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