[WIP] Split query execution into hot and cold paths

src/librustc/dep_graph/graph.rs

@@ -1122,6 +1122,7 @@ impl CurrentDepGraph {
 }
 
 impl DepGraphData {
+    #[inline]
     fn read_index(&self, source: DepNodeIndex) {
         ty::tls::with_context_opt(|icx| {
             let icx = if let Some(icx) = icx { icx } else { return };

src/librustc/ty/context.rs

@@ -1688,6 +1688,7 @@ pub mod tls {
 
     /// Gets the pointer to the current `ImplicitCtxt`.
     #[cfg(not(parallel_compiler))]
+    #[inline]
     fn get_tlv() -> usize {
         TLV.with(|tlv| tlv.get())
     }

src/librustc/ty/query/plumbing.rs

@@ -12,10 +12,8 @@ use crate::ty::{self, TyCtxt};
 #[cfg(not(parallel_compiler))]
 use rustc_data_structures::cold_path;
 use rustc_data_structures::fx::{FxHashMap, FxHasher};
-#[cfg(parallel_compiler)]
-use rustc_data_structures::profiling::TimingGuard;
 use rustc_data_structures::sharded::Sharded;
-use rustc_data_structures::sync::Lock;
+use rustc_data_structures::sync::{Lock, LockGuard};
 use rustc_data_structures::thin_vec::ThinVec;
 use rustc_errors::{struct_span_err, Diagnostic, DiagnosticBuilder, FatalError, Handler, Level};
 use rustc_span::source_map::DUMMY_SP;
@@ -70,6 +68,12 @@ impl<'tcx, M: QueryConfig<'tcx>> Default for QueryCache<'tcx, M> {
     }
 }
 
+/// Values used when checking a query cache which can be reused on a cache-miss to execute the query.
+pub(super) struct QueryLookup<'tcx, Q: QueryDescription<'tcx>> {
+    shard: usize,
+    lock: LockGuard<'tcx, QueryCache<'tcx, Q>>,
+}
+
 /// A type representing the responsibility to execute the job in the `job` field.
 /// This will poison the relevant query if dropped.
 pub(super) struct JobOwner<'a, 'tcx, Q: QueryDescription<'tcx>> {
@@ -81,119 +85,91 @@ pub(super) struct JobOwner<'a, 'tcx, Q: QueryDescription<'tcx>> {
 impl<'a, 'tcx, Q: QueryDescription<'tcx>> JobOwner<'a, 'tcx, Q> {
     /// Either gets a `JobOwner` corresponding the query, allowing us to
     /// start executing the query, or returns with the result of the query.
-    /// If the query is executing elsewhere, this will wait for it.
+    /// This function assumes that `try_get_cached` is already called and returned `lookup`.
+    /// If the query is executing elsewhere, this will wait for it and return the result.
     /// If the query panicked, this will silently panic.
     ///
     /// This function is inlined because that results in a noticeable speed-up
     /// for some compile-time benchmarks.
     #[inline(always)]
-    pub(super) fn try_get(tcx: TyCtxt<'tcx>, span: Span, key: &Q::Key) -> TryGetJob<'a, 'tcx, Q> {
-        // Handling the `query_blocked_prof_timer` is a bit weird because of the
-        // control flow in this function: Blocking is implemented by
-        // awaiting a running job and, once that is done, entering the loop below
-        // again from the top. In that second iteration we will hit the
-        // cache which provides us with the information we need for
-        // finishing the "query-blocked" event.
-        //
-        // We thus allocate `query_blocked_prof_timer` outside the loop,
-        // initialize it during the first iteration and finish it during the
-        // second iteration.
-        #[cfg(parallel_compiler)]
-        let mut query_blocked_prof_timer: Option<TimingGuard<'_>> = None;
-
-        let cache = Q::query_cache(tcx);
-        loop {
-            // We compute the key's hash once and then use it for both the
-            // shard lookup and the hashmap lookup. This relies on the fact
-            // that both of them use `FxHasher`.
-            let mut state = FxHasher::default();
-            key.hash(&mut state);
-            let key_hash = state.finish();
-
-            let shard = cache.get_shard_index_by_hash(key_hash);
-            let mut lock_guard = cache.get_shard_by_index(shard).lock();
-            let lock = &mut *lock_guard;
-
-            if let Some((_, value)) =
-                lock.results.raw_entry().from_key_hashed_nocheck(key_hash, key)
-            {
-                if unlikely!(tcx.prof.enabled()) {
-                    tcx.prof.query_cache_hit(value.index.into());
+    pub(super) fn try_start(
+        tcx: TyCtxt<'tcx>,
+        span: Span,
+        key: &Q::Key,
+        mut lookup: QueryLookup<'tcx, Q>,
+    ) -> TryGetJob<'a, 'tcx, Q> {
+        let lock = &mut *lookup.lock;
+
+        let (latch, mut _query_blocked_prof_timer) = match lock.active.entry((*key).clone()) {
+            Entry::Occupied(mut entry) => {
+                match entry.get_mut() {
+                    QueryResult::Started(job) => {
+                        // 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 = if cfg!(parallel_compiler) {
+                            Some(tcx.prof.query_blocked())
+                        } else {
+                            None
+                        };
 
