Relay-Version: version B 2.10 5/3/83; site utzoo.UUCP
Path: utzoo!mnetor!uunet!seismo!columbia!rutgers!mit-eddie!uw-beaver!fluke!kurt
From: kurt@tc.fluke.COM (Kurt Guntheroth)
Newsgroups: comp.misc
Subject: Better compilers (1.3 wish)
Message-ID: <1428@sputnik.COM>
Date: Fri, 7-Aug-87 11:44:32 EDT
Article-I.D.: sputnik.1428
Posted: Fri Aug  7 11:44:32 1987
Date-Received: Sun, 9-Aug-87 02:11:26 EDT
References: <8707190424.AA10158@cogsci.berkeley.edu> <434@sugar.UUCP> <3664@well.UUCP> <4564@jade.BERKELEY.EDU>
Sender: news@tc.fluke.COM
Organization: John Fluke Mfg. Co., Inc., Everett, WA
Lines: 34

I agree.  The best optimizing compilers are now better than the best
human-coded assembly language for program more than a few hundred lines
long.  We should be spending our collective time writing better compilers
rather than recoding in assembler.  By the way, the best compilers are
better than humans now because they do things to the code that are
considered bad style for a human-maintainable assembler source, like reusing
the same storage location for different variables, storing the same quantity
in different locations/registers during the program run, moving code around,
etc.

By the way, it is hard to optimize c very highly because of things like
pointers and lax type checking.  In principle, the best-possible object code
from a pascal program will be tighter than the best-possible object code
from a c program doing the same thing.  The tricks c uses to permit a
relatively crummy compiler to produce decent code hobble a really good
compiler.

And now an anecdote.  I once worked for a small computer manufacturer you've
never heard of.  We had a sales rep in California who thought he was the
hottest thing ever to invoke an assembler.  He had a customer who neeeded a
large hierarchical database sorted every day.  Unfortunate thing was, the
sort took about 30 hours to crunch the day's data.  This guy used an
undocumented and difficult to use entry point to the operating system to
permit him to code an in-memory sort in (sort of) microcode.  He managed to
get the sort down to 15 hours.  During this time I was working on a standard
sort for our new database.  My bosses said "call this guy up, he knows
everything about sorting."  Well, to make a long story sort (uh, short), he
was using an O(n squared) sort.  In 2 weeks I coded a completely general
purpose merge sort in the structured assembler that substituted for a
programming language.  I sorted this guy's database in about an hour and a
half.  So what's the big deal about assembler?  Algorithms and tuning is
where it's at.

Kurt Guntheroth