[PGO] Memory intrinsic calls optimization based on profiled size

[xur]

Summary

This patch optimizes two memory intrinsic operations: memset and memcpy based on the profiled size of the operation. The high level transformation is like:

mem_op(..., size)
==>
switch (size) {
  case s1:
     mem_op(..., s1);
     goto merge_bb;
  case s2:
     mem_op(..., s2);
     goto merge_bb;
  ...
  default:
     mem_op(..., size);
     goto merge_bb;
  }
merge_bb:

There are a few internal options that controls when the optimization takes place.

This patch in an incremental patch based on
https://reviews.llvm.org/D28965

[davidxl]

The optimization pass should be split into two phases as IC promotion. The annotation part should probably be merged with the instrumentation patch. The transformation patch should be done in the same pass as IndirectCallPromotion.

[xur]

In D28966#694902, @davidxl wrote:

The optimization pass should be split into two phases as IC promotion. The annotation part should probably be merged with the instrumentation patch. The transformation patch should be done in the same pass as IndirectCallPromotion.

Is there a good reason for doing the transformation late? Here I don't do the annotation and instead, I do the transformation directly in the same pass. The main reason we have annotations in indirect-call-promotion is we need to call it late (in LTO or ThinLTO). Another reason I'm reluctant to use annotation is that we need to maintain/update it (for inline and unroll, for example).

-Rong

[xur]

Integrated David's suggestion to make the optimization a stand along pass (a function pass).
It currently resides in IndirectCallPromotion.cpp. I will have a follow-up patch to change IndirectCallPromotion.cpp to a more appropriate name.

-Rong

[davidxl]

lib/Transforms/Instrumentation/IndirectCallPromotion.cpp

804

Add a comment on the kinds.

NormalGroup--> NonLargeGroup?

809

Is it correct to guard the parsing here?

Should the parsing be called at the start of the pass with PreciseStart etc as a member variable?

839

DC --> RemainCount?

858

Should it use TotalCount here?

[xur]

lib/Transforms/Instrumentation/IndirectCallPromotion.cpp

809

moved to class MemOPSizeOpt

858

The percent threshold use the ratio of current count and remain count.
I refactor the code a bit (by using a function). The new code also compares the count with the count-threshold.

[xur]

This new patch integrated David's review suggestion to make this optimization a separated pass.
This pass will be called after inlining to avoid the potential interference to in-lining decisions. It will use both the metadata annotation for memop size profiling and the enclosing BB's edge count to find the actual count.

[davidxl]

lib/Transforms/Instrumentation/IndirectCallPromotion.cpp

835

Use APInt or SaturatingMultipleAdd to deal with possible overflow.

850

Should check the actual count.

884

can shrink the MemOpScaleCount check into getScaledCount

[xur]

integrated David's review comments

[davidxl]

lgtm

include/llvm/InitializePasses.h

 void initializeObjCARCOptPass(PassRegistry&);
 void initializeOptimizationRemarkEmitterWrapperPassPass(PassRegistry&);
 void initializeOptimizePHIsPass(PassRegistry&);
 void initializePAEvalPass(PassRegistry&);
 void initializePEIPass(PassRegistry&);
 void initializePGOIndirectCallPromotionLegacyPassPass(PassRegistry&);
 void initializePGOInstrumentationGenLegacyPassPass(PassRegistry&);
 void initializePGOInstrumentationUseLegacyPassPass(PassRegistry&);
+void initializePGOMemOPSizeOptLegacyPassPass(PassRegistry&);
 void initializePHIEliminationPass(PassRegistry&);
 void initializePartialInlinerLegacyPassPass(PassRegistry&);
 void initializePartiallyInlineLibCallsLegacyPassPass(PassRegistry&);
 void initializePatchableFunctionPass(PassRegistry&);
 void initializePeepholeOptimizerPass(PassRegistry&);
 void initializePhysicalRegisterUsageInfoPass(PassRegistry&);
 void initializePlaceBackedgeSafepointsImplPass(PassRegistry&);
 void initializePlaceSafepointsPass(PassRegistry&);

include/llvm/LinkAllPasses.h

ForcePassLinking() {
       (void) llvm::createDomOnlyPrinterPass();
       (void) llvm::createDomPrinterPass();
       (void) llvm::createDomOnlyViewerPass();
       (void) llvm::createDomViewerPass();
       (void) llvm::createGCOVProfilerPass();
       (void) llvm::createPGOInstrumentationGenLegacyPass();
       (void) llvm::createPGOInstrumentationUseLegacyPass();
       (void) llvm::createPGOIndirectCallPromotionLegacyPass();
+      (void) llvm::createPGOMemOPSizeOptLegacyPass();
       (void) llvm::createInstrProfilingLegacyPass();
       (void) llvm::createFunctionImportPass();
       (void) llvm::createFunctionInliningPass();
       (void) llvm::createAlwaysInlinerLegacyPass();
       (void) llvm::createGlobalDCEPass();
       (void) llvm::createGlobalOptimizerPass();
       (void) llvm::createGlobalsAAWrapperPass();
       (void) llvm::createGuardWideningPass();

include/llvm/ProfileData/InstrProf.h

 // InstrProfilingBuffer.c.
 struct Header {
 #define INSTR_PROF_RAW_HEADER(Type, Name, Init) const Type Name;
 #include "llvm/ProfileData/InstrProfData.inc"
 };
 
 } // end namespace RawInstrProf
 
+// Parse MemOP Size range option.
+void getMemOPSizeRangeFromOption(std::string Str, int64_t &RangeStart,
+                                 int64_t &RangeLast);
+
 } // end namespace llvm
 
 #endif // LLVM_PROFILEDATA_INSTRPROF_H

include/llvm/Transforms/InstrProfiling.h

private:
   };
   DenseMap<GlobalVariable *, PerFunctionProfileData> ProfileDataMap;
   std::vector<GlobalValue *> UsedVars;
   std::vector<GlobalVariable *> ReferencedNames;
   GlobalVariable *NamesVar;
   size_t NamesSize;
 
