Greetings!
Raymond Toy <toy.raymond@gmail.com> writes:
> On Mon, Feb 26, 2024 at 5:28 AM Camm Maguire <camm@maguirefamily.org> wrote:
>
> Greetings, and thanks to all for the very helpful feedback!
>
> I'd also like to note in passing that the excellent SBCL out of the box
> triggers an error, which we separately refer to as an 'exception trap',
> when NaN is passed to the comparison functions, e.g. '=. And low and
> behold, gcl does the same when 'trapping' is turned on, but gcl turns
> trapping off 'out of the box'. NaN does not trigger an exception nor
>
> I hope you will consider changing the default to trap on invalid
> operations so that instead of returning NaN, an error is signaled. I
> certainly hate this behaviour, because after you do a long set of
> numerical computations, all the results are NaN. I've just wasted a
> whole bunch of time. This is super-annying in _javascript_ where you
> don't have an option of disabling traps. Now you get to go through
> all the code and either put prints or checks for NaN to see what
> caused it. What a colossal waste of time when the computer could have
> told me.
>
This is good to hear, as I was of the opposite impression that most
users wanted NaNs to propagate freely.
Maybe you should take a poll or something. It could be that I'm just a terrible numerical programmer. But in my experience, it's fairly easy to produce an infinity somewhere which then gets subtracted from infinity that then becomes NaN which then poisons all further computations.
> error when passed to '+,'*,'cos etc. but returns NaN. This is the
> current state of play. What follows is a discussion of what it means
> and the best way to fit NaN into the common-lisp type lattice so it does
> not get in the way of the compiler.
>
> In summary, NaN seems indistinguishable to me from an ignorable error
> indicator at the hardware level. It is not conceptually a 'valid input'
> to any of these functions, but is a clever design to propagate the error
> properly when ignored. Conceptually, it seems no different from
> 'ignore-errors in common-lisp. So really the spec is not all that far
> away from IEEE.
>
> I suppose that that is true, but I don't think any Lisp programmer
> would wrap their entire programs in a ignore-errors form. That's
> usually one done on a very limited scope. Having traps off to get
> NaN, is basically the same idea.
>
Wrapping code in ignore-errors is totally unnecessary. I can insert a
specific handler within GCL's type-error mechanism to skip the error
when testing NaN depending on a global variable setting. One question
is whether such a setting governing skipping a lisp error should be the
same as or distinct from the setting enabling hardware traps.
I was just saying having traps off is like if someone wrapped their entire lisp program in ignore-errors. But maybe that's me; I like traps on to find numerical mistakes when they occur, not at the end where everything is NaN.
> As for common-lisp types, the obvious path is what GCL and others
> currently implement, double precision NaN is of type 'long-float and
> hence 'number. This actually means that lisp will *stray* from IEEE and
>
> I think one thing that confuses me, is that for gcl, single-float =
> double-float = double precision. short-float is single precision. I
> think long-float is double-precision. I think all other lisps have
> single-float = single precision, double-float = double precision.
> Short-float is usually the same as single-float. Long-float is
> usually double-float, if they don't have a separate long-float (like
> clisp).
>
> So whenever you say long-float, I sometimes think of actual
> long-floats like clisp, but you're actually talking about
> double-precision floats.
Thanks, this needs to be cleaned up. I think single-float should be
short-float (e.g. 4 bytes).
Gcl's choice is certainly allowed by the CLHS. It's just confusing that the names don't match the IEEE precision names and most other lisps. Fortunately, I think all other lisps treat short-float as single-precision; I don't think anyone supports a short-float that is not IEEE single precision.