🚀 go-pugleaf

Does gcc 2.0 have any sort of "noalias" provision?  Or can it automatically
assume such a thing in some cases?


Background from a non computer-scientist:

As far as I know a "noalias" facility is usually useful for numerical codes:


Consider:

void affine_transformation(int N, double *A, double B, double *C, double *X)
{
	/* implement vector operation:  X= A + B * C
	int i;

	for (i=1; i <= N; i++) {

		X[i] = A[i] + B * C[i];

	}
	
}


The problem comes about, because as arrays are passed as pointers in C,
the   X[i] may happen to lie at the same address as A[j], or something
like that.  As C is defined in terms of a serial model of execution, the
compiler has to follow the steps explicitly.

Of course, in most code, this never happens.  But because it's legal
in C, it can inhibit optimization.  A "noalias" facility allows one
to somehow let the compiler know that this isn't a possiblity.

FORTRAN, for example, disallows this aliasing by assumption
(it will generate "wrong" results if they arguments overlap) , and so
can still generate better code than C compilers for situations like
these.


Specifically, I have some hand-translated C versions of LINPACK routines.

I think they run slower than FORTRAN for this reason.  However, if I could
specify noalias, and declare vector routines "inline" (I actually do
this already) things might be better.



Matt Kennel
mbk@inls1.ucsd.edu