mGCA: error on region vars lowering

[Human9000-bit]

This also needs to error on min_generic_const_args feature

cc @BoxyUwU (didn't r her as she is unavailable rn)

Fixes #157147

[rustbot]

HIR ty lowering was modified

cc @fmease

[rustbot]

Failed to set assignee to her: invalid assignee

Note: Only org members with at least the repository "read" role, users with write permissions, or people who have commented on the PR may be assigned.

[Human9000-bit]

@rustbot reroll

[reddevilmidzy]

And if you wrap it in backticks, it probably won't be recognized as a command.😄

View changes since this review

[BoxyUwU]

r? BoxyUwU

[BoxyUwU]

I expect that a similar case should be possible without min_generic_const_args, instead using generic_const_parameter_types e.g. something like Foo<'a, const N: &'a ()> and then Foo<'_, { &() }> . In general I don't see a problem extending this fix to happen even on stable as it should never be meaningfully reachable (though it would be good to have a test case where it actually matters)

[Human9000-bit]

Is that what you meant?

I found a case with generic_const_parameter_types, but it seems that there is no obvious case to trigger the ICE without const generics features
(lowk used ai to find gcpt case)

[Human9000-bit]

it seems that that clanker ai hasn't checked that it triggers ICE even on non-incremental. Is there a problem with that?

EDIT: ofc there is. Removed the gcpt case

[BoxyUwU]

I actually am somewhat confused here, I thought that the has_free_regions check would trigger if there are region variables in the expected type. For example the following does error:

#![feature(generic_const_parameter_types, adt_const_params, const_param_ty_unchecked)]

fn evil<T, const R: T>() {}

fn foo<const N: usize>() {
    evil::<&u32, { &1_u32 }>();
}

what's the difference in expected types of the { &1_u32 } anon const in this example, and the const { &0_u8 } anon const in the mgca example? I would expect both of these to ICE not just the mGCA one

[BoxyUwU]

@rustbot author

[Human9000-bit]

I would expect both of these to ICE not just the mGCA one

Your example is correct in terms of it containing region var, but it being caught by existing check:

rust/compiler/rustc_hir_analysis/src/hir_ty_lowering/mod.rs

Lines 2313 to 2314 in 4f8ea98

	if tcx.features().generic_const_parameter_types()
	&& (ty.has_free_regions() || ty.has_erased_regions())

, because gcpt enabled and const having free regions. Does that make sense?
@rustbot ready

[BoxyUwU]

hmm, what I meant is why does the ICEing example not get caught by that existing check. it too should have free regions in the type. i got confused before though and I realise now it's jsut because the check for this stuff only happens under gcpt but mgca also allows for gcpts implicitly.

I also realise now there's operator precedence jank at play in this diff 😅 :

if tcx.features().generic_const_parameter_types()
    && (ty.has_free_regions() || ty.has_erased_regions())
    || ty.has_infer_regions()
{

is actually

if (tcx.features().generic_const_parameter_types()
    && (ty.has_free_regions() || ty.has_erased_regions()))
    || ty.has_infer_regions()
{

I think your original diff was fine- making it be ``if (mgca || gcpt) && (has_free_regions || has_erased_regions)` seems correct enough to me 🤔