[LLVMdev] GSOC Project Proposal: Profile-guided optimizations

[Gratian Lup]

Hi!
My name is Gratian and I would like to participate to GSOC 2011. I'm
interested in profile-guided optimizations, and I want to implement two
optimizations that can bring tangible benefits for most applications:
profile-guided function inlining and basic block positioning. Inlining can
be greatly improved if we take into consideration how many times the
function we want to inline was actually called. Functions that are not
called at all, or are called infrequently should not be inlined, while the
ones that are frequently called should have a higher chance of being
inlined. The algorithm I want to use is based on a benefit/cost ratio, so it
can take advantage of the existing InlineCost analysis. The algorithm is a
variation of the one found in JikesRVM, of course adapted and tuned for
LLVM. According to the paper, the performance of the application can be up
to 57% better than with inlining done without any profile info.

The second optimization is intended to replace the current
BasicBlockPlacement pass, which uses a naive algorithm, with an algorithm
that performs better in practice. It will be based on Algo2 from
Pettis&Hansen, while the current one is Algo1. I think this is more useful
than forming superblocks, because they may actually degrade the performance
if new optimization opportunities are not discovered (the size of the code
increases). If you think that superblocks are more useful, I could change
the proposal (I read two papers about superblock formation and optimization,
so I'm a bit familiar). The only problem would be that this, together with
inlining, may be too much for the scope of GSOC.
Any suggestions are welcome.

Gratian
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20110323/89b127ad/attachment.html>