Re: [Bug-apl] First function with axis specification

[Frederick Pitts]

Jürgen,

Not to issue AXIS ERROR's on primitive functions where IBM APL2 does seems to be a fairly large deviation from Gnu-APL's claim to be IBM APL2 compliant. Yes, having the ability to use axis specifications with defined functions is a desirable extension to the language as long as the feature does not interfere with the base language. Breaking syntax checking looks like interference. With defined functions, the person using the function is probably the person who wrote it, so he/she knows the feature is there. With primitive functions the documentation is sketchy and the user depends on the interpreter to point out syntax errors. Trying to use a axis specification with a primitive that does support it is a syntax error (at least, the IBM APL2 documentation says it is).

Section 3.13 'Axis arguments in defined functions' of the apl.html documentation does not give a hint of the non-existence of syntax checking for primitive functions with respect to axis specification support. Maybe a warning to that effect and a list of Gnu-APL primitives supporting axis specifications can be added to the section.

Regards,

Fred

Hi Fred,

this is actually a language extension which has a deeper background.

I define a language extension to be a (different) behaviour of the interpreter when
the standard or the IBM language reference return an error. Sometimes extensions
are introduced or adopted from other interpreters (for example ⍳1 2 3 taken from
Dyalog) or because the language becomes more general.

So what you see is a language extension and the reason is this.

1. GNU APL supports axis arguments for defined functions. This is so that you can
write proper wrapper functions around APL primitives. Since defined functions can
not be overloaded you need a way to write a wrapper that can emulate the optional
axis argument of the wrapped apl primitive. I was always annoyed by not being able
to write such wrappers, in particular for benchmarking purposes.

2. Since the axis argument is optional in all apl primitives that can have one, the axis
is also optional in GNU APL defined functions. You can check for its presence with
⎕NC just like for optional left arguments.

3. Another concept in GNU APL is that there shall be as few differences as possible
between primitive functions (including ⎕-functions) and defined functions. A consequence
of that (which simplifies the parser a lot) is that also all primitive function can have
an axis argument. The default behaviour in both cases is to ignore the axis argument.
That is what you see for ↑[] but also for all other primitives that have no axis variant.
You can change that by changing e.g.

   virtual Token eval_AXB(Value_P A, Value_P X, Value_P B)
      { return eval_AB(A, B); }

to:

   virtual Token eval_AXB(Value_P A, Value_P X, Value_P B)
      { SYNTAX_ERROR; }


in all eval_XXX() functions declared in Function.hh that have an X (for axis) in their name. Since all functions are derived from this class it will change all primitives, but
also all defined functions and operators. I personally would not do it, though.

This language extension has been in GNU APL since almost day one. Therefore I
cannot predict how often it is used (you can also do some dirty tricks with it). For
that reason I believe it is better not to change it.

Best Regards,
Jürgen Sauermann

On 07/04/2017 07:42 PM, Frederick Pitts wrote:

Hello Jürgen,

	The IBM 'APL2 Programming: Language Reference' suggests that
axis specification is not allowed with the first function in that
'Figure 7. Functions and Operators That Allow Axis Specification', page
  45, does not list 'First' by the '↑' symbol, whereas ravel, catenate
and laminate are listed by the ',' symbol. The text under Figure 7 says
an AXIS ERROR occurs if a function or operator is not in the list.

	ISO/IEC 13751:2000(E)' contains 'Table 4: The Form Table',
pages 71-74, which claims to list all syntactic-units for which
evaluation sequences exist.  The first function with axis specification
 does not appear in that table.

	Yet Gnu-APL allows

	↑[ 2 ] 3 3 3 ⍳ 27
1

which is the first scalar in the ravel of the argument and the axis
specification is being ignored.

	Neither IBM or ISO specifications say that the first function
with axis specification is part of the language.  I respectfully submit
that Gnu-APL should report an attempt to use first function with axis
specification as a syntax error.

Regards,

Fred