1 //===- PGOInstrumentation.cpp - MST-based PGO Instrumentation -------------===// 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 PGO instrumentation using a minimum spanning tree based 10 // on the following paper: 11 // [1] Donald E. Knuth, Francis R. Stevenson. Optimal measurement of points 12 // for program frequency counts. BIT Numerical Mathematics 1973, Volume 13, 13 // Issue 3, pp 313-322 14 // The idea of the algorithm based on the fact that for each node (except for 15 // the entry and exit), the sum of incoming edge counts equals the sum of 16 // outgoing edge counts. The count of edge on spanning tree can be derived from 17 // those edges not on the spanning tree. Knuth proves this method instruments 18 // the minimum number of edges. 19 // 20 // The minimal spanning tree here is actually a maximum weight tree -- on-tree 21 // edges have higher frequencies (more likely to execute). The idea is to 22 // instrument those less frequently executed edges to reduce the runtime 23 // overhead of instrumented binaries. 24 // 25 // This file contains two passes: 26 // (1) Pass PGOInstrumentationGen which instruments the IR to generate edge 27 // count profile, and generates the instrumentation for indirect call 28 // profiling. 29 // (2) Pass PGOInstrumentationUse which reads the edge count profile and 30 // annotates the branch weights. It also reads the indirect call value 31 // profiling records and annotate the indirect call instructions. 32 // 33 // To get the precise counter information, These two passes need to invoke at 34 // the same compilation point (so they see the same IR). For pass 35 // PGOInstrumentationGen, the real work is done in instrumentOneFunc(). For 36 // pass PGOInstrumentationUse, the real work in done in class PGOUseFunc and 37 // the profile is opened in module level and passed to each PGOUseFunc instance. 38 // The shared code for PGOInstrumentationGen and PGOInstrumentationUse is put 39 // in class FuncPGOInstrumentation. 40 // 41 // Class PGOEdge represents a CFG edge and some auxiliary information. Class 42 // BBInfo contains auxiliary information for each BB. These two classes are used 43 // in pass PGOInstrumentationGen. Class PGOUseEdge and UseBBInfo are the derived 44 // class of PGOEdge and BBInfo, respectively. They contains extra data structure 45 // used in populating profile counters. 46 // The MST implementation is in Class CFGMST (CFGMST.h). 47 // 48 //===----------------------------------------------------------------------===// 49 50 #include "llvm/Transforms/Instrumentation/PGOInstrumentation.h" 51 #include "CFGMST.h" 52 #include "ValueProfileCollector.h" 53 #include "llvm/ADT/APInt.h" 54 #include "llvm/ADT/ArrayRef.h" 55 #include "llvm/ADT/MapVector.h" 56 #include "llvm/ADT/STLExtras.h" 57 #include "llvm/ADT/SmallVector.h" 58 #include "llvm/ADT/Statistic.h" 59 #include "llvm/ADT/StringRef.h" 60 #include "llvm/ADT/Triple.h" 61 #include "llvm/ADT/Twine.h" 62 #include "llvm/ADT/iterator.h" 63 #include "llvm/ADT/iterator_range.h" 64 #include "llvm/Analysis/BlockFrequencyInfo.h" 65 #include "llvm/Analysis/BranchProbabilityInfo.h" 66 #include "llvm/Analysis/CFG.h" 67 #include "llvm/Analysis/EHPersonalities.h" 68 #include "llvm/Analysis/LoopInfo.h" 69 #include "llvm/Analysis/OptimizationRemarkEmitter.h" 70 #include "llvm/Analysis/ProfileSummaryInfo.h" 71 #include "llvm/IR/Attributes.h" 72 #include "llvm/IR/BasicBlock.h" 73 #include "llvm/IR/CFG.h" 74 #include "llvm/IR/Comdat.h" 75 #include "llvm/IR/Constant.h" 76 #include "llvm/IR/Constants.h" 77 #include "llvm/IR/DiagnosticInfo.h" 78 #include "llvm/IR/Dominators.h" 79 #include "llvm/IR/Function.h" 80 #include "llvm/IR/GlobalAlias.h" 81 #include "llvm/IR/GlobalValue.h" 82 #include "llvm/IR/GlobalVariable.h" 83 #include "llvm/IR/IRBuilder.h" 84 #include "llvm/IR/InstVisitor.h" 85 #include "llvm/IR/InstrTypes.h" 86 #include "llvm/IR/Instruction.h" 87 #include "llvm/IR/Instructions.h" 88 #include "llvm/IR/IntrinsicInst.h" 89 #include "llvm/IR/Intrinsics.h" 90 #include "llvm/IR/LLVMContext.h" 91 #include "llvm/IR/MDBuilder.h" 92 #include "llvm/IR/Module.h" 93 #include "llvm/IR/PassManager.h" 94 #include "llvm/IR/ProfileSummary.h" 95 #include "llvm/IR/Type.h" 96 #include "llvm/IR/Value.h" 97 #include "llvm/InitializePasses.h" 98 #include "llvm/Pass.h" 99 #include "llvm/ProfileData/InstrProf.h" 100 #include "llvm/ProfileData/InstrProfReader.h" 101 #include "llvm/Support/BranchProbability.h" 102 #include "llvm/Support/CRC.h" 103 #include "llvm/Support/Casting.h" 104 #include "llvm/Support/CommandLine.h" 105 #include "llvm/Support/DOTGraphTraits.h" 106 #include "llvm/Support/Debug.h" 107 #include "llvm/Support/Error.h" 108 #include "llvm/Support/ErrorHandling.h" 109 #include "llvm/Support/GraphWriter.h" 110 #include "llvm/Support/raw_ostream.h" 111 #include "llvm/Transforms/Instrumentation.h" 112 #include "llvm/Transforms/Utils/BasicBlockUtils.h" 113 #include "llvm/Transforms/Utils/MisExpect.h" 114 #include <algorithm> 115 #include <cassert> 116 #include <cstdint> 117 #include <memory> 118 #include <numeric> 119 #include <string> 120 #include <unordered_map> 121 #include <utility> 122 #include <vector> 123 124 using namespace llvm; 125 using ProfileCount = Function::ProfileCount; 126 using VPCandidateInfo = ValueProfileCollector::CandidateInfo; 127 128 #define DEBUG_TYPE "pgo-instrumentation" 129 130 STATISTIC(NumOfPGOInstrument, "Number of edges instrumented."); 131 STATISTIC(NumOfPGOSelectInsts, "Number of select instruction instrumented."); 132 STATISTIC(NumOfPGOMemIntrinsics, "Number of mem intrinsics instrumented."); 133 STATISTIC(NumOfPGOEdge, "Number of edges."); 134 STATISTIC(NumOfPGOBB, "Number of basic-blocks."); 135 STATISTIC(NumOfPGOSplit, "Number of critical edge splits."); 136 STATISTIC(NumOfPGOFunc, "Number of functions having valid profile counts."); 137 STATISTIC(NumOfPGOMismatch, "Number of functions having mismatch profile."); 138 STATISTIC(NumOfPGOMissing, "Number of functions without profile."); 139 STATISTIC(NumOfPGOICall, "Number of indirect call value instrumentations."); 140 STATISTIC(NumOfCSPGOInstrument, "Number of edges instrumented in CSPGO."); 141 STATISTIC(NumOfCSPGOSelectInsts, 142 "Number of select instruction instrumented in CSPGO."); 143 STATISTIC(NumOfCSPGOMemIntrinsics, 144 "Number of mem intrinsics instrumented in CSPGO."); 145 STATISTIC(NumOfCSPGOEdge, "Number of edges in CSPGO."); 146 STATISTIC(NumOfCSPGOBB, "Number of basic-blocks in CSPGO."); 147 STATISTIC(NumOfCSPGOSplit, "Number of critical edge splits in CSPGO."); 148 STATISTIC(NumOfCSPGOFunc, 149 "Number of functions having valid profile counts in CSPGO."); 150 STATISTIC(NumOfCSPGOMismatch, 151 "Number of functions having mismatch profile in CSPGO."); 152 STATISTIC(NumOfCSPGOMissing, "Number of functions without profile in CSPGO."); 153 154 // Command line option to specify the file to read profile from. This is 155 // mainly used for testing. 156 static cl::opt<std::string> 157 PGOTestProfileFile("pgo-test-profile-file", cl::init(""), cl::Hidden, 158 cl::value_desc("filename"), 159 cl::desc("Specify the path of profile data file. This is" 160 "mainly for test purpose.")); 161 static cl::opt<std::string> PGOTestProfileRemappingFile( 162 "pgo-test-profile-remapping-file", cl::init(""), cl::Hidden, 163 cl::value_desc("filename"), 164 cl::desc("Specify the path of profile remapping file. This is mainly for " 165 "test purpose.")); 166 167 // Command line option to disable value profiling. The default is false: 168 // i.e. value profiling is enabled by default. This is for debug purpose. 169 static cl::opt<bool> DisableValueProfiling("disable-vp", cl::init(false), 170 cl::Hidden, 171 cl::desc("Disable Value Profiling")); 172 173 // Command line option to set the maximum number of VP annotations to write to 174 // the metadata for a single indirect call callsite. 175 static cl::opt<unsigned> MaxNumAnnotations( 176 "icp-max-annotations", cl::init(3), cl::Hidden, cl::ZeroOrMore, 177 cl::desc("Max number of annotations for a single indirect " 178 "call callsite")); 179 180 // Command line option to set the maximum number of value annotations 181 // to write to the metadata for a single memop intrinsic. 182 static cl::opt<unsigned> MaxNumMemOPAnnotations( 183 "memop-max-annotations", cl::init(4), cl::Hidden, cl::ZeroOrMore, 184 cl::desc("Max number of preicise value annotations for a single memop" 185 "intrinsic")); 186 187 // Command line option to control appending FunctionHash to the name of a COMDAT 188 // function. This is to avoid the hash mismatch caused by the preinliner. 