blob: 0c824f231028892675d3c846563bde496feb0fd4 [file] [log] [blame]
// Copyright (c) 2020, the Dart project authors. Please see the AUTHORS file
// for details. All rights reserved. Use of this source code is governed by a
// BSD-style license that can be found in the LICENSE file.
/// This library contains utilities for reading and analyzing snapshot profiles
/// produced by `--write-v8-snapshot-profile-to` VM flag.
library vm_snapshot_analysis.v8_profile;
import 'package:meta/meta.dart';
import 'package:vm_snapshot_analysis/src/dominators.dart' as dominators;
import 'package:vm_snapshot_analysis/name.dart';
import 'package:vm_snapshot_analysis/program_info.dart';
/// This class represents snapshot graph.
/// Note that we do not eagerly deserialize the graph, instead we provide helper
/// methods and wrapper objects to work with serialized representation.
class Snapshot {
final Meta meta;
final int nodeCount;
final int edgeCount;
/// Serialized flat representation of nodes in the graph. Each node occupies
/// [meta.nodeFieldCount] consecutive elements of the list.
final List _nodes;
/// Serialized flat representation of edges between nodes. Each edge occupies
/// [meta.edgeFieldCount] consecutive elements of the list. All outgoing edges
/// for a node are serialized consecutively, number of outgoing edges is given
/// by the value at index [meta.nodeEdgeCountIndex] inside the node.
final List _edges;
/// Auxiliary array which gives starting index of edges (in the [_edges] list)
/// for the given node index.
final List<int> _edgesStartIndexForNode;
List<int> _dominators;
final List strings;
Snapshot._(this.meta, this.nodeCount, this.edgeCount, this._nodes,
this._edges, this.strings, this._edgesStartIndexForNode);
/// Return node with the given index.
Node nodeAt(int index) {
assert(index >= 0, 'Node index should be positive: $index');
return Node._(snapshot: this, index: index);
/// Return all nodes in the snapshot.
Iterable<Node> get nodes => Iterable.generate(nodeCount, nodeAt);
/// Return dominator node for the given node [n].
Node dominatorOf(Node n) {
_dominators ??= _computeDominators(this);
return nodeAt(_dominators[n.index]);
/// Returns true if the given JSON object is likely to be a serialized
/// snapshot using V8 heap snapshot format.
static bool isV8HeapSnapshot(Object m) =>
m is Map<String, dynamic> && m.containsKey('snapshot');
/// Construct [Snapshot] object from the given JSON object.
factory Snapshot.fromJson(Map<String, dynamic> m) {
// Extract meta information first.
final meta = Meta._fromJson(m['snapshot']['meta']);
final nodes = m['nodes'];
// Build an array of starting indexes of edges for each node.
final edgesStartIndexForNode = <int>[0];
int nextStartIndex = 0;
for (var i = meta.nodeEdgeCountIndex;
i < nodes.length;
i += meta.nodeFieldCount) {
nextStartIndex += nodes[i];
return Snapshot._(
/// Meta-information about the serialized snapshot.
/// Describes the structure of serialized nodes and edges by giving indexes of
/// the various fields.
class Meta {
final int nodeTypeIndex;
final int nodeNameIndex;
final int nodeIdIndex;
final int nodeSelfSizeIndex;
final int nodeEdgeCountIndex;
final int nodeFieldCount;
final int edgeTypeIndex;
final int edgeNameOrIndexIndex;
final int edgeToNodeIndex;
final int edgeFieldCount;
final List<String> nodeTypes;
final List<String> edgeTypes;
factory Meta._fromJson(Map<String, dynamic> m) {
final nodeFields = m['node_fields'];
final nodeTypes = m['node_types'].first.cast<String>();
final edgeFields = m['edge_fields'];
final edgeTypes = m['edge_types'].first.cast<String>();
return Meta._(
nodeTypeIndex: nodeFields.indexOf('type'),
nodeNameIndex: nodeFields.indexOf('name'),
nodeIdIndex: nodeFields.indexOf('id'),
nodeSelfSizeIndex: nodeFields.indexOf('self_size'),
nodeEdgeCountIndex: nodeFields.indexOf('edge_count'),
nodeFieldCount: nodeFields.length,
edgeTypeIndex: edgeFields.indexOf('type'),
edgeNameOrIndexIndex: edgeFields.indexOf('name_or_index'),
edgeToNodeIndex: edgeFields.indexOf('to_node'),
edgeFieldCount: edgeFields.length,
nodeTypes: nodeTypes,
edgeTypes: edgeTypes);
/// Edge from [Node] to [Node] in the [Snapshot] graph.
