diff --git a/src/pkg/runtime/386/asm.s b/src/pkg/runtime/386/asm.s
index c6df4f069a7..add14be964f 100644
--- a/src/pkg/runtime/386/asm.s
+++ b/src/pkg/runtime/386/asm.s
@@ -479,12 +479,16 @@ TEXT runtime·stackcheck(SB), 7, $0
 TEXT runtime·memclr(SB),7,$0
 	MOVL	4(SP), DI		// arg 1 addr
 	MOVL	8(SP), CX		// arg 2 count
-	ADDL	$3, CX
+	MOVL	CX, BX
+	ANDL	$3, BX
 	SHRL	$2, CX
 	MOVL	$0, AX
 	CLD
 	REP
 	STOSL
+	MOVL	BX, CX
+	REP
+	STOSB
 	RET
 
 TEXT runtime·getcallerpc(SB),7,$0
diff --git a/src/pkg/runtime/amd64/asm.s b/src/pkg/runtime/amd64/asm.s
index 2b16587f8cf..6ac84c408d6 100644
--- a/src/pkg/runtime/amd64/asm.s
+++ b/src/pkg/runtime/amd64/asm.s
@@ -527,12 +527,16 @@ TEXT runtime·stackcheck(SB), 7, $0
 TEXT runtime·memclr(SB),7,$0
 	MOVQ	8(SP), DI		// arg 1 addr
 	MOVQ	16(SP), CX		// arg 2 count
-	ADDQ	$7, CX
+	MOVQ	CX, BX
+	ANDQ	$7, BX
 	SHRQ	$3, CX
 	MOVQ	$0, AX
 	CLD
 	REP
 	STOSQ
+	MOVQ	BX, CX
+	REP
+	STOSB
 	RET
 
 TEXT runtime·getcallerpc(SB),7,$0
diff --git a/src/pkg/runtime/runtime.c b/src/pkg/runtime/runtime.c
index a16ca417e6f..00116c001a0 100644
--- a/src/pkg/runtime/runtime.c
+++ b/src/pkg/runtime/runtime.c
@@ -381,18 +381,11 @@ memprint(uint32 s, void *a)
 static void
 memcopy(uint32 s, void *a, void *b)
 {
-	byte *ba, *bb;
-	uint32 i;
-
-	ba = a;
-	bb = b;
-	if(bb == nil) {
-		for(i=0; i<s; i++)
-			ba[i] = 0;
+	if(b == nil) {
+		runtime·memclr(a,s);
 		return;
 	}
-	for(i=0; i<s; i++)
-		ba[i] = bb[i];
+	runtime·memmove(a,b,s);
 }
 
 static uint32