-                    #[cfg(parallel_compiler)]
-                    {
-                        if let Some(prof_timer) = query_blocked_prof_timer.take() {
-                            prof_timer.finish_with_query_invocation_id(value.index.into());
-                        }
-                    }
-                }
+                        // Create the id of the job we're waiting for
+                        let id = QueryJobId::new(job.id, lookup.shard, Q::dep_kind());
 
-                let result = (value.value.clone(), value.index);
-                #[cfg(debug_assertions)]
-                {
-                    lock.cache_hits += 1;
+                        (job.latch(id), _query_blocked_prof_timer)
+                    }
+                    QueryResult::Poisoned => FatalError.raise(),
                 }
-                return TryGetJob::JobCompleted(result);
             }
+            Entry::Vacant(entry) => {
+                // No job entry for this query. Return a new one to be started later.
 
-            let latch = match lock.active.entry((*key).clone()) {
-                Entry::Occupied(mut entry) => {
-                    match entry.get_mut() {
-                        QueryResult::Started(job) => {
-                            // 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.
-                            #[cfg(parallel_compiler)]
-                            {
-                                query_blocked_prof_timer = Some(tcx.prof.query_blocked());
-                            }
-
-                            // Create the id of the job we're waiting for
-                            let id = QueryJobId::new(job.id, shard, Q::dep_kind());
+                // Generate an id unique within this shard.
+                let id = lock.jobs.checked_add(1).unwrap();
+                lock.jobs = id;
+                let id = QueryShardJobId(NonZeroU32::new(id).unwrap());
 
-                            job.latch(id)
-                        }
-                        QueryResult::Poisoned => FatalError.raise(),
-                    }
-                }
-                Entry::Vacant(entry) => {
-                    // No job entry for this query. Return a new one to be started later.
+                let global_id = QueryJobId::new(id, lookup.shard, Q::dep_kind());
 
-                    // Generate an id unique within this shard.
-                    let id = lock.jobs.checked_add(1).unwrap();
-                    lock.jobs = id;
-                    let id = QueryShardJobId(NonZeroU32::new(id).unwrap());
+                let job = tls::with_related_context(tcx, |icx| QueryJob::new(id, span, icx.query));
 
-                    let global_id = QueryJobId::new(id, shard, Q::dep_kind());
+                entry.insert(QueryResult::Started(job));
 
-                    let job =
-                        tls::with_related_context(tcx, |icx| QueryJob::new(id, span, icx.query));
+                let owner =
+                    JobOwner { cache: Q::query_cache(tcx), id: global_id, key: (*key).clone() };
+                return TryGetJob::NotYetStarted(owner);
+            }
+        };
+        mem::drop(lookup.lock);
 