class Edge {
final Snapshot snapshot;
/// Index of this [Edge] within the [snapshot].
final int index;
Edge._({this.snapshot, this.index});
String get type => snapshot
.meta.edgeTypes[snapshot._edges[_offset + snapshot.meta.edgeTypeIndex]];
Node get target {
return Node._(
snapshot: snapshot,
index: snapshot._edges[_offset + snapshot.meta.edgeToNodeIndex] ~/
String get name {
final nameOrIndex =
snapshot._edges[_offset + snapshot.meta.edgeNameOrIndexIndex];
return type == 'property' ? snapshot.strings[nameOrIndex] : '@$nameOrIndex';
String toString() {
final nameOrIndex =
snapshot._edges[_offset + snapshot.meta.edgeNameOrIndexIndex];
return {
'type': type,
type == 'property' ? snapshot.strings[nameOrIndex] : nameOrIndex,
'toNode': target.index,
/// Offset into [Snapshot._edges] list at which this edge begins.
int get _offset => index * snapshot.meta.edgeFieldCount;
/// Node in the [Snapshot] graph.
class Node {
final Snapshot snapshot;
/// Index of this [Node] within the [snapshot].
final int index;
Node._({this.snapshot, this.index});
int get edgeCount =>
snapshot._nodes[_offset + snapshot.meta.nodeEdgeCountIndex];
String get type => snapshot
.meta.nodeTypes[snapshot._nodes[_offset + snapshot.meta.nodeTypeIndex]];
String get name =>
snapshot.strings[snapshot._nodes[_offset + snapshot.meta.nodeNameIndex]];
int get selfSize =>
snapshot._nodes[_offset + snapshot.meta.nodeSelfSizeIndex];
int get id => snapshot._nodes[_offset + snapshot.meta.nodeIdIndex];
/// Returns all outgoing edges for this node.
Iterable<Edge> get edges sync* {
var firstEdgeIndex = snapshot._edgesStartIndexForNode[index];
for (var i = 0, n = edgeCount; i < n; i++) {
yield Edge._(snapshot: snapshot, index: firstEdgeIndex + i);
String toString() {
return {
'type': type,
'name': name,
'id': id,
'selfSize': selfSize,
'edges': edges.toList(),
/// Returns the target of an outgoing edge with the given name (if any).
Node operator [](String edgeName) => this
.firstWhere((e) => == edgeName, orElse: () => null)
bool operator ==(Object other) {
return other is Node && other.index == index;
int get hashCode => this.index.hashCode;
/// Offset into [Snapshot._nodes] list at which this node begins.
int get _offset => index * snapshot.meta.nodeFieldCount;
/// Class representing information about V8 snapshot profile in relation
/// to a [ProgramInfo] structure that was derived from it.
class SnapshotInfo {
final Snapshot snapshot;
final List<ProgramInfoNode> infoNodes;
final Map<int, int> _ownerOf;
SnapshotInfo._(this.snapshot, this.infoNodes, this._ownerOf);
ProgramInfoNode ownerOf(Node node) =>
infoNodes[_ownerOf[node.index] ?? ProgramInfo.unknownId];
ProgramInfo toProgramInfo(Snapshot snap,
{bool collapseAnonymousClosures = false}) {
return _ProgramInfoBuilder(
collapseAnonymousClosures: collapseAnonymousClosures)
class _ProgramInfoBuilder {
final bool collapseAnonymousClosures;
final program = ProgramInfo();
final List<ProgramInfoNode> infoNodes = [];
/// Mapping between snapshot [Node] index and id of [ProgramInfoNode] which
/// own this node.
final Map<int, int> ownerOf = {};
/// Mapping between snapshot [Node] indices and corresponding
/// [ProgramInfoNode] objects. Note that multiple snapshot nodes might be
/// mapped to a single [ProgramInfoNode] (e.g. when anonymous closures are
/// collapsed).
final Map<int, ProgramInfoNode> infoNodeByIndex = {};
// Mapping between package names and corresponding [ProgramInfoNode] objects
// representing those packages.
final Map<String, ProgramInfoNode> infoNodeForPackage = {};
/// Owners of some [Node] are determined by the program structure and not
/// by their reachability through the graph. For example, an owner of a
/// function is a class that contains it, even though the function can
/// also be reachable from another function through object pool.
final Set<int> nodesWithFrozenOwner = {};
/// Cache used to optimize common ancestor operation on [ProgramInfoNode] ids.
