- Semantic was strange: shallow checking for pointers, deep checking for
  structs (because no real solution in case of cycles was implemented)
- Implementation was incomplete/buggy
- Nobody was using it anyway

This is the conservative and fast solution. We would get better results
if we would follow all Sync predecessors and ensure there is a store on
all of them (or after a join point).

- use c99 features
- no unexpected and unnecessary tuple removing/graph modifying
- translate some comments from german to english

This should fix testsuite/opt/switch_ll.c

They weren't used in a sensible way and their semantic for
non-arithmetic ops was unclear.
Also implicitely fixes a bunch of bugs where these where set but no op_index
was set at the same time.

oparity_unary, oparity_binary require opindex to be set, which wasn't
always the case.
Also people seem to expect certain semantics from unops/binops, like
being arithmetic operations. So we shouldn't automatically set
oparity_XXX.

This fixes opt/combo5.c.

Set magic parameters according to
http://stackoverflow.com/questions/1711631/how-do-i-improve-the-performance-of-sqlite.

Also simplified the code.

Remove some unnecessary/unused functions. Simplify the implementation to
pass around const char* (which is a little bit faster and also nicer in
gdb)