[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: dependency info
Craig has done a lot of work in this area.
I wonder what he would say?
I'm not sure how to do a hybrid scheme that would work.
At 1:51 PM 9/26/95, Jecel Mattos de Assumpcao Jr. wrote:
>For some strange reason I've never liked the dependency information
>that Self has to keep on each compiled method. I would guess that
>the overhead of keeping all those lists up to date is not very
>high and the "spy" shows that the memory used is not too exagerated,
>so I can't give a good reason for this dislike.
>If we didn't have this information at all, then it would still be
>possible to make Self work by flushing the whole compiled code cache
>after any programming primitive, right? The result would not be
>very good - pauses of over 15 seconds could be expected on a
>Sparc 2 ( see Urs' PhD thesis on page 111, item 13 in the table ).
>How about a hybrid solution ( it seems we always go for these ):
>if we could devide the objects into a few large groups ( less than
>32, for example ) we could use a single word in each compiled
>method to tag it as depending on objects of one or more groups.
>When an object is changed, all methods depending on its group
>would be flushed. Object groups are *much* less precise than
>individual slots, but the difference in results might not be as
>great as it would seem at first.
>Any thoughts on this?