ICE when trying to match on non-PartialEq slice. #61188
Comments
matthewjasper
added
the
I-ICE
jonas-schievink
added
C-bug
|
CC #53708 |
pnkfelix
added
the
A-MIR
|
Ah ha! If you add Which may be something a user might infer from the message ...
... but probably not 🤣 |
|
triage: P-high; removing nomination. I want to know why we are not catching the failure to fulfill |
|
I suspect EDIT: I should've read more carefully, looks like |
|
assigning to self with intent to try to mentor. |
jonas-schievink
added
the
E-mentor
|
I started looking at this, with the intent of writing mentorship notes. I'm going to try to jot down notes on my process as I go, with the intent of showing a mentee what kinds of experiments one might do to figure out how to narrow down the code to the point of interest for finding a fix. To be clear, I don't consider myself an expert in the MIR generation code; I've done some work in it, but I am at a point where I need to reacquaint myself with it before I'd feel comfortable hacking on it. So, here's the steps I've done so far:
Look at that comment (added back in df3de7b, many thanks @matthewjasper!); the fact that you see three types that really can be hard-coded into the compiler, and then one special case of That's where I decided I wanted to start taking some notes, since arguably I was supposed to be making mentorship instructions for this, but if I get further into the weeds without documenting the context, I won't be able to make sensible instructions. |
|
To be clear, the reason that I saw the Either:
|
|
And reading over rust-lang/rfcs#1445 led me to find that the current code (linked below) for linting use of rust/src/librustc_mir/hair/pattern/mod.rs Lines 988 to 1012 in ef064d2 |
|
Filed #62307 to track the insufficient check for |
|
(the change suggested in PR #62339 fixes this and has a regression test for it.) |
…l-match-check, r=nikomatsakis use visitor for #[structural_match] check This changes the code so that we recur down the structure of a type of a const (rather than just inspecting at a shallow one or two levels) when we are looking to see if it has an ADT that did not derive `PartialEq` and `Eq`. Fix #61188 Fix #62307 Cc #62336
|
Reopening because there are still cases where we ICE: #![deny(indirect_structural_match)]
#[derive(PartialEq, Eq)]
enum O<T> {
Some(*const T), // Can also use PhantomData<T>
None,
}
struct B;
const C: &[O<B>] = &[O::None];
pub fn foo() {
let x = O::None;
match &[x][..] {
C => (),
_ => (),
}
}We might want to move to a trait based approach as in #63248 (comment) |
|
visiting for triage. I agree with @matthewjasper about potentially moving to a trait-based approach. part of me is tempted to jump into implementing that. But it probably makes more sense for me to write up mentorship instructions, maybe. |
pnkfelix
added
E-needs-mentor
and removed
E-mentor
pnkfelix
added
E-mentor
|
triage: this is probably easy to resolve but I have not had time myself to write up mentorship instructions. Marking bug accordingly (i.e. "need-mentor" is to be interpreted as "if you are willing to help make some instructions, I'd love it...") |
|
Note: moving to a trait-based approach alone won't resolve every ICE. Not without regressing other stable code, least. In particular:
I discovered this while exploring cases related to #63438 and #63479 Here is a concrete example (play): fn assert_partialeq<T: PartialEq>(_: &T) {}
fn rquux(_: &()) { }
const RQ: fn(&()) = rquux;
const LRQ: &[fn(&())] = &[];
fn main() {
// assert_partialeq(&RQ); (RQ isn't PartialEq, and this will say so.)
match RQ {
RQ => println!("RQ"),
_ => {}
};
let input: &[fn(&())] = &[][..];
match input {
LRQ => println!("LRQ"), // compiles if you comment this out
&[] => println!("matched inlined pattern"),
_ => {}
};
}(The "obvious solution" of just requiring const patterns to always implement |
…ructural` trait. As far as I can tell this preserves previous observable stable behavior, *including* the ICE from rust-lang#61188. (In other words, further work is necessary to reap benefits of switching to a trait.)
…ewjasper's example. specific example is from [comment][]: [comment]: rust-lang#61188 (comment) ```rust #![deny(indirect_structural_match)] #[derive(PartialEq, Eq)] enum O<T> { Some(*const T), // Can also use PhantomData<T> None, } struct B; const C: &[O<B>] = &[O::None]; pub fn foo() { let x = O::None; match &[x][..] { C => (), _ => (), } } ```
|
(closing this bug so we can let issue #65466 be dedicated to tracking the more subtle problem laid out in @matthewjasper 's comment above.) |
The following code will result in an ICE in
codegenMIR construction:playground
Error and Backtrace
The text was updated successfully, but these errors were encountered: