| Age | Commit message (Collapse) | Author |
|---|
|
|
|
* Replacing __gnu_cxx::hash_map with std::unordered_map.
* Replacing __gnu_cxx::hash_set with std::unordered_set.
* Replacing __gnu_cxx::hash with std::hash.
* Adding missing includes.
|
|
|
|
|
|
- The options replayStream, lemmaInputChannel, lemmaOutputChannel have been removed due to their datatypes. These datatypes were previously pointers to types that were not usable from the options/ library.
- The option replayLog has been removed due to inconsistent memory management.
- SmtGlobals is a class that wraps a pointer to each of these removed options. These can each be set independently.
- There is a single SmtGlobals per SmtEngine with the lifetime of the SmtEngine.
- A pointer to this is freely given to the user of an SmtEngine to parameterize the solver after construction.
- Selected classes have been given a copy of this pointer in their constructors.
- Removed the dependence on Node from Result. Moving Result back into util/.
|
|
What to Know As a User:
A number of files have moved. Users that include files in the public API in more refined ways than using #include <cvc4.h> should consult which files have moved. Note though that some files may move again after being cleaned up. A number of small tweaks have been made to the swig interfaces that may cause issues. Please file bug reports for any problems.
The Problem:
The build order of CVC4 used to be [roughly] specified as:
options < expr < util < libcvc4 < parsers < main
Each of these had their own directories and their own Makefile.am files. With the exception of the util/ directory, each of the subdirectories built exactly one convenience library. The util/ directory additionally built a statistics library. While the order above was partially correct, the build order was more complicated as options/Makefile.am executed building the sources for expr/Makefile.am as part of its BUILT_SOURCES phase. This options/Makefile.am also build the options/h and options.cpp files in other directories. There were cyclical library dependencies between the first four above libraries. All of these aspects combined to make options extremely brittle and hard to develop. Maintaining these between clang versus gcc, and bazel versus autotools has become increasing unpredictable.
The Solution:
To address these cyclic build problems, I am simplifying the build process. Here are the main things that have to happen:
1. util/ will be split into 3 separate directories: base, util, and smt_util. Each will have their own library and Makefile.am file.
2. Dependencies for options/ will be moved into options/. If a type appears as an option, this file will be moved into options.
3. All of the old options_handlers.h files have been refactored.
4. Some files have moved from util into expr/ to resolve cycles. Some of these moves are temporary.
5. I am removing the libstatistics library.
The constraints that the CVC4 build system will eventually satisfy are:
- The include order for both the .h and .cpp files for a directory must respect the order libraries are built. For example, a file in options/ cannot include from the expr/ directory. This includes built source files such as those coming from */kinds files and */options files.
- The types definitions must also respect the build order. Forward type declarations will be allowed in exceptional, justified cases.
- The Makefile.am for a directory cannot generate a file outside of the directory it controls. (Or call another Makefile.am except through subdirectory calls.)
- One library per Makefile.am.
- No extra copies of libraries will be built for the purpose of distinguishing between external and internal visibility in libraries for building parser/ or main/ libraries and binaries. Any function used by parser/ and main/ will be labeled with CVC4_PUBLIC and be in a public API. (AFAICT, libstatistics was being built exactly to skirt this.)
The build order of CVC4 can now be [roughly] specified as
base < options < util < expr < smt_util < libcvc4 < parsers < main
The distinction between "base < options < util < expr" are currently clean. The relationship between expr and the subsequent directories/libraries are not yet clean.
More details about the directories:
base/
The new directory base/ contains the shared utilities that are absolutely crucial to starting cvc4. The list currently includes just: cvc4_assert.{h,cpp}, output.{h,cpp}, exception.{h,cpp}, and tls.{h, h.in, cpp}. These are things that are required everywhere.
options/
The options/ directory is self contained.
- It contains all of the enums that appear as options. This includes things like theory/bv/bitblast_mode.h .
- There are exactly 4 classes that handled currently using forward declarations currently to this: LogicInfo, LemmaInputChannel, LemmaOutputChannel, and CommandSequence. These will all be removed from options.
