sync/atomic: fix new swap on arm linux

TBR=dvyukov
CC=golang-dev
https://golang.org/cl/12920043

src/pkg/sync/atomic/64bit_arm.go

@@ -37,7 +37,7 @@ func addUint64(val *uint64, delta uint64) (new uint64) {
 
 func swapUint64(addr *uint64, new uint64) (old uint64) {
 	for {
-		old := *addr
+		old = *addr
 		if CompareAndSwapUint64(addr, old, new) {
 			break
 		}

src/pkg/sync/atomic/asm_linux_arm.s

@@ -88,9 +88,10 @@ TEXT ·SwapUint32(SB),NOSPLIT,$0-12
 	MOVW	new+4(FP), R1
 swaploop1:
 	MOVW	0(R2), R0
+	MOVW	R0, R4 // cas smashes R0
 	BL	cas<>(SB)
 	BCC	swaploop1
-	MOVW	R0, old+8(FP)
+	MOVW	R4, old+8(FP)
 	RET
 
 TEXT ·SwapUintptr(SB),NOSPLIT,$0