Fix boxing of Nothing / Add explicit type annotation to Val, Let

[b-studios]

No description provided.

[b-studios]

why do you need to translate one let into two here?

[marzipankaiser]

Because the coerce expects a Pure [1] expression but the binding is not.
So we introduce two bindings:

• One for the "original" value
• One for the boxed/unboxed value

AFAICT there is no binding construct for expressions, is there?
Alterantively, we could use something like Run(Let(orig, origTpe, transform(binding), coerce(...))).

[1]: This is mostly to cut down on cases, and since we will use the value as a parameter to Make anyway (so an intermediate value is needed).

[b-studios]

This pattern shows up quite often, how can we simplify?

[marzipankaiser]

Something like foo(v.symbol.tpe, transformed.tpe) might be possible, equivalent to a
v.symbol.tpe.map(transform).getOrElse(transformed.tpe).
Though I'm not sure it's worth it.

I am unsure if we can abstract over more easily since the type of binding/transformed is different in most cases (Stmt for val, Expr for let). - This also means that the call to transform (on the binding) is a different one.

[b-studios]

Sadly, this still doesn't solve the LLVM issue for me on #449

Even when merging this PR here I still get

opt: ./out/strings.ll:13764:9: error: multiple definition of local value named 'tmp31441_4389'
        %tmp31441_4389 = load %Pos, ptr %tmp31441_4389p_4562
        ^

on

run --backend llvm examples/stdlib/string/strings.effekt

[marzipankaiser]

@phischu @b-studios
The next minified example is:

interface MyException {
  def myraise(): Nothing
}

def myDefault[T]{d: => T}{body: => T / MyException} = try { body() } with MyException {
  def myraise() = d()
}

def main() = {
  try {
    myDefault { val t = do myraise(); t }{ 0 };
    ()
  } with MyException {
    def myraise() = ()
  }
}

[marzipankaiser]

I think the problem is here:

effekt/effekt/shared/src/main/scala/effekt/core/PolymorphismBoxing.scala

Line 187 in 2739c47

instantiate(calleeT, targs).call(calleeT, vargs map transform, bargs map transform)

effekt/effekt/shared/src/main/scala/effekt/core/PolymorphismBoxing.scala

Line 218 in 2739c47

instantiate(b, targs).callDirect(b, vargs map transform, bargs map transform)

effekt/effekt/shared/src/main/scala/effekt/core/PolymorphismBoxing.scala

Line 226 in 2739c47

case Pure.PureApp(b, targs, vargs) => instantiate(b, targs).callPure(b, vargs map transform)

Where now, we also need to check the paramaters and potentially box/unbox those.
This is only due to subtyping, which was added recently (for parametric polymorphism, representations of ValueTypes match up after instantiation, and BlockTypes do, too - since we do not instantiate implicitly! - this is however no longer true for subtyping).

[marzipankaiser]

@b-studios Now, the StdlibLLVMTests pass for me (when merging fix/boxing-nothing with refactor/stdlib).

[b-studios]

Can you add the new minified example to the hotfix?

[b-studios]

This suggests to me that using PolymorphismBoxing to also deal with Nothing and Any was maybe not a good idea.

[b-studios]

I merged it into #449 and it indeed resolves the problems with the LLVM backend there. Let's keep boxing on our list and try to clean this up in the future.