|
This is a fairly large commit (around 140 files and 7000 lines of code
impacted), it will cause some troubles for sure (I've listed the know
regressions below, there is bound to be more).
At this state of developpement it brings few features to the user, as
the old tactics were
ported with no change. Changes are on the side of the developer mostly.
Here comes a list of the major changes. I will stay brief, but the code
is hopefully well documented so that it is reasonably easy to infer the
details from it.
Feature developer-side:
* Primitives for a "real" refine tactic (generating a goal for each
evar).
* Abstract type of tactics, goals and proofs
* Tactics can act on several goals (formally all the focused goals). An
interesting consequence of this is that the tactical (. ; [ . | ... ])
can be separated in two
tacticals (. ; .) and ( [ . | ... ] ) (although there is a conflict for
this particular syntax). We can also imagine a tactic to reorder the
goals.
* Possibility for a tactic to pass a value to following tactics (a
typical example is
an intro function which tells the following tactics which name it
introduced).
* backtracking primitives for tactics (it is now possible to implement a
tactical '+'
with (a+b);c equivalent to (a;c+b;c) (itself equivalent to
(a;c||b;c)). This is a valuable
tool to implement tactics like "auto" without nowing of the
implementation of tactics.
* A notion of proof modes, which allows to dynamically change the parser
for tactics. It is controlled at user level with the keywords Set
Default Proof Mode (this is the proof mode which is loaded at the start
of each proof) and Proof Mode (switches the proof mode of the current
proof) to control them.
* A new primitive Evd.fold_undefined which operates like an Evd.fold,
except it only goes through the evars whose body is Evar_empty. This is
a common operation throughout the code,
some of the fold-and-test-if-empty occurences have been replaced by
fold_undefined. For now,
it is only implemented as a fold-and-test, but we expect to have some
optimisations coming some day, as there can be a lot of evars in an
evar_map with this new implementation (I've observed a couple of
thousands), whereas there are rarely more than a dozen undefined ones.
Folding being a linear operation, this might result in a significant
speed-up.
* The declarative mode has been moved into the plugins. This is made
possible by the proof mode feature. I tried to document it so that it
can serve as a tutorial for a tactic mode plugin.
Features user-side:
* Unfocus does not go back to the root of the proof if several Focus-s
have been performed.
It only goes back to the point where it was last focused.
* experimental (non-documented) support of keywords
BeginSubproof/EndSubproof:
BeginSubproof focuses on first goal, one can unfocus only with
EndSubproof, and only
if the proof is completed for that goal.
* experimental (non-documented) support for bullets ('+', '-' and '*')
they act as hierarchical BeginSubproof/EndSubproof:
First time one uses '+' (for instance) it focuses on first goal, when
the subproof is
completed, one can use '+' again which unfocuses and focuses on next
first goal.
Meanwhile, one cas use '*' (for instance) to focus more deeply.
Known regressions:
* The xml plugin had some functions related to proof trees. As the
structure of proof changed significantly, they do not work anymore.
* I do not know how to implement info or show script in this new engine.
Actually I don't even know what they were suppose to actually mean in
earlier versions either. I wager they would require some calm thinking
before going back to work.
* Declarative mode not entirely working (in particular proofs by
induction need to be restored).
* A bug in the inversion tactic (observed in some contributions)
* A bug in Program (observed in some contributions)
* Minor change in the 'old' type of tactics causing some contributions
to fail.
* Compilation time takes about 10-15% longer for unknown reasons (I
suspect it might be linked to the fact that I don't perform any
reduction at QED-s, and also to some linear operations on evar_map-s
(see Evd.fold_undefined above)).
git-svn-id: svn+ssh://scm.gforge.inria.fr/svn/coq/trunk@12961 85f007b7-540e-0410-9357-904b9bb8a0f7
|
|
1- Management of the name-space in a modular development / sharing of non-logical objects.
2- Performance of atomic module operations (adding a module to the environment, subtyping ...).
1-
There are 3 module constructions which derive equalities on fields from a module to another:
Let P be a module path and foo a field of P
Module M := P.
Module M.
Include P.
...
End M.
Declare Module K : S with Module M := P.
In this 3 cases we don't want to be bothered by the duplication of names.
Of course, M.foo delta reduce to P.foo but many non-logical features of coq
do not work modulo conversion (they use eq_constr or constr_pat object).
To engender a transparent name-space (ie using P.foo or M.foo is the same thing)
we quotient the name-space by the equivalence relation on names induced by the
3 constructions above.
To implement this, the types constant and mutual_inductive are now couples of
kernel_names. The first projection correspond to the name used by the user and the second
projection to the canonical name, for example the internal name of M.foo is
(M.foo,P.foo).
So:
*************************************************************************************
* Use the eq_(con,mind,constructor,gr,egr...) function and not = on names values *
*************************************************************************************
Map and Set indexed on names are ordered on user name for the kernel side
and on canonical name outside. Thus we have sharing of notation, hints... for free
(also for a posteriori declaration of them, ex: a notation on M.foo will be
avaible on P.foo). If you want to use this, use the appropriate compare function
defined in name.ml or libnames.ml.
2-
No more time explosion (i hoppe) when using modules i have re-implemented atomic
module operations so that they are all linear in the size of the module. We also
have no more unique identifier (internal module names) for modules, it is now based
on a section_path like mechanism => we have less substitutions to perform at require,
module closing and subtyping but we pre-compute more information hence if we instanciate
several functors then we have bigger vo.
Last thing, the checker will not work well on vo(s) that contains one of the 3 constructions
above, i will work on it soon...
git-svn-id: svn+ssh://scm.gforge.inria.fr/svn/coq/trunk@12406 85f007b7-540e-0410-9357-904b9bb8a0f7
|
