| Age | Commit message (Collapse) | Author |
|---|
|
|
|
|
|
|
|
In TSatProof<Solver>::finalizeProof(), we got a clause from the clause
allocator, called resolveUnit() and then size() on the clause. The problem is
that resolveUnit() can reallocate memory (and there is even a comment warning
about that in finalizeProof()), which invalidates the clause. This commit gets
the clause again from the clause allocator before calling size().
|
|
cbqi, tptp parser), fix full saturation instantiation to not loop in rare case, update regressions, update casc scripts.
|
|
|
|
This commit fixes bug 821. As written in the description of the bug, the issue
is that `id` is not being set on one of the paths in addClause(), specifically
in the case where all but one literal are assigned false and the remaining
literal is assigned true. In that case, we are not actually adding anything and
set the `id` to `ClauseIdUndef`.
|
|
Fix uninitialized value
|
|
|
|
|
|
As pointed out by @cpitclaudel in pull request #172, we are using the same
handler for SIGSEGV and SIGILL and ill_handler() is unused. This commit changes
the SIGILL handler to ill_handler().
|
|
|
|
|
|
This commit fixes bug 811. Bug 811 was caused because tokens were referring to
a buffer that was reallocated and thus the pointers were not valid anymore.
Background:
The buffered input stream avoids copying the whole input stream before handing
it to ANTLR (in contrast to the non-buffered input stream that first copies
everything into a buffer). This enables interactivity (e.g. with kind2) and may
save memory.
CVC4 uses it when reading from stdin in competition mode for the application
track (the incremental benchmarks) and in non-competition mode. To set the
CVC4_SMTCOMP_APPLICATION_TRACK flag, the {C,CXX}FLAGS have to be modified at
configure time.
Solution:
This commit fixes the issue by changing how a stream gets buffered. Instead of
storing the stream into a single buffer, CVC4 now stores each line in a
separate buffer, making sure that they do not have to move, keeping tokens
valid. The commit adds the LineBuffer class for managing those buffers. It
further modifies CVC4's LA and consume functions to use line number and
position within a line to index into the line buffer. This allows us to use the
standard mark()/etc. functions because they automatically store and
restore that state. The solution also (arguably) simplifies the code.
Disadvantages:
Tokens split across lines would cause problems (seems reasonable to me). One
allocation per line.
Alternatives considered:
Pull request 162 by Tim was a first attempt to solve the problem. The issues
with this solution are: memory usage (old versions of the buffer do not get
deleted), tokens split across buffers would be problematic, and
mark()/rewind()/etc. would have to be overwritten for the approach to work.
I had a partially working fix that used indexes into the stream instead of
pointers to memory. The solution stored the content of the stream into a
segmented buffer (lines were not guaranteed to be consecutive in memory. This
approach was working for basic use cases but had the following issues: ugly
casting (the solution requires casting the index to a pointer and storing it in
the input stream's nextChar because that's where ANTLR is taking the location
information from when creating a token), more modifications (not only would
this solution require overwriting more functions of the input stream such as
substr, it also requires changes to the use of GETCHARINDEX() in the Smt2
parser and AntlrInput::tokenText() for example), more complex code.
|
|
Parse 'is', 'match' differently for non-DT input
|
|
This commit fixes bug 819 by making d_unitConflictId context dependent and adds a test case.
|
|
In SMT 2.6, Datatypes are being introduced and they come
with testers (indexed identifier of the form (_ is c)) and
match expressions. This lead to failures in UFIDL
benchmarks in SMT-LIB because they declare the function
'is'. This commit changes the parser s.t. it does not
consider 'is' and 'match' special tokens unless the theory
of datatypes is enabled.
|
|
|
|
|
|
|
|
cbqi for datatypes with uninterpreted sort subfields. Simplify fmc model construction.
|
|
This fixes bug 820. The issue was that (a udiv a) got rewriten to 1, which is
not correct when a is 0 (the result is all ones in that case). Even with the
--bv-div-zero-const flag disabled, the UdivSelf rewrite was incorrect because
it was applied to BITVECTOR_UDIV_TOTAL, which is "defined to be the all-ones
bit pattern, if divisor is 0" according to src/theory/bv/kinds . The commit
adds instead an optimization that returns all ones if the divisor of a
BITVECTOR_UDIV_TOTAL is zero.
|
|
|
|
|
|
|
|
|
|
proof infrastructure (e.g., quantifiers)
|
|
|
|
Minor changes to smt comp script.
|
|
in regressions.
|
|
added the option dump-unsat-cores-full for printing the entire core, as before
|
|
|
|
This commit fixes an issue where the ResChain in `d_resolutionChains` gets
deleted here:
https://github.com/CVC4/CVC4/blob/master/src/proof/sat_proof_implementation.h#L729
The condition immediately after is false because the condition on line 727 is
true. Thus, `d_resolutionChains` now has a deleted entry for `id`.
When CVC4 afterwards gets the ResChain associated with `id` in
`checkResolution()`, it accesses the deleted entry:
https://github.com/CVC4/CVC4/blob/master/src/proof/sat_proof_implementation.h#L303
|
|
|
|
|
|
Conditional coverage
|
|
Avoid tokenizing FP tokens in non-FP input
|
|
This commit addresses bug 807. CVC4 was parsing floating-point related tokens
such as NaN as floating-point tokens even for inputs that do not use the FP
theory, which lead to failing SMT-LIB benchmarks that declare functions named
NaN.
|
|
The Windows build was missing the `print_statistics()` function, this commit
moves the function out of the `#ifndef __WIN32__` guard.
|
|
Minor fix in safe_print function
|
|
This commit fixes two issues reported by Coverity:
- Fixes the check whether the buffer is full in safe_print_hex
- Removes dead code in safe_print for floating-point values
Additionally, it fixes an issue reported by Andy where the names of the
statistics were printed as "<unsupported>" due to calling the const char*
version instead of the std::string version of safe_print.
Finally, this fixes an issue where --segv-spin would not print the program name
because it was a const char*. The program name is now stored as a string.
NOTE: As a side effect, the last part also fixes Coverity issue 1362944, which
has been in CVC4 for a long time.
|
|
|
|
Fix type checks for relation operators
|
|
This commit fixes an assertion error when applying transpose or transitive
closure to a set instead of a relation. Instead it now prints a parse error.
|
|
Fix minor bug in sets rewriter
|
|
As reported by Coverity, one of the switches in the sets rewriter had a missing
break. This could lead to an assertion failure when rewriting the cardinality
of a transpose as in the test case included in this commit.
|
|
|
|
Fix condition in upwards closure check for sets
|
|
Coverity reported this mismatched iterator.
|
|
|
