* first attempt with codex. Codex notes:
What changed:
- Each tree node now tracks:
- active worker count
- lease epoch
- cancel epoch
- get_cube() now hands each worker an explicit lease: (node, epoch, cancel_epoch).
- try_split() and backtrack() now operate on that lease, and the batch manager releases the worker’s lease under the tree lock before mutating the
node.
- If another worker closes the leased node or subtree, the batch manager cancels only the workers whose current leased nodes are now closed.
- Workers detect canceled leases after check(), reset their local cancel flag, abandon the stale lease, and continue instead of turning that into a
global exception.
- The “reopen immediately into the open queue” policy is preserved. I did not add a barrier waiting for all workers on a node to finish.
- Active-worker accounting is now separate from the open/active/closed scheduling status, so reopening a node no longer erases the fact that other
workers are still on it.
I also updated search_tree bookkeeping so:
- closure bumps node cancel/lease epochs
- active-node counting uses actual active-worker presence, not just status == active
* fix(parallel-smt): gate split/backtrack by lease epoch
What it changes:
- util/search_tree.h
- bumps node epoch on split
- threads epoch through should_split(...) and try_split(...)
- always records effort, but only split/reopen if the lease epoch still matches
- smt/smt_parallel.cpp
- requires is_lease_valid(..., lease.epoch) before backtrack(...)
- passes lease.epoch into m_search_tree.try_split(...)
* clean up code and add some comments
* fix bug about backtracking condition being too strict: The epoch guard should not block backtrack(...) the same way it blocks try_split(...). A stale worker that proves UNSAT for n should still be able to
close n, and that closure should then cancel the other workers on n and its subtree.
I changed smt/smt_parallel.cpp accordingly:
- try_split(...) still uses epoch to reject stale structural splits
- backtrack(...) no longer requires is_lease_valid(..., epoch); it only requires that the lease is not already canceled
So the intended asymmetry is now restored:
- stale split: reject
- stale unsat/backtrack: allow closure, then cancel affected workers
* ablate to no backtracking on stale leases
* revert codex change about exception handling
* remove old code
* ablate backtracking gate
* attempt to fix linux crashes
* try to fix bug about active worker counts/lease accounting. current policy should hold: - stale leases: release/decrement
- canceled leases: do not release/decrement (just ignore since we have an invariant that canceled leases mean closed nodes that are never revisited
* delay premature root activation
* fix major semantic bug about threads continually choosing the root if their lease is reset
* fix cancellation to unknown status
* fix very bad bug about all threads needing to start at the root
* ablate active ranking: now nodes are only reopened if they are truly inactive (active worker count is 0)
* fix some bugs about leases
* ablate adding static effort only
* fix some bugs about leases
* don't explode effort for portfolio nodes
* fix: still accumulate per-node effort, but don't over-accumulate on portfolio solves
* restore dynamically scaled effort
* lease cancellation doens't touch rlimit now, it just sets max conflicts to 0. also fix a VERY BAD BUG about effort never being updated until all leases are done on a node, which meant we never left the root
* cross-thread modification of max conflicts is unsafe, so create an atomic lease canceled variable that's ch
ecked in ctx where max conflicts is also checked
* move atomic lease check in the context to the more global get_cancel_flag function
* Fix new SIGSEV. issue: The root cause: get_cancel_flag() is called from within propagation loops (mid-BCP, mid-equality-propagation, mid-atom-propagation). When it returns true there, the solver exits early and leaves the context in an intermediate state —
propagation queues partially processed, theory state potentially inconsistent with boolean state.
For the global cancel (m.limit().cancel()), this is harmless: the worker exits entirely and the context is destroyed. Intermediate state doesn't matter.
For a lease cancel, the context is reused — the worker gets a new cube and calls ctx->check() again on the same context object. Re-entering check() on a context interrupted mid-propagation causes it to access that corrupted intermediate
state → SIGSEGV.
The m_max_conflicts check is the only checkpoint that's safe for re-entry: it only fires post-conflict-resolution, pre-decision, when propagation queues are empty and theory state is consistent.
Fix: Remove m_lease_canceled from get_cancel_flag(). Keep it only at safe, between-phase checkpoints where the context is in a known-consistent state. The result is two safe checkpoints for m_lease_canceled: after each conflict (post-resolution, queues empty) and before each theory final check (not yet entered the theory). Neither interrupts the solver mid-mutation. The SIGSEGV should be
gone, and NIA performance should improve because long theory final checks (where NIA burns most time) are now preemptable before they start.
