cyclic trait inheritance leads to a rustc stack overflow

[nrc]

trait t1 : t2 {
}
trait t2 : t1 {
}

or even

trait t : t {
}

gives a stack overflow:

task 'rustc' has overflowed its stack
./build.sh: line 1: 22351 Illegal instruction (core dumped) /home/nick/rust2/x86_64-unknown-linux-gnu/stage1/bin/rustc $1.rs -o $1

[nrc]

I would expect an error message.

[huonw]

Seems similar to #8762 and possibly #9095.

[nrc]

Yeah, I think we need to be checking for logical cycles in a lot of places.

[nikomatsakis]

cc me and #5527

[lilyball]

cc me

[nikomatsakis]

This is related to the work on trait-reform and where clauses, actually. The immediate cause for this is that typeck's collect phase attempts to gather both the bounds and the set of type parameters in one step. It should do a first-pass that just collects, for each nominal type, what its type parameter names are. Then a second pass to gather bounds. This is natural with where clauses. The next problem will occur when gathering up "implied" trait bounds -- if we're not careful, we'll get a cycle there too, as we go from A to B and back again. But we can patch that too, both by checking for recursion and reporting an error (no matter what), but also by seeing if we've already visited one of the type-trait-pairs and avoiding recursion in that case (this would allow us to reach a fixed point).

[nikomatsakis]

I'll fix this soon.