Re: [Valgrind-developers] DynamoRIO

[Nicholas N. <nj...@ca...>]

Oh, another thing:

On 18 Nov 2002, Jeremy Fitzhardinge wrote:

> > It's at www.cag.lcs.mit.edu/dynamorio/.  The slides from the ASPLOS
> > tutorial are very interesting.  The whole thing was only released in June.
>
> Looks interesting.  It seems their biggest performance win was using BB
> chaining.  They also implemented a trace cache, though it seems to be a
> fair bit of complexity for hardly any incremental improvement.

The tutorial is a bit unclear, but I think the figures in the first part
don't account for any optimisations within the traces.  So it is likely
underestimating the advantage of the trace cache.

Interestingly, the original version of Dynamo, implemented on HPPA-RISC,
actually sped up most programs by up to 20%.  There's no clear reason why
the x86 version mostly slows them down;  maybe it's just an
engineering/maturity thing.

N