verilator/src/V3Trace.cpp

// -*- mode: C++; c-file-style: "cc-mode" -*-
//*************************************************************************
// DESCRIPTION: Verilator: Waves tracing
//
// Code available from: https://verilator.org
//
//*************************************************************************
//
// Copyright 2003-2020 by Wilson Snyder. This program is free software; you
// can redistribute it and/or modify it under the terms of either the GNU
// Lesser General Public License Version 3 or the Perl Artistic License
// Version 2.0.
// SPDX-License-Identifier: LGPL-3.0-only OR Artistic-2.0
//
//*************************************************************************
// V3Trace's Transformations:
//
//  Examine whole design and build a graph describing which function call
//  may result in a write to a traced variable. This is done in 2 passes:
//
//  Pass 1:
//      Add vertices for TraceDecl, CFunc, CCall and VarRef nodes, add
//      edges from CCall -> CFunc, VarRef -> TraceDecl, also add edges
//      for public entry points to CFuncs (these are like a spontaneous
//      call)
//
//  Pass 2:
//      Add edges from CFunc -> VarRef being written
//
//  Finally:
//      Process graph to determine when traced variables can change, allocate
//      activity flags, insert nodes to set activity flags, allocate signal
//      numbers (codes), and construct the full and incremental trace
//      functions, together with all other trace support functions.
//
//*************************************************************************

#include "config_build.h"
#include "verilatedos.h"

#include "V3Global.h"
#include "V3Trace.h"
#include "V3EmitCBase.h"
#include "V3Graph.h"
#include "V3Hashed.h"
#include "V3Stats.h"

#include <map>
#include <set>

//######################################################################
// Graph vertexes

class TraceActivityVertex : public V3GraphVertex {
    AstNode* const m_insertp;
    vlsint32_t m_activityCode;
    bool m_slow;  // If always slow, we can use the same code
public:
    enum { ACTIVITY_NEVER = ((1UL << 31) - 1) };
    enum { ACTIVITY_ALWAYS = ((1UL << 31) - 2) };
    enum { ACTIVITY_SLOW = 0 };
    TraceActivityVertex(V3Graph* graphp, AstNode* nodep, bool slow)
        : V3GraphVertex(graphp)
        , m_insertp(nodep) {
        m_activityCode = 0;
        m_slow = slow;
    }
    TraceActivityVertex(V3Graph* graphp, vlsint32_t code)
        : V3GraphVertex(graphp)
        , m_insertp(NULL) {
        m_activityCode = code;
        m_slow = false;
    }
    virtual ~TraceActivityVertex() {}
    // ACCESSORS
    AstNode* insertp() const {
        if (!m_insertp) v3fatalSrc("Null insertp; probably called on a special always/slow.");
        return m_insertp;
    }
    virtual string name() const {
        if (activityAlways()) {
            return "*ALWAYS*";
        } else {
            return (string(slow() ? "*SLOW* " : "")) + insertp()->name();
        }
    }
    virtual string dotColor() const { return slow() ? "yellowGreen" : "green"; }
    vlsint32_t activityCode() const { return m_activityCode; }
    bool activityAlways() const { return activityCode() == ACTIVITY_ALWAYS; }
    void activityCode(vlsint32_t code) { m_activityCode = code; }
    bool slow() const { return m_slow; }
    void slow(bool flag) {
        if (!flag) m_slow = false;
    }
};

class TraceCFuncVertex : public V3GraphVertex {
    AstCFunc* m_nodep;

public:
    TraceCFuncVertex(V3Graph* graphp, AstCFunc* nodep)
        : V3GraphVertex(graphp)
        , m_nodep(nodep) {}
    virtual ~TraceCFuncVertex() {}
    // ACCESSORS
    AstCFunc* nodep() const { return m_nodep; }
    virtual string name() const { return nodep()->name(); }
    virtual string dotColor() const { return "yellow"; }
    virtual FileLine* fileline() const { return nodep()->fileline(); }
};

class TraceTraceVertex : public V3GraphVertex {
    AstTraceDecl* const m_nodep;  // TRACEINC this represents
    // NULL, or other vertex with the real code() that duplicates this one
    TraceTraceVertex* m_duplicatep;

public:
    TraceTraceVertex(V3Graph* graphp, AstTraceDecl* nodep)
        : V3GraphVertex(graphp)
        , m_nodep(nodep)
        , m_duplicatep(NULL) {}
    virtual ~TraceTraceVertex() {}
    // ACCESSORS
    AstTraceDecl* nodep() const { return m_nodep; }
    virtual string name() const { return nodep()->name(); }
    virtual string dotColor() const { return "red"; }
    virtual FileLine* fileline() const { return nodep()->fileline(); }
    TraceTraceVertex* duplicatep() const { return m_duplicatep; }
    void duplicatep(TraceTraceVertex* dupp) {
        UASSERT_OBJ(!duplicatep(), nodep(), "Assigning duplicatep() to already duplicated node");
        m_duplicatep = dupp;
    }
};

class TraceVarVertex : public V3GraphVertex {
    AstVarScope* m_nodep;

public:
    TraceVarVertex(V3Graph* graphp, AstVarScope* nodep)
        : V3GraphVertex(graphp)
        , m_nodep(nodep) {}
    virtual ~TraceVarVertex() {}
    // ACCESSORS
    AstVarScope* nodep() const { return m_nodep; }
    virtual string name() const { return nodep()->name(); }
    virtual string dotColor() const { return "skyblue"; }
    virtual FileLine* fileline() const { return nodep()->fileline(); }
};

