Age | Commit message (Collapse) | Author |
|
|
|
|
|
|
|
|
|
Some quick background: CVC4 has two primary contexts: the assertion
context (which corresponds to the push/pops issued by the user) and the
decision/SAT context (which corresponds to the push/pops done by the
user and each decision made by MiniSat.
Before 2cc0e2c6a691fb6d30ed8879832b49bc1f277d77, we had an additonal
push/pop when doing the actual solving. With these removed, it could
happen that we get a wrong result when doing incremental solving:
```
...
; check-sat call returns sat, the decision level in the SAT solver is
; non-zero at this point
(check-sat)
; Solver::addClause_() determines that we cannot add the clause to the
; SAT solver directly because the decision level is non-zero (i.e. the
; clause is treated like a theory lemma), thus it is added to the lemmas
; list.
(assert false)
; The solver stores some information about the current state, including
; the value of the "ok" flag in Solver, which indicates whether the
; current constraints are unsatisfiable. Note that "ok" is true at this
; point.
(push)
; Now, in Solver::updateLemmas(), we add clauses from the lemmas list.
; The problem is that empty clauses (which "false" is after removing "false"
; literals) and unit clauses are not added like normal clauses. Instead,
; the empty clause, essentially just results in the "ok" flag being set to
; false (unit clauses are added to the decision trail).
(check-sat)
; Here, the solver restores the information stored during
; (push), including the "ok" flag, which is now true again.
(pop)
; At this point, the solver has "forgotten" about (assert false) since
; the "ok" flag is back to true and it answers sat.
(check-sat)
```
There are multiple ways to look at the problem and to fix it:
- One could argue that an input assertion should always be added
directly to the clauses in the SAT solver, i.e. the solver should
always be at decision level 0 when we are adding input clauses. The
advantage of this is that it is relatively easy to implement,
corresponds to what the original MiniSat code does (it calls
Solver::cancelUntil(0) after solving), and is no worse than what we had
before commit 2cc0e2c6a691fb6d30ed8879832b49bc1f277d77 (push/pop do a
strict superset of what resetting the decision at the current assertion
level does). The disadvantage is that we might throw away some useful work.
- One could argue that is fine that (assert false) is treated like a
theory lemma. The advantage being that it might result in more efficient
solving and the disadvantage being that it is much trickier to
implement.
Unfortunately, it is not quite clear from the code what the original
intention was but we decided to implement the first solution. This
commit exposes a method in MiniSat to reset the decisions at the current
assertion level. We call this method from SmtEngine after solving.
Resetting the decisions directly after solving while we are still in
MiniSat does not work because it causes issues with datastructures in
the SMT solver that use the decision context (e.g.
TheoryEngine::d_incomplete seems to be reset too early, resulting in us
answering sat instead of unknown).
|
|
Adds missing override keywords.
|
|
|
|
|
|
context/ classes, and their subclasses. Fixes compilation issues with clang 3.5 and -std=c++11 'exception specification of overriding function is more lax than base version' for a couple of different classes.
|
|
|
|
|
|
- The functionality the get the StatisticsRegistry attached to the SmtEngine was previously through StatisticsRegistry::current(). This is the dominant StatisticsRegistry in the code. (There is another StatisticsRegistry attached to the NodeManager.) Having this be a static function on StatisticsRegistry requires the use of an SmtEngine in the wrong compilation unit.
- Usages of StatisticsRegistry::current() that were visible in prop/{bvminisat,minisat} has been removed. A pointer to the relevant StatisticsRegistry should be passed instead into the constructor.
- The function StatisticsRegistry::current() has been replaced by SmtScope::currentStatisticsRegistry(). SmtScope is in the libcvc4 package, where SmtEngine is available in the compilation unit.
- The function smtStatisticsRegistry() is a synonym for SmtScope::currentStatisticsRegistry() in smt/smt_statistics_registry.h. This header has fewer include dependencies than the one for SmtScope.
- Correspondingly, the static functions StatisticsRegistry::{registerStat, unregisterStat} have been removed. One should instead use smtStatisticsRegistry()->{registerStat,unregisterStat} instead.
- The KEEP_STATISTIC macro has been moved into smt/smt_statistics_registry.h.
- Documents the reason StatisticsRegistry is CVC4_PUBLIC. This lets me remove the warning I added.
- Removing most operators for timespec from statistics_registry.h file. These a bit error prone in clang.
- Most of the really confusing ifdef's in util/statistics_registry.h are gone.
|
|
|
|
|
|
Signed-off-by: Morgan Deters <mdeters@cs.nyu.edu>
|
|
|
|
Thanks Johannes Kanig for the report.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
* the newVar interface of the sat solver now changed to include (isTheoryLiteral, preRegister, canEliminate)
* when bitblast-eager all bv atoms are (theory=false, prereg = true, canelim = true)
* bitblast-eager implies decision=internal
|
|
|
|
Adds TheoryQuantifiers and TheoryRewriteRules, QuantifiersEngine, and other infrastructure.
Adds theory instantiators to many theories.
Adds the UF strong solver.
|
|
(no performace or search behavior changes expected)
|
|
* notifications are now through the interface subclass instead of a template
* notifications include constants being merged
* changed contextNotifyObj::notify to contextNotifyObj::contextNotifyPop so it's more descriptive and doesn't clutter methods when subclassed
* sat solver now has explicit methods to make true and false constants
* 0-level literals are removed from explanations of propagations
|
|
|
|
|
|
The available SAT solvers can be seen with the --show-sat-solvers option.
|
|
|
|
|
|
|