| Age | Commit message (Collapse) | Author |
|---|
|
To obtain a.vo one can now:
1) coqtop -quick -compile a
2) coqtop -vi2vo a.vi
To make that possible the .vo structure has been complicated. It is now
made of 5 segments.
| vo | vi | vi2vo | contents
--------------+------+-----+-------+------------------------------------
lib | Yes | Yes | Yes | libstack (modules, notations,...)
opauqe_univs | No | Yes | Yes | constraints coming from opaque proofs
discharge | No | Yes | No | data needed to close sections
tasks | No | Yes | No | STM tasks to produce proof terms
opaque_proofs | Yes | Yes | Yes | proof terms
--------------+------+-----+-------+------------------------------------
This means one can load only the strictly necessay parts. Usually one
does not load the tasks segment of a .vi nor the opaque_proof segment of
a .vo, unless one is turning a .vi into a .vo, in which case he load
all the segments.
Optional segments are marshalled as None. But for lib, all segments
are Array.t of:
| type
--------------+---------------------------------------------------------
lib | a list of Libobject.obj (n'importe quoi)
opauqe_univs | Univ.consraints Future.computation
discharge | what Cooking.cook_constr needs
tasks | Stm.tasks (a task is system_state * vernacexpr list)
opaque_proofs | Term.constr Future.computation
--------------+------+-----+-------+------------------------------------
Invariant: all Future.computation in a vo file (obtained by a vi2vo
compilation or not) have been terminated with Future.join (or
Future.sink). This means they are values (inside a box).
This invariant does not hold for vi files. E.g. opauqe_proofs can be
dangling Future.computation (i.e. NotHere exception). The vi2vo
compilation step will replace them by true values.
Rationale for opaque_univs: in the vi2vo transformation we want to reuse
the lib segment. Hence the missing pieces have to be put on the side,
not inside. Opaque proof terms are already in a separte segment.
Universe constraints are not, hence the new opauqe_univs segment. Such
segment, if present in a .vo file, is always loaded, and
Declare.open_constant will add to the environment the constraints stored
there. For regular constants this is not necessay since the constraints
are already in their enclosing module (and also in the constant_body).
With vi2vo the constraints coming from the proof are not in the
constant_body (hence not in the enclosing module) but there and are
added to the environment explicitly by Declare.open_constant.
Rationale for discharge: vi2vo produces a proof term in its original
context (in the middle of a section). Then it has to discharge the
object. This segment contains the data that is needed in order to do
so. It is morally the input that Lib.close_section passes to Cooking
(via the insane rewinding of libstack, GlobalRecipe, etc chain).
Checksums: the checksum of .vi and a .vo obtain from it is the same.
This means that if if b.vo has been compiled using a.vi, and then
a.vi is compiled into a.vo, Require Import b works (and recursively
loads a.vo).
|
|
The sharing introduced by this commit is now correct, since a reference used by
the native compiler has been removed from constant_body.
This reverts commit 413f5fcd4bf581ff3ea4694c193d637589c7d54f.
|
|
For now, this reference (renamed to link_info) has been moved to the
environment (for constants and inductive types). But this is only a first step
towards making the native compiler more functional.
|
|
Also, the future chain that reaches the kernel is greedy.
If the user executes step by step, then the error is raised immediately.
|
|
This optimization was undone because the kernel type checking was
not a pure functions (it was accessing the conv_oracle state imperatively).
Now that the conv_oracle state is part of env, the optimization can be
restored. This was the cause of the increase in memory consumption, since
it was forcing to keep a copy of the system state for every proof, even the
ones that are not delayed/delegated to slaves.
git-svn-id: svn+ssh://scm.gforge.inria.fr/svn/coq/trunk@16963 85f007b7-540e-0410-9357-904b9bb8a0f7
|
|
A future always carries a fix_exn with it: a function that enriches
an exception with the state in which the error occurs and also a safe
state close to it where one could backtrack.
A future can be in two states: Ongoing or Finished.
The latter state is obtained by Future.join and after that the future
can be safely marshalled.
An Ongoing future can be marshalled, but its value is lost. This makes
it possible to send the environment to a slave process without
pre-processing it to drop all unfinished proofs (they are dropped
automatically in some sense).
git-svn-id: svn+ssh://scm.gforge.inria.fr/svn/coq/trunk@16892 85f007b7-540e-0410-9357-904b9bb8a0f7
|
|
git-svn-id: svn+ssh://scm.gforge.inria.fr/svn/coq/trunk@16806 85f007b7-540e-0410-9357-904b9bb8a0f7
|
|
If a constant is defined as transparent, not only its side effects
(opaque sub proofs as in abstract, and transparent ind schemes)
are declared globally, but the ones that are schemes are also declared
as such.
The only sub optimal thing is that the code handling in a special
way the side effects of transparent constants is in declare.ml that
does not see ind_tables.ml, hence a forward ref to a function is used.
