===================================================================
RCS file: /home/cvs/OpenXM_contrib2/asir2000/engine/distm.c,v
retrieving revision 1.6
retrieving revision 1.17
diff -u -p -r1.6 -r1.17
--- OpenXM_contrib2/asir2000/engine/distm.c	2000/08/22 05:04:05	1.6
+++ OpenXM_contrib2/asir2000/engine/distm.c	2003/07/23 01:22:50	1.17
@@ -45,38 +45,15 @@
  * DEVELOPER SHALL HAVE NO LIABILITY IN CONNECTION WITH THE USE,
  * PERFORMANCE OR NON-PERFORMANCE OF THE SOFTWARE.
  *
- * $OpenXM: OpenXM_contrib2/asir2000/engine/distm.c,v 1.5 2000/08/21 08:31:27 noro Exp $ 
+ * $OpenXM: OpenXM_contrib2/asir2000/engine/distm.c,v 1.16 2003/07/22 10:11:43 noro Exp $ 
 */
 #include "ca.h"
 #include "inline.h"
 
-#define NV(p) ((p)->nv)
-#define C(p) ((p)->c)
-#if 0
-#define ITOS(p) (((unsigned int)(p))>>1)
-#define STOI(i) ((P)((((unsigned int)(i))<<1)|1))
-#else
-#define ITOS(p) (((unsigned int)(p))&0x7fffffff)
-#define STOI(i) ((P)((unsigned int)(i)|0x80000000))
-#endif
-
 extern int (*cmpdl)();
 extern int do_weyl;
 
-void comm_mulmd();
-void weyl_mulmd();
-void weyl_mulmdm();
-void weyl_mulmmm();
-void _comm_mulmd();
-void _weyl_mulmd();
-void _weyl_mulmmm();
-void _weyl_mulmdm();
-
-void ptomd(vl,mod,dvl,p,pr)
-VL vl,dvl;
-int mod;
-P p;
-DP *pr;
+void ptomd(VL vl,int mod,VL dvl,P p,DP *pr)
 {
 	P t;
 
@@ -84,11 +61,7 @@ DP *pr;
 	mptomd(vl,mod,dvl,t,pr);
 }
 
-void mptomd(vl,mod,dvl,p,pr)
-VL vl,dvl;
-int mod;
-P p;
-DP *pr;
+void mptomd(VL vl,int mod,VL dvl,P p,DP *pr)
 {
 	int n,i;
 	VL tvl;
@@ -118,7 +91,8 @@ DP *pr;
 			} else {
 				for ( dc = DC(p), s = 0; dc; dc = NEXT(dc) ) {
 					mptomd(vl,mod,dvl,COEF(dc),&t);
-					NEWDL(d,n); d->td = QTOS(DEG(dc)); d->d[i] = d->td;
+					NEWDL(d,n); d->d[i] = QTOS(DEG(dc));
+					d->td = MUL_WEIGHT(d->d[i],i);
 					NEWMP(m); m->dl = d; C(m) = (P)ONEM; NEXT(m) = 0; MKDP(n,m,u);
 					comm_mulmd(vl,mod,t,u,&r); addmd(vl,mod,r,s,&t); s = t;
 				}
@@ -128,11 +102,7 @@ DP *pr;
 	}
 }
 
-void mdtop(vl,mod,dvl,p,pr)
-VL vl,dvl;
-int mod;
-DP p;
-P *pr;
+void mdtop(VL vl,int mod,VL dvl,DP p,P *pr)
 {
 	int n,i;
 	DL d;
@@ -156,10 +126,7 @@ P *pr;
 	}
 }
 
-void addmd(vl,mod,p1,p2,pr)
-VL vl;
-int mod;
-DP p1,p2,*pr;
+void addmd(VL vl,int mod,DP p1,DP p2,DP *pr)
 {
 	int n;
 	MP m1,m2,mr,mr0;
@@ -217,10 +184,7 @@ DP p1,p2,*pr;
 	}
 }
 
-void submd(vl,mod,p1,p2,pr)
-VL vl;
-int mod;
-DP p1,p2,*pr;
+void submd(VL vl,int mod,DP p1,DP p2,DP *pr)
 {
 	DP t;
 
@@ -231,9 +195,7 @@ DP p1,p2,*pr;
 	}
 }
 
-void chsgnmd(mod,p,pr)
-int mod;
-DP p,*pr;
+void chsgnmd(int mod,DP p,DP *pr)
 {
 	MP m,mr,mr0;
 
