graphite2/src/Pass.cpp

/*  GRAPHITE2 LICENSING

    Copyright 2010, SIL International
    All rights reserved.

    This library is free software; you can redistribute it and/or modify
    it under the terms of the GNU Lesser General Public License as published
    by the Free Software Foundation; either version 2.1 of License, or
    (at your option) any later version.

    This program is distributed in the hope that it will be useful,
    but WITHOUT ANY WARRANTY; without even the implied warranty of
    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
    Lesser General Public License for more details.

    You should also have received a copy of the GNU Lesser General Public
    License along with this library in the file named "LICENSE".
    If not, write to the Free Software Foundation, 51 Franklin Street,
    Suite 500, Boston, MA 02110-1335, USA or visit their web page on the
    internet at http://www.fsf.org/licenses/lgpl.html.

Alternatively, the contents of this file may be used under the terms of the
Mozilla Public License (http://mozilla.org/MPL) or the GNU General Public
License, as published by the Free Software Foundation, either version 2
of the License or (at your option) any later version.
*/
#include "inc/Main.h"
#include "inc/debug.h"
#include "inc/Endian.h"
#include "inc/Pass.h"
#include <cstring>
#include <cstdlib>
#include <cassert>
#include <cmath>
#include "inc/Segment.h"
#include "inc/Code.h"
#include "inc/Rule.h"
#include "inc/Error.h"
#include "inc/Collider.h"

using namespace graphite2;
using vm::Machine;
typedef Machine::Code  Code;

enum KernCollison
{
    None       = 0,
    CrossSpace = 1,
    InWord     = 2,
    reserved   = 3
};

Pass::Pass()
: m_silf(0),
  m_cols(0),
  m_rules(0),
  m_ruleMap(0),
  m_startStates(0),
  m_transitions(0),
  m_states(0),
  m_codes(0),
  m_progs(0),
  m_numCollRuns(0),
  m_kernColls(0),
  m_iMaxLoop(0),
  m_numGlyphs(0),
  m_numRules(0),
  m_numStates(0),
  m_numTransition(0),
  m_numSuccess(0),
  m_successStart(0),
  m_numColumns(0),
  m_minPreCtxt(0),
  m_maxPreCtxt(0),
  m_colThreshold(0),
  m_isReverseDir(false)
{
}

Pass::~Pass()
{
    free(m_cols);
    free(m_startStates);
    free(m_transitions);
    free(m_states);
    free(m_ruleMap);

    if (m_rules) delete [] m_rules;
    if (m_codes) delete [] m_codes;
    free(m_progs);
}

bool Pass::readPass(const byte * const pass_start, size_t pass_length, size_t subtable_base,
        GR_MAYBE_UNUSED Face & face, passtype pt, GR_MAYBE_UNUSED uint32 version, Error &e)
{
    const byte * p              = pass_start,
               * const pass_end = p + pass_length;
    size_t numRanges;

    if (e.test(pass_length < 40, E_BADPASSLENGTH)) return face.error(e);
    // Read in basic values
    const byte flags = be::read<byte>(p);
    if (e.test((flags & 0x1f) &&
            (pt < PASS_TYPE_POSITIONING || !m_silf->aCollision() || !face.glyphs().hasBoxes() || !(m_silf->flags() & 0x20)),
            E_BADCOLLISIONPASS))
        return face.error(e);
    m_numCollRuns = flags & 0x7;
    m_kernColls   = (flags >> 3) & 0x3;
    m_isReverseDir = (flags >> 5) & 0x1;
    m_iMaxLoop = be::read<byte>(p);
    if (m_iMaxLoop < 1) m_iMaxLoop = 1;
    be::skip<byte>(p,2); // skip maxContext & maxBackup
    m_numRules = be::read<uint16>(p);
    if (e.test(!m_numRules && m_numCollRuns == 0, E_BADEMPTYPASS)) return face.error(e);
    be::skip<uint16>(p);   // fsmOffset - not sure why we would want this
    const byte * const pcCode = pass_start + be::read<uint32>(p) - subtable_base,
               * const rcCode = pass_start + be::read<uint32>(p) - subtable_base,
               * const aCode  = pass_start + be::read<uint32>(p) - subtable_base;
    be::skip<uint32>(p);
    m_numStates = be::read<uint16>(p);
    m_numTransition = be::read<uint16>(p);
    m_numSuccess = be::read<uint16>(p);
    m_numColumns = be::read<uint16>(p);
    numRanges = be::read<uint16>(p);
    be::skip<uint16>(p, 3); // skip searchRange, entrySelector & rangeShift.
    assert(p - pass_start == 40);
    // Perform some sanity checks.
    if ( e.test(m_numTransition > m_numStates, E_BADNUMTRANS)
            || e.test(m_numSuccess > m_numStates, E_BADNUMSUCCESS)
            || e.test(m_numSuccess + m_numTransition < m_numStates, E_BADNUMSTATES)
            || e.test(m_numRules && numRanges == 0, E_NORANGES)
            || e.test(m_numColumns > 0x7FFF, E_BADNUMCOLUMNS))
        return face.error(e);

    m_successStart = m_numStates - m_numSuccess;
    // test for beyond end - 1 to account for reading uint16
    if (e.test(p + numRanges * 6 - 2 > pass_end, E_BADPASSLENGTH)) return face.error(e);
    m_numGlyphs = be::peek<uint16>(p + numRanges * 6 - 4) + 1;
    // Calculate the start of various arrays.
    const byte * const ranges = p;
    be::skip<uint16>(p, numRanges*3);
    const byte * const o_rule_map = p;
    be::skip<uint16>(p, m_numSuccess + 1);

