#pragma once #include #include #include #include #include #include #include #include #include #include #include #include namespace DB { namespace ErrorCodes { extern const int BAD_TYPE_OF_FIELD; extern const int BAD_GET; extern const int NOT_IMPLEMENTED; extern const int LOGICAL_ERROR; extern const int ILLEGAL_TYPE_OF_ARGUMENT; } template struct NearestFieldTypeImpl; template using NearestFieldType = typename NearestFieldTypeImpl::Type; class Field; using FieldVector = std::vector; /// Array and Tuple use the same storage type -- FieldVector, but we declare /// distinct types for them, so that the caller can choose whether it wants to /// construct a Field of Array or a Tuple type. An alternative approach would be /// to construct both of these types from FieldVector, and have the caller /// specify the desired Field type explicitly. #define DEFINE_FIELD_VECTOR(X) \ struct X : public FieldVector \ { \ using FieldVector::FieldVector; \ } DEFINE_FIELD_VECTOR(Array); DEFINE_FIELD_VECTOR(Tuple); #undef DEFINE_FIELD_VECTOR struct AggregateFunctionStateData { String name; /// Name with arguments. String data; bool operator < (const AggregateFunctionStateData &) const { throw Exception("Operator < is not implemented for AggregateFunctionStateData.", ErrorCodes::ILLEGAL_TYPE_OF_ARGUMENT); } bool operator <= (const AggregateFunctionStateData &) const { throw Exception("Operator <= is not implemented for AggregateFunctionStateData.", ErrorCodes::ILLEGAL_TYPE_OF_ARGUMENT); } bool operator > (const AggregateFunctionStateData &) const { throw Exception("Operator > is not implemented for AggregateFunctionStateData.", ErrorCodes::ILLEGAL_TYPE_OF_ARGUMENT); } bool operator >= (const AggregateFunctionStateData &) const { throw Exception("Operator >= is not implemented for AggregateFunctionStateData.", ErrorCodes::ILLEGAL_TYPE_OF_ARGUMENT); } bool operator == (const AggregateFunctionStateData & rhs) const { if (name != rhs.name) throw Exception("Comparing aggregate functions with different types: " + name + " and " + rhs.name, ErrorCodes::ILLEGAL_TYPE_OF_ARGUMENT); return data == rhs.data; } }; template bool decimalEqual(T x, T y, UInt32 x_scale, UInt32 y_scale); template bool decimalLess(T x, T y, UInt32 x_scale, UInt32 y_scale); template bool decimalLessOrEqual(T x, T y, UInt32 x_scale, UInt32 y_scale); template class DecimalField { public: DecimalField(T value, UInt32 scale_) : dec(value), scale(scale_) {} operator T() const { return dec; } T getValue() const { return dec; } T getScaleMultiplier() const { return T::getScaleMultiplier(scale); } UInt32 getScale() const { return scale; } template bool operator < (const DecimalField & r) const { using MaxType = std::conditional_t<(sizeof(T) > sizeof(U)), T, U>; return decimalLess(dec, r.getValue(), scale, r.getScale()); } template bool operator <= (const DecimalField & r) const { using MaxType = std::conditional_t<(sizeof(T) > sizeof(U)), T, U>; return decimalLessOrEqual(dec, r.getValue(), scale, r.getScale()); } template bool operator == (const DecimalField & r) const { using MaxType = std::conditional_t<(sizeof(T) > sizeof(U)), T, U>; return decimalEqual(dec, r.getValue(), scale, r.getScale()); } template bool operator > (const DecimalField & r) const { return r < *this; } template bool operator >= (const DecimalField & r) const { return r <= * this; } template bool operator != (const DecimalField & r) const { return !(*this == r); } const DecimalField & operator += (const DecimalField & r) { if (scale != r.getScale()) throw Exception("Add different decimal fields", ErrorCodes::LOGICAL_ERROR); dec += r.getValue(); return *this; } const DecimalField & operator -= (const DecimalField & r) { if (scale != r.getScale()) throw Exception("Sub different decimal fields", ErrorCodes::LOGICAL_ERROR); dec -= r.getValue(); return *this; } private: T dec; UInt32 scale; }; /// char may be signed or unsigned, and behave identically to signed char or unsigned char, /// but they are always three different types. /// signedness of char is different in Linux on x86 and Linux on ARM. template <> struct NearestFieldTypeImpl { using Type = std::conditional_t, Int64, UInt64>; }; template <> struct NearestFieldTypeImpl { using Type = Int64; }; template <> struct NearestFieldTypeImpl { using Type = UInt64; }; template <> struct NearestFieldTypeImpl { using Type = UInt64; }; template <> struct NearestFieldTypeImpl { using