On Wed, 2 Jun 2021 17:12:21 GMT, Stuart Marks <smarks@openjdk.org> wrote:

A quick search reveals that Guava has a public subclass of SimpleImmutableEntry: https://guava.dev/releases/30.1.1-jre/api/docs/com/google/common/cache/Remov...

There are possibly others. It doesn't seem worth the incompatibility to me, as it would break stuff in order to have only a "cleaner" meaning for "immutable." Also, there are other classes in the JDK that claim they are immutable but which can be subclassed, e.g., BigInteger. I don't see a good way of fixing this.

Thanks, i've done some digging too, I foolishly hoped that nobody would subclass a class with `Immutable` in its name, oh boy i was wrong :) So documenting the existing behavior seems the only possible way. ------------- PR: https://git.openjdk.java.net/jdk/pull/4295

----- Mail original -----

De: "Peter Levart" <peter.levart@gmail.com> À: "Rémi Forax" <forax@openjdk.java.net>, "core-libs-dev" <core-libs-dev@openjdk.java.net> Envoyé: Jeudi 3 Juin 2021 20:49:05 Objet: Re: RFR: 8199318: add idempotent copy operation for Map.Entry

On 02/06/2021 19:24, Rémi Forax wrote:

...

I foolishly hoped that nobody would subclass a class with `Immutable` in its name, oh boy i was wrong:)

There's nothing wrong in subclassing an immutable class. As long as the subclass keeps being immutable. Was it subclassed and made mutable by that?

It has to be immutable all the way up, you have the same issue if the subclass is not final. If you filter out guava an google cache, github get you 12 pages of result and only one stupid code https://github.com/klayders/interlude/blob/95fd59a911d2baa8ac36ae6b877955aa4... A lot of code uses SimpleImmutableEntry as a pair, my hope is that the introduction of records will clean that practice. That said, there is also several broken codes, mostly two issues, either a.equals(n) is not equivalent to b.equals(a) or a.equals(b) is not equivalent to a.compareTo(b) == 0. I kind of regret that the compiler does not provide automatically an implementation of compareTo if the record implements Comparable. People sucks at writing compareTo and the resulting bugs are hard to find/reproduce.

Peter

Rémi

De: "John Rose" <john.r.rose@oracle.com> À: "Remi Forax" <forax@univ-mlv.fr> Cc: "Peter Levart" <peter.levart@gmail.com>, "Rémi Forax" <forax@openjdk.java.net>, "core-libs-dev" <core-libs-dev@openjdk.java.net> Envoyé: Jeudi 3 Juin 2021 22:51:28 Objet: Re: RFR: 8199318: add idempotent copy operation for Map.Entry

On Jun 3, 2021, at 12:46 PM, Remi Forax < [ mailto:forax@univ-mlv.fr | forax@univ-mlv.fr ] > wrote:

...

I kind of regret that the compiler does not provide automatically an implementation of compareTo if the record implements Comparable. People sucks at writing compareTo and the resulting bugs are hard to find/reproduce.

That’s a slippery slope. IIRC we consciously stopped before that step.

That said, there are other ways to fix this. We should have utilities (maybe in the JDK but not the JLS) which build such methods and make it easy for users to grab onto them. Maybe something like this:

interface ComparableRecord<T extends Record & ComparableRecord<T>> extends Comparable<T> { … }

record Foo(int x, String y) implements ComparableRecord<Foo> { … }

[ http://cr.openjdk.java.net/~jrose/draft/ComparableRecord.java | http://cr.openjdk.java.net/~jrose/draft/ComparableRecord.java ]

The main issue with this kind of code is that the JIT does not see through the ClassValue. Tweaking a little bit your code, I get (It's a PITA that we have to use a raw type to workaround circularly defined parameter type) import java.util.ArrayList; import java.util.Comparator; import java.util.List; import java.util.stream.Stream; @SuppressWarnings({"rawtypes","unchecked"}) interface ComparableRecord<T extends Record & ComparableRecord<T>> extends Comparable<T> { @Override default int compareTo(T that) { if (this.getClass() != that.getClass()) { throw new IllegalArgumentException("not same class"); } return COMPARE_TO_METHODS.get(this.getClass()).compare(this, that); } static final ClassValue<Comparator<Object>> COMPARE_TO_METHODS = new ClassValue<>() { @Override protected Comparator<Object> computeValue(Class<?> type) { return Stream.of(type.getRecordComponents()) .map(component -> { var accessor = component.getAccessor(); return Comparator.<Object, Comparable>comparing(r -> { try { return (Comparable<?>) accessor.invoke(r); } catch (ReflectiveOperationException ex) { throw new IllegalArgumentException(ex); } }); }) .reduce((r1, r2) -> 0, Comparator::thenComparing, (_1, _2) -> { throw null; }); } }; static void main(String[] args) { record Foo(int x, String y) implements ComparableRecord<Foo> { } var list = Stream.of(new Foo(2, "foo"), new Foo(2, "bar")) .sorted().toList(); System.out.println(list); } }

