The Essence of JavaScript Arjun Guha, Claudiu Saftoiu, and Shriram Krishnamurthi.

The Essence of JavaScript

Arjun Guha, Claudiu Saftoiu, and Shriram Krishnamurthi

2

"JavaScript has much in common with Scheme […] Because of this deep similarity …"

( )

3

function bar(x) { return function() { var x = x; return x; };}

var f = bar(200);f() 200

function bar(x) { return function() { var x = x; return x; };}

var f = bar(200);f() undefined

4

var x = 0;var y = 900;

function baz(obj) { with (obj) { x = y; }}

baz({ y: 100 });x 100

var myObj = { x : 0 };baz(myObj);x 100myObj.x 900

Is JavaScript Even Lexically Scoped?

5

"JavaScript has much in common with Scheme […] Because of this deep similarity …"

No help to researchers studying Web security, building JavaScript analyses, etc.

6

Bad

Arjun Guha, Claudiu Saftoiu, and Shriram Krishnamurthi

weirdness

7

nytimes.com is a JavaScript mashup

<script lang="javascript" src="http://ad.doubleclick.net/..."><script lang="javascript" src="http://ad.linkstorms.com/...">

8

function safeLookup(obj, fd) { if (fd === "eval") { throw "cannot access eval"; } else { return obj[fd]; }}

window["ev" + "al"] window["eval"] vulnerability

safeLookup(window, "ev" + "al") safeLookup(window, "eval") * exception

Malicious 3rd party code

ADsafe / Caja / Facebook JavaScript Syntactic Checks +

Inserted Runtime Checks

Runtime Safety Check

“Sanitized” 3rd party code

BUGGY

9

How can we reason about JavaScript?

• The JavaScript standard (ECMA-262). 200 pages of prose and pseudocode.

• Maffeis, Mitchell, and Taly. An Operational Semantics for JavaScript. 70 pages of semantics.

We need a tractable semantics

10

11

12

13

14

The Essence of JavaScript:Functions, Prototype-Based Objects, State,

Control Operators, and Primitives

15

Thank You!

Questions?

16

What about the bad parts?

Thanks, Emery Berger

17

In practice most development effort goes into the “noise” that researchers abstract � �away […]. [M]inimalistic subsets give rise to a nice and simple formalization, whereas language implementers actually need help formalizing the rough edges of the language, not the beautiful and clean subset.

Erik Meijer.Confessions of a Used Programming Language Salesman.

OOPSLA 2007.

18

What about the bad parts?scope objects, with, switch, return, var, continue, for, do-while, for-in, implicit type conversions, function statements, named function expressions, function objects, "constructors", new-expressions, sparse "arrays", this keyword, toString(), valueOf(), variable-arity, Function.caller, Function.callee, the standard library, etc.

syntactic sugar

Thanks, Emery Berger

We implement desugaring (1,000 LOC)

19

Desugaring is Compositional*

desugar(e1 + e2) = C [ desugar(e1), desugar(e2) ]

desugar(obj[field]) = C [ desugar(obj), desugar(field) ]

etc.

program context, inserted by desugaring

*except for with statements

20

JavaScript program λJS programdesugar

Chrome,Firefox,Rhino

100LOCinterpreter

(Desugaring is Total) For all JavaScript programs e, is desugar(e) defined?

(Desugar Commutes with Eval) For all JavaScript programs e, does desugar(JS-eval(e)) = λJS-eval(desugar(e))?

theiranswer

ouranswer

21

Syntactic Form Occurrences (approx.)

with blocks 15

var statements 500

try blocks 20

if and switch statements 90

functions 200

typeof and instanceof 35

new expressions 50

Math library functions 15

5,400 lines of the Mozilla JavaScript test suite:

22

/* if F, G are inverse functions and x==y, this should return 1 */function match(x, y, F, G) { switch (x) { case F(G(y)): return 1; default: return 0; }}

test_case("A", match(17, f(fInverse(17)), f, fInverse)), 1);test_case("B", match(17, 2000, f, fInverse), 0);test_case("C", match(1, 1, Math.exp, Math.log), 1);test_case("D", match(1, 200, Math.exp, Math.log), 0);test_case("E", match(1, 1, Math.sin, Math.cos), 1);

23

$ ./test_firefox.sh tests/Statements/switch-001.jsBUGNUMBER: (none)STATUS: Testing the switch statement PASSED! Section A of test PASSED! Section B of test PASSED! Section C of test PASSED! Section D of test PASSED! Section E of test

$ ./test_lambdajs.sh tests/Statements/switch-001.jsBUGNUMBER: (none)STATUS: Testing the switch statement PASSED! Section A of test PASSED! Section B of test PASSED! Section C of test PASSED! Section D of test PASSED! Section E of test

Our semantics produces exactly the same result

24

Syntactic Form Occurrences (approx.)

with blocks 15

var statements 500

try blocks 20

if and switch statements 90

functions 200

typeof and instanceof 35

new expressions 50

Math library functions 15

5,400 lines of the Mozilla JavaScript test suite:

scalable strategy: add more tests

equivalent under diff

25

Recent JavaScript Research

• Staged Information Flow for JavaScript. PLDI’09.• GateKeeper. USENIX’09.• Static Analysis for Ajax Intrusion Detection.

WWW’09.• Type Analysis for JavaScript. SAS’09.• Object Views: Fine-Grained Sharing in

Browsers. WWW’10.• …

Proofs?desugar to λJS

do proofs for λJS

build tools for λJS

26

function safeLookup(obj, fd) { if (fd === "eval") { throw "cannot access eval"; } else { return obj[fd]; }}

function safeLookup(obj, fd) { if (fd === "eval") { throw "cannot access eval"; } else { return obj[fd.toString()]; }}

Implicit call in JavaScriptExplicit call in λJS

badObj ={toString:

function () {return "eval"}}

window[badObj] safeLookup(window, badObj) window[badObj.toString()] window[(function () return "eval")()] window["eval"]

27

Conclusion

• λJS is tractable and good for soundness proofs

• desugar is executable, so semantics-based tools can handle real source

• Used in Typed JavaScript, flow analyses, security type systems (JS source lang. too big, too implicit)

• λJS sets a new semantics standard: testing