[racket] Optimizations at the library level.

From: Neil Van Dyke (neil at neilvandyke.org)
Date: Wed Sep 1 05:01:19 EDT 2010

Noel Welsh wrote at 08/31/2010 04:36 PM:
> In general this is a huge problem (equivalent, I think, to the "sufficiently smart compiler" problem, which itself is equivalent to solving AI).
>   

What Noel said.  Though, it might be a useful thing for the Racket 
interpreter (compiler) to have hooks of some kind for pluggable 
optimizations.  Useful both for practical use and for increasing how 
Racket can be used as a language research platform.

As a user, would be nice if I could get a matrix library without special 
compiler optimizations by doing:

(require (planet foo/matrix))

Or, if I indicated that I wanted the special compiler optimizations 
provided by the module, like:

(require/hooks (planet foo/matrix))

then special optimization code from the foo/matrix module can hook 
itself into a compiler phase for my using module.  In this example, 
perhaps in a phase after macro expansion and with all (current) 
references resolved.  (I'm thinking of hooks covering only intra-module 
stuff for the using module, no access into other modules.)

I don't know how feasible it would be to provide these hooks in a 
particular version of Racket.  Regarding keeping such hooks from 
encumbering evolution of future versions of Racket, would it suffice for 
the compiler to not be obligated to call the hooks and, if it does call, 
to announce its version when calling the hooks?  The hook code would be 
responsible for doing the right thing (or nothing) for that particular 
Racket version.  So, in this example, when a new version of Racket comes 
out, the foo/matrix would likely run without requested hook 
optimizations until the developer of foo/matrix looks at any changes 
made to the internal representations exposed by the hooks in that 
version of Racket, and makes any necessary changes to the 
optimizations.  Importantly, Racket would make no commitment to maintain 
the representations between versions, nor to even to document any of the 
representations.

It's fortunate that Matthew has seemingly infinite time for all the work 
that people would like him to do. :)

I still like the idea of doing the matrix optimizations in pure Racket 
with internally-optimizing ADTs with lazy operations, rather than with 
syntax extension mini-language, an entire front-end language, or 
compiler hooks.

-- 
http://www.neilvandyke.org/


Posted on the users mailing list.