This is the mail archive of the 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]
Other format: [Raw text]

Re: Basic block reordering algorithm

On Wednesday 13 April 2005 20:46, Pat Haugen wrote:
> Steven Bosscher <> wrote on 04/13/2005 09:39:55 AM:
> > On Wednesday 13 April 2005 00:18, Pat Haugen wrote:
> > > When we have a test block gating whether a loop should be
> > > entered, the new block frequency check causes the code to pick the
> non-loop
> > > path as the next block to add to the trace since the loop header block
> has
> > > a higher frequency, and hence the loop gets moved out of line.
> >
> > Have you tried with the patch to re-enable the loop header copying
> > predictor:
> > It sounds like you might want to try that first...
> That does indeed fix the case when there is no profile data present. But if
> we have real profile data which presents a 50/50 probability on the gating
> block then we're back to the same situation of moving the loop out of line
> (although I guess one can always come up with scenarios to cause heuristics
> to look bad, which many times may not be worth worrying about).

Well, if there is a 50/50 probability, does it matter if the
loop is moved out of line?  There is a 50/50 chance that you
will not enter the loop, so moving it inline hurts the other,
equally likely path.

The problem with your original proposal is that computing
post-dominance information really is expensive.  Depending
on how often this 50/50 case happens, in a real profile, it
may or may not be worth the cost do as you suggested.  My
guess is that it doesn't happen very often and it isn't worth
it.  Maybe it is, you'd have to try.

With some luck we will have the infrastructure in place RSN
to preserve loop information over passes (at the very least
as a "loop region"). When that stuff is available, we could
teach bb-reorder to use that information to solve your problem
properly (and fix the find_traces_1_round hack you noted).


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