[LLVMdev] RFC: Loop versioning for LICM

Nema, Ashutosh Ashutosh.Nema at amd.com
Tue Mar 3 01:29:01 PST 2015

Hi Adam,

Thanks for looking into LoopVersioning work.

I have gone through recent LoopAccessAnalysis changes and found some of the stuff
overlaps (i.e. runtime memory check, loop access analysis etc.).  LoopVersioning can
use some of the things from LAA.

LoopVersioning is a memory check based multi versioning optimization, it simply creates
aggressive alias version of loop preceded by a memory check. It’s not concerned about
the order of instructions and detailed dependency check that LoopVectorizer does.
It does some basic loop structure check, loop instruction checks & memory checks.

In general found LAA work is more inclined towards LoopVectorizer.
Found some of the possible reusable functions are biased towards LoopVectorizer,
they has specific condition checks for it.

It’s good to make some of the classes & function more generic and reusable.
Will be covering some of the points in this mail.

While creating runtime check I have found, some of the things are not getting considered.
1) No need to check if two read only pointers intersect.
2) Only need to check pointers between two different dependency sets.
3) Only need to check pointers in the same alias set

I’m sure if we like this to be used by other optimization then not all optimization appreciate
above checks. Specifically LoopVersioning does not care about this, it expects all the pointers
in a loop should be considered for a memory check. Also it does not care about different
dependency set & different alias sets.

I suggest we can make these checks optional, and give flexibility to users of this class to set it.

For the same suggesting following change:
class RuntimeCheckEmitter {
  /// Consider readonly pointer intersection in memcheck
  bool CheckReadOnlyPointersIntersection;
  /// Consider pointers in same dependency sets for memcheck.
  bool CheckPointersInSameDependencySet;
  /// Consider pointers in different Alias sets for memcheck
  bool CheckPointersInDifferentAliasSet

Add the above 3 variables to class, and allow users of this class to set it.

In "RuntimeCheckEmitter::addRuntimeCheck" following 3 condition needs to
controlled by above conditional variables.

   Change Following Check:
      // No need to check if two readonly pointers intersect.
      if (!PtrRtCheck->IsWritePtr[i] && !PtrRtCheck->IsWritePtr[j])
      // No need to check if two readonly pointers intersect.
      if (!CheckReadOnlyPointersIntersection && !PtrRtCheck->IsWritePtr[i] &&

   Change Following Check:
      // Only need to check pointers between two different dependency sets.
      if (PtrRtCheck->DependencySetId[i] == PtrRtCheck->DependencySetId[j])
      // Only need to check pointers between two different dependency sets.
      if (!CheckPointersInSameDependencySet &&
             PtrRtCheck->DependencySetId[i] == PtrRtCheck->DependencySetId[j])

   Change Following Check:
      // Only need to check pointers in the same alias set.
      if (PtrRtCheck->AliasSetId[i] != PtrRtCheck->AliasSetId[j])
      // Only need to check pointers in the same alias set.
      if (!CheckPointersInDifferentAliasSet &&
            PtrRtCheck->AliasSetId[i] != PtrRtCheck->AliasSetId[j])

By this we allowing RuntimeCheckEmitter as more flexible and providing user
more control to use it.

Here again its very specific to LoopVectorizer.
The way it handles stores & loads may not be appreciated by other optimization
expecting other treatment. I suggest we should think on flexibility for user to
override load & store handling. We can provide virtual methods for load & store
handling (i.e. analyzeLoads & analyzeStores). Also some of the optimization may not
like call instruction, or further they like to analyze call. We should also think on those
lines to make some provision.

AccessAnalysis & LoopAccessAnalysis are tied up dependency check, If some analysis
needs same functionality except dependency check then there should be provision available.
i.e. LoopVersioning needs similar stuff except dependency analysis, for now possibility is
extend & rewrite functions by removing dependency checks.


From: Adam Nemet [mailto:anemet at apple.com]
Sent: Friday, February 27, 2015 12:40 AM
To: Nema, Ashutosh
Cc: llvmdev at cs.uiuc.edu
Subject: Re: [LLVMdev] RFC: Loop versioning for LICM

Hi Ashutosh,

Have you been following the recent Loop Access Analysis work?  LAA was split out from the Loop Vectorizer that have been performing the kind of loop versioning that you describe.  The main reason was to be able to share this functionality with other passes.

Loop Access Analysis is an analysis pass that computes basic memory dependence and the runtime checks.  The versioning decision and then performing the transformation are left to the transform passes using this analysis.

If we decide that a stand-alone memcheck-based loop-versioning is desired we should probably use this analysis and possibly extend it instead of duplicating the code.


