Factoring algorithms on rtl ?

[Joern Rennecke]

> We would very much like to continue this work, because even in this very
> simple code motion algorithm we have extra gain beside the mentioned
> algorithm when optimizing for size. I expect significant reductions from
> the more sophisticated factorings. Do you support this?

I wonder how to make sure that local registers are actually local to
the common sequences.  Do we have accurate REG_DEAD notes at the
point you want to do this factoring?