251284 - SpiderMonkey: RegExp.prototype should be of class "Object"!?

Status: RESOLVED WONTFIX

(Core :: JavaScript Engine, defect, P3)

Description

[Brendan Eich [:brendan]]

ECMA-262 Edition 3 specifies that RegExp.prototype is an object of class
"Object", not of the class for whose constructor it is the prototype, namely
"RegExp".  This seems like a bug in the spec to me.

This is a big honking incompatibility, but ECMA-262 mandates it for
- arguments objects, but they have no class constructor (10.1.8)
- RegExp.prototype (15.10.6)

but not for:
- Function.prototype (15.3.4)
- Array.prototype (15.4.4)
- String.prototype (15.5.4)
- Boolean.prototype (15.6.4)
- Number.prototype (15.7.4)
- Date.prototype (15.9.5)
- Error.prototype (15.11.4, see also 15.11.7.7 for Error subtypes)

I'm likely to WONTFIX this and add to the ECMA-262 errata page at
http://www.mozilla.org/js/language/E262-3-errata.html.

What does Rhino do?

/be

Comment 1

[Igor Bukanov]

In Rhino [[Class]] for RegExp.prototype is "RegExp" and it would require rather
untrivial efforts to make that [[Object]].

Comment 2

[Waldemar Horwat]

This behavior is deliberate in ECMA-262 Edition 3.  RegExp.prototype doesn't
have enough properties or behaviors defined to have class "RegExp".  For
example, this way the spec gets away with not specifying [[Match]] for
RegExp.prototype.

Comment 3

[Brendan Eich [:brendan]]

This requires special case code in any implementation that provides common
mechanism for {Function,Array,String,Boolean,Number,Date,*Error}.prototype. 
That seems worse to me, because of real-world costs, than any spec uncleanliness
that would accrue from a RegExp.prototype whose [[Class]] was "RegExp".

What does JScript do?  How about KJS (Safari)?

/be

Comment 4

[Brendan Eich [:brendan]]

IE6: alert(RegExp.prototype) => "//".

KJS in Safari: alert(RegExp.prototype) => "", and further testing shows that
RegExp.prototype.test is an "(internal function)", to with the same as /hi/.test.

I haven't tested Opera; cc'ing Ian.

If no major browser follows the spec here, and even now with 3 major browsers
disagreeing, I'm going to lobby ECMA TG1 to change the spec.  Inconvenience for
the spec writers can't justify the code bloat and ugly special case for all the
implementations, in my view.

/be

Comment 5

[Brendan Eich [:brendan]]

Ugh, s/to with/to wit/ in my last comment.

/be

Comment 6

[Hixie (not reading bugmail)]

Opera seems to do the same as IE (returns "//" for RegExp.prototype). Is there 
anything else I should test? "RegExp" returns a native code function with 9 
members ($7, $8, $9, $1, $2, ... $6, respectively). RegExp.multiline, RegExp.
rightContext, etc, return undefined (unlike in Mozilla). RegExp.prototype.test 
returns a native code function, RegExp.prototype.multiline returns undefined.

Comment 7

[Brendan Eich [:brendan]]

Hixie: thanks, the main point is that RegExp.prototype is an object whose
[[Class]] is "RegExp" in Opera, just as in IE, Safari, and Mozilla's two JS
engines (btw, the better toString() result for an empty regexp is "/(?:)/",
because it can be round-tripped through eval).

So no major browser's JS implementation matches the ECMA spec here, and making
all four do so adds code bloat and bug habitat, breaking backward compatibility
slightly just for a spec simplicity issue.  I'd like to add an erratum for
15.10.6 in http://www.mozilla.org/js/language/E262-3-errata.html, and lobby ECMA
TG1.  Any objections?

/be

Comment 8

[Brendan Eich [:brendan]]

I'll bring this up at the May 19 ECMA TG1 meeting.

/be

Comment 9

[Brendan Eich [:brendan]]

Comment 8 was a while ago. The reference implementation from http://es-lang.org/ agrees with browser implementations. WONTFIX.

/be