Path: chuka.playstation.co.uk!news
From: "Russ Williams" <russ@algorithm.demon.co.uk>
Newsgroups: scee.yaroze.programming.3d_graphics
Subject: Re: More Pointless Optimisation Tricks (sin, cos
Date: Fri, 24 Oct 1997 20:13:32 +0100
Organization: PlayStation Net Yaroze (SCEE)
Lines: 48
Message-ID: <62qrli$ju02@chuka.playstation.co.uk>
References: <344647ee.14591228@news.playstation.co.uk> <625jjg$e3p24@chuka.playstation.co.uk> <3446d0a3.18235002@news.playstation.co.uk> <627afa$e3p28@chuka.playstation.co.uk> <62aiaf$e3p29@chuka.playstation.co.uk>
NNTP-Posting-Host: algorithm.demon.co.uk
X-Newsreader: Microsoft Outlook Express 4.71.1712.3
X-MimeOLE: Produced By Microsoft MimeOLE V4.71.1712.3

Graeme Evans wrote in message <62aiaf$e3p29@chuka.playstation.co.uk>...
>
>SCEE wrote in message <627afa$e3p28@chuka.playstation.co.uk>...
>>I've done the swap thing before on PC, and it was only quicker if when
>>doing >the normal swap via temp the temp had to be declared every
>>time.As for PSX, it'll definetly be slower, you could write a little
inline
>>assembler routine.
>>
>>Stuart
>
>Doesn't it depend on register spills? Does an XOR take more cycles
>than a register move?


I doubt it. Most RISC cores execute simple arithmetic / logic ops in one
(pipelined) clock.

The XOR swap would assemble to something like:
    xor v0,v0,v1
    *

    xor v1,v0,v1
    *

    xor v0,v0,v1
    ...

The * instructions may need to be nops. I don't know if the R3000
needs a delay slot for flow dependencies, but the code can't be
any faster - you can't xor with a value until it's been calculated...

That should execute in 2.5/3 clocks (assuming 2-issue superscalar)

The move method might be:
    add a0,zero,v0
    or v0,zero,v1

    xor v1,zero,a0
    ...
(ie: 1.5/2 clocks)

Apparently, the R3000 has no 'move' instruction. Wierd!

---
Russ