I was referring to this essay:

http://www.dreamsongs.com/WIB.html

What I'm trying to get at, to make myself clear, is that in the CLR
model, opcodes are compiled by the JIT compiler into machine code; yet
your instructions manipulate opcodes - that is, they manipulate the
source code of the JIT compiler. (This is what I was referring to with
the Lisp reference.)

How would it allocate registers? How are arguments passed in? How does
it compose, yet ensure that the parts composed don't clobber each
other's registers and locals?

Efficient allocation of registers is probably the most important thing a
compiler does that relies on seeing as much of the source structure as
possible before doing stuff, and is also one of the more important
issues in efficient compilation.

My point being though, that it will have to enter the compiler every
time, rather than just once, if the operations created through
composition are to be efficient. The alternative is to have a fast
interpreter, maybe using your self-described naive scheme where simple
cookie-cutter machine code blocks get pasted together.

I will say that I know much thought has gone into this kind of thing in
the Lisp world, and I'm not deeply familiar with optimization techniques
that have been applied there, so I'm actually unqualified to object to
it. But I brought it up anyway, since no one else is reacting :)

I was musing out loud about possibilities for efficient compilation. I
can't get specific without trying to design something that would work,
which would take too much time out of my day, unfortunately. Though, see
the thoughts below...

Yes, I would say that interpreting would typically be faster than an
efficient compiler for less-complex inputs, i.e. dependent on a low 'n'.
But when I think about it now, it's like layering e.g. the Hotspot JIT
inside itself; that is, the JIT generates code that interprets yet falls
over to (re)compilation if enough reps are done...

You'll have to forgive me, I'm a little slow for new ideas :)

Well now, technically, a Lisp compiler could evaluate that at compile
time :P

The main limiting factors on the speed of typical CLR & C#
implementation of above snippet are the cost of calling through a
delegate, and the opaqueness of a delegate to optimization - that is,
when the compiler is processing the definitions of Map and Fold, it
can't use information about the actual arguments to optimize as much as
it could if things were coded as a loop. That isn't to say that Map &
Fold aka Reduce don't have interesting optimization strategies in their
own right, the point is that the extra indirection of delegates is a
black box for the compiler.

I'm not sure what you're suggesting will improve that specifically
though, it's more oriented towards dynamic composition of new functions,
as I see it - and that case certainly isn't served well by the current
.NET delegate paradigm, because every composition accumulates delegate
invocation costs. Let me make that concrete:

 bool P(bool value);
 
 P Not(P pred)
   { return delegate(bool x) { return !pred(x); }; }
 P And(P left, T right)
   { return delegate(bool x) { return left(x) && right(x); }; }
 // ...

If someone wants to compose a predicate dynamically in C#/.NET,
currently they accumulate these costs. They can mitigate these costs
with the new stuff coming in C# 3.0 though, with LINQ Func<> trees,
which can be passed to a compiler and get the benefit of full
compilation. However, that forces the user to make the interpretation /
compilation tradeoff themselves.

OK. Now that I've thought about it more, it looks interesting, and with
the nesting of JIT compilation, it could be made efficient in both the
seldom-called and repeatedly-called cases. Have you considered modifying
e.g. Mono or one of the other open-source VM implementations to try and
get something along these lines? Because it seems to me that this kind
of thing could indeed make the cases above better, e.g. lambda
constructions could be improved.

I'm curious: can the compose operation itself be composed? :) That would
be something which would be particularly painful to handle in the LINQ
sense, although applications for it don't exactly jump out at me.

-- Barry