This reverts commit 26e639c4.
We cannot benefit from bitinfo if we want *exactly* one bit.

- Managed resources within the analysis itself
- Let constbits take care of the obstack to store the bit information

Mux((a & (1 << c)) != (b & (1 << c)), 0, 1) => (a ^ b) >> c.

Didn't notice this before because gcc doesn't report a warning for this,
clang does.

- document this correctly
- correctly test with tarval_is_constant() if you expect a constant
  value.

This avoids using reserved identifiers

These 2 functions have a special role as we cannot fold the whole node
with computed_value() because of mode_T result. The exising code was
broken as it tried to apply value_of() on a mode_T node.

Otherwise the bitinfo result was sometimes worse than the
computed_value() results which was not expected by the code in iropt.

The simple x+(-x) case worked already because it was converted to a x-x,
but the more complicated case (x-y) + (y-x) did not work. While
certainly the reassoc could perform better here, adding this rule
as a localopt doesn't hurt.

This fixes opt/fehler231.c and opt/fehler232.c.

The modes may differ.
E.g. execution nodes use a boolean.

Especially, do not confirm consistent out edges.
This fixes ack/t7.c on SPARC.

Apparently it is optional on linux (just often used because the RIP
encoding is often shorter than a global encoding).

Along the way:
- Simplified code to sharpen Cmp relations
- Do not remove unreachable code by using bit information anymore

This fixes opt/fehler230.c.