Hi Stuart,
I think that what matters are the methods that execute in tight loops.
Some thousand array clones that would be performed while interpreting
would be barely noticeable. I agree that JMH benchmarks presented here
are not a representative for real-world tight loops. Especially because
successful EA is very sensitive to various arrangements. But this can
also be viewed as an opportunity to tune the JIT exactly for the
situations that arise with such vararg collection factories. We would
get a two-fold improvement: smaller API footprint and better
optimization of non-escaping vararg arrays that might benefit other code
too. OTOH I understand that this would be a larger undertaking involving
more resources.
Regards, Peter
On 11/09/2015 09:44 PM, Stuart Marks wrote:
Hi Michael, Peter,
I admit to not having followed all the benchmarking discussion
(travel, Devoxx) but I did want to respond to this:
Peter wrote:
So it looks like that there's no need to burden the public API with
explicit argument overloads.
These are JMH benchmarks, which take care to warm up the JIT, so that
everything is nicely optimized, inlined, escape-analyzed, etc., right?
If so, I don't see this at all being representative of how these APIs
will be used. They won't be called in tight loops that give the JVM a
chance to optimize them. Instead, they'll be used for initializing
fields, including static fields:
class MyClass {
static final Set<String> set = Set.of("a", "b", "c");
...
}
This call to Set.of() will (probably) be called exactly once when
MyClass is loaded. It'll probably be run in interpretive mode. If the
API becomes popular, there will be lots of calls like this in static
initializers and instance initializers. Most of these will be called
once or relatively few times, but there may be a lot of such calls. I
think most of these would be run by the interpreter.
How do we assess the impact of fixed-args vs varargs overloads for
cases like this?
s'marks