//######################################################################
// Trace state, as a visitor of each AstNode

class TraceVisitor : public EmitCBaseVisitor {
private:
    // NODE STATE
    // V3Hashed
    //  Ast*::user4()                   // V3Hashed calculation
    // Cleared entire netlist
    //  AstCFunc::user1()               // V3GraphVertex* for this node
    //  AstTraceDecl::user1()           // V3GraphVertex* for this node
    //  AstVarScope::user1()            // V3GraphVertex* for this node
    //  AstCCall::user2()               // bool; walked next list for other ccalls
    //  Ast*::user3()                   // TraceActivityVertex* for this node
    AstUser1InUse m_inuser1;
    AstUser2InUse m_inuser2;
    AstUser3InUse m_inuser3;
    // AstUser4InUse     In V3Hashed

    // STATE
    AstNodeModule* m_topModp;  // Module to add variables to
    AstScope* m_topScopep;  // Scope to add variables to
    AstCFunc* m_funcp;  // C function adding to graph
    AstTraceDecl* m_tracep;  // Trace function adding to graph
    AstVarScope* m_activityVscp;  // Activity variable
    uint32_t m_activityNumber;  // Count of fields in activity variable
    uint32_t m_code;  // Trace ident code# being assigned
    V3Graph m_graph;  // Var/CFunc tracking
    TraceActivityVertex* const m_alwaysVtxp;  // "Always trace" vertex
    bool m_finding;  // Pass one of algorithm?

    VDouble0 m_statChgSigs;  // Statistic tracking
    VDouble0 m_statUniqSigs;  // Statistic tracking
    VDouble0 m_statUniqCodes;  // Statistic tracking

    // All activity numbers applying to a given trace
    typedef std::set<uint32_t> ActCodeSet;
    // For activity set, what traces apply
    typedef std::multimap<ActCodeSet, TraceTraceVertex*> TraceVec;

    // METHODS
    VL_DEBUG_FUNC;  // Declare debug()

    void detectDuplicates() {
        UINFO(9, "Finding duplicates\n");
        // Note uses user4
        V3Hashed hashed;  // Duplicate code detection
        // Hash all of the values the traceIncs need
        for (const V3GraphVertex* itp = m_graph.verticesBeginp(); itp;
             itp = itp->verticesNextp()) {
            if (const TraceTraceVertex* const vvertexp
                = dynamic_cast<const TraceTraceVertex*>(itp)) {
                const AstTraceDecl* const nodep = vvertexp->nodep();
                if (nodep->valuep()) {
                    UASSERT_OBJ(nodep->valuep()->backp() == nodep, nodep,
                                "Trace duplicate back needs consistency,"
                                " so we can map duplicates back to TRACEINCs");
                    hashed.hash(nodep->valuep());
                    UINFO(8, "  Hashed " << std::hex << hashed.nodeHash(nodep->valuep()) << " "
                                         << nodep << endl);

                    // Just keep one node in the map and point all duplicates to this node
                    if (hashed.findDuplicate(nodep->valuep()) == hashed.end()) {
                        hashed.hashAndInsert(nodep->valuep());
                    }
                }
            }
        }
        // Find if there are any duplicates
        for (V3GraphVertex* itp = m_graph.verticesBeginp(); itp; itp = itp->verticesNextp()) {
            if (TraceTraceVertex* const vvertexp = dynamic_cast<TraceTraceVertex*>(itp)) {
                AstTraceDecl* const nodep = vvertexp->nodep();
                if (nodep->valuep() && !vvertexp->duplicatep()) {
                    V3Hashed::iterator dupit = hashed.findDuplicate(nodep->valuep());
                    if (dupit != hashed.end()) {
                        const AstTraceDecl* const dupDeclp
                            = VN_CAST_CONST(hashed.iteratorNodep(dupit)->backp(), TraceDecl);
                        UASSERT_OBJ(dupDeclp, nodep, "Trace duplicate of wrong type");
                        TraceTraceVertex* const dupvertexp
                            = dynamic_cast<TraceTraceVertex*>(dupDeclp->user1u().toGraphVertex());
                        UINFO(8, "  Orig " << nodep << endl);
                        UINFO(8, "   dup " << dupDeclp << endl);
                        // Mark the hashed node as the original and our
                        // iterating node as duplicated
                        vvertexp->duplicatep(dupvertexp);
                    }
                }
            }
        }
        hashed.clear();
    }

