Maybe? I think (<.) is not properly specified for large floats. The
dictionary seems to imply that the result of (<.a) is the largest
integer b such that (b<:a) using tolerant comparison. But that's
certainly not the correct behavior of floor: shown below, for (a=2^45),
we would have to set (<.a) equal to (a+2)!

   ]b =. <.2^45
35184372088832
   (b+2) <: 2^45
1
   b =!.0 ]2^45
1

In fact, the J source uses the C code (x=jfloor(0.5+v), x-TGT(x,v)) to
compute tolerant floor for integer ranges. jfloor is a not-tolerant
floor which accepts and returns a float, and TGT(x,v) is J's tolerant
(x>v). This guarantees that the result of (<.a) is no more than one away
from (a). A proper definition from this code would be:

If y is tolerantly equal to an integer (including integers expressed in
floating point), (<.y) is the integer closest to y. Otherwise, (<.y) is
the largest integer less than y. The result of (<.) is of integer type
if and only if it is between 2^-63 and 2^63-1, inclusive.

So under this (reasonable, in my opinion) interpretation of (<.), J is
correct. In any case, I think it's important exactly because of
confusion like this to have a utility that performs conversions with
very precisely specified behavior.

Marshall

On Sat, Feb 06, 2016 at 08:22:37PM -0500, Raul Miller wrote:
> Yeah, it's a bug. But I'm pretty sure it's a bug in the interpreter,
> which your code is simply reporting on.
> 
> -- 
> Raul
> 
> 
> On Sat, Feb 6, 2016 at 7:53 PM, Marshall Lochbaum <[email protected]> 
> wrote:
> > The spec is that if there is a noun of the given type that matches, it
> > should be returned. So this is actually a bug, given:
> >
> >    MAX_ptype_
> > 9223372036854775807
> >    MAX_ptype_ -: 432.1-~2^63
> > 1
> >
> > (Note that (431-~2^63) is exactly the same as (2^63), as one can verify
> > with -:!.0).
> >
> >
> > However, converting floats to integers only up to comparison tolerance
> > doesn't seem right since it allows us to convert much larger numbers
> > to the same value MAX_ptype_ (up to (+/2^63 19)). I considered at one
> > point requiring
> >
> >    y -: (3!:0 y) typecast x typecast y
> >
> > but that doesn't seem to make sense either, as it prevents sending a
> > number like (<:2^62), which is close to but not exactly equal to a
> > float, to that float:
> >
> >    (, 4 typecast 8 typecast ]) <:2<.@^62
> > 4611686018427387903 4611686018427387904
> >    (-: 4 typecast 8 typecast ]) <:2<.@^62
> > 0
> >
> > I will make typecast follow the current spec (with relatively liberal
> > float->int conversion). However, it might make sense to also have a
> > strict conversion function for numbers that obeys the additional rule,
> > and only converts an int if there's a float that uniquely represents it.
> >
> > Marshall
> >
> > On Sat, Feb 06, 2016 at 06:59:06PM -0500, Raul Miller wrote:
> >> Hmm... I guess throwing an error works for cases like
> >>    4 typecast (2^63)-432.1
> >>
> >> It would be nice if that particular integer could be converted, but
> >> mostly no one should care.
> >>
> >> Anyways, nice piece of work (though I'll have to admit I did not
> >> exercise it thoroughly - I am only assuming it works well for many
> >> cases).
> >>
> >> Thanks,
> >>
> >> --
> >> Raul
> >>
> >>
> >> On Sat, Feb 6, 2016 at 4:38 PM, Marshall Lochbaum <[email protected]> 
> >> wrote:
> >> > Based on the recent discussions about integers and floats, I decided to
> >> > make a general tool to convert between various J types. It's hosted
> >> > here:
> >> >
> >> > https://github.com/mlochbaum/JScripts/blob/master/Misc/typecast.ijs
> >> >
> >> > The script defines typecast, a verb to convert between types while
> >> > maintaining J equivalence (-:), and numcast, a verb which converts
> >> > between numeric types and guarantees no errors. Both of these operations
> >> > are quite error-prone, so it's good to have a reference with (lightly,
> >> > at the moment) tested verbs to perform them.
> >> >
> >> > My goal is to eventually have a glossary of ways to convert between
> >> > types. I will probably make typecast suggest alternative conversion
> >> > methods if it fails on a particular argument. So I'm curious if anyone
> >> > knows of interesting and useful ways to convert types. The ones I know
> >> > of are:
> >> >
> >> > - Conversions (3!:x) from numbers to bytes
> >> > - (0&~:) to convert numbers to booleans
> >> > - (mod 2^64) conversion from exact to 64-bit integer
> >> > - Representations (5!:x) from nouns (or other parts of speech) to
> >> >   strings or boxes.
> >> > - Gerund representation for a verb
> >> > - The sneaky trick which boxes a part of speech directly
> >> >
> >> > Any others?
> >> >
> >> > Marshall
> >> > ----------------------------------------------------------------------
> >> > For information about J forums see http://www.jsoftware.com/forums.htm
> >> ----------------------------------------------------------------------
> >> For information about J forums see http://www.jsoftware.com/forums.htm
> > ----------------------------------------------------------------------
> > For information about J forums see http://www.jsoftware.com/forums.htm
> ----------------------------------------------------------------------
> For information about J forums see http://www.jsoftware.com/forums.htm
----------------------------------------------------------------------
For information about J forums see http://www.jsoftware.com/forums.htm

Reply via email to