[LLVMdev] Improving loop vectorizer support for loops with a volatile iteration variable

[Hyojin Sung]

Hi all,

I would like to propose an improvement of the “almost dead” block elimination in Transforms/Local.cpp so that it will preserve the canonical loop form for loops with a volatile iteration variable.

*** Problem statement
Nested loops in LCALS Subset B (https://codesign.llnl.gov/LCALS.php) are not vectorized with LLVM -O3 because the LLVM loop vectorizer fails the test whether the loop latch and exiting block of a loop is the same. The loops are vectorizable, and get vectorized with LLVM -O2 and also with other commercial compilers (icc, xlc).

*** Details
These loops ended up with different loop latch and exiting block after a series of optimizations including loop unswitching, jump threading, simplify-the-CFG, and loop simplify. The fundamental problem here is that the above optimizations cannot recognize a loop with a volatile iteration variable and do not preserve its canonical loop structure.

(1) Loop unswitching generates several empty placeholder BBs only with PHI nodes after separating out a shorter path with no inner loop execution from a standard path.

(2) Jump threading and simplify-the-CFG passes independently calls TryToSimplifyUnconditionalBranchFromEmptyBlock() in Transforms/Utils/Local.cpp to get rid of almost empty BBs.

(3) TryToSimplifyUnconditionalBranchFromEmtpyBlock() eliminates the placeholder BBs after loop unswitching and merges them into subsequent blocks including the header of the inner loop. Before eliminating the blocks, the function checks if the block is a loop header by looking at its PHI nodes so that it can be saved, but the test fails with the loops with a volatile iteration variable. The outer loop is now collapsed into the inner loop with multiple backedges.

(4) LoopSimplify checks if a loop with multiple backedges can be separated to nested loops by looking at PHI nodes in the loop header. The test fails with the loops with a volatile iteration variable.

(5) LoopSimplify then creates a unique backedge block for the loop, and the loop now has a different loop latch (the unique backedge block created in (3)) and exiting block (a block where the volatile outer loop variable is incremented and tested).  

*** Proposed solutions
(1) Make LoopInfo available in Jump Threading and Simplify-the-CFG passes and use LoopInfo to test whether an almost empty BB is a loop header. If yes, do not eliminate the BB.
      Pros: Leverages existing analysis results, small code changes / Cons: Add pass dependencies

(2) Instead of using LoopInfo, iterate through BB’s to identify backedges and loop headers in TryToSimplifyUnconditionalBranchFromEmptyBlock() and use the results to test if a BB is a loop header. If yes, do not eliminate the BB. Jump Threading has functions that do similar cheap loop analysis.
      Pros: No need to depend on external analysis results / Cons: More lines to be added

*** Summary
I would like to propose an improvement of the “almost dead” block elimination in Transforms/Local.cpp so that it will preserve the canonical loop form for loops with a volatile iteration variable. On top of the current algorithm relying on PHI nodes to recognize loops, actual loop analysis to test if a BB belongs to a loop and etc. can be used.

[Martin J. O'Riordan]

Regarding your proposed solution (2) - could this be further abstracted from just the volatile iteration variables so that all conditional dependencies between BBs could be tracked?  For VLIW architectures with predication support, it is often advantageous to either clone a BB predecessor (with appropriate predication) to each of the successor BBs, or to move a dependency from successor BBs into the predecessor BBs (with appropriate predication).  The approach you propose might help with this kind of ILP problem.  Currently LLVM does not seem to do a great job at tracking this kind of ILP data edge dependency across BBs.  I can achieve this now, but only by using a large amount of custom coding and I would prefer to use a more generalised abstraction.

 

            MartinO

[Chandler Carruth]

On Wed, Jul 15, 2015 at 12:55 PM Hyojin Sung <hs...@us.ibm.com> wrote:

Hi all,

I would like to propose an improvement of the “almost dead” block elimination in Transforms/Local.cpp so that it will preserve the canonical loop form for loops with a volatile iteration variable.

*** Problem statement
Nested loops in LCALS Subset B (https://codesign.llnl.gov/LCALS.php) are not vectorized with LLVM -O3 because the LLVM loop vectorizer fails the test whether the loop latch and exiting block of a loop is the same. The loops are vectorizable, and get vectorized with LLVM -O2

I would be interested to know why -O2 succeeds here.

 

and also with other commercial compilers (icc, xlc).

*** Details
These loops ended up with different loop latch and exiting block after a series of optimizations including loop unswitching, jump threading, simplify-the-CFG, and loop simplify. The fundamental problem here is that the above optimizations cannot recognize a loop with a volatile iteration variable and do not preserve its canonical loop structure.

Ok, meta-level question first:

Why do we care about performance of loops with a volatile iteration variable? That seems both counter-intuitive and unlikely to be a useful goal. We simply don't optimize volatile operations well in *any* part of the optimizer, and I'm not sure why we need to start trying to fix that. This seems like an irreparably broken benchmark, but perhaps there is a motivation I don't yet see.

Assuming that sufficient motivation arises to try to fix this, see my comments below:

 

(1) Loop unswitching generates several empty placeholder BBs only with PHI nodes after separating out a shorter path with no inner loop execution from a standard path.

(2) Jump threading and simplify-the-CFG passes independently calls TryToSimplifyUnconditionalBranchFromEmptyBlock() in Transforms/Utils/Local.cpp to get rid of almost empty BBs.

(3) TryToSimplifyUnconditionalBranchFromEmtpyBlock() eliminates the placeholder BBs after loop unswitching and merges them into subsequent blocks including the header of the inner loop. Before eliminating the blocks, the function checks if the block is a loop header by looking at its PHI nodes so that it can be saved, but the test fails with the loops with a volatile iteration variable.

Why does this fail for a volatile iteration variable but not for a non-volatile one? I think understanding that will be key to understanding how it should be fixed.

[Sean Silva]

On Wed, Jul 15, 2015 at 5:34 PM, Chandler Carruth <chan...@google.com> wrote:

On Wed, Jul 15, 2015 at 12:55 PM Hyojin Sung <hs...@us.ibm.com> wrote:

Hi all,

I would like to propose an improvement of the “almost dead” block elimination in Transforms/Local.cpp so that it will preserve the canonical loop form for loops with a volatile iteration variable.

*** Problem statement
Nested loops in LCALS Subset B (https://codesign.llnl.gov/LCALS.php) are not vectorized with LLVM -O3 because the LLVM loop vectorizer fails the test whether the loop latch and exiting block of a loop is the same. The loops are vectorizable, and get vectorized with LLVM -O2

I would be interested to know why -O2 succeeds here.

 

and also with other commercial compilers (icc, xlc).

*** Details
These loops ended up with different loop latch and exiting block after a series of optimizations including loop unswitching, jump threading, simplify-the-CFG, and loop simplify. The fundamental problem here is that the above optimizations cannot recognize a loop with a volatile iteration variable and do not preserve its canonical loop structure.

Ok, meta-level question first:

Why do we care about performance of loops with a volatile iteration variable? That seems both counter-intuitive and unlikely to be a useful goal. We simply don't optimize volatile operations well in *any* part of the optimizer, and I'm not sure why we need to start trying to fix that. This seems like an irreparably broken benchmark, but perhaps there is a motivation I don't yet see.

A quick look at the tarball on the linked site suggests that the volatile iteration variable is done on purpose so that the outer "run thing N times and take the average" loop can't be optimized.

-- Sean Silva

 

Assuming that sufficient motivation arises to try to fix this, see my comments below:

 

(1) Loop unswitching generates several empty placeholder BBs only with PHI nodes after separating out a shorter path with no inner loop execution from a standard path.

(2) Jump threading and simplify-the-CFG passes independently calls TryToSimplifyUnconditionalBranchFromEmptyBlock() in Transforms/Utils/Local.cpp to get rid of almost empty BBs.

(3) TryToSimplifyUnconditionalBranchFromEmtpyBlock() eliminates the placeholder BBs after loop unswitching and merges them into subsequent blocks including the header of the inner loop. Before eliminating the blocks, the function checks if the block is a loop header by looking at its PHI nodes so that it can be saved, but the test fails with the loops with a volatile iteration variable.

