Cleaned some code and removed some warning messages

[dyninst.git] / paradynd / src / metricFocusNode.C

/*
 * Copyright (c) 1996 Barton P. Miller
 * 
 * We provide the Paradyn Parallel Performance Tools (below
 * described as Paradyn") on an AS IS basis, and do not warrant its
 * validity or performance.  We reserve the right to update, modify,
 * or discontinue this software at any time.  We shall have no
 * obligation to supply such updates or modifications or any other
 * form of support to you.
 * 
 * This license is for research uses.  For such uses, there is no
 * charge. We define "research use" to mean you may freely use it
 * inside your organization for whatever purposes you see fit. But you
 * may not re-distribute Paradyn or parts of Paradyn, in any form
 * source or binary (including derivatives), electronic or otherwise,
 * to any other organization or entity without our permission.
 * 
 * (for other uses, please contact us at paradyn@cs.wisc.edu)
 * 
 * All warranties, including without limitation, any warranty of
 * merchantability or fitness for a particular purpose, are hereby
 * excluded.
 * 
 * By your use of Paradyn, you understand and agree that we (or any
 * other person or entity with proprietary rights in Paradyn) are
 * under no obligation to provide either maintenance services,
 * update services, notices of latent defects, or correction of
 * defects for Paradyn.
 * 
 * Even if advised of the possibility of such damages, under no
 * circumstances shall we (or any other person or entity with
 * proprietary rights in the software licensed hereunder) be liable
 * to you or any third party for direct, indirect, or consequential
 * damages of any character regardless of type of action, including,
 * without limitation, loss of profits, loss of use, loss of good
 * will, or computer failure or malfunction.  You agree to indemnify
 * us (and any other person or entity with proprietary rights in the
 * software licensed hereunder) for any and all liability it may
 * incur to third parties resulting from your use of Paradyn.
 */

#include "util/h/headers.h"
#include <limits.h>
#include <assert.h>

#include "rtinst/h/rtinst.h"
#include "rtinst/h/trace.h"
#include "util/h/aggregateSample.h"
#include "symtab.h"
#include "process.h"
#include "inst.h"
#include "instP.h"
#include "dyninstP.h"
#include "ast.h"
#include "util.h"
#include "comm.h"
#include "internalMetrics.h"
#include <strstream.h>
#include "init.h"
#include "perfStream.h"
#include "main.h"
#include "stats.h"
#include "dynrpc.h"
#include "paradynd/src/mdld.h"
#include "util/h/Timer.h"
#include "showerror.h"
#include "costmetrics.h"
#include "metric.h"
#include "util/h/debugOstream.h"

// The following vrbles were defined in process.C:
extern debug_ostream attach_cerr;
extern debug_ostream inferiorrpc_cerr;
extern debug_ostream shmsample_cerr;
extern debug_ostream forkexec_cerr;
extern debug_ostream metric_cerr;

extern int activeCT;
extern int memCT;
extern int inferiorMemAvailable;
extern vector<unsigned> getAllTrampsAtPoint(instInstance *instance);
static int internalMetricCounterId = 0;

void flush_batch_buffer();
void batchSampleData(int mid, double startTimeStamp, double endTimeStamp,
                     double value, unsigned val_weight, bool internal_metric);

#ifdef sparc_tmc_cmost7_3
extern int getNumberOfCPUs();
#endif

double currentPredictedCost = 0.0;

dictionary_hash <unsigned, metricDefinitionNode*> midToMiMap(uiHash);
   // maps low-level counter-ids to metricDefinitionNodes

unsigned mdnHash(const metricDefinitionNode *&mdn) {
  return ((unsigned)mdn) >> 2; // assume all addrs are 4-byte aligned
//  return ((unsigned) mdn);
}

unsigned componentMdnPtrHash(metricDefinitionNode * const &ptr) {
   // maybe assert that "ptr" isn't for an aggregate mi
   return string::hash(ptr->getFullName());
}


dictionary_hash<unsigned, metricDefinitionNode*> allMIs(uiHash);
dictionary_hash<string, metricDefinitionNode*> allMIComponents(string::hash);
vector<internalMetric*> internalMetric::allInternalMetrics;

// used to indicate the mi is no longer used.
#define DELETED_MI 1
#define MILLION 1000000.0

bool mdl_internal_metric_data(const string& metric_name, mdl_inst_data& result) {
  unsigned size = internalMetric::allInternalMetrics.size();
  for (unsigned u=0; u<size; u++) {
    internalMetric *theIMetric = internalMetric::allInternalMetrics[u];
    if (theIMetric->name() == metric_name) {
      result.aggregate = theIMetric->aggregate();
      result.style = theIMetric->style();
      return true;
    }
  }

  for (unsigned u2=0; u2< costMetric::allCostMetrics.size(); u2++) {
    if (costMetric::allCostMetrics[u2]->name() == metric_name) {
      result.aggregate = costMetric::allCostMetrics[u2]->aggregate();
      result.style = costMetric::allCostMetrics[u2]->style();
      return true;
    }
  }

  return (mdl_metric_data(metric_name, result));
}

// for non-aggregate metrics
metricDefinitionNode::metricDefinitionNode(process *p, const string& met_name, 
                                   const vector< vector<string> >& foc,
                                   const vector< vector<string> >& component_foc,
                                   const string& component_flat_name, int agg_style)
: aggregate_(false), 
  aggOp(agg_style), // CM5 metrics need aggOp to be set
  inserted_(false), installed_(false), met_(met_name),
  focus_(foc), component_focus(component_foc),
  flat_name_(component_flat_name),
  aggSample(0),
  cumulativeValue(0.0), samples(0),
  id_(-1), originalCost_(0.0), proc_(p)
{
  mdl_inst_data md;
  bool aflag;
  aflag=mdl_internal_metric_data(met_name, md);
  assert(aflag);
  style_ = md.style;
}

// for aggregate metrics
metricDefinitionNode::metricDefinitionNode(const string& metric_name,
                                           const vector< vector<string> >& foc,
                                           const string& cat_name, 
                                           vector<metricDefinitionNode*>& parts,
                                           int agg_op)
: aggregate_(true), aggOp(agg_op), inserted_(false),  installed_(false),
  met_(metric_name), focus_(foc),
  flat_name_(cat_name), components(parts),
  aggSample(agg_op),
  id_(-1), originalCost_(0.0), proc_(NULL)
{
  unsigned p_size = parts.size();
  for (unsigned u=0; u<p_size; u++) {
    metricDefinitionNode *mi = parts[u];
    mi->aggregators += this;
    mi->samples += aggSample.newComponent();
  }
}

// check for "special" metrics that are computed directly by paradynd 
// if a cost of an internal metric is asked for, enable=false
metricDefinitionNode *doInternalMetric(vector< vector<string> >& canon_focus,
                                       vector< vector<string> >& component_canon_focus,
                                       string& metric_name, string& flat_name,
                                       bool enable, bool& matched)
{
  // called by createMetricInstance, below.
  // return values:
  //   a valid metricDefinitionNode* when successful
  //   -1 --> enable was false
  //   -2 --> not legal to instrument this focus
  //   NULL --> a more serious error (probably metric-is-unknown)

  matched = false;
  metricDefinitionNode *mn = 0; 

  // check to see if this is an internal metric
  unsigned im_size = internalMetric::allInternalMetrics.size();
  for (unsigned im_index=0; im_index<im_size; im_index++){
    internalMetric *theIMetric = internalMetric::allInternalMetrics[im_index];
    if (theIMetric->name() == metric_name) {
      matched = true;
      if (!enable)
         return (metricDefinitionNode*)-1;

      if (!theIMetric->legalToInst(canon_focus))
         // Paradyn will handle this case and report appropriate error msg
         return (metricDefinitionNode*)-2;

      mn = new metricDefinitionNode(NULL, metric_name, canon_focus,
                                    component_canon_focus,
                                    flat_name, theIMetric->aggregate());
      assert(mn);

      theIMetric->enableNewInstance(mn);
      return(mn);
    }
  }

  // check to see if this is a cost metric
  for (unsigned i=0; i < costMetric::allCostMetrics.size(); i++){
     if(costMetric::allCostMetrics[i]->name() == metric_name){
          matched = true;
          if (!enable) return (metricDefinitionNode*)-1;
          costMetric *nc = costMetric::allCostMetrics[i];
          if (!nc->legalToInst(canon_focus)) return (metricDefinitionNode*)-2;

          mn = new metricDefinitionNode(NULL, metric_name, canon_focus,
                                        component_canon_focus,
                                        flat_name, nc->aggregate());
          assert(mn);

          nc->enable(mn); 

          return(mn);
     }
  }

  // No matches found among internal or cost metrics
  return NULL;
}

// the following should probably be made a public static member fn of class metric
string metricAndCanonFocus2FlatName(const string &metricName,
                                    const vector< vector<string> > &canonFocus) {
   string result = metricName;

   for (unsigned hierarchy=0; hierarchy < canonFocus.size(); hierarchy++)
      for (unsigned component=0; component < canonFocus[hierarchy].size();
           component++)
         result += canonFocus[hierarchy][component];

   return result;
}

// the following should probably be made a public static member fn of class metric
static bool focus2CanonicalFocus(const vector<unsigned> &focus,
                                 vector< vector<string> > &canonFocus,
                                 bool important) {
   // takes in "focus", writes to "canonFocus".  Returns true iff successful.
   // if "important" is false, don't print error msg on failure (called by guessCost();
   // no mi is really being created)

   vector< vector<string> > unCanonFocus;
   if (!resource::foc_to_strings(unCanonFocus, focus, important)) { // writes to unCanonFocus
      if (important)
         cerr << "focus2CanonicalFocus failed since resource::foc_to_strings failed" << endl;
      return false;
   }

   resource::make_canonical(unCanonFocus, canonFocus);

   return true;
}

static void print_focus(debug_ostream &os, vector< vector<string> > &focus) {
   for (unsigned a=0; a < focus.size(); a++) {
      for (unsigned b=0; b < focus[a].size(); b++)
         os << '/' << focus[a][b];

      if (a < focus.size()-1)
         os << ',';
   }
   os << endl;
}

metricDefinitionNode *createMetricInstance(string& metric_name, 
                                           vector<u_int>& focus,
                                           bool enable, // true if for real; false for guessCost()
                                           bool& internal)
{
    vector< vector<string> > canonicalFocus;
    if (!focus2CanonicalFocus(focus, canonicalFocus, enable)) {
       if (enable) // for real, so an error msg is appropriate
          cerr << "createMetricInstance failed because focus2CanonicalFocus failed" << endl;

       return NULL;
    }

    string flat_name = metricAndCanonFocus2FlatName(metric_name, canonicalFocus);

    // first see if it is already defined.
    dictionary_hash_iter<unsigned, metricDefinitionNode*> mdi(allMIs);

/*
 * See if we can find the requested metric instance.
 *   Currently this is only used to cache structs built for cost requests 
 *   which are then instantiated.  This could be used as a general system
 *   to request find sub-metrics that are already.defines and use them to
 *   reduce cost.  This would require adding the componenets of an aggregate
 *   into the allMIs list since searching tends to be top down, not bottom
 *   up.  This would also require adding a ref count to many of the structures
 *   so they only get deleted when we are really done with them.
 *
 */

    unsigned key;
    metricDefinitionNode *mi= NULL;

    // TODO -- a dictionary search here will be much faster
    while (mdi.next(key, mi))
      if (mi->getFullName() == flat_name) {
        metric_cerr << "createMetricInstance: mi with flat_name " << flat_name << " already exists! using it" << endl;
        return mi; // this metricDefinitionNode has already been defined
      }

    if (mdl_can_do(metric_name)) {
      internal = false;

