[rwar]

The "simplifications" in the program that can be achieved by a supercompiler

Redundant code elimination. Performing operations on data that are known at the time of supercompilation. Removing intermediate data structures. Transforming multi-pass algorithms to one-pass algorithms.

http://sites.google.com/site/keldyshscp/Home/supercompilerco...

[gte910h]

So basically, complicated optimizations that will often be opaque to a human (not trying to be adversarial, just understand it)?

[scott_s]

Quite the opposite, if I understand the concept. It's more trying to apply the kinds of optimizations that a human would make when reasoning about how a program actually runs. When we read code, we often think "Oh, I get it - it's just trying to do X." A good supercompiler would just emit the code to do X.

I can see this coming up when you, say, make library calls in certain ways that don't need the full functionality of the algorithms. Maybe one paramater is fixed, or the sequence of the calls is really to achieve something else.

The original paper is freely available off of citeseer, "The Concept of a Supercompiler": http://citeseer.ist.psu.edu/viewdoc/download;jsessionid=B373...

I'm currently reading through it. I find it useful to forget the term "supercompiler," since that implies to me a "meta-compiler," which I don't think is an accurate way of describing the technique. This could be incorporated into a regular compiler, you'd just want this as the first optimization phase. (I think.)

[gte910h]

Thanks for going through it; I'd honestly like a less researchery article than that one to read. That wasn't my field when I was a researcher, and I'm not particularly interested in wading through articles written for a academic audience which may or may not allow me to draw the correct conclusions from trying to read through it.

"Yet another optimizer phase" is what I'm readying here.

[scott_s]

Except I've only made it through the introduction, and I'm still not clear on whether or not their technique implies a compile, execute, profile, re-compile cycle.

One major difficulty with academic articles is that they have to "sell" their contribution in order to get the paper accepted. This entails taking the approach to its logical extension, and talking up the impact it can potentially have. Otherwise, reviewers may reject it for not being novel enough. But, this sort of approach is often detrimental to understand what, exactly, the researchers actually did. I know that when I read a research paper, there's often an "aha" moment when I realize "Oh, they just took XYZ and did ABC to it" which can either be followed up with "that's lame" or "that's neat!"

[crux_]

I'm pretty certain its role is that of a transformer from programs in an intermediate language to optimized programs in that same intermediate language. Hence, a pipeline internal to a compiler.

The use of the word "traces" may be the source of your confusion in light the rise of tracing JITs (there that actually execute a program). I think a better word would be "abstract interpretation".

[scott_s]

These are the passages from the intro that made me say that:

"So a supercompiler does not transform the program by steps; it controls and observes (SUPERvises) the running of the machine that is represented by the program; let us call this machine M,. In observing the operation of M1, the supercompiler COMPILES a program which describes the activities of M1, but it makes shortcuts and whatever clever tricks it knows in order to produce the same effect as M1, but faster."

"A supercompiler would run M, in a general form, with unknown values of variables, and create a graph of states and transitions between possible configurations of the computing system. However, this process (called driving) can usually go on infinitely."

This implies, to me, a compile, execute, profile, re-compile cycle. But, that conflicts with the understanding of the technique that I described above. If that does not describe a compile, execute, profile, re-compile cycle, then I'd like to know both what, exactly, the author meant by the above, and why he chose to phrase it as such.

[crux_]

I don't think it has much to do with human-readability; but rather with:

- automatically finding and applying optimizations

- with a general approach, rather than a big database of hard-coded optimizations.

I'm not sure how (or how well) it all works, though... but it's the sort of thing I find interesting, so here's hoping for some free time sooner or later. ;)

[gte910h]

If it's "yet another thing I run after I am done writing code" that's just a optimizer IMO, not a different sort of thing.

If it actually changes the code I have written, and then leaves it in that state for some reason 1> I want to know about that and 2> I want to know about if the trade offs of using it are worth it for the payments we make via reduced readability and malleability.

[crux_]

It's a component for your compiler, not something that would likely exist in a standalone form except for demonstration & research.

(For the purposes of above comment, JIT interpreters absolutely count as compilers.)

And yes, it's 'just' an optimizer. It so happens to be a fairly interesting approach.

[thesz]

I think you're right. This is a method that produces optimizations from evaluation rewrite rules.

[derleth]

> Performing operations on data that are known at the time of supercompilation.

Isn't this known as partial application? For example, I have a function of five arguments, and I know three of them, partial application optimizes the function based on the three knowns while retaining full generality for the two unknowns.

[thesz]

This is just one of transformations performed by supercompilation.

I recently discovered distillation [1], which could transform quadratic and even more complex programs into linear ones (ie, perform non-linear complexity reduction).

So there's more in supercompilation that partial evaluation.

[1]: http://meta2010.pereslavl.ru/accepted-papers/paper-info-2.ht...

[dkersten]

Distillation is cool! Thanks for the link.

Geoff Hamilton lectured my classes in computability, parsing, automata and compilers. He certainly knows his stuff.