On 11/01/2015 04:35 PM, Martin Nowak via dmd-internals wrote:
On 10/27/2015 10:45 PM, Andrei Alexandrescu wrote:
Oddly enough I want to dissuade you from doing RC - per my post "Safe
reference counting cannot be implemented as a library" it's impossible
to make library-level RC safe, which decreases its interest
dramatically. We plan to do language-level reference counting, which, if
we succeed, would obviate any RC library for classes.
However, working on uniqueness is great! We'll either see how far we can
get within the current language, or figure the minimal language
additions needed to make it work.
Well unique classes also cannot be implemented safely.
There's a less pleasant possibility: have the unique object NOT offer a
pointer/reference to the held resource, but only offer a method such as
.relinquish() that destructively reads it. So:
Unique!int x = makeUnique!int(42);
int* p = x.relinquish;
assert(x.empty);
assert(*p == 42);
So... relinquish destructively gets the raw pointer from the unique
object and puts the responsibility of managing it straight on the
shoulders of the user.
Now I think @safe reference counting is great but we don't know if or
when any of those ideas pan out. So I'd rather implement a small dub
package now than waiting for the ultimate solution. If that gets laters
obsoleted by a safer and better dmd/phobos implementation, all the better.
I think that's a great route to follow. A carefully written library
implementation will inform any study
Andrei
_______________________________________________
dmd-internals mailing list
[email protected]
http://lists.puremagic.com/mailman/listinfo/dmd-internals
