/*++ Copyright (c) 2021 Microsoft Corporation Module Name: polysat unsigned <= constraints Author: Nikolaj Bjorner (nbjorner) 2021-03-19 Jakob Rath 2021-04-06 Notes: Canonical representation of equation p == 0 is the constraint p <= 0. The alternatives p < 1, -1 <= q, q > -2 are eliminated. Rewrite rules to simplify expressions. In the following let k, k1, k2 be values. - k1 <= k2 ==> 0 <= 0 if k1 <= k2 - k1 <= k2 ==> 1 <= 0 if k1 > k2 - 0 <= p ==> 0 <= 0 - p <= 0 ==> 1 <= 0 if p is never zero due to parity - p <= -1 ==> 0 <= 0 - k <= p ==> p - k <= - k - 1 - k*2^n*p <= 0 ==> 2^n*p <= 0 if k is odd, leading coeffient is always a power of 2. Note: the rules will rewrite alternative formulations of equations: - -1 <= p ==> p + 1 <= 0 - 1 <= p ==> p - 1 <= -2 Rewrite rules on signed constraints: - p > -2 ==> p + 1 <= 0 - p <= -2 ==> p + 1 > 0 At this point, all equations are in canonical form. TODO: clause simplifications: - p + k <= p ==> p + k <= k & p != 0 for k != 0 - p*q = 0 ==> p = 0 or q = 0 applies to any factoring - 2*p <= 2*q ==> (p >= 2^n-1 & q < 2^n-1) or (p >= 2^n-1 = q >= 2^n-1 & p <= q) ==> (p >= 2^n-1 => q < 2^n-1 or p <= q) & (p < 2^n-1 => p <= q) & (p < 2^n-1 => q < 2^n-1) - 3*p <= 3*q ==> ? Note: case p <= p + k is already covered because we test (lhs - rhs).is_val() It can be seen as an instance of lemma 5.2 of Supratik and John. The following forms are equivalent: p <= q p <= p - q - 1 q - p <= q q - p <= -p - 1 -q - 1 <= -p - 1 -q - 1 <= p - q - 1 Useful lemmas: p <= q && q+1 != 0 ==> p+1 <= q+1 p <= q && p != 0 ==> -q <= -p --*/ #include "sat/smt/polysat/polysat_constraints.h" #include "sat/smt/polysat/polysat_ule.h" #define LOG(_msg_) verbose_stream() << _msg_ << "\n" namespace polysat { // Simplify lhs <= rhs. // // NOTE: the result should not depend on the initial value of is_positive; // the purpose of is_positive is to allow flipping the sign as part of a rewrite rule. static void simplify_impl(bool& is_positive, pdd& lhs, pdd& rhs) { SASSERT_EQ(lhs.power_of_2(), rhs.power_of_2()); unsigned const N = lhs.power_of_2(); // 0 <= p --> 0 <= 0 if (lhs.is_zero()) { rhs = 0; return; } // p <= -1 --> 0 <= 0 if (rhs.is_max()) { lhs = 0, rhs = 0; return; } // p <= p --> 0 <= 0 if (lhs == rhs) { lhs = 0, rhs = 0; return; } // Evaluate constants if (lhs.is_val() && rhs.is_val()) { if (lhs.val() <= rhs.val()) lhs = 0, rhs = 0; else lhs = 0, rhs = 0, is_positive = !is_positive; return; } // Try to reduce the number of variables on one side using one of these rules: // // p <= q --> p <= p - q - 1 // p <= q --> q - p <= q // // Possible alternative to 1: // p <= q --> q - p <= -p - 1 // Possible alternative to 2: // p <= q --> -q-1 <= p - q - 1 // // Example: // // x <= x + y --> x <= - y - 1 // x + y <= x --> -y <= x if (!lhs.is_val() && !rhs.is_val()) { unsigned const lhs_vars = lhs.free_vars().size(); unsigned const rhs_vars = rhs.free_vars().size(); unsigned const diff_vars = (lhs - rhs).free_vars().size(); if (diff_vars < lhs_vars || diff_vars < rhs_vars) { LOG("reduce