Re: Ideas for Ada 200X

[Hyman Rosen <hyrosen@mail.com>]

Russ wrote:
> Interesting. Where can I learn more about this?

Here's a general explanation of how expression templates
work in C++. The language features used are the ability
to overload the array indexing and function call operators
and automatic type deduction and instantiation of function
templates.

We declare our matrix types, and whatever other types we
need. We define the arithmetic and indexing operators on
these types not to actually do anything except note the
operator and references to the operands, in effect making
a runtime parse tree of the expression. Finally, when an
actual value is required, suc has upon assignment, the
entire structure can be evaluated, with the compiler doing
massive inlining to achieve the efficiency of handwritten
code.

For example, suppose we have the matrix expression C = A + B.
When A + B is evaluated, the result is a special "addition"
object which holds references to A and B. This object has an
overloaded indexing operator, such that when add_obj(i,j) is
requested, it produces a(i,j) + b(i,j), where a and b are the
saved operand references. The Matrix class has an assignment
operator which simply (or complexly, sometimes) iterates,
doing c(i,j) = rhs(i,j). This works for expressions of any
complexity, as long as the compiler is willing to inline. The
iteration pattern itself can be modified to take advantage of
cache coherency; the Blitz++ library includes the ability to
traverse multidimensional matrices using a space-filling curve
to stay local.