    void graphSimplify(bool initial) {
        if (initial) {
            // Remove all variable nodes
            for (V3GraphVertex *nextp, *itp = m_graph.verticesBeginp(); itp; itp = nextp) {
                nextp = itp->verticesNextp();
                if (TraceVarVertex* const vvertexp = dynamic_cast<TraceVarVertex*>(itp)) {
                    vvertexp->rerouteEdges(&m_graph);
                    vvertexp->unlinkDelete(&m_graph);
                }
            }
            // Remove multiple variables connecting funcs to traces
            // We do this twice, as then we have fewer edges to multiply out in the below
            // expansion.
            m_graph.removeRedundantEdges(&V3GraphEdge::followAlwaysTrue);
            // Remove all Cfunc nodes
            for (V3GraphVertex *nextp, *itp = m_graph.verticesBeginp(); itp; itp = nextp) {
                nextp = itp->verticesNextp();
                if (TraceCFuncVertex* const vvertexp = dynamic_cast<TraceCFuncVertex*>(itp)) {
                    vvertexp->rerouteEdges(&m_graph);
                    vvertexp->unlinkDelete(&m_graph);
                }
            }
        }

        // Remove multiple variables connecting funcs to traces
        m_graph.removeRedundantEdges(&V3GraphEdge::followAlwaysTrue);

        // If there are any edges from a always, keep only the always
        for (const V3GraphVertex* itp = m_graph.verticesBeginp(); itp;
             itp = itp->verticesNextp()) {
            if (const TraceTraceVertex* const vvertexp
                = dynamic_cast<const TraceTraceVertex*>(itp)) {
                // Search for the incoming always edge
                const V3GraphEdge* alwaysEdgep = NULL;
                for (const V3GraphEdge* edgep = vvertexp->inBeginp(); edgep;
                     edgep = edgep->inNextp()) {
                    const TraceActivityVertex* const actVtxp
                        = dynamic_cast<const TraceActivityVertex*>(edgep->fromp());
                    UASSERT_OBJ(actVtxp, vvertexp->nodep(),
                                "Tracing a node with FROM non activity");
                    if (actVtxp->activityAlways()) {
                        alwaysEdgep = edgep;
                        break;
                    }
                }
                // If always edge exists, remove all other edges
                if (alwaysEdgep) {
                    for (V3GraphEdge *nextp, *edgep = vvertexp->inBeginp(); edgep; edgep = nextp) {
                        nextp = edgep->inNextp();
                        if (edgep != alwaysEdgep) edgep->unlinkDelete();
                    }
                }
            }
        }

        // Activity points with no outputs can be removed
        for (V3GraphVertex *nextp, *itp = m_graph.verticesBeginp(); itp; itp = nextp) {
            nextp = itp->verticesNextp();
            if (TraceActivityVertex* const vtxp = dynamic_cast<TraceActivityVertex*>(itp)) {
                // Leave in the always vertex for later use.
                if (vtxp != m_alwaysVtxp && !vtxp->outBeginp()) { vtxp->unlinkDelete(&m_graph); }
            }
        }
    }

    uint32_t assignactivityNumbers() {
        uint32_t activityNumber = 1;  // Note 0 indicates "slow"
        for (V3GraphVertex* itp = m_graph.verticesBeginp(); itp; itp = itp->verticesNextp()) {
            if (TraceActivityVertex* const vvertexp = dynamic_cast<TraceActivityVertex*>(itp)) {
                if (vvertexp != m_alwaysVtxp) {
                    if (vvertexp->slow()) {
                        vvertexp->activityCode(TraceActivityVertex::ACTIVITY_SLOW);
                    } else {
                        vvertexp->activityCode(activityNumber++);
                    }
                }
            }
        }
        return activityNumber;
    }

    void sortTraces(TraceVec& traces, uint32_t& nFullCodes, uint32_t& nChgCodes) {
        // Populate sort structure
        traces.clear();
        nFullCodes = 0;
        nChgCodes = 0;
        for (V3GraphVertex* itp = m_graph.verticesBeginp(); itp; itp = itp->verticesNextp()) {
            if (TraceTraceVertex* const vtxp = dynamic_cast<TraceTraceVertex*>(itp)) {
                ActCodeSet actSet;
                UINFO(9, "  Add to sort: " << vtxp << endl);
                if (debug() >= 9) vtxp->nodep()->dumpTree(cout, "-   trnode: ");
                for (const V3GraphEdge* edgep = vtxp->inBeginp(); edgep;
                     edgep = edgep->inNextp()) {
                    const TraceActivityVertex* const cfvertexp
                        = dynamic_cast<const TraceActivityVertex*>(edgep->fromp());
                    UASSERT_OBJ(cfvertexp, vtxp->nodep(),
                                "Should have been function pointing to this trace");
                    UINFO(9, "   Activity: " << cfvertexp << endl);
                    if (cfvertexp->activityAlways()) {
                        // If code 0, we always trace; ignore other codes
                        actSet.insert(TraceActivityVertex::ACTIVITY_ALWAYS);
                    } else {
                        actSet.insert(cfvertexp->activityCode());
                    }
                }
                UASSERT_OBJ(actSet.count(TraceActivityVertex::ACTIVITY_ALWAYS) == 0
                                || actSet.size() == 1,
                            vtxp->nodep(), "Always active trace has further triggers");
                // Count nodes
                if (!vtxp->duplicatep()) {
                    const uint32_t inc = vtxp->nodep()->codeInc();
                    nFullCodes += inc;
                    if (!actSet.empty()) nChgCodes += inc;
                }
                // If a trace doesn't have activity, it's constant, and we
                // don't need to track changes on it.
                // We put constants and non-changers last, as then the
                // prevvalue vector is more compacted
                if (actSet.empty()) actSet.insert(TraceActivityVertex::ACTIVITY_NEVER);
                traces.insert(make_pair(actSet, vtxp));
            }
        }
    }

