threads.c

Go to the documentation of this file.

 
/* #[ License : */
/*
 *   Copyright (C) 1984-2026 J.A.M. Vermaseren
 *   When using this file you are requested to refer to the publication
 *   J.A.M.Vermaseren "New features of FORM" math-ph/0010025
 *   This is considered a matter of courtesy as the development was paid
 *   for by FOM the Dutch physics granting agency and we would like to
 *   be able to track its scientific use to convince FOM of its value
 *   for the community.
 *
 *   This file is part of FORM.
 *
 *   FORM is free software: you can redistribute it and/or modify it under the
 *   terms of the GNU General Public License as published by the Free Software
 *   Foundation, either version 3 of the License, or (at your option) any later
 *   version.
 *
 *   FORM is distributed in the hope that it will be useful, but WITHOUT ANY
 *   WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS
 *   FOR A PARTICULAR PURPOSE.  See the GNU General Public License for more
 *   details.
 *
 *   You should have received a copy of the GNU General Public License along
 *   with FORM.  If not, see <http://www.gnu.org/licenses/>.
 */
/* #] License : */ 
 
#ifdef WITHPTHREADS
 
#define WHOLEBRACKETS
/*
    #[ Variables :
 
    The sortbot additions are from 17-may-2007 and after. They constitute
    an attempt to make the final merge sorting faster for the master.
    This way the master has only one compare per term.
    It does add some complexity, but the final merge routine (SortBotMasterMerge)
    is much simpler for the sortbots.
*/
 
#include "form3.h"
#include <math.h>
 
#ifdef WITH_ALARM
// This is only required if we are blocking SIG_ALRM in the worker threads.
#include <signal.h>
#endif
 
#ifdef WITHFLOAT
#include <gmp.h>
 
int PackFloat(WORD *,mpf_t);
int UnpackFloat(mpf_t, WORD *);
void RatToFloat(mpf_t result, UWORD *formrat, int ratsize);
#endif
 
static int numberofthreads;
static int numberofworkers;
static int identityofthreads = 0;
static int *listofavailables;
static int topofavailables = 0;
static pthread_key_t identitykey;
static INILOCK(numberofthreadslock)
static INILOCK(availabilitylock)
static pthread_t *threadpointers = 0;
static pthread_mutex_t *wakeuplocks;
static pthread_mutex_t *wakeupmasterthreadlocks;
static pthread_cond_t *wakeupconditions;
static pthread_condattr_t *wakeupconditionattributes;
static int *wakeup;
static int *wakeupmasterthread;
static INILOCK(wakeupmasterlock)
static pthread_cond_t wakeupmasterconditions = PTHREAD_COND_INITIALIZER;
static pthread_cond_t *wakeupmasterthreadconditions;
static int wakeupmaster = 0;
static int identityretval;
/* static INILOCK(clearclocklock) */
static LONG *timerinfo;
static LONG *sumtimerinfo;
static int numberclaimed;
 
static THREADBUCKET **threadbuckets, **freebuckets;
static int numthreadbuckets;
static int numberoffullbuckets;
 
/* static int numberbusy = 0; */
 
INILOCK(dummylock)
INIRWLOCK(dummyrwlock)
static pthread_cond_t dummywakeupcondition = PTHREAD_COND_INITIALIZER;
 
static POSITION SortBotPosition;
static INILOCK(wakeupsortbotlock)
static pthread_cond_t wakeupsortbotconditions = PTHREAD_COND_INITIALIZER;
static int topsortbotavailables = 0;
static LONG numberofterms;
 
/*
    #] Variables : 
    #[ Identity :
        #[ StartIdentity :
*/
void StartIdentity(void)
{
    pthread_key_create(&identitykey,FinishIdentity);
}
 
/*
        #] StartIdentity : 
        #[ FinishIdentity :
*/
void FinishIdentity(void *keyp)
{
    DUMMYUSE(keyp);
/*  free(keyp); */
}
 
/*
        #] FinishIdentity : 
        #[ SetIdentity :
*/
int SetIdentity(int *identityretval)
{
/*
#ifdef _MSC_VER
    printf("addr %d\n",&numberofthreadslock);
    printf("size %d\n",sizeof(numberofthreadslock));
#endif
*/
    LOCK(numberofthreadslock);
    *identityretval = identityofthreads++;
    UNLOCK(numberofthreadslock);
    pthread_setspecific(identitykey,(void *)identityretval);
    return(*identityretval);
}
 
/*
        #] SetIdentity : 
        #[ WhoAmI :
*/
 
int WhoAmI(void)
{
    int *identity;
/*
    First a fast exit for when there is at most one thread
*/
    if ( identityofthreads <= 1 ) return(0);
/*
    Now the reading of the key.
    According to the book the statement should read:
 
    pthread_getspecific(identitykey,(void **)(&identity));
 
    but according to the information in pthread.h it is:
*/
    identity = (int *)pthread_getspecific(identitykey);
    return(*identity);
}
 
/*
        #] WhoAmI : 
        #[ BeginIdentities :
*/
void BeginIdentities(void)
{
    StartIdentity();
    SetIdentity(&identityretval);
}
 
/*
        #] BeginIdentities : 
    #] Identity : 
    #[ StartHandleLock :
*/
void StartHandleLock(void)
{
    AM.handlelock = dummyrwlock;
}
 
/*
    #] StartHandleLock : 
    #[ GetNumberOfSortBots :
*/
int GetNumberOfSortBots(int numberofworkers)
{
    if (numberofworkers <= 1)
        return 0;
 
    return numberofworkers - 1;
}
/*
    #] GetNumberOfSortBots :
    #[ StartAllThreads :
*/
int StartAllThreads(int number)
{
    int identity, j, dummy, mul;
    ALLPRIVATES *B;
    pthread_t thethread;
    identity = WhoAmI();
 
 
    timerinfo = (LONG *)Malloc1(sizeof(LONG)*number*2,"timerinfo");
    sumtimerinfo = (LONG *)Malloc1(sizeof(LONG)*number*2,"sumtimerinfo");
    for ( j = 0; j < number*2; j++ ) { timerinfo[j] = 0; sumtimerinfo[j] = 0; }
    mul = 2;
 
    listofavailables = (int *)Malloc1(sizeof(int)*(number+1),"listofavailables");
    threadpointers = (pthread_t *)Malloc1(sizeof(pthread_t)*number*mul,"threadpointers");
    AB = (ALLPRIVATES **)Malloc1(sizeof(ALLPRIVATES *)*number*mul,"Private structs");
 
    wakeup = (int *)Malloc1(sizeof(int)*number*mul,"wakeup");
    wakeuplocks = (pthread_mutex_t *)Malloc1(sizeof(pthread_mutex_t)*number*mul,"wakeuplocks");
    wakeupconditions = (pthread_cond_t *)Malloc1(sizeof(pthread_cond_t)*number*mul,"wakeupconditions");
    wakeupconditionattributes = (pthread_condattr_t *)
            Malloc1(sizeof(pthread_condattr_t)*number*mul,"wakeupconditionattributes");
 
    wakeupmasterthread = (int *)Malloc1(sizeof(int)*number*mul,"wakeupmasterthread");
    wakeupmasterthreadlocks = (pthread_mutex_t *)Malloc1(sizeof(pthread_mutex_t)*number*mul,"wakeupmasterthreadlocks");
    wakeupmasterthreadconditions = (pthread_cond_t *)Malloc1(sizeof(pthread_cond_t)*number*mul,"wakeupmasterthread");
 
    numberofthreads = number;
    numberofworkers = number - 1;
    threadpointers[identity] = pthread_self();
    topofavailables = 0;
 
#ifdef WITH_ALARM
    /* During thread creation, we block SIGALRM on the main thread. The created
       threads will inherit this. This is required for #timeout to work properly
       in TFORM: only the main thread should recieve SIGALRM. */
    sigset_t sig_set;
    sigemptyset(&sig_set);
    sigaddset(&sig_set, SIGALRM);
    pthread_sigmask(SIG_BLOCK, &sig_set, NULL);
#endif
 
    for ( j = 1; j < number; j++ ) {
        if ( pthread_create(&thethread,NULL,RunThread,(void *)(&dummy)) )
            goto failure;
    }
/*
    Now we initialize the master at the same time that the workers are doing so.
*/
    B = InitializeOneThread(identity);
    AR.infile = &(AR.Fscr[0]);
    AR.outfile = &(AR.Fscr[1]);
    AR.hidefile = &(AR.Fscr[2]);
    AM.sbuflock = dummylock;
    AS.inputslock = dummylock;
    AS.outputslock = dummylock;
    AS.MaxExprSizeLock = dummylock;
    AP.PreVarLock = dummylock;
    AC.halfmodlock = dummylock;
    MakeThreadBuckets(number,0);
/*
    Now we wait for the workers to finish their startup.
    We don't want to initialize the sortbots yet and run the risk that 
    some of them may end up with a lower number than one of the workers.
*/
    MasterWaitAll();
 
    for ( j = 0; j < GetNumberOfSortBots(numberofworkers); j++ ) {
        if ( pthread_create(&thethread,NULL,RunSortBot,(void *)(&dummy)) )
            goto failure;
    }
    MasterWaitAllSortBots();
    DefineSortBotTree();
    IniSortBlocks(number-1);
    AS.MasterSort = 0;
    AM.storefilelock = dummylock;
 
#ifdef WITH_ALARM
    /* Now we allow the main thread to recieve SIGALRM again. */
    pthread_sigmask(SIG_UNBLOCK, &sig_set, NULL);
#endif
 
/*
MesPrint("AB = %x %x %x  %d",AB[0],AB[1],AB[2], identityofthreads);
*/
    return(0);
/* INTERNAL_ERROR_EXCL_START */
failure:
    MesPrint("!>Cannot start %d threads",number);
    Terminate(-1);
    return(-1);
/* INTERNAL_ERROR_EXCL_STOP */
}
 
/*
    #] StartAllThreads : 
    #[ InitializeOneThread :
*/
UBYTE *scratchname[] = { (UBYTE *)"scratchsize",
                         (UBYTE *)"scratchsize",
                         (UBYTE *)"hidesize" };
ALLPRIVATES *InitializeOneThread(int identity)
{
    WORD *t, *ScratchBuf;
    int i, j, bsize, *bp;
    LONG ScratchSize[3], IOsize;
    ALLPRIVATES *B;
    UBYTE *s;
 
    wakeup[identity] = 0;
    wakeuplocks[identity] = dummylock;
    pthread_condattr_init(&(wakeupconditionattributes[identity]));
    pthread_condattr_setpshared(&(wakeupconditionattributes[identity]),PTHREAD_PROCESS_PRIVATE);
    wakeupconditions[identity] = dummywakeupcondition;
    pthread_cond_init(&(wakeupconditions[identity]),&(wakeupconditionattributes[identity]));
    wakeupmasterthread[identity] = 0;
    wakeupmasterthreadlocks[identity] = dummylock;
    wakeupmasterthreadconditions[identity] = dummywakeupcondition;
 
    bsize = sizeof(ALLPRIVATES);
    bsize = (bsize+sizeof(int)-1)/sizeof(int);
    B = (ALLPRIVATES *)Malloc1(sizeof(int)*bsize,"B struct");
    for ( bp = (int *)B, j = 0; j < bsize; j++ ) *bp++ = 0;
 
    AB[identity] = B;
/*
            12-jun-2007 JV:
    For the timing one has to know a few things:
    The POSIX standard is that there is only a single process ID and that
    getrusage returns the time of all the threads together.
    Under Linux there are two methods though: The older LinuxThreads and NPTL.
    LinuxThreads gives each thread its own process id. This makes that we
    can time the threads with getrusage, and hence this was done. Under NPTL
    this has been 'corrected' and suddenly getruage doesn't work anymore the
    way it used to. Now we need
        clock_gettime(CLOCK_THREAD_CPUTIME_ID,&timing)
    which is declared in <time.h> and we need -lrt extra in the link statement.
    (this is at least the case on blade02 at DESY-Zeuthen).
    See also the code in tools.c at the routine Timer.
    We may still have to include more stuff there.
*/
    if ( identity > 0 ) TimeCPU(0);
 
 
    if ( identity > numberofworkers ) {
/*
        Some workspace is needed when we have a PolyFun and we have to add
        two terms and the new result is going to be longer than the old result.
*/
        LONG length = AM.WorkSize*sizeof(WORD)/8+AM.MaxTer*2;
        AT.WorkSpace = (WORD *)Malloc1(length,"WorkSpace");
        AT.WorkTop = AT.WorkSpace + length/sizeof(WORD);
        AT.WorkPointer = AT.WorkSpace;
        AT.identity = identity;
/*
        The SB struct gets treated in IniSortBlocks.
        The SortBotIn variables will be defined DefineSortBotTree.
*/
        if ( AN.SoScratC == 0 ) {
            AN.SoScratC = (UWORD *)Malloc1(2*(AM.MaxTal+2)*sizeof(UWORD),"Scratch in SortBot");
        }
        AT.SS = (SORTING *)Malloc1(sizeof(SORTING),"dummy sort buffer");
        AT.SS->PolyFlag = 0;
 
        AT.comsym[0] = 8;
        AT.comsym[1] = SYMBOL;
        AT.comsym[2] = 4;
        AT.comsym[3] = 0;
        AT.comsym[4] = 1;
        AT.comsym[5] = 1;
        AT.comsym[6] = 1;
        AT.comsym[7] = 3;
        AT.comnum[0] = 4;
        AT.comnum[1] = 1;
        AT.comnum[2] = 1;
        AT.comnum[3] = 3;
        AT.comfun[0] = FUNHEAD+4;
        AT.comfun[1] = FUNCTION;
        AT.comfun[2] = FUNHEAD;
        AT.comfun[3] = 0;
#if FUNHEAD > 3
        for ( i = 4; i <= FUNHEAD; i++ )
            AT.comfun[i] = 0;
#endif
        AT.comfun[FUNHEAD+1] = 1;
        AT.comfun[FUNHEAD+2] = 1;
        AT.comfun[FUNHEAD+3] = 3;
        AT.comind[0] = 7;
        AT.comind[1] = INDEX;
        AT.comind[2] = 3;
        AT.comind[3] = 0;
        AT.comind[4] = 1;
        AT.comind[5] = 1;
        AT.comind[6] = 3;
 
        AT.inprimelist = -1;
        AT.sizeprimelist = 0;
        AT.primelist = 0;
        AT.bracketinfo = 0;
        
        AR.CompareRoutine = (COMPAREDUMMY)(&Compare1);
 
        AR.sLevel = 0;
        AR.wranfia = 0;
        AR.wranfcall = 0;
        AR.wranfnpair1 = NPAIR1;
        AR.wranfnpair2 = NPAIR2;
        AN.NumFunSorts = 5;
        AN.MaxFunSorts = 5;
        AN.SplitScratch = 0;
        AN.SplitScratchSize = AN.InScratch = 0;
        AN.SplitScratch1 = 0;
        AN.SplitScratchSize1 = AN.InScratch1 = 0;
 
        AN.FunSorts = (SORTING **)Malloc1((AN.NumFunSorts+1)*sizeof(SORTING *),"FunSort pointers");
        for ( i = 0; i <= AN.NumFunSorts; i++ ) AN.FunSorts[i] = 0;
        AN.FunSorts[0] = AT.S0 = AT.SS;
        AN.idfunctionflag = 0;
        AN.tryterm = 0;
 
        return(B);
    }
    if ( identity == 0 && AN.SoScratC == 0 ) {
        AN.SoScratC = (UWORD *)Malloc1(2*(AM.MaxTal+2)*sizeof(UWORD),"Scratch in SortBot");
    }
    AR.CurDum = AM.IndDum;
    for ( j = 0; j < 3; j++ ) {
        if ( identity == 0 ) {
            if ( j == 2 ) {
                ScratchSize[j] = AM.HideSize;
            }
            else {
                ScratchSize[j] = AM.ScratSize;
            }
            if ( ScratchSize[j] < 10*AM.MaxTer ) ScratchSize[j] = 10 * AM.MaxTer;
        }
        else {
/*
            ScratchSize[j] = AM.ScratSize / (numberofthreads-1);
            ScratchSize[j] = ScratchSize[j] / 20;
            if ( ScratchSize[j] < 10*AM.MaxTer ) ScratchSize[j] = 10 * AM.MaxTer;
*/
            if ( j == 1 ) ScratchSize[j] = AM.ThreadScratOutSize;
            else          ScratchSize[j] = AM.ThreadScratSize;
            if ( ScratchSize[j] < 4*AM.MaxTer ) ScratchSize[j] = 4 * AM.MaxTer;
            AR.Fscr[j].name = 0;
        }
        ScratchSize[j] = ( ScratchSize[j] + 255 ) / 256;
        ScratchSize[j] = ScratchSize[j] * 256;
        ScratchBuf = (WORD *)Malloc1(ScratchSize[j]*sizeof(WORD),(char *)(scratchname[j]));
        AR.Fscr[j].POsize = ScratchSize[j] * sizeof(WORD);
        AR.Fscr[j].POfull = AR.Fscr[j].POfill = AR.Fscr[j].PObuffer = ScratchBuf;
        AR.Fscr[j].POstop = AR.Fscr[j].PObuffer + ScratchSize[j];
        PUTZERO(AR.Fscr[j].POposition);
        AR.Fscr[j].pthreadslock = dummylock;
        AR.Fscr[j].wPOsize = AR.Fscr[j].POsize;
        AR.Fscr[j].wPObuffer = AR.Fscr[j].PObuffer;
        AR.Fscr[j].wPOfill = AR.Fscr[j].POfill;
        AR.Fscr[j].wPOfull = AR.Fscr[j].POfull;
        AR.Fscr[j].wPOstop = AR.Fscr[j].POstop;
    }
    AR.InInBuf = 0;
    AR.InHiBuf = 0;
    AR.Fscr[0].handle = -1;
    AR.Fscr[1].handle = -1;
    AR.Fscr[2].handle = -1;
    AR.FoStage4[0].handle = -1;
    AR.FoStage4[1].handle = -1;
    IOsize = AM.S0->file.POsize;
#ifdef WITHZLIB
    AR.FoStage4[0].ziosize = IOsize;
    AR.FoStage4[1].ziosize = IOsize;
    AR.FoStage4[0].ziobuffer = 0;
    AR.FoStage4[1].ziobuffer = 0;
#endif  
    AR.FoStage4[0].POsize  = ((IOsize+sizeof(WORD)-1)/sizeof(WORD))*sizeof(WORD);
    AR.FoStage4[1].POsize  = ((IOsize+sizeof(WORD)-1)/sizeof(WORD))*sizeof(WORD);
 
    AR.hidefile = &(AR.Fscr[2]);
    AR.StoreData.Handle = -1;
    AR.SortType = AC.SortType;
 
    AN.IndDum = AM.IndDum;
 
    if ( identity > 0 ) {
        s = (UBYTE *)(FG.fname); i = 0;
        while ( *s ) { s++; i++; }
        s = (UBYTE *)Malloc1(sizeof(char)*(i+12),"name for Fscr[0] file");
        snprintf((char *)s,i+12,"%s.%d",FG.fname,identity);
        s[i-3] = 's'; s[i-2] = 'c'; s[i-1] = '0';
        AR.Fscr[0].name = (char *)s;
        s = (UBYTE *)(FG.fname); i = 0;
        while ( *s ) { s++; i++; }
        s = (UBYTE *)Malloc1(sizeof(char)*(i+12),"name for Fscr[1] file");
        snprintf((char *)s,i+12,"%s.%d",FG.fname,identity);
        s[i-3] = 's'; s[i-2] = 'c'; s[i-1] = '1';
        AR.Fscr[1].name = (char *)s;
    }
 
    AR.CompressBuffer = (WORD *)Malloc1((AM.CompressSize+10)*sizeof(WORD),"compresssize");
    AR.ComprTop = AR.CompressBuffer + AM.CompressSize;
    AR.CompareRoutine = (COMPAREDUMMY)(&Compare1);
/*
    Here we make all allocations for the struct AT
    (which is AB[identity].T or B->T with B = AB+identity).
*/
    AT.WorkSpace = (WORD *)Malloc1(AM.WorkSize*sizeof(WORD),"WorkSpace");
    AT.WorkTop = AT.WorkSpace + AM.WorkSize;
    AT.WorkPointer = AT.WorkSpace;
 
