Bug 85752 – Truncating multiplication on integers should not OSR exit every time

• Component: JavaScriptCore
• Status: RESOLVED FIXED

[Filip Pizlo]

Say you do a * b where the result is larger than what can fit in an int32, but you use the result in integer contexts only (like [_] | 0).  Currently the DFG will "prove" that the multiplication will produce integers despite overflowing.  But then the multiplication will speculate no overflow, which will typically fail, since in this case a * b does overflow.

Multiplication should be smarter about this.  If you do a * b and we know that it overflows then in the general case our best bet is to perform the multiplication using the FPU.  But if we know something about a or b, for example if we really have something like a * 42, then we know that integer truncation would be equivalent to double truncation - and hence we can perform an integer multiplication and skip the overflow check.

[Filip Pizlo]

Created attachment 140441 [details]
the patch

[Gavin Barraclough]

Comment on attachment 140441 [details]
the patch

View in context: https://bugs.webkit.org/attachment.cgi?id=140441&action=review

r-, because I really do think we need a comment here – but nice optimization, looks great otherwise.

> Source/JavaScriptCore/dfg/DFGGraph.h:512
> +        int32_t twoToThe17 = 131072;

Why 2^17?

A number with an absolute value less than 2^17 is representable by 17 bits.  The result of multiplying a value representable by 17 bits by 2^31 (the largest possible absolute value of a signed integer) is representable in a 48 bit mantissa (with additional representation for the sign).  Doubles support 53 bits of mantissa precision.  Put the other way around, for the result of a multiply to be constrained within a 53 bits mantissa, if I know that the absolute value of one operand is no more than 2^31 then the second operand may have a magnitude of +/- 2^22 (less than/greater than – exclusive!).

I think 2^22 is correct here, but either way this really needs a comment. :-)

Also, I'd suggest this would slightly nicer written ''int32_t  twoToThe17 = 1 << 17;".

[Filip Pizlo]

(In reply to comment #2)
> (From update of attachment 140441 [details])
> View in context: https://bugs.webkit.org/attachment.cgi?id=140441&action=review
> 
> r-, because I really do think we need a comment here – but nice optimization, looks great otherwise.
> 
> > Source/JavaScriptCore/dfg/DFGGraph.h:512
> > +        int32_t twoToThe17 = 131072;
> 
> Why 2^17?

Yeah, that's sloppy.  It was the largest I tested and I figured it was good enough for the optimization to be effective.  I will make this more rational!

[Filip Pizlo]

Created attachment 140444 [details]
the patch

Incorporated Gavin's feedback.

[Filip Pizlo]

Landed in http://trac.webkit.org/changeset/116264

[Filip Pizlo]

Merged in http://trac.webkit.org/changeset/117993