Why does this fail for a volatile iteration variable but not for a non-volatile one? I think understanding that will be key to understanding how it should be fixed.

_______________________________________________
LLVM Developers mailing list
LLV...@cs.uiuc.edu         http://llvm.cs.uiuc.edu
http://lists.cs.uiuc.edu/mailman/listinfo/llvmdev

[Hal Finkel]

From: "Chandler Carruth" <chan...@google.com>
To: "Hyojin Sung" <hs...@us.ibm.com>, llv...@cs.uiuc.edu
Sent: Wednesday, July 15, 2015 7:34:54 PM
Subject: Re: [LLVMdev] Improving loop vectorizer support for loops with a volatile iteration variable

On Wed, Jul 15, 2015 at 12:55 PM Hyojin Sung <hs...@us.ibm.com> wrote:

Hi all,

I would like to propose an improvement of the “almost dead” block elimination in Transforms/Local.cpp so that it will preserve the canonical loop form for loops with a volatile iteration variable.

*** Problem statement
Nested loops in LCALS Subset B (https://codesign.llnl.gov/LCALS.php) are not vectorized with LLVM -O3 because the LLVM loop vectorizer fails the test whether the loop latch and exiting block of a loop is the same. The loops are vectorizable, and get vectorized with LLVM -O2

I would be interested to know why -O2 succeeds here.

 

and also with other commercial compilers (icc, xlc).

*** Details
These loops ended up with different loop latch and exiting block after a series of optimizations including loop unswitching, jump threading, simplify-the-CFG, and loop simplify. The fundamental problem here is that the above optimizations cannot recognize a loop with a volatile iteration variable and do not preserve its canonical loop structure.

Ok, meta-level question first:

Why do we care about performance of loops with a volatile iteration variable?

I don't think we do, however, I think that misses the point. In this case, the volatile iteration variable is just an easy way to expose this problem that we have with nested loop canonicalization and the vectorizer. To be specific:

This we vectorizer just fine:

void foo1(float * restrict x, float * restrict y, float * restrict z) {
  for (int i = 0; i < 1000; ++i) {
    for (int j = 0; j < 1600; ++j) {
      x[j] = y[j] + z[j];
    }
  }
}

And, indeed, this we don't (the only change is adding volatile on i):

void foo2(float * restrict x, float * restrict y, float * restrict z) {
  for (volatile int i = 0; i < 1000; ++i) {
    for (int j = 0; j < 1600; ++j) {
      x[j] = y[j] + z[j];
    }
  }
}

However, this we don't either, and that's a big problem:

int done(float *x, float *y, float *z);
void foo3(float * restrict x, float * restrict y, float * restrict z) {
  while (!done(x, y, z)) {
    for (int j = 0; j < 1600; ++j) {
      x[j] = y[j] + z[j];
    }
  }
}

And the underlying reason is the same. The IR at the point in time when the loop vectorizer runs looks like this:

define void @foo3(float* noalias %x, float* noalias %y, float* noalias %z) #0 {
entry:
  %call.14 = tail call i32 @done(float* %x, float* %y, float* %z) #1
  %lnot.15 = icmp eq i32 %call.14, 0
  br i1 %lnot.15, label %for.body.preheader, label %while.end

for.body.preheader:                               ; preds = %entry
  br label %for.body

while.cond.loopexit:                              ; preds = %for.body
  %call = tail call i32 @done(float* %x, float* %y, float* %z) #1
  %lnot = icmp eq i32 %call, 0
  br i1 %lnot, label %for.body.backedge, label %while.end.loopexit

for.body:                                         ; preds = %for.body.backedge, %for.body.preheader
  %indvars.iv = phi i64 [ 0, %for.body.preheader ], [ %indvars.iv.be, %for.body.backedge ]
  ...
  %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1
  %exitcond = icmp eq i64 %indvars.iv.next, 1600
  br i1 %exitcond, label %while.cond.loopexit, label %for.body.backedge

for.body.backedge:                                ; preds = %for.body, %while.cond.loopexit
  %indvars.iv.be = phi i64 [ %indvars.iv.next, %for.body ], [ 0, %while.cond.loopexit ]
  br label %for.body

while.end.loopexit:                               ; preds = %while.cond.loopexit
  br label %while.end

while.end:                                        ; preds = %while.end.loopexit, %entry
  ret void
}

and we can currently only vectorize loops where the loop latch is also the loop's exiting block. In this case, as in the case with the volatile variable, vectorization is blocked by this constraint (here the backedge is from the terminator of %for.body.backedge, but the loop exiting block is %for.body). The check in the vectorizer is explicit:

  // We only handle bottom-tested loops, i.e. loop in which the condition is
  // checked at the end of each iteration. With that we can assume that all
  // instructions in the loop are executed the same number of times.
  if (TheLoop->getExitingBlock() != TheLoop->getLoopLatch()) {
    ...

 -Hal

That seems both counter-intuitive and unlikely to be a useful goal. We simply don't optimize volatile operations well in *any* part of the optimizer, and I'm not sure why we need to start trying to fix that. This seems like an irreparably broken benchmark, but perhaps there is a motivation I don't yet see.

Assuming that sufficient motivation arises to try to fix this, see my comments below:

 

(1) Loop unswitching generates several empty placeholder BBs only with PHI nodes after separating out a shorter path with no inner loop execution from a standard path.

(2) Jump threading and simplify-the-CFG passes independently calls TryToSimplifyUnconditionalBranchFromEmptyBlock() in Transforms/Utils/Local.cpp to get rid of almost empty BBs.

(3) TryToSimplifyUnconditionalBranchFromEmtpyBlock() eliminates the placeholder BBs after loop unswitching and merges them into subsequent blocks including the header of the inner loop. Before eliminating the blocks, the function checks if the block is a loop header by looking at its PHI nodes so that it can be saved, but the test fails with the loops with a volatile iteration variable.

Why does this fail for a volatile iteration variable but not for a non-volatile one? I think understanding that will be key to understanding how it should be fixed.

_______________________________________________
LLVM Developers mailing list
LLV...@cs.uiuc.edu         http://llvm.cs.uiuc.edu
http://lists.cs.uiuc.edu/mailman/listinfo/llvmdev

--

Hal Finkel
Assistant Computational Scientist
Leadership Computing Facility
Argonne National Laboratory

[Chandler Carruth]

Thanks for the detailed explanation. This makes much more sense why we need to handle it. I think its much better to look at nested loops of this form than anything to do with volatile -- the latter is too prone to other random optimizations turning off.

Regarding this problem, it would be interesting to know based on this explanation what the desired fix would be. I see at least these options:

1) Canonicalize loops harder to make them look the way the vectorizer wants. If this can be done without causing significant problems, it seems likely the best approach.

2) Teach the vectorizer to vectorize without this constraint by instead establishing the actual invariant it cares about.

Maybe there is another strategy? 

[Hal Finkel]

I agree. In this case, we could certainly fold the trivial %for.body.backedge block into %for.body, meaning transforming this:

for.body.backedge:                                ; preds = %while.cond.loopexit, %for.body

  %indvars.iv.be = phi i64 [ %indvars.iv.next, %for.body ], [ 0, %while.cond.loopexit ]
  br label %for.body

for.body:                                         ; preds = %for.body.backedge, %for.body.preheader
  %indvars.iv = phi i64 [ 0, %for.body.preheader ], [ %indvars.iv.be, %for.body.backedge ]
  ...
  %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1
  %exitcond = icmp eq i64 %indvars.iv.next, 1600
  br i1 %exitcond, label %while.cond.loopexit, label %for.body.backedge

into this:

for.body:                                         ; preds = %for.body.backedge, %for.body.preheader

  %indvars.iv.be = phi i64 [ %indvars.iv.next, %for.body ], [ 0, %while.cond.loopexit ]

  %indvars.iv = phi i64 [ 0, %for.body.preheader ], [ %indvars.iv.be, %for.body ]

  ...
  %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1
  %exitcond = icmp eq i64 %indvars.iv.next, 1600
  br i1 %exitcond, label %while.cond.loopexit, label %for.body

and this seems pretty trivial when %for.body.backedge is completely empty (as in this case), but if it had non-PHI instructions in it on which the existing PHIs in %for.body depended, then maybe this becomes less trivial?

