Simon Marlow wrote:
I suppose you might argue that "extra precision is always good".
Well... I'm having a hard time thinking of a situation where it isn't. I realize that people want reproducibility, I'm just not convinced that they should. The situations where optimization flags make a significant difference in the result are situations where the original result wasn't trustworthy to begin with. Reproducible rounding gives you a false sense of security in those cases, and doesn't make a difference otherwise. If you know IEEE 754 inside out, and IEEE double-precision is exactly what you want, then it's nice if your programming language can cater to that, but along those lines it'd be nice if Haskell supported rounding modes, trap handlers, and so on.
I suppose this problem could be solved the way the analogous Int problem is solved, with a machine-precision Float and portable/reproducible Float32 and Float64.
The x87 isn't the only FPU with this problem. The PowerPC has a fused (double*double)+double operation that the optimizer can't use if you need reproducibility, because it doesn't round the intermediate product.
LongDouble would be fine, but storing intermediate Doubles in 80 bits is bad.
The x87 registers are actually 82 bits wide, so even using LongDouble doesn't guarantee reproducibility. Maybe there needs to be a Float80 also. (Available only on x86, presumably.)
-- Ben _______________________________________________ Glasgow-haskell-users mailing list Glasgow-haskell-users@haskell.org http://www.haskell.org/mailman/listinfo/glasgow-haskell-users
