Re: g77 performance for real*complex

[Toon Moene <toon at moene dot indiv dot nluug dot nl>]

June-Yub Lee wrote:

> Maybe you alreay know that the problem of real*complex but

No, I hadn't realised this.

> let me present a simplest example:
> > cat t.f
>         real*8 x
>         complex*16 z1, z2
>         x = 1.1d0
>         z1 = (2.2d0, 3.3d0)
>         z2 = x*z1
>         end
> 
> The program needs only TWO double precision multiplication (fmuld):
> x*dble(z1) for the real part, x*dimag(z1) for the imaginary part.
> However, the g77 (2.95.1) generates 2*fmuld & fsubd for the real part
> and 2*fmuld & faddd for the imaginary part. And it comsumes 6 flops
> for real*complex even under -O2 optimization mode.

The reason for this behaviour is simple:  the DOUBLE PRECISION variable
X first is converted to the DOUBLE COMPLEX variable (1.1d0, 0.0d0)
before the DOUBLE COMPLEX multiplication (1.1d0, 0.0d0) * (2.2d0, 3.3d0)
is done [according to the promotion rules of the Fortran Standard].  The
latter needs 4 multiplications and 2 additions 1) (one with a minus
sign).  Apparently, the SUN compiler knows how to make use of the fact
that the imaginary part of a real number is zero.

Clever trick - if I have some spare time (hah :-) I might look into
teaching the Fortran frontend such tricks.

1) (a + b i) * (c + d i) = a * c - b * d + (a * d + b * c) i

-- 
Toon Moene - mailto:toon@moene.indiv.nluug.nl - phoneto: +31 346 214290
Saturnushof 14, 3738 XG  Maartensdijk, The Netherlands
GNU Fortran 77: http://gcc.gnu.org/onlinedocs/g77_news.html
GNU Fortran 95: http://g95.sourceforge.net/ (under construction)