rustc: Disallow autoderef for bottom type !

[mdup]

Fix issue #13847.

[alexcrichton]

Thanks for looking into this! While this does indeed fix the ICE from appearing, it may not be the desired behavior that we want. In most other use cases the bottom type will unify and work with all other types. For example, if one branch of an if expression has the bottom type, the other branch can have whatever type it wants.

The desired semantics here may be to "just work" and continue to propagate the bottom type outwards, but this is definitely another alternative! It may not matter too much in the long run because this seems pretty rare.

[lilyball]

I agree, propagating bottom seems like the more correct approach.

[nikomatsakis]

cc me

[nikomatsakis]

I am not sure if propagating bottom is good in this case. It makes sense in (some) other cases, because the semantics of what code would execute are not dependent on what type the ! value winds up being. This is not the case for method calls or (similarly) generic functions.

[nikomatsakis]

To be honest, I actually wonder if we should include bottom in the type system at all. it is not particularly like other types. I think I would be happier overall if we just assigned the "type" ! to a fresh inference variable (but still used the information about what must fail when doing dataflow). Hmm.

[mdup]

On a short term basis, my patch fixes the ICE, so that might be desirable. On the other hand, it doesn't explain why the assumptions in middle/typeck/check/method.rs:confirm_candidate() are not satisfied (hence the bug() call). It may be a smell of some other failure not discovered yet, so I guess it needs to be addressed for the future.

[lilyball]

@nikomatsakis By "a fresh inference variable" do you mean that an expression typed as ! would just become an unconstrained variable that can end up unifying with another type (e.g. so ! could end up being inferred as int)?

[nikomatsakis]

On Mon, Jun 02, 2014 at 01:05:44PM -0700, Kevin Ballard wrote:

@nikomatsakis By "a fresh inference variable" do you mean that an expression typed as ! would just become an unconstrained variable that can end up unifying with another type (e.g. so ! could end up being inferred as int)?

Yes. This is what ML compilers do, afaik.

[alexcrichton]

Closing due to inactivity, but feel free to reopen with a rebase!