On Thu, 23 Mar 2000, Jeffrey B. Siegal wrote:
> Here's an interesting bit on /. (links to ArsTechnica) about how HPs
> dynamic-recompiler optimizes code at run-time. If such a thing were
> built into a virtualizer, virtual machines could potentially run
> *faster* than the host!
This has been commented on slashdot, but I think it best to restate it
here.
Basically, what Dynamo does is compile/interpret/nativize/whip/bludgeon
code with runtime information. Dynamo can have performance increases
because it has this runtime information; really where this could shine is
in cache coloring of code segments. But I digress.
A compiler that can take in a trace and use run-time information to
compile the program can do better than one that doesn't. The advantage of
Dynamo is that the trace is uses is the trace you're producing.
I think it's also important to point out that HP's compilers really aren't
that good. Improving them isn't a tremendous feat. Not a lot of
information is given on the performance increases: are they with a
rewritten binary, or do they include the overhead for rewriting?
Binary rewriting has been around for some time, as JIT compilers show. If
it were such a powerful technique, there would have been a stir before. I
think they're hiding some details.
Phil