2) Teach the vectorizer to vectorize without this constraint by instead establishing the actual invariant it cares about.

It really cares that there's no code that comes in between the latch and the exit, because such code is not really part of the loop (it only runs once), or code in between the exit and the latch (because such code runs in one fewer iterations than the code before the exit). At least nothing with side effects I presume.

 -Hal

[Hal Finkel]

Although, based on what I said below, the case with instructions there we can't currently vectorize anyway for more-fundamental reasons.

 -Hal

[Hal Finkel]

Also worth pointing out that SimplifyCFG does this exact transformation. The vectorizer never sees it, however, because LoopSimplify prefers this form with the separate backedge block that the vectorizer can't handle.

[Hyojin Sung]

Chandler Carruth ---07/15/2015 08:35:09 PM---On Wed, Jul 15, 2015 at 12:55 PM Hyojin Sung <hs...@us.ibm.com> wrote: > Hi all,

From: Chandler Carruth <chan...@google.com>
To: Hyojin Sung/Watson/IBM@IBMUS, llv...@cs.uiuc.edu
Date: 07/15/2015 08:35 PM
Subject: Re: [LLVMdev] Improving loop vectorizer support for loops with a volatile iteration variable

---

On Wed, Jul 15, 2015 at 12:55 PM Hyojin Sung <hs...@us.ibm.com> wrote:

Hi all,

I would like to propose an improvement of the “almost dead” block elimination in Transforms/Local.cpp so that it will preserve the canonical loop form for loops with a volatile iteration variable.

*** Problem statement
Nested loops in LCALS Subset B (https://codesign.llnl.gov/LCALS.php) are not vectorized with LLVM -O3 because the LLVM loop vectorizer fails the test whether the loop latch and exiting block of a loop is the same. The loops are vectorizable, and get vectorized with LLVM -O2

I would be interested to know why -O2 succeeds here.

-O2 does not perform loop unswitching which creates artificial empty placeholder blocks in the outer loop. As long as incrementing and testing the volatile iteration variable is kept only in the original BB, the block does not get eliminated.  

[Hyojin Sung]

Hi,

I discussed the issue offline with Hal, and would like to clarify what is exactly going on, what are trade-offs for different solutions, and ask for more feedback on my proposed solution (http://reviews.llvm.org/D11728). I will use the example from Hal's post:

void foo2(float * restrict x, float * restrict y, float * restrict z) {
  for (volatile int i = 0; i < 1000; ++i) {
    for (int j = 0; j < 1600; ++j) {
      x[j] = y[j] + z[j];
    }
  }
}

IR after the first loop simplify: A preheader is created.

; Function Attrs: nounwind
define void @foo2(float* noalias nocapture %x, float* noalias nocapture readonly %y, float* noalias nocapture readonly %z) #0 {
entry:
  %i = alloca i32, align 4
  tail call void @llvm.dbg.value(metadata float* %x, i64 0, metadata !11, metadata !25), !dbg !26
  tail call void @llvm.dbg.value(metadata float* %y, i64 0, metadata !12, metadata !25), !dbg !27
  tail call void @llvm.dbg.value(metadata float* %z, i64 0, metadata !13, metadata !25), !dbg !28
  %i.0.i.0..sroa_cast = bitcast i32* %i to i8*
  call void @llvm.lifetime.start(i64 4, i8* %i.0.i.0..sroa_cast)
  tail call void @llvm.dbg.value(metadata i32 0, i64 0, metadata !14, metadata !25), !dbg !29
  store volatile i32 0, i32* %i, align 4, !dbg !29
  br label %for.cond, !dbg !30

for.cond:                                         ; preds = %for.cond.cleanup.3, %entry
  tail call void @llvm.dbg.value(metadata i32* %i, i64 0, metadata !14, metadata !25), !dbg !29
  %i.0.i.0. = load volatile i32, i32* %i, align 4, !dbg !31
  %cmp = icmp slt i32 %i.0.i.0., 1000, !dbg !34
  br i1 %cmp, label %for.cond.1.preheader, label %for.cond.cleanup, !dbg !35

for.cond.1.preheader:                             ; preds = %for.cond
  br label %for.cond.1, !dbg !36

for.cond.cleanup:                                 ; preds = %for.cond
  call void @llvm.lifetime.end(i64 4, i8* %i.0.i.0..sroa_cast)
  ret void, !dbg !38

for.cond.1:                                       ; preds = %for.cond.1.preheader, %for.body.4
  %j.0 = phi i32 [ %inc, %for.body.4 ], [ 0, %for.cond.1.preheader ]
  %cmp2 = icmp slt i32 %j.0, 1600, !dbg !36
  br i1 %cmp2, label %for.body.4, label %for.cond.cleanup.3, !dbg !39

for.cond.cleanup.3:                               ; preds = %for.cond.1
  tail call void @llvm.dbg.value(metadata i32* %i, i64 0, metadata !14, metadata !25), !dbg !29
  %i.0.i.0.17 = load volatile i32, i32* %i, align 4, !dbg !40
  %inc10 = add nsw i32 %i.0.i.0.17, 1, !dbg !40
  tail call void @llvm.dbg.value(metadata i32 %inc10, i64 0, metadata !14, metadata !25), !dbg !29
  store volatile i32 %inc10, i32* %i, align 4, !dbg !40
  br label %for.cond, !dbg !41

for.body.4:                                       ; preds = %for.cond.1
  %idxprom = sext i32 %j.0 to i64, !dbg !42
  %arrayidx = getelementptr inbounds float, float* %y, i64 %idxprom, !dbg !42
  %0 = load float, float* %arrayidx, align 4, !dbg !42, !tbaa !44
  %arrayidx6 = getelementptr inbounds float, float* %z, i64 %idxprom, !dbg !48
  %1 = load float, float* %arrayidx6, align 4, !dbg !48, !tbaa !44
  %add = fadd float %0, %1, !dbg !49
  %arrayidx8 = getelementptr inbounds float, float* %x, i64 %idxprom, !dbg !50
  store float %add, float* %arrayidx8, align 4, !dbg !51, !tbaa !44
  %inc = add nsw i32 %j.0, 1, !dbg !52
  tail call void @llvm.dbg.value(metadata i32 %inc, i64 0, metadata !18, metadata !25), !dbg !53
  br label %for.cond.1, !dbg !54
}

IR after loop rotation: After loop rotation, a rotated preheader (for.cond.1.preheader.lr.ph) is created. A test for (i < 1000) is added at the end of "entry" block. If true, the control jumps unconditionally to "for.body.4" through "for.cond.1.preheader.lr.ph" and "for.cond.1.preheader". You can see that these two blocks ("for.cond.1.preheader.lr.ph" and "for.cond.1.preheader") are practically empty, and they will get eliminated later by Jump Threading and/or Simplify-the-CFG. *IF* the outer loop has a non-volatile induction variable, the loop will not be rotated in the first place as "for.cond.1.preheader" has a PHI node for "i", and these blocks will not be eliminated.

