Files
ARMSX2/common/Semaphore.cpp
T
pstef 2e39fcc216 Optimization: wait on the address instead of spinning the pipeline
Each of the three spin-then-sleep semaphore loops watches a single
atomic word, but ShortSpin() has no way to know that: it spends its
share of SPIN_TIME_NS in batches of eight isb, and every one of those is
a pipeline flush. arm64 can watch the word itself, so ShortSpinOn()
does, and the loops that hold their whole predicate in one word take it.

Hosts that cannot watch an address keep the old spin.

Inspired by Whatcookie's work on arm64 for rpcs3.
2026-08-09 11:20:53 +02:00

215 lines
6.0 KiB
C++

// SPDX-FileCopyrightText: 2002-2026 PCSX2 Dev Team
// SPDX-License-Identifier: GPL-3.0+
#include "common/Threading.h"
#include "common/Assertions.h"
#include "common/HostSys.h"
#ifdef _WIN32
#include "common/RedtapeWindows.h"
#endif
#include <limits>
// --------------------------------------------------------------------------------------
// Semaphore Implementations
// --------------------------------------------------------------------------------------
bool Threading::WorkSema::CheckForWork()
{
s32 value = m_state.load(std::memory_order_relaxed);
pxAssert(!IsDead(value));
// we want to switch to the running state, but preserve the waiting empty bit for RUNNING_N -> RUNNING_0
// otherwise, we clear the waiting flag (since we're notifying the waiter that we're empty below)
while (!m_state.compare_exchange_weak(value,
IsReadyForSleep(value) ? STATE_RUNNING_0 : (value & STATE_FLAG_WAITING_EMPTY),
std::memory_order_acq_rel, std::memory_order_relaxed))
{
}
// if we're not empty, we have work to do
if (!IsReadyForSleep(value))
return true;
// this means we're empty, so notify any waiters
if (value & STATE_FLAG_WAITING_EMPTY)
m_empty_sema.Post();
// no work to do
return false;
}
void Threading::WorkSema::WaitForWork()
{
// State change:
// SLEEPING, SPINNING: This is the worker thread and it's clearly not asleep or spinning, so these states should be impossible
// RUNNING_0: Change state to SLEEPING, wake up thread if WAITING_EMPTY
// RUNNING_N: Change state to RUNNING_0 (and preserve WAITING_EMPTY flag)
s32 value = m_state.load(std::memory_order_relaxed);
pxAssert(!IsDead(value));
while (!m_state.compare_exchange_weak(value, NextStateWaitForWork(value), std::memory_order_acq_rel, std::memory_order_relaxed))
;
if (IsReadyForSleep(value))
{
if (value & STATE_FLAG_WAITING_EMPTY)
m_empty_sema.Post();
m_sema.Wait();
// Acknowledge any additional work added between wake up request and getting here
m_state.fetch_and(STATE_FLAG_WAITING_EMPTY, std::memory_order_acquire);
}
}
void Threading::WorkSema::WaitForWorkWithSpin()
{
s32 value = m_state.load(std::memory_order_relaxed);
pxAssert(!IsDead(value));
while (IsReadyForSleep(value))
{
if (m_state.compare_exchange_weak(value, STATE_SPINNING, std::memory_order_release, std::memory_order_relaxed))
{
if (value & STATE_FLAG_WAITING_EMPTY)
m_empty_sema.Post();
value = STATE_SPINNING;
break;
}
}
u32 waited = 0;
while (value < 0)
{
if (waited > SPIN_TIME_NS)
{
if (!m_state.compare_exchange_weak(value, STATE_SLEEPING, std::memory_order_relaxed))
continue;
m_sema.Wait();
break;
}
waited += ShortSpinOn(m_state, value);
value = m_state.load(std::memory_order_relaxed);
}
