…ess.

Add a timeout to all such instances.

Thread-local C++ objects should be a portable way to get thread destructors.

The slp_* functions need to be compiled as C because we call them from assembly.

forward declarations should also be c linkage, otherwise MSVC complains.

This eliminates the complexity of green_updatecurrent, I believe. Much more can be simplified as well.

Still need to add the thread cleanup functions.

Currently, we're somewhat slower than the global variable case, as
expected.

| Benchmark                    | gl-1.1.2 | tl1                   |
|------------------------------|:--------:|:---------------------:|
| create a greenlet            | 68.7 ns  | 88.0 ns: 1.28x slower |
| switch between two greenlets | 301 ns   | 435 ns: 1.45x slower  |
| getcurrent single thread     | 17.9 ns  | 31.8 ns: 1.78x slower |
| chain(100000)                | 256 ms   | 275 ms: 1.08x slower  |
| Geometric mean               | (ref)    | 1.37x slower          |

Also some Windows debugging.

Amusingly, this only caused issues on 64-bit Python 3.9 and 3.10 an Windows.

…global in g_switchstack into a single copy, thus messing up the stack.

Try accessing it with the global variable each time.

This makes clang, at least, look it up each time, even with optimizations on.

We're now using Py_AddPendingCall to clean up some thread-local things that don't clean themselves up. I'm still working to be sure we break cycles.

You need special syntax to return a pointer in a noinline function in MSVC.

Because the parser ignores, without warning, __declspec after a * character.
https://docs.microsoft.com/en-us/cpp/cpp/declspec?view=msvc-160

Improve the debugging to not do that. Also limit it to a suspected issue for now.

This should let us restore support for older compilers.

I haven't tested this on other platforms or with a more complex app like gevent yet though.

…s we take.

Also add a testing API to be sure cleanups have finished.

…f the compile on old MSVC.

may need to force linking with kernel32

This way we don't have to grow the greenlet class by a pointer that is almost always empty. Also brings more type safety to the code.

…pilation error on MSVC.

…re it.

It passes for me locally.

…i to test this.

…e can save just this one.

Unfortunately, one still does and we have to manually cleanup a few others. Those can be fixed later.

This is less error prone.