      /* select the processes that should be instrumented. We skip process
         that have exited, and processes that have been created but are not
         completely initialized yet.
         If we try to insert instrumentation in a process that is not ready
         yet, we get a core dump.
         A process is ready when it is not in neonatal state and the 
         isBootstrappedYet returns true.
      */
      vector<process*> procs;

      for (unsigned u = 0; u < processVec.size(); u++) {
        if (processVec[u]->status()==exited || processVec[u]->status()==neonatal
            || processVec[u]->isBootstrappedYet())
          procs += processVec[u];
      }

      if (procs.size() == 0) {
        // there are no processes to instrument
        return NULL;
      }

      mi = mdl_do(canonicalFocus, metric_name, flat_name, procs, false);
      if (mi == NULL) {
         metric_cerr << "createMetricInstance failed since mdl_do failed" << endl;
         metric_cerr << "metric name was " << metric_name << "; focus was ";
         print_focus(metric_cerr, canonicalFocus);
      }
    } else {
      bool matched;
      mi=doInternalMetric(canonicalFocus,
                          canonicalFocus, // is this right for component_canon_focus???
                          metric_name,flat_name,enable,matched);
         // NULL on serious error; -1 if enable was false; -2 if illegal to instr with
         // given focus [many internal metrics work only for whole program]

      if (mi == (metricDefinitionNode*)-2) {
         metric_cerr << "createMetricInstance: internal metric " << metric_name << " isn't defined for focus: ";
         print_focus(metric_cerr, canonicalFocus);
         mi = NULL; // straighten up the return value
      }
      else if (mi == (metricDefinitionNode*)-1) {
         assert(!enable); // no error msg needed
         mi = NULL; // straighten up the return value
      }
      else if (mi == NULL) {
         // more serious error...do a printout
         metric_cerr << "createMetricInstance failed since doInternalMetric failed" << endl;
         metric_cerr << "metric name was " << metric_name << "; focus was ";
         print_focus(metric_cerr, canonicalFocus);
      }

      internal = true;
    }

    return mi;
}


// propagate this metric instance to process p.
// p is a process that started after the metric instance was created
// note: don't call this routine for a process started via fork or exec, just
// for processes started the "normal" way.
// "this" is an aggregate mi, not a component one.

void metricDefinitionNode::propagateToNewProcess(process *p) {
  unsigned comp_size = components.size();

  if (comp_size == 0)
    return; // if there are no components, shouldn't the mi be fried?

  for (unsigned u = 0; u < comp_size; u++) {
    if (components[u]->proc() == p) {
      // The metric is already enabled for this process. This case can 
      // happen when we are starting several processes at the same time.
      // (explain...?)
      return;
    }
  }

  bool internal = false;

  metricDefinitionNode *mi = NULL;
     // an aggregate (not component) mi, though we know that it'll contain just
     // one component.  It's that one component that we're really interested in.
  if (mdl_can_do(met_)) {
      // Make the unique ID for this metric/focus visible in MDL.
      string vname = "$globalId";
      mdl_env::add(vname, false, MDL_T_INT);
      mdl_env::set(this->getMId(), vname);

      vector<process *> vp(1,p);
      mi = mdl_do(focus_, met_, flat_name_, vp, false);
  } else {
    // internal and cost metrics don't need to be propagated (um, is this correct?)
    mi = NULL;
  }

  if (mi) { // successfully created new mi
    assert(mi->components.size() == 1);
    metricDefinitionNode *theNewComponent = mi->components[0];

    components += theNewComponent;
    theNewComponent->aggregators[0] = this;
    theNewComponent->samples[0] = aggSample.newComponent();
    if (!internal) {
      theNewComponent->insertInstrumentation();
      theNewComponent->checkAndInstallInstrumentation();
    }

    // update cost
    const float cost = mi->cost();
    if (cost > originalCost_) {
      currentPredictedCost += cost - originalCost_;
      originalCost_ = cost;
    }

    mi->components.resize(0); // protect the new component
    delete mi;
  }
}

metricDefinitionNode* metricDefinitionNode::handleExec() {
   // called by handleExec(), below.  See that routine for documentation.
   // "this" is a component mi.

   // If this component mi can be (re-)enabled in the new (post-exec) process, then do
   // so.  Else, remove the component mi from aggregators, etc.  Returns new component
   // mi if successful, NULL otherwise.

   assert(!aggregate_);

   // How can we tell if the mi can be inserted into the "new" (post-exec) process?
   // A component mi is basically a set of instReqNodes and dataReqNodes.  The latter
   // don't restrict what can be inserted (is this right?); the instReqNodes hold the
   // key -- we should look at the functions (instPoint's) where code (whose contents
   // are in AstNode's) would be inserted.  Now clearly, the instPoint's must be
   // checked -- if any one doesn't exist, then the instReqNode and hence the component
   // mi doesn't belong in the post-exec process.  But what about the AstNode's?
   // Should the code that gets inserted be subject to a similar test?  Probably, but
   // we currently don't do it.

   // BUT: Even if a process contains a function in both the pre-exec and post-exec
   // stages, we must assume that the function is IN A DIFFERENT LOCATION IN
   // THE ADDRESS SPACE.  Ick.  So the instPoint's can't be trusted and must
   // be recalculated from scratch.  In that regard, this routine is similar to
   // propagateToNewProcess(), which propagates aggregate mi's to a brand new
   // process (but which doesn't work for processes started via fork or exec).
   // The lesson learned is to (ick, ick, ick) call mdl_do() all over again.
   // This gets really confusing when you consider that a component mi can belong
   // to several aggregate mi's (e.g. if we represent cpu time for proc 100 then
   // we can belong to cpu/whole and cpu/proc-100); for which aggregate mi should
   // we run mdl_do?  Any will do, so we can pick arbitrarily (is this right?).

   // QUESTION: What about internal or cost metrics???  They have aggregate and
   //           component mi's just like normal metrics, right?  If that's so, then
   //           they must be propagated too!   NOT YET IMPLEMENTED!!!

   metricDefinitionNode *aggregateMI = this->aggregators[0];
   metricDefinitionNode *resultCompMI = NULL; // so far...

   const bool internal = !mdl_can_do(aggregateMI->met_);
   if (internal)
      return NULL; // NOT YET IMPLEMENTED

   // try to propagate the mi
   // note: the following code is mostly stolen from propagateToNewProcess(); blame
   //       it for any bugs :)

   // Make the unique ID for this metric/focus visible in MDL. (?)
   string vname = "$globalId";
   mdl_env::add(vname, false, MDL_T_INT);
   mdl_env::set(aggregateMI->getMId(), vname);

   vector<process*> vp(1, this->proc());
   metricDefinitionNode *tempAggMI = mdl_do(aggregateMI->focus_,
                                            aggregateMI->met_,
                                            aggregateMI->flat_name_,
                                            vp,
                                            true // --> fry existing component MI
                                            );
   if (tempAggMI == NULL)
      return NULL; // failure

   assert(tempAggMI->aggregate_);

   // okay, it looks like we successfully created a new aggregate mi.
   // Of course, we're just interested in the (single) component mi contained
   // within it; it'll replace "this".

   assert(tempAggMI->components.size() == 1);
   resultCompMI = tempAggMI->components[0];

   resultCompMI->aggregators.resize(0);
   resultCompMI->samples.resize(0);

   // For each aggregator, go back and find where "this" was a component mi.
   // When found, replace the ptr to "this" with "theNewComponent".
   unsigned num_aggregators = aggregators.size();
   assert(num_aggregators > 0);
   for (unsigned agglcv=0; agglcv < num_aggregators; agglcv++) {
      metricDefinitionNode *aggMI = aggregators[agglcv];

      bool found=false;
      for (unsigned complcv=0; complcv < aggMI->components.size(); complcv++) {
         if (aggMI->components[complcv] == this) {
            aggMI->components[complcv] = resultCompMI;

            resultCompMI->aggregators += aggMI;
            resultCompMI->samples     += aggMI->aggSample.newComponent();
            
            aggMI->aggSample.removeComponent(this->samples[agglcv]);
            
            found=true;
            break;
         }
      }
      assert(found);
   }

   // Now let's actually insert the instrumentation:
   if (!internal) {
      resultCompMI->insertInstrumentation();
      resultCompMI->checkAndInstallInstrumentation();
   }

   // And fry "tempAggMI", but make sure "resultCompMI" isn't fried when we do so
   tempAggMI->components.resize(0); // protect resultCompMI
   delete tempAggMI; // good riddance; you were an ugly hack to begin with

   return resultCompMI;
}

void metricDefinitionNode::handleExec(process *proc) {
   // a static member fn.
   // handling exec is tricky.  At the time this routine is called, the "new" process
   // has been bootstrapped and is ready for stuff to get inserted.  No mi's have yet
   // been propagated, and the data structures (allMIs, allMIComponents, etc.) are still
   // in their old, pre-exec state, so they show component mi's enabled for this
   // process, even though they're not (at least not yet).  This routines brings things
   // up-to-date.
   //
   // Algorithm: loop thru all component mi's for this process.  If it is possible to
   // propagate it to the "new" (post-exec) process, then do so.  If not, fry the
   // component mi.  An example where a component mi can no longer fit is an mi
   // specific to, say, function foo(), which (thanks to the exec syscall) no longer
   // exists in this process.  Note that the exec syscall changed the addr space enough
   // so even if a given routine foo() is present in both the pre-exec and post-exec
   // process, we must assume that it has MOVED TO A NEW LOCATION, thus making
   // the component mi's instReqNode's instPoint out-of-date.  Ick.

   vector<metricDefinitionNode*> miComponents = allMIComponents.values();
   for (unsigned lcv=0; lcv < miComponents.size(); lcv++) {
      metricDefinitionNode *componentMI = miComponents[lcv];
      if (componentMI->proc() != proc)
         continue;

      forkexec_cerr << "calling handleExec for component "
                    << componentMI->flat_name_ << endl;

      metricDefinitionNode *replaceWithComponentMI = componentMI->handleExec();

      if (replaceWithComponentMI == NULL) {
         forkexec_cerr << "handleExec for component " << componentMI->flat_name_
                       << " failed, so not propagating it" << endl;
         componentMI->removeThisInstance(); // propagation failed; fry component mi
      }
      else {
         forkexec_cerr << "handleExec for component " << componentMI->flat_name_
                       << " succeeded...it has been propagated" << endl;
         // new component mi has already been inserted in place of old component mi
         // in all of its aggregate's component lists.  So, not much left to do,
         // except to update allMIComponents.

         assert(replaceWithComponentMI->flat_name_ == componentMI->flat_name_);

         delete componentMI; // old component mi (dtor removes it from allMIComponents)
         assert(!allMIComponents.defines(replaceWithComponentMI->flat_name_));
         allMIComponents[replaceWithComponentMI->flat_name_] = replaceWithComponentMI;
      }
   }
}

// called when all components have been removed (because the processes have exited
// or exec'd) from "this".  "this" is an aggregate mi.
void metricDefinitionNode::endOfDataCollection() {
  assert(components.size() == 0);

  // flush aggregateSamples
  sampleInterval ret = aggSample.aggregateValues();

  while (ret.valid) {
    assert(ret.end > ret.start);
    assert(ret.start >= (firstRecordTime/MILLION));
    assert(ret.end >= (firstRecordTime/MILLION));
    batchSampleData(id_, ret.start, ret.end, ret.value,
                    aggSample.numComponents(),false);
    ret = aggSample.aggregateValues();
  }
  flush_batch_buffer();
  tp->endOfDataCollection(id_);
}

// remove a component from an aggregate.
// "this" is an aggregate mi; "comp" is a component mi.
void metricDefinitionNode::removeFromAggregate(metricDefinitionNode *comp) {
  unsigned size = components.size();
  for (unsigned u = 0; u < size; u++) {
    if (components[u] == comp) {
      delete components[u];
      components[u] = NULL;
      components[u] = components[size-1];
      components.resize(size-1);
      if (size == 1) {
        endOfDataCollection();
      }
      return;
    }
  }
  // should always find the right component 
  assert(0);
}

// remove this component mi from all aggregators it is a component of.
// if the aggregate mi no longer has any components then fry the mi aggregate mi.
// called by removeFromMetricInstances, below, when a process exits (or exec's)
void metricDefinitionNode::removeThisInstance() {
  assert(!aggregate_);

  // first, remove from allMIComponents (this is new --- is it right?)
  assert(allMIComponents.defines(flat_name_));
  allMIComponents.undef(flat_name_);

  assert(aggregators.size() == samples.size());
  unsigned aggr_size = aggregators.size();
  assert(aggr_size > 0);

  //for (unsigned u = 0; u < aggr_size; u++) {
  for (unsigned u = 0; u < aggregators.size() && u < samples.size(); u++) {
    aggregators[u]->aggSample.removeComponent(samples[u]);
    aggregators[u]->removeFromAggregate(this); 
  }
}


// Called when a process exits, to remove the component associated to proc 
// from all metric instances.  (If, after an exec, we never want to carry over
// mi's from the pre-exec, then this routine will work there, too.  But we try to
// carry over mi's whenever appropriate.)
// Remove the aggregate metric instances that don't have any components left
void removeFromMetricInstances(process *proc) {
    // Loop through all of the _component_ mi's; for each with component process
    // of "proc", remove the component mi from its aggregate mi.
    // Note: imho, there should be a *per-process* vector of mi-components.

