Question to GC experts: NO_SCAN for a part of the block

[Steven Schveighoffer via Digitalmars-d]

On 6/6/17 8:34 AM, Stanislav Blinov wrote:
> On Tuesday, 6 June 2017 at 12:07:28 UTC, Steven Schveighoffer wrote:
>> On 6/6/17 6:36 AM, Stanislav Blinov wrote:
>>> On Tuesday, 6 June 2017 at 09:57:31 UTC, Steven Schveighoffer wrote:
>>>
>>>> No, 2 allocations instead of N.
>>>
>>> Oh, I misunderstood you. You mean two blocks total, for scanned and
>>> non-scanned data. This could indeed work for cases when more than two
>>> arrays are needed. Still, it's one extra allocation.
>>>
>>
>> True. But O(2) is better than O(N) :)
>>
>
> Heh, this isn't malloc. We could end up in a situation when even one
> allocation is slower than N, depending on when they happen :)
>
> Actually, separate blocks might be a decent choice: there's another
> problem with sharing blocks for different data in that finalizers can't
> be run. So such a function would either not support types with
> finalizers at all, or would have to allocate separate blocks for such
> types anyway. Don't know, need to actually implement and measure if it's
> even worth the trouble.

Looking at your code, you have allocated separate blocks for the objects 
and the bytes, but your objects still are just pointers (e.g. C.sizeof 
== (void *).sizeof). You are allocating a new block for each class 
instance anyway in the loop inside main.

I don't think there is any support in the GC to run the finalizers of an 
array of class instances (read: an array of class instance storage that 
you point at), so this whole exercise may be moot if you are using 
classes :)

-Steve