@@ -249,10 +211,7 @@ DP p,*pr;
 	}
 }
 
-void mulmd(vl,mod,p1,p2,pr)
-VL vl;
-int mod;
-DP p1,p2,*pr;
+void mulmd(VL vl,int mod,DP p1,DP p2,DP *pr)
 {
 	if ( !do_weyl )
 		comm_mulmd(vl,mod,p1,p2,pr);
@@ -260,10 +219,7 @@ DP p1,p2,*pr;
 		weyl_mulmd(vl,mod,p1,p2,pr);
 }
 
-void comm_mulmd(vl,mod,p1,p2,pr)
-VL vl;
-int mod;
-DP p1,p2,*pr;
+void comm_mulmd(VL vl,int mod,DP p1,DP p2,DP *pr)
 {
 	MP m;
 	DP s,t,u;
@@ -299,14 +255,11 @@ DP p1,p2,*pr;
 	}
 }
 
-void weyl_mulmd(vl,mod,p1,p2,pr)
-VL vl;
-int mod;
-DP p1,p2,*pr;
+void weyl_mulmd(VL vl,int mod,DP p1,DP p2,DP *pr)
 {
 	MP m;
 	DP s,t,u;
-	int i,l,l1;
+	int i,l;
 	static MP *w;
 	static int wlen;
 
@@ -333,12 +286,7 @@ DP p1,p2,*pr;
 	}
 }
 
-void mulmdm(vl,mod,p,m0,pr)
-VL vl;
-int mod;
-DP p;
-MP m0;
-DP *pr;
+void mulmdm(VL vl,int mod,DP p,MP m0,DP *pr)
 {
 	MP m,mr,mr0;
 	P c;
@@ -365,12 +313,7 @@ DP *pr;
 	}
 }
 
-void weyl_mulmdm(vl,mod,p,m0,pr)
-VL vl;
-int mod;
-DP p;
-MP m0;
-DP *pr;
+void weyl_mulmdm(VL vl,int mod,DP p,MP m0,DP *pr)
 {
 	DP r,t,t1;
 	MP m;
@@ -402,17 +345,12 @@ DP *pr;
 
 /* m0 = x0^d0*x1^d1*... * dx0^d(n/2)*dx1^d(n/2+1)*... */
 
-void weyl_mulmmm(vl,mod,m0,m1,n,pr)
-VL vl;
-int mod;
-MP m0,m1;
-int n;
-DP *pr;
+void weyl_mulmmm(VL vl,int mod,MP m0,MP m1,int n,DP *pr)
 {
-	MP m,mr,mr0;
+	MP mr,mr0;
 	MQ mq;
 	DP r,t,t1;
-	P c,c0,c1,cc;
+	P c,c0,c1;
 	DL d,d0,d1;
 	int i,j,a,b,k,l,n2,s,min;
 	static int *tab;
@@ -437,8 +375,12 @@ DP *pr;
 		for ( i = 0; i < n2; i++ ) {
 			a = d0->d[i]; b = d1->d[n2+i];
 			k = d0->d[n2+i]; l = d1->d[i];
+
 			/* degree of xi^a*(Di^k*xi^l)*Di^b */
-			s = a+k+l+b;
+			a += l;
+			b += k;
+			s = MUL_WEIGHT(a,i)+MUL_WEIGHT(b,n2+i);
+
 			/* compute xi^a*(Di^k*xi^l)*Di^b */
 			min = MIN(k,l);
 
@@ -451,16 +393,16 @@ DP *pr;
 			if ( n & 1 )
 				for ( mr0 = 0, j = 0; j <= min; j++ ) {
 					NEXTMP(mr0,mr); NEWDL(d,n);
-					d->d[i] = l-j+a; d->d[n2+i] = k-j+b;
+					d->d[i] = a-j; d->d[n2+i] = b-j;
 					d->td = s;
-					d->d[n-1] = s-(d->d[i]+d->d[n2+i]); 
+					d->d[n-1] = s-(MUL_WEIGHT(a-j,i)+MUL_WEIGHT(b-j,n2+i));
 					STOMQ(tab[j],mq); mr->c = (P)mq; mr->dl = d;
 				}
 			else
 				for ( mr0 = 0, s = 0, j = 0; j <= min; j++ ) {
 					NEXTMP(mr0,mr); NEWDL(d,n);
-					d->d[i] = l-j+a; d->d[n2+i] = k-j+b;
-					d->td = d->d[i]+d->d[n2+i]; /* XXX */
+					d->d[i] = a-j; d->d[n2+i] = b-j;
+					d->td = MUL_WEIGHT(a-j,i)+MUL_WEIGHT(b-j,n2+i); /* XXX */
 					s = MAX(s,d->td); /* XXX */
 					STOMQ(tab[j],mq); mr->c = (P)mq; mr->dl = d;
 				}
@@ -476,12 +418,7 @@ DP *pr;
 	}
 }
 
