This PR proposes changes to existing test262 tests to allow them to pass under Hardened JavaScript (see Secure ECMAScript proposal and Hardened JavaScript). Moddable uses Hardened JavaScript for JavaScript runtimes on resource constrained embedded devices, including those targeted by ECMA-419.
The changes fall into four groups:
1. Replace use of new Date() with new Date(1970). Scripts running inside a Compartment cannot retrieve the current time, so new Date() throws but new Date(1970) succeeds. Very few tests need the current time, but instead simply need a Date instance.
2. Use Object.defineProperty instead of setting existing built-in properties directly, such as toString and toValue. In Hardened JavaScript, prototypes of built-in objects are frozen. Consequently, setting properties of an instance that exist on the prototype throw (Hardened JavaScript is always in strict mode).
3. Eliminate use of Math.random(). Scripts running inside a Compartment cannot generate random numbers. One test identified so far uses Math.random() in a way that can easily be replaced with a counter.
4. Narrow the scope of exception tests. Consider the following
assert.throws(TypeError, () => {
var s1 = new Date();
s1.toString = Boolean.prototype.toString;
s1.toString();
});
This test passes, but only because new Date() fails by throwing a TypeError. If the invocation of the Date constructor is resolved by (1) above, then the assignment to toString fails as per (2) above. The script should be modified as below to ensure that assert.throws only tests the intended statement, s1.toString(). The modified script tests the intended functionality and passes under Hardened JavaScript
var s1 = new Date(1970);
Object.defineProperty(s1, "toString", {
value: Boolean.prototype.toString
});
assert.throws(TypeError, () => {
s1.toString();
});
This is an initial PR to begin the process of adapting test262 for use with Hardened JavaScript. Further changes are expected, with the vast majority likely to fall into the four groups described above.
Thank you to gibson042, kriskowal, and erights for their advice on this work.
For the statement level test, the inner class name is not initialized
at the time the decorators evaluate, resulting in a ReferenceError
that the declaration can not be accessed prior to initialization.
Similar, non-decorator code, like:
class C {
static dec() {}
static {
this.x = C.dec();
class C {}
}
}
also results in a ReferenceError.
For the expression level test, the var C is undefined at the time the
decorators are evaluated, resulting in TypeError while trying to access
a member of undefined.
Similar, non-decorator code, like:
var C = class {
static f() {};
static {
this.x = C.f();
}
}
also results in a TypeError.
While we're at it, use assert() instead of assert.sameValue() for brevity,
if we are not specifically testing that the return value of hasOwnProperty
is the value true or false; and add more informative assertion messages to
help with debugging.
In some cases, the Object.hasOwnProperty.call could be replaced with
verifyProperty(), if the property descriptor was also being verified at
the same time.
This fixes some tests that were faulty to begin with: a common mistake was
Object.hasOwnProperty(obj, prop) which is probably going to return false
when that's not what you want.
The only instances left of `Object.hasOwnProperty` are one regression test
in implementation-contributed which I can't tell if it was intentionally
needed to trigger the regression, and a few instances of
`Object.hasOwnProperty('prototype')` which would defeat the purpose to
convert into `Object.prototype.hasOwnProperty.call(Object, 'prototype')`
form.
Closes: #3524
Some tests which include function declarations designed to verify
behavior do not reference those functions. Insert the references
necessary for those functions to serve their intended purpose.
This reverts commit b690cb67be9b487eb10156c03e2c00869e88cc9d, reversing
changes made to 50dd431dffe5cf86e9064a652d6b01dbbe542cf0. This is
necessary because the reverted changeset reduced coverage by an unknown
extent.
Since https://github.com/tc39/ecma262/pull/1490, the "length" and "name"
properties of a class are defined before any static methods. This is
tested by #2057, in test/language/computed-property-names of all places.
At the same time, static methods with "name" as the name would overwrite
the original property, but retain the original property enumeration
order. This was not previously tested. In fact, the overwriting behavior
was not tested at all for the "length" property.
This commit mends both holes in test coverage.
* Add tests for "Class Static Init. Blocks" proposal
This proposal is currently at "stage 3" in TC39's standardization
process.
* fixup! Add tests for "Class Static Init. Blocks" proposal
* Correct identifier reference
* Update tests for grammar
* Update tests for identifiers
* Add tests for scope derivation
Several tests for getters and setters claim to check for an early
SyntaxError regarding mixing static and non-static propeties with the
same name. However, the tests trigger another issue: the getters have no
method body; they're missing curlies.
Fix the tests to test only the intended SyntaxError, not unrelated
SyntaxError-s.
