➕ using Google Benchmark #921

benchmarks/thirdparty/benchmark/src/CMakeLists.txt

@@ -0,0 +1,87 @@
+# Allow the source files to find headers in src/
+include_directories(${PROJECT_SOURCE_DIR}/src)
+
+if (DEFINED BENCHMARK_CXX_LINKER_FLAGS)
+  list(APPEND CMAKE_SHARED_LINKER_FLAGS ${BENCHMARK_CXX_LINKER_FLAGS})
+  list(APPEND CMAKE_MODULE_LINKER_FLAGS ${BENCHMARK_CXX_LINKER_FLAGS})
+endif()
+
+file(GLOB
+  SOURCE_FILES
+    *.cc
+    ${PROJECT_SOURCE_DIR}/include/benchmark/*.h
+    ${CMAKE_CURRENT_SOURCE_DIR}/*.h)
+
+add_library(benchmark ${SOURCE_FILES})
+set_target_properties(benchmark PROPERTIES
+  OUTPUT_NAME "benchmark"
+  VERSION ${GENERIC_LIB_VERSION}
+  SOVERSION ${GENERIC_LIB_SOVERSION}
+)
+target_include_directories(benchmark PUBLIC
+    $<BUILD_INTERFACE:${CMAKE_CURRENT_SOURCE_DIR}/../include>
+    )
+
+# Link threads.
+target_link_libraries(benchmark  ${BENCHMARK_CXX_LIBRARIES} ${CMAKE_THREAD_LIBS_INIT})
+find_library(LIBRT rt)
+if(LIBRT)
+  target_link_libraries(benchmark ${LIBRT})
+endif()
+
+# We need extra libraries on Windows
+if(${CMAKE_SYSTEM_NAME} MATCHES "Windows")
+  target_link_libraries(benchmark Shlwapi)
+endif()
+
+set(include_install_dir "include")
+set(lib_install_dir "lib/")
+set(bin_install_dir "bin/")
+set(config_install_dir "lib/cmake/${PROJECT_NAME}")
+set(pkgconfig_install_dir "lib/pkgconfig")
+
+set(generated_dir "${CMAKE_CURRENT_BINARY_DIR}/generated")
+
+set(version_config "${generated_dir}/${PROJECT_NAME}ConfigVersion.cmake")
+set(project_config "${generated_dir}/${PROJECT_NAME}Config.cmake")
+set(pkg_config "${generated_dir}/${PROJECT_NAME}.pc")
+set(targets_export_name "${PROJECT_NAME}Targets")
+
+set(namespace "${PROJECT_NAME}::")
+
+include(CMakePackageConfigHelpers)
+write_basic_package_version_file(
+    "${version_config}" VERSION ${GIT_VERSION} COMPATIBILITY SameMajorVersion
+)
+
+configure_file("${PROJECT_SOURCE_DIR}/cmake/Config.cmake.in" "${project_config}" @ONLY)
+configure_file("${PROJECT_SOURCE_DIR}/cmake/benchmark.pc.in" "${pkg_config}" @ONLY)
+
+if (BENCHMARK_ENABLE_INSTALL)
+  # Install target (will install the library to specified CMAKE_INSTALL_PREFIX variable)
+  install(
+    TARGETS benchmark
+    EXPORT ${targets_export_name}
+    ARCHIVE DESTINATION ${lib_install_dir}
+    LIBRARY DESTINATION ${lib_install_dir}
+    RUNTIME DESTINATION ${bin_install_dir}
+    INCLUDES DESTINATION ${include_install_dir})
+
+  install(
+    DIRECTORY "${PROJECT_SOURCE_DIR}/include/benchmark"
+    DESTINATION ${include_install_dir}
+    FILES_MATCHING PATTERN "*.*h")
+
+  install(
+      FILES "${project_config}" "${version_config}"
+      DESTINATION "${config_install_dir}")
+
+  install(
+      FILES "${pkg_config}"
+      DESTINATION "${pkgconfig_install_dir}")
+
+  install(
+      EXPORT "${targets_export_name}"
+      NAMESPACE "${namespace}"
+      DESTINATION "${config_install_dir}")
+endif()

benchmarks/thirdparty/benchmark/src/arraysize.h

@@ -0,0 +1,33 @@
+#ifndef BENCHMARK_ARRAYSIZE_H_
+#define BENCHMARK_ARRAYSIZE_H_
+
+#include "internal_macros.h"
+
+namespace benchmark {
+namespace internal {
+// The arraysize(arr) macro returns the # of elements in an array arr.
+// The expression is a compile-time constant, and therefore can be
+// used in defining new arrays, for example.  If you use arraysize on
+// a pointer by mistake, you will get a compile-time error.
+//
+
+// This template function declaration is used in defining arraysize.
+// Note that the function doesn't need an implementation, as we only
+// use its type.
+template <typename T, size_t N>
+char (&ArraySizeHelper(T (&array)[N]))[N];
+
+// That gcc wants both of these prototypes seems mysterious. VC, for
+// its part, can't decide which to use (another mystery). Matching of
+// template overloads: the final frontier.
+#ifndef COMPILER_MSVC
+template <typename T, size_t N>
+char (&ArraySizeHelper(const T (&array)[N]))[N];
+#endif
+
+#define arraysize(array) (sizeof(::benchmark::internal::ArraySizeHelper(array)))
+
+}  // end namespace internal
+}  // end namespace benchmark
+
+#endif  // BENCHMARK_ARRAYSIZE_H_

benchmarks/thirdparty/benchmark/src/benchmark.cc

@@ -0,0 +1,714 @@
+// Copyright 2015 Google Inc. All rights reserved.
+//
+// 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
+//
+//     http://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 "benchmark/benchmark.h"
+#include "benchmark_api_internal.h"
+#include "internal_macros.h"
+
+#ifndef BENCHMARK_OS_WINDOWS
+#include <sys/resource.h>
+#include <sys/time.h>
+#include <unistd.h>
+#endif
+
+#include <algorithm>
+#include <atomic>
+#include <condition_variable>
+#include <cstdio>
+#include <cstdlib>
+#include <cstring>
+#include <fstream>
+#include <iostream>
+#include <memory>
+#include <thread>
+
+#include "check.h"
+#include "colorprint.h"
+#include "commandlineflags.h"
+#include "complexity.h"
+#include "counter.h"
+#include "internal_macros.h"
+#include "log.h"
+#include "mutex.h"
+#include "re.h"
+#include "statistics.h"
+#include "string_util.h"
+#include "timers.h"
+
+DEFINE_bool(benchmark_list_tests, false,
+            "Print a list of benchmarks. This option overrides all other "
+            "options.");
+
+DEFINE_string(benchmark_filter, ".",
+              "A regular expression that specifies the set of benchmarks "
+              "to execute.  If this flag is empty, no benchmarks are run.  "
+              "If this flag is the string \"all\", all benchmarks linked "
+              "into the process are run.");
+
+DEFINE_double(benchmark_min_time, 0.5,
+              "Minimum number of seconds we should run benchmark before "
+              "results are considered significant.  For cpu-time based "
+              "tests, this is the lower bound on the total cpu time "
+              "used by all threads that make up the test.  For real-time "
+              "based tests, this is the lower bound on the elapsed time "
+              "of the benchmark execution, regardless of number of "
+              "threads.");
+
+DEFINE_int32(benchmark_repetitions, 1,
+             "The number of runs of each benchmark. If greater than 1, the "
+             "mean and standard deviation of the runs will be reported.");
+
+DEFINE_bool(benchmark_report_aggregates_only, false,
+            "Report the result of each benchmark repetitions. When 'true' is "
+            "specified only the mean, standard deviation, and other statistics "
+            "are reported for repeated benchmarks.");
+
+DEFINE_string(benchmark_format, "console",
+              "The format to use for console output. Valid values are "
+              "'console', 'json', or 'csv'.");
+
+DEFINE_string(benchmark_out_format, "json",
+              "The format to use for file output. Valid values are "
+              "'console', 'json', or 'csv'.");
+
+DEFINE_string(benchmark_out, "", "The file to write additonal output to");
+
+DEFINE_string(benchmark_color, "auto",
+              "Whether to use colors in the output.  Valid values: "
+              "'true'/'yes'/1, 'false'/'no'/0, and 'auto'. 'auto' means to use "
+              "colors if the output is being sent to a terminal and the TERM "
+              "environment variable is set to a terminal type that supports "
+              "colors.");
+
+DEFINE_bool(benchmark_counters_tabular, false,
+            "Whether to use tabular format when printing user counters to "
+            "the console.  Valid values: 'true'/'yes'/1, 'false'/'no'/0."
+            "Defaults to false.");
+
+DEFINE_int32(v, 0, "The level of verbose logging to output");
+
+namespace benchmark {
+
+namespace {
+static const size_t kMaxIterations = 1000000000;
+}  // end namespace
+
+namespace internal {
+
+void UseCharPointer(char const volatile*) {}
+
+class ThreadManager {
+ public:
+  ThreadManager(int num_threads)
+      : alive_threads_(num_threads), start_stop_barrier_(num_threads) {}
+
+  Mutex& GetBenchmarkMutex() const RETURN_CAPABILITY(benchmark_mutex_) {
+    return benchmark_mutex_;
+  }
+
+  bool StartStopBarrier() EXCLUDES(end_cond_mutex_) {
+    return start_stop_barrier_.wait();
+  }
+
+  void NotifyThreadComplete() EXCLUDES(end_cond_mutex_) {
+    start_stop_barrier_.removeThread();
+    if (--alive_threads_ == 0) {
+      MutexLock lock(end_cond_mutex_);
+      end_condition_.notify_all();
+    }
+  }
+
+  void WaitForAllThreads() EXCLUDES(end_cond_mutex_) {
+    MutexLock lock(end_cond_mutex_);
+    end_condition_.wait(lock.native_handle(),
+                        [this]() { return alive_threads_ == 0; });
+  }
+
+ public:
+  struct Result {
+    double real_time_used = 0;
+    double cpu_time_used = 0;
+    double manual_time_used = 0;
+    int64_t bytes_processed = 0;
+    int64_t items_processed = 0;
+    int complexity_n = 0;
+    std::string report_label_;
+    std::string error_message_;
+    bool has_error_ = false;
+    UserCounters counters;
+  };
+  GUARDED_BY(GetBenchmarkMutex()) Result results;
+
+ private:
+  mutable Mutex benchmark_mutex_;
+  std::atomic<int> alive_threads_;
+  Barrier start_stop_barrier_;
+  Mutex end_cond_mutex_;
+  Condition end_condition_;
+};
+
+// Timer management class
+class ThreadTimer {
+ public:
+  ThreadTimer() = default;
+
+  // Called by each thread
+  void StartTimer() {
+    running_ = true;
+    start_real_time_ = ChronoClockNow();
+    start_cpu_time_ = ThreadCPUUsage();
+  }
+
+  // Called by each thread
+  void StopTimer() {
+    CHECK(running_);
+    running_ = false;
+    real_time_used_ += ChronoClockNow() - start_real_time_;
+    // Floating point error can result in the subtraction producing a negative
+    // time. Guard against that.
+    cpu_time_used_ += std::max<double>(ThreadCPUUsage() - start_cpu_time_, 0);
+  }
+
+  // Called by each thread
+  void SetIterationTime(double seconds) { manual_time_used_ += seconds; }
+
+  bool running() const { return running_; }
+
+  // REQUIRES: timer is not running
+  double real_time_used() {
+    CHECK(!running_);
+    return real_time_used_;
+  }
+
+  // REQUIRES: timer is not running
+  double cpu_time_used() {
+    CHECK(!running_);
+    return cpu_time_used_;
+  }
+
+  // REQUIRES: timer is not running
+  double manual_time_used() {
+    CHECK(!running_);
+    return manual_time_used_;
+  }
+
+ private:
+  bool running_ = false;        // Is the timer running
+  double start_real_time_ = 0;  // If running_
+  double start_cpu_time_ = 0;   // If running_
+
+  // Accumulated time so far (does not contain current slice if running_)
+  double real_time_used_ = 0;
+  double cpu_time_used_ = 0;
+  // Manually set iteration time. User sets this with SetIterationTime(seconds).
+  double manual_time_used_ = 0;
+};
+
+namespace {
+
+BenchmarkReporter::Run CreateRunReport(
+    const benchmark::internal::Benchmark::Instance& b,
+    const internal::ThreadManager::Result& results, size_t iters,
+    double seconds) {
+  // Create report about this benchmark run.
+  BenchmarkReporter::Run report;
+
+  report.benchmark_name = b.name;
+  report.error_occurred = results.has_error_;
+  report.error_message = results.error_message_;
+  report.report_label = results.report_label_;
+  // Report the total iterations across all threads.
+  report.iterations = static_cast<int64_t>(iters) * b.threads;
+  report.time_unit = b.time_unit;
+
+  if (!report.error_occurred) {
+    double bytes_per_second = 0;
+    if (results.bytes_processed > 0 && seconds > 0.0) {
+      bytes_per_second = (results.bytes_processed / seconds);
+    }
+    double items_per_second = 0;
+    if (results.items_processed > 0 && seconds > 0.0) {
+      items_per_second = (results.items_processed / seconds);
+    }
+
+    if (b.use_manual_time) {
+      report.real_accumulated_time = results.manual_time_used;
+    } else {
+      report.real_accumulated_time = results.real_time_used;
+    }
+    report.cpu_accumulated_time = results.cpu_time_used;
+    report.bytes_per_second = bytes_per_second;
+    report.items_per_second = items_per_second;
+    report.complexity_n = results.complexity_n;
+    report.complexity = b.complexity;
+    report.complexity_lambda = b.complexity_lambda;
+    report.statistics = b.statistics;
+    report.counters = results.counters;
+    internal::Finish(&report.counters, seconds, b.threads);
+  }
+  return report;
+}
+
+// Execute one thread of benchmark b for the specified number of iterations.
+// Adds the stats collected for the thread into *total.
+void RunInThread(const benchmark::internal::Benchmark::Instance* b,
+                 size_t iters, int thread_id,
+                 internal::ThreadManager* manager) {
+  internal::ThreadTimer timer;
+  State st(iters, b->arg, thread_id, b->threads, &timer, manager);
+  b->benchmark->Run(st);
+  CHECK(st.iterations() == st.max_iterations)
+      << "Benchmark returned before State::KeepRunning() returned false!";
+  {
+    MutexLock l(manager->GetBenchmarkMutex());
+    internal::ThreadManager::Result& results = manager->results;
+    results.cpu_time_used += timer.cpu_time_used();
+    results.real_time_used += timer.real_time_used();
+    results.manual_time_used += timer.manual_time_used();
+    results.bytes_processed += st.bytes_processed();
+    results.items_processed += st.items_processed();
+    results.complexity_n += st.complexity_length_n();
+    internal::Increment(&results.counters, st.counters);
+  }
+  manager->NotifyThreadComplete();
+}
+
+std::vector<BenchmarkReporter::Run> RunBenchmark(
+    const benchmark::internal::Benchmark::Instance& b,
+    std::vector<BenchmarkReporter::Run>* complexity_reports) {
+  std::vector<BenchmarkReporter::Run> reports;  // return value
+
+  const bool has_explicit_iteration_count = b.iterations != 0;
+  size_t iters = has_explicit_iteration_count ? b.iterations : 1;
+  std::unique_ptr<internal::ThreadManager> manager;
+  std::vector<std::thread> pool(b.threads - 1);
+  const int repeats =
+      b.repetitions != 0 ? b.repetitions : FLAGS_benchmark_repetitions;
+  const bool report_aggregates_only =
+      repeats != 1 &&
+      (b.report_mode == internal::RM_Unspecified
+           ? FLAGS_benchmark_report_aggregates_only
+           : b.report_mode == internal::RM_ReportAggregatesOnly);
+  for (int repetition_num = 0; repetition_num < repeats; repetition_num++) {
+    for (;;) {
+      // Try benchmark
+      VLOG(2) << "Running " << b.name << " for " << iters << "\n";
+
+      manager.reset(new internal::ThreadManager(b.threads));
+      for (std::size_t ti = 0; ti < pool.size(); ++ti) {
+        pool[ti] = std::thread(&RunInThread, &b, iters,
+                               static_cast<int>(ti + 1), manager.get());
+      }
+      RunInThread(&b, iters, 0, manager.get());
+      manager->WaitForAllThreads();
+      for (std::thread& thread : pool) thread.join();
+      internal::ThreadManager::Result results;
+      {
+        MutexLock l(manager->GetBenchmarkMutex());
+        results = manager->results;
+      }
+      manager.reset();
+      // Adjust real/manual time stats since they were reported per thread.
+      results.real_time_used /= b.threads;
+      results.manual_time_used /= b.threads;
+
+      VLOG(2) << "Ran in " << results.cpu_time_used << "/"
+              << results.real_time_used << "\n";
+
+      // Base decisions off of real time if requested by this benchmark.
+      double seconds = results.cpu_time_used;
+      if (b.use_manual_time) {
+        seconds = results.manual_time_used;
+      } else if (b.use_real_time) {
+        seconds = results.real_time_used;
+      }
+
+      const double min_time =
+          !IsZero(b.min_time) ? b.min_time : FLAGS_benchmark_min_time;
+
+      // Determine if this run should be reported; Either it has
+      // run for a sufficient amount of time or because an error was reported.
+      const bool should_report =  repetition_num > 0
+        || has_explicit_iteration_count // An exact iteration count was requested
+        || results.has_error_
+        || iters >= kMaxIterations
+        || seconds >= min_time // the elapsed time is large enough
+        // CPU time is specified but the elapsed real time greatly exceeds the
+        // minimum time. Note that user provided timers are except from this
+        // sanity check.
+        || ((results.real_time_used >= 5 * min_time) && !b.use_manual_time);
+
+      if (should_report) {
+        BenchmarkReporter::Run report =
+            CreateRunReport(b, results, iters, seconds);
+        if (!report.error_occurred && b.complexity != oNone)
+          complexity_reports->push_back(report);
+        reports.push_back(report);
+        break;
+      }
+
+      // See how much iterations should be increased by
+      // Note: Avoid division by zero with max(seconds, 1ns).
+      double multiplier = min_time * 1.4 / std::max(seconds, 1e-9);
+      // If our last run was at least 10% of FLAGS_benchmark_min_time then we
+      // use the multiplier directly. Otherwise we use at most 10 times
+      // expansion.
+      // NOTE: When the last run was at least 10% of the min time the max
+      // expansion should be 14x.
+      bool is_significant = (seconds / min_time) > 0.1;
+      multiplier = is_significant ? multiplier : std::min(10.0, multiplier);
+      if (multiplier <= 1.0) multiplier = 2.0;
+      double next_iters = std::max(multiplier * iters, iters + 1.0);
+      if (next_iters > kMaxIterations) {
+        next_iters = kMaxIterations;
+      }
+      VLOG(3) << "Next iters: " << next_iters << ", " << multiplier << "\n";
+      iters = static_cast<int>(next_iters + 0.5);
+    }
+  }
+  // Calculate additional statistics
+  auto stat_reports = ComputeStats(reports);
+  if ((b.complexity != oNone) && b.last_benchmark_instance) {
+    auto additional_run_stats = ComputeBigO(*complexity_reports);
+    stat_reports.insert(stat_reports.end(), additional_run_stats.begin(),
+                        additional_run_stats.end());
+    complexity_reports->clear();
+  }
+
+  if (report_aggregates_only) reports.clear();
+  reports.insert(reports.end(), stat_reports.begin(), stat_reports.end());
+  return reports;
+}
+
+}  // namespace
+}  // namespace internal
+
+State::State(size_t max_iters, const std::vector<int>& ranges, int thread_i,
+             int n_threads, internal::ThreadTimer* timer,
+             internal::ThreadManager* manager)
+    : started_(false),
+      finished_(false),
+      total_iterations_(max_iters + 1),
+      range_(ranges),
+      bytes_processed_(0),
+      items_processed_(0),
+      complexity_n_(0),
+      error_occurred_(false),
+      counters(),
+      thread_index(thread_i),
+      threads(n_threads),
+      max_iterations(max_iters),
+      timer_(timer),
+      manager_(manager) {
+  CHECK(max_iterations != 0) << "At least one iteration must be run";
+  CHECK(total_iterations_ != 0) << "max iterations wrapped around";
+  CHECK_LT(thread_index, threads) << "thread_index must be less than threads";
+}
+
+void State::PauseTiming() {
+  // Add in time accumulated so far
+  CHECK(started_ && !finished_ && !error_occurred_);
+  timer_->StopTimer();
+}
+
+void State::ResumeTiming() {
+  CHECK(started_ && !finished_ && !error_occurred_);
+  timer_->StartTimer();
+}
+
+void State::SkipWithError(const char* msg) {
+  CHECK(msg);
+  error_occurred_ = true;
+  {
+    MutexLock l(manager_->GetBenchmarkMutex());
+    if (manager_->results.has_error_ == false) {
+      manager_->results.error_message_ = msg;
+      manager_->results.has_error_ = true;
+    }
+  }
+  total_iterations_ = 1;
+  if (timer_->running()) timer_->StopTimer();
+}
+
+void State::SetIterationTime(double seconds) {
+  timer_->SetIterationTime(seconds);
+}
+
+void State::SetLabel(const char* label) {
+  MutexLock l(manager_->GetBenchmarkMutex());
+  manager_->results.report_label_ = label;
+}
+
+void State::StartKeepRunning() {
+  CHECK(!started_ && !finished_);
+  started_ = true;
+  manager_->StartStopBarrier();
+  if (!error_occurred_) ResumeTiming();
+}
+
+void State::FinishKeepRunning() {
+  CHECK(started_ && (!finished_ || error_occurred_));
+  if (!error_occurred_) {
+    PauseTiming();
+  }
+  // Total iterations has now wrapped around zero. Fix this.
+  total_iterations_ = 1;
+  finished_ = true;
+  manager_->StartStopBarrier();
+}
+
+namespace internal {
+namespace {
+
+void RunBenchmarks(const std::vector<Benchmark::Instance>& benchmarks,
+                           BenchmarkReporter* console_reporter,
+                           BenchmarkReporter* file_reporter) {
+  // Note the file_reporter can be null.
+  CHECK(console_reporter != nullptr);
+
+  // Determine the width of the name field using a minimum width of 10.
+  bool has_repetitions = FLAGS_benchmark_repetitions > 1;
+  size_t name_field_width = 10;
+  size_t stat_field_width = 0;
+  for (const Benchmark::Instance& benchmark : benchmarks) {
+    name_field_width =
+        std::max<size_t>(name_field_width, benchmark.name.size());
+    has_repetitions |= benchmark.repetitions > 1;
+
+    for(const auto& Stat : *benchmark.statistics)
+      stat_field_width = std::max<size_t>(stat_field_width, Stat.name_.size());
+  }
+  if (has_repetitions) name_field_width += 1 + stat_field_width;
+
+  // Print header here
+  BenchmarkReporter::Context context;
+  context.name_field_width = name_field_width;
+
+  // Keep track of runing times of all instances of current benchmark
+  std::vector<BenchmarkReporter::Run> complexity_reports;
+
+  // We flush streams after invoking reporter methods that write to them. This
+  // ensures users get timely updates even when streams are not line-buffered.
+  auto flushStreams = [](BenchmarkReporter* reporter) {
+    if (!reporter) return;
+    std::flush(reporter->GetOutputStream());
+    std::flush(reporter->GetErrorStream());
+  };
+
+  if (console_reporter->ReportContext(context) &&
+      (!file_reporter || file_reporter->ReportContext(context))) {
+    flushStreams(console_reporter);
+    flushStreams(file_reporter);
+    for (const auto& benchmark : benchmarks) {
+      std::vector<BenchmarkReporter::Run> reports =
+          RunBenchmark(benchmark, &complexity_reports);
+      console_reporter->ReportRuns(reports);
+      if (file_reporter) file_reporter->ReportRuns(reports);
+      flushStreams(console_reporter);
+      flushStreams(file_reporter);
+    }
+  }
+  console_reporter->Finalize();
+  if (file_reporter) file_reporter->Finalize();
+  flushStreams(console_reporter);
+  flushStreams(file_reporter);
+}
+
+std::unique_ptr<BenchmarkReporter> CreateReporter(
+    std::string const& name, ConsoleReporter::OutputOptions output_opts) {
+  typedef std::unique_ptr<BenchmarkReporter> PtrType;
+  if (name == "console") {
+    return PtrType(new ConsoleReporter(output_opts));
+  } else if (name == "json") {
+    return PtrType(new JSONReporter);
+  } else if (name == "csv") {
+    return PtrType(new CSVReporter);
+  } else {
+    std::cerr << "Unexpected format: '" << name << "'\n";
+    std::exit(1);
+  }
+}
+
+}  // end namespace
+
+bool IsZero(double n) {
+  return std::abs(n) < std::numeric_limits<double>::epsilon();
+}
+
+ConsoleReporter::OutputOptions GetOutputOptions(bool force_no_color) {
+  int output_opts = ConsoleReporter::OO_Defaults;
+  if ((FLAGS_benchmark_color == "auto" && IsColorTerminal()) ||
+      IsTruthyFlagValue(FLAGS_benchmark_color)) {
+    output_opts |= ConsoleReporter::OO_Color;
+  } else {
+    output_opts &= ~ConsoleReporter::OO_Color;
+  }
+  if(force_no_color) {
+    output_opts &= ~ConsoleReporter::OO_Color;
+  }
+  if(FLAGS_benchmark_counters_tabular) {
+    output_opts |= ConsoleReporter::OO_Tabular;
+  } else {
+    output_opts &= ~ConsoleReporter::OO_Tabular;
+  }
+  return static_cast< ConsoleReporter::OutputOptions >(output_opts);
+}
+
+}  // end namespace internal
+
+size_t RunSpecifiedBenchmarks() {
+  return RunSpecifiedBenchmarks(nullptr, nullptr);
+}
+
+size_t RunSpecifiedBenchmarks(BenchmarkReporter* console_reporter) {
+  return RunSpecifiedBenchmarks(console_reporter, nullptr);
+}
+
+size_t RunSpecifiedBenchmarks(BenchmarkReporter* console_reporter,
+                              BenchmarkReporter* file_reporter) {
+  std::string spec = FLAGS_benchmark_filter;
+  if (spec.empty() || spec == "all")
+    spec = ".";  // Regexp that matches all benchmarks
+
+  // Setup the reporters
+  std::ofstream output_file;
+  std::unique_ptr<BenchmarkReporter> default_console_reporter;
+  std::unique_ptr<BenchmarkReporter> default_file_reporter;
+  if (!console_reporter) {
+    default_console_reporter = internal::CreateReporter(
+          FLAGS_benchmark_format, internal::GetOutputOptions());
+    console_reporter = default_console_reporter.get();
+  }
+  auto& Out = console_reporter->GetOutputStream();
+  auto& Err = console_reporter->GetErrorStream();
+
+  std::string const& fname = FLAGS_benchmark_out;
+  if (fname.empty() && file_reporter) {
+    Err << "A custom file reporter was provided but "
+           "--benchmark_out=<file> was not specified."
+        << std::endl;
+    std::exit(1);
+  }
+  if (!fname.empty()) {
+    output_file.open(fname);
+    if (!output_file.is_open()) {
+      Err << "invalid file name: '" << fname << std::endl;
+      std::exit(1);
+    }
+    if (!file_reporter) {
+      default_file_reporter = internal::CreateReporter(
+          FLAGS_benchmark_out_format, ConsoleReporter::OO_None);
+      file_reporter = default_file_reporter.get();
+    }
+    file_reporter->SetOutputStream(&output_file);
+    file_reporter->SetErrorStream(&output_file);
+  }
+
+  std::vector<internal::Benchmark::Instance> benchmarks;
+  if (!FindBenchmarksInternal(spec, &benchmarks, &Err)) return 0;
+
+  if (benchmarks.empty()) {
+    Err << "Failed to match any benchmarks against regex: " << spec << "\n";
+    return 0;
+  }
+
+  if (FLAGS_benchmark_list_tests) {
+    for (auto const& benchmark : benchmarks) Out << benchmark.name << "\n";
+  } else {
+    internal::RunBenchmarks(benchmarks, console_reporter, file_reporter);
+  }
+
+  return benchmarks.size();
+}
+
+namespace internal {
+
+void PrintUsageAndExit() {
+  fprintf(stdout,
+          "benchmark"
+          " [--benchmark_list_tests={true|false}]\n"
+          "          [--benchmark_filter=<regex>]\n"
+          "          [--benchmark_min_time=<min_time>]\n"
+          "          [--benchmark_repetitions=<num_repetitions>]\n"
+          "          [--benchmark_report_aggregates_only={true|false}\n"
+          "          [--benchmark_format=<console|json|csv>]\n"
+          "          [--benchmark_out=<filename>]\n"
+          "          [--benchmark_out_format=<json|console|csv>]\n"
+          "          [--benchmark_color={auto|true|false}]\n"
+          "          [--benchmark_counters_tabular={true|false}]\n"
+          "          [--v=<verbosity>]\n");
+  exit(0);
+}
+
+void ParseCommandLineFlags(int* argc, char** argv) {
+  using namespace benchmark;
+  for (int i = 1; i < *argc; ++i) {
+    if (ParseBoolFlag(argv[i], "benchmark_list_tests",
+                      &FLAGS_benchmark_list_tests) ||
+        ParseStringFlag(argv[i], "benchmark_filter", &FLAGS_benchmark_filter) ||
+        ParseDoubleFlag(argv[i], "benchmark_min_time",
+                        &FLAGS_benchmark_min_time) ||
+        ParseInt32Flag(argv[i], "benchmark_repetitions",
+                       &FLAGS_benchmark_repetitions) ||
+        ParseBoolFlag(argv[i], "benchmark_report_aggregates_only",
+                      &FLAGS_benchmark_report_aggregates_only) ||
+        ParseStringFlag(argv[i], "benchmark_format", &FLAGS_benchmark_format) ||
+        ParseStringFlag(argv[i], "benchmark_out", &FLAGS_benchmark_out) ||
+        ParseStringFlag(argv[i], "benchmark_out_format",
+                        &FLAGS_benchmark_out_format) ||
+        ParseStringFlag(argv[i], "benchmark_color", &FLAGS_benchmark_color) ||
+        // "color_print" is the deprecated name for "benchmark_color".
+        // TODO: Remove this.
+        ParseStringFlag(argv[i], "color_print", &FLAGS_benchmark_color) ||
+        ParseBoolFlag(argv[i], "benchmark_counters_tabular",
+                        &FLAGS_benchmark_counters_tabular) ||
+        ParseInt32Flag(argv[i], "v", &FLAGS_v)) {
+      for (int j = i; j != *argc - 1; ++j) argv[j] = argv[j + 1];
+
+      --(*argc);
+      --i;
+    } else if (IsFlag(argv[i], "help")) {
+      PrintUsageAndExit();
+    }
+  }
+  for (auto const* flag :
+       {&FLAGS_benchmark_format, &FLAGS_benchmark_out_format})
+    if (*flag != "console" && *flag != "json" && *flag != "csv") {
+      PrintUsageAndExit();
+    }
+  if (FLAGS_benchmark_color.empty()) {
+    PrintUsageAndExit();
+  }
+}
+
+int InitializeStreams() {
+  static std::ios_base::Init init;
+  return 0;
+}
+
+}  // end namespace internal
+
+void Initialize(int* argc, char** argv) {
+  internal::ParseCommandLineFlags(argc, argv);
+  internal::LogLevel() = FLAGS_v;
+}
+
+bool ReportUnrecognizedArguments(int argc, char** argv) {
+  for (int i = 1; i < argc; ++i) {
+    fprintf(stderr, "%s: error: unrecognized command-line flag: %s\n", argv[0], argv[i]);
+  }
+  return argc > 1;
+}
+
+}  // end namespace benchmark

