mqt-core 3.3.2__cp313-cp313t-win_amd64.whl

mqt/core/include/mqt-core/boost/multiprecision/rational_adaptor.hpp

@@ -0,0 +1,1289 @@
+///////////////////////////////////////////////////////////////
+//  Copyright 2020 John Maddock. Distributed under the Boost
+//  Software License, Version 1.0. (See accompanying file
+//  LICENSE_1_0.txt or copy at https://www.boost.org/LICENSE_1_0.txt
+
+#ifndef BOOST_MP_RATIONAL_ADAPTOR_HPP
+#define BOOST_MP_RATIONAL_ADAPTOR_HPP
+
+#include <boost/multiprecision/number.hpp>
+#include <boost/multiprecision/detail/hash.hpp>
+#include <boost/multiprecision/detail/float128_functions.hpp>
+#include <boost/multiprecision/detail/no_exceptions_support.hpp>
+
+namespace boost {
+namespace multiprecision {
+namespace backends {
+
+template <class Backend>
+struct rational_adaptor
+{
+   //
+   // Each backend need to declare 3 type lists which declare the types
+   // with which this can interoperate.  These lists must at least contain
+   // the widest type in each category - so "long long" must be the final
+   // type in the signed_types list for example.  Any narrower types if not
+   // present in the list will get promoted to the next wider type that is
+   // in the list whenever mixed arithmetic involving that type is encountered.
+   //
+   typedef typename Backend::signed_types    signed_types;
+   typedef typename Backend::unsigned_types  unsigned_types;
+   typedef typename Backend::float_types     float_types;
+
+   typedef typename std::tuple_element<0, unsigned_types>::type ui_type;
+
+   static Backend get_one()
+   {
+      Backend t;
+      t = static_cast<ui_type>(1);
+      return t;
+   }
+   static Backend get_zero()
+   {
+      Backend t;
+      t = static_cast<ui_type>(0);
+      return t;
+   }
+
+   static const Backend& one()
+   {
+      static const Backend result(get_one());
+      return result;
+   }
+   static const Backend& zero()
+   {
+      static const Backend result(get_zero());
+      return result;
+   }
+
+   void normalize()
+   {
+      using default_ops::eval_gcd;
+      using default_ops::eval_eq;
+      using default_ops::eval_divide;
+      using default_ops::eval_get_sign;
+
+      int s = eval_get_sign(m_denom);
+
+      if(s == 0)
+      {
+         BOOST_MP_THROW_EXCEPTION(std::overflow_error("Integer division by zero"));
+      }
+      else if (s < 0)
+      {
+         m_num.negate();
+         m_denom.negate();
+      }
+
+      Backend g, t;
+      eval_gcd(g, m_num, m_denom);
+      if (!eval_eq(g, one()))
+      {
+         eval_divide(t, m_num, g);
+         m_num.swap(t);
+         eval_divide(t, m_denom, g);
+         m_denom = std::move(t);
+      }
+   }
+
+   // We must have a default constructor:
+   rational_adaptor()
+      : m_num(zero()), m_denom(one()) {}
+
+   rational_adaptor(const rational_adaptor& o) : m_num(o.m_num), m_denom(o.m_denom) {}
+   rational_adaptor(rational_adaptor&& o) = default;
+
+   // Optional constructors, we can make this type slightly more efficient
+   // by providing constructors from any type we can handle natively.
+   // These will also cause number<> to be implicitly constructible
+   // from these types unless we make such constructors explicit.
+   //
+   template <class Arithmetic>
+   rational_adaptor(const Arithmetic& val, typename std::enable_if<std::is_constructible<Backend, Arithmetic>::value && !std::is_floating_point<Arithmetic>::value>::type const* = nullptr)
+      : m_num(val), m_denom(one()) {}
+
+   //
+   // Pass-through 2-arg construction of components:
+   //
+   template <class T, class U>
+   rational_adaptor(const T& a, const U& b, typename std::enable_if<std::is_constructible<Backend, T const&>::value && std::is_constructible<Backend, U const&>::value>::type const* = nullptr)
+      : m_num(a), m_denom(b) 
+   {
+      normalize();
+   }
+   template <class T, class U>
+   rational_adaptor(T&& a, const U& b, typename std::enable_if<std::is_constructible<Backend, T>::value && std::is_constructible<Backend, U>::value>::type const* = nullptr)
+      : m_num(static_cast<T&&>(a)), m_denom(b) 
+   {
+      normalize();
+   }
+   template <class T, class U>
+   rational_adaptor(T&& a, U&& b, typename std::enable_if<std::is_constructible<Backend, T>::value && std::is_constructible<Backend, U>::value>::type const* = nullptr)
+      : m_num(static_cast<T&&>(a)), m_denom(static_cast<U&&>(b)) 
+   {
+      normalize();
+   }
+   template <class T, class U>
+   rational_adaptor(const T& a, U&& b, typename std::enable_if<std::is_constructible<Backend, T>::value && std::is_constructible<Backend, U>::value>::type const* = nullptr)
+      : m_num(a), m_denom(static_cast<U&&>(b)) 
+   {
+      normalize();
+   }
+   //
+   // In the absense of converting constructors, operator= takes the strain.
+   // In addition to the usual suspects, there must be one operator= for each type
+   // listed in signed_types, unsigned_types, and float_types plus a string constructor.