    vector<metricDefinitionNode *> MIs = allMIComponents.values();
    for (unsigned j = 0; j < MIs.size(); j++) {
      if (MIs[j]->proc() == proc)
        MIs[j]->removeThisInstance();
    }
    costMetric::removeProcessFromAll(proc); // what about internal metrics?
}

/* *************************************************************************** */

// obligatory definition of static member vrble:
int metricDefinitionNode::counterId=0;

dataReqNode *metricDefinitionNode::addSampledIntCounter(int initialValue) {
   dataReqNode *result=NULL;

#ifdef SHM_SAMPLING
   // shared memory sampling of a reported intCounter
   result = new sampledShmIntCounterReqNode(initialValue,
                                            metricDefinitionNode::counterId);
      // implicit conversion to base class
#else
   // non-shared-memory sampling of a reported intCounter
   result = new sampledIntCounterReqNode(initialValue,
                                         metricDefinitionNode::counterId);
      // implicit conversion to base class
#endif

   assert(result);
   
   metricDefinitionNode::counterId++;

   internalMetricCounterId = metricDefinitionNode::counterId;

   dataRequests += result;
   return result;
}

dataReqNode *metricDefinitionNode::addUnSampledIntCounter(int initialValue) {
   // sampling of a non-reported intCounter (probably just a predicate)
   // NOTE: In the future, we should probably put un-sampled intcounters
   // into shared-memory when SHM_SAMPLING is defined.  After all, the shared
   // memory heap is faster.
   dataReqNode *result = new nonSampledIntCounterReqNode
                         (initialValue, metricDefinitionNode::counterId);
      // implicit conversion to base class
   assert(result);

   metricDefinitionNode::counterId++;

   internalMetricCounterId = metricDefinitionNode::counterId;

   dataRequests += result;
   return result;
};

dataReqNode *metricDefinitionNode::addWallTimer() {
   dataReqNode *result = NULL;

#ifdef SHM_SAMPLING
   result = new sampledShmWallTimerReqNode(metricDefinitionNode::counterId);
      // implicit conversion to base class
#else
   result = new sampledTimerReqNode(wallTime, metricDefinitionNode::counterId);
      // implicit conversion to base class
#endif

   assert(result);

   metricDefinitionNode::counterId++;

   internalMetricCounterId = metricDefinitionNode::counterId;

   dataRequests += result;
   return result;
}

dataReqNode *metricDefinitionNode::addProcessTimer() {
   dataReqNode *result = NULL;

#ifdef SHM_SAMPLING
   result = new sampledShmProcTimerReqNode(metricDefinitionNode::counterId);
      // implicit conversion to base class
#else
   result = new sampledTimerReqNode(processTime, metricDefinitionNode::counterId);
      // implicit conversion to base class
#endif

   assert(result);

   metricDefinitionNode::counterId++;

   internalMetricCounterId = metricDefinitionNode::counterId;

   dataRequests += result;
   return result;
};

/* *************************************************************************** */

// called when a process forks (by handleFork(), below). "this" is a (component)
// mi in the parent process. Duplicate it for the child, with appropriate
// changes (i.e. the pid of the component focus name differs), and return the newly
// created child mi.  "map" maps all instInstance's of the parent to those copied into
// the child.
// 
// Note how beautifully everything falls into place.  Consider the case of alarm
// sampling with cpu/whole program.  Then comes the fork.  The parent process has
// (0) a tTimer structure allocated in a specific location in the inferior heap,
// (1) instrumentation @ main to call startTimer on that ptr, (2) instrumentation in
// DYNINSTsampleValues() to call DYNINSTreportTimer on that ptr.
// The child process of fork will have ALL of these things in the exact same locations,
// which is correct.  We want the timer to be located in the same spot; we want
// DYNINSTreportTimer to be called on the same pointer; and main() hasn't moved.
//
// So there's not much to do here.  We create a new component mi (with same flat name
// as in the parent, except for a different pid), and call "forkProcess" for all
// dataReqNodes and instReqNodes, none of which have to do anything titanic.

metricDefinitionNode *metricDefinitionNode::forkProcess(process *child,
                        const dictionary_hash<instInstance*,instInstance*> &map) const {
    // The "focus_" member vrble stays the same, because it was always for the
    // metric as a whole, and not for some component.
    //
    // But two things must change, because they were component-specific (and the
    // component has changed processes):
    // (1) the flat name
    // (2) the component focus (not to be confused with plain focus_)
    //
    // For example, instead of
    // "/Code/foo.c/myfunc, /Process/100, ...", we should have
    // "/Code/foo.c/myfunc, /Process/101, ...", because the pid of the child
    // differs from that of the parent.

    // The resource structure of a given process is found in the "rid"
    // field of class process.
    const resource *parentResource = child->getParent()->rid;
    const string &parentPartName = parentResource->part_name();

    const resource *childResource = child->rid;
    const string &childPartName = childResource->part_name();

    vector< vector<string> > newComponentFocus = this->component_focus;
       // we'll change the process, but not the machine name.
    bool foundProcess = false;

    for (unsigned hier=0; hier < component_focus.size(); hier++) {
       if (component_focus[hier][0] == "Process") {
          foundProcess = true;
          assert(component_focus[hier].size() == 2);
             // since a component focus is by definition specific to some process

          assert(component_focus[hier][1] == parentPartName);

          // change the process:
          newComponentFocus[hier][1] = childPartName;
          break;
       }
    }
    assert(foundProcess);
    
    string newComponentFlatName = metricAndCanonFocus2FlatName(met_, newComponentFocus);

    metricDefinitionNode *mi =
        new metricDefinitionNode(child,
                         met_, // metric name doesn't change
                         focus_, // focus doesn't change (tho component focus will)
                         newComponentFocus, // this is a change
                         newComponentFlatName, // this is a change
                         aggOp // no change
                         );
    assert(mi);

    metricDefinitionNode::counterId++;

    forkexec_cerr << "metricDefinitionNode::forkProcess -- component flat name for parent is " << flat_name_ << "; for child is " << mi->flat_name_ << endl;

    internalMetricCounterId = metricDefinitionNode::counterId;

    assert(!allMIComponents.defines(newComponentFlatName));
    allMIComponents[newComponentFlatName] = mi;

    // Duplicate the dataReqNodes:
    for (unsigned u = 0; u < dataRequests.size(); u++) {
       // must add to midToMiMap[] before dup() to avoid some assert fails
       const int newCounterId = metricDefinitionNode::counterId++;
          // no relation to mi->getMId();
       forkexec_cerr << "forked dataReqNode going into midToMiMap with id " << newCounterId << endl;
       assert(!midToMiMap.defines(newCounterId));
       midToMiMap[newCounterId] = mi;
       
       dataReqNode *newNode = dataRequests[u]->dup(child, mi, newCounterId, map);
         // remember, dup() is a virtual fn, so the right dup() and hence the
         // right fork-ctor is called.
       assert(newNode);

       mi->dataRequests += newNode;
    }

    // Duplicate the instReqNodes:
    for (unsigned u = 0; u < instRequests.size(); u++) {
      mi->instRequests += instReqNode::forkProcess(instRequests[u], map);
    }

    mi->inserted_ = true;

    return mi;
}

bool metricDefinitionNode::unFork(dictionary_hash<instInstance*, instInstance*> &map,
                                  bool unForkInstRequests,
                                  bool unForkDataRequests) {
   // see below handleFork() for explanation of why this routine is needed.
   // "this" is a component mi for the parent process; we need to remove copied
   // instrumentation from the _child_ process.
   // Returns true iff the instrumentation was removed in the child (would be false
   // if it's not safe to remove the instrumentation in the child because it was
   // active.)

   // "map" maps instInstances from the parent process to instInstances in the child
   // process.

   // We loop thru the instReqNodes of the parent process, unforking each.
   // In addition, we need to unfork the dataReqNodes, because the alarm-sampled
   // ones instrument DYNINSTsampleValues.

   bool result = true;

   if (unForkInstRequests)
      for (unsigned lcv=0; lcv < instRequests.size(); lcv++)
         if (!instRequests[lcv].unFork(map))
            result = false; // failure

   if (unForkDataRequests)
      for (unsigned lcv=0; lcv < dataRequests.size(); lcv++)
         if (!dataRequests[lcv]->unFork(map))
            result = false; // failure

   return result;
}


// called by forkProcess of context.C, just after the fork-constructor was
// called for the child process.
void metricDefinitionNode::handleFork(const process *parent, process *child,
                              dictionary_hash<instInstance*,instInstance*> &map) {
   // "map" defines a mapping from all instInstance's of the parent process to
   // the copied one in the child process.  Some of the child process's ones may
   // get fried by this routine, as it detects that instrumentation has been copied
   // (by the fork syscall, which we have no control over) which doesn't belong in
   // the child process and therefore gets deleted manually.
  
   // Remember that a given component can be shared by multiple aggregator-mi's,
   // so be careful about duplicating a component twice.  Since we loop through
   // component mi's instead of aggregate mi's, it's no problem.  Note that it's
   // possible that only a subset of a component-mi's aggregators should get the newly
   // created child component mi.

   vector<metricDefinitionNode *> allComponents = allMIComponents.values();
   for (unsigned complcv=0; complcv < allComponents.size(); complcv++) {
      metricDefinitionNode *comp = allComponents[complcv];

      // duplicate the component (create a new one) if it belongs in the
      // child process.  It belongs if any of its aggregate mi's should be
      // propagated to the child process.  An aggregate mi should be propagated
      // if it wasn't refined to some process.

      bool shouldBePropagated = false; // so far
      bool shouldBeUnforkedIfNotPropagated = false; // so far
      assert(comp->aggregators.size() > 0);
      for (unsigned agglcv=0; agglcv < comp->aggregators.size(); agglcv++) {
         metricDefinitionNode *aggMI = comp->aggregators[agglcv];

         if (aggMI->focus_[resource::process].size() == 1) {
            // wasn't specific to any process
            shouldBeUnforkedIfNotPropagated = false; // we'll definitely be using it
            shouldBePropagated = true;
            break;
         }
         else if (comp->proc() == parent)
            // was specific to parent process, so fork() copied it into the child,
            // unless it was an internal or cost metric, in which case there was nothing
            // for fork to copy.
            if (!internalMetric::isInternalMetric(aggMI->getMetName()) &&
                !costMetric::isCostMetric(aggMI->getMetName()))
               shouldBeUnforkedIfNotPropagated = true;
         else
            // was specific to other process, so nothing is in the child for it yet
            ;
      }

      if (!shouldBePropagated && shouldBeUnforkedIfNotPropagated) {
         // this component mi isn't gonna be propagated to the child process, but
         // the fork syscall left some residue in the child.  Delete that residue now.
         assert(comp->proc() == parent);
         comp->unFork(map, true, true); // also modifies 'map' to remove items
      }

      if (!shouldBePropagated)
         continue;

      // Okay, it's time to propagate this component mi to the subset of its aggregate
      // mi's which weren't refined to a specific process.  If we've gotten to this
      // point, then there _is_ at least one such aggregate.
      assert(shouldBePropagated);
      metricDefinitionNode *newComp = comp->forkProcess(child, map);
         // copies instr (well, fork() does this for us), allocs ctr/timer space,
         // initializes.  Basically, copies dataReqNode's and instReqNode's.

      bool foundAgg = false;
      for (unsigned agglcv=0; agglcv < comp->aggregators.size(); agglcv++) {
         metricDefinitionNode *aggMI = comp->aggregators[agglcv];
         if (aggMI->focus_[resource::process].size() == 1) {
            // this aggregate mi wasn't specific to any process, so it gets the new
            // child component.
            aggMI->components += newComp;
            newComp->aggregators += aggMI;
            newComp->samples     += aggMI->aggSample.newComponent();
            foundAgg = true;
         }
      }
      assert(foundAgg);
   }
}

bool metricDefinitionNode::anythingToManuallyTrigger() const {
   if (aggregate_) {
      for (unsigned i=0; i < components.size(); i++)
         if (components[i]->anythingToManuallyTrigger())
            return true;
      return false;
   }
   else {
      for (unsigned i=0; i < instRequests.size(); i++)
         if (instRequests[i].anythingToManuallyTrigger())
            return true;
      return false;
   }

   assert(false);
}

void metricDefinitionNode::manuallyTrigger() {
   assert(anythingToManuallyTrigger());

   if (aggregate_) {
      for (unsigned i=0; i < components.size(); i++)
         components[i]->manuallyTrigger();
   }
   else {
      for (unsigned i=0; i < instRequests.size(); i++)
         if (instRequests[i].anythingToManuallyTrigger())
            if (!instRequests[i].triggerNow(proc())) {
               cerr << "manual trigger failed for an inst request" << endl;
            }
   }
}


