Upgrade to GAP 4.10

[nthiery]

GAP 4.9 comes with a completely rewritten build system that will simplify our packaging. In particular, libGAP no longer needs to be a separate package.

What the branch does:

• Remove the libgap spkg

• Update the gap spkg to the new build system and build and install libgap

• Replace gap.shi.patch by a plain gap startup script for Sage

Rationale: GAP used to provide a startup shell script. The GAP devs are in the process of getting rid of it and provide a very minimal one. They recommend to just write our own rather than patching it.

• Update a few doctests w.r.t. changes of output of some GAP functions

• Possibly controversial: The new libgap currently *does not come* with symbol rewriting (foobar -> libGAP_foobar). This avoids messing around with GAP's sources; in particular opening the door for using a stock GAP from the OS distribution. However there always is a risk of name conflicts. And indeed, GAP's enums T_INT, T_FLOAT, ... conflict with Python's constants defined in structmember.h. This is hopefully not actually a problem in practice due to the way how Cython orders includes.

Something similar was started by Volker at #19915.

• Removes configure.patch: it was patching configure for better GMP detection under Cygwin (#13954). This should not be needed anymore with the new build system and use of --with-gmp. If it is, upstream asked for it to be reported and they will fix it.

• Revert #19726 (not needed anymore)

Status: currently broken - crashes deep inside GAP error handling system after few simple commands.

Basic tests on libgap:

sage: libgap.eval("GAPInfo.Version")
sage: libgap.DihedralGroup(10).CharacterTable()
CharacterTable( <pc group of size 10 with 2 generators> )
sage: libgap.Group(libgap.eval("[(1,2,3),(1,2)]")).Size()
6

Running most relevant tests:

sage -tp 8 sage/groups sage/libs/gap

Current status: lots of errors

Testing packages with dynamic loading (e.g. IO):

Install IO:

cd $SAGE/local/gap/latest/pkg
wget http://www.gap-system.org/pub/gap/gap4/tar.gz/packages/io-4.4.6.tar.gz
tar xvf /tmp/io-4.4.6.tar.gz
mv io-4.4.6 io
cd io
./configure
make

Test it locally:

cd ../..
./gap -l .
gap> LoadPackage("IO");
true

This does not yet work:

sage: libgap.LoadPackage("IO")
ValueError: libGAP: Error, module '/opt/sage-git/local/gap/latest/pkg/io/bin/x86_64-pc-linux-gnu-gcc-default64/io.so' not found

This should be fixed once GAP's gap binary is built on top of libgap. See: ​https://github.com/markuspf/gap/issues/1.

Note:

• Max Horn reviewed the list of GAP symbols we use in Sage; some have already changed in 4.9. See this ​pad for notes.

Tarball: ​https://www.gap-system.org/pub/gap/gap-4.10/tar.bz2/gap-4.10.0.tar.bz2

[git]

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

​1c645e0	#22626: remove GAP's symbol prefixing in libgap: libGAP_Foo -> Foo
​1ecc67b	#22626: doctest update w.r.t. minor changes of output in GAP
​e8ebbca	#22626: GMP detection patch for cygwin should not be needed anymore
​fd65b06	#22626: Remove libgap spkg
​9511733	#22626: replace patch for GAP's startup script template in favor of a custom script
​550625e	#22626: remove GAP's symbol prefixing in libgap: libGAP_Foo -> Foo, and workaround GAP <-> Python symbol conflict
​a9c859c	#22626: updated gap spkg w.r.t. GAP's devel version and its new build system; also include compilation and installation of libgap
​3011ac0	Merge branch 'develop' into t/22626/upgrade_to_gap_4_9

[git]

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

​234c54b

#22626: revert #19726 as it won't be needed for gap 4.9

[git]

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

​b301e4c	#22626: ReST fix
​431845f	#22626: better work around for the GAP-Python name clash on T_INT, ...

[nthiery]

New commits:

​b301e4c	#22626: ReST fix
​431845f	#22626: better work around for the GAP-Python name clash on T_INT, ...

[git]

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

​7c04025

22626: document the work around for the GAP-Python name clash on T_INT, ...

[fbissey]

So if I understand the mechanics involved, at the moment you have to configure and install gap. Then clean and configure and install for libgap. So it is still behaving like "two packages" with the same source code. Do we have an idea on when we'll have just have gap linking on libgap which would also solve this particular problem?

[nthiery]

Salut François,

Replying to fbissey:

So if I understand the mechanics involved, at the moment you have to configure and install gap. Then clean and configure and install for libgap. So it is still behaving like "two packages" with the same source code.

Exactly.

Do we have an idea on when we'll have just have gap linking on libgap which would also solve this particular problem?

The GAP developers are well aware of that and planning to implement it. They have not yet done yet just to be more incremental; their new build system is a big PR already :-) I would assume that this will be done before GAP 4.9 which they are planning for a couple months from now.

See: ​https://github.com/markuspf/gap/issues/2

Btw: feel free to comment / expand on the list there. I'll advertise this issue for additional feedback on the sage-packaging mailing list.

[jdemeyer]

Replying to fbissey:

So if I understand the mechanics involved, at the moment you have to configure and install gap. Then clean and configure and install for libgap. So it is still behaving like "two packages" with the same source code.

For this reason, I wonder why you didn't keep the separation GAP + libGAP in Sage too. If it effectively behaves as two packages, it seems more natural to keep it as two packages in Sage too. Imagine for example that we need to patch GAP but not libGAP or conversely, that would be harder with the current approach. That being said, this is mostly bikeshedding. So, if the current setup works well, there might be no reason to change it.

[dimpase]

Replying to jdemeyer:

Replying to fbissey:

So if I understand the mechanics involved, at the moment you have to configure and install gap. Then clean and configure and install for libgap. So it is still behaving like "two packages" with the same source code.

For this reason, I wonder why you didn't keep the separation GAP + libGAP in Sage too. If it effectively behaves as two packages, it seems more natural to keep it as two packages in Sage too. Imagine for example that we need to patch GAP but not libGAP or conversely, that would be harder with the current approach. That being said, this is mostly bikeshedding. So, if the current setup works well, there might be no reason to change it.

Two completely separate packages would also mean that gap_packages should come in two different flavours, etc. I'd rather keep it simple and do not multiply these instances (but rather hope that libGAP will support all of GAP packages soon).

[jdemeyer]

Replying to dimpase:

Two completely separate packages would also mean that gap_packages should come in two different flavours, etc.

So, you are saying that we need to install gap_packages twice too? Once for GAP and once for libGAP?

[nthiery]

Replying to jdemeyer:

For this reason, I wonder why you didn't keep the separation GAP + libGAP in Sage too.

I wanted to experiment with how far we were from having a proper single package, so as to give early feedback to the gap developers on any sticking points. And indeed, linking gap to libgap is basically the only sticking point. Since this point will be most likely resolved by the time this ticket gets merged in, we might as well shoot directly for the "right thing".

[dimpase]

Replying to jdemeyer:

Replying to dimpase:

Two completely separate packages would also mean that gap_packages should come in two different flavours, etc.

So, you are saying that we need to install gap_packages twice too? Once for GAP and once for libGAP?

no, not at all---I mean to say that there exist GAP packages (not part of Sage ATM) that break libGAP now. As some of these packages are very useful, it's important to have this fixed (but not at expense of having 2 separate gap_packages).

[nthiery]

Replying to dimpase:

As some of these packages are very useful, it's important to have this fixed

YES. In fact that is my main motivation to work on the libgap integration: I badly want to use Semigroups that requires IO! We went through the details with Markus and Max, and the shared aim is indeed to have this resolved with the upgrade to 4.9.

[jdemeyer]

Is there any news here?

[slelievre]

GAP 4.9.0, a beta release for GAP 4.9, has been released.

This means we can start working on building Sage with it.

• Release notes: ​https://github.com/gap-system/gap/wiki/GAP-4.9-release-notes

[dimpase]

Do you know if they now provide a replacement for libGAP?

[slelievre]

Cc-ing @alexk and @markuspf.

[nthiery]

Replying to dimpase:

Do you know if they now provide a replacement for libGAP?

