Partially specialize simd for unscoped enumerations

[mattkretz]

Given simd<T, A> where T is an unscoped enum (U is a shorthand for std::underlying_type_t<T>):

• mask_type is simd<U, A>::mask_type
• implicit broadcast ctor from T
• explicit cvt ctor from simd<U, A> (enables explicit broadcast from U and int)
• P????: explicit cvt ctor from simd<T1, A1>, requires simd_size_v<T, A> == simd_size_v<T1, A1>
• generator ctor, where generator return type must be T
• explicit load ctor and member function from const T* and const U*
• store member function to T* and U*
• subscript access to elements of T
• unary operators !, +, and -, defined in such a way that program-defined non-member overloads of those operators take precedence. ! returns mask_type, + and - return simd<U, A>.
• implicit conversion operator to simd<U, A>
• binary operators +, -, *, /, %, &, |, ^, <<, >>, ==, !=, <, <=, >, and >=
  • all comparison operators return mask_type, the remaining binary operators return simd<U, A>
  • one of the operands has to be simd<T, A>, the other operand has to be convertible to simd<U, A> (which includes simd<T, A>)
  • program-defined non-member overloads of these operators take precedence
• additional overloads for << and >> where the type of the second operand is int
• operator?: overloads for bool and mask_type
  • the bool overload is short-circuiting and defines simd<U, A> as the common type for simd<T, A> and any type convertible to simd<U, A>
  • the mask_type overload blends two objects of type simd<T, A>.
  • program-defined non-member overloads of the mask_type conditional operator take precedence
• all constructors and member functions are constexpr except loads and stores
• not defined: (can be provided by program-defined overloads)
  • compound assignment
  • increment and decrement