Re: call_insn's and argument locations

[law at redhat dot com]

  In message <200107021704.NAA23364@makai.watson.ibm.com>you write:
  > 	As Jeff said, most of the discussions to date have centered on the
  > new mid-level IR being an idealized form of RTL, without all of the
  > CLOBBER, cc0, etc. baggage.
That's what I think makes the most sense.

  > 	The very broad view, as far as I understand it, would move some
  > optimizations into the tree phase, then drop down into the mid-level IR
  > and SSA optimizations, then introduce more of the machine dependencies of
  > the current RTL and the current optimizations which still are relevant.
And that's precisely the model I've got in mind.   I've been spending most
of the last few weeks working through issues related to the mid level RTL
and SSA-RTL optimizers.

Note that this model still has room for Tree-SSA and optimizations using
it.  In fact, there are classes of optimizations which would be best done
at that level.

This model also has room for the targets to twiddle RTL in ways which
are cost effective for the target.  Introduction of target sensitive
sequences happens as we lower from the generic RTL to target RTL.



jeff