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/OptimizationRemarkEmitter.h" 24 #include "llvm/Analysis/TargetLibraryInfo.h" 25 #include "llvm/IR/BasicBlock.h" 26 #include "llvm/IR/DerivedTypes.h" 27 #include "llvm/IR/Dominators.h" 28 #include "llvm/IR/Function.h" 29 #include "llvm/IR/IRBuilder.h" 30 #include "llvm/IR/InstVisitor.h" 31 #include "llvm/IR/Instruction.h" 32 #include "llvm/IR/Instructions.h" 33 #include "llvm/IR/LLVMContext.h" 34 #include "llvm/IR/PassManager.h" 35 #include "llvm/IR/Type.h" 36 #include "llvm/ProfileData/InstrProf.h" 37 #define INSTR_PROF_VALUE_PROF_MEMOP_API 38 #include "llvm/ProfileData/InstrProfData.inc" 39 #include "llvm/Support/Casting.h" 40 #include "llvm/Support/CommandLine.h" 41 #include "llvm/Support/Debug.h" 42 #include "llvm/Support/ErrorHandling.h" 43 #include "llvm/Support/MathExtras.h" 44 #include "llvm/Transforms/Instrumentation/PGOInstrumentation.h" 45 #include "llvm/Transforms/Utils/BasicBlockUtils.h" 46 #include <cassert> 47 #include <cstdint> 48 #include <vector> 49 50 using namespace llvm; 51 52 #define DEBUG_TYPE "pgo-memop-opt" 53 54 STATISTIC(NumOfPGOMemOPOpt, "Number of memop intrinsics optimized."); 55 STATISTIC(NumOfPGOMemOPAnnotate, "Number of memop intrinsics annotated."); 56 57 // The minimum call count to optimize memory intrinsic calls. 58 static cl::opt<unsigned> 59 MemOPCountThreshold("pgo-memop-count-threshold", cl::Hidden, cl::init(1000), 60 cl::desc("The minimum count to optimize memory " 61 "intrinsic calls")); 62 63 // Command line option to disable memory intrinsic optimization. The default is 64 // false. This is for debug purpose. 65 static cl::opt<bool> DisableMemOPOPT("disable-memop-opt", cl::init(false), 66 cl::Hidden, cl::desc("Disable optimize")); 67 68 // The percent threshold to optimize memory intrinsic calls. 69 static cl::opt<unsigned> 70 MemOPPercentThreshold("pgo-memop-percent-threshold", cl::init(40), 71 cl::Hidden, cl::ZeroOrMore, 72 cl::desc("The percentage threshold for the " 73 "memory intrinsic calls optimization")); 74 75 // Maximum number of versions for optimizing memory intrinsic call. 76 static cl::opt<unsigned> 77 MemOPMaxVersion("pgo-memop-max-version", cl::init(3), cl::Hidden, 78 cl::ZeroOrMore, 79 cl::desc("The max version for the optimized memory " 80 " intrinsic calls")); 81 82 // Scale the counts from the annotation using the BB count value. 83 static cl::opt<bool> 84 MemOPScaleCount("pgo-memop-scale-count", cl::init(true), cl::Hidden, 85 cl::desc("Scale the memop size counts using the basic " 86 " block count value")); 87 88 cl::opt<bool> 89 MemOPOptMemcmpBcmp("pgo-memop-optimize-memcmp-bcmp", cl::init(true), 90 cl::Hidden, 91 cl::desc("Size-specialize memcmp and bcmp calls")); 92 93 static cl::opt<unsigned> 94 MemOpMaxOptSize("memop-value-prof-max-opt-size", cl::Hidden, cl::init(128), 95 cl::desc("Optimize the memop size <= this value")); 96 97 namespace { 98 99 static const char *getMIName(const MemIntrinsic *MI) { 100 switch (MI->getIntrinsicID()) { 101 case Intrinsic::memcpy: 102 return "memcpy"; 103 case Intrinsic::memmove: 104 return "memmove"; 105 case Intrinsic::memset: 106 return "memset"; 107 default: 108 return "unknown"; 109 } 110 } 111 112 // A class that abstracts a memop (memcpy, memmove, memset, memcmp and bcmp). 