1 //===- BasicBlockUtils.cpp - BasicBlock Utilities --------------------------==// 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 family of functions perform manipulations on basic blocks, and 10 // instructions contained within basic blocks. 11 // 12 //===----------------------------------------------------------------------===// 13 14 #include "llvm/Transforms/Utils/BasicBlockUtils.h" 15 #include "llvm/ADT/ArrayRef.h" 16 #include "llvm/ADT/SmallPtrSet.h" 17 #include "llvm/ADT/SmallVector.h" 18 #include "llvm/ADT/Twine.h" 19 #include "llvm/Analysis/CFG.h" 20 #include "llvm/Analysis/DomTreeUpdater.h" 21 #include "llvm/Analysis/LoopInfo.h" 22 #include "llvm/Analysis/MemoryDependenceAnalysis.h" 23 #include "llvm/Analysis/MemorySSAUpdater.h" 24 #include "llvm/IR/BasicBlock.h" 25 #include "llvm/IR/CFG.h" 26 #include "llvm/IR/Constants.h" 27 #include "llvm/IR/DebugInfo.h" 28 #include "llvm/IR/DebugInfoMetadata.h" 29 #include "llvm/IR/Dominators.h" 30 #include "llvm/IR/Function.h" 31 #include "llvm/IR/IRBuilder.h" 32 #include "llvm/IR/InstrTypes.h" 33 #include "llvm/IR/Instruction.h" 34 #include "llvm/IR/Instructions.h" 35 #include "llvm/IR/LLVMContext.h" 36 #include "llvm/IR/Type.h" 37 #include "llvm/IR/User.h" 38 #include "llvm/IR/Value.h" 39 #include "llvm/IR/ValueHandle.h" 40 #include "llvm/Support/Casting.h" 41 #include "llvm/Support/CommandLine.h" 42 #include "llvm/Support/Debug.h" 43 #include "llvm/Support/raw_ostream.h" 44 #include "llvm/Transforms/Utils/Local.h" 45 #include <cassert> 46 #include <cstdint> 47 #include <string> 48 #include <utility> 49 #include <vector> 50 51 using namespace llvm; 52 53 #define DEBUG_TYPE "basicblock-utils" 54 55 static cl::opt<unsigned> MaxDeoptOrUnreachableSuccessorCheckDepth( 56 "max-deopt-or-unreachable-succ-check-depth", cl::init(8), cl::Hidden, 57 cl::desc("Set the maximum path length when checking whether a basic block " 58 "is followed by a block that either has a terminating " 59 "deoptimizing call or is terminated with an unreachable")); 60 61 void llvm::detachDeadBlocks( 62 ArrayRef<BasicBlock *> BBs, 63 SmallVectorImpl<DominatorTree::UpdateType> *Updates, 64 bool KeepOneInputPHIs) { 65 for (auto *BB : BBs) { 66 // Loop through all of our successors and make sure they know that one 67 // of their predecessors is going away. 68 SmallPtrSet<BasicBlock *, 4> UniqueSuccessors; 69 for (BasicBlock *Succ : successors(BB)) { 70 Succ->removePredecessor(BB, KeepOneInputPHIs); 71 if (Updates && UniqueSuccessors.insert(Succ).second) 72 Updates->push_back({DominatorTree::Delete, BB, Succ}); 73 } 74 75 // Zap all the instructions in the block. 76 while (!BB->empty()) { 77 Instruction &I = BB->back(); 78 // If this instruction is used, replace uses with an arbitrary value. 79 // Because control flow can't get here, we don't care what we replace the 80 // value with. Note that since this block is unreachable, and all values 81 // contained within it must dominate their uses, that all uses will 82 // eventually be removed (they are themselves dead). 