/*++ Copyright (c) 2015 Microsoft Corporation Module Name: qe_arith.cpp Abstract: Simple projection function for real arithmetic based on Loos-W. Author: Nikolaj Bjorner (nbjorner) 2013-09-12 Revision History: Moved projection functionality to model_based_opt module. 2016-06-26 --*/ #include "qe/mbp/mbp_arith.h" #include "ast/ast_util.h" #include "ast/arith_decl_plugin.h" #include "ast/ast_pp.h" #include "ast/rewriter/th_rewriter.h" #include "ast/expr_functors.h" #include "ast/rewriter/expr_safe_replace.h" #include "math/simplex/model_based_opt.h" #include "model/model_evaluator.h" #include "model/model_smt2_pp.h" #include "model/model_v2_pp.h" namespace mbp { struct arith_project_plugin::imp { ast_manager& m; arith_util a; bool m_check_purified { true }; // check that variables are properly pure bool m_apply_projection { false }; imp(ast_manager& m) : m(m), a(m) {} ~imp() {} void insert_mul(expr* x, rational const& v, obj_map& ts) { // TRACE("qe", tout << "Adding variable " << mk_pp(x, m) << " " << v << "\n";); rational w; if (ts.find(x, w)) ts.insert(x, w + v); else ts.insert(x, v); } // // extract linear inequalities from literal 'lit' into the model-based optimization manager 'mbo'. // It uses the current model to choose values for conditionals and it primes mbo with the current // interpretation of sub-expressions that are treated as variables for mbo. // bool linearize(opt::model_based_opt& mbo, model_evaluator& eval, expr* lit, expr_ref_vector& fmls, obj_map& tids) { obj_map ts; rational c(0), mul(1); expr_ref t(m); opt::ineq_type ty = opt::t_le; expr* e1, *e2; DEBUG_CODE(expr_ref val(m); eval(lit, val); CTRACE("qe", !m.is_true(val), tout << mk_pp(lit, m) << " := " << val << "\n";); SASSERT(m.limit().is_canceled() || !m.is_false(val));); if (!m.inc()) return false; TRACE("opt", tout << mk_pp(lit, m) << " " << a.is_lt(lit) << " " << a.is_gt(lit) << "\n";); bool is_not = m.is_not(lit, lit); if (is_not) { mul.neg(); } SASSERT(!m.is_not(lit)); if ((a.is_le(lit, e1, e2) || a.is_ge(lit, e2, e1))) { linearize(mbo, eval, mul, e1, c, fmls, ts, tids); linearize(mbo, eval, -mul, e2, c, fmls, ts, tids); ty = is_not ? opt::t_lt : opt::t_le; } else if ((a.is_lt(lit, e1, e2) || a.is_gt(lit, e2, e1))) { linearize(mbo, eval, mul, e1, c, fmls, ts, tids); linearize(mbo, eval, -mul, e2, c, fmls, ts, tids); ty = is_not ? opt::t_le: opt::t_lt; } else if (m.is_eq(lit, e1, e2) && !is_not && is_arith(e1)) { linearize(mbo, eval, mul, e1, c, fmls, ts, tids); linearize(mbo, eval, -mul, e2, c, fmls, ts, tids); ty = opt::t_eq; } else if (m.is_eq(lit, e1, e2) && is_not && is_arith(e1)) { rational r1, r2; expr_ref val1 = eval(e1); expr_ref val2 = eval(e2); //TRACE("qe", tout << mk_pp(e1, m) << " " << val1 << "\n";); //TRACE("qe", tout << mk_pp(e2, m) << " " << val2 << "\n";); if (!a.is_numeral(val1, r1)) return false; if (!a.is_numeral(val2, r2)) return false; SASSERT(r1 != r2); if (r1 < r2) { std::swap(e1, e2); } ty = opt::t_lt; linearize(mbo, eval, mul, e1, c, fmls, ts, tids); linearize(mbo, eval, -mul, e2, c, fmls, ts, tids); } else if (m.is_distinct(lit) && !is_not && is_arith(to_app(lit)->get_arg(0))) { expr_ref