lib/Transforms/Instrumentation/PGOMemOPSizeOpt.cpp

   1 //===-- PGOMemOPSizeOpt.cpp - Optimizations based on value profiling ===//
   2 //
   3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
   4 // See https://llvm.org/LICENSE.txt for license information.
   5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
   6 //
   7 //===----------------------------------------------------------------------===//
   8 //
   9 // This file implements the transformation that optimizes memory intrinsics
  10 // such as memcpy using the size value profile. When memory intrinsic size
  11 // value profile metadata is available, a single memory intrinsic is expanded
  12 // to a sequence of guarded specialized versions that are called with the
  13 // hottest size(s), for later expansion into more optimal inline sequences.
  14 //
  15 //===----------------------------------------------------------------------===//
  16
  17 #include "llvm/ADT/ArrayRef.h"
  18 #include "llvm/ADT/Statistic.h"
  19 #include "llvm/ADT/StringRef.h"
  20 #include "llvm/ADT/Twine.h"
  21 #include "llvm/Analysis/BlockFrequencyInfo.h"
  22 #include "llvm/Analysis/DomTreeUpdater.h"
  23 #include "llvm/Analysis/GlobalsModRef.h"
  24 #include "llvm/Analysis/OptimizationRemarkEmitter.h"
  25 #include "llvm/IR/BasicBlock.h"
  26 #include "llvm/IR/CallSite.h"
  27 #include "llvm/IR/DerivedTypes.h"
  28 #include "llvm/IR/Dominators.h"
  29 #include "llvm/IR/Function.h"
  30 #include "llvm/IR/IRBuilder.h"
  31 #include "llvm/IR/InstVisitor.h"
  32 #include "llvm/IR/InstrTypes.h"
  33 #include "llvm/IR/Instruction.h"
  34 #include "llvm/IR/Instructions.h"
  35 #include "llvm/IR/LLVMContext.h"
  36 #include "llvm/IR/PassManager.h"
  37 #include "llvm/IR/Type.h"
  38 #include "llvm/Pass.h"
  39 #include "llvm/PassRegistry.h"
  40 #include "llvm/PassSupport.h"
  41 #include "llvm/ProfileData/InstrProf.h"
  42 #include "llvm/Support/Casting.h"
  43 #include "llvm/Support/CommandLine.h"
  44 #include "llvm/Support/Debug.h"
  45 #include "llvm/Support/ErrorHandling.h"
  46 #include "llvm/Support/MathExtras.h"
  47 #include "llvm/Transforms/Instrumentation.h"
  48 #include "llvm/Transforms/Instrumentation/PGOInstrumentation.h"
  49 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
  50 #include <cassert>
  51 #include <cstdint>
  52 #include <vector>
  53
  54 using namespace llvm;
  55
  56 #define DEBUG_TYPE "pgo-memop-opt"
  57
  58 STATISTIC(NumOfPGOMemOPOpt, "Number of memop intrinsics optimized.");
  59 STATISTIC(NumOfPGOMemOPAnnotate, "Number of memop intrinsics annotated.");
  60
  61 // The minimum call count to optimize memory intrinsic calls.
  62 static cl::opt<unsigned>
  63     MemOPCountThreshold("pgo-memop-count-threshold", cl::Hidden, cl::ZeroOrMore,
  64                         cl::init(1000),
  65                         cl::desc("The minimum count to optimize memory "
  66                                  "intrinsic calls"));
  67
  68 // Command line option to disable memory intrinsic optimization. The default is
  69 // false. This is for debug purpose.
  70 static cl::opt<bool> DisableMemOPOPT("disable-memop-opt", cl::init(false),
  71                                      cl::Hidden, cl::desc("Disable optimize"));
  72
  73 // The percent threshold to optimize memory intrinsic calls.
  74 static cl::opt<unsigned>
  75     MemOPPercentThreshold("pgo-memop-percent-threshold", cl::init(40),
  76                           cl::Hidden, cl::ZeroOrMore,
  77                           cl::desc("The percentage threshold for the "
  78                                    "memory intrinsic calls optimization"));
  79
  80 // Maximum number of versions for optimizing memory intrinsic call.
