z3/lib/dl_mk_unbound_compressor.cpp

/*++
Copyright (c) 2006 Microsoft Corporation

Module Name:

    dl_mk_unbound_compressor.cpp

Abstract:

    <abstract>

Author:

    Krystof Hoder (t-khoder) 2010-10-04.

Revision History:

--*/

#include<utility>
#include<sstream>
#include"dl_mk_unbound_compressor.h"

namespace datalog {

    mk_unbound_compressor::mk_unbound_compressor(context & ctx) :
        plugin(500),
        m_context(ctx),
        m_manager(ctx.get_manager()),
        m_rules(ctx.get_rule_manager()),
        m_pinned(m_manager) {
    }

    void mk_unbound_compressor::reset() {
        m_rules.reset();
        m_todo.reset();
        m_in_progress.reset();
        m_map.reset();
        m_pinned.reset();
    }

    bool mk_unbound_compressor::is_unbound_argument(rule * r, unsigned head_index) {
        app * head = r->get_head();
        expr * head_arg = head->get_arg(head_index);
        if (!is_var(head_arg)) {
            return false;
        }
        unsigned var_idx = to_var(head_arg)->get_idx();

        var_idx_set tail_vars;
        collect_tail_vars(m_manager, r, tail_vars);

        return tail_vars.contains(var_idx);
    }

    void mk_unbound_compressor::add_task(func_decl * pred, unsigned arg_index) {
        c_info ci = c_info(pred, arg_index);
        if (m_map.contains(ci)) {
            return; //this task was already added
        }

        unsigned parent_arity = pred->get_arity();
        sort * const * parent_domain = pred->get_domain();
        symbol const& parent_name = pred->get_name();
        unsigned arity = parent_arity-1;
        ptr_vector<sort> domain;
        for (unsigned i = 0; i < parent_arity; i++) {
            if (i != arg_index) {
                domain.push_back(parent_domain[i]);
            }
        }

        std::stringstream name_suffix;
        name_suffix << "compr_arg_" << arg_index;

        func_decl * cpred = m_context.mk_fresh_head_predicate(parent_name, symbol(name_suffix.str().c_str()), 
            arity, domain.c_ptr(), pred);
        m_pinned.push_back(cpred);

        m_todo.push_back(ci);
        m_map.insert(ci, cpred);
    }

    void mk_unbound_compressor::detect_tasks(unsigned rule_index) {
        rule * r = m_rules.get(rule_index);
        var_idx_set tail_vars;
        collect_tail_vars(m_manager, r, tail_vars);

        app * head = r->get_head();
        func_decl * head_pred = head->get_decl();
        if (m_context.is_output_predicate(head_pred)) {
            //we don't compress output predicates
            return;
        }

        unsigned n = head_pred->get_arity();

        var_counter head_var_counter;
        head_var_counter.count_vars(m_manager, head, 1);

        for (unsigned i=0; i<n; i++) {
            expr * arg = head->get_arg(i);
            if (!is_var(arg)) {
                continue;
            }
            unsigned var_idx = to_var(arg)->get_idx();
            if (!tail_vars.contains(var_idx)) {
                //unbound

                unsigned occurence_cnt = head_var_counter.get(var_idx);
                SASSERT(occurence_cnt>0);
                if (occurence_cnt == 1) {
                    TRACE("dl", r->display(m_context, tout << "Compress: "););
                    add_task(head_pred, i);
                    return; //we compress out the unbound arguments one by one
                }
            }
        }
    }

    void mk_unbound_compressor::try_compress(unsigned rule_index) {
    start:
        rule * r = m_rules.get(rule_index);
        var_idx_set tail_vars;
        collect_tail_vars(m_manager, r, tail_vars);
        
        app * head = r->get_head();
        func_decl * head_pred = head->get_decl();
        unsigned head_arity = head_pred->get_arity();

        var_counter head_var_counter;
        head_var_counter.count_vars(m_manager, head);

        unsigned arg_index;
        for (arg_index = 0; arg_index < head_arity; arg_index++) {
            expr * arg = head->get_arg(arg_index);
            if (!is_var(arg)) {
                continue;
            }
            unsigned var_idx = to_var(arg)->get_idx();
            if (!tail_vars.contains(var_idx)) {
                //unbound
                unsigned occurence_cnt = head_var_counter.get(var_idx);
                SASSERT(occurence_cnt>0);
                if ( occurence_cnt==1 && m_in_progress.contains(c_info(head_pred, arg_index)) ) {
                    //we have found what to compress
                    break;
                }
            }
        }
        if (arg_index == head_arity) {
            //we didn't find anything to compress
            return;
        }
        SASSERT(arg_index<head_arity);
        c_info ci(head_pred, arg_index);
        func_decl * cpred;
        TRUSTME( m_map.find(ci, cpred) );
        ptr_vector<expr> cargs;
        for (unsigned i=0; i<head_arity; i++) {
            if (i != arg_index) {
                cargs.push_back(head->get_arg(i));
            }
        }

