Fixes for extended and quad precision checking. Add CI cheks with fpm

[jalvesz]

After looking at #819 I realized that there were several issues:

1. in the _specialfunctions_gamma files, there were this kind of fypp macros:

#:set WITH_QP = False
#:set WITH_XDP = False
#:include "common.fypp"

Which implied that the different kind definitions were being redefined. In parallel, this means that some files were correctly preprocessed and others not, depending on when the problematic files were being preprocessed.
2. many tests are not using fypp "properly" so I refactored those that I catched after the modifications
3. Seems like there is no CI checking xdp, so I propose to use the fpm CI to check xdp and qp with gfortran.
4. I don't know if test-drive is not built with xdp and qp, but the call check( ... ) using the tolerance argument thr had to be modified to avoid passing xdp or qp arguments.

I think this is ready for consideration @perazz @jvdp1 @nsh-envs

[jvdp1]

Thank you @jalvesz .
I only have a minor commnet, regarding the robustness of the definition of the variable RC_KINDS_TYPES.

[perazz]

I agree with @jvdp1: great work @jalvesz, thank you for finding this out! I don't want to have you rewrite your PR but here is some thought.
Elsewhere because LAPACK also does not support xdp, I've used this approach:

stdlib/src/stdlib_linalg_determinant.fypp

Lines 17 to 18 in 42182b0

	#:for rk,rt in RC_KINDS_TYPES
	#:if rk!="xdp"

So here, it could be something like this?

#:if rk not in ["xdp","qp"]

Also: the special functions just want to be for the "core" types (32- and 64-bit).
So I'm not suggesting to do it in this PR, but would there be a way we could unify this once and forall? For example, by having some CORE_REAL_KINDS defined in common.fypp as opposed to all REAL_KINDS?

stdlib/include/common.fypp

Lines 21 to 28 in 42182b0

	#! Real kinds to be considered during templating
	#:set REAL_KINDS = ["sp", "dp"]
	#:if WITH_XDP
	#:set REAL_KINDS = REAL_KINDS + ["xdp"]
	#:endif
	#:if WITH_QP
	#:set REAL_KINDS = REAL_KINDS + ["qp"]
	#:endif

[awvwgk]

For test-drive to accept extended double and quadruple precision the C preprocessor macros WITH_XDP=1 and WITH_QP=1 need to be set when compiling test-drive.

[jvdp1]

Thank ypu @jalvesz .

[jvdp1]

LGTM.
Did you find a way to compile test-drive with/out QP and XDP if supported or not?

[jalvesz]

@jvdp1 I didn't have time last week to look at that. I guess this should be done by propagating the fypp macros backwards in the dependency hierarchy but right now I do not see clearly how to do that without touching the build of test-drive itself or transferring the python preprocessing script to fpm.

[jvdp1]

I guess it should be also in the CMake files. I will have a look.
Edit: the flags are already used for test-drive with CMake

[jalvesz]

Should we then have one of the cmake builds test xdp and qp instead of the fpm build?

[jvdp1]

Yes, good idea!
But the issue might still be there if someone try the tests with fpm, right? Or do I miss something?

[jalvesz]

Yes, it would still be there. The ideas on top of my head are: add the python preprocessing to test-drive as well or add it to fpm (it might be the most robust solution but I'm guessing not the easiest)

Edit: I looked at test-drive sources, so since it uses only C preprocessing and not fypp this should be actually easier. I'll check how to propagate the flags from the stdlib fpm build.

[jalvesz]

Passing the corresponding flags in the build fpm test --profile release --flag '-DWITH_XDP -DWITH_QP' are indeed enough to backpropagate the option.

[jvdp1]

Thank you @jalvesz . Pending a minor comment, I think this PR can be merged.

[jvdp1]

@jalvesz It sounds it is ready to be merged. We will wait a few more days, and then merge it. Thank you.