[dev.power64] 5g,6g,8g,9g: debug prints for regopt pass 6 and paint2

Theses were very helpful in understanding the regions and
register selection when porting regopt to 9g.  Add them to the
other compilers (and improve 9g's successor debug print).

LGTM=rsc
R=rsc
CC=golang-codereviews
https://golang.org/cl/174130043

src/cmd/5g/opt.h

@@ -160,7 +160,7 @@ void	prop(Reg*, Bits, Bits);
 void	synch(Reg*, Bits);
 uint32	allreg(uint32, Rgn*);
 void	paint1(Reg*, int);
-uint32	paint2(Reg*, int);
+uint32	paint2(Reg*, int, int);
 void	paint3(Reg*, int, uint32, int);
 void	addreg(Adr*, int);
 void	dumpit(char *str, Flow *r0, int);

src/cmd/5g/reg.c

@@ -454,9 +454,13 @@ brk:
 	 * replace code (paint3)
 	 */
 	rgp = region;
+	if(debug['R'] && debug['v'])
+		print("\nregisterizing\n");
 	for(i=0; i<nregion; i++) {
+		if(debug['R'] && debug['v'])
+			print("region %d: cost %d varno %d enter %d\n", i, rgp->cost, rgp->varno, rgp->enter->f.prog->pc);
 		bit = blsh(rgp->varno);
-		vreg = paint2(rgp->enter, rgp->varno);
+		vreg = paint2(rgp->enter, rgp->varno, 0);
 		vreg = allreg(vreg, rgp);
 		if(debug['R']) {
 			if(rgp->regno >= NREG)
@@ -477,9 +481,6 @@ brk:
 		rgp++;
 	}
 
-	if(debug['R'] && debug['v'])
-		dumpit("pass6", &firstr->f, 1);
-
 	/*
 	 * free aux structures. peep allocates new ones.
 	 */
@@ -488,6 +489,15 @@ brk:
 	flowend(g);
 	firstr = R;
 
+	if(debug['R'] && debug['v']) {
+		// Rebuild flow graph, since we inserted instructions
+		g = flowstart(firstp, sizeof(Reg));
+		firstr = (Reg*)g->start;
+		dumpit("pass6", &firstr->f, 1);
+		flowend(g);
+		firstr = R;
+	}
+
 	/*
 	 * pass 7
 	 * peep-hole on basic block
@@ -1189,7 +1199,7 @@ paint1(Reg *r, int bn)
 }
 
 uint32
-paint2(Reg *r, int bn)
+paint2(Reg *r, int bn, int depth)
 {
 	Reg *r1;
 	int z;
@@ -1213,6 +1223,9 @@ paint2(Reg *r, int bn)
 		r = r1;
 	}
 	for(;;) {
+		if(debug['R'] && debug['v'])
+			print("  paint2 %d %P\n", depth, r->f.prog);
+
 		r->act.b[z] &= ~bb;
 
 		vreg |= r->regu;
@@ -1220,14 +1233,14 @@ paint2(Reg *r, int bn)
 		if(r->refbehind.b[z] & bb)
 			for(r1 = (Reg*)r->f.p2; r1 != R; r1 = (Reg*)r1->f.p2link)
 				if(r1->refahead.b[z] & bb)
-					vreg |= paint2(r1, bn);
+					vreg |= paint2(r1, bn, depth+1);
 
 		if(!(r->refahead.b[z] & bb))
 			break;
 		r1 = (Reg*)r->f.s2;
 		if(r1 != R)
 			if(r1->refbehind.b[z] & bb)
-				vreg |= paint2(r1, bn);
+				vreg |= paint2(r1, bn, depth+1);
 		r = (Reg*)r->f.s1;
 		if(r == R)
 			break;
@@ -1344,6 +1357,8 @@ RtoB(int r)
 int
 BtoR(uint32 b)
 {
+	// TODO Allow R0 and R1, but be careful with a 0 return
+	// TODO Allow R9. Only R10 is reserved now (just g, not m).
 	b &= 0x11fcL;	// excluded R9 and R10 for m and g, but not R12
 	if(b == 0)
 		return 0;
@@ -1442,12 +1457,14 @@ dumpit(char *str, Flow *r0, int isreg)
 				print(" (only)");
 			print("\n");
 		}
-//		r1 = r->s1;
+		// Print successors if it's not just the next one
-//		if(r1 != nil) {
+		if(r->s1 != r->link || r->s2 != nil) {
-//			print("	succ:");
+			print("	succ:");
-//			for(; r1 != R; r1 = r1->s1)
+			if(r->s1 != nil)
-//				print(" %.4ud", (int)r1->prog->pc);
+				print(" %.4ud", (int)r->s1->prog->pc);
-//			print("\n");
+			if(r->s2 != nil)
-//		}
+				print(" %.4ud", (int)r->s2->prog->pc);
+			print("\n");
+		}
 	}
 }

