Improve performance of LimitChunkCountPlugin a lot

This commit is contained in:
Tobias Koppers 2019-09-24 16:16:36 +02:00
parent adfa484921
commit 6f1c685071
3 changed files with 427 additions and 35 deletions

View File

@ -385,7 +385,7 @@ class Chunk {
/**
*
* @param {Chunk} otherChunk the chunk to integrate with
* @param {ModuleReason} reason reason why the module is being integrated
* @param {string} reason reason why the module is being integrated
* @returns {boolean} returns true or false if integration succeeds or fails
*/
integrate(otherChunk, reason) {
@ -462,6 +462,11 @@ class Chunk {
return false;
}
/**
* @param {Chunk} a chunk
* @param {Chunk} b chunk
* @returns {boolean} true, if a is always available when b is reached
*/
const isAvailable = (a, b) => {
const queue = new Set(b.groupsIterable);
for (const chunkGroup of queue) {

View File

@ -6,8 +6,33 @@
const validateOptions = require("schema-utils");
const schema = require("../../schemas/plugins/optimize/LimitChunkCountPlugin.json");
const LazyBucketSortedSet = require("../util/LazyBucketSortedSet");
/** @typedef {import("../../declarations/plugins/optimize/LimitChunkCountPlugin").LimitChunkCountPluginOptions} LimitChunkCountPluginOptions */
/** @typedef {import("../Chunk")} Chunk */
/** @typedef {import("../Compiler")} Compiler */
/**
* @typedef {Object} ChunkCombination
* @property {boolean} deleted this is set to true when combination was removed
* @property {number} sizeDiff
* @property {number} integratedSize
* @property {Chunk} a
* @property {Chunk} b
* @property {number} aIdx
* @property {number} bIdx
* @property {number} aSize
* @property {number} bSize
*/
const addToSetMap = (map, key, value) => {
const set = map.get(key);
if (set === undefined) {
map.set(key, new Set([value]));
} else {
set.add(value);
}
};
class LimitChunkCountPlugin {
/**
@ -19,6 +44,11 @@ class LimitChunkCountPlugin {
validateOptions(schema, options, "Limit Chunk Count Plugin");
this.options = options;
}
/**
* @param {Compiler} compiler the webpack compiler
* @returns {void}
*/
apply(compiler) {
const options = this.options;
compiler.hooks.compilation.tap("LimitChunkCountPlugin", compilation => {
@ -30,47 +60,169 @@ class LimitChunkCountPlugin {
if (maxChunks < 1) return;
if (chunks.length <= maxChunks) return;
let remainingChunksToMerge = chunks.length - maxChunks;
// order chunks in a deterministic way
const orderedChunks = chunks.slice().sort((a, b) => a.compareTo(b));
const sortedExtendedPairCombinations = orderedChunks
.reduce((combinations, a, idx) => {
// create combination pairs
for (let i = 0; i < idx; i++) {
const b = orderedChunks[i];
combinations.push([b, a]);
}
return combinations;
}, [])
.map(pair => {
// extend combination pairs with size and integrated size
const a = pair[0].size(options);
const b = pair[1].size(options);
const ab = pair[0].integratedSize(pair[1], options);
return [a + b - ab, ab, pair[0], pair[1], a, b];
})
.filter(extendedPair => {
// create a lazy sorted data structure to keep all combinations
// this is large. Size = chunks * (chunks - 1) / 2
// It uses a multi layer bucket sort plus normal sort in the last layer
// It's also lazy so only accessed buckets are sorted
const combinations = new LazyBucketSortedSet(
// Layer 1: ordered by largest size benefit
c => c.sizeDiff,
(a, b) => b - a,
// Layer 2: ordered by smallest combined size
c => c.integratedSize,
(a, b) => a - b,
// Layer 3: ordered by position difference in orderedChunk (-> to be deterministic)
c => c.bIdx - c.aIdx,
(a, b) => a - b,
// Layer 4: ordered by position in orderedChunk (-> to be deterministic)
(a, b) => a.bIdx - b.bIdx
);
// we keep a mappng from chunk to all combinations
// but this mapping is not kept up-to-date with deletions
// so `deleted` flag need to be considered when iterating this
/** @type {Map<Chunk, Set<ChunkCombination>>} */
const combinationsByChunk = new Map();
orderedChunks.forEach((b, bIdx) => {
// create combination pairs with size and integrated size
