Apparent regression in implied bounds

[nikomatsakis]

In the RFC 1214 branch, I found that at some point after a rebase (where a new snapshot was introduced), a new implied bound was needed. I believe this is specific to stage0, but I am not sure. I didn't have time to minimize or dig deeply, so i've marked the offending bound for now with a FIXME.

[nikomatsakis]

OK, so, stranger and stranger. Now that we have a new stage0, this where clause that was previously required seems...to lead to an error? If I remove it, though, things work. Bizarre.

[nikomatsakis]

Seems like I should put some effort into trying to narrow down the test case better.

[nikomatsakis]

This test seems to fail.

use std::cell::Cell;
use std::marker::PhantomData;

pub trait Delegate<'tcx> { }

pub struct InferCtxt<'a, 'tcx: 'a> {
    x: PhantomData<&'a Cell<&'tcx ()>>
}

pub struct MemCategorizationContext<'t, 'a: 't, 'tcx : 'a> {
    x: &'t InferCtxt<'a, 'tcx>,
}

pub struct ExprUseVisitor<'d, 't, 'a: 't, 'tcx:'a+'d> {
    typer: &'t InferCtxt<'a, 'tcx>,
    mc: MemCategorizationContext<'t, 'a, 'tcx>,
    delegate: &'d mut (Delegate<'tcx>+'d),
}

impl<'d,'t,'a,'tcx> ExprUseVisitor<'d,'t,'a,'tcx> {
    pub fn new(delegate: &'d mut Delegate<'tcx>,
               typer: &'t InferCtxt<'a, 'tcx>)
               -> ExprUseVisitor<'d,'t,'a,'tcx>
    {
        ExprUseVisitor {
            typer: typer,
            mc: MemCategorizationContext::new(typer),
            delegate: delegate,
        }
    }
}

impl<'t, 'a,'tcx> MemCategorizationContext<'t, 'a, 'tcx> {
    pub fn new(typer: &'t InferCtxt<'a, 'tcx>) -> MemCategorizationContext<'t, 'a, 'tcx> {
        MemCategorizationContext { x: typer }
    }
}

fn main() { }

[nikomatsakis]

This MAY be related to object defaults -- if I change &'d mut Delegate<'tcx> to &'d mut (Delegate<'tcx>+'d), it works.

[nikomatsakis]

OK, this is definitely some kind of wacky node-id confusion. If I change &'d mut Delegate<'tcx> to &'d mut (Delegate<'tcx>), the test passes!

[nikomatsakis]

OK, I found the problem. The problem is that the lub_free_regions code is really not very smart. In particular, for some 'a and 'b pair, it considers where 'a <= 'b or 'b <= 'a and otherwise falls back to 'static. It does not consider that there might be some other 'c such that 'a <= 'c and 'b <= 'c (in this case, 'tcx is such a region).

[nikomatsakis]

I believe that adding parens makes things work simply because it permutes the ordering of various operations in region inference, causing slightly different LUB operations to arise in which the right variables are seen at the right time. I haven't fully verified this.

[nikomatsakis]

(And, of course, we are not guaranteed a "best" answer here, so this is another case where region inference is (currently) bound to be somewhat inadequate.)