Skip to content

Experiment: Only track fingerprints for queries with reconstructible dep-nodes. #118667

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

Closed
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
10 changes: 5 additions & 5 deletions compiler/rustc_middle/src/query/mod.rs
Original file line number Diff line number Diff line change
@@ -1066,7 +1066,7 @@ rustc_queries! {
"const-evaluating + checking `{}`",
key.value.display(tcx)
}
cache_on_disk_if { true }
// cache_on_disk_if { true } Cannot cache this anymore
}

/// Evaluates const items or anonymous constants
@@ -1081,7 +1081,7 @@ rustc_queries! {
"simplifying constant for the type system `{}`",
key.value.display(tcx)
}
cache_on_disk_if { true }
// cache_on_disk_if { true } Cannot cache this anymore
}

/// Evaluate a constant and convert it to a type level constant or
@@ -1148,7 +1148,7 @@ rustc_queries! {
/// look up the correct symbol name of instances from upstream crates.
query symbol_name(key: ty::Instance<'tcx>) -> ty::SymbolName<'tcx> {
desc { "computing the symbol for `{}`", key }
cache_on_disk_if { true }
// cache_on_disk_if { true } Cannot cache this anymore
}

query def_kind(def_id: DefId) -> DefKind {
@@ -1284,7 +1284,7 @@ rustc_queries! {
query codegen_select_candidate(
key: (ty::ParamEnv<'tcx>, ty::TraitRef<'tcx>)
) -> Result<&'tcx ImplSource<'tcx, ()>, CodegenObligationError> {
cache_on_disk_if { true }
// cache_on_disk_if { true } Cannot cache this anymore
desc { |tcx| "computing candidate for `{}`", key.1 }
}

@@ -1894,7 +1894,7 @@ rustc_queries! {
}

query unused_generic_params(key: ty::InstanceDef<'tcx>) -> UnusedGenericParams {
cache_on_disk_if { key.def_id().is_local() }
// cache_on_disk_if { key.def_id().is_local() } Cannot cache this anymore
desc {
|tcx| "determining which generic parameters are unused by `{}`",
tcx.def_path_str(key.def_id())
171 changes: 94 additions & 77 deletions compiler/rustc_query_system/src/dep_graph/graph.rs
Original file line number Diff line number Diff line change
@@ -133,7 +133,7 @@ impl<D: Deps> DepGraph<D> {
let colors = DepNodeColorMap::new(prev_graph_node_count);

// Instantiate a dependy-less node only once for anonymous queries.
let _green_node_index = current.intern_new_node(
let _green_node_index = current.alloc_new_node(
profiler,
DepNode { kind: D::DEP_KIND_NULL, hash: current.anon_id_seed.into() },
EdgesVec::new(),
@@ -272,6 +272,7 @@ impl<D: Deps> DepGraph<D> {
D::with_deps(TaskDepsRef::Forbid, op)
}

// FIXME(sparse_fps): Document
#[inline(always)]
pub fn with_task<Ctxt: HasDepContext<Deps = D>, A: Debug, R>(
&self,
@@ -287,6 +288,7 @@ impl<D: Deps> DepGraph<D> {
}
}

// FIXME(sparse_fps): Document
pub fn with_anon_task<Tcx: DepContext<Deps = D>, OP, R>(
&self,
cx: Tcx,
@@ -297,7 +299,7 @@ impl<D: Deps> DepGraph<D> {
OP: FnOnce() -> R,
{
match self.data() {
Some(data) => data.with_anon_task(cx, dep_kind, op),
Some(data) => data.with_anon_task(cx, dep_kind, true, op),
None => (op(), self.next_virtual_depnode_index()),
}
}
@@ -395,61 +397,71 @@ impl<D: Deps> DepGraphData<D> {
(result, dep_node_index)
}

