Next steps for `as(X, type)`

[perrydv]

Issue #185 opened discussion on as, a DSL feature to provide some run-time type flexibility.

PR #195 adds this feature, with updates to what it covers and not.

The best way to see what it currently covers is to work through specificOp_tests/test-as.R.

Broadly:

• as(X, type) treats X as type, where type is a basic type such as numericVector.
• This works on the LHS or RHS and should work with any other Eigen operations.
• When X has a known type, implementation is roughly:
  • with same-type scalars (X and type involve the same scalar element type)
    • On RHS or LHS
      • with no indexing after the as(), return an Eigen::TensorMap (e.g. to view a matrix as a vector).
      • with indexing after the as(), return a StridedTensorMap (an Eigen extension in nCompiler)
  • with cross-type scalars
    • on LHS
      • return a "CastingProxy" object that holds an STM for a Proxy copy that is updated on exit (destruction)
    • on RHS
      • return a "RHSCastProxy" object that holds a TM or STM (if non-indexed or indexed, respectively) and applies a static_cast when accessing elements.

These are all determined from ETaccessorTyped::asTyped, called from nC_as.

Because some cases need a Proxy class to hold a copy of X, all cases are implemented with a proxy layer so that there is a coding standard where an extra () is needed to access the Eigen-compatible object. e.g. nC_as<type>(X)(). (Due to delicacy of Eigen chains going out of scope, the extra () must be local.)

Future steps or to-dos for discussion include:

• Support true scalar types as inputs or target types.
• Allow as(x, 'integer') to maintain whatever nDim x already has.
• Allow as(x, 'vector') to maintain whatever scalar type x already has
• Use for argument passing, e.g. to pass one type for another by a copy, ref, or blockRef.
• When as is on the LHS, we are not also getting the benefit of flex_() assignment. Check on that.
• LHS proxies (also without indexing) might sometimes not need to make the copy upon creation but rather just allocate memory to be assigned into.