Easy & huge GC optimizations

[Rainer Schuetze via Digitalmars-d]

On 23.05.2014 20:52, Etienne wrote:
> On 2014-05-23 2:14 PM, Rainer Schuetze wrote:
>> BTW: How do you detect the sample pointers are alive? Or do you mean
>> just the roots?
>
> You store a void** and the original value, dereference it to see if it's
> the same value as the original. Loop through 20 of those if you have
> 500, and you update them during marking by taking 1 in 20.
>

Looking at the address where the reference is stored gives you a hint 
about modifications, but does not say anything about whether the object 
that contains this reference is still alive.

> There's mathematical proof that you if you don't have 5 of those dead,
> you're great to go and don't need to collect.
>

You still have to scan/mark the whole memory. Collecting unmarked memory 
is the easy part that doesn't cost too much time.

> Being flexible enough, you can skip 9 in 10 collections altogether imo
>
> see this:
> https://github.com/D-Programming-Language/druntime/pull/803
>
> also mathematical proof is at page 154-183 of:
> http://books.google.ca/books?id=b-XFrpBQ7d0C&lpg=PA119&pg=PA154#v=onepage&q&f=false
>

AFAICT your test case does not measure garbage collection, but manual 
memory management using the GC as the memory manager. delete/free are 
not meant to be called by user code as these are unsafe operations.