Know yourengines velocity2011

Know Your Engines
How to Make Your JavaScript Fast

Dave Mandelin
June 15, 2011
O’Reilly Velocity

5 years of progress...
10

JavaScript
7.5 C
run time vs. C

5

2.5

0
2006 2008 2011

one program on one popular browser:
10x faster!

...lost in an instant!
function f() {
var sum = 0;
for (var i = 0; i < N; ++i) {
sum += i;
}
}

function f() {
eval(“”);
var sum = 0;
for (var i = 0; i < N; ++i) {
sum += i;
}
}

...lost in an instant!
function f() { 80
var sum = 0;
for (var i = 0; i < N; ++i) {
sum += i; 60
}
} 40

20

function f() {
eval(“”); 0
without eval with eval
var sum = 0;
for (var i = 0; i < N; ++i) {
sum += i; with eval(“”) up to
}
} 10x slower!

Making JavaScript Fast
Or, Not Making JavaScript Slow

How JITs make JavaScript not slow

How not to ruin animation with pauses

How to write JavaScript that’s not slow

Inside the 2006 JS Engine
DOM

Standard
Front End Interpreter Library

Garbage
Collector

// JavaScript source
e.innerHTML = n + “ items”; DOM

Standard

Garbage
Collector


Standard

// bytecode (AST in some engines) Garbage
tmp_0 = add var_1 str_3 Collector
setprop var_0 ‘innerHTML’ tmp_0


Standard
Run the bytecode



Standard
Run the bytecode
Reclaim memory


Set innerHTML


Standard
Run the bytecode
Reclaim memory


Why it’s hard to make JS fast

Because
JavaScript is an untyped language.

untyped = no type declarations

Operations in an untyped language
x = y + z can mean many things

• if y and z are numbers, numeric addition

• if y and z are strings, concatenation

• and many other cases; y and z can have different types

Engine-Internal Types
JS engines use ﬁner-grained types internally.

JavaScript type

number

object


JavaScript type Engine type

number 32-bit* integer
64-bit ﬂoating-point

object




{ a: 1 }
{ a: 1, b: 2 }
object
{ a: get ... }
{ a: 1, __proto__ = new C }




{ a: 1 }
{ a: 1, b: 2 } Different
object
{ a: get ... } shapes
{ a: 1, __proto__ = new C }

Values in an untyped language
Because JavaScript is untyped, the interpreter needs boxed values.

Boxed Unboxed
Purpose Storage Computation
Examples (INT, 55) 55

(STRING, “foo”) “foo”

Deﬁnition (type tag, C++ value) C++ value

only boxed values can be stored in variables,
only unboxed values can be computed with (+, *, etc)

Running Code in the Interpreter
Here’s what the interpreter must do to execute x = y + z:


‣ read the operation x = y + z from memory


‣ read the boxed inputs y and z from memory


‣ check the types of y and z and choose the action


‣ unbox y and z


‣ unbox y and z
‣ execute the action


‣ unbox y and z
‣ box the output x


‣ unbox y and z
‣ write the boxed output x to memory


‣ unbox y and z
This is the only real work!
‣ write the boxed output x to memory


‣ unbox y and z
This is the only real work!
‣ box the output x Everything else is
‣ write the boxed output x to memory overhead.

Garbage
Collector DOM
Interpreter
JavaScript source Standard
Library

Front End

bytecode/AST

Garbage
Collector DOM
Interpreter
Library

Front End JIT Compiler

Compile to x86/x64/ARM

bytecode/AST

Garbage
Collector DOM
Interpreter
Library

Fast! x86/x64/ARM

CPU

bytecode/AST

Garbage
Collector DOM
Interpreter
Library

Fast! x86/x64/ARM

CPU
Type-Specializing
bytecode/AST JIT Compiler Ultra Fast!

THE Garbage
DOM
Collector
SLOW ZONE Interpreter
Library

Fast! x86/x64/ARM

CPU
Type-Specializing
bytecode/AST JIT Compiler Ultra Fast!

