Re: An accurate, pauseless & deterministic GC for C++
"Joe Seigh" <jseigh_01@xemaps.com> wrote in message
news:XZU0j.1586$Mr.1194@trnddc04...
Chris Thomasson wrote:
"Joe Seigh" <jseigh_01@xemaps.com> wrote in message
news:ZmZ%i.2695$Jy1.83@trndny02...
[...]
I figured it was some form of RCU w/ memory barriers in the quiescent
states.
That by itself doesn't let you elide memory barriers. Whether you can
elide
a membar depends on the specific situation. And the proofs can be a bit
complicated depending on the situation (at least for SMR+RCU).
Yeah. For instance, the polling logic needs to do some things (e.g., add
extra epoch) to allow you to have release semantics. If your doing your fast
SMR, the polling logic needs to scan the hazard pointers in a way that gives
them acquire semantics. Been there done that. ;^)
I'm guessing from the docs, the refcounting is a bit like atomic_ptr with
the distinguishing of local non-shared vs. global heap shared reference
counts.
Thats what I am leaning to. I have not examined the refcount logic in the
source-code I downloaded enough to make a definitive answer. The first thing
I will check for is wether or not it uses interlocking instructions. If not,
then it has to be a form of distributed counting. If so, I will compare and
contrast against the method I created for vZOOM ref-counting:
http://groups.google.com/group/comp.programming.threads/browse_frm/thread/1216861ac568b2be
With refcounting, it was always safe to use a raw reference as long as you
had
(owned) one refcount. There seems to be a wrapper class that is the
logical
equivalent.
Indeed. Its also safe to take and use a non-NULL raw "read" reference w/
dd-load from a data-structure being protected by a proxy collector:
pc_read& read = proxy_read(&pc_anchor);
// inside a read-only epoch
//_______________________________________
node* n = load_depends(&shared_lifo);
while (n) {
node* const nx = load_depends(&n->nx);
read_only_process(n);
n = nx;
}
proxy_release(read);
The burden of proof is on him. He needs to show how his stuff is
different
from prior art. That means explicitly listing the prior art and
contrasting
it with his stuff.
Yup. He also has to do it with fewer claims than before; the USPO issued new
rules that effect the number of claims you can have. Here the are:
___________________________________________________________________
(A) - No more than twenty-five total claims.
(B) - No more than five independent claims.
(C) - No more than two continuations per-patent.
(D) - Identify all co-pending patent applications.
___________________________________________________________________
Luckily, my vZOOM stuff passes all the above. Although, I was wondering how
RCU is going to explain the numerous continuations? I think that is going to
be a problem...