Abseil Common Libraries (C++) (grcp 依赖) https://abseil.io/
You can not select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.

1999 lines
63 KiB

Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
// Copyright 2020 The Abseil Authors.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// https://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#include "absl/strings/cord.h"
#include <algorithm>
#include <atomic>
#include <cstddef>
#include <cstdio>
#include <cstdlib>
#include <iomanip>
#include <iostream>
#include <limits>
#include <ostream>
#include <sstream>
#include <type_traits>
#include <unordered_set>
#include <vector>
#include "absl/base/casts.h"
#include "absl/base/internal/raw_logging.h"
#include "absl/base/macros.h"
#include "absl/base/port.h"
#include "absl/container/fixed_array.h"
#include "absl/container/inlined_vector.h"
#include "absl/strings/escaping.h"
#include "absl/strings/internal/cord_internal.h"
#include "absl/strings/internal/cord_rep_flat.h"
Export of internal Abseil changes -- c68f1886f5e8fd90eb0c2d2e68feaf00a7cdacda by CJ Johnson <johnsoncj@google.com>: Introduce absl::Cleanup to the OSS repo PiperOrigin-RevId: 354583156 -- 17030cf388e10f7eb959e3e566326d1072ce392e by Abseil Team <absl-team@google.com>: Internal change only PiperOrigin-RevId: 354574953 -- e979d7236d4f3252e79ddda6739b67a9a326bf6d by CJ Johnson <johnsoncj@google.com>: Internal change PiperOrigin-RevId: 354545297 -- 7ea02b3783f7f49ef97d86a8f6580a19cc57df14 by Abseil Team <absl-team@google.com>: Pre-allocate memory for vectors where the size is known. PiperOrigin-RevId: 354344576 -- 9246c7cb11f1d6444f79ebe25acc69a8a9b870e0 by Matt Kulukundis <kfm@google.com>: Add support for Elbrus 2000 (e2k) Import of https://github.com/abseil/abseil-cpp/pull/889 PiperOrigin-RevId: 354344013 -- 0fc93d359cc1fb307552e917b37b7b2e7eed822f by Abseil Team <absl-team@google.com>: Integrate CordRepRing logic into cord (but do not enable it) PiperOrigin-RevId: 354312238 -- eda05622f7da71466723acb33403f783529df24b by Abseil Team <absl-team@google.com>: Protect ignore diagnostic with "__has_warning". PiperOrigin-RevId: 354112334 -- 47716c5d8fb10efa4fdd801d28bac414c6f8ec32 by Abseil Team <absl-team@google.com>: Rearrange InlinedVector copy constructor and destructor to treat a few special cases inline and then tail-call a non-inlined routine for the rest. In particular, we optimize for empty vectors in both cases. Added a couple of benchmarks that copy either an InlVec<int64> or an InlVec<InlVec<int64>>. Speed difference: ``` BM_CopyTrivial/0 0.92ns +- 0% 0.47ns +- 0% -48.91% (p=0.000 n=11+12) BM_CopyTrivial/1 0.92ns +- 0% 1.15ns +- 0% +25.00% (p=0.000 n=10+9) BM_CopyTrivial/8 8.57ns +- 0% 10.72ns +- 1% +25.16% (p=0.000 n=10+12) BM_CopyNonTrivial/0 3.21ns +- 0% 0.70ns +- 0% -78.23% (p=0.000 n=12+10) BM_CopyNonTrivial/1 5.88ns +- 1% 5.51ns +- 0% -6.28% (p=0.000 n=10+8) BM_CopyNonTrivial/8 21.5ns +- 1% 15.2ns +- 2% -29.23% (p=0.000 n=12+12) ``` Note: the slowdowns are a few cycles which is expected given the procedure call added in that case. We decided this is a good tradeoff given the code size reductions and the more significant speedups for empty vectors. Size difference (as measured by nm): ``` BM_CopyTrivial from 1048 bytes to 326 bytes. BM_CopyNonTrivial from 749 bytes to 470 bytes. ``` Code size for a large binary drops by ~500KB (from 349415719 to 348906015 348906191). All of the benchmarks that showed a significant difference: Ones that improve with this CL: ``` BM_CopyNonTrivial/0 3.21ns +- 0% 0.70ns +- 0% -78.23% (p=0.000 n=12+10) BM_InlinedVectorFillString/0 0.93ns +- 0% 0.24ns +- 0% -74.19% (p=0.000 n=12+10) BM_InlinedVectorAssignments/1 10.5ns +- 0% 4.1ns +- 0% -60.64% (p=0.000 n=11+10) BM_InlinedVectorAssignments/2 10.7ns +- 0% 4.4ns +- 0% -59.08% (p=0.000 n=11+11) BM_CopyTrivial/0 0.92ns +- 0% 0.47ns +- 0% -48.91% (p=0.000 n=11+12) BM_CopyNonTrivial/8 21.5ns +- 1% 15.2ns +- 2% -29.23% (p=0.000 n=12+12) BM_StdVectorEmpty 0.47ns +- 1% 0.35ns +- 0% -24.73% (p=0.000 n=12+12) BM_StdVectorSize 0.46ns +- 2% 0.35ns +- 0% -24.32% (p=0.000 n=12+12) BM_SwapElements<LargeCopyableOnly>/0 3.44ns +- 0% 2.76ns +- 1% -19.83% (p=0.000 n=11+11) BM_InlinedVectorFillRange/256 20.7ns +- 1% 17.8ns +- 0% -14.08% (p=0.000 n=12+9) BM_CopyNonTrivial/1 5.88ns +- 1% 5.51ns +- 0% -6.28% (p=0.000 n=10+8) BM_SwapElements<LargeCopyableMovable>/1 4.19ns +- 0% 3.95ns +- 1% -5.63% (p=0.000 n=11+12) BM_SwapElements<LargeCopyableMovableSwappable>/1 4.18ns +- 0% 3.99ns +- 0% -4.70% (p=0.000 n=9+11) BM_SwapElements<LargeCopyableMovable>/0 2.41ns +- 0% 2.31ns +- 0% -4.45% (p=0.000 n=12+12) BM_InlinedVectorFillRange/64 8.25ns +- 0% 8.04ns +- 0% -2.51% (p=0.000 n=12+11) BM_SwapElements<LargeCopyableOnly>/1 82.4ns +- 0% 81.5ns +- 0% -1.06% (p=0.000 n=12+12) ``` Ones that get worse with this CL: ``` BM_CopyTrivial/1 0.92ns +- 0% 1.15ns +- 0% +25.00% (p=0.000 n=10+9) BM_CopyTrivial/8 8.57ns +- 0% 10.72ns +- 1% +25.16% (p=0.000 n=10+12) BM_SwapElements<LargeCopyableMovableSwappable>/512 1.48ns +- 1% 1.66ns +- 1% +11.88% (p=0.000 n=12+12) BM_InlinedVectorFillString/1 11.5ns +- 0% 12.8ns +- 1% +11.62% (p=0.000 n=12+11) BM_SwapElements<LargeCopyableMovableSwappable>/64 1.48ns +- 2% 1.66ns +- 1% +11.66% (p=0.000 n=12+11) BM_SwapElements<LargeCopyableMovableSwappable>/1k 1.48ns +- 1% 1.65ns +- 2% +11.32% (p=0.000 n=12+12) BM_SwapElements<LargeCopyableMovable>/512 1.48ns +- 2% 1.58ns +- 4% +6.62% (p=0.000 n=11+12) BM_SwapElements<LargeCopyableMovable>/1k 1.49ns +- 2% 1.58ns +- 3% +6.05% (p=0.000 n=12+12) BM_SwapElements<LargeCopyableMovable>/64 1.48ns +- 2% 1.57ns +- 4% +6.04% (p=0.000 n=11+12) BM_InlinedVectorFillRange/1 4.81ns +- 0% 5.05ns +- 0% +4.83% (p=0.000 n=11+11) BM_InlinedVectorFillString/8 79.4ns +- 1% 83.1ns +- 1% +4.64% (p=0.000 n=10+12) BM_StdVectorFillString/1 16.3ns +- 0% 16.6ns +- 0% +2.13% (p=0.000 n=11+8) ``` PiperOrigin-RevId: 353906786 -- 8e26518b3cec9c598e5e9573c46c3bd1b03a67ef by Abseil Team <absl-team@google.com>: Internal change PiperOrigin-RevId: 353737330 -- f206ae0983e58c9904ed8b8f05f9caf564a446be by Matt Kulukundis <kfm@google.com>: Import of CCTZ from GitHub. PiperOrigin-RevId: 353682256 GitOrigin-RevId: c68f1886f5e8fd90eb0c2d2e68feaf00a7cdacda Change-Id: I5790c1036c4f543c701d1039848fabf7ae881ad8
4 years ago
#include "absl/strings/internal/cord_rep_ring.h"
Export of internal Abseil changes -- 341670bce317dd6af8d3c066970230591a47e80c by Martijn Vels <mvels@google.com>: Change GetStack() and GetParentStack() to return absl::Span PiperOrigin-RevId: 368765721 -- 6aaab9536d6957303c7aba100c3afaa6fb0ea2c8 by Martijn Vels <mvels@google.com>: Remove locking from parent stack. This change removes the need to lock all access to `parent_stack' by making the 'copy constructor' logic specify the 'copied from' CordzInfo (where available) to the TrackCord function, after which parent_stack is immutable. PiperOrigin-RevId: 368760630 -- b19e2059cada35a8ede994833018edac94de6ddc by Martijn Vels <mvels@google.com>: Add cordz instrumentation to Cord PiperOrigin-RevId: 368746225 -- 67b8bbf980f0f4e1db79aa32968e9a715a09b51a by Martijn Vels <mvels@google.com>: Create ABSL_INTERNAL_CORDZ_ENABLED define controlling when Cordz code is enabled There are specific builds and condtions under which we don't support cordz sampling, which is per this change represented by ABSL_INTERNAL_CORDZ_ENABLED being defined. PiperOrigin-RevId: 368731603 -- 8cbfe0e3169637a620f4b66ad2bc2ce340879cb0 by Martijn Vels <mvels@google.com>: Add a `rep` property to CordzInfo to be managed by Cord logic. This change adds a `rep` property to CordzInfo, which is intended to be used by collection logic. Mini design: Cord invokes TrackCord() providing the active 'root' cordrep of the newly sampled Cord, returning a CordzInfo with a weak (uncounted) reference to this root. Cord invokes `SetCordRep()` each time the root cordrep of the sampled Cord is updated while holding `mutex()`. Cord must also obtain `mutex()` _before_ removing a reference on the old root. i.e.: Cord must guarantee that the (weak) reference held in CordzInfo is at all times valid. CordzInfo collection code can then safely obtain a (reference counted) rep pointer by adding a reference to `rep_` while holding `mutex()`. This requires only a very brief critical section inside CordzInfo logic, minimizing contention with concurrent Cord updates. Cord code should typically obtain and hold `mutex()` for the entirety of each mutating Cord operation on a sampled cord. As Cord is thread compatible, it never competes on the lock with any other thread. The only possible concurrent access is from Cordz collection code, which should be a relatively rare event. PiperOrigin-RevId: 368673758 -- 1255120dce2bdd6b4205a34a0e555e0b74b6152f by Martijn Vels <mvels@google.com>: Remove 'depth' from active recorded metrics. Going forward we do not 'live' record depth (and size), but will observe these at collection time only. PiperOrigin-RevId: 368636572 -- 83e5146e35f221736b49e9f0a8805f8c159a51db by Martijn Vels <mvels@google.com>: Make cordz targets visible in OSS PiperOrigin-RevId: 368615010 -- dcb16a4f1239151f0a8c70a8cfeb29dabbd113b8 by Martijn Vels <mvels@google.com>: Internal cleanup PiperOrigin-RevId: 368514666 GitOrigin-RevId: 341670bce317dd6af8d3c066970230591a47e80c Change-Id: I94cecfbbd441eb386f99fc5186c468a7a5538862
4 years ago
#include "absl/strings/internal/cordz_statistics.h"
Export of internal Abseil changes -- ac1df60490c9583e475e22de7adfc40023196fbf by Martijn Vels <mvels@google.com>: Change Cord constructor(string_view) to explicit make_tree and Cordz tracking This CL changes the ctor to use an easier to maintain model where Cord code explicitly invokes Cordz update or new / tree logic, which avoids the ambiguity of the 'branched' InlineRep::set_tree code. This removes the need to equip InlineRep with 'MethodIdentifier' or other necessary call info, and also is a cleaner model: InlineRep is carrying too much code now that should plainly sit in Cord, especially with all internal abstractions having moved to InlineData. See child CL(s) for desired state PiperOrigin-RevId: 369433619 -- b665af7f586e6c679a8b27d4f78d5a1d2b596058 by Abseil Team <absl-team@google.com>: Rename the 'Compare' template type to 'LessThan', as the passed-in function is expected to act like operator<. It is worth avoiding confusion with std::compare, which returns an int (-1/0/1), as due to implicit casting this can lead to hard-to-spot bugs. PiperOrigin-RevId: 369391118 -- c3c775269cad0f4982ec63f3616dd78bb9e52dca by Martijn Vels <mvels@google.com>: Integrate CordzUpdateTracker into CordzInfo PiperOrigin-RevId: 369348824 -- 771d81ed357496c117179e1daec76eba5155932d by Martijn Vels <mvels@google.com>: Replace mutex() with Lock() / Unlock() function Mini design future tracking of CordzInfo sampled cords: CordzInfo holds a CordRep* reference without a reference count. Cord is responsible for synchronizing updates for sampled cords such that the CordRep* contained in CordzInfo is at all times valid. This is done by scoping Lock() and Unlock() calls around the code modifying the code of a sampled cord. For example (using the future CL CordzUpdateScope()): CordzInfo* cordz_info = get_cordz_info(); CordzUpdateScope scope(cordz_info, CordzUpdateTracker::kRemovePrefix); CordRep* rep = RemovePrefixImpl(root); set_tree(rep); if (cordz_info) { cordz_info->SetCordRep(rep); } On CordzInfo::Unlock(), if the internal rep is null, the cord is no longer sampled, and CordzInfo will be deleted. Thus any update resulting in the Cord being inlined will automatically no longer be sampled. PiperOrigin-RevId: 369338802 -- 5563c12df04a1e965a03b50bdd032739c55c0706 by Martijn Vels <mvels@google.com>: Add UpdateTracker to CordzStatistics PiperOrigin-RevId: 369318178 -- 6b4d8463722a3e55a3e8f6cb3741a41055e7f83e by Martijn Vels <mvels@google.com>: Add kClear, kConstructor* and kUnknown values and fix typo PiperOrigin-RevId: 369297163 -- 041adcbc929789d6d53371a8236840fc350e1eeb by Derek Mauro <dmauro@google.com>: Switch from malloc to operator new in pool_urbg.cc so it can only fail by throwing/aborting PiperOrigin-RevId: 369274087 -- 5d97a5f43e3f2d02d0a5bbe586d93b5751812981 by Benjamin Barenblat <bbaren@google.com>: Correct Thumb function bound computation in the symbolizer On 32-bit ARM, all functions are aligned to multiples of two bytes, and the lowest-order bit in a function’s address is ignored by the CPU when computing branch targets. That bit is still present in instructions and ELF symbol tables, though; it’s repurposed to indicate whether the function contains ARM or Thumb code. If the symbolizer doesn’t ignore that bit, it will believe Thumb functions have boundaries that are off by one byte, so instruct the symbolizer to null out the lowest-order bit after retrieving it from the symbol table. PiperOrigin-RevId: 369254082 -- 462bb307c6cc332c1e2c3adb5f0cad51804bf937 by Derek Mauro <dmauro@google.com>: Add a check for malloc failure in pool_urbg.cc GitHub #940 PiperOrigin-RevId: 369238100 GitOrigin-RevId: ac1df60490c9583e475e22de7adfc40023196fbf Change-Id: Ic6ec91c62cd3a0031f6a75a43a83da959ece2d25
4 years ago
#include "absl/strings/internal/cordz_update_tracker.h"
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
#include "absl/strings/internal/resize_uninitialized.h"
#include "absl/strings/str_cat.h"
#include "absl/strings/str_format.h"
#include "absl/strings/str_join.h"
#include "absl/strings/string_view.h"
namespace absl {
ABSL_NAMESPACE_BEGIN
using ::absl::cord_internal::CordRep;
using ::absl::cord_internal::CordRepConcat;
using ::absl::cord_internal::CordRepExternal;
using ::absl::cord_internal::CordRepFlat;
Export of internal Abseil changes -- c68f1886f5e8fd90eb0c2d2e68feaf00a7cdacda by CJ Johnson <johnsoncj@google.com>: Introduce absl::Cleanup to the OSS repo PiperOrigin-RevId: 354583156 -- 17030cf388e10f7eb959e3e566326d1072ce392e by Abseil Team <absl-team@google.com>: Internal change only PiperOrigin-RevId: 354574953 -- e979d7236d4f3252e79ddda6739b67a9a326bf6d by CJ Johnson <johnsoncj@google.com>: Internal change PiperOrigin-RevId: 354545297 -- 7ea02b3783f7f49ef97d86a8f6580a19cc57df14 by Abseil Team <absl-team@google.com>: Pre-allocate memory for vectors where the size is known. PiperOrigin-RevId: 354344576 -- 9246c7cb11f1d6444f79ebe25acc69a8a9b870e0 by Matt Kulukundis <kfm@google.com>: Add support for Elbrus 2000 (e2k) Import of https://github.com/abseil/abseil-cpp/pull/889 PiperOrigin-RevId: 354344013 -- 0fc93d359cc1fb307552e917b37b7b2e7eed822f by Abseil Team <absl-team@google.com>: Integrate CordRepRing logic into cord (but do not enable it) PiperOrigin-RevId: 354312238 -- eda05622f7da71466723acb33403f783529df24b by Abseil Team <absl-team@google.com>: Protect ignore diagnostic with "__has_warning". PiperOrigin-RevId: 354112334 -- 47716c5d8fb10efa4fdd801d28bac414c6f8ec32 by Abseil Team <absl-team@google.com>: Rearrange InlinedVector copy constructor and destructor to treat a few special cases inline and then tail-call a non-inlined routine for the rest. In particular, we optimize for empty vectors in both cases. Added a couple of benchmarks that copy either an InlVec<int64> or an InlVec<InlVec<int64>>. Speed difference: ``` BM_CopyTrivial/0 0.92ns +- 0% 0.47ns +- 0% -48.91% (p=0.000 n=11+12) BM_CopyTrivial/1 0.92ns +- 0% 1.15ns +- 0% +25.00% (p=0.000 n=10+9) BM_CopyTrivial/8 8.57ns +- 0% 10.72ns +- 1% +25.16% (p=0.000 n=10+12) BM_CopyNonTrivial/0 3.21ns +- 0% 0.70ns +- 0% -78.23% (p=0.000 n=12+10) BM_CopyNonTrivial/1 5.88ns +- 1% 5.51ns +- 0% -6.28% (p=0.000 n=10+8) BM_CopyNonTrivial/8 21.5ns +- 1% 15.2ns +- 2% -29.23% (p=0.000 n=12+12) ``` Note: the slowdowns are a few cycles which is expected given the procedure call added in that case. We decided this is a good tradeoff given the code size reductions and the more significant speedups for empty vectors. Size difference (as measured by nm): ``` BM_CopyTrivial from 1048 bytes to 326 bytes. BM_CopyNonTrivial from 749 bytes to 470 bytes. ``` Code size for a large binary drops by ~500KB (from 349415719 to 348906015 348906191). All of the benchmarks that showed a significant difference: Ones that improve with this CL: ``` BM_CopyNonTrivial/0 3.21ns +- 0% 0.70ns +- 0% -78.23% (p=0.000 n=12+10) BM_InlinedVectorFillString/0 0.93ns +- 0% 0.24ns +- 0% -74.19% (p=0.000 n=12+10) BM_InlinedVectorAssignments/1 10.5ns +- 0% 4.1ns +- 0% -60.64% (p=0.000 n=11+10) BM_InlinedVectorAssignments/2 10.7ns +- 0% 4.4ns +- 0% -59.08% (p=0.000 n=11+11) BM_CopyTrivial/0 0.92ns +- 0% 0.47ns +- 0% -48.91% (p=0.000 n=11+12) BM_CopyNonTrivial/8 21.5ns +- 1% 15.2ns +- 2% -29.23% (p=0.000 n=12+12) BM_StdVectorEmpty 0.47ns +- 1% 0.35ns +- 0% -24.73% (p=0.000 n=12+12) BM_StdVectorSize 0.46ns +- 2% 0.35ns +- 0% -24.32% (p=0.000 n=12+12) BM_SwapElements<LargeCopyableOnly>/0 3.44ns +- 0% 2.76ns +- 1% -19.83% (p=0.000 n=11+11) BM_InlinedVectorFillRange/256 20.7ns +- 1% 17.8ns +- 0% -14.08% (p=0.000 n=12+9) BM_CopyNonTrivial/1 5.88ns +- 1% 5.51ns +- 0% -6.28% (p=0.000 n=10+8) BM_SwapElements<LargeCopyableMovable>/1 4.19ns +- 0% 3.95ns +- 1% -5.63% (p=0.000 n=11+12) BM_SwapElements<LargeCopyableMovableSwappable>/1 4.18ns +- 0% 3.99ns +- 0% -4.70% (p=0.000 n=9+11) BM_SwapElements<LargeCopyableMovable>/0 2.41ns +- 0% 2.31ns +- 0% -4.45% (p=0.000 n=12+12) BM_InlinedVectorFillRange/64 8.25ns +- 0% 8.04ns +- 0% -2.51% (p=0.000 n=12+11) BM_SwapElements<LargeCopyableOnly>/1 82.4ns +- 0% 81.5ns +- 0% -1.06% (p=0.000 n=12+12) ``` Ones that get worse with this CL: ``` BM_CopyTrivial/1 0.92ns +- 0% 1.15ns +- 0% +25.00% (p=0.000 n=10+9) BM_CopyTrivial/8 8.57ns +- 0% 10.72ns +- 1% +25.16% (p=0.000 n=10+12) BM_SwapElements<LargeCopyableMovableSwappable>/512 1.48ns +- 1% 1.66ns +- 1% +11.88% (p=0.000 n=12+12) BM_InlinedVectorFillString/1 11.5ns +- 0% 12.8ns +- 1% +11.62% (p=0.000 n=12+11) BM_SwapElements<LargeCopyableMovableSwappable>/64 1.48ns +- 2% 1.66ns +- 1% +11.66% (p=0.000 n=12+11) BM_SwapElements<LargeCopyableMovableSwappable>/1k 1.48ns +- 1% 1.65ns +- 2% +11.32% (p=0.000 n=12+12) BM_SwapElements<LargeCopyableMovable>/512 1.48ns +- 2% 1.58ns +- 4% +6.62% (p=0.000 n=11+12) BM_SwapElements<LargeCopyableMovable>/1k 1.49ns +- 2% 1.58ns +- 3% +6.05% (p=0.000 n=12+12) BM_SwapElements<LargeCopyableMovable>/64 1.48ns +- 2% 1.57ns +- 4% +6.04% (p=0.000 n=11+12) BM_InlinedVectorFillRange/1 4.81ns +- 0% 5.05ns +- 0% +4.83% (p=0.000 n=11+11) BM_InlinedVectorFillString/8 79.4ns +- 1% 83.1ns +- 1% +4.64% (p=0.000 n=10+12) BM_StdVectorFillString/1 16.3ns +- 0% 16.6ns +- 0% +2.13% (p=0.000 n=11+8) ``` PiperOrigin-RevId: 353906786 -- 8e26518b3cec9c598e5e9573c46c3bd1b03a67ef by Abseil Team <absl-team@google.com>: Internal change PiperOrigin-RevId: 353737330 -- f206ae0983e58c9904ed8b8f05f9caf564a446be by Matt Kulukundis <kfm@google.com>: Import of CCTZ from GitHub. PiperOrigin-RevId: 353682256 GitOrigin-RevId: c68f1886f5e8fd90eb0c2d2e68feaf00a7cdacda Change-Id: I5790c1036c4f543c701d1039848fabf7ae881ad8
4 years ago
using ::absl::cord_internal::CordRepRing;
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
using ::absl::cord_internal::CordRepSubstring;
using ::absl::cord_internal::kMinFlatLength;
using ::absl::cord_internal::kMaxFlatLength;
using ::absl::cord_internal::CONCAT;
using ::absl::cord_internal::EXTERNAL;
using ::absl::cord_internal::FLAT;
using ::absl::cord_internal::RING;
using ::absl::cord_internal::SUBSTRING;
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
Export of internal Abseil changes -- 6e9f93888bbe6718997ed90bbd049f1f3572b066 by Abseil Team <absl-team@google.com>: Fix status to be safe for self move assignment. PiperOrigin-RevId: 346815392 -- 35cae74a977f258e81dfbe925fb5a34cb6632036 by Gennadiy Rozental <rogeeff@google.com>: Eliminate unnecessary access to the global vars. PiperOrigin-RevId: 346777168 -- e7e786c243069060f5d6c1c945decb4b0b83f95b by Andy Getzendanner <durandal@google.com>: Internal change. PiperOrigin-RevId: 346685656 -- 4ccd41c48f1a83cfa20b3ea534f743dd7d788376 by Abseil Team <absl-team@google.com>: Move CordRep Ref() and Unref() logic into cord_internal.h This change moves Ref() and Unref() logic out of cord.cc into cord_internal. The main purpose is to make upcoming ring buffer changes easier to isolate from existing cord.cc code. Notice that this removes the nullptr check from Unref() and now requires it to be non null (which held true most times). We may need to rethink if the 'unref unlikely one' is the common case: are cordreps most likely shared? This may be something between 'root' and non root nodes, i.e., is it more likely for leaf / flat nodes in large cords to be shared than top level cordreps being shared? Vice versa? Benchmarks say that we mostly shouldn't care, the caveat being that atomic ops seem more expensive on upcoming archs (arcadia) so we should error on the side of an extra IsOne() branch saving us single owned atomic ops. PiperOrigin-RevId: 346676121 -- f0babab103b9e60d61ba09482d468985e43eceb3 by Samuel Benzaquen <sbenza@google.com>: Fix iterator based constructor and `.insert` members to only require EmplaceConstructible as the standard specifies. PiperOrigin-RevId: 346616707 -- 8f48eedda02277f9c96a88ed7726e34b557cce20 by Evan Brown <ezb@google.com>: Fix a bug in binary_search_impl when there's a transparent, three-way comparator that has different equivalence classes for different lookup types. Add a new can_have_multiple_equivalent_keys method to share the common logic for these cases. PiperOrigin-RevId: 346605948 -- 649183cb3cc9383431de9c81fb1c0f885d4001ae by Abseil Team <absl-team@google.com>: Add benchmark for accessing a Duration flag. PiperOrigin-RevId: 346594843 -- fefdb046520871af63ce2229e2f7cccfc0483dea by Abseil Team <absl-team@google.com>: Restructure CordReader for upcoming ring buffer changes. PiperOrigin-RevId: 346410642 -- 8b2f50e7da0ebab06ead5f94e366e984ca23cb6a by Abseil Team <absl-team@google.com>: Wire in an internal-only flag to toggle upcoming ring buffer changes on/off for experimentation. PiperOrigin-RevId: 346199111 GitOrigin-RevId: 6e9f93888bbe6718997ed90bbd049f1f3572b066 Change-Id: I8f34866b25a79209cb5448bbb28dd3044111d2e9
4 years ago
using ::absl::cord_internal::kInlinedVectorSize;
using ::absl::cord_internal::kMaxBytesToCopy;
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
constexpr uint64_t Fibonacci(unsigned char n, uint64_t a = 0, uint64_t b = 1) {
return n == 0 ? a : Fibonacci(n - 1, b, a + b);
}
static_assert(Fibonacci(63) == 6557470319842,
"Fibonacci values computed incorrectly");
// Minimum length required for a given depth tree -- a tree is considered
// balanced if
// length(t) >= min_length[depth(t)]
// The root node depth is allowed to become twice as large to reduce rebalancing
// for larger strings (see IsRootBalanced).
static constexpr uint64_t min_length[] = {
Fibonacci(2), Fibonacci(3), Fibonacci(4), Fibonacci(5),
Fibonacci(6), Fibonacci(7), Fibonacci(8), Fibonacci(9),
Fibonacci(10), Fibonacci(11), Fibonacci(12), Fibonacci(13),
Fibonacci(14), Fibonacci(15), Fibonacci(16), Fibonacci(17),
Fibonacci(18), Fibonacci(19), Fibonacci(20), Fibonacci(21),
Fibonacci(22), Fibonacci(23), Fibonacci(24), Fibonacci(25),
Fibonacci(26), Fibonacci(27), Fibonacci(28), Fibonacci(29),
Fibonacci(30), Fibonacci(31), Fibonacci(32), Fibonacci(33),
Fibonacci(34), Fibonacci(35), Fibonacci(36), Fibonacci(37),
Fibonacci(38), Fibonacci(39), Fibonacci(40), Fibonacci(41),
Fibonacci(42), Fibonacci(43), Fibonacci(44), Fibonacci(45),
Fibonacci(46), Fibonacci(47),
0xffffffffffffffffull, // Avoid overflow
};
static const int kMinLengthSize = ABSL_ARRAYSIZE(min_length);
Export of internal Abseil changes -- c68f1886f5e8fd90eb0c2d2e68feaf00a7cdacda by CJ Johnson <johnsoncj@google.com>: Introduce absl::Cleanup to the OSS repo PiperOrigin-RevId: 354583156 -- 17030cf388e10f7eb959e3e566326d1072ce392e by Abseil Team <absl-team@google.com>: Internal change only PiperOrigin-RevId: 354574953 -- e979d7236d4f3252e79ddda6739b67a9a326bf6d by CJ Johnson <johnsoncj@google.com>: Internal change PiperOrigin-RevId: 354545297 -- 7ea02b3783f7f49ef97d86a8f6580a19cc57df14 by Abseil Team <absl-team@google.com>: Pre-allocate memory for vectors where the size is known. PiperOrigin-RevId: 354344576 -- 9246c7cb11f1d6444f79ebe25acc69a8a9b870e0 by Matt Kulukundis <kfm@google.com>: Add support for Elbrus 2000 (e2k) Import of https://github.com/abseil/abseil-cpp/pull/889 PiperOrigin-RevId: 354344013 -- 0fc93d359cc1fb307552e917b37b7b2e7eed822f by Abseil Team <absl-team@google.com>: Integrate CordRepRing logic into cord (but do not enable it) PiperOrigin-RevId: 354312238 -- eda05622f7da71466723acb33403f783529df24b by Abseil Team <absl-team@google.com>: Protect ignore diagnostic with "__has_warning". PiperOrigin-RevId: 354112334 -- 47716c5d8fb10efa4fdd801d28bac414c6f8ec32 by Abseil Team <absl-team@google.com>: Rearrange InlinedVector copy constructor and destructor to treat a few special cases inline and then tail-call a non-inlined routine for the rest. In particular, we optimize for empty vectors in both cases. Added a couple of benchmarks that copy either an InlVec<int64> or an InlVec<InlVec<int64>>. Speed difference: ``` BM_CopyTrivial/0 0.92ns +- 0% 0.47ns +- 0% -48.91% (p=0.000 n=11+12) BM_CopyTrivial/1 0.92ns +- 0% 1.15ns +- 0% +25.00% (p=0.000 n=10+9) BM_CopyTrivial/8 8.57ns +- 0% 10.72ns +- 1% +25.16% (p=0.000 n=10+12) BM_CopyNonTrivial/0 3.21ns +- 0% 0.70ns +- 0% -78.23% (p=0.000 n=12+10) BM_CopyNonTrivial/1 5.88ns +- 1% 5.51ns +- 0% -6.28% (p=0.000 n=10+8) BM_CopyNonTrivial/8 21.5ns +- 1% 15.2ns +- 2% -29.23% (p=0.000 n=12+12) ``` Note: the slowdowns are a few cycles which is expected given the procedure call added in that case. We decided this is a good tradeoff given the code size reductions and the more significant speedups for empty vectors. Size difference (as measured by nm): ``` BM_CopyTrivial from 1048 bytes to 326 bytes. BM_CopyNonTrivial from 749 bytes to 470 bytes. ``` Code size for a large binary drops by ~500KB (from 349415719 to 348906015 348906191). All of the benchmarks that showed a significant difference: Ones that improve with this CL: ``` BM_CopyNonTrivial/0 3.21ns +- 0% 0.70ns +- 0% -78.23% (p=0.000 n=12+10) BM_InlinedVectorFillString/0 0.93ns +- 0% 0.24ns +- 0% -74.19% (p=0.000 n=12+10) BM_InlinedVectorAssignments/1 10.5ns +- 0% 4.1ns +- 0% -60.64% (p=0.000 n=11+10) BM_InlinedVectorAssignments/2 10.7ns +- 0% 4.4ns +- 0% -59.08% (p=0.000 n=11+11) BM_CopyTrivial/0 0.92ns +- 0% 0.47ns +- 0% -48.91% (p=0.000 n=11+12) BM_CopyNonTrivial/8 21.5ns +- 1% 15.2ns +- 2% -29.23% (p=0.000 n=12+12) BM_StdVectorEmpty 0.47ns +- 1% 0.35ns +- 0% -24.73% (p=0.000 n=12+12) BM_StdVectorSize 0.46ns +- 2% 0.35ns +- 0% -24.32% (p=0.000 n=12+12) BM_SwapElements<LargeCopyableOnly>/0 3.44ns +- 0% 2.76ns +- 1% -19.83% (p=0.000 n=11+11) BM_InlinedVectorFillRange/256 20.7ns +- 1% 17.8ns +- 0% -14.08% (p=0.000 n=12+9) BM_CopyNonTrivial/1 5.88ns +- 1% 5.51ns +- 0% -6.28% (p=0.000 n=10+8) BM_SwapElements<LargeCopyableMovable>/1 4.19ns +- 0% 3.95ns +- 1% -5.63% (p=0.000 n=11+12) BM_SwapElements<LargeCopyableMovableSwappable>/1 4.18ns +- 0% 3.99ns +- 0% -4.70% (p=0.000 n=9+11) BM_SwapElements<LargeCopyableMovable>/0 2.41ns +- 0% 2.31ns +- 0% -4.45% (p=0.000 n=12+12) BM_InlinedVectorFillRange/64 8.25ns +- 0% 8.04ns +- 0% -2.51% (p=0.000 n=12+11) BM_SwapElements<LargeCopyableOnly>/1 82.4ns +- 0% 81.5ns +- 0% -1.06% (p=0.000 n=12+12) ``` Ones that get worse with this CL: ``` BM_CopyTrivial/1 0.92ns +- 0% 1.15ns +- 0% +25.00% (p=0.000 n=10+9) BM_CopyTrivial/8 8.57ns +- 0% 10.72ns +- 1% +25.16% (p=0.000 n=10+12) BM_SwapElements<LargeCopyableMovableSwappable>/512 1.48ns +- 1% 1.66ns +- 1% +11.88% (p=0.000 n=12+12) BM_InlinedVectorFillString/1 11.5ns +- 0% 12.8ns +- 1% +11.62% (p=0.000 n=12+11) BM_SwapElements<LargeCopyableMovableSwappable>/64 1.48ns +- 2% 1.66ns +- 1% +11.66% (p=0.000 n=12+11) BM_SwapElements<LargeCopyableMovableSwappable>/1k 1.48ns +- 1% 1.65ns +- 2% +11.32% (p=0.000 n=12+12) BM_SwapElements<LargeCopyableMovable>/512 1.48ns +- 2% 1.58ns +- 4% +6.62% (p=0.000 n=11+12) BM_SwapElements<LargeCopyableMovable>/1k 1.49ns +- 2% 1.58ns +- 3% +6.05% (p=0.000 n=12+12) BM_SwapElements<LargeCopyableMovable>/64 1.48ns +- 2% 1.57ns +- 4% +6.04% (p=0.000 n=11+12) BM_InlinedVectorFillRange/1 4.81ns +- 0% 5.05ns +- 0% +4.83% (p=0.000 n=11+11) BM_InlinedVectorFillString/8 79.4ns +- 1% 83.1ns +- 1% +4.64% (p=0.000 n=10+12) BM_StdVectorFillString/1 16.3ns +- 0% 16.6ns +- 0% +2.13% (p=0.000 n=11+8) ``` PiperOrigin-RevId: 353906786 -- 8e26518b3cec9c598e5e9573c46c3bd1b03a67ef by Abseil Team <absl-team@google.com>: Internal change PiperOrigin-RevId: 353737330 -- f206ae0983e58c9904ed8b8f05f9caf564a446be by Matt Kulukundis <kfm@google.com>: Import of CCTZ from GitHub. PiperOrigin-RevId: 353682256 GitOrigin-RevId: c68f1886f5e8fd90eb0c2d2e68feaf00a7cdacda Change-Id: I5790c1036c4f543c701d1039848fabf7ae881ad8
4 years ago
static inline bool cord_ring_enabled() {
return cord_internal::cord_ring_buffer_enabled.load(
std::memory_order_relaxed);
}
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
static inline bool IsRootBalanced(CordRep* node) {
if (node->tag != CONCAT) {
return true;
} else if (node->concat()->depth() <= 15) {
return true;
} else if (node->concat()->depth() > kMinLengthSize) {
return false;
} else {
// Allow depth to become twice as large as implied by fibonacci rule to
// reduce rebalancing for larger strings.
return (node->length >= min_length[node->concat()->depth() / 2]);
}
}
static CordRep* Rebalance(CordRep* node);
Export of internal Abseil changes -- c68f1886f5e8fd90eb0c2d2e68feaf00a7cdacda by CJ Johnson <johnsoncj@google.com>: Introduce absl::Cleanup to the OSS repo PiperOrigin-RevId: 354583156 -- 17030cf388e10f7eb959e3e566326d1072ce392e by Abseil Team <absl-team@google.com>: Internal change only PiperOrigin-RevId: 354574953 -- e979d7236d4f3252e79ddda6739b67a9a326bf6d by CJ Johnson <johnsoncj@google.com>: Internal change PiperOrigin-RevId: 354545297 -- 7ea02b3783f7f49ef97d86a8f6580a19cc57df14 by Abseil Team <absl-team@google.com>: Pre-allocate memory for vectors where the size is known. PiperOrigin-RevId: 354344576 -- 9246c7cb11f1d6444f79ebe25acc69a8a9b870e0 by Matt Kulukundis <kfm@google.com>: Add support for Elbrus 2000 (e2k) Import of https://github.com/abseil/abseil-cpp/pull/889 PiperOrigin-RevId: 354344013 -- 0fc93d359cc1fb307552e917b37b7b2e7eed822f by Abseil Team <absl-team@google.com>: Integrate CordRepRing logic into cord (but do not enable it) PiperOrigin-RevId: 354312238 -- eda05622f7da71466723acb33403f783529df24b by Abseil Team <absl-team@google.com>: Protect ignore diagnostic with "__has_warning". PiperOrigin-RevId: 354112334 -- 47716c5d8fb10efa4fdd801d28bac414c6f8ec32 by Abseil Team <absl-team@google.com>: Rearrange InlinedVector copy constructor and destructor to treat a few special cases inline and then tail-call a non-inlined routine for the rest. In particular, we optimize for empty vectors in both cases. Added a couple of benchmarks that copy either an InlVec<int64> or an InlVec<InlVec<int64>>. Speed difference: ``` BM_CopyTrivial/0 0.92ns +- 0% 0.47ns +- 0% -48.91% (p=0.000 n=11+12) BM_CopyTrivial/1 0.92ns +- 0% 1.15ns +- 0% +25.00% (p=0.000 n=10+9) BM_CopyTrivial/8 8.57ns +- 0% 10.72ns +- 1% +25.16% (p=0.000 n=10+12) BM_CopyNonTrivial/0 3.21ns +- 0% 0.70ns +- 0% -78.23% (p=0.000 n=12+10) BM_CopyNonTrivial/1 5.88ns +- 1% 5.51ns +- 0% -6.28% (p=0.000 n=10+8) BM_CopyNonTrivial/8 21.5ns +- 1% 15.2ns +- 2% -29.23% (p=0.000 n=12+12) ``` Note: the slowdowns are a few cycles which is expected given the procedure call added in that case. We decided this is a good tradeoff given the code size reductions and the more significant speedups for empty vectors. Size difference (as measured by nm): ``` BM_CopyTrivial from 1048 bytes to 326 bytes. BM_CopyNonTrivial from 749 bytes to 470 bytes. ``` Code size for a large binary drops by ~500KB (from 349415719 to 348906015 348906191). All of the benchmarks that showed a significant difference: Ones that improve with this CL: ``` BM_CopyNonTrivial/0 3.21ns +- 0% 0.70ns +- 0% -78.23% (p=0.000 n=12+10) BM_InlinedVectorFillString/0 0.93ns +- 0% 0.24ns +- 0% -74.19% (p=0.000 n=12+10) BM_InlinedVectorAssignments/1 10.5ns +- 0% 4.1ns +- 0% -60.64% (p=0.000 n=11+10) BM_InlinedVectorAssignments/2 10.7ns +- 0% 4.4ns +- 0% -59.08% (p=0.000 n=11+11) BM_CopyTrivial/0 0.92ns +- 0% 0.47ns +- 0% -48.91% (p=0.000 n=11+12) BM_CopyNonTrivial/8 21.5ns +- 1% 15.2ns +- 2% -29.23% (p=0.000 n=12+12) BM_StdVectorEmpty 0.47ns +- 1% 0.35ns +- 0% -24.73% (p=0.000 n=12+12) BM_StdVectorSize 0.46ns +- 2% 0.35ns +- 0% -24.32% (p=0.000 n=12+12) BM_SwapElements<LargeCopyableOnly>/0 3.44ns +- 0% 2.76ns +- 1% -19.83% (p=0.000 n=11+11) BM_InlinedVectorFillRange/256 20.7ns +- 1% 17.8ns +- 0% -14.08% (p=0.000 n=12+9) BM_CopyNonTrivial/1 5.88ns +- 1% 5.51ns +- 0% -6.28% (p=0.000 n=10+8) BM_SwapElements<LargeCopyableMovable>/1 4.19ns +- 0% 3.95ns +- 1% -5.63% (p=0.000 n=11+12) BM_SwapElements<LargeCopyableMovableSwappable>/1 4.18ns +- 0% 3.99ns +- 0% -4.70% (p=0.000 n=9+11) BM_SwapElements<LargeCopyableMovable>/0 2.41ns +- 0% 2.31ns +- 0% -4.45% (p=0.000 n=12+12) BM_InlinedVectorFillRange/64 8.25ns +- 0% 8.04ns +- 0% -2.51% (p=0.000 n=12+11) BM_SwapElements<LargeCopyableOnly>/1 82.4ns +- 0% 81.5ns +- 0% -1.06% (p=0.000 n=12+12) ``` Ones that get worse with this CL: ``` BM_CopyTrivial/1 0.92ns +- 0% 1.15ns +- 0% +25.00% (p=0.000 n=10+9) BM_CopyTrivial/8 8.57ns +- 0% 10.72ns +- 1% +25.16% (p=0.000 n=10+12) BM_SwapElements<LargeCopyableMovableSwappable>/512 1.48ns +- 1% 1.66ns +- 1% +11.88% (p=0.000 n=12+12) BM_InlinedVectorFillString/1 11.5ns +- 0% 12.8ns +- 1% +11.62% (p=0.000 n=12+11) BM_SwapElements<LargeCopyableMovableSwappable>/64 1.48ns +- 2% 1.66ns +- 1% +11.66% (p=0.000 n=12+11) BM_SwapElements<LargeCopyableMovableSwappable>/1k 1.48ns +- 1% 1.65ns +- 2% +11.32% (p=0.000 n=12+12) BM_SwapElements<LargeCopyableMovable>/512 1.48ns +- 2% 1.58ns +- 4% +6.62% (p=0.000 n=11+12) BM_SwapElements<LargeCopyableMovable>/1k 1.49ns +- 2% 1.58ns +- 3% +6.05% (p=0.000 n=12+12) BM_SwapElements<LargeCopyableMovable>/64 1.48ns +- 2% 1.57ns +- 4% +6.04% (p=0.000 n=11+12) BM_InlinedVectorFillRange/1 4.81ns +- 0% 5.05ns +- 0% +4.83% (p=0.000 n=11+11) BM_InlinedVectorFillString/8 79.4ns +- 1% 83.1ns +- 1% +4.64% (p=0.000 n=10+12) BM_StdVectorFillString/1 16.3ns +- 0% 16.6ns +- 0% +2.13% (p=0.000 n=11+8) ``` PiperOrigin-RevId: 353906786 -- 8e26518b3cec9c598e5e9573c46c3bd1b03a67ef by Abseil Team <absl-team@google.com>: Internal change PiperOrigin-RevId: 353737330 -- f206ae0983e58c9904ed8b8f05f9caf564a446be by Matt Kulukundis <kfm@google.com>: Import of CCTZ from GitHub. PiperOrigin-RevId: 353682256 GitOrigin-RevId: c68f1886f5e8fd90eb0c2d2e68feaf00a7cdacda Change-Id: I5790c1036c4f543c701d1039848fabf7ae881ad8
4 years ago
static void DumpNode(CordRep* rep, bool include_data, std::ostream* os,
int indent = 0);
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
static bool VerifyNode(CordRep* root, CordRep* start_node,
bool full_validation);
static inline CordRep* VerifyTree(CordRep* node) {
// Verification is expensive, so only do it in debug mode.
// Even in debug mode we normally do only light validation.
// If you are debugging Cord itself, you should define the
// macro EXTRA_CORD_VALIDATION, e.g. by adding
// --copt=-DEXTRA_CORD_VALIDATION to the blaze line.
#ifdef EXTRA_CORD_VALIDATION
assert(node == nullptr || VerifyNode(node, node, /*full_validation=*/true));
#else // EXTRA_CORD_VALIDATION
assert(node == nullptr || VerifyNode(node, node, /*full_validation=*/false));
#endif // EXTRA_CORD_VALIDATION
static_cast<void>(&VerifyNode);
return node;
}
// Return the depth of a node
static int Depth(const CordRep* rep) {
if (rep->tag == CONCAT) {
return rep->concat()->depth();
} else {
return 0;
}
}
static void SetConcatChildren(CordRepConcat* concat, CordRep* left,
CordRep* right) {
concat->left = left;
concat->right = right;
concat->length = left->length + right->length;
concat->set_depth(1 + std::max(Depth(left), Depth(right)));
}
// Create a concatenation of the specified nodes.
// Does not change the refcounts of "left" and "right".
// The returned node has a refcount of 1.
static CordRep* RawConcat(CordRep* left, CordRep* right) {
// Avoid making degenerate concat nodes (one child is empty)
Export of internal Abseil changes -- 6e9f93888bbe6718997ed90bbd049f1f3572b066 by Abseil Team <absl-team@google.com>: Fix status to be safe for self move assignment. PiperOrigin-RevId: 346815392 -- 35cae74a977f258e81dfbe925fb5a34cb6632036 by Gennadiy Rozental <rogeeff@google.com>: Eliminate unnecessary access to the global vars. PiperOrigin-RevId: 346777168 -- e7e786c243069060f5d6c1c945decb4b0b83f95b by Andy Getzendanner <durandal@google.com>: Internal change. PiperOrigin-RevId: 346685656 -- 4ccd41c48f1a83cfa20b3ea534f743dd7d788376 by Abseil Team <absl-team@google.com>: Move CordRep Ref() and Unref() logic into cord_internal.h This change moves Ref() and Unref() logic out of cord.cc into cord_internal. The main purpose is to make upcoming ring buffer changes easier to isolate from existing cord.cc code. Notice that this removes the nullptr check from Unref() and now requires it to be non null (which held true most times). We may need to rethink if the 'unref unlikely one' is the common case: are cordreps most likely shared? This may be something between 'root' and non root nodes, i.e., is it more likely for leaf / flat nodes in large cords to be shared than top level cordreps being shared? Vice versa? Benchmarks say that we mostly shouldn't care, the caveat being that atomic ops seem more expensive on upcoming archs (arcadia) so we should error on the side of an extra IsOne() branch saving us single owned atomic ops. PiperOrigin-RevId: 346676121 -- f0babab103b9e60d61ba09482d468985e43eceb3 by Samuel Benzaquen <sbenza@google.com>: Fix iterator based constructor and `.insert` members to only require EmplaceConstructible as the standard specifies. PiperOrigin-RevId: 346616707 -- 8f48eedda02277f9c96a88ed7726e34b557cce20 by Evan Brown <ezb@google.com>: Fix a bug in binary_search_impl when there's a transparent, three-way comparator that has different equivalence classes for different lookup types. Add a new can_have_multiple_equivalent_keys method to share the common logic for these cases. PiperOrigin-RevId: 346605948 -- 649183cb3cc9383431de9c81fb1c0f885d4001ae by Abseil Team <absl-team@google.com>: Add benchmark for accessing a Duration flag. PiperOrigin-RevId: 346594843 -- fefdb046520871af63ce2229e2f7cccfc0483dea by Abseil Team <absl-team@google.com>: Restructure CordReader for upcoming ring buffer changes. PiperOrigin-RevId: 346410642 -- 8b2f50e7da0ebab06ead5f94e366e984ca23cb6a by Abseil Team <absl-team@google.com>: Wire in an internal-only flag to toggle upcoming ring buffer changes on/off for experimentation. PiperOrigin-RevId: 346199111 GitOrigin-RevId: 6e9f93888bbe6718997ed90bbd049f1f3572b066 Change-Id: I8f34866b25a79209cb5448bbb28dd3044111d2e9
4 years ago
if (left == nullptr) return right;
if (right == nullptr) return left;
if (left->length == 0) {
CordRep::Unref(left);
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
return right;
}
Export of internal Abseil changes -- 6e9f93888bbe6718997ed90bbd049f1f3572b066 by Abseil Team <absl-team@google.com>: Fix status to be safe for self move assignment. PiperOrigin-RevId: 346815392 -- 35cae74a977f258e81dfbe925fb5a34cb6632036 by Gennadiy Rozental <rogeeff@google.com>: Eliminate unnecessary access to the global vars. PiperOrigin-RevId: 346777168 -- e7e786c243069060f5d6c1c945decb4b0b83f95b by Andy Getzendanner <durandal@google.com>: Internal change. PiperOrigin-RevId: 346685656 -- 4ccd41c48f1a83cfa20b3ea534f743dd7d788376 by Abseil Team <absl-team@google.com>: Move CordRep Ref() and Unref() logic into cord_internal.h This change moves Ref() and Unref() logic out of cord.cc into cord_internal. The main purpose is to make upcoming ring buffer changes easier to isolate from existing cord.cc code. Notice that this removes the nullptr check from Unref() and now requires it to be non null (which held true most times). We may need to rethink if the 'unref unlikely one' is the common case: are cordreps most likely shared? This may be something between 'root' and non root nodes, i.e., is it more likely for leaf / flat nodes in large cords to be shared than top level cordreps being shared? Vice versa? Benchmarks say that we mostly shouldn't care, the caveat being that atomic ops seem more expensive on upcoming archs (arcadia) so we should error on the side of an extra IsOne() branch saving us single owned atomic ops. PiperOrigin-RevId: 346676121 -- f0babab103b9e60d61ba09482d468985e43eceb3 by Samuel Benzaquen <sbenza@google.com>: Fix iterator based constructor and `.insert` members to only require EmplaceConstructible as the standard specifies. PiperOrigin-RevId: 346616707 -- 8f48eedda02277f9c96a88ed7726e34b557cce20 by Evan Brown <ezb@google.com>: Fix a bug in binary_search_impl when there's a transparent, three-way comparator that has different equivalence classes for different lookup types. Add a new can_have_multiple_equivalent_keys method to share the common logic for these cases. PiperOrigin-RevId: 346605948 -- 649183cb3cc9383431de9c81fb1c0f885d4001ae by Abseil Team <absl-team@google.com>: Add benchmark for accessing a Duration flag. PiperOrigin-RevId: 346594843 -- fefdb046520871af63ce2229e2f7cccfc0483dea by Abseil Team <absl-team@google.com>: Restructure CordReader for upcoming ring buffer changes. PiperOrigin-RevId: 346410642 -- 8b2f50e7da0ebab06ead5f94e366e984ca23cb6a by Abseil Team <absl-team@google.com>: Wire in an internal-only flag to toggle upcoming ring buffer changes on/off for experimentation. PiperOrigin-RevId: 346199111 GitOrigin-RevId: 6e9f93888bbe6718997ed90bbd049f1f3572b066 Change-Id: I8f34866b25a79209cb5448bbb28dd3044111d2e9
4 years ago
if (right->length == 0) {
CordRep::Unref(right);
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
return left;
}
CordRepConcat* rep = new CordRepConcat();
rep->tag = CONCAT;
SetConcatChildren(rep, left, right);
return rep;
}
static CordRep* Concat(CordRep* left, CordRep* right) {
CordRep* rep = RawConcat(left, right);
if (rep != nullptr && !IsRootBalanced(rep)) {
rep = Rebalance(rep);
}
return VerifyTree(rep);
}
// Make a balanced tree out of an array of leaf nodes.
static CordRep* MakeBalancedTree(CordRep** reps, size_t n) {
// Make repeated passes over the array, merging adjacent pairs
// until we are left with just a single node.
while (n > 1) {
size_t dst = 0;
for (size_t src = 0; src < n; src += 2) {
if (src + 1 < n) {
reps[dst] = Concat(reps[src], reps[src + 1]);
} else {
reps[dst] = reps[src];
}
dst++;
}
n = dst;
}
return reps[0];
}
Export of internal Abseil changes -- c68f1886f5e8fd90eb0c2d2e68feaf00a7cdacda by CJ Johnson <johnsoncj@google.com>: Introduce absl::Cleanup to the OSS repo PiperOrigin-RevId: 354583156 -- 17030cf388e10f7eb959e3e566326d1072ce392e by Abseil Team <absl-team@google.com>: Internal change only PiperOrigin-RevId: 354574953 -- e979d7236d4f3252e79ddda6739b67a9a326bf6d by CJ Johnson <johnsoncj@google.com>: Internal change PiperOrigin-RevId: 354545297 -- 7ea02b3783f7f49ef97d86a8f6580a19cc57df14 by Abseil Team <absl-team@google.com>: Pre-allocate memory for vectors where the size is known. PiperOrigin-RevId: 354344576 -- 9246c7cb11f1d6444f79ebe25acc69a8a9b870e0 by Matt Kulukundis <kfm@google.com>: Add support for Elbrus 2000 (e2k) Import of https://github.com/abseil/abseil-cpp/pull/889 PiperOrigin-RevId: 354344013 -- 0fc93d359cc1fb307552e917b37b7b2e7eed822f by Abseil Team <absl-team@google.com>: Integrate CordRepRing logic into cord (but do not enable it) PiperOrigin-RevId: 354312238 -- eda05622f7da71466723acb33403f783529df24b by Abseil Team <absl-team@google.com>: Protect ignore diagnostic with "__has_warning". PiperOrigin-RevId: 354112334 -- 47716c5d8fb10efa4fdd801d28bac414c6f8ec32 by Abseil Team <absl-team@google.com>: Rearrange InlinedVector copy constructor and destructor to treat a few special cases inline and then tail-call a non-inlined routine for the rest. In particular, we optimize for empty vectors in both cases. Added a couple of benchmarks that copy either an InlVec<int64> or an InlVec<InlVec<int64>>. Speed difference: ``` BM_CopyTrivial/0 0.92ns +- 0% 0.47ns +- 0% -48.91% (p=0.000 n=11+12) BM_CopyTrivial/1 0.92ns +- 0% 1.15ns +- 0% +25.00% (p=0.000 n=10+9) BM_CopyTrivial/8 8.57ns +- 0% 10.72ns +- 1% +25.16% (p=0.000 n=10+12) BM_CopyNonTrivial/0 3.21ns +- 0% 0.70ns +- 0% -78.23% (p=0.000 n=12+10) BM_CopyNonTrivial/1 5.88ns +- 1% 5.51ns +- 0% -6.28% (p=0.000 n=10+8) BM_CopyNonTrivial/8 21.5ns +- 1% 15.2ns +- 2% -29.23% (p=0.000 n=12+12) ``` Note: the slowdowns are a few cycles which is expected given the procedure call added in that case. We decided this is a good tradeoff given the code size reductions and the more significant speedups for empty vectors. Size difference (as measured by nm): ``` BM_CopyTrivial from 1048 bytes to 326 bytes. BM_CopyNonTrivial from 749 bytes to 470 bytes. ``` Code size for a large binary drops by ~500KB (from 349415719 to 348906015 348906191). All of the benchmarks that showed a significant difference: Ones that improve with this CL: ``` BM_CopyNonTrivial/0 3.21ns +- 0% 0.70ns +- 0% -78.23% (p=0.000 n=12+10) BM_InlinedVectorFillString/0 0.93ns +- 0% 0.24ns +- 0% -74.19% (p=0.000 n=12+10) BM_InlinedVectorAssignments/1 10.5ns +- 0% 4.1ns +- 0% -60.64% (p=0.000 n=11+10) BM_InlinedVectorAssignments/2 10.7ns +- 0% 4.4ns +- 0% -59.08% (p=0.000 n=11+11) BM_CopyTrivial/0 0.92ns +- 0% 0.47ns +- 0% -48.91% (p=0.000 n=11+12) BM_CopyNonTrivial/8 21.5ns +- 1% 15.2ns +- 2% -29.23% (p=0.000 n=12+12) BM_StdVectorEmpty 0.47ns +- 1% 0.35ns +- 0% -24.73% (p=0.000 n=12+12) BM_StdVectorSize 0.46ns +- 2% 0.35ns +- 0% -24.32% (p=0.000 n=12+12) BM_SwapElements<LargeCopyableOnly>/0 3.44ns +- 0% 2.76ns +- 1% -19.83% (p=0.000 n=11+11) BM_InlinedVectorFillRange/256 20.7ns +- 1% 17.8ns +- 0% -14.08% (p=0.000 n=12+9) BM_CopyNonTrivial/1 5.88ns +- 1% 5.51ns +- 0% -6.28% (p=0.000 n=10+8) BM_SwapElements<LargeCopyableMovable>/1 4.19ns +- 0% 3.95ns +- 1% -5.63% (p=0.000 n=11+12) BM_SwapElements<LargeCopyableMovableSwappable>/1 4.18ns +- 0% 3.99ns +- 0% -4.70% (p=0.000 n=9+11) BM_SwapElements<LargeCopyableMovable>/0 2.41ns +- 0% 2.31ns +- 0% -4.45% (p=0.000 n=12+12) BM_InlinedVectorFillRange/64 8.25ns +- 0% 8.04ns +- 0% -2.51% (p=0.000 n=12+11) BM_SwapElements<LargeCopyableOnly>/1 82.4ns +- 0% 81.5ns +- 0% -1.06% (p=0.000 n=12+12) ``` Ones that get worse with this CL: ``` BM_CopyTrivial/1 0.92ns +- 0% 1.15ns +- 0% +25.00% (p=0.000 n=10+9) BM_CopyTrivial/8 8.57ns +- 0% 10.72ns +- 1% +25.16% (p=0.000 n=10+12) BM_SwapElements<LargeCopyableMovableSwappable>/512 1.48ns +- 1% 1.66ns +- 1% +11.88% (p=0.000 n=12+12) BM_InlinedVectorFillString/1 11.5ns +- 0% 12.8ns +- 1% +11.62% (p=0.000 n=12+11) BM_SwapElements<LargeCopyableMovableSwappable>/64 1.48ns +- 2% 1.66ns +- 1% +11.66% (p=0.000 n=12+11) BM_SwapElements<LargeCopyableMovableSwappable>/1k 1.48ns +- 1% 1.65ns +- 2% +11.32% (p=0.000 n=12+12) BM_SwapElements<LargeCopyableMovable>/512 1.48ns +- 2% 1.58ns +- 4% +6.62% (p=0.000 n=11+12) BM_SwapElements<LargeCopyableMovable>/1k 1.49ns +- 2% 1.58ns +- 3% +6.05% (p=0.000 n=12+12) BM_SwapElements<LargeCopyableMovable>/64 1.48ns +- 2% 1.57ns +- 4% +6.04% (p=0.000 n=11+12) BM_InlinedVectorFillRange/1 4.81ns +- 0% 5.05ns +- 0% +4.83% (p=0.000 n=11+11) BM_InlinedVectorFillString/8 79.4ns +- 1% 83.1ns +- 1% +4.64% (p=0.000 n=10+12) BM_StdVectorFillString/1 16.3ns +- 0% 16.6ns +- 0% +2.13% (p=0.000 n=11+8) ``` PiperOrigin-RevId: 353906786 -- 8e26518b3cec9c598e5e9573c46c3bd1b03a67ef by Abseil Team <absl-team@google.com>: Internal change PiperOrigin-RevId: 353737330 -- f206ae0983e58c9904ed8b8f05f9caf564a446be by Matt Kulukundis <kfm@google.com>: Import of CCTZ from GitHub. PiperOrigin-RevId: 353682256 GitOrigin-RevId: c68f1886f5e8fd90eb0c2d2e68feaf00a7cdacda Change-Id: I5790c1036c4f543c701d1039848fabf7ae881ad8
4 years ago
static CordRepFlat* CreateFlat(const char* data, size_t length,
size_t alloc_hint) {
CordRepFlat* flat = CordRepFlat::New(length + alloc_hint);
flat->length = length;
memcpy(flat->Data(), data, length);
return flat;
}
// Creates a new flat or ringbuffer out of the specified array.
// The returned node has a refcount of 1.
static CordRep* RingNewTree(const char* data, size_t length,
size_t alloc_hint) {
if (length <= kMaxFlatLength) {
return CreateFlat(data, length, alloc_hint);
}
CordRepFlat* flat = CreateFlat(data, kMaxFlatLength, 0);
data += kMaxFlatLength;
length -= kMaxFlatLength;
size_t extra = (length - 1) / kMaxFlatLength + 1;
auto* root = CordRepRing::Create(flat, extra);
return CordRepRing::Append(root, {data, length}, alloc_hint);
}
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
// Create a new tree out of the specified array.