// FIXME(sparse_fps): Document
/// Executes something within an "anonymous" task, that is, a task the
/// `DepNode` of which is determined by the list of inputs it read from.
pub(crate) fn with_anon_task<Tcx: DepContext<Deps = D>, OP, R>(
&self,
cx: Tcx,
dep_kind: DepKind,
intern: bool,
op: OP,
) -> (R, DepNodeIndex)
where
OP: FnOnce() -> R,
{
debug_assert!(!cx.is_eval_always(dep_kind));

let task_deps = Lock::new(TaskDeps::default());
let result = D::with_deps(TaskDepsRef::Allow(&task_deps), op);
let task_deps = task_deps.into_inner();
let task_deps = task_deps.reads;

let dep_node_index = match task_deps.len() {
0 => {
// Because the dep-node id of anon nodes is computed from the sets of its
// dependencies we already know what the ID of this dependency-less node is
// going to be (i.e. equal to the precomputed
// `SINGLETON_DEPENDENCYLESS_ANON_NODE`). As a consequence we can skip creating
// a `StableHasher` and sending the node through interning.
DepNodeIndex::SINGLETON_DEPENDENCYLESS_ANON_NODE
}
1 => {
// When there is only one dependency, don't bother creating a node.
task_deps[0]
}
_ => {
// The dep node indices are hashed here instead of hashing the dep nodes of the
// dependencies. These indices may refer to different nodes per session, but this isn't
// a problem here because we that ensure the final dep node hash is per session only by
// combining it with the per session random number `anon_id_seed`. This hash only need
// to map the dependencies to a single value on a per session basis.
let mut hasher = StableHasher::new();
task_deps.hash(&mut hasher);

let target_dep_node = DepNode {
kind: dep_kind,
// Fingerprint::combine() is faster than sending Fingerprint
// through the StableHasher (at least as long as StableHasher
// is so slow).
hash: self.current.anon_id_seed.combine(hasher.finish()).into(),
};

self.current.intern_new_node(
cx.profiler(),
target_dep_node,
task_deps,
Fingerprint::ZERO,
)
let dep_node_index = if intern {
// FIXME(sparse_fp): what is this assertion about?
debug_assert!(!cx.is_eval_always(dep_kind));

match task_deps.len() {
0 => {
// Because the dep-node id of anon nodes is computed from the sets of its
// dependencies we already know what the ID of this dependency-less node is
// going to be (i.e. equal to the precomputed
// `SINGLETON_DEPENDENCYLESS_ANON_NODE`). As a consequence we can skip creating
// a `StableHasher` and sending the node through interning.
DepNodeIndex::SINGLETON_DEPENDENCYLESS_ANON_NODE
}
1 => {
// When there is only one dependency, don't bother creating a node.
task_deps[0]
}
_ => {
// The dep node indices are hashed here instead of hashing the dep nodes of the
// dependencies. These indices may refer to different nodes per session, but this isn't
// a problem here because we that ensure the final dep node hash is per session only by
// combining it with the per session random number `anon_id_seed`. This hash only need
// to map the dependencies to a single value on a per session basis.
let mut hasher = StableHasher::new();
task_deps.hash(&mut hasher);
dep_kind.hash(&mut hasher);

let dedup_fingerprint: Fingerprint = hasher.finish();

match self
.current
.interned_node_to_index
.lock_shard_by_value(&dedup_fingerprint)
.entry(dedup_fingerprint)
{
Entry::Occupied(entry) => *entry.get(),
Entry::Vacant(entry) => {
let dep_node_index =
self.current.alloc_anon_node(cx.profiler(), dep_kind, task_deps);
entry.insert(dep_node_index);
dep_node_index
}
}
}
}
} else {
self.current.alloc_anon_node(cx.profiler(), dep_kind, task_deps)
};

(result, dep_node_index)
@@ -616,18 +628,20 @@ impl<D: Deps> DepGraph<D> {
}

impl<D: Deps> DepGraphData<D> {
#[inline]
fn dep_node_index_of_opt(&self, dep_node: &DepNode) -> Option<DepNodeIndex> {
if let Some(prev_index) = self.previous.node_to_index_opt(dep_node) {
self.current.prev_index_to_index.lock()[prev_index]
} else {
self.current.new_node_to_index.lock_shard_by_value(dep_node).get(dep_node).copied()
}
}
// #[inline]
// fn dep_node_index_of_opt(&self, dep_node: &DepNode) -> Option<DepNodeIndex> {
// if let Some(prev_index) = self.previous.node_to_index_opt(dep_node) {
// self.current.prev_index_to_index.lock()[prev_index]
// } else {
// self.current.interned_node_to_index.lock_shard_by_value(dep_node).get(dep_node).copied()
// }
// }

