Rustc doesn't unify inference variables with subtype bounds:

https://github.com/rust-lang/rust/blob/b776d1c3e3db8befabb123ebb1e46c3531eaed46/compiler/rustc_infer/src/infer/sub.rs#L87-L112

Looking at unification, the only difference here with rustc in this area is that {int_var} <: {ty_var} is turned into {int_var} = {ty_var} (and likewise with floats, and with the subtype relation reversed).

Looking at unification, the only difference here with rustc in this area is that {int_var} <: {ty_var} is turned into {int_var} = {ty_var} (and likewise with floats, and with the subtype relation reversed).

Ugh yeah, that's what I was missing and what the proper change is.

Actually, that alone is not enough to make the test pass :/

I'm not sure if the test case with two arbitrary variables needs to pass. One of the variables will eventually be resolved if the program is going to compile.

I wondered about that. I did test with the original test from @flodiebold with the int T and that didn't pass.

I don't love this solution. Like @matthewjasper, I'm not sure that the test with two arbitrary variables needs to pass. @jackh726 can you share what you tried with integer variables? In rustc, we definitely consider integer variables "more specific" than normal variables (so e.g. a general variable would be "assigned" the value of an integer variable).

Yeah, I updated.

So, to elaborate on why only always unifying non-general inference vars doesn't work:
First, we start off trying to unify our goal Implemented(MyClosure<"rust" for<0> [?0 := (&'static ?0), ?1 := 0]>: FnOnce<1<(&'static ?1i)>>) with our (instantiated) clause Implemented(MyClosure<"rust" for<0> [?0 := ?2, ?1 := 0]>: FnOnce<1<?2>>).

From there, we try to unify MyClosure<"rust" for<0> [?0 := (&'static ?0), ?1 := 0]> and MyClosure<"rust" for<0> [?0 := ?2, ?1 := 0]> first. When ?2 tries to get unified with (&'static ?0), it is generalized and set to (&'?3 ?4). From the next unification, we get ?4 < ?0 (and the reverse too, since both Covariant and Contravariant, since we're in Invariant).

By this point, when we unify 1<(&'static ?1i)> and 1<?2>, it's too late, and we've already added the subtype goals.

Okay, so we probably just need to actually resolve inference vars when trying to simplify a subtype goal...

Okay so that worked.

oops, I forgot, we use bors now don't we... d'oh