modified to use a different 'panic' strategy, and use

new clash-table in tables.c when possible, and some other tuning

mach/proto/ncg/codegen.c

@@ -62,8 +62,8 @@ unsigned codegen(codep,ply,toplevel,costlimit,forced) byte *codep; unsigned cost
 	int inscoerc=0;
 	int procarg[2];
 #ifdef ALLOW_NEXTEM
-	int paniced;
-	char *savebp;
+	static int paniced;
+	char *savebp = 0;
 #endif
 	state_t state;
 #define SAVEST	savestatus(&state)
@@ -128,7 +128,7 @@ unsigned codegen(codep,ply,toplevel,costlimit,forced) byte *codep; unsigned cost
 #ifndef ALLOW_NEXTEM
 	bp = nextem(toplevel);
 #else
-	paniced=0;
+	if (toplevel) paniced=0;
 	savebp = nextem(toplevel);
     panic:
 	bp = savebp;
@@ -171,9 +171,18 @@ if (Debug)
 					if (dist<mindistance) {
 						if(dist==0)
 							goto gotit;
+#ifdef ALLOW_NEXTEM
+						if (! paniced) {
+#endif
 						npos=0;
 						mindistance = dist;
+#ifdef ALLOW_NEXTEM
+						}
+#endif
 					}
+#ifdef ALLOW_NEXTEM
+					if (dist < MAXINT)
+#endif
 					pos[npos++] = cindex;
 				}
 			}
@@ -198,8 +207,9 @@ if (Debug)
 					RESTST;
 				}
 				FREEST;
-				if (totalcost+mincost>costlimit)
+				if (totalcost+mincost>costlimit) {
 					BROKE();
+				}
 			} else {
 				cindex = pos[0];
 			}
@@ -344,11 +354,11 @@ if(Debug>1) fprintf(stderr,"Pattern too long, %d with only %d items on stack\n",
 					myfree(regls[j]);
 #ifndef ALLOW_NEXTEM
 				assert(!toplevel);
-				BROKE();
 #else
 				assert(!(toplevel&&paniced));
-				goto normalfailed;
+				if (paniced) goto normalfailed;
 #endif
+				BROKE();
 			}
 			if (cp->c3_prop<0) {
 				totalcost+=docoerc(tp,cp,ply,toplevel,0);
@@ -372,15 +382,21 @@ if(Debug>1) fprintf(stderr,"Pattern too long, %d with only %d items on stack\n",
 	besttup=0;
 	for (; tup != 0; tup = ntup) {
 #ifndef NDEBUG
-if(Debug>1) fprintf(stderr,"Next tuple %d,%d,%d,%d\n",
+if(Debug>1) { fprintf(stderr,"Next tuple %d,%d,%d,%d\n",
 			tup->p_rar[0],
 			tup->p_rar[1],
 			tup->p_rar[2],
 			tup->p_rar[3]);
+		fprintf(stderr, "totalcost = %u, costlimit = %u, mincost = %u\n",
+			totalcost, costlimit, mincost);
+	}
 #endif
 		ntup = tup->p_next;
 		for (i=0,t=0;i<nregneeded && t<mincost; i++)
 			t += docoerc(regtp[i],regcp[i],ply,FALSE,tup->p_rar[i]);
+#ifndef NDEBUG
+if (Debug > 1) fprintf(stderr, "cost after coercions: %u\n", t);
+#endif
 		if ( t<mincost && tokpatlen<=stackheight ) {
 #ifndef NDEBUG
 			if (Debug>2)
@@ -415,17 +431,7 @@ normalfailed:	if (stackpad!=tokpatlen) {
 			goto nextmatch;
 		}
 		totalcost += mincost;
-#ifndef ALLOW_NEXTEM
-		BROKE();
-#else
-		if (toplevel && !paniced) {
-			stackheight=0;
-			paniced++;
-			DEBUG("PANIC!");
-			goto panic;
-		} else
-			BROKE();
-#endif
+                BROKE();
 	}
 	for (i=0;i<nregneeded;i++)
 		totalcost += docoerc(regtp[i],regcp[i],ply,toplevel,besttup->p_rar[i]);
@@ -826,13 +832,25 @@ normalfailed:	if (stackpad!=tokpatlen) {
     case DO_RETURN:
 	DEBUG("RETURN");
 	assert(origcp!=startupcode);
-    doreturn:
 #ifndef NDEBUG
 	level--;
 #endif
 	return(totalcost);
 	}
 	}
+    doreturn:
+#ifdef ALLOW_NEXTEM
+	if (toplevel && totalcost == INFINITY && ! paniced) {
+		totalcost += stackupto(&fakestack[stackheight-1], ply, toplevel);
+		paniced = 1;
+		DEBUG("PANIC!");
+		goto panic;
+	}
+#endif
+#ifndef NDEBUG
+	level--;
+#endif
+	return(totalcost);
 }
 
