diff --git a/check/fixes.frm b/check/fixes.frm
index 3ad37168..b1fefdc5 100644
--- a/check/fixes.frm
+++ b/check/fixes.frm
@@ -2162,13 +2162,123 @@ Print;
 .end
 # Only for 64-bit systems. Otherwise "Output term too large".
 #require wordsize == 4
-# For now it fails because
-#   "Currently Stage 4 sorts are not allowed for function arguments or $ variables."
-assert runtime_error?
-#assert succeeded?
-#assert result("test1") =~ expr("0")
-#assert result("test2") =~ expr("g(0)")
+assert warning?("StageSort in sub-buffer: recommend increasing setup sub-buffer sizes")
+assert result("test1") =~ expr("0")
+assert result("test2") =~ expr("g(0)")
 *--#] Issue211 : 
+*--#[ Issue215_1 :
+#-
+
+#: SubTermsInSmall 16
+#: SubLargePatches 4
+#: SubFilePatches 4
+
+Off statistics;
+Off threadstats;
+
+CFunction f;
+Symbol j,x;
+
+#define BLOWUP "1000"
+
+Local argument = f(x) + f(2*x);
+Argument f;
+	Identify x = sum_(j,1,`BLOWUP',x*j/`BLOWUP');
+EndArgument;
+Identify f(x?) = x;
+Identify x = 1;
+Print;
+.end
+assert warning?("StageSort in sub-buffer: recommend increasing setup sub-buffer sizes")
+assert result("argument") =~ expr("3003/2")
+*--#] Issue215_1 :
+*--#[ Issue215_2 :
+#-
+
+#: SubTermsInSmall 16
+#: SubLargePatches 4
+#: SubFilePatches 4
+
+Off statistics;
+Off threadstats;
+
+Symbol j,x;
+
+#define BLOWUP "1000"
+
+#$dol = x;
+Local dollar = 1;
+Inside $dol;
+	Identify x = sum_(j,1,`BLOWUP',x*j/`BLOWUP');
+EndInside;
+Multiply $dol;
+Identify x = 1;
+Print;
+ModuleOption local $dol;
+.end
+assert warning?("StageSort in sub-buffer: recommend increasing setup sub-buffer sizes")
+assert result("dollar") =~ expr("1001/2")
+*--#] Issue215_2 :
+*--#[ Issue215_3 :
+#-
+
+#: SubTermsInSmall 16
+#: SubLargePatches 4
+#: SubFilePatches 4
+
+Off statistics;
+Off threadstats;
+
+Symbol j,x;
+
+#define BLOWUP "1000"
+
+Local term = x;
+Term;
+	Identify x = sum_(j,1,`BLOWUP',x*j/`BLOWUP');
+EndTerm;
+Identify x = 1;
+Print;
+.end
+assert warning?("StageSort in sub-buffer: recommend increasing setup sub-buffer sizes")
+assert result("term") =~ expr("1001/2")
+*--#] Issue215_3 :
+*--#[ Issue215_4 :
+#-
+
+#: TermsInSmall 16
+#: LargePatches 4
+#: FilePatches 4
+
+#: SubTermsInSmall 16
+#: SubLargePatches 4
+#: SubFilePatches 4
+
+Off statistics;
+Off threadstats;
+
+CFunction f;
+Symbol i,j,x;
+
+#define N "500"
+#define BLOWUP "1000"
+
+Local nested = {`N'*(`N'+1)/2}
+	#do i = 1,`N'
+		- f(x^`i')
+	#enddo
+	;
+
+Argument f;
+	Identify x^i? = sum_(j,1,`BLOWUP',x^i*i/`BLOWUP');
+EndArgument;
+Identify f(x?) = x;
+Identify x^i? = 1;
+Print;
+.end
+assert warning?("StageSort in sub-buffer: recommend increasing setup sub-buffer sizes")
+assert result("nested") =~ expr("0")
+*--#] Issue215_4 :
 *--#[ Issue222 :
 * accessing #factdollar factors causes program termination
 Symbol x;
diff --git a/doc/manual/sorting.tex b/doc/manual/sorting.tex
index ed5146b6..adfa26a1 100644
--- a/doc/manual/sorting.tex
+++ b/doc/manual/sorting.tex
@@ -254,10 +254,10 @@ \chapter{Sorting and statistics}
 the chapter on the setup \ref{setup}. This chapter gives also all default 
 values.
