Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Match type bounded by type with argument(s) cannot be reduced to satisfy =:= check #17149

Closed
tribbloid opened this issue Mar 25, 2023 · 12 comments · Fixed by #17180
Closed

Match type bounded by type with argument(s) cannot be reduced to satisfy =:= check #17149

tribbloid opened this issue Mar 25, 2023 · 12 comments · Fixed by #17180

Comments

@tribbloid
Copy link

tribbloid commented Mar 25, 2023

Pointed out by @DmytroMitin in the following post:

https://stackoverflow.com/questions/75837775/in-scala-2-or-3-is-there-a-higher-kind-argument-extractor-without-using-match-t/75842531#75842531

  type Ext[S <: Seq[_]] = S match {
    case Seq[t] => t
  }

implicitly[Ext[Seq[Int]] =:= Int] // e.scala: Cannot prove that e.Ext[Seq[Int]] =:= Int

If I change the definition of Ext as follow, the consequence will differ:

  type Ext[S] = S match {
    case Seq[t] => t
  } // success!

  type Ext[S <: Any] = S match {
    case Seq[t] => t
  } // success!

  type Ext[S <: Seq[Any]] = S match {
    case Seq[t] => t
  } // same error
@Decel
Copy link
Collaborator

Decel commented Mar 27, 2023

For some reason, we don't get a reduction:

adapt to subtype A =:= A !<:< Ext[Seq[Int]] =:= Int
>>>> StoredError: Found:    A =:= A
Required: Ext[Seq[Int]] =:= Int

where:    A is a type variable
adapt to subtype A =:= A !<:< Ext[Seq[Int]] =:= Int
>>>> StoredError: Found:    A =:= A
Required: Ext[Seq[Int]] =:= Int

if we use [S <: Seq[Int]] everything works, because during isSubType, we have an upper bound of Int, instead of Any, so for example:

==> isSubType Ext[Seq[Int]] <:< Nothing?
  ==> isSubType Int <:< Nothing?
  <== isSubType Int <:< Nothing = false
==> isSubType Ext[Seq[Int]] <:< Nothing?
 ==> isSubType Any <:< Nothing?
 <== isSubType Any <:< Nothing = false

So, this works:

type Ext[S <: Seq[Any]] = S match
  case Seq[t] => t

val a = summon[Ext[Seq[Int]] =:= Any]

@Decel
Copy link
Collaborator

Decel commented Mar 27, 2023

And, unsurprisingly, works in 3.1.3 -- Looks like a duplicate of #15926.

@sjrd
Copy link
Member

sjrd commented Mar 27, 2023

Also works in tasty-query 😁

@Decel
Copy link
Collaborator

Decel commented Mar 28, 2023

The whole problematic tree looks like this:

==> S match case Seq[t] => t?
  ==> isSubType S <:< Seq[t]?
  <== isSubType S <:< Seq[t] = false
  ==> isSubType S <:< Seq[t]?
    ==> isSubType Seq[Any] <:< Seq[t] (left is approximated)?
      ==> isSubType Any <:< t?
        ==> isSubType Any <:< Nothing?
        <== isSubType Any <:< Nothing = false
        ==> isSubType Any <:< Any?
          ==> isSubType scala.type <:< (scala : scala.type)?
          <== isSubType scala.type <:< (scala : scala.type) = true
        <== isSubType Any <:< Any = true
      <== isSubType Any <:< t = true
    <== isSubType Seq[Any] <:< Seq[t] (left is approximated) = true
  <== isSubType S <:< Seq[t] = true
  ==> isSubType Any <:< Any?
    ==> isSubType (scala : scala.type) <:< scala.type?
    <== isSubType (scala : scala.type) <:< scala.type = true
  <== isSubType Any <:< Any = true
<== S match case Seq[t] => t = Reduced(Any)

So, in the end, we get:

type Ext = ([S <: Seq[Any]] =>> Any)

😢

@dwijnand
Copy link
Member

Yeah, I looked at that trace again, from that duplicate. My original fix was to disallow, using your trace, S widening to Seq[Any]. But last week I was wondering whether the issue was the loss of approximation state that was to blame.

@Decel
Copy link
Collaborator

Decel commented Mar 29, 2023

Yeah, I looked at that trace again, from that duplicate. My original fix was to disallow, using your trace, S widening to Seq[Any]. But last week I was wondering whether the issue was the loss of approximation state that was to blame.

Disallowing it seems too restrictive, -- and if we try to widen it to Any, then we may violate type bounds. I can't see a nice workaround here, it's a fascinatingly bad situation.

I am not sure about the approximation state -- what is it supposed to look like? Seems to me, the approximation is correct because anything other than Seq[Any] will be a bounds violation, but not sure if I understand you completely.

@dwijnand
Copy link
Member

dwijnand commented Mar 29, 2023

but not sure if I understand you completely.

ApproxState maintains whether the LHS has been widened and/or the RHS has been narrowed, and that affects, for instance, whether to consider it as a constraint bound. Here I'm thinking of the check if canConstrain(tp2) && !approx.low.

In the trace it (sometimes) appears as "(left is approximated)" (or right), and it disappears when checking args. I was wondering if that was the problem.