* fix new inconsistent theory bug: The problem is returning FC_GIVEUP from inside final_check() after some theories have already run final_check_eh() and pushed propagations into the queue. Those pending propagations reference context state that gets invalidated on the next check() call → SIGSEGV. The fix: check m_lease_canceled before entering final_check() in bounded_search(), never from inside it. That way the context is always in a clean pre-final-check state when we bail out. This is safe: decide() returned false (all variables assigned, no pending propagations), theories haven't been touched yet, context is in a fully consistent state. For NIA, this is still a meaningful win — we avoid entering expensive arithmetic final checks entirely when the lease is already canceled.
* remove second lease cancel check in smt_context, not sure it's safe. only check where we do the max conflicts check
* check epoch match in release_lease_unlocked
* restore exception handling logic to master branch
* restore reslimit cancels since the bug appears to be latent
---------
Co-authored-by: Ilana Shapiro <ilanashapiro@Ilanas-MacBook-Pro.local>
Co-authored-by: Ilana Shapiro <ilanashapiro@Mac.lan1>
Fixes a double-free (SIGSEGV in mpz_manager::del) in
algebraic_numbers::manager:👿:del_poly, reached through the
destruction of nlsat::evaluator's scoped_anum_vector members on a
subsequent call to nra::solver:👿:reset.
Root cause: sort_roots runs std::sort over a numeral_vector with a
comparator (lt_proc -> manager::lt -> compare_core) that legitimately
throws when the reslimit fires mid-comparison. libc++'s insertion sort
shifts elements via move-assignment inside its inner loop, and because
anum previously had only compiler-generated shallow copy/move (both
just copied m_cell without nulling the source), a throw between two
consecutive shifts could leave two vector slots pointing at the same
algebraic_cell. When the owning scoped_anum_vector was later destroyed
it del'd the same cell twice, reading through a freed chunk whose
first bytes had been overwritten by small_object_allocator's free-list
next pointer.
Fix: give anum proper move constructor and move assignment that
transfer the tagged m_cell pointer and null the source. Copy stays
a shallow handle copy (ownership is still tracked externally by the
manager / owning vector, as before). With the new move, every
intermediate state of sort's move-via-tmp sequence has at most one
slot referencing any given cell, so a throwing comparator can leak
the in-flight tmp cell but cannot produce aliased slots and therefore
cannot cause the downstream double-free.
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* Fix broken term_comparer in m_normalized_terms_to_columns lookup
The `m_normalized_terms_to_columns` map in `lar_solver` uses a
`term_comparer` that delegates to `lar_term::operator==`, which
intentionally returns `false` (with comment "take care not to create
identical terms"). This makes `fetch_normalized_term_column` unable to
find any term, rendering the Horner module's `interval_from_term`
bounds-recovery path dead code.
History: `lar_term::operator==` returning `false` has been present since
the original "merge LRA" commit (911b24784, 2018). The
`m_normalized_terms_to_columns` lookup was added later (dfe0e856,
c95f66e0, Aug 2019) as "toward fetching existing terms intervals from
lar_solver". The initial code had `lp_assert(find == end)` on
registration (always true with broken ==) and `lp_assert(find != end)`
on deregister (always false). The very next commit (207c1c50, one day
later) removed both asserts, replacing them with soft checks. The
`term_comparer` struct delegating to `operator==` was introduced during
a later PIMPL refactor (b375faa77).
Fix: Replace the `term_comparer` implementation with a structural
comparison that checks size and then verifies each coefficient-variable
pair via `coeffs().find_core()`. This is localized to the
`m_normalized_terms_to_columns` map and does not change
`lar_term::operator==`, preserving its intentional semantics elsewhere.
Validated: on a QF_UFNIA benchmark, `interval_from_term` lookups go
from 0/573 successful to 34/573 successful. Unit test added for the
`fetch_normalized_term_column` round-trip.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* Disable operator== for lar_term
The operator== for lar_term was never intended to be used.
This changes physically disables it to identify what happens to depend
on the operator.
* Work around missing lar_term==
Previous commit disabled lar_term==. This is the only use of the
operator that seems meaningful. Changed it to compare by references
instead.
Compiles, but not sure this is the best solution.
* replace with e
Signed-off-by: Nikolaj Bjorner <nbjorner@microsoft.com>
* Delete unused ineq::operator==
The operator is unused, so there is no need to figure what is
the best fix for it.
