Garabage collector have to resolve managed pointers to their objects when a specified managed pointer refers to the heap. In rotor's implementation this is expensive task. I shell try to investigate whether it's possible in some (many) cases to resolve managed pointer in a way faster then scanning heap.
I believe in many cases source object of managed pointer are refered in the same or some ancestor's stack frame as managed pointer.
...
I've modified JIT "promotion" callback in garbage collector in order it would try to resolve managed pointers using objects that occur in the stack frame and only if this algorithm would fail, would resolve managed pointers with "find in the heap" algorithm.
The goal of this implementation is:
In short algorithm works as following:
The algorithm has really short path for object, just a several simple conditions (in average is only couple) and a bit more complicated logic for managed pointer, but it inteded to resolve it!
There were several aspects in testing:
It was pretty easy to create more fast sample: if we have object that contains field of structure type and call any method on this field value we get "this" as managed pointer; or even more simple example when we pass field address as ref argument to the function.
The advantage in such examples varies from -0 to 10% this mainly depends on how fast original "find in the heap" algorithm finds the object. -0 hear means that in extremely successful case "find in the heap" algorithm finds the result just from the first try.
On the other hand when stack does not contains managed pointer (to heap) at all results are just equal.
And finally there is a class of cases when the resolving algorithm fails and "find in the heap" algorithm is invoked. In this case result could be about 2% - 3% slower. I believe however that patterns that lead to such a slow program are quite unnatural.
And finnaly I was trying to find samples in bcl that prove my hypothesis. To my disappointment I've found that bcl carefully avoids using managed pointers (directly and inderectly) at all! This is probably why there is conception in bcl that value types are init only. As result accounting probability of appearing of managed pointer in code the final result is very very similar to original implementation. It's looks like I've made code bigger without anysignificant result!
I admit that the problem I've discussed(?) here probably is not so sharp as I allowed in the first place. All related code I've separated in branch "Resolving-interiors" in vm and fjit subprojects. I still belive that this realy was beautiful algorithm.
Share your thoughts with as about this article.