Running Code with the JIT
All Major
The basic JIT compiler on x = y + z: Browsers

‣ read the inputs y and z from memory
‣ unbox y and z
‣ write the output x to memory

All Major

‣ read the operation x = y + z from memory CPU does it for us!
‣ unbox y and z

All Major

‣ read the operation x = y + z from memory CPU does it for us!
‣ unbox y and z
‣ write the output x to memory JIT code can keep
things in registers

Choosing the action in the JIT

• Many cases for operators like +


• Engines generate fast JIT code for “common cases”
• number + number

• string + string


• Engines generate fast JIT code for “common cases”
• number + number

• string + string

• “Rare cases” run in the slow zone
• number + undeﬁned

JITs for Regular Expressions
All Major
Browsers
• There is a separate JIT for regular expressions

• Regular expressions are generally faster than manual search

• Still in the slow zone:

• Some complex regexes (example: backreferences)

• Building result arrays (test much faster than exec)

Object Properties
function f(obj) {
return obj.a + 1;
}

Object Properties
function f(obj) {
return obj.a + 1;
}

• Need to search obj for a property named a slow

Object Properties
function f(obj) {
return obj.a + 1;
}


• May need to search prototype chain up several levels super-slow

Object Properties
function f(obj) {
return obj.a + 1;
}


• May need to search prototype chain up several levels super-slow

• Finally, once we’ve found it, get the property value fast!

ICs: a mini-JIT for objects All Major
Browsers

Browsers

• Properties become fast with inline caching (we prefer IC)

Browsers


• Basic plan:

Browsers


• Basic plan:
1. First time around, search for the property in the Slow Zone

Browsers


• Basic plan:
2. But record the steps done to actually get the property

Browsers


• Basic plan:
2. But record the steps done to actually get the property
3. Then JIT a little piece of code that does just that

ICs: Example
Example Code
var obj1 = { a: 1, b: 2, c: 3 };
var obj2 = { b: 2 };

function f(obj) {
return obj.b + 1;
}

ICs: Example
Example Code
var obj1 = { a: 1, b: 2, c: 3 };
var obj2 = { b: 2 };

function f(obj) {
return obj.b + 1;
}

Generated JIT Code
...
jump slowPropAccess slowPropAccess:
continue_1: ... set up call
... call ICGetProp ; C++ Slow Zone
jump continue_1

ICs: Example
Example Code
shape=12, in position 1
var obj1 = { a: 1, b: 2, c: 3 };
var obj2 = { b: 2 };

function f(obj) {
return obj.b + 1;
}

Generated JIT Code
...
jump continue_1

ICs: Example
Example Code icStub_1:
shape=12, in position 1 compare obj.shape, 12
var obj1 = { a: 1, b: 2, c: 3 }; jumpIfFalse slowPropAccess
var obj2 = { b: 2 }; load obj.props[1]
jump continue_1
function f(obj) {
return obj.b + 1;
}

Generated JIT Code
...
jump continue_1

ICs: Example
shape=15, in position 0 jump continue_1
function f(obj) {
return obj.b + 1;
}

Generated JIT Code
...
jump continue_1

ICs: Example
shape=15, in position 0 jump continue_1
function f(obj) {
return obj.b + 1;
} icStub_2:
compare obj.shape, 15
jumpIfFalse slowPropAccess
Generated JIT Code load obj.props[0]
jump continue_1

...
jump continue_1

These are fast because of ICs
Global Variable Access
var q = 4;
var r;

function f(obj) {
r = q;
}

Global Variable Access
var q = 4;
var r;

function f(obj) {
r = q;
}

Direct Property Access
var obj1 = { a: 1, b: 2, c: 3 };
var obj2 = { b: 2 };

function f(obj) {
obj2.b = obj1.c;
}

Global Variable Access Closure Variable Access
var q = 4; var f = function() {
var r; var x = 1;
var g = function() {
function f(obj) { var sum = 0;
r = q; for (var i = 0; i < N; ++i) {
} sum += x;
}
return sum;
Direct Property Access }
return g();
var obj1 = { a: 1, b: 2, c: 3 }; }
var obj2 = { b: 2 };

function f(obj) {
obj2.b = obj1.c;
}

Prototypes don’t hurt much
function A(x) {
this.x = x;
}

function B(y) {
this.y = y;
}
B.prototype = new A;

function C(z) {
this.z = z;
}
C.prototype = new B;

new A
function A(x) {
this.x = x;
}
new B
function B(y) { proto
this.y = y;
}
new C(1)