In principle yes: the branch I posted here last April was using their stock alpha version. Now the next step is to revive the branch and adapt it if needed.

I can work on this, but probably not at once. Any volunteer to take over welcome.

Markus: do you foresee any major changes in how GAP's lib is to be built that could affect what I was doing last April?

[dimpase]

I just noticed that in 4.9 smallgrps and transgrps are released under a GPL-compatible license. This is great! We really should make them (i.e. database_gap) standard Sage package, and do away with the extra hurdle of having database_gap optional.

[jdemeyer]

If you're willing to wait 2 months, this might be a good task for the Cernay workshop ​https://github.com/OpenDreamKit/OpenDreamKit/issues/251

[nthiery]

Replying to jdemeyer:

If you're willing to wait 2 months, this might be a good task for the Cernay workshop ​https://github.com/OpenDreamKit/OpenDreamKit/issues/251

Yes indeed; it will be helpful to have GAP people under hand. I may just become impatient before that for some research project of mine, in which case I'll have a head start :-)

[nthiery]

Replying to dimpase:

I just noticed that in 4.9 smallgrps and transgrps are released under a GPL-compatible license. This is great! We really should make them (i.e. database_gap) standard Sage package, and do away with the extra hurdle of having database_gap optional.

Yes! I can't wait for all the simplifications this will bring to us!

[jdemeyer]

Replying to fbissey:

So if I understand the mechanics involved, at the moment you have to configure and install gap. Then clean and configure and install for libgap. So it is still behaving like "two packages".

With the past and future modifications to the Sage build system in mind, I now object more strongly than before to treating two separate packages (GAP + libGAP) as one package.

The reason is that we now try to separate the build and install stages of packages. But that is incompatible with the recipe here of first "build and install GAP" and then "build and install libGAP".

[jdemeyer]

For an example of two Sage packages with the same sources, see gcc and gfortran. IMHO, you should do the same for GAP and libGAP (unless they are merged upstream but my impression is that this has not happened yet).

[fbissey]

Replying to jdemeyer:

Replying to fbissey:

So if I understand the mechanics involved, at the moment you have to configure and install gap. Then clean and configure and install for libgap. So it is still behaving like "two packages".

With the past and future modifications to the Sage build system in mind, I now object more strongly than before to treating two separate packages (GAP + libGAP) as one package.

The reason is that we now try to separate the build and install stages of packages. But that is incompatible with the recipe here of first "build and install GAP" and then "build and install libGAP".

Hopefully we will be getting something more serious at final release time, like with pari where you build libpari and then the gp executable. That would be the ideal scenario where it is really one package. Have to see how far they have gone with 4.9.

[jdemeyer]

A lot of changes on this ticket (also the changes likely to lead to merge conflicts) are related to the unprefixing of libGAP symbols. To ease development, maybe we should try to do that separately from the upgrade to GAP 4.9.

I was thinking to do something like #19915 but changing only the Sage source code (not libGAP) and undoing the prefixing with macros like #define SomeGapFunction libGAP_SomeGapFunction. What do you think?

[dimpase]

there is also libgap unprefixing in Python/Cython? that has to be done. Perhaps separating these from C level unprefixing would help making it smoother.

[jdemeyer]

Replying to dimpase:

there is also libgap unprefixing in Python/Cython? that has to be done. Perhaps separating these from C level unprefixing would help making it smoother.

Yes, that is what I meant. Creating a separate ticket for unprefixing only at the Cython level by using macros to translate between the unprefixed names in Cython and the prefixed names in libGAP.

[dimpase]

Replying to jdemeyer:

Replying to dimpase:

there is also libgap unprefixing in Python/Cython? that has to be done. Perhaps separating these from C level unprefixing would help making it smoother.

Yes, that is what I meant. Creating a separate ticket for unprefixing only at the Cython level by using macros to translate between the unprefixed names in Cython and the prefixed names in libGAP.

I am not sure I understand. Currently in Cython on can do something like

    from sage.libs.gap.libgap import libgap
    g=libgap.ProjectiveGeneralLinearGroup(3,3)

How will this change under what you propose?

[jdemeyer]

Replying to dimpase:

    from sage.libs.gap.libgap import libgap
    g=libgap.ProjectiveGeneralLinearGroup(3,3)

How will this change under what you propose?

The Python interface won't change at all. I am talking about how sage.libs.gap calls libGAP.

[jdemeyer]

As far as I can tell, GAP 4.9 has not been released. Does anybody know how close we are to an actual release?

[slelievre]

I think GAP 4.9.0 is considered a public beta for GAP 4.9.

The list of releases at

• ​https://www.gap-system.org/Releases/

still displays GAP 4.8.10 as the latest release.

In the list of past releases there, the GAP 4.8, 4.7, 4.6, 4.5 series start at GAP 4.8.2, 4.7.2, 4.6.2, 4.5.4, so probably GAP 4.x.y with those x and lower y were considered beta.

Still, seeing how 4.9.0 works with Sage would be nice.

[jdemeyer]

Where is the GAP source tarball which is supposed to be used here?

[git]

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

​72b5066	Unprefixed libGAP interface
​b7278a1	Upgrade to GAP 4.9

[dimpase]

Replying to jdemeyer:

Where is the GAP source tarball which is supposed to be used here?

here: ​https://www.gap-system.org/pub/gap/gap-4.9/beta/

[jdemeyer]

Replying to nthiery:

The branch attached to this ticket updates Sage to run on top of ​a branch of GAP by Markus Pfeiffer that adds libgap compilation and ​will be merged soon in the devel version of GAP.

It seems that this branch which "will be merged soon" is still not merged. It would be good to have an idea of whether or not that will happen.

[git]

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

​d28b575	Unprefixed libGAP interface
​daf8754	Unprefixed libGAP interface
​40e644a	Upgrade to GAP 4.9

[nthiery]

According to ​https://github.com/gap-system/gap/pull/1205#issuecomment-385670578, Markus's libgap branch won't be merged in GAP 4.9: it's not yet ready enough. Bummer.

We are discussing with Sebastian and Thomas what we can do to move forward here. They are very interested as well by libgap for the integration between GAP and Julia.

Tentative steps:

• As a warm up, upgrade our current gap and libgap spkg to 4.8.10
• Upgrade our current gap and libgap spkg to 4.9; with the many changes to the build system, this may get more tricky
• To make the switch to GAP's libgap less urgent, investigate whether we could change our current libgap spkg to not prefix GAP symbols, in order to be able to use GAP's dynamic packages; see #25273
• On the GAP side: move toward more incremental changes. In particular, extract the changesets from Markus branch that are non controversial and ready, and get them merged in GAP (4.9? 4.10?)

[slelievre]

​GAP 4.9.1 was released in May 2018.

This is the first stable release of GAP 4.9 (GAP 4.9.0 was a beta for GAP 4.9).

[gh-antonio-rojas]

I'm confused - I just built GAP 4.9.1 from the upstream tarball and 'make install' *does* install /usr/lib/libgap.so and libgap's headers. Is this a different libgap than the one mentioned here to be included in 4.10?

[slelievre]

One comment at pull request 1205 in GAP's github repo said GAP won't ship libGAP before GAP 4.10:

• ​GAP discussion: no libgap before GAP 4.10

Maybe this was changed though.

[jdemeyer]

Has anybody asked GAP developers for a policy on the prefixing?

In Cernay, I proposed that GAP should prefix some/all of their names like T_INT -> GAP_T_INT. For compiling GAP (+ GAP packages) itself, we can use a file with defines like

#define  T_INT  GAP_T_INT

to ensure backwards compatibility. This way, in GAP itself, one could use either T_INT or GAP_T_INT. Outside of GAP, only GAP_T_INT would work.

I think this is relatively easy to implement (I'm willing to do it) but I don't know what upstream GAP thinks of it.

[slelievre]

​GAP 4.9.2 was released in July 2018.

[nthiery]

Hi,

For the record: I experimented a bit a month ago and a bit more this morning with Dima with the latest gap master that has make libgap and make install-libgap rules. See the following branch, which builds libgap from GAP's pristine sources:

​https://git.sagemath.org/sage.git/log/?h=u/nthiery/gap-libgap

Note that, at this stage, Sage won't start with it; maybe even not compile.

