Type Class / Witness

[Remi Forax]

In the future, we want to retrofit primitive to be value types.

Type classes are one of the solutions for that.

Rémi


On August 8, 2025 9:12:28 PM UTC, David Alayachew <davidalayachew at gmail.com> wrote:
>Woah, Type Classes!
>
>Ok, design talk aside, are you saying type classes are something you all
>have decided to do? Is that a feature, even if only for value classes, that
>we can safely assume is on the road map?
>
>All previous discussion on this made it sound like it's being strongly
>considered, but not on the road map.
>
>
>On Fri, Aug 8, 2025, 4:02 PM Brian Goetz <brian.goetz at oracle.com> wrote:
>
>> This is a valid observation, but you are getting way ahead of yourself.
>> My talk this week was about general mechanisms; we are not designing this
>> feature yet, and discussions about it are premature (and likely
>> counterproductive, as people chime in with questions about something that
>> is not yet designed.)
>>
>> > On Aug 8, 2025, at 12:06 PM, Remi Forax <forax at univ-mlv.fr> wrote:
>> >
>> > Hello,
>> > for those who where not at the JVMLS, it may make sense to re-read this
>> email later,
>> > when the talk of Brian on Type class / witness is visible on youtube.
>> > I write this now because i'm sure i will totally forget about that in
>> few days.
>> >
>> > During the talk, Brian said that witness are restricted to be used only
>> by value classes so we avoid the operator overloading hell that you can see
>> in other languages.
>> >
>> > I think this restriction can be bypassed.
>> >
>> > The problem is that you can always wrap any class into a value class
>> organized as a monad (containing a superposition of states) thus bypassing
>> that restriction.
>> >
>> >
>> > As an example, let say we want to abuse of the >> operator to emulate
>> C++ iostream.
>> > Let suppose we have this witness interface defined in the JDK (it is by
>> example used by Integer)
>> >
>> > witness interface BitwiseNumeric<V> {
>> >  V rightShift(V v1, V v2);
>> >
>> >  // other methods
>> > }
>> >
>> > And the witness interface for exact conversion also defined in the JDK
>> >
>> > witness interface Conversion<U, V> {
>> >  U convert(V value);
>> > }
>> >
>> >
>> > So the trick is to create a value class that encode either a Writer or
>> any value you want to add to the writer as a monad,
>> > The exact conversion is used to wrap any value into the IOStream, the
>> righShift is used to consume the wrapped value using a side-effect.
>> >
>> > value class IOStream {
>> > private final Writer writer;
>> > private final Object o;
>> >
>> > private IOStream(Writer writer, Object o) { this.writer = writer; this.o
>> = o; super(); }
>> >
>> > public IOWriter(Writer! writer) { this(writer, null); }
>> >
>> > public static final witness BitwiseNumeric<IOStream!> BITWISE_NUMERIC =
>> new BitwiseNumeric<>() {
>> >   public IOStream! rightShift(IOStream! v1, IOStream! v2) {
>> >     if (v1.writer == null || v2.writer != null) {
>> >       throw new ISE();
>> >     }
>> >     v1.writer.append("" + v2.o);
>> >     return v1;
>> >   }
>> >
>> >   // other methods throw UnsupportedOperationException
>> > };
>> >
>> > public static final witness Conversion<IOStream!, Integer!> FROM_INTEGER
>> = new Conversion<>() {
>> >   IOStream! convert(Integer! value) {
>> >     return new IOStream(null, value);
>> >   }
>> > };
>> > }
>> >
>> > So one can write
>> >  IOStream stream = new IOStream(writer);
>> >  var _ = stream >> 2 >> 3;
>> >
>> > I think this is a general mechanism so any operator can be abused that
>> way.
>> >
>> > regards,
>> > Rémi
>>
>>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://mail.openjdk.org/pipermail/amber-spec-experts/attachments/20250808/65bcc5a0/attachment.htm>