[PATCH] D22305: [BasicAA] Strip phi nodes, when all incoming values are the same.

Thu Jul 14 19:49:14 PDT 2016

I can measure compile time impact of (2) and (3) on power. A couple of
questions/remarks.

1- As for (2) I don't think there will be a significant compile time cost.
If we do not reduce the phi to its unique incoming value, we will end up in
aliasPHI and we do this check there. It is just too late for some cases
like the problem here, and we get conservative result. Do I miss something
here?

2- I will use Daniel's patch for (3). Depending on how expensive it is, we
can look at finer grain cache clean ups, as Hal suggests.

3- I will modify my patch for (2) and modify stripPointerCasts to include
"seeing through phi". I wanted to separate that move as a next step, but
Hal disagreed on the code review.

If you disagree with this way of doing the experiments, please let me know.

Thanks
Ehsan

From:	Hal Finkel <hfinkel at anl.gov>
To:	Daniel Berlin <dberlin at dberlin.org>
Cc:	llvm-commits <llvm-commits at lists.llvm.org>, <reviews+D22305
            +public+92ca108e50bc4651 at reviews.llvm.org>, Ehsan A
            Amiri/Toronto/IBM at IBMCA
Date:	07/14/2016 09:35 PM
Subject:	Re: [PATCH] D22305: [BasicAA] Strip phi nodes, when all
            incoming values are the same.

 From: "Daniel Berlin" <dberlin at dberlin.org>
 To: "Hal Finkel" <hfinkel at anl.gov>
 Cc: "llvm-commits" <llvm-commits at lists.llvm.org>, reviews+D22305+public
 +92ca108e50bc4651 at reviews.llvm.org, "Ehsan A Amiri" <amehsan at ca.ibm.com>
 Sent: Thursday, July 14, 2016 8:30:09 PM
 Subject: Re: [PATCH] D22305: [BasicAA] Strip phi nodes, when all incoming
 values are the same.

  (2) is, in theory, the right thing to do. Even if we were to consider
  uniform PHIs to be anti-canonical, and thus something which should be
  simplified, we can't simplify often enough to prevent these from blocking
  useful analysis work.

 FWIW, i'm fine with this approach if our approach is going to be as you
 say - that we will not simplify often enough.

 Right now, as i said, we simplify *everywhere*, and every one of those
 calls will eliminate this phi node.

 So it's only *this particular path* that misses all those calls.

 For example, if the alias check had gone through a gep of a phi anywhere,
 it would simplify the phi as part of getunderlyingobject, etc.

  Arbitrary uses of RAUW can create these PHIs, and we can't (and probably
  shouldn't) run InstCombine in between every other pass. This is a local
  pattern that stripPointerCasts, and similar functions, can look through.

 Fine with this as long as we maybe stop trying to simplify instructions 8
 or 9 times, and instead do it once (max) per instructions, and make this
 part of it.

 (IE This would mean we would have SimplifyAndGetUnderlyingObject and
 GetUnderlyingObject, and we simplify once and call the latter or
 something, or whatever. Not suggesting we decide this second, just
 suggesting that we agree if this is going to be our general approach).
This is hard because we don't cache the simplifications in any way. It is
not like we're updating the IR when we discover some simplification; we're
only using the simplified version in place. I'm not sure how to fix this.
Maybe we should run InstSimplify a lot more often. It is not as expensive
was InstCombine by a large margin (IIRC).

  (3) is also, in theory, the right thing to do. The memdep cache, by
  necessity, caches negative results. Each GVN iteration, however, performs
  "information revealing" operations which can make the cached results more
  conservative than a new query's results.

 Yes.
 It's theoretically possible to make it less expensive than blowing away
 the whole cache, but so far, experience has told me that fully maintaining
 the cache becomes slower than redoing the queries :P
I'm not sure how much this helps, but we could/should only clear out the
MayAlias results.

 -Hal

  Now we actually need to measure the costs.

   -Hal

   (Note: i've attached a patch for 3 in case anyone wants to see the
   compile time cost)

   On Thu, Jul 14, 2016 at 2:12 PM, Daniel Berlin <dberlin at dberlin.org>
   wrote:
    Note: This already had GVN run once on it, do you have the one before
    that?