- Functionality of the options_handlers.h files has been moved into smt/smt_options_handler.h. The options library itself only uses an interface class defined in options/options_handler_interface.h. We are now using virtual dispatch to avoid using inlined functions as was previously done.
- The */options_handlers.h files have been removed.
- The generated smt/smt_options.cpp file has been be replaced by pushing the functionality that was generated into: options/options_handler_{get,set}_option_template.cpp . The non-generated functionality was moved into smt_engine.cpp.
- All of the options files have been moved from their directories into options/. This means includes like theory/arith/options.h have changed to change to options/arith_options.h .
util/
The util/ directory continues to contain core utility classes that may be used [almost] everywhere. The exception is that these are not used by options/ or base/. This includes things like rational and integer. These may not use anything in expr/ or libcvc4. A number of files have been moved out of this directory as they have cyclic dependencies graph with exprs and types. The build process up to this directory is currently clean.
expr/
The expr/ directory continues to be the home of expressions. The major change is files moving from util/ moving into expr/. The reason for this is that these files form a cycle with files in expr/.
- An example is datatype.h. This includes "expr/expr.h", "expr/type.h" while "expr/command.h" includes datatype.h.
- Another example is predicate.h. This uses expr.h and is also declared in a kinds file and thus appears in kinds.h.
- The rule of thumb is if expr/ pulls it in it needs to be independent of expr/, in which case it is in util/, or it is not, in which case it is pulled into expr/.
- Some files do not have a strong justification currently. Result, ResourceManager and SExpr can be moved back into util/ once the iostream manipulation routines are refactored out of the Node and Expr classes.
- Note the kinds files are expected to remain in the theory/ directories. These are only read in order to build sources.
- This directory is not yet clean. It contains forward references into libcvc4 such as the printer. It also makes some classes used by main/ and parser CVC4_PUBLIC.
smt_util/
The smt_util/ directory contains those utility classes which require exprs, but expr/ does not require them. These are mostly utilities for working with expressions and nodes. Examples include ite_removal.h, LemmaInputChannel and LemmaOutputChannel.
What is up next:
- A number of new #warning "TODO: ..." items have been scattered throughout the code as reminders to myself. Help with these issues is welcomed.
- The expr/ directory needs to be cleaned up in a similar to options/. Before this happens statistics needs to be cleaned up.
|
|
|
|
This commit adds tracking of Farkas coefficients to proof enabled builds in the theory of linear
real arithmetic when proofs are enabled. There could be some performance changes due to subtly
different search paths being taken.
Additional bug fixes:
- Polynomial::exactDivide did not satisfy preconditions to the Monomial constructor.
To prevent future problems, Monomials should now be made via one of the mkMonomial functions.
- Fixes a bug in SumOfInfeasibilitiesSPD::greedyConflictSubsets().
There was a way to use a row twice in the construction of the conflicts.
This was violating an assumption in the Tableau when constructing the intermediate rows.
Constraints:
- To enable proofs, all conflicts and propagations are designed to go through the Constraint system
before they are converted to externally understandable conflicts and propagations in the form
of Node.
- Constraints must now be given a reason for marking them as true that corresponds to a proof.
- Constraints should now be marked as being true by one of the impliedbyX functions.
- Each impliedByX function has an ArithProofType associated with it.
- Each call to an impliedByX function stores a context dependent ConstraintRule object
to track the proof.
- After marking the node as true the caller should either try to propagate the constraint or raise
a conflict.
- There are no more special cases for marking a node as being true when its negation has a proof
vs. when the negation does not have a proof. One must now explicitly pass in a inConflict flag
to the impliedByX (and similar functions).
