Interview with InformIT part 2/3

[Walter Bright]

Leandro Lucarella wrote:
> Walter Bright, el 18 de agosto a las 10:08 me escribiste:
>> bearophile wrote:
>>> Currently in the D2 GC there is no notion of pinned/unpinned class instances,
>>> but eventually an attribute as @pinned may be added to D3, plus its related
>>> semantics. It adds complexity to the language and it needs to interact with
>>> the GC, so it will get useful as the D GC becomes more modern (with different
>>> zones for newly allocated objects, etc).
>> There is no need for a pin attribute, the gc can determine if a
>> class needs pinning or not.
> 
> As long as the precise heap scanning patch is applied (and much better
> if we can manage to scan the static data precisely too). Otherwise you
> simply just can't move stuff around because you don't know what is
> a pointer and what is not (thus you can't update pointer that point to
> moved stuff).

Hence the objects with ambiguous references to them get automatically pinned by 
the gc. I've implemented such a gc before, it works fine. It's called a "mostly 
copying collector".

Besides, any scheme where you have to manually mark pinnable objects is doomed 
to have disastrously subtle and hard-to-find bugs.