// The returned node has a refcount of 1.
static CordRep* NewTree(const char* data,
size_t length,
size_t alloc_hint) {
if (length == 0) return nullptr;
Export of internal Abseil changes -- c68f1886f5e8fd90eb0c2d2e68feaf00a7cdacda by CJ Johnson <johnsoncj@google.com>: Introduce absl::Cleanup to the OSS repo PiperOrigin-RevId: 354583156 -- 17030cf388e10f7eb959e3e566326d1072ce392e by Abseil Team <absl-team@google.com>: Internal change only PiperOrigin-RevId: 354574953 -- e979d7236d4f3252e79ddda6739b67a9a326bf6d by CJ Johnson <johnsoncj@google.com>: Internal change PiperOrigin-RevId: 354545297 -- 7ea02b3783f7f49ef97d86a8f6580a19cc57df14 by Abseil Team <absl-team@google.com>: Pre-allocate memory for vectors where the size is known. PiperOrigin-RevId: 354344576 -- 9246c7cb11f1d6444f79ebe25acc69a8a9b870e0 by Matt Kulukundis <kfm@google.com>: Add support for Elbrus 2000 (e2k) Import of https://github.com/abseil/abseil-cpp/pull/889 PiperOrigin-RevId: 354344013 -- 0fc93d359cc1fb307552e917b37b7b2e7eed822f by Abseil Team <absl-team@google.com>: Integrate CordRepRing logic into cord (but do not enable it) PiperOrigin-RevId: 354312238 -- eda05622f7da71466723acb33403f783529df24b by Abseil Team <absl-team@google.com>: Protect ignore diagnostic with "__has_warning". PiperOrigin-RevId: 354112334 -- 47716c5d8fb10efa4fdd801d28bac414c6f8ec32 by Abseil Team <absl-team@google.com>: Rearrange InlinedVector copy constructor and destructor to treat a few special cases inline and then tail-call a non-inlined routine for the rest. In particular, we optimize for empty vectors in both cases. Added a couple of benchmarks that copy either an InlVec<int64> or an InlVec<InlVec<int64>>. Speed difference: ``` BM_CopyTrivial/0 0.92ns +- 0% 0.47ns +- 0% -48.91% (p=0.000 n=11+12) BM_CopyTrivial/1 0.92ns +- 0% 1.15ns +- 0% +25.00% (p=0.000 n=10+9) BM_CopyTrivial/8 8.57ns +- 0% 10.72ns +- 1% +25.16% (p=0.000 n=10+12) BM_CopyNonTrivial/0 3.21ns +- 0% 0.70ns +- 0% -78.23% (p=0.000 n=12+10) BM_CopyNonTrivial/1 5.88ns +- 1% 5.51ns +- 0% -6.28% (p=0.000 n=10+8) BM_CopyNonTrivial/8 21.5ns +- 1% 15.2ns +- 2% -29.23% (p=0.000 n=12+12) ``` Note: the slowdowns are a few cycles which is expected given the procedure call added in that case. We decided this is a good tradeoff given the code size reductions and the more significant speedups for empty vectors. Size difference (as measured by nm): ``` BM_CopyTrivial from 1048 bytes to 326 bytes. BM_CopyNonTrivial from 749 bytes to 470 bytes. ``` Code size for a large binary drops by ~500KB (from 349415719 to 348906015 348906191). All of the benchmarks that showed a significant difference: Ones that improve with this CL: ``` BM_CopyNonTrivial/0 3.21ns +- 0% 0.70ns +- 0% -78.23% (p=0.000 n=12+10) BM_InlinedVectorFillString/0 0.93ns +- 0% 0.24ns +- 0% -74.19% (p=0.000 n=12+10) BM_InlinedVectorAssignments/1 10.5ns +- 0% 4.1ns +- 0% -60.64% (p=0.000 n=11+10) BM_InlinedVectorAssignments/2 10.7ns +- 0% 4.4ns +- 0% -59.08% (p=0.000 n=11+11) BM_CopyTrivial/0 0.92ns +- 0% 0.47ns +- 0% -48.91% (p=0.000 n=11+12) BM_CopyNonTrivial/8 21.5ns +- 1% 15.2ns +- 2% -29.23% (p=0.000 n=12+12) BM_StdVectorEmpty 0.47ns +- 1% 0.35ns +- 0% -24.73% (p=0.000 n=12+12) BM_StdVectorSize 0.46ns +- 2% 0.35ns +- 0% -24.32% (p=0.000 n=12+12) BM_SwapElements<LargeCopyableOnly>/0 3.44ns +- 0% 2.76ns +- 1% -19.83% (p=0.000 n=11+11) BM_InlinedVectorFillRange/256 20.7ns +- 1% 17.8ns +- 0% -14.08% (p=0.000 n=12+9) BM_CopyNonTrivial/1 5.88ns +- 1% 5.51ns +- 0% -6.28% (p=0.000 n=10+8) BM_SwapElements<LargeCopyableMovable>/1 4.19ns +- 0% 3.95ns +- 1% -5.63% (p=0.000 n=11+12) BM_SwapElements<LargeCopyableMovableSwappable>/1 4.18ns +- 0% 3.99ns +- 0% -4.70% (p=0.000 n=9+11) BM_SwapElements<LargeCopyableMovable>/0 2.41ns +- 0% 2.31ns +- 0% -4.45% (p=0.000 n=12+12) BM_InlinedVectorFillRange/64 8.25ns +- 0% 8.04ns +- 0% -2.51% (p=0.000 n=12+11) BM_SwapElements<LargeCopyableOnly>/1 82.4ns +- 0% 81.5ns +- 0% -1.06% (p=0.000 n=12+12) ``` Ones that get worse with this CL: ``` BM_CopyTrivial/1 0.92ns +- 0% 1.15ns +- 0% +25.00% (p=0.000 n=10+9) BM_CopyTrivial/8 8.57ns +- 0% 10.72ns +- 1% +25.16% (p=0.000 n=10+12) BM_SwapElements<LargeCopyableMovableSwappable>/512 1.48ns +- 1% 1.66ns +- 1% +11.88% (p=0.000 n=12+12) BM_InlinedVectorFillString/1 11.5ns +- 0% 12.8ns +- 1% +11.62% (p=0.000 n=12+11) BM_SwapElements<LargeCopyableMovableSwappable>/64 1.48ns +- 2% 1.66ns +- 1% +11.66% (p=0.000 n=12+11) BM_SwapElements<LargeCopyableMovableSwappable>/1k 1.48ns +- 1% 1.65ns +- 2% +11.32% (p=0.000 n=12+12) BM_SwapElements<LargeCopyableMovable>/512 1.48ns +- 2% 1.58ns +- 4% +6.62% (p=0.000 n=11+12) BM_SwapElements<LargeCopyableMovable>/1k 1.49ns +- 2% 1.58ns +- 3% +6.05% (p=0.000 n=12+12) BM_SwapElements<LargeCopyableMovable>/64 1.48ns +- 2% 1.57ns +- 4% +6.04% (p=0.000 n=11+12) BM_InlinedVectorFillRange/1 4.81ns +- 0% 5.05ns +- 0% +4.83% (p=0.000 n=11+11) BM_InlinedVectorFillString/8 79.4ns +- 1% 83.1ns +- 1% +4.64% (p=0.000 n=10+12) BM_StdVectorFillString/1 16.3ns +- 0% 16.6ns +- 0% +2.13% (p=0.000 n=11+8) ``` PiperOrigin-RevId: 353906786 -- 8e26518b3cec9c598e5e9573c46c3bd1b03a67ef by Abseil Team <absl-team@google.com>: Internal change PiperOrigin-RevId: 353737330 -- f206ae0983e58c9904ed8b8f05f9caf564a446be by Matt Kulukundis <kfm@google.com>: Import of CCTZ from GitHub. PiperOrigin-RevId: 353682256 GitOrigin-RevId: c68f1886f5e8fd90eb0c2d2e68feaf00a7cdacda Change-Id: I5790c1036c4f543c701d1039848fabf7ae881ad8
4 years ago
if (cord_ring_enabled()) {
return RingNewTree(data, length, alloc_hint);
}
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
absl::FixedArray<CordRep*> reps((length - 1) / kMaxFlatLength + 1);
size_t n = 0;
do {
const size_t len = std::min(length, kMaxFlatLength);
Export of internal Abseil changes -- a0491c8d790972cd80e2d720fe1fdf5f711a6f1a by Greg Falcon <gfalcon@google.com>: Stop directly accessing CordRepFlat data via CordRep::data. The old pattern of access breaks the `CordRep` type abstraction; since `CordRep::data` is not in general guaranteed to contain the chunk's data, we shouldn't access it that way. This incidentally adds an assertion check (via the flat() accessor) that the CordRep is indeed flat on each such access, but a manual inspection of the code, as well as the fact that this code currently works, suggest that this is always true.) PiperOrigin-RevId: 351592344 -- f40c3b43ca5b1d7e23cd45f1ffac1783105ac1a3 by Abseil Team <absl-team@google.com>: Revert 18abb2902b9f06c63a968b24d3dda785ebf99a22 PiperOrigin-RevId: 351523518 -- 18abb2902b9f06c63a968b24d3dda785ebf99a22 by Abseil Team <absl-team@google.com>: Internal change PiperOrigin-RevId: 351512412 -- 9b881602d45e95e06089792c7627cd56528a255a by Abseil Team <absl-team@google.com>: Keep time's global state in a cacheline-aligned structure. Keeping the global state as separate global variables results in two issues: 1) False sharing with adjacent global data (e.g., cycle clock source), since the global fields are updated every O(10usec). 2) The hot global fields (e.g., seq and samples) can reside on different cache lines. To fix this, simply wrap the global data in a ABSL_CACHE_ALIGNED structure. This is similar to what we do for MutexGlobals. PiperOrigin-RevId: 351389466 GitOrigin-RevId: a0491c8d790972cd80e2d720fe1fdf5f711a6f1a Change-Id: Ie0fa80112043381cd37c84e2ab2b7334839f54b5
4 years ago
CordRepFlat* rep = CordRepFlat::New(len + alloc_hint);
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
rep->length = len;
Export of internal Abseil changes -- a0491c8d790972cd80e2d720fe1fdf5f711a6f1a by Greg Falcon <gfalcon@google.com>: Stop directly accessing CordRepFlat data via CordRep::data. The old pattern of access breaks the `CordRep` type abstraction; since `CordRep::data` is not in general guaranteed to contain the chunk's data, we shouldn't access it that way. This incidentally adds an assertion check (via the flat() accessor) that the CordRep is indeed flat on each such access, but a manual inspection of the code, as well as the fact that this code currently works, suggest that this is always true.) PiperOrigin-RevId: 351592344 -- f40c3b43ca5b1d7e23cd45f1ffac1783105ac1a3 by Abseil Team <absl-team@google.com>: Revert 18abb2902b9f06c63a968b24d3dda785ebf99a22 PiperOrigin-RevId: 351523518 -- 18abb2902b9f06c63a968b24d3dda785ebf99a22 by Abseil Team <absl-team@google.com>: Internal change PiperOrigin-RevId: 351512412 -- 9b881602d45e95e06089792c7627cd56528a255a by Abseil Team <absl-team@google.com>: Keep time's global state in a cacheline-aligned structure. Keeping the global state as separate global variables results in two issues: 1) False sharing with adjacent global data (e.g., cycle clock source), since the global fields are updated every O(10usec). 2) The hot global fields (e.g., seq and samples) can reside on different cache lines. To fix this, simply wrap the global data in a ABSL_CACHE_ALIGNED structure. This is similar to what we do for MutexGlobals. PiperOrigin-RevId: 351389466 GitOrigin-RevId: a0491c8d790972cd80e2d720fe1fdf5f711a6f1a Change-Id: Ie0fa80112043381cd37c84e2ab2b7334839f54b5
4 years ago
memcpy(rep->Data(), data, len);
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
reps[n++] = VerifyTree(rep);
data += len;
length -= len;
} while (length != 0);
return MakeBalancedTree(reps.data(), n);
}
namespace cord_internal {
void InitializeCordRepExternal(absl::string_view data, CordRepExternal* rep) {
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
assert(!data.empty());
rep->length = data.size();
rep->tag = EXTERNAL;
rep->base = data.data();
VerifyTree(rep);
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
}
} // namespace cord_internal
static CordRep* NewSubstring(CordRep* child, size_t offset, size_t length) {
// Never create empty substring nodes
if (length == 0) {
Export of internal Abseil changes -- 6e9f93888bbe6718997ed90bbd049f1f3572b066 by Abseil Team <absl-team@google.com>: Fix status to be safe for self move assignment. PiperOrigin-RevId: 346815392 -- 35cae74a977f258e81dfbe925fb5a34cb6632036 by Gennadiy Rozental <rogeeff@google.com>: Eliminate unnecessary access to the global vars. PiperOrigin-RevId: 346777168 -- e7e786c243069060f5d6c1c945decb4b0b83f95b by Andy Getzendanner <durandal@google.com>: Internal change. PiperOrigin-RevId: 346685656 -- 4ccd41c48f1a83cfa20b3ea534f743dd7d788376 by Abseil Team <absl-team@google.com>: Move CordRep Ref() and Unref() logic into cord_internal.h This change moves Ref() and Unref() logic out of cord.cc into cord_internal. The main purpose is to make upcoming ring buffer changes easier to isolate from existing cord.cc code. Notice that this removes the nullptr check from Unref() and now requires it to be non null (which held true most times). We may need to rethink if the 'unref unlikely one' is the common case: are cordreps most likely shared? This may be something between 'root' and non root nodes, i.e., is it more likely for leaf / flat nodes in large cords to be shared than top level cordreps being shared? Vice versa? Benchmarks say that we mostly shouldn't care, the caveat being that atomic ops seem more expensive on upcoming archs (arcadia) so we should error on the side of an extra IsOne() branch saving us single owned atomic ops. PiperOrigin-RevId: 346676121 -- f0babab103b9e60d61ba09482d468985e43eceb3 by Samuel Benzaquen <sbenza@google.com>: Fix iterator based constructor and `.insert` members to only require EmplaceConstructible as the standard specifies. PiperOrigin-RevId: 346616707 -- 8f48eedda02277f9c96a88ed7726e34b557cce20 by Evan Brown <ezb@google.com>: Fix a bug in binary_search_impl when there's a transparent, three-way comparator that has different equivalence classes for different lookup types. Add a new can_have_multiple_equivalent_keys method to share the common logic for these cases. PiperOrigin-RevId: 346605948 -- 649183cb3cc9383431de9c81fb1c0f885d4001ae by Abseil Team <absl-team@google.com>: Add benchmark for accessing a Duration flag. PiperOrigin-RevId: 346594843 -- fefdb046520871af63ce2229e2f7cccfc0483dea by Abseil Team <absl-team@google.com>: Restructure CordReader for upcoming ring buffer changes. PiperOrigin-RevId: 346410642 -- 8b2f50e7da0ebab06ead5f94e366e984ca23cb6a by Abseil Team <absl-team@google.com>: Wire in an internal-only flag to toggle upcoming ring buffer changes on/off for experimentation. PiperOrigin-RevId: 346199111 GitOrigin-RevId: 6e9f93888bbe6718997ed90bbd049f1f3572b066 Change-Id: I8f34866b25a79209cb5448bbb28dd3044111d2e9
4 years ago
CordRep::Unref(child);
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
return nullptr;
} else {
CordRepSubstring* rep = new CordRepSubstring();
assert((offset + length) <= child->length);
rep->length = length;
rep->tag = SUBSTRING;
rep->start = offset;
rep->child = child;
return VerifyTree(rep);
}
}
// --------------------------------------------------------------------
// Cord::InlineRep functions
Export of internal Abseil changes -- 739f9fb80212c21c015fec473e9e29803a156ef9 by Derek Mauro <dmauro@google.com>: Define FlagStateInterface::~FlagStateInterface() in the translation unit in which it is actually declared Fixes #717 PiperOrigin-RevId: 319083605 -- 913ef1f23113268b22d636d3ae3b992862efdb1a by Derek Mauro <dmauro@google.com>: Fix ABSL_LOCK_RETURNED statement PiperOrigin-RevId: 319078667 -- a43b1413da1770d638147c73e7e1693cfaf869c7 by Derek Mauro <dmauro@google.com>: Fix redeclaration ‘absl::Cord::InlineRep::kMaxInline’, which differs in ‘constexpr’ Fixes #725 PiperOrigin-RevId: 319060910 -- 1ad7d491a80f6c9de78a6fc20f09b7765d224503 by Abseil Team <absl-team@google.com>: Make absl SpinLock trivially destructible when possible PiperOrigin-RevId: 319049456 -- 659ecad3578dfa669854a82279fa590002bdb37f by Derek Mauro <dmauro@google.com>: Remove the static initialization of global variables used by absl::Mutex as requested by Chromium PiperOrigin-RevId: 319031204 -- 609c491d8bb4f8bb3b44c5a4c0bee51c583df24c by Abseil Team <absl-team@google.com>: Add implementation of %a and %A to absl::StrFormat. Prior to this it just fell back to sprintf. PiperOrigin-RevId: 318888039 -- 5e8ae6392bcd135248aac14c4b9f2a5116868678 by Abseil Team <absl-team@google.com>: Google-internal changes only. PiperOrigin-RevId: 318857077 -- 4a2578e33e8442954e29e5f0380ddfcf0f033f0d by Greg Falcon <gfalcon@google.com>: Change of enum constants to accommodate internal change. PiperOrigin-RevId: 318844716 -- 4b578b102816260c213675759f4c15911735578a by Abseil Team <absl-team@google.com>: Internal change PiperOrigin-RevId: 318704453 -- 0ee82fd24d548b260c9229fa1f54571dae1dfa24 by Gennadiy Rozental <rogeeff@google.com>: Allow lookup of retired flags. At the moment we issue warning on attempt to find a retired flag. This way we can't even check if flag is retired without issuing the warning. With this change we will only issue the warning if one tries to access any functionality of retired flag but it's name "is retired" status, and type. PiperOrigin-RevId: 318605017 -- 3e35fe9b4c79f636fa328c59e2aabb93e29b7c99 by Abseil Team <absl-team@google.com>: Fix error return from InstallSymbolDecorator(). PiperOrigin-RevId: 318490405 -- ae46063f3eb2998cb961f62a359d932e5908a4bc by Abseil Team <absl-team@google.com>: Do not make copies of iterated collection elements into the loop variable. PiperOrigin-RevId: 318423139 -- d06a075a12aab5f6ab98474677ce50d588b21de3 by Abseil Team <absl-team@google.com>: add missing word making the error code better English PiperOrigin-RevId: 318335052 GitOrigin-RevId: 739f9fb80212c21c015fec473e9e29803a156ef9 Change-Id: Id77a0a4b1959036b00555deef40e82d51884fbc1
4 years ago
constexpr unsigned char Cord::InlineRep::kMaxInline;
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
inline void Cord::InlineRep::set_data(const char* data, size_t n,
bool nullify_tail) {
static_assert(kMaxInline == 15, "set_data is hard-coded for a length of 15");
Export of internal Abseil changes -- 642ab296a2c9629c44f3f2ce6911cd2488bcf416 by Derek Mauro <dmauro@google.com>: Remove an obsolete check in CMakeLists.txt PiperOrigin-RevId: 352852564 -- ce78cb96bcfd162737dbcf35005da3d1d6a3486b by Abseil Team <absl-team@google.com>: Clarify that the calling *thread* must have locked the mutex in order to unlock it. PiperOrigin-RevId: 352801804 -- 24e1f5f72756046f5265abf618e951c341f09b8d by Derek Mauro <dmauro@google.com>: Fixes failing CMake string comparisons https://cmake.org/cmake/help/latest/policy/CMP0054.html Fixes #791 PiperOrigin-RevId: 352791054 -- 0ac10bc3f4dca2c4c4b51d7b8196a2eaee9537a1 by Abseil Team <absl-team@google.com>: Introduce CordRepRing class This change introduces the CordRepRing class that implements all the lower level / internal implementation for upcoming CordRepRing ring buffer support in cord. PiperOrigin-RevId: 352771994 -- 4bd36dda61760785844f0f29f26d90cc18046f75 by Abseil Team <absl-team@google.com>: Optimize InlineData representation for cord sampling (cordz) This CL changes InlineData to allow us to store a (future) Cordz Info pointer directly into the inline representation: - make InlineData a class that provides a public API to set the active union members (tree or chars) and safely access that data. - change 'tree' and 'profiled' bits to be the 2 least significant bits, allowing us 62 continquous bits for storing a Cordz Info pointer. PiperOrigin-RevId: 352642411 -- dc55ba71bbce0e6a83e05a453990c51ac3d68426 by Mark Barolak <mbar@google.com>: Add unit test coverage for the mutating overload of absl::AsciiStrToLower. PiperOrigin-RevId: 352626006 GitOrigin-RevId: 642ab296a2c9629c44f3f2ce6911cd2488bcf416 Change-Id: I6c5929dd830d3c630e14e7fd5387fc3e25a69100
4 years ago
cord_internal::SmallMemmove(data_.as_chars(), data, n, nullify_tail);
set_inline_size(n);
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
}
inline char* Cord::InlineRep::set_data(size_t n) {
assert(n <= kMaxInline);
ResetToEmpty();
Export of internal Abseil changes -- 642ab296a2c9629c44f3f2ce6911cd2488bcf416 by Derek Mauro <dmauro@google.com>: Remove an obsolete check in CMakeLists.txt PiperOrigin-RevId: 352852564 -- ce78cb96bcfd162737dbcf35005da3d1d6a3486b by Abseil Team <absl-team@google.com>: Clarify that the calling *thread* must have locked the mutex in order to unlock it. PiperOrigin-RevId: 352801804 -- 24e1f5f72756046f5265abf618e951c341f09b8d by Derek Mauro <dmauro@google.com>: Fixes failing CMake string comparisons https://cmake.org/cmake/help/latest/policy/CMP0054.html Fixes #791 PiperOrigin-RevId: 352791054 -- 0ac10bc3f4dca2c4c4b51d7b8196a2eaee9537a1 by Abseil Team <absl-team@google.com>: Introduce CordRepRing class This change introduces the CordRepRing class that implements all the lower level / internal implementation for upcoming CordRepRing ring buffer support in cord. PiperOrigin-RevId: 352771994 -- 4bd36dda61760785844f0f29f26d90cc18046f75 by Abseil Team <absl-team@google.com>: Optimize InlineData representation for cord sampling (cordz) This CL changes InlineData to allow us to store a (future) Cordz Info pointer directly into the inline representation: - make InlineData a class that provides a public API to set the active union members (tree or chars) and safely access that data. - change 'tree' and 'profiled' bits to be the 2 least significant bits, allowing us 62 continquous bits for storing a Cordz Info pointer. PiperOrigin-RevId: 352642411 -- dc55ba71bbce0e6a83e05a453990c51ac3d68426 by Mark Barolak <mbar@google.com>: Add unit test coverage for the mutating overload of absl::AsciiStrToLower. PiperOrigin-RevId: 352626006 GitOrigin-RevId: 642ab296a2c9629c44f3f2ce6911cd2488bcf416 Change-Id: I6c5929dd830d3c630e14e7fd5387fc3e25a69100
4 years ago
set_inline_size(n);
return data_.as_chars();
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
}
inline CordRep* Cord::InlineRep::force_tree(size_t extra_hint) {
Export of internal Abseil changes -- 642ab296a2c9629c44f3f2ce6911cd2488bcf416 by Derek Mauro <dmauro@google.com>: Remove an obsolete check in CMakeLists.txt PiperOrigin-RevId: 352852564 -- ce78cb96bcfd162737dbcf35005da3d1d6a3486b by Abseil Team <absl-team@google.com>: Clarify that the calling *thread* must have locked the mutex in order to unlock it. PiperOrigin-RevId: 352801804 -- 24e1f5f72756046f5265abf618e951c341f09b8d by Derek Mauro <dmauro@google.com>: Fixes failing CMake string comparisons https://cmake.org/cmake/help/latest/policy/CMP0054.html Fixes #791 PiperOrigin-RevId: 352791054 -- 0ac10bc3f4dca2c4c4b51d7b8196a2eaee9537a1 by Abseil Team <absl-team@google.com>: Introduce CordRepRing class This change introduces the CordRepRing class that implements all the lower level / internal implementation for upcoming CordRepRing ring buffer support in cord. PiperOrigin-RevId: 352771994 -- 4bd36dda61760785844f0f29f26d90cc18046f75 by Abseil Team <absl-team@google.com>: Optimize InlineData representation for cord sampling (cordz) This CL changes InlineData to allow us to store a (future) Cordz Info pointer directly into the inline representation: - make InlineData a class that provides a public API to set the active union members (tree or chars) and safely access that data. - change 'tree' and 'profiled' bits to be the 2 least significant bits, allowing us 62 continquous bits for storing a Cordz Info pointer. PiperOrigin-RevId: 352642411 -- dc55ba71bbce0e6a83e05a453990c51ac3d68426 by Mark Barolak <mbar@google.com>: Add unit test coverage for the mutating overload of absl::AsciiStrToLower. PiperOrigin-RevId: 352626006 GitOrigin-RevId: 642ab296a2c9629c44f3f2ce6911cd2488bcf416 Change-Id: I6c5929dd830d3c630e14e7fd5387fc3e25a69100
4 years ago
if (data_.is_tree()) {
return data_.as_tree();
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
}
Export of internal Abseil changes -- 642ab296a2c9629c44f3f2ce6911cd2488bcf416 by Derek Mauro <dmauro@google.com>: Remove an obsolete check in CMakeLists.txt PiperOrigin-RevId: 352852564 -- ce78cb96bcfd162737dbcf35005da3d1d6a3486b by Abseil Team <absl-team@google.com>: Clarify that the calling *thread* must have locked the mutex in order to unlock it. PiperOrigin-RevId: 352801804 -- 24e1f5f72756046f5265abf618e951c341f09b8d by Derek Mauro <dmauro@google.com>: Fixes failing CMake string comparisons https://cmake.org/cmake/help/latest/policy/CMP0054.html Fixes #791 PiperOrigin-RevId: 352791054 -- 0ac10bc3f4dca2c4c4b51d7b8196a2eaee9537a1 by Abseil Team <absl-team@google.com>: Introduce CordRepRing class This change introduces the CordRepRing class that implements all the lower level / internal implementation for upcoming CordRepRing ring buffer support in cord. PiperOrigin-RevId: 352771994 -- 4bd36dda61760785844f0f29f26d90cc18046f75 by Abseil Team <absl-team@google.com>: Optimize InlineData representation for cord sampling (cordz) This CL changes InlineData to allow us to store a (future) Cordz Info pointer directly into the inline representation: - make InlineData a class that provides a public API to set the active union members (tree or chars) and safely access that data. - change 'tree' and 'profiled' bits to be the 2 least significant bits, allowing us 62 continquous bits for storing a Cordz Info pointer. PiperOrigin-RevId: 352642411 -- dc55ba71bbce0e6a83e05a453990c51ac3d68426 by Mark Barolak <mbar@google.com>: Add unit test coverage for the mutating overload of absl::AsciiStrToLower. PiperOrigin-RevId: 352626006 GitOrigin-RevId: 642ab296a2c9629c44f3f2ce6911cd2488bcf416 Change-Id: I6c5929dd830d3c630e14e7fd5387fc3e25a69100
4 years ago
size_t len = inline_size();
Export of internal Abseil changes -- a0491c8d790972cd80e2d720fe1fdf5f711a6f1a by Greg Falcon <gfalcon@google.com>: Stop directly accessing CordRepFlat data via CordRep::data. The old pattern of access breaks the `CordRep` type abstraction; since `CordRep::data` is not in general guaranteed to contain the chunk's data, we shouldn't access it that way. This incidentally adds an assertion check (via the flat() accessor) that the CordRep is indeed flat on each such access, but a manual inspection of the code, as well as the fact that this code currently works, suggest that this is always true.) PiperOrigin-RevId: 351592344 -- f40c3b43ca5b1d7e23cd45f1ffac1783105ac1a3 by Abseil Team <absl-team@google.com>: Revert 18abb2902b9f06c63a968b24d3dda785ebf99a22 PiperOrigin-RevId: 351523518 -- 18abb2902b9f06c63a968b24d3dda785ebf99a22 by Abseil Team <absl-team@google.com>: Internal change PiperOrigin-RevId: 351512412 -- 9b881602d45e95e06089792c7627cd56528a255a by Abseil Team <absl-team@google.com>: Keep time's global state in a cacheline-aligned structure. Keeping the global state as separate global variables results in two issues: 1) False sharing with adjacent global data (e.g., cycle clock source), since the global fields are updated every O(10usec). 2) The hot global fields (e.g., seq and samples) can reside on different cache lines. To fix this, simply wrap the global data in a ABSL_CACHE_ALIGNED structure. This is similar to what we do for MutexGlobals. PiperOrigin-RevId: 351389466 GitOrigin-RevId: a0491c8d790972cd80e2d720fe1fdf5f711a6f1a Change-Id: Ie0fa80112043381cd37c84e2ab2b7334839f54b5
4 years ago
CordRepFlat* result = CordRepFlat::New(len + extra_hint);
result->length = len;
Export of internal Abseil changes -- 642ab296a2c9629c44f3f2ce6911cd2488bcf416 by Derek Mauro <dmauro@google.com>: Remove an obsolete check in CMakeLists.txt PiperOrigin-RevId: 352852564 -- ce78cb96bcfd162737dbcf35005da3d1d6a3486b by Abseil Team <absl-team@google.com>: Clarify that the calling *thread* must have locked the mutex in order to unlock it. PiperOrigin-RevId: 352801804 -- 24e1f5f72756046f5265abf618e951c341f09b8d by Derek Mauro <dmauro@google.com>: Fixes failing CMake string comparisons https://cmake.org/cmake/help/latest/policy/CMP0054.html Fixes #791 PiperOrigin-RevId: 352791054 -- 0ac10bc3f4dca2c4c4b51d7b8196a2eaee9537a1 by Abseil Team <absl-team@google.com>: Introduce CordRepRing class This change introduces the CordRepRing class that implements all the lower level / internal implementation for upcoming CordRepRing ring buffer support in cord. PiperOrigin-RevId: 352771994 -- 4bd36dda61760785844f0f29f26d90cc18046f75 by Abseil Team <absl-team@google.com>: Optimize InlineData representation for cord sampling (cordz) This CL changes InlineData to allow us to store a (future) Cordz Info pointer directly into the inline representation: - make InlineData a class that provides a public API to set the active union members (tree or chars) and safely access that data. - change 'tree' and 'profiled' bits to be the 2 least significant bits, allowing us 62 continquous bits for storing a Cordz Info pointer. PiperOrigin-RevId: 352642411 -- dc55ba71bbce0e6a83e05a453990c51ac3d68426 by Mark Barolak <mbar@google.com>: Add unit test coverage for the mutating overload of absl::AsciiStrToLower. PiperOrigin-RevId: 352626006 GitOrigin-RevId: 642ab296a2c9629c44f3f2ce6911cd2488bcf416 Change-Id: I6c5929dd830d3c630e14e7fd5387fc3e25a69100
4 years ago
static_assert(kMinFlatLength >= sizeof(data_), "");
memcpy(result->Data(), data_.as_chars(), sizeof(data_));
set_tree(result);
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
return result;
}
inline void Cord::InlineRep::reduce_size(size_t n) {
Export of internal Abseil changes -- 642ab296a2c9629c44f3f2ce6911cd2488bcf416 by Derek Mauro <dmauro@google.com>: Remove an obsolete check in CMakeLists.txt PiperOrigin-RevId: 352852564 -- ce78cb96bcfd162737dbcf35005da3d1d6a3486b by Abseil Team <absl-team@google.com>: Clarify that the calling *thread* must have locked the mutex in order to unlock it. PiperOrigin-RevId: 352801804 -- 24e1f5f72756046f5265abf618e951c341f09b8d by Derek Mauro <dmauro@google.com>: Fixes failing CMake string comparisons https://cmake.org/cmake/help/latest/policy/CMP0054.html Fixes #791 PiperOrigin-RevId: 352791054 -- 0ac10bc3f4dca2c4c4b51d7b8196a2eaee9537a1 by Abseil Team <absl-team@google.com>: Introduce CordRepRing class This change introduces the CordRepRing class that implements all the lower level / internal implementation for upcoming CordRepRing ring buffer support in cord. PiperOrigin-RevId: 352771994 -- 4bd36dda61760785844f0f29f26d90cc18046f75 by Abseil Team <absl-team@google.com>: Optimize InlineData representation for cord sampling (cordz) This CL changes InlineData to allow us to store a (future) Cordz Info pointer directly into the inline representation: - make InlineData a class that provides a public API to set the active union members (tree or chars) and safely access that data. - change 'tree' and 'profiled' bits to be the 2 least significant bits, allowing us 62 continquous bits for storing a Cordz Info pointer. PiperOrigin-RevId: 352642411 -- dc55ba71bbce0e6a83e05a453990c51ac3d68426 by Mark Barolak <mbar@google.com>: Add unit test coverage for the mutating overload of absl::AsciiStrToLower. PiperOrigin-RevId: 352626006 GitOrigin-RevId: 642ab296a2c9629c44f3f2ce6911cd2488bcf416 Change-Id: I6c5929dd830d3c630e14e7fd5387fc3e25a69100
4 years ago
size_t tag = inline_size();
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
assert(tag <= kMaxInline);
assert(tag >= n);
tag -= n;
Export of internal Abseil changes -- 642ab296a2c9629c44f3f2ce6911cd2488bcf416 by Derek Mauro <dmauro@google.com>: Remove an obsolete check in CMakeLists.txt PiperOrigin-RevId: 352852564 -- ce78cb96bcfd162737dbcf35005da3d1d6a3486b by Abseil Team <absl-team@google.com>: Clarify that the calling *thread* must have locked the mutex in order to unlock it. PiperOrigin-RevId: 352801804 -- 24e1f5f72756046f5265abf618e951c341f09b8d by Derek Mauro <dmauro@google.com>: Fixes failing CMake string comparisons https://cmake.org/cmake/help/latest/policy/CMP0054.html Fixes #791 PiperOrigin-RevId: 352791054 -- 0ac10bc3f4dca2c4c4b51d7b8196a2eaee9537a1 by Abseil Team <absl-team@google.com>: Introduce CordRepRing class This change introduces the CordRepRing class that implements all the lower level / internal implementation for upcoming CordRepRing ring buffer support in cord. PiperOrigin-RevId: 352771994 -- 4bd36dda61760785844f0f29f26d90cc18046f75 by Abseil Team <absl-team@google.com>: Optimize InlineData representation for cord sampling (cordz) This CL changes InlineData to allow us to store a (future) Cordz Info pointer directly into the inline representation: - make InlineData a class that provides a public API to set the active union members (tree or chars) and safely access that data. - change 'tree' and 'profiled' bits to be the 2 least significant bits, allowing us 62 continquous bits for storing a Cordz Info pointer. PiperOrigin-RevId: 352642411 -- dc55ba71bbce0e6a83e05a453990c51ac3d68426 by Mark Barolak <mbar@google.com>: Add unit test coverage for the mutating overload of absl::AsciiStrToLower. PiperOrigin-RevId: 352626006 GitOrigin-RevId: 642ab296a2c9629c44f3f2ce6911cd2488bcf416 Change-Id: I6c5929dd830d3c630e14e7fd5387fc3e25a69100
4 years ago
memset(data_.as_chars() + tag, 0, n);
set_inline_size(static_cast<char>(tag));
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
}
inline void Cord::InlineRep::remove_prefix(size_t n) {
Export of internal Abseil changes -- 642ab296a2c9629c44f3f2ce6911cd2488bcf416 by Derek Mauro <dmauro@google.com>: Remove an obsolete check in CMakeLists.txt PiperOrigin-RevId: 352852564 -- ce78cb96bcfd162737dbcf35005da3d1d6a3486b by Abseil Team <absl-team@google.com>: Clarify that the calling *thread* must have locked the mutex in order to unlock it. PiperOrigin-RevId: 352801804 -- 24e1f5f72756046f5265abf618e951c341f09b8d by Derek Mauro <dmauro@google.com>: Fixes failing CMake string comparisons https://cmake.org/cmake/help/latest/policy/CMP0054.html Fixes #791 PiperOrigin-RevId: 352791054 -- 0ac10bc3f4dca2c4c4b51d7b8196a2eaee9537a1 by Abseil Team <absl-team@google.com>: Introduce CordRepRing class This change introduces the CordRepRing class that implements all the lower level / internal implementation for upcoming CordRepRing ring buffer support in cord. PiperOrigin-RevId: 352771994 -- 4bd36dda61760785844f0f29f26d90cc18046f75 by Abseil Team <absl-team@google.com>: Optimize InlineData representation for cord sampling (cordz) This CL changes InlineData to allow us to store a (future) Cordz Info pointer directly into the inline representation: - make InlineData a class that provides a public API to set the active union members (tree or chars) and safely access that data. - change 'tree' and 'profiled' bits to be the 2 least significant bits, allowing us 62 continquous bits for storing a Cordz Info pointer. PiperOrigin-RevId: 352642411 -- dc55ba71bbce0e6a83e05a453990c51ac3d68426 by Mark Barolak <mbar@google.com>: Add unit test coverage for the mutating overload of absl::AsciiStrToLower. PiperOrigin-RevId: 352626006 GitOrigin-RevId: 642ab296a2c9629c44f3f2ce6911cd2488bcf416 Change-Id: I6c5929dd830d3c630e14e7fd5387fc3e25a69100
4 years ago
cord_internal::SmallMemmove(data_.as_chars(), data_.as_chars() + n,
inline_size() - n);
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
reduce_size(n);
}
Export of internal Abseil changes -- c68f1886f5e8fd90eb0c2d2e68feaf00a7cdacda by CJ Johnson <johnsoncj@google.com>: Introduce absl::Cleanup to the OSS repo PiperOrigin-RevId: 354583156 -- 17030cf388e10f7eb959e3e566326d1072ce392e by Abseil Team <absl-team@google.com>: Internal change only PiperOrigin-RevId: 354574953 -- e979d7236d4f3252e79ddda6739b67a9a326bf6d by CJ Johnson <johnsoncj@google.com>: Internal change PiperOrigin-RevId: 354545297 -- 7ea02b3783f7f49ef97d86a8f6580a19cc57df14 by Abseil Team <absl-team@google.com>: Pre-allocate memory for vectors where the size is known. PiperOrigin-RevId: 354344576 -- 9246c7cb11f1d6444f79ebe25acc69a8a9b870e0 by Matt Kulukundis <kfm@google.com>: Add support for Elbrus 2000 (e2k) Import of https://github.com/abseil/abseil-cpp/pull/889 PiperOrigin-RevId: 354344013 -- 0fc93d359cc1fb307552e917b37b7b2e7eed822f by Abseil Team <absl-team@google.com>: Integrate CordRepRing logic into cord (but do not enable it) PiperOrigin-RevId: 354312238 -- eda05622f7da71466723acb33403f783529df24b by Abseil Team <absl-team@google.com>: Protect ignore diagnostic with "__has_warning". PiperOrigin-RevId: 354112334 -- 47716c5d8fb10efa4fdd801d28bac414c6f8ec32 by Abseil Team <absl-team@google.com>: Rearrange InlinedVector copy constructor and destructor to treat a few special cases inline and then tail-call a non-inlined routine for the rest. In particular, we optimize for empty vectors in both cases. Added a couple of benchmarks that copy either an InlVec<int64> or an InlVec<InlVec<int64>>. Speed difference: ``` BM_CopyTrivial/0 0.92ns +- 0% 0.47ns +- 0% -48.91% (p=0.000 n=11+12) BM_CopyTrivial/1 0.92ns +- 0% 1.15ns +- 0% +25.00% (p=0.000 n=10+9) BM_CopyTrivial/8 8.57ns +- 0% 10.72ns +- 1% +25.16% (p=0.000 n=10+12) BM_CopyNonTrivial/0 3.21ns +- 0% 0.70ns +- 0% -78.23% (p=0.000 n=12+10) BM_CopyNonTrivial/1 5.88ns +- 1% 5.51ns +- 0% -6.28% (p=0.000 n=10+8) BM_CopyNonTrivial/8 21.5ns +- 1% 15.2ns +- 2% -29.23% (p=0.000 n=12+12) ``` Note: the slowdowns are a few cycles which is expected given the procedure call added in that case. We decided this is a good tradeoff given the code size reductions and the more significant speedups for empty vectors. Size difference (as measured by nm): ``` BM_CopyTrivial from 1048 bytes to 326 bytes. BM_CopyNonTrivial from 749 bytes to 470 bytes. ``` Code size for a large binary drops by ~500KB (from 349415719 to 348906015 348906191). All of the benchmarks that showed a significant difference: Ones that improve with this CL: ``` BM_CopyNonTrivial/0 3.21ns +- 0% 0.70ns +- 0% -78.23% (p=0.000 n=12+10) BM_InlinedVectorFillString/0 0.93ns +- 0% 0.24ns +- 0% -74.19% (p=0.000 n=12+10) BM_InlinedVectorAssignments/1 10.5ns +- 0% 4.1ns +- 0% -60.64% (p=0.000 n=11+10) BM_InlinedVectorAssignments/2 10.7ns +- 0% 4.4ns +- 0% -59.08% (p=0.000 n=11+11) BM_CopyTrivial/0 0.92ns +- 0% 0.47ns +- 0% -48.91% (p=0.000 n=11+12) BM_CopyNonTrivial/8 21.5ns +- 1% 15.2ns +- 2% -29.23% (p=0.000 n=12+12) BM_StdVectorEmpty 0.47ns +- 1% 0.35ns +- 0% -24.73% (p=0.000 n=12+12) BM_StdVectorSize 0.46ns +- 2% 0.35ns +- 0% -24.32% (p=0.000 n=12+12) BM_SwapElements<LargeCopyableOnly>/0 3.44ns +- 0% 2.76ns +- 1% -19.83% (p=0.000 n=11+11) BM_InlinedVectorFillRange/256 20.7ns +- 1% 17.8ns +- 0% -14.08% (p=0.000 n=12+9) BM_CopyNonTrivial/1 5.88ns +- 1% 5.51ns +- 0% -6.28% (p=0.000 n=10+8) BM_SwapElements<LargeCopyableMovable>/1 4.19ns +- 0% 3.95ns +- 1% -5.63% (p=0.000 n=11+12) BM_SwapElements<LargeCopyableMovableSwappable>/1 4.18ns +- 0% 3.99ns +- 0% -4.70% (p=0.000 n=9+11) BM_SwapElements<LargeCopyableMovable>/0 2.41ns +- 0% 2.31ns +- 0% -4.45% (p=0.000 n=12+12) BM_InlinedVectorFillRange/64 8.25ns +- 0% 8.04ns +- 0% -2.51% (p=0.000 n=12+11) BM_SwapElements<LargeCopyableOnly>/1 82.4ns +- 0% 81.5ns +- 0% -1.06% (p=0.000 n=12+12) ``` Ones that get worse with this CL: ``` BM_CopyTrivial/1 0.92ns +- 0% 1.15ns +- 0% +25.00% (p=0.000 n=10+9) BM_CopyTrivial/8 8.57ns +- 0% 10.72ns +- 1% +25.16% (p=0.000 n=10+12) BM_SwapElements<LargeCopyableMovableSwappable>/512 1.48ns +- 1% 1.66ns +- 1% +11.88% (p=0.000 n=12+12) BM_InlinedVectorFillString/1 11.5ns +- 0% 12.8ns +- 1% +11.62% (p=0.000 n=12+11) BM_SwapElements<LargeCopyableMovableSwappable>/64 1.48ns +- 2% 1.66ns +- 1% +11.66% (p=0.000 n=12+11) BM_SwapElements<LargeCopyableMovableSwappable>/1k 1.48ns +- 1% 1.65ns +- 2% +11.32% (p=0.000 n=12+12) BM_SwapElements<LargeCopyableMovable>/512 1.48ns +- 2% 1.58ns +- 4% +6.62% (p=0.000 n=11+12) BM_SwapElements<LargeCopyableMovable>/1k 1.49ns +- 2% 1.58ns +- 3% +6.05% (p=0.000 n=12+12) BM_SwapElements<LargeCopyableMovable>/64 1.48ns +- 2% 1.57ns +- 4% +6.04% (p=0.000 n=11+12) BM_InlinedVectorFillRange/1 4.81ns +- 0% 5.05ns +- 0% +4.83% (p=0.000 n=11+11) BM_InlinedVectorFillString/8 79.4ns +- 1% 83.1ns +- 1% +4.64% (p=0.000 n=10+12) BM_StdVectorFillString/1 16.3ns +- 0% 16.6ns +- 0% +2.13% (p=0.000 n=11+8) ``` PiperOrigin-RevId: 353906786 -- 8e26518b3cec9c598e5e9573c46c3bd1b03a67ef by Abseil Team <absl-team@google.com>: Internal change PiperOrigin-RevId: 353737330 -- f206ae0983e58c9904ed8b8f05f9caf564a446be by Matt Kulukundis <kfm@google.com>: Import of CCTZ from GitHub. PiperOrigin-RevId: 353682256 GitOrigin-RevId: c68f1886f5e8fd90eb0c2d2e68feaf00a7cdacda Change-Id: I5790c1036c4f543c701d1039848fabf7ae881ad8
4 years ago
// Returns `rep` converted into a CordRepRing.
// Directly returns `rep` if `rep` is already a CordRepRing.
static CordRepRing* ForceRing(CordRep* rep, size_t extra) {
return (rep->tag == RING) ? rep->ring() : CordRepRing::Create(rep, extra);
}
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
void Cord::InlineRep::AppendTree(CordRep* tree) {
if (tree == nullptr) return;
Export of internal Abseil changes -- 642ab296a2c9629c44f3f2ce6911cd2488bcf416 by Derek Mauro <dmauro@google.com>: Remove an obsolete check in CMakeLists.txt PiperOrigin-RevId: 352852564 -- ce78cb96bcfd162737dbcf35005da3d1d6a3486b by Abseil Team <absl-team@google.com>: Clarify that the calling *thread* must have locked the mutex in order to unlock it. PiperOrigin-RevId: 352801804 -- 24e1f5f72756046f5265abf618e951c341f09b8d by Derek Mauro <dmauro@google.com>: Fixes failing CMake string comparisons https://cmake.org/cmake/help/latest/policy/CMP0054.html Fixes #791 PiperOrigin-RevId: 352791054 -- 0ac10bc3f4dca2c4c4b51d7b8196a2eaee9537a1 by Abseil Team <absl-team@google.com>: Introduce CordRepRing class This change introduces the CordRepRing class that implements all the lower level / internal implementation for upcoming CordRepRing ring buffer support in cord. PiperOrigin-RevId: 352771994 -- 4bd36dda61760785844f0f29f26d90cc18046f75 by Abseil Team <absl-team@google.com>: Optimize InlineData representation for cord sampling (cordz) This CL changes InlineData to allow us to store a (future) Cordz Info pointer directly into the inline representation: - make InlineData a class that provides a public API to set the active union members (tree or chars) and safely access that data. - change 'tree' and 'profiled' bits to be the 2 least significant bits, allowing us 62 continquous bits for storing a Cordz Info pointer. PiperOrigin-RevId: 352642411 -- dc55ba71bbce0e6a83e05a453990c51ac3d68426 by Mark Barolak <mbar@google.com>: Add unit test coverage for the mutating overload of absl::AsciiStrToLower. PiperOrigin-RevId: 352626006 GitOrigin-RevId: 642ab296a2c9629c44f3f2ce6911cd2488bcf416 Change-Id: I6c5929dd830d3c630e14e7fd5387fc3e25a69100
4 years ago
if (data_.is_empty()) {
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
set_tree(tree);
Export of internal Abseil changes -- c68f1886f5e8fd90eb0c2d2e68feaf00a7cdacda by CJ Johnson <johnsoncj@google.com>: Introduce absl::Cleanup to the OSS repo PiperOrigin-RevId: 354583156 -- 17030cf388e10f7eb959e3e566326d1072ce392e by Abseil Team <absl-team@google.com>: Internal change only PiperOrigin-RevId: 354574953 -- e979d7236d4f3252e79ddda6739b67a9a326bf6d by CJ Johnson <johnsoncj@google.com>: Internal change PiperOrigin-RevId: 354545297 -- 7ea02b3783f7f49ef97d86a8f6580a19cc57df14 by Abseil Team <absl-team@google.com>: Pre-allocate memory for vectors where the size is known. PiperOrigin-RevId: 354344576 -- 9246c7cb11f1d6444f79ebe25acc69a8a9b870e0 by Matt Kulukundis <kfm@google.com>: Add support for Elbrus 2000 (e2k) Import of https://github.com/abseil/abseil-cpp/pull/889 PiperOrigin-RevId: 354344013 -- 0fc93d359cc1fb307552e917b37b7b2e7eed822f by Abseil Team <absl-team@google.com>: Integrate CordRepRing logic into cord (but do not enable it) PiperOrigin-RevId: 354312238 -- eda05622f7da71466723acb33403f783529df24b by Abseil Team <absl-team@google.com>: Protect ignore diagnostic with "__has_warning". PiperOrigin-RevId: 354112334 -- 47716c5d8fb10efa4fdd801d28bac414c6f8ec32 by Abseil Team <absl-team@google.com>: Rearrange InlinedVector copy constructor and destructor to treat a few special cases inline and then tail-call a non-inlined routine for the rest. In particular, we optimize for empty vectors in both cases. Added a couple of benchmarks that copy either an InlVec<int64> or an InlVec<InlVec<int64>>. Speed difference: ``` BM_CopyTrivial/0 0.92ns +- 0% 0.47ns +- 0% -48.91% (p=0.000 n=11+12) BM_CopyTrivial/1 0.92ns +- 0% 1.15ns +- 0% +25.00% (p=0.000 n=10+9) BM_CopyTrivial/8 8.57ns +- 0% 10.72ns +- 1% +25.16% (p=0.000 n=10+12) BM_CopyNonTrivial/0 3.21ns +- 0% 0.70ns +- 0% -78.23% (p=0.000 n=12+10) BM_CopyNonTrivial/1 5.88ns +- 1% 5.51ns +- 0% -6.28% (p=0.000 n=10+8) BM_CopyNonTrivial/8 21.5ns +- 1% 15.2ns +- 2% -29.23% (p=0.000 n=12+12) ``` Note: the slowdowns are a few cycles which is expected given the procedure call added in that case. We decided this is a good tradeoff given the code size reductions and the more significant speedups for empty vectors. Size difference (as measured by nm): ``` BM_CopyTrivial from 1048 bytes to 326 bytes. BM_CopyNonTrivial from 749 bytes to 470 bytes. ``` Code size for a large binary drops by ~500KB (from 349415719 to 348906015 348906191). All of the benchmarks that showed a significant difference: Ones that improve with this CL: ``` BM_CopyNonTrivial/0 3.21ns +- 0% 0.70ns +- 0% -78.23% (p=0.000 n=12+10) BM_InlinedVectorFillString/0 0.93ns +- 0% 0.24ns +- 0% -74.19% (p=0.000 n=12+10) BM_InlinedVectorAssignments/1 10.5ns +- 0% 4.1ns +- 0% -60.64% (p=0.000 n=11+10) BM_InlinedVectorAssignments/2 10.7ns +- 0% 4.4ns +- 0% -59.08% (p=0.000 n=11+11) BM_CopyTrivial/0 0.92ns +- 0% 0.47ns +- 0% -48.91% (p=0.000 n=11+12) BM_CopyNonTrivial/8 21.5ns +- 1% 15.2ns +- 2% -29.23% (p=0.000 n=12+12) BM_StdVectorEmpty 0.47ns +- 1% 0.35ns +- 0% -24.73% (p=0.000 n=12+12) BM_StdVectorSize 0.46ns +- 2% 0.35ns +- 0% -24.32% (p=0.000 n=12+12) BM_SwapElements<LargeCopyableOnly>/0 3.44ns +- 0% 2.76ns +- 1% -19.83% (p=0.000 n=11+11) BM_InlinedVectorFillRange/256 20.7ns +- 1% 17.8ns +- 0% -14.08% (p=0.000 n=12+9) BM_CopyNonTrivial/1 5.88ns +- 1% 5.51ns +- 0% -6.28% (p=0.000 n=10+8) BM_SwapElements<LargeCopyableMovable>/1 4.19ns +- 0% 3.95ns +- 1% -5.63% (p=0.000 n=11+12) BM_SwapElements<LargeCopyableMovableSwappable>/1 4.18ns +- 0% 3.99ns +- 0% -4.70% (p=0.000 n=9+11) BM_SwapElements<LargeCopyableMovable>/0 2.41ns +- 0% 2.31ns +- 0% -4.45% (p=0.000 n=12+12) BM_InlinedVectorFillRange/64 8.25ns +- 0% 8.04ns +- 0% -2.51% (p=0.000 n=12+11) BM_SwapElements<LargeCopyableOnly>/1 82.4ns +- 0% 81.5ns +- 0% -1.06% (p=0.000 n=12+12) ``` Ones that get worse with this CL: ``` BM_CopyTrivial/1 0.92ns +- 0% 1.15ns +- 0% +25.00% (p=0.000 n=10+9) BM_CopyTrivial/8 8.57ns +- 0% 10.72ns +- 1% +25.16% (p=0.000 n=10+12) BM_SwapElements<LargeCopyableMovableSwappable>/512 1.48ns +- 1% 1.66ns +- 1% +11.88% (p=0.000 n=12+12) BM_InlinedVectorFillString/1 11.5ns +- 0% 12.8ns +- 1% +11.62% (p=0.000 n=12+11) BM_SwapElements<LargeCopyableMovableSwappable>/64 1.48ns +- 2% 1.66ns +- 1% +11.66% (p=0.000 n=12+11) BM_SwapElements<LargeCopyableMovableSwappable>/1k 1.48ns +- 1% 1.65ns +- 2% +11.32% (p=0.000 n=12+12) BM_SwapElements<LargeCopyableMovable>/512 1.48ns +- 2% 1.58ns +- 4% +6.62% (p=0.000 n=11+12) BM_SwapElements<LargeCopyableMovable>/1k 1.49ns +- 2% 1.58ns +- 3% +6.05% (p=0.000 n=12+12) BM_SwapElements<LargeCopyableMovable>/64 1.48ns +- 2% 1.57ns +- 4% +6.04% (p=0.000 n=11+12) BM_InlinedVectorFillRange/1 4.81ns +- 0% 5.05ns +- 0% +4.83% (p=0.000 n=11+11) BM_InlinedVectorFillString/8 79.4ns +- 1% 83.1ns +- 1% +4.64% (p=0.000 n=10+12) BM_StdVectorFillString/1 16.3ns +- 0% 16.6ns +- 0% +2.13% (p=0.000 n=11+8) ``` PiperOrigin-RevId: 353906786 -- 8e26518b3cec9c598e5e9573c46c3bd1b03a67ef by Abseil Team <absl-team@google.com>: Internal change PiperOrigin-RevId: 353737330 -- f206ae0983e58c9904ed8b8f05f9caf564a446be by Matt Kulukundis <kfm@google.com>: Import of CCTZ from GitHub. PiperOrigin-RevId: 353682256 GitOrigin-RevId: c68f1886f5e8fd90eb0c2d2e68feaf00a7cdacda Change-Id: I5790c1036c4f543c701d1039848fabf7ae881ad8
4 years ago
} else if (cord_ring_enabled()) {
set_tree(CordRepRing::Append(ForceRing(force_tree(0), 1), tree));
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
} else {
set_tree(Concat(force_tree(0), tree));
}
}
void Cord::InlineRep::PrependTree(CordRep* tree) {
assert(tree != nullptr);
Export of internal Abseil changes -- 642ab296a2c9629c44f3f2ce6911cd2488bcf416 by Derek Mauro <dmauro@google.com>: Remove an obsolete check in CMakeLists.txt PiperOrigin-RevId: 352852564 -- ce78cb96bcfd162737dbcf35005da3d1d6a3486b by Abseil Team <absl-team@google.com>: Clarify that the calling *thread* must have locked the mutex in order to unlock it. PiperOrigin-RevId: 352801804 -- 24e1f5f72756046f5265abf618e951c341f09b8d by Derek Mauro <dmauro@google.com>: Fixes failing CMake string comparisons https://cmake.org/cmake/help/latest/policy/CMP0054.html Fixes #791 PiperOrigin-RevId: 352791054 -- 0ac10bc3f4dca2c4c4b51d7b8196a2eaee9537a1 by Abseil Team <absl-team@google.com>: Introduce CordRepRing class This change introduces the CordRepRing class that implements all the lower level / internal implementation for upcoming CordRepRing ring buffer support in cord. PiperOrigin-RevId: 352771994 -- 4bd36dda61760785844f0f29f26d90cc18046f75 by Abseil Team <absl-team@google.com>: Optimize InlineData representation for cord sampling (cordz) This CL changes InlineData to allow us to store a (future) Cordz Info pointer directly into the inline representation: - make InlineData a class that provides a public API to set the active union members (tree or chars) and safely access that data. - change 'tree' and 'profiled' bits to be the 2 least significant bits, allowing us 62 continquous bits for storing a Cordz Info pointer. PiperOrigin-RevId: 352642411 -- dc55ba71bbce0e6a83e05a453990c51ac3d68426 by Mark Barolak <mbar@google.com>: Add unit test coverage for the mutating overload of absl::AsciiStrToLower. PiperOrigin-RevId: 352626006 GitOrigin-RevId: 642ab296a2c9629c44f3f2ce6911cd2488bcf416 Change-Id: I6c5929dd830d3c630e14e7fd5387fc3e25a69100
4 years ago
if (data_.is_empty()) {
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
set_tree(tree);
Export of internal Abseil changes -- c68f1886f5e8fd90eb0c2d2e68feaf00a7cdacda by CJ Johnson <johnsoncj@google.com>: Introduce absl::Cleanup to the OSS repo PiperOrigin-RevId: 354583156 -- 17030cf388e10f7eb959e3e566326d1072ce392e by Abseil Team <absl-team@google.com>: Internal change only PiperOrigin-RevId: 354574953 -- e979d7236d4f3252e79ddda6739b67a9a326bf6d by CJ Johnson <johnsoncj@google.com>: Internal change PiperOrigin-RevId: 354545297 -- 7ea02b3783f7f49ef97d86a8f6580a19cc57df14 by Abseil Team <absl-team@google.com>: Pre-allocate memory for vectors where the size is known. PiperOrigin-RevId: 354344576 -- 9246c7cb11f1d6444f79ebe25acc69a8a9b870e0 by Matt Kulukundis <kfm@google.com>: Add support for Elbrus 2000 (e2k) Import of https://github.com/abseil/abseil-cpp/pull/889 PiperOrigin-RevId: 354344013 -- 0fc93d359cc1fb307552e917b37b7b2e7eed822f by Abseil Team <absl-team@google.com>: Integrate CordRepRing logic into cord (but do not enable it) PiperOrigin-RevId: 354312238 -- eda05622f7da71466723acb33403f783529df24b by Abseil Team <absl-team@google.com>: Protect ignore diagnostic with "__has_warning". PiperOrigin-RevId: 354112334 -- 47716c5d8fb10efa4fdd801d28bac414c6f8ec32 by Abseil Team <absl-team@google.com>: Rearrange InlinedVector copy constructor and destructor to treat a few special cases inline and then tail-call a non-inlined routine for the rest. In particular, we optimize for empty vectors in both cases. Added a couple of benchmarks that copy either an InlVec<int64> or an InlVec<InlVec<int64>>. Speed difference: ``` BM_CopyTrivial/0 0.92ns +- 0% 0.47ns +- 0% -48.91% (p=0.000 n=11+12) BM_CopyTrivial/1 0.92ns +- 0% 1.15ns +- 0% +25.00% (p=0.000 n=10+9) BM_CopyTrivial/8 8.57ns +- 0% 10.72ns +- 1% +25.16% (p=0.000 n=10+12) BM_CopyNonTrivial/0 3.21ns +- 0% 0.70ns +- 0% -78.23% (p=0.000 n=12+10) BM_CopyNonTrivial/1 5.88ns +- 1% 5.51ns +- 0% -6.28% (p=0.000 n=10+8) BM_CopyNonTrivial/8 21.5ns +- 1% 15.2ns +- 2% -29.23% (p=0.000 n=12+12) ``` Note: the slowdowns are a few cycles which is expected given the procedure call added in that case. We decided this is a good tradeoff given the code size reductions and the more significant speedups for empty vectors. Size difference (as measured by nm): ``` BM_CopyTrivial from 1048 bytes to 326 bytes. BM_CopyNonTrivial from 749 bytes to 470 bytes. ``` Code size for a large binary drops by ~500KB (from 349415719 to 348906015 348906191). All of the benchmarks that showed a significant difference: Ones that improve with this CL: ``` BM_CopyNonTrivial/0 3.21ns +- 0% 0.70ns +- 0% -78.23% (p=0.000 n=12+10) BM_InlinedVectorFillString/0 0.93ns +- 0% 0.24ns +- 0% -74.19% (p=0.000 n=12+10) BM_InlinedVectorAssignments/1 10.5ns +- 0% 4.1ns +- 0% -60.64% (p=0.000 n=11+10) BM_InlinedVectorAssignments/2 10.7ns +- 0% 4.4ns +- 0% -59.08% (p=0.000 n=11+11) BM_CopyTrivial/0 0.92ns +- 0% 0.47ns +- 0% -48.91% (p=0.000 n=11+12) BM_CopyNonTrivial/8 21.5ns +- 1% 15.2ns +- 2% -29.23% (p=0.000 n=12+12) BM_StdVectorEmpty 0.47ns +- 1% 0.35ns +- 0% -24.73% (p=0.000 n=12+12) BM_StdVectorSize 0.46ns +- 2% 0.35ns +- 0% -24.32% (p=0.000 n=12+12) BM_SwapElements<LargeCopyableOnly>/0 3.44ns +- 0% 2.76ns +- 1% -19.83% (p=0.000 n=11+11) BM_InlinedVectorFillRange/256 20.7ns +- 1% 17.8ns +- 0% -14.08% (p=0.000 n=12+9) BM_CopyNonTrivial/1 5.88ns +- 1% 5.51ns +- 0% -6.28% (p=0.000 n=10+8) BM_SwapElements<LargeCopyableMovable>/1 4.19ns +- 0% 3.95ns +- 1% -5.63% (p=0.000 n=11+12) BM_SwapElements<LargeCopyableMovableSwappable>/1 4.18ns +- 0% 3.99ns +- 0% -4.70% (p=0.000 n=9+11) BM_SwapElements<LargeCopyableMovable>/0 2.41ns +- 0% 2.31ns +- 0% -4.45% (p=0.000 n=12+12) BM_InlinedVectorFillRange/64 8.25ns +- 0% 8.04ns +- 0% -2.51% (p=0.000 n=12+11) BM_SwapElements<LargeCopyableOnly>/1 82.4ns +- 0% 81.5ns +- 0% -1.06% (p=0.000 n=12+12) ``` Ones that get worse with this CL: ``` BM_CopyTrivial/1 0.92ns +- 0% 1.15ns +- 0% +25.00% (p=0.000 n=10+9) BM_CopyTrivial/8 8.57ns +- 0% 10.72ns +- 1% +25.16% (p=0.000 n=10+12) BM_SwapElements<LargeCopyableMovableSwappable>/512 1.48ns +- 1% 1.66ns +- 1% +11.88% (p=0.000 n=12+12) BM_InlinedVectorFillString/1 11.5ns +- 0% 12.8ns +- 1% +11.62% (p=0.000 n=12+11) BM_SwapElements<LargeCopyableMovableSwappable>/64 1.48ns +- 2% 1.66ns +- 1% +11.66% (p=0.000 n=12+11) BM_SwapElements<LargeCopyableMovableSwappable>/1k 1.48ns +- 1% 1.65ns +- 2% +11.32% (p=0.000 n=12+12) BM_SwapElements<LargeCopyableMovable>/512 1.48ns +- 2% 1.58ns +- 4% +6.62% (p=0.000 n=11+12) BM_SwapElements<LargeCopyableMovable>/1k 1.49ns +- 2% 1.58ns +- 3% +6.05% (p=0.000 n=12+12) BM_SwapElements<LargeCopyableMovable>/64 1.48ns +- 2% 1.57ns +- 4% +6.04% (p=0.000 n=11+12) BM_InlinedVectorFillRange/1 4.81ns +- 0% 5.05ns +- 0% +4.83% (p=0.000 n=11+11) BM_InlinedVectorFillString/8 79.4ns +- 1% 83.1ns +- 1% +4.64% (p=0.000 n=10+12) BM_StdVectorFillString/1 16.3ns +- 0% 16.6ns +- 0% +2.13% (p=0.000 n=11+8) ``` PiperOrigin-RevId: 353906786 -- 8e26518b3cec9c598e5e9573c46c3bd1b03a67ef by Abseil Team <absl-team@google.com>: Internal change PiperOrigin-RevId: 353737330 -- f206ae0983e58c9904ed8b8f05f9caf564a446be by Matt Kulukundis <kfm@google.com>: Import of CCTZ from GitHub. PiperOrigin-RevId: 353682256 GitOrigin-RevId: c68f1886f5e8fd90eb0c2d2e68feaf00a7cdacda Change-Id: I5790c1036c4f543c701d1039848fabf7ae881ad8
4 years ago
} else if (cord_ring_enabled()) {
set_tree(CordRepRing::Prepend(ForceRing(force_tree(0), 1), tree));
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
} else {
set_tree(Concat(tree, force_tree(0)));
}
}
// Searches for a non-full flat node at the rightmost leaf of the tree. If a
// suitable leaf is found, the function will update the length field for all
// nodes to account for the size increase. The append region address will be
// written to region and the actual size increase will be written to size.
static inline bool PrepareAppendRegion(CordRep* root, char** region,
size_t* size, size_t max_length) {
Export of internal Abseil changes -- c68f1886f5e8fd90eb0c2d2e68feaf00a7cdacda by CJ Johnson <johnsoncj@google.com>: Introduce absl::Cleanup to the OSS repo PiperOrigin-RevId: 354583156 -- 17030cf388e10f7eb959e3e566326d1072ce392e by Abseil Team <absl-team@google.com>: Internal change only PiperOrigin-RevId: 354574953 -- e979d7236d4f3252e79ddda6739b67a9a326bf6d by CJ Johnson <johnsoncj@google.com>: Internal change PiperOrigin-RevId: 354545297 -- 7ea02b3783f7f49ef97d86a8f6580a19cc57df14 by Abseil Team <absl-team@google.com>: Pre-allocate memory for vectors where the size is known. PiperOrigin-RevId: 354344576 -- 9246c7cb11f1d6444f79ebe25acc69a8a9b870e0 by Matt Kulukundis <kfm@google.com>: Add support for Elbrus 2000 (e2k) Import of https://github.com/abseil/abseil-cpp/pull/889 PiperOrigin-RevId: 354344013 -- 0fc93d359cc1fb307552e917b37b7b2e7eed822f by Abseil Team <absl-team@google.com>: Integrate CordRepRing logic into cord (but do not enable it) PiperOrigin-RevId: 354312238 -- eda05622f7da71466723acb33403f783529df24b by Abseil Team <absl-team@google.com>: Protect ignore diagnostic with "__has_warning". PiperOrigin-RevId: 354112334 -- 47716c5d8fb10efa4fdd801d28bac414c6f8ec32 by Abseil Team <absl-team@google.com>: Rearrange InlinedVector copy constructor and destructor to treat a few special cases inline and then tail-call a non-inlined routine for the rest. In particular, we optimize for empty vectors in both cases. Added a couple of benchmarks that copy either an InlVec<int64> or an InlVec<InlVec<int64>>. Speed difference: ``` BM_CopyTrivial/0 0.92ns +- 0% 0.47ns +- 0% -48.91% (p=0.000 n=11+12) BM_CopyTrivial/1 0.92ns +- 0% 1.15ns +- 0% +25.00% (p=0.000 n=10+9) BM_CopyTrivial/8 8.57ns +- 0% 10.72ns +- 1% +25.16% (p=0.000 n=10+12) BM_CopyNonTrivial/0 3.21ns +- 0% 0.70ns +- 0% -78.23% (p=0.000 n=12+10) BM_CopyNonTrivial/1 5.88ns +- 1% 5.51ns +- 0% -6.28% (p=0.000 n=10+8) BM_CopyNonTrivial/8 21.5ns +- 1% 15.2ns +- 2% -29.23% (p=0.000 n=12+12) ``` Note: the slowdowns are a few cycles which is expected given the procedure call added in that case. We decided this is a good tradeoff given the code size reductions and the more significant speedups for empty vectors. Size difference (as measured by nm): ``` BM_CopyTrivial from 1048 bytes to 326 bytes. BM_CopyNonTrivial from 749 bytes to 470 bytes. ``` Code size for a large binary drops by ~500KB (from 349415719 to 348906015 348906191). All of the benchmarks that showed a significant difference: Ones that improve with this CL: ``` BM_CopyNonTrivial/0 3.21ns +- 0% 0.70ns +- 0% -78.23% (p=0.000 n=12+10) BM_InlinedVectorFillString/0 0.93ns +- 0% 0.24ns +- 0% -74.19% (p=0.000 n=12+10) BM_InlinedVectorAssignments/1 10.5ns +- 0% 4.1ns +- 0% -60.64% (p=0.000 n=11+10) BM_InlinedVectorAssignments/2 10.7ns +- 0% 4.4ns +- 0% -59.08% (p=0.000 n=11+11) BM_CopyTrivial/0 0.92ns +- 0% 0.47ns +- 0% -48.91% (p=0.000 n=11+12) BM_CopyNonTrivial/8 21.5ns +- 1% 15.2ns +- 2% -29.23% (p=0.000 n=12+12) BM_StdVectorEmpty 0.47ns +- 1% 0.35ns +- 0% -24.73% (p=0.000 n=12+12) BM_StdVectorSize 0.46ns +- 2% 0.35ns +- 0% -24.32% (p=0.000 n=12+12) BM_SwapElements<LargeCopyableOnly>/0 3.44ns +- 0% 2.76ns +- 1% -19.83% (p=0.000 n=11+11) BM_InlinedVectorFillRange/256 20.7ns +- 1% 17.8ns +- 0% -14.08% (p=0.000 n=12+9) BM_CopyNonTrivial/1 5.88ns +- 1% 5.51ns +- 0% -6.28% (p=0.000 n=10+8) BM_SwapElements<LargeCopyableMovable>/1 4.19ns +- 0% 3.95ns +- 1% -5.63% (p=0.000 n=11+12) BM_SwapElements<LargeCopyableMovableSwappable>/1 4.18ns +- 0% 3.99ns +- 0% -4.70% (p=0.000 n=9+11) BM_SwapElements<LargeCopyableMovable>/0 2.41ns +- 0% 2.31ns +- 0% -4.45% (p=0.000 n=12+12) BM_InlinedVectorFillRange/64 8.25ns +- 0% 8.04ns +- 0% -2.51% (p=0.000 n=12+11) BM_SwapElements<LargeCopyableOnly>/1 82.4ns +- 0% 81.5ns +- 0% -1.06% (p=0.000 n=12+12) ``` Ones that get worse with this CL: ``` BM_CopyTrivial/1 0.92ns +- 0% 1.15ns +- 0% +25.00% (p=0.000 n=10+9) BM_CopyTrivial/8 8.57ns +- 0% 10.72ns +- 1% +25.16% (p=0.000 n=10+12) BM_SwapElements<LargeCopyableMovableSwappable>/512 1.48ns +- 1% 1.66ns +- 1% +11.88% (p=0.000 n=12+12) BM_InlinedVectorFillString/1 11.5ns +- 0% 12.8ns +- 1% +11.62% (p=0.000 n=12+11) BM_SwapElements<LargeCopyableMovableSwappable>/64 1.48ns +- 2% 1.66ns +- 1% +11.66% (p=0.000 n=12+11) BM_SwapElements<LargeCopyableMovableSwappable>/1k 1.48ns +- 1% 1.65ns +- 2% +11.32% (p=0.000 n=12+12) BM_SwapElements<LargeCopyableMovable>/512 1.48ns +- 2% 1.58ns +- 4% +6.62% (p=0.000 n=11+12) BM_SwapElements<LargeCopyableMovable>/1k 1.49ns +- 2% 1.58ns +- 3% +6.05% (p=0.000 n=12+12) BM_SwapElements<LargeCopyableMovable>/64 1.48ns +- 2% 1.57ns +- 4% +6.04% (p=0.000 n=11+12) BM_InlinedVectorFillRange/1 4.81ns +- 0% 5.05ns +- 0% +4.83% (p=0.000 n=11+11) BM_InlinedVectorFillString/8 79.4ns +- 1% 83.1ns +- 1% +4.64% (p=0.000 n=10+12) BM_StdVectorFillString/1 16.3ns +- 0% 16.6ns +- 0% +2.13% (p=0.000 n=11+8) ``` PiperOrigin-RevId: 353906786 -- 8e26518b3cec9c598e5e9573c46c3bd1b03a67ef by Abseil Team <absl-team@google.com>: Internal change PiperOrigin-RevId: 353737330 -- f206ae0983e58c9904ed8b8f05f9caf564a446be by Matt Kulukundis <kfm@google.com>: Import of CCTZ from GitHub. PiperOrigin-RevId: 353682256 GitOrigin-RevId: c68f1886f5e8fd90eb0c2d2e68feaf00a7cdacda Change-Id: I5790c1036c4f543c701d1039848fabf7ae881ad8
4 years ago
if (root->tag == RING && root->refcount.IsOne()) {
Span<char> span = root->ring()->GetAppendBuffer(max_length);
if (!span.empty()) {
*region = span.data();
*size = span.size();
return true;
}
}
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
// Search down the right-hand path for a non-full FLAT node.