-void mulmdc(vl,mod,p,c,pr)
-VL vl;
-int mod;
-DP p;
-P c;
-DP *pr;
+void mulmdc(VL vl,int mod,DP p,P c,DP *pr)
 {
 	MP m,mr,mr0;
 	int t;
@@ -505,12 +442,7 @@ DP *pr;
 	}
 }
 
-void divsmdc(vl,mod,p,c,pr)
-VL vl;
-int mod;
-DP p;
-P c;
-DP *pr;
+void divsmdc(VL vl,int mod,DP p,P c,DP *pr)
 {
 	MP m,mr,mr0;
 
@@ -528,43 +460,100 @@ DP *pr;
 	}
 }
 
-void _mdtop(vl,mod,dvl,p,pr)
-VL vl,dvl;
-int mod;
-DP p;
-P *pr;
+void _dtop_mod(VL vl,VL dvl,DP p,P *pr)
 {
 	int n,i;
 	DL d;
 	MP m;
 	P r,s,t,u,w;
 	Q q;
-	MQ tq;
 	VL tvl;
 
 	if ( !p )
 		*pr = 0;
 	else {
 		for ( n = p->nv, m = BDY(p), s = 0; m; m = NEXT(m) ) {
-			STOMQ(ITOS(C(m)),tq); t = (P)tq;
+			i = ITOS(m->c); STOQ(i,q); t = (P)q;
 			for ( i = 0, d = m->dl, tvl = dvl;
 				i < n; tvl = NEXT(tvl), i++ ) {
-				MKV(tvl->v,r); STOQ(d->d[i],q); pwrmp(vl,mod,r,q,&u);
-				mulmp(vl,mod,t,u,&w); t = w;
+				MKV(tvl->v,r); STOQ(d->d[i],q); pwrp(vl,r,q,&u);
+				mulp(vl,t,u,&w); t = w;
 			}
-			addmp(vl,mod,s,t,&u); s = u;
+			addp(vl,s,t,&u); s = u;
 		}
-		mptop(s,pr);
+		*pr = s;
 	}
 }
 
-void _addmd(vl,mod,p1,p2,pr)
-VL vl;
-int mod;
-DP p1,p2,*pr;
+void _dp_mod(DP p,int mod,NODE subst,DP *rp)
 {
+	MP m,mr,mr0;
+	P t,s;
+	NODE tn;
+
+	if ( !p )
+		*rp = 0;
+	else {
+		for ( mr0 = 0, m = BDY(p); m; m = NEXT(m) ) {
+			for ( tn = subst, s = m->c; tn; tn = NEXT(NEXT(tn)) ) {
+				substp(CO,s,BDY(tn),BDY(NEXT(tn)),&t);
+				s = t;
+			}
+			ptomp(mod,s,&t);
+			if ( t ) {
+				NEXTMP(mr0,mr); mr->c = STOI(CONT((MQ)t)); mr->dl = m->dl;
+			}
+		}
+		if ( mr0 ) {
+			NEXT(mr) = 0; MKDP(p->nv,mr0,*rp); (*rp)->sugar = p->sugar;
+		} else
+			*rp = 0;
+	}
+}
+
+void _dp_monic(DP p,int mod,DP *rp)
+{
+	MP m,mr,mr0;
+	int c,c1;
+
+	if ( !p )
+		*rp = 0;
+	else {
+		c = invm(ITOS(BDY(p)->c),mod);
+		for ( mr0 = 0, m = BDY(p); m; m = NEXT(m) ) {
+			c1 = dmar(ITOS(m->c),c,0,mod);
+			NEXTMP(mr0,mr); mr->c = STOI(c1); mr->dl = m->dl;
+		}
+		NEXT(mr) = 0; MKDP(p->nv,mr0,*rp); (*rp)->sugar = p->sugar;
+	}
+}
+
+void _printdp(DP d)
+{
+	int n,i;
+	MP m;
+	DL dl;
+
+	if ( !d ) {
+		printf("0"); return;
+	}
+	for ( n = d->nv, m = BDY(d); m; m = NEXT(m) ) {
+		printf("%d*<<",ITOS(m->c));
+		for ( i = 0, dl = m->dl; i < n-1; i++ )
+			printf("%d,",dl->d[i]);
+		printf("%d",dl->d[i]);
+		printf(">>");
+		if ( NEXT(m) )
+			printf("+");
+	}
+}
+
+/* merge p1 and p2 into pr */
+
+void addmd_destructive(int mod,DP p1,DP p2,DP *pr)
+{
 	int n;
-	MP m1,m2,mr,mr0;
+	MP m1,m2,mr,mr0,s;
 	int t;
 