   // The start value of precise value profile range for memory intrinsic sizes.
-  const int64_t DefaultMemOPSizeRangeStart = 0;
   int64_t MemOPSizeRangeStart;
   // The end value of precise value profile range for memory intrinsic sizes.
-  const int64_t DefaultMemOPSizeRangeLast = 8;
   int64_t MemOPSizeRangeLast;
-  int64_t MemOPSizeLargeVal;
 
   bool isMachO() const;
 
   /// Get the section name for the counter variables.
   StringRef getCountersSection() const;
 
   /// Get the section name for the name variables.
   StringRef getNameSection() const;
private:
   void emitRuntimeHook();
 
   /// Add uses of our data variables and runtime hook.
   void emitUses();
 
   /// Create a static initializer for our data, on platforms that need it,
   /// and for any profile output file that was specified.
   void emitInitialization();
-
-  /// Helper funtion that parsing the MemOPSize value profile options
-  void getMemOPSizeOptions();
 };
 
 } // end namespace llvm
 
 #endif // LLVM_TRANSFORMS_INSTRPROFILING_H

include/llvm/Transforms/Instrumentation.h

ModulePass *createGCOVProfilerPass(const GCOVOptions &Options =
                                    GCOVOptions::getDefault());
 
 // PGO Instrumention
 ModulePass *createPGOInstrumentationGenLegacyPass();
 ModulePass *
 createPGOInstrumentationUseLegacyPass(StringRef Filename = StringRef(""));
 ModulePass *createPGOIndirectCallPromotionLegacyPass(bool InLTO = false,
                                                      bool SamplePGO = false);
+FunctionPass *createPGOMemOPSizeOptLegacyPass();
 
 // Helper function to check if it is legal to promote indirect call \p Inst
 // to a direct call of function \p F. Stores the reason in \p Reason.
 bool isLegalToPromote(Instruction *Inst, Function *F, const char **Reason);
 
 // Helper function that transforms Inst (either an indirect-call instruction, or
 // an invoke instruction , to a conditional call to F. This is like:
 //     if (Inst.CalledValue == F)

include/llvm/Transforms/PGOInstrumentation.h

PGOIndirectCallPromotion(bool IsInLTO = false, bool SamplePGO = false)
       : InLTO(IsInLTO), SamplePGO(SamplePGO) {}
   PreservedAnalyses run(Module &M, ModuleAnalysisManager &AM);
 
 private:
   bool InLTO;
   bool SamplePGO;
 };
 
+/// The profile size based optimization pass for memory intrinsics.
+class PGOMemOPSizeOpt : public PassInfoMixin<PGOMemOPSizeOpt> {
+public:
+  PGOMemOPSizeOpt() {}
+  PreservedAnalyses run(Function &F, FunctionAnalysisManager &AM);
+};
+
+void setProfMetadata(Module *M, Instruction *TI, ArrayRef<uint64_t> EdgeCounts,
+                     uint64_t MaxCount);
+
 } // End llvm namespace
 #endif

lib/Passes/PassRegistry.def

 FUNCTION_PASS("newgvn", NewGVNPass())
 FUNCTION_PASS("jump-threading", JumpThreadingPass())
 FUNCTION_PASS("partially-inline-libcalls", PartiallyInlineLibCallsPass())
 FUNCTION_PASS("lcssa", LCSSAPass())
 FUNCTION_PASS("loop-data-prefetch", LoopDataPrefetchPass())
 FUNCTION_PASS("loop-load-elim", LoopLoadEliminationPass())
 FUNCTION_PASS("loop-distribute", LoopDistributePass())
 FUNCTION_PASS("loop-vectorize", LoopVectorizePass())
+FUNCTION_PASS("pgo-memop-opt", PGOMemOPSizeOpt())
 FUNCTION_PASS("print", PrintFunctionPass(dbgs()))
 FUNCTION_PASS("print<assumptions>", AssumptionPrinterPass(dbgs()))
 FUNCTION_PASS("print<block-freq>", BlockFrequencyPrinterPass(dbgs()))
 FUNCTION_PASS("print<branch-prob>", BranchProbabilityPrinterPass(dbgs()))
 FUNCTION_PASS("print<domtree>", DominatorTreePrinterPass(dbgs()))
 FUNCTION_PASS("print<postdomtree>", PostDominatorTreePrinterPass(dbgs()))
 FUNCTION_PASS("print<demanded-bits>", DemandedBitsPrinterPass(dbgs()))
 FUNCTION_PASS("print<domfrontier>", DominanceFrontierPrinterPass(dbgs()))

lib/ProfileData/InstrProf.cpp

bool canRenameComdatFunc(const Function &F, bool CheckAddressTaken) {
   // For AvailableExternallyLinkage functions.
   if (!F.hasComdat()) {
     assert(F.getLinkage() == GlobalValue::AvailableExternallyLinkage);
     return true;
   }
   return true;
 }
 
+// Parse the value profile options.
+void getMemOPSizeRangeFromOption(std::string MemOPSizeRange,
+                                 int64_t &RangeStart, int64_t &RangeLast) {
+  static const int64_t DefaultMemOPSizeRangeStart = 0;
+  static const int64_t DefaultMemOPSizeRangeLast = 8;
+  RangeStart = DefaultMemOPSizeRangeStart;
+  RangeLast = DefaultMemOPSizeRangeLast;
+
+  if (!MemOPSizeRange.empty()) {
+    auto Pos = MemOPSizeRange.find(":");
+    if (Pos != std::string::npos) {
+      if (Pos > 0)
+        RangeStart = atoi(MemOPSizeRange.substr(0, Pos).c_str());
+      if (Pos < MemOPSizeRange.size() - 1)
+        RangeLast = atoi(MemOPSizeRange.substr(Pos + 1).c_str());
+    } else
+      RangeLast = atoi(MemOPSizeRange.c_str());
+  }
+  assert(RangeLast >= RangeStart);
+}
+
 } // end namespace llvm

lib/Transforms/IPO/PassManagerBuilder.cpp

void PassManagerBuilder::addFunctionSimplificationPasses(
   MPM.add(createCorrelatedValuePropagationPass()); // Propagate conditionals
   MPM.add(createCFGSimplificationPass());     // Merge & remove BBs
   // Combine silly seq's
   addInstructionCombiningPass(MPM);
   if (SizeLevel == 0 && !DisableLibCallsShrinkWrap)
     MPM.add(createLibCallsShrinkWrapPass());
   addExtensionsToPM(EP_Peephole, MPM);
 
