// input and value tensors. It also accepts the copy callback provided by
// pluggable vendor to do the copying of the tensors and the update callback to
// apply the arithmetic operation. The caller takes ownership of the `source`,
// `dest`, `tensor` and `value` tensors and is responsible for freeing them with
// TF_DeleteTensor.
TF_CAPI_EXPORT extern void TF_AssignUpdateVariable(

  let constructor = "mlir::quant::stablehlo::CreateQuantizeWeightPass()";
}

def UnfuseMhloBatchNormPass : Pass<"stablehlo-unfuse-mhlo-batch-norm", "mlir::func::FuncOp"> {
  let summary = "Unfuses batch normalization into arithmetic ops.";
}

def LiftQuantizableSpotsAsFunctionsPass : Pass<"stablehlo-lift-quantizable-spots-as-functions", "mlir::ModuleOp"> {
  let summary = "Replace quantization candidates with composite functions into the module.";

  let constructor = "CreateIdentifyDilatedConvPass()";
  let dependentDialects = ["TFL::TensorFlowLiteDialect"];
}

def GetArithmeticCountPass : Pass<"tfl-get-arithmetic-count", "mlir::func::FuncOp"> {
  let summary = "Calculate arithmetic count for tfl operations.";
  let constructor = "CreateGetArithmeticCountPass()";
  let dependentDialects = ["TFL::TensorFlowLiteDialect"];
}

		// Otherwise use the raw address, so that things like relative
		// jumps show up as JMP 0x123 instead of JMP f+10(SB).
		// It is usually easier to search for 0x123 than to do the mental
		// arithmetic to find f+10.
		addr := pc + uint64(inst.Len) + uint64(a)
		if s, base := symname(addr); s != "" && addr == base {
			return fmt.Sprintf("%s(SB)", s)
		}
		return fmt.Sprintf("%#x", addr)

	case Imm:

	}

	if hex {
		f, err := atofHex(s[:n], &float32info, mantissa, exp, neg, trunc)
		return float32(f), n, err
	}

	if optimize {
		// Try pure floating-point arithmetic conversion, and if that fails,
		// the Eisel-Lemire algorithm.
		if !trunc {
			if f, ok := atof32exact(mantissa, exp, neg); ok {
				return f, n, nil
			}
		}
		f, ok := eiselLemire32(mantissa, exp, neg)

 public:
  StringRef getArgument() const final { return "unfuse-mhlo-batch-norm-pass"; }
  StringRef getDescription() const final {
    return "Unfuses MHLO batch norm inference op into arithmetic ops";
  }
  void getDependentDialects(::mlir::DialectRegistry &registry) const override {
    // TODO(b/299514833): Remove TensorFlowDialect usage.
    registry.insert<shape::ShapeDialect, mlir::TF::TensorFlowDialect>();

		{5e-304 / 1e20, 'g', -1, "5e-324"},
		{-5e-304 / 1e20, 'g', -1, "-5e-324"},
		{fdiv(5e-304, 1e20), 'g', -1, "5e-324"},   // avoid constant arithmetic
		{fdiv(-5e-304, 1e20), 'g', -1, "-5e-324"}, // avoid constant arithmetic

		{32, 'g', -1, "32"},
		{32, 'g', 0, "3e+01"},

		{100, 'x', -1, "0x1.9p+06"},

		// {math.NaN(), 'g', -1, "NaN"},  // Float doesn't support NaNs

	_ = unsafe.Sizeof(func() { type _ [iota /* ERROR "iota outside constant decl" */ ]byte })
	_ = unsafe.Sizeof(func() { func() int { return iota /* ERROR "iota outside constant decl" */ }() })
)

// constant arithmetic precision and rounding must lead to expected (integer) results
var _ = []int64{
	0.0005 * 1e9,
	0.001 * 1e9,
	0.005 * 1e9,
	0.01 * 1e9,
	0.05 * 1e9,
	0.1 * 1e9,
	0.5 * 1e9,
	1 * 1e9,
	5 * 1e9,

  /// construct an Initializer. Initializer stores the value it got constructed
  /// with in a Tensor object.
  struct Initializer {
    /// Construct from a scalar value of an arithmetic type or a type that can
    /// be converted to a string (eg. a string literal).
    template <typename T, typename = typename std::enable_if<
                              std::is_arithmetic<T>::value ||

    // We extend the recursive expression for the one-variable case at Art of Computer Programming
    // vol. 2, Knuth, 4.2.2, (16) to the two-variable case. We have two value series x_i and y_i.
    // We define the arithmetic means X_n = 1/n \sum_{i=1}^n x_i, and Y_n = 1/n \sum_{i=1}^n y_i.
    // We also define the sum of the products of the differences from the means
    //           C_n = \sum_{i=1}^n x_i y_i - n X_n Y_n