Path: utzoo!mnetor!uunet!mcvax!dik
From: dik@cwi.nl (Dik T. Winter)
Newsgroups: comp.lang.c
Subject: Re: More on noalias
Message-ID: <7488@boring.cwi.nl>
Date: 12 Jan 88 22:28:52 GMT
References: <10129@mimsy.UUCP> <3433@megaron.arizona.edu>
Organization: CWI, Amsterdam
Lines: 49

In article <3433@megaron.arizona.edu> mike@arizona.edu (Mike Coffin) writes:
 >  	void
 >  	vector_add /* maybe */ (int *a, int i, int j, int k, int n)
 >  	{
 >  		if (i==j || i==k) return;
 > 
 > 		/* now the compiler can infer that i!=j and i!=k */
 >  		for (; --n >= 0; i++, j++, k++)
 >  			a[i] = a[j] + a[k];
 >  	}
 > 
and supposes that this allows vectorization (just like the parent article
by Chris Torek).

First the conditions are wrong (if i equals both j and k, vectorization
is possible; no vectorization is possible if (j-n) < i < j, or
similar for i and k).  But that is beside the point; the equivalent
Fortran code will not vectorize on any machine that I know (Cray, CDC 205,
NEC SX/2, Fujitsu), because of this aliasing problem.  A better example is:
	void
	vector_add /* maybe */ (int *a, *b, *c, int n)
	{
	int i;
		for(i = 0; i < n; i++)
			a[i] = b[i] + c[i];
	}
This will vectorize in Fortran, because the standard states that aliasing
is not allowed here if a is involved.  It will not vectorize in C.
Adding a noalias for a will solve the problem (if I understand it right).

Of course the statement
		if((a > b && a < b+n) || (a > c && a < c+n)) exit(1);
(or EXIT_FAILURE) will help also, but the existence of such a statement
(or a similar one) is very difficult to check, because although in this case
the condition is extremely simple, it might be much more complex in
other cases. (What is the condition if the loop reads:
	int i, j;
		for(i = 0, j = 0; i < n; j += i, i++)
			a[j] = b[j] + c[j];
not very common, but also not unheard of.  It vectorizes on a number of
machines.)

Note that this example also clarifies Doug Gwyn's remark that noalias
changes the semantics.  Especially if the programmer lies.  (Without
noalias the result is well-defined, with it is undefined.)
-- 
dik t. winter, cwi, amsterdam, nederland
INTERNET   : dik@cwi.nl
BITNET/EARN: dik@mcvax