On Feb 26, 2015, at 2:31 AM, Nema, Ashutosh <Ashutosh.Nema at amd.com<mailto:Ashutosh.Nema at amd.com>> wrote:

I like to propose a new loop multi versioning optimization for LICM.
For now I kept this for LICM only, but it can be used in multiple places.
The main motivation is to allow optimizations stuck because of memory
alias dependencies. Most of the time when alias analysis is unsure about
memory access and it says may-alias. This un surety from alias analysis restrict
some of the memory based optimizations to proceed further.
We observed some cases with LICM, where things are beyond aliasing.
In cases where alias analysis is unsure we like to use loop versioning as an alternative.

Loop Versioning will creates version of the loop with aggressive alias and the other
with conservative (default) alias. Aggressive alias version of loop will have all the
memory access marked as no-alias. These two version of loop will be preceded by a
memory runtime check. This runtime check consists of bound checks for all unique memory
accessed in loop, and it ensures aliasing of memory. Based on this check result at runtime
any of the loops gets executed, if memory is non aliased then aggressive aliasing loop
gets executed, else when memory is aliased then non aggressive aliased version gets executed.

By setting no-alias to memory accessed in aggressive alias version of loop, enable other
optimization to continue further.

Following are the top level steps:

1) Perform loop do versioning feasibility check.
2) If loop is a candidate for versioning then create a memory bound check, by considering
     all the memory access in loop body.
3) Clone original loop and set all memory access as no-alias in new loop.
4) Set original loop & versioned loop as a branch target of runtime check result.
5) Call LICM on aggressive alias versioned of loop(For now LICM is scheduled later and not directly
     called from LoopVersioning pass).

Consider following test:

     1  int foo(int * var1, int * var2, int * var3, unsigned itr) {
     2    unsigned i = 0, j = 0;
     3    for(; i < itr; i++) {
     4      for(; j < itr; j++) {
     5        var1[j] = itr + i;
     6        var3[i] = var1[j] + var3[i];
     7      }
     8    }
     9  }

At line #6 store to var3 can be moved out by LICM(promoteLoopAccessesToScalars)
but because of alias analysis un surety about memory access it unable to move it out.

After Loop versioning IR:

<Versioned Loop>
for.body3.loopVersion:                            ; preds = %for.body3.loopVersion.preheader, %for.body3.loopVersion
  %indvars.iv.loopVersion = phi i64 [ %indvars.iv.next.loopVersion, %for.body3.loopVersion ], [ %2, %for.body3.loopVersion.preheader ]
  %arrayidx.loopVersion = getelementptr inbounds i32* %var1, i64 %indvars.iv.loopVersion
  store i32 %add, i32* %arrayidx.loopVersion, align 4, !tbaa !1, !alias.scope !11, !noalias !11
  %indvars.iv.next.loopVersion = add nuw nsw i64 %indvars.iv.loopVersion, 1
  %lftr.wideiv.loopVersion = trunc i64 %indvars.iv.loopVersion to i32
  %exitcond.loopVersion = icmp eq i32 %lftr.wideiv.loopVersion, %0
  br i1 %exitcond.loopVersion, label %for.inc11.loopexit38, label %for.body3.loopVersion

<Original Loop>
for.body3:                                        ; preds = %for.body3.lr.ph, %for.body3
  %indvars.iv = phi i64 [ %indvars.iv.next, %for.body3 ], [ %2, %for.body3.lr.ph ]
  %arrayidx = getelementptr inbounds i32* %var1, i64 %indvars.iv
  store i32 %add, i32* %arrayidx, align 4, !tbaa !1
  %8 = load i32* %arrayidx7, align 4, !tbaa !1
  %add8 = add nsw i32 %8, %add
  store i32 %add8, i32* %arrayidx7, align 4, !tbaa !1
  %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1
  %lftr.wideiv = trunc i64 %indvars.iv to i32
  %exitcond = icmp eq i32 %lftr.wideiv, %0
  br i1 %exitcond, label %for.inc11, label %for.body3

In versioned loop difference is visible, 1 store has moved out.

Following are some high level details about current implementation:

-  LoopVersioning
LoopVersioning is main class which holds multi versioning functionality.

- LoopVersioning :: isVersioningBeneficial
Its member to ‘LoopVersioning’
Does feasibility check for loop versioning.
a) Checks layout of loop.
b) Instruction level check.
c) memory checks.

- LoopVersioning :: versionizeLoop
a) Clone original loo
b) Create a runtime memory check.
c) Add both loops under runtime check results target.

- RuntimeMemoryCheck
This class take cares runtime memory check.

- RuntimeMemoryCheck ::createRuntimeCheck
It creates runtime memory check.

In this patch used maximum loop nest threshold as 2, and maximum number
of pointers in runtime memory check as 5.

Later I like to make this as a utility so others can use it.

Requesting to go through patch for detailed approach.
Patch available at http://reviews.llvm.org/D7900

Suggestions are comments are welcome.

LLVM Developers mailing list
LLVMdev at cs.uiuc.edu<mailto:LLVMdev at cs.uiuc.edu>         http://llvm.cs.uiuc.edu<http://llvm.cs.uiuc.edu/>

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20150303/b10feddf/attachment.html>

More information about the llvm-dev mailing list