environments manual states otherwise, slw uses the whole contents of
the register instead of bits 27-31 as shift count (rlwnm doesn't)
* fixed generation of offset inside normal sets where bits have to be
inserted
* some bugfixes
* powerpc.inc is almost complete (only fillchar/word/dword, get_frame etc
and the class helpers are still needed
- removed unnecessary register saving in set.inc (thanks to compilerproc)
* use registers reserved for parameters as much as possible instead of
those reserved for local vars (since those have to be saved by the
called anyway, while the ones for local vars have to be saved by the
callee)