CordRep* dst = root;
while (dst->tag == CONCAT && dst->refcount.IsOne()) {
dst = dst->concat()->right;
}
if (dst->tag < FLAT || !dst->refcount.IsOne()) {
*region = nullptr;
*size = 0;
return false;
}
const size_t in_use = dst->length;
Export of internal Abseil changes -- ff793052bd01e1e4fcf639f94d7c30c4855a9372 by Evan Brown <ezb@google.com>: Roll forward of btree_iterator refactoring. PiperOrigin-RevId: 346116047 -- 17984679f16e3e2139b0f14fa76f4a6ca16a3ef9 by Chris Kennelly <ckennelly@google.com>: Extend absl::StrContains to accept single character needles. Single characters are more efficient to search for. Extending this API allows the abseil-string-find-str-contains Clang Tidy to include this pattern. The C++ committee has adopted http://wg21.link/P1679 for inclusion in C++23. PiperOrigin-RevId: 346095060 -- ef20b31c501b1dcaa25e244fd8f8aa43dec09bd6 by Jorg Brown <jorg@google.com>: Internal change for cord ring PiperOrigin-RevId: 346087545 -- b70f2c1cb77fc9e733a126e790967d45c5fd1dc7 by Derek Mauro <dmauro@google.com>: Release layout_benchmark PiperOrigin-RevId: 345968909 -- 3a0eda337ee43622f92cfe14c2aa06f72dc71ee5 by Derek Mauro <dmauro@google.com>: Release raw_hash_set_probe_benchmark PiperOrigin-RevId: 345965969 -- abffdb4bb241a2264cb4e73a6262b660bb10447d by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 345733599 -- 7c9e24a71188df945be17fe98f700bdb51f81b16 by Derek Mauro <dmauro@google.com>: Release hash_benchmark PiperOrigin-RevId: 345721635 -- d68f33f17f9a8cd3f6da8eee3870bdb46402cdc8 by Derek Mauro <dmauro@google.com>: Release raw_hash_set_benchmark PiperOrigin-RevId: 345708384 -- 6e6c547d4d1327b226c0ffe8ff34d0aa103ce24b by Abseil Team <absl-team@google.com>: Updates the implementation of InlinedVector to accurately express the value-initialization semantics of the default constructor PiperOrigin-RevId: 345548260 -- 1532424deda97d468444c217cc0fa4614099c7c1 by Evan Brown <ezb@google.com>: Rollback btree_iterator refactoring. PiperOrigin-RevId: 345543900 GitOrigin-RevId: ff793052bd01e1e4fcf639f94d7c30c4855a9372 Change-Id: I719831981fd056de41939f9addfee3d85e3b49b2
4 years ago
const size_t capacity = dst->flat()->Capacity();
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
if (in_use == capacity) {
*region = nullptr;
*size = 0;
return false;
}
size_t size_increase = std::min(capacity - in_use, max_length);
// We need to update the length fields for all nodes, including the leaf node.
for (CordRep* rep = root; rep != dst; rep = rep->concat()->right) {
rep->length += size_increase;
}
dst->length += size_increase;
Export of internal Abseil changes -- a0491c8d790972cd80e2d720fe1fdf5f711a6f1a by Greg Falcon <gfalcon@google.com>: Stop directly accessing CordRepFlat data via CordRep::data. The old pattern of access breaks the `CordRep` type abstraction; since `CordRep::data` is not in general guaranteed to contain the chunk's data, we shouldn't access it that way. This incidentally adds an assertion check (via the flat() accessor) that the CordRep is indeed flat on each such access, but a manual inspection of the code, as well as the fact that this code currently works, suggest that this is always true.) PiperOrigin-RevId: 351592344 -- f40c3b43ca5b1d7e23cd45f1ffac1783105ac1a3 by Abseil Team <absl-team@google.com>: Revert 18abb2902b9f06c63a968b24d3dda785ebf99a22 PiperOrigin-RevId: 351523518 -- 18abb2902b9f06c63a968b24d3dda785ebf99a22 by Abseil Team <absl-team@google.com>: Internal change PiperOrigin-RevId: 351512412 -- 9b881602d45e95e06089792c7627cd56528a255a by Abseil Team <absl-team@google.com>: Keep time's global state in a cacheline-aligned structure. Keeping the global state as separate global variables results in two issues: 1) False sharing with adjacent global data (e.g., cycle clock source), since the global fields are updated every O(10usec). 2) The hot global fields (e.g., seq and samples) can reside on different cache lines. To fix this, simply wrap the global data in a ABSL_CACHE_ALIGNED structure. This is similar to what we do for MutexGlobals. PiperOrigin-RevId: 351389466 GitOrigin-RevId: a0491c8d790972cd80e2d720fe1fdf5f711a6f1a Change-Id: Ie0fa80112043381cd37c84e2ab2b7334839f54b5
4 years ago
*region = dst->flat()->Data() + in_use;
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
*size = size_increase;
return true;
}
void Cord::InlineRep::GetAppendRegion(char** region, size_t* size,
size_t max_length) {
if (max_length == 0) {
*region = nullptr;
*size = 0;
return;
}
// Try to fit in the inline buffer if possible.
Export of internal Abseil changes -- 642ab296a2c9629c44f3f2ce6911cd2488bcf416 by Derek Mauro <dmauro@google.com>: Remove an obsolete check in CMakeLists.txt PiperOrigin-RevId: 352852564 -- ce78cb96bcfd162737dbcf35005da3d1d6a3486b by Abseil Team <absl-team@google.com>: Clarify that the calling *thread* must have locked the mutex in order to unlock it. PiperOrigin-RevId: 352801804 -- 24e1f5f72756046f5265abf618e951c341f09b8d by Derek Mauro <dmauro@google.com>: Fixes failing CMake string comparisons https://cmake.org/cmake/help/latest/policy/CMP0054.html Fixes #791 PiperOrigin-RevId: 352791054 -- 0ac10bc3f4dca2c4c4b51d7b8196a2eaee9537a1 by Abseil Team <absl-team@google.com>: Introduce CordRepRing class This change introduces the CordRepRing class that implements all the lower level / internal implementation for upcoming CordRepRing ring buffer support in cord. PiperOrigin-RevId: 352771994 -- 4bd36dda61760785844f0f29f26d90cc18046f75 by Abseil Team <absl-team@google.com>: Optimize InlineData representation for cord sampling (cordz) This CL changes InlineData to allow us to store a (future) Cordz Info pointer directly into the inline representation: - make InlineData a class that provides a public API to set the active union members (tree or chars) and safely access that data. - change 'tree' and 'profiled' bits to be the 2 least significant bits, allowing us 62 continquous bits for storing a Cordz Info pointer. PiperOrigin-RevId: 352642411 -- dc55ba71bbce0e6a83e05a453990c51ac3d68426 by Mark Barolak <mbar@google.com>: Add unit test coverage for the mutating overload of absl::AsciiStrToLower. PiperOrigin-RevId: 352626006 GitOrigin-RevId: 642ab296a2c9629c44f3f2ce6911cd2488bcf416 Change-Id: I6c5929dd830d3c630e14e7fd5387fc3e25a69100
4 years ago
if (!is_tree()) {
size_t inline_length = inline_size();
if (max_length <= kMaxInline - inline_length) {
*region = data_.as_chars() + inline_length;
*size = max_length;
set_inline_size(inline_length + max_length);
return;
}
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
}
CordRep* root = force_tree(max_length);
if (PrepareAppendRegion(root, region, size, max_length)) {
Export of internal Abseil changes -- 341670bce317dd6af8d3c066970230591a47e80c by Martijn Vels <mvels@google.com>: Change GetStack() and GetParentStack() to return absl::Span PiperOrigin-RevId: 368765721 -- 6aaab9536d6957303c7aba100c3afaa6fb0ea2c8 by Martijn Vels <mvels@google.com>: Remove locking from parent stack. This change removes the need to lock all access to `parent_stack' by making the 'copy constructor' logic specify the 'copied from' CordzInfo (where available) to the TrackCord function, after which parent_stack is immutable. PiperOrigin-RevId: 368760630 -- b19e2059cada35a8ede994833018edac94de6ddc by Martijn Vels <mvels@google.com>: Add cordz instrumentation to Cord PiperOrigin-RevId: 368746225 -- 67b8bbf980f0f4e1db79aa32968e9a715a09b51a by Martijn Vels <mvels@google.com>: Create ABSL_INTERNAL_CORDZ_ENABLED define controlling when Cordz code is enabled There are specific builds and condtions under which we don't support cordz sampling, which is per this change represented by ABSL_INTERNAL_CORDZ_ENABLED being defined. PiperOrigin-RevId: 368731603 -- 8cbfe0e3169637a620f4b66ad2bc2ce340879cb0 by Martijn Vels <mvels@google.com>: Add a `rep` property to CordzInfo to be managed by Cord logic. This change adds a `rep` property to CordzInfo, which is intended to be used by collection logic. Mini design: Cord invokes TrackCord() providing the active 'root' cordrep of the newly sampled Cord, returning a CordzInfo with a weak (uncounted) reference to this root. Cord invokes `SetCordRep()` each time the root cordrep of the sampled Cord is updated while holding `mutex()`. Cord must also obtain `mutex()` _before_ removing a reference on the old root. i.e.: Cord must guarantee that the (weak) reference held in CordzInfo is at all times valid. CordzInfo collection code can then safely obtain a (reference counted) rep pointer by adding a reference to `rep_` while holding `mutex()`. This requires only a very brief critical section inside CordzInfo logic, minimizing contention with concurrent Cord updates. Cord code should typically obtain and hold `mutex()` for the entirety of each mutating Cord operation on a sampled cord. As Cord is thread compatible, it never competes on the lock with any other thread. The only possible concurrent access is from Cordz collection code, which should be a relatively rare event. PiperOrigin-RevId: 368673758 -- 1255120dce2bdd6b4205a34a0e555e0b74b6152f by Martijn Vels <mvels@google.com>: Remove 'depth' from active recorded metrics. Going forward we do not 'live' record depth (and size), but will observe these at collection time only. PiperOrigin-RevId: 368636572 -- 83e5146e35f221736b49e9f0a8805f8c159a51db by Martijn Vels <mvels@google.com>: Make cordz targets visible in OSS PiperOrigin-RevId: 368615010 -- dcb16a4f1239151f0a8c70a8cfeb29dabbd113b8 by Martijn Vels <mvels@google.com>: Internal cleanup PiperOrigin-RevId: 368514666 GitOrigin-RevId: 341670bce317dd6af8d3c066970230591a47e80c Change-Id: I94cecfbbd441eb386f99fc5186c468a7a5538862
4 years ago
UpdateCordzStatistics();
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
return;
}
// Allocate new node.
CordRepFlat* new_node =
CordRepFlat::New(std::max(static_cast<size_t>(root->length), max_length));
new_node->length = std::min(new_node->Capacity(), max_length);
Export of internal Abseil changes -- a0491c8d790972cd80e2d720fe1fdf5f711a6f1a by Greg Falcon <gfalcon@google.com>: Stop directly accessing CordRepFlat data via CordRep::data. The old pattern of access breaks the `CordRep` type abstraction; since `CordRep::data` is not in general guaranteed to contain the chunk's data, we shouldn't access it that way. This incidentally adds an assertion check (via the flat() accessor) that the CordRep is indeed flat on each such access, but a manual inspection of the code, as well as the fact that this code currently works, suggest that this is always true.) PiperOrigin-RevId: 351592344 -- f40c3b43ca5b1d7e23cd45f1ffac1783105ac1a3 by Abseil Team <absl-team@google.com>: Revert 18abb2902b9f06c63a968b24d3dda785ebf99a22 PiperOrigin-RevId: 351523518 -- 18abb2902b9f06c63a968b24d3dda785ebf99a22 by Abseil Team <absl-team@google.com>: Internal change PiperOrigin-RevId: 351512412 -- 9b881602d45e95e06089792c7627cd56528a255a by Abseil Team <absl-team@google.com>: Keep time's global state in a cacheline-aligned structure. Keeping the global state as separate global variables results in two issues: 1) False sharing with adjacent global data (e.g., cycle clock source), since the global fields are updated every O(10usec). 2) The hot global fields (e.g., seq and samples) can reside on different cache lines. To fix this, simply wrap the global data in a ABSL_CACHE_ALIGNED structure. This is similar to what we do for MutexGlobals. PiperOrigin-RevId: 351389466 GitOrigin-RevId: a0491c8d790972cd80e2d720fe1fdf5f711a6f1a Change-Id: Ie0fa80112043381cd37c84e2ab2b7334839f54b5
4 years ago
*region = new_node->Data();
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
*size = new_node->length;
Export of internal Abseil changes -- c68f1886f5e8fd90eb0c2d2e68feaf00a7cdacda by CJ Johnson <johnsoncj@google.com>: Introduce absl::Cleanup to the OSS repo PiperOrigin-RevId: 354583156 -- 17030cf388e10f7eb959e3e566326d1072ce392e by Abseil Team <absl-team@google.com>: Internal change only PiperOrigin-RevId: 354574953 -- e979d7236d4f3252e79ddda6739b67a9a326bf6d by CJ Johnson <johnsoncj@google.com>: Internal change PiperOrigin-RevId: 354545297 -- 7ea02b3783f7f49ef97d86a8f6580a19cc57df14 by Abseil Team <absl-team@google.com>: Pre-allocate memory for vectors where the size is known. PiperOrigin-RevId: 354344576 -- 9246c7cb11f1d6444f79ebe25acc69a8a9b870e0 by Matt Kulukundis <kfm@google.com>: Add support for Elbrus 2000 (e2k) Import of https://github.com/abseil/abseil-cpp/pull/889 PiperOrigin-RevId: 354344013 -- 0fc93d359cc1fb307552e917b37b7b2e7eed822f by Abseil Team <absl-team@google.com>: Integrate CordRepRing logic into cord (but do not enable it) PiperOrigin-RevId: 354312238 -- eda05622f7da71466723acb33403f783529df24b by Abseil Team <absl-team@google.com>: Protect ignore diagnostic with "__has_warning". PiperOrigin-RevId: 354112334 -- 47716c5d8fb10efa4fdd801d28bac414c6f8ec32 by Abseil Team <absl-team@google.com>: Rearrange InlinedVector copy constructor and destructor to treat a few special cases inline and then tail-call a non-inlined routine for the rest. In particular, we optimize for empty vectors in both cases. Added a couple of benchmarks that copy either an InlVec<int64> or an InlVec<InlVec<int64>>. Speed difference: ``` BM_CopyTrivial/0 0.92ns +- 0% 0.47ns +- 0% -48.91% (p=0.000 n=11+12) BM_CopyTrivial/1 0.92ns +- 0% 1.15ns +- 0% +25.00% (p=0.000 n=10+9) BM_CopyTrivial/8 8.57ns +- 0% 10.72ns +- 1% +25.16% (p=0.000 n=10+12) BM_CopyNonTrivial/0 3.21ns +- 0% 0.70ns +- 0% -78.23% (p=0.000 n=12+10) BM_CopyNonTrivial/1 5.88ns +- 1% 5.51ns +- 0% -6.28% (p=0.000 n=10+8) BM_CopyNonTrivial/8 21.5ns +- 1% 15.2ns +- 2% -29.23% (p=0.000 n=12+12) ``` Note: the slowdowns are a few cycles which is expected given the procedure call added in that case. We decided this is a good tradeoff given the code size reductions and the more significant speedups for empty vectors. Size difference (as measured by nm): ``` BM_CopyTrivial from 1048 bytes to 326 bytes. BM_CopyNonTrivial from 749 bytes to 470 bytes. ``` Code size for a large binary drops by ~500KB (from 349415719 to 348906015 348906191). All of the benchmarks that showed a significant difference: Ones that improve with this CL: ``` BM_CopyNonTrivial/0 3.21ns +- 0% 0.70ns +- 0% -78.23% (p=0.000 n=12+10) BM_InlinedVectorFillString/0 0.93ns +- 0% 0.24ns +- 0% -74.19% (p=0.000 n=12+10) BM_InlinedVectorAssignments/1 10.5ns +- 0% 4.1ns +- 0% -60.64% (p=0.000 n=11+10) BM_InlinedVectorAssignments/2 10.7ns +- 0% 4.4ns +- 0% -59.08% (p=0.000 n=11+11) BM_CopyTrivial/0 0.92ns +- 0% 0.47ns +- 0% -48.91% (p=0.000 n=11+12) BM_CopyNonTrivial/8 21.5ns +- 1% 15.2ns +- 2% -29.23% (p=0.000 n=12+12) BM_StdVectorEmpty 0.47ns +- 1% 0.35ns +- 0% -24.73% (p=0.000 n=12+12) BM_StdVectorSize 0.46ns +- 2% 0.35ns +- 0% -24.32% (p=0.000 n=12+12) BM_SwapElements<LargeCopyableOnly>/0 3.44ns +- 0% 2.76ns +- 1% -19.83% (p=0.000 n=11+11) BM_InlinedVectorFillRange/256 20.7ns +- 1% 17.8ns +- 0% -14.08% (p=0.000 n=12+9) BM_CopyNonTrivial/1 5.88ns +- 1% 5.51ns +- 0% -6.28% (p=0.000 n=10+8) BM_SwapElements<LargeCopyableMovable>/1 4.19ns +- 0% 3.95ns +- 1% -5.63% (p=0.000 n=11+12) BM_SwapElements<LargeCopyableMovableSwappable>/1 4.18ns +- 0% 3.99ns +- 0% -4.70% (p=0.000 n=9+11) BM_SwapElements<LargeCopyableMovable>/0 2.41ns +- 0% 2.31ns +- 0% -4.45% (p=0.000 n=12+12) BM_InlinedVectorFillRange/64 8.25ns +- 0% 8.04ns +- 0% -2.51% (p=0.000 n=12+11) BM_SwapElements<LargeCopyableOnly>/1 82.4ns +- 0% 81.5ns +- 0% -1.06% (p=0.000 n=12+12) ``` Ones that get worse with this CL: ``` BM_CopyTrivial/1 0.92ns +- 0% 1.15ns +- 0% +25.00% (p=0.000 n=10+9) BM_CopyTrivial/8 8.57ns +- 0% 10.72ns +- 1% +25.16% (p=0.000 n=10+12) BM_SwapElements<LargeCopyableMovableSwappable>/512 1.48ns +- 1% 1.66ns +- 1% +11.88% (p=0.000 n=12+12) BM_InlinedVectorFillString/1 11.5ns +- 0% 12.8ns +- 1% +11.62% (p=0.000 n=12+11) BM_SwapElements<LargeCopyableMovableSwappable>/64 1.48ns +- 2% 1.66ns +- 1% +11.66% (p=0.000 n=12+11) BM_SwapElements<LargeCopyableMovableSwappable>/1k 1.48ns +- 1% 1.65ns +- 2% +11.32% (p=0.000 n=12+12) BM_SwapElements<LargeCopyableMovable>/512 1.48ns +- 2% 1.58ns +- 4% +6.62% (p=0.000 n=11+12) BM_SwapElements<LargeCopyableMovable>/1k 1.49ns +- 2% 1.58ns +- 3% +6.05% (p=0.000 n=12+12) BM_SwapElements<LargeCopyableMovable>/64 1.48ns +- 2% 1.57ns +- 4% +6.04% (p=0.000 n=11+12) BM_InlinedVectorFillRange/1 4.81ns +- 0% 5.05ns +- 0% +4.83% (p=0.000 n=11+11) BM_InlinedVectorFillString/8 79.4ns +- 1% 83.1ns +- 1% +4.64% (p=0.000 n=10+12) BM_StdVectorFillString/1 16.3ns +- 0% 16.6ns +- 0% +2.13% (p=0.000 n=11+8) ``` PiperOrigin-RevId: 353906786 -- 8e26518b3cec9c598e5e9573c46c3bd1b03a67ef by Abseil Team <absl-team@google.com>: Internal change PiperOrigin-RevId: 353737330 -- f206ae0983e58c9904ed8b8f05f9caf564a446be by Matt Kulukundis <kfm@google.com>: Import of CCTZ from GitHub. PiperOrigin-RevId: 353682256 GitOrigin-RevId: c68f1886f5e8fd90eb0c2d2e68feaf00a7cdacda Change-Id: I5790c1036c4f543c701d1039848fabf7ae881ad8
4 years ago
if (cord_ring_enabled()) {
replace_tree(CordRepRing::Append(ForceRing(root, 1), new_node));
return;
}
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
replace_tree(Concat(root, new_node));
}
void Cord::InlineRep::GetAppendRegion(char** region, size_t* size) {
const size_t max_length = std::numeric_limits<size_t>::max();
// Try to fit in the inline buffer if possible.
Export of internal Abseil changes -- 642ab296a2c9629c44f3f2ce6911cd2488bcf416 by Derek Mauro <dmauro@google.com>: Remove an obsolete check in CMakeLists.txt PiperOrigin-RevId: 352852564 -- ce78cb96bcfd162737dbcf35005da3d1d6a3486b by Abseil Team <absl-team@google.com>: Clarify that the calling *thread* must have locked the mutex in order to unlock it. PiperOrigin-RevId: 352801804 -- 24e1f5f72756046f5265abf618e951c341f09b8d by Derek Mauro <dmauro@google.com>: Fixes failing CMake string comparisons https://cmake.org/cmake/help/latest/policy/CMP0054.html Fixes #791 PiperOrigin-RevId: 352791054 -- 0ac10bc3f4dca2c4c4b51d7b8196a2eaee9537a1 by Abseil Team <absl-team@google.com>: Introduce CordRepRing class This change introduces the CordRepRing class that implements all the lower level / internal implementation for upcoming CordRepRing ring buffer support in cord. PiperOrigin-RevId: 352771994 -- 4bd36dda61760785844f0f29f26d90cc18046f75 by Abseil Team <absl-team@google.com>: Optimize InlineData representation for cord sampling (cordz) This CL changes InlineData to allow us to store a (future) Cordz Info pointer directly into the inline representation: - make InlineData a class that provides a public API to set the active union members (tree or chars) and safely access that data. - change 'tree' and 'profiled' bits to be the 2 least significant bits, allowing us 62 continquous bits for storing a Cordz Info pointer. PiperOrigin-RevId: 352642411 -- dc55ba71bbce0e6a83e05a453990c51ac3d68426 by Mark Barolak <mbar@google.com>: Add unit test coverage for the mutating overload of absl::AsciiStrToLower. PiperOrigin-RevId: 352626006 GitOrigin-RevId: 642ab296a2c9629c44f3f2ce6911cd2488bcf416 Change-Id: I6c5929dd830d3c630e14e7fd5387fc3e25a69100
4 years ago
if (!data_.is_tree()) {
size_t inline_length = inline_size();
if (inline_length < kMaxInline) {
*region = data_.as_chars() + inline_length;
*size = kMaxInline - inline_length;
set_inline_size(kMaxInline);
return;
}
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
}
CordRep* root = force_tree(max_length);
if (PrepareAppendRegion(root, region, size, max_length)) {
Export of internal Abseil changes -- 341670bce317dd6af8d3c066970230591a47e80c by Martijn Vels <mvels@google.com>: Change GetStack() and GetParentStack() to return absl::Span PiperOrigin-RevId: 368765721 -- 6aaab9536d6957303c7aba100c3afaa6fb0ea2c8 by Martijn Vels <mvels@google.com>: Remove locking from parent stack. This change removes the need to lock all access to `parent_stack' by making the 'copy constructor' logic specify the 'copied from' CordzInfo (where available) to the TrackCord function, after which parent_stack is immutable. PiperOrigin-RevId: 368760630 -- b19e2059cada35a8ede994833018edac94de6ddc by Martijn Vels <mvels@google.com>: Add cordz instrumentation to Cord PiperOrigin-RevId: 368746225 -- 67b8bbf980f0f4e1db79aa32968e9a715a09b51a by Martijn Vels <mvels@google.com>: Create ABSL_INTERNAL_CORDZ_ENABLED define controlling when Cordz code is enabled There are specific builds and condtions under which we don't support cordz sampling, which is per this change represented by ABSL_INTERNAL_CORDZ_ENABLED being defined. PiperOrigin-RevId: 368731603 -- 8cbfe0e3169637a620f4b66ad2bc2ce340879cb0 by Martijn Vels <mvels@google.com>: Add a `rep` property to CordzInfo to be managed by Cord logic. This change adds a `rep` property to CordzInfo, which is intended to be used by collection logic. Mini design: Cord invokes TrackCord() providing the active 'root' cordrep of the newly sampled Cord, returning a CordzInfo with a weak (uncounted) reference to this root. Cord invokes `SetCordRep()` each time the root cordrep of the sampled Cord is updated while holding `mutex()`. Cord must also obtain `mutex()` _before_ removing a reference on the old root. i.e.: Cord must guarantee that the (weak) reference held in CordzInfo is at all times valid. CordzInfo collection code can then safely obtain a (reference counted) rep pointer by adding a reference to `rep_` while holding `mutex()`. This requires only a very brief critical section inside CordzInfo logic, minimizing contention with concurrent Cord updates. Cord code should typically obtain and hold `mutex()` for the entirety of each mutating Cord operation on a sampled cord. As Cord is thread compatible, it never competes on the lock with any other thread. The only possible concurrent access is from Cordz collection code, which should be a relatively rare event. PiperOrigin-RevId: 368673758 -- 1255120dce2bdd6b4205a34a0e555e0b74b6152f by Martijn Vels <mvels@google.com>: Remove 'depth' from active recorded metrics. Going forward we do not 'live' record depth (and size), but will observe these at collection time only. PiperOrigin-RevId: 368636572 -- 83e5146e35f221736b49e9f0a8805f8c159a51db by Martijn Vels <mvels@google.com>: Make cordz targets visible in OSS PiperOrigin-RevId: 368615010 -- dcb16a4f1239151f0a8c70a8cfeb29dabbd113b8 by Martijn Vels <mvels@google.com>: Internal cleanup PiperOrigin-RevId: 368514666 GitOrigin-RevId: 341670bce317dd6af8d3c066970230591a47e80c Change-Id: I94cecfbbd441eb386f99fc5186c468a7a5538862
4 years ago
UpdateCordzStatistics();
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
return;
}
// Allocate new node.
CordRepFlat* new_node = CordRepFlat::New(root->length);
new_node->length = new_node->Capacity();
Export of internal Abseil changes -- a0491c8d790972cd80e2d720fe1fdf5f711a6f1a by Greg Falcon <gfalcon@google.com>: Stop directly accessing CordRepFlat data via CordRep::data. The old pattern of access breaks the `CordRep` type abstraction; since `CordRep::data` is not in general guaranteed to contain the chunk's data, we shouldn't access it that way. This incidentally adds an assertion check (via the flat() accessor) that the CordRep is indeed flat on each such access, but a manual inspection of the code, as well as the fact that this code currently works, suggest that this is always true.) PiperOrigin-RevId: 351592344 -- f40c3b43ca5b1d7e23cd45f1ffac1783105ac1a3 by Abseil Team <absl-team@google.com>: Revert 18abb2902b9f06c63a968b24d3dda785ebf99a22 PiperOrigin-RevId: 351523518 -- 18abb2902b9f06c63a968b24d3dda785ebf99a22 by Abseil Team <absl-team@google.com>: Internal change PiperOrigin-RevId: 351512412 -- 9b881602d45e95e06089792c7627cd56528a255a by Abseil Team <absl-team@google.com>: Keep time's global state in a cacheline-aligned structure. Keeping the global state as separate global variables results in two issues: 1) False sharing with adjacent global data (e.g., cycle clock source), since the global fields are updated every O(10usec). 2) The hot global fields (e.g., seq and samples) can reside on different cache lines. To fix this, simply wrap the global data in a ABSL_CACHE_ALIGNED structure. This is similar to what we do for MutexGlobals. PiperOrigin-RevId: 351389466 GitOrigin-RevId: a0491c8d790972cd80e2d720fe1fdf5f711a6f1a Change-Id: Ie0fa80112043381cd37c84e2ab2b7334839f54b5
4 years ago
*region = new_node->Data();
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
*size = new_node->length;
Export of internal Abseil changes -- c68f1886f5e8fd90eb0c2d2e68feaf00a7cdacda by CJ Johnson <johnsoncj@google.com>: Introduce absl::Cleanup to the OSS repo PiperOrigin-RevId: 354583156 -- 17030cf388e10f7eb959e3e566326d1072ce392e by Abseil Team <absl-team@google.com>: Internal change only PiperOrigin-RevId: 354574953 -- e979d7236d4f3252e79ddda6739b67a9a326bf6d by CJ Johnson <johnsoncj@google.com>: Internal change PiperOrigin-RevId: 354545297 -- 7ea02b3783f7f49ef97d86a8f6580a19cc57df14 by Abseil Team <absl-team@google.com>: Pre-allocate memory for vectors where the size is known. PiperOrigin-RevId: 354344576 -- 9246c7cb11f1d6444f79ebe25acc69a8a9b870e0 by Matt Kulukundis <kfm@google.com>: Add support for Elbrus 2000 (e2k) Import of https://github.com/abseil/abseil-cpp/pull/889 PiperOrigin-RevId: 354344013 -- 0fc93d359cc1fb307552e917b37b7b2e7eed822f by Abseil Team <absl-team@google.com>: Integrate CordRepRing logic into cord (but do not enable it) PiperOrigin-RevId: 354312238 -- eda05622f7da71466723acb33403f783529df24b by Abseil Team <absl-team@google.com>: Protect ignore diagnostic with "__has_warning". PiperOrigin-RevId: 354112334 -- 47716c5d8fb10efa4fdd801d28bac414c6f8ec32 by Abseil Team <absl-team@google.com>: Rearrange InlinedVector copy constructor and destructor to treat a few special cases inline and then tail-call a non-inlined routine for the rest. In particular, we optimize for empty vectors in both cases. Added a couple of benchmarks that copy either an InlVec<int64> or an InlVec<InlVec<int64>>. Speed difference: ``` BM_CopyTrivial/0 0.92ns +- 0% 0.47ns +- 0% -48.91% (p=0.000 n=11+12) BM_CopyTrivial/1 0.92ns +- 0% 1.15ns +- 0% +25.00% (p=0.000 n=10+9) BM_CopyTrivial/8 8.57ns +- 0% 10.72ns +- 1% +25.16% (p=0.000 n=10+12) BM_CopyNonTrivial/0 3.21ns +- 0% 0.70ns +- 0% -78.23% (p=0.000 n=12+10) BM_CopyNonTrivial/1 5.88ns +- 1% 5.51ns +- 0% -6.28% (p=0.000 n=10+8) BM_CopyNonTrivial/8 21.5ns +- 1% 15.2ns +- 2% -29.23% (p=0.000 n=12+12) ``` Note: the slowdowns are a few cycles which is expected given the procedure call added in that case. We decided this is a good tradeoff given the code size reductions and the more significant speedups for empty vectors. Size difference (as measured by nm): ``` BM_CopyTrivial from 1048 bytes to 326 bytes. BM_CopyNonTrivial from 749 bytes to 470 bytes. ``` Code size for a large binary drops by ~500KB (from 349415719 to 348906015 348906191). All of the benchmarks that showed a significant difference: Ones that improve with this CL: ``` BM_CopyNonTrivial/0 3.21ns +- 0% 0.70ns +- 0% -78.23% (p=0.000 n=12+10) BM_InlinedVectorFillString/0 0.93ns +- 0% 0.24ns +- 0% -74.19% (p=0.000 n=12+10) BM_InlinedVectorAssignments/1 10.5ns +- 0% 4.1ns +- 0% -60.64% (p=0.000 n=11+10) BM_InlinedVectorAssignments/2 10.7ns +- 0% 4.4ns +- 0% -59.08% (p=0.000 n=11+11) BM_CopyTrivial/0 0.92ns +- 0% 0.47ns +- 0% -48.91% (p=0.000 n=11+12) BM_CopyNonTrivial/8 21.5ns +- 1% 15.2ns +- 2% -29.23% (p=0.000 n=12+12) BM_StdVectorEmpty 0.47ns +- 1% 0.35ns +- 0% -24.73% (p=0.000 n=12+12) BM_StdVectorSize 0.46ns +- 2% 0.35ns +- 0% -24.32% (p=0.000 n=12+12) BM_SwapElements<LargeCopyableOnly>/0 3.44ns +- 0% 2.76ns +- 1% -19.83% (p=0.000 n=11+11) BM_InlinedVectorFillRange/256 20.7ns +- 1% 17.8ns +- 0% -14.08% (p=0.000 n=12+9) BM_CopyNonTrivial/1 5.88ns +- 1% 5.51ns +- 0% -6.28% (p=0.000 n=10+8) BM_SwapElements<LargeCopyableMovable>/1 4.19ns +- 0% 3.95ns +- 1% -5.63% (p=0.000 n=11+12) BM_SwapElements<LargeCopyableMovableSwappable>/1 4.18ns +- 0% 3.99ns +- 0% -4.70% (p=0.000 n=9+11) BM_SwapElements<LargeCopyableMovable>/0 2.41ns +- 0% 2.31ns +- 0% -4.45% (p=0.000 n=12+12) BM_InlinedVectorFillRange/64 8.25ns +- 0% 8.04ns +- 0% -2.51% (p=0.000 n=12+11) BM_SwapElements<LargeCopyableOnly>/1 82.4ns +- 0% 81.5ns +- 0% -1.06% (p=0.000 n=12+12) ``` Ones that get worse with this CL: ``` BM_CopyTrivial/1 0.92ns +- 0% 1.15ns +- 0% +25.00% (p=0.000 n=10+9) BM_CopyTrivial/8 8.57ns +- 0% 10.72ns +- 1% +25.16% (p=0.000 n=10+12) BM_SwapElements<LargeCopyableMovableSwappable>/512 1.48ns +- 1% 1.66ns +- 1% +11.88% (p=0.000 n=12+12) BM_InlinedVectorFillString/1 11.5ns +- 0% 12.8ns +- 1% +11.62% (p=0.000 n=12+11) BM_SwapElements<LargeCopyableMovableSwappable>/64 1.48ns +- 2% 1.66ns +- 1% +11.66% (p=0.000 n=12+11) BM_SwapElements<LargeCopyableMovableSwappable>/1k 1.48ns +- 1% 1.65ns +- 2% +11.32% (p=0.000 n=12+12) BM_SwapElements<LargeCopyableMovable>/512 1.48ns +- 2% 1.58ns +- 4% +6.62% (p=0.000 n=11+12) BM_SwapElements<LargeCopyableMovable>/1k 1.49ns +- 2% 1.58ns +- 3% +6.05% (p=0.000 n=12+12) BM_SwapElements<LargeCopyableMovable>/64 1.48ns +- 2% 1.57ns +- 4% +6.04% (p=0.000 n=11+12) BM_InlinedVectorFillRange/1 4.81ns +- 0% 5.05ns +- 0% +4.83% (p=0.000 n=11+11) BM_InlinedVectorFillString/8 79.4ns +- 1% 83.1ns +- 1% +4.64% (p=0.000 n=10+12) BM_StdVectorFillString/1 16.3ns +- 0% 16.6ns +- 0% +2.13% (p=0.000 n=11+8) ``` PiperOrigin-RevId: 353906786 -- 8e26518b3cec9c598e5e9573c46c3bd1b03a67ef by Abseil Team <absl-team@google.com>: Internal change PiperOrigin-RevId: 353737330 -- f206ae0983e58c9904ed8b8f05f9caf564a446be by Matt Kulukundis <kfm@google.com>: Import of CCTZ from GitHub. PiperOrigin-RevId: 353682256 GitOrigin-RevId: c68f1886f5e8fd90eb0c2d2e68feaf00a7cdacda Change-Id: I5790c1036c4f543c701d1039848fabf7ae881ad8
4 years ago
if (cord_ring_enabled()) {
replace_tree(CordRepRing::Append(ForceRing(root, 1), new_node));
return;
}
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
replace_tree(Concat(root, new_node));
}
// If the rep is a leaf, this will increment the value at total_mem_usage and
// will return true.
static bool RepMemoryUsageLeaf(const CordRep* rep, size_t* total_mem_usage) {
if (rep->tag >= FLAT) {
Export of internal Abseil changes -- ff793052bd01e1e4fcf639f94d7c30c4855a9372 by Evan Brown <ezb@google.com>: Roll forward of btree_iterator refactoring. PiperOrigin-RevId: 346116047 -- 17984679f16e3e2139b0f14fa76f4a6ca16a3ef9 by Chris Kennelly <ckennelly@google.com>: Extend absl::StrContains to accept single character needles. Single characters are more efficient to search for. Extending this API allows the abseil-string-find-str-contains Clang Tidy to include this pattern. The C++ committee has adopted http://wg21.link/P1679 for inclusion in C++23. PiperOrigin-RevId: 346095060 -- ef20b31c501b1dcaa25e244fd8f8aa43dec09bd6 by Jorg Brown <jorg@google.com>: Internal change for cord ring PiperOrigin-RevId: 346087545 -- b70f2c1cb77fc9e733a126e790967d45c5fd1dc7 by Derek Mauro <dmauro@google.com>: Release layout_benchmark PiperOrigin-RevId: 345968909 -- 3a0eda337ee43622f92cfe14c2aa06f72dc71ee5 by Derek Mauro <dmauro@google.com>: Release raw_hash_set_probe_benchmark PiperOrigin-RevId: 345965969 -- abffdb4bb241a2264cb4e73a6262b660bb10447d by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 345733599 -- 7c9e24a71188df945be17fe98f700bdb51f81b16 by Derek Mauro <dmauro@google.com>: Release hash_benchmark PiperOrigin-RevId: 345721635 -- d68f33f17f9a8cd3f6da8eee3870bdb46402cdc8 by Derek Mauro <dmauro@google.com>: Release raw_hash_set_benchmark PiperOrigin-RevId: 345708384 -- 6e6c547d4d1327b226c0ffe8ff34d0aa103ce24b by Abseil Team <absl-team@google.com>: Updates the implementation of InlinedVector to accurately express the value-initialization semantics of the default constructor PiperOrigin-RevId: 345548260 -- 1532424deda97d468444c217cc0fa4614099c7c1 by Evan Brown <ezb@google.com>: Rollback btree_iterator refactoring. PiperOrigin-RevId: 345543900 GitOrigin-RevId: ff793052bd01e1e4fcf639f94d7c30c4855a9372 Change-Id: I719831981fd056de41939f9addfee3d85e3b49b2
4 years ago
*total_mem_usage += rep->flat()->AllocatedSize();
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
return true;
}
if (rep->tag == EXTERNAL) {
*total_mem_usage += sizeof(CordRepConcat) + rep->length;
return true;
}
return false;
}
Export of internal Abseil changes -- 341670bce317dd6af8d3c066970230591a47e80c by Martijn Vels <mvels@google.com>: Change GetStack() and GetParentStack() to return absl::Span PiperOrigin-RevId: 368765721 -- 6aaab9536d6957303c7aba100c3afaa6fb0ea2c8 by Martijn Vels <mvels@google.com>: Remove locking from parent stack. This change removes the need to lock all access to `parent_stack' by making the 'copy constructor' logic specify the 'copied from' CordzInfo (where available) to the TrackCord function, after which parent_stack is immutable. PiperOrigin-RevId: 368760630 -- b19e2059cada35a8ede994833018edac94de6ddc by Martijn Vels <mvels@google.com>: Add cordz instrumentation to Cord PiperOrigin-RevId: 368746225 -- 67b8bbf980f0f4e1db79aa32968e9a715a09b51a by Martijn Vels <mvels@google.com>: Create ABSL_INTERNAL_CORDZ_ENABLED define controlling when Cordz code is enabled There are specific builds and condtions under which we don't support cordz sampling, which is per this change represented by ABSL_INTERNAL_CORDZ_ENABLED being defined. PiperOrigin-RevId: 368731603 -- 8cbfe0e3169637a620f4b66ad2bc2ce340879cb0 by Martijn Vels <mvels@google.com>: Add a `rep` property to CordzInfo to be managed by Cord logic. This change adds a `rep` property to CordzInfo, which is intended to be used by collection logic. Mini design: Cord invokes TrackCord() providing the active 'root' cordrep of the newly sampled Cord, returning a CordzInfo with a weak (uncounted) reference to this root. Cord invokes `SetCordRep()` each time the root cordrep of the sampled Cord is updated while holding `mutex()`. Cord must also obtain `mutex()` _before_ removing a reference on the old root. i.e.: Cord must guarantee that the (weak) reference held in CordzInfo is at all times valid. CordzInfo collection code can then safely obtain a (reference counted) rep pointer by adding a reference to `rep_` while holding `mutex()`. This requires only a very brief critical section inside CordzInfo logic, minimizing contention with concurrent Cord updates. Cord code should typically obtain and hold `mutex()` for the entirety of each mutating Cord operation on a sampled cord. As Cord is thread compatible, it never competes on the lock with any other thread. The only possible concurrent access is from Cordz collection code, which should be a relatively rare event. PiperOrigin-RevId: 368673758 -- 1255120dce2bdd6b4205a34a0e555e0b74b6152f by Martijn Vels <mvels@google.com>: Remove 'depth' from active recorded metrics. Going forward we do not 'live' record depth (and size), but will observe these at collection time only. PiperOrigin-RevId: 368636572 -- 83e5146e35f221736b49e9f0a8805f8c159a51db by Martijn Vels <mvels@google.com>: Make cordz targets visible in OSS PiperOrigin-RevId: 368615010 -- dcb16a4f1239151f0a8c70a8cfeb29dabbd113b8 by Martijn Vels <mvels@google.com>: Internal cleanup PiperOrigin-RevId: 368514666 GitOrigin-RevId: 341670bce317dd6af8d3c066970230591a47e80c Change-Id: I94cecfbbd441eb386f99fc5186c468a7a5538862
4 years ago
void Cord::InlineRep::StartProfiling() {
CordRep* tree = as_tree();
auto* cordz_info = absl::cord_internal::CordzInfo::TrackCord(tree);
set_cordz_info(cordz_info);
cordz_info->RecordMetrics(size());
}
void Cord::InlineRep::StartProfiling(const Cord::InlineRep& src) {
CordRep* tree = as_tree();
absl::cord_internal::CordzInfo* parent =
src.is_profiled() ? src.cordz_info() : nullptr;
auto* cordz_info = absl::cord_internal::CordzInfo::TrackCord(tree, parent);
set_cordz_info(cordz_info);
cordz_info->RecordMetrics(size());
}
void Cord::InlineRep::UpdateCordzStatisticsSlow() {
CordRep* tree = as_tree();
data_.cordz_info()->RecordMetrics(tree->length);
}
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
void Cord::InlineRep::AssignSlow(const Cord::InlineRep& src) {
Export of internal Abseil changes -- 6cfac39ea1266f01f195de5eb4c9a6fc9ea9b20a by Matt Kulukundis <kfm@google.com>: Fix a typo PiperOrigin-RevId: 366174890 -- 6ee8c58647aef171d394e59fa06f9bf8cd0306ec by Derek Mauro <dmauro@google.com>: Adds `ABSL_ATTRIBUTE_LIFETIME_BOUND` and applies it to the `const std::string&` constructor of `absl::string_view`. Compilers that support this attribute will emit a warning if the parameter does not have sufficient lifetime. PiperOrigin-RevId: 366027738 -- b944427d96e4b436b8fa0fe396c2a1118dbbbd13 by Abseil Team <absl-team@google.com>: Calls to `ResetToEmpty()` from `ClearSlow` use ~3% of the time in assignment. However, `ClearSlow()` is only used in contexts where `data_` is immediately reassigned. Rename `ClearSlow()` into `UnrefTree()` and remove `data_` resetting. PiperOrigin-RevId: 365977213 -- 7428b3147a5672c8bb55649efa3a1cfe19b52a8b by Abseil Team <absl-team@google.com>: Fix CordRepRing diabolical growth The 'Mutable' function in CordRepRing was over-eager in doubling capacity, which lead to 'ludicrous' growth in the diabolical test case as added to cord_test. This CL fixes the doubling for growing shared reps, and tempers CordRepRing growth for non shared capacity to 1.5 instead of 2, which is more inline with a conservative growth we also have in tree cord. After this change, CordRepRing no longer swamps the heap into the shadow realm, and is in effect reducing the memory used compared to the tree implementation. With a diabolical 5000 bytes growth pattern: Tree cord: 1523520 bytes Ring cord: 274232 bytes PiperOrigin-RevId: 365915857 -- f24d4aee48b03c5a7980664df903f947cbb198e8 by Andy Getzendanner <durandal@google.com>: Import of CCTZ from GitHub. PiperOrigin-RevId: 365873932 GitOrigin-RevId: 6cfac39ea1266f01f195de5eb4c9a6fc9ea9b20a Change-Id: I9737aa215ac732c9785a1d0032c77aba62330f12
4 years ago
UnrefTree();
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
data_ = src.data_;
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
if (is_tree()) {
Export of internal Abseil changes -- 6e9f93888bbe6718997ed90bbd049f1f3572b066 by Abseil Team <absl-team@google.com>: Fix status to be safe for self move assignment. PiperOrigin-RevId: 346815392 -- 35cae74a977f258e81dfbe925fb5a34cb6632036 by Gennadiy Rozental <rogeeff@google.com>: Eliminate unnecessary access to the global vars. PiperOrigin-RevId: 346777168 -- e7e786c243069060f5d6c1c945decb4b0b83f95b by Andy Getzendanner <durandal@google.com>: Internal change. PiperOrigin-RevId: 346685656 -- 4ccd41c48f1a83cfa20b3ea534f743dd7d788376 by Abseil Team <absl-team@google.com>: Move CordRep Ref() and Unref() logic into cord_internal.h This change moves Ref() and Unref() logic out of cord.cc into cord_internal. The main purpose is to make upcoming ring buffer changes easier to isolate from existing cord.cc code. Notice that this removes the nullptr check from Unref() and now requires it to be non null (which held true most times). We may need to rethink if the 'unref unlikely one' is the common case: are cordreps most likely shared? This may be something between 'root' and non root nodes, i.e., is it more likely for leaf / flat nodes in large cords to be shared than top level cordreps being shared? Vice versa? Benchmarks say that we mostly shouldn't care, the caveat being that atomic ops seem more expensive on upcoming archs (arcadia) so we should error on the side of an extra IsOne() branch saving us single owned atomic ops. PiperOrigin-RevId: 346676121 -- f0babab103b9e60d61ba09482d468985e43eceb3 by Samuel Benzaquen <sbenza@google.com>: Fix iterator based constructor and `.insert` members to only require EmplaceConstructible as the standard specifies. PiperOrigin-RevId: 346616707 -- 8f48eedda02277f9c96a88ed7726e34b557cce20 by Evan Brown <ezb@google.com>: Fix a bug in binary_search_impl when there's a transparent, three-way comparator that has different equivalence classes for different lookup types. Add a new can_have_multiple_equivalent_keys method to share the common logic for these cases. PiperOrigin-RevId: 346605948 -- 649183cb3cc9383431de9c81fb1c0f885d4001ae by Abseil Team <absl-team@google.com>: Add benchmark for accessing a Duration flag. PiperOrigin-RevId: 346594843 -- fefdb046520871af63ce2229e2f7cccfc0483dea by Abseil Team <absl-team@google.com>: Restructure CordReader for upcoming ring buffer changes. PiperOrigin-RevId: 346410642 -- 8b2f50e7da0ebab06ead5f94e366e984ca23cb6a by Abseil Team <absl-team@google.com>: Wire in an internal-only flag to toggle upcoming ring buffer changes on/off for experimentation. PiperOrigin-RevId: 346199111 GitOrigin-RevId: 6e9f93888bbe6718997ed90bbd049f1f3572b066 Change-Id: I8f34866b25a79209cb5448bbb28dd3044111d2e9
4 years ago
CordRep::Ref(tree());
Export of internal Abseil changes -- 4ff721439234e91caf6f7b772e5f554e7dd423c8 by Benjamin Barenblat <bbaren@google.com>: Remove endian-sensitivity from hash slow path Prior to this commit, the Abseil hash fast path was endian-agnostic, but the slow path assumed a little-endian platform. Change the slow path to be endian-correct, ensuring that values produced by the fast and slow paths are equal even on big-endian systems. PiperOrigin-RevId: 355424258 -- 7f4fe1aa4de46ad0a2ef19fa9c061fc12a7391ed by Abseil Team <absl-team@google.com>: Directly store CordzInfo in the InlineData data contents of InlineRep This greatly reduces the cost of coping and moving cords. Especially the move constructor and move assignment are now back to lean loads and stores without needing any CordzInfo lookups for tracked cords. PiperOrigin-RevId: 355409161 -- 3ca4ca84ed6d98f1e383ffd8d12c28876e905bb3 by Abseil Team <absl-team@google.com>: Add #include <unordered_map> PiperOrigin-RevId: 355386114 -- 30b0ffad0621971b3135148fcc9e183b0dd2a6bb by Abseil Team <absl-team@google.com>: Optimize Cord copy constructor This change avoids double stores of the Cord copy constructor from the zero init of the InlineData / InlineRep contents followed by the assignment and inlines the copy constructor. PiperOrigin-RevId: 355287939 -- 0c043fa7b6e41ca7cefc5edc1e17ad46223e4e77 by CJ Johnson <johnsoncj@google.com>: Now that the absl::Cleanup example returns absl::Status, since we decided on absl::FailedPreconditionError, the precondition should be a positive statement and then the check should be failure to adhere to that positive statement PiperOrigin-RevId: 355216923 -- 9ed922ca5d28fe8790ec6bc0837cf39fbcc92896 by Gennadiy Rozental <rogeeff@google.com>: Do not set mvsc linker flags for clang-cl (fixes #874) Import of https://github.com/abseil/abseil-cpp/pull/891 PiperOrigin-RevId: 355199380 GitOrigin-RevId: 4ff721439234e91caf6f7b772e5f554e7dd423c8 Change-Id: I3d9d2383549720d7a91f9108dfcd979ad6632fce
4 years ago
clear_cordz_info();
Export of internal Abseil changes -- 341670bce317dd6af8d3c066970230591a47e80c by Martijn Vels <mvels@google.com>: Change GetStack() and GetParentStack() to return absl::Span PiperOrigin-RevId: 368765721 -- 6aaab9536d6957303c7aba100c3afaa6fb0ea2c8 by Martijn Vels <mvels@google.com>: Remove locking from parent stack. This change removes the need to lock all access to `parent_stack' by making the 'copy constructor' logic specify the 'copied from' CordzInfo (where available) to the TrackCord function, after which parent_stack is immutable. PiperOrigin-RevId: 368760630 -- b19e2059cada35a8ede994833018edac94de6ddc by Martijn Vels <mvels@google.com>: Add cordz instrumentation to Cord PiperOrigin-RevId: 368746225 -- 67b8bbf980f0f4e1db79aa32968e9a715a09b51a by Martijn Vels <mvels@google.com>: Create ABSL_INTERNAL_CORDZ_ENABLED define controlling when Cordz code is enabled There are specific builds and condtions under which we don't support cordz sampling, which is per this change represented by ABSL_INTERNAL_CORDZ_ENABLED being defined. PiperOrigin-RevId: 368731603 -- 8cbfe0e3169637a620f4b66ad2bc2ce340879cb0 by Martijn Vels <mvels@google.com>: Add a `rep` property to CordzInfo to be managed by Cord logic. This change adds a `rep` property to CordzInfo, which is intended to be used by collection logic. Mini design: Cord invokes TrackCord() providing the active 'root' cordrep of the newly sampled Cord, returning a CordzInfo with a weak (uncounted) reference to this root. Cord invokes `SetCordRep()` each time the root cordrep of the sampled Cord is updated while holding `mutex()`. Cord must also obtain `mutex()` _before_ removing a reference on the old root. i.e.: Cord must guarantee that the (weak) reference held in CordzInfo is at all times valid. CordzInfo collection code can then safely obtain a (reference counted) rep pointer by adding a reference to `rep_` while holding `mutex()`. This requires only a very brief critical section inside CordzInfo logic, minimizing contention with concurrent Cord updates. Cord code should typically obtain and hold `mutex()` for the entirety of each mutating Cord operation on a sampled cord. As Cord is thread compatible, it never competes on the lock with any other thread. The only possible concurrent access is from Cordz collection code, which should be a relatively rare event. PiperOrigin-RevId: 368673758 -- 1255120dce2bdd6b4205a34a0e555e0b74b6152f by Martijn Vels <mvels@google.com>: Remove 'depth' from active recorded metrics. Going forward we do not 'live' record depth (and size), but will observe these at collection time only. PiperOrigin-RevId: 368636572 -- 83e5146e35f221736b49e9f0a8805f8c159a51db by Martijn Vels <mvels@google.com>: Make cordz targets visible in OSS PiperOrigin-RevId: 368615010 -- dcb16a4f1239151f0a8c70a8cfeb29dabbd113b8 by Martijn Vels <mvels@google.com>: Internal cleanup PiperOrigin-RevId: 368514666 GitOrigin-RevId: 341670bce317dd6af8d3c066970230591a47e80c Change-Id: I94cecfbbd441eb386f99fc5186c468a7a5538862
4 years ago
if (ABSL_PREDICT_FALSE(should_profile())) {
StartProfiling(src);
}
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
}
}
Export of internal Abseil changes -- 6cfac39ea1266f01f195de5eb4c9a6fc9ea9b20a by Matt Kulukundis <kfm@google.com>: Fix a typo PiperOrigin-RevId: 366174890 -- 6ee8c58647aef171d394e59fa06f9bf8cd0306ec by Derek Mauro <dmauro@google.com>: Adds `ABSL_ATTRIBUTE_LIFETIME_BOUND` and applies it to the `const std::string&` constructor of `absl::string_view`. Compilers that support this attribute will emit a warning if the parameter does not have sufficient lifetime. PiperOrigin-RevId: 366027738 -- b944427d96e4b436b8fa0fe396c2a1118dbbbd13 by Abseil Team <absl-team@google.com>: Calls to `ResetToEmpty()` from `ClearSlow` use ~3% of the time in assignment. However, `ClearSlow()` is only used in contexts where `data_` is immediately reassigned. Rename `ClearSlow()` into `UnrefTree()` and remove `data_` resetting. PiperOrigin-RevId: 365977213 -- 7428b3147a5672c8bb55649efa3a1cfe19b52a8b by Abseil Team <absl-team@google.com>: Fix CordRepRing diabolical growth The 'Mutable' function in CordRepRing was over-eager in doubling capacity, which lead to 'ludicrous' growth in the diabolical test case as added to cord_test. This CL fixes the doubling for growing shared reps, and tempers CordRepRing growth for non shared capacity to 1.5 instead of 2, which is more inline with a conservative growth we also have in tree cord. After this change, CordRepRing no longer swamps the heap into the shadow realm, and is in effect reducing the memory used compared to the tree implementation. With a diabolical 5000 bytes growth pattern: Tree cord: 1523520 bytes Ring cord: 274232 bytes PiperOrigin-RevId: 365915857 -- f24d4aee48b03c5a7980664df903f947cbb198e8 by Andy Getzendanner <durandal@google.com>: Import of CCTZ from GitHub. PiperOrigin-RevId: 365873932 GitOrigin-RevId: 6cfac39ea1266f01f195de5eb4c9a6fc9ea9b20a Change-Id: I9737aa215ac732c9785a1d0032c77aba62330f12
4 years ago
void Cord::InlineRep::UnrefTree() {
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
if (is_tree()) {
Export of internal Abseil changes -- 341670bce317dd6af8d3c066970230591a47e80c by Martijn Vels <mvels@google.com>: Change GetStack() and GetParentStack() to return absl::Span PiperOrigin-RevId: 368765721 -- 6aaab9536d6957303c7aba100c3afaa6fb0ea2c8 by Martijn Vels <mvels@google.com>: Remove locking from parent stack. This change removes the need to lock all access to `parent_stack' by making the 'copy constructor' logic specify the 'copied from' CordzInfo (where available) to the TrackCord function, after which parent_stack is immutable. PiperOrigin-RevId: 368760630 -- b19e2059cada35a8ede994833018edac94de6ddc by Martijn Vels <mvels@google.com>: Add cordz instrumentation to Cord PiperOrigin-RevId: 368746225 -- 67b8bbf980f0f4e1db79aa32968e9a715a09b51a by Martijn Vels <mvels@google.com>: Create ABSL_INTERNAL_CORDZ_ENABLED define controlling when Cordz code is enabled There are specific builds and condtions under which we don't support cordz sampling, which is per this change represented by ABSL_INTERNAL_CORDZ_ENABLED being defined. PiperOrigin-RevId: 368731603 -- 8cbfe0e3169637a620f4b66ad2bc2ce340879cb0 by Martijn Vels <mvels@google.com>: Add a `rep` property to CordzInfo to be managed by Cord logic. This change adds a `rep` property to CordzInfo, which is intended to be used by collection logic. Mini design: Cord invokes TrackCord() providing the active 'root' cordrep of the newly sampled Cord, returning a CordzInfo with a weak (uncounted) reference to this root. Cord invokes `SetCordRep()` each time the root cordrep of the sampled Cord is updated while holding `mutex()`. Cord must also obtain `mutex()` _before_ removing a reference on the old root. i.e.: Cord must guarantee that the (weak) reference held in CordzInfo is at all times valid. CordzInfo collection code can then safely obtain a (reference counted) rep pointer by adding a reference to `rep_` while holding `mutex()`. This requires only a very brief critical section inside CordzInfo logic, minimizing contention with concurrent Cord updates. Cord code should typically obtain and hold `mutex()` for the entirety of each mutating Cord operation on a sampled cord. As Cord is thread compatible, it never competes on the lock with any other thread. The only possible concurrent access is from Cordz collection code, which should be a relatively rare event. PiperOrigin-RevId: 368673758 -- 1255120dce2bdd6b4205a34a0e555e0b74b6152f by Martijn Vels <mvels@google.com>: Remove 'depth' from active recorded metrics. Going forward we do not 'live' record depth (and size), but will observe these at collection time only. PiperOrigin-RevId: 368636572 -- 83e5146e35f221736b49e9f0a8805f8c159a51db by Martijn Vels <mvels@google.com>: Make cordz targets visible in OSS PiperOrigin-RevId: 368615010 -- dcb16a4f1239151f0a8c70a8cfeb29dabbd113b8 by Martijn Vels <mvels@google.com>: Internal cleanup PiperOrigin-RevId: 368514666 GitOrigin-RevId: 341670bce317dd6af8d3c066970230591a47e80c Change-Id: I94cecfbbd441eb386f99fc5186c468a7a5538862
4 years ago
if (ABSL_PREDICT_FALSE(is_profiled())) {
absl::cord_internal::CordzInfo::UntrackCord(cordz_info());
}
Export of internal Abseil changes -- 6e9f93888bbe6718997ed90bbd049f1f3572b066 by Abseil Team <absl-team@google.com>: Fix status to be safe for self move assignment. PiperOrigin-RevId: 346815392 -- 35cae74a977f258e81dfbe925fb5a34cb6632036 by Gennadiy Rozental <rogeeff@google.com>: Eliminate unnecessary access to the global vars. PiperOrigin-RevId: 346777168 -- e7e786c243069060f5d6c1c945decb4b0b83f95b by Andy Getzendanner <durandal@google.com>: Internal change. PiperOrigin-RevId: 346685656 -- 4ccd41c48f1a83cfa20b3ea534f743dd7d788376 by Abseil Team <absl-team@google.com>: Move CordRep Ref() and Unref() logic into cord_internal.h This change moves Ref() and Unref() logic out of cord.cc into cord_internal. The main purpose is to make upcoming ring buffer changes easier to isolate from existing cord.cc code. Notice that this removes the nullptr check from Unref() and now requires it to be non null (which held true most times). We may need to rethink if the 'unref unlikely one' is the common case: are cordreps most likely shared? This may be something between 'root' and non root nodes, i.e., is it more likely for leaf / flat nodes in large cords to be shared than top level cordreps being shared? Vice versa? Benchmarks say that we mostly shouldn't care, the caveat being that atomic ops seem more expensive on upcoming archs (arcadia) so we should error on the side of an extra IsOne() branch saving us single owned atomic ops. PiperOrigin-RevId: 346676121 -- f0babab103b9e60d61ba09482d468985e43eceb3 by Samuel Benzaquen <sbenza@google.com>: Fix iterator based constructor and `.insert` members to only require EmplaceConstructible as the standard specifies. PiperOrigin-RevId: 346616707 -- 8f48eedda02277f9c96a88ed7726e34b557cce20 by Evan Brown <ezb@google.com>: Fix a bug in binary_search_impl when there's a transparent, three-way comparator that has different equivalence classes for different lookup types. Add a new can_have_multiple_equivalent_keys method to share the common logic for these cases. PiperOrigin-RevId: 346605948 -- 649183cb3cc9383431de9c81fb1c0f885d4001ae by Abseil Team <absl-team@google.com>: Add benchmark for accessing a Duration flag. PiperOrigin-RevId: 346594843 -- fefdb046520871af63ce2229e2f7cccfc0483dea by Abseil Team <absl-team@google.com>: Restructure CordReader for upcoming ring buffer changes. PiperOrigin-RevId: 346410642 -- 8b2f50e7da0ebab06ead5f94e366e984ca23cb6a by Abseil Team <absl-team@google.com>: Wire in an internal-only flag to toggle upcoming ring buffer changes on/off for experimentation. PiperOrigin-RevId: 346199111 GitOrigin-RevId: 6e9f93888bbe6718997ed90bbd049f1f3572b066 Change-Id: I8f34866b25a79209cb5448bbb28dd3044111d2e9
4 years ago
CordRep::Unref(tree());
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
}
}
// --------------------------------------------------------------------
// Constructors and destructors
Cord::Cord(absl::string_view src) {
const size_t n = src.size();
if (n <= InlineRep::kMaxInline) {
contents_.set_data(src.data(), n, false);
} else {
Export of internal Abseil changes -- ac1df60490c9583e475e22de7adfc40023196fbf by Martijn Vels <mvels@google.com>: Change Cord constructor(string_view) to explicit make_tree and Cordz tracking This CL changes the ctor to use an easier to maintain model where Cord code explicitly invokes Cordz update or new / tree logic, which avoids the ambiguity of the 'branched' InlineRep::set_tree code. This removes the need to equip InlineRep with 'MethodIdentifier' or other necessary call info, and also is a cleaner model: InlineRep is carrying too much code now that should plainly sit in Cord, especially with all internal abstractions having moved to InlineData. See child CL(s) for desired state PiperOrigin-RevId: 369433619 -- b665af7f586e6c679a8b27d4f78d5a1d2b596058 by Abseil Team <absl-team@google.com>: Rename the 'Compare' template type to 'LessThan', as the passed-in function is expected to act like operator<. It is worth avoiding confusion with std::compare, which returns an int (-1/0/1), as due to implicit casting this can lead to hard-to-spot bugs. PiperOrigin-RevId: 369391118 -- c3c775269cad0f4982ec63f3616dd78bb9e52dca by Martijn Vels <mvels@google.com>: Integrate CordzUpdateTracker into CordzInfo PiperOrigin-RevId: 369348824 -- 771d81ed357496c117179e1daec76eba5155932d by Martijn Vels <mvels@google.com>: Replace mutex() with Lock() / Unlock() function Mini design future tracking of CordzInfo sampled cords: CordzInfo holds a CordRep* reference without a reference count. Cord is responsible for synchronizing updates for sampled cords such that the CordRep* contained in CordzInfo is at all times valid. This is done by scoping Lock() and Unlock() calls around the code modifying the code of a sampled cord. For example (using the future CL CordzUpdateScope()): CordzInfo* cordz_info = get_cordz_info(); CordzUpdateScope scope(cordz_info, CordzUpdateTracker::kRemovePrefix); CordRep* rep = RemovePrefixImpl(root); set_tree(rep); if (cordz_info) { cordz_info->SetCordRep(rep); } On CordzInfo::Unlock(), if the internal rep is null, the cord is no longer sampled, and CordzInfo will be deleted. Thus any update resulting in the Cord being inlined will automatically no longer be sampled. PiperOrigin-RevId: 369338802 -- 5563c12df04a1e965a03b50bdd032739c55c0706 by Martijn Vels <mvels@google.com>: Add UpdateTracker to CordzStatistics PiperOrigin-RevId: 369318178 -- 6b4d8463722a3e55a3e8f6cb3741a41055e7f83e by Martijn Vels <mvels@google.com>: Add kClear, kConstructor* and kUnknown values and fix typo PiperOrigin-RevId: 369297163 -- 041adcbc929789d6d53371a8236840fc350e1eeb by Derek Mauro <dmauro@google.com>: Switch from malloc to operator new in pool_urbg.cc so it can only fail by throwing/aborting PiperOrigin-RevId: 369274087 -- 5d97a5f43e3f2d02d0a5bbe586d93b5751812981 by Benjamin Barenblat <bbaren@google.com>: Correct Thumb function bound computation in the symbolizer On 32-bit ARM, all functions are aligned to multiples of two bytes, and the lowest-order bit in a function’s address is ignored by the CPU when computing branch targets. That bit is still present in instructions and ELF symbol tables, though; it’s repurposed to indicate whether the function contains ARM or Thumb code. If the symbolizer doesn’t ignore that bit, it will believe Thumb functions have boundaries that are off by one byte, so instruct the symbolizer to null out the lowest-order bit after retrieving it from the symbol table. PiperOrigin-RevId: 369254082 -- 462bb307c6cc332c1e2c3adb5f0cad51804bf937 by Derek Mauro <dmauro@google.com>: Add a check for malloc failure in pool_urbg.cc GitHub #940 PiperOrigin-RevId: 369238100 GitOrigin-RevId: ac1df60490c9583e475e22de7adfc40023196fbf Change-Id: Ic6ec91c62cd3a0031f6a75a43a83da959ece2d25
4 years ago
contents_.data_.make_tree(NewTree(src.data(), n, 0));
if (ABSL_PREDICT_FALSE(absl::cord_internal::cordz_should_profile())) {
contents_.StartProfiling();
}
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
}
}
Export of internal Abseil changes -- 7d0468a6610ed85586d5c87fd65de8dac5118923 by Derek Mauro <dmauro@google.com>: Import of CCTZ from GitHub. PiperOrigin-RevId: 313226473 -- 1131ef6d116f5ce7d46537a82f300ea06dcaaa53 by Gennadiy Rozental <rogeeff@google.com>: Migrate internal interface to use mutable references. PiperOrigin-RevId: 312931131 -- 96225212a9f5fbd0b38c71fe65539164992c7c3b by Laramie Leavitt <lar@google.com>: Remove random/internal/distributions.h This file was something of an historical artifact. All of the related code has either been removed or migraged, and so the only remaining type belongs with uniform_helper.h, as it is used to infer the return type of the absl::Uniform method in a few cases. PiperOrigin-RevId: 312878173 -- 6dcbd5be58ad425e08740ff64088373ee7fe4a72 by Mark Barolak <mbar@google.com>: Release the StrFormat test case for Cords to open source. PiperOrigin-RevId: 312707974 -- 34484d18dfb63a0a7ad6e2aaeb570e33592968be by Abseil Team <absl-team@google.com>: Let Cord::Cord(string&&), Cord::operator=(string&&), Cord::Append(string&&), and Cord::Prepend(string&&) steal string data and embed it into the Cord as a single external chunk, instead of copying it into flat chunks (at most 4083-byte each). Stealing string data is faster, but it creates a long chunk, which leads to a higher more memory usage if its subcords are created and outlive the whole Cord. These functions revert to copying the data if any of the following conditions holds: - string size is at most kMaxBytesToCopy (511), to avoid the overhead of an external chunk for short strings; - less than half of string capacity is used, to avoid pinning to much unused memory. PiperOrigin-RevId: 312683785 GitOrigin-RevId: 7d0468a6610ed85586d5c87fd65de8dac5118923 Change-Id: If79b5a1dfe6d53a8ddddbc7da84338f11fc4cfa3
5 years ago
template <typename T, Cord::EnableIfString<T>>
Cord::Cord(T&& src) {
if (
// String is short: copy data to avoid external block overhead.