    AT.Nest = (NESTING)Malloc1((LONG)sizeof(struct NeStInG)*AM.maxFlevels,"functionlevels");
    AT.NestStop = AT.Nest + AM.maxFlevels;
    AT.NestPoin = AT.Nest;
 
    AT.WildMask = (WORD *)Malloc1((LONG)AM.MaxWildcards*sizeof(WORD),"maxwildcards");
 
    LOCK(availabilitylock);
    AT.ebufnum = inicbufs();        /* Buffer for extras during execution */
    AT.fbufnum = inicbufs();        /* Buffer for caching in factorization */
    AT.allbufnum = inicbufs();      /* Buffer for id,all */
    AT.aebufnum = inicbufs();       /* Buffer for id,all */
    UNLOCK(availabilitylock);
 
    AT.RepCount = (int *)Malloc1((LONG)((AM.RepMax+3)*sizeof(int)),"repeat buffers");
    AN.RepPoint = AT.RepCount;
    AN.polysortflag = 0;
    AN.subsubveto = 0;
    AN.tryterm = 0;
    AT.RepTop = AT.RepCount + AM.RepMax;
 
    AT.WildArgTaken = (WORD *)Malloc1((LONG)AC.WildcardBufferSize*sizeof(WORD)/2
                ,"argument list names");
    AT.WildcardBufferSize = AC.WildcardBufferSize;
    AT.previousEfactor = 0;
 
    AT.identity = identity;
    AT.LoadBalancing = 0;
/*
    Still to do: the SS stuff.
                 the Fscr[3]
                 the FoStage4[2]
*/
    if ( AT.WorkSpace == 0 ||
         AT.Nest == 0 ||
         AT.WildMask == 0 ||
         AT.RepCount == 0 ||
         AT.WildArgTaken == 0 ) goto OnError;
/*
    And initializations
*/
    AT.comsym[0] = 8;
    AT.comsym[1] = SYMBOL;
    AT.comsym[2] = 4;
    AT.comsym[3] = 0;
    AT.comsym[4] = 1;
    AT.comsym[5] = 1;
    AT.comsym[6] = 1;
    AT.comsym[7] = 3;
    AT.comnum[0] = 4;
    AT.comnum[1] = 1;
    AT.comnum[2] = 1;
    AT.comnum[3] = 3;
    AT.comfun[0] = FUNHEAD+4;
    AT.comfun[1] = FUNCTION;
    AT.comfun[2] = FUNHEAD;
    AT.comfun[3] = 0;
#if FUNHEAD > 3
    for ( i = 4; i <= FUNHEAD; i++ )
        AT.comfun[i] = 0;
#endif
    AT.comfun[FUNHEAD+1] = 1;
    AT.comfun[FUNHEAD+2] = 1;
    AT.comfun[FUNHEAD+3] = 3;
    AT.comind[0] = 7;
    AT.comind[1] = INDEX;
    AT.comind[2] = 3;
    AT.comind[3] = 0;
    AT.comind[4] = 1;
    AT.comind[5] = 1;
    AT.comind[6] = 3;
    AT.locwildvalue[0] = SUBEXPRESSION;
    AT.locwildvalue[1] = SUBEXPSIZE;
    for ( i = 2; i < SUBEXPSIZE; i++ ) AT.locwildvalue[i] = 0;
    AT.mulpat[0] = TYPEMULT;
    AT.mulpat[1] = SUBEXPSIZE+3;
    AT.mulpat[2] = 0;
    AT.mulpat[3] = SUBEXPRESSION;
    AT.mulpat[4] = SUBEXPSIZE;
    AT.mulpat[5] = 0;
    AT.mulpat[6] = 1;
    for ( i = 7; i < SUBEXPSIZE+5; i++ ) AT.mulpat[i] = 0;
    AT.proexp[0] = SUBEXPSIZE+4;
    AT.proexp[1] = EXPRESSION;
    AT.proexp[2] = SUBEXPSIZE;
    AT.proexp[3] = -1;
    AT.proexp[4] = 1;
    for ( i = 5; i < SUBEXPSIZE+1; i++ ) AT.proexp[i] = 0;
    AT.proexp[SUBEXPSIZE+1] = 1;
    AT.proexp[SUBEXPSIZE+2] = 1;
    AT.proexp[SUBEXPSIZE+3] = 3;
    AT.proexp[SUBEXPSIZE+4] = 0;
    AT.dummysubexp[0] = SUBEXPRESSION;
    AT.dummysubexp[1] = SUBEXPSIZE+4;
    for ( i = 2; i < SUBEXPSIZE; i++ ) AT.dummysubexp[i] = 0;
    AT.dummysubexp[SUBEXPSIZE] = WILDDUMMY;
    AT.dummysubexp[SUBEXPSIZE+1] = 4;
    AT.dummysubexp[SUBEXPSIZE+2] = 0;
    AT.dummysubexp[SUBEXPSIZE+3] = 0;
 
    AT.MinVecArg[0] = 7+ARGHEAD;
    AT.MinVecArg[ARGHEAD] = 7;
    AT.MinVecArg[1+ARGHEAD] = INDEX;
    AT.MinVecArg[2+ARGHEAD] = 3;
    AT.MinVecArg[3+ARGHEAD] = 0;
    AT.MinVecArg[4+ARGHEAD] = 1;
    AT.MinVecArg[5+ARGHEAD] = 1;
    AT.MinVecArg[6+ARGHEAD] = -3;
    t = AT.FunArg;
    *t++ = 4+ARGHEAD+FUNHEAD;
    for ( i = 1; i < ARGHEAD; i++ ) *t++ = 0;
    *t++ = 4+FUNHEAD;
    *t++ = 0;
    *t++ = FUNHEAD;
    for ( i = 2; i < FUNHEAD; i++ ) *t++ = 0;
    *t++ = 1; *t++ = 1; *t++ = 3;
 
    AT.inprimelist = -1;
    AT.sizeprimelist = 0;
    AT.primelist = 0;
    AT.nfac = AT.nBer = 0;
    AT.factorials = 0;
    AT.bernoullis = 0;
    AR.wranfia = 0;
    AR.wranfcall = 0;
    AR.wranfnpair1 = NPAIR1;
    AR.wranfnpair2 = NPAIR2;
    AR.wranfseed = 0;
 
    AT.NormData = Malloc1(sizeof(*(AT.NormData)), "NormData thread pointers");
    AT.NormDataSize = 1;
    AT.NormData[0] = AllocNormData();
    AT.NormDepth = 0;
 
    AN.SplitScratch = 0;
    AN.SplitScratchSize = AN.InScratch = 0;
    AN.SplitScratch1 = 0;
    AN.SplitScratchSize1 = AN.InScratch1 = 0;
/*
    Now the sort buffers. They depend on which thread. The master
    inherits the sortbuffer from AM.S0
*/
    if ( identity == 0 ) {
        AT.S0 = AM.S0;
    }
    else {
/*
        For the moment we don't have special settings.
        They may become costly in virtual memory.
*/
        AT.S0 = AllocSort(AM.S0->LargeSize*sizeof(WORD)/numberofworkers
                         ,AM.S0->SmallSize*sizeof(WORD)/numberofworkers
                         ,AM.S0->SmallEsize*sizeof(WORD)/numberofworkers
                         ,AM.S0->TermsInSmall
                         ,AM.S0->MaxPatches
/*                       ,AM.S0->MaxPatches/numberofworkers  */
                         ,AM.S0->MaxFpatches/numberofworkers
                         ,AM.S0->file.POsize
                         ,0);
    }
    AR.CompressPointer = AR.CompressBuffer;
/*
    Install the store caches (15-aug-2006 JV)
*/
    AT.StoreCache = AT.StoreCacheAlloc = 0;
    if ( AM.NumStoreCaches > 0 ) {
        STORECACHE sa, sb;
        LONG size;
        size = sizeof(struct StOrEcAcHe)+AM.SizeStoreCache;
        size = ((size-1)/sizeof(size_t)+1)*sizeof(size_t);
        AT.StoreCacheAlloc = (STORECACHE)Malloc1(size*AM.NumStoreCaches,"StoreCaches");
        sa = AT.StoreCache = AT.StoreCacheAlloc;
        for ( i = 0; i < AM.NumStoreCaches; i++ ) {
            sb = (STORECACHE)(void *)((UBYTE *)sa+size);
            if ( i == AM.NumStoreCaches-1 ) {
                sa->next = 0;
            }
            else {
                sa->next = sb;
            }
            SETBASEPOSITION(sa->position,-1);
            SETBASEPOSITION(sa->toppos,-1);
            sa = sb;
        }       
    }
 
    ReserveTempFiles(2);
    return(B);
/* INTERNAL_ERROR_EXCL_START */
OnError:;
    MLOCK(ErrorMessageLock);
    MesPrint("!>Error initializing thread %d",identity);
    MUNLOCK(ErrorMessageLock);
    Terminate(-1);
    return(B);
/* INTERNAL_ERROR_EXCL_STOP */
}
 
/*
    #] InitializeOneThread : 
    #[ FinalizeOneThread :
*/
void FinalizeOneThread(int identity)
{
    timerinfo[identity] = TimeCPU(1);
}
 
/*
    #] FinalizeOneThread : 
    #[ ClearAllThreads :
*/
void ClearAllThreads(void)
{
    int i;
    MasterWaitAll();
    for ( i = 1; i <= numberofworkers; i++ ) {
        WakeupThread(i,CLEARCLOCK);
    }
    for ( i = numberofworkers+1; i <= numberofworkers+GetNumberOfSortBots(numberofworkers); i++ ) {
        WakeupThread(i,CLEARCLOCK);
    }
}
 
/*
    #] ClearAllThreads : 
    #[ TerminateAllThreads :
*/
void TerminateAllThreads(void)
{
    int i;
    for ( i = 1; i <= numberofworkers; i++ ) {
        GetThread(i);
        WakeupThread(i,TERMINATETHREAD);
    }
    for ( i = numberofworkers+1; i <= numberofworkers+GetNumberOfSortBots(numberofworkers); i++ ) {
        WakeupThread(i,TERMINATETHREAD);
    }
    for ( i = 1; i <= numberofworkers; i++ ) {
        pthread_join(threadpointers[i],NULL);
    }
    for ( i = numberofworkers+1; i <= numberofworkers+GetNumberOfSortBots(numberofworkers); i++ ) {
        pthread_join(threadpointers[i],NULL);
    }
}
 
/*
    #] TerminateAllThreads : 
    #[ MakeThreadBuckets :
*/
int MakeThreadBuckets(int number, int par)
{
    int i;
    LONG sizethreadbuckets;
    THREADBUCKET *thr;
 
    // Here we divide by 4, which has been the behaviour with the default
    // AC.ThreadBucketSize for a long time.
    // MAXTER is the default value of AM.MaxTer, in units of sizeof(WORD).
    sizethreadbuckets = (AC.ThreadBucketSize*MAXTER)/4;
    // Now we scale up the buffer logarithmically with the user AM.MaxTer.
    // If we scale linearly, we end up with really enormous buffers here
    // when AM.MaxTer is of the order of millions of WORDs.
    float scale = 1.0;
    if ( AM.MaxTer/sizeof(WORD) > MAXTER ) {
        scale += log(((float)AM.MaxTer/sizeof(WORD))/MAXTER);
    }
    sizethreadbuckets = (LONG)((float)sizethreadbuckets*scale);
    // Nonetheless, we must fit at least BUCKETMINTERMS terms in each bucket!
    // So the buffer will eventually scale linearly with MaxTer anyway, but
    // much less aggressively than the old code.
    sizethreadbuckets = MaX((ULONG)sizethreadbuckets, BUCKETMINTERMS*AM.MaxTer/sizeof(WORD));
    
    if ( par == 0 ) {
        numthreadbuckets = 2*(number-1);
        threadbuckets = (THREADBUCKET **)Malloc1(numthreadbuckets*sizeof(THREADBUCKET *),"threadbuckets");
        freebuckets   = (THREADBUCKET **)Malloc1(numthreadbuckets*sizeof(THREADBUCKET *),"freebuckets");
    }
    if ( par > 0 ) {
        if ( sizethreadbuckets <= threadbuckets[0]->threadbuffersize ) return(0);
        for ( i = 0; i < numthreadbuckets; i++ ) {
            thr = threadbuckets[i];
            M_free(thr->deferbuffer,"deferbuffer");
            M_free(thr->threadbuffer,"threadbuffer");
            M_free(thr->compressbuffer,"compressbuffer");
        }
    }
    else {
        for ( i = 0; i < numthreadbuckets; i++ ) {
            threadbuckets[i] = (THREADBUCKET *)Malloc1(sizeof(THREADBUCKET),"threadbuckets");
            threadbuckets[i]->lock = dummylock;
        }
    }
    for ( i = 0; i < numthreadbuckets; i++ ) {
        thr = threadbuckets[i];
        thr->threadbuffersize = sizethreadbuckets;
        // This buffer does not need to be so large, start it at MaxTer.
        // We'll double it if necessary.
        thr->compressbuffersize = AM.MaxTer/sizeof(WORD);
        thr->free = BUCKETFREE;
        thr->deferbuffer = (POSITION *)Malloc1((AC.ThreadBucketSize+1)*sizeof(POSITION),"deferbuffer");
        thr->threadbuffer   = (WORD *)Malloc1(sizethreadbuckets*sizeof(WORD),"threadbuffer");
        thr->compressbuffer = (WORD *)Malloc1(thr->compressbuffersize*sizeof(WORD),"compressbuffer");
        thr->busy = BUCKETPREPARINGTERM;
        thr->usenum = thr->totnum = 0;
        thr->type = BUCKETDOINGTERMS;
    }
    return(0);
}
 
/*
    #] MakeThreadBuckets : 
    #[ GetTimerInfo :
*/
 
/* UNFINISHED_FEATURE_EXCL_START */
int GetTimerInfo(LONG** ti,LONG** sti)
{
    *ti = timerinfo;
    *sti = sumtimerinfo;
    return (numberofworkers + GetNumberOfSortBots(numberofworkers) + 1);
}
/* UNFINISHED_FEATURE_EXCL_STOP */
/*
    #] GetTimerInfo : 
    #[ WriteTimerInfo :
*/
 
/* UNFINISHED_FEATURE_EXCL_START */
void WriteTimerInfo(LONG* ti,LONG* sti)
{
    int i;
    int max = (numberofworkers + GetNumberOfSortBots(numberofworkers) + 1);
    for ( i=0; i<max; ++i ) {
        timerinfo[i] = ti[i];
        sumtimerinfo[i] = sti[i];
    }
}
/* UNFINISHED_FEATURE_EXCL_STOP */
/*
    #] WriteTimerInfo : 
    #[ GetWorkerTimes :
*/
LONG GetWorkerTimes(void)
{
    LONG retval = 0;
    int i;
    for ( i = 1; i <= numberofworkers; i++ ) retval += timerinfo[i] + sumtimerinfo[i];
    for ( i = numberofworkers+1; i <= numberofworkers+GetNumberOfSortBots(numberofworkers); i++ )
        retval += timerinfo[i] + sumtimerinfo[i];
    return(retval);
}
 
/*
    #] GetWorkerTimes : 
    #[ UpdateOneThread :
*/
int UpdateOneThread(int identity)
{
    ALLPRIVATES *B = AB[identity], *B0 = AB[0];
    AR.GetFile = AR0.GetFile;
    AR.KeptInHold = AR0.KeptInHold;
    AR.CurExpr = AR0.CurExpr;
    AR.SortType = AC.SortType;
    if ( AT.WildcardBufferSize < AC.WildcardBufferSize ) {
        M_free(AT.WildArgTaken,"argument list names");
        AT.WildcardBufferSize = AC.WildcardBufferSize;
        AT.WildArgTaken = (WORD *)Malloc1((LONG)AC.WildcardBufferSize*sizeof(WORD)/2
                ,"argument list names");
        if ( AT.WildArgTaken == 0 ) return(-1);
    }
    return(0);
}
 
/*
    #] UpdateOneThread : 
    #[ LoadOneThread :
*/
int LoadOneThread(int from, int identity, THREADBUCKET *thr, int par)
{
    WORD *t1, *t2;
    ALLPRIVATES *B = AB[identity], *B0 = AB[from];
 
    AR.DefPosition = AR0.DefPosition;
    AR.NoCompress = AR0.NoCompress;
    AR.gzipCompress = AR0.gzipCompress;
    AR.BracketOn = AR0.BracketOn;
    AR.CurDum = AR0.CurDum;
    AR.DeferFlag = AR0.DeferFlag;
    AR.TePos = 0;
    AR.sLevel = AR0.sLevel;
    AR.Stage4Name = AR0.Stage4Name;
    AR.GetOneFile = AR0.GetOneFile;
    AR.PolyFun = AR0.PolyFun;
    AR.PolyFunInv = AR0.PolyFunInv;
    AR.PolyFunType = AR0.PolyFunType;
    AR.PolyFunExp = AR0.PolyFunExp;
    AR.PolyFunVar = AR0.PolyFunVar;
    AR.PolyFunPow = AR0.PolyFunPow;
    AR.Eside = AR0.Eside;
    AR.Cnumlhs = AR0.Cnumlhs;
/*
    AR.MaxBracket = AR0.MaxBracket;
 
    The compressbuffer contents are mainly relevant for keep brackets
    We should do this only if there is a keep brackets statement
    We may however still need the compressbuffer for expressions in the rhs.
*/
    if ( par >= 1 ) {
/*
        We may not need this %%%%% 7-apr-2006
*/
        t1 = AR.CompressBuffer; t2 = AR0.CompressBuffer;
        while ( t2 < AR0.CompressPointer ) *t1++ = *t2++;
        AR.CompressPointer = t1;
 
    }
    else {
        AR.CompressPointer = AR.CompressBuffer;
    }
    if ( AR.DeferFlag ) {
        if ( AR.infile->handle < 0 ) {
            AR.infile->POfill = AR0.infile->POfill;
        }
        else {
/*
            We have to set the value of POposition to something that will
            force a read in the first try.
*/
            AR.infile->POfull = AR.infile->POfill = AR.infile->PObuffer;
        }
    }
    if ( par == 0 ) {
        AN.threadbuck = thr;
        AN.ninterms = thr->firstterm;
    }
    else if ( par == 1 ) {
        WORD *tstop;
        t1 = thr->threadbuffer; tstop = t1 + *t1;
        t2 = AT.WorkPointer;
        while ( t1 < tstop ) *t2++ = *t1++;
        AN.ninterms = thr->firstterm;
    }
    AN.TeInFun = 0;
    AN.ncmod = AC.ncmod;
    AT.BrackBuf = AT0.BrackBuf;
    AT.bracketindexflag = AT0.bracketindexflag;
    AN.PolyFunTodo = 0;
/*
    The relevant variables and the term are in their place.
    There is nothing more to do.
*/
    return(0);
}
 
/*
    #] LoadOneThread : 
    #[ BalanceRunThread :
*/
/* UNFINISHED_FEATURE_EXCL_START */
int BalanceRunThread(PHEAD int identity, WORD *term, WORD level)
{
    GETBIDENTITY
    ALLPRIVATES *BB;
    WORD *t, *tt;
    int i, *ti, *tti;
 
    LoadOneThread(AT.identity,identity,0,2);
/*
    Extra loading if needed. Quantities changed in Generator.
    Like the level that has to be passed.
*/
    BB = AB[identity];
    BB->R.level = level;
    BB->T.TMbuff = AT.TMbuff;
    ti = AT.RepCount; tti = BB->T.RepCount;
    i = AN.RepPoint - AT.RepCount;
    BB->N.RepPoint = BB->T.RepCount + i;
    for ( ; i >= 0; i-- ) tti[i] = ti[i];
 
    t = term; i = *term;
    tt = BB->T.WorkSpace;
    NCOPY(tt,t,i);
    BB->T.WorkPointer = tt;
 
