diff options
author | Abseil Team <absl-team@google.com> | 2019-11-13 08:54:32 -0800 |
---|---|---|
committer | Andy Getz <durandal@google.com> | 2019-11-13 15:04:44 -0500 |
commit | fa8c75182fbfdeddb2485fc0d53baeda3f40b7a3 (patch) | |
tree | cf973bc4dee81f08cb7edd811353d944d10fc5f9 /absl/base | |
parent | 85092b4b648ca729c6263c4a302a41dfff28705e (diff) |
Export of internal Abseil changes
--
049ac45508e335c6f010f2d28d71016b9fa65b4e by Derek Mauro <dmauro@google.com>:
Fix librt detection
PiperOrigin-RevId: 280207723
--
6382c3a9fb2643af9dc031f92ca846c4a78e249c by Andy Getzendanner <durandal@google.com>:
Fix Conan builds
Import of https://github.com/abseil/abseil-cpp/pull/400
PiperOrigin-RevId: 280025424
--
aebcd52b1686ac82663a8d0193b60d0122a43372 by Samuel Benzaquen <sbenza@google.com>:
Enable the assertion in the iterator's operator== and operator!=
PiperOrigin-RevId: 279998951
--
5b61d909e2159ac6fd45e0e456818db1e725ecd1 by Derek Mauro <dmauro@google.com>:
Add best effort support for compiling much of Abseil with MinGW.
This involves disabling ABSL_ATTRIBUTE_WEAK and adding link flags.
A change to CCTZ is still necessary.
Tests were not run yet, but most of them now build.
PiperOrigin-RevId: 279966541
--
4336f8b10cff906e2defdd7d1d449cde4907da5d by Abseil Team <absl-team@google.com>:
Add comments and relax memory orders in base_internal::CallOnceImpl.
Add a comment to document the memory order guarantee if
base_internal::SpinLockWait() is called and returns kOnceDone.
Add a comment for the load/store sequence in base_internal::CallOnceImpl
based on Mike Burrows' explanation.
The atomic load of 'control' in the #ifndef NDEBUG block does not need
std::memory_order_acquire. It can use std::memory_order_relaxed.
The atomic compare_exchange_strong of 'control' does not need
std::memory_order_acquire in the success case. It can use
std::memory_order_relaxed.
PiperOrigin-RevId: 279814155
--
407de3a5e9af957cded54a136ca0468bde620d4d by Abseil Team <absl-team@google.com>:
Added a script to generate abseil.podspec from all BUILD.bazel files automatically.
PiperOrigin-RevId: 279811441
--
26139497d4a363d6c7bc989c554da593e8819a07 by Derek Mauro <dmauro@google.com>:
Add missing copyright and Apache License to //absl/functional/BUILD.bazel
PiperOrigin-RevId: 279795227
--
98ed625b02af6e5834edf52a920d8ca2dab4cd90 by Matt Kulukundis <kfm@google.com>:
Switch the implementation of hashtablez to *only* work on platforms that have a
PER_THREAD_TLS.
The old case is very slow (global mutex) and nobody collects data from that
configuration anyway.
PiperOrigin-RevId: 279775149
--
07225900ef672c005c38f467ad3f92f38d0922b3 by Derek Mauro <dmauro@google.com>:
Remove the minumum glibc version check
PiperOrigin-RevId: 279750412
--
ec09956a951b4f52228ecc81968b8db7ae19ed15 by Derek Mauro <dmauro@google.com>:
CMake only: link with -lrt to support older glibc versions
PiperOrigin-RevId: 279741661
--
97b113fb2e8246f6152c36330ba13793b37154b6 by Xiaoyi Zhang <zhangxy@google.com>:
Internal change.
PiperOrigin-RevId: 279390188
--
ca8f72f2721546cc9b01bd01b2ea144962e6e0c5 by Andy Getzendanner <durandal@google.com>:
Expose PutTwoDigits for internal use within Abseil.
PiperOrigin-RevId: 279374239
--
14c6384cc03bbdfdefd2e4b635f104af5dd7e026 by Derek Mauro <dmauro@google.com>:
Remove log_severity sources from the base target.
They are already compiled as part of a separate library.
