Hoist/sink malloc/free's in LICM.

[nicholas]

Summary

Hoist/sink malloc/free's in LICM.

Adds a new method on Instruction that answers the query of whether the given Instruction might allocate or free memory. Unlike builtin info which answers whether they are an allocation/deallocation call, this method indicates whether allocation behaviour is possible.

Adds a new analysis on loops called LoopAllocationInfo. This is tightly tied to the internal implementation of LICM, but is exposed in a public API because the relevant parts of LICM's externals, hoistRegion and sinkRegion, are also a public API.

One potentially surprising aspect of this optimization is that it is correct even if the loop exits in the middle, after the malloc and before the free. As long as there is only one live malloc at a time, there is no visible difference in behaviour. It is sufficient to show that we've found free() calls which cover all possible access to the loop backedge (the other option is that the code had a pre-existing double-free, which is undefined behaviour). The loop may thus exit with the malloc either allocated or freed, and we track which loop exits blocks should have frees added to them and which ones should not.

[asbirlea]

A couple of quick comments.
Would you mind splitting the cleanups and bug fix (which will be straightforward to approve) and rebase this on top of that?

llvm/include/llvm/Analysis/LoopAllocationInfo.h

30

Could you include the comment from the description here? "This analysis is tightly coupled to the internal implementation of the LICM transform."

llvm/test/Transforms/LICM/allocs.ll

3

Could you also add a RUN line with -enable-mssa-loop-dependency?

[reames]

I've only skimmed through the high level comments so far, so if if any of this is addressed inline, just say so.

I was wondering why you phrased this as a combination of hoisting and sinking as opposed to promotion. We have the scalar promotion path, and the basic transform you're doing feels a lot like promotion. I suspect that many of the legality aspects will be common.

llvm/include/llvm/IR/Instruction.h

539 ↗

(On Diff #193152)

These feel like they need a bit more specification. In particular, are there any expectations around how new memory is returned? Is a deallocation routine allowed to free a random pointer read from a global?

(i.e. which set of locations are we talking about freeing and allocating?)

llvm/lib/IR/Instruction.cpp

562 ↗

(On Diff #193152)

Err, using a blacklist here feels *really* dangerous.

[nicholas]

I've only skimmed through the high level comments so far, so if if any of this is addressed inline, just say so.

I was wondering why you phrased this as a combination of hoisting and sinking as opposed to promotion. We have the scalar promotion path, and the basic transform you're doing feels a lot like promotion. I suspect that many of the legality aspects will be common.

I didn't frame it as promotion simply because I view promotion as a transform from one thing to another (such as heap to stack, stack to scalar) whereas this optimization was just hoisting out the allocation and sinking the matching deallocations.

The safety checks we need are "is there a path from this malloc back to this malloc without passing through one of the frees we identified?" and "for each exit block, can we see that all paths either pass through one-or-more frees or that all paths do not pass through any frees?". We have no need of MemorySSA or AliasSets like the memory-to-scalar promotion does.

Comparing this code with LICM's promotion does reveal one bug. We might try to insert into catchswitch blocks.

Should I pull it out of hoisting and sinking and into a post-pass like LICM's scalar promotion? Initially I had imagined that it would participate in hoist and sink's queries of whether "are all arguments to this instruction loop invariant (outside the loop)" as hoists instructions iterately, it would hoist malloc like any other instruction. In practice that couldn't happen because the hoist is deferred until sink time. We do need the malloc hoist to occur after hoistRegion in order to make the malloc size argument invariant. So, maybe there's no reason not to place it after hoist and sink? We currently piggy-back on hoist and sink's linear scans, but doing them again is merely a constant factor. It would eliminate the need to expose LoopAllocationInfo as a public interface, which is a benefit.

llvm/include/llvm/IR/Instruction.h

539 ↗

(On Diff #193152)

Acknowledged.

These exist to help us maintain a C++-language invariant that we may not increase the total amount of heap memory in use, so, malloc+free+malloc may not be transformed into malloc+malloc+free. We simply need to avoid moving memory allocations around each other (though we can sort a continuous block of memory allocations if there's no frees in between). As such, there's no consideration of how the allocated memory would be returned or which memory gets deallocated (similar to how mayReadOrWriteMemory doesn't). We also don't really talk about what happens when an intrinsic allocates and deallocates internally, I think ultimately whether the intrinsic wants to be treated as something we can move other allocation/deallocation operations around should be left to the intrinsic author. It may not matter to the Objective-C runtime for instance. (But what about Objective-C++?)

I think this will end up getting resolved when we address your comments on llvm/lib/IR/Instruction.cpp:562.

llvm/lib/IR/Instruction.cpp

562 ↗

(On Diff #193152)

Uhh, you aren't suggesting that I actually list them all, right?

The reason I put it here in Instruction instead of hidden away in LICM/LoopAllocationInfo is to make it somewhat clearer that this is a global property that you'll need to update when adding an instruction or intrinsic.

How about if we update Intrinsics.td to include a Mallocs and Frees (similar to Throws) IntrinsicProperty? That makes it opt-in when declaring the intrinsic? (Speaking of which, why Throws and not IntrThrows? Should mine be IntrMallocs/IntrFrees or Mallocs/Frees?)

Also, can a readnone function malloc? Is there any way we could make this stricter?

[nicholas]

In D60056#1450884, @nicholas wrote:

Should I pull it out of hoisting and sinking and into a post-pass like LICM's scalar promotion? Initially I had imagined that it would participate in hoist and sink's queries of whether "are all arguments to this instruction loop invariant (outside the loop)" as hoists instructions iterately, it would hoist malloc like any other instruction. In practice that couldn't happen because the hoist is deferred until sink time.

Er, I'm mistaken. It does, because sinkRegion happens first and doesn't call isLoopInvariant, then hoistRegion happens second and that uses isLoopInvariant. So in theory if you have something like ptrtoint of the malloc (maybe an alignment check?), we could hoist that out of the loop too. So it does interlace with the rest of hoistRegion, and that's important for optimization power. I'll add a test to that effect shortly.

[nicholas]

Fixed creation of invalid IR when the loop exit block we need to insert a free call in, is a catchswitch block.

[reames]

llvm/lib/Analysis/LoopAllocationInfo.cpp

93

As discussed offline, bug example:
for (int i = 0; i < N; i++) {

throw_if(requested_size > TOO_BIG);
a = malloc(requested_size);
free(a);

}

[reames]

Other items discussed offline:

• the instruction usage feels to generic with too few uses. I suggested (but will not require) either making them computed properties of an intrinsic, or finding other uses.
• my comment about using a blacklist is clearly wrong. You'd suggested using an attribute in Instrinsics.td. I was fine with that, but it felt like possible overkill.

llvm/include/llvm/Analysis/LoopAllocationInfo.h

92

It really looks like you have a named tuple hidding here.

AllocationInfoEntry?

llvm/lib/Transforms/Scalar/LICM.cpp

862

As debated offline, naming here is problematic.

• addDeallocations -> locationsNeedingFrees? newAllocPlacement?
• toSink -> freesForMalloc

(key piece: avoid action names)

A comment would be help as well.

Pull out the cast to CallInst

865

You might wish to either a) union the source locations, or b) restrict this to non -g

See applyMergedLocation on Instruction.

[reames]

minor comments on interface, LICM part, and tests. I got interrupted and need to come back and finish reviewing the analysis implementation.

llvm/include/llvm/Analysis/LoopAllocationInfo.h

77

either "may ... ?" or "return true if ...". (i.e make a question a question)

82

Should we assert that CI is a malloc within the specified loop?

106

Not sure what the last part of the comment means. Reword?

llvm/lib/Transforms/Scalar/LICM.cpp

535

What about invokes? (Use callbase)

llvm/test/Transforms/LICM/allocs.ll

567

This test is odd. I don't see anything preventing us from reordering the volatile load and malloc. Was this an attempt to test for a side exit? If so, simply using an unknown call before the malloc would be preferred.

[nicholas]

llvm/include/llvm/Analysis/LoopAllocationInfo.h

77

Done, used "return true iff".

