| Age | Commit message (Collapse) | Author |
|---|
|
duplicates in kernel side effects. They were chosen according to an equality
that was quite irrelevant, and as expected this patch did not break the
test-suite.
|
|
|
|
|
|
|
|
|
|
optimization purposes. I replaced their use with the under-approximation
of pointer equality.
|
|
This meas that the list of future_constraints in safe_env is empty,
meaning that nothing was delayed.
|
|
Make this module deal only with opaque proofs.
Make discharging/substitution invariant more explicit via a third constructor.
|
|
When you resume the compilation of a .vi file, you want to
avoid collisions on fresh names.
|
|
|
|
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).
|
|
subst1 case.
|
|
|
|
1. Only apply last Zupdates
2. Better smartmap with state.
|
|
|
|
|
|
I restored this in the kernel, and added it to the checker. There is one last
source of non-uniformity, which is the Sort case in the checker (was not
present in the kernel). I don't know what this case covers, so it should be
reviewed.
|
|
|
|
Thanks again Maximes.
This time the C value was stored in the env_(named|rel)_val of
the environment
|
|
Kudos to Maximes for finding the culprit in no time!
Values of type 'Pre_env.key' store in the OCaml state the 'address'
of an already evaluated constant in the VM's C state. Such values
are not sent to work processes. The worker is going to re-evaluate the
constant, but just once, since the cache is cleared only when the env is
marshalled (via ephemerons).
|
|
Proof using can be followed by:
- All : all variables
- Type : all variables occurring in the type
- expr:
- (a b .. c) : set
- expr + expr : set union
- expr - expr : set difference
- -expr : set complement (All - expr)
Exceptions:
- a singleton set can be written without parentheses. This also allows
the implementation of named sets sharing the same name space of
section hyps ans write
- bla - x : where bla is defined as (a b .. x y) elsewhere.
- if expr is just a set, then parentheses can be omitted
This module also implements some AI to tell the user how he could
decorate "Proof" with a "using BLA" clause.
Finally, one can Set Default Proof Using "str" to any string that is
used whenever the "using ..." part is missing. The coding of this
sucks a little since it is the parser that applies the default.
|
|
|
|
-async-proofs off
the system behaves as in 8.4
-async-proofs lazy
proofs are delayed (when possible) but never processed in parallel
-async-proofs on
proofs are processed in parallel (when possible). The number of
workers is 1, can be changed with -async-proofs-j. Extra options to
the worker process can be given with -async-proofs-worker-flags.
The default for batch compilation used to be "lazy", now it is "off".
The "lazy" default was there to test the machinery, but it makes very
little sense in a batch scenario. If you process things sequentially,
you'd better do them immediately instead of accumulating everything in
memory until the end of the file and only then force all lazy computations.
The default for -ideslave was and still is "on". It becomes dynamically
"lazy" on a per task (proof) basis if the worker dies badly.
Note that by passing "-async-proofs on" to coqc one can produce a .vo
exploiting multiple workers. But this is rarely profitable given
that master-to-worker communication is inefficient (i.e. it really
depends on the size of proofs v.s. size of system state).
|
|
File format:
The .vo file format changed:
- after the magic number there are 3 segments. A segment is made of 3
components: bynary int, an ocaml value, a digest. The binary int
is the position of the digest, so that one can skip the value without
unmarshalling it
- the first segment is the library, as before
- the second segment is the STM task list
- the third segment is the opaque table, as before
A .vo file has a complete opaque table (all proof terms are there).
A .vi file follows the same format of a .vo file, but some entries
in the opaque table are missing. A proof task is stocked instead.
Utilities:
coqc: option -quick generates a .vi insted of a .vo
coq_makefile: target quick to generate all .vi
coqdep: generate deps for .vi files too
votour: can browse .vi files too, the first question is which segment
should be read
coqchk: rejects .vi files
|
|
This has nothing to do with the kernel itself, but it is
the place where this piece of data is inferred.
|
|
Let proof terms are stocked in the named_context that is used directly
everywhere, hence there is no way to stock a Future proof term there.
|
|
Experimental. Turned out to be much harder to implement than I thought. The main
issue is that the reification in the native compiler and the VM is not quite
untyped. Indeed, type annotations for lambdas have to be reconstructed. Hence,
when reifying an application u = t a1 ... an, the type of t has to be known or
reconstructed. It is always possible to do so in plain CIC, when u is in normal
form and its type is known. However, with partial terms this may no longer be
the case, as in: ?1 a1 ... an. So we also compile and evaluate the type of
evars and metas.
