[LoopUnroll] Avoid unnecessary DT recomputation.

[mzolotukhin]

Summary

When we completely unroll a loop, it's pretty easy to update DT in-place and
thus avoid rebuilding it. DT recalculation is one of the most time-consuming
tasks in loop-unroll, so avoiding it at least in case of full unroll should be
beneficial.

On some extreme (but still real-world) tests this patch improves compile time by
~2x.

[mzolotukhin]

• s/UniqueExit/Exit/ - the exit block doesn't need to be unique.

[chandlerc]

Only high-level question is whether all of these cases have to be handled when the loop is in simplified form.

lib/Transforms/Utils/LoopUnroll.cpp

115–117

Range based for-loop?

533–534

There is a range based function as well for this.

536

I'm surprised this much work is necessary even when the loop is in simplified form?

562–568

Merge the two ifs? And update the comment?

[mzolotukhin]

Thanks for the feedback, I'll update the patch soon.

Michael

lib/Transforms/Utils/LoopUnroll.cpp

536

The reason we're doing this is that dominator for an exit block could change after unrolling. Consider a diamond-like loop body with header H, side blocks A and B, and latchblock L. Suppose B is exiting to E.

If B is the immediate dominator of E before unrolling, it's not the case after unrolling - we'll have several blocks exiting to E, so we have to actually find their common dominator.

Probably, there is a more efficient way of doing this, but even in this form it's a pure win over what we have now.

[chandlerc]

lib/Transforms/Utils/LoopUnroll.cpp

536

Ahh, I see.

So, the reason that this seemed odd to me is that all of these blocks that now branch to E come from unrolled copies of the loop, and so they should all have the same IDom -- the IDom of B from the first copy of the loop (which I think pretty much has to be the header, but I've not thought very hard about that).

But have we done any CFG simplification during unrolling at this point? (I know we talked about that, not sure any of it landed...) If so, that would of course potentially invalidate the idea of basing this purely on the loop structure and structural nature of unrolling.

It's not so much that this is ever going to be expensive at runtime (the domtree should make this pattern quite fast), it was just that I wanted to understand the complexity.

I think explaining some of the context of how to think about this code in comments would be very useful here.

562–568

Also, what do you think about running verifyDomTree here to help flush out any bugs?

[chandlerc]

Also LGTM with the comments addressed. Also happy to chat further.

[mzolotukhin]

Hi Chandler,

Some replies from me are inline. I'm heading off for the weekend now, probably update the patch and add some comments on Monday. Thanks for your feedback!

Michael

lib/Transforms/Utils/LoopUnroll.cpp

536

all of these blocks that now branch to E come from unrolled copies of the loop, and so they should all have the same IDom -- the IDom of B from the first copy of the loop

I don't think it's correct. IDom of B might be H, but doesn't have to be (you can imagine a diamond in diamond structure to prove it).

To explain it better I'll try to use ASCII mad skills here. Here is our (slightly modified) original loop body:

  (H)
   |
   v
  (I)
  / \
 v   v
(A) (B) --> (E)
  \ /
   v
  (L)

Here IDom(B) = I, IDom(E) = B.

After unrolling we'll have:

  (H)
   |
   v
  (I)
  / \
 v   v
(A) (B) ------
  \ /         \
   v          |
  (L)         |
   |          |
   v          |
  (H')        |
   |          |
   v          |
  (I')        |
  /  \        |
 v    v       v
(A') (B') -> (E)
  \  /
   vv
  (L')

In the unrolled loop IDom(B) = I, IDom(B') = I', IDom(E) = NearestCommonDominator(B', B) = I. Pleas note, that it doesn't have to be the header.

That said, I see what you meant by using structural nature of unrolling - we do exploit it when we assign dominators for cloned blocks.

As for the CFG simplification - we perform some folding right after this, in FoldBlockIntoPredecessor, which I also updated in this patch.

562–568

I did have

else 
  DEBUG(DT->verifyDomTree());

here (and I run tests with it), but if I keep it in, it'll regress compile time for Asserts=On builds. If that's fine, I can restore it.

