* vcsn/algos/make-context.hh: .
* vcsn/alphabets/char.hh: .
* vcsn/alphabets/setalpha.hh: .
* vcsn/core/rat/info.hh: .
* vcsn/core/rat/less-than.hh: .
* vcsn/core/rat/printer.hh: .
* vcsn/core/rat/ratexp.hh: .
* vcsn/core/rat/transpose.hh: .
* vcsn/core/rat/visitor.hh: .
* vcsn/ctx/context.hh: .
* vcsn/dyn/polynomial.hh: .
* vcsn/dyn/polynomialset.hh: .
* vcsn/dyn/ratexp.hh: .
* vcsn/dyn/ratexpset.hh: .
* vcsn/dyn/weight.hh: .
* vcsn/empty.hh: .
* vcsn/labelset/genset-labelset.hh: .
* vcsn/labelset/letterset.hh: .
* vcsn/labelset/nullableset.hh: .
* vcsn/labelset/oneset.hh: .
* vcsn/labelset/wordset.hh: .

This is more troublesome.  Consider a public (visibility=default)
class template 'derived<T>', parameterized with private
(visibility=invisible) type 'payload'.  The resulting type
'derived<payload>' is uniquely defined for G++, and _not_ for Clang++.
As a result, dynamic_cast fails with Clang++, not with G++.

We have this problem with concrete_abstract_weight<WeightSet>: Clang
requires the WeightSet to be public.

See http://stackoverflow.com/questions/19496643.

* vcsn/core/rat/ratexpset.hh, vcsn/dyn/weight.hh,
* vcsn/dyn/weightset.hh, vcsn/weights/b.hh, vcsn/weights/f2.hh,
* vcsn/weights/polynomialset.hh, vcsn/weights/q.hh, vcsn/weights/r.hh,
* vcsn/weights/z.hh, vcsn/weights/zmin.hh:
Declare these classes/class templates as public.

Clang does not support namespace attributes.  A pity.

* vcsn/misc/attributes.hh (ATTRIBUTE_DLLEXPORT, ATTRIBUTE_DLLIMPORT): New
for Clang.
* vcsn/dyn/algos.hh, vcsn/misc/escape.hh, vcsn/misc/stream.hh,
* lib/vcsn/rat/read.hh: Do not use the namespace attributes.

* Makefile.am: here.
* lib/local.mk, lib/vcsn/local.mk: Simplify.

* vcsn/ctx/lal_char_b.hh: here.

* vcsn/core/transition.hh: here.
Touching b.hh should not trigger the recompilation of all the contexts!

* vcsn/dyn/automaton.hh, vcsn/dyn/context.hh,
* vcsn/dyn/polynomial.hh, vcsn/dyn/ratexp.hh,
* vcsn/dyn/weight.hh: Here.

* vcsn/algos/aut-to-exp.hh, vcsn/algos/product.hh: Beware of signedness.

* configure.ac: here.
For Clang.

The implementation was buggy:

  $ for i in 1 2 3 4 5 6 7 8
    do
     printf "$i "
     vcsn-power -f share/vcsn/lal_char_z/c1.gv $i | vcsn-evaluate -f - ba
    done
  1 2
  2 4
  3 8
  4 16
  5 64
  6 64
  7 512
  8 256

Interestingly, it is not always a clear speed-up to use the
square-based exponentiation, plain iterative products might be faster.
This would require a more thorough bench evaluation.  Other factors
(especially the implementation of transitions in mutable_automaton)
are expected to interfere.

This is "vcsn power -O null -f binary.gv $i":
          8    9    10   11   12   13   14   15     16
products 0.05 0.06 0.11 0.17 0.45 1.54 6.43 28.65 145.95
squares  0.06 0.05 0.06 0.08 0.19 0.71 3.51 19.10 207.46

Score:
  7.93s (0.05s+7.88s): ladybird 21 | determinize -O null
  0.10s (0.05s+0.05s): de-bruijn 20 | evaluate -O null -f - aaaaaaaaaaaaaaaaaaaaa
 11.65s (0.85s+10.80s): thompson -C "lan_char(a)_b" -Ee "a?{2000}" | proper -O null
 21.30s (9.52s+11.78s): standard -Ee "(a+b+c+d)?{100}" | aut-to-exp -O null
  0.51s (0.05s+0.46s): standard -C "lal_char(ab)_z" -Ee "(a+b)*b(<2>a+<2>b)*" | power -O null -f- 10
  0.05s (0.05s): standard -E -e "(a?){70}" -o a70.gv
 12.69s (12.69s): product -O null -f a70.gv a70.gv
  0.19s (0.19s): shuffle -O null -f a70.gv a70.gv
 15.87s (15.87s): infiltrate -O null -f a70.gv a70.gv

* vcsn/ctx/context.hh (operator=): New.
* vcsn/core/mutable_automaton.hh (context_): No longer const, to allow...
(operator=): the definition of a move-assignment operator.
* vcsn/algos/product.hh (power): Fix the implementation.

