Since PDL_Long * pos is the RETVAL

I don't think so...

Reading http://perldoc.perl.org/perlxs.html#The-NO_INIT-Keyword again, I see I have misread

After checking the generated Basic/Core/Core.c, I'm guessing it is an regression

Also, it seems to me that any initializations using NO_INIT need to go after CODE

Reading http://perldoc.perl.org/perlxs.html#The-NO_INIT-Keyword again, I see I have misread

When I read that link it seems to me that the current PDL code is totally misusing the NO_INIT keyword.
Looks to me that the variable 'pos' is simply a variable that has a value assigned to it - which, one suspects, is (by itself) insufficient to make it an "output" variable at all.

Also, it seems to me that any initializations using NO_INIT need to go after CODE

No, it needs to go *before* CODE. (Perhaps that's what you meant.) With both 2.2210 and 3.03, when I put PDL_Long * pos = NO_INIT; in the CODE section I get an error such as:

Core.xs:673: error: `NO_INIT' undeclared (first use in this function)
[download]

Cheers,
Rob

Update: This has now reached the point where I'm thinking I should probably present the function in its entirety. So here it is as it currently appears in Core.xs (which is autogenerated from Core.xs.PL):

SV *
at_c(x,position)
   pdl*    x
   PDL_Long * pos = NO_INIT
   CODE:
    int npos, ipos;
    double result;

      pdl_make_physvaffine( x );

    pos = pdl_packdims( ST(1), &npos);
    
    if (pos == NULL || npos < x->ndims)
       croak("Invalid position");

    /*  allow additional trailing indices
     *  which must be all zero, i.e. a
     *  [3,1,5] piddle is treated as an [3,1,5,1,1,1,....]
     *  infinite dim piddle
     */
    for (ipos=x->ndims; ipos<npos; ipos++)
      if (pos[ipos] != 0)
         croak("Invalid position");

    result=pdl_at(PDL_REPRP(x), x->datatype, pos, x->dims,
        (PDL_VAFFOK(x) ? x->vafftrans->incs : x->dimincs), PDL_REPROFF
+S(x),
    x->ndims);

    SET_RETVAL_NV(x) ;

    OUTPUT:
     RETVAL
[download]

To reiterate - that will work fine as is with 2.2210, but only works with 3.03 if there is a typemap entry for 'PDL_Long *'. (With 3.03 it doesn't seem to matter what that typemap entry specifies - as long as there is one.)
If I don't amend the typemap, then to make that Core.xs work with both 2.2210 and 3.03 I have to remove the PDL_Long * pos = NO_INIT and begin the CODE section with a PDL_Long * pos; declaration.

When I read that link it seems to me that the current PDL code is totally misusing the NO_INIT keyword.

Yes, this.

No, it needs to go *before* CODE. (Perhaps that's what you meant.)

No, I meant after CODE, without NO_INIT , like you did in your tests, PDL_Long * pos;