1 //===- CGSCCPassManager.cpp - Managing & running CGSCC passes -------------===// 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 #include "llvm/Analysis/CGSCCPassManager.h" 10 #include "llvm/ADT/ArrayRef.h" 11 #include "llvm/ADT/Optional.h" 12 #include "llvm/ADT/STLExtras.h" 13 #include "llvm/ADT/SetVector.h" 14 #include "llvm/ADT/SmallPtrSet.h" 15 #include "llvm/ADT/SmallVector.h" 16 #include "llvm/ADT/iterator_range.h" 17 #include "llvm/Analysis/LazyCallGraph.h" 18 #include "llvm/IR/Constant.h" 19 #include "llvm/IR/InstIterator.h" 20 #include "llvm/IR/Instruction.h" 21 #include "llvm/IR/PassManager.h" 22 #include "llvm/IR/PassManagerImpl.h" 23 #include "llvm/IR/ValueHandle.h" 24 #include "llvm/Support/Casting.h" 25 #include "llvm/Support/CommandLine.h" 26 #include "llvm/Support/Debug.h" 27 #include "llvm/Support/ErrorHandling.h" 28 #include "llvm/Support/TimeProfiler.h" 29 #include "llvm/Support/raw_ostream.h" 30 #include <algorithm> 31 #include <cassert> 32 #include <iterator> 33 34 #define DEBUG_TYPE "cgscc" 35 36 using namespace llvm; 37 38 // Explicit template instantiations and specialization definitions for core 39 // template typedefs. 40 namespace llvm { 41 42 static cl::opt<bool> AbortOnMaxDevirtIterationsReached( 43 "abort-on-max-devirt-iterations-reached", 44 cl::desc("Abort when the max iterations for devirtualization CGSCC repeat " 45 "pass is reached")); 46 47 // Explicit instantiations for the core proxy templates. 48 template class AllAnalysesOn<LazyCallGraph::SCC>; 49 template class AnalysisManager<LazyCallGraph::SCC, LazyCallGraph &>; 50 template class PassManager<LazyCallGraph::SCC, CGSCCAnalysisManager, 51 LazyCallGraph &, CGSCCUpdateResult &>; 52 template class InnerAnalysisManagerProxy<CGSCCAnalysisManager, Module>; 53 template class OuterAnalysisManagerProxy<ModuleAnalysisManager, 54 LazyCallGraph::SCC, LazyCallGraph &>; 55 template class OuterAnalysisManagerProxy<CGSCCAnalysisManager, Function>; 56 57 /// Explicitly specialize the pass manager run method to handle call graph 58 /// updates. 59 template <> 60 PreservedAnalyses 61 PassManager<LazyCallGraph::SCC, CGSCCAnalysisManager, LazyCallGraph &, 62 CGSCCUpdateResult &>::run(LazyCallGraph::SCC &InitialC, 63 CGSCCAnalysisManager &AM, 64 LazyCallGraph &G, CGSCCUpdateResult &UR) { 65 // Request PassInstrumentation from analysis manager, will use it to run 66 // instrumenting callbacks for the passes later. 67 PassInstrumentation PI = 68 AM.getResult<PassInstrumentationAnalysis>(InitialC, G); 69 70 PreservedAnalyses PA = PreservedAnalyses::all(); 71 72 if (DebugLogging) 73 dbgs() << "Starting CGSCC pass manager run.\n"; 74 75 // The SCC may be refined while we are running passes over it, so set up 76 // a pointer that we can update. 77 LazyCallGraph::SCC *C = &InitialC; 78 79 // Get Function analysis manager from its proxy. 80 FunctionAnalysisManager &FAM = 81 AM.getCachedResult<FunctionAnalysisManagerCGSCCProxy>(*C)->getManager(); 82 83 for (auto &Pass : Passes) { 84 // Check the PassInstrumentation's BeforePass callbacks before running the 85 // pass, skip its execution completely if asked to (callback returns false). 86 if (!PI.runBeforePass(*Pass, *C)) 87 continue; 88 89 PreservedAnalyses PassPA; 90 { 91 TimeTraceScope TimeScope(Pass->name()); 92 PassPA = Pass->run(*C, AM, G, UR); 93 } 94 95 if (UR.InvalidatedSCCs.count(C)) 96 PI.runAfterPassInvalidated<LazyCallGraph::SCC>(*Pass, PassPA); 97 else 98 PI.runAfterPass<LazyCallGraph::SCC>(*Pass, *C, PassPA); 99 100 // Update the SCC if necessary. 101 C = UR.UpdatedC ? UR.UpdatedC : C; 102 if (UR.UpdatedC) { 103 // If C is updated, also create a proxy and update FAM inside the result. 104 auto *ResultFAMCP = 105 &AM.getResult<FunctionAnalysisManagerCGSCCProxy>(*C, G); 106 ResultFAMCP->updateFAM(FAM); 107 } 108 109 // If the CGSCC pass wasn't able to provide a valid updated SCC, the 110 // current SCC may simply need to be skipped if invalid. 111 if (UR.InvalidatedSCCs.count(C)) { 112 LLVM_DEBUG(dbgs() << "Skipping invalidated root or island SCC!\n"); 113 break; 114 } 115 // Check that we didn't miss any update scenario. 116 assert(C->begin() != C->end() && "Cannot have an empty SCC!"); 117 118 // Update the analysis manager as each pass runs and potentially 119 // invalidates analyses. 120 AM.invalidate(*C, PassPA); 121 122 // Finally, we intersect the final preserved analyses to compute the 123 // aggregate preserved set for this pass manager. 124 PA.intersect(std::move(PassPA)); 125 126 // FIXME: Historically, the pass managers all called the LLVM context's 127 // yield function here. We don't have a generic way to acquire the 128 // context and it isn't yet clear what the right pattern is for yielding 129 // in the new pass manager so it is currently omitted. 130 // ...getContext().yield(); 131 } 132 133 // Before we mark all of *this* SCC's analyses as preserved below, intersect 134 // this with the cross-SCC preserved analysis set. This is used to allow 135 // CGSCC passes to mutate ancestor SCCs and still trigger proper invalidation 136 // for them. 137 UR.CrossSCCPA.intersect(PA); 138 139 // Invalidation was handled after each pass in the above loop for the current 140 // SCC. Therefore, the remaining analysis results in the AnalysisManager are 141 // preserved. We mark this with a set so that we don't need to inspect each 142 // one individually. 143 PA.preserveSet<AllAnalysesOn<LazyCallGraph::SCC>>(); 144 145 if (DebugLogging) 146 dbgs() << "Finished CGSCC pass manager run.\n"; 147 148 return PA; 149 } 150 151 PreservedAnalyses 152 ModuleToPostOrderCGSCCPassAdaptor::run(Module &M, ModuleAnalysisManager &AM) { 153 // Setup the CGSCC analysis manager from its proxy. 154 CGSCCAnalysisManager &CGAM = 155 AM.getResult<CGSCCAnalysisManagerModuleProxy>(M).getManager(); 156 157 // Get the call graph for this module. 158 LazyCallGraph &CG = AM.getResult<LazyCallGraphAnalysis>(M); 159 160 // Get Function analysis manager from its proxy. 161 FunctionAnalysisManager &FAM = 162 AM.getCachedResult<FunctionAnalysisManagerModuleProxy>(M)->getManager(); 163 164 // We keep worklists to allow us to push more work onto the pass manager as 165 // the passes are run. 166 SmallPriorityWorklist<LazyCallGraph::RefSCC *, 1> RCWorklist; 167 SmallPriorityWorklist<LazyCallGraph::SCC *, 1> CWorklist; 168 169 // Keep sets for invalidated SCCs and RefSCCs that should be skipped when 170 // iterating off the worklists. 