[SimpleLoopUnswitch] Unswitch by experimental.guard intrinsics

[mkazantsev]

Summary

This patch adds support of llvm.experimental.guard intrinsics to non-trivial
simple loop unswitching. These intrinsics represent implicit control flow which
has pretty much the same semantics as usual conditional branches. The
algorithm of dealing with them is following:

• Consider guards as unswitching candidates;
• If a guard is considered the best candidate, turn it into a branch;
• Apply normal unswitching algorithm on this branch.

The patch has no compile time effect on code that does not contain any guards.

[mkazantsev]

Typo fix.

[chandlerc]

While this is a neat way to handle this, I'm a bit worried about the effects of it...

Does normal unswitch do the same thing currently?

Have you looked at how hard it would be to handle this case directly rather than by introducing a branch?

lib/Transforms/Scalar/SimpleLoopUnswitch.cpp

2253–2255 ↗

(On Diff #171250)

What happens if the guard isn't unswitched? Will anything clean up the branch?

2446–2447 ↗

(On Diff #171250)

Needs clang-format?

[mkazantsev]

The old version of unswitching only needs a condition to unswitch on, not a particular user of this condition. That's why guards and branches can be handled in the same manner there. New one, however, also needs info about the particular user (which is currently always a terminator instruction). The boons it gives to us is that we can make a surgical DT and LI update and not invalidate these analyzes as a whole. The underlying logic in unswitchNontrivialInvariants deeply specializes on the terminator type (either branch of switch), and introducing a new type of a user (which is also a non-terminator) will make us write a lot of ugly duplicating code or mess up the existing code. It will also be bug prone.

I think it's not worth doing because I hope that we will be able to get rid of llvm.experimental.guard at all once we have D51207 merged. In that case, guards will be expressed as normal branches. I put a lot of faith into this patch, that's why I don't want to mess up the existing code (or duplicate it) just to support something that isn't going to live long.

lib/Transforms/Scalar/SimpleLoopUnswitch.cpp

2253–2255 ↗

(On Diff #171250)

Nope. As far as I'm aware, unswitchNontrivialInvariants can only return false if one of successors starts with CleanupPadInst, which is not the case we have after this split. I believe that in practice unswitching should be always successful after this transform. I placed a TODO to follow-up on this in the future (I think we need to assert that fact).

[mkazantsev]

Fixed formatting.

[mkazantsev]

lib/Transforms/Scalar/SimpleLoopUnswitch.cpp

2253–2255 ↗

(On Diff #171250)

The only way how unswitch may fail is marked as FIXME:

// We cannot unswitch if exit blocks contain a cleanuppad instruction as we
// don't know how to split those exit blocks.
// FIXME: We should teach SplitBlock to handle this and remove this
// restriction.
for (auto *ExitBB : ExitBlocks)
  if (isa<CleanupPadInst>(ExitBB->getFirstNonPHI()))
    return false;

I think it's more or less OK to not heal the guard if unswitching failed, provided that there is a plan to either make it unfailable or migrate to new guards representation in D51207.

[mkazantsev]

lib/Transforms/Scalar/SimpleLoopUnswitch.cpp

2476 ↗

(On Diff #171256)

We should signal that we did this and return Changed.

[mkazantsev]

Fixed Changed reporting for case when lowering was successful and unswitching was not.

[chandlerc]

In D53744#1276992, @mkazantsev wrote:

The old version of unswitching only needs a condition to unswitch on, not a particular user of this condition. That's why guards and branches can be handled in the same manner there. New one, however, also needs info about the particular user (which is currently always a terminator instruction). The boons it gives to us is that we can make a surgical DT and LI update and not invalidate these analyzes as a whole. The underlying logic in unswitchNontrivialInvariants deeply specializes on the terminator type (either branch of switch), and introducing a new type of a user (which is also a non-terminator) will make us write a lot of ugly duplicating code or mess up the existing code. It will also be bug prone.

