compiler/rustc_query_impl/src/execution.rs

use std::hash::Hash;
use std::mem::ManuallyDrop;
use std::num::NonZero;
use std::sync::Arc;

use parking_lot::{Condvar, Mutex};
use rustc_data_structures::hash_table::Entry;
use rustc_data_structures::{defer, outline, sharded, sync};
use rustc_errors::FatalError;
use rustc_middle::dep_graph::{
    DepGraphData, DepNode, DepNodeIndex, DepNodeKey, SerializedDepNodeIndex,
};
use rustc_middle::query::{
    ActiveKeyStatus, QueryCache, QueryCycle, QueryJob, QueryJobId, QueryLatch, QueryMode,
    QueryState, QueryVTable, QueryWaiter,
};
use rustc_middle::ty::TyCtxt;
use rustc_middle::ty::tls::{self, ImplicitCtxt};
use rustc_middle::verify_ich::incremental_verify_ich;
use rustc_span::def_id::LOCAL_CRATE;
use rustc_span::{DUMMY_SP, Span};
use rustc_structures::Limit;

use crate::diagnostics::{QueryOverflow, QueryOverflowNote};
use crate::handle_cycle_error;
use crate::incremental::should_verify_loaded_value;
use crate::job::{
    CollectActiveJobsKind, collect_active_query_jobs, find_cycle_in_stack, find_dep_kind_root,
};

#[inline]
fn equivalent_key<K: Eq, V>(k: K) -> impl Fn(&(K, V)) -> bool {
    move |x| x.0 == k
}

#[cold]
#[inline(never)]
fn handle_cycle<'tcx, C: QueryCache>(
    query: &'tcx QueryVTable<'tcx, C>,
    tcx: TyCtxt<'tcx>,
    key: C::Key,
    cycle: QueryCycle<'tcx>,
) -> C::Value {
    let nested;
    {
        let mut nesting = tcx.query_system.cycle_handler_nesting.lock();
        nested = match *nesting {
            0 => false,
            1 => true,
            _ => {
                // Don't print further nested errors to avoid cases of infinite recursion
                tcx.dcx().delayed_bug("doubly nested cycle error").raise_fatal()
            }
        };
        *nesting += 1;
    }
    let _guard = defer(|| *tcx.query_system.cycle_handler_nesting.lock() -= 1);

    let error = handle_cycle_error::create_cycle_error(tcx, &cycle, nested);

    if nested {
        // Avoid custom handlers and only use the robust `create_cycle_error` for nested cycle errors
        handle_cycle_error::default(error)
    } else {
        (query.handle_cycle_error_fn)(tcx, key, cycle, error)
    }
}

/// Signals to waiters that the query is complete.
///
/// This does nothing for single threaded rustc, as there are no concurrent jobs which could be
/// waiting on us.
#[inline]
fn signal_complete(job: QueryJob<'_>) {
    if let Some(latch) = job.latch {
        // Set the latch and resume all waiters on it.
        let mut waiters_guard = latch.waiters.lock();
        let waiters = waiters_guard.take().unwrap(); // mark the latch as complete
        let registry = rustc_thread_pool::Registry::current();
        for waiter in waiters {
            rustc_thread_pool::mark_unblocked(&registry);
            waiter.condvar.notify_one();
        }
    }
}

/// Guard object representing the responsibility to execute a query job and
/// mark it as completed.
///
/// This will poison the relevant query key if it is dropped without calling
/// [`Self::complete`].
struct ActiveJobGuard<'tcx, K>
where
    K: Eq + Hash + Copy,
{
    state: &'tcx QueryState<'tcx, K>,
    key: K,
    key_hash: u64,
}

impl<'tcx, K> ActiveJobGuard<'tcx, K>
where
    K: Eq + Hash + Copy,
{
    /// Completes the query by updating the query cache with the `result`,
    /// signals the waiter, and forgets the guard so it won't poison the query.
    fn complete<C>(self, cache: &C, value: C::Value, dep_node_index: DepNodeIndex)
    where
        C: QueryCache<Key = K>,
    {
        // Mark as complete before we remove the job from the active state
        // so no other thread can re-execute this query.
        cache.complete(self.key, value, dep_node_index);

        let mut this = ManuallyDrop::new(self);

