Open crusso opened 4 years ago
Using specifier "%.16g" not "%f" yields better result, but I don't know if that's good enough.
rts/float.c
export as_ptr float_fmt(double a) {
extern int snprintf(char *__restrict, size_t, const char *__restrict, ...);
char buf[50]; // corresponds to 150 bits of pure floatness, room for 64 bits needed
// const int chars = snprintf(buf, sizeof buf, "%f", a);
const int chars = snprintf(buf, sizeof buf, "%.17g", a);
return text_of_ptr_size(buf, chars);
}
@rossberg any advice here?
Yes, %.17g
is enough to roundtrip, I've used that in the Wasm interpreter, and it works for all the many corner cases in the test suite that exercise bit-precise rounding behaviour (at least OCaml's implementation of the formatting, which I assume is inherited from C).
However, in practice, users may not want to see umteen zeroes all the time. We'll have to add some richer formatting functions; toText should probably default to %f
.
There's also %h
, of course.
Something like toText(f : Float, prec : ?Nat8)
would be easily realisable (in the compiler, for the interpreter probably hairier).
@ggreif, I think it should be a separate function from toText, though. But then we'd probably want s.th like SML's realfmt type, maybe extended with a hex
case for %h
.
In the interpreter that ought to be trivial (using OCaml's Printf).
@rossberg In OCaml we can't build the format string dynamically, can we?
val fmt : string = "%.9g"
# Printf.sprintf fmt 3.14159;;
Error: This expression has type string but an expression was expected of type
('a -> 'b, unit, string) format =
('a -> 'b, unit, string, string, string, string)
CamlinternalFormatBasics.format6
Following SML's basis library, here is what I propose: a function Float.toFormattedText : (Float, Format) -> Text
, where the type Float.Format
is defined as follows:
type Format = {
#fix : Nat; // like "%.*f"
#exp : Nat; // like "%.*e"
#gen : Nat; // like "%.*g"
#hex : Nat; // like "%.*h" a.k.a. "%.*a"
#exact; // like "%.17g"
}
(Making precision non-optional, since that doesn't buy you much in this form of description.)
@ggreif, no, but a simple switch over the 4 or 5 formats is enough. The precision can be turned into a dynamic argument by using *
.
That's what I mean by "hairier".
How is that "hairy"?
It has more hair :-)
But we'll still need some supporting primitives for this to fly, right? (I'll come up with a PR in dfinity/motoko soon.)
Perhaps just extend float_fmt with a mode and a precision arg?
Great, thanks guys. I'll leave that in your hands then @ggreif.
I totally misunderstood what "%.*f"
means. I thought that the *
is purely meta-notation. Another thing learned...
Now I see why you considered it hairy. :)
Now I see why you considered it hairy. :)
Precisely.
Is there any way we can not have toFormattedText
at all and make it very very clear in the documentation that a simple toText
is not for displaying to users? Formatting floating point numbers for human consumption is a locale-dependent function, as there are many different ways that different cultures format numbers with varying thousands separators, digit groupings, decimal separators, rounding rules, etc. A separate number formatter class that takes a locale would be ideal and would keep floats and natural numbers clean and small.
see rts/float.c