[pypy-dev] Proof that pypy is feasible

[Armin Rigo]

Hello Edward,

On Thu, Apr 08, 2004 at 08:21:22AM -0500, Edward K. Ream wrote:
> "Rumors have it that the secret goal is being faster-than-C which is
> nonsense, isn't it?"  My thinking about this has changed since then as the
> result of reading the following two excellent HP tech reports:
> 
> http://www.hpl.hp.com/techreports/1999/HPL-1999-77.pdf
> http://www.hpl.hp.com/techreports/1999/HPL-1999-78.pdf

The Dynamo project.  It is excellent because it shows that you can optimize in
software at run-time a program that has been so-called super-optimized by a
compiler, thus essentially showing that you can beat decades of efforts in
static compilation with a run-time tool.

For me it is a proof-of-concept, but I'm even more interested in the next
logical step: if the compiler produced a "pseudo-binary" or something with
explicit support for this later phase, e.g. pseudo-code with optimization
hints instead of machine code, it would probably be even better.  This is the
direction in which I see this going.

I wouldn't like it too much if the hardware directly supported this kind of
technique, because it has a lot to gain from more extreme flexibility than
hardware can provide.  In other words, making a processor even more complex
than they already are looks like a bad idea.  In my opinions processors should
be extremely dumb (no memory management, for example).

An interesting middle ground, however, are processors with a built-in software
layer over their core, like the Code Morphing one you mentioned which
interprets x86 code over a different, simpler core.  I like these approaches
because I think that ultimately the software interpretation layer will be
moved out of the processor and made part of the OS.


A bientôt,

Armin.