189 static cl::opt<bool> DoComdatRenaming( 190 "do-comdat-renaming", cl::init(false), cl::Hidden, 191 cl::desc("Append function hash to the name of COMDAT function to avoid " 192 "function hash mismatch due to the preinliner")); 193 194 // Command line option to enable/disable the warning about missing profile 195 // information. 196 static cl::opt<bool> 197 PGOWarnMissing("pgo-warn-missing-function", cl::init(false), cl::Hidden, 198 cl::desc("Use this option to turn on/off " 199 "warnings about missing profile data for " 200 "functions.")); 201 202 // Command line option to enable/disable the warning about a hash mismatch in 203 // the profile data. 204 static cl::opt<bool> 205 NoPGOWarnMismatch("no-pgo-warn-mismatch", cl::init(false), cl::Hidden, 206 cl::desc("Use this option to turn off/on " 207 "warnings about profile cfg mismatch.")); 208 209 // Command line option to enable/disable the warning about a hash mismatch in 210 // the profile data for Comdat functions, which often turns out to be false 211 // positive due to the pre-instrumentation inline. 212 static cl::opt<bool> 213 NoPGOWarnMismatchComdat("no-pgo-warn-mismatch-comdat", cl::init(true), 214 cl::Hidden, 215 cl::desc("The option is used to turn on/off " 216 "warnings about hash mismatch for comdat " 217 "functions.")); 218 219 // Command line option to enable/disable select instruction instrumentation. 220 static cl::opt<bool> 221 PGOInstrSelect("pgo-instr-select", cl::init(true), cl::Hidden, 222 cl::desc("Use this option to turn on/off SELECT " 223 "instruction instrumentation. ")); 224 225 // Command line option to turn on CFG dot or text dump of raw profile counts 226 static cl::opt<PGOViewCountsType> PGOViewRawCounts( 227 "pgo-view-raw-counts", cl::Hidden, 228 cl::desc("A boolean option to show CFG dag or text " 229 "with raw profile counts from " 230 "profile data. See also option " 231 "-pgo-view-counts. To limit graph " 232 "display to only one function, use " 233 "filtering option -view-bfi-func-name."), 234 cl::values(clEnumValN(PGOVCT_None, "none", "do not show."), 235 clEnumValN(PGOVCT_Graph, "graph", "show a graph."), 236 clEnumValN(PGOVCT_Text, "text", "show in text."))); 237 238 // Command line option to enable/disable memop intrinsic call.size profiling. 239 static cl::opt<bool> 240 PGOInstrMemOP("pgo-instr-memop", cl::init(true), cl::Hidden, 241 cl::desc("Use this option to turn on/off " 242 "memory intrinsic size profiling.")); 243 244 // Emit branch probability as optimization remarks. 245 static cl::opt<bool> 246 EmitBranchProbability("pgo-emit-branch-prob", cl::init(false), cl::Hidden, 247 cl::desc("When this option is on, the annotated " 248 "branch probability will be emitted as " 249 "optimization remarks: -{Rpass|" 250 "pass-remarks}=pgo-instrumentation")); 251 252 // Command line option to turn on CFG dot dump after profile annotation. 253 // Defined in Analysis/BlockFrequencyInfo.cpp: -pgo-view-counts 254 extern cl::opt<PGOViewCountsType> PGOViewCounts; 255 256 // Command line option to specify the name of the function for CFG dump 257 // Defined in Analysis/BlockFrequencyInfo.cpp: -view-bfi-func-name= 258 extern cl::opt<std::string> ViewBlockFreqFuncName; 259 260 // Return a string describing the branch condition that can be 261 // used in static branch probability heuristics: 262 static std::string getBranchCondString(Instruction *TI) { 263 BranchInst *BI = dyn_cast<BranchInst>(TI); 264 if (!BI || !BI->isConditional()) 265 return std::string(); 266 267 Value *Cond = BI->getCondition(); 268 ICmpInst *CI = dyn_cast<ICmpInst>(Cond); 269 if (!CI) 270 return std::string(); 271 272 std::string result; 273 raw_string_ostream OS(result); 274 OS << CmpInst::getPredicateName(CI->getPredicate()) << "_"; 275 CI->getOperand(0)->getType()->print(OS, true); 276 277 Value *RHS = CI->getOperand(1); 278 ConstantInt *CV = dyn_cast<ConstantInt>(RHS); 279 if (CV) { 280 if (CV->isZero()) 281 OS << "_Zero"; 282 else if (CV->isOne()) 283 OS << "_One"; 284 else if (CV->isMinusOne()) 285 OS << "_MinusOne"; 286 else 287 OS << "_Const"; 288 } 289 OS.flush(); 290 return result; 291 } 292 293 static const char *ValueProfKindDescr[] = { 294 #define VALUE_PROF_KIND(Enumerator, Value, Descr) Descr, 295 #include "llvm/ProfileData/InstrProfData.inc" 296 }; 297 298 namespace { 299 300 /// The select instruction visitor plays three roles specified 301 /// by the mode. In \c VM_counting mode, it simply counts the number of 302 /// select instructions. In \c VM_instrument mode, it inserts code to count 303 /// the number times TrueValue of select is taken. In \c VM_annotate mode, 304 /// it reads the profile data and annotate the select instruction with metadata. 305 enum VisitMode { VM_counting, VM_instrument, VM_annotate }; 306 class PGOUseFunc; 307 308 /// Instruction Visitor class to visit select instructions. 309 struct SelectInstVisitor : public InstVisitor<SelectInstVisitor> { 310 Function &F; 311 unsigned NSIs = 0; // Number of select instructions instrumented. 312 VisitMode Mode = VM_counting; // Visiting mode. 313 unsigned *CurCtrIdx = nullptr; // Pointer to current counter index. 314 unsigned TotalNumCtrs = 0; // Total number of counters 315 GlobalVariable *FuncNameVar = nullptr; 316 uint64_t FuncHash = 0; 317 PGOUseFunc *UseFunc = nullptr; 318 319 SelectInstVisitor(Function &Func) : F(Func) {} 320 321 void countSelects(Function &Func) { 322 NSIs = 0; 323 Mode = VM_counting; 324 visit(Func); 325 } 326 327 // Visit the IR stream and instrument all select instructions. \p 328 // Ind is a pointer to the counter index variable; \p TotalNC 329 // is the total number of counters; \p FNV is the pointer to the 330 // PGO function name var; \p FHash is the function hash. 331 void instrumentSelects(Function &Func, unsigned *Ind, unsigned TotalNC, 332 GlobalVariable *FNV, uint64_t FHash) { 333 Mode = VM_instrument; 334 CurCtrIdx = Ind; 335 TotalNumCtrs = TotalNC; 336 FuncHash = FHash; 337 FuncNameVar = FNV; 338 visit(Func); 339 } 340 341 // Visit the IR stream and annotate all select instructions. 342 void annotateSelects(Function &Func, PGOUseFunc *UF, unsigned *Ind) { 343 Mode = VM_annotate; 344 UseFunc = UF; 345 CurCtrIdx = Ind; 346 visit(Func); 347 } 348 349 void instrumentOneSelectInst(SelectInst &SI); 350 void annotateOneSelectInst(SelectInst &SI); 351 352 // Visit \p SI instruction and perform tasks according to visit mode. 353 void visitSelectInst(SelectInst &SI); 354 355 // Return the number of select instructions. This needs be called after 356 // countSelects(). 357 unsigned getNumOfSelectInsts() const { return NSIs; } 358 }; 359 360 361 class PGOInstrumentationGenLegacyPass : public ModulePass { 362 public: 363 static char ID; 364 365 PGOInstrumentationGenLegacyPass(bool IsCS = false) 366 : ModulePass(ID), IsCS(IsCS) { 367 initializePGOInstrumentationGenLegacyPassPass( 368 *PassRegistry::getPassRegistry()); 369 } 370 371 StringRef getPassName() const override { return "PGOInstrumentationGenPass"; } 372 373 private: 374 // Is this is context-sensitive instrumentation. 375 bool IsCS; 376 bool runOnModule(Module &M) override; 377 378 void getAnalysisUsage(AnalysisUsage &AU) const override { 379 AU.addRequired<BlockFrequencyInfoWrapperPass>(); 380 } 381 }; 382 383 class PGOInstrumentationUseLegacyPass : public ModulePass { 384 public: 385 static char ID; 386 387 // Provide the profile filename as the parameter. 388 PGOInstrumentationUseLegacyPass(std::string Filename = "", bool IsCS = false) 389 : ModulePass(ID), ProfileFileName(std::move(Filename)), IsCS(IsCS) { 390 if (!PGOTestProfileFile.empty()) 391 ProfileFileName = PGOTestProfileFile; 392 initializePGOInstrumentationUseLegacyPassPass( 393 *PassRegistry::getPassRegistry()); 394 } 395 396 StringRef getPassName() const override { return "PGOInstrumentationUsePass"; } 397 398 private: 399 std::string ProfileFileName; 400 // Is this is context-sensitive instrumentation use. 401 bool IsCS; 402 403 bool runOnModule(Module &M) override; 404 405 void getAnalysisUsage(AnalysisUsage &AU) const override { 406 AU.addRequired<ProfileSummaryInfoWrapperPass>(); 407 AU.addRequired<BlockFrequencyInfoWrapperPass>(); 408 } 409 }; 410 411 class PGOInstrumentationGenCreateVarLegacyPass : public ModulePass { 412 public: 413 static char ID; 414 StringRef getPassName() const override { 415 return "PGOInstrumentationGenCreateVarPass"; 416 } 417 PGOInstrumentationGenCreateVarLegacyPass(std::string CSInstrName = "") 418 : ModulePass(ID), InstrProfileOutput(CSInstrName) { 419 initializePGOInstrumentationGenCreateVarLegacyPassPass( 420 *PassRegistry::getPassRegistry()); 421 } 422 423 private: 424 bool runOnModule(Module &M) override { 425 createProfileFileNameVar(M, InstrProfileOutput); 426 createIRLevelProfileFlagVar(M, true); 427 return