Easy & huge GC optimizations

[Rainer Schuetze via Digitalmars-d]

On 22.05.2014 18:42, Etienne wrote:
> I was thinking of how the GC could be optimized further and came across
> some sweet flags that are barely used throughout Phobos in favor of a
> witch-hunting against the GC:
>
> GC.BlkAttr.NO_SCAN
> GC.BlkAttr.NO_INTERIOR
>
> When using the NO_SCAN attribute with GC.setAttr(p, NO_SCAN), you're
> basically doing removeRange on a GC allocation. It's never going to scan
> the memory in it, but the memory will stay alive if pointers are found
> pointing to anything inside it. This is very useful for strings! But I
> can't even find an example of a string with this flag. I'm totally baffled.
>
> When using NO_INTERIOR attribute, you're telling the GC that nothing can
> point to the inside of the allocation if it's bigger than 4096 bytes,
> and to completely ignore scanning its contents in such case.
>
> With these 2 attributes, one could write a simple recursive function in
> phobos that adjusts the flags on an object's allocations based on the
> type info.
>
> Tuple!(int, int, int, string)[] bigTupleArray;
> bigTupleArray.optimizeGC(); // sets NO_SCAN on ints, and on the pointee
> of string
>
> Thoughts?

rt.lifetime makes heavy use of "NO_SCAN" for all array handling 
(including strings). It uses the compiler generated flag inside the 
TypeInfo. Classes also deal with this flag (see _d_newclass).

I'm not sure where allocating a struct ends up, so maybe there is some 
potential there.

"NO_INTERIOR" is currently only used for the hash array used by 
associative arrays. It is a bit dangerous to use as any pointer,slice or 
register still operating on the array is ignored, so collecting it might 
corrupt your memory.