Define library dependencies in .pxd files

[jdemeyer]

Rather than defining libraries in module_list.py or in .pyx files, we should add them to the .pxd files where the declarations are.

In this branch, we do this as proof-of-concept for pari and gmp. Other tickets can be opened later for other libraries. We still leave the libraries explicitly in module_list.py when needed as dependency of another library, for example mpfr depends on gmp.

One annoying part is that the order of libraries is not defined, so we need to manually re-order them using one global library order. Since the order of libraries currently specified in src/module_list.py is far from a DAG, many extensions will be compiled with a different library order than before.

To clean up, we also remove the explicit mention of stdc++ for c++ code (which is automatically added anyway).

---

These are all modules where the list of libraries was actually changed:

These were probably underlinked before:

• sage/rings/polynomial/pbori: added gmp
• sage/rings/real_arb (optional): added gmp

These use inline functions or macros of libraries but don't use exported functions. So they don't need to link to the library, but it's added anyway:

• sage/rings/rational: added pari
• sage/rings/padics/common_conversion: added pari

Modules using the Sage/PARI interface but not directly using PARI will now link against pari. Strictly speaking, this is a bug but it's hard to avoid:

• sage/rings/padics/padic_capped_absolute_element: added pari
• sage/rings/padics/padic_capped_relative_element: added pari
• sage/rings/padics/padic_fixed_mod_element: added pari
• sage/rings/real_double: added pari

This branch adds gmp for all modules where bitsets are used. Some bitset functions require GMP, others do not. These modules were only using the ones not implemented using GMP:

• sage/combinat/debruijn_sequence: added gmp
• sage/combinat/words/word_char: added gmp
• sage/graphs/base/static_sparse_backend: added gmp
• sage/graphs/weakly_chordal: added gmp

Unused libraries are removed:

• sage/graphs/graph_generators_pyx: removed gmp
• sage/groups/semimonomial_transformations/semimonomial_transformation: removed gmp
• sage/modular/modsym/p1list: removed gmp
• sage/modules/vector_mod2_dense: removed gmp

---

We need 1 patch to upstream Cython to fix dependency handling:

• ​https://github.com/cython/cython/pull/381 (accepted)

[jdemeyer]

New commits:

​2ed346b	Remove unneeded inclusions of cdefs.pxi
​c1eb2b8	Add Cython patches to improve handling of dependencies
​49029fd	Define some library dependencies in .pxd/.pxi files

[git]

Branch pushed to git repo; I updated commit sha1. This was a forced push. New commits:

​57e9278	More removals of cdefs.pxi
​370106d	Even more removals of cdefs
​ca87ba3	Add Cython patches to improve handling of dependencies
​a1ad51c	Define some library dependencies in .pxd/.pxi files

[git]

Branch pushed to git repo; I updated commit sha1. New commits:

​941fa74

Merge tag '6.8.beta0' into t/18450/define_library_dependencies_in__pxd_files

[jdemeyer]

To potential reviewers: if the upstream status of the second Cython patch is what prevents you from reviewing this ticket, I can add a quick work-around for that. Just let me know.

[jpflori]

Please do!

I had a look at the upstream ticket and did not get any strong feeling yet.

[git]

Branch pushed to git repo; I updated commit sha1. New commits:

​ae16699	Merge tag '6.8.beta2' into HEAD
​8b83481	Move pari/decl.pxi to pari/paridecl.pxd

[gouezel]

With the patch, mpfr is not picked when compiling partitions_c (on cygwin) Part of the log:

build/temp.cygwin-1.7.35-x86_64-2.7/sage/combinat/partitions_c.o: dans la fonction « compute_current_precision »:
/home/Sebastien/sage13/sage/src/sage/combinat/partitions_c.cc:448: référence indéfinie vers « __imp_mpfr_init2 »
/home/Sebastien/sage13/sage/src/sage/combinat/partitions_c.cc:448:(.text+0x16): relocalisation tronquée pour concorder avec la taille: R_X86_64_PC32 vers le symbole indéfini __imp_mpfr_init2
/home/Sebastien/sage13/sage/src/sage/combinat/partitions_c.cc:452: référence indéfinie vers « __imp_mpfr_set_d »

Full log on request.

[jdemeyer]

Right, it's partitions_c.cc which needs those libraries, not the Cython code...

[git]

Branch pushed to git repo; I updated commit sha1. New commits:

​1c72dd6

partitions_c.cc needs gmp and mpfr libraries

[git]

Branch pushed to git repo; I updated commit sha1. New commits:

​87be62b

Correct Jonathan Bober's name

[jpflori]

Coule you please open a follow up ticket for the Pari linking issue?

[jdemeyer]

You mean the fact that pari gets added as library when it's not needed? That's now #18583.

[gouezel]

With the patch, linking problem of 'libecm' with 'gmp' Part of log:

gcc -shared -Wl,--enable-auto-image-base -L/home/Sebastien/sage13/sage/local/lib build/temp.cygwin-1.7.35-x86_64-2.7/build/cythonized/sage/libs/libecm.o -L/home/Sebastien/sage13/sage/local/lib -L/home/Sebastien/sage13/sage/local/lib/python2.7/config -L/home/Sebastien/sage13/sage/local/lib -lcsage -lgmp -lecm -lpython2.7 -o build/lib.cygwin-1.7.35-x86_64-2.7/sage/libs/libecm.dll
/home/Sebastien/sage13/sage/local/lib/libecm.a(libecm_la-factor.o): dans la fonction « ecm_init »:
/home/Sebastien/sage13/sage/local/var/tmp/sage/build/ecm-6.4.4/src/factor.c:32: référence indéfinie vers « __imp___gmpz_init_set_ui »

[jpflori]

There is a lgmp but before lecm and the linker is picky.

[git]

Branch pushed to git repo; I updated commit sha1. New commits:

​0be796e

Add ecm library before gmp

[jdemeyer]

Right, I forgot to add ecm.

[gouezel]

Good to go.

[jdemeyer]

Thanks for the review!

[leif]

Just out of curiosity: Are there follow-ups moving further things into *.pxd?

(More precisely, into # distutils: ... declarations.)