Re: [9fans] "ridiculous benchmarks"-r-us

[Roman Shaposhnick <rvs@sun.com>]

On Fri, Jan 09, 2004 at 10:42:49AM -0700, mirtchov@cpsc.ucalgary.ca wrote:
> Rog at vitanuova already caught how stupid the following article is
> (linked to from slashdot):
>
> 	http://osnews.com/story.php?news_id=5602
>
> It took me much longer -- until I looked at their "C" code (term used
> very loosely) trying to compile it on Plan 9...  There are multiple
> gems in it including ++ on a double, mid-function variable definitions
> (C-code, right?), log10(0.0), impossible to look at I/O, etc.  In
> fact, I'm attaching the benchmark for your viewing pleasure :)

  Benchmarking nowdays, is a real pain in the kazoo, if you know what I
  mean. Benchmarking compiler itself is no way of demonstrating performance
  of the real application where heavily templated library like libm or
  even libc can make all the difference. The trends I've seen seem to be
  quite amusing when compilers are made to recognize parts of the AST
  (including calls to stuff like sqrt and such) and replace them with
  the predefine assembler template. For evert supported combination
  of arch. switches. Painful.

  The interesting part, of course is when we start considering JITs like
  HotSpot for example. Just recently I was quite amused to see what
  JDK 1.4 did for the simplest fibonacci. Quite educational, I might add.
  Somehow, I feel, that for the best performance possible one simply *has*
  consider execution environment with JIT-modifiable code. No way around it.

Thanks,
Roman.