On 04/25/2014 06:13 PM, Peter Levart wrote:
Hi Paul,
On 04/25/2014 02:12 PM, Paul Sandoz wrote:
On Apr 23, 2014, at 11:06 AM, Peter Levart<[email protected]> wrote:
Hi,
I propose a patch for:
https://bugs.openjdk.java.net/browse/JDK-8040892
+1, nice use of putIfAbsent. Just double checking, did you run the stream unit
tests?
Yes, all 60 /java/util/stream jtreg tests pass.
I think key step was to separate map merge and accumulate functions for the
toMap* methods not accepting a value merge function. Now that is the case one
could still use merge, but that would be less efficient than putIfAbsent since
that does not require a lambda capturing the key.
That fix caused me to think that we should update the value merge function
accepting toMap* methods for the case when the merge function returns null, as
removing an entry would be most undesirable.
Yes, removing doesn't make sense. It's undeterministic. For example,
if the merge function always returns null, then duplicate keys result
in entry being absent, triplicate keys result in entry (the last one
merged) being present, quadruple keys result in entry being absent
again, etc ...
Hi Paul,
To make any sense of null return from the mergeFunction, which could
mean that an entry with that key is absent from resulting map, such
Collector would have to keep some more state - the collecting map (which
is also returned at the end) and the set of removed keys. For example:
public static <T, K, U, M extends Map<K, U>>
Collector<T, ?, M> toMap(Function<? super T, ? extends K> keyMapper,
Function<? super T, ? extends U> valueMapper,
BinaryOperator<U> mergeFunction,
Supplier<M> mapSupplier) {
class State {
final M map = mapSupplier.get();
final Set<K> removedKeys = new HashSet<>();
M map() { return map; }
}
BiConsumer<State, T> accumulator = (state, element) -> {
K k = keyMapper.apply(element);
if (!state.removedKeys.contains(k)) {
U u = state.map.merge(k,
valueMapper.apply(element), mergeFunction);
if (u == null) state.removedKeys.add(k);
}
};
BinaryOperator<State> merger = (state1, state2) -> {
state1.map.keySet().removeAll(state2.removedKeys);
state2.map.keySet().removeAll(state1.removedKeys);
state1.removedKeys.addAll(state2.removedKeys);
for (Map.Entry<K,U> e : state2.map.entrySet()) {
U u = state1.map.merge(e.getKey(), e.getValue(),
mergeFunction);
if (u == null) state1.removedKeys.add(e.getKey());
}
return state1;
};
return new CollectorImpl<>(State::new, accumulator, merger,
State::map, CH_ID);
}
...but the concurrent variant would be tricky to implement.
Regards, Peter
A simple fix would be to throw an NPE if the merge function returns null, if it
is acceptable to induce a side-effect of a removal, which seems reasonable
given a partial result would any be returned if a null check is performed on
the result of the merge function, otherwise that merge function requires
wrapping e.g. mergeFunction.andThen(Objects::requireNonNull).
I think that allowing null return from mergeFunction to remove the
entry from the (maybe intermediate) map is just confusing. Throwing
NPE is the most sensible thing we can do here.
Regards, Peter
Paul.
Here's the webrev:
http://cr.openjdk.java.net/~plevart/jdk9-dev/Collectors.duplicateKey/webrev.02/
There has already been a discussion on the lambda-dev about this bug:
http://mail.openjdk.java.net/pipermail/lambda-dev/2014-April/012005.html
Initially a fix has been proposed which used a private RuntimeException subclass thrown
by the merger function in order to transport the values that attempted to be merged out
of Map.merge() method to an outer context where they could be combined with the key to
produce a friendlier "duplicate key" message. It turns out that there is an
alternative approach that doesn't require exception acrobatics - using Map.putIfAbsent()
instead of Map.merge(). I checked both methods in Map defaults and in HashMap
implementations for possible semantic differences. Same behaviour can be achieved by
requiring that the value passed to Map.putIfAbsent(key, value) is non-null and throwing
NPE if it isn't.
Regards, Peter
