835417 - Mark DOM getters as pure when they are

Status: RESOLVED FIXED

Description

[Boris Zbarsky [:bzbarsky, bz on IRC]]

This is similar to bug 747289 but promises less.  A pure getter promises that calling it has no side-effects (e.g. it doesn't flush and will never run script), but does not promise that the value returned is the same no matter what.  So it can be reordered wrt slot sets, but not wrt DOM methods or setters.

A preliminary attempt at this for Element.id gives me http://dromaeo.com/?id=189130,189131 because it makes Dromaeo's .id test completely useless....  But I expect this to help in real-life code that touches things like .className too.

Now one issue is that _most_ DOM getters are pure, but to be safe-by-default we should probably explicitly mark them so.  That means an unfortunate lot of annotation.  :(

Comment 1

[Boris Zbarsky [:bzbarsky, bz on IRC]]

Attachment: part 1. Mark Node.namespaceURI as not throwing, since [Constant] things aren't allowed to throw (though we were not enforcing that).

Comment 2

[Boris Zbarsky [:bzbarsky, bz on IRC]]

Attachment: part 2. Implement WebIDL parser and codegen support for marking things pure.

Comment 3

[Boris Zbarsky [:bzbarsky, bz on IRC]]

Attachment: part 3. Add the notion of aliasing DOM stuff to MIR and flag MGetDOMProperty with the right alias set if it's pure.

Comment 4

[Boris Zbarsky [:bzbarsky, bz on IRC]]

Attachment: part 4. Flag a bunch of DOM getters as [Pure].

Comment 5

[Boris Zbarsky [:bzbarsky, bz on IRC]]

Oh, and I'm not doing methods yet because jandem tells me there are complications with trying to move around MCalls.  But I really want to sort those out, and then we can do this for a bunch of methods too.

Comment 6

[Jan de Mooij [:jandem]]

Comment on attachment 707171 [details] [diff] [review]
part 3.  Add the notion of aliasing DOM stuff to MIR and flag MGetDOMProperty with the right alias set if it's pure.

Review of attachment 707171 [details] [diff] [review]:
-----------------------------------------------------------------

Nice!

Comment 7

[Boris Zbarsky [:bzbarsky, bz on IRC]]

Attachment: part 1. Mark Node.namespaceURI as not throwing, since [Constant] things aren't allowed to throw (though we were not enforcing that).

Comment 8

[Peter Van der Beken [:peterv]]

Comment on attachment 707167 [details] [diff] [review]
part 2.  Implement WebIDL parser and codegen support for marking things pure.

Review of attachment 707167 [details] [diff] [review]:
-----------------------------------------------------------------

::: dom/bindings/Codegen.py
@@ +4682,5 @@
>                  sig = sigs[0]
> +                if (len(sig[1]) != 0 or
> +                    not infallibleForMember(self.member, sig[0], self.descriptor)):
> +                    # We have arguments or our return-value boxing can fail
> +                    methodInfal = False

I think you should just do the |methodInfal = "infallible" in self.descriptor.getExtendedAttributes(self.member)| bit here in an else block.

Comment 9

[Boris Zbarsky [:bzbarsky, bz on IRC]]

   https://hg.mozilla.org/integration/mozilla-inbound/rev/fce9b4a08399
   https://hg.mozilla.org/integration/mozilla-inbound/rev/f0bd3d538233
   https://hg.mozilla.org/integration/mozilla-inbound/rev/0a21a1eeda35
   https://hg.mozilla.org/integration/mozilla-inbound/rev/d8a95f5b67d7

Comment 10

[Boris Zbarsky [:bzbarsky, bz on IRC]]

Documented at https://developer.mozilla.org/en-US/docs/Mozilla/WebIDL_bindings#.5BPure.5D

Comment 11

[Ryan VanderMeulen [:RyanVM]]

https://hg.mozilla.org/mozilla-central/rev/fce9b4a08399
https://hg.mozilla.org/mozilla-central/rev/f0bd3d538233
https://hg.mozilla.org/mozilla-central/rev/0a21a1eeda35
https://hg.mozilla.org/mozilla-central/rev/d8a95f5b67d7