PiperOrigin-RevId: 279372619
--
3c5d926c718f8bf394e3bee87b6ba8d94601e0d3 by Abseil Team <absl-team@google.com>:
s/indepdent/independent/g in SimpleAtof's documentation.
PiperOrigin-RevId: 279350836
--
de2c44be8a8edf9efa1fe2007cba3564f3e5b0b8 by Abseil Team <absl-team@google.com>:
Internal change
PiperOrigin-RevId: 279346990
--
2ba078341423fcf6d0ba5ca1831f86570a26e615 by Samuel Benzaquen <sbenza@google.com>:
Add hash support for std::wstring, std::u16string and std::u32string.
PiperOrigin-RevId: 279320672
--
3272d3ffcfa55283a04f90e5868701912da95ef7 by Andy Soffer <asoffer@google.com>:
Removing a bunch of __restricts that amount to no performance differences. One
of these is the cause of https://github.com/abseil/abseil-cpp/issues/396. In
particular, in one of the Vector128Store functions, restricts on two pointers
that were indeed aliased seems to be the root cause of the issues.
Closes #396
PiperOrigin-RevId: 279318999
--
342f338ab31cc24344d5de8f28cf455bbb629a17 by Jorg Brown <jorg@google.com>:
Support uint128 in SimpleAtoi
PiperOrigin-RevId: 279234038
--
81cb0a04cf2dc4515d303679fc60968712191571 by Derek Mauro <dmauro@google.com>:
Change the check for futex availability to support older Linux systems
PiperOrigin-RevId: 279147079
--
cb4ca4aa4c8d2d710a5d483c56c4ce4f979e14b1 by Abseil Team <absl-team@google.com>:
Add IWYU pragma: export for int128 .inc files.
PiperOrigin-RevId: 279107098
--
b8df86ef610c366729f07326c726f3e34817b4dd by Abseil Team <absl-team@google.com>:
An optimization for Waiter::Post() in the SEM waiter mode.
Like the FUTEX waiter mode, Waiter::Post() only needs to call Poke() if
it incremented the atomic variable from 0.
PiperOrigin-RevId: 279086133
GitOrigin-RevId: 049ac45508e335c6f010f2d28d71016b9fa65b4e
Change-Id: I4c1a4073fff62cb6a1fcb1c104aa7d62dad588c2
Diffstat (limited to 'absl/base')
-rw-r--r-- | absl/base/BUILD.bazel | 4 | ||||
-rw-r--r-- | absl/base/CMakeLists.txt | 8 | ||||
-rw-r--r-- | absl/base/attributes.h | 6 | ||||
-rw-r--r-- | absl/base/call_once.h | 13 | ||||
-rw-r--r-- | absl/base/policy_checks.h | 10 |
5 files changed, 24 insertions, 17 deletions
diff --git a/absl/base/BUILD.bazel b/absl/base/BUILD.bazel index 6fc712d9..7e234bc6 100644 --- a/absl/base/BUILD.bazel +++ b/absl/base/BUILD.bazel @@ -192,7 +192,9 @@ cc_library( ], copts = ABSL_DEFAULT_COPTS, linkopts = select({ - "//absl:windows": [], + "//absl:windows": [ + "-DEFAULTLIB:shlwapi.lib", + ], "//conditions:default": ["-pthread"], }) + ABSL_DEFAULT_LINKOPTS, deps = [ diff --git a/absl/base/CMakeLists.txt b/absl/base/CMakeLists.txt index 9550cdb2..7ab6955e 100644 --- a/absl/base/CMakeLists.txt +++ b/absl/base/CMakeLists.txt @@ -14,6 +14,8 @@ # limitations under the License. # +find_library(LIBRT rt) + absl_cc_library( NAME atomic_hook @@ -163,16 +165,18 @@ absl_cc_library( "internal/thread_identity.h" "internal/tsan_mutex_interface.h" "internal/unscaledcycleclock.h" - "log_severity.h" SRCS "internal/cycleclock.cc" "internal/spinlock.cc" "internal/sysinfo.cc" "internal/thread_identity.cc" "internal/unscaledcycleclock.cc" - "log_severity.cc" COPTS ${ABSL_DEFAULT_COPTS} + LINKOPTS + ${ABSL_DEFAULT_LINKOPTS} + $<$<BOOL:${LIBRT}>:${LIBRT}> + $<$<BOOL:${MINGW}>:"shlwapi"> DEPS absl::atomic_hook absl::base_internal diff --git a/absl/base/attributes.h b/absl/base/attributes.h index 7b7656a8..acd1c526 100644 --- a/absl/base/attributes.h +++ b/absl/base/attributes.h @@ -158,9 +158,11 @@ // Weak attributes currently do not work properly in LLVM's Windows backend, // so disable them there. See https://bugs.llvm.org/show_bug.cgi?id=37598 // for further information. -#if (ABSL_HAVE_ATTRIBUTE(weak) || \ +// The MinGW compiler doesn't complain about the weak attribute until the link +// step, presumably because Windows doesn't use ELF binaries. +#if (ABSL_HAVE_ATTRIBUTE(weak) || \ (defined(__GNUC__) && !defined(__clang__))) && \ - !