    void graphOptimize() {
        // Assign initial activity numbers to activity vertices
        assignactivityNumbers();

        // Sort the traces by activity sets
        TraceVec traces;
        uint32_t unused1, unused2;
        sortTraces(traces, unused1, unused2);

        // For each activity set with only a small number of signals, make those
        // signals always traced, as it's cheaper to check a few value changes
        // than to test a lot of activity flags
        TraceVec::iterator it = traces.begin();
        const TraceVec::iterator end = traces.end();
        while (it != end) {
            TraceVec::iterator head = it;
            // Count the value comparisons in all traces under this activity set
            uint32_t valueComparisons = 0;
            const ActCodeSet& actSet = it->first;
            for (; it != end && it->first == actSet; ++it) {
                valueComparisons += it->second->nodep()->codeInc();
            }
            // Leave alone always changing, never changing and signals only set in slow code
            if (actSet.count(TraceActivityVertex::ACTIVITY_ALWAYS)) continue;
            if (actSet.count(TraceActivityVertex::ACTIVITY_NEVER)) continue;
            if (actSet.count(TraceActivityVertex::ACTIVITY_SLOW) && actSet.size() == 1) continue;
            // If the value comparisons are cheaper to perform than checking the
            // activity flags make the signals always traced. Note this cost
            // equation is heuristic.
            if (valueComparisons < actSet.size() * 2) {
                for (; head != it; ++head) {
                    new V3GraphEdge(&m_graph, m_alwaysVtxp, head->second, 1);
                }
            }
        }

        graphSimplify(false);
    }

    AstNode* selectActivity(FileLine* flp, uint32_t acode, bool lvalue) {
        return new AstArraySel(flp, new AstVarRef(flp, m_activityVscp, lvalue), acode);
    }

    void createActivityFlags() {
        // Assign final activity numbers
        m_activityNumber = assignactivityNumbers();

        // Create an array of bytes, not a bit vector, as they can be set
        // atomically by mtasks, and are cheaper to set (not need for
        // read-modify-write on the C type), and the speed of the tracing code
        // is the same on largish designs.
        FileLine* const flp = m_topScopep->fileline();
        AstNodeDType* const newScalarDtp = new AstBasicDType(flp, VFlagLogicPacked(), 1);
        v3Global.rootp()->typeTablep()->addTypesp(newScalarDtp);
        AstRange* const newArange = new AstRange(flp, VNumRange(m_activityNumber - 1, 0, false));
        AstNodeDType* const newArrDtp = new AstUnpackArrayDType(flp, newScalarDtp, newArange);
        v3Global.rootp()->typeTablep()->addTypesp(newArrDtp);
        AstVar* const newvarp
            = new AstVar(flp, AstVarType::MODULETEMP, "__Vm_traceActivity", newArrDtp);
        m_topModp->addStmtp(newvarp);
        AstVarScope* const newvscp = new AstVarScope(flp, m_topScopep, newvarp);
        m_topScopep->addVarp(newvscp);
        m_activityVscp = newvscp;

        // Insert activity setters
        for (const V3GraphVertex* itp = m_graph.verticesBeginp(); itp;
             itp = itp->verticesNextp()) {
            if (const TraceActivityVertex* const vtxp
                = dynamic_cast<const TraceActivityVertex*>(itp)) {
                if (!vtxp->activityAlways()) {
                    FileLine* const fl = vtxp->insertp()->fileline();
                    const uint32_t acode = vtxp->activityCode();
                    AstAssign* const setterp
                        = new AstAssign(fl, selectActivity(fl, acode, true),
                                        new AstConst(fl, AstConst::LogicTrue()));
                    if (AstCCall* const callp = VN_CAST(vtxp->insertp(), CCall)) {
                        callp->addNextHere(setterp);
                    } else if (AstCFunc* const funcp = VN_CAST(vtxp->insertp(), CFunc)) {
                        funcp->addStmtsp(setterp);
                    } else {
                        vtxp->insertp()->v3fatalSrc("Bad trace activity vertex");
                    }
                }
            }
        }
    }

