| Age | Commit message (Collapse) | Author |
|---|
|
This commit removes the support for the CVC language and converts all *.cvc regression tests to SMT-LIBv2.
|
|
This implements several variants of lazy proof checking in the core proof checker.
Note this extends the ProofNode class with an additional Boolean d_provenChecked indicating whether the d_proven field was checked by the underlying proof checker.
This PR updates the default proof checking mode to lazy. The previous default can now be enabled by --proof-check=eager-simple.
|
|
This enables the new implementation of justification heuristic by default.
Fixes #5454, fixes #5785. Fixes wishues 114, 115, 149, 160.
|
|
This updates all regressions that pass check-unsat-cores to enable check-unsat-cores. This includes any incremental benchmark, which was disabled in run_regression.py previously.
It adds --no-check-unsat-cores to a few corner benchmarks that were previously disabled based on --incremental.
It also reverts a change to when proofs are disabled: options like sygus-inference should not permit proofs (or unsat cores).
|
|
This PR improves the integration of the CAD solver with the nl-ext solver in a simple way: we simply use a few of the simple linearization lemmas in combination with CAD by default, significantly improving the performance on QF_NRA.
|
|
|
|
This makes it so that we attempt evaluation + rewriting on applications of operators that do not always evaluate, and return constants in case the evaluation was successful.
This fixes warnings for check-models on 43 of our regressions, and also uncovered one regression where our model was wrong but check-models silently succeeded. I've opened CVC4/cvc4-projects#276 for fixing the latter.
|
|
Linear arithmetic simply tried to branch on the first integer variable that had a non-integral assignment.
If it holds stale variables that are not part of the current input, these branches can be emitted and are processed by the solver, but the resulting new assertions will not be considered relevant and thus not added to the theory.
As it still triggers a new theory check, linear arithmetic repeats the same procedure and causes an infinite loop.
This PR explicitly tracks the currently relevant nodes by storing all preregistered nodes and only allows branching on variables from this set.
Fixes #6146.
|
|
This commit removes parser and printer support for old SMT-LIB standards and also converts all regression tests to 2.6.
|
|
Moves regressions taking >4 seconds (summing all configs) in debug to regress1.
|
|
Also moves several proof-specific options to proof_options.
|
|
This enables more strict handling of operators div, mod and abs
for Integer arithmetic logics.
More strict handling for '/' for Real arithmetic logics is more involved
and should be done in the parser -- instead at solving time, like is
currently done for checking that the application * is in the linear
fragment. The latter should be checked in the parser, too.
This is postponed to a later PR.
|
|
Was causing arithmetic to process a Boolean equality when --arith-rewrite-equalities is true.
Fixes #5761.
|
|
Arith has a normal form for literals, which the rewriter computes.
Nonetheless, arith sometimes gets (and ultimately propagates) non-normal
literals. It normalizes them internally and goes about its business.
However, when asked for an explanation, it must prove the non-normal
literal, rather than the normal one.
Also includes a regression for the bug, courtesy of Andy.
Co-authored-by: Andrew Reynolds andrew-reynolds@uiowa.edu
|
|
This adds a net +82 regressions to regress[0-2] and adds several additional disabled regressions to regress3 and regress4. This involved fixing the status on several regressions, and ensuring CMakeLists.txt includes all files (exactly once) in the test/regress/ subdirectory.
It also moves several regressions to the proper regression levels (those that take >30 seconds in debug are moved to regress3+).
|
|
This updates the preprocessor so that expand definitions does not expand theory symbols at the beginning of preprocessing.
This also restores the previous expandDefinitions method in arithmetic, which is required for correctly interpreting division by zero in models, but should not be applied at the beginning of preprocessing. Moreover it ensures that only partial operators are eliminated in arithmetic expandDefinitions, which required an additional argument partialOnly to arith::OperatorElim.
This adds -q to suppress warnings for many quantified regressions which now emit warnings with --check-model. This will be addressed later as part of CVC4/cvc4-wishues#43.
The purpose of this PR is two-fold:
(1) Currently our responses to get-value are incorrect for partial operators like div, mod, seq.nth since partial operators can be left unevaluated.
