rolw has no constraints for the registers.
Also it is faster on some processors.
On the other hand, xchgb is shorter, so try to replace rolw in the peephole pass, when optimizing for size and the register permits.

This fixes opt/fehler298.c.

Otherwise, the code would be broken.

This fixes opt/fehler299.c.

The optimization does not work properly for relation_true and relation_false.

* Parameter entities are shifted by the correct amount.
* All Return nodes are lowered, if there is more than one.