    WakeupThread(identity,HIGHERLEVELGENERATION);
 
    return(0);
}
/* UNFINISHED_FEATURE_EXCL_STOP */
/*
    #] BalanceRunThread : 
    #[ SetWorkerFiles :
*/
void SetWorkerFiles(void)
{
    int id;
    ALLPRIVATES *B, *B0 = AB[0];
    for ( id = 1; id < AM.totalnumberofthreads; id++ ) {
        B = AB[id];
        AR.infile = &(AR.Fscr[0]);
        AR.outfile = &(AR.Fscr[1]);
        AR.hidefile = &(AR.Fscr[2]);
        AR.infile->handle = AR0.infile->handle;
        AR.hidefile->handle = AR0.hidefile->handle;
        if ( AR.infile->handle < 0 ) {
            AR.infile->PObuffer = AR0.infile->PObuffer;
            AR.infile->POstop = AR0.infile->POstop;
            AR.infile->POfill = AR0.infile->POfill;
            AR.infile->POfull = AR0.infile->POfull;
            AR.infile->POsize = AR0.infile->POsize;
            AR.InInBuf = AR0.InInBuf;
            AR.infile->POposition = AR0.infile->POposition;
            AR.infile->filesize = AR0.infile->filesize;
        }
        else {
            AR.infile->PObuffer = AR.infile->wPObuffer;
            AR.infile->POstop = AR.infile->wPOstop;
            AR.infile->POfill = AR.infile->wPOfill;
            AR.infile->POfull = AR.infile->wPOfull;
            AR.infile->POsize = AR.infile->wPOsize;
            AR.InInBuf = 0;
            PUTZERO(AR.infile->POposition);
        }
/*
        If there is some writing, it betters happens to ones own outfile.
        Currently this is to be done only for InParallel.
        Merging of the outputs is then done by the CopyExpression routine.
*/
        {
            AR.outfile->PObuffer = AR.outfile->wPObuffer;
            AR.outfile->POstop = AR.outfile->wPOstop;
            AR.outfile->POfill = AR.outfile->wPOfill;
            AR.outfile->POfull = AR.outfile->wPOfull;
            AR.outfile->POsize = AR.outfile->wPOsize;
            PUTZERO(AR.outfile->POposition);
        }
        if ( AR.hidefile->handle < 0 ) {
            AR.hidefile->PObuffer = AR0.hidefile->PObuffer;
            AR.hidefile->POstop = AR0.hidefile->POstop;
            AR.hidefile->POfill = AR0.hidefile->POfill;
            AR.hidefile->POfull = AR0.hidefile->POfull;
            AR.hidefile->POsize = AR0.hidefile->POsize;
            AR.InHiBuf = AR0.InHiBuf;
            AR.hidefile->POposition = AR0.hidefile->POposition;
            AR.hidefile->filesize = AR0.hidefile->filesize;
        }
        else {
            AR.hidefile->PObuffer = AR.hidefile->wPObuffer;
            AR.hidefile->POstop = AR.hidefile->wPOstop;
            AR.hidefile->POfill = AR.hidefile->wPOfill;
            AR.hidefile->POfull = AR.hidefile->wPOfull;
            AR.hidefile->POsize = AR.hidefile->wPOsize;
            AR.InHiBuf = 0;
            PUTZERO(AR.hidefile->POposition);
        }
    }
    if ( AR0.StoreData.dirtyflag ) {
        for ( id = 1; id < AM.totalnumberofthreads; id++ ) {
            B = AB[id];
            AR.StoreData = AR0.StoreData;
        }
    }
}
 
/*
    #] SetWorkerFiles : 
    #[ RunThread :
*/
void *RunThread(void *dummy)
{
    WORD *term, *ttin, *tt, *ttco, *oldwork;
    int identity, wakeupsignal, identityretv, i, tobereleased, errorcode;
    ALLPRIVATES *B;
    THREADBUCKET *thr;
    POSITION *ppdef;
    EXPRESSIONS e;
    DUMMYUSE(dummy);
    identity = SetIdentity(&identityretv);
    threadpointers[identity] = pthread_self();
    B = InitializeOneThread(identity);
    while ( ( wakeupsignal = ThreadWait(identity) ) > 0 ) {
        switch ( wakeupsignal ) {
/*
            #[ STARTNEWEXPRESSION :
*/
            case STARTNEWEXPRESSION:
/*
                Set up the sort routines etc.
                Start with getting some buffers synchronized with the compiler
*/
                if ( UpdateOneThread(identity) ) {
/* INTERNAL_ERROR_EXCL_START */
                    MLOCK(ErrorMessageLock);
                    MesPrint("!>Update error in starting expression in thread %d in module %d",identity,AC.CModule);
                    MUNLOCK(ErrorMessageLock);
                    Terminate(-1);
/* INTERNAL_ERROR_EXCL_STOP */
                }
                AR.DeferFlag = AC.ComDefer;
                AR.sLevel = AS.sLevel;
                AR.MaxDum = AM.IndDum;
                AR.expchanged = AB[0]->R.expchanged;
                AR.expflags = AB[0]->R.expflags;
                AR.PolyFun = AB[0]->R.PolyFun;
                AR.PolyFunInv = AB[0]->R.PolyFunInv;
                AR.PolyFunType = AB[0]->R.PolyFunType;
                AR.PolyFunExp = AB[0]->R.PolyFunExp;
                AR.PolyFunVar = AB[0]->R.PolyFunVar;
                AR.PolyFunPow = AB[0]->R.PolyFunPow;
/*
                Now fire up the sort buffer.
*/
                NewSort(BHEAD0);
                break;
/*
            #] STARTNEWEXPRESSION : 
            #[ LOWESTLEVELGENERATION :
*/
            case LOWESTLEVELGENERATION:
#ifdef INNERTEST
                if ( AC.InnerTest ) {
                    if ( StrCmp(AC.TestValue,(UBYTE *)INNERTEST) == 0 ) {
                        MesPrint("Testing(Worker%d): value = %s",AT.identity,AC.TestValue);
                    }
                }
#endif
                e = Expressions + AR.CurExpr;
                thr = AN.threadbuck;
                ppdef = thr->deferbuffer;
                ttin = thr->threadbuffer;
                ttco = thr->compressbuffer;
                term = AT.WorkPointer;
                thr->usenum = 0;
                tobereleased = 0;
                AN.inputnumber = thr->firstterm;
                AN.ninterms = thr->firstterm;
                do {
                  thr->usenum++;    /* For if the master wants to steal the bucket */
                  tt = term; i = *ttin;
                  NCOPY(tt,ttin,i);
                  AT.WorkPointer = tt;
                  if ( AR.DeferFlag ) {
                    tt = AR.CompressBuffer; i = *ttco;
                    NCOPY(tt,ttco,i);
                    AR.CompressPointer = tt;
                    AR.DefPosition = ppdef[0]; ppdef++;
                  }
                  if ( thr->free == BUCKETTERMINATED ) {
/*
                    The next statement allows the master to steal the bucket
                    for load balancing purposes. We do still execute the current
                    term, but afterwards we drop out.
                    Once we have written the release code, we cannot use this
                    bucket anymore. Hence the exit to the label bucketstolen.
*/
                    if ( thr->usenum == thr->totnum ) {
                        thr->free = BUCKETCOMINGFREE;
                    }
                    else {
                        thr->free = BUCKETRELEASED;
                        tobereleased = 1;
                    }
                  }
/*
                    What if we want to steal and we set thr->free while
                    the thread is inside the next code for a long time?
                  if ( AT.LoadBalancing ) {
*/
                    LOCK(thr->lock);
                    thr->busy = BUCKETDOINGTERM;
                    UNLOCK(thr->lock);
/*
                  }
                  else {
                    thr->busy = BUCKETDOINGTERM;
                  }
*/
                  AN.RepPoint = AT.RepCount + 1;
 
                  if ( ( e->vflags & ISFACTORIZED ) != 0 && term[1] == HAAKJE ) {
                    StoreTerm(BHEAD term);
                  }
                  else {
                  if ( AR.DeferFlag ) {
                    AR.CurDum = AN.IndDum = Expressions[AR.CurExpr].numdummies + AM.IndDum;
                  }
                  else {
                    AN.IndDum = AM.IndDum;
                    AR.CurDum = ReNumber(BHEAD term);
                  }
                  if ( AC.SymChangeFlag ) MarkDirty(term,DIRTYSYMFLAG);
                  if ( AN.ncmod ) {
                    if ( ( AC.modmode & ALSOFUNARGS ) != 0 ) MarkDirty(term,DIRTYFLAG);
                    else if ( AR.PolyFun ) PolyFunDirty(BHEAD term);
                  }
                  else if ( AC.PolyRatFunChanged ) PolyFunDirty(BHEAD term);
                  if ( ( AP.PreDebug & THREADSDEBUG ) != 0 ) {
                    MLOCK(ErrorMessageLock);
                    MesPrint("Thread %w executing term:");
                    PrintTerm(term,"LLG");
                    MUNLOCK(ErrorMessageLock);
                  }
                  if ( ( AR.PolyFunType == 2 ) && ( AC.PolyRatFunChanged == 0 )
                        && ( e->status == LOCALEXPRESSION || e->status == GLOBALEXPRESSION ) ) {
                        PolyFunClean(BHEAD term);
                  }
                  if ( Generator(BHEAD term,0) ) {
/* INTERNAL_ERROR_EXCL_START */
                    LowerSortLevel();
                    MLOCK(ErrorMessageLock);
                    MesPrint("!>Error in processing one term in thread %d in module %d",identity,AC.CModule);
                    MUNLOCK(ErrorMessageLock);
                    Terminate(-1);
/* INTERNAL_ERROR_EXCL_STOP */
                  }
                  AN.ninterms++;
                  }
/*                if ( AT.LoadBalancing ) { */
                    LOCK(thr->lock);
                    thr->busy = BUCKETPREPARINGTERM;
                    UNLOCK(thr->lock);
/*
                  }
                  else {
                    thr->busy = BUCKETPREPARINGTERM;
                  }
*/
                  if ( thr->free == BUCKETTERMINATED ) {
                    if ( thr->usenum == thr->totnum ) {
                        thr->free = BUCKETCOMINGFREE;
                    }
                    else {
                        thr->free = BUCKETRELEASED;
                        tobereleased = 1;
                    }
                  }
                  if ( tobereleased ) goto bucketstolen;
                } while ( *ttin );
                thr->free = BUCKETCOMINGFREE;
bucketstolen:;
/*              if ( AT.LoadBalancing ) { */
                    LOCK(thr->lock);
                    thr->busy = BUCKETTOBERELEASED;
                    UNLOCK(thr->lock);
/*              }
                else {
                    thr->busy = BUCKETTOBERELEASED;
                }
*/
                AT.WorkPointer = term;
                break;
/*
            #] LOWESTLEVELGENERATION : 
            #[ FINISHEXPRESSION :
*/
            case CLAIMOUTPUT:
                LOCK(AT.SB.MasterBlockLock[1]);
                break;
            case FINISHEXPRESSION:
/*
                Finish the sort
 
                Start with claiming the first block
                Once we have claimed it we can let the master know that
                everything is all right.
*/
                LOCK(AT.SB.MasterBlockLock[1]);
                ThreadClaimedBlock(identity);
/*
                Entry for when we work with sortbots
*/
                /* fall through */
            case FINISHEXPRESSION2:
/*
                Now we may need here an fsync on the sort file
*/
                if ( AC.ThreadSortFileSynch ) {
                  if ( AT.S0->file.handle >= 0 ) {
                    SynchFile(AT.S0->file.handle);
                  }
                }
                AT.SB.FillBlock = 1;
                AT.SB.MasterFill[1] = AT.SB.MasterStart[1];
                errorcode = EndSort(BHEAD AT.S0->sBuffer,0);
                UNLOCK(AT.SB.MasterBlockLock[AT.SB.FillBlock]);
                UpdateMaxSize();
                if ( errorcode ) {
                    MLOCK(ErrorMessageLock);
                    MesPrint("Error terminating sort in thread %d in module %d",identity,AC.CModule);
                    MUNLOCK(ErrorMessageLock);
                    Terminate(-1);
                }
                break;
/*
            #] FINISHEXPRESSION : 
            #[ CLEANUPEXPRESSION :
*/
            case CLEANUPEXPRESSION:
/*
                Cleanup everything and wait for the next expression
*/
                if ( AR.outfile->handle >= 0 ) {
                    CloseFile(AR.outfile->handle);
                    AR.outfile->handle = -1;
                    remove(AR.outfile->name);
                    AR.outfile->POfill = AR.outfile->POfull = AR.outfile->PObuffer;
                    PUTZERO(AR.outfile->POposition);
                    PUTZERO(AR.outfile->filesize);
                }
                else {
                    AR.outfile->POfill = AR.outfile->POfull = AR.outfile->PObuffer;
                    PUTZERO(AR.outfile->POposition);
                    PUTZERO(AR.outfile->filesize);
                }
                {
                    CBUF *C = cbuf+AT.ebufnum;
                    WORD **w, ii;
                    if ( C->numrhs > 0 || C->numlhs > 0 ) {
                        if ( C->rhs ) {
                            w = C->rhs; ii = C->numrhs;
                            do { *w++ = 0; } while ( --ii > 0 );
                        }
                        if ( C->lhs ) {
                            w = C->lhs; ii = C->numlhs;
                            do { *w++ = 0; } while ( --ii > 0 );
                        }
                        C->numlhs = C->numrhs = 0;
                        ClearTree(AT.ebufnum);
                        C->Pointer = C->Buffer;
                    }
                }
                break;
/*
            #] CLEANUPEXPRESSION : 
            #[ HIGHERLEVELGENERATION :
*/
            case HIGHERLEVELGENERATION:
/*
                When foliating halfway the tree.
                This should only be needed in a second level load balancing
*/
                term = AT.WorkSpace; AT.WorkPointer = term + *term;
                if ( Generator(BHEAD term,AR.level) ) {
                    LowerSortLevel();
                    MLOCK(ErrorMessageLock);
                    MesPrint("Error in load balancing one term at level %d in thread %d in module %d",AR.level,AT.identity,AC.CModule);
                    MUNLOCK(ErrorMessageLock);
                    Terminate(-1);
                }
                AT.WorkPointer = term;
                break;
/*
            #] HIGHERLEVELGENERATION : 
            #[ STARTNEWMODULE :
*/
            case STARTNEWMODULE:
/*
                For resetting variables.
*/
                SpecialCleanup(B);
                break;
/*
            #] STARTNEWMODULE : 
            #[ TERMINATETHREAD :
*/
            case TERMINATETHREAD:
                goto EndOfThread;
/*
            #] TERMINATETHREAD : 
            #[ DOONEEXPRESSION :
 
                When a thread has to do a complete (not too big) expression.
                The number of the expression to be done is in AR.exprtodo.
                The code is mostly taken from Processor. The only difference
                is with what to do with the output.
                The output should go to the scratch buffer of the worker
                (which is free at the right moment). If this buffer is too
                small we have a problem. We could write to file or give the
                master what we have and from now on the master has to collect
                pieces until things are complete.
                Note: this assumes that the expressions don't keep their order.
                If they have to keep their order, don't use this feature.
*/
            case DOONEEXPRESSION: {
 
                POSITION position, outposition;
                FILEHANDLE *fi, *fout, *oldoutfile;
                LONG dd = 0;
                WORD oldBracketOn = AR.BracketOn;
                WORD *oldBrackBuf = AT.BrackBuf;
                WORD oldbracketindexflag = AT.bracketindexflag;
                WORD fromspectator = 0;
                e = Expressions + AR.exprtodo;
                i = AR.exprtodo;
                AR.CurExpr = i;
                AR.SortType = AC.SortType;
                AR.expchanged = 0;
                if ( ( e->vflags & ISFACTORIZED ) != 0 ) {
                    AR.BracketOn = 1;
                    AT.BrackBuf = AM.BracketFactors;
                    AT.bracketindexflag = 1;
                }
 
                position = AS.OldOnFile[i];
                if ( e->status == HIDDENLEXPRESSION || e->status == HIDDENGEXPRESSION
                    || e->status == UNHIDELEXPRESSION || e->status == UNHIDEGEXPRESSION ) {
                    AR.GetFile = 2; fi = AR.hidefile;
                }
                else {
                    AR.GetFile = 0; fi = AR.infile;
                }
/*
                PUTZERO(fi->POposition);
                if ( fi->handle >= 0 ) {
                    fi->POfill = fi->POfull = fi->PObuffer;
                }
*/
                SetScratch(fi,&position);
                term = oldwork = AT.WorkPointer;
                AR.CompressPointer = AR.CompressBuffer;
                AR.CompressPointer[0] = 0;
                AR.KeptInHold = 0;
                if ( GetTerm(BHEAD term) <= 0 ) {
                    MLOCK(ErrorMessageLock);
                    MesPrint("Expression %d has problems in scratchfile (t)",i);
                    MUNLOCK(ErrorMessageLock);
                    Terminate(-1);
                }
                if ( AT.bracketindexflag > 0 ) OpenBracketIndex(i);
                term[3] = i;
                if ( term[5] < 0 ) {
                    fromspectator = -term[5];
                    PUTZERO(AM.SpectatorFiles[fromspectator-1].readpos);
                    term[5] = AC.cbufnum;
                }
                PUTZERO(outposition);
                fout = AR.outfile;
                fout->POfill = fout->POfull = fout->PObuffer;
                fout->POposition = outposition;
                if ( fout->handle >= 0 ) {
                    fout->POposition = outposition;
                }
/*
                The next statement is needed because we need the system
                to believe that the expression is at position zero for
                the moment. In this worker, with no memory of other expressions,
                it is. This is needed for when a bracket index is made
                because there e->onfile is an offset. Afterwards, when the
                expression is written to its final location in the masters
                output e->onfile will get its real value.
*/
                PUTZERO(e->onfile);
                if ( PutOut(BHEAD term,&outposition,fout,0) < 0 ) goto ProcErr;
 
                AR.DeferFlag = AC.ComDefer;
 