(2) The preprocessor should have the opportunity to rewrite and eliminate extended operators before they are expanded. This is required for addressing performance issues for non-linear arithmetic. It is also required for ensuring that trigger selection can be done properly for datatype selectors (to be addressed on a later PR).
|
|
This PR makes it so that extended arithmetic operators are not expanded during expand definitions. Instead they are eliminated at theory preprocessing, which occurs as the last step of preprocessing.
The motivation for this is three fold:
(1) Some quantified LIA benchmarks lead to CVC4 failing to eliminate div functions from quantifier instantiation, this leads to spurious logic failures. A regression is included, which now is correctly solved.
(2) We should allow better rewriting and preprocessing for extended arithmetic functions, especially for div/mod which is important for many users of QF_NIA.
(3) More generally,Theory::expandDefinitions will be deleted. All functionalities in expandDefinitions should move to Theory::ppRewrite.
This changes impacts many benchmarks that involve non-linear and quantifiers:
Many benchmarks (as expected) give a warning during check-models since (/ n 0) cannot be evaluated. I've added -q to disable these warnings. Fully addressing this is part of an agenda to improve our support for --check-models.
Several fuzzing benchmarks involving NL+quantifiers now time out. However, several can be solved by --sygus-inst, which is now the preferred instantiation strategy for NL+quantifiers.
2 other non-linear regressions time out, which are disabled in this PR.
one sygus-inference benchmark (regress1/sygus/issue3498.smt2), now correctly times out (previously it did not timeout since the preprocessor was unable to apply sygus-inference and resorted to normal methods, now sygus-inference can apply but as expected times out).
Fixes #5237.
|
|
The arithmetic inference manager did rewrite conflicts, which lead to nodes from the conflict not being assumptions (but rewritten assumptions) triggering an assertion. This rewrite is now removed.
Furthermore rewrites triggering the same assertion within the icp subsolver have been refactored to avoid this issue.
|
|
This reduces the use of recursion in term formula removal module. The recursion depth is now limited to the number of term positions on a path where a formula must be removed, instead of being limited to the overall formula depth. This PR also fixes some documentation.
Notice for its main cache d_tfCache, this class uses a CDInsertHashMap (instead of a CDHashMap) which does not allow inserting more than once, hence an auxliary "processedChildren" vector had to be introduced to track whether we have processed children. It's not clear to me whether we should just use the more standard CDHashMap, which would simplify this.
One non-linear arithmetic regression went unsat -> unknown, I added --nl-ext-tplanes to fix this.
This should fix #4889. It is possible to further reduce the recursion in this pass, which can be done on a followup PR if needed.
|
|
In preparation of removing the old proof module, this commit changes the
regression runner to not add the flag --check-proofs anymore when
running the regressions.
|
|
Fixes #4367. We set the theoryof-mode depending on whether sharing is
enabled or not. However, we were checking whether sharing was enabled
before theory widening, leading to unexpected results. This commit moves
the check after widening the theories.
For the benchmark in the issue, setting the theoryof-mode before theory
widening lead to problems because of the following:
The main solver checks the condition for enabling term-based
theoryof-mode, decides not to do it because sharing is not enabled
Main solver adds UF to the logic
Main solver does a check-sat all
Unsat core check runs, sees both UF and NRA enabled, and switches
to term-based mode
Main solver gets to second check-sat call, now the theoryof-mode is
suddenly changed, which leads to problems in the equality engine
This commit fixes the issue in this particular instance but it is important
to note that it does not address the issue of the subsolver changing
settings in general. This can only really be solved by separating the
options from the ExprManager/NodeManager and having separate
options for each SmtEngine/Solver instance.
|
|
Fixes #4525. The actual problem in the issue is not that the unsat core
is satisfiable but that our unsat core check is not working as intended.
Our unsat core check uses the same `ExprManager` as the main `SmtEngine`
and thus also shares the same attributes for nodes. However, since we
have a different `SmtEngine` instance for the check, we also have
different instances of `TheoryArith`. This leads to the check failing
due to the following:
- Our only input assertion is `(> (cot 0.0) (/ 1 0)))`.
- `TheoryArith::expandDefinition()` gets called on `(> (cot 0.0) (/ 1
0))` but nothing happens.
- `TheoryArith::expandDefinition()` gets called on `(/ 1 0)`, which gets
expanded as expected but no attribute is set because it happens in a
simple `TheoryArith::eliminateOperators()` call.
