Re: [Valgrind-developers] summary of patches

[Julian S. <js...@ac...>]

Hi,

> Before:
> --10356--  transtab: new        8,362 (300,800 -> 2,120,621; ratio
> 70:10)

> --14932--  transtab: new        8,359 (300,880 -> 2,119,507; ratio
> 70:10) [0 scs]

--tool=memcheck stresses the register allocator and spiller much more
than --tool=none, so it is a better test of any optimisations.  I suggest
you try it.

Also, maybe use the test cases in perf/; at least bz2.c, fbench.c and
ffbench.c ("make perf").  These are conveniently provided for exactly
this kind of reason :-)

> FWIW: I'm running this on an intel mac with vmware fusion. I wasn't
> using the CPU outside vmware fusion (only mild web browsing) so the
> test shouldn't have been skewed too much.

Hmm, getting reliable timing numbers even on a real machine is hard
enough.  On a VM is likely to be more difficult.  (The VMM introduces
more sources of measurement "noise" ..)

J