Path: utzoo!utgpu!water!watmath!clyde!rutgers!lll-lcc!ames!ucbcad!zodiac!sri-spam!sri-unix!quintus!ok
From: ok@quintus.UUCP (Richard A. O'Keefe)
Newsgroups: comp.lang.c
Subject: Re: a couple quickies
Summary: loop should execute 0 times
Message-ID: <510@cresswell.quintus.UUCP>
Date: 8 Jan 88 01:25:30 GMT
References: <11145@brl-adm.ARPA>
Organization: Quintus Computer Systems, Mountain View, CA
Lines: 49

In article <11145@brl-adm.ARPA>, PEPRBV%CFAAMP.BITNET@husc6.harvard.EDU (Bob Babcock) writes:
> TLIMONCE%DREW.BITNET@CUNYVM.CUNY.EDU asks
> >>Question: Under ANSI C,
> >>                    for (i = 0; i == 100; i++) foo(&i);
> >>If foo() does something like: "*i = 0" will this loop ever complete?
> >>I hope this is classified as undefined.
> 
> Seems to me that this is well defined,  and the loop should never
> terminate.   But I'm not sure that the answer doesn't change if i
> is declared noalias.

I thought a loop like this was supposed to be equivalent to
    {
	i = 0;
	while (i == 100) {
	    foo(&i);
	    i++;
	}
    }
Since 0 != 100, the body of the loop should never be executed, so it
had *better* terminate!

Just to keep the 'noalias' pot boiling,
    I asked what makes C so special that it needs 'noalias' when Fortran
    doesn't, and someone said "Pointers!".
Unfortunately, that is not an adequate answer.  It has been known for a
long time that pointers and array elements cause exactly the same problems
with respect to aliasing.  (For example, a Euclid program is, strictly
speaking, legal only if the I =\= J verification conditions the compiler
generates to express "no aliasing through array indices" are provable.)
What I mean by this is that if we have

	float a[1000];

	float fred(int i, int j)
	    {
		float x;
		x = a[i];
		a[j] -= 1.0;
		return x + a[i];
	    }

Is it safe to translate the last expression as x+x?  Only if i != j.
Ok, this is a contrived example, but the opportunity arises again and
again in typical numerical code.  (It is a really big problem for
vectorising compilers.)  So tell me, how do I use the 'noalias'
directive to tell a C compiler it is safe to assume i != j here?
And if 'noalias' cannot be used to tell the compiler that there is no
aliasing here, what good is it?