Re: What's wrong with C++?

["Dmitry A. Kazakov" <mailbox@dmitry-kazakov.de>]

On Thu, 17 Feb 2011 12:27:11 -0500, Peter C. Chapin wrote:

> On Thu, 17 Feb 2011, Dmitry A. Kazakov wrote:
> 
>> The second note is that if types are not the things from the problem space 
>> (of which the compiler knows little or nothing), then what kind of stuff 
>> is there in their place? I presume that nothing.
> 
> You can still define the types you need. Inference just means that you 
> typically don't have to bother with explicit declarations: create the 
> objects you need (of the types you need) and let the compiler work out the 
> details.

How the compiler may know that 1.7 is magnetic flux? I don't mean if the
compiler derive type constraints:

   Text : String := "abcd";

That is OK to me.

>> Now the main objection is computability and decidability. Only "simple" 
>> things can be inferred. The compiler cannot solve problems beyond 
>> Turing-completeness...
> 
> It sounds like you are taking "type inference" to mean that the compiler 
> figures out pre-conditions, post-conditions, etc. I'm just talking about the 
> compiler recovering type information based on how entities are used.

I doubt this were possible in a rich type system where types are
consistently cloned. E.g. in Ada:

   type My_Integer is new Integer;

The properties of both types are same. The above is a typical case of
reuse. It seems that reuse is incompatible with inference. There is nothing
that tells us that two things are indeed same or not. Maybe the same would
be inferred from both, maybe not, how do I know?

BTW, inference is incompatible with information hiding and modularity. In
order to infer you have to lay everything open, not a good software
engineering practice. Even if you say that inferring were optional. Let
some part of the system rely on that "optional" inference. That
automatically makes it mandatory. What happens if in the process of
maintenance inferring stop working because of small incremental changes?

> The algorithms for doing this are well understood and they are decidable.

Maybe, like pattern matching is, decidable but incomprehensive. It is
impossible for the reader to guess the language matched by given pattern.
The program shall be understandably by small incremental efforts on the
reader side. Non-trivial inference requires qualitative jumps, either you
can infer what the compiler does or not.

-- 
Regards,
Dmitry A. Kazakov
http://www.dmitry-kazakov.de