Can pattern methods be total ?

[Brian Goetz]

I don't necessarily mind if we get to a world where member patterns can 
be total.  (On the other hand, I don't mind if we never do, either.)  
The "dtors are total, others are partial" is a lump move, which is a 
good-enough-for-now-and-maybe-forever approximation.  If we are 
considering allowing totality to be denoted elsewhere, adding the 
ability to modify patterns later as total is an entirely reasonable 
option to preserve, and I don't think we've closed that off.

That said, I don't think that means we can shift the axes as you 
suggest.  Deconstructors are different from instance patterns in far 
more ways than just being total:

  - Deconstructor patterns are not inherited
  - Deconstructor patterns do not have names
  - Deconstructor patterns cannot be overridden, or abstract
  - Deconstructor patterns cannot have input parameters

Only the last still makes sense if you s/deconstructor/total/. 
Constructors are special in the language, for better or worse; their 
duals are too.

Now, you might say that the word "deconstructor" is fussy and that you 
prefer to use pattern without a name/arg list:

     pattern(BINDINGS) { ... } // dtor
     pattern(BINDINGS) name(ARGS) { ... } // method pattern

That's bikeshed stuff, but its a possibility to discuss once the model 
is nailed down.  (The word "deconstructor" would be more palatable if we 
used "constructor" for constructors, rather than the class name.  The 
C++ syntax "~C()" (un-C) would actually be good for deconstructors if it 
weren't polluted by C++.)

We need to define override and method selection rules anyway; existing 
rules for methods take modifiers into account (can't override instance 
with static, or public with private), so totality constraints fit pretty 
easily.

On 1/25/2021 7:32 AM, Remi Forax wrote:
> The last document sent by Brian split patterns in two categories, deconstructor and method patterns.
> I would like to propose another way to see patterns, along the axis of partial vs total pattern.
>
> A total pattern is a pattern that always succeed and bind the bindings.
> A partial pattern is a pattern that may succeed or not and if it succeed bind the bindings.
>
> Currently Brian proposes that all destructors are total patterns and all method patterns are partial patterns.
> While i agree for the former, as a user I will be surprising if "instanceof Point(var x, var y)" can fail not because it's not a Point but because the deconstruction did not succeed.
> I disagree with the later because having total method patterns is useful, by example you may want to be able to see a Point in cartesian coordinates or in polar coordinates. You can not use a deconstructor here because in both cases the bindings are two doubles, so you need a name to disambiguate between them.
> Something like
>    class Point {
>      ...
>      total pattern [double, double] polar() { ... }
>      total pattern [double, double] cartesian() { ... }
>    }
>    
> Here "total" means that the compiler can reject patterns with the same prefix that both use different total nested patterns on the same class. It also means that the compiler will enforce that body on the pattern method always "assign" values to the bindings (also allowing infinte loops and throws).
>
> If we go in that direction, we also need to introduce a rule for overriding, if a pattern method is declared total, all its overriden methods must be marked total (enforced by both the compiler and the VM).
> So if a pattern method is not total it can be overriden by a pattern method marked total but the reverse is not possible.
> Note that in term of compilation strategy this may requires a bridge method so we may decide to have a stricter rule, i.e. a pattern method that override a pattern method must have the same "totality".
>
> Another reason to know if a pattern is total or not is when we will introduce patterns at local variables declaration site.
>    Point(var x, var y) = point;
>
> I think only total patterns should be allowed here.
> So
>    Point.polar(var x, var y) = point;
> is ok but
>    Optional.of(var value) = optional;
> is not, because optional can be Optional.empty().
>
>
> So can pattern methods be total ?
>
> Rémi