function C(z) {
this.z = z;
}

new A
function A(x) {
this.x = x;
}
new B
this.y = y;
}
new C(1) new C(2)

function C(z) {
this.z = z;
}

new A
function A(x) {
this.x = x;
}
new B
this.y = y;
}
new C(1) new C(2) new C(3)

function C(z) {
this.z = z;
}

new A
function A(x) {
this.x = x;
}
new B
this.y = y;
}

function C(z) {
this.z = z;
Shape of new C objects determines prototype
}

new A
function A(x) {
this.x = x;
}
new B
this.y = y;
}

function C(z) {
this.z = z;
Shape of new C objects determines prototype
}
C.prototype = new B; -> IC can generate code that checks shape,
then reads directly from prototype without walking

Many Shapes Slow Down ICs
What happens if many shapes of obj are passed to f?
function f(obj) {
return obj.p;
}

ICs end up looking like this:

function f(obj) {
return obj.p;
}

jumpIf shape != 12
read for shape 12

function f(obj) {
return obj.p;
}

jumpIf shape != 12
read for shape 12

jumpIf shape != 15
read for shape 15

function f(obj) {
return obj.p;
}

jumpIf shape != 12
read for shape 12

jumpIf shape != 15
read for shape 15

jumpIf shape != 6
read for shape 6

function f(obj) {
return obj.p;
}

ICs end up looking like this: ...
jumpIf shape != 12 jumpIf shape != 16
read for shape 12 read for shape 16



Many shapes in practice
100
IE IE Slow Zone for 2+ shapes
Opera
Chrome
75
Opera # of shapes doesn’t matter!
nanoseconds/iteration

Firefox
Safari

50 Chrome more shapes -> slower

Firefox
25 slower with more shapes,
but levels off in Slow Zone
Safari
0
1 2 8 16 32 100 200
# of shapes at property read site

Deeply Nested Closures are Slower
var f = function() {
var x;
var h = function() {
var y;
var i = function () {
var j = function() {
z = x + y;

var f = function() { f call object
var x;
var h = function() { h call object
var y;
var j = function() { j call object
z = x + y;
First call to f

var f = function() { f call object f call object
var x;
var h = function() { h call object h call object
var y;
var j = function() { j call object j call object
z = x + y;
First call to f Second call to f

var x;
var y;
z = x + y;

• Prototype chains don’t slow us down, but deep closure nesting does. Why?

var x;
var y;
z = x + y;


• Every call to f generates a unique closure object to hold x.

var x;
var y;
z = x + y;


• Every call to f generates a unique closure object to hold x.

• The engine must walk up to x each time

Properties in the Slow Zone
Undeﬁned Property
(Fast on Firefox, Chrome)

var a = {};
a.x;

Undeﬁned Property
(Fast on Firefox, Chrome)

var a = {};
a.x;

DOM Access
(I only tested .id, so take with a grain of salt--
other properties may differ)

var a = document.getByElementId(“foo”);
a.id;

Undeﬁned Property Scripted Getter
(Fast on Firefox, Chrome) (Fast on IE)

var a = {}; var a = { x: get() { return 1; } };
a.x; a.x;

DOM Access
other properties may differ)

a.id;

Undeﬁned Property Scripted Getter
(Fast on Firefox, Chrome) (Fast on IE)

var a = {}; var a = { x: get() { return 1; } };
a.x; a.x;

DOM Access Scripted Setter
other properties may differ) var a = { x: set(y) { this.x_ = y; } };
a.x = 1;
a.id;

The Type-Specializing JIT
Firefox 3.5+
(Tracemonkey)

Chrome 11+
(Crankshaft)

Types FTW!
If only JavaScript had type declarations...

Types FTW!

➡ The JIT would know the type of every local variable

Types FTW!

➡ Know exactly what action to use (no type checks)

Types FTW!

➡ Local variables don’t need to be boxed (or unboxed)

Types FTW!