[nthiery]

Has anybody asked GAP developers for a policy on the prefixing?

In Cernay, I proposed that GAP should prefix some/all of their names like T_INT -> GAP_T_INT. For compiling GAP (+ GAP packages) itself, we can use a file with defines like

#define  T_INT  GAP_T_INT

to ensure backwards compatibility. This way, in GAP itself, one could use either T_INT or GAP_T_INT. Outside of GAP, only GAP_T_INT would work.

I think this is relatively easy to implement (I'm willing to do it) but I don't know what upstream GAP thinks of it.

That's the idea indeed. Eventually they will provide a standalone header file that will define the official GAP API, and where all symbols will be prefixed. To bootstrap the process, we (well mostly Max!) reviewed and annotated the symbols that are used in Sage:

​https://hackmd.io/emNi76svSWCh1fBeLKqPdA#

[gh-timokau]

Thank you for working on this! Sage using an up-to-date gap and doing away with the additional libgap package will simplify packaging.

[vdelecroix]

update milestone 8.3 -> 8.4

[slelievre]

GAP 4.9.3 was released.

• ​GAP 4.9.3 release announcement

[dimpase]

note that the prefix of the tarball should start with src/ - else one gets weird error messages from the python's tar.

GAPs make install target now works in a preliminary form - on Linux for sure.

[dimpase]

This is WIP - builds, but libgap.pyx interaction loop has to be rewritten using new API.

Also, sage --gap is broken

gap: hmm, I cannot find 'lib/init.g' maybe use option '-l <gaproot>'?

---

New commits:

​c766ad6	Merge branch 'u/jdemeyer/upgrade_to_gap_4_9' of trac.sagemath.org:sage into libgap
​5152637	building GAP and libGAP from GAP
​6bed83a	changes for using libGAP from GAP - part 1
​f069122	removing util*

[dimpase]

oops, one still needs initialize() from util.pyx, in some form...

[git]

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

​8c28e13	Revert "removing util*"
​9c22e72	few comments

[nthiery]

Great to see the progress! Sounds like a productive week in Siegen. Thanks so much.

[git]

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

​14919d3

update util and the GAP package

[git]

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

​b3dc304	Merge branch 'develop' into libgap
​8a8a146	use MarkBag instead of MARK_BAG
​c5a5d4f	no more *_BAG, and a call signature fix
​d13fc80	more compilation fixes
​8b5ccbe	removed debugging utils (they need a rewrite)

[git]

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

​a0616de

adding --nointeract to libgap startup options

[dimpase]

at this point I am at

sage: libgap(1)
Error, Variable: '_rich_repr_' must have a value
Error, Variable: '_ipython_canary_method_should_not_exist_' must have a\
 value
[... several more errors like above]
gap: panic, could not open *errout*!

[git]

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

​9cc327f

adding --nointeract to libgap startup options

[dimpase]

In the current branch I am able to run

sage: g=libgap.SymmetricGroup(4)
sage: print(g.Order())
24
sage: i=g.Order().sage()
sage: i
24

So it's outputting GAP object that is broken ATM.

I'll try fixing the "old" GAP interface, perhaps it will help.

[nthiery]

Replying to dimpase:

In the current branch I am able to run

sage: g=libgap.SymmetricGroup(4)
sage: print(g.Order())
24
sage: i=g.Order().sage()
sage: i
24

Nice! Making good progress!

So it's outputting GAP object that is broken ATM. I'll try fixing the "old" GAP interface, perhaps it will help.

I am not sure what's broken above?

[dimpase]

Replying to nthiery:

So it's outputting GAP object that is broken ATM. I'll try fixing the "old" GAP interface, perhaps it will help.

I am not sure what's broken above?

see comment:86 - displaying (lib)GAP objects is broken. Also,

sage: gap(1)
** Gap crashed or quit executing 'SetUserPreference("HistoryMaxLines", 30);' **
Restarting Gap and trying again
** Gap crashed or quit executing '\$sage1:=1;;' **
...

the latter is probably due to some scripts gotten messed up by the upgrade.

And error handling in libgap interface needs work, too.

[dimpase]

libgap's display problem seems to be ipython-specific. Indeed

$ ./sage --python
Python 2.7.15 (default, Sep 15 2018, 05:05:36) 
[GCC 8.2.0] on linux2
Type "help", "copyright", "credits" or "license" for more information.
>>> from sage.all import *
>>> libgap(1)
1

[dimpase]

Some discussions related to libgap API:

• ​https://github.com/gap-system/gap/issues/2864
• ​https://github.com/gap-system/gap/issues/2839

and an accepted PR to fix a bug that was breaking libgap/Sage interface:

​https://github.com/gap-system/gap/pull/2840

[git]

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

​ece8cfc

fix gap() interface and gap_console, removed some old stuff

[git]

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

​043fb95

fixing all but one doctests in interfaces/gap.py

[dimpase]

One error - tab completion does not work on GAP objects:

sage -t --long --warn-long 49.3 src/sage/interfaces/gap.py
**********************************************************************
File "src/sage/interfaces/gap.py", line 1628, in sage.interfaces.gap.GapElement._tab_completion
Failed example:
    'Centralizer' in s5._tab_completion()
Exception raised:
    Traceback (most recent call last):
      File "/mnt/opt/Sage/sage-dev/local/lib/python2.7/site-packages/sage/doctest/forker.py", line 659, in _run
        self.compile_and_execute(example, compiler, test.globs)
      File "/mnt/opt/Sage/sage-dev/local/lib/python2.7/site-packages/sage/doctest/forker.py", line 1070, in compile_and_execute
        exec(compiled, globs)
      File "<doctest sage.interfaces.gap.GapElement._tab_completion[1]>", line 1, in <module>
        'Centralizer' in s5._tab_completion()
      File "sage/misc/cachefunc.pyx", line 2316, in sage.misc.cachefunc.CachedMethodCallerNoArgs.__call__ (build/cythonized/sage/misc/cachefunc.c:13467)
        self.cache = f(self._instance)
      File "/mnt/opt/Sage/sage-dev/local/lib/python2.7/site-packages/sage/interfaces/gap.py", line 1633, in _tab_completion
        v = P.eval(r'\$SAGE.OperationsAdmittingFirstArgument(%s)'%self.name())
      File "/mnt/opt/Sage/sage-dev/local/lib/python2.7/site-packages/sage/interfaces/gap.py", line 583, in eval
        result = Expect.eval(self, input_line, **kwds)
      File "/mnt/opt/Sage/sage-dev/local/lib/python2.7/site-packages/sage/interfaces/expect.py", line 1354, in eval
        for L in code.split('\n') if L != ''])
      File "/mnt/opt/Sage/sage-dev/local/lib/python2.7/site-packages/sage/interfaces/gap.py", line 789, in _eval_line
        raise RuntimeError(message)
    RuntimeError: Gap produced error output
    Error, no method found! For debugging hints type ?Recovery from NoMethodFound
    Error, no 1st choice method found for `Iterator' on 1 arguments

       executing \$SAGE.OperationsAdmittingFirstArgument(x);
**********************************************************************
1 item had failures:
   1 of   3 in sage.interfaces.gap.GapElement._tab_completion

[git]

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

​0128112

few important Cython types fixed

[git]

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

​c4d128e	pass the correct error callback function
​732cbdd	removing libgap_enter/exit

[dimpase]

in gap.* interface, tab completion is broken due to outdated code in src/ext/gap/sage.g: with GAP 4.10 one gets

gap> Read("sage.g");
gap> s5:=SymmetricGroup(5);
Sym( [ 1 .. 5 ] )
gap> \$SAGE.OperationsAdmittingFirstArgument(s5);
Error, no method found! For debugging hints type ?Recovery from NoMethodFound
Error, no 1st choice method found for `Iterator' on 1 arguments at /mnt/opt/Sage/sage-dev/local/gap/latest/lib/methsel2.g:250 called from
OPERATIONS[i + 1] at sage.g:69 called from
<function "LastReadValue">( <arguments> )
 called from read-eval loop at *stdin*:4
type 'quit;' to quit to outer loop

in GAP 4.8.6 this works:

gap> Read("sage.g");
gap> s5:=SymmetricGroup(5);
Sym( [ 1 .. 5 ] )
gap> \$SAGE.OperationsAdmittingFirstArgument(s5);
[ <Operation "ViewObj">, <Operation "ViewString">, <Operation "NameFunction">, <Operation "SetNameFunction">, 
...

