new convention, direction bit is

always left cleared. changed
compiler generated memcpy and
memset to assume CLD.

R=r
OCL=21215
CL=21215

src/cmd/6g/cgen.c

@@ -874,22 +874,29 @@ sgen(Node *n, Node *ns, int32 w)
 			gins(AREP, N, N);	// repeat
 			gins(AMOVSQ, N, N);	// MOVQ *(SI)-,*(DI)-
 		}
-		// for future optimization
 		// we leave with the flag clear
 		gins(ACLD, N, N);
 	} else {
 		// normal direction
-		gins(ACLD, N, N);		// clear direction flag
-		if(q > 0) {
+		if(q >= 4) {
 			gconreg(AMOVQ, q, D_CX);
 			gins(AREP, N, N);	// repeat
 			gins(AMOVSQ, N, N);	// MOVQ *(SI)+,*(DI)+
+		} else
+		while(q > 0) {
+			gins(AMOVSQ, N, N);	// MOVQ *(SI)+,*(DI)+
+			q--;
 		}
 
-		if(c > 0) {
+		if(c >= 4) {
 			gconreg(AMOVQ, c, D_CX);
 			gins(AREP, N, N);	// repeat
 			gins(AMOVSB, N, N);	// MOVB *(SI)+,*(DI)+
+		
+		} else
+		while(c > 0) {
+			gins(AMOVSB, N, N);	// MOVB *(SI)+,*(DI)+
+			c--;
 		}
 	}
 }

src/cmd/6g/gen.c

@@ -1037,7 +1037,7 @@ cgen_as(Node *nl, Node *nr, int op)
 {
 	Node nc, n1;
 	Type *tl;
-	uint32 w, c;
+	uint32 w, c, q;
 	int iszer;
 
 	if(nl == N)
@@ -1058,31 +1058,32 @@ cgen_as(Node *nl, Node *nr, int op)
 			if(debug['g'])
 				dump("\nclearfat", nl);
 
-			if(nl->type->width < 0)
-				fatal("clearfat %T %lld", nl->type, nl->type->width);
 			w = nl->type->width;
+			c = w % 8;	// bytes
+			q = w / 8;	// quads
 
-			if(w > 0)
-				gconreg(AMOVQ, 0, D_AX);
+			gconreg(AMOVQ, 0, D_AX);
+			nodreg(&n1, types[tptr], D_DI);
+			agen(nl, &n1);
 
-			if(w > 0) {
-				nodreg(&n1, types[tptr], D_DI);
-				agen(nl, &n1);
-				gins(ACLD, N, N);	// clear direction flag
-			}
-
-			c = w / 8;
-			if(c > 0) {
-				gconreg(AMOVQ, c, D_CX);
+			if(q >= 4) {
+				gconreg(AMOVQ, q, D_CX);
 				gins(AREP, N, N);	// repeat
 				gins(ASTOSQ, N, N);	// STOQ AL,*(DI)+
+			} else
+			while(q > 0) {
+				gins(ASTOSQ, N, N);	// STOQ AL,*(DI)+
+				q--;
 			}
 
-			c = w % 8;
-			if(c > 0) {
+			if(c >= 4) {
 				gconreg(AMOVQ, c, D_CX);
 				gins(AREP, N, N);	// repeat
 				gins(ASTOSB, N, N);	// STOB AL,*(DI)+
+			} else
+			while(c > 0) {
+				gins(ASTOSB, N, N);	// STOB AL,*(DI)+
+				c--;
 			}
 			goto ret;
 		}

src/runtime/rt0_amd64.s

@@ -26,6 +26,7 @@ TEXT	_rt0_amd64(SB),7,$-8
 	MOVQ	AX, 0(R15)		// 0(R15) is stack limit (w 104b guard)
 	MOVQ	SP, 8(R15)		// 8(R15) is base
 
+	CLD				// convention is D is always left cleared
 	CALL	check(SB)
 
 	MOVL	16(SP), AX		// copy argc