Re: 19991023 ICE compiling XFree86

[Jeffrey A Law <law at cygnus dot com>]

  In message <19991026211023.A20666@cygnus.com>you write:
  > > I believe that was the whole point behind the unusual definition of
  > > find_basic_block (I'm guessing this originally came from reorg.c).
  > 
  > Yeah it did.  I can't imagine how it used to work though.  The
  > incorrect answer I was seeing wasn't even vaugely close.
  > 
  > So, like, what does reorg need this resource stuff for if it
  > doesn't have to be right?  Does it get some indication that the
  > data isn't valid and just gives up?  I guess some study is in
  > order....
It needs to track what registers are live on both paths of conditional jumps
so that it knows what insns are always safe to put in a delay slot or
what insns are safe to put in an annul-true or annul-false slot.

If it gets too confused it is supposed to give up and assume the worst.

It's always been some of the worst code in reorg/resource.


I actually looked at replacing cross-jump completely with something similar
to code hoisting.  The basic ideas are the same, except that we move stuff down
in the cfg to post dominators.  Anyway, I never got around to working through
all the details (it's assignment motion and needs to deal with hard regs too).


jeff