Path: utzoo!utgpu!watserv1!watmath!att!cbnewsi!yam
From: yam@cbnewsi.att.com (toshihiko.yamakami)
Newsgroups: comp.arch
Subject: Re: Compiler Costs
Message-ID: <1990Jul12.182552.26715@cbnewsi.att.com>
Date: 12 Jul 90 18:25:52 GMT
References: <628@dg.dg.com>
Organization: AT&T Bell Laboratories
Lines: 43

I have little experience on assembler coding,
so I might make rather silly comments.

From article <628@dg.dg.com>, by Publius@dg.dg.com (Publius):
> Many of the optimization problems belong to the "NP-complete" or even
> the "Turing-uncomputable" categories.  Thus, no matter how smart the
> global optimizer is, as long as it is equivalent to a Turing machine,
> it can not achieve the FULL optimization for EVERY situation.

I always wonder why human beings can do 'Turing-uncomputable' computation.
If human beings can do, why can't a program simulate it?
If anyone is kind enough to mail me the reason, I will appreciate it.

> If a large program spends most of the CPU time in a tiny routine, then,
> go ahead and hand-code it in assembler and do an exhaustive optimization,
> as long as you isolate it and document it comprehensively.

It is "comp.arch", so I would like to comment on RISC chips
and optimization. Following discussion may not be applied to CISC.
I believe it was true that hand-code in assembler was fast
than compiler code in CISC era.
however, some of RISC chip's instruction set is so simple,
so I wonder what people can do in hand on such a code.
For example, let's consider MIPS R2000(or R3000).
Last year, I tried to my first assembler optimization in my life.
Our group needed the enciphering performance of 1 Mbit/sec/MIPS.
To achieve it, I was forced to do some hand optimization.
To be honest, I don't like assembler coding(I am biased, you know :-) ).
I tried to re-write compiler output code on MC68030, SPARC, R2000.
I could improve performance about three times in MC68030, 5-10% on SPARC,
nothing(to be honest, negativbe) on R2000.
To my surprise, every time I rewrote the R2000 code, the performance
was decreased, and I gave up rewriting.
After that, I believe it is almost impossible to improve
performance in hand-coding on some types of RISC chips.

I am happy if I can hear other assembler experts's hand-coding experience
on MIPS or similar RISC chips.

-- 
Toshihiko YAMAKAMI(NTT, Japan) Resident visitor in Bell Labs until Feb 1991
 Room 4G634, AT&T Bell Labs, Crawfords Corner Rd. Holmdel, NJ 07733-1988
 Tel:(201)949-5742	e-mail: yam@vax135.att.com (was: yam@nttmhs.ntt.jp)