113 struct MemOp { 114 Instruction *I; 115 MemOp(MemIntrinsic *MI) : I(MI) {} 116 MemOp(CallInst *CI) : I(CI) {} 117 MemIntrinsic *asMI() { return dyn_cast<MemIntrinsic>(I); } 118 CallInst *asCI() { return cast<CallInst>(I); } 119 MemOp clone() { 120 if (auto MI = asMI()) 121 return MemOp(cast<MemIntrinsic>(MI->clone())); 122 return MemOp(cast<CallInst>(asCI()->clone())); 123 } 124 Value *getLength() { 125 if (auto MI = asMI()) 126 return MI->getLength(); 127 return asCI()->getArgOperand(2); 128 } 129 void setLength(Value *Length) { 130 if (auto MI = asMI()) 131 return MI->setLength(Length); 132 asCI()->setArgOperand(2, Length); 133 } 134 StringRef getFuncName() { 135 if (auto MI = asMI()) 136 return MI->getCalledFunction()->getName(); 137 return asCI()->getCalledFunction()->getName(); 138 } 139 bool isMemmove() { 140 if (auto MI = asMI()) 141 if (MI->getIntrinsicID() == Intrinsic::memmove) 142 return true; 143 return false; 144 } 145 bool isMemcmp(TargetLibraryInfo &TLI) { 146 LibFunc Func; 147 if (asMI() == nullptr && TLI.getLibFunc(*asCI(), Func) && 148 Func == LibFunc_memcmp) { 149 return true; 150 } 151 return false; 152 } 153 bool isBcmp(TargetLibraryInfo &TLI) { 154 LibFunc Func; 155 if (asMI() == nullptr && TLI.getLibFunc(*asCI(), Func) && 156 Func == LibFunc_bcmp) { 157 return true; 158 } 159 return false; 160 } 161 const char *getName(TargetLibraryInfo &TLI) { 162 if (auto MI = asMI()) 163 return getMIName(MI); 164 LibFunc Func; 165 if (TLI.getLibFunc(*asCI(), Func)) { 166 if (Func == LibFunc_memcmp) 167 return "memcmp"; 168 if (Func == LibFunc_bcmp) 169 return "bcmp"; 170 } 171 llvm_unreachable("Must be MemIntrinsic or memcmp/bcmp CallInst"); 172 return nullptr; 173 } 174 }; 175 176 class MemOPSizeOpt : public InstVisitor<MemOPSizeOpt> { 177 public: 178 MemOPSizeOpt(Function &Func, BlockFrequencyInfo &BFI, 179 OptimizationRemarkEmitter &ORE, DominatorTree *DT, 180 TargetLibraryInfo &TLI) 181 : Func(Func), BFI(BFI), ORE(ORE), DT(DT), TLI(TLI), Changed(false) { 182 ValueDataArray = 183 std::make_unique<InstrProfValueData[]>(INSTR_PROF_NUM_BUCKETS); 184 } 185 bool isChanged() const { return Changed; } 186 void perform() { 187 WorkList.clear(); 188 visit(Func); 189 190 for (auto &MO : WorkList) { 191 ++NumOfPGOMemOPAnnotate; 192 if (perform(MO)) { 193 Changed = true; 194 ++NumOfPGOMemOPOpt; 195 LLVM_DEBUG(dbgs() << "MemOP call: " << MO.getFuncName() 196 << "is Transformed.\n"); 197 } 198 } 199 } 200 201 void visitMemIntrinsic(MemIntrinsic &MI) { 202 Value *Length = MI.getLength(); 203 // Not perform on constant length calls. 