The original regexp took 'unsigned short *' and captured only the
'short'. This changed allows it to capture 'unsigned short'.

Since fitgauss1d[r] has GenericTypes=>[D], it is safe to assume
that everything will be Double, so using fabs is preferred.

The switch statement was written long before those datatypes were
in PDL, and the compiler was nicely complaining about their absense.

Perl's 'warn' function had been undef'd because of problems with the gsl code's
'warn' variable. But PDL::PP inserts code into XS files that calls
Perl's 'warn' function.  To avoid this I removed the #undef warn and
changed the gsl variable name from warn to gslwarn.

Some function prototypes that get generated into SlatecProtos.h
have return types float or double, but most were just missing, and
the compiler assumed them to be int. The function defs all do
return int, so I just made it explicit.

A few functions (SDstart ... SDend) did not have function prototypes
in hdf.h, this other header file was necessary.

…tructor.t

If a user mistakenly passed a perl scalar argument to qsort and
its ilk, the SV would be mistaken for a piddle and passed to
the underlying generic_qsort etc routines. Not surprisingly,
those routine's array index tricks would cause a segfault. This
commit causes qsort, qsorti, qsortvec, and qsortveci to barf
if an argument is passed with incorrect dimensions.

Some care is taken to allow trivial sorts.

2 tests for each routine have been added, and some existing
tests labeled.

Hopefully this will compile on Strawberry. Not sure why the original
worked OK on my Mac w/ clang.

The changes in <#188> broke the
build for all cases of `BADVAL_USENAN` xor `BADVAL_PER_PDL`.

This was because the checking for badvalues when copying in
`pdl_kludge_copy_*` did not take into account these variations in
accessing and setting the badvalue.

For example, a simple comparison using `==` is not valid when NaN
values are being compared. Here we need to use `!finite(...)` to make
the comparison if the type can use NaN values.

The example in the docs was using the matrix v directly instead of
the transpose.  Doesn't matter for 2x2 matrices, but does for larger.
Hopefully this clears things up.

I also added some tests for svd: square matrices, and rectangular
matrices where m>n and m<n.

Some comments taken from Nash (1990), some are my own to aid in
understanding this algorithm.

The code to determine whether the output size agreed with the
format string size was only activated when the format string was
input as the first argument, not in the options hash. This change
activates it for both cases.

Added some tests as well.

The list is generated by Types.pm.PL, but gets put into pdl.h and
pdlsimple.h. This was motivated by code in Bad.c that assigns to
a variable of type pdl_datatypes the value -1 (in some error condition
that is unlikely to be reached). This jives with the ANYVAL_FROM_CTYPE
macro in pdl.h.  Thanks to Chris Marshall and David Mertens for the
pointers.  See pdl-devel 2016-05-05 discussion.

This was motivated because we should have been using labs or llabs,
the correct function to call was system-dependent. This avoides all
that by replacing the abs() call with multiplication by the
variable's sign.

fabs is fine for doubles, but not for anything else: this inserts
the correct function (abs, labs, llabs, etc) based on the datatype
used in the complex division. I hope.

This silences a compiler warning about a printf wanting double ("%f")
but we passed an Anyval.

Most prototypes get inserted into SlatecProtos.h by defslatec.
But polfit_ and dpolft_ do not, so they need to be added manually.