+   //
+   rational_adaptor& operator=(const rational_adaptor& o) = default;
+   rational_adaptor& operator=(rational_adaptor&& o) = default;
+   template <class Arithmetic>
+   inline typename std::enable_if<!std::is_floating_point<Arithmetic>::value, rational_adaptor&>::type operator=(const Arithmetic& i)
+   {
+      m_num = i;
+      m_denom = one();
+      return *this;
+   }
+   rational_adaptor& operator=(const char* s)
+   {
+      using default_ops::eval_eq;
+
+      std::string                        s1;
+      multiprecision::number<Backend>    v1, v2;
+      char                               c;
+      bool                               have_hex = false;
+      const char* p = s; // saved for later
+
+      while ((0 != (c = *s)) && (c == 'x' || c == 'X' || c == '-' || c == '+' || (c >= '0' && c <= '9') || (have_hex && (c >= 'a' && c <= 'f')) || (have_hex && (c >= 'A' && c <= 'F'))))
+      {
+         if (c == 'x' || c == 'X')
+            have_hex = true;
+         s1.append(1, c);
+         ++s;
+      }
+      v1.assign(s1);
+      s1.erase();
+      if (c == '/')
+      {
+         ++s;
+         while ((0 != (c = *s)) && (c == 'x' || c == 'X' || c == '-' || c == '+' || (c >= '0' && c <= '9') || (have_hex && (c >= 'a' && c <= 'f')) || (have_hex && (c >= 'A' && c <= 'F'))))
+         {
+            if (c == 'x' || c == 'X')
+               have_hex = true;
+            s1.append(1, c);
+            ++s;
+         }
+         v2.assign(s1);
+      }
+      else
+         v2 = 1;
+      if (*s)
+      {
+         BOOST_MP_THROW_EXCEPTION(std::runtime_error(std::string("Could not parse the string \"") + p + std::string("\" as a valid rational number.")));
+      }
+      multiprecision::number<Backend> gcd;
+      eval_gcd(gcd.backend(), v1.backend(), v2.backend());
+      if (!eval_eq(gcd.backend(), one()))
+      {
+         v1 /= gcd;
+         v2 /= gcd;
+      }
+      num() = std::move(std::move(v1).backend());
+      denom() = std::move(std::move(v2).backend());
+      return *this;
+   }
+   template <class Float>
+   typename std::enable_if<std::is_floating_point<Float>::value, rational_adaptor&>::type operator=(Float i)
+   {
+      using default_ops::eval_eq;
+      BOOST_MP_FLOAT128_USING using std::floor; using std::frexp; using std::ldexp;
+
+      int   e;
+      Float f = frexp(i, &e);
+#ifdef BOOST_HAS_FLOAT128
+      f = ldexp(f, std::is_same<float128_type, Float>::value ? 113 : std::numeric_limits<Float>::digits);
+      e -= std::is_same<float128_type, Float>::value ? 113 : std::numeric_limits<Float>::digits;
+#else
+      f = ldexp(f, std::numeric_limits<Float>::digits);
+      e -= std::numeric_limits<Float>::digits;
+#endif
+      number<Backend> num(f);
+      number<Backend> denom(1u);
+      if (e > 0)
+      {
+         num <<= e;
+      }
+      else if (e < 0)
+      {
+         denom <<= -e;
+      }
+      number<Backend> gcd;
+      eval_gcd(gcd.backend(), num.backend(), denom.backend());
+      if (!eval_eq(gcd.backend(), one()))
+      {
+         num /= gcd;
+         denom /= gcd;
+      }
+      this->num() = std::move(std::move(num).backend());
+      this->denom() = std::move(std::move(denom).backend());
+      return *this;
+   }
+
+   void swap(rational_adaptor& o)
+   {
+      m_num.swap(o.m_num);
+      m_denom.swap(o.m_denom);
+   }
+   std::string str(std::streamsize digits, std::ios_base::fmtflags f) const
+   {
+      using default_ops::eval_eq;
+      //
+      // We format the string ourselves so we can match what GMP's mpq type does:
+      //
+      std::string result = num().str(digits, f);
+      if (!eval_eq(denom(), one()))
+      {
+         result.append(1, '/');
+         result.append(denom().str(digits, f));
+      }
+      return result;
+   }
+   void negate()
+   {
+      m_num.negate();
+   }
+   int compare(const rational_adaptor& o) const
+   {
+      std::ptrdiff_t s1 = eval_get_sign(*this);
+      std::ptrdiff_t s2 = eval_get_sign(o);
+      if (s1 != s2)
+      {
+         return s1 < s2 ? -1 : 1;
+      }
+      else if (s1 == 0)
+         return 0; // both zero.
+
+      bool neg = false;
+      if (s1 >= 0)
+      {
+         s1 = eval_msb(num()) + eval_msb(o.denom());
+         s2 = eval_msb(o.num()) + eval_msb(denom());
+      }
+      else
+      {
+         Backend t(num());
+         t.negate();
+         s1 = eval_msb(t) + eval_msb(o.denom());
+         t = o.num();
+         t.negate();
+         s2 = eval_msb(t) + eval_msb(denom());
+         neg = true;
+      }
+      s1 -= s2;
+      if (s1 < -1)
+         return neg ? 1 : -1;
+      else if (s1 > 1)
+         return neg ? -1 : 1;
+
+      Backend t1, t2;
+      eval_multiply(t1, num(), o.denom());
+      eval_multiply(t2, o.num(), denom());
+      return t1.compare(t2);
+   }
+   //
+   // Comparison with arithmetic types, default just constructs a temporary:
+   //
+   template <class A>
+   typename std::enable_if<boost::multiprecision::detail::is_arithmetic<A>::value, int>::type compare(A i) const
+   {
+      rational_adaptor t;
+      t = i;  //  Note: construct directly from i if supported.
