Optimization of gfortran for execution speed

[Paul Thomas]

Andrew,

> A way to speed/fixing this up, is to make struct array1_real4 for each 
> array
> at the beginning of the function, even though they might not be used.
> You should be able to let DCE do its job of cleanuping up unused structs.

I was going to start trying to understand how to do that... I might be a 
little while!  Can you give me a pointer, please?

> Note you should test to make sure this speeds up other programs that just
> small benchmarks then this.

I have no intention of sticking to little benchmarks. As I made plain, 
the problem emerged from two of the Polyhedron Benchmarks and Steven 
Bosscher tells me that one of the spec benchmarks could likely benefit.  
The reason that I reduced it was that I could not quite believe what the 
tests on the Polyhedron programs were telling me!  I very much want o 
sort this out so that I can use gfortran for production programs.

What I have identified is, of course general to all function calls.  
However dot_product happens to be very sensitive to this over head 
because the vectors being dotted often are of dimension 3, for example.....

Best regards

Paul