; Function Attrs: nounwind
define void @foo2(float* noalias nocapture %x, float* noalias nocapture readonly %y, float* noalias nocapture readonly %z) #0 {
entry:
  %i = alloca i32, align 4
  tail call void @llvm.dbg.value(metadata float* %x, i64 0, metadata !11, metadata !25), !dbg !26
  tail call void @llvm.dbg.value(metadata float* %y, i64 0, metadata !12, metadata !25), !dbg !27
  tail call void @llvm.dbg.value(metadata float* %z, i64 0, metadata !13, metadata !25), !dbg !28
  %i.0.i.0..sroa_cast = bitcast i32* %i to i8*
  call void @llvm.lifetime.start(i64 4, i8* %i.0.i.0..sroa_cast)
  tail call void @llvm.dbg.value(metadata i32 0, i64 0, metadata !14, metadata !25), !dbg !29
  store volatile i32 0, i32* %i, align 4, !dbg !29
  tail call void @llvm.dbg.value(metadata i32* %i, i64 0, metadata !14, metadata !25), !dbg !29
  %i.0.i.0..21 = load volatile i32, i32* %i, align 4, !dbg !30
  %cmp.22 = icmp slt i32 %i.0.i.0..21, 1000, !dbg !33
  br i1 %cmp.22, label %for.cond.1.preheader.lr.ph, label %for.cond.cleanup, !dbg !34

for.cond.1.preheader.lr.ph:                       ; preds = %entry
  br label %for.cond.1.preheader, !dbg !34

for.cond.1.preheader:                             ; preds = %for.cond.1.preheader.lr.ph, %for.cond.cleanup.3
  br label %for.body.4, !dbg !35

for.cond.for.cond.cleanup_crit_edge:              ; preds = %for.cond.cleanup.3
  br label %for.cond.cleanup, !dbg !34

for.cond.cleanup:                                 ; preds = %for.cond.for.cond.cleanup_crit_edge, %entry
  call void @llvm.lifetime.end(i64 4, i8* %i.0.i.0..sroa_cast)
  ret void, !dbg !36

for.cond.cleanup.3:                               ; preds = %for.body.4
  tail call void @llvm.dbg.value(metadata i32* %i, i64 0, metadata !14, metadata !25), !dbg !29
  %i.0.i.0.17 = load volatile i32, i32* %i, align 4, !dbg !37
  %inc10 = add nsw i32 %i.0.i.0.17, 1, !dbg !37
  tail call void @llvm.dbg.value(metadata i32 %inc10, i64 0, metadata !14, metadata !25), !dbg !29
  store volatile i32 %inc10, i32* %i, align 4, !dbg !37
  tail call void @llvm.dbg.value(metadata i32* %i, i64 0, metadata !14, metadata !25), !dbg !29
  %i.0.i.0. = load volatile i32, i32* %i, align 4, !dbg !30
  %cmp = icmp slt i32 %i.0.i.0., 1000, !dbg !33
  br i1 %cmp, label %for.cond.1.preheader, label %for.cond.for.cond.cleanup_crit_edge, !dbg !34

for.body.4:                                       ; preds = %for.cond.1.preheader, %for.body.4
  %j.020 = phi i32 [ 0, %for.cond.1.preheader ], [ %inc, %for.body.4 ]
  %idxprom = sext i32 %j.020 to i64, !dbg !38
  %arrayidx = getelementptr inbounds float, float* %y, i64 %idxprom, !dbg !38
  %0 = load float, float* %arrayidx, align 4, !dbg !38, !tbaa !41
  %arrayidx6 = getelementptr inbounds float, float* %z, i64 %idxprom, !dbg !45
  %1 = load float, float* %arrayidx6, align 4, !dbg !45, !tbaa !41
  %add = fadd float %0, %1, !dbg !46
  %arrayidx8 = getelementptr inbounds float, float* %x, i64 %idxprom, !dbg !47
  store float %add, float* %arrayidx8, align 4, !dbg !48, !tbaa !41
  %inc = add nsw i32 %j.020, 1, !dbg !49
  tail call void @llvm.dbg.value(metadata i32 %inc, i64 0, metadata !18, metadata !25), !dbg !50
  %cmp2 = icmp slt i32 %inc, 1600, !dbg !51
  br i1 %cmp2, label %for.body.4, label %for.cond.cleanup.3, !dbg !35


After Jump Threading: "for.cond.1.preheader.lr.ph" and "for.cond.1.preheader" are merged into "for.body.4" by TryToSimplifyUnconditionalBranchFromEmptyBlock() in Transforms/Utils/Local.cpp. Now "for.body.4" has three incoming edges (two backedges).

; Function Attrs: nounwind
define void @foo2(float* noalias nocapture %x, float* noalias nocapture readonly %y, float* noalias nocapture readonly %z) #0 {
entry:
  %i = alloca i32, align 4
  tail call void @llvm.dbg.value(metadata float* %x, i64 0, metadata !11, metadata !25), !dbg !26
  tail call void @llvm.dbg.value(metadata float* %y, i64 0, metadata !12, metadata !25), !dbg !27
  tail call void @llvm.dbg.value(metadata float* %z, i64 0, metadata !13, metadata !25), !dbg !28
  %i.0.i.0..sroa_cast = bitcast i32* %i to i8*
  call void @llvm.lifetime.start(i64 4, i8* %i.0.i.0..sroa_cast)
  tail call void @llvm.dbg.value(metadata i32 0, i64 0, metadata !14, metadata !25), !dbg !29
  store volatile i32 0, i32* %i, align 4, !dbg !29
  tail call void @llvm.dbg.value(metadata i32* %i, i64 0, metadata !14, metadata !25), !dbg !29
  %i.0.i.0..21 = load volatile i32, i32* %i, align 4, !dbg !30
  %cmp.22 = icmp slt i32 %i.0.i.0..21, 1000, !dbg !33
  br i1 %cmp.22, label %for.body.4, label %for.cond.cleanup, !dbg !34

for.cond.cleanup:                                 ; preds = %for.cond.cleanup.3, %entry
  call void @llvm.lifetime.end(i64 4, i8* %i.0.i.0..sroa_cast)
  ret void, !dbg !35

for.cond.cleanup.3:                               ; preds = %for.body.4
  tail call void @llvm.dbg.value(metadata i32* %i, i64 0, metadata !14, metadata !25), !dbg !29
  %i.0.i.0.17 = load volatile i32, i32* %i, align 4, !dbg !36
  %inc10 = add nsw i32 %i.0.i.0.17, 1, !dbg !36
  tail call void @llvm.dbg.value(metadata i32 %inc10, i64 0, metadata !14, metadata !25), !dbg !29
  store volatile i32 %inc10, i32* %i, align 4, !dbg !36
  tail call void @llvm.dbg.value(metadata i32* %i, i64 0, metadata !14, metadata !25), !dbg !29
  %i.0.i.0. = load volatile i32, i32* %i, align 4, !dbg !30
  %cmp = icmp slt i32 %i.0.i.0., 1000, !dbg !33
  br i1 %cmp, label %for.body.4, label %for.cond.cleanup, !dbg !34

for.body.4:                                       ; preds = %for.cond.cleanup.3, %entry, %for.body.4
  %indvars.iv = phi i64 [ %indvars.iv.next, %for.body.4 ], [ 0, %entry ], [ 0, %for.cond.cleanup.3 ]
  %arrayidx = getelementptr inbounds float, float* %y, i64 %indvars.iv, !dbg !37
  %0 = load float, float* %arrayidx, align 4, !dbg !37, !tbaa !40
  %arrayidx6 = getelementptr inbounds float, float* %z, i64 %indvars.iv, !dbg !44
  %1 = load float, float* %arrayidx6, align 4, !dbg !44, !tbaa !40
  %add = fadd float %0, %1, !dbg !45
  %arrayidx8 = getelementptr inbounds float, float* %x, i64 %indvars.iv, !dbg !46
  store float %add, float* %arrayidx8, align 4, !dbg !47, !tbaa !40
  %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1, !dbg !48
  %exitcond = icmp eq i64 %indvars.iv.next, 1600, !dbg !48
  br i1 %exitcond, label %for.cond.cleanup.3, label %for.body.4, !dbg !48


After another loop simplify: Loop simplify tries to separate out nested loops but fails to do so with this loop since it does not has a PHI node for the outer loop variable. Instead, it creates a backedge block.

