Skip to content

performance: Optimize shared_mutex and fix C++20 modular build errors #7007

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
hkaiser merged 12 commits into from
May 10, 2026
+106 −17
Merged

performance: Optimize shared_mutex and fix C++20 modular build errors #7007

Show file tree
Hide file tree
Changes from 9 commits
Commits
Show all changes
12 commits
Select commit Hold shift + click to select a range
d5bf3c3
performance: Optimize shared_mutex with lock-free fast paths and stru…
Mar 24, 2026
6acdd90
performance: Address maintainer feedback on shared_mutex spin loops a…
arpittkhandelwal Apr 16, 2026
47de6f4
performance: Refine set_state short-circuit and reuse state from fail…
arpittkhandelwal Apr 17, 2026
a371a39
performance: Replace all redundant state.load with s=s1 in spin loops
arpittkhandelwal Apr 18, 2026
f00d4da
performance: Fix clang-format line wrap in set_state
arpittkhandelwal Apr 18, 2026
13ad2b2
performance: Eliminate redundant atomic refcounts in shared_mutex wra…
arpittkhandelwal Apr 19, 2026
5c3a426
style: Fix clang-format line wrapping
arpittkhandelwal Apr 19, 2026
38d7f22
performance: Final optimized shared_mutex with lifetime safety refcounts
arpittkhandelwal Apr 19, 2026
e7605f4
style: Match clang-format requirements for compare_exchange_strong
arpittkhandelwal Apr 19, 2026
1ed78e7
style: fix clang-format issues in shared_mutex.hpp
arpittkhandelwal Apr 29, 2026
81041ec
performance: Simplify unlock_shared and add missing include to benchmark
arpittkhandelwal May 6, 2026
cb7f0af
style: Fix clang-format issues in shared_mutex.hpp and shared_mutex_o…
arpittkhandelwal May 6, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
68 changes: 51 additions & 17 deletions libs/core/synchronization/include/hpx/synchronization/shared_mutex.hpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -77,20 +77,18 @@ namespace hpx::detail {
bool set_state(shared_state& s1, shared_state& s) noexcept
{
++s.data.tag;
return s1.value == state.load(std::memory_order_relaxed).value &&
state.compare_exchange_strong(s1, s, std::memory_order_release);
return state.compare_exchange_strong(
s1, s, std::memory_order_release, std::memory_order_relaxed);
}

bool set_state(shared_state& s1, shared_state& s,
std::unique_lock<mutex_type>& lk) noexcept
{
if (s1.value != state.load(std::memory_order_relaxed).value)
return false;

++s.data.tag;

lk = std::unique_lock<mutex_type>(state_change);
if (state.compare_exchange_strong(s1, s, std::memory_order_release))
if (state.compare_exchange_strong(
s1, s, std::memory_order_release, std::memory_order_relaxed))
return true;

lk.unlock();
Expand Down Expand Up @@ -121,9 +119,9 @@ namespace hpx::detail {

bool try_lock_shared()
{
auto s = state.load(std::memory_order_acquire);
while (true)
{
auto s = state.load(std::memory_order_acquire);
if (s.data.exclusive || s.data.exclusive_waiting_blocked)
{
return false;
Expand All @@ -136,15 +134,37 @@ namespace hpx::detail {
{
break;
}
s = s1;
}
Comment thread
hkaiser marked this conversation as resolved.
return true;
}
Comment thread
hkaiser marked this conversation as resolved.

