for pb.Next() {
			atomic.And(&x[63], i)
			i++
		}
	})
}

func BenchmarkOr8(b *testing.B) {
	var x [512]uint8 // give byte its own cache line
	sink = &x
	for i := 0; i < b.N; i++ {
		atomic.Or8(&x[255], uint8(i))
	}
}

func BenchmarkOr(b *testing.B) {
	var x [128]uint32 // give x its own cache line
	sink = &x
	for i := 0; i < b.N; i++ {
		atomic.Or(&x[63], uint32(i))
	}
}

//go:noescape
func LoadAcq64(ptr *uint64) uint64

//go:noescape
func LoadAcquintptr(ptr *uintptr) uintptr

//go:noescape
func And8(ptr *uint8, val uint8)

//go:noescape
func Or8(ptr *uint8, val uint8)

// NOTE: Do not add atomicxor8 (XOR is not idempotent).

//go:noescape
func And(ptr *uint32, val uint32)

//go:noescape
func Or(ptr *uint32, val uint32)

//go:noescape

//go:noescape
func LoadAcquintptr(ptr *uintptr) uintptr

//go:noescape
func And8(ptr *uint8, val uint8)

//go:noescape
func And(ptr *uint32, val uint32)

//go:noescape
func Or8(ptr *uint8, val uint8)

//go:noescape
func Or(ptr *uint32, val uint32)

//go:noescape
func And32(ptr *uint32, val uint32) uint32

//go:noescape
func Or32(ptr *uint32, val uint32) uint32

//go:noescape

TEXT ·StoreRel(SB), NOSPLIT, $0-12
	JMP	·Store(SB)

TEXT ·StoreRel64(SB), NOSPLIT, $0-16
	JMP	·Store64(SB)

TEXT ·StoreReluintptr(SB), NOSPLIT, $0-16
	JMP	·Store64(SB)

// void	·Or8(byte volatile*, byte);
TEXT ·Or8(SB), NOSPLIT, $0-9
	MOVQ	ptr+0(FP), AX
	MOVB	val+8(FP), BX
	LOCK
	ORB	BX, (AX)
	RET

// void	·And8(byte volatile*, byte);
TEXT ·And8(SB), NOSPLIT, $0-9
	MOVQ	ptr+0(FP), AX

	*ptr = val
}

//go:nosplit
//go:noinline
func StoreReluintptr(ptr *uintptr, val uintptr) {
	*ptr = val
}

//go:noescape
func And8(ptr *uint8, val uint8)

//go:noescape
func Or8(ptr *uint8, val uint8)

// NOTE: Do not add atomicxor8 (XOR is not idempotent).

//go:noescape
func And(ptr *uint32, val uint32)

//go:noescape
func Or(ptr *uint32, val uint32)

//go:noescape

//go:noescape
func LoadAcq64(ptr *uint64) uint64

//go:noescape
func LoadAcquintptr(ptr *uintptr) uintptr

//go:noescape
func And8(ptr *uint8, val uint8)

//go:noescape
func Or8(ptr *uint8, val uint8)

// NOTE: Do not add atomicxor8 (XOR is not idempotent).

//go:noescape
func And(ptr *uint32, val uint32)

//go:noescape
func Or(ptr *uint32, val uint32)

//go:noescape

//go:noescape
func LoadAcq(ptr *uint32) uint32

//go:noescape
func LoadAcq64(ptr *uint64) uint64

//go:noescape
func LoadAcquintptr(ptr *uintptr) uintptr

//go:noescape
func Or8(ptr *uint8, val uint8)

//go:noescape
func And8(ptr *uint8, val uint8)

//go:noescape
func And(ptr *uint32, val uint32)

//go:noescape
func Or(ptr *uint32, val uint32)

//go:noescape

TEXT ·StorepNoWB(SB),NOSPLIT,$0-8
	JMP	·Store(SB)

TEXT ·StoreRel(SB),NOSPLIT,$0-8
	JMP	·Store(SB)

TEXT ·StoreReluintptr(SB),NOSPLIT,$0-8
	JMP	·Store(SB)

// void	Or8(byte volatile*, byte);
TEXT ·Or8(SB),NOSPLIT,$0-5
	MOVW	ptr+0(FP), R1
	MOVBU	val+4(FP), R2
	MOVW	$~3, R3	// Align ptr down to 4 bytes so we can use 32-bit load/store.
	AND	R1, R3
#ifdef GOARCH_mips
	// Big endian.  ptr = ptr ^ 3

//go:noescape
func Load64(ptr *uint64) uint64

//go:nosplit
//go:noinline
func Load8(ptr *uint8) uint8 {
	return *ptr
}

//go:noescape
func And8(ptr *uint8, val uint8)

//go:noescape
func Or8(ptr *uint8, val uint8)

//go:noescape
func And(ptr *uint32, val uint32)

//go:noescape
func Or(ptr *uint32, val uint32)

//go:noescape
func And32(ptr *uint32, val uint32) uint32

//go:noescape

	mask := byte(1 << ((obj / heapArenaBytes) % 8))
	bytep := &arena.checkmarks.b[arenaWord]

	if atomic.Load8(bytep)&mask != 0 {
		// Already checkmarked.
		return true
	}

	atomic.Or8(bytep, mask)
	return false