On Sat, Oct 06, 2012 at 12:15:16PM +0200, Aurelien Jarno wrote: > On Fri, Oct 05, 2012 at 05:00:04PM -0700, Richard Henderson wrote: > > This applies with or without the sparc-compare patch set I > > recently sent, and it works with the same set of tests. > > > > I've not had time to do true benchmarking on this, but it > > does reduce the size of the generated code: > > Experience proves that there is not a direct relation between size of > the generated code and the resulting emulation speed, sometimes smaller > code means slower emulation (when the code generation/optimization takes > too much time), and sometimes bigger code might means faster emulation > (think about replacing some helpers by TCG code). > > As from the user point of view what is important is the emulation speed, > I think benchmarks (even simple ones like measuring the boot time of a > guest) are essential for this kind of patch. >
For what it worth, I measure a 4% slow down booting a sparc64 guest on a Core-i5 2500 machine. I guess the memcpy() of the register windows is more expensive that the gain on the TCG side, though it should be probably be confirmed using some profiling tools. -- Aurelien Jarno GPG: 1024D/F1BCDB73 aurel...@aurel32.net http://www.aurel32.net
