3
0
Fork 0
mirror of https://github.com/Z3Prover/z3 synced 2026-06-29 03:48:51 +00:00

feat(seq::bisim): hoist all ITEs to top in bisim derivative leaves

In the bisimulation equivalence path, derivative leaves are now
extracted via the new seq_rewriter::brz_derivative_cofactors
(derive::derivative_cofactors), which computes the symbolic
derivative and enumerates its reachable leaves in fully ITE-hoisted
normal form: every if-then-else over the input character (including
ones previously buried under a concat or union) is hoisted to the top
via decompose_ite, infeasible minterms are pruned, and unions are kept
intact as single states. Each leaf is therefore a ground regex free of
(:var 0), so its nullability is always decidable.

This replaces collect_leaves (which only split top-level ITEs and left
buried (:var 0) ITEs inside leaves), the root cause of bisim returning
l_undef and falling through to the slow theory solver.

Validation on the regex-equivalence corpus (1523 files, -T:5, 8 workers):
- vs master: total solved 1394 vs 1378 (+16), soft_timeouts 129 vs 145,
  0 soundness disagreements (was 18 -> 5 -> 0).
- vs derive: +242 solved (1394 vs 1152), 25.4% faster on commonly-solved
  files, fixes 18 soundness disagreements, only 6 regressions.
- corpus wall time halved (172s vs 332s/349s).
- All 91 unit tests pass, including seq_regex_bisim.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
This commit is contained in:
Margus Veanes 2026-06-25 08:40:32 +03:00
parent a2b73b0ee6
commit ee20c9963b
5 changed files with 52 additions and 42 deletions

View file

@ -204,6 +204,25 @@ namespace seq {
*/
void get_cofactors(expr* ele, expr* r, expr_ref_pair_vector& result);
/**
* Compute the symbolic derivative of r and enumerate its reachable
* leaves in fully ITE-hoisted normal form.
*
* Concretely this returns, for every feasible minterm (character
* class) of δ(r), a pair (path_condition, target_regex). Every
* if-then-else over the input character (including ones that would
* otherwise be buried under a concat/union) is hoisted to the top
* via the same path/interval pruning used by the derivative engine,
* so each target_regex is free of (:var 0) and its nullability is
* always decidable. Unions are kept intact as single leaves (a
* union leaf denotes a single bisimulation state). Infeasible
* minterms are pruned, so all returned leaves are reachable.
*
* This is the entry point the regex_bisim equivalence procedure
* uses: it consumes the target_regex of each pair and ignores the
* (redundant) path condition.
*/
void derivative_cofactors(expr* r, expr_ref_pair_vector& result);
void set_antimirov(bool flag) {
m_antimirov_derivative = flag;