simonmarhaskell at gmail.com
Thu May 11 09:57:30 EDT 2006
Ben Rudiak-Gould wrote:
> John Meacham wrote:
>> also, incidentally, for anyone on x86 that cares about math performance,
>> use -optc-fsse2 to make it use the much nicer math coprocessor available
>> on modern x86 cpus.
> I object to its characterization as "nicer". It's faster, but *lower
> precision*. It worries me that people are so blithely abandoning those
> extra bits in the name of speed. A few years from now there's going to
> be an expensive engineering failure, and months of investigation will
> reveal that it was because a math library was compiled for SSE2. Be
> careful out there.
On the other hand, keeping intermediate Doubles to 80-bit precision is
both (a) non-portable and (b) unpredictable (the programmer doesn't know
which intermediates are going to be stored in 80 bits, and turning on
optimisation will probably make a difference).
I suppose you might argue that "extra precision is always good". But I
like it when Haskell programs give the same results, regardless of the
platform, compilation strategy, and level of optimisation.
> (And while I'm on the subject, Haskell should have a LongDouble type.)
LongDouble would be fine, but storing intermediate Doubles in 80 bits is
More information about the Glasgow-haskell-users