stepping into where-clauses during normalization may be productive

[lcnr]

View all comments

fixesrust-lang/trait-system-refactor-initiative#273, see that issue for more info.

Whether stepping into a where-clause is productive depends not on whether we're proving a NormalizesTo or Trait goal, but instead on how both the impl and the cycle rely on it.

In the example in tests/ui/traits/next-solver/cycles/normalizes-to-is-not-productive-2.rs this is just a productive use given the way @Nadrieril and I are thinking about it right now.

We're changing such cycles to be ambiguous for now, so this does not commit us to anything.

This previously caused a lot of breakage when normalizing where-clauses, e.g. rust-lang/trait-system-refactor-initiative#176, however with #158643 that is no longer an issue. We will need to figure out what to do here if we want to properly fix ParamEnv normalization in the future

r? @BoxyUwU or @nikomatsakis

[BoxyUwU]

@rustbot author

pending figuring out how breaking this is

[rustbot]

Reminder, once the PR becomes ready for a review, use @rustbot ready.

[Randl]

Not sure if you're already aware but this PR ICEs on the following

//@ revisions: current next//@ ignore-compare-mode-next-solver (explicit revisions)//@[next] compile-flags: -Znext-solver//@ edition: 2021//@ compile-flags: --crate-type=lib//@ build-passuse std::future::Future;use std::pin::Pin;use std::sync::Arc;use std::task::Context;use std::task::Poll;structBuffer;typeResult<T> = std::result::Result<T,()>;typeBoxedFuture<'a,T> = Pin<Box<dynFuture<Output = T> + Send + 'a>>;traitRead:Unpin + Send{fnread(&mutself) -> implFuture<Output = Result<Buffer>>;}traitReadDyn:Unpin + Send + Sync{}typeReader = Box<dynReadDyn>;traitAccess:Send + Sync + Unpin{typeReader;fnread(&self) -> implFuture<Output = Result<(u32,Self::Reader)>> + Send;}traitAccessDyn:Send + Sync + Unpin{fnread_dyn(&self) -> BoxedFuture<'_,Result<(u32,Reader)>>;}implAccessfordynAccessDyn{typeReader = Reader;asyncfnread(&self) -> Result<(u32,Self::Reader)>{self.read_dyn().await}}impl<T:Access + ?Sized>AccessforArc<T>{typeReader = T::Reader;fnread(&self) -> implFuture<Output = Result<(u32,Self::Reader)>> + Send{async{self.as_ref().read().await}}}structReadContext{acc:Arc<dynAccessDyn>,}structReadGenerator{ctx:Arc<ReadContext>,}implReadGenerator{asyncfnnext_reader(&self) -> Result<Option<Reader>>{let(_, r) = self.ctx.acc.read().await?;Ok(Some(r))}}traitStream{typeItem;fnpoll_next(self:Pin<&mutSelf>,cx:&mutContext<'_>) -> Poll<Option<Self::Item>>;}enumTwoWays<A,B>{One(A),Two(B),}impl<A:Read,B:Read>ReadforTwoWays<A,B>{asyncfnread(&mutself) -> Result<Buffer>{matchself{TwoWays::One(v) => v.read().await,TwoWays::Two(v) => v.read().await,}}}structStreamingReader{generator:ReadGenerator,}implReadforStreamingReader{asyncfnread(&mutself) -> Result<Buffer>{let _ = self.generator.next_reader().await;loop{}}}structChunkedReader;implReadforChunkedReader{asyncfnread(&mutself) -> Result<Buffer>{loop{}}}enumState{Idle(Option<TwoWays<StreamingReader,ChunkedReader>>),Reading(Pin<Box<dynFuture<Output = (TwoWays<StreamingReader,ChunkedReader>,Result<Buffer>)> + Send>>),}structBufferStream{state:State,}implStreamforBufferStream{typeItem = Result<()>;fnpoll_next(self:Pin<&mutSelf>,_cx:&mutContext<'_>) -> Poll<Option<Self::Item>>{let this = unsafe{self.get_unchecked_mut()};loop{match&mut this.state{State::Idle(reader) => {letmut reader = reader.take().unwrap();let fut = async{let ret = reader.read().await;(reader, ret)};
this.state = State::Reading(Box::pin(fut));}State::Reading(_) => returnPoll::Pending,}}}}

with

error: internal compiler error: compiler/rustc_mir_transform/src/validate.rs:81:25: broken MIR in Item(DefId(0:86 ~ async_block_box_pin_unsize_broken_mir[e897]::{impl#6}::poll_next)) (after phase change to runtime-optimized) at bb8[0]:
Unsize coercion, but `std::pin::Pin<std::boxed::Box<{async block@/Users/evgeniizh/RustroverProjects/rust/tests/ui/traits/next-solver/async-block-box-pin-unsize-broken-mir.rs:131:31: 131:36}>>` isn't coercible to `std::pin::Pin<std::boxed::Box<dyn std::future::Future<Output = (TwoWays<StreamingReader, ChunkedReader>, std::result::Result<Buffer, ()>)> + std::marker::Send>>`
--> /Users/evgeniizh/RustroverProjects/rust/tests/ui/traits/next-solver/async-block-box-pin-unsize-broken-mir.rs:135:49
|
LL | this.state = State::Reading(Box::pin(fut));
| ^^^^^^^^^^^^^
thread 'rustc' (4832989) panicked at compiler/rustc_mir_transform/src/validate.rs:81:25:

while current main doesn't.

[Randl]

Hm never mind looks like it passes after rebase on main

[BoxyUwU]

@bors delegate+

r=me after rebasing

@rustbot author

[rust-bors]

✌️ @lcnr, you can now approve this pull request!

If @BoxyUwU told you to "r=me" after making some further change, then please make that change and post @bors r=BoxyUwU.

View changes since this delegation.

[rustbot]

This PR was rebased onto a different main commit. Here's a range-diff highlighting what actually changed.

Rebasing is a normal part of keeping PRs up to date, so no action is needed—this note is just to help reviewers.

[lcnr]

@bors r=BoxyUwU rollup (next-solver only)