        // Drop everything without poisoning the query.
        this.drop_and_maybe_poison(/* poison */ false);
    }

    fn drop_and_maybe_poison(&mut self, poison: bool) {
        let status = {
            let mut shard = self.state.active.lock_shard_by_hash(self.key_hash);
            match shard.find_entry(self.key_hash, equivalent_key(self.key)) {
                Err(_) => {
                    // Note: we must not panic while holding the lock, because unwinding also looks
                    // at this map, which can result in a double panic. So drop it first.
                    drop(shard);
                    panic!();
                }
                Ok(occupied) => {
                    let ((key, status), vacant) = occupied.remove();
                    if poison {
                        vacant.insert((key, ActiveKeyStatus::Poisoned));
                    }
                    status
                }
            }
        };

        // Also signal the completion of the job, so waiters will continue execution.
        match status {
            ActiveKeyStatus::Started(job) => signal_complete(job),
            ActiveKeyStatus::Poisoned => panic!(),
        }
    }
}

impl<'tcx, K> Drop for ActiveJobGuard<'tcx, K>
where
    K: Eq + Hash + Copy,
{
    #[inline(never)]
    #[cold]
    fn drop(&mut self) {
        // Poison the query so jobs waiting on it panic.
        self.drop_and_maybe_poison(/* poison */ true);
    }
}

#[cold]
#[inline(never)]
fn find_and_handle_cycle<'tcx, C: QueryCache>(
    query: &'tcx QueryVTable<'tcx, C>,
    tcx: TyCtxt<'tcx>,
    key: C::Key,
    try_execute: QueryJobId,
    span: Span,
) -> (C::Value, Option<DepNodeIndex>) {
    // Ensure there were no errors collecting all active jobs.
    // We need the complete map to ensure we find a cycle to break.
    let job_map = collect_active_query_jobs(tcx, CollectActiveJobsKind::FullNoContention);

    let cycle = find_cycle_in_stack(try_execute, job_map, &current_query_job(), span);
    (handle_cycle(query, tcx, key, cycle), None)
}

/// Awaits for the query job to complete.
fn latch_wait_on<'tcx>(
    latch: &QueryLatch<'tcx>,
    query: Option<QueryJobId>,
    span: Span,
) -> Result<(), QueryCycle<'tcx>> {
    let mut waiters_guard = latch.waiters.lock();
    let Some(waiters) = &mut *waiters_guard else {
        return Ok(()); // already complete
    };

    let waiter = Arc::new(QueryWaiter {
        parent: query,
        span,
        cycle: Mutex::new(None),
        condvar: Condvar::new(),
    });

    // We push the waiter on to the `waiters` list. It can be accessed inside the `wait` call
    // below, by 1) the `signal_complete` function or 2) by deadlock detection. Both of these will
    // remove it from the `waiters` list before resuming this thread.
    waiters.push(Arc::clone(&waiter));

    // Awaits the caller on this latch by blocking the current thread. If this detects a deadlock
    // and the deadlock handler wants to resume this thread we have to be in the `wait` call. This
    // is ensured by the deadlock handler getting the waiters lock.
    rustc_thread_pool::mark_blocked_and_wait(|| {
        waiter.condvar.wait(&mut waiters_guard);
        // Release the lock before we potentially block when acquiring jobserver token.
        drop(waiters_guard);
    });

    // FIXME: Get rid of this lock. We have ownership of the QueryWaiter although another thread
    // may still have a Arc reference so we cannot use Arc::get_mut.
    let mut cycle = waiter.cycle.lock();
    match cycle.take() {
        None => Ok(()),
        Some(cycle) => Err(cycle),
    }
}

#[inline(always)]
fn wait_for_query<'tcx, C: QueryCache>(
    query: &'tcx QueryVTable<'tcx, C>,
    tcx: TyCtxt<'tcx>,
    span: Span,
    key: C::Key,
    key_hash: u64,
    latch: QueryLatch<'tcx>,
    current: Option<QueryJobId>,
) -> (C::Value, Option<DepNodeIndex>) {
    // For parallel queries, we'll block and wait until the query running
    // in another thread has completed. Record how long we wait in the
    // self-profiler.
    let query_blocked_prof_timer = tcx.prof.query_blocked();

