TEXT ·StoreRel(SB),NOSPLIT,$0-8
	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

// Atomically:
//	if(*val == old){
//		*val = new;
//		return 1;
//	} else
//		return 0;
TEXT ·Casp1(SB), NOSPLIT, $0-25
	JMP	·Cas64(SB)

// uint32 xadd(uint32 volatile *ptr, int32 delta)
// Atomically:
//	*val += delta;
//	return *val;
TEXT ·Xadd(SB), NOSPLIT, $0-20
	MOVV	ptr+0(FP), R4
	MOVW	delta+8(FP), R5
	DBAR
	LL	(R4), R6
	ADDU	R6, R5, R7
	MOVV	R7, R6
	SC	R7, (R4)
	BEQ	R7, -4(PC)

		}
	}
}

//go:nosplit
func And64(addr *uint64, val uint64) (old uint64) {
	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

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

func BenchmarkXadd(b *testing.B) {
	var x uint32
	ptr := &x
	b.RunParallel(func(pb *testing.PB) {
		for pb.Next() {
			atomic.Xadd(ptr, 1)
		}
	})
}

func BenchmarkXadd64(b *testing.B) {
	var x uint64
	ptr := &x
	b.RunParallel(func(pb *testing.PB) {
		for pb.Next() {
			atomic.Xadd64(ptr, 1)
		}
	})
}

		if ret == _EWOULDBLOCK {
			return -1
		}
	}
}

//go:nosplit
func semawakeup(mp *m) {
	atomic.Xadd(&mp.waitsemacount, 1)
	ret := thrwakeup(uintptr(unsafe.Pointer(&mp.waitsemacount)), 1)
	if ret != 0 && ret != _ESRCH {
		// semawakeup can be called on signal stack.
		systemstack(func() {

	ROL:       true,
	ROR:       true,
	SAR:       true,
	SBB:       true,
	SHL:       true,
	SHLD:      true,
	SHR:       true,
	SHRD:      true,
	SUB:       true,
	TEST:      true,
	XADD:      true,
	XCHG:      true,
	XOR:       true,
}

var plan9Reg = [...]string{
	AL:   "AL",
	CL:   "CL",
	BL:   "BL",
	DL:   "DL",
	AH:   "AH",
	CH:   "CH",
	BH:   "BH",

// CHECK-NEXT:    %[[ARG0]] = f32[] parameter(0)
// CHECK-NEXT:    %[[ARG1]] = f32[] parameter(1)
// CHECK-NEXT:    [[ADD:%.*]] = f32[] add(f32[] %[[ARG0]], f32[] %[[ARG1]])
// CHECK-NEXT:    ROOT %tuple.{{[0-9]+}} = (f32[]) tuple(f32[] [[ADD]])
// CHECK-NEXT:  }

// CHECK:       // InputMapping {0, 1}
// CHECK-NEXT:  // XlaInputShape f32[]
// CHECK-NEXT:  // XlaInputShape f32[]

# RUN: tf_tfl_translate -tf-input-arrays=input0,input1 -tf-input-shapes=4:4 -tf-input-data-types=DT_INT32,DT_INT32 -tf-output-arrays=Add %s -o - | flatbuffer_to_string - | FileCheck %s

# Add two tensor<4xi32> inputs and return the result

node {
  name: "Add"
  op: "Add"
  input: "input0"
  input: "input1"
  attr {
    key: "T"
    value {
      type: DT_INT32
    }
  }
}
node {
  name: "input0"
  op: "Placeholder"
  attr {

	VADD	V18.S4, V4.S4, V16.S4
	HASHUPDATEPARITY

	VADD	V18.S4, V5.S4, V16.S4
	HASHUPDATEPARITY

	VADD	V18.S4, V6.S4, V16.S4
	HASHUPDATEPARITY

	VADD	V18.S4, V7.S4, V16.S4
	HASHUPDATEPARITY

	SUB	$64, R3, R3                                  // message length - 64bytes, then compare with 64bytes
	VADD	V2.S4, V0.S4, V0.S4
	VADD	V1.S4, V20.S4, V20.S4
	CBNZ	R3, blockloop

sha1ret:

	WORD	$0x4D40E8DD

	VADD	V20.S4, V4.S4, V4.S4
	VADD	V21.S4, V5.S4, V5.S4
	VADD	V22.S4, V6.S4, V6.S4
	VADD	V23.S4, V7.S4, V7.S4
	VADD	V24.S4, V8.S4, V8.S4
	VADD	V25.S4, V9.S4, V9.S4
	VADD	V26.S4, V10.S4, V10.S4
	VADD	V27.S4, V11.S4, V11.S4
	VADD	V28.S4, V12.S4, V12.S4
	VADD	V29.S4, V13.S4, V13.S4
	VADD	V30.S4, V14.S4, V14.S4
	VADD	V31.S4, V15.S4, V15.S4

	VZIP1	V1.S4, V0.S4, V16.S4