    On Thu, Jul 14, 2016 at 1:53 PM, Ehsan A Amiri <amehsan at ca.ibm.com>
    wrote:
      I can see the problem with the following command line and attached
      file

      opt -gvn t.ll

      (See attached file: t.ll)

      My clang is almost a week old.

      Inactive hide details for Daniel Berlin ---07/14/2016 04:27:43
      PM---Actually, can you please attach a .ll file and an opt comma
      Daniel Berlin ---07/14/2016 04:27:43 PM---Actually, can you please
      attach a .ll file and an opt command line that reproduces the
      problem?

      From: Daniel Berlin <dberlin at dberlin.org>
      To: Ehsan A Amiri/Toronto/IBM at IBMCA
      Cc: Hal Finkel <hfinkel at anl.gov>, llvm-commits <
      llvm-commits at lists.llvm.org>, reviews+D22305+public
      +92ca108e50bc4651 at reviews.llvm.org
      Date: 07/14/2016 04:27 PM
      Subject: Re: [PATCH] D22305: [BasicAA] Strip phi nodes, when all
      incoming values are the same.

      Actually, can you please attach a .ll file and an opt command line
      that reproduces the problem?
      The clang command line you have is very sensitive to versions, etc.

      I cannot get your issue to reproduce with the clang i have installed
      that can target powerpc-linux (and the issue does not reproduce with
      your testcase on x86) :)

      While debugging a bit, note that there is at least one obvious bug in
      GVN that may affect this, by inspection:

      When GVN splits a critical edge, it never adds the new block to the
      iteration order (at all), even though it inserts into it.
      So they will not get processed until the next iteration of GVN on the
      function, even though they have code in them.  While this is okay
      from a correctness standpoint, it may block optimization of certain
      things (including the cases you've discovered).  In practice, there
      is no way to perfectly solve that without pre-splitting all critical
      edges, but you should get the same effect if we throw the critical
      edge block and then it's successors (including the current blcok)
      into bbvect after the current block again.

      It is also missing a real phi simplification.
      While simplifyinstruction will check if all arguments are trivially
      the same, that is not the real test that should be performed.

      It should be doing VN.lookup on each argument and seeing if they come
      up with the the same value number.

      Once you attach the .ll file, i'll fix these and see if it fixes your
      testcase, and if not, debug further.

      On Thu, Jul 14, 2016 at 10:11 AM, Daniel Berlin <dberlin at dberlin.org>
      wrote:

            On Thu, Jul 14, 2016 at 7:09 AM, Ehsan A Amiri <
            amehsan at ca.ibm.com> wrote:
            This is order of events:
            This order cannot be correct if the solution I gave (or adding
            simplification to *some place in GVN*) does not work.  Or GVN
            is broken in other ways.

            1) GVN starts looking at the function. At this point the phi
            node has two different incoming values.
            2) GVN performs an RAUW. The phi is converted to the one that
            has two identincal incoming values.

            At this point, it should now process the phi instruction again
            before it processes the load, because it is doing a reverse
            postorder traversal.
            When it did that, the phi should have been simplified
            So why did that not happen?

            Given the complexity of fixing the real problem,

            Look, i understand why you want to just fix this in AA and be
            done with it.
            Really, I do.

            I understand you have spent a lot of time on this bug, and I
            greatly appreciate that.
            But I really want to understand what is going on before we try
            to actually fix it.
            I have a good understanding of what happens once the bad answer
            gets into memdep (and thank you for that!), but i still have
            trouble seeing why it lived long enough to get there.

            To that end, so you don't have to spend more time running
            around for me,  i'll take over this bug, and either figure out
            why GVN lets this PHI live to the point it gets an AA query
            about it (and fix it/decide it can't be fixed), or commit the
            AA patch for you if we decide it can't be fixed.

            My ETA is by friday.

            I assume the testcase in the bug is the one we are still using,
            right?
            (If not, if you can attach it, that would be helpful)

  --
  Hal Finkel
  Assistant Computational Scientist
  Leadership Computing Facility
  Argonne National Laboratory

--
Hal Finkel
Assistant Computational Scientist
Leadership Computing Facility
Argonne National Laboratory
[attachment "graycol.gif" deleted by Ehsan A Amiri/Toronto/IBM]
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20160714/9d06329a/attachment.html>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: graycol.gif
Type: image/gif
Size: 105 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20160714/9d06329a/attachment.gif>