82

We don't presently store the right member variables to assert that, and I don't really want to add DEBUG-only member variables. The condition on this function is that mayHoist(CI) must be true. I've added an assert that CI is in entries, which is the condition under which this function would have UB.

llvm/lib/Analysis/LoopAllocationInfo.cpp

93

I've updated this to use isGuaranteedToTransferExecutionToSuccessor on the instructions leading up to the allocation. The "throw_if" example couldn't happen because any opaque function might also malloc or free, so the test @test16 uses a volatile load instead.

llvm/lib/Transforms/Scalar/LICM.cpp

497

Rename this to something clearer, such as "ScanForFrees".

535

Done.

This loop skips the terminator, so it can't be an invoke.

Which is, indeed, a miscompile bug. Fixed above where we initialize ScanForFrees and added @test19.

Changed it to CallBase here anyways.

862

Is picking the first one out of toSink/freesForMalloc a problem? The order depends on the use-list ordering. I don't know what LLVM's current rules on that are, I know that there's https://llvm.org/docs/LangRef.html#use-list-order-directives . Is it considered a bug to have an optimization whose result depends on use list order?

862

Is it possible for a free call to have an operand bundle on a malloc or free call? In CloneInstructionInExitBlock, LICM updates the funclet operand bundle for the new location in the CFG.

llvm/test/Transforms/LICM/allocs.ll

567

This test shows the difference from using isGuaranteedToTranferExecution in analyzeLoop. Before that change, we would have hoisted the malloc above the volatile load. The same test written with a function would not have shown any difference simply because there is no way to create a function that might-throw but can't-malloc/free (without adding new attributes to LLVM). It would also just be @test3 again.

I think the new behaviour tested for here is correct. Suppose you have a system where out of memory terminates the program, and the volatile store is triggering some external action (moves the robot arm). Suppose the programmer is using volatile to make sure the external system is in the safe state before doing the malloc that may terminate, and wants the malloc to complete before beginning any more operations. The compiler should not reorder these actions.

[nicholas]

This update adds a failing test @test21 which demonstrates a miscompile. While we check the block with the free in it for potentially-malloc'ing instructions, we don't examine all the possible paths from where that free was pre-transform to the new location of the free post-transform.

It's not immediately clear to me how to fix that efficiently. It's possible to label each loop block with whether it contains a potentially-malloc'ing instruction, then find all paths from frees to exit blocks which will have frees added to them and do not pass through any backedge or back through the header. We can simplify this path scan, knowing that all exits reachable after the free call are exits where we will insert a free.

[reames]

Assumed inactive, cleaning out review list, please readd if needed.

llvm/include/llvm/Analysis/LoopAllocationInfo.h

+//===- LoopAllocationInfo.h - memory allocations inside loops ---*- C++ -*-===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef LLVM_ANALYSIS_LOOPALLOCATIONINFO_H
+#define LLVM_ANALYSIS_LOOPALLOCATIONINFO_H
+
+#include "llvm/ADT/DenseMap.h"
+#include "llvm/ADT/DenseSet.h"
+#include "llvm/ADT/SmallVector.h"
+
+namespace llvm {
+
+class BasicBlock;
+class CallInst;
+class DominatorTree;
+class Loop;
+class LoopInfo;
+class TargetLibraryInfo;
+
+/// Track ptr = allocfn to one or more freefn(ptr) relationships within the
+/// loop.
+///
+/// This analysis stores information about allocation and deallocation calls
+/// where the pointer is only live within the loop. Its design is tighly coupled
+/// to the internal implementation of LICM, but is exposed in a public API
+/// because the hoistRegion and sinkRegion functions in LICM are also exposed.
+///
+/// An allocation must be both hoisted and sunk, or neither. LICM performs
+/// sinking first, then hoisting. As sinking and hoisting is performed, more
+/// instructions will appear to be loop invariant, making more opportunities.
+/// For this reason we split the analysis into two pieces:
+///   1. for a given pointer that is only live in the loop, retain a list its
+///      matching deallocating instructions.
+///   2. during sinking, if we would sink the deallocation, don't, but inform
+///      the LoopAllocationInfo. If all deallocations would have been sunk, and
+///      hoistSink would hoist the allocation, perform both the hoist and sink
+///      the various deallocation instructions.
+///
+/// We don't need guaranteed execution of any deallocation function during
+/// unwinding. Before this transformation, an unwind mid-loop would leak one
+/// pointer that was allocated at the start of the loop. After this
+/// transformation, it would leak the one pointer allocated before the loop
+/// began. There is no difference visible to the caller.
+///
+/// We do have to ensure that the maximum amount of allocated memory is not
+/// raised by this transformation. To ensure this, we stop looking for
+/// allocations after seeing an opaque function call. For example:
+///   loop_top:
+///     %ptr1 = call i8* @malloc(i32 %loop_invariant1)  ;; candidate
+///     %ptr2 = call i8* @malloc(i32 %loop_variant2)    ;; not candidate
+///     %ptr3 = call i8* @malloc(i32 %loop_invariant3)  ;; candidate
+///     call void @opaque()                             ;; stop scanning
+///     %ptr5 = call i8* @malloc(i32 %loop_invariant5)  ;; not candidate
+///
+/// The same idea applies in reverse order to @free calls at the loop exits:
+///   loop_exiting:
+///      call void @free(%ptr1)  ;; not candidate
+///      call void @opaque()     ;; stop scanning
+///      call void @free(%ptr2)  ;; %ptr2 is candidate
+///                              ;; start here and scan upwards
+class LoopAllocationInfo {
+public:
+  void analyzeLoop(LoopInfo *LI, DominatorTree *DT, TargetLibraryInfo *TLI,
+                   Loop *CurLoop);
+
+  // LICM has determined that this deallocation call may be sunk.
+  void addSafeToSink(const CallInst *CI) {
+    if (DeallocationToAllocation.count(CI))
+      SafeToSink.insert(CI);
+  }
+
+  // Given an allocation call, are all the safety conditions met to hoist this
+  // allocation and sink its deallocations.
+  bool mayHoist(const CallInst *CI) const;
+
+  // Retrieve the list of free calls in the loop matching a given allocation.
+  const SmallVector<CallInst *, 16> &FreesForMalloc(CallInst *CI) const {
+    return Entries.find(CI)->second.Deallocations;
+  }
+
+  // Retrieve the list of loop exit blocks that are only entered after the
+  // allocation has been deallocated. All other loop exit blocks are reached
+  // with the allocation still allocated.
+  const SmallVector<BasicBlock *, 16> &
+  ExitBlocksWithPointerFreed(CallInst *CI) const {
+    return Entries.find(CI)->second.ExitBlocksToAddDeallocationsTo;
+  }
+
+  // Returns whether the loop has any potentially-hoistable allocations.
+  bool empty() const { return Entries.empty(); }
+
+private:
+  struct AllocationInfoEntry {
+    AllocationInfoEntry(
+        SmallVector<CallInst *, 16> &&Deallocations,
+        SmallVector<BasicBlock *, 16> &&ExitBlocksToAddDeallocationsTo)
+        : Deallocations(Deallocations),
+          ExitBlocksToAddDeallocationsTo(ExitBlocksToAddDeallocationsTo) {}
+
+    // Known deallocations of this pointer. These are guaranteed to deallocate
+    // the pointer if executed. (Excludes "free(expr ? ptr1 : nullptr);".)
+    SmallVector<CallInst *, 16> Deallocations;
+
+    // Loop exit blocks which can only be entered with the pointer already freed.
+    SmallVector<BasicBlock *, 16> ExitBlocksToAddDeallocationsTo;
+  };
+
+  // Store deallocation info for a given allocation. Keyed on the allocation
+  // call.
+  DenseMap<const CallInst *, AllocationInfoEntry> Entries;
+
+  // Look up allocation matching a given deallocation.
+  DenseMap<const CallInst *, const CallInst *> DeallocationToAllocation;
+
+  // Deallocations that LICM has determined are safe to sink. They can be sunk
+  // only if the whole collection of Deallocations in an entry are sunk and its
+  // allocation are matching is hoisted.
+  DenseSet<const CallInst *> SafeToSink;
+};
+
+} // namespace llvm
+
+#endif

llvm/include/llvm/Transforms/Utils/LoopUtils.h

 
 namespace llvm {
 
 class AliasSet;
 class AliasSetTracker;
 class BasicBlock;
 class DataLayout;
 class Loop;
+class LoopAllocationInfo;
 class LoopInfo;
 class MemoryAccess;
 class MemorySSAUpdater;
 class OptimizationRemarkEmitter;
 class PredicatedScalarEvolution;
 class PredIteratorCache;
 class ScalarEvolution;
 class SCEV;

