931489 - use range analysis to eliminate NaN code

Status: RESOLVED FIXED

(Core :: JavaScript Engine: JIT, enhancement)

Description

[Dan Gohman [:sunfish]]

Attachment: range-ordered-compare.patch

We can use range analysis to eliminate several NaN and negative-zero checks. The following are a series of patches to implement this for a few things. The main one that's interesting is the one for MCompare. The other two, for MNot and MPowHalf, are less interesting so feel free to decline them if you don't think they're worth the bother.

Comment 1

[Dan Gohman [:sunfish]]

Attachment: range-pow-half.patch

Comment 2

[Dan Gohman [:sunfish]]

Attachment: range-numeric-not.patch

Comment 3

[Nicolas B. Pierron [:nbp]]

Comment on attachment 822919 [details] [diff] [review]
range-ordered-compare.patch

Review of attachment 822919 [details] [diff] [review]:
-----------------------------------------------------------------

Rename the flag to canBeNaN (don't forget to invert the meaning of it), and r=me.
I wonder how we managed to keep this around for a while.

::: js/src/jit/MIR.h
@@ +2140,5 @@
>    private:
>      CompareType compareType_;
>      JSOp jsop_;
>      bool operandMightEmulateUndefined_;
> +    bool alwaysOrdered_;

alwaysOrdered_ is not clear if you are thinking in terms of JavaScript, I think it makes more sense to rename it canBeNaN_.

Comment 4

[Nicolas B. Pierron [:nbp]]

(In reply to Nicolas B. Pierron [:nbp] from comment #3)
> Comment on attachment 822919 [details] [diff] [review]
> range-ordered-compare.patch
> 
> Review of attachment 822919 [details] [diff] [review]:
> -----------------------------------------------------------------
> alwaysOrdered_ is not clear if you are thinking in terms of JavaScript, I
> think it makes more sense to rename it canBeNaN_.

After looking at the second patch, I think it makes more sense to use the same naming as in range-pow-half (operandsAreNeverNaN_), as we are not dealing with the return type of the MCompare.

Comment 5

[Nicolas B. Pierron [:nbp]]

Comment on attachment 822922 [details] [diff] [review]
range-pow-half.patch

Review of attachment 822922 [details] [diff] [review]:
-----------------------------------------------------------------

::: js/src/jit/shared/CodeGenerator-x86-shared.cpp
@@ +477,4 @@
>  
> +        Assembler::DoubleCondition cond = Assembler::DoubleNotEqualOrUnordered;
> +        if (ins->mir()->operandIsNeverNaN())
> +            cond = Assembler::DoubleNotEqual;

Is that a performance issue?
Can we just avoid adding this extra flag and always used the …OrUnordered condition?

Comment 6

[Dan Gohman [:sunfish]]

(In reply to Nicolas B. Pierron [:nbp] from comment #4)
> After looking at the second patch, I think it makes more sense to use the
> same naming as in range-pow-half (operandsAreNeverNaN_), as we are not
> dealing with the return type of the MCompare.

Done.

https://hg.mozilla.org/integration/mozilla-inbound/rev/e9ff16009983
https://hg.mozilla.org/integration/mozilla-inbound/rev/3d175be0fcb0

Comment 7

[Dan Gohman [:sunfish]]

(In reply to Nicolas B. Pierron [:nbp] from comment #5)
> Comment on attachment 822922 [details] [diff] [review]
> range-pow-half.patch
> 
> Review of attachment 822922 [details] [diff] [review]:
> -----------------------------------------------------------------
> 
> ::: js/src/jit/shared/CodeGenerator-x86-shared.cpp
> @@ +477,4 @@
> >  
> > +        Assembler::DoubleCondition cond = Assembler::DoubleNotEqualOrUnordered;
> > +        if (ins->mir()->operandIsNeverNaN())
> > +            cond = Assembler::DoubleNotEqual;
> 
> Is that a performance issue?
> Can we just avoid adding this extra flag and always used the …OrUnordered
> condition?

Yes, it is a performance issue. DoubleNotEqual is more efficient than DoubleNotEqualOrUnordered on x86/x64 because it doesn't seed a separate NaN check.

Comment 8

[Phil Ringnalda (:philor)]

https://hg.mozilla.org/mozilla-central/rev/e9ff16009983
https://hg.mozilla.org/mozilla-central/rev/3d175be0fcb0

Comment 9

[Dan Gohman [:sunfish]]

Comment on attachment 822922 [details] [diff] [review]
range-pow-half.patch

(In reply to Dan Gohman [:sunfish] from comment #7)
> (In reply to Nicolas B. Pierron [:nbp] from comment #5)
> > Comment on attachment 822922 [details] [diff] [review]
> > range-pow-half.patch
> > 
> > Review of attachment 822922 [details] [diff] [review]:
> > -----------------------------------------------------------------
> > 
> > ::: js/src/jit/shared/CodeGenerator-x86-shared.cpp
> > @@ +477,4 @@
> > >  
> > > +        Assembler::DoubleCondition cond = Assembler::DoubleNotEqualOrUnordered;
> > > +        if (ins->mir()->operandIsNeverNaN())
> > > +            cond = Assembler::DoubleNotEqual;
> > 
> > Is that a performance issue?
> > Can we just avoid adding this extra flag and always used the …OrUnordered
> > condition?
> 
> Yes, it is a performance issue. DoubleNotEqual is more efficient than
> DoubleNotEqualOrUnordered on x86/x64 because it doesn't seed a separate NaN
> check.

The question answered, I'm re-requesting review, so I don't loose track of this. It's not urgent.

Comment 10

[Dan Gohman [:sunfish]]

https://hg.mozilla.org/integration/mozilla-inbound/rev/91d286c5eecc

Comment 11

[Ryan VanderMeulen [:RyanVM]]

https://hg.mozilla.org/mozilla-central/rev/91d286c5eecc