Ad hoc type restriction

[Archie Cobbs]

Ethan McCue <ethan at mccue.dev> wrote:

> However there is nothing conceptually preventing the tools validating
> @NonNull usage from also emitting an error until you have inserted a known
> precheck.
> ...
> But for other single-value invariants, like your @PhoneNumber example, it
> seems fairly practical. Especially since, as a general rule, arbitrary cost
> computations really shouldn't be invisible. How would one know if (@B A) is
> going to thread invocations of some validation method everywhere?


This is why 3rd party tools aren't as good as having the compiler handle
it, because the compiler is in a position to provide both stronger and more
efficient guarantees - think generic types and runtime erasure.
Compiler-supported typing allows the developer to move the burden of proof
from the method receiving a parameter to the code invoking that method, and
onward back up the call chain, so that validations tend to occur "early",
when they are first known to be true, instead of "late" at the (many more)
points in the code where someone actually cares that they are true.

So if phone numbers are central to your application, and they are passed
around and used all over the place as type @PhoneNumber String, then they
will only need to actually be validated at a few application entry points,
not at the start of every method that has a phone number as a parameter. In
other words, the annotation is ideally not a "to-do" list but rather an
"it's already done" list.

The guarantee that the compiler would then provide is ideally on the same
level as with generics: while it's being provided by the compiler, not the
JVM, so you can always get around it if you try hard enough (native code,
reflection, class file switcheroo, etc.), as long as you "follow the rules"
you get the guarantee - or if not, an error or at least a warning.

Brian Goetz <brian.goetz at oracle.com> wrote:

> I think the best bet for making this usable would be some mechanism like a
> "view", likely only on value types, that would erase down to the underlying
> wrapped type, but interpose yourself on construction, and provided a
> conversion from T to RefinedT that verified the requirement.  But this is
> both nontrivial and presumes a lot of stuff we don't even have yet...
>

I think that is close to what I was imagining. It seems like it could be
done with fairly minimal impact/disruption...? No need for wrappers or
views.

But first just to be clear, what I'm getting at here is a fairly narrow
idea, i.e., what relatively simple thing might the compiler do, with a
worthwhile cost/benefit ratio, to make it easier for developers to reason
about the correctness of their code when "type restriction" is being used,
either formally or informally (meaning, if you're using an int to pass
around the size of collection, you're doing informal type restriction).

What's the benefit? Type restriction is fairly pervasive, and yet because
Java doesn't make it very easy to do, it's often not being done at all, and
this ends up adding to the amount of manual work developers must do to
prove to themselves their code is correct. The more of this burden the
compiler could take on, the bigger the benefit would be.

What's the cost? That depends on the solution of course.

To me the giant poster-child for this kind of pragmatic language addition
is generics. It had all kinds of minor flaws from the point of view of
language design, but the problem it addressed was so pervasive, and the new
tool it provided to developers for verifying the correctness of their code
was so powerful, that nobody thinks it wasn't worth the trade-off.

OK let me throw out two straw-man proposals. I'll just assume these are
stupid/naive ideas with major flaws. Hopefully they can at least help map
out the usable territory - if any exists.

*Proposal #1*

This one is very simple, but provides a weaker guarantee.

   1. The compiler recognizes and tracks "type restriction annotations",
   which are type annotations having the meta-annotation @TypeRestriction
   2. For all operations assigning some value v of type S to type T:
      1. If a type restriction annotation A is present on T but not S, the
      compiler generates a warning in the new lint category
      "type-restriction"

That's it. A cast like var pn = (@PhoneNumber String)input functions simply
as a developer assertion that the type restriction has been verified, but
the compiler does not actually check this. There is no change to the
generated bytecode. If the developer chooses to write a validation method
that takes a string, validates it (or throws an exception), and then
returns the validated string, that method will need to be annotated with
@SuppressWarnings("type-restriction") because of the cast in front of the
return statement.

Guarantee provided: Proper type restriction as long as "type-restriction"
warnings are enabled and not emitted. However, this is a "fail slow"
guarantee: it's easy to defeat (just cast!). So if you write a method that
takes a @PhoneNumber String parameter that is passed an invalid value, you
won't find out until something goes wrong later down the line (or never).
In other words, *your* code will be correct, but you have to be trusting of
any code that *invokes* your code, which in practice is not always a sound
strategy.

*Proposal #2*

This is proposal is more complex but provides a stronger guarantee:

   1. The compiler recognizes and tracks "type restriction annotations",
   which have the meta-annotation @TypeRestriction
      1. The annotation specifies a user-supplied "constructor" class
      providing a user-defined construction/validation method validate(v)
      2. We add class TypeRestrictionException extends RuntimeException and
      encourage validate() methods to throw (some subclass of) it
   2. For all operations assigning some value v of type S to type T:
      1. If a type restriction annotation A is present on T but not S, the
      compiler generates a "type-restriction" warning AND adds an implicit
      cast added (see next step)
   3. For every cast like  var pn = (@PhoneNumber String)"+15105551212"  the
   compiler inserts bytecode to invoke the appropriate enforcer validate(v)
   method
   4. The JLS rules for method resolution, type inference, etc., do not
   change (that would be way over-complicating things)
      1. Two methods void dial(String pn) and void dial(@PhoneNumber String
      pn) will still collide

Guarantee provided: Proper type restriction unless you are going to
extremes (native code, reflection, runtime classfile switcheroo, etc.).
This is a "fail fast" guarantee: errors are caught at the moment an invalid
value is assigned to a type-restricted variable. If your method parameters
have the annotation, you don't have to trust 3rd party code that calls
those methods (as long as it was compiled properly). I.e., same level of
guarantee as generics.

These are by no means complete or particularly elegant solutions from a
language design point of view. They are pragmatic and relatively
unobtrusive add-ons, using existing language concepts, to get us most of
what we want, which is:

   - User-defined "custom" type restrictions with compile-time
   checking/enforcement
      - As with generics, the goal is not language perfection, but rather
      making it easier for developers to reason about correctness
   - Compile-time guarantees that type restricted values in source files
   will be actually type restricted at runtime
   - Efficient implementation
      - Validation only happens "when necessary"
      - No JVM changes needed (erasure)
   - No changes to language syntax; existing source files are 100% backward
   compatible

The developer side of me says that the cost/benefit ratio of something like
this would be worthwhile, in spite of its pragmatic nature, simply because
the problem being addressed seems so pervasive. I felt the same way about
generics (which was a much bigger change addressing a much bigger pervasive
problem).

But I'm sure there are things I'm missing... ?

-Archie

-- 
Archie L. Cobbs
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://mail.openjdk.org/pipermail/amber-dev/attachments/20251013/a6919440/attachment.htm>