// Clear back to STATE_RUNNING_0 (but preserve waiting empty flag)
m_state.fetch_and(STATE_FLAG_WAITING_EMPTY, std::memory_order_acquire);
}
bool Threading::WorkSema::WaitForEmpty()
{
s32 value = m_state.load(std::memory_order_acquire);
while (true)
{
if (value < 0)
return !IsDead(value); // STATE_SLEEPING or STATE_SPINNING, queue is empty!
// Note: We technically only need memory_order_acquire on *failure* (because that's when we could leave without sleeping), but libstdc++ still asserts on failure < success
if (m_state.compare_exchange_weak(value, value | STATE_FLAG_WAITING_EMPTY, std::memory_order_acquire))
break;
}
pxAssertMsg(!(value & STATE_FLAG_WAITING_EMPTY), "Multiple threads attempted to wait for empty (not currently supported)");
m_empty_sema.Wait();
return !IsDead(m_state.load(std::memory_order_relaxed));
}
bool Threading::WorkSema::WaitForEmptyWithSpin()
{
s32 value = m_state.load(std::memory_order_acquire);
u32 waited = 0;
while (true)
{
if (value < 0)
return !IsDead(value); // STATE_SLEEPING or STATE_SPINNING, queue is empty!
if (waited > SPIN_TIME_NS && m_state.compare_exchange_weak(value, value | STATE_FLAG_WAITING_EMPTY, std::memory_order_acquire))
break;
waited += ShortSpinOn(m_state, value);
value = m_state.load(std::memory_order_acquire);
}
pxAssertMsg(!(value & STATE_FLAG_WAITING_EMPTY), "Multiple threads attempted to wait for empty (not currently supported)");
m_empty_sema.Wait();
return !IsDead(m_state.load(std::memory_order_relaxed));
}
void Threading::WorkSema::Kill()
{
s32 value = m_state.exchange(std::numeric_limits<s32>::min(), std::memory_order_release);
if (value & STATE_FLAG_WAITING_EMPTY)
m_empty_sema.Post();
}
void Threading::WorkSema::Reset()
{
m_state = STATE_RUNNING_0;
}
void Threading::UserspaceSemaphore::WaitWithSpin()
{
// See header for rationale. The peek-and-CAS spin path is the win — when
// the producer is about to Post (within ~50µs), we acquire in user space
// and skip the futex syscall entirely.
int32_t counter = m_counter.load(std::memory_order_relaxed);
u32 waited = 0;
while (true)
{
while (counter > 0)
{
if (m_counter.compare_exchange_weak(counter, counter - 1,
std::memory_order_acquire, std::memory_order_relaxed))
{
return;
}
}
if (waited >= SPIN_TIME_NS)
break;
waited += ShortSpinOn(m_counter, counter);
counter = m_counter.load(std::memory_order_relaxed);
}
// Spin window expired — block in the kernel (same as plain Wait()).
if (m_counter.fetch_sub(1, std::memory_order_acquire) <= 0)
m_sema.Wait();
}
#if !defined(__APPLE__) // macOS implementations are in DarwinThreads
Threading::KernelSemaphore::KernelSemaphore()
{
#ifdef _WIN32
m_sema = CreateSemaphore(nullptr, 0, LONG_MAX, nullptr);
#else
sem_init(&m_sema, false, 0);
#endif
}
Threading::KernelSemaphore::~KernelSemaphore()
{
#ifdef _WIN32
CloseHandle(m_sema);
#else
sem_destroy(&m_sema);
#endif
}
void Threading::KernelSemaphore::Post()
{
#ifdef _WIN32
ReleaseSemaphore(m_sema, 1, nullptr);
#else
sem_post(&m_sema);
#endif
}
void Threading::KernelSemaphore::Wait()
{
#ifdef _WIN32
WaitForSingleObject(m_sema, INFINITE);
#else
sem_wait(&m_sema);
#endif
}
bool Threading::KernelSemaphore::TryWait()
{
#ifdef _WIN32
return WaitForSingleObject(m_sema, 0) == WAIT_OBJECT_0;
#else
return sem_trywait(&m_sema) == 0;
#endif
}
#endif