benchmarks/thirdparty/benchmark/src/benchmark_api_internal.h

@@ -0,0 +1,47 @@
+#ifndef BENCHMARK_API_INTERNAL_H
+#define BENCHMARK_API_INTERNAL_H
+
+#include "benchmark/benchmark.h"
+
+#include <cmath>
+#include <iosfwd>
+#include <limits>
+#include <string>
+#include <vector>
+
+namespace benchmark {
+namespace internal {
+
+// Information kept per benchmark we may want to run
+struct Benchmark::Instance {
+  std::string name;
+  Benchmark* benchmark;
+  ReportMode report_mode;
+  std::vector<int> arg;
+  TimeUnit time_unit;
+  int range_multiplier;
+  bool use_real_time;
+  bool use_manual_time;
+  BigO complexity;
+  BigOFunc* complexity_lambda;
+  UserCounters counters;
+  const std::vector<Statistics>* statistics;
+  bool last_benchmark_instance;
+  int repetitions;
+  double min_time;
+  size_t iterations;
+  int threads;  // Number of concurrent threads to us
+};
+
+bool FindBenchmarksInternal(const std::string& re,
+                            std::vector<Benchmark::Instance>* benchmarks,
+                            std::ostream* Err);
+
+bool IsZero(double n);
+
+ConsoleReporter::OutputOptions GetOutputOptions(bool force_no_color = false);
+
+}  // end namespace internal
+}  // end namespace benchmark
+
+#endif  // BENCHMARK_API_INTERNAL_H

benchmarks/thirdparty/benchmark/src/benchmark_register.cc

@@ -0,0 +1,476 @@
+// Copyright 2015 Google Inc. All rights reserved.
+//
+// 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
+//
+//     http://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 "benchmark/benchmark.h"
+#include "benchmark_api_internal.h"
+#include "internal_macros.h"
+
+#ifndef BENCHMARK_OS_WINDOWS
+#include <sys/resource.h>
+#include <sys/time.h>
+#include <unistd.h>
+#endif
+
+#include <algorithm>
+#include <atomic>
+#include <condition_variable>
+#include <cstdio>
+#include <cstdlib>
+#include <cstring>
+#include <fstream>
+#include <iostream>
+#include <memory>
+#include <sstream>
+#include <thread>
+
+#include "check.h"
+#include "commandlineflags.h"
+#include "complexity.h"
+#include "statistics.h"
+#include "log.h"
+#include "mutex.h"
+#include "re.h"
+#include "string_util.h"
+#include "timers.h"
+
+namespace benchmark {
+
+namespace {
+// For non-dense Range, intermediate values are powers of kRangeMultiplier.
+static const int kRangeMultiplier = 8;
+// The size of a benchmark family determines is the number of inputs to repeat
+// the benchmark on. If this is "large" then warn the user during configuration.
+static const size_t kMaxFamilySize = 100;
+}  // end namespace
+
+namespace internal {
+
+//=============================================================================//
+//                         BenchmarkFamilies
+//=============================================================================//
+
+// Class for managing registered benchmarks.  Note that each registered
+// benchmark identifies a family of related benchmarks to run.
+class BenchmarkFamilies {
+ public:
+  static BenchmarkFamilies* GetInstance();
+
+  // Registers a benchmark family and returns the index assigned to it.
+  size_t AddBenchmark(std::unique_ptr<Benchmark> family);
+
+  // Clear all registered benchmark families.
+  void ClearBenchmarks();
+
+  // Extract the list of benchmark instances that match the specified
+  // regular expression.
+  bool FindBenchmarks(const std::string& re,
+                      std::vector<Benchmark::Instance>* benchmarks,
+                      std::ostream* Err);
+
+ private:
+  BenchmarkFamilies() {}
+
+  std::vector<std::unique_ptr<Benchmark>> families_;
+  Mutex mutex_;
+};
+
+BenchmarkFamilies* BenchmarkFamilies::GetInstance() {
+  static BenchmarkFamilies instance;
+  return &instance;
+}
+
+size_t BenchmarkFamilies::AddBenchmark(std::unique_ptr<Benchmark> family) {
+  MutexLock l(mutex_);
+  size_t index = families_.size();
+  families_.push_back(std::move(family));
+  return index;
+}
+
+void BenchmarkFamilies::ClearBenchmarks() {
+  MutexLock l(mutex_);
+  families_.clear();
+  families_.shrink_to_fit();
+}
+
+bool BenchmarkFamilies::FindBenchmarks(
+    const std::string& spec, std::vector<Benchmark::Instance>* benchmarks,
+    std::ostream* ErrStream) {
+  CHECK(ErrStream);
+  auto& Err = *ErrStream;
+  // Make regular expression out of command-line flag
+  std::string error_msg;
+  Regex re;
+  if (!re.Init(spec, &error_msg)) {
+    Err << "Could not compile benchmark re: " << error_msg << std::endl;
+    return false;
+  }
+
+  // Special list of thread counts to use when none are specified
+  const std::vector<int> one_thread = {1};
+
+  MutexLock l(mutex_);
+  for (std::unique_ptr<Benchmark>& family : families_) {
+    // Family was deleted or benchmark doesn't match
+    if (!family) continue;
+
+    if (family->ArgsCnt() == -1) {
+      family->Args({});
+    }
+    const std::vector<int>* thread_counts =
+        (family->thread_counts_.empty()
+             ? &one_thread
+             : &static_cast<const std::vector<int>&>(family->thread_counts_));
+    const size_t family_size = family->args_.size() * thread_counts->size();
+    // The benchmark will be run at least 'family_size' different inputs.
+    // If 'family_size' is very large warn the user.
+    if (family_size > kMaxFamilySize) {
+      Err << "The number of inputs is very large. " << family->name_
+          << " will be repeated at least " << family_size << " times.\n";
+    }
+    // reserve in the special case the regex ".", since we know the final
+    // family size.
+    if (spec == ".") benchmarks->reserve(family_size);
+
+    for (auto const& args : family->args_) {
+      for (int num_threads : *thread_counts) {
+        Benchmark::Instance instance;
+        instance.name = family->name_;
+        instance.benchmark = family.get();
+        instance.report_mode = family->report_mode_;
+        instance.arg = args;
+        instance.time_unit = family->time_unit_;
+        instance.range_multiplier = family->range_multiplier_;
+        instance.min_time = family->min_time_;
+        instance.iterations = family->iterations_;
+        instance.repetitions = family->repetitions_;
+        instance.use_real_time = family->use_real_time_;
+        instance.use_manual_time = family->use_manual_time_;
+        instance.complexity = family->complexity_;
+        instance.complexity_lambda = family->complexity_lambda_;
+        instance.statistics = &family->statistics_;
+        instance.threads = num_threads;
+
+        // Add arguments to instance name
+        size_t arg_i = 0;
+        for (auto const& arg : args) {
+          instance.name += "/";
+
+          if (arg_i < family->arg_names_.size()) {
+            const auto& arg_name = family->arg_names_[arg_i];
+            if (!arg_name.empty()) {
+              instance.name +=
+                  StringPrintF("%s:", family->arg_names_[arg_i].c_str());
+            }
+          }
+          
+          instance.name += StringPrintF("%d", arg);
+          ++arg_i;
+        }
+
+        if (!IsZero(family->min_time_))
+          instance.name += StringPrintF("/min_time:%0.3f", family->min_time_);
+        if (family->iterations_ != 0)
+          instance.name += StringPrintF("/iterations:%d", family->iterations_);
+        if (family->repetitions_ != 0)
+          instance.name += StringPrintF("/repeats:%d", family->repetitions_);
+
+        if (family->use_manual_time_) {
+          instance.name += "/manual_time";
+        } else if (family->use_real_time_) {
+          instance.name += "/real_time";
+        }
+
+        // Add the number of threads used to the name
+        if (!family->thread_counts_.empty()) {
+          instance.name += StringPrintF("/threads:%d", instance.threads);
+        }
+
+        if (re.Match(instance.name)) {
+          instance.last_benchmark_instance = (&args == &family->args_.back());
+          benchmarks->push_back(std::move(instance));
+        }
+      }
+    }
+  }
+  return true;
+}
+
+Benchmark* RegisterBenchmarkInternal(Benchmark* bench) {
+  std::unique_ptr<Benchmark> bench_ptr(bench);
+  BenchmarkFamilies* families = BenchmarkFamilies::GetInstance();
+  families->AddBenchmark(std::move(bench_ptr));
+  return bench;
+}
+
+// FIXME: This function is a hack so that benchmark.cc can access
+// `BenchmarkFamilies`
+bool FindBenchmarksInternal(const std::string& re,
+                            std::vector<Benchmark::Instance>* benchmarks,
+                            std::ostream* Err) {
+  return BenchmarkFamilies::GetInstance()->FindBenchmarks(re, benchmarks, Err);
+}
+
+//=============================================================================//
+//                               Benchmark
+//=============================================================================//
+
+Benchmark::Benchmark(const char* name)
+    : name_(name),
+      report_mode_(RM_Unspecified),
+      time_unit_(kNanosecond),
+      range_multiplier_(kRangeMultiplier),
+      min_time_(0),
+      iterations_(0),
+      repetitions_(0),
+      use_real_time_(false),
+      use_manual_time_(false),
+      complexity_(oNone),
+      complexity_lambda_(nullptr) {
+  ComputeStatistics("mean", StatisticsMean);
+  ComputeStatistics("median", StatisticsMedian);
+  ComputeStatistics("stddev", StatisticsStdDev);
+}
+
+Benchmark::~Benchmark() {}
+
+void Benchmark::AddRange(std::vector<int>* dst, int lo, int hi, int mult) {
+  CHECK_GE(lo, 0);
+  CHECK_GE(hi, lo);
+  CHECK_GE(mult, 2);
+
+  // Add "lo"
+  dst->push_back(lo);
+
+  static const int kint32max = std::numeric_limits<int32_t>::max();
+
+  // Now space out the benchmarks in multiples of "mult"
+  for (int32_t i = 1; i < kint32max / mult; i *= mult) {
+    if (i >= hi) break;
+    if (i > lo) {
+      dst->push_back(i);
+    }
+  }
+  // Add "hi" (if different from "lo")
+  if (hi != lo) {
+    dst->push_back(hi);
+  }
+}
+
+Benchmark* Benchmark::Arg(int x) {
+  CHECK(ArgsCnt() == -1 || ArgsCnt() == 1);
+  args_.push_back({x});
+  return this;
+}
+
+Benchmark* Benchmark::Unit(TimeUnit unit) {
+  time_unit_ = unit;
+  return this;
+}
+
+Benchmark* Benchmark::Range(int start, int limit) {
+  CHECK(ArgsCnt() == -1 || ArgsCnt() == 1);
+  std::vector<int> arglist;
+  AddRange(&arglist, start, limit, range_multiplier_);
+
+  for (int i : arglist) {
+    args_.push_back({i});
+  }
+  return this;
+}
+
+Benchmark* Benchmark::Ranges(const std::vector<std::pair<int, int>>& ranges) {
+  CHECK(ArgsCnt() == -1 || ArgsCnt() == static_cast<int>(ranges.size()));
+  std::vector<std::vector<int>> arglists(ranges.size());
+  std::size_t total = 1;
+  for (std::size_t i = 0; i < ranges.size(); i++) {
+    AddRange(&arglists[i], ranges[i].first, ranges[i].second,
+             range_multiplier_);
+    total *= arglists[i].size();
+  }
+
+  std::vector<std::size_t> ctr(arglists.size(), 0);
+
+  for (std::size_t i = 0; i < total; i++) {
+    std::vector<int> tmp;
+    tmp.reserve(arglists.size());
+
+    for (std::size_t j = 0; j < arglists.size(); j++) {
+      tmp.push_back(arglists[j].at(ctr[j]));
+    }
+
+    args_.push_back(std::move(tmp));
+
+    for (std::size_t j = 0; j < arglists.size(); j++) {
+      if (ctr[j] + 1 < arglists[j].size()) {
+        ++ctr[j];
+        break;
+      }
+      ctr[j] = 0;
+    }
+  }
+  return this;
+}
+
+Benchmark* Benchmark::ArgName(const std::string& name) {
+  CHECK(ArgsCnt() == -1 || ArgsCnt() == 1);
+  arg_names_ = {name};
+  return this;
+}
+
+Benchmark* Benchmark::ArgNames(const std::vector<std::string>& names) {
+  CHECK(ArgsCnt() == -1 || ArgsCnt() == static_cast<int>(names.size()));
+  arg_names_ = names;
+  return this;
+}
+
+Benchmark* Benchmark::DenseRange(int start, int limit, int step) {
+  CHECK(ArgsCnt() == -1 || ArgsCnt() == 1);
+  CHECK_GE(start, 0);
+  CHECK_LE(start, limit);
+  for (int arg = start; arg <= limit; arg += step) {
+    args_.push_back({arg});
+  }
+  return this;
+}
+
+Benchmark* Benchmark::Args(const std::vector<int>& args) {
+  CHECK(ArgsCnt() == -1 || ArgsCnt() == static_cast<int>(args.size()));
+  args_.push_back(args);
+  return this;
+}
+
+Benchmark* Benchmark::Apply(void (*custom_arguments)(Benchmark* benchmark)) {
+  custom_arguments(this);
+  return this;
+}
+
+Benchmark* Benchmark::RangeMultiplier(int multiplier) {
+  CHECK(multiplier > 1);
+  range_multiplier_ = multiplier;
+  return this;
+}
+
+
+Benchmark* Benchmark::MinTime(double t) {
+  CHECK(t > 0.0);
+  CHECK(iterations_ == 0);
+  min_time_ = t;
+  return this;
+}
+
+
+Benchmark* Benchmark::Iterations(size_t n) {
+  CHECK(n > 0);
+  CHECK(IsZero(min_time_));
+  iterations_ = n;
+  return this;
+}
+
+Benchmark* Benchmark::Repetitions(int n) {
+  CHECK(n > 0);
+  repetitions_ = n;
+  return this;
+}
+
+Benchmark* Benchmark::ReportAggregatesOnly(bool value) {
+  report_mode_ = value ? RM_ReportAggregatesOnly : RM_Default;
+  return this;
+}
+
+Benchmark* Benchmark::UseRealTime() {
+  CHECK(!use_manual_time_)
+      << "Cannot set UseRealTime and UseManualTime simultaneously.";
+  use_real_time_ = true;
+  return this;
+}
+
+Benchmark* Benchmark::UseManualTime() {
+  CHECK(!use_real_time_)
+      << "Cannot set UseRealTime and UseManualTime simultaneously.";
+  use_manual_time_ = true;
+  return this;
+}
+
+Benchmark* Benchmark::Complexity(BigO complexity) {
+  complexity_ = complexity;
+  return this;
+}
+
+Benchmark* Benchmark::Complexity(BigOFunc* complexity) {
+  complexity_lambda_ = complexity;
+  complexity_ = oLambda;
+  return this;
+}
+
+Benchmark* Benchmark::ComputeStatistics(std::string name,
+                                        StatisticsFunc* statistics) {
+  statistics_.emplace_back(name, statistics);
+  return this;
+}
+
+Benchmark* Benchmark::Threads(int t) {
+  CHECK_GT(t, 0);
+  thread_counts_.push_back(t);
+  return this;
+}
+
+Benchmark* Benchmark::ThreadRange(int min_threads, int max_threads) {
+  CHECK_GT(min_threads, 0);
+  CHECK_GE(max_threads, min_threads);
+
+  AddRange(&thread_counts_, min_threads, max_threads, 2);
+  return this;
+}
+
+Benchmark* Benchmark::DenseThreadRange(int min_threads, int max_threads,
+                                       int stride) {
+  CHECK_GT(min_threads, 0);
+  CHECK_GE(max_threads, min_threads);
+  CHECK_GE(stride, 1);
+
+  for (auto i = min_threads; i < max_threads; i += stride) {
+    thread_counts_.push_back(i);
+  }
+  thread_counts_.push_back(max_threads);
+  return this;
+}
+
+Benchmark* Benchmark::ThreadPerCpu() {
+  thread_counts_.push_back(CPUInfo::Get().num_cpus);
+  return this;
+}
+
+void Benchmark::SetName(const char* name) { name_ = name; }
+
+int Benchmark::ArgsCnt() const {
+  if (args_.empty()) {
+    if (arg_names_.empty()) return -1;
+    return static_cast<int>(arg_names_.size());
+  }
+  return static_cast<int>(args_.front().size());
+}
+
+//=============================================================================//
+//                            FunctionBenchmark
+//=============================================================================//
+
+void FunctionBenchmark::Run(State& st) { func_(st); }
+
+}  // end namespace internal
+
+void ClearRegisteredBenchmarks() {
+  internal::BenchmarkFamilies::GetInstance()->ClearBenchmarks();
+}
+
+}  // end namespace benchmark

