Use estimated number of optimized insns in unroll-threshold computation.
[oota-llvm.git] / lib / Transforms / Scalar / LoopUnrollPass.cpp
index 2e94a44cad7282b325589f679403b23b3dc9eeed..f691beb721971f8ba830b7e6be94964f2dbc275a 100644 (file)
 //===----------------------------------------------------------------------===//
 
 #include "llvm/Transforms/Scalar.h"
+#include "llvm/Analysis/AssumptionCache.h"
 #include "llvm/Analysis/CodeMetrics.h"
 #include "llvm/Analysis/LoopPass.h"
 #include "llvm/Analysis/ScalarEvolution.h"
+#include "llvm/Analysis/ScalarEvolutionExpressions.h"
 #include "llvm/Analysis/TargetTransformInfo.h"
 #include "llvm/IR/DataLayout.h"
 #include "llvm/IR/DiagnosticInfo.h"
@@ -26,6 +28,8 @@
 #include "llvm/Support/Debug.h"
 #include "llvm/Support/raw_ostream.h"
 #include "llvm/Transforms/Utils/UnrollLoop.h"
+#include "llvm/IR/InstVisitor.h"
+#include "llvm/Analysis/InstructionSimplify.h"
 #include <climits>
 
 using namespace llvm;
@@ -36,6 +40,22 @@ static cl::opt<unsigned>
 UnrollThreshold("unroll-threshold", cl::init(150), cl::Hidden,
   cl::desc("The cut-off point for automatic loop unrolling"));
 
+static cl::opt<unsigned> UnrollMaxIterationsCountToAnalyze(
+    "unroll-max-iteration-count-to-analyze", cl::init(1000), cl::Hidden,
+    cl::desc("Don't allow loop unrolling to simulate more than this number of"
+             "iterations when checking full unroll profitability"));
+
+static cl::opt<unsigned> UnrollMinPercentOfOptimized(
+    "unroll-percent-of-optimized-for-complete-unroll", cl::init(20), cl::Hidden,
+    cl::desc("If complete unrolling could trigger further optimizations, and, "
+             "by that, remove the given percent of instructions, perform the "
+             "complete unroll even if it's beyond the threshold"));
+
+static cl::opt<unsigned> UnrollAbsoluteThreshold(
+    "unroll-absolute-threshold", cl::init(2000), cl::Hidden,
+    cl::desc("Don't unroll if the unrolled size is bigger than this threshold,"
+             " even if we can remove big portion of instructions later."));
+
 static cl::opt<unsigned>
 UnrollCount("unroll-count", cl::init(0), cl::Hidden,
   cl::desc("Use this unroll count for all loops including those with "
@@ -61,11 +81,16 @@ namespace {
     static char ID; // Pass ID, replacement for typeid
     LoopUnroll(int T = -1, int C = -1, int P = -1, int R = -1) : LoopPass(ID) {
       CurrentThreshold = (T == -1) ? UnrollThreshold : unsigned(T);
+      CurrentAbsoluteThreshold = UnrollAbsoluteThreshold;
+      CurrentMinPercentOfOptimized = UnrollMinPercentOfOptimized;
       CurrentCount = (C == -1) ? UnrollCount : unsigned(C);
       CurrentAllowPartial = (P == -1) ? UnrollAllowPartial : (bool)P;
       CurrentRuntime = (R == -1) ? UnrollRuntime : (bool)R;
 
       UserThreshold = (T != -1) || (UnrollThreshold.getNumOccurrences() > 0);
+      UserAbsoluteThreshold = (UnrollAbsoluteThreshold.getNumOccurrences() > 0);
+      UserPercentOfOptimized =
+          (UnrollMinPercentOfOptimized.getNumOccurrences() > 0);
       UserAllowPartial = (P != -1) ||
                          (UnrollAllowPartial.getNumOccurrences() > 0);
       UserRuntime = (R != -1) || (UnrollRuntime.getNumOccurrences() > 0);
@@ -89,10 +114,16 @@ namespace {
 
     unsigned CurrentCount;
     unsigned CurrentThreshold;
+    unsigned CurrentAbsoluteThreshold;
+    unsigned CurrentMinPercentOfOptimized;
     bool     CurrentAllowPartial;
     bool     CurrentRuntime;
     bool     UserCount;            // CurrentCount is user-specified.
     bool     UserThreshold;        // CurrentThreshold is user-specified.
+    bool UserAbsoluteThreshold;    // CurrentAbsoluteThreshold is
+                                   // user-specified.
+    bool UserPercentOfOptimized;   // CurrentMinPercentOfOptimized is
+                                   // user-specified.
     bool     UserAllowPartial;     // CurrentAllowPartial is user-specified.
     bool     UserRuntime;          // CurrentRuntime is user-specified.
 
