646938 - JM: x === x fails when x is -NaN

Status: RESOLVED FIXED

(Core :: JavaScript Engine, defect)

Description

[Jan de Mooij [:jandem]]

--
function f() {
    var x = -[NaN][0];
    print(x === x);
}
f();
--
With -m -a this prints |true| instead of |false|. 

I think jsop_neg flips the first bit and this fails the check in jsop_stricteq.

Comment 1

[Jan de Mooij [:jandem]]

Attachment: Patch

This patch fixes the NaN-check in stricteq to ignore the sign bit. The alternative is to add a NaN-check to jsop_neg but that seems more expensive than this patch.

Comment 2

[Brendan Eich [:brendan]]

IIRC neg is very rare, eq is quite common. Not sure about stricteq vs. neg but I would not be surprised if stricteq was more frequent, statically and dynamically. dmandelin, any data?

/be

Comment 3

[Jan de Mooij [:jandem]]

(In reply to comment #2)
> IIRC neg is very rare, eq is quite common. Not sure about stricteq vs. neg but
> I would not be surprised if stricteq was more frequent, statically and
> dynamically.

This is about the stricteq case where lhs and rhs have the same backing (same local or copy of same local). Thinking about this more, I should probably fix jsop_neg first and benchmark.

Comment 4

[Jan de Mooij [:jandem]]

Comment on attachment 523710 [details] [diff] [review]
Patch

OK this is interesting. GCC implements -(double) like JM does (xor with sign mask). This means that fixing jsop_neg is not enough as we can still get a -NaN from the interpreter.

Comment 5

[David Mandelin [:dmandelin]]

(In reply to comment #2)
> IIRC neg is very rare, eq is quite common. Not sure about stricteq vs. neg but
> I would not be surprised if stricteq was more frequent, statically and
> dynamically. dmandelin, any data?

In the V8 benchmarks, neg is rare, and eq/stricteq are common. In SunSpider (which we care less about than we used to), stricteq is rare; eq and neg are both fairly common, but eq runs 2x as many times as neg.

Comment 6

[David Anderson [:dvander] - inactive, e-mail if emergency]

Comment on attachment 523710 [details] [diff] [review]
Patch

bonus points if you change that r11 to scratchRegister ;)

Comment 7

[Jan de Mooij [:jandem]]

Attachment: Patch

scratchRegister was protected, this patch makes it public like stackPointerRegister.

Comment 8

[Jan de Mooij [:jandem]]

http://hg.mozilla.org/tracemonkey/rev/de7b0f3323c1

Comment 9

[Chris Leary [:cdleary] (not checking bugmail)]

http://hg.mozilla.org/mozilla-central/rev/de7b0f3323c1