    // More sanity checks
    if (e.test(reinterpret_cast<const byte *>(o_rule_map + m_numSuccess*sizeof(uint16)) > pass_end
            || p > pass_end, E_BADRULEMAPLEN))
        return face.error(e);
    const size_t numEntries = be::peek<uint16>(o_rule_map + m_numSuccess*sizeof(uint16));
    const byte * const   rule_map = p;
    be::skip<uint16>(p, numEntries);

    if (e.test(p + 2*sizeof(uint8) > pass_end, E_BADPASSLENGTH)) return face.error(e);
    m_minPreCtxt = be::read<uint8>(p);
    m_maxPreCtxt = be::read<uint8>(p);
    if (e.test(m_minPreCtxt > m_maxPreCtxt, E_BADCTXTLENBOUNDS)) return face.error(e);
    const byte * const start_states = p;
    be::skip<int16>(p, m_maxPreCtxt - m_minPreCtxt + 1);
    const uint16 * const sort_keys = reinterpret_cast<const uint16 *>(p);
    be::skip<uint16>(p, m_numRules);
    const byte * const precontext = p;
    be::skip<byte>(p, m_numRules);

    if (e.test(p + sizeof(uint16) + sizeof(uint8) > pass_end, E_BADCTXTLENS)) return face.error(e);
    m_colThreshold = be::read<uint8>(p);
    if (m_colThreshold == 0) m_colThreshold = 10;       // A default
    const size_t pass_constraint_len = be::read<uint16>(p);

    const uint16 * const o_constraint = reinterpret_cast<const uint16 *>(p);
    be::skip<uint16>(p, m_numRules + 1);
    const uint16 * const o_actions = reinterpret_cast<const uint16 *>(p);
    be::skip<uint16>(p, m_numRules + 1);
    const byte * const states = p;
    if (e.test(2u*m_numTransition*m_numColumns >= (unsigned)(pass_end - p), E_BADPASSLENGTH)
            || e.test(p >= pass_end, E_BADPASSLENGTH))
        return face.error(e);
    be::skip<int16>(p, m_numTransition*m_numColumns);
    be::skip<uint8>(p);
    if (e.test(p != pcCode, E_BADPASSCCODEPTR)) return face.error(e);
    be::skip<byte>(p, pass_constraint_len);
    if (e.test(p != rcCode, E_BADRULECCODEPTR)
        || e.test(size_t(rcCode - pcCode) != pass_constraint_len, E_BADCCODELEN)) return face.error(e);
    be::skip<byte>(p, be::peek<uint16>(o_constraint + m_numRules));
    if (e.test(p != aCode, E_BADACTIONCODEPTR)) return face.error(e);
    be::skip<byte>(p, be::peek<uint16>(o_actions + m_numRules));

    // We should be at the end or within the pass
    if (e.test(p > pass_end, E_BADPASSLENGTH)) return face.error(e);

    // Load the pass constraint if there is one.
    if (pass_constraint_len)
    {
        face.error_context(face.error_context() + 1);
        m_cPConstraint = vm::Machine::Code(true, pcCode, pcCode + pass_constraint_len,
                                  precontext[0], be::peek<uint16>(sort_keys), *m_silf, face, PASS_TYPE_UNKNOWN);
        if (e.test(!m_cPConstraint, E_OUTOFMEM)
                || e.test(m_cPConstraint.status() != Code::loaded, m_cPConstraint.status() + E_CODEFAILURE))
            return face.error(e);
        face.error_context(face.error_context() - 1);
    }
    if (m_numRules)
    {
        if (!readRanges(ranges, numRanges, e)) return face.error(e);
        if (!readRules(rule_map, numEntries,  precontext, sort_keys,
                   o_constraint, rcCode, o_actions, aCode, face, pt, e)) return false;
    }
#ifdef GRAPHITE2_TELEMETRY
    telemetry::category _states_cat(face.tele.states);
#endif
    return m_numRules ? readStates(start_states, states, o_rule_map, face, e) : true;
}


bool Pass::readRules(const byte * rule_map, const size_t num_entries,
                     const byte *precontext, const uint16 * sort_key,
                     const uint16 * o_constraint, const byte *rc_data,
                     const uint16 * o_action,     const byte * ac_data,
                     Face & face, passtype pt, Error &e)
{
    const byte * const ac_data_end = ac_data + be::peek<uint16>(o_action + m_numRules);
    const byte * const rc_data_end = rc_data + be::peek<uint16>(o_constraint + m_numRules);

    precontext += m_numRules;
    sort_key   += m_numRules;
    o_constraint += m_numRules;
    o_action += m_numRules;

    // Load rules.
    const byte * ac_begin = 0, * rc_begin = 0,
               * ac_end = ac_data + be::peek<uint16>(o_action),
               * rc_end = rc_data + be::peek<uint16>(o_constraint);

    // Allocate pools
    m_rules = new Rule [m_numRules];
    m_codes = new Code [m_numRules*2];
    int totalSlots = 0;
    const uint16 *tsort = sort_key;
    for (int i = 0; i < m_numRules; ++i)
        totalSlots += be::peek<uint16>(--tsort);
    const size_t prog_pool_sz = vm::Machine::Code::estimateCodeDataOut(ac_end - ac_data + rc_end - rc_data, 2 * m_numRules, totalSlots);
    m_progs = gralloc<byte>(prog_pool_sz);
    byte * prog_pool_free = m_progs,
         * prog_pool_end  = m_progs + prog_pool_sz;
    if (e.test(!(m_rules && m_codes && m_progs), E_OUTOFMEM)) return face.error(e);

