--
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
--
ee19e203eca970ff88e8f25ce4e19c32e143b988 by Jon Cohen <cohenjon@google.com>:
Exception safety testing no longer uses absl::optional
PiperOrigin-RevId: 220336204
--
460666eb0b316a8b4aeedc589644d53b05251bd1 by Derek Mauro <dmauro@google.com>:
Rework SwissTable SSE2 support
- Use SSE2 on MSVC when available
https://github.com/abseil/abseil-cpp/issues/210
- Emulate _mm_cmpgt_epi8 with other SSE2 instructions when using
-funsigned-char under GCC
https://github.com/abseil/abseil-cpp/issues/209
PiperOrigin-RevId: 220312351
--
1f4318ecedf8d539b7b698eb803d613ad6b69278 by Abseil Team <absl-team@google.com>:
Change CollectPerfectRatios to use 10 trials to smooth out the outliers in the
sample.
PiperOrigin-RevId: 220286579
--
6755abc2673553a7f578bb29c6e9ca8d991bc9c8 by Abseil Team <absl-team@google.com>:
Internal change
PiperOrigin-RevId: 220274307
--
8645b6187329ebf0aaf3c2de2888ba44466cd879 by Abseil Team <absl-team@google.com>:
* #endif for a header guard should reference the guard macro in a comment
PiperOrigin-RevId: 220206868
--
3987a7ad11319230910931cd2468b60b3fd1b85c by Gennadiy Civil <misterg@google.com>:
Internal Change
PiperOrigin-RevId: 220136674
--
cc908c1db2ee0d4523dc813e33f600583bb986c5 by Abseil Team <absl-team@google.com>:
absl: fix backoff logic in SpinLockWait
There are 3 bugs in loop variable handling:
1. It starts with 0, but AbslInternalSpinLockDelay ignores loop == 0.
So it does not actually wait when it should.
2. loop is incremented after successful state changes,
but it should not (why would be increase backoff delay after that?).
3. loop is incremented after CAS failures,
but it should not (why would be increase backoff delay after that?).
Use the same handling of loop as used in SpinLock.
PiperOrigin-RevId: 220136079
--
a0a1c6ef5910ebd28e07215d7df03cc0da0b3eed by Abseil Team <absl-team@google.com>:
absl: relax unnecessarily strong memory ordering in SpinLock::SlowLock
We don't need to acquire visibility over anything when setting kSpinLockSleeper.
Replace the confusing and unnecessarily strong memory order with relaxed.
PiperOrigin-RevId: 220023380
--
c50858b51af28b9fca1a62616324f85f3e84ea74 by Tom Manshreck <shreck@google.com>:
Update comments in flat_hash_map, node_hash_{set, map} and the containers developer guide
PiperOrigin-RevId: 219938692
--
e87b7d1a5f61e165b1c44d3b16d8d967197cdfce by CJ Johnson <johnsoncj@google.com>:
Rearranges the public methods of InlinedVector and cleans up the comments
PiperOrigin-RevId: 219896257
--
f3234c466f792e0fc4bfd21fc7919dba5e679375 by CJ Johnson <johnsoncj@google.com>:
Adds branch prediction to exceptional early exit cases of inlined vector's API
PiperOrigin-RevId: 219887173
--
4dfccf1a81ca0425912d3da25a8470f78c532ce4 by CJ Johnson <johnsoncj@google.com>:
Fixes the InlinedVector public interface to use the allocator type references instead of assuming the type
Also cleans up some cruft in formatting and comments
PiperOrigin-RevId: 219878876
--
4bb6a2b892abb10bd6a424db7e94ed8640802470 by Tom Manshreck <shreck@google.com>:
Add comments on constructor and assignment operator support to flat_hash_set
PiperOrigin-RevId: 219825338
--
c23f973e2f7f4feea0da36bf8a9c3f8a8954bb74 by Abseil Team <absl-team@google.com>:
Import of CCTZ from GitHub.
PiperOrigin-RevId: 219823847
GitOrigin-RevId: ee19e203eca970ff88e8f25ce4e19c32e143b988
Change-Id: I288c927ca481dc57340420dbb4c278a05cf15e83
NixOS 18.09 switches from slim to something else (lightdm?) as the
default display manager, however DBUS is broken in that manager if the
session is ever restarted.
This reverts back to slim, which may not look as fancy by default but
it actually works.
--
4e224c85c3730398919fc5195cb1fc7a752e6e4f by Mark Barolak <mbar@google.com>:
Update some references to "StringPiece" to say "string_view" instead.
PiperOrigin-RevId: 219693697
--
6bdc925a3db5e97f1f8a404bdfda2e47e48f7b9a by Abseil Team <absl-team@google.com>:
Disable weak symbols for the Windows backend of LLVM, since they are currently buggy. See https://bugs.llvm.org/show_bug.cgi?id=37598 for more information.
PiperOrigin-RevId: 219676493
--
5823f495036181191f435efa4c45d60ca3160145 by Derek Mauro <dmauro@google.com>:
Don't use the SSE2 implementation of container_internal::Group
with -funsigned-char under GCC.
This is a workaround for https://github.com/abseil/abseil-cpp/issues/209.
_mm_cmpgt_epi8 is broken under GCC with -funsigned-char.