@@ -102,15 +133,16 @@ namespace {
     /// loop preheaders be inserted into the CFG...
     ///
     void getAnalysisUsage(AnalysisUsage &AU) const override {
-      AU.addRequired<LoopInfo>();
-      AU.addPreserved<LoopInfo>();
+      AU.addRequired<AssumptionCacheTracker>();
+      AU.addRequired<LoopInfoWrapperPass>();
+      AU.addPreserved<LoopInfoWrapperPass>();
       AU.addRequiredID(LoopSimplifyID);
       AU.addPreservedID(LoopSimplifyID);
       AU.addRequiredID(LCSSAID);
       AU.addPreservedID(LCSSAID);
       AU.addRequired<ScalarEvolution>();
       AU.addPreserved<ScalarEvolution>();
-      AU.addRequired<TargetTransformInfo>();
+      AU.addRequired<TargetTransformInfoWrapperPass>();
       // FIXME: Loop unroll requires LCSSA. And LCSSA requires dom info.
       // If loop unroll does not preserve dom info then LCSSA pass on next
       // loop will receive invalid dom info.
@@ -123,6 +155,8 @@ namespace {
     void getUnrollingPreferences(Loop *L, const TargetTransformInfo &TTI,
                                  TargetTransformInfo::UnrollingPreferences &UP) {
       UP.Threshold = CurrentThreshold;
+      UP.AbsoluteThreshold = CurrentAbsoluteThreshold;
+      UP.MinPercentOfOptimized = CurrentMinPercentOfOptimized;
       UP.OptSizeThreshold = OptSizeUnrollThreshold;
       UP.PartialThreshold = CurrentThreshold;
       UP.PartialOptSizeThreshold = OptSizeUnrollThreshold;
@@ -149,13 +183,33 @@ namespace {
     // unrolled loops respectively.
     void selectThresholds(const Loop *L, bool HasPragma,
                           const TargetTransformInfo::UnrollingPreferences &UP,
-                          unsigned &Threshold, unsigned &PartialThreshold) {
+                          unsigned &Threshold, unsigned &PartialThreshold,
+                          unsigned NumberOfOptimizedInstructions) {
       // Determine the current unrolling threshold.  While this is
       // normally set from UnrollThreshold, it is overridden to a
       // smaller value if the current function is marked as
       // optimize-for-size, and the unroll threshold was not user
       // specified.
       Threshold = UserThreshold ? CurrentThreshold : UP.Threshold;
+
+      // If we are allowed to completely unroll if we can remove M% of
+      // instructions, and we know that with complete unrolling we'll be able
+      // to kill N instructions, then we can afford to completely unroll loops
+      // with unrolled size up to N*100/M.
+      // Adjust the threshold according to that:
+      unsigned PercentOfOptimizedForCompleteUnroll =
+          UserPercentOfOptimized ? CurrentMinPercentOfOptimized
+                                 : UP.MinPercentOfOptimized;
+      unsigned AbsoluteThreshold = UserAbsoluteThreshold
+                                       ? CurrentAbsoluteThreshold
+                                       : UP.AbsoluteThreshold;
+      if (PercentOfOptimizedForCompleteUnroll)
+        Threshold = std::max<unsigned>(Threshold,
+                                       NumberOfOptimizedInstructions * 100 /
+                                           PercentOfOptimizedForCompleteUnroll);
+      // But don't allow unrolling loops bigger than absolute threshold.
+      Threshold = std::min<unsigned>(Threshold, AbsoluteThreshold);
+
       PartialThreshold = UserThreshold ? CurrentThreshold : UP.PartialThreshold;
       if (!UserThreshold &&
           L->getHeader()->getParent()->getAttributes().
@@ -181,8 +235,9 @@ namespace {
 
 char LoopUnroll::ID = 0;
 INITIALIZE_PASS_BEGIN(LoopUnroll, "loop-unroll", "Unroll loops", false, false)
-INITIALIZE_AG_DEPENDENCY(TargetTransformInfo)
-INITIALIZE_PASS_DEPENDENCY(LoopInfo)
+INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass)
+INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
+INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
 INITIALIZE_PASS_DEPENDENCY(LCSSA)
 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
@@ -197,14 +252,337 @@ Pass *llvm::createSimpleLoopUnrollPass() {
   return llvm::createLoopUnrollPass(-1, -1, 0, 0);
 }
 
