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

Export of internal Abseil changes.

Browse source 
--
3f04cd3c25a99df91ff913977b8c5b343532db5d by Abseil Team <absl-team@google.com>:

Stricter memory order constraints for CycleClock callback.

PiperOrigin-RevId: 242670115

--
216db48375306490f1722a11aaf33080939d9f2f by Abseil Team <absl-team@google.com>:

internal/optional.h: move macro from types/optional.h

ABSL_OPTIONAL_USE_INHERITING_CONSTRUCTORS is only used within this file.
additionally check the macro with #ifdef rather than #if, fixes -Wundef
warning:
'ABSL_OPTIONAL_USE_INHERITING_CONSTRUCTORS' is not defined, evaluates to 0
PiperOrigin-RevId: 242548205

--
fbe22e7d8dc5c0b3d43ac26297e97ddbaeab3d39 by Samuel Benzaquen <sbenza@google.com>:

Implement %f natively for any input.
It evaluates the input at runtime and allocates stack space accordingly.

This removes a potential fallback into snprintf, improves performance, and removes all memory allocations in this formatting path.

PiperOrigin-RevId: 242531736

--
1458f9ba2a79ef0534e46527cd34770dee54164d by Greg Falcon <gfalcon@google.com>:

Add explicit check for NVCC in compressed_tuple.h.

NVCC claims to be MSVC, but does not implement this MSVC attribute.

PiperOrigin-RevId: 242513453
GitOrigin-RevId: 3f04cd3c25a99df91ff913977b8c5b343532db5d
Change-Id: I0742e8619c5248c7607961113e406486bc0e279b
This commit is contained in:
Abseil Team 2019-04-09 08:22:32 -07:00 • committed by Shaindel Schwartz
parent 044da8a29c
commit dbae8764fb
8 changed files with 78 additions and 584 deletions
Download patch file Download diff file Expand all files Collapse all files

18
absl/base/internal/cycleclock.cc
View file

@ -55,10 +55,23 @@ static constexpr int32_t kShift = 2;
static constexpr double kFrequencyScale = 1.0 / (1 << kShift);
static std::atomic<CycleClockSourceFunc> cycle_clock_source;
CycleClockSourceFunc LoadCycleClockSource() {
// Optimize for the common case (no callback) by first doing a relaxed load;
// this is significantly faster on non-x86 platforms.
if (cycle_clock_source.load(std::memory_order_relaxed) == nullptr) {
return nullptr;
}
// This corresponds to the store(std::memory_order_release) in
// CycleClockSource::Register, and makes sure that any updates made prior to
// registering the callback are visible to this thread before the callback is
// invoked.
return cycle_clock_source.load(std::memory_order_acquire);
}
} // namespace
int64_t CycleClock::Now() {
auto fn = cycle_clock_source.load(std::memory_order_relaxed);
auto fn = LoadCycleClockSource();
if (fn == nullptr) {
return base_internal::UnscaledCycleClock::Now() >> kShift;
}
@ -70,7 +83,8 @@ double CycleClock::Frequency() {
}
void CycleClockSource::Register(CycleClockSourceFunc source) {
cycle_clock_source.store(source, std::memory_order_relaxed);
// Corresponds to the load(std::memory_order_acquire) in LoadCycleClockSource.
cycle_clock_source.store(source, std::memory_order_release);
}
#else