for.cond.cleanup.3:                               ; preds = %for.body.4
  tail call void @llvm.dbg.value(metadata i32* %i, i64 0, metadata !14, metadata !25), !dbg !29
  %i.0.i.0.17 = load volatile i32, i32* %i, align 4, !dbg !39
  %inc10 = add nsw i32 %i.0.i.0.17, 1, !dbg !39
  tail call void @llvm.dbg.value(metadata i32 %inc10, i64 0, metadata !14, metadata !25), !dbg !29
  store volatile i32 %inc10, i32* %i, align 4, !dbg !39
  tail call void @llvm.dbg.value(metadata i32* %i, i64 0, metadata !14, metadata !25), !dbg !29
  %i.0.i.0. = load volatile i32, i32* %i, align 4, !dbg !30
  %cmp = icmp slt i32 %i.0.i.0., 1000, !dbg !33
  br i1 %cmp, label %for.body.4.backedge, label %for.cond.cleanup.loopexit, !dbg !34

for.body.4:                                       ; preds = %for.body.4.backedge, %for.body.4.preheader
  %indvars.iv = phi i64 [ 0, %for.body.4.preheader ], [ %indvars.iv.be, %for.body.4.backedge ]
  %arrayidx = getelementptr inbounds float, float* %y, i64 %indvars.iv, !dbg !35
  %0 = load float, float* %arrayidx, align 4, !dbg !35, !tbaa !40
  %arrayidx6 = getelementptr inbounds float, float* %z, i64 %indvars.iv, !dbg !44
  %1 = load float, float* %arrayidx6, align 4, !dbg !44, !tbaa !40
  %add = fadd float %0, %1, !dbg !45
  %arrayidx8 = getelementptr inbounds float, float* %x, i64 %indvars.iv, !dbg !46
  store float %add, float* %arrayidx8, align 4, !dbg !47, !tbaa !40
  %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1, !dbg !48
  %exitcond = icmp eq i64 %indvars.iv.next, 1600, !dbg !48
  br i1 %exitcond, label %for.cond.cleanup.3, label %for.body.4.backedge, !dbg !48

for.body.4.backedge:                              ; preds = %for.body.4, %for.cond.cleanup.3
  %indvars.iv.be = phi i64 [ %indvars.iv.next, %for.body.4 ], [ 0, %for.cond.cleanup.3 ]
  br label %for.body.4x


LLVM  loop vectorizer rejects to vectorize any loop for which a loop latch (for.body.4.backedge) is different from a loop exiting block (for.cond.cleanup.3). The loop vectorizer can assume that all instructions in the loop are executed the same number of times with the test.

I believe a fix is in order in one way or another because the example is simple and common enough and vectorized by other compilers. We may approach it by either (1) preventing loops from being collapsed in the first place or (2) teaching loop vectorizer to handle collapsed loops. For (2), we may need to allow loop vectorizer to forego the assumption and handle the loop as it is. The assumption seems fundamental to many of the vectorization algorithms, so it will require extensive updates or may end up with reverting the loop back to a properly nested form. The downside of (1) is that it may slow down common optimization passes that are repeatedly executed before vectorization.  

My patch (http://reviews.llvm.org/D11728) is a prototype fix for (1) that modifies Jump Threading and Simplify-the-CFG to not eliminate an empty loop header BB even when the loop does not have a PHI node for its induction variable. The details can be found at http://reviews.llvm.org/D11728. I would welcome and appreciate any comments or feedback.

Best,
Hyojin


Hal Finkel ---07/16/2015 03:19:24 AM-------- Original Message ----- > From: "Hal Finkel" <hfi...@anl.gov>

Argonne National Laboratory_______________________________________________

LLVM Developers mailing list
LLV...@cs.uiuc.edu        

http://llvm.cs.uiuc.edu
http://lists.cs.uiuc.edu/mailman/listinfo/llvmdev

[Gerolf Hoflehner]

I see a more fundamental problem and perhaps this example can serve as a stepping stone towards a solution.

There is a desired property: In this case, loop is vectorizable.

There are N compiler transformations. These transformations must either establish the property or keep it invariant,  but never destroy it.

If there is agreement to this then the first step is to have the analysis of ‘is vectorizable’ runnable after every transformation and report violations in detail, likely under a flag. Then comparing a set of loops not vectorizable with clang but with other compilers (gcc, clang, …) should shape precise ideas on normal forms for the vectorizer and general improvements to transformations to the keep/establish the invariant/desired property.

I’m worried that the one example at time approach over time confuscates matters within transformations resulting in less maintainable code. 

Gerolf

<graycol.gif>Hal Finkel ---07/16/2015 03:19:24 AM-------- Original Message ----- > From: "Hal Finkel" <hfi...@anl.gov>

[Gerolf Hoflehner via llvm-dev]

On Aug 5, 2015, at 9:08 PM, Gerolf Hoflehner <ghofl...@apple.com> wrote:

Ok, assuming a verifier/checker was in place the information could look like:

...

LV: Not vectorizing: Cannot prove legality.

LV: Legal to vectorize. — after loop rotation

LV: Legal to vectorize.

...

LV: Legal to vectorize.

LV: Not vectorizing: Cannot prove legality. — after Jump Threading (later also — after CFG Simplify)

Note that in your example the second round of loop rotation fails making a loop vectorizable (from the comment that is its purpose, though!). It might be more appropriate to generalize that approach to any transformation that makes a loop vectorizable (not just rotation) rather than tweaking jump threading + simplification.

I’m convinced there are more cases like yours and the following framework would give a systematic way to handle them:

a) a checker for is_vectorizable (from refactoring the legal class in the loop vectorizer - but properly, not hacked like my prototype …) 

b) invoke checker optionally. The checker can be used for testing, analysis and - independently - as a service for loop transformations (eg. if the loop is vectorizable already, perhaps a transformation is not needed or vice versa). I’m not sure about the architecture yet. Perhaps the pass manager could run a checker automatically after every pass when an option is set. Adding extra calls after every pass is too clumsy.

c) fix violators if appropriate and/or add transformations that morph a loop into a vectorizable form when possible (especially when it has been in vectorizable shape at some point).

-Gerolf

On Aug 4, 2015, at 8:04 PM, Gerolf Hoflehner <ghofl...@apple.com> wrote:

The approach would be to run the legal check before + after transformations on demand. For loops that “should” be vectorized it would give the means to isolate the transformation that introduces a barrier. I’ll see that can get a dirty prototype to illustrate this on your test case.  I could give a warning/remark when the code was (loop-) vectorizable at some point but not anymore when the vectorizer actually kicks in and speed-up diagnosis when comparing good/bad cases (like in your instance w/ and w/o the volatile modifier).

Gerolf

On Aug 4, 2015, at 8:40 AM, Hyojin Sung <hs...@us.ibm.com> wrote:

Hi Gerolf,

Thanks for the comment. Yes, I agree that compiler transformation as a whole should not render vectorizable loops unvectorizable. However, some passes may temporarily make a loop unvectorizable (e.g., by aggressive GVN or DCE/DBE) then later passes may revert the effect. Trying to keep the invariants after each pass may prevent useful optimizations. Regarding your concern about the example-based approach, LoopVectorizationLegality class gives a good idea about what will be such invariants and they are mostly related to check for canonical loop structures. I think that there are only a handful of transformation passes that may disrupt loop structures mainly through aggressively eliminating BB's. Focusing on such transformations may be sufficient to prevent prior transformations from destroying vectorizable loops. I'd be glad to hear more of your thoughts.

Best,
Hyojin


<graycol.gif>Gerolf Hoflehner ---08/03/2015 06:14:22 PM---I see a more fundamental problem and perhaps this example can serve as a stepping stone towards a so

From: Gerolf Hoflehner <ghofl...@apple.com>
To: Hyojin Sung/Watson/IBM@IBMUS
Cc: Hal Finkel <hfi...@anl.gov>, llvmdev...@cs.uiuc.edu, llv...@cs.uiuc.edu
Date: 08/03/2015 06:14 PM
Subject: Re: [LLVMdev] Improving loop vectorizer support for loops with a volatile iteration variable

[Hyojin Sung via llvm-dev]

Inlined.

Gerolf Hoflehner ---08/06/2015 12:11:45 AM---> On Aug 5, 2015, at 9:08 PM, Gerolf Hoflehner <ghofl...@apple.com> wrote: >

From: Gerolf Hoflehner <ghofl...@apple.com>
To: Hyojin Sung/Watson/IBM@IBMUS
Cc: Hal Finkel <hfi...@anl.gov>, llvm...@lists.llvm.org
Date: 08/06/2015 12:11 AM
Subject: Re: [LLVMdev] Improving loop vectorizer support for loops with a volatile iteration variable

---

On Aug 5, 2015, at 9:08 PM, Gerolf Hoflehner <ghofl...@apple.com> wrote:

Ok, assuming a verifier/checker was in place the information could look like:

...
LV: Not vectorizing: Cannot prove legality.
LV: Legal to vectorize. — after loop rotation
LV: Legal to vectorize.
...
LV: Legal to vectorize.
LV: Not vectorizing: Cannot prove legality. — after Jump Threading (later also — after CFG Simplify)

Note that in your example the second round of loop rotation fails making a loop vectorizable (from the comment that is its purpose, though!). It might be more appropriate to generalize that approach to any transformation that makes a loop vectorizable (not just rotation) rather than tweaking jump threading + simplification.

