From: Damien Doligez <damien.doligez@inria.fr>
> > I believe the problem with failwith is solvable, albeit rather
> > complicated. The idea is that you want to be warned when you apply a
> > function of type (\forall 'a. 'a) to something, because no such
> > function may exist, so that this application will never actually take
> > place.
> 
> More generally, you want to output a warning whenever the computation
> of such a value is not immediately followed by a join in the control
> flow graph, because at that point you know you're compiling dead code.
> 
> Then you would also get a warning for things like this:
> 
>    failwith "foo";
>    print_string "hello world"
> 
> or
> 
>    f (a, b, failwith "foo", c, d)
> 
> etc.
> 
> Don't ask me to implement it, though.

This is not specially hard to implement case by case.
The problem is rather that the technique I use, based on type
inference, is not foolproof (you can avoid the warning with a type
annotation for instance) and is wrong in presence of Obj.magic.
So the question is in which cases having a warning is worth the
inconvenience and the extra code in the compiler.

I would say your first example is reasonable (this may be a consequence
of a dangling then), but much less the second one (where is the
ambiguity?)

Jacques Garrigue