// startCollecting is called by dynRPC::enableDataCollection (or enableDataCollection2)
// in dynrpc.C
// startCollecting is a friend of metricDefinitionNode; can it be
// made a member function of metricDefinitionNode instead?
// Especially since it clearly is an integral part of the class;
// in particular, it sets the crucial vrble "id_"
int startCollecting(string& metric_name, vector<u_int>& focus, int id,
                    vector<process *> &procsToCont)
{
    bool internal = false;

    // Make the unique ID for this metric/focus visible in MDL.
    string vname = "$globalId";
    mdl_env::add(vname, false, MDL_T_INT);
    mdl_env::set(id, vname);

    metricDefinitionNode *mi = createMetricInstance(metric_name, focus,
                                                    true, internal);
       // calls mdl_do()
    if (!mi) {
       cerr << "startCollecting failed because createMetricInstance failed" << endl;
       return(-1);
    }

    mi->id_ = id;

    assert(!allMIs.defines(mi->id_));
    allMIs[mi->id_] = mi;

    const float cost = mi->cost();
    mi->originalCost_ = cost;

    currentPredictedCost += cost;

#ifdef ndef
    // enable timing stuff: also code in insertInstrumentation()
    u_int start_size = test_heapsize;
    printf("ENABLE: %d %s %s\n",start_size,
        (mi->getMetName()).string_of(),
        (mi->getFullName()).string_of());
    static timer inTimer;
    inTimer.start();
#endif


    if (!internal) {

        // pause processes that are running and add them to procsToCont.
        // We don't rerun the processes after we insert instrumentation,
        // this will be done by our caller, after all instrumentation
        // has been inserted.
        for (unsigned u = 0; u < mi->components.size(); u++) {
          process *p = mi->components[u]->proc();
          if (p->status() == running && p->pause()) {
            procsToCont += p;
          }
        }


        mi->insertInstrumentation(); // calls pause and unpause (this could be a bug, since the next line should be allowed to execute before the unpause!!!)
        mi->checkAndInstallInstrumentation();

        // Now that the timers and counters have been allocated on the heap, and
        // the instrumentation added, we can manually execute instrumentation
        // we may have missed at $start.entry.  But has the process been paused
        // all this time?  Hopefully so; otherwise things can get screwy.

        if (mi->anythingToManuallyTrigger()) {
           process *theProc = mi->components[0]->proc();
           assert(theProc);

           bool alreadyRunning = (theProc->status_ == running);

           if (alreadyRunning)
              theProc->pause();

           mi->manuallyTrigger();

           if (alreadyRunning)
              theProc->continueProc(); // the continue will trigger our code
           else
              ; // the next time the process continues, we'll trigger our code
        }
    }

#ifdef ndef
    inTimer.stop();
    if(!start_size) start_size = test_heapsize;
    printf("It took %f:user %f:system %f:wall seconds heap_left: %d used %d\n"
                , inTimer.usecs(), inTimer.ssecs(), inTimer.wsecs(),
                test_heapsize,start_size-test_heapsize);
#endif

    metResPairsEnabled++;
    return(mi->id_);
}

float guessCost(string& metric_name, vector<u_int>& focus) {
   // called by dynrpc.C (getPredictedDataCost())
    bool internal;
    metricDefinitionNode *mi = createMetricInstance(metric_name, focus, false, internal);
    if (!mi) {
       //metric_cerr << "guessCost returning 0.0 since createMetricInstance failed" << endl;
       return(0.0);
    }

    float cost = mi->cost();
    // delete the metric instance, if it is not being used 
    if (!allMIs.defines(mi->getMId()))
      delete mi;

    return(cost);
}

bool metricDefinitionNode::insertInstrumentation()
{
    // returns true iff successful
    if (inserted_)
       return true;

    inserted_ = true;

    if (aggregate_) {
        unsigned c_size = components.size();
        for (unsigned u=0; u<c_size; u++)
          if (!components[u]->insertInstrumentation())
             return false; // shouldn't we try to undo what's already put in?
    } else {
      bool needToCont = proc_->status() == running;
      bool res = proc_->pause();
      if (!res)
        return false;

      // Loop thru "dataRequests", an array of (ptrs to) dataReqNode:
      // Here we allocate ctrs/timers in the inferior heap but don't
      // stick in any code, except (if appropriate) that we'll instrument the
      // application's alarm-handler when not shm sampling.
      unsigned size = dataRequests.size();
      for (unsigned u=0; u<size; u++) {
        // the following allocs an object in inferior heap and arranges for
        // it to be alarm sampled, if appropriate.
        if (!dataRequests[u]->insertInstrumentation(proc_, this))
           return false; // shouldn't we try to undo what's already put in?

        unsigned mid = dataRequests[u]->getSampleId();
        assert(!midToMiMap.defines(mid));
        midToMiMap[mid] = this;
      }

      // Loop thru "instRequests", an array of instReqNode:
      // (Here we insert code instrumentation, tramps, etc. via addInstFunc())
      for (unsigned u1=0; u1<instRequests.size(); u1++) {
          // NEW: the following may also manually trigger the instrumentation
          // via inferiorRPC.
          returnInstance *retInst=NULL;
          if (!instRequests[u1].insertInstrumentation(proc_, retInst))
             return false; // shouldn't we try to undo what's already put in?

          if (retInst)
            returnInsts += retInst;
      }

      if (needToCont)
         proc_->continueProc();
    }

    return(true);
}

bool metricDefinitionNode::checkAndInstallInstrumentation() {
    bool needToCont = false;

    if (installed_) return(true);

    installed_ = true;

    if (aggregate_) {
        unsigned c_size = components.size();
        for (unsigned u=0; u<c_size; u++)
            components[u]->checkAndInstallInstrumentation();
    } else {
        needToCont = proc_->status() == running;
        if (!proc_->pause()) {
            cerr << "checkAnd... pause failed" << endl; cerr.flush();
            return false;
        }

        vector<Address> pc = proc_->walkStack();

        // for(u_int i=0; i < pc.size(); i++){
        //     printf("frame %d: pc = 0x%x\n",i,pc[i]);
        // }

        unsigned rsize = returnInsts.size();
        u_int max_index = 0;  // first frame where it is safe to install instr
        bool delay_install = false; // true if some instr. needs to be delayed 
        vector<bool> delay_elm(rsize); // wch instr. to delay
        // for each inst point walk the stack to determine if it can be
        // inserted now (it can if it is not currently on the stack)
        // If some can not be inserted, then find the first safe point on
        // the stack where all can be inserted, and set a break point  
        for (unsigned u=0; u<rsize; u++) {
            u_int index = 0;
            bool installSafe = returnInsts[u] -> checkReturnInstance(pc,index);
            if ((!installSafe) && (index > max_index)) max_index = index;
            
            if (installSafe) {
                //cerr << "installSafe!" << endl;
                returnInsts[u] -> installReturnInstance(proc_);
                delay_elm[u] = false;
            } else {
                delay_install = true;
                delay_elm[u] = true;
            }
        }
        if(delay_install){
            // get rid of pathological cases...caused by threaded applications 
            // TODO: this should be fixed to do something smarter
            if((max_index > 0) && ((max_index+1) >= pc.size())){
               max_index--;
               //printf("max_index changed: %d\n",max_index);
            }
            if((max_index > 0) && (pc[max_index+1] == 0)){
               max_index--;
               //printf("max_index changed: %d\n",max_index);
            }
            Address pc2 = pc[max_index+1];
            for(u_int i=0; i < rsize; i++){
                if(delay_elm[i]){
                    returnInsts[i]->addToReturnWaitingList(pc2, proc_);
                }
            }
        }

        if (needToCont) proc_->continueProc();
    }
    return(true);
}

float metricDefinitionNode::cost() const
{
    float ret = 0.0;
    if (aggregate_) {
        unsigned c_size = components.size();
        for (unsigned u=0; u<c_size; u++) {
          float nc = components[u]->cost();
          if (nc > ret) ret = nc;
        }
    } else {
      for (unsigned u=0; u<instRequests.size(); u++)
        ret += instRequests[u].cost(proc_);
    }
    return(ret);
}

void metricDefinitionNode::disable()
{
    // check for internal metrics

    unsigned ai_size = internalMetric::allInternalMetrics.size();
    for (unsigned u=0; u<ai_size; u++) {
      internalMetric *theIMetric = internalMetric::allInternalMetrics[u];
      if (theIMetric->disableByMetricDefinitionNode(this)) {
        //logLine("disabled internal metric\n");
        return;
      }
    }

    // check for cost metrics
    for (unsigned i=0; i<costMetric::allCostMetrics.size(); i++){
      if (costMetric::allCostMetrics[i]->node == this) {
        costMetric::allCostMetrics[i]->disable();
        //logLine("disabled cost metric\n");
        return;
    }}

    if (!inserted_) return;

    inserted_ = false;
    if (aggregate_) {
        /* disable components of aggregate metrics */
        // unsigned c_size = components.size();
        //for (unsigned u=0; u<c_size; u++) {
        for (unsigned u=0; u<components.size(); u++) {
          //components[u]->disable();
          metricDefinitionNode *m = components[u];
          unsigned aggr_size = m->aggregators.size();
          assert(aggr_size == m->samples.size());
          for (unsigned u1=0; u1 < aggr_size; u1++) {
            if (m->aggregators[u1] == this) {
              m->aggregators[u1] = m->aggregators[aggr_size-1];
              m->aggregators.resize(aggr_size-1);
              m->samples[u1] = m->samples[aggr_size-1];
              m->samples.resize(aggr_size-1);
              break;
            }
          }
          assert(m->aggregators.size() == aggr_size-1);
          // disable component only if it is not being shared
          if (aggr_size == 1) {
            m->disable();
          }
        }

    } else {
      vector<unsigVecType> pointsToCheck;
      for (unsigned u1=0; u1<instRequests.size(); u1++) {
        unsigVecType pointsForThisRequest = 
            getAllTrampsAtPoint(instRequests[u1].getInstance());
        pointsToCheck += pointsForThisRequest;

        instRequests[u1].disable(pointsForThisRequest); // calls deleteInst()
      }

      for (unsigned u=0; u<dataRequests.size(); u++) {
        unsigned mid = dataRequests[u]->getSampleId();
        dataRequests[u]->disable(proc_, pointsToCheck); // deinstrument
        assert(midToMiMap.defines(mid));
        midToMiMap.undef(mid);
      }
    }
}

void metricDefinitionNode::removeComponent(metricDefinitionNode *comp) {
    assert(!comp->aggregate_);
    unsigned aggr_size = comp->aggregators.size();
    unsigned found = aggr_size;

    if (aggr_size == 0) {
      delete comp;
      return;
    }

    // component has more than one aggregator. Remove this from list of aggregators
    for (unsigned u = 0; u < aggr_size; u++) {
      if (comp->aggregators[u] == this) {
        found = u;
        break;
      }
    }
    if (found == aggr_size)
     return;
    assert(found < aggr_size);
    assert(aggr_size == comp->samples.size());
    comp->aggregators[found] = comp->aggregators[aggr_size-1];
    comp->aggregators.resize(aggr_size-1);
    comp->samples[found] = comp->samples[aggr_size-1];
    comp->samples.resize(aggr_size-1);

    if (aggr_size == 1) {
      delete comp;
      return;
    }

}

metricDefinitionNode::~metricDefinitionNode()
{
    if (aggregate_) {
        /* delete components of aggregate metrics */
        unsigned c_size = components.size();
        for (unsigned u=0; u<c_size; u++)
          removeComponent(components[u]);
          //delete components[u];
        components.resize(0);
    } else {
      allMIComponents.undef(flat_name_);
      unsigned size = dataRequests.size();
      for (unsigned u=0; u<size; u++)
        delete dataRequests[u];
    }
}


// NOTE: This stuff (flush_batch_buffer() and batchSampleData()) belongs
//       in perfStream.C; this is an inappropriate file.

