We now restart the optimization as soon as we changed an operand or the
relation. This makes the whole optimization less error-prone.

This fixes opt/fehler309.c.

This allows the use the function in more cases.

The involved conversion are all same size and won't create extra target
instructions. So we can safely perform the transformtion at the Add/Sub
level to get a normalizing effect. This also fixes a problem where an
AddP was constructed with the wrong offset_mode.

This fixes firm-testsuite/opt/fehler307.c

We can get a proper mode for pointer-pointer from
get_reference_offset_mode().

Also choose a signed integer type which is more typical for offsets.

Other localopts/confirm construction already optimize this case, so
testing a==b is enough to check for equality.

This flag can be set on entities where the actual address isn't
known/compared against other address. This also fixes a bug where we
used the mach-o csrings section for global char arrays which need
distinct addresses.

It doesn't matter for correctness but looks nicer.

As of the switch to C99 libfirm uses too many features that are not
supported by microsoft visual c and I don't see any hope to support it
again.
Windows users can still switch to a gcc based toolchain (cygwin or msys)
or may have better luck with a proper C99 compiler (clang, icc, future
versions of msvc?)

It combines sched_add_before() and be_peephole_exchange().

Default initialisation just happens to set the correct value.