false; 428 } 429 std::string InstrProfileOutput; 430 }; 431 432 } // end anonymous namespace 433 434 char PGOInstrumentationGenLegacyPass::ID = 0; 435 436 INITIALIZE_PASS_BEGIN(PGOInstrumentationGenLegacyPass, "pgo-instr-gen", 437 "PGO instrumentation.", false, false) 438 INITIALIZE_PASS_DEPENDENCY(BlockFrequencyInfoWrapperPass) 439 INITIALIZE_PASS_DEPENDENCY(BranchProbabilityInfoWrapperPass) 440 INITIALIZE_PASS_END(PGOInstrumentationGenLegacyPass, "pgo-instr-gen", 441 "PGO instrumentation.", false, false) 442 443 ModulePass *llvm::createPGOInstrumentationGenLegacyPass(bool IsCS) { 444 return new PGOInstrumentationGenLegacyPass(IsCS); 445 } 446 447 char PGOInstrumentationUseLegacyPass::ID = 0; 448 449 INITIALIZE_PASS_BEGIN(PGOInstrumentationUseLegacyPass, "pgo-instr-use", 450 "Read PGO instrumentation profile.", false, false) 451 INITIALIZE_PASS_DEPENDENCY(BlockFrequencyInfoWrapperPass) 452 INITIALIZE_PASS_DEPENDENCY(BranchProbabilityInfoWrapperPass) 453 INITIALIZE_PASS_DEPENDENCY(ProfileSummaryInfoWrapperPass) 454 INITIALIZE_PASS_END(PGOInstrumentationUseLegacyPass, "pgo-instr-use", 455 "Read PGO instrumentation profile.", false, false) 456 457 ModulePass *llvm::createPGOInstrumentationUseLegacyPass(StringRef Filename, 458 bool IsCS) { 459 return new PGOInstrumentationUseLegacyPass(Filename.str(), IsCS); 460 } 461 462 char PGOInstrumentationGenCreateVarLegacyPass::ID = 0; 463 464 INITIALIZE_PASS(PGOInstrumentationGenCreateVarLegacyPass, 465 "pgo-instr-gen-create-var", 466 "Create PGO instrumentation version variable for CSPGO.", false, 467 false) 468 469 ModulePass * 470 llvm::createPGOInstrumentationGenCreateVarLegacyPass(StringRef CSInstrName) { 471 return new PGOInstrumentationGenCreateVarLegacyPass(std::string(CSInstrName)); 472 } 473 474 namespace { 475 476 /// An MST based instrumentation for PGO 477 /// 478 /// Implements a Minimum Spanning Tree (MST) based instrumentation for PGO 479 /// in the function level. 480 struct PGOEdge { 481 // This class implements the CFG edges. Note the CFG can be a multi-graph. 482 // So there might be multiple edges with same SrcBB and DestBB. 483 const BasicBlock *SrcBB; 484 const BasicBlock *DestBB; 485 uint64_t Weight; 486 bool InMST = false; 487 bool Removed = false; 488 bool IsCritical = false; 489 490 PGOEdge(const BasicBlock *Src, const BasicBlock *Dest, uint64_t W = 1) 491 : SrcBB(Src), DestBB(Dest), Weight(W) {} 492 493 // Return the information string of an edge. 494 const std::string infoString() const { 495 return (Twine(Removed ? "-" : " ") + (InMST ? " " : "*") + 496 (IsCritical ? "c" : " ") + " W=" + Twine(Weight)).str(); 497 } 498 }; 499 500 // This class stores the auxiliary information for each BB. 501 struct BBInfo { 502 BBInfo *Group; 503 uint32_t Index; 504 uint32_t Rank = 0; 505 506 BBInfo(unsigned IX) : Group(this), Index(IX) {} 507 508 // Return the information string of this object. 509 const std::string infoString() const { 510 return (Twine("Index=") + Twine(Index)).str(); 511 } 512 513 // Empty function -- only applicable to UseBBInfo. 514 void addOutEdge(PGOEdge *E LLVM_ATTRIBUTE_UNUSED) {} 515 516 // Empty function -- only applicable to UseBBInfo. 517 void addInEdge(PGOEdge *E LLVM_ATTRIBUTE_UNUSED) {} 518 }; 519 520 // This class implements the CFG edges. Note the CFG can be a multi-graph. 521 template <class Edge, class BBInfo> class FuncPGOInstrumentation { 522 private: 523 Function &F; 524 525 // Is this is context-sensitive instrumentation. 526 bool IsCS; 527 528 // A map that stores the Comdat group in function F. 529 std::unordered_multimap<Comdat *, GlobalValue *> &ComdatMembers; 530 531 ValueProfileCollector VPC; 532 533 void computeCFGHash(); 534 void renameComdatFunction(); 535 536 public: 537 std::vector<std::vector<VPCandidateInfo>> ValueSites; 538 SelectInstVisitor SIVisitor; 539 std::string FuncName; 540 GlobalVariable *FuncNameVar; 541 542 // CFG hash value for this function. 543 uint64_t FunctionHash = 0; 544 545 // The Minimum Spanning Tree of function CFG. 546 CFGMST<Edge, BBInfo> MST; 547 548 // Collect all the BBs that will be instrumented, and store them in 549 // InstrumentBBs. 550 void getInstrumentBBs(std::vector<BasicBlock *> &InstrumentBBs); 551 552 // Give an edge, find the BB that will be instrumented. 553 // Return nullptr if there is no BB to be instrumented. 554 BasicBlock *getInstrBB(Edge *E); 555 556 // Return the auxiliary BB information. 557 BBInfo &getBBInfo(const BasicBlock *BB) const { return MST.getBBInfo(BB); } 558 559 // Return the auxiliary BB information if available. 560 BBInfo *findBBInfo(const BasicBlock *BB) const { return MST.findBBInfo(BB); } 561 562 // Dump edges and BB information. 563 void dumpInfo(std::string Str = "") const { 564 MST.dumpEdges(dbgs(), Twine("Dump Function ") + FuncName + " Hash: " + 565 Twine(FunctionHash) + "\t" + Str); 566 } 567 568 FuncPGOInstrumentation( 569 Function &Func, 570 std::unordered_multimap<Comdat *, GlobalValue *> &ComdatMembers, 571 bool CreateGlobalVar = false, BranchProbabilityInfo *BPI = nullptr, 572 BlockFrequencyInfo *BFI = nullptr, bool IsCS = false) 573 : F(Func), IsCS(IsCS), ComdatMembers(ComdatMembers), VPC(Func), 574 ValueSites(IPVK_Last + 1), SIVisitor(Func), MST(F, BPI, BFI) { 575 // This should be done before CFG hash computation. 576 SIVisitor.countSelects(Func); 577 ValueSites[IPVK_MemOPSize] = VPC.get(IPVK_MemOPSize); 578 if (!IsCS) { 579 NumOfPGOSelectInsts += SIVisitor.getNumOfSelectInsts(); 580 NumOfPGOMemIntrinsics += ValueSites[IPVK_MemOPSize].size(); 581 NumOfPGOBB += MST.BBInfos.size(); 582 ValueSites[IPVK_IndirectCallTarget] = VPC.get(IPVK_IndirectCallTarget); 583 } else { 584 NumOfCSPGOSelectInsts += SIVisitor.getNumOfSelectInsts(); 585 NumOfCSPGOMemIntrinsics += ValueSites[IPVK_MemOPSize].size(); 586 NumOfCSPGOBB += MST.BBInfos.size(); 587 } 588 589 FuncName = getPGOFuncName(F); 590 computeCFGHash(); 591 if (!ComdatMembers.empty()) 592 renameComdatFunction(); 593 LLVM_DEBUG(dumpInfo("after CFGMST")); 594 595 for (auto &E : MST.AllEdges) { 596 if (E->Removed) 597 continue; 598 IsCS ? NumOfCSPGOEdge++ : NumOfPGOEdge++; 599 if (!E->InMST) 600 IsCS ? NumOfCSPGOInstrument++ : NumOfPGOInstrument++; 601 } 602 603 if (CreateGlobalVar) 604 FuncNameVar = createPGOFuncNameVar(F, FuncName); 605 } 606 }; 607 608 } // end anonymous namespace 609 610 // Compute Hash value for the CFG: the lower 32 bits are CRC32 of the index 611 // value of each BB in the CFG. The higher 32 bits record the number of edges. 612 template <class Edge, class BBInfo> 613 void FuncPGOInstrumentation<Edge, BBInfo>::computeCFGHash() { 614 std::vector<uint8_t> Indexes; 615 JamCRC JC; 616 for (auto &BB : F) { 617 const Instruction *TI = BB.getTerminator(); 618 for (unsigned I = 0, E = TI->getNumSuccessors(); I != E; ++I) { 619 BasicBlock *Succ = TI->getSuccessor(I); 620 auto BI = findBBInfo(Succ); 621 if (BI == nullptr) 622 continue; 623 uint32_t Index = BI->Index; 624 for (int J = 0; J < 4; J++) 625 Indexes.push_back((uint8_t)(Index >> (J * 8))); 626 } 627 } 628 JC.update(Indexes); 629 630 // Hash format for context sensitive profile. Reserve 4 bits for other 631 // information. 632 FunctionHash = (uint64_t)SIVisitor.getNumOfSelectInsts() << 56 | 633 (uint64_t)ValueSites[IPVK_IndirectCallTarget].size() << 48 | 634 //(uint64_t)ValueSites[IPVK_MemOPSize].size() << 40 | 635 (uint64_t)MST.AllEdges.size() << 32 | JC.getCRC(); 636 // Reserve bit 60-63 for other information purpose. 637 FunctionHash &= 0x0FFFFFFFFFFFFFFF; 638 if (IsCS) 639 NamedInstrProfRecord::setCSFlagInHash(FunctionHash); 640 LLVM_DEBUG(dbgs() << "Function Hash Computation for " << F.getName() << ":\n" 641 << " CRC = " << JC.getCRC() 642 << ", Selects = " << SIVisitor.getNumOfSelectInsts() 643 << ", Edges = " << MST.AllEdges.size() << ", ICSites = " 644 << ValueSites[IPVK_IndirectCallTarget].size() 645 << ", Hash = " << FunctionHash << "\n";); 646 } 647 648 // Check if we can safely rename this Comdat function. 649 static bool canRenameComdat( 650 Function &F, 651 std::unordered_multimap<Comdat *, GlobalValue *> &ComdatMembers) { 652 if (!DoComdatRenaming || !canRenameComdatFunc(F, true)) 653 return false; 654 655 // FIXME: Current only handle those Comdat groups that only containing one 656 // function and function aliases. 657 // (1) For a Comdat group containing multiple functions, we need to have a 658 // unique postfix based on the hashes for each function. There is a 659 // non-trivial code refactoring to do this efficiently. 