-
Notifications
You must be signed in to change notification settings - Fork 12.7k
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
minor *dyn
cast cleanup
#131898
minor *dyn
cast cleanup
#131898
Conversation
- remove a redundant check, because we always emit the "better diagnostic" now - clean up the comments a bit
@bors r+ rollup |
// Since ptr-to-ptr casts, unlike unsizing coercions, cannot change | ||
// the pointer metadata, dropping a the principal in a (non-coercion) | ||
// cast is currently not allowed. |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Why can they not change metadata?
And why would you need to change it anyway? Any vtable is valid for auto traits.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
=This may have just been an oversight of mine when authoring the dyn Trait -> dyn Auto
PR.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Why can they not change metadata?
It's not that they inherently cannot, they just do not as they are currently implemented and would need to be changed for this to work.
And why would you need to change it anyway? Any vtable is valid for auto traits.
Because our UB checks currently require the vtable to match exactly. That's what the sentence above is trying to explain.
Currently, this is considered UB and Miri will report it as such:
#![feature(core_intrinsics, custom_mir)]
trait Trait {}
impl Trait for () {}
#[custom_mir(dialect = "built")]
fn cast(ptr: *const (dyn Trait + Send)) -> *const dyn Send {
use std::intrinsics::mir::*;
mir! {
{
RET = CastPtrToPtr(ptr);
Return()
}
}
}
fn main() {
let ptr: *const (dyn Trait + Send) = &();
let ptr = cast(ptr);
}
error: Undefined Behavior: constructing invalid value: wrong trait in wide pointer vtable: expected `std::marker::Send`, but encountered `Trait + std::marker::Send`
--> src/main.rs:11:13
|
11 | RET = CastPtrToPtr(ptr);
| ^^^^^^^^^^^^^^^^^^^^^^^ constructing invalid value: wrong trait in wide pointer vtable: expected `std::marker::Send`, but encountered `Trait + std::marker::Send`
|
So our options here are
- changing the opsem of a PtrToPtr cast in MIR
- lowering this HIR cast to something other than a single MIR cast
- declaring that the vtable must no longer match exactly (allowing
dyn Trait
vtables in*const dyn Auto
) and adjusting the UB checks accordingly- this would also mean adjusting the
*const (dyn Trait + Auto)
to*const dyn Auto
upcast coercions to no longer change the vtable (which they currently semantically do, even though it gets codegenned to a no-op)
- this would also mean adjusting the
Either way, this comment is trying to explain that we need to do more than just change this to return Ok(CastKind::PtrPtrCast)
here.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
=This may have just been an oversight of mine when authoring the
dyn Trait -> dyn Auto
PR.
@compiler-errors I'm not following, what oversight and PR are you talking about here?
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
I'd say that we should relax the requirement. Either by allowing empty traits (!has_own_existential_vtable_entries
) to have any vtable, or by making it actually compare vtables (something like own_existential_vtable_entries_iter(actual).chain(repeat(Vacant)).zip(own_existential_vtable_entries_iter(expected)).all(|a, b| a == b)
)..
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
I would say that this is fine, while both casts guarantee to return a pointer with a valid vtable for dyn Send
, there is no guarantee that it will or will not be valid for anything else.
But also, the upcast coercions don't change the vtable either.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
But also, the upcast coercions don't change the vtable either.
They do in const eval / miri where this is currently considered UB:
trait Trait {}
impl Trait for () {}
fn main() {
let ptr: *const (dyn Trait + Send) = &();
let ptr2: *const dyn Send = ptr as *const dyn Send;
let _: *const (dyn Trait + Send) = unsafe { std::mem::transmute(ptr2) };
}
error: Undefined Behavior: constructing invalid value: wrong trait in wide pointer vtable: expected `Trait + std::marker::Send`, but encountered `std::marker::Send`
--> src/main.rs:8:49
|
8 | let _: *const (dyn Trait + Send) = unsafe { std::mem::transmute(ptr2) };
| ^^^^^^^^^^^^^^^^^^^^^^^^^ constructing invalid value: wrong trait in wide pointer vtable: expected `Trait + std::marker::Send`, but encountered `std::marker::Send`
|
(you need -Zextra-const-ub-checks
to trigger the check in CTFE)
there is no guarantee that it will or will not be valid
For the sake of guarantees for stable code I think it makes sense to leave it unspecified whether this transmute is valid.
But for someone who writes unsafe code and tests it with miri I think it would be very confusing that the first case is UB, but the second case is not -- both of these are an as
cast followed by a transmute, nothing here makes it obvious that the as
cast actually does something different for each case.
Also we don't need to find a solution here: In this PR I'm just trying to add a comment to explain that there is a problem that needs to be solved before we can allow dropping the principal in non-coercion casts, and that allowing these casts won't be a one-line fix.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
What's a "non-coercion cast"? I assumed that foo as type
will trigger coercions as required to perform the cast.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Yes, a T as U
cast will always trigger a coercion if possible. With "non-coercion" cast, I mean a cast that did not trigger a coercion, i.e. the case where it's not possible to coerce from T
to U
. Ferrocence calls these "specialized casts", but I don't think we have an official term for them?
@bors r- because the comment needs more clarification |
d9eb6d8
to
46cc5e9
Compare
I think it's fine that we've documented the status quo. @bors r+ |
Rollup of 10 pull requests Successful merges: - rust-lang#130225 (Rename Receiver -> LegacyReceiver) - rust-lang#131169 (Fix `target_vendor` in QNX Neutrino targets) - rust-lang#131623 (misc cleanups) - rust-lang#131756 (Deeply normalize `TypeTrace` when reporting type error in new solver) - rust-lang#131898 (minor `*dyn` cast cleanup) - rust-lang#131909 (Prevent overflowing enum cast from ICEing) - rust-lang#131930 (Don't allow test revisions that conflict with built in cfgs) - rust-lang#131956 (coverage: Pass coverage mappings to LLVM as separate structs) - rust-lang#132076 (HashStable for rustc_feature::Features: stop hashing compile-time constant) - rust-lang#132088 (Print safety correctly in extern static items) r? `@ghost` `@rustbot` modify labels: rollup
Rollup merge of rust-lang#131898 - lukas-code:ptr-cast-cleanup, r=compiler-errors minor `*dyn` cast cleanup Small follow-up to rust-lang#130234 to remove a redundant check and clean up comments. No functional changes. Also, explain why casts cannot drop the principal even though coercions can, and add a test because apparently we didn't have one already. r? `@WaffleLapkin` or `@compiler-errors`
Also see rust-lang/reference#1661 |
Small follow-up to #130234 to remove a redundant check and clean up comments. No functional changes.
Also, explain why casts cannot drop the principal even though coercions can, and add a test because apparently we didn't have one already.
r? @WaffleLapkin or @compiler-errors