+  // Optimize memory intrinsic calls based on the profiled size information.
+  if (SizeLevel == 0)
+    MPM.add(createPGOMemOPSizeOptLegacyPass());
+
   MPM.add(createTailCallEliminationPass()); // Eliminate tail calls
   MPM.add(createCFGSimplificationPass());     // Merge & remove BBs
   MPM.add(createReassociatePass());           // Reassociate expressions
   // Rotate Loop - disable header duplication at -Oz
   MPM.add(createLoopRotatePass(SizeLevel == 2 ? 0 : -1));
   MPM.add(createLICMPass());                  // Hoist loop invariants
   MPM.add(createLoopUnswitchPass(SizeLevel || OptLevel < 3, DivergentTarget));
   MPM.add(createCFGSimplificationPass());

lib/Transforms/Instrumentation/IndirectCallPromotion.cpp

-//===-- IndirectCallPromotion.cpp - Promote indirect calls to direct calls ===//
+//===-- IndirectCallPromotion.cpp - Optimizations based on value profiling ===//
 //
 //                      The LLVM Compiler Infrastructure
 //
 // This file is distributed under the University of Illinois Open Source
 // License. See LICENSE.TXT for details.
 //
 //===----------------------------------------------------------------------===//
 //
 // This file implements the transformation that promotes indirect calls to
 // conditional direct calls when the indirect-call value profile metadata is
 // available.
 //
 //===----------------------------------------------------------------------===//
 
 #include "llvm/ADT/ArrayRef.h"
 #include "llvm/ADT/Statistic.h"
 #include "llvm/ADT/StringRef.h"
 #include "llvm/ADT/Twine.h"
+#include "llvm/Analysis/BlockFrequencyInfo.h"
 #include "llvm/Analysis/IndirectCallPromotionAnalysis.h"
 #include "llvm/Analysis/IndirectCallSiteVisitor.h"
 #include "llvm/IR/BasicBlock.h"
 #include "llvm/IR/CallSite.h"
 #include "llvm/IR/DerivedTypes.h"
 #include "llvm/IR/DiagnosticInfo.h"
 #include "llvm/IR/Function.h"
 #include "llvm/IR/IRBuilder.h"
 #include "llvm/IR/InstrTypes.h"
 #include "llvm/IR/Instruction.h"
 #include "llvm/IR/Instructions.h"
 #include "llvm/IR/LLVMContext.h"
 #include "llvm/IR/MDBuilder.h"
 #include "llvm/IR/PassManager.h"
 #include "llvm/IR/Type.h"
 #include "llvm/Pass.h"
 #include "llvm/PassRegistry.h"
 #include "llvm/PassSupport.h"
 #include "llvm/ProfileData/InstrProf.h"
 #include "llvm/Support/Casting.h"
 #include "llvm/Support/CommandLine.h"
 #include "llvm/Support/Debug.h"
 #include "llvm/Support/ErrorHandling.h"
+#include "llvm/Support/MathExtras.h"
 #include "llvm/Transforms/Instrumentation.h"
 #include "llvm/Transforms/PGOInstrumentation.h"
 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
 #include <cassert>
 #include <cstdint>
 #include <vector>
 
 using namespace llvm;
 
 #define DEBUG_TYPE "pgo-icall-prom"
 
 STATISTIC(NumOfPGOICallPromotion, "Number of indirect call promotions.");
 STATISTIC(NumOfPGOICallsites, "Number of indirect call candidate sites.");
+STATISTIC(NumOfPGOMemOPOpt, "Number of memop intrinsics optimized.");
+STATISTIC(NumOfPGOMemOPAnnotate, "Number of memop intrinsics annotated.");
 
 // Command line option to disable indirect-call promotion with the default as
 // false. This is for debug purpose.
 static cl::opt<bool> DisableICP("disable-icp", cl::init(false), cl::Hidden,
                                 cl::desc("Disable indirect call promotion"));
 
 // Set the cutoff value for the promotion. If the value is other than 0, we
 // stop the transformation once the total number of promotions equals the cutoff
static cl::opt<bool> ICPInvokeOnly("icp-invoke-only", cl::init(false),
                                             "invoke instruction only"));
 
 // Dump the function level IR if the transformation happened in this
 // function. For debug use only.
 static cl::opt<bool>
     ICPDUMPAFTER("icp-dumpafter", cl::init(false), cl::Hidden,
                  cl::desc("Dump IR after transformation happens"));
 
+// The minimum call count to optimize memory intrinsic calls.
+static cl::opt<unsigned>
+    MemOPCountThreshold("pgo-memop-count-threshold", cl::Hidden, cl::ZeroOrMore,
+                        cl::init(1000),
+                        cl::desc("The minimum count to optimize memory "
+                                 "intrinsic calls"));
+
+// Command line option to disable memory intrinsic optimization. The default is
+// false. This is for debug purpose.
+static cl::opt<bool> DisableMemOPOPT("disable-memop-opt", cl::init(false),
+                                     cl::Hidden, cl::desc("Disable optimize"));
+
+// The percent threshold to optimize memory intrinsic calls.
+static cl::opt<unsigned>
+    MemOPPercentThreshold("pgo-memop-percent-threshold", cl::init(40),
+                          cl::Hidden, cl::ZeroOrMore,
+                          cl::desc("The percentage threshold for the "
+                                   "memory intrinsic calls optimization"));
+
+// Maximum number of versions for optimizing memory intrinsic call.
+static cl::opt<unsigned>
+    MemOPMaxVersion("pgo-memop-max-version", cl::init(3), cl::Hidden,
+                    cl::ZeroOrMore,
+                    cl::desc("The max version for the optimized memory "
+                             " intrinsic calls"));
+
+// Scale the counts from the annotation using the BB count value.
+static cl::opt<bool>
+    MemOPScaleCount("pgo-memop-scale-count", cl::init(true), cl::Hidden,
+                    cl::desc("Scale the memop size counts using the basic "
+                             " block count value"));
+
+// This option set the rangge of precise profile memop sizes.
+extern cl::opt<std::string> MemOPSizeRange;
+
+// This option set the value that groups large memop sizes
+extern cl::opt<unsigned> MemOPSizeLarge;
+
 namespace {
 class PGOIndirectCallPromotionLegacyPass : public ModulePass {
 public:
   static char ID;
 