660 // (2) Variables can not be renamed, so we can not rename Comdat function in a 661 // group including global vars. 662 Comdat *C = F.getComdat(); 663 for (auto &&CM : make_range(ComdatMembers.equal_range(C))) { 664 if (dyn_cast<GlobalAlias>(CM.second)) 665 continue; 666 Function *FM = dyn_cast<Function>(CM.second); 667 if (FM != &F) 668 return false; 669 } 670 return true; 671 } 672 673 // Append the CFGHash to the Comdat function name. 674 template <class Edge, class BBInfo> 675 void FuncPGOInstrumentation<Edge, BBInfo>::renameComdatFunction() { 676 if (!canRenameComdat(F, ComdatMembers)) 677 return; 678 std::string OrigName = F.getName().str(); 679 std::string NewFuncName = 680 Twine(F.getName() + "." + Twine(FunctionHash)).str(); 681 F.setName(Twine(NewFuncName)); 682 GlobalAlias::create(GlobalValue::WeakAnyLinkage, OrigName, &F); 683 FuncName = Twine(FuncName + "." + Twine(FunctionHash)).str(); 684 Comdat *NewComdat; 685 Module *M = F.getParent(); 686 // For AvailableExternallyLinkage functions, change the linkage to 687 // LinkOnceODR and put them into comdat. This is because after renaming, there 688 // is no backup external copy available for the function. 689 if (!F.hasComdat()) { 690 assert(F.getLinkage() == GlobalValue::AvailableExternallyLinkage); 691 NewComdat = M->getOrInsertComdat(StringRef(NewFuncName)); 692 F.setLinkage(GlobalValue::LinkOnceODRLinkage); 693 F.setComdat(NewComdat); 694 return; 695 } 696 697 // This function belongs to a single function Comdat group. 698 Comdat *OrigComdat = F.getComdat(); 699 std::string NewComdatName = 700 Twine(OrigComdat->getName() + "." + Twine(FunctionHash)).str(); 701 NewComdat = M->getOrInsertComdat(StringRef(NewComdatName)); 702 NewComdat->setSelectionKind(OrigComdat->getSelectionKind()); 703 704 for (auto &&CM : make_range(ComdatMembers.equal_range(OrigComdat))) { 705 if (GlobalAlias *GA = dyn_cast<GlobalAlias>(CM.second)) { 706 // For aliases, change the name directly. 707 assert(dyn_cast<Function>(GA->getAliasee()->stripPointerCasts()) == &F); 708 std::string OrigGAName = GA->getName().str(); 709 GA->setName(Twine(GA->getName() + "." + Twine(FunctionHash))); 710 GlobalAlias::create(GlobalValue::WeakAnyLinkage, OrigGAName, GA); 711 continue; 712 } 713 // Must be a function. 714 Function *CF = dyn_cast<Function>(CM.second); 715 assert(CF); 716 CF->setComdat(NewComdat); 717 } 718 } 719 720 // Collect all the BBs that will be instruments and return them in 721 // InstrumentBBs and setup InEdges/OutEdge for UseBBInfo. 722 template <class Edge, class BBInfo> 723 void FuncPGOInstrumentation<Edge, BBInfo>::getInstrumentBBs( 724 std::vector<BasicBlock *> &InstrumentBBs) { 725 // Use a worklist as we will update the vector during the iteration. 726 std::vector<Edge *> EdgeList; 727 EdgeList.reserve(MST.AllEdges.size()); 728 for (auto &E : MST.AllEdges) 729 EdgeList.push_back(E.get()); 730 731 for (auto &E : EdgeList) { 732 BasicBlock *InstrBB = getInstrBB(E); 733 if (InstrBB) 734 InstrumentBBs.push_back(InstrBB); 735 } 736 737 // Set up InEdges/OutEdges for all BBs. 738 for (auto &E : MST.AllEdges) { 739 if (E->Removed) 740 continue; 741 const BasicBlock *SrcBB = E->SrcBB; 742 const BasicBlock *DestBB = E->DestBB; 743 BBInfo &SrcInfo = getBBInfo(SrcBB); 744 BBInfo &DestInfo = getBBInfo(DestBB); 745 SrcInfo.addOutEdge(E.get()); 746 DestInfo.addInEdge(E.get()); 747 } 748 } 749 750 // Given a CFG E to be instrumented, find which BB to place the instrumented 751 // code. The function will split the critical edge if necessary. 752 template <class Edge, class BBInfo> 753 BasicBlock *FuncPGOInstrumentation<Edge, BBInfo>::getInstrBB(Edge *E) { 754 if (E->InMST || E->Removed) 755 return nullptr; 756 757 BasicBlock *SrcBB = const_cast<BasicBlock *>(E->SrcBB); 758 BasicBlock *DestBB = const_cast<BasicBlock *>(E->DestBB); 759 // For a fake edge, instrument the real BB. 760 if (SrcBB == nullptr) 761 return DestBB; 762 if (DestBB == nullptr) 763 return SrcBB; 764 765 auto canInstrument = [](BasicBlock *BB) -> BasicBlock * { 766 // There are basic blocks (such as catchswitch) cannot be instrumented. 767 // If the returned first insertion point is the end of BB, skip this BB. 768 if (BB->getFirstInsertionPt() == BB->end()) 769 return nullptr; 770 return BB; 771 }; 772 773 // Instrument the SrcBB if it has a single successor, 774 // otherwise, the DestBB if this is not a critical edge. 775 Instruction *TI = SrcBB->getTerminator(); 776 if (TI->getNumSuccessors() <= 1) 777 return canInstrument(SrcBB); 778 if (!E->IsCritical) 779 return canInstrument(DestBB); 780 781 unsigned SuccNum = GetSuccessorNumber(SrcBB, DestBB); 782 BasicBlock *InstrBB = SplitCriticalEdge(TI, SuccNum); 783 if (!InstrBB) { 784 LLVM_DEBUG( 785 dbgs() << "Fail to split critical edge: not instrument this edge.\n"); 786 return nullptr; 787 } 788 // For a critical edge, we have to split. Instrument the newly 789 // created BB. 790 IsCS ? NumOfCSPGOSplit++ : NumOfPGOSplit++; 791 LLVM_DEBUG(dbgs() << "Split critical edge: " << getBBInfo(SrcBB).Index 792 << " --> " << getBBInfo(DestBB).Index << "\n"); 793 // Need to add two new edges. First one: Add new edge of SrcBB->InstrBB. 794 MST.addEdge(SrcBB, InstrBB, 0); 795 // Second one: Add new edge of InstrBB->DestBB. 796 Edge &NewEdge1 = MST.addEdge(InstrBB, DestBB, 0); 797 NewEdge1.InMST = true; 798 E->Removed = true; 799 800 return canInstrument(InstrBB); 801 } 802 803 // When generating value profiling calls on Windows routines that make use of 804 // handler funclets for exception processing an operand bundle needs to attached 805 // to the called function. This routine will set \p OpBundles to contain the 806 // funclet information, if any is needed, that should be placed on the generated 807 // value profiling call for the value profile candidate call. 808 static void 809 populateEHOperandBundle(VPCandidateInfo &Cand, 810 DenseMap<BasicBlock *, ColorVector> &BlockColors, 811 SmallVectorImpl<OperandBundleDef> &OpBundles) { 812 auto *OrigCall = dyn_cast<CallBase>(Cand.AnnotatedInst); 813 if (OrigCall && !isa<IntrinsicInst>(OrigCall)) { 814 // The instrumentation call should belong to the same funclet as a 815 // non-intrinsic call, so just copy the operand bundle, if any exists. 816 Optional<OperandBundleUse> ParentFunclet = 817 OrigCall->getOperandBundle(LLVMContext::OB_funclet); 818 if (ParentFunclet) 819 OpBundles.emplace_back(OperandBundleDef(*ParentFunclet)); 820 } else { 821 // Intrinsics or other instructions do not get funclet information from the 822 // front-end. Need to use the BlockColors that was computed by the routine 823 // colorEHFunclets to determine whether a funclet is needed. 824 if (!BlockColors.empty()) { 825 const ColorVector &CV = BlockColors.find(OrigCall->getParent())->second; 826 assert(CV.size() == 1 && "non-unique color for block!"); 827 Instruction *EHPad = CV.front()->getFirstNonPHI(); 828 if (EHPad->isEHPad()) 829 OpBundles.emplace_back("funclet", EHPad); 830 } 831 } 832 } 833 834 // Visit all edge and instrument the edges not in MST, and do value profiling. 835 // Critical edges will be split. 836 static void instrumentOneFunc( 837 Function &F, Module *M, BranchProbabilityInfo *BPI, BlockFrequencyInfo *BFI, 838 std::unordered_multimap<Comdat *, GlobalValue *> &ComdatMembers, 839 bool IsCS) { 840 // Split indirectbr critical edges here before computing the MST rather than 841 // later in getInstrBB() to avoid invalidating it. 842 SplitIndirectBrCriticalEdges(F, BPI, BFI); 843 844 FuncPGOInstrumentation<PGOEdge, BBInfo> FuncInfo(F, ComdatMembers, true, BPI, 845 BFI, IsCS); 846 std::vector<BasicBlock *> InstrumentBBs; 847 FuncInfo.getInstrumentBBs(InstrumentBBs); 848 unsigned NumCounters = 849 InstrumentBBs.size() + FuncInfo.SIVisitor.getNumOfSelectInsts(); 850 851 uint32_t I = 0; 852 Type *I8PtrTy = Type::getInt8PtrTy(M->getContext()); 853 for (auto *InstrBB : InstrumentBBs) { 854 IRBuilder<> Builder(InstrBB, InstrBB->getFirstInsertionPt()); 855 assert(Builder.GetInsertPoint() != InstrBB->end() && 856 "Cannot get the Instrumentation point"); 857 Builder.CreateCall( 858 Intrinsic::getDeclaration(M, Intrinsic::instrprof_increment), 859 {ConstantExpr::getBitCast(FuncInfo.FuncNameVar, I8PtrTy), 860 Builder.getInt64(FuncInfo.FunctionHash), Builder.getInt32(NumCounters), 861 Builder.getInt32(I++)}); 862 } 863 864 // Now instrument select instructions: 865 FuncInfo.SIVisitor.instrumentSelects(F, &I, NumCounters, FuncInfo.FuncNameVar, 866 FuncInfo.FunctionHash); 867 assert(I == NumCounters); 868 869 if (DisableValueProfiling) 870 return; 871 872 NumOfPGOICall += FuncInfo.ValueSites[IPVK_IndirectCallTarget].size(); 873 874 // Intrinsic function calls do not have funclet operand bundles needed for 875 // Windows exception handling attached to them. However, if value profiling is 876 // inserted for one of these calls, then a funclet value will need to be set 877 // on the instrumentation call based on the funclet coloring. 