171 SmallPtrSet<LazyCallGraph::RefSCC *, 4> InvalidRefSCCSet; 172 SmallPtrSet<LazyCallGraph::SCC *, 4> InvalidSCCSet; 173 174 SmallDenseSet<std::pair<LazyCallGraph::Node *, LazyCallGraph::SCC *>, 4> 175 InlinedInternalEdges; 176 177 CGSCCUpdateResult UR = { 178 RCWorklist, CWorklist, InvalidRefSCCSet, InvalidSCCSet, 179 nullptr, nullptr, PreservedAnalyses::all(), InlinedInternalEdges, 180 {}}; 181 182 // Request PassInstrumentation from analysis manager, will use it to run 183 // instrumenting callbacks for the passes later. 184 PassInstrumentation PI = AM.getResult<PassInstrumentationAnalysis>(M); 185 186 PreservedAnalyses PA = PreservedAnalyses::all(); 187 CG.buildRefSCCs(); 188 for (auto RCI = CG.postorder_ref_scc_begin(), 189 RCE = CG.postorder_ref_scc_end(); 190 RCI != RCE;) { 191 assert(RCWorklist.empty() && 192 "Should always start with an empty RefSCC worklist"); 193 // The postorder_ref_sccs range we are walking is lazily constructed, so 194 // we only push the first one onto the worklist. The worklist allows us 195 // to capture *new* RefSCCs created during transformations. 196 // 197 // We really want to form RefSCCs lazily because that makes them cheaper 198 // to update as the program is simplified and allows us to have greater 199 // cache locality as forming a RefSCC touches all the parts of all the 200 // functions within that RefSCC. 201 // 202 // We also eagerly increment the iterator to the next position because 203 // the CGSCC passes below may delete the current RefSCC. 204 RCWorklist.insert(&*RCI++); 205 206 do { 207 LazyCallGraph::RefSCC *RC = RCWorklist.pop_back_val(); 208 if (InvalidRefSCCSet.count(RC)) { 209 LLVM_DEBUG(dbgs() << "Skipping an invalid RefSCC...\n"); 210 continue; 211 } 212 213 assert(CWorklist.empty() && 214 "Should always start with an empty SCC worklist"); 215 216 LLVM_DEBUG(dbgs() << "Running an SCC pass across the RefSCC: " << *RC 217 << "\n"); 218 219 // The top of the worklist may *also* be the same SCC we just ran over 220 // (and invalidated for). Keep track of that last SCC we processed due 221 // to SCC update to avoid redundant processing when an SCC is both just 222 // updated itself and at the top of the worklist. 223 LazyCallGraph::SCC *LastUpdatedC = nullptr; 224 225 // Push the initial SCCs in reverse post-order as we'll pop off the 226 // back and so see this in post-order. 227 for (LazyCallGraph::SCC &C : llvm::reverse(*RC)) 228 CWorklist.insert(&C); 229 230 do { 231 LazyCallGraph::SCC *C = CWorklist.pop_back_val(); 232 // Due to call graph mutations, we may have invalid SCCs or SCCs from 233 // other RefSCCs in the worklist. The invalid ones are dead and the 234 // other RefSCCs should be queued above, so we just need to skip both 235 // scenarios here. 236 if (InvalidSCCSet.count(C)) { 237 LLVM_DEBUG(dbgs() << "Skipping an invalid SCC...\n"); 238 continue; 239 } 240 if (LastUpdatedC == C) { 241 LLVM_DEBUG(dbgs() << "Skipping redundant run on SCC: " << *C << "\n"); 242 continue; 243 } 244 if (&C->getOuterRefSCC() != RC) { 245 LLVM_DEBUG(dbgs() << "Skipping an SCC that is now part of some other " 246 "RefSCC...\n"); 247 continue; 248 } 249 250 // Ensure we can proxy analysis updates from the CGSCC analysis manager 251 // into the the Function analysis manager by getting a proxy here. 252 // This also needs to update the FunctionAnalysisManager, as this may be 253 // the first time we see this SCC. 254 CGAM.getResult<FunctionAnalysisManagerCGSCCProxy>(*C, CG).updateFAM( 255 FAM); 256 257 // Each time we visit a new SCC pulled off the worklist, 258 // a transformation of a child SCC may have also modified this parent 259 // and invalidated analyses. So we invalidate using the update record's 260 // cross-SCC preserved set. This preserved set is intersected by any 261 // CGSCC pass that handles invalidation (primarily pass managers) prior 262 // to marking its SCC as preserved. That lets us track everything that 263 // might need invalidation across SCCs without excessive invalidations 264 // on a single SCC. 265 // 266 // This essentially allows SCC passes to freely invalidate analyses 267 // of any ancestor SCC. If this becomes detrimental to successfully 268 // caching analyses, we could force each SCC pass to manually 269 // invalidate the analyses for any SCCs other than themselves which 270 // are mutated. However, that seems to lose the robustness of the 271 // pass-manager driven invalidation scheme. 272 CGAM.invalidate(*C, UR.CrossSCCPA); 273 274 do { 275 // Check that we didn't miss any update scenario. 276 assert(!InvalidSCCSet.count(C) && "Processing an invalid SCC!"); 277 assert(C->begin() != C->end() && "Cannot have an empty SCC!"); 278 assert(&C->getOuterRefSCC() == RC && 279 "Processing an SCC in a different RefSCC!"); 280 281 LastUpdatedC = UR.UpdatedC; 282 UR.UpdatedRC = nullptr; 283 UR.UpdatedC = nullptr; 284 285 // Check the PassInstrumentation's BeforePass callbacks before 286 // running the pass, skip its execution completely if asked to 287 // (callback returns false). 288 if (!PI.runBeforePass<LazyCallGraph::SCC>(*Pass, *C)) 289 continue; 290 291 PreservedAnalyses PassPA; 292 { 293 TimeTraceScope TimeScope(Pass->name()); 294 PassPA = Pass->run(*C, CGAM, CG, UR); 295 } 296 297 if (UR.InvalidatedSCCs.count(C)) 298 PI.runAfterPassInvalidated<LazyCallGraph::SCC>(*Pass, PassPA); 299 else 300 PI.runAfterPass<LazyCallGraph::SCC>(*Pass, *C, PassPA); 301 302 // Update the SCC and RefSCC if necessary. 303 C = UR.UpdatedC ? UR.UpdatedC : C; 304 RC = UR.UpdatedRC ? UR.UpdatedRC : RC; 305 306 if (UR.UpdatedC) { 307 // If we're updating the SCC, also update the FAM inside the proxy's 308 // result. 309 CGAM.getResult<FunctionAnalysisManagerCGSCCProxy>(*C, CG).updateFAM( 310 FAM); 311 } 312 313 // If the CGSCC pass wasn't able to provide a valid updated SCC, 314 // the current SCC may simply need to be skipped if invalid. 315 if (UR.InvalidatedSCCs.count(C)) { 316 LLVM_DEBUG(dbgs() << "Skipping invalidated root or island SCC!