(defined(__llvm__) && defined(_WIN32)) + !(defined(__llvm__) && defined(_WIN32)) && !defined(__MINGW32__) #undef ABSL_ATTRIBUTE_WEAK #define ABSL_ATTRIBUTE_WEAK __attribute__((weak)) #define ABSL_HAVE_ATTRIBUTE_WEAK 1 diff --git a/absl/base/call_once.h b/absl/base/call_once.h index 4aa6360c..e1614e51 100644 --- a/absl/base/call_once.h +++ b/absl/base/call_once.h @@ -148,7 +148,7 @@ void CallOnceImpl(std::atomic<uint32_t>* control, Args&&... args) { #ifndef NDEBUG { - uint32_t old_control = control->load(std::memory_order_acquire); + uint32_t old_control = control->load(std::memory_order_relaxed); if (old_control != kOnceInit && old_control != kOnceRunning && old_control != kOnceWaiter && @@ -166,14 +166,23 @@ void CallOnceImpl(std::atomic<uint32_t>* control, // Must do this before potentially modifying control word's state. base_internal::SchedulingHelper maybe_disable_scheduling(scheduling_mode); // Short circuit the simplest case to avoid procedure call overhead. + // The base_internal::SpinLockWait() call returns either kOnceInit or + // kOnceDone. If it returns kOnceDone, it must have loaded the control word + // with std::memory_order_acquire and seen a value of kOnceDone. uint32_t old_control = kOnceInit; if (control->compare_exchange_strong(old_control, kOnceRunning, - std::memory_order_acquire, std::memory_order_relaxed) || base_internal::SpinLockWait(control, ABSL_ARRAYSIZE(trans), trans, scheduling_mode) == kOnceInit) { base_internal::Invoke(std::forward<Callable>(fn), std::forward<Args>(args)...); + // The call to SpinLockWake below is an optimization, because the waiter + // in SpinLockWait is waiting with a short timeout. The atomic load/store + // sequence is slightly faster than an atomic exchange: + // old_control = control->exchange(base_internal::kOnceDone, + // std::memory_order_release); + // We opt for a slightly faster case when there are no waiters, in spite + // of longer tail latency when there are waiters. old_control = control->load(std::memory_order_relaxed); control->store(base_internal::kOnceDone, std::memory_order_release); if (old_control == base_internal::kOnceWaiter) { diff --git a/absl/base/policy_checks.h b/absl/base/policy_checks.h index 699fb1a2..4dfa49e5 100644 --- a/absl/base/policy_checks.h +++ b/absl/base/policy_checks.h @@ -82,16 +82,6 @@ // Standard Library Check // ----------------------------------------------------------------------------- -// We have chosen glibc 2.12 as the minimum as it was tagged for release -// in May, 2010 and includes some functionality used in Google software -// (for instance pthread_setname_np): -// https://sourceware.org/ml/libc-alpha/2010-05/msg00000.html -#if defined(__GLIBC__) && defined(__GLIBC_PREREQ) -#if !__GLIBC_PREREQ(2, 12) -#error "Minimum required version of glibc is 2.12." -#endif -#endif - #if defined(_STLPORT_VERSION) #error "STLPort is not supported." #endif |