again2:
	if q0 >= b || q0*vn0 > b*rhat+un0 {
		q0--
		rhat += vn1
		if rhat < b {
			goto again2
		}
	}

	return q1*b + q0, (un21*b + un0 - q0*v) >> s
}

func fadd32(x, y uint32) uint32 {
	return f64to32(fadd64(f32to64(x), f32to64(y)))
}

func fmul32(x, y uint32) uint32 {
	return f64to32(fmul64(f32to64(x), f32to64(y)))
}

func fdiv32(x, y uint32) uint32 {

(Mul64 x y) =>
	(Int64Make
		(Add32 <typ.UInt32>
			(Mul32 <typ.UInt32> (Int64Lo x) (Int64Hi y))
			(Add32 <typ.UInt32>
				(Mul32 <typ.UInt32> (Int64Hi x) (Int64Lo y))
				(Select0 <typ.UInt32> (Mul32uhilo (Int64Lo x) (Int64Lo y)))))
		(Select1 <typ.UInt32> (Mul32uhilo (Int64Lo x) (Int64Lo y))))

(And64 x y) =>
	(Int64Make
		(And32 <typ.UInt32> (Int64Hi x) (Int64Hi y))
		(And32 <typ.UInt32> (Int64Lo x) (Int64Lo y)))

  // 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>

	fail64("/", div64_ssa, a, b, 0.75)
	fail64("neg", neg64_ssa, a, b, -7)

	fail32("+", add32_ssa, c, d, 7.0)
	fail32("*", mul32_ssa, c, d, 12.0)
	fail32("-", sub32_ssa, c, d, -1.0)
	fail32("/", div32_ssa, c, d, 0.75)
	fail32("neg", neg32_ssa, c, d, -7)

	// denorm-squared should underflow to zero.
	fail32("*", mul32_ssa, tiny, tiny, 0)

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 {

		addr1 := v1.UnsafeAddr()
		addr2 := v2.UnsafeAddr()
		if addr1 > addr2 {
			// Canonicalize order to reduce number of entries in visited.
			addr1, addr2 = addr2, addr1
		}

		// Short circuit if references are identical ...
		if addr1 == addr2 {
			return true
		}

		// ... or already seen
		typ := v1.Type()
		v := visit{addr1, addr2, typ}
		if visited[v] {
			return true
		}

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

		map1, map2 = fake, fake
	}
	return
}

func TestConvertCPUProfile(t *testing.T) {
	addr1, addr2, map1, map2 := testPCs(t)

	b := []uint64{
		3, 0, 500, // hz = 500
		5, 0, 10, uint64(addr1 + 1), uint64(addr1 + 2), // 10 samples in addr1
		5, 0, 40, uint64(addr2 + 1), uint64(addr2 + 2), // 40 samples in addr2
		5, 0, 10, uint64(addr1 + 1), uint64(addr1 + 2), // 10 samples in addr1
	}

  // CHECK-NEXT: "tf.AddV2"({{.*}}, [[v]]) {device = "/CPU:0"} : (tensor<i32>, tensor<i32>) -> tensor<i32>
  // CHECK-NEXT: return
  %const = "tf.Const"() {device = "/CPU:0", value = dense<0> : tensor<i32>} : () -> tensor<i32>
  %x = "tf.AddV2"(%const, %const) {device = "/CPU:0"} : (tensor<i32>, tensor<i32>) -> tensor<i32>
  %r = "tf.AddV2"(%arg, %x) {device = "/CPU:0"} : (tensor<i32>, tensor<i32>) -> tensor<i32>