\n"); 317 break; 318 } 319 // Check that we didn't miss any update scenario. 320 assert(C->begin() != C->end() && "Cannot have an empty SCC!"); 321 322 // We handle invalidating the CGSCC analysis manager's information 323 // for the (potentially updated) SCC here. Note that any other SCCs 324 // whose structure has changed should have been invalidated by 325 // whatever was updating the call graph. This SCC gets invalidated 326 // late as it contains the nodes that were actively being 327 // processed. 328 CGAM.invalidate(*C, PassPA); 329 330 // Then intersect the preserved set so that invalidation of module 331 // analyses will eventually occur when the module pass completes. 332 // Also intersect with the cross-SCC preserved set to capture any 333 // cross-SCC invalidation. 334 UR.CrossSCCPA.intersect(PassPA); 335 PA.intersect(std::move(PassPA)); 336 337 // The pass may have restructured the call graph and refined the 338 // current SCC and/or RefSCC. We need to update our current SCC and 339 // RefSCC pointers to follow these. Also, when the current SCC is 340 // refined, re-run the SCC pass over the newly refined SCC in order 341 // to observe the most precise SCC model available. This inherently 342 // cannot cycle excessively as it only happens when we split SCCs 343 // apart, at most converging on a DAG of single nodes. 344 // FIXME: If we ever start having RefSCC passes, we'll want to 345 // iterate there too. 346 if (UR.UpdatedC) 347 LLVM_DEBUG(dbgs() 348 << "Re-running SCC passes after a refinement of the " 349 "current SCC: " 350 << *UR.UpdatedC << "\n"); 351 352 // Note that both `C` and `RC` may at this point refer to deleted, 353 // invalid SCC and RefSCCs respectively. But we will short circuit 354 // the processing when we check them in the loop above. 355 } while (UR.UpdatedC); 356 } while (!CWorklist.empty()); 357 358 // We only need to keep internal inlined edge information within 359 // a RefSCC, clear it to save on space and let the next time we visit 360 // any of these functions have a fresh start. 361 InlinedInternalEdges.clear(); 362 } while (!RCWorklist.empty()); 363 } 364 365 // By definition we preserve the call garph, all SCC analyses, and the 366 // analysis proxies by handling them above and in any nested pass managers. 367 PA.preserveSet<AllAnalysesOn<LazyCallGraph::SCC>>(); 368 PA.preserve<LazyCallGraphAnalysis>(); 369 PA.preserve<CGSCCAnalysisManagerModuleProxy>(); 370 PA.preserve<FunctionAnalysisManagerModuleProxy>(); 371 return PA; 372 } 373 374 PreservedAnalyses DevirtSCCRepeatedPass::run(LazyCallGraph::SCC &InitialC, 375 CGSCCAnalysisManager &AM, 376 LazyCallGraph &CG, 377 CGSCCUpdateResult &UR) { 378 PreservedAnalyses PA = PreservedAnalyses::all(); 379 PassInstrumentation PI = 380 AM.getResult<PassInstrumentationAnalysis>(InitialC, CG); 381 382 // The SCC may be refined while we are running passes over it, so set up 383 // a pointer that we can update. 384 LazyCallGraph::SCC *C = &InitialC; 385 386 // Struct to track the counts of direct and indirect calls in each function 387 // of the SCC. 388 struct CallCount { 389 int Direct; 390 int Indirect; 391 }; 392 393 // Put value handles on all of the indirect calls and return the number of 394 // direct calls for each function in the SCC. 395 auto ScanSCC = [](LazyCallGraph::SCC &C, 396 SmallMapVector<Value *, WeakTrackingVH, 16> &CallHandles) { 397 assert(CallHandles.empty() && "Must start with a clear set of handles."); 398 399 SmallDenseMap<Function *, CallCount> CallCounts; 400 CallCount CountLocal = {0, 0}; 401 for (LazyCallGraph::Node &N : C) { 402 CallCount &Count = 403 CallCounts.insert(std::make_pair(&N.getFunction(), CountLocal)) 404 .first->second; 405 for (Instruction &I : instructions(N.getFunction())) 406 if (auto *CB = dyn_cast<CallBase>(&I)) { 407 if (CB->getCalledFunction()) { 408 ++Count.Direct; 409 } else { 410 ++Count.Indirect; 411 CallHandles.insert({CB, WeakTrackingVH(CB)}); 412 } 413 } 414 } 415 416 return CallCounts; 417 }; 418 419 UR.IndirectVHs.clear(); 420 // Populate the initial call handles and get the initial call counts. 421 auto CallCounts = ScanSCC(*C, UR.IndirectVHs); 422 423 for (int Iteration = 0;; ++Iteration) { 424 if (!PI.runBeforePass<LazyCallGraph::SCC>(*Pass, *C)) 425 continue; 426 427 PreservedAnalyses PassPA = Pass->run(*C, AM, CG, UR); 428 429 if (UR.InvalidatedSCCs.count(C)) 430 PI.runAfterPassInvalidated<LazyCallGraph::SCC>(*Pass, PassPA); 431 else 432 PI.runAfterPass<LazyCallGraph::SCC>(*Pass, *C, PassPA); 433 434 // If the SCC structure has changed, bail immediately and let the outer 435 // CGSCC layer handle any iteration to reflect the refined structure. 436 if (UR.UpdatedC && UR.UpdatedC != C) { 437 PA.intersect(std::move(PassPA)); 438 break; 439 } 440 441 // Check that we didn't miss any update scenario. 442 assert(!UR.InvalidatedSCCs.count(C) && "Processing an invalid SCC!"); 443 assert(C->begin() != C->end() && "Cannot have an empty SCC!"); 444 445 // Check whether any of the handles were devirtualized. 446 bool Devirt = llvm::any_of(UR.IndirectVHs, [](auto &P) -> bool { 447 if (P.second) { 448 if (CallBase *CB = dyn_cast<CallBase>(P.second)) { 449 if (CB->getCalledFunction()) { 450 LLVM_DEBUG(dbgs() << "Found devirtualized call: " << *CB << "\n"); 451 return true; 452 } 453 } 454 } 455 return false; 456 }); 457 458 // Rescan to build up a new set of handles and count how many direct 459 // calls remain. If we decide to iterate, this also sets up the input to 460 // the next iteration. 461 UR.IndirectVHs.clear(); 462 auto NewCallCounts = ScanSCC(*C, UR.IndirectVHs); 463 464 // If we haven't found an explicit devirtualization already see if we 465 // have decreased the number of indirect calls and increased the number 466 // of direct calls for any function in the SCC. This can be fooled by all 467 // manner of transformations such as DCE and other things, but seems to 468 // work well in practice. 469 if (!Devirt) 470 // Iterate over the keys in NewCallCounts, if Function also exists in 471 // CallCounts, make the check below. 