204 if (isa<ConstantInt>(Length)) 205 return; 206 WorkList.push_back(MemOp(&MI)); 207 } 208 209 void visitCallInst(CallInst &CI) { 210 LibFunc Func; 211 if (TLI.getLibFunc(CI, Func) && 212 (Func == LibFunc_memcmp || Func == LibFunc_bcmp) && 213 !isa<ConstantInt>(CI.getArgOperand(2))) { 214 WorkList.push_back(MemOp(&CI)); 215 } 216 } 217 218 private: 219 Function &Func; 220 BlockFrequencyInfo &BFI; 221 OptimizationRemarkEmitter &ORE; 222 DominatorTree *DT; 223 TargetLibraryInfo &TLI; 224 bool Changed; 225 std::vector<MemOp> WorkList; 226 // The space to read the profile annotation. 227 std::unique_ptr<InstrProfValueData[]> ValueDataArray; 228 bool perform(MemOp MO); 229 }; 230 231 static bool isProfitable(uint64_t Count, uint64_t TotalCount) { 232 assert(Count <= TotalCount); 233 if (Count < MemOPCountThreshold) 234 return false; 235 if (Count < TotalCount * MemOPPercentThreshold / 100) 236 return false; 237 return true; 238 } 239 240 static inline uint64_t getScaledCount(uint64_t Count, uint64_t Num, 241 uint64_t Denom) { 242 if (!MemOPScaleCount) 243 return Count; 244 bool Overflowed; 245 uint64_t ScaleCount = SaturatingMultiply(Count, Num, &Overflowed); 246 return ScaleCount / Denom; 247 } 248 249 bool MemOPSizeOpt::perform(MemOp MO) { 250 assert(MO.I); 251 if (MO.isMemmove()) 252 return false; 253 if (!MemOPOptMemcmpBcmp && (MO.isMemcmp(TLI) || MO.isBcmp(TLI))) 254 return false; 255 256 uint32_t NumVals, MaxNumVals = INSTR_PROF_NUM_BUCKETS; 257 uint64_t TotalCount; 258 if (!getValueProfDataFromInst(*MO.I, IPVK_MemOPSize, MaxNumVals, 259 ValueDataArray.get(), NumVals, TotalCount)) 260 return false; 261 262 uint64_t ActualCount = TotalCount; 263 uint64_t SavedTotalCount = TotalCount; 264 if (MemOPScaleCount) { 265 auto BBEdgeCount = BFI.getBlockProfileCount(MO.I->getParent()); 266 if (!BBEdgeCount) 267 return false; 268 ActualCount = *BBEdgeCount; 269 } 270 271 ArrayRef<InstrProfValueData> VDs(ValueDataArray.get(), NumVals); 272 LLVM_DEBUG(dbgs() << "Read one memory intrinsic profile with count " 273 << ActualCount << "\n"); 274 LLVM_DEBUG( 275 for (auto &VD 276 : VDs) { dbgs() << " (" << VD.Value << "," << VD.Count << ")\n"; }); 277 278 if (ActualCount < MemOPCountThreshold) 279 return false; 280 // Skip if the total value profiled count is 0, in which case we can't 281 // scale up the counts properly (and there is no profitable transformation). 282 if (TotalCount == 0) 283 return false; 284 285 TotalCount = ActualCount; 286 if (MemOPScaleCount) 287 LLVM_DEBUG(dbgs() << "Scale counts: numerator = " << ActualCount 288 << " denominator = " << SavedTotalCount << "\n"); 289 290 // Keeping track of the count of the default case: 291 uint64_t RemainCount = TotalCount; 292 uint64_t SavedRemainCount = SavedTotalCount; 293 SmallVector<uint64_t, 16> SizeIds; 294 SmallVector<uint64_t, 16> CaseCounts; 295 uint64_t MaxCount = 0; 296 unsigned Version = 0; 297 int64_t LastV = -1; 298 // Default case is in the front -- save the slot here. 299 CaseCounts.push_back(0); 300 SmallVector<InstrProfValueData, 24> RemainingVDs; 301 for (auto I = VDs.begin(), E = VDs.end(); I != E; ++I) { 302 auto &VD = *I; 303 int64_t V = VD.Value; 304 uint64_t C = VD.Count; 305 if (MemOPScaleCount) 306 C = getScaledCount(C, ActualCount, SavedTotalCount); 307 308 if (!InstrProfIsSingleValRange(V) || V > MemOpMaxOptSize) { 309 RemainingVDs.push_back(VD); 310 continue; 311 } 312 313 // ValueCounts are sorted on the count. Break at the first un-profitable 314 // value. 