878 DenseMap<BasicBlock *, ColorVector> BlockColors; 879 if (F.hasPersonalityFn() && 880 isFuncletEHPersonality(classifyEHPersonality(F.getPersonalityFn()))) 881 BlockColors = colorEHFunclets(F); 882 883 // For each VP Kind, walk the VP candidates and instrument each one. 884 for (uint32_t Kind = IPVK_First; Kind <= IPVK_Last; ++Kind) { 885 unsigned SiteIndex = 0; 886 if (Kind == IPVK_MemOPSize && !PGOInstrMemOP) 887 continue; 888 889 for (VPCandidateInfo Cand : FuncInfo.ValueSites[Kind]) { 890 LLVM_DEBUG(dbgs() << "Instrument one VP " << ValueProfKindDescr[Kind] 891 << " site: CallSite Index = " << SiteIndex << "\n"); 892 893 IRBuilder<> Builder(Cand.InsertPt); 894 assert(Builder.GetInsertPoint() != Cand.InsertPt->getParent()->end() && 895 "Cannot get the Instrumentation point"); 896 897 Value *ToProfile = nullptr; 898 if (Cand.V->getType()->isIntegerTy()) 899 ToProfile = Builder.CreateZExtOrTrunc(Cand.V, Builder.getInt64Ty()); 900 else if (Cand.V->getType()->isPointerTy()) 901 ToProfile = Builder.CreatePtrToInt(Cand.V, Builder.getInt64Ty()); 902 assert(ToProfile && "value profiling Value is of unexpected type"); 903 904 SmallVector<OperandBundleDef, 1> OpBundles; 905 populateEHOperandBundle(Cand, BlockColors, OpBundles); 906 Builder.CreateCall( 907 Intrinsic::getDeclaration(M, Intrinsic::instrprof_value_profile), 908 {ConstantExpr::getBitCast(FuncInfo.FuncNameVar, I8PtrTy), 909 Builder.getInt64(FuncInfo.FunctionHash), ToProfile, 910 Builder.getInt32(Kind), Builder.getInt32(SiteIndex++)}, 911 OpBundles); 912 } 913 } // IPVK_First <= Kind <= IPVK_Last 914 } 915 916 namespace { 917 918 // This class represents a CFG edge in profile use compilation. 919 struct PGOUseEdge : public PGOEdge { 920 bool CountValid = false; 921 uint64_t CountValue = 0; 922 923 PGOUseEdge(const BasicBlock *Src, const BasicBlock *Dest, uint64_t W = 1) 924 : PGOEdge(Src, Dest, W) {} 925 926 // Set edge count value 927 void setEdgeCount(uint64_t Value) { 928 CountValue = Value; 929 CountValid = true; 930 } 931 932 // Return the information string for this object. 933 const std::string infoString() const { 934 if (!CountValid) 935 return PGOEdge::infoString(); 936 return (Twine(PGOEdge::infoString()) + " Count=" + Twine(CountValue)) 937 .str(); 938 } 939 }; 940 941 using DirectEdges = SmallVector<PGOUseEdge *, 2>; 942 943 // This class stores the auxiliary information for each BB. 944 struct UseBBInfo : public BBInfo { 945 uint64_t CountValue = 0; 946 bool CountValid; 947 int32_t UnknownCountInEdge = 0; 948 int32_t UnknownCountOutEdge = 0; 949 DirectEdges InEdges; 950 DirectEdges OutEdges; 951 952 UseBBInfo(unsigned IX) : BBInfo(IX), CountValid(false) {} 953 954 UseBBInfo(unsigned IX, uint64_t C) 955 : BBInfo(IX), CountValue(C), CountValid(true) {} 956 957 // Set the profile count value for this BB. 958 void setBBInfoCount(uint64_t Value) { 959 CountValue = Value; 960 CountValid = true; 961 } 962 963 // Return the information string of this object. 964 const std::string infoString() const { 965 if (!CountValid) 966 return BBInfo::infoString(); 967 return (Twine(BBInfo::infoString()) + " Count=" + Twine(CountValue)).str(); 968 } 969 970 // Add an OutEdge and update the edge count. 971 void addOutEdge(PGOUseEdge *E) { 972 OutEdges.push_back(E); 973 UnknownCountOutEdge++; 974 } 975 976 // Add an InEdge and update the edge count. 977 void addInEdge(PGOUseEdge *E) { 978 InEdges.push_back(E); 979 UnknownCountInEdge++; 980 } 981 }; 982 983 } // end anonymous namespace 984 985 // Sum up the count values for all the edges. 986 static uint64_t sumEdgeCount(const ArrayRef<PGOUseEdge *> Edges) { 987 uint64_t Total = 0; 988 for (auto &E : Edges) { 989 if (E->Removed) 990 continue; 991 Total += E->CountValue; 992 } 993 return Total; 994 } 995 996 namespace { 997 998 class PGOUseFunc { 999 public: 1000 PGOUseFunc(Function &Func, Module *Modu, 1001 std::unordered_multimap<Comdat *, GlobalValue *> &ComdatMembers, 1002 BranchProbabilityInfo *BPI, BlockFrequencyInfo *BFIin, 1003 ProfileSummaryInfo *PSI, bool IsCS) 1004 : F(Func), M(Modu), BFI(BFIin), PSI(PSI), 1005 FuncInfo(Func, ComdatMembers, false, BPI, BFIin, IsCS), 1006 FreqAttr(FFA_Normal), IsCS(IsCS) {} 1007 1008 // Read counts for the instrumented BB from profile. 1009 bool readCounters(IndexedInstrProfReader *PGOReader, bool &AllZeros); 1010 1011 // Populate the counts for all BBs. 1012 void populateCounters(); 1013 1014 // Set the branch weights based on the count values. 1015 void setBranchWeights(); 1016 1017 // Annotate the value profile call sites for all value kind. 1018 void annotateValueSites(); 1019 1020 // Annotate the value profile call sites for one value kind. 1021 void annotateValueSites(uint32_t Kind); 1022 1023 // Annotate the irreducible loop header weights. 1024 void annotateIrrLoopHeaderWeights(); 1025 1026 // The hotness of the function from the profile count. 1027 enum FuncFreqAttr { FFA_Normal, FFA_Cold, FFA_Hot }; 1028 1029 // Return the function hotness from the profile. 1030 FuncFreqAttr getFuncFreqAttr() const { return FreqAttr; } 1031 1032 // Return the function hash. 1033 uint64_t getFuncHash() const { return FuncInfo.FunctionHash; } 1034 1035 // Return the profile record for this function; 1036 InstrProfRecord &getProfileRecord() { return ProfileRecord; } 1037 1038 // Return the auxiliary BB information. 1039 UseBBInfo &getBBInfo(const BasicBlock *BB) const { 1040 return FuncInfo.getBBInfo(BB); 1041 } 1042 1043 // Return the auxiliary BB information if available. 1044 UseBBInfo *findBBInfo(const BasicBlock *BB) const { 1045 return FuncInfo.findBBInfo(BB); 1046 } 1047 1048 Function &getFunc() const { return F; } 1049 1050 void dumpInfo(std::string Str = "") const { 1051 FuncInfo.dumpInfo(Str); 1052 } 1053 1054 uint64_t getProgramMaxCount() const { return ProgramMaxCount; } 1055 private: 1056 Function &F; 1057 Module *M; 1058 BlockFrequencyInfo *BFI; 1059 ProfileSummaryInfo *PSI; 1060 1061 // This member stores the shared information with class PGOGenFunc. 1062 FuncPGOInstrumentation<PGOUseEdge, UseBBInfo> FuncInfo; 1063 1064 // The maximum count value in the profile. This is only used in PGO use 1065 // compilation. 1066 uint64_t ProgramMaxCount; 1067 1068 // Position of counter that remains to be read. 1069 uint32_t CountPosition = 0; 1070 1071 // Total size of the profile count for this function. 1072 uint32_t ProfileCountSize = 0; 1073 1074 // ProfileRecord for this function. 1075 InstrProfRecord ProfileRecord; 1076 1077 // Function hotness info derived from profile. 1078 FuncFreqAttr FreqAttr; 1079 1080 // Is to use the context sensitive profile. 1081 bool IsCS; 1082 1083 // Find the Instrumented BB and set the value. Return false on error. 1084 bool setInstrumentedCounts(const std::vector<uint64_t> &CountFromProfile); 1085 1086 // Set the edge counter value for the unknown edge -- there should be only 1087 // one unknown edge. 1088 void setEdgeCount(DirectEdges &Edges, uint64_t Value); 1089 1090 // Return FuncName string; 1091 const std::string getFuncName() const { return FuncInfo.FuncName; } 1092 1093 // Set the hot/cold inline hints based on the count values. 1094 // FIXME: This function should be removed once the functionality in 1095 // the inliner is implemented. 1096 void markFunctionAttributes(uint64_t EntryCount, uint64_t MaxCount) { 1097 if (PSI->isHotCount(EntryCount)) 1098 FreqAttr = FFA_Hot; 1099 else if (PSI->isColdCount(MaxCount)) 1100 FreqAttr = FFA_Cold; 1101 } 1102 }; 1103 1104 } // end anonymous namespace 1105 1106 // Visit all the edges and assign the count value for the instrumented 1107 // edges and the BB. Return false on error. 1108 bool PGOUseFunc::setInstrumentedCounts( 1109 const std::vector<uint64_t> &CountFromProfile) { 1110 1111 std::vector<BasicBlock *> InstrumentBBs; 1112 FuncInfo.getInstrumentBBs(InstrumentBBs); 1113 unsigned NumCounters = 1114 InstrumentBBs.size() + FuncInfo.SIVisitor.getNumOfSelectInsts(); 1115 // The number of counters here should match the number of counters 1116 // in profile. Return if they mismatch. 1117 if (NumCounters != CountFromProfile.size()) { 1118 return false; 1119 } 1120 // Set the profile count to the Instrumented BBs. 1121 uint32_t I = 0; 1122 for (BasicBlock *InstrBB : InstrumentBBs) { 1123 uint64_t CountValue = CountFromProfile[I++]; 1124 UseBBInfo &Info = getBBInfo(InstrBB); 1125 Info.setBBInfoCount(CountValue); 1126 } 1127 ProfileCountSize = CountFromProfile.size(); 1128 CountPosition = I; 1129 1130 // Set the edge count and update the count of unknown edges for BBs. 1131 auto setEdgeCount = [this](PGOUseEdge *E, uint64_t Value) -> void { 1132 E->setEdgeCount(Value); 1133 this->getBBInfo(E->SrcBB).UnknownCountOutEdge--; 1134 this->getBBInfo(E->DestBB).UnknownCountInEdge--; 1135 }; 1136 1137 // Set the profile count the Instrumented edges. There are BBs that not in 1138 // MST but not instrumented. Need to set the edge count value so that we can 1139 // populate the profile counts later. 