83 if (!I.use_empty()) 84 I.replaceAllUsesWith(PoisonValue::get(I.getType())); 85 BB->back().eraseFromParent(); 86 } 87 new UnreachableInst(BB->getContext(), BB); 88 assert(BB->size() == 1 && 89 isa<UnreachableInst>(BB->getTerminator()) && 90 "The successor list of BB isn't empty before " 91 "applying corresponding DTU updates."); 92 } 93 } 94 95 void llvm::DeleteDeadBlock(BasicBlock *BB, DomTreeUpdater *DTU, 96 bool KeepOneInputPHIs) { 97 DeleteDeadBlocks({BB}, DTU, KeepOneInputPHIs); 98 } 99 100 void llvm::DeleteDeadBlocks(ArrayRef <BasicBlock *> BBs, DomTreeUpdater *DTU, 101 bool KeepOneInputPHIs) { 102 #ifndef NDEBUG 103 // Make sure that all predecessors of each dead block is also dead. 104 SmallPtrSet<BasicBlock *, 4> Dead(llvm::from_range, BBs); 105 assert(Dead.size() == BBs.size() && "Duplicating blocks?"); 106 for (auto *BB : Dead) 107 for (BasicBlock *Pred : predecessors(BB)) 108 assert(Dead.count(Pred) && "All predecessors must be dead!"); 109 #endif 110 111 SmallVector<DominatorTree::UpdateType, 4> Updates; 112 detachDeadBlocks(BBs, DTU ? &Updates : nullptr, KeepOneInputPHIs); 113 114 if (DTU) 115 DTU->applyUpdates(Updates); 116 117 for (BasicBlock *BB : BBs) 118 if (DTU) 119 DTU->deleteBB(BB); 120 else 121 BB->eraseFromParent(); 122 } 123 124 bool llvm::EliminateUnreachableBlocks(Function &F, DomTreeUpdater *DTU, 125 bool KeepOneInputPHIs) { 126 df_iterator_default_set<BasicBlock*> Reachable; 127 128 // Mark all reachable blocks. 129 for (BasicBlock *BB : depth_first_ext(&F, Reachable)) 130 (void)BB/* Mark all reachable blocks */; 131 132 // Collect all dead blocks. 133 std::vector<BasicBlock*> DeadBlocks; 134 for (BasicBlock &BB : F) 135 if (!Reachable.count(&BB)) 136 DeadBlocks.push_back(&BB); 137 138 // Delete the dead blocks. 139 DeleteDeadBlocks(DeadBlocks, DTU, KeepOneInputPHIs); 140 141 return !DeadBlocks.empty(); 142 } 143 144 bool llvm::FoldSingleEntryPHINodes(BasicBlock *BB, 145 MemoryDependenceResults *MemDep) { 146 if (!isa<PHINode>(BB->begin())) 147 return false; 148 149 while (PHINode *PN = dyn_cast<PHINode>(BB->begin())) { 150 if (PN->getIncomingValue(0) != PN) 151 PN->replaceAllUsesWith(PN->getIncomingValue(0)); 152 else 153 PN->replaceAllUsesWith(PoisonValue::get(PN->getType())); 154 155 if (MemDep) 156 MemDep->removeInstruction(PN); // Memdep updates AA itself. 157 158 PN->eraseFromParent(); 159 } 160 return true; 161 } 162 163 bool llvm::DeleteDeadPHIs(BasicBlock *BB, const TargetLibraryInfo *TLI, 164 MemorySSAUpdater *MSSAU) { 165 // Recursively deleting a PHI may cause multiple PHIs to be deleted 166 // or RAUW'd undef, so use an array of WeakTrackingVH for the PHIs to delete. 167 SmallVector<WeakTrackingVH, 8> PHIs(llvm::make_pointer_range(BB->phis())); 168 169 bool Changed = false; 170 for (const auto &PHI : PHIs) 171 if (PHINode *PN = dyn_cast_or_null<PHINode>(PHI.operator Value *())) 172 Changed |= RecursivelyDeleteDeadPHINode(PN, TLI, MSSAU); 173 174 return Changed; 175 } 176 177 bool llvm::MergeBlockIntoPredecessor(BasicBlock *BB, DomTreeUpdater *DTU, 178 LoopInfo *LI, MemorySSAUpdater *MSSAU, 179 MemoryDependenceResults *MemDep, 180 bool PredecessorWithTwoSuccessors, 181 DominatorTree *DT) { 182 if (BB->hasAddressTaken()) 183 return false; 184 185 // Can't merge if there are multiple predecessors, or no predecessors. 186 BasicBlock *PredBB = BB->getUniquePredecessor(); 187 if (!PredBB) return false; 188 189 // Don't break self-loops. 190 if (PredBB == BB) return false; 191 192 // Don't break unwinding instructions or terminators with other side-effects. 193 Instruction *PTI = PredBB->getTerminator(); 194 if (PTI->isSpecialTerminator() || PTI->mayHaveSideEffects()) 195 return false; 196 197 // Can't merge if there are multiple distinct successors. 198 if (!PredecessorWithTwoSuccessors && PredBB->getUniqueSuccessor() != BB