472 for (auto &Pair : NewCallCounts) { 473 auto &CallCountNew = Pair.second; 474 auto CountIt = CallCounts.find(Pair.first); 475 if (CountIt != CallCounts.end()) { 476 const auto &CallCountOld = CountIt->second; 477 if (CallCountOld.Indirect > CallCountNew.Indirect && 478 CallCountOld.Direct < CallCountNew.Direct) { 479 Devirt = true; 480 break; 481 } 482 } 483 } 484 485 if (!Devirt) { 486 PA.intersect(std::move(PassPA)); 487 break; 488 } 489 490 // Otherwise, if we've already hit our max, we're done. 491 if (Iteration >= MaxIterations) { 492 maxDevirtIterationsReached(); 493 LLVM_DEBUG( 494 dbgs() << "Found another devirtualization after hitting the max " 495 "number of repetitions (" 496 << MaxIterations << ") on SCC: " << *C << "\n"); 497 PA.intersect(std::move(PassPA)); 498 break; 499 } 500 501 LLVM_DEBUG( 502 dbgs() << "Repeating an SCC pass after finding a devirtualization in: " 503 << *C << "\n"); 504 505 // Move over the new call counts in preparation for iterating. 506 CallCounts = std::move(NewCallCounts); 507 508 // Update the analysis manager with each run and intersect the total set 509 // of preserved analyses so we're ready to iterate. 510 AM.invalidate(*C, PassPA); 511 512 PA.intersect(std::move(PassPA)); 513 } 514 515 // Note that we don't add any preserved entries here unlike a more normal 516 // "pass manager" because we only handle invalidation *between* iterations, 517 // not after the last iteration. 518 return PA; 519 } 520 521 PreservedAnalyses CGSCCToFunctionPassAdaptor::run(LazyCallGraph::SCC &C, 522 CGSCCAnalysisManager &AM, 523 LazyCallGraph &CG, 524 CGSCCUpdateResult &UR) { 525 // Setup the function analysis manager from its proxy. 526 FunctionAnalysisManager &FAM = 527 AM.getResult<FunctionAnalysisManagerCGSCCProxy>(C, CG).getManager(); 528 529 SmallVector<LazyCallGraph::Node *, 4> Nodes; 530 for (LazyCallGraph::Node &N : C) 531 Nodes.push_back(&N); 532 533 // The SCC may get split while we are optimizing functions due to deleting 534 // edges. If this happens, the current SCC can shift, so keep track of 535 // a pointer we can overwrite. 536 LazyCallGraph::SCC *CurrentC = &C; 537 538 LLVM_DEBUG(dbgs() << "Running function passes across an SCC: " << C << "\n"); 539 540 PreservedAnalyses PA = PreservedAnalyses::all(); 541 for (LazyCallGraph::Node *N : Nodes) { 542 // Skip nodes from other SCCs. These may have been split out during 543 // processing. We'll eventually visit those SCCs and pick up the nodes 544 // there. 545 if (CG.lookupSCC(*N) != CurrentC) 546 continue; 547 548 Function &F = N->getFunction(); 549 550 PassInstrumentation PI = FAM.getResult<PassInstrumentationAnalysis>(F); 551 if (!PI.runBeforePass<Function>(*Pass, F)) 552 continue; 553 554 PreservedAnalyses PassPA; 555 { 556 TimeTraceScope TimeScope(Pass->name()); 557 PassPA = Pass->run(F, FAM); 558 } 559 560 PI.runAfterPass<Function>(*Pass, F, PassPA); 561 562 // We know that the function pass couldn't have invalidated any other 563 // function's analyses (that's the contract of a function pass), so 564 // directly handle the function analysis manager's invalidation here. 565 FAM.invalidate(F, PassPA); 566 567 // Then intersect the preserved set so that invalidation of module 568 // analyses will eventually occur when the module pass completes. 569 PA.intersect(std::move(PassPA)); 570 571 // If the call graph hasn't been preserved, update it based on this 572 // function pass. This may also update the current SCC to point to 573 // a smaller, more refined SCC. 574 auto PAC = PA.getChecker<LazyCallGraphAnalysis>(); 575 if (!PAC.preserved() && !PAC.preservedSet<AllAnalysesOn<Module>>()) { 576 CurrentC = &updateCGAndAnalysisManagerForFunctionPass(CG, *CurrentC, *N, 577 AM, UR, FAM); 578 assert(CG.lookupSCC(*N) == CurrentC && 579 "Current SCC not updated to the SCC containing the current node!"); 580 } 581 } 582 583 // By definition we preserve the proxy. And we preserve all analyses on 584 // Functions. This precludes *any* invalidation of function analyses by the 585 // proxy, but that's OK because we've taken care to invalidate analyses in 586 // the function analysis manager incrementally above. 587 PA.preserveSet<AllAnalysesOn<Function>>(); 588 PA.preserve<FunctionAnalysisManagerCGSCCProxy>(); 589 590 // We've also ensured that we updated the call graph along the way. 591 PA.preserve<LazyCallGraphAnalysis>(); 592 593 return PA; 594 } 595 596 bool CGSCCAnalysisManagerModuleProxy::Result::invalidate( 597 Module &M, const PreservedAnalyses &PA, 598 ModuleAnalysisManager::Invalidator &Inv) { 599 // If literally everything is preserved, we're done. 600 if (PA.areAllPreserved()) 601 return false; // This is still a valid proxy. 602 603 // If this proxy or the call graph is going to be invalidated, we also need 604 // to clear all the keys coming from that analysis. 605 // 606 // We also directly invalidate the FAM's module proxy if necessary, and if 607 // that proxy isn't preserved we can't preserve this proxy either. We rely on 608 // it to handle module -> function analysis invalidation in the face of 609 // structural changes and so if it's unavailable we conservatively clear the 610 // entire SCC layer as well rather than trying to do invalidation ourselves. 611 auto PAC = PA.getChecker<CGSCCAnalysisManagerModuleProxy>(); 612 if (!(PAC.preserved() || PAC.preservedSet<AllAnalysesOn<Module>>()) || 613 Inv.invalidate<LazyCallGraphAnalysis>(M, PA) || 614 Inv.invalidate<FunctionAnalysisManagerModuleProxy>(M, PA)) { 615 InnerAM->clear(); 616 617 // And the proxy itself should be marked as invalid so that we can observe 618 // the new call graph. This isn't strictly necessary because we cheat 619 // above, but is still useful. 620 return true; 621 } 622 623 // Directly check if the relevant set is preserved so we can short circuit 624 // invalidating SCCs below. 625 bool AreSCCAnalysesPreserved = 626 PA.allAnalysesInSetPreserved<AllAnalysesOn<LazyCallGraph::SCC>>(); 627 628 // Ok, we have a graph, so we can propagate the invalidation down into it. 629 G->buildRefSCCs(); 630 for (auto &RC : G->postorder_ref_sccs()) 631 for (auto &C : RC) { 632 Optional<PreservedAnalyses> InnerPA; 633 634 // Check to see whether the preserved set needs to be adjusted based on 635 // module-level analysis invalidation triggering deferred invalidation 636 // for this SCC. 637 if (auto *OuterProxy = 638 InnerAM->getCachedResult<ModuleAnalysisManagerCGSCCProxy>(C)) 639 for (const auto &OuterInvalidationPair : 640 OuterProxy->getOuterInvalidations()) { 641 AnalysisKey *OuterAnalysisID = OuterInvalidationPair.first; 642 const auto &InnerAnalysisIDs = OuterInvalidationPair.second; 643 if (Inv.invalidate(OuterAnalysisID, M, PA)) { 644 if (!InnerPA) 645 InnerPA = PA; 646 for (AnalysisKey *InnerAnalysisID : InnerAnalysisIDs) 647 InnerPA->abandon(InnerAnalysisID); 648 } 649 } 650 651 // Check if we needed a custom PA set. If so we'll need to run the inner 652 // invalidation. 