 
-There is one restriction on the sorting of function arguments and 
-\$-expressions: They are not allowed to go into the stage4 sorting. Any 
-such attempt will result in an error message and the suggestion to raise 
-the size of the buffers for this type of sorting.
+Sorting these `sub-buffers' may also run into a stage 4 sort when necessary.
+This usually implies that performance will be severely affected, hence the
+first time this happens during a program a warning will be printed, with the
+suggestion to increase the sizes of the sub-buffers.
 
 When \FORM\ is running in parallel mode (either \TFORM\ or \ParFORM) each worker 
 will need its own buffers. In \ParFORM\ in which the processors each control 
diff --git a/sources/checkpoint.c b/sources/checkpoint.c
index 343170d6..8216e77f 100644
--- a/sources/checkpoint.c
+++ b/sources/checkpoint.c
@@ -1184,7 +1184,6 @@ static void print_R()
 	MesPrint("%d", AR.DeferFlag);
 	MesPrint("%d", AR.TePos);
 	MesPrint("%d", AR.sLevel);
-	MesPrint("%d", AR.Stage4Name);
 	MesPrint("%d", AR.GetOneFile);
 	MesPrint("%d", AR.PolyFun);
 	MesPrint("%d", AR.PolyFunInv);
@@ -2333,7 +2332,6 @@ int DoRecovery(int *moduletype)
 	R_SET(AR.DeferFlag, WORD);
 	R_SET(AR.TePos, WORD);
 	R_SET(AR.sLevel, WORD);
-	R_SET(AR.Stage4Name, WORD);
 	R_SET(AR.GetOneFile, WORD);
 	R_SET(AR.PolyFun, WORD);
 	R_SET(AR.PolyFunInv, WORD);
@@ -2957,7 +2955,6 @@ static int DoSnapshot(int moduletype)
 	S_WRITE_B(&AR.DeferFlag, sizeof(WORD));
 	S_WRITE_B(&AR.TePos, sizeof(WORD));
 	S_WRITE_B(&AR.sLevel, sizeof(WORD));
-	S_WRITE_B(&AR.Stage4Name, sizeof(WORD));
 	S_WRITE_B(&AR.GetOneFile, sizeof(WORD));
 	S_WRITE_B(&AR.PolyFun, sizeof(WORD));
 	S_WRITE_B(&AR.PolyFunInv, sizeof(WORD));
diff --git a/sources/proces.c b/sources/proces.c
index b0fc8f74..b7b11edf 100644
--- a/sources/proces.c
+++ b/sources/proces.c
@@ -205,6 +205,7 @@ int Processor(void)
 #endif
 		AS.CollectOverFlag = 0;
 		AR.expchanged = 0;
+		AT.SubSortStage4Warning = 0;
 		if ( i == last ) LastExpression = 1;
 		else             LastExpression = 0;
 		if ( e->inmem ) {
diff --git a/sources/setfile.c b/sources/setfile.c
index d5ecee9e..1ef3b69e 100644
--- a/sources/setfile.c
+++ b/sources/setfile.c
@@ -417,10 +417,10 @@ int AllocSetups(void)
 	SETUPPARAMETERS *sp;
 	LONG LargeSize, SmallSize, SmallEsize, TermsInSmall, IOsize;
 	int MaxPatches, MaxFpatches, error = 0, i, size;
-	UBYTE *s;
 #ifndef WITHPTHREADS
 	int j;
 #endif
+	UBYTE *s;
 	sp = GetSetupPar((UBYTE *)"threads");
 	if ( sp->value > 0 ) AM.totalnumberofthreads = sp->value+1;
 
@@ -588,19 +588,14 @@ int AllocSetups(void)
 	/* AM.S0->file.ziosize was already set to a (larger) value by AllocSort, here it is re-set. */
 #ifdef WITHZLIB
 	AM.S0->file.ziosize = IOsize;
+	AM.S0->FoStage4.ziosize = IOsize;
 #ifndef WITHPTHREADS
-	AR.FoStage4[0].ziosize = IOsize;
-	AR.FoStage4[1].ziosize = IOsize;
 	AT.S0 = AM.S0;
 #endif
 #else
 #ifndef WITHPTHREADS
 	AT.S0 = AM.S0;
 #endif
-#endif
-#ifndef WITHPTHREADS
-	AR.FoStage4[0].POsize   = ((IOsize+sizeof(WORD)-1)/sizeof(WORD))*sizeof(WORD);