        app_ref chead(m_manager.mk_app(cpred, head_arity-1, cargs.c_ptr()), m_manager);

        if (r->get_tail_size()==0 && m_context.get_rule_manager().is_fact(chead)) {
            m_non_empty_rels.insert(cpred);
            m_context.add_fact(chead);
            //remove the rule that became fact by placing the last rule on its place
            m_head_occurrence_ctr.dec(m_rules.get(rule_index)->get_head()->get_decl());
            m_rules.set(rule_index, m_rules.get(m_rules.size()-1));
            m_rules.shrink(m_rules.size()-1);
            //since we moved the last rule to rule_index, we have to try to compress it as well
            if (rule_index<m_rules.size()) {
                goto start;
            }
        }
        else {
            rule_ref new_rule(m_context.get_rule_manager().mk(r, chead), m_context.get_rule_manager());
            new_rule->set_accounting_parent_object(m_context, r);

            m_head_occurrence_ctr.dec(m_rules.get(rule_index)->get_head()->get_decl());
            m_rules.set(rule_index, new_rule);
            m_head_occurrence_ctr.inc(m_rules.get(rule_index)->get_head()->get_decl());
            detect_tasks(rule_index);
        }

        m_modified = true;
    }

    void mk_unbound_compressor::mk_decompression_rule(rule * r, unsigned tail_index, unsigned arg_index,
        rule_ref& res)
    {
        app * orig_dtail = r->get_tail(tail_index); //dtail ~ decompressed tail
        c_info ci(orig_dtail->get_decl(), arg_index);
        func_decl * dtail_pred;
        TRUSTME( m_map.find(ci, dtail_pred) );
        ptr_vector<expr> dtail_args;
        unsigned orig_dtail_arity = orig_dtail->get_num_args();
        for (unsigned i=0;i<orig_dtail_arity;i++) {
            if (i != arg_index) {
                dtail_args.push_back(orig_dtail->get_arg(i));
            }
        }
        SASSERT(dtail_args.size()==dtail_pred->get_arity());
        app_ref dtail(m_manager.mk_app(dtail_pred, dtail_args.size(), dtail_args.c_ptr()), m_manager);

        svector<bool> tails_negated;
        app_ref_vector tails(m_manager);
        unsigned tail_len = r->get_tail_size();
        for (unsigned i=0; i<tail_len; i++) {
            tails_negated.push_back(r->is_neg_tail(i));
            if (i==tail_index && !r->is_neg_tail(i)) {
                tails.push_back(dtail);
            }
            else {
                tails.push_back(r->get_tail(i));
            }
        }

        // Accumulate negated filtered rule instead 
        // of replacing the original predicate.
        if (r->is_neg_tail(tail_index)) {
            tails_negated.push_back(true);
            tails.push_back(dtail);
        }

        res = m_context.get_rule_manager().mk( r->get_head(), tails.size(), tails.c_ptr(), tails_negated.c_ptr());
        res->set_accounting_parent_object(m_context, r);
        m_context.get_rule_manager().fix_unbound_vars(res, true);
    }

    void mk_unbound_compressor::add_decompression_rule(rule * r, unsigned tail_index, unsigned arg_index) {
        rule_ref new_rule(m_context.get_rule_manager());
        mk_decompression_rule(r, tail_index, arg_index, new_rule);

        unsigned new_rule_index = m_rules.size();
        m_rules.push_back(new_rule);
        m_head_occurrence_ctr.inc(new_rule->get_head()->get_decl());


        detect_tasks(new_rule_index);

        m_modified = true;

        //TODO: avoid rule duplicity
        //If two predicates are compressed in a rule, applying decompression 
        //to the results can cause a rule being added multiple times:
        //P:- R(x,y), S(x,y)
        //is decompressed into rules
        //P:- R1(x), S(x,y)
        //P:- R(x,y), S1(x)
        //and each of these rules is again decompressed giving the same rule
        //P:- R1(x), S1(x)
        //P:- R1(x), S1(x)
    }

    void mk_unbound_compressor::replace_by_decompression_rule(unsigned rule_index, unsigned tail_index, unsigned arg_index)
    {
        rule * r = m_rules.get(rule_index);

        rule_ref new_rule(m_context.get_rule_manager());
        mk_decompression_rule(r, tail_index, arg_index, new_rule);

        m_rules.set(rule_index, new_rule);
        