653 if (InnerPA) { 654 InnerAM->invalidate(C, *InnerPA); 655 continue; 656 } 657 658 // Otherwise we only need to do invalidation if the original PA set didn't 659 // preserve all SCC analyses. 660 if (!AreSCCAnalysesPreserved) 661 InnerAM->invalidate(C, PA); 662 } 663 664 // Return false to indicate that this result is still a valid proxy. 665 return false; 666 } 667 668 template <> 669 CGSCCAnalysisManagerModuleProxy::Result 670 CGSCCAnalysisManagerModuleProxy::run(Module &M, ModuleAnalysisManager &AM) { 671 // Force the Function analysis manager to also be available so that it can 672 // be accessed in an SCC analysis and proxied onward to function passes. 673 // FIXME: It is pretty awkward to just drop the result here and assert that 674 // we can find it again later. 675 (void)AM.getResult<FunctionAnalysisManagerModuleProxy>(M); 676 677 return Result(*InnerAM, AM.getResult<LazyCallGraphAnalysis>(M)); 678 } 679 680 AnalysisKey FunctionAnalysisManagerCGSCCProxy::Key; 681 682 FunctionAnalysisManagerCGSCCProxy::Result 683 FunctionAnalysisManagerCGSCCProxy::run(LazyCallGraph::SCC &C, 684 CGSCCAnalysisManager &AM, 685 LazyCallGraph &CG) { 686 // Note: unconditionally getting checking that the proxy exists may get it at 687 // this point. There are cases when this is being run unnecessarily, but 688 // it is cheap and having the assertion in place is more valuable. 689 auto &MAMProxy = AM.getResult<ModuleAnalysisManagerCGSCCProxy>(C, CG); 690 Module &M = *C.begin()->getFunction().getParent(); 691 bool ProxyExists = 692 MAMProxy.cachedResultExists<FunctionAnalysisManagerModuleProxy>(M); 693 assert(ProxyExists && 694 "The CGSCC pass manager requires that the FAM module proxy is run " 695 "on the module prior to entering the CGSCC walk"); 696 (void)ProxyExists; 697 698 // We just return an empty result. The caller will use the updateFAM interface 699 // to correctly register the relevant FunctionAnalysisManager based on the 700 // context in which this proxy is run. 701 return Result(); 702 } 703 704 bool FunctionAnalysisManagerCGSCCProxy::Result::invalidate( 705 LazyCallGraph::SCC &C, const PreservedAnalyses &PA, 706 CGSCCAnalysisManager::Invalidator &Inv) { 707 // If literally everything is preserved, we're done. 708 if (PA.areAllPreserved()) 709 return false; // This is still a valid proxy. 710 711 // All updates to preserve valid results are done below, so we don't need to 712 // invalidate this proxy. 713 // 714 // Note that in order to preserve this proxy, a module pass must ensure that 715 // the FAM has been completely updated to handle the deletion of functions. 716 // Specifically, any FAM-cached results for those functions need to have been 717 // forcibly cleared. When preserved, this proxy will only invalidate results 718 // cached on functions *still in the module* at the end of the module pass. 719 auto PAC = PA.getChecker<FunctionAnalysisManagerCGSCCProxy>(); 720 if (!PAC.preserved() && !PAC.preservedSet<AllAnalysesOn<LazyCallGraph::SCC>>()) { 721 for (LazyCallGraph::Node &N : C) 722 FAM->clear(N.getFunction(), N.getFunction().getName()); 723 724 return false; 725 } 726 727 // Directly check if the relevant set is preserved. 728 bool AreFunctionAnalysesPreserved = 729 PA.allAnalysesInSetPreserved<AllAnalysesOn<Function>>(); 730 731 // Now walk all the functions to see if any inner analysis invalidation is 732 // necessary. 733 for (LazyCallGraph::Node &N : C) { 734 Function &F = N.getFunction(); 735 Optional<PreservedAnalyses> FunctionPA; 736 737 // Check to see whether the preserved set needs to be pruned based on 738 // SCC-level analysis invalidation that triggers deferred invalidation 739 // registered with the outer analysis manager proxy for this function. 740 if (auto *OuterProxy = 741 FAM->getCachedResult<CGSCCAnalysisManagerFunctionProxy>(F)) 742 for (const auto &OuterInvalidationPair : 743 OuterProxy->getOuterInvalidations()) { 744 AnalysisKey *OuterAnalysisID = OuterInvalidationPair.first; 745 const auto &InnerAnalysisIDs = OuterInvalidationPair.second; 746 if (Inv.invalidate(OuterAnalysisID, C, PA)) { 747 if (!FunctionPA) 748 FunctionPA = PA; 749 for (AnalysisKey *InnerAnalysisID : InnerAnalysisIDs) 750 FunctionPA->abandon(InnerAnalysisID); 751 } 752 } 753 754 // Check if we needed a custom PA set, and if so we'll need to run the 755 // inner invalidation. 756 if (FunctionPA) { 757 FAM->invalidate(F, *FunctionPA); 758 continue; 759 } 760 761 // Otherwise we only need to do invalidation if the original PA set didn't 762 // preserve all function analyses. 763 if (!AreFunctionAnalysesPreserved) 764 FAM->invalidate(F, PA); 765 } 766 767 // Return false to indicate that this result is still a valid proxy. 768 return false; 769 } 770 771 } // end namespace llvm 772 773 /// When a new SCC is created for the graph we first update the 774 /// FunctionAnalysisManager in the Proxy's result. 775 /// As there might be function analysis results cached for the functions now in 776 /// that SCC, two forms of updates are required. 777 /// 778 /// First, a proxy from the SCC to the FunctionAnalysisManager needs to be 779 /// created so that any subsequent invalidation events to the SCC are 780 /// propagated to the function analysis results cached for functions within it. 781 /// 782 /// Second, if any of the functions within the SCC have analysis results with 783 /// outer analysis dependencies, then those dependencies would point to the 784 /// *wrong* SCC's analysis result. We forcibly invalidate the necessary 785 /// function analyses so that they don't retain stale handles. 