+      return compare(t);
+   }
+
+   Backend& num() { return m_num; }
+   const Backend& num()const { return m_num; }
+   Backend& denom() { return m_denom; }
+   const Backend& denom()const { return m_denom; }
+
+   #ifndef BOOST_MP_STANDALONE
+   template <class Archive>
+   void serialize(Archive& ar, const std::integral_constant<bool, true>&)
+   {
+      // Saving
+      number<Backend> n(num()), d(denom());
+      ar& boost::make_nvp("numerator", n);
+      ar& boost::make_nvp("denominator", d);
+   }
+   template <class Archive>
+   void serialize(Archive& ar, const std::integral_constant<bool, false>&)
+   {
+      // Loading
+      number<Backend> n, d;
+      ar& boost::make_nvp("numerator", n);
+      ar& boost::make_nvp("denominator", d);
+      num() = n.backend();
+      denom() = d.backend();
+   }
+   template <class Archive>
+   void serialize(Archive& ar, const unsigned int /*version*/)
+   {
+      using tag = typename Archive::is_saving;
+      using saving_tag = std::integral_constant<bool, tag::value>;
+      serialize(ar, saving_tag());
+   }
+   #endif // BOOST_MP_STANDALONE
+   
+ private:
+   Backend m_num, m_denom;
+};
+
+//
+// Helpers:
+//
+template <class T>
+inline constexpr typename std::enable_if<std::numeric_limits<T>::is_specialized && !std::numeric_limits<T>::is_signed, bool>::type
+is_minus_one(const T&)
+{
+   return false;
+}
+template <class T>
+inline constexpr typename std::enable_if<!std::numeric_limits<T>::is_specialized || std::numeric_limits<T>::is_signed, bool>::type
+is_minus_one(const T& val)
+{
+   return val == -1;
+}
+
+//
+// Required non-members:
+//
+template <class Backend> 
+inline void eval_add(rational_adaptor<Backend>& a, const rational_adaptor<Backend>& b)
+{
+   eval_add_subtract_imp(a, a, b, true);
+}
+template <class Backend> 
+inline void eval_subtract(rational_adaptor<Backend>& a, const rational_adaptor<Backend>& b)
+{
+   eval_add_subtract_imp(a, a, b, false);
+}
+
+template <class Backend> 
+inline void eval_multiply(rational_adaptor<Backend>& a, const rational_adaptor<Backend>& b)
+{
+   eval_multiply_imp(a, a, b.num(), b.denom());
+}
+
+template <class Backend> 
+void eval_divide(rational_adaptor<Backend>& a, const rational_adaptor<Backend>& b)
+{
+   using default_ops::eval_divide;
+   rational_adaptor<Backend> t;
+   eval_divide(t, a, b);
+   a = std::move(t);
+}
+//
+// Conversions:
+//
+template <class R, class IntBackend>
+inline typename std::enable_if<number_category<R>::value == number_kind_floating_point>::type eval_convert_to(R* result, const rational_adaptor<IntBackend>& backend)
+{
+   //
+   // The generic conversion is as good as anything we can write here:
+   //
+   ::boost::multiprecision::detail::generic_convert_rational_to_float(*result, backend);
+}
+
+template <class R, class IntBackend>
+inline typename std::enable_if<(number_category<R>::value != number_kind_integer) && (number_category<R>::value != number_kind_floating_point) && !std::is_enum<R>::value>::type eval_convert_to(R* result, const rational_adaptor<IntBackend>& backend)
+{
+   using default_ops::eval_convert_to;
+   R d;
+   eval_convert_to(result, backend.num());
+   eval_convert_to(&d, backend.denom());
+   *result /= d;
+}
+
+template <class R, class Backend>
+inline typename std::enable_if<number_category<R>::value == number_kind_integer>::type eval_convert_to(R* result, const rational_adaptor<Backend>& backend)
+{
+   using default_ops::eval_divide;
+   using default_ops::eval_convert_to;
+   Backend t;
+   eval_divide(t, backend.num(), backend.denom());
+   eval_convert_to(result, t);
+}
+
+//
+// Hashing support, not strictly required, but it is used in our tests:
+//
+template <class Backend>
+inline std::size_t hash_value(const rational_adaptor<Backend>& arg)
+{
+   std::size_t result = hash_value(arg.num());
+   std::size_t result2 = hash_value(arg.denom());
+   boost::multiprecision::detail::hash_combine(result, result2);
+   return result;
+}
+//
+// assign_components:
+//
+template <class Backend>
+void assign_components(rational_adaptor<Backend>& result, Backend const& a, Backend const& b)
+{
+   using default_ops::eval_gcd;
+   using default_ops::eval_divide;
+   using default_ops::eval_eq;
+   using default_ops::eval_is_zero;
+   using default_ops::eval_get_sign;
+
+   if (eval_is_zero(b))
+   {
+      BOOST_MP_THROW_EXCEPTION(std::overflow_error("Integer division by zero"));
+   }
+   Backend g;
+   eval_gcd(g, a, b);
+   if (eval_eq(g, rational_adaptor<Backend>::one()))
+   {
+      result.num() = a;
+      result.denom() = b;
+   }
+   else
+   {
+      eval_divide(result.num(), a, g);
+      eval_divide(result.denom(), b, g);
+   }
+   if (eval_get_sign(result.denom()) < 0)
+   {
+      result.num().negate();
+      result.denom().negate();
+   }
+}
+//
+// Again for arithmetic types, overload for whatever arithmetic types are directly supported:
+//
+template <class Backend, class Arithmetic1, class Arithmetic2>
+inline void assign_components(rational_adaptor<Backend>& result, const Arithmetic1& a, typename std::enable_if<std::is_arithmetic<Arithmetic1>::value && std::is_arithmetic<Arithmetic2>::value, const Arithmetic2&>::type b)
+{
+   using default_ops::eval_gcd;
+   using default_ops::eval_divide;
+   using default_ops::eval_eq;
+
+   if (b == 0)
+   {
+      BOOST_MP_THROW_EXCEPTION(std::overflow_error("Integer division by zero"));
+   }
+
+   Backend g;
+   result.num()   = a;
+   eval_gcd(g, result.num(), b);