        //we don't update the m_head_occurrence_ctr because the head predicate doesn't change

        detect_tasks(rule_index);

        m_modified = true;
    }

    void mk_unbound_compressor::add_decompression_rules(unsigned rule_index) {

        unsigned_vector compressed_tail_pred_arg_indexes;

        //this value is updated inside the loop if replace_by_decompression_rule is called
        rule_ref r(m_rules.get(rule_index), m_context.get_rule_manager());

        unsigned utail_len = r->get_uninterpreted_tail_size();
        unsigned tail_index=0;
        while (tail_index<utail_len) {
            app * t = r->get_tail(tail_index);
            func_decl * t_pred = t->get_decl();
            unsigned t_arity = t_pred->get_arity();
            bool is_negated_predicate = r->is_neg_tail(tail_index);
            compressed_tail_pred_arg_indexes.reset();
            for (unsigned arg_index=0; arg_index<t_arity; arg_index++) {
                c_info ci(t_pred, arg_index);
                if (m_in_progress.contains(ci)) {
                    compressed_tail_pred_arg_indexes.push_back(arg_index);
                }
            }
            bool orig_rule_replaced = false;
            while (!compressed_tail_pred_arg_indexes.empty()) {
                unsigned arg_index = compressed_tail_pred_arg_indexes.back();
                compressed_tail_pred_arg_indexes.pop_back();

                bool can_remove_orig_rule = 
                    compressed_tail_pred_arg_indexes.empty() &&
                    !m_non_empty_rels.contains(t_pred) &&
                    m_head_occurrence_ctr.get(t_pred)==0;                

                if (can_remove_orig_rule || is_negated_predicate) {
                    replace_by_decompression_rule(rule_index, tail_index, arg_index);
                    orig_rule_replaced = true;
                }
                else {
                    add_decompression_rule(r, tail_index, arg_index);
                }
            }
            if (orig_rule_replaced) {
                //update r with the new rule
                rule * new_rule = m_rules.get(rule_index);
                SASSERT(new_rule->get_uninterpreted_tail_size() >= utail_len);
                //here we check that the rule replacement didn't affect other uninterpreted literals 
                //in the tail (aside of variable renaming)
                SASSERT(tail_index==0 || 
                    new_rule->get_tail(tail_index-1)->get_decl()==r->get_tail(tail_index-1)->get_decl());

                r = new_rule;

                //we have replaced the original rule, with one that has different
                //content of the tail_index -th tail. we will therefore not do
                //tail_index++, so that we examine the new tail literal as well
            }
            else {
                tail_index++;
            }
        }
    }

    rule_set * mk_unbound_compressor::operator()(rule_set const & source, model_converter_ref& mc, proof_converter_ref& pc) {
        // TODO mc, pc
        m_modified = false;

        m_context.get_rmanager().collect_non_empty_predicates(m_non_empty_rels);

        unsigned init_rule_cnt = source.get_num_rules();
        SASSERT(m_rules.empty());
        for (unsigned i=0; i<init_rule_cnt; i++) {
            rule * r = source.get_rule(i);
            m_rules.push_back(r);
            m_head_occurrence_ctr.inc(r->get_head()->get_decl());
        }

        for (unsigned i=0; i<init_rule_cnt; i++) {
            detect_tasks(i);
        }

        while (!m_todo.empty()) {
            m_in_progress.reset();
            while (!m_todo.empty()) {
                m_in_progress.insert(m_todo.back());
                m_todo.pop_back();
            }
            unsigned rule_index = 0;
            while (rule_index<m_rules.size()) {
                try_compress(rule_index); //m_rules.size() can change here
                if (rule_index<m_rules.size()) {
                    add_decompression_rules(rule_index); //m_rules.size() can change here
                }
                rule_index++;
            }
        }

        rule_set * result = static_cast<rule_set *>(0);
        if (m_modified) {
            result = alloc(rule_set, m_context);
            unsigned fin_rule_cnt = m_rules.size();
            for (unsigned i=0; i<fin_rule_cnt; i++) {
                result->add_rule(m_rules.get(i));
            }
        }
        reset();
        return result;
    }


};