786 static void updateNewSCCFunctionAnalyses(LazyCallGraph::SCC &C, 787 LazyCallGraph &G, 788 CGSCCAnalysisManager &AM, 789 FunctionAnalysisManager &FAM) { 790 AM.getResult<FunctionAnalysisManagerCGSCCProxy>(C, G).updateFAM(FAM); 791 792 // Now walk the functions in this SCC and invalidate any function analysis 793 // results that might have outer dependencies on an SCC analysis. 794 for (LazyCallGraph::Node &N : C) { 795 Function &F = N.getFunction(); 796 797 auto *OuterProxy = 798 FAM.getCachedResult<CGSCCAnalysisManagerFunctionProxy>(F); 799 if (!OuterProxy) 800 // No outer analyses were queried, nothing to do. 801 continue; 802 803 // Forcibly abandon all the inner analyses with dependencies, but 804 // invalidate nothing else. 805 auto PA = PreservedAnalyses::all(); 806 for (const auto &OuterInvalidationPair : 807 OuterProxy->getOuterInvalidations()) { 808 const auto &InnerAnalysisIDs = OuterInvalidationPair.second; 809 for (AnalysisKey *InnerAnalysisID : InnerAnalysisIDs) 810 PA.abandon(InnerAnalysisID); 811 } 812 813 // Now invalidate anything we found. 814 FAM.invalidate(F, PA); 815 } 816 } 817 818 void llvm::maxDevirtIterationsReached() { 819 if (AbortOnMaxDevirtIterationsReached) 820 report_fatal_error("Max devirtualization iterations reached"); 821 } 822 823 /// Helper function to update both the \c CGSCCAnalysisManager \p AM and the \c 824 /// CGSCCPassManager's \c CGSCCUpdateResult \p UR based on a range of newly 825 /// added SCCs. 826 /// 827 /// The range of new SCCs must be in postorder already. The SCC they were split 828 /// out of must be provided as \p C. The current node being mutated and 829 /// triggering updates must be passed as \p N. 830 /// 831 /// This function returns the SCC containing \p N. This will be either \p C if 832 /// no new SCCs have been split out, or it will be the new SCC containing \p N. 833 template <typename SCCRangeT> 834 static LazyCallGraph::SCC * 835 incorporateNewSCCRange(const SCCRangeT &NewSCCRange, LazyCallGraph &G, 836 LazyCallGraph::Node &N, LazyCallGraph::SCC *C, 837 CGSCCAnalysisManager &AM, CGSCCUpdateResult &UR) { 838 using SCC = LazyCallGraph::SCC; 839 840 if (NewSCCRange.begin() == NewSCCRange.end()) 841 return C; 842 843 // Add the current SCC to the worklist as its shape has changed. 844 UR.CWorklist.insert(C); 845 LLVM_DEBUG(dbgs() << "Enqueuing the existing SCC in the worklist:" << *C 846 << "\n"); 847 848 SCC *OldC = C; 849 850 // Update the current SCC. Note that if we have new SCCs, this must actually 851 // change the SCC. 852 assert(C != &*NewSCCRange.begin() && 853 "Cannot insert new SCCs without changing current SCC!"); 854 C = &*NewSCCRange.begin(); 855 assert(G.lookupSCC(N) == C && "Failed to update current SCC!"); 856 857 // If we had a cached FAM proxy originally, we will want to create more of 858 // them for each SCC that was split off. 859 FunctionAnalysisManager *FAM = nullptr; 860 if (auto *FAMProxy = 861 AM.getCachedResult<FunctionAnalysisManagerCGSCCProxy>(*OldC)) 862 FAM = &FAMProxy->getManager(); 863 864 // We need to propagate an invalidation call to all but the newly current SCC 865 // because the outer pass manager won't do that for us after splitting them. 866 // FIXME: We should accept a PreservedAnalysis from the CG updater so that if 867 // there are preserved analysis we can avoid invalidating them here for 868 // split-off SCCs. 869 // We know however that this will preserve any FAM proxy so go ahead and mark 870 // that. 871 PreservedAnalyses PA; 872 PA.preserve<FunctionAnalysisManagerCGSCCProxy>(); 873 AM.invalidate(*OldC, PA); 874 875 // Ensure the now-current SCC's function analyses are updated. 876 if (FAM) 877 updateNewSCCFunctionAnalyses(*C, G, AM, *FAM); 878 879 for (SCC &NewC : llvm::reverse(make_range(std::next(NewSCCRange.begin()), 880 NewSCCRange.end()))) { 881 assert(C != &NewC && "No need to re-visit the current SCC!"); 882 assert(OldC != &NewC && "Already handled the original SCC!"); 883 UR.CWorklist.insert(&NewC); 884 LLVM_DEBUG(dbgs() << "Enqueuing a newly formed SCC:" << NewC << "\n"); 885 886 // Ensure new SCCs' function analyses are updated. 887 if (FAM) 888 updateNewSCCFunctionAnalyses(NewC, G, AM, *FAM); 889 890 // Also propagate a normal invalidation to the new SCC as only the current 891 // will get one from the pass manager infrastructure. 892 AM.invalidate(NewC, PA); 893 } 894 return C; 895 } 896 897 static LazyCallGraph::SCC &updateCGAndAnalysisManagerForPass( 898 LazyCallGraph &G, LazyCallGraph::SCC &InitialC, LazyCallGraph::Node &N, 899 CGSCCAnalysisManager &AM, CGSCCUpdateResult &UR, 900 FunctionAnalysisManager &FAM, bool FunctionPass) { 901 using Node = LazyCallGraph::Node; 902 using Edge = LazyCallGraph::Edge; 903 using SCC = LazyCallGraph::SCC; 904 using RefSCC = LazyCallGraph::RefSCC; 905 906 RefSCC &InitialRC = InitialC.getOuterRefSCC(); 907 SCC *C = &InitialC; 908 RefSCC *RC = &InitialRC; 909 Function &F = N.getFunction(); 910 911 // Walk the function body and build up the set of retained, promoted, and 912 // demoted edges. 913 SmallVector<Constant *, 16> Worklist; 914 SmallPtrSet<Constant *, 16> Visited; 915 SmallPtrSet<Node *, 16> RetainedEdges; 916 SmallSetVector<Node *, 4> PromotedRefTargets; 917 SmallSetVector<Node *, 4> DemotedCallTargets; 918 SmallSetVector<Node *, 4> NewCallEdges; 919 SmallSetVector<Node *, 4> NewRefEdges; 920 SmallSetVector<Node *, 4> NewNodes; 921 922 // First walk the function and handle all called functions. We do this first 923 // because if there is a single call edge, whether there are ref edges is 924 // irrelevant. 925 for (Instruction &I : instructions(F)) { 926 if (auto *CB = dyn_cast<CallBase>(&I)) { 927 if (Function *Callee = CB->getCalledFunction()) { 928 if (Visited.insert(Callee).second && !Callee->isDeclaration()) { 929 Node *CalleeN = G.lookup(*Callee); 930 if (!CalleeN) { 931 CalleeN = &G.get(*Callee); 932 NewNodes.insert(CalleeN); 933 } 934 Edge *E = N->lookup(*CalleeN); 935 assert((E || !FunctionPass) && 936 "No function transformations should introduce *new* " 937 "call edges! Any new calls should be modeled as " 938 "promoted existing ref edges!"); 939 bool Inserted = RetainedEdges.insert(CalleeN).second; 940 (void)Inserted; 941 assert(Inserted && "We should never visit a function twice."); 942 if (!E) 943 NewCallEdges.insert(CalleeN); 944 else if (!E->isCall()) 945 PromotedRefTargets.insert(CalleeN); 946 } 947 } else { 948 // We can miss devirtualization if an indirect call is created then 949 // promoted before updateCGAndAnalysisManagerForPass runs. 950 auto *Entry = UR.IndirectVHs.find(CB); 951 if (Entry == UR.IndirectVHs.end()) 952 UR.IndirectVHs.insert({CB, WeakTrackingVH(CB)}); 953 else if (!Entry->second) 954 Entry->second = WeakTrackingVH(CB); 955 } 956 } 957 } 958 959 // Now walk all references. 