For example,this is now longer both:
void setAssertedToTheTheory(TNode witness);
void setAssertedToTheTheoryWithNegationTrue(TNode witness);
There is just:
void setAssertedToTheTheory(TNode witness, bool inConflict);
|
|
|
|
master. See the CAV14 submission for an explanation of the changes to the integer solver's behavior. If compiled against the our custom extension of glpk, https://github.com/timothy-king/glpk-cut-log, this should have substantial differences in behavior. This should have moderate performance differences for linear real and integer arithmetic even if these features are disabled.
|
|
|
|
|
|
|
|
|
|
|
|
just the header comments at the top, though. Don't update to this rev if
you don't have time for a complete rebuild, and exclude this rev if you
want to see what's new across a range of commits.
(this commit was certified error- and warning-free by the test-and-commit script.)
|
|
The external interface (e.g., what's answered by ExprManager::getStatistics() and SmtEngine::getStatistics()) is a snapshot of the current statistics (rather than a reference to the actual StatisticsRegistry).
The StatisticsRegistry is now internal-only. However, it's built as a convenience library so that the parser and driver can use it too (by re-linking against it).
This is part of the ongoing effort to clean up the public interface.
(this commit was certified error- and warning-free by the test-and-commit script.)
|
|
1. changes the way options are declared (see http://church.cims.nyu.edu/wiki/Options)
2. moves module-specific options enumerations (SimplificationMode, DecisionMode, ArithUnateLemmaMode, etc.) to their own header files, also they are no longer inside the Options:: class namespace.
3. includes many SMT-LIBv2 compliance fixes, especially to (set-option..) and (get-option..)
The biggest syntactical changes (outside of adding new options) you'll notice are in accessing and setting options:
* to access an option, write (e.g.) options::unconstrainedSimp() instead of Options::current()->unconstrainedSimp.
* to determine if an option value was set by the user, check (e.g.) options::unconstrainedSimp.wasSetByUser().
* ensure that you have the option available (you have to #include the right module's options.h file, e.g. #include "theory/uf/options.h" for UF options)
*** this point is important. If you access an option and it tells you the option doesn't exist, you aren't #including the appropriate options.h header file ***
Note that if you want an option to be directly set (i.e., other than via command-line parsing or SmtEngine::setOption()), you need to mark the option :read-write in its options file (otherwise it's read-only), and you then write (e.g.) options::unconstrainedSimp.set(true).
Adding new options is incredibly simple for primitive types (int, unsigned, bool, string, double). For option settings that you need to turn into a member of an enumerated type, you write a custom "handler" for the option---this is no additional work than it was before, and there are many examples to copy from (a good one is stringToSimplificationMode() in src/smt/options_handlers.h).
Benefits of the new options system include:
1. changes to options declarations don't require a full-source rebuild (you only have to rebuild those sources that depend on the set of options that changed).
2. lots of sanity checks (that the same option isn't declared twice, that option values are in range for their type, that all options are documented properly, etc.)
3. consistency: Boolean-valued option --foo gets a --no-foo automatically, documentation is generated consistently, the option-parsing matches the documented option name, etc.
4. setting options programmatically via SmtEngine::setOption() is enabled, and behaves the same as command-line equivalents (including checking the value is in range, etc.)
5. the notion of options being "set by the user" is now primitive; you can use (e.g.) options::unconstrainedSimp.wasSetByUser() instead of having to use (and maintain) a separate Boolean option for the purpose
I've taken lots of care not to break anything. Hopefully, I've succeeded in that.
|
|
\file tags corrected, copyright added to files that had it missing, etc.
I ensured that I didn't change any code with this commit, and even tested on the cluster to be doubly sure:
http://church.cims.nyu.edu/regress-results/compare_jobs.php?job_id=4655&reference_id=4646&p=0
|
|
ArithPriorityQueue::reduce(), ::begin() and ::end().
|
|
reenqueued so that the queue was a superset of the failing assertions. This adds a super expensive debug routine unenqueuedVariablesAreConsistent() that catches this bug. This is enabled when -d arith::consistency is turned on. make check passes with this flag enabled.
|
|
restricted-simplex.
|
|
single check call.
- This increases the number of substitutions that ppAssert can solve on integer equations.
|
|
- Splits the functionality of having a sparse matrix of Ts and a solved matrix of rationals in tableau.
- Splits ArithVarSet into DenseMap and CDDenseSet and simplifies the code.
