Don't simplify |/& by taking lubs/glbs for explicitly declared types.

[odersky]

A lub might return a different type (i.e. singleton types are widenened).
We should not do this if the type is given explicitly.

This is a more conservative version of #844. Review by @smarter.
@smarter can you push your tests on top of this PR? Thanks!

[smarter]

/rebuild

[smarter]

• With this patch, we still assign the type of OrTypeTree using |(https://github.com/dotty-staging/dotty/blob/change-simplify-types/src/dotty/tools/dotc/typer/TypeAssigner.scala#L369) which calls lub, so we're still losing singletons.
• Even if this is fixed, this patch does not help in situations like this:

val foo: a.type | b.type = a
val bar: a.type | b.type = foo

The type of the tree corresponding to foo in the right-hand side of bar is going to be adapted, and since the type is not written explicitly it will be simplified to Int, so we'll get a type mismatch.

• Since we call Type#simplified when instantiating tvars (https://github.com/dotty-staging/dotty/blob/change-simplify-types/src/dotty/tools/dotc/core/Types.scala#L2475), we can probably run into a situation where we instantiate a type variable in a way that result in a type mismatch, but I'm too jetlagged to think up an example right now.

[odersky]

I think all this is correct. But what can we do? Disallow unions of
singletons in types?

On Fri, Oct 30, 2015 at 10:05 AM, Guillaume Martres <
[email protected]> wrote:

• With this patch, we still assign the type of OrTypeTree using |(
  https://github.com/dotty-staging/dotty/blob/change-simplify-types/src/dotty/tools/dotc/typer/TypeAssigner.scala#L369)
  which calls lub, so we're still losing singletons.
• Even if this is fixed, this patch does not help in situations like
  this:

val foo: a.type | b.type = aval bar: a.type | b.type = foo

The type of the tree corresponding to foo in the right-hand side of bar
is going to be adapted, and since the type is not written explicitly it
will be simplified to Int, so we'll get a type mismatch.

• Since we call Type#simplified when instantiating tvars (
  https://github.com/dotty-staging/dotty/blob/change-simplify-types/src/dotty/tools/dotc/core/Types.scala#L2475),
  we can probably run into a situation where we instantiate a type variable
  in a way that result in a type mismatch, but I'm too jetlagged to think up
  an example right now.

—
Reply to this email directly or view it on GitHub
#885 (comment).

Martin Odersky
EPFL

[smarter]

Well, #844 seems to work fine, do you see any cases where it might do something wrong?

[odersky]

The danger with #844 is that it produces too large types and will make
inference very slow or even cause it to blow up. Hard to test for it.

On Fri, Oct 30, 2015 at 10:15 AM, Guillaume Martres <
[email protected]> wrote:

Well, #844 #844 seems to work
fine, do you see any cases where it might do something wrong?

—
Reply to this email directly or view it on GitHub
#885 (comment).

Martin Odersky
EPFL

[odersky]

I now believe we are on the wrong track altogether. Messing with the meaning of lubs gets us into trouble. If we want lubs to widen singleton types we have to disallow union types consisting of singleton types. Everything else would be too operational.