//////////////////////////////////////////////////////////////////////////////
// Buffer the samples before we actually send it                            //
//      Send it when the buffers are full                                   //
//      or, send it when the last sample in the interval has arrived.       //
//////////////////////////////////////////////////////////////////////////////

const unsigned SAMPLE_BUFFER_SIZE = (1*1024)/sizeof(T_dyninstRPC::batch_buffer_entry);
bool BURST_HAS_COMPLETED = false;
   // set to true after a burst (after a processTraceStream(), or sampleNodes for
   // the CM5), which will force the buffer to be flushed before it fills up
   // (if not, we'd have bad response time)

vector<T_dyninstRPC::batch_buffer_entry> theBatchBuffer (SAMPLE_BUFFER_SIZE);
unsigned int batch_buffer_next=0;

// The following routines (flush_batch_buffer() and batchSampleData() are
// in an inappropriate src file...move somewhere more appropriate)
void flush_batch_buffer() {
   // don't need to flush if the batch had no data (this does happen; see
   // perfStream.C)
   if (batch_buffer_next == 0)
      return;

   // alloc buffer of the exact size to make communication
   // more efficient.  Why don't we send theBatchBuffer with a count?
   // This would work but would always (in the igen call) copy the entire
   // vector.  This solution has the downside of calling new but is not too bad
   // and is clean.
   vector<T_dyninstRPC::batch_buffer_entry> copyBatchBuffer(batch_buffer_next);
   assert(copyBatchBuffer.size() <= theBatchBuffer.size());
   for (unsigned i=0; i< batch_buffer_next; i++) {
      copyBatchBuffer[i] = theBatchBuffer[i];
   }

#ifdef FREEDEBUG
timeStamp t1,t2;
t1=getCurrentTime(false);
#endif

   // Now let's do the actual igen call!
   tp->batchSampleDataCallbackFunc(0, copyBatchBuffer);

#ifdef FREEDEBUG
t2=getCurrentTime(false);
if ((float)(t2-t1) > 15.0) {
sprintf(errorLine,"++--++ TEST ++--++ batchSampleDataCallbackFunc took %5.2f secs, size=%d, Kbytes=%5.2f\n",(float)(t2-t1),sizeof(T_dyninstRPC::batch_buffer_entry),(float)(sizeof(T_dyninstRPC::batch_buffer_entry)*copyBatchBuffer.size()/1024.0));
logLine(errorLine);
}
#endif

   BURST_HAS_COMPLETED = false;
   batch_buffer_next = 0;
}

void batchSampleData(int mid, double startTimeStamp,
                     double endTimeStamp, double value, unsigned val_weight,
                     bool internal_metric) 
{
   // This routine is called where we used to call tp->sampleDataCallbackFunc.
   // We buffer things up and eventually call tp->batchSampleDataCallbackFunc

#ifdef notdef
   char myLogBuffer[120] ;
   sprintf(myLogBuffer, "mid %d, value %g\n", mid, value) ;
   logLine(myLogBuffer) ;
#endif

   // Flush the buffer if (1) it is full, or (2) for good response time, after
   // a burst of data:
   if (batch_buffer_next >= SAMPLE_BUFFER_SIZE || BURST_HAS_COMPLETED)
      flush_batch_buffer();

   // Now let's batch this entry.
   T_dyninstRPC::batch_buffer_entry &theEntry = theBatchBuffer[batch_buffer_next];
   theEntry.mid = mid;
   theEntry.startTimeStamp = startTimeStamp;
   theEntry.endTimeStamp = endTimeStamp;
   theEntry.value = value;
   theEntry.weight = val_weight;
   theEntry.internal_met = internal_metric;
   batch_buffer_next++;
}

void metricDefinitionNode::forwardSimpleValue(timeStamp start, timeStamp end,
                                       sampleValue value, unsigned weight,
                                       bool internal_met)
{
  // TODO mdc
    assert(start + 0.000001 >= (firstRecordTime/MILLION));
    assert(end >= (firstRecordTime/MILLION));
    assert(end > start);

    batchSampleData(id_, start, end, value, weight, internal_met);
}

void metricDefinitionNode::updateValue(time64 wallTime, 
                                       sampleValue value)
{
    timeStamp sampleTime = wallTime / 1000000.0;
       // note: we can probably do integer division by million quicker

    assert(value >= -0.01);

    // TODO -- is this ok?
    // TODO -- do sampledFuncs work ?
    if (style_ == EventCounter) { 

      // only use delta from last sample.
      if (value < cumulativeValue) {
        if ((value/cumulativeValue) < 0.99999) {
          assert((value + 0.0001)  >= cumulativeValue);
        } else {
          // floating point rounding error ignore
          cumulativeValue = value;
        }
      }

      //        if (value + 0.0001 < cumulativeValue)
      //           printf ("WARNING:  sample went backwards!!!!!\n");
      value -= cumulativeValue;
      cumulativeValue += value;
    } 

    //
    // If style==EventCounter then value is changed. Otherwise, it keeps the
    // the current "value" (e.g. SampledFunction case). That's why it is not
    // necessary to have an special case for SampledFunction.
    //

    assert(samples.size() == aggregators.size());
    for (unsigned u = 0; u < samples.size(); u++) {
      if (samples[u]->firstValueReceived())
        samples[u]->newValue(sampleTime, value);
      else {
        samples[u]->startTime(sampleTime);
      }
      aggregators[u]->updateAggregateComponent();
    }
}

void metricDefinitionNode::updateAggregateComponent()
{
    // currently called (only) by the above routine
    sampleInterval ret = aggSample.aggregateValues();
    if (ret.valid) {
        assert(ret.end > ret.start);
        assert(ret.start + 0.000001 >= (firstRecordTime/MILLION));
        assert(ret.end >= (firstRecordTime/MILLION));
        batchSampleData(id_, ret.start, ret.end, ret.value,
                        aggSample.numComponents(),false);
    }
//    else {
//        metric_cerr << "sorry, ret.valid false so not batching sample data" << endl;
//    }
}

//
// Costs are now reported to paradyn like other metrics (ie. we are not
// calling reportInternalMetrics to deliver cost values, instead we wait
// until we have received a new interval of cost data from each process)
// note: this only works for the CM5 because all cost metrics are sumed
// at the daemons and at paradyn, otherwise the CM5 needs its own version
// of this routine that uses the same aggregate method as the one for paradyn 
//
#ifndef SHM_SAMPLING
void processCost(process *proc, traceHeader *h, costUpdate *s)
{
    // we can probably do integer division by million quicker.
    timeStamp newSampleTime = (h->wall / 1000000.0);
    timeStamp newProcessTime = (h->process / 1000000.0);

    timeStamp lastProcessTime = 
                        totalPredictedCost->getLastSampleProcessTime(proc); 

    // find the portion of uninstrumented time for this interval
    double unInstTime = ((newProcessTime - lastProcessTime) 
                         / (1+currentPredictedCost));
    // update predicted cost
    // note: currentPredictedCost is the same for all processes 
    //       this should be changed to be computed on a per process basis
    sampleValue newPredCost = totalPredictedCost->getCumulativeValue(proc);
    newPredCost += (float)(currentPredictedCost*unInstTime); 
    totalPredictedCost->updateValue(proc,newPredCost,
                                    newSampleTime,newProcessTime);
    // update observed cost 
    observed_cost->updateValue(proc,s->obsCostIdeal,
                               newSampleTime,newProcessTime);

    // update smooth observed cost
    smooth_obs_cost->updateSmoothValue(proc,s->obsCostIdeal,
                                 newSampleTime,newProcessTime);
}
#endif

#ifndef SHM_SAMPLING
void processSample(int /* pid */, traceHeader *h, traceSample *s)
{
    // called from processTraceStream (perfStream.C) when a TR_SAMPLE record
    // has arrived from the appl.

    unsigned mid = s->id.id; // low-level counterId (see primitives.C)

    static long long firstWall = 0;

    static bool firstTime = true;

    if (firstTime) {
       firstWall = h->wall;
    }

    metricDefinitionNode *mi; // filled in by find() if found
    if (!midToMiMap.find(mid, mi)) { // low-level counterId to metricDefinitionNode
       metric_cerr << "TR_SAMPLE id " << s->id.id << " not for valid mi...discarding" << endl;
       return;
    }

//    metric_cerr << "FROM pid " << pid << " got value " << s->value << " for id " << s->id.id << endl;

    //    sprintf(errorLine, "sample id %d at time %8.6f = %f\n", s->id.id, 
    //  ((double) *(int*) &h->wall) + (*(((int*) &h->wall)+1))/1000000.0, s->value);
    //    logLine(errorLine);
    mi->updateValue(h->wall, s->value);
    samplesDelivered++;
}
#endif

/*
 * functions to operate on inst request graph.
 *
 */
instReqNode::instReqNode(instPoint *iPoint,
                         AstNode *iAst,
                         callWhen  iWhen,
                         callOrder o, bool iManuallyTrigger) {
    point = iPoint;
    when = iWhen;
    order = o;
    instance = NULL; // set when insertInstrumentation() calls addInstFunc()
    ast = assignAst(iAst);
    manuallyTrigger = iManuallyTrigger;
    assert(point);
}

instReqNode instReqNode::forkProcess(const instReqNode &parentNode,
                             const dictionary_hash<instInstance*,instInstance*> &map) {
    instReqNode ret = instReqNode(parentNode.point, parentNode.ast, parentNode.when,
                                  parentNode.order,
                                  false // don't manually trigger
                                  );

    if (!map.find(parentNode.instance, ret.instance)) // writes to ret.instance
       assert(false);

    return ret;
}

bool instReqNode::unFork(dictionary_hash<instInstance*,instInstance*> &map) const {
   // The fork syscall duplicates all trampolines from the parent into the child. For
   // those mi's which we don't want to propagate to the child, this creates a
   // problem.  We need to remove instrumentation code from the child.  This routine
   // does that.
   //
   // "this" represents an instReqNode in the PARENT process.
   // "map" maps all instInstance*'s of the parent process to instInstance*'s in the
   // child process.  We modify "map" by setting a value to NULL.

   instInstance *parentInstance = getInstance();
   
   instInstance *childInstance;
   if (!map.find(parentInstance, childInstance)) // writes to childInstance
      assert(false);

   vector<unsigned> pointsToCheck; // is it right leaving this empty on a fork()???
   deleteInst(childInstance, pointsToCheck);

   map[parentInstance] = NULL; // since we've deleted...

   return true; // success
}

bool instReqNode::insertInstrumentation(process *theProc,
                                        returnInstance *&retInstance) 
{
    // NEW: We may manually trigger the instrumentation, via a call to postRPCtoDo()

    // addInstFunc() is one of the key routines in all paradynd.
    // It installs a base tramp at the point (if needed), generates code
    // for the tramp, calls inferiorMalloc() in the text heap to get space for it,
    // and actually inserts the instrumentation.
    instance = addInstFunc(theProc, point, ast, when, order,
                           false, // false --> don't exclude cost
                           retInstance);

    return (instance != NULL);
}

void instReqNode::disable(const vector<unsigned> &pointsToCheck)
{
    deleteInst(instance, pointsToCheck);
    instance = NULL;
}

instReqNode::~instReqNode()
{
    instance = NULL;
    removeAst(ast);
}

float instReqNode::cost(process *theProc) const
{
    float value;
    float unitCost;
    float frequency;
    int unitCostInCycles;

    unitCostInCycles = ast->cost() + getPointCost(theProc, point) +
                       getInsnCost(trampPreamble) + getInsnCost(trampTrailer);
    // printf("unit cost = %d cycles\n", unitCostInCycles);
    unitCost = unitCostInCycles/ cyclesPerSecond;
    frequency = getPointFrequency(point);
    value = unitCost * frequency;
    return(value);
}

bool instReqNode::triggerNow(process *theProc) {
   assert(manuallyTrigger);

   theProc->postRPCtoDo(ast, false, // don't skip cost
                        NULL, // no callback fn needed
                        NULL
                        );
      // the rpc will be launched with a call to launchRPCifAppropriate()
      // in the main loop (perfStream.C)

   return true;
}

/* ************************************************************************* */

#ifndef SHM_SAMPLING
sampledIntCounterReqNode::sampledIntCounterReqNode(int iValue, int iCounterId) :
                                                  dataReqNode() {
   theSampleId = iCounterId;
   initialValue = iValue;

   // The following fields are NULL until insertInstrumentation()
   counterPtr = NULL;
   sampler = NULL;

#if defined(MT_THREAD)
   position_=0;
#endif
}

sampledIntCounterReqNode::sampledIntCounterReqNode(const sampledIntCounterReqNode &src,
                                                   process *childProc,
                                                   metricDefinitionNode *,
                                                   int iCounterId,
                                                   const dictionary_hash<instInstance*,instInstance*> &map) {
   // a dup() routine (call after a fork())
   counterPtr = src.counterPtr; // assumes addr spaces have been dup()d.

   if (!map.find(src.sampler, this->sampler)) // writes to this->sampler
      assert(false);

   theSampleId = iCounterId;
 
   intCounter temp;
   temp.id.id = this->theSampleId;
   temp.value = initialValue;
   writeToInferiorHeap(childProc, temp);

#if defined(MT_THREAD)
   position_=0;
#endif
}

dataReqNode *
sampledIntCounterReqNode::dup(process *childProc,
                              metricDefinitionNode *mi,
                              int iCounterId,
                              const dictionary_hash<instInstance*,instInstance*> &map
                              ) const {
   // duplicate 'this' (allocate w/ new) and return.  Call after a fork().

   sampledIntCounterReqNode *tmp;
   tmp = new sampledIntCounterReqNode(*this, childProc, mi, iCounterId, map);
      // fork ctor

#if defined(MT_THREAD)
   // initialize position for new counter id with the same value as the
   // position for the "parent"
   assert(childProc);
   process *parent = childProc->getParent();
   assert(parent);
   // NOTE: this needs to be done for every thread, not just for threads[0]
   Thread *thr = parent->threads[0];
   childProc->threads[0]->CTvector->dup(tmp->theSampleId, iCounterId, thr, tmp->position_);
#endif 
  
   return tmp;
}

bool sampledIntCounterReqNode::insertInstrumentation(process *theProc,
                                                     metricDefinitionNode *) {
   // Remember counterPtr and sampler are NULL until this routine
   // gets called.
   counterPtr = (intCounter*)inferiorMalloc(theProc, sizeof(intCounter), dataHeap);
   if (counterPtr == NULL)
      return false; // failure!