val(m); rational r; app* alit = to_app(lit); vector > nums; for (expr* arg : *alit) { val = eval(arg); TRACE("qe", tout << mk_pp(arg, m) << " " << val << "\n";); if (!a.is_numeral(val, r)) return false; nums.push_back(std::make_pair(arg, r)); } std::sort(nums.begin(), nums.end(), compare_second()); for (unsigned i = 0; i + 1 < nums.size(); ++i) { SASSERT(nums[i].second < nums[i+1].second); expr_ref fml(a.mk_lt(nums[i].first, nums[i+1].first), m); if (!linearize(mbo, eval, fml, fmls, tids)) { return false; } } return true; } else if (m.is_distinct(lit) && is_not && is_arith(to_app(lit)->get_arg(0))) { // find the two arguments that are equal. // linearize these. map values; bool found_eq = false; for (unsigned i = 0; !found_eq && i < to_app(lit)->get_num_args(); ++i) { expr* arg1 = to_app(lit)->get_arg(i), *arg2 = nullptr; rational r; expr_ref val = eval(arg1); TRACE("qe", tout << mk_pp(arg1, m) << " " << val << "\n";); if (!a.is_numeral(val, r)) return false; if (values.find(r, arg2)) { ty = opt::t_eq; linearize(mbo, eval, mul, arg1, c, fmls, ts, tids); linearize(mbo, eval, -mul, arg2, c, fmls, ts, tids); found_eq = true; } else { values.insert(r, arg1); } } SASSERT(found_eq); } else { TRACE("qe", tout << "Skipping " << mk_pp(lit, m) << "\n";); return false; } vars coeffs; extract_coefficients(mbo, eval, ts, tids, coeffs); mbo.add_constraint(coeffs, c, ty); return true; } // // convert linear arithmetic term into an inequality for mbo. // void linearize(opt::model_based_opt& mbo, model_evaluator& eval, rational const& mul, expr* t, rational& c, expr_ref_vector& fmls, obj_map& ts, obj_map& tids) { expr* t1, *t2, *t3; rational mul1; expr_ref val(m); if (a.is_mul(t, t1, t2) && is_numeral(t1, mul1)) linearize(mbo, eval, mul* mul1, t2, c, fmls, ts, tids); else if (a.is_mul(t, t1, t2) && is_numeral(t2, mul1)) linearize(mbo, eval, mul* mul1, t1, c, fmls, ts, tids); else if (a.is_uminus(t, t1)) linearize(mbo, eval, -mul, t1, c, fmls, ts, tids); else if (a.is_numeral(t, mul1)) c += mul * mul1; else if (a.is_add(t)) { for (expr* arg : *to_app(t)) linearize(mbo, eval, mul, arg, c, fmls, ts, tids); } else if (a.is_sub(t, t1, t2)) { linearize(mbo, eval, mul, t1, c, fmls, ts, tids); linearize(mbo, eval, -mul, t2, c, fmls, ts, tids); } else if (m.is_ite(t, t1, t2, t3)) { val = eval(t1); TRACE("qe", tout << mk_pp(t1, m) << " := " << val << "\n";); if (m.is_true(val)) { linearize(mbo, eval, mul, t2, c, fmls, ts, tids); fmls.push_back(t1); } else if (m.is_false(val)) { expr_ref not_t1(mk_not(m, t1), m); fmls.push_back(not_t1); linearize(mbo, eval, mul, t3, c, fmls, ts, tids); } else { throw default_exception("mbp evaluation didn't produce a truth value"); } } else if (a.is_mod(t, t1, t2) && is_numeral(t2, mul1) && !mul1.is_zero()) { rational r; val = eval(t); if (!a.is_numeral(val, r)) { throw default_exception("mbp evaluation didn't produce an integer"); } c += mul*r; // t1 mod mul1 == r rational c0(-r), mul0(1); obj_map ts0; linearize(mbo, eval, mul0, t1, c0, fmls, ts0, tids); vars coeffs; extract_coefficients(mbo, eval, ts0, tids, coeffs); mbo.add_divides(coeffs, c0, mul1); } else { insert_mul(t, mul, ts); } } bool is_numeral(expr* t, rational& r) { return a.is_extended_numeral(t, r); } struct compare_second { bool operator()(std::pair const& a, std::pair const& b) const { return a.second < b.second; } }; bool is_arith(expr* e) { return a.is_int_real(e); } rational n_sign(rational const& b) { return rational(b.is_pos()?