I think it's not worth doing because I hope that we will be able to get rid of llvm.experimental.guard at all once we have D51207 merged. In that case, guards will be expressed as normal branches. I put a lot of faith into this patch, that's why I don't want to mess up the existing code (or duplicate it) just to support something that isn't going to live long.

Thanks for the background context on why you're going with this approach. I think it makes total sense in that context.

I think it might be good to write some of this information down in the documentation next to the code. Something along the lines of:

/// FIXME: Eventually, the use of this routine to handle guard intrinsics should be removed in favor of non-implicit control flow intrinsics, or re-visited to ensure we have a sustainable approach. The approach here of converting to branches is intended to be a simple and hopefully not long-term mechanism to support the existing users of guard intrinsics.

[chandlerc]

lib/Transforms/Scalar/SimpleLoopUnswitch.cpp

2253–2255 ↗

(On Diff #171250)

I think you should just go ahead and assert this.

Because this code is *creating* the branch, it can (and does) ensure that the resulting branch does not go to a block with a cleanup pad. The invariant that this is an unswitchable condition should always hold and we should just verify it.

Then we don't even need to discuss cleanups, etc.

[mkazantsev]

lib/Transforms/Scalar/SimpleLoopUnswitch.cpp

2253–2255 ↗

(On Diff #171250)

Unfortunately, unswitching breaks is *any* exit block has a cleanup. :( I was able to construct such test:

define void @test_cleanuppad(i1 %cond, i32 %N) uwtable ssp personality i8* bitcast (i32 (...)* @__gxx_personality_v0 to i8*) {

entry:
  br label %loop

loop:
  %iv = phi i32 [ 0, %entry ], [ %iv.next, %loop ]
  call void (i1, ...) @llvm.experimental.guard(i1 %cond) [ "deopt"() ]
  %iv.next = add i32 %iv, 1
  invoke void @may_throw(i32 %iv) to label %loop unwind label %exit

exit:
  %cp = cleanuppad within none []
  cleanupret from %cp unwind to caller

}

I think we can make the cleanup check before we make any transform, and therefore we guarantee that the unswitching is always successful.

[chandlerc]

lib/Transforms/Scalar/SimpleLoopUnswitch.cpp

2253–2255 ↗

(On Diff #171250)

Ah, nice test case, and yeah I like the idea of making this layer check the conditions necessary.

[mkazantsev]

Rebased on top of https://reviews.llvm.org/D53747. Now we know for sure that we have a point after which the unswitching will succeed, and we don't need to worry about turning guard into a branch and then failing to unswitch.

[chandlerc]

Generally looks fine, two minor adjustments below and LGTM once the underlying cleanuppad change lands.

lib/Transforms/Scalar/SimpleLoopUnswitch.cpp

2241–2242 ↗

(On Diff #171265)

Update comment to reflect that this is now accurate?

2258–2265 ↗

(On Diff #171265)

I assume this is trying to avoid the cost of looking for a guard intrinsic? We already walk all the instructions in the loop several times in this routine, so I'm not sure this matters much in practice. I'd just skip this and check the flag below.

[mkazantsev]

lib/Transforms/Scalar/SimpleLoopUnswitch.cpp

2479 ↗

(On Diff #171265)

I just realized that we should add the newly-created exit block to the array here.

[mkazantsev]

Added the new exit block to the vector. Though it was pretty straightforward, I will re-run my fuzz testing to make sure everything is fine now.

lib/Transforms/Scalar/SimpleLoopUnswitch.cpp

2258–2265 ↗

(On Diff #171265)

Maybe we'll remove it in future, but I see no harm in it.