    // With parallel queries we might just have to wait on some other thread.
    let result = latch_wait_on(&latch, current, span);

    match result {
        Ok(()) => {
            let Some((v, index)) = query.cache.lookup(&key) else {
                outline(|| {
                    // We didn't find the query result in the query cache. Check if it was
                    // poisoned due to a panic instead.
                    let shard = query.state.active.lock_shard_by_hash(key_hash);
                    match shard.find(key_hash, equivalent_key(key)) {
                        // The query we waited on panicked. Continue unwinding here.
                        Some((_, ActiveKeyStatus::Poisoned)) => FatalError.raise(),
                        _ => panic!(
                            "query '{}' result must be in the cache or the query must be poisoned after a wait",
                            query.name
                        ),
                    }
                })
            };

            tcx.prof.query_cache_hit(index.into());
            query_blocked_prof_timer.finish_with_query_invocation_id(index.into());

            (v, Some(index))
        }
        Err(cycle) => (handle_cycle(query, tcx, key, cycle), None),
    }
}

#[inline]
fn next_job_id<'tcx>(tcx: TyCtxt<'tcx>) -> QueryJobId {
    QueryJobId(
        NonZero::new(tcx.query_system.jobs.fetch_add(1, std::sync::atomic::Ordering::Relaxed))
            .unwrap(),
    )
}

#[inline]
fn current_query_job() -> Option<QueryJobId> {
    tls::with_context(|icx| icx.query)
}

/// Shared main part of both [`execute_query_incr_inner`] and [`execute_query_non_incr_inner`].
#[inline(never)]
fn try_execute_query<'tcx, C: QueryCache, const INCR: bool>(
    query: &'tcx QueryVTable<'tcx, C>,
    tcx: TyCtxt<'tcx>,
    span: Span,
    key: C::Key,
    dep_node: Option<DepNode>, // `None` for non-incremental, `Some` for incremental
) -> (C::Value, Option<DepNodeIndex>) {
    let key_hash = sharded::make_hash(&key);
    let mut state_lock = query.state.active.lock_shard_by_hash(key_hash);

    // For the parallel compiler we need to check both the query cache and query state structures
    // while holding the state lock to ensure that 1) the query has not yet completed and 2) the
    // query is not still executing. Without checking the query cache here, we can end up
    // re-executing the query since `try_start` only checks that the query is not currently
    // executing, but another thread may have already completed the query and stores it result
    // in the query cache.
    if tcx.sess.opts.jobs.frontend.is_some() {
        if let Some((value, index)) = query.cache.lookup(&key) {
            tcx.prof.query_cache_hit(index.into());
            return (value, Some(index));
        }
    }

    let current_job_id = current_query_job();

    match state_lock.entry(key_hash, equivalent_key(key), |(k, _)| sharded::make_hash(k)) {
        Entry::Vacant(entry) => {
            // Nothing has computed or is computing the query, so we start a new job and insert it
            // in the state map.
            let id = next_job_id(tcx);
            let job = QueryJob::new(id, span, current_job_id);
            entry.insert((key, ActiveKeyStatus::Started(job)));

            // Drop the lock before we start executing the query.
            drop(state_lock);

            // Set up a guard object that will automatically poison the query if a
            // panic occurs while executing the query (or any intermediate plumbing).
            let job_guard = ActiveJobGuard { state: &query.state, key, key_hash };

            // Delegate to another function to actually execute the query job.
            let (value, dep_node_index) = if INCR {
                execute_job_incr(query, tcx, key, dep_node.unwrap(), id)
            } else {
                execute_job_non_incr(query, tcx, key, id)
            };

            if query.feedable {
                check_feedable_consistency(tcx, query, key, &value);
            }

            // Tell the guard to insert `value` in the cache and remove the status entry from
            // `query.state`.
            job_guard.complete(&query.cache, value, dep_node_index);

            (value, Some(dep_node_index))
        }
        Entry::Occupied(mut entry) => {
            match &mut entry.get_mut().1 {
                ActiveKeyStatus::Started(job) => {
                    if sync::is_dyn_thread_safe() {
                        // Get the latch out
                        let latch = job.latch.get_or_insert_with(QueryLatch::new).clone();
                        drop(state_lock);