    Rule * r = m_rules + m_numRules - 1;
    for (size_t n = m_numRules; r >= m_rules; --n, --r, ac_end = ac_begin, rc_end = rc_begin)
    {
        face.error_context((face.error_context() & 0xFFFF00) + EC_ARULE + int((n - 1) << 24));
        r->preContext = *--precontext;
        r->sort       = be::peek<uint16>(--sort_key);
#ifndef NDEBUG
        r->rule_idx   = uint16(n - 1);
#endif
        if (r->sort > 63 || r->preContext >= r->sort || r->preContext > m_maxPreCtxt || r->preContext < m_minPreCtxt)
            return false;
        ac_begin      = ac_data + be::peek<uint16>(--o_action);
        --o_constraint;
        rc_begin      = be::peek<uint16>(o_constraint) ? rc_data + be::peek<uint16>(o_constraint) : rc_end;

        if (ac_begin > ac_end || ac_begin > ac_data_end || ac_end > ac_data_end
                || rc_begin > rc_end || rc_begin > rc_data_end || rc_end > rc_data_end
                || vm::Machine::Code::estimateCodeDataOut(ac_end - ac_begin + rc_end - rc_begin, 2, r->sort) > size_t(prog_pool_end - prog_pool_free))
            return false;
        r->action     = new (m_codes+n*2-2) vm::Machine::Code(false, ac_begin, ac_end, r->preContext, r->sort, *m_silf, face, pt, &prog_pool_free);
        r->constraint = new (m_codes+n*2-1) vm::Machine::Code(true,  rc_begin, rc_end, r->preContext, r->sort, *m_silf, face, pt, &prog_pool_free);

        if (e.test(!r->action || !r->constraint, E_OUTOFMEM)
                || e.test(r->action->status() != Code::loaded, r->action->status() + E_CODEFAILURE)
                || e.test(r->constraint->status() != Code::loaded, r->constraint->status() + E_CODEFAILURE)
                || e.test(!r->constraint->immutable(), E_MUTABLECCODE))
            return face.error(e);
    }

    byte * const moved_progs = prog_pool_free > m_progs ? static_cast<byte *>(realloc(m_progs, prog_pool_free - m_progs)) : 0;
    if (e.test(!moved_progs, E_OUTOFMEM))
    {
        free(m_progs);
        m_progs = 0;
        return face.error(e);
    }

    if (moved_progs != m_progs)
    {
        for (Code * c = m_codes, * const ce = c + m_numRules*2; c != ce; ++c)
        {
            c->externalProgramMoved(moved_progs - m_progs);
        }
        m_progs = moved_progs;
    }

    // Load the rule entries map
    face.error_context((face.error_context() & 0xFFFF00) + EC_APASS);
    //TODO: Coverity: 1315804: FORWARD_NULL
    RuleEntry * re = m_ruleMap = gralloc<RuleEntry>(num_entries);
    if (e.test(!re, E_OUTOFMEM)) return face.error(e);
    for (size_t n = num_entries; n; --n, ++re)
    {
        const ptrdiff_t rn = be::read<uint16>(rule_map);
        if (e.test(rn >= m_numRules, E_BADRULENUM))  return face.error(e);
        re->rule = m_rules + rn;
    }

    return true;
}

static int cmpRuleEntry(const void *a, const void *b) { return (*(RuleEntry *)a < *(RuleEntry *)b ? -1 :
                                                                (*(RuleEntry *)b < *(RuleEntry *)a ? 1 : 0)); }

bool Pass::readStates(const byte * starts, const byte *states, const byte * o_rule_map, GR_MAYBE_UNUSED Face & face, Error &e)
{
#ifdef GRAPHITE2_TELEMETRY
    telemetry::category _states_cat(face.tele.starts);
#endif
    m_startStates = gralloc<uint16>(m_maxPreCtxt - m_minPreCtxt + 1);
#ifdef GRAPHITE2_TELEMETRY
    telemetry::set_category(face.tele.states);
#endif
    m_states      = gralloc<State>(m_numStates);
#ifdef GRAPHITE2_TELEMETRY
    telemetry::set_category(face.tele.transitions);
#endif
    m_transitions      = gralloc<uint16>(m_numTransition * m_numColumns);

    if (e.test(!m_startStates || !m_states || !m_transitions, E_OUTOFMEM)) return face.error(e);
    // load start states
    for (uint16 * s = m_startStates,
                * const s_end = s + m_maxPreCtxt - m_minPreCtxt + 1; s != s_end; ++s)
    {
        *s = be::read<uint16>(starts);
        if (e.test(*s >= m_numStates, E_BADSTATE))
        {
            face.error_context((face.error_context() & 0xFFFF00) + EC_ASTARTS + int((s - m_startStates) << 24));
            return face.error(e); // true;
        }
    }

    // load state transition table.
    for (uint16 * t = m_transitions,
                * const t_end = t + m_numTransition*m_numColumns; t != t_end; ++t)
    {
        *t = be::read<uint16>(states);
        if (e.test(*t >= m_numStates, E_BADSTATE))
        {
            face.error_context((face.error_context() & 0xFFFF00) + EC_ATRANS + int(((t - m_transitions) / m_numColumns) << 8));
            return face.error(e);
        }
    }

