r? @davidtwco

rustbot has assigned @davidtwco.
They will have a look at your PR within the next two weeks and either review your PR or reassign to another reviewer.

Use r? to explicitly pick a reviewer

Can you please add a regression test for this

Can you please add a regression test for this

Right now, tests don't create real ICEs as it's being complex to write such test. But I have wrote two test files who documents the issue.

Ah, there is no MCVE... Honestly, I'm not sure if just removing debug_asserts is enough to fix this. Have you found a root cause of this issue or something?

This is the wrong way to approach this. MIR inlining is an optimization, so it's optional. Those debug-assertions ensure that we can reason about what codegen does, which helps soundness, so it's mandatory.

Could you investigate disabling instances of MIR inlining that would trigger this bug?

Ah, there is no MCVE... Honestly, I'm not sure if just removing debug_asserts is enough to fix this. Have you found a root cause of this issue or something?

With complex type projections like <Child<'c, S, T> as MaybeTotalScope>::MaybeTotalTimestamp , the compiler needs to normalize this to S::MaybeTotalTimestamp but during aggressive inlining, the type normalization happens during codegen (late in compilation) and then normalization fails → SignatureMismatch error and then ICE occurs.

Could you investigate disabling instances of MIR inlining that would trigger this bug?

It will return Unimplemented that prevents MIR inlining - when codegen can't resolve the instance, the inliner skips it.

When local is backend-ref (too big), we must use Memory. Earlier it assumed this never happens after field projection, but it can with multi-variant enums, so here second ice occurs.