- No performance loss!
|
|
Adds DecisionEngine and an abstract class DecisionStrategy
which other strategies will derive from eventually.
Full revision summary of merged commits:
r3241 merge from trunk
r3240 fix
r3239 WIP
r3238 JH, CVC3 code: 5% done -- 5% translated
r3237 JH groundwork
r3236 make make regrss pass
r3234 hueristic->heuristic
r3229 JustificationHeuristic: EOD-WIP
r3228 DecisionEngine: hookup assetions
r3227 move ITE outside simplifyAssertions
r3226 DecisionStrategy abstract class
r3222 DecisionEngine: begin
|
|
Below is a highlight of the changes:
- This introduces a new normal form to arithmetic.
-- Equalities and disequalities are in solved form.
Roughly speaking this means: (= x (+ y z)) is in normal form.
(See the comments in normal_form.h for what this formally requires.)
-- The normal form for inequality atoms always uses GEQ and GT instead of GEQ and LEQ.
Integer atoms always use GEQ.
- Constraint was added to TheoryArith.
-- A constraint is a triple of (k x v) where:
--- k is the type of the constraint (either LowerBound, UpperBound, Equality or Disequality),
--- x is an ArithVar, and
--- v is a DeltaRational value.
-- Constraints are always attached to a ConstraintDatabase.
-- A Constraint has its negation in the ConstraintDatabase [at least for now].
-- Every constraint belongs to a set of constraints for each ArithVar sorted by the delta rational values.
-- This set can be iterated over and provides efficient access to other constraints for this variable.
-- A literal may be attached to a constraint.
-- Constraints with attached literals may be marked as being asserted to the theory (sat context dependent).
-- Constraints can be propagated.
-- Every constraint has a proof (sat context dependent).
-- Proofs can be explained for either conflicts or propagations (if the node was propagated). (These proofs may be different.)
-- Equalities and disequalities can be marked as being split (user context dependent)
- This removes and replaces:
-- src/theory/arith/arith_prop_manager.*
-- src/theory/arith/atom_database.*
-- src/theory/arith/ordered_set.h
- Added isZero(), isOne() and isNegativeOne() to Rational and Integer.
- Added operator+ to CDList::const_iterator.
- Added const_iterator to CDQueue.
- Changes to regression tests.
|
|
- Improved the checks in AssertLower and AssertUpper so that redundant bounds cause less work.
- Because of the above change, d_constantIntegerVariables now cannot have duplicate elements enqueued. This allows removing d_varsInDioSolver.
- Fix to an assertion in CDQueue.
- Implements a CDArithVarSet using a vector of booleans and CDList.
- Refactored ArithVar out of arith_utilities.h. Miscellaneous cleanup of arithmetic.
|
|
This is a significant refactoring of code.
- r2820
-- Refactors Simplex so that it does significantly fewer functions.
-- Adds the LinearEqualityModule for handling update and pivotAndUpdate and other utility functions that require access to both the tableau and partial model.
-- Some of the code for propagation has moved to TheoryArith.
-r2826
-- Small changes to documentation and removes the no longer implemented deduceLowerBound() and deduceUpperBound().
- r2827
-- Adds isZero() to Rational. Adds cmp to DeltaRational.
- r2831
-- Refactored comparison to upper and lower in the partial model to use DeltaRational::cmp.
-- Refactored AssertUpper and AssertLower in TheoryArith to include functionality that has weaseled into TheoryArith::assertionCases.
|
|
Dumping infrastructure. Can dump preprocessed queries and clauses. Can
also dump queries (for testing with another solver) to see if any conflicts
are missed, T-propagations are missed, all lemmas are T-valid, etc. For a
full list of options see --dump=help.
CUDD building much cleaner.
Documentation and assertion fixes.
Printer improvements, printing of commands in language-defined way, etc.
Typechecker stuff in expr package now autogenerated, no need to manually
edit the expr package when adding a new theory.
CVC3 compatibility layer (builds as libcompat).
SWIG detection and language binding support (infrastructure).
