6384206: Phis which are later unneeded are impairing our ability to inline based on static types
Summary:
Reviewed-by:

This changes C2's parser into a real reverse post order parser.
Previously C2 used a pre order which was close to RPO but not quite
the same. RPO parsing help guarantee that as often as possible all
predecessors of a block are parsed before we start parsing it. This
helps code fold up more quickly and feeds more exact types into
inlining. There's also a new lightweight bytecode analysis during
ciTypeFlow that identifies phis at loop heads which aren't needed.
This also helps with our inlining decisions. It's really a band aid
until we introduce real post parse inlining.

Additionally ciTypeFlow now performs real loop detection which is used
to drive loop head cloning in ciTypeFlow. Previously the cloning was
based on bcis which was unreliable and the cloning could introduce
irreducible loops.

I added a new assert in loop opts to check that we don't create
irreducible loops during optimization but unfortunately it currently
fails periodically because of the split_flow_path optimization in
cfgnode.cpp. Until I can do something about that I'm going to leave
the assert commented out.

I also modified CTW so that the CompileTheWorldStartAt flag only
controls when compilation begins, which help with reproducing bugs
under CTW. If you start loading later then it's often the case that
you don't end up with the same inlining decisions in other compiles.
I also changed it to make the code not entrant after the compile which
helps keep the code cache from filling up. After this change only 3
of the jar files managed to fill up the code cache during CTW.

I added some LogCompilation output to track the loops we encounter
during compilation. The current output looks like this:

<loop_tree>
<loop idx='7598' >
<loop idx='8114' inner_loop='1' pre_loop='7614' >
</loop>
<loop idx='9506' inner_loop='1' main_loop='9506' >
</loop>
<loop idx='8049' inner_loop='1' post_loop='7614' >
</loop>
</loop>
<loop idx='8265' inner_loop='1' pre_loop='7630' >
</loop>
<loop idx='9118' inner_loop='1' main_loop='9118' >
</loop>
<loop idx='8189' inner_loop='1' post_loop='7630' >
</loop>
<loop idx='7601' inner_loop='1' >
</loop>
</loop_tree>

I haven't finished the performance results yet though mostly they
should be neutral and that's what I've seen so far. The main effect
of this is to get static binding of call sites for which we previously
relied on type profiling to get the right results. In all the
interesting cases I'd looked at, previously we generally got the right
answer but now we consistently get the right answer so it might make
our performance more stable.

Tested with CTW, runThese and nsk on i586, amd64, sparc and sparcv9.