642772 - TM: [infer failure] Missing type at #3:00006 pushed 0: String.split

Status: RESOLVED FIXED

Description

[Christian Holler (:decoder)]

Created attachment 520178 [details]
shell testcase, unpack, chdir and run main.js with options "-n -a"

The attached test case (run main.js with -n -a) crashes on TI tip, tested on 64
bit.

Comment 1

[Brian Hackett (:bhackett)]

This looks like a bug in TM related to lazy standard classes.

JS_EnumerateStandardClasses calls obj->nativeContains for each standard atom to figure out whether that class needs initialization.  If the class had previously been initialized and then deleted, a new version of it (and all its associated natives, etc.) will be created and attached to the global, confusing inference (which thinks that compileAndGo code will only ever have a single String/Number/Boolean/etc.)

With a TM shell:

js> String
function String() {[native code]}
js> delete String
true
js> String
typein:3: ReferenceError: String is not defined
js> for (x in this) {}
js> String
function String() {[native code]}

Also:

var one = String.prototype.split;
delete String;
var two = ("foo").split;
for (x in this) {}
var three = ("foo").split;
print(one === two);
print(one === three);

d8/jsc:
true
true

js:
true
false

Comment 5

[Paul Biggar]

Created attachment 527628 [details] [diff] [review]
Kinda ugly fix

I'm not 100% happy with this solution, but it works. I tried the way you suggested, which was to add a prototype and a constructor, but that was user visible and didn't match the spec (in that JSON.prototype was not undefined anymore).

I have more tests to add here, from the duplicated bugs.

Comment 6

[Paul Biggar]

Comment on attachment 527628 [details] [diff] [review]
Kinda ugly fix

Gonna see if I can't fix:

delete JSON;
for (x in this);
assertEq(JSON, undefined);

while I'm at it.

Comment 7

[Paul Biggar]

Created attachment 527776 [details] [diff] [review]
updated

This makes it more explicit what we're doing, and handles one case we didn't before, and adds many more tests. I didn't include the tests from the duplicates - I think I need to try them on the JM branch first.

Note this patch is for TM.

Comment 8

[Brian Hackett (:bhackett)]

Comment on attachment 527776 [details] [diff] [review]
updated


>+        /*
>+         * If the constructor is undefined, then it hasn't been initialized
>+         * (see also js::MarkInitializedNoProto.
>+         */

Missing a ')', also MarkInitializedNoProto doesn't seem to exist.

>+void
>+DefineNoConstructorAndPrototype(JSContext* cx, JSProtoKey key)
>+{
>+    /*
>+     * Lazy standard classes need a way to indicate if they have been
>+     * initialized. Otherwise, when we delete them, we might accidentally
>+     * recreate them via a lazy initialization. We use the presence of a ctor
>+     * or proto in the globalObject's slot to indicate that they've been
>+     * constructed, but this only works for classes which have a proto and
>+     * ctor. Classes which don't have one can call this function.
>+     *
>+     * We use True so that it's obvious what we're doing (instead of, say, Null).
>+     * If the class has no constructor (JSON and Math), then our test
>+     */
>+    if (cx->globalObject->getReservedSlot(key) == UndefinedValue())
>+        cx->globalObject->setSlot(key, BooleanValue(true));
>+}*

DefineNoConstructorAndPrototype should have a better name ("if it doesn't define a constructor and prototype, then what *does* it do?").  MarkStandardClassInitializedNoProto?  Also, the last sentence in the comment trails off.

>-fails-if(xulRuntime.shell) script regress-633741.js
>+script regress-633741.js

Nice.

I think Proxy follows the same pattern as Math and JSON and should mark the class as initialized.  No others I'm aware of, but you might want to double check.

r+ with these addressed.

Comment 9

[Paul Biggar]

Created attachment 528335 [details] [diff] [review]
Add proxy and reflect

Just sending this for another round of review since I've change how it works a bit. You were right about proxy, but Reflect was also like that, and I needed to refactor things a little bit to make that work in a non-ugly way.

Comment 10

[Brian Hackett (:bhackett)]

Comment on attachment 528335 [details] [diff] [review]
Add proxy and reflect

Review of attachment 528335 [details] [diff] [review]:

::: js/src/jsobj.h
@@ +1549,5 @@
                               JSPropertySpec *ps, JSFunctionSpec *fs,
                               JSPropertySpec *static_ps, JSFunctionSpec *static_fs);
+
+bool
+isStandardClassResolved(JSObject *obj, js::Class *clasp);

I think the convention here is that non-member functions should start with an uppercase.

Comment 11

[Paul Biggar]

http://hg.mozilla.org/tracemonkey/rev/881c06ebc5ee

Comment 13

[Christian Holler (:decoder)]

Brian merged this and other changes to TI at 

http://hg.mozilla.org/projects/jaegermonkey/rev/02c4a0f752f3

and I can no longer reproduce this on TI trunk, so I assume this is fixed now also on that branch :)

Comment 14

[Christian Holler (:decoder)]

A testcase for this bug was automatically identified at js/src/jit-test/tests/basic/bug642772-2.js.