//
    // Therefore we compute this as
    //
    //   (z + 1/2 - t(z) / log(t(z))) * log(t(z)).
    //
    // log_y = log_sqrt_two_pi + (z + one_half - t / log_t) * log_t + Log(x);
    Value t_div_log_t = rewriter.create<DivOp>(loc, t, log_t);
    Value one_half_minus_t_div_log_t =
        rewriter.create<SubOp>(loc, one_half, t_div_log_t);
    Value z_plus_one_half_minus_t_div_log_t =

	for i := 0; i < len(vf); i++ {
		a := Abs(vf[i])
		if f := Log10(a); !veryclose(log10[i], f) {
			t.Errorf("Log10(%g) = %g, want %g", a, f, log10[i])
		}
	}
	if f := Log10(E); f != Log10E {
		t.Errorf("Log10(%g) = %g, want %g", E, f, Log10E)
	}
	for i := 0; i < len(vflogSC); i++ {
		if f := Log10(vflogSC[i]); !alike(logSC[i], f) {
			t.Errorf("Log10(%g) = %g, want %g", vflogSC[i], f, logSC[i])
		}
	}
}

    // real(x) < 0 is fine for the complex case
    auto log_x = Where3(scope, NotEqual(scope, x, zero), Log(scope, x),
                        ZerosLike(scope, x));
    auto gy_1 = Mul(scope, Mul(scope, grad, z), log_x);
    return BinaryGradCommon(scope, op, grad_outputs, gx_1, gy_1);
  } else {
    // There's no sensible real value to return if x < 0, so return 0
    auto log_x = Where3(scope, Greater(scope, x, zero), Log(scope, x),

	// The probability distribution function is mean*exp(-mean*x), so the CDF is
	// p = 1 - exp(-mean*x), so
	// q = 1 - p == exp(-mean*x)
	// log_e(q) = -mean*x
	// -log_e(q)/mean = x
	// x = -log_e(q) * mean
	// x = log_2(q) * (-log_e(2)) * mean    ; Using log_2 for efficiency
	const randomBitCount = 26
	q := cheaprandn(1<<randomBitCount) + 1
	qlog := fastlog2(float64(q)) - randomBitCount
	if qlog > 0 {

            "Atan", "Atanh", "Ceil", "Cos", "Cosh", "Sin", "Exp", "Expm1",
            "Floor", "IsFinite", "IsInf", "IsNan", "Inv", "Reciprocal", "Log",
            "Log1p", "Invert", "LogicalNot", "Ndtri", "Neg", "Rint", "Round",
            "Rsqrt", "Sigmoid", "Sign", "Sinh", "Softplus", "Softsign", "Sqrt",
            "Square", "Tan", "Tanh", "Real", "Imag", "Erf", "Erfc", "Erfinv",

	// Make sure this stays in sync with the compiler's definition.
	count     int // # live cells == size of map.  Must be first (used by len() builtin)
	flags     uint8
	B         uint8  // log_2 of # of buckets (can hold up to loadFactor * 2^B items)
	noverflow uint16 // approximate number of overflow buckets; see incrnoverflow for details
	hash0     uint32 // hash seed

	return int64(ntz64(^x))
}

// logX returns logarithm of n base 2.
// n must be a positive power of 2 (isPowerOfTwoX returns true).
func log8(n int8) int64 {
	return int64(bits.Len8(uint8(n))) - 1
}
func log16(n int16) int64 {
	return int64(bits.Len16(uint16(n))) - 1
}
func log32(n int32) int64 {
	return int64(bits.Len32(uint32(n))) - 1
}
func log64(n int64) int64 {
	return int64(bits.Len64(uint64(n))) - 1
}

  // CHECK:  %[[IS_ZERO:.*]] = "tf.Equal"(%[[X]], %[[ZERO]]) <{incompatible_shape_error = true}> : (tensor<*xf32>, tensor<f32>) -> tensor<*xi1>
  // CHECK:  %[[LOG:.*]] = "tf.Log1p"(%[[Y]]) : (tensor<*xf32>) -> tensor<*xf32>
  // CHECK:  %[[MUL:.*]] = "tf.Mul"(%[[X]], %[[LOG]]) : (tensor<*xf32>, tensor<*xf32>) -> tensor<*xf32>