   // initialize the intCounter in the inferior heap
   intCounter temp;
   temp.id.id = this->theSampleId;
   temp.value = this->initialValue;

   writeToInferiorHeap(theProc, temp);

   pdFunction *sampleFunction = theProc->findOneFunction("DYNINSTsampleValues");
   assert(sampleFunction);

   AstNode *ast, *tmp;
   tmp = new AstNode(AstNode::Constant, counterPtr);
   ast = new AstNode("DYNINSTreportCounter", tmp);
   removeAst(tmp);

   instPoint *func_entry = sampleFunction->funcEntry(theProc);
   sampler = addInstFunc(theProc, func_entry,
                         ast, callPreInsn, orderLastAtPoint, false);
   removeAst(ast);

#if defined(MT_THREAD)
   updateCounterTimerVectorMT(theProc,this->theSampleId,this->position_,getInferiorPtr());
#endif

   return true; // success
}

void sampledIntCounterReqNode::disable(process *theProc,
                                       const vector<unsigVecType> &pointsToCheck) {
   // We used to remove the sample id from midToMiMap here but now the caller is
   // responsible for that.

   // Remove instrumentation added to DYNINSTsampleValues(), if necessary:
   if (sampler != NULL)
      ::deleteInst(sampler, getAllTrampsAtPoint(sampler));

   // Deallocate space for intCounter in the inferior heap:
   assert(counterPtr != NULL);
   inferiorFree(theProc, (unsigned)counterPtr, dataHeap, pointsToCheck);

#if defined(MT_THREAD)
   Thread *thr = theProc->threads[0];
   thr->CTvector->remove(this->theSampleId, this->position_);
#endif
}

void sampledIntCounterReqNode::writeToInferiorHeap(process *theProc,
                                                   const intCounter &dataSrc) const {
   // using the contents of "dataSrc", write to the inferior heap at loc
   // "counterPtr" via proc->writeDataSpace()
   assert(counterPtr);
   theProc->writeDataSpace(counterPtr, sizeof(intCounter), &dataSrc);
}

bool sampledIntCounterReqNode::
unFork(dictionary_hash<instInstance*,instInstance*> &map) {
   instInstance *parentSamplerInstance = this->sampler;

   instInstance *childSamplerInstance;
   if (!map.find(parentSamplerInstance, childSamplerInstance))
      assert(false);

   vector<unsigned> pointsToCheck; // empty on purpose
   deleteInst(childSamplerInstance, pointsToCheck);

   map[parentSamplerInstance] = NULL;

   return true;
}
                                      
#endif

/* ************************************************************************* */

#ifdef SHM_SAMPLING

sampledShmIntCounterReqNode::sampledShmIntCounterReqNode(int iValue, int iCounterId) :
                                                  dataReqNode() {
   theSampleId = iCounterId;
   initialValue = iValue;

   // The following fields are NULL until insertInstrumentation()
   allocatedIndex = UINT_MAX;
   inferiorCounterPtr = NULL;

#if defined(MT_THREAD)
   position_=0;
#endif
}

sampledShmIntCounterReqNode::
sampledShmIntCounterReqNode(const sampledShmIntCounterReqNode &src,
                            process *childProc, metricDefinitionNode *mi,
                            int iCounterId) {
   // a dup() routine (call after a fork())
   // Assumes that "childProc" has been copied already (e.g., the shm seg was copied).

   // Note that the index w/in the inferior heap remains the same, so setting the
   // new inferiorCounterPtr isn't too hard.  Actually, it's trivial, since other code
   // ensures that the new shm segment is placed in exactly the same virtual mem location
   // as the previous one.
   //
   // Note that the fastInferiorHeap class's fork ctor will have already copied the
   // actual data; we need to fill in new meta-data (new houseKeeping entries).

   this->allocatedIndex = src.allocatedIndex;

   this->theSampleId = iCounterId;  // this is different from the parent's value
   this->initialValue = src.initialValue;

   fastInferiorHeap<intCounterHK, intCounter> &theHeap =
                    childProc->getInferiorIntCounters();

   // since the new shm seg is placed in exactly the same memory location as the old
   // one, nothing here should change.
   intCounter *oldInferiorCounterPtr = src.inferiorCounterPtr;
   inferiorCounterPtr = theHeap.index2InferiorAddr(allocatedIndex);
   assert(inferiorCounterPtr == oldInferiorCounterPtr);

   // write to the raw item in the inferior heap:
   intCounter *localCounterPtr = theHeap.index2LocalAddr(allocatedIndex);
   const intCounter *localSrcCounterPtr = childProc->getParent()->getInferiorIntCounters().index2LocalAddr(allocatedIndex);
   localCounterPtr->value = initialValue;
   localCounterPtr->id.id = theSampleId;
   localCounterPtr->theSpinner = localSrcCounterPtr->theSpinner;
      // in case we're in the middle of an operation

   // write HK for this intCounter:
   // Note: we don't assert anything about mi->getMId(), because that id has no
   // relation to the ids we work with (theSampleId).  In fact, we (the sampling code)
   // just don't ever care what mi->getMId() is.
   assert(theSampleId >= 0);
   assert(midToMiMap.defines(theSampleId));
   assert(midToMiMap[theSampleId] == mi);
   intCounterHK iHKValue(theSampleId, mi);
      // the mi differs from the mi of the parent; theSampleId differs too.
   theHeap.initializeHKAfterFork(allocatedIndex, iHKValue);

#if defined(MT_THREAD)
   position_=0;
#endif
}

dataReqNode *
sampledShmIntCounterReqNode::dup(process *childProc,
                                 metricDefinitionNode *mi,
                                 int iCounterId,
                                 const dictionary_hash<instInstance*,instInstance*> &
                                 ) const {
   // duplicate 'this' (allocate w/ new) and return.  Call after a fork().

   sampledShmIntCounterReqNode *tmp;
   tmp = new sampledShmIntCounterReqNode(*this, childProc, mi, iCounterId);
      // fork ctor

#if defined(MT_THREAD)
   // initialize position for new counter id with the same value as the
   // position for the "parent"
   assert(childProc);
   const process *parent = childProc->getParent();
   assert(parent);
   Thread *thr = parent->threads[0];
   childProc->threads[0]->CTvector->dup(tmp->theSampleId, iCounterId, thr, tmp->position_);
#endif

   return tmp;
}

bool sampledShmIntCounterReqNode::insertInstrumentation(process *theProc,
                                                        metricDefinitionNode *iMi) {
   // Remember counterPtr is NULL until this routine gets called.
   // WARNING: there will be an assert failure if the applic hasn't yet attached to the
   //          shm segment!!!

   // initialize the intCounter in the inferior heap
   intCounter iValue;
   iValue.id.id = this->theSampleId;
   iValue.value = this->initialValue; // what about initializing 'theSpinner'???

   intCounterHK iHKValue(this->theSampleId, iMi);

   fastInferiorHeap<intCounterHK, intCounter> &theShmHeap =
          theProc->getInferiorIntCounters();

   if (!theShmHeap.alloc(iValue, iHKValue, this->allocatedIndex))
      return false; // failure

   inferiorCounterPtr = theShmHeap.getBaseAddrInApplic() + allocatedIndex;
      // ptr arith.  Now we know where in the inferior heap this counter is
      // attached to, so getInferiorPtr() can work ok.

   assert(inferiorCounterPtr == theShmHeap.index2InferiorAddr(allocatedIndex));
      // just a check for fun

#if defined(MT_THREAD)
   updateCounterTimerVectorMT(theProc,this->theSampleId,this->position_,getInferiorPtr());
#endif

   return true; // success
}

void sampledShmIntCounterReqNode::disable(process *theProc,
                                          const vector<unsigVecType> &pointsToCheck) {
   // We used to remove the sample id from midToMiMap here but now the caller is
   // responsible for that.

   fastInferiorHeap<intCounterHK, intCounter> &theShmHeap =
          theProc->getInferiorIntCounters();

   // Remove from inferior heap; make sure we won't be sampled any more:
   vector<unsigned> trampsMaybeUsing;
   for (unsigned pointlcv=0; pointlcv < pointsToCheck.size(); pointlcv++)
      for (unsigned tramplcv=0; tramplcv < pointsToCheck[pointlcv].size(); tramplcv++)
         trampsMaybeUsing += pointsToCheck[pointlcv][tramplcv];

   theShmHeap.makePendingFree(allocatedIndex, trampsMaybeUsing);

#if defined(MT_THREAD)
    Thread *thr = theProc->threads[0];
    thr->CTvector->remove(this->theSampleId, this->position_);
#endif
}

#endif

/* ************************************************************************* */

nonSampledIntCounterReqNode::nonSampledIntCounterReqNode(int iValue, int iCounterId) :
                                                  dataReqNode() {
   theSampleId = iCounterId;
   initialValue = iValue;

   // The following fields are NULL until insertInstrumentation()
   counterPtr = NULL;

#if defined(MT_THREAD)
   position_=0;
#endif
}

nonSampledIntCounterReqNode::
nonSampledIntCounterReqNode(const nonSampledIntCounterReqNode &src,
                            process *childProc, metricDefinitionNode *,
                            int iCounterId) {
   // a dup() routine (call after a fork())
   counterPtr = src.counterPtr; // assumes addr spaces have been dup()d.
   initialValue = src.initialValue;
   theSampleId = iCounterId;
 
   intCounter temp;
   temp.id.id = this->theSampleId;
   temp.value = this->initialValue;
   writeToInferiorHeap(childProc, temp);

#if defined(MT_THREAD)
   position_=0;
#endif
}

dataReqNode *
nonSampledIntCounterReqNode::dup(process *childProc,
                                 metricDefinitionNode *mi,
                                 int iCounterId,
                                 const dictionary_hash<instInstance*,instInstance*> &
                                 ) const {
   // duplicate 'this' (allocate w/ new) and return.  Call after a fork().

   nonSampledIntCounterReqNode *tmp;
   tmp = new nonSampledIntCounterReqNode(*this, childProc, mi, iCounterId);
      // fork ctor

#if defined(MT_THREAD)
   // initialize position for new counter id with the same value as the
   // position for the "parent"
   assert(childProc);
   const process *parent = childProc->getParent();
   assert(parent);
   Thread *thr = parent->threads[0];
   childProc->threads[0]->CTvector->dup(tmp->theSampleId, iCounterId, thr, tmp->position_);
#endif

   return tmp;
}

bool nonSampledIntCounterReqNode::insertInstrumentation(process *theProc,
                                                        metricDefinitionNode *) {
   // Remember counterPtr is NULL until this routine gets called.
   counterPtr = (intCounter*)inferiorMalloc(theProc, sizeof(intCounter), dataHeap);
   if (counterPtr == NULL)
      return false; // failure!

   // initialize the intCounter in the inferior heap
   intCounter temp;
   temp.id.id = this->theSampleId;
   temp.value = this->initialValue;

   writeToInferiorHeap(theProc, temp);

#if defined(MT_THREAD)
   updateCounterTimerVectorMT(theProc,this->theSampleId,this->position_,getInferiorPtr());
#endif

   return true; // success
}

void nonSampledIntCounterReqNode::disable(process *theProc,
                                          const vector<unsigVecType> &pointsToCheck) {
   // We used to remove the sample id from midToMiMap here but now the caller is
   // responsible for that.