Support for some Z3 extended commands (like datatypes) in SMT-LIBv2 mode
(when not in compliance mode).
Copyright and file headers regenerated.
|
|
|
|
- This adds code for bounds refinement, and conflict weakening.
- This adds util/boolean_simplification.h.
- This adds a propagation manager to theory of arithmetic.
- Propagation is disabled by default.
- Propagation can be enabled by the command line flag "--enable-arithmetic-propagation"
- Propagation interacts *heavily* with rewriting equalities, and will work best if the command line flag "--rewrite-arithmetic-equalities" is enabled.
|
|
* Makes Options an "omnipresent thread-local global" (like the notion
of the "current NodeManager" was already). Options::current() accesses
this structure.
* Removes Options from constructors and data structures everywhere
(this cleans up a lot of things).
* No longer uses StatisticsRegistry statically. An instance of the
registry is created and linked to a NodeManager.
* StatisticsRegistry::current() is similar to Options::current(), but
the pointer is stowed in the NodeManager (rather than stored)
* The static functions of StatisticsRegistry have been left, for backward
compatibility; they now use the "current" statistics registry.
* SmtEngine::getStatisticsRegistry() is a public accessor for the
registry; this is needed by main() to reach in and get the registry,
for flushing statistics at the end.
|
|
|
|
importantly removes an unintentional line of code that had it pivoting more times than intended before rechecking the queue. Importantly, it does this without losing any examples with rewrite-equality enabled. This adds a parameter NUM_CHECKS which determines how many times the queue chould be checked during difference mode. A value of 10 for NUM_CHECKS has been empirically determined to be good in practice. See jobs 1815, 1824, 1825, 1821, 1814.
|
|
|
|
|
|
- Adds ArithStaticLearner. Consolidates and cleans up the code for static learning in arithmetic. Static learning is now associated with a small amount of state between calls. This is used to track the data for the miplib trick. The goal is to make this inference work without relying on the fact that all of the miplib problem is asserted under the same AND node.
- This commit contains miscellaneous other arithmetic cleanup.
|
|
phases of looking for additional conflicts during and after the heuristic pivoting stage. (For the expected performance gain, comparing jobs 1676 and 1643 gives a rough idea.)
|
|
PreferenceFunction allows for specifying how to choose between two nonbasic variables for which should become basic during the selectSlack(...) function. This partially addresses a point brought up by Dejan during the Code Review. (Unfortunately, function pointers are involved in the implementation. Because of this, I have had Morgan review this code before check-in.)
|
|
are sent to the sat solver during theory propagation. The lemmas currently come from additional conflicts that are discovered by findConflictOnTheQueue(...).
|
|
- Adds 3 choices of heuristic variable orders to use in ArithPriorityQueue.
- Adds the pivot-rule command line option.
|
|
heuristic round has been completed. This happens immediately before switching to the variable order round.
|
|
an abstraction to the previous priority queues representing the 2 different pivoting rules.
|
|
- Removed a bug in row counting in row counting.
|
|
- The Tableau is now in charge of managing what variables are basic in a unified manner. Specifically, TheoryArith::d_basicManager was merged into Tableau::d_basicVariables.
|
|
|
|
ejection, and is now superfluous.
|
|
will follow shortly.
|
|
- Preparing to remove row ejection from the code base!
- Checks for conflicts immediately after a pivot to avoid potentially wasteful search.
- Added arithvar_set.h. This replaces ArithVarSet that was previously in the Tableau, and ArithVarDenseSet.
- Removes variables that have no preregistered bounds during presolve().
- Theory::isLeafOf() currently returns true for atoms. (I was unaware.) I modified Variable::isMember() to account for this exclude atoms.
- Added statistics all over the place.
This commit effects both boolean search and simplex search so expect running times to go all over the place. The time differences should be roughly as follows:
http://goedel.cims.nyu.edu/regress-results/compare_jobs.php?job_id=1486&reference_id=1447&p=10&category=1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20,21,29
|
|
termination bug again. *le sigh*)
|
