[patch] shark llvm 2.6 JIT improved optimization level

[Gary Benson]

Andrew Haley wrote:
> Xerxes Rånby wrote:
> > Andrew Haley skrev:
> > > Xerxes Rånby wrote:
> > > > The LLVM 2.6 JIT api have changed in order to give developers
> > > > more fine-grained control of the optimization level of the JIT.
> > > > http://lists.cs.uiuc.edu/pipermail/llvmdev/2009-April/021990.html
> > >
> > > OK, I'll bite.  :-)
> > >
> > > This option should be accessible from the command line.
> > 
> > I have committed the patch with a change to make it use the same
> > optimization level as before since shark currently only have one
> > optimization level.
> > 
> > If we want to be able to use the command line to set the
> > optimization level then it could be implemented in a number
> > of ways:
> > 
> > We could add a new -XX option like -XX:SharkOptimization=Aggressive
> > 
> > We could as well make shark hotspot respond to the "standard"
> > -client and -server options in order to set the optimization
> > suitable for responsive user interaction or bulk server
> > throughput.
> 
> I don't think that -client and -server really give us what we need.
> In order to find out what optimization levels we should use we need
> to be able to benchmark with varying levels.  If there is a Java
> parameter that corresponds with the LLVM levels, than we can use
> that, but I don't think there is.
> 
> It might be, for example, that Aggressive optimization is pointless.
> I don't know.

My preference would be for -XX:SharkOptimization=Aggressive.  The
problem with -client and -server is that they already have a meaning,
and they're handled in a completely different way to all other command
line options, so getting them through to Shark would involve some
pretty invasive modifications.

Cheers,
Gary

-- 
http://gbenson.net/