   // Deallocate space for intCounter in the inferior heap:
   assert(counterPtr != NULL);
   inferiorFree(theProc, (unsigned)counterPtr, dataHeap, pointsToCheck);

#if defined(MT_THREAD)
    Thread *thr = theProc->threads[0];
    thr->CTvector->remove(this->theSampleId, this->position_);
#endif
}

void nonSampledIntCounterReqNode::writeToInferiorHeap(process *theProc,
                                                      const intCounter &dataSrc) const {
   // using the contents of "dataSrc", write to the inferior heap at loc
   // "counterPtr" via proc->writeDataSpace()
   assert(counterPtr);
   theProc->writeDataSpace(counterPtr, sizeof(intCounter), &dataSrc);
}

/* ****************************************************************** */

#ifndef SHM_SAMPLING
sampledTimerReqNode::sampledTimerReqNode(timerType iType, int iCounterId) :
                                                 dataReqNode() {
   theSampleId = iCounterId;
   theTimerType = iType;

   // The following fields are NULL until insertInstrumentatoin():
   timerPtr = NULL;
   sampler  = NULL;

#if defined(MT_THREAD)
   position_=0;
#endif
}

sampledTimerReqNode::sampledTimerReqNode(const sampledTimerReqNode &src,
                                         process *childProc,
                                         metricDefinitionNode *,
                                         int iCounterId,
                                         const dictionary_hash<instInstance*,instInstance*> &map) {
   // a dup()-like routine; call after a fork()
   timerPtr = src.timerPtr; // assumes addr spaces have been dup()'d

   if (!map.find(src.sampler, this->sampler)) // writes to this->sampler
      assert(false);

   assert(sampler); // makes sense; timers are always sampled, whereas intCounters
                    // might be just non-sampled predicates.
   
   theSampleId = iCounterId;
   theTimerType = src.theTimerType;

   tTimer temp;
   P_memset(&temp, '\0', sizeof(tTimer)); /* is this needed? */
   temp.id.id = this->theSampleId;
   temp.type = this->theTimerType;
   temp.normalize = 1000000;
   writeToInferiorHeap(childProc, temp);

#if defined(MT_THREAD)
   position_=0;
#endif

   // WARNING: shouldn't we be resetting the raw value to count=0, start=0,
   //          total = src.initialValue ???  On the other hand, it's not that
   //          simple -- if the timer is active in the parent, then it'll be active
   //          in the child.  So how about setting count to src.count, start=now,
   //          total=0 ???
}

dataReqNode *
sampledTimerReqNode::dup(process *childProc, metricDefinitionNode *mi,
                         int iCounterId,
                         const dictionary_hash<instInstance*,instInstance*> &map
                         ) const {
   // duplicate 'this' (allocate w/ new) and return.  Call after a fork().

   sampledTimerReqNode *result = new sampledTimerReqNode(*this, childProc, mi, iCounterId, map);
      // fork ctor
   if (!result)
      return NULL; // on failure, return w/o incrementing counterId

#if defined(MT_THREAD)
   // initialize position for new counter id with the same value as the
   // position for the "parent"
   assert(childProc);
   const process *parent = childProc->getParent();
   assert(parent);
   Thread *thr = parent->threads[0];
   childProc->threads[0]->CTvector->dup(result->theSampleId, iCounterId, thr, result->position_);
#endif

   return result;
}

bool sampledTimerReqNode::insertInstrumentation(process *theProc,
                                                metricDefinitionNode *) {
   timerPtr = (tTimer *)inferiorMalloc(theProc, sizeof(tTimer), dataHeap);
   if (timerPtr == NULL)
      return false; // failure!

   // Now let's initialize the newly allocated tTimer in the inferior heap:
   tTimer temp;
   P_memset(&temp, '\0', sizeof(tTimer));
   temp.id.id = this->theSampleId;
   temp.type = this->theTimerType;
   temp.normalize = 1000000;
   writeToInferiorHeap(theProc, temp);

   // Now instrument DYNINSTreportTimer:
   pdFunction *sampleFunction = theProc->findOneFunction("DYNINSTsampleValues");
   assert(sampleFunction);

   AstNode *ast, *tmp;
   tmp = new AstNode(AstNode::Constant, timerPtr);
   ast = new AstNode("DYNINSTreportTimer", tmp);
   removeAst(tmp);

   instPoint *func_entry = sampleFunction->funcEntry(theProc);
   sampler = addInstFunc(theProc, func_entry, ast,
                         callPreInsn, orderLastAtPoint, false);
   removeAst(ast);

#if defined(MT_THREAD)
   updateCounterTimerVectorMT(theProc,this->theSampleId,this->position_,getInferiorPtr());
#endif

   return true; // successful
}

void sampledTimerReqNode::disable(process *theProc,
                                  const vector<unsigVecType> &pointsToCheck) {
   // We used to remove the sample id from midToMiMap here but now the caller is
   // responsible for that.

   // Remove instrumentation added to DYNINSTsampleValues(), if necessary:
   if (sampler != NULL)
      ::deleteInst(sampler, getAllTrampsAtPoint(sampler));

   // Deallocate space for tTimer in the inferior heap:
   assert(timerPtr);
   inferiorFree(theProc, (unsigned)timerPtr, dataHeap, pointsToCheck);

#if defined(MT_THREAD)
    Thread *thr = theProc->threads[0];
    thr->CTvector->remove(this->theSampleId, this->position_);
#endif
}

void sampledTimerReqNode::writeToInferiorHeap(process *theProc,
                                              const tTimer &dataSrc) const {
   // using contents of "dataSrc", a local copy of the data,
   // write to inferior heap at loc "timerPtr" via proc->writeDataSpace()
   assert(timerPtr);
   theProc->writeDataSpace(timerPtr, sizeof(tTimer), &dataSrc);
}

bool sampledTimerReqNode::
unFork(dictionary_hash<instInstance*,instInstance*> &map) {
   instInstance *parentSamplerInstance = sampler;

   instInstance *childSamplerInstance;
   if (!map.find(parentSamplerInstance, childSamplerInstance))
      assert(false);

   vector<unsigned> pointsToCheck; // empty
   deleteInst(childSamplerInstance, pointsToCheck);

   map[parentSamplerInstance] = NULL; // since we've deleted...

   return true;
}
                                 
#endif

/* ****************************************************************** */

#ifdef SHM_SAMPLING
sampledShmWallTimerReqNode::sampledShmWallTimerReqNode(int iCounterId) :
                                                 dataReqNode() {
   theSampleId = iCounterId;

   // The following fields are NULL until insertInstrumentation():
   allocatedIndex = UINT_MAX;
   inferiorTimerPtr = NULL;
}

sampledShmWallTimerReqNode::
sampledShmWallTimerReqNode(const sampledShmWallTimerReqNode &src,
                           process *childProc,
                           metricDefinitionNode *mi,
                           int iCounterId) {
   // a dup()-like routine; call after a fork().
   // Assumes that the "childProc" has been duplicated already

   // Note that the index w/in the inferior heap remains the same, so setting the new
   // inferiorTimerPtr isn't too hard.  Actually, it's trivial, since other code
   // ensures that the new shm segment is placed in exactly the same virtual mem loc
   // as the previous one.
   //
   // Note that the fastInferiorHeap class's fork ctor will have already copied the
   // actual data; we need to fill in new meta-data (new houseKeeping entries).

   allocatedIndex = src.allocatedIndex;
   theSampleId = iCounterId;
   assert(theSampleId != src.theSampleId);

   fastInferiorHeap<wallTimerHK, tTimer> &theHeap =
                    childProc->getInferiorWallTimers();

   // since the new shm seg is placed in exactly the same memory location as the old
   // one, nothing here should change.
   tTimer *oldInferiorTimerPtr = src.inferiorTimerPtr;
   inferiorTimerPtr = theHeap.index2InferiorAddr(allocatedIndex);
   assert(inferiorTimerPtr == oldInferiorTimerPtr);

   // Write new raw value in the inferior heap:
   // we set localTimerPtr as follows: protector1 and procetor2 should be copied from
   //    src. total should be reset to 0.  start should be set to now if active else 0.
   //    counter should be copied from the source.
   // NOTE: SINCE WE COPY FROM THE SOURCE, IT'S IMPORTANT THAT ON A FORK, BOTH THE
   //       PARENT AND CHILD ARE PAUSED UNTIL WE COPY THINGS OVER.  THAT THE CHILD IS
   //       PAUSED IS NOTHING NEW; THAT THE PARENT SHOULD BE PAUSED IS NEW NEWS!

   tTimer *localTimerPtr = theHeap.index2LocalAddr(allocatedIndex);
   const tTimer *srcTimerPtr = childProc->getParent()->getInferiorWallTimers().index2LocalAddr(allocatedIndex);

   localTimerPtr->total = 0;
   localTimerPtr->counter = srcTimerPtr->counter;
   localTimerPtr->id.id   = theSampleId;
   localTimerPtr->protector1 = srcTimerPtr->protector1;
   localTimerPtr->protector2 = srcTimerPtr->protector2;

   if (localTimerPtr->counter == 0)
      // inactive timer...this is the easy case to copy
      localTimerPtr->start = 0; // undefined, really
   else
      // active timer...don't copy the start time from the source...make it 'now'
      localTimerPtr->start = getCurrWallTime();

   // write new HK for this tTimer:
   // Note: we don't assert anything about mi->getMId(), because that id has no
   // relation to the ids we work with (theSampleId).  In fact, we (the sampling code)
   // just don't ever care what mi->getMId() is.
   assert(theSampleId >= 0);
   assert(midToMiMap.defines(theSampleId));
   assert(midToMiMap[theSampleId] == mi);
   wallTimerHK iHKValue(theSampleId, mi, 0); // is last param right?
      // the mi should differ from the mi of the parent; theSampleId differs too.
   theHeap.initializeHKAfterFork(allocatedIndex, iHKValue);

#if defined(MT_THREAD)
   position_=0;
#endif
}

dataReqNode *
sampledShmWallTimerReqNode::dup(process *childProc,
                                metricDefinitionNode *mi,
                                int iCounterId,
                                const dictionary_hash<instInstance*,instInstance*> &
                                ) const {
   // duplicate 'this' (allocate w/ new) and return.  Call after a fork().

   sampledShmWallTimerReqNode *tmp;
   tmp = new sampledShmWallTimerReqNode(*this, childProc, mi, iCounterId);
      // fork constructor

#if defined(MT_THREAD)
   // initialize position for new counter id with the same value as the
   // position for the "parent"
   assert(childProc);
   const process *parent = childProc->getParent();
   assert(parent);
   Thread *thr = parent->threads[0];
   childProc->threads[0]->CTvector->dup(tmp->theSampleId, iCounterId, thr, tmp->position_);
#endif

   return tmp;
}

bool sampledShmWallTimerReqNode::insertInstrumentation(process *theProc,
                                                       metricDefinitionNode *iMi) {
   // Remember inferiorTimerPtr is NULL until this routine gets called.
   // WARNING: there will be an assert failure if the applic hasn't yet attached to the
   //          shm segment!!!

