State of C99 _Complex?

[Scott Robert Ladd]

My experiments suggest that GCC 3.3 generates slow code when it comes to 
implementing the C99 _Complex types on the Pentium 4.

I'm not sure if the problem is in my use of gcc (i.e., I missed a subtle 
compile-time switch), or if _Complex support is simply "weak" at this point.

I've tried several complex-dependent applications, and all have 
performed slowly when compiled with gcc 3.3.

A case in point involves multiplication and addition of double-precision 
complex numbers. The C version of the code, using double complex, runs 
in 630 seconds, while the C++ version of the code, using complex<double> 
runs in 285 seconds.

In other words: The intrinsic C types are twice as slow as the C++ 
template implementation. And that just doesn't seem *right*. ;)

The performance characteristics appear to be constant across different 
sets of compiler switches. I've used a number of different options 
(-mfpmath=sse, -mfpmath=387, ffast-math, etc.), to no avail; gcc's 
complex code is just very, very slow.

I suppose I should try a Fortran 77 program, to see if g77 suffers from 
the same problem.

Is this a known issue?

I can dig in deeper into the assembler output if no one else has a 
handle on this.

-- 
Scott Robert Ladd
Coyote Gulch Productions (http://www.coyotegulch.com)