Re: Guile v2 and GC

[David Kastrup]

Han-Wen Nienhuys <address@hidden> writes:

> On Tue, Dec 6, 2016 at 6:54 PM, David Kastrup <address@hidden> wrote:
>> Han-Wen Nienhuys <address@hidden> writes:
>>
>>> Hello David,
>>>
>>> I saw that you made (some) changes for dealing with the Guile v2
>>> garbage collector. what is the overall status of this?
>>
>> The known problems are patched up.
>
> so we have unknown problems? Or is it just working?

Not enough data.

>>> Have you considered defining an operator new() for the smob types? It
>>> would call scm_gc_malloc. Wouldnt that take care of the whole messy
>>> derived_mark business?
>>
>> The whole messy derived_mark business is there for a reason.  LilyPond's
>> data structures are large and LilyPond fills almost all the available
>> address space on 32bit systems.
>
> It is there because GUILE 1.6 did not support anything else.
>
>> In spite of our targeted mark passes, we still had the occasional
>> "Object should be dead" message, just from conservatively marking the
>> stack (and explicitly garbage collecting at a known low tide point).
>
> Conservative garbage collection when the heap size is close to address
> space is always a losing proposition, so we should not aim for it;
> People that want to use a lot of RAM should buy a 64-bit system.
>
>> And you are not going to win this by scm_gc_alloc-ing the memory of the
>> smob types anyway: you need custom alloc overrides for all of the STL
>> class arrays containing SCM values of some kind as well.
>
> AFAICT There are actually very few of them.
>
>> And of course you know that _lot_ of SCM values are not even stored
>> explicitly but as structure pointers that still need ->self () calls to
>> actually get a recognizable SCM value out.
>
> well, they could be stored as SCM with conversion function.

Feel free to go for it.  I consider it a heap of new trouble, but then I
am not the one who'll have to deal with the fallout any more.

And you have to be aware that there are very few people who will be able
to deal with the kind of trouble low-level memory allocation can cause.
Basically you'll be out on your own.

And Guile-2.x is good for a number of surprises of its own.

-- 
David Kastrup