 readcodebytes() {

mach/proto/ncg/data.h

@@ -29,6 +29,9 @@ struct reginfo {
 #if MAXMEMBERS!=0
 	int     r_members[MAXMEMBERS];  /* register contained within this reg */
 	short	r_clash[REGSETSIZE];	/* set of clashing registers */
+	int	r_iclash;		/* index in clashlist; other represen-
+					   tation of r_clash
+					*/
 #endif
 	int     r_refcount;             /* Times in use */
 	token_t r_contents;             /* Current contents */

mach/proto/ncg/move.c

@@ -22,9 +22,9 @@ unsigned costcalc();
 
 move(tp1,tp2,ply,toplevel,maxcost) token_p tp1,tp2; unsigned maxcost; {
 	register move_p mp;
-	register unsigned t;
+	unsigned t;
 	register struct reginfo *rp;
-	tkdef_p tdp;
+	register byte *tdpb;
 	int i;
 	unsigned codegen();
 
@@ -36,32 +36,14 @@ move(tp1,tp2,ply,toplevel,maxcost) token_p tp1,tp2; unsigned maxcost; {
 				    &machregs[tp2->t_att[0].ar].r_contents) &&
 			      machregs[tp1->t_att[0].ar].r_contents.t_token!=0)
 				return(0);
-			erasereg(tp2->t_att[0].ar);
-			machregs[tp2->t_att[0].ar].r_contents =
-			  machregs[tp1->t_att[0].ar].r_contents ;
-
 		} else {
 			if (eqtoken(&machregs[tp2->t_att[0].ar].r_contents,tp1))
 				return(0);
-			erasereg(tp2->t_att[0].ar);
-			machregs[tp2->t_att[0].ar].r_contents = *tp1;
-		}
-		for (rp=machregs+1;rp<machregs+NREGS;rp++) {
-			if (rp->r_contents.t_token == 0)
-				continue;
-			assert(rp->r_contents.t_token > 0);
-			tdp = &tokens[rp->r_contents.t_token];
-			for (i=0;i<TOKENSIZE;i++)
-				if (tdp->t_type[i] == EV_REG &&
-				    clash(rp->r_contents.t_att[i].ar,tp2->t_att[0].ar)) {
-					erasereg(rp-machregs);
-					break;
-				}
 		}
+		erasereg(tp2->t_att[0].ar);
 	} else if (tp1->t_token == -1) {
 		if (eqtoken(tp2,&machregs[tp1->t_att[0].ar].r_contents))
 			return(0);
-		machregs[tp1->t_att[0].ar].r_contents = *tp2;
 	}
 	/*
 	 * If we arrive here the move must really be executed
@@ -89,6 +71,27 @@ move(tp1,tp2,ply,toplevel,maxcost) token_p tp1,tp2; unsigned maxcost; {
 	t = codegen(&coderules[mp->m_cindex],ply,toplevel,maxcost,0);
 	tokpatlen -= 2;
 	stackheight -= 2;
+	if (tp2->t_token == -1) {
+		rp = &machregs[tp2->t_att[0].ar];
+		if (tp1->t_token == -1) {
+			rp->r_contents =
+			  machregs[tp1->t_att[0].ar].r_contents ;
+		}
+		else	rp->r_contents = *tp1;
+		if (rp->r_contents.t_token > 0) {
+			tdpb = &(tokens[rp->r_contents.t_token].t_type[0]);
+			for (i=0;i<TOKENSIZE;i++)
+				if (*tdpb++ == EV_REG &&
+				    clash(rp->r_contents.t_att[i].ar,tp2->t_att[0].ar)) {
+					rp->r_contents.t_token = 0;
+					for (i = 0; i < TOKENSIZE; i++)
+						rp->r_contents.t_att[i].aw = 0;
+					break;
+				}
+		}
+	}
+	else if (tp1->t_token == -1)
+		machregs[tp1->t_att[0].ar].r_contents = *tp2;
 	return(t);
 }
 