* Remove lp tests that use ineq::operator==
---------
Signed-off-by: Nikolaj Bjorner <nbjorner@microsoft.com>
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Co-authored-by: Nikolaj Bjorner <nbjorner@microsoft.com>
* fix: address issues 1,2,4,5 and add Goal API to Go bindings
Issue 2 (Go): Add Substitute, SubstituteVars, SubstituteFuns to Expr
Issue 4 (Go): Add GetDecl, NumArgs, Arg to Expr for AST app introspection
Goal API (Go): Add IsInconsistent and ToDimacsString to Goal
ASTVector (Go): Add public Size, Get, String methods
ASTMap (Go): Add ASTMap type with full CRUD API in spacer.go
Issue 1 (Go): Add Spacer fixedpoint methods QueryFromLvl, GetGroundSatAnswer,
GetRulesAlongTrace, GetRuleNamesAlongTrace, AddInvariant, GetReachable
Issue 1 (Go): Add context-level QE functions ModelExtrapolate, QeLite,
QeModelProject, QeModelProjectSkolem, QeModelProjectWithWitness
Issue 5 (OCaml): Add substitute_funs to z3.ml and z3.mli
Agent-Logs-Url: https://github.com/Z3Prover/z3/sessions/afa18588-47af-4720-8cea-55fe0544ae55
Co-authored-by: NikolajBjorner <3085284+NikolajBjorner@users.noreply.github.com>
* fix: add substitute_funs to Expr module sig in z3.ml
The internal sig...end block in z3.ml (the module type declaration for Expr)
was missing val substitute_funs, causing OCaml compiler error:
The value substitute_funs is required but not provided
Agent-Logs-Url: https://github.com/Z3Prover/z3/sessions/c6662702-46a3-4aa0-b225-d6b73c2a2505
Co-authored-by: NikolajBjorner <3085284+NikolajBjorner@users.noreply.github.com>
---------
Co-authored-by: copilot-swe-agent[bot] <198982749+Copilot@users.noreply.github.com>
Co-authored-by: NikolajBjorner <3085284+NikolajBjorner@users.noreply.github.com>
The is_mod handler in theory_lra called ensure_nla(), which
unnecessarily created the NLA solver for pure linear problems, causing
the optimizer to return a finite value instead of -infinity.
Fix: check `m_nla` instead of calling `ensure_nla()`, matching the
pattern used by the is_idiv handler. The mod division is only registered
when NLA is already active due to nonlinear terms.
Update mod_factor tests to use QF_NIA logic and assert the mul term
before the mod term so that internalize_mul triggers ensure_nla() before
mod internalization.
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* Refactor parallel search tree to use global node selection (SMTS-style) instead of DFS traversal.
Introduce effort-based prioritization, allow activation of any open node, and add controlled/gated
expansion to prevent over-partitioning and improve load balancing.
* clean up code
* ablations
* ablations2: effort
* ablations2: activation
* ablations3: more activations
* ablations4: visit all nodes before splitting
* throttle tree size min is based on workers not activated nodes
* ablate random throttling
* ablate nonlinear effort
* clean up code
* ablate throttle
* ablate where add_effort is
* reset
* clean up a function and add comment
---------
Co-authored-by: Ilana Shapiro <ilanashapiro@Ilanas-MBP.localdomain>
Co-authored-by: Ilana Shapiro <ilanashapiro@Ilanas-MacBook-Pro.local>
Co-authored-by: Ilana Shapiro <ilanashapiro@Ilanas-MBP.lan1>
When a monic x*y has a factor x with mod(x, p) = 0 (fixed), propagate
mod(x*y, p) = 0. This enables Z3 to prove divisibility properties like
x mod p = 0 => (x*y) mod p = 0, which previously timed out even for
p = 2. The lemma fires in the NLA divisions check and allows Gröbner
basis and LIA to subsequently derive distributivity of div over addition.
Extends division tuples from (q, x, y) to (q, x, y, r) to track the
mod lpvar. Also registers bounded divisions from the mod internalization
path in theory_lra, not just the idiv path.
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
root_function's move constructor and move assignment were doing deep
copies of algebraic numbers via anum_manager::set() instead of proper
moves. During std::vector reallocation (emplace) and std::sort, this
caused massive allocation churn that corrupted the heap.
Fixes:
1. Move constructor: use std::move(other.val) for proper swap semantics.
2. Move assignment: use val.swap(other.val) instead of deep copy.
3. Add friend swap() for ADL so std::sort uses efficient swaps.
4. Sort root_function partitions via index permutation + swap cycles
instead of std::sort directly on root_function objects.
