[fix] check for invalid axis in concat

[kchanqvq]

Previously this is not checked and may silently give wrong result:

> (concat (list (asarray '(1 2 3))
                     (asarray '(4 5 6)))
               :axis 1)
#<STANDARD-DENSE-ARRAY :ROW-MAJOR 3 DOUBLE-FLOAT
    1.000
    4.000
    5.000
   {10042CB033}>

[digikar99]

Error checks should use (policy-cond:with-expectations (= 0 safety) ...) as in the following:

numericals/src/utils/concat.lisp

Lines 37 to 40 in 0ba5f2c

	(policy-cond:with-expectations (cl:= 0 safety)
	((assertion (let ((out-shape (out-shape-for-concat axis array-likes)))
	(equal out-shape (narray-dimensions out)))
	(out)))

This allows the checks to be skipped if the code is inline-compiled with (safety 0) optimization.

[kchanqvq]

That's good to know! Although, is this necessary in this case? It is a single comparison which is very likely much cheaper than anything else in out-shape-for-concat. Moreover I take it that out-shape-for-concat is not intended to be used in high performance code anyway -- one use is inside policy-cond:with-expectations, the other is for allocating zeros array at runtime. Thoughts?

[digikar99]

That's a good argument :). But then I'd atleast like an assert with non-nil places, so users can recover from the error without unwinding the stack.

[kchanqvq]

Good point about error recovering!

[digikar99]

Actually, this looks like a bad point for error recovery. The recovery through the change of values should happen in the polymorph itself, since those changes need to be visible to the rest of the algorithm. You can revert to the previous simple error and mention this error recovery as a TODO comment in the polymorph body. Later, someone could better the error recovery.

[digikar99]

Thanks!