@@ -101,10 +104,7 @@ setcc(tp) token_p tp; {
 
 test(tp,ply,toplevel,maxcost) token_p tp; unsigned maxcost; {
 	register test_p mp;
-	register unsigned t;
-	register struct reginfo *rp;
-	tkdef_p tdp;
-	int i;
+	unsigned t;
 	unsigned codegen();
 
 	if (cocoreg.t_token!=0) {

mach/proto/ncg/reg.c

@@ -59,30 +59,33 @@ getrefcount(regno) {
 }
 
 erasereg(regno) {
-	register struct reginfo *rp;
+	register struct reginfo *rp = &machregs[regno];
 	register int i;
 
-	rp = &machregs[regno];
+#if MAXMEMBERS==0
 	rp->r_contents.t_token = 0;
 	for (i=0;i<TOKENSIZE;i++)
 		rp->r_contents.t_att[i].aw = 0;
 
-#if MAXMEMBERS==0
 	awayreg(regno);
 #else
-	for (rp=machregs+1;rp<machregs+NREGS;rp++)
-		if (rp->r_clash[regno>>4]&(1<<(regno&017))) {
-			rp->r_contents.t_token = 0;
-			for (i=0;i<TOKENSIZE;i++)
-				rp->r_contents.t_att[i].aw = 0;
-			awayreg(rp-machregs);
-		}
+	extern short clashlist[];
+	register short *sp = &clashlist[rp->r_iclash];
+
+	rp->r_contents.t_token = 0;
+	while (*sp) {
+		rp = &machregs[*sp];
+		rp->r_contents.t_token = 0;
+		for (i=0;i<TOKENSIZE;i++)
+			rp->r_contents.t_att[i].aw = 0;
+		awayreg(*sp++);
+	}
 #endif
 }
 
 awayreg(regno) {
 	register struct reginfo *rp;
-	register tkdef_p tdp;
+	register byte *tdpb;
 	register i;
 
 	/* Now erase recursively all registers containing
@@ -90,14 +93,27 @@ awayreg(regno) {
 	 */
 	for (rp=machregs;rp<machregs+NREGS;rp++) {
 		if (rp->r_contents.t_token == -1) {
-			if (rp->r_contents.t_att[0].ar == regno)
-				erasereg(rp-machregs);
+			if (rp->r_contents.t_att[0].ar == regno) {
+				/* erasereg(rp-machregs);
+				   replaced by the following three
+				   lines
+				*/
+				rp->r_contents.t_token = 0;
+				for (i=0;i<TOKENSIZE;i++)
+					rp->r_contents.t_att[i].aw = 0;
+			}
 		} else if (rp->r_contents.t_token > 0) {
-			tdp= & tokens[rp->r_contents.t_token];
+			tdpb= & (tokens[rp->r_contents.t_token].t_type[0]);
 			for (i=0;i<TOKENSIZE;i++)
-				if (tdp->t_type[i] == EV_REG && 
+				if (*tdpb++ == EV_REG && 
 				    rp->r_contents.t_att[i].ar == regno) {
-					erasereg(rp-machregs);
+					/* erasereg(rp-machregs);
+					   replaced by the following three
+					   lines
+					*/
+					rp->r_contents.t_token = 0;
+					for (i=0;i<TOKENSIZE;i++)
+						rp->r_contents.t_att[i].aw = 0;
 					break;
 				}
 		}
@@ -137,7 +153,7 @@ inctcount(regno) {
 chkregs() {
 	register struct reginfo *rp;
 	register token_p tp;
-	register tkdef_p tdp;
+	register byte *tdpb;
 	int i;
 
 	for (rp=machregs+1;rp<machregs+NREGS;rp++) {
@@ -147,9 +163,9 @@ chkregs() {
 		if (tp->t_token == -1)
 			inctcount(tp->t_att[0].ar);
 		else {
-			tdp = &tokens[tp->t_token];
+			tdpb = &(tokens[tp->t_token].t_type[0]);
 			for (i=0;i<TOKENSIZE;i++)
-				if (tdp->t_type[i]==EV_REG)
+				if (*tdpb++==EV_REG)
 					inctcount(tp->t_att[i].ar);
 		}
 	}