(Previous discussion continued)
Re: SRFI-47 -- types and declarations. Aubrey Jaffer 04 Dec 2003 15:50 UTC

Re: SRFI-47 -- types and declarations. Aubrey Jaffer 04 Dec 2003 15:50 UTC

 | Date: Tue, 25 Nov 2003 16:49:51 -0500
 | From: Taylor Campbell <xxxxxx@evdev.ath.cx>
 | On Thursday, Nov 20, 2003, at 18:56 US/Eastern, Aubrey Jaffer wrote:
 | > I will remove the read syntax from SRFI-47, which removes many of
 | > the objections.
 | Why not provide SRFI 10 reader constructors?  This kind of thing is
 | _exactly_ what SRFI 10 is for!

Array syntax can be the subject of its own SRFI.

 | Also, I don't think my other questions, those not regarding reader
 | syntax, were addressed: will there be a follow-up complete array
 | library to deprecate SRFI 43 and some of SRFI 13?

SRFI-43 contains many procedures which do not generalize to rank 2 and
higher arrays.  So an array-based SRFI would not replace SRFI-43.
SLIB's array-for-each module defines:

 - Procedure: array-map!   array0 proc array1 ...
 - Function:  array-map prototype proc array1 array2 ...
 - Function:  array-for-each proc array0 ...
 - Function:  array-indexes       array
 - Procedure: array-index-map!    array proc
 - Procedure: array-copy!  source destination

The argument order to array-copy! seems inconsistent with the others.

Sequence order is an essential idea in SRFI-13 (and for many functions
in SRFI-43).  Defining procedures which make sense only for rank-1
arrays in a multidimensional array package is poor practice.

One might consider a rank-2 array of characters to be text on a page.
I am unaware of Unicode or other character standards covering such
constructs.  But with most Schemes still limited to single byte
characters, we need to resolve the plethora of difficult string issues
before going where none have gone before.

 | and ARRAY=? is less general than it could be: why isn't it more
 | like SRFI 1's LIST= or SRFI 43's VECTOR= ?

ARRAY=? was added to facilitate incorporation of arrays into EQUAL?  A
more complete solution would be to redefine EQUAL? instead.

 | I have also come up with another question: why is there no ARRAY,
 | analogous to VECTOR as MAKE-ARRAY is analogous to MAKE-VECTOR?

VECTOR and STRING return arrays.

The VECTOR, LIST, and STRING functions allow one to construct those
objects on the basis of arguments alone -- not requiring explicit
construction of intermediate lists.  This property was important to
some of the rrrs-authors; they did not want argument vectors to
necessarily be lists.

To do the same for arrays would require the dimensions of the array be
passed as arguments to the ARRAY function.  Passing them without using
a list or vector would have them as several arguments.  But there are
a variable number of them, so the rank must also be an argument.  This
results in a form reminiscent of run-length coding:

 - Function: array prototype j k0 ... k{j-1} obj0 ...
     J should be an exact nonnegative integer specifying the rank of
     the array.  K0 ... K{J-1} should be exact nonnegative integers
     specifying the dimensions of the array.  OBJ0 ... should be
     (* K0 ... K{J-1}) elements to be stored in the new array in
     row-major order.

     `Array' returns a newly allocated array of type PROTOTYPE whose
     elements contain the given OBJ0 ... arguments.

I can think of no Scheme precedent for having a variable number of
(dimension) arguments followed by a variable number of other
arguments; but it is well-defined.  The lack of list or other
structure to organize the obj0 .. arguments would probably make the
ARRAY function difficult to use for all but the smallest arrays, but
it could be useful for those.  The prototype functions already
generate one and zero element arrays.

SCM has a function list->uniform-array which organizes the elements
using lists (which should be renamed list->array):

 - Function: list->array rank prot lst
     Returns a uniform array of the type indicated by prototype PROT
     with elements the same as those of LST.  Elements must be of the
     appropriate type, no coercions are done.

     In, for example, the case of a rank-2 array, LST must be a list of
     lists, all of the same length.  The length of LST will be the
     first dimension of the result array, and the length of each
     element the second dimension.

     If RANK is zero, LST, which need not be a list, is the single
     element of the returned array.

Which of these approaches makes the most sense?