GC Precision

bearophile bearophileHUGS at lycos.com
Thu Oct 29 03:47:31 PDT 2009


dsimcha:

> A moving GC, one that doesn't stop the world on collection,
> and one that's fully precise including stack would be nice, but they're several
> orders of magnitude less important and would also have more ripple effects.

I agree that here doing something simple now is better than doing nothing or doing something very refined in an unknown future.

And in future things may be improved. In D objects are always managed by reference, and I think that most programs don't alter or cast such references to something else (objects allocated on memory specified by the programmer, and scoped objects allocated on the stack may be excluded from this). So I think that it can be safe to move objects, to compact the heap.

So you may have 5 memory zones:
- C heap. (The type system of the D compiler may see the C-heap pointers and D-heap pointers as two different types, as I have proposed in the past. So you need casts if you want to mix them, and the compiler can use the D moving heap in a safer way).
- Pinned D heap for everything can't be moved, like structs managed by pointers (eventually a D compiler can infer at compile time that some structs too may be moved around, because their pointer is used only in clean ways (you don't need to introduce struct references for this)). I think SafeD modules will not use this heap a lot (but they can use unsafe modules that may use pinned objects. Is D safety transitive? I think it is not, so from a SafeD module you can call and use an unsafe module);
- Old object generation managed with compaction (I think there's no need for the permanent objects zone in D);
- Two "from" and "to" zones for the young generation, that don't use a true compaction strategy, young objects bounce between them;
- New generation Eden where new object allocations happen managed as a memory arena.

All this has the disadvantage of requiring a more complex GC, and probably requiring 2-4 times more RAM at runtime. It hopefully has the advantage of allowing new programmers, that have learnt Java at university, to program in almost like in Java. (I have found a not-synthetic Java benchmark program that converted to D is something like 18 times slower on LDC. I'll put the code in my site in the following days).

Bye,
bearophile



More information about the Digitalmars-d mailing list