I've asked ​here, as this is due to format change of internal GAP data...

[dimpase]

Very similar outdated code is in OperationInspector.operations in sage/libs/gap/operations.py.

[git]

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

​4121824

fix for tab completion for GAP interface

[git]

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

​1d60fdd

fix tab completion in libgap interface

[dimpase]

broken error handling:

sage: a=libgap(1)
sage: getattr(a,'_rich_repr_',None)
Error, Variable: '_rich_repr_' must have a value
sage: 

(with a default argument, getattr should just return it, with raising anything)

[dimpase]

Please see ​https://github.com/gap-system/gap/issues/2874 for the current sticking point, which is error handling and processing of error messages. Basically, the question is whether it's possible to retain the behaviour of libGAP without patches to the GAP's kernel.

I'd like someone who knows more about libGAP error handling to have a look at badly butchered by me parts in the branch dealing with it, and suggest what should be done.

Another (easier) thing to fix is GAP and libgap documentation display.

[git]

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

​e352f7e

modest progress - can do libgap(1)

[dimpase]

the interaction with ipython (which tries to get things like _rich_repr_ from GAP to display GAP objects) is far from done. The current problem:

sage: a=libgap(41)
sage: a
41
sage: b=libgap(42)
---------------------------------------------------------------------------
ValueError                                Traceback (most recent call last)
...
ValueError: libGAP: PrintTo: cannot open stream for output
Error, Length: <list> must be a list (not a boolean or fail)
...

this looks like a bug in my new (bad) implementation of GAP objects printing in Sage.

[dimpase]

perhaps we should rather try rebasing libGAP to GAP 4.10, which looks a less daunting task now than a total replacement with "native" GAP's libgap, get it working, and then replace all the libgap_* functions by the native ones.

[git]

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

​406cae6

still the interaction with ipython is broken

[dimpase]

On the present branch, I get

sage: a=libgap(1)
sage: a
*** trouble in PRINT_OR_APPEND_TO_STREAM***
...
*** trouble in PRINT_OR_APPEND_TO_STREAM***
1

after this GAP is pretty much nuked. This happens during ipython's attempts to pretty-print a, so that it tries to find _rich_repr_ etc. Specifically, at least on the system I am running this, the trouble starts not immediately, but where _rich_png_ (or the previous thing in the loop) is checked for by get_real_method in IPython/utils/dir2.py.

If I run the same from ./sage --python (after loading Sage there as usual) at least this much (and more) works.

I don't have a good idea how to debug this, so I'd like to pass the token now.

[dimpase]

oops, wrong ticket

[embray]

As long as we're upgrading GAP we should make sure DESTDIR builds work now; see #25044 which this conflicts with. But if this can get done soon we can skip #25044 and incorporate it into this. I'll give you a patch for it if you want.

I'm working on testing this on Cygwin now.

[embray]

Okay, I am again confused as to what upstream tarball to use, because the one in the ticket description no longer matches the checksum.

[dimpase]

I just checked that md5=017738e0ba9e166a19084d68123d3e1f as in GAP package checksum matches the md5 of ​https://github.com/dimpase/gap/releases/download/dev/gap-4.dev.tar.gz

Perhaps your download just got corrupted?

[thansen]

GAP 4.9.3 was uploaded to Debian unstable yesterday, the sagemath Debian package is broken until we can use GAP 4.9. Is the patch here already enough to use libgap from GAP 4.9 directly? It does not apply cleanly to sage 8.4. Or would you say updating the separate libgap to 4.9 is a better idea?

[dimpase]

Replying to thansen:

GAP 4.9.3 was uploaded to Debian unstable yesterday, the sagemath Debian package is broken until we can use GAP 4.9. Is the patch here already enough to use libgap from GAP 4.9 directly? It does not apply cleanly to sage 8.4.

the patch here is very far from a working one.

Or would you say updating the separate libgap to 4.9 is a better idea?

If you could manage this, it would be great and very useful, but it's a very nontrivial amount of work, due to substantial changes in GAP internals that happened in GAP 4.9.

[thansen]

I see. If there was progress on this from you guys it would be very much appreciated. The freeze period for the next Debian release starts in January. If we can't fix the problem in time, sage will not be in the next Debian release, which will be the stable release for two years.

[jdemeyer]

Replying to thansen:

The freeze period for the next Debian release starts in January.

For those of us who are not Debian experts, where is this documented? And what kind of changes are still allowed during the "freeze" period? For example, suppose for the sake of argument that this SageMath ticket is only merged in February, could it still be added as a patch to the Debian SageMath package?

[thansen]

It is documented here: ​https://release.debian.org/buster/freeze_policy.html

The broken sagemath will eventually be automatically removed from Debian testing, which will become the next Debian release. After February 12 it cannot migrate back to testing even if it is fixed. The earlier transition freeze on January 12 means that no more library updates with SONAME change can be done (which might be relevant for a sage update too).

[thansen]

About the example you mentioned: This ticket does not need to be merged, we just need a usable patch. It seems that the gap maintainer is open to providing libgap.la, see ​https://bugs.debian.org/912862 . It is not yet clear if gap 4.10 is released in time to be included before the freeze.

[embray]

This ticket mentions GAP 4.10, but should we not be focusing instead on GAP 4.9? Or would getting Sage working against GAP 4.10 imply support for 4.9 as well? I would want to check against both, and starting with the existing release (4.9) seems simpler...

Is there something at the system integration level I could be able to help with? ISTM this needs to be a high priority.

[embray]

Replying to thansen:

About the example you mentioned: This ticket does not need to be merged, we just need a usable patch. It seems that the gap maintainer is open to providing libgap.la, see ​https://bugs.debian.org/912862 . It is not yet clear if gap 4.10 is released in time to be included before the freeze.

If that's the case, for the sake of integration, it might make more sense to pin GAP in Debia to 4.9 and focus on getting a Sage out that works against GAP 4.9 instead of focusing on 4.10. Given the amount of time left to do this, waiting on an uncertain GAP 4.10 package(s) landing in Debian seems a bit risky...

[dimpase]

I think getting Sage's libGAP work with GAP 4.9 and/or 4.10 is much harder than in used to be in 4.8.x, due to major changes to GAP internals (although I have not tried it yet).

[nthiery]

Posting on behalf of Alexander Konovalov who does not yet have a trac account:

​http://www.gap-system.org/pub/gap/gap-4.10/

has now GAP 4.10.0. The website will be updated in due course, but that means that GAP 4.10 is out.

That's excellent news. Now it's up to Sage to catch up.

[embray]

Great! From what I saw yesterday it seemed like there were some doubts about how soon 4.10 would be released. But this being the case I rescind my previous comment.

[embray]

Replying to git:

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

​406cae6

still the interaction with ipython is broken

If I'm to understand correctly, you just added this patch for debugging purposes, right? It's not intended to go into the final product is it? I'm going to look into this problem.

[dimpase]

Replying to embray:

Replying to git:

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

​406cae6

still the interaction with ipython is broken

If I'm to understand correctly, you just added this patch for debugging purposes, right? It's not intended to go into the final product is it? I'm going to look into this problem.

That's right, it's for debugging purposes (although it merely shows that something is going wrong, not more :-)). You might like to coordinate this with Jeroen, as he's going to work on this branch too...

[embray]

I'm updating it right now to the 4.10 release. Jeroen is probably a better person to work on the LibGAP integration just because he has more experience with it. But if there any specific segfaults or anything I could look at I will. I'll take a look, for example, at the tab-completion issue...

[embray]

I'm also rebasing this branch on current develop, which needs a bit of work...