5. Reserve rfunc vector before emplace in add_linear_approximations().
6. Reserve lhalf/uhalf vectors before collecting root functions.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
The m_cache.reset(), m_assignment.reset(), m_lo.reset(), m_hi.reset()
calls added to clear() in commit 481eb0327 cause heap corruption when
clear() is called from the destructor. The cache reset frees polynomials
while the polynomial manager still holds references to them, corrupting
the heap. This manifests as 'corrupted double-linked list' crashes
during nlsat solver destruction in the nra check path.
The reset() method already has these calls for solver reuse. The
destructor path via clear() should not duplicate them, as implicit
member destruction handles cleanup in the correct order.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Reset polynomial cache and assignments in nlsat::solver:👿:clear()
to prevent use-after-free when the solver is destroyed. The missing
resets caused heap corruption when check_assignment's
compute_linear_explanation created cached polynomials and root atoms
that outlived the solver's other data structures during destruction.
Also fix _scoped_numeral_vector copy constructor to read from other
instead of uninitialized self.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Calling factor_sqf_pp recursively on Hensel-lifted factors corrupts
shared mutable state in the polynomial manager, m_m2pos, m_som_buffer,
m_cheap_som_buffer, m_tmp1, etc., causing assertion violations:
- polynomial.cpp:473 id < m_m2pos.size()
- upolynomial.cpp:2624 sign_a == -sign_b
Use factor_1_sqf_pp/factor_2_sqf_pp for small degrees, push directly
for larger degrees. These don't conflict with the outer call's buffers.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
- Fix memory leaks: use scoped_numeral instead of raw numeral for
evaluation points, ensuring cleanup on exceptions
- Precompute lc_inv before the Hensel lifting loop instead of
recomputing each iteration
- Use scoped_numeral_vector for eval_vals for consistency with codebase
- Move eval_values and candidate_primes to static constexpr class-level
- Document limitations: single-prime Hensel lifting, contiguous factor
splits only, pseudo-division lc-power caveat
- Condense Bezout derivation comment to 4-line summary
- Fix README to say Hensel lifting instead of GCD recovery
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Replace the stub factor_n_sqf_pp (TODO: invoke Dejan's procedure) with a
working implementation using bivariate Hensel lifting:
- Evaluate away extra variables to reduce to bivariate
- Factor the univariate specialization
- Lift univariate factors to bivariate via linear Hensel lifting in Zp[x]
- Verify lifted factors multiply to original over Z[x,y]
- For >2 variables, check bivariate factors divide the original polynomial
Tests: (x0+x1)(x0+2x1)(x0+3x1) now correctly factors into 3 linear factors.
All 89 unit tests pass in both release and debug builds.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
* Initial plan
* simplify extract_var_bound in qe_lite_tactic.cpp via operator normalization
Co-authored-by: NikolajBjorner <3085284+NikolajBjorner@users.noreply.github.com>
* Add defensive check for integer type in lhs
Added a defensive check for integer type in lhs before proceeding with inequality checks.
* Update qe_lite_tactic.cpp
* Fix utility function call for integer check
---------
Co-authored-by: copilot-swe-agent[bot] <198982749+Copilot@users.noreply.github.com>
Co-authored-by: NikolajBjorner <3085284+NikolajBjorner@users.noreply.github.com>
Co-authored-by: Nikolaj Bjorner <nbjorner@microsoft.com>
update_lower_lex updates m_lower for subsequent objectives with saved
values from the current model. Reset m_lower[i] and m_upper[i] to
their initial values before optimizing each objective so earlier
objectives do not contaminate later ones.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Z3_ast_to_string returns a pointer to an internal buffer that is
overwritten on the next call. Store results in std::string immediately
to avoid reading a stale, garbled buffer.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
geometric_lex's update_lower_lex updates m_lower for all subsequent
objectives with saved values from the current model. In box mode this
contaminates later objectives' starting bounds, causing platform-dependent
results. Save and restore m_lower/m_upper across iterations so each
objective starts from a clean state.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
In box mode (opt.priority=box), each objective should be optimized
independently. Previously, box() called geometric_opt() which optimizes
all objectives together using a shared disjunction of bounds. This caused
adding/removing an objective to change the optimal values of other
objectives.
Fix: Rewrite box() to optimize each objective in its own push/pop scope
using geometric_lex, ensuring complete isolation between objectives.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
resultant vanishes during a nested isolate_roots call. The mathematical
invariant that the resultant cannot vanish again after recovery does not
hold in all cases, e.g. with certain nonlinear real arithmetic formulas.