➡ Local variables don’t need to be boxed (or unboxed)
We call this kind of JIT a
type-specializing JIT

But JS doesn’t have types

• Problem: JS doesn’t have type declarations
• won’t have them any time soon
• we don’t want to wait


• Solution: run the program for a bit, monitor types


• Solution: run the program for a bit, monitor types
• Then recompile optimized for those types

Running with the Type-Specializing JIT
Firefox 3.5+
On x = y + z: Chrome 11+
‣ unbox y and z

Further Optimization 1
Automatic Inlining

original code
function getPop(city) {
return popdata[city.id];
}

for (var i = 0; i < N; ++i) {
total += getPop(city);
}

Automatic Inlining

original code JIT compiles as if
function getPop(city) {
you wrote this
return popdata[city.id];
} for (var i = 0; i < N; ++i) {
total += popdata[city.id];
for (var i = 0; i < N; ++i) { }
total += getPop(city);
}

Loop Invariant Code Motion (LICM, “hoisting”)

original code

for (var i = 0; i < N; ++i) {
total += a[i] *
(1 + options.tax);
}

Loop Invariant Code Motion (LICM, “hoisting”)

original code JIT compiles as if
you wrote this
for (var i = 0; i < N; ++i) { var f = 1 + options.tax;
total += a[i] * for (var i = 0; i < N; ++i) {
(1 + options.tax); total += a[i] * f;
} }

Optimize Only Hot Code
• Type-specializing JITs can have a hefty startup cost
• Need to collect the type information

• Advanced compiler optimizations take longer to run

Optimize Only Hot Code
• Type-specializing JITs can have a hefty startup cost
• Need to collect the type information

• Advanced compiler optimizations take longer to run

• Therefore, type specialization is applied selectively
• Only on hot code

• Tracemonkey: hot = 70 iterations

• Crankshaft: hot = according to a proﬁler

• Only if judged to be worthwhile (incomprehensible heuristics)

Current Limitations
• What happens if the types change after compiling?
• Just a few changes -> recompile, slight slowdown
• Many changes -> give up and deoptimize to basic JIT

Current Limitations
• Array elements, object properties, and closed-over variables
• Usually still boxed
• Still need to check type and unbox on get, box on set
• Typed arrays might help, but support is not always there yet

Current Limitations
• Array elements, object properties, and closed-over variables
• Usually still boxed
• Still need to check type and unbox on get, box on set
• Typed arrays might help, but support is not always there yet
• JS semantics require overﬂow checks for integer math

Type Inference for JITs
Current
Research
@Mozilla

Type Inference

• Trying to get rid of the last few instances of boxing
(from before: array and object properties)

Type Inference

• Trying to get rid of the last few instances of boxing
(from before: array and object properties)

• Idea: use static program analysis to prove types

• of object props, array elements, called functions

• or, almost prove types, and also prove minimal checks needed

Type Inference Example
var a = [];
for (var i = 0; i < N; ++i) {
a[i] = i * i;
]

var sum = 0;
for (var i = 0; i < N; ++i) {
sum += a[i];
}

Type inference gets this...

var a = [];
for (var i = 0; i < N; ++i) {
a[i] = i * i;
]

var sum = 0;
for (var i = 0; i < N; ++i) {
sum += a[i];
}

Type inference gets this...
“i is always a number,
so i * i is always a number,
so a[_] is always a number!”

var a = []; var a = [];
for (var i = 0; i < N; ++i) { for (var i = 0; i < N; ++i) {
a[i] = i * i; if (i % 2)
] a[i] = i * i;
else
var sum = 0; a[i] = “foo”;
for (var i = 0; i < N; ++i) { ]
sum += a[i];
} var sum = 0;
for (var i = 0; i < N; ++i) {
if (i % 2)
Type inference gets this... sum += a[i];
}
“i is always a number,
so i * i is always a number, ...but not this.
so a[_] is always a number!”

Type-stable JavaScript
The key to running faster in future JITs is

type-stable JavaScript.

This means JavaScript where you could
declare a single engine-internal type for each variable.