+static bool IsLoadFromConstantInitializer(Value *V) {
+  if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V))
+    if (GV->isConstant() && GV->hasDefinitiveInitializer())
+      return GV->getInitializer();
+  return false;
+}
+
+struct FindConstantPointers {
+  bool LoadCanBeConstantFolded;
+  bool IndexIsConstant;
+  APInt Step;
+  APInt StartValue;
+  Value *BaseAddress;
+  const Loop *L;
+  ScalarEvolution &SE;
+  FindConstantPointers(const Loop *loop, ScalarEvolution &SE)
+      : LoadCanBeConstantFolded(true), IndexIsConstant(true), L(loop), SE(SE) {}
+
+  bool follow(const SCEV *S) {
+    if (const SCEVUnknown *SC = dyn_cast<SCEVUnknown>(S)) {
+      // We've reached the leaf node of SCEV, it's most probably just a
+      // variable. Now it's time to see if it corresponds to a global constant
+      // global (in which case we can eliminate the load), or not.
+      BaseAddress = SC->getValue();
+      LoadCanBeConstantFolded =
+          IndexIsConstant && IsLoadFromConstantInitializer(BaseAddress);
+      return false;
+    }
+    if (isa<SCEVConstant>(S))
+      return true;
+    if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
+      // If the current SCEV expression is AddRec, and its loop isn't the loop
+      // we are about to unroll, then we won't get a constant address after
+      // unrolling, and thus, won't be able to eliminate the load.
+      if (AR->getLoop() != L)
+        return IndexIsConstant = false;
+      // If the step isn't constant, we won't get constant addresses in unrolled
+      // version. Bail out.
+      if (const SCEVConstant *StepSE =
+              dyn_cast<SCEVConstant>(AR->getStepRecurrence(SE)))
+        Step = StepSE->getValue()->getValue();
+      else
+        return IndexIsConstant = false;
+
+      return IndexIsConstant;
+    }
+    // If Result is true, continue traversal.
+    // Otherwise, we have found something that prevents us from (possible) load
+    // elimination.
+    return IndexIsConstant;
+  }
+  bool isDone() const { return !IndexIsConstant; }
+};
+
+// This class is used to get an estimate of the optimization effects that we
+// could get from complete loop unrolling. It comes from the fact that some
+// loads might be replaced with concrete constant values and that could trigger
+// a chain of instruction simplifications.
+//
+// E.g. we might have:
+//   int a[] = {0, 1, 0};
+//   v = 0;
+//   for (i = 0; i < 3; i ++)
+//     v += b[i]*a[i];
+// If we completely unroll the loop, we would get:
+//   v = b[0]*a[0] + b[1]*a[1] + b[2]*a[2]
+// Which then will be simplified to:
+//   v = b[0]* 0 + b[1]* 1 + b[2]* 0
+// And finally:
+//   v = b[1]
+class UnrollAnalyzer : public InstVisitor<UnrollAnalyzer, bool> {
+  typedef InstVisitor<UnrollAnalyzer, bool> Base;
+  friend class InstVisitor<UnrollAnalyzer, bool>;
+
+  const Loop *L;
+  unsigned TripCount;
+  ScalarEvolution &SE;
+  const TargetTransformInfo &TTI;
+  unsigned NumberOfOptimizedInstructions;
+
+  DenseMap<Value *, Constant *> SimplifiedValues;
+  DenseMap<LoadInst *, Value *> LoadBaseAddresses;
+  SmallPtrSet<Instruction *, 32> CountedInsns;
+
+  // Provide base case for our instruction visit.
+  bool visitInstruction(Instruction &I) { return false; };
+  // TODO: We should also visit ICmp, FCmp, GetElementPtr, Trunc, ZExt, SExt,
+  // FPTrunc, FPExt, FPToUI, FPToSI, UIToFP, SIToFP, BitCast, Select,
+  // ExtractElement, InsertElement, ShuffleVector, ExtractValue, InsertValue.
+  //
+  // Probaly it's worth to hoist the code for estimating the simplifications
+  // effects to a separate class, since we have a very similar code in
+  // InlineCost already.
+  bool visitBinaryOperator(BinaryOperator &I) {
+    Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
+    if (!isa<Constant>(LHS))
+      if (Constant *SimpleLHS = SimplifiedValues.lookup(LHS))
+        LHS = SimpleLHS;
+    if (!isa<Constant>(RHS))
+      if (Constant *SimpleRHS = SimplifiedValues.lookup(RHS))
+        RHS = SimpleRHS;
+    Value *SimpleV = nullptr;
+    if (auto FI = dyn_cast<FPMathOperator>(&I))
+      SimpleV =
+          SimplifyFPBinOp(I.getOpcode(), LHS, RHS, FI->getFastMathFlags());
+    else
+      SimpleV = SimplifyBinOp(I.getOpcode(), LHS, RHS);
+
