Late change to JEP 433

[forax at univ-mlv.fr]

> From: "Brian Goetz" <brian.goetz at oracle.com>
> To: "Remi Forax" <forax at univ-mlv.fr>
> Cc: "John Rose" <john.r.rose at oracle.com>, "Alex Buckley"
> <alex.buckley at oracle.com>, "amber-spec-experts"
> <amber-spec-experts at openjdk.java.net>
> Sent: Tuesday, November 15, 2022 12:35:24 AM
> Subject: Re: Late change to JEP 433

> I win my bet! I had my money on "If you send this, Remi will surely attempt to
> reopen his pet NPE issue." We have been through this all before, so unless you
> have something dramatically new and compelling, we're not reopening this issue
> now.

> As a reminder, not all causes of MatchException come from separate compilation
> anomalies. Some come from match remainder, which is just an ordinary domain
> error, and separating the two is only possible in the easy cases.

> I am not sure you fully understand how this works, otherwise you wouldn't keep
> raising the same issue after getting the same explanation. (For example, the
> "NPE can be thrown organically" claim you make here, which you've made before,
> is wrong; it is possible a later pattern can match Box(null), and we do not
> throw ME until all choices are exhausted.)
Let say we have, 
record Box(Box b) { } 

and 
Box box = ... 
switch(box) { 
case Box(Box(var value)) -> ... // 1 
case Box value2 -> ... // 2 
} 

you can translate it line by line that way, 
if (box == null) { 
throw new NPE(); 
} 
if (box instanceof Box $b1) { 
Box $b2 = $b1.b; 
if ($b2 instanceof Box value) { 
... // 1 
} 
} 
if (box instanceof Box value2) { 
... // 2 
} 
throw new MatchException(...) 

but 
- the first null check is not really needed if later you ask for the content of the box, 
- the instanceofs "box instanceof Box $b1" and "box instanceof Box value2" are useless and you do not need to do that check twice 
a better code is 

Box $b2 = box.b; // will throw a NPE 
if ($b2 == null) { 
... // 2 
} 
Box value2 = $b2; 
... // 1 

You can observe that like the switch on strings or the switch on enums, the NPE comes "organically" from asking the content of the box. 

The "throw new MatchException" disappears because there is no need for a catch all anymore, this is also true if the last instanceof on a subtype of a sealed type is transformed to a cast instead of an instanceof. 

But generating such code requires to merge subsequent patterns like Jan has tried to do in a recent patch, something i believe we will have to do anyway because the current perf is not that great. 

Rémi 

> On 11/14/2022 6:23 PM, [ mailto:forax at univ-mlv.fr | forax at univ-mlv.fr ] wrote:

>>> From: "Brian Goetz" [ mailto:brian.goetz at oracle.com | <brian.goetz at oracle.com> ]
>>> To: "Remi Forax" [ mailto:forax at univ-mlv.fr | <forax at univ-mlv.fr> ] , "John
>>> Rose" [ mailto:john.r.rose at oracle.com | <john.r.rose at oracle.com> ]
>>> Cc: "Alex Buckley" [ mailto:alex.buckley at oracle.com | <alex.buckley at oracle.com>
>>> ] , "amber-spec-experts" [ mailto:amber-spec-experts at openjdk.java.net |
>>> <amber-spec-experts at openjdk.java.net> ]
>>> Sent: Monday, November 14, 2022 11:40:27 PM
>>> Subject: Re: Late change to JEP 433

>>> Its MatchException. Error would not be appropriate, since this is the same
>>> exception that gets used for remainder (e.g., Box(Box(var x)) against a target
>>> of Box(null)).
>> I still think we should throw a NPE when a destructuring something which is
>> null.

>> It will make the semantics of MatchException very similar to
>> LambdaConversionException, i.e. a lot of users will never see it because it can
>> be thrown only when there is a separate compilation issue or a bug in the
>> deconstructor.

>> If you take a look to the implementation, the NPEs can be thrown organically as
>> the result of calling the deconstructor (or the accessor methods) on null,
>> instead of adding a nullcheck that goto to the location where the
>> MatchException is thrown.

>> Rémi

>>> On 11/14/2022 5:31 PM, Remi Forax wrote:

>>>> I'm confused, is it MatchError or MatchException ?

>>>> Because if it's an error instead of an exception, it may be less an issue in
>>>> term of backward compatibility but it is not what is proposed, right ?

>>>> Rémi

>>>> ----- Original Message -----

>>>>> From: "John Rose" [ mailto:john.r.rose at oracle.com | <john.r.rose at oracle.com> ]
>>>>> To: "Alex Buckley" [ mailto:alex.buckley at oracle.com | <alex.buckley at oracle.com>
>>>>> ] Cc: "amber-spec-experts" [ mailto:amber-spec-experts at openjdk.java.net |
>>>>> <amber-spec-experts at openjdk.java.net> ] Sent: Monday, November 14, 2022 8:24:04
>>>>> PM
>>>>> Subject: Re: Late change to JEP 433

>>>>> On 14 Nov 2022, at 10:56, Alex Buckley wrote:

>>>>>> … ICCE can hand over to MatchException …

>>>>> Precisely; I agree that it is time for this to happen.  Thanks, Alex for
>>>>> reminding us of the history and lineage of ICCE, and why it doesn’t make sense
>>>>> for switch statements (not even classic switch-over-enum).

>>>>> As a program linkage error, ICCE is necessarily uncommunicative when applied to
>>>>> a misconfigured switch statement.  Using a MatchError is on the other hand
>>>>> highly informative:  The user (and possibly try/catch logic surrounding the
>>>>> failure) knows exactly what happened, that a set of matches expected to succeed
>>>>> has failed.

>>>>> So, here’s a possible knock-on advantage if we hand off from ICCE to ME:  If
>>>>> there is further development of the concept of “a set of matches which is
>>>>> required to succeed”, the error processing can continue to be unified under the
>>>>> ME.  Brian’s ideas about “let-statements” entail a single pattern which is
>>>>> required to match; ME is surely the right way to signal failure (if not
>>>>> something more specific like CCE or NPE, which is an interesting
>>>>> side-conversation).  Or, if we ever did Haskell-style method overloads that
>>>>> discriminate arguments by means of patterns, surely they would desugar to
>>>>> omnibus methods that start with switches; once again ME surely makes sense as a
>>>>> way to signal inapplicability of such match-based methods.

>>>>> — John

>>>>> P.S. More speculatively, and probably a bridge too far, would be to employ
>>>>> MatchException as a part of a meta-language protocol that defines how sets of
>>>>> patterns compose, and in particular how one part of a composite signals failure
>>>>> to the whole composite.  (Surely there is some future use for X in methods :
>>>>> method handles :: patterns : X; that’s what I mean by a meta-language
>>>>> protocol.)  I say this is a bridge too far because Java exceptions are not a
>>>>> very good tool for normally-frequent control flow, and also because ME, like
>>>>> NPE or CCE, probably best signals a failure of the programmer’s settled
>>>>> intentions about some code, rather than signaling an alternative control path
>>>>> (like if/else).  Still, I wanted to point this out because in other languages
>>>>> exception-like concepts are used to convey backtracking out of composite
>>>>> control flow patterns, and if we decided to try this out for Java,
>>>>> MatchExpression would raise its hand and say “pick me!”
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://mail.openjdk.org/pipermail/amber-spec-observers/attachments/20221115/ffd7d0af/attachment.htm>