Am 2016-01-11 um 19:45 schrieb Attila Szegedi:
Specifically in relation to the remark that JSType.isNumber check is getting
long… I think we could be even more conservative if we wished to, namely only
recognizing Integer and Double as wrapped numbers. Within Nashorn, byte, short,
float, and long don't occur naturally, only int and double occur. Their
wrappers (Byte, Short, etc.) therefore also can’t appear. The only way to
introduce them is through POJO methods with such return types (e.g.
Byte.valueOf()). I’m not sure we’re under an obligation to consider those
equivalent to JS numbers. We could declare that Byte, Short, Float, and Long
are not numbers just host objects, albeit ones that JSType.toNumber understands.
I guess we could do that, but we would (IMO) needlessly change the rules
we set up with Nashorn in JDK8. I think that making Java number
primitives equivalent to JS numbers is actually a nice feature.
Unfortunately it doesn't really work for longs because JS happens to use
double as its only number type, but I don't think that means we should
throw out the idea as a whole.
Also, an important point here is that we can't really safely distinguish
between primitive number types and object wrappers in Nashorn. A
java.lang.Short or java.lang.Float may just be a primitive that happens
to be handled in an object context.
Actually, I can take this train of thought in an even more drastic direction.
Consider this: we would be at liberty to internally box numbers any way we want
and not rely on java.lang.Double and friends at all! It’s just convenient to
box JS primitive number values (and booleans, BTW) using readily available
java.lang.* boxes, but we could have a NashornNumber and NashornBoolean classes
for internal boxing of JS primitive numbers and booleans, when they need to be
boxed. Then we could even allow ourselves to not recognize any Java primitive
wrapper objects as special cases of JS primitive values at all. The thought is
actually not so far fetched in the sense that we already have ConsString too,
an alternative internal string representation that is not a java.lang.String.
We’d just have to take care that such NashornNumber boxes are handled in all
places where ConsString is handled and converted into a Double when passing
from Nashorn to outside world.
That would provide a 100% consistent way of handling wrapper classes: they’re
all just POJOs at that point, no special treatment whatsoever.
I think what I wrote above also applies here - that seamlessness is
actually nice and that we don't really control whether a numeric value
from Java is handled as primitive or as wrapper object in Nashorn - that
may vary on how JavaScript code was used/optimized/compiled before, what
the code does with it, etc.
In Rhino, we had explicit wrapper classes for things to be treated as
objects (e.g. a java.lang.Double would be wrapped in a JavaObject
wrapper or whatever it was called). Since Nashorn does not use object
wrappers, I think it would be hard to distinguish between primitive
numbers and object wrappers in this way. It would als feel kind of
arbitrary to me.
Hannes
Attila.
On Jan 11, 2016, at 1:48 PM, Hannes Wallnoefer <hannes.wallnoe...@oracle.com>
wrote:
You are right of course, there needs to be consistency between typeof operator
and treatment as JS numbers.
This is in fact an unpleasant problem to solve. I've struggled trying to fix
this without breaking any existing code, but I've come to the conclusion that
it is not possible. Since we can't treat all Java longs/Longs as JS numbers,
we'd have to differentiate depending on whether the value can be represented as
double without losing precision.
In a way we already do this with optimistic types, but I consider it more a bug
than a feature. It's weird (and error prone) if the return value for a Java
method returning long is reported as number or object depending on the actual
value.
So I think the right thing to do is draw a clear line between which Java
primitive/wrapper types represent JS numbers and which don't. I've uploaded a
new webrev that implements this:
http://cr.openjdk.java.net/~hannesw/8143896/webrev.01/
Note that the only types to be treated as JS numbers are the direct wrapper
classes for Java primitives that can be fully represented as doubles. This
means also things like AtomicInteger and DoubleAdder will be reported and
treated as objects. I think that's the correct thing to do as they are not
primitive numbers in the first place. They are still converted to numbers when
used in such a context in JS. So I think the only place where this change is a
actually painful/surprising is longs.
Unfortunately the check for number type in JSType.isNumber gets a bit long as
we have to individually check for all primitive wrapper classes. I've done
extensive benchmarking and I don't think it has an impact on performance. In
any way, I wouldn't know how to handle this differently.
Let me know what you think.
Hannes
Am 2016-01-04 um 05:00 schrieb Sundararajan Athijegannathan:
I think I already commented on this webrev -- that we need to cover tests for
BigInteger, BigDecimal.
Also, I'm not sure linking Double and Int by nashorn primitive linkers is the right
solution. AtomicInteger, DoubleAdder etc. are all Number subtypes. We return
"number" when typeof is used on any Number subtype.
Now, that means JS code will see these as 'number' type objects -- yet
Number.prototype methods won't work on those!! I know this is hard problem --
we also have another (somewhat related) BigDecimal, BigInteger toString /
String conversion issue. We need to discuss this.
-Sundar
On 1/2/2016 8:29 PM, Attila Szegedi wrote:
+1
On Dec 18, 2015, at 3:54 PM, Hannes Wallnoefer <hannes.wallnoe...@oracle.com>
wrote:
Please review JDK-8143896: java.lang.Long is implicitly converted to double
http://cr.openjdk.java.net/~hannesw/8143896/webrev/
Thanks,
Hannes