   PGOIndirectCallPromotionLegacyPass(bool InLTO = false, bool SamplePGO = false)
       : ModulePass(ID), InLTO(InLTO), SamplePGO(SamplePGO) {
     initializePGOIndirectCallPromotionLegacyPassPass(
         *PassRegistry::getPassRegistry());
   }
 
   StringRef getPassName() const override { return "PGOIndirectCallPromotion"; }
 
 private:
   bool runOnModule(Module &M) override;
 
   // If this pass is called in LTO. We need to special handling the PGOFuncName
   // for the static variables due to LTO's internalization.
   bool InLTO;
 
   // If this pass is called in SamplePGO. We need to add the prof metadata to
   // the promoted direct call.
   bool SamplePGO;
 };
+
+class PGOMemOPSizeOptLegacyPass : public FunctionPass {
+public:
+  static char ID;
+
+  PGOMemOPSizeOptLegacyPass() : FunctionPass(ID) {
+    initializePGOMemOPSizeOptLegacyPassPass(*PassRegistry::getPassRegistry());
+  }
+
+  StringRef getPassName() const override { return "PGOMemOPSize"; }
+
+private:
+  bool runOnFunction(Function &F) override;
+  void getAnalysisUsage(AnalysisUsage &AU) const override {
+    AU.addRequired<BlockFrequencyInfoWrapperPass>();
+  }
+};
 } // end anonymous namespace
 
 char PGOIndirectCallPromotionLegacyPass::ID = 0;
 INITIALIZE_PASS(PGOIndirectCallPromotionLegacyPass, "pgo-icall-prom",
                 "Use PGO instrumentation profile to promote indirect calls to "
                 "direct calls.",
                 false, false)
 
 ModulePass *llvm::createPGOIndirectCallPromotionLegacyPass(bool InLTO,
                                                            bool SamplePGO) {
   return new PGOIndirectCallPromotionLegacyPass(InLTO, SamplePGO);
 }
 
+char PGOMemOPSizeOptLegacyPass::ID = 0;
+INITIALIZE_PASS_BEGIN(PGOMemOPSizeOptLegacyPass, "pgo-memop-opt",
+                      "Optimize memory intrinsic using its size value profile",
+                      false, false)
+INITIALIZE_PASS_DEPENDENCY(BlockFrequencyInfoWrapperPass)
+INITIALIZE_PASS_END(PGOMemOPSizeOptLegacyPass, "pgo-memop-opt",
+                    "Optimize memory intrinsic using its size value profile",
+                    false, false)
+
+FunctionPass *llvm::createPGOMemOPSizeOptLegacyPass() {
+  return new PGOMemOPSizeOptLegacyPass();
+}
+
 namespace {
 // The class for main data structure to promote indirect calls to conditional
 // direct calls.
 class ICallPromotionFunc {
 private:
   Function &F;
   Module *M;
 

 PreservedAnalyses PGOIndirectCallPromotion::run(Module &M,
                                                 ModuleAnalysisManager &AM) {
   if (!promoteIndirectCalls(M, InLTO | ICPLTOMode,
                             SamplePGO | ICPSamplePGOMode))
     return PreservedAnalyses::all();
 