960 for (Instruction &I : instructions(F)) 961 for (Value *Op : I.operand_values()) 962 if (auto *OpC = dyn_cast<Constant>(Op)) 963 if (Visited.insert(OpC).second) 964 Worklist.push_back(OpC); 965 966 auto VisitRef = [&](Function &Referee) { 967 Node *RefereeN = G.lookup(Referee); 968 if (!RefereeN) { 969 RefereeN = &G.get(Referee); 970 NewNodes.insert(RefereeN); 971 } 972 Edge *E = N->lookup(*RefereeN); 973 assert((E || !FunctionPass) && 974 "No function transformations should introduce *new* ref " 975 "edges! Any new ref edges would require IPO which " 976 "function passes aren't allowed to do!"); 977 bool Inserted = RetainedEdges.insert(RefereeN).second; 978 (void)Inserted; 979 assert(Inserted && "We should never visit a function twice."); 980 if (!E) 981 NewRefEdges.insert(RefereeN); 982 else if (E->isCall()) 983 DemotedCallTargets.insert(RefereeN); 984 }; 985 LazyCallGraph::visitReferences(Worklist, Visited, VisitRef); 986 987 for (Node *NewNode : NewNodes) 988 G.initNode(*NewNode, *C); 989 990 // Handle new ref edges. 991 for (Node *RefTarget : NewRefEdges) { 992 SCC &TargetC = *G.lookupSCC(*RefTarget); 993 RefSCC &TargetRC = TargetC.getOuterRefSCC(); 994 (void)TargetRC; 995 // TODO: This only allows trivial edges to be added for now. 996 assert((RC == &TargetRC || 997 RC->isAncestorOf(TargetRC)) && "New ref edge is not trivial!"); 998 RC->insertTrivialRefEdge(N, *RefTarget); 999 } 1000 1001 // Handle new call edges. 1002 for (Node *CallTarget : NewCallEdges) { 1003 SCC &TargetC = *G.lookupSCC(*CallTarget); 1004 RefSCC &TargetRC = TargetC.getOuterRefSCC(); 1005 (void)TargetRC; 1006 // TODO: This only allows trivial edges to be added for now. 1007 assert((RC == &TargetRC || 1008 RC->isAncestorOf(TargetRC)) && "New call edge is not trivial!"); 1009 // Add a trivial ref edge to be promoted later on alongside 1010 // PromotedRefTargets. 1011 RC->insertTrivialRefEdge(N, *CallTarget); 1012 } 1013 1014 // Include synthetic reference edges to known, defined lib functions. 1015 for (auto *LibFn : G.getLibFunctions()) 1016 // While the list of lib functions doesn't have repeats, don't re-visit 1017 // anything handled above. 1018 if (!Visited.count(LibFn)) 1019 VisitRef(*LibFn); 1020 1021 // First remove all of the edges that are no longer present in this function. 1022 // The first step makes these edges uniformly ref edges and accumulates them 1023 // into a separate data structure so removal doesn't invalidate anything. 1024 SmallVector<Node *, 4> DeadTargets; 1025 for (Edge &E : *N) { 1026 if (RetainedEdges.count(&E.getNode())) 1027 continue; 1028 1029 SCC &TargetC = *G.lookupSCC(E.getNode()); 1030 RefSCC &TargetRC = TargetC.getOuterRefSCC(); 1031 if (&TargetRC == RC && E.isCall()) { 1032 if (C != &TargetC) { 1033 // For separate SCCs this is trivial. 1034 RC->switchTrivialInternalEdgeToRef(N, E.getNode()); 1035 } else { 1036 // Now update the call graph. 1037 C = incorporateNewSCCRange(RC->switchInternalEdgeToRef(N, E.getNode()), 1038 G, N, C, AM, UR); 1039 } 1040 } 1041 1042 // Now that this is ready for actual removal, put it into our list. 1043 DeadTargets.push_back(&E.getNode()); 1044 } 1045 // Remove the easy cases quickly and actually pull them out of our list. 1046 DeadTargets.erase( 1047 llvm::remove_if(DeadTargets, 1048 [&](Node *TargetN) { 1049 SCC &TargetC = *G.lookupSCC(*TargetN); 1050 RefSCC &TargetRC = TargetC.getOuterRefSCC(); 1051 1052 // We can't trivially remove internal targets, so skip 1053 // those. 1054 if (&TargetRC == RC) 1055 return false; 1056 1057 RC->removeOutgoingEdge(N, *TargetN); 1058 LLVM_DEBUG(dbgs() << "Deleting outgoing edge from '" 1059 << N << "' to '" << TargetN << "'\n"); 1060 return true; 1061 }), 1062 DeadTargets.end()); 1063 1064 // Now do a batch removal of the internal ref edges left. 1065 auto NewRefSCCs = RC->removeInternalRefEdge(N, DeadTargets); 1066 if (!NewRefSCCs.empty()) { 1067 // The old RefSCC is dead, mark it as such. 1068 UR.InvalidatedRefSCCs.insert(RC); 1069 1070 // Note that we don't bother to invalidate analyses as ref-edge 1071 // connectivity is not really observable in any way and is intended 1072 // exclusively to be used for ordering of transforms rather than for 1073 // analysis conclusions. 1074 1075 // Update RC to the "bottom". 1076 assert(G.lookupSCC(N) == C && "Changed the SCC when splitting RefSCCs!"); 1077 RC = &C->getOuterRefSCC(); 1078 assert(G.lookupRefSCC(N) == RC && "Failed to update current RefSCC!"); 1079 1080 // The RC worklist is in reverse postorder, so we enqueue the new ones in 1081 // RPO except for the one which contains the source node as that is the 1082 // "bottom" we will continue processing in the bottom-up walk. 1083 assert(NewRefSCCs.front() == RC && 1084 "New current RefSCC not first in the returned list!"); 1085 for (RefSCC *NewRC : llvm::reverse(make_range(std::next(NewRefSCCs.begin()), 1086 NewRefSCCs.end()))) { 1087 assert(NewRC != RC && "Should not encounter the current RefSCC further " 1088 "in the postorder list of new RefSCCs."); 1089 UR.RCWorklist.insert(NewRC); 1090 LLVM_DEBUG(dbgs() << "Enqueuing a new RefSCC in the update worklist: " 1091 << *NewRC << "\n"); 1092 } 1093 } 1094 1095 // Next demote all the call edges that are now ref edges. This helps make 1096 // the SCCs small which should minimize the work below as we don't want to 1097 // form cycles that this would break. 1098 for (Node *RefTarget : DemotedCallTargets) { 1099 SCC &TargetC = *G.lookupSCC(*RefTarget); 1100 RefSCC &TargetRC = TargetC.getOuterRefSCC(); 1101 1102 // The easy case is when the target RefSCC is not this RefSCC. This is 1103 // only supported when the target RefSCC is a child of this RefSCC. 