* tests/tafkit/power.dir/binary^0.gv, tests/tafkit/power.dir/binary^1.gv:
New, extracted from...
* tests/tafkit/power.chk: Here.
Adjust.
Check the previous failures (e.g., 2^5 => 32, not 64).

* vcsn/algos/product.hh: Here.
* vcsn/ctx/instantiate.hh (register_commutative_semiring_kind_functions): Add.
* vcsn/algos/info.hh (is_ambiguous): Adjust enable_if condition.

* tests/tafkit/infiltrate.chk, tests/tafkit/product.chk,
* tests/tafkit/shuffle.chk: Verify failure on non-commutative weightsets.

* vcsn/core/rat/ratexpset.hh: Here.
* vcsn/weights/b.hh: Here.
* vcsn/weights/f2.hh: Here.
* vcsn/weights/polynomialset.hh: Here.
* vcsn/weights/q.hh: Here.
* vcsn/weights/r.hh: Here.
* vcsn/weights/z.hh: Here.
* vcsn/weights/zmin.hh: Here.

The infiltration algorithm is essentially a combination of product and
shuffle.  Factor the common operations into private methods.

* vcsn/algos/product.hh: Here.

* vcsn/algos/product.hh (infiltrate): Add algorithm.
* vcsn/ctx/instantiate.hh: Adjust.
* vcsn/dyn/algos.hh: Adjust.
* lib/vcsn/algos/product.cc: Adjust.
* bin/vcsn-infiltrate.cc: New.
* bin/local.mk: Adjust.

* tests/tafkit/infiltrate.chk: New.
* tests/tafkit/local.mk: Adjust.

* vcsn/algos/derive.hh (derived_terms): New.
* vcsn/ctx/instantiate.hh: Instantiate.
* vcsn/dyn/algos.hh (dyn::derived_terms): Declare.
* lib/vcsn/algos/derive.cc: Implement it.

* bin/vcsn-derived-terms.cc: New.
* bin/local.mk: Compile it.

* tests/tafkit/derive.dir/e1-dt.gv: New.
* tests/tafkit/local.mk: Adjust.
* tests/tafkit/derive.chk: Check derived-terms.

Currently it share the one of constant-term, but since more algorithms
will live in derive.hh, split it not.

* lib/vcsn/algos/constant-term.cc (derive): Move to...
* lib/vcsn/algos/derive.cc: New.
* lib/vcsn/local.mk: here.

New algorithm, which provides the user with a means to control the
order of state elimination.  There are still issues to treat:
- we need the actual state number, not the one displayed by dot
- we need better control on the state number
- we need to provide a dyn:: access to state selection

* vcsn/algos/aut_to_exp.hh (dyn::eliminate_state): New.
* lib/vcsn/algos/aut-to-exp.cc, vcsn/ctx/instantiate.hh,
* vcsn/dyn/algos.hh: Adjust.
* bin/vcsn-eliminate-state.cc: New.
* bin/local.mk, common.mk: Adjust.

* tests/tafkit/aut-to-exp.dir/lao-elim-0.gv,
* tests/tafkit/aut-to-exp.dir/lao-elim-1.gv,
* tests/tafkit/aut-to-exp.dir/lao-elim-2.gv,
* tests/tafkit/aut-to-exp.dir/lao.gv: New test files.
* tests/tafkit/aut-to-exp.chk: Use them.

* configure.ac: Let's go for "vaucanson 2b.XX" as numbering scheme for betas.
* ports/Portfile: New.

Instead of explicitly hiding the symbols that must not be exported, do
the converse: require that by default symbols are _not_ exported, but
in turn explicitly flag what symbols must be exported.

The vcsn/misc/export.hh file provide a symbol, LIBVCSN_API, which must
be explicitly used to flag symbols to export, _and_ used to import
them in callers.

Of course all the dyn/algos.hh signature must be exported, which is
easy to write thanks to the namespace attributes syntax from GCC.
Likewise for algorithms in vcsn/misc that are actually implemented in
*.cc files.

What was not really expected is that explicit class template
instantiations must also be flagged with LIBVCSN_API (see for instance
its uses in vcsn/ctx/instantiate.hh).