[mehdi_amini]

lib/Transforms/Utils/LoopUnroll.cpp

536

It is not necessarily the header, but it is still IDom(B), do you have an example where IDom(E) != IDom(B)? (Yeah, I love your asciiart-fu ;))

[mzolotukhin]

lib/Transforms/Utils/LoopUnroll.cpp

536

One example is that IDom(E) might be B itself. Also, if you construct an example with nested diamonds, you'll probably be able to get IDom(E) != IDom(B).

But after thinking about it, I think IDom(E) = NearestCommonDominator(B, H'). It looks like obvious to me now, but it's too late on Friday to actually accurately prove it:)

[mzolotukhin]

• Rebase on TOT.
• Optimize dom-info update for exit blocks.
• Add DT verification.
• Add more comments.

[mzolotukhin]

Hi Mehdi, Chandler,

I rewrote the part about updating dominators for exit block. Could you please take a look one more time?

Thanks,
Michael

[chandlerc]

Awesome, thanks Michael. And yea, I'm very happy to have the verify in here until someone complains about it. This is too easy to get wrong, and the only way we'll find the really awesome test cases are with the verify left in place.

[mehdi_amini]

lib/Transforms/Utils/LoopUnroll.cpp

115–117

The above comment on range based for-loop still applies I think?

[mzolotukhin]

Ouch, I missed that, I'll commit this as a follow-up. Thanks!

Michael

[mehdi_amini]

lib/Transforms/Utils/LoopUnroll.cpp

121

Random thought: couldn't the pattern above being implemented more efficiently inside the DT (by breaking the invariant during the process): no need for a temporary vector and the ability to reserve the space in the new dominator (could just directly append the source children vector to the new one).
This pattern seems to appear at multiple places in the codebase (looked very quickly).

[mehdi_amini]

llvm/trunk/lib/Transforms/Utils/LoopUnroll.cpp

579 ↗

(On Diff #48755)

If DT is null, this will always break (the test will be false, the else branch taken, and the nullptr dereferenced).
Clang can be smart and eliminate your null check as well, deducing that it is not possible for it to be null.

[mzolotukhin]

llvm/trunk/lib/Transforms/Utils/LoopUnroll.cpp

579 ↗

(On Diff #48755)

Good catch! Thank, Mehdi!

lib/Transforms/Utils/LoopUnroll.cpp

 /// only has one predecessor, and that predecessor only has one successor.
 /// The LoopInfo Analysis that is passed will be kept consistent.  If folding is
 /// successful references to the containing loop must be removed from
 /// ScalarEvolution by calling ScalarEvolution::forgetLoop because SE may have
 /// references to the eliminated BB.  The argument ForgottenLoops contains a set
 /// of loops that have already been forgotten to prevent redundant, expensive
 /// calls to ScalarEvolution::forgetLoop.  Returns the new combined block.
 static BasicBlock *
-FoldBlockIntoPredecessor(BasicBlock *BB, LoopInfo* LI, ScalarEvolution *SE,
-                         SmallPtrSetImpl<Loop *> &ForgottenLoops) {
+FoldBlockIntoPredecessor(BasicBlock *BB, LoopInfo *LI, ScalarEvolution *SE,
+                         SmallPtrSetImpl<Loop *> &ForgottenLoops,
+                         DominatorTree *DT) {
   // Merge basic blocks into their predecessor if there is only one distinct
   // pred, and if there is only one distinct successor of the predecessor, and
   // if there are no PHI nodes.
   BasicBlock *OnlyPred = BB->getSinglePredecessor();
   if (!OnlyPred) return nullptr;
 
   if (OnlyPred->getTerminator()->getNumSuccessors() != 1)
     return nullptr;
FoldBlockIntoPredecessor(BasicBlock *BB, LoopInfo *LI, ScalarEvolution *SE,
   BB->replaceAllUsesWith(OnlyPred);
 
   // Move all definitions in the successor to the predecessor...
   OnlyPred->getInstList().splice(OnlyPred->end(), BB->getInstList());
 
   // OldName will be valid until erased.
   StringRef OldName = BB->getName();
 
-  // Erase basic block from the function...
+  // Erase the old block and update dominator info.
+  if (DT)
+    if (DomTreeNode *DTN = DT->getNode(BB)) {
+      DomTreeNode *PredDTN = DT->getNode(OnlyPred);
+      SmallVector<DomTreeNode *, 8> Children(DTN->begin(), DTN->end());
+      for (SmallVectorImpl<DomTreeNode *>::iterator DI = Children.begin(),
+                                                    DE = Children.end();
+           DI != DE; ++DI)
+        DT->changeImmediateDominator(*DI, PredDTN);
+
+      DT->eraseNode(BB);
+    }
 
