1 //===- CGSCCPassManager.h - Call graph pass management ----------*- C++ -*-===// 2 // 3 // The LLVM Compiler Infrastructure 4 // 5 // This file is distributed under the University of Illinois Open Source 6 // License. See LICENSE.TXT for details. 7 // 8 //===----------------------------------------------------------------------===// 9 /// \file 10 /// 11 /// This header provides classes for managing passes over SCCs of the call 12 /// graph. These passes form an important component of LLVM's interprocedural 13 /// optimizations. Because they operate on the SCCs of the call graph, and they 14 /// traverse the graph in post-order, they can effectively do pair-wise 15 /// interprocedural optimizations for all call edges in the program while 16 /// incrementally refining it and improving the context of these pair-wise 17 /// optimizations. At each call site edge, the callee has already been 18 /// optimized as much as is possible. This in turn allows very accurate 19 /// analysis of it for IPO. 20 /// 21 /// A secondary more general goal is to be able to isolate optimization on 22 /// unrelated parts of the IR module. This is useful to ensure our 23 /// optimizations are principled and don't miss oportunities where refinement 24 /// of one part of the module influence transformations in another part of the 25 /// module. But this is also useful if we want to parallelize the optimizations 26 /// across common large module graph shapes which tend to be very wide and have 27 /// large regions of unrelated cliques. 28 /// 29 /// To satisfy these goals, we use the LazyCallGraph which provides two graphs 30 /// nested inside each other (and built lazily from the bottom-up): the call 31 /// graph proper, and a reference graph. The reference graph is super set of 32 /// the call graph and is a conservative approximation of what could through 33 /// scalar or CGSCC transforms *become* the call graph. Using this allows us to 34 /// ensure we optimize functions prior to them being introduced into the call 35 /// graph by devirtualization or other technique, and thus ensures that 36 /// subsequent pair-wise interprocedural optimizations observe the optimized 37 /// form of these functions. The (potentially transitive) reference 38 /// reachability used by the reference graph is a conservative approximation 39 /// that still allows us to have independent regions of the graph. 40 /// 41 /// FIXME: There is one major drawback of the reference graph: in its naive 42 /// form it is quadratic because it contains a distinct edge for each 43 /// (potentially indirect) reference, even if are all through some common 44 /// global table of function pointers. This can be fixed in a number of ways 45 /// that essentially preserve enough of the normalization. While it isn't 46 /// expected to completely preclude the usability of this, it will need to be 47 /// addressed. 48 /// 49 /// 50 /// All of these issues are made substantially more complex in the face of 51 /// mutations to the call graph while optimization passes are being run. When 52 /// mutations to the call graph occur we want to achieve two different things: 53 /// 54 /// - We need to update the call graph in-flight and invalidate analyses 55 /// cached on entities in the graph. Because of the cache-based analysis 56 /// design of the pass manager, it is essential to have stable identities for 57 /// the elements of the IR that passes traverse, and to invalidate any 58 /// analyses cached on these elements as the mutations take place. 59 /// 60 /// - We want to preserve the incremental and post-order traversal of the 61 /// graph even as it is refined and mutated. This means we want optimization 62 /// to observe the most refined form of the call graph and to do so in 63 /// post-order. 64 /// 65 /// To address this, the CGSCC manager uses both worklists that can be expanded 66 /// by passes which transform the IR, and provides invalidation tests to skip 67 /// entries that become dead. This extra data is provided to every SCC pass so 68 /// that it can carefully update the manager's traversal as the call graph 69 /// mutates. 70 /// 71 /// We also provide support for running function passes within the CGSCC walk, 72 /// and there we provide automatic update of the call graph including of the 73 /// pass manager to reflect call graph changes that fall out naturally as part 74 /// of scalar transformations. 