                AR.sLevel = AB[0]->R.sLevel;
                term = AT.WorkPointer;
                NewSort(BHEAD0);
                AR.MaxDum = AM.IndDum;
                AN.ninterms = 0;
                if ( fromspectator ) {
                 while ( GetFromSpectator(term,fromspectator-1) ) {
                  AT.WorkPointer = term + *term;
                  AN.RepPoint = AT.RepCount + 1;
                  AN.IndDum = AM.IndDum;
                  AR.CurDum = ReNumber(BHEAD term);
                  if ( AC.SymChangeFlag ) MarkDirty(term,DIRTYSYMFLAG);
                  if ( AN.ncmod ) {
                    if ( ( AC.modmode & ALSOFUNARGS ) != 0 ) MarkDirty(term,DIRTYFLAG);
                    else if ( AR.PolyFun ) PolyFunDirty(BHEAD term);
                  }
                  else if ( AC.PolyRatFunChanged ) PolyFunDirty(BHEAD term);
                  if ( ( AR.PolyFunType == 2 ) && ( AC.PolyRatFunChanged == 0 )
                        && ( e->status == LOCALEXPRESSION || e->status == GLOBALEXPRESSION ) ) {
                        PolyFunClean(BHEAD term);
                  }
                  if ( Generator(BHEAD term,0) ) {
                    LowerSortLevel(); goto ProcErr;
                  }
                 }
                }
                else {
                 while ( GetTerm(BHEAD term) ) {
                  SeekScratch(fi,&position);
                  AN.ninterms++; dd = AN.deferskipped;
                  if ( ( e->vflags & ISFACTORIZED ) != 0 && term[1] == HAAKJE ) {
                      StoreTerm(BHEAD term);
                  }
                  else {
                  if ( AC.CollectFun && *term <= (AM.MaxTer/(2*(LONG)sizeof(WORD))) ) {
                    if ( GetMoreTerms(term) < 0 ) {
                      LowerSortLevel(); goto ProcErr;
                    }
                    SeekScratch(fi,&position);
                  }
                  AT.WorkPointer = term + *term;
                  AN.RepPoint = AT.RepCount + 1;
                  if ( AR.DeferFlag ) {
                    AR.CurDum = AN.IndDum = Expressions[AR.exprtodo].numdummies;
                  }
                  else {
                    AN.IndDum = AM.IndDum;
                    AR.CurDum = ReNumber(BHEAD term);
                  }
                  if ( AC.SymChangeFlag ) MarkDirty(term,DIRTYSYMFLAG);
                  if ( AN.ncmod ) {
                    if ( ( AC.modmode & ALSOFUNARGS ) != 0 ) MarkDirty(term,DIRTYFLAG);
                    else if ( AR.PolyFun ) PolyFunDirty(BHEAD term);
                  }
                  else if ( AC.PolyRatFunChanged ) PolyFunDirty(BHEAD term);
                  if ( ( AR.PolyFunType == 2 ) && ( AC.PolyRatFunChanged == 0 )
                        && ( e->status == LOCALEXPRESSION || e->status == GLOBALEXPRESSION ) ) {
                        PolyFunClean(BHEAD term);
                  }
                  if ( Generator(BHEAD term,0) ) {
                    LowerSortLevel(); goto ProcErr;
                  }
                  AN.ninterms += dd;
                  }
                  SetScratch(fi,&position);
                  if ( fi == AR.hidefile ) {
                    AR.InHiBuf = (fi->POfull-fi->PObuffer)
                        -DIFBASE(position,fi->POposition)/sizeof(WORD);
                  }
                  else {
                    AR.InInBuf = (fi->POfull-fi->PObuffer)
                        -DIFBASE(position,fi->POposition)/sizeof(WORD);
                  }
                 }
                }
                AN.ninterms += dd;
                if ( EndSort(BHEAD AT.S0->sBuffer,0) < 0 ) goto ProcErr;
                e->numdummies = AR.MaxDum - AM.IndDum;
                AR.BracketOn = oldBracketOn;
                AT.BrackBuf = oldBrackBuf;
                if ( ( e->vflags & TOBEFACTORED ) != 0 )
                        poly_factorize_expression(e);
                else if ( ( ( e->vflags & TOBEUNFACTORED ) != 0 )
                 && ( ( e->vflags & ISFACTORIZED ) != 0 ) )
                        poly_unfactorize_expression(e);
                if ( AT.S0->TermsLeft )   e->vflags &= ~ISZERO;
                else                      e->vflags |= ISZERO;
                if ( AR.expchanged == 0 ) e->vflags |= ISUNMODIFIED;
                if ( AT.S0->TermsLeft ) AR.expflags |= ISZERO;
                if ( AR.expchanged )    AR.expflags |= ISUNMODIFIED;
                AR.GetFile = 0;
                AT.bracketindexflag = oldbracketindexflag;
/*
                Now copy the whole thing from fout to AR0.outfile
                Do this in one go to keep the lock occupied as short as possible
*/
                SeekScratch(fout,&outposition);
                LOCK(AS.outputslock);
                oldoutfile = AB[0]->R.outfile;
                if ( e->status == INTOHIDELEXPRESSION || e->status == INTOHIDEGEXPRESSION ) {
                    AB[0]->R.outfile = AB[0]->R.hidefile;
                }
                SeekScratch(AB[0]->R.outfile,&position);
                e->onfile = position;
                if ( CopyExpression(fout,AB[0]->R.outfile) < 0 ) {
                    AB[0]->R.outfile = oldoutfile;
                    UNLOCK(AS.outputslock);
                    MLOCK(ErrorMessageLock);
                    MesPrint("Error copying output of 'InParallel' expression to master. Thread: %d",identity);
                    MUNLOCK(ErrorMessageLock);
                    goto ProcErr;
                }
                AB[0]->R.outfile = oldoutfile;
                AB[0]->R.expflags = AR.expflags;
                UNLOCK(AS.outputslock);
 
                if ( fout->handle >= 0 ) {  /* Now get rid of the file */
                    CloseFile(fout->handle);
                    fout->handle = -1;
                    remove(fout->name);
                    PUTZERO(fout->POposition);
                    PUTZERO(fout->filesize);
                    fout->POfill = fout->POfull = fout->PObuffer;
                }
                UpdateMaxSize();
 
                AT.WorkPointer = oldwork;
 
                } break;
/*
            #] DOONEEXPRESSION : 
            #[ DOBRACKETS :
 
                In case we have a bracket index we can have the worker treat
                one or more of the entries in the bracket index.
                The advantage is that identical terms will meet each other
                sooner in the sorting and hence fewer compares will be needed.
                Also this way the master doesn't need to fill the buckets.
                The main problem is the load balancing which can become very
                bad when there is a long tail without things outside the bracket.
                
                We get sent:
                1: The number of the first bracket to be done
                2: The number of the last bracket to be done
*/
            case DOBRACKETS: {
                BRACKETINFO *binfo;
                BRACKETINDEX *bi;
                FILEHANDLE *fi;
                POSITION stoppos,where;
                e = Expressions + AR.CurExpr;
                binfo = e->bracketinfo;
                thr = AN.threadbuck;
                bi = &(binfo->indexbuffer[thr->firstbracket]);
                if ( AR.GetFile == 2 ) fi = AR.hidefile;
                else                   fi = AR.infile;
                where = bi->start;
                ADD2POS(where,AS.OldOnFile[AR.CurExpr]);
                SetScratch(fi,&(where));
                stoppos = binfo->indexbuffer[thr->lastbracket].next;
                ADD2POS(stoppos,AS.OldOnFile[AR.CurExpr]);
                AN.ninterms = thr->firstterm;
/*
                Now we have to put the 'value' of the bracket in the
                Compress buffer.
*/
                ttco = AR.CompressBuffer;
                tt = binfo->bracketbuffer + bi->bracket;
                i = *tt;
                NCOPY(ttco,tt,i)
                AR.CompressPointer = ttco;
                term = AT.WorkPointer;
                while ( GetTerm(BHEAD term) ) {
                    SeekScratch(fi,&where);
                    AT.WorkPointer = term + *term;
                    AN.IndDum = AM.IndDum;
                    AR.CurDum = ReNumber(BHEAD term);
                    if ( AC.SymChangeFlag ) MarkDirty(term,DIRTYSYMFLAG);
                    if ( AN.ncmod ) {
                        if ( ( AC.modmode & ALSOFUNARGS ) != 0 ) MarkDirty(term,DIRTYFLAG);
                        else if ( AR.PolyFun ) PolyFunDirty(BHEAD term);
                    }
                    else if ( AC.PolyRatFunChanged ) PolyFunDirty(BHEAD term);
                    if ( ( AR.PolyFunType == 2 ) && ( AC.PolyRatFunChanged == 0 )
                        && ( e->status == LOCALEXPRESSION || e->status == GLOBALEXPRESSION ) ) {
                        PolyFunClean(BHEAD term);
                    }
                    if ( ( AP.PreDebug & THREADSDEBUG ) != 0 ) {
                        MLOCK(ErrorMessageLock);
                        MesPrint("Thread %w executing term:");
                        PrintTerm(term,"DoBrackets");
                        MUNLOCK(ErrorMessageLock);
                    }
                    AT.WorkPointer = term + *term;
                    if ( Generator(BHEAD term,0) ) {
                        LowerSortLevel();
                        MLOCK(ErrorMessageLock);
                        MesPrint("Error in processing one term in thread %d in module %d",identity,AC.CModule);
                        MUNLOCK(ErrorMessageLock);
                        Terminate(-1);
                    }
                    AN.ninterms++;
                    SetScratch(fi,&(where));
                    if ( ISGEPOS(where,stoppos) ) break;
                }
                AT.WorkPointer = term;
                thr->free = BUCKETCOMINGFREE;
                break;
            }
/*
            #] DOBRACKETS : 
            #[ CLEARCLOCK :
 
            The program only comes here after a .clear
*/
            case CLEARCLOCK:
/*              LOCK(clearclocklock); */
                sumtimerinfo[identity] += TimeCPU(1);
                timerinfo[identity] = TimeCPU(0);
/*              UNLOCK(clearclocklock); */
                break;
/*
            #] CLEARCLOCK : 
            #[ MCTSEXPANDTREE :
*/
            case MCTSEXPANDTREE:
                AT.optimtimes = AB[0]->T.optimtimes;
                find_Horner_MCTS_expand_tree();
                break;
/*
            #] MCTSEXPANDTREE : 
            #[ OPTIMIZEEXPRESSION :
*/
            case OPTIMIZEEXPRESSION:
                optimize_expression_given_Horner();
                break;
/*
            #] OPTIMIZEEXPRESSION : 
*/
            default:
/* INTERNAL_ERROR_EXCL_START */
                MLOCK(ErrorMessageLock);
                MesPrint("!>Illegal wakeup signal %d for thread %d",wakeupsignal,identity);
                MUNLOCK(ErrorMessageLock);
                Terminate(-1);
                break;
/* INTERNAL_ERROR_EXCL_STOP */
        }
        /* we need the following update in case we are using checkpoints. then we
           need to readjust the clocks when recovering using this information */
        timerinfo[identity] = TimeCPU(1);
    }
EndOfThread:;
/*
    This is the end of the thread. We cleanup and exit.
    If we are using flint, call the per-thread cleanup function. This keep valgrind happy.
*/
#ifdef WITHFLINT
    flint_final_cleanup_thread();
#endif
    FinalizeOneThread(identity);
    return(0);
ProcErr:
    Terminate(-1);
    return(0);
}
 
/*
    #] RunThread : 
    #[ RunSortBot :
*/
void *RunSortBot(void *dummy)
{
    int identity, wakeupsignal, identityretv;
    ALLPRIVATES *B, *BB;
    DUMMYUSE(dummy);
    identity = SetIdentity(&identityretv);
    threadpointers[identity] = pthread_self();
    B = InitializeOneThread(identity);
    while ( ( wakeupsignal = SortBotWait(identity) ) > 0 ) {
        switch ( wakeupsignal ) {
/*
            #[ INISORTBOT :
*/
            case INISORTBOT:
                AR.CompressBuffer = AB[0]->R.CompressBuffer;
                AR.ComprTop = AB[0]->R.ComprTop;
                AR.CompressPointer = AB[0]->R.CompressPointer;
                AR.CurExpr = AB[0]->R.CurExpr;
                AR.PolyFun = AB[0]->R.PolyFun;
                AR.PolyFunInv = AB[0]->R.PolyFunInv;
                AR.PolyFunType = AB[0]->R.PolyFunType;
                AR.PolyFunExp = AB[0]->R.PolyFunExp;
                AR.PolyFunVar = AB[0]->R.PolyFunVar;
                AR.PolyFunPow = AB[0]->R.PolyFunPow;
                AR.SortType = AC.SortType;
                if ( AR.PolyFun == 0 ) { AT.SS->PolyFlag = 0; }
                else if ( AR.PolyFunType == 1 ) { AT.SS->PolyFlag = 1; }
                else if ( AR.PolyFunType == 2 ) {
                    if ( AR.PolyFunExp == 2
                      || AR.PolyFunExp == 3 ) AT.SS->PolyFlag = 1;
                    else                      AT.SS->PolyFlag = 2;
                }
                AT.SS->PolyWise = 0;
                AN.ncmod = AC.ncmod;
                LOCK(AT.SB.MasterBlockLock[1]);
                BB = AB[AT.SortBotIn1];
                LOCK(BB->T.SB.MasterBlockLock[BB->T.SB.MasterNumBlocks]);
                if ( AT.SortBotIn2 >= 0 ) {
                    BB = AB[AT.SortBotIn2];
                    LOCK(BB->T.SB.MasterBlockLock[BB->T.SB.MasterNumBlocks]);
                }
                AT.SB.FillBlock = 1;
                AT.SB.MasterFill[1] = AT.SB.MasterStart[1];
                SETBASEPOSITION(AN.theposition,0);
                // Reset the sortbot comparison count
                AT.SS->verbComparisons = 0;
                // Reset the maximal term size count
                AT.SS->verbMaxTermSize = 0;
                break;
/*
            #] INISORTBOT : 
            #[ RUNSORTBOT :
*/
            case RUNSORTBOT:
                SortBotMerge(B);
                break;
/*
            #] RUNSORTBOT : 
            #[ TERMINATETHREAD :
*/
            case TERMINATETHREAD:
                goto EndOfThread;
/*
            #] TERMINATETHREAD : 
            #[ CLEARCLOCK :
 
            The program only comes here after a .clear
*/
            case CLEARCLOCK:
/*              LOCK(clearclocklock); */
                sumtimerinfo[identity] += TimeCPU(1);
                timerinfo[identity] = TimeCPU(0);
/*              UNLOCK(clearclocklock); */
                break;
/*
            #] CLEARCLOCK : 
*/
            default:
/* INTERNAL_ERROR_EXCL_START */
                MLOCK(ErrorMessageLock);
                MesPrint("!>Illegal wakeup signal %d for thread %d",wakeupsignal,identity);
                MUNLOCK(ErrorMessageLock);
                Terminate(-1);
                break;
/* INTERNAL_ERROR_EXCL_STOP */
        }
    }
EndOfThread:;
/*
    This is the end of the thread. We cleanup and exit.
    If we are using flint, call the per-thread cleanup function. This keep valgrind happy.
*/
#ifdef WITHFLINT
    flint_final_cleanup_thread();
#endif
    FinalizeOneThread(identity);
    return(0);
}
 
 
/*
    #] RunSortBot : 
    #[ IAmAvailable :
*/
void IAmAvailable(int identity)
{
    int top;
    LOCK(availabilitylock);
    top = topofavailables;
    listofavailables[topofavailables++] = identity;
    if ( top == 0 ) {
        UNLOCK(availabilitylock);
        LOCK(wakeupmasterlock);
        wakeupmaster = identity;
        pthread_cond_signal(&wakeupmasterconditions);
        UNLOCK(wakeupmasterlock);
    }
    else {
        UNLOCK(availabilitylock);
    }
}
 
/*
    #] IAmAvailable : 
    #[ GetAvailableThread :
*/
int GetAvailableThread(void)
{
    int retval = -1;
    LOCK(availabilitylock);
    if ( topofavailables > 0 ) retval = listofavailables[--topofavailables];
    UNLOCK(availabilitylock);
    if ( retval >= 0 ) {
/*
        Make sure the thread is indeed waiting and not between
        saying that it is available and starting to wait.
*/
        LOCK(wakeuplocks[retval]);
        UNLOCK(wakeuplocks[retval]);
    }
    return(retval);
}
 
/*
    #] GetAvailableThread : 
    #[ ConditionalGetAvailableThread :
*/
int ConditionalGetAvailableThread(void)
{
    int retval = -1;
    if ( topofavailables > 0 ) {
        LOCK(availabilitylock);
        if ( topofavailables > 0 ) {
            retval = listofavailables[--topofavailables];
        }
        UNLOCK(availabilitylock);
        if ( retval >= 0 ) {
/*
            Make sure the thread is indeed waiting and not between
            saying that it is available and starting to wait.
*/
            LOCK(wakeuplocks[retval]);
            UNLOCK(wakeuplocks[retval]);
        }
    }
    return(retval);
}
 
/*
    #] ConditionalGetAvailableThread : 
    #[ GetThread :
*/
int GetThread(int identity)
{
    int retval = -1, j;
    LOCK(availabilitylock);
    for ( j = 0; j < topofavailables; j++ ) {
        if ( identity == listofavailables[j] ) break;
    }
    if ( j < topofavailables ) {
        --topofavailables;
        for ( ; j < topofavailables; j++ ) {
            listofavailables[j] = listofavailables[j+1];
        }
        retval = identity;
    }
    UNLOCK(availabilitylock);
    return(retval);
}
 
/*
    #] GetThread : 
    #[ ThreadWait :
*/
int ThreadWait(int identity)
{
    int retval, top, j;
    LOCK(wakeuplocks[identity]);
    LOCK(availabilitylock);
    top = topofavailables;
    for ( j = topofavailables; j > 0; j-- )
        listofavailables[j] = listofavailables[j-1];
    listofavailables[0] = identity;
    topofavailables++;
    if ( top == 0 || topofavailables == numberofworkers ) {
        UNLOCK(availabilitylock);
        LOCK(wakeupmasterlock);
        wakeupmaster = identity;
        pthread_cond_signal(&wakeupmasterconditions);
        UNLOCK(wakeupmasterlock);
    }
    else {
        UNLOCK(availabilitylock);
    }
    while ( wakeup[identity] == 0 ) {
        pthread_cond_wait(&(wakeupconditions[identity]),&(wakeuplocks[identity]));
    }
    retval = wakeup[identity];
    wakeup[identity] = 0;
    UNLOCK(wakeuplocks[identity]);
    return(retval);
}
 
/*
    #] ThreadWait : 
    #[ SortBotWait :
*/
 
int SortBotWait(int identity)
{
    int retval;
    LOCK(wakeuplocks[identity]);
    LOCK(availabilitylock);
    topsortbotavailables++;
    if ( topsortbotavailables >= GetNumberOfSortBots(numberofworkers) ) {
        UNLOCK(availabilitylock);
        LOCK(wakeupsortbotlock);
        wakeupmaster = identity;
        pthread_cond_signal(&wakeupsortbotconditions);
        UNLOCK(wakeupsortbotlock);
    }
    else {
        UNLOCK(availabilitylock);
    }
    while ( wakeup[identity] == 0 ) {
        pthread_cond_wait(&(wakeupconditions[identity]),&(wakeuplocks[identity]));
    }
    retval = wakeup[identity];
    wakeup[identity] = 0;
    UNLOCK(wakeuplocks[identity]);
    return(retval);
}
 
 
/*
    #] SortBotWait : 
    #[ ThreadClaimedBlock :
*/
int ThreadClaimedBlock(int identity)
{
    LOCK(availabilitylock);
    numberclaimed++;    
    if ( numberclaimed >= numberofworkers ) {
        UNLOCK(availabilitylock);
        LOCK(wakeupmasterlock);
        wakeupmaster = identity;
        pthread_cond_signal(&wakeupmasterconditions);
        UNLOCK(wakeupmasterlock);
    }
    else {
        UNLOCK(availabilitylock);
    }
    return(0);
}
 
/*
    #] ThreadClaimedBlock : 
    #[ MasterWait :
*/
int MasterWait(void)
{
    int retval;
    LOCK(wakeupmasterlock);
    while ( wakeupmaster == 0 ) {
        pthread_cond_wait(&wakeupmasterconditions,&wakeupmasterlock);
    }
    retval = wakeupmaster;
    wakeupmaster = 0;
    UNLOCK(wakeupmasterlock);
    return(retval);
}
 
/*
    #] MasterWait : 
    #[ MasterWaitThread :
*/
int MasterWaitThread(int identity)
{
    int retval;
    LOCK(wakeupmasterthreadlocks[identity]);
    while ( wakeupmasterthread[identity] == 0 ) {
        pthread_cond_wait(&(wakeupmasterthreadconditions[identity])
                ,&(wakeupmasterthreadlocks[identity]));
    }
    retval = wakeupmasterthread[identity];
    wakeupmasterthread[identity] = 0;
    UNLOCK(wakeupmasterthreadlocks[identity]);
    return(retval);
}
 
/*
    #] MasterWaitThread : 
    #[ MasterWaitAll :
*/
void MasterWaitAll(void)
{
    LOCK(wakeupmasterlock);
    while ( topofavailables < numberofworkers ) {
        pthread_cond_wait(&wakeupmasterconditions,&wakeupmasterlock);
    }
    UNLOCK(wakeupmasterlock);
    return;
}
 
/*
    #] MasterWaitAll : 
    #[ MasterWaitAllSortBots :
*/
 
 
void MasterWaitAllSortBots(void)
{
    LOCK(wakeupsortbotlock);
    while ( topsortbotavailables < GetNumberOfSortBots(numberofworkers) ) {
        pthread_cond_wait(&wakeupsortbotconditions,&wakeupsortbotlock);
    }
    UNLOCK(wakeupsortbotlock);
    return;
}
 
 
/*
    #] MasterWaitAllSortBots : 
    #[ MasterWaitAllBlocks :
*/
void MasterWaitAllBlocks(void)
{
    LOCK(wakeupmasterlock);
    while ( numberclaimed < numberofworkers ) {
        pthread_cond_wait(&wakeupmasterconditions,&wakeupmasterlock);
    }
    UNLOCK(wakeupmasterlock);
    return;
}
 