315 if (!isProfitable(C, RemainCount)) { 316 RemainingVDs.insert(RemainingVDs.end(), I, E); 317 break; 318 } 319 320 if (V == LastV) { 321 LLVM_DEBUG(dbgs() << "Invalid Profile Data in Function " << Func.getName() 322 << ": Two consecutive, identical values in MemOp value" 323 "counts.\n"); 324 return false; 325 } 326 327 LastV = V; 328 329 SizeIds.push_back(V); 330 CaseCounts.push_back(C); 331 if (C > MaxCount) 332 MaxCount = C; 333 334 assert(RemainCount >= C); 335 RemainCount -= C; 336 assert(SavedRemainCount >= VD.Count); 337 SavedRemainCount -= VD.Count; 338 339 if (++Version >= MemOPMaxVersion && MemOPMaxVersion != 0) { 340 RemainingVDs.insert(RemainingVDs.end(), I + 1, E); 341 break; 342 } 343 } 344 345 if (Version == 0) 346 return false; 347 348 CaseCounts[0] = RemainCount; 349 if (RemainCount > MaxCount) 350 MaxCount = RemainCount; 351 352 uint64_t SumForOpt = TotalCount - RemainCount; 353 354 LLVM_DEBUG(dbgs() << "Optimize one memory intrinsic call to " << Version 355 << " Versions (covering " << SumForOpt << " out of " 356 << TotalCount << ")\n"); 357 358 // mem_op(..., size) 359 // ==> 360 // switch (size) { 361 // case s1: 362 // mem_op(..., s1); 363 // goto merge_bb; 364 // case s2: 365 // mem_op(..., s2); 366 // goto merge_bb; 367 // ... 368 // default: 369 // mem_op(..., size); 370 // goto merge_bb; 371 // } 372 // merge_bb: 373 374 BasicBlock *BB = MO.I->getParent(); 375 LLVM_DEBUG(dbgs() << "\n\n== Basic Block Before ==\n"); 376 LLVM_DEBUG(dbgs() << *BB << "\n"); 377 auto OrigBBFreq = BFI.getBlockFreq(BB); 378 379 BasicBlock *DefaultBB = SplitBlock(BB, MO.I, DT); 380 BasicBlock::iterator It(*MO.I); 381 ++It; 382 assert(It != DefaultBB->end()); 383 BasicBlock *MergeBB = SplitBlock(DefaultBB, &(*It), DT); 384 MergeBB->setName("MemOP.Merge"); 385 BFI.setBlockFreq(MergeBB, OrigBBFreq.getFrequency()); 386 DefaultBB->setName("MemOP.Default"); 387 388 DomTreeUpdater DTU(DT, DomTreeUpdater::UpdateStrategy::Eager); 389 auto &Ctx = Func.getContext(); 390 IRBuilder<> IRB(BB); 391 BB->getTerminator()->eraseFromParent(); 392 Value *SizeVar = MO.getLength(); 393 SwitchInst *SI = IRB.CreateSwitch(SizeVar, DefaultBB, SizeIds.size()); 394 Type *MemOpTy = MO.I->getType(); 395 PHINode *PHI = nullptr; 396 if (!MemOpTy->isVoidTy()) { 397 // Insert a phi for the return values at the merge block. 398 IRBuilder<> IRBM(MergeBB->getFirstNonPHI()); 399 PHI = IRBM.CreatePHI(MemOpTy, SizeIds.size() + 1, "MemOP.RVMerge"); 400 MO.I->replaceAllUsesWith(PHI); 401 PHI->addIncoming(MO.I, DefaultBB); 402 } 403 404 // Clear the value profile data. 405 MO.I->setMetadata(LLVMContext::MD_prof, nullptr); 406 // If all promoted, we don't need the MD.prof metadata. 