softFloatOps = map[ssa.Op]sfRtCallDef{
		ssa.OpAdd32F: {typecheck.LookupRuntimeFunc("fadd32"), types.TFLOAT32},
		ssa.OpAdd64F: {typecheck.LookupRuntimeFunc("fadd64"), types.TFLOAT64},
		ssa.OpSub32F: {typecheck.LookupRuntimeFunc("fadd32"), types.TFLOAT32},
		ssa.OpSub64F: {typecheck.LookupRuntimeFunc("fadd64"), types.TFLOAT64},
		ssa.OpMul32F: {typecheck.LookupRuntimeFunc("fmul32"), types.TFLOAT32},

// (x-C) to (x+(-C)).

// x + (C + z) -> C + (x + z)
(Add64 (Add64 i:(Const64 <t>) z) x) && (z.Op != OpConst64 && x.Op != OpConst64) => (Add64 i (Add64 <t> z x))
(Add32 (Add32 i:(Const32 <t>) z) x) && (z.Op != OpConst32 && x.Op != OpConst32) => (Add32 i (Add32 <t> z x))
(Add16 (Add16 i:(Const16 <t>) z) x) && (z.Op != OpConst16 && x.Op != OpConst16) => (Add16 i (Add16 <t> z x))

	uaddr := uintptr(unsafe.Pointer(addr))
	addr32 := (*uint32)(unsafe.Pointer(uaddr &^ 3))
	word := uint32(v) << ((uaddr & 3) * 8) // little endian
	for {
		old := *addr32
		if Cas(addr32, old, old|word) {
			return
		}
	}
}

//go:nosplit
func And8(addr *uint8, v uint8) {
	// Align down to 4 bytes and use 32-bit CAS.
	uaddr := uintptr(unsafe.Pointer(addr))
	addr32 := (*uint32)(unsafe.Pointer(uaddr &^ 3))

  // CHECK-NEXT: %[[ADDN2:.*]] = "tf.AddN"(%[[ADDN]], %[[ZEROS_LIKE]]) : (tensor<10xf32>, tensor<10xf32>) -> tensor<10xf32>
  %addn2 = "tf.AddN"(%addn, %zeros-like) : (tensor<!tf_type.variant<tensor<f32>>>, tensor<!tf_type.variant<tensor<f32>>>) -> tensor<!tf_type.variant<tensor<f32>>>
  %stack = "tf.TensorListStack"(%addn2, %elem_shape) : (tensor<!tf_type.variant<tensor<f32>>>, tensor<0xi32>) -> tensor<10xf32>

					continue
				}
				v.reset(OpAdd32)
				v.AddArg2(y, z)
				return true
			}
		}
		break
	}
	// match: (Add32 (Add32 i:(Const32 <t>) z) x)
	// cond: (z.Op != OpConst32 && x.Op != OpConst32)
	// result: (Add32 i (Add32 <t> z x))
	for {
		for _i0 := 0; _i0 <= 1; _i0, v_0, v_1 = _i0+1, v_1, v_0 {
			if v_0.Op != OpAdd32 {
				continue
			}
			_ = v_0.Args[1]

// CHECK-DAG: [[cst_0:%.*]] = "tf.Const"() <{value = dense<2> : tensor<i32>}>
// CHECK: [[r0:%.*]] = "tf.AddV2"([[x]], [[cst]])
// CHECK: [[r1:%.*]] = "tf.AddV2"([[y]], [[r0]])
// CHECK: [[r2:%.*]] = "tf.AddV2"([[x]], [[cst_0]])
// CHECK: [[r3:%.*]] = "tf.AddV2"([[y]], [[r2]])
// CHECK: return [[r0]], [[r1]], [[r3]]

// CHECK-LABEL: func @merge_stateless

func parseIndVar(ind *Value) (min, inc, nxt *Value) {
	if ind.Op != OpPhi {
		return
	}

	if n := ind.Args[0]; (n.Op == OpAdd64 || n.Op == OpAdd32 || n.Op == OpAdd16 || n.Op == OpAdd8) && (n.Args[0] == ind || n.Args[1] == ind) {
		min, nxt = ind.Args[1], n
	} else if n := ind.Args[1]; (n.Op == OpAdd64 || n.Op == OpAdd32 || n.Op == OpAdd16 || n.Op == OpAdd8) && (n.Args[0] == ind || n.Args[1] == ind) {
		min, nxt = ind.Args[0], n
	} else {

	lock(&globalRand.lock)
	if !globalRand.init {
		fatal("randinit missed")
	}
	globalRand.state.Reseed()
	unlock(&globalRand.lock)
}

// rand32 is uint32(rand()), called from compiler-generated code.
//
//go:nosplit
func rand32() uint32 {
	return uint32(rand())
}

// rand returns a random uint64 from the per-m chacha8 state.
// Do not change signature: used via linkname from other packages.

func Add(x, y, carry uint) (sum, carryOut uint) {
	if UintSize == 32 {
		s32, c32 := Add32(uint32(x), uint32(y), uint32(carry))
		return uint(s32), uint(c32)
	}
	s64, c64 := Add64(uint64(x), uint64(y), uint64(carry))
	return uint(s64), uint(c64)
}

// Add32 returns the sum with carry of x, y and carry: sum = x + y + carry.
// The carry input must be 0 or 1; otherwise the behavior is undefined.

		// bit
		OpCtz8, OpCtz16, OpCtz32, OpCtz64,
		// mask
		OpSlicemask,
		// safety check
		OpIsNonNil,
		// not
		OpNot:
		return true
	case
		// add
		OpAdd64, OpAdd32, OpAdd16, OpAdd8,
		OpAdd32F, OpAdd64F,
		// sub
		OpSub64, OpSub32, OpSub16, OpSub8,
		OpSub32F, OpSub64F,
		// mul
		OpMul64, OpMul32, OpMul16, OpMul8,
		OpMul32F, OpMul64F,
		// div