mark caller-saved registers as undefined on function entry

_From [derek.br...@gmail.com](https://code.google.com/u/117968039472581148324/) on December 10, 2010 17:57:54_

PR 425630

from PR 425498:

Brings up false negative problems that that same Valgrind paper says are very difficult to
avoid (not surprisingly): if we could assume the calling convention and
identify procedure entries we could mark scratch registers (caller-saved)
as undefined on entry.

Filing in case we want to try to implement this in a conservative manner.

_Original issue: http://code.google.com/p/drmemory/issues/detail?id=132_