75 /// 76 /// The patterns used to ensure the goals of post-order visitation of the fully 77 /// refined graph: 78 /// 79 /// 1) Sink toward the "bottom" as the graph is refined. This means that any 80 /// iteration continues in some valid post-order sequence after the mutation 81 /// has altered the structure. 82 /// 83 /// 2) Enqueue in post-order, including the current entity. If the current 84 /// entity's shape changes, it and everything after it in post-order needs 85 /// to be visited to observe that shape. 86 /// 87 //===----------------------------------------------------------------------===// 88 89 #ifndef LLVM_ANALYSIS_CGSCCPASSMANAGER_H 90 #define LLVM_ANALYSIS_CGSCCPASSMANAGER_H 91 92 #include "llvm/ADT/DenseSet.h" 93 #include "llvm/ADT/PriorityWorklist.h" 94 #include "llvm/ADT/STLExtras.h" 95 #include "llvm/ADT/SmallPtrSet.h" 96 #include "llvm/ADT/SmallVector.h" 97 #include "llvm/Analysis/LazyCallGraph.h" 98 #include "llvm/IR/CallSite.h" 99 #include "llvm/IR/Function.h" 100 #include "llvm/IR/InstIterator.h" 101 #include "llvm/IR/PassManager.h" 102 #include "llvm/IR/ValueHandle.h" 103 #include "llvm/Support/Debug.h" 104 #include "llvm/Support/raw_ostream.h" 105 #include <algorithm> 106 #include <cassert> 107 #include <utility> 108 109 namespace llvm { 110 111 struct CGSCCUpdateResult; 112 class Module; 113 114 // Allow debug logging in this inline function. 115 #define DEBUG_TYPE "cgscc" 116 117 /// Extern template declaration for the analysis set for this IR unit. 118 extern template class AllAnalysesOn<LazyCallGraph::SCC>; 119 120 extern template class AnalysisManager<LazyCallGraph::SCC, LazyCallGraph &>; 121 122 /// \brief The CGSCC analysis manager. 123 /// 124 /// See the documentation for the AnalysisManager template for detail 125 /// documentation. This type serves as a convenient way to refer to this 126 /// construct in the adaptors and proxies used to integrate this into the larger 127 /// pass manager infrastructure. 128 using CGSCCAnalysisManager = 129 AnalysisManager<LazyCallGraph::SCC, LazyCallGraph &>; 130 131 // Explicit specialization and instantiation declarations for the pass manager. 132 // See the comments on the definition of the specialization for details on how 133 // it differs from the primary template. 134 template <> 135 PreservedAnalyses 136 PassManager<LazyCallGraph::SCC, CGSCCAnalysisManager, LazyCallGraph &, 137 CGSCCUpdateResult &>::run(LazyCallGraph::SCC &InitialC, 138 CGSCCAnalysisManager &AM, 139 LazyCallGraph &G, CGSCCUpdateResult &UR); 140 extern template class PassManager<LazyCallGraph::SCC, CGSCCAnalysisManager, 141 LazyCallGraph &, CGSCCUpdateResult &>; 142 143 /// \brief The CGSCC pass manager. 144 /// 145 /// See the documentation for the PassManager template for details. It runs 146 /// a sequence of SCC passes over each SCC that the manager is run over. This 147 /// type serves as a convenient way to refer to this construct. 148 using CGSCCPassManager = 149 PassManager<LazyCallGraph::SCC, CGSCCAnalysisManager, LazyCallGraph &, 150 CGSCCUpdateResult &>; 151 152 /// An explicit specialization of the require analysis template pass. 153 template <typename AnalysisT> 154 struct RequireAnalysisPass<AnalysisT, LazyCallGraph::SCC, CGSCCAnalysisManager, 155 LazyCallGraph &, CGSCCUpdateResult &> 156 : PassInfoMixin<RequireAnalysisPass<AnalysisT, LazyCallGraph::SCC, 157 CGSCCAnalysisManager, LazyCallGraph &, 158 CGSCCUpdateResult &>> { 159 PreservedAnalyses run(LazyCallGraph::SCC &C, CGSCCAnalysisManager &AM, 160 LazyCallGraph &CG, CGSCCUpdateResult &) { 161 (void)AM.template getResult<AnalysisT>(C, CG); 162 return PreservedAnalyses::all(); 163 } 164 }; 165 166 /// A proxy from a \c CGSCCAnalysisManager to a \c Module. 167 using CGSCCAnalysisManagerModuleProxy = 168 InnerAnalysisManagerProxy<CGSCCAnalysisManager, Module>; 169 170 /// We need a specialized result for the \c CGSCCAnalysisManagerModuleProxy so 171 /// it can have access to the call graph in order to walk all the SCCs when 172 /// invalidating things. 173 template <> class CGSCCAnalysisManagerModuleProxy::Result { 174 public: 175 explicit Result(CGSCCAnalysisManager &InnerAM, LazyCallGraph &G) 176 : InnerAM(&InnerAM), G(&G) {} 177 178 /// \brief Accessor for the analysis manager. 179 CGSCCAnalysisManager &getManager() { return *InnerAM; } 180 181 /// \brief Handler for invalidation of the Module. 182 /// 183 /// If the proxy analysis itself is preserved, then we assume that the set of 184 /// SCCs in the Module hasn't changed. Thus any pointers to SCCs in the 185 /// CGSCCAnalysisManager are still valid, and we don't need to call \c clear 186 /// on the CGSCCAnalysisManager. 