On Tue, 29 May 2001, Chan, Sun C wrote:
>
>
> > -----Original Message-----
> > From: Peng Zhao [mailto:pengzhao@xxxxxxxxxxxxxx]
> > Sent: Thursday, May 24, 2001 9:21 PM
> > To: sgi
> > Subject: several times of instrumentation? (sorry for the
> > previous email
> > without title)
> >
> >
> >
> > Hi,
> >
> > I noticed that Pro64 make several instrumentations during the
> > backend phase (e.g. just before VHO, before LNO, before WOPT
> > and cg etc).
> > Is it because that transformation makes the annotation
> > imprecise and we
> > need to read the profiling result again? How to ensure the profiling
> > result compatible with the transformed code?
>
> In fact, the correct phrase should be Pro64 "can" instrument
> during the phases you mentioned (not cg though). In reality, it does
> that before VHO. The design is so that you can do it in any phase.
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
So, there are still possibilities that the feedback
information is read in the later phase than VHO. Am I right? Does this
implies that the instrumentation conducted in the first compilation must
happen at the same phase of the annotation in the second compilation?
And the instrumentation or annotation can only take place once during the
compilation, although Pro64 can do it in several phases. Right?
You mean that the instrumentation or annotation is only conducted
before VHO even when various optimization flags are switched
on(i.e. other calls in later phases to annotation are never invoked)?
>
> >
> > E.g, I want to do partial inlinining in the backend
> > phase. And my
> > initial plan of doing it is just after the first annotation
> > (there still
> > needs some more investigation for the feasibility, of course your
> > suggestions are quite welcomed). After the partial
>
> You meant during IPA?
No. in the VHO phase ( just before the lowering from VH whirl to H
whirl). It is after the annotation.
>
> > inlining, the caller and callee change definitely. Is there
> > any problem
> > that later annotation?
>
> The question is how you update the info after partial inlining. Since its
> a form of cloning, you need to make some guestimation. The feedback is
> designed for being able to verify before and after each optimization and
> assert that the information is still "consistent". The update heuristic
> is strictly up to the optimization phase since each has its own peculiar
> requirement and hence different ways to deal with it. Pro64 can also
> do some minor patching when annotation is inconsistent. When it cannot
> patch,
> it will invalidate a portion of the information.
> > I guess the problem exists for other optimizations.In current Pro64,
> > how the similar problems are addressed?
> >
> > Thanks.
>
>
--
Regards
Peng
Peng Zhao pengzhao@xxxxxxxxxxxxxx
http://www.cs.ualberta.ca/~pengzhao
TEL (Lab): (780)492-3725 Lab: CSC251
|