Mostly optimizing TLS accesses to bring local heap allocation performance
closer to that of oldsched. It's not completely at parity but removing the
branches involved in supporting oldsched and optimizing pthread_get/setspecific
to instead use our dedicated TCB slot will probably make up for it.
This pull request converts the scheduler from a naive shared queue scheduler to a naive workstealing scheduler. The deque is still a queue inside a lock, but there is still a substantial performance gain. Fiddling with the messaging benchmark I got a ~10x speedup and observed massively reduced memory usage.
There are still *many* locations for optimization, but based on my experience so far it is a clear performance win as it is now.
In the first commit it is obvious why some of the barriers can be changed to ```Relaxed```, but it is not as obvious for the once I changed in ```kill.rs```. The rationale for those is documented as part of the documenting commit.
Also the last commit is a temporary hack to prevent kill signals from being received in taskgroup cleanup code, which could be fixed in a more principled way once the old runtime is gone.
old design the TLS held the scheduler struct, and the scheduler struct
held the active task. This posed all sorts of weird problems due to
how we wanted to use the contents of TLS. The cleaner approach is to
leave the active task in TLS and have the task hold the scheduler. To
make this work out the scheduler has to run inside a regular task, and
then once that is the case the context switching code is massively
simplified, as instead of three possible paths there is only one. The
logical flow is also easier to follow, as the scheduler struct acts
somewhat like a "token" indicating what is active.
These changes also necessitated changing a large number of runtime
tests, and rewriting most of the runtime testing helpers.
Polish level is "low", as I will very soon start on more scheduler
changes that will require wiping the polish off. That being said there
should be sufficient comments around anything complex to make this
entirely respectable as a standalone commit.
Change the former repetition::
for 5.times { }
to::
do 5.times { }
.times() cannot be broken with `break` or `return` anymore; for those
cases, use a numerical range loop instead.
This removes a bunch of options from the task builder interface that are irrelevant to the new scheduler and were generally unused anyway. It also bumps the stack size of new scheduler tasks so that there's enough room to run rustc and changes the interface to `Thread` to not implicitly join threads on destruction, but instead require an explicit, and mandatory, call to `join`.