187 /// 188 /// Regardless of whether this analysis is marked as preserved, all of the 189 /// analyses in the \c CGSCCAnalysisManager are potentially invalidated based 190 /// on the set of preserved analyses. 191 bool invalidate(Module &M, const PreservedAnalyses &PA, 192 ModuleAnalysisManager::Invalidator &Inv); 193 194 private: 195 CGSCCAnalysisManager *InnerAM; 196 LazyCallGraph *G; 197 }; 198 199 /// Provide a specialized run method for the \c CGSCCAnalysisManagerModuleProxy 200 /// so it can pass the lazy call graph to the result. 201 template <> 202 CGSCCAnalysisManagerModuleProxy::Result 203 CGSCCAnalysisManagerModuleProxy::run(Module &M, ModuleAnalysisManager &AM); 204 205 // Ensure the \c CGSCCAnalysisManagerModuleProxy is provided as an extern 206 // template. 207 extern template class InnerAnalysisManagerProxy<CGSCCAnalysisManager, Module>; 208 209 extern template class OuterAnalysisManagerProxy< 210 ModuleAnalysisManager, LazyCallGraph::SCC, LazyCallGraph &>; 211 212 /// A proxy from a \c ModuleAnalysisManager to an \c SCC. 213 using ModuleAnalysisManagerCGSCCProxy = 214 OuterAnalysisManagerProxy<ModuleAnalysisManager, LazyCallGraph::SCC, 215 LazyCallGraph &>; 216 217 /// Support structure for SCC passes to communicate updates the call graph back 218 /// to the CGSCC pass manager infrsatructure. 219 /// 220 /// The CGSCC pass manager runs SCC passes which are allowed to update the call 221 /// graph and SCC structures. This means the structure the pass manager works 222 /// on is mutating underneath it. In order to support that, there needs to be 223 /// careful communication about the precise nature and ramifications of these 224 /// updates to the pass management infrastructure. 225 /// 226 /// All SCC passes will have to accept a reference to the management layer's 227 /// update result struct and use it to reflect the results of any CG updates 228 /// performed. 229 /// 230 /// Passes which do not change the call graph structure in any way can just 231 /// ignore this argument to their run method. 232 struct CGSCCUpdateResult { 233 /// Worklist of the RefSCCs queued for processing. 234 /// 235 /// When a pass refines the graph and creates new RefSCCs or causes them to 236 /// have a different shape or set of component SCCs it should add the RefSCCs 237 /// to this worklist so that we visit them in the refined form. 238 /// 239 /// This worklist is in reverse post-order, as we pop off the back in order 240 /// to observe RefSCCs in post-order. When adding RefSCCs, clients should add 241 /// them in reverse post-order. 242 SmallPriorityWorklist<LazyCallGraph::RefSCC *, 1> &RCWorklist; 243 244 /// Worklist of the SCCs queued for processing. 245 /// 246 /// When a pass refines the graph and creates new SCCs or causes them to have 247 /// a different shape or set of component functions it should add the SCCs to 248 /// this worklist so that we visit them in the refined form. 249 /// 250 /// Note that if the SCCs are part of a RefSCC that is added to the \c 251 /// RCWorklist, they don't need to be added here as visiting the RefSCC will 252 /// be sufficient to re-visit the SCCs within it. 253 /// 254 /// This worklist is in reverse post-order, as we pop off the back in order 255 /// to observe SCCs in post-order. When adding SCCs, clients should add them 256 /// in reverse post-order. 257 SmallPriorityWorklist<LazyCallGraph::SCC *, 1> &CWorklist; 258 259 /// The set of invalidated RefSCCs which should be skipped if they are found 260 /// in \c RCWorklist. 261 /// 262 /// This is used to quickly prune out RefSCCs when they get deleted and 263 /// happen to already be on the worklist. We use this primarily to avoid 264 /// scanning the list and removing entries from it. 265 SmallPtrSetImpl<LazyCallGraph::RefSCC *> &InvalidatedRefSCCs; 266 267 /// The set of invalidated SCCs which should be skipped if they are found 268 /// in \c CWorklist. 269 /// 270 /// This is used to quickly prune out SCCs when they get deleted and happen 271 /// to already be on the worklist. We use this primarily to avoid scanning 272 /// the list and removing entries from it. 273 SmallPtrSetImpl<LazyCallGraph::SCC *> &InvalidatedSCCs; 274 275 /// If non-null, the updated current \c RefSCC being processed. 