RFR: updated draft API for JEP 269 Convenience Collection Factories
Stuart Marks
stuart.marks at oracle.com
Fri Nov 6 18:28:47 UTC 2015
On 11/6/15 5:12 AM, Michael Hixson wrote:
> + static <E> List<E> of(E... es) {
> + for (E e : es) {
> + Objects.requireNonNull(e);
> + }
> + // NOTE: this can allow a null element to slip through
> + return Collections.unmodifiableList(Arrays.asList(es));
> + }
>
> Even as a skeletal implementation, this one has to be changed to be
> truly immutable, right? It currently returns a view of the (mutable)
> argument array rather than new storage. Sorry for not providing a
> proper test:
Good catch! Funnily I had noticed the TOCTOU case that allowed null elements in
the array to slip through, but not that the array itself was still modifiable
from the outside. Anyway, I'll fix this. No worries about the test.
> Has anyone been able to quantify the advantage of having these
> overloads as opposed to having the varargs versions only? Is it a
> matter of performance?
>
> I ask because the overloads seem like warts on the APIs (which is a
> shame -- List and Set are such important APIs). I'm imagining a
> future where:
>
> 1. We add these overloads for performance gains now.
> 2. But they're all skeletal implementations that aren't that perfomant
> anyway. Efficient versions don't make it into Java SE 9. People that
> care a lot about performance avoid using these ones.
> 3. A few years later, varargs performance or some other language / VM
> / compiler-level change renders the overloads obsolete.
Yeah, the overloads seem like warts on the API, though probably necessary ones.
At present, and for the forseeable future, varargs calls allocate an array on
the heap, whereas fixed-args calls do not. I don't know how to quantify the
difference though. Certainly the cost of allocation and initialization is borne
in-line. Then there is the cost of collection. Collecting short-lived objects is
cheap (but not free). There is also the possibility of escape analysis
eliminating the allocation. This seems unlikely to me; certainly not something
to be relied upon.
The most likely possible future optimization is "frozen arrays," part of the
"Arrays 2.0" stuff that John Rose has talked about. This is basically about
immutable arrays. Here, the possibility is to eliminate the defensive copy, if
the array created to hold the varargs arguments is made immutable. (This will
require some adjustment on the callee side, as yet unspecified.) There's still
an array, though. And a defensive copy would still have to be made if the caller
passes an actual array, as opposed to a varargs list.
While I can't quantify it, I do think there's an expense to creating the varargs
array, and there is only a possibility to reduce (but not eliminate) its cost in
future JDK releases. This cost is entirely avoided by fixed-args overloads.
(There is the cost of cluttering up the API, though.)
Turning to the skeletal vs. optimized implementation, my plan is certainly to
ensure that the optimized implementations get into JDK 9. Of course, plans can
change. If the APIs get in without the optimized implementations, I think the
big attractor will still be the convenience of using these static factory
methods as opposed to conventional code. They're no slower than conventional
code, and the space consumed is the same. So I think they'll be popular even if
the space efficiency benefits aren't there initially.
When the optimized implementations do get in, callers will benefit, even without
recompilation. Thus there is some present value added based on potential future
benefits.
There is always the set of possible future events that cause something not to
work out, but I think pursuing the approach I've outlined has a good chance of
benefiting the platform in the long term.
s'marks
More information about the core-libs-dev
mailing list