To clarify, the loop is vectorizable before and after loop rotation. At least for the example, loop rotation does not affect the vectorizability of a loop. The LLVM loop vectorizer and vectorization legality checker rely on loops in a canonical form. So if there's no interfering passes, a check should be able to tell if a loop is vectorizable after loop simplify.

I’m convinced there are more cases like yours and the following framework would give a systematic way to handle them:

a) a checker for is_vectorizable (from refactoring the legal class in the loop vectorizer - but properly, not hacked like my prototype …)
b) invoke checker optionally. The checker can be used for testing, analysis and - independently - as a service for loop transformations (eg. if the loop is vectorizable already, perhaps a transformation is not needed or vice versa). I’m not sure about the architecture yet. Perhaps the pass manager could run a checker automatically after every pass when an option is set. Adding extra calls after every pass is too clumsy.

c) fix violators if appropriate and/or add transformations that morph a loop into a vectorizable form when possible (especially when it has been in vectorizable shape at some point).

We can design a more general form of checker/verifier that is optionally executed after LLVM passes can help with identifying interfering transformations. We can use the checker only for debugging (reporting which transformation is interfering with others) or for dynamically triggering or reverting transformations to maintain desired properties like vectorizability as you said. If the checker dynamically affects actual pass executions, then I think it can require extensive changes to the incremental way LLVM applies transformations.


Regards,
Hyojin

[Hal Finkel via llvm-dev]

Hi Gerolf,

I think we have several (perhaps separable) issues here:

1. Do we have a canonical form for loops, preserved through the optimizer, that allows naturally-constructed loop nests to remain separable?

2. Do we forbid non-lowering transformations that turn vectorizable loops into non-vectorizable loops?

3. How do we detect cases where transformations cause a negative answer to either of the above?

Now (2) is a subset of (1), but (1) applies as a general structural property to a much larger class of transformations. Hyojin has found, and proposed a fix for, two important instances of (1) by manual investigation, and I think we should move forward with reviewing those (the actual Phabricator differential did not cc llvm-commits, etc. and I've commented there asking her to open a new one).

Regarding a more-automatic system for finding these kinds of problems, I'm certainly in favor. We could certainly run the vectorizer's legality check, add metadata to loops found to be vectorizable (or use some other state-keeping mechanism), and then check again later that those loops are still considered legally vectorizable (we could also check that we never move from a profitable to unprofitable state). I suspect this would be much too expensive to run aggressively by default, but we could certainly have a mode for us to use that did this.

Part of the challenge, however, is that loops are often taken in and our of 'LoopSimplify' form, and we normally modify the IR in LoopSimplify before the vectorizer's analysis is runs (and, similarly before other passes that operate on loops). As a result, we (might) need to be judicious regarding where to do these checks (maybe we should do them only after LoopSimplify runs?).

Regardless, I think that we should work on documenting what our canonical loop form is, and specifically include restrictions necessary to preserve (1). For (2), this would be harder (what the vectorizer will or will not vectorize will likely change much more over time), but detecting problems here is certainly also very important.

Thanks again,
Hal

----- Original Message -----
> From: "Gerolf Hoflehner" <ghofl...@apple.com>
> To: "Hyojin Sung" <hs...@us.ibm.com>
> Cc: "Hal Finkel" <hfi...@anl.gov>, llv...@cs.uiuc.edu, llvmdev...@cs.uiuc.edu
> Sent: Wednesday, August 5, 2015 11:08:11 PM
> Subject: Re: [LLVMdev] Improving loop vectorizer support for loops with a volatile iteration variable
>

> Ok, assuming a verifier/checker was in place the information could
> look like:
>
>
>
> ...
> LV: Not vectorizing: Cannot prove legality.
> LV: Legal to vectorize. — after loop rotation
> LV: Legal to vectorize.
> ...
> LV: Legal to vectorize.
> LV: Not vectorizing: Cannot prove legality. — after Jump Threading
> (later also — after CFG Simplify)
>
>
> Note that in your example the second round of loop rotation fails
> making a loop vectorizable (from the comment that is its purpose,
> though!). It might be more appropriate to generalize that approach
> to any transformation that makes a loop vectorizable (not just
> rotation) rather than tweaking jump threading + simplification.
>
>

> I’m convinced there are more cases like yours and the following
> framework would give a systematic way to handle them:
>
>
> a) a checker for is_vectorizable (from refactoring the legal class in
> the loop vectorizer - but properly, not hacked like my prototype …)
> b) invoke checker optionally. The checker can be used for testing,
> analysis and - independently - as a service for loop transformations
> (eg. if the loop is vectorizable already, perhaps a transformation
> is not needed or vice versa). I’m not sure about the architecture
> yet. Perhaps the pass manager could run a checker automatically
> after every pass when an option is set. Adding extra calls after
> every pass is too clumsy.
> c) fix violators if appropriate and/or add transformations that morph
> a loop into a vectorizable form when possible (especially when it
> has been in vectorizable shape at some point).
>
>
>
>

> I see a more fundamental problem and perhaps this example can serve
> as a stepping stone towards a solution.
>
> There is a desired property: In this case, loop is vectorizable.
> There are N compiler transformations. These transformations must
> either establish the property or keep it invariant, but never
> destroy it.
>
> If there is agreement to this then the first step is to have the
> analysis of ‘is vectorizable’ runnable after every transformation
> and report violations in detail, likely under a flag. Then comparing
> a set of loops not vectorizable with clang but with other compilers
> (gcc, clang, …) should shape precise ideas on normal forms for the
> vectorizer and general improvements to transformations to the
> keep/establish the invariant/desired property.
>
> I’m worried that the one example at time approach over time
> confuscates matters within transformations resulting in less
> maintainable code.
>
> Gerolf
>
>
>
>
> On Aug 3, 2015, at 2:33 PM, Hyojin Sung < hs...@us.ibm.com > wrote:
>
> Hi, I discussed the issue offline with Hal, and would like to clarify
> what is exactly going on, what are trade-offs for different
> solutions, and ask for more feedback on my proposed solution (

> http://reviews.llvm.org/D11728 ). I will use the example from Hal's

> 0, %for.body.4.preheader ], [ % indvars.iv.be , %for.body.4.backedge

> ] %arrayidx = getelementptr inbounds float, float* %y, i64
> %indvars.iv, !dbg !35 %0 = load float, float* %arrayidx, align 4,
> !dbg !35, !tbaa !40 %arrayidx6 = getelementptr inbounds float,
> float* %z, i64 %indvars.iv, !dbg !44 %1 = load float, float*
> %arrayidx6, align 4, !dbg !44, !tbaa !40 %add = fadd float %0, %1,
> !dbg !45 %arrayidx8 = getelementptr inbounds float, float* %x, i64
> %indvars.iv, !dbg !46 store float %add, float* %arrayidx8, align 4,
> !dbg !47, !tbaa !40 %indvars.iv.next = add nuw nsw i64 %indvars.iv,
> 1, !dbg !48 %exitcond = icmp eq i64 %indvars.iv.next, 1600, !dbg !48
> br i1 %exitcond, label %for.cond.cleanup.3, label
> %for.body.4.backedge, !dbg !48 for.body.4.backedge: ; preds =

