riot-rs-threads: cortex-m: add compiler fence around wfi dance

[kaspar030]

Fixes Cortex-M sched() not seeing changes to the runqueue.

See #310 (comment).

As to why the fence is necessary, I cannot explain. The critical section should actually imply a fence. In this case, the runqueue accesses are heavily inlined, so maybe something is off with the optimizer. Or, we're doing something wrong with how we access THREADS, which is a lot more likely than a compiler bug. 😏

[elenaf9]

Or, we're doing something wrong with how we access THREADS, which is a lot more likely than a compiler bug.

 loop {
-        if let Some(res) = critical_section::with(|cs| {
-            let threads = unsafe { &mut *THREADS.as_ptr(cs) };
+        if let Some(res) = THREADS.with_mut(|mut threads| {
             let next_pid = match threads.runqueue.get_next() {

Would also fixes this issue. But as discussed in #187 (comment), this negatively impacts performance. My guess is that because it internally borrows the value instead of operating on the pointer, some optimizations are prevented?

Also, I tested using compiler_fence instead of fence, which didn't fix the issue. So I guess it's on the hardware side rather than the compiler?

Either way, I tested you fix and it works for me 👍.

[kaspar030]

Would also fixes this issue. But as discussed in #187 (comment), this negatively impacts performance. My guess is that because it internally borrows the value instead of operating on the pointer, some optimizations are prevented?

Could be ... We'll need to get back to this at some point.

Thanks for testing! 👍