IMO, ind_tables has no reason to stay in toplevel/.
git-svn-id: svn+ssh://scm.gforge.inria.fr/svn/coq/trunk@16747 85f007b7-540e-0410-9357-904b9bb8a0f7
|
|
It seems that it's critical for the native compiler that
a fresh (ref NotLinked) is created during substitution.
git-svn-id: svn+ssh://scm.gforge.inria.fr/svn/coq/trunk@16719 85f007b7-540e-0410-9357-904b9bb8a0f7
|
|
We try harder to preserve pointer equality when substituting.
This will probably have little effect (for instance the
constr_substituted are anyway _never_ substituted in place),
but it cannot harm.
Two particular cases:
- we try to share (and maintain shared) mind_user_lc and mind_nf_lc
- we try to share (and maintain shared) mod_expr and mod_type
TODO: check that native compiler is still ok, since we might have
less (ref NotLinked) now. Having references in constant_body and co
is anyway a Very Bad Idea (TM).
git-svn-id: svn+ssh://scm.gforge.inria.fr/svn/coq/trunk@16711 85f007b7-540e-0410-9357-904b9bb8a0f7
|
|
Without this, the stdlib vo files are +30% larger
git-svn-id: svn+ssh://scm.gforge.inria.fr/svn/coq/trunk@16707 85f007b7-540e-0410-9357-904b9bb8a0f7
|
|
Non-opaque-constant's side effects are processed before
the constant enters the kernel and global constants are
generated for them (as before, but not by side effect in
the middle of the proof construction).
This makes sense because proofs ending with Defined have
to be run immediately, so the list of side effects is
immediately available.
These side effects are type checked again.
To fix that the idea of kernel signatures could be employed.
git-svn-id: svn+ssh://scm.gforge.inria.fr/svn/coq/trunk@16702 85f007b7-540e-0410-9357-904b9bb8a0f7
|
|
It can be:
`Yes Full data, in a state that can be marshalled
`No Full data, good for Undo only
`Shallow Partial data, marshallable, good for slave processes
git-svn-id: svn+ssh://scm.gforge.inria.fr/svn/coq/trunk@16682 85f007b7-540e-0410-9357-904b9bb8a0f7
|
|
The process_transaction function adds a new edge to the Dag without
executing the transaction (when possible).
The observe id function runs the transactions necessary to reach to the
state id. Transaction being on a merged branch are not executed but
stored into a future.
The finish function calls observe on the tip of the current branch.
Imperative modifications to the environment made by some tactics are
now explicitly declared by the tactic and modeled as let-in/beta-redexes
at the root of the proof term. An example is the abstract tactic.
This is the work described in the Coq Workshop 2012 paper.
Coq is compile with thread support from now on.
git-svn-id: svn+ssh://scm.gforge.inria.fr/svn/coq/trunk@16674 85f007b7-540e-0410-9357-904b9bb8a0f7
|
|
git-svn-id: svn+ssh://scm.gforge.inria.fr/svn/coq/trunk@16463 85f007b7-540e-0410-9357-904b9bb8a0f7
|
|
Get rid of the LightenLibrary hack : no more last-minute
collect of opaque terms and Obj.magic tricks. Instead, we
make coqc accumulate the opaque terms as soon as constant_bodies
are created outside sections. In these cases, the opaque
terms are placed in a special table, and some (DirPath.t * int)
are used as indexes in constant_body. In an interactive session,
the local opaque terms stay directly stored in the constant_body.
The structure of .vo file stays similar : magic number, regular
library structure, digest of the first part, array of opaque terms.
In addition, we now have a final checksum for checking the
integrity of the whole .vo file. The other difference is that
lazy_constr aren't changed into int indexes in .vo files, but are
now coded as (substitution list * DirPath.t * int). In particular
this approach allows to refer to opaque terms from another
library. This (and accumulating substitutions in lazy_constr)
seems to greatly help decreasing the size of opaque tables :
-20% of vo size on the standard library :-). The compilation times
are slightly better, but that can be statistic noise.
The -force-load-proofs isn't active anymore : it behaves now
just like -lazy-load-proofs. The -dont-load-proofs mode has
slightly changed : opaque terms aren't seen as axioms anymore,
but accessing their bodies will raise an error.
Btw, API change : Declareops.body_of_constant now produces directly
a constr option instead of a constr_substituted option
git-svn-id: svn+ssh://scm.gforge.inria.fr/svn/coq/trunk@16382 85f007b7-540e-0410-9357-904b9bb8a0f7
|
|
- constr_substituted and lazy_constr are now in a dedicated kernel/lazyconstr.ml
- the functions that were in declarations.ml (mostly substitution utilities
and hashcons) are now in kernel/declareops.ml
git-svn-id: svn+ssh://scm.gforge.inria.fr/svn/coq/trunk@16250 85f007b7-540e-0410-9357-904b9bb8a0f7
|