+   if (eval_eq(g, rational_adaptor<Backend>::one()))
+   {
+      result.denom() = b;
+   }
+   else
+   {
+      eval_divide(result.num(), g);
+      eval_divide(result.denom(), b, g);
+   }
+   if (eval_get_sign(result.denom()) < 0)
+   {
+      result.num().negate();
+      result.denom().negate();
+   }
+}
+template <class Backend, class Arithmetic1, class Arithmetic2>
+inline void assign_components(rational_adaptor<Backend>& result, const Arithmetic1& a, typename std::enable_if<!std::is_arithmetic<Arithmetic1>::value || !std::is_arithmetic<Arithmetic2>::value, const Arithmetic2&>::type b)
+{
+   using default_ops::eval_gcd;
+   using default_ops::eval_divide;
+   using default_ops::eval_eq;
+
+   Backend g;
+   result.num()   = a;
+   result.denom() = b;
+
+   if (eval_get_sign(result.denom()) == 0)
+   {
+      BOOST_MP_THROW_EXCEPTION(std::overflow_error("Integer division by zero"));
+   }
+
+   eval_gcd(g, result.num(), result.denom());
+   if (!eval_eq(g, rational_adaptor<Backend>::one()))
+   {
+      eval_divide(result.num(), g);
+      eval_divide(result.denom(), g);
+   }
+   if (eval_get_sign(result.denom()) < 0)
+   {
+      result.num().negate();
+      result.denom().negate();
+   }
+}
+//
+// Optional comparison operators:
+//
+template <class Backend>
+inline bool eval_is_zero(const rational_adaptor<Backend>& arg)
+{
+   using default_ops::eval_is_zero;
+   return eval_is_zero(arg.num());
+}
+
+template <class Backend>
+inline int eval_get_sign(const rational_adaptor<Backend>& arg)
+{
+   using default_ops::eval_get_sign;
+   return eval_get_sign(arg.num());
+}
+
+template <class Backend>
+inline bool eval_eq(const rational_adaptor<Backend>& a, const rational_adaptor<Backend>& b)
+{
+   using default_ops::eval_eq;
+   return eval_eq(a.num(), b.num()) && eval_eq(a.denom(), b.denom());
+}
+
+template <class Backend, class Arithmetic>
+inline typename std::enable_if<std::is_convertible<Arithmetic, Backend>::value&& std::is_integral<Arithmetic>::value, bool>::type 
+   eval_eq(const rational_adaptor<Backend>& a, Arithmetic b)
+{
+   using default_ops::eval_eq;
+   return eval_eq(a.denom(), rational_adaptor<Backend>::one()) && eval_eq(a.num(), b);
+}
+
+template <class Backend, class Arithmetic>
+inline typename std::enable_if<std::is_convertible<Arithmetic, Backend>::value&& std::is_integral<Arithmetic>::value, bool>::type 
+   eval_eq(Arithmetic b, const rational_adaptor<Backend>& a)
+{
+   using default_ops::eval_eq;
+   return eval_eq(a.denom(), rational_adaptor<Backend>::one()) && eval_eq(a.num(), b);
+}
+
+//
+// Arithmetic operations, starting with addition:
+//
+template <class Backend, class Arithmetic> 
+void eval_add_subtract_imp(rational_adaptor<Backend>& result, const Arithmetic& arg, bool isaddition)
+{
+   using default_ops::eval_multiply;
+   using default_ops::eval_divide;
+   using default_ops::eval_add;
+   using default_ops::eval_gcd;
+   Backend t;
+   eval_multiply(t, result.denom(), arg);
+   if (isaddition)
+      eval_add(result.num(), t);
+   else
+      eval_subtract(result.num(), t);
+   //
+   // There is no need to re-normalize here, we have 
+   // (a + bm) / b
+   // and gcd(a + bm, b) = gcd(a, b) = 1
+   //
+   /*
+   eval_gcd(t, result.num(), result.denom());
+   if (!eval_eq(t, rational_adaptor<Backend>::one()) != 0)
+   {
+      Backend t2;
+      eval_divide(t2, result.num(), t);
+      t2.swap(result.num());
+      eval_divide(t2, result.denom(), t);
+      t2.swap(result.denom());
+   }
+   */
+}
+
+template <class Backend, class Arithmetic> 
+inline typename std::enable_if<std::is_convertible<Arithmetic, Backend>::value && (std::is_integral<Arithmetic>::value || std::is_same<Arithmetic, Backend>::value)>::type
+   eval_add(rational_adaptor<Backend>& result, const Arithmetic& arg)
+{
+   eval_add_subtract_imp(result, arg, true);
+}
+
+template <class Backend, class Arithmetic> 
+inline typename std::enable_if<std::is_convertible<Arithmetic, Backend>::value && (std::is_integral<Arithmetic>::value || std::is_same<Arithmetic, Backend>::value)>::type
+   eval_subtract(rational_adaptor<Backend>& result, const Arithmetic& arg)
+{
+   eval_add_subtract_imp(result, arg, false);
+}
+
+template <class Backend>
+void eval_add_subtract_imp(rational_adaptor<Backend>& result, const rational_adaptor<Backend>& a, const rational_adaptor<Backend>& b, bool isaddition)
+{
+   using default_ops::eval_eq;
+   using default_ops::eval_multiply;
+   using default_ops::eval_divide;
+   using default_ops::eval_add;
+   using default_ops::eval_subtract;
+   //
+   // Let  a = an/ad
+   //      b = bn/bd
+   //      g = gcd(ad, bd)
+   // result = rn/rd
+   //
+   // Then:
+   // rn = an * (bd/g) + bn * (ad/g)
+   // rd = ad * (bd/g)
+   //    = (ad/g) * (bd/g) * g
+   //
+   // And the whole thing can then be rescaled by
+   //      gcd(rn, g)
+   //
+   Backend gcd, t1, t2, t3, t4;
+   //
+   // Begin by getting the gcd of the 2 denominators:
+   //
+   eval_gcd(gcd, a.denom(), b.denom());
+   //
+   // Do we have gcd > 1:
+   //
+   if (!eval_eq(gcd, rational_adaptor<Backend>::one()))
+   {
+      //
+      // Scale the denominators by gcd, and put the results in t1 and t2:
+      //
+      eval_divide(t1, b.denom(), gcd);
+      eval_divide(t2, a.denom(), gcd);
+      //
+      // multiply the numerators by the scale denominators and put the results in t3, t4:
+      //
+      eval_multiply(t3, a.num(), t1);
+      eval_multiply(t4, b.num(), t2);
+      //
+      // Add them up:
+      //
