Age | Commit message (Collapse) | Author |
|
|
|
Breaking an edge between the sat solver and command.h.
|
|
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.
|
|
Signed-off-by: Morgan Deters <mdeters@cs.nyu.edu>
|
|
* support for new commands meta-info, declare-const, echo, get-model,
reset, and reset-assertions
* support for set-option :global-declarations
* support for set-option :produce-assertions
* support for set-option :reproducible-resource-limit
* support for get-info :assertion-stack-levels
* support for set-info :smt-lib-version 2.5
* ascribe types for abstract values (the new 2.5 standard clarifies that
this is required)
* SMT-LIB v2.5 string literals (we still support 2.0 string literals when
in 2.0 mode)
What's still to do:
* check-sat-assumptions/get-unsat-assumptions (still being hotly debated).
Also set-option :produce-unsat-assumptions.
* define-fun-rec doesn't allow mutual recursion
* All options should be restored to defaults with (reset) command.
(Currently :incremental and maybe others get "stuck" due to late driver
integration.)
|
|
support elsewhere).
|
|
|
|
generation; fix bug 285.
* segfaults/assert-fails in proof-generation fixed, including bug 285
* added --check-proofs to automatically check proofs, like --check-models (but only for UF/SAT at present)
* proof generation now works in portfolio (but *not* --check-proofs, since LFSC code uses globals)
* proofs are *not* yet supported in incremental mode
* added --dump-proofs to dump out proofs, like --dump-models
* run_regression script now runs with --check-proofs where appropriate
* options scripts now support :link-smt for SMT options, like :link for command-line
|
|
|
|
|
|
|
|
|
|
|
|
|
|
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.)
|
|
* ITE removal fixed to be context-dependent (on UserContext).
Resolves incrementality bugs 376 and 396 (which had given wrong answers).
* some bugfixes for incrementality that Dejan found (fixes bug 394)
* fix for bug in SmtEngine::getValue() where definitions weren't respected
(partially resolves bug 411, but get-model is still broken).
* change status of microwave21.ec.minimized.smt2 (it's actually unsat, but
was labeled sat); re-enable it for "make regress"
Also:
* --check-model doesn't fail if quantified assertions don't simplify away.
* fix some examples, and the Java system test, for the disappearance of the
BoolExpr class
* add copy constructor to array type enumerator (the type enumerator
framework requires copy ctors, and the automatically-generated copy ctor
was copying pointers that were then deleted, leaving dangling pointers in
the copy and causing segfaults)
* --dump=assertions now implies --dump=skolems
* --dump=assertions:pre-<PASS> and --dump=assertions:post-<PASS> now allow
you to dump before/after a particular preprocessing pass. E.g.,
--dump=assertions:pre-ite-removal or --dump=assertions:post-static-learning.
"--dump=assertions" by itself is after all preprocessing, just before CNF
conversion.
* minor fixes to dumping output
* include Model in language bindings
Minor refactoring/misc:
* fix compiler warning in src/theory/model.cpp
* remove unnecessary SmtEngine::printModel().
* mkoptions script doesn't give progress output if stdout isn't a terminal
(e.g., if it's written to a log, or piped through less(1), or whatever).
* add some type enumerator unit tests
* de-emphasize --parse-only and --preprocess-only (they aren't really "common"
options)
* fix some exception throw() specifications in SmtEngine
* minor documentation clarifications
|
|
|
|
(this commit was certified error- and warning-free by the test-and-commit script.)
|
|
meeting last week. The SmtEngine now subscribes to NodeManager events,
does appropriate dumping of variable declarations, and notifies the Model
class.
The way to create a skolem is now:
nodeManager->mkSkolem("myvar_$$", TypeNode, "is a variable created by the theory of Foo")
The first argument is the name of the skolem, and the (optional) "$$" is a
placeholder for the node id (to get a unique name). Without a "$$", a "_$$"
is automatically appended to the given name.
The second argument is the type.
The (optional, but recommended) third argument is a comment, used by the
dump infrastructure to indicate what the variable is for / who owns it.
An optional fourth argument (not shown) allows you to specify flags that
control the behavior (e.g., don't do notification, and/or don't make a
unique name). Look at the documentation for details on these.
