This function is super high in the profile because there is a high contention on the lock. Best to remove it.

oops, forgot to rebase over your de-virtualization.

(forgot to qref)

Comment on attachment 8340450 [details] [diff] [review] Stop locking when getting the preferred samplerate from the AudioStream. r= Review of attachment 8340450 [details] [diff] [review]: ----------------------------------------------------------------- ::: content/media/AudioStream.h @@ +382,5 @@ > static cubeb* sCubebContext; > > // Prefered samplerate, in Hz (characteristic of the > // hardware/mixer/platform/API used). > static uint32_t sPreferredSampleRate; Move this above the mutex comment since it no longer applies.

So, this regressed bug 723793. Oops.

I'm going to back this out for now, reopen the bug, and we'll fix it a different way here. https://hg.mozilla.org/integration/mozilla-inbound/rev/66b206715056 Sorry for missing this in the review. :-/

So, backing up a step, why are we calling PreferredSampleRate() so often that it matters how fast it is? Looks like the only two callers are: - in the constructor for WebAudio's AudioContext, which shouldn't be particularly performance critical - in MediaStreamGraphImpl::RunThread, inside the processing loop, via IdealAudioRate(), which is probably where the problem stems from Does the MSG expect the ideal rate to change? If not, we can hoist the call to IdealRate() out of the loop. If it does, we can teach the IdealRate() implementation that the result from PreferredSampleRate() won't change and therefore is safe to cache.

Aside from that question, we could try something like this. Store the value in an Atomic, and only take the lock when necessary to initialize the value.

FWIW, I don't know if this patch will help, since an atomic shouldn't be much cheaper than an uncontended lock on platforms with decent locks.

Maybe this is better?

Comment on attachment 8345422 [details] [diff] [review] patch Looks good. - // Get the preferred samplerate for this platform, or fallback to something "sample rate" + // Queries the samplerate the hardware/mixer runs at, and stores it. ditto + // Get the aformentionned sample rate. Does not lock. aforementioned

Backed out for crashes. https://tbpl.mozilla.org/?tree=Mozilla-Inbound&rev=63ae0ef30e3f

Relanded with the obvious error fixed. (|| instead of && and use the unlocked version of GetCubebContext) https://hg.mozilla.org/integration/mozilla-inbound/rev/bc103864507b