(< complex) and friends

(< complex) and friends

[Bill Schottstaedt]

guile> (< 1.0+1.0i)
#t
guile> (< 2.0 1.0+1.0i)

Backtrace:
In standard input:
   2: 0* [< 2.0 {1.0+1.0i}]

standard input:2:1: In procedure < in expression (< 2.0 1.0+1.0i):
standard input:2:1: Wrong type: 1.0+1.0i
ABORT: (wrong-type-arg)


(Happens also with > <= >=)

Re: (< complex) and friends

[Neil Jerram]

2008/8/11 Bill Schottstaedt <bil@ccrma.stanford.edu>:
> guile> (< 1.0+1.0i)
> #t
> guile> (< 2.0 1.0+1.0i)
>
> Backtrace:
> In standard input:
>   2: 0* [< 2.0 {1.0+1.0i}]
>
> standard input:2:1: In procedure < in expression (< 2.0 1.0+1.0i):
> standard input:2:1: Wrong type: 1.0+1.0i
> ABORT: (wrong-type-arg)

My guess (without actually looking at the code) is that

- using < etc with complex numbers will always give a wrong type arg error

- the single arg case is being optimized before reaching the check for
a complex number.

What behaviour would you prefer?  (Perhaps to say that < etc. only
compare the real parts of the provided numbers?  Is there any
precedent for that in other Schemes or standards?)

      Neil

Re: (< complex) and friends

[Neil Jerram]

This is mostly for the record, as I'm afraid I don't have any fix yet...

2008/8/12 Neil Jerram <neiljerram@googlemail.com>:
> 2008/8/11 Bill Schottstaedt <bil@ccrma.stanford.edu>:
>> guile> (< 1.0+1.0i)
>> #t

> My guess (without actually looking at the code) is that

> - the single arg case is being optimized before reaching the check for
> a complex number.

My guess was correct.  This comes from < being implemented as a
scm_tc7_rpsubr, which means that there is an underlying C primitive
(scm_less_p) that takes 2 args, and the evaluator deals with calling
this however many times are needed.  When there's just 1 arg, the
evaluator shortcuts the process and returns #t.

Instead of just returning #t, I guess there should be a type predicate
call.  It's not obvious where one would store the type predicate
though, in the subr structure...

As a workaround, you could use:

(let ((libguile-< <))
  (set! < (lambda args
	    (or (null? args)
		(not (null? (cdr args)))
		(real? (car args))
		(error "wrong type arg" (car args)))
	    (apply libguile-< args))))

Would that help?

Regards,
     Neil

Re: (< complex) and friends

[Bill Schottstaedt]

There are actually more problems in that area; for example:

guile> (< 2 1 1.0+3.0i)
#f

which perhaps should be an error also.  My druthers for the
one arg case is to simply disallow it -- why does Guile support
stuff like:

guile> (<)
#t

Re: (< complex) and friends

[Jon Wilson]

Say you want to both filter out most elements from a list and also make 
sure that the remaining elements are sorted.

(apply < (filter some-predicate original-list))

If (<) didn't evaluate to #t, then you'd have to put code in there to 
check the length of the value you get from filter, which makes things 
uglier.  Also, having (<) => #t doesn't  hurt anything.

Bill Schottstaedt wrote:
> There are actually more problems in that area; for example:
> 
> guile> (< 2 1 1.0+3.0i)
> #f
> 
> which perhaps should be an error also.  My druthers for the
> one arg case is to simply disallow it -- why does Guile support
> stuff like:
> 
> guile> (<)
> #t
> 
> 
> 
> 
> 

Re: (< complex) and friends

[Andy Wingo]

On Wed 17 Sep 2008 23:45, "Bill Schottstaedt" <bil@ccrma.Stanford.EDU> writes:

> why does Guile support stuff like:
>
> guile> (<)
> #t

I was going to say it's in the standard, but on second look it only
specifies "(< x1 x2 x3...)". And:

    Welcome to MzScheme v372 [3m], Copyright (c) 2004-2007 PLT Scheme Inc.
    > (<)
    <: expects at least 2 arguments, given 0

(PLT Scheme Inc? Anyway.)

Perhaps better to follow others' lead in this respect.

Andy
-- 
http://wingolog.org/

Re: (< complex) and friends

[Bill Schottstaedt]

A couple similar cases:

guile> (< 2 1 "hi") 
#f

guile> (* 0 "hi")     
0

On the (apply < ...) business, I'd rather get an error than have something
broken go by just because "it is prettier".  That null list is going to trip you...