 	if ( !p1 )
@@ -578,16 +567,18 @@ DP p1,p2,*pr;
 					t = (ITOS(C(m1))+ITOS(C(m2))) - mod;
 					if ( t < 0 )
 						t += mod;
+					s = m1; m1 = NEXT(m1);
 					if ( t ) {
-						NEXTMP(mr0,mr); mr->dl = m1->dl; C(mr) = STOI(t);
+						NEXTMP2(mr0,mr,s); C(mr) = STOI(t);
 					}
-					m1 = NEXT(m1); m2 = NEXT(m2); break;
+					s = m2; m2 = NEXT(m2);
+					break;
 				case 1:
-					NEXTMP(mr0,mr); mr->dl = m1->dl; C(mr) = C(m1);
-					m1 = NEXT(m1); break;
+					s = m1; m1 = NEXT(m1); NEXTMP2(mr0,mr,s);
+					break;
 				case -1:
-					NEXTMP(mr0,mr); mr->dl = m2->dl; C(mr) = C(m2);
-					m2 = NEXT(m2); break;
+					s = m2; m2 = NEXT(m2); NEXTMP2(mr0,mr,s);
+					break;
 			}
 		if ( !mr0 )
 			if ( m1 )
@@ -610,53 +601,15 @@ DP p1,p2,*pr;
 	}
 }
 
-void _submd(vl,mod,p1,p2,pr)
-VL vl;
-int mod;
-DP p1,p2,*pr;
+void mulmd_dup(int mod,DP p1,DP p2,DP *pr)
 {
-	DP t;
-
-	if ( !p2 )
-		*pr = p1;
-	else {
-		_chsgnmd(mod,p2,&t); _addmd(vl,mod,p1,t,pr);
-	}
-}
-
-void _chsgnmd(mod,p,pr)
-int mod;
-DP p,*pr;
-{
-	MP m,mr,mr0;
-
-	if ( !p )
-		*pr = 0;
-	else {
-		for ( mr0 = 0, m = BDY(p); m; m = NEXT(m) ) {
-			NEXTMP(mr0,mr); C(mr) = STOI(mod - ITOS(C(m))); mr->dl = m->dl;
-		}
-		NEXT(mr) = 0; MKDP(NV(p),mr0,*pr);
-		if ( *pr )
-			(*pr)->sugar = p->sugar;
-	}
-}
-
-void _mulmd(vl,mod,p1,p2,pr)
-VL vl;
-int mod;
-DP p1,p2,*pr;
-{
 	if ( !do_weyl )
-		_comm_mulmd(vl,mod,p1,p2,pr);
+		comm_mulmd_dup(mod,p1,p2,pr);
 	else
-		_weyl_mulmd(vl,mod,p1,p2,pr);
+		weyl_mulmd_dup(mod,p1,p2,pr);
 }
 
-void _comm_mulmd(vl,mod,p1,p2,pr)
-VL vl;
-int mod;
-DP p1,p2,*pr;
+void comm_mulmd_dup(int mod,DP p1,DP p2,DP *pr)
 {
 	MP m;
 	DP s,t,u;
@@ -681,53 +634,46 @@ DP p1,p2,*pr;
 		for ( m = BDY(p2), i = 0; i < l; m = NEXT(m), i++ )
 			w[i] = m;
 		for ( s = 0, i = l-1; i >= 0; i-- ) {
-			_mulmdm(vl,mod,p1,w[i],&t); _addmd(vl,mod,s,t,&u); s = u;
+			mulmdm_dup(mod,p1,w[i],&t); addmd_destructive(mod,s,t,&u); s = u;
 		}
 		bzero(w,l*sizeof(MP));
 		*pr = s;
 	}
 }
 
-void _weyl_mulmd(vl,mod,p1,p2,pr)
-VL vl;
-int mod;
-DP p1,p2,*pr;
+void weyl_mulmd_dup(int mod,DP p1,DP p2,DP *pr)
 {
 	MP m;
 	DP s,t,u;
-	int i,l,l1;
+	int i,l;
 	static MP *w;
 	static int wlen;
 
 	if ( !p1 || !p2 )
 		*pr = 0;
 	else {
-		for ( m = BDY(p2), l = 0; m; m = NEXT(m), l++ );
+		for ( m = BDY(p1), l = 0; m; m = NEXT(m), l++ );
 		if ( l > wlen ) {
 			if ( w ) GC_free(w);
 			w = (MP *)MALLOC(l*sizeof(MP));
 			wlen = l;
 		}
-		for ( m = BDY(p2), i = 0; i < l; m = NEXT(m), i++ )
+		for ( m = BDY(p1), i = 0; i < l; m = NEXT(m), i++ )
 			w[i] = m;
 		for ( s = 0, i = l-1; i >= 0; i-- ) {
-			_weyl_mulmdm(vl,mod,p1,w[i],&t); _addmd(vl,mod,s,t,&u); s = u;
+			weyl_mulmdm_dup(mod,w[i],p2,&t); addmd_destructive(mod,s,t,&u); s = u;
 		}
 		bzero(w,l*sizeof(MP));
 		*pr = s;
 	}
 }
 