Type-stable JS: examples
Type-stable
var g = 34;

var o1 = { a: 56 };
var o2 = { a: 99 };

for (var i = 0; i < 10; ++i) {
var o = i % 2 ? o1 : o2;
g += o.a;
}

g = 0;

Type-stable NOT type-stable
var g = 34; var g = 34;

var o1 = { a: 56 }; var o1 = { a: 56 };
var o2 = { a: 99 }; var o2 = { z: 22, a: 56 };

for (var i = 0; i < 10; ++i) { for (var i = 0; i < 10; ++i) {
var o = i % 2 ? o1 : o2; var o = i % 2 ? o1 : o2;
g += o.a; g += o.a;
} }

g = 0; g = “hello”;

What Allocates Memory?
Objects
new Object();
new MyConstructor();
{ a: 4, b: 5 }
Object.create();

Arrays
new Array();
[ 1, 2, 3, 4 ];

Strings
new String(“hello”);
“<p>” + e.innerHTML + “</p>”

Objects Function Objects
new Object(); var x = function () { ... }
new MyConstructor(); new Function(code);
{ a: 4, b: 5 }
Object.create();

Arrays
new Array();
[ 1, 2, 3, 4 ];

Strings

{ a: 4, b: 5 }
Object.create();

Arrays Closure Environments
new Array(); function outer(name) {
[ 1, 2, 3, 4 ]; var x = name;
return function inner() {
return “Hi, “ + name;
Strings }
}

{ a: 4, b: 5 }
Object.create();

Arrays Closure Environments
new Array(); function outer(name) {
[ 1, 2, 3, 4 ]; var x = name;
return function inner() {
return “Hi, “ + name;
Strings }
}
new String(“hello”); name is stored in an
“<p>” + e.innerHTML + “</p>” implicitly created object!

GC Pauses Your Program!
Time JavaScript GC Running
Running JS Paused

Running JS Paused

• Basic GC algorithm (mark and sweep)
• Traverse all reachable objects (from locals, window, DOM)

• Recycle objects that are not reachable

Running JS Paused



• The JS program is paused during GC for safe traversal

Running JS Paused



• The JS program is paused during GC for safe traversal
• Pauses may be long: 100 ms or more
• Serious problem for animation

• Can also be a drag on general performance

Reducing Pauses with Science 1
Generational GC Chrome

Idea: Optimize for creating many short-lived objects


Create objects in a frequently collected nursery area


Promote long-lived objects to a rarely collected tenured area



JavaScript GC Running
Simple GC
Running JS Paused



Simple GC
Running JS Paused

JavaScript
Generational GC
Running



Simple GC
Running JS Paused

JavaScript
Generational GC
Running

nursery collection (<100 us)



Simple GC
Running JS Paused

JavaScript
Generational GC
Running

tenured collection



Simple GC
Running JS Paused

JavaScript
Generational GC fewer pauses!
Running

tenured collection

Generational GC by Example
scavenging young generation (aka nursery)

mark-and-sweep tenured generation

Message
Message
Array


Point


Message
Message
Array


Point Point


Message
Message
Array


Point Point Line


Message
Message
Array


Point Point Line
a b


Message
Message
Array


Point Point Line Point
a b


Message
Message
Array


Point Point Line Point Message
a b


Message
Message
Array


Point Point Line Point Message Point
a b


Message
Message
Array


Point Point Line Point Point
a b


Message
Message
Message
Array



Message
Message
Message
Array

Reducing Pauses with Science 1I
Current
Incremental GC Research
@Mozilla

Current
@Mozilla
Idea: Do a little bit of GC traversal at a time

Current
@Mozilla

Simple GC
Running JS Paused

Current
@Mozilla

Simple GC
Running JS Paused

Incremental GC

Current
@Mozilla

Simple GC
Running JS Paused

Incremental GC shorter pauses!

Reducing Pauses in Practice
• For all GCs
• Fewer live objects -> shorter pauses (if not incremental),
less time spent in GC

• For all GCs

• For simple GCs
• Lower allocation rate (objects/second) -> less frequent pauses

• For all GCs

• For simple GCs
• Lower allocation rate (objects/second) -> less frequent pauses

• For generational GCs
• Short-lived objects don’t affect pause frequency

• Long-lived objects cost extra (promotion = copying)

JavaScript Engines in Practice

Performance Faults
• Performance fault: when a tiny change hurts performance

• Sometimes, just makes one statement slower

• Other times, deoptimizes the entire function!

• Reasons we have performance faults

• bug, tends to get quickly

• “rare” case, will get ﬁxed if not rare

• hard to optimize, RSN...