-                    entry.insert(QueryResult::Started(job));
+        // If we are single-threaded we know that we have cycle error,
+        // so we just return the error.
+        #[cfg(not(parallel_compiler))]
+        return TryGetJob::Cycle(cold_path(|| {
+            Q::handle_cycle_error(tcx, latch.find_cycle_in_stack(tcx, span))
+        }));
 
-                    let owner = JobOwner { cache, id: global_id, key: (*key).clone() };
-                    return TryGetJob::NotYetStarted(owner);
-                }
-            };
-            mem::drop(lock_guard);
+        // With parallel queries we might just have to wait on some other
+        // thread.
+        #[cfg(parallel_compiler)]
+        {
+            let result = latch.wait_on(tcx, span);
 
-            // If we are single-threaded we know that we have cycle error,
-            // so we just return the error.
-            #[cfg(not(parallel_compiler))]
-            return TryGetJob::Cycle(cold_path(|| {
-                Q::handle_cycle_error(tcx, latch.find_cycle_in_stack(tcx, span))
-            }));
+            if let Err(cycle) = result {
+                return TryGetJob::Cycle(Q::handle_cycle_error(tcx, cycle));
+            }
 
-            // With parallel queries we might just have to wait on some other
-            // thread.
-            #[cfg(parallel_compiler)]
-            {
-                let result = latch.wait_on(tcx, span);
+            let cached = tcx.try_get_cached::<Q, _, _, _>(
+                key,
+                |(value, index)| (value.clone(), index),
+                |_, _| panic!("value must be in cache after waiting"),
+            );
 
-                if let Err(cycle) = result {
-                    return TryGetJob::Cycle(Q::handle_cycle_error(tcx, cycle));
-                }
+            if let Some(prof_timer) = _query_blocked_prof_timer.take() {
+                prof_timer.finish_with_query_invocation_id(cached.1.into());
             }
+
+            return TryGetJob::JobCompleted(cached);
         }
     }
 
@@ -269,6 +245,7 @@ pub(super) enum TryGetJob<'a, 'tcx, D: QueryDescription<'tcx>> {
     /// The query was already completed.
     /// Returns the result of the query and its dep-node index
     /// if it succeeded or a cycle error if it failed.
+    #[cfg(parallel_compiler)]
     JobCompleted((D::Value, DepNodeIndex)),
 
     /// Trying to execute the query resulted in a cycle.
@@ -396,13 +373,78 @@ impl<'tcx> TyCtxt<'tcx> {
         eprintln!("end of query stack");
     }
 
+    /// Checks if the query is already computed and in the cache.
+    /// It returns the shard index and a lock guard to the shard,
+    /// which will be used if the query is not in the cache and we need
+    /// to compute it.
+    #[inline(always)]
+    fn try_get_cached<Q, R, OnHit, OnMiss>(
+        self,
+        key: Q::Key,
+        // `on_hit` can be called while holding a lock to the query cache
+        on_hit: OnHit,
+        on_miss: OnMiss,
+    ) -> R
+    where
+        Q: QueryDescription<'tcx> + 'tcx,
+        OnHit: FnOnce(&Q::Value, DepNodeIndex) -> R,
+        OnMiss: FnOnce(Q::Key, QueryLookup<'tcx, Q>) -> R,
+    {
+        let cache = Q::query_cache(self);
+
+        // We compute the key's hash once and then use it for both the
+        // shard lookup and the hashmap lookup. This relies on the fact
+        // that both of them use `FxHasher`.
+        let mut state = FxHasher::default();
+        key.hash(&mut state);
+        let key_hash = state.finish();
+
+        let shard = cache.get_shard_index_by_hash(key_hash);
+        let mut lock_guard = cache.get_shard_by_index(shard).lock();
+        let lock = &mut *lock_guard;
+
+        let result = lock.results.raw_entry().from_key_hashed_nocheck(key_hash, &key);
+
+        if let Some((_, value)) = result {
+            if unlikely!(self.prof.enabled()) {
+                self.prof.query_cache_hit(value.index.into());
+            }
+
+            on_hit(&value.value, value.index)
+        } else {
+            on_miss(key, QueryLookup { lock: lock_guard, shard })
+        }
+    }
+
     #[inline(never)]
-    pub(super) fn get_query<Q: QueryDescription<'tcx>>(self, span: Span, key: Q::Key) -> Q::Value {
+    pub(super) fn get_query<Q: QueryDescription<'tcx> + 'tcx>(
+        self,
+        span: Span,
+        key: Q::Key,
+    ) -> Q::Value {
         debug!("ty::query::get_query<{}>(key={:?}, span={:?})", Q::NAME, key, span);
 