llvm/trunk/lib/Transforms/Scalar/SimpleLoopUnswitch.cpp

 #include "llvm/ADT/SetVector.h"
 #include "llvm/ADT/SmallPtrSet.h"
 #include "llvm/ADT/SmallVector.h"
 #include "llvm/ADT/Statistic.h"
 #include "llvm/ADT/Twine.h"
 #include "llvm/Analysis/AssumptionCache.h"
 #include "llvm/Analysis/CFG.h"
 #include "llvm/Analysis/CodeMetrics.h"
+#include "llvm/Analysis/GuardUtils.h"
 #include "llvm/Analysis/InstructionSimplify.h"
 #include "llvm/Analysis/LoopAnalysisManager.h"
 #include "llvm/Analysis/LoopInfo.h"
 #include "llvm/Analysis/LoopIterator.h"
 #include "llvm/Analysis/LoopPass.h"
 #include "llvm/Analysis/Utils/Local.h"
 #include "llvm/IR/BasicBlock.h"
 #include "llvm/IR/Constant.h"

 #include <utility>
 
 #define DEBUG_TYPE "simple-loop-unswitch"
 
 using namespace llvm;
 
 STATISTIC(NumBranches, "Number of branches unswitched");
 STATISTIC(NumSwitches, "Number of switches unswitched");
+STATISTIC(NumGuards, "Number of guards turned into branches for unswitching");
 STATISTIC(NumTrivial, "Number of unswitches that are trivial");
 
 static cl::opt<bool> EnableNonTrivialUnswitch(
     "enable-nontrivial-unswitch", cl::init(false), cl::Hidden,
     cl::desc("Forcibly enables non-trivial loop unswitching rather than "
              "following the configuration passed into the pass."));
 
 static cl::opt<int>
     UnswitchThreshold("unswitch-threshold", cl::init(50), cl::Hidden,
                       cl::desc("The cost threshold for unswitching a loop."));
 
+static cl::opt<bool> UnswitchGuards(
+    "simple-loop-unswitch-guards", cl::init(true), cl::Hidden,
+    cl::desc("If enabled, simple loop unswitching will also consider "
+             "llvm.experimental.guard intrinsics as unswitch candidates."));
+
 /// Collect all of the loop invariant input values transitively used by the
 /// homogeneous instruction graph from a given root.
 ///
 /// This essentially walks from a root recursively through loop variant operands
 /// which have the exact same opcode and finds all inputs which are loop
 /// invariant. For some operations these can be re-associated and unswitched out
 /// of the loop entirely.
 static TinyPtrVector<Value *>
int Cost = std::accumulate(
         return Sum + computeDomSubtreeCost(*ChildN, BBCostMap, DTCostMap);
       });
   bool Inserted = DTCostMap.insert({&N, Cost}).second;
   (void)Inserted;
   assert(Inserted && "Should not insert a node while visiting children!");
   return Cost;
 }
 