[dimpase]

Replying to nthiery:

Posting on behalf of Alexander Konovalov who does not yet have a trac account:

​http://www.gap-system.org/pub/gap/gap-4.10/

Alex should be aware that he can post on trac using his github account!

[embray]

Another thing that it seems still hasn't been done is to update the spkg-install for GAP to use DESTDIR-based installation. I'll just work on that simultaneously with this. If we have to upgrade GAP anyways that should be part of the process (and may expose some flaws in GAP's new installation process...)

[embray]

Needed to modify build/make/deps to directly make GAP a build dependency of sagelib as well, which is now needed (and remove the obsolete reference to the libgap spkg).

[jdemeyer]

Replying to embray:

I'm updating it right now to the 4.10 release.

Are you actually doing that? I haven't seen any update on the branch here.

[jdemeyer]

Thanks to this Cython PR (by me!), the libGAP T_INT thing might not be an issue anymore: ​https://github.com/cython/cython/pull/2006

Cython will still include <structmember.h> but only after all user code. It's still very fragile to have two different meanings for T_INT in the same source file (one enum value and one #define), but it should work for now. Of course, nobody can promise that it will remain working in the future.

[embray]

Replying to jdemeyer:

Replying to embray:

I'm updating it right now to the 4.10 release.

Are you actually doing that? I haven't seen any update on the branch here.

Just locally. It doesn't even build for me yet so no point in pushing.

[embray]

I should clarify: GAP itself compiles just fine. But "installation", such that there is any, is still a partially-broken headache, and I'm experimenting how best to incorporate a GAP installation into Sage.

For now I'll probably keep with something similar to what we have currently, of having a GAP_DIR under $SAGE_LOCAL/gap. But I'm trying to decide, in that case, what the best thing to do with libgap might be.

Though I'm also trying to see if we can do something entirely different with the installation. There's something close to a standard "make install" but it isn't quite right, and I don't know how it works with GAP packages.

[embray]

It seems that currently make install actually just installs the libtool wrapper script for gap, and not the plain gap executable. I'm not sure if this is intentional or not--normally this would not make a whole lot of sense I think. Though one advantage to it is that the libtool wrapper script adds the "GAP_ROOT" that it was built in to PATH, and thus it seems capable of finding its way to the original GAP_ROOT it was installed from. This doesn't really work though, from the perspective of system packaging. The paths hard-coded in the libtool wrapper script are to the build directory itself, so when we move all the GAP_ROOT files to a new location such as $SAGE_LOCAL/gap/gap-4.10 it doesn't work properly with that GAP_ROOT.

Better to do what we did traditionally of having bin/gap as a shell script that wraps the real gap executable and passes an appropriate -l flag. Dima's branch already has something to that effect, though rather than patching bin/gap.sh I think we should just install our own custom gap script, such as the one that's currently in this branch but unused.

I'm also experimenting with adding the actual executable under something like bin/gap-bin. I don't think we need all the files from the GAP source tree, much less build artifacts. I think we can have a minimal GAP_ROOT of sorts (e.g. under share/gap) that includes grp/, lib/, pkg/ at a minimum... Maybe doc/? And possibly also src/ (so that PageSource can work for kernel functions) but I'm not sure if that's needed or not...

[embray]

So far this "relatively minimal" GAP install seems to be working pretty well, and is much cleaner. I'm sure there are some glitches here and there, but I don't have the expertise to really know how to break it. A lot of the test suites under tst/ seem to work, or mostly work.

[dimpase]

Replying to embray:

So far this "relatively minimal" GAP install seems to be working pretty well, and is much cleaner. I'm sure there are some glitches here and there, but I don't have the expertise to really know how to break it. A lot of the test suites under tst/ seem to work, or mostly work.

Breaking libGAP should be easy, but GAP itself should be OK, I think.

[embray]

FWIW I'm also developing and testing this on Cygwin, just for some extra wind in my face.

[embray]

Note: Also need to manually install config.h into the headers. There's even a note in GAP's Makefile: # TODO: take care of config.h :(

[embray]

Progress on my branch so far ​here.

Currently it builds, but is broken (trying to do anything with libgap just hangs).

[jdemeyer]

Replying to embray:

Just locally. It doesn't even build for me yet so no point in pushing.

Depends. If it's an improvement over Dima's branch...

The question is basically: if I want to work on the libGAP interface, on which branch should I work?

[dimpase]

I'm building Erik's branch now, once done, I'll be able to say how well it does its job (Erik works on Windows, so this makes things even more interesting for him).

[dimpase]

Erik's branch does refer to *.gz tarball, but GAP has .*bz2 tarball. Thus

--- a/build/pkgs/gap/checksums.ini
+++ b/build/pkgs/gap/checksums.ini
@@ -1,4 +1,4 @@
-tarball=gap-VERSION.tar.gz
-sha1=ab95077df0775bc39dbf13becdc05818935a53d2
-md5=4d3cd775a7dcf5582b8b4dbc2b484bd4
-cksum=1444102593
+tarball=gap-VERSION.tar.bz2
+sha1=60d0b6b0127758da4270569c6ff42ad2364cac7e
+md5=7a94c54d2b1190d6471084b8fcbda6a9
+cksum=40709676

[dimpase]

Apart from this little issue, Erik's branch (u/embray/spkgs/gap-410) is OK---in the sense I can basically do what I was able to do on my branch. E.g. (after not forgetting to remove the stale (lib)GAP workspaces - all bets are off if they are present, e.g. it might be why Erik's job hang, probably something to rectify in the installation prcedure) I can do

sage: a=libgap(1)
sage: b=libgap(1)
sage: c=a+b
sage: c.sage()
2

and then get a segfault as follows:

sage: a
---------------------------------------------------------------------------
SignalError                               Traceback (most recent call last)
<ipython-input-5-3f786850e387> in <module>()
----> 1 a

/home/scratch2/dimpase/sage/sage/local/lib/python2.7/site-packages/IPython/core/displayhook.pyc in __call__(self, result)
    244             self.start_displayhook()
    245             self.write_output_prompt()
--> 246             format_dict, md_dict = self.compute_format_data(result)
    247             self.update_user_ns(result)
    248             self.fill_exec_result(result)

/home/scratch2/dimpase/sage/sage/local/lib/python2.7/site-packages/IPython/core/displayhook.pyc in compute_format_data(self, result)
    148 
    149         """
--> 150         return self.shell.display_formatter.format(result)
    151 
    152     # This can be set to True by the write_output_prompt method in a subclass

/home/scratch2/dimpase/sage/sage/local/lib/python2.7/site-packages/sage/repl/display/formatter.pyc in format(self, obj, include, exclude)
    211             exclude = self.default_mime()
    212         ipy_format, ipy_metadata = super(SageDisplayFormatter, self).format(
--> 213             obj, include=include, exclude=exclude)
    214         if not ipy_format:
    215             return sage_format, sage_metadata

/home/scratch2/dimpase/sage/sage/local/lib/python2.7/site-packages/IPython/core/formatters.pyc in format(self, obj, include, exclude)
    171             md = None
    172             try:
--> 173                 data = formatter(obj)
    174             except:
    175                 # FIXME: log the exception

<decorator-gen-9> in __call__(self, obj)

/home/scratch2/dimpase/sage/sage/local/lib/python2.7/site-packages/IPython/core/formatters.pyc in catch_format_error(method, self, *args, **kwargs)
    215     """show traceback on failed format call"""
    216     try:
--> 217         r = method(self, *args, **kwargs)
    218     except NotImplementedError:
    219         # don't warn on NotImplementedErrors

/home/scratch2/dimpase/sage/sage/local/lib/python2.7/site-packages/IPython/core/formatters.pyc in __call__(self, obj)
    334                 return printer(obj)
    335             # Finally look for special method names
--> 336             method = get_real_method(obj, self.print_method)
    337             if method is not None:
    338                 return method()

/home/scratch2/dimpase/sage/sage/local/lib/python2.7/site-packages/IPython/utils/dir2.pyc in get_real_method(obj, name)
     72 
     73     try:
---> 74         m = getattr(obj, name, None)
     75     except Exception:
     76         return None

/home/scratch2/dimpase/sage/sage/local/lib/python2.7/site-packages/sage/libs/gap/element.pyx in sage.libs.gap.element.GapElement.__getattr__ (build/cythonized/sage/libs/gap/element.c:6741)()
    581         try:
    582             proxy = make_GapElement_MethodProxy\
--> 583                 (self.parent(), gap_eval(name), self)
    584         except ValueError:
    585             raise AttributeError('name "'+str(name)+'" is not defined in GAP.')

