On Sat Feb 22, 2025 at 12:21 AM CET, G. Branden Robinson wrote:
> At 2025-02-22T00:11:13+0100, onf wrote:
> > By the way, Heirloom troff has floating point registers.
>
> Just converting the type backing groff's registers from `int` to `long`
> is a significant refactor, because it implicates other things like the
> formatting of diagnostic messages. I found this out when taking a stab
> at it a few years ago, assuming it would be easy.
>
> Floats aren't always better. They have variable instead of fixed
> relative error, for example. So I wouldn't just assume that replacing
> `int` with `double`--I assume no one would advocate `float`--would be an
> unambiguous win.
I wouldn't think so, either. I just thought it might be relevant to the
topic. I likewise think that troff's integer arithmetic is usually
sufficient, and the combination of base-2 floating point tradeoffs and
rise in internal complexity might not be worth it.
With that said, while problems with floating point roundoff errors are
well known, I've been wondering for a while whether the errors resulting
from repeated truncation are any better, or if they could in fact be
worse.
> Then one wonders, "why not both?"
> [...]
Well, that's what Heirloom troff does... it has both integer and
floating point registers, and it's in fact using the float data
type for the floating point ones :)
/* troff/tdef.h */
extern struct numtab {
int r; /* name */
short fmt;
int inc;
int val;
struct numtab *link;
int aln;
int nlink;
float fval;
float finc;
enum flags flags;
} *numtab;
~ onf
