In particular this allows ifconv when a PhiM is present, but nothing happens with the memory on the two paths.

This is a workaround for keepalive edges which are probably broken
(adding the stuff to the End node is wrong, it should get added to the
 endless loop somehow)

SymConst_addr_ent are never zero for example, also there's no need to
give up after seeing a confirm

I aligning them to a multiple of 100 now, so that when code generated in
an earlier functions only changes lightly it is very likely that the
block numbers in later functions stay the same.

Performing it as localopt was hairy as the optimisation dependet on
correct iredges values which were misleading/wrong during construction
of new stuff.

The middleend should optimize all these cases, it does so for 99% of
them already. There are just some cases where ands get transformed into
conversions which the middleend doesn't catch yet.

This needed some rewriting as ia32 previously checked already
transformed nodes while sparc was checking nodes before transformation.
Change ia32 to check nodes before transformation too. Also improved some
code in sparc+ia32.

The hashmap used to point to blocks which have already been deleted.

This means floatingpoint calculation in a firm graph must be performed
in the mode specified (and may not be performed in a wider mode as
before).