  81 static cl::opt<unsigned>
  82     MemOPMaxVersion("pgo-memop-max-version", cl::init(3), cl::Hidden,
  83                     cl::ZeroOrMore,
  84                     cl::desc("The max version for the optimized memory "
  85                              " intrinsic calls"));
  86
  87 // Scale the counts from the annotation using the BB count value.
  88 static cl::opt<bool>
  89     MemOPScaleCount("pgo-memop-scale-count", cl::init(true), cl::Hidden,
  90                     cl::desc("Scale the memop size counts using the basic "
  91                              " block count value"));
  92
  93 // This option sets the rangge of precise profile memop sizes.
  94 extern cl::opt<std::string> MemOPSizeRange;
  95
  96 // This option sets the value that groups large memop sizes
  97 extern cl::opt<unsigned> MemOPSizeLarge;
  98
  99 namespace {
 100 class PGOMemOPSizeOptLegacyPass : public FunctionPass {
 101 public:
 102   static char ID;
 103
 104   PGOMemOPSizeOptLegacyPass() : FunctionPass(ID) {
 105     initializePGOMemOPSizeOptLegacyPassPass(*PassRegistry::getPassRegistry());
 106   }
 107
 108   StringRef getPassName() const override { return "PGOMemOPSize"; }
 109
 110 private:
 111   bool runOnFunction(Function &F) override;
 112   void getAnalysisUsage(AnalysisUsage &AU) const override {
 113     AU.addRequired<BlockFrequencyInfoWrapperPass>();
 114     AU.addRequired<OptimizationRemarkEmitterWrapperPass>();
 115     AU.addPreserved<GlobalsAAWrapperPass>();
 116     AU.addPreserved<DominatorTreeWrapperPass>();
 117   }
 118 };
 119 } // end anonymous namespace
 120
 121 char PGOMemOPSizeOptLegacyPass::ID = 0;
 122 INITIALIZE_PASS_BEGIN(PGOMemOPSizeOptLegacyPass, "pgo-memop-opt",
 123                       "Optimize memory intrinsic using its size value profile",
 124                       false, false)
 125 INITIALIZE_PASS_DEPENDENCY(BlockFrequencyInfoWrapperPass)
 126 INITIALIZE_PASS_END(PGOMemOPSizeOptLegacyPass, "pgo-memop-opt",
 127                     "Optimize memory intrinsic using its size value profile",
 128                     false, false)
 129
 130 FunctionPass *llvm::createPGOMemOPSizeOptLegacyPass() {
 131   return new PGOMemOPSizeOptLegacyPass();
 132 }
 133
 134 namespace {
 135 class MemOPSizeOpt : public InstVisitor<MemOPSizeOpt> {
 136 public:
 137   MemOPSizeOpt(Function &Func, BlockFrequencyInfo &BFI,
 138                OptimizationRemarkEmitter &ORE, DominatorTree *DT)
 139       : Func(Func), BFI(BFI), ORE(ORE), DT(DT), Changed(false) {
 140     ValueDataArray =
 141         std::make_unique<InstrProfValueData[]>(MemOPMaxVersion + 2);
 142     // Get the MemOPSize range information from option MemOPSizeRange,
 143     getMemOPSizeRangeFromOption(MemOPSizeRange, PreciseRangeStart,
 144                                 PreciseRangeLast);
 145   }
 146   bool isChanged() const { return Changed; }
 147   void perform() {
 148     WorkList.clear();
 149     visit(Func);
 150
 151     for (auto &MI : WorkList) {
 152       ++NumOfPGOMemOPAnnotate;
 153       if (perform(MI)) {
 154         Changed = true;
 155         ++NumOfPGOMemOPOpt;
 156         LLVM_DEBUG(dbgs() << "MemOP call: "
 157                           << MI->getCalledFunction()->getName()
 158                           << "is Transformed.\n");
 159       }
 160     }
 161   }
 162
 163   void visitMemIntrinsic(MemIntrinsic &MI) {
 164     Value *Length = MI.getLength();
 165     // Not perform on constant length calls.
 166     if (dyn_cast<ConstantInt>(Length))
 167       return;
 168     WorkList.push_back(&MI);
 169   }
 170
 171 private:
 172   Function &Func;
 173   BlockFrequencyInfo &BFI;
 174   OptimizationRemarkEmitter &ORE;
 175   DominatorTree *DT;
 176   bool Changed;
 177   std::vector<MemIntrinsic *> WorkList;
 178   // Start of the previse range.