    AstCFunc* newCFunc(AstCFuncType type, AstCFunc* callfromp, AstCFunc* regp, int& funcNump) {
        // Create new function
        string name;
        switch (type) {
        case AstCFuncType::TRACE_FULL: name = "traceFullTop"; break;
        case AstCFuncType::TRACE_FULL_SUB: name = "traceFullSub"; break;
        case AstCFuncType::TRACE_CHANGE: name = "traceChgTop"; break;
        case AstCFuncType::TRACE_CHANGE_SUB: name = "traceChgSub"; break;
        default: m_topScopep->v3fatalSrc("Bad trace function type");
        }
        name += cvtToStr(funcNump++);
        FileLine* const flp = m_topScopep->fileline();
        AstCFunc* const funcp = new AstCFunc(flp, name, m_topScopep);
        const string argTypes("void* userp, " + v3Global.opt.traceClassBase() + "* tracep");
        funcp->argTypes(argTypes);
        funcp->funcType(type);
        funcp->slow(type == AstCFuncType::TRACE_FULL || type == AstCFuncType::TRACE_FULL_SUB);
        funcp->symProlog(true);
        funcp->declPrivate(true);
        // Add it to top scope
        m_topScopep->addActivep(funcp);
        // Add call to new function
        if (callfromp) {
            AstCCall* callp = new AstCCall(funcp->fileline(), funcp);
            callp->argTypes("userp, tracep");
            callfromp->addStmtsp(callp);
        }
        // Register function
        if (regp) {
            string registration = "tracep->add";
            if (type == AstCFuncType::TRACE_FULL) {
                registration += "Full";
            } else if (type == AstCFuncType::TRACE_CHANGE) {
                registration += "Chg";
            } else {
                funcp->v3fatal("Don't know how to register this type of function");
            }
            registration += "Cb(&" + protect(name) + ", __VlSymsp);\n";
            AstCStmt* const stmtp = new AstCStmt(flp, registration);
            regp->addStmtsp(stmtp);
        }
        // Add global activity check to TRACE_CHANGE functions
        if (type == AstCFuncType::TRACE_CHANGE) {
            funcp->addInitsp(
                new AstCStmt(flp, string("if (VL_UNLIKELY(!vlSymsp->__Vm_activity)) return;\n")));
        }
        // Done
        UINFO(5, "  newCFunc " << funcp << endl);
        return funcp;
    }

    void assignDeclCode(AstTraceDecl* nodep) {
        if (!nodep->code()) {
            nodep->code(m_code);
            m_code += nodep->codeInc();
            m_statUniqCodes += nodep->codeInc();
            ++m_statUniqSigs;
        }
    }

    void createFullTraceFunction(const TraceVec& traces, uint32_t nAllCodes, uint32_t parallelism,
                                 AstCFunc* regFuncp) {
        const int splitLimit = v3Global.opt.outputSplitCTrace() ? v3Global.opt.outputSplitCTrace()
                                                                : std::numeric_limits<int>::max();

        int topFuncNum = 0;
        int subFuncNum = 0;
        TraceVec::const_iterator it = traces.begin();
        while (it != traces.end()) {
            AstCFunc* topFuncp = NULL;
            AstCFunc* subFuncp = NULL;
            int subStmts = 0;
            const uint32_t maxCodes = (nAllCodes + parallelism - 1) / parallelism;
            uint32_t nCodes = 0;
            for (; nCodes < maxCodes && it != traces.end(); ++it) {
                const TraceTraceVertex* const vtxp = it->second;
                AstTraceDecl* const declp = vtxp->nodep();
                if (const TraceTraceVertex* const canonVtxp = vtxp->duplicatep()) {
                    // This is a duplicate trace node. We will assign the signal
                    // number to the canonical node, and emit this as an alias, so
                    // no need to create a TraceInc node.
                    AstTraceDecl* const canonDeclp = canonVtxp->nodep();
                    UASSERT_OBJ(!canonVtxp->duplicatep(), canonDeclp,
                                "Canonical node is a duplicate");
                    assignDeclCode(canonDeclp);
                    declp->code(canonDeclp->code());
                } else {
                    // This is a canonical trace node. Assign signal number and
                    // add a TraceInc node to the full dump function.
                    assignDeclCode(declp);

                    // Create top function if not yet created
                    if (!topFuncp) {
                        topFuncp = newCFunc(AstCFuncType::TRACE_FULL, NULL, regFuncp, topFuncNum);
                    }

                    // Crate new sub function if required
                    if (!subFuncp || subStmts > splitLimit) {
                        subStmts = 0;
                        subFuncp
                            = newCFunc(AstCFuncType::TRACE_FULL_SUB, topFuncp, NULL, subFuncNum);
                    }