    State * s = m_states,
          * const success_begin = m_states + m_numStates - m_numSuccess;
    const RuleEntry * rule_map_end = m_ruleMap + be::peek<uint16>(o_rule_map + m_numSuccess*sizeof(uint16));
    for (size_t n = m_numStates; n; --n, ++s)
    {
        RuleEntry * const begin = s < success_begin ? 0 : m_ruleMap + be::read<uint16>(o_rule_map),
                  * const end   = s < success_begin ? 0 : m_ruleMap + be::peek<uint16>(o_rule_map);

        if (e.test(begin >= rule_map_end || end > rule_map_end || begin > end, E_BADRULEMAPPING))
        {
            face.error_context((face.error_context() & 0xFFFF00) + EC_ARULEMAP + int(n << 24));
            return face.error(e);
        }
        s->rules = begin;
        s->rules_end = (end - begin <= FiniteStateMachine::MAX_RULES)? end :
            begin + FiniteStateMachine::MAX_RULES;
        if (begin)      // keep UBSan happy can't call qsort with null begin
            qsort(begin, end - begin, sizeof(RuleEntry), &cmpRuleEntry);
    }

    return true;
}

bool Pass::readRanges(const byte * ranges, size_t num_ranges, Error &e)
{
    m_cols = gralloc<uint16>(m_numGlyphs);
    if (e.test(!m_cols, E_OUTOFMEM)) return false;
    memset(m_cols, 0xFF, m_numGlyphs * sizeof(uint16));
    for (size_t n = num_ranges; n; --n)
    {
        uint16     * ci     = m_cols + be::read<uint16>(ranges),
                   * ci_end = m_cols + be::read<uint16>(ranges) + 1,
                     col    = be::read<uint16>(ranges);

        if (e.test(ci >= ci_end || ci_end > m_cols+m_numGlyphs || col >= m_numColumns, E_BADRANGE))
            return false;

        // A glyph must only belong to one column at a time
        while (ci != ci_end && *ci == 0xffff)
            *ci++ = col;

        if (e.test(ci != ci_end, E_BADRANGE))
            return false;
    }
    return true;
}


bool Pass::runGraphite(vm::Machine & m, FiniteStateMachine & fsm, bool reverse) const
{
    Slot *s = m.slotMap().segment.first();
    if (!s || !testPassConstraint(m)) return true;
    if (reverse)
    {
        m.slotMap().segment.reverseSlots();
        s = m.slotMap().segment.first();
    }
    if (m_numRules)
    {
        Slot *currHigh = s->next();

#if !defined GRAPHITE2_NTRACING
        if (fsm.dbgout)  *fsm.dbgout << "rules" << json::array;
        json::closer rules_array_closer(fsm.dbgout);
#endif

        m.slotMap().highwater(currHigh);
        int lc = m_iMaxLoop;
        do
        {
            findNDoRule(s, m, fsm);
            if (m.status() != Machine::finished) return false;
            if (s && (s == m.slotMap().highwater() || m.slotMap().highpassed() || --lc == 0)) {
                if (!lc)
                    s = m.slotMap().highwater();
                lc = m_iMaxLoop;
                if (s)
                    m.slotMap().highwater(s->next());
            }
        } while (s);
    }
    //TODO: Use enums for flags
    const bool collisions = m_numCollRuns || m_kernColls;

    if (!collisions || !m.slotMap().segment.hasCollisionInfo())
        return true;

    if (m_numCollRuns)
    {
        if (!(m.slotMap().segment.flags() & Segment::SEG_INITCOLLISIONS))
        {
            m.slotMap().segment.positionSlots(0, 0, 0, m.slotMap().dir(), true);
//            m.slotMap().segment.flags(m.slotMap().segment.flags() | Segment::SEG_INITCOLLISIONS);
        }
        if (!collisionShift(&m.slotMap().segment, m.slotMap().dir(), fsm.dbgout))
            return false;
    }
    if ((m_kernColls) && !collisionKern(&m.slotMap().segment, m.slotMap().dir(), fsm.dbgout))
        return false;
    if (collisions && !collisionFinish(&m.slotMap().segment, fsm.dbgout))
        return false;
    return true;
}

bool Pass::runFSM(FiniteStateMachine& fsm, Slot * slot) const
{
    fsm.reset(slot, m_maxPreCtxt);
    if (fsm.slots.context() < m_minPreCtxt)
        return false;

    uint16 state = m_startStates[m_maxPreCtxt - fsm.slots.context()];
    uint8  free_slots = SlotMap::MAX_SLOTS;
    do
    {
        fsm.slots.pushSlot(slot);
        if (slot->gid() >= m_numGlyphs
         || m_cols[slot->gid()] == 0xffffU
         || --free_slots == 0
         || state >= m_numTransition)
            return free_slots != 0;

        const uint16 * transitions = m_transitions + state*m_numColumns;
        state = transitions[m_cols[slot->gid()]];
        if (state >= m_successStart)
            fsm.rules.accumulate_rules(m_states[state]);

        slot = slot->next();
    } while (state != 0 && slot);

    fsm.slots.pushSlot(slot);
    return true;
}

#if !defined GRAPHITE2_NTRACING

inline
Slot * input_slot(const SlotMap &  slots, const int n)
{
    Slot * s = slots[slots.context() + n];
    if (!s->isCopied())     return s;

    return s->prev() ? s->prev()->next() : (s->next() ? s->next()->prev() : slots.segment.last());
}

inline
Slot * output_slot(const SlotMap &  slots, const int n)
{
    Slot * s = slots[slots.context() + n - 1];
    return s ? s->next() : slots.segment.first();
}

