rustc never finishes after emitting error diagnostic

[djkoloski]

This issue was identified in Fuchsia. We're were unable to move this issue out of tree for reproduction.

Adding a relatively innocuous HashMap<u32, String> field to an existing struct causes the compiler to hang forever:

pub struct Kernel {
    pub new_map: HashMap<u32, String>,
    /// The Zircon job object that holds the processes running in this kernel.
    pub job: zx::Job,

Right before it hangs, it prints:

error[E0412]: cannot find type `HashMap` in this scope
  --> ../../src/starnix/kernel/task/kernel.rs:61:18
   |
61 |     pub new_map: HashMap<u32, String>,
   |                  ^^^^^^^ not found in this scope
   |
help: consider importing this struct
   |
5  + use std::collections::HashMap;
   |

The trailing pipe seems to suggest that some additional diagnostic information is intended to be printed. Adding use std::collections::HashMap; to the file fixes the error and causes the compilation to finish normally.

By attaching to a debug build with gdb, we were able to capture a full call stack. We noticed that the call stack contained many repetitions of this particular set of frames:

#92 rustc_trait_selection::traits::select::{impl#2}::evaluate_candidate::{closure#0} () at compiler/rustc_trait_selection/src/traits/select/mod.rs:1307
#93 rustc_trait_selection::traits::select::SelectionContext::evaluate_candidate (self=0x7f5461f66208, stack=0x7f5461eeda70, candidate=0x7f5461ee9270) at compiler/rustc_trait_selection/src/traits/select/mod.rs:1296
#94 0x00007f546b026c16 in rustc_trait_selection::traits::select::SelectionContext::evaluate_stack (self=0x7f5461f66208, stack=0x7f5461eeda70) at compiler/rustc_trait_selection/src/traits/select/mod.rs:1270
#95 0x00007f546ae96301 in rustc_trait_selection::traits::select::{impl#2}::evaluate_trait_predicate_recursively::{closure#0}::{closure#1} (this=0x7f5461f66208) at compiler/rustc_trait_selection/src/traits/select/mod.rs:1092
#96 rustc_trait_selection::traits::select::{impl#2}::in_task::{closure#0}<rustc_trait_selection::traits::select::{impl#2}::evaluate_trait_predicate_recursively::{closure#0}::{closure_env#1}, core::result::Result<rustc_middle::traits::select::EvaluationResult, rustc_middle::traits::select::OverflowError>> () at compiler/rustc_trait_selection/src/traits/select/mod.rs:1440
#97 rustc_query_system::dep_graph::graph::DepGraph<rustc_middle::dep_graph::DepsType>::with_anon_task<rustc_middle::dep_graph::DepsType, rustc_middle::ty::context::TyCtxt, rustc_trait_selection::traits::select::{impl#2}::in_task::{closure_env#0}<rustc_trait_selection::traits::select::{impl#2}::evaluate_trait_predicate_recursively::{closure#0}::{closure_env#1}, core::result::Result<rustc_middle::traits::select::EvaluationResult, rustc_middle::traits::select::OverflowError>>, core::result::Result<rustc_middle::traits::select::EvaluationResult, rustc_middle::traits::select::OverflowError>> (self=0x7f5461f6a4f8, cx=..., dep_kind=..., op=...) at compiler/rustc_query_system/src/dep_graph/graph.rs:303
#98 0x00007f546b042687 in rustc_trait_selection::traits::select::SelectionContext::in_task<rustc_trait_selection::traits::select::{impl#2}::evaluate_trait_predicate_recursively::{closure#0}::{closure_env#1}, core::result::Result<rustc_middle::traits::select::EvaluationResult, rustc_middle::traits::select::OverflowError>> (self=0x7f5461f66208, op=...) at compiler/rustc_trait_selection/src/traits/select/mod.rs:1440
#99 rustc_trait_selection::traits::select::{impl#2}::evaluate_trait_predicate_recursively::{closure#0} () at compiler/rustc_trait_selection/src/traits/select/mod.rs:1091