   return PreservedAnalyses::none();
 }
+
+namespace {
+class MemOPSizeOpt : public InstVisitor<MemOPSizeOpt> {
+public:
+  MemOPSizeOpt(Function &Func, BlockFrequencyInfo &BFI)
+      : Func(Func), BFI(BFI), Changed(false) {
+    ValueDataArray =
+        llvm::make_unique<InstrProfValueData[]>(MemOPMaxVersion + 2);
+    // Get the MemOPSize range information from option MemOPSizeRange,
+    getMemOPSizeRangeFromOption(MemOPSizeRange, PreciseRangeStart,
+                                PreciseRangeLast);
+  }
+  bool isChanged() const { return Changed; }
+  void perform() {
+    WorkList.clear();
+    visit(Func);
+
+    for (auto &MI : WorkList) {
+      ++NumOfPGOMemOPAnnotate;
+      if (perform(MI)) {
+        Changed = true;
+        ++NumOfPGOMemOPOpt;
+        DEBUG(dbgs() << "MemOP calls: " << MI->getCalledFunction()->getName()
+                     << "is Transformed.\n");
+      }
+    }
+  }
+
+  void visitMemIntrinsic(MemIntrinsic &MI) {
+    Value *Length = MI.getLength();
+    // Not perform on constant length calls.
+    if (dyn_cast<ConstantInt>(Length))
+      return;
+    WorkList.push_back(&MI);
+  }
+
+private:
+  Function &Func;
+  BlockFrequencyInfo &BFI;
+  bool Changed;
+  std::vector<MemIntrinsic *> WorkList;
+  // Start of the previse range.
+  int64_t PreciseRangeStart;
+  // Last value of the previse range.
+  int64_t PreciseRangeLast;
+  // The space to read the profile annotation.
+  std::unique_ptr<InstrProfValueData[]> ValueDataArray;
+  bool perform(MemIntrinsic *MI);
+
+  // This kind shows which group the value falls in. For PreciseValue, we have
+  // the profile count for that value. LargeGroup groups the values that are in
+  // range [LargeValue, +inf). NonLargeGroup groups the rest of values.
+  enum MemOPSizeKind { PreciseValue, NonLargeGroup, LargeGroup };
+
+  MemOPSizeKind getMemOPSizeKind(int64_t Value) const {
+    if (Value == MemOPSizeLarge && MemOPSizeLarge != 0)
+      return LargeGroup;
+    if (Value == PreciseRangeLast + 1)
+      return NonLargeGroup;
+    return PreciseValue;
+  }
+};
+
+static const char *getMIName(const MemIntrinsic *MI) {
+  switch (MI->getIntrinsicID()) {
+  case Intrinsic::memcpy:
+    return "memcpy";
+  case Intrinsic::memmove:
+    return "memmove";
+  case Intrinsic::memset:
+    return "memset";
+  default:
+    return "unknown";
+  }
+}
+
+static bool isProfitable(uint64_t Count, uint64_t TotalCount) {
+  assert(Count <= TotalCount);
+  if (Count < MemOPCountThreshold)
+    return false;
+  if (Count < TotalCount * MemOPPercentThreshold / 100)
+    return false;
+  return true;
+}
+
+static inline uint64_t getScaledCount(uint64_t Count, uint64_t Num,
+                                      uint64_t Denom) {
+  if (!MemOPScaleCount)
+    return Count;
+  bool Overflowed;
+  uint64_t ScaleCount = SaturatingMultiply(Count, Num, &Overflowed);
+  return ScaleCount / Denom;
+}
+
+bool MemOPSizeOpt::perform(MemIntrinsic *MI) {
+  assert(MI);
+  if (MI->getIntrinsicID() == Intrinsic::memmove)
+    return false;
+
+  uint32_t NumVals, MaxNumPromotions = MemOPMaxVersion + 2;
+  uint64_t TotalCount;
+  if (!getValueProfDataFromInst(*MI, IPVK_MemOPSize, MaxNumPromotions,
+                                ValueDataArray.get(), NumVals, TotalCount))
+    return false;
+
+  uint64_t ActualCount = TotalCount;
+  uint64_t SavedTotalCount = TotalCount;
+  if (MemOPScaleCount) {
+    auto BBEdgeCount = BFI.getBlockProfileCount(MI->getParent());
+    if (!BBEdgeCount)
+      return false;
+    ActualCount = *BBEdgeCount;
+  }
+
+  if (ActualCount < MemOPCountThreshold)
+    return false;
+
+  ArrayRef<InstrProfValueData> VDs(ValueDataArray.get(), NumVals);
+  TotalCount = ActualCount;
+  if (MemOPScaleCount)
+    DEBUG(dbgs() << "Scale counts: numberator = " << ActualCount
+                 << " denominator = " << SavedTotalCount << "\n");
+
+  // Keeping track of the count of the default case:
+  uint64_t RemainCount = TotalCount;
+  SmallVector<uint64_t, 16> SizeIds;
+  SmallVector<uint64_t, 16> CaseCounts;
+  uint64_t MaxCount = 0;
+  unsigned Version = 0;
+  // Default case is in the front -- save the slot here.
+  CaseCounts.push_back(0);
+  for (auto &VD : VDs) {
+    int64_t V = VD.Value;
+    uint64_t C = VD.Count;
+    if (MemOPScaleCount)
+      C = getScaledCount(C, ActualCount, SavedTotalCount);
+
+    // Only care precise value here.
+    if (getMemOPSizeKind(V) != PreciseValue)
+      continue;
+
+    // ValueCount are sorted on the count. Break at the first un-profitable
+    // value.
+    if (!isProfitable(C, RemainCount))
+      break;
+
+    SizeIds.push_back(V);
+    CaseCounts.push_back(C);
+    if (C > MaxCount)
+      MaxCount = C;
+
+    assert(RemainCount >= C);
+    RemainCount -= C;
+
+    if (++Version > MemOPMaxVersion && MemOPMaxVersion != 0)
+      break;
+  }
+
+  if (Version == 0)
+    return false;
+
+  CaseCounts[0] = RemainCount;
+  if (RemainCount > MaxCount)
+    MaxCount = RemainCount;
+
+  uint64_t SumForOpt = TotalCount - RemainCount;
+  DEBUG(dbgs() << "Read one memory intrinsic profile: " << SumForOpt << " vs "
+               << TotalCount << "\n");
+  DEBUG(
+      for (auto &VD
+           : VDs) { dbgs() << "  (" << VD.Value << "," << VD.Count << ")\n"; });
+
+  DEBUG(dbgs() << "Optimize one memory intrinsic call to " << Version
+               << " Versions\n");
+
+  // mem_op(..., size)
+  // ==>
+  // switch (size) {
+  //   case s1:
+  //      mem_op(..., s1);
+  //      goto merge_bb;
+  //   case s2:
+  //      mem_op(..., s2);
+  //      goto merge_bb;
+  //   ...
+  //   default:
+  //      mem_op(..., size);
+  //      goto merge_bb;
+  // }
+  // merge_bb:
+
+  BasicBlock *BB = MI->getParent();
+  DEBUG(dbgs() << "\n\n== Basic Block Before ==\n");
+  DEBUG(dbgs() << *BB << "\n");
+
+  BasicBlock *DefaultBB = SplitBlock(BB, MI);
+  BasicBlock::iterator It(*MI);
+  ++It;
+  assert(It != DefaultBB->end());
+  BasicBlock *MergeBB = SplitBlock(DefaultBB, &(*It));
+  DefaultBB->setName("MemOP.Default");
+  MergeBB->setName("MemOP.Merge");
+
+  auto &Ctx = Func.getContext();
+  IRBuilder<> IRB(BB);
+  BB->getTerminator()->eraseFromParent();
+  Value *SizeVar = MI->getLength();
+  SwitchInst *SI = IRB.CreateSwitch(SizeVar, DefaultBB, SizeIds.size());
+
+  // Clear the value profile data.
+  MI->setMetadata(LLVMContext::MD_prof, nullptr);
+
+  DEBUG(dbgs() << "\n\n== Basic Block After==\n");
+
+  for (uint64_t SizeId : SizeIds) {
+    ConstantInt *CaseSizeId = ConstantInt::get(Type::getInt64Ty(Ctx), SizeId);
+    BasicBlock *CaseBB = BasicBlock::Create(
+        Ctx, Twine("MemOP.Case.") + Twine(SizeId), &Func, DefaultBB);
+    Instruction *NewInst = MI->clone();
+    // Fix the argument.
+    dyn_cast<MemIntrinsic>(NewInst)->setLength(CaseSizeId);
+    CaseBB->getInstList().push_back(NewInst);
+    IRBuilder<> IRBCase(CaseBB);
+    IRBCase.CreateBr(MergeBB);
+    SI->addCase(CaseSizeId, CaseBB);
+    DEBUG(dbgs() << *CaseBB << "\n");
+  }
+  setProfMetadata(Func.getParent(), SI, CaseCounts, MaxCount);
+
+  DEBUG(dbgs() << *BB << "\n");
+  DEBUG(dbgs() << *DefaultBB << "\n");
+  DEBUG(dbgs() << *MergeBB << "\n");
+
+  emitOptimizationRemark(Func.getContext(), "memop-opt", Func,
+                         MI->getDebugLoc(),
+                         Twine("optimize ") + getMIName(MI) + " with count " +
+                             Twine(SumForOpt) + " out of " + Twine(TotalCount) +
+                             " for " + Twine(Version) + " versions");
+
+  return true;
+}
+} // namespace
+
+static bool PGOMemOPSizeOptImpl(Function &F, BlockFrequencyInfo &BFI) {
+  if (DisableMemOPOPT)
+    return false;
+
+  if (F.hasFnAttribute(Attribute::OptimizeForSize))
+    return false;
+  MemOPSizeOpt MemOPSizeOpt(F, BFI);
+  MemOPSizeOpt.perform();
+  return MemOPSizeOpt.isChanged();
+}
+
+bool PGOMemOPSizeOptLegacyPass::runOnFunction(Function &F) {
+  BlockFrequencyInfo &BFI =
+      getAnalysis<BlockFrequencyInfoWrapperPass>().getBFI();
+  return PGOMemOPSizeOptImpl(F, BFI);
+}
+
+namespace llvm {
+char &PGOMemOPSizeOptID = PGOMemOPSizeOptLegacyPass::ID;
+
+PreservedAnalyses PGOMemOPSizeOpt::run(Function &F,
+                                       FunctionAnalysisManager &FAM) {
+  auto &BFI = FAM.getResult<BlockFrequencyAnalysis>(F);
+  bool Changed = PGOMemOPSizeOptImpl(F, BFI);
+  if (!Changed)
+    return PreservedAnalyses::all();
+  return PreservedAnalyses::none();
+}
+} // namespace llvm