/*
    #] MasterWaitAllBlocks : 
    #[ WakeupThread :
*/
void WakeupThread(int identity, int signalnumber)
{
    if ( signalnumber == 0 ) {
/* INTERNAL_ERROR_EXCL_START */
        MLOCK(ErrorMessageLock);
        MesPrint("!>Illegal wakeup signal for thread %d",identity);
        MUNLOCK(ErrorMessageLock);
        Terminate(-1);
/* INTERNAL_ERROR_EXCL_STOP */
    }
    LOCK(wakeuplocks[identity]);
    wakeup[identity] = signalnumber;
    pthread_cond_signal(&(wakeupconditions[identity]));
    UNLOCK(wakeuplocks[identity]);
}
 
/*
    #] WakeupThread : 
    #[ WakeupMasterFromThread :
*/
void WakeupMasterFromThread(int identity, int signalnumber)
{
    if ( signalnumber == 0 ) {
/* INTERNAL_ERROR_EXCL_START */
        MLOCK(ErrorMessageLock);
        MesPrint("!>Illegal wakeup signal for master %d",identity);
        MUNLOCK(ErrorMessageLock);
        Terminate(-1);
/* INTERNAL_ERROR_EXCL_STOP */
    }
    LOCK(wakeupmasterthreadlocks[identity]);
    wakeupmasterthread[identity] = signalnumber;
    pthread_cond_signal(&(wakeupmasterthreadconditions[identity]));
    UNLOCK(wakeupmasterthreadlocks[identity]);
}
 
/*
    #] WakeupMasterFromThread : 
    #[ SendOneBucket :
*/
int SendOneBucket(int type)
{
    ALLPRIVATES *B0 = AB[0];
    THREADBUCKET *thr = 0;
    int j, k, id;
    for ( j = 0; j < numthreadbuckets; j++ ) {
        if ( threadbuckets[j]->free == BUCKETFILLED ) {
            thr = threadbuckets[j];
            for ( k = j+1; k < numthreadbuckets; k++ )
                threadbuckets[k-1] = threadbuckets[k];
            threadbuckets[numthreadbuckets-1] = thr;
            break;
        }
    }
    AN0.ninterms++;
    while ( ( id = GetAvailableThread() ) < 0 ) { MasterWait(); }
/*
    Prepare the thread. Give it the term and variables.
*/
    LoadOneThread(0,id,thr,0);
    thr->busy = BUCKETASSIGNED;
    thr->free = BUCKETINUSE;
    numberoffullbuckets--;
/*
    And signal the thread to run.
    Form now on we may only interfere with this bucket
    1: after it has been marked BUCKETCOMINGFREE
    2: when thr->busy == BUCKETDOINGTERM and then only when protected by
       thr->lock. This would be for load balancing.
*/
    WakeupThread(id,type);
/*  AN0.ninterms += thr->ddterms; */
    return(0);
}
 
/*
    #] SendOneBucket : 
    #[ InParallelProcessor :
*/
int InParallelProcessor(void)
{
    GETIDENTITY
    int i, id, retval = 0, num = 0;
    EXPRESSIONS e;
    if ( numberofworkers >= 2 ) {
        SetWorkerFiles();
        for ( i = 0; i < NumExpressions; i++ ) {
            e = Expressions+i;
            if ( e->partodo <= 0 ) continue;
            if ( e->status == LOCALEXPRESSION || e->status == GLOBALEXPRESSION
            || e->status == UNHIDELEXPRESSION || e->status == UNHIDEGEXPRESSION
            || e->status == INTOHIDELEXPRESSION || e->status == INTOHIDEGEXPRESSION ) {
            }
            else {
                e->partodo = 0;
                continue;
            }
            if ( e->counter == 0 ) { /* Expression with zero terms */
                e->partodo = 0;
                continue;
            }
/*
            This expression should go to an idle worker
*/
            while ( ( id = GetAvailableThread() ) < 0 ) { MasterWait(); }
            LoadOneThread(0,id,0,-1);
            AB[id]->R.exprtodo = i;
            WakeupThread(id,DOONEEXPRESSION);
            num++;
        }
/*
        Now we have to wait for all workers to finish
*/
        if ( num > 0 ) MasterWaitAll();
 
        if ( AC.CollectFun ) AR.DeferFlag = 0;
    }
    else {
        for ( i = 0; i < NumExpressions; i++ ) {
            Expressions[i].partodo = 0;
        }
    }
    return(retval);
}
 
/*
    #] InParallelProcessor : 
    #[ ThreadsProcessor :
*/
int ThreadsProcessor(EXPRESSIONS e, WORD LastExpression, WORD fromspectator)
{
    ALLPRIVATES *B0 = AB[0], *B = B0;
    int id, oldgzipCompress, endofinput = 0, j, still, k, defcount = 0, bra = 0, first = 1;
    LONG dd = 0, ddd, thrbufsiz, thrbufsiz0, thrbufsiz2, numbucket = 0, numpasses;
    LONG num, i;
    WORD *oldworkpointer = AT0.WorkPointer, *tt, *ttco = 0, *t1 = 0, ter, *tstop = 0, *t2;
    THREADBUCKET *thr = 0;
    FILEHANDLE *oldoutfile = AR0.outfile;
    GETTERM GetTermP = &GetTerm;
    POSITION eonfile = AS.OldOnFile[e-Expressions];
    numberoffullbuckets = 0;
/*
    Start up all threads. The lock needs to be around the whole loop
    to keep processes from terminating quickly and putting themselves
    in the list of available threads again.
*/
    AM.tracebackflag = 1;
 
    AS.sLevel = AR0.sLevel;
    LOCK(availabilitylock);
    topofavailables = 0;
    for ( id = 1; id <= numberofworkers; id++ ) {
        WakeupThread(id,STARTNEWEXPRESSION);
    }
    UNLOCK(availabilitylock);
    NewSort(BHEAD0);
    AN0.ninterms = 1;
/*
    Now for redefine
*/
    if ( AC.numpfirstnum > 0 ) {
        for ( j = 0; j < AC.numpfirstnum; j++ ) {
            AC.inputnumbers[j] = -1;
        }
    }
    MasterWaitAll();
/*
    Determine a reasonable bucketsize.
    This is based on the value of AC.ThreadBucketSize and the number
    of terms. We want at least 5 buckets per worker at the moment.
    Some research should show whether this is reasonable.
 
    The number of terms in the expression is in e->counter
*/
    thrbufsiz2 = thrbufsiz = AC.ThreadBucketSize-1;
    /* CopySpectator expressions have a one-term expression prototype in the
     * scratch file, regardless of how many terms the spectator contains. Don't
     * use this prototype to determine bucket size, or we'll end up with single-term
     * buckets and bad performance. Just use thrbufsiz in this case. */
    if ( ! fromspectator && ( e->counter / ( numberofworkers * 5 ) ) < thrbufsiz ) {
        thrbufsiz = e->counter / ( numberofworkers * 5 ) - 1;
        if ( thrbufsiz < 0 ) thrbufsiz = 0;
    }
    thrbufsiz0 = thrbufsiz;
    numpasses = 5; /* this is just for trying */
    thrbufsiz = thrbufsiz0 / (2 << numpasses);
/*
    Mark all buckets as free and take the first.
*/
    for ( j = 0; j < numthreadbuckets; j++ )
        threadbuckets[j]->free = BUCKETFREE;
    thr = threadbuckets[0];
/*
    #[ Whole brackets :
 
    First we look whether we have to work with entire brackets
    This is the case when there is a non-NULL address in e->bracketinfo.
    Of course we shouldn't have interference from a collect or keep statement.
*/
#ifdef WHOLEBRACKETS
    if ( e->bracketinfo && AC.CollectFun == 0 && AR0.DeferFlag == 0 ) {
        FILEHANDLE *curfile;
        int didone = 0;
        LONG num, n;
        AN0.expr = e;
        for ( n = 0; n < e->bracketinfo->indexfill; n++ ) {
            num = TreatIndexEntry(B0,n);
            if ( num > 0 ) {
                didone = 1;
/*
                This bracket can be sent off.
                1: Look for an empty bucket
*/
ReTry:;
                for ( j = 0; j < numthreadbuckets; j++ ) {
                    switch ( threadbuckets[j]->free ) {
                        case BUCKETFREE:
                            thr = threadbuckets[j];
                            goto Found1;
                        case BUCKETCOMINGFREE:
                            thr = threadbuckets[j];
                            thr->free = BUCKETFREE;
                            for ( k = j+1; k < numthreadbuckets; k++ )
                                threadbuckets[k-1] = threadbuckets[k];
                            threadbuckets[numthreadbuckets-1] = thr;
                            j--;
                            break;
                        default:
                            break;
                    }
                }
Found1:;
                if ( j < numthreadbuckets ) {
/*
                    Found an empty bucket. Fill it.
*/
                    thr->firstbracket = n;
                    thr->lastbracket = n + num - 1;
                    thr->type = BUCKETDOINGBRACKET;
                    thr->free = BUCKETFILLED;
                    thr->firstterm = AN0.ninterms;
                    for ( j = n; j < n+num; j++ ) {
                        AN0.ninterms += e->bracketinfo->indexbuffer[j].termsinbracket;
                    }
                    n += num-1;
                    numberoffullbuckets++;
                    if ( topofavailables > 0 ) {
                        SendOneBucket(DOBRACKETS);
                    }
                }
/*
                    All buckets are in use.
                    Look/wait for an idle worker. Give it a bucket.
                    After that, retry for a bucket
*/
                else {
                    while ( topofavailables <= 0 ) {
                        MasterWait();
                    }
                    SendOneBucket(DOBRACKETS);
                    goto ReTry;
                }
            }
        }
        if ( didone ) {
/*
            And now put the input back in the original position.
*/
            switch ( e->status ) {
                case UNHIDELEXPRESSION:
                case UNHIDEGEXPRESSION:
                case DROPHLEXPRESSION:
                case DROPHGEXPRESSION:
                case HIDDENLEXPRESSION:
                case HIDDENGEXPRESSION:
                    curfile = AR0.hidefile;
                    break;
                default:
                    curfile = AR0.infile;
                    break;
            }
            SetScratch(curfile,&eonfile);
            GetTerm(B0,AT0.WorkPointer);
/*
            Now we point the GetTerm that is used to the one that is selective
*/
            GetTermP = &GetTerm2;
/*
            Next wait till there is a bucket available and initialize thr to it.
*/
            for(;;) {
                for ( j = 0; j < numthreadbuckets; j++ ) {
                    switch ( threadbuckets[j]->free ) {
                        case BUCKETFREE:
                            thr = threadbuckets[j];
                            goto Found2;
                        case BUCKETCOMINGFREE:
                            thr = threadbuckets[j];
                            thr->free = BUCKETFREE;
                            for ( k = j+1; k < numthreadbuckets; k++ )
                                threadbuckets[k-1] = threadbuckets[k];
                            threadbuckets[numthreadbuckets-1] = thr;
                            j--;
                            break;
                        default:
                            break;
                    }
                }
                while ( topofavailables <= 0 ) {
                    MasterWait();
                }
                while ( topofavailables > 0 && numberoffullbuckets > 0 ) {
                    SendOneBucket(DOBRACKETS);
                }
            }
Found2:;
            while ( numberoffullbuckets > 0 ) {
                while ( topofavailables <= 0 ) {
                    MasterWait();
                }
                while ( topofavailables > 0 && numberoffullbuckets > 0 ) {
                    SendOneBucket(DOBRACKETS);
                }
            }
/*
            Disable the 'warming up' with smaller buckets.
 
            numpasses = 0;
            thrbufsiz = thrbufsiz0;
*/
            AN0.lastinindex = -1;
        }
        MasterWaitAll();
    }
#endif
/*
    #] Whole brackets : 
 
    Now the loop to start a bucket
*/
    for(;;) {
        if ( fromspectator ) {
            ter = GetFromSpectator(thr->threadbuffer,fromspectator-1);
            if ( ter == 0 ) fromspectator = 0;
        }
        else {
            ter = GetTermP(B0,thr->threadbuffer);
/*
            At this point we could check whether the input term is
            just an expression that resides in a scratch file.
            If this is the case we should store the current input info
            (file and buffer content) and redirect the input.
            At the end we can go back to where we were.
            There are two possibilities: we are in the same scratchfile
            as the main input and the other expression is still in the
            input buffer, or we have to do some reading. The reading is
            of course done in GetTerm.
 
            How to set this up can also be studied in TestSub (in file proces.c)
            where it checks for EXPRESSION.
*/
        }
        if ( ter < 0 ) break;
        if ( ter == 0 ) { endofinput = 1; goto Finalize; }
        dd = AN0.deferskipped;
        if ( AR0.DeferFlag ) {
            defcount = 0;
            thr->deferbuffer[defcount++] = AR0.DefPosition;
            ttco = thr->compressbuffer; t1 = AR0.CompressBuffer; j = *t1;
            while ( thr->compressbuffersize <= j ) {
                // the compressbuffer is not large enough!
                WORD *top = thr->compressbuffer+thr->compressbuffersize;
                DoubleBuffer((void**)&(thr->compressbuffer),(void**)&(top),
                    sizeof(*(thr->compressbuffer)), "double compressbuffer");
                ttco = thr->compressbuffer;
                thr->compressbuffersize *= 2;
            }
            NCOPY(ttco,t1,j);
        }
        else if ( first && ( AC.CollectFun == 0 ) ) { /* Brackets ? */
            first = 0;
            t1 = tstop = thr->threadbuffer;
            tstop += *tstop; tstop -= ABS(tstop[-1]);
            t1++;
            while ( t1 < tstop ) {
                if ( t1[0] == HAAKJE ) { bra = 1; break; }
                t1 += t1[1];
            }
            t1 = thr->threadbuffer;
        }
/*
        Check whether we have a collect,function going. If so execute it.
*/
        if ( AC.CollectFun && *(thr->threadbuffer) < (AM.MaxTer/((LONG)sizeof(WORD))-10) ) {
            if ( ( dd = GetMoreTerms(thr->threadbuffer) ) < 0 ) {
                LowerSortLevel(); goto ProcErr;
            }
        }
/*
        Check whether we have a priority task:
*/
        if ( topofavailables > 0 && numberoffullbuckets > 0 ) SendOneBucket(LOWESTLEVELGENERATION);
/*
        Now put more terms in the bucket. Position tt after the first term
*/
        tt = thr->threadbuffer; tt += *tt;
        thr->totnum = 1;
        thr->usenum = 0;
/*
        Next we worry about the 'slow startup' in which we make the initial
        buckets smaller, so that we get all threads busy as soon as possible.
*/
        if ( numpasses > 0 ) {
            numbucket++;
            if ( numbucket >= numberofworkers ) {
                numbucket = 0;
                numpasses--;
                if ( numpasses == 0 ) thrbufsiz = thrbufsiz0;
                else                  thrbufsiz = thrbufsiz0 / (2 << numpasses);
            }
            thrbufsiz2 = thrbufsiz + thrbufsiz/5; /* for completing brackets */
        }
/*
        we have already 1+dd terms
*/
        while ( ( dd < thrbufsiz ) &&
            ( tt - thr->threadbuffer ) < ( thr->threadbuffersize - AM.MaxTer/((LONG)sizeof(WORD)) - 2 ) ) {
/*
            First check:
*/
            if ( topofavailables > 0 && numberoffullbuckets > 0 ) SendOneBucket(LOWESTLEVELGENERATION);
/*
            There is room in the bucket. Fill yet another term.
*/
            if ( ( fromspectator
                ? GetFromSpectator(tt,fromspectator-1)
                : GetTermP(B0,tt) ) == 0 ) { endofinput = 1; break; }
            dd++;
            thr->totnum++;
            dd += AN0.deferskipped;
            if ( AR0.DeferFlag ) {
                thr->deferbuffer[defcount++] = AR0.DefPosition;
                t1 = AR0.CompressBuffer; j = *t1;
                while ( thr->compressbuffer+thr->compressbuffersize-ttco <= j ) {
                    // the compressbuffer is not large enough!
                    const ptrdiff_t oldoffset = ttco - thr->compressbuffer;
                    WORD *top = thr->compressbuffer+thr->compressbuffersize;
                    DoubleBuffer((void**)&(thr->compressbuffer),(void**)&(top),
                        sizeof(*(thr->compressbuffer)), "double compressbuffer");
                    ttco = thr->compressbuffer + oldoffset;
                    thr->compressbuffersize *= 2;
                }
                NCOPY(ttco,t1,j);
            }
            if ( AC.CollectFun && *tt < (AM.MaxTer/((LONG)sizeof(WORD))-10) ) {
                if ( ( ddd = GetMoreTerms(tt) ) < 0 ) {
                    LowerSortLevel(); goto ProcErr;
                }
                dd += ddd;
            }
            t1 = tt;
            tt += *tt;
        }
/*
        Check whether there are regular brackets and if we have no DeferFlag
        and no collect, we try to add more terms till we finish the current
        bracket. We should however not overdo it. Let us say: up to 20%
        more terms are allowed.
*/
        if ( bra ) {
            tstop = t1 + *t1; tstop -= ABS(tstop[-1]);
            t2 = t1+1;
            while ( t2 < tstop ) {
                if ( t2[0] == HAAKJE ) { break; }
                t2 += t2[1];
            }
            if ( t2[0] == HAAKJE ) {
              t2 += t2[1]; num = t2 - t1;
              while ( ( dd < thrbufsiz2 ) &&
                ( tt - thr->threadbuffer ) < ( thr->threadbuffersize - AM.MaxTer - 2 ) ) {
/*
                First check:
*/
                if ( topofavailables > 0 && numberoffullbuckets > 0 ) SendOneBucket(LOWESTLEVELGENERATION);
/*
                There is room in the bucket. Fill yet another term.
*/
                if ( ( fromspectator
                    ? GetFromSpectator(tt,fromspectator-1)
                    : GetTermP(B0,tt) ) == 0 ) { endofinput = 1; break; }
/*
                Same bracket?
*/
                tstop = tt + *tt; tstop -= ABS(tstop[-1]);
                if ( tstop-tt < num ) { /* Different: abort */
                    AR0.KeptInHold = 1;
                    break;
                }
                for ( i = 1; i < num; i++ ) {
                    if ( t1[i] != tt[i] ) break;
                }
                if ( i < num ) { /* Different: abort */
                    AR0.KeptInHold = 1;
                    break;
                }
/*
                Same bracket. We need this term.
*/
                dd++;
                thr->totnum++;
                tt += *tt;
              }
            }
        }
        thr->ddterms = dd; /* total number of terms including keep brackets */
        thr->firstterm = AN0.ninterms;
        AN0.ninterms += dd;
        *tt = 0;           /* mark end of bucket */
        thr->free = BUCKETFILLED;
        thr->type = BUCKETDOINGTERMS;
        numberoffullbuckets++;
        if ( topofavailables <= 0 && endofinput == 0 ) {
/*
            Problem: topofavailables may already be > 0, but the
            thread has not yet gone into waiting. Can the signal get lost?
            How can we tell that a thread is waiting for a signal?
 
