ppc regression

[Jeffrey A Law]

  In message < 9906181517.AA31982@marc.watson.ibm.com >you write:
  > 	I originally was considering a splitter that was all RTL, letting
  > the compiler figure out the most efficient way to add four to the address,
  > but I can place the C code in the define_insn for the time being.
This is probably the best short term.  We can write splitters for all the
appropriate double-word moves in the mainline tree at our leisure.

  > I would
  > not think that I would need all of the refers_to_regno_p() stuff if I were
  > using a splitter as the compiler would allocate registers appropriately
  > given its other uses.  No?
I believe you still need the refers_to_regno_p stuff.

If I understand the pattern correctly, you only want to split some of the
cases -- and to determine when you want to split the insn you need to know the
hard register assignments.

Thus, the splitter would not be active until after register allocation and
you can not depend on the allocators re-allocating registers for you.

jeff