- `TheoryArith::expandDefinition()` on `(cot (/ 0 1))` first expands to
`(/ 1 0)` (not cached) and then we expand it recursively to the
expected term and cache `(/ 1 0) ---> (ite (= 0 0) (divByZero 1.0) (/
1 0))`.
Up to this point, things are suboptimal but there are no correctness
issues. The problem starts when we do the same process in the unsat core
check:
- Our only input assertion is again `(> (cot 0.0) (/ 1 0)))`.
- `TheoryArith::expandDefinition()` gets called on `(> (cot 0.0) (/ 1
0))` but nothing happens.
- `TheoryArith::expandDefinition()` gets called on `(/ 1 0)`, which gets
expanded as expected but no attribute is set or checked because it
happens in a simple `TheoryArith::eliminateOperators()` call. Note
that the skolem introduced here for the division-by-zero case is
different from the skolem introduced above because this is in a
different `TheoryArith` instance that does not know the existing
skolems.
- `TheoryArith::expandDefinition()` on `(cot (/ 0 1))` first expands to
`(/ 1 0)` (not cached) and then we use the cache from our solving call
to expand it `(/ 1 0) ---> (ite (= 0 0) (divByZero 1.0) (/ 1 0))`.
Note that `divByZero` here is the skolem from the main solver.
As a result, the preprocessed assertions mix skolems from the main
`SmtEngine` with the `SmtEngine` of the unsat core check, making the
constraints satisfiable.
To solve this problem, this commit removes the caching-by-attribute
mechanism. The reason for removing it is that it is currently
ineffective (since most eliminations do not seem to be cached) and there
are caches at other levels, e.g. when expanding definitions. If we deem
the operator elimination to be a bottleneck, we can introduce a similar
mechanism at the level of `TheoryArith`.
|
|
This is a major refactor of how operators are eliminated in arithmetic. Currently there were (at least) two things wrong:
(1) ppRewriteTerm sent lemmas on the output channel. This behavior is incompatible with how preprocessing works. In particular, this caused unconstrained simplification to be unaware of terms from such lemmas, leading to incorrect "sat" answers.
(2) Lemmas used to eliminate certain "div-like" terms were processed in a context-independent way. However, lemmas should be cached in a user-context-dependent way. This was leading to incorrect "sat" answers in incremental.
The solution to these issues is to eliminate operators via the construction of witness terms. No lemmas are sent out, and instead these lemmas are the consequence of term formula removal in the standard way.
As a result of the refactor, 2 quantifiers regressions time out due to infinite branch and bound issues (one only during --check-unsat-cores). These appear to be random and I've changed the options to avoid these issues. 3 others now have check-model warnings, which I've added --quiet to. Improving check-model will be addressed on a future PR.
This PR is not required for SMT COMP since we have workarounds that avoid both the incorrect behaviors in our scripts.
Also notice that --rewrite-divk is effectively now enabled by default always.
Fixes #4484, fixes #4486, fixes #4481.
|
|
This renames api::Solver::checkValidAssuming to checkEntailed and
removes api::Solver::checkValid. Internally, SmtEngine::query is renamed
to SmtEngine::checkEntailed, and these changes are further propagated to
the Result class.
|
|
Fixes #3914. The pass was only applicable to inputs with UFs that were
exclusively applied to single integer values. This limitation seems to
make the preprocessing pass not very useful in practice and it is
subsumed by our Ackermannization pass, which can remove UFs from more
complex inputs. Thus, this commit removes the preprocessing pass.
|
|
Switches arith_proof.cpp from th_lra to th_lira.
Changes:
Eliminate the d_realMode hack.
instead: modify printOwnedTermAsType prints as integers OR
reals, depending on expectedType.
simultaneously: write printOwnedTermAsType more concisely
also: reimplement printOwnedSort.
Change to the LIRA axioms:
Because they reason about bound types using side conditions, we
no longer need to worry about choosing the correct strictness for
our axiom.
This allows us to cut out a lot of code, rewriting & shrinking
printTheoryLemmaProof.
They also have different names.
This requires us to change a lot of string literals
enable proof-checking for many tests.
|
|
|
|
* Add more arith proof regression tests
These tests are designed to test interesting cases of arithmetic proofs,
such as mixing integers and reals and tightening bounds.