benchmarks/thirdparty/benchmark/src/check.h

@@ -0,0 +1,79 @@
+#ifndef CHECK_H_
+#define CHECK_H_
+
+#include <cstdlib>
+#include <ostream>
+#include <cmath>
+
+#include "internal_macros.h"
+#include "log.h"
+
+namespace benchmark {
+namespace internal {
+
+typedef void(AbortHandlerT)();
+
+inline AbortHandlerT*& GetAbortHandler() {
+  static AbortHandlerT* handler = &std::abort;
+  return handler;
+}
+
+BENCHMARK_NORETURN inline void CallAbortHandler() {
+  GetAbortHandler()();
+  std::abort();  // fallback to enforce noreturn
+}
+
+// CheckHandler is the class constructed by failing CHECK macros. CheckHandler
+// will log information about the failures and abort when it is destructed.
+class CheckHandler {
+ public:
+  CheckHandler(const char* check, const char* file, const char* func, int line)
+      : log_(GetErrorLogInstance()) {
+    log_ << file << ":" << line << ": " << func << ": Check `" << check
+         << "' failed. ";
+  }
+
+  LogType& GetLog() { return log_; }
+
+  BENCHMARK_NORETURN ~CheckHandler() BENCHMARK_NOEXCEPT_OP(false) {
+    log_ << std::endl;
+    CallAbortHandler();
+  }
+
+  CheckHandler& operator=(const CheckHandler&) = delete;
+  CheckHandler(const CheckHandler&) = delete;
+  CheckHandler() = delete;
+
+ private:
+  LogType& log_;
+};
+
+}  // end namespace internal
+}  // end namespace benchmark
+
+// The CHECK macro returns a std::ostream object that can have extra information
+// written to it.
+#ifndef NDEBUG
+#define CHECK(b)                                                             \
+  (b ? ::benchmark::internal::GetNullLogInstance()                           \
+     : ::benchmark::internal::CheckHandler(#b, __FILE__, __func__, __LINE__) \
+           .GetLog())
+#else
+#define CHECK(b) ::benchmark::internal::GetNullLogInstance()
+#endif
+
+#define CHECK_EQ(a, b) CHECK((a) == (b))
+#define CHECK_NE(a, b) CHECK((a) != (b))
+#define CHECK_GE(a, b) CHECK((a) >= (b))
+#define CHECK_LE(a, b) CHECK((a) <= (b))
+#define CHECK_GT(a, b) CHECK((a) > (b))
+#define CHECK_LT(a, b) CHECK((a) < (b))
+
+#define CHECK_FLOAT_EQ(a, b, eps) CHECK(std::fabs((a) - (b)) <  (eps))
+#define CHECK_FLOAT_NE(a, b, eps) CHECK(std::fabs((a) - (b)) >= (eps))
+#define CHECK_FLOAT_GE(a, b, eps) CHECK((a) - (b) > -(eps))
+#define CHECK_FLOAT_LE(a, b, eps) CHECK((b) - (a) > -(eps))
+#define CHECK_FLOAT_GT(a, b, eps) CHECK((a) - (b) >  (eps))
+#define CHECK_FLOAT_LT(a, b, eps) CHECK((b) - (a) >  (eps))
+
+#endif  // CHECK_H_

benchmarks/thirdparty/benchmark/src/colorprint.cc

@@ -0,0 +1,188 @@
+// Copyright 2015 Google Inc. All rights reserved.
+//
+// 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
+//
+//     http://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 "colorprint.h"
+
+#include <cstdarg>
+#include <cstdio>
+#include <cstdlib>
+#include <cstring>
+#include <memory>
+#include <string>
+
+#include "check.h"
+#include "internal_macros.h"
+
+#ifdef BENCHMARK_OS_WINDOWS
+#include <Windows.h>
+#include <io.h>
+#else
+#include <unistd.h>
+#endif  // BENCHMARK_OS_WINDOWS
+
+namespace benchmark {
+namespace {
+#ifdef BENCHMARK_OS_WINDOWS
+typedef WORD PlatformColorCode;
+#else
+typedef const char* PlatformColorCode;
+#endif
+
+PlatformColorCode GetPlatformColorCode(LogColor color) {
+#ifdef BENCHMARK_OS_WINDOWS
+  switch (color) {
+    case COLOR_RED:
+      return FOREGROUND_RED;
+    case COLOR_GREEN:
+      return FOREGROUND_GREEN;
+    case COLOR_YELLOW:
+      return FOREGROUND_RED | FOREGROUND_GREEN;
+    case COLOR_BLUE:
+      return FOREGROUND_BLUE;
+    case COLOR_MAGENTA:
+      return FOREGROUND_BLUE | FOREGROUND_RED;
+    case COLOR_CYAN:
+      return FOREGROUND_BLUE | FOREGROUND_GREEN;
+    case COLOR_WHITE:  // fall through to default
+    default:
+      return 0;
+  }
+#else
+  switch (color) {
+    case COLOR_RED:
+      return "1";
+    case COLOR_GREEN:
+      return "2";
+    case COLOR_YELLOW:
+      return "3";
+    case COLOR_BLUE:
+      return "4";
+    case COLOR_MAGENTA:
+      return "5";
+    case COLOR_CYAN:
+      return "6";
+    case COLOR_WHITE:
+      return "7";
+    default:
+      return nullptr;
+  };
+#endif
+}
+
+}  // end namespace
+
+std::string FormatString(const char* msg, va_list args) {
+  // we might need a second shot at this, so pre-emptivly make a copy
+  va_list args_cp;
+  va_copy(args_cp, args);
+
+  std::size_t size = 256;
+  char local_buff[256];
+  auto ret = vsnprintf(local_buff, size, msg, args_cp);
+
+  va_end(args_cp);
+
+  // currently there is no error handling for failure, so this is hack.
+  CHECK(ret >= 0);
+
+  if (ret == 0)  // handle empty expansion
+    return {};
+  else if (static_cast<size_t>(ret) < size)
+    return local_buff;
+  else {
+    // we did not provide a long enough buffer on our first attempt.
+    size = (size_t)ret + 1;  // + 1 for the null byte
+    std::unique_ptr<char[]> buff(new char[size]);
+    ret = vsnprintf(buff.get(), size, msg, args);
+    CHECK(ret > 0 && ((size_t)ret) < size);
+    return buff.get();
+  }
+}
+
+std::string FormatString(const char* msg, ...) {
+  va_list args;
+  va_start(args, msg);
+  auto tmp = FormatString(msg, args);
+  va_end(args);
+  return tmp;
+}
+
+void ColorPrintf(std::ostream& out, LogColor color, const char* fmt, ...) {
+  va_list args;
+  va_start(args, fmt);
+  ColorPrintf(out, color, fmt, args);
+  va_end(args);
+}
+
+void ColorPrintf(std::ostream& out, LogColor color, const char* fmt,
+                 va_list args) {
+#ifdef BENCHMARK_OS_WINDOWS
+  ((void)out);  // suppress unused warning
+
+  const HANDLE stdout_handle = GetStdHandle(STD_OUTPUT_HANDLE);
+
+  // Gets the current text color.
+  CONSOLE_SCREEN_BUFFER_INFO buffer_info;
+  GetConsoleScreenBufferInfo(stdout_handle, &buffer_info);
+  const WORD old_color_attrs = buffer_info.wAttributes;
+
+  // We need to flush the stream buffers into the console before each
+  // SetConsoleTextAttribute call lest it affect the text that is already
+  // printed but has not yet reached the console.
+  fflush(stdout);
+  SetConsoleTextAttribute(stdout_handle,
+                          GetPlatformColorCode(color) | FOREGROUND_INTENSITY);
+  vprintf(fmt, args);
+
+  fflush(stdout);
+  // Restores the text color.
+  SetConsoleTextAttribute(stdout_handle, old_color_attrs);
+#else
+  const char* color_code = GetPlatformColorCode(color);
+  if (color_code) out << FormatString("\033[0;3%sm", color_code);
+  out << FormatString(fmt, args) << "\033[m";
+#endif
+}
+
+bool IsColorTerminal() {
+#if BENCHMARK_OS_WINDOWS
+  // On Windows the TERM variable is usually not set, but the
+  // console there does support colors.
+  return 0 != _isatty(_fileno(stdout));
+#else
+  // On non-Windows platforms, we rely on the TERM variable. This list of
+  // supported TERM values is copied from Google Test:
+  // <https://github.com/google/googletest/blob/master/googletest/src/gtest.cc#L2925>.
+  const char* const SUPPORTED_TERM_VALUES[] = {
+      "xterm",         "xterm-color",     "xterm-256color",
+      "screen",        "screen-256color", "tmux",
+      "tmux-256color", "rxvt-unicode",    "rxvt-unicode-256color",
+      "linux",         "cygwin",
+  };
+
+  const char* const term = getenv("TERM");
+
+  bool term_supports_color = false;
+  for (const char* candidate : SUPPORTED_TERM_VALUES) {
+    if (term && 0 == strcmp(term, candidate)) {
+      term_supports_color = true;
+      break;
+    }
+  }
+
+  return 0 != isatty(fileno(stdout)) && term_supports_color;
+#endif  // BENCHMARK_OS_WINDOWS
+}
+
+}  // end namespace benchmark

benchmarks/thirdparty/benchmark/src/colorprint.h

@@ -0,0 +1,33 @@
+#ifndef BENCHMARK_COLORPRINT_H_
+#define BENCHMARK_COLORPRINT_H_
+
+#include <cstdarg>
+#include <iostream>
+#include <string>
+
+namespace benchmark {
+enum LogColor {
+  COLOR_DEFAULT,
+  COLOR_RED,
+  COLOR_GREEN,
+  COLOR_YELLOW,
+  COLOR_BLUE,
+  COLOR_MAGENTA,
+  COLOR_CYAN,
+  COLOR_WHITE
+};
+
+std::string FormatString(const char* msg, va_list args);
+std::string FormatString(const char* msg, ...);
+
+void ColorPrintf(std::ostream& out, LogColor color, const char* fmt,
+                 va_list args);
+void ColorPrintf(std::ostream& out, LogColor color, const char* fmt, ...);
+
+// Returns true if stdout appears to be a terminal that supports colored
+// output, false otherwise.
+bool IsColorTerminal();
+
+}  // end namespace benchmark
+
+#endif  // BENCHMARK_COLORPRINT_H_

benchmarks/thirdparty/benchmark/src/commandlineflags.cc

@@ -0,0 +1,218 @@
+// Copyright 2015 Google Inc. All rights reserved.
+//
+// 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
+//
+//     http://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 "commandlineflags.h"
+
+#include <cctype>
+#include <cstdlib>
+#include <cstring>
+#include <iostream>
+#include <limits>
+
+namespace benchmark {
+// Parses 'str' for a 32-bit signed integer.  If successful, writes
+// the result to *value and returns true; otherwise leaves *value
+// unchanged and returns false.
+bool ParseInt32(const std::string& src_text, const char* str, int32_t* value) {
+  // Parses the environment variable as a decimal integer.
+  char* end = nullptr;
+  const long long_value = strtol(str, &end, 10);  // NOLINT
+
+  // Has strtol() consumed all characters in the string?
+  if (*end != '\0') {
+    // No - an invalid character was encountered.
+    std::cerr << src_text << " is expected to be a 32-bit integer, "
+              << "but actually has value \"" << str << "\".\n";
+    return false;
+  }
+
+  // Is the parsed value in the range of an Int32?
+  const int32_t result = static_cast<int32_t>(long_value);
+  if (long_value == std::numeric_limits<long>::max() ||
+      long_value == std::numeric_limits<long>::min() ||
+      // The parsed value overflows as a long.  (strtol() returns
+      // LONG_MAX or LONG_MIN when the input overflows.)
+      result != long_value
+      // The parsed value overflows as an Int32.
+      ) {
+    std::cerr << src_text << " is expected to be a 32-bit integer, "
+              << "but actually has value \"" << str << "\", "
+              << "which overflows.\n";
+    return false;
+  }
+
+  *value = result;
+  return true;
+}
+
+// Parses 'str' for a double.  If successful, writes the result to *value and
+// returns true; otherwise leaves *value unchanged and returns false.
+bool ParseDouble(const std::string& src_text, const char* str, double* value) {
+  // Parses the environment variable as a decimal integer.
+  char* end = nullptr;
+  const double double_value = strtod(str, &end);  // NOLINT
+
+  // Has strtol() consumed all characters in the string?
+  if (*end != '\0') {
+    // No - an invalid character was encountered.
+    std::cerr << src_text << " is expected to be a double, "
+              << "but actually has value \"" << str << "\".\n";
+    return false;
+  }
+
+  *value = double_value;
+  return true;
+}
+
+// Returns the name of the environment variable corresponding to the
+// given flag.  For example, FlagToEnvVar("foo") will return
+// "BENCHMARK_FOO" in the open-source version.
+static std::string FlagToEnvVar(const char* flag) {
+  const std::string flag_str(flag);
+
+  std::string env_var;
+  for (size_t i = 0; i != flag_str.length(); ++i)
+    env_var += static_cast<char>(::toupper(flag_str.c_str()[i]));
+
+  return "BENCHMARK_" + env_var;
+}
+
+// Reads and returns the Boolean environment variable corresponding to
+// the given flag; if it's not set, returns default_value.
+//
+// The value is considered true iff it's not "0".
+bool BoolFromEnv(const char* flag, bool default_value) {
+  const std::string env_var = FlagToEnvVar(flag);
+  const char* const string_value = getenv(env_var.c_str());
+  return string_value == nullptr ? default_value
+                                 : strcmp(string_value, "0") != 0;
+}
+
+// Reads and returns a 32-bit integer stored in the environment
+// variable corresponding to the given flag; if it isn't set or
+// doesn't represent a valid 32-bit integer, returns default_value.
+int32_t Int32FromEnv(const char* flag, int32_t default_value) {
+  const std::string env_var = FlagToEnvVar(flag);
+  const char* const string_value = getenv(env_var.c_str());
+  if (string_value == nullptr) {
+    // The environment variable is not set.
+    return default_value;
+  }
+
+  int32_t result = default_value;
+  if (!ParseInt32(std::string("Environment variable ") + env_var, string_value,
+                  &result)) {
+    std::cout << "The default value " << default_value << " is used.\n";
+    return default_value;
+  }
+
+  return result;
+}
+
+// Reads and returns the string environment variable corresponding to
+// the given flag; if it's not set, returns default_value.
+const char* StringFromEnv(const char* flag, const char* default_value) {
+  const std::string env_var = FlagToEnvVar(flag);
+  const char* const value = getenv(env_var.c_str());
+  return value == nullptr ? default_value : value;
+}
+
+// Parses a string as a command line flag.  The string should have
+// the format "--flag=value".  When def_optional is true, the "=value"
+// part can be omitted.
+//
+// Returns the value of the flag, or nullptr if the parsing failed.
+const char* ParseFlagValue(const char* str, const char* flag,
+                           bool def_optional) {
+  // str and flag must not be nullptr.
+  if (str == nullptr || flag == nullptr) return nullptr;
+
+  // The flag must start with "--".
+  const std::string flag_str = std::string("--") + std::string(flag);
+  const size_t flag_len = flag_str.length();
+  if (strncmp(str, flag_str.c_str(), flag_len) != 0) return nullptr;
+
+  // Skips the flag name.
+  const char* flag_end = str + flag_len;
+
+  // When def_optional is true, it's OK to not have a "=value" part.
+  if (def_optional && (flag_end[0] == '\0')) return flag_end;
+
+  // If def_optional is true and there are more characters after the
+  // flag name, or if def_optional is false, there must be a '=' after
+  // the flag name.
+  if (flag_end[0] != '=') return nullptr;
+
+  // Returns the string after "=".
+  return flag_end + 1;
+}
+
+bool ParseBoolFlag(const char* str, const char* flag, bool* value) {
+  // Gets the value of the flag as a string.
+  const char* const value_str = ParseFlagValue(str, flag, true);
+
+  // Aborts if the parsing failed.
+  if (value_str == nullptr) return false;
+
+  // Converts the string value to a bool.
+  *value = IsTruthyFlagValue(value_str);
+  return true;
+}
+
+bool ParseInt32Flag(const char* str, const char* flag, int32_t* value) {
+  // Gets the value of the flag as a string.
+  const char* const value_str = ParseFlagValue(str, flag, false);
+
+  // Aborts if the parsing failed.
+  if (value_str == nullptr) return false;
+
+  // Sets *value to the value of the flag.
+  return ParseInt32(std::string("The value of flag --") + flag, value_str,
+                    value);
+}
+
+bool ParseDoubleFlag(const char* str, const char* flag, double* value) {
+  // Gets the value of the flag as a string.
+  const char* const value_str = ParseFlagValue(str, flag, false);
+
+  // Aborts if the parsing failed.
+  if (value_str == nullptr) return false;
+
+  // Sets *value to the value of the flag.
+  return ParseDouble(std::string("The value of flag --") + flag, value_str,
+                     value);
+}
+
+bool ParseStringFlag(const char* str, const char* flag, std::string* value) {
+  // Gets the value of the flag as a string.
+  const char* const value_str = ParseFlagValue(str, flag, false);
+
+  // Aborts if the parsing failed.
+  if (value_str == nullptr) return false;
+
+  *value = value_str;
+  return true;
+}
+
+bool IsFlag(const char* str, const char* flag) {
+  return (ParseFlagValue(str, flag, true) != nullptr);
+}
+
+bool IsTruthyFlagValue(const std::string& value) {
+  if (value.empty()) return true;
+  char ch = value[0];
+  return isalnum(ch) &&
+         !(ch == '0' || ch == 'f' || ch == 'F' || ch == 'n' || ch == 'N');
+}
+}  // end namespace benchmark