for (let aIdx = 0; aIdx < bIdx; aIdx++) {
const a = orderedChunks[aIdx];
const integratedSize = a.integratedSize(b, options);
// filter pairs that do not have an integratedSize
// meaning they can NOT be integrated!
return extendedPair[1] !== false;
})
.sort((a, b) => {
// sadly javascript does an inplace sort here
// sort them by size
const diff1 = b[0] - a[0];
if (diff1 !== 0) return diff1;
const diff2 = a[1] - b[1];
if (diff2 !== 0) return diff2;
const diff3 = a[2].compareTo(b[2]);
if (diff3 !== 0) return diff3;
return a[3].compareTo(b[3]);
});
if (integratedSize === false) continue;
const pair = sortedExtendedPairCombinations[0];
const aSize = a.size(options);
const bSize = b.size(options);
const c = {
deleted: false,
sizeDiff: aSize + bSize - integratedSize,
integratedSize,
a,
b,
aIdx,
bIdx,
aSize,
bSize
};
combinations.add(c);
addToSetMap(combinationsByChunk, a, c);
addToSetMap(combinationsByChunk, b, c);
}
return combinations;
});
if (pair && pair[2].integrate(pair[3], "limit")) {
chunks.splice(chunks.indexOf(pair[3]), 1);
return true;
// list of modified chunks during this run
// combinations affected by this change are skipped to allow
// futher optimizations
/** @type {Set<Chunk>} */
const modifiedChunks = new Set();
let changed = false;
// eslint-disable-next-line no-constant-condition
loop: while (true) {
const combination = combinations.popFirst();
if (combination === undefined) break;
combination.deleted = true;
const { a, b, integratedSize } = combination;
// skip over pair when
// one of the already merged chunks is a parent of one of the chunks
if (modifiedChunks.size > 0) {
const queue = new Set(a.groupsIterable);
for (const group of b.groupsIterable) {
queue.add(group);
}
for (const group of queue) {
for (const mChunk of modifiedChunks) {
if (mChunk !== a && mChunk !== b && mChunk.isInGroup(group)) {
// This is a potential pair which needs recalculation
// We can't do that now, but it merge before following pairs
// so we leave space for it, and consider chunks as modified
// just for the worse case
remainingChunksToMerge--;
if (remainingChunksToMerge <= 0) break loop;
modifiedChunks.add(a);
modifiedChunks.add(b);
continue loop;
}
}
for (const parent of group.parentsIterable) {
queue.add(parent);
}
}
}
// merge the chunks
if (a.integrate(b, "limit")) {
chunks.splice(chunks.indexOf(b), 1);
// flag chunk a as modified as further optimization are possible for all children here
modifiedChunks.add(a);
changed = true;
remainingChunksToMerge--;
if (remainingChunksToMerge <= 0) break;
// Update all affected combinations
// delete all combination with the removed chunk
// we will use combinations with the kept chunk instead
for (const combination of combinationsByChunk.get(b)) {
if (combination.deleted) continue;
combination.deleted = true;
combinations.delete(combination);
}
// Update combinations with the kept chunk with new sizes
for (const combination of combinationsByChunk.get(a)) {
if (combination.deleted) continue;
if (combination.a === a) {
// Update size
const newIntegratedSize = a.integratedSize(
combination.b,
options
);
if (newIntegratedSize === false) {
combination.deleted = true;
combinations.delete(combination);
continue;
}
const finishUpdate = combinations.startUpdate(combination);
combination.integratedSize = newIntegratedSize;
combination.aSize = integratedSize;
combination.sizeDiff =
combination.bSize + integratedSize - newIntegratedSize;
finishUpdate();
} else if (combination.b === a) {
// Update size
const newIntegratedSize = combination.a.integratedSize(
a,
options
);
if (newIntegratedSize === false) {
combination.deleted = true;
combinations.delete(combination);
continue;
}
const finishUpdate = combinations.startUpdate(combination);
combination.integratedSize = newIntegratedSize;
combination.bSize = integratedSize;
combination.sizeDiff =
integratedSize + combination.aSize - newIntegratedSize;
finishUpdate();
}
}
}
}
if (changed) return true;
}
);
});

