mirrors/z3
3
0
Fork 0
mirror of https://github.com/Z3Prover/z3 synced 2026-05-16 23:25:36 +00:00
Code Activity

Add global backbones to parallel architecture (#9343)

Browse source 
* setting up new backbone experiment

* fix phase scores bug

* debug crash from negated atoms

* backbone thread/global backbones in progress, does NOT compile yet

* debug, still need to add backbones worker as a new thread

* setting up complicated condition variable thing for backbones worker thread

* debug

* debug lock contention

* it's a little messy, but change how i'm checking backbones by initiating with batch check

* don't split on global backbones, share global backbones once detected. still need to prune search tree with backbones

* close global backbone branches in search tree

* fix backbone ranking (take average of bb age over cubes and incorporate hits/num cubes the bb appears in

* add stats to backbone experiment

* gate the backbones experiment by local vs global

* update stats and fix bug about unsat core size=1 means global backbone

* phase negation ablation

* unforce phase ablation

* reset ablations

* add todo notes

* fix backbone aging

* first draft of Janota Alg 7

* process exactly 10 bb candidates in each batch

* fixing the Janota algorithm

* add backbone stats for Janota algorithm

* fix bug about global backbones not being checked unless local is also true

* hopefully fix bug about closing global backbones in search tree

* fix another bug in janota alg

* report random seed for debug

* print random seed for debug

* refactor janota alg code, still can't repro the crash

* fix some bugs in the janota algorithm

* try to fix weird memory leak thing with ramon/linux

* revert fix, it didn't work

* add second backbones thread

* increase chunk size when undef

* fix how the 2 backbone threads work on batches (they each race to finish the same batch). this was very complicated to code due to thread synchronization and while it runs there may be bugs

* update how we report stats for backbones

* first draft of doing the bb threads in neg and pos mode, needs revising

* fix some bugs in the positive version of the bb check, still need to review

* debug some more things in the positive bb worker

* keep bb candidates sorted, increase batch and chunk size

* try to resolve a couple of bugs

* fix very bad bug about backbones workers not doing anything

* ablate positive backbone thread

* fix how we record backbones in positive mode (shouldn't impact previous run)

* clarify code about adding found backbones

* add back the positive bb thread

* try to fix the random segfault bug + ablate the postiive bb thread again

* clean up logs

* share clauses with bb threads

* fixes

Signed-off-by: Nikolaj Bjorner <nbjorner@microsoft.com>

* resolve deadlock

* add comment about SAT bb case

* todo comments

* complete TODOs in code, still need to debug bb threads

* debug bb threads, add bb_positive thread back in

* ablate bb_positive thread

* style

* configure num bb threads as param

* enable sat and unsat mode

Signed-off-by: Nikolaj Bjorner <nbjorner@microsoft.com>

* updates

Signed-off-by: Nikolaj Bjorner <nbjorner@microsoft.com>

* remove while true

Signed-off-by: Nikolaj Bjorner <nbjorner@microsoft.com>

* updates

* try to fix rewriter_exception bug

* possibly reduce code under lock when only 1 bb thread

* add some copilot-suggested optimizations

* add copilot suggestions to fix condition variable synchronization with bb threads

* revert changes that are too messy with the code

* ablate collect clauses

* ablate condition variable logic changes

* ablate reset batch

* revert ablation

* remove m_batch_in_progress that makes the bb threads wait until both have exited the batch after one signals cancel (can be long if one is stuck in ctx check)

* sharing theory lemmas

* finish setup for search tree thread modes, and fix local bb setup to pull from the global pool

* variable renames

* update bb hyperparams after copilot (hopefully??) ran tuning experiments

* fix possible AST manager bug

* ablate collect clauses

* remove bb collect shared clauses

* fix local bb experiment bug and reinstate collect clauses for global bb

* local bb cands are thread-local ablation

* remove thread-local local bb ablation

* fix bug in nonthread-local bb experiment

* fix more nonthreadlocal bb bugs

* try to fix local bb bug

* AST manager mismatch bugfix

* attempt to fix another canonicalization bug

* try another bugfix

* try another bugfix

* try yet another bugfix

* thread local bb ablation

* ablate force phase

* ablate set activity

* undo ablations since apparently it's not forcing phase or boosting activities

* remove old experiments

* try guarding m_birthdate size

* try to fix several bugs including with m_birthdate initialization and how we're storing original phases

* one more bugfix

* remove local bb experiment after negative signals on experiments, and change bb ranking to VSIDS scores as opposed to phase

* select bb polarity based on phase, not VSIDS

* 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 smts bugfix git merge issues with backtrack

* 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

* fix merge

* revert codex change about exception handling

* fix linux bugs

* ablate backtrack gating

* attempt to fix linux crashes

* ablate backtracking on global bb

* the rare bb bug appears to be from creating the synthetic lease for a bb node and then backtracking on the synthetic lease. this is an attempt to fix it

* clean up code

* 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

* clean up merge from cherry pick

* tighten which nodes we detect for proven global bb closure (only detect nonclosed nodes)

* fix cancel to unknown exception on bb code

* 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.

* ablate lease cancel check in ctx final theory check due to crash (??)

* gate bb-specific code behind param

* try some possible bugfixes for the sigsev

* ablate some bugfixes

* remove second lease cancel check in smt_context, not sure it's safe. only check where we do the max conflicts check

* restore exception handling logic to master branch

* restore reslimit cancels since the bug appears to be latent

* add bookkeeping for race condition of multiple lease cancels on a single node (messes with reslimit)

* restore unrelated code to master

---------

Signed-off-by: Nikolaj Bjorner <nbjorner@microsoft.com>
Co-authored-by: Ilana Shapiro <ilanashapiro@Ilanas-MacBook-Pro.local>
Co-authored-by: Ilana Shapiro <ilanashapiro@Ilanas-MBP.lan1>
Co-authored-by: Nikolaj Bjorner <nbjorner@microsoft.com>
Co-authored-by: Ilana Shapiro <ilanashapiro@Mac.localdomain>
Co-authored-by: Ilana Shapiro <ilanashapiro@Ilanas-MBP.localdomain>
Co-authored-by: Ilana Shapiro <ilanashapiro@Mac.lan1>
This commit is contained in:
Ilana Shapiro 2026-04-20 09:22:07 -07:00 • committed by GitHub
parent 75039d631c
commit 19e95f40af
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
9 changed files with 750 additions and 33 deletions
Download patch file Download diff file Expand all files Collapse all files

17
src/util/search_tree.h
View file

@ -170,7 +170,7 @@ namespace search_tree {
// Used for tree expansion throttling policy in should_split()
// SMTS says set to num workers, but our experiments show a big regression
// Leaving at 0 for now, but making it confirgurable for future experimentation
// Leaving at 0 for now, but making it configurable for future experimentation
unsigned m_min_tree_size = 0;
struct candidate {
@ -524,6 +524,21 @@ namespace search_tree {
return m_root.get();
}
void find_nonclosed_nodes_with_literal(literal const& lit, ptr_vector<node<Config>>& out) {
find_nonclosed_nodes_with_literal_rec(m_root.get(), lit, out);
}
void find_nonclosed_nodes_with_literal_rec(node<Config>* n, literal const& lit, ptr_vector<node<Config>>& out) {
if (!n)
return;
if (!Config::literal_is_null(n->get_literal()) && n->get_literal() == lit && n->get_status() != status::closed)
out.push_back(n);
find_nonclosed_nodes_with_literal_rec(n->left(), lit, out);
find_nonclosed_nodes_with_literal_rec(n->right(), lit, out);
}
void dec_active_workers(node<Config>* n) {
if (!n)
return;