RE: [Algorithms] Floating point comparison.

[Stefan B. <ste...@te...>]

> While mucking about in the debugger I noticed that comparing two
floating
> point numbers for equality actually generates FPU instructions rather
than
> just performing an integer comparison.
> <...> 
> So the fact that the compiler feels motivated to use FPU instructions
> either means it's lame (unlikely) or there are floating point numbers
with
> differing bit patterns that are equal. Other than +0 and -0 I can't
think
> of any.

  Well I don't have the CPU manuals at hand, but I would imagine a
comparing a NAN with another NAN should always yield a false result, and
likewise for INF?

  As an aside, I would recommend not explicitly hard-coding float-as-int
comparisons in your code when you optimise... on some architectures
(MIPS etc) float comparisons are quite fast, and by using integer
comparisons you might actually end up slowing things down unless you're
just comparing variables that you know are stored in memory (i.e. not
results of recent computations). Personally I use some simple inline
functions for quick compares.

Cheers,
Stef
--
Stefan Boberg, R&D Manager, Team17 Software Ltd.
bo...@te...