 179   int64_t PreciseRangeStart;
 180   // Last value of the previse range.
 181   int64_t PreciseRangeLast;
 182   // The space to read the profile annotation.
 183   std::unique_ptr<InstrProfValueData[]> ValueDataArray;
 184   bool perform(MemIntrinsic *MI);
 185
 186   // This kind shows which group the value falls in. For PreciseValue, we have
 187   // the profile count for that value. LargeGroup groups the values that are in
 188   // range [LargeValue, +inf). NonLargeGroup groups the rest of values.
 189   enum MemOPSizeKind { PreciseValue, NonLargeGroup, LargeGroup };
 190
 191   MemOPSizeKind getMemOPSizeKind(int64_t Value) const {
 192     if (Value == MemOPSizeLarge && MemOPSizeLarge != 0)
 193       return LargeGroup;
 194     if (Value == PreciseRangeLast + 1)
 195       return NonLargeGroup;
 196     return PreciseValue;
 197   }
 198 };
 199
 200 static const char *getMIName(const MemIntrinsic *MI) {
 201   switch (MI->getIntrinsicID()) {
 202   case Intrinsic::memcpy:
 203     return "memcpy";
 204   case Intrinsic::memmove:
 205     return "memmove";
 206   case Intrinsic::memset:
 207     return "memset";
 208   default:
 209     return "unknown";
 210   }
 211 }
 212
 213 static bool isProfitable(uint64_t Count, uint64_t TotalCount) {
 214   assert(Count <= TotalCount);
 215   if (Count < MemOPCountThreshold)
 216     return false;
 217   if (Count < TotalCount * MemOPPercentThreshold / 100)
 218     return false;
 219   return true;
 220 }
 221
 222 static inline uint64_t getScaledCount(uint64_t Count, uint64_t Num,
 223                                       uint64_t Denom) {
 224   if (!MemOPScaleCount)
 225     return Count;
 226   bool Overflowed;
 227   uint64_t ScaleCount = SaturatingMultiply(Count, Num, &Overflowed);
 228   return ScaleCount / Denom;
 229 }
 230
 231 bool MemOPSizeOpt::perform(MemIntrinsic *MI) {
 232   assert(MI);
 233   if (MI->getIntrinsicID() == Intrinsic::memmove)
 234     return false;
 235
 236   uint32_t NumVals, MaxNumPromotions = MemOPMaxVersion + 2;
 237   uint64_t TotalCount;
 238   if (!getValueProfDataFromInst(*MI, IPVK_MemOPSize, MaxNumPromotions,
 239                                 ValueDataArray.get(), NumVals, TotalCount))
 240     return false;
 241
 242   uint64_t ActualCount = TotalCount;
 243   uint64_t SavedTotalCount = TotalCount;
 244   if (MemOPScaleCount) {
 245     auto BBEdgeCount = BFI.getBlockProfileCount(MI->getParent());
 246     if (!BBEdgeCount)
 247       return false;
 248     ActualCount = *BBEdgeCount;
 249   }
 250
 251   ArrayRef<InstrProfValueData> VDs(ValueDataArray.get(), NumVals);
 252   LLVM_DEBUG(dbgs() << "Read one memory intrinsic profile with count "
 253                     << ActualCount << "\n");
 254   LLVM_DEBUG(
 255       for (auto &VD
 256            : VDs) { dbgs() << "  (" << VD.Value << "," << VD.Count << ")\n"; });
 257
 258   if (ActualCount < MemOPCountThreshold)
 259     return false;
 260   // Skip if the total value profiled count is 0, in which case we can't
 261   // scale up the counts properly (and there is no profitable transformation).
 262   if (TotalCount == 0)
 263     return false;
 264
 265   TotalCount = ActualCount;
 266   if (MemOPScaleCount)
 267     LLVM_DEBUG(dbgs() << "Scale counts: numerator = " << ActualCount
 268                       << " denominator = " << SavedTotalCount << "\n");
 269
 270   // Keeping track of the count of the default case:
 271   uint64_t RemainCount = TotalCount;
 272   uint64_t SavedRemainCount = SavedTotalCount;
 273   SmallVector<uint64_t, 16> SizeIds;
 274   SmallVector<uint64_t, 16> CaseCounts;
 275   uint64_t MaxCount = 0;
 276   unsigned Version = 0;
 277   // Default case is in the front -- save the slot here.