#100 rustc_trait_selection::traits::select::SelectionContext::evaluate_trait_predicate_recursively (self=<optimized out>, previous_stack=..., obligation=...) at compiler/rustc_trait_selection/src/traits/select/mod.rs:1046
#101 0x00007f546b040468 in rustc_trait_selection::traits::select::{impl#2}::evaluate_predicate_recursively::{closure#0}::{closure#0} () at compiler/rustc_trait_selection/src/traits/select/mod.rs:686
#102 0x00007f546b03fd09 in stacker::maybe_grow<core::result::Result<rustc_middle::traits::select::EvaluationResult, rustc_middle::traits::select::OverflowError>, rustc_trait_selection::traits::select::{impl#2}::evaluate_predicate_recursively::{closure#0}::{closure_env#0}> (red_zone=102400, stack_size=1048576, callback=...) at /usr/local/google/home/dkoloski/.cargo/registry/src/index.crates.io-6f17d22bba15001f/stacker-0.1.15/src/lib.rs:55
#103 rustc_data_structures::stack::ensure_sufficient_stack<core::result::Result<rustc_middle::traits::select::EvaluationResult, rustc_middle::traits::select::OverflowError>, rustc_trait_selection::traits::select::{impl#2}::evaluate_predicate_recursively::{closure#0}::{closure_env#0}> (f=...) at compiler/rustc_data_structures/src/stack.rs:17
#104 rustc_trait_selection::traits::select::{impl#2}::evaluate_predicate_recursively::{closure#0} () at compiler/rustc_trait_selection/src/traits/select/mod.rs:679
#105 rustc_trait_selection::traits::select::SelectionContext::evaluate_predicate_recursively (self=<optimized out>, previous_stack=..., obligation=...) at compiler/rustc_trait_selection/src/traits/select/mod.rs:658
#106 0x00007f546b03f4bc in rustc_trait_selection::traits::select::SelectionContext::evaluate_predicates_recursively<alloc::vec::into_iter::IntoIter<rustc_infer::traits::Obligation<rustc_middle::ty::Predicate>, alloc::alloc::Global>> (self=0x7f5461f66208, stack=..., predicates=...) at compiler/rustc_trait_selection/src/traits/select/mod.rs:646
#107 0x00007f546ae35d6f in rustc_trait_selection::traits::select::{impl#2}::evaluate_candidate::{closure#0}::{closure#0} (this=0x7f5461f66208) at compiler/rustc_trait_selection/src/traits/select/mod.rs:1312
#108 rustc_trait_selection::traits::select::{impl#2}::evaluation_probe::{closure#0}<rustc_trait_selection::traits::select::{impl#2}::evaluate_candidate::{closure#0}::{closure_env#0}> (snapshot=0x7f5461eee2f0) at compiler/rustc_trait_selection/src/traits/select/mod.rs:612
#109 rustc_infer::infer::InferCtxt::probe<core::result::Result<rustc_middle::traits::select::EvaluationResult, rustc_middle::traits::select::OverflowError>, rustc_trait_selection::traits::select::{impl#2}::evaluation_probe::{closure_env#0}<rustc_trait_selection::traits::select::{impl#2}::evaluate_candidate::{closure#0}::{closure_env#0}>> (self=0x7f5461f662d8, f=...) at compiler/rustc_infer/src/infer/mod.rs:885
#110 0x00007f546b044753 in rustc_trait_selection::traits::select::SelectionContext::evaluation_probe<rustc_trait_selection::traits::select::{impl#2}::evaluate_candidate::{closure#0}::{closure_env#0}> (self=0x7f5461f66208, op=...) at compiler/rustc_trait_selection/src/traits/select/mod.rs:610

In our case, the total number of frames reached about 2500. It looks like this may be hitting some kind of exponential computation edge case.

Potentially related: #116914

Meta

rustc --version --verbose:

rustc 1.75.0-nightly (9d1e4b787 2023-10-11)
binary: rustc
commit-hash: 9d1e4b7870f0aecb9f53e71f3cca3529b21d677a
commit-date: 2023-10-11
host: x86_64-unknown-linux-gnu
release: 1.75.0-nightly
LLVM version: 17.0.2

This is a custom-built toolchain following our toolchain build instructions.