+/// Turns a llvm.experimental.guard intrinsic into implicit control flow branch,
+/// making the following replacement:
+///
+///   <code before guard>
+///   call void (i1, ...) @llvm.experimental.guard(i1 %cond) [ "deopt"() ]
+///   <code after guard>
+///
+/// into
+///
+///   <code before guard>
+///   br i1 %cond, label %guarded, label %deopt
+///
+/// guarded:
+///   <code after guard>
+///
+/// deopt:
+///   call void (i1, ...) @llvm.experimental.guard(i1 false) [ "deopt"() ]
+///   unreachable
+///
+/// It also makes all relevant DT and LI updates, so that all structures are in
+/// valid state after this transform.
+static BranchInst *
+turnGuardIntoBranch(IntrinsicInst *GI, Loop &L,
+                    SmallVectorImpl<BasicBlock *> &ExitBlocks,
+                    DominatorTree &DT, LoopInfo &LI) {
+  SmallVector<DominatorTree::UpdateType, 4> DTUpdates;
+  LLVM_DEBUG(dbgs() << "Turning " << *GI << " into a branch.\n");
+  BasicBlock *CheckBB = GI->getParent();
+
+  // Remove all CheckBB's successors from DomTree. A block can be seen among
+  // successors more than once, but for DomTree it should be added only once.
+  SmallPtrSet<BasicBlock *, 4> Successors;
+  for (auto *Succ : successors(CheckBB))
+    if (Successors.insert(Succ).second)
+      DTUpdates.push_back({DominatorTree::Delete, CheckBB, Succ});
+
+  Instruction *DeoptBlockTerm =
+      SplitBlockAndInsertIfThen(GI->getArgOperand(0), GI, true);
+  BranchInst *CheckBI = cast<BranchInst>(CheckBB->getTerminator());
+  // SplitBlockAndInsertIfThen inserts control flow that branches to
+  // DeoptBlockTerm if the condition is true.  We want the opposite.
+  CheckBI->swapSuccessors();
+
+  BasicBlock *GuardedBlock = CheckBI->getSuccessor(0);
+  GuardedBlock->setName("guarded");
+  CheckBI->getSuccessor(1)->setName("deopt");
+
+  // We now have a new exit block.
+  ExitBlocks.push_back(CheckBI->getSuccessor(1));
+
+  GI->moveBefore(DeoptBlockTerm);
+  GI->setArgOperand(0, ConstantInt::getFalse(GI->getContext()));
+
+  // Add new successors of CheckBB into DomTree.
+  for (auto *Succ : successors(CheckBB))
+    DTUpdates.push_back({DominatorTree::Insert, CheckBB, Succ});
+
+  // Now the blocks that used to be CheckBB's successors are GuardedBlock's
+  // successors.
+  for (auto *Succ : Successors)
+    DTUpdates.push_back({DominatorTree::Insert, GuardedBlock, Succ});
+
+  // Make proper changes to DT.
+  DT.applyUpdates(DTUpdates);
+  // Inform LI of a new loop block.
+  L.addBasicBlockToLoop(GuardedBlock, LI);
+
+  ++NumGuards;
+  return CheckBI;
+}
+
 static bool
 unswitchBestCondition(Loop &L, DominatorTree &DT, LoopInfo &LI,
                       AssumptionCache &AC, TargetTransformInfo &TTI,
                       function_ref<void(bool, ArrayRef<Loop *>)> UnswitchCB,
                       ScalarEvolution *SE) {
   // Collect all invariant conditions within this loop (as opposed to an inner
   // loop which would be handled when visiting that inner loop).
   SmallVector<std::pair<Instruction *, TinyPtrVector<Value *>>, 4>
       UnswitchCandidates;
+
+  // Whether or not we should also collect guards in the loop.
+  bool CollectGuards = false;
+  if (UnswitchGuards) {
+    auto *GuardDecl = L.getHeader()->getParent()->getParent()->getFunction(
+        Intrinsic::getName(Intrinsic::experimental_guard));
+    if (GuardDecl && !GuardDecl->use_empty())
+      CollectGuards = true;
+  }
+
   for (auto *BB : L.blocks()) {
     if (LI.getLoopFor(BB) != &L)
       continue;
 
+    if (CollectGuards)
+      for (auto &I : *BB)
+        if (isGuard(&I)) {
+          auto *Cond = cast<IntrinsicInst>(&I)->getArgOperand(0);
+          // TODO: Support AND, OR conditions and partial unswitching.
+          if (!isa<Constant>(Cond) && L.isLoopInvariant(Cond))
+            UnswitchCandidates.push_back({&I, {Cond}});
+        }
+
     if (auto *SI = dyn_cast<SwitchInst>(BB->getTerminator())) {
       // We can only consider fully loop-invariant switch conditions as we need
       // to completely eliminate the switch after unswitching.
       if (!isa<Constant>(SI->getCondition()) &&
           L.isLoopInvariant(SI->getCondition()))
         UnswitchCandidates.push_back({SI, {SI->getCondition()}});
       continue;
     }
for (BasicBlock *SuccBB : successors(&BB)) {
         assert(Cost >= 0 &&
                "Non-duplicated cost should never exceed total loop cost!");
       }
     }
 