 278   CaseCounts.push_back(0);
 279   for (auto &VD : VDs) {
 280     int64_t V = VD.Value;
 281     uint64_t C = VD.Count;
 282     if (MemOPScaleCount)
 283       C = getScaledCount(C, ActualCount, SavedTotalCount);
 284
 285     // Only care precise value here.
 286     if (getMemOPSizeKind(V) != PreciseValue)
 287       continue;
 288
 289     // ValueCounts are sorted on the count. Break at the first un-profitable
 290     // value.
 291     if (!isProfitable(C, RemainCount))
 292       break;
 293
 294     SizeIds.push_back(V);
 295     CaseCounts.push_back(C);
 296     if (C > MaxCount)
 297       MaxCount = C;
 298
 299     assert(RemainCount >= C);
 300     RemainCount -= C;
 301     assert(SavedRemainCount >= VD.Count);
 302     SavedRemainCount -= VD.Count;
 303
 304     if (++Version > MemOPMaxVersion && MemOPMaxVersion != 0)
 305       break;
 306   }
 307
 308   if (Version == 0)
 309     return false;
 310
 311   CaseCounts[0] = RemainCount;
 312   if (RemainCount > MaxCount)
 313     MaxCount = RemainCount;
 314
 315   uint64_t SumForOpt = TotalCount - RemainCount;
 316
 317   LLVM_DEBUG(dbgs() << "Optimize one memory intrinsic call to " << Version
 318                     << " Versions (covering " << SumForOpt << " out of "
 319                     << TotalCount << ")\n");
 320
 321   // mem_op(..., size)
 322   // ==>
 323   // switch (size) {
 324   //   case s1:
 325   //      mem_op(..., s1);
 326   //      goto merge_bb;
 327   //   case s2:
 328   //      mem_op(..., s2);
 329   //      goto merge_bb;
 330   //   ...
 331   //   default:
 332   //      mem_op(..., size);
 333   //      goto merge_bb;
 334   // }
 335   // merge_bb:
 336
 337   BasicBlock *BB = MI->getParent();
 338   LLVM_DEBUG(dbgs() << "\n\n== Basic Block Before ==\n");
 339   LLVM_DEBUG(dbgs() << *BB << "\n");
 340   auto OrigBBFreq = BFI.getBlockFreq(BB);
 341
 342   BasicBlock *DefaultBB = SplitBlock(BB, MI, DT);
 343   BasicBlock::iterator It(*MI);
 344   ++It;
 345   assert(It != DefaultBB->end());
 346   BasicBlock *MergeBB = SplitBlock(DefaultBB, &(*It), DT);
 347   MergeBB->setName("MemOP.Merge");
 348   BFI.setBlockFreq(MergeBB, OrigBBFreq.getFrequency());
 349   DefaultBB->setName("MemOP.Default");
 350
 351   DomTreeUpdater DTU(DT, DomTreeUpdater::UpdateStrategy::Eager);
 352   auto &Ctx = Func.getContext();
 353   IRBuilder<> IRB(BB);
 354   BB->getTerminator()->eraseFromParent();
 355   Value *SizeVar = MI->getLength();
 356   SwitchInst *SI = IRB.CreateSwitch(SizeVar, DefaultBB, SizeIds.size());
 357
 358   // Clear the value profile data.
 359   MI->setMetadata(LLVMContext::MD_prof, nullptr);
 360   // If all promoted, we don't need the MD.prof metadata.
 361   if (SavedRemainCount > 0 || Version != NumVals)
 362     // Otherwise we need update with the un-promoted records back.
 363     annotateValueSite(*Func.getParent(), *MI, VDs.slice(Version),
 364                       SavedRemainCount, IPVK_MemOPSize, NumVals);
 365
 366   LLVM_DEBUG(dbgs() << "\n\n== Basic Block After==\n");
 367
 368   std::vector<DominatorTree::UpdateType> Updates;
 369   if (DT)
 370     Updates.reserve(2 * SizeIds.size());
 371
 372   for (uint64_t SizeId : SizeIds) {
 373     BasicBlock *CaseBB = BasicBlock::Create(
 374         Ctx, Twine("MemOP.Case.") + Twine(SizeId), &Func, DefaultBB);
 375     Instruction *NewInst = MI->clone();
 376     // Fix the argument.