-	AR.FoStage4[1].POsize   = ((IOsize+sizeof(WORD)-1)/sizeof(WORD))*sizeof(WORD);
 #endif
 	sp = GetSetupPar((UBYTE *)"subsmallsize");
 	AM.SSmallSize = sp->value;
@@ -998,6 +993,17 @@ SORTING *AllocSort(LONG inLargeSize, LONG inSmallSize, LONG inSmallEsize, LONG i
 #ifdef WITHZLIB
 	sort->file.ziosize = IObuffersize*sizeof(WORD);
 	sort->file.ziobuffer = 0;
+#endif
+	FILEHANDLE *stage4 = &(sort->FoStage4);
+	memset(stage4,0,sizeof(*stage4));
+	stage4->POsize = sort->file.POsize;
+	stage4->handle = -1;
+#ifdef WITHPTHREADS
+	stage4->pthreadslock = dummylock;
+#endif
+#ifdef WITHZLIB
+	stage4->ziosize = sort->file.ziosize;
+	stage4->ziobuffer = 0;
 #endif
 	if ( AM.S0 != 0 ) {
 		sort->file.name = (char *)(sort->file.PObuffer + IObuffersize);
diff --git a/sources/sort.c b/sources/sort.c
index 69de71e7..520008d7 100644
--- a/sources/sort.c
+++ b/sources/sort.c
@@ -872,7 +872,6 @@ LONG EndSort(PHEAD WORD *buffer, int par)
 #endif
 			}
 		}
-		AR.Stage4Name = 0;
 #ifdef WITHPTHREADS
 		if ( AS.MasterSort && AC.ThreadSortFileSynch ) {
 			if ( S->file.handle >= 0 ) {
@@ -3539,7 +3538,7 @@ int MergePatches(WORD par)
 	}
 #endif
 	fin = &S->file;
-	fout = &(AR.FoStage4[0]);
+	fout = &(S->FoStage4);
 NewMerge:
 	coef = AN.SoScratC;
 	poin = S->poina; poin2 = S->poin2a;
@@ -3598,14 +3597,8 @@ int MergePatches(WORD par)
 	else if ( par == 0 && S->stage4 > 0 ) {
 /*
 		We will have to do our job more than once.
-		Input is from S->file and output will go to AR.FoStage4.
+		Input is from S->file and output will go to S->FoStage4.
 		The file corresponding to this last one must be made now.
-*/
-		AR.Stage4Name ^= 1;
-/*
-		s = (UBYTE *)(fout->name); while ( *s ) s++;
-		if ( AR.Stage4Name ) s[-1] += 1;
-		else                s[-1] -= 1;
 */
 		S->iPatches = S->fPatches;
 		S->fPatches = S->inPatches;
@@ -4221,14 +4214,6 @@ int MergePatches(WORD par)
 			goto ConMer;
 		}
 		else {
-/*
-			if ( fin == &(AR.FoStage4[0]) ) {
-				s = (UBYTE *)(fin->name); while ( *s ) s++;
-				if ( AR.Stage4Name == 1 ) s[-1] -= 1;
-				else                      s[-1] += 1;
-			}
-*/
-/*			TruncateFile(fin->handle); */
 			UpdateMaxSize();
 #ifdef WITHZLIB
 			ClearSortGZIP(fin);
@@ -4240,13 +4225,6 @@ int MergePatches(WORD par)
 			MesPrint("%w MergePatches removed in file %s",fin->name);
 			MUNLOCK(ErrorMessageLock);
 #endif
-/*
-			if ( fin == &(AR.FoStage4[0]) ) {
-				s = (UBYTE *)(fin->name); while ( *s ) s++;
-				if ( AR.Stage4Name == 1 ) s[-1] += 1;
-				else                      s[-1] -= 1;
-			}
-*/
 			fin->handle = -1;
 			{ FILEHANDLE *ff = fin; fin = fout; fout = ff; }
 			PUTZERO(S->SizeInFile[0]);
@@ -4440,35 +4418,26 @@ void StageSort(FILEHANDLE *fout)
 	SORTING *S = AT.SS;
 	if ( S->fPatchN >= S->MaxFpatches ) {
 		POSITION position;
-		if ( S != AT.S0 ) {
-/*
-			There are no proper provisions for stage 4 or higher sorts
-			for function arguments and $ variables. The reason:
-			The current code maps out the patches, based on the size of
-			the buffers in the FoStage4 structs, while they are used
-			inside the S->file struct that may have far smaller buffers.
