ICE bootstrapping rustc with NLL enabled

[nikomatsakis]

When I try to bootstrap rustc with NLL enabled, I get an odd ICE compiling libsyntax_ext:

error: internal compiler error: broken MIR in NodeId(16117) (_368 = move _369): bad assignment (for<'cx, 'r, 's> fn(&'cx mut syntax::ext::base::ExtCtxt<'r>, syntax_pos::Span, &'s [syntax::tokenstream::To\
kenTree]) -> std::boxed::Box<syntax::ext::base::MacResult + 'cx> = for<'r, 's, 't0> fn(&'r mut syntax::ext::base::ExtCtxt<'s>, syntax_pos::Span, &'t0 [syntax::tokenstream::TokenTree]) -> std::boxed::Box<\
syntax::ext::base::MacResult>): RegionsOverlyPolymorphic(BrNamed(crate11:DefIndex(1:1012), 'cx), '_#729r)
   --> src/libsyntax_ext/lib.rs:65:44
    |
65  |                           expander: Box::new($f as MacroExpanderFn),
    |                                              ^^^^^^^^^^^^^^^^^^^^^
...
92  | /     register! {
93  | |         line: expand_line,
94  | |         __rust_unstable_column: expand_column_gated,
95  | |         column: expand_column,
...   |
112 | |         compile_error: compile_error::expand_compile_error,
113 | |     }
    | |_____- in this macro invocation

[Aaron1011]

I'd like to work on this.

[nikomatsakis]

@Aaron1011 any progress? want me to take a look at all?

[Aaron1011]

@nikomatsakis: I haven't yet been able to reproduce this, due to #48129 stopping compilation. I'll try to work around that issue locally, and see if I can reproduce + investigate the original issue.

[Aaron1011]

@nikomatsakis: I ran across another issue while trying to reach the ICE: #48224

[Aaron1011]

@nikomatsakis: Until the intermediate issues are resolved, I don't think I'll be able to reproduce this.

[nikomatsakis]

@Aaron1011 this branch #48372 clears the way

[nikomatsakis]

Hmm, but on that branch I am now encountering the AddAssign problems I guess. So we're still blocked on #48129

[nikomatsakis]

Actually, the issue-48071 branch on my repo reproduces the problem. You have to build like so:

RUSTFLAGS_STAGE_NOT_0='-Znll -Zborrowck=mir -Ztwo-phase-borrows' ./x.py build --stage 2 

[nikomatsakis]

Error that I get is (lightly edited):

error: internal compiler error: broken MIR in NodeId(16117) (_368 = move _369): bad assignment (

for<'cx, 'r, 's> fn(&'cx mut syntax::ext::base::ExtCtxt<'r>, syntax_pos::Span, &'s [syntax::tokenstream::TokenTree]) -> std::boxed::Box<syntax::ext::base::MacResult + 'cx>
for<'r, 's, 't0> fn(&'r mut syntax::ext::base::ExtCtxt<'s>, syntax_pos::Span, &'t0 [syntax::tokenstream::TokenTree]) -> std::boxed::Box<syntax::ext::base::MacResult>): 

RegionsOverlyPolymorphic(BrNamed(crate11:DefIndex(1:1012), 'cx), '_#729r)

Removing explicit paths:

for<'cx, 'r, 's> fn(&'cx mut ExtCtxt<'r>, Span, &'s [TokenTree]) -> Box<MacResult + 'cx>
for<'r, 's, 't0> fn(&'r mut ExtCtxt<'s>, Span, &'t0 [TokenTree]) -> Box<MacResult>)

The problem seems to be the return type.

[nikomatsakis]

Minimized:

#![allow(warnings)]
#![feature(dyn_trait)]
#![feature(nll)]

trait Foo {
    fn foo(&self) { }
}

impl Foo for () {
}

type MakeFooFn = for<'a> fn(&'a u8) -> Box<dyn Foo + 'a>;

fn make_foo(x: &u8) -> Box<dyn Foo + 'static> {
    Box::new(())
}

fn main() {
    let x: MakeFooFn = make_foo as MakeFooFn;
}

[nikomatsakis]

I think the problem is that there is a missing coercion of sorts, or else that this code in the NLL type checker is too strict:

rust/src/librustc_mir/borrow_check/nll/type_check/mod.rs

Lines 1308 to 1330 in 27a046e

	CastKind::ReifyFnPointer => {
	let fn_sig = op.ty(mir, tcx).fn_sig(tcx);
	// The type that we see in the fcx is like
	// `foo::<'a, 'b>`, where `foo` is the path to a
	// function definition. When we extract the
	// signature, it comes from the `fn_sig` query,
	// and hence may contain unnormalized results.
	let fn_sig = self.normalize(&fn_sig, location);
	let ty_fn_ptr_from = tcx.mk_fn_ptr(fn_sig);
	if let Err(terr) = self.eq_types(ty_fn_ptr_from, ty, location.at_self()) {
	span_mirbug!(
	self,
	rvalue,
	"equating {:?} with {:?} yields {:?}",
	ty_fn_ptr_from,
	ty,
	terr
	);
	}
	}

In particular, it constructs the "reified" form of the fn pointer in ty_fn_ptr_from and then requires that the target type be equal to that. But in this case, the target type is a subtype of that.

[nikomatsakis]

(Hmm, I have a vague memory of solving a similar problem somewhere else, where we were rolling multiple coercions into one. It's sort of debatable what's right here, I suppose -- probably we should change the NLL type checker to use sub_types.)

[nikomatsakis]

OK well that proposed change didn't actually fix anything =)

[nikomatsakis]

Oh, interesting. The problem, I believe, is #33684 -- which @sgrif and I have been working on a PR to fix, independently.

[nikomatsakis]

Tagging as NLL-deferred -- this is blocked on @sgrif's work at the moment and there exists a workaround for within rustc.

[nikomatsakis]

Hmm. I'm going to mark this as high priority because it blocks bootstrap. I think it's a bit tricky to fix. While bike riding recently though I had some ideas here -- I realized that we could probably write a custom bit of subtyping code used only (initially) by the MIR type checker. The hope would be that this is faster than canonicalizing etc. This code could use a universe-based check. This would sidestep the limitations of our existing subtyping algorithm which are biting us here. It would also help eliminate a blocker for chalk integration later. I may take a stab at this.

[nikomatsakis]

OK, my branch is now working and passing tests. I want to clean it up some and perhaps try to improve its performance though.

[pnkfelix]

visited for triage. this should be fixed via PR #52488 (which is WIP but hopefully will be ready soon).