+    if (SimpleV && CountedInsns.insert(&I).second)
+      NumberOfOptimizedInstructions += TTI.getUserCost(&I);
+
+    if (Constant *C = dyn_cast_or_null<Constant>(SimpleV)) {
+      SimplifiedValues[&I] = C;
+      return true;
+    }
+    return false;
+  }
+
+  Constant *computeLoadValue(LoadInst *LI, unsigned Iteration) {
+    if (!LI)
+      return nullptr;
+    Value *BaseAddr = LoadBaseAddresses[LI];
+    if (!BaseAddr)
+      return nullptr;
+
+    auto GV = dyn_cast<GlobalVariable>(BaseAddr);
+    if (!GV)
+      return nullptr;
+
+    ConstantDataSequential *CDS =
+        dyn_cast<ConstantDataSequential>(GV->getInitializer());
+    if (!CDS)
+      return nullptr;
+
+    const SCEV *BaseAddrSE = SE.getSCEV(BaseAddr);
+    const SCEV *S = SE.getSCEV(LI->getPointerOperand());
+    const SCEV *OffSE = SE.getMinusSCEV(S, BaseAddrSE);
+
+    APInt StepC, StartC;
+    const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(OffSE);
+    if (!AR)
+      return nullptr;
+
+    if (const SCEVConstant *StepSE =
+            dyn_cast<SCEVConstant>(AR->getStepRecurrence(SE)))
+      StepC = StepSE->getValue()->getValue();
+    else
+      return nullptr;
+
+    if (const SCEVConstant *StartSE = dyn_cast<SCEVConstant>(AR->getStart()))
+      StartC = StartSE->getValue()->getValue();
+    else
+      return nullptr;
+
+    unsigned ElemSize = CDS->getElementType()->getPrimitiveSizeInBits() / 8U;
+    unsigned Start = StartC.getLimitedValue();
+    unsigned Step = StepC.getLimitedValue();
+
+    unsigned Index = (Start + Step * Iteration) / ElemSize;
+    if (Index >= CDS->getNumElements())
+      return nullptr;
+
+    Constant *CV = CDS->getElementAsConstant(Index);
+
+    return CV;
+  }
+
+public:
+  UnrollAnalyzer(const Loop *L, unsigned TripCount, ScalarEvolution &SE,
+                 const TargetTransformInfo &TTI)
+      : L(L), TripCount(TripCount), SE(SE), TTI(TTI),
+        NumberOfOptimizedInstructions(0) {}
+
+  // Visit all loads the loop L, and for those that, after complete loop
+  // unrolling, would have a constant address and it will point to a known
+  // constant initializer, record its base address for future use.  It is used
+  // when we estimate number of potentially simplified instructions.
+  void FindConstFoldableLoads() {
+    for (auto BB : L->getBlocks()) {
+      for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
+        if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
+          if (!LI->isSimple())
+            continue;
+          Value *AddrOp = LI->getPointerOperand();
+          const SCEV *S = SE.getSCEV(AddrOp);
+          FindConstantPointers Visitor(L, SE);
+          SCEVTraversal<FindConstantPointers> T(Visitor);
+          T.visitAll(S);
+          if (Visitor.IndexIsConstant && Visitor.LoadCanBeConstantFolded) {
+            LoadBaseAddresses[LI] = Visitor.BaseAddress;
+          }
+        }
+      }
+    }
+  }
+
+  // Given a list of loads that could be constant-folded (LoadBaseAddresses),
+  // estimate number of optimized instructions after substituting the concrete
+  // values for the given Iteration.
+  // Fill in SimplifiedInsns map for future use in DCE-estimation.
+  unsigned EstimateNumberOfSimplifiedInsns(unsigned Iteration) {
+    SmallVector<Instruction *, 8> Worklist;
+    SimplifiedValues.clear();
+    CountedInsns.clear();
+
+    NumberOfOptimizedInstructions = 0;
+    // We start by adding all loads to the worklist.
+    for (auto LoadDescr : LoadBaseAddresses) {
+      LoadInst *LI = LoadDescr.first;
+      SimplifiedValues[LI] = computeLoadValue(LI, Iteration);
+      if (CountedInsns.insert(LI).second)
+        NumberOfOptimizedInstructions += TTI.getUserCost(LI);
+
+      for (auto U : LI->users()) {
+        Instruction *UI = dyn_cast<Instruction>(U);
+        if (!UI)
+          continue;
+        if (!L->contains(UI))
+          continue;
+        Worklist.push_back(UI);
+      }
+    }
+
+    // And then we try to simplify every user of every instruction from the
+    // worklist. If we do simplify a user, add it to the worklist to process
+    // its users as well.
+    while (!Worklist.empty()) {
+      Instruction *I = Worklist.pop_back_val();
+      if (!visit(I))
+        continue;
+      for (auto U : I->users()) {
