sjlj eh flow patch

[Jeffrey A Law]

  In message < 19971110192354.31792@dot.cygnus.com >you write:
  > The following patch generates more optimal basic block flows when
  > using sjlj by taking into account the fact that we know that longjmp
  > will go through a different path than the original call to setjmp.
  > 
  > It also gets rid of the stupid call to __dummy.  There is one backend
  > that needs to be fixed up because of this (MIPS); I'd been meaning to
  > do it but forgot.  If this patch is accepted for the post-1.00 release
  > I'll finish that bit.
I think this should go in the mainline instead of the release branch.

jeff