                        // Only call `wait_for_query` if we're using a Rayon thread pool
                        // as it will attempt to mark the worker thread as blocked.
                        wait_for_query(query, tcx, span, key, key_hash, latch, current_job_id)
                    } else {
                        let id = job.id;
                        drop(state_lock);

                        // If we are single-threaded we know that we have cycle error,
                        // so we just return the error.
                        find_and_handle_cycle(query, tcx, key, id, span)
                    }
                }
                ActiveKeyStatus::Poisoned => FatalError.raise(),
            }
        }
    }
}

#[inline(always)]
fn check_feedable_consistency<'tcx, C: QueryCache>(
    tcx: TyCtxt<'tcx>,
    query: &'tcx QueryVTable<'tcx, C>,
    key: C::Key,
    value: &C::Value,
) {
    // We should not compute queries that also got a value via feeding.
    // This can't happen, as query feeding adds the very dependencies to the fed query
    // as its feeding query had. So if the fed query is red, so is its feeder, which will
    // get evaluated first, and re-feed the query.
    let Some((cached_value, _)) = query.cache.lookup(&key) else { return };

    let Some(hash_value_fn) = query.hash_value_fn else {
        panic!(
            "no_hash fed query later has its value computed.\n\
            Remove `no_hash` modifier to allow recomputation.\n\
            The already cached value: {}",
            (query.format_value)(&cached_value)
        );
    };

    let (old_hash, new_hash) = tcx.with_stable_hashing_context(|mut hcx| {
        (hash_value_fn(&mut hcx, &cached_value), hash_value_fn(&mut hcx, value))
    });
    let formatter = query.format_value;
    if old_hash != new_hash {
        // We have an inconsistency. This can happen if one of the two
        // results is tainted by errors.
        assert!(
            tcx.dcx().has_errors().is_some(),
            "Computed query value for {:?}({:?}) is inconsistent with fed value,\n\
                computed={:#?}\nfed={:#?}",
            query.dep_kind,
            key,
            formatter(value),
            formatter(&cached_value),
        );
    }
}

fn depth_limit_error<'tcx>(tcx: TyCtxt<'tcx>, job: QueryJobId) {
    let job_map = collect_active_query_jobs(tcx, CollectActiveJobsKind::Full);
    let (span, desc, depth) = find_dep_kind_root(tcx, job, job_map);

    let suggested_limit = match tcx.recursion_limit() {
        Limit(0) => Limit(2),
        limit => limit * 2,
    };

    tcx.dcx().emit_fatal(QueryOverflow {
        span,
        note: QueryOverflowNote { desc, depth },
        suggested_limit,
        crate_name: tcx.crate_name(LOCAL_CRATE),
    });
}

/// Executes a job by changing the `ImplicitCtxt` to point to the new query job while it executes.
#[inline(always)]
fn start_query<R>(job_id: QueryJobId, depth_limit: bool, compute: impl FnOnce() -> R) -> R {
    tls::with_context(move |icx| {
        if depth_limit && !icx.tcx.recursion_limit().value_within_limit(icx.query_depth) {
            depth_limit_error(icx.tcx, job_id);
        }

        // Update the `ImplicitCtxt` to point to our new query job.
        let icx = ImplicitCtxt {
            query: Some(job_id),
            query_depth: icx.query_depth + if depth_limit { 1 } else { 0 },
            ..*icx
        };

        // Use the `ImplicitCtxt` while we execute the query.
        tls::enter_context(&icx, compute)
    })
}

// Fast path for when incr. comp. is off.
#[inline(always)]
fn execute_job_non_incr<'tcx, C: QueryCache>(
    query: &'tcx QueryVTable<'tcx, C>,
    tcx: TyCtxt<'tcx>,
    key: C::Key,
    job_id: QueryJobId,
) -> (C::Value, DepNodeIndex) {
    debug_assert!(!tcx.dep_graph.is_fully_enabled());

    let prof_timer = tcx.prof.query_provider();
    // Call the query provider.
    let value = start_query(job_id, query.depth_limit, || (query.invoke_provider_fn)(tcx, key));
    let dep_node_index = tcx.dep_graph.next_virtual_depnode_index();
    prof_timer.finish_with_query_invocation_id(dep_node_index.into());