> %for.body.4, %for.cond.cleanup.3 % indvars.iv.be = phi i64 [

> %indvars.iv.next, %for.body.4 ], [ 0, %for.cond.cleanup.3 ] br label
> %for.body.4x LLVM loop vectorizer rejects to vectorize any loop for
> which a loop latch (for.body.4.backedge) is different from a loop
> exiting block (for.cond.cleanup.3). The loop vectorizer can assume
> that all instructions in the loop are executed the same number of
> times with the test. I believe a fix is in order in one way or
> another because the example is simple and common enough and
> vectorized by other compilers. We may approach it by either (1)
> preventing loops from being collapsed in the first place or (2)
> teaching loop vectorizer to handle collapsed loops. For (2), we may
> need to allow loop vectorizer to forego the assumption and handle
> the loop as it is. The assumption seems fundamental to many of the
> vectorization algorithms, so it will require extensive updates or
> may end up with reverting the loop back to a properly nested form.
> The downside of (1) is that it may slow down common optimization
> passes that are repeatedly executed before vectorization. My patch (

> http://reviews.llvm.org/D11728 ) is a prototype fix for (1) that

> modifies Jump Threading and Simplify-the-CFG to not eliminate an
> empty loop header BB even when the loop does not have a PHI node for
> its induction variable. The details can be found at

> http://reviews.llvm.org/D11728 . I would welcome and appreciate any

> comments or feedback. Best, Hyojin <graycol.gif> Hal Finkel
> ---07/16/2015 03:19:24 AM-------- Original Message ----- > From:
> "Hal Finkel" < hfi...@anl.gov > From: Hal Finkel < hfi...@anl.gov
> > To: Chandler Carruth < chan...@google.com > Cc:
> llv...@cs.uiuc.edu Date: 07/16/2015 03:19 AM Subject: Re: [LLVMdev]
> Improving loop vectorizer support for loops with a volatile
> iteration variable Sent by: llvmdev...@cs.uiuc.edu
> ----- Original Message -----
>
>
> From: "Hal Finkel" < hfi...@anl.gov > To: "Chandler Carruth" <
> chan...@google.com > Cc: llv...@cs.uiuc.edu Sent: Thursday, July
> 16, 2015 1:58:02 AM Subject: Re: [LLVMdev] Improving loop vectorizer
> support for loops with a volatile iteration variable
>

> From: "Hal Finkel" < hfi...@anl.gov > To: "Chandler Carruth" <
> chan...@google.com > Cc: llv...@cs.uiuc.edu Sent: Thursday, July
> 16, 2015 1:46:42 AM Subject: Re: [LLVMdev] Improving loop vectorizer
> support for loops with a volatile iteration variable
>

> From: "Chandler Carruth" < chan...@google.com > To: "Hal Finkel" <
> hfi...@anl.gov > Cc: "Hyojin Sung" < hs...@us.ibm.com >,
> llv...@cs.uiuc.edu Sent: Thursday, July 16, 2015 1:06:03 AM
> Subject: Re: [LLVMdev] Improving loop vectorizer support for loops
> with a volatile iteration variable On Wed, Jul 15, 2015 at 6:36 PM
> Hal Finkel < hfi...@anl.gov > wrote:
>
>
>
> From: "Chandler Carruth" < chan...@google.com > To: "Hyojin Sung" <
> hs...@us.ibm.com >, llv...@cs.uiuc.edu Sent: Wednesday, July 15,
> 2015 7:34:54 PM Subject: Re: [LLVMdev] Improving loop vectorizer
> support for loops with a volatile iteration variable On Wed, Jul 15,
> 2015 at 12:55 PM Hyojin Sung < hs...@us.ibm.com > wrote:
>
> Hi all, I would like to propose an improvement of the “almost dead”
> block elimination in Transforms/Local.cpp so that it will preserve
> the canonical loop form for loops with a volatile iteration
> variable. *** Problem statement Nested loops in LCALS Subset B (

> https://codesign.llnl.gov/LCALS.php ) are not vectorized with LLVM

> ], [ % indvars.iv.be , %for.body.backedge ] ... %indvars.iv.next =

> add nuw nsw i64 %indvars.iv, 1 %exitcond = icmp eq i64
> %indvars.iv.next, 1600 br i1 %exitcond, label %while.cond.loopexit,
> label %for.body.backedge for.body.backedge: ; preds = %for.body,

> %while.cond.loopexit % indvars.iv.be = phi i64 [ %indvars.iv.next,

> %for.body ], [ 0, %while.cond.loopexit ] br label %for.body
> while.end.loopexit: ; preds = %while.cond.loopexit br label
> %while.end while.end: ; preds = %while.end.loopexit, %entry ret void
> } and we can currently only vectorize loops where the loop latch is
> also the loop's exiting block. In this case, as in the case with the
> volatile variable, vectorization is blocked by this constraint (here
> the backedge is from the terminator of %for.body.backedge, but the
> loop exiting block is %for.body). The check in the vectorizer is
> explicit: // We only handle bottom-tested loops, i.e. loop in which
> the condition is // checked at the end of each iteration. With that
> we can assume that all // instructions in the loop are executed the
> same number of times. if (TheLoop->getExitingBlock() !=
> TheLoop->getLoopLatch()) { ... Thanks for the detailed explanation.
> This makes much more sense why we need to handle it. I think its
> much better to look at nested loops of this form than anything to do
> with volatile -- the latter is too prone to other random
> optimizations turning off. Regarding this problem, it would be
> interesting to know based on this explanation what the desired fix
> would be. I see at least these options: 1) Canonicalize loops harder
> to make them look the way the vectorizer wants. If this can be done
> without causing significant problems, it seems likely the best
> approach. I agree. In this case, we could certainly fold the trivial
> %for.body.backedge block into %for.body, meaning transforming this:
> for.body.backedge: ; preds = %while.cond.loopexit, %for.body %

> indvars.iv.be = phi i64 [ %indvars.iv.next, %for.body ], [ 0,

> %while.cond.loopexit ] br label %for.body for.body: ; preds =
> %for.body.backedge, %for.body.preheader %indvars.iv = phi i64 [ 0,

> %for.body.preheader ], [ % indvars.iv.be , %for.body.backedge ] ...

> %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1 %exitcond = icmp
> eq i64 %indvars.iv.next, 1600 br i1 %exitcond, label
> %while.cond.loopexit, label %for.body.backedge into this: for.body:

> ; preds = %for.body.backedge, %for.body.preheader % indvars.iv.be =

llvm...@lists.llvm.org http://llvm.cs.uiuc.edu
http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev

[Gerolf Hoflehner via llvm-dev]

Hi Hal,

Please see below. I like your break down. 

On Aug 13, 2015, at 11:00 AM, Hal Finkel <hfi...@anl.gov> wrote:

Hi Gerolf,

I think we have several (perhaps separable) issues here:

1. Do we have a canonical form for loops, preserved through the optimizer, that allows naturally-constructed loop nests to remain separable?

Are you going to take a stab at the documentation like you mention below?

2. Do we forbid non-lowering transformations that turn vectorizable loops into non-vectorizable loops?

This is the long-term architecture question which I believe requires more data and investigation. My preference for now is to be flexible. Then it is OK to allow vec -> non-vec trafos, but it should be done consciously (e.g. based on a cost model or for some other well understood reason). Therefore the first step in answering that question is providing the means to discover and reason about such transformations.

The current solution is that the first instance of loop rotation attempts to create a vectorizable loop. Then some transformations may break that property. So another instance of rotation is invoked before the vectorizer. But in some cases like Hyojin’s that still fails. So it is a brittle design and the story behind it looks like this (and I’m speculating w/o any history digging, so all of this might be totally off):

a) initial  design - transform a loop into vectorizable form at some point and keep it that way (first call to loop rotation).

b) surprise 1 - some transformations don’t play along. Probably the initial answers were something like let’s patch them ( so loops stay vectorizable)

c) surprise 2 - over time the patches were getting ugly. So let’s patch everything with a new pass (second call to loop rotation) and “be done with it."

We are now at d) surprise 3 - after all patches the design space has still not been covered. Perhaps worse, how much of it has been covered? 

Possible alternatives: Why not have a general pass before the vectorizer to generate a vectorizable loop? Or  have a helper class for it in the vectorizer (check legality - if not legal transform - if successful vectorize otherwise give up)?

3. How do we detect cases where transformations cause a negative answer to either of the above?

Now (2) is a subset of (1), but (1) applies as a general structural property to a much larger class of transformations. Hyojin has found, and proposed a fix for, two important instances of (1) by manual investigation, and I think we should move forward with reviewing those (the actual Phabricator differential did not cc llvm-commits, etc. and I've commented there asking her to open a new one).