     // Now scale the cost by the number of unique successors minus one. We
     // subtract one because there is already at least one copy of the entire
     // loop. This is computing the new cost of unswitching a condition.
-    assert(Visited.size() > 1 &&
+    // Note that guards always have 2 unique successors that are implicit and
+    // will be materialized if we decide to unswitch it.
+    int SuccessorsCount = isGuard(&TI) ? 2 : Visited.size();
+    assert(SuccessorsCount > 1 &&
            "Cannot unswitch a condition without multiple distinct successors!");
-    return Cost * (Visited.size() - 1);
+    return Cost * (SuccessorsCount - 1);
   };
   Instruction *BestUnswitchTI = nullptr;
   int BestUnswitchCost;
   ArrayRef<Value *> BestUnswitchInvariants;
   for (auto &TerminatorAndInvariants : UnswitchCandidates) {
     Instruction &TI = *TerminatorAndInvariants.first;
     ArrayRef<Value *> Invariants = TerminatorAndInvariants.second;
     BranchInst *BI = dyn_cast<BranchInst>(&TI);
unswitchBestCondition(Loop &L, DominatorTree &DT, LoopInfo &LI,
   }
 
   if (BestUnswitchCost >= UnswitchThreshold) {
     LLVM_DEBUG(dbgs() << "Cannot unswitch, lowest cost found: "
                       << BestUnswitchCost << "\n");
     return false;
   }
 
+  // If the best candidate is a guard, turn it into a branch.
+  if (isGuard(BestUnswitchTI))
+    BestUnswitchTI = turnGuardIntoBranch(cast<IntrinsicInst>(BestUnswitchTI), L,
+                                         ExitBlocks, DT, LI);
+
   LLVM_DEBUG(dbgs() << "  Unswitching non-trivial (cost = "
                     << BestUnswitchCost << ") terminator: " << *BestUnswitchTI
                     << "\n");
   unswitchNontrivialInvariants(L, *BestUnswitchTI, BestUnswitchInvariants,
                                ExitBlocks, DT, LI, AC, UnswitchCB, SE);
   return true;
 }
 