#endif //!defined GRAPHITE2_NTRACING

void Pass::findNDoRule(Slot * & slot, Machine &m, FiniteStateMachine & fsm) const
{
    assert(slot);

    if (runFSM(fsm, slot))
    {
        // Search for the first rule which passes the constraint
        const RuleEntry *        r = fsm.rules.begin(),
                        * const re = fsm.rules.end();
        while (r != re && !testConstraint(*r->rule, m))
        {
            ++r;
            if (m.status() != Machine::finished)
                return;
        }

#if !defined GRAPHITE2_NTRACING
        if (fsm.dbgout)
        {
            if (fsm.rules.size() != 0)
            {
                *fsm.dbgout << json::item << json::object;
                dumpRuleEventConsidered(fsm, *r);
                if (r != re)
                {
                    const int adv = doAction(r->rule->action, slot, m);
                    dumpRuleEventOutput(fsm, *r->rule, slot);
                    if (r->rule->action->deletes()) fsm.slots.collectGarbage(slot);
                    adjustSlot(adv, slot, fsm.slots);
                    *fsm.dbgout << "cursor" << objectid(dslot(&fsm.slots.segment, slot))
                            << json::close; // Close RuelEvent object

                    return;
                }
                else
                {
                    *fsm.dbgout << json::close  // close "considered" array
                            << "output" << json::null
                            << "cursor" << objectid(dslot(&fsm.slots.segment, slot->next()))
                            << json::close;
                }
            }
        }
        else
#endif
        {
            if (r != re)
            {
                const int adv = doAction(r->rule->action, slot, m);
                if (m.status() != Machine::finished) return;
                if (r->rule->action->deletes()) fsm.slots.collectGarbage(slot);
                adjustSlot(adv, slot, fsm.slots);
                return;
            }
        }
    }

    slot = slot->next();
    return;
}

#if !defined GRAPHITE2_NTRACING

void Pass::dumpRuleEventConsidered(const FiniteStateMachine & fsm, const RuleEntry & re) const
{
    *fsm.dbgout << "considered" << json::array;
    for (const RuleEntry *r = fsm.rules.begin(); r != &re; ++r)
    {
        if (r->rule->preContext > fsm.slots.context())
            continue;
        *fsm.dbgout << json::flat << json::object
                    << "id" << r->rule - m_rules
                    << "failed" << true
                    << "input" << json::flat << json::object
                        << "start" << objectid(dslot(&fsm.slots.segment, input_slot(fsm.slots, -r->rule->preContext)))
                        << "length" << r->rule->sort
                        << json::close  // close "input"
                    << json::close; // close Rule object
    }
}


void Pass::dumpRuleEventOutput(const FiniteStateMachine & fsm, const Rule & r, Slot * const last_slot) const
{
    *fsm.dbgout     << json::item << json::flat << json::object
                        << "id"     << &r - m_rules
                        << "failed" << false
                        << "input" << json::flat << json::object
                            << "start" << objectid(dslot(&fsm.slots.segment, input_slot(fsm.slots, 0)))
                            << "length" << r.sort - r.preContext
                            << json::close // close "input"
                        << json::close  // close Rule object
                << json::close // close considered array
                << "output" << json::object
                    << "range" << json::flat << json::object
                        << "start"  << objectid(dslot(&fsm.slots.segment, input_slot(fsm.slots, 0)))
                        << "end"    << objectid(dslot(&fsm.slots.segment, last_slot))
                    << json::close // close "input"
                    << "slots"  << json::array;
    const Position rsb_prepos = last_slot ? last_slot->origin() : fsm.slots.segment.advance();
    fsm.slots.segment.positionSlots(0, 0, 0, fsm.slots.segment.currdir());

    for(Slot * slot = output_slot(fsm.slots, 0); slot != last_slot; slot = slot->next())
        *fsm.dbgout     << dslot(&fsm.slots.segment, slot);
    *fsm.dbgout         << json::close  // close "slots"
                    << "postshift"  << (last_slot ? last_slot->origin() : fsm.slots.segment.advance()) - rsb_prepos
                << json::close;         // close "output" object

}

#endif


inline
bool Pass::testPassConstraint(Machine & m) const
{
    if (!m_cPConstraint) return true;

    assert(m_cPConstraint.constraint());

    m.slotMap().reset(*m.slotMap().segment.first(), 0);
    m.slotMap().pushSlot(m.slotMap().segment.first());
    vm::slotref * map = m.slotMap().begin();
    const uint32 ret = m_cPConstraint.run(m, map);

#if !defined GRAPHITE2_NTRACING
    json * const dbgout = m.slotMap().segment.getFace()->logger();
    if (dbgout)
        *dbgout << "constraint" << (ret && m.status() == Machine::finished);
#endif

    return ret && m.status() == Machine::finished;
}


bool Pass::testConstraint(const Rule & r, Machine & m) const
{
    const uint16 curr_context = m.slotMap().context();
    if (unsigned(r.sort + curr_context - r.preContext) > m.slotMap().size()
        || curr_context - r.preContext < 0) return false;

    vm::slotref * map = m.slotMap().begin() + curr_context - r.preContext;
    if (map[r.sort - 1] == 0)
        return false;

    if (!*r.constraint) return true;
    assert(r.constraint->constraint());
    for (int n = r.sort; n && map; --n, ++map)
    {
        if (!*map) continue;
        const int32 ret = r.constraint->run(m, map);
        if (!ret || m.status() != Machine::finished)
            return false;
    }

