GitLab

In Python 3.7, the re module changes the name of its _pattern_type
attribute to "Pattern", this patch handles both versions

In C++14, static constexpr expressions that are odr-used must be
declared in the compilation unit at namespace scope even if they are
defined and assigned to in the class.

The `indices` static constexpr attribute of tuple_automaton_impl was
missing that namespace-level declaration, some compiler were fine with
it, but clang 8 isn't, unless compiling with -O1 or higher.

Fix #216

More recent versions of bison generate some switch blocks using a
`YYCASE_` macro. This macro made the fuse-switch script stop outputing,
resulting in a botched generated file.

This commit make the fuse-switch ignore those switch blocks and keep
them as is.

fix #215

Younes Khoudli authored Mar 18, 2019 and Paul Hervot committed Aug 18, 2019

Boost 1.69 introduced an ADL barrier on boost::distance(), so the call
that depended on it failed to build. Simply use qualified name lookup
instead.

See also https://github.com/boostorg/range/pull/75

.

* vcsn/core/rat/printer.hh: Fix.
Reported-by: Baptiste Esteban <baptiste.esteban@epita.fr>

Fix #214

That's only an initial stab.

* tests/tools/help.chk: s/$@/"$@"/.

No visible performance changes.

* tests/python/proper.dir/lan-poly.3.fail.gv,
* tests/python/proper.dir/lao-r.in.gv: These are actually daut files.

We parse in `lan, q`, we parse `<1/2>` as `<1/2>$`, not as `<1/2>\e`.
This is to support pre- and post-transition labels (e.g., `$ 0 <1/2>`).
Unfortunately it's easy, in particular with lao, to forget to write
`\e`, and I have been bitten: I tried to get the weight-series

    context = lao, expressionset<lal, q>
    $ -> 0
    0 -> 1 <a>
    1 -> 1 <b>
    1 -> 2 <c>
    2 -> $

But I got `<a+c>`.  The culprit was actually proper/is_proper: both
consider that automaton to be proper.  And the real culprit is
actually the user: there are indeed no spontaneous transitions in this
automaton: they are labeled with `$` because we did not use `\e`.
This error is too easy, so be sure to be kind to the user, and map the
special label to one when building the automaton.

It turns out that the test suite itself contained such an error: a
proper automaton still had a pseudo-spontaneous transition.

* vcsn/algos/edit-automaton.hh: Map $ to \e when adding an inner
transition.
* tests/python/proper.dir/lao-r.out.gv: Fix.

* vcsn/algos/edit-automaton.hh: there is not reason to accept
`$ 0 <1/2>\e`, spell it `$ 0 <1/2>`.

That "no" is useful for configure's logs, but not in our configuration
file.

A user has, in his config:

    ipython: no
    jupyter: /opt/local/var/macports/build/_opt_local_var_macports_sources_rsync.macports.org_macports_release_tarballs_ports_devel_vcsn/vcsn/work/vcsn-2.7/build-aux/bin/missing jupyter

This is not helpful in anyway.  In particular, `missing` is unlikely
to still be there when the user fires `vcsn jupyter`.

Leave a big fat nothing instead.

* tests/bin/ipynbdoctest.py: Let's try to look like the actual output
from IPython.  It's output is pretty-printed, but it does not look
like what I get using pprint.  They seem use the json module.

* tests/bin/update-test: Enable patches based on diffs.

It now fails when an exception escapes.

It took me way too long to understand the new failure...

To match the fact that dictionaries are sorted in Python 3.7, IPython
no longer sort the dictionaries when displaying
them (https://github.com/ipython/ipython/issues/10110).

So build them in the sorted order, so that our notebook tests produce
the expected output.

I wanted to know where OpenFST stands on is-deterministic when there
are spontaneous transitions.  The answer is: Open FST does not care
about them wrt determinism: the `one` label is treated like any other
label.

This is something I was considering for Vcsn too...  But I'm uneasy
with this, so I think I will rather introduce a new property P, and
define deterministic as having P and no spontaneous transitions.

See also #196.

* python/vcsn/automaton.py (fstinfo): New.
* tests/python/efsm.py: Check.

PATH etc. must be ready before being used.

I recently broke the install `vcsn`, as a result, (almost) nothing
works properly.  I didn't notice, because `make check` works, and we
run only it.

So try one installcheck for each arch, and speed/debug.

A new target "Release", which prepares some of the repetitive tasks.

* doc/notebooks/Trouble-Shooting.ipynb: Rename as...
* doc/notebooks/Troubleshooting.ipynb: this.

Rather than calling equal _and_ less.

* tests/bin/vcsn.in (VCSN_DOCDIR): Define it.

Let's use a nice and clear syntax.

Rename the formats efsm.bzip2 and efsm.lzma.

Now that we support compressed efsm, let's compress the efsm
files of sms2fr.  Not only will this save space, it should also
help with sites that claim that Vcsn is 98.3% of shell-scripts...

The results with Xz are impressive: graphemic.efsm is reduced to
5% of its original size, and syntactic.efsm 12%.

    -rw-r--r--   1 akim  staff  93099091 21 mar 23:46 graphemic.efsm
    -rw-r--r--   1 akim  staff  11713613 21 mar 19:05 graphemic.efsm.bz2
    -rw-r--r--   1 akim  staff   5502716 21 mar 23:22 graphemic.efsm.xz
    -rw-r--r--   1 akim  staff  62289670 21 mar 23:46 syntactic.efsm
    -rw-r--r--   1 akim  staff  13055960 13 mar 19:44 syntactic.efsm.bz2
    -rw-r--r--   1 akim  staff   7916576 21 mar 23:24 syntactic.efsm.xz

And Vcsn is now 68.8% C++, 20.6% Python and 3.5% shell-script!

While at it:
- use int not char for options
- don't use references to keep the automata
- don't force the format to efsm: let vcsn guess the format