benchmarks/thirdparty/benchmark/src/commandlineflags.h

@@ -0,0 +1,79 @@
+#ifndef BENCHMARK_COMMANDLINEFLAGS_H_
+#define BENCHMARK_COMMANDLINEFLAGS_H_
+
+#include <cstdint>
+#include <string>
+
+// Macro for referencing flags.
+#define FLAG(name) FLAGS_##name
+
+// Macros for declaring flags.
+#define DECLARE_bool(name) extern bool FLAG(name)
+#define DECLARE_int32(name) extern int32_t FLAG(name)
+#define DECLARE_int64(name) extern int64_t FLAG(name)
+#define DECLARE_double(name) extern double FLAG(name)
+#define DECLARE_string(name) extern std::string FLAG(name)
+
+// Macros for defining flags.
+#define DEFINE_bool(name, default_val, doc) bool FLAG(name) = (default_val)
+#define DEFINE_int32(name, default_val, doc) int32_t FLAG(name) = (default_val)
+#define DEFINE_int64(name, default_val, doc) int64_t FLAG(name) = (default_val)
+#define DEFINE_double(name, default_val, doc) double FLAG(name) = (default_val)
+#define DEFINE_string(name, default_val, doc) \
+  std::string FLAG(name) = (default_val)
+
+namespace benchmark {
+// Parses 'str' for a 32-bit signed integer.  If successful, writes the result
+// to *value and returns true; otherwise leaves *value unchanged and returns
+// false.
+bool ParseInt32(const std::string& src_text, const char* str, int32_t* value);
+
+// Parses a bool/Int32/string from the environment variable
+// corresponding to the given Google Test flag.
+bool BoolFromEnv(const char* flag, bool default_val);
+int32_t Int32FromEnv(const char* flag, int32_t default_val);
+double DoubleFromEnv(const char* flag, double default_val);
+const char* StringFromEnv(const char* flag, const char* default_val);
+
+// Parses a string for a bool flag, in the form of either
+// "--flag=value" or "--flag".
+//
+// In the former case, the value is taken as true if it passes IsTruthyValue().
+//
+// In the latter case, the value is taken as true.
+//
+// On success, stores the value of the flag in *value, and returns
+// true.  On failure, returns false without changing *value.
+bool ParseBoolFlag(const char* str, const char* flag, bool* value);
+
+// Parses a string for an Int32 flag, in the form of
+// "--flag=value".
+//
+// On success, stores the value of the flag in *value, and returns
+// true.  On failure, returns false without changing *value.
+bool ParseInt32Flag(const char* str, const char* flag, int32_t* value);
+
+// Parses a string for a Double flag, in the form of
+// "--flag=value".
+//
+// On success, stores the value of the flag in *value, and returns
+// true.  On failure, returns false without changing *value.
+bool ParseDoubleFlag(const char* str, const char* flag, double* value);
+
+// Parses a string for a string flag, in the form of
+// "--flag=value".
+//
+// On success, stores the value of the flag in *value, and returns
+// true.  On failure, returns false without changing *value.
+bool ParseStringFlag(const char* str, const char* flag, std::string* value);
+
+// Returns true if the string matches the flag.
+bool IsFlag(const char* str, const char* flag);
+
+// Returns true unless value starts with one of: '0', 'f', 'F', 'n' or 'N', or
+// some non-alphanumeric character. As a special case, also returns true if
+// value is the empty string.
+bool IsTruthyFlagValue(const std::string& value);
+}  // end namespace benchmark
+
+#endif  // BENCHMARK_COMMANDLINEFLAGS_H_

benchmarks/thirdparty/benchmark/src/complexity.cc

@@ -0,0 +1,220 @@
+// Copyright 2016 Ismael Jimenez Martinez. All rights reserved.
+//
+// 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
+//
+//     http://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.
+
+// Source project : https://github.com/ismaelJimenez/cpp.leastsq
+// Adapted to be used with google benchmark
+
+#include "benchmark/benchmark.h"
+
+#include <algorithm>
+#include <cmath>
+#include "check.h"
+#include "complexity.h"
+
+namespace benchmark {
+
+// Internal function to calculate the different scalability forms
+BigOFunc* FittingCurve(BigO complexity) {
+  switch (complexity) {
+    case oN:
+      return [](int n) -> double { return n; };
+    case oNSquared:
+      return [](int n) -> double { return std::pow(n, 2); };
+    case oNCubed:
+      return [](int n) -> double { return std::pow(n, 3); };
+    case oLogN:
+      return [](int n) { return log2(n); };
+    case oNLogN:
+      return [](int n) { return n * log2(n); };
+    case o1:
+    default:
+      return [](int) { return 1.0; };
+  }
+}
+
+// Function to return an string for the calculated complexity
+std::string GetBigOString(BigO complexity) {
+  switch (complexity) {
+    case oN:
+      return "N";
+    case oNSquared:
+      return "N^2";
+    case oNCubed:
+      return "N^3";
+    case oLogN:
+      return "lgN";
+    case oNLogN:
+      return "NlgN";
+    case o1:
+      return "(1)";
+    default:
+      return "f(N)";
+  }
+}
+
+// Find the coefficient for the high-order term in the running time, by
+// minimizing the sum of squares of relative error, for the fitting curve
+// given by the lambda expresion.
+//   - n             : Vector containing the size of the benchmark tests.
+//   - time          : Vector containing the times for the benchmark tests.
+//   - fitting_curve : lambda expresion (e.g. [](int n) {return n; };).
+
+// For a deeper explanation on the algorithm logic, look the README file at
+// http://github.com/ismaelJimenez/Minimal-Cpp-Least-Squared-Fit
+
+LeastSq MinimalLeastSq(const std::vector<int>& n,
+                       const std::vector<double>& time,
+                       BigOFunc* fitting_curve) {
+  double sigma_gn = 0.0;
+  double sigma_gn_squared = 0.0;
+  double sigma_time = 0.0;
+  double sigma_time_gn = 0.0;
+
+  // Calculate least square fitting parameter
+  for (size_t i = 0; i < n.size(); ++i) {
+    double gn_i = fitting_curve(n[i]);
+    sigma_gn += gn_i;
+    sigma_gn_squared += gn_i * gn_i;
+    sigma_time += time[i];
+    sigma_time_gn += time[i] * gn_i;
+  }
+
+  LeastSq result;
+  result.complexity = oLambda;
+
+  // Calculate complexity.
+  result.coef = sigma_time_gn / sigma_gn_squared;
+
+  // Calculate RMS
+  double rms = 0.0;
+  for (size_t i = 0; i < n.size(); ++i) {
+    double fit = result.coef * fitting_curve(n[i]);
+    rms += pow((time[i] - fit), 2);
+  }
+
+  // Normalized RMS by the mean of the observed values
+  double mean = sigma_time / n.size();
+  result.rms = sqrt(rms / n.size()) / mean;
+
+  return result;
+}
+
+// Find the coefficient for the high-order term in the running time, by
+// minimizing the sum of squares of relative error.
+//   - n          : Vector containing the size of the benchmark tests.
+//   - time       : Vector containing the times for the benchmark tests.
+//   - complexity : If different than oAuto, the fitting curve will stick to
+//                  this one. If it is oAuto, it will be calculated the best
+//                  fitting curve.
+LeastSq MinimalLeastSq(const std::vector<int>& n,
+                       const std::vector<double>& time, const BigO complexity) {
+  CHECK_EQ(n.size(), time.size());
+  CHECK_GE(n.size(), 2);  // Do not compute fitting curve is less than two
+                          // benchmark runs are given
+  CHECK_NE(complexity, oNone);
+
+  LeastSq best_fit;
+
+  if (complexity == oAuto) {
+    std::vector<BigO> fit_curves = {oLogN, oN, oNLogN, oNSquared, oNCubed};
+
+    // Take o1 as default best fitting curve
+    best_fit = MinimalLeastSq(n, time, FittingCurve(o1));
+    best_fit.complexity = o1;
+
+    // Compute all possible fitting curves and stick to the best one
+    for (const auto& fit : fit_curves) {
+      LeastSq current_fit = MinimalLeastSq(n, time, FittingCurve(fit));
+      if (current_fit.rms < best_fit.rms) {
+        best_fit = current_fit;
+        best_fit.complexity = fit;
+      }
+    }
+  } else {
+    best_fit = MinimalLeastSq(n, time, FittingCurve(complexity));
+    best_fit.complexity = complexity;
+  }
+
+  return best_fit;
+}
+
+std::vector<BenchmarkReporter::Run> ComputeBigO(
+    const std::vector<BenchmarkReporter::Run>& reports) {
+  typedef BenchmarkReporter::Run Run;
+  std::vector<Run> results;
+
+  if (reports.size() < 2) return results;
+
+  // Accumulators.
+  std::vector<int> n;
+  std::vector<double> real_time;
+  std::vector<double> cpu_time;
+
+  // Populate the accumulators.
+  for (const Run& run : reports) {
+    CHECK_GT(run.complexity_n, 0) << "Did you forget to call SetComplexityN?";
+    n.push_back(run.complexity_n);
+    real_time.push_back(run.real_accumulated_time / run.iterations);
+    cpu_time.push_back(run.cpu_accumulated_time / run.iterations);
+  }
+
+  LeastSq result_cpu;
+  LeastSq result_real;
+
+  if (reports[0].complexity == oLambda) {
+    result_cpu = MinimalLeastSq(n, cpu_time, reports[0].complexity_lambda);
+    result_real = MinimalLeastSq(n, real_time, reports[0].complexity_lambda);
+  } else {
+    result_cpu = MinimalLeastSq(n, cpu_time, reports[0].complexity);
+    result_real = MinimalLeastSq(n, real_time, result_cpu.complexity);
+  }
+  std::string benchmark_name =
+      reports[0].benchmark_name.substr(0, reports[0].benchmark_name.find('/'));
+
+  // Get the data from the accumulator to BenchmarkReporter::Run's.
+  Run big_o;
+  big_o.benchmark_name = benchmark_name + "_BigO";
+  big_o.iterations = 0;
+  big_o.real_accumulated_time = result_real.coef;
+  big_o.cpu_accumulated_time = result_cpu.coef;
+  big_o.report_big_o = true;
+  big_o.complexity = result_cpu.complexity;
+
+  // All the time results are reported after being multiplied by the
+  // time unit multiplier. But since RMS is a relative quantity it
+  // should not be multiplied at all. So, here, we _divide_ it by the
+  // multiplier so that when it is multiplied later the result is the
+  // correct one.
+  double multiplier = GetTimeUnitMultiplier(reports[0].time_unit);
+
+  // Only add label to mean/stddev if it is same for all runs
+  Run rms;
+  big_o.report_label = reports[0].report_label;
+  rms.benchmark_name = benchmark_name + "_RMS";
+  rms.report_label = big_o.report_label;
+  rms.iterations = 0;
+  rms.real_accumulated_time = result_real.rms / multiplier;
+  rms.cpu_accumulated_time = result_cpu.rms / multiplier;
+  rms.report_rms = true;
+  rms.complexity = result_cpu.complexity;
+  // don't forget to keep the time unit, or we won't be able to
+  // recover the correct value.
+  rms.time_unit = reports[0].time_unit;
+
+  results.push_back(big_o);
+  results.push_back(rms);
+  return results;
+}
+
+}  // end namespace benchmark

benchmarks/thirdparty/benchmark/src/complexity.h

@@ -0,0 +1,55 @@
+// Copyright 2016 Ismael Jimenez Martinez. All rights reserved.
+//
+// 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
+//
+//     http://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.
+
+// Source project : https://github.com/ismaelJimenez/cpp.leastsq
+// Adapted to be used with google benchmark
+
+#ifndef COMPLEXITY_H_
+#define COMPLEXITY_H_
+
+#include <string>
+#include <vector>
+
+#include "benchmark/benchmark.h"
+
+namespace benchmark {
+
+// Return a vector containing the bigO and RMS information for the specified
+// list of reports. If 'reports.size() < 2' an empty vector is returned.
+std::vector<BenchmarkReporter::Run> ComputeBigO(
+    const std::vector<BenchmarkReporter::Run>& reports);
+
+// This data structure will contain the result returned by MinimalLeastSq
+//   - coef        : Estimated coeficient for the high-order term as
+//                   interpolated from data.
+//   - rms         : Normalized Root Mean Squared Error.
+//   - complexity  : Scalability form (e.g. oN, oNLogN). In case a scalability
+//                   form has been provided to MinimalLeastSq this will return
+//                   the same value. In case BigO::oAuto has been selected, this
+//                   parameter will return the best fitting curve detected.
+
+struct LeastSq {
+  LeastSq() : coef(0.0), rms(0.0), complexity(oNone) {}
+
+  double coef;
+  double rms;
+  BigO complexity;
+};
+
+// Function to return an string for the calculated complexity
+std::string GetBigOString(BigO complexity);
+
+}  // end namespace benchmark
+
+#endif  // COMPLEXITY_H_

benchmarks/thirdparty/benchmark/src/console_reporter.cc

@@ -0,0 +1,182 @@
+// Copyright 2015 Google Inc. All rights reserved.
+//
+// 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
+//
+//     http://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 "benchmark/benchmark.h"
+#include "complexity.h"
+#include "counter.h"
+
+#include <algorithm>
+#include <cstdint>
+#include <cstdio>
+#include <iostream>
+#include <string>
+#include <tuple>
+#include <vector>
+
+#include "check.h"
+#include "colorprint.h"
+#include "commandlineflags.h"
+#include "internal_macros.h"
+#include "string_util.h"
+#include "timers.h"
+
+namespace benchmark {
+
+bool ConsoleReporter::ReportContext(const Context& context) {
+  name_field_width_ = context.name_field_width;
+  printed_header_ = false;
+  prev_counters_.clear();
+
+  PrintBasicContext(&GetErrorStream(), context);
+
+#ifdef BENCHMARK_OS_WINDOWS
+  if ((output_options_ & OO_Color) && &std::cout != &GetOutputStream()) {
+    GetErrorStream()
+        << "Color printing is only supported for stdout on windows."
+           " Disabling color printing\n";
+    output_options_ = static_cast< OutputOptions >(output_options_ & ~OO_Color);
+  }
+#endif
+
+  return true;
+}
+
+void ConsoleReporter::PrintHeader(const Run& run) {
+  std::string str = FormatString("%-*s %13s %13s %10s", static_cast<int>(name_field_width_),
+                                 "Benchmark", "Time", "CPU", "Iterations");
+  if(!run.counters.empty()) {
+    if(output_options_ & OO_Tabular) {
+      for(auto const& c : run.counters) {
+        str += FormatString(" %10s", c.first.c_str());
+      }
+    } else {
+      str += " UserCounters...";
+    }
+  }
+  str += "\n";
+  std::string line = std::string(str.length(), '-');
+  GetOutputStream() << line << "\n" << str << line << "\n";
+}
+
+void ConsoleReporter::ReportRuns(const std::vector<Run>& reports) {
+  for (const auto& run : reports) {
+    // print the header:
+    // --- if none was printed yet
+    bool print_header = !printed_header_;
+    // --- or if the format is tabular and this run
+    //     has different fields from the prev header
+    print_header |= (output_options_ & OO_Tabular) &&
+                    (!internal::SameNames(run.counters, prev_counters_));
+    if (print_header) {
+      printed_header_ = true;
+      prev_counters_ = run.counters;
+      PrintHeader(run);
+    }
+    // As an alternative to printing the headers like this, we could sort
+    // the benchmarks by header and then print. But this would require
+    // waiting for the full results before printing, or printing twice.
+    PrintRunData(run);
+  }
+}
+
+static void IgnoreColorPrint(std::ostream& out, LogColor, const char* fmt,
+                             ...) {
+  va_list args;
+  va_start(args, fmt);
+  out << FormatString(fmt, args);
+  va_end(args);
+}
+
+void ConsoleReporter::PrintRunData(const Run& result) {
+  typedef void(PrinterFn)(std::ostream&, LogColor, const char*, ...);
+  auto& Out = GetOutputStream();
+  PrinterFn* printer = (output_options_ & OO_Color) ?
+                         (PrinterFn*)ColorPrintf : IgnoreColorPrint;
+  auto name_color =
+      (result.report_big_o || result.report_rms) ? COLOR_BLUE : COLOR_GREEN;
+  printer(Out, name_color, "%-*s ", name_field_width_,
+          result.benchmark_name.c_str());
+
+  if (result.error_occurred) {
+    printer(Out, COLOR_RED, "ERROR OCCURRED: \'%s\'",
+            result.error_message.c_str());
+    printer(Out, COLOR_DEFAULT, "\n");
+    return;
+  }
+  // Format bytes per second
+  std::string rate;
+  if (result.bytes_per_second > 0) {
+    rate = StrCat(" ", HumanReadableNumber(result.bytes_per_second), "B/s");
+  }
+
+  // Format items per second
+  std::string items;
+  if (result.items_per_second > 0) {
+    items =
+        StrCat(" ", HumanReadableNumber(result.items_per_second), " items/s");
+  }
+
+  const double real_time = result.GetAdjustedRealTime();
+  const double cpu_time = result.GetAdjustedCPUTime();
+
+  if (result.report_big_o) {
+    std::string big_o = GetBigOString(result.complexity);
+    printer(Out, COLOR_YELLOW, "%10.2f %s %10.2f %s ", real_time, big_o.c_str(),
+            cpu_time, big_o.c_str());
+  } else if (result.report_rms) {
+    printer(Out, COLOR_YELLOW, "%10.0f %% %10.0f %% ", real_time * 100,
+            cpu_time * 100);
+  } else {
+    const char* timeLabel = GetTimeUnitString(result.time_unit);
+    printer(Out, COLOR_YELLOW, "%10.0f %s %10.0f %s ", real_time, timeLabel,
+            cpu_time, timeLabel);
+  }
+
+  if (!result.report_big_o && !result.report_rms) {
+    printer(Out, COLOR_CYAN, "%10lld", result.iterations);
+  }
+
+  for (auto& c : result.counters) {
+    const std::size_t cNameLen = std::max(std::string::size_type(10),
+                                          c.first.length());
+    auto const& s = HumanReadableNumber(c.second.value, 1000);
+    if (output_options_ & OO_Tabular) {
+      if (c.second.flags & Counter::kIsRate) {
+        printer(Out, COLOR_DEFAULT, " %*s/s", cNameLen - 2, s.c_str());
+      } else {
+        printer(Out, COLOR_DEFAULT, " %*s", cNameLen, s.c_str());
+      }
+    } else {
+      const char* unit = (c.second.flags & Counter::kIsRate) ? "/s" : "";
+      printer(Out, COLOR_DEFAULT, " %s=%s%s", c.first.c_str(), s.c_str(),
+              unit);
+    }
+  }
+
+  if (!rate.empty()) {
+    printer(Out, COLOR_DEFAULT, " %*s", 13, rate.c_str());
+  }
+
+  if (!items.empty()) {
+    printer(Out, COLOR_DEFAULT, " %*s", 18, items.c_str());
+  }
+
+  if (!result.report_label.empty()) {
+    printer(Out, COLOR_DEFAULT, " %s", result.report_label.c_str());
+  }
+
+  printer(Out, COLOR_DEFAULT, "\n");
+}
+
+}  // end namespace benchmark

