Path: utzoo!attcan!uunet!convex!concave!loafman
From: loafman@concave.uucp (Kenneth W. Loafman)
Newsgroups: comp.sys.ibm.pc
Subject: Re: Microsoft Vs. Borland; benchmarks!
Summary: /Oa is broke.
Message-ID: <644@convex.UUCP>
Date: 14 Oct 88 00:07:32 GMT
References: <876@galaxy> <8254@haddock.ima.isc.com> <788@euraiv1.UUCP> <9088@haddock.ima.isc.com> <789@euraiv1.UUCP>
Sender: news@convex.UUCP
Reply-To: loafman@convex.UUCP (Kenneth W. Loafman)
Organization: Convex Computer Corporation, Richardson, Tx.
Lines: 21

In article <789@euraiv1.UUCP> evas@euraiv1.UUCP (Eelco van Asperen) writes:
>
>Oops; you're quite right on that one. -Ox implies -Oa, ie. "no-aliasing"
>and that's dangerous; most code won't be compiled correctly.
>Apparently, most benchmarks can be compiled with -Ox and run correctly;
>do I hear somebody grumbling about benchmarks and their value ;-) ?
>
Actually /Oa is broken in ways that have absolutely nothing to do with
aliasing.  I have one case where MSC 5.1 trashed a counter variable in a
double nested loop.  It was declared as automatic int.  Problem went away
when I turned off the /Oa option.  There was no possibility of aliasing in
that particular piece of code.

-----
Kenneth W. Loafman @ Convex Computer Corp, Dallas, Texas
UUCP:  {allegra,uiucdcs,ctvax}!convex!loafman
Disclaimer:  Well!  I never...
-----
Kenneth W. Loafman @ Convex Computer Corp, Dallas, Texas
UUCP:  {allegra,uiucdcs,ctvax}!convex!loafman
Disclaimer:  Well!  I never...