number of varables"); // verbose_stream() << "IN: " << ule_pp(to_lbool(is_positive), lhs, rhs) << "\n"; if (lhs_vars <= rhs_vars) rhs = lhs - rhs - 1; else lhs = rhs - lhs; // verbose_stream() << "OUT: " << ule_pp(to_lbool(is_positive), lhs, rhs) << "\n"; } } // -p + k <= k --> p <= k if (rhs.is_val() && !rhs.is_zero() && lhs.offset() == rhs.val()) { LOG("-p + k <= k --> p <= k"); lhs = rhs - lhs; } // k <= p + k --> p <= -k-1 if (lhs.is_val() && !lhs.is_zero() && lhs.val() == rhs.offset()) { LOG("k <= p + k --> p <= -k-1"); pdd k = lhs; lhs = rhs - lhs; rhs = -k - 1; } // k <= -p --> p-1 <= -k-1 if (lhs.is_val() && rhs.leading_coefficient().get_bit(N - 1) && !rhs.offset().is_zero()) { LOG("k <= -p --> p-1 <= -k-1"); pdd k = lhs; lhs = -(rhs + 1); rhs = -k - 1; } // -p <= k --> -k-1 <= p-1 // if (rhs.is_val() && lhs.leading_coefficient() > rational::power_of_two(N - 1) && !lhs.offset().is_zero()) { if (rhs.is_val() && lhs.leading_coefficient().get_bit(N - 1) && !lhs.offset().is_zero()) { LOG("-p <= k --> -k-1 <= p-1"); pdd k = rhs; rhs = -(lhs + 1); lhs = -k - 1; } // NOTE: do not use pdd operations in conditions when comparing pdd values. // e.g.: "lhs.offset() == (rhs + 1).val()" is problematic with the following evaluation: // 1. return reference into pdd_manager::m_values from lhs.offset() // 2. compute rhs+1, which may reallocate pdd_manager::m_values // 3. now the reference returned from lhs.offset() may be invalid pdd const rhs_plus_one = rhs + 1; // p - k <= -k - 1 --> k <= p // TODO: potential bug here: first call offset(), then rhs+1 has to reallocate pdd_manager::m_values, then the reference to offset is broken. if (rhs.is_val() && !rhs.is_zero() && lhs.offset() == rhs_plus_one.val()) { LOG("p - k <= -k - 1 --> k <= p"); pdd k = -(rhs + 1); rhs = lhs + k; lhs = k; } pdd const lhs_minus_one = lhs - 1; // k <= 2^(N-1)*p + q + k-1 --> k <= 2^(N-1)*p - q if (lhs.is_val() && rhs.leading_coefficient() == rational::power_of_two(N-1) && rhs.offset() == lhs_minus_one.val()) { LOG("k <= 2^(N-1)*p + q + k-1 --> k <= 2^(N-1)*p - q"); rhs = (lhs - 1) - rhs; } // -1 <= p --> p + 1 <= 0 if (lhs.is_max()) { lhs = rhs + 1; rhs = 0; } // 1 <= p --> p > 0 if (lhs.is_one()) { lhs = rhs; rhs = 0; is_positive = !is_positive; } // p > -2 --> p + 1 <= 0 // p <= -2 --> p + 1 > 0 if (rhs.is_val() && !rhs.is_zero() && (rhs + 2).is_zero()) { // Note: rhs.is_zero() iff rhs.manager().power_of_2() == 1 (the rewrite is not wrong for M=2, but useless) lhs = lhs + 1; rhs = 0; is_positive = !is_positive; } // 2p + 1 <= 0 --> 0 < 0 if (rhs.is_zero() && lhs.is_never_zero()) { lhs = 0; is_positive = !is_positive; return; } // a*p + q <= 0 --> p + a^-1*q <= 0 for a odd if (rhs.is_zero() && !lhs.leading_coefficient().is_power_of_two()) { rational lc = lhs.leading_coefficient(); rational x, y; gcd(lc, lhs.manager().two_to_N(), x, y); if (x.is_neg()) x = mod(x, lhs.manager().two_to_N()); lhs *= x; SASSERT(lhs.leading_coefficient().is_power_of_two()); } } // simplify_impl } namespace polysat { ule_constraint::ule_constraint(pdd const& l, pdd const& r) : m_lhs(l), m_rhs(r) { SASSERT_EQ(m_lhs.power_of_2(), m_rhs.power_of_2()); vars().append(m_lhs.free_vars()); for (auto v : m_rhs.free_vars()) if (!vars().contains(v)) vars().push_back(v); } void ule_constraint::simplify(bool& is_positive, pdd& lhs, pdd& rhs) { SASSERT_EQ(lhs.power_of_2(), rhs.power_of_2()); #ifndef NDEBUG bool const old_is_positive = is_positive; pdd const old_lhs = lhs; pdd const old_rhs = rhs; #endif simplify_impl(is_positive, lhs, rhs); #ifndef NDEBUG if (old_is_positive != is_positive || old_lhs != lhs || old_rhs != rhs) { ule_pp const old_ule(to_lbool(old_is_positive), old_lhs, old_rhs); ule_pp const new_ule(to_lbool(is_positive), lhs, rhs); // always-false and always-true constraints should be rewritten to 0 != 0 and 0 == 0, respectively. if (is_always_false(old_is_positive, old_lhs, old_rhs)) { SASSERT(!is_positive); SASSERT(lhs.is_zero()); SASSERT(rhs.is_zero()); } if (is_always_true(old_is_positive, old_lhs, old_rhs)) { SASSERT(is_positive); SASSERT(lhs.is_zero()); SASSERT(rhs.is_zero()); } } SASSERT(is_simplified(lhs, rhs)); // rewriting should be idempotent #endif } bool ule_constraint::is_simplified(pdd const& lhs0, pdd const& rhs0) { bool const pos0 = true; bool pos1 = pos0; pdd lhs1 = lhs0; pdd rhs1 = rhs0; simplify_impl(pos1, lhs1, rhs1); bool const is_simplified = (pos1 == pos0 && lhs1 == lhs0 && rhs1 == rhs0); DEBUG_CODE({ // check that simplification doesn't depend on initial sign bool pos2 = !pos0; pdd lhs2 = lhs0; pdd rhs2 = rhs0; simplify_impl(pos2, lhs2, rhs2); SASSERT_EQ(pos2, !pos1); SASSERT_EQ(lhs2, lhs1); SASSERT_EQ(rhs2, rhs1); }); return is_simplified; } std::ostream& ule_constraint::display(std::ostream& out, lbool status, pdd const& lhs, pdd const& rhs) { out << lhs; if (rhs.is_zero() && status == l_true) out << " == "; else if (rhs.is_zero() && status == l_false) out << " != "; else if (status == l_true) out << " <= "; else if (status == l_false) out << " > "; else out << " <=/> "; return out << rhs; } std::ostream& ule_constraint::display(std::ostream& out, lbool status) const { return display(out, status, m_lhs, m_rhs); } std::ostream& ule_constraint::display(std::ostream& out) const { return display(out, l_true, m_lhs, m_rhs); } // Evaluate lhs <= rhs lbool ule_constraint::eval(pdd const& lhs, pdd const& rhs) { // NOTE: don't assume simplifications here because we also call this on partially substituted constraints if (lhs.is_zero()) return l_true; // 0 <= p if (lhs == rhs) return l_true; // p <= p if (rhs.is_max()) return l_true; // p <= -1 if (rhs.is_zero() && lhs.is_never_zero()) return l_false; // p <= 0 implies p == 0 if (lhs.is_one() && rhs.is_never_zero()) return l_true; // 1 <= p implies p != 0 if (lhs.is_val() && rhs.is_val()) return to_lbool(lhs.val() <= rhs.val()); return l_undef; } lbool ule_constraint::eval() const { return eval(lhs(), rhs()); } lbool ule_constraint::eval(assignment const& a) const { return eval(a.apply_to(lhs()), a.apply_to(rhs())); } }