    // Sanity: Fingerprint the key and the result to assert they don't contain anything unhashable.
    if cfg!(debug_assertions) {
        let _ = key.to_fingerprint(tcx);
        if let Some(hash_value_fn) = query.hash_value_fn {
            tcx.with_stable_hashing_context(|mut hcx| {
                hash_value_fn(&mut hcx, &value);
            });
        }
    }

    (value, dep_node_index)
}

#[inline(always)]
fn execute_job_incr<'tcx, C: QueryCache>(
    query: &'tcx QueryVTable<'tcx, C>,
    tcx: TyCtxt<'tcx>,
    key: C::Key,
    dep_node: DepNode,
    job_id: QueryJobId,
) -> (C::Value, DepNodeIndex) {
    let dep_graph_data =
        tcx.dep_graph.data().expect("should always be present in incremental mode");

    if !query.eval_always {
        // The diagnostics for this query will be promoted to the current session during
        // `try_mark_green()`, so we can ignore them here.
        if let Some(ret) = start_query(job_id, false, || try {
            let (prev_index, dep_node_index) = dep_graph_data.try_mark_green(tcx, &dep_node)?;
            let value = load_from_disk_or_invoke_provider_green(
                tcx,
                dep_graph_data,
                query,
                key,
                &dep_node,
                prev_index,
                dep_node_index,
            );
            (value, dep_node_index)
        }) {
            return ret;
        }
    }

    let prof_timer = tcx.prof.query_provider();

    let (result, dep_node_index) = start_query(job_id, query.depth_limit, || {
        // Call the query provider.
        dep_graph_data.with_task(
            dep_node,
            tcx,
            || (query.invoke_provider_fn)(tcx, key),
            query.hash_value_fn,
        )
    });

    prof_timer.finish_with_query_invocation_id(dep_node_index.into());

    (result, dep_node_index)
}

/// Given that the dep node for this query+key is green, obtain a value for it
/// by loading one from disk if possible, or by invoking its query provider if
/// necessary.
#[inline(always)]
fn load_from_disk_or_invoke_provider_green<'tcx, C: QueryCache>(
    tcx: TyCtxt<'tcx>,
    dep_graph_data: &DepGraphData,
    query: &'tcx QueryVTable<'tcx, C>,
    key: C::Key,
    dep_node: &DepNode,
    prev_index: SerializedDepNodeIndex,
    dep_node_index: DepNodeIndex,
) -> C::Value {
    // Note this function can be called concurrently from the same query
    // We must ensure that this is handled correctly.

    debug_assert!(dep_graph_data.is_index_green(prev_index));

    // First try to load the result from the on-disk cache. Some things are never cached on disk.
    let try_value = if query.will_cache_on_disk_for_key(key) {
        let prof_timer = tcx.prof.incr_cache_loading();
        let value = (query.try_load_from_disk_fn)(tcx, prev_index);
        prof_timer.finish_with_query_invocation_id(dep_node_index.into());
        value
    } else {
        None
    };
    let (value, verify) = match try_value {
        Some(value) => {
            if std::intrinsics::unlikely(tcx.sess.opts.unstable_opts.query_dep_graph) {
                dep_graph_data.mark_debug_loaded_from_disk(*dep_node)
            }

            let verify = should_verify_loaded_value(tcx, dep_graph_data, dep_node.key_fingerprint);

            (value, verify)
        }
        None => {
            // We could not load a result from the on-disk cache, so recompute. The dep-graph for
            // this computation is already in-place, so we can just call the query provider.
            let prof_timer = tcx.prof.query_provider();
            let value = tcx.dep_graph.with_ignore(|| (query.invoke_provider_fn)(tcx, key));
            prof_timer.finish_with_query_invocation_id(dep_node_index.into());