/home/scratch2/dimpase/sage/sage/local/lib/python2.7/site-packages/sage/libs/gap/util.pyx in sage.libs.gap.util.gap_eval (build/cythonized/sage/libs/gap/util.c:5412)()
    310 #    print("gap_string: "+gap_string+"\n")
    311     try:
--> 312             sig_on()
    313             result = GAP_EvalString(cmd)
    314             nresults = LEN_LIST(result)

SignalError: Segmentation fault

This is the sort of crashes I traced to ipython trying to do a rich display of a, basically, making GAP repeatedly report errors.

It seems that now I can reproduce this hanging, it seems that libgap workspace does not load correctly, so to start Sage and libgap I have to manually remove it in ~/.sage/gap/.

[fbissey]

Replying to embray:

It seems that currently make install actually just installs the libtool wrapper script for gap, and not the plain gap executable. I'm not sure if this is intentional or not--normally this would not make a whole lot of sense I think. Though one advantage to it is that the libtool wrapper script adds the "GAP_ROOT" that it was built in to PATH, and thus it seems capable of finding its way to the original GAP_ROOT it was installed from. This doesn't really work though, from the perspective of system packaging. The paths hard-coded in the libtool wrapper script are to the build directory itself, so when we move all the GAP_ROOT files to a new location such as $SAGE_LOCAL/gap/gap-4.10 it doesn't work properly with that GAP_ROOT.

Better to do what we did traditionally of having bin/gap as a shell script that wraps the real gap executable and passes an appropriate -l flag. Dima's branch already has something to that effect, though rather than patching bin/gap.sh I think we should just install our own custom gap script, such as the one that's currently in this branch but unused.

I'm also experimenting with adding the actual executable under something like bin/gap-bin. I don't think we need all the files from the GAP source tree, much less build artifacts. I think we can have a minimal GAP_ROOT of sorts (e.g. under share/gap) that includes grp/, lib/, pkg/ at a minimum... Maybe doc/? And possibly also src/ (so that PageSource can work for kernel functions) but I'm not sure if that's needed or not...

This is not my experience here with just plain gap-4.10.0 package

fbissey@moonloop ~/sandbox/local-gap/bin $ ldd -r gap
        linux-vdso.so.1 (0x00007ffc039c2000)
        libgmp.so.10 => /usr/lib64/libgmp.so.10 (0x00007fa163a4a000)
        libz.so.1 => /lib64/libz.so.1 (0x00007fa163833000)
        libreadline.so.7 => /lib64/libreadline.so.7 (0x00007fa1635e6000)
        libm.so.6 => /lib64/libm.so.6 (0x00007fa16325c000)
        libdl.so.2 => /lib64/libdl.so.2 (0x00007fa163058000)
        libutil.so.1 => /lib64/libutil.so.1 (0x00007fa162e55000)
        libc.so.6 => /lib64/libc.so.6 (0x00007fa162a8d000)
        libncurses.so.6 => /lib64/libncurses.so.6 (0x00007fa162831000)
        /lib64/ld-linux-x86-64.so.2 (0x00007fa1645f2000)
fbissey@moonloop ~/sandbox/local-gap/bin $ file gap
gap: ELF 64-bit LSB pie executable, x86-64, version 1 (SYSV), dynamically linked, interpreter /lib64/ld-linux-x86-64.so.2, for GNU/Linux 3.2.0, with debug_info, not stripped

I just get plain executable binaries not libtool scripts after install. Both with the tarball in the description and the one from github (which is lighter because it doesn't include gap packages).

[embray]

I was getting outright hanging as soon as libgap is initialized (the sage.libs.gap.util.initialize function. It would hang so badly that I could only terminate the process with a SIGKILL, and it would leave my terminal in some invalid state (I have to type reset to get it back--the characters aren't even echo'd). I should note, this was on Cygwin. I will try now on Linux and see if I get any further.

Regardless, libgap probably shouldn't be doing any monkeying around with the terminal, and that might be part of the problem on Cygwin.

Dima, from where did you download the 4.10.0 tarball? The one I got from gap-system.org was a tar.gz as in my branch.

[embray]

Replying to fbissey:

This is not my experience here with just plain gap-4.10.0 package

fbissey@moonloop ~/sandbox/local-gap/bin $ ldd -r gap
        linux-vdso.so.1 (0x00007ffc039c2000)
        libgmp.so.10 => /usr/lib64/libgmp.so.10 (0x00007fa163a4a000)
        libz.so.1 => /lib64/libz.so.1 (0x00007fa163833000)
        libreadline.so.7 => /lib64/libreadline.so.7 (0x00007fa1635e6000)
        libm.so.6 => /lib64/libm.so.6 (0x00007fa16325c000)
        libdl.so.2 => /lib64/libdl.so.2 (0x00007fa163058000)
        libutil.so.1 => /lib64/libutil.so.1 (0x00007fa162e55000)
        libc.so.6 => /lib64/libc.so.6 (0x00007fa162a8d000)
        libncurses.so.6 => /lib64/libncurses.so.6 (0x00007fa162831000)
        /lib64/ld-linux-x86-64.so.2 (0x00007fa1645f2000)
fbissey@moonloop ~/sandbox/local-gap/bin $ file gap
gap: ELF 64-bit LSB pie executable, x86-64, version 1 (SYSV), dynamically linked, interpreter /lib64/ld-linux-x86-64.so.2, for GNU/Linux 3.2.0, with debug_info, not stripped

I just get plain executable binaries not libtool scripts after install. Both with the tarball in the description and the one from github (which is lighter because it doesn't include gap packages).

I can't tell exactly what file you're looking at, but if the bin/ directory you reference above is the one that gets created in the gap source three, that's the normal executable. I'm taking about the gap that gets installed to $PREFIX when you run make install.

[embray]

It's possible this is only a bug on Cygwin too. I haven't played with it on Linux yet, but I am beginning to now.

[embray]

On Linux I was able to reproduce Dima's results. The first libgap() call is a little slow, which is to be expected--the GAP interface is initialized lazily. After that it works up to outputting a, whereas a.sage() and the like works. I also get the tab-completion segfault.

Jeroen: do you have anything to report from your investigation so far? I'd like to look into these segfaults but I don't want to duplicate your effort if you've already made significant progress.

[embray]

Ooh, interesting. After my initial crash I started Sage up again, tried libgap(1), and it hanged just like on Cygwin. The process just sits without any CPU usage and does not respond to ctrl-C. It does respond to a SIGTERM. Could this have something to do with a corrupt GAP workspace?

[embray]

I suspected something like this: GAP is installing its own readline callbacks, overriding the Python REPL's. I think it should not be doing that at all if --nointeract.

[dimpase]

Replying to embray:

On Linux I was able to reproduce Dima's results. The first libgap() call is a little slow, which is to be expected--the GAP interface is initialized lazily. After that it works up to outputting a, whereas a.sage() and the like works. I also get the tab-completion segfault.

Jeroen: do you have anything to report from your investigation so far? I'd like to look into these segfaults but I don't want to duplicate your effort if you've already made significant progress.

I suspect that I forgot that we must not hold pointers to GAP objects - as GAP’s GC is moving things, this would be asking for trouble. So this could be a rich source of bugs I introduced in my branch :(

[dimpase]

​https://github.com/gap-system/gap/pull/2988 is introducing more things to libgap API, have a look (and say)

[dimpase]

GAP 4.10 tarballs are here: ​http://www.gap-system.org/Releases/

[embray]

Apparently passing the -E argument disables readline support so we should do that. Trying it now.