-			By itself that might still be repairable, but it goes completely
-			wrong when during the sort polyRatFuns have to be added and they
-			would go into stage4 (very rare but possible).
-			The only really correct solution would be to put FoStage4 structs
-			in all sort levels. Messy. (JV 8-oct-2018).
-*/
-			MLOCK(ErrorMessageLock);
-			MesPrint("Currently Stage 4 sorts are not allowed for function arguments or $ variables.");
-			MesPrint("Please increase correspondingsorting parameters (sub-) in the setup.");
-			MUNLOCK(ErrorMessageLock);
-			Terminate(-1);
-		}
 		PUTZERO(position);
-		MLOCK(ErrorMessageLock);
+		if ( S == AT.S0 ) {
+			MLOCK(ErrorMessageLock);
 #ifdef WITHPTHREADS
-		MesPrint("StageSort in thread %d",identity);
+			MesPrint("StageSort in thread %d",identity);
 #elif defined(WITHMPI)
-		MesPrint("StageSort in process %d",PF.me);
+			MesPrint("StageSort in process %d",PF.me);
 #else
-		MesPrint("StageSort");
+			MesPrint("StageSort");
 #endif
-		MUNLOCK(ErrorMessageLock);
+			MUNLOCK(ErrorMessageLock);
+		}
+		else {
+			if ( AT.SubSortStage4Warning == 0 ) {
+				AT.SubSortStage4Warning = 1;
+				MLOCK(ErrorMessageLock);
+				Warning("StageSort in sub-buffer: recommend increasing setup sub-buffer sizes");
+				MUNLOCK(ErrorMessageLock);
+			}
+		}
 		SeekFile(fout->handle,&position,SEEK_END);
 /*
 		No extra compression data has to be written.
@@ -4492,25 +4461,16 @@ void StageSort(FILEHANDLE *fout)
 		S->fPatches[0] = fout->filesize;
 		S->fPatchN = 0;
 
-		if ( AR.FoStage4[0].PObuffer == 0 ) {
-			AR.FoStage4[0].PObuffer = (WORD *)Malloc1(AR.FoStage4[0].POsize*sizeof(WORD)
-												,"Stage 4 buffer");
-			AR.FoStage4[0].POfill   = AR.FoStage4[0].PObuffer;
-			AR.FoStage4[0].POstop   = AR.FoStage4[0].PObuffer
-						 + AR.FoStage4[0].POsize/sizeof(WORD);
-#ifdef WITHPTHREADS
-			AR.FoStage4[0].pthreadslock = dummylock;
-#endif
-		}
-		if ( AR.FoStage4[1].PObuffer == 0 ) {
-			AR.FoStage4[1].PObuffer = (WORD *)Malloc1(AR.FoStage4[1].POsize*sizeof(WORD)
-												,"Stage 4 buffer");
-			AR.FoStage4[1].POfill   = AR.FoStage4[1].PObuffer;
-			AR.FoStage4[1].POstop   = AR.FoStage4[1].PObuffer
-						 + AR.FoStage4[1].POsize/sizeof(WORD);
-#ifdef WITHPTHREADS
-			AR.FoStage4[1].pthreadslock = dummylock;
-#endif
+		{
+			FILEHANDLE *stage4 = &(S->FoStage4);
+			if ( stage4->PObuffer == 0 ) {
+				size_t namesize = strlen(S->file.name) + 4;
+				stage4->PObuffer = (WORD *)Malloc1(stage4->POsize + namesize, "Stage 4 buffer");
+				stage4->POfill = stage4->POfull = stage4->PObuffer;
+				stage4->POstop = stage4->PObuffer + stage4->POsize/sizeof(WORD);
+				stage4->name = (char *)((UBYTE *)stage4->PObuffer + stage4->POsize);
+				snprintf(stage4->name,namesize,"%s.4a",S->file.name);
+			}
 		}
 		S->stage4 = 1;
 	}
@@ -4644,6 +4604,18 @@ void CleanUpSort(int num)
 		for ( i = minnum; i < AN.NumFunSorts; i++ ) {
 			S = AN.FunSorts[i];
 			if ( S ) {
+				if ( S->FoStage4.handle >= 0 ) {
+					UpdateMaxSize();
+#ifdef WITHZLIB
+					ClearSortGZIP(&(S->FoStage4));
+#endif
+					CloseFile(S->FoStage4.handle);
+					remove(S->FoStage4.name);
+					S->FoStage4.handle = -1;
+				}
+				if ( S->FoStage4.PObuffer ) {