lib/Transforms/Instrumentation/InstrProfiling.cpp

 #include <cstddef>
 #include <cstdint>
 #include <string>
 
 using namespace llvm;
 
 #define DEBUG_TYPE "instrprof"
 
+// The start and end values of precise value profile range for memory
+// intrinsic sizes
+cl::opt<std::string> MemOPSizeRange(
+    "memop-size-range",
+    cl::desc("Set the range of size in memory intrinsic calls to be profiled "
+             "precisely, in a format of <start_val>:<end_val>"),
+    cl::init(""));
+
+// The value that considered to be large value in  memory intrinsic.
+cl::opt<unsigned> MemOPSizeLarge(
+    "memop-size-large",
+    cl::desc("Set large value thresthold in memory intrinsic size profiling. "
+             "Value of 0 disables the large value profiling."),
+    cl::init(8192));
+
 namespace {
 
 cl::opt<bool> DoNameCompression("enable-name-compression",
                                 cl::desc("Enable name string compression"),
                                 cl::init(true));
 
 cl::opt<bool> DoHashBasedCounterSplit(
     "hash-based-counter-split",
cl::opt<double> NumCountersPerValueSite(
     cl::desc("The average number of profile counters allocated "
              "per value profiling site."),
     // This is set to a very small value because in real programs, only
     // a very small percentage of value sites have non-zero targets, e.g, 1/30.
     // For those sites with non-zero profile, the average number of targets
     // is usually smaller than 2.
     cl::init(1.0));
 
-cl::opt<std::string> MemOPSizeRange(
-    "memop-size-range",
-    cl::desc("Set the range of size in memory intrinsic calls to be profiled "
-             "precisely, in a format of <start_val>:<end_val>"),
-    cl::init(""));
-cl::opt<unsigned> MemOPSizeLarge(
-    "memop-size-large",
-    cl::desc("Set large value thresthold in memory intrinsic size profiling. "
-             "Value of 0 disables the large value profiling."),
-    cl::init(8192));
-
 class InstrProfilingLegacyPass : public ModulePass {
   InstrProfiling InstrProf;
 
 public:
   static char ID;
 
   InstrProfilingLegacyPass() : ModulePass(ID) {}
   InstrProfilingLegacyPass(const InstrProfOptions &Options)
bool InstrProfiling::run(Module &M, const TargetLibraryInfo &TLI) {
   bool MadeChange = false;
 
   this->M = &M;
   this->TLI = &TLI;
   NamesVar = nullptr;
   NamesSize = 0;
   ProfileDataMap.clear();
   UsedVars.clear();
-  getMemOPSizeOptions();
+  getMemOPSizeRangeFromOption(MemOPSizeRange, MemOPSizeRangeStart,
+                              MemOPSizeRangeLast);
 
   // We did not know how many value sites there would be inside
   // the instrumented function. This is counting the number of instrumented
   // target value sites to enter it as field in the profile data variable.
   for (Function &F : M) {
     InstrProfIncrementInst *FirstProfIncInst = nullptr;
     for (BasicBlock &BB : F)
       for (auto I = BB.begin(), E = BB.end(); I != E; I++)
bool IsRange = (Ind->getValueKind()->getZExtValue() ==
                   llvm::InstrProfValueKind::IPVK_MemOPSize);
   CallInst *Call = nullptr;
   if (!IsRange) {
     Value *Args[3] = {Ind->getTargetValue(),
                       Builder.CreateBitCast(DataVar, Builder.getInt8PtrTy()),
                       Builder.getInt32(Index)};
     Call = Builder.CreateCall(getOrInsertValueProfilingCall(*M, *TLI), Args);
   } else {
-    Value *Args[6] = {Ind->getTargetValue(),
-                      Builder.CreateBitCast(DataVar, Builder.getInt8PtrTy()),
-                      Builder.getInt32(Index),
-                      Builder.getInt64(MemOPSizeRangeStart),
-                      Builder.getInt64(MemOPSizeRangeLast),
-                      Builder.getInt64(MemOPSizeLargeVal)};
+    Value *Args[6] = {
+        Ind->getTargetValue(),
+        Builder.getInt64(MemOPSizeLarge == 0 ? INT64_MIN : MemOPSizeLarge)};
     Call =
         Builder.CreateCall(getOrInsertValueProfilingCall(*M, *TLI, true), Args);
   }
   if (auto AK = TLI->getExtAttrForI32Param(false))
     Call->addAttribute(3, AK);
   Ind->replaceAllUsesWith(Call);
   Ind->eraseFromParent();
 }
void InstrProfiling::emitInitialization() {
   // Add the basic block and the necessary calls.
   IRBuilder<> IRB(BasicBlock::Create(M->getContext(), "", F));
   if (RegisterF)
     IRB.CreateCall(RegisterF, {});
   IRB.CreateRetVoid();
 
