From: Andrei Alexandrescu (andrei@cs.washington.edu)
Date: Mon Nov 25 2002 - 17:37:31 PST
It would be nice, time permitting, to continue discussing how to allow the
programmer to express optimizations in source code, notably loop fusion with
user-defined operators. Here are a couple of references:
1. The original AOP paper discusses AOP within the context of optimizing
image transformation algos:
http://www.parc.xerox.com/csl/groups/sda/publications/papers/Kiczales-ECOOP9
7/for-web.pdf
2. A paper on expression templates for implementing operators:
http://osl.iu.edu/~tveldhui/papers/Expression-Templates/exprtmpl.html
The page has references to a couple of libraries that boast FORTRAN-like
performance with regular infix notation for vector and matrix operations. In
essence, what happens is that each operator returns a type that encodes the
semantics of that operator. The object's state contains references to the
operands. For example:
Matrix A, B, C, D;
A = B + C + D;
Here, B + C + D creates an object of type Plus<Matrix, Plus<Matrix, Matrix>
>. So as you see, the AST is in the type. (The type generated by more
complex expressions will be appropriatedly hairy.) In addition, each Plus
object stores two references to its operands. During the assignment, one
iteration is done calling the indexing operator on Plus. That operator
evaluates the indexing operator on its two operands. This compile-time
recursive invocation of indexing yields efficient code.
So my question is: what language mechanism would allow the programmer to
express such (and related) transformations in a more natural manner?
It seems indeed (see http://www.xpdeveloper.com/cgi-bin/wiki.cgi?XPmeetsAOP)
that AspectJ indeed cannot do even simple loop fusion. This comes as an
embarrassing limitation of AspectJ, especially considering that
user-controlled optimizations are the motivation for AOP and one of its main
applications.
Andrei
_______________________________________________
Cecil mailing list
Cecil@cs.washington.edu
http://mailman.cs.washington.edu/mailman/listinfo/cecil
This archive was generated by hypermail 2.1.5 : Mon Nov 25 2002 - 17:41:25 PST