On my machine, using 4.8, I get a significant reduction of the library
sizes (roughly 20%):

  920K -> 732K liblal_char_b.0.dylib
  1,1M -> 788K liblal_char_br.0.dylib
  840K -> 660K liblal_char_q.0.dylib
  828K -> 656K liblal_char_r.0.dylib
  840K -> 660K liblal_char_z.0.dylib
  848K -> 664K liblal_char_zmin.0.dylib
  1,1M -> 800K liblal_char_zr.0.dylib
  1,2M -> 912K liblal_char_zrr.0.dylib
  864K -> 680K liblan_char_b.0.dylib
  844K -> 660K liblan_char_r.0.dylib
  832K -> 648K liblan_char_z.0.dylib
  1,1M -> 812K liblan_char_zr.0.dylib
  816K -> 628K liblao_br.0.dylib
  572K -> 464K liblao_z.0.dylib
  768K -> 604K liblaw_char_b.0.dylib
  964K -> 740K liblaw_char_br.0.dylib
  796K -> 628K liblaw_char_q.0.dylib
  780K -> 620K liblaw_char_r.0.dylib
  768K -> 608K liblaw_char_z.0.dylib
  780K -> 616K liblaw_char_zmin.0.dylib
  968K -> 748K liblaw_char_zr.0.dylib
  1,1M -> 848K liblaw_char_zrr.0.dylib
  764K -> 720K libvcsn.0.dylib

   20M -> 16M  TOTAL

There is no compilation speed-up (or I failed to measure it).  However
we _might_ benefit from a small speed-up when launching TAF-Kit (as the
dynamic loader as a simpler job to do):

Before:
  8.53s (0.08s+8.45s): ladybird 21 | determinize -O null
  0.17s (0.08s+0.09s): de-bruijn 20 | evaluate -O null -f - aaaaaaaaaaaaaaaaaaaaa
 12.84s (1.03s+11.81s): thompson -C "lan_char(a)_b" -Ee "a?{2000}" | proper -O null
 23.30s (10.37s+12.93s): standard -Ee "(a+b+c+d)?{100}" | aut-to-exp -O null
  7.27s (0.07s+7.20s): standard -C "lal_char(ab)_z" -Ee "(a+b)*b(<2>a+<2>b)*" | power -O null -f- 10
  0.07s (0.07s): standard -E -e "(a?){70}" -o a70.gv
 13.63s (13.63s): product -O null -f a70.gv a70.gv

After:
  7.94s (0.05s+7.89s): ladybird 21 | determinize -O null
  0.10s (0.05s+0.05s): de-bruijn 20 | evaluate -O null -f - aaaaaaaaaaaaaaaaaaaaa
 12.04s (0.93s+11.11s): thompson -C "lan_char(a)_b" -Ee "a?{2000}" | proper -O null
 22.69s (10.16s+12.53s): standard -Ee "(a+b+c+d)?{100}" | aut-to-exp -O null
  6.72s (0.05s+6.67s): standard -C "lal_char(ab)_z" -Ee "(a+b)*b(<2>a+<2>b)*" | power -O null -f- 10
  0.05s (0.05s): standard -E -e "(a?){70}" -o a70.gv
 13.01s (13.01s): product -O null -f a70.gv a70.gv

Granted, the differences are small and might be insignificant.

* configure.ac: Check if the compiler supports -fvisibility=hidden.
* lib/local.mk, lib/vcsn/local.mk: Use it to compile libvcsn and
the context libraries.

* lib/ctx/libctx.cc: Enable the exportation of the symbols.

* vcsn/misc/export.hh: New.
* vcsn/local.mk: Adjust.

* lib/vcsn/dot/driver.hh, lib/vcsn/rat/driver.hh,
* lib/vcsn/rat/read.hh, vcsn/core/mutable_automaton.hh,
* vcsn/ctx/instantiate.hh, vcsn/dyn/algos.hh, vcsn/dyn/automaton.hh,
* vcsn/dyn/context.hh, vcsn/dyn/fwd.hh, vcsn/dyn/ratexp.hh,
* vcsn/dyn/weight.hh, vcsn/misc/escape.hh, vcsn/misc/stream.hh:
Flag the symbols that need to be exported.

* vcsn/misc/attributes.hh: Sort.
Improve layout.
(ATTRIBUTE_DLLEXPORT, ATTRIBUTE_DLLIMPORT): New.

* vcsn/algos/determinize.hh, vcsn/algos/product.hh (origins_t)
(print): New.
Use them.

Decided during Vaucanson meeting 2013-06-25.

No visible impact of the additional test in the score.

