Warnings are not reported if there are errors

[paulyoung]

I merged master (at 9cb2c23) into paulyoung/language-server/#461 and started seeing the following issue:

It's sort of the opposite to the issue I reported in #440.

[paulyoung]

Just to verify that it's not related to the LSP work, here it is from the command line:

λ cat Warnings.as && result/bin/asc -c Warnings.as
let foo = [1, true, ""];
let bar = if true 1 else "2";
// let baz = qux;

Warnings.as:1.11-1.24: warning, this array has type [Any] because elements have inconsistent types
Warnings.as:2.11-2.29: warning, this if has type Any because branches have inconsistent types,
true produces
  Nat
false produces
  Text

λ cat Warnings.as && result/bin/asc -c Warnings.as
let foo = [1, true, ""];
let bar = if true 1 else "2";
let baz = qux;

Warnings.as:3.11-3.14: type error, unbound variable qux

[paulyoung]

Same thing with --check instead of -c.

[nomeata]

I don’t think we can solve this in all generality. The type checker usually just aborts if there is an error, to avoid further touching bad code. There are specific points where we allow the type checker to continue in a different part of the code (so I assume if the warning and the error are happening in separate blocks, both will be returned). But in this case, it can’t even know the type of baz, so it will not look at the recursive block at at all. (Remember that all definitions are mutually recursive.)

[nomeata]

See:

~/dfinity/actorscript/src $ cat warnings.as ; echo; ./asc -c warnings.as 
let foo = [1, true, ""];
let bar = if true 1 else "2";
let baz = qux;

warnings.as:3.11-3.14: type error, unbound variable qux

~/dfinity/actorscript/src $ cat warnings.as ; echo; ./asc -c warnings.as 
{
let foo = [1, true, ""];
let bar = if true 1 else "2";
};
{
let baz = qux;
};

warnings.as:2.11-2.24: warning, this array has type [Any] because elements have inconsistent types
warnings.as:3.11-3.29: warning, this if has type Any because branches have inconsistent types,
true produces
  Nat
false produces
  Text
warnings.as:6.11-6.14: type error, unbound variable qux

I think we can close this, as in its full generality this is unfixable (but of course still try to make the type checker recover from errors where possible).

[nomeata]

Hmm. We can actually resolve names before type checking. So we could do a separate name resolution pass (like GHC does, called the renamer there) and identify the SCC of the mutual groups, and then do type-checking separately on them. This would contain unrecoverable errors much more, and maybe even improve type inference. But it would be a major refactoring, so probably not worth right now.