Allow case constants as nested patterns?
Brian Goetz
brian.goetz at oracle.com
Wed Nov 8 14:26:19 UTC 2023
I think you are probably thinking that "taking what is already possible
at the top level and allowing it in nested position" is somehow
considerably simpler than "constant patterns", and therefore a suitable
stopgap, but that's not the case. (Not only is it essentially the same
problem, but a stopgap always leaves the possibility we will foreclose
on getting to the real thing.)
It is an unfortunate artifact of transition (ideally, temporary) that a
switch can mix "constant cases" and "pattern cases"; this division adds
spec complexity and user friction for no direct benefit.
As to null, since `null` is a literal just like `0`, we would hope that
this would fit cleanly into constant patterns as well.
On 11/8/2023 7:35 AM, Sebastian Fischer wrote:
> Thank you both for your responses!
>
> I understand that more diverse kinds of nested patterns interact with
> multiple other planned extensions and in order to better understand
> those interactions, nested patterns have been restricted to record
> patterns for now.
>
> My question was less about constant patterns (or other unfinished
> extensions) and more about taking what is already possible at the
> top-level of a switch and allowing it also in nested pattern positions
> (probably caused by my hope that the notion of nested patterns could
> be extended incrementally with each new extension at the top-level).
> The example of refactoring a nested pattern into nested switches
> illustrates my intuition quite well: that nested patterns would have
> the same semantics as corresponding nested switches.
>
> The idea of unifying top-level with nested possibilities raises the
> question whether null patterns or even guards should be available in
> nested patterns. But I restricted my question to CaseConstant (in the
> currently supported form) because that is the most obvious gap and it
> complicates refactorings of empty records into enum constants.
> Personally, I think nested guards could be quite confusing but there
> might be a case for nested null patterns.
>
> Your summary helped me get an overview of previous considerations.
> I'll try to find more details in the archives and am looking forward
> to future discussions in the context of the mentioned JEPs.
>
> Sebastian
>
> On Tue, Nov 7, 2023 at 3:30 PM Brian Goetz <brian.goetz at oracle.com> wrote:
>
> To add to this: there are several categories of questions that we
> wanted to settle before considering constant patterns (which would
> (almost?) complete the rehabilitation of switch):
>
> - Syntactic. If `0` is both a literal and a pattern, then
> locutions like `f(0)` can either be method invocations or pattern
> matches. It was not obvious in the early days whether such
> ambiguities would pose roadblocks for using patterns in other
> contexts. This is seeming more viable now but we didn't want to
> foreclose this path too early.
>
> - User perception. Related to the above, if it is not obvious
> what `f(0)` means, even if the grammar can be parsed
> unambiguously, it might make code harder to read. So we might
> hold out for a syntax that is somewhat wordier but more obvious.
>
> - Semantics. At the beginning, it was very much not clear (to
> the point where our first ideas would have been pretty wrong) what
> the semantics of matching a constant are. JEP 455 has been
> working through those issues, so we believe we now have a firm
> foundation for the meaning of `0` as a pattern.
>
> - Alignment with switch. The "obvious" meaning of constant
> patterns may or may not align with the treatment of constants in
> switch today; this needs to be worked through carefully, otherwise
> refactoring a switch like:
>
> case Foo(k)
> case Foo(k')
> case Foo(k'')
>
> to a switch like
>
> case Foo(var x) {
> switch (x) {
> case k
> case k'
> case k''
> }
> }
>
> will have subtle asymmetries.
>
> So to sum this up, constant patterns are an "obvious" addition,
> but which turn out to be very much non-obvious when you consider
> all the connections with other language features and future
> language directions. So we've been waiting for these to settle.
>
> On 11/7/2023 4:37 AM, Gavin Bierman wrote:
>> Hi Sebastian,
>>
>> Yes, there has been some discussion about adding constant patterns if you look through the archive. There are interesting syntactic issues, but IIRC we have been waiting for the primitive type patterns feature to settle so we have a proper basis to deal with any type conversion questions. (For example, you write switch(new Box(b)) { case Box(0) -> … } where b is a byte and 0 is an integer constant…)
>>
>> So, it’s on our radar :-)
>>
>> Gavin
>>
>>> On 2 Nov 2023, at 11:13, Sebastian Fischer<mail at sebfisch.de> <mailto:mail at sebfisch.de> wrote:
>>>
>>> Hello.
>>>
>>> I would like to improve my understanding of the interaction of two JEPs: Pattern matching for switch and Record patterns.
>>>
>>> I am new to this list so might have missed previous discussion about this but could not find what I was looking for in the archives.
>>>
>>> JEP 441 shows the following grammar for switch labels.
>>>
>>> SwitchLabel:
>>> case CaseConstant { , CaseConstant }
>>> case null [, default]
>>> case Pattern [ Guard ]
>>> default
>>>
>>> JEP 440 shows the following grammar for patterns.
>>>
>>> Pattern:
>>> TypePattern
>>> RecordPattern
>>>
>>> TypePattern:
>>> LocalVariableDeclaration
>>>
>>> RecordPattern:
>>> ReferenceType ( [ PatternList ] )
>>>
>>> PatternList :
>>> Pattern { , Pattern }
>>>
>>> As a consequence it is not possible to have case constants in the pattern list of a record pattern. For example consider the following definitions modeling (a simplified version of) boolean expressions.
>>>
>>> sealed interface BoolExpr {
>>> enum Constant implements BoolExpr { TRUE, FALSE }
>>> record And(BoolExpr left, BoolExpr right) implements BoolExpr {}
>>>
>>> default BoolExpr recursively(UnaryOperator<BoolExpr> transform) {
>>> return transform.apply(switch (this) {
>>> case Constant c -> c;
>>> case And(var left, var right) -> new And(
>>> left.recursively(transform),
>>> right.recursively(transform)
>>> );
>>> });
>>> }
>>>
>>> default BoolExpr shortCircuit() {
>>> return recursively(expr -> switch (expr) {
>>> case And(var left, var unused)
>>> when left == Constant.FALSE -> Constant.FALSE;
>>> default ->
>>> expr;
>>> });
>>> }
>>> }
>>>
>>> In the definition of `shortCircuit` (which does partial evaluation) we need to use a guard to check if the left argument of `And` is false.
>>>
>>> Instead of using an enum, we can model constants as empty records as follows.
>>>
>>> sealed interface Constant extends BoolExpr permits True, False {}
>>> record True() implements Constant {}
>>> record False() implements Constant {}
>>>
>>> Now we can write `shortCircuit` using a nested record pattern.
>>>
>>> default BoolExpr shortCircuit() {
>>> return recursively(expr -> switch (expr) {
>>> case And(False(), var unused) -> new False();
>>> default -> expr;
>>> });
>>> }
>>>
>>> It would be convenient to be able to use a nested pattern instead of a guard with the original definition using an enum for constants. Here is a hypothetical implementation of `shortCircuit` that is currently not supported.
>>>
>>> default BoolExpr shortCircuit() {
>>> return recursively(expr -> switch (expr) {
>>> case And(Constant.FALSE, var unused) -> Constant.FALSE;
>>> default -> expr;
>>> });
>>> }
>>>
>>> What are potential problems with allowing case constants as nested patterns?
>>>
>>> Kind regards,
>>> Sebastian
>>>
>>>
>>>
>>>
>>>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://mail.openjdk.org/pipermail/amber-dev/attachments/20231108/a95ad1c5/attachment-0001.htm>
More information about the amber-dev
mailing list