Re: GCC's -ffast-math behavior

[xunxun <xunxun1982@xxxxxxxxx>]

于 2012/2/9 18:35, xunxun 写道:
> 于 2012/2/9 18:29, Andrew Haley 写道:
> > On 02/09/2012 10:20 AM, xunxun wrote:
> > > I use the code main.c (test sin speed)
> > >
> > > On Win7 64bit, gcc 4.6.2 32bit
> > >
> > > gcc -O3 -ffast-math main.c -o main.exe
> > >
> > > run main.exe will cost 6.853s.
> > >
> > > When linking with intel libM no fastmath
> > >
> > > gcc -O3 main.c -o main.exe libmmt.lib libircmt.lib
> > >
> > > run main.exe will cost 4.367s.
> > Ah okay, I get it now, its sin, not exp.  I suspect that it's the
> > argument reduction step that's slowing you down.
> >
> > Andrew.
> I think so. I don't know whether I may change main.c 's sin to exp
>
> It will be also slow.
And I think it's related with -funsafe-math-optimizations

If I only use the option, gcc also don't generate the symbol

The fastmath other option

 -fno-math-errno, -ffinite-math-only, -fno-rounding-math, -fno-signaling-nans, and -fcx-limited-range

don't have the effect.

--
Best Regards,
xunxun