On Wed, Sep 16, 2009 at 11:57:04PM +0200, Lennart Poettering wrote:
> On Wed, 16.09.09 15:15, Daniel Mack (daniel@caiaq.de) wrote:
>
> > + s = pa_xnew(pa_semaphore, 1);
> > + MPCreateSemaphore(UINT_MAX, value, &(s->sema));
> > + pa_assert(s->sema != 0);
>
> Hmm, I'd prefer if the ret val of MPCreateSemaphore() would be checked
> here.
>
> Also I find it a bit weird checking for s->sema, though not
> initializing it to 0 in the beginning. If the call actually failed,
> then the assert will check uninitialized memory. Also, comparing
> pointers with 0 sucks. That should be NULL.
>
> Given that this can not realisitically fail, only in OOM or OOM-like
> situations in which case we abort anyway it mght be enough just writing:
>
> pa_assert_se(MPCreateSemaphore(UINT_MAX, value, &s->sema) == 0);
>
> (Assuming that success is signalled by retval == 0 on MacOSX)
>
> > +void pa_semaphore_free(pa_semaphore *s) {
> > + pa_assert(s);
> > + MPDeleteSemaphore(s->sema);
>
> Same here.
>
> > + pa_xfree(s);
> > +}
> > +
> > +void pa_semaphore_post(pa_semaphore *s) {
> > + pa_assert(s);
> > + MPSignalSemaphore(s->sema);
>
> And here.
>
> > +}
> > +
> > +void pa_semaphore_wait(pa_semaphore *s) {
> > + pa_assert(s);
> > + /* should probably check return value (-ve is error), noErr is ok. */
> > + MPWaitOnSemaphore(s->sema, kDurationForever);
>
> And here.
Ok, done. See the patch below.
Daniel
>From 26df2fbae6d9215a3ae084876fb5f79e4d9cf4f0 Mon Sep 17 00:00:00 2001
From: Kim Lester <kim@dfusion.com.au>
Date: Wed, 16 Sep 2009 09:23:39 +0800
Subject: [PATCH] Mac OS X: add semaphore implementation
On Wed, Sep 16, 2009 at 11:48:58PM +0200, Lennart Poettering wrote:
> On Wed, 16.09.09 15:15, Daniel Mack (daniel@caiaq.de) wrote:
>
> > From: Kim Lester <kim@dfusion.com.au>
> >
> > OS X does not define clockid_t or clock_gettime() and friends.
> > Add a wrapper to fix this.
>
> Hmpf. I am not particularly happy with this. This adds a lot of
> unnecessary compat code. We don't actually need implementations of
> clock_getres(). All we need is some kind of check whether system
> timers are accurate or whether they are rounded up to scheduling
> slices. On Linux we do that check with clock_getres(), but all the
> information it returns is actually not intertesting at all. We just
> check if this is below some trheshold, that's all.
>
> clock_settime() we don't use at all! We shouldn't carry compat code
> for that.
>
> And clock_gettime we don't really need either. We need some kind of
> accurate system timers (preferably monotonic), and on Linux we use
> clock_gettime() for that. But we already have a fallback there for
> gettimeofday().
>
> Or in other words, the current APIs pa_rtclock_get(),
> pa_rtclock_hrtimer() is supposed to be the abstract API that has
> different backends on different systems. I'd very much prefer if any
> MacOS specific code would simply be plugged in there instead of
> creating various new abstraction interfaces!
Ok - what about the version below? I don't particularily like the
Daniel
>From 9f0a051953ec354ccdb8aa44a9845c408b26ae0b Mon Sep 17 00:00:00 2001
From: Kim Lester <kim@dfusion.com.au>
Date: Wed, 16 Sep 2009 14:40:01 +0800
Subject: [PATCH] Implement pa_rtclock_get() and pa_rtclock_hrtimer() for Darwin
OS X does not define clockid_t or clock_gettime() and friends.
Add wrappers to fix this. Based on a patch from Kim Lester
<kim@dfusion.com.au>.
Tweak the constraints a little so that register starved 32bit systems
can select a stack variable for the channel paramter instead of reusing one of
the registers we're using in the code.
We can reorder the algortihm a bit like we do for sse so that we
don't need the contants and masking instructions. Saves 2 instructions
for the mmx code.