-1:1); } bool operator()(model& model, app* v, app_ref_vector& vars, expr_ref_vector& lits) { app_ref_vector vs(m); vs.push_back(v); project(model, vs, lits, false); return vs.empty(); } typedef opt::model_based_opt::var var; typedef opt::model_based_opt::row row; typedef vector vars; expr_ref var2expr(ptr_vector const& index2expr, var const& v) { expr_ref t(index2expr[v.m_id], m); if (!v.m_coeff.is_one()) { t = a.mk_mul(a.mk_numeral(v.m_coeff, a.is_int(t)), t); } return t; } vector project(model& model, app_ref_vector& vars, expr_ref_vector& fmls, bool compute_def) { bool has_arith = false; for (expr* v : vars) has_arith |= is_arith(v); if (!has_arith) return vector(); model_evaluator eval(model); TRACE("qe", tout << model;); eval.set_model_completion(true); compute_def |= m_apply_projection; opt::model_based_opt mbo; obj_map tids; expr_ref_vector pinned(m); unsigned j = 0; TRACE("qe", tout << "vars: " << vars << "\n"; for (expr* f : fmls) tout << mk_pp(f, m) << " := " << model(f) << "\n";); for (unsigned i = 0; i < fmls.size(); ++i) { expr* fml = fmls.get(i); if (!linearize(mbo, eval, fml, fmls, tids)) { TRACE("qe", tout << "could not linearize: " << mk_pp(fml, m) << "\n";); fmls[j++] = fml; } else { pinned.push_back(fml); } } fmls.shrink(j); TRACE("qe", tout << "formulas\n" << fmls << "\n"; for (auto [e, id] : tids) tout << mk_pp(e, m) << " -> " << id << "\n";); // fmls holds residue, // mbo holds linear inequalities that are in scope // collect variables in residue an in tids. // filter variables that are absent from residue. // project those. // collect result of projection // return those to fmls. expr_mark var_mark, fmls_mark; for (app * v : vars) { var_mark.mark(v); if (is_arith(v) && !tids.contains(v)) { rational r; expr_ref val = eval(v); if (!m.inc()) return vector(); if (!a.is_numeral(val, r)) throw default_exception("evaluation did not produce a numeral"); TRACE("qe", tout << mk_pp(v, m) << " " << val << "\n";); tids.insert(v, mbo.add_var(r, a.is_int(v))); } } if (m_check_purified) { for (expr* fml : fmls) mark_rec(fmls_mark, fml); for (auto& kv : tids) { expr* e = kv.m_key; if (!var_mark.is_marked(e)) { mark_rec(fmls_mark, e); } } } ptr_vector index2expr; for (auto& kv : tids) index2expr.setx(kv.m_value, kv.m_key, nullptr); j = 0; unsigned_vector real_vars; for (app* v : vars) { if (is_arith(v) && !fmls_mark.is_marked(v)) real_vars.push_back(tids.find(v)); else vars[j++] = v; } vars.shrink(j); TRACE("qe", tout << "remaining vars: " << vars << "\n"; for (unsigned v : real_vars) tout << "v" << v << " " << mk_pp(index2expr[v], m) << "\n"; mbo.display(tout);); vector defs = mbo.project(real_vars.size(), real_vars.data(), compute_def); vector rows; mbo.get_live_rows(rows); rows2fmls(rows, index2expr, fmls); TRACE("qe", mbo.display(tout << "mbo result\n"); for (auto const& d : defs) tout << "def: " << d << "\n"; tout << fmls << "\n";); vector result; if (compute_def) optdefs2mbpdef(defs, index2expr, real_vars, result); if (m_apply_projection) apply_projection(result, fmls); return result; } void optdefs2mbpdef(vector const& defs, ptr_vector const& index2expr, unsigned_vector const& real_vars, vector& result) { SASSERT(defs.size() == real_vars.size()); for (unsigned i = 0; i < defs.size(); ++i) { auto const& d = defs[i]; expr* x = index2expr[real_vars[i]]; bool is_int = a.is_int(x); expr_ref_vector ts(m); expr_ref t(m); for (var const& v : d.m_vars) ts.push_back(var2expr(index2expr, v)); if (!d.m_coeff.is_zero()) ts.push_back(a.mk_numeral(d.m_coeff, is_int)); if (ts.empty()) ts.push_back(a.mk_numeral(rational(0), is_int)); t = mk_add(ts); if (!d.m_div.is_one() && is_int) t = a.mk_idiv(t, a.mk_numeral(d.m_div, is_int)); else if (!d.m_div.is_one() && !is_int) t = a.mk_div(t, a.mk_numeral(d.m_div, is_int)); result.push_back(def(expr_ref(x, m), t)); } } void rows2fmls(vector const& rows, ptr_vector const& index2expr, expr_ref_vector& fmls) { for (row const& r : rows) { expr_ref_vector ts(m); expr_ref t(m), s(m), val(m); if (r.m_vars.empty()) { continue; } if (r.m_vars.size() == 1 && r.m_vars[0].m_coeff.is_neg() && r.m_type != opt::t_mod) { var const& v = r.m_vars[0]; t = index2expr[v.m_id]; if (!v.m_coeff.is_minus_one()) { t = a.mk_mul(a.mk_numeral(-v.m_coeff, a.is_int(t)), t); } s = a.mk_numeral(r.m_coeff, a.is_int(t)); switch (r.m_type) { case opt::t_lt: t = a.mk_gt(t, s); break; case opt::t_le: t = a.mk_ge(t, s); break; case opt::t_eq: t = a.mk_eq(t, s); break; default: UNREACHABLE(); } fmls.push_back(t); continue; } for (var const& v : r.m_vars) { t = index2expr[v.m_id]; if (!v.m_coeff.is_one()) { t = a.mk_mul(a.mk_numeral(v.m_coeff, a.is_int(t)), t); } ts.push_back(t); } t = mk_add(ts); s = a.mk_numeral(-r.m_coeff, r.m_coeff.is_int() && a.is_int(t)); switch (r.m_type) { case opt::t_lt: t = a.mk_lt(t, s); break; case opt::t_le: t = a.mk_le(t, s); break; case opt::t_eq: t = a.mk_eq(t, s); break; case opt::t_mod: if (!r.m_coeff.is_zero()) { t = a.mk_sub(t, s); } t = a.mk_eq(a.mk_mod(t, a.mk_numeral(r.m_mod, true)), a.mk_int(0)); break; } fmls.push_back(t); } } expr_ref mk_add(expr_ref_vector const& ts) { return a.mk_add_simplify(ts); } opt::inf_eps maximize(expr_ref_vector const& fmls0, model& mdl, app* t, expr_ref& ge, expr_ref& gt) { SASSERT(a.is_real(t)); expr_ref_vector fmls(fmls0); opt::model_based_opt mbo; opt::inf_eps value; obj_map ts; obj_map tids; model_evaluator eval(mdl); // extract objective function. vars coeffs; rational c(0), mul(1); linearize(mbo, eval, mul, t, c, fmls, ts, tids); extract_coefficients(mbo, eval, ts, tids, coeffs); mbo.set_objective(coeffs, c); SASSERT(validate_model(eval, fmls0)); // extract linear constraints for (expr * fml : fmls) { linearize(mbo, eval, fml, fmls, tids); } // find optimal value value = mbo.maximize(); // update model to use new values that satisfy optimality ptr_vector vars; for (auto& kv : tids) { expr* e = kv.m_key; if (is_uninterp_const(e)) { unsigned id = kv.m_value; func_decl* f = to_app(e)->get_decl(); expr_ref