#[inline]
fn dep_node_exists(&self, dep_node: &DepNode) -> bool {
self.dep_node_index_of_opt(dep_node).is_some()
fn dep_node_exists(&self, _dep_node: &DepNode) -> bool {
// FIXME(sparse_fps): bring back assertions
//self.dep_node_index_of_opt(dep_node).is_some()
false
}

fn node_color(&self, dep_node: &DepNode) -> Option<DepNodeColor> {
@@ -1071,7 +1085,7 @@ rustc_index::newtype_index! {
/// first, and `data` second.
pub(super) struct CurrentDepGraph<D: Deps> {
encoder: Steal<GraphEncoder<D>>,
new_node_to_index: Sharded<FxHashMap<DepNode, DepNodeIndex>>,
interned_node_to_index: Sharded<FxHashMap<Fingerprint, DepNodeIndex>>,
prev_index_to_index: Lock<IndexVec<SerializedDepNodeIndex, Option<DepNodeIndex>>>,

/// This is used to verify that fingerprints do not change between the creation of a node
@@ -1152,7 +1166,7 @@ impl<D: Deps> CurrentDepGraph<D> {
record_graph,
record_stats,
)),
new_node_to_index: Sharded::new(|| {
interned_node_to_index: Sharded::new(|| {
FxHashMap::with_capacity_and_hasher(
new_node_count_estimate / sharded::shards(),
Default::default(),
@@ -1182,29 +1196,30 @@ impl<D: Deps> CurrentDepGraph<D> {
/// Writes the node to the current dep-graph and allocates a `DepNodeIndex` for it.
/// Assumes that this is a node that has no equivalent in the previous dep-graph.
#[inline(always)]
fn intern_new_node(
fn alloc_new_node(
&self,
profiler: &SelfProfilerRef,
key: DepNode,
edges: EdgesVec,
current_fingerprint: Fingerprint,
) -> DepNodeIndex {
let dep_node_index = match self.new_node_to_index.lock_shard_by_value(&key).entry(key) {
Entry::Occupied(entry) => *entry.get(),
Entry::Vacant(entry) => {
let dep_node_index =
self.encoder.borrow().send(profiler, key, current_fingerprint, edges);
entry.insert(dep_node_index);
dep_node_index
}
};

let dep_node_index = self.encoder.borrow().send(profiler, key, current_fingerprint, edges);
#[cfg(debug_assertions)]
self.record_edge(dep_node_index, key, current_fingerprint);

dep_node_index
}

#[inline(always)]
fn alloc_anon_node(
&self,
profiler: &SelfProfilerRef,
dep_kind: DepKind,
edges: EdgesVec,
) -> DepNodeIndex {
self.encoder.borrow().send_anon_node(profiler, dep_kind, edges)
}

fn intern_node(
&self,
profiler: &SelfProfilerRef,
@@ -1262,7 +1277,7 @@ impl<D: Deps> CurrentDepGraph<D> {
let fingerprint = fingerprint.unwrap_or(Fingerprint::ZERO);

// This is a new node: it didn't exist in the previous compilation session.
let dep_node_index = self.intern_new_node(profiler, key, edges, fingerprint);
let dep_node_index = self.alloc_new_node(profiler, key, edges, fingerprint);

(dep_node_index, None)
}
@@ -1274,7 +1289,8 @@ impl<D: Deps> CurrentDepGraph<D> {
prev_graph: &SerializedDepGraph,
prev_index: SerializedDepNodeIndex,
) -> DepNodeIndex {
self.debug_assert_not_in_new_nodes(prev_graph, prev_index);
// FIXME(sparse_fp): restore assertions
// self.debug_assert_not_in_new_nodes(prev_graph, prev_index);

let mut prev_index_to_index = self.prev_index_to_index.lock();

