in 6g -r (the rsc flag),

rewrite heap pointer moves as xchg.

R=ken
OCL=22665
CL=22665

src/cmd/6g/cgen.c

@@ -7,6 +7,7 @@
 /*
  * generate:
  *	res = n;
+ * simplifies and calls gmove.
  */
 void
 cgen(Node *n, Node *res)

src/cmd/6g/gsubr.c

@@ -511,15 +511,49 @@ gmove(Node *f, Node *t)
 		goto st;
 	case TINT32:
 	case TUINT32:
-	case TPTR32:
 		a = AMOVL;
 		goto st;
 	case TINT64:
 	case TUINT64:
-	case TPTR64:
 		a = AMOVQ;
 		goto st;
 
+	case TPTR32:
+	case TPTR64:
+		/*
+		 * store to pointer.
+		 */
+		if(tt == TPTR32)
+			a = AMOVL;
+		else
+			a = AMOVQ;
+		switch(t->op) {
+		default:
+			dump("gmove to", t);
+			fatal("gmove t %O", t->op);
+
+		case OINDREG:
+			if(t->val.u.reg != D_SP)
+				goto refcount;
+			break;
+
+		case ONAME:
+			switch(t->class) {
+			default:
+				dump("gmove", t);
+				fatal("gmove t %O class %d reg %R", t->op, t->class, t->val.u.reg);
+			case PEXTERN:
+			case PSTATIC:
+				goto refcount;
+				break;
+			case PAUTO:
+			case PPARAM:
+				break;
+			}
+			break;
+		}
+		goto st;
+
 	st:
 		if(f->op == OCONST) {
 			gins(a, f, t);
@@ -532,6 +566,22 @@ gmove(Node *f, Node *t)
 		regfree(&nod);
 		return;
 
+	refcount:
+		if(!debug['r'])
+			goto st;
+		// for now, mark ref count updates with AXCHGQ.
+		// using a temporary on the left, so no semantic
+		// changes.  code is likely slower, but still correct.
+		if(t64)
+			a = AXCHGQ;
+		else
+			a = AXCHGL;
+		regalloc(&nod, t->type, f);
+		gmove(f, &nod);
+		gins(a, &nod, t);
+		regfree(&nod);
+		return;
+
 	case TFLOAT32:
 		a = AMOVSS;
 		goto fst;

src/cmd/6g/reg.c

@@ -182,6 +182,19 @@ regopt(Prog *firstp)
 			for(z=0; z<BITS; z++)
 				r->use1.b[z] |= bit.b[z];
 			break;
+
+		/*
+		 * left side read+write
+		 */
+		case AXCHGB:
+		case AXCHGW:
+		case AXCHGL:
+		case AXCHGQ:
+			for(z=0; z<BITS; z++) {
+				r->use1.b[z] |= bit.b[z];
+				r->set.b[z] |= bit.b[z];
+			}
+			break;
 		}
 
 		bit = mkvar(r, &p->to);
@@ -313,6 +326,11 @@ regopt(Prog *firstp)
 		case ASBBL:
 		case ASBBQ:
 
+		case AXCHGB:
+		case AXCHGW:
+		case AXCHGL:
+		case AXCHGQ:
+
 		case AADDSD:
 		case AADDSS:
 		case ACMPSD: