x86 jitter? (One more try!)
Henrik.Gedenryd at lucs.lu.se
Sat Apr 14 11:23:29 UTC 2001
Tim Rowledge wrote:
> It would be entirely possible to make an improvement to the overall
> performance of the system (as measured by larger benchmark code) that
> actually lead to a lower value for the simple bytecodes/sec test. I
> suspect that would be harder w.r.t the send/return test.
Forgive me if I misunderstood you, but that was an unusually roundabout way
of saying that the sends figure is the one to care about, since it is quite
relevant to real-world performance, whereas the bytecodes figure is not.
However, it might be added, as far as the JIT is concerned its sends value
increase does not reflect the increase in real-world performance that you
gain from using the JIT.
I for one would be interested in info on how to tweak JIT parameters to
improve real-world performance. I can imagine that such things as changing
the size of the JIT cache will be useful for real programs that don't stay
within a handful of methods.
More information about the Squeak-dev