Path: utzoo!utgpu!jarvis.csri.toronto.edu!mailrus!ncar!ico!ism780c!haddock!karl
From: karl@haddock.ima.isc.com (Karl Heuer)
Newsgroups: comp.lang.c
Subject: Re: faster bcopy using duffs device
Keywords: loop unrolling, optimize, hacks
Message-ID: <14558@haddock.ima.isc.com>
Date: 8 Sep 89 20:45:42 GMT
References: <5180@portia.Stanford.EDU> <19473@mimsy.UUCP>
Reply-To: karl@haddock.ima.isc.com (Karl Heuer)
Organization: Interactive Systems, Cambridge, MA 02138-5302
Lines: 13

In article <19473@mimsy.UUCP> chris@mimsy.UUCP (Chris Torek) writes:
>bcopy() should be written in assembly (on most processors), put in
>a library, and forgotten about...

Yes.  (And it should be called memcpy(), that being the standard name; of
course bcopy() can be an alternate synonym if it's done properly.)  But if you
really want to blow the socks off the benchmarks, you should have the compiler
recognize memcpy() as a builtin.  This not only allows it to be inlined, but
also makes possible certain compile-time optimizations such as constant length
move (for which unrolling might be a win) and known-alignment pointers (for
which it could safely copy larger-sized chunks).

Karl W. Z. Heuer (ima!haddock!karl or karl@haddock.isc.com), The Walking Lint