benchmarks/thirdparty/benchmark/src/counter.cc

@@ -0,0 +1,68 @@
+// Copyright 2015 Google Inc. All rights reserved.
+//
+// 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
+//
+//     http://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 "counter.h"
+
+namespace benchmark {
+namespace internal {
+
+double Finish(Counter const& c, double cpu_time, double num_threads) {
+  double v = c.value;
+  if (c.flags & Counter::kIsRate) {
+    v /= cpu_time;
+  }
+  if (c.flags & Counter::kAvgThreads) {
+    v /= num_threads;
+  }
+  return v;
+}
+
+void Finish(UserCounters *l, double cpu_time, double num_threads) {
+  for (auto &c : *l) {
+    c.second.value = Finish(c.second, cpu_time, num_threads);
+  }
+}
+
+void Increment(UserCounters *l, UserCounters const& r) {
+  // add counters present in both or just in *l
+  for (auto &c : *l) {
+    auto it = r.find(c.first);
+    if (it != r.end()) {
+      c.second.value = c.second + it->second;
+    }
+  }
+  // add counters present in r, but not in *l
+  for (auto const &tc : r) {
+    auto it = l->find(tc.first);
+    if (it == l->end()) {
+      (*l)[tc.first] = tc.second;
+    }
+  }
+}
+
+bool SameNames(UserCounters const& l, UserCounters const& r) {
+  if (&l == &r) return true;
+  if (l.size() != r.size()) {
+    return false;
+  }
+  for (auto const& c : l) {
+    if (r.find(c.first) == r.end()) {
+      return false;
+    }
+  }
+  return true;
+}
+
+} // end namespace internal
+} // end namespace benchmark

benchmarks/thirdparty/benchmark/src/counter.h

@@ -0,0 +1,26 @@
+// Copyright 2015 Google Inc. All rights reserved.
+//
+// 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
+//
+//     http://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 "benchmark/benchmark.h"
+
+namespace benchmark {
+
+// these counter-related functions are hidden to reduce API surface.
+namespace internal {
+void Finish(UserCounters *l, double time, double num_threads);
+void Increment(UserCounters *l, UserCounters const& r);
+bool SameNames(UserCounters const& l, UserCounters const& r);
+} // end namespace internal
+
+} //end namespace benchmark

benchmarks/thirdparty/benchmark/src/csv_reporter.cc

@@ -0,0 +1,149 @@
+// Copyright 2015 Google Inc. All rights reserved.
+//
+// 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
+//
+//     http://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 "benchmark/benchmark.h"
+#include "complexity.h"
+
+#include <algorithm>
+#include <cstdint>
+#include <iostream>
+#include <string>
+#include <tuple>
+#include <vector>
+
+#include "string_util.h"
+#include "timers.h"
+#include "check.h"
+
+// File format reference: http://edoceo.com/utilitas/csv-file-format.
+
+namespace benchmark {
+
+namespace {
+std::vector<std::string> elements = {
+    "name",           "iterations",       "real_time",        "cpu_time",
+    "time_unit",      "bytes_per_second", "items_per_second", "label",
+    "error_occurred", "error_message"};
+}  // namespace
+
+bool CSVReporter::ReportContext(const Context& context) {
+  PrintBasicContext(&GetErrorStream(), context);
+  return true;
+}
+
+void CSVReporter::ReportRuns(const std::vector<Run> & reports) {
+  std::ostream& Out = GetOutputStream();
+
+  if (!printed_header_) {
+    // save the names of all the user counters
+    for (const auto& run : reports) {
+      for (const auto& cnt : run.counters) {
+        user_counter_names_.insert(cnt.first);
+      }
+    }
+
+    // print the header
+    for (auto B = elements.begin(); B != elements.end();) {
+      Out << *B++;
+      if (B != elements.end()) Out << ",";
+    }
+    for (auto B = user_counter_names_.begin(); B != user_counter_names_.end();) {
+      Out << ",\"" << *B++ << "\"";
+    }
+    Out << "\n";
+
+    printed_header_ = true;
+  } else {
+    // check that all the current counters are saved in the name set
+    for (const auto& run : reports) {
+      for (const auto& cnt : run.counters) {
+        CHECK(user_counter_names_.find(cnt.first) != user_counter_names_.end())
+              << "All counters must be present in each run. "
+              << "Counter named \"" << cnt.first
+              << "\" was not in a run after being added to the header";
+      }
+    }
+  }
+
+  // print results for each run
+  for (const auto& run : reports) {
+    PrintRunData(run);
+  }
+
+}
+
+void CSVReporter::PrintRunData(const Run & run) {
+  std::ostream& Out = GetOutputStream();
+
+  // Field with embedded double-quote characters must be doubled and the field
+  // delimited with double-quotes.
+  std::string name = run.benchmark_name;
+  ReplaceAll(&name, "\"", "\"\"");
+  Out << '"' << name << "\",";
+  if (run.error_occurred) {
+    Out << std::string(elements.size() - 3, ',');
+    Out << "true,";
+    std::string msg = run.error_message;
+    ReplaceAll(&msg, "\"", "\"\"");
+    Out << '"' << msg << "\"\n";
+    return;
+  }
+
+  // Do not print iteration on bigO and RMS report
+  if (!run.report_big_o && !run.report_rms) {
+    Out << run.iterations;
+  }
+  Out << ",";
+
+  Out << run.GetAdjustedRealTime() << ",";
+  Out << run.GetAdjustedCPUTime() << ",";
+
+  // Do not print timeLabel on bigO and RMS report
+  if (run.report_big_o) {
+    Out << GetBigOString(run.complexity);
+  } else if (!run.report_rms) {
+    Out << GetTimeUnitString(run.time_unit);
+  }
+  Out << ",";
+
+  if (run.bytes_per_second > 0.0) {
+    Out << run.bytes_per_second;
+  }
+  Out << ",";
+  if (run.items_per_second > 0.0) {
+    Out << run.items_per_second;
+  }
+  Out << ",";
+  if (!run.report_label.empty()) {
+    // Field with embedded double-quote characters must be doubled and the field
+    // delimited with double-quotes.
+    std::string label = run.report_label;
+    ReplaceAll(&label, "\"", "\"\"");
+    Out << "\"" << label << "\"";
+  }
+  Out << ",,";  // for error_occurred and error_message
+
+  // Print user counters
+  for (const auto &ucn : user_counter_names_) {
+    auto it = run.counters.find(ucn);
+    if(it == run.counters.end()) {
+      Out << ",";
+    } else {
+      Out << "," << it->second;
+    }
+  }
+  Out << '\n';
+}
+
+}  // end namespace benchmark

benchmarks/thirdparty/benchmark/src/cycleclock.h

@@ -0,0 +1,172 @@
+// ----------------------------------------------------------------------
+// CycleClock
+//    A CycleClock tells you the current time in Cycles.  The "time"
+//    is actually time since power-on.  This is like time() but doesn't
+//    involve a system call and is much more precise.
+//
+// NOTE: Not all cpu/platform/kernel combinations guarantee that this
+// clock increments at a constant rate or is synchronized across all logical
+// cpus in a system.
+//
+// If you need the above guarantees, please consider using a different
+// API. There are efforts to provide an interface which provides a millisecond
+// granularity and implemented as a memory read. A memory read is generally
+// cheaper than the CycleClock for many architectures.
+//
+// Also, in some out of order CPU implementations, the CycleClock is not
+// serializing. So if you're trying to count at cycles granularity, your
+// data might be inaccurate due to out of order instruction execution.
+// ----------------------------------------------------------------------
+
+#ifndef BENCHMARK_CYCLECLOCK_H_
+#define BENCHMARK_CYCLECLOCK_H_
+
+#include <cstdint>
+
+#include "benchmark/benchmark.h"
+#include "internal_macros.h"
+
+#if defined(BENCHMARK_OS_MACOSX)
+#include <mach/mach_time.h>
+#endif
+// For MSVC, we want to use '_asm rdtsc' when possible (since it works
+// with even ancient MSVC compilers), and when not possible the
+// __rdtsc intrinsic, declared in <intrin.h>.  Unfortunately, in some
+// environments, <windows.h> and <intrin.h> have conflicting
+// declarations of some other intrinsics, breaking compilation.
+// Therefore, we simply declare __rdtsc ourselves. See also
+// http://connect.microsoft.com/VisualStudio/feedback/details/262047
+#if defined(COMPILER_MSVC) && !defined(_M_IX86)
+extern "C" uint64_t __rdtsc();
+#pragma intrinsic(__rdtsc)
+#endif
+
+#ifndef BENCHMARK_OS_WINDOWS
+#include <sys/time.h>
+#include <time.h>
+#endif
+
+#ifdef BENCHMARK_OS_EMSCRIPTEN
+#include <emscripten.h>
+#endif
+
+namespace benchmark {
+// NOTE: only i386 and x86_64 have been well tested.
+// PPC, sparc, alpha, and ia64 are based on
+//    http://peter.kuscsik.com/wordpress/?p=14
+// with modifications by m3b.  See also
+//    https://setisvn.ssl.berkeley.edu/svn/lib/fftw-3.0.1/kernel/cycle.h
+namespace cycleclock {
+// This should return the number of cycles since power-on.  Thread-safe.
+inline BENCHMARK_ALWAYS_INLINE int64_t Now() {
+#if defined(BENCHMARK_OS_MACOSX)
+  // this goes at the top because we need ALL Macs, regardless of
+  // architecture, to return the number of "mach time units" that
+  // have passed since startup.  See sysinfo.cc where
+  // InitializeSystemInfo() sets the supposed cpu clock frequency of
+  // macs to the number of mach time units per second, not actual
+  // CPU clock frequency (which can change in the face of CPU
+  // frequency scaling).  Also note that when the Mac sleeps, this
+  // counter pauses; it does not continue counting, nor does it
+  // reset to zero.
+  return mach_absolute_time();
+#elif defined(BENCHMARK_OS_EMSCRIPTEN)
+  // this goes above x86-specific code because old versions of Emscripten
+  // define __x86_64__, although they have nothing to do with it.
+  return static_cast<int64_t>(emscripten_get_now() * 1e+6);
+#elif defined(__i386__)
+  int64_t ret;
+  __asm__ volatile("rdtsc" : "=A"(ret));
+  return ret;
+#elif defined(__x86_64__) || defined(__amd64__)
+  uint64_t low, high;
+  __asm__ volatile("rdtsc" : "=a"(low), "=d"(high));
+  return (high << 32) | low;
+#elif defined(__powerpc__) || defined(__ppc__)
+  // This returns a time-base, which is not always precisely a cycle-count.
+  int64_t tbl, tbu0, tbu1;
+  asm("mftbu %0" : "=r"(tbu0));
+  asm("mftb  %0" : "=r"(tbl));
+  asm("mftbu %0" : "=r"(tbu1));
+  tbl &= -static_cast<int64_t>(tbu0 == tbu1);
+  // high 32 bits in tbu1; low 32 bits in tbl  (tbu0 is garbage)
+  return (tbu1 << 32) | tbl;
+#elif defined(__sparc__)
+  int64_t tick;
+  asm(".byte 0x83, 0x41, 0x00, 0x00");
+  asm("mov   %%g1, %0" : "=r"(tick));
+  return tick;
+#elif defined(__ia64__)
+  int64_t itc;
+  asm("mov %0 = ar.itc" : "=r"(itc));
+  return itc;
+#elif defined(COMPILER_MSVC) && defined(_M_IX86)
+  // Older MSVC compilers (like 7.x) don't seem to support the
+  // __rdtsc intrinsic properly, so I prefer to use _asm instead
+  // when I know it will work.  Otherwise, I'll use __rdtsc and hope
+  // the code is being compiled with a non-ancient compiler.
+  _asm rdtsc
+#elif defined(COMPILER_MSVC)
+  return __rdtsc();
+#elif defined(BENCHMARK_OS_NACL)
+  // Native Client validator on x86/x86-64 allows RDTSC instructions,
+  // and this case is handled above. Native Client validator on ARM
+  // rejects MRC instructions (used in the ARM-specific sequence below),
+  // so we handle it here. Portable Native Client compiles to
+  // architecture-agnostic bytecode, which doesn't provide any
+  // cycle counter access mnemonics.
+
+  // Native Client does not provide any API to access cycle counter.
+  // Use clock_gettime(CLOCK_MONOTONIC, ...) instead of gettimeofday
+  // because is provides nanosecond resolution (which is noticable at
+  // least for PNaCl modules running on x86 Mac & Linux).
+  // Initialize to always return 0 if clock_gettime fails.
+  struct timespec ts = { 0, 0 };
+  clock_gettime(CLOCK_MONOTONIC, &ts);
+  return static_cast<int64_t>(ts.tv_sec) * 1000000000 + ts.tv_nsec;
+#elif defined(__aarch64__)
+  // System timer of ARMv8 runs at a different frequency than the CPU's.
+  // The frequency is fixed, typically in the range 1-50MHz.  It can be
+  // read at CNTFRQ special register.  We assume the OS has set up
+  // the virtual timer properly.
+  int64_t virtual_timer_value;
+  asm volatile("mrs %0, cntvct_el0" : "=r"(virtual_timer_value));
+  return virtual_timer_value;
+#elif defined(__ARM_ARCH)
+  // V6 is the earliest arch that has a standard cyclecount
+  // Native Client validator doesn't allow MRC instructions.
+#if (__ARM_ARCH >= 6)
+  uint32_t pmccntr;
+  uint32_t pmuseren;
+  uint32_t pmcntenset;
+  // Read the user mode perf monitor counter access permissions.
+  asm volatile("mrc p15, 0, %0, c9, c14, 0" : "=r"(pmuseren));
+  if (pmuseren & 1) {  // Allows reading perfmon counters for user mode code.
+    asm volatile("mrc p15, 0, %0, c9, c12, 1" : "=r"(pmcntenset));
+    if (pmcntenset & 0x80000000ul) {  // Is it counting?
+      asm volatile("mrc p15, 0, %0, c9, c13, 0" : "=r"(pmccntr));
+      // The counter is set up to count every 64th cycle
+      return static_cast<int64_t>(pmccntr) * 64;  // Should optimize to << 6
+    }
+  }
+#endif
+  struct timeval tv;
+  gettimeofday(&tv, nullptr);
+  return static_cast<int64_t>(tv.tv_sec) * 1000000 + tv.tv_usec;
+#elif defined(__mips__)
+  // mips apparently only allows rdtsc for superusers, so we fall
+  // back to gettimeofday.  It's possible clock_gettime would be better.
+  struct timeval tv;
+  gettimeofday(&tv, nullptr);
+  return static_cast<int64_t>(tv.tv_sec) * 1000000 + tv.tv_usec;
+#else
+// The soft failover to a generic implementation is automatic only for ARM.
+// For other platforms the developer is expected to make an attempt to create
+// a fast implementation and use generic version if nothing better is available.
+#error You need to define CycleTimer for your OS and CPU
+#endif
+}
+}  // end namespace cycleclock
+}  // end namespace benchmark
+
+#endif  // BENCHMARK_CYCLECLOCK_H_

benchmarks/thirdparty/benchmark/src/internal_macros.h

@@ -0,0 +1,82 @@
+#ifndef BENCHMARK_INTERNAL_MACROS_H_
+#define BENCHMARK_INTERNAL_MACROS_H_
+
+#include "benchmark/benchmark.h"
+
+#ifndef __has_feature
+#define __has_feature(x) 0
+#endif
+#ifndef __has_builtin
+#define __has_builtin(x) 0
+#endif
+
+#if defined(__clang__)
+  #if !defined(COMPILER_CLANG)
+    #define COMPILER_CLANG
+  #endif
+#elif defined(_MSC_VER)
+  #if !defined(COMPILER_MSVC)
+    #define COMPILER_MSVC
+  #endif
+#elif defined(__GNUC__)
+  #if !defined(COMPILER_GCC)
+    #define COMPILER_GCC
+  #endif
+#endif
+
+#if __has_feature(cxx_attributes)
+  #define BENCHMARK_NORETURN [[noreturn]]
+#elif defined(__GNUC__)
+  #define BENCHMARK_NORETURN __attribute__((noreturn))
+#elif defined(COMPILER_MSVC)
+  #define BENCHMARK_NORETURN __declspec(noreturn)
+#else
+  #define BENCHMARK_NORETURN
+#endif
+
+#if defined(__CYGWIN__)
+  #define BENCHMARK_OS_CYGWIN 1
+#elif defined(_WIN32)
+  #define BENCHMARK_OS_WINDOWS 1
+#elif defined(__APPLE__)
+  #include "TargetConditionals.h"
+  #if defined(TARGET_OS_MAC)
+    #define BENCHMARK_OS_MACOSX 1
+    #if defined(TARGET_OS_IPHONE)
+      #define BENCHMARK_OS_IOS 1
+    #endif
+  #endif
+#elif defined(__FreeBSD__)
+  #define BENCHMARK_OS_FREEBSD 1
+#elif defined(__NetBSD__)
+  #define BENCHMARK_OS_NETBSD 1
+#elif defined(__linux__)
+  #define BENCHMARK_OS_LINUX 1
+#elif defined(__native_client__)
+  #define BENCHMARK_OS_NACL 1
+#elif defined(EMSCRIPTEN)
+  #define BENCHMARK_OS_EMSCRIPTEN 1
+#elif defined(__rtems__)
+  #define BENCHMARK_OS_RTEMS 1
+#endif
+
+#if !__has_feature(cxx_exceptions) && !defined(__cpp_exceptions) \
+     && !defined(__EXCEPTIONS)
+  #define BENCHMARK_HAS_NO_EXCEPTIONS
+#endif
+
+#if defined(COMPILER_CLANG) || defined(COMPILER_GCC)
+  #define BENCHMARK_MAYBE_UNUSED __attribute__((unused))
+#else
+  #define BENCHMARK_MAYBE_UNUSED
+#endif
+
+#if defined(COMPILER_GCC) || __has_builtin(__builtin_unreachable)
+  #define BENCHMARK_UNREACHABLE() __builtin_unreachable()
+#elif defined(COMPILER_MSVC)
+  #define BENCHMARK_UNREACHABLE() __assume(false)
+#else
+  #define BENCHMARK_UNREACHABLE() ((void)0)
+#endif
+
+#endif  // BENCHMARK_INTERNAL_MACROS_H_

benchmarks/thirdparty/benchmark/src/json_reporter.cc

@@ -0,0 +1,201 @@
+// Copyright 2015 Google Inc. All rights reserved.
+//
+// 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
+//
+//     http://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 "benchmark/benchmark.h"
+#include "complexity.h"
+
+#include <algorithm>
+#include <cstdint>
+#include <iostream>
+#include <string>
+#include <tuple>
+#include <vector>
+#include <iomanip> // for setprecision
+#include <limits>
+
+#include "string_util.h"
+#include "timers.h"
+
+namespace benchmark {
+
+namespace {
+
+std::string FormatKV(std::string const& key, std::string const& value) {
+  return StringPrintF("\"%s\": \"%s\"", key.c_str(), value.c_str());
+}
+
+std::string FormatKV(std::string const& key, const char* value) {
+  return StringPrintF("\"%s\": \"%s\"", key.c_str(), value);
+}
+
+std::string FormatKV(std::string const& key, bool value) {
+  return StringPrintF("\"%s\": %s", key.c_str(), value ? "true" : "false");
+}
+
+std::string FormatKV(std::string const& key, int64_t value) {
+  std::stringstream ss;
+  ss << '"' << key << "\": " << value;
+  return ss.str();
+}
+
+std::string FormatKV(std::string const& key, double value) {
+  std::stringstream ss;
+  ss << '"' << key << "\": ";
+
+  const auto max_digits10 = std::numeric_limits<decltype (value)>::max_digits10;
+  const auto max_fractional_digits10 = max_digits10 - 1;
+
+  ss << std::scientific << std::setprecision(max_fractional_digits10) << value;
+  return ss.str();
+}
+
+int64_t RoundDouble(double v) { return static_cast<int64_t>(v + 0.5); }
+
+}  // end namespace
+
+bool JSONReporter::ReportContext(const Context& context) {
+  std::ostream& out = GetOutputStream();
+
+  out << "{\n";
+  std::string inner_indent(2, ' ');
+
+  // Open context block and print context information.
+  out << inner_indent << "\"context\": {\n";
+  std::string indent(4, ' ');
+
+  std::string walltime_value = LocalDateTimeString();
+  out << indent << FormatKV("date", walltime_value) << ",\n";
+
+  CPUInfo const& info = context.cpu_info;
+  out << indent << FormatKV("num_cpus", static_cast<int64_t>(info.num_cpus))
+      << ",\n";
+  out << indent
+      << FormatKV("mhz_per_cpu",
+                  RoundDouble(info.cycles_per_second / 1000000.0))
+      << ",\n";
+  out << indent << FormatKV("cpu_scaling_enabled", info.scaling_enabled)
+      << ",\n";
+
+  out << indent << "\"caches\": [\n";
+  indent = std::string(6, ' ');
+  std::string cache_indent(8, ' ');
+  for (size_t i = 0; i < info.caches.size(); ++i) {
+    auto& CI = info.caches[i];
+    out << indent << "{\n";
+    out << cache_indent << FormatKV("type", CI.type) << ",\n";
+    out << cache_indent << FormatKV("level", static_cast<int64_t>(CI.level))
+        << ",\n";
+    out << cache_indent
+        << FormatKV("size", static_cast<int64_t>(CI.size) * 1000u) << ",\n";
+    out << cache_indent
+        << FormatKV("num_sharing", static_cast<int64_t>(CI.num_sharing))
+        << "\n";
+    out << indent << "}";
+    if (i != info.caches.size() - 1) out << ",";
+    out << "\n";
+  }
+  indent = std::string(4, ' ');
+  out << indent << "],\n";
+
+#if defined(NDEBUG)
+  const char build_type[] = "release";
+#else
+  const char build_type[] = "debug";
+#endif
+  out << indent << FormatKV("library_build_type", build_type) << "\n";
+  // Close context block and open the list of benchmarks.
+  out << inner_indent << "},\n";
+  out << inner_indent << "\"benchmarks\": [\n";
+  return true;
+}
+
+void JSONReporter::ReportRuns(std::vector<Run> const& reports) {
+  if (reports.empty()) {
+    return;
+  }
+  std::string indent(4, ' ');
+  std::ostream& out = GetOutputStream();
+  if (!first_report_) {
+    out << ",\n";
+  }
+  first_report_ = false;
+
+  for (auto it = reports.begin(); it != reports.end(); ++it) {
+    out << indent << "{\n";
+    PrintRunData(*it);
+    out << indent << '}';
+    auto it_cp = it;
+    if (++it_cp != reports.end()) {
+      out << ",\n";
+    }
+  }
+}
+
+void JSONReporter::Finalize() {
+  // Close the list of benchmarks and the top level object.
+  GetOutputStream() << "\n  ]\n}\n";
+}
+
+void JSONReporter::PrintRunData(Run const& run) {
+  std::string indent(6, ' ');
+  std::ostream& out = GetOutputStream();
+  out << indent << FormatKV("name", run.benchmark_name) << ",\n";
+  if (run.error_occurred) {
+    out << indent << FormatKV("error_occurred", run.error_occurred) << ",\n";
+    out << indent << FormatKV("error_message", run.error_message) << ",\n";
+  }
+  if (!run.report_big_o && !run.report_rms) {
+    out << indent << FormatKV("iterations", run.iterations) << ",\n";
+    out << indent
+        << FormatKV("real_time", run.GetAdjustedRealTime())
+        << ",\n";
+    out << indent
+        << FormatKV("cpu_time", run.GetAdjustedCPUTime());
+    out << ",\n"
+        << indent << FormatKV("time_unit", GetTimeUnitString(run.time_unit));
+  } else if (run.report_big_o) {
+    out << indent
+        << FormatKV("cpu_coefficient", run.GetAdjustedCPUTime())
+        << ",\n";
+    out << indent
+        << FormatKV("real_coefficient", run.GetAdjustedRealTime())
+        << ",\n";
+    out << indent << FormatKV("big_o", GetBigOString(run.complexity)) << ",\n";
+    out << indent << FormatKV("time_unit", GetTimeUnitString(run.time_unit));
+  } else if (run.report_rms) {
+    out << indent
+        << FormatKV("rms", run.GetAdjustedCPUTime());
+  }
+  if (run.bytes_per_second > 0.0) {
+    out << ",\n"
+        << indent
+        << FormatKV("bytes_per_second", run.bytes_per_second);
+  }
+  if (run.items_per_second > 0.0) {
+    out << ",\n"
+        << indent
+        << FormatKV("items_per_second", run.items_per_second);
+  }
+  for(auto &c : run.counters) {
+    out << ",\n"
+        << indent
+        << FormatKV(c.first, c.second);
+  }
+  if (!run.report_label.empty()) {
+    out << ",\n" << indent << FormatKV("label", run.report_label);
+  }
+  out << '\n';
+}
+
+} // end namespace benchmark