    return true;
}


void SlotMap::collectGarbage(Slot * &aSlot)
{
    for(Slot **s = begin(), *const *const se = end() - 1; s != se; ++s) {
        Slot *& slot = *s;
        if(slot && (slot->isDeleted() || slot->isCopied()))
        {
            if (slot == aSlot)
                aSlot = slot->prev() ? slot->prev() : slot->next();
            segment.freeSlot(slot);
        }
    }
}


int Pass::doAction(const Code *codeptr, Slot * & slot_out, vm::Machine & m) const
{
    assert(codeptr);
    if (!*codeptr) return 0;
    SlotMap   & smap = m.slotMap();
    vm::slotref * map = &smap[smap.context()];
    smap.highpassed(false);

    int32 ret = codeptr->run(m, map);

    if (m.status() != Machine::finished)
    {
        slot_out = NULL;
        smap.highwater(0);
        return 0;
    }

    slot_out = *map;
    return ret;
}


void Pass::adjustSlot(int delta, Slot * & slot_out, SlotMap & smap) const
{
    if (!slot_out)
    {
        if (smap.highpassed() || slot_out == smap.highwater())
        {
            slot_out = smap.segment.last();
            ++delta;
            if (!smap.highwater() || smap.highwater() == slot_out)
                smap.highpassed(false);
        }
        else
        {
            slot_out = smap.segment.first();
            --delta;
        }
    }
    if (delta < 0)
    {
        while (++delta <= 0 && slot_out)
        {
            slot_out = slot_out->prev();
            if (smap.highpassed() && smap.highwater() == slot_out)
                smap.highpassed(false);
        }
    }
    else if (delta > 0)
    {
        while (--delta >= 0 && slot_out)
        {
            if (slot_out == smap.highwater() && slot_out)
                smap.highpassed(true);
            slot_out = slot_out->next();
        }
    }
}

bool Pass::collisionShift(Segment *seg, int dir, json * const dbgout) const
{
    ShiftCollider shiftcoll(dbgout);
    // bool isfirst = true;
    bool hasCollisions = false;
    Slot *start = seg->first();      // turn on collision fixing for the first slot
    Slot *end = NULL;
    bool moved = false;

#if !defined GRAPHITE2_NTRACING
    if (dbgout)
        *dbgout << "collisions" << json::array
            << json::flat << json::object << "num-loops" << m_numCollRuns << json::close;
#endif

    while (start)
    {
#if !defined GRAPHITE2_NTRACING
        if (dbgout)  *dbgout << json::object << "phase" << "1" << "moves" << json::array;
#endif
        hasCollisions = false;
        end = NULL;
        // phase 1 : position shiftable glyphs, ignoring kernable glyphs
        for (Slot *s = start; s; s = s->next())
        {
            const SlotCollision * c = seg->collisionInfo(s);
            if (start && (c->flags() & (SlotCollision::COLL_FIX | SlotCollision::COLL_KERN)) == SlotCollision::COLL_FIX
                      && !resolveCollisions(seg, s, start, shiftcoll, false, dir, moved, hasCollisions, dbgout))
                return false;
            if (s != start && (c->flags() & SlotCollision::COLL_END))
            {
                end = s->next();
                break;
            }
        }

#if !defined GRAPHITE2_NTRACING
        if (dbgout)
            *dbgout << json::close << json::close; // phase-1
#endif

        // phase 2 : loop until happy.
        for (int i = 0; i < m_numCollRuns - 1; ++i)
        {
            if (hasCollisions || moved)
            {

#if !defined GRAPHITE2_NTRACING
                if (dbgout)
                    *dbgout << json::object << "phase" << "2a" << "loop" << i << "moves" << json::array;
#endif
                // phase 2a : if any shiftable glyphs are in collision, iterate backwards,
                // fixing them and ignoring other non-collided glyphs. Note that this handles ONLY
                // glyphs that are actually in collision from phases 1 or 2b, and working backwards
                // has the intended effect of breaking logjams.
                if (hasCollisions)
                {
                    hasCollisions = false;
                    #if 0
                    moved = true;
                    for (Slot *s = start; s != end; s = s->next())
                    {
                        SlotCollision * c = seg->collisionInfo(s);
                        c->setShift(Position(0, 0));
                    }
                    #endif
                    Slot *lend = end ? end->prev() : seg->last();
                    Slot *lstart = start->prev();
                    for (Slot *s = lend; s != lstart; s = s->prev())
                    {
                        SlotCollision * c = seg->collisionInfo(s);
                        if (start && (c->flags() & (SlotCollision::COLL_FIX | SlotCollision::COLL_KERN | SlotCollision::COLL_ISCOL))
                                        == (SlotCollision::COLL_FIX | SlotCollision::COLL_ISCOL)) // ONLY if this glyph is still colliding
                        {
                            if (!resolveCollisions(seg, s, lend, shiftcoll, true, dir, moved, hasCollisions, dbgout))
                                return false;
                            c->setFlags(c->flags() | SlotCollision::COLL_TEMPLOCK);
                        }
                    }
                }

#if !defined GRAPHITE2_NTRACING
                if (dbgout)
                    *dbgout << json::close << json::close // phase 2a
                        << json::object << "phase" << "2b" << "loop" << i << "moves" << json::array;
#endif