src.size() <= kMaxBytesToCopy ||
// String is wasteful: copy data to avoid pinning too much unused memory.
src.size() < src.capacity() / 2
) {
if (src.size() <= InlineRep::kMaxInline) {
contents_.set_data(src.data(), src.size(), false);
} else {
contents_.set_tree(NewTree(src.data(), src.size(), 0));
}
} else {
struct StringReleaser {
void operator()(absl::string_view /* data */) {}
std::string data;
};
const absl::string_view original_data = src;
auto* rep = static_cast<
::absl::cord_internal::CordRepExternalImpl<StringReleaser>*>(
absl::cord_internal::NewExternalRep(
original_data, StringReleaser{std::forward<T>(src)}));
Export of internal Abseil changes -- 7d0468a6610ed85586d5c87fd65de8dac5118923 by Derek Mauro <dmauro@google.com>: Import of CCTZ from GitHub. PiperOrigin-RevId: 313226473 -- 1131ef6d116f5ce7d46537a82f300ea06dcaaa53 by Gennadiy Rozental <rogeeff@google.com>: Migrate internal interface to use mutable references. PiperOrigin-RevId: 312931131 -- 96225212a9f5fbd0b38c71fe65539164992c7c3b by Laramie Leavitt <lar@google.com>: Remove random/internal/distributions.h This file was something of an historical artifact. All of the related code has either been removed or migraged, and so the only remaining type belongs with uniform_helper.h, as it is used to infer the return type of the absl::Uniform method in a few cases. PiperOrigin-RevId: 312878173 -- 6dcbd5be58ad425e08740ff64088373ee7fe4a72 by Mark Barolak <mbar@google.com>: Release the StrFormat test case for Cords to open source. PiperOrigin-RevId: 312707974 -- 34484d18dfb63a0a7ad6e2aaeb570e33592968be by Abseil Team <absl-team@google.com>: Let Cord::Cord(string&&), Cord::operator=(string&&), Cord::Append(string&&), and Cord::Prepend(string&&) steal string data and embed it into the Cord as a single external chunk, instead of copying it into flat chunks (at most 4083-byte each). Stealing string data is faster, but it creates a long chunk, which leads to a higher more memory usage if its subcords are created and outlive the whole Cord. These functions revert to copying the data if any of the following conditions holds: - string size is at most kMaxBytesToCopy (511), to avoid the overhead of an external chunk for short strings; - less than half of string capacity is used, to avoid pinning to much unused memory. PiperOrigin-RevId: 312683785 GitOrigin-RevId: 7d0468a6610ed85586d5c87fd65de8dac5118923 Change-Id: If79b5a1dfe6d53a8ddddbc7da84338f11fc4cfa3
5 years ago
// Moving src may have invalidated its data pointer, so adjust it.
rep->base = rep->template get<0>().data.data();
Export of internal Abseil changes -- 7d0468a6610ed85586d5c87fd65de8dac5118923 by Derek Mauro <dmauro@google.com>: Import of CCTZ from GitHub. PiperOrigin-RevId: 313226473 -- 1131ef6d116f5ce7d46537a82f300ea06dcaaa53 by Gennadiy Rozental <rogeeff@google.com>: Migrate internal interface to use mutable references. PiperOrigin-RevId: 312931131 -- 96225212a9f5fbd0b38c71fe65539164992c7c3b by Laramie Leavitt <lar@google.com>: Remove random/internal/distributions.h This file was something of an historical artifact. All of the related code has either been removed or migraged, and so the only remaining type belongs with uniform_helper.h, as it is used to infer the return type of the absl::Uniform method in a few cases. PiperOrigin-RevId: 312878173 -- 6dcbd5be58ad425e08740ff64088373ee7fe4a72 by Mark Barolak <mbar@google.com>: Release the StrFormat test case for Cords to open source. PiperOrigin-RevId: 312707974 -- 34484d18dfb63a0a7ad6e2aaeb570e33592968be by Abseil Team <absl-team@google.com>: Let Cord::Cord(string&&), Cord::operator=(string&&), Cord::Append(string&&), and Cord::Prepend(string&&) steal string data and embed it into the Cord as a single external chunk, instead of copying it into flat chunks (at most 4083-byte each). Stealing string data is faster, but it creates a long chunk, which leads to a higher more memory usage if its subcords are created and outlive the whole Cord. These functions revert to copying the data if any of the following conditions holds: - string size is at most kMaxBytesToCopy (511), to avoid the overhead of an external chunk for short strings; - less than half of string capacity is used, to avoid pinning to much unused memory. PiperOrigin-RevId: 312683785 GitOrigin-RevId: 7d0468a6610ed85586d5c87fd65de8dac5118923 Change-Id: If79b5a1dfe6d53a8ddddbc7da84338f11fc4cfa3
5 years ago
contents_.set_tree(rep);
}
}
template Cord::Cord(std::string&& src);
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
// The destruction code is separate so that the compiler can determine
// that it does not need to call the destructor on a moved-from Cord.
void Cord::DestroyCordSlow() {
Export of internal Abseil changes -- 341670bce317dd6af8d3c066970230591a47e80c by Martijn Vels <mvels@google.com>: Change GetStack() and GetParentStack() to return absl::Span PiperOrigin-RevId: 368765721 -- 6aaab9536d6957303c7aba100c3afaa6fb0ea2c8 by Martijn Vels <mvels@google.com>: Remove locking from parent stack. This change removes the need to lock all access to `parent_stack' by making the 'copy constructor' logic specify the 'copied from' CordzInfo (where available) to the TrackCord function, after which parent_stack is immutable. PiperOrigin-RevId: 368760630 -- b19e2059cada35a8ede994833018edac94de6ddc by Martijn Vels <mvels@google.com>: Add cordz instrumentation to Cord PiperOrigin-RevId: 368746225 -- 67b8bbf980f0f4e1db79aa32968e9a715a09b51a by Martijn Vels <mvels@google.com>: Create ABSL_INTERNAL_CORDZ_ENABLED define controlling when Cordz code is enabled There are specific builds and condtions under which we don't support cordz sampling, which is per this change represented by ABSL_INTERNAL_CORDZ_ENABLED being defined. PiperOrigin-RevId: 368731603 -- 8cbfe0e3169637a620f4b66ad2bc2ce340879cb0 by Martijn Vels <mvels@google.com>: Add a `rep` property to CordzInfo to be managed by Cord logic. This change adds a `rep` property to CordzInfo, which is intended to be used by collection logic. Mini design: Cord invokes TrackCord() providing the active 'root' cordrep of the newly sampled Cord, returning a CordzInfo with a weak (uncounted) reference to this root. Cord invokes `SetCordRep()` each time the root cordrep of the sampled Cord is updated while holding `mutex()`. Cord must also obtain `mutex()` _before_ removing a reference on the old root. i.e.: Cord must guarantee that the (weak) reference held in CordzInfo is at all times valid. CordzInfo collection code can then safely obtain a (reference counted) rep pointer by adding a reference to `rep_` while holding `mutex()`. This requires only a very brief critical section inside CordzInfo logic, minimizing contention with concurrent Cord updates. Cord code should typically obtain and hold `mutex()` for the entirety of each mutating Cord operation on a sampled cord. As Cord is thread compatible, it never competes on the lock with any other thread. The only possible concurrent access is from Cordz collection code, which should be a relatively rare event. PiperOrigin-RevId: 368673758 -- 1255120dce2bdd6b4205a34a0e555e0b74b6152f by Martijn Vels <mvels@google.com>: Remove 'depth' from active recorded metrics. Going forward we do not 'live' record depth (and size), but will observe these at collection time only. PiperOrigin-RevId: 368636572 -- 83e5146e35f221736b49e9f0a8805f8c159a51db by Martijn Vels <mvels@google.com>: Make cordz targets visible in OSS PiperOrigin-RevId: 368615010 -- dcb16a4f1239151f0a8c70a8cfeb29dabbd113b8 by Martijn Vels <mvels@google.com>: Internal cleanup PiperOrigin-RevId: 368514666 GitOrigin-RevId: 341670bce317dd6af8d3c066970230591a47e80c Change-Id: I94cecfbbd441eb386f99fc5186c468a7a5538862
4 years ago
if (ABSL_PREDICT_FALSE(contents_.is_profiled())) {
absl::cord_internal::CordzInfo::UntrackCord(contents_.cordz_info());
}
Export of internal Abseil changes -- 6e9f93888bbe6718997ed90bbd049f1f3572b066 by Abseil Team <absl-team@google.com>: Fix status to be safe for self move assignment. PiperOrigin-RevId: 346815392 -- 35cae74a977f258e81dfbe925fb5a34cb6632036 by Gennadiy Rozental <rogeeff@google.com>: Eliminate unnecessary access to the global vars. PiperOrigin-RevId: 346777168 -- e7e786c243069060f5d6c1c945decb4b0b83f95b by Andy Getzendanner <durandal@google.com>: Internal change. PiperOrigin-RevId: 346685656 -- 4ccd41c48f1a83cfa20b3ea534f743dd7d788376 by Abseil Team <absl-team@google.com>: Move CordRep Ref() and Unref() logic into cord_internal.h This change moves Ref() and Unref() logic out of cord.cc into cord_internal. The main purpose is to make upcoming ring buffer changes easier to isolate from existing cord.cc code. Notice that this removes the nullptr check from Unref() and now requires it to be non null (which held true most times). We may need to rethink if the 'unref unlikely one' is the common case: are cordreps most likely shared? This may be something between 'root' and non root nodes, i.e., is it more likely for leaf / flat nodes in large cords to be shared than top level cordreps being shared? Vice versa? Benchmarks say that we mostly shouldn't care, the caveat being that atomic ops seem more expensive on upcoming archs (arcadia) so we should error on the side of an extra IsOne() branch saving us single owned atomic ops. PiperOrigin-RevId: 346676121 -- f0babab103b9e60d61ba09482d468985e43eceb3 by Samuel Benzaquen <sbenza@google.com>: Fix iterator based constructor and `.insert` members to only require EmplaceConstructible as the standard specifies. PiperOrigin-RevId: 346616707 -- 8f48eedda02277f9c96a88ed7726e34b557cce20 by Evan Brown <ezb@google.com>: Fix a bug in binary_search_impl when there's a transparent, three-way comparator that has different equivalence classes for different lookup types. Add a new can_have_multiple_equivalent_keys method to share the common logic for these cases. PiperOrigin-RevId: 346605948 -- 649183cb3cc9383431de9c81fb1c0f885d4001ae by Abseil Team <absl-team@google.com>: Add benchmark for accessing a Duration flag. PiperOrigin-RevId: 346594843 -- fefdb046520871af63ce2229e2f7cccfc0483dea by Abseil Team <absl-team@google.com>: Restructure CordReader for upcoming ring buffer changes. PiperOrigin-RevId: 346410642 -- 8b2f50e7da0ebab06ead5f94e366e984ca23cb6a by Abseil Team <absl-team@google.com>: Wire in an internal-only flag to toggle upcoming ring buffer changes on/off for experimentation. PiperOrigin-RevId: 346199111 GitOrigin-RevId: 6e9f93888bbe6718997ed90bbd049f1f3572b066 Change-Id: I8f34866b25a79209cb5448bbb28dd3044111d2e9
4 years ago
if (CordRep* tree = contents_.tree()) {
CordRep::Unref(VerifyTree(tree));
}
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
}
// --------------------------------------------------------------------
// Mutators
void Cord::Clear() {
Export of internal Abseil changes -- 6e9f93888bbe6718997ed90bbd049f1f3572b066 by Abseil Team <absl-team@google.com>: Fix status to be safe for self move assignment. PiperOrigin-RevId: 346815392 -- 35cae74a977f258e81dfbe925fb5a34cb6632036 by Gennadiy Rozental <rogeeff@google.com>: Eliminate unnecessary access to the global vars. PiperOrigin-RevId: 346777168 -- e7e786c243069060f5d6c1c945decb4b0b83f95b by Andy Getzendanner <durandal@google.com>: Internal change. PiperOrigin-RevId: 346685656 -- 4ccd41c48f1a83cfa20b3ea534f743dd7d788376 by Abseil Team <absl-team@google.com>: Move CordRep Ref() and Unref() logic into cord_internal.h This change moves Ref() and Unref() logic out of cord.cc into cord_internal. The main purpose is to make upcoming ring buffer changes easier to isolate from existing cord.cc code. Notice that this removes the nullptr check from Unref() and now requires it to be non null (which held true most times). We may need to rethink if the 'unref unlikely one' is the common case: are cordreps most likely shared? This may be something between 'root' and non root nodes, i.e., is it more likely for leaf / flat nodes in large cords to be shared than top level cordreps being shared? Vice versa? Benchmarks say that we mostly shouldn't care, the caveat being that atomic ops seem more expensive on upcoming archs (arcadia) so we should error on the side of an extra IsOne() branch saving us single owned atomic ops. PiperOrigin-RevId: 346676121 -- f0babab103b9e60d61ba09482d468985e43eceb3 by Samuel Benzaquen <sbenza@google.com>: Fix iterator based constructor and `.insert` members to only require EmplaceConstructible as the standard specifies. PiperOrigin-RevId: 346616707 -- 8f48eedda02277f9c96a88ed7726e34b557cce20 by Evan Brown <ezb@google.com>: Fix a bug in binary_search_impl when there's a transparent, three-way comparator that has different equivalence classes for different lookup types. Add a new can_have_multiple_equivalent_keys method to share the common logic for these cases. PiperOrigin-RevId: 346605948 -- 649183cb3cc9383431de9c81fb1c0f885d4001ae by Abseil Team <absl-team@google.com>: Add benchmark for accessing a Duration flag. PiperOrigin-RevId: 346594843 -- fefdb046520871af63ce2229e2f7cccfc0483dea by Abseil Team <absl-team@google.com>: Restructure CordReader for upcoming ring buffer changes. PiperOrigin-RevId: 346410642 -- 8b2f50e7da0ebab06ead5f94e366e984ca23cb6a by Abseil Team <absl-team@google.com>: Wire in an internal-only flag to toggle upcoming ring buffer changes on/off for experimentation. PiperOrigin-RevId: 346199111 GitOrigin-RevId: 6e9f93888bbe6718997ed90bbd049f1f3572b066 Change-Id: I8f34866b25a79209cb5448bbb28dd3044111d2e9
4 years ago
if (CordRep* tree = contents_.clear()) {
CordRep::Unref(tree);
}
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
}
Cord& Cord::operator=(absl::string_view src) {
Export of internal Abseil changes -- 341670bce317dd6af8d3c066970230591a47e80c by Martijn Vels <mvels@google.com>: Change GetStack() and GetParentStack() to return absl::Span PiperOrigin-RevId: 368765721 -- 6aaab9536d6957303c7aba100c3afaa6fb0ea2c8 by Martijn Vels <mvels@google.com>: Remove locking from parent stack. This change removes the need to lock all access to `parent_stack' by making the 'copy constructor' logic specify the 'copied from' CordzInfo (where available) to the TrackCord function, after which parent_stack is immutable. PiperOrigin-RevId: 368760630 -- b19e2059cada35a8ede994833018edac94de6ddc by Martijn Vels <mvels@google.com>: Add cordz instrumentation to Cord PiperOrigin-RevId: 368746225 -- 67b8bbf980f0f4e1db79aa32968e9a715a09b51a by Martijn Vels <mvels@google.com>: Create ABSL_INTERNAL_CORDZ_ENABLED define controlling when Cordz code is enabled There are specific builds and condtions under which we don't support cordz sampling, which is per this change represented by ABSL_INTERNAL_CORDZ_ENABLED being defined. PiperOrigin-RevId: 368731603 -- 8cbfe0e3169637a620f4b66ad2bc2ce340879cb0 by Martijn Vels <mvels@google.com>: Add a `rep` property to CordzInfo to be managed by Cord logic. This change adds a `rep` property to CordzInfo, which is intended to be used by collection logic. Mini design: Cord invokes TrackCord() providing the active 'root' cordrep of the newly sampled Cord, returning a CordzInfo with a weak (uncounted) reference to this root. Cord invokes `SetCordRep()` each time the root cordrep of the sampled Cord is updated while holding `mutex()`. Cord must also obtain `mutex()` _before_ removing a reference on the old root. i.e.: Cord must guarantee that the (weak) reference held in CordzInfo is at all times valid. CordzInfo collection code can then safely obtain a (reference counted) rep pointer by adding a reference to `rep_` while holding `mutex()`. This requires only a very brief critical section inside CordzInfo logic, minimizing contention with concurrent Cord updates. Cord code should typically obtain and hold `mutex()` for the entirety of each mutating Cord operation on a sampled cord. As Cord is thread compatible, it never competes on the lock with any other thread. The only possible concurrent access is from Cordz collection code, which should be a relatively rare event. PiperOrigin-RevId: 368673758 -- 1255120dce2bdd6b4205a34a0e555e0b74b6152f by Martijn Vels <mvels@google.com>: Remove 'depth' from active recorded metrics. Going forward we do not 'live' record depth (and size), but will observe these at collection time only. PiperOrigin-RevId: 368636572 -- 83e5146e35f221736b49e9f0a8805f8c159a51db by Martijn Vels <mvels@google.com>: Make cordz targets visible in OSS PiperOrigin-RevId: 368615010 -- dcb16a4f1239151f0a8c70a8cfeb29dabbd113b8 by Martijn Vels <mvels@google.com>: Internal cleanup PiperOrigin-RevId: 368514666 GitOrigin-RevId: 341670bce317dd6af8d3c066970230591a47e80c Change-Id: I94cecfbbd441eb386f99fc5186c468a7a5538862
4 years ago
if (ABSL_PREDICT_FALSE(contents_.is_profiled())) {
absl::cord_internal::CordzInfo::UntrackCord(contents_.cordz_info());
contents_.clear_cordz_info();
}
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
const char* data = src.data();
size_t length = src.size();
CordRep* tree = contents_.tree();
if (length <= InlineRep::kMaxInline) {
// Embed into this->contents_
contents_.set_data(data, length, true);
Export of internal Abseil changes -- 6e9f93888bbe6718997ed90bbd049f1f3572b066 by Abseil Team <absl-team@google.com>: Fix status to be safe for self move assignment. PiperOrigin-RevId: 346815392 -- 35cae74a977f258e81dfbe925fb5a34cb6632036 by Gennadiy Rozental <rogeeff@google.com>: Eliminate unnecessary access to the global vars. PiperOrigin-RevId: 346777168 -- e7e786c243069060f5d6c1c945decb4b0b83f95b by Andy Getzendanner <durandal@google.com>: Internal change. PiperOrigin-RevId: 346685656 -- 4ccd41c48f1a83cfa20b3ea534f743dd7d788376 by Abseil Team <absl-team@google.com>: Move CordRep Ref() and Unref() logic into cord_internal.h This change moves Ref() and Unref() logic out of cord.cc into cord_internal. The main purpose is to make upcoming ring buffer changes easier to isolate from existing cord.cc code. Notice that this removes the nullptr check from Unref() and now requires it to be non null (which held true most times). We may need to rethink if the 'unref unlikely one' is the common case: are cordreps most likely shared? This may be something between 'root' and non root nodes, i.e., is it more likely for leaf / flat nodes in large cords to be shared than top level cordreps being shared? Vice versa? Benchmarks say that we mostly shouldn't care, the caveat being that atomic ops seem more expensive on upcoming archs (arcadia) so we should error on the side of an extra IsOne() branch saving us single owned atomic ops. PiperOrigin-RevId: 346676121 -- f0babab103b9e60d61ba09482d468985e43eceb3 by Samuel Benzaquen <sbenza@google.com>: Fix iterator based constructor and `.insert` members to only require EmplaceConstructible as the standard specifies. PiperOrigin-RevId: 346616707 -- 8f48eedda02277f9c96a88ed7726e34b557cce20 by Evan Brown <ezb@google.com>: Fix a bug in binary_search_impl when there's a transparent, three-way comparator that has different equivalence classes for different lookup types. Add a new can_have_multiple_equivalent_keys method to share the common logic for these cases. PiperOrigin-RevId: 346605948 -- 649183cb3cc9383431de9c81fb1c0f885d4001ae by Abseil Team <absl-team@google.com>: Add benchmark for accessing a Duration flag. PiperOrigin-RevId: 346594843 -- fefdb046520871af63ce2229e2f7cccfc0483dea by Abseil Team <absl-team@google.com>: Restructure CordReader for upcoming ring buffer changes. PiperOrigin-RevId: 346410642 -- 8b2f50e7da0ebab06ead5f94e366e984ca23cb6a by Abseil Team <absl-team@google.com>: Wire in an internal-only flag to toggle upcoming ring buffer changes on/off for experimentation. PiperOrigin-RevId: 346199111 GitOrigin-RevId: 6e9f93888bbe6718997ed90bbd049f1f3572b066 Change-Id: I8f34866b25a79209cb5448bbb28dd3044111d2e9
4 years ago
if (tree) CordRep::Unref(tree);
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
return *this;
}
if (tree != nullptr && tree->tag >= FLAT &&
Export of internal Abseil changes -- ff793052bd01e1e4fcf639f94d7c30c4855a9372 by Evan Brown <ezb@google.com>: Roll forward of btree_iterator refactoring. PiperOrigin-RevId: 346116047 -- 17984679f16e3e2139b0f14fa76f4a6ca16a3ef9 by Chris Kennelly <ckennelly@google.com>: Extend absl::StrContains to accept single character needles. Single characters are more efficient to search for. Extending this API allows the abseil-string-find-str-contains Clang Tidy to include this pattern. The C++ committee has adopted http://wg21.link/P1679 for inclusion in C++23. PiperOrigin-RevId: 346095060 -- ef20b31c501b1dcaa25e244fd8f8aa43dec09bd6 by Jorg Brown <jorg@google.com>: Internal change for cord ring PiperOrigin-RevId: 346087545 -- b70f2c1cb77fc9e733a126e790967d45c5fd1dc7 by Derek Mauro <dmauro@google.com>: Release layout_benchmark PiperOrigin-RevId: 345968909 -- 3a0eda337ee43622f92cfe14c2aa06f72dc71ee5 by Derek Mauro <dmauro@google.com>: Release raw_hash_set_probe_benchmark PiperOrigin-RevId: 345965969 -- abffdb4bb241a2264cb4e73a6262b660bb10447d by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 345733599 -- 7c9e24a71188df945be17fe98f700bdb51f81b16 by Derek Mauro <dmauro@google.com>: Release hash_benchmark PiperOrigin-RevId: 345721635 -- d68f33f17f9a8cd3f6da8eee3870bdb46402cdc8 by Derek Mauro <dmauro@google.com>: Release raw_hash_set_benchmark PiperOrigin-RevId: 345708384 -- 6e6c547d4d1327b226c0ffe8ff34d0aa103ce24b by Abseil Team <absl-team@google.com>: Updates the implementation of InlinedVector to accurately express the value-initialization semantics of the default constructor PiperOrigin-RevId: 345548260 -- 1532424deda97d468444c217cc0fa4614099c7c1 by Evan Brown <ezb@google.com>: Rollback btree_iterator refactoring. PiperOrigin-RevId: 345543900 GitOrigin-RevId: ff793052bd01e1e4fcf639f94d7c30c4855a9372 Change-Id: I719831981fd056de41939f9addfee3d85e3b49b2
4 years ago
tree->flat()->Capacity() >= length &&
tree->refcount.IsOne()) {
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
// Copy in place if the existing FLAT node is reusable.
Export of internal Abseil changes -- a0491c8d790972cd80e2d720fe1fdf5f711a6f1a by Greg Falcon <gfalcon@google.com>: Stop directly accessing CordRepFlat data via CordRep::data. The old pattern of access breaks the `CordRep` type abstraction; since `CordRep::data` is not in general guaranteed to contain the chunk's data, we shouldn't access it that way. This incidentally adds an assertion check (via the flat() accessor) that the CordRep is indeed flat on each such access, but a manual inspection of the code, as well as the fact that this code currently works, suggest that this is always true.) PiperOrigin-RevId: 351592344 -- f40c3b43ca5b1d7e23cd45f1ffac1783105ac1a3 by Abseil Team <absl-team@google.com>: Revert 18abb2902b9f06c63a968b24d3dda785ebf99a22 PiperOrigin-RevId: 351523518 -- 18abb2902b9f06c63a968b24d3dda785ebf99a22 by Abseil Team <absl-team@google.com>: Internal change PiperOrigin-RevId: 351512412 -- 9b881602d45e95e06089792c7627cd56528a255a by Abseil Team <absl-team@google.com>: Keep time's global state in a cacheline-aligned structure. Keeping the global state as separate global variables results in two issues: 1) False sharing with adjacent global data (e.g., cycle clock source), since the global fields are updated every O(10usec). 2) The hot global fields (e.g., seq and samples) can reside on different cache lines. To fix this, simply wrap the global data in a ABSL_CACHE_ALIGNED structure. This is similar to what we do for MutexGlobals. PiperOrigin-RevId: 351389466 GitOrigin-RevId: a0491c8d790972cd80e2d720fe1fdf5f711a6f1a Change-Id: Ie0fa80112043381cd37c84e2ab2b7334839f54b5
4 years ago
memmove(tree->flat()->Data(), data, length);
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
tree->length = length;
VerifyTree(tree);
return *this;
}
contents_.set_tree(NewTree(data, length, 0));
Export of internal Abseil changes -- 6e9f93888bbe6718997ed90bbd049f1f3572b066 by Abseil Team <absl-team@google.com>: Fix status to be safe for self move assignment. PiperOrigin-RevId: 346815392 -- 35cae74a977f258e81dfbe925fb5a34cb6632036 by Gennadiy Rozental <rogeeff@google.com>: Eliminate unnecessary access to the global vars. PiperOrigin-RevId: 346777168 -- e7e786c243069060f5d6c1c945decb4b0b83f95b by Andy Getzendanner <durandal@google.com>: Internal change. PiperOrigin-RevId: 346685656 -- 4ccd41c48f1a83cfa20b3ea534f743dd7d788376 by Abseil Team <absl-team@google.com>: Move CordRep Ref() and Unref() logic into cord_internal.h This change moves Ref() and Unref() logic out of cord.cc into cord_internal. The main purpose is to make upcoming ring buffer changes easier to isolate from existing cord.cc code. Notice that this removes the nullptr check from Unref() and now requires it to be non null (which held true most times). We may need to rethink if the 'unref unlikely one' is the common case: are cordreps most likely shared? This may be something between 'root' and non root nodes, i.e., is it more likely for leaf / flat nodes in large cords to be shared than top level cordreps being shared? Vice versa? Benchmarks say that we mostly shouldn't care, the caveat being that atomic ops seem more expensive on upcoming archs (arcadia) so we should error on the side of an extra IsOne() branch saving us single owned atomic ops. PiperOrigin-RevId: 346676121 -- f0babab103b9e60d61ba09482d468985e43eceb3 by Samuel Benzaquen <sbenza@google.com>: Fix iterator based constructor and `.insert` members to only require EmplaceConstructible as the standard specifies. PiperOrigin-RevId: 346616707 -- 8f48eedda02277f9c96a88ed7726e34b557cce20 by Evan Brown <ezb@google.com>: Fix a bug in binary_search_impl when there's a transparent, three-way comparator that has different equivalence classes for different lookup types. Add a new can_have_multiple_equivalent_keys method to share the common logic for these cases. PiperOrigin-RevId: 346605948 -- 649183cb3cc9383431de9c81fb1c0f885d4001ae by Abseil Team <absl-team@google.com>: Add benchmark for accessing a Duration flag. PiperOrigin-RevId: 346594843 -- fefdb046520871af63ce2229e2f7cccfc0483dea by Abseil Team <absl-team@google.com>: Restructure CordReader for upcoming ring buffer changes. PiperOrigin-RevId: 346410642 -- 8b2f50e7da0ebab06ead5f94e366e984ca23cb6a by Abseil Team <absl-team@google.com>: Wire in an internal-only flag to toggle upcoming ring buffer changes on/off for experimentation. PiperOrigin-RevId: 346199111 GitOrigin-RevId: 6e9f93888bbe6718997ed90bbd049f1f3572b066 Change-Id: I8f34866b25a79209cb5448bbb28dd3044111d2e9
4 years ago
if (tree) CordRep::Unref(tree);
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
return *this;
}
Export of internal Abseil changes -- 7d0468a6610ed85586d5c87fd65de8dac5118923 by Derek Mauro <dmauro@google.com>: Import of CCTZ from GitHub. PiperOrigin-RevId: 313226473 -- 1131ef6d116f5ce7d46537a82f300ea06dcaaa53 by Gennadiy Rozental <rogeeff@google.com>: Migrate internal interface to use mutable references. PiperOrigin-RevId: 312931131 -- 96225212a9f5fbd0b38c71fe65539164992c7c3b by Laramie Leavitt <lar@google.com>: Remove random/internal/distributions.h This file was something of an historical artifact. All of the related code has either been removed or migraged, and so the only remaining type belongs with uniform_helper.h, as it is used to infer the return type of the absl::Uniform method in a few cases. PiperOrigin-RevId: 312878173 -- 6dcbd5be58ad425e08740ff64088373ee7fe4a72 by Mark Barolak <mbar@google.com>: Release the StrFormat test case for Cords to open source. PiperOrigin-RevId: 312707974 -- 34484d18dfb63a0a7ad6e2aaeb570e33592968be by Abseil Team <absl-team@google.com>: Let Cord::Cord(string&&), Cord::operator=(string&&), Cord::Append(string&&), and Cord::Prepend(string&&) steal string data and embed it into the Cord as a single external chunk, instead of copying it into flat chunks (at most 4083-byte each). Stealing string data is faster, but it creates a long chunk, which leads to a higher more memory usage if its subcords are created and outlive the whole Cord. These functions revert to copying the data if any of the following conditions holds: - string size is at most kMaxBytesToCopy (511), to avoid the overhead of an external chunk for short strings; - less than half of string capacity is used, to avoid pinning to much unused memory. PiperOrigin-RevId: 312683785 GitOrigin-RevId: 7d0468a6610ed85586d5c87fd65de8dac5118923 Change-Id: If79b5a1dfe6d53a8ddddbc7da84338f11fc4cfa3
5 years ago
template <typename T, Cord::EnableIfString<T>>
Cord& Cord::operator=(T&& src) {
if (src.size() <= kMaxBytesToCopy) {
*this = absl::string_view(src);
} else {
*this = Cord(std::forward<T>(src));
Export of internal Abseil changes -- 7d0468a6610ed85586d5c87fd65de8dac5118923 by Derek Mauro <dmauro@google.com>: Import of CCTZ from GitHub. PiperOrigin-RevId: 313226473 -- 1131ef6d116f5ce7d46537a82f300ea06dcaaa53 by Gennadiy Rozental <rogeeff@google.com>: Migrate internal interface to use mutable references. PiperOrigin-RevId: 312931131 -- 96225212a9f5fbd0b38c71fe65539164992c7c3b by Laramie Leavitt <lar@google.com>: Remove random/internal/distributions.h This file was something of an historical artifact. All of the related code has either been removed or migraged, and so the only remaining type belongs with uniform_helper.h, as it is used to infer the return type of the absl::Uniform method in a few cases. PiperOrigin-RevId: 312878173 -- 6dcbd5be58ad425e08740ff64088373ee7fe4a72 by Mark Barolak <mbar@google.com>: Release the StrFormat test case for Cords to open source. PiperOrigin-RevId: 312707974 -- 34484d18dfb63a0a7ad6e2aaeb570e33592968be by Abseil Team <absl-team@google.com>: Let Cord::Cord(string&&), Cord::operator=(string&&), Cord::Append(string&&), and Cord::Prepend(string&&) steal string data and embed it into the Cord as a single external chunk, instead of copying it into flat chunks (at most 4083-byte each). Stealing string data is faster, but it creates a long chunk, which leads to a higher more memory usage if its subcords are created and outlive the whole Cord. These functions revert to copying the data if any of the following conditions holds: - string size is at most kMaxBytesToCopy (511), to avoid the overhead of an external chunk for short strings; - less than half of string capacity is used, to avoid pinning to much unused memory. PiperOrigin-RevId: 312683785 GitOrigin-RevId: 7d0468a6610ed85586d5c87fd65de8dac5118923 Change-Id: If79b5a1dfe6d53a8ddddbc7da84338f11fc4cfa3
5 years ago
}
return *this;
}
template Cord& Cord::operator=(std::string&& src);
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
// TODO(sanjay): Move to Cord::InlineRep section of file. For now,
// we keep it here to make diffs easier.
void Cord::InlineRep::AppendArray(const char* src_data, size_t src_size) {
if (src_size == 0) return; // memcpy(_, nullptr, 0) is undefined.
size_t appended = 0;
Export of internal Abseil changes -- 642ab296a2c9629c44f3f2ce6911cd2488bcf416 by Derek Mauro <dmauro@google.com>: Remove an obsolete check in CMakeLists.txt PiperOrigin-RevId: 352852564 -- ce78cb96bcfd162737dbcf35005da3d1d6a3486b by Abseil Team <absl-team@google.com>: Clarify that the calling *thread* must have locked the mutex in order to unlock it. PiperOrigin-RevId: 352801804 -- 24e1f5f72756046f5265abf618e951c341f09b8d by Derek Mauro <dmauro@google.com>: Fixes failing CMake string comparisons https://cmake.org/cmake/help/latest/policy/CMP0054.html Fixes #791 PiperOrigin-RevId: 352791054 -- 0ac10bc3f4dca2c4c4b51d7b8196a2eaee9537a1 by Abseil Team <absl-team@google.com>: Introduce CordRepRing class This change introduces the CordRepRing class that implements all the lower level / internal implementation for upcoming CordRepRing ring buffer support in cord. PiperOrigin-RevId: 352771994 -- 4bd36dda61760785844f0f29f26d90cc18046f75 by Abseil Team <absl-team@google.com>: Optimize InlineData representation for cord sampling (cordz) This CL changes InlineData to allow us to store a (future) Cordz Info pointer directly into the inline representation: - make InlineData a class that provides a public API to set the active union members (tree or chars) and safely access that data. - change 'tree' and 'profiled' bits to be the 2 least significant bits, allowing us 62 continquous bits for storing a Cordz Info pointer. PiperOrigin-RevId: 352642411 -- dc55ba71bbce0e6a83e05a453990c51ac3d68426 by Mark Barolak <mbar@google.com>: Add unit test coverage for the mutating overload of absl::AsciiStrToLower. PiperOrigin-RevId: 352626006 GitOrigin-RevId: 642ab296a2c9629c44f3f2ce6911cd2488bcf416 Change-Id: I6c5929dd830d3c630e14e7fd5387fc3e25a69100
4 years ago
CordRep* root = nullptr;
if (is_tree()) {
root = data_.as_tree();
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
char* region;
if (PrepareAppendRegion(root, &region, &appended, src_size)) {
memcpy(region, src_data, appended);
Export of internal Abseil changes -- 341670bce317dd6af8d3c066970230591a47e80c by Martijn Vels <mvels@google.com>: Change GetStack() and GetParentStack() to return absl::Span PiperOrigin-RevId: 368765721 -- 6aaab9536d6957303c7aba100c3afaa6fb0ea2c8 by Martijn Vels <mvels@google.com>: Remove locking from parent stack. This change removes the need to lock all access to `parent_stack' by making the 'copy constructor' logic specify the 'copied from' CordzInfo (where available) to the TrackCord function, after which parent_stack is immutable. PiperOrigin-RevId: 368760630 -- b19e2059cada35a8ede994833018edac94de6ddc by Martijn Vels <mvels@google.com>: Add cordz instrumentation to Cord PiperOrigin-RevId: 368746225 -- 67b8bbf980f0f4e1db79aa32968e9a715a09b51a by Martijn Vels <mvels@google.com>: Create ABSL_INTERNAL_CORDZ_ENABLED define controlling when Cordz code is enabled There are specific builds and condtions under which we don't support cordz sampling, which is per this change represented by ABSL_INTERNAL_CORDZ_ENABLED being defined. PiperOrigin-RevId: 368731603 -- 8cbfe0e3169637a620f4b66ad2bc2ce340879cb0 by Martijn Vels <mvels@google.com>: Add a `rep` property to CordzInfo to be managed by Cord logic. This change adds a `rep` property to CordzInfo, which is intended to be used by collection logic. Mini design: Cord invokes TrackCord() providing the active 'root' cordrep of the newly sampled Cord, returning a CordzInfo with a weak (uncounted) reference to this root. Cord invokes `SetCordRep()` each time the root cordrep of the sampled Cord is updated while holding `mutex()`. Cord must also obtain `mutex()` _before_ removing a reference on the old root. i.e.: Cord must guarantee that the (weak) reference held in CordzInfo is at all times valid. CordzInfo collection code can then safely obtain a (reference counted) rep pointer by adding a reference to `rep_` while holding `mutex()`. This requires only a very brief critical section inside CordzInfo logic, minimizing contention with concurrent Cord updates. Cord code should typically obtain and hold `mutex()` for the entirety of each mutating Cord operation on a sampled cord. As Cord is thread compatible, it never competes on the lock with any other thread. The only possible concurrent access is from Cordz collection code, which should be a relatively rare event. PiperOrigin-RevId: 368673758 -- 1255120dce2bdd6b4205a34a0e555e0b74b6152f by Martijn Vels <mvels@google.com>: Remove 'depth' from active recorded metrics. Going forward we do not 'live' record depth (and size), but will observe these at collection time only. PiperOrigin-RevId: 368636572 -- 83e5146e35f221736b49e9f0a8805f8c159a51db by Martijn Vels <mvels@google.com>: Make cordz targets visible in OSS PiperOrigin-RevId: 368615010 -- dcb16a4f1239151f0a8c70a8cfeb29dabbd113b8 by Martijn Vels <mvels@google.com>: Internal cleanup PiperOrigin-RevId: 368514666 GitOrigin-RevId: 341670bce317dd6af8d3c066970230591a47e80c Change-Id: I94cecfbbd441eb386f99fc5186c468a7a5538862
4 years ago
UpdateCordzStatistics();
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
}
} else {
Export of internal Abseil changes -- 642ab296a2c9629c44f3f2ce6911cd2488bcf416 by Derek Mauro <dmauro@google.com>: Remove an obsolete check in CMakeLists.txt PiperOrigin-RevId: 352852564 -- ce78cb96bcfd162737dbcf35005da3d1d6a3486b by Abseil Team <absl-team@google.com>: Clarify that the calling *thread* must have locked the mutex in order to unlock it. PiperOrigin-RevId: 352801804 -- 24e1f5f72756046f5265abf618e951c341f09b8d by Derek Mauro <dmauro@google.com>: Fixes failing CMake string comparisons https://cmake.org/cmake/help/latest/policy/CMP0054.html Fixes #791 PiperOrigin-RevId: 352791054 -- 0ac10bc3f4dca2c4c4b51d7b8196a2eaee9537a1 by Abseil Team <absl-team@google.com>: Introduce CordRepRing class This change introduces the CordRepRing class that implements all the lower level / internal implementation for upcoming CordRepRing ring buffer support in cord. PiperOrigin-RevId: 352771994 -- 4bd36dda61760785844f0f29f26d90cc18046f75 by Abseil Team <absl-team@google.com>: Optimize InlineData representation for cord sampling (cordz) This CL changes InlineData to allow us to store a (future) Cordz Info pointer directly into the inline representation: - make InlineData a class that provides a public API to set the active union members (tree or chars) and safely access that data. - change 'tree' and 'profiled' bits to be the 2 least significant bits, allowing us 62 continquous bits for storing a Cordz Info pointer. PiperOrigin-RevId: 352642411 -- dc55ba71bbce0e6a83e05a453990c51ac3d68426 by Mark Barolak <mbar@google.com>: Add unit test coverage for the mutating overload of absl::AsciiStrToLower. PiperOrigin-RevId: 352626006 GitOrigin-RevId: 642ab296a2c9629c44f3f2ce6911cd2488bcf416 Change-Id: I6c5929dd830d3c630e14e7fd5387fc3e25a69100
4 years ago
// Try to fit in the inline buffer if possible.
size_t inline_length = inline_size();
if (src_size <= kMaxInline - inline_length) {
// Append new data to embedded array
memcpy(data_.as_chars() + inline_length, src_data, src_size);
set_inline_size(inline_length + src_size);
return;
}
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
// It is possible that src_data == data_, but when we transition from an
// InlineRep to a tree we need to assign data_ = root via set_tree. To
// avoid corrupting the source data before we copy it, delay calling
// set_tree until after we've copied data.
// We are going from an inline size to beyond inline size. Make the new size
// either double the inlined size, or the added size + 10%.
const size_t size1 = inline_length * 2 + src_size;
const size_t size2 = inline_length + src_size / 10;
root = CordRepFlat::New(std::max<size_t>(size1, size2));
appended = std::min(
Export of internal Abseil changes -- ff793052bd01e1e4fcf639f94d7c30c4855a9372 by Evan Brown <ezb@google.com>: Roll forward of btree_iterator refactoring. PiperOrigin-RevId: 346116047 -- 17984679f16e3e2139b0f14fa76f4a6ca16a3ef9 by Chris Kennelly <ckennelly@google.com>: Extend absl::StrContains to accept single character needles. Single characters are more efficient to search for. Extending this API allows the abseil-string-find-str-contains Clang Tidy to include this pattern. The C++ committee has adopted http://wg21.link/P1679 for inclusion in C++23. PiperOrigin-RevId: 346095060 -- ef20b31c501b1dcaa25e244fd8f8aa43dec09bd6 by Jorg Brown <jorg@google.com>: Internal change for cord ring PiperOrigin-RevId: 346087545 -- b70f2c1cb77fc9e733a126e790967d45c5fd1dc7 by Derek Mauro <dmauro@google.com>: Release layout_benchmark PiperOrigin-RevId: 345968909 -- 3a0eda337ee43622f92cfe14c2aa06f72dc71ee5 by Derek Mauro <dmauro@google.com>: Release raw_hash_set_probe_benchmark PiperOrigin-RevId: 345965969 -- abffdb4bb241a2264cb4e73a6262b660bb10447d by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 345733599 -- 7c9e24a71188df945be17fe98f700bdb51f81b16 by Derek Mauro <dmauro@google.com>: Release hash_benchmark PiperOrigin-RevId: 345721635 -- d68f33f17f9a8cd3f6da8eee3870bdb46402cdc8 by Derek Mauro <dmauro@google.com>: Release raw_hash_set_benchmark PiperOrigin-RevId: 345708384 -- 6e6c547d4d1327b226c0ffe8ff34d0aa103ce24b by Abseil Team <absl-team@google.com>: Updates the implementation of InlinedVector to accurately express the value-initialization semantics of the default constructor PiperOrigin-RevId: 345548260 -- 1532424deda97d468444c217cc0fa4614099c7c1 by Evan Brown <ezb@google.com>: Rollback btree_iterator refactoring. PiperOrigin-RevId: 345543900 GitOrigin-RevId: ff793052bd01e1e4fcf639f94d7c30c4855a9372 Change-Id: I719831981fd056de41939f9addfee3d85e3b49b2
4 years ago
src_size, root->flat()->Capacity() - inline_length);
Export of internal Abseil changes -- 642ab296a2c9629c44f3f2ce6911cd2488bcf416 by Derek Mauro <dmauro@google.com>: Remove an obsolete check in CMakeLists.txt PiperOrigin-RevId: 352852564 -- ce78cb96bcfd162737dbcf35005da3d1d6a3486b by Abseil Team <absl-team@google.com>: Clarify that the calling *thread* must have locked the mutex in order to unlock it. PiperOrigin-RevId: 352801804 -- 24e1f5f72756046f5265abf618e951c341f09b8d by Derek Mauro <dmauro@google.com>: Fixes failing CMake string comparisons https://cmake.org/cmake/help/latest/policy/CMP0054.html Fixes #791 PiperOrigin-RevId: 352791054 -- 0ac10bc3f4dca2c4c4b51d7b8196a2eaee9537a1 by Abseil Team <absl-team@google.com>: Introduce CordRepRing class This change introduces the CordRepRing class that implements all the lower level / internal implementation for upcoming CordRepRing ring buffer support in cord. PiperOrigin-RevId: 352771994 -- 4bd36dda61760785844f0f29f26d90cc18046f75 by Abseil Team <absl-team@google.com>: Optimize InlineData representation for cord sampling (cordz) This CL changes InlineData to allow us to store a (future) Cordz Info pointer directly into the inline representation: - make InlineData a class that provides a public API to set the active union members (tree or chars) and safely access that data. - change 'tree' and 'profiled' bits to be the 2 least significant bits, allowing us 62 continquous bits for storing a Cordz Info pointer. PiperOrigin-RevId: 352642411 -- dc55ba71bbce0e6a83e05a453990c51ac3d68426 by Mark Barolak <mbar@google.com>: Add unit test coverage for the mutating overload of absl::AsciiStrToLower. PiperOrigin-RevId: 352626006 GitOrigin-RevId: 642ab296a2c9629c44f3f2ce6911cd2488bcf416 Change-Id: I6c5929dd830d3c630e14e7fd5387fc3e25a69100
4 years ago
memcpy(root->flat()->Data(), data_.as_chars(), inline_length);
Export of internal Abseil changes -- a0491c8d790972cd80e2d720fe1fdf5f711a6f1a by Greg Falcon <gfalcon@google.com>: Stop directly accessing CordRepFlat data via CordRep::data. The old pattern of access breaks the `CordRep` type abstraction; since `CordRep::data` is not in general guaranteed to contain the chunk's data, we shouldn't access it that way. This incidentally adds an assertion check (via the flat() accessor) that the CordRep is indeed flat on each such access, but a manual inspection of the code, as well as the fact that this code currently works, suggest that this is always true.) PiperOrigin-RevId: 351592344 -- f40c3b43ca5b1d7e23cd45f1ffac1783105ac1a3 by Abseil Team <absl-team@google.com>: Revert 18abb2902b9f06c63a968b24d3dda785ebf99a22 PiperOrigin-RevId: 351523518 -- 18abb2902b9f06c63a968b24d3dda785ebf99a22 by Abseil Team <absl-team@google.com>: Internal change PiperOrigin-RevId: 351512412 -- 9b881602d45e95e06089792c7627cd56528a255a by Abseil Team <absl-team@google.com>: Keep time's global state in a cacheline-aligned structure. Keeping the global state as separate global variables results in two issues: 1) False sharing with adjacent global data (e.g., cycle clock source), since the global fields are updated every O(10usec). 2) The hot global fields (e.g., seq and samples) can reside on different cache lines. To fix this, simply wrap the global data in a ABSL_CACHE_ALIGNED structure. This is similar to what we do for MutexGlobals. PiperOrigin-RevId: 351389466 GitOrigin-RevId: a0491c8d790972cd80e2d720fe1fdf5f711a6f1a Change-Id: Ie0fa80112043381cd37c84e2ab2b7334839f54b5
4 years ago
memcpy(root->flat()->Data() + inline_length, src_data, appended);
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
root->length = inline_length + appended;
set_tree(root);
}
src_data += appended;
src_size -= appended;
if (src_size == 0) {
return;
}
Export of internal Abseil changes -- c68f1886f5e8fd90eb0c2d2e68feaf00a7cdacda by CJ Johnson <johnsoncj@google.com>: Introduce absl::Cleanup to the OSS repo PiperOrigin-RevId: 354583156 -- 17030cf388e10f7eb959e3e566326d1072ce392e by Abseil Team <absl-team@google.com>: Internal change only PiperOrigin-RevId: 354574953 -- e979d7236d4f3252e79ddda6739b67a9a326bf6d by CJ Johnson <johnsoncj@google.com>: Internal change PiperOrigin-RevId: 354545297 -- 7ea02b3783f7f49ef97d86a8f6580a19cc57df14 by Abseil Team <absl-team@google.com>: Pre-allocate memory for vectors where the size is known. PiperOrigin-RevId: 354344576 -- 9246c7cb11f1d6444f79ebe25acc69a8a9b870e0 by Matt Kulukundis <kfm@google.com>: Add support for Elbrus 2000 (e2k) Import of https://github.com/abseil/abseil-cpp/pull/889 PiperOrigin-RevId: 354344013 -- 0fc93d359cc1fb307552e917b37b7b2e7eed822f by Abseil Team <absl-team@google.com>: Integrate CordRepRing logic into cord (but do not enable it) PiperOrigin-RevId: 354312238 -- eda05622f7da71466723acb33403f783529df24b by Abseil Team <absl-team@google.com>: Protect ignore diagnostic with "__has_warning". PiperOrigin-RevId: 354112334 -- 47716c5d8fb10efa4fdd801d28bac414c6f8ec32 by Abseil Team <absl-team@google.com>: Rearrange InlinedVector copy constructor and destructor to treat a few special cases inline and then tail-call a non-inlined routine for the rest. In particular, we optimize for empty vectors in both cases. Added a couple of benchmarks that copy either an InlVec<int64> or an InlVec<InlVec<int64>>. Speed difference: ``` BM_CopyTrivial/0 0.92ns +- 0% 0.47ns +- 0% -48.91% (p=0.000 n=11+12) BM_CopyTrivial/1 0.92ns +- 0% 1.15ns +- 0% +25.00% (p=0.000 n=10+9) BM_CopyTrivial/8 8.57ns +- 0% 10.72ns +- 1% +25.16% (p=0.000 n=10+12) BM_CopyNonTrivial/0 3.21ns +- 0% 0.70ns +- 0% -78.23% (p=0.000 n=12+10) BM_CopyNonTrivial/1 5.88ns +- 1% 5.51ns +- 0% -6.28% (p=0.000 n=10+8) BM_CopyNonTrivial/8 21.5ns +- 1% 15.2ns +- 2% -29.23% (p=0.000 n=12+12) ``` Note: the slowdowns are a few cycles which is expected given the procedure call added in that case. We decided this is a good tradeoff given the code size reductions and the more significant speedups for empty vectors. Size difference (as measured by nm): ``` BM_CopyTrivial from 1048 bytes to 326 bytes. BM_CopyNonTrivial from 749 bytes to 470 bytes. ``` Code size for a large binary drops by ~500KB (from 349415719 to 348906015 348906191). All of the benchmarks that showed a significant difference: Ones that improve with this CL: ``` BM_CopyNonTrivial/0 3.21ns +- 0% 0.70ns +- 0% -78.23% (p=0.000 n=12+10) BM_InlinedVectorFillString/0 0.93ns +- 0% 0.24ns +- 0% -74.19% (p=0.000 n=12+10) BM_InlinedVectorAssignments/1 10.5ns +- 0% 4.1ns +- 0% -60.64% (p=0.000 n=11+10) BM_InlinedVectorAssignments/2 10.7ns +- 0% 4.4ns +- 0% -59.08% (p=0.000 n=11+11) BM_CopyTrivial/0 0.92ns +- 0% 0.47ns +- 0% -48.91% (p=0.000 n=11+12) BM_CopyNonTrivial/8 21.5ns +- 1% 15.2ns +- 2% -29.23% (p=0.000 n=12+12) BM_StdVectorEmpty 0.47ns +- 1% 0.35ns +- 0% -24.73% (p=0.000 n=12+12) BM_StdVectorSize 0.46ns +- 2% 0.35ns +- 0% -24.32% (p=0.000 n=12+12) BM_SwapElements<LargeCopyableOnly>/0 3.44ns +- 0% 2.76ns +- 1% -19.83% (p=0.000 n=11+11) BM_InlinedVectorFillRange/256 20.7ns +- 1% 17.8ns +- 0% -14.08% (p=0.000 n=12+9) BM_CopyNonTrivial/1 5.88ns +- 1% 5.51ns +- 0% -6.28% (p=0.000 n=10+8) BM_SwapElements<LargeCopyableMovable>/1 4.19ns +- 0% 3.95ns +- 1% -5.63% (p=0.000 n=11+12) BM_SwapElements<LargeCopyableMovableSwappable>/1 4.18ns +- 0% 3.99ns +- 0% -4.70% (p=0.000 n=9+11) BM_SwapElements<LargeCopyableMovable>/0 2.41ns +- 0% 2.31ns +- 0% -4.45% (p=0.000 n=12+12) BM_InlinedVectorFillRange/64 8.25ns +- 0% 8.04ns +- 0% -2.51% (p=0.000 n=12+11) BM_SwapElements<LargeCopyableOnly>/1 82.4ns +- 0% 81.5ns +- 0% -1.06% (p=0.000 n=12+12) ``` Ones that get worse with this CL: ``` BM_CopyTrivial/1 0.92ns +- 0% 1.15ns +- 0% +25.00% (p=0.000 n=10+9) BM_CopyTrivial/8 8.57ns +- 0% 10.72ns +- 1% +25.16% (p=0.000 n=10+12) BM_SwapElements<LargeCopyableMovableSwappable>/512 1.48ns +- 1% 1.66ns +- 1% +11.88% (p=0.000 n=12+12) BM_InlinedVectorFillString/1 11.5ns +- 0% 12.8ns +- 1% +11.62% (p=0.000 n=12+11) BM_SwapElements<LargeCopyableMovableSwappable>/64 1.48ns +- 2% 1.66ns +- 1% +11.66% (p=0.000 n=12+11) BM_SwapElements<LargeCopyableMovableSwappable>/1k 1.48ns +- 1% 1.65ns +- 2% +11.32% (p=0.000 n=12+12) BM_SwapElements<LargeCopyableMovable>/512 1.48ns +- 2% 1.58ns +- 4% +6.62% (p=0.000 n=11+12) BM_SwapElements<LargeCopyableMovable>/1k 1.49ns +- 2% 1.58ns +- 3% +6.05% (p=0.000 n=12+12) BM_SwapElements<LargeCopyableMovable>/64 1.48ns +- 2% 1.57ns +- 4% +6.04% (p=0.000 n=11+12) BM_InlinedVectorFillRange/1 4.81ns +- 0% 5.05ns +- 0% +4.83% (p=0.000 n=11+11) BM_InlinedVectorFillString/8 79.4ns +- 1% 83.1ns +- 1% +4.64% (p=0.000 n=10+12) BM_StdVectorFillString/1 16.3ns +- 0% 16.6ns +- 0% +2.13% (p=0.000 n=11+8) ``` PiperOrigin-RevId: 353906786 -- 8e26518b3cec9c598e5e9573c46c3bd1b03a67ef by Abseil Team <absl-team@google.com>: Internal change PiperOrigin-RevId: 353737330 -- f206ae0983e58c9904ed8b8f05f9caf564a446be by Matt Kulukundis <kfm@google.com>: Import of CCTZ from GitHub. PiperOrigin-RevId: 353682256 GitOrigin-RevId: c68f1886f5e8fd90eb0c2d2e68feaf00a7cdacda Change-Id: I5790c1036c4f543c701d1039848fabf7ae881ad8
4 years ago
if (cord_ring_enabled()) {
absl::string_view data(src_data, src_size);
root = ForceRing(root, (data.size() - 1) / kMaxFlatLength + 1);
replace_tree(CordRepRing::Append(root->ring(), data));
return;
}
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
// Use new block(s) for any remaining bytes that were not handled above.
// Alloc extra memory only if the right child of the root of the new tree is
// going to be a FLAT node, which will permit further inplace appends.
size_t length = src_size;
if (src_size < kMaxFlatLength) {
// The new length is either
// - old size + 10%
// - old_size + src_size
// This will cause a reasonable conservative step-up in size that is still
// large enough to avoid excessive amounts of small fragments being added.
length = std::max<size_t>(root->length / 10, src_size);
}
set_tree(Concat(root, NewTree(src_data, src_size, length - src_size)));
}
inline CordRep* Cord::TakeRep() const& {
Export of internal Abseil changes -- 6e9f93888bbe6718997ed90bbd049f1f3572b066 by Abseil Team <absl-team@google.com>: Fix status to be safe for self move assignment. PiperOrigin-RevId: 346815392 -- 35cae74a977f258e81dfbe925fb5a34cb6632036 by Gennadiy Rozental <rogeeff@google.com>: Eliminate unnecessary access to the global vars. PiperOrigin-RevId: 346777168 -- e7e786c243069060f5d6c1c945decb4b0b83f95b by Andy Getzendanner <durandal@google.com>: Internal change. PiperOrigin-RevId: 346685656 -- 4ccd41c48f1a83cfa20b3ea534f743dd7d788376 by Abseil Team <absl-team@google.com>: Move CordRep Ref() and Unref() logic into cord_internal.h This change moves Ref() and Unref() logic out of cord.cc into cord_internal. The main purpose is to make upcoming ring buffer changes easier to isolate from existing cord.cc code. Notice that this removes the nullptr check from Unref() and now requires it to be non null (which held true most times). We may need to rethink if the 'unref unlikely one' is the common case: are cordreps most likely shared? This may be something between 'root' and non root nodes, i.e., is it more likely for leaf / flat nodes in large cords to be shared than top level cordreps being shared? Vice versa? Benchmarks say that we mostly shouldn't care, the caveat being that atomic ops seem more expensive on upcoming archs (arcadia) so we should error on the side of an extra IsOne() branch saving us single owned atomic ops. PiperOrigin-RevId: 346676121 -- f0babab103b9e60d61ba09482d468985e43eceb3 by Samuel Benzaquen <sbenza@google.com>: Fix iterator based constructor and `.insert` members to only require EmplaceConstructible as the standard specifies. PiperOrigin-RevId: 346616707 -- 8f48eedda02277f9c96a88ed7726e34b557cce20 by Evan Brown <ezb@google.com>: Fix a bug in binary_search_impl when there's a transparent, three-way comparator that has different equivalence classes for different lookup types. Add a new can_have_multiple_equivalent_keys method to share the common logic for these cases. PiperOrigin-RevId: 346605948 -- 649183cb3cc9383431de9c81fb1c0f885d4001ae by Abseil Team <absl-team@google.com>: Add benchmark for accessing a Duration flag. PiperOrigin-RevId: 346594843 -- fefdb046520871af63ce2229e2f7cccfc0483dea by Abseil Team <absl-team@google.com>: Restructure CordReader for upcoming ring buffer changes. PiperOrigin-RevId: 346410642 -- 8b2f50e7da0ebab06ead5f94e366e984ca23cb6a by Abseil Team <absl-team@google.com>: Wire in an internal-only flag to toggle upcoming ring buffer changes on/off for experimentation. PiperOrigin-RevId: 346199111 GitOrigin-RevId: 6e9f93888bbe6718997ed90bbd049f1f3572b066 Change-Id: I8f34866b25a79209cb5448bbb28dd3044111d2e9
4 years ago
return CordRep::Ref(contents_.tree());
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
}
inline CordRep* Cord::TakeRep() && {
CordRep* rep = contents_.tree();
contents_.clear();
return rep;
}
template <typename C>
inline void Cord::AppendImpl(C&& src) {
if (empty()) {
// In case of an empty destination avoid allocating a new node, do not copy
// data.
*this = std::forward<C>(src);
return;
}
// For short cords, it is faster to copy data if there is room in dst.
const size_t src_size = src.contents_.size();
if (src_size <= kMaxBytesToCopy) {
CordRep* src_tree = src.contents_.tree();
if (src_tree == nullptr) {
// src has embedded data.
contents_.AppendArray(src.contents_.data(), src_size);
return;
}
if (src_tree->tag >= FLAT) {
// src tree just has one flat node.
Export of internal Abseil changes -- a0491c8d790972cd80e2d720fe1fdf5f711a6f1a by Greg Falcon <gfalcon@google.com>: Stop directly accessing CordRepFlat data via CordRep::data. The old pattern of access breaks the `CordRep` type abstraction; since `CordRep::data` is not in general guaranteed to contain the chunk's data, we shouldn't access it that way. This incidentally adds an assertion check (via the flat() accessor) that the CordRep is indeed flat on each such access, but a manual inspection of the code, as well as the fact that this code currently works, suggest that this is always true.) PiperOrigin-RevId: 351592344 -- f40c3b43ca5b1d7e23cd45f1ffac1783105ac1a3 by Abseil Team <absl-team@google.com>: Revert 18abb2902b9f06c63a968b24d3dda785ebf99a22 PiperOrigin-RevId: 351523518 -- 18abb2902b9f06c63a968b24d3dda785ebf99a22 by Abseil Team <absl-team@google.com>: Internal change PiperOrigin-RevId: 351512412 -- 9b881602d45e95e06089792c7627cd56528a255a by Abseil Team <absl-team@google.com>: Keep time's global state in a cacheline-aligned structure. Keeping the global state as separate global variables results in two issues: 1) False sharing with adjacent global data (e.g., cycle clock source), since the global fields are updated every O(10usec). 2) The hot global fields (e.g., seq and samples) can reside on different cache lines. To fix this, simply wrap the global data in a ABSL_CACHE_ALIGNED structure. This is similar to what we do for MutexGlobals. PiperOrigin-RevId: 351389466 GitOrigin-RevId: a0491c8d790972cd80e2d720fe1fdf5f711a6f1a Change-Id: Ie0fa80112043381cd37c84e2ab2b7334839f54b5
4 years ago
contents_.AppendArray(src_tree->flat()->Data(), src_size);
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
return;
}
if (&src == this) {
// ChunkIterator below assumes that src is not modified during traversal.