benchmarks/thirdparty/benchmark/src/log.h

@@ -0,0 +1,73 @@
+#ifndef BENCHMARK_LOG_H_
+#define BENCHMARK_LOG_H_
+
+#include <iostream>
+#include <ostream>
+
+#include "benchmark/benchmark.h"
+
+namespace benchmark {
+namespace internal {
+
+typedef std::basic_ostream<char>&(EndLType)(std::basic_ostream<char>&);
+
+class LogType {
+  friend LogType& GetNullLogInstance();
+  friend LogType& GetErrorLogInstance();
+
+  // FIXME: Add locking to output.
+  template <class Tp>
+  friend LogType& operator<<(LogType&, Tp const&);
+  friend LogType& operator<<(LogType&, EndLType*);
+
+ private:
+  LogType(std::ostream* out) : out_(out) {}
+  std::ostream* out_;
+  BENCHMARK_DISALLOW_COPY_AND_ASSIGN(LogType);
+};
+
+template <class Tp>
+LogType& operator<<(LogType& log, Tp const& value) {
+  if (log.out_) {
+    *log.out_ << value;
+  }
+  return log;
+}
+
+inline LogType& operator<<(LogType& log, EndLType* m) {
+  if (log.out_) {
+    *log.out_ << m;
+  }
+  return log;
+}
+
+inline int& LogLevel() {
+  static int log_level = 0;
+  return log_level;
+}
+
+inline LogType& GetNullLogInstance() {
+  static LogType log(nullptr);
+  return log;
+}
+
+inline LogType& GetErrorLogInstance() {
+  static LogType log(&std::clog);
+  return log;
+}
+
+inline LogType& GetLogInstanceForLevel(int level) {
+  if (level <= LogLevel()) {
+    return GetErrorLogInstance();
+  }
+  return GetNullLogInstance();
+}
+
+}  // end namespace internal
+}  // end namespace benchmark
+
+#define VLOG(x)                                                               \
+  (::benchmark::internal::GetLogInstanceForLevel(x) << "-- LOG(" << x << "):" \
+                                                                         " ")
+
+#endif

benchmarks/thirdparty/benchmark/src/mutex.h

@@ -0,0 +1,155 @@
+#ifndef BENCHMARK_MUTEX_H_
+#define BENCHMARK_MUTEX_H_
+
+#include <condition_variable>
+#include <mutex>
+
+#include "check.h"
+
+// Enable thread safety attributes only with clang.
+// The attributes can be safely erased when compiling with other compilers.
+#if defined(HAVE_THREAD_SAFETY_ATTRIBUTES)
+#define THREAD_ANNOTATION_ATTRIBUTE__(x) __attribute__((x))
+#else
+#define THREAD_ANNOTATION_ATTRIBUTE__(x)  // no-op
+#endif
+
+#define CAPABILITY(x) THREAD_ANNOTATION_ATTRIBUTE__(capability(x))
+
+#define SCOPED_CAPABILITY THREAD_ANNOTATION_ATTRIBUTE__(scoped_lockable)
+
+#define GUARDED_BY(x) THREAD_ANNOTATION_ATTRIBUTE__(guarded_by(x))
+
+#define PT_GUARDED_BY(x) THREAD_ANNOTATION_ATTRIBUTE__(pt_guarded_by(x))
+
+#define ACQUIRED_BEFORE(...) \
+  THREAD_ANNOTATION_ATTRIBUTE__(acquired_before(__VA_ARGS__))
+
+#define ACQUIRED_AFTER(...) \
+  THREAD_ANNOTATION_ATTRIBUTE__(acquired_after(__VA_ARGS__))
+
+#define REQUIRES(...) \
+  THREAD_ANNOTATION_ATTRIBUTE__(requires_capability(__VA_ARGS__))
+
+#define REQUIRES_SHARED(...) \
+  THREAD_ANNOTATION_ATTRIBUTE__(requires_shared_capability(__VA_ARGS__))
+
+#define ACQUIRE(...) \
+  THREAD_ANNOTATION_ATTRIBUTE__(acquire_capability(__VA_ARGS__))
+
+#define ACQUIRE_SHARED(...) \
+  THREAD_ANNOTATION_ATTRIBUTE__(acquire_shared_capability(__VA_ARGS__))
+
+#define RELEASE(...) \
+  THREAD_ANNOTATION_ATTRIBUTE__(release_capability(__VA_ARGS__))
+
+#define RELEASE_SHARED(...) \
+  THREAD_ANNOTATION_ATTRIBUTE__(release_shared_capability(__VA_ARGS__))
+
+#define TRY_ACQUIRE(...) \
+  THREAD_ANNOTATION_ATTRIBUTE__(try_acquire_capability(__VA_ARGS__))
+
+#define TRY_ACQUIRE_SHARED(...) \
+  THREAD_ANNOTATION_ATTRIBUTE__(try_acquire_shared_capability(__VA_ARGS__))
+
+#define EXCLUDES(...) THREAD_ANNOTATION_ATTRIBUTE__(locks_excluded(__VA_ARGS__))
+
+#define ASSERT_CAPABILITY(x) THREAD_ANNOTATION_ATTRIBUTE__(assert_capability(x))
+
+#define ASSERT_SHARED_CAPABILITY(x) \
+  THREAD_ANNOTATION_ATTRIBUTE__(assert_shared_capability(x))
+
+#define RETURN_CAPABILITY(x) THREAD_ANNOTATION_ATTRIBUTE__(lock_returned(x))
+
+#define NO_THREAD_SAFETY_ANALYSIS \
+  THREAD_ANNOTATION_ATTRIBUTE__(no_thread_safety_analysis)
+
+namespace benchmark {
+
+typedef std::condition_variable Condition;
+
+// NOTE: Wrappers for std::mutex and std::unique_lock are provided so that
+// we can annotate them with thread safety attributes and use the
+// -Wthread-safety warning with clang. The standard library types cannot be
+// used directly because they do not provided the required annotations.
+class CAPABILITY("mutex") Mutex {
+ public:
+  Mutex() {}
+
+  void lock() ACQUIRE() { mut_.lock(); }
+  void unlock() RELEASE() { mut_.unlock(); }
+  std::mutex& native_handle() { return mut_; }
+
+ private:
+  std::mutex mut_;
+};
+
+class SCOPED_CAPABILITY MutexLock {
+  typedef std::unique_lock<std::mutex> MutexLockImp;
+
+ public:
+  MutexLock(Mutex& m) ACQUIRE(m) : ml_(m.native_handle()) {}
+  ~MutexLock() RELEASE() {}
+  MutexLockImp& native_handle() { return ml_; }
+
+ private:
+  MutexLockImp ml_;
+};
+
+class Barrier {
+ public:
+  Barrier(int num_threads) : running_threads_(num_threads) {}
+
+  // Called by each thread
+  bool wait() EXCLUDES(lock_) {
+    bool last_thread = false;
+    {
+      MutexLock ml(lock_);
+      last_thread = createBarrier(ml);
+    }
+    if (last_thread) phase_condition_.notify_all();
+    return last_thread;
+  }
+
+  void removeThread() EXCLUDES(lock_) {
+    MutexLock ml(lock_);
+    --running_threads_;
+    if (entered_ != 0) phase_condition_.notify_all();
+  }
+
+ private:
+  Mutex lock_;
+  Condition phase_condition_;
+  int running_threads_;
+
+  // State for barrier management
+  int phase_number_ = 0;
+  int entered_ = 0;  // Number of threads that have entered this barrier
+
+  // Enter the barrier and wait until all other threads have also
+  // entered the barrier.  Returns iff this is the last thread to
+  // enter the barrier.
+  bool createBarrier(MutexLock& ml) REQUIRES(lock_) {
+    CHECK_LT(entered_, running_threads_);
+    entered_++;
+    if (entered_ < running_threads_) {
+      // Wait for all threads to enter
+      int phase_number_cp = phase_number_;
+      auto cb = [this, phase_number_cp]() {
+        return this->phase_number_ > phase_number_cp ||
+               entered_ == running_threads_;  // A thread has aborted in error
+      };
+      phase_condition_.wait(ml.native_handle(), cb);
+      if (phase_number_ > phase_number_cp) return false;
+      // else (running_threads_ == entered_) and we are the last thread.
+    }
+    // Last thread has reached the barrier
+    phase_number_++;
+    entered_ = 0;
+    return true;
+  }
+};
+
+}  // end namespace benchmark
+
+#endif  // BENCHMARK_MUTEX_H_

benchmarks/thirdparty/benchmark/src/re.h

@@ -0,0 +1,140 @@
+// Copyright 2015 Google Inc. All rights reserved.
+//
+// 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
+//
+//     http://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.
+
+#ifndef BENCHMARK_RE_H_
+#define BENCHMARK_RE_H_
+
+#include "internal_macros.h"
+
+// Prefer C regex libraries when compiling w/o exceptions so that we can
+// correctly report errors.
+#if defined(BENCHMARK_HAS_NO_EXCEPTIONS) && defined(HAVE_STD_REGEX) && \
+    (defined(HAVE_GNU_POSIX_REGEX) || defined(HAVE_POSIX_REGEX))
+#undef HAVE_STD_REGEX
+#endif
+
+#if defined(HAVE_STD_REGEX)
+#include <regex>
+#elif defined(HAVE_GNU_POSIX_REGEX)
+#include <gnuregex.h>
+#elif defined(HAVE_POSIX_REGEX)
+#include <regex.h>
+#else
+#error No regular expression backend was found!
+#endif
+#include <string>
+
+#include "check.h"
+
+namespace benchmark {
+
+// A wrapper around the POSIX regular expression API that provides automatic
+// cleanup
+class Regex {
+ public:
+  Regex() : init_(false) {}
+
+  ~Regex();
+
+  // Compile a regular expression matcher from spec.  Returns true on success.
+  //
+  // On failure (and if error is not nullptr), error is populated with a human
+  // readable error message if an error occurs.
+  bool Init(const std::string& spec, std::string* error);
+
+  // Returns whether str matches the compiled regular expression.
+  bool Match(const std::string& str);
+
+ private:
+  bool init_;
+// Underlying regular expression object
+#if defined(HAVE_STD_REGEX)
+  std::regex re_;
+#elif defined(HAVE_POSIX_REGEX) || defined(HAVE_GNU_POSIX_REGEX)
+  regex_t re_;
+#else
+#error No regular expression backend implementation available
+#endif
+};
+
+#if defined(HAVE_STD_REGEX)
+
+inline bool Regex::Init(const std::string& spec, std::string* error) {
+#ifdef BENCHMARK_HAS_NO_EXCEPTIONS
+  ((void)error); // suppress unused warning
+#else
+  try {
+#endif
+    re_ = std::regex(spec, std::regex_constants::extended);
+    init_ = true;
+#ifndef BENCHMARK_HAS_NO_EXCEPTIONS
+  } catch (const std::regex_error& e) {
+    if (error) {
+      *error = e.what();
+    }
+  }
+#endif
+  return init_;
+}
+
+inline Regex::~Regex() {}
+
+inline bool Regex::Match(const std::string& str) {
+  if (!init_) {
+    return false;
+  }
+  return std::regex_search(str, re_);
+}
+
+#else
+inline bool Regex::Init(const std::string& spec, std::string* error) {
+  int ec = regcomp(&re_, spec.c_str(), REG_EXTENDED | REG_NOSUB);
+  if (ec != 0) {
+    if (error) {
+      size_t needed = regerror(ec, &re_, nullptr, 0);
+      char* errbuf = new char[needed];
+      regerror(ec, &re_, errbuf, needed);
+
+      // regerror returns the number of bytes necessary to null terminate
+      // the string, so we move that when assigning to error.
+      CHECK_NE(needed, 0);
+      error->assign(errbuf, needed - 1);
+
+      delete[] errbuf;
+    }
+
+    return false;
+  }
+
+  init_ = true;
+  return true;
+}
+
+inline Regex::~Regex() {
+  if (init_) {
+    regfree(&re_);
+  }
+}
+
+inline bool Regex::Match(const std::string& str) {
+  if (!init_) {
+    return false;
+  }
+  return regexec(&re_, str.c_str(), 0, nullptr, 0) == 0;
+}
+#endif
+
+}  // end namespace benchmark
+
+#endif  // BENCHMARK_RE_H_

benchmarks/thirdparty/benchmark/src/reporter.cc

@@ -0,0 +1,81 @@
+// Copyright 2015 Google Inc. All rights reserved.
+//
+// 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
+//
+//     http://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 "benchmark/benchmark.h"
+#include "timers.h"
+
+#include <cstdlib>
+
+#include <iostream>
+#include <tuple>
+#include <vector>
+
+#include "check.h"
+
+namespace benchmark {
+
+BenchmarkReporter::BenchmarkReporter()
+    : output_stream_(&std::cout), error_stream_(&std::cerr) {}
+
+BenchmarkReporter::~BenchmarkReporter() {}
+
+void BenchmarkReporter::PrintBasicContext(std::ostream *out,
+                                          Context const &context) {
+  CHECK(out) << "cannot be null";
+  auto &Out = *out;
+
+  Out << LocalDateTimeString() << "\n";
+
+  const CPUInfo &info = context.cpu_info;
+  Out << "Run on (" << info.num_cpus << " X "
+      << (info.cycles_per_second / 1000000.0) << " MHz CPU "
+      << ((info.num_cpus > 1) ? "s" : "") << ")\n";
+  if (info.caches.size() != 0) {
+    Out << "CPU Caches:\n";
+    for (auto &CInfo : info.caches) {
+      Out << "  L" << CInfo.level << " " << CInfo.type << " "
+          << (CInfo.size / 1000) << "K";
+      if (CInfo.num_sharing != 0)
+        Out << " (x" << (info.num_cpus / CInfo.num_sharing) << ")";
+      Out << "\n";
+    }
+  }
+
+  if (info.scaling_enabled) {
+    Out << "***WARNING*** CPU scaling is enabled, the benchmark "
+           "real time measurements may be noisy and will incur extra "
+           "overhead.\n";
+  }
+
+#ifndef NDEBUG
+  Out << "***WARNING*** Library was built as DEBUG. Timings may be "
+         "affected.\n";
+#endif
+}
+
+BenchmarkReporter::Context::Context() : cpu_info(CPUInfo::Get()) {}
+
+double BenchmarkReporter::Run::GetAdjustedRealTime() const {
+  double new_time = real_accumulated_time * GetTimeUnitMultiplier(time_unit);
+  if (iterations != 0) new_time /= static_cast<double>(iterations);
+  return new_time;
+}
+
+double BenchmarkReporter::Run::GetAdjustedCPUTime() const {
+  double new_time = cpu_accumulated_time * GetTimeUnitMultiplier(time_unit);
+  if (iterations != 0) new_time /= static_cast<double>(iterations);
+  return new_time;
+}
+
+}  // end namespace benchmark

benchmarks/thirdparty/benchmark/src/sleep.cc

@@ -0,0 +1,51 @@
+// Copyright 2015 Google Inc. All rights reserved.
+//
+// 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
+//
+//     http://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 "sleep.h"
+
+#include <cerrno>
+#include <cstdlib>
+#include <ctime>
+
+#include "internal_macros.h"
+
+#ifdef BENCHMARK_OS_WINDOWS
+#include <Windows.h>
+#endif
+
+namespace benchmark {
+#ifdef BENCHMARK_OS_WINDOWS
+// Window's Sleep takes milliseconds argument.
+void SleepForMilliseconds(int milliseconds) { Sleep(milliseconds); }
+void SleepForSeconds(double seconds) {
+  SleepForMilliseconds(static_cast<int>(kNumMillisPerSecond * seconds));
+}
+#else   // BENCHMARK_OS_WINDOWS
+void SleepForMicroseconds(int microseconds) {
+  struct timespec sleep_time;
+  sleep_time.tv_sec = microseconds / kNumMicrosPerSecond;
+  sleep_time.tv_nsec = (microseconds % kNumMicrosPerSecond) * kNumNanosPerMicro;
+  while (nanosleep(&sleep_time, &sleep_time) != 0 && errno == EINTR)
+    ;  // Ignore signals and wait for the full interval to elapse.
+}
+
+void SleepForMilliseconds(int milliseconds) {
+  SleepForMicroseconds(milliseconds * kNumMicrosPerMilli);
+}
+
+void SleepForSeconds(double seconds) {
+  SleepForMicroseconds(static_cast<int>(seconds * kNumMicrosPerSecond));
+}
+#endif  // BENCHMARK_OS_WINDOWS
+}  // end namespace benchmark

benchmarks/thirdparty/benchmark/src/sleep.h

@@ -0,0 +1,15 @@
+#ifndef BENCHMARK_SLEEP_H_
+#define BENCHMARK_SLEEP_H_
+
+namespace benchmark {
+const int kNumMillisPerSecond = 1000;
+const int kNumMicrosPerMilli = 1000;
+const int kNumMicrosPerSecond = kNumMillisPerSecond * 1000;
+const int kNumNanosPerMicro = 1000;
+const int kNumNanosPerSecond = kNumNanosPerMicro * kNumMicrosPerSecond;
+
+void SleepForMilliseconds(int milliseconds);
+void SleepForSeconds(double seconds);
+}  // end namespace benchmark
+
+#endif  // BENCHMARK_SLEEP_H_