+        Instruction *UI = dyn_cast<Instruction>(U);
+        if (!UI)
+          continue;
+        if (!L->contains(UI))
+          continue;
+        Worklist.push_back(UI);
+      }
+    }
+    return NumberOfOptimizedInstructions;
+  }
+
+  // Given a list of potentially simplifed instructions, estimate number of
+  // instructions that would become dead if we do perform the simplification.
+  unsigned EstimateNumberOfDeadInsns() {
+    NumberOfOptimizedInstructions = 0;
+    SmallVector<Instruction *, 8> Worklist;
+    DenseMap<Instruction *, bool> DeadInstructions;
+    // Start by initializing worklist with simplified instructions.
+    for (auto Folded : SimplifiedValues) {
+      if (auto FoldedInsn = dyn_cast<Instruction>(Folded.first)) {
+        Worklist.push_back(FoldedInsn);
+        DeadInstructions[FoldedInsn] = true;
+      }
+    }
+    // If a definition of an insn is only used by simplified or dead
+    // instructions, it's also dead. Check defs of all instructions from the
+    // worklist.
+    while (!Worklist.empty()) {
+      Instruction *FoldedInsn = Worklist.pop_back_val();
+      for (Value *Op : FoldedInsn->operands()) {
+        if (auto I = dyn_cast<Instruction>(Op)) {
+          if (!L->contains(I))
+            continue;
+          if (SimplifiedValues[I])
+            continue; // This insn has been counted already.
+          if (I->getNumUses() == 0)
+            continue;
+          bool AllUsersFolded = true;
+          for (auto U : I->users()) {
+            Instruction *UI = dyn_cast<Instruction>(U);
+            if (!SimplifiedValues[UI] && !DeadInstructions[UI]) {
+              AllUsersFolded = false;
+              break;
+            }
+          }
+          if (AllUsersFolded) {
+            NumberOfOptimizedInstructions += TTI.getUserCost(I);
+            Worklist.push_back(I);
+            DeadInstructions[I] = true;
+          }
+        }
+      }
+    }
+    return NumberOfOptimizedInstructions;
+  }
+};
+
+// Complete loop unrolling can make some loads constant, and we need to know if
+// that would expose any further optimization opportunities.
+// This routine estimates this optimization effect and returns the number of
+// instructions, that potentially might be optimized away.
+static unsigned
+ApproximateNumberOfOptimizedInstructions(const Loop *L, ScalarEvolution &SE,
+                                         unsigned TripCount,
+                                         const TargetTransformInfo &TTI) {
+  if (!TripCount)
+    return 0;
+
+  UnrollAnalyzer UA(L, TripCount, SE, TTI);
+  UA.FindConstFoldableLoads();
+
+  // Estimate number of instructions, that could be simplified if we replace a
+  // load with the corresponding constant. Since the same load will take
+  // different values on different iterations, we have to go through all loop's
+  // iterations here. To limit ourselves here, we check only first N
+  // iterations, and then scale the found number, if necessary.
+  unsigned IterationsNumberForEstimate =
+      std::min<unsigned>(UnrollMaxIterationsCountToAnalyze, TripCount);
+  unsigned NumberOfOptimizedInstructions = 0;
+  for (unsigned i = 0; i < IterationsNumberForEstimate; ++i) {
+    NumberOfOptimizedInstructions += UA.EstimateNumberOfSimplifiedInsns(i);
+    NumberOfOptimizedInstructions += UA.EstimateNumberOfDeadInsns();
+  }
+  NumberOfOptimizedInstructions *= TripCount / IterationsNumberForEstimate;
+
+  return NumberOfOptimizedInstructions;
+}
+
 /// ApproximateLoopSize - Approximate the size of the loop.
 static unsigned ApproximateLoopSize(const Loop *L, unsigned &NumCalls,
                                     bool &NotDuplicatable,
-                                    const TargetTransformInfo &TTI) {
+                                    const TargetTransformInfo &TTI,
+                                    AssumptionCache *AC) {
+  SmallPtrSet<const Value *, 32> EphValues;
+  CodeMetrics::collectEphemeralValues(L, AC, EphValues);
+
   CodeMetrics Metrics;
   for (Loop::block_iterator I = L->block_begin(), E = L->block_end();
        I != E; ++I)
-    Metrics.analyzeBasicBlock(*I, TTI);
+    Metrics.analyzeBasicBlock(*I, TTI, EphValues);
   NumCalls = Metrics.NumInlineCandidates;
   NotDuplicatable = Metrics.notDuplicatable;
 
