native 0.0.1
Vectors, masks and wide register packs for C++26
Loading...
Searching...
No Matches
x86 WAITPKG/MWAITX: waiting and spin loops

x86 instruction sets

Why use it

A busy loop consumes execution resources while it waits for another thread. PAUSE gives the processor a hint; monitored waits can suspend until there is activity near an address, and TPAUSE can wait for a TSC deadline without monitoring memory. These are useful building blocks for short waits where a full scheduler handoff would cost too much.

Operations

Import native.x86.wait and link native::minimal. The x86 and main hubs also export the module.

Operation Behavior
spin::mwait() Issue one PAUSE; ignore the timer argument
mwaitx::monitor(p) Arm MONITORX for the address
mwaitx::mwait(timer) Issue MWAITX with a 32-bit timer operand
umwait::monitor(p) Arm UMONITOR for the address
umwait::mwait(deadline) Issue UMWAIT with an absolute 64-bit TSC deadline; return its status byte
tpause<Arch, Control>(deadline) Wait on an absolute 64-bit TSC deadline without arming a monitor; return its status byte
wait_until<W>(p, predicate) Check the predicate, arm the monitor, check again, and wait; repeat until the predicate holds
with_waiter(callback) Invoke the callback with a supported waiter type: prefer mwaitx, then umwait, then spin

mwaitx::supported and umwait::supported report admission of their respective features. spin::supported is true. spin::monitor(p) does nothing.

native::with_waiter([&]<native::waiter W> {
native::wait_until<W>(p, predicate);
});
native::waiter
some way to wait for a value to change
Definition native.x86.wait.ccm:31
native::wait_until
void wait_until(volatile auto *p, auto f) noexcept(noexcept(!f(p)) &&noexcept(W::monitor(p)) &&noexcept(W::mwait(0))) native_blocking
Wait until a predicate holds about a given memory location.
Definition native.x86.wait.ccm:47
native::with_waiter
auto with_waiter(auto k) noexcept(noexcept(k.template operator()<mwaitx >()) &&noexcept(k.template operator()<umwait >()) &&noexcept(k.template operator()<spin >()))
finds an appropriate waiter for the current CPU Prefer MWAITX, then WAITPKG, otherwise PAUSE....
Definition native.x86.wait.ccm:115

Caveats

These operations provide no memory ordering. The predicate must perform the atomic loads and synchronization your algorithm needs; volatile is not inter-thread synchronization. Wakeups may be spurious, so always recheck the condition. The OS can limit a wait, and monitor activity need not mean that the particular value you care about has changed.

MWAITX and WAITPKG are independent features and require no SIMD register state. Check supported before using a waiter directly. The MWAITX timer is not a TSC deadline. The wrapper passes zero for its extension flags, so passing a nonzero timer alone does not enable a timed wait.

For tpause, Arch defaults to the module's compiler baseline and must contain x86_feature::waitpkg; the caller must enable the waitpkg target and admit it before execution. Control=0 permits C0.2; the default Control=1 requests C0.1. Waits have no constant-evaluation substitute.