The algebraic_exception propagates cleanly through the nlsat solver and
tactic layers which already catch z3_exception.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
After qe-light's equation solver (eq_der) eliminates variables from
linear equations, remaining bounded integer quantifiers may still have
non-unit coefficients that prevent Fourier-Motzkin elimination.
Add a bounded quantifier expansion step: when the remaining quantified
integer variables all have explicit finite bounds and the product of
domain sizes is <= 10000, expand the quantifier into a finite
disjunction. This turns e.g. exists y0 in [0,10), y1 in [0,15): P(x,y0,y1)
into P(x,0,0) | P(x,0,1) | ... | P(x,9,14), which is 150 disjuncts.
The SMT solver handles the resulting quantifier-free formula instantly,
whereas the previous QSAT/MBP approach timed out due to weak integer
projections from the (|a|-1)*(|b|-1) slack in Fourier-Motzkin resolution
with non-unit coefficients.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Always print each test's captured output, not just for failures.
This preserves backward compatibility:
- PASS appears on its own line per test, as before
- ASAN/UBSAN reports from any test appear in captured logs
- timeit output is preserved for all tests
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Parallel mode (/j) is now the default. Use /seq to force serial execution.
Child processes are invoked with /seq to prevent recursive parallelism.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Refactor src/test/main.cpp to support parallel test execution:
- Add /j[:N] flag to run tests in parallel using N jobs (default: number of cores)
- Use process-based parallelism: each test runs as a child process,
avoiding thread-safety issues with global state like enable_debug/enable_trace
- Output is captured per-test and printed atomically, so different tests never mix
- Provide summary with pass/fail counts, wall time, and failed test names
- Refactor test list into X-macros for single source of truth
- Fix pre-existing bug where serial /a mode ran each test argc times
Platform support:
- Unix (Linux/macOS/FreeBSD): popen/pclose with WEXITSTATUS
- Windows: _popen/_pclose
- Emscripten: parallel disabled (no threading support)
- Works with both SINGLE_THREAD and multi-threaded builds
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
m_fixed.insert(e) was placed before the check_long_strings guard,
causing check_fixed_length(false, false) to mark variables with
len > 20 as processed without actually decomposing them. The
subsequent check_fixed_length(false, true) then skipped them.
Move the insertion after the guard so variables are only marked
as fixed once they are actually decomposed.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
The add_axiom optimization that skips adding clauses when a literal is
already true was unsound: the satisfying literal could be retracted by
backtracking, leaving the axiom clause missing. This caused the solver
to miss propagations, e.g., not propagating indexof(a,s) = -1 when
contains(a,s) becomes false after backtracking.
Fix: only skip the clause if the satisfying literal is assigned at
base level (scope 0), where it can never be retracted.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
In multi-threaded solving, IF_VERBOSE(0, ...) in found_non_diff_logic_expr
was always acquiring the global g_verbose_mux mutex (since verbosity >= 0 is
always true) while holding it for potentially expensive mk_pp() calls. This
caused catastrophic lock contention when multiple threads internalized atoms.
Change IF_VERBOSE(0, ...) to IF_VERBOSE(2, ...) in both theory_diff_logic_def.h
and theory_dense_diff_logic_def.h. The diagnostic message is still available at
verbosity level 2 (-v:2), but is no longer printed (or locked) at the default
verbosity level, eliminating the contention.
Co-authored-by: levnach <5377127+levnach@users.noreply.github.com>
Test tst_box_mod_opt verifies that maximize (mod (- (* 232 a)) 256)
returns 248 when using box priority with multiple objectives.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Push/pop isolation in maximize_objectives1 (added for #7677) can corrupt
LP column values between objectives. For non-linear objectives like mod,
the LP maximize call may return stale values after a preceding
objective's push/pop cycle.
Fix: save the baseline model before the push/pop loop and use it as a
floor for each objective's value. Extract two helpers:
- maximize_objective_isolated: push/pop + save/restore per objective
- update_from_baseline_model: adopt baseline model value when it is better
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
When the LP optimizer returns the same blocker expression in successive
iterations of geometric_lex (e.g., due to nonlinear constraints like
mod/to_int preventing the LP from exploring the full feasible region),
the loop now falls back to using the model-based lower bound to push
harder instead of breaking immediately.
This fixes the case where minimize(3*a) incorrectly returned -162
while minimize(a) correctly returned -infinity with the same constraints.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