Type = UInt64; }; template <> struct NearestFieldTypeImpl { using Type = UInt64; }; template <> struct NearestFieldTypeImpl { using Type = UInt128; }; template <> struct NearestFieldTypeImpl { using Type = UInt128; }; template <> struct NearestFieldTypeImpl { using Type = Int64; }; template <> struct NearestFieldTypeImpl { using Type = Int64; }; /// long and long long are always different types that may behave identically or not. /// This is different on Linux and Mac. template <> struct NearestFieldTypeImpl { using Type = Int64; }; template <> struct NearestFieldTypeImpl { using Type = Int64; }; template <> struct NearestFieldTypeImpl { using Type = UInt64; }; template <> struct NearestFieldTypeImpl { using Type = UInt64; }; template <> struct NearestFieldTypeImpl { using Type = Int128; }; template <> struct NearestFieldTypeImpl { using Type = DecimalField; }; template <> struct NearestFieldTypeImpl { using Type = DecimalField; }; template <> struct NearestFieldTypeImpl { using Type = DecimalField; }; template <> struct NearestFieldTypeImpl> { using Type = DecimalField; }; template <> struct NearestFieldTypeImpl> { using Type = DecimalField; }; template <> struct NearestFieldTypeImpl> { using Type = DecimalField; }; template <> struct NearestFieldTypeImpl { using Type = Float64; }; template <> struct NearestFieldTypeImpl { using Type = Float64; }; template <> struct NearestFieldTypeImpl { using Type = String; }; template <> struct NearestFieldTypeImpl { using Type = String; }; template <> struct NearestFieldTypeImpl { using Type = Array; }; template <> struct NearestFieldTypeImpl { using Type = Tuple; }; template <> struct NearestFieldTypeImpl { using Type = UInt64; }; template <> struct NearestFieldTypeImpl { using Type = Null; }; template <> struct NearestFieldTypeImpl { using Type = AggregateFunctionStateData; }; // For enum types, use the field type that corresponds to their underlying type. template struct NearestFieldTypeImpl>> { using Type = NearestFieldType>; }; /** 32 is enough. Round number is used for alignment and for better arithmetic inside std::vector. * NOTE: Actually, sizeof(std::string) is 32 when using libc++, so Field is 40 bytes. */ #define DBMS_MIN_FIELD_SIZE 32 /** Discriminated union of several types. * Made for replacement of `boost::variant` * is not generalized, * but somewhat more efficient, and simpler. * * Used to represent a single value of one of several types in memory. * Warning! Prefer to use chunks of columns instead of single values. See Column.h */ class Field { public: struct Types { /// Type tag. enum Which { Null = 0, UInt64 = 1, Int64 = 2, Float64 = 3, UInt128 = 4, Int128 = 5, /// Non-POD types. String = 16, Array = 17, Tuple = 18, Decimal32 = 19, Decimal64 = 20, Decimal128 = 21, AggregateFunctionState = 22, }; static const int MIN_NON_POD = 16; static const char * toString(Which which) { switch (which) { case Null: return "Null"; case UInt64: return "UInt64"; case UInt128: return "UInt128"; case Int64: return "Int64"; case Int128: return "Int128"; case Float64: return "Float64"; case String: return "String"; case Array: return "Array"; case Tuple: return "Tuple"; case Decimal32: return "Decimal32"; case Decimal64: return "Decimal64"; case Decimal128: return "Decimal128"; case AggregateFunctionState: return "AggregateFunctionState"; } throw Exception("Bad type of Field", ErrorCodes::BAD_TYPE_OF_FIELD); } }; /// Returns an identifier for the type or vice versa. template struct TypeToEnum; template struct EnumToType; static bool IsDecimal(Types::Which which) { return which >= Types::Decimal32 && which <= Types::Decimal128; } Field() : which(Types::Null) { } /** Despite the presence of a template constructor, this constructor is still needed, * since, in its absence, the compiler will still generate the default constructor. */ Field(const Field & rhs) { create(rhs); } Field(Field && rhs) { create(std::move(rhs)); } template Field(T && rhs, std::enable_if_t, Field>, void *> = nullptr); /// Create a string inplace. Field(const char * data, size_t size) { create(data, size); } Field(const unsigned char * data, size_t size) { create(data, size); } /// NOTE In case when field already has string type, more direct assign is possible. void assignString(const char * data, size_t size) { destroy(); create(data, size); } void assignString(const unsigned char * data, size_t size) { destroy(); create(data, size); } Field & operator= (const Field & rhs) { if (this != &rhs) { if (which != rhs.which) { destroy(); create(rhs); } else assign(rhs); /// This assigns string or vector without deallocation of existing buffer. } return *this; } Field & operator= (Field && rhs) { if (this != &rhs) { if (which != rhs.which) { destroy(); create(std::move(rhs)); } else assign(std::move(rhs)); } return *this; } template std::enable_if_t, Field>, Field &> operator= (T && rhs); ~Field() { destroy(); } Types::Which getType() const { return which; } const char * getTypeName() const { return Types::toString(which); } bool isNull() const { return which == Types::Null; } template T & get(); template const T & get() const { auto mutable_this = const_cast *>(this); return mutable_this->get(); } template T & reinterpret(); template const T & reinterpret() const { auto mutable_this = const_cast *>(this); return mutable_this->reinterpret(); } template bool tryGet(T & result) { const Types::Which requested = TypeToEnum>::value; if (which != requested) return false; result = get(); return true; } template bool tryGet(T & result) const { const Types::Which requested = TypeToEnum>::value; if (which != requested) return false; result = get(); return true; } template T & safeGet() { const Types::Which requested = TypeToEnum>::value; if (which != requested) throw Exception("Bad get: has " + std::string(getTypeName()) + ", requested " + std::string(Types::toString(requested)), ErrorCodes::BAD_GET); return get(); } template const T & safeGet() const { const Types::Which requested = TypeToEnum>::value; if (which != requested) throw Exception("Bad get: has " + std::string(getTypeName()) + ", requested " + std::string(Types::toString(requested)), ErrorCodes::BAD_GET); return get(); } bool operator< (const Field & rhs) const { if (which < rhs.which) return true; if (which > rhs.which) return false; switch (which) { case Types::Null: return false; case Types::UInt64: return get() < rhs.get(); case Types::UInt128: return get() < rhs.get(); case Types::Int64: return get() < rhs.get(); case Types::Int128: return get() < rhs.get(); case Types::Float64: return get() < rhs.get(); case Types::String: return get() < rhs.get(); case Types::Array: return get() < rhs.get(); case Types::Tuple: return get() < rhs.get(); case Types::Decimal32: return get>() < rhs.get>(); case Types::Decimal64: return get>() < rhs.get>(); case Types::Decimal128: return get>() < rhs.get>(); case Types::AggregateFunctionState: return get() < rhs.get(); } throw Exception("Bad type of Field", ErrorCodes::BAD_TYPE_OF_FIELD); } bool operator> (const Field & rhs) const { return rhs < *this; } bool operator<= (const Field & rhs) const { if (which < rhs.which) return true; if (which > rhs.which) return false; switch (which) { case Types::Null: return true; case Types::UInt64: return get() <= rhs.get(); case Types::UInt128: return get() <= rhs.get(); case Types::Int64: return get() <= rhs.get(); case Types::Int128: return get() <= rhs.get(); case Types::Float64: return get() <= rhs.get(); case Types::String: return get() <= rhs.get(); case Types::Array: return get() <= rhs.get(); case Types::Tuple: return get() <= rhs.get(); case Types::Decimal32: return get>() <= rhs.get>(); case Types::Decimal64: return get>() <= rhs.get>(); case Types::Decimal128: return get>() <= rhs.get>(); case Types::AggregateFunctionState: return get() <= rhs.get(); } throw Exception("Bad type of Field", ErrorCodes::BAD_TYPE_OF_FIELD); } bool operator>= (const Field & rhs) const { return rhs <= *this; } // More like bitwise equality as opposed to semantic equality: // Null equals Null and NaN equals NaN. bool operator== (const Field & rhs) const { if (which != rhs.which) return false; switch (which) { case Types::Null: return true; case Types::UInt64: return get() == rhs.get(); case Types::Int64: return get() == rhs.get(); case Types::Float64: { // Compare as UInt64 so that NaNs compare as equal. return reinterpret() == rhs.reinterpret(); } case Types::String: return get() == rhs.get(); case Types::Array: return get() == rhs.get(); case Types::Tuple: return get() == rhs.get(); case Types::UInt128: return get() == rhs.get(); case Types::Int128: return get() == rhs.get(); case Types::Decimal32: return get>() == rhs.get>(); case Types::Decimal64: return get>() == rhs.get>(); case Types::Decimal128: return get>() == rhs.get>(); case Types::AggregateFunctionState: return get() == rhs.get(); } throw Exception("Bad type of Field", ErrorCodes::BAD_TYPE_OF_FIELD); } bool operator!