 /// uses before definitions, allowing us to sink a loop body in one pass without
 /// iteration. Takes DomTreeNode, AliasAnalysis, LoopInfo, DominatorTree,
 /// DataLayout, TargetLibraryInfo, Loop, AliasSet information for all
 /// instructions of the loop and loop safety information as
 /// arguments. Diagnostics is emitted via \p ORE. It returns changed status.
 bool sinkRegion(DomTreeNode *, AliasAnalysis *, LoopInfo *, DominatorTree *,
                 TargetLibraryInfo *, TargetTransformInfo *, Loop *,
                 AliasSetTracker *, MemorySSAUpdater *, ICFLoopSafetyInfo *,
-                bool, int &, OptimizationRemarkEmitter *);
+                LoopAllocationInfo *, bool, int &, OptimizationRemarkEmitter *);
 
 /// Walk the specified region of the CFG (defined by all blocks
 /// dominated by the specified block, and that are in the current loop) in depth
 /// first order w.r.t the DominatorTree.  This allows us to visit definitions
 /// before uses, allowing us to hoist a loop body in one pass without iteration.
 /// Takes DomTreeNode, AliasAnalysis, LoopInfo, DominatorTree, DataLayout,
 /// TargetLibraryInfo, Loop, AliasSet information for all instructions of the
 /// loop and loop safety information as arguments. Diagnostics is emitted via \p
 /// ORE. It returns changed status.
 bool hoistRegion(DomTreeNode *, AliasAnalysis *, LoopInfo *, DominatorTree *,
                  TargetLibraryInfo *, Loop *, AliasSetTracker *,
-                 MemorySSAUpdater *, ICFLoopSafetyInfo *, bool, int &,
-                 OptimizationRemarkEmitter *);
+                 MemorySSAUpdater *, ICFLoopSafetyInfo *, LoopAllocationInfo *,
+                 bool, int &, OptimizationRemarkEmitter *);
 
 /// This function deletes dead loops. The caller of this function needs to
 /// guarantee that the loop is infact dead.
 /// The function requires a bunch or prerequisites to be present:
 ///   - The loop needs to be in LCSSA form
 ///   - The loop needs to have a Preheader
 ///   - A unique dedicated exit block must exist
 ///

llvm/lib/Analysis/CMakeLists.txt

add_llvm_library(LLVMAnalysis
   LazyBranchProbabilityInfo.cpp
   LazyBlockFrequencyInfo.cpp
   LazyCallGraph.cpp
   LazyValueInfo.cpp
   LegacyDivergenceAnalysis.cpp
   Lint.cpp
   Loads.cpp
   LoopAccessAnalysis.cpp
+  LoopAllocationInfo.cpp
   LoopAnalysisManager.cpp
   LoopUnrollAnalyzer.cpp
   LoopInfo.cpp
   LoopPass.cpp
   MemDepPrinter.cpp
   MemDerefPrinter.cpp
   MemoryBuiltins.cpp
   MemoryDependenceAnalysis.cpp

