[Fwd: Re: Performance of locally copied members ?]
wimmer at ssw.jku.at
Mon May 3 17:30:57 PDT 2010
>> Client doesn't perform loop invariant code motion so caching by
>> hand will certainly make a difference for hot loops.
> Is this a feasible improvement for C1?
Thomas Wuerthinger implemented a limited form of loop invariant code
motion as part of the array bounds check elimination work. It is part
of the global value numbering and optimized all the loops that the
current global value numbering also optimizes, i.e., call-free blocks
with a maximum length of a few basic blocks.
More information about the hotspot-compiler-dev