+					M_free(S->FoStage4.PObuffer, "CleanUpSort: Stage 4 buffer");
+				}
 				if ( S->file.handle >= 0 ) {
 /*					TruncateFile(S->file.handle); */
 					UpdateMaxSize();
@@ -4684,6 +4656,15 @@ void CleanUpSort(int num)
 		if ( num == 0 ) {
 			S = AN.FunSorts[0];
 			if ( S ) {
+				if ( S->FoStage4.handle >= 0 ) {
+					UpdateMaxSize();
+#ifdef WITHZLIB
+					ClearSortGZIP(&(S->FoStage4));
+#endif
+					CloseFile(S->FoStage4.handle);
+					remove(S->FoStage4.name);
+					S->FoStage4.handle = -1;
+				}
 				if ( S->file.handle >= 0 ) {
 /*					TruncateFile(S->file.handle); */
 					UpdateMaxSize();
@@ -4702,22 +4683,6 @@ void CleanUpSort(int num)
 			}
 		}
 	}
-	for ( i = 0; i < 2; i++ ) {
-		if ( AR.FoStage4[i].handle >= 0 ) {
-			UpdateMaxSize();
-#ifdef WITHZLIB
-			ClearSortGZIP(&(AR.FoStage4[i]));
-#endif
-			CloseFile(AR.FoStage4[i].handle);
-			remove(AR.FoStage4[i].name);
-			AR.FoStage4[i].handle = -1;
-#ifdef GZIPDEBUG
-			MLOCK(ErrorMessageLock);
-			MesPrint("%w CleanUpSort removed stage4 file %s",AR.FoStage4[i].name);
-			MUNLOCK(ErrorMessageLock);
-#endif
-		}
-	}
 }
 
 /*
diff --git a/sources/startup.c b/sources/startup.c
index 126b7d3b..96fdc6bc 100644
--- a/sources/startup.c
+++ b/sources/startup.c
@@ -850,7 +850,7 @@ classic:;
 	*t = 0;
 
 /*
-	Now we should assign a name to the main sort file and the two stage 4 files.
+	Now we should assign a name to the main sort file.
 */
 	AM.S0->file.name = (char *)Malloc1(sizeof(char)*(i+DEFAULTFNAMELENGTH),"name for temporary files");
 	s = (UBYTE *)AM.S0->file.name;
@@ -877,27 +877,13 @@ classic:;
 	}
 #endif
 /*
-	With the stage4 and scratch file names we have to be a bit more careful.
+	With the scratch file names we have to be a bit more careful.
 	They are to be allocated after the threads are initialized when there
 	are threads of course.
 */
 	if ( par == 0 ) {
-		s = (UBYTE *)((void *)(FG.fname2)); i = 0;
-		while ( *s ) { s++; i++; }
-		/* +1 for null terminator */
-		s = (UBYTE *)Malloc1(sizeof(char)*(i+1),"name for stage4 file a");
-		AR.FoStage4[1].name = (char *)s;
-		t = (UBYTE *)FG.fname2;
-		while ( *t ) *s++ = *t++;
-		s[-2] = '4'; s[-1] = 'a'; *s = 0;
 		s = (UBYTE *)((void *)(FG.fname)); i = 0;
 		while ( *s ) { s++; i++; }
-		/* +1 for null terminator */
-		s = (UBYTE *)Malloc1(sizeof(char)*(i+1),"name for stage4 file b");
-		AR.FoStage4[0].name = (char *)s;
-		t = (UBYTE *)FG.fname;
-		while ( *t ) *s++ = *t++;
-		s[-2] = '4'; s[-1] = 'b'; *s = 0;
 		for ( j = 0; j < 3; j++ ) {
 			/* +1 for null terminator */
 			s = (UBYTE *)Malloc1(sizeof(char)*(i+1),"name for scratch file");
@@ -909,23 +895,8 @@ classic:;
 	}
 #ifdef WITHPTHREADS
 	else if ( par == 2 ) {
-		size_t tname;
-		s = (UBYTE *)((void *)(FG.fname2)); i = 0;
-		while ( *s ) { s++; i++; }
-		/* +1 for null terminator, +10 for 32bit int, +1 for "." */
-		tname = sizeof(char)*(i+12);
-		s = (UBYTE *)Malloc1(tname,"name for stage4 file a");
-		snprintf((char *)s,tname,"%s.%d",FG.fname2,AT.identity);
-		s[i-2] = '4'; s[i-1] = 'a';
-		AR.FoStage4[1].name = (char *)s;
 		s = (UBYTE *)((void *)(FG.fname)); i = 0;
 		while ( *s ) { s++; i++; }
-		/* +1 for null terminator, +10 for 32bit int, +1 for "." */