+      if (isaddition)
+         eval_add(t3, t4);
+      else
+         eval_subtract(t3, t4);
+      //
+      // Get the gcd of gcd and our numerator (t3):
+      //
+      eval_gcd(t4, t3, gcd);
+      if (eval_eq(t4, rational_adaptor<Backend>::one()))
+      {
+         result.num() = t3;
+         eval_multiply(result.denom(), t1, a.denom());
+      }
+      else
+      {
+         //
+         // Uncommon case where gcd is not 1, divide the numerator
+         // and the denominator terms by the new gcd.  Note we perform division
+         // on the existing gcd value as this is the smallest of the 3 denominator
+         // terms we'll be multiplying together, so there's a good chance it's a
+         // single limb value already:
+         //
+         eval_divide(result.num(), t3, t4);
+         eval_divide(t3, gcd, t4);
+         eval_multiply(t4, t1, t2);
+         eval_multiply(result.denom(), t4, t3);
+      }
+   }
+   else
+   {
+      //
+      // Most common case (approx 60%) where gcd is one:
+      //
+      eval_multiply(t1, a.num(), b.denom());
+      eval_multiply(t2, a.denom(), b.num());
+      if (isaddition)
+         eval_add(result.num(), t1, t2);
+      else
+         eval_subtract(result.num(), t1, t2);
+      eval_multiply(result.denom(), a.denom(), b.denom());
+   }
+}
+
+
+template <class Backend>
+inline void eval_add(rational_adaptor<Backend>& result, const rational_adaptor<Backend>& a, const rational_adaptor<Backend>& b)
+{
+   eval_add_subtract_imp(result, a, b, true);
+}
+template <class Backend>
+inline void eval_subtract(rational_adaptor<Backend>& result, const rational_adaptor<Backend>& a, const rational_adaptor<Backend>& b)
+{
+   eval_add_subtract_imp(result, a, b, false);
+}
+
+template <class Backend, class Arithmetic>
+void eval_add_subtract_imp(rational_adaptor<Backend>& result, const rational_adaptor<Backend>& a, const Arithmetic& b, bool isaddition)
+{
+   using default_ops::eval_add;
+   using default_ops::eval_subtract;
+   using default_ops::eval_multiply;
+
+   if (&result == &a)
+      return eval_add_subtract_imp(result, b, isaddition);
+
+   eval_multiply(result.num(), a.denom(), b);
+   if (isaddition)
+      eval_add(result.num(), a.num());
+   else
+      BOOST_IF_CONSTEXPR(std::numeric_limits<Backend>::is_signed == false)
+   {
+      Backend t;
+      eval_subtract(t, a.num(), result.num());
+      result.num() = std::move(t);
+   }
+   else
+   {
+      eval_subtract(result.num(), a.num());
+      result.negate();
+   }
+   result.denom() = a.denom();
+   //
+   // There is no need to re-normalize here, we have 
+   // (a + bm) / b
+   // and gcd(a + bm, b) = gcd(a, b) = 1
+   //
+}
+template <class Backend, class Arithmetic>
+inline typename std::enable_if<std::is_convertible<Arithmetic, Backend>::value && (std::is_integral<Arithmetic>::value || std::is_same<Arithmetic, Backend>::value)>::type
+   eval_add(rational_adaptor<Backend>& result, const rational_adaptor<Backend>& a, const Arithmetic& b)
+{
+   eval_add_subtract_imp(result, a, b, true);
+}
+template <class Backend, class Arithmetic>
+inline typename std::enable_if<std::is_convertible<Arithmetic, Backend>::value && (std::is_integral<Arithmetic>::value || std::is_same<Arithmetic, Backend>::value)>::type
+   eval_subtract(rational_adaptor<Backend>& result, const rational_adaptor<Backend>& a, const Arithmetic& b)
+{
+   eval_add_subtract_imp(result, a, b, false);
+}
+
+//
+// Multiplication:
+//
+template <class Backend> 
+void eval_multiply_imp(rational_adaptor<Backend>& result, const rational_adaptor<Backend>& a, const Backend& b_num, const Backend& b_denom)
+{
+   using default_ops::eval_multiply;
+   using default_ops::eval_divide;
+   using default_ops::eval_gcd;
+   using default_ops::eval_get_sign;
+   using default_ops::eval_eq;
+
+   Backend gcd_left, gcd_right, t1, t2;
+   eval_gcd(gcd_left, a.num(), b_denom);
+   eval_gcd(gcd_right, b_num, a.denom());
+   //
+   // Unit gcd's are the most likely case:
+   //
+   bool b_left = eval_eq(gcd_left, rational_adaptor<Backend>::one());
+   bool b_right = eval_eq(gcd_right, rational_adaptor<Backend>::one());
+
+   if (b_left && b_right)
+   {
+      eval_multiply(result.num(), a.num(), b_num);
+      eval_multiply(result.denom(), a.denom(), b_denom);
+   }
+   else if (b_left)
+   {
+      eval_divide(t2, b_num, gcd_right);
+      eval_multiply(result.num(), a.num(), t2);
+      eval_divide(t1, a.denom(), gcd_right);
+      eval_multiply(result.denom(), t1, b_denom);
+   }
+   else if (b_right)
+   {
+      eval_divide(t1, a.num(), gcd_left);
+      eval_multiply(result.num(), t1, b_num);
+      eval_divide(t2, b_denom, gcd_left);
+      eval_multiply(result.denom(), a.denom(), t2);
+   }
+   else
+   {
+      eval_divide(t1, a.num(), gcd_left);
+      eval_divide(t2, b_num, gcd_right);
+      eval_multiply(result.num(), t1, t2);
+      eval_divide(t1, a.denom(), gcd_right);
+      eval_divide(t2, b_denom, gcd_left);
+      eval_multiply(result.denom(), t1, t2);
+   }
+   //
+   // We may have b_denom negative if this is actually division, if so just correct things now:
+   //
+   if (eval_get_sign(b_denom) < 0)
+   {
+      result.num().negate();
+      result.denom().negate();
+   }
+}
+
+template <class Backend> 
+void eval_multiply(rational_adaptor<Backend>& result, const rational_adaptor<Backend>& a, const rational_adaptor<Backend>& b)
+{
+   using default_ops::eval_multiply;
+
+   if (&a == &b)
+   {
+      // squaring, gcd's are 1:
+      eval_multiply(result.num(), a.num(), b.num());
+      eval_multiply(result.denom(), a.denom(), b.denom());
+      return;
+   }
+   eval_multiply_imp(result, a, b.num(), b.denom());
+}
+
+template <class Backend, class Arithmetic> 
+void eval_multiply_imp(Backend& result_num, Backend& result_denom, Arithmetic arg)
+{
+   if (arg == 0)
+   {
+      result_num = rational_adaptor<Backend>::zero();
+      result_denom = rational_adaptor<Backend>::one();
+      return;
+   }
+   else if (arg == 1)
+      return;
+
+   using default_ops::eval_multiply;
+   using default_ops::eval_divide;
+   using default_ops::eval_gcd;
+   using default_ops::eval_convert_to;
+
+   Backend gcd, t;
+   Arithmetic integer_gcd;
+   eval_gcd(gcd, result_denom, arg);
+   eval_convert_to(&integer_gcd, gcd);
+   arg /= integer_gcd;
+   if (boost::multiprecision::detail::unsigned_abs(arg) > 1)
+   {
+      eval_multiply(t, result_num, arg);
+      result_num = std::move(t);
+   }
+   else if (is_minus_one(arg))
+      result_num.negate();
+   if (integer_gcd > 1)
+   {
+      eval_divide(t, result_denom, integer_gcd);
+      result_denom = std::move(t);
+   }
+}
+template <class Backend> 
+void eval_multiply_imp(Backend& result_num, Backend& result_denom, Backend arg)
+{
+   using default_ops::eval_multiply;
+   using default_ops::eval_divide;
+   using default_ops::eval_gcd;
+   using default_ops::eval_convert_to;
+   using default_ops::eval_is_zero;
+   using default_ops::eval_eq;
+   using default_ops::eval_get_sign;
+
+   if (eval_is_zero(arg))
+   {
+      result_num = rational_adaptor<Backend>::zero();
+      result_denom = rational_adaptor<Backend>::one();
+      return;
+   }
+   else if (eval_eq(arg, rational_adaptor<Backend>::one()))
+      return;
+
+   Backend gcd, t;
+   eval_gcd(gcd, result_denom, arg);
+   if (!eval_eq(gcd, rational_adaptor<Backend>::one()))
+   {
+      eval_divide(t, arg, gcd);
+      arg = t;
+   }
+   else
+      t = arg;
+   if (eval_get_sign(arg) < 0)
+      t.negate();
+
+   if (!eval_eq(t, rational_adaptor<Backend>::one()))
+   {
+      eval_multiply(t, result_num, arg);
+      result_num = std::move(t);
+   }
+   else if (eval_get_sign(arg) < 0)
+      result_num.negate();
+   if (!eval_eq(gcd, rational_adaptor<Backend>::one()))
+   {
+      eval_divide(t, result_denom, gcd);
+      result_denom = std::move(t);
+   }
+}
+
+template <class Backend, class Arithmetic> 
+inline typename std::enable_if<std::is_convertible<Arithmetic, Backend>::value && (std::is_integral<Arithmetic>::value || std::is_same<Arithmetic, Backend>::value)>::type
+   eval_multiply(rational_adaptor<Backend>& result, const Arithmetic& arg)
+{
+   eval_multiply_imp(result.num(), result.denom(), arg);
+}
+
+template <class Backend, class Arithmetic> 
+typename std::enable_if<std::is_convertible<Arithmetic, Backend>::value && std::is_integral<Arithmetic>::value>::type
+   eval_multiply_imp(rational_adaptor<Backend>& result, const Backend& a_num, const Backend& a_denom, Arithmetic b)
+{
+   if (b == 0)
+   {
+      result.num() = rational_adaptor<Backend>::zero();
+      result.denom() = rational_adaptor<Backend>::one();
+      return;
+   }
+   else if (b == 1)
+   {
+      result.num() = a_num;
+      result.denom() = a_denom;
+      return;
+   }
+
+   using default_ops::eval_multiply;
+   using default_ops::eval_divide;
+   using default_ops::eval_gcd;
+   using default_ops::eval_convert_to;
+
+   Backend gcd;
+   Arithmetic integer_gcd;
+   eval_gcd(gcd, a_denom, b);
+   eval_convert_to(&integer_gcd, gcd);
+   b /= integer_gcd;
+   if (boost::multiprecision::detail::unsigned_abs(b) > 1)
+      eval_multiply(result.num(), a_num, b);
+   else if (is_minus_one(b))
+   {
+      result.num() = a_num;
+      result.num().negate();
+   }
+   else
+      result.num() = a_num;
+   if (integer_gcd > 1)
+      eval_divide(result.denom(), a_denom, integer_gcd);
+   else
+      result.denom() = a_denom;
+}
+template <class Backend> 
+inline void eval_multiply_imp(rational_adaptor<Backend>& result, const Backend& a_num, const Backend& a_denom, const Backend& b)
+{
+   result.num() = a_num;
+   result.denom() = a_denom;
+   eval_multiply_imp(result.num(), result.denom(), b);
+}
+
+template <class Backend, class Arithmetic> 
+inline typename std::enable_if<std::is_convertible<Arithmetic, Backend>::value && (std::is_integral<Arithmetic>::value || std::is_same<Arithmetic, Backend>::value)>::type
+   eval_multiply(rational_adaptor<Backend>& result, const rational_adaptor<Backend>& a, const Arithmetic& b)
+{
+   if (&result == &a)
+      return eval_multiply(result, b);
+
+   eval_multiply_imp(result, a.num(), a.denom(), b);
+}
+
+template <class Backend, class Arithmetic> 
+inline typename std::enable_if<std::is_convertible<Arithmetic, Backend>::value && (std::is_integral<Arithmetic>::value || std::is_same<Arithmetic, Backend>::value)>::type
+   eval_multiply(rational_adaptor<Backend>& result, const Arithmetic& b, const rational_adaptor<Backend>& a)
+{
+   return eval_multiply(result, a, b);
+}
+
+//
+// Division:
+//
+template <class Backend>
+inline void eval_divide(rational_adaptor<Backend>& result, const rational_adaptor<Backend>& a, const rational_adaptor<Backend>& b)
+{
+   using default_ops::eval_multiply;
+   using default_ops::eval_get_sign;
+
+   if (eval_get_sign(b.num()) == 0)
+   {
+      BOOST_MP_THROW_EXCEPTION(std::overflow_error("Integer division by zero"));
+      return;
+   }
+   if (&a == &b)
+   {
+      // Huh? Really?