@@ -212,8 +590,11 @@ static unsigned ApproximateLoopSize(const Loop *L, unsigned &NumCalls,
 
   // Don't allow an estimate of size zero.  This would allows unrolling of loops
   // with huge iteration counts, which is a compile time problem even if it's
-  // not a problem for code quality.
-  if (LoopSize == 0) LoopSize = 1;
+  // not a problem for code quality. Also, the code using this size may assume
+  // that each loop has at least three instructions (likely a conditional
+  // branch, a comparison feeding that branch, and some kind of loop increment
+  // feeding that comparison instruction).
+  LoopSize = std::max(LoopSize, 3u);
 
   return LoopSize;
 }
@@ -221,48 +602,31 @@ static unsigned ApproximateLoopSize(const Loop *L, unsigned &NumCalls,
 // Returns the loop hint metadata node with the given name (for example,
 // "llvm.loop.unroll.count").  If no such metadata node exists, then nullptr is
 // returned.
-const MDNode *GetUnrollMetadata(const Loop *L, StringRef Name) {
-  MDNode *LoopID = L->getLoopID();
-  if (!LoopID)
-    return nullptr;
-
-  // First operand should refer to the loop id itself.
-  assert(LoopID->getNumOperands() > 0 && "requires at least one operand");
-  assert(LoopID->getOperand(0) == LoopID && "invalid loop id");
-
-  for (unsigned i = 1, e = LoopID->getNumOperands(); i < e; ++i) {
-    const MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i));
-    if (!MD)
-      continue;
-
-    const MDString *S = dyn_cast<MDString>(MD->getOperand(0));
-    if (!S)
-      continue;
-
-    if (Name.equals(S->getString()))
-      return MD;
-  }
+static MDNode *GetUnrollMetadataForLoop(const Loop *L, StringRef Name) {
+  if (MDNode *LoopID = L->getLoopID())
+    return GetUnrollMetadata(LoopID, Name);
   return nullptr;
 }
 
 // Returns true if the loop has an unroll(full) pragma.
 static bool HasUnrollFullPragma(const Loop *L) {
-  return GetUnrollMetadata(L, "llvm.loop.unroll.full");
+  return GetUnrollMetadataForLoop(L, "llvm.loop.unroll.full");
 }
 
