ICE with mutually recursive infinite-sized type and enum from another module

[XavilPergis]

I'm getting a very touchy ICE on rustc 1.33.0-nightly (9eac38634 2018-12-31) running on x86_64-unknown-linux-gnu. I'm not 100% sure what it has to do with...

Here is the minimum example I can trigger the ICE with:

// in `crate::signature` / `src/signature.rs`
use crate::field::BaseType;

pub enum ObjectType {
    Class(ClassTypeSignature),
    Array(TypeSignature),
    TypeVariable(()),
}

pub struct ClassTypeSignature {
    pub package: (),
    pub class: (),
    pub inner: (),
}

pub enum TypeSignature {
    Base(BaseType),
    Object(ObjectType),
}

// in `crate::field` / `src/field.rs`
#[derive(Copy, Clone, Debug, Eq, PartialEq, Hash)]
pub enum BaseType {
    Byte,
    Char,
    Double,
    Float,
    Int,
    Long,
    Short,
    Boolean,
}

I'm expecting to see an error about an infinitely-sized type, but instead the compiler crashes. Changes that make the compiler not ICE:

• Moving field::BaseType into signature
• Making ObjectType/TypeSignature not mutually recursive
• Rearranging or removing any field/variant from any of these structs/enums
• Making ObjectType::Class not contain ClassTypeSignature

Output of RUST_BACKTRACE=1 cargo build: https://gist.github.com/XavilPergis/102774c6c542403db10a351ff013f356
Original code that ICE'd: https://gist.github.com/XavilPergis/dafc2bb478f18cee700f78e16eea81cd

[estebank]

The error is raised here because it is expected that an error should have been emitted before reaching this code branch, presumably by force_from_dep_node:

rust/src/librustc/dep_graph/graph.rs

Lines 632 to 642 in ec19464

	None => {
	if !tcx.sess.has_errors() {
	bug!("try_mark_green() - Forcing the DepNode \
	should have set its color")
	} else {
	// If the query we just forced has resulted
	// in some kind of compilation error, we
	// don't expect that the corresponding
	// dep-node color has been updated.
	}
	}

It seems that the check was added in #49088 to get around similar errors. I wonder if completely removing the check (ignoring the None case and returning None, like in the Red case) would cause any bugs, but looking at the commit that originally introduced this check, we used to return None; always there.

CC @michaelwoerister

[XavilPergis]

Here's the code in my repository, by the way: https://github.com/XavilPergis/rvm/blob/ice/class/src/signature.rs

[jakubadamw]

@XavilPergis, I tried to compile your ice branch with the 2018-12-31 nightly and it produced the expected errors with no ICEs. Same with the newest nightly and the stable 1.36.