            All threads are busy. Try to load up another bucket.
            In the future we could be more sophisticated.
            At the moment we load a complete bucket which could be
            1000 terms or even more.
            In principle it is better to keep a full bucket ready
            and check after each term we put in the next bucket. That
            way we don't waste time of the workers.
*/
            for ( j = 0; j < numthreadbuckets; j++ ) {
                switch ( threadbuckets[j]->free ) {
                    case BUCKETFREE:
                        thr = threadbuckets[j];
                        if ( !endofinput ) goto NextBucket;
/*
                        If we are at the end of the input we mark
                        the free buckets in a special way. That way
                        we don't keep running into them.
*/
                        thr->free = BUCKETATEND;
                        break;
                    case BUCKETCOMINGFREE:
                        thr = threadbuckets[j];
                        thr->free = BUCKETFREE;
/*
                        Bucket has just been finished.
                        Put at the end of the list. We don't want
                        an early bucket to wait to be treated last.
*/
                        for ( k = j+1; k < numthreadbuckets; k++ )
                            threadbuckets[k-1] = threadbuckets[k];
                        threadbuckets[numthreadbuckets-1] = thr;
                        j--;   /* we have to redo the same number j. */
                        break;
                    default:
                        break;
                }
            }
/*
            We have no free bucket or we are at the end.
            The only thing we can do now is wait for a worker to come free,
            provided there are still buckets to send.
*/
        }
/*
        Look for the next bucket to send. There is at least one full bucket!
*/
        for ( j = 0; j < numthreadbuckets; j++ ) {
            if ( threadbuckets[j]->free == BUCKETFILLED ) {
                thr = threadbuckets[j];
                for ( k = j+1; k < numthreadbuckets; k++ )
                    threadbuckets[k-1] = threadbuckets[k];
                threadbuckets[numthreadbuckets-1] = thr;
                break;
            }
        }
/*
        Wait for a thread to become available
        The bucket we are going to use is in thr.
*/
DoBucket:;
        AN0.ninterms++;
        while ( ( id = GetAvailableThread() ) < 0 ) { MasterWait(); }
/*
        Prepare the thread. Give it the term and variables.
*/
        LoadOneThread(0,id,thr,0);
        LOCK(thr->lock);
        thr->busy = BUCKETASSIGNED;
        UNLOCK(thr->lock);
        thr->free = BUCKETINUSE;
        numberoffullbuckets--;
/*
        And signal the thread to run.
        Form now on we may only interfere with this bucket
        1: after it has been marked BUCKETCOMINGFREE
        2: when thr->busy == BUCKETDOINGTERM and then only when protected by
           thr->lock. This would be for load balancing.
*/
        WakeupThread(id,LOWESTLEVELGENERATION);
/*      AN0.ninterms += thr->ddterms; */
/*
        Now look whether there is another bucket filled and a worker available
*/
        if ( topofavailables > 0 ) {  /* there is a worker */
            for ( j = 0; j < numthreadbuckets; j++ ) {
                if ( threadbuckets[j]->free == BUCKETFILLED ) {
                    thr = threadbuckets[j];
                    for ( k = j+1; k < numthreadbuckets; k++ )
                        threadbuckets[k-1] = threadbuckets[k];
                    threadbuckets[numthreadbuckets-1] = thr;
                    goto DoBucket; /* and we found a bucket */
                }
            }
/*
            no bucket is loaded but there is a thread available
            find a bucket to load. If there is none (all are USED or ATEND)
            we jump out of the loop.
*/
            for ( j = 0; j < numthreadbuckets; j++ ) {
                switch ( threadbuckets[j]->free ) {
                    case BUCKETFREE:
                        thr = threadbuckets[j];
                        if ( !endofinput ) goto NextBucket;
                        thr->free = BUCKETATEND;
                        break;
                    case BUCKETCOMINGFREE:
                        thr = threadbuckets[j];
                        if ( endofinput ) {
                            thr->free = BUCKETATEND;
                        }
                        else {
                            thr->free = BUCKETFREE;
                            for ( k = j+1; k < numthreadbuckets; k++ )
                                threadbuckets[k-1] = threadbuckets[k];
                            threadbuckets[numthreadbuckets-1] = thr;
                            j--;
                        }
                        break;
                    default:
                        break;
                }
            }
            if ( j >= numthreadbuckets ) break;
        }
        else {
/*
            No worker available.
            Look for a bucket to load.
            Its number will be in "still"
*/
Finalize:;
            still = -1;
            for ( j = 0; j < numthreadbuckets; j++ ) {
                switch ( threadbuckets[j]->free ) {
                    case BUCKETFREE:
                        thr = threadbuckets[j];
                        if ( !endofinput ) goto NextBucket;
                        thr->free = BUCKETATEND;
                        break;
                    case BUCKETCOMINGFREE:
                        thr = threadbuckets[j];
                        if ( endofinput ) thr->free = BUCKETATEND;
                        else {
                            thr->free = BUCKETFREE;
                            for ( k = j+1; k < numthreadbuckets; k++ )
                                threadbuckets[k-1] = threadbuckets[k];
                            threadbuckets[numthreadbuckets-1] = thr;
                            j--;
                        }
                        break;
                    case BUCKETFILLED:
                        if ( still < 0 ) still = j;
                        break;
                    default:
                        break;
                }
            }
            if ( still < 0 ) {
/*
                No buckets to be executed and no buckets FREE.
                We must be at the end. Break out of the loop.
*/
                break;
            }
            thr = threadbuckets[still];
            for ( k = still+1; k < numthreadbuckets; k++ )
                threadbuckets[k-1] = threadbuckets[k];
            threadbuckets[numthreadbuckets-1] = thr;
            goto DoBucket;
        }
NextBucket:;
    }
/*
    Now the stage one load balancing.
    If the load has been readjusted we have again filled buckets.
    In that case we jump back in the loop.
 
    Tricky point: when do the workers see the new value of AT.LoadBalancing?
    It should activate the locks on thr->busy
*/
    if ( AC.ThreadBalancing ) {
        for ( id = 1; id <= numberofworkers; id++ ) {
            AB[id]->T.LoadBalancing = 1;
        }
        if ( LoadReadjusted() ) goto Finalize;
        for ( id = 1; id <= numberofworkers; id++ ) {
            AB[id]->T.LoadBalancing = 0;
        }
    }
    if ( AC.ThreadBalancing ) {
/*
        The AS.Balancing flag should have Generator look for
        free workers and apply the "buro" method.
 
        There is still a serious problem.
        When for instance a sum_, there may be space created in a local
        compiler buffer for a wildcard substitution or whatever.
        Compiler buffer execution scribble space.....
        This isn't copied along?
        Look up ebufnum. There are 12 places with AddRHS!
        Problem: one process allocates in ebuf. Then term is given to
        other process. It would like to use from this ebuf, but the sender
        finishes first and removes the ebuf (and/or overwrites it).
 
        Other problem: local $ variables aren't copied along.
*/
        AS.Balancing = 0;
    }
    MasterWaitAll();
    AS.Balancing = 0;
/*
    When we deal with the last expression we can now remove the input
    scratch file. This saves potentially much disk space (up to 1/3)
*/
    if ( LastExpression ) {
        UpdateMaxSize();
        if ( AR0.infile->handle >= 0 ) {
            CloseFile(AR0.infile->handle);
            AR0.infile->handle = -1;
            remove(AR0.infile->name);
            PUTZERO(AR0.infile->POposition);
            AR0.infile->POfill = AR0.infile->POfull = AR0.infile->PObuffer;
        }
    }
/*
    We order the threads to finish in the SortBotMasterMerge routine
    It will start with waiting for all threads to finish.
    One could make an administration in which threads that have
    finished can start already with the final sort but
    1: The load balancing should not make this super urgent
    2: It would definitely not be very compatible with the second
       stage load balancing.
*/
    oldgzipCompress = AR0.gzipCompress;
    AR0.gzipCompress = 0;
    if ( AR0.outtohide ) AR0.outfile = AR0.hidefile;
    if ( SortBotMasterMerge() < 0 ) {
        if ( AR0.outtohide ) AR0.outfile = oldoutfile;
        AR0.gzipCompress = oldgzipCompress;
        goto ProcErr;
    }
    if ( AR0.outtohide ) AR0.outfile = oldoutfile;
    AR0.gzipCompress = oldgzipCompress;
/*
    Now wait for all threads to be ready to give them the cleaning up signal.
    With the new SortBotMasterMerge routine we can do the cleanup already automatically
    avoiding having to send these signals.
*/
    MasterWaitAll();
    AR0.sLevel--;
    for ( id = 1; id < AM.totalnumberofthreads; id++ ) {
        if ( GetThread(id) > 0 ) WakeupThread(id,CLEANUPEXPRESSION);
    }
    e->numdummies = 0;
    for ( id = 1; id < AM.totalnumberofthreads; id++ ) {
        if ( AB[id]->R.MaxDum - AM.IndDum > e->numdummies )
            e->numdummies = AB[id]->R.MaxDum - AM.IndDum;
        AR0.expchanged |= AB[id]->R.expchanged;
    }
/*
    And wait for all to be clean.
*/
    MasterWaitAll();
    AT0.WorkPointer = oldworkpointer;
    return(0);
ProcErr:;
    return(-1);
}
 
/*
    #] ThreadsProcessor : 
    #[ LoadReadjusted :
*/
int LoadReadjusted(void)
{
    ALLPRIVATES *B0 = AB[0];
    THREADBUCKET *thr = 0, *thrtogo = 0;
    int numtogo, numfree, numbusy, n, nperbucket, extra, i, j, u, bus;
    LONG numinput;
    WORD *t1, *c1, *t2, *c2, *t3;
/*
    Start with waiting for at least one free processor.
    We don't want the master competing for time when all are busy.
*/
    while ( topofavailables <= 0 ) MasterWait();
/*
    Now look for the fullest bucket and make a list of free buckets
    The bad part is that most numbers can change at any moment.
*/
restart:;
    numtogo = 0;
    numfree = 0;
    numbusy = 0;
    for ( j = 0; j < numthreadbuckets; j++ ) {
        thr = threadbuckets[j];
        if ( thr->free == BUCKETFREE || thr->free == BUCKETATEND
        || thr->free == BUCKETCOMINGFREE ) {
            freebuckets[numfree++] = thr;
        }
        else if ( thr->type != BUCKETDOINGTERMS ) {}
        else if ( thr->totnum > 1 ) { /* never steal from a bucket with one term */
            LOCK(thr->lock);
            bus = thr->busy;
            UNLOCK(thr->lock);
            if ( thr->free == BUCKETINUSE ) {
                n = thr->totnum-thr->usenum;
                if ( bus == BUCKETASSIGNED ) numbusy++;
                else if ( ( bus != BUCKETASSIGNED )
                       && ( n > numtogo ) ) {
                    numtogo = n;
                    thrtogo = thr;
                }
            }
            else if ( bus == BUCKETTOBERELEASED
             && thr->free == BUCKETRELEASED ) {
                freebuckets[numfree++] = thr;
                thr->free = BUCKETATEND;
                LOCK(thr->lock);
                thr->busy = BUCKETPREPARINGTERM;
                UNLOCK(thr->lock);
            }
        }
    }
    if ( numfree == 0 ) return(0); /* serious problem */
    if ( numtogo > 0 ) {   /* provisionally there is something to be stolen */
        thr = thrtogo;
/*
        If the number has changed there is good progress.
        Maybe there is another thread that needs assistance.
        We start all over.
*/
        if ( thr->totnum-thr->usenum < numtogo ) goto restart;
/*
        If the thread is in the term loading phase
        (thr->busy == BUCKETPREPARINGTERM) we better stay away from it.
        We wait now for the thread to be busy, and don't allow it
        now to drop out of this state till we are done here.
        This all depends on whether AT.LoadBalancing == 1 is seen by
        the thread.
*/
        LOCK(thr->lock);
        if ( thr->busy != BUCKETDOINGTERM ) {
            UNLOCK(thr->lock);
            goto restart;
        }
        if ( thr->totnum-thr->usenum < numtogo ) {
            UNLOCK(thr->lock);
            goto restart;
        }
        thr->free = BUCKETTERMINATED;
/*
        The above will signal the thread we want to terminate.
        Next all effort goes into making sure the landing is soft.
        Unfortunately we don't want to wait for a signal, because the thread
        may be working for a long time on a single term.
*/
        if ( thr->usenum == thr->totnum ) {
/*
            Terminated in the mean time or by now working on the
            last term. Try again.
*/
            thr->free = BUCKETATEND;
            UNLOCK(thr->lock);
            goto restart;
        }
        goto intercepted;
    }
/* This has always been commented. Indeed no lock is held here. */
/*  UNLOCK(thr->lock); */
    if ( numbusy > 0 ) {
        /* JD: this avoids large runtimes for tform tests under valgrind.
           What seems to happen is we return from here, goto Finalize, and
           end up in LoadReadjusted again without the threads having a
           chance to update their busy status. Then we end up here again.
           Sleep the thread for, say, 1us to allow threads to aquire the lock. */
        struct timespec sleeptime;
        sleeptime.tv_sec = 0;
        sleeptime.tv_nsec = 1000L;
        nanosleep(&sleeptime, NULL);
        return(1); /* Wait a bit.... */
    }
    return(0);
intercepted:;
/*
    We intercepted one successfully. Now it becomes interesting. Action:
    1: determine how many terms per free bucket.
    2: find the first untreated term.
    3: put the terms in the free buckets.
 
    Remember: we still have the lock to avoid interference from the thread
    that is being robbed. We were holding it and then jumped here with
    goto intercepted.
*/
    numinput = thr->firstterm + thr->usenum;
    nperbucket = numtogo / numfree;
    extra = numtogo - nperbucket*numfree;
    if ( AR0.DeferFlag ) {
        t1 = thr->threadbuffer; c1 = thr->compressbuffer; u = thr->usenum;
        for ( n = 0; n < thr->usenum; n++ ) { t1 += *t1; c1 += *c1; }
        t3 = t1;
        if ( extra > 0 ) {
          for ( i = 0; i < extra; i++ ) {
            thrtogo = freebuckets[i];
            t2 = thrtogo->threadbuffer;
            c2 = thrtogo->compressbuffer;
            thrtogo->free = BUCKETFILLED;
            thrtogo->type = BUCKETDOINGTERMS;
            thrtogo->totnum = nperbucket+1;
            thrtogo->ddterms = 0;
            thrtogo->usenum = 0;
            thrtogo->busy = BUCKETASSIGNED;
            thrtogo->firstterm = numinput;
            numinput += nperbucket+1;
            for ( n = 0; n <= nperbucket; n++ ) {
                j = *t1; NCOPY(t2,t1,j);
                j = *c1; NCOPY(c2,c1,j);
                thrtogo->deferbuffer[n] = thr->deferbuffer[u++];
            }
            *t2 = *c2 = 0;
          }
        }
        if ( nperbucket > 0 ) {
          for ( i = extra; i < numfree; i++ ) {
            thrtogo = freebuckets[i];
            t2 = thrtogo->threadbuffer;
            c2 = thrtogo->compressbuffer;
            thrtogo->free = BUCKETFILLED;
            thrtogo->type = BUCKETDOINGTERMS;
            thrtogo->totnum = nperbucket;
            thrtogo->ddterms = 0;
            thrtogo->usenum = 0;
            thrtogo->busy = BUCKETASSIGNED;
            thrtogo->firstterm = numinput;
            numinput += nperbucket;
            for ( n = 0; n < nperbucket; n++ ) {
                j = *t1; NCOPY(t2,t1,j);
                j = *c1; NCOPY(c2,c1,j);
                thrtogo->deferbuffer[n] = thr->deferbuffer[u++];
            }
            *t2 = *c2 = 0;
          }
        }
    }
    else {
        t1 = thr->threadbuffer;
        for ( n = 0; n < thr->usenum; n++ ) { t1 += *t1; }
        t3 = t1;
        if ( extra > 0 ) {
          for ( i = 0; i < extra; i++ ) {
            thrtogo = freebuckets[i];
            t2 = thrtogo->threadbuffer;
            thrtogo->free = BUCKETFILLED;
            thrtogo->type = BUCKETDOINGTERMS;
            thrtogo->totnum = nperbucket+1;
            thrtogo->ddterms = 0;
            thrtogo->usenum = 0;
            thrtogo->busy = BUCKETASSIGNED;
            thrtogo->firstterm = numinput;
            numinput += nperbucket+1;
            for ( n = 0; n <= nperbucket; n++ ) {
                j = *t1; NCOPY(t2,t1,j);
            }
            *t2 = 0;
          }
        }
        if ( nperbucket > 0 ) {
          for ( i = extra; i < numfree; i++ ) {
            thrtogo = freebuckets[i];
            t2 = thrtogo->threadbuffer;
            thrtogo->free = BUCKETFILLED;
            thrtogo->type = BUCKETDOINGTERMS;
            thrtogo->totnum = nperbucket;
            thrtogo->ddterms = 0;
            thrtogo->usenum = 0;
            thrtogo->busy = BUCKETASSIGNED;
            thrtogo->firstterm = numinput;
            numinput += nperbucket;
            for ( n = 0; n < nperbucket; n++ ) {
                j = *t1; NCOPY(t2,t1,j);
            }
            *t2 = 0;
          }
        }
    }
    *t3 = 0;   /* This is some form of extra insurance */
    if ( thr->free == BUCKETRELEASED && thr->busy == BUCKETTOBERELEASED ) {
        thr->free = BUCKETATEND; thr->busy = BUCKETPREPARINGTERM;
    }
    UNLOCK(thr->lock);
    return(1);
}
 
/*
    #] LoadReadjusted : 
    #[ SortStrategy :
*/
/*
    #] SortStrategy : 
    #[ PutToMaster :
*/
int PutToMaster(PHEAD WORD *term)
{
    int i,j,nexti,ret = 0;
    int urgent = 0;
    WORD *t, *fill, *top, zero = 0;
    if ( term == 0 ) { /* Mark the end of the expression */
        t = &zero; j = 1;
    }
    else {
        t = term; ret = j = *term;
        if ( j == 0 ) { j = 1; } /* Just in case there is a spurious end */
    }
    i = AT.SB.FillBlock;          /* The block we are working at */
    fill = AT.SB.MasterFill[i];     /* Where we are filling */
    top = AT.SB.MasterStop[i];      /* End of the block */
 
    // If there is space in the block, and we have already written MINWRITENUMBEROFTERMS,
    // determine if the reading thread is waiting for us by trying to lock the previous
    // block. If we manage to lock it, then we still have time to continue filling this
    // block. If we can't lock it, the reading thread is waiting for us and we should
    // move to the next block ASAP.
    if ( j < top - fill && AT.SB.BlockTerms[i] > MINWRITENUMBEROFTERMS ) {
        const int prev = ( i == 1 ? AT.SB.MasterNumBlocks : i-1 );
        if ( ! pthread_mutex_trylock(&(AT.SB.MasterBlockLock[prev])) ) {
            UNLOCK(AT.SB.MasterBlockLock[prev]);
        }
        else {
            urgent = 1;
        }
    }
 
    // If the term doesn't fit in the current block, or a thread is waiting for us
    // (and we've already written at least MINWRITENUMBEROFTERMS), move to the next:
    if ( ( j >= top - fill ) || urgent ) {
        nexti = i+1;
        if ( nexti > AT.SB.MasterNumBlocks ) {
            nexti = 1;
        }
        LOCK(AT.SB.MasterBlockLock[nexti]);
        UNLOCK(AT.SB.MasterBlockLock[i]);
        AT.SB.MasterFill[i] = AT.SB.MasterStart[i];
        AT.SB.FillBlock = i = nexti;
        fill = AT.SB.MasterStart[i];
        top = AT.SB.MasterStop[i];
        if ( AT.SB.BlockTerms[i] != 0 ) {
            // In this case, there has been an accounting error in a previous use
            // of this block. Blocks that have been read from and unlocked, should
            // have BlockTerms == 0.
/* INTERNAL_ERROR_EXCL_START */
            MLOCK(ErrorMessageLock);
            MesPrint("!>Error in PutToMaster, starting a block with BlockTerms != 0");
            MUNLOCK(ErrorMessageLock);
            Terminate(-1);
/* INTERNAL_ERROR_EXCL_STOP */
        }
    }
 
    NCOPY(fill, t, j);
    AT.SB.BlockTerms[i]++;
    AT.SB.MasterFill[i] = fill;
    return(ret);
}
 
/*
    #] PutToMaster : 
    #[ SortBotOut :
*/
 
 
int
SortBotOut(PHEAD WORD *term)
{
    WORD im;
 
    if ( AT.identity != 0 ) return(PutToMaster(BHEAD term));
 
    if ( term == 0 ) {
        if ( FlushOut(&SortBotPosition,AR.outfile,1) ) return(-1);
        ADDPOS(AT.SS->SizeInFile[0],1);
        return(0);
    }
    else {
        numberofterms++;
        if ( ( im = PutOut(BHEAD term,&SortBotPosition,AR.outfile,1) ) < 0 ) {
/* INTERNAL_ERROR_EXCL_START */
            MLOCK(ErrorMessageLock);
            MesPrint("Called from SortBotMasterMerge/SortBotOut");
            MUNLOCK(ErrorMessageLock);
            return(-1);
/* INTERNAL_ERROR_EXCL_STOP */
        }
        ADDPOS(AT.SS->SizeInFile[0],im);
        return(im);
    }
}
 
/*
    #] SortBotOut : 
    #[ SortBotMasterMerge :
*/
 
 
int SortBotMasterMerge(void)
{
    FILEHANDLE *fin, *fout;
    ALLPRIVATES *B = AB[0], *BB;
    POSITION position;
    SORTING *S = AT.SS;
    int i, j;
/*
    Get the sortbots get to claim their writing blocks.
    We have to wait till all have been claimed because they also have to
    claim the last writing blocks of the workers to prevent the head of
    the circular buffer to overrun the tail.
 