-void _mulmdm(vl,mod,p,m0,pr)
-VL vl;
-int mod;
-DP p;
-MP m0;
-DP *pr;
+void mulmdm_dup(int mod,DP p,MP m0,DP *pr)
 {
 	MP m,mr,mr0;
-	DL d;
-	int c,n,r;
+	DL d,dt,dm;
+	int c,n,i;
+	int *pt,*p1,*p2;
 
 	if ( !p )
 		*pr = 0;
@@ -736,7 +682,11 @@ DP *pr;
 			m; m = NEXT(m) ) {
 			NEXTMP(mr0,mr);
 			C(mr) = STOI(dmar(ITOS(C(m)),c,0,mod));
-			adddl(n,m->dl,d,&mr->dl);
+			NEWDL_NOINIT(dt,n); mr->dl = dt;
+			dm = m->dl;
+			dt->td = d->td + dm->td;
+			for ( i = 0, pt = dt->d, p1=d->d, p2 = dm->d; i < n; i++ )
+				*pt++ = *p1++ + *p2++;
 		}
 		NEXT(mr) = 0; MKDP(NV(p),mr0,*pr);
 		if ( *pr )
@@ -744,18 +694,16 @@ DP *pr;
 	}
 }
 
-void _weyl_mulmdm(vl,mod,p,m0,pr)
-VL vl;
-int mod;
-DP p;
-MP m0;
-DP *pr;
+void weyl_mulmdm_dup(int mod,MP m0,DP p,DP *pr)
 {
 	DP r,t,t1;
 	MP m;
-	int n,l,i;
-	static MP *w;
+	DL d0;
+	int n,n2,l,i,j,tlen;
+	static MP *w,*psum;
+	static struct cdlm *tab;
 	static int wlen;
+	static int rtlen;
 
 	if ( !p )
 		*pr = 0;
@@ -768,11 +716,34 @@ DP *pr;
 		}
 		for ( m = BDY(p), i = 0; i < l; m = NEXT(m), i++ )
 			w[i] = m;
-		for ( r = 0, i = l-1, n = NV(p); i >= 0; i-- ) {
-			_weyl_mulmmm(vl,mod,w[i],m0,n,&t);
-			_addmd(vl,mod,r,t,&t1); r = t1;
+		n = NV(p); n2 = n>>1;
+		d0 = m0->dl;
+
+		for ( i = 0, tlen = 1; i < n2; i++ )
+			tlen *= d0->d[n2+i]+1;
+		if ( tlen > rtlen ) {
+			if ( tab ) GC_free(tab);
+			if ( psum ) GC_free(psum);
+			rtlen = tlen;
+			tab = (struct cdlm *)MALLOC(rtlen*sizeof(struct cdlm));
+			psum = (MP *)MALLOC(rtlen*sizeof(MP));
 		}
-		bzero(w,l*sizeof(MP));
+		bzero(psum,tlen*sizeof(MP));
+		for ( i = l-1; i >= 0; i-- ) {
+			bzero(tab,tlen*sizeof(struct cdlm));
+			weyl_mulmmm_dup(mod,m0,w[i],n,tab,tlen);
+			for ( j = 0; j < tlen; j++ ) {
+				if ( tab[j].c ) {
+					NEWMP(m); m->dl = tab[j].d;
+					C(m) = STOI(tab[j].c); NEXT(m) = psum[j];
+					psum[j] = m;
+				}
+			}
+		}
+		for ( j = tlen-1, r = 0; j >= 0; j-- ) 
+			if ( psum[j] ) {
+				MKDP(n,psum[j],t); addmd_destructive(mod,r,t,&t1); r = t1;
+			}
 		if ( r )
 			r->sugar = p->sugar + m0->dl->td;
 		*pr = r;
@@ -781,255 +752,136 @@ DP *pr;
 
 /* m0 = x0^d0*x1^d1*... * dx0^d(n/2)*dx1^d(n/2+1)*... */
 
-void _weyl_mulmmm(vl,mod,m0,m1,n,pr)
-VL vl;
-int mod;
-MP m0,m1;
-int n;
-DP *pr;
+void weyl_mulmmm_dup(int mod,MP m0,MP m1,int n,struct cdlm *rtab,int rtablen)
 {
-	MP m,mr,mr0;
-	DP r,t,t1;
-	int c,c0,c1,cc;
-	DL d,d0,d1;
-	int i,j,a,b,k,l,n2,s,min,h;
-	static int *tab;
+	int c,c0,c1;
+	DL d,d0,d1,dt;
+	int i,j,a,b,k,l,n2,s,min,curlen;
+	struct cdlm *p;
+	static int *ctab;
+	static struct cdlm *tab;
 	static int tablen;
+	static struct cdlm *tmptab;
+	static int tmptablen;
 