View File

@ -0,0 +1,235 @@
/*
MIT License http://www.opensource.org/licenses/mit-license.php
Author Tobias Koppers @sokra
*/
"use strict";
const SortableSet = require("./SortableSet");
/**
* @template T
* @template K
* Multi layer bucket sorted set
* Supports adding non-existing items (DO NOT ADD ITEM TWICE)
* Supports removing exiting items (DO NOT REMOVE ITEM NOT IN SET)
* Supports popping the first items according to defined order
* Supports iterating all items without order
* Supports updating an item in an efficient way
* Supports size property, which is the number of items
* Items are lazy partially sorted when needed
*/
class LazyBucketSortedSet {
/**
* @param {function(T): K} getKey function to get key from item
* @param {function(K, K): number} comparator comparator to sort keys
* @param {...((function(T): any) | (function(any, any): number))} args more pairs of getKey and comparator plus optional final comparator for the last layer
*/
constructor(getKey, comparator, ...args) {
this._getKey = getKey;
this._innerArgs = args;
this._leaf = args.length <= 1;
this._keys = new SortableSet(undefined, comparator);
/** @type {Map<K, LazyBucketSortedSet<T, any> | SortableSet<T>>} */
this._map = new Map();
this._unsortedItems = new Set();
this.size = 0;
}
/**
* @param {T} item an item
* @returns {void}
*/
add(item) {
this.size++;
this._unsortedItems.add(item);
}
/**
* @param {K} key key of item
* @param {T} item the item
* @returns {void}
*/
_addInternal(key, item) {
let entry = this._map.get(key);
if (entry === undefined) {
entry = this._leaf
? new SortableSet(undefined, this._innerArgs[0])
: new /** @type {any} */ (LazyBucketSortedSet)(...this._innerArgs);
this._keys.add(key);
this._map.set(key, entry);
}
entry.add(item);
}
/**
* @param {T} item an item
* @returns {void}
*/
delete(item) {
this.size--;
if (this._unsortedItems.has(item)) {
this._unsortedItems.delete(item);
return;
}
const key = this._getKey(item);
const entry = this._map.get(key);
entry.delete(item);
if (entry.size === 0) {
this._deleteKey(key);
}
}
/**
* @param {K} key key to be removed
* @returns {void}
*/
_deleteKey(key) {
this._keys.delete(key);
this._map.delete(key);
}
/**
* @returns {T | undefined} an item
*/
popFirst() {
if (this.size === 0) return undefined;
this.size--;
if (this._unsortedItems.size > 0) {
for (const item of this._unsortedItems) {
const key = this._getKey(item);
this._addInternal(key, item);
}
this._unsortedItems.clear();
}
this._keys.sort();
const key = this._keys.values().next().value;
const entry = this._map.get(key);
if (this._leaf) {
const leafEntry = /** @type {SortableSet<T>} */ (entry);
leafEntry.sort();
const item = leafEntry.values().next().value;
leafEntry.delete(item);
if (leafEntry.size === 0) {
this._deleteKey(key);
}
return item;
} else {
const nodeEntry = /** @type {LazyBucketSortedSet<T, any>} */ (entry);
const item = nodeEntry.popFirst();
if (nodeEntry.size === 0) {
this._deleteKey(key);
}
return item;
}
}
/**
* @param {T} item to be updated item
* @returns {function(true=): void} finish update
*/
startUpdate(item) {
if (this._unsortedItems.has(item)) {
return remove => {
if (remove) {
this._unsortedItems.delete(item);
this.size--;
return;
}
};
}
const key = this._getKey(item);
if (this._leaf) {
const oldEntry = /** @type {SortableSet<T>} */ (this._map.get(key));
return remove => {
if (remove) {
this.size--;
oldEntry.delete(item);
if (oldEntry.size === 0) {
this._deleteKey(key);
}
return;
}
const newKey = this._getKey(item);
if (key === newKey) {
// This flags the sortable set as unordered
oldEntry.add(item);
} else {
oldEntry.delete(item);
if (oldEntry.size === 0) {
this._deleteKey(key);
}
this._addInternal(newKey, item);
}
};
} else {
const oldEntry = /** @type {LazyBucketSortedSet<T, any>} */ (this._map.get(
key
));
const finishUpdate = oldEntry.startUpdate(item);
return remove => {
if (remove) {
this.size--;
finishUpdate(true);
if (oldEntry.size === 0) {
this._deleteKey(key);
}
return;
}
const newKey = this._getKey(item);
if (key === newKey) {
finishUpdate();
} else {
finishUpdate(true);
if (oldEntry.size === 0) {
this._deleteKey(key);
}
this._addInternal(newKey, item);
}
};
}
}
/**
* @param {Iterator<T>[]} iterators list of iterators to append to
* @returns {void}
*/
_appendIterators(iterators) {
if (this._unsortedItems.size > 0)
iterators.push(this._unsortedItems[Symbol.iterator]());
for (const key of this._keys) {
const entry = this._map.get(key);
if (this._leaf) {
const leafEntry = /** @type {SortableSet<T>} */ (entry);
const iterator = leafEntry[Symbol.iterator]();
iterators.push(iterator);
} else {
const nodeEntry = /** @type {LazyBucketSortedSet<T, any>} */ (entry);
nodeEntry._appendIterators(iterators);
}
}
}
/**
* @returns {Iterator<T>} the iterator
*/
[Symbol.iterator]() {
const iterators = [];
this._appendIterators(iterators);
iterators.reverse();
let currentIterator = iterators.pop();
return {
next: () => {
const res = currentIterator.next();
if (res.done) {
if (iterators.length === 0) return res;
currentIterator = iterators.pop();
return currentIterator.next();
}
return res;
}
};
}
}
module.exports = LazyBucketSortedSet;