— John

On Jun 3, 2021, at 3:17 PM, forax@univ-mlv.fr<mailto:forax@univ-mlv.fr> wrote: ________________________________ De: "John Rose" <john.r.rose@oracle.com<mailto:r.rose@oracle.com>> À: "Remi Forax" <forax@univ-mlv.fr<mailto:forax@univ-mlv.fr>> Cc: "Peter Levart" <peter.levart@gmail.com<mailto:peter.levart@gmail.com>>, "Rémi Forax" <forax@openjdk.java.net<mailto:forax@openjdk.java.net>>, "core-libs-dev" <core-libs-dev@openjdk.java.net<mailto:core-libs-dev@openjdk.java.net>> Envoyé: Jeudi 3 Juin 2021 22:51:28 Objet: Re: RFR: 8199318: add idempotent copy operation for Map.Entry ... interface ComparableRecord<T extends Record & ComparableRecord<T>> extends Comparable<T> { … } record Foo(int x, String y) implements ComparableRecord<Foo> { … } http://cr.openjdk.java.net/~jrose/draft/ComparableRecord.java [repost with a link] The main issue with this kind of code is that the JIT does not see through the ClassValue. Wow, it’s almost as if we’ve discussed this already! ;-) So, https://bugs.openjdk.java.net/browse/JDK-8238260 Part of my agenda here is finding more reasons why JDK-8238260 deserves some love. Currently, the translation strategy for Records requires a lot of boilerplate generated into each subclass for toString, equals, and hashCode. It is partially declarative, because it uses indy. So, yay for that. But it is also a bad smell (IMO) that the trick needs to be played in each subclass. If ClassValue were performant, we could have just one method in Record for each of toString, equals, and hashCode, and have them DTRT. The user code would then be able to call super.toString() to get the standard behavior in a refining wrapper method in a subclass. Looking further, it would be nice to have methods which (a) inherit from supers as reusable code ought to, but (b) re-specialize themselves once per subclass indy-style. This is a VM trick I hope to look into after we do specialization. For now, ClassValue is the way to simulate that. Tweaking a little bit your code, I get https://gist.github.com/forax/e76367e1a90bf011692ee9bec65ff0f8<https://urldefense.com/v3/__https://gist.github.com/forax/e76367e1a90bf011692ee9bec65ff0f8__;!!GqivPVa7Brio!MheW9rHDHXlXYNKUju7v5G0vXlpj1YOoDWFjG9AcpnXnVz2TxlMYDM7i86yFtT_B$> (It's a PITA that we have to use a raw type to workaround circularly defined parameter type) I tried to DTRT and got into Inference Hell, surrounded by a swarms of unfriendly wildcards with pitchforks. Glad it wasn’t just me. Inspired by your more whole-hearted use of streams to build the code, I updated my example. Thanks. — John

On Wed, 2 Jun 2021 17:54:06 GMT, Stuart Marks <smarks@openjdk.org> wrote:

...

I'm fixing this along with a couple intertwined issues.

1. Add Map.Entry::copyOf as an idempotent copy operation.

2. Clarify that AbstractMap.SimpleImmutableEntry is itself unmodifiable (not really immutable) but that subclasses can be modifiable.

3. Clarify some confusing, historical wording about Map.Entry instances being obtainable only via iteration of a Map's entry-set view. This was never really true, since anyone could implement the Map.Entry interface, and it certainly was not true since JDK 1.6 when SimpleEntry and SimpleImmutableEntry were added.

Stuart Marks has updated the pull request incrementally with one additional commit since the last revision:

Tiny editorial tweaks.

src/java.base/share/classes/java/util/Map.java line 396:

394: 395: /** 396: * A map entry (key-value pair). The Entry may be unmodifiable, or the

In that case then maybe it should be `{@code Entry}` in both places where `entry` was replaced with `Entry` ? ------------- PR: https://git.openjdk.java.net/jdk/pull/4295

On Wed, 2 Jun 2021 17:54:06 GMT, Stuart Marks <smarks@openjdk.org> wrote:

...

I'm fixing this along with a couple intertwined issues.

1. Add Map.Entry::copyOf as an idempotent copy operation.

2. Clarify that AbstractMap.SimpleImmutableEntry is itself unmodifiable (not really immutable) but that subclasses can be modifiable.

3. Clarify some confusing, historical wording about Map.Entry instances being obtainable only via iteration of a Map's entry-set view. This was never really true, since anyone could implement the Map.Entry interface, and it certainly was not true since JDK 1.6 when SimpleEntry and SimpleImmutableEntry were added.

Stuart Marks has updated the pull request incrementally with one additional commit since the last revision:

Tiny editorial tweaks.

Marked as reviewed by psandoz (Reviewer). ------------- PR: https://git.openjdk.java.net/jdk/pull/4295