-	if ( !m0 || !m1 )
-		*pr = 0;
-	else {
-		c0 = ITOS(C(m0)); c1 = ITOS(C(m1));
-		c = dmar(c0,c1,0,mod);
-		d0 = m0->dl; d1 = m1->dl;
-		n2 = n>>1;
-
-		NEWDL(d,n);
-		if ( n & 1 )
-			/* offset of h-degree */
-			d->td = d->d[n-1] = d0->d[n-1]+d1->d[n-1];
-		else
-			d->td = 0;
-		NEWMP(mr); mr->c = STOI(c); mr->dl = d; NEXT(mr) = 0;
-		MKDP(n,mr,r); r->sugar = d->td;
-
-		/* homogenized computation; dx-xd=h^2 */
-		for ( i = 0; i < n2; i++ ) {
-			a = d0->d[i]; b = d1->d[n2+i];
-			k = d0->d[n2+i]; l = d1->d[i];
-			/* degree of xi^a*(Di^k*xi^l)*Di^b */
-			s = a+k+l+b;
-			/* compute xi^a*(Di^k*xi^l)*Di^b */
-			min = MIN(k,l);
-
-			if ( min+1 > tablen ) {
-				if ( tab ) GC_free(tab);
-				tab = (int *)MALLOC((min+1)*sizeof(int));
-				tablen = min+1;
-			}
-			mkwcm(k,l,mod,tab);
-			if ( n & 1 )
-				for ( mr0 = 0, j = 0; j <= min; j++ ) {
-					NEXTMP(mr0,mr); NEWDL(d,n);
-					d->d[i] = l-j+a; d->d[n2+i] = k-j+b;
-					d->td = s;
-					d->d[n-1] = s-(d->d[i]+d->d[n2+i]); 
-					mr->c = STOI(tab[j]); mr->dl = d;
-				}
-			else
-				for ( mr0 = 0, s = 0, j = 0; j <= min; j++ ) {
-					NEXTMP(mr0,mr); NEWDL(d,n);
-					d->d[i] = l-j+a; d->d[n2+i] = k-j+b;
-					d->td = d->d[i]+d->d[n2+i]; /* XXX */
-					s = MAX(s,d->td); /* XXX */
-					mr->c = STOI(tab[j]); mr->dl = d;
-				}
-			bzero(tab,(min+1)*sizeof(int));
-			if ( mr0 )
-				NEXT(mr) = 0;
-			MKDP(n,mr0,t);
-			if ( t )
-				t->sugar = s;
-			_comm_mulmd(vl,mod,r,t,&t1); r = t1;
-		}
-		*pr = r;
+	if ( !m0 || !m1 ) {
+		rtab[0].c = 0;
+		rtab[0].d = 0;
+		return;
 	}
-}
+	c0 = ITOS(C(m0)); c1 = ITOS(C(m1));
+	c = dmar(c0,c1,0,mod);
+	d0 = m0->dl; d1 = m1->dl;
+	n2 = n>>1;
+	curlen = 1;
 
-void _dtop_mod(vl,dvl,p,pr)
-VL vl,dvl;
-DP p;
-P *pr;
-{
-	int n,i;
-	DL d;
-	MP m;
-	P r,s,t,u,w;
-	Q q;
-	VL tvl;
+	NEWDL(d,n);
+	if ( n & 1 )
+		/* offset of h-degree */
+		d->td = d->d[n-1] = d0->d[n-1]+d1->d[n-1];
+	else
+		d->td = 0;
+	rtab[0].c = c;
+	rtab[0].d = d;
 
-	if ( !p )
-		*pr = 0;
-	else {
-		for ( n = p->nv, m = BDY(p), s = 0; m; m = NEXT(m) ) {
-			i = ITOS(m->c); STOQ(i,q); t = (P)q;
-			for ( i = 0, d = m->dl, tvl = dvl;
-				i < n; tvl = NEXT(tvl), i++ ) {
-				MKV(tvl->v,r); STOQ(d->d[i],q); pwrp(vl,r,q,&u);
-				mulp(vl,t,u,&w); t = w;
-			}
-			addp(vl,s,t,&u); s = u;
-		}
-		*pr = s;
+	if ( rtablen > tmptablen ) {
+		if ( tmptab ) GC_free(tmptab);
+		tmptab = (struct cdlm *)MALLOC(rtablen*sizeof(struct cdlm));
+		tmptablen = rtablen; 
 	}
-}
 
