Automatically detect and test optional dependencies

[robertwb]

I'd deferred actually fixing the broken doctests to #13884, #13885, and #2120 so we can start using this (and not have everything optional broken).

---

Superseded / fixed by #18558, #20182.

[jdemeyer]

Checking just for the existence of gcc is insufficient, you would have a lot of false positives with missing/broken assemblers/linkers... You actually need to compile something, preferably a shared library.

[jdemeyer]

For the internet test: could you use www.sagemath.org instead of 173.194.33.3. In any case, an IP address is bad because

1. you don't test DNS, which is required.
2. it can change.

[jdemeyer]

Also please add a check for g++ and gfortran.

[jdemeyer]

Instead of relying on which, simply execute the command and catch OSError.

[jdemeyer]

Command lines for gcc, g++, gfortran compiling a shared library from /dev/null:

gcc -shared -fpic -x c /dev/null -o /dev/null
g++ -shared -fpic -x c++ /dev/null -o /dev/null
gfortran -cpp -ffree-form -shared -fpic -x f95 /dev/null -o /dev/null

I have no idea how to do any of this with other compilers such as clang.

[jdemeyer]

I think it would be good to implement either caching or lazy evaluation (only check when actually encountering # optional - foo). Otherwise, somebody with a slow internet connection could see a substantial time added to every doctest run.

[jdemeyer]

For programs like magma, does Sage require a specific version? If so, you should check for that version.

[robertwb]

I would be surprised if many people that have gcc have a broken gcc, but it couldn't hurt to check (and g++, etc.) Same with internet--it's rare for people to be connected but not be able to resolve DNS. (If this is lazy, latency less of an issue.)

Good points. I'll refactor this to make things lazy and otherwise improve the patch.

[robertwb]

Apply only 13540-optional-tests.2.patch

[robertwb]

Apply 13540-optional-tests.2.patch to sage-local repo.

[vbraun]

The patch is great but a great number of the # optional - internet doctests fail because they haven't been doctested in a while... Any opinion on how to proceeed?

[leif]

IMHO we shouldn't hardcode the names of the executables (gcc, g++, gfortran etc.), but respect for example CC, CXX, FC/F77/F95.

(Also wonder whether "# optional - gcc" really means GCC, or some C compiler, perhaps at least capable of compiling Cython-generated code.)

[leif]

Replying to leif:

IMHO we shouldn't hardcode the names of the executables (gcc, g++, gfortran etc.), but respect for example CC, CXX, FC/F77/F95.

As an alternative, we might put their settings "during build" (whatever that means) into dummy/wrapper scripts in $SAGE_LOCAL/bin/. (I don't really have an opion on that, but tend to be against such. It might be convenient to "ordinary" users, or perhaps for automated testing, but annoying or confusing to developers.)

[robertwb]

For all intents and purposes, at the moment, gcc really means gcc. And building a Cython extension with a different compiler than Python was compiled with (and the relevant libraries, for C++ in particular) gets really messy.

In any case, the point of this ticket is to introduce a framework that lets us test optional doctests automatically, as they the current setup means they're never tested and so quickly broken. We could do something more sophisticated with CC, etc. but gcc, at the very least, provides a common enough dependency to actually exercise this code for most (probably all) developers (and it's not near as brittle as others).

[leif]

Replying to robertwb:

For all intents and purposes, at the moment, gcc really means gcc. And building a Cython extension with a different compiler than Python was compiled with (and the relevant libraries, for C++ in particular) gets really messy.

We might then even use Python's / distutils' settings for testing tool availability (although there's no variable for the Fortran compiler AFAIK). Not sure what to do with bdists (in case they don't contain a pre-built GCC as well).

In any case, the point of this ticket is to introduce a framework that lets us test optional doctests automatically, as they the current setup means they're never tested and so quickly broken. We could do something more sophisticated with CC, etc. but gcc, at the very least, provides a common enough dependency to actually exercise this code for most (probably all) developers (and it's not near as brittle as others).

Yes. Nevertheless, I don't think I'm the only one that almost always sets CC, CXX etc., in order to use different versions of GCC.

Support for other compilers (e.g. clang), probably also flags to pass, is certainly out of the scope of this ticket. (It would be convenient to have something like config.status, i.e., a mechanism for storing and restoring the build environment, and/or also a ".sagerc" local / specific to the Sage installation, one overriding the other...)

[jdemeyer]

How is this supposed to interact with the --optional and --only-optional command line options? I have a hard time understanding the new code.

[jdemeyer]

This would obviously need to be rebased to #12415.

[ncohen]

(just to follow what's happening. This ticket would be a nice change indeed O_o)

[leif]

Magma doctest failures now get fixed at #16322.

[ncohen]

Hello everybody,

A branch that does that has been created at #18558 (needs review)

Nathann

[klee]

This is now "duplicate/wontfix" as the replacement #20182 is now merged.

[slelievre]

Is this ticket solved by the other tickets mentioned in the last comments? Does any comment in the discussion beg for a follow-up ticket?

[klee]

Replying to slelievre:

Is this ticket solved by the other tickets mentioned in the last comments?

Yes.

Does any comment in the discussion beg for a follow-up ticket?

I don't think so. Most of the comments in this discussion are 4 years old, and so perhaps invalid now.

[leif]

4-year old => invalid

ROFL.

[embray]

Determined to be invalid/duplicate/wontfix (closing as "wontfix" as a catch-all resolution).