                    // Add TraceInc node
                    AstTraceInc* const incp = new AstTraceInc(declp->fileline(), declp, true);
                    subFuncp->addStmtsp(incp);
                    subStmts += EmitCBaseCounterVisitor(incp).count();

                    // Track partitioning
                    nCodes += declp->codeInc();
                }
            }
            if (topFuncp) {  // might be NULL if all trailing entries were duplicates
                UINFO(5, "traceFullTop" << topFuncNum - 1 << " codes: " << nCodes << "/"
                                        << maxCodes << endl);
            }
        }
    }

    void createChgTraceFunctions(const TraceVec& traces, uint32_t nAllCodes, uint32_t parallelism,
                                 AstCFunc* regFuncp) {
        const int splitLimit = v3Global.opt.outputSplitCTrace() ? v3Global.opt.outputSplitCTrace()
                                                                : std::numeric_limits<int>::max();
        int topFuncNum = 0;
        int subFuncNum = 0;
        TraceVec::const_iterator it = traces.begin();
        while (it != traces.end()) {
            AstCFunc* topFuncp = NULL;
            AstCFunc* subFuncp = NULL;
            int subStmts = 0;
            const uint32_t maxCodes = (nAllCodes + parallelism - 1) / parallelism;
            uint32_t nCodes = 0;
            const ActCodeSet* prevActSet = NULL;
            AstIf* ifp = NULL;
            for (; nCodes < maxCodes && it != traces.end(); ++it) {
                const TraceTraceVertex* const vtxp = it->second;
                // This is a duplicate decl, no need to add it to incremental dump
                if (vtxp->duplicatep()) continue;
                const ActCodeSet& actSet = it->first;
                // Traced value never changes, no need to add it to incremental dump
                if (actSet.count(TraceActivityVertex::ACTIVITY_NEVER)) continue;

                // Create top function if not yet created
                if (!topFuncp) {
                    topFuncp = newCFunc(AstCFuncType::TRACE_CHANGE, NULL, regFuncp, topFuncNum);
                }

                // Crate new sub function if required
                if (!subFuncp || subStmts > splitLimit) {
                    subStmts = 0;
                    subFuncp
                        = newCFunc(AstCFuncType::TRACE_CHANGE_SUB, topFuncp, NULL, subFuncNum);
                    prevActSet = NULL;
                    ifp = NULL;
                }

                // If required, create the conditional node checking the activity flags
                if (!prevActSet || actSet != *prevActSet) {
                    FileLine* const flp = m_topScopep->fileline();
                    bool always = actSet.count(TraceActivityVertex::ACTIVITY_ALWAYS) != 0;
                    AstNode* condp = NULL;
                    if (always) {
                        condp = new AstConst(flp, 1);  // Always true, will be folded later
                    } else {
                        for (ActCodeSet::iterator it = actSet.begin(); it != actSet.end(); ++it) {
                            const uint32_t actCode = *it;
                            AstNode* const selp = selectActivity(flp, actCode, false);
                            condp = condp ? new AstOr(flp, condp, selp) : selp;
                        }
                    }
                    ifp = new AstIf(flp, condp, NULL, NULL);
                    if (!always) { ifp->branchPred(VBranchPred::BP_UNLIKELY); }
                    subFuncp->addStmtsp(ifp);
                    subStmts += EmitCBaseCounterVisitor(ifp).count();
                    prevActSet = &actSet;
                }

                // Add TraceInc node
                AstTraceDecl* const declp = vtxp->nodep();
                AstTraceInc* const incp = new AstTraceInc(declp->fileline(), declp, false);
                ifp->addIfsp(incp);
                subStmts += EmitCBaseCounterVisitor(incp).count();

                // Track partitioning
                nCodes += declp->codeInc();
            }
            if (topFuncp) {  // might be NULL if all trailing entries were duplicates/constants
                UINFO(5, "traceChgTop" << topFuncNum - 1 << " codes: " << nCodes << "/" << maxCodes
                                       << endl);
            }
        }
    }

    void createCleanupFunction(AstCFunc* regFuncp) {
        FileLine* const fl = m_topScopep->fileline();
        AstCFunc* const cleanupFuncp = new AstCFunc(fl, "traceCleanup", m_topScopep);
        const string argTypes("void* userp, " + v3Global.opt.traceClassBase() + "* /*unused*/");
        cleanupFuncp->argTypes(argTypes);
        cleanupFuncp->funcType(AstCFuncType::TRACE_CLEANUP);
        cleanupFuncp->slow(false);
        cleanupFuncp->symProlog(true);
        cleanupFuncp->declPrivate(true);
        m_topScopep->addActivep(cleanupFuncp);

        // Register it
        regFuncp->addStmtsp(new AstCStmt(
            fl, string("tracep->addCleanupCb(&" + protect("traceCleanup") + ", __VlSymsp);\n")));

        // Clear global activity flag
        cleanupFuncp->addStmtsp(
            new AstCStmt(m_topScopep->fileline(), string("vlSymsp->__Vm_activity = false;\n")));