276 /// 277 /// This is set when a graph refinement takes place an the "current" point in 278 /// the graph moves "down" or earlier in the post-order walk. This will often 279 /// cause the "current" RefSCC to be a newly created RefSCC object and the 280 /// old one to be added to the above worklist. When that happens, this 281 /// pointer is non-null and can be used to continue processing the "top" of 282 /// the post-order walk. 283 LazyCallGraph::RefSCC *UpdatedRC; 284 285 /// If non-null, the updated current \c SCC being processed. 286 /// 287 /// This is set when a graph refinement takes place an the "current" point in 288 /// the graph moves "down" or earlier in the post-order walk. This will often 289 /// cause the "current" SCC to be a newly created SCC object and the old one 290 /// to be added to the above worklist. When that happens, this pointer is 291 /// non-null and can be used to continue processing the "top" of the 292 /// post-order walk. 293 LazyCallGraph::SCC *UpdatedC; 294 295 /// A hacky area where the inliner can retain history about inlining 296 /// decisions that mutated the call graph's SCC structure in order to avoid 297 /// infinite inlining. See the comments in the inliner's CG update logic. 298 /// 299 /// FIXME: Keeping this here seems like a big layering issue, we should look 300 /// for a better technique. 301 SmallDenseSet<std::pair<LazyCallGraph::Node *, LazyCallGraph::SCC *>, 4> 302 &InlinedInternalEdges; 303 }; 304 305 /// \brief The core module pass which does a post-order walk of the SCCs and 306 /// runs a CGSCC pass over each one. 307 /// 308 /// Designed to allow composition of a CGSCCPass(Manager) and 309 /// a ModulePassManager. Note that this pass must be run with a module analysis 310 /// manager as it uses the LazyCallGraph analysis. It will also run the 311 /// \c CGSCCAnalysisManagerModuleProxy analysis prior to running the CGSCC 312 /// pass over the module to enable a \c FunctionAnalysisManager to be used 313 /// within this run safely. 314 template <typename CGSCCPassT> 315 class ModuleToPostOrderCGSCCPassAdaptor 316 : public PassInfoMixin<ModuleToPostOrderCGSCCPassAdaptor<CGSCCPassT>> { 317 public: 318 explicit ModuleToPostOrderCGSCCPassAdaptor(CGSCCPassT Pass) 319 : Pass(std::move(Pass)) {} 320 321 // We have to explicitly define all the special member functions because MSVC 322 // refuses to generate them. 323 ModuleToPostOrderCGSCCPassAdaptor( 324 const ModuleToPostOrderCGSCCPassAdaptor &Arg) 325 : Pass(Arg.Pass) {} 326 327 ModuleToPostOrderCGSCCPassAdaptor(ModuleToPostOrderCGSCCPassAdaptor &&Arg) 328 : Pass(std::move(Arg.Pass)) {} 329 330 friend void swap(ModuleToPostOrderCGSCCPassAdaptor &LHS, 331 ModuleToPostOrderCGSCCPassAdaptor &RHS) { 332 std::swap(LHS.Pass, RHS.Pass); 333 } 334 335 ModuleToPostOrderCGSCCPassAdaptor & 336 operator=(ModuleToPostOrderCGSCCPassAdaptor RHS) { 337 swap(*this, RHS); 338 return *this; 339 } 340 341 /// \brief Runs the CGSCC pass across every SCC in the module. 342 PreservedAnalyses run(Module &M, ModuleAnalysisManager &AM) { 343 // Setup the CGSCC analysis manager from its proxy. 344 CGSCCAnalysisManager &CGAM = 345 AM.getResult<CGSCCAnalysisManagerModuleProxy>(M).getManager(); 346 347 // Get the call graph for this module. 348 LazyCallGraph &CG = AM.getResult<LazyCallGraphAnalysis>(M); 349 350 // We keep worklists to allow us to push more work onto the pass manager as 351 // the passes are run. 352 SmallPriorityWorklist<LazyCallGraph::RefSCC *, 1> RCWorklist; 353 SmallPriorityWorklist<LazyCallGraph::SCC *, 1> CWorklist; 354 355 // Keep sets for invalidated SCCs and RefSCCs that should be skipped when 356 // iterating off the worklists. 357 SmallPtrSet<LazyCallGraph::RefSCC *, 4> InvalidRefSCCSet; 358 SmallPtrSet<LazyCallGraph::SCC *, 4> InvalidSCCSet; 359 360 SmallDenseSet<std::pair<LazyCallGraph::Node *, LazyCallGraph::SCC *>, 4> 361 InlinedInternalEdges; 362 363 CGSCCUpdateResult UR = {RCWorklist, CWorklist, InvalidRefSCCSet, 364 InvalidSCCSet, nullptr, nullptr, 365 InlinedInternalEdges}; 366 367 PreservedAnalyses PA = PreservedAnalyses::all(); 368 CG.buildRefSCCs(); 369 for (auto RCI = CG.postorder_ref_scc_begin(), 370 RCE = CG.postorder_ref_scc_end(); 371 RCI != RCE;) { 372 assert(RCWorklist.empty() && 373 "Should always start with an empty RefSCC worklist"); 374 // The postorder_ref_sccs range we are walking is lazily constructed, so 375 // we only push the first one onto the worklist. The worklist allows us 376 // to capture *new* RefSCCs created during transformations. 