PhantomData not behaving as advertised

[dylanede]

As far as I can tell, from the point of view of lifetimes, PhantomData<T> should behave identically to T. This is unfortunately not the case in practice.

This function causes a mismatched type error on x:

fn cast_phantom<'a, 'b, T>(x: PhantomData<&'a T>) -> PhantomData<&'b T> where 'a : 'b {
    x
}

Whereas this similar function does not:

fn cast_actual<'a, 'b, T>(x: &'a T) -> &'b T where 'a : 'b {
    x
}

The variance rules of lifetimes should make both functions valid.

[pnkfelix]

cc @nikomatsakis

[japaric]

Have you tried with master? AFAIK the current nightly (dfc5c0f Wed Feb 18 07:50:21 2015) doesn't include the variance PR (#22541 look for the roll-up merge of #22286) which landed on 522d09d Thu Feb 19 13:36:59 2015

[japaric]

I can confirm cast_phantom compiles with rustc 1.0.0-dev (522d09dfe 2015-02-19) (built 2015-02-20) (on my ARM board).

[dylanede]

@japaric Ah, I was using the playpen to check that it still behaved the same on master. I guess next time I have an issue I better actually compile from master instead. It just takes so long though.

[japaric]

FWIW, multirust is now providing nightlies (built 2015-02-20). It seems the playpen/playbot is still using (built 2015-02-18). From what I've seen multirust always provides the newest nightlies first, whereas travis-ci and the playpen tend to lag a little behind (1-2 days).