[compiler-errors]

Can you share the top and bottom 200 frames (or ideally all of them)? The ones you shared aren't really useful if trying to diagnose exactly how we got here.

[djkoloski]

Here's a gist of all of the frames in the callstack I captured. I added it to the main issue description as well.

[erickt]

After letting the compiler run for 30+ minutes, we managed to get a rustc-ice during selection which comes with a very large backtrace.

[erickt]

I ran our build again, and it produced essentially the same ice, so good news it's repeatable. The downside is it was compiling for about 7 hours for the ICE to occur :)

[djkoloski]

After some bisecting and code reduction, it looks like this may be related to #87012

[lqd]

Example starnix_core reduction:

use std::collections::{BTreeMap, HashMap, HashSet};
use std::sync::{Arc, RwLock, Weak};

pub struct Kernel {
    pub pids: RwLock<PidTable>,
}
pub struct PidTable {
    table: HashMap<u32, PidEntry>,
}
struct PidEntry {
    process_group: Option<Weak<ProcessGroup>>,
}

pub struct ProcessGroupMutableState {
    thread_groups: BTreeMap<u32, Weak<ThreadGroup>>,
}
pub struct ProcessGroup {
    mutable_state: RwLock<ProcessGroupMutableState>,
}
pub struct Task {
    fs: Option<Arc<FsContext>>,
}
pub struct ThreadGroup {
    pub kernel: Arc<Kernel>,
    mutable_state: RwLock<ThreadGroupMutableState>,
}
pub struct ThreadGroupMutableState {
    tasks: BTreeMap<u32, TaskContainer>,
}
pub struct TaskContainer(Arc<Task>);
pub struct FileSystem {
    pub kernel: Weak<Kernel>,
}
struct FsContextState {
    namespace: Arc<Namespace>,
}
pub struct FsContext {
    state: RwLock<FsContextState>,
}

pub struct Namespace {
    root_mount: Arc<Mount>,
}
pub struct Mount {
    fs: Arc<FileSystem>,
    state: RwLock<MountState>,
}
pub struct MountState {
    submounts: HashMap<Arc<ERROR>, Arc<Mount>>,
    peer_group_: Option<(Arc<PeerGroup>, Arc<Mount>)>, // +6s
    upstream_: Option<(Weak<PeerGroup>, Arc<Mount>)>, // +5s
}
struct PeerGroup {
    state: RwLock<PeerGroupState>,
}
struct PeerGroupState {
    downstream: HashSet<Arc<Mount>>,
}
pub trait SocketOps: Send + Sync {} // + Sync = +5s
struct UEventNetlinkSocket {
    kernel: Arc<Kernel>,
}
impl SocketOps for UEventNetlinkSocket {} // load-bearing

fn main() {}

The above is:

• slow using the old solver, when there's the error
• fast using the old solver, without the non-existent ERROR type
• fast using the new solver, with or without the non-existent ERROR type

This behavior started in 1.73:

Version	Item	Self time	% of total time	Time	Item count
1.70.0	evaluate_obligation	646.14µs	3.441	1.24ms	186
1.71.0	evaluate_obligation	699.79µs	3.830	1.37ms	186
1.72.0	evaluate_obligation	737.36µs	3.974	1.43ms	186
1.73.0	evaluate_obligation	5.88s	99.738	5.88s	177
1.74.0	evaluate_obligation	5.86s	99.738	5.86s	177
1.75.0	evaluate_obligation	5.63s	99.736	5.63s	176
1.76.0	evaluate_obligation	5.57s	99.739	5.57s	176
1.77.0	evaluate_obligation	5.59s	99.676	5.59s	167
1.78.0	evaluate_obligation	5.59s	99.673	5.59s	168

I'm pretty sure it appeared in nightly-2023-07-29 (commit range), where #113312 sticks out as particularly relevant (though it's a soundness fix), since the slowest evaluations are related to auto traits:

evaluate_obligation, value: TraitPredicate(<UEventNetlinkSocket as std::marker::Sync>, took 3299 ms
evaluate_obligation, value: TraitPredicate(<UEventNetlinkSocket as std::marker::Send>, took 3732 ms