benchmarks/thirdparty/benchmark/src/statistics.cc

@@ -0,0 +1,175 @@
+// Copyright 2016 Ismael Jimenez Martinez. All rights reserved.
+// Copyright 2017 Roman Lebedev. All rights reserved.
+//
+// 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
+//
+//     http://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 "benchmark/benchmark.h"
+
+#include <algorithm>
+#include <cmath>
+#include <string>
+#include <vector>
+#include <numeric>
+#include "check.h"
+#include "statistics.h"
+
+namespace benchmark {
+
+auto StatisticsSum = [](const std::vector<double>& v) {
+  return std::accumulate(v.begin(), v.end(), 0.0);
+};
+
+double StatisticsMean(const std::vector<double>& v) {
+  if (v.size() == 0) return 0.0;
+  return StatisticsSum(v) * (1.0 / v.size());
+}
+
+double StatisticsMedian(const std::vector<double>& v) {
+  if (v.size() < 3) return StatisticsMean(v);
+  std::vector<double> partial;
+  // we need roundDown(count/2)+1 slots
+  partial.resize(1 + (v.size() / 2));
+  std::partial_sort_copy(v.begin(), v.end(), partial.begin(), partial.end());
+  // did we have odd number of samples?
+  // if yes, then the last element of partially-sorted vector is the median
+  // it no, then the average of the last two elements is the median
+  if(v.size() % 2 == 1)
+    return partial.back();
+  return (partial[partial.size() - 2] + partial[partial.size() - 1]) / 2.0;
+}
+
+// Return the sum of the squares of this sample set
+auto SumSquares = [](const std::vector<double>& v) {
+  return std::inner_product(v.begin(), v.end(), v.begin(), 0.0);
+};
+
+auto Sqr = [](const double dat) { return dat * dat; };
+auto Sqrt = [](const double dat) {
+  // Avoid NaN due to imprecision in the calculations
+  if (dat < 0.0) return 0.0;
+  return std::sqrt(dat);
+};
+
+double StatisticsStdDev(const std::vector<double>& v) {
+  const auto mean = StatisticsMean(v);
+  if (v.size() == 0) return mean;
+
+  // Sample standard deviation is undefined for n = 1
+  if (v.size() == 1)
+    return 0.0;
+
+  const double avg_squares = SumSquares(v) * (1.0 / v.size());
+  return Sqrt(v.size() / (v.size() - 1.0) * (avg_squares - Sqr(mean)));
+}
+
+std::vector<BenchmarkReporter::Run> ComputeStats(
+    const std::vector<BenchmarkReporter::Run>& reports) {
+  typedef BenchmarkReporter::Run Run;
+  std::vector<Run> results;
+
+  auto error_count =
+      std::count_if(reports.begin(), reports.end(),
+                    [](Run const& run) { return run.error_occurred; });
+
+  if (reports.size() - error_count < 2) {
+    // We don't report aggregated data if there was a single run.
+    return results;
+  }
+
+  // Accumulators.
+  std::vector<double> real_accumulated_time_stat;
+  std::vector<double> cpu_accumulated_time_stat;
+  std::vector<double> bytes_per_second_stat;
+  std::vector<double> items_per_second_stat;
+
+  real_accumulated_time_stat.reserve(reports.size());
+  cpu_accumulated_time_stat.reserve(reports.size());
+  bytes_per_second_stat.reserve(reports.size());
+  items_per_second_stat.reserve(reports.size());
+
+  // All repetitions should be run with the same number of iterations so we
+  // can take this information from the first benchmark.
+  int64_t const run_iterations = reports.front().iterations;
+  // create stats for user counters
+  struct CounterStat {
+    Counter c;
+    std::vector<double> s;
+  };
+  std::map< std::string, CounterStat > counter_stats;
+  for(Run const& r : reports) {
+    for(auto const& cnt : r.counters) {
+      auto it = counter_stats.find(cnt.first);
+      if(it == counter_stats.end()) {
+        counter_stats.insert({cnt.first, {cnt.second, std::vector<double>{}}});
+        it = counter_stats.find(cnt.first);
+        it->second.s.reserve(reports.size());
+      } else {
+        CHECK_EQ(counter_stats[cnt.first].c.flags, cnt.second.flags);
+      }
+    }
+  }
+
+  // Populate the accumulators.
+  for (Run const& run : reports) {
+    CHECK_EQ(reports[0].benchmark_name, run.benchmark_name);
+    CHECK_EQ(run_iterations, run.iterations);
+    if (run.error_occurred) continue;
+    real_accumulated_time_stat.emplace_back(run.real_accumulated_time);
+    cpu_accumulated_time_stat.emplace_back(run.cpu_accumulated_time);
+    items_per_second_stat.emplace_back(run.items_per_second);
+    bytes_per_second_stat.emplace_back(run.bytes_per_second);
+    // user counters
+    for(auto const& cnt : run.counters) {
+      auto it = counter_stats.find(cnt.first);
+      CHECK_NE(it, counter_stats.end());
+      it->second.s.emplace_back(cnt.second);
+    }
+  }
+
+  // Only add label if it is same for all runs
+  std::string report_label = reports[0].report_label;
+  for (std::size_t i = 1; i < reports.size(); i++) {
+    if (reports[i].report_label != report_label) {
+      report_label = "";
+      break;
+    }
+  }
+
+  for(const auto& Stat : *reports[0].statistics) {
+    // Get the data from the accumulator to BenchmarkReporter::Run's.
+    Run data;
+    data.benchmark_name = reports[0].benchmark_name + "_" + Stat.name_;
+    data.report_label = report_label;
+    data.iterations = run_iterations;
+
+    data.real_accumulated_time = Stat.compute_(real_accumulated_time_stat);
+    data.cpu_accumulated_time = Stat.compute_(cpu_accumulated_time_stat);
+    data.bytes_per_second = Stat.compute_(bytes_per_second_stat);
+    data.items_per_second = Stat.compute_(items_per_second_stat);
+
+    data.time_unit = reports[0].time_unit;
+
+    // user counters
+    for(auto const& kv : counter_stats) {
+      const auto uc_stat = Stat.compute_(kv.second.s);
+      auto c = Counter(uc_stat, counter_stats[kv.first].c.flags);
+      data.counters[kv.first] = c;
+    }
+
+    results.push_back(data);
+  }
+
+  return results;
+}
+
+}  // end namespace benchmark

benchmarks/thirdparty/benchmark/src/statistics.h

@@ -0,0 +1,37 @@
+// Copyright 2016 Ismael Jimenez Martinez. All rights reserved.
+// Copyright 2017 Roman Lebedev. All rights reserved.
+//
+// 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
+//
+//     http://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.
+
+#ifndef STATISTICS_H_
+#define STATISTICS_H_
+
+#include <vector>
+
+#include "benchmark/benchmark.h"
+
+namespace benchmark {
+
+// Return a vector containing the mean, median and standard devation information
+// (and any user-specified info) for the specified list of reports. If 'reports'
+// contains less than two non-errored runs an empty vector is returned
+std::vector<BenchmarkReporter::Run> ComputeStats(
+    const std::vector<BenchmarkReporter::Run>& reports);
+
+double StatisticsMean(const std::vector<double>& v);
+double StatisticsMedian(const std::vector<double>& v);
+double StatisticsStdDev(const std::vector<double>& v);
+
+}  // end namespace benchmark
+
+#endif  // STATISTICS_H_

benchmarks/thirdparty/benchmark/src/string_util.cc

@@ -0,0 +1,172 @@
+#include "string_util.h"
+
+#include <array>
+#include <cmath>
+#include <cstdarg>
+#include <cstdio>
+#include <memory>
+#include <sstream>
+
+#include "arraysize.h"
+
+namespace benchmark {
+namespace {
+
+// kilo, Mega, Giga, Tera, Peta, Exa, Zetta, Yotta.
+const char kBigSIUnits[] = "kMGTPEZY";
+// Kibi, Mebi, Gibi, Tebi, Pebi, Exbi, Zebi, Yobi.
+const char kBigIECUnits[] = "KMGTPEZY";
+// milli, micro, nano, pico, femto, atto, zepto, yocto.
+const char kSmallSIUnits[] = "munpfazy";
+
+// We require that all three arrays have the same size.
+static_assert(arraysize(kBigSIUnits) == arraysize(kBigIECUnits),
+              "SI and IEC unit arrays must be the same size");
+static_assert(arraysize(kSmallSIUnits) == arraysize(kBigSIUnits),
+              "Small SI and Big SI unit arrays must be the same size");
+
+static const int64_t kUnitsSize = arraysize(kBigSIUnits);
+
+void ToExponentAndMantissa(double val, double thresh, int precision,
+                           double one_k, std::string* mantissa,
+                           int64_t* exponent) {
+  std::stringstream mantissa_stream;
+
+  if (val < 0) {
+    mantissa_stream << "-";
+    val = -val;
+  }
+
+  // Adjust threshold so that it never excludes things which can't be rendered
+  // in 'precision' digits.
+  const double adjusted_threshold =
+      std::max(thresh, 1.0 / std::pow(10.0, precision));
+  const double big_threshold = adjusted_threshold * one_k;
+  const double small_threshold = adjusted_threshold;
+  // Values in ]simple_threshold,small_threshold[ will be printed as-is
+  const double simple_threshold = 0.01;
+
+  if (val > big_threshold) {
+    // Positive powers
+    double scaled = val;
+    for (size_t i = 0; i < arraysize(kBigSIUnits); ++i) {
+      scaled /= one_k;
+      if (scaled <= big_threshold) {
+        mantissa_stream << scaled;
+        *exponent = i + 1;
+        *mantissa = mantissa_stream.str();
+        return;
+      }
+    }
+    mantissa_stream << val;
+    *exponent = 0;
+  } else if (val < small_threshold) {
+    // Negative powers
+    if (val < simple_threshold) {
+      double scaled = val;
+      for (size_t i = 0; i < arraysize(kSmallSIUnits); ++i) {
+        scaled *= one_k;
+        if (scaled >= small_threshold) {
+          mantissa_stream << scaled;
+          *exponent = -static_cast<int64_t>(i + 1);
+          *mantissa = mantissa_stream.str();
+          return;
+        }
+      }
+    }
+    mantissa_stream << val;
+    *exponent = 0;
+  } else {
+    mantissa_stream << val;
+    *exponent = 0;
+  }
+  *mantissa = mantissa_stream.str();
+}
+
+std::string ExponentToPrefix(int64_t exponent, bool iec) {
+  if (exponent == 0) return "";
+
+  const int64_t index = (exponent > 0 ? exponent - 1 : -exponent - 1);
+  if (index >= kUnitsSize) return "";
+
+  const char* array =
+      (exponent > 0 ? (iec ? kBigIECUnits : kBigSIUnits) : kSmallSIUnits);
+  if (iec)
+    return array[index] + std::string("i");
+  else
+    return std::string(1, array[index]);
+}
+
+std::string ToBinaryStringFullySpecified(double value, double threshold,
+                                         int precision, double one_k = 1024.0) {
+  std::string mantissa;
+  int64_t exponent;
+  ToExponentAndMantissa(value, threshold, precision, one_k, &mantissa,
+                        &exponent);
+  return mantissa + ExponentToPrefix(exponent, false);
+}
+
+}  // end namespace
+
+void AppendHumanReadable(int n, std::string* str) {
+  std::stringstream ss;
+  // Round down to the nearest SI prefix.
+  ss << ToBinaryStringFullySpecified(n, 1.0, 0);
+  *str += ss.str();
+}
+
+std::string HumanReadableNumber(double n, double one_k) {
+  // 1.1 means that figures up to 1.1k should be shown with the next unit down;
+  // this softens edge effects.
+  // 1 means that we should show one decimal place of precision.
+  return ToBinaryStringFullySpecified(n, 1.1, 1, one_k);
+}
+
+std::string StringPrintFImp(const char* msg, va_list args) {
+  // we might need a second shot at this, so pre-emptivly make a copy
+  va_list args_cp;
+  va_copy(args_cp, args);
+
+  // TODO(ericwf): use std::array for first attempt to avoid one memory
+  // allocation guess what the size might be
+  std::array<char, 256> local_buff;
+  std::size_t size = local_buff.size();
+  // 2015-10-08: vsnprintf is used instead of snd::vsnprintf due to a limitation
+  // in the android-ndk
+  auto ret = vsnprintf(local_buff.data(), size, msg, args_cp);
+
+  va_end(args_cp);
+
+  // handle empty expansion
+  if (ret == 0) return std::string{};
+  if (static_cast<std::size_t>(ret) < size)
+    return std::string(local_buff.data());
+
+  // we did not provide a long enough buffer on our first attempt.
+  // add 1 to size to account for null-byte in size cast to prevent overflow
+  size = static_cast<std::size_t>(ret) + 1;
+  auto buff_ptr = std::unique_ptr<char[]>(new char[size]);
+  // 2015-10-08: vsnprintf is used instead of snd::vsnprintf due to a limitation
+  // in the android-ndk
+  ret = vsnprintf(buff_ptr.get(), size, msg, args);
+  return std::string(buff_ptr.get());
+}
+
+std::string StringPrintF(const char* format, ...) {
+  va_list args;
+  va_start(args, format);
+  std::string tmp = StringPrintFImp(format, args);
+  va_end(args);
+  return tmp;
+}
+
+void ReplaceAll(std::string* str, const std::string& from,
+                const std::string& to) {
+  std::size_t start = 0;
+  while ((start = str->find(from, start)) != std::string::npos) {
+    str->replace(start, from.length(), to);
+    start += to.length();
+  }
+}
+
+}  // end namespace benchmark

benchmarks/thirdparty/benchmark/src/string_util.h

@@ -0,0 +1,40 @@
+#ifndef BENCHMARK_STRING_UTIL_H_
+#define BENCHMARK_STRING_UTIL_H_
+
+#include <sstream>
+#include <string>
+#include <utility>
+#include "internal_macros.h"
+
+namespace benchmark {
+
+void AppendHumanReadable(int n, std::string* str);
+
+std::string HumanReadableNumber(double n, double one_k = 1024.0);
+
+std::string StringPrintF(const char* format, ...);
+
+inline std::ostream& StringCatImp(std::ostream& out) BENCHMARK_NOEXCEPT {
+  return out;
+}
+
+template <class First, class... Rest>
+inline std::ostream& StringCatImp(std::ostream& out, First&& f,
+                                  Rest&&... rest) {
+  out << std::forward<First>(f);
+  return StringCatImp(out, std::forward<Rest>(rest)...);
+}
+
+template <class... Args>
+inline std::string StrCat(Args&&... args) {
+  std::ostringstream ss;
+  StringCatImp(ss, std::forward<Args>(args)...);
+  return ss.str();
+}
+
+void ReplaceAll(std::string* str, const std::string& from,
+                const std::string& to);
+
+}  // end namespace benchmark
+
+#endif  // BENCHMARK_STRING_UTIL_H_