src/cmd/6g/opt.h

@@ -156,7 +156,7 @@ void	prop(Reg*, Bits, Bits);
 void	synch(Reg*, Bits);
 uint32	allreg(uint32, Rgn*);
 void	paint1(Reg*, int);
-uint32	paint2(Reg*, int);
+uint32	paint2(Reg*, int, int);
 void	paint3(Reg*, int, uint32, int);
 void	addreg(Adr*, int);
 void	dumpone(Flow*, int);

src/cmd/6g/reg.c

@@ -389,9 +389,13 @@ brk:
 	 * replace code (paint3)
 	 */
 	rgp = region;
+	if(debug['R'] && debug['v'])
+		print("\nregisterizing\n");
 	for(i=0; i<nregion; i++) {
+		if(debug['R'] && debug['v'])
+			print("region %d: cost %d varno %d enter %d\n", i, rgp->cost, rgp->varno, rgp->enter->f.prog->pc);
 		bit = blsh(rgp->varno);
-		vreg = paint2(rgp->enter, rgp->varno);
+		vreg = paint2(rgp->enter, rgp->varno, 0);
 		vreg = allreg(vreg, rgp);
 		if(rgp->regno != 0) {
 			if(debug['R'] && debug['v']) {
@@ -406,9 +410,6 @@ brk:
 		rgp++;
 	}
 
-	if(debug['R'] && debug['v'])
-		dumpit("pass6", &firstr->f, 1);
-	
 	/*
 	 * free aux structures. peep allocates new ones.
 	 */
@@ -417,6 +418,15 @@ brk:
 	flowend(g);
 	firstr = R;
 
+	if(debug['R'] && debug['v']) {
+		// Rebuild flow graph, since we inserted instructions
+		g = flowstart(firstp, sizeof(Reg));
+		firstr = (Reg*)g->start;
+		dumpit("pass6", &firstr->f, 1);
+		flowend(g);
+		firstr = R;
+	}
+
 	/*
 	 * pass 7
 	 * peep-hole on basic block
@@ -1020,7 +1030,7 @@ paint1(Reg *r, int bn)
 }
 
 uint32
-paint2(Reg *r, int bn)
+paint2(Reg *r, int bn, int depth)
 {
 	Reg *r1;
 	int z;
@@ -1044,6 +1054,9 @@ paint2(Reg *r, int bn)
 		r = r1;
 	}
 	for(;;) {
+		if(debug['R'] && debug['v'])
+			print("  paint2 %d %P\n", depth, r->f.prog);
+
 		r->act.b[z] &= ~bb;
 
 		vreg |= r->regu;
@@ -1051,14 +1064,14 @@ paint2(Reg *r, int bn)
 		if(r->refbehind.b[z] & bb)
 			for(r1 = (Reg*)r->f.p2; r1 != R; r1 = (Reg*)r1->f.p2link)
 				if(r1->refahead.b[z] & bb)
-					vreg |= paint2(r1, bn);
+					vreg |= paint2(r1, bn, depth+1);
 
 		if(!(r->refahead.b[z] & bb))
 			break;
 		r1 = (Reg*)r->f.s2;
 		if(r1 != R)
 			if(r1->refbehind.b[z] & bb)
-				vreg |= paint2(r1, bn);
+				vreg |= paint2(r1, bn, depth+1);
 		r = (Reg*)r->f.s1;
 		if(r == R)
 			break;
@@ -1259,12 +1272,14 @@ dumpit(char *str, Flow *r0, int isreg)
 				print(" %.4ud", (int)r1->prog->pc);
 			print("\n");
 		}
-//		r1 = r->s1;
+		// Print successors if it's not just the next one
-//		if(r1 != R) {
+		if(r->s1 != r->link || r->s2 != nil) {
-//			print("	succ:");
+			print("	succ:");
-//			for(; r1 != R; r1 = r1->s1)
+			if(r->s1 != nil)
-//				print(" %.4ud", (int)r1->prog->pc);
+				print(" %.4ud", (int)r->s1->prog->pc);
-//			print("\n");
+			if(r->s2 != nil)
-//		}
+				print(" %.4ud", (int)r->s2->prog->pc);
+			print("\n");
+		}
 	}
 }

src/cmd/8g/opt.h

@@ -174,7 +174,7 @@ void	loopit(Reg*, int32);
 void	synch(Reg*, Bits);
 uint32	allreg(uint32, Rgn*);
 void	paint1(Reg*, int);
-uint32	paint2(Reg*, int);
+uint32	paint2(Reg*, int, int);
 void	paint3(Reg*, int, uint32, int);
 void	addreg(Adr*, int);
 void	dumpone(Flow*, int);