Right now, they have the --no-check-proofs flag set, which prevents them
from testing the proof machinery. However, once we check that machinery
into master, we'll remove that flag, thus enabling the full effect of
the tests.
* A few comments explaining things.
* Add another tightening test
* Add new test to CMake
* No --no-check-models. There are no models anyway.
* Delete smt-lib-version, per Yoni
|
|
|
|
* Bugfix: convert ifte arms to formulas for printing
We have two kinds of ITEs in our LFSC proofs:
* ite: for sort-typed expressions
* ifte: for formulas
Say that we have a Bool-sorted ITE. We had machinery for emitting an
`ifte` for it, but this machinery didn't actually convert the arms of
the ITE into formulas... Facepalm.
Fixed now.
* Test the lifting of ITEs from arithmetic.
This test verifies that booleans ITEs are correctly lifted to formula
ITEs in LRA proofs.
It used to fail, but now passes.
* clang-format
* Typos.
* Add test to CMake
* Set --check-proofs in test
* Address Yoni
* Expand printsAsBool documentation
* Assert ITE typing soundness
* Assert a subtype relation for ITEs, not equality
* Update src/proof/arith_proof.h
Thanks Yoni!
Co-Authored-By: yoni206 <yoni206@users.noreply.github.com>
Co-authored-by: yoni206 <yoni206@users.noreply.github.com>
|
|
|
|
Fixes #1399.
|
|
The ackermannization process is currently already support general theories rather than specifically for BV. In this pull request, an option has been added to turn on ackermannization independently.
|
|
This commit removes the SMT1 parser infrastructure and adds the SMT2 translations of the SMT1 regression tests. For now this commit removes regression test regress3/pp-regfile.smt since the SMT2 translation has a file size of 887M (vs. 172K for the SMT1 version).
Fixes #2948 and fixes #1313.
|
|
|
|
|
|
|
|
Currently, we can optionally specify an *.expect file with the metadata
of a regression test. This commit removes that option because it was not
widely used, adds maintenance overhead and makes the transition to a new
build system more cumbersome. Regression files can still be fed to a
solver without removing the metadata first since they are in comments of
the corresponding input format (note that this was not always the case,
it changed in efc6163629c6c5de446eccfe81777c93829995d5).
|
|
|
|
In SmtEngine::setDefaults() we were setting options such as
--simplifciation=none for unsat cores but not proofs. Producing unsat
cores relies on the same infrastructure as proofs and should be a subset
of the same work in most cases. Thus, this commit changes
SmtEngine::setDefaults() to set the same options for proofs as we
previously only did for unsat cores. Additionally, it changes the
function to only set the simplificationMode to SIMPLIFICATION_MODE_BATCH
if proofs and unsat cores are not enabled. Fixes issue #1953.
|
|
This commit refactors code generation for options. It uses a new configuration format for defining options (*.toml) and a new Python script mkoptions.py to generate the source code and option documentation.
The option behavior did not change for most of the options, except that for bool --enable-/--disable- long options enable/disable was removed. E.g. --enable-miplib-trick and --disable-miplib-trick got changed to --miplib-trick and --no-miplib-trick.
This commit fixes also an issues with set-option/get-option via the SMT2 interface. Before long options were only accessible if the name included the =ARG part.
|
|
Until now, regression tests were split across tens of different
Makefile.am, which required a lot of code duplication and does not
really seem to be in the spirit of automake. If we want to change the
LOG_COMPILER/LOG_DRIVER for example, we have to change every single
Makefile.am, which is cumbersome (I was able to get something
semi-working by exporting those variables but it didn't seem very
clean). Additionally, it made the output of the regression tests fairly
verbose and split the output across multiple log files. Finally
it also limited parallelism when running the regression tests (this fix lowers
the time it takes to run regression level 1 from 3m to 1m45s on my
machine with 16 threads).
This commit moves all the regression tests into
test/regress/Makefile.tests and changes test/regress/Makefile.am to deal
with this new structure. Finally, it changes how the test summary in
test/Makefile.am is produced: instead of relying on the log files for
the subdirectories, it greps for the test results in the log files of
the individual tests. Not the most elegant solution but we should
probably anyway delegate that task to a Python script at some point.
|
|
|
|
sygus), update regressions.
|
|
|
|
|
|
|
|
|
|
|
|
commit for nlAlgMaster.
|