In particular, the above means you can't just do a mkSkolem(boolType) for
example---you have to specify a name and (hopefully also,
but it's optional) a comment. This leads to easier debugging than the
anonymous skolems before, since we'll be able to track where the skolems
came from.
Much of the Model and Dump stuff, as well as some Command stuff, is cleaned up
by this commit. Some remains to be cleaned up.
(this commit was certified error- and warning-free by the test-and-commit script.)
|
|
To enable, use --check-models. Turning on the option can be done in debug or optimized builds, regardless of whether normal assertions are on or not. This is to allow us to check the generated models in long-running queries, and might be useful to end users as a double-check too.
By default, --check-models is quiet (no output unless it detects a problem). That allows regression runs to pass unless there are problems:
make regress CVC4_REGRESSION_ARGS=--check-models
To see it work, use -v in addition to --check-models.
There may still be bugs in the feature itself, but already I've found some apparent model-generation bugs (and discussed with Andy) from this feature, so it seems useful in its current state.
--check-models turns on what SMT-LIBv2 calls "interactive mode" (which keeps the list of user assertions around), and also implies --produce-models. This version does NOT require incremental-mode, which one design did (the one mentioned in yesterday's meeting).
Also:
* TheoryUF::collectModelInfo() now generates UninterpretedConstants (rather than non-constants)
* The UF rewriter now reduces (APPLY_UF (LAMBDA...) args...), and treats uninterpreted constants correctly (e.g. uc_U_1 != uc_U_2)
* The SubstitutionMap now supports substitutions of operators for paramaterized kinds (e.g., function symbols)
|
|
collectModelInfo with fullModel argument, most theory-specific implementation out of the model class, model printer relegated to printer classes. Also updates to finite mode finding, modifications to datatypes making them compatible with theory combination, support for theory-specific handling of user attributes, refactoring of uf models
|
|
|
|
|
|
|
|
By default, common subexpressions are dagified if they appear > 1 time and are not constants or variables.
This can be changed with --default-expr-dag=N --- N is a threshold such that if the subexpression occurs > N
times, it is dagified; a setting of 0 turns off dagification entirely.
If you notice strange dumping behavior (taking too long to print anything, e.g.), revert to the old behavior
with --default-expr-dag=0 and let me know of the problem.
|
|
* with a patched SWIG, the ocaml bindings build correctly.
** I will provide my patch to the SWIG dev team.
* fixed some class interfaces to play more nicely with SWIG.
* php, perl, tcl now work; examples added.
* improved binding module building and installation.
Also:
Stop #defining NULL ((void*) 0). This has been in cvc4_public.h for
a long, long time, I forget why I added it in the first place, and
it's a very, very bad idea. In C++, certain things are permitted
for NULL that aren't permitted for ((void*) 0), like for instance
implicit conversion to any pointer type. We didn't see an issue
here (until now, when interfacing with SWIG), because GCC is usually
pretty smart at working around such a broken #definition of NULL.
But that's fragile.
New exception-free Command architecture. Previously, some command
invocations were wrapped in a try {} catch() {} and printed out an
error. This is much more consistent now. Each Command invocation
results in a CommandStatus. The status can be "unsupported",
"error", or "success" (these are each derived classes, though, not
strings, so that they can be easily printed in a language-specific
way... e.g., in SMT-LIBv2, they are printed in a manner consistent
with the spec, and "success" is not printed if the print-success
option is off.) All Command functionality are now no-throw
functions, which @cconway reports is a Good Thing for Google
(where all C++ exceptions are suspect), and also I think is much
cleaner than the old way in this instance.
Added an --smtlib2 option that enables an "SMT-LIBv2 compliance
mode"---really it just sets a few other options like strictParsing,
inputLanguage, and printSuccess. In the future we might put other
options into a compliance mode, or we might choose to make it the
default.
|
|
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.
|
|
|
|
|
|
implemented. This new infrastructure removes support for pretty-printing
(even in the AST language) an Expr with reference count 0. Previously,
this was supported in a few places internally to the expr package, for
example in NodeBuilder. (Now, a NodeBuilder cannot be prettyprinted, you
must extract the Node before printing it.)
|