for {
		old = *addr
		if Cas64(addr, old, old&val) {
			return old
		}
	}
}

//go:noescape
func Xadd(ptr *uint32, delta int32) uint32

//go:noescape
func Xadduintptr(ptr *uintptr, delta uintptr) uintptr

//go:noescape
func Xchg(ptr *uint32, new uint32) uint32

//go:noescape
func Xchguintptr(ptr *uintptr, new uintptr) uintptr

//go:noescape

TEXT ·Loadint32(SB), NOSPLIT, $0-12
	BR	·Load(SB)

// func Loadint64(ptr *int64) int64
TEXT ·Loadint64(SB), NOSPLIT, $0-16
	BR	·Load64(SB)

// func Xadduintptr(ptr *uintptr, delta uintptr) uintptr
TEXT ·Xadduintptr(SB), NOSPLIT, $0-24
	BR	·Xadd64(SB)

// func Xaddint32(ptr *int32, delta int32) int32
TEXT ·Xaddint32(SB), NOSPLIT, $0-20
	BR	·Xadd(SB)

TEXT ·Xadd64(SB), NOSPLIT, $0-24
	MOV	ptr+0(FP), A0
	MOV	delta+8(FP), A1
	AMOADDD A1, (A0), A2
	ADD	A2, A1, A0
	MOV	A0, ret+16(FP)
	RET

// func Xadduintptr(ptr *uintptr, delta uintptr) uintptr
TEXT ·Xadduintptr(SB), NOSPLIT, $0-24
	JMP	·Xadd64(SB)

// func Xchgint32(ptr *int32, new int32) int32
TEXT ·Xchgint32(SB), NOSPLIT, $0-20
	JMP	·Xchg(SB)

	MOVQ	val+8(FP), CX
casloop:
	MOVQ 	CX, DX
	MOVQ	(BX), AX
	ANDQ	AX, DX
	LOCK
	CMPXCHGQ	DX, (BX)
	JNZ casloop
	MOVQ 	AX, ret+16(FP)
	RET

// func Anduintptr(addr *uintptr, v uintptr) old uintptr
TEXT ·Anduintptr(SB), NOSPLIT, $0-24
	JMP	·And64(SB)

// func Oruintptr(addr *uintptr, v uintptr) old uintptr
TEXT ·Oruintptr(SB), NOSPLIT, $0-24

//
//go:nosplit
func Xadd(val *uint32, delta int32) uint32 {
	for {
		oval := *val
		nval := oval + uint32(delta)
		if Cas(val, oval, nval) {
			return nval
		}
	}
}

//go:noescape
func Xadduintptr(ptr *uintptr, delta uintptr) uintptr

//go:nosplit
func Xchg(addr *uint32, v uint32) uint32 {
	for {
		old := *addr
		if Cas(addr, old, v) {
			return old
		}
	}
}

//go:nosplit

	MOVW	ptr+0(FP), R1
	MOVW	val+4(FP), R2

	SYNC
	LL	(R1), R3
	AND	R2, R3, R4
	SC	R4, (R1)
	BEQ	R4, -4(PC)
	SYNC
	MOVW	R3, ret+8(FP)
	RET

// func Anduintptr(addr *uintptr, v uintptr) old uintptr
TEXT ·Anduintptr(SB), NOSPLIT, $0-12
	JMP	·And32(SB)

// func Oruintptr(addr *uintptr, v uintptr) old uintptr
TEXT ·Oruintptr(SB), NOSPLIT, $0-12
	JMP	·Or32(SB)

TEXT ·spinLock(SB),NOSPLIT,$0-4

	MOVV	ptr+0(FP), R4
	MOVV	val+8(FP), R5
	DBAR
	LLV	(R4), R6
	AND	R5, R6, R7
	SCV	R7, (R4)
	BEQ	R7, -4(PC)
	DBAR
	MOVV R6, ret+16(FP)
	RET

// func Anduintptr(addr *uintptr, v uintptr) old uintptr
TEXT ·Anduintptr(SB), NOSPLIT, $0-24
	JMP	·And64(SB)

// func Oruintptr(addr *uintptr, v uintptr) old uintptr
TEXT ·Oruintptr(SB), NOSPLIT, $0-24
	JMP	·Or64(SB)

	MOVL	AX, BX
	MOVL	DX, CX
	ORL	SI, BX
	ORL	DI, CX
	LOCK
	CMPXCHG8B	0(BP)
	JNZ casloop
	MOVL	AX, ret_lo+12(FP)
	MOVL	DX, ret_hi+16(FP)
	RET

// func Anduintptr(addr *uintptr, v uintptr) old uintptr
TEXT ·Anduintptr(SB), NOSPLIT, $0-12
	JMP	·And32(SB)

// func Oruintptr(addr *uintptr, v uintptr) old uintptr
TEXT ·Oruintptr(SB), NOSPLIT, $0-12

	MOVV	ptr+0(FP), R1
	MOVV	val+8(FP), R2

	SYNC
	LLV	(R1), R3
	AND	R2, R3, R4
	SCV	R4, (R1)
	BEQ	R4, -3(PC)
	SYNC
	MOVV	R3, ret+16(FP)
	RET

// func Anduintptr(addr *uintptr, v uintptr) old uintptr
TEXT ·Anduintptr(SB), NOSPLIT, $0-24
	JMP	·And64(SB)

// func Oruintptr(addr *uintptr, v uintptr) old uintptr
TEXT ·Oruintptr(SB), NOSPLIT, $0-24
	JMP	·Or64(SB)

	B	·Store(SB)

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

TEXT ·Xaddint32(SB),NOSPLIT,$0-12
	B	·Xadd(SB)

TEXT ·Xaddint64(SB),NOSPLIT,$-4-20
	B	·Xadd64(SB)

TEXT ·Xadduintptr(SB),NOSPLIT,$0-12
	B	·Xadd(SB)

TEXT ·Xchgint32(SB),NOSPLIT,$0-12
	B	·Xchg(SB)

TEXT ·Xchgint64(SB),NOSPLIT,$-4-20
	B	·Xchg64(SB)

// 64-bit atomics