1140 for (auto &E : FuncInfo.MST.AllEdges) { 1141 if (E->Removed || E->InMST) 1142 continue; 1143 const BasicBlock *SrcBB = E->SrcBB; 1144 UseBBInfo &SrcInfo = getBBInfo(SrcBB); 1145 1146 // If only one out-edge, the edge profile count should be the same as BB 1147 // profile count. 1148 if (SrcInfo.CountValid && SrcInfo.OutEdges.size() == 1) 1149 setEdgeCount(E.get(), SrcInfo.CountValue); 1150 else { 1151 const BasicBlock *DestBB = E->DestBB; 1152 UseBBInfo &DestInfo = getBBInfo(DestBB); 1153 // If only one in-edge, the edge profile count should be the same as BB 1154 // profile count. 1155 if (DestInfo.CountValid && DestInfo.InEdges.size() == 1) 1156 setEdgeCount(E.get(), DestInfo.CountValue); 1157 } 1158 if (E->CountValid) 1159 continue; 1160 // E's count should have been set from profile. If not, this meenas E skips 1161 // the instrumentation. We set the count to 0. 1162 setEdgeCount(E.get(), 0); 1163 } 1164 return true; 1165 } 1166 1167 // Set the count value for the unknown edge. There should be one and only one 1168 // unknown edge in Edges vector. 1169 void PGOUseFunc::setEdgeCount(DirectEdges &Edges, uint64_t Value) { 1170 for (auto &E : Edges) { 1171 if (E->CountValid) 1172 continue; 1173 E->setEdgeCount(Value); 1174 1175 getBBInfo(E->SrcBB).UnknownCountOutEdge--; 1176 getBBInfo(E->DestBB).UnknownCountInEdge--; 1177 return; 1178 } 1179 llvm_unreachable("Cannot find the unknown count edge"); 1180 } 1181 1182 // Read the profile from ProfileFileName and assign the value to the 1183 // instrumented BB and the edges. This function also updates ProgramMaxCount. 1184 // Return true if the profile are successfully read, and false on errors. 1185 bool PGOUseFunc::readCounters(IndexedInstrProfReader *PGOReader, bool &AllZeros) { 1186 auto &Ctx = M->getContext(); 1187 Expected<InstrProfRecord> Result = 1188 PGOReader->getInstrProfRecord(FuncInfo.FuncName, FuncInfo.FunctionHash); 1189 if (Error E = Result.takeError()) { 1190 handleAllErrors(std::move(E), [&](const InstrProfError &IPE) { 1191 auto Err = IPE.get(); 1192 bool SkipWarning = false; 1193 LLVM_DEBUG(dbgs() << "Error in reading profile for Func " 1194 << FuncInfo.FuncName << ": "); 1195 if (Err == instrprof_error::unknown_function) { 1196 IsCS ? NumOfCSPGOMissing++ : NumOfPGOMissing++; 1197 SkipWarning = !PGOWarnMissing; 1198 LLVM_DEBUG(dbgs() << "unknown function"); 1199 } else if (Err == instrprof_error::hash_mismatch || 1200 Err == instrprof_error::malformed) { 1201 IsCS ? NumOfCSPGOMismatch++ : NumOfPGOMismatch++; 1202 SkipWarning = 1203 NoPGOWarnMismatch || 1204 (NoPGOWarnMismatchComdat && 1205 (F.hasComdat() || 1206 F.getLinkage() == GlobalValue::AvailableExternallyLinkage)); 1207 LLVM_DEBUG(dbgs() << "hash mismatch (skip=" << SkipWarning << ")"); 1208 } 1209 1210 LLVM_DEBUG(dbgs() << " IsCS=" << IsCS << "\n"); 1211 if (SkipWarning) 1212 return; 1213 1214 std::string Msg = IPE.message() + std::string(" ") + F.getName().str() + 1215 std::string(" Hash = ") + 1216 std::to_string(FuncInfo.FunctionHash); 1217 1218 Ctx.diagnose( 1219 DiagnosticInfoPGOProfile(M->getName().data(), Msg, DS_Warning)); 1220 }); 1221 return false; 1222 } 1223 ProfileRecord = std::move(Result.get()); 1224 std::vector<uint64_t> &CountFromProfile = ProfileRecord.Counts; 1225 1226 IsCS ? NumOfCSPGOFunc++ : NumOfPGOFunc++; 1227 LLVM_DEBUG(dbgs() << CountFromProfile.size() << " counts\n"); 1228 uint64_t ValueSum = 0; 1229 for (unsigned I = 0, S = CountFromProfile.size(); I < S; I++) { 1230 LLVM_DEBUG(dbgs() << " " << I << ": " << CountFromProfile[I] << "\n"); 1231 ValueSum += CountFromProfile[I]; 1232 } 1233 AllZeros = (ValueSum == 0); 1234 1235 LLVM_DEBUG(dbgs() << "SUM = " << ValueSum << "\n"); 1236 1237 getBBInfo(nullptr).UnknownCountOutEdge = 2; 1238 getBBInfo(nullptr).UnknownCountInEdge = 2; 1239 1240 if (!setInstrumentedCounts(CountFromProfile)) { 1241 LLVM_DEBUG( 1242 dbgs() << "Inconsistent number of counts, skipping this function"); 1243 Ctx.diagnose(DiagnosticInfoPGOProfile( 1244 M->getName().data(), 1245 Twine("Inconsistent number of counts in ") + F.getName().str() 1246 + Twine(": the profile may be stale or there is a function name collision."), 1247 DS_Warning)); 1248 return false; 1249 } 1250 ProgramMaxCount = PGOReader->getMaximumFunctionCount(IsCS); 1251 return true; 1252 } 1253 1254 // Populate the counters from instrumented BBs to all BBs. 1255 // In the end of this operation, all BBs should have a valid count value. 1256 void PGOUseFunc::populateCounters() { 1257 bool Changes = true; 1258 unsigned NumPasses = 0; 1259 while (Changes) { 1260 NumPasses++; 1261 Changes = false; 1262 1263 // For efficient traversal, it's better to start from the end as most 1264 // of the instrumented edges are at the end. 1265 for (auto &BB : reverse(F)) { 1266 UseBBInfo *Count = findBBInfo(&BB); 1267 if (Count == nullptr) 1268 continue; 1269 if (!Count->CountValid) { 1270 if (Count->UnknownCountOutEdge == 0) { 1271 Count->CountValue = sumEdgeCount(Count->OutEdges); 1272 Count->CountValid = true; 1273 Changes = true; 1274 } else if (Count->UnknownCountInEdge == 0) { 1275 Count->CountValue = sumEdgeCount(Count->InEdges); 1276 Count->CountValid = true; 1277 Changes = true; 1278 } 1279 } 1280 if (Count->CountValid) { 1281 if (Count->UnknownCountOutEdge == 1) { 1282 uint64_t Total = 0; 1283 uint64_t OutSum = sumEdgeCount(Count->OutEdges); 1284 // If the one of the successor block can early terminate (no-return), 1285 // we can end up with situation where out edge sum count is larger as 1286 // the source BB's count is collected by a post-dominated block. 1287 if (Count->CountValue > OutSum) 1288 Total = Count->CountValue - OutSum; 1289 setEdgeCount(Count->OutEdges, Total); 1290 Changes = true; 1291 } 1292 if (Count->UnknownCountInEdge == 1) { 1293 uint64_t Total = 0; 1294 uint64_t InSum = sumEdgeCount(Count->InEdges); 1295 if (Count->CountValue > InSum) 1296 Total = Count->CountValue - InSum; 1297 setEdgeCount(Count->InEdges, Total); 1298 Changes = true; 1299 } 1300 } 1301 } 1302 } 1303 1304 LLVM_DEBUG(dbgs() << "Populate counts in " << NumPasses << " passes.\n"); 1305 #ifndef NDEBUG 1306 // Assert every BB has a valid counter. 1307 for (auto &BB : F) { 1308 auto BI = findBBInfo(&BB); 1309 if (BI == nullptr) 1310 continue; 1311 assert(BI->CountValid && "BB count is not valid"); 1312 } 1313 #endif 1314 uint64_t FuncEntryCount = getBBInfo(&*F.begin()).CountValue; 1315 F.setEntryCount(ProfileCount(FuncEntryCount, Function::PCT_Real)); 1316 uint64_t FuncMaxCount = FuncEntryCount; 1317 for (auto &BB : F) { 1318 auto BI = findBBInfo(&BB); 1319 if (BI == nullptr) 1320 continue; 1321 FuncMaxCount = std::max(FuncMaxCount, BI->CountValue); 1322 } 1323 markFunctionAttributes(FuncEntryCount, FuncMaxCount); 1324 1325 // Now annotate select instructions 1326 FuncInfo.SIVisitor.annotateSelects(F, this, &CountPosition); 1327 assert(CountPosition == ProfileCountSize); 1328 1329 LLVM_DEBUG(FuncInfo.dumpInfo("after reading profile.")); 1330 } 1331 1332 // Assign the scaled count values to the BB with multiple out edges. 1333 void PGOUseFunc::setBranchWeights() { 1334 // Generate MD_prof metadata for every branch instruction. 1335 LLVM_DEBUG(dbgs() << "\nSetting branch weights for func " << F.getName() 1336 << " IsCS=" << IsCS << "\n"); 1337 for (auto &BB : F) { 1338 Instruction *TI = BB.getTerminator(); 1339 if (TI->getNumSuccessors() < 2) 1340 continue; 1341 if (!(isa<BranchInst>(TI) || isa<SwitchInst>(TI) || 1342 isa<IndirectBrInst>(TI))) 1343 continue; 1344 1345 if (getBBInfo(&BB).CountValue == 0) 1346 continue; 1347 1348 // We have a non-zero Branch BB. 1349 const UseBBInfo &BBCountInfo = getBBInfo(&BB); 1350 unsigned Size = BBCountInfo.OutEdges.size(); 1351 SmallVector<uint64_t, 2> EdgeCounts(Size, 0); 1352 uint64_t MaxCount = 0; 1353 for (unsigned s = 0; s < Size; s++) { 1354 const PGOUseEdge *E = BBCountInfo.OutEdges[s]; 1355 const BasicBlock *SrcBB = E->SrcBB; 1356 const BasicBlock *DestBB = E->DestBB; 1357 if (DestBB == nullptr) 1358 continue; 1359 unsigned SuccNum = GetSuccessorNumber(SrcBB, DestBB); 1360 uint64_t EdgeCount = E->CountValue; 1361 if (EdgeCount > MaxCount) 1362 MaxCount = EdgeCount; 1363 EdgeCounts[SuccNum] = EdgeCount; 1364 } 1365 setProfMetadata(M, TI, EdgeCounts, MaxCount); 1366 } 1367 } 1368 1369 static bool isIndirectBrTarget(BasicBlock *BB) { 1370 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) { 1371 if (isa<IndirectBrInst>((*PI)->getTerminator())) 1372 return true; 1373 } 1374 return false; 1375 } 1376 1377 void PGOUseFunc::annotateIrrLoopHeaderWeights() { 1378 LLVM_DEBUG(dbgs() << "\nAnnotating irreducible loop header weights.