  8.35s (0.08s+8.27s): ladybird 21 | determinize -O null
  0.14s (0.07s+0.07s): de-bruijn 20 | evaluate -O null -f - aaaaaaaaaaaaaaaaaaaaa
 11.66s (0.86s+10.80s): thompson -C "lan_char(a)_b" -Ee "a?{2000}" | proper -O null
 21.32s (9.54s+11.78s): standard -Ee "(a+b+c+d)?{100}" | aut-to-exp -O null
  6.73s (0.07s+6.66s): standard -C "lal_char(ab)_z" -Ee "(a+b)*b(<2>a+<2>b)*" | power -O null -f- 10
  0.07s (0.07s): standard -E -e "(a?){70}" -o a70.gv
 13.33s (13.33s): product -O null -f a70.gv a70.gv

* vcsn/algos/determinize.hh (determinize): Accept a argument specifying
whether to force the result to be complete.
By the way, when VCSN_DETERMINIZE is set, output the origins.
* bin/vcsn-determinize.cc, lib/vcsn/algos/determinize.cc,
* vcsn/algos/are-equivalent.hh, vcsn/ctx/lal_char_b.hh,
* vcsn/dyn/algos.hh: Adjust.

Adjust the TAF-Kit test suite to check the origins, now that they are
generated in the output.

* tests/tafkit/determinize.dir/a-det-0.gv: New.
* tests/tafkit/determinize.dir/a-det.gv: Rename as...
* tests/tafkit/determinize.dir/a-det-1.gv: this.
* tests/tafkit/determinize.dir/empty-det-0.gv: New.
* tests/tafkit/determinize.dir/empty-det.gv: Rename as...
* tests/tafkit/determinize.dir/empty-det-1.gv: this.
* tests/tafkit/determinize.dir/epsilon-det-0.gv: New.
* tests/tafkit/determinize.dir/epsilon-det.gv: Rename as...
* tests/tafkit/determinize.dir/epsilon-det-1.gv: this.
* tests/tafkit/determinize.chk, tests/tafkit/local.mk: Adjust.

* tests/tafkit/transpose.chk: Adjust: require a complete result.

Now that the TAF-Kit test suite checks the origins, and the fact that
the result is complete or not, the "unit" test for determinize is
useless.

* tests/unit/determinize.cc: Remove.
* tests/unit/determinize.chk: Remove.
* tests/unit/local.mk: Adjust.

* NEWS.txt: Advertise.

fixup! determinize: no longer force the result to be complete

* vcsn/algos/derive.hh: here.
And use it.

* lib/vcsn/dot/scan.ll, lib/vcsn/rat/scan.ll: here.

make_automaton_editor is somewhat special: it is parameterized by an
Automaton type, but takes a context name as argument.

Turn this into something more consistent with the other algorithms.

* vcsn/algos/edit-automaton.hh (make_automaton_editor): Take a context
as parameter type.
* vcsn/ctx/instantiate.hh: Simplify accordingly.

See previous commit for details.

* vcsn/algos/dot.hh, vcsn/algos/is-valid.hh, vcsn/algos/lift.hh:
Name the bridges after the dyn type.
* vcsn/ctx/instantiate.hh: Simplify.

There is really no reason to force the bridges to be overloaded (e.g.,
using "standard" for two routines, one with automaton as input, and
the other with "ratexp" as input), as these functions are never seen
by the user.

However, it does make the instantiatiation and use by the system more
complex (e.g., the REGISTER macros cannot be used).

Relax these constraints, and use the same name for the bridge, and for
its associated typedef (standard -> standard_t, standard_exp ->
standard_exp_t).

* vcsn/algos/edit-automaton.hh, vcsn/factory/ladybird.hh: Comment change.
* vcsn/algos/print.hh, vcsn/algos/standard.hh: Use the same names for
bridges and types.
* vcsn/ctx/lan_char_b.hh, vcsn/ctx/instantiate.hh: Simplify.

* vcsn/algos/enumerate.hh: Remove a dead FIXME.
* vcsn/factory/double-ring.hh: Use state_t instead of unsigned.

* lib/vcsn/algos/left-mult.cc (vname): Move to...
* lib/vcsn/algos/registry.hh (vname): here, and generalize it to any type.

* lib/vcsn/algos/are-equivalent.cc, lib/vcsn/algos/concatenate.cc,
* lib/vcsn/algos/left-mult.cc, lib/vcsn/algos/product.cc,
* lib/vcsn/algos/sum.cc, lib/vcsn/algos/union.cc: Use it.
Interestingly some of these calls were wrong, and using "rhs, rhs" types
instead of "lhs, rhs".

* vcsn/weights/entryset.hh: Remove.
* vcsn/local.mk: Adjust.

* vcsn/algos/edit-automaton.hh (entryset_): Replace this entryset with...
(ps_): this polynomialset.
And name changes for consistency.

polynomialset was hard-wired for classes that have a genset, which
rules out oneset (and tupleset etc.).  Just use the labelset's conv.
This also requires to apply to polynomialset changes made on entryset
by commit e5008147.

* vcsn/weights/polynomialset.hh (conv): here.

* vcsn/weights/polynomialset.hh: polynomials are no longer about
word_t only, but really for label_t.