This still has to be tested more extensively, but the correction of the kernel
native conversion (on terms without evars or metas) should not be impacted.
Much of this could be reused for the VM.
|
|
|
|
|
|
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.
|
|
To make this possible the state id has to reach the kernel.
Hence definition_entry has an extra field, and many files had
to be fixed.
|
|
test-suite pass.
|
|
* vars.mli was mentionning Term instead of Constr, leading to a dep cycle
* Having a file named toplevel/toplevel.ml isn't a good idea when we also
have a toplevel/toplevel.mllib that ought to produce a toplevel.cma.
We rename toplevel.ml into Coqloop.ml
* Extra cleanup of toplevel.mllib :
- Ppextra isn't anywhere around (?!)
- Ppvernac was mentionned twice, and rather belongs to printing.mllib anyway
- Vernacexpr is a .mli and shouldn't appear in an .mllib
* During the link, printing.cma now comes after parsing.cma (Ppvernac uses Egramml)
* A few extra -threads in ocamlbuild files (btw, it's a bit sad to need -thread
for coqchk).
|
|
function is sufficient to skip the undefined variables.
|
|
I chose n to be 10000 iterations. It might be big, but a slave, to check
for a termination request, has to pass the ball to the thread that
sends "regularly" Ticks to the master process. Thread.yield is a
system call, so we have to do it very rarely.
|
|
|
|
|
|
Contrarily to my machine results, it seems that it tore down the performance of
Coq on benchmarks.
git-svn-id: svn+ssh://scm.gforge.inria.fr/svn/coq/trunk@17091 85f007b7-540e-0410-9357-904b9bb8a0f7
|
|
This was quite a severe performance bottleneck.
Ideally, this data structure should be put into contexts, but the relevant
type is transparent... For now, we stick to this inelegant workaround.
git-svn-id: svn+ssh://scm.gforge.inria.fr/svn/coq/trunk@17086 85f007b7-540e-0410-9357-904b9bb8a0f7
|
|
git-svn-id: svn+ssh://scm.gforge.inria.fr/svn/coq/trunk@17065 85f007b7-540e-0410-9357-904b9bb8a0f7
|
|
cases, which are precisely term manipulation.
git-svn-id: svn+ssh://scm.gforge.inria.fr/svn/coq/trunk@17054 85f007b7-540e-0410-9357-904b9bb8a0f7
|
|
git-svn-id: svn+ssh://scm.gforge.inria.fr/svn/coq/trunk@17049 85f007b7-540e-0410-9357-904b9bb8a0f7
|
|
git-svn-id: svn+ssh://scm.gforge.inria.fr/svn/coq/trunk@16966 85f007b7-540e-0410-9357-904b9bb8a0f7
|
|
This commit should fix the contrib ProjectiveGeometry
This update_delta_resolver is in fact a sequential composition
(resolver1 then resolver2). The earlier version was strangely
favoring the bindings coming from resolver2 over the bindings
coming from (resolver1 chained with resolver2). So any inlining
information stored in resolver1 could be discarded savagely.
Apparently, this situation wasn't occurring in practice until
recently, when I started to do lots of Mod_subst.join for
improving the size of modular libobjects in the vo's.
So, when combining two resolvers now :
- Inline(_,None) is the weakest information, it's just a
declaration that we intend to inline this kn someday if possible
(i.e. when we'll have a transparent implementation for it).
This kind of Inline is only relevant inside a module type.
- Equiv(_) should only appear in modules (after some Include)
so it should be ok if it takes precedence over any Inline(_,None)
remaining in the other resolver.
- Inline(_,Some _) is there after functor application
(cf inline_delta_resolver) : we've done the inlining,
so we don't care anymore about other Equiv(_) or Inline(_,None)
informations about this kn, since we have anyway a new body
for it.
git-svn-id: svn+ssh://scm.gforge.inria.fr/svn/coq/trunk@16965 85f007b7-540e-0410-9357-904b9bb8a0f7
|
|
equality, maybe impeding hashconsing.
git-svn-id: svn+ssh://scm.gforge.inria.fr/svn/coq/trunk@16964 85f007b7-540e-0410-9357-904b9bb8a0f7
|
|
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
|
|
But for vm, the kernel should be functional now
git-svn-id: svn+ssh://scm.gforge.inria.fr/svn/coq/trunk@16961 85f007b7-540e-0410-9357-904b9bb8a0f7
|
|
git-svn-id: svn+ssh://scm.gforge.inria.fr/svn/coq/trunk@16959 85f007b7-540e-0410-9357-904b9bb8a0f7
|