Re: (< complex) and friends

[Jon Wilson]

Here's a different way to think about it that makes the (<) case, as 
well as perhaps (< 2 1 "hi"), make sense quite naturally.  < is a 
procedure that returns #f if any of its arguments is less than or equal 
to the preceding argument.  Otherwise, it returns #t.

This definition makes the case (< a b) extend quite naturally to (< a b 
c ...), as well as to (< a) and (<).

And anyway, what harm does (<) really do?  Why do you think that it is 
"broken"?

Many functions which expect a list (fold, map) or a number of arguments 
(+, *) accept an empty list or no args as degenerate cases, and give 
sensible results.  When you write a procedure which expects a list or 
uses a rest arg, you generally handle the degenerate cases gracefully 
(instead of raising an error).  Why should < not be the same?

Bill Schottstaedt wrote:
> A couple similar cases:
> 
> guile> (< 2 1 "hi") 
> #f
> 
> guile> (* 0 "hi")     
> 0
> 
> On the (apply < ...) business, I'd rather get an error than have something
> broken go by just because "it is prettier".  That null list is going to trip you...
> 
> 
> 
> 

Re: (< complex) and friends

[Neil Jerram]

2008/9/19 Jon Wilson <jsw@wilsonjc.us>:
> Here's a different way to think about it that makes the (<) case, as well as
> perhaps (< 2 1 "hi"), make sense quite naturally. [...]

Thanks everyone for discussing these cases.  It seems to me that
although it may seem curious, Guile's current behavior doesn't
actually cause any significant trouble, and so doesn't need fixing.

But I may well be wrong about the practical impact.  If there are
realistic scenarios where the current behaviour turns out to be
dangerous, or to matter in some way, please describe those, and I'll
look again.

Regards,
       Neil

Re: (< complex) and friends

[Bill Schottstaedt]

map does not accept no list:

guile> (map (lambda (a) a))

Backtrace:
In standard input:
   1: 0* [map #<procedure #f (a)>]

standard input:1:1: In procedure map in expression (map (lambda # a)):
standard input:1:1: Wrong number of arguments to #<primitive-generic map>
ABORT: (wrong-number-of-args)


+ and * have a "natural" identity (0 and 1), so it is not silly that (+ a) is
(+ a 0) and (+) is (+ 0 0), but #t does not strike me as a natural identity
for < -- (< 1 1) is #f.   But I hate these kinds of discussions, so I defer...

Re: (< complex) and friends

[Neil Jerram]

2008/9/19 Bill Schottstaedt <bil@ccrma.stanford.edu>:
> map does not accept no list:
>
> guile> (map (lambda (a) a))
>
> Backtrace:
> In standard input:
>   1: 0* [map #<procedure #f (a)>]
>
> standard input:1:1: In procedure map in expression (map (lambda # a)):
> standard input:1:1: Wrong number of arguments to #<primitive-generic map>
> ABORT: (wrong-number-of-args)

Right.  The number of lists given to map should be the same as the
number of parameters the lambda takes.

Were you saying that this is a bug, or is this just to set up an
analogy with the following?

> + and * have a "natural" identity (0 and 1), so it is not silly that (+ a) is
> (+ a 0) and (+) is (+ 0 0), but #t does not strike me as a natural identity
> for < -- (< 1 1) is #f.   But I hate these kinds of discussions, so I defer...

The value of (<) seems natural to me... but I admit that I can't
explain really clearly why that is!

And then there's (and) and (or) !

Nevertheless, I still don't see any significant impact anywhere here.
Please explain further if you do.

Regards,
     Neil

Re: (< complex) and friends

[Jon Wilson]

Bill Schottstaedt wrote:
> map does not accept no list:
But it does accept the empty list, which is what I intended to say, but 
didn't make clear.

> 
> guile> (map (lambda (a) a))
> 
> Backtrace:
> In standard input:
>    1: 0* [map #<procedure #f (a)>]
> 
> standard input:1:1: In procedure map in expression (map (lambda # a)):
> standard input:1:1: Wrong number of arguments to #<primitive-generic map>
> ABORT: (wrong-number-of-args)
> 
> 
> + and * have a "natural" identity (0 and 1), so it is not silly that (+ a) is
> (+ a 0) and (+) is (+ 0 0), but #t does not strike me as a natural identity
> for < -- (< 1 1) is #f.   But I hate these kinds of discussions, so I defer...

A natural identity for a predicate like this is either #t or #f.  The 
choice is really left up to the implementor, but should probably match 
other implementations (at least those that have some sort of identity) 
for the sake of consistency, and should most especially match the 
language in the specification or the manual (which should match each 
other where they overlap, of course!).  I'm not too concerned with which 
way the degenerate case goes (except in that it matches language and 
convention), but I do feel that the degenerate case should not be an error.
Regards,
Jon