   appendToGlobalCtors(*M, F, 0);
 }
-
-void InstrProfiling::getMemOPSizeOptions() {
-  // Parse the value profile options.
-  MemOPSizeRangeStart = DefaultMemOPSizeRangeStart;
-  MemOPSizeRangeLast = DefaultMemOPSizeRangeLast;
-  if (!MemOPSizeRange.empty()) {
-    auto Pos = MemOPSizeRange.find(":");
-    if (Pos != std::string::npos) {
-      if (Pos > 0)
-        MemOPSizeRangeStart = atoi(MemOPSizeRange.substr(0, Pos).c_str());
-      if (Pos < MemOPSizeRange.size() - 1)
-        MemOPSizeRangeLast = atoi(MemOPSizeRange.substr(Pos + 1).c_str());
-    } else
-      MemOPSizeRangeLast = atoi(MemOPSizeRange.c_str());
-  }
-  assert(MemOPSizeRangeLast >= MemOPSizeRangeStart);
-
-  MemOPSizeLargeVal = MemOPSizeLarge;
-  if (MemOPSizeLargeVal == 0)
-    MemOPSizeLargeVal = INT64_MIN;
-}

lib/Transforms/Instrumentation/Instrumentation.cpp

 void llvm::initializeInstrumentation(PassRegistry &Registry) {
   initializeAddressSanitizerPass(Registry);
   initializeAddressSanitizerModulePass(Registry);
   initializeBoundsCheckingPass(Registry);
   initializeGCOVProfilerLegacyPassPass(Registry);
   initializePGOInstrumentationGenLegacyPassPass(Registry);
   initializePGOInstrumentationUseLegacyPassPass(Registry);
   initializePGOIndirectCallPromotionLegacyPassPass(Registry);
+  initializePGOMemOPSizeOptLegacyPassPass(Registry);
   initializeInstrProfilingLegacyPassPass(Registry);
   initializeMemorySanitizerPass(Registry);
   initializeThreadSanitizerPass(Registry);
   initializeSanitizerCoverageModulePass(Registry);
   initializeDataFlowSanitizerPass(Registry);
   initializeEfficiencySanitizerPass(Registry);
 }
 
 /// LLVMInitializeInstrumentation - C binding for
 /// initializeInstrumentation.
 void LLVMInitializeInstrumentation(LLVMPassRegistryRef R) {
   initializeInstrumentation(*unwrap(R));
 }

lib/Transforms/Instrumentation/PGOInstrumentation.cpp

#endif
 
   // Now annotate select instructions
   FuncInfo.SIVisitor.annotateSelects(F, this, &CountPosition);
   assert(CountPosition == ProfileCountSize);
 
   DEBUG(FuncInfo.dumpInfo("after reading profile."));
 }
 
-static void setProfMetadata(Module *M, Instruction *TI,
-                            ArrayRef<uint64_t> EdgeCounts, uint64_t MaxCount) {
-  MDBuilder MDB(M->getContext());
-  assert(MaxCount > 0 && "Bad max count");
-  uint64_t Scale = calculateCountScale(MaxCount);
-  SmallVector<unsigned, 4> Weights;
-  for (const auto &ECI : EdgeCounts)
-    Weights.push_back(scaleBranchCount(ECI, Scale));
-
-  DEBUG(dbgs() << "Weight is: ";
-        for (const auto &W : Weights) { dbgs() << W << " "; }
-        dbgs() << "\n";);
-  TI->setMetadata(llvm::LLVMContext::MD_prof, MDB.createBranchWeights(Weights));
-}
-
 // Assign the scaled count values to the BB with multiple out edges.
 void PGOUseFunc::setBranchWeights() {
   // Generate MD_prof metadata for every branch instruction.
   DEBUG(dbgs() << "\nSetting branch weights.\n");
   for (auto &BB : F) {
     TerminatorInst *TI = BB.getTerminator();
     if (TI->getNumSuccessors() < 2)
       continue;
bool PGOInstrumentationUseLegacyPass::runOnModule(Module &M) {
   auto LookupBFI = [this](Function &F) {
     return &this->getAnalysis<BlockFrequencyInfoWrapperPass>(F).getBFI();
   };
 
   return annotateAllFunctions(M, ProfileFileName, LookupBPI, LookupBFI);
 }
 
 namespace llvm {
+void setProfMetadata(Module *M, Instruction *TI,
+                     ArrayRef<uint64_t> EdgeCounts, uint64_t MaxCount) {
+  MDBuilder MDB(M->getContext());
+  assert(MaxCount > 0 && "Bad max count");
+  uint64_t Scale = calculateCountScale(MaxCount);
+  SmallVector<unsigned, 4> Weights;
+  for (const auto &ECI : EdgeCounts)
+    Weights.push_back(scaleBranchCount(ECI, Scale));
+
+  DEBUG(dbgs() << "Weight is: ";
+        for (const auto &W : Weights) { dbgs() << W << " "; }
+        dbgs() << "\n";);
+  TI->setMetadata(llvm::LLVMContext::MD_prof, MDB.createBranchWeights(Weights));
+}
+
 template <> struct GraphTraits<PGOUseFunc *> {
   typedef const BasicBlock *NodeRef;
   typedef succ_const_iterator ChildIteratorType;
   typedef pointer_iterator<Function::const_iterator> nodes_iterator;
 