-void _dp_red_mod(p1,p2,mod,rp)
-DP p1,p2;
-int mod;
-DP *rp;
-{
-	int i,n;
-	DL d1,d2,d;
-	MP m;
-	DP t,s;
-	int c,c1;
+	for ( i = 0; i < n2; i++ ) {
+		a = d0->d[i]; b = d1->d[n2+i];
+		k = d0->d[n2+i]; l = d1->d[i];
 
-	n = p1->nv; d1 = BDY(p1)->dl; d2 = BDY(p2)->dl;
-	NEWDL(d,n); d->td = d1->td - d2->td;
-	for ( i = 0; i < n; i++ )
-		d->d[i] = d1->d[i]-d2->d[i];
-	c = invm(ITOS(BDY(p2)->c),mod); c1 = dmar(c,ITOS(BDY(p1)->c),0,mod);
-	NEWMP(m); m->dl = d; m->c = STOI(mod-c1); NEXT(m) = 0;
-	MKDP(n,m,s); s->sugar = d->td;
-	_mulmd(CO,mod,s,p2,&t); _addmd(CO,mod,p1,t,rp);
-}
+		/* degree of xi^a*(Di^k*xi^l)*Di^b */
+		a += l;
+		b += k;
+		s = MUL_WEIGHT(a,i)+MUL_WEIGHT(b,n2+i);
 
-void _dp_mod(p,mod,subst,rp)
-DP p;
-int mod;
-NODE subst;
-DP *rp;
-{
-	MP m,mr,mr0;
-	P t,s;
-	NODE tn;
-
-	if ( !p )
-		*rp = 0;
-	else {
-		for ( mr0 = 0, m = BDY(p); m; m = NEXT(m) ) {
-			for ( tn = subst, s = m->c; tn; tn = NEXT(NEXT(tn)) ) {
-				substp(CO,s,BDY(tn),BDY(NEXT(tn)),&t);
-				s = t;
+		if ( !k || !l ) {
+			for ( j = 0, p = rtab; j < curlen; j++, p++ ) {
+				if ( p->c ) {
+					dt = p->d;
+					dt->d[i] = a;
+					dt->d[n2+i] = b;
+					dt->td += s;
+				}
 			}
-			ptomp(mod,s,&t);
-			if ( t ) {
-				NEXTMP(mr0,mr); mr->c = STOI(CONT((MQ)t)); mr->dl = m->dl;
-			}
+			curlen *= k+1;
+			continue;
 		}
-		if ( mr0 ) {
-			NEXT(mr) = 0; MKDP(p->nv,mr0,*rp); (*rp)->sugar = p->sugar;
-		} else
-			*rp = 0;
+		if ( k+1 > tablen ) {
+			if ( tab ) GC_free(tab);
+			if ( ctab ) GC_free(ctab);
+			tablen = k+1;
+			tab = (struct cdlm *)MALLOC(tablen*sizeof(struct cdlm));
+			ctab = (int *)MALLOC(tablen*sizeof(int));
+		}
+		/* compute xi^a*(Di^k*xi^l)*Di^b */
+		min = MIN(k,l);
+		mkwcm(k,l,mod,ctab);
+		bzero(tab,(k+1)*sizeof(struct cdlm));
+		/* n&1 != 0 => homogenized computation; dx-xd=h^2 */
+		if ( n & 1 )
+			for ( j = 0; j <= min; j++ ) {
+				NEWDL(d,n);
+				d->d[i] = a-j; d->d[n2+i] = b-j;
+				d->td = s;
+				d->d[n-1] = s-(MUL_WEIGHT(a-j,i)+MUL_WEIGHT(b-j,n2+i));
+				tab[j].d = d;
+				tab[j].c = ctab[j];
+			}
+		else
+			for ( j = 0; j <= min; j++ ) {
+				NEWDL(d,n);
+				d->d[i] = a-j; d->d[n2+i] = b-j;
+				d->td = MUL_WEIGHT(a-j,i)+MUL_WEIGHT(b-j,n2+i); /* XXX */
+				tab[j].d = d;
+				tab[j].c = ctab[j];
+			}
+		comm_mulmd_tab_destructive(mod,n,rtab,curlen,tab,k+1);
+		curlen *= k+1;
 	}
 }
 
-void _dp_monic(p,mod,rp)
-DP p;
-int mod;
-DP *rp;
+void comm_mulmd_tab_destructive(int mod,int nv,struct cdlm *t,int n,struct cdlm *t1,int n1)
 {
-	MP m,mr,mr0;
-	int c,c1;
-	NODE tn;
+	int i,j;
+	struct cdlm *p;
+	int c;
+	DL d;
 
