Re: -ffast-math and gfortran

[Iwan Kawrakow <iwan at irs dot phy dot nrc dot ca>]

Steve,

> On Sunday 04 June 2006 18:58, Steve Kargl wrote:
> Go back in the thread and see the very simple program I wrote.
> It is well-behaved without -ffast-math but generates NaNs with
> this option.  
>
> Apparently, you do not use complex data types. 

I do sometimes.

> Or, do you expand all multiplications and divisions into the
> well-known numerically stable algorithms?  
> 
Yes, I always do. Based on the experience that in the long 
run it saves you time hunting for "unexpected" behavior. 
But one does not need to go to complex math to find 
examples where "well-behaved" programs fail. 
Everybody who writes code that deals with floating 
point arithmetic should understand the limitations posed 
by their limited precision and range. 

To come back to your example:

program h
  complex z
  z = cmplx(1.e-25,1.e-25)
  call div(z,z)
end program h
subroutine div(a,b)
  complex a, b, c
  c = a / b
  print *, c
end subroutine div

gfortran h.f90 && ./a.out
(  1.000000    ,  0.000000    )
gfortran -ffast-math  h.f90 && ./a.out
(           NaN,           NaN)
gfortran -fno-math-errno -funsafe-math-optimizations -fno-trapping-math -ffinite-math-only -fno-rounding-math -fno-signaling-nans 
h.f90 && ./a.out
(  1.000000    ,  0.000000    )
The last line uses all flags set by -ffast-math except fcx-limited-range. 

Someone who does not understand why your program behaves the way 
it does with the different options should look for another job. ;-)

As long as the fast math option is not removed altogether 
and the switch change is documented, your can call 
it  -frisky-math or whatever else you like :-)

Ivan