Re: [Patch match.pd] Add a simplify rule for x * copysign (1.0, y);

[Michael Matz <matz@suse.de>]

Hi,

On Thu, 1 Oct 2015, James Greenhalgh wrote:

> > >  x * copysign (1.0, y)
> > > 
> > >  x ^ (y & (1 << sign_bit_position))
> > 
> > Also I think this can only be done for finite and non trapping types. 
> 
> That may be well true, I swithered either way and went for no checks, 
> but I'd happily go back on that and wrap this in something suitable 
> restrictive if I need to.

I don't think that's necessary.  copysign (1.0, y) is always 1.0 or -1.0, 
even with y being a NaN or inf.  Additionally copysign is allowed to not 
signal even if y is a sNaN.  That leaves only the form of x to doubt.  If 
x is inf all is well (multiplying by +-1.0 is defined and both sequences 
get the same result), if x is NaN the result must be a NaN, and it is in 
both cases.  The catch is that strictly speaking (NaN * -1.0) needs to 
deliver NaN, not -NaN (operations involving quiet NaNs need to provide 
one of the input NaNs as result), and here both are not equivalent.  OTOH 
the sign of NaNs isn't specified, so I think we could reasonably decide to 
not care about this case (it would have to be checked if the hardware 
multiplication even follows that rule, otherwise it's moot anyway).

And yes, also on x86-64 cores the new sequence would be better (or at 
least as good; the latency of xor[sp][sd] is less than or equal to mul), 
but that only is the case if the arithmetic really happen in SSE 
registers, not in integer ones, and this isn't the case right now.  Hmpf.


Ciao,
Michael.