(2.11.0-M8) whitebox macros are now first typechecked against outerPt

[xeno-by]

Even though whitebox macros are supposed to be used to produce expansions
that refine advertised return types of their macro definitions, sometimes
those more precise types aren’t picked up by the typechecker.

It all started with Travis generating structural types with macros
and noticing that typer needs an extra nudge in order to make generated
members accessible to the outside world. I didn’t understand the mechanism
of the phenomenon back then, and after some time I just gave up.

Afterwards, when this issue had been brought up again in a different
StackOverflow question, we discussed it at reflection meeting, figured out
that typedBlock provides some special treatment to anonymous classes,
and it became clear that the first macro typecheck (the one that types
the expansion against the return type of the corresponding macro def)
is at fault here.

The thing is that if we have a block that stands for a desugard anonymous
class instantiation, and we typecheck it with expected type different from
WildcardType, then typer isn’t going to include decls of the anonymous class
in the resulting structural type: https://github.com/scala/scala/blob/master/src/compiler/scala/tools/nsc/typechecker/Typers.scala#L2350.
I tried to figure it out at https://groups.google.com/forum/#!topic/scala-internals/eXQt-BPm4i8,
but couldn’t dispel the mystery, so again I just gave up.

But today I had a profound WAT experience that finally tipped the scales.
It turns out that if we typecheck an if, providing a suitable pt, then
the resulting type of an if is going to be that pt, even though the lub
of the branch types might be more precise. I’m sure that reasons for this
behavior are also beyond my understanding, so I decided to sidestep this problem.

upd. Here’s Jason’s clarification: Doing thing differently would require
us to believe that "'Tis better to have lubbed and lost than never to have
lubbed at all." But the desire for efficiency trumps such sentimentality.

Now expansions of whitebox macros are first typechecked against outerPt,
the expected type that comes from the enclosing context, before being
typechecked against innerPt, the expected type that comes from the return
type of the macro def. This means that now outerPt provides the correct
expected type for the initial, most important typecheck, which makes
types more precise.

[xeno-by]

review @retronym @travisbrown @DarkDimius

[retronym]

I’m sure that reasons for this behavior are also beyond my understanding, so I decided to sidestep this problem.

Doing thing differently would require us to believe that "'Tis better to have lubbed and lost than never to have lubbed at all." But the desire for efficiency trumps such sentimentality.

[adriaanm]

should this be a tristate instead? third state being "dunno"

[xeno-by]

I agree, duplication here is suboptimal. Therefore I just removed the isWhitebox method, because we actually don't need it.

[xeno-by]

@retronym Applied your suggestion, everything works.

[retronym]

Not sure if this matters, but we should consider this sort of thing:

class M[A: ClassTag] { println(implicitly[ClassTag[A]]) }
def impl: Tree = q"new M"
def apply: M[Any] = macro impl
{apply; () }

Under this scheme, we'll infer A = Nothing.

[xeno-by]

@retronym Thanks for bringing this up! I think we can live with that.

[paulp]

@xeno-by @retronym

It turns out that if we typecheck an if, providing a suitable pt, then the resulting type of an if is going to be that pt, even though the lub of the branch types might be more precise. I’m sure that reasons for this behavior are also beyond my understanding, so I decided to sidestep this problem.

Isn't he talking about the special case hacks in typedIf? The comment doesn't give much indication it's about performance, at least not primarily. It sounds more like bug avoidance shenanigans.

// in principle we should pack the types of each branch before lubbing, but lub doesn't really work for existentials anyway
// in the special (though common) case where the types are equal, it pays to pack before comparing
// especially virtpatmat needs more aggressive unification of skolemized types
// this breaks src/library/scala/collection/immutable/TrieIterator.scala
// annotated types need to be lubbed regardless (at least, continations break if you by pass them like this)

[retronym]

Not quite: it's this part that avoids lubbing the branches:

def finish(ownType: Type) = treeCopy.If(tree, cond1, thenp1, elsep1) setType ownType
if (isFullyDefined(pt))
  finish(pt)

So, in (if (true) "" else ""): Any, the If has type Any, rather than String.

[retronym]

LGTM

[jvican]

@retronym Is it on your roadmap to look at the root issues here so that we can avoid the double typechecking of whitebox macros?

[retronym]

That's not on my radar. Is there a canonical example of the problem?

[jvican]

The problem regarding lubs is the responsible for the reason why we typecheck whitebox macros three times instead of two, so I was wondering if it was worth having a look at it again so that we could get rid of the extra typechecker. Ideally, we would only typecheck whitebox macros once (as blackbox macros), but I don't know if that's technically feasible.