[erlang-questions] dialyzer/typer and exceptions

[Daniel Goertzen]

Thank you Kostis.  I studied Haskell some time ago so I guess I was trying
to make Erlang match up with that.  I need to let go of that and recognize
Erlang is different.

Dan.


On Sat, Oct 2, 2010 at 3:36 AM, Kostis Sagonas <kostis@REDACTED> wrote:

> Daniel Goertzen wrote:
>
>> While experimenting with dialyzer and typer, I discovered that "catch"
>> acts
>> as type inference barrier.
>>
>
> Not sure what you mean by "type inference barrier".
>
> What the type inference algorithm does is to approximate the types (i.e.,
> set of terms) for which a function will "succeed" (i.e., not raise an
> exception). What 'catch' does is it catches exceptions and thus allows
> functions to be used with more terms.
>
> For example, while the function:
>
>  foo(X) -> X + 42.
>
> will return a value only for numbers and thus has type signature
>
>  -spec foo(number()) -> number().
>
> the function:
>
>  bar() -> catch (X + 22).
>
> will return some value for all terms, not only for numbers, so it has the
> type signature:
>
>  -spec bar(any()) -> any().
>
>
>   As an example, gen_server:call/2 uses catch in
>> its first line and the resulting inferred type signature is:
>>
>> -spec call(_,_) -> any().
>>
>> Is there a way to get dialyzer/typer to ignore catches and assume the
>> non-exception case, or is that opening pandora's box?
>>
>
> The answer to your question is "of course there is such a way", but the
> question then is what exactly would this type inference be computing. The
> current one tries to model the operational semantics of Erlang, not those of
> some ideal(ized) language.
>
>
>  Also, are there any other constructs that are known to confound
>> dialyzer/typer?
>>
>
> I think confound is a very strong term here, but the answer is that catch
> is the only such construct.  What also destroys type information is
> catch-all clauses/cases.
>
> Independently of type inference though catch is a terrible construct and
> should be avoided in modern Erlang; try-catch is a much better construct for
> various reasons and dialyzer handles this much better.
>
> Kostis
>