-		tname = sizeof(char)*(i+12);
-		s = (UBYTE *)Malloc1(tname,"name for stage4 file b");
-		snprintf((char *)s,tname,"%s.%d",FG.fname,AT.identity);
-		s[i-2] = '4'; s[i-1] = 'b';
-		AR.FoStage4[0].name = (char *)s;
 		if ( AT.identity == 0 ) {
 			for ( j = 0; j < 3; j++ ) {
 				/* +1 for null terminator */
@@ -1496,8 +1467,6 @@ void IniVars(void)
 	AM.gOutNumberType = RATIONALMODE;
 #ifdef WITHZLIB
 	AR.gzipCompress = GZIPDEFAULT;
-	AR.FoStage4[0].ziobuffer = 0;
-	AR.FoStage4[1].ziobuffer = 0;
 #ifdef WITHZSTD
 	/* Zstd compression is on by default, if we have compiled with it */
 	ZWRAP_useZSTDcompression(1);
diff --git a/sources/structs.h b/sources/structs.h
index 707e3e0f..d92fd15c 100644
--- a/sources/structs.h
+++ b/sources/structs.h
@@ -1113,6 +1113,7 @@ typedef struct PaRtI {
 
 typedef struct sOrT {
     FILEHANDLE file;            /* The own sort file */
+    FILEHANDLE FoStage4;        /* Temporary file for stage 4 and higher */
     POSITION SizeInFile[3];     /* Sizes in the various files */
     POSITION OldPosIn;          /* Sort file fill positions */
     POSITION OldPosOut;
@@ -1990,7 +1991,6 @@ struct S_const {
 struct R_const {
     FILEDATA    StoreData;         /* (O) */
     FILEHANDLE  Fscr[3];           /* (R) Dollars etc play with it too */
-    FILEHANDLE  FoStage4[2];       /* (R) In Sort. Stage 4. */
     POSITION DefPosition;          /* (R) Deferred position of keep brackets. */
     FILEHANDLE *infile;            /* (R) Points alternatingly to Fscr[0] or Fscr[1] */
     FILEHANDLE *outfile;           /* (R) Points alternatingly to Fscr[1] or Fscr[0] */
@@ -2033,7 +2033,6 @@ struct R_const {
     WORD    DeferFlag;             /* (R) For deferred brackets */
     WORD    TePos;                 /* (R) */
     WORD    sLevel;                /* (R) Sorting level */
-    WORD    Stage4Name;            /* (R) Sorting only */
     WORD    GetOneFile;            /* (R) Getting from hide or regular */
     WORD    PolyFun;               /* (C) Number of the PolyFun function */
     WORD    PolyFunInv;            /* (C) Number of the Inverse of the PolyFun function */
@@ -2152,6 +2151,7 @@ struct T_const {
     int     numpoly;
     int     LeaveNegative;
     int     TrimPower;             /* Indicates trimming in polyratfun expansion */
+    int     SubSortStage4Warning;  /* Has the sub-sort stage 4 warning been printed? */
     WORD    small_power_maxx;      /*     size of the cache for small powers  */
     WORD    small_power_maxn;      /*     size of the cache for small powers */
     WORD    dummysubexp[SUBEXPSIZE+4]; /* () used in normal.c */
diff --git a/sources/threads.c b/sources/threads.c
index 9dc5760c..2479b74c 100644
--- a/sources/threads.c
+++ b/sources/threads.c
@@ -402,7 +402,7 @@ ALLPRIVATES *InitializeOneThread(int identity)
 {
 	WORD *t, *ScratchBuf;
 	int i, j, bsize, *bp;
-	LONG ScratchSize[3], IOsize;
+	LONG ScratchSize[3];
 	ALLPRIVATES *B;
 	UBYTE *s;
 
@@ -563,18 +563,6 @@ ALLPRIVATES *InitializeOneThread(int identity)
 	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;
@@ -637,7 +625,6 @@ ALLPRIVATES *InitializeOneThread(int identity)
 /*
 	Still to do: the SS stuff.
