899652 - eliminate static constructors for WebIDL constant tables on constexpr-supporting compilers

Status: RESOLVED FIXED

(Core :: DOM: Core & HTML, defect)

Comment 1

[Nathan Froyd [:froydnj]]

Attachment: use JS::UndefinedValue instead of JSVAL_VOID to encourage constexpr-ness

Waldo informed me over IRC yesterday that JS::UndefinedValue() should be used instead
of JSVAL_VOID.  Since the former is a function and can be constexpr-folded, let's use
that instead.

Comment 2

[Nathan Froyd [:froydnj]]

Attachment: modify bindings codegen to make webidl constant arrays constexpr

Some simple modifications to make generatePrefableArray attach the appropriate
constexpr qualifier when necessary.  Note that this depends on the code in bug
899309.

I'm double-checking right now with Try whether the arrays actually need
to be constexpr in order to fold them to static data with GCC.

Comment 3

[Boris Zbarsky [:bzbarsky]]

Comment on attachment 783221 [details] [diff] [review]
use JS::UndefinedValue instead of JSVAL_VOID to encourage constexpr-ness

Oy, JSVAL_VOID is no longer a constant expression?  :(  Why did it get changed to not be the same as UndefinedValue()??

In any case, this is definitely a good change.  r=me

That said, should we file a followup to mark UndefinedValue as MOZ_CONSTEXPR?

Comment 4

[Boris Zbarsky [:bzbarsky]]

> That said, should we file a followup to mark UndefinedValue as MOZ_CONSTEXPR?

Ah, I see, bug 899309 is that.

Comment 5

[Nathan Froyd [:froydnj]]

(In reply to Boris Zbarsky (:bz) from comment #3)
> Oy, JSVAL_VOID is no longer a constant expression?  :(  Why did it get
> changed to not be the same as UndefinedValue()??

JSVAL_VOID has never (?) been a constant expression:

http://mxr.mozilla.org/mozilla-central/source/js/src/jsapi.h#1175

Maybe you are thinking of JSID_VOID?

Comment 6

[Boris Zbarsky [:bzbarsky]]

> JSVAL_VOID has never (?) been a constant expression:

As of rev 5a3e49205389, the definition in opt builds was:

#define JSVAL_VOID   BUILD_JSVAL(JSVAL_TAG_UNDEFINED, 0)

(and at all points before that JSVAL_VOID was definitely a constant integer value) but this got changed to the current setup in bug 684526, looks like....  I just missed it at the time.

We should go through Gecko code and get rid of all uses of JSVAL_*.

Comment 7

[Boris Zbarsky [:bzbarsky]]

Comment on attachment 783224 [details] [diff] [review]
modify bindings codegen to make webidl constant arrays constexpr

>+        arraySpecType = ("JS_VALUE_CONSTEXPR_VAR " if useValueConstexpr else "const ") + specType

If we can just always have the JS_VALUE_CONSTEXPR_VAR here, that would be great; then we don't need the new useValueConstexpr argument.

r=me either way.

Comment 8

[Nathan Froyd [:froydnj]]

Comment on attachment 783224 [details] [diff] [review]
modify bindings codegen to make webidl constant arrays constexpr

It looks like bare 'const' is sufficient; I must have thought constexpr was needed based on some earlier bogus tests.  So that's a nice bonus.  I'll just be checking in the first part.

Comment 9

[Carsten Book [:Tomcat]]

https://hg.mozilla.org/mozilla-central/rev/b0e1bdb61af3