+      result.num() = result.denom() = rational_adaptor<Backend>::one();
+      return;
+   }
+   if (&result == &b)
+   {
+      rational_adaptor<Backend> t(b);
+      return eval_divide(result, a, t);
+   }
+   eval_multiply_imp(result, a, b.denom(), b.num());
+}
+
+template <class Backend, class Arithmetic> 
+inline typename std::enable_if<std::is_convertible<Arithmetic, Backend>::value && (std::is_integral<Arithmetic>::value || std::is_same<Arithmetic, Backend>::value)>::type
+   eval_divide(rational_adaptor<Backend>& result, const Arithmetic& b, const rational_adaptor<Backend>& a)
+{
+   using default_ops::eval_get_sign;
+
+   if (eval_get_sign(a.num()) == 0)
+   {
+      BOOST_MP_THROW_EXCEPTION(std::overflow_error("Integer division by zero"));
+      return;
+   }
+   if (&a == &result)
+   {
+      eval_multiply_imp(result.denom(), result.num(), b);
+      result.num().swap(result.denom());
+   }
+   else
+      eval_multiply_imp(result, a.denom(), a.num(), b);
+
+   if (eval_get_sign(result.denom()) < 0)
+   {
+      result.num().negate();
+      result.denom().negate();
+   }
+}
+
+template <class Backend, class Arithmetic>
+typename std::enable_if<std::is_convertible<Arithmetic, Backend>::value && std::is_integral<Arithmetic>::value>::type
+eval_divide(rational_adaptor<Backend>& result, Arithmetic arg)
+{
+   if (arg == 0)
+   {
+      BOOST_MP_THROW_EXCEPTION(std::overflow_error("Integer division by zero"));
+      return;
+   }
+   else if (arg == 1)
+      return;
+   else if (is_minus_one(arg))
+   {
+      result.negate();
+      return;
+   }
+   if (eval_get_sign(result) == 0)
+   {
+      return;
+   }
+
+
+   using default_ops::eval_multiply;
+   using default_ops::eval_gcd;
+   using default_ops::eval_convert_to;
+   using default_ops::eval_divide;
+
+   Backend gcd, t;
+   Arithmetic integer_gcd;
+   eval_gcd(gcd, result.num(), arg);
+   eval_convert_to(&integer_gcd, gcd);
+   arg /= integer_gcd;
+
+   eval_multiply(t, result.denom(), boost::multiprecision::detail::unsigned_abs(arg));
+   result.denom() = std::move(t);
+   if (arg < 0)
+   {
+      result.num().negate();
+   }
+   if (integer_gcd > 1)
+   {
+      eval_divide(t, result.num(), integer_gcd);
+      result.num() = std::move(t);
+   }
+}
+template <class Backend>
+void eval_divide(rational_adaptor<Backend>& result, const rational_adaptor<Backend>& a, Backend arg)
+{
+   using default_ops::eval_multiply;
+   using default_ops::eval_gcd;
+   using default_ops::eval_convert_to;
+   using default_ops::eval_divide;
+   using default_ops::eval_is_zero;
+   using default_ops::eval_eq;
+   using default_ops::eval_get_sign;
+
+   if (eval_is_zero(arg))
+   {
+      BOOST_MP_THROW_EXCEPTION(std::overflow_error("Integer division by zero"));
+      return;
+   }
+   else if (eval_eq(a, rational_adaptor<Backend>::one()) || (eval_get_sign(a) == 0))
+   {
+      if (&result != &a)
+         result = a;
+      result.denom() = arg;
+      return;
+   }
+
+   Backend gcd, u_arg, t;
+   eval_gcd(gcd, a.num(), arg);
+   bool has_unit_gcd = eval_eq(gcd, rational_adaptor<Backend>::one());
+   if (!has_unit_gcd)
+   {
+      eval_divide(u_arg, arg, gcd);
+      arg = u_arg;
+   }
+   else
+      u_arg = arg;
+   if (eval_get_sign(u_arg) < 0)
+      u_arg.negate();
+
+   eval_multiply(t, a.denom(), u_arg);
+   result.denom() = std::move(t);
+   
+   if (!has_unit_gcd)
+   {
+      eval_divide(t, a.num(), gcd);
+      result.num() = std::move(t);
+   }
+   else if (&result != &a)
+      result.num() = a.num();
+
+   if (eval_get_sign(arg) < 0)
+   {
+      result.num().negate();
+   }
+}
+template <class Backend>
+void eval_divide(rational_adaptor<Backend>& result, const Backend& arg)
+{
+   eval_divide(result, result, arg);
+}
+
+template <class Backend, class Arithmetic>
+typename std::enable_if<std::is_convertible<Arithmetic, Backend>::value && std::is_integral<Arithmetic>::value>::type
+   eval_divide(rational_adaptor<Backend>& result, const rational_adaptor<Backend>& a, Arithmetic arg)
+{
+   if (&result == &a)
+      return eval_divide(result, arg);
+   if (arg == 0)
+   {
+      BOOST_MP_THROW_EXCEPTION(std::overflow_error("Integer division by zero"));
+      return;
+   }
+   else if (arg == 1)
+   {
+      result = a;
+      return;
+   }
+   else if (is_minus_one(arg))
+   {
+      result = a;
+      result.num().negate();
+      return;