407 if (SavedRemainCount > 0 || Version != NumVals) { 408 // Otherwise we need update with the un-promoted records back. 409 ArrayRef<InstrProfValueData> RemVDs(RemainingVDs); 410 annotateValueSite(*Func.getParent(), *MO.I, RemVDs, SavedRemainCount, 411 IPVK_MemOPSize, NumVals); 412 } 413 414 LLVM_DEBUG(dbgs() << "\n\n== Basic Block After==\n"); 415 416 std::vector<DominatorTree::UpdateType> Updates; 417 if (DT) 418 Updates.reserve(2 * SizeIds.size()); 419 420 for (uint64_t SizeId : SizeIds) { 421 BasicBlock *CaseBB = BasicBlock::Create( 422 Ctx, Twine("MemOP.Case.") + Twine(SizeId), &Func, DefaultBB); 423 MemOp NewMO = MO.clone(); 424 // Fix the argument. 425 auto *SizeType = dyn_cast<IntegerType>(NewMO.getLength()->getType()); 426 assert(SizeType && "Expected integer type size argument."); 427 ConstantInt *CaseSizeId = ConstantInt::get(SizeType, SizeId); 428 NewMO.setLength(CaseSizeId); 429 CaseBB->getInstList().push_back(NewMO.I); 430 IRBuilder<> IRBCase(CaseBB); 431 IRBCase.CreateBr(MergeBB); 432 SI->addCase(CaseSizeId, CaseBB); 433 if (!MemOpTy->isVoidTy()) 434 PHI->addIncoming(NewMO.I, CaseBB); 435 if (DT) { 436 Updates.push_back({DominatorTree::Insert, CaseBB, MergeBB}); 437 Updates.push_back({DominatorTree::Insert, BB, CaseBB}); 438 } 439 LLVM_DEBUG(dbgs() << *CaseBB << "\n"); 440 } 441 DTU.applyUpdates(Updates); 442 Updates.clear(); 443 444 setProfMetadata(Func.getParent(), SI, CaseCounts, MaxCount); 445 446 LLVM_DEBUG(dbgs() << *BB << "\n"); 447 LLVM_DEBUG(dbgs() << *DefaultBB << "\n"); 448 LLVM_DEBUG(dbgs() << *MergeBB << "\n"); 449 450 ORE.emit([&]() { 451 using namespace ore; 452 return OptimizationRemark(DEBUG_TYPE, "memopt-opt", MO.I) 453 << "optimized " << NV("Memop", MO.getName(TLI)) << " with count " 454 << NV("Count", SumForOpt) << " out of " << NV("Total", TotalCount) 455 << " for " << NV("Versions", Version) << " versions"; 456 }); 457 458 return true; 459 } 460 } // namespace 461 462 static bool PGOMemOPSizeOptImpl(Function &F, BlockFrequencyInfo &BFI, 463 OptimizationRemarkEmitter &ORE, 464 DominatorTree *DT, TargetLibraryInfo &TLI) { 465 if (DisableMemOPOPT) 466 return false; 467 468 if (F.hasFnAttribute(Attribute::OptimizeForSize)) 469 return false; 470 MemOPSizeOpt MemOPSizeOpt(F, BFI, ORE, DT, TLI); 471 MemOPSizeOpt.perform(); 472 return MemOPSizeOpt.isChanged(); 473 } 474 475 PreservedAnalyses PGOMemOPSizeOpt::run(Function &F, 476 FunctionAnalysisManager &FAM) { 477 auto &BFI = FAM.getResult<BlockFrequencyAnalysis>(F); 478 auto &ORE = FAM.getResult<OptimizationRemarkEmitterAnalysis>(F); 479 auto *DT = FAM.getCachedResult<DominatorTreeAnalysis>(F); 480 auto &TLI = FAM.getResult<TargetLibraryAnalysis>(F); 481 bool Changed = PGOMemOPSizeOptImpl(F, BFI, ORE, DT, TLI); 482 if (!Changed) 483 return PreservedAnalyses::all(); 484 auto PA = PreservedAnalyses(); 485 PA.preserve<DominatorTreeAnalysis>(); 486 return PA; 487 } 488