Mutex speedups cont'd

1 files changed, 80 insertions, 37 deletions

diff --git a/ChangeLog b/ChangeLog
index ec65d84..c1fe46f 100644
--- a/ChangeLog
+++ b/ChangeLog
@@ -1,45 +1,88 @@
-2004-10-08  Ross Johnson  <rpj at callisto.canberra.edu.au>
+2004-10-15  Ross Johnson  <rpj at callisto.canberra.edu.au>
 
-	* pthread_mutex_destroy.c (pthread_mutex_destroy): Critical Section
-	element is no longer required.
-	* pthread_mutex_init.c (pthread_mutex_init): Likewise.
-	* pthread_mutex_lock.c (pthread_mutex_lock): New algorithm following Drepper's
-	paper at http://people.redhat.com/drepper/futex.pdf, but using the existing
-	semaphore in place of the futex described in the paper. Idea suggested by
-	Alexander Terekhov - see:
-	http://sources.redhat.com/ml/pthreads-win32/2003/msg00108.html
-	* pthread_mutex_timedlock.c pthread_mutex_timedlock(): Similarly.
-	* pthread_mutex_trylock.c (pthread_mutex_trylock): Similarly.
-	* pthread_mutex_unlock.c (pthread_mutex_unlock): Similarly.
-	* pthread_barrier_wait.c (pthread_barrier_wait): Use inlined version of
-	InterlockedCompareExchange() if possible - determined at build-time.
-	* pthread_spin_destroy.c pthread_spin_destroy(): Likewise.
-	* pthread_spin_lock.c pthread_spin_lock():Likewise.
-	* pthread_spin_trylock.c (pthread_spin_trylock):Likewise.
-	* pthread_spin_unlock.c (pthread_spin_unlock):Likewise.
-	* ptw32_InterlockedCompareExchange.c: Sets up macro for inlined use.
-	* implement.h (pthread_mutex_t_): Remove Critical Section element.
-	(PTW32_INTERLOCKED_COMPARE_EXCHANGE): Set to default non-inlined version of
-	InterlockedCompareExchange().
-	* private.c: Include ptw32_InterlockedCompareExchange.c first for inlining.
-	* GNUmakefile: Add commandline option to use inlined InterlockedCompareExchange().
-	* Makefile: Likewise.
+	* implement.h (othread_mutex_t_): Use an event in place of
+	the POSIX semaphore.
+	* pthread_mutex_init.c: Create the event; remove semaphore init.
+	* pthread_mutex_destroy.c: Delete the event.
+	* pthread_mutex_lock.c: Replace the semaphore wait with the event wait.
+	* pthread_mutex_trylock.c: Likewise.
+	* pthread_mutex_timedlock.c: Likewise.
+	* pthread_mutex_unlock.c: Set the event.
+	
+2004-10-14  Ross Johnson  <rpj at callisto.canberra.edu.au>

+
+	* pthread_mutex_lock.c (pthread_mutex_lock): New algorithm using
+	Terekhov's xchg based variation of Drepper's cmpxchg model.
+	Theoretically, xchg uses fewer clock cycles than cmpxchg (using IA-32
+	as a reference), however, in my opinion bus locking dominates the
+	equation on smp systems, so the model with the least number of bus
+	lock operations in the execution path should win, which is Terekhov's
+	variant. On IA-32 uni-processor systems, it's faster to use the
+	CMPXCHG instruction without locking the bus than to use the XCHG
+	instruction, which always locks the bus. This makes the two variants
+	equal for the non-contended lock (fast lane) execution path on up
+	IA-32. Testing shows that the xchg variant is faster on up IA-32 as
+	well if the test forces higher lock contention frequency, even though
+	kernel calls should be dominating the times (on up IA-32, both
+	variants used CMPXCHG instructions and neither locked the bus).
+	* pthread_mutex_timedlock.c pthread_mutex_timedlock(): Similarly.

+	* pthread_mutex_trylock.c (pthread_mutex_trylock): Similarly.

+	* pthread_mutex_unlock.c (pthread_mutex_unlock): Similarly.

+	* ptw32_InterlockedCompareExchange.c (ptw32_InterlockExchange): New
+	function.
+	(PTW32_INTERLOCKED_EXCHANGE): Sets up macro to use inlined
+	ptw32_InterlockedExchange.

+	* implement.h (PTW32_INTERLOCKED_EXCHANGE): Set default to
+	InterlockedExchange().