+   }
+
+   if (eval_get_sign(a) == 0)
+   {
+      result = a;
+      return;
+   }
+
+   using default_ops::eval_multiply;
+   using default_ops::eval_divide;
+   using default_ops::eval_gcd;
+   using default_ops::eval_convert_to;
+
+   Backend gcd;
+   Arithmetic integer_gcd;
+   eval_gcd(gcd, a.num(), arg);
+   eval_convert_to(&integer_gcd, gcd);
+   arg /= integer_gcd;
+   eval_multiply(result.denom(), a.denom(), boost::multiprecision::detail::unsigned_abs(arg));
+
+   if (integer_gcd > 1)
+   {
+      eval_divide(result.num(), a.num(), integer_gcd);
+   }
+   else
+      result.num() = a.num();
+   if (arg < 0)
+   {
+      result.num().negate();
+   }
+}
+
+//
+// Increment and decrement:
+//
+template <class Backend> 
+inline void eval_increment(rational_adaptor<Backend>& arg)
+{
+   using default_ops::eval_add;
+   eval_add(arg.num(), arg.denom());
+}
+template <class Backend> 
+inline void eval_decrement(rational_adaptor<Backend>& arg)
+{
+   using default_ops::eval_subtract;
+   eval_subtract(arg.num(), arg.denom());
+}
+
+//
+// abs:
+//
+template <class Backend> 
+inline void eval_abs(rational_adaptor<Backend>& result, const rational_adaptor<Backend>& arg)
+{
+   using default_ops::eval_abs;
+   eval_abs(result.num(), arg.num());
+   result.denom() = arg.denom();
+}
+
+} // namespace backends
+
+//
+// Define a category for this number type, one of:
+// 
+//    number_kind_integer
+//    number_kind_floating_point
+//    number_kind_rational
+//    number_kind_fixed_point
+//    number_kind_complex
+//
+template<class Backend>
+struct number_category<rational_adaptor<Backend> > : public std::integral_constant<int, number_kind_rational>
+{};
+
+template <class Backend, expression_template_option ExpressionTemplates>
+struct component_type<number<rational_adaptor<Backend>, ExpressionTemplates> >
+{
+   typedef number<Backend, ExpressionTemplates> type;
+};
+
+template <class IntBackend, expression_template_option ET>
+inline number<IntBackend, ET> numerator(const number<rational_adaptor<IntBackend>, ET>& val)
+{
+   return val.backend().num();
+}
+template <class IntBackend, expression_template_option ET>
+inline number<IntBackend, ET> denominator(const number<rational_adaptor<IntBackend>, ET>& val)
+{
+   return val.backend().denom();
+}
+
+template <class Backend>
+struct is_unsigned_number<rational_adaptor<Backend> > : public is_unsigned_number<Backend>
+{};
+
+
+}} // namespace boost::multiprecision
+
+namespace std {
+
+   template <class IntBackend, boost::multiprecision::expression_template_option ExpressionTemplates>
+   class numeric_limits<boost::multiprecision::number<boost::multiprecision::rational_adaptor<IntBackend>, ExpressionTemplates> > : public std::numeric_limits<boost::multiprecision::number<IntBackend, ExpressionTemplates> >
+   {
+      using base_type = std::numeric_limits<boost::multiprecision::number<IntBackend> >;
+      using number_type = boost::multiprecision::number<boost::multiprecision::rational_adaptor<IntBackend> >;
+
+   public:
+      static constexpr bool is_integer = false;
+      static constexpr bool is_exact = true;
+      static constexpr      number_type(min)() { return (base_type::min)(); }
+      static constexpr      number_type(max)() { return (base_type::max)(); }
+      static constexpr number_type lowest() { return -(max)(); }
+      static constexpr number_type epsilon() { return base_type::epsilon(); }
+      static constexpr number_type round_error() { return epsilon() / 2; }
+      static constexpr number_type infinity() { return base_type::infinity(); }
+      static constexpr number_type quiet_NaN() { return base_type::quiet_NaN(); }
+      static constexpr number_type signaling_NaN() { return base_type::signaling_NaN(); }
+      static constexpr number_type denorm_min() { return base_type::denorm_min(); }
+   };
+
+   template <class IntBackend, boost::multiprecision::expression_template_option ExpressionTemplates>
+   constexpr bool numeric_limits<boost::multiprecision::number<boost::multiprecision::rational_adaptor<IntBackend>, ExpressionTemplates> >::is_integer;
+   template <class IntBackend, boost::multiprecision::expression_template_option ExpressionTemplates>
+   constexpr bool numeric_limits<boost::multiprecision::number<boost::multiprecision::rational_adaptor<IntBackend>, ExpressionTemplates> >::is_exact;
+
+} // namespace std
+
+#endif