   // initialize the tTimer in the inferior heap
   tTimer iValue;
   P_memset(&iValue, '\0', sizeof(tTimer));
   iValue.id.id = this->theSampleId;

   wallTimerHK iHKValue(this->theSampleId, iMi, 0);

   fastInferiorHeap<wallTimerHK, tTimer> &theShmHeap =
          theProc->getInferiorWallTimers();

   if (!theShmHeap.alloc(iValue, iHKValue, this->allocatedIndex))
      return false; // failure

   inferiorTimerPtr = theShmHeap.getBaseAddrInApplic() + allocatedIndex;
      // ptr arith.  Now we know where in the inferior heap this counter is
      // attached to, so getInferiorPtr() can work ok.

   assert(inferiorTimerPtr == theShmHeap.index2InferiorAddr(allocatedIndex));
      // just a check for fun

#if defined(MT_THREAD)
   updateCounterTimerVectorMT(theProc,this->theSampleId,this->position_,getInferiorPtr());
#endif

   return true;
}

void sampledShmWallTimerReqNode::disable(process *theProc,
                                         const vector<unsigVecType> &pointsToCheck) {
   // We used to remove the sample id from midToMiMap here but now the caller is
   // responsible for that.

   fastInferiorHeap<wallTimerHK, tTimer> &theShmHeap = theProc->getInferiorWallTimers();

   // Remove from inferior heap; make sure we won't be sampled any more:
   vector<unsigned> trampsMaybeUsing;
   for (unsigned pointlcv=0; pointlcv < pointsToCheck.size(); pointlcv++)
      for (unsigned tramplcv=0; tramplcv < pointsToCheck[pointlcv].size(); tramplcv++)
         trampsMaybeUsing += pointsToCheck[pointlcv][tramplcv];

   theShmHeap.makePendingFree(allocatedIndex, trampsMaybeUsing);

#if defined(MT_THREAD)
    Thread *thr = theProc->threads[0];
    thr->CTvector->remove(this->theSampleId, this->position_);
#endif
}

/* ****************************************************************** */

sampledShmProcTimerReqNode::sampledShmProcTimerReqNode(int iCounterId) :
                                                 dataReqNode() {
   theSampleId = iCounterId;

   // The following fields are NULL until insertInstrumentatoin():
   allocatedIndex = UINT_MAX;
   inferiorTimerPtr = NULL;

#if defined(MT_THREAD)
   position_=0;
#endif
}

sampledShmProcTimerReqNode::
sampledShmProcTimerReqNode(const sampledShmProcTimerReqNode &src,
                           process *childProc,
                           metricDefinitionNode *mi,
                           int iCounterId) {
   // a dup()-like routine; call after a fork()
   // Assumes that the "childProc" has been duplicated already

   // Note that the index w/in the inferior heap remains the same, so setting the new
   // inferiorTimerPtr isn't too hard.  Actually, it's trivial, since other code
   // ensures that the new shm segment is placed in exactly the same virtual mem loc
   // as the previous one.
   //
   // Note that the fastInferiorHeap class's fork ctor will have already copied the
   // actual data; we need to fill in new meta-data (new houseKeeping entries).

   allocatedIndex = src.allocatedIndex;
   theSampleId = iCounterId;
   assert(theSampleId != src.theSampleId);

   fastInferiorHeap<processTimerHK, tTimer> &theHeap =
                    childProc->getInferiorProcessTimers();

   // since the new shm seg is placed in exactly the same memory location as the old
   // one, nothing here should change.
   tTimer *oldInferiorTimerPtr = src.inferiorTimerPtr;
   inferiorTimerPtr = theHeap.index2InferiorAddr(allocatedIndex);
   assert(inferiorTimerPtr == oldInferiorTimerPtr);

   // Write new raw value:
   // we set localTimerPtr as follows: protector1 and procetor2 should be copied from
   //    src. total should be reset to 0.  start should be set to now if active else 0.
   //    counter should be copied from the source.
   // NOTE: SINCE WE COPY FROM THE SOURCE, IT'S IMPORTANT THAT ON A FORK, BOTH THE
   //       PARENT AND CHILD ARE PAUSED UNTIL WE COPY THINGS OVER.  THAT THE CHILD IS
   //       PAUSED IS NOTHING NEW; THAT THE PARENT SHOULD BE PAUSED IS NEW NEWS!

   tTimer *localTimerPtr = theHeap.index2LocalAddr(allocatedIndex);
   const tTimer *srcTimerPtr = childProc->getParent()->getInferiorProcessTimers().index2LocalAddr(allocatedIndex);

   localTimerPtr->total = 0;
   localTimerPtr->counter = srcTimerPtr->counter;
   localTimerPtr->id.id   = theSampleId;
   localTimerPtr->protector1 = srcTimerPtr->protector1;
   localTimerPtr->protector2 = srcTimerPtr->protector2;

   if (localTimerPtr->counter == 0)
      // inactive timer...this is the easy case to copy
      localTimerPtr->start = 0; // undefined, really
   else
      // active timer...don't copy the start time from the source...make it 'now'
      localTimerPtr->start = childProc->getInferiorProcessCPUtime();

   // Write new HK for this tTimer:
   // Note: we don't assert anything about mi->getMId(), because that id has no
   // relation to the ids we work with (theSampleId).  In fact, we (the sampling code)
   // just don't ever care what mi->getMId() is.
   assert(theSampleId >= 0);
   assert(midToMiMap.defines(theSampleId));
   assert(midToMiMap[theSampleId] == mi);
   processTimerHK iHKValue(theSampleId, mi, 0); // is last param right?
      // the mi differs from the mi of the parent; theSampleId differs too.
   theHeap.initializeHKAfterFork(allocatedIndex, iHKValue);

#if defined(MT_THREAD)
   position_=0;
#endif
}

dataReqNode *
sampledShmProcTimerReqNode::dup(process *childProc,
                                metricDefinitionNode *mi,
                                int iCounterId,
                                const dictionary_hash<instInstance*,instInstance*> &
                                ) const {
   // duplicate 'this' (allocate w/ new) and return.  Call after a fork().

   sampledShmProcTimerReqNode *tmp;
   tmp = new sampledShmProcTimerReqNode(*this, childProc, mi, iCounterId);
      // fork constructor

#if defined(MT_THREAD)
   // initialize position for new counter id with the same value as the
   // position for the "parent"
   assert(childProc);
   const process *parent = childProc->getParent();
   assert(parent);
   Thread *thr = parent->threads[0];
   childProc->threads[0]->CTvector->dup(tmp->theSampleId, iCounterId, thr, tmp->position_);
#endif

   return tmp;
}

bool sampledShmProcTimerReqNode::insertInstrumentation(process *theProc,
                                                       metricDefinitionNode *iMi) {
   // Remember inferiorTimerPtr is NULL until this routine gets called.
   // WARNING: there will be an assert failure if the applic hasn't yet attached to the
   //          shm segment!!!

   // initialize the tTimer in the inferior heap
   tTimer iValue;
   P_memset(&iValue, '\0', sizeof(tTimer));
   iValue.id.id = this->theSampleId;

   processTimerHK iHKValue(this->theSampleId, iMi, 0);

   fastInferiorHeap<processTimerHK, tTimer> &theShmHeap =
          theProc->getInferiorProcessTimers();

   if (!theShmHeap.alloc(iValue, iHKValue, this->allocatedIndex))
      return false; // failure

   inferiorTimerPtr = theShmHeap.getBaseAddrInApplic() + allocatedIndex;
      // ptr arith.  Now we know where in the inferior heap this counter is
      // attached to, so getInferiorPtr() can work ok.

   assert(inferiorTimerPtr == theShmHeap.index2InferiorAddr(allocatedIndex));
      // just a check for fun

#if defined(MT_THREAD)
   updateCounterTimerVectorMT(theProc,this->theSampleId,this->position_,getInferiorPtr());
#endif

   return true;
}

void sampledShmProcTimerReqNode::disable(process *theProc,
                                         const vector<unsigVecType> &pointsToCheck) {
   // We used to remove the sample id from midToMiMap here but now the caller is
   // responsible for that.

   fastInferiorHeap<processTimerHK, tTimer> &theShmHeap =
          theProc->getInferiorProcessTimers();

   // Remove from inferior heap; make sure we won't be sampled any more:
   vector<unsigned> trampsMaybeUsing;
   for (unsigned pointlcv=0; pointlcv < pointsToCheck.size(); pointlcv++)
      for (unsigned tramplcv=0; tramplcv < pointsToCheck[pointlcv].size(); tramplcv++)
         trampsMaybeUsing += pointsToCheck[pointlcv][tramplcv];

   theShmHeap.makePendingFree(allocatedIndex, trampsMaybeUsing);

#if defined(MT_THREAD)
   Thread *thr = theProc->threads[0];
   thr->CTvector->remove(this->theSampleId, this->position_);
#endif
}
#endif

/* **************************** */

void reportInternalMetrics(bool force) 
{
    if (isApplicationPaused())
       return; // we don't sample when paused (is this right?)

    static timeStamp end=0.0;

    // see if we have a sample to establish time base.
    if (!firstRecordTime) {
       cerr << "reportInternalMetrics: no because firstRecordTime==0" << endl;
       return;
    }

    if (end==0.0)
        end = (timeStamp)firstRecordTime/MILLION;

    const timeStamp now = getCurrentTime(false);

    //  check if it is time for a sample
    if (!force && now < end + samplingRate)  {
//        cerr << "reportInternalMetrics: no because now < end + samplingRate (end=" << end << "; samplingRate=" << samplingRate << "; now=" << now << ")" << endl;
//      cerr << "difference is " << (end+samplingRate-now) << endl;
        return;
    }

    timeStamp start = end;
    end = now;

    // TODO -- clean me up, please

    unsigned ai_size = internalMetric::allInternalMetrics.size();
    for (unsigned u=0; u<ai_size; u++) {
      internalMetric *theIMetric = internalMetric::allInternalMetrics[u];
      // Loop thru all enabled instances of this internal metric...

      for (unsigned v=0; v < theIMetric->num_enabled_instances(); v++) {
        internalMetric::eachInstance &theInstance = theIMetric->getEnabledInstance(v);
           // not "const" since bumpCumulativeValueBy() may be called

        sampleValue value = 0;
        if (theIMetric->name() == "active_processes") {
          //value = (end - start) * activeProcesses;
          value = (end - start) * theInstance.getValue();
        } else if (theIMetric->name() == "bucket_width") {
          //value = (end - start)* theInstance.getValue();
          // I would prefer to use (end-start) * theInstance.getValue(); however,
          // we've had some problems getting setValue() called in time, thus
          // leaving us with getValues() of 0 sometimes.  See longer comment in dynrpc.C --ari
          extern float currSamplingRate;
          value = (end - start) * currSamplingRate;
        } else if (theIMetric->name() == "number_of_cpus") {
          value = (end - start) * numberOfCPUs;
        } else if (theIMetric->name() == "total_CT") {
          value = (end - start) * internalMetricCounterId;
          assert(value>=0.0);
        } else if (theIMetric->name() == "active_CT") {
          value = (end - start) * activeCT;
          assert(value>=0.0);
        } else if (theIMetric->name() == "mem_CT") {
          value = (end - start) * memCT;
          assert(value>=0.0);
        } else if (theIMetric->name() == "infHeapMemAvailable") {
          value = (end - start) * inferiorMemAvailable;
          assert(value>=0.0);
        } else if (theIMetric->style() == EventCounter) {
          value = theInstance.getValue();
          // assert((value + 0.0001)  >= imp->cumulativeValue);
          value -= theInstance.getCumulativeValue();
          theInstance.bumpCumulativeValueBy(value);
        } else if (theIMetric->style() == SampledFunction) {
          value = theInstance.getValue();
        }

        theInstance.report(start, end, value);
           // calls metricDefinitionNode->forwardSimpleValue()
      }
    }
}

void disableAllInternalMetrics() {
    for (unsigned u=0; u < internalMetric::allInternalMetrics.size(); u++) {
      internalMetric *theIMetric = internalMetric::allInternalMetrics[u];

      // Now loop thru all the enabled instances of this internal metric...
      while (theIMetric->num_enabled_instances() > 0) {
        internalMetric::eachInstance &theInstance = theIMetric->getEnabledInstance(0);
        tp->endOfDataCollection(theInstance.getMId());
        theIMetric->disableInstance(0);
      }
    }  
}

#if defined(MT_THREAD)

void dataReqNode::updateCounterTimerVectorMT(process *proc, int CTid, 
                                             unsigned &position, 
                                             unsigned CTaddr)
{
#if defined(MT_DEBUG)
    sprintf(errorLine,"(pid=%d) Creating new Counter/Timer, CTid is %d\n",proc->pid, CTid);
    logLine(errorLine);
#endif

    unsigned tableAddr,addr;
    bool ok, err;

    assert(CTaddr && proc);

    // Getting thread table address
    tableAddr = proc->findInternalAddress("DYNINSTthreadTable",true, err);
    assert(!err);

#if defined(MT_DEBUG)
    sprintf(errorLine,"DYNINSTthreadTable address is 0x%x\n",tableAddr);
    logLine(errorLine);
#endif

    // Current thread
    // This should be a parameter to this function for the multi-threaded case
    Thread *thr = proc->threads[0];

    // Find the right CTvector address for this thread in the threadTable
    tableAddr += thr->get_pos()*sizeof(unsigned);

    // Check if the CTvector has enough room for another element. If not,
    // allocate more memory for it.
    // NOTE: for the final implementation, this has to be done for every
    // thread.
    ok = thr->CTvector->update(tableAddr);
    assert(ok);

    // Read address of vector of counter/timers for this thread
    proc->readDataSpace((caddr_t) tableAddr, 
                        sizeof(unsigned),
                        (caddr_t) &addr,true);
    assert(addr);

    thr->CTvector->add(CTid, position);

    // Save pointer to Counter/Timer in table of counter/timers 
    // (using CTid as offset)
    addr += sizeof(unsigned)*(position);
    proc->writeDataSpace((caddr_t) addr, sizeof(unsigned),
                         (caddr_t) &CTaddr);
}

#endif