https://gcc.gnu.org/bugzilla/show_bug.cgi?id=107753
kargl at gcc dot gnu.org changed:
What |Removed |Added
----------------------------------------------------------------------------
Priority|P3 |P4
--- Comment #10 from kargl at gcc dot gnu.org ---
(In reply to Steve Kargl from comment #9)
> On Fri, Nov 18, 2022 at 11:24:29PM +0000, sgk at troutmask dot
> apl.washington.edu wrote:
> >
> > Does anyone know what is meant by "Fortran rules"? F66 does not
> > have any particular algorithm specified. I'll look at F77 shortly.
> >
>
> Well, I hunted down the origins of -fcx-fortran-rules.
>
> https://gcc.gnu.org/bugzilla/show_bug.cgi?id=29549
>
> So, it appears to be an optimization, where Smith's algorithm
> will fail for extreme values of the real and imaginary parts
> of the complex number.
So, I wrong a dirty little program to time complex division.
program foo
use timerm, only : rdtsc
implicit none
integer, parameter :: n = 1024*1024, dp = kind(1.d0)
real(dp) re(n), im(n)
complex(dp) x(n)
integer i
integer(8) t0, t
t = 0
do i = 1, 10
call random_number(re)
call random_number(im)
x = cmplx(4 * re, 10 * im,8)
t0 = rdtsc()
x = x / x
t = t + (rdtsc() - t0)
end do
print '(G0,1X,G0)', x(1)
print *, real(t,10) / 10 / n
end program foo
Compiled with gfortran with its current method of doing division
(i.e., -fcx-fortran-rules), I see roughly 44.5 clock ticks per
division. If run with a patched gfortran that uses the method that
the C compiler uses, I get about 62 ticks per division. So, using
the stricter method impacts performance.
I'll note that gfortran unilaterally enforces -fcx-fortran-rules,
i.e., -fno-cx-fortran-rules has no effect. Perhaps, gfortran
could be given a new -fcx-division=XXX option, where XXX is one of
naive --> what -ffast-math does (flags_complex_method = 0)
smith --> what -fcx-fortran-rules (flags_complex_method = 1)
strict -> default C behavior (flags_complex_method = 2)
