This was spurred a bit by the need to fix 67892, which is a problem with 
the old threader's support for threading across backedges.

I've indicated in the past that I'd like to simply drop that code.  It's 
fragile and generally just a PITA.  The functionality would be replaced 
by Sebastian's FSM jump threading code.

Sebastian's code should also form the basis for a jump threader that 
operates independently of DOM/VRP with the ultimate goal of totally 
dropping the old threader.  That's obviously further out, but that's 
where I'm going.

There are things Sebastian's code doesn't handle, some of which I 
already know how to fix, some of which are still TBD.  But there's no 
reason to not start this process.

The first step in my mind is to get Sebastian's code factored out of 
tree-ssa-threadedge.c.  tree-ssa-threadedge.c will still use a single 
entry point in Sebastian's code during the transition.

That single entry point can also be used by a distinct jump threading 
pass (I've already done this as a proof of concept locally).

Once factored out, I can start enhancing Sebastian's code under the hood 
to start handling the missing cases with an eye towards handling enough 
cases that I can remove the old threader's support for threading across 
backedges (thus fixing 67892 along the way).

This patch strictly moves bits out of tree-ssa-threadedge.c into a new 
file tree-ssa-threadbackwards.[ch] and minimizes the headers included by 
tree-ssa-threadedge.c.  I chose "backwards" over "fsm" as I see this 
going to a place that is far more general than the FSM optimization, 
basically a backwards walking jump threader.

The patch changes no logic or even the "fsm" references in the existing 
code.

Bootstrapped and regression tested on x86_64-linux-gnu.  Installed on 
the trunk.

jeff