\n"); 1379 // Find irr loop headers 1380 for (auto &BB : F) { 1381 // As a heuristic also annotate indrectbr targets as they have a high chance 1382 // to become an irreducible loop header after the indirectbr tail 1383 // duplication. 1384 if (BFI->isIrrLoopHeader(&BB) || isIndirectBrTarget(&BB)) { 1385 Instruction *TI = BB.getTerminator(); 1386 const UseBBInfo &BBCountInfo = getBBInfo(&BB); 1387 setIrrLoopHeaderMetadata(M, TI, BBCountInfo.CountValue); 1388 } 1389 } 1390 } 1391 1392 void SelectInstVisitor::instrumentOneSelectInst(SelectInst &SI) { 1393 Module *M = F.getParent(); 1394 IRBuilder<> Builder(&SI); 1395 Type *Int64Ty = Builder.getInt64Ty(); 1396 Type *I8PtrTy = Builder.getInt8PtrTy(); 1397 auto *Step = Builder.CreateZExt(SI.getCondition(), Int64Ty); 1398 Builder.CreateCall( 1399 Intrinsic::getDeclaration(M, Intrinsic::instrprof_increment_step), 1400 {ConstantExpr::getBitCast(FuncNameVar, I8PtrTy), 1401 Builder.getInt64(FuncHash), Builder.getInt32(TotalNumCtrs), 1402 Builder.getInt32(*CurCtrIdx), Step}); 1403 ++(*CurCtrIdx); 1404 } 1405 1406 void SelectInstVisitor::annotateOneSelectInst(SelectInst &SI) { 1407 std::vector<uint64_t> &CountFromProfile = UseFunc->getProfileRecord().Counts; 1408 assert(*CurCtrIdx < CountFromProfile.size() && 1409 "Out of bound access of counters"); 1410 uint64_t SCounts[2]; 1411 SCounts[0] = CountFromProfile[*CurCtrIdx]; // True count 1412 ++(*CurCtrIdx); 1413 uint64_t TotalCount = 0; 1414 auto BI = UseFunc->findBBInfo(SI.getParent()); 1415 if (BI != nullptr) 1416 TotalCount = BI->CountValue; 1417 // False Count 1418 SCounts[1] = (TotalCount > SCounts[0] ? TotalCount - SCounts[0] : 0); 1419 uint64_t MaxCount = std::max(SCounts[0], SCounts[1]); 1420 if (MaxCount) 1421 setProfMetadata(F.getParent(), &SI, SCounts, MaxCount); 1422 } 1423 1424 void SelectInstVisitor::visitSelectInst(SelectInst &SI) { 1425 if (!PGOInstrSelect) 1426 return; 1427 // FIXME: do not handle this yet. 1428 if (SI.getCondition()->getType()->isVectorTy()) 1429 return; 1430 1431 switch (Mode) { 1432 case VM_counting: 1433 NSIs++; 1434 return; 1435 case VM_instrument: 1436 instrumentOneSelectInst(SI); 1437 return; 1438 case VM_annotate: 1439 annotateOneSelectInst(SI); 1440 return; 1441 } 1442 1443 llvm_unreachable("Unknown visiting mode"); 1444 } 1445 1446 // Traverse all valuesites and annotate the instructions for all value kind. 1447 void PGOUseFunc::annotateValueSites() { 1448 if (DisableValueProfiling) 1449 return; 1450 1451 // Create the PGOFuncName meta data. 1452 createPGOFuncNameMetadata(F, FuncInfo.FuncName); 1453 1454 for (uint32_t Kind = IPVK_First; Kind <= IPVK_Last; ++Kind) 1455 annotateValueSites(Kind); 1456 } 1457 1458 // Annotate the instructions for a specific value kind. 1459 void PGOUseFunc::annotateValueSites(uint32_t Kind) { 1460 assert(Kind <= IPVK_Last); 1461 unsigned ValueSiteIndex = 0; 1462 auto &ValueSites = FuncInfo.ValueSites[Kind]; 1463 unsigned NumValueSites = ProfileRecord.getNumValueSites(Kind); 1464 if (NumValueSites != ValueSites.size()) { 1465 auto &Ctx = M->getContext(); 1466 Ctx.diagnose(DiagnosticInfoPGOProfile( 1467 M->getName().data(), 1468 Twine("Inconsistent number of value sites for ") + 1469 Twine(ValueProfKindDescr[Kind]) + 1470 Twine(" profiling in \"") + F.getName().str() + 1471 Twine("\", possibly due to the use of a stale profile."), 1472 DS_Warning)); 1473 return; 1474 } 1475 1476 for (VPCandidateInfo &I : ValueSites) { 1477 LLVM_DEBUG(dbgs() << "Read one value site profile (kind = " << Kind 1478 << "): Index = " << ValueSiteIndex << " out of " 1479 << NumValueSites << "\n"); 1480 annotateValueSite(*M, *I.AnnotatedInst, ProfileRecord, 1481 static_cast<InstrProfValueKind>(Kind), ValueSiteIndex, 1482 Kind == IPVK_MemOPSize ? MaxNumMemOPAnnotations 1483 : MaxNumAnnotations); 1484 ValueSiteIndex++; 1485 } 1486 } 1487 1488 // Collect the set of members for each Comdat in module M and store 1489 // in ComdatMembers. 1490 static void collectComdatMembers( 1491 Module &M, 1492 std::unordered_multimap<Comdat *, GlobalValue *> &ComdatMembers) { 1493 if (!DoComdatRenaming) 1494 return; 1495 for (Function &F : M) 1496 if (Comdat *C = F.getComdat()) 1497 ComdatMembers.insert(std::make_pair(C, &F)); 1498 for (GlobalVariable &GV : M.globals()) 1499 if (Comdat *C = GV.getComdat()) 1500 ComdatMembers.insert(std::make_pair(C, &GV)); 1501 for (GlobalAlias &GA : M.aliases()) 1502 if (Comdat *C = GA.getComdat()) 1503 ComdatMembers.insert(std::make_pair(C, &GA)); 1504 } 1505 1506 static bool InstrumentAllFunctions( 1507 Module &M, function_ref<BranchProbabilityInfo *(Function &)> LookupBPI, 1508 function_ref<BlockFrequencyInfo *(Function &)> LookupBFI, bool IsCS) { 1509 // For the context-sensitve instrumentation, we should have a separated pass 1510 // (before LTO/ThinLTO linking) to create these variables. 1511 if (!IsCS) 1512 createIRLevelProfileFlagVar(M, /* IsCS */ false); 1513 std::unordered_multimap<Comdat *, GlobalValue *> ComdatMembers; 1514 collectComdatMembers(M, ComdatMembers); 1515 1516 for (auto &F : M) { 1517 if (F.isDeclaration()) 1518 continue; 1519 auto *BPI = LookupBPI(F); 1520 auto *BFI = LookupBFI(F); 1521 instrumentOneFunc(F, &M, BPI, BFI, ComdatMembers, IsCS); 1522 } 1523 return true; 1524 } 1525 1526 PreservedAnalyses 1527 PGOInstrumentationGenCreateVar::run(Module &M, ModuleAnalysisManager &AM) { 1528 createProfileFileNameVar(M, CSInstrName); 1529 createIRLevelProfileFlagVar(M, /* IsCS */ true); 1530 return PreservedAnalyses::all(); 1531 } 1532 1533 bool PGOInstrumentationGenLegacyPass::runOnModule(Module &M) { 1534 if (skipModule(M)) 1535 return false; 1536 1537 auto LookupBPI = [this](Function &F) { 1538 return &this->getAnalysis<BranchProbabilityInfoWrapperPass>(F).getBPI(); 1539 }; 1540 auto LookupBFI = [this](Function &F) { 1541 return &this->getAnalysis<BlockFrequencyInfoWrapperPass>(F).getBFI(); 1542 }; 1543 return InstrumentAllFunctions(M, LookupBPI, LookupBFI, IsCS); 1544 } 1545 1546 PreservedAnalyses PGOInstrumentationGen::run(Module &M, 1547 ModuleAnalysisManager &AM) { 1548 auto &FAM = AM.getResult<FunctionAnalysisManagerModuleProxy>(M).getManager(); 1549 auto LookupBPI = [&FAM](Function &F) { 1550 return &FAM.getResult<BranchProbabilityAnalysis>(F); 1551 }; 1552 1553 auto LookupBFI = [&FAM](Function &F) { 1554 return &FAM.getResult<BlockFrequencyAnalysis>(F); 1555 }; 1556 1557 if (!InstrumentAllFunctions(M, LookupBPI, LookupBFI, IsCS)) 1558 return PreservedAnalyses::all(); 1559 1560 return PreservedAnalyses::none(); 1561 } 1562 1563 static bool annotateAllFunctions( 1564 Module &M, StringRef ProfileFileName, StringRef ProfileRemappingFileName, 1565 function_ref<BranchProbabilityInfo *(Function &)> LookupBPI, 1566 function_ref<BlockFrequencyInfo *(Function &)> LookupBFI, 1567 ProfileSummaryInfo *PSI, bool IsCS) { 1568 LLVM_DEBUG(dbgs() << "Read in profile counters: "); 1569 auto &Ctx = M.getContext(); 1570 // Read the counter array from file. 1571 auto ReaderOrErr = 1572 IndexedInstrProfReader::create(ProfileFileName, ProfileRemappingFileName); 1573 if (Error E = ReaderOrErr.takeError()) { 1574 handleAllErrors(std::move(E), [&](const ErrorInfoBase &EI) { 1575 Ctx.diagnose( 1576 DiagnosticInfoPGOProfile(ProfileFileName.data(), EI.message())); 1577 }); 1578 return false; 1579 } 1580 1581 std::unique_ptr<IndexedInstrProfReader> PGOReader = 1582 std::move(ReaderOrErr.get()); 1583 if (!PGOReader) { 1584 Ctx.diagnose(DiagnosticInfoPGOProfile(ProfileFileName.data(), 1585 StringRef("Cannot get PGOReader"))); 1586 return false; 1587 } 1588 if (!PGOReader->hasCSIRLevelProfile() && IsCS) 1589 return false; 1590 1591 // TODO: might need to change the warning once the clang option is finalized. 1592 if (!PGOReader->isIRLevelProfile()) { 1593 Ctx.diagnose(DiagnosticInfoPGOProfile( 1594 ProfileFileName.data(), "Not an IR level instrumentation profile")); 1595 return false; 1596 } 1597 1598 // Add the profile summary (read from the header of the indexed summary) here 1599 // so that we can use it below when reading counters (which checks if the 1600 // function should be marked with a cold or inlinehint attribute). 