    Before waiting we can do some needed initializations.
    Also the master has to claim the last writing blocks of its input.
*/
    topsortbotavailables = 0;
    for ( i = numberofworkers+1; i <= numberofworkers+GetNumberOfSortBots(numberofworkers); i++ ) {
        WakeupThread(i,INISORTBOT);
    }
 
    AS.MasterSort = 1;
    fout = AR.outfile;
    numberofterms = 0;
    AR.CompressPointer[0] = 0;
    numberclaimed = 0;
    BB = AB[AT.SortBotIn1];
    LOCK(BB->T.SB.MasterBlockLock[BB->T.SB.MasterNumBlocks]);
    if ( AT.SortBotIn2 >= 0 ) {
        BB = AB[AT.SortBotIn2];
        LOCK(BB->T.SB.MasterBlockLock[BB->T.SB.MasterNumBlocks]);
    }
 
    MasterWaitAllSortBots();
/*
    Now we can start up the workers. They will claim their writing blocks.
    Here the master will wait till all writing blocks have been claimed.
*/
    for ( i = 1; i <= numberofworkers; i++ ) {
        j = GetThread(i);
        WakeupThread(i,FINISHEXPRESSION);
    }
/*
    Prepare the output file in the mean time.
*/
    if ( fout->handle >= 0 ) {
        PUTZERO(SortBotPosition);
        SeekFile(fout->handle,&SortBotPosition,SEEK_END);
        ADDPOS(SortBotPosition,((fout->POfill-fout->PObuffer)*sizeof(WORD)));
    }
    else {
        SETBASEPOSITION(SortBotPosition,(fout->POfill-fout->PObuffer)*sizeof(WORD));
    }
    MasterWaitAllBlocks();
/*
    Now we can start the sortbots after which the master goes in
    sortbot mode to do its part of the job (the very final merge and
    the writing to output file).
*/
    topsortbotavailables = 0;
    for ( i = numberofworkers+1; i <= numberofworkers+GetNumberOfSortBots(numberofworkers); i++ ) {
        WakeupThread(i,RUNSORTBOT);
    }
    if ( SortBotMerge(BHEAD0) ) {
/* INTERNAL_ERROR_EXCL_START */
        MLOCK(ErrorMessageLock);
        MesPrint("!>Called from SortBotMasterMerge");
        MUNLOCK(ErrorMessageLock);
        AS.MasterSort = 0;
        return(-1);
/* INTERNAL_ERROR_EXCL_STOP */
    }
/*
    And next the cleanup
*/
    if ( S->file.handle >= 0 )
    {
        fin = &S->file;
        CloseFile(fin->handle);
        remove(fin->name);
        fin->handle = -1;
    }
    position = S->SizeInFile[0];
    MULPOS(position,sizeof(WORD));
 
    // Collect global sort statistics information from the threads.
    // The total GenTerms is the sum of the thread GenTerms.
    // The total small/large buffer sort info is the sum of the thread info.
    // The total comparison count is the sum of the thread and sortbot counts.
    // The total unsorted size is the sum of the total generated terms sizes
    // The total maximal term size is the maximum of all maximal term sizes
    S->GenTerms = 0;
    for ( j = 1; j <= numberofworkers; j++ ) {
        S->GenTerms += AB[j]->T.SS->GenTerms;
        S->verbComparisons += AB[j]->T.SS->verbComparisons;
        if ( S->verbMaxTermSize < AB[j]->T.SS->verbMaxTermSize )
            S->verbMaxTermSize = AB[j]->T.SS->verbMaxTermSize;
        S->verbSBsortTerms += AB[j]->T.SS->verbSBsortTerms;
        S->verbSBsortCap += AB[j]->T.SS->verbSBsortCap;
        S->verbLBsortPatches += AB[j]->T.SS->verbLBsortPatches;
        S->verbLBsortCap += AB[j]->T.SS->verbLBsortCap;
        S->verbUnsortedSize += AB[j]->T.SS->verbUnsortedSize;
    }
    for ( j = numberofworkers+1; j <= numberofworkers+GetNumberOfSortBots(numberofworkers); j++ ) {
        S->verbComparisons += AB[j]->T.SS->verbComparisons;
        if ( S->verbMaxTermSize < AB[j]->T.SS->verbMaxTermSize )
            S->verbMaxTermSize = AB[j]->T.SS->verbMaxTermSize;
    }
 
    S->TermsLeft = numberofterms;
    WriteStats(&position,STATSPOSTSORT,NOCHECKLOGTYPE);
    Expressions[AR.CurExpr].counter = S->TermsLeft;
    Expressions[AR.CurExpr].size = position;
    AS.MasterSort = 0;
/*
    The next statement is to prevent one of the sortbots not having
    completely cleaned up before the next module starts.
    If this statement is omitted every once in a while one of the sortbots
    is still running when the next expression starts and misses its
    initialization. The result is usually disastrous.
*/
    MasterWaitAllSortBots();
 
    return(0);
}
 
 
/*
    #] SortBotMasterMerge : 
    #[ SortBotMerge :
*/
 
 
int SortBotMerge(PHEAD0)
{
    GETBIDENTITY
    ALLPRIVATES *Bin1, *Bin2;
    Bin1 = AB[AT.SortBotIn1];
    if ( AT.SortBotIn2 >= 0 ) {
        Bin2 = AB[AT.SortBotIn2];
    }
    else {
        // Incorrect dereference of this will cause a crash: all accesses must
        // be protected with with "if ( AT.SortBotIn2 >= 0 )"
        Bin2 = NULL;
    }
    WORD *term1, *term2, *wp;
    int blin1, blin2;   /* Current block numbers */
    int error = 0;
    WORD l1, l2, *m1, *m2, *w, r1, r2, r3, r33, r31, *tt1, ii;
    WORD *to, *from, im, c;
    UWORD *coef;
    SORTING *S = AT.SS;
#ifdef WITHFLOAT
    WORD *fun1, *fun2, *fun3, *tstop1, *tstop2, size1, size2, size3, l3, jj, *m3;
#endif
/*
    Set the pointers to the input terms and the output space
*/
    coef = AN.SoScratC;
    blin1 = 1;
    blin2 = 1;
    if ( AT.identity == 0 ) {
        wp = AT.WorkPointer;
    }
    else {
        wp = AT.WorkPointer = AT.WorkSpace;
    }
/*
    Get the locks for reading the input
    This means that we can start once these locks have been cleared
    which means that there will be input.
*/
    LOCK(Bin1->T.SB.MasterBlockLock[blin1]);
    if ( AT.SortBotIn2 >= 0 ) {
        LOCK(Bin2->T.SB.MasterBlockLock[blin2]);
    }
    term1 = Bin1->T.SB.MasterStart[blin1];
    WORD zeroterm = 0;
    term2 = &zeroterm;
    if ( AT.SortBotIn2 >= 0 ) {
        term2 = Bin2->T.SB.MasterStart[blin2];
    }
    AT.SB.FillBlock = 1;
/*
    Now the main loop. Keep going until one of the two hits the end.
*/
    while ( *term1 && *term2 ) {
        if ( ( c = CompareTerms(BHEAD term1,term2,(WORD)0) ) > 0 ) {
/*
            #[ One is smallest :
*/
            Bin1->T.SB.BlockTerms[blin1]--;
            if ( SortBotOut(BHEAD term1) < 0 ) {
/* INTERNAL_ERROR_EXCL_START */
                MLOCK(ErrorMessageLock);
                MesPrint("!>Called from SortBotMerge with thread = %d",AT.identity);
                MUNLOCK(ErrorMessageLock);
                error = -1;
                goto ReturnError;
/* INTERNAL_ERROR_EXCL_STOP */
            }
            term1 += *term1;
            if ( Bin1->T.SB.BlockTerms[blin1] == 0 ) {
                if ( blin1 == 1 ) {
                    UNLOCK(Bin1->T.SB.MasterBlockLock[Bin1->T.SB.MasterNumBlocks]);
                }
                else {
                    UNLOCK(Bin1->T.SB.MasterBlockLock[blin1-1]);
                }
                if ( blin1 == Bin1->T.SB.MasterNumBlocks ) {
                    blin1 = 1;
                }
                else {
                    blin1++;
                }
                LOCK(Bin1->T.SB.MasterBlockLock[blin1]);
                Bin1->T.SB.MasterBlock = blin1;
                term1 = Bin1->T.SB.MasterStart[blin1];
            }
/*
            #] One is smallest : 
*/
        }
        else if ( c < 0 ) {
/*
            #[ Two is smallest :
*/
            Bin2->T.SB.BlockTerms[blin2]--;
            if ( SortBotOut(BHEAD term2) < 0 ) {
/* INTERNAL_ERROR_EXCL_START */
                MLOCK(ErrorMessageLock);
                MesPrint("!>Called from SortBotMerge with thread = %d",AT.identity);
                MUNLOCK(ErrorMessageLock);
                error = -1;
                goto ReturnError;
/* INTERNAL_ERROR_EXCL_STOP */
            }
next2:
            term2 += *term2;
            if ( Bin2->T.SB.BlockTerms[blin2] == 0 ) {
                if ( blin2 == 1 ) {
                    UNLOCK(Bin2->T.SB.MasterBlockLock[Bin2->T.SB.MasterNumBlocks]);
                }
                else {
                    UNLOCK(Bin2->T.SB.MasterBlockLock[blin2-1]);
                }
                if ( blin2 == Bin2->T.SB.MasterNumBlocks ) {
                    blin2 = 1;
                }
                else {
                    blin2++;
                }
                LOCK(Bin2->T.SB.MasterBlockLock[blin2]);
                Bin2->T.SB.MasterBlock = blin2;
                term2 = Bin2->T.SB.MasterStart[blin2];
            }
/*
            #] Two is smallest : 
*/
        }
        else {
/*
            #[ Equal :
*/
            Bin1->T.SB.BlockTerms[blin1]--;
            Bin2->T.SB.BlockTerms[blin2]--;
            l1 = *( m1 = term1 );
            l2 = *( m2 = term2 );
            if ( S->PolyWise ) {  /* Here we work with PolyFun */
                tt1 = m1;
                m1 += S->PolyWise;
                m2 += S->PolyWise;
                if ( S->PolyFlag == 2 ) {
                    AT.WorkPointer = wp;
                    w = poly_ratfun_add(BHEAD m1,m2);
                    if ( *tt1 + w[1] - m1[1] > AM.MaxTer/((LONG)sizeof(WORD)) ) {
                        MLOCK(ErrorMessageLock);
                        MesPrint("Term too complex in PolyRatFun addition. MaxTermSize of %10l is too small",AM.MaxTer);
                        MUNLOCK(ErrorMessageLock);
                        Terminate(-1);
                    }
                    AT.WorkPointer = wp;
                    if ( w[FUNHEAD] == -SNUMBER && w[FUNHEAD+1] == 0 && w[1] > FUNHEAD ) {
                        goto cancelled;
                    }
                }
                else {
                    w = wp;
                    if ( w + m1[1] + m2[1] > AT.WorkTop ) {
                        MLOCK(ErrorMessageLock);
                        MesPrint("SortBotMerge(%d): A Maxtermsize of %10l is too small",
                                AT.identity,AM.MaxTer/sizeof(WORD));
                        MesPrint("m1[1] = %d, m2[1] = %d, Space = %l",m1[1],m2[1],(LONG)(AT.WorkTop-wp));
                        PrintTerm(term1,"term1");
                        PrintTerm(term2,"term2");
                        MesPrint("PolyWise = %d",S->PolyWise);
                        MUNLOCK(ErrorMessageLock);
                        Terminate(-1);
                    }
                    AddArgs(BHEAD m1,m2,w);
                }
                r1 = w[1];
                if ( r1 <= FUNHEAD
                    || ( w[FUNHEAD] == -SNUMBER && w[FUNHEAD+1] == 0 ) )
                         { goto cancelled; }
                if ( r1 == m1[1] ) {
                    NCOPY(m1,w,r1);
                }
                else if ( r1 < m1[1] ) {
                    r2 = m1[1] - r1;
                    m2 = w + r1;
                    m1 += m1[1];
                    while ( --r1 >= 0 ) *--m1 = *--m2;
                    m2 = m1 - r2;
                    r1 = S->PolyWise;
                    while ( --r1 >= 0 ) *--m1 = *--m2;
                    *m1 -= r2;
                    term1 = m1;
                }
                else {
                    // Here we are writing the new merged term *before* the original start of term1.
                    // We can always do this, since before term1 there is previous term data of this
                    // block, or the previous block, for which we are holding a lock. This requires
                    // the existence of "block 0", if term1 is the first term of block 1!
                    // It also requires the blocks to be contiguous in memory; we can't allocate
                    // separate memory regions for each block without larger-scale changes.
                    r2 = r1 - m1[1];
                    m2 = tt1 - r2;
                    r1 = S->PolyWise;
                    m1 = tt1;
                    *m1 += r2;
                    term1 = m2;
                    NCOPY(m2,m1,r1);
                    r1 = w[1];
                    NCOPY(m2,w,r1);
                }
            }
#ifdef WITHFLOAT
            else if ( AT.SortFloatMode ) {
/*
                The terms are in m1/term1 and m2/term2 and their length in l1 and l2.
                We have to locate the floats which have already been
                verified in the compare routine. Once we have the new
                term we can jump to the code that writes away the
                result after adding the rationals.
*/
                tstop1 = m1+l1; size1 = tstop1[-1]; tstop1 -= ABS(size1);
                tstop2 = m2+l2; size2 = tstop2[-1]; tstop2 -= ABS(size2);
                if ( AT.SortFloatMode == 3 ) {
                    fun1 = m1+1;
                    while ( fun1[0] != FLOATFUN && fun1+fun1[1] < tstop1 ) {
                        fun1 += fun1[1];
                    }
                    if ( size1 < 0 ) fun1[FUNHEAD+3] = -fun1[FUNHEAD+3];
                    UnpackFloat(aux1,fun1);
                    fun2 = m2+1;
                    while ( fun2[0] != FLOATFUN && fun2+fun2[1] < tstop2 ) {
                        fun2 += fun2[1];
                    }
                    if ( size2 < 0 ) fun2[FUNHEAD+3] = -fun2[FUNHEAD+3];
                    UnpackFloat(aux2,fun2);
                }
                else if ( AT.SortFloatMode == 1 ) {
                    fun1 = m1+1;
                    while ( fun1[0] != FLOATFUN && fun1+fun1[1] < tstop1 ) {
                        fun1 += fun1[1];
                    }
                    if ( size1 < 0 ) fun1[FUNHEAD+3] = -fun1[FUNHEAD+3];
                    UnpackFloat(aux1,fun1);
                    fun2 = tstop2;
                    RatToFloat(aux2,(UWORD *)fun2,size2);
                }
                else if ( AT.SortFloatMode == 2 ) {
                    fun1 = tstop1;
                    RatToFloat(aux1,(UWORD *)fun1,size1);
                    fun2 = m2+1;
                    while ( fun2[0] != FLOATFUN && fun2+fun2[1] < tstop2 ) {
                        fun2 += fun2[1];
                    }
                    if ( size2 < 0 ) fun2[FUNHEAD+3] = -fun2[FUNHEAD+3];
                    UnpackFloat(aux2,fun2);
                }
                else {
/* INTERNAL_ERROR_EXCL_START */
                    MLOCK(ErrorMessageLock);
                    MesPrint("!>Illegal value %d for AT.SortFloatMode in SortBotMerge.",AT.SortFloatMode);
                    MUNLOCK(ErrorMessageLock);
                    Terminate(-1);
                    return(0);
/* INTERNAL_ERROR_EXCL_STOP */
                }
                mpf_add(aux3,aux1,aux2);
                size3 = mpf_sgn(aux3);
                if ( size3 == 0 ) { /* Cancelling! Rare! */
                    goto cancelled;
                }
                else if ( size3 < 0 ) mpf_neg(aux3,aux3);
                fun3 = TermMalloc("SortBotMasterMerge");
                PackFloat(fun3,aux3);
                l3 = fun3[1]+(fun1-m1)+3; /* new size */
 
                if ( l3 != l1 ) {
/*
                    Copy it all to wp
*/
                    if ( (l3)*((LONG)sizeof(WORD)) >= AM.MaxTer ) {
                        MLOCK(ErrorMessageLock);
                        MesPrint("SortBotMasterMerge: Coefficient overflow during sort");
                        MUNLOCK(ErrorMessageLock);
                        goto ReturnError;
                    }
                    m3 = wp; m2 = term1;
                    while ( m2 < fun1 ) *m3++ = *m2++;
                    for ( jj = 0; jj < fun3[1]; jj++ ) *m3++ = fun3[jj];
                    *m3++ = 1; *m3++ = 1;
                    *m3++ = size3 < 0 ? -3: 3;
                    *wp = m3-wp;
                    TermFree(fun3,"SortBotMasterMerge");
                    goto PutOutwp;
                }
                for ( jj = 0; jj < fun3[1]; jj++ ) fun1[jj] = fun3[jj];
                fun1 += fun3[1];
                *fun1++ = 1; *fun1++ = 1;
                *fun1++ = size3 < 0 ? -3: 3;
                *term1 = fun1-term1;
                TermFree(fun3,"SortBotMasterMerge");
            }
#endif
            else {
                r1 = *( m1 += l1 - 1 );
                m1 -= ABS(r1) - 1;
                r1 = ( ( r1 > 0 ) ? (r1-1) : (r1+1) ) >> 1;
                r2 = *( m2 += l2 - 1 );
                m2 -= ABS(r2) - 1;
                r2 = ( ( r2 > 0 ) ? (r2-1) : (r2+1) ) >> 1;
 
                if ( AddRat(BHEAD (UWORD *)m1,r1,(UWORD *)m2,r2,coef,&r3) ) {
                    MLOCK(ErrorMessageLock);
                    MesCall("SortBotMerge");
                    MUNLOCK(ErrorMessageLock);
                    SETERROR(-1)
                }
 
                if ( AN.ncmod != 0 ) {
                    if ( ( AC.modmode & POSNEG ) != 0 ) {
                        NormalModulus(coef,&r3);
                    }
                    else if ( BigLong(coef,r3,(UWORD *)AC.cmod,ABS(AN.ncmod)) >= 0 ) {
                        SubPLon(coef,r3,(UWORD *)AC.cmod,ABS(AN.ncmod),coef,&r3);
                        coef[r3] = 1;
                        for ( ii = 1; ii < r3; ii++ ) coef[r3+ii] = 0;
                    }
                }
                if ( !r3 ) { goto cancelled; }
                r3 *= 2;
                r33 = ( r3 > 0 ) ? ( r3 + 1 ) : ( r3 - 1 );
                if ( r3 < 0 ) r3 = -r3;
                if ( r1 < 0 ) r1 = -r1;
                r1 *= 2;
                r31 = r3 - r1;
                if ( !r31 ) {       /* copy coef into term1 */
                    m2 = (WORD *)coef; im = r3;
                    NCOPY(m1,m2,im);
                    *m1 = r33;
                }
                else {
                    to = wp; from = term1;
                    while ( from < m1 ) *to++ = *from++;
                    from = (WORD *)coef; im = r3;
                    NCOPY(to,from,im);
                    *to++ = r33;
                    wp[0] = to - wp;
PutOutwp:
                    if ( SortBotOut(BHEAD wp) < 0 ) {
/* INTERNAL_ERROR_EXCL_START */
                        MLOCK(ErrorMessageLock);
                        MesPrint("!>Called from SortBotMerge with thread = %d",AT.identity);
                        MUNLOCK(ErrorMessageLock);
                        error = -1;
                        goto ReturnError;
/* INTERNAL_ERROR_EXCL_STOP */
                    }
                    goto cancelled;
                }
            }
            if ( SortBotOut(BHEAD term1) < 0 ) {
/* INTERNAL_ERROR_EXCL_START */
                MLOCK(ErrorMessageLock);
                MesPrint("!>Called from SortBotMerge with thread = %d",AT.identity);
                MUNLOCK(ErrorMessageLock);
                error = -1;
                goto ReturnError;
/* INTERNAL_ERROR_EXCL_STOP */
            }
cancelled:;     /* Now we need two new terms */
            term1 += *term1;
            if ( Bin1->T.SB.BlockTerms[blin1] == 0 ) {
 