bool try_unlock_shared_fast()
{
auto s = state.load(std::memory_order_acquire);
while (true)
{
if (s.data.exclusive || s.data.exclusive_waiting_blocked ||
s.data.upgrade || s.data.shared_count <= 1)
{
return false;
}

auto s1 = s;
--s.data.shared_count;
if (set_state(s1, s))
{
return true;
}
s = s1;
}
}

void unlock_shared()
{
auto s = state.load(std::memory_order_acquire);
while (true)
{
auto s = state.load(std::memory_order_acquire);
auto s1 = s;

if (--s.data.shared_count == 0)
Expand Down Expand Up @@ -184,14 +204,15 @@ namespace hpx::detail {
{
break;
}
s = s1;
}
}

void lock()
{
auto s = state.load(std::memory_order_acquire);
while (true)
{
auto s = state.load(std::memory_order_acquire);
while (s.data.shared_count != 0 || s.data.exclusive)
{
auto s1 = s;
Expand All @@ -214,14 +235,15 @@ namespace hpx::detail {
{
break;
}
s = s1;
}
}

bool try_lock()
{
auto s = state.load(std::memory_order_acquire);
while (true)
{
auto s = state.load(std::memory_order_acquire);
if (s.data.shared_count || s.data.exclusive)
{
return false;
Expand All @@ -234,15 +256,16 @@ namespace hpx::detail {
{
break;
}
s = s1;
}
return true;
}

void unlock()
{
auto s = state.load(std::memory_order_acquire);
while (true)
{
auto s = state.load(std::memory_order_acquire);
auto s1 = s;

s.data.exclusive = false;
Expand All @@ -255,6 +278,7 @@ namespace hpx::detail {
release_waiters(lk);
break;
}
s = s1;
}
}

Expand Down Expand Up @@ -287,9 +311,9 @@ namespace hpx::detail {

bool try_lock_upgrade()
{
auto s = state.load(std::memory_order_acquire);
while (true)
{
auto s = state.load(std::memory_order_acquire);
if (s.data.exclusive || s.data.exclusive_waiting_blocked ||
s.data.upgrade)
{
Expand All @@ -304,15 +328,16 @@ namespace hpx::detail {
{
break;
}
s = s1;
}
return true;
}

void unlock_upgrade()
{
auto s = state.load(std::memory_order_acquire);
while (true)
{
auto s = state.load(std::memory_order_acquire);
auto s1 = s;

bool release = false;
Expand All @@ -337,6 +362,7 @@ namespace hpx::detail {
{
break;
}
s = s1;
}
}

Expand Down Expand Up @@ -384,9 +410,9 @@ namespace hpx::detail {

void unlock_and_lock_upgrade()
{
auto s = state.load(std::memory_order_acquire);
while (true)
{
auto s = state.load(std::memory_order_acquire);
auto s1 = s;

s.data.exclusive = false;
Expand All @@ -401,14 +427,15 @@ namespace hpx::detail {
release_waiters(lk);
break;
}
s = s1;
}
}

void unlock_and_lock_shared()
{
auto s = state.load(std::memory_order_acquire);
while (true)
{
auto s = state.load(std::memory_order_acquire);
auto s1 = s;

s.data.exclusive = false;
Expand All @@ -422,14 +449,15 @@ namespace hpx::detail {
release_waiters(lk);
break;
}
s = s1;
}
}

bool try_unlock_shared_and_lock()
{
auto s = state.load(std::memory_order_acquire);
while (true)
{
auto s = state.load(std::memory_order_acquire);
if (s.data.exclusive || s.data.exclusive_waiting_blocked ||
s.data.upgrade || s.data.shared_count != 1)
{
Expand All @@ -444,15 +472,16 @@ namespace hpx::detail {
{
break;
}
s = s1;
}
return true;
}

void unlock_upgrade_and_lock_shared()
{
auto s = state.load(std::memory_order_acquire);
while (true)
{
auto s = state.load(std::memory_order_acquire);
auto s1 = s;

s.data.exclusive_waiting_blocked = false;
Expand All @@ -465,6 +494,7 @@ namespace hpx::detail {
release_waiters(lk);
break;
}
s = s1;
}
}

Expand Down Expand Up @@ -510,6 +540,8 @@ namespace hpx::detail {
void lock_shared()
{
auto data = data_;
if (data->try_lock_shared())
return;
data->lock_shared();
}