            (value, true)
        }
    };

    if verify {
        // Verify that re-running the query produced a result with the expected hash.
        // This catches bugs in query implementations, turning them into ICEs.
        // For example, a query might sort its result by `DefId` - since `DefId`s are
        // not stable across compilation sessions, the result could get up getting sorted
        // in a different order when the query is re-run, even though all of the inputs
        // (e.g. `DefPathHash` values) were green.
        //
        // See issue #82920 for an example of a miscompilation that would get turned into
        // an ICE by this check
        incremental_verify_ich(
            tcx,
            dep_graph_data,
            &value,
            prev_index,
            query.hash_value_fn,
            query.format_value,
        );
    }

    value
}

/// Checks whether a `tcx.ensure_ok()` query call can
/// return early without actually trying to execute.
///
/// This only makes sense during incremental compilation, because it relies
/// on having the dependency graph (and in some cases a disk-cached value)
/// from the previous incr-comp session.
#[inline(never)]
fn ensure_can_skip_execution<'tcx, C: QueryCache>(
    query: &'tcx QueryVTable<'tcx, C>,
    tcx: TyCtxt<'tcx>,
    dep_node: DepNode,
) -> bool {
    // Queries with `eval_always` should never skip execution.
    if query.eval_always {
        return false;
    }

    match tcx.dep_graph.try_mark_green(tcx, &dep_node) {
        None => {
            // A None return from `try_mark_green` means that this is either
            // a new dep node or that the dep node has already been marked red.
            // Either way, we can't call `dep_graph.read()` as we don't have the
            // DepNodeIndex. We must invoke the query itself. The performance cost
            // this introduces should be negligible as we'll immediately hit the
            // in-memory cache, or another query down the line will.
            false
        }
        Some((_, dep_node_index)) => {
            tcx.dep_graph.read_index(dep_node_index);
            tcx.prof.query_cache_hit(dep_node_index.into());

            // We can skip execution for this key if the
            // node is green. It must have succeeded in the previous
            // session, and therefore would succeed in the current session
            // if executed.
            true
        }
    }
}

/// Called by a macro-generated impl of [`QueryVTable::execute_query_fn`],
/// in non-incremental mode.
#[inline(always)]
pub(super) fn execute_query_non_incr_inner<'tcx, C: QueryCache>(
    query: &'tcx QueryVTable<'tcx, C>,
    tcx: TyCtxt<'tcx>,
    span: Span,
    key: C::Key,
) -> C::Value {
    try_execute_query::<C, false>(query, tcx, span, key, None).0
}

/// Called by a macro-generated impl of [`QueryVTable::execute_query_fn`],
/// in incremental mode.
#[inline(always)]
pub(super) fn execute_query_incr_inner<'tcx, C: QueryCache>(
    query: &'tcx QueryVTable<'tcx, C>,
    tcx: TyCtxt<'tcx>,
    span: Span,
    key: C::Key,
    mode: QueryMode,
) -> Option<C::Value> {
    let dep_node = DepNode::construct(tcx, query.dep_kind, &key);

    // Check if query execution can be skipped, for `ensure_ok`.
    if let QueryMode::EnsureOk = mode
        && ensure_can_skip_execution(query, tcx, dep_node)
    {
        return None;
    }

    let (result, dep_node_index) =
        try_execute_query::<C, true>(query, tcx, span, key, Some(dep_node));
    if let Some(dep_node_index) = dep_node_index {
        tcx.dep_graph.read_index(dep_node_index)
    }
    Some(result)
}

/// Inner implementation of [`DepKindVTable::force_from_dep_node_fn`][force_fn]
/// for query nodes.
///
/// [force_fn]: rustc_middle::dep_graph::DepKindVTable::force_from_dep_node_fn
pub(crate) fn force_query_dep_node<'tcx, C: QueryCache>(
    tcx: TyCtxt<'tcx>,
    query: &'tcx QueryVTable<'tcx, C>,
    dep_node: DepNode,
) -> bool {
    let Some(key) = C::Key::try_recover_key(tcx, &dep_node) else {
        // We couldn't recover a key from the node's key fingerprint.
        // Tell the caller that we couldn't force the node.
        return false;
    };

    try_execute_query::<C, true>(query, tcx, DUMMY_SP, key, Some(dep_node));

    // We did manage to recover a key and force the node, though it's up to
    // the caller to check whether the node ended up marked red or green.
    true
}