src/cmd/8g/reg.c

@@ -358,18 +358,19 @@ brk:
 	 * replace code (paint3)
 	 */
 	rgp = region;
+	if(debug['R'] && debug['v'])
+		print("\nregisterizing\n");
 	for(i=0; i<nregion; i++) {
+		if(debug['R'] && debug['v'])
+			print("region %d: cost %d varno %d enter %d\n", i, rgp->cost, rgp->varno, rgp->enter->f.prog->pc);
 		bit = blsh(rgp->varno);
-		vreg = paint2(rgp->enter, rgp->varno);
+		vreg = paint2(rgp->enter, rgp->varno, 0);
 		vreg = allreg(vreg, rgp);
 		if(rgp->regno != 0)
 			paint3(rgp->enter, rgp->varno, vreg, rgp->regno);
 		rgp++;
 	}
 
-	if(debug['R'] && debug['v'])
-		dumpit("pass6", &firstr->f, 1);
-
 	/*
 	 * free aux structures. peep allocates new ones.
 	 */
@@ -378,6 +379,15 @@ brk:
 	flowend(g);
 	firstr = R;
 
+	if(debug['R'] && debug['v']) {
+		// Rebuild flow graph, since we inserted instructions
+		g = flowstart(firstp, sizeof(Reg));
+		firstr = (Reg*)g->start;
+		dumpit("pass6", &firstr->f, 1);
+		flowend(g);
+		firstr = R;
+	}
+
 	/*
 	 * pass 7
 	 * peep-hole on basic block
@@ -996,7 +1006,7 @@ paint1(Reg *r, int bn)
 }
 
 uint32
-paint2(Reg *r, int bn)
+paint2(Reg *r, int bn, int depth)
 {
 	Reg *r1;
 	int z;
@@ -1020,6 +1030,9 @@ paint2(Reg *r, int bn)
 		r = r1;
 	}
 	for(;;) {
+		if(debug['R'] && debug['v'])
+			print("  paint2 %d %P\n", depth, r->f.prog);
+
 		r->act.b[z] &= ~bb;
 
 		vreg |= r->regu;
@@ -1027,14 +1040,14 @@ paint2(Reg *r, int bn)
 		if(r->refbehind.b[z] & bb)
 			for(r1 = (Reg*)r->f.p2; r1 != R; r1 = (Reg*)r1->f.p2link)
 				if(r1->refahead.b[z] & bb)
-					vreg |= paint2(r1, bn);
+					vreg |= paint2(r1, bn, depth+1);
 
 		if(!(r->refahead.b[z] & bb))
 			break;
 		r1 = (Reg*)r->f.s2;
 		if(r1 != R)
 			if(r1->refbehind.b[z] & bb)
-				vreg |= paint2(r1, bn);
+				vreg |= paint2(r1, bn, depth+1);
 		r = (Reg*)r->f.s1;
 		if(r == R)
 			break;
@@ -1227,12 +1240,14 @@ dumpit(char *str, Flow *r0, int isreg)
 				print(" %.4ud", (int)r1->prog->pc);
 			print("\n");
 		}
-//		r1 = r->s1;
+		// Print successors if it's not just the next one
-//		if(r1 != nil) {
+		if(r->s1 != r->link || r->s2 != nil) {
-//			print("	succ:");
+			print("	succ:");
-//			for(; r1 != R; r1 = r1->s1)
+			if(r->s1 != nil)
-//				print(" %.4ud", (int)r1->prog->pc);
+				print(" %.4ud", (int)r->s1->prog->pc);
-//			print("\n");
+			if(r->s2 != nil)
-//		}
+				print(" %.4ud", (int)r->s2->prog->pc);
+			print("\n");
+		}
 	}
 }

src/cmd/9g/reg.c

@@ -1322,7 +1322,6 @@ void
 dumpit(char *str, Flow *r0, int isreg)
 {
 	Flow *r, *r1;
-	int s1v, s2v;
 
 	print("\n%s\n", str);
 	for(r = r0; r != nil; r = r->link) {
@@ -1334,10 +1333,8 @@ dumpit(char *str, Flow *r0, int isreg)
 				print(" %.4ud", (int)r1->prog->pc);
 			print("\n");
 		}
-		// If at least one successor is "interesting", print both
+		// Print successors if it's not just the next one
-		s1v = (r->s1 != nil) && (r->s1->prog != r->prog->link);
+		if(r->s1 != r->link || r->s2 != nil) {
-		s2v = (r->s2 != nil) && (r->s2->prog != r->prog->link);
-		if(s1v || s2v) {
 			print("	succ:");
 			if(r->s1 != nil)
 				print(" %.4ud", (int)r->s1->prog->pc);