1601 M.setProfileSummary(PGOReader->getSummary(IsCS).getMD(M.getContext()), 1602 IsCS ? ProfileSummary::PSK_CSInstr 1603 : ProfileSummary::PSK_Instr); 1604 1605 std::unordered_multimap<Comdat *, GlobalValue *> ComdatMembers; 1606 collectComdatMembers(M, ComdatMembers); 1607 std::vector<Function *> HotFunctions; 1608 std::vector<Function *> ColdFunctions; 1609 for (auto &F : M) { 1610 if (F.isDeclaration()) 1611 continue; 1612 auto *BPI = LookupBPI(F); 1613 auto *BFI = LookupBFI(F); 1614 // Split indirectbr critical edges here before computing the MST rather than 1615 // later in getInstrBB() to avoid invalidating it. 1616 SplitIndirectBrCriticalEdges(F, BPI, BFI); 1617 PGOUseFunc Func(F, &M, ComdatMembers, BPI, BFI, PSI, IsCS); 1618 bool AllZeros = false; 1619 if (!Func.readCounters(PGOReader.get(), AllZeros)) 1620 continue; 1621 if (AllZeros) { 1622 F.setEntryCount(ProfileCount(0, Function::PCT_Real)); 1623 if (Func.getProgramMaxCount() != 0) 1624 ColdFunctions.push_back(&F); 1625 continue; 1626 } 1627 Func.populateCounters(); 1628 Func.setBranchWeights(); 1629 Func.annotateValueSites(); 1630 Func.annotateIrrLoopHeaderWeights(); 1631 PGOUseFunc::FuncFreqAttr FreqAttr = Func.getFuncFreqAttr(); 1632 if (FreqAttr == PGOUseFunc::FFA_Cold) 1633 ColdFunctions.push_back(&F); 1634 else if (FreqAttr == PGOUseFunc::FFA_Hot) 1635 HotFunctions.push_back(&F); 1636 if (PGOViewCounts != PGOVCT_None && 1637 (ViewBlockFreqFuncName.empty() || 1638 F.getName().equals(ViewBlockFreqFuncName))) { 1639 LoopInfo LI{DominatorTree(F)}; 1640 std::unique_ptr<BranchProbabilityInfo> NewBPI = 1641 std::make_unique<BranchProbabilityInfo>(F, LI); 1642 std::unique_ptr<BlockFrequencyInfo> NewBFI = 1643 std::make_unique<BlockFrequencyInfo>(F, *NewBPI, LI); 1644 if (PGOViewCounts == PGOVCT_Graph) 1645 NewBFI->view(); 1646 else if (PGOViewCounts == PGOVCT_Text) { 1647 dbgs() << "pgo-view-counts: " << Func.getFunc().getName() << "\n"; 1648 NewBFI->print(dbgs()); 1649 } 1650 } 1651 if (PGOViewRawCounts != PGOVCT_None && 1652 (ViewBlockFreqFuncName.empty() || 1653 F.getName().equals(ViewBlockFreqFuncName))) { 1654 if (PGOViewRawCounts == PGOVCT_Graph) 1655 if (ViewBlockFreqFuncName.empty()) 1656 WriteGraph(&Func, Twine("PGORawCounts_") + Func.getFunc().getName()); 1657 else 1658 ViewGraph(&Func, Twine("PGORawCounts_") + Func.getFunc().getName()); 1659 else if (PGOViewRawCounts == PGOVCT_Text) { 1660 dbgs() << "pgo-view-raw-counts: " << Func.getFunc().getName() << "\n"; 1661 Func.dumpInfo(); 1662 } 1663 } 1664 } 1665 1666 // Set function hotness attribute from the profile. 1667 // We have to apply these attributes at the end because their presence 1668 // can affect the BranchProbabilityInfo of any callers, resulting in an 1669 // inconsistent MST between prof-gen and prof-use. 1670 for (auto &F : HotFunctions) { 1671 F->addFnAttr(Attribute::InlineHint); 1672 LLVM_DEBUG(dbgs() << "Set inline attribute to function: " << F->getName() 1673 << "\n"); 1674 } 1675 for (auto &F : ColdFunctions) { 1676 F->addFnAttr(Attribute::Cold); 1677 LLVM_DEBUG(dbgs() << "Set cold attribute to function: " << F->getName() 1678 << "\n"); 1679 } 1680 return true; 1681 } 1682 1683 PGOInstrumentationUse::PGOInstrumentationUse(std::string Filename, 1684 std::string RemappingFilename, 1685 bool IsCS) 1686 : ProfileFileName(std::move(Filename)), 1687 ProfileRemappingFileName(std::move(RemappingFilename)), IsCS(IsCS) { 1688 if (!PGOTestProfileFile.empty()) 1689 ProfileFileName = PGOTestProfileFile; 1690 if (!PGOTestProfileRemappingFile.empty()) 1691 ProfileRemappingFileName = PGOTestProfileRemappingFile; 1692 } 1693 1694 PreservedAnalyses PGOInstrumentationUse::run(Module &M, 1695 ModuleAnalysisManager &AM) { 1696 1697 auto &FAM = AM.getResult<FunctionAnalysisManagerModuleProxy>(M).getManager(); 1698 auto LookupBPI = [&FAM](Function &F) { 1699 return &FAM.getResult<BranchProbabilityAnalysis>(F); 1700 }; 1701 1702 auto LookupBFI = [&FAM](Function &F) { 1703 return &FAM.getResult<BlockFrequencyAnalysis>(F); 1704 }; 1705 1706 auto *PSI = &AM.getResult<ProfileSummaryAnalysis>(M); 1707 1708 if (!annotateAllFunctions(M, ProfileFileName, ProfileRemappingFileName, 1709 LookupBPI, LookupBFI, PSI, IsCS)) 1710 return PreservedAnalyses::all(); 1711 1712 return PreservedAnalyses::none(); 1713 } 1714 1715 bool PGOInstrumentationUseLegacyPass::runOnModule(Module &M) { 1716 if (skipModule(M)) 1717 return false; 1718 1719 auto LookupBPI = [this](Function &F) { 1720 return &this->getAnalysis<BranchProbabilityInfoWrapperPass>(F).getBPI(); 1721 }; 1722 auto LookupBFI = [this](Function &F) { 1723 return &this->getAnalysis<BlockFrequencyInfoWrapperPass>(F).getBFI(); 1724 }; 1725 1726 auto *PSI = &getAnalysis<ProfileSummaryInfoWrapperPass>().getPSI(); 1727 return annotateAllFunctions(M, ProfileFileName, "", LookupBPI, LookupBFI, PSI, 1728 IsCS); 1729 } 1730 1731 static std::string getSimpleNodeName(const BasicBlock *Node) { 1732 if (!Node->getName().empty()) 1733 return std::string(Node->getName()); 1734 1735 std::string SimpleNodeName; 1736 raw_string_ostream OS(SimpleNodeName); 1737 Node->printAsOperand(OS, false); 1738 return OS.str(); 1739 } 1740 1741 void llvm::setProfMetadata(Module *M, Instruction *TI, 1742 ArrayRef<uint64_t> EdgeCounts, 1743 uint64_t MaxCount) { 1744 MDBuilder MDB(M->getContext()); 1745 assert(MaxCount > 0 && "Bad max count"); 1746 uint64_t Scale = calculateCountScale(MaxCount); 1747 SmallVector<unsigned, 4> Weights; 1748 for (const auto &ECI : EdgeCounts) 1749 Weights.push_back(scaleBranchCount(ECI, Scale)); 1750 1751 LLVM_DEBUG(dbgs() << "Weight is: "; for (const auto &W 1752 : Weights) { 1753 dbgs() << W << " "; 1754 } dbgs() << "\n";); 1755 1756 misexpect::verifyMisExpect(TI, Weights, TI->getContext()); 1757 1758 TI->setMetadata(LLVMContext::MD_prof, MDB.createBranchWeights(Weights)); 1759 if (EmitBranchProbability) { 1760 std::string BrCondStr = getBranchCondString(TI); 1761 if (BrCondStr.empty()) 1762 return; 1763 1764 uint64_t WSum = 1765 std::accumulate(Weights.begin(), Weights.end(), (uint64_t)0, 1766 [](uint64_t w1, uint64_t w2) { return w1 + w2; }); 1767 uint64_t TotalCount = 1768 std::accumulate(EdgeCounts.begin(), EdgeCounts.end(), (uint64_t)0, 1769 [](uint64_t c1, uint64_t c2) { return c1 + c2; }); 1770 Scale = calculateCountScale(WSum); 1771 BranchProbability BP(scaleBranchCount(Weights[0], Scale), 1772 scaleBranchCount(WSum, Scale)); 1773 std::string BranchProbStr; 1774 raw_string_ostream OS(BranchProbStr); 1775 OS << BP; 1776 OS << " (total count : " << TotalCount << ")"; 1777 OS.flush(); 1778 Function *F = TI->getParent()->getParent(); 1779 OptimizationRemarkEmitter ORE(F); 1780 ORE.emit([&]() { 1781 return OptimizationRemark(DEBUG_TYPE, "pgo-instrumentation", TI) 1782 << BrCondStr << " is true with probability : " << BranchProbStr; 1783 }); 1784 } 1785 } 1786 1787 namespace llvm { 1788 1789 void setIrrLoopHeaderMetadata(Module *M, Instruction *TI, uint64_t Count) { 1790 MDBuilder MDB(M->getContext()); 1791 TI->setMetadata(llvm::LLVMContext::MD_irr_loop, 1792 MDB.createIrrLoopHeaderWeight(Count)); 1793 } 1794 1795 template <> struct GraphTraits<PGOUseFunc *> { 1796 using NodeRef = const BasicBlock *; 1797 using ChildIteratorType = const_succ_iterator; 1798 using nodes_iterator = pointer_iterator<Function::const_iterator>; 1799 1800 static NodeRef getEntryNode(const PGOUseFunc *G) { 1801 return &G->getFunc().front(); 1802 } 1803 1804 static ChildIteratorType child_begin(const NodeRef N) { 1805 return succ_begin(N); 1806 } 1807 1808 static ChildIteratorType child_end(const NodeRef N) { return succ_end(N); } 1809 1810 static nodes_iterator nodes_begin(const PGOUseFunc *G) { 1811 return nodes_iterator(G->getFunc().begin()); 1812 } 1813 1814 static nodes_iterator nodes_end(const PGOUseFunc *G) { 1815 return nodes_iterator(G->getFunc().end()); 1816 } 1817 }; 1818 1819 template <> struct DOTGraphTraits<PGOUseFunc *> : DefaultDOTGraphTraits { 1820 explicit DOTGraphTraits(bool isSimple = false) 1821 : DefaultDOTGraphTraits(isSimple) {} 1822 1823 static std::string getGraphName(const PGOUseFunc *G) { 1824 return std::string(G->getFunc().getName()); 1825 } 1826 1827 std::string getNodeLabel(const BasicBlock *Node, const PGOUseFunc *Graph) { 1828 std::string Result; 1829 raw_string_ostream OS(Result); 1830 1831 OS << getSimpleNodeName(Node) << ":\\l"; 1832 UseBBInfo *BI = Graph->findBBInfo(Node); 1833 OS << "Count : "; 1834 if (BI && BI->CountValid) 1835 OS << BI->CountValue << "\\l"; 1836 else 1837 OS << "Unknown\\l"; 1838 1839 if (!PGOInstrSelect) 1840 return Result; 1841 1842 for (auto BI = Node->begin(); BI != Node->end(); ++BI) { 1843 auto *I = &*BI; 1844 if (!isa<SelectInst>(I)) 1845 continue; 1846 // Display scaled counts for SELECT instruction: 1847 OS << "SELECT : { T = "; 1848 uint64_t TC, FC; 1849 bool HasProf = I->extractProfMetadata(TC, FC); 1850 if (!HasProf) 1851 OS << "Unknown, F = Unknown }\\l"; 1852 else 1853 OS << TC << ", F = " << FC << " }\\l"; 1854 } 1855 return Result; 1856 } 1857 }; 1858 1859 } // end namespace llvm 1860