C++ runtime and hyperobjects

[VoxSciurorum]

Convert runtime to C++. Create two new reducer kinds for C++ code.

This change must be paired with the corresponding compiler change.

[neboat]

As far as I can tell, hash is always initialized to 0 in a new bucket. Why not specify that default initialization in the definition of bucket?

[VoxSciurorum]

The default value in the struct definition is 0. I think I didn't want to mix aggregate and constructor style initialization.

[neboat]

As far as I can tell, the particular mixing of aggregate and default initialization that would happen here, if this line didn't initialize .hash explicitly, has well-defined behavior as of C++11.

I think it's actually cleaner to rely on the default initialization of .hash in places like this, rather than explicitly initialize it. In my mind, the default initialization of .hash encodes the correct value of .hash for any new bucket, which is dictated by the hypertable structure and not the call to the registration/lookup function itself. Letting the default initialization handle .hash allows for less code duplication and less error-prone code.

[VoxSciurorum]

I removed the explicit initialization.

[neboat]

The other reducer_register functions all set view to the key. Why is this one different?

[VoxSciurorum]

In the 0-argument variant lookup calls return extra as a reducer_base * and ignore view. In the 1- and 2-argument variants lookup calls return view and ignore extra. This difference is to allow reducer_base to be a virtual base class of the view type.

[neboat]

Should the size() function be const? Is there a use case where this needs to not be const? If not, then it seems better to make this const and prevent worries about needing to synchronize concurrent calls to this function.

[VoxSciurorum]

In C++ a const method is still allowed to modify state. If you want size() to be const for documentation purposes ("we prefer that the method not have side effects") I can make it const. Only the runtime should be calling size().