Append(Cord(src));
return;
}
// TODO(mec): Should we only do this if "dst" has space?
for (absl::string_view chunk : src.Chunks()) {
Append(chunk);
}
return;
}
Export of internal Abseil changes -- 6e9f93888bbe6718997ed90bbd049f1f3572b066 by Abseil Team <absl-team@google.com>: Fix status to be safe for self move assignment. PiperOrigin-RevId: 346815392 -- 35cae74a977f258e81dfbe925fb5a34cb6632036 by Gennadiy Rozental <rogeeff@google.com>: Eliminate unnecessary access to the global vars. PiperOrigin-RevId: 346777168 -- e7e786c243069060f5d6c1c945decb4b0b83f95b by Andy Getzendanner <durandal@google.com>: Internal change. PiperOrigin-RevId: 346685656 -- 4ccd41c48f1a83cfa20b3ea534f743dd7d788376 by Abseil Team <absl-team@google.com>: Move CordRep Ref() and Unref() logic into cord_internal.h This change moves Ref() and Unref() logic out of cord.cc into cord_internal. The main purpose is to make upcoming ring buffer changes easier to isolate from existing cord.cc code. Notice that this removes the nullptr check from Unref() and now requires it to be non null (which held true most times). We may need to rethink if the 'unref unlikely one' is the common case: are cordreps most likely shared? This may be something between 'root' and non root nodes, i.e., is it more likely for leaf / flat nodes in large cords to be shared than top level cordreps being shared? Vice versa? Benchmarks say that we mostly shouldn't care, the caveat being that atomic ops seem more expensive on upcoming archs (arcadia) so we should error on the side of an extra IsOne() branch saving us single owned atomic ops. PiperOrigin-RevId: 346676121 -- f0babab103b9e60d61ba09482d468985e43eceb3 by Samuel Benzaquen <sbenza@google.com>: Fix iterator based constructor and `.insert` members to only require EmplaceConstructible as the standard specifies. PiperOrigin-RevId: 346616707 -- 8f48eedda02277f9c96a88ed7726e34b557cce20 by Evan Brown <ezb@google.com>: Fix a bug in binary_search_impl when there's a transparent, three-way comparator that has different equivalence classes for different lookup types. Add a new can_have_multiple_equivalent_keys method to share the common logic for these cases. PiperOrigin-RevId: 346605948 -- 649183cb3cc9383431de9c81fb1c0f885d4001ae by Abseil Team <absl-team@google.com>: Add benchmark for accessing a Duration flag. PiperOrigin-RevId: 346594843 -- fefdb046520871af63ce2229e2f7cccfc0483dea by Abseil Team <absl-team@google.com>: Restructure CordReader for upcoming ring buffer changes. PiperOrigin-RevId: 346410642 -- 8b2f50e7da0ebab06ead5f94e366e984ca23cb6a by Abseil Team <absl-team@google.com>: Wire in an internal-only flag to toggle upcoming ring buffer changes on/off for experimentation. PiperOrigin-RevId: 346199111 GitOrigin-RevId: 6e9f93888bbe6718997ed90bbd049f1f3572b066 Change-Id: I8f34866b25a79209cb5448bbb28dd3044111d2e9
4 years ago
// Guaranteed to be a tree (kMaxBytesToCopy > kInlinedSize)
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
contents_.AppendTree(std::forward<C>(src).TakeRep());
}
void Cord::Append(const Cord& src) { AppendImpl(src); }
void Cord::Append(Cord&& src) { AppendImpl(std::move(src)); }
Export of internal Abseil changes -- 7d0468a6610ed85586d5c87fd65de8dac5118923 by Derek Mauro <dmauro@google.com>: Import of CCTZ from GitHub. PiperOrigin-RevId: 313226473 -- 1131ef6d116f5ce7d46537a82f300ea06dcaaa53 by Gennadiy Rozental <rogeeff@google.com>: Migrate internal interface to use mutable references. PiperOrigin-RevId: 312931131 -- 96225212a9f5fbd0b38c71fe65539164992c7c3b by Laramie Leavitt <lar@google.com>: Remove random/internal/distributions.h This file was something of an historical artifact. All of the related code has either been removed or migraged, and so the only remaining type belongs with uniform_helper.h, as it is used to infer the return type of the absl::Uniform method in a few cases. PiperOrigin-RevId: 312878173 -- 6dcbd5be58ad425e08740ff64088373ee7fe4a72 by Mark Barolak <mbar@google.com>: Release the StrFormat test case for Cords to open source. PiperOrigin-RevId: 312707974 -- 34484d18dfb63a0a7ad6e2aaeb570e33592968be by Abseil Team <absl-team@google.com>: Let Cord::Cord(string&&), Cord::operator=(string&&), Cord::Append(string&&), and Cord::Prepend(string&&) steal string data and embed it into the Cord as a single external chunk, instead of copying it into flat chunks (at most 4083-byte each). Stealing string data is faster, but it creates a long chunk, which leads to a higher more memory usage if its subcords are created and outlive the whole Cord. These functions revert to copying the data if any of the following conditions holds: - string size is at most kMaxBytesToCopy (511), to avoid the overhead of an external chunk for short strings; - less than half of string capacity is used, to avoid pinning to much unused memory. PiperOrigin-RevId: 312683785 GitOrigin-RevId: 7d0468a6610ed85586d5c87fd65de8dac5118923 Change-Id: If79b5a1dfe6d53a8ddddbc7da84338f11fc4cfa3
5 years ago
template <typename T, Cord::EnableIfString<T>>
void Cord::Append(T&& src) {
if (src.size() <= kMaxBytesToCopy) {
Append(absl::string_view(src));
} else {
Append(Cord(std::forward<T>(src)));
Export of internal Abseil changes -- 7d0468a6610ed85586d5c87fd65de8dac5118923 by Derek Mauro <dmauro@google.com>: Import of CCTZ from GitHub. PiperOrigin-RevId: 313226473 -- 1131ef6d116f5ce7d46537a82f300ea06dcaaa53 by Gennadiy Rozental <rogeeff@google.com>: Migrate internal interface to use mutable references. PiperOrigin-RevId: 312931131 -- 96225212a9f5fbd0b38c71fe65539164992c7c3b by Laramie Leavitt <lar@google.com>: Remove random/internal/distributions.h This file was something of an historical artifact. All of the related code has either been removed or migraged, and so the only remaining type belongs with uniform_helper.h, as it is used to infer the return type of the absl::Uniform method in a few cases. PiperOrigin-RevId: 312878173 -- 6dcbd5be58ad425e08740ff64088373ee7fe4a72 by Mark Barolak <mbar@google.com>: Release the StrFormat test case for Cords to open source. PiperOrigin-RevId: 312707974 -- 34484d18dfb63a0a7ad6e2aaeb570e33592968be by Abseil Team <absl-team@google.com>: Let Cord::Cord(string&&), Cord::operator=(string&&), Cord::Append(string&&), and Cord::Prepend(string&&) steal string data and embed it into the Cord as a single external chunk, instead of copying it into flat chunks (at most 4083-byte each). Stealing string data is faster, but it creates a long chunk, which leads to a higher more memory usage if its subcords are created and outlive the whole Cord. These functions revert to copying the data if any of the following conditions holds: - string size is at most kMaxBytesToCopy (511), to avoid the overhead of an external chunk for short strings; - less than half of string capacity is used, to avoid pinning to much unused memory. PiperOrigin-RevId: 312683785 GitOrigin-RevId: 7d0468a6610ed85586d5c87fd65de8dac5118923 Change-Id: If79b5a1dfe6d53a8ddddbc7da84338f11fc4cfa3
5 years ago
}
}
template void Cord::Append(std::string&& src);
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
void Cord::Prepend(const Cord& src) {
CordRep* src_tree = src.contents_.tree();
if (src_tree != nullptr) {
Export of internal Abseil changes -- 6e9f93888bbe6718997ed90bbd049f1f3572b066 by Abseil Team <absl-team@google.com>: Fix status to be safe for self move assignment. PiperOrigin-RevId: 346815392 -- 35cae74a977f258e81dfbe925fb5a34cb6632036 by Gennadiy Rozental <rogeeff@google.com>: Eliminate unnecessary access to the global vars. PiperOrigin-RevId: 346777168 -- e7e786c243069060f5d6c1c945decb4b0b83f95b by Andy Getzendanner <durandal@google.com>: Internal change. PiperOrigin-RevId: 346685656 -- 4ccd41c48f1a83cfa20b3ea534f743dd7d788376 by Abseil Team <absl-team@google.com>: Move CordRep Ref() and Unref() logic into cord_internal.h This change moves Ref() and Unref() logic out of cord.cc into cord_internal. The main purpose is to make upcoming ring buffer changes easier to isolate from existing cord.cc code. Notice that this removes the nullptr check from Unref() and now requires it to be non null (which held true most times). We may need to rethink if the 'unref unlikely one' is the common case: are cordreps most likely shared? This may be something between 'root' and non root nodes, i.e., is it more likely for leaf / flat nodes in large cords to be shared than top level cordreps being shared? Vice versa? Benchmarks say that we mostly shouldn't care, the caveat being that atomic ops seem more expensive on upcoming archs (arcadia) so we should error on the side of an extra IsOne() branch saving us single owned atomic ops. PiperOrigin-RevId: 346676121 -- f0babab103b9e60d61ba09482d468985e43eceb3 by Samuel Benzaquen <sbenza@google.com>: Fix iterator based constructor and `.insert` members to only require EmplaceConstructible as the standard specifies. PiperOrigin-RevId: 346616707 -- 8f48eedda02277f9c96a88ed7726e34b557cce20 by Evan Brown <ezb@google.com>: Fix a bug in binary_search_impl when there's a transparent, three-way comparator that has different equivalence classes for different lookup types. Add a new can_have_multiple_equivalent_keys method to share the common logic for these cases. PiperOrigin-RevId: 346605948 -- 649183cb3cc9383431de9c81fb1c0f885d4001ae by Abseil Team <absl-team@google.com>: Add benchmark for accessing a Duration flag. PiperOrigin-RevId: 346594843 -- fefdb046520871af63ce2229e2f7cccfc0483dea by Abseil Team <absl-team@google.com>: Restructure CordReader for upcoming ring buffer changes. PiperOrigin-RevId: 346410642 -- 8b2f50e7da0ebab06ead5f94e366e984ca23cb6a by Abseil Team <absl-team@google.com>: Wire in an internal-only flag to toggle upcoming ring buffer changes on/off for experimentation. PiperOrigin-RevId: 346199111 GitOrigin-RevId: 6e9f93888bbe6718997ed90bbd049f1f3572b066 Change-Id: I8f34866b25a79209cb5448bbb28dd3044111d2e9
4 years ago
CordRep::Ref(src_tree);
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
contents_.PrependTree(src_tree);
return;
}
// `src` cord is inlined.
absl::string_view src_contents(src.contents_.data(), src.contents_.size());
return Prepend(src_contents);
}
void Cord::Prepend(absl::string_view src) {
if (src.empty()) return; // memcpy(_, nullptr, 0) is undefined.
Export of internal Abseil changes -- 642ab296a2c9629c44f3f2ce6911cd2488bcf416 by Derek Mauro <dmauro@google.com>: Remove an obsolete check in CMakeLists.txt PiperOrigin-RevId: 352852564 -- ce78cb96bcfd162737dbcf35005da3d1d6a3486b by Abseil Team <absl-team@google.com>: Clarify that the calling *thread* must have locked the mutex in order to unlock it. PiperOrigin-RevId: 352801804 -- 24e1f5f72756046f5265abf618e951c341f09b8d by Derek Mauro <dmauro@google.com>: Fixes failing CMake string comparisons https://cmake.org/cmake/help/latest/policy/CMP0054.html Fixes #791 PiperOrigin-RevId: 352791054 -- 0ac10bc3f4dca2c4c4b51d7b8196a2eaee9537a1 by Abseil Team <absl-team@google.com>: Introduce CordRepRing class This change introduces the CordRepRing class that implements all the lower level / internal implementation for upcoming CordRepRing ring buffer support in cord. PiperOrigin-RevId: 352771994 -- 4bd36dda61760785844f0f29f26d90cc18046f75 by Abseil Team <absl-team@google.com>: Optimize InlineData representation for cord sampling (cordz) This CL changes InlineData to allow us to store a (future) Cordz Info pointer directly into the inline representation: - make InlineData a class that provides a public API to set the active union members (tree or chars) and safely access that data. - change 'tree' and 'profiled' bits to be the 2 least significant bits, allowing us 62 continquous bits for storing a Cordz Info pointer. PiperOrigin-RevId: 352642411 -- dc55ba71bbce0e6a83e05a453990c51ac3d68426 by Mark Barolak <mbar@google.com>: Add unit test coverage for the mutating overload of absl::AsciiStrToLower. PiperOrigin-RevId: 352626006 GitOrigin-RevId: 642ab296a2c9629c44f3f2ce6911cd2488bcf416 Change-Id: I6c5929dd830d3c630e14e7fd5387fc3e25a69100
4 years ago
if (!contents_.is_tree()) {
size_t cur_size = contents_.inline_size();
if (cur_size + src.size() <= InlineRep::kMaxInline) {
// Use embedded storage.
char data[InlineRep::kMaxInline + 1] = {0};
memcpy(data, src.data(), src.size());
memcpy(data + src.size(), contents_.data(), cur_size);
memcpy(contents_.data_.as_chars(), data, InlineRep::kMaxInline + 1);
contents_.set_inline_size(cur_size + src.size());
return;
}
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
}
Export of internal Abseil changes -- 642ab296a2c9629c44f3f2ce6911cd2488bcf416 by Derek Mauro <dmauro@google.com>: Remove an obsolete check in CMakeLists.txt PiperOrigin-RevId: 352852564 -- ce78cb96bcfd162737dbcf35005da3d1d6a3486b by Abseil Team <absl-team@google.com>: Clarify that the calling *thread* must have locked the mutex in order to unlock it. PiperOrigin-RevId: 352801804 -- 24e1f5f72756046f5265abf618e951c341f09b8d by Derek Mauro <dmauro@google.com>: Fixes failing CMake string comparisons https://cmake.org/cmake/help/latest/policy/CMP0054.html Fixes #791 PiperOrigin-RevId: 352791054 -- 0ac10bc3f4dca2c4c4b51d7b8196a2eaee9537a1 by Abseil Team <absl-team@google.com>: Introduce CordRepRing class This change introduces the CordRepRing class that implements all the lower level / internal implementation for upcoming CordRepRing ring buffer support in cord. PiperOrigin-RevId: 352771994 -- 4bd36dda61760785844f0f29f26d90cc18046f75 by Abseil Team <absl-team@google.com>: Optimize InlineData representation for cord sampling (cordz) This CL changes InlineData to allow us to store a (future) Cordz Info pointer directly into the inline representation: - make InlineData a class that provides a public API to set the active union members (tree or chars) and safely access that data. - change 'tree' and 'profiled' bits to be the 2 least significant bits, allowing us 62 continquous bits for storing a Cordz Info pointer. PiperOrigin-RevId: 352642411 -- dc55ba71bbce0e6a83e05a453990c51ac3d68426 by Mark Barolak <mbar@google.com>: Add unit test coverage for the mutating overload of absl::AsciiStrToLower. PiperOrigin-RevId: 352626006 GitOrigin-RevId: 642ab296a2c9629c44f3f2ce6911cd2488bcf416 Change-Id: I6c5929dd830d3c630e14e7fd5387fc3e25a69100
4 years ago
contents_.PrependTree(NewTree(src.data(), src.size(), 0));
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
}
Export of internal Abseil changes -- 7d0468a6610ed85586d5c87fd65de8dac5118923 by Derek Mauro <dmauro@google.com>: Import of CCTZ from GitHub. PiperOrigin-RevId: 313226473 -- 1131ef6d116f5ce7d46537a82f300ea06dcaaa53 by Gennadiy Rozental <rogeeff@google.com>: Migrate internal interface to use mutable references. PiperOrigin-RevId: 312931131 -- 96225212a9f5fbd0b38c71fe65539164992c7c3b by Laramie Leavitt <lar@google.com>: Remove random/internal/distributions.h This file was something of an historical artifact. All of the related code has either been removed or migraged, and so the only remaining type belongs with uniform_helper.h, as it is used to infer the return type of the absl::Uniform method in a few cases. PiperOrigin-RevId: 312878173 -- 6dcbd5be58ad425e08740ff64088373ee7fe4a72 by Mark Barolak <mbar@google.com>: Release the StrFormat test case for Cords to open source. PiperOrigin-RevId: 312707974 -- 34484d18dfb63a0a7ad6e2aaeb570e33592968be by Abseil Team <absl-team@google.com>: Let Cord::Cord(string&&), Cord::operator=(string&&), Cord::Append(string&&), and Cord::Prepend(string&&) steal string data and embed it into the Cord as a single external chunk, instead of copying it into flat chunks (at most 4083-byte each). Stealing string data is faster, but it creates a long chunk, which leads to a higher more memory usage if its subcords are created and outlive the whole Cord. These functions revert to copying the data if any of the following conditions holds: - string size is at most kMaxBytesToCopy (511), to avoid the overhead of an external chunk for short strings; - less than half of string capacity is used, to avoid pinning to much unused memory. PiperOrigin-RevId: 312683785 GitOrigin-RevId: 7d0468a6610ed85586d5c87fd65de8dac5118923 Change-Id: If79b5a1dfe6d53a8ddddbc7da84338f11fc4cfa3
5 years ago
template <typename T, Cord::EnableIfString<T>>
inline void Cord::Prepend(T&& src) {
if (src.size() <= kMaxBytesToCopy) {
Prepend(absl::string_view(src));
} else {
Prepend(Cord(std::forward<T>(src)));
Export of internal Abseil changes -- 7d0468a6610ed85586d5c87fd65de8dac5118923 by Derek Mauro <dmauro@google.com>: Import of CCTZ from GitHub. PiperOrigin-RevId: 313226473 -- 1131ef6d116f5ce7d46537a82f300ea06dcaaa53 by Gennadiy Rozental <rogeeff@google.com>: Migrate internal interface to use mutable references. PiperOrigin-RevId: 312931131 -- 96225212a9f5fbd0b38c71fe65539164992c7c3b by Laramie Leavitt <lar@google.com>: Remove random/internal/distributions.h This file was something of an historical artifact. All of the related code has either been removed or migraged, and so the only remaining type belongs with uniform_helper.h, as it is used to infer the return type of the absl::Uniform method in a few cases. PiperOrigin-RevId: 312878173 -- 6dcbd5be58ad425e08740ff64088373ee7fe4a72 by Mark Barolak <mbar@google.com>: Release the StrFormat test case for Cords to open source. PiperOrigin-RevId: 312707974 -- 34484d18dfb63a0a7ad6e2aaeb570e33592968be by Abseil Team <absl-team@google.com>: Let Cord::Cord(string&&), Cord::operator=(string&&), Cord::Append(string&&), and Cord::Prepend(string&&) steal string data and embed it into the Cord as a single external chunk, instead of copying it into flat chunks (at most 4083-byte each). Stealing string data is faster, but it creates a long chunk, which leads to a higher more memory usage if its subcords are created and outlive the whole Cord. These functions revert to copying the data if any of the following conditions holds: - string size is at most kMaxBytesToCopy (511), to avoid the overhead of an external chunk for short strings; - less than half of string capacity is used, to avoid pinning to much unused memory. PiperOrigin-RevId: 312683785 GitOrigin-RevId: 7d0468a6610ed85586d5c87fd65de8dac5118923 Change-Id: If79b5a1dfe6d53a8ddddbc7da84338f11fc4cfa3
5 years ago
}
}
template void Cord::Prepend(std::string&& src);
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
static CordRep* RemovePrefixFrom(CordRep* node, size_t n) {
if (n >= node->length) return nullptr;
Export of internal Abseil changes -- 6e9f93888bbe6718997ed90bbd049f1f3572b066 by Abseil Team <absl-team@google.com>: Fix status to be safe for self move assignment. PiperOrigin-RevId: 346815392 -- 35cae74a977f258e81dfbe925fb5a34cb6632036 by Gennadiy Rozental <rogeeff@google.com>: Eliminate unnecessary access to the global vars. PiperOrigin-RevId: 346777168 -- e7e786c243069060f5d6c1c945decb4b0b83f95b by Andy Getzendanner <durandal@google.com>: Internal change. PiperOrigin-RevId: 346685656 -- 4ccd41c48f1a83cfa20b3ea534f743dd7d788376 by Abseil Team <absl-team@google.com>: Move CordRep Ref() and Unref() logic into cord_internal.h This change moves Ref() and Unref() logic out of cord.cc into cord_internal. The main purpose is to make upcoming ring buffer changes easier to isolate from existing cord.cc code. Notice that this removes the nullptr check from Unref() and now requires it to be non null (which held true most times). We may need to rethink if the 'unref unlikely one' is the common case: are cordreps most likely shared? This may be something between 'root' and non root nodes, i.e., is it more likely for leaf / flat nodes in large cords to be shared than top level cordreps being shared? Vice versa? Benchmarks say that we mostly shouldn't care, the caveat being that atomic ops seem more expensive on upcoming archs (arcadia) so we should error on the side of an extra IsOne() branch saving us single owned atomic ops. PiperOrigin-RevId: 346676121 -- f0babab103b9e60d61ba09482d468985e43eceb3 by Samuel Benzaquen <sbenza@google.com>: Fix iterator based constructor and `.insert` members to only require EmplaceConstructible as the standard specifies. PiperOrigin-RevId: 346616707 -- 8f48eedda02277f9c96a88ed7726e34b557cce20 by Evan Brown <ezb@google.com>: Fix a bug in binary_search_impl when there's a transparent, three-way comparator that has different equivalence classes for different lookup types. Add a new can_have_multiple_equivalent_keys method to share the common logic for these cases. PiperOrigin-RevId: 346605948 -- 649183cb3cc9383431de9c81fb1c0f885d4001ae by Abseil Team <absl-team@google.com>: Add benchmark for accessing a Duration flag. PiperOrigin-RevId: 346594843 -- fefdb046520871af63ce2229e2f7cccfc0483dea by Abseil Team <absl-team@google.com>: Restructure CordReader for upcoming ring buffer changes. PiperOrigin-RevId: 346410642 -- 8b2f50e7da0ebab06ead5f94e366e984ca23cb6a by Abseil Team <absl-team@google.com>: Wire in an internal-only flag to toggle upcoming ring buffer changes on/off for experimentation. PiperOrigin-RevId: 346199111 GitOrigin-RevId: 6e9f93888bbe6718997ed90bbd049f1f3572b066 Change-Id: I8f34866b25a79209cb5448bbb28dd3044111d2e9
4 years ago
if (n == 0) return CordRep::Ref(node);
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
absl::InlinedVector<CordRep*, kInlinedVectorSize> rhs_stack;
while (node->tag == CONCAT) {
assert(n <= node->length);
if (n < node->concat()->left->length) {
// Push right to stack, descend left.
rhs_stack.push_back(node->concat()->right);
node = node->concat()->left;
} else {
// Drop left, descend right.
n -= node->concat()->left->length;
node = node->concat()->right;
}
}
assert(n <= node->length);
if (n == 0) {
Export of internal Abseil changes -- 6e9f93888bbe6718997ed90bbd049f1f3572b066 by Abseil Team <absl-team@google.com>: Fix status to be safe for self move assignment. PiperOrigin-RevId: 346815392 -- 35cae74a977f258e81dfbe925fb5a34cb6632036 by Gennadiy Rozental <rogeeff@google.com>: Eliminate unnecessary access to the global vars. PiperOrigin-RevId: 346777168 -- e7e786c243069060f5d6c1c945decb4b0b83f95b by Andy Getzendanner <durandal@google.com>: Internal change. PiperOrigin-RevId: 346685656 -- 4ccd41c48f1a83cfa20b3ea534f743dd7d788376 by Abseil Team <absl-team@google.com>: Move CordRep Ref() and Unref() logic into cord_internal.h This change moves Ref() and Unref() logic out of cord.cc into cord_internal. The main purpose is to make upcoming ring buffer changes easier to isolate from existing cord.cc code. Notice that this removes the nullptr check from Unref() and now requires it to be non null (which held true most times). We may need to rethink if the 'unref unlikely one' is the common case: are cordreps most likely shared? This may be something between 'root' and non root nodes, i.e., is it more likely for leaf / flat nodes in large cords to be shared than top level cordreps being shared? Vice versa? Benchmarks say that we mostly shouldn't care, the caveat being that atomic ops seem more expensive on upcoming archs (arcadia) so we should error on the side of an extra IsOne() branch saving us single owned atomic ops. PiperOrigin-RevId: 346676121 -- f0babab103b9e60d61ba09482d468985e43eceb3 by Samuel Benzaquen <sbenza@google.com>: Fix iterator based constructor and `.insert` members to only require EmplaceConstructible as the standard specifies. PiperOrigin-RevId: 346616707 -- 8f48eedda02277f9c96a88ed7726e34b557cce20 by Evan Brown <ezb@google.com>: Fix a bug in binary_search_impl when there's a transparent, three-way comparator that has different equivalence classes for different lookup types. Add a new can_have_multiple_equivalent_keys method to share the common logic for these cases. PiperOrigin-RevId: 346605948 -- 649183cb3cc9383431de9c81fb1c0f885d4001ae by Abseil Team <absl-team@google.com>: Add benchmark for accessing a Duration flag. PiperOrigin-RevId: 346594843 -- fefdb046520871af63ce2229e2f7cccfc0483dea by Abseil Team <absl-team@google.com>: Restructure CordReader for upcoming ring buffer changes. PiperOrigin-RevId: 346410642 -- 8b2f50e7da0ebab06ead5f94e366e984ca23cb6a by Abseil Team <absl-team@google.com>: Wire in an internal-only flag to toggle upcoming ring buffer changes on/off for experimentation. PiperOrigin-RevId: 346199111 GitOrigin-RevId: 6e9f93888bbe6718997ed90bbd049f1f3572b066 Change-Id: I8f34866b25a79209cb5448bbb28dd3044111d2e9
4 years ago
CordRep::Ref(node);
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
} else {
size_t start = n;
size_t len = node->length - n;
if (node->tag == SUBSTRING) {
// Consider in-place update of node, similar to in RemoveSuffixFrom().
start += node->substring()->start;
node = node->substring()->child;
}
Export of internal Abseil changes -- 6e9f93888bbe6718997ed90bbd049f1f3572b066 by Abseil Team <absl-team@google.com>: Fix status to be safe for self move assignment. PiperOrigin-RevId: 346815392 -- 35cae74a977f258e81dfbe925fb5a34cb6632036 by Gennadiy Rozental <rogeeff@google.com>: Eliminate unnecessary access to the global vars. PiperOrigin-RevId: 346777168 -- e7e786c243069060f5d6c1c945decb4b0b83f95b by Andy Getzendanner <durandal@google.com>: Internal change. PiperOrigin-RevId: 346685656 -- 4ccd41c48f1a83cfa20b3ea534f743dd7d788376 by Abseil Team <absl-team@google.com>: Move CordRep Ref() and Unref() logic into cord_internal.h This change moves Ref() and Unref() logic out of cord.cc into cord_internal. The main purpose is to make upcoming ring buffer changes easier to isolate from existing cord.cc code. Notice that this removes the nullptr check from Unref() and now requires it to be non null (which held true most times). We may need to rethink if the 'unref unlikely one' is the common case: are cordreps most likely shared? This may be something between 'root' and non root nodes, i.e., is it more likely for leaf / flat nodes in large cords to be shared than top level cordreps being shared? Vice versa? Benchmarks say that we mostly shouldn't care, the caveat being that atomic ops seem more expensive on upcoming archs (arcadia) so we should error on the side of an extra IsOne() branch saving us single owned atomic ops. PiperOrigin-RevId: 346676121 -- f0babab103b9e60d61ba09482d468985e43eceb3 by Samuel Benzaquen <sbenza@google.com>: Fix iterator based constructor and `.insert` members to only require EmplaceConstructible as the standard specifies. PiperOrigin-RevId: 346616707 -- 8f48eedda02277f9c96a88ed7726e34b557cce20 by Evan Brown <ezb@google.com>: Fix a bug in binary_search_impl when there's a transparent, three-way comparator that has different equivalence classes for different lookup types. Add a new can_have_multiple_equivalent_keys method to share the common logic for these cases. PiperOrigin-RevId: 346605948 -- 649183cb3cc9383431de9c81fb1c0f885d4001ae by Abseil Team <absl-team@google.com>: Add benchmark for accessing a Duration flag. PiperOrigin-RevId: 346594843 -- fefdb046520871af63ce2229e2f7cccfc0483dea by Abseil Team <absl-team@google.com>: Restructure CordReader for upcoming ring buffer changes. PiperOrigin-RevId: 346410642 -- 8b2f50e7da0ebab06ead5f94e366e984ca23cb6a by Abseil Team <absl-team@google.com>: Wire in an internal-only flag to toggle upcoming ring buffer changes on/off for experimentation. PiperOrigin-RevId: 346199111 GitOrigin-RevId: 6e9f93888bbe6718997ed90bbd049f1f3572b066 Change-Id: I8f34866b25a79209cb5448bbb28dd3044111d2e9
4 years ago
node = NewSubstring(CordRep::Ref(node), start, len);
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
}
while (!rhs_stack.empty()) {
Export of internal Abseil changes -- 6e9f93888bbe6718997ed90bbd049f1f3572b066 by Abseil Team <absl-team@google.com>: Fix status to be safe for self move assignment. PiperOrigin-RevId: 346815392 -- 35cae74a977f258e81dfbe925fb5a34cb6632036 by Gennadiy Rozental <rogeeff@google.com>: Eliminate unnecessary access to the global vars. PiperOrigin-RevId: 346777168 -- e7e786c243069060f5d6c1c945decb4b0b83f95b by Andy Getzendanner <durandal@google.com>: Internal change. PiperOrigin-RevId: 346685656 -- 4ccd41c48f1a83cfa20b3ea534f743dd7d788376 by Abseil Team <absl-team@google.com>: Move CordRep Ref() and Unref() logic into cord_internal.h This change moves Ref() and Unref() logic out of cord.cc into cord_internal. The main purpose is to make upcoming ring buffer changes easier to isolate from existing cord.cc code. Notice that this removes the nullptr check from Unref() and now requires it to be non null (which held true most times). We may need to rethink if the 'unref unlikely one' is the common case: are cordreps most likely shared? This may be something between 'root' and non root nodes, i.e., is it more likely for leaf / flat nodes in large cords to be shared than top level cordreps being shared? Vice versa? Benchmarks say that we mostly shouldn't care, the caveat being that atomic ops seem more expensive on upcoming archs (arcadia) so we should error on the side of an extra IsOne() branch saving us single owned atomic ops. PiperOrigin-RevId: 346676121 -- f0babab103b9e60d61ba09482d468985e43eceb3 by Samuel Benzaquen <sbenza@google.com>: Fix iterator based constructor and `.insert` members to only require EmplaceConstructible as the standard specifies. PiperOrigin-RevId: 346616707 -- 8f48eedda02277f9c96a88ed7726e34b557cce20 by Evan Brown <ezb@google.com>: Fix a bug in binary_search_impl when there's a transparent, three-way comparator that has different equivalence classes for different lookup types. Add a new can_have_multiple_equivalent_keys method to share the common logic for these cases. PiperOrigin-RevId: 346605948 -- 649183cb3cc9383431de9c81fb1c0f885d4001ae by Abseil Team <absl-team@google.com>: Add benchmark for accessing a Duration flag. PiperOrigin-RevId: 346594843 -- fefdb046520871af63ce2229e2f7cccfc0483dea by Abseil Team <absl-team@google.com>: Restructure CordReader for upcoming ring buffer changes. PiperOrigin-RevId: 346410642 -- 8b2f50e7da0ebab06ead5f94e366e984ca23cb6a by Abseil Team <absl-team@google.com>: Wire in an internal-only flag to toggle upcoming ring buffer changes on/off for experimentation. PiperOrigin-RevId: 346199111 GitOrigin-RevId: 6e9f93888bbe6718997ed90bbd049f1f3572b066 Change-Id: I8f34866b25a79209cb5448bbb28dd3044111d2e9
4 years ago
node = Concat(node, CordRep::Ref(rhs_stack.back()));
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
rhs_stack.pop_back();
}
return node;
}
// RemoveSuffixFrom() is very similar to RemovePrefixFrom(), with the
// exception that removing a suffix has an optimization where a node may be
// edited in place iff that node and all its ancestors have a refcount of 1.
static CordRep* RemoveSuffixFrom(CordRep* node, size_t n) {
if (n >= node->length) return nullptr;
Export of internal Abseil changes -- 6e9f93888bbe6718997ed90bbd049f1f3572b066 by Abseil Team <absl-team@google.com>: Fix status to be safe for self move assignment. PiperOrigin-RevId: 346815392 -- 35cae74a977f258e81dfbe925fb5a34cb6632036 by Gennadiy Rozental <rogeeff@google.com>: Eliminate unnecessary access to the global vars. PiperOrigin-RevId: 346777168 -- e7e786c243069060f5d6c1c945decb4b0b83f95b by Andy Getzendanner <durandal@google.com>: Internal change. PiperOrigin-RevId: 346685656 -- 4ccd41c48f1a83cfa20b3ea534f743dd7d788376 by Abseil Team <absl-team@google.com>: Move CordRep Ref() and Unref() logic into cord_internal.h This change moves Ref() and Unref() logic out of cord.cc into cord_internal. The main purpose is to make upcoming ring buffer changes easier to isolate from existing cord.cc code. Notice that this removes the nullptr check from Unref() and now requires it to be non null (which held true most times). We may need to rethink if the 'unref unlikely one' is the common case: are cordreps most likely shared? This may be something between 'root' and non root nodes, i.e., is it more likely for leaf / flat nodes in large cords to be shared than top level cordreps being shared? Vice versa? Benchmarks say that we mostly shouldn't care, the caveat being that atomic ops seem more expensive on upcoming archs (arcadia) so we should error on the side of an extra IsOne() branch saving us single owned atomic ops. PiperOrigin-RevId: 346676121 -- f0babab103b9e60d61ba09482d468985e43eceb3 by Samuel Benzaquen <sbenza@google.com>: Fix iterator based constructor and `.insert` members to only require EmplaceConstructible as the standard specifies. PiperOrigin-RevId: 346616707 -- 8f48eedda02277f9c96a88ed7726e34b557cce20 by Evan Brown <ezb@google.com>: Fix a bug in binary_search_impl when there's a transparent, three-way comparator that has different equivalence classes for different lookup types. Add a new can_have_multiple_equivalent_keys method to share the common logic for these cases. PiperOrigin-RevId: 346605948 -- 649183cb3cc9383431de9c81fb1c0f885d4001ae by Abseil Team <absl-team@google.com>: Add benchmark for accessing a Duration flag. PiperOrigin-RevId: 346594843 -- fefdb046520871af63ce2229e2f7cccfc0483dea by Abseil Team <absl-team@google.com>: Restructure CordReader for upcoming ring buffer changes. PiperOrigin-RevId: 346410642 -- 8b2f50e7da0ebab06ead5f94e366e984ca23cb6a by Abseil Team <absl-team@google.com>: Wire in an internal-only flag to toggle upcoming ring buffer changes on/off for experimentation. PiperOrigin-RevId: 346199111 GitOrigin-RevId: 6e9f93888bbe6718997ed90bbd049f1f3572b066 Change-Id: I8f34866b25a79209cb5448bbb28dd3044111d2e9
4 years ago
if (n == 0) return CordRep::Ref(node);
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
absl::InlinedVector<CordRep*, kInlinedVectorSize> lhs_stack;
bool inplace_ok = node->refcount.IsOne();
while (node->tag == CONCAT) {
assert(n <= node->length);
if (n < node->concat()->right->length) {
// Push left to stack, descend right.
lhs_stack.push_back(node->concat()->left);
node = node->concat()->right;
} else {
// Drop right, descend left.
n -= node->concat()->right->length;
node = node->concat()->left;
}
inplace_ok = inplace_ok && node->refcount.IsOne();
}
assert(n <= node->length);
if (n == 0) {
Export of internal Abseil changes -- 6e9f93888bbe6718997ed90bbd049f1f3572b066 by Abseil Team <absl-team@google.com>: Fix status to be safe for self move assignment. PiperOrigin-RevId: 346815392 -- 35cae74a977f258e81dfbe925fb5a34cb6632036 by Gennadiy Rozental <rogeeff@google.com>: Eliminate unnecessary access to the global vars. PiperOrigin-RevId: 346777168 -- e7e786c243069060f5d6c1c945decb4b0b83f95b by Andy Getzendanner <durandal@google.com>: Internal change. PiperOrigin-RevId: 346685656 -- 4ccd41c48f1a83cfa20b3ea534f743dd7d788376 by Abseil Team <absl-team@google.com>: Move CordRep Ref() and Unref() logic into cord_internal.h This change moves Ref() and Unref() logic out of cord.cc into cord_internal. The main purpose is to make upcoming ring buffer changes easier to isolate from existing cord.cc code. Notice that this removes the nullptr check from Unref() and now requires it to be non null (which held true most times). We may need to rethink if the 'unref unlikely one' is the common case: are cordreps most likely shared? This may be something between 'root' and non root nodes, i.e., is it more likely for leaf / flat nodes in large cords to be shared than top level cordreps being shared? Vice versa? Benchmarks say that we mostly shouldn't care, the caveat being that atomic ops seem more expensive on upcoming archs (arcadia) so we should error on the side of an extra IsOne() branch saving us single owned atomic ops. PiperOrigin-RevId: 346676121 -- f0babab103b9e60d61ba09482d468985e43eceb3 by Samuel Benzaquen <sbenza@google.com>: Fix iterator based constructor and `.insert` members to only require EmplaceConstructible as the standard specifies. PiperOrigin-RevId: 346616707 -- 8f48eedda02277f9c96a88ed7726e34b557cce20 by Evan Brown <ezb@google.com>: Fix a bug in binary_search_impl when there's a transparent, three-way comparator that has different equivalence classes for different lookup types. Add a new can_have_multiple_equivalent_keys method to share the common logic for these cases. PiperOrigin-RevId: 346605948 -- 649183cb3cc9383431de9c81fb1c0f885d4001ae by Abseil Team <absl-team@google.com>: Add benchmark for accessing a Duration flag. PiperOrigin-RevId: 346594843 -- fefdb046520871af63ce2229e2f7cccfc0483dea by Abseil Team <absl-team@google.com>: Restructure CordReader for upcoming ring buffer changes. PiperOrigin-RevId: 346410642 -- 8b2f50e7da0ebab06ead5f94e366e984ca23cb6a by Abseil Team <absl-team@google.com>: Wire in an internal-only flag to toggle upcoming ring buffer changes on/off for experimentation. PiperOrigin-RevId: 346199111 GitOrigin-RevId: 6e9f93888bbe6718997ed90bbd049f1f3572b066 Change-Id: I8f34866b25a79209cb5448bbb28dd3044111d2e9
4 years ago
CordRep::Ref(node);
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
} else if (inplace_ok && node->tag != EXTERNAL) {
// Consider making a new buffer if the current node capacity is much
// larger than the new length.
Export of internal Abseil changes -- 6e9f93888bbe6718997ed90bbd049f1f3572b066 by Abseil Team <absl-team@google.com>: Fix status to be safe for self move assignment. PiperOrigin-RevId: 346815392 -- 35cae74a977f258e81dfbe925fb5a34cb6632036 by Gennadiy Rozental <rogeeff@google.com>: Eliminate unnecessary access to the global vars. PiperOrigin-RevId: 346777168 -- e7e786c243069060f5d6c1c945decb4b0b83f95b by Andy Getzendanner <durandal@google.com>: Internal change. PiperOrigin-RevId: 346685656 -- 4ccd41c48f1a83cfa20b3ea534f743dd7d788376 by Abseil Team <absl-team@google.com>: Move CordRep Ref() and Unref() logic into cord_internal.h This change moves Ref() and Unref() logic out of cord.cc into cord_internal. The main purpose is to make upcoming ring buffer changes easier to isolate from existing cord.cc code. Notice that this removes the nullptr check from Unref() and now requires it to be non null (which held true most times). We may need to rethink if the 'unref unlikely one' is the common case: are cordreps most likely shared? This may be something between 'root' and non root nodes, i.e., is it more likely for leaf / flat nodes in large cords to be shared than top level cordreps being shared? Vice versa? Benchmarks say that we mostly shouldn't care, the caveat being that atomic ops seem more expensive on upcoming archs (arcadia) so we should error on the side of an extra IsOne() branch saving us single owned atomic ops. PiperOrigin-RevId: 346676121 -- f0babab103b9e60d61ba09482d468985e43eceb3 by Samuel Benzaquen <sbenza@google.com>: Fix iterator based constructor and `.insert` members to only require EmplaceConstructible as the standard specifies. PiperOrigin-RevId: 346616707 -- 8f48eedda02277f9c96a88ed7726e34b557cce20 by Evan Brown <ezb@google.com>: Fix a bug in binary_search_impl when there's a transparent, three-way comparator that has different equivalence classes for different lookup types. Add a new can_have_multiple_equivalent_keys method to share the common logic for these cases. PiperOrigin-RevId: 346605948 -- 649183cb3cc9383431de9c81fb1c0f885d4001ae by Abseil Team <absl-team@google.com>: Add benchmark for accessing a Duration flag. PiperOrigin-RevId: 346594843 -- fefdb046520871af63ce2229e2f7cccfc0483dea by Abseil Team <absl-team@google.com>: Restructure CordReader for upcoming ring buffer changes. PiperOrigin-RevId: 346410642 -- 8b2f50e7da0ebab06ead5f94e366e984ca23cb6a by Abseil Team <absl-team@google.com>: Wire in an internal-only flag to toggle upcoming ring buffer changes on/off for experimentation. PiperOrigin-RevId: 346199111 GitOrigin-RevId: 6e9f93888bbe6718997ed90bbd049f1f3572b066 Change-Id: I8f34866b25a79209cb5448bbb28dd3044111d2e9
4 years ago
CordRep::Ref(node);
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
node->length -= n;
} else {
size_t start = 0;
size_t len = node->length - n;
if (node->tag == SUBSTRING) {
start = node->substring()->start;
node = node->substring()->child;
}
Export of internal Abseil changes -- 6e9f93888bbe6718997ed90bbd049f1f3572b066 by Abseil Team <absl-team@google.com>: Fix status to be safe for self move assignment. PiperOrigin-RevId: 346815392 -- 35cae74a977f258e81dfbe925fb5a34cb6632036 by Gennadiy Rozental <rogeeff@google.com>: Eliminate unnecessary access to the global vars. PiperOrigin-RevId: 346777168 -- e7e786c243069060f5d6c1c945decb4b0b83f95b by Andy Getzendanner <durandal@google.com>: Internal change. PiperOrigin-RevId: 346685656 -- 4ccd41c48f1a83cfa20b3ea534f743dd7d788376 by Abseil Team <absl-team@google.com>: Move CordRep Ref() and Unref() logic into cord_internal.h This change moves Ref() and Unref() logic out of cord.cc into cord_internal. The main purpose is to make upcoming ring buffer changes easier to isolate from existing cord.cc code. Notice that this removes the nullptr check from Unref() and now requires it to be non null (which held true most times). We may need to rethink if the 'unref unlikely one' is the common case: are cordreps most likely shared? This may be something between 'root' and non root nodes, i.e., is it more likely for leaf / flat nodes in large cords to be shared than top level cordreps being shared? Vice versa? Benchmarks say that we mostly shouldn't care, the caveat being that atomic ops seem more expensive on upcoming archs (arcadia) so we should error on the side of an extra IsOne() branch saving us single owned atomic ops. PiperOrigin-RevId: 346676121 -- f0babab103b9e60d61ba09482d468985e43eceb3 by Samuel Benzaquen <sbenza@google.com>: Fix iterator based constructor and `.insert` members to only require EmplaceConstructible as the standard specifies. PiperOrigin-RevId: 346616707 -- 8f48eedda02277f9c96a88ed7726e34b557cce20 by Evan Brown <ezb@google.com>: Fix a bug in binary_search_impl when there's a transparent, three-way comparator that has different equivalence classes for different lookup types. Add a new can_have_multiple_equivalent_keys method to share the common logic for these cases. PiperOrigin-RevId: 346605948 -- 649183cb3cc9383431de9c81fb1c0f885d4001ae by Abseil Team <absl-team@google.com>: Add benchmark for accessing a Duration flag. PiperOrigin-RevId: 346594843 -- fefdb046520871af63ce2229e2f7cccfc0483dea by Abseil Team <absl-team@google.com>: Restructure CordReader for upcoming ring buffer changes. PiperOrigin-RevId: 346410642 -- 8b2f50e7da0ebab06ead5f94e366e984ca23cb6a by Abseil Team <absl-team@google.com>: Wire in an internal-only flag to toggle upcoming ring buffer changes on/off for experimentation. PiperOrigin-RevId: 346199111 GitOrigin-RevId: 6e9f93888bbe6718997ed90bbd049f1f3572b066 Change-Id: I8f34866b25a79209cb5448bbb28dd3044111d2e9
4 years ago
node = NewSubstring(CordRep::Ref(node), start, len);
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
}
while (!lhs_stack.empty()) {
Export of internal Abseil changes -- 6e9f93888bbe6718997ed90bbd049f1f3572b066 by Abseil Team <absl-team@google.com>: Fix status to be safe for self move assignment. PiperOrigin-RevId: 346815392 -- 35cae74a977f258e81dfbe925fb5a34cb6632036 by Gennadiy Rozental <rogeeff@google.com>: Eliminate unnecessary access to the global vars. PiperOrigin-RevId: 346777168 -- e7e786c243069060f5d6c1c945decb4b0b83f95b by Andy Getzendanner <durandal@google.com>: Internal change. PiperOrigin-RevId: 346685656 -- 4ccd41c48f1a83cfa20b3ea534f743dd7d788376 by Abseil Team <absl-team@google.com>: Move CordRep Ref() and Unref() logic into cord_internal.h This change moves Ref() and Unref() logic out of cord.cc into cord_internal. The main purpose is to make upcoming ring buffer changes easier to isolate from existing cord.cc code. Notice that this removes the nullptr check from Unref() and now requires it to be non null (which held true most times). We may need to rethink if the 'unref unlikely one' is the common case: are cordreps most likely shared? This may be something between 'root' and non root nodes, i.e., is it more likely for leaf / flat nodes in large cords to be shared than top level cordreps being shared? Vice versa? Benchmarks say that we mostly shouldn't care, the caveat being that atomic ops seem more expensive on upcoming archs (arcadia) so we should error on the side of an extra IsOne() branch saving us single owned atomic ops. PiperOrigin-RevId: 346676121 -- f0babab103b9e60d61ba09482d468985e43eceb3 by Samuel Benzaquen <sbenza@google.com>: Fix iterator based constructor and `.insert` members to only require EmplaceConstructible as the standard specifies. PiperOrigin-RevId: 346616707 -- 8f48eedda02277f9c96a88ed7726e34b557cce20 by Evan Brown <ezb@google.com>: Fix a bug in binary_search_impl when there's a transparent, three-way comparator that has different equivalence classes for different lookup types. Add a new can_have_multiple_equivalent_keys method to share the common logic for these cases. PiperOrigin-RevId: 346605948 -- 649183cb3cc9383431de9c81fb1c0f885d4001ae by Abseil Team <absl-team@google.com>: Add benchmark for accessing a Duration flag. PiperOrigin-RevId: 346594843 -- fefdb046520871af63ce2229e2f7cccfc0483dea by Abseil Team <absl-team@google.com>: Restructure CordReader for upcoming ring buffer changes. PiperOrigin-RevId: 346410642 -- 8b2f50e7da0ebab06ead5f94e366e984ca23cb6a by Abseil Team <absl-team@google.com>: Wire in an internal-only flag to toggle upcoming ring buffer changes on/off for experimentation. PiperOrigin-RevId: 346199111 GitOrigin-RevId: 6e9f93888bbe6718997ed90bbd049f1f3572b066 Change-Id: I8f34866b25a79209cb5448bbb28dd3044111d2e9
4 years ago
node = Concat(CordRep::Ref(lhs_stack.back()), node);
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
lhs_stack.pop_back();
}
return node;
}
void Cord::RemovePrefix(size_t n) {
ABSL_INTERNAL_CHECK(n <= size(),
absl::StrCat("Requested prefix size ", n,
" exceeds Cord's size ", size()));
CordRep* tree = contents_.tree();
if (tree == nullptr) {
contents_.remove_prefix(n);
Export of internal Abseil changes -- c68f1886f5e8fd90eb0c2d2e68feaf00a7cdacda by CJ Johnson <johnsoncj@google.com>: Introduce absl::Cleanup to the OSS repo PiperOrigin-RevId: 354583156 -- 17030cf388e10f7eb959e3e566326d1072ce392e by Abseil Team <absl-team@google.com>: Internal change only PiperOrigin-RevId: 354574953 -- e979d7236d4f3252e79ddda6739b67a9a326bf6d by CJ Johnson <johnsoncj@google.com>: Internal change PiperOrigin-RevId: 354545297 -- 7ea02b3783f7f49ef97d86a8f6580a19cc57df14 by Abseil Team <absl-team@google.com>: Pre-allocate memory for vectors where the size is known. PiperOrigin-RevId: 354344576 -- 9246c7cb11f1d6444f79ebe25acc69a8a9b870e0 by Matt Kulukundis <kfm@google.com>: Add support for Elbrus 2000 (e2k) Import of https://github.com/abseil/abseil-cpp/pull/889 PiperOrigin-RevId: 354344013 -- 0fc93d359cc1fb307552e917b37b7b2e7eed822f by Abseil Team <absl-team@google.com>: Integrate CordRepRing logic into cord (but do not enable it) PiperOrigin-RevId: 354312238 -- eda05622f7da71466723acb33403f783529df24b by Abseil Team <absl-team@google.com>: Protect ignore diagnostic with "__has_warning". PiperOrigin-RevId: 354112334 -- 47716c5d8fb10efa4fdd801d28bac414c6f8ec32 by Abseil Team <absl-team@google.com>: Rearrange InlinedVector copy constructor and destructor to treat a few special cases inline and then tail-call a non-inlined routine for the rest. In particular, we optimize for empty vectors in both cases. Added a couple of benchmarks that copy either an InlVec<int64> or an InlVec<InlVec<int64>>. Speed difference: ``` BM_CopyTrivial/0 0.92ns +- 0% 0.47ns +- 0% -48.91% (p=0.000 n=11+12) BM_CopyTrivial/1 0.92ns +- 0% 1.15ns +- 0% +25.00% (p=0.000 n=10+9) BM_CopyTrivial/8 8.57ns +- 0% 10.72ns +- 1% +25.16% (p=0.000 n=10+12) BM_CopyNonTrivial/0 3.21ns +- 0% 0.70ns +- 0% -78.23% (p=0.000 n=12+10) BM_CopyNonTrivial/1 5.88ns +- 1% 5.51ns +- 0% -6.28% (p=0.000 n=10+8) BM_CopyNonTrivial/8 21.5ns +- 1% 15.2ns +- 2% -29.23% (p=0.000 n=12+12) ``` Note: the slowdowns are a few cycles which is expected given the procedure call added in that case. We decided this is a good tradeoff given the code size reductions and the more significant speedups for empty vectors. Size difference (as measured by nm): ``` BM_CopyTrivial from 1048 bytes to 326 bytes. BM_CopyNonTrivial from 749 bytes to 470 bytes. ``` Code size for a large binary drops by ~500KB (from 349415719 to 348906015 348906191). All of the benchmarks that showed a significant difference: Ones that improve with this CL: ``` BM_CopyNonTrivial/0 3.21ns +- 0% 0.70ns +- 0% -78.23% (p=0.000 n=12+10) BM_InlinedVectorFillString/0 0.93ns +- 0% 0.24ns +- 0% -74.19% (p=0.000 n=12+10) BM_InlinedVectorAssignments/1 10.5ns +- 0% 4.1ns +- 0% -60.64% (p=0.000 n=11+10) BM_InlinedVectorAssignments/2 10.7ns +- 0% 4.4ns +- 0% -59.08% (p=0.000 n=11+11) BM_CopyTrivial/0 0.92ns +- 0% 0.47ns +- 0% -48.91% (p=0.000 n=11+12) BM_CopyNonTrivial/8 21.5ns +- 1% 15.2ns +- 2% -29.23% (p=0.000 n=12+12) BM_StdVectorEmpty 0.47ns +- 1% 0.35ns +- 0% -24.73% (p=0.000 n=12+12) BM_StdVectorSize 0.46ns +- 2% 0.35ns +- 0% -24.32% (p=0.000 n=12+12) BM_SwapElements<LargeCopyableOnly>/0 3.44ns +- 0% 2.76ns +- 1% -19.83% (p=0.000 n=11+11) BM_InlinedVectorFillRange/256 20.7ns +- 1% 17.8ns +- 0% -14.08% (p=0.000 n=12+9) BM_CopyNonTrivial/1 5.88ns +- 1% 5.51ns +- 0% -6.28% (p=0.000 n=10+8) BM_SwapElements<LargeCopyableMovable>/1 4.19ns +- 0% 3.95ns +- 1% -5.63% (p=0.000 n=11+12) BM_SwapElements<LargeCopyableMovableSwappable>/1 4.18ns +- 0% 3.99ns +- 0% -4.70% (p=0.000 n=9+11) BM_SwapElements<LargeCopyableMovable>/0 2.41ns +- 0% 2.31ns +- 0% -4.45% (p=0.000 n=12+12) BM_InlinedVectorFillRange/64 8.25ns +- 0% 8.04ns +- 0% -2.51% (p=0.000 n=12+11) BM_SwapElements<LargeCopyableOnly>/1 82.4ns +- 0% 81.5ns +- 0% -1.06% (p=0.000 n=12+12) ``` Ones that get worse with this CL: ``` BM_CopyTrivial/1 0.92ns +- 0% 1.15ns +- 0% +25.00% (p=0.000 n=10+9) BM_CopyTrivial/8 8.57ns +- 0% 10.72ns +- 1% +25.16% (p=0.000 n=10+12) BM_SwapElements<LargeCopyableMovableSwappable>/512 1.48ns +- 1% 1.66ns +- 1% +11.88% (p=0.000 n=12+12) BM_InlinedVectorFillString/1 11.5ns +- 0% 12.8ns +- 1% +11.62% (p=0.000 n=12+11) BM_SwapElements<LargeCopyableMovableSwappable>/64 1.48ns +- 2% 1.66ns +- 1% +11.66% (p=0.000 n=12+11) BM_SwapElements<LargeCopyableMovableSwappable>/1k 1.48ns +- 1% 1.65ns +- 2% +11.32% (p=0.000 n=12+12) BM_SwapElements<LargeCopyableMovable>/512 1.48ns +- 2% 1.58ns +- 4% +6.62% (p=0.000 n=11+12) BM_SwapElements<LargeCopyableMovable>/1k 1.49ns +- 2% 1.58ns +- 3% +6.05% (p=0.000 n=12+12) BM_SwapElements<LargeCopyableMovable>/64 1.48ns +- 2% 1.57ns +- 4% +6.04% (p=0.000 n=11+12) BM_InlinedVectorFillRange/1 4.81ns +- 0% 5.05ns +- 0% +4.83% (p=0.000 n=11+11) BM_InlinedVectorFillString/8 79.4ns +- 1% 83.1ns +- 1% +4.64% (p=0.000 n=10+12) BM_StdVectorFillString/1 16.3ns +- 0% 16.6ns +- 0% +2.13% (p=0.000 n=11+8) ``` PiperOrigin-RevId: 353906786 -- 8e26518b3cec9c598e5e9573c46c3bd1b03a67ef by Abseil Team <absl-team@google.com>: Internal change PiperOrigin-RevId: 353737330 -- f206ae0983e58c9904ed8b8f05f9caf564a446be by Matt Kulukundis <kfm@google.com>: Import of CCTZ from GitHub. PiperOrigin-RevId: 353682256 GitOrigin-RevId: c68f1886f5e8fd90eb0c2d2e68feaf00a7cdacda Change-Id: I5790c1036c4f543c701d1039848fabf7ae881ad8
4 years ago
} else if (tree->tag == RING) {
contents_.replace_tree(CordRepRing::RemovePrefix(tree->ring(), n));
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
} else {
CordRep* newrep = RemovePrefixFrom(tree, n);
Export of internal Abseil changes -- 6e9f93888bbe6718997ed90bbd049f1f3572b066 by Abseil Team <absl-team@google.com>: Fix status to be safe for self move assignment. PiperOrigin-RevId: 346815392 -- 35cae74a977f258e81dfbe925fb5a34cb6632036 by Gennadiy Rozental <rogeeff@google.com>: Eliminate unnecessary access to the global vars. PiperOrigin-RevId: 346777168 -- e7e786c243069060f5d6c1c945decb4b0b83f95b by Andy Getzendanner <durandal@google.com>: Internal change. PiperOrigin-RevId: 346685656 -- 4ccd41c48f1a83cfa20b3ea534f743dd7d788376 by Abseil Team <absl-team@google.com>: Move CordRep Ref() and Unref() logic into cord_internal.h This change moves Ref() and Unref() logic out of cord.cc into cord_internal. The main purpose is to make upcoming ring buffer changes easier to isolate from existing cord.cc code. Notice that this removes the nullptr check from Unref() and now requires it to be non null (which held true most times). We may need to rethink if the 'unref unlikely one' is the common case: are cordreps most likely shared? This may be something between 'root' and non root nodes, i.e., is it more likely for leaf / flat nodes in large cords to be shared than top level cordreps being shared? Vice versa? Benchmarks say that we mostly shouldn't care, the caveat being that atomic ops seem more expensive on upcoming archs (arcadia) so we should error on the side of an extra IsOne() branch saving us single owned atomic ops. PiperOrigin-RevId: 346676121 -- f0babab103b9e60d61ba09482d468985e43eceb3 by Samuel Benzaquen <sbenza@google.com>: Fix iterator based constructor and `.insert` members to only require EmplaceConstructible as the standard specifies. PiperOrigin-RevId: 346616707 -- 8f48eedda02277f9c96a88ed7726e34b557cce20 by Evan Brown <ezb@google.com>: Fix a bug in binary_search_impl when there's a transparent, three-way comparator that has different equivalence classes for different lookup types. Add a new can_have_multiple_equivalent_keys method to share the common logic for these cases. PiperOrigin-RevId: 346605948 -- 649183cb3cc9383431de9c81fb1c0f885d4001ae by Abseil Team <absl-team@google.com>: Add benchmark for accessing a Duration flag. PiperOrigin-RevId: 346594843 -- fefdb046520871af63ce2229e2f7cccfc0483dea by Abseil Team <absl-team@google.com>: Restructure CordReader for upcoming ring buffer changes. PiperOrigin-RevId: 346410642 -- 8b2f50e7da0ebab06ead5f94e366e984ca23cb6a by Abseil Team <absl-team@google.com>: Wire in an internal-only flag to toggle upcoming ring buffer changes on/off for experimentation. PiperOrigin-RevId: 346199111 GitOrigin-RevId: 6e9f93888bbe6718997ed90bbd049f1f3572b066 Change-Id: I8f34866b25a79209cb5448bbb28dd3044111d2e9
4 years ago
CordRep::Unref(tree);
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
contents_.replace_tree(VerifyTree(newrep));
}
}
void Cord::RemoveSuffix(size_t n) {
ABSL_INTERNAL_CHECK(n <= size(),
absl::StrCat("Requested suffix size ", n,
" exceeds Cord's size ", size()));
CordRep* tree = contents_.tree();
if (tree == nullptr) {
contents_.reduce_size(n);
Export of internal Abseil changes -- c68f1886f5e8fd90eb0c2d2e68feaf00a7cdacda by CJ Johnson <johnsoncj@google.com>: Introduce absl::Cleanup to the OSS repo PiperOrigin-RevId: 354583156 -- 17030cf388e10f7eb959e3e566326d1072ce392e by Abseil Team <absl-team@google.com>: Internal change only PiperOrigin-RevId: 354574953 -- e979d7236d4f3252e79ddda6739b67a9a326bf6d by CJ Johnson <johnsoncj@google.com>: Internal change PiperOrigin-RevId: 354545297 -- 7ea02b3783f7f49ef97d86a8f6580a19cc57df14 by Abseil Team <absl-team@google.com>: Pre-allocate memory for vectors where the size is known. PiperOrigin-RevId: 354344576 -- 9246c7cb11f1d6444f79ebe25acc69a8a9b870e0 by Matt Kulukundis <kfm@google.com>: Add support for Elbrus 2000 (e2k) Import of https://github.com/abseil/abseil-cpp/pull/889 PiperOrigin-RevId: 354344013 -- 0fc93d359cc1fb307552e917b37b7b2e7eed822f by Abseil Team <absl-team@google.com>: Integrate CordRepRing logic into cord (but do not enable it) PiperOrigin-RevId: 354312238 -- eda05622f7da71466723acb33403f783529df24b by Abseil Team <absl-team@google.com>: Protect ignore diagnostic with "__has_warning". PiperOrigin-RevId: 354112334 -- 47716c5d8fb10efa4fdd801d28bac414c6f8ec32 by Abseil Team <absl-team@google.com>: Rearrange InlinedVector copy constructor and destructor to treat a few special cases inline and then tail-call a non-inlined routine for the rest. In particular, we optimize for empty vectors in both cases. Added a couple of benchmarks that copy either an InlVec<int64> or an InlVec<InlVec<int64>>. Speed difference: ``` BM_CopyTrivial/0 0.92ns +- 0% 0.47ns +- 0% -48.91% (p=0.000 n=11+12) BM_CopyTrivial/1 0.92ns +- 0% 1.15ns +- 0% +25.00% (p=0.000 n=10+9) BM_CopyTrivial/8 8.57ns +- 0% 10.72ns +- 1% +25.16% (p=0.000 n=10+12) BM_CopyNonTrivial/0 3.21ns +- 0% 0.70ns +- 0% -78.23% (p=0.000 n=12+10) BM_CopyNonTrivial/1 5.88ns +- 1% 5.51ns +- 0% -6.28% (p=0.000 n=10+8) BM_CopyNonTrivial/8 21.5ns +- 1% 15.2ns +- 2% -29.23% (p=0.000 n=12+12) ``` Note: the slowdowns are a few cycles which is expected given the procedure call added in that case. We decided this is a good tradeoff given the code size reductions and the more significant speedups for empty vectors. Size difference (as measured by nm): ``` BM_CopyTrivial from 1048 bytes to 326 bytes. BM_CopyNonTrivial from 749 bytes to 470 bytes. ``` Code size for a large binary drops by ~500KB (from 349415719 to 348906015 348906191). All of the benchmarks that showed a significant difference: Ones that improve with this CL: ``` BM_CopyNonTrivial/0 3.21ns +- 0% 0.70ns +- 0% -78.23% (p=0.000 n=12+10) BM_InlinedVectorFillString/0 0.93ns +- 0% 0.24ns +- 0% -74.19% (p=0.000 n=12+10) BM_InlinedVectorAssignments/1 10.5ns +- 0% 4.1ns +- 0% -60.64% (p=0.000 n=11+10) BM_InlinedVectorAssignments/2 10.7ns +- 0% 4.4ns +- 0% -59.08% (p=0.000 n=11+11) BM_CopyTrivial/0 0.92ns +- 0% 0.47ns +- 0% -48.91% (p=0.000 n=11+12) BM_CopyNonTrivial/8 21.5ns +- 1% 15.2ns +- 2% -29.23% (p=0.000 n=12+12) BM_StdVectorEmpty 0.47ns +- 1% 0.35ns +- 0% -24.73% (p=0.000 n=12+12) BM_StdVectorSize 0.46ns +- 2% 0.35ns +- 0% -24.32% (p=0.000 n=12+12) BM_SwapElements<LargeCopyableOnly>/0 3.44ns +- 0% 2.76ns +- 1% -19.83% (p=0.000 n=11+11) BM_InlinedVectorFillRange/256 20.7ns +- 1% 17.8ns +- 0% -14.08% (p=0.000 n=12+9) BM_CopyNonTrivial/1 5.88ns +- 1% 5.51ns +- 0% -6.28% (p=0.000 n=10+8) BM_SwapElements<LargeCopyableMovable>/1 4.19ns +- 0% 3.95ns +- 1% -5.63% (p=0.000 n=11+12) BM_SwapElements<LargeCopyableMovableSwappable>/1 4.18ns +- 0% 3.99ns +- 0% -4.70% (p=0.000 n=9+11) BM_SwapElements<LargeCopyableMovable>/0 2.41ns +- 0% 2.31ns +- 0% -4.45% (p=0.000 n=12+12) BM_InlinedVectorFillRange/64 8.25ns +- 0% 8.04ns +- 0% -2.51% (p=0.000 n=12+11) BM_SwapElements<LargeCopyableOnly>/1 82.4ns +- 0% 81.5ns +- 0% -1.06% (p=0.000 n=12+12) ``` Ones that get worse with this CL: ``` BM_CopyTrivial/1 0.92ns +- 0% 1.15ns +- 0% +25.00% (p=0.000 n=10+9) BM_CopyTrivial/8 8.57ns +- 0% 10.72ns +- 1% +25.16% (p=0.000 n=10+12) BM_SwapElements<LargeCopyableMovableSwappable>/512 1.48ns +- 1% 1.66ns +- 1% +11.88% (p=0.000 n=12+12) BM_InlinedVectorFillString/1 11.5ns +- 0% 12.8ns +- 1% +11.62% (p=0.000 n=12+11) BM_SwapElements<LargeCopyableMovableSwappable>/64 1.48ns +- 2% 1.66ns +- 1% +11.66% (p=0.000 n=12+11) BM_SwapElements<LargeCopyableMovableSwappable>/1k 1.48ns +- 1% 1.65ns +- 2% +11.32% (p=0.000 n=12+12) BM_SwapElements<LargeCopyableMovable>/512 1.48ns +- 2% 1.58ns +- 4% +6.62% (p=0.000 n=11+12) BM_SwapElements<LargeCopyableMovable>/1k 1.49ns +- 2% 1.58ns +- 3% +6.05% (p=0.000 n=12+12) BM_SwapElements<LargeCopyableMovable>/64 1.48ns +- 2% 1.57ns +- 4% +6.04% (p=0.000 n=11+12) BM_InlinedVectorFillRange/1 4.81ns +- 0% 5.05ns +- 0% +4.83% (p=0.000 n=11+11) BM_InlinedVectorFillString/8 79.4ns +- 1% 83.1ns +- 1% +4.64% (p=0.000 n=10+12) BM_StdVectorFillString/1 16.3ns +- 0% 16.6ns +- 0% +2.13% (p=0.000 n=11+8) ``` PiperOrigin-RevId: 353906786 -- 8e26518b3cec9c598e5e9573c46c3bd1b03a67ef by Abseil Team <absl-team@google.com>: Internal change PiperOrigin-RevId: 353737330 -- f206ae0983e58c9904ed8b8f05f9caf564a446be by Matt Kulukundis <kfm@google.com>: Import of CCTZ from GitHub. PiperOrigin-RevId: 353682256 GitOrigin-RevId: c68f1886f5e8fd90eb0c2d2e68feaf00a7cdacda Change-Id: I5790c1036c4f543c701d1039848fabf7ae881ad8
4 years ago
} else if (tree->tag == RING) {
contents_.replace_tree(CordRepRing::RemoveSuffix(tree->ring(), n));
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
} else {
CordRep* newrep = RemoveSuffixFrom(tree, n);
Export of internal Abseil changes -- 6e9f93888bbe6718997ed90bbd049f1f3572b066 by Abseil Team <absl-team@google.com>: Fix status to be safe for self move assignment. PiperOrigin-RevId: 346815392 -- 35cae74a977f258e81dfbe925fb5a34cb6632036 by Gennadiy Rozental <rogeeff@google.com>: Eliminate unnecessary access to the global vars. PiperOrigin-RevId: 346777168 -- e7e786c243069060f5d6c1c945decb4b0b83f95b by Andy Getzendanner <durandal@google.com>: Internal change. PiperOrigin-RevId: 346685656 -- 4ccd41c48f1a83cfa20b3ea534f743dd7d788376 by Abseil Team <absl-team@google.com>: Move CordRep Ref() and Unref() logic into cord_internal.h This change moves Ref() and Unref() logic out of cord.cc into cord_internal. The main purpose is to make upcoming ring buffer changes easier to isolate from existing cord.cc code. Notice that this removes the nullptr check from Unref() and now requires it to be non null (which held true most times). We may need to rethink if the 'unref unlikely one' is the common case: are cordreps most likely shared? This may be something between 'root' and non root nodes, i.e., is it more likely for leaf / flat nodes in large cords to be shared than top level cordreps being shared? Vice versa? Benchmarks say that we mostly shouldn't care, the caveat being that atomic ops seem more expensive on upcoming archs (arcadia) so we should error on the side of an extra IsOne() branch saving us single owned atomic ops. PiperOrigin-RevId: 346676121 -- f0babab103b9e60d61ba09482d468985e43eceb3 by Samuel Benzaquen <sbenza@google.com>: Fix iterator based constructor and `.insert` members to only require EmplaceConstructible as the standard specifies. PiperOrigin-RevId: 346616707 -- 8f48eedda02277f9c96a88ed7726e34b557cce20 by Evan Brown <ezb@google.com>: Fix a bug in binary_search_impl when there's a transparent, three-way comparator that has different equivalence classes for different lookup types. Add a new can_have_multiple_equivalent_keys method to share the common logic for these cases. PiperOrigin-RevId: 346605948 -- 649183cb3cc9383431de9c81fb1c0f885d4001ae by Abseil Team <absl-team@google.com>: Add benchmark for accessing a Duration flag. PiperOrigin-RevId: 346594843 -- fefdb046520871af63ce2229e2f7cccfc0483dea by Abseil Team <absl-team@google.com>: Restructure CordReader for upcoming ring buffer changes. PiperOrigin-RevId: 346410642 -- 8b2f50e7da0ebab06ead5f94e366e984ca23cb6a by Abseil Team <absl-team@google.com>: Wire in an internal-only flag to toggle upcoming ring buffer changes on/off for experimentation. PiperOrigin-RevId: 346199111 GitOrigin-RevId: 6e9f93888bbe6718997ed90bbd049f1f3572b066 Change-Id: I8f34866b25a79209cb5448bbb28dd3044111d2e9
4 years ago
CordRep::Unref(tree);
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
contents_.replace_tree(VerifyTree(newrep));
}
}
// Work item for NewSubRange().
