RE: [Open64-devel] RE: [Ipf-orc-support] ORC optimization level and involved optimizations

[Peng T. <tu...@te...>]

Very good points.  I haven't thought about these.

Since you wrote the whole DG package in the compiler,
you certainly know the intricacy and trade-offs better than
anyone.

I CCed the whole list to make sure my previous comments
don't lead some poor graduate students into a life long
endeavor :-)

Peng.
  -----Original Message-----
  From: Dror Maydan [mailto:ma...@te...]
  Sent: Wednesday, June 19, 2002 10:09 AM
  To: Peng Tu
  Subject: Re: [Open64-devel] RE: [Ipf-orc-support] ORC optimization level
and involved optimizations


  Peng Tu wrote:

    The function is in Build_CG_Dependence_Graph in lnopt_main.cxx.  It is
invoked after LNO
    finishs its transformations.  Current CG only understands the stripped
DG so it must be
    stripped.  However, you can certainly keep the original DG and pass it
to CG.  You will
    have to change the CG to use it.

    I would suggest you add a new interface to the DG (dep_graph.h) such
that the current
    CG will still get only the inner loop information (stripping
on-the-fly), whereas your new
    CG can use the full graph.

    Read the comments in dep_graph.h.  Another thing to keep in mind is that
the dependence
    graph passes through the global optimizer.  When the optimizer CSE
load/store, it preserve
    the DG.  It should work just fine, but may need some small fixes here
and there.
  I'd be nervous.  What happens if the optimizer deletes the outer loop
because it has only one iteration.  What happens if a load is moved out of
an inner loop into an outer.

  Dror


    Peng.

      -----Original Message-----
      From: rong [mailto:ro...@ca...]
      Sent: Wednesday, June 19, 2002 6:26 AM
      To: Peng Tu
      Cc: ope...@li...
      Subject: Re: [Open64-devel] RE: [Ipf-orc-support] ORC optimization
level and involved optimizations


      Thanks for the answer. Do you know exactly where the DDG for outer
loops are stripped? Before  Backend_Processing or after it? Do we must strip
it or we can preserve it?

      Thanks.

      Rong

      Peng Tu wrote:

It is not clear what you meant by multi-dimensional dependence.At -O3, LNO
builds array dependence graph for multi-dimensionalarrays in single and
nested loops.  The dependence graph is usedby LNO to perform loop
transformations. However, because the SWP only works on the innermost loop
of nestedloops. The dependence graph for outer loops are stripped before
itis passed to the SWP.I don't know if ORC has been updated with
multi-dimensional arrayidentification codes in LNO.  The GCC frontend always
flattens arrayaccess into a linear address expression.  The original LNO
arrayformation routine in the Pro64 release was built for the EDG
frontendwhere array subscripts are preserved.Peng.
-----Original Message-----Message: 1Date: Mon, 17 Jun 2002
15:49:19 -0400From: rong <ro...@ca...>To: "Chan, Sun C"
<sun...@in...>CC: "'Won So'" <ws...@eo...>,
ipf...@li...,
"'ope...@li...'"
<ope...@li...>Subject: Re: [Open64-devel] RE:
[Ipf-orc-support] ORC optimization level and involved optimiz	 ations.!!<
br>--------------020604090800050803020207Content-Type: text/plain;
charset=us-ascii; format=flowedContent-Transfer-Encoding: 7bitHello. I have
another question: in level O3, you mean swp gets multi-dimensional
dependence information from LNO? So it can be used not only for single
loops, but also nested loops. Am I right?Thanks.Rong
----------------------------------------------------------------------------
Bringing you mounds of caffeinated joy                   >>>
http://thinkgeek.com/sf
<<<_______________________________________________Open64-devel mailing
lis...@li...https://lists.sourceforge.net/lists/li
stinfo/open64-devel