377 // 378 // We really want to form RefSCCs lazily because that makes them cheaper 379 // to update as the program is simplified and allows us to have greater 380 // cache locality as forming a RefSCC touches all the parts of all the 381 // functions within that RefSCC. 382 // 383 // We also eagerly increment the iterator to the next position because 384 // the CGSCC passes below may delete the current RefSCC. 385 RCWorklist.insert(&*RCI++); 386 387 do { 388 LazyCallGraph::RefSCC *RC = RCWorklist.pop_back_val(); 389 if (InvalidRefSCCSet.count(RC)) { 390 DEBUG(dbgs() << "Skipping an invalid RefSCC...\n"); 391 continue; 392 } 393 394 assert(CWorklist.empty() && 395 "Should always start with an empty SCC worklist"); 396 397 DEBUG(dbgs() << "Running an SCC pass across the RefSCC: " << *RC 398 << "\n"); 399 400 // Push the initial SCCs in reverse post-order as we'll pop off the the 401 // back and so see this in post-order. 402 for (LazyCallGraph::SCC &C : llvm::reverse(*RC)) 403 CWorklist.insert(&C); 404 405 do { 406 LazyCallGraph::SCC *C = CWorklist.pop_back_val(); 407 // Due to call graph mutations, we may have invalid SCCs or SCCs from 408 // other RefSCCs in the worklist. The invalid ones are dead and the 409 // other RefSCCs should be queued above, so we just need to skip both 410 // scenarios here. 411 if (InvalidSCCSet.count(C)) { 412 DEBUG(dbgs() << "Skipping an invalid SCC...\n"); 413 continue; 414 } 415 if (&C->getOuterRefSCC() != RC) { 416 DEBUG(dbgs() << "Skipping an SCC that is now part of some other " 417 "RefSCC...\n"); 418 continue; 419 } 420 421 do { 422 // Check that we didn't miss any update scenario. 423 assert(!InvalidSCCSet.count(C) && "Processing an invalid SCC!"); 424 assert(C->begin() != C->end() && "Cannot have an empty SCC!"); 425 assert(&C->getOuterRefSCC() == RC && 426 "Processing an SCC in a different RefSCC!"); 427 428 UR.UpdatedRC = nullptr; 429 UR.UpdatedC = nullptr; 430 PreservedAnalyses PassPA = Pass.run(*C, CGAM, CG, UR); 431 432 // Update the SCC and RefSCC if necessary. 433 C = UR.UpdatedC ? UR.UpdatedC : C; 434 RC = UR.UpdatedRC ? UR.UpdatedRC : RC; 435 436 // If the CGSCC pass wasn't able to provide a valid updated SCC, 437 // the current SCC may simply need to be skipped if invalid. 438 if (UR.InvalidatedSCCs.count(C)) { 439 DEBUG(dbgs() << "Skipping invalidated root or island SCC!\n"); 440 break; 441 } 442 // Check that we didn't miss any update scenario. 443 assert(C->begin() != C->end() && "Cannot have an empty SCC!"); 444 445 // We handle invalidating the CGSCC analysis manager's information 446 // for the (potentially updated) SCC here. Note that any other SCCs 447 // whose structure has changed should have been invalidated by 448 // whatever was updating the call graph. This SCC gets invalidated 449 // late as it contains the nodes that were actively being 450 // processed. 451 CGAM.invalidate(*C, PassPA); 452 453 // Then intersect the preserved set so that invalidation of module 454 // analyses will eventually occur when the module pass completes. 455 PA.intersect(std::move(PassPA)); 456 457 // The pass may have restructured the call graph and refined the 458 // current SCC and/or RefSCC. We need to update our current SCC and 459 // RefSCC pointers to follow these. Also, when the current SCC is 460 // refined, re-run the SCC pass over the newly refined SCC in order 461 // to observe the most precise SCC model available. This inherently 462 // cannot cycle excessively as it only happens when we split SCCs 463 // apart, at most converging on a DAG of single nodes. 464 // FIXME: If we ever start having RefSCC passes, we'll want to 465 // iterate there too. 466 if (UR.UpdatedC) 467 DEBUG(dbgs() << "Re-running SCC passes after a refinement of the " 468 "current SCC: " 469 << *UR.UpdatedC << "\n"); 470 471 // Note that both `C` and `RC` may at this point refer to deleted, 472 // invalid SCC and RefSCCs respectively. But we will short circuit 473 // the processing when we check them in the loop above. 474 } while (UR.UpdatedC); 475 } while (!CWorklist.empty()); 476 477 // We only need to keep internal inlined edge information within 478 // a RefSCC, clear it to save on space and let the next time we visit 479 // any of these functions have a fresh start. 480 InlinedInternalEdges.clear(); 481 } while (!RCWorklist.empty()); 482 } 483 484 // By definition we preserve the call garph, all SCC analyses, and the 485 // analysis proxies by handling them above and in any nested pass managers. 