-        let job = match JobOwner::try_get(self, span, &key) {
+        self.try_get_cached::<Q, _, _, _>(
+            key,
+            |value, index| {
+                self.dep_graph.read_index(index);
+                value.clone()
+            },
+            |key, lookup| self.try_execute_query(span, key, lookup),
+        )
+    }
+
+    #[inline(always)]
+    pub(super) fn try_execute_query<Q: QueryDescription<'tcx>>(
+        self,
+        span: Span,
+        key: Q::Key,
+        lookup: QueryLookup<'tcx, Q>,
+    ) -> Q::Value {
+        let job = match JobOwner::try_start(self, span, &key, lookup) {
             TryGetJob::NotYetStarted(job) => job,
             TryGetJob::Cycle(result) => return result,
+            #[cfg(parallel_compiler)]
             TryGetJob::JobCompleted((v, index)) => {
                 self.dep_graph.read_index(index);
                 return v;
@@ -615,7 +657,7 @@ impl<'tcx> TyCtxt<'tcx> {
     /// side-effects -- e.g., in order to report errors for erroneous programs.
     ///
     /// Note: The optimization is only available during incr. comp.
-    pub(super) fn ensure_query<Q: QueryDescription<'tcx>>(self, key: Q::Key) -> () {
+    pub(super) fn ensure_query<Q: QueryDescription<'tcx> + 'tcx>(self, key: Q::Key) -> () {
         if Q::EVAL_ALWAYS {
             let _ = self.get_query::<Q>(DUMMY_SP, key);
             return;
@@ -643,14 +685,30 @@ impl<'tcx> TyCtxt<'tcx> {
     }
 
     #[allow(dead_code)]
-    fn force_query<Q: QueryDescription<'tcx>>(self, key: Q::Key, span: Span, dep_node: DepNode) {
+    fn force_query<Q: QueryDescription<'tcx> + 'tcx>(
+        self,
+        key: Q::Key,
+        span: Span,
+        dep_node: DepNode,
+    ) {
         // We may be concurrently trying both execute and force a query.
         // Ensure that only one of them runs the query.
-        let job = match JobOwner::try_get(self, span, &key) {
-            TryGetJob::NotYetStarted(job) => job,
-            TryGetJob::Cycle(_) | TryGetJob::JobCompleted(_) => return,
-        };
-        self.force_query_with_job::<Q>(key, job, dep_node);
+
+        self.try_get_cached::<Q, _, _, _>(
+            key,
+            |_, _| {
+                // Cache hit, do nothing
+            },
+            |key, lookup| {
+                let job = match JobOwner::try_start(self, span, &key, lookup) {
+                    TryGetJob::NotYetStarted(job) => job,
+                    TryGetJob::Cycle(_) => return,
+                    #[cfg(parallel_compiler)]
+                    TryGetJob::JobCompleted(_) => return,
+                };
+                self.force_query_with_job::<Q>(key, job, dep_node);
+            },
+        );
     }
 }