 // Returns true if the loop has an unroll(disable) pragma.
 static bool HasUnrollDisablePragma(const Loop *L) {
-  return GetUnrollMetadata(L, "llvm.loop.unroll.disable");
+  return GetUnrollMetadataForLoop(L, "llvm.loop.unroll.disable");
 }
 
 // If loop has an unroll_count pragma return the (necessarily
 // positive) value from the pragma.  Otherwise return 0.
 static unsigned UnrollCountPragmaValue(const Loop *L) {
-  const MDNode *MD = GetUnrollMetadata(L, "llvm.loop.unroll.count");
+  MDNode *MD = GetUnrollMetadataForLoop(L, "llvm.loop.unroll.count");
   if (MD) {
     assert(MD->getNumOperands() == 2 &&
            "Unroll count hint metadata should have two operands.");
-    unsigned Count = cast<ConstantInt>(MD->getOperand(1))->getZExtValue();
+    unsigned Count =
+        mdconst::extract<ConstantInt>(MD->getOperand(1))->getZExtValue();
     assert(Count >= 1 && "Unroll count must be positive.");
     return Count;
   }
@@ -278,9 +642,9 @@ static void SetLoopAlreadyUnrolled(Loop *L) {
   if (!LoopID) return;
 
   // First remove any existing loop unrolling metadata.
-  SmallVector<Value *, 4> Vals;
+  SmallVector<Metadata *, 4> MDs;
   // Reserve first location for self reference to the LoopID metadata node.
-  Vals.push_back(nullptr);
+  MDs.push_back(nullptr);
   for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
     bool IsUnrollMetadata = false;
     MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i));
@@ -288,21 +652,21 @@ static void SetLoopAlreadyUnrolled(Loop *L) {
       const MDString *S = dyn_cast<MDString>(MD->getOperand(0));
       IsUnrollMetadata = S && S->getString().startswith("llvm.loop.unroll.");
     }
-    if (!IsUnrollMetadata) Vals.push_back(LoopID->getOperand(i));
+    if (!IsUnrollMetadata)
+      MDs.push_back(LoopID->getOperand(i));
   }
 
   // Add unroll(disable) metadata to disable future unrolling.
   LLVMContext &Context = L->getHeader()->getContext();
-  SmallVector<Value *, 1> DisableOperands;
+  SmallVector<Metadata *, 1> DisableOperands;
   DisableOperands.push_back(MDString::get(Context, "llvm.loop.unroll.disable"));
   MDNode *DisableNode = MDNode::get(Context, DisableOperands);
-  Vals.push_back(DisableNode);
+  MDs.push_back(DisableNode);
 
-  MDNode *NewLoopID = MDNode::get(Context, Vals);
+  MDNode *NewLoopID = MDNode::get(Context, MDs);
   // Set operand 0 to refer to the loop id itself.
   NewLoopID->replaceOperandWith(0, NewLoopID);
   L->setLoopID(NewLoopID);
-  LoopID->replaceAllUsesWith(NewLoopID);
 }
 
 unsigned LoopUnroll::selectUnrollCount(
@@ -349,9 +713,13 @@ bool LoopUnroll::runOnLoop(Loop *L, LPPassManager &LPM) {
   if (skipOptnoneFunction(L))
     return false;
 
-  LoopInfo *LI = &getAnalysis<LoopInfo>();
+  Function &F = *L->getHeader()->getParent();
+
+  LoopInfo *LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
   ScalarEvolution *SE = &getAnalysis<ScalarEvolution>();
-  const TargetTransformInfo &TTI = getAnalysis<TargetTransformInfo>();
+  const TargetTransformInfo &TTI =
+      getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
+  auto &AC = getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
 
