maintaining (ugly and error-prone) hash tables.

That change revealed an untriggered bug in EvalClock.check_args: it was wrong
to add in subst the substitutions made of the parameters and the arguments (it
is enough to unify the clocks of the pars and of the args). For instance, consider
the node (in should_work/clock/clock.lus)

   node clock5(x :  bool; y:  bool when x; z: bool when y)

and the call

  z2 = clock5(a, b when a, c when e);

I was adding y/b in the subst, which was wrong.

Other minor changes:
 - move const_to_val_eff from Eff to UnifyClock.
 - GetEff.translate_val_exp now returns a substitution, in order to be able
   to unify clock vars and propagate the resulting substitution.

info is attached to all  val_exp. Moreover, this allow to perfrom all
the  type checking  in  EvalType.f, instead  of  doin part  of it  in
GetEff.translate_val_exp.

maintaining (ugly and error-prone) hash tables.

constants that  are not  on the base  clock, add the  suitable "when"
statement if it was not present  in the source code (the lic (and the
ec/lv4) backend  does not  perform such kind  of clock  inference for
constants.

instead of silently generate  (syntactically) wrong lustre v4 (or ec).

(versus raising an internal error...).

the moment, I just don't generate such operator anymore (i generate the
polymorphic version Lustre::lt instead). I don't really know what is the
better.

fixing the v4 backend : print "+" instead of "Lustre::plus" (and ditto for the others).

In  short, the  rationale for  this change,  is that  it is  having a
recursive node_exp is
- useless,
- too complicated,
- wrong w.r.t. nesting iterator calls

In long:

- It is useless because, at the  Eff level, a node cannot call itself
  via one  of its  static arg (which  was where the  recursivity came
  from).

- and indeed, it  is much simpler to consider that  a static arg node
  can only be ident.long that identifies a node alias.

  This  means of  course, that  nested iterators  have  been unnested
  before,   inventing  alias   node  names   along  the   way...  And
  polymorphism makes thing difficult once again.

- But the  *big* problem  with a recursive  node_exp is that  it make
  things very  complicated to (lic)dump nested  iterator call because
  of polymorphism!   Actually, it  even makes thing  complicated when
  the iterators were themselves not nested in the source code !

  Some ugly  things were done in  LicDump to unnest  those calls when
  printing node_exp.   But this uglyness  have a price:  tricky code,
  and bugs! Indeed, nested iterators calls were wong for example when
  using  the --inline-iterator  mode (but  i would  not  be surprised
  that is wrong in other cases...).

Hence,  LicDump  is  simpler,  but  of course  LazyCompiler  is  more
complicated.  But  this  is  reasonable:  a  pretty-printer  is  not
supposed to be complicated.

To  do  that, we  centralized  the ident  name  creation  into a  new
dedicated module Name.

+ fix should_work/NONREG/param_node.lus, we contained a combinational loop.

The rationale for this change is that this is necessary for constants
appearing static  arg to  be replaced. One  problem is that  they are
handled in exactly  the same way as top-level  constants, which could
be left  un-expanded. I could make  something in order  to not expand
those top-levelconstants, but is it worth the trouble ?

To do that,  I have changed sligthly the  representation of Eff.ARRAY
(the elements  are now  attached to the  constructor itself,  and the
operands  is empty)  and  the one  of  array constant  (we attach  to
Array_const_eff  a  list  instead  of  an  array,  for  the  sake  of
homogeneity) with what is done in Eff.val_eff.

(--inline-iterators) to activate it.

nb : do not inline completely nested iterator calls (yet, cf TODO).

change the parser and accept only clock expressions after a when.

A clock is now made of 2 idents: one for the clock constructor, and
one for the clock variable.

simple, and also avoids  3 shift/reduce conflicts. It also make
the pragma stuff more homogeneous  and simple to explain in the
manual.

Moreover, pramgma  are no  more attached to  AST nodes,  but to
lexemes.

Note that by polymorphic here I mean *really* polymorphic, since they
can handle tuples, which is no the case of, let's say, "<".

splitting expressions.

as many  new local  variables as necessary  so that an  expression is
made at most of one operator.

The rational  for that  is to obtain  a lic  code that is  trivial to
clock check (nested node calls, for example, make it less simple).

The old behavior can still be obtained using --keep-nested-calls.

During  that  change,   I  realised  that  I  did   not  clock  check
asserts. Hence, I have also added this check.

test/should_work/Pascal/t1.lus

+ typos

Add some tests with various kinds of expr that we can have after a when.

Not yet implemented (assert false): iterators, struct

Add a UnifyClock module, and rename Unify into UnifyType.

nb : a lot of test are now broken, because
   - the clock checking is now plugged ;-)
   - iterators, struct are not yet implemented

are defined exactly once.

This  algo is  naive and does not  work with  slices of  slices with
negative steps.  I commit it  before trying a  new one to  get better
non-reg test.

Fix some non-reg test in the should_work dir that this new check revealed.

It  does not  work, I  have changed  my mind  to use  a  more general
clocking algorithm based on  unification. I commit that current state
just in case I change my mind again...

the parameters  input types are  compatible with Unify.f.   Apply (if
necessary) the resulting substitution to the output parameters.

Also,  when checking array  concat, try  to unify  the types  of both
arrays instead of just checking their equality.

This of course triggers some errors in the non-reg tests.  One of
this error is due to a bug in the parser, where the list of parameter
was inversed twice. I've also fixed that in this change.