[embray]

The issue with "corrupt" (?) gap workspaces seems to affect old versions too. After a crash, I switched over to a Sage build with GAP 4.8 just to compare something, and calling libgap hanged again. I have to rm -rf the old workspace to get it to work again. Nevermind; was still running the development Sage. Old Sage does not appear to be impacted.

[embray]

This comment from Markus seems relevant: ​https://github.com/gap-system/gap/pull/2702#issuecomment-414960197 I am observing something that sounds like what he describes, and it would explain the terminal non-responsiveness. I'm not exactly sure where or why this is happening though.

[embray]

-E fixed problems with readline, but we're still having problems in error handling. Somehow the error handling code--particularly manipulation of the ERROR_OUTPUT stream, results in a stack overflow \o/

[dimpase]

libgap is started in sage.libs.gap.util.initialize(), and you see that

argv[11] = "--nointeract"

i.e. --nointeract is passed. So you add -E to this list, right?

The GAP ​patch you disabled in your branch makes the stack overflow you see a bit more transparent, I think...

[dimpase]

in my branch in sage.libs.gap.util.initialize() I have cdef int argc = 11, and this seems to be off by 1. Is it the root of all the evil?

[embray]

Replying to dimpase:

in my branch in sage.libs.gap.util.initialize() I have cdef int argc = 11, and this seems to be off by 1. Is it the root of all the evil?

No, that seems right to me...? No, you're right. It's off by one.

[embray]

Replying to dimpase:

libgap is started in sage.libs.gap.util.initialize(), and you see that

argv[11] = "--nointeract"

i.e. --nointeract is passed. So you add -E to this list, right?

Basically. I replaced -T with -E, since -T is implied by --no-interact (which should probably also imply -E, but it doesn't).

However, as you pointed out, there's an off-by-one and now I'm not so sure if --no-interact was being processed at all.

[embray]

There's what I would consider a bug in GAP, where the function PRINT_OR_APPEND_TO_STREAM goes:

    /* try to open the file for output                                     */
    i = OpenOutputStream(stream);
    if ( ! i ) {
        ErrorQuit( "%s: cannot open stream for output", (Int)funcname, 0L );
        return 0;
    }

but if this is called to print to the ERROR_OUTPUT stream during ErrorQuit then we get an infinite recursion. This is the root cause of the segfault's we're seeing. What I'm not sure about is why OpenOutputStream would fail, since we close ERROR_OUTPUT after each error, and then re-open it as a new stream. That part bothers me. But I think GAP should check here if stream == ERROR_OUTPUT and if so reset ERROR_OUTPUT to its default value as a last resort, or something to that effect.

[dimpase]

Erik, do you have your updated branch somewhere online? My last attempt to build your branch ends with a glorious collision of GAP's T_INT and Python's T_INT...

Yes, I agree that PRINT_OR_APPEND_TO_STREAM is a problem, and the patch in my branch I mentioned in comment 166 provides a crude workaround. I already mentioned this to GAP devs more than once, see e.g. ​https://github.com/gap-system/gap/issues/2487

[embray]

No, I don't have an updated branch. I'm not seeing that problem before.

I think I have some idea what the issue is. The problem is that because we call sig_error() in our error handler, we jump out of GAP's eval loop before it can perform certain cleanup, including closing the output stream that it's using to capture output in GAP_EvalString.

So we're leaving things in a slightly fragile state; and more importantly we're not cleaning up after ourselves. I believe that's the case here. I wonder if we should forgo sig_error() entirely and just let GAP_EvalString return naturally. The error handler can just record somewhere that an error occurred, and we can raise an exception from within gap_eval`.

[embray]

Yep. This patch fixes it:

src/sage/libs/gap/gap_includes.pxd

@@ -129 +129 @@
     ctypedef void (*CallbackFunc)()
     void GAP_Initialize(int argc, char ** argv, char ** env,
         CallbackFunc, CallbackFunc)
-    Obj GAP_EvalString(const char *)
+    Obj GAP_EvalString(const char *) except *
     Obj GAP_ValueGlobalVariable(const char *)
 #    void libgap_start_interaction(char* inputline)
 #    char* libgap_get_output()

src/sage/libs/gap/util.pyx

@@ -221 +221 @@
     argv[8] = "64m"
 
     argv[9] = "-q"    # no prompt!
-    argv[10] = "-T"    # no debug loop
-    argv[11] = "--nointeract"
+    argv[10] = "-E"    # don't use readline as this will interfere with Python
+    argv[11] = "--nointeract"  # Implies -T
     argv[12] = NULL
     cdef int argc = 11   # argv[argc] must be NULL
 
@@ -237 +237 @@
 
     # Initialize GAP and capture any error messages
     # The initialization just prints error and does not use the error handler
-    # libgap_initialize(argc, argv)
     try:
         GAP_Initialize(argc, argv, environ, &gasman_callback, &error_handler)
     except RuntimeError as msg:
         raise RuntimeError('libGAP initialization failed\n' + msg)
 
-    # gap_error_msg = char_to_str(libgap_get_output())
-    GAP_EvalString('libgap_errout := \"\"; ERROR_OUTPUT := OutputTextString(libgap_errout, false);')
-
-    # The error handler is called if a GAP evaluation fails, e.g. 1/0
-    # libgap_set_error_handler(&error_handler)
+    # Set the ERROR_OUTPUT global in GAP to an output stream in which to
+    # receive error output
+    GAP_EvalString('libgap_errout := ""; '
+                   'ERROR_OUTPUT := OutputTextString(libgap_errout, false);')
 
     # Prepare global GAP variable to hold temporary GAP objects
     global reference_holder
@@ -307 +305 @@
     # so that Cython doesn't dereference it before libGAP is done with
     # its contents.
     cmd = str_to_bytes(gap_string + ';\n')
-#    print("gap_string: "+gap_string+"\n")
     try:
-            sig_on()
-            result = GAP_EvalString(cmd)
-            nresults = LEN_LIST(result)
-#            print("nresults="+str(nresults)+"\n")
-            if nresults > 1: # to mimick the old libGAP
-                raise ValueError('can only evaluate a single statement')
-            result = ELM_LIST(result, 1) # 1-indexed!
-#            print("result's length: "+str(LEN_LIST(result))+"\n")
-            if ELM_LIST(result, 1) != GAP_True:
-                # libgap_call_error_handler()
-                        print("An error occurred, but libGAP has no handler set")
-                        return GAP_False # needs work
-            sig_off()
+        sig_on()
+        result = GAP_EvalString(cmd)
+
+        # If an error occurred in GAP_EvalString we won't even get
+        # here if the error handler was set; but in case it wasn't
+        # let's still check the result...
+        nresults = LEN_LIST(result)
+        if nresults > 1:  # to mimick the old libGAP
+            # TODO: Get rid of this restriction eventually?
+            raise ValueError('can only evaluate a single statement')
+
+        result = ELM_LIST(result, 1) # 1-indexed!
+        if ELM_LIST(result, 1) != GAP_True:
+            raise RuntimeError("an error occurred, but libGAP has no "
+                               "error handler set")
     except RuntimeError as msg:
-            raise ValueError('libGAP: '+str(msg).strip())
+        raise ValueError(f'libGAP: {msg}')
+    finally:
+        sig_off()
 
     return ELM_LIST(result, 2)
 
@@ -357 +358 @@
 
 cdef void error_handler():
     """
-    The libgap error handler
+    The libgap error handler.
 