 	             the Fscr[3]
-	             the FoStage4[2]
 */
 	if ( AT.WorkSpace == 0 ||
 	     AT.Nest == 0 ||
@@ -1076,7 +1063,6 @@ int LoadOneThread(int from, int identity, THREADBUCKET *thr, int par)
 	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;
@@ -1320,6 +1306,7 @@ void *RunThread(void *dummy)
 				AR.PolyFunExp = AB[0]->R.PolyFunExp;
 				AR.PolyFunVar = AB[0]->R.PolyFunVar;
 				AR.PolyFunPow = AB[0]->R.PolyFunPow;
+				AT.SubSortStage4Warning = 0;
 /*
 				Now fire up the sort buffer.
 */
@@ -1608,6 +1595,7 @@ bucketstolen:;
 				AR.CurExpr = i;
 				AR.SortType = AC.SortType;
 				AR.expchanged = 0;
+				AT.SubSortStage4Warning = 0;
 				if ( ( e->vflags & ISFACTORIZED ) != 0 ) {
 					AR.BracketOn = 1;
 					AT.BrackBuf = AM.BracketFactors;
@@ -4773,4 +4761,4 @@ extern void optimize_expression_given_Horner_threaded(void) {
   	#] optimize_expression_given_Horner_threaded : 
 */
 
-#endif
\ No newline at end of file
+#endif
diff --git a/sources/tools.c b/sources/tools.c
index 65d3df53..dd2eaf0a 100644
--- a/sources/tools.c
+++ b/sources/tools.c
@@ -966,8 +966,6 @@ void StartFiles(void)
 	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;
 	AR.infile = &(AR.Fscr[0]);
 	AR.outfile = &(AR.Fscr[1]);
 	AR.hidefile = &(AR.Fscr[2]);
@@ -1614,8 +1612,9 @@ int CloseChannel(char *name)
 void UpdateMaxSize(void)
 {
 	POSITION position, sumsize;
-	int i;
+	int i, k;
 	FILEHANDLE *scr;
+	SORTING *sort;
 #ifdef WITHMPI
 	/* Currently, it works only on the master. The sort files on the slaves
 	 * are ignored. (TU 11 Oct 2011) */
@@ -1649,29 +1648,30 @@ void UpdateMaxSize(void)
 		ALLPRIVATES *B;
 		for ( j = 0; j < AM.totalnumberofthreads; j++ ) {
 			B = AB[j];
-			if ( AT.SS && AT.SS->file.handle >= 0 ) {
-				position = AT.SS->file.filesize;
-/*
-MLOCK(ErrorMessageLock);
-MesPrint("%d: %10p",j,&(AT.SS->file.filesize));
-MUNLOCK(ErrorMessageLock);
-*/
-				ADD2POS(sumsize,position);
-			}
-			if ( AR.FoStage4[0].handle >= 0 ) {
-				position = AR.FoStage4[0].filesize;
-				ADD2POS(sumsize,position);
+			for ( k = 0; k < AN.NumFunSorts; k++ ) {
+				sort = AN.FunSorts[k];
+				if ( sort && sort->file.handle >= 0 ) {
+					position = sort->file.filesize;
+					ADD2POS(sumsize,position);
+				}
+				if ( sort && sort->FoStage4.handle >= 0 ) {
+					position = sort->FoStage4.filesize;
+					ADD2POS(sumsize,position);
+				}
 			}
 		}
 	}
 #else
-	if ( AT.SS && AT.SS->file.handle >= 0 ) {
-		position = AT.SS->file.filesize;
-		ADD2POS(sumsize,position);
-	}
-	if ( AR.FoStage4[0].handle >= 0 ) {
-		position = AR.FoStage4[0].filesize;
-		ADD2POS(sumsize,position);
+	for ( k = 0; k < AN.NumFunSorts; k++ ) {
+		sort = AN.FunSorts[k];
+		if ( sort && sort->file.handle >= 0 ) {
+			position = sort->file.filesize;
+			ADD2POS(sumsize,position);
+		}
+		if ( sort && sort->FoStage4.handle >= 0 ) {
+			position = sort->FoStage4.filesize;
+			ADD2POS(sumsize,position);
+		}
 	}
 #endif
 /*