+	* Makefile: Building using /Ob2 so that asm sections within inline
+	functions are inlined.

+
+2004-10-08  Ross Johnson  <rpj at callisto.canberra.edu.au>

+
+	* pthread_mutex_destroy.c (pthread_mutex_destroy): Critical Section

+	element is no longer required.

+	* pthread_mutex_init.c (pthread_mutex_init): Likewise.

+	* pthread_mutex_lock.c (pthread_mutex_lock): New algorithm following
+	Drepper's paper at http://people.redhat.com/drepper/futex.pdf, but
+	using the existing semaphore in place of the futex described in the
+	paper. Idea suggested by Alexander Terekhov - see:

+	http://sources.redhat.com/ml/pthreads-win32/2003/msg00108.html

+	* pthread_mutex_timedlock.c pthread_mutex_timedlock(): Similarly.

+	* pthread_mutex_trylock.c (pthread_mutex_trylock): Similarly.

+	* pthread_mutex_unlock.c (pthread_mutex_unlock): Similarly.

+	* pthread_barrier_wait.c (pthread_barrier_wait): Use inlined version
+	of InterlockedCompareExchange() if possible - determined at
+	build-time.

+	* pthread_spin_destroy.c pthread_spin_destroy(): Likewise.

+	* pthread_spin_lock.c pthread_spin_lock():Likewise.

+	* pthread_spin_trylock.c (pthread_spin_trylock):Likewise.

+	* pthread_spin_unlock.c (pthread_spin_unlock):Likewise.

+	* ptw32_InterlockedCompareExchange.c: Sets up macro for inlined use.

+	* implement.h (pthread_mutex_t_): Remove Critical Section element.

+	(PTW32_INTERLOCKED_COMPARE_EXCHANGE): Set to default non-inlined
+	version of InterlockedCompareExchange().

+	* private.c: Include ptw32_InterlockedCompareExchange.c first for
+	inlining.

+	* GNUmakefile: Add commandline option to use inlined
+	InterlockedCompareExchange().

+	* Makefile: Likewise.

 
 2004-09-27  Ross Johnson  <rpj at callisto.canberra.edu.au>

 
-	* pthread_mutex_lock.c (pthread_mutex_lock): Separate PTHREAD_MUTEX_NORMAL

-	logic since we do not need to keep or check some state required by other

-	mutex types; do not check mutex pointer arg for validity - leave this to

-	the system since we are only checking for NULL pointers. This should improve

-	speed of NORMAL mutexes and marginally improve speed of other type.

+	* pthread_mutex_lock.c (pthread_mutex_lock): Separate
+	PTHREAD_MUTEX_NORMAL logic since we do not need to keep or check some
+	state required by other mutex types; do not check mutex pointer arg
+	for validity - leave this to the system since we are only checking
+	for NULL pointers. This should improve speed of NORMAL mutexes and
+	marginally improve speed of other type.

 	* pthread_mutex_trylock.c (pthread_mutex_trylock): Likewise.

 	* pthread_mutex_unlock.c (pthread_mutex_unlock): Likewise; also avoid

-	entering the critical section for the no-waiters case, with approx. 30%

-	reduction in lock/unlock overhead for this case..

+	entering the critical section for the no-waiters case, with approx.
+	30% reduction in lock/unlock overhead for this case.

 	* pthread_mutex_timedlock.c (pthread_mutex_timedlock): Likewise; also

-	no longer keeps mutex if post-timeout second attempt succeeds - this will

-	assist applications that wish to impose strict lock deadlines, rather than

-	simply to escape from frozen locks.

+	no longer keeps mutex if post-timeout second attempt succeeds - this
+	will assist applications that wish to impose strict lock deadlines,
+	rather than simply to escape from frozen locks.

 
 2004-09-09  Tristan Savatier  <tristan at mpegtv.com>

 	* pthread.h (struct pthread_once_t_): Qualify the 'done' element

@@ -49,8 +92,8 @@
 	[Maintainer's note: the race condition is harmless on SPU systems

 	and only a problem on MPU systems if concurrent access results in an

 	exception (presumably generated by a hardware interrupt). There are

-	other instances of similar harmless race conditions that have not been

-	identified as issues.]

+	other instances of similar harmless race conditions that have not
+	been identified as issues.]

 
 2004-09-09  Ross Johnson  <rpj at callisto.canberra.edu.au>