This is the mail archive of the
gcc@gcc.gnu.org
mailing list for the GCC project.
Re: Tree-profiling branch status
On May 17, 2004, at 8:44 AM, Richard Guenther wrote:
On Mon, 17 May 2004, Jan Hubicka wrote:
On Mon, 17 May 2004, Jan Hubicka wrote:
What's going to happen with the tree-ssa-20020619-branch based
branches
now that the tree-ssa branch is closed? I know only of lno and
tree-profiling, but maybe there are others. I first thought they
would
perhaps be merged into tree-ssa-20020619-branch and development
continues
there.
The overall plan for tree-profiling branch has been to implemenet
the
CFG transparent expansion, inlining and to build CFG just once
before
inlining so the profile can be maintained in it.
The CFG transparent expansion is currently done and I believe that
it is
quite stable (it needs testing on more wide variety of platforms).
The only major problem is the fact that mudflap is completely
broken but
rewriting it into CFG aware version is not major task I believe.
The branch also contains fixed branch predictors that can be run
early,
GCOV support and the tree-profiling itself.
The CFG transparent inlining is still being worked on, Stuart will
be
able to say more details ;)
What do you mean with CFG transparent inlining? You mean we can
inline
GIMPLE then and do basic optimization before inlining? That would be
cool. Anyway, I submitted the leafify patch for inclusion two weeks
ago,
Yes, basically the idea is that we can build CFG, do early
optimization
phasses and build the callgraph during analysis phase.
Then we can decide inlining and do inlinnig already on the gimple form
with CFG built that would allow profile to be already load, doing
partial inlining and similar tricks.
I forgot to add Stuart into CC, he has basic patch for this working,
but
I didn't had chance to read it in detail yet.
Stuart, if you can summarize the status and how much work you think it
will be needed, it would be cool.
It's that I'd like to get rid of __attribute__((leafify)), and if we
can
do some early loop analysis before inlining, we could tune inlining
heuristics (or switch them off) based on loop-nest and
predicted/profiled
iteration count. Do you plan to inline in SSA form or just before
going
into SSA?
Yes, that's the current plan. However,
Or do gimple -> SSA -> gimple before inlining?
I suspect this would work too, and it would be an easy experiment to
try.
stuart hastings
Apple Computer