        // Clear fine grained activity flags
        for (uint32_t i = 0; i < m_activityNumber; ++i) {
            AstNode* const clrp = new AstAssign(fl, selectActivity(fl, i, true),
                                                new AstConst(fl, AstConst::LogicFalse()));
            cleanupFuncp->addStmtsp(clrp);
        }
    }

    void createTraceFunctions() {
        // Detect and remove duplicate values
        detectDuplicates();

        // Simplify & optimize the graph
        if (debug() >= 6) m_graph.dumpDotFilePrefixed("trace_pre");
        graphSimplify(true);
        if (debug() >= 6) m_graph.dumpDotFilePrefixed("trace_simplified");
        graphOptimize();
        if (debug() >= 6) m_graph.dumpDotFilePrefixed("trace_optimized");

        // Create the fine grained activity flags
        createActivityFlags();

        // Form a sorted list of the traces we are interested in
        TraceVec traces;  // The sorted traces
        // We will split functions such that each have to dump roughly the same amount of data
        // for this we need to keep tack of the number of codes used by the trace functions.
        uint32_t nFullCodes = 0;  // Number of non-duplicate codes (need to go into full* dump)
        uint32_t nChgCodes = 0;  // Number of non-consant codes (need to go in to chg* dump)
        sortTraces(traces, nFullCodes, nChgCodes);

        UINFO(5, "nFullCodes: " << nFullCodes << " nChgCodes: " << nChgCodes << endl);

        // Our keys are now sorted to have same activity number adjacent, then
        // by trace order. (Better would be execution order for cache
        // efficiency....) Last are constants and non-changers, as then the
        // last value vector is more compact

        // Create the trace registration function
        AstCFunc* const regFuncp
            = new AstCFunc(m_topScopep->fileline(), "traceRegister", m_topScopep);
        regFuncp->argTypes(v3Global.opt.traceClassBase() + "* tracep");
        regFuncp->funcType(AstCFuncType::TRACE_REGISTER);
        regFuncp->slow(true);
        regFuncp->isStatic(false);
        regFuncp->declPrivate(true);
        m_topScopep->addActivep(regFuncp);

        const int parallelism = 1;  // Note: will bump this later, code below works for any value

        // Create the full dump functions, also allocates signal numbers
        createFullTraceFunction(traces, nFullCodes, parallelism, regFuncp);

        // Create the incremental dump functions
        createChgTraceFunctions(traces, nChgCodes, parallelism, regFuncp);

        // Remove refs to traced values from TraceDecl nodes, these have now moved under
        // TraceInc
        for (TraceVec::iterator it = traces.begin(); it != traces.end(); ++it) {
            AstNode* const valuep = it->second->nodep()->valuep();
            valuep->unlinkFrBack();
            valuep->deleteTree();
        }

        // Create the trace cleanup function clearing the activity flags
        createCleanupFunction(regFuncp);
    }

    TraceCFuncVertex* getCFuncVertexp(AstCFunc* nodep) {
        TraceCFuncVertex* vertexp
            = dynamic_cast<TraceCFuncVertex*>(nodep->user1u().toGraphVertex());
        if (!vertexp) {
            vertexp = new TraceCFuncVertex(&m_graph, nodep);
            nodep->user1p(vertexp);
        }
        return vertexp;
    }
    TraceActivityVertex* getActivityVertexp(AstNode* nodep, bool slow) {
        TraceActivityVertex* vertexp
            = dynamic_cast<TraceActivityVertex*>(nodep->user3u().toGraphVertex());
        if (!vertexp) {
            vertexp = new TraceActivityVertex(&m_graph, nodep, slow);
            nodep->user3p(vertexp);
        }
        vertexp->slow(slow);
        return vertexp;
    }

    // VISITORS
    virtual void visit(AstNetlist* nodep) VL_OVERRIDE {
        m_code = 1;  // Multiple TopScopes will require fixing how code#s
        // are assigned as duplicate varscopes must result in the same tracing code#.

        // Add vertexes for all TraceDecl, and edges from VARs each trace looks at
        m_finding = false;
        iterateChildren(nodep);

        // Add vertexes for all CFUNCs, and edges to VARs the func sets
        m_finding = true;
        iterateChildren(nodep);