/// See [findCommonAncestor] method.
final Map<int, int> commonAncestorCache = {};
/// Recover [ProgramInfo] structure from the snapshot profile.
/// This is done via a simple graph traversal: first all nodes representing
/// objects with clear ownership (like libraries, classes, functions) are
/// discovered and corresponding [ProgramInfoNode] objects are created for
/// them. Then the rest of the snapshot is attributed to one of these nodes
/// based on reachability (ignoring reachability from normal snapshot roots):
/// let `R(n)` be a set of [ProgramInfoNode] objects from which a given
/// snapshot node `n` is reachable. Then we define an owner of `n` to be
/// a lowest common ancestor of all nodes in `R(n)`.
/// Nodes which are not reachable from any normal [ProgramInfoNode] are
/// attributed to special `@unknown` [ProgramInfoNode].
ProgramInfo build(Snapshot snap) {
// Create ProgramInfoNode for every snapshot node representing an element
// of the program structure (e.g. a library, a class, a function).
// Propagate the ownership information across the edges.
final worklist = ownerOf.keys.toList();
while (worklist.isNotEmpty) {
final node = snap.nodeAt(worklist.removeLast());
final sourceOwner = ownerOf[node.index];
for (var e in node.edges) {
final target =;
if (!nodesWithFrozenOwner.contains(target.index)) {
final targetOwner = ownerOf[target.index];
final updatedOwner = findCommonAncestor(sourceOwner, targetOwner);
if (updatedOwner != targetOwner) {
ownerOf[target.index] = updatedOwner;
// Now attribute sizes from the snapshot to nodes that own them.
for (var node in snap.nodes) {
if (node.selfSize > 0) {
final owner = infoNodes[ownerOf[node.index] ?? ProgramInfo.unknownId];
owner.size = (owner.size ?? 0) + node.selfSize;
program.snapshotInfo = SnapshotInfo._(snap, infoNodes, ownerOf);
return program;
ProgramInfoNode getInfoNodeFor(Node node) {
var info = infoNodeByIndex[node.index];
if (info == null) {
info = createInfoNodeFor(node);
if (info != null) {
// Snapshot nodes which represent the program structure can't change
// their owner during iteration - their owner is frozen and is given
// by the program structure.
// Note that [ProgramInfoNode] owns its corresponding [Snapshot] node
// because we want the size of the snapshot node to be attributed to
// the info node itself.
ownerOf[node.index] =;
// Handle some nodes specially.
switch (node.type) {
case 'Code':
// Freeze ownership of the Instructions object.
final instructions = node['<instructions>'];
ownerOf[instructions.index] =
case 'Library':
// Freeze ownership of the Script objects owned by this library.
final scripts = node['owned_scripts_'];
if (scripts != null) {
for (var e in scripts.edges) {
if ( == 'Script') {
ownerOf[] =
return info;
ProgramInfoNode createInfoNodeFor(Node node) {
switch (node.type) {
case 'Code':
var owner = node['owner_'];
if (owner.type != 'Type') {
if (owner.type == 'WeakSerializationReference') {
owner = node[':owner_'];
final ownerNode =
owner.type == 'Null' ? program.stubs : getInfoNodeFor(owner);
if (owner.type == 'Function') {
// For normal functions we just attribute Code object and all
// objects dominated by it to the function itself.
return ownerNode;
// For stubs we create a dummy functionNode that is going to own
// all objects dominated by it.
return makeInfoNode(node.index,
name:, parent: ownerNode, type: NodeType.functionNode);
case 'Function':
if ( != '<anonymous signature>') {
var owner = node['owner_'];
// Artificial nodes will not have data_ field.
if (node['data_']?.type == 'ClosureData') {
owner = node['data_']['parent_function_'];
return makeInfoNode(node.index,
parent: getInfoNodeFor(owner),
type: NodeType.functionNode);
case 'PatchClass':
return getInfoNodeFor(node['patched_class_']);
case 'Class':
if (node['library_'] != null) {
return makeInfoNode(node.index,
parent: getInfoNodeFor(node['library_']) ?? program.root,
type: NodeType.classNode);
case 'Library':
// Create fake owner node for the package which contains this library.
