This is probably beyond me.  I took a look and I couldn't follow all the twists and turns between different templates for do_single_type_concat and single_type_concat in pt-tm-const.[cc|h].  It seems like eventually there is just an octave_value_extract<T> (...), but not clear to me how to promote some of that code up in to pt-eval.cc.  Alternatively, it looks like there are occasional optimizations in the concat routines.  Maybe they should be checking for the 1x1 case and optimizing there.

I wonder if this could be handled (and would make sense to handle) in octave::tree_evaluator::visit_matrix with a special case. If the dimensions of the matrix expression are 1×1, there is no need to do any concatenation or resizing of the arguments. The result would just be the single object in the matrix expression as-is.

If this were only done for the 'else' case (non-numeric types) this should be safe. If it were done for all types it would break the case where '[1:10]' is turned into a normal Matrix, I don't think we want that.

Is this just a special case?  What happens in Matlab if you try to reshape a function handle?  Sample code:

x = @sin;
reshape (x, [1, 1])
reshape (x, [2, 1])
reshape (x, [0, 1])

As Kai pointed out, it might be very simple to overload resize() for function handles to allow 1x1.

Confirmed, Matlab 2017a treats this case like there were no brackets:

>> a = [@() true]

a =

  function_handle with value:

    @()true

>> a = @() true

a =

  function_handle with value:

    @()true

>> a = {@() true}

a =

  cell

    @()true

>> b = [@() true; @() false]
Error using vertcat
Nonscalar arrays of function handles are not allowed; use cell arrays instead.

>> b = {@() true; @() false}

b =

  2×1 cell array

    @()true
    @()false

vs. Octave (hg id

>> a = [@() true]
error: octave_base_value::resize (): wrong type argument 'function handle'
>> a = @() true
a =

@() true

>> a = {@() true}
a =
{
  [1,1] =

@() true

}

>> b = [@() true; @() false]
error: octave_base_value::resize (): wrong type argument 'function handle'
>> b = {@() true; @() false}
b =
{
  [1,1] =

@() true

  [2,1] =

@() false

}

This is because Octave's `octave_fcn_handle`

https://hg.savannah.gnu.org/hgweb/octave/file/ec3d37eeafa2/libinterp/octave-value/ov-fcn-handle.h#l45

does not overwrite `octave_base_value::resize()`

https://hg.savannah.gnu.org/hgweb/octave/file/ec3d37eeafa2/libinterp/octave-value/ov-base.cc#l364

Maybe you can implement this method to just return itself, if the dimension vector is (1,1).

Anyways, I was curious, where did you find/use this border case?