val(a.mk_numeral(mbo.get_value(id), false), m); mdl.register_decl(f, val); } else { TRACE("qe", tout << "omitting model update for non-uninterpreted constant " << mk_pp(e, m) << "\n";); } } expr_ref val(a.mk_numeral(value.get_rational(), false), m); expr_ref tval = eval(t); // update the predicate 'bound' which forces larger values when 'strict' is true. // strict: bound := valuue < t // !strict: bound := value <= t if (!value.is_finite()) { ge = a.mk_ge(t, tval); gt = m.mk_false(); } else if (value.get_infinitesimal().is_neg()) { ge = a.mk_ge(t, tval); gt = a.mk_ge(t, val); } else { ge = a.mk_ge(t, val); gt = a.mk_gt(t, val); } SASSERT(validate_model(eval, fmls0)); return value; } bool validate_model(model_evaluator& eval, expr_ref_vector const& fmls) { bool valid = true; for (expr* fml : fmls) { expr_ref val = eval(fml); if (!m.is_true(val)) { valid = false; TRACE("qe", tout << mk_pp(fml, m) << " := " << val << "\n";); } } return valid; } void extract_coefficients(opt::model_based_opt& mbo, model_evaluator& eval, obj_map const& ts, obj_map& tids, vars& coeffs) { coeffs.reset(); eval.set_model_completion(true); for (auto& kv : ts) { unsigned id; expr* v = kv.m_key; if (!tids.find(v, id)) { rational r; expr_ref val = eval(v); if (!a.is_numeral(val, r)) { TRACE("qe", tout << eval.get_model() << "\n";); throw default_exception("mbp evaluation was only partial"); } id = mbo.add_var(r, a.is_int(v)); tids.insert(v, id); } CTRACE("qe", kv.m_value.is_zero(), tout << mk_pp(v, m) << " has coefficeint 0\n";); if (!kv.m_value.is_zero()) { coeffs.push_back(var(id, kv.m_value)); } } } void apply_projection(vector& defs, expr_ref_vector& fmls) { if (fmls.empty() || defs.empty()) return; expr_safe_replace subst(m); for (auto const& d : defs) subst.insert(d.var, d.term); unsigned j = 0; expr_ref tmp(m); for (expr* fml : fmls) { subst(fml, tmp); fmls[j++] = tmp; } } }; arith_project_plugin::arith_project_plugin(ast_manager& m):project_plugin(m) { m_imp = alloc(imp, m); } arith_project_plugin::~arith_project_plugin() { dealloc(m_imp); } bool arith_project_plugin::operator()(model& model, app* var, app_ref_vector& vars, expr_ref_vector& lits) { return (*m_imp)(model, var, vars, lits); } void arith_project_plugin::operator()(model& model, app_ref_vector& vars, expr_ref_vector& lits) { m_imp->project(model, vars, lits, false); } vector arith_project_plugin::project(model& model, app_ref_vector& vars, expr_ref_vector& lits) { return m_imp->project(model, vars, lits, true); } void arith_project_plugin::set_check_purified(bool check_purified) { m_imp->m_check_purified = check_purified; } void arith_project_plugin::set_apply_projection(bool f) { m_imp->m_apply_projection = f; } family_id arith_project_plugin::get_family_id() { return m_imp->a.get_family_id(); } opt::inf_eps arith_project_plugin::maximize(expr_ref_vector const& fmls, model& mdl, app* t, expr_ref& ge, expr_ref& gt) { return m_imp->maximize(fmls, mdl, t, ge, gt); } bool arith_project(model& model, app* var, expr_ref_vector& lits) { ast_manager& m = lits.get_manager(); arith_project_plugin ap(m); app_ref_vector vars(m); return ap(model, var, vars, lits); } }