QLatch: optimize to avoid a syscall when no one is waitingHEADdev

We do it by using the sign bit of the atomic to indicate that there are
no waiters waiting. That way, the counter only becomes a zero when all
expected counters have counted down *and* there is at least one waiter.

That means the countDown() code remains unchanged. On x86-64 and
AArch64, there is no change in the number of instructions in the inline
portion of wait() either. The non-inline portion uses a
__atomic_and_fetch() instead of atomic_fetch_and() so compilers will
generate LOCK AND for x86-64, LDCLR for AArch64 ARMv8.1, and AMOAND.W
for RISC-V.

This is more efficient than the Standard Libraries's current
implementations, which use a separate, global atomic out of a pool,
meaning that there could be a collision between two or more latches (or
any other users std::atomic waiting, such as std::semaphore and
std::barrier). Coupled with the fact that we futexWait() on the actual
latch's address (something libc++ can't / won't do), this implementation
should be overall much more efficient.

Change-Id: Ib5ce7a497e034ebabb2cfffd1761b02a44d548d3
Reviewed-by: Mårten Nordheim <[email protected]>

0 files changed, 0 insertions, 0 deletions