no spinning when waiting, just exit; repr(c) job

src/praetor/mod.rs

@@ -600,6 +600,7 @@ mod job {
         error: ManuallyDrop<Option<Box<dyn Any + Send + 'static>>>,
     }
 
+    #[repr(C)]
     pub struct Job<T = ()> {
         /// tagged pointer, 8-aligned
         harness_and_state: TaggedAtomicPtr<usize, 3>,
@@ -763,8 +764,7 @@ mod job {
                         // after sleeping, state should be `Finished`
                     }
                     Err(state) => {
-                        // debug_assert_ne!(state, JobState::Pending as usize);
-
+                        // job finished under us, check if it was successful
                         if state == JobState::Finished as usize {
                             let err = unsafe { (&mut *self.err_or_link.get()).error.take() };
 
@@ -1070,7 +1070,8 @@ impl JobCounter {
 
     /// must only be called once
     pub unsafe fn wait(&self) {
-        _ = self.waker.lock().insert(std::thread::current());
+        // SAFETY: this is only called once, so the waker is guaranteed to be None.
+        assert!(self.waker.lock().replace(std::thread::current()).is_none());
 
         let count = self.jobs_pending.load(Ordering::SeqCst);
         if count > 0 {
@@ -1364,20 +1365,22 @@ impl WorkerThread {
                         continue 'outer;
                     }
                     None => {
-                        // TODO: spin2win
                         tracing::trace!("waiting for shared job, thread id: {:?}", self.index);
 
                         // TODO: wait on latch? if we have something that can
                         // signal being done, e.g. can be waited on instead of
                         // shared jobs, we should wait on it instead, but we
                         // would also want to receive shared jobs still?
+                        // Spin? probably just wastes CPU time.
                         // self.context.shared_job.wait(&mut guard);
                         // if spin.spin() {
                         //     // wait for more shared jobs.
                         //     // self.context.shared_job.wait(&mut guard);
                         //     return;
                         // }
-                        std::thread::yield_now();
+                        // Yield? same as spinning, really, so just exit and let the upstream use wait
+                        // std::thread::yield_now();
+                        return;
                     }
                 }
             }