llvm/lib/Analysis/LoopAllocationInfo.cpp

+//===- LoopAllocationInfo.cpp - memory allocations inside loops -----------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+
+#include "llvm/Analysis/LoopAllocationInfo.h"
+
+#include "llvm/ADT/BitVector.h"
+#include "llvm/ADT/DenseMap.h"
+#include "llvm/ADT/DenseSet.h"
+#include "llvm/ADT/STLExtras.h"
+#include "llvm/Analysis/CFG.h"
+#include "llvm/Analysis/LoopInfo.h"
+#include "llvm/Analysis/MemoryBuiltins.h"
+#include "llvm/Analysis/ValueTracking.h"
+#include "llvm/IR/BasicBlock.h"
+#include "llvm/IR/CFG.h"
+#include "llvm/IR/Dominators.h"
+#include "llvm/IR/Instruction.h"
+#include "llvm/IR/Instructions.h"
+#include "llvm/IR/PatternMatch.h"
+#include "llvm/Support/Casting.h"
+
+namespace llvm {
+
+/// Track ptr = allocfn to one or more freefn(ptr) relationships within the
+/// loop.
+///
+/// This analysis stores information about allocation and deallocation calls
+/// where the pointer is only live within the loop.
+///
+/// An allocation must be both hoisted and sunk, or neither. LICM performs
+/// sinking first, then hoisting. As sinking and hoisting is performed, more
+/// instructions will appear to be loop invariant, making more opportunities.
+/// For this reason we split the analysis into two pieces:
+///   1. for a given pointer that is only live in the loop, retain a list its
+///      matching deallocating instructions.
+///   2. during sinking, if we would sink the deallocation, don't, but inform
+///      the LoopAllocationInfo. If all deallocations would have been sunk, and
+///      hoistSink would hoist the allocation, perform both the hoist and sink
+///      the various deallocation instructions.
+///
+/// We don't need guaranteed execution of any deallocation function during
+/// unwinding. Before this transformation, an unwind mid-loop would leak one
+/// pointer that was allocated at the start of the loop. After this
+/// transformation, it would leak the one pointer allocated before the loop
+/// began. There is no difference visible to the caller.
+///
+/// We do have to ensure that the maximum amount of allocated memory is not
+/// raised by this transformation. To ensure this, we stop looking for
+/// allocations after seeing an opaque function call. For example:
+///   loop_top:
+///     %ptr1 = call i8* @malloc(i32 %loop_invariant1)  ;; candidate
+///     %ptr2 = call i8* @malloc(i32 %loop_variant2)    ;; not candidate
+///     %ptr3 = call i8* @malloc(i32 %loop_invariant3)  ;; candidate
+///     call void @opaque()                             ;; stop scanning
+///     %ptr5 = call i8* @malloc(i32 %loop_invariant5)  ;; not candidate
+///
+/// The same idea applies in reverse order to @free calls at the loop exits:
+///   loop_exiting:
+///      call void @free(%ptr1)  ;; not candidate
+///      call void @opaque()     ;; stop scanning
+///      call void @free(%ptr2)  ;; %ptr2 is candidate
+///                              ;; start here and scan upwards
+void LoopAllocationInfo::analyzeLoop(LoopInfo *LI, DominatorTree *DT,
+                                     TargetLibraryInfo *TLI, Loop *CurLoop) {
+  assert(LI != nullptr && DT != nullptr && CurLoop != nullptr &&
+         "Unexpected input to LoopAllocationInfo::analyzeLoop.");
+  if (!TLI)
+    return;
+
+  SmallVector<BasicBlock *, 32> LatchBlocks;
+  CurLoop->getLoopLatches(LatchBlocks);
+  if (LatchBlocks.empty())
+    return;
+
+  SmallVector<BasicBlock *, 32> ExitBlocks;
+  CurLoop->getUniqueExitBlocks(ExitBlocks);
+  if (ExitBlocks.empty())
+    return;
+
+  // Unlike hoistRegion, we only scan the header block. If the allocation is in
+  // a later block, it must be behind some sort of branch which implies that
+  // it's conditional on something or else another optimization would've removed
+  // the unconditional branch earlier).
+  BasicBlock *HeaderBB = CurLoop->getHeader();
+  SmallPtrSet<BasicBlock *, 1> HeaderBlock;
+  HeaderBlock.insert(HeaderBB);
+  for (BasicBlock::iterator I = HeaderBB->begin(), E = HeaderBB->end(); I != E;
+       ++I) {
+    // Find a block of allocations that are guaranteed to execute, ignoring the
+    // fact that allocation functions are themselves able to exit the program or
+    // unwind.
+
+    // There are no allocation functions which are guaranteed to continue to
+    // execute the next instruction.
+    if (isGuaranteedToTransferExecutionToSuccessor(&*I)) {
+      // These two intrinsics and increase and decrease heap size respectively.
+      // Don't hoist over an allocating intrinsic in order to ensure that any
+      // out-of-memory conditions will occur on the same instruction, in case
+      // there are differences in how the instructions handle OOM. Don't hoist
+      // over a deallocating intrinsic to avoid increasing the maximum heap
+      // allocated.
+      if (match(&*I, PatternMatch::m_Intrinsic<
+                         Intrinsic::objc_autoreleasePoolPush>()) ||
+          match(
+              &*I,
+              PatternMatch::m_Intrinsic<Intrinsic::objc_autoreleasePoolPop>()))
+        break;
+
+      continue;
+    }
+
+    if (!llvm::isAllocationFn(&*I, TLI))
+      break;
+    auto CI = cast<CallInst>(I);
+
+    // Find places in the loop where the pointer is certainly freed. The places
+    // we identify must free this allocation, but there may be any number of
+    // other places that could free our pointer which we miss.
+    //
+    // We use this to ensure that the allocation is freed before we go around
+    // any loop latch.
+    SmallVector<CallInst *, 16> FreeCalls;
+    constexpr size_t UseVisitThreshold = 20;
+    SmallVector<Instruction *, UseVisitThreshold> Worklist;
+    SmallPtrSet<Instruction *, UseVisitThreshold> Visited;
+    auto Enqueue = [&](const Instruction *I) {
+      for (const Use &U : I->uses()) {
+        Instruction *User = cast<Instruction>(U.getUser());
+        if (!CurLoop->contains(User))
+          continue;
+        if (!Visited.insert(User).second)
+          continue;
+        if (Visited.size() == UseVisitThreshold)
+          return;
+        Worklist.push_back(User);
+      }
+    };
+    Enqueue(CI);
+    do {
+      Instruction *I = Worklist.pop_back_val();
+      if (isa<BitCastInst>(I)) {
+        Enqueue(I);
+      } else if (auto *GEP = dyn_cast<GetElementPtrInst>(I)) {
+        if (GEP->hasAllZeroIndices())
+          Enqueue(GEP);
+      } else if (auto *PN = dyn_cast<PHINode>(I)) {
+        if (PN->hasConstantOrUndefValue())
+          Enqueue(I);
+      } else if (llvm::isFreeCall(I, TLI)) {
+        FreeCalls.push_back(cast<CallInst>(I));
+      }
+    } while (!Worklist.empty() && Visited.size() < UseVisitThreshold);
+
+    if (FreeCalls.empty())
+      continue;
+
+    SmallPtrSet<BasicBlock *, 32> FreeCallBBs;
+    for (auto FreeCall : FreeCalls)
+      FreeCallBBs.insert(FreeCall->getParent());
+
+    bool DoNotTransform = false;
+
+    // Check that the frees we found cover all the latches.
+    for (BasicBlock *LatchBB : LatchBlocks) {
+      if (FreeCallBBs.count(LatchBB))
+        continue;
+      SmallVector<BasicBlock *, 32> HeaderWorklist;
+      HeaderWorklist.push_back(HeaderBB);
+      if (llvm::isPotentiallyReachableFromMany(HeaderWorklist, LatchBB,
+                                               &FreeCallBBs, DT, LI)) {
+        DoNotTransform = true;
+        break;
+      }
+    }
+    if (DoNotTransform)
+      continue;
+
+    // Determine whether each exit block is reached with the pointer freed or
+    // not. If it can be reached with the pointer conditionally freed, we abort
+    // the transform.
+    SmallVector<BasicBlock *, 16> InsertDeallocation;
+    for (auto ExitBB : ExitBlocks) {
+      SmallVector<BasicBlock *, 32> HeaderWorklist, FreeCallWorklist;
+      HeaderWorklist.push_back(HeaderBB);
+      FreeCallWorklist.insert(FreeCallWorklist.end(), FreeCallBBs.begin(),
+                              FreeCallBBs.end());
+      // Starting at the header, is there a path to ExitBB without passing
+      // through any frees?
+      bool ReachExitBBWithoutFree = llvm::isPotentiallyReachableFromMany(
+          HeaderWorklist, ExitBB, &FreeCallBBs, DT, LI);
+      // Starting from each of the frees, is there a path to ExitBB without
+      // passing back through the header, or is this a single-block loop where
+      // the malloc and free are in both in the header?
+      bool ReachExitBBWithFree =
+          FreeCallBBs.count(HeaderBB) ||
+          llvm::isPotentiallyReachableFromMany(FreeCallWorklist, ExitBB,
+                                               &HeaderBlock, DT, LI);
+      assert((ReachExitBBWithoutFree || ReachExitBBWithFree) &&
+             "loop exit blocks not reachable from loop");
+      if (ReachExitBBWithoutFree && ReachExitBBWithFree) {
+        // These exist paths to this exit BB where the allocation is freed and
+        // not freed.
+        DoNotTransform = true;
+        break;
+      }
+      if (ReachExitBBWithFree) {
+        // Some blocks, such as those with a catchswitch instruction, can't
+        // have a free call inserted in them. Give up transforming the
+        // allocation in that case.
+        if (ExitBB->getFirstInsertionPt() == ExitBB->end()) {
+          DoNotTransform = true;
+          break;
+        }
+        // We'll need to insert a free-call into this exit BB if we hoist the
+        // allocation out of the loop.
+        InsertDeallocation.push_back(ExitBB);
+      }
+    }
+    if (DoNotTransform)
+      continue;
+
+    for (const CallInst *FreeCall : FreeCalls)
+      DeallocationToAllocation[FreeCall] = CI;
+    Entries.try_emplace(CI, std::move(FreeCalls),
+                        std::move(InsertDeallocation));
+  }
+}
+
+// Given an allocation call, are all the safety conditions met to hoist this
+// allocation and sink its deallocations.
+bool LoopAllocationInfo::mayHoist(const CallInst *CI) const {
+  auto it = Entries.find(CI);
+  if (it == Entries.end())
+    return false;
+  for (const CallInst *FreeCall : it->second.Deallocations) {
+    if (!SafeToSink.count(FreeCall))
+      return false;
+  }
+  return true;
+}
+
+} // namespace llvm

llvm/lib/Transforms/Scalar/LICM.cpp

 #include "llvm/Analysis/AliasAnalysis.h"
 #include "llvm/Analysis/AliasSetTracker.h"
 #include "llvm/Analysis/BasicAliasAnalysis.h"
 #include "llvm/Analysis/CaptureTracking.h"
 #include "llvm/Analysis/ConstantFolding.h"
 #include "llvm/Analysis/GlobalsModRef.h"
 #include "llvm/Analysis/GuardUtils.h"
 #include "llvm/Analysis/Loads.h"
+#include "llvm/Analysis/LoopAllocationInfo.h"
 #include "llvm/Analysis/LoopInfo.h"
 #include "llvm/Analysis/LoopIterator.h"
 #include "llvm/Analysis/LoopPass.h"
 #include "llvm/Analysis/MemoryBuiltins.h"
 #include "llvm/Analysis/MemorySSA.h"
 #include "llvm/Analysis/MemorySSAUpdater.h"
 #include "llvm/Analysis/OptimizationRemarkEmitter.h"
 #include "llvm/Analysis/ScalarEvolution.h"
 #include "llvm/Analysis/ScalarEvolutionAliasAnalysis.h"
 #include "llvm/Analysis/TargetLibraryInfo.h"
 #include "llvm/Analysis/ValueTracking.h"
 #include "llvm/IR/CFG.h"
 #include "llvm/IR/Constants.h"
 #include "llvm/IR/DataLayout.h"
 #include "llvm/IR/DerivedTypes.h"
 #include "llvm/IR/Dominators.h"
+#include "llvm/IR/IRBuilder.h"
 #include "llvm/IR/Instructions.h"
 #include "llvm/IR/IntrinsicInst.h"
 #include "llvm/IR/LLVMContext.h"
 #include "llvm/IR/Metadata.h"
 #include "llvm/IR/PatternMatch.h"
 #include "llvm/IR/PredIteratorCache.h"
 #include "llvm/Support/CommandLine.h"
 #include "llvm/Support/Debug.h"
bool LoopInvariantCodeMotion::runOnLoop(
 