   static NodeRef getEntryNode(const PGOUseFunc *G) {
     return &G->getFunc().front();
   }

test/Transforms/PGOProfile/memop_size_opt.ll

+; RUN: opt < %s -passes=pgo-memop-opt -pgo-memop-count-threshold=90 -pgo-memop-percent-threshold=15 -S | FileCheck %s --check-prefix=MEMOP_OPT
+; RUN: opt < %s -pgo-memop-opt -pgo-memop-count-threshold=90 -pgo-memop-percent-threshold=15 -S | FileCheck %s --check-prefix=MEMOP_OPT
+
+target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
+target triple = "x86_64-unknown-linux-gnu"
+
+define void @foo(i8* %dst, i8* %src, i32* %a, i32 %n) !prof !27 {
+entry:
+  br label %for.cond
+
+for.cond:
+  %i.0 = phi i32 [ 0, %entry ], [ %inc5, %for.inc4 ]
+  %cmp = icmp slt i32 %i.0, %n
+  br i1 %cmp, label %for.body, label %for.end6, !prof !28
+
+for.body:
+  br label %for.cond1
+
+for.cond1:
+  %j.0 = phi i32 [ 0, %for.body ], [ %inc, %for.inc ]
+  %idx.ext = sext i32 %i.0 to i64
+  %add.ptr = getelementptr inbounds i32, i32* %a, i64 %idx.ext
+  %0 = load i32, i32* %add.ptr, align 4
+  %cmp2 = icmp slt i32 %j.0, %0
+  br i1 %cmp2, label %for.body3, label %for.end, !prof !29
+
+for.body3:
+  %add = add nsw i32 %i.0, 1
+  %conv = sext i32 %add to i64
+  call void @llvm.memcpy.p0i8.p0i8.i64(i8* %dst, i8* %src, i64 %conv, i32 1, i1 false), !prof !30
+  br label %for.inc
+
+; MEMOP_OPT:  switch i64 %conv, label %[[Default_LABEL:.*]] [
+; MEMOP_OPT:    i64 1, label %[[CASE_1_LABEL:.*]]
+; MEMOP_OPT:  ], !prof [[SWITCH_BW:![0-9]+]] 
+; MEMOP_OPT: [[CASE_1_LABEL]]:
+; MEMOP_OPT:   call void @llvm.memcpy.p0i8.p0i8.i64(i8* %dst, i8* %src, i64 1, i32 1, i1 false)
+; MEMOP_OPT:   br label %[[MERGE_LABEL:.*]]
+; MEMOP_OPT: [[Default_LABEL]]:
+; MEMOP_OPT:   call void @llvm.memcpy.p0i8.p0i8.i64(i8* %dst, i8* %src, i64 %conv, i32 1, i1 false)
+; MEMOP_OPT-NOT:   call void @llvm.memcpy.p0i8.p0i8.i64(i8* %dst, i8* %src, i64 %conv, i32 1, i1 false), !prof
+; MEMOP_OPT:   br label %[[MERGE_LABEL]]
+; MEMOP_OPT: [[MERGE_LABEL]]:
+; MEMOP_OPT:   br label %for.inc
+; MEMOP_OPT: [[SWITCH_BW]] = !{!"branch_weights", i32 457, i32 99}
+
+for.inc:
+  %inc = add nsw i32 %j.0, 1
+  br label %for.cond1
+
+for.end:
+  br label %for.inc4
+
+for.inc4:
+  %inc5 = add nsw i32 %i.0, 1
+  br label %for.cond
+
+for.end6:
+  ret void
+}
+
+!llvm.module.flags = !{!0}
+
+!0 = !{i32 1, !"ProfileSummary", !1}
+!1 = !{!2, !3, !4, !5, !6, !7, !8, !9}
+!2 = !{!"ProfileFormat", !"InstrProf"}
+!3 = !{!"TotalCount", i64 579}
+!4 = !{!"MaxCount", i64 556}
+!5 = !{!"MaxInternalCount", i64 20}
+!6 = !{!"MaxFunctionCount", i64 556}
+!7 = !{!"NumCounts", i64 6}
+!8 = !{!"NumFunctions", i64 3}
+!9 = !{!"DetailedSummary", !10}
+!10 = !{!11, !12, !13, !14, !15, !16, !16, !17, !17, !18, !19, !20, !21, !22, !23, !24, !25, !26}
+!11 = !{i32 10000, i64 556, i32 1}
+!12 = !{i32 100000, i64 556, i32 1}
+!13 = !{i32 200000, i64 556, i32 1}
+!14 = !{i32 300000, i64 556, i32 1}
+!15 = !{i32 400000, i64 556, i32 1}
+!16 = !{i32 500000, i64 556, i32 1}
+!17 = !{i32 600000, i64 556, i32 1}
+!18 = !{i32 700000, i64 556, i32 1}
+!19 = !{i32 800000, i64 556, i32 1}
+!20 = !{i32 900000, i64 556, i32 1}
+!21 = !{i32 950000, i64 556, i32 1}
+!22 = !{i32 990000, i64 20, i32 2}
+!23 = !{i32 999000, i64 1, i32 5}
+!24 = !{i32 999900, i64 1, i32 5}
+!25 = !{i32 999990, i64 1, i32 5}
+!26 = !{i32 999999, i64 1, i32 5}
+!27 = !{!"function_entry_count", i64 1}
+!28 = !{!"branch_weights", i32 20, i32 1}
+!29 = !{!"branch_weights", i32 556, i32 20}
+!30 = !{!"VP", i32 1, i64 556, i64 1, i64 99, i64 2, i64 88, i64 3, i64 77, i64 9, i64 72, i64 4, i64 66, i64 5, i64 55, i64 6, i64 44, i64 7, i64 33, i64 8, i64 22}
+
+declare void @llvm.lifetime.start(i64, i8* nocapture)
+
+declare void @llvm.memcpy.p0i8.p0i8.i64(i8* nocapture writeonly, i8* nocapture readonly, i64, i32, i1)
+
+declare void @llvm.lifetime.end(i64, i8* nocapture)