This is the mail archive of the gcc-patches@gcc.gnu.org mailing list for the GCC project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]

Re: moving constants out of loops



  In message <9806302327.AA14100@vlsi1.ultra.nyu.edu>you write:
  >     Furthermore, we found that exposing more givs and invariants wasn't
  >     always a win.  In fact, it was a huge loss on the PA for a while (see
  >     alias discussion below).
  > 
  > My feeling is a bit different here.  There is currently code that attempts
  > to decide which givs are worth doing something about. I feel that the
  > more choice it has to pick from, the better job it wil do.  But that's
  > not relevant in this case, since what's being found aren't "givs",
  > but constants.
Now that we agree that we're really enabling the loop optimizer to
find more givs hopefully my comment makes more sense.  ie, the
existing alias code in gcc2 isn't up to the task of dealing with
the more complex givs (in particular those used for memory addresses).

The net result is severe impact on the scheduler's ability to reorder
memory references within loops.


jeff


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]