486 PA.preserveSet<AllAnalysesOn<LazyCallGraph::SCC>>(); 487 PA.preserve<LazyCallGraphAnalysis>(); 488 PA.preserve<CGSCCAnalysisManagerModuleProxy>(); 489 PA.preserve<FunctionAnalysisManagerModuleProxy>(); 490 return PA; 491 } 492 493 private: 494 CGSCCPassT Pass; 495 }; 496 497 /// \brief A function to deduce a function pass type and wrap it in the 498 /// templated adaptor. 499 template <typename CGSCCPassT> 500 ModuleToPostOrderCGSCCPassAdaptor<CGSCCPassT> 501 createModuleToPostOrderCGSCCPassAdaptor(CGSCCPassT Pass) { 502 return ModuleToPostOrderCGSCCPassAdaptor<CGSCCPassT>(std::move(Pass)); 503 } 504 505 /// A proxy from a \c FunctionAnalysisManager to an \c SCC. 506 /// 507 /// When a module pass runs and triggers invalidation, both the CGSCC and 508 /// Function analysis manager proxies on the module get an invalidation event. 509 /// We don't want to fully duplicate responsibility for most of the 510 /// invalidation logic. Instead, this layer is only responsible for SCC-local 511 /// invalidation events. We work with the module's FunctionAnalysisManager to 512 /// invalidate function analyses. 513 class FunctionAnalysisManagerCGSCCProxy 514 : public AnalysisInfoMixin<FunctionAnalysisManagerCGSCCProxy> { 515 public: 516 class Result { 517 public: 518 explicit Result(FunctionAnalysisManager &FAM) : FAM(&FAM) {} 519 520 /// \brief Accessor for the analysis manager. 521 FunctionAnalysisManager &getManager() { return *FAM; } 522 523 bool invalidate(LazyCallGraph::SCC &C, const PreservedAnalyses &PA, 524 CGSCCAnalysisManager::Invalidator &Inv); 525 526 private: 527 FunctionAnalysisManager *FAM; 528 }; 529 530 /// Computes the \c FunctionAnalysisManager and stores it in the result proxy. 531 Result run(LazyCallGraph::SCC &C, CGSCCAnalysisManager &AM, LazyCallGraph &); 532 533 private: 534 friend AnalysisInfoMixin<FunctionAnalysisManagerCGSCCProxy>; 535 536 static AnalysisKey Key; 537 }; 538 539 extern template class OuterAnalysisManagerProxy<CGSCCAnalysisManager, Function>; 540 541 /// A proxy from a \c CGSCCAnalysisManager to a \c Function. 542 using CGSCCAnalysisManagerFunctionProxy = 543 OuterAnalysisManagerProxy<CGSCCAnalysisManager, Function>; 544 545 /// Helper to update the call graph after running a function pass. 546 /// 547 /// Function passes can only mutate the call graph in specific ways. This 548 /// routine provides a helper that updates the call graph in those ways 549 /// including returning whether any changes were made and populating a CG 550 /// update result struct for the overall CGSCC walk. 551 LazyCallGraph::SCC &updateCGAndAnalysisManagerForFunctionPass( 552 LazyCallGraph &G, LazyCallGraph::SCC &C, LazyCallGraph::Node &N, 553 CGSCCAnalysisManager &AM, CGSCCUpdateResult &UR); 554 555 /// \brief Adaptor that maps from a SCC to its functions. 556 /// 557 /// Designed to allow composition of a FunctionPass(Manager) and 558 /// a CGSCCPassManager. Note that if this pass is constructed with a pointer 559 /// to a \c CGSCCAnalysisManager it will run the 560 /// \c FunctionAnalysisManagerCGSCCProxy analysis prior to running the function 561 /// pass over the SCC to enable a \c FunctionAnalysisManager to be used 562 /// within this run safely. 563 template <typename FunctionPassT> 564 class CGSCCToFunctionPassAdaptor 565 : public PassInfoMixin<CGSCCToFunctionPassAdaptor<FunctionPassT>> { 566 public: 567 explicit CGSCCToFunctionPassAdaptor(FunctionPassT Pass) 568 : Pass(std::move(Pass)) {} 569 570 // We have to explicitly define all the special member functions because MSVC 571 // refuses to generate them. 572 CGSCCToFunctionPassAdaptor(const CGSCCToFunctionPassAdaptor &Arg) 573 : Pass(Arg.Pass) {} 574 575 CGSCCToFunctionPassAdaptor(CGSCCToFunctionPassAdaptor &&Arg) 576 : Pass(std::move(Arg.Pass)) {} 577 578 friend void swap(CGSCCToFunctionPassAdaptor &LHS, 579 CGSCCToFunctionPassAdaptor &RHS) { 580 std::swap(LHS.Pass, RHS.Pass); 581 } 582 583 CGSCCToFunctionPassAdaptor &operator=(CGSCCToFunctionPassAdaptor RHS) { 584 swap(*this, RHS); 585 return *this; 586 } 587 588 /// \brief Runs the function pass across every function in the module. 589 PreservedAnalyses run(LazyCallGraph::SCC &C, CGSCCAnalysisManager &AM, 590 LazyCallGraph &CG, CGSCCUpdateResult &UR) { 591 // Setup the function analysis manager from its proxy. 592 FunctionAnalysisManager &FAM = 593 AM.getResult<FunctionAnalysisManagerCGSCCProxy>(C, CG).getManager(); 594 595 SmallVector<LazyCallGraph::Node *, 4> Nodes; 596 for (LazyCallGraph::Node &N : C) 597 Nodes.push_back(&N); 598 599 // The SCC may get split while we are optimizing functions due to deleting 600 // edges. If this happens, the current SCC can shift, so keep track of 601 // a pointer we can overwrite. 