@@ -1296,18 +1312,19 @@ impl<D: Deps> CurrentDepGraph<D> {
}
}

#[inline]
fn debug_assert_not_in_new_nodes(
&self,
prev_graph: &SerializedDepGraph,
prev_index: SerializedDepNodeIndex,
) {
let node = &prev_graph.index_to_node(prev_index);
debug_assert!(
!self.new_node_to_index.lock_shard_by_value(node).contains_key(node),
"node from previous graph present in new node collection"
);
}
// FIXME(sparse_fp): restore assertions
// #[inline]
// fn debug_assert_not_in_new_nodes(
// &self,
// prev_graph: &SerializedDepGraph,
// prev_index: SerializedDepNodeIndex,
// ) {
// let node = &prev_graph.index_to_node(prev_index);
// debug_assert!(
// !self.interned_node_to_index.lock_shard_by_value(node).contains_key(node),
// "node from previous graph present in new node collection"
// );
// }
}

#[derive(Debug, Clone, Copy)]
85 changes: 69 additions & 16 deletions compiler/rustc_query_system/src/dep_graph/serialized.rs
Original file line number Diff line number Diff line change
@@ -259,8 +259,16 @@ impl SerializedDepGraph {
.map(|_| UnhashMap::with_capacity_and_hasher(d.read_u32() as usize, Default::default()))
.collect();

let anon_node_fingerprint: PackedFingerprint = anon_node_fingerprint().into();

for (idx, node) in nodes.iter_enumerated() {
index[node.kind.as_usize()].insert(node.hash, idx);
// FIXME(sparse_fps): Filter out anon nodes from the reverse index
// by looking for the special DepNode::hash. This
// is a hack, we should not store depnodes and
// fingerprints for these to begin with.
if node.hash != anon_node_fingerprint {
index[node.kind.as_usize()].insert(node.hash, idx);
}
}

SerializedDepGraph { nodes, fingerprints, edge_list_indices, edge_list_data, index }
@@ -311,7 +319,14 @@ impl<D: Deps> SerializedNodeHeader<D> {
fn new(node_info: &NodeInfo) -> Self {
debug_assert_eq!(Self::TOTAL_BITS, Self::LEN_BITS + Self::WIDTH_BITS + Self::KIND_BITS);

let NodeInfo { node, fingerprint, edges } = node_info;
let (node, fingerprint, edges) = match node_info {
NodeInfo::Full(node, fingerprint, edges) => (*node, *fingerprint, edges),
NodeInfo::Anon(kind, edges) => (
DepNode { kind: *kind, hash: anon_node_fingerprint().into() },
anon_node_fingerprint(),
edges,
),
};

let mut head = node.kind.as_inner();

@@ -336,10 +351,10 @@ impl<D: Deps> SerializedNodeHeader<D> {
#[cfg(debug_assertions)]
{
let res = Self { bytes, _marker: PhantomData };
assert_eq!(node_info.fingerprint, res.fingerprint());
assert_eq!(node_info.node, res.node());
assert_eq!(node_info.fingerprint(), res.fingerprint());
assert_eq!(node_info.node(), res.node());
if let Some(len) = res.len() {
assert_eq!(node_info.edges.len(), len);
assert_eq!(node_info.edges().len(), len);
}
}
Self { bytes, _marker: PhantomData }
@@ -394,23 +409,45 @@ impl<D: Deps> SerializedNodeHeader<D> {
}

#[derive(Debug)]
struct NodeInfo {
node: DepNode,
fingerprint: Fingerprint,
edges: EdgesVec,
enum NodeInfo {
Full(DepNode, Fingerprint, EdgesVec),
Anon(DepKind, EdgesVec),
}

impl NodeInfo {
#[cfg(debug_assertions)]
fn fingerprint(&self) -> Fingerprint {
match *self {
NodeInfo::Full(_, fingerprint, _) => fingerprint,
NodeInfo::Anon(_, _) => anon_node_fingerprint(),
}
}

#[inline]
fn node(&self) -> DepNode {
match *self {
NodeInfo::Full(node, _, _) => node,
NodeInfo::Anon(kind, _) => DepNode { kind, hash: anon_node_fingerprint().into() },
}
}