 377     MemIntrinsic * MemI = dyn_cast<MemIntrinsic>(NewInst);
 378     IntegerType *SizeType = dyn_cast<IntegerType>(MemI->getLength()->getType());
 379     assert(SizeType && "Expected integer type size argument.");
 380     ConstantInt *CaseSizeId = ConstantInt::get(SizeType, SizeId);
 381     MemI->setLength(CaseSizeId);
 382     CaseBB->getInstList().push_back(NewInst);
 383     IRBuilder<> IRBCase(CaseBB);
 384     IRBCase.CreateBr(MergeBB);
 385     SI->addCase(CaseSizeId, CaseBB);
 386     if (DT) {
 387       Updates.push_back({DominatorTree::Insert, CaseBB, MergeBB});
 388       Updates.push_back({DominatorTree::Insert, BB, CaseBB});
 389     }
 390     LLVM_DEBUG(dbgs() << *CaseBB << "\n");
 391   }
 392   DTU.applyUpdates(Updates);
 393   Updates.clear();
 394
 395   setProfMetadata(Func.getParent(), SI, CaseCounts, MaxCount);
 396
 397   LLVM_DEBUG(dbgs() << *BB << "\n");
 398   LLVM_DEBUG(dbgs() << *DefaultBB << "\n");
 399   LLVM_DEBUG(dbgs() << *MergeBB << "\n");
 400
 401   ORE.emit([&]() {
 402     using namespace ore;
 403     return OptimizationRemark(DEBUG_TYPE, "memopt-opt", MI)
 404              << "optimized " << NV("Intrinsic", StringRef(getMIName(MI)))
 405              << " with count " << NV("Count", SumForOpt) << " out of "
 406              << NV("Total", TotalCount) << " for " << NV("Versions", Version)
 407              << " versions";
 408   });
 409
 410   return true;
 411 }
 412 } // namespace
 413
 414 static bool PGOMemOPSizeOptImpl(Function &F, BlockFrequencyInfo &BFI,
 415                                 OptimizationRemarkEmitter &ORE,
 416                                 DominatorTree *DT) {
 417   if (DisableMemOPOPT)
 418     return false;
 419
 420   if (F.hasFnAttribute(Attribute::OptimizeForSize))
 421     return false;
 422   MemOPSizeOpt MemOPSizeOpt(F, BFI, ORE, DT);
 423   MemOPSizeOpt.perform();
 424   return MemOPSizeOpt.isChanged();
 425 }
 426
 427 bool PGOMemOPSizeOptLegacyPass::runOnFunction(Function &F) {
 428   BlockFrequencyInfo &BFI =
 429       getAnalysis<BlockFrequencyInfoWrapperPass>().getBFI();
 430   auto &ORE = getAnalysis<OptimizationRemarkEmitterWrapperPass>().getORE();
 431   auto *DTWP = getAnalysisIfAvailable<DominatorTreeWrapperPass>();
 432   DominatorTree *DT = DTWP ? &DTWP->getDomTree() : nullptr;
 433   return PGOMemOPSizeOptImpl(F, BFI, ORE, DT);
 434 }
 435
 436 namespace llvm {
 437 char &PGOMemOPSizeOptID = PGOMemOPSizeOptLegacyPass::ID;
 438
 439 PreservedAnalyses PGOMemOPSizeOpt::run(Function &F,
 440                                        FunctionAnalysisManager &FAM) {
 441   auto &BFI = FAM.getResult<BlockFrequencyAnalysis>(F);
 442   auto &ORE = FAM.getResult<OptimizationRemarkEmitterAnalysis>(F);
 443   auto *DT = FAM.getCachedResult<DominatorTreeAnalysis>(F);
 444   bool Changed = PGOMemOPSizeOptImpl(F, BFI, ORE, DT);
 445   if (!Changed)
 446     return PreservedAnalyses::all();
 447   auto PA = PreservedAnalyses();
 448   PA.preserve<GlobalsAA>();
 449   PA.preserve<DominatorTreeAnalysis>();
 450   return PA;
 451 }
 452 } // namespace llvm