On 26.11.2009 23:26, Eliot Miranda wrote:


Hi Levente,

On Thu, Nov 26, 2009 at 12:01 PM, Levente Uzonyi <leves@elte.hu> wrote:

On Thu, 26 Nov 2009, Eliot Miranda wrote:

I wouldn't bother trying to fix the primitive.  Both in VisualWorks and at Teleplace the experience has been the same, getting rid of the primitives for next, nextPut: and atEnd sped things up.  The system is simply too complex nowadays for these primitives to pay for themselves because they only cover a small number of cases.  The primitives support Array and ByteString bow there are so many different kinds of arrays being streamed over that the cost of primitive failures outweigh the benefits of primitive successes.

We are using these primitives in our own stream implementation like this:

next

       <primitive: 65>
       [ position < readLimit ] whileFalse: [ self receiveData ].
       ^buffer at: (position := position + 1)

And it does make a difference (buffer is a 8kiB sized ByteArray). Would this be faster without the primitive?

I can't guarantee it.  You'd have to measure.  All I can say is that we found it better to get rid of the stream primitives in VisualWorks and in the Cog JIT.  Looking at the Squeak VM code I think the primitives will still win in the interpreter; after all they use the same machinery as at: and at:put:.  But in a JIT they'll likely loose.

Since the Cog JIT isn't available yet I'm not really being helpful.  I should think before I blurt.  Apologies.

Eliot


Levente

FWIW, I'd take a JIT anyday over this bug getting fixed :)
On the other hand, it's a one-word fix which merely improves performace while we're waiting (unless someone decides to remove the primtive calls from the image altogether, then it merely becomes obsolete), so I hardly see the harm in applying it.

Cheers,
Henry

PS. Speaking of unneccessary overhead, what about that collection class == ByteString part for _every_ stream nextPut: , when ByteString at:put: will handle the case correctly anyways if collection is in fact a ByteString? Not to mention how it performs for ByteString currencly, seeing as the primitive now _does_ fail every time.