= (const Field & rhs) const { return !(*this == rhs); } /// Field is template parameter, to allow universal reference for field, /// that is useful for const and non-const . template static auto dispatch(F && f, FieldRef && field) { switch (field.which) { case Types::Null: return f(field.template get()); case Types::UInt64: return f(field.template get()); case Types::UInt128: return f(field.template get()); case Types::Int64: return f(field.template get()); case Types::Float64: return f(field.template get()); case Types::String: return f(field.template get()); case Types::Array: return f(field.template get()); case Types::Tuple: return f(field.template get()); case Types::Decimal32: return f(field.template get>()); case Types::Decimal64: return f(field.template get>()); case Types::Decimal128: return f(field.template get>()); case Types::AggregateFunctionState: return f(field.template get()); case Types::Int128: // TODO: investigate where we need Int128 Fields. There are no // field visitors that support them, and they only arise indirectly // in some functions that use Decimal columns: they get the // underlying Field value with get(). Probably should be // switched to DecimalField, but this is a whole endeavor in itself. throw Exception("Unexpected Int128 in Field::dispatch()", ErrorCodes::LOGICAL_ERROR); } // GCC 9 complains that control reaches the end, despite that we handle // all the cases above (maybe because of throw?). Return something to // silence it. Null null{}; return f(null); } private: std::aligned_union_t, DecimalField, DecimalField, AggregateFunctionStateData > storage; Types::Which which; /// Assuming there was no allocated state or it was deallocated (see destroy). template void createConcrete(T && x) { using UnqualifiedType = std::decay_t; // In both Field and PODArray, small types may be stored as wider types, // e.g. char is stored as UInt64. Field can return this extended value // with get(). To avoid uninitialized results from get(), // we must initialize the entire wide stored type, and not just the // nominal type. using StorageType = NearestFieldType; new (&storage) StorageType(std::forward(x)); which = TypeToEnum::value; } /// Assuming same types. template void assignConcrete(T && x) { using JustT = std::decay_t; assert(which == TypeToEnum::value); JustT * MAY_ALIAS ptr = reinterpret_cast(&storage); *ptr = std::forward(x); } void create(const Field & x) { dispatch([this] (auto & value) { createConcrete(value); }, x); } void create(Field && x) { dispatch([this] (auto & value) { createConcrete(std::move(value)); }, x); } void assign(const Field & x) { dispatch([this] (auto & value) { assignConcrete(value); }, x); } void assign(Field && x) { dispatch([this] (auto & value) { assignConcrete(std::move(value)); }, x); } void create(const char * data, size_t size) { new (&storage) String(data, size); which = Types::String; } void create(const unsigned char * data, size_t size) { create(reinterpret_cast(data), size); } ALWAYS_INLINE void destroy() { if (which < Types::MIN_NON_POD) return; switch (which) { case Types::String: destroy(); break; case Types::Array: destroy(); break; case Types::Tuple: destroy(); break; case Types::AggregateFunctionState: destroy(); break; default: break; } which = Types::Null; /// for exception safety in subsequent calls to destroy and create, when create fails. } template void destroy() { T * MAY_ALIAS ptr = reinterpret_cast(&storage); ptr->~T(); } }; #undef DBMS_MIN_FIELD_SIZE template <> struct Field::TypeToEnum { static const Types::Which value = Types::Null; }; template <> struct Field::TypeToEnum { static const Types::Which value = Types::UInt64; }; template <> struct Field::TypeToEnum { static const Types::Which value = Types::UInt128; }; template <> struct Field::TypeToEnum { static const Types::Which value = Types::Int64; }; template <> struct Field::TypeToEnum { static const Types::Which value = Types::Int128; }; template <> struct Field::TypeToEnum { static const Types::Which value = Types::Float64; }; template <> struct Field::TypeToEnum { static const Types::Which value = Types::String; }; template <> struct Field::TypeToEnum { static const Types::Which value = Types::Array; }; template <> struct Field::TypeToEnum { static