Barriers fixed and tested more extensively.

	* barrier.c: Fix several bugs in all routines. Now passes
	tests/barrier5.c which is fairly rigorous. There is still
	a non-optimal work-around for a race condition between
	the barrier breeched event signal and event wait. Basically
	the last (signalling) thread to hit the barrier yields
	to allow any other threads, which may have lost the race,
	to complete.

tests/ChangeLog:
	* barrier3.c: Fixed.
	* barrier4.c: Fixed.
	* barrier5.c: New; proves that all threads in the group
	reaching the barrier wait and then resume together. Repeats
	the test using groups of 1 to 16 threads. Each group of
	threads must negotiate a large number of barriers (10000).
	* spin4.c: Fixed.
	* test.h (error_string): Modified the success (0) value.

1 files changed, 1 insertions, 1 deletions

diff --git a/implement.h b/implement.h
index 5a146c0..5b30d1a 100644
--- a/implement.h
+++ b/implement.h
@@ -189,8 +189,8 @@ struct pthread_spinlock_t_ {
 struct pthread_barrier_t_ {
   LONG nCurrentBarrierHeight;
   LONG nInitialBarrierHeight;
-  sem_t semBarrierBreeched;
   pthread_mutex_t mtxExclusiveAccess;
+  HANDLE eventBarrierBreeched;
 };
 
 struct pthread_barrierattr_t_ {