602 LazyCallGraph::SCC *CurrentC = &C; 603 604 DEBUG(dbgs() << "Running function passes across an SCC: " << C << "\n"); 605 606 PreservedAnalyses PA = PreservedAnalyses::all(); 607 for (LazyCallGraph::Node *N : Nodes) { 608 // Skip nodes from other SCCs. These may have been split out during 609 // processing. We'll eventually visit those SCCs and pick up the nodes 610 // there. 611 if (CG.lookupSCC(*N) != CurrentC) 612 continue; 613 614 PreservedAnalyses PassPA = Pass.run(N->getFunction(), FAM); 615 616 // We know that the function pass couldn't have invalidated any other 617 // function's analyses (that's the contract of a function pass), so 618 // directly handle the function analysis manager's invalidation here. 619 FAM.invalidate(N->getFunction(), PassPA); 620 621 // Then intersect the preserved set so that invalidation of module 622 // analyses will eventually occur when the module pass completes. 623 PA.intersect(std::move(PassPA)); 624 625 // If the call graph hasn't been preserved, update it based on this 626 // function pass. This may also update the current SCC to point to 627 // a smaller, more refined SCC. 628 auto PAC = PA.getChecker<LazyCallGraphAnalysis>(); 629 if (!PAC.preserved() && !PAC.preservedSet<AllAnalysesOn<Module>>()) { 630 CurrentC = &updateCGAndAnalysisManagerForFunctionPass(CG, *CurrentC, *N, 631 AM, UR); 632 assert( 633 CG.lookupSCC(*N) == CurrentC && 634 "Current SCC not updated to the SCC containing the current node!"); 635 } 636 } 637 638 // By definition we preserve the proxy. And we preserve all analyses on 639 // Functions. This precludes *any* invalidation of function analyses by the 640 // proxy, but that's OK because we've taken care to invalidate analyses in 641 // the function analysis manager incrementally above. 642 PA.preserveSet<AllAnalysesOn<Function>>(); 643 PA.preserve<FunctionAnalysisManagerCGSCCProxy>(); 644 645 // We've also ensured that we updated the call graph along the way. 646 PA.preserve<LazyCallGraphAnalysis>(); 647 648 return PA; 649 } 650 651 private: 652 FunctionPassT Pass; 653 }; 654 655 /// \brief A function to deduce a function pass type and wrap it in the 656 /// templated adaptor. 657 template <typename FunctionPassT> 658 CGSCCToFunctionPassAdaptor<FunctionPassT> 659 createCGSCCToFunctionPassAdaptor(FunctionPassT Pass) { 660 return CGSCCToFunctionPassAdaptor<FunctionPassT>(std::move(Pass)); 661 } 662 663 /// A helper that repeats an SCC pass each time an indirect call is refined to 664 /// a direct call by that pass. 665 /// 666 /// While the CGSCC pass manager works to re-visit SCCs and RefSCCs as they 667 /// change shape, we may also want to repeat an SCC pass if it simply refines 668 /// an indirect call to a direct call, even if doing so does not alter the 669 /// shape of the graph. Note that this only pertains to direct calls to 670 /// functions where IPO across the SCC may be able to compute more precise 671 /// results. For intrinsics, we assume scalar optimizations already can fully 672 /// reason about them. 673 /// 674 /// This repetition has the potential to be very large however, as each one 675 /// might refine a single call site. As a consequence, in practice we use an 676 /// upper bound on the number of repetitions to limit things. 677 template <typename PassT> 678 class DevirtSCCRepeatedPass 679 : public PassInfoMixin<DevirtSCCRepeatedPass<PassT>> { 680 public: 681 explicit DevirtSCCRepeatedPass(PassT Pass, int MaxIterations) 682 : Pass(std::move(Pass)), MaxIterations(MaxIterations) {} 683 684 /// Runs the wrapped pass up to \c MaxIterations on the SCC, iterating 685 /// whenever an indirect call is refined. 686 PreservedAnalyses run(LazyCallGraph::SCC &InitialC, CGSCCAnalysisManager &AM, 687 LazyCallGraph &CG, CGSCCUpdateResult &UR) { 688 PreservedAnalyses PA = PreservedAnalyses::all(); 689 690 // The SCC may be refined while we are running passes over it, so set up 691 // a pointer that we can update. 692 LazyCallGraph::SCC *C = &InitialC; 693 694 // Collect value handles for all of the indirect call sites. 