const Types::Which value = Types::Tuple; }; template <> struct Field::TypeToEnum>{ static const Types::Which value = Types::Decimal32; }; template <> struct Field::TypeToEnum>{ static const Types::Which value = Types::Decimal64; }; template <> struct Field::TypeToEnum>{ static const Types::Which value = Types::Decimal128; }; template <> struct Field::TypeToEnum{ static const Types::Which value = Types::AggregateFunctionState; }; template <> struct Field::EnumToType { using Type = Null; }; template <> struct Field::EnumToType { using Type = UInt64; }; template <> struct Field::EnumToType { using Type = UInt128; }; template <> struct Field::EnumToType { using Type = Int64; }; template <> struct Field::EnumToType { using Type = Int128; }; template <> struct Field::EnumToType { using Type = Float64; }; template <> struct Field::EnumToType { using Type = String; }; template <> struct Field::EnumToType { using Type = Array; }; template <> struct Field::EnumToType { using Type = Tuple; }; template <> struct Field::EnumToType { using Type = DecimalField; }; template <> struct Field::EnumToType { using Type = DecimalField; }; template <> struct Field::EnumToType { using Type = DecimalField; }; template <> struct Field::EnumToType { using Type = DecimalField; }; template T & Field::get() { using ValueType = std::decay_t; //assert(TypeToEnum>::value == which); ValueType * MAY_ALIAS ptr = reinterpret_cast(&storage); return *ptr; } template T & Field::reinterpret() { using ValueType = std::decay_t; ValueType * MAY_ALIAS ptr = reinterpret_cast(&storage); return *ptr; } template T get(const Field & field) { return field.template get(); } template T get(Field & field) { return field.template get(); } template T safeGet(const Field & field) { return field.template safeGet(); } template T safeGet(Field & field) { return field.template safeGet(); } template <> struct TypeName { static std::string get() { return "Array"; } }; template <> struct TypeName { static std::string get() { return "Tuple"; } }; template <> struct TypeName { static std::string get() { return "AggregateFunctionState"; } }; template decltype(auto) castToNearestFieldType(T && x) { using U = NearestFieldType>; if constexpr (std::is_same_v, U>) return std::forward(x); else return U(x); } /// This (rather tricky) code is to avoid ambiguity in expressions like /// Field f = 1; /// instead of /// Field f = Int64(1); /// Things to note: /// 1. float <--> int needs explicit cast /// 2. customized types needs explicit cast template Field::Field(T && rhs, std::enable_if_t, Field>, void *>) { auto && val = castToNearestFieldType(std::forward(rhs)); createConcrete(std::forward(val)); } template std::enable_if_t, Field>, Field &> Field::operator= (T && rhs) { auto && val = castToNearestFieldType(std::forward(rhs)); using U = decltype(val); if (which != TypeToEnum>::value) { destroy(); createConcrete(std::forward(val)); } else assignConcrete(std::forward(val)); return *this; } class ReadBuffer; class WriteBuffer; /// It is assumed that all elements of the array have the same type. void readBinary(Array & x, ReadBuffer & buf); [[noreturn]] inline void readText(Array &, ReadBuffer &) { throw Exception("Cannot read Array.", ErrorCodes::NOT_IMPLEMENTED); } [[noreturn]] inline void readQuoted(Array &, ReadBuffer &) { throw Exception("Cannot read Array.", ErrorCodes::NOT_IMPLEMENTED); } /// It is assumed that all elements of the array have the same type. /// Also write size and type into buf. UInt64 and Int64 is written in variadic size form void writeBinary(const Array & x, WriteBuffer & buf); void writeText(const Array & x, WriteBuffer & buf); [[noreturn]] inline void writeQuoted(const Array &, WriteBuffer &) { throw Exception("Cannot write Array quoted.", ErrorCodes::NOT_IMPLEMENTED); } void readBinary(Tuple & x, ReadBuffer & buf); [[noreturn]] inline void readText(Tuple &, ReadBuffer &) { throw Exception("Cannot read Tuple.", ErrorCodes::NOT_IMPLEMENTED); } [[noreturn]] inline void readQuoted(Tuple &, ReadBuffer &) { throw Exception("Cannot read Tuple.", ErrorCodes::NOT_IMPLEMENTED); } void writeBinary(const Tuple & x, WriteBuffer & buf); void writeText(const Tuple & x, WriteBuffer & buf); void writeFieldText(const Field & x, WriteBuffer & buf); [[noreturn]] inline void writeQuoted(const Tuple &, WriteBuffer &) { throw Exception("Cannot write Tuple quoted.", ErrorCodes::NOT_IMPLEMENTED); } }