[llvm] r342055 - [GVNHoist] computeInsertionPoints() miscalculates IDF

[Alexandros Lamprineas via llvm-commits]

Hi Philip,


Thanks for your suggestions. I indeed performed as many proactive checks as I could every time I re-enabled GVNHoist, but unfortunately buildbots are always configured slightly different from our local environment.


I've been running both csmith and llvm-opt-fuzzer overnight for a couple of days already, but none of them has caught anything interesting for GVNHoist so far. A bootstrap build of clang with UbSan enabled seems to be the most intense test for GVNHoist. It has revealed interesting bugs that have been already addressed. On that note, I'll try to re-enable it again today.


Regards,

Alexandros


________________________________
From: Philip Reames <listmail at philipreames.com>
Sent: Thursday, September 13, 2018 1:06 AM
To: Alexandros Lamprineas; llvm-commits at lists.llvm.org
Subject: Re: [llvm] r342055 - [GVNHoist] computeInsertionPoints() miscalculates IDF


I've noticed another round of work recently on fixing issues in GVNHoist.  First, let me start by saying, thank you for being responsive to issues and disabling quickly while investigation is in progress.  That makes this much less disruptive for the rest of the community.

However, I would like to suggest a change in approach from the current "enable and see what breaks" approach.

We have multiple community fuzzing efforts currently active.  Many of the issues being found seem like things that fuzzers would be good at uncovering.  Can I request that you make a proactive effort to leverage testing infrastructure such a fuzzers before re-enabling GVNHoist by default?

Specifically, we have the following tooling:

  *   csmith is available and I believe regularly run.
  *   We have the llvm-opt-fuzzer infrastructure which is run continuously on OSSFuzz.  Adapting this to exercise GVNHoist should be straight-forward.  (see https://llvm.org/docs/FuzzingLLVM.html)
  *   We (Azul) have a Java fuzzer which has found several subtle issues in GVN itself.  We can run a few cycles for you at some point, but since exposing the failing test cases is very manual for us, I'd ask you explore the other options first.

An alternate approach would be to identify one or more large projects - say Chromium, Clang itself -, build with GVN hoist enabled, and debug any failures which result.  Lest this is not misread, let me be quick to say that I haven't followed this effort closely.  It's entirely possible this has already been done and this is simply not visible to me as a casual observer.  If so, it might be helpful to spell out what has already been done and link to it from future re-enable commits.

Philip

On 09/12/2018 07:28 AM, Alexandros Lamprineas via llvm-commits wrote:

Author: alelab01
Date: Wed Sep 12 07:28:23 2018
New Revision: 342055

URL: http://llvm.org/viewvc/llvm-project?rev=342055&view=rev
Log:
[GVNHoist] computeInsertionPoints() miscalculates IDF

Fix for https://bugs.llvm.org/show_bug.cgi?id=38912.

In GVNHoist::computeInsertionPoints() we iterate over the Value
Numbers and calculate the Iterated Dominance Frontiers without
clearing the IDFBlocks vector first. IDFBlocks ends up accumulating
an insane number of basic blocks, which bloats the compilation time
of SemaChecking.cpp with ubsan enabled.

Differential Revision: https://reviews.llvm.org/D51980

Modified:
    llvm/trunk/lib/Transforms/Scalar/GVNHoist.cpp

Modified: llvm/trunk/lib/Transforms/Scalar/GVNHoist.cpp
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Transforms/Scalar/GVNHoist.cpp?rev=342055&r1=342054&r2=342055&view=diff
==============================================================================
--- llvm/trunk/lib/Transforms/Scalar/GVNHoist.cpp (original)
+++ llvm/trunk/lib/Transforms/Scalar/GVNHoist.cpp Wed Sep 12 07:28:23 2018
@@ -155,7 +155,6 @@ struct CHIArg {

 using CHIIt = SmallVectorImpl<CHIArg>::iterator;
 using CHIArgs = iterator_range<CHIIt>;
-using CHICache = DenseMap<BasicBlock *, SmallPtrSet<Instruction *, 4>>;
 using OutValuesType = DenseMap<BasicBlock *, SmallVector<CHIArg, 2>>;
 using InValuesType =
     DenseMap<BasicBlock *, SmallVector<std::pair<VNType, Instruction *>, 2>>;
@@ -767,7 +766,6 @@ private:
     ReverseIDFCalculator IDFs(*PDT);
     OutValuesType OutValue;
     InValuesType InValue;
-    CHICache CachedCHIs;
     for (const auto &R : Ranks) {
       const SmallVecInsn &V = Map.lookup(R);
       if (V.size() < 2)
@@ -786,6 +784,7 @@ private:
       // which currently have dead terminators that are control
       // dependence sources of a block which is in NewLiveBlocks.
       IDFs.setDefiningBlocks(VNBlocks);
+      IDFBlocks.clear();
       IDFs.calculate(IDFBlocks);

       // Make a map of BB vs instructions to be hoisted.
@@ -798,8 +797,7 @@ private:
         for (unsigned i = 0; i < V.size(); ++i) {
           CHIArg C = {VN, nullptr, nullptr};
            // Ignore spurious PDFs.
-          if (DT->properlyDominates(IDFB, V[i]->getParent()) &&
-              CachedCHIs[IDFB].insert(V[i]).second) {
+          if (DT->properlyDominates(IDFB, V[i]->getParent())) {
             OutValue[IDFB].push_back(C);
             LLVM_DEBUG(dbgs() << "\nInsertion a CHI for BB: " << IDFB->getName()
                               << ", for Insn: " << *V[i]);


_______________________________________________
llvm-commits mailing list
llvm-commits at lists.llvm.org<mailto:llvm-commits at lists.llvm.org>
http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-commits


IMPORTANT NOTICE: The contents of this email and any attachments are confidential and may also be privileged. If you are not the intended recipient, please notify the sender immediately and do not disclose the contents to any other person, use it for any purpose, or store or copy the information in any medium. Thank you.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20180917/9e47de09/attachment.html>