   // ScalarEvolution holds references to loop exit blocks.
   if (SE) {
     if (Loop *L = LI->getLoopFor(BB)) {
       if (ForgottenLoops.insert(L).second)
         SE->forgetLoop(L);
     }
   }
for (LoopBlocksDFS::RPOIterator BB = BlockBegin; BB != BlockEnd; ++BB) {
       // we can insert the proper branches later.
       if (*BB == Header)
         Headers.push_back(New);
       if (*BB == LatchBlock)
         Latches.push_back(New);
 
       NewBlocks.push_back(New);
       UnrolledLoopBlocks.push_back(New);
+
+      // Update DomTree
+      if (DT) {
+        if (*BB == Header)
+          DT->addNewBlock(New, Latches[It - 1]);
+        else {
+          auto BBDomNode = DT->getNode(*BB);
+          auto BBIDom = BBDomNode->getIDom();
+          BasicBlock *OriginalBBIDom = BBIDom->getBlock();
+          DT->addNewBlock(
+              New, cast<BasicBlock>(LastValueMap[cast<Value>(OriginalBBIDom)]));
+        }
+      }
     }
 
     // Remap all instructions in the most recent iteration
     for (unsigned i = 0; i < NewBlocks.size(); ++i)
       for (BasicBlock::iterator I = NewBlocks[i]->begin(),
            E = NewBlocks[i]->end(); I != E; ++I)
         ::RemapInstruction(&*I, LastValueMap);
   }
if (NeedConditional) {
           }
         }
       }
       // Replace the conditional branch with an unconditional one.
       BranchInst::Create(Dest, Term);
       Term->eraseFromParent();
     }
   }
+  // Update dominators of loop exit blocks
+  if (DT) {
+    for (auto Exit : ExitBlocks) {
+      BasicBlock *ExitIDom = nullptr;
+      for (auto BI = pred_begin(Exit), BE = pred_end(Exit); BI != BE; BI++) {
+        ExitIDom =
+            ExitIDom ? DT->findNearestCommonDominator(ExitIDom, *BI) : *BI;
+      }
+      DT->changeImmediateDominator(Exit, ExitIDom);
+    }
+  }
 
   // Merge adjacent basic blocks, if possible.
   SmallPtrSet<Loop *, 4> ForgottenLoops;
   for (unsigned i = 0, e = Latches.size(); i != e; ++i) {
     BranchInst *Term = cast<BranchInst>(Latches[i]->getTerminator());
     if (Term->isUnconditional()) {
       BasicBlock *Dest = Term->getSuccessor(0);
-      if (BasicBlock *Fold = FoldBlockIntoPredecessor(Dest, LI, SE,
-                                                      ForgottenLoops)) {
+      if (BasicBlock *Fold =
+              FoldBlockIntoPredecessor(Dest, LI, SE, ForgottenLoops, DT)) {
         // Dest has been folded into Fold. Update our worklists accordingly.
         std::replace(Latches.begin(), Latches.end(), Dest, Fold);
         UnrolledLoopBlocks.erase(std::remove(UnrolledLoopBlocks.begin(),
                                              UnrolledLoopBlocks.end(), Dest),
                                  UnrolledLoopBlocks.end());
       }
     }
   }
 
   // FIXME: We could register any cloned assumptions instead of clearing the
   // whole function's cache.
   AC->clear();
 
+  if (DT) {
     // FIXME: Reconstruct dom info, because it is not preserved properly.
     // Incrementally updating domtree after loop unrolling would be easy.
-  if (DT)
+    if (!CompletelyUnroll)
       DT->recalculate(*L->getHeader()->getParent());
+  }
 
   // Simplify any new induction variables in the partially unrolled loop.
   if (SE && !CompletelyUnroll) {
     SmallVector<WeakVH, 16> DeadInsts;
     simplifyLoopIVs(L, SE, DT, LI, DeadInsts);
 
     // Aggressively clean up dead instructions that simplifyLoopIVs already
     // identified. Any remaining should be cleaned up below.
     while (!DeadInsts.empty())