struct SubRange {
SubRange(CordRep* a_node, size_t a_pos, size_t a_n)
: node(a_node), pos(a_pos), n(a_n) {}
CordRep* node; // nullptr means concat last 2 results.
size_t pos;
size_t n;
};
static CordRep* NewSubRange(CordRep* node, size_t pos, size_t n) {
absl::InlinedVector<CordRep*, kInlinedVectorSize> results;
absl::InlinedVector<SubRange, kInlinedVectorSize> todo;
todo.push_back(SubRange(node, pos, n));
do {
const SubRange& sr = todo.back();
node = sr.node;
pos = sr.pos;
n = sr.n;
todo.pop_back();
if (node == nullptr) {
assert(results.size() >= 2);
CordRep* right = results.back();
results.pop_back();
CordRep* left = results.back();
results.pop_back();
results.push_back(Concat(left, right));
} else if (pos == 0 && n == node->length) {
Export of internal Abseil changes -- 6e9f93888bbe6718997ed90bbd049f1f3572b066 by Abseil Team <absl-team@google.com>: Fix status to be safe for self move assignment. PiperOrigin-RevId: 346815392 -- 35cae74a977f258e81dfbe925fb5a34cb6632036 by Gennadiy Rozental <rogeeff@google.com>: Eliminate unnecessary access to the global vars. PiperOrigin-RevId: 346777168 -- e7e786c243069060f5d6c1c945decb4b0b83f95b by Andy Getzendanner <durandal@google.com>: Internal change. PiperOrigin-RevId: 346685656 -- 4ccd41c48f1a83cfa20b3ea534f743dd7d788376 by Abseil Team <absl-team@google.com>: Move CordRep Ref() and Unref() logic into cord_internal.h This change moves Ref() and Unref() logic out of cord.cc into cord_internal. The main purpose is to make upcoming ring buffer changes easier to isolate from existing cord.cc code. Notice that this removes the nullptr check from Unref() and now requires it to be non null (which held true most times). We may need to rethink if the 'unref unlikely one' is the common case: are cordreps most likely shared? This may be something between 'root' and non root nodes, i.e., is it more likely for leaf / flat nodes in large cords to be shared than top level cordreps being shared? Vice versa? Benchmarks say that we mostly shouldn't care, the caveat being that atomic ops seem more expensive on upcoming archs (arcadia) so we should error on the side of an extra IsOne() branch saving us single owned atomic ops. PiperOrigin-RevId: 346676121 -- f0babab103b9e60d61ba09482d468985e43eceb3 by Samuel Benzaquen <sbenza@google.com>: Fix iterator based constructor and `.insert` members to only require EmplaceConstructible as the standard specifies. PiperOrigin-RevId: 346616707 -- 8f48eedda02277f9c96a88ed7726e34b557cce20 by Evan Brown <ezb@google.com>: Fix a bug in binary_search_impl when there's a transparent, three-way comparator that has different equivalence classes for different lookup types. Add a new can_have_multiple_equivalent_keys method to share the common logic for these cases. PiperOrigin-RevId: 346605948 -- 649183cb3cc9383431de9c81fb1c0f885d4001ae by Abseil Team <absl-team@google.com>: Add benchmark for accessing a Duration flag. PiperOrigin-RevId: 346594843 -- fefdb046520871af63ce2229e2f7cccfc0483dea by Abseil Team <absl-team@google.com>: Restructure CordReader for upcoming ring buffer changes. PiperOrigin-RevId: 346410642 -- 8b2f50e7da0ebab06ead5f94e366e984ca23cb6a by Abseil Team <absl-team@google.com>: Wire in an internal-only flag to toggle upcoming ring buffer changes on/off for experimentation. PiperOrigin-RevId: 346199111 GitOrigin-RevId: 6e9f93888bbe6718997ed90bbd049f1f3572b066 Change-Id: I8f34866b25a79209cb5448bbb28dd3044111d2e9
4 years ago
results.push_back(CordRep::Ref(node));
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
} else if (node->tag != CONCAT) {
if (node->tag == SUBSTRING) {
pos += node->substring()->start;
node = node->substring()->child;
}
Export of internal Abseil changes -- 6e9f93888bbe6718997ed90bbd049f1f3572b066 by Abseil Team <absl-team@google.com>: Fix status to be safe for self move assignment. PiperOrigin-RevId: 346815392 -- 35cae74a977f258e81dfbe925fb5a34cb6632036 by Gennadiy Rozental <rogeeff@google.com>: Eliminate unnecessary access to the global vars. PiperOrigin-RevId: 346777168 -- e7e786c243069060f5d6c1c945decb4b0b83f95b by Andy Getzendanner <durandal@google.com>: Internal change. PiperOrigin-RevId: 346685656 -- 4ccd41c48f1a83cfa20b3ea534f743dd7d788376 by Abseil Team <absl-team@google.com>: Move CordRep Ref() and Unref() logic into cord_internal.h This change moves Ref() and Unref() logic out of cord.cc into cord_internal. The main purpose is to make upcoming ring buffer changes easier to isolate from existing cord.cc code. Notice that this removes the nullptr check from Unref() and now requires it to be non null (which held true most times). We may need to rethink if the 'unref unlikely one' is the common case: are cordreps most likely shared? This may be something between 'root' and non root nodes, i.e., is it more likely for leaf / flat nodes in large cords to be shared than top level cordreps being shared? Vice versa? Benchmarks say that we mostly shouldn't care, the caveat being that atomic ops seem more expensive on upcoming archs (arcadia) so we should error on the side of an extra IsOne() branch saving us single owned atomic ops. PiperOrigin-RevId: 346676121 -- f0babab103b9e60d61ba09482d468985e43eceb3 by Samuel Benzaquen <sbenza@google.com>: Fix iterator based constructor and `.insert` members to only require EmplaceConstructible as the standard specifies. PiperOrigin-RevId: 346616707 -- 8f48eedda02277f9c96a88ed7726e34b557cce20 by Evan Brown <ezb@google.com>: Fix a bug in binary_search_impl when there's a transparent, three-way comparator that has different equivalence classes for different lookup types. Add a new can_have_multiple_equivalent_keys method to share the common logic for these cases. PiperOrigin-RevId: 346605948 -- 649183cb3cc9383431de9c81fb1c0f885d4001ae by Abseil Team <absl-team@google.com>: Add benchmark for accessing a Duration flag. PiperOrigin-RevId: 346594843 -- fefdb046520871af63ce2229e2f7cccfc0483dea by Abseil Team <absl-team@google.com>: Restructure CordReader for upcoming ring buffer changes. PiperOrigin-RevId: 346410642 -- 8b2f50e7da0ebab06ead5f94e366e984ca23cb6a by Abseil Team <absl-team@google.com>: Wire in an internal-only flag to toggle upcoming ring buffer changes on/off for experimentation. PiperOrigin-RevId: 346199111 GitOrigin-RevId: 6e9f93888bbe6718997ed90bbd049f1f3572b066 Change-Id: I8f34866b25a79209cb5448bbb28dd3044111d2e9
4 years ago
results.push_back(NewSubstring(CordRep::Ref(node), pos, n));
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
} else if (pos + n <= node->concat()->left->length) {
todo.push_back(SubRange(node->concat()->left, pos, n));
} else if (pos >= node->concat()->left->length) {
pos -= node->concat()->left->length;
todo.push_back(SubRange(node->concat()->right, pos, n));
} else {
size_t left_n = node->concat()->left->length - pos;
todo.push_back(SubRange(nullptr, 0, 0)); // Concat()
todo.push_back(SubRange(node->concat()->right, 0, n - left_n));
todo.push_back(SubRange(node->concat()->left, pos, left_n));
}
} while (!todo.empty());
assert(results.size() == 1);
return results[0];
}
Cord Cord::Subcord(size_t pos, size_t new_size) const {
Cord sub_cord;
size_t length = size();
if (pos > length) pos = length;
if (new_size > length - pos) new_size = length - pos;
CordRep* tree = contents_.tree();
if (tree == nullptr) {
// sub_cord is newly constructed, no need to re-zero-out the tail of
// contents_ memory.
sub_cord.contents_.set_data(contents_.data() + pos, new_size, false);
} else if (new_size == 0) {
// We want to return empty subcord, so nothing to do.
} else if (new_size <= InlineRep::kMaxInline) {
Cord::ChunkIterator it = chunk_begin();
it.AdvanceBytes(pos);
Export of internal Abseil changes -- 642ab296a2c9629c44f3f2ce6911cd2488bcf416 by Derek Mauro <dmauro@google.com>: Remove an obsolete check in CMakeLists.txt PiperOrigin-RevId: 352852564 -- ce78cb96bcfd162737dbcf35005da3d1d6a3486b by Abseil Team <absl-team@google.com>: Clarify that the calling *thread* must have locked the mutex in order to unlock it. PiperOrigin-RevId: 352801804 -- 24e1f5f72756046f5265abf618e951c341f09b8d by Derek Mauro <dmauro@google.com>: Fixes failing CMake string comparisons https://cmake.org/cmake/help/latest/policy/CMP0054.html Fixes #791 PiperOrigin-RevId: 352791054 -- 0ac10bc3f4dca2c4c4b51d7b8196a2eaee9537a1 by Abseil Team <absl-team@google.com>: Introduce CordRepRing class This change introduces the CordRepRing class that implements all the lower level / internal implementation for upcoming CordRepRing ring buffer support in cord. PiperOrigin-RevId: 352771994 -- 4bd36dda61760785844f0f29f26d90cc18046f75 by Abseil Team <absl-team@google.com>: Optimize InlineData representation for cord sampling (cordz) This CL changes InlineData to allow us to store a (future) Cordz Info pointer directly into the inline representation: - make InlineData a class that provides a public API to set the active union members (tree or chars) and safely access that data. - change 'tree' and 'profiled' bits to be the 2 least significant bits, allowing us 62 continquous bits for storing a Cordz Info pointer. PiperOrigin-RevId: 352642411 -- dc55ba71bbce0e6a83e05a453990c51ac3d68426 by Mark Barolak <mbar@google.com>: Add unit test coverage for the mutating overload of absl::AsciiStrToLower. PiperOrigin-RevId: 352626006 GitOrigin-RevId: 642ab296a2c9629c44f3f2ce6911cd2488bcf416 Change-Id: I6c5929dd830d3c630e14e7fd5387fc3e25a69100
4 years ago
char* dest = sub_cord.contents_.data_.as_chars();
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
size_t remaining_size = new_size;
while (remaining_size > it->size()) {
cord_internal::SmallMemmove(dest, it->data(), it->size());
remaining_size -= it->size();
dest += it->size();
++it;
}
cord_internal::SmallMemmove(dest, it->data(), remaining_size);
Export of internal Abseil changes -- 642ab296a2c9629c44f3f2ce6911cd2488bcf416 by Derek Mauro <dmauro@google.com>: Remove an obsolete check in CMakeLists.txt PiperOrigin-RevId: 352852564 -- ce78cb96bcfd162737dbcf35005da3d1d6a3486b by Abseil Team <absl-team@google.com>: Clarify that the calling *thread* must have locked the mutex in order to unlock it. PiperOrigin-RevId: 352801804 -- 24e1f5f72756046f5265abf618e951c341f09b8d by Derek Mauro <dmauro@google.com>: Fixes failing CMake string comparisons https://cmake.org/cmake/help/latest/policy/CMP0054.html Fixes #791 PiperOrigin-RevId: 352791054 -- 0ac10bc3f4dca2c4c4b51d7b8196a2eaee9537a1 by Abseil Team <absl-team@google.com>: Introduce CordRepRing class This change introduces the CordRepRing class that implements all the lower level / internal implementation for upcoming CordRepRing ring buffer support in cord. PiperOrigin-RevId: 352771994 -- 4bd36dda61760785844f0f29f26d90cc18046f75 by Abseil Team <absl-team@google.com>: Optimize InlineData representation for cord sampling (cordz) This CL changes InlineData to allow us to store a (future) Cordz Info pointer directly into the inline representation: - make InlineData a class that provides a public API to set the active union members (tree or chars) and safely access that data. - change 'tree' and 'profiled' bits to be the 2 least significant bits, allowing us 62 continquous bits for storing a Cordz Info pointer. PiperOrigin-RevId: 352642411 -- dc55ba71bbce0e6a83e05a453990c51ac3d68426 by Mark Barolak <mbar@google.com>: Add unit test coverage for the mutating overload of absl::AsciiStrToLower. PiperOrigin-RevId: 352626006 GitOrigin-RevId: 642ab296a2c9629c44f3f2ce6911cd2488bcf416 Change-Id: I6c5929dd830d3c630e14e7fd5387fc3e25a69100
4 years ago
sub_cord.contents_.set_inline_size(new_size);
Export of internal Abseil changes -- c68f1886f5e8fd90eb0c2d2e68feaf00a7cdacda by CJ Johnson <johnsoncj@google.com>: Introduce absl::Cleanup to the OSS repo PiperOrigin-RevId: 354583156 -- 17030cf388e10f7eb959e3e566326d1072ce392e by Abseil Team <absl-team@google.com>: Internal change only PiperOrigin-RevId: 354574953 -- e979d7236d4f3252e79ddda6739b67a9a326bf6d by CJ Johnson <johnsoncj@google.com>: Internal change PiperOrigin-RevId: 354545297 -- 7ea02b3783f7f49ef97d86a8f6580a19cc57df14 by Abseil Team <absl-team@google.com>: Pre-allocate memory for vectors where the size is known. PiperOrigin-RevId: 354344576 -- 9246c7cb11f1d6444f79ebe25acc69a8a9b870e0 by Matt Kulukundis <kfm@google.com>: Add support for Elbrus 2000 (e2k) Import of https://github.com/abseil/abseil-cpp/pull/889 PiperOrigin-RevId: 354344013 -- 0fc93d359cc1fb307552e917b37b7b2e7eed822f by Abseil Team <absl-team@google.com>: Integrate CordRepRing logic into cord (but do not enable it) PiperOrigin-RevId: 354312238 -- eda05622f7da71466723acb33403f783529df24b by Abseil Team <absl-team@google.com>: Protect ignore diagnostic with "__has_warning". PiperOrigin-RevId: 354112334 -- 47716c5d8fb10efa4fdd801d28bac414c6f8ec32 by Abseil Team <absl-team@google.com>: Rearrange InlinedVector copy constructor and destructor to treat a few special cases inline and then tail-call a non-inlined routine for the rest. In particular, we optimize for empty vectors in both cases. Added a couple of benchmarks that copy either an InlVec<int64> or an InlVec<InlVec<int64>>. Speed difference: ``` BM_CopyTrivial/0 0.92ns +- 0% 0.47ns +- 0% -48.91% (p=0.000 n=11+12) BM_CopyTrivial/1 0.92ns +- 0% 1.15ns +- 0% +25.00% (p=0.000 n=10+9) BM_CopyTrivial/8 8.57ns +- 0% 10.72ns +- 1% +25.16% (p=0.000 n=10+12) BM_CopyNonTrivial/0 3.21ns +- 0% 0.70ns +- 0% -78.23% (p=0.000 n=12+10) BM_CopyNonTrivial/1 5.88ns +- 1% 5.51ns +- 0% -6.28% (p=0.000 n=10+8) BM_CopyNonTrivial/8 21.5ns +- 1% 15.2ns +- 2% -29.23% (p=0.000 n=12+12) ``` Note: the slowdowns are a few cycles which is expected given the procedure call added in that case. We decided this is a good tradeoff given the code size reductions and the more significant speedups for empty vectors. Size difference (as measured by nm): ``` BM_CopyTrivial from 1048 bytes to 326 bytes. BM_CopyNonTrivial from 749 bytes to 470 bytes. ``` Code size for a large binary drops by ~500KB (from 349415719 to 348906015 348906191). All of the benchmarks that showed a significant difference: Ones that improve with this CL: ``` BM_CopyNonTrivial/0 3.21ns +- 0% 0.70ns +- 0% -78.23% (p=0.000 n=12+10) BM_InlinedVectorFillString/0 0.93ns +- 0% 0.24ns +- 0% -74.19% (p=0.000 n=12+10) BM_InlinedVectorAssignments/1 10.5ns +- 0% 4.1ns +- 0% -60.64% (p=0.000 n=11+10) BM_InlinedVectorAssignments/2 10.7ns +- 0% 4.4ns +- 0% -59.08% (p=0.000 n=11+11) BM_CopyTrivial/0 0.92ns +- 0% 0.47ns +- 0% -48.91% (p=0.000 n=11+12) BM_CopyNonTrivial/8 21.5ns +- 1% 15.2ns +- 2% -29.23% (p=0.000 n=12+12) BM_StdVectorEmpty 0.47ns +- 1% 0.35ns +- 0% -24.73% (p=0.000 n=12+12) BM_StdVectorSize 0.46ns +- 2% 0.35ns +- 0% -24.32% (p=0.000 n=12+12) BM_SwapElements<LargeCopyableOnly>/0 3.44ns +- 0% 2.76ns +- 1% -19.83% (p=0.000 n=11+11) BM_InlinedVectorFillRange/256 20.7ns +- 1% 17.8ns +- 0% -14.08% (p=0.000 n=12+9) BM_CopyNonTrivial/1 5.88ns +- 1% 5.51ns +- 0% -6.28% (p=0.000 n=10+8) BM_SwapElements<LargeCopyableMovable>/1 4.19ns +- 0% 3.95ns +- 1% -5.63% (p=0.000 n=11+12) BM_SwapElements<LargeCopyableMovableSwappable>/1 4.18ns +- 0% 3.99ns +- 0% -4.70% (p=0.000 n=9+11) BM_SwapElements<LargeCopyableMovable>/0 2.41ns +- 0% 2.31ns +- 0% -4.45% (p=0.000 n=12+12) BM_InlinedVectorFillRange/64 8.25ns +- 0% 8.04ns +- 0% -2.51% (p=0.000 n=12+11) BM_SwapElements<LargeCopyableOnly>/1 82.4ns +- 0% 81.5ns +- 0% -1.06% (p=0.000 n=12+12) ``` Ones that get worse with this CL: ``` BM_CopyTrivial/1 0.92ns +- 0% 1.15ns +- 0% +25.00% (p=0.000 n=10+9) BM_CopyTrivial/8 8.57ns +- 0% 10.72ns +- 1% +25.16% (p=0.000 n=10+12) BM_SwapElements<LargeCopyableMovableSwappable>/512 1.48ns +- 1% 1.66ns +- 1% +11.88% (p=0.000 n=12+12) BM_InlinedVectorFillString/1 11.5ns +- 0% 12.8ns +- 1% +11.62% (p=0.000 n=12+11) BM_SwapElements<LargeCopyableMovableSwappable>/64 1.48ns +- 2% 1.66ns +- 1% +11.66% (p=0.000 n=12+11) BM_SwapElements<LargeCopyableMovableSwappable>/1k 1.48ns +- 1% 1.65ns +- 2% +11.32% (p=0.000 n=12+12) BM_SwapElements<LargeCopyableMovable>/512 1.48ns +- 2% 1.58ns +- 4% +6.62% (p=0.000 n=11+12) BM_SwapElements<LargeCopyableMovable>/1k 1.49ns +- 2% 1.58ns +- 3% +6.05% (p=0.000 n=12+12) BM_SwapElements<LargeCopyableMovable>/64 1.48ns +- 2% 1.57ns +- 4% +6.04% (p=0.000 n=11+12) BM_InlinedVectorFillRange/1 4.81ns +- 0% 5.05ns +- 0% +4.83% (p=0.000 n=11+11) BM_InlinedVectorFillString/8 79.4ns +- 1% 83.1ns +- 1% +4.64% (p=0.000 n=10+12) BM_StdVectorFillString/1 16.3ns +- 0% 16.6ns +- 0% +2.13% (p=0.000 n=11+8) ``` PiperOrigin-RevId: 353906786 -- 8e26518b3cec9c598e5e9573c46c3bd1b03a67ef by Abseil Team <absl-team@google.com>: Internal change PiperOrigin-RevId: 353737330 -- f206ae0983e58c9904ed8b8f05f9caf564a446be by Matt Kulukundis <kfm@google.com>: Import of CCTZ from GitHub. PiperOrigin-RevId: 353682256 GitOrigin-RevId: c68f1886f5e8fd90eb0c2d2e68feaf00a7cdacda Change-Id: I5790c1036c4f543c701d1039848fabf7ae881ad8
4 years ago
} else if (tree->tag == RING) {
tree = CordRepRing::SubRing(CordRep::Ref(tree)->ring(), pos, new_size);
sub_cord.contents_.set_tree(tree);
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
} else {
sub_cord.contents_.set_tree(NewSubRange(tree, pos, new_size));
}
return sub_cord;
}
// --------------------------------------------------------------------
// Balancing
class CordForest {
public:
explicit CordForest(size_t length)
: root_length_(length), trees_(kMinLengthSize, nullptr) {}
void Build(CordRep* cord_root) {
std::vector<CordRep*> pending = {cord_root};
while (!pending.empty()) {
CordRep* node = pending.back();
pending.pop_back();
CheckNode(node);
if (ABSL_PREDICT_FALSE(node->tag != CONCAT)) {
AddNode(node);
continue;
}
CordRepConcat* concat_node = node->concat();
if (concat_node->depth() >= kMinLengthSize ||
concat_node->length < min_length[concat_node->depth()]) {
pending.push_back(concat_node->right);
pending.push_back(concat_node->left);
if (concat_node->refcount.IsOne()) {
concat_node->left = concat_freelist_;
concat_freelist_ = concat_node;
} else {
Export of internal Abseil changes -- 6e9f93888bbe6718997ed90bbd049f1f3572b066 by Abseil Team <absl-team@google.com>: Fix status to be safe for self move assignment. PiperOrigin-RevId: 346815392 -- 35cae74a977f258e81dfbe925fb5a34cb6632036 by Gennadiy Rozental <rogeeff@google.com>: Eliminate unnecessary access to the global vars. PiperOrigin-RevId: 346777168 -- e7e786c243069060f5d6c1c945decb4b0b83f95b by Andy Getzendanner <durandal@google.com>: Internal change. PiperOrigin-RevId: 346685656 -- 4ccd41c48f1a83cfa20b3ea534f743dd7d788376 by Abseil Team <absl-team@google.com>: Move CordRep Ref() and Unref() logic into cord_internal.h This change moves Ref() and Unref() logic out of cord.cc into cord_internal. The main purpose is to make upcoming ring buffer changes easier to isolate from existing cord.cc code. Notice that this removes the nullptr check from Unref() and now requires it to be non null (which held true most times). We may need to rethink if the 'unref unlikely one' is the common case: are cordreps most likely shared? This may be something between 'root' and non root nodes, i.e., is it more likely for leaf / flat nodes in large cords to be shared than top level cordreps being shared? Vice versa? Benchmarks say that we mostly shouldn't care, the caveat being that atomic ops seem more expensive on upcoming archs (arcadia) so we should error on the side of an extra IsOne() branch saving us single owned atomic ops. PiperOrigin-RevId: 346676121 -- f0babab103b9e60d61ba09482d468985e43eceb3 by Samuel Benzaquen <sbenza@google.com>: Fix iterator based constructor and `.insert` members to only require EmplaceConstructible as the standard specifies. PiperOrigin-RevId: 346616707 -- 8f48eedda02277f9c96a88ed7726e34b557cce20 by Evan Brown <ezb@google.com>: Fix a bug in binary_search_impl when there's a transparent, three-way comparator that has different equivalence classes for different lookup types. Add a new can_have_multiple_equivalent_keys method to share the common logic for these cases. PiperOrigin-RevId: 346605948 -- 649183cb3cc9383431de9c81fb1c0f885d4001ae by Abseil Team <absl-team@google.com>: Add benchmark for accessing a Duration flag. PiperOrigin-RevId: 346594843 -- fefdb046520871af63ce2229e2f7cccfc0483dea by Abseil Team <absl-team@google.com>: Restructure CordReader for upcoming ring buffer changes. PiperOrigin-RevId: 346410642 -- 8b2f50e7da0ebab06ead5f94e366e984ca23cb6a by Abseil Team <absl-team@google.com>: Wire in an internal-only flag to toggle upcoming ring buffer changes on/off for experimentation. PiperOrigin-RevId: 346199111 GitOrigin-RevId: 6e9f93888bbe6718997ed90bbd049f1f3572b066 Change-Id: I8f34866b25a79209cb5448bbb28dd3044111d2e9
4 years ago
CordRep::Ref(concat_node->right);
CordRep::Ref(concat_node->left);
CordRep::Unref(concat_node);
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
}
} else {
AddNode(node);
}
}
}
CordRep* ConcatNodes() {
CordRep* sum = nullptr;
for (auto* node : trees_) {
if (node == nullptr) continue;
sum = PrependNode(node, sum);
root_length_ -= node->length;
if (root_length_ == 0) break;
}
ABSL_INTERNAL_CHECK(sum != nullptr, "Failed to locate sum node");
return VerifyTree(sum);
}
private:
CordRep* AppendNode(CordRep* node, CordRep* sum) {
return (sum == nullptr) ? node : MakeConcat(sum, node);
}
CordRep* PrependNode(CordRep* node, CordRep* sum) {
return (sum == nullptr) ? node : MakeConcat(node, sum);
}
void AddNode(CordRep* node) {
CordRep* sum = nullptr;
// Collect together everything with which we will merge with node
int i = 0;
for (; node->length > min_length[i + 1]; ++i) {
auto& tree_at_i = trees_[i];
if (tree_at_i == nullptr) continue;
sum = PrependNode(tree_at_i, sum);
tree_at_i = nullptr;
}
sum = AppendNode(node, sum);
// Insert sum into appropriate place in the forest
for (; sum->length >= min_length[i]; ++i) {
auto& tree_at_i = trees_[i];
if (tree_at_i == nullptr) continue;
sum = MakeConcat(tree_at_i, sum);
tree_at_i = nullptr;
}
// min_length[0] == 1, which means sum->length >= min_length[0]
assert(i > 0);
trees_[i - 1] = sum;
}
// Make concat node trying to resue existing CordRepConcat nodes we
// already collected in the concat_freelist_.
CordRep* MakeConcat(CordRep* left, CordRep* right) {
if (concat_freelist_ == nullptr) return RawConcat(left, right);
CordRepConcat* rep = concat_freelist_;
if (concat_freelist_->left == nullptr) {
concat_freelist_ = nullptr;
} else {
concat_freelist_ = concat_freelist_->left->concat();
}
SetConcatChildren(rep, left, right);
return rep;
}
static void CheckNode(CordRep* node) {
ABSL_INTERNAL_CHECK(node->length != 0u, "");
if (node->tag == CONCAT) {
ABSL_INTERNAL_CHECK(node->concat()->left != nullptr, "");
ABSL_INTERNAL_CHECK(node->concat()->right != nullptr, "");
ABSL_INTERNAL_CHECK(node->length == (node->concat()->left->length +
node->concat()->right->length),
"");
}
}
size_t root_length_;
// use an inlined vector instead of a flat array to get bounds checking
absl::InlinedVector<CordRep*, kInlinedVectorSize> trees_;
// List of concat nodes we can re-use for Cord balancing.
CordRepConcat* concat_freelist_ = nullptr;
};
static CordRep* Rebalance(CordRep* node) {
VerifyTree(node);
assert(node->tag == CONCAT);
if (node->length == 0) {
return nullptr;
}
CordForest forest(node->length);
forest.Build(node);
return forest.ConcatNodes();
}
// --------------------------------------------------------------------
// Comparators
namespace {
int ClampResult(int memcmp_res) {
return static_cast<int>(memcmp_res > 0) - static_cast<int>(memcmp_res < 0);
}
int CompareChunks(absl::string_view* lhs, absl::string_view* rhs,
size_t* size_to_compare) {
size_t compared_size = std::min(lhs->size(), rhs->size());
assert(*size_to_compare >= compared_size);
*size_to_compare -= compared_size;
int memcmp_res = ::memcmp(lhs->data(), rhs->data(), compared_size);
if (memcmp_res != 0) return memcmp_res;
lhs->remove_prefix(compared_size);
rhs->remove_prefix(compared_size);
return 0;
}
// This overload set computes comparison results from memcmp result. This
// interface is used inside GenericCompare below. Differet implementations
// are specialized for int and bool. For int we clamp result to {-1, 0, 1}
// set. For bool we just interested in "value == 0".
template <typename ResultType>
ResultType ComputeCompareResult(int memcmp_res) {
return ClampResult(memcmp_res);
}
template <>
bool ComputeCompareResult<bool>(int memcmp_res) {
return memcmp_res == 0;
}
} // namespace
// Helper routine. Locates the first flat chunk of the Cord without
// initializing the iterator.
inline absl::string_view Cord::InlineRep::FindFlatStartPiece() const {
Export of internal Abseil changes -- 642ab296a2c9629c44f3f2ce6911cd2488bcf416 by Derek Mauro <dmauro@google.com>: Remove an obsolete check in CMakeLists.txt PiperOrigin-RevId: 352852564 -- ce78cb96bcfd162737dbcf35005da3d1d6a3486b by Abseil Team <absl-team@google.com>: Clarify that the calling *thread* must have locked the mutex in order to unlock it. PiperOrigin-RevId: 352801804 -- 24e1f5f72756046f5265abf618e951c341f09b8d by Derek Mauro <dmauro@google.com>: Fixes failing CMake string comparisons https://cmake.org/cmake/help/latest/policy/CMP0054.html Fixes #791 PiperOrigin-RevId: 352791054 -- 0ac10bc3f4dca2c4c4b51d7b8196a2eaee9537a1 by Abseil Team <absl-team@google.com>: Introduce CordRepRing class This change introduces the CordRepRing class that implements all the lower level / internal implementation for upcoming CordRepRing ring buffer support in cord. PiperOrigin-RevId: 352771994 -- 4bd36dda61760785844f0f29f26d90cc18046f75 by Abseil Team <absl-team@google.com>: Optimize InlineData representation for cord sampling (cordz) This CL changes InlineData to allow us to store a (future) Cordz Info pointer directly into the inline representation: - make InlineData a class that provides a public API to set the active union members (tree or chars) and safely access that data. - change 'tree' and 'profiled' bits to be the 2 least significant bits, allowing us 62 continquous bits for storing a Cordz Info pointer. PiperOrigin-RevId: 352642411 -- dc55ba71bbce0e6a83e05a453990c51ac3d68426 by Mark Barolak <mbar@google.com>: Add unit test coverage for the mutating overload of absl::AsciiStrToLower. PiperOrigin-RevId: 352626006 GitOrigin-RevId: 642ab296a2c9629c44f3f2ce6911cd2488bcf416 Change-Id: I6c5929dd830d3c630e14e7fd5387fc3e25a69100
4 years ago
if (!is_tree()) {
return absl::string_view(data_.as_chars(), data_.inline_size());
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
}
CordRep* node = tree();
if (node->tag >= FLAT) {
Export of internal Abseil changes -- a0491c8d790972cd80e2d720fe1fdf5f711a6f1a by Greg Falcon <gfalcon@google.com>: Stop directly accessing CordRepFlat data via CordRep::data. The old pattern of access breaks the `CordRep` type abstraction; since `CordRep::data` is not in general guaranteed to contain the chunk's data, we shouldn't access it that way. This incidentally adds an assertion check (via the flat() accessor) that the CordRep is indeed flat on each such access, but a manual inspection of the code, as well as the fact that this code currently works, suggest that this is always true.) PiperOrigin-RevId: 351592344 -- f40c3b43ca5b1d7e23cd45f1ffac1783105ac1a3 by Abseil Team <absl-team@google.com>: Revert 18abb2902b9f06c63a968b24d3dda785ebf99a22 PiperOrigin-RevId: 351523518 -- 18abb2902b9f06c63a968b24d3dda785ebf99a22 by Abseil Team <absl-team@google.com>: Internal change PiperOrigin-RevId: 351512412 -- 9b881602d45e95e06089792c7627cd56528a255a by Abseil Team <absl-team@google.com>: Keep time's global state in a cacheline-aligned structure. Keeping the global state as separate global variables results in two issues: 1) False sharing with adjacent global data (e.g., cycle clock source), since the global fields are updated every O(10usec). 2) The hot global fields (e.g., seq and samples) can reside on different cache lines. To fix this, simply wrap the global data in a ABSL_CACHE_ALIGNED structure. This is similar to what we do for MutexGlobals. PiperOrigin-RevId: 351389466 GitOrigin-RevId: a0491c8d790972cd80e2d720fe1fdf5f711a6f1a Change-Id: Ie0fa80112043381cd37c84e2ab2b7334839f54b5
4 years ago
return absl::string_view(node->flat()->Data(), node->length);
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
}
if (node->tag == EXTERNAL) {
return absl::string_view(node->external()->base, node->length);
}
Export of internal Abseil changes -- c68f1886f5e8fd90eb0c2d2e68feaf00a7cdacda by CJ Johnson <johnsoncj@google.com>: Introduce absl::Cleanup to the OSS repo PiperOrigin-RevId: 354583156 -- 17030cf388e10f7eb959e3e566326d1072ce392e by Abseil Team <absl-team@google.com>: Internal change only PiperOrigin-RevId: 354574953 -- e979d7236d4f3252e79ddda6739b67a9a326bf6d by CJ Johnson <johnsoncj@google.com>: Internal change PiperOrigin-RevId: 354545297 -- 7ea02b3783f7f49ef97d86a8f6580a19cc57df14 by Abseil Team <absl-team@google.com>: Pre-allocate memory for vectors where the size is known. PiperOrigin-RevId: 354344576 -- 9246c7cb11f1d6444f79ebe25acc69a8a9b870e0 by Matt Kulukundis <kfm@google.com>: Add support for Elbrus 2000 (e2k) Import of https://github.com/abseil/abseil-cpp/pull/889 PiperOrigin-RevId: 354344013 -- 0fc93d359cc1fb307552e917b37b7b2e7eed822f by Abseil Team <absl-team@google.com>: Integrate CordRepRing logic into cord (but do not enable it) PiperOrigin-RevId: 354312238 -- eda05622f7da71466723acb33403f783529df24b by Abseil Team <absl-team@google.com>: Protect ignore diagnostic with "__has_warning". PiperOrigin-RevId: 354112334 -- 47716c5d8fb10efa4fdd801d28bac414c6f8ec32 by Abseil Team <absl-team@google.com>: Rearrange InlinedVector copy constructor and destructor to treat a few special cases inline and then tail-call a non-inlined routine for the rest. In particular, we optimize for empty vectors in both cases. Added a couple of benchmarks that copy either an InlVec<int64> or an InlVec<InlVec<int64>>. Speed difference: ``` BM_CopyTrivial/0 0.92ns +- 0% 0.47ns +- 0% -48.91% (p=0.000 n=11+12) BM_CopyTrivial/1 0.92ns +- 0% 1.15ns +- 0% +25.00% (p=0.000 n=10+9) BM_CopyTrivial/8 8.57ns +- 0% 10.72ns +- 1% +25.16% (p=0.000 n=10+12) BM_CopyNonTrivial/0 3.21ns +- 0% 0.70ns +- 0% -78.23% (p=0.000 n=12+10) BM_CopyNonTrivial/1 5.88ns +- 1% 5.51ns +- 0% -6.28% (p=0.000 n=10+8) BM_CopyNonTrivial/8 21.5ns +- 1% 15.2ns +- 2% -29.23% (p=0.000 n=12+12) ``` Note: the slowdowns are a few cycles which is expected given the procedure call added in that case. We decided this is a good tradeoff given the code size reductions and the more significant speedups for empty vectors. Size difference (as measured by nm): ``` BM_CopyTrivial from 1048 bytes to 326 bytes. BM_CopyNonTrivial from 749 bytes to 470 bytes. ``` Code size for a large binary drops by ~500KB (from 349415719 to 348906015 348906191). All of the benchmarks that showed a significant difference: Ones that improve with this CL: ``` BM_CopyNonTrivial/0 3.21ns +- 0% 0.70ns +- 0% -78.23% (p=0.000 n=12+10) BM_InlinedVectorFillString/0 0.93ns +- 0% 0.24ns +- 0% -74.19% (p=0.000 n=12+10) BM_InlinedVectorAssignments/1 10.5ns +- 0% 4.1ns +- 0% -60.64% (p=0.000 n=11+10) BM_InlinedVectorAssignments/2 10.7ns +- 0% 4.4ns +- 0% -59.08% (p=0.000 n=11+11) BM_CopyTrivial/0 0.92ns +- 0% 0.47ns +- 0% -48.91% (p=0.000 n=11+12) BM_CopyNonTrivial/8 21.5ns +- 1% 15.2ns +- 2% -29.23% (p=0.000 n=12+12) BM_StdVectorEmpty 0.47ns +- 1% 0.35ns +- 0% -24.73% (p=0.000 n=12+12) BM_StdVectorSize 0.46ns +- 2% 0.35ns +- 0% -24.32% (p=0.000 n=12+12) BM_SwapElements<LargeCopyableOnly>/0 3.44ns +- 0% 2.76ns +- 1% -19.83% (p=0.000 n=11+11) BM_InlinedVectorFillRange/256 20.7ns +- 1% 17.8ns +- 0% -14.08% (p=0.000 n=12+9) BM_CopyNonTrivial/1 5.88ns +- 1% 5.51ns +- 0% -6.28% (p=0.000 n=10+8) BM_SwapElements<LargeCopyableMovable>/1 4.19ns +- 0% 3.95ns +- 1% -5.63% (p=0.000 n=11+12) BM_SwapElements<LargeCopyableMovableSwappable>/1 4.18ns +- 0% 3.99ns +- 0% -4.70% (p=0.000 n=9+11) BM_SwapElements<LargeCopyableMovable>/0 2.41ns +- 0% 2.31ns +- 0% -4.45% (p=0.000 n=12+12) BM_InlinedVectorFillRange/64 8.25ns +- 0% 8.04ns +- 0% -2.51% (p=0.000 n=12+11) BM_SwapElements<LargeCopyableOnly>/1 82.4ns +- 0% 81.5ns +- 0% -1.06% (p=0.000 n=12+12) ``` Ones that get worse with this CL: ``` BM_CopyTrivial/1 0.92ns +- 0% 1.15ns +- 0% +25.00% (p=0.000 n=10+9) BM_CopyTrivial/8 8.57ns +- 0% 10.72ns +- 1% +25.16% (p=0.000 n=10+12) BM_SwapElements<LargeCopyableMovableSwappable>/512 1.48ns +- 1% 1.66ns +- 1% +11.88% (p=0.000 n=12+12) BM_InlinedVectorFillString/1 11.5ns +- 0% 12.8ns +- 1% +11.62% (p=0.000 n=12+11) BM_SwapElements<LargeCopyableMovableSwappable>/64 1.48ns +- 2% 1.66ns +- 1% +11.66% (p=0.000 n=12+11) BM_SwapElements<LargeCopyableMovableSwappable>/1k 1.48ns +- 1% 1.65ns +- 2% +11.32% (p=0.000 n=12+12) BM_SwapElements<LargeCopyableMovable>/512 1.48ns +- 2% 1.58ns +- 4% +6.62% (p=0.000 n=11+12) BM_SwapElements<LargeCopyableMovable>/1k 1.49ns +- 2% 1.58ns +- 3% +6.05% (p=0.000 n=12+12) BM_SwapElements<LargeCopyableMovable>/64 1.48ns +- 2% 1.57ns +- 4% +6.04% (p=0.000 n=11+12) BM_InlinedVectorFillRange/1 4.81ns +- 0% 5.05ns +- 0% +4.83% (p=0.000 n=11+11) BM_InlinedVectorFillString/8 79.4ns +- 1% 83.1ns +- 1% +4.64% (p=0.000 n=10+12) BM_StdVectorFillString/1 16.3ns +- 0% 16.6ns +- 0% +2.13% (p=0.000 n=11+8) ``` PiperOrigin-RevId: 353906786 -- 8e26518b3cec9c598e5e9573c46c3bd1b03a67ef by Abseil Team <absl-team@google.com>: Internal change PiperOrigin-RevId: 353737330 -- f206ae0983e58c9904ed8b8f05f9caf564a446be by Matt Kulukundis <kfm@google.com>: Import of CCTZ from GitHub. PiperOrigin-RevId: 353682256 GitOrigin-RevId: c68f1886f5e8fd90eb0c2d2e68feaf00a7cdacda Change-Id: I5790c1036c4f543c701d1039848fabf7ae881ad8
4 years ago
if (node->tag == RING) {
return node->ring()->entry_data(node->ring()->head());
}
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
// Walk down the left branches until we hit a non-CONCAT node.
while (node->tag == CONCAT) {
node = node->concat()->left;
}
// Get the child node if we encounter a SUBSTRING.
size_t offset = 0;
size_t length = node->length;
assert(length != 0);
if (node->tag == SUBSTRING) {
offset = node->substring()->start;
node = node->substring()->child;
}
if (node->tag >= FLAT) {
Export of internal Abseil changes -- a0491c8d790972cd80e2d720fe1fdf5f711a6f1a by Greg Falcon <gfalcon@google.com>: Stop directly accessing CordRepFlat data via CordRep::data. The old pattern of access breaks the `CordRep` type abstraction; since `CordRep::data` is not in general guaranteed to contain the chunk's data, we shouldn't access it that way. This incidentally adds an assertion check (via the flat() accessor) that the CordRep is indeed flat on each such access, but a manual inspection of the code, as well as the fact that this code currently works, suggest that this is always true.) PiperOrigin-RevId: 351592344 -- f40c3b43ca5b1d7e23cd45f1ffac1783105ac1a3 by Abseil Team <absl-team@google.com>: Revert 18abb2902b9f06c63a968b24d3dda785ebf99a22 PiperOrigin-RevId: 351523518 -- 18abb2902b9f06c63a968b24d3dda785ebf99a22 by Abseil Team <absl-team@google.com>: Internal change PiperOrigin-RevId: 351512412 -- 9b881602d45e95e06089792c7627cd56528a255a by Abseil Team <absl-team@google.com>: Keep time's global state in a cacheline-aligned structure. Keeping the global state as separate global variables results in two issues: 1) False sharing with adjacent global data (e.g., cycle clock source), since the global fields are updated every O(10usec). 2) The hot global fields (e.g., seq and samples) can reside on different cache lines. To fix this, simply wrap the global data in a ABSL_CACHE_ALIGNED structure. This is similar to what we do for MutexGlobals. PiperOrigin-RevId: 351389466 GitOrigin-RevId: a0491c8d790972cd80e2d720fe1fdf5f711a6f1a Change-Id: Ie0fa80112043381cd37c84e2ab2b7334839f54b5
4 years ago
return absl::string_view(node->flat()->Data() + offset, length);
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
}
assert((node->tag == EXTERNAL) && "Expect FLAT or EXTERNAL node here");
return absl::string_view(node->external()->base + offset, length);
}
inline int Cord::CompareSlowPath(absl::string_view rhs, size_t compared_size,
size_t size_to_compare) const {
auto advance = [](Cord::ChunkIterator* it, absl::string_view* chunk) {
if (!chunk->empty()) return true;
++*it;
if (it->bytes_remaining_ == 0) return false;
*chunk = **it;
return true;
};
Cord::ChunkIterator lhs_it = chunk_begin();
// compared_size is inside first chunk.
absl::string_view lhs_chunk =
(lhs_it.bytes_remaining_ != 0) ? *lhs_it : absl::string_view();
assert(compared_size <= lhs_chunk.size());
assert(compared_size <= rhs.size());
lhs_chunk.remove_prefix(compared_size);
rhs.remove_prefix(compared_size);
size_to_compare -= compared_size; // skip already compared size.
while (advance(&lhs_it, &lhs_chunk) && !rhs.empty()) {
int comparison_result = CompareChunks(&lhs_chunk, &rhs, &size_to_compare);
if (comparison_result != 0) return comparison_result;
if (size_to_compare == 0) return 0;
}
return static_cast<int>(rhs.empty()) - static_cast<int>(lhs_chunk.empty());
}
inline int Cord::CompareSlowPath(const Cord& rhs, size_t compared_size,
size_t size_to_compare) const {
auto advance = [](Cord::ChunkIterator* it, absl::string_view* chunk) {
if (!chunk->empty()) return true;
++*it;
if (it->bytes_remaining_ == 0) return false;
*chunk = **it;
return true;
};
Cord::ChunkIterator lhs_it = chunk_begin();
Cord::ChunkIterator rhs_it = rhs.chunk_begin();
// compared_size is inside both first chunks.
absl::string_view lhs_chunk =
(lhs_it.bytes_remaining_ != 0) ? *lhs_it : absl::string_view();
absl::string_view rhs_chunk =
(rhs_it.bytes_remaining_ != 0) ? *rhs_it : absl::string_view();
assert(compared_size <= lhs_chunk.size());
assert(compared_size <= rhs_chunk.size());
lhs_chunk.remove_prefix(compared_size);
rhs_chunk.remove_prefix(compared_size);
size_to_compare -= compared_size; // skip already compared size.
while (advance(&lhs_it, &lhs_chunk) && advance(&rhs_it, &rhs_chunk)) {
int memcmp_res = CompareChunks(&lhs_chunk, &rhs_chunk, &size_to_compare);
if (memcmp_res != 0) return memcmp_res;
if (size_to_compare == 0) return 0;
}
return static_cast<int>(rhs_chunk.empty()) -
static_cast<int>(lhs_chunk.empty());
}
inline absl::string_view Cord::GetFirstChunk(const Cord& c) {
return c.contents_.FindFlatStartPiece();
}
inline absl::string_view Cord::GetFirstChunk(absl::string_view sv) {
return sv;
}
// Compares up to 'size_to_compare' bytes of 'lhs' with 'rhs'. It is assumed
// that 'size_to_compare' is greater that size of smallest of first chunks.
template <typename ResultType, typename RHS>
ResultType GenericCompare(const Cord& lhs, const RHS& rhs,
size_t size_to_compare) {
absl::string_view lhs_chunk = Cord::GetFirstChunk(lhs);
absl::string_view rhs_chunk = Cord::GetFirstChunk(rhs);
size_t compared_size = std::min(lhs_chunk.size(), rhs_chunk.size());
assert(size_to_compare >= compared_size);
int memcmp_res = ::memcmp(lhs_chunk.data(), rhs_chunk.data(), compared_size);
if (compared_size == size_to_compare || memcmp_res != 0) {
return ComputeCompareResult<ResultType>(memcmp_res);
}
return ComputeCompareResult<ResultType>(
lhs.CompareSlowPath(rhs, compared_size, size_to_compare));
}
bool Cord::EqualsImpl(absl::string_view rhs, size_t size_to_compare) const {
return GenericCompare<bool>(*this, rhs, size_to_compare);
}
bool Cord::EqualsImpl(const Cord& rhs, size_t size_to_compare) const {
return GenericCompare<bool>(*this, rhs, size_to_compare);
}
template <typename RHS>
inline int SharedCompareImpl(const Cord& lhs, const RHS& rhs) {
size_t lhs_size = lhs.size();
size_t rhs_size = rhs.size();
if (lhs_size == rhs_size) {
return GenericCompare<int>(lhs, rhs, lhs_size);
}
if (lhs_size < rhs_size) {
auto data_comp_res = GenericCompare<int>(lhs, rhs, lhs_size);
return data_comp_res == 0 ? -1 : data_comp_res;
}
auto data_comp_res = GenericCompare<int>(lhs, rhs, rhs_size);
return data_comp_res == 0 ? +1 : data_comp_res;
}
int Cord::Compare(absl::string_view rhs) const {
return SharedCompareImpl(*this, rhs);
}
int Cord::CompareImpl(const Cord& rhs) const {
return SharedCompareImpl(*this, rhs);
}
bool Cord::EndsWith(absl::string_view rhs) const {
size_t my_size = size();
size_t rhs_size = rhs.size();
if (my_size < rhs_size) return false;
Cord tmp(*this);
tmp.RemovePrefix(my_size - rhs_size);
return tmp.EqualsImpl(rhs, rhs_size);
}
bool Cord::EndsWith(const Cord& rhs) const {
size_t my_size = size();
size_t rhs_size = rhs.size();
if (my_size < rhs_size) return false;
Cord tmp(*this);
tmp.RemovePrefix(my_size - rhs_size);
return tmp.EqualsImpl(rhs, rhs_size);
}
// --------------------------------------------------------------------
// Misc.
Cord::operator std::string() const {
std::string s;
absl::CopyCordToString(*this, &s);
return s;
}
void CopyCordToString(const Cord& src, std::string* dst) {
if (!src.contents_.is_tree()) {
src.contents_.CopyTo(dst);
} else {
absl::strings_internal::STLStringResizeUninitialized(dst, src.size());
src.CopyToArraySlowPath(&(*dst)[0]);
}
}
void Cord::CopyToArraySlowPath(char* dst) const {
assert(contents_.is_tree());
absl::string_view fragment;
if (GetFlatAux(contents_.tree(), &fragment)) {
memcpy(dst, fragment.data(), fragment.size());
return;
}
for (absl::string_view chunk : Chunks()) {
memcpy(dst, chunk.data(), chunk.size());
dst += chunk.size();
}
}
Export of internal Abseil changes -- 0bfa836596a9c787a2f0bdc283011dd1f6810c6e by Benjamin Barenblat <bbaren@google.com>: Ignore missing CPU frequency on more architectures Linux on MIPS, PA-RISC, RISC-V, and SystemZ doesn’t expose the nominal CPU frequency via /sys, so don’t worry if `NominalCPUFrequency` returns 1.0 on those platforms. Some POWER machines expose the CPU frequency; others do not. Since we can’t predict which type of machine the tests will run on, simply disable testing for `NominalCPUFrequency` on POWER. PiperOrigin-RevId: 347079873 -- 492b6834ed4a07cbc3abccd846f7e37d8c556ee5 by Benjamin Barenblat <bbaren@google.com>: Use ABSL_HAVE_THREAD_LOCAL macro instead of copying code Reduce code duplication by checking the ABSL_HAVE_THREAD_LOCAL macro instead of copying code from base/config.h. PiperOrigin-RevId: 347079561 -- 8d656efce4da9cb032094377e58493d98427a536 by Abseil Team <absl-team@google.com>: Rollback PiperOrigin-RevId: 347078779 -- 221bc69ec6dd7e2777ffcff6942584f979ef6382 by Abseil Team <absl-team@google.com>: Add flag for 'shallow subcord' feature for experimental ring buffer rollout There is a potential trade-off of CPU cost vs over-sharing cord data for subcord of large cords. This flag allows making subcords shallow for ringbuffers (with a potential larger waste of referenced source cords), which allows us to make subcord fast for this apps that do no persist (unmodified / plain copied) sub cords. This change also introduces constants for the default settings, intended to keep the internal cord settings concistent with external flags. PiperOrigin-RevId: 347053271 -- 00a56c24293566734009f6bf2169a83fb37a35ba by Abseil Team <absl-team@google.com>: Revert the usage of variant<> in Cord iterator and reader. The introduction of the variant may lead to some missed compiler optimizations. PiperOrigin-RevId: 347053041 -- c7b7b5ed7e3ab46b1e75b80f1a7de0bda26c8f70 by Chris Kennelly <ckennelly@google.com>: Release library for integer power-of-2 functions and bit counting. PiperOrigin-RevId: 347035065 -- 5a035c0d9840b251967f9e7039fc6a4e01dd52f3 by Abseil Team <absl-team@google.com>: Restructure Cord::ChunkIterator for future ring buffer support. PiperOrigin-RevId: 346890054 GitOrigin-RevId: 0bfa836596a9c787a2f0bdc283011dd1f6810c6e Change-Id: I3a58e2a44cb4c6f2116c43e2a4ccbc319d3ccecf
4 years ago
Cord::ChunkIterator& Cord::ChunkIterator::AdvanceStack() {
auto& stack_of_right_children = stack_of_right_children_;
Export of internal Abseil changes -- 0bfa836596a9c787a2f0bdc283011dd1f6810c6e by Benjamin Barenblat <bbaren@google.com>: Ignore missing CPU frequency on more architectures Linux on MIPS, PA-RISC, RISC-V, and SystemZ doesn’t expose the nominal CPU frequency via /sys, so don’t worry if `NominalCPUFrequency` returns 1.0 on those platforms. Some POWER machines expose the CPU frequency; others do not. Since we can’t predict which type of machine the tests will run on, simply disable testing for `NominalCPUFrequency` on POWER. PiperOrigin-RevId: 347079873 -- 492b6834ed4a07cbc3abccd846f7e37d8c556ee5 by Benjamin Barenblat <bbaren@google.com>: Use ABSL_HAVE_THREAD_LOCAL macro instead of copying code Reduce code duplication by checking the ABSL_HAVE_THREAD_LOCAL macro instead of copying code from base/config.h. PiperOrigin-RevId: 347079561 -- 8d656efce4da9cb032094377e58493d98427a536 by Abseil Team <absl-team@google.com>: Rollback PiperOrigin-RevId: 347078779 -- 221bc69ec6dd7e2777ffcff6942584f979ef6382 by Abseil Team <absl-team@google.com>: Add flag for 'shallow subcord' feature for experimental ring buffer rollout There is a potential trade-off of CPU cost vs over-sharing cord data for subcord of large cords. This flag allows making subcords shallow for ringbuffers (with a potential larger waste of referenced source cords), which allows us to make subcord fast for this apps that do no persist (unmodified / plain copied) sub cords. This change also introduces constants for the default settings, intended to keep the internal cord settings concistent with external flags. PiperOrigin-RevId: 347053271 -- 00a56c24293566734009f6bf2169a83fb37a35ba by Abseil Team <absl-team@google.com>: Revert the usage of variant<> in Cord iterator and reader. The introduction of the variant may lead to some missed compiler optimizations. PiperOrigin-RevId: 347053041 -- c7b7b5ed7e3ab46b1e75b80f1a7de0bda26c8f70 by Chris Kennelly <ckennelly@google.com>: Release library for integer power-of-2 functions and bit counting. PiperOrigin-RevId: 347035065 -- 5a035c0d9840b251967f9e7039fc6a4e01dd52f3 by Abseil Team <absl-team@google.com>: Restructure Cord::ChunkIterator for future ring buffer support. PiperOrigin-RevId: 346890054 GitOrigin-RevId: 0bfa836596a9c787a2f0bdc283011dd1f6810c6e Change-Id: I3a58e2a44cb4c6f2116c43e2a4ccbc319d3ccecf
4 years ago
if (stack_of_right_children.empty()) {
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
assert(!current_chunk_.empty()); // Called on invalid iterator.
// We have reached the end of the Cord.
return *this;
}
// Process the next node on the stack.