llvm/trunk/test/Transforms/SimpleLoopUnswitch/guards.ll

+; RUN: opt -passes='loop(unswitch),verify<loops>' -enable-nontrivial-unswitch -simple-loop-unswitch-guards -S < %s | FileCheck %s
+; RUN: opt -simple-loop-unswitch -enable-nontrivial-unswitch -simple-loop-unswitch-guards -S < %s | FileCheck %s
+
+declare void @llvm.experimental.guard(i1, ...)
+
+define void @test_simple_case(i1 %cond, i32 %N) {
+; CHECK-LABEL: @test_simple_case(
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    br i1 [[COND:%.*]], label [[ENTRY_SPLIT_US:%.*]], label [[ENTRY_SPLIT:%.*]]
+; CHECK:       entry.split.us:
+; CHECK-NEXT:    br label [[LOOP_US:%.*]]
+; CHECK:       loop.us:
+; CHECK-NEXT:    [[IV_US:%.*]] = phi i32 [ 0, [[ENTRY_SPLIT_US]] ], [ [[IV_NEXT_US:%.*]], [[GUARDED_US:%.*]] ]
+; CHECK-NEXT:    br label [[GUARDED_US]]
+; CHECK:       guarded.us:
+; CHECK-NEXT:    [[IV_NEXT_US]] = add i32 [[IV_US]], 1
+; CHECK-NEXT:    [[LOOP_COND_US:%.*]] = icmp slt i32 [[IV_NEXT_US]], [[N:%.*]]
+; CHECK-NEXT:    br i1 [[LOOP_COND_US]], label [[LOOP_US]], label [[EXIT_SPLIT_US:%.*]]
+; CHECK:       deopt:
+; CHECK-NEXT:    call void (i1, ...) @llvm.experimental.guard(i1 false) [ "deopt"() ]
+; CHECK-NEXT:    unreachable
+;
+
+entry:
+  br label %loop
+
+loop:
+  %iv = phi i32 [ 0, %entry ], [ %iv.next, %loop ]
+  call void (i1, ...) @llvm.experimental.guard(i1 %cond) [ "deopt"() ]
+  %iv.next = add i32 %iv, 1
+  %loop.cond = icmp slt i32 %iv.next, %N
+  br i1 %loop.cond, label %loop, label %exit
+
+exit:
+  ret void
+}
+
+define void @test_two_guards(i1 %cond1, i1 %cond2, i32 %N) {
+; CHECK-LABEL: @test_two_guards(
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    br i1 [[COND1:%.*]], label [[ENTRY_SPLIT_US:%.*]], label [[ENTRY_SPLIT:%.*]]
+; CHECK:       entry.split.us:
+; CHECK-NEXT:    br i1 [[COND2:%.*]], label [[ENTRY_SPLIT_US_SPLIT_US:%.*]], label [[ENTRY_SPLIT_US_SPLIT:%.*]]
+; CHECK:       entry.split.us.split.us:
+; CHECK-NEXT:    br label [[LOOP_US_US:%.*]]
+; CHECK:       loop.us.us:
+; CHECK-NEXT:    [[IV_US_US:%.*]] = phi i32 [ 0, [[ENTRY_SPLIT_US_SPLIT_US]] ], [ [[IV_NEXT_US_US:%.*]], [[GUARDED_US2:%.*]] ]
+; CHECK-NEXT:    br label [[GUARDED_US_US:%.*]]
+; CHECK:       guarded.us.us:
+; CHECK-NEXT:    br label [[GUARDED_US2]]
+; CHECK:       guarded.us2:
+; CHECK-NEXT:    [[IV_NEXT_US_US]] = add i32 [[IV_US_US]], 1
+; CHECK-NEXT:    [[LOOP_COND_US_US:%.*]] = icmp slt i32 [[IV_NEXT_US_US]], [[N:%.*]]
+; CHECK-NEXT:    br i1 [[LOOP_COND_US_US]], label [[LOOP_US_US]], label [[EXIT_SPLIT_US_SPLIT_US:%.*]]
+; CHECK:       deopt1:
+; CHECK-NEXT:    call void (i1, ...) @llvm.experimental.guard(i1 false) [ "deopt"() ]
+; CHECK-NEXT:    unreachable
+; CHECK:       deopt:
+; CHECK-NEXT:    call void (i1, ...) @llvm.experimental.guard(i1 false) [ "deopt"() ]
+; CHECK-NEXT:    unreachable
+; CHECK:       exit:
+; CHECK-NEXT:    ret void
+;
+
+entry:
+  br label %loop
+
+loop:
+  %iv = phi i32 [ 0, %entry ], [ %iv.next, %loop ]
+  call void (i1, ...) @llvm.experimental.guard(i1 %cond1) [ "deopt"() ]
