Newsgroups: comp.lang.c
Path: utzoo!utgpu!jarvis.csri.toronto.edu!dgp.toronto.edu!flaps
From: flaps@dgp.toronto.edu (Alan J Rosenthal)
Subject: Re: Suitably aligned pointers
Message-ID: <8810151524.AA16480@explorer.dgp.toronto.edu>
Organization: University of Toronto
References: <345@marob.MASA.COM>
Date: Sat, 15 Oct 88 11:24:47 EDT


In article <345@marob.MASA.COM> daveh@marob.UUCP (Dave Hammond) writes:
>The problem is the routines don't return a `suitably aligned' pointer,
>so the accuracy of casting the return to other than char * is in doubt.
>
>Can someone explain the technique used to suitably align a pointer ?

Well, it can't be done very nicely in a portable way.  It can be done very
easily in a machine-dependent way:  just do this:
	if ((long)p & 1)
	    p++;
for example to enforce even alignment.  Of course, your memory routine will
then have to allocate n + 1 bytes, counting the one you just skipped.  (Don't
assume your machines require even alignment, check it out.)

To make it a little more portable, you have to over-align, so to speak -- there
is no way to figure out the alignment requirement but there is a way to figure
out something which must be a multiple of the alignment requirement.  Take all
types you can think of (whether or not, for example, you must include all of
pointer-to-char, pointer-to-pointer-to-char, and pointer-to-int, can not be
determined, and since there are an infinite number of types this problem is not
solvable in general).  Make a union of them, and then take sizeof it.  If
you're worried that someone might have taken the ansi draft literally and left
out necessary padding of unions, wrap it in a struct first and take sizeof the
struct.

It is not possible to eliminate the assumption that a pointer can fit into some
integral type, although I suppose you could use unsigned long if you think that
would help.  On machines on which a pointer cannot fit into any integral type,
I suppose that a little assembly language is required.

ajr

--
#define __STDC__ 0