Export of internal Abseil changes -- 0bfa836596a9c787a2f0bdc283011dd1f6810c6e by Benjamin Barenblat <bbaren@google.com>: Ignore missing CPU frequency on more architectures Linux on MIPS, PA-RISC, RISC-V, and SystemZ doesn’t expose the nominal CPU frequency via /sys, so don’t worry if `NominalCPUFrequency` returns 1.0 on those platforms. Some POWER machines expose the CPU frequency; others do not. Since we can’t predict which type of machine the tests will run on, simply disable testing for `NominalCPUFrequency` on POWER. PiperOrigin-RevId: 347079873 -- 492b6834ed4a07cbc3abccd846f7e37d8c556ee5 by Benjamin Barenblat <bbaren@google.com>: Use ABSL_HAVE_THREAD_LOCAL macro instead of copying code Reduce code duplication by checking the ABSL_HAVE_THREAD_LOCAL macro instead of copying code from base/config.h. PiperOrigin-RevId: 347079561 -- 8d656efce4da9cb032094377e58493d98427a536 by Abseil Team <absl-team@google.com>: Rollback PiperOrigin-RevId: 347078779 -- 221bc69ec6dd7e2777ffcff6942584f979ef6382 by Abseil Team <absl-team@google.com>: Add flag for 'shallow subcord' feature for experimental ring buffer rollout There is a potential trade-off of CPU cost vs over-sharing cord data for subcord of large cords. This flag allows making subcords shallow for ringbuffers (with a potential larger waste of referenced source cords), which allows us to make subcord fast for this apps that do no persist (unmodified / plain copied) sub cords. This change also introduces constants for the default settings, intended to keep the internal cord settings concistent with external flags. PiperOrigin-RevId: 347053271 -- 00a56c24293566734009f6bf2169a83fb37a35ba by Abseil Team <absl-team@google.com>: Revert the usage of variant<> in Cord iterator and reader. The introduction of the variant may lead to some missed compiler optimizations. PiperOrigin-RevId: 347053041 -- c7b7b5ed7e3ab46b1e75b80f1a7de0bda26c8f70 by Chris Kennelly <ckennelly@google.com>: Release library for integer power-of-2 functions and bit counting. PiperOrigin-RevId: 347035065 -- 5a035c0d9840b251967f9e7039fc6a4e01dd52f3 by Abseil Team <absl-team@google.com>: Restructure Cord::ChunkIterator for future ring buffer support. PiperOrigin-RevId: 346890054 GitOrigin-RevId: 0bfa836596a9c787a2f0bdc283011dd1f6810c6e Change-Id: I3a58e2a44cb4c6f2116c43e2a4ccbc319d3ccecf
4 years ago
CordRep* node = stack_of_right_children.back();
stack_of_right_children.pop_back();
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
// Walk down the left branches until we hit a non-CONCAT node. Save the
// right children to the stack for subsequent traversal.
while (node->tag == CONCAT) {
Export of internal Abseil changes -- 0bfa836596a9c787a2f0bdc283011dd1f6810c6e by Benjamin Barenblat <bbaren@google.com>: Ignore missing CPU frequency on more architectures Linux on MIPS, PA-RISC, RISC-V, and SystemZ doesn’t expose the nominal CPU frequency via /sys, so don’t worry if `NominalCPUFrequency` returns 1.0 on those platforms. Some POWER machines expose the CPU frequency; others do not. Since we can’t predict which type of machine the tests will run on, simply disable testing for `NominalCPUFrequency` on POWER. PiperOrigin-RevId: 347079873 -- 492b6834ed4a07cbc3abccd846f7e37d8c556ee5 by Benjamin Barenblat <bbaren@google.com>: Use ABSL_HAVE_THREAD_LOCAL macro instead of copying code Reduce code duplication by checking the ABSL_HAVE_THREAD_LOCAL macro instead of copying code from base/config.h. PiperOrigin-RevId: 347079561 -- 8d656efce4da9cb032094377e58493d98427a536 by Abseil Team <absl-team@google.com>: Rollback PiperOrigin-RevId: 347078779 -- 221bc69ec6dd7e2777ffcff6942584f979ef6382 by Abseil Team <absl-team@google.com>: Add flag for 'shallow subcord' feature for experimental ring buffer rollout There is a potential trade-off of CPU cost vs over-sharing cord data for subcord of large cords. This flag allows making subcords shallow for ringbuffers (with a potential larger waste of referenced source cords), which allows us to make subcord fast for this apps that do no persist (unmodified / plain copied) sub cords. This change also introduces constants for the default settings, intended to keep the internal cord settings concistent with external flags. PiperOrigin-RevId: 347053271 -- 00a56c24293566734009f6bf2169a83fb37a35ba by Abseil Team <absl-team@google.com>: Revert the usage of variant<> in Cord iterator and reader. The introduction of the variant may lead to some missed compiler optimizations. PiperOrigin-RevId: 347053041 -- c7b7b5ed7e3ab46b1e75b80f1a7de0bda26c8f70 by Chris Kennelly <ckennelly@google.com>: Release library for integer power-of-2 functions and bit counting. PiperOrigin-RevId: 347035065 -- 5a035c0d9840b251967f9e7039fc6a4e01dd52f3 by Abseil Team <absl-team@google.com>: Restructure Cord::ChunkIterator for future ring buffer support. PiperOrigin-RevId: 346890054 GitOrigin-RevId: 0bfa836596a9c787a2f0bdc283011dd1f6810c6e Change-Id: I3a58e2a44cb4c6f2116c43e2a4ccbc319d3ccecf
4 years ago
stack_of_right_children.push_back(node->concat()->right);
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
node = node->concat()->left;
}
// Get the child node if we encounter a SUBSTRING.
size_t offset = 0;
size_t length = node->length;
if (node->tag == SUBSTRING) {
offset = node->substring()->start;
node = node->substring()->child;
}
assert(node->tag == EXTERNAL || node->tag >= FLAT);
assert(length != 0);
const char* data =
Export of internal Abseil changes -- a0491c8d790972cd80e2d720fe1fdf5f711a6f1a by Greg Falcon <gfalcon@google.com>: Stop directly accessing CordRepFlat data via CordRep::data. The old pattern of access breaks the `CordRep` type abstraction; since `CordRep::data` is not in general guaranteed to contain the chunk's data, we shouldn't access it that way. This incidentally adds an assertion check (via the flat() accessor) that the CordRep is indeed flat on each such access, but a manual inspection of the code, as well as the fact that this code currently works, suggest that this is always true.) PiperOrigin-RevId: 351592344 -- f40c3b43ca5b1d7e23cd45f1ffac1783105ac1a3 by Abseil Team <absl-team@google.com>: Revert 18abb2902b9f06c63a968b24d3dda785ebf99a22 PiperOrigin-RevId: 351523518 -- 18abb2902b9f06c63a968b24d3dda785ebf99a22 by Abseil Team <absl-team@google.com>: Internal change PiperOrigin-RevId: 351512412 -- 9b881602d45e95e06089792c7627cd56528a255a by Abseil Team <absl-team@google.com>: Keep time's global state in a cacheline-aligned structure. Keeping the global state as separate global variables results in two issues: 1) False sharing with adjacent global data (e.g., cycle clock source), since the global fields are updated every O(10usec). 2) The hot global fields (e.g., seq and samples) can reside on different cache lines. To fix this, simply wrap the global data in a ABSL_CACHE_ALIGNED structure. This is similar to what we do for MutexGlobals. PiperOrigin-RevId: 351389466 GitOrigin-RevId: a0491c8d790972cd80e2d720fe1fdf5f711a6f1a Change-Id: Ie0fa80112043381cd37c84e2ab2b7334839f54b5
4 years ago
node->tag == EXTERNAL ? node->external()->base : node->flat()->Data();
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
current_chunk_ = absl::string_view(data + offset, length);
current_leaf_ = node;
return *this;
}
Cord Cord::ChunkIterator::AdvanceAndReadBytes(size_t n) {
ABSL_HARDENING_ASSERT(bytes_remaining_ >= n &&
"Attempted to iterate past `end()`");
Cord subcord;
if (n <= InlineRep::kMaxInline) {
// Range to read fits in inline data. Flatten it.
char* data = subcord.contents_.set_data(n);
while (n > current_chunk_.size()) {
memcpy(data, current_chunk_.data(), current_chunk_.size());
data += current_chunk_.size();
n -= current_chunk_.size();
++*this;
}
memcpy(data, current_chunk_.data(), n);
if (n < current_chunk_.size()) {
RemoveChunkPrefix(n);
} else if (n > 0) {
++*this;
}
return subcord;
}
Export of internal Abseil changes -- c68f1886f5e8fd90eb0c2d2e68feaf00a7cdacda by CJ Johnson <johnsoncj@google.com>: Introduce absl::Cleanup to the OSS repo PiperOrigin-RevId: 354583156 -- 17030cf388e10f7eb959e3e566326d1072ce392e by Abseil Team <absl-team@google.com>: Internal change only PiperOrigin-RevId: 354574953 -- e979d7236d4f3252e79ddda6739b67a9a326bf6d by CJ Johnson <johnsoncj@google.com>: Internal change PiperOrigin-RevId: 354545297 -- 7ea02b3783f7f49ef97d86a8f6580a19cc57df14 by Abseil Team <absl-team@google.com>: Pre-allocate memory for vectors where the size is known. PiperOrigin-RevId: 354344576 -- 9246c7cb11f1d6444f79ebe25acc69a8a9b870e0 by Matt Kulukundis <kfm@google.com>: Add support for Elbrus 2000 (e2k) Import of https://github.com/abseil/abseil-cpp/pull/889 PiperOrigin-RevId: 354344013 -- 0fc93d359cc1fb307552e917b37b7b2e7eed822f by Abseil Team <absl-team@google.com>: Integrate CordRepRing logic into cord (but do not enable it) PiperOrigin-RevId: 354312238 -- eda05622f7da71466723acb33403f783529df24b by Abseil Team <absl-team@google.com>: Protect ignore diagnostic with "__has_warning". PiperOrigin-RevId: 354112334 -- 47716c5d8fb10efa4fdd801d28bac414c6f8ec32 by Abseil Team <absl-team@google.com>: Rearrange InlinedVector copy constructor and destructor to treat a few special cases inline and then tail-call a non-inlined routine for the rest. In particular, we optimize for empty vectors in both cases. Added a couple of benchmarks that copy either an InlVec<int64> or an InlVec<InlVec<int64>>. Speed difference: ``` BM_CopyTrivial/0 0.92ns +- 0% 0.47ns +- 0% -48.91% (p=0.000 n=11+12) BM_CopyTrivial/1 0.92ns +- 0% 1.15ns +- 0% +25.00% (p=0.000 n=10+9) BM_CopyTrivial/8 8.57ns +- 0% 10.72ns +- 1% +25.16% (p=0.000 n=10+12) BM_CopyNonTrivial/0 3.21ns +- 0% 0.70ns +- 0% -78.23% (p=0.000 n=12+10) BM_CopyNonTrivial/1 5.88ns +- 1% 5.51ns +- 0% -6.28% (p=0.000 n=10+8) BM_CopyNonTrivial/8 21.5ns +- 1% 15.2ns +- 2% -29.23% (p=0.000 n=12+12) ``` Note: the slowdowns are a few cycles which is expected given the procedure call added in that case. We decided this is a good tradeoff given the code size reductions and the more significant speedups for empty vectors. Size difference (as measured by nm): ``` BM_CopyTrivial from 1048 bytes to 326 bytes. BM_CopyNonTrivial from 749 bytes to 470 bytes. ``` Code size for a large binary drops by ~500KB (from 349415719 to 348906015 348906191). All of the benchmarks that showed a significant difference: Ones that improve with this CL: ``` BM_CopyNonTrivial/0 3.21ns +- 0% 0.70ns +- 0% -78.23% (p=0.000 n=12+10) BM_InlinedVectorFillString/0 0.93ns +- 0% 0.24ns +- 0% -74.19% (p=0.000 n=12+10) BM_InlinedVectorAssignments/1 10.5ns +- 0% 4.1ns +- 0% -60.64% (p=0.000 n=11+10) BM_InlinedVectorAssignments/2 10.7ns +- 0% 4.4ns +- 0% -59.08% (p=0.000 n=11+11) BM_CopyTrivial/0 0.92ns +- 0% 0.47ns +- 0% -48.91% (p=0.000 n=11+12) BM_CopyNonTrivial/8 21.5ns +- 1% 15.2ns +- 2% -29.23% (p=0.000 n=12+12) BM_StdVectorEmpty 0.47ns +- 1% 0.35ns +- 0% -24.73% (p=0.000 n=12+12) BM_StdVectorSize 0.46ns +- 2% 0.35ns +- 0% -24.32% (p=0.000 n=12+12) BM_SwapElements<LargeCopyableOnly>/0 3.44ns +- 0% 2.76ns +- 1% -19.83% (p=0.000 n=11+11) BM_InlinedVectorFillRange/256 20.7ns +- 1% 17.8ns +- 0% -14.08% (p=0.000 n=12+9) BM_CopyNonTrivial/1 5.88ns +- 1% 5.51ns +- 0% -6.28% (p=0.000 n=10+8) BM_SwapElements<LargeCopyableMovable>/1 4.19ns +- 0% 3.95ns +- 1% -5.63% (p=0.000 n=11+12) BM_SwapElements<LargeCopyableMovableSwappable>/1 4.18ns +- 0% 3.99ns +- 0% -4.70% (p=0.000 n=9+11) BM_SwapElements<LargeCopyableMovable>/0 2.41ns +- 0% 2.31ns +- 0% -4.45% (p=0.000 n=12+12) BM_InlinedVectorFillRange/64 8.25ns +- 0% 8.04ns +- 0% -2.51% (p=0.000 n=12+11) BM_SwapElements<LargeCopyableOnly>/1 82.4ns +- 0% 81.5ns +- 0% -1.06% (p=0.000 n=12+12) ``` Ones that get worse with this CL: ``` BM_CopyTrivial/1 0.92ns +- 0% 1.15ns +- 0% +25.00% (p=0.000 n=10+9) BM_CopyTrivial/8 8.57ns +- 0% 10.72ns +- 1% +25.16% (p=0.000 n=10+12) BM_SwapElements<LargeCopyableMovableSwappable>/512 1.48ns +- 1% 1.66ns +- 1% +11.88% (p=0.000 n=12+12) BM_InlinedVectorFillString/1 11.5ns +- 0% 12.8ns +- 1% +11.62% (p=0.000 n=12+11) BM_SwapElements<LargeCopyableMovableSwappable>/64 1.48ns +- 2% 1.66ns +- 1% +11.66% (p=0.000 n=12+11) BM_SwapElements<LargeCopyableMovableSwappable>/1k 1.48ns +- 1% 1.65ns +- 2% +11.32% (p=0.000 n=12+12) BM_SwapElements<LargeCopyableMovable>/512 1.48ns +- 2% 1.58ns +- 4% +6.62% (p=0.000 n=11+12) BM_SwapElements<LargeCopyableMovable>/1k 1.49ns +- 2% 1.58ns +- 3% +6.05% (p=0.000 n=12+12) BM_SwapElements<LargeCopyableMovable>/64 1.48ns +- 2% 1.57ns +- 4% +6.04% (p=0.000 n=11+12) BM_InlinedVectorFillRange/1 4.81ns +- 0% 5.05ns +- 0% +4.83% (p=0.000 n=11+11) BM_InlinedVectorFillString/8 79.4ns +- 1% 83.1ns +- 1% +4.64% (p=0.000 n=10+12) BM_StdVectorFillString/1 16.3ns +- 0% 16.6ns +- 0% +2.13% (p=0.000 n=11+8) ``` PiperOrigin-RevId: 353906786 -- 8e26518b3cec9c598e5e9573c46c3bd1b03a67ef by Abseil Team <absl-team@google.com>: Internal change PiperOrigin-RevId: 353737330 -- f206ae0983e58c9904ed8b8f05f9caf564a446be by Matt Kulukundis <kfm@google.com>: Import of CCTZ from GitHub. PiperOrigin-RevId: 353682256 GitOrigin-RevId: c68f1886f5e8fd90eb0c2d2e68feaf00a7cdacda Change-Id: I5790c1036c4f543c701d1039848fabf7ae881ad8
4 years ago
if (ring_reader_) {
size_t chunk_size = current_chunk_.size();
if (n <= chunk_size && n <= kMaxBytesToCopy) {
subcord = Cord(current_chunk_.substr(0, n));
} else {
auto* ring = CordRep::Ref(ring_reader_.ring())->ring();
size_t offset = ring_reader_.length() - bytes_remaining_;
subcord.contents_.set_tree(CordRepRing::SubRing(ring, offset, n));
}
if (n < chunk_size) {
bytes_remaining_ -= n;
current_chunk_.remove_prefix(n);
} else {
AdvanceBytesRing(n);
}
return subcord;
}
auto& stack_of_right_children = stack_of_right_children_;
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
if (n < current_chunk_.size()) {
// Range to read is a proper subrange of the current chunk.
assert(current_leaf_ != nullptr);
Export of internal Abseil changes -- 6e9f93888bbe6718997ed90bbd049f1f3572b066 by Abseil Team <absl-team@google.com>: Fix status to be safe for self move assignment. PiperOrigin-RevId: 346815392 -- 35cae74a977f258e81dfbe925fb5a34cb6632036 by Gennadiy Rozental <rogeeff@google.com>: Eliminate unnecessary access to the global vars. PiperOrigin-RevId: 346777168 -- e7e786c243069060f5d6c1c945decb4b0b83f95b by Andy Getzendanner <durandal@google.com>: Internal change. PiperOrigin-RevId: 346685656 -- 4ccd41c48f1a83cfa20b3ea534f743dd7d788376 by Abseil Team <absl-team@google.com>: Move CordRep Ref() and Unref() logic into cord_internal.h This change moves Ref() and Unref() logic out of cord.cc into cord_internal. The main purpose is to make upcoming ring buffer changes easier to isolate from existing cord.cc code. Notice that this removes the nullptr check from Unref() and now requires it to be non null (which held true most times). We may need to rethink if the 'unref unlikely one' is the common case: are cordreps most likely shared? This may be something between 'root' and non root nodes, i.e., is it more likely for leaf / flat nodes in large cords to be shared than top level cordreps being shared? Vice versa? Benchmarks say that we mostly shouldn't care, the caveat being that atomic ops seem more expensive on upcoming archs (arcadia) so we should error on the side of an extra IsOne() branch saving us single owned atomic ops. PiperOrigin-RevId: 346676121 -- f0babab103b9e60d61ba09482d468985e43eceb3 by Samuel Benzaquen <sbenza@google.com>: Fix iterator based constructor and `.insert` members to only require EmplaceConstructible as the standard specifies. PiperOrigin-RevId: 346616707 -- 8f48eedda02277f9c96a88ed7726e34b557cce20 by Evan Brown <ezb@google.com>: Fix a bug in binary_search_impl when there's a transparent, three-way comparator that has different equivalence classes for different lookup types. Add a new can_have_multiple_equivalent_keys method to share the common logic for these cases. PiperOrigin-RevId: 346605948 -- 649183cb3cc9383431de9c81fb1c0f885d4001ae by Abseil Team <absl-team@google.com>: Add benchmark for accessing a Duration flag. PiperOrigin-RevId: 346594843 -- fefdb046520871af63ce2229e2f7cccfc0483dea by Abseil Team <absl-team@google.com>: Restructure CordReader for upcoming ring buffer changes. PiperOrigin-RevId: 346410642 -- 8b2f50e7da0ebab06ead5f94e366e984ca23cb6a by Abseil Team <absl-team@google.com>: Wire in an internal-only flag to toggle upcoming ring buffer changes on/off for experimentation. PiperOrigin-RevId: 346199111 GitOrigin-RevId: 6e9f93888bbe6718997ed90bbd049f1f3572b066 Change-Id: I8f34866b25a79209cb5448bbb28dd3044111d2e9
4 years ago
CordRep* subnode = CordRep::Ref(current_leaf_);
Export of internal Abseil changes -- a0491c8d790972cd80e2d720fe1fdf5f711a6f1a by Greg Falcon <gfalcon@google.com>: Stop directly accessing CordRepFlat data via CordRep::data. The old pattern of access breaks the `CordRep` type abstraction; since `CordRep::data` is not in general guaranteed to contain the chunk's data, we shouldn't access it that way. This incidentally adds an assertion check (via the flat() accessor) that the CordRep is indeed flat on each such access, but a manual inspection of the code, as well as the fact that this code currently works, suggest that this is always true.) PiperOrigin-RevId: 351592344 -- f40c3b43ca5b1d7e23cd45f1ffac1783105ac1a3 by Abseil Team <absl-team@google.com>: Revert 18abb2902b9f06c63a968b24d3dda785ebf99a22 PiperOrigin-RevId: 351523518 -- 18abb2902b9f06c63a968b24d3dda785ebf99a22 by Abseil Team <absl-team@google.com>: Internal change PiperOrigin-RevId: 351512412 -- 9b881602d45e95e06089792c7627cd56528a255a by Abseil Team <absl-team@google.com>: Keep time's global state in a cacheline-aligned structure. Keeping the global state as separate global variables results in two issues: 1) False sharing with adjacent global data (e.g., cycle clock source), since the global fields are updated every O(10usec). 2) The hot global fields (e.g., seq and samples) can reside on different cache lines. To fix this, simply wrap the global data in a ABSL_CACHE_ALIGNED structure. This is similar to what we do for MutexGlobals. PiperOrigin-RevId: 351389466 GitOrigin-RevId: a0491c8d790972cd80e2d720fe1fdf5f711a6f1a Change-Id: Ie0fa80112043381cd37c84e2ab2b7334839f54b5
4 years ago
const char* data = subnode->tag == EXTERNAL ? subnode->external()->base
: subnode->flat()->Data();
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
subnode = NewSubstring(subnode, current_chunk_.data() - data, n);
subcord.contents_.set_tree(VerifyTree(subnode));
RemoveChunkPrefix(n);
return subcord;
}
// Range to read begins with a proper subrange of the current chunk.
assert(!current_chunk_.empty());
assert(current_leaf_ != nullptr);
Export of internal Abseil changes -- 6e9f93888bbe6718997ed90bbd049f1f3572b066 by Abseil Team <absl-team@google.com>: Fix status to be safe for self move assignment. PiperOrigin-RevId: 346815392 -- 35cae74a977f258e81dfbe925fb5a34cb6632036 by Gennadiy Rozental <rogeeff@google.com>: Eliminate unnecessary access to the global vars. PiperOrigin-RevId: 346777168 -- e7e786c243069060f5d6c1c945decb4b0b83f95b by Andy Getzendanner <durandal@google.com>: Internal change. PiperOrigin-RevId: 346685656 -- 4ccd41c48f1a83cfa20b3ea534f743dd7d788376 by Abseil Team <absl-team@google.com>: Move CordRep Ref() and Unref() logic into cord_internal.h This change moves Ref() and Unref() logic out of cord.cc into cord_internal. The main purpose is to make upcoming ring buffer changes easier to isolate from existing cord.cc code. Notice that this removes the nullptr check from Unref() and now requires it to be non null (which held true most times). We may need to rethink if the 'unref unlikely one' is the common case: are cordreps most likely shared? This may be something between 'root' and non root nodes, i.e., is it more likely for leaf / flat nodes in large cords to be shared than top level cordreps being shared? Vice versa? Benchmarks say that we mostly shouldn't care, the caveat being that atomic ops seem more expensive on upcoming archs (arcadia) so we should error on the side of an extra IsOne() branch saving us single owned atomic ops. PiperOrigin-RevId: 346676121 -- f0babab103b9e60d61ba09482d468985e43eceb3 by Samuel Benzaquen <sbenza@google.com>: Fix iterator based constructor and `.insert` members to only require EmplaceConstructible as the standard specifies. PiperOrigin-RevId: 346616707 -- 8f48eedda02277f9c96a88ed7726e34b557cce20 by Evan Brown <ezb@google.com>: Fix a bug in binary_search_impl when there's a transparent, three-way comparator that has different equivalence classes for different lookup types. Add a new can_have_multiple_equivalent_keys method to share the common logic for these cases. PiperOrigin-RevId: 346605948 -- 649183cb3cc9383431de9c81fb1c0f885d4001ae by Abseil Team <absl-team@google.com>: Add benchmark for accessing a Duration flag. PiperOrigin-RevId: 346594843 -- fefdb046520871af63ce2229e2f7cccfc0483dea by Abseil Team <absl-team@google.com>: Restructure CordReader for upcoming ring buffer changes. PiperOrigin-RevId: 346410642 -- 8b2f50e7da0ebab06ead5f94e366e984ca23cb6a by Abseil Team <absl-team@google.com>: Wire in an internal-only flag to toggle upcoming ring buffer changes on/off for experimentation. PiperOrigin-RevId: 346199111 GitOrigin-RevId: 6e9f93888bbe6718997ed90bbd049f1f3572b066 Change-Id: I8f34866b25a79209cb5448bbb28dd3044111d2e9
4 years ago
CordRep* subnode = CordRep::Ref(current_leaf_);
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
if (current_chunk_.size() < subnode->length) {
Export of internal Abseil changes -- a0491c8d790972cd80e2d720fe1fdf5f711a6f1a by Greg Falcon <gfalcon@google.com>: Stop directly accessing CordRepFlat data via CordRep::data. The old pattern of access breaks the `CordRep` type abstraction; since `CordRep::data` is not in general guaranteed to contain the chunk's data, we shouldn't access it that way. This incidentally adds an assertion check (via the flat() accessor) that the CordRep is indeed flat on each such access, but a manual inspection of the code, as well as the fact that this code currently works, suggest that this is always true.) PiperOrigin-RevId: 351592344 -- f40c3b43ca5b1d7e23cd45f1ffac1783105ac1a3 by Abseil Team <absl-team@google.com>: Revert 18abb2902b9f06c63a968b24d3dda785ebf99a22 PiperOrigin-RevId: 351523518 -- 18abb2902b9f06c63a968b24d3dda785ebf99a22 by Abseil Team <absl-team@google.com>: Internal change PiperOrigin-RevId: 351512412 -- 9b881602d45e95e06089792c7627cd56528a255a by Abseil Team <absl-team@google.com>: Keep time's global state in a cacheline-aligned structure. Keeping the global state as separate global variables results in two issues: 1) False sharing with adjacent global data (e.g., cycle clock source), since the global fields are updated every O(10usec). 2) The hot global fields (e.g., seq and samples) can reside on different cache lines. To fix this, simply wrap the global data in a ABSL_CACHE_ALIGNED structure. This is similar to what we do for MutexGlobals. PiperOrigin-RevId: 351389466 GitOrigin-RevId: a0491c8d790972cd80e2d720fe1fdf5f711a6f1a Change-Id: Ie0fa80112043381cd37c84e2ab2b7334839f54b5
4 years ago
const char* data = subnode->tag == EXTERNAL ? subnode->external()->base
: subnode->flat()->Data();
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
subnode = NewSubstring(subnode, current_chunk_.data() - data,
current_chunk_.size());
}
n -= current_chunk_.size();
bytes_remaining_ -= current_chunk_.size();
// Process the next node(s) on the stack, reading whole subtrees depending on
// their length and how many bytes we are advancing.
CordRep* node = nullptr;
Export of internal Abseil changes -- 0bfa836596a9c787a2f0bdc283011dd1f6810c6e by Benjamin Barenblat <bbaren@google.com>: Ignore missing CPU frequency on more architectures Linux on MIPS, PA-RISC, RISC-V, and SystemZ doesn’t expose the nominal CPU frequency via /sys, so don’t worry if `NominalCPUFrequency` returns 1.0 on those platforms. Some POWER machines expose the CPU frequency; others do not. Since we can’t predict which type of machine the tests will run on, simply disable testing for `NominalCPUFrequency` on POWER. PiperOrigin-RevId: 347079873 -- 492b6834ed4a07cbc3abccd846f7e37d8c556ee5 by Benjamin Barenblat <bbaren@google.com>: Use ABSL_HAVE_THREAD_LOCAL macro instead of copying code Reduce code duplication by checking the ABSL_HAVE_THREAD_LOCAL macro instead of copying code from base/config.h. PiperOrigin-RevId: 347079561 -- 8d656efce4da9cb032094377e58493d98427a536 by Abseil Team <absl-team@google.com>: Rollback PiperOrigin-RevId: 347078779 -- 221bc69ec6dd7e2777ffcff6942584f979ef6382 by Abseil Team <absl-team@google.com>: Add flag for 'shallow subcord' feature for experimental ring buffer rollout There is a potential trade-off of CPU cost vs over-sharing cord data for subcord of large cords. This flag allows making subcords shallow for ringbuffers (with a potential larger waste of referenced source cords), which allows us to make subcord fast for this apps that do no persist (unmodified / plain copied) sub cords. This change also introduces constants for the default settings, intended to keep the internal cord settings concistent with external flags. PiperOrigin-RevId: 347053271 -- 00a56c24293566734009f6bf2169a83fb37a35ba by Abseil Team <absl-team@google.com>: Revert the usage of variant<> in Cord iterator and reader. The introduction of the variant may lead to some missed compiler optimizations. PiperOrigin-RevId: 347053041 -- c7b7b5ed7e3ab46b1e75b80f1a7de0bda26c8f70 by Chris Kennelly <ckennelly@google.com>: Release library for integer power-of-2 functions and bit counting. PiperOrigin-RevId: 347035065 -- 5a035c0d9840b251967f9e7039fc6a4e01dd52f3 by Abseil Team <absl-team@google.com>: Restructure Cord::ChunkIterator for future ring buffer support. PiperOrigin-RevId: 346890054 GitOrigin-RevId: 0bfa836596a9c787a2f0bdc283011dd1f6810c6e Change-Id: I3a58e2a44cb4c6f2116c43e2a4ccbc319d3ccecf
4 years ago
while (!stack_of_right_children.empty()) {
node = stack_of_right_children.back();
stack_of_right_children.pop_back();
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
if (node->length > n) break;
// TODO(qrczak): This might unnecessarily recreate existing concat nodes.
// Avoiding that would need pretty complicated logic (instead of
Export of internal Abseil changes -- 0bfa836596a9c787a2f0bdc283011dd1f6810c6e by Benjamin Barenblat <bbaren@google.com>: Ignore missing CPU frequency on more architectures Linux on MIPS, PA-RISC, RISC-V, and SystemZ doesn’t expose the nominal CPU frequency via /sys, so don’t worry if `NominalCPUFrequency` returns 1.0 on those platforms. Some POWER machines expose the CPU frequency; others do not. Since we can’t predict which type of machine the tests will run on, simply disable testing for `NominalCPUFrequency` on POWER. PiperOrigin-RevId: 347079873 -- 492b6834ed4a07cbc3abccd846f7e37d8c556ee5 by Benjamin Barenblat <bbaren@google.com>: Use ABSL_HAVE_THREAD_LOCAL macro instead of copying code Reduce code duplication by checking the ABSL_HAVE_THREAD_LOCAL macro instead of copying code from base/config.h. PiperOrigin-RevId: 347079561 -- 8d656efce4da9cb032094377e58493d98427a536 by Abseil Team <absl-team@google.com>: Rollback PiperOrigin-RevId: 347078779 -- 221bc69ec6dd7e2777ffcff6942584f979ef6382 by Abseil Team <absl-team@google.com>: Add flag for 'shallow subcord' feature for experimental ring buffer rollout There is a potential trade-off of CPU cost vs over-sharing cord data for subcord of large cords. This flag allows making subcords shallow for ringbuffers (with a potential larger waste of referenced source cords), which allows us to make subcord fast for this apps that do no persist (unmodified / plain copied) sub cords. This change also introduces constants for the default settings, intended to keep the internal cord settings concistent with external flags. PiperOrigin-RevId: 347053271 -- 00a56c24293566734009f6bf2169a83fb37a35ba by Abseil Team <absl-team@google.com>: Revert the usage of variant<> in Cord iterator and reader. The introduction of the variant may lead to some missed compiler optimizations. PiperOrigin-RevId: 347053041 -- c7b7b5ed7e3ab46b1e75b80f1a7de0bda26c8f70 by Chris Kennelly <ckennelly@google.com>: Release library for integer power-of-2 functions and bit counting. PiperOrigin-RevId: 347035065 -- 5a035c0d9840b251967f9e7039fc6a4e01dd52f3 by Abseil Team <absl-team@google.com>: Restructure Cord::ChunkIterator for future ring buffer support. PiperOrigin-RevId: 346890054 GitOrigin-RevId: 0bfa836596a9c787a2f0bdc283011dd1f6810c6e Change-Id: I3a58e2a44cb4c6f2116c43e2a4ccbc319d3ccecf
4 years ago
// current_leaf, keep current_subtree_ which points to the highest node
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
// such that the current leaf can be found on the path of left children
// starting from current_subtree_; delay creating subnode while node is
// below current_subtree_; find the proper node along the path of left
// children starting from current_subtree_ if this loop exits while staying
// below current_subtree_; etc.; alternatively, push parents instead of
// right children on the stack).
Export of internal Abseil changes -- 6e9f93888bbe6718997ed90bbd049f1f3572b066 by Abseil Team <absl-team@google.com>: Fix status to be safe for self move assignment. PiperOrigin-RevId: 346815392 -- 35cae74a977f258e81dfbe925fb5a34cb6632036 by Gennadiy Rozental <rogeeff@google.com>: Eliminate unnecessary access to the global vars. PiperOrigin-RevId: 346777168 -- e7e786c243069060f5d6c1c945decb4b0b83f95b by Andy Getzendanner <durandal@google.com>: Internal change. PiperOrigin-RevId: 346685656 -- 4ccd41c48f1a83cfa20b3ea534f743dd7d788376 by Abseil Team <absl-team@google.com>: Move CordRep Ref() and Unref() logic into cord_internal.h This change moves Ref() and Unref() logic out of cord.cc into cord_internal. The main purpose is to make upcoming ring buffer changes easier to isolate from existing cord.cc code. Notice that this removes the nullptr check from Unref() and now requires it to be non null (which held true most times). We may need to rethink if the 'unref unlikely one' is the common case: are cordreps most likely shared? This may be something between 'root' and non root nodes, i.e., is it more likely for leaf / flat nodes in large cords to be shared than top level cordreps being shared? Vice versa? Benchmarks say that we mostly shouldn't care, the caveat being that atomic ops seem more expensive on upcoming archs (arcadia) so we should error on the side of an extra IsOne() branch saving us single owned atomic ops. PiperOrigin-RevId: 346676121 -- f0babab103b9e60d61ba09482d468985e43eceb3 by Samuel Benzaquen <sbenza@google.com>: Fix iterator based constructor and `.insert` members to only require EmplaceConstructible as the standard specifies. PiperOrigin-RevId: 346616707 -- 8f48eedda02277f9c96a88ed7726e34b557cce20 by Evan Brown <ezb@google.com>: Fix a bug in binary_search_impl when there's a transparent, three-way comparator that has different equivalence classes for different lookup types. Add a new can_have_multiple_equivalent_keys method to share the common logic for these cases. PiperOrigin-RevId: 346605948 -- 649183cb3cc9383431de9c81fb1c0f885d4001ae by Abseil Team <absl-team@google.com>: Add benchmark for accessing a Duration flag. PiperOrigin-RevId: 346594843 -- fefdb046520871af63ce2229e2f7cccfc0483dea by Abseil Team <absl-team@google.com>: Restructure CordReader for upcoming ring buffer changes. PiperOrigin-RevId: 346410642 -- 8b2f50e7da0ebab06ead5f94e366e984ca23cb6a by Abseil Team <absl-team@google.com>: Wire in an internal-only flag to toggle upcoming ring buffer changes on/off for experimentation. PiperOrigin-RevId: 346199111 GitOrigin-RevId: 6e9f93888bbe6718997ed90bbd049f1f3572b066 Change-Id: I8f34866b25a79209cb5448bbb28dd3044111d2e9
4 years ago
subnode = Concat(subnode, CordRep::Ref(node));
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
n -= node->length;
bytes_remaining_ -= node->length;
node = nullptr;
}
if (node == nullptr) {
// We have reached the end of the Cord.
assert(bytes_remaining_ == 0);
subcord.contents_.set_tree(VerifyTree(subnode));
return subcord;
}
// Walk down the appropriate branches until we hit a non-CONCAT node. Save the
// right children to the stack for subsequent traversal.
while (node->tag == CONCAT) {
if (node->concat()->left->length > n) {
// Push right, descend left.
Export of internal Abseil changes -- 0bfa836596a9c787a2f0bdc283011dd1f6810c6e by Benjamin Barenblat <bbaren@google.com>: Ignore missing CPU frequency on more architectures Linux on MIPS, PA-RISC, RISC-V, and SystemZ doesn’t expose the nominal CPU frequency via /sys, so don’t worry if `NominalCPUFrequency` returns 1.0 on those platforms. Some POWER machines expose the CPU frequency; others do not. Since we can’t predict which type of machine the tests will run on, simply disable testing for `NominalCPUFrequency` on POWER. PiperOrigin-RevId: 347079873 -- 492b6834ed4a07cbc3abccd846f7e37d8c556ee5 by Benjamin Barenblat <bbaren@google.com>: Use ABSL_HAVE_THREAD_LOCAL macro instead of copying code Reduce code duplication by checking the ABSL_HAVE_THREAD_LOCAL macro instead of copying code from base/config.h. PiperOrigin-RevId: 347079561 -- 8d656efce4da9cb032094377e58493d98427a536 by Abseil Team <absl-team@google.com>: Rollback PiperOrigin-RevId: 347078779 -- 221bc69ec6dd7e2777ffcff6942584f979ef6382 by Abseil Team <absl-team@google.com>: Add flag for 'shallow subcord' feature for experimental ring buffer rollout There is a potential trade-off of CPU cost vs over-sharing cord data for subcord of large cords. This flag allows making subcords shallow for ringbuffers (with a potential larger waste of referenced source cords), which allows us to make subcord fast for this apps that do no persist (unmodified / plain copied) sub cords. This change also introduces constants for the default settings, intended to keep the internal cord settings concistent with external flags. PiperOrigin-RevId: 347053271 -- 00a56c24293566734009f6bf2169a83fb37a35ba by Abseil Team <absl-team@google.com>: Revert the usage of variant<> in Cord iterator and reader. The introduction of the variant may lead to some missed compiler optimizations. PiperOrigin-RevId: 347053041 -- c7b7b5ed7e3ab46b1e75b80f1a7de0bda26c8f70 by Chris Kennelly <ckennelly@google.com>: Release library for integer power-of-2 functions and bit counting. PiperOrigin-RevId: 347035065 -- 5a035c0d9840b251967f9e7039fc6a4e01dd52f3 by Abseil Team <absl-team@google.com>: Restructure Cord::ChunkIterator for future ring buffer support. PiperOrigin-RevId: 346890054 GitOrigin-RevId: 0bfa836596a9c787a2f0bdc283011dd1f6810c6e Change-Id: I3a58e2a44cb4c6f2116c43e2a4ccbc319d3ccecf
4 years ago
stack_of_right_children.push_back(node->concat()->right);
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
node = node->concat()->left;
} else {
// Read left, descend right.
Export of internal Abseil changes -- 6e9f93888bbe6718997ed90bbd049f1f3572b066 by Abseil Team <absl-team@google.com>: Fix status to be safe for self move assignment. PiperOrigin-RevId: 346815392 -- 35cae74a977f258e81dfbe925fb5a34cb6632036 by Gennadiy Rozental <rogeeff@google.com>: Eliminate unnecessary access to the global vars. PiperOrigin-RevId: 346777168 -- e7e786c243069060f5d6c1c945decb4b0b83f95b by Andy Getzendanner <durandal@google.com>: Internal change. PiperOrigin-RevId: 346685656 -- 4ccd41c48f1a83cfa20b3ea534f743dd7d788376 by Abseil Team <absl-team@google.com>: Move CordRep Ref() and Unref() logic into cord_internal.h This change moves Ref() and Unref() logic out of cord.cc into cord_internal. The main purpose is to make upcoming ring buffer changes easier to isolate from existing cord.cc code. Notice that this removes the nullptr check from Unref() and now requires it to be non null (which held true most times). We may need to rethink if the 'unref unlikely one' is the common case: are cordreps most likely shared? This may be something between 'root' and non root nodes, i.e., is it more likely for leaf / flat nodes in large cords to be shared than top level cordreps being shared? Vice versa? Benchmarks say that we mostly shouldn't care, the caveat being that atomic ops seem more expensive on upcoming archs (arcadia) so we should error on the side of an extra IsOne() branch saving us single owned atomic ops. PiperOrigin-RevId: 346676121 -- f0babab103b9e60d61ba09482d468985e43eceb3 by Samuel Benzaquen <sbenza@google.com>: Fix iterator based constructor and `.insert` members to only require EmplaceConstructible as the standard specifies. PiperOrigin-RevId: 346616707 -- 8f48eedda02277f9c96a88ed7726e34b557cce20 by Evan Brown <ezb@google.com>: Fix a bug in binary_search_impl when there's a transparent, three-way comparator that has different equivalence classes for different lookup types. Add a new can_have_multiple_equivalent_keys method to share the common logic for these cases. PiperOrigin-RevId: 346605948 -- 649183cb3cc9383431de9c81fb1c0f885d4001ae by Abseil Team <absl-team@google.com>: Add benchmark for accessing a Duration flag. PiperOrigin-RevId: 346594843 -- fefdb046520871af63ce2229e2f7cccfc0483dea by Abseil Team <absl-team@google.com>: Restructure CordReader for upcoming ring buffer changes. PiperOrigin-RevId: 346410642 -- 8b2f50e7da0ebab06ead5f94e366e984ca23cb6a by Abseil Team <absl-team@google.com>: Wire in an internal-only flag to toggle upcoming ring buffer changes on/off for experimentation. PiperOrigin-RevId: 346199111 GitOrigin-RevId: 6e9f93888bbe6718997ed90bbd049f1f3572b066 Change-Id: I8f34866b25a79209cb5448bbb28dd3044111d2e9
4 years ago
subnode = Concat(subnode, CordRep::Ref(node->concat()->left));
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
n -= node->concat()->left->length;
bytes_remaining_ -= node->concat()->left->length;
node = node->concat()->right;
}
}
// Get the child node if we encounter a SUBSTRING.
size_t offset = 0;
size_t length = node->length;
if (node->tag == SUBSTRING) {
offset = node->substring()->start;
node = node->substring()->child;
}
// Range to read ends with a proper (possibly empty) subrange of the current
// chunk.
assert(node->tag == EXTERNAL || node->tag >= FLAT);
assert(length > n);
Export of internal Abseil changes -- 6e9f93888bbe6718997ed90bbd049f1f3572b066 by Abseil Team <absl-team@google.com>: Fix status to be safe for self move assignment. PiperOrigin-RevId: 346815392 -- 35cae74a977f258e81dfbe925fb5a34cb6632036 by Gennadiy Rozental <rogeeff@google.com>: Eliminate unnecessary access to the global vars. PiperOrigin-RevId: 346777168 -- e7e786c243069060f5d6c1c945decb4b0b83f95b by Andy Getzendanner <durandal@google.com>: Internal change. PiperOrigin-RevId: 346685656 -- 4ccd41c48f1a83cfa20b3ea534f743dd7d788376 by Abseil Team <absl-team@google.com>: Move CordRep Ref() and Unref() logic into cord_internal.h This change moves Ref() and Unref() logic out of cord.cc into cord_internal. The main purpose is to make upcoming ring buffer changes easier to isolate from existing cord.cc code. Notice that this removes the nullptr check from Unref() and now requires it to be non null (which held true most times). We may need to rethink if the 'unref unlikely one' is the common case: are cordreps most likely shared? This may be something between 'root' and non root nodes, i.e., is it more likely for leaf / flat nodes in large cords to be shared than top level cordreps being shared? Vice versa? Benchmarks say that we mostly shouldn't care, the caveat being that atomic ops seem more expensive on upcoming archs (arcadia) so we should error on the side of an extra IsOne() branch saving us single owned atomic ops. PiperOrigin-RevId: 346676121 -- f0babab103b9e60d61ba09482d468985e43eceb3 by Samuel Benzaquen <sbenza@google.com>: Fix iterator based constructor and `.insert` members to only require EmplaceConstructible as the standard specifies. PiperOrigin-RevId: 346616707 -- 8f48eedda02277f9c96a88ed7726e34b557cce20 by Evan Brown <ezb@google.com>: Fix a bug in binary_search_impl when there's a transparent, three-way comparator that has different equivalence classes for different lookup types. Add a new can_have_multiple_equivalent_keys method to share the common logic for these cases. PiperOrigin-RevId: 346605948 -- 649183cb3cc9383431de9c81fb1c0f885d4001ae by Abseil Team <absl-team@google.com>: Add benchmark for accessing a Duration flag. PiperOrigin-RevId: 346594843 -- fefdb046520871af63ce2229e2f7cccfc0483dea by Abseil Team <absl-team@google.com>: Restructure CordReader for upcoming ring buffer changes. PiperOrigin-RevId: 346410642 -- 8b2f50e7da0ebab06ead5f94e366e984ca23cb6a by Abseil Team <absl-team@google.com>: Wire in an internal-only flag to toggle upcoming ring buffer changes on/off for experimentation. PiperOrigin-RevId: 346199111 GitOrigin-RevId: 6e9f93888bbe6718997ed90bbd049f1f3572b066 Change-Id: I8f34866b25a79209cb5448bbb28dd3044111d2e9
4 years ago
if (n > 0) {
subnode = Concat(subnode, NewSubstring(CordRep::Ref(node), offset, n));
}
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
const char* data =
Export of internal Abseil changes -- a0491c8d790972cd80e2d720fe1fdf5f711a6f1a by Greg Falcon <gfalcon@google.com>: Stop directly accessing CordRepFlat data via CordRep::data. The old pattern of access breaks the `CordRep` type abstraction; since `CordRep::data` is not in general guaranteed to contain the chunk's data, we shouldn't access it that way. This incidentally adds an assertion check (via the flat() accessor) that the CordRep is indeed flat on each such access, but a manual inspection of the code, as well as the fact that this code currently works, suggest that this is always true.) PiperOrigin-RevId: 351592344 -- f40c3b43ca5b1d7e23cd45f1ffac1783105ac1a3 by Abseil Team <absl-team@google.com>: Revert 18abb2902b9f06c63a968b24d3dda785ebf99a22 PiperOrigin-RevId: 351523518 -- 18abb2902b9f06c63a968b24d3dda785ebf99a22 by Abseil Team <absl-team@google.com>: Internal change PiperOrigin-RevId: 351512412 -- 9b881602d45e95e06089792c7627cd56528a255a by Abseil Team <absl-team@google.com>: Keep time's global state in a cacheline-aligned structure. Keeping the global state as separate global variables results in two issues: 1) False sharing with adjacent global data (e.g., cycle clock source), since the global fields are updated every O(10usec). 2) The hot global fields (e.g., seq and samples) can reside on different cache lines. To fix this, simply wrap the global data in a ABSL_CACHE_ALIGNED structure. This is similar to what we do for MutexGlobals. PiperOrigin-RevId: 351389466 GitOrigin-RevId: a0491c8d790972cd80e2d720fe1fdf5f711a6f1a Change-Id: Ie0fa80112043381cd37c84e2ab2b7334839f54b5
4 years ago
node->tag == EXTERNAL ? node->external()->base : node->flat()->Data();
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
current_chunk_ = absl::string_view(data + offset + n, length - n);
current_leaf_ = node;
bytes_remaining_ -= n;
subcord.contents_.set_tree(VerifyTree(subnode));
return subcord;
}
void Cord::ChunkIterator::AdvanceBytesSlowPath(size_t n) {
assert(bytes_remaining_ >= n && "Attempted to iterate past `end()`");
assert(n >= current_chunk_.size()); // This should only be called when
// iterating to a new node.
n -= current_chunk_.size();
bytes_remaining_ -= current_chunk_.size();
if (stack_of_right_children_.empty()) {
Export of internal Abseil changes -- 0bfa836596a9c787a2f0bdc283011dd1f6810c6e by Benjamin Barenblat <bbaren@google.com>: Ignore missing CPU frequency on more architectures Linux on MIPS, PA-RISC, RISC-V, and SystemZ doesn’t expose the nominal CPU frequency via /sys, so don’t worry if `NominalCPUFrequency` returns 1.0 on those platforms. Some POWER machines expose the CPU frequency; others do not. Since we can’t predict which type of machine the tests will run on, simply disable testing for `NominalCPUFrequency` on POWER. PiperOrigin-RevId: 347079873 -- 492b6834ed4a07cbc3abccd846f7e37d8c556ee5 by Benjamin Barenblat <bbaren@google.com>: Use ABSL_HAVE_THREAD_LOCAL macro instead of copying code Reduce code duplication by checking the ABSL_HAVE_THREAD_LOCAL macro instead of copying code from base/config.h. PiperOrigin-RevId: 347079561 -- 8d656efce4da9cb032094377e58493d98427a536 by Abseil Team <absl-team@google.com>: Rollback PiperOrigin-RevId: 347078779 -- 221bc69ec6dd7e2777ffcff6942584f979ef6382 by Abseil Team <absl-team@google.com>: Add flag for 'shallow subcord' feature for experimental ring buffer rollout There is a potential trade-off of CPU cost vs over-sharing cord data for subcord of large cords. This flag allows making subcords shallow for ringbuffers (with a potential larger waste of referenced source cords), which allows us to make subcord fast for this apps that do no persist (unmodified / plain copied) sub cords. This change also introduces constants for the default settings, intended to keep the internal cord settings concistent with external flags. PiperOrigin-RevId: 347053271 -- 00a56c24293566734009f6bf2169a83fb37a35ba by Abseil Team <absl-team@google.com>: Revert the usage of variant<> in Cord iterator and reader. The introduction of the variant may lead to some missed compiler optimizations. PiperOrigin-RevId: 347053041 -- c7b7b5ed7e3ab46b1e75b80f1a7de0bda26c8f70 by Chris Kennelly <ckennelly@google.com>: Release library for integer power-of-2 functions and bit counting. PiperOrigin-RevId: 347035065 -- 5a035c0d9840b251967f9e7039fc6a4e01dd52f3 by Abseil Team <absl-team@google.com>: Restructure Cord::ChunkIterator for future ring buffer support. PiperOrigin-RevId: 346890054 GitOrigin-RevId: 0bfa836596a9c787a2f0bdc283011dd1f6810c6e Change-Id: I3a58e2a44cb4c6f2116c43e2a4ccbc319d3ccecf
4 years ago
// We have reached the end of the Cord.
assert(bytes_remaining_ == 0);
return;
}
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
// Process the next node(s) on the stack, skipping whole subtrees depending on
// their length and how many bytes we are advancing.
CordRep* node = nullptr;
auto& stack_of_right_children = stack_of_right_children_;
Export of internal Abseil changes -- 0bfa836596a9c787a2f0bdc283011dd1f6810c6e by Benjamin Barenblat <bbaren@google.com>: Ignore missing CPU frequency on more architectures Linux on MIPS, PA-RISC, RISC-V, and SystemZ doesn’t expose the nominal CPU frequency via /sys, so don’t worry if `NominalCPUFrequency` returns 1.0 on those platforms. Some POWER machines expose the CPU frequency; others do not. Since we can’t predict which type of machine the tests will run on, simply disable testing for `NominalCPUFrequency` on POWER. PiperOrigin-RevId: 347079873 -- 492b6834ed4a07cbc3abccd846f7e37d8c556ee5 by Benjamin Barenblat <bbaren@google.com>: Use ABSL_HAVE_THREAD_LOCAL macro instead of copying code Reduce code duplication by checking the ABSL_HAVE_THREAD_LOCAL macro instead of copying code from base/config.h. PiperOrigin-RevId: 347079561 -- 8d656efce4da9cb032094377e58493d98427a536 by Abseil Team <absl-team@google.com>: Rollback PiperOrigin-RevId: 347078779 -- 221bc69ec6dd7e2777ffcff6942584f979ef6382 by Abseil Team <absl-team@google.com>: Add flag for 'shallow subcord' feature for experimental ring buffer rollout There is a potential trade-off of CPU cost vs over-sharing cord data for subcord of large cords. This flag allows making subcords shallow for ringbuffers (with a potential larger waste of referenced source cords), which allows us to make subcord fast for this apps that do no persist (unmodified / plain copied) sub cords. This change also introduces constants for the default settings, intended to keep the internal cord settings concistent with external flags. PiperOrigin-RevId: 347053271 -- 00a56c24293566734009f6bf2169a83fb37a35ba by Abseil Team <absl-team@google.com>: Revert the usage of variant<> in Cord iterator and reader. The introduction of the variant may lead to some missed compiler optimizations. PiperOrigin-RevId: 347053041 -- c7b7b5ed7e3ab46b1e75b80f1a7de0bda26c8f70 by Chris Kennelly <ckennelly@google.com>: Release library for integer power-of-2 functions and bit counting. PiperOrigin-RevId: 347035065 -- 5a035c0d9840b251967f9e7039fc6a4e01dd52f3 by Abseil Team <absl-team@google.com>: Restructure Cord::ChunkIterator for future ring buffer support. PiperOrigin-RevId: 346890054 GitOrigin-RevId: 0bfa836596a9c787a2f0bdc283011dd1f6810c6e Change-Id: I3a58e2a44cb4c6f2116c43e2a4ccbc319d3ccecf
4 years ago
while (!stack_of_right_children.empty()) {
node = stack_of_right_children.back();
stack_of_right_children.pop_back();
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
if (node->length > n) break;
n -= node->length;
bytes_remaining_ -= node->length;
node = nullptr;
}
if (node == nullptr) {
// We have reached the end of the Cord.
assert(bytes_remaining_ == 0);
return;
}
// Walk down the appropriate branches until we hit a non-CONCAT node. Save the
// right children to the stack for subsequent traversal.
while (node->tag == CONCAT) {
if (node->concat()->left->length > n) {
// Push right, descend left.
Export of internal Abseil changes -- 0bfa836596a9c787a2f0bdc283011dd1f6810c6e by Benjamin Barenblat <bbaren@google.com>: Ignore missing CPU frequency on more architectures Linux on MIPS, PA-RISC, RISC-V, and SystemZ doesn’t expose the nominal CPU frequency via /sys, so don’t worry if `NominalCPUFrequency` returns 1.0 on those platforms. Some POWER machines expose the CPU frequency; others do not. Since we can’t predict which type of machine the tests will run on, simply disable testing for `NominalCPUFrequency` on POWER. PiperOrigin-RevId: 347079873 -- 492b6834ed4a07cbc3abccd846f7e37d8c556ee5 by Benjamin Barenblat <bbaren@google.com>: Use ABSL_HAVE_THREAD_LOCAL macro instead of copying code Reduce code duplication by checking the ABSL_HAVE_THREAD_LOCAL macro instead of copying code from base/config.h. PiperOrigin-RevId: 347079561 -- 8d656efce4da9cb032094377e58493d98427a536 by Abseil Team <absl-team@google.com>: Rollback PiperOrigin-RevId: 347078779 -- 221bc69ec6dd7e2777ffcff6942584f979ef6382 by Abseil Team <absl-team@google.com>: Add flag for 'shallow subcord' feature for experimental ring buffer rollout There is a potential trade-off of CPU cost vs over-sharing cord data for subcord of large cords. This flag allows making subcords shallow for ringbuffers (with a potential larger waste of referenced source cords), which allows us to make subcord fast for this apps that do no persist (unmodified / plain copied) sub cords. This change also introduces constants for the default settings, intended to keep the internal cord settings concistent with external flags. PiperOrigin-RevId: 347053271 -- 00a56c24293566734009f6bf2169a83fb37a35ba by Abseil Team <absl-team@google.com>: Revert the usage of variant<> in Cord iterator and reader. The introduction of the variant may lead to some missed compiler optimizations. PiperOrigin-RevId: 347053041 -- c7b7b5ed7e3ab46b1e75b80f1a7de0bda26c8f70 by Chris Kennelly <ckennelly@google.com>: Release library for integer power-of-2 functions and bit counting. PiperOrigin-RevId: 347035065 -- 5a035c0d9840b251967f9e7039fc6a4e01dd52f3 by Abseil Team <absl-team@google.com>: Restructure Cord::ChunkIterator for future ring buffer support. PiperOrigin-RevId: 346890054 GitOrigin-RevId: 0bfa836596a9c787a2f0bdc283011dd1f6810c6e Change-Id: I3a58e2a44cb4c6f2116c43e2a4ccbc319d3ccecf
4 years ago
stack_of_right_children.push_back(node->concat()->right);
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
node = node->concat()->left;
} else {
// Skip left, descend right.
n -= node->concat()->left->length;
bytes_remaining_ -= node->concat()->left->length;
node = node->concat()->right;
}
}
// Get the child node if we encounter a SUBSTRING.
size_t offset = 0;
size_t length = node->length;
if (node->tag == SUBSTRING) {
offset = node->substring()->start;
node = node->substring()->child;
}
assert(node->tag == EXTERNAL || node->tag >= FLAT);
assert(length > n);
const char* data =
Export of internal Abseil changes -- a0491c8d790972cd80e2d720fe1fdf5f711a6f1a by Greg Falcon <gfalcon@google.com>: Stop directly accessing CordRepFlat data via CordRep::data. The old pattern of access breaks the `CordRep` type abstraction; since `CordRep::data` is not in general guaranteed to contain the chunk's data, we shouldn't access it that way. This incidentally adds an assertion check (via the flat() accessor) that the CordRep is indeed flat on each such access, but a manual inspection of the code, as well as the fact that this code currently works, suggest that this is always true.) PiperOrigin-RevId: 351592344 -- f40c3b43ca5b1d7e23cd45f1ffac1783105ac1a3 by Abseil Team <absl-team@google.com>: Revert 18abb2902b9f06c63a968b24d3dda785ebf99a22 PiperOrigin-RevId: 351523518 -- 18abb2902b9f06c63a968b24d3dda785ebf99a22 by Abseil Team <absl-team@google.com>: Internal change PiperOrigin-RevId: 351512412 -- 9b881602d45e95e06089792c7627cd56528a255a by Abseil Team <absl-team@google.com>: Keep time's global state in a cacheline-aligned structure. Keeping the global state as separate global variables results in two issues: 1) False sharing with adjacent global data (e.g., cycle clock source), since the global fields are updated every O(10usec). 2) The hot global fields (e.g., seq and samples) can reside on different cache lines. To fix this, simply wrap the global data in a ABSL_CACHE_ALIGNED structure. This is similar to what we do for MutexGlobals. PiperOrigin-RevId: 351389466 GitOrigin-RevId: a0491c8d790972cd80e2d720fe1fdf5f711a6f1a Change-Id: Ie0fa80112043381cd37c84e2ab2b7334839f54b5
4 years ago
node->tag == EXTERNAL ? node->external()->base : node->flat()->Data();
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
current_chunk_ = absl::string_view(data + offset + n, length - n);
current_leaf_ = node;
bytes_remaining_ -= n;
}
char Cord::operator[](size_t i) const {
ABSL_HARDENING_ASSERT(i < size());
size_t offset = i;
const CordRep* rep = contents_.tree();
if (rep == nullptr) {
return contents_.data()[i];
}
while (true) {
assert(rep != nullptr);
assert(offset < rep->length);
if (rep->tag >= FLAT) {
// Get the "i"th character directly from the flat array.
Export of internal Abseil changes -- a0491c8d790972cd80e2d720fe1fdf5f711a6f1a by Greg Falcon <gfalcon@google.com>: Stop directly accessing CordRepFlat data via CordRep::data. The old pattern of access breaks the `CordRep` type abstraction; since `CordRep::data` is not in general guaranteed to contain the chunk's data, we shouldn't access it that way. This incidentally adds an assertion check (via the flat() accessor) that the CordRep is indeed flat on each such access, but a manual inspection of the code, as well as the fact that this code currently works, suggest that this is always true.) PiperOrigin-RevId: 351592344 -- f40c3b43ca5b1d7e23cd45f1ffac1783105ac1a3 by Abseil Team <absl-team@google.com>: Revert 18abb2902b9f06c63a968b24d3dda785ebf99a22 PiperOrigin-RevId: 351523518 -- 18abb2902b9f06c63a968b24d3dda785ebf99a22 by Abseil Team <absl-team@google.com>: Internal change PiperOrigin-RevId: 351512412 -- 9b881602d45e95e06089792c7627cd56528a255a by Abseil Team <absl-team@google.com>: Keep time's global state in a cacheline-aligned structure. Keeping the global state as separate global variables results in two issues: 1) False sharing with adjacent global data (e.g., cycle clock source), since the global fields are updated every O(10usec). 2) The hot global fields (e.g., seq and samples) can reside on different cache lines. To fix this, simply wrap the global data in a ABSL_CACHE_ALIGNED structure. This is similar to what we do for MutexGlobals. PiperOrigin-RevId: 351389466 GitOrigin-RevId: a0491c8d790972cd80e2d720fe1fdf5f711a6f1a Change-Id: Ie0fa80112043381cd37c84e2ab2b7334839f54b5
4 years ago
return rep->flat()->Data()[offset];
Export of internal Abseil changes -- c68f1886f5e8fd90eb0c2d2e68feaf00a7cdacda by CJ Johnson <johnsoncj@google.com>: Introduce absl::Cleanup to the OSS repo PiperOrigin-RevId: 354583156 -- 17030cf388e10f7eb959e3e566326d1072ce392e by Abseil Team <absl-team@google.com>: Internal change only PiperOrigin-RevId: 354574953 -- e979d7236d4f3252e79ddda6739b67a9a326bf6d by CJ Johnson <johnsoncj@google.com>: Internal change PiperOrigin-RevId: 354545297 -- 7ea02b3783f7f49ef97d86a8f6580a19cc57df14 by Abseil Team <absl-team@google.com>: Pre-allocate memory for vectors where the size is known. PiperOrigin-RevId: 354344576 -- 9246c7cb11f1d6444f79ebe25acc69a8a9b870e0 by Matt Kulukundis <kfm@google.com>: Add support for Elbrus 2000 (e2k) Import of https://github.com/abseil/abseil-cpp/pull/889 PiperOrigin-RevId: 354344013 -- 0fc93d359cc1fb307552e917b37b7b2e7eed822f by Abseil Team <absl-team@google.com>: Integrate CordRepRing logic into cord (but do not enable it) PiperOrigin-RevId: 354312238 -- eda05622f7da71466723acb33403f783529df24b by Abseil Team <absl-team@google.com>: Protect ignore diagnostic with "__has_warning". PiperOrigin-RevId: 354112334 -- 47716c5d8fb10efa4fdd801d28bac414c6f8ec32 by Abseil Team <absl-team@google.com>: Rearrange InlinedVector copy constructor and destructor to treat a few special cases inline and then tail-call a non-inlined routine for the rest. In particular, we optimize for empty vectors in both cases. Added a couple of benchmarks that copy either an InlVec<int64> or an InlVec<InlVec<int64>>. Speed difference: ``` BM_CopyTrivial/0 0.92ns +- 0% 0.47ns +- 0% -48.91% (p=0.000 n=11+12) BM_CopyTrivial/1 0.92ns +- 0% 1.15ns +- 0% +25.00% (p=0.000 n=10+9) BM_CopyTrivial/8 8.57ns +- 0% 10.72ns +- 1% +25.16% (p=0.000 n=10+12) BM_CopyNonTrivial/0 3.21ns +- 0% 0.70ns +- 0% -78.23% (p=0.000 n=12+10) BM_CopyNonTrivial/1 5.88ns +- 1% 5.51ns +- 0% -6.28% (p=0.000 n=10+8) BM_CopyNonTrivial/8 21.5ns +- 1% 15.2ns +- 2% -29.23% (p=0.000 n=12+12) ``` Note: the slowdowns are a few cycles which is expected given the procedure call added in that case. We decided this is a good tradeoff given the code size reductions and the more significant speedups for empty vectors. Size difference (as measured by nm): ``` BM_CopyTrivial from 1048 bytes to 326 bytes. BM_CopyNonTrivial from 749 bytes to 470 bytes. ``` Code size for a large binary drops by ~500KB (from 349415719 to 348906015 348906191). All of the benchmarks that showed a significant difference: Ones that improve with this CL: ``` BM_CopyNonTrivial/0 3.21ns +- 0% 0.70ns +- 0% -78.23% (p=0.000 n=12+10) BM_InlinedVectorFillString/0 0.93ns +- 0% 0.24ns +- 0% -74.19% (p=0.000 n=12+10) BM_InlinedVectorAssignments/1 10.5ns +- 0% 4.1ns +- 0% -60.64% (p=0.000 n=11+10) BM_InlinedVectorAssignments/2 10.7ns +- 0% 4.4ns +- 0% -59.08% (p=0.000 n=11+11) BM_CopyTrivial/0 0.92ns +- 0% 0.47ns +- 0% -48.91% (p=0.000 n=11+12) BM_CopyNonTrivial/8 21.5ns +- 1% 15.2ns +- 2% -29.23% (p=0.000 n=12+12) BM_StdVectorEmpty 0.47ns +- 1% 0.35ns +- 0% -24.73% (p=0.000 n=12+12) BM_StdVectorSize 0.46ns +- 2% 0.35ns +- 0% -24.32% (p=0.000 n=12+12) BM_SwapElements<LargeCopyableOnly>/0 3.44ns +- 0% 2.76ns +- 1% -19.83% (p=0.000 n=11+11) BM_InlinedVectorFillRange/256 20.7ns +- 1% 17.8ns +- 0% -14.08% (p=0.000 n=12+9) BM_CopyNonTrivial/1 5.88ns +- 1% 5.51ns +- 0% -6.28% (p=0.000 n=10+8) BM_SwapElements<LargeCopyableMovable>/1 4.19ns +- 0% 3.95ns +- 1% -5.63% (p=0.000 n=11+12) BM_SwapElements<LargeCopyableMovableSwappable>/1 4.18ns +- 0% 3.99ns +- 0% -4.70% (p=0.000 n=9+11) BM_SwapElements<LargeCopyableMovable>/0 2.41ns +- 0% 2.31ns +- 0% -4.45% (p=0.000 n=12+12) BM_InlinedVectorFillRange/64 8.25ns +- 0% 8.04ns +- 0% -2.51% (p=0.000 n=12+11) BM_SwapElements<LargeCopyableOnly>/1 82.4ns +- 0% 81.5ns +- 0% -1.06% (p=0.000 n=12+12) ``` Ones that get worse with this CL: ``` BM_CopyTrivial/1 0.92ns +- 0% 1.15ns +- 0% +25.00% (p=0.000 n=10+9) BM_CopyTrivial/8 8.57ns +- 0% 10.72ns +- 1% +25.16% (p=0.000 n=10+12) BM_SwapElements<LargeCopyableMovableSwappable>/512 1.48ns +- 1% 1.66ns +- 1% +11.88% (p=0.000 n=12+12) BM_InlinedVectorFillString/1 11.5ns +- 0% 12.8ns +- 1% +11.62% (p=0.000 n=12+11) BM_SwapElements<LargeCopyableMovableSwappable>/64 1.48ns +- 2% 1.66ns +- 1% +11.66% (p=0.000 n=12+11) BM_SwapElements<LargeCopyableMovableSwappable>/1k 1.48ns +- 1% 1.65ns +- 2% +11.32% (p=0.000 n=12+12) BM_SwapElements<LargeCopyableMovable>/512 1.48ns +- 2% 1.58ns +- 4% +6.62% (p=0.000 n=11+12) BM_SwapElements<LargeCopyableMovable>/1k 1.49ns +- 2% 1.58ns +- 3% +6.05% (p=0.000 n=12+12) BM_SwapElements<LargeCopyableMovable>/64 1.48ns +- 2% 1.57ns +- 4% +6.04% (p=0.000 n=11+12) BM_InlinedVectorFillRange/1 4.81ns +- 0% 5.05ns +- 0% +4.83% (p=0.000 n=11+11) BM_InlinedVectorFillString/8 79.4ns +- 1% 83.1ns +- 1% +4.64% (p=0.000 n=10+12) BM_StdVectorFillString/1 16.3ns +- 0% 16.6ns +- 0% +2.13% (p=0.000 n=11+8) ``` PiperOrigin-RevId: 353906786 -- 8e26518b3cec9c598e5e9573c46c3bd1b03a67ef by Abseil Team <absl-team@google.com>: Internal change PiperOrigin-RevId: 353737330 -- f206ae0983e58c9904ed8b8f05f9caf564a446be by Matt Kulukundis <kfm@google.com>: Import of CCTZ from GitHub. PiperOrigin-RevId: 353682256 GitOrigin-RevId: c68f1886f5e8fd90eb0c2d2e68feaf00a7cdacda Change-Id: I5790c1036c4f543c701d1039848fabf7ae881ad8
4 years ago
} else if (rep->tag == RING) {
return rep->ring()->GetCharacter(offset);
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
} else if (rep->tag == EXTERNAL) {
// Get the "i"th character from the external array.
return rep->external()->base[offset];
} else if (rep->tag == CONCAT) {
// Recursively branch to the side of the concatenation that the "i"th
// character is on.
size_t left_length = rep->concat()->left->length;
if (offset < left_length) {
rep = rep->concat()->left;
} else {
offset -= left_length;
rep = rep->concat()->right;
}
} else {
// This must be a substring a node, so bypass it to get to the child.
assert(rep->tag == SUBSTRING);
offset += rep->substring()->start;
rep = rep->substring()->child;
}
}
}
absl::string_view Cord::FlattenSlowPath() {
size_t total_size = size();
CordRep* new_rep;
char* new_buffer;
// Try to put the contents into a new flat rep. If they won't fit in the
// biggest possible flat node, use an external rep instead.
