|Pathologically Eclectic Rubbish Lister|
Re^3: Quickly detecting variable writesby sfink (Deacon)
|on Jan 06, 2007 at 17:11 UTC||Need Help??|
In your example, unless += is being overloaded, $x will always be "written to" if @kabluther contains anything--even if the value added is 0. In which case, knowing whether $x 'was written to', is equivalent to knowing 'if @kabluther was non-empty'.Yes. And in that particular case, that is exactly what I want.
If += is overloaded, then you're already paying the penalty of overloading and adding a flag internally to the overload method and a second method to query the flag will add negligable extra overhead.Oops, sorry, bad example. Yes, you are correct. But that depends on the type of value being used. Values could be overloaded blessed array refs, as in the example above. Those are used for 2d, 3d, and 4d vectors. But other common types are booleans, integers, floats, doubles, and strings, as well as a bunch of more heavyweight types (images, vectors, sets, ...) None of those other lightweight types are overloaded or otherwise special, and they are probably the more important ones to suppress the false updates for. Especially booleans, which tend to be used as triggers.
To answer your question: my system is comprised of a tree of nodes, where each node contains a collection of attributes. Attributes can be either inputs or outputs. All processing is done by connecting the inputs and outputs together. An external trigger (clock ticks, camera frames, keyboard events, ...) causes a cascade of updates throughout the graph.
Attributes that are bound together share a value, so the value is communicated from one node to the next. But in addition, there is an explicit call saying "I updated this attribute" that invokes the callbacks on all input attributes that are bound to that output. So you can think of it as a graph where the edges propagate both values and control (in the form of update notifications).
One type of node is a Perl script node. It can have whatever attributes it likes, and can bind any of them to Perl variables using the ": Attr(...)" syntax in the example in another post, or by a lower-level notation to retrieve the value which doesn't try to do any magical auto-updating stuff.
So the decision of what value to assign to some attribute is different than whether to mark it as updated. Usually, the two go hand in hand. But there are many cases where triggering unnecessary update notifications will result in a huge amount of wasted work. So, for example, if a Perl script node has six boolean output attributes, you really don't want to mark every one of them as updated every time anything changes. And you can't tell from the value; it is perfectly reasonable to repeatedly set an output attribute to 'false' and mark it as updated over and over again.
You always have the option of dropping back to a lower level, and marking every update explicitly. I am trying to come up with an easier-to-use interface that removes the need for the marking in most cases, preferably without adding too much overhead. The basic idea is that 95% of the time, if you write to a variable then you want to mark it updated. Most of those writes will end up changing the variable's value, but the cost of missing an update (the rest of the tree does not get triggered) is much worse than the cost of doing an unneeded updated (usually just redundant processing), so value comparisons are conservative in the wrong direction.
Looking at the length of this reply, I think you may have been correct to regret persisting!