benchmarks/thirdparty/benchmark/src/sysinfo.cc

@@ -0,0 +1,517 @@
+// Copyright 2015 Google Inc. All rights reserved.
+//
+// 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
+//
+//     http://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 "internal_macros.h"
+
+#ifdef BENCHMARK_OS_WINDOWS
+#include <Shlwapi.h>
+#include <VersionHelpers.h>
+#include <Windows.h>
+#else
+#include <fcntl.h>
+#include <sys/resource.h>
+#include <sys/time.h>
+#include <sys/types.h>  // this header must be included before 'sys/sysctl.h' to avoid compilation error on FreeBSD
+#include <unistd.h>
+#if defined BENCHMARK_OS_FREEBSD || defined BENCHMARK_OS_MACOSX || \
+    defined BENCHMARK_OS_NETBSD
+#define BENCHMARK_HAS_SYSCTL
+#include <sys/sysctl.h>
+#endif
+#endif
+
+#include <algorithm>
+#include <array>
+#include <bitset>
+#include <cerrno>
+#include <climits>
+#include <cstdint>
+#include <cstdio>
+#include <cstdlib>
+#include <cstring>
+#include <fstream>
+#include <iostream>
+#include <iterator>
+#include <limits>
+#include <memory>
+#include <sstream>
+
+#include "check.h"
+#include "cycleclock.h"
+#include "internal_macros.h"
+#include "log.h"
+#include "sleep.h"
+#include "string_util.h"
+
+namespace benchmark {
+namespace {
+
+void PrintImp(std::ostream& out) { out << std::endl; }
+
+template <class First, class... Rest>
+void PrintImp(std::ostream& out, First&& f, Rest&&... rest) {
+  out << std::forward<First>(f);
+  PrintImp(out, std::forward<Rest>(rest)...);
+}
+
+template <class... Args>
+BENCHMARK_NORETURN void PrintErrorAndDie(Args&&... args) {
+  PrintImp(std::cerr, std::forward<Args>(args)...);
+  std::exit(EXIT_FAILURE);
+}
+
+#ifdef BENCHMARK_HAS_SYSCTL
+
+/// ValueUnion - A type used to correctly alias the byte-for-byte output of
+/// `sysctl` with the result type it's to be interpreted as.
+struct ValueUnion {
+  union DataT {
+    uint32_t uint32_value;
+    uint64_t uint64_value;
+    // For correct aliasing of union members from bytes.
+    char bytes[8];
+  };
+  using DataPtr = std::unique_ptr<DataT, decltype(&std::free)>;
+
+  // The size of the data union member + its trailing array size.
+  size_t Size;
+  DataPtr Buff;
+
+ public:
+  ValueUnion() : Size(0), Buff(nullptr, &std::free) {}
+
+  explicit ValueUnion(size_t BuffSize)
+      : Size(sizeof(DataT) + BuffSize),
+        Buff(::new (std::malloc(Size)) DataT(), &std::free) {}
+
+  ValueUnion(ValueUnion&& other) = default;
+
+  explicit operator bool() const { return bool(Buff); }
+
+  char* data() const { return Buff->bytes; }
+
+  std::string GetAsString() const { return std::string(data()); }
+
+  int64_t GetAsInteger() const {
+    if (Size == sizeof(Buff->uint32_value))
+      return static_cast<int32_t>(Buff->uint32_value);
+    else if (Size == sizeof(Buff->uint64_value))
+      return static_cast<int64_t>(Buff->uint64_value);
+    BENCHMARK_UNREACHABLE();
+  }
+
+  uint64_t GetAsUnsigned() const {
+    if (Size == sizeof(Buff->uint32_value))
+      return Buff->uint32_value;
+    else if (Size == sizeof(Buff->uint64_value))
+      return Buff->uint64_value;
+    BENCHMARK_UNREACHABLE();
+  }
+
+  template <class T, int N>
+  std::array<T, N> GetAsArray() {
+    const int ArrSize = sizeof(T) * N;
+    CHECK_LE(ArrSize, Size);
+    std::array<T, N> Arr;
+    std::memcpy(Arr.data(), data(), ArrSize);
+    return Arr;
+  }
+};
+
+ValueUnion GetSysctlImp(std::string const& Name) {
+  size_t CurBuffSize = 0;
+  if (sysctlbyname(Name.c_str(), nullptr, &CurBuffSize, nullptr, 0) == -1)
+    return ValueUnion();
+
+  ValueUnion buff(CurBuffSize);
+  if (sysctlbyname(Name.c_str(), buff.data(), &buff.Size, nullptr, 0) == 0)
+    return buff;
+  return ValueUnion();
+}
+
+BENCHMARK_MAYBE_UNUSED
+bool GetSysctl(std::string const& Name, std::string* Out) {
+  Out->clear();
+  auto Buff = GetSysctlImp(Name);
+  if (!Buff) return false;
+  Out->assign(Buff.data());
+  return true;
+}
+
+template <class Tp,
+          class = typename std::enable_if<std::is_integral<Tp>::value>::type>
+bool GetSysctl(std::string const& Name, Tp* Out) {
+  *Out = 0;
+  auto Buff = GetSysctlImp(Name);
+  if (!Buff) return false;
+  *Out = static_cast<Tp>(Buff.GetAsUnsigned());
+  return true;
+}
+
+template <class Tp, size_t N>
+bool GetSysctl(std::string const& Name, std::array<Tp, N>* Out) {
+  auto Buff = GetSysctlImp(Name);
+  if (!Buff) return false;
+  *Out = Buff.GetAsArray<Tp, N>();
+  return true;
+}
+#endif
+
+template <class ArgT>
+bool ReadFromFile(std::string const& fname, ArgT* arg) {
+  *arg = ArgT();
+  std::ifstream f(fname.c_str());
+  if (!f.is_open()) return false;
+  f >> *arg;
+  return f.good();
+}
+
+bool CpuScalingEnabled(int num_cpus) {
+  // We don't have a valid CPU count, so don't even bother.
+  if (num_cpus <= 0) return false;
+#ifndef BENCHMARK_OS_WINDOWS
+  // On Linux, the CPUfreq subsystem exposes CPU information as files on the
+  // local file system. If reading the exported files fails, then we may not be
+  // running on Linux, so we silently ignore all the read errors.
+  std::string res;
+  for (int cpu = 0; cpu < num_cpus; ++cpu) {
+    std::string governor_file =
+        StrCat("/sys/devices/system/cpu/cpu", cpu, "/cpufreq/scaling_governor");
+    if (ReadFromFile(governor_file, &res) && res != "performance") return true;
+  }
+#endif
+  return false;
+}
+
+int CountSetBitsInCPUMap(std::string Val) {
+  auto CountBits = [](std::string Part) {
+    using CPUMask = std::bitset<sizeof(std::uintptr_t) * CHAR_BIT>;
+    Part = "0x" + Part;
+    CPUMask Mask(std::stoul(Part, nullptr, 16));
+    return static_cast<int>(Mask.count());
+  };
+  size_t Pos;
+  int total = 0;
+  while ((Pos = Val.find(',')) != std::string::npos) {
+    total += CountBits(Val.substr(0, Pos));
+    Val = Val.substr(Pos + 1);
+  }
+  if (!Val.empty()) {
+    total += CountBits(Val);
+  }
+  return total;
+}
+
+BENCHMARK_MAYBE_UNUSED
+std::vector<CPUInfo::CacheInfo> GetCacheSizesFromKVFS() {
+  std::vector<CPUInfo::CacheInfo> res;
+  std::string dir = "/sys/devices/system/cpu/cpu0/cache/";
+  int Idx = 0;
+  while (true) {
+    CPUInfo::CacheInfo info;
+    std::string FPath = StrCat(dir, "index", Idx++, "/");
+    std::ifstream f(StrCat(FPath, "size").c_str());
+    if (!f.is_open()) break;
+    std::string suffix;
+    f >> info.size;
+    if (f.fail())
+      PrintErrorAndDie("Failed while reading file '", FPath, "size'");
+    if (f.good()) {
+      f >> suffix;
+      if (f.bad())
+        PrintErrorAndDie(
+            "Invalid cache size format: failed to read size suffix");
+      else if (f && suffix != "K")
+        PrintErrorAndDie("Invalid cache size format: Expected bytes ", suffix);
+      else if (suffix == "K")
+        info.size *= 1000;
+    }
+    if (!ReadFromFile(StrCat(FPath, "type"), &info.type))
+      PrintErrorAndDie("Failed to read from file ", FPath, "type");
+    if (!ReadFromFile(StrCat(FPath, "level"), &info.level))
+      PrintErrorAndDie("Failed to read from file ", FPath, "level");
+    std::string map_str;
+    if (!ReadFromFile(StrCat(FPath, "shared_cpu_map"), &map_str))
+      PrintErrorAndDie("Failed to read from file ", FPath, "shared_cpu_map");
+    info.num_sharing = CountSetBitsInCPUMap(map_str);
+    res.push_back(info);
+  }
+
+  return res;
+}
+
+#ifdef BENCHMARK_OS_MACOSX
+std::vector<CPUInfo::CacheInfo> GetCacheSizesMacOSX() {
+  std::vector<CPUInfo::CacheInfo> res;
+  std::array<uint64_t, 4> CacheCounts{{0, 0, 0, 0}};
+  GetSysctl("hw.cacheconfig", &CacheCounts);
+
+  struct {
+    std::string name;
+    std::string type;
+    int level;
+    size_t num_sharing;
+  } Cases[] = {{"hw.l1dcachesize", "Data", 1, CacheCounts[1]},
+               {"hw.l1icachesize", "Instruction", 1, CacheCounts[1]},
+               {"hw.l2cachesize", "Unified", 2, CacheCounts[2]},
+               {"hw.l3cachesize", "Unified", 3, CacheCounts[3]}};
+  for (auto& C : Cases) {
+    int val;
+    if (!GetSysctl(C.name, &val)) continue;
+    CPUInfo::CacheInfo info;
+    info.type = C.type;
+    info.level = C.level;
+    info.size = val;
+    info.num_sharing = static_cast<int>(C.num_sharing);
+    res.push_back(std::move(info));
+  }
+  return res;
+}
+#elif defined(BENCHMARK_OS_WINDOWS)
+std::vector<CPUInfo::CacheInfo> GetCacheSizesWindows() {
+  std::vector<CPUInfo::CacheInfo> res;
+  DWORD buffer_size = 0;
+  using PInfo = SYSTEM_LOGICAL_PROCESSOR_INFORMATION;
+  using CInfo = CACHE_DESCRIPTOR;
+
+  using UPtr = std::unique_ptr<PInfo, decltype(&std::free)>;
+  GetLogicalProcessorInformation(nullptr, &buffer_size);
+  UPtr buff((PInfo*)malloc(buffer_size), &std::free);
+  if (!GetLogicalProcessorInformation(buff.get(), &buffer_size))
+    PrintErrorAndDie("Failed during call to GetLogicalProcessorInformation: ",
+                     GetLastError());
+
+  PInfo* it = buff.get();
+  PInfo* end = buff.get() + (buffer_size / sizeof(PInfo));
+
+  for (; it != end; ++it) {
+    if (it->Relationship != RelationCache) continue;
+    using BitSet = std::bitset<sizeof(ULONG_PTR) * CHAR_BIT>;
+    BitSet B(it->ProcessorMask);
+    // To prevent duplicates, only consider caches where CPU 0 is specified
+    if (!B.test(0)) continue;
+    CInfo* Cache = &it->Cache;
+    CPUInfo::CacheInfo C;
+    C.num_sharing = B.count();
+    C.level = Cache->Level;
+    C.size = Cache->Size;
+    switch (Cache->Type) {
+      case CacheUnified:
+        C.type = "Unified";
+        break;
+      case CacheInstruction:
+        C.type = "Instruction";
+        break;
+      case CacheData:
+        C.type = "Data";
+        break;
+      case CacheTrace:
+        C.type = "Trace";
+        break;
+      default:
+        C.type = "Unknown";
+        break;
+    }
+    res.push_back(C);
+  }
+  return res;
+}
+#endif
+
+std::vector<CPUInfo::CacheInfo> GetCacheSizes() {
+#ifdef BENCHMARK_OS_MACOSX
+  return GetCacheSizesMacOSX();
+#elif defined(BENCHMARK_OS_WINDOWS)
+  return GetCacheSizesWindows();
+#else
+  return GetCacheSizesFromKVFS();
+#endif
+}
+
+int GetNumCPUs() {
+#ifdef BENCHMARK_HAS_SYSCTL
+  int NumCPU = -1;
+  if (GetSysctl("hw.ncpu", &NumCPU)) return NumCPU;
+  fprintf(stderr, "Err: %s\n", strerror(errno));
+  std::exit(EXIT_FAILURE);
+#elif defined(BENCHMARK_OS_WINDOWS)
+  SYSTEM_INFO sysinfo;
+  // Use memset as opposed to = {} to avoid GCC missing initializer false
+  // positives.
+  std::memset(&sysinfo, 0, sizeof(SYSTEM_INFO));
+  GetSystemInfo(&sysinfo);
+  return sysinfo.dwNumberOfProcessors;  // number of logical
+                                        // processors in the current
+                                        // group
+#else
+  int NumCPUs = 0;
+  int MaxID = -1;
+  std::ifstream f("/proc/cpuinfo");
+  if (!f.is_open()) {
+    std::cerr << "failed to open /proc/cpuinfo\n";
+    return -1;
+  }
+  const std::string Key = "processor";
+  std::string ln;
+  while (std::getline(f, ln)) {
+    if (ln.empty()) continue;
+    size_t SplitIdx = ln.find(':');
+    std::string value;
+    if (SplitIdx != std::string::npos) value = ln.substr(SplitIdx + 1);
+    if (ln.size() >= Key.size() && ln.compare(0, Key.size(), Key) == 0) {
+      NumCPUs++;
+      if (!value.empty()) {
+        int CurID = std::stoi(value);
+        MaxID = std::max(CurID, MaxID);
+      }
+    }
+  }
+  if (f.bad()) {
+    std::cerr << "Failure reading /proc/cpuinfo\n";
+    return -1;
+  }
+  if (!f.eof()) {
+    std::cerr << "Failed to read to end of /proc/cpuinfo\n";
+    return -1;
+  }
+  f.close();
+
+  if ((MaxID + 1) != NumCPUs) {
+    fprintf(stderr,
+            "CPU ID assignments in /proc/cpuinfo seem messed up."
+            " This is usually caused by a bad BIOS.\n");
+  }
+  return NumCPUs;
+#endif
+  BENCHMARK_UNREACHABLE();
+}
+
+double GetCPUCyclesPerSecond() {
+#if defined BENCHMARK_OS_LINUX || defined BENCHMARK_OS_CYGWIN
+  long freq;
+
+  // If the kernel is exporting the tsc frequency use that. There are issues
+  // where cpuinfo_max_freq cannot be relied on because the BIOS may be
+  // exporintg an invalid p-state (on x86) or p-states may be used to put the
+  // processor in a new mode (turbo mode). Essentially, those frequencies
+  // cannot always be relied upon. The same reasons apply to /proc/cpuinfo as
+  // well.
+  if (ReadFromFile("/sys/devices/system/cpu/cpu0/tsc_freq_khz", &freq)
+      // If CPU scaling is in effect, we want to use the *maximum* frequency,
+      // not whatever CPU speed some random processor happens to be using now.
+      || ReadFromFile("/sys/devices/system/cpu/cpu0/cpufreq/cpuinfo_max_freq",
+                      &freq)) {
+    // The value is in kHz (as the file name suggests).  For example, on a
+    // 2GHz warpstation, the file contains the value "2000000".
+    return freq * 1000.0;
+  }
+
+  const double error_value = -1;
+  double bogo_clock = error_value;
+
+  std::ifstream f("/proc/cpuinfo");
+  if (!f.is_open()) {
+    std::cerr << "failed to open /proc/cpuinfo\n";
+    return error_value;
+  }
+
+  auto startsWithKey = [](std::string const& Value, std::string const& Key) {
+    if (Key.size() > Value.size()) return false;
+    auto Cmp = [&](char X, char Y) {
+      return std::tolower(X) == std::tolower(Y);
+    };
+    return std::equal(Key.begin(), Key.end(), Value.begin(), Cmp);
+  };
+
+  std::string ln;
+  while (std::getline(f, ln)) {
+    if (ln.empty()) continue;
+    size_t SplitIdx = ln.find(':');
+    std::string value;
+    if (SplitIdx != std::string::npos) value = ln.substr(SplitIdx + 1);
+    // When parsing the "cpu MHz" and "bogomips" (fallback) entries, we only
+    // accept postive values. Some environments (virtual machines) report zero,
+    // which would cause infinite looping in WallTime_Init.
+    if (startsWithKey(ln, "cpu MHz")) {
+      if (!value.empty()) {
+        double cycles_per_second = std::stod(value) * 1000000.0;
+        if (cycles_per_second > 0) return cycles_per_second;
+      }
+    } else if (startsWithKey(ln, "bogomips")) {
+      if (!value.empty()) {
+        bogo_clock = std::stod(value) * 1000000.0;
+        if (bogo_clock < 0.0) bogo_clock = error_value;
+      }
+    }
+  }
+  if (f.bad()) {
+    std::cerr << "Failure reading /proc/cpuinfo\n";
+    return error_value;
+  }
+  if (!f.eof()) {
+    std::cerr << "Failed to read to end of /proc/cpuinfo\n";
+    return error_value;
+  }
+  f.close();
+  // If we found the bogomips clock, but nothing better, we'll use it (but
+  // we're not happy about it); otherwise, fallback to the rough estimation
+  // below.
+  if (bogo_clock >= 0.0) return bogo_clock;
+
+#elif defined BENCHMARK_HAS_SYSCTL
+  constexpr auto* FreqStr =
+#if defined(BENCHMARK_OS_FREEBSD) || defined(BENCHMARK_OS_NETBSD)
+      "machdep.tsc_freq";
+#else
+      "hw.cpufrequency";
+#endif
+  unsigned long long hz = 0;
+  if (GetSysctl(FreqStr, &hz)) return hz;
+
+  fprintf(stderr, "Unable to determine clock rate from sysctl: %s: %s\n",
+          FreqStr, strerror(errno));
+
+#elif defined BENCHMARK_OS_WINDOWS
+  // In NT, read MHz from the registry. If we fail to do so or we're in win9x
+  // then make a crude estimate.
+  DWORD data, data_size = sizeof(data);
+  if (IsWindowsXPOrGreater() &&
+      SUCCEEDED(
+          SHGetValueA(HKEY_LOCAL_MACHINE,
+                      "HARDWARE\\DESCRIPTION\\System\\CentralProcessor\\0",
+                      "~MHz", nullptr, &data, &data_size)))
+    return static_cast<double>((int64_t)data *
+                               (int64_t)(1000 * 1000));  // was mhz
+#endif
+  // If we've fallen through, attempt to roughly estimate the CPU clock rate.
+  const int estimate_time_ms = 1000;
+  const auto start_ticks = cycleclock::Now();
+  SleepForMilliseconds(estimate_time_ms);
+  return static_cast<double>(cycleclock::Now() - start_ticks);
+}
+
+}  // end namespace
+
+const CPUInfo& CPUInfo::Get() {
+  static const CPUInfo* info = new CPUInfo();
+  return *info;
+}
+
+CPUInfo::CPUInfo()
+    : num_cpus(GetNumCPUs()),
+      cycles_per_second(GetCPUCyclesPerSecond()),
+      caches(GetCacheSizes()),
+      scaling_enabled(CpuScalingEnabled(num_cpus)) {}
+
+}  // end namespace benchmark

benchmarks/thirdparty/benchmark/src/timers.cc

@@ -0,0 +1,212 @@
+// Copyright 2015 Google Inc. All rights reserved.
+//
+// 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
+//
+//     http://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 "timers.h"
+#include "internal_macros.h"
+
+#ifdef BENCHMARK_OS_WINDOWS
+#include <Shlwapi.h>
+#include <VersionHelpers.h>
+#include <Windows.h>
+#else
+#include <fcntl.h>
+#include <sys/resource.h>
+#include <sys/time.h>
+#include <sys/types.h>  // this header must be included before 'sys/sysctl.h' to avoid compilation error on FreeBSD
+#include <unistd.h>
+#if defined BENCHMARK_OS_FREEBSD || defined BENCHMARK_OS_MACOSX
+#include <sys/sysctl.h>
+#endif
+#if defined(BENCHMARK_OS_MACOSX)
+#include <mach/mach_init.h>
+#include <mach/mach_port.h>
+#include <mach/thread_act.h>
+#endif
+#endif
+
+#ifdef BENCHMARK_OS_EMSCRIPTEN
+#include <emscripten.h>
+#endif
+
+#include <cerrno>
+#include <cstdint>
+#include <cstdio>
+#include <cstdlib>
+#include <cstring>
+#include <ctime>
+#include <iostream>
+#include <limits>
+#include <mutex>
+
+#include "check.h"
+#include "log.h"
+#include "sleep.h"
+#include "string_util.h"
+
+namespace benchmark {
+
+// Suppress unused warnings on helper functions.
+#if defined(__GNUC__)
+#pragma GCC diagnostic ignored "-Wunused-function"
+#endif
+
+namespace {
+#if defined(BENCHMARK_OS_WINDOWS)
+double MakeTime(FILETIME const& kernel_time, FILETIME const& user_time) {
+  ULARGE_INTEGER kernel;
+  ULARGE_INTEGER user;
+  kernel.HighPart = kernel_time.dwHighDateTime;
+  kernel.LowPart = kernel_time.dwLowDateTime;
+  user.HighPart = user_time.dwHighDateTime;
+  user.LowPart = user_time.dwLowDateTime;
+  return (static_cast<double>(kernel.QuadPart) +
+          static_cast<double>(user.QuadPart)) *
+         1e-7;
+}
+#else
+double MakeTime(struct rusage const& ru) {
+  return (static_cast<double>(ru.ru_utime.tv_sec) +
+          static_cast<double>(ru.ru_utime.tv_usec) * 1e-6 +
+          static_cast<double>(ru.ru_stime.tv_sec) +
+          static_cast<double>(ru.ru_stime.tv_usec) * 1e-6);
+}
+#endif
+#if defined(BENCHMARK_OS_MACOSX)
+double MakeTime(thread_basic_info_data_t const& info) {
+  return (static_cast<double>(info.user_time.seconds) +
+          static_cast<double>(info.user_time.microseconds) * 1e-6 +
+          static_cast<double>(info.system_time.seconds) +
+          static_cast<double>(info.system_time.microseconds) * 1e-6);
+}
+#endif
+#if defined(CLOCK_PROCESS_CPUTIME_ID) || defined(CLOCK_THREAD_CPUTIME_ID)
+double MakeTime(struct timespec const& ts) {
+  return ts.tv_sec + (static_cast<double>(ts.tv_nsec) * 1e-9);
+}
+#endif
+
+BENCHMARK_NORETURN static void DiagnoseAndExit(const char* msg) {
+  std::cerr << "ERROR: " << msg << std::endl;
+  std::exit(EXIT_FAILURE);
+}
+
+}  // end namespace
+
+double ProcessCPUUsage() {
+#if defined(BENCHMARK_OS_WINDOWS)
+  HANDLE proc = GetCurrentProcess();
+  FILETIME creation_time;
+  FILETIME exit_time;
+  FILETIME kernel_time;
+  FILETIME user_time;
+  if (GetProcessTimes(proc, &creation_time, &exit_time, &kernel_time,
+                      &user_time))
+    return MakeTime(kernel_time, user_time);
+  DiagnoseAndExit("GetProccessTimes() failed");
+#elif defined(BENCHMARK_OS_EMSCRIPTEN)
+  // clock_gettime(CLOCK_PROCESS_CPUTIME_ID, ...) returns 0 on Emscripten.
+  // Use Emscripten-specific API. Reported CPU time would be exactly the
+  // same as total time, but this is ok because there aren't long-latency
+  // syncronous system calls in Emscripten.
+  return emscripten_get_now() * 1e-3;
+#elif defined(CLOCK_PROCESS_CPUTIME_ID) && !defined(BENCHMARK_OS_MACOSX)
+  // FIXME We want to use clock_gettime, but its not available in MacOS 10.11. See
+  // https://github.com/google/benchmark/pull/292
+  struct timespec spec;
+  if (clock_gettime(CLOCK_PROCESS_CPUTIME_ID, &spec) == 0)
+    return MakeTime(spec);
+  DiagnoseAndExit("clock_gettime(CLOCK_PROCESS_CPUTIME_ID, ...) failed");
+#else
+  struct rusage ru;
+  if (getrusage(RUSAGE_SELF, &ru) == 0) return MakeTime(ru);
+  DiagnoseAndExit("getrusage(RUSAGE_SELF, ...) failed");
+#endif
+}
+
+double ThreadCPUUsage() {
+#if defined(BENCHMARK_OS_WINDOWS)
+  HANDLE this_thread = GetCurrentThread();
+  FILETIME creation_time;
+  FILETIME exit_time;
+  FILETIME kernel_time;
+  FILETIME user_time;
+  GetThreadTimes(this_thread, &creation_time, &exit_time, &kernel_time,
+                 &user_time);
+  return MakeTime(kernel_time, user_time);
+#elif defined(BENCHMARK_OS_MACOSX)
+  // FIXME We want to use clock_gettime, but its not available in MacOS 10.11. See
+  // https://github.com/google/benchmark/pull/292
+  mach_msg_type_number_t count = THREAD_BASIC_INFO_COUNT;
+  thread_basic_info_data_t info;
+  mach_port_t thread = pthread_mach_thread_np(pthread_self());
+  if (thread_info(thread, THREAD_BASIC_INFO, (thread_info_t)&info, &count) ==
+      KERN_SUCCESS) {
+    return MakeTime(info);
+  }
+  DiagnoseAndExit("ThreadCPUUsage() failed when evaluating thread_info");
+#elif defined(BENCHMARK_OS_EMSCRIPTEN)
+  // Emscripten doesn't support traditional threads
+  return ProcessCPUUsage();
+#elif defined(BENCHMARK_OS_RTEMS)
+  // RTEMS doesn't support CLOCK_THREAD_CPUTIME_ID. See
+  // https://github.com/RTEMS/rtems/blob/master/cpukit/posix/src/clockgettime.c
+  return ProcessCPUUsage();
+#elif defined(CLOCK_THREAD_CPUTIME_ID)
+  struct timespec ts;
+  if (clock_gettime(CLOCK_THREAD_CPUTIME_ID, &ts) == 0) return MakeTime(ts);
+  DiagnoseAndExit("clock_gettime(CLOCK_THREAD_CPUTIME_ID, ...) failed");
+#else
+#error Per-thread timing is not available on your system.
+#endif
+}
+
+namespace {
+
+std::string DateTimeString(bool local) {
+  typedef std::chrono::system_clock Clock;
+  std::time_t now = Clock::to_time_t(Clock::now());
+  const std::size_t kStorageSize = 128;
+  char storage[kStorageSize];
+  std::size_t written;
+
+  if (local) {
+#if defined(BENCHMARK_OS_WINDOWS)
+    written =
+        std::strftime(storage, sizeof(storage), "%x %X", ::localtime(&now));
+#else
+    std::tm timeinfo;
+    std::memset(&timeinfo, 0, sizeof(std::tm));
+    ::localtime_r(&now, &timeinfo);
+    written = std::strftime(storage, sizeof(storage), "%F %T", &timeinfo);
+#endif
+  } else {
+#if defined(BENCHMARK_OS_WINDOWS)
+    written = std::strftime(storage, sizeof(storage), "%x %X", ::gmtime(&now));
+#else
+    std::tm timeinfo;
+    std::memset(&timeinfo, 0, sizeof(std::tm));
+    ::gmtime_r(&now, &timeinfo);
+    written = std::strftime(storage, sizeof(storage), "%F %T", &timeinfo);
+#endif
+  }
+  CHECK(written < kStorageSize);
+  ((void)written);  // prevent unused variable in optimized mode.
+  return std::string(storage);
+}
+
+}  // end namespace
+
+std::string LocalDateTimeString() { return DateTimeString(true); }
+
+}  // end namespace benchmark

benchmarks/thirdparty/benchmark/src/timers.h

@@ -0,0 +1,48 @@
+#ifndef BENCHMARK_TIMERS_H
+#define BENCHMARK_TIMERS_H
+
+#include <chrono>
+#include <string>
+
+namespace benchmark {
+
+// Return the CPU usage of the current process
+double ProcessCPUUsage();
+
+// Return the CPU usage of the children of the current process
+double ChildrenCPUUsage();
+
+// Return the CPU usage of the current thread
+double ThreadCPUUsage();
+
+#if defined(HAVE_STEADY_CLOCK)
+template <bool HighResIsSteady = std::chrono::high_resolution_clock::is_steady>
+struct ChooseSteadyClock {
+  typedef std::chrono::high_resolution_clock type;
+};
+
+template <>
+struct ChooseSteadyClock<false> {
+  typedef std::chrono::steady_clock type;
+};
+#endif
+
+struct ChooseClockType {
+#if defined(HAVE_STEADY_CLOCK)
+  typedef ChooseSteadyClock<>::type type;
+#else
+  typedef std::chrono::high_resolution_clock type;
+#endif
+};
+
+inline double ChronoClockNow() {
+  typedef ChooseClockType::type ClockType;
+  using FpSeconds = std::chrono::duration<double, std::chrono::seconds::period>;
+  return FpSeconds(ClockType::now().time_since_epoch()).count();
+}
+
+std::string LocalDateTimeString();
+
+}  // end namespace benchmark
+
+#endif  // BENCHMARK_TIMERS_H