                // phase 2b : redo basic diacritic positioning pass for ALL glyphs. Each successive loop adjusts
                // glyphs from their current adjusted position, which has the effect of gradually minimizing the
                // resulting adjustment; ie, the final result will be gradually closer to the original location.
                // Also it allows more flexibility in the final adjustment, since it is moving along the
                // possible 8 vectors from successively different starting locations.
                if (moved)
                {
                    moved = false;
                    for (Slot *s = start; s != end; s = s->next())
                    {
                        SlotCollision * c = seg->collisionInfo(s);
                        if (start && (c->flags() & (SlotCollision::COLL_FIX | SlotCollision::COLL_TEMPLOCK
                                                        | SlotCollision::COLL_KERN)) == SlotCollision::COLL_FIX
                                  && !resolveCollisions(seg, s, start, shiftcoll, false, dir, moved, hasCollisions, dbgout))
                            return false;
                        else if (c->flags() & SlotCollision::COLL_TEMPLOCK)
                            c->setFlags(c->flags() & ~SlotCollision::COLL_TEMPLOCK);
                    }
                }
        //      if (!hasCollisions) // no, don't leave yet because phase 2b will continue to improve things
        //          break;
#if !defined GRAPHITE2_NTRACING
                if (dbgout)
                    *dbgout << json::close << json::close; // phase 2
#endif
            }
        }
        if (!end)
            break;
        start = NULL;
        for (Slot *s = end->prev(); s; s = s->next())
        {
            if (seg->collisionInfo(s)->flags() & SlotCollision::COLL_START)
            {
                start = s;
                break;
            }
        }
    }
    return true;
}

bool Pass::collisionKern(Segment *seg, int dir, json * const dbgout) const
{
    Slot *start = seg->first();
    float ymin = 1e38f;
    float ymax = -1e38f;
    const GlyphCache &gc = seg->getFace()->glyphs();

    // phase 3 : handle kerning of clusters
#if !defined GRAPHITE2_NTRACING
    if (dbgout)
        *dbgout << json::object << "phase" << "3" << "moves" << json::array;
#endif

    for (Slot *s = seg->first(); s; s = s->next())
    {
        if (!gc.check(s->gid()))
            return false;
        const SlotCollision * c = seg->collisionInfo(s);
        const Rect &bbox = seg->theGlyphBBoxTemporary(s->gid());
        float y = s->origin().y + c->shift().y;
        if (!(c->flags() & SlotCollision::COLL_ISSPACE))
        {
            ymax = max(y + bbox.tr.y, ymax);
            ymin = min(y + bbox.bl.y, ymin);
        }
        if (start && (c->flags() & (SlotCollision::COLL_KERN | SlotCollision::COLL_FIX))
                        == (SlotCollision::COLL_KERN | SlotCollision::COLL_FIX))
            resolveKern(seg, s, start, dir, ymin, ymax, dbgout);
        if (c->flags() & SlotCollision::COLL_END)
            start = NULL;
        if (c->flags() & SlotCollision::COLL_START)
            start = s;
    }

#if !defined GRAPHITE2_NTRACING
    if (dbgout)
        *dbgout << json::close << json::close; // phase 3
#endif
    return true;
}

bool Pass::collisionFinish(Segment *seg, GR_MAYBE_UNUSED json * const dbgout) const
{
    for (Slot *s = seg->first(); s; s = s->next())
    {
        SlotCollision *c = seg->collisionInfo(s);
        if (c->shift().x != 0 || c->shift().y != 0)
        {
            const Position newOffset = c->shift();
            const Position nullPosition(0, 0);
            c->setOffset(newOffset + c->offset());
            c->setShift(nullPosition);
        }
    }
//    seg->positionSlots();

#if !defined GRAPHITE2_NTRACING
        if (dbgout)
            *dbgout << json::close;
#endif
    return true;
}

// Can slot s be kerned, or is it attached to something that can be kerned?
static bool inKernCluster(Segment *seg, Slot *s)
{
    SlotCollision *c = seg->collisionInfo(s);
    if (c->flags() & SlotCollision::COLL_KERN /** && c->flags() & SlotCollision::COLL_FIX **/ )
        return true;
    while (s->attachedTo())
    {
        s = s->attachedTo();
        c = seg->collisionInfo(s);
        if (c->flags() & SlotCollision::COLL_KERN /** && c->flags() & SlotCollision::COLL_FIX **/ )
            return true;
    }
    return false;
}

// Fix collisions for the given slot.
// Return true if everything was fixed, false if there are still collisions remaining.
// isRev means be we are processing backwards.
bool Pass::resolveCollisions(Segment *seg, Slot *slotFix, Slot *start,
        ShiftCollider &coll, GR_MAYBE_UNUSED bool isRev, int dir, bool &moved, bool &hasCol,
        json * const dbgout) const
{
    Slot * nbor;  // neighboring slot
    SlotCollision *cFix = seg->collisionInfo(slotFix);
    if (!coll.initSlot(seg, slotFix, cFix->limit(), cFix->margin(), cFix->marginWt(),
            cFix->shift(), cFix->offset(), dir, dbgout))
        return false;
    bool collides = false;
    // When we're processing forward, ignore kernable glyphs that preceed the target glyph.
    // When processing backward, don't ignore these until we pass slotFix.
    bool ignoreForKern = !isRev;
    bool rtl = dir & 1;
    Slot *base = slotFix;
    while (base->attachedTo())
        base = base->attachedTo();
    Position zero(0., 0.);