-	if ( !p )
-		*rp = 0;
-	else {
-		c = invm(ITOS(BDY(p)->c),mod);
-		for ( mr0 = 0, m = BDY(p); m; m = NEXT(m) ) {
-			c1 = dmar(ITOS(m->c),c,0,mod);
-			NEXTMP(mr0,mr); mr->c = STOI(c1); mr->dl = m->dl;
+	for ( j = 1, p = t+n; j < n1; j++ ) {
+		c = t1[j].c;
+		d = t1[j].d;
+		if ( !c )
+			break;
+		for ( i = 0; i < n; i++, p++ ) {
+			if ( t[i].c ) {
+				p->c = dmar(t[i].c,c,0,mod);
+				adddl_dup(nv,t[i].d,d,&p->d);
+			}
 		}
-		NEXT(mr) = 0; MKDP(p->nv,mr0,*rp); (*rp)->sugar = p->sugar;
 	}
+	c = t1[0].c;
+	d = t1[0].d;
+	for ( i = 0, p = t; i < n; i++, p++ )
+		if ( t[i].c ) {
+			p->c = dmar(t[i].c,c,0,mod);
+			/* t[i].d += d */
+			adddl_destructive(nv,t[i].d,d);
+		}
 }
 
-void _dp_sp_mod(p1,p2,mod,rp)
-DP p1,p2;
-int mod;
-DP *rp;
+void adddl_dup(int n,DL d1,DL d2,DL *dr)
 {
-	int i,n,td;
-	int *w;
-	DL d1,d2,d;
-	MP m;
-	DP t,s,u;
+	DL dt;
+	int i;
 
-	n = p1->nv; d1 = BDY(p1)->dl; d2 = BDY(p2)->dl;
-	w = (int *)ALLOCA(n*sizeof(int));
-	for ( i = 0, td = 0; i < n; i++ ) {
-		w[i] = MAX(d1->d[i],d2->d[i]); td += w[i];
-	}
-	NEWDL(d,n); d->td = td - d1->td;
+	NEWDL(dt,n);
+	*dr = dt;
+	dt->td = d1->td + d2->td;
 	for ( i = 0; i < n; i++ )
-		d->d[i] = w[i] - d1->d[i];
-	NEWMP(m); m->dl = d; m->c = BDY(p2)->c; NEXT(m) = 0;
-	MKDP(n,m,s); s->sugar = d->td; _mulmd(CO,mod,s,p1,&t);
-	NEWDL(d,n); d->td = td - d2->td;
-	for ( i = 0; i < n; i++ )
-		d->d[i] = w[i] - d2->d[i];
-	NEWMP(m); m->dl = d; m->c = STOI(mod - ITOS(BDY(p1)->c)); NEXT(m) = 0;
-	MKDP(n,m,s); s->sugar = d->td; _mulmd(CO,mod,s,p2,&u);
-	_addmd(CO,mod,t,u,rp);
-}
-
-void _dp_sp_component_mod(p1,p2,mod,f1,f2)
-DP p1,p2;
-int mod;
-DP *f1,*f2;
-{
-	int i,n,td;
-	int *w;
-	DL d1,d2,d;
-	MP m;
-	DP t,s,u;
-
-	n = p1->nv; d1 = BDY(p1)->dl; d2 = BDY(p2)->dl;
-	w = (int *)ALLOCA(n*sizeof(int));
-	for ( i = 0, td = 0; i < n; i++ ) {
-		w[i] = MAX(d1->d[i],d2->d[i]); td += w[i];
-	}
-	NEWDL(d,n); d->td = td - d1->td;
-	for ( i = 0; i < n; i++ )
-		d->d[i] = w[i] - d1->d[i];
-	NEWMP(m); m->dl = d; m->c = BDY(p2)->c; NEXT(m) = 0;
-	MKDP(n,m,s); s->sugar = d->td; _mulmd(CO,mod,s,p1,f1);
-	NEWDL(d,n); d->td = td - d2->td;
-	for ( i = 0; i < n; i++ )
-		d->d[i] = w[i] - d2->d[i];
-	NEWMP(m); m->dl = d; m->c = BDY(p1)->c; NEXT(m) = 0;
-	MKDP(n,m,s); s->sugar = d->td; _mulmd(CO,mod,s,p2,f2);
-}
-
-void _printdp(d)
-DP d;
-{
-	int n,i;
-	MP m;
-	DL dl;
-
-	if ( !d ) {
-		printf("0"); return;
-	}
-	for ( n = d->nv, m = BDY(d); m; m = NEXT(m) ) {
-		printf("%d*<<",ITOS(m->c));
-		for ( i = 0, dl = m->dl; i < n-1; i++ )
-			printf("%d,",dl->d[i]);
-		printf("%d",dl->d[i]);
-		printf(">>");
-		if ( NEXT(m) )
-			printf("+");
-	}
+		dt->d[i] = d1->d[i]+d2->d[i];
 }