1104 if (&TargetRC != RC) { 1105 assert(RC->isAncestorOf(TargetRC) && 1106 "Cannot potentially form RefSCC cycles here!"); 1107 RC->switchOutgoingEdgeToRef(N, *RefTarget); 1108 LLVM_DEBUG(dbgs() << "Switch outgoing call edge to a ref edge from '" << N 1109 << "' to '" << *RefTarget << "'\n"); 1110 continue; 1111 } 1112 1113 // We are switching an internal call edge to a ref edge. This may split up 1114 // some SCCs. 1115 if (C != &TargetC) { 1116 // For separate SCCs this is trivial. 1117 RC->switchTrivialInternalEdgeToRef(N, *RefTarget); 1118 continue; 1119 } 1120 1121 // Now update the call graph. 1122 C = incorporateNewSCCRange(RC->switchInternalEdgeToRef(N, *RefTarget), G, N, 1123 C, AM, UR); 1124 } 1125 1126 // We added a ref edge earlier for new call edges, promote those to call edges 1127 // alongside PromotedRefTargets. 1128 for (Node *E : NewCallEdges) 1129 PromotedRefTargets.insert(E); 1130 1131 // Now promote ref edges into call edges. 1132 for (Node *CallTarget : PromotedRefTargets) { 1133 SCC &TargetC = *G.lookupSCC(*CallTarget); 1134 RefSCC &TargetRC = TargetC.getOuterRefSCC(); 1135 1136 // The easy case is when the target RefSCC is not this RefSCC. This is 1137 // only supported when the target RefSCC is a child of this RefSCC. 1138 if (&TargetRC != RC) { 1139 assert(RC->isAncestorOf(TargetRC) && 1140 "Cannot potentially form RefSCC cycles here!"); 1141 RC->switchOutgoingEdgeToCall(N, *CallTarget); 1142 LLVM_DEBUG(dbgs() << "Switch outgoing ref edge to a call edge from '" << N 1143 << "' to '" << *CallTarget << "'\n"); 1144 continue; 1145 } 1146 LLVM_DEBUG(dbgs() << "Switch an internal ref edge to a call edge from '" 1147 << N << "' to '" << *CallTarget << "'\n"); 1148 1149 // Otherwise we are switching an internal ref edge to a call edge. This 1150 // may merge away some SCCs, and we add those to the UpdateResult. We also 1151 // need to make sure to update the worklist in the event SCCs have moved 1152 // before the current one in the post-order sequence 1153 bool HasFunctionAnalysisProxy = false; 1154 auto InitialSCCIndex = RC->find(*C) - RC->begin(); 1155 bool FormedCycle = RC->switchInternalEdgeToCall( 1156 N, *CallTarget, [&](ArrayRef<SCC *> MergedSCCs) { 1157 for (SCC *MergedC : MergedSCCs) { 1158 assert(MergedC != &TargetC && "Cannot merge away the target SCC!"); 1159 1160 HasFunctionAnalysisProxy |= 1161 AM.getCachedResult<FunctionAnalysisManagerCGSCCProxy>( 1162 *MergedC) != nullptr; 1163 1164 // Mark that this SCC will no longer be valid. 1165 UR.InvalidatedSCCs.insert(MergedC); 1166 1167 // FIXME: We should really do a 'clear' here to forcibly release 1168 // memory, but we don't have a good way of doing that and 1169 // preserving the function analyses. 1170 auto PA = PreservedAnalyses::allInSet<AllAnalysesOn<Function>>(); 1171 PA.preserve<FunctionAnalysisManagerCGSCCProxy>(); 1172 AM.invalidate(*MergedC, PA); 1173 } 1174 }); 1175 1176 // If we formed a cycle by creating this call, we need to update more data 1177 // structures. 1178 if (FormedCycle) { 1179 C = &TargetC; 1180 assert(G.lookupSCC(N) == C && "Failed to update current SCC!"); 1181 1182 // If one of the invalidated SCCs had a cached proxy to a function 1183 // analysis manager, we need to create a proxy in the new current SCC as 1184 // the invalidated SCCs had their functions moved. 1185 if (HasFunctionAnalysisProxy) 1186 AM.getResult<FunctionAnalysisManagerCGSCCProxy>(*C, G).updateFAM(FAM); 1187 1188 // Any analyses cached for this SCC are no longer precise as the shape 1189 // has changed by introducing this cycle. However, we have taken care to 1190 // update the proxies so it remains valide. 1191 auto PA = PreservedAnalyses::allInSet<AllAnalysesOn<Function>>(); 1192 PA.preserve<FunctionAnalysisManagerCGSCCProxy>(); 1193 AM.invalidate(*C, PA); 1194 } 1195 auto NewSCCIndex = RC->find(*C) - RC->begin(); 1196 // If we have actually moved an SCC to be topologically "below" the current 1197 // one due to merging, we will need to revisit the current SCC after 1198 // visiting those moved SCCs. 1199 // 1200 // It is critical that we *do not* revisit the current SCC unless we 1201 // actually move SCCs in the process of merging because otherwise we may 1202 // form a cycle where an SCC is split apart, merged, split, merged and so 1203 // on infinitely. 1204 if (InitialSCCIndex < NewSCCIndex) { 1205 // Put our current SCC back onto the worklist as we'll visit other SCCs 1206 // that are now definitively ordered prior to the current one in the 1207 // post-order sequence, and may end up observing more precise context to 1208 // optimize the current SCC. 1209 UR.CWorklist.insert(C); 1210 LLVM_DEBUG(dbgs() << "Enqueuing the existing SCC in the worklist: " << *C 1211 << "\n"); 1212 // Enqueue in reverse order as we pop off the back of the worklist. 1213 for (SCC &MovedC : llvm::reverse(make_range(RC->begin() + InitialSCCIndex, 1214 RC->begin() + NewSCCIndex))) { 1215 UR.CWorklist.insert(&MovedC); 1216 LLVM_DEBUG(dbgs() << "Enqueuing a newly earlier in post-order SCC: " 1217 << MovedC << "\n"); 1218 } 1219 } 1220 } 1221 1222 assert(!UR.InvalidatedSCCs.count(C) && "Invalidated the current SCC!"); 1223 assert(!UR.InvalidatedRefSCCs.count(RC) && "Invalidated the current RefSCC!"); 1224 assert(&C->getOuterRefSCC() == RC && "Current SCC not in current RefSCC!"); 1225 1226 // Record the current RefSCC and SCC for higher layers of the CGSCC pass 1227 // manager now that all the updates have been applied. 1228 if (RC != &InitialRC) 1229 UR.UpdatedRC = RC; 1230 if (C != &InitialC) 1231 UR.UpdatedC = C; 1232 1233 return *C; 1234 } 1235 1236 LazyCallGraph::SCC &llvm::updateCGAndAnalysisManagerForFunctionPass( 1237 LazyCallGraph &G, LazyCallGraph::SCC &InitialC, LazyCallGraph::Node &N, 1238 CGSCCAnalysisManager &AM, CGSCCUpdateResult &UR, 1239 FunctionAnalysisManager &FAM) { 1240 return updateCGAndAnalysisManagerForPass(G, InitialC, N, AM, UR, FAM, 1241 /* FunctionPass */ true); 1242 } 1243 LazyCallGraph::SCC &llvm::updateCGAndAnalysisManagerForCGSCCPass( 1244 LazyCallGraph &G, LazyCallGraph::SCC &InitialC, LazyCallGraph::Node &N, 1245 CGSCCAnalysisManager &AM, CGSCCUpdateResult &UR, 1246 FunctionAnalysisManager &FAM) { 1247 return updateCGAndAnalysisManagerForPass(G, InitialC, N, AM, UR, FAM, 1248 /* FunctionPass */ false); 1249 } 1250