Strings
• In the Slow Zone, but some things are faster than you might think

Strings

• .substring() is fast, O(1)

• Don’t need to copy characters, just point within original

Strings

• .substring() is fast, O(1)

• Don’t need to copy characters, just point within original

• Concatenation is also optimized

• Batch up inputs in a rope or concat tree, concat all at once

• Performance fault: prepending (Chrome, Opera)

Strings

• .substring() is fast, O(1) // Prepending example
var s = “”;
•Don’t need to copy characters, just point iwithin<original {
for (var = 0; i 100; ++i)
s = i + s;
• Concatenation is also optimized }

• Batch up inputs in a rope or concat tree, concat all at once

• Performance fault: prepending (Chrome, Opera)

Arrays
fast: dense array
var a = [];
Want a fast array?
for (var i = 0; i < 100; ++i) {
a[i] = 0; ‣ Make sure it’s dense
}
‣ 0..N fill or push fill is always dense
3-15x slower: sparse array ‣ Huge gaps are always sparse
var a = []; ‣ N..0 fill is sparse on Firefox
a[10000] = 0;
for (var i = 0; i < 100; ++i) {
a[i] = 0;
‣ adding a named property is sparse
on Firefox, IE
}
a.x = 7; // Fx, IE only

Iteration over Arrays

fastest: index iteration
// This runs in all in JIT code,
// so it’s really fast.
for (var i = 0; i < a.length; ++i) {
sum += a[i];
}

Iteration over Arrays
3-15x slower: functional style
// This makes N function calls,
fastest: index iteration // and most JITs don’t optimize
// through C++ reduce().
sum = a.reduce(function(a, b) {
// This runs in all in JIT code,
return a + b; });
// so it’s really fast.
for (var i = 0; i < a.length; ++i) {
sum += a[i];
}
20-80x slower: for-in
// This calls a C++ function to
// navigate the property list.
for (var i in a) {
sum += a[i];
}

Functions
• Function calls use ICs, so they are fast

• Manual inlining can still help sometimes

• Key performance faults:

• f.call() - 1.3-35x slower than f()
• f.apply() - 5-50x slower than f()
• arguments - often very slow, but varies

Creating Objects
Creating objects is slow
Doesn’t matter too much how you create or populate

Creating Objects
Creating objects is slow
Doesn’t matter too much how you create or populate

Exception: Constructors on Chrome are fast
function Cons(x, y, z) {
this.x = x;
this.y = y;
this.z = z;
}

for (var i = 0; i < N; ++i)
new Cons(i, i + 1, i * 2);

OOP Styling
Prototype
function Point(x, y) {
this.x = x;
this.y = y;
}
Point.prototype = {
distance: function(pt2) ...

OOP Styling
Prototype Information-Hiding
this.x = x; function Point(x, y) {
this.y = y; return {
} distance: function(pt2) ...
Point.prototype = { }
distance: function(pt2) ... }

OOP Styling

Instance Methods
this.x = x;
this.y = y;
this.distance = function(pt2) ...
}

OOP Styling

Prototype style is much faster to create Instance Methods
(each closure creates a function object) function Point(x, y) {
this.x = x;
this.y = y;
}

OOP Styling

Prototype style is much faster to create Instance Methods
(each closure creates a function object) function Point(x, y) {
this.x = x;
this.y = y;
}
Using the objects is about the same

Exceptions
• Exceptions assumed to be rare in perf-sensitive code

• running a try statement is free on most browers

• throw/catch is really slow

• There are many performance faults around exceptions

• just having a try statement deoptimizes on some browers

• try-ﬁnally is perf fault on some

eval and with
Short version:
Do not use anywhere near performance sensitive code!

Mind-Bogglingly Awful Still Terrible
5-100x slower than using a function call 2-10x slower than without eval
var sum = 0; var sum = 0;
for (var i = 0; i < N; ++i) { eval(“”);
sum = eval(“sum + i”); for (var i = 0; i < N; ++i) {
} sum = eval(“sum + i”);
}

Top 5 Things to Know
5. Avoid eval, with, exceptions near perf-senstive code

4. Avoid creating objects in hot loops

3. Use dense arrays (know what causes sparseness)

2. Write type-stable code

2. Write type-stable code
1. ...

Talk To Us

JS engine developers want to help you. Tell us about:

• Performance faults you run into

• Exciting apps that require fast JS

• Anything interesting you discover about JS performance

Know yourengines velocity2011

More Related Content

Know yourengines velocity2011

Editor's Notes