    // Look for collisions with the neighboring glyphs.
    for (nbor = start; nbor; nbor = isRev ? nbor->prev() : nbor->next())
    {
        SlotCollision *cNbor = seg->collisionInfo(nbor);
        bool sameCluster = nbor->isChildOf(base);
        if (nbor != slotFix         						// don't process if this is the slot of interest
                      && !(cNbor->ignore())    				// don't process if ignoring
                      && (nbor == base || sameCluster       // process if in the same cluster as slotFix
                            || !inKernCluster(seg, nbor))   // or this cluster is not to be kerned
//                            || (rtl ^ ignoreForKern))       // or it comes before(ltr) or after(rtl)
                      && (!isRev    // if processing forwards then good to merge otherwise only:
                            || !(cNbor->flags() & SlotCollision::COLL_FIX)     // merge in immovable stuff
                            || ((cNbor->flags() & SlotCollision::COLL_KERN) && !sameCluster)     // ignore other kernable clusters
                            || (cNbor->flags() & SlotCollision::COLL_ISCOL))   // test against other collided glyphs
                      && !coll.mergeSlot(seg, nbor, cNbor, cNbor->shift(), !ignoreForKern, sameCluster, collides, false, dbgout))
            return false;
        else if (nbor == slotFix)
            // Switching sides of this glyph - if we were ignoring kernable stuff before, don't anymore.
            ignoreForKern = !ignoreForKern;

        if (nbor != start && (cNbor->flags() & (isRev ? SlotCollision::COLL_START : SlotCollision::COLL_END)))
            break;
    }
    bool isCol = false;
    if (collides || cFix->shift().x != 0.f || cFix->shift().y != 0.f)
    {
        Position shift = coll.resolve(seg, isCol, dbgout);
        // isCol has been set to true if a collision remains.
        if (std::fabs(shift.x) < 1e38f && std::fabs(shift.y) < 1e38f)
        {
            if (sqr(shift.x-cFix->shift().x) + sqr(shift.y-cFix->shift().y) >= m_colThreshold * m_colThreshold)
                moved = true;
            cFix->setShift(shift);
            if (slotFix->firstChild())
            {
                Rect bbox;
                Position here = slotFix->origin() + shift;
                float clusterMin = here.x;
                slotFix->firstChild()->finalise(seg, NULL, here, bbox, 0, clusterMin, rtl, false);
            }
        }
    }
    else
    {
        // This glyph is not colliding with anything.
#if !defined GRAPHITE2_NTRACING
        if (dbgout)
        {
            *dbgout << json::object
                            << "missed" << objectid(dslot(seg, slotFix));
            coll.outputJsonDbg(dbgout, seg, -1);
            *dbgout << json::close;
        }
#endif
    }

    // Set the is-collision flag bit.
    if (isCol)
    { cFix->setFlags(cFix->flags() | SlotCollision::COLL_ISCOL | SlotCollision::COLL_KNOWN); }
    else
    { cFix->setFlags((cFix->flags() & ~SlotCollision::COLL_ISCOL) | SlotCollision::COLL_KNOWN); }
    hasCol |= isCol;
    return true;
}

float Pass::resolveKern(Segment *seg, Slot *slotFix, GR_MAYBE_UNUSED Slot *start, int dir,
    float &ymin, float &ymax, json *const dbgout) const
{
    Slot *nbor; // neighboring slot
    float currSpace = 0.;
    bool collides = false;
    unsigned int space_count = 0;
    Slot *base = slotFix;
    while (base->attachedTo())
        base = base->attachedTo();
    SlotCollision *cFix = seg->collisionInfo(base);
    const GlyphCache &gc = seg->getFace()->glyphs();
    const Rect &bbb = seg->theGlyphBBoxTemporary(slotFix->gid());
    const float by = slotFix->origin().y + cFix->shift().y;

    if (base != slotFix)
    {
        cFix->setFlags(cFix->flags() | SlotCollision::COLL_KERN | SlotCollision::COLL_FIX);
        return 0;
    }
    bool seenEnd = (cFix->flags() & SlotCollision::COLL_END) != 0;
    bool isInit = false;
    KernCollider coll(dbgout);

    ymax = max(by + bbb.tr.y, ymax);
    ymin = min(by + bbb.bl.y, ymin);
    for (nbor = slotFix->next(); nbor; nbor = nbor->next())
    {
        if (nbor->isChildOf(base))
            continue;
        if (!gc.check(nbor->gid()))
            return 0.;
        const Rect &bb = seg->theGlyphBBoxTemporary(nbor->gid());
        SlotCollision *cNbor = seg->collisionInfo(nbor);
        if ((bb.bl.y == 0.f && bb.tr.y == 0.f) || (cNbor->flags() & SlotCollision::COLL_ISSPACE))
        {
            if (m_kernColls == InWord)
                break;
            // Add space for a space glyph.
            currSpace += nbor->advance();
            ++space_count;
        }
        else
        {
            space_count = 0;
            if (nbor != slotFix && !cNbor->ignore())
            {
                seenEnd = true;
                if (!isInit)
                {
                    if (!coll.initSlot(seg, slotFix, cFix->limit(), cFix->margin(),
                                    cFix->shift(), cFix->offset(), dir, ymin, ymax, dbgout))
                        return 0.;
                    isInit = true;
                }
                collides |= coll.mergeSlot(seg, nbor, cNbor->shift(), currSpace, dir, dbgout);
            }
        }
        if (cNbor->flags() & SlotCollision::COLL_END)
        {
            if (seenEnd && space_count < 2)
                break;
            else
                seenEnd = true;
        }
    }
    if (collides)
    {
        Position mv = coll.resolve(seg, slotFix, dir, dbgout);
        coll.shift(mv, dir);
        Position delta = slotFix->advancePos() + mv - cFix->shift();
        slotFix->advance(delta);
        cFix->setShift(mv);
        return mv.x;
    }
    return 0.;
}