   BasicBlock *Header = L->getHeader();
   DEBUG(dbgs() << "Loop Unroll: F[" << Header->getParent()->getName()
@@ -370,13 +738,15 @@ bool LoopUnroll::runOnLoop(Loop *L, LPPassManager &LPM) {
   // Find trip count and trip multiple if count is not available
   unsigned TripCount = 0;
   unsigned TripMultiple = 1;
-  // Find "latch trip count". UnrollLoop assumes that control cannot exit
-  // via the loop latch on any iteration prior to TripCount. The loop may exit
-  // early via an earlier branch.
-  BasicBlock *LatchBlock = L->getLoopLatch();
-  if (LatchBlock) {
-    TripCount = SE->getSmallConstantTripCount(L, LatchBlock);
-    TripMultiple = SE->getSmallConstantTripMultiple(L, LatchBlock);
+  // If there are multiple exiting blocks but one of them is the latch, use the
+  // latch for the trip count estimation. Otherwise insist on a single exiting
+  // block for the trip count estimation.
+  BasicBlock *ExitingBlock = L->getLoopLatch();
+  if (!ExitingBlock || !L->isLoopExiting(ExitingBlock))
+    ExitingBlock = L->getExitingBlock();
+  if (ExitingBlock) {
+    TripCount = SE->getSmallConstantTripCount(L, ExitingBlock);
+    TripMultiple = SE->getSmallConstantTripMultiple(L, ExitingBlock);
   }
 
   // Select an initial unroll count.  This may be reduced later based
@@ -388,9 +758,13 @@ bool LoopUnroll::runOnLoop(Loop *L, LPPassManager &LPM) {
   unsigned NumInlineCandidates;
   bool notDuplicatable;
   unsigned LoopSize =
-      ApproximateLoopSize(L, NumInlineCandidates, notDuplicatable, TTI);
+      ApproximateLoopSize(L, NumInlineCandidates, notDuplicatable, TTI, &AC);
   DEBUG(dbgs() << "  Loop Size = " << LoopSize << "\n");
-  uint64_t UnrolledSize = (uint64_t)LoopSize * Count;
+
+  // When computing the unrolled size, note that the conditional branch on the
+  // backedge and the comparison feeding it are not replicated like the rest of
+  // the loop body (which is why 2 is subtracted).
+  uint64_t UnrolledSize = (uint64_t)(LoopSize-2) * Count + 2;
   if (notDuplicatable) {
     DEBUG(dbgs() << "  Not unrolling loop which contains non-duplicatable"
                  << " instructions.\n");
@@ -401,8 +775,14 @@ bool LoopUnroll::runOnLoop(Loop *L, LPPassManager &LPM) {
     return false;
   }
 
+  unsigned NumberOfOptimizedInstructions =
+      ApproximateNumberOfOptimizedInstructions(L, *SE, TripCount, TTI);
+  DEBUG(dbgs() << "  Complete unrolling could save: "
+               << NumberOfOptimizedInstructions << "\n");
+
   unsigned Threshold, PartialThreshold;
-  selectThresholds(L, HasPragma, UP, Threshold, PartialThreshold);
+  selectThresholds(L, HasPragma, UP, Threshold, PartialThreshold,
+                   NumberOfOptimizedInstructions);
 
   // Given Count, TripCount and thresholds determine the type of
   // unrolling which is to be performed.
@@ -435,7 +815,7 @@ bool LoopUnroll::runOnLoop(Loop *L, LPPassManager &LPM) {
     }
     if (PartialThreshold != NoThreshold && UnrolledSize > PartialThreshold) {
       // Reduce unroll count to be modulo of TripCount for partial unrolling.
-      Count = PartialThreshold / LoopSize;
+      Count = (std::max(PartialThreshold, 3u)-2) / (LoopSize-2);
       while (Count != 0 && TripCount % Count != 0)
         Count--;
     }
@@ -449,7 +829,7 @@ bool LoopUnroll::runOnLoop(Loop *L, LPPassManager &LPM) {
     // the original count which satisfies the threshold limit.
     while (Count != 0 && UnrolledSize > PartialThreshold) {
       Count >>= 1;
-      UnrolledSize = LoopSize * Count;
+      UnrolledSize = (LoopSize-2) * Count + 2;
     }
     if (Count > UP.MaxCount)
       Count = UP.MaxCount;
@@ -494,7 +874,8 @@ bool LoopUnroll::runOnLoop(Loop *L, LPPassManager &LPM) {
   }
 
   // Unroll the loop.
-  if (!UnrollLoop(L, Count, TripCount, AllowRuntime, TripMultiple, LI, this, &LPM))
+  if (!UnrollLoop(L, Count, TripCount, AllowRuntime, TripMultiple, LI, this,
+                  &LPM, &AC))
     return false;
 
   return true;