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=87853
PiperOrigin-RevId: 219666066
GitOrigin-RevId: 4e224c85c3730398919fc5195cb1fc7a752e6e4f
Change-Id: I2f115d0256576cf476ae73a9464c21d4106a2a56
--
9e8aa654630015ea8221703b0ea10dd1a47a848f by Abseil Team <absl-team@google.com>:
Fix typo
PiperOrigin-RevId: 219474910
--
b1621572fb5a326642766f9f0f16abc3620dd6e8 by Xiaoyi Zhang <zhangxy@google.com>:
Applies https://github.com/abseil/abseil-cpp/pull/194 to fix raw_hash_map
build issues for Windows VS 15.8.
PiperOrigin-RevId: 219473484
--
a61df296bc5449261c6deccb6410df83d560d210 by Jon Cohen <cohenjon@google.com>:
Remove reference to ABSL_COMPILE_CXXFLAGS in our cmakelists files. It wasn't used, but we will soon instead use an INERFACE target with target_compile_options for the compile options needed for Abseil.
PiperOrigin-RevId: 219352641
--
0d9536a26c03229df3e782e78a8211e450259af2 by Abseil Team <absl-team@google.com>:
Add missing dependencies on base:core_headers
PiperOrigin-RevId: 219320098
--
5693e836e4ec4ce23315f3a7846c12c64cf3d5ab by Abseil Team <absl-team@google.com>:
Remove dynamic_annotations from core_headers, and add core_headers to CMakeLists.
PiperOrigin-RevId: 219306959
GitOrigin-RevId: 9e8aa654630015ea8221703b0ea10dd1a47a848f
Change-Id: Ie52b9fa7ab1285b6846ceea5503caf744e7303d8
Previously, config would only be read from XDG_CONFIG_HOME. This change
allows reading config from additional directories, which enables e.g.
per-project binary caches or chroot stores with the help of direnv.
--
ba4dd47492748bd630462eb68b7959037fc6a11a by Abseil Team <absl-team@google.com>:
Work around nvcc 9.0 compiler bug for open-source Tensorflow build.
With the current implementation, when I (unintentionally and transitively)
include absl/types/optional.h in a CUDA compilation unit, I get the following
nvcc error message:
INFO: From Compiling tensorflow/core/kernels/crop_and_resize_op_gpu.cu.cc:
external/com_google_absl/absl/types/optional.h: In member function 'void absl::optional_internal::optional_data_dtor_base<T, <anonymous> >::destruct()':
external/com_google_absl/absl/types/optional.h:185:50: error: '__T0' was not declared in this scope
data_.~T();
I've also seen similar compilation failures online, for flat_hash_map:
https://devtalk.nvidia.com/default/topic/1042599/nvcc-preprocessor-bug-causes-compilation-failure/
The bug is always around unnamed template parameters. Therefore, the workaround is to make them named.
PiperOrigin-RevId: 219208288
--
dad2f40cb2e8d5017660985ef6fb57f3c3cdcc80 by CJ Johnson <johnsoncj@google.com>:
Adds internal macros for catching and throwing unknown exception types
PiperOrigin-RevId: 219207362
--
0a9840328d2d86e8420b853435fdbf1f7a19d931 by Abseil Team <absl-team@google.com>:
Fix typo in mutex.h comments.
PiperOrigin-RevId: 219199397
--
0d576dc7597564210bfdf91518075064756f0bf4 by Matt Calabrese <calabrese@google.com>:
Internal change.
PiperOrigin-RevId: 219185475
--
66be156095571959fb19a76da8ad0b53ec37658e by Abseil Team <absl-team@google.com>:
Fix alignment conformance for VS 2017 >= 15.8 (fix#193)
PiperOrigin-RevId: 219129894
--
a6e1825a12587945f8194677ccfdcaba6f7aad1d by Abseil Team <absl-team@google.com>:
Reapply PR #173
PiperOrigin-RevId: 219129361
--
cf72ade4881b25acc6ccaea468f69793a0fdce32 by Abseil Team <absl-team@google.com>:
Update .gitignore
PiperOrigin-RevId: 219127495
--
0537490c6348a2cb489abe15638928ac5aa6982a by Jon Cohen <cohenjon@google.com>:
Small refactor and reformat of error messages from the exception safety test framework.
PiperOrigin-RevId: 218927773
--
4c556ca45fa25698ad12002a00c713aeceefab73 by CJ Johnson <johnsoncj@google.com>:
Updates the inlined vector swap tests to check for number of moves that took place if available
PiperOrigin-RevId: 218900777
--
dcbfda0021a1e6dfa9586986b1269c06ec394053 by Mark Barolak <mbar@google.com>:
Add parens around calls to std::numeric_limits<>::min and
std::numeric_limits<>::max to prevent compilation errors on Windows platforms
where min and max are defined as macros.
PiperOrigin-RevId: 218888700
GitOrigin-RevId: ba4dd47492748bd630462eb68b7959037fc6a11a
Change-Id: I0e393958eb8cb501b85f6114979f6d4d86ed996c
The use of TransferManager has several issues, including that it
doesn't allow setting a Content-Encoding without a patch, and it
doesn't handle exceptions in worker threads (causing termination on
memory allocation failure).
Fixes#2493.
This commit partially reverts 48662d151b. When
copying from an older store (in my case a store running Nix 1.11.7), nix would
throw errors about there being no hash. This is fixed by recalculating the hash.
--
dabd5614eec687a27bcba28e1d98e84ce716f281 by Abseil Team <absl-team@google.com>:
Demonstrate that std::bitset is supported by AbslHash, both with a comment and
a test.
PiperOrigin-RevId: 218782040
GitOrigin-RevId: dabd5614eec687a27bcba28e1d98e84ce716f281
Change-Id: I777d5e030ba8c6b8a2a353e29ace87484caa811f