+  call void (i1, ...) @llvm.experimental.guard(i1 %cond2) [ "deopt"() ]
+  %iv.next = add i32 %iv, 1
+  %loop.cond = icmp slt i32 %iv.next, %N
+  br i1 %loop.cond, label %loop, label %exit
+
+exit:
+  ret void
+}
+
+define void @test_conditional_guards(i1 %cond, i32 %N) {
+; CHECK-LABEL: @test_conditional_guards(
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    br i1 [[COND:%.*]], label [[ENTRY_SPLIT_US:%.*]], label [[ENTRY_SPLIT:%.*]]
+; CHECK:       entry.split.us:
+; CHECK-NEXT:    br label [[LOOP_US:%.*]]
+; CHECK:       loop.us:
+; CHECK-NEXT:    [[IV_US:%.*]] = phi i32 [ 0, [[ENTRY_SPLIT_US]] ], [ [[IV_NEXT_US:%.*]], [[BACKEDGE_US:%.*]] ]
+; CHECK-NEXT:    [[CONDITION_US:%.*]] = icmp eq i32 [[IV_US]], 123
+; CHECK-NEXT:    br i1 [[CONDITION_US]], label [[GUARD_US:%.*]], label [[BACKEDGE_US]]
+; CHECK:       guard.us:
+; CHECK-NEXT:    br label [[GUARDED_US:%.*]]
+; CHECK:       backedge.us:
+; CHECK-NEXT:    [[IV_NEXT_US]] = add i32 [[IV_US]], 1
+; CHECK-NEXT:    [[LOOP_COND_US:%.*]] = icmp slt i32 [[IV_NEXT_US]], [[N:%.*]]
+; CHECK-NEXT:    br i1 [[LOOP_COND_US]], label [[LOOP_US]], label [[EXIT_SPLIT_US:%.*]]
+; CHECK:       loop:
+; CHECK-NEXT:    [[IV:%.*]] = phi i32 [ 0, [[ENTRY_SPLIT]] ], [ [[IV_NEXT:%.*]], [[BACKEDGE:%.*]] ]
+; CHECK-NEXT:    [[CONDITION:%.*]] = icmp eq i32 [[IV]], 123
+; CHECK-NEXT:    br i1 [[CONDITION]], label [[GUARD:%.*]], label [[BACKEDGE]]
+; CHECK:       guard:
+; CHECK-NEXT:    br label [[DEOPT:%.*]]
+; CHECK:       deopt:
+; CHECK-NEXT:    call void (i1, ...) @llvm.experimental.guard(i1 false) [ "deopt"() ]
+; CHECK-NEXT:    unreachable
+; CHECK:       backedge:
+; CHECK-NEXT:    [[IV_NEXT]] = add i32 [[IV]], 1
+; CHECK-NEXT:    [[LOOP_COND:%.*]] = icmp slt i32 [[IV_NEXT]], [[N]]
+; CHECK-NEXT:    br i1 [[LOOP_COND]], label %loop, label [[EXIT_SPLIT:%.*]]
+;
+
+entry:
+  br label %loop
+
+loop:
+  %iv = phi i32 [ 0, %entry ], [ %iv.next, %backedge ]
+  %condition = icmp eq i32 %iv, 123
+  br i1 %condition, label %guard, label %backedge
+
+guard:
+  call void (i1, ...) @llvm.experimental.guard(i1 %cond) [ "deopt"() ]
+  br label %backedge
+
+backedge:
+  %iv.next = add i32 %iv, 1
+  %loop.cond = icmp slt i32 %iv.next, %N
+  br i1 %loop.cond, label %loop, label %exit
+
+exit:
+  ret void
+}
+
+define void @test_nested_loop(i1 %cond, i32 %N) {
+; CHECK-LABEL: @test_nested_loop(
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    br i1 [[COND:%.*]], label [[ENTRY_SPLIT:%.*]], label [[OUTER_LOOP_SPLIT:%.*]]
+; CHECK:       entry.split:
+; CHECK-NEXT:    br label [[OUTER_LOOP:%.*]]
+; CHECK:       outer_loop:
+; CHECK-NEXT:    br label [[OUTER_LOOP_SPLIT_US:%.*]]
+; CHECK:       outer_loop.split.us:
+; CHECK-NEXT:    br label [[LOOP_US:%.*]]
+; CHECK:       loop.us:
+; CHECK-NEXT:    [[IV_US:%.*]] = phi i32 [ 0, [[OUTER_LOOP_SPLIT_US]] ], [ [[IV_NEXT_US:%.*]], [[GUARDED_US:%.*]] ]
+; CHECK-NEXT:    br label [[GUARDED_US]]
+; CHECK:       guarded.us:
+; CHECK-NEXT:    [[IV_NEXT_US]] = add i32 [[IV_US]], 1
+; CHECK-NEXT:    [[LOOP_COND_US:%.*]] = icmp slt i32 [[IV_NEXT_US]], [[N:%.*]]