                if ( blin1 == 1 ) {
                    UNLOCK(Bin1->T.SB.MasterBlockLock[Bin1->T.SB.MasterNumBlocks]);
                }
                else {
                    UNLOCK(Bin1->T.SB.MasterBlockLock[blin1-1]);
                }
                if ( blin1 == Bin1->T.SB.MasterNumBlocks ) {
                    blin1 = 1;
                }
                else {
                    blin1++;
                }
                LOCK(Bin1->T.SB.MasterBlockLock[blin1]);
                Bin1->T.SB.MasterBlock = blin1;
                term1 = Bin1->T.SB.MasterStart[blin1];
            }
            goto next2;
/*
            #] Equal : 
*/
        }
    }
/*
    Copy the tail
*/
    if ( *term1 ) {
/*
            #[ Tail in one :
*/
        while ( *term1 ) {
            Bin1->T.SB.BlockTerms[blin1]--;
            if ( SortBotOut(BHEAD term1) < 0 ) {
/* INTERNAL_ERROR_EXCL_START */
                MLOCK(ErrorMessageLock);
                MesPrint("!>Called from SortBotMerge with thread = %d",AT.identity);
                MUNLOCK(ErrorMessageLock);
                error = -1;
                goto ReturnError;
/* INTERNAL_ERROR_EXCL_STOP */
            }
            if ( Bin1->T.SB.BlockTerms[blin1] == 0 ) {
 
                if ( blin1 == 1 ) {
                    UNLOCK(Bin1->T.SB.MasterBlockLock[Bin1->T.SB.MasterNumBlocks]);
                }
                else {
                    UNLOCK(Bin1->T.SB.MasterBlockLock[blin1-1]);
                }
                if ( blin1 == Bin1->T.SB.MasterNumBlocks ) {
                    blin1 = 1;
                }
                else {
                    blin1++;
                }
                LOCK(Bin1->T.SB.MasterBlockLock[blin1]);
                Bin1->T.SB.MasterBlock = blin1;
                term1 = Bin1->T.SB.MasterStart[blin1];
            }
            else {
                term1 += *term1;
            }
        }
/*
            #] Tail in one : 
*/
    }
    else if ( *term2 ) {
/*
            #[ Tail in two :
*/
        while ( *term2 ) {
            Bin2->T.SB.BlockTerms[blin2]--;
            if ( SortBotOut(BHEAD term2) < 0 ) {
/* INTERNAL_ERROR_EXCL_START */
                MLOCK(ErrorMessageLock);
                MesPrint("!>Called from SortBotMerge with thread = %d",AT.identity);
                MUNLOCK(ErrorMessageLock);
                error = -1;
                goto ReturnError;
/* INTERNAL_ERROR_EXCL_STOP */
            }
            if ( Bin2->T.SB.BlockTerms[blin2] == 0 ) {
 
                if ( blin2 == 1 ) {
                    UNLOCK(Bin2->T.SB.MasterBlockLock[Bin2->T.SB.MasterNumBlocks]);
                }
                else {
                    UNLOCK(Bin2->T.SB.MasterBlockLock[blin2-1]);
                }
                if ( blin2 == Bin2->T.SB.MasterNumBlocks ) {
                    blin2 = 1;
                }
                else {
                    blin2++;
                }
                LOCK(Bin2->T.SB.MasterBlockLock[blin2]);
                Bin2->T.SB.MasterBlock = blin2;
                term2 = Bin2->T.SB.MasterStart[blin2];
            }
            else {
                term2 += *term2;
            }
        }
/*
            #] Tail in two : 
*/
    }
 
    // Both streams have hit the end-of-stream marker "0". We still need to
    // decrement the BlockTerms counters a final time, the marker is included
    // in the count.
    Bin1->T.SB.BlockTerms[blin1]--;
    if ( AT.SortBotIn2 >= 0 ) {
        Bin2->T.SB.BlockTerms[blin2]--;
    }
 
    SortBotOut(BHEAD 0);
ReturnError:;
/*
    Release all locks.
*/
    UNLOCK(Bin1->T.SB.MasterBlockLock[blin1]);
    if ( blin1 > 1 ) {
        UNLOCK(Bin1->T.SB.MasterBlockLock[blin1-1]);
    }
    else {
        UNLOCK(Bin1->T.SB.MasterBlockLock[Bin1->T.SB.MasterNumBlocks]);
    }
    if ( AT.SortBotIn2 >= 0 ) {
        UNLOCK(Bin2->T.SB.MasterBlockLock[blin2]);
        if ( blin2 > 1 ) {
            UNLOCK(Bin2->T.SB.MasterBlockLock[blin2-1]);
        }
        else {
            UNLOCK(Bin2->T.SB.MasterBlockLock[Bin2->T.SB.MasterNumBlocks]);
        }
    }
    if ( AT.identity > 0 ) {
        UNLOCK(AT.SB.MasterBlockLock[AT.SB.FillBlock]);
    }
/*
    And that was all folks
*/
    return(error);
}
 
 
/*
    #] SortBotMerge : 
    #[ IniSortBlocks :
*/
 
static int SortBlocksInitialized = 0;
 
int IniSortBlocks(int numworkers)
{
    ALLPRIVATES *B;
    SORTING *S;
    LONG totalsize, workersize, blocksize, numberofterms;
    int maxter, id, j;
    int numberofblocks = NUMBEROFBLOCKSINSORT, numparts;
    WORD *w;
 
    if ( SortBlocksInitialized ) return(0);
    SortBlocksInitialized = 1;
    if ( numworkers == 0 ) return(0);
 
    if ( GetNumberOfSortBots(numworkers) > 0 ) {
        numparts = numworkers + GetNumberOfSortBots(numberofworkers);
        numberofblocks = numberofblocks/2;
    }
    else {
        numparts = numworkers;
    }
    S = AM.S0;
    totalsize = S->LargeSize + S->SmallEsize;
    workersize = totalsize / numparts;
    maxter = AM.MaxTer/sizeof(WORD);
    blocksize = ( workersize - maxter )/numberofblocks;
    numberofterms = blocksize / maxter;
    if ( numberofterms < MINIMUMNUMBEROFTERMS ) {
/*
        This should have been taken care of in RecalcSetups.
*/
/* INTERNAL_ERROR_EXCL_START */
        MesPrint("!>We have a problem with the size of the blocks in IniSortBlocks");
        Terminate(-1);
/* INTERNAL_ERROR_EXCL_STOP */
    }
/*
    Layout:  For each worker
                block 0: size is maxter WORDS
                numberofblocks blocks of size blocksize WORDS
*/
    w = S->lBuffer;
    if ( w == 0 ) w = S->sBuffer;
    for ( id = 1; id <= numparts; id++ ) {
        B = AB[id];
        AT.SB.MasterBlockLock = (pthread_mutex_t *)Malloc1(
            sizeof(pthread_mutex_t)*(numberofblocks+1),"MasterBlockLock");
        AT.SB.MasterStart = (WORD **)Malloc1(sizeof(WORD *)*(numberofblocks+1)*3,"MasterBlock");
        AT.SB.MasterFill = AT.SB.MasterStart + (numberofblocks+1);
        AT.SB.MasterStop = AT.SB.MasterFill  + (numberofblocks+1);
        AT.SB.MasterNumBlocks = numberofblocks;
        AT.SB.BlockTerms = (LONG*)Malloc1(sizeof(LONG)*(numberofblocks+1),"BlockTerms");
        AT.SB.MasterBlock = 0;
        AT.SB.FillBlock = 0;
        AT.SB.MasterFill[0] = AT.SB.MasterStart[0] = w;
        AT.SB.BlockTerms[0] = 0;
        w += maxter;
        AT.SB.MasterStop[0] = w;
        AT.SB.MasterBlockLock[0] = dummylock;
        for ( j = 1; j <= numberofblocks; j++ ) {
            AT.SB.MasterFill[j] = AT.SB.MasterStart[j] = w;
            AT.SB.BlockTerms[j] = 0;
            w += blocksize;
            AT.SB.MasterStop[j] = w;
            AT.SB.MasterBlockLock[j] = dummylock;
        }
    }
    if ( w > S->sTop2 ) {
/* INTERNAL_ERROR_EXCL_START */
        MesPrint("!>Counting problem in IniSortBlocks");
        Terminate(-1);
/* INTERNAL_ERROR_EXCL_STOP */
    }
    return(0);
}
 
/*
    #] IniSortBlocks : 
    #[ UpdateSortBlocks :
*/
 
int UpdateSortBlocks(int numworkers)
{
    ALLPRIVATES *B;
    SORTING *S;
    LONG totalsize, workersize, blocksize, numberofterms;
    int maxter, id, j;
    int numberofblocks = NUMBEROFBLOCKSINSORT, numparts;
    WORD *w;
 
    if ( numworkers == 0 ) return(0);
 
    if ( GetNumberOfSortBots(numworkers) > 0) {
        numparts = numworkers + GetNumberOfSortBots(numberofworkers);
        numberofblocks = numberofblocks/2;
    }
    else {
        numparts = numworkers;
    }
    S = AM.S0;
    totalsize = S->LargeSize + S->SmallEsize;
    workersize = totalsize / numparts;
    maxter = AM.MaxTer/sizeof(WORD);
    blocksize = ( workersize - maxter )/numberofblocks;
    numberofterms = blocksize / maxter;
    if ( numberofterms < MINIMUMNUMBEROFTERMS ) {
/*
        This should have been taken care of in RecalcSetups.
*/
/* INTERNAL_ERROR_EXCL_START */
        MesPrint("!>We have a problem with the size of the blocks in UpdateSortBlocks");
        Terminate(-1);
/* INTERNAL_ERROR_EXCL_STOP */
    }
/*
    Layout:  For each worker
                block 0: size is maxter WORDS
                numberofblocks blocks of size blocksize WORDS
*/
    w = S->lBuffer;
    if ( w == 0 ) w = S->sBuffer;
    for ( id = 1; id <= numparts; id++ ) {
        B = AB[id];
        AT.SB.MasterFill[0] = AT.SB.MasterStart[0] = w;
        w += maxter;
        AT.SB.MasterStop[0] = w;
        for ( j = 1; j <= numberofblocks; j++ ) {
            AT.SB.MasterFill[j] = AT.SB.MasterStart[j] = w;
            w += blocksize;
            AT.SB.MasterStop[j] = w;
        }
    }
    if ( w > S->sTop2 ) {
/* INTERNAL_ERROR_EXCL_START */
        MesPrint("!>Counting problem in UpdateSortBlocks");
        Terminate(-1);
/* INTERNAL_ERROR_EXCL_STOP */
    }
    return(0);
}
 
/*
    #] UpdateSortBlocks : 
    #[ DefineSortBotTree :
*/
 
 
void DefineSortBotTree(void)
{
    ALLPRIVATES *B;
    int n, i, from;
    n = numberofworkers+GetNumberOfSortBots(numberofworkers);
    for ( i = numberofworkers+1, from = 1; i <= n; i++ ) {
        B = AB[i];
        AT.SortBotIn1 = from++;
        AT.SortBotIn2 = from++;
    }
    B = AB[0];
    AT.SortBotIn1 = from++;
    AT.SortBotIn2 = -1;
}
 
 
/*
    #] DefineSortBotTree : 
    #[ GetTerm2 :
 
    Routine does a GetTerm but only when a bracket index is involved and
    only from brackets that have been judged not suitable for treatment
    as complete brackets by a single worker. Whether or not a bracket should
    be treated by a single worker is decided by TreatIndexEntry
*/
 
WORD GetTerm2(PHEAD WORD *term)
{
    WORD *ttco, *tt, retval;
    LONG n,i;
    FILEHANDLE *fi;
    EXPRESSIONS e = AN.expr;
    BRACKETINFO *b  = e->bracketinfo;
    BRACKETINDEX *bi = b->indexbuffer;
    POSITION where, eonfile = AS.OldOnFile[e-Expressions], bstart, bnext;
/*
    1: Get the current position.
*/
    switch ( e->status ) {
        case UNHIDELEXPRESSION:
        case UNHIDEGEXPRESSION:
        case DROPHLEXPRESSION:
        case DROPHGEXPRESSION:
        case HIDDENLEXPRESSION:
        case HIDDENGEXPRESSION:
            fi = AR.hidefile;
            break;
        default:
            fi = AR.infile;
            break;
    }
    if ( AR.KeptInHold ) {
        retval = GetTerm(BHEAD term);
        return(retval);
    }
    SeekScratch(fi,&where);
    if ( AN.lastinindex < 0 ) {
/*
        We have to test whether we have to do the first bracket
*/
        if ( ( n = TreatIndexEntry(BHEAD 0) ) <= 0 ) {
            AN.lastinindex = n;
            where = bi[n].start;
            ADD2POS(where,eonfile);
            SetScratch(fi,&where);
/*
            Put the bracket in the Compress buffer.
*/
            ttco = AR.CompressBuffer;
            tt = b->bracketbuffer + bi[0].bracket;
            i = *tt;
            NCOPY(ttco,tt,i)
            AR.CompressPointer = ttco;
            retval = GetTerm(BHEAD term);
            return(retval);
        }
        else AN.lastinindex = n-1;
    }
/*
    2: Find the corresponding index number
       a: test whether it is in the current bracket
*/
    n = AN.lastinindex;
    bstart = bi[n].start;
    ADD2POS(bstart,eonfile);
    bnext = bi[n].next;
    ADD2POS(bnext,eonfile);
    if ( ISLESSPOS(bstart,where) && ISLESSPOS(where,bnext) ) {
        retval = GetTerm(BHEAD term);
        return(retval);
    }
    for ( n++ ; n < b->indexfill; n++ ) {
        i = TreatIndexEntry(BHEAD n);
        if ( i <= 0 ) {
/*
            Put the bracket in the Compress buffer.
*/
            ttco = AR.CompressBuffer;
            tt = b->bracketbuffer + bi[n].bracket;
            i = *tt;
            NCOPY(ttco,tt,i)
            AR.CompressPointer = ttco;
            AN.lastinindex = n;
            where = bi[n].start;
            ADD2POS(where,eonfile);
            SetScratch(fi,&(where));
            retval = GetTerm(BHEAD term);
            return(retval);
        }
        else n += i - 1;
    }
    return(0);
}
 
/*
    #] GetTerm2 : 
    #[ TreatIndexEntry :
*/
int TreatIndexEntry(PHEAD LONG n)
{
    BRACKETINFO *b  = AN.expr->bracketinfo;
    LONG numbra = b->indexfill - 1, i;
    LONG totterms;
    BRACKETINDEX *bi;
    POSITION pos1, average;
/*
    1: number of the bracket should be such that there is one bucket
       for each worker remaining.
*/
    if ( ( numbra - n ) <= numberofworkers ) return(0);
/*
    2: size of the bracket contents should be less than what remains in
       the expression divided by the number of workers.
*/
    bi = b->indexbuffer;
    DIFPOS(pos1,bi[numbra].next,bi[n].next);  /* Size of what remains */
    BASEPOSITION(average) = DIVPOS(pos1,(3*numberofworkers));
    DIFPOS(pos1,bi[n].next,bi[n].start);      /* Size of the current bracket */
    if ( ISLESSPOS(average,pos1) ) return(0);
/*
    It passes.
    Now check whether we can do more brackets
*/
    totterms = bi->termsinbracket;
    if ( totterms > 2*AC.ThreadBucketSize ) return(1);
    for ( i = 1; i < numbra-n; i++ ) {
        DIFPOS(pos1,bi[n+i].next,bi[n].start); /* Size of the combined brackets */
        if ( ISLESSPOS(average,pos1) ) return(i);
        totterms += bi->termsinbracket;
        if ( totterms > 2*AC.ThreadBucketSize ) return(i+1);
    }
/*
    We have a problem at the end of the system. Just do one.
*/
    return(1);
}
 
/*
    #] TreatIndexEntry : 
    #[ SetHideFiles :
*/
 
void SetHideFiles(void) {
    int i;
    ALLPRIVATES *B, *B0 = AB[0];
    for ( i = 1; i <= numberofworkers; i++ ) {
        B = AB[i];
        AR.hidefile->handle = AR0.hidefile->handle;
        if ( AR.hidefile->handle < 0 ) {
            AR.hidefile->PObuffer = AR0.hidefile->PObuffer;
            AR.hidefile->POstop = AR0.hidefile->POstop;
            AR.hidefile->POfill = AR0.hidefile->POfill;
            AR.hidefile->POfull = AR0.hidefile->POfull;
            AR.hidefile->POsize = AR0.hidefile->POsize;
            AR.hidefile->POposition = AR0.hidefile->POposition;
            AR.hidefile->filesize = AR0.hidefile->filesize;
        }
        else {
            AR.hidefile->PObuffer = AR.hidefile->wPObuffer;
            AR.hidefile->POstop = AR.hidefile->wPOstop;
            AR.hidefile->POfill = AR.hidefile->wPOfill;
            AR.hidefile->POfull = AR.hidefile->wPOfull;
            AR.hidefile->POsize = AR.hidefile->wPOsize;
            PUTZERO(AR.hidefile->POposition);
        }
    }
}
 
/*
    #] SetHideFiles : 
    #[ IniFbufs :
*/
 
void IniFbufs(void)
{
    int i;
    for ( i = 0; i < AM.totalnumberofthreads; i++ ) {
        IniFbuffer(AB[i]->T.fbufnum);
    }
}
 
/*
    #] IniFbufs : 
    #[ SetMods :
*/
 
void SetMods(void)
{
    ALLPRIVATES *B;
    int i, n, j;
    for ( j = 0; j < AM.totalnumberofthreads; j++ ) {
        B = AB[j];
        AN.ncmod = AC.ncmod;
        if ( AN.cmod != 0 ) M_free(AN.cmod,"AN.cmod");
        n = ABS(AN.ncmod);
        AN.cmod = (UWORD *)Malloc1(sizeof(WORD)*n,"AN.cmod");
        for ( i = 0; i < n; i++ ) AN.cmod[i] = AC.cmod[i];
    }
}
 
/*
    #] SetMods : 
    #[ UnSetMods :
*/
 
void UnSetMods(void)
{
    ALLPRIVATES *B;
    int j;
    for ( j = 0; j < AM.totalnumberofthreads; j++ ) {
        B = AB[j];
        if ( AN.cmod != 0 ) M_free(AN.cmod,"AN.cmod");
        AN.cmod = 0;
    }
}
 
/*
    #] UnSetMods : 
    #[ find_Horner_MCTS_expand_tree_threaded :
*/
 
void find_Horner_MCTS_expand_tree_threaded(void) {
    int id;
    while (( id = GetAvailableThread() ) < 0)
        MasterWait();   
    WakeupThread(id,MCTSEXPANDTREE);
}
 
/*
    #] find_Horner_MCTS_expand_tree_threaded : 
    #[ optimize_expression_given_Horner_threaded :
*/
 
extern void optimize_expression_given_Horner_threaded(void) {
    int id;
    while (( id = GetAvailableThread() ) < 0)
        MasterWait();   
    WakeupThread(id,OPTIMIZEEXPRESSION);
}
 
/*
    #] optimize_expression_given_Horner_threaded : 
*/
 
#endif