Path: csiph.com!v102.xanadu-bbs.net!xanadu-bbs.net!feeder.erje.net!eu.feeder.erje.net!fu-berlin.de!dotsrc.org!filter.dotsrc.org!news.dotsrc.org!not-for-mail
X-Newsreader: iForth 2.0 console (October 21, 2006)
From: mhx@iae.nl (Marcel Hendrix)
Subject: Re: revisiting a prgramming problem
Newsgroups: comp.lang.forth
Message-ID: <07078910938433@frunobulax.edu>
Date: Sat, 25 Oct 2014 14:28:28 +0200
References: <544a2a5e$0$6868$e4fe514c@dreader36.news.xs4all.nl>
Lines: 82
Organization: SunSITE.dk - Supporting Open source
NNTP-Posting-Host: 82.157.147.127
X-Trace: news.sunsite.dk DXC=>D11<e9ScK^07<M:5AA`SVYSB=nbEKnk[aHGbJQVQ?QSAFQDNO0X`FR05[\_Lj_;YY6D3KhK?DWD]XA1XFke>T<R
X-Complaints-To: staff@sunsite.dk
Xref: csiph.com comp.lang.forth:34135

albert@spenarnc.xs4all.nl (Albert van der Horst) writes Re: revisiting a prgramming problem

> In article <78081413938433@frunobulax.edu>, Marcel Hendrix <mhx@iae.nl> wrote:
> <SNIP>

>>But the 'best' (defined as: no jumps/compares) is of course:

>>      \ 22 ins, 0 jumps
>>
>>: sqr ( a -- a^2 ) dup * ;
>>: suosq ( a b c -- d )
>>  sqr swap sqr rot sqr 3dup ( -- a^2 b^2 c^2 a^2 b^2 c^2 )
>>  + >r + r> max
>>    >r + r> max ;

> Correct me of I'm wrong, but getting the squares first,
> seems to fail if any of the numbers is negative.

Who specified that? Anyway, it is only three instructions more:

: suosq ( z y x -- a )
  2dup 
  MAX SQR -rot ( -- z u^2 y x )
  MIN rot  
  MAX SQR + ;

\ 25 ins, no jumps
FORTH> ' suosq idis
$01229480  : sumsq
$0122948A  pop           rbx
$0122948B  mov           rdi, [rsp] qword
$0122948F  push          rbx
$01229490  mov           rcx, rdi
$01229493  cmp           rbx, rcx
$01229496  cmovl         rbx, rcx
$0122949A  imul          rbx, rbx
$0122949E  pop           rdi
$0122949F  pop           rax
$012294A0  push          rbx
$012294A1  mov           rbx, rdi
$012294A4  mov           rcx, rax
$012294A7  cmp           rbx, rcx
$012294AA  cmovg         rbx, rcx
$012294AE  pop           rdi
$012294AF  pop           rax
$012294B0  push          rdi
$012294B1  mov           rcx, rbx
$012294B4  mov           rbx, rax
$012294B7  cmp           rbx, rcx
$012294BA  cmovl         rbx, rcx
$012294BE  imul          rbx, rbx
$012294C2  pop           rdi
$012294C3  lea           rbx, [rdi rbx*1] qword
$012294C7  push          rbx
$012294C8  ;


> As is usual with your benchmarks the conclusion fails.

> What is your opinion, impression or advice based on all the
> work you've done. (Not that it is not appreciated that you
> present results in great detail, and be specific about
> the Forth it runs on. )

My advice is to write critical code as straightforward and clearly as 
possible. Current Forth compilers use all the tricks in the book 
(as demonstrated - remove all stack words, remove branches, remove flags). 
You can't beat these compilers with a CODE definition unless your first 
name is Terje. 
However, always try a few algorithms and check the output of the compiler 
for critical work. For example, I rejected a few algorithms that used 
locals because there were more items than the current iForth compiler 
can handle.

The basic assumption up to now has been that conditional jumps are 
bad for the run-time. I didn't test that here, and it might very well be
wrong. Anyway, I suspect the difference will be so small that it only 
matters for those that are using optimizing compilers (and then in only 
extreme cases).

-marcel