final packageName = packageOf(;
return makeInfoNode(node.index,
parent: packageName !=
? packageOwner(packageName)
: program.root,
type: NodeType.libraryNode);
case 'Field':
return makeInfoNode(node.index,
parent: getInfoNodeFor(node['owner_']),
type: NodeType.other);
return null;
ProgramInfoNode makeInfoNode(int index,
{@required ProgramInfoNode parent,
@required String name,
@required NodeType type}) {
assert(parent != null,
'Trying to create node of type ${type} with ${name} and no parent.');
assert(name != null);
name = Name(name).scrubbed;
if (collapseAnonymousClosures) {
name = Name.collapse(name);
final node = program.makeNode(name: name, parent: parent, type: type);
if ( == infoNodes.length) {
if (index != null) {
infoNodeByIndex[index] = node;
return node;
ProgramInfoNode packageOwner(String packageName) =>
() => makeInfoNode(null,
name: packageName,
type: NodeType.packageNode,
parent: program.root));
/// Create a single key from two node ids.
/// Note that this operation is commutative, because common ancestor of A and
/// B is the same as common ancestor of B and A.
static int ancestorCacheKey(int a, int b) {
if (a > b) {
return b << 32 | a;
} else {
return a << 32 | b;
/// Returns id of a common ancestor between [ProgramInfoNode] with [idA] and
/// [idB].
int findCommonAncestor(int idA, int idB) {
if (idA == null) {
return idB;
if (idB == null) {
return idA;
if (idA == idB) {
return idA;
// If either are shared - then result is shared.
if (idA == ProgramInfo.rootId || idB == ProgramInfo.rootId) {
return ProgramInfo.rootId;
final infoA = infoNodes[idA];
final infoB = infoNodes[idB];
final key = ancestorCacheKey(idA, idB);
var ancestor = commonAncestorCache[key];
if (ancestor == null) {
commonAncestorCache[key] =
ancestor = findCommonAncestorImpl(infoA, infoB).id;
return ancestor;
static List<ProgramInfoNode> pathToRoot(ProgramInfoNode node) {
final path = <ProgramInfoNode>[];
while (node != null) {
node = node.parent;
return path;
static ProgramInfoNode findCommonAncestorImpl(
ProgramInfoNode a, ProgramInfoNode b) {
final pathA = pathToRoot(a);
final pathB = pathToRoot(b);
var i = pathA.length - 1, j = pathB.length - 1;
while (i > 0 && j > 0 && (pathA[i - 1] == pathB[j - 1])) {
assert(pathA[i] == pathB[j]);
return pathA[i];
final bucketLegend = '''
IMPORTANT: Dart AOT snapshot is a serialized representation of Dart VM heap.
Outside of few specific cases (e.g. an object representing a library clearly
originates from the library it represents) there is no well defined relationship
between snapshot bytes and a specific method/class/library to which these
bytes can be attributed with certainty. This snapshot analysis tool tries
to attribute bytes to specific program structure elements based on their
reachability from objects with well defined origin - meaning that this analysis
has some margin of error and imprecision.
- @other bucket denotes bytes attributed to entities outside of the current
granularity. For example, when breaking down the size by method name there
might be bytes which exist outside of any specific symbol - in which case
they will be attributed to @other.
- @stubs bucket accumulates bytes attributed to stubs (pieces of machine code
produced by the VM for internal purposes).
- @shared bucket accumulates bytes shared between otherwise unrelated program
- @unknown bucket accumulates bytes which are not reachable from any program
structure nodes (usually VM internal objects).
/// Compute dominator tree of the graph.
/// The code for dominator tree computation is taken verbatim from the
/// native compiler (see runtime/vm/compiler/backend/
List<int> _computeDominators(Snapshot snap) {
final predecessors = List<Object>.filled(snap.nodeCount, null);
void addPred(int n, int p) {
if (predecessors[n] == null) {
predecessors[n] = p;
} else if (predecessors[n] is int) {
predecessors[n] = <int>[predecessors[n], p];
} else {
(predecessors[n] as List<int>).add(p);
Iterable<int> predOf(int n) sync* {
final ps = predecessors[n];
if (ps is int) {
yield ps;
} else if (ps is List<int>) {
yield* ps;
return dominators.computeDominators(
size: snap.nodeCount,
root: snap.nodes.first.index,
succ: (n) => snap.nodeAt(n) =>,
predOf: predOf,
handleEdge: addPred);