
private:
std::mutex* mu_;
};
template <typename MutexType>
void BM_Contended(benchmark::State& state) {
struct Shared {
MutexType mu;
int data = 0;
};
static auto* shared = new Shared;
int local = 0;
for (auto _ : state) {
// Here we model both local work outside of the critical section as well as
// some work inside of the critical section. The idea is to capture some
// more or less realisitic contention levels.
// If contention is too low, the benchmark won't measure anything useful.
// If contention is unrealistically high, the benchmark will favor
// bad mutex implementations that block and otherwise distract threads
// from the mutex and shared state for as much as possible.
// To achieve this amount of local work is multiplied by number of threads
// to keep ratio between local work and critical section approximately
// equal regardless of number of threads.
DelayNs(100 * state.threads, &local);
RaiiLocker<MutexType> locker(&shared->mu);
DelayNs(state.range(0), &shared->data);
}
}
BENCHMARK_TEMPLATE(BM_Contended, absl::Mutex)
->UseRealTime()
// ThreadPerCpu poorly handles non-power-of-two CPU counts.
->Threads(1)
->Threads(2)
->Threads(4)
->Threads(6)
->Threads(8)
->Threads(12)
->Threads(16)
->Threads(24)
->Threads(32)
->Threads(48)
->Threads(64)
->Threads(96)
->Threads(128)
->Threads(192)
->Threads(256)
// Some empirically chosen amounts of work in critical section.
// 1 is low contention, 200 is high contention and few values in between.
->Arg(1)
->Arg(20)
->Arg(50)
->Arg(200);
BENCHMARK_TEMPLATE(BM_Contended, absl::base_internal::SpinLock)
->UseRealTime()
// ThreadPerCpu poorly handles non-power-of-two CPU counts.
->Threads(1)
->Threads(2)
->Threads(4)
->Threads(6)
->Threads(8)
->Threads(12)
->Threads(16)
->Threads(24)
->Threads(32)
->Threads(48)
->Threads(64)
->Threads(96)
->Threads(128)
->Threads(192)
->Threads(256)
// Some empirically chosen amounts of work in critical section.
// 1 is low contention, 200 is high contention and few values in between.
->Arg(1)
->Arg(20)
->Arg(50)
->Arg(200);
BENCHMARK_TEMPLATE(BM_Contended, std::mutex)
->UseRealTime()
// ThreadPerCpu poorly handles non-power-of-two CPU counts.
->Threads(1)
->Threads(2)
->Threads(4)
->Threads(6)
->Threads(8)
->Threads(12)
->Threads(16)
->Threads(24)
->Threads(32)
->Threads(48)
->Threads(64)
->Threads(96)
->Threads(128)
->Threads(192)
->Threads(256)
// Some empirically chosen amounts of work in critical section.
// 1 is low contention, 200 is high contention and few values in between.
->Arg(1)
->Arg(20)
->Arg(50)
->Arg(200);
// Measure the overhead of conditions on mutex release (when they must be
// evaluated). Mutex has (some) support for equivalence classes allowing
// Conditions with the same function/argument to potentially not be multiply
@ -82,13 +220,4 @@ constexpr int kMaxConditionWaiters = 1024;
#endif
BENCHMARK(BM_ConditionWaiters)->RangePair(0, 2, 1, kMaxConditionWaiters);
void BM_ContendedMutex(benchmark::State& state) {
static absl::Mutex* mu = new absl::Mutex;
for (auto _ : state) {
absl::MutexLock lock(mu);
}
}
BENCHMARK(BM_ContendedMutex)->Threads(1);
BENCHMARK(BM_ContendedMutex)->ThreadPerCpu();
} // namespace

1
absl/synchronization/notification.h
View file

@ -52,6 +52,7 @@
#include <atomic>
#include "absl/base/macros.h"
#include "absl/synchronization/mutex.h"
#include "absl/time/time.h"