Expand All @@ -522,6 +554,8 @@ namespace hpx::detail {
void unlock_shared()
{
auto data = data_;
if (data->try_unlock_shared_fast())
return;
Comment thread
arpittkhandelwal marked this conversation as resolved.
Outdated
data->unlock_shared();
}

Expand Down
2 changes: 2 additions & 0 deletions tests/performance/local/CMakeLists.txt
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -25,6 +25,7 @@ set(benchmarks
skynet
wait_all_timings
benchmark_stealing
shared_mutex_overhead
)

set(timed_task_spawn_SOURCES activate_counters.cpp)
Expand Down Expand Up @@ -144,6 +145,7 @@ set(print_heterogeneous_payloads_PARAMETERS NO_HPX_MAIN)
set(skynet_PARAMETERS NO_HPX_MAIN)
set(timed_task_spawn_PARAMETERS NO_HPX_MAIN)
set(benchmark_stealing_PARAMETERS NO_HPX_MAIN)
set(shared_mutex_overhead_PARAMETERS NO_HPX_MAIN)
set(hpx_tls_overhead_PARAMETERS NO_HPX_MAIN)
Comment thread
arpittkhandelwal marked this conversation as resolved.
set(native_tls_overhead_PARAMETERS NO_HPX_MAIN)
set(coroutines_call_overhead_PARAMETERS NO_HPX_MAIN)
Expand Down
75 changes: 75 additions & 0 deletions tests/performance/local/shared_mutex_overhead.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,75 @@
// (C) Copyright 2026 Arpit Khandelwal
//
// SPDX-License-Identifier: BSL-1.0
// Distributed under the Boost Software License, Version 1.0. (See accompanying
// file LICENSE_1_0.txt or copy at http://www.boost.org/LICENSE_1_0.txt)

#include <hpx/config.hpp>
#include <hpx/hpx.hpp>
#include <hpx/hpx_init.hpp>
#include <hpx/include/util.hpp>
#include <hpx/modules/testing.hpp>
#include <hpx/synchronization/shared_mutex.hpp>

#include <cstdint>
#include <iostream>
Comment thread
hkaiser marked this conversation as resolved.
#include <vector>

std::uint64_t num_iterations = 100000;
std::uint64_t reader_threads = 4;

hpx::shared_mutex mtx;

void reader()
{
for (std::uint64_t i = 0; i < num_iterations; ++i)
{
std::shared_lock<hpx::shared_mutex> l(mtx);
}
Comment thread
arpittkhandelwal marked this conversation as resolved.
}

int hpx_main(hpx::program_options::variables_map& vm)
{
num_iterations = vm["iterations"].as<std::uint64_t>();
reader_threads = hpx::get_num_worker_threads();

std::cout << "Starting benchmark with " << reader_threads << " threads..."
<< std::endl;

std::vector<hpx::future<void>> futures;
futures.reserve(reader_threads);

hpx::chrono::high_resolution_timer walltime;

for (std::uint64_t i = 0; i < reader_threads; ++i)
{
futures.push_back(hpx::async(&reader));
}

hpx::wait_all(futures);

double const duration = walltime.elapsed();

std::cout << "Total time: " << duration << " seconds" << std::endl;
std::cout << "Average time per reader thread: " << duration / reader_threads
<< " seconds" << std::endl;

hpx::util::print_cdash_timing("SharedMutexOverhead", duration);

return hpx::local::finalize();
}

int main(int argc, char* argv[])
{
hpx::program_options::options_description cmdline(
"usage: " HPX_APPLICATION_STRING " [options]");

cmdline.add_options()("iterations",
hpx::program_options::value<std::uint64_t>()->default_value(100000),
"number of iterations per thread");

hpx::local::init_params init_args;
init_args.desc_cmdline = cmdline;

return hpx::local::init(hpx_main, argc, argv, init_args);
}
Loading