Internal Compiler Error on stable (Broken MIR) #96512
Comments
conradludgate
added
C-bug
|
I've discovered that destructuring the enum variant outside of the async block fixes the issue |
|
@conradludgate amazing that you did reduce your testcase like that. struct Variant;
enum Enum {
Variant(Variant),
}
fn spawn<T>(_: T)
where
T: std::future::Future Send 'static,
T::Output: Send 'static,
{
}
async fn dispatch(_: Variant) {}
fn main() {
let eh = Enum::Variant(Variant);
spawn(async move {
let Enum::Variant(value) = eh;
dispatch(value).await;
});
}Regressed in rollup #91555 probably in #91450 searched nightlies: from nightly-2021-12-01 to nightly-2022-04-24 bisected with cargo-bisect-rustc v0.6.0Host triple: x86_64-unknown-linux-gnu cargo bisect-rustc --end=2022-04-24 --without-cargo --access=github --regress=ice |
|
Reduced a bit more (requires an edition with precise captures, I think): struct Struct;
enum Enum { Variant(Struct) }
fn main() {
let _enum = Enum::Variant(Struct);
|| {
let Enum::Variant(_value) = _enum;
};
} |
|
If no one is yet investigating this, I'd like to pick it up |
|
So far, I've found 2 issues that cause this behaviour. The first is that it's having trouble in this context to get the field type of the single variant enum: Fix is easy, - ty::Adt(adt_def, substs) if !adt_def.is_enum() => {
ty::Adt(adt_def, substs) if adt_def.variants().len() == 1 => {The second issue is that the MIR is being badly formed. Looking at the traces, I see Which hints to me that the code is not being read as let Enum::Variant(_value) = _enum;but instead let Enum::Variant(Struct(_value)) = _enum; |
Yeah, it looks like constructed MIR is incorrect. One issue seems to be in There seems to be one more issue when constructing the capture - the downcasting operation is completely missing. While the documentation for MIR
I think that is an outdated information. |
|
I also just noticed the Downcast issue. In my traces, I do see the variant projection using this Downcast for just a single variant, and the suspect PR changes does seem to ignore all Downcast projections (there's a little bit of discussion taking place on zulip https://rust-lang.zulipchat.com/#narrow/stream/182449-t-compiler.2Fhelp/topic/reading.20MIR.20.28.2396512.29/near/280855883) |
|
Should the fix be
|
|
@arora-aman or @roxelo might be in a better position to give advice here. I think we should keep and / or add downcast where they are currently missing (e.g., when constructing the closure), and fix how projections corresponding to captured prefix are removed. As far the latter aspect is concerned, I think the issue is in the following lines ( rust/compiler/rustc_mir_build/src/build/expr/as_place.rs Lines 271 to 278 in 905fd73 |
I vaguely recall we relied on this to detect if we had a multi variant enum, in which case we would completely capture the root variable. Edit: Ignore that, scrolling into the code bit more, there is comment that Downcast can be expected for single-variant enums |
|
@conradludgate are you still planning to work on this issue? |
|
Not currently, I spent a few full days looking but I don't want to risk burning out debugging something that I don't completely understand 😓. I hope I at least narrowed it down a bit |
…-aman Fix precise field capture of univariant enums When constructing a MIR from a THIR field expression, introduce an additional downcast projection before accessing a field of an enum. When rebasing a place builder on top of a captured place, account for the fact that a single HIR enum field projection corresponds to two MIR projection elements: a downcast element and a field element. Fixes rust-lang#95271. Fixes rust-lang#96299. Fixes rust-lang#96512. Fixes rust-lang#97378. r? `@nikomatsakis` `@arora-aman`
Code
There's only so much that I can share. The code that seems to be erroring is similar to
Meta
rustc --version --verbose:Error output
Backtrace
The text was updated successfully, but these errors were encountered: