Improve the set of LLVM optimization passes

[GoogleCodeExporter]

We currently use a minimal set of LLVM optimization passes [1]. We should
explore more passes and other pass orderings. Sources for these:

- LLVM's opt tool [2].
- MacRuby [3]. I notice that MacRuby is using one or two passes that we
aren't, like createTailCallEliminationPass() and createReassociatePass().
- Rubinius [4]. Looks like Rubinius is also using passes we aren't.

We should evaluate their optimization sets against perf.py

[1] -
http://code.google.com/p/unladen-swallow/source/browse/trunk/Python/global_llvm_
data.cc
[2] -
http://llvm.org/viewvc/llvm-project/llvm/trunk/include/llvm/Support/StandardPass
es.h
[3] -
http://svn.macosforge.org/repository/ruby/MacRuby/branches/experimental/vm.cpp
[4] -
http://github.com/evanphx/rubinius/blob/4773e97cbd04d50f41aa58e6bb2b61a550187fc7
/vm/llvm/jit.cpp

Original issue reported on code.google.com by collinw on 12 Jun 2009 at 11:04

[GoogleCodeExporter]

Adding the opt -O3 flags as a local -O3 level. At the moment, these don't appear
faster, but that's probably partly because issue 57 isn't done.

Original comment by jyass...@gmail.com on 17 Jun 2009 at 6:18

• Changed state: Started

[GoogleCodeExporter]

Submitted as r664. They still didn't speed things up. I'm going to stop working 
on 
this until type feedback gives the optimizers more information to play with.

Original comment by jyass...@gmail.com on 22 Jun 2009 at 6:36

• Changed state: Accepted
• Added labels: Priority-Medium, Release-2009Q3
• Removed labels: Priority-High, Release-2009Q2

[GoogleCodeExporter]

Original comment by collinw on 29 Sep 2009 at 5:57

• Added labels: Release-2009Q4

[GoogleCodeExporter]

Original comment by collinw on 6 Jan 2010 at 11:42

• Removed labels: Release-2009Q4

[GoogleCodeExporter]

r1044 dropped the unused and unproven -O3 optimization passes. We can always 
pull 
them back out of svn if we need them.

Original comment by collinw on 27 Jan 2010 at 1:50