695 SmallVector<WeakTrackingVH, 8> CallHandles; 696 697 // Struct to track the counts of direct and indirect calls in each function 698 // of the SCC. 699 struct CallCount { 700 int Direct; 701 int Indirect; 702 }; 703 704 // Put value handles on all of the indirect calls and return the number of 705 // direct calls for each function in the SCC. 706 auto ScanSCC = [](LazyCallGraph::SCC &C, 707 SmallVectorImpl<WeakTrackingVH> &CallHandles) { 708 assert(CallHandles.empty() && "Must start with a clear set of handles."); 709 710 SmallVector<CallCount, 4> CallCounts; 711 for (LazyCallGraph::Node &N : C) { 712 CallCounts.push_back({0, 0}); 713 CallCount &Count = CallCounts.back(); 714 for (Instruction &I : instructions(N.getFunction())) 715 if (auto CS = CallSite(&I)) { 716 if (CS.getCalledFunction()) { 717 ++Count.Direct; 718 } else { 719 ++Count.Indirect; 720 CallHandles.push_back(WeakTrackingVH(&I)); 721 } 722 } 723 } 724 725 return CallCounts; 726 }; 727 728 // Populate the initial call handles and get the initial call counts. 729 auto CallCounts = ScanSCC(*C, CallHandles); 730 731 for (int Iteration = 0;; ++Iteration) { 732 PreservedAnalyses PassPA = Pass.run(*C, AM, CG, UR); 733 734 // If the SCC structure has changed, bail immediately and let the outer 735 // CGSCC layer handle any iteration to reflect the refined structure. 736 if (UR.UpdatedC && UR.UpdatedC != C) { 737 PA.intersect(std::move(PassPA)); 738 break; 739 } 740 741 // Check that we didn't miss any update scenario. 742 assert(!UR.InvalidatedSCCs.count(C) && "Processing an invalid SCC!"); 743 assert(C->begin() != C->end() && "Cannot have an empty SCC!"); 744 assert((int)CallCounts.size() == C->size() && 745 "Cannot have changed the size of the SCC!"); 746 747 // Check whether any of the handles were devirtualized. 748 auto IsDevirtualizedHandle = [&](WeakTrackingVH &CallH) { 749 if (!CallH) 750 return false; 751 auto CS = CallSite(CallH); 752 if (!CS) 753 return false; 754 755 // If the call is still indirect, leave it alone. 756 Function *F = CS.getCalledFunction(); 757 if (!F) 758 return false; 759 760 DEBUG(dbgs() << "Found devirutalized call from " 761 << CS.getParent()->getParent()->getName() << " to " 762 << F->getName() << "\n"); 763 764 // We now have a direct call where previously we had an indirect call, 765 // so iterate to process this devirtualization site. 766 return true; 767 }; 768 bool Devirt = llvm::any_of(CallHandles, IsDevirtualizedHandle); 769 770 // Rescan to build up a new set of handles and count how many direct 771 // calls remain. If we decide to iterate, this also sets up the input to 772 // the next iteration. 773 CallHandles.clear(); 774 auto NewCallCounts = ScanSCC(*C, CallHandles); 775 776 // If we haven't found an explicit devirtualization already see if we 777 // have decreased the number of indirect calls and increased the number 778 // of direct calls for any function in the SCC. This can be fooled by all 779 // manner of transformations such as DCE and other things, but seems to 780 // work well in practice. 781 if (!Devirt) 782 for (int i = 0, Size = C->size(); i < Size; ++i) 783 if (CallCounts[i].Indirect > NewCallCounts[i].Indirect && 784 CallCounts[i].Direct < NewCallCounts[i].Direct) { 785 Devirt = true; 786 break; 787 } 788 789 if (!Devirt) { 790 PA.intersect(std::move(PassPA)); 791 break; 792 } 793 794 // Otherwise, if we've already hit our max, we're done. 795 if (Iteration >= MaxIterations) { 796 DEBUG(dbgs() << "Found another devirtualization after hitting the max " 797 "number of repetitions (" 798 << MaxIterations << ") on SCC: " << *C << "\n"); 799 PA.intersect(std::move(PassPA)); 800 break; 801 } 802 803 DEBUG(dbgs() 804 << "Repeating an SCC pass after finding a devirtualization in: " 805 << *C << "\n"); 806 807 // Move over the new call counts in preparation for iterating. 808 CallCounts = std::move(NewCallCounts); 809 810 // Update the analysis manager with each run and intersect the total set 811 // of preserved analyses so we're ready to iterate. 812 AM.invalidate(*C, PassPA); 813 PA.intersect(std::move(PassPA)); 814 } 815 816 // Note that we don't add any preserved entries here unlike a more normal 817 // "pass manager" because we only handle invalidation *between* iterations, 818 // not after the last iteration. 819 return PA; 820 } 821 822 private: 823 PassT Pass; 824 int MaxIterations; 825 }; 826 827 /// \brief A function to deduce a function pass type and wrap it in the 828 /// templated adaptor. 829 template <typename PassT> 830 DevirtSCCRepeatedPass<PassT> createDevirtSCCRepeatedPass(PassT Pass, 831 int MaxIterations) { 832 return DevirtSCCRepeatedPass<PassT>(std::move(Pass), MaxIterations); 833 } 834 835 // Clear out the debug logging macro. 836 #undef DEBUG_TYPE 837 838 } // end namespace llvm 839 840 #endif // LLVM_ANALYSIS_CGSCCPASSMANAGER_H 841