#[inline]
fn edges(&self) -> &EdgesVec {
match self {
NodeInfo::Full(_, _, edges) | NodeInfo::Anon(_, edges) => edges,
}
}

fn encode<D: Deps>(&self, e: &mut FileEncoder) {
let header = SerializedNodeHeader::<D>::new(self);
e.write_array(header.bytes);

if header.len().is_none() {
e.emit_usize(self.edges.len());
e.emit_usize(self.edges().len());
}

let bytes_per_index = header.bytes_per_index();
for node_index in self.edges.iter() {
for node_index in self.edges().iter() {
e.write_with(|dest| {
*dest = node_index.as_u32().to_le_bytes();
bytes_per_index
@@ -455,20 +492,20 @@ impl<D: Deps> EncoderState<D> {
) -> DepNodeIndex {
let index = DepNodeIndex::new(self.total_node_count);
self.total_node_count += 1;
self.kind_stats[node.node.kind.as_usize()] += 1;
self.kind_stats[node.node().kind.as_usize()] += 1;

let edge_count = node.edges.len();
let edge_count = node.edges().len();
self.total_edge_count += edge_count;

if let Some(record_graph) = &record_graph {
// Do not ICE when a query is called from within `with_query`.
if let Some(record_graph) = &mut record_graph.try_lock() {
record_graph.push(index, node.node, &node.edges);
record_graph.push(index, node.node(), node.edges());
}
}

if let Some(stats) = &mut self.stats {
let kind = node.node.kind;
let kind = node.node().kind;

let stat = stats.entry(kind).or_insert(Stat { kind, node_counter: 0, edge_counter: 0 });
stat.node_counter += 1;
@@ -597,7 +634,18 @@ impl<D: Deps> GraphEncoder<D> {
edges: EdgesVec,
) -> DepNodeIndex {
let _prof_timer = profiler.generic_activity("incr_comp_encode_dep_graph");
let node = NodeInfo { node, fingerprint, edges };
let node = NodeInfo::Full(node, fingerprint, edges);
self.status.lock().encode_node(&node, &self.record_graph)
}

pub(crate) fn send_anon_node(
&self,
profiler: &SelfProfilerRef,
dep_kind: DepKind,
edges: EdgesVec,
) -> DepNodeIndex {
let _prof_timer = profiler.generic_activity("incr_comp_encode_dep_graph");
let node = NodeInfo::Anon(dep_kind, edges);
self.status.lock().encode_node(&node, &self.record_graph)
}

@@ -606,3 +654,8 @@ impl<D: Deps> GraphEncoder<D> {
self.status.into_inner().finish(profiler)
}
}

#[inline]
pub fn anon_node_fingerprint() -> Fingerprint {
Fingerprint::new(0x01234567_89abcdef, 0xfedcba98_7654321)
}
40 changes: 23 additions & 17 deletions compiler/rustc_query_system/src/query/plumbing.rs
Original file line number Diff line number Diff line change
@@ -491,7 +491,10 @@ where
Q: QueryConfig<Qcx>,
Qcx: QueryContext,
{
if !query.anon() && !query.eval_always() {
let is_reconstructible = !query.anon()
&& qcx.dep_context().dep_kind_info(query.dep_kind()).force_from_dep_node.is_some();

if is_reconstructible && !query.eval_always() {
// `to_dep_node` is expensive for some `DepKind`s.
let dep_node =
dep_node_opt.get_or_insert_with(|| query.construct_dep_node(*qcx.dep_context(), &key));
@@ -510,23 +513,26 @@ where

let (result, dep_node_index) =
qcx.start_query(job_id, query.depth_limit(), Some(&diagnostics), || {
if query.anon() {
return dep_graph_data.with_anon_task(*qcx.dep_context(), query.dep_kind(), || {
query.compute(qcx, key)
});
if is_reconstructible {
// `to_dep_node` is expensive for some `DepKind`s.
let dep_node = dep_node_opt
.unwrap_or_else(|| query.construct_dep_node(*qcx.dep_context(), &key));

dep_graph_data.with_task(
dep_node,
(qcx, query),
key,
|(qcx, query), key| query.compute(qcx, key),
query.hash_result(),
)
} else {
dep_graph_data.with_anon_task(
*qcx.dep_context(),
query.dep_kind(),
query.anon(),
|| query.compute(qcx, key),
)
}

// `to_dep_node` is expensive for some `DepKind`s.
let dep_node =
dep_node_opt.unwrap_or_else(|| query.construct_dep_node(*qcx.dep_context(), &key));

dep_graph_data.with_task(
dep_node,
(qcx, query),
key,
|(qcx, query), key| query.compute(qcx, key),
query.hash_result(),
)
});

prof_timer.finish_with_query_invocation_id(dep_node_index.into());