if (total_size <= kMaxFlatLength) {
new_rep = CordRepFlat::New(total_size);
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
new_rep->length = total_size;
Export of internal Abseil changes -- a0491c8d790972cd80e2d720fe1fdf5f711a6f1a by Greg Falcon <gfalcon@google.com>: Stop directly accessing CordRepFlat data via CordRep::data. The old pattern of access breaks the `CordRep` type abstraction; since `CordRep::data` is not in general guaranteed to contain the chunk's data, we shouldn't access it that way. This incidentally adds an assertion check (via the flat() accessor) that the CordRep is indeed flat on each such access, but a manual inspection of the code, as well as the fact that this code currently works, suggest that this is always true.) PiperOrigin-RevId: 351592344 -- f40c3b43ca5b1d7e23cd45f1ffac1783105ac1a3 by Abseil Team <absl-team@google.com>: Revert 18abb2902b9f06c63a968b24d3dda785ebf99a22 PiperOrigin-RevId: 351523518 -- 18abb2902b9f06c63a968b24d3dda785ebf99a22 by Abseil Team <absl-team@google.com>: Internal change PiperOrigin-RevId: 351512412 -- 9b881602d45e95e06089792c7627cd56528a255a by Abseil Team <absl-team@google.com>: Keep time's global state in a cacheline-aligned structure. Keeping the global state as separate global variables results in two issues: 1) False sharing with adjacent global data (e.g., cycle clock source), since the global fields are updated every O(10usec). 2) The hot global fields (e.g., seq and samples) can reside on different cache lines. To fix this, simply wrap the global data in a ABSL_CACHE_ALIGNED structure. This is similar to what we do for MutexGlobals. PiperOrigin-RevId: 351389466 GitOrigin-RevId: a0491c8d790972cd80e2d720fe1fdf5f711a6f1a Change-Id: Ie0fa80112043381cd37c84e2ab2b7334839f54b5
4 years ago
new_buffer = new_rep->flat()->Data();
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
CopyToArraySlowPath(new_buffer);
} else {
new_buffer = std::allocator<char>().allocate(total_size);
CopyToArraySlowPath(new_buffer);
new_rep = absl::cord_internal::NewExternalRep(
absl::string_view(new_buffer, total_size), [](absl::string_view s) {
std::allocator<char>().deallocate(const_cast<char*>(s.data()),
s.size());
});
}
Export of internal Abseil changes -- 6e9f93888bbe6718997ed90bbd049f1f3572b066 by Abseil Team <absl-team@google.com>: Fix status to be safe for self move assignment. PiperOrigin-RevId: 346815392 -- 35cae74a977f258e81dfbe925fb5a34cb6632036 by Gennadiy Rozental <rogeeff@google.com>: Eliminate unnecessary access to the global vars. PiperOrigin-RevId: 346777168 -- e7e786c243069060f5d6c1c945decb4b0b83f95b by Andy Getzendanner <durandal@google.com>: Internal change. PiperOrigin-RevId: 346685656 -- 4ccd41c48f1a83cfa20b3ea534f743dd7d788376 by Abseil Team <absl-team@google.com>: Move CordRep Ref() and Unref() logic into cord_internal.h This change moves Ref() and Unref() logic out of cord.cc into cord_internal. The main purpose is to make upcoming ring buffer changes easier to isolate from existing cord.cc code. Notice that this removes the nullptr check from Unref() and now requires it to be non null (which held true most times). We may need to rethink if the 'unref unlikely one' is the common case: are cordreps most likely shared? This may be something between 'root' and non root nodes, i.e., is it more likely for leaf / flat nodes in large cords to be shared than top level cordreps being shared? Vice versa? Benchmarks say that we mostly shouldn't care, the caveat being that atomic ops seem more expensive on upcoming archs (arcadia) so we should error on the side of an extra IsOne() branch saving us single owned atomic ops. PiperOrigin-RevId: 346676121 -- f0babab103b9e60d61ba09482d468985e43eceb3 by Samuel Benzaquen <sbenza@google.com>: Fix iterator based constructor and `.insert` members to only require EmplaceConstructible as the standard specifies. PiperOrigin-RevId: 346616707 -- 8f48eedda02277f9c96a88ed7726e34b557cce20 by Evan Brown <ezb@google.com>: Fix a bug in binary_search_impl when there's a transparent, three-way comparator that has different equivalence classes for different lookup types. Add a new can_have_multiple_equivalent_keys method to share the common logic for these cases. PiperOrigin-RevId: 346605948 -- 649183cb3cc9383431de9c81fb1c0f885d4001ae by Abseil Team <absl-team@google.com>: Add benchmark for accessing a Duration flag. PiperOrigin-RevId: 346594843 -- fefdb046520871af63ce2229e2f7cccfc0483dea by Abseil Team <absl-team@google.com>: Restructure CordReader for upcoming ring buffer changes. PiperOrigin-RevId: 346410642 -- 8b2f50e7da0ebab06ead5f94e366e984ca23cb6a by Abseil Team <absl-team@google.com>: Wire in an internal-only flag to toggle upcoming ring buffer changes on/off for experimentation. PiperOrigin-RevId: 346199111 GitOrigin-RevId: 6e9f93888bbe6718997ed90bbd049f1f3572b066 Change-Id: I8f34866b25a79209cb5448bbb28dd3044111d2e9
4 years ago
if (CordRep* tree = contents_.tree()) {
CordRep::Unref(tree);
}
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
contents_.set_tree(new_rep);
return absl::string_view(new_buffer, total_size);
}
/* static */ bool Cord::GetFlatAux(CordRep* rep, absl::string_view* fragment) {
assert(rep != nullptr);
if (rep->tag >= FLAT) {
Export of internal Abseil changes -- a0491c8d790972cd80e2d720fe1fdf5f711a6f1a by Greg Falcon <gfalcon@google.com>: Stop directly accessing CordRepFlat data via CordRep::data. The old pattern of access breaks the `CordRep` type abstraction; since `CordRep::data` is not in general guaranteed to contain the chunk's data, we shouldn't access it that way. This incidentally adds an assertion check (via the flat() accessor) that the CordRep is indeed flat on each such access, but a manual inspection of the code, as well as the fact that this code currently works, suggest that this is always true.) PiperOrigin-RevId: 351592344 -- f40c3b43ca5b1d7e23cd45f1ffac1783105ac1a3 by Abseil Team <absl-team@google.com>: Revert 18abb2902b9f06c63a968b24d3dda785ebf99a22 PiperOrigin-RevId: 351523518 -- 18abb2902b9f06c63a968b24d3dda785ebf99a22 by Abseil Team <absl-team@google.com>: Internal change PiperOrigin-RevId: 351512412 -- 9b881602d45e95e06089792c7627cd56528a255a by Abseil Team <absl-team@google.com>: Keep time's global state in a cacheline-aligned structure. Keeping the global state as separate global variables results in two issues: 1) False sharing with adjacent global data (e.g., cycle clock source), since the global fields are updated every O(10usec). 2) The hot global fields (e.g., seq and samples) can reside on different cache lines. To fix this, simply wrap the global data in a ABSL_CACHE_ALIGNED structure. This is similar to what we do for MutexGlobals. PiperOrigin-RevId: 351389466 GitOrigin-RevId: a0491c8d790972cd80e2d720fe1fdf5f711a6f1a Change-Id: Ie0fa80112043381cd37c84e2ab2b7334839f54b5
4 years ago
*fragment = absl::string_view(rep->flat()->Data(), rep->length);
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
return true;
} else if (rep->tag == EXTERNAL) {
*fragment = absl::string_view(rep->external()->base, rep->length);
return true;
} else if (rep->tag == RING) {
return rep->ring()->IsFlat(fragment);
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
} else if (rep->tag == SUBSTRING) {
CordRep* child = rep->substring()->child;
if (child->tag >= FLAT) {
Export of internal Abseil changes -- a0491c8d790972cd80e2d720fe1fdf5f711a6f1a by Greg Falcon <gfalcon@google.com>: Stop directly accessing CordRepFlat data via CordRep::data. The old pattern of access breaks the `CordRep` type abstraction; since `CordRep::data` is not in general guaranteed to contain the chunk's data, we shouldn't access it that way. This incidentally adds an assertion check (via the flat() accessor) that the CordRep is indeed flat on each such access, but a manual inspection of the code, as well as the fact that this code currently works, suggest that this is always true.) PiperOrigin-RevId: 351592344 -- f40c3b43ca5b1d7e23cd45f1ffac1783105ac1a3 by Abseil Team <absl-team@google.com>: Revert 18abb2902b9f06c63a968b24d3dda785ebf99a22 PiperOrigin-RevId: 351523518 -- 18abb2902b9f06c63a968b24d3dda785ebf99a22 by Abseil Team <absl-team@google.com>: Internal change PiperOrigin-RevId: 351512412 -- 9b881602d45e95e06089792c7627cd56528a255a by Abseil Team <absl-team@google.com>: Keep time's global state in a cacheline-aligned structure. Keeping the global state as separate global variables results in two issues: 1) False sharing with adjacent global data (e.g., cycle clock source), since the global fields are updated every O(10usec). 2) The hot global fields (e.g., seq and samples) can reside on different cache lines. To fix this, simply wrap the global data in a ABSL_CACHE_ALIGNED structure. This is similar to what we do for MutexGlobals. PiperOrigin-RevId: 351389466 GitOrigin-RevId: a0491c8d790972cd80e2d720fe1fdf5f711a6f1a Change-Id: Ie0fa80112043381cd37c84e2ab2b7334839f54b5
4 years ago
*fragment = absl::string_view(
child->flat()->Data() + rep->substring()->start, rep->length);
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
return true;
} else if (child->tag == EXTERNAL) {
*fragment = absl::string_view(
child->external()->base + rep->substring()->start, rep->length);
return true;
} else if (child->tag == RING) {
return child->ring()->IsFlat(rep->substring()->start, rep->length,
fragment);
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
}
}
return false;
}
/* static */ void Cord::ForEachChunkAux(
absl::cord_internal::CordRep* rep,
absl::FunctionRef<void(absl::string_view)> callback) {
Export of internal Abseil changes -- c68f1886f5e8fd90eb0c2d2e68feaf00a7cdacda by CJ Johnson <johnsoncj@google.com>: Introduce absl::Cleanup to the OSS repo PiperOrigin-RevId: 354583156 -- 17030cf388e10f7eb959e3e566326d1072ce392e by Abseil Team <absl-team@google.com>: Internal change only PiperOrigin-RevId: 354574953 -- e979d7236d4f3252e79ddda6739b67a9a326bf6d by CJ Johnson <johnsoncj@google.com>: Internal change PiperOrigin-RevId: 354545297 -- 7ea02b3783f7f49ef97d86a8f6580a19cc57df14 by Abseil Team <absl-team@google.com>: Pre-allocate memory for vectors where the size is known. PiperOrigin-RevId: 354344576 -- 9246c7cb11f1d6444f79ebe25acc69a8a9b870e0 by Matt Kulukundis <kfm@google.com>: Add support for Elbrus 2000 (e2k) Import of https://github.com/abseil/abseil-cpp/pull/889 PiperOrigin-RevId: 354344013 -- 0fc93d359cc1fb307552e917b37b7b2e7eed822f by Abseil Team <absl-team@google.com>: Integrate CordRepRing logic into cord (but do not enable it) PiperOrigin-RevId: 354312238 -- eda05622f7da71466723acb33403f783529df24b by Abseil Team <absl-team@google.com>: Protect ignore diagnostic with "__has_warning". PiperOrigin-RevId: 354112334 -- 47716c5d8fb10efa4fdd801d28bac414c6f8ec32 by Abseil Team <absl-team@google.com>: Rearrange InlinedVector copy constructor and destructor to treat a few special cases inline and then tail-call a non-inlined routine for the rest. In particular, we optimize for empty vectors in both cases. Added a couple of benchmarks that copy either an InlVec<int64> or an InlVec<InlVec<int64>>. Speed difference: ``` BM_CopyTrivial/0 0.92ns +- 0% 0.47ns +- 0% -48.91% (p=0.000 n=11+12) BM_CopyTrivial/1 0.92ns +- 0% 1.15ns +- 0% +25.00% (p=0.000 n=10+9) BM_CopyTrivial/8 8.57ns +- 0% 10.72ns +- 1% +25.16% (p=0.000 n=10+12) BM_CopyNonTrivial/0 3.21ns +- 0% 0.70ns +- 0% -78.23% (p=0.000 n=12+10) BM_CopyNonTrivial/1 5.88ns +- 1% 5.51ns +- 0% -6.28% (p=0.000 n=10+8) BM_CopyNonTrivial/8 21.5ns +- 1% 15.2ns +- 2% -29.23% (p=0.000 n=12+12) ``` Note: the slowdowns are a few cycles which is expected given the procedure call added in that case. We decided this is a good tradeoff given the code size reductions and the more significant speedups for empty vectors. Size difference (as measured by nm): ``` BM_CopyTrivial from 1048 bytes to 326 bytes. BM_CopyNonTrivial from 749 bytes to 470 bytes. ``` Code size for a large binary drops by ~500KB (from 349415719 to 348906015 348906191). All of the benchmarks that showed a significant difference: Ones that improve with this CL: ``` BM_CopyNonTrivial/0 3.21ns +- 0% 0.70ns +- 0% -78.23% (p=0.000 n=12+10) BM_InlinedVectorFillString/0 0.93ns +- 0% 0.24ns +- 0% -74.19% (p=0.000 n=12+10) BM_InlinedVectorAssignments/1 10.5ns +- 0% 4.1ns +- 0% -60.64% (p=0.000 n=11+10) BM_InlinedVectorAssignments/2 10.7ns +- 0% 4.4ns +- 0% -59.08% (p=0.000 n=11+11) BM_CopyTrivial/0 0.92ns +- 0% 0.47ns +- 0% -48.91% (p=0.000 n=11+12) BM_CopyNonTrivial/8 21.5ns +- 1% 15.2ns +- 2% -29.23% (p=0.000 n=12+12) BM_StdVectorEmpty 0.47ns +- 1% 0.35ns +- 0% -24.73% (p=0.000 n=12+12) BM_StdVectorSize 0.46ns +- 2% 0.35ns +- 0% -24.32% (p=0.000 n=12+12) BM_SwapElements<LargeCopyableOnly>/0 3.44ns +- 0% 2.76ns +- 1% -19.83% (p=0.000 n=11+11) BM_InlinedVectorFillRange/256 20.7ns +- 1% 17.8ns +- 0% -14.08% (p=0.000 n=12+9) BM_CopyNonTrivial/1 5.88ns +- 1% 5.51ns +- 0% -6.28% (p=0.000 n=10+8) BM_SwapElements<LargeCopyableMovable>/1 4.19ns +- 0% 3.95ns +- 1% -5.63% (p=0.000 n=11+12) BM_SwapElements<LargeCopyableMovableSwappable>/1 4.18ns +- 0% 3.99ns +- 0% -4.70% (p=0.000 n=9+11) BM_SwapElements<LargeCopyableMovable>/0 2.41ns +- 0% 2.31ns +- 0% -4.45% (p=0.000 n=12+12) BM_InlinedVectorFillRange/64 8.25ns +- 0% 8.04ns +- 0% -2.51% (p=0.000 n=12+11) BM_SwapElements<LargeCopyableOnly>/1 82.4ns +- 0% 81.5ns +- 0% -1.06% (p=0.000 n=12+12) ``` Ones that get worse with this CL: ``` BM_CopyTrivial/1 0.92ns +- 0% 1.15ns +- 0% +25.00% (p=0.000 n=10+9) BM_CopyTrivial/8 8.57ns +- 0% 10.72ns +- 1% +25.16% (p=0.000 n=10+12) BM_SwapElements<LargeCopyableMovableSwappable>/512 1.48ns +- 1% 1.66ns +- 1% +11.88% (p=0.000 n=12+12) BM_InlinedVectorFillString/1 11.5ns +- 0% 12.8ns +- 1% +11.62% (p=0.000 n=12+11) BM_SwapElements<LargeCopyableMovableSwappable>/64 1.48ns +- 2% 1.66ns +- 1% +11.66% (p=0.000 n=12+11) BM_SwapElements<LargeCopyableMovableSwappable>/1k 1.48ns +- 1% 1.65ns +- 2% +11.32% (p=0.000 n=12+12) BM_SwapElements<LargeCopyableMovable>/512 1.48ns +- 2% 1.58ns +- 4% +6.62% (p=0.000 n=11+12) BM_SwapElements<LargeCopyableMovable>/1k 1.49ns +- 2% 1.58ns +- 3% +6.05% (p=0.000 n=12+12) BM_SwapElements<LargeCopyableMovable>/64 1.48ns +- 2% 1.57ns +- 4% +6.04% (p=0.000 n=11+12) BM_InlinedVectorFillRange/1 4.81ns +- 0% 5.05ns +- 0% +4.83% (p=0.000 n=11+11) BM_InlinedVectorFillString/8 79.4ns +- 1% 83.1ns +- 1% +4.64% (p=0.000 n=10+12) BM_StdVectorFillString/1 16.3ns +- 0% 16.6ns +- 0% +2.13% (p=0.000 n=11+8) ``` PiperOrigin-RevId: 353906786 -- 8e26518b3cec9c598e5e9573c46c3bd1b03a67ef by Abseil Team <absl-team@google.com>: Internal change PiperOrigin-RevId: 353737330 -- f206ae0983e58c9904ed8b8f05f9caf564a446be by Matt Kulukundis <kfm@google.com>: Import of CCTZ from GitHub. PiperOrigin-RevId: 353682256 GitOrigin-RevId: c68f1886f5e8fd90eb0c2d2e68feaf00a7cdacda Change-Id: I5790c1036c4f543c701d1039848fabf7ae881ad8
4 years ago
if (rep->tag == RING) {
ChunkIterator it(rep), end;
while (it != end) {
callback(*it);
++it;
}
return;
}
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
assert(rep != nullptr);
int stack_pos = 0;
constexpr int stack_max = 128;
// Stack of right branches for tree traversal
absl::cord_internal::CordRep* stack[stack_max];
absl::cord_internal::CordRep* current_node = rep;
while (true) {
if (current_node->tag == CONCAT) {
if (stack_pos == stack_max) {
// There's no more room on our stack array to add another right branch,
// and the idea is to avoid allocations, so call this function
// recursively to navigate this subtree further. (This is not something
// we expect to happen in practice).
ForEachChunkAux(current_node, callback);
// Pop the next right branch and iterate.
current_node = stack[--stack_pos];
continue;
} else {
// Save the right branch for later traversal and continue down the left
// branch.
stack[stack_pos++] = current_node->concat()->right;
current_node = current_node->concat()->left;
continue;
}
}
// This is a leaf node, so invoke our callback.
absl::string_view chunk;
bool success = GetFlatAux(current_node, &chunk);
assert(success);
if (success) {
callback(chunk);
}
if (stack_pos == 0) {
// end of traversal
return;
}
current_node = stack[--stack_pos];
}
}
Export of internal Abseil changes -- c68f1886f5e8fd90eb0c2d2e68feaf00a7cdacda by CJ Johnson <johnsoncj@google.com>: Introduce absl::Cleanup to the OSS repo PiperOrigin-RevId: 354583156 -- 17030cf388e10f7eb959e3e566326d1072ce392e by Abseil Team <absl-team@google.com>: Internal change only PiperOrigin-RevId: 354574953 -- e979d7236d4f3252e79ddda6739b67a9a326bf6d by CJ Johnson <johnsoncj@google.com>: Internal change PiperOrigin-RevId: 354545297 -- 7ea02b3783f7f49ef97d86a8f6580a19cc57df14 by Abseil Team <absl-team@google.com>: Pre-allocate memory for vectors where the size is known. PiperOrigin-RevId: 354344576 -- 9246c7cb11f1d6444f79ebe25acc69a8a9b870e0 by Matt Kulukundis <kfm@google.com>: Add support for Elbrus 2000 (e2k) Import of https://github.com/abseil/abseil-cpp/pull/889 PiperOrigin-RevId: 354344013 -- 0fc93d359cc1fb307552e917b37b7b2e7eed822f by Abseil Team <absl-team@google.com>: Integrate CordRepRing logic into cord (but do not enable it) PiperOrigin-RevId: 354312238 -- eda05622f7da71466723acb33403f783529df24b by Abseil Team <absl-team@google.com>: Protect ignore diagnostic with "__has_warning". PiperOrigin-RevId: 354112334 -- 47716c5d8fb10efa4fdd801d28bac414c6f8ec32 by Abseil Team <absl-team@google.com>: Rearrange InlinedVector copy constructor and destructor to treat a few special cases inline and then tail-call a non-inlined routine for the rest. In particular, we optimize for empty vectors in both cases. Added a couple of benchmarks that copy either an InlVec<int64> or an InlVec<InlVec<int64>>. Speed difference: ``` BM_CopyTrivial/0 0.92ns +- 0% 0.47ns +- 0% -48.91% (p=0.000 n=11+12) BM_CopyTrivial/1 0.92ns +- 0% 1.15ns +- 0% +25.00% (p=0.000 n=10+9) BM_CopyTrivial/8 8.57ns +- 0% 10.72ns +- 1% +25.16% (p=0.000 n=10+12) BM_CopyNonTrivial/0 3.21ns +- 0% 0.70ns +- 0% -78.23% (p=0.000 n=12+10) BM_CopyNonTrivial/1 5.88ns +- 1% 5.51ns +- 0% -6.28% (p=0.000 n=10+8) BM_CopyNonTrivial/8 21.5ns +- 1% 15.2ns +- 2% -29.23% (p=0.000 n=12+12) ``` Note: the slowdowns are a few cycles which is expected given the procedure call added in that case. We decided this is a good tradeoff given the code size reductions and the more significant speedups for empty vectors. Size difference (as measured by nm): ``` BM_CopyTrivial from 1048 bytes to 326 bytes. BM_CopyNonTrivial from 749 bytes to 470 bytes. ``` Code size for a large binary drops by ~500KB (from 349415719 to 348906015 348906191). All of the benchmarks that showed a significant difference: Ones that improve with this CL: ``` BM_CopyNonTrivial/0 3.21ns +- 0% 0.70ns +- 0% -78.23% (p=0.000 n=12+10) BM_InlinedVectorFillString/0 0.93ns +- 0% 0.24ns +- 0% -74.19% (p=0.000 n=12+10) BM_InlinedVectorAssignments/1 10.5ns +- 0% 4.1ns +- 0% -60.64% (p=0.000 n=11+10) BM_InlinedVectorAssignments/2 10.7ns +- 0% 4.4ns +- 0% -59.08% (p=0.000 n=11+11) BM_CopyTrivial/0 0.92ns +- 0% 0.47ns +- 0% -48.91% (p=0.000 n=11+12) BM_CopyNonTrivial/8 21.5ns +- 1% 15.2ns +- 2% -29.23% (p=0.000 n=12+12) BM_StdVectorEmpty 0.47ns +- 1% 0.35ns +- 0% -24.73% (p=0.000 n=12+12) BM_StdVectorSize 0.46ns +- 2% 0.35ns +- 0% -24.32% (p=0.000 n=12+12) BM_SwapElements<LargeCopyableOnly>/0 3.44ns +- 0% 2.76ns +- 1% -19.83% (p=0.000 n=11+11) BM_InlinedVectorFillRange/256 20.7ns +- 1% 17.8ns +- 0% -14.08% (p=0.000 n=12+9) BM_CopyNonTrivial/1 5.88ns +- 1% 5.51ns +- 0% -6.28% (p=0.000 n=10+8) BM_SwapElements<LargeCopyableMovable>/1 4.19ns +- 0% 3.95ns +- 1% -5.63% (p=0.000 n=11+12) BM_SwapElements<LargeCopyableMovableSwappable>/1 4.18ns +- 0% 3.99ns +- 0% -4.70% (p=0.000 n=9+11) BM_SwapElements<LargeCopyableMovable>/0 2.41ns +- 0% 2.31ns +- 0% -4.45% (p=0.000 n=12+12) BM_InlinedVectorFillRange/64 8.25ns +- 0% 8.04ns +- 0% -2.51% (p=0.000 n=12+11) BM_SwapElements<LargeCopyableOnly>/1 82.4ns +- 0% 81.5ns +- 0% -1.06% (p=0.000 n=12+12) ``` Ones that get worse with this CL: ``` BM_CopyTrivial/1 0.92ns +- 0% 1.15ns +- 0% +25.00% (p=0.000 n=10+9) BM_CopyTrivial/8 8.57ns +- 0% 10.72ns +- 1% +25.16% (p=0.000 n=10+12) BM_SwapElements<LargeCopyableMovableSwappable>/512 1.48ns +- 1% 1.66ns +- 1% +11.88% (p=0.000 n=12+12) BM_InlinedVectorFillString/1 11.5ns +- 0% 12.8ns +- 1% +11.62% (p=0.000 n=12+11) BM_SwapElements<LargeCopyableMovableSwappable>/64 1.48ns +- 2% 1.66ns +- 1% +11.66% (p=0.000 n=12+11) BM_SwapElements<LargeCopyableMovableSwappable>/1k 1.48ns +- 1% 1.65ns +- 2% +11.32% (p=0.000 n=12+12) BM_SwapElements<LargeCopyableMovable>/512 1.48ns +- 2% 1.58ns +- 4% +6.62% (p=0.000 n=11+12) BM_SwapElements<LargeCopyableMovable>/1k 1.49ns +- 2% 1.58ns +- 3% +6.05% (p=0.000 n=12+12) BM_SwapElements<LargeCopyableMovable>/64 1.48ns +- 2% 1.57ns +- 4% +6.04% (p=0.000 n=11+12) BM_InlinedVectorFillRange/1 4.81ns +- 0% 5.05ns +- 0% +4.83% (p=0.000 n=11+11) BM_InlinedVectorFillString/8 79.4ns +- 1% 83.1ns +- 1% +4.64% (p=0.000 n=10+12) BM_StdVectorFillString/1 16.3ns +- 0% 16.6ns +- 0% +2.13% (p=0.000 n=11+8) ``` PiperOrigin-RevId: 353906786 -- 8e26518b3cec9c598e5e9573c46c3bd1b03a67ef by Abseil Team <absl-team@google.com>: Internal change PiperOrigin-RevId: 353737330 -- f206ae0983e58c9904ed8b8f05f9caf564a446be by Matt Kulukundis <kfm@google.com>: Import of CCTZ from GitHub. PiperOrigin-RevId: 353682256 GitOrigin-RevId: c68f1886f5e8fd90eb0c2d2e68feaf00a7cdacda Change-Id: I5790c1036c4f543c701d1039848fabf7ae881ad8
4 years ago
static void DumpNode(CordRep* rep, bool include_data, std::ostream* os,
int indent) {
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
const int kIndentStep = 1;
absl::InlinedVector<CordRep*, kInlinedVectorSize> stack;
absl::InlinedVector<int, kInlinedVectorSize> indents;
for (;;) {
*os << std::setw(3) << rep->refcount.Get();
*os << " " << std::setw(7) << rep->length;
*os << " [";
if (include_data) *os << static_cast<void*>(rep);
*os << "]";
*os << " " << (IsRootBalanced(rep) ? 'b' : 'u');
*os << " " << std::setw(indent) << "";
if (rep->tag == CONCAT) {
*os << "CONCAT depth=" << Depth(rep) << "\n";
indent += kIndentStep;
indents.push_back(indent);
stack.push_back(rep->concat()->right);
rep = rep->concat()->left;
} else if (rep->tag == SUBSTRING) {
*os << "SUBSTRING @ " << rep->substring()->start << "\n";
indent += kIndentStep;
rep = rep->substring()->child;
Export of internal Abseil changes -- c68f1886f5e8fd90eb0c2d2e68feaf00a7cdacda by CJ Johnson <johnsoncj@google.com>: Introduce absl::Cleanup to the OSS repo PiperOrigin-RevId: 354583156 -- 17030cf388e10f7eb959e3e566326d1072ce392e by Abseil Team <absl-team@google.com>: Internal change only PiperOrigin-RevId: 354574953 -- e979d7236d4f3252e79ddda6739b67a9a326bf6d by CJ Johnson <johnsoncj@google.com>: Internal change PiperOrigin-RevId: 354545297 -- 7ea02b3783f7f49ef97d86a8f6580a19cc57df14 by Abseil Team <absl-team@google.com>: Pre-allocate memory for vectors where the size is known. PiperOrigin-RevId: 354344576 -- 9246c7cb11f1d6444f79ebe25acc69a8a9b870e0 by Matt Kulukundis <kfm@google.com>: Add support for Elbrus 2000 (e2k) Import of https://github.com/abseil/abseil-cpp/pull/889 PiperOrigin-RevId: 354344013 -- 0fc93d359cc1fb307552e917b37b7b2e7eed822f by Abseil Team <absl-team@google.com>: Integrate CordRepRing logic into cord (but do not enable it) PiperOrigin-RevId: 354312238 -- eda05622f7da71466723acb33403f783529df24b by Abseil Team <absl-team@google.com>: Protect ignore diagnostic with "__has_warning". PiperOrigin-RevId: 354112334 -- 47716c5d8fb10efa4fdd801d28bac414c6f8ec32 by Abseil Team <absl-team@google.com>: Rearrange InlinedVector copy constructor and destructor to treat a few special cases inline and then tail-call a non-inlined routine for the rest. In particular, we optimize for empty vectors in both cases. Added a couple of benchmarks that copy either an InlVec<int64> or an InlVec<InlVec<int64>>. Speed difference: ``` BM_CopyTrivial/0 0.92ns +- 0% 0.47ns +- 0% -48.91% (p=0.000 n=11+12) BM_CopyTrivial/1 0.92ns +- 0% 1.15ns +- 0% +25.00% (p=0.000 n=10+9) BM_CopyTrivial/8 8.57ns +- 0% 10.72ns +- 1% +25.16% (p=0.000 n=10+12) BM_CopyNonTrivial/0 3.21ns +- 0% 0.70ns +- 0% -78.23% (p=0.000 n=12+10) BM_CopyNonTrivial/1 5.88ns +- 1% 5.51ns +- 0% -6.28% (p=0.000 n=10+8) BM_CopyNonTrivial/8 21.5ns +- 1% 15.2ns +- 2% -29.23% (p=0.000 n=12+12) ``` Note: the slowdowns are a few cycles which is expected given the procedure call added in that case. We decided this is a good tradeoff given the code size reductions and the more significant speedups for empty vectors. Size difference (as measured by nm): ``` BM_CopyTrivial from 1048 bytes to 326 bytes. BM_CopyNonTrivial from 749 bytes to 470 bytes. ``` Code size for a large binary drops by ~500KB (from 349415719 to 348906015 348906191). All of the benchmarks that showed a significant difference: Ones that improve with this CL: ``` BM_CopyNonTrivial/0 3.21ns +- 0% 0.70ns +- 0% -78.23% (p=0.000 n=12+10) BM_InlinedVectorFillString/0 0.93ns +- 0% 0.24ns +- 0% -74.19% (p=0.000 n=12+10) BM_InlinedVectorAssignments/1 10.5ns +- 0% 4.1ns +- 0% -60.64% (p=0.000 n=11+10) BM_InlinedVectorAssignments/2 10.7ns +- 0% 4.4ns +- 0% -59.08% (p=0.000 n=11+11) BM_CopyTrivial/0 0.92ns +- 0% 0.47ns +- 0% -48.91% (p=0.000 n=11+12) BM_CopyNonTrivial/8 21.5ns +- 1% 15.2ns +- 2% -29.23% (p=0.000 n=12+12) BM_StdVectorEmpty 0.47ns +- 1% 0.35ns +- 0% -24.73% (p=0.000 n=12+12) BM_StdVectorSize 0.46ns +- 2% 0.35ns +- 0% -24.32% (p=0.000 n=12+12) BM_SwapElements<LargeCopyableOnly>/0 3.44ns +- 0% 2.76ns +- 1% -19.83% (p=0.000 n=11+11) BM_InlinedVectorFillRange/256 20.7ns +- 1% 17.8ns +- 0% -14.08% (p=0.000 n=12+9) BM_CopyNonTrivial/1 5.88ns +- 1% 5.51ns +- 0% -6.28% (p=0.000 n=10+8) BM_SwapElements<LargeCopyableMovable>/1 4.19ns +- 0% 3.95ns +- 1% -5.63% (p=0.000 n=11+12) BM_SwapElements<LargeCopyableMovableSwappable>/1 4.18ns +- 0% 3.99ns +- 0% -4.70% (p=0.000 n=9+11) BM_SwapElements<LargeCopyableMovable>/0 2.41ns +- 0% 2.31ns +- 0% -4.45% (p=0.000 n=12+12) BM_InlinedVectorFillRange/64 8.25ns +- 0% 8.04ns +- 0% -2.51% (p=0.000 n=12+11) BM_SwapElements<LargeCopyableOnly>/1 82.4ns +- 0% 81.5ns +- 0% -1.06% (p=0.000 n=12+12) ``` Ones that get worse with this CL: ``` BM_CopyTrivial/1 0.92ns +- 0% 1.15ns +- 0% +25.00% (p=0.000 n=10+9) BM_CopyTrivial/8 8.57ns +- 0% 10.72ns +- 1% +25.16% (p=0.000 n=10+12) BM_SwapElements<LargeCopyableMovableSwappable>/512 1.48ns +- 1% 1.66ns +- 1% +11.88% (p=0.000 n=12+12) BM_InlinedVectorFillString/1 11.5ns +- 0% 12.8ns +- 1% +11.62% (p=0.000 n=12+11) BM_SwapElements<LargeCopyableMovableSwappable>/64 1.48ns +- 2% 1.66ns +- 1% +11.66% (p=0.000 n=12+11) BM_SwapElements<LargeCopyableMovableSwappable>/1k 1.48ns +- 1% 1.65ns +- 2% +11.32% (p=0.000 n=12+12) BM_SwapElements<LargeCopyableMovable>/512 1.48ns +- 2% 1.58ns +- 4% +6.62% (p=0.000 n=11+12) BM_SwapElements<LargeCopyableMovable>/1k 1.49ns +- 2% 1.58ns +- 3% +6.05% (p=0.000 n=12+12) BM_SwapElements<LargeCopyableMovable>/64 1.48ns +- 2% 1.57ns +- 4% +6.04% (p=0.000 n=11+12) BM_InlinedVectorFillRange/1 4.81ns +- 0% 5.05ns +- 0% +4.83% (p=0.000 n=11+11) BM_InlinedVectorFillString/8 79.4ns +- 1% 83.1ns +- 1% +4.64% (p=0.000 n=10+12) BM_StdVectorFillString/1 16.3ns +- 0% 16.6ns +- 0% +2.13% (p=0.000 n=11+8) ``` PiperOrigin-RevId: 353906786 -- 8e26518b3cec9c598e5e9573c46c3bd1b03a67ef by Abseil Team <absl-team@google.com>: Internal change PiperOrigin-RevId: 353737330 -- f206ae0983e58c9904ed8b8f05f9caf564a446be by Matt Kulukundis <kfm@google.com>: Import of CCTZ from GitHub. PiperOrigin-RevId: 353682256 GitOrigin-RevId: c68f1886f5e8fd90eb0c2d2e68feaf00a7cdacda Change-Id: I5790c1036c4f543c701d1039848fabf7ae881ad8
4 years ago
} else { // Leaf or ring
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
if (rep->tag == EXTERNAL) {
*os << "EXTERNAL [";
if (include_data)
*os << absl::CEscape(std::string(rep->external()->base, rep->length));
*os << "]\n";
Export of internal Abseil changes -- c68f1886f5e8fd90eb0c2d2e68feaf00a7cdacda by CJ Johnson <johnsoncj@google.com>: Introduce absl::Cleanup to the OSS repo PiperOrigin-RevId: 354583156 -- 17030cf388e10f7eb959e3e566326d1072ce392e by Abseil Team <absl-team@google.com>: Internal change only PiperOrigin-RevId: 354574953 -- e979d7236d4f3252e79ddda6739b67a9a326bf6d by CJ Johnson <johnsoncj@google.com>: Internal change PiperOrigin-RevId: 354545297 -- 7ea02b3783f7f49ef97d86a8f6580a19cc57df14 by Abseil Team <absl-team@google.com>: Pre-allocate memory for vectors where the size is known. PiperOrigin-RevId: 354344576 -- 9246c7cb11f1d6444f79ebe25acc69a8a9b870e0 by Matt Kulukundis <kfm@google.com>: Add support for Elbrus 2000 (e2k) Import of https://github.com/abseil/abseil-cpp/pull/889 PiperOrigin-RevId: 354344013 -- 0fc93d359cc1fb307552e917b37b7b2e7eed822f by Abseil Team <absl-team@google.com>: Integrate CordRepRing logic into cord (but do not enable it) PiperOrigin-RevId: 354312238 -- eda05622f7da71466723acb33403f783529df24b by Abseil Team <absl-team@google.com>: Protect ignore diagnostic with "__has_warning". PiperOrigin-RevId: 354112334 -- 47716c5d8fb10efa4fdd801d28bac414c6f8ec32 by Abseil Team <absl-team@google.com>: Rearrange InlinedVector copy constructor and destructor to treat a few special cases inline and then tail-call a non-inlined routine for the rest. In particular, we optimize for empty vectors in both cases. Added a couple of benchmarks that copy either an InlVec<int64> or an InlVec<InlVec<int64>>. Speed difference: ``` BM_CopyTrivial/0 0.92ns +- 0% 0.47ns +- 0% -48.91% (p=0.000 n=11+12) BM_CopyTrivial/1 0.92ns +- 0% 1.15ns +- 0% +25.00% (p=0.000 n=10+9) BM_CopyTrivial/8 8.57ns +- 0% 10.72ns +- 1% +25.16% (p=0.000 n=10+12) BM_CopyNonTrivial/0 3.21ns +- 0% 0.70ns +- 0% -78.23% (p=0.000 n=12+10) BM_CopyNonTrivial/1 5.88ns +- 1% 5.51ns +- 0% -6.28% (p=0.000 n=10+8) BM_CopyNonTrivial/8 21.5ns +- 1% 15.2ns +- 2% -29.23% (p=0.000 n=12+12) ``` Note: the slowdowns are a few cycles which is expected given the procedure call added in that case. We decided this is a good tradeoff given the code size reductions and the more significant speedups for empty vectors. Size difference (as measured by nm): ``` BM_CopyTrivial from 1048 bytes to 326 bytes. BM_CopyNonTrivial from 749 bytes to 470 bytes. ``` Code size for a large binary drops by ~500KB (from 349415719 to 348906015 348906191). All of the benchmarks that showed a significant difference: Ones that improve with this CL: ``` BM_CopyNonTrivial/0 3.21ns +- 0% 0.70ns +- 0% -78.23% (p=0.000 n=12+10) BM_InlinedVectorFillString/0 0.93ns +- 0% 0.24ns +- 0% -74.19% (p=0.000 n=12+10) BM_InlinedVectorAssignments/1 10.5ns +- 0% 4.1ns +- 0% -60.64% (p=0.000 n=11+10) BM_InlinedVectorAssignments/2 10.7ns +- 0% 4.4ns +- 0% -59.08% (p=0.000 n=11+11) BM_CopyTrivial/0 0.92ns +- 0% 0.47ns +- 0% -48.91% (p=0.000 n=11+12) BM_CopyNonTrivial/8 21.5ns +- 1% 15.2ns +- 2% -29.23% (p=0.000 n=12+12) BM_StdVectorEmpty 0.47ns +- 1% 0.35ns +- 0% -24.73% (p=0.000 n=12+12) BM_StdVectorSize 0.46ns +- 2% 0.35ns +- 0% -24.32% (p=0.000 n=12+12) BM_SwapElements<LargeCopyableOnly>/0 3.44ns +- 0% 2.76ns +- 1% -19.83% (p=0.000 n=11+11) BM_InlinedVectorFillRange/256 20.7ns +- 1% 17.8ns +- 0% -14.08% (p=0.000 n=12+9) BM_CopyNonTrivial/1 5.88ns +- 1% 5.51ns +- 0% -6.28% (p=0.000 n=10+8) BM_SwapElements<LargeCopyableMovable>/1 4.19ns +- 0% 3.95ns +- 1% -5.63% (p=0.000 n=11+12) BM_SwapElements<LargeCopyableMovableSwappable>/1 4.18ns +- 0% 3.99ns +- 0% -4.70% (p=0.000 n=9+11) BM_SwapElements<LargeCopyableMovable>/0 2.41ns +- 0% 2.31ns +- 0% -4.45% (p=0.000 n=12+12) BM_InlinedVectorFillRange/64 8.25ns +- 0% 8.04ns +- 0% -2.51% (p=0.000 n=12+11) BM_SwapElements<LargeCopyableOnly>/1 82.4ns +- 0% 81.5ns +- 0% -1.06% (p=0.000 n=12+12) ``` Ones that get worse with this CL: ``` BM_CopyTrivial/1 0.92ns +- 0% 1.15ns +- 0% +25.00% (p=0.000 n=10+9) BM_CopyTrivial/8 8.57ns +- 0% 10.72ns +- 1% +25.16% (p=0.000 n=10+12) BM_SwapElements<LargeCopyableMovableSwappable>/512 1.48ns +- 1% 1.66ns +- 1% +11.88% (p=0.000 n=12+12) BM_InlinedVectorFillString/1 11.5ns +- 0% 12.8ns +- 1% +11.62% (p=0.000 n=12+11) BM_SwapElements<LargeCopyableMovableSwappable>/64 1.48ns +- 2% 1.66ns +- 1% +11.66% (p=0.000 n=12+11) BM_SwapElements<LargeCopyableMovableSwappable>/1k 1.48ns +- 1% 1.65ns +- 2% +11.32% (p=0.000 n=12+12) BM_SwapElements<LargeCopyableMovable>/512 1.48ns +- 2% 1.58ns +- 4% +6.62% (p=0.000 n=11+12) BM_SwapElements<LargeCopyableMovable>/1k 1.49ns +- 2% 1.58ns +- 3% +6.05% (p=0.000 n=12+12) BM_SwapElements<LargeCopyableMovable>/64 1.48ns +- 2% 1.57ns +- 4% +6.04% (p=0.000 n=11+12) BM_InlinedVectorFillRange/1 4.81ns +- 0% 5.05ns +- 0% +4.83% (p=0.000 n=11+11) BM_InlinedVectorFillString/8 79.4ns +- 1% 83.1ns +- 1% +4.64% (p=0.000 n=10+12) BM_StdVectorFillString/1 16.3ns +- 0% 16.6ns +- 0% +2.13% (p=0.000 n=11+8) ``` PiperOrigin-RevId: 353906786 -- 8e26518b3cec9c598e5e9573c46c3bd1b03a67ef by Abseil Team <absl-team@google.com>: Internal change PiperOrigin-RevId: 353737330 -- f206ae0983e58c9904ed8b8f05f9caf564a446be by Matt Kulukundis <kfm@google.com>: Import of CCTZ from GitHub. PiperOrigin-RevId: 353682256 GitOrigin-RevId: c68f1886f5e8fd90eb0c2d2e68feaf00a7cdacda Change-Id: I5790c1036c4f543c701d1039848fabf7ae881ad8
4 years ago
} else if (rep->tag >= FLAT) {
Export of internal Abseil changes -- ff793052bd01e1e4fcf639f94d7c30c4855a9372 by Evan Brown <ezb@google.com>: Roll forward of btree_iterator refactoring. PiperOrigin-RevId: 346116047 -- 17984679f16e3e2139b0f14fa76f4a6ca16a3ef9 by Chris Kennelly <ckennelly@google.com>: Extend absl::StrContains to accept single character needles. Single characters are more efficient to search for. Extending this API allows the abseil-string-find-str-contains Clang Tidy to include this pattern. The C++ committee has adopted http://wg21.link/P1679 for inclusion in C++23. PiperOrigin-RevId: 346095060 -- ef20b31c501b1dcaa25e244fd8f8aa43dec09bd6 by Jorg Brown <jorg@google.com>: Internal change for cord ring PiperOrigin-RevId: 346087545 -- b70f2c1cb77fc9e733a126e790967d45c5fd1dc7 by Derek Mauro <dmauro@google.com>: Release layout_benchmark PiperOrigin-RevId: 345968909 -- 3a0eda337ee43622f92cfe14c2aa06f72dc71ee5 by Derek Mauro <dmauro@google.com>: Release raw_hash_set_probe_benchmark PiperOrigin-RevId: 345965969 -- abffdb4bb241a2264cb4e73a6262b660bb10447d by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 345733599 -- 7c9e24a71188df945be17fe98f700bdb51f81b16 by Derek Mauro <dmauro@google.com>: Release hash_benchmark PiperOrigin-RevId: 345721635 -- d68f33f17f9a8cd3f6da8eee3870bdb46402cdc8 by Derek Mauro <dmauro@google.com>: Release raw_hash_set_benchmark PiperOrigin-RevId: 345708384 -- 6e6c547d4d1327b226c0ffe8ff34d0aa103ce24b by Abseil Team <absl-team@google.com>: Updates the implementation of InlinedVector to accurately express the value-initialization semantics of the default constructor PiperOrigin-RevId: 345548260 -- 1532424deda97d468444c217cc0fa4614099c7c1 by Evan Brown <ezb@google.com>: Rollback btree_iterator refactoring. PiperOrigin-RevId: 345543900 GitOrigin-RevId: ff793052bd01e1e4fcf639f94d7c30c4855a9372 Change-Id: I719831981fd056de41939f9addfee3d85e3b49b2
4 years ago
*os << "FLAT cap=" << rep->flat()->Capacity()
<< " [";
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
if (include_data)
Export of internal Abseil changes -- a0491c8d790972cd80e2d720fe1fdf5f711a6f1a by Greg Falcon <gfalcon@google.com>: Stop directly accessing CordRepFlat data via CordRep::data. The old pattern of access breaks the `CordRep` type abstraction; since `CordRep::data` is not in general guaranteed to contain the chunk's data, we shouldn't access it that way. This incidentally adds an assertion check (via the flat() accessor) that the CordRep is indeed flat on each such access, but a manual inspection of the code, as well as the fact that this code currently works, suggest that this is always true.) PiperOrigin-RevId: 351592344 -- f40c3b43ca5b1d7e23cd45f1ffac1783105ac1a3 by Abseil Team <absl-team@google.com>: Revert 18abb2902b9f06c63a968b24d3dda785ebf99a22 PiperOrigin-RevId: 351523518 -- 18abb2902b9f06c63a968b24d3dda785ebf99a22 by Abseil Team <absl-team@google.com>: Internal change PiperOrigin-RevId: 351512412 -- 9b881602d45e95e06089792c7627cd56528a255a by Abseil Team <absl-team@google.com>: Keep time's global state in a cacheline-aligned structure. Keeping the global state as separate global variables results in two issues: 1) False sharing with adjacent global data (e.g., cycle clock source), since the global fields are updated every O(10usec). 2) The hot global fields (e.g., seq and samples) can reside on different cache lines. To fix this, simply wrap the global data in a ABSL_CACHE_ALIGNED structure. This is similar to what we do for MutexGlobals. PiperOrigin-RevId: 351389466 GitOrigin-RevId: a0491c8d790972cd80e2d720fe1fdf5f711a6f1a Change-Id: Ie0fa80112043381cd37c84e2ab2b7334839f54b5
4 years ago
*os << absl::CEscape(std::string(rep->flat()->Data(), rep->length));
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
*os << "]\n";
Export of internal Abseil changes -- c68f1886f5e8fd90eb0c2d2e68feaf00a7cdacda by CJ Johnson <johnsoncj@google.com>: Introduce absl::Cleanup to the OSS repo PiperOrigin-RevId: 354583156 -- 17030cf388e10f7eb959e3e566326d1072ce392e by Abseil Team <absl-team@google.com>: Internal change only PiperOrigin-RevId: 354574953 -- e979d7236d4f3252e79ddda6739b67a9a326bf6d by CJ Johnson <johnsoncj@google.com>: Internal change PiperOrigin-RevId: 354545297 -- 7ea02b3783f7f49ef97d86a8f6580a19cc57df14 by Abseil Team <absl-team@google.com>: Pre-allocate memory for vectors where the size is known. PiperOrigin-RevId: 354344576 -- 9246c7cb11f1d6444f79ebe25acc69a8a9b870e0 by Matt Kulukundis <kfm@google.com>: Add support for Elbrus 2000 (e2k) Import of https://github.com/abseil/abseil-cpp/pull/889 PiperOrigin-RevId: 354344013 -- 0fc93d359cc1fb307552e917b37b7b2e7eed822f by Abseil Team <absl-team@google.com>: Integrate CordRepRing logic into cord (but do not enable it) PiperOrigin-RevId: 354312238 -- eda05622f7da71466723acb33403f783529df24b by Abseil Team <absl-team@google.com>: Protect ignore diagnostic with "__has_warning". PiperOrigin-RevId: 354112334 -- 47716c5d8fb10efa4fdd801d28bac414c6f8ec32 by Abseil Team <absl-team@google.com>: Rearrange InlinedVector copy constructor and destructor to treat a few special cases inline and then tail-call a non-inlined routine for the rest. In particular, we optimize for empty vectors in both cases. Added a couple of benchmarks that copy either an InlVec<int64> or an InlVec<InlVec<int64>>. Speed difference: ``` BM_CopyTrivial/0 0.92ns +- 0% 0.47ns +- 0% -48.91% (p=0.000 n=11+12) BM_CopyTrivial/1 0.92ns +- 0% 1.15ns +- 0% +25.00% (p=0.000 n=10+9) BM_CopyTrivial/8 8.57ns +- 0% 10.72ns +- 1% +25.16% (p=0.000 n=10+12) BM_CopyNonTrivial/0 3.21ns +- 0% 0.70ns +- 0% -78.23% (p=0.000 n=12+10) BM_CopyNonTrivial/1 5.88ns +- 1% 5.51ns +- 0% -6.28% (p=0.000 n=10+8) BM_CopyNonTrivial/8 21.5ns +- 1% 15.2ns +- 2% -29.23% (p=0.000 n=12+12) ``` Note: the slowdowns are a few cycles which is expected given the procedure call added in that case. We decided this is a good tradeoff given the code size reductions and the more significant speedups for empty vectors. Size difference (as measured by nm): ``` BM_CopyTrivial from 1048 bytes to 326 bytes. BM_CopyNonTrivial from 749 bytes to 470 bytes. ``` Code size for a large binary drops by ~500KB (from 349415719 to 348906015 348906191). All of the benchmarks that showed a significant difference: Ones that improve with this CL: ``` BM_CopyNonTrivial/0 3.21ns +- 0% 0.70ns +- 0% -78.23% (p=0.000 n=12+10) BM_InlinedVectorFillString/0 0.93ns +- 0% 0.24ns +- 0% -74.19% (p=0.000 n=12+10) BM_InlinedVectorAssignments/1 10.5ns +- 0% 4.1ns +- 0% -60.64% (p=0.000 n=11+10) BM_InlinedVectorAssignments/2 10.7ns +- 0% 4.4ns +- 0% -59.08% (p=0.000 n=11+11) BM_CopyTrivial/0 0.92ns +- 0% 0.47ns +- 0% -48.91% (p=0.000 n=11+12) BM_CopyNonTrivial/8 21.5ns +- 1% 15.2ns +- 2% -29.23% (p=0.000 n=12+12) BM_StdVectorEmpty 0.47ns +- 1% 0.35ns +- 0% -24.73% (p=0.000 n=12+12) BM_StdVectorSize 0.46ns +- 2% 0.35ns +- 0% -24.32% (p=0.000 n=12+12) BM_SwapElements<LargeCopyableOnly>/0 3.44ns +- 0% 2.76ns +- 1% -19.83% (p=0.000 n=11+11) BM_InlinedVectorFillRange/256 20.7ns +- 1% 17.8ns +- 0% -14.08% (p=0.000 n=12+9) BM_CopyNonTrivial/1 5.88ns +- 1% 5.51ns +- 0% -6.28% (p=0.000 n=10+8) BM_SwapElements<LargeCopyableMovable>/1 4.19ns +- 0% 3.95ns +- 1% -5.63% (p=0.000 n=11+12) BM_SwapElements<LargeCopyableMovableSwappable>/1 4.18ns +- 0% 3.99ns +- 0% -4.70% (p=0.000 n=9+11) BM_SwapElements<LargeCopyableMovable>/0 2.41ns +- 0% 2.31ns +- 0% -4.45% (p=0.000 n=12+12) BM_InlinedVectorFillRange/64 8.25ns +- 0% 8.04ns +- 0% -2.51% (p=0.000 n=12+11) BM_SwapElements<LargeCopyableOnly>/1 82.4ns +- 0% 81.5ns +- 0% -1.06% (p=0.000 n=12+12) ``` Ones that get worse with this CL: ``` BM_CopyTrivial/1 0.92ns +- 0% 1.15ns +- 0% +25.00% (p=0.000 n=10+9) BM_CopyTrivial/8 8.57ns +- 0% 10.72ns +- 1% +25.16% (p=0.000 n=10+12) BM_SwapElements<LargeCopyableMovableSwappable>/512 1.48ns +- 1% 1.66ns +- 1% +11.88% (p=0.000 n=12+12) BM_InlinedVectorFillString/1 11.5ns +- 0% 12.8ns +- 1% +11.62% (p=0.000 n=12+11) BM_SwapElements<LargeCopyableMovableSwappable>/64 1.48ns +- 2% 1.66ns +- 1% +11.66% (p=0.000 n=12+11) BM_SwapElements<LargeCopyableMovableSwappable>/1k 1.48ns +- 1% 1.65ns +- 2% +11.32% (p=0.000 n=12+12) BM_SwapElements<LargeCopyableMovable>/512 1.48ns +- 2% 1.58ns +- 4% +6.62% (p=0.000 n=11+12) BM_SwapElements<LargeCopyableMovable>/1k 1.49ns +- 2% 1.58ns +- 3% +6.05% (p=0.000 n=12+12) BM_SwapElements<LargeCopyableMovable>/64 1.48ns +- 2% 1.57ns +- 4% +6.04% (p=0.000 n=11+12) BM_InlinedVectorFillRange/1 4.81ns +- 0% 5.05ns +- 0% +4.83% (p=0.000 n=11+11) BM_InlinedVectorFillString/8 79.4ns +- 1% 83.1ns +- 1% +4.64% (p=0.000 n=10+12) BM_StdVectorFillString/1 16.3ns +- 0% 16.6ns +- 0% +2.13% (p=0.000 n=11+8) ``` PiperOrigin-RevId: 353906786 -- 8e26518b3cec9c598e5e9573c46c3bd1b03a67ef by Abseil Team <absl-team@google.com>: Internal change PiperOrigin-RevId: 353737330 -- f206ae0983e58c9904ed8b8f05f9caf564a446be by Matt Kulukundis <kfm@google.com>: Import of CCTZ from GitHub. PiperOrigin-RevId: 353682256 GitOrigin-RevId: c68f1886f5e8fd90eb0c2d2e68feaf00a7cdacda Change-Id: I5790c1036c4f543c701d1039848fabf7ae881ad8
4 years ago
} else {
assert(rep->tag == RING);
auto* ring = rep->ring();
*os << "RING, entries = " << ring->entries() << "\n";
CordRepRing::index_type head = ring->head();
do {
DumpNode(ring->entry_child(head), include_data, os,
indent + kIndentStep);
head = ring->advance(head);;
} while (head != ring->tail());
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
}
if (stack.empty()) break;
rep = stack.back();
stack.pop_back();
indent = indents.back();
indents.pop_back();
}
}
ABSL_INTERNAL_CHECK(indents.empty(), "");
}
static std::string ReportError(CordRep* root, CordRep* node) {
std::ostringstream buf;
buf << "Error at node " << node << " in:";
DumpNode(root, true, &buf);
return buf.str();
}
static bool VerifyNode(CordRep* root, CordRep* start_node,
bool full_validation) {
absl::InlinedVector<CordRep*, 2> worklist;
worklist.push_back(start_node);
do {
CordRep* node = worklist.back();
worklist.pop_back();
ABSL_INTERNAL_CHECK(node != nullptr, ReportError(root, node));
if (node != root) {
ABSL_INTERNAL_CHECK(node->length != 0, ReportError(root, node));
}
if (node->tag == CONCAT) {
ABSL_INTERNAL_CHECK(node->concat()->left != nullptr,
ReportError(root, node));
ABSL_INTERNAL_CHECK(node->concat()->right != nullptr,
ReportError(root, node));
ABSL_INTERNAL_CHECK((node->length == node->concat()->left->length +
node->concat()->right->length),
ReportError(root, node));
if (full_validation) {
worklist.push_back(node->concat()->right);
worklist.push_back(node->concat()->left);
}
} else if (node->tag >= FLAT) {
ABSL_INTERNAL_CHECK(
Export of internal Abseil changes -- ff793052bd01e1e4fcf639f94d7c30c4855a9372 by Evan Brown <ezb@google.com>: Roll forward of btree_iterator refactoring. PiperOrigin-RevId: 346116047 -- 17984679f16e3e2139b0f14fa76f4a6ca16a3ef9 by Chris Kennelly <ckennelly@google.com>: Extend absl::StrContains to accept single character needles. Single characters are more efficient to search for. Extending this API allows the abseil-string-find-str-contains Clang Tidy to include this pattern. The C++ committee has adopted http://wg21.link/P1679 for inclusion in C++23. PiperOrigin-RevId: 346095060 -- ef20b31c501b1dcaa25e244fd8f8aa43dec09bd6 by Jorg Brown <jorg@google.com>: Internal change for cord ring PiperOrigin-RevId: 346087545 -- b70f2c1cb77fc9e733a126e790967d45c5fd1dc7 by Derek Mauro <dmauro@google.com>: Release layout_benchmark PiperOrigin-RevId: 345968909 -- 3a0eda337ee43622f92cfe14c2aa06f72dc71ee5 by Derek Mauro <dmauro@google.com>: Release raw_hash_set_probe_benchmark PiperOrigin-RevId: 345965969 -- abffdb4bb241a2264cb4e73a6262b660bb10447d by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 345733599 -- 7c9e24a71188df945be17fe98f700bdb51f81b16 by Derek Mauro <dmauro@google.com>: Release hash_benchmark PiperOrigin-RevId: 345721635 -- d68f33f17f9a8cd3f6da8eee3870bdb46402cdc8 by Derek Mauro <dmauro@google.com>: Release raw_hash_set_benchmark PiperOrigin-RevId: 345708384 -- 6e6c547d4d1327b226c0ffe8ff34d0aa103ce24b by Abseil Team <absl-team@google.com>: Updates the implementation of InlinedVector to accurately express the value-initialization semantics of the default constructor PiperOrigin-RevId: 345548260 -- 1532424deda97d468444c217cc0fa4614099c7c1 by Evan Brown <ezb@google.com>: Rollback btree_iterator refactoring. PiperOrigin-RevId: 345543900 GitOrigin-RevId: ff793052bd01e1e4fcf639f94d7c30c4855a9372 Change-Id: I719831981fd056de41939f9addfee3d85e3b49b2
4 years ago
node->length <= node->flat()->Capacity(),
ReportError(root, node));
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
} else if (node->tag == EXTERNAL) {
ABSL_INTERNAL_CHECK(node->external()->base != nullptr,
ReportError(root, node));
} else if (node->tag == SUBSTRING) {
ABSL_INTERNAL_CHECK(
node->substring()->start < node->substring()->child->length,
ReportError(root, node));
ABSL_INTERNAL_CHECK(node->substring()->start + node->length <=
node->substring()->child->length,
ReportError(root, node));
}
} while (!worklist.empty());
return true;
}
// Traverses the tree and computes the total memory allocated.
/* static */ size_t Cord::MemoryUsageAux(const CordRep* rep) {
size_t total_mem_usage = 0;
// Allow a quick exit for the common case that the root is a leaf.
if (RepMemoryUsageLeaf(rep, &total_mem_usage)) {
return total_mem_usage;
}
// Iterate over the tree. cur_node is never a leaf node and leaf nodes will
// never be appended to tree_stack. This reduces overhead from manipulating
// tree_stack.
absl::InlinedVector<const CordRep*, kInlinedVectorSize> tree_stack;
const CordRep* cur_node = rep;
while (true) {
const CordRep* next_node = nullptr;
if (cur_node->tag == CONCAT) {
total_mem_usage += sizeof(CordRepConcat);
const CordRep* left = cur_node->concat()->left;
if (!RepMemoryUsageLeaf(left, &total_mem_usage)) {
next_node = left;
}
const CordRep* right = cur_node->concat()->right;
if (!RepMemoryUsageLeaf(right, &total_mem_usage)) {
if (next_node) {
tree_stack.push_back(next_node);
}
next_node = right;
}
Export of internal Abseil changes -- c68f1886f5e8fd90eb0c2d2e68feaf00a7cdacda by CJ Johnson <johnsoncj@google.com>: Introduce absl::Cleanup to the OSS repo PiperOrigin-RevId: 354583156 -- 17030cf388e10f7eb959e3e566326d1072ce392e by Abseil Team <absl-team@google.com>: Internal change only PiperOrigin-RevId: 354574953 -- e979d7236d4f3252e79ddda6739b67a9a326bf6d by CJ Johnson <johnsoncj@google.com>: Internal change PiperOrigin-RevId: 354545297 -- 7ea02b3783f7f49ef97d86a8f6580a19cc57df14 by Abseil Team <absl-team@google.com>: Pre-allocate memory for vectors where the size is known. PiperOrigin-RevId: 354344576 -- 9246c7cb11f1d6444f79ebe25acc69a8a9b870e0 by Matt Kulukundis <kfm@google.com>: Add support for Elbrus 2000 (e2k) Import of https://github.com/abseil/abseil-cpp/pull/889 PiperOrigin-RevId: 354344013 -- 0fc93d359cc1fb307552e917b37b7b2e7eed822f by Abseil Team <absl-team@google.com>: Integrate CordRepRing logic into cord (but do not enable it) PiperOrigin-RevId: 354312238 -- eda05622f7da71466723acb33403f783529df24b by Abseil Team <absl-team@google.com>: Protect ignore diagnostic with "__has_warning". PiperOrigin-RevId: 354112334 -- 47716c5d8fb10efa4fdd801d28bac414c6f8ec32 by Abseil Team <absl-team@google.com>: Rearrange InlinedVector copy constructor and destructor to treat a few special cases inline and then tail-call a non-inlined routine for the rest. In particular, we optimize for empty vectors in both cases. Added a couple of benchmarks that copy either an InlVec<int64> or an InlVec<InlVec<int64>>. Speed difference: ``` BM_CopyTrivial/0 0.92ns +- 0% 0.47ns +- 0% -48.91% (p=0.000 n=11+12) BM_CopyTrivial/1 0.92ns +- 0% 1.15ns +- 0% +25.00% (p=0.000 n=10+9) BM_CopyTrivial/8 8.57ns +- 0% 10.72ns +- 1% +25.16% (p=0.000 n=10+12) BM_CopyNonTrivial/0 3.21ns +- 0% 0.70ns +- 0% -78.23% (p=0.000 n=12+10) BM_CopyNonTrivial/1 5.88ns +- 1% 5.51ns +- 0% -6.28% (p=0.000 n=10+8) BM_CopyNonTrivial/8 21.5ns +- 1% 15.2ns +- 2% -29.23% (p=0.000 n=12+12) ``` Note: the slowdowns are a few cycles which is expected given the procedure call added in that case. We decided this is a good tradeoff given the code size reductions and the more significant speedups for empty vectors. Size difference (as measured by nm): ``` BM_CopyTrivial from 1048 bytes to 326 bytes. BM_CopyNonTrivial from 749 bytes to 470 bytes. ``` Code size for a large binary drops by ~500KB (from 349415719 to 348906015 348906191). All of the benchmarks that showed a significant difference: Ones that improve with this CL: ``` BM_CopyNonTrivial/0 3.21ns +- 0% 0.70ns +- 0% -78.23% (p=0.000 n=12+10) BM_InlinedVectorFillString/0 0.93ns +- 0% 0.24ns +- 0% -74.19% (p=0.000 n=12+10) BM_InlinedVectorAssignments/1 10.5ns +- 0% 4.1ns +- 0% -60.64% (p=0.000 n=11+10) BM_InlinedVectorAssignments/2 10.7ns +- 0% 4.4ns +- 0% -59.08% (p=0.000 n=11+11) BM_CopyTrivial/0 0.92ns +- 0% 0.47ns +- 0% -48.91% (p=0.000 n=11+12) BM_CopyNonTrivial/8 21.5ns +- 1% 15.2ns +- 2% -29.23% (p=0.000 n=12+12) BM_StdVectorEmpty 0.47ns +- 1% 0.35ns +- 0% -24.73% (p=0.000 n=12+12) BM_StdVectorSize 0.46ns +- 2% 0.35ns +- 0% -24.32% (p=0.000 n=12+12) BM_SwapElements<LargeCopyableOnly>/0 3.44ns +- 0% 2.76ns +- 1% -19.83% (p=0.000 n=11+11) BM_InlinedVectorFillRange/256 20.7ns +- 1% 17.8ns +- 0% -14.08% (p=0.000 n=12+9) BM_CopyNonTrivial/1 5.88ns +- 1% 5.51ns +- 0% -6.28% (p=0.000 n=10+8) BM_SwapElements<LargeCopyableMovable>/1 4.19ns +- 0% 3.95ns +- 1% -5.63% (p=0.000 n=11+12) BM_SwapElements<LargeCopyableMovableSwappable>/1 4.18ns +- 0% 3.99ns +- 0% -4.70% (p=0.000 n=9+11) BM_SwapElements<LargeCopyableMovable>/0 2.41ns +- 0% 2.31ns +- 0% -4.45% (p=0.000 n=12+12) BM_InlinedVectorFillRange/64 8.25ns +- 0% 8.04ns +- 0% -2.51% (p=0.000 n=12+11) BM_SwapElements<LargeCopyableOnly>/1 82.4ns +- 0% 81.5ns +- 0% -1.06% (p=0.000 n=12+12) ``` Ones that get worse with this CL: ``` BM_CopyTrivial/1 0.92ns +- 0% 1.15ns +- 0% +25.00% (p=0.000 n=10+9) BM_CopyTrivial/8 8.57ns +- 0% 10.72ns +- 1% +25.16% (p=0.000 n=10+12) BM_SwapElements<LargeCopyableMovableSwappable>/512 1.48ns +- 1% 1.66ns +- 1% +11.88% (p=0.000 n=12+12) BM_InlinedVectorFillString/1 11.5ns +- 0% 12.8ns +- 1% +11.62% (p=0.000 n=12+11) BM_SwapElements<LargeCopyableMovableSwappable>/64 1.48ns +- 2% 1.66ns +- 1% +11.66% (p=0.000 n=12+11) BM_SwapElements<LargeCopyableMovableSwappable>/1k 1.48ns +- 1% 1.65ns +- 2% +11.32% (p=0.000 n=12+12) BM_SwapElements<LargeCopyableMovable>/512 1.48ns +- 2% 1.58ns +- 4% +6.62% (p=0.000 n=11+12) BM_SwapElements<LargeCopyableMovable>/1k 1.49ns +- 2% 1.58ns +- 3% +6.05% (p=0.000 n=12+12) BM_SwapElements<LargeCopyableMovable>/64 1.48ns +- 2% 1.57ns +- 4% +6.04% (p=0.000 n=11+12) BM_InlinedVectorFillRange/1 4.81ns +- 0% 5.05ns +- 0% +4.83% (p=0.000 n=11+11) BM_InlinedVectorFillString/8 79.4ns +- 1% 83.1ns +- 1% +4.64% (p=0.000 n=10+12) BM_StdVectorFillString/1 16.3ns +- 0% 16.6ns +- 0% +2.13% (p=0.000 n=11+8) ``` PiperOrigin-RevId: 353906786 -- 8e26518b3cec9c598e5e9573c46c3bd1b03a67ef by Abseil Team <absl-team@google.com>: Internal change PiperOrigin-RevId: 353737330 -- f206ae0983e58c9904ed8b8f05f9caf564a446be by Matt Kulukundis <kfm@google.com>: Import of CCTZ from GitHub. PiperOrigin-RevId: 353682256 GitOrigin-RevId: c68f1886f5e8fd90eb0c2d2e68feaf00a7cdacda Change-Id: I5790c1036c4f543c701d1039848fabf7ae881ad8
4 years ago
} else if (cur_node->tag == RING) {
total_mem_usage += CordRepRing::AllocSize(cur_node->ring()->capacity());
const CordRepRing* ring = cur_node->ring();
CordRepRing::index_type pos = ring->head(), tail = ring->tail();
do {
CordRep* node = ring->entry_child(pos);
assert(node->tag >= FLAT || node->tag == EXTERNAL);
RepMemoryUsageLeaf(node, &total_mem_usage);
} while ((pos = ring->advance(pos)) != tail);
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
} else {
// Since cur_node is not a leaf or a concat node it must be a substring.
assert(cur_node->tag == SUBSTRING);
total_mem_usage += sizeof(CordRepSubstring);
next_node = cur_node->substring()->child;
if (RepMemoryUsageLeaf(next_node, &total_mem_usage)) {
next_node = nullptr;
}
}
if (!next_node) {
if (tree_stack.empty()) {
return total_mem_usage;
}
next_node = tree_stack.back();
tree_stack.pop_back();
}
cur_node = next_node;
}
}
std::ostream& operator<<(std::ostream& out, const Cord& cord) {
for (absl::string_view chunk : cord.Chunks()) {
out.write(chunk.data(), chunk.size());
}
return out;
}
namespace strings_internal {
size_t CordTestAccess::FlatOverhead() { return cord_internal::kFlatOverhead; }
size_t CordTestAccess::MaxFlatLength() { return cord_internal::kMaxFlatLength; }
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
size_t CordTestAccess::FlatTagToLength(uint8_t tag) {
return cord_internal::TagToLength(tag);
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
}
uint8_t CordTestAccess::LengthToTag(size_t s) {
ABSL_INTERNAL_CHECK(s <= kMaxFlatLength, absl::StrCat("Invalid length ", s));
return cord_internal::AllocatedSizeToTag(s + cord_internal::kFlatOverhead);
Export of internal Abseil changes -- f012012ef78234a6a4585321b67d7b7c92ebc266 by Laramie Leavitt <lar@google.com>: Slight restructuring of absl/random/internal randen implementation. Convert round-keys.inc into randen_round_keys.cc file. Consistently use a 128-bit pointer type for internal method parameters. This allows simpler pointer arithmetic in C++ & permits removal of some constants and casts. Remove some redundancy in comments & constexpr variables. Specifically, all references to Randen algorithm parameters use RandenTraits; duplication in RandenSlow removed. PiperOrigin-RevId: 312190313 -- dc8b42e054046741e9ed65335bfdface997c6063 by Abseil Team <absl-team@google.com>: Internal change. PiperOrigin-RevId: 312167304 -- f13d248fafaf206492c1362c3574031aea3abaf7 by Matthew Brown <matthewbr@google.com>: Cleanup StrFormat extensions a little. PiperOrigin-RevId: 312166336 -- 9d9117589667afe2332bb7ad42bc967ca7c54502 by Derek Mauro <dmauro@google.com>: Internal change PiperOrigin-RevId: 312105213 -- 9a12b9b3aa0e59b8ee6cf9408ed0029045543a9b by Abseil Team <absl-team@google.com>: Complete IGNORE_TYPE macro renaming. PiperOrigin-RevId: 311999699 -- 64756f20d61021d999bd0d4c15e9ad3857382f57 by Gennadiy Rozental <rogeeff@google.com>: Switch to fixed bytes specific default value. This fixes the Abseil Flags for big endian platforms. PiperOrigin-RevId: 311844448 -- bdbe6b5b29791dbc3816ada1828458b3010ff1e9 by Laramie Leavitt <lar@google.com>: Change many distribution tests to use pcg_engine as a deterministic source of entropy. It's reasonable to test that the BitGen itself has good entropy, however when testing the cross product of all random distributions x all the architecture variations x all submitted changes results in a large number of tests. In order to account for these failures while still using good entropy requires that our allowed sigma need to account for all of these independent tests. Our current sigma values are too restrictive, and we see a lot of failures, so we have to either relax the sigma values or convert some of the statistical tests to use deterministic values. This changelist does the latter. PiperOrigin-RevId: 311840096 GitOrigin-RevId: f012012ef78234a6a4585321b67d7b7c92ebc266 Change-Id: Ic84886f38ff30d7d72c126e9b63c9a61eb729a1a
5 years ago
}
size_t CordTestAccess::SizeofCordRepConcat() { return sizeof(CordRepConcat); }
size_t CordTestAccess::SizeofCordRepExternal() {
return sizeof(CordRepExternal);
}
size_t CordTestAccess::SizeofCordRepSubstring() {
return sizeof(CordRepSubstring);
}
} // namespace strings_internal
ABSL_NAMESPACE_END
} // namespace absl