        // Create the trace functions and insert them into the tree
        createTraceFunctions();
    }
    virtual void visit(AstNodeModule* nodep) VL_OVERRIDE {
        if (nodep->isTop()) m_topModp = nodep;
        iterateChildren(nodep);
    }
    virtual void visit(AstTopScope* nodep) VL_OVERRIDE {
        AstScope* const scopep = nodep->scopep();
        UASSERT_OBJ(scopep, nodep, "No scope found on top level");
        m_topScopep = scopep;
        iterateChildren(nodep);
    }
    virtual void visit(AstCCall* nodep) VL_OVERRIDE {
        UINFO(8, "   CCALL " << nodep << endl);
        if (!m_finding && !nodep->user2()) {
            // See if there are other calls in same statement list;
            // If so, all funcs might share the same activity code
            TraceActivityVertex* const activityVtxp
                = getActivityVertexp(nodep, nodep->funcp()->slow());
            for (AstNode* nextp = nodep; nextp; nextp = nextp->nextp()) {
                if (AstCCall* const ccallp = VN_CAST(nextp, CCall)) {
                    ccallp->user2(true);  // Processed
                    UINFO(8, "     SubCCALL " << ccallp << endl);
                    V3GraphVertex* const ccallFuncVtxp = getCFuncVertexp(ccallp->funcp());
                    activityVtxp->slow(ccallp->funcp()->slow());
                    new V3GraphEdge(&m_graph, activityVtxp, ccallFuncVtxp, 1);
                }
            }
        }
        iterateChildren(nodep);
    }
    virtual void visit(AstCFunc* nodep) VL_OVERRIDE {
        UINFO(8, "   CFUNC " << nodep << endl);
        V3GraphVertex* const funcVtxp = getCFuncVertexp(nodep);
        if (!m_finding) {  // If public, we need a unique activity code to allow for sets
                           // directly in this func
            if (nodep->funcPublic() || nodep->dpiExport() || nodep == v3Global.rootp()->evalp()) {
                V3GraphVertex* const activityVtxp = getActivityVertexp(nodep, nodep->slow());
                new V3GraphEdge(&m_graph, activityVtxp, funcVtxp, 1);
            }
        }
        m_funcp = nodep;
        iterateChildren(nodep);
        m_funcp = NULL;
    }
    virtual void visit(AstTraceDecl* nodep) VL_OVERRIDE {
        UINFO(8, "   TRACE " << nodep << endl);
        if (!m_finding) {
            V3GraphVertex* const vertexp = new TraceTraceVertex(&m_graph, nodep);
            nodep->user1p(vertexp);

            UASSERT_OBJ(m_funcp, nodep, "Trace not under func");
            m_tracep = nodep;
            iterateChildren(nodep);
            m_tracep = NULL;
        }
    }
    virtual void visit(AstVarRef* nodep) VL_OVERRIDE {
        if (m_tracep) {
            UASSERT_OBJ(nodep->varScopep(), nodep, "No var scope?");
            UASSERT_OBJ(!nodep->lvalue(), nodep, "Lvalue in trace?  Should be const.");
            V3GraphVertex* varVtxp = nodep->varScopep()->user1u().toGraphVertex();
            if (!varVtxp) {
                varVtxp = new TraceVarVertex(&m_graph, nodep->varScopep());
                nodep->varScopep()->user1p(varVtxp);
            }
            V3GraphVertex* const traceVtxp = m_tracep->user1u().toGraphVertex();
            new V3GraphEdge(&m_graph, varVtxp, traceVtxp, 1);
            if (nodep->varp()->isPrimaryInish()  // Always need to trace primary inputs
                || nodep->varp()->isSigPublic()) {  // Or ones user can change
                new V3GraphEdge(&m_graph, m_alwaysVtxp, traceVtxp, 1);
            }
        } else if (m_funcp && m_finding && nodep->lvalue()) {
            UASSERT_OBJ(nodep->varScopep(), nodep, "No var scope?");
            V3GraphVertex* const funcVtxp = getCFuncVertexp(m_funcp);
            V3GraphVertex* const varVtxp = nodep->varScopep()->user1u().toGraphVertex();
            if (varVtxp) {  // else we're not tracing this signal
                new V3GraphEdge(&m_graph, funcVtxp, varVtxp, 1);
            }
        }
    }
    //--------------------
    virtual void visit(AstNode* nodep) VL_OVERRIDE { iterateChildren(nodep); }

public:
    // CONSTRUCTORS
    explicit TraceVisitor(AstNetlist* nodep)
        : m_alwaysVtxp(new TraceActivityVertex(&m_graph, TraceActivityVertex::ACTIVITY_ALWAYS)) {
        m_funcp = NULL;
        m_tracep = NULL;
        m_topModp = NULL;
        m_topScopep = NULL;
        m_finding = false;
        m_activityVscp = NULL;
        m_activityNumber = 0;
        m_code = 0;
        iterate(nodep);
    }
    virtual ~TraceVisitor() {
        V3Stats::addStat("Tracing, Unique changing signals", m_statChgSigs);
        V3Stats::addStat("Tracing, Unique traced signals", m_statUniqSigs);
        V3Stats::addStat("Tracing, Unique trace codes", m_statUniqCodes);
    }
};

//######################################################################
// Trace class functions

void V3Trace::traceAll(AstNetlist* nodep) {
    UINFO(2, __FUNCTION__ << ": " << endl);
    { TraceVisitor visitor(nodep); }  // Destruct before checking
    V3Global::dumpCheckGlobalTree("trace", 0, v3Global.opt.dumpTreeLevel(__FILE__) >= 3);
}