I guess I can see good reasons for signing up on your proposal. Similar patches have been accepted in the past, so why come out of the woods and pick on this one? My concern is maintainability. When there are implicit dependences between two passes/phases in the compiler they will break. In my view we are at d) above and that would favor a design overhaul. 

Why are these loops important? Catching important loops for a small maintenance overhead (which hopefully is temporary) would be a reasonable trade-off.

Regarding a more-automatic system for finding these kinds of problems, I'm certainly in favor. We could certainly run the vectorizer's legality check, add metadata to loops found to be vectorizable (or use some other state-keeping mechanism), and then check again later that those loops are still considered legally vectorizable (we could also check that we never move from a profitable to unprofitable state). I suspect this would be much too expensive to run aggressively by default, but we could certainly have a mode for us to use that did this.

Agree that it would be too heavy for default mode. I’ll see that I can clean up my patch so then we can find the best way to make use of it.

Part of the challenge, however, is that loops are often taken in and our of 'LoopSimplify' form, and we normally modify the IR in LoopSimplify before the vectorizer's analysis is runs (and, similarly before other passes that operate on loops). As a result, we (might) need to be judicious regarding where to do these checks (maybe we should do them only after LoopSimplify runs?).

Sure.

Regardless, I think that we should work on documenting what our canonical loop form is, and specifically include restrictions necessary to preserve (1). For (2), this would be harder (what the vectorizer will or will not vectorize will likely change much more over time), but detecting problems here is certainly also very important.

Ok, please see 1. above.

[Hal Finkel via llvm-dev]

From: "Gerolf Hoflehner" <ghofl...@apple.com>
To: "Hal Finkel" <hfi...@anl.gov>
Cc: "Hyojin Sung" <hs...@us.ibm.com>, "llvm-dev" <llvm...@lists.llvm.org>
Sent: Thursday, August 13, 2015 7:50:27 PM
Subject: Re: [LLVMdev] Improving loop vectorizer support for loops with a volatile iteration variable

Hi Hal,

Please see below. I like your break down. 

On Aug 13, 2015, at 11:00 AM, Hal Finkel <hfi...@anl.gov> wrote:

Hi Gerolf,

I think we have several (perhaps separable) issues here:

1. Do we have a canonical form for loops, preserved through the optimizer, that allows naturally-constructed loop nests to remain separable?

Are you going to take a stab at the documentation like you mention below?

I'm really not the right person to do this; but I'm happy to help. Chandler also told me that he would help. The basic structure is described in the comments at the top of LoopSimplify.cpp, and there are some additional comments regarding nested loops interspersed with the code. Then there is LCSSA (which is described briefly in LCSSA.cpp). Starting from those pieces likely makes sense.

2. Do we forbid non-lowering transformations that turn vectorizable loops into non-vectorizable loops?

This is the long-term architecture question which I believe requires more data and investigation. My preference for now is to be flexible. Then it is OK to allow vec -> non-vec trafos, but it should be done consciously (e.g. based on a cost model or for some other well understood reason). Therefore the first step in answering that question is providing the means to discover and reason about such transformations.

The current solution is that the first instance of loop rotation attempts to create a vectorizable loop. Then some transformations may break that property. So another instance of rotation is invoked before the vectorizer. But in some cases like Hyojin’s that still fails. So it is a brittle design and the story behind it looks like this (and I’m speculating w/o any history digging, so all of this might be totally off):

a) initial  design - transform a loop into vectorizable form at some point and keep it that way (first call to loop rotation).

b) surprise 1 - some transformations don’t play along. Probably the initial answers were something like let’s patch them ( so loops stay vectorizable)

c) surprise 2 - over time the patches were getting ugly. So let’s patch everything with a new pass (second call to loop rotation) and “be done with it."

We are now at d) surprise 3 - after all patches the design space has still not been covered. Perhaps worse, how much of it has been covered? 

Possible alternatives: Why not have a general pass before the vectorizer to generate a vectorizable loop? Or  have a helper class for it in the vectorizer (check legality - if not legal transform - if successful vectorize otherwise give up)?

We already have LoopSimplify that runs prior to the vectorizer, so enhancing that certainly seems like one option. As I recall, Hyojin investigated that as one possible option.

3. How do we detect cases where transformations cause a negative answer to either of the above?

Now (2) is a subset of (1), but (1) applies as a general structural property to a much larger class of transformations. Hyojin has found, and proposed a fix for, two important instances of (1) by manual investigation, and I think we should move forward with reviewing those (the actual Phabricator differential did not cc llvm-commits, etc. and I've commented there asking her to open a new one).

I guess I can see good reasons for signing up on your proposal. Similar patches have been accepted in the past, so why come out of the woods and pick on this one? My concern is maintainability. When there are implicit dependences between two passes/phases in the compiler they will break. In my view we are at d) above and that would favor a design overhaul. 

Why are these loops important? Catching important loops for a small maintenance overhead (which hopefully is temporary) would be a reasonable trade-off.

They're important because they show up in real performance-sensitive code. Not the volatile-variable case (I suspect that's relegated to benchmarks), but I've certainly seen loop nests that look something like:

while(!done(...)) {
  for (...)
    ... [vectorizable]
}

which is common in iterative solvers, and the vectorizer is crippled, as I discovered, on these loops for the same reason. Regarding maintainability, I think we need to decide on the right future direction, and accept a fix in line with that direction.

Regarding a more-automatic system for finding these kinds of problems, I'm certainly in favor. We could certainly run the vectorizer's legality check, add metadata to loops found to be vectorizable (or use some other state-keeping mechanism), and then check again later that those loops are still considered legally vectorizable (we could also check that we never move from a profitable to unprofitable state). I suspect this would be much too expensive to run aggressively by default, but we could certainly have a mode for us to use that did this.

Agree that it would be too heavy for default mode. I’ll see that I can clean up my patch so then we can find the best way to make use of it.

Sounds good.

Thanks again,
Hal

Part of the challenge, however, is that loops are often taken in and our of 'LoopSimplify' form, and we normally modify the IR in LoopSimplify before the vectorizer's analysis is runs (and, similarly before other passes that operate on loops). As a result, we (might) need to be judicious regarding where to do these checks (maybe we should do them only after LoopSimplify runs?).

Sure.

Regardless, I think that we should work on documenting what our canonical loop form is, and specifically include restrictions necessary to preserve (1). For (2), this would be harder (what the vectorizer will or will not vectorize will likely change much more over time), but detecting problems here is certainly also very important.

Ok, please see 1. above.

Thanks again,
Hal

---

[Philip Reames via llvm-dev]

On 08/18/2015 08:52 PM, Hal Finkel via llvm-dev wrote:

---

From: "Gerolf Hoflehner" <ghofl...@apple.com>
To: "Hal Finkel" <hfi...@anl.gov>
Cc: "Hyojin Sung" <hs...@us.ibm.com>, "llvm-dev" <llvm...@lists.llvm.org>
Sent: Thursday, August 13, 2015 7:50:27 PM
Subject: Re: [LLVMdev] Improving loop vectorizer support for loops with a volatile iteration variable

Hi Hal,

Please see below. I like your break down. 

On Aug 13, 2015, at 11:00 AM, Hal Finkel <hfi...@anl.gov> wrote:

Hi Gerolf,

I think we have several (perhaps separable) issues here:

1. Do we have a canonical form for loops, preserved through the optimizer, that allows naturally-constructed loop nests to remain separable?

Are you going to take a stab at the documentation like you mention below?

I'm really not the right person to do this; but I'm happy to help. Chandler also told me that he would help. The basic structure is described in the comments at the top of LoopSimplify.cpp, and there are some additional comments regarding nested loops interspersed with the code. Then there is LCSSA (which is described briefly in LCSSA.cpp). Starting from those pieces likely makes sense.

I'm happy to volunteer myself and Sanjoy to help out here.  If someone wants to take a first pass, we're happy to help revise and point out the exceptions to the general rules we've encountered.  Having this documented would be very helpful.

Philip