sync/atomic: make unaligned 64-bit atomic operation panic on ARM

use MOVW.NE instead of BEQ and MOVW.

R=golang-dev, dave, rsc, daniel.morsing
CC=golang-dev
https://golang.org/cl/7718043

src/pkg/sync/atomic/asm_arm.s

@@ -29,6 +29,10 @@ casfail:
 TEXT ·armCompareAndSwapUint64(SB),7,$0
 	BL	fastCheck64<>(SB)
 	MOVW	addr+0(FP), R1
+	// make unaligned atomic access panic
+	AND.S	$7, R1, R2
+	BEQ 	2(PC)
+	MOVW	R2, (R2)
 	MOVW	oldlo+4(FP), R2
 	MOVW	oldhi+8(FP), R3
 	MOVW	newlo+12(FP), R4
@@ -67,6 +71,10 @@ addloop:
 TEXT ·armAddUint64(SB),7,$0
 	BL	fastCheck64<>(SB)
 	MOVW	addr+0(FP), R1
+	// make unaligned atomic access panic
+	AND.S	$7, R1, R2
+	BEQ 	2(PC)
+	MOVW	R2, (R2)
 	MOVW	deltalo+4(FP), R2
 	MOVW	deltahi+8(FP), R3
 add64loop:
@@ -84,6 +92,10 @@ add64loop:
 TEXT ·armLoadUint64(SB),7,$0
 	BL	fastCheck64<>(SB)
 	MOVW	addr+0(FP), R1
+	// make unaligned atomic access panic
+	AND.S	$7, R1, R2
+	BEQ 	2(PC)
+	MOVW	R2, (R2)
 load64loop:
 	LDREXD	(R1), R2	// loads R2 and R3
 	STREXD	R2, (R1), R0	// stores R2 and R3
@@ -96,6 +108,10 @@ load64loop:
 TEXT ·armStoreUint64(SB),7,$0
 	BL	fastCheck64<>(SB)
 	MOVW	addr+0(FP), R1
+	// make unaligned atomic access panic
+	AND.S	$7, R1, R2
+	BEQ 	2(PC)
+	MOVW	R2, (R2)
 	MOVW	vallo+4(FP), R2
 	MOVW	valhi+8(FP), R3
 store64loop:

src/pkg/sync/atomic/asm_linux_arm.s

@@ -80,6 +80,10 @@ TEXT cas64<>(SB),7,$0
 TEXT kernelCAS64<>(SB),7,$0
 	// int (*__kuser_cmpxchg64_t)(const int64_t *oldval, const int64_t *newval, volatile int64_t *ptr);
 	MOVW	addr+0(FP), R2 // ptr
+	// make unaligned atomic access panic
+	AND.S	$7, R2, R1
+	BEQ 	2(PC)
+	MOVW	R1, (R1)
 	MOVW	$4(FP), R0 // oldval
 	MOVW	$12(FP), R1 // newval
 	BL		cas64<>(SB)
@@ -91,6 +95,10 @@ TEXT kernelCAS64<>(SB),7,$0
 TEXT generalCAS64<>(SB),7,$20
 	// bool runtime·cas64(uint64 volatile *addr, uint64 *old, uint64 new)
 	MOVW	addr+0(FP), R0
+	// make unaligned atomic access panic
+	AND.S	$7, R0, R1
+	BEQ 	2(PC)
+	MOVW	R1, (R1)
 	MOVW	R0, 4(R13)
 	MOVW	$4(FP), R1 // oldval
 	MOVW	R1, 8(R13)

src/pkg/sync/atomic/atomic_test.go

@@ -1189,10 +1189,10 @@ func shouldPanic(t *testing.T, name string, f func()) {
 
 func TestUnaligned64(t *testing.T) {
 	// Unaligned 64-bit atomics on 32-bit systems are
-	// a continual source of pain. Test that on 386 they crash
+	// a continual source of pain. Test that on 32-bit systems they crash
 	// instead of failing silently.
-	if runtime.GOARCH != "386" {
-		t.Skip("test only runs on 386")
+	if unsafe.Sizeof(int(0)) != 4 {
+		t.Skip("test only runs on 32-bit systems")
 	}
 
 	x := make([]uint32, 4)