+; CHECK-NEXT:    br i1 [[LOOP_COND_US]], label [[LOOP_US]], label [[OUTER_BACKEDGE_SPLIT_US:%.*]]
+; CHECK:       outer_backedge.split.us:
+; CHECK-NEXT:    br label [[OUTER_BACKEDGE:%.*]]
+; CHECK:       deopt:
+; CHECK-NEXT:    call void (i1, ...) @llvm.experimental.guard(i1 false) [ "deopt"() ]
+; CHECK-NEXT:    unreachable
+; CHECK:       outer_backedge:
+; CHECK-NEXT:    br i1 false, label [[OUTER_LOOP]], label [[EXIT:%.*]]
+;
+
+entry:
+  br label %outer_loop
+
+outer_loop:
+  br label %loop
+
+loop:
+  %iv = phi i32 [ 0, %outer_loop ], [ %iv.next, %loop ]
+  call void (i1, ...) @llvm.experimental.guard(i1 %cond) [ "deopt"() ]
+  %iv.next = add i32 %iv, 1
+  %loop.cond = icmp slt i32 %iv.next, %N
+  br i1 %loop.cond, label %loop, label %outer_backedge
+
+outer_backedge:
+  br i1 undef, label %outer_loop, label %exit
+
+exit:
+  ret void
+}
+
+define void @test_sibling_loops(i1 %cond1, i1 %cond2, i32 %N) {
+; CHECK-LABEL: @test_sibling_loops(
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    br i1 [[COND1:%.*]], label [[ENTRY_SPLIT_US:%.*]], label [[ENTRY_SPLIT:%.*]]
+; CHECK:         [[IV1_US:%.*]] = phi i32 [ 0, [[ENTRY_SPLIT_US]] ], [ [[IV1_NEXT_US:%.*]], [[GUARDED_US:%.*]] ]
+; CHECK-NEXT:    br label [[GUARDED_US]]
+; CHECK:         call void (i1, ...) @llvm.experimental.guard(i1 false) [ "deopt"() ]
+; CHECK-NEXT:    unreachable
+; CHECK:         [[IV2_US:%.*]] = phi i32 [ 0, [[BETWEEN:%.*]] ], [ [[IV1_NEXT_US2:%.*]], [[GUARDED_US2:%.*]] ]
+; CHECK-NEXT:    br label [[GUARDED_US2]]
+; CHECK:         call void (i1, ...) @llvm.experimental.guard(i1 false) [ "deopt"() ]
+; CHECK-NEXT:    unreachable
+;
+
+entry:
+  br label %loop1
+
+loop1:
+  %iv1 = phi i32 [ 0, %entry ], [ %iv1.next, %loop1 ]
+  call void (i1, ...) @llvm.experimental.guard(i1 %cond1) [ "deopt"() ]
+  %iv1.next = add i32 %iv1, 1
+  %loop1.cond = icmp slt i32 %iv1.next, %N
+  br i1 %loop1.cond, label %loop1, label %between
+
+between:
+  br label %loop2
+
+loop2:
+  %iv2 = phi i32 [ 0, %between ], [ %iv2.next, %loop2 ]
+  call void (i1, ...) @llvm.experimental.guard(i1 %cond2) [ "deopt"() ]
+  %iv2.next = add i32 %iv2, 1
+  %loop2.cond = icmp slt i32 %iv2.next, %N
+  br i1 %loop2.cond, label %loop2, label %exit
+
+exit:
+  ret void
+}
+
+; Check that we don't do anything because of cleanuppad.
+; CHECK-LABEL: @test_cleanuppad(
+; CHECK:       call void (i1, ...) @llvm.experimental.guard(i1 %cond) [ "deopt"() ]
+; CHECK-NOT:   call void (i1, ...) @llvm.experimental.guard(
+define void @test_cleanuppad(i1 %cond, i32 %N) personality i32 (...)* @__CxxFrameHandler3 {
+
+entry:
+  br label %loop
+
+loop:
+  %iv = phi i32 [ 0, %entry ], [ %iv.next, %loop ]
+  call void (i1, ...) @llvm.experimental.guard(i1 %cond) [ "deopt"() ]
+  %iv.next = add i32 %iv, 1
+  invoke void @may_throw(i32 %iv) to label %loop unwind label %exit
+
+exit:
+  %cp = cleanuppad within none []
+  cleanupret from %cp unwind to caller
+
+}
+
+declare void @may_throw(i32 %i)
+declare i32 @__CxxFrameHandler3(...)