On 4/7/2012 8:26 PM, Sean Kelly wrote:
On Apr 7, 2012, at 9:45 AM, Rainer Schuetze<r.sagita...@gmx.de> wrote:
On 4/6/2012 6:20 PM, deadalnix wrote:
Le 06/04/2012 18:07, Andrei Alexandrescu a écrit :
A few more samples of people's perception of the two languages:
http://news.ycombinator.com/item?id=3805302
Andrei
I did some measurement on that point for D lately :
http://www.deadalnix.me/2012/03/05/impact-of-64bits-vs-32bits-when-using-non-precise-gc/
I studied the GC a bit more and noticed a possible issue:
- memory allocations are aligned up to a power of 2<= page size
- the memory area beyond the actually requested size is left untouched when
allocating
- when the memory is collected, it is also untouched
- the marking of references during collection does not know the requested size,
so it scans the full memory block
Result: When a collected memory block is reused by a smaller allocation, there
might still be false pointers in the unused area.
When I clear this data, my first impression is that it has improved the
situation, but not enough. I'll have to create some non-interactive test to
verify.
The unused area is typically zeroed out on allocation. Check GC.mallocNoSync..
That's where I added it ;-) But in fact it is in malloc/calloc and a few
more places, though I haven't checked whether all the other calls of
mallocNoSync handle the situation correctly. Almost all calls are to
malloc anyway, so my patch doesn't change anything.