@Decel
Copy link
Collaborator

Decel commented Mar 29, 2023

I looked into 3.1.3 trace:

==> isSubType S <:< Seq[t]?
<== isSubType S <:< Seq[t] = false
==> isSubType Any <:< Any?
  ==> isSubType scala.type <:< (scala : scala.type)?
  <== isSubType scala.type <:< (scala : scala.type) = true
<== isSubType Any <:< Any = true
==> isSubType Nothing <:< Any?
<== isSubType Nothing <:< Any = true

So it's widening to Any instead of Seq[Any] and in 3.1.3 the following code works:

type Ext[S <: Seq[Int]] = S match
  case Seq[t] => t

val a = summon[Ext[Seq[String]] =:= String]

UPD:

Ah, that's #17168. If we use implicitly instead we run into error.

  |val a = implicitly[Ext[Seq[String]] =:= String]
  |                                               ^
  |                            Cannot prove that Ext[Seq[String]] =:= String.

@dwijnand
Copy link
Member

dwijnand commented Mar 29, 2023

Yeah, not widening the abstract type S to its bounds Seq[Int] is what I meant by "my" original fix. But it was really to see whether reverting that part (back to how Olivier had it) fixed that issue, which it did. Seth did a good job attempting to narrate and crosslink this saga in #15926 (comment).

@dwijnand
Copy link
Member

Disallowing it seems too restrictive

Going back to this comment, I think disallowing is correct: we don't want the match type to reduce to the upper bound, because then we lose the tie to the actually passed type. We want the String in Seq[String], not Any, even though Any is String's upper bound. The match type, as it's defined using an abstract type in the match alias's type lambda, should stay an unreduced match type (because reduction gets stuck).

@dwijnand dwijnand linked a pull request Mar 29, 2023 that will close this issue
@odersky
Copy link
Contributor

odersky commented Mar 29, 2023

I had a look at this with @Decel. The critical PR which led to this situation is #15423, which itself follows up on several previous PRs that go back and forth on the issue. Essentially, there are two interlinked questions

  1. Are we allowed to widen an abstract type or type parameter to its bound when reducing a matchtype?
  2. Is match type reduction allowed to instantiate variables in patterns?

Originally, the answer was yes to both, but one can easily show that that's unsound. Then #13780 disallowed (1), but that caused a lot of regressions of reasonable code that worked before. Then #15423 answered the question by trying two different strategies as shown in this snippet in TypeComparer#matchCases

      if caseLambda.exists && matches(canWidenAbstract = false) then
        redux(canApprox = true)
      else if matches(canWidenAbstract = true) then
        redux(canApprox = false)

So, we first say no to (1), and if that fails we try again saying yes to (1), but restricting how type variables get instantiated.
The necessary restriction is as follows:

If we widen a type S <: T to T then any type variables we instantiate would get exactly the same instantiations if we used some subtype of T for S.

But that's super hard to check, so we need an approximation. The approximation we used was that, if at the end of the comparison, a type variable was constrained to a single point, then we can instantiate the type variable to that point. Might look reasonable, but is unsound. The problem was in the example here. We have:

  S <: Seq[_]   in the scrutinee
  S <: Seq[t]    in the pattern match

We widen S to Seq[_] which gives us Seq[_] <: Seq[t], which gives us t = Any. The same result would be obtained if we hard started with the bound S <: Seq[Any]. But we could have instantiated S to the subtype Seq[Int] instead which would give t = Int. So we are not allowed to instantiate t to Any, after all, even though it is uniquely determined from Seq[_] <: Seq[t].

Not yet sure what to do about this.

odersky added a commit to dotty-staging/dotty that referenced this issue Mar 29, 2023
Refine criterion when to widen types in match type reduction.

We now do not widen if the compared-against type contains
variant named type parameters. This is intended to fix the
soundness problem in scala#17149.

Fixes scala#17149

Todos:

 - [ ] Check & fix neg test failures
 - [ ] Add more tests
 - [ ] Also consider approximating abstract types to lower bounds.
       This is completely missing so far. There are neither tests
       nor an implementation.
 - [ ] Update the docs on match types to explain what goes on here.
odersky added a commit to dotty-staging/dotty that referenced this issue Mar 29, 2023
Refine criterion when to widen types in match type reduction.

We now do not widen if the compared-against type contains
variant named type parameters. This is intended to fix the
soundness problem in scala#17149.

Fixes scala#17149
Fixes scala#15926

Todos:

 - [ ] Check & fix neg test failures
 - [ ] Add more tests
 - [ ] Also consider approximating abstract types to lower bounds.
       This is completely missing so far. There are neither tests
       nor an implementation.
 - [ ] Update the docs on match types to explain what goes on here.
@tribbloid tribbloid changed the title Match type bounded by type with argument(s) cannot be normalised to satisfy =:= check Match type bounded by type with argument(s) cannot be reduced to satisfy =:= check Mar 29, 2023
@tribbloid
Copy link
Author

tribbloid commented May 3, 2023

I also have a simplified dual form:

      trait SGen {

        type S <: Seq[_]

        type Ext = S match {
          case Seq[t] => t
        }
      }

      object SeqIntGen extends SGen {

        type S = Seq[Int]
      }

      implicitly[SeqIntGen.Ext =:= Int] // still broken

The fix should theoretically work on both cases

dwijnand pushed a commit to dotty-staging/dotty that referenced this issue May 24, 2023
Refine criterion when to widen types in match type reduction.

We now do not widen if the compared-against type contains
variant named type parameters. This is intended to fix the
soundness problem in scala#17149.

Fixes scala#17149
Fixes scala#15926

Todos:

 - [ ] Check & fix neg test failures
 - [ ] Add more tests
 - [ ] Also consider approximating abstract types to lower bounds.
       This is completely missing so far. There are neither tests
       nor an implementation.
 - [ ] Update the docs on match types to explain what goes on here.
@Kordyjan Kordyjan added this to the 3.4.0 milestone Aug 1, 2023
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Projects
None yet
6 participants