Path: utzoo!utgpu!news-server.csri.toronto.edu!rpi!usc!elroy.jpl.nasa.gov!swrinde!mips!dimacs.rutgers.edu!rutgers!cmcl2!lanl!cochiti.lanl.gov!jlg
From: jlg@cochiti.lanl.gov (Jim Giles)
Newsgroups: comp.lang.c
Subject: Re: low level optimization
Message-ID: <22021@lanl.gov>
Date: 19 Apr 91 22:52:08 GMT
References: <21846@lanl.gov> <1991Apr18.233807.19552@zoo.toronto.edu> <21868@lanl.gov> <3906@inews.intel.com>
Sender: news@lanl.gov
Organization: Los Alamos National Laboratory
Lines: 25



> You're so hyped up about the fact that separate compilation
> prevents use of this optimization, but you haven't been
> so vocal about the case where you deliberately alias one
> of your arrays, which also blows your optimization.

Since I _never_ deliberately alias array arguments either to each
other or to globals to which the procedure I'm calling has access,
this is not an issue to me.  In _most_ code (not just mine) _most_
of the arrays that you pass around as parameters are not aliased
and _most_ C compilers forgo a perfectly valid optimization in
order to cater to the much rarer case where someone had "deliberately"
aliased something.  Deliberate aliasing happens all the time in
processing recursive data structures (like linked lists, trees, graphs,
etc.) but for array manipulation it is almost always a serious error -
no matter how carefully the compiler works.  It would be nice if C
let me pass arrays to procedures _without_ converting them to pointers
and it would be nice if the compiler were free to assume that array
parameters were not aliased to each other or to any globals.  If they
become aliased to any locals, that could only happen through loaclly
visible action that the compiler can see without interprocedure
analysis.

J. Giles