-    We call ``sig_error()`` which causes us to jump back to the Sage
-    signal handler. Since we wrap libGAP C calls in ``sig_on`` /
-    ``sig_off`` blocks, this then jumps back to the ``sig_on`` where
-    the ``RuntimeError`` we raise here will be seen.
+    If an error occurred we set a RuntimeError; when the original
+    GAP_EvalString returns this exception will be raised.
+
+    TODO: We should probably prevent re-entering this function if we
+    are already handling an error; if there is an error in our stream
+    handling code below it could result in a stack overflow.
     """
     cdef Obj r
+    cdef char *msg
+
+    # TODO: Do we need/want this ClearError??
+    ClearError()
+
+    # Close the error stream: This flushes any remaining output and closes
+    # the stream for further writing; reset ERROR_OUTPUT to something sane
+    # just in case (trying to print to a closed stream segfaults GAP)
+    GAP_EvalString('CloseStream(ERROR_OUTPUT); '
+                   'ERROR_OUTPUT := "*errout*"; '
+                   'MakeImmutable(libgap_errout);');
     r = GAP_ValueGlobalVariable("libgap_errout")
-    GAP_EvalString("CloseStream(ERROR_OUTPUT);");
-    GAP_EvalString('libgap_errout := \"\"; ERROR_OUTPUT := OutputTextString(libgap_errout, false);')
 
-    msg_py = char_to_str(CSTR_STRING(r))
-    msg_py = msg_py.replace('For debugging hints type ?Recovery from NoMethodFound\n', '')
+    # Grab a pointer to the C string underlying the GAP string libgap_errout
+    # then copy it to a Python str (char_to_str contains an implicit strcpy)
+    msg = CSTR_STRING(r)
+    if msg != NULL:
+        msg_py = char_to_str(msg)
+        msg_py = msg_py.replace('For debugging hints type ?Recovery from '
+                                'NoMethodFound\n', '')
+    else:
+        # Shouldn't happen but just in case...
+        msg_py = "An unknown error occurred in libGAP"
+
+    # Reset ERROR_OUTPUT with a new text string stream
+    GAP_EvalString('libgap_errout := ""; '
+                   'ERROR_OUTPUT := OutputTextString(libgap_errout, false);')
+
     PyErr_SetObject(RuntimeError, msg_py)
-    ClearError()
-    sig_error()
 
 
 ############################################################################

Now printing GAP elements works, and it also fixes tab-completion. So a lot of basic stuff works with this. I make no guarantees about memory leaks, reference counting, etc.

[dimpase]

you still have wrong argc count on line 227.

[dimpase]

But otherwise it's a lot of progress, great!!!

[jdemeyer]

Replying to embray:

Jeroen: do you have anything to report from your investigation so far? I'd like to look into these segfaults but I don't want to duplicate your effort if you've already made significant progress.

No, on the contrary: I'm not doing anything for the moment since I don't want to interfere were your (Erik and Dima) work.

[jdemeyer]

Replying to dimpase:

My last attempt to build your branch ends with a glorious collision of GAP's T_INT and Python's T_INT...

Which branch is that? According to my earlier analysis, that shouldn't happen. So I'd like to see that branch.

[jdemeyer]

Replying to embray:

The problem is that because we call sig_error() in our error handler, we jump out of GAP's eval loop before it can perform certain cleanup, including closing the output stream that it's using to capture output in GAP_EvalString.

If so, the problem is not limited to sig_error() but also to interrupts. So it seems like your proposed solution of not using sig_error() will break interrupts, which I do not consider acceptable.

Would there be a way of patching GAP to make GAP_EvalString stateless?

[embray]

I don't know what you mean that it would "break interrupts". Either way handling an interrupt with cysignals and not returning to GAP is going to break things.

There's definitely still a problem somewhere:

sage: b = libgap.eval('Sleep(5);')
Caught error at top-most level, probably quit from library loading

It sleeps, then displays that message immediately and quits the process. Strangely, the only source I can find for that message is at the end of InitializeGap which should have already exited. I think there is still some strange things going on in GAP's error handling...

On the other hand if I do:

sage: f = libgap.eval('Sleep;')
sage: f
<Gap function "Sleep">
sage: f(5)

it works. But when I hit Ctrl-C:

sage: f(10)
^CCaught error at top-most level, probably quit from library loading

and the process quits again.

[jdemeyer]

Replying to embray:

I don't know what you mean that it would "break interrupts".

(A) If the keep the sig_on() statements, then CTRL-C would break GAP in the same way as sig_error() breaks GAP.

(B) If you remove the sig_on() statements, then GAP can no longer be interrupted with CTRL-C.

I consider both alternatives not acceptable.

[embray]

I'm not even sure yet how GAP handles signals. Do you know anything about it? I haven't looked into it yet.

[embray]

There is a RegisterSyLongjmpObserver that might be useful. I'm not sure if it's used at all in signal handling though.

[jdemeyer]

Looking at the GAP sources right now... GAP seems to handle interrupts similar to CPython, by setting some global state. The function below is called by the SIGINT handler:

/****************************************************************************
**
*F  InterruptExecStat() . . . . . . . . interrupt the execution of statements
**
**  'InterruptExecStat'  interrupts the execution of   statements at the next
**  possible moment.  It is called from 'SyAnsIntr' if an interrupt signal is
**  received.  It is never called on systems that do not support signals.  On
**  those systems the executors test 'SyIsIntr' at regular intervals.
**
**  'InterruptExecStat' changes all entries   in the executor  dispatch table
**  'ExecStatFuncs'  to point to  'ExecIntrStat',  which changes  the entries
**  back, calls 'Error', and redispatches after a return from the break-loop.
*/
void InterruptExecStat ( void )
{
    /* remember the original entries from the table 'ExecStatFuncs'        */
    STATE(CurrExecStatFuncs) = IntrExecStatFuncs;
}

[embray]

Okay. GAP's own signal handling (which just handles SIGINT) uses InterruptExecStat to signal to the eval loop that an interrupt has occurred, so it will break out and return an error at the first opportunity. We can probably make use of that.

[embray]

I see we've got our noses on the same trail :) I'm going to step away for a bit anyways.

[jdemeyer]

Replying to embray:

We can probably make use of that.

This will need new hooks in cysignals though. Also, all sig_on() and sig_off() statements should be removed from the GAP interface.

Given that this is not the first priority, I'll let you guys work further. We can deal with interrupts once we have a mostly-finished GAP interface.

[embray]

Replying to jdemeyer:

Replying to embray:

We can probably make use of that.

This will need new hooks in cysignals though. Also, all sig_on() and sig_off() statements should be removed from the GAP interface.

I was wondering about that. I'm not sure cysignals is needed at all really, though if it can provide a wrapper interface to make things easier that might be nice. I think it would be sufficient, when entering and leaving GAP_EvalString, to just register/unregister a custom signal handler, for SIGINT at the very least.

We also need to be careful around GAP_Initialize to restore any existing SIGINT handler that gets clobbered by GAP's initialization...

[jdemeyer]

Replying to embray:

to just register/unregister a custom signal handler, for SIGINT at the very least.

Since that involves system calls, that would be less efficient than doing something within cysignals. I bet that this would cause a noticable slowdown for a calculation like 1 + 1.

[embray]

Replying to embray:

I don't know what you mean that it would "break interrupts". Either way handling an interrupt with cysignals and not returning to GAP is going to break things.

There's definitely still a problem somewhere:

sage: b = libgap.eval('Sleep(5);')
Caught error at top-most level, probably quit from library loading

It sleeps, then displays that message immediately and quits the process. Strangely, the only source I can find for that message is at the end of InitializeGap which should have already exited. I think there is still some strange things going on in GAP's error handling...

The reason for this appears to be a bug in how we're processing the result list from GAP_EvalString--in particular it is the equivalent of a Python IndexError on the list. GAP is treating this as a sort of internal error, meaning it jumps to the last invocation of its TRY_IF_NO_ERROR/CATCH_ERROR macros. But since in our code we're not in a TRY/CATCH, perversely the last TRY is back in InitializeGap, so it invokes a longjmp back there.

It seems we can work around that by using the ELM_PLIST function directly rather than the higher-level generic ELM_LIST. The lower-level function does not invoke GAP error handling. I've added another comment about this here: ​https://github.com/gap-system/gap/pull/2988/files#r233021437

[embray]

Another annoyance--albeit one that I'm not sure if Sage's libGAP handles either--is that if GAP Panic(...)s it will just exit(1) the process and I'm not sure if there's any way around that currently...

[dimpase]

Replying to jdemeyer:

Replying to dimpase:

My last attempt to build your branch ends with a glorious collision of GAP's T_INT and Python's T_INT...

Which branch is that? According to my earlier analysis, that shouldn't happen. So I'd like to see that branch.

Well, it was something cooked up a bit, it could be I just messed things up. I hope Erik can update his branch and share it, so that I can try again...