   // Get the preheader block to move instructions into...
   BasicBlock *Preheader = L->getLoopPreheader();
 
   // Compute loop safety information.
   ICFLoopSafetyInfo SafetyInfo(DT);
   SafetyInfo.computeLoopSafetyInfo(L);
 
+  // Scan for matching malloc/free pairs.
+  LoopAllocationInfo LAI;
+  if (L->hasDedicatedExits() && Preheader)
+    LAI.analyzeLoop(LI, DT, TLI, L);
+
   // We want to visit all of the instructions in this loop... that are not parts
   // of our subloops (they have already had their invariants hoisted out of
   // their loop, into this loop, so there is no need to process the BODIES of
   // the subloops).
   //
   // Traverse the body of the loop in depth first order on the dominator tree so
   // that we are guaranteed to see definitions before we see uses.  This allows
   // us to sink instructions in one pass, without iteration.  After sinking
   // instructions, we perform another pass to hoist them out of the loop.
   //
   if (L->hasDedicatedExits())
     Changed |= sinkRegion(DT->getNode(L->getHeader()), AA, LI, DT, TLI, TTI, L,
-                          CurAST.get(), MSSAU.get(), &SafetyInfo,
+                          CurAST.get(), MSSAU.get(), &SafetyInfo, &LAI,
                           NoOfMemAccTooLarge, LicmMssaOptCounter, ORE);
   if (Preheader)
     Changed |= hoistRegion(DT->getNode(L->getHeader()), AA, LI, DT, TLI, L,
-                           CurAST.get(), MSSAU.get(), &SafetyInfo,
+                           CurAST.get(), MSSAU.get(), &SafetyInfo, &LAI,
                            NoOfMemAccTooLarge, LicmMssaOptCounter, ORE);
 
   // Now that all loop invariants have been removed from the loop, promote any
   // memory references to scalars that we can.
   // Don't sink stores from loops without dedicated block exits. Exits
   // containing indirect branches are not transformed by loop simplify,
   // make sure we catch that. An additional load may be generated in the
   // preheader for SSA updater, so also avoid sinking when no preheader

 /// the specified block, and that are in the current loop) in reverse depth
 /// first order w.r.t the DominatorTree.  This allows us to visit uses before
 /// definitions, allowing us to sink a loop body in one pass without iteration.
 ///
 bool llvm::sinkRegion(DomTreeNode *N, AliasAnalysis *AA, LoopInfo *LI,
                       DominatorTree *DT, TargetLibraryInfo *TLI,
                       TargetTransformInfo *TTI, Loop *CurLoop,
                       AliasSetTracker *CurAST, MemorySSAUpdater *MSSAU,
-                      ICFLoopSafetyInfo *SafetyInfo, bool NoOfMemAccTooLarge,
-                      int &LicmMssaOptCounter, OptimizationRemarkEmitter *ORE) {
+                      ICFLoopSafetyInfo *SafetyInfo, LoopAllocationInfo *LAI,
+                      bool NoOfMemAccTooLarge, int &LicmMssaOptCounter,
+                      OptimizationRemarkEmitter *ORE) {
 
   // Verify inputs.
   assert(N != nullptr && AA != nullptr && LI != nullptr && DT != nullptr &&
-         CurLoop != nullptr && SafetyInfo != nullptr &&
+         CurLoop != nullptr && SafetyInfo != nullptr && LAI != nullptr &&
          "Unexpected input to sinkRegion.");
   assert(((CurAST != nullptr) ^ (MSSAU != nullptr)) &&
          "Either AliasSetTracker or MemorySSA should be initialized.");
 
   // We want to visit children before parents. We will enque all the parents
   // before their children in the worklist and process the worklist in reverse
   // order.
   SmallVector<DomTreeNode *, 16> Worklist = collectChildrenInLoop(N, CurLoop);
 
   bool Changed = false;
   for (DomTreeNode *DTN : reverse(Worklist)) {
     BasicBlock *BB = DTN->getBlock();
     // Only need to process the contents of this block if it is not part of a
     // subloop (which would already have been processed).
     if (inSubLoop(BB, CurLoop, LI))
       continue;
 
+    bool ScanForFrees = !LAI->empty();
     for (BasicBlock::iterator II = BB->end(); II != BB->begin();) {
       Instruction &I = *--II;
 
       // If the instruction is dead, we would try to sink it because it isn't
       // used in the loop, instead, just delete it.
       if (isInstructionTriviallyDead(&I, TLI)) {
         LLVM_DEBUG(dbgs() << "LICM deleting dead inst: " << I << '\n');
         salvageDebugInfo(I);
for (BasicBlock::iterator II = BB->end(); II != BB->begin();) {
         if (sink(I, LI, DT, CurLoop, SafetyInfo, MSSAU, ORE)) {
           if (!FreeInLoop) {
             ++II;
             eraseInstruction(I, *SafetyInfo, CurAST, MSSAU);
           }
           Changed = true;
         }
       }
+
+      if (ScanForFrees) {
+        if (llvm::isFreeCall(&I, TLI)) {
+          LAI->addSafeToSink(cast<CallInst>(&I));
+        } else if (isa<CallInst>(I) &&
+                   (!isa<IntrinsicInst>(I) ||
+                    match(&I, PatternMatch::m_Intrinsic<
+                                  Intrinsic::objc_autoreleasePoolPush>()))) {
+          // This call might allocate memory. Don't move frees which are before
+          // it to after it.
+          ScanForFrees = false;
+        }
+      }
     }
   }
   if (MSSAU && VerifyMemorySSA)
     MSSAU->getMemorySSA()->verifyMemorySSA();
   return Changed;
 }
 
 namespace {

 /// Walk the specified region of the CFG (defined by all blocks dominated by
 /// the specified block, and that are in the current loop) in depth first
 /// order w.r.t the DominatorTree.  This allows us to visit definitions before
 /// uses, allowing us to hoist a loop body in one pass without iteration.
 ///
 bool llvm::hoistRegion(DomTreeNode *N, AliasAnalysis *AA, LoopInfo *LI,
                        DominatorTree *DT, TargetLibraryInfo *TLI, Loop *CurLoop,
                        AliasSetTracker *CurAST, MemorySSAUpdater *MSSAU,
-                       ICFLoopSafetyInfo *SafetyInfo, bool NoOfMemAccTooLarge,
-                       int &LicmMssaOptCounter,
+                       ICFLoopSafetyInfo *SafetyInfo, LoopAllocationInfo *LAI,
+                       bool NoOfMemAccTooLarge, int &LicmMssaOptCounter,
                        OptimizationRemarkEmitter *ORE) {
   // Verify inputs.
   assert(N != nullptr && AA != nullptr && LI != nullptr && DT != nullptr &&
-         CurLoop != nullptr && SafetyInfo != nullptr &&
+         CurLoop != nullptr && SafetyInfo != nullptr && LAI != nullptr &&
          "Unexpected input to hoistRegion.");
   assert(((CurAST != nullptr) ^ (MSSAU != nullptr)) &&
          "Either AliasSetTracker or MemorySSA should be initialized.");
 
   ControlFlowHoister CFH(LI, DT, CurLoop, MSSAU);
 
   // Keep track of instructions that have been hoisted, as they may need to be
   // re-hoisted if they end up not dominating all of their uses.
for (BasicBlock::iterator II = BB->begin(), E = BB->end(); II != E;) {
               CurLoop->getLoopPreheader()->getTerminator())) {
         hoist(I, DT, CurLoop, CFH.getOrCreateHoistedBlock(BB), SafetyInfo,
               MSSAU, ORE);
         HoistedInstructions.push_back(&I);
         Changed = true;
         continue;
       }
 
+      // If it's a malloc/free pair, try hoisting it out to the preheader. We
+      // don't need to worry about intervening unwinding function calls in this
+      // case, so we can be more aggressive than for generic instructions.
+      if (CurLoop->hasLoopInvariantOperands(&I) &&
+          llvm::isAllocationFn(&I, TLI) && LAI->mayHoist(cast<CallInst>(&I))) {
+        IRBuilder<>(&I).CreateLifetimeStart(&I);
+        hoist(I, DT, CurLoop, CFH.getOrCreateHoistedBlock(BB), SafetyInfo,
+              MSSAU, ORE);
+        HoistedInstructions.push_back(&I);
+        auto CI = cast<CallInst>(&I);
+
+        // One deallocation will be cloned into all the necessary exit blocks
+        // then the deallocations will be deleted.
+        CallInst *Pattern = LAI->FreesForMalloc(CI)[0];
+        Pattern->dropUnknownNonDebugMetadata();
+
+        // In case it is freeing a copy of the pointer not available in all
+        // exits of the loop.
+        Pattern->setArgOperand(0, CI);
+
+        // Merge debug locations. We don't track which frees have paths to which
+        // exit blocks, so we merge all the debug info to put on all the frees
+        // being placed in the exit blocks.
+        for (auto FreeCall : LAI->FreesForMalloc(CI)) {
+          Pattern->applyMergedLocation(Pattern->getDebugLoc(),
+                                       FreeCall->getDebugLoc());
+        }
+
+        // Insert frees into the loop exit blocks where we know the pointer
+        // would have been freed.
+        for (auto ExitBlock : LAI->ExitBlocksWithPointerFreed(CI)) {
+          Pattern->clone()->insertBefore(&*ExitBlock->getFirstInsertionPt());
+        }
+
+        // Replace the original frees with a @lifetime.end marker.
+        for (auto FreeCall : LAI->FreesForMalloc(CI)) {
+          IRBuilder<>(FreeCall).CreateLifetimeEnd(&I);
+          eraseInstruction(*FreeCall, *SafetyInfo, CurAST, MSSAU);
+        }
+
+        Changed = true;
+        continue;
+      }
+
       // Attempt to remove floating point division out of the loop by
       // converting it to a reciprocal multiplication.
       if (I.getOpcode() == Instruction::FDiv &&
           CurLoop->isLoopInvariant(I.getOperand(1)) &&
           I.hasAllowReciprocal()) {
         auto Divisor = I.getOperand(1);
         auto One = llvm::ConstantFP::get(Divisor->getType(), 1.0);
         auto ReciprocalDivisor = BinaryOperator::CreateFDiv(One, Divisor);

llvm/test/Transforms/LICM/allocs.ll

+; RUN: opt -S -licm < %s | FileCheck %s
+; RUN: opt -S -enable-mssa-loop-dependency -licm < %s | FileCheck %s
+
+declare i8* @malloc(i32) nounwind
+declare void @free(i8* nocapture) nounwind
+
+declare void @use(i8*)
+declare i1 @expr()
+
+define void @test1(i32 %loop_invariant) {
+; CHECK-LABEL: @test1
+header:
+; CHECK-LABEL: header:
+; CHECK: @malloc
+; CHECK: br label %loop
+  br label %loop
+
+loop:
+; CHECK-LABEL: loop:
+; CHECK-NOT: @malloc
+; CHECK: @llvm.lifetime.start
+; CHECK-NOT: @malloc
+; CHECK: @use
+; CHECK-NOT: @free
+; CHECK: @llvm.lifetime.end
+; CHECK-NOT: @free
+  %ptr = call i8* @malloc(i32 %loop_invariant)
+  call void @use(i8* %ptr)
+  call void @free(i8* %ptr)
+  br i1 undef, label %loop, label %loopexit
+
+loopexit:
+; CHECK-LABEL: loopexit:
+; CHECK: @free
+  ret void
+}
+
+declare void @mightmalloc()
+
+define void @test2(i32 %loop_invariant) {
+; CHECK-LABEL: @test2
+header:
+; CHECK-LABEL: header:
+; CHECK-NOT: @malloc
+; CHECK: br label %loop
+  br label %loop
+
+loop:
+; CHECK-LABEL: loop:
+; CHECK: @malloc
+; CHECK: @use
+; CHECK: @free
+; CHECK: @mightmalloc
+  %ptr = call i8* @malloc(i32 %loop_invariant)
+  call void @use(i8* %ptr)
+  call void @free(i8* %ptr)
+  call void @mightmalloc()
+  br i1 undef, label %loop, label %loopexit
+
+loopexit:
+; CHECK-LABEL: loopexit:
+; CHECK-NOT: @free
+  ret void
+}
+
+define void @test3(i32 %loop_invariant) {
+; CHECK-LABEL: @test3
+header:
+; CHECK-LABEL: header:
+; CHECK-NOT: @malloc
+; CHECK: br label %loop
+  br label %loop
+
+loop:
+; CHECK-LABEL: loop:
+; CHECK: @mightmalloc
+; CHECK: @malloc
+; CHECK: @use
+; CHECK: @free
+  call void @mightmalloc()
+  %ptr = call i8* @malloc(i32 %loop_invariant)
+  call void @use(i8* %ptr)
+  call void @free(i8* %ptr)
+  br i1 undef, label %loop, label %loopexit
+
+loopexit:
+; CHECK-LABEL: loopexit:
+; CHECK-NOT: @free
+  ret void
+}
+
+define void @test4(i32 %loop_invariant) {
+; CHECK-LABEL: @test4
+header:
+; CHECK-LABEL: header:
+; CHECK: @malloc
+; CHECK: @malloc
+; CHECK: @malloc
+; CHECK: br label %loop
+  br label %loop
+
+loop:
+; CHECK-LABEL: loop:
+; CHECK-NOT: @malloc
+; CHECK: @llvm.lifetime.start
+; CHECK: @llvm.lifetime.start
+; CHECK: @llvm.lifetime.start
+; CHECK: @use
+; CHECK: @use
+; CHECK: @use
+; CHECK: @llvm.lifetime.end
+; CHECK: @llvm.lifetime.end
+; CHECK: @llvm.lifetime.end
+; CHECK-NOT: @free
+  %ptr1 = call i8* @malloc(i32 %loop_invariant)
+  %ptr2 = call i8* @malloc(i32 %loop_invariant)
+  %ptr3 = call i8* @malloc(i32 %loop_invariant)
+  call void @use(i8* %ptr1)
+  call void @use(i8* %ptr2)
+  call void @use(i8* %ptr3)
+  call void @free(i8* %ptr1)
+  call void @free(i8* %ptr2)
+  call void @free(i8* %ptr3)
+  br i1 undef, label %loop, label %loopexit
+
+loopexit:
+; CHECK-LABEL: loopexit:
+; CHECK: @free
+; CHECK: @free
+; CHECK: @free
+  ret void
+}
+
+define void @test5(i32 %loop_invariant) {
+; CHECK-LABEL: @test5
+header:
+; CHECK-LABEL: header:
+; CHECK: @malloc
+; CHECK: @malloc
+; CHECK: @malloc
+; CHECK: br label %loop
+  br label %loop
+
+loop:
+; CHECK-LABEL: loop:
+; CHECK-NOT: @malloc
+; CHECK: @llvm.lifetime.start
+; CHECK: @llvm.lifetime.start
+; CHECK: @llvm.lifetime.start
+; CHECK: @use
+; CHECK: @use
+; CHECK: @use
+; CHECK: @llvm.lifetime.end
+; CHECK: @llvm.lifetime.end
+; CHECK: @llvm.lifetime.end
+; CHECK-NOT: @free
+  %ptr3 = call i8* @malloc(i32 %loop_invariant)
+  %ptr2 = call i8* @malloc(i32 %loop_invariant)
+  %ptr1 = call i8* @malloc(i32 %loop_invariant)
+  call void @use(i8* %ptr1)
+  call void @use(i8* %ptr2)
+  call void @use(i8* %ptr3)
+  call void @free(i8* %ptr1)
+  call void @free(i8* %ptr2)
+  call void @free(i8* %ptr3)
+  br i1 undef, label %loop, label %loopexit
+
+loopexit:
+; CHECK-LABEL: loopexit:
+; CHECK: @free
+; CHECK: @free
+; CHECK: @free
+  ret void
+}
+
+define void @test6(i32 %loop_invariant) {
+; CHECK-LABEL: @test6
+header:
+; CHECK-LABEL: header:
+; CHECK: @malloc
+; CHECK: @malloc
+; CHECK-NOT: @malloc
+; CHECK: br label %loop
+  br label %loop
+
+loop:
+; CHECK-LABEL: loop:
+; CHECK: @llvm.lifetime.start
+; CHECK: @malloc
+; CHECK: @llvm.lifetime.start
+; CHECK: @use
+; CHECK: @use
+; CHECK: @use
+; CHECK: @llvm.lifetime.end
+; CHECK: @free
+; CHECK: @llvm.lifetime.end
+  %i = phi i32 [%loop_invariant, %header], [%i.1, %loop]
+  %ptr1 = call i8* @malloc(i32 %loop_invariant)
+  %ptr2 = call i8* @malloc(i32 %i)
+  %ptr3 = call i8* @malloc(i32 %loop_invariant)
+  call void @use(i8* %ptr1)
+  call void @use(i8* %ptr2)
+  call void @use(i8* %ptr3)
+  call void @free(i8* %ptr1)
+  call void @free(i8* %ptr2)
+  call void @free(i8* %ptr3)
+  %i.1 = add i32 %i, 1
+  br i1 undef, label %loop, label %loopexit
+
+loopexit:
+; CHECK-LABEL: loopexit:
+; CHECK: @free
+; CHECK: @free
+; CHECK-NOT: @free
+; CHECK: ret void
+  ret void
+}
+
+define void @test7(i32 %loop_invariant) {
+; CHECK-LABEL: @test7
+header:
+; CHECK-LABEL: header:
+; CHECK: @malloc
+; CHECK: br label %loop
+  br label %loop
+
+loop:
+; CHECK-LABEL: loop:
+; CHECK-NOT: @malloc
+; CHECK: @llvm.lifetime.start
+; CHECK-NOT: @malloc
+; CHECK: @use
+; CHECK-NOT: @free
+; CHECK-NOT: @llvm.lifetime.end
+  %i = phi i32 [%loop_invariant, %header], [%i.1, %loop.2], [%i.1, %loop.3]
+  %ptr = call i8* @malloc(i32 %loop_invariant)
+  call void @use(i8* %ptr)
+  %i.1 = add i32 %i, 1
+  %cmp = icmp eq i32 %i.1, 123
+  br i1 %cmp, label %loop.2, label %loop.3
+
+loop.2:
+; CHECK-LABEL: loop.2:
+; CHECK-NOT: @free
+; CHECK: @llvm.lifetime.end
+; CHECK-NOT: @free
+  %continue.1 = call i1 @expr()
+  call void @free(i8* %ptr)
+  br i1 %continue.1, label %loop, label %loopexit
+
+loop.3:
+; CHECK-LABEL: loop.3:
+; CHECK-NOT: @free
+; CHECK: @llvm.lifetime.end
+; CHECK-NOT: @free
+  %continue.2 = call i1 @expr()
+  call void @free(i8* %ptr)
+  br i1 %continue.2, label %loop, label %loopexit
+
+loopexit:
+; CHECK-LABEL: loopexit:
+; CHECK: @free
+  ret void
+}
+
+declare void @assert_fail()
+
+define void @test8(i32 %loop_invariant) {
+; CHECK-LABEL: @test8
+header:
+; CHECK-LABEL: header:
+; CHECK: @malloc
+; CHECK: br label %loop
+  br label %loop
+
+loop:
+; CHECK-LABEL: loop:
+; CHECK-NOT: @malloc
+; CHECK: @llvm.lifetime.start
+; CHECK-NOT: @malloc
+  %ptr = call i8* @malloc(i32 %loop_invariant)
+  %fail = call i1 @expr()
+  br i1 %fail, label %assert_fail, label %assert_pass
+
+assert_fail:
+  call void @assert_fail()
+  unreachable
+
+assert_pass:
+; CHECK-LABEL: assert_pass:
+; CHECK-NOT: @free
+; CHECK: @llvm.lifetime.end
+; CHECK-NOT: @free
+  call void @free(i8* %ptr)
+  br i1 undef, label %loop, label %loopexit
+
+loopexit:
+; CHECK-LABEL: loopexit:
+; CHECK: @free
+  ret void
+}
+
+declare void @usei64(i64)
+
+define void @test9(i32 %loop_invariant) {
+; CHECK-LABEL: @test9
+header:
+; CHECK-LABEL: header:
+; CHECK: add
+; CHECK: @malloc
+; CHECK: ptrtoint
+; CHECK: and
+; CHECK-NOT: usei64
+; CHECK: br label %loop
+  br label %loop
+
+loop:
+; CHECK-LABEL: loop:
+; CHECK-NOT: @malloc
+; CHECK: @llvm.lifetime.start
+; CHECK-NOT: @malloc
+; CHECK: usei64
+  %size = add i32 %loop_invariant, 1
+  %ptr = call i8* @malloc(i32 %size)
+  %ptrbits = ptrtoint i8* %ptr to i64
+  %aligncheck = and i64 %ptrbits, 4
+  call void @usei64(i64%aligncheck)
+  %fail = call i1 @expr()
+  br i1 %fail, label %assert_fail, label %assert_pass
+
+assert_fail:
+  call void @assert_fail()
+  unreachable
+
+assert_pass:
+; CHECK-LABEL: assert_pass:
+; CHECK-NOT: @free
+; CHECK: @llvm.lifetime.end
+; CHECK-NOT: @free
+  call void @free(i8* %ptr)
+  br i1 undef, label %loop, label %loopexit
+
+loopexit:
+; CHECK-LABEL: loopexit:
+; CHECK: @free
+  ret void
+}
+
+declare x86_stdcallcc void @_CxxThrowException(i8*, i8*)
+
+declare i32 @__CxxFrameHandler3(...)
+
+define void @test10(i32 %loop_invariant) personality i32 (...)* @__CxxFrameHandler3 {
+; CHECK-LABEL: @test10
+header:
+  br label %loop
+
+loop:
+; CHECK-LABEL: loop:
+; CHECK: @malloc
+; CHECK: @use
+; CHECK: @free
+  %ptr = call i8* @malloc(i32 %loop_invariant)
+  call void @use(i8* %ptr)
+  call void @free(i8* %ptr)
+  invoke void @_CxxThrowException(i8* null, i8* null)
+          to label %continue unwind label %catch.dispatch
+
+continue:
+  %take.backedge = call i1 @expr()
+  br i1 %take.backedge, label %loop, label %loopexit
+
+catch.dispatch:
+  %cs = catchswitch within none [label %catch] unwind to caller
+
+catch:                                            ; preds = %catch.dispatch
+  %cp = catchpad within %cs [i8* null, i32 64, i8* null]
+  catchret from %cp to label %loopexit
+
+loopexit:
+  ret void
+}
+
+define void @test11(i32 %loop_invariant) personality i32 (...)* @__CxxFrameHandler3 {
+; CHECK-LABEL: @test11
+header:
+; CHECK-LABEL: header:
+; CHECK: @malloc
+; CHECK: br label %loop
+  br label %loop
+
+loop:
+; CHECK-LABEL: loop:
+; CHECK-NOT: @malloc
+; CHECK: @llvm.lifetime.start
+; CHECK-NOT: @malloc
+; CHECK: @use
+  %ptr = call i8* @malloc(i32 %loop_invariant)
+  call void @use(i8* %ptr)
+  invoke void @_CxxThrowException(i8* null, i8* null)
+          to label %continue unwind label %catch.dispatch
+
+continue:
+; CHECK_LABEL: continue:
+; CHECK-NOT: free
+; CHECK: lifetime.end
+; CHECK-NOT: free
+  %take.backedge = call i1 @expr()
+  call void @free(i8* %ptr)
+  br i1 %take.backedge, label %loop, label %loopexit
+
+catch.dispatch:
+; CHECK-LABEL: catch.dispatch:
+; CHECK-NEXT: catchswitch
+  %cs = catchswitch within none [label %catch] unwind to caller
+
+catch:                                            ; preds = %catch.dispatch
+  %cp = catchpad within %cs [i8* null, i32 64, i8* null]
+  catchret from %cp to label %loopexit
+
+loopexit:
+; CHECK-LABEL: loopexit:
+; CHECK: @free
+  ret void
+}
+
+declare float @llvm.log.f32(float %Val)
+
+define void @test12(i32 %loop_invariant) {
+; CHECK-LABEL: @test1
+header:
+; CHECK-LABEL: header:
+; CHECK: @malloc
+; CHECK: br label %loop
+  br label %loop
+
+loop:
+; CHECK-LABEL: loop:
+; CHECK-NOT: @malloc
+; CHECK: @llvm.lifetime.start
+; CHECK-NOT: @malloc
+; CHECK: @use
+; CHECK-NOT: @free
+; CHECK: @llvm.lifetime.end
+; CHECK-NOT: @free
+  call float @llvm.log.f32(float 0.0)
+  %ptr = call i8* @malloc(i32 %loop_invariant)
+  call void @use(i8* %ptr)
+  call float @llvm.log.f32(float 0.0)
+  call void @free(i8* %ptr)
+  call float @llvm.log.f32(float 0.0)
+  br i1 undef, label %loop, label %loopexit
+
+loopexit:
+; CHECK-LABEL: loopexit:
+; CHECK: @free
+  ret void
+}
+
+declare void @llvm.objc.autoreleasePoolPop(i8*)
+declare i8* @llvm.objc.autoreleasePoolPush()
+
+define void @test13(i32 %loop_invariant) {
+; CHECK-LABEL: @test13
+header:
+; CHECK-LABEL: header:
+; CHECK-NOT: @malloc
+; CHECK: br label %loop
+  br label %loop
+
+loop:
+; CHECK-LABEL: loop:
+; CHECK: @malloc
+; CHECK: @use
+; CHECK: @free
+  %managed_ptr = call i8* @llvm.objc.autoreleasePoolPush()
+  %ptr = call i8* @malloc(i32 %loop_invariant)
+  call void @use(i8* %ptr)
+  call void @free(i8* %ptr)
+  br i1 undef, label %loop, label %loopexit
+
+loopexit:
+; CHECK-LABEL: loopexit:
+; CHECK-NOT: @free
+  ret void
+}
+
+define void @test14(i32 %loop_invariant) {
+; CHECK-LABEL: @test14
+header:
+; CHECK-LABEL: header:
+; CHECK-NOT: @malloc
+; CHECK: br label %loop
+  br label %loop
+
+loop:
+; CHECK-LABEL: loop:
+; CHECK: @malloc
+; CHECK: @use
+; CHECK: @free
+  %ptr = call i8* @malloc(i32 %loop_invariant)
+  call void @use(i8* %ptr)
+  call void @free(i8* %ptr)
+  %managed_ptr = call i8* @llvm.objc.autoreleasePoolPush()
+  br i1 undef, label %loop, label %loopexit
+
+loopexit:
+; CHECK-LABEL: loopexit:
+; CHECK-NOT: @free
+  ret void
+}
+
+declare i8 @expr8()
+
+define void @test15(i32 %loop_invariant) {
+; CHECK-LABEL: @test15
+header:
+; CHECK-LABEL: header:
+; CHECK: @malloc
+; CHECK: br label %loop
+  br label %loop
+
+loop:
+  %ptr = call i8* @malloc(i32 %loop_invariant)
+  call void @use(i8* %ptr)
+  %Val = call i8 @expr8()
+  switch i8 %Val, label %b1 [ i8 2, label %b2
+                              i8 3, label %b3 ]
+
+b1:
+  call void @usei64(i64 1)
+  %a = call i1 @expr()
+  call void @free(i8* %ptr), !dbg !6
+  br i1 %a, label %loop, label %loopexit
+
+b2:
+  call void @usei64(i64 2)
+  %b = call i1 @expr()
+  call void @free(i8* %ptr), !dbg !7
+  br i1 %b, label %loop, label %loopexit
+
+b3:
+  call void @usei64(i64 3)
+  %c = call i1 @expr()
+  call void @free(i8* %ptr), !dbg !8
+  br i1 %c, label %loop, label %loopexit
+
+loopexit:
+; CHECK-LABEL: loopexit:
+; CHECK: @free{{.*}}, !dbg [[test15MergedLoc:![0-9]+]]
+  ret void
+}
+
+define void @test16(i32 %loop_invariant, i8* %vptr) {
+; CHECK-LABEL: @test16
+header:
+; CHECK-LABEL: header:
+; CHECK-NOT: @malloc
+; CHECK: br label %loop
+  br label %loop
+
+loop:
+; CHECK-LABEL: loop:
+; CHECK: @malloc
+; CHECK: @use
+; CHECK: @free
+  %Val = load volatile i8, i8* %vptr
+  %ptr = call i8* @malloc(i32 %loop_invariant)
+  call void @use(i8* %ptr)
+  call void @free(i8* %ptr)
+  br i1 undef, label %loop, label %loopexit
+
+loopexit:
+; CHECK-LABEL: loopexit:
+; CHECK-NOT: @free
+  ret void
+}
+
+define void @test17(i32 %loop_invariant) {
+; CHECK-LABEL: @test17
+header:
+; CHECK-LABEL: header:
+; CHECK: @malloc
+; CHECK: bitcast
+; CHECK: getelementptr
+; CHECK: br label %loop
+  br label %loop
+
+loop:
+; CHECK-LABEL: loop:
+; CHECK-NOT: @malloc
+; CHECK: @llvm.lifetime.start
+; CHECK-NOT: @malloc
+  %ptr = call i8* @malloc(i32 %loop_invariant)
+  %ptr2 = bitcast i8* %ptr to i8*
+  %ptr3 = getelementptr i8, i8* %ptr2, i32 0
+br label %nextloopblock
+
+nextloopblock:
+  %ptr4 = phi i8* [%ptr3, %loop]
+  call void @free(i8* %ptr4)
+  br i1 undef, label %loop, label %loopexit
+
+loopexit:
+; CHECK-LABEL: loopexit:
+; CHECK: @free
+  ret void
+}
+
+define void @test18(i32 %loop_invariant) {
+; CHECK-LABEL: @test18
+header:
+; CHECK-LABEL: header:
+; CHECK: @malloc
+; CHECK: br label %loop
+  br label %loop
+
+loop:
+; CHECK-LABEL: loop:
+; CHECK-NOT: @malloc
+; CHECK: @llvm.lifetime.start
+; CHECK-NOT: @malloc
+  %ptr = call i8* @malloc(i32 %loop_invariant)
+  %continue = call i1 @expr()
+  %branch = call i1 @expr()
+  br i1 %branch, label %nextbb1, label %nextbb2
+
+nextbb1:
+  %ptr2 = bitcast i8* %ptr to i8*
+  call void @free(i8* %ptr2)
+  br label %nextloopblock
+
+nextbb2:
+  call void @free(i8* %ptr)
+  br label %nextloopblock
+
+nextloopblock:
+  br i1 %continue, label %loop, label %loopexit
+
+loopexit:
+; CHECK-LABEL: loopexit:
+; CHECK-NOT: undef
+; CHECK: @free
+  ret void
+}
+
+!llvm.module.flags = !{!0, !1, !2}
+!llvm.dbg.cu = !{!3}
+
+; CHECK: [[test15MergedLoc]] = !DILocation(line: 0
+!0 = !{i32 2, !"Dwarf Version", i32 4}
+!1 = !{i32 2, !"Debug Info Version", i32 3}
+!2 = !{i32 1, !"PIC Level", i32 2}
+!3 = distinct !DICompileUnit(language: DW_LANG_C99, file: !4)
+!4 = !DIFile(filename: "allocs.ll", directory: "test/Transforms/LICM")
+!5 = distinct !DISubprogram(name: "test/Transforms/LICM/allocs.ll", unit: !3)
+!6 = !DILocation(line: 1, column: 10, scope: !5)
+!7 = !DILocation(line: 2, column: 20, scope: !5)
+!8 = !DILocation(line: 3, column: 30, scope: !5)