My last question

My last question

[Richard Kenner]

I see the way it was done was to avoid reassigning stack slots to objects
of a different alias set.  That will indeed solve the problem but will
also have the effect, for most purposes, of disabling the sharing of temporary
slots.  So I'd like to find some other approach if we can.

Re: My last question

[Mark Mitchell]

>>>>> "Richard" == Richard Kenner <kenner@vlsi1.ultra.nyu.edu> writes:

    Richard> I see the way it was done was to avoid reassigning stack
    Richard> slots to objects of a different alias set.  That will
    Richard> indeed solve the problem but will also have the effect,
    Richard> for most purposes, of disabling the sharing of temporary
    Richard> slots.  So I'd like to find some other approach if we
    Richard> can.

Yes.  I've argued that stack slots should be assigned like registers.
There should be stack-pseudos for a while, and then there should be
stack-allocation.  In fact, in the long run, I'm not sure that
stack-pseudos should even be different from ordinary pseudos.  One of
the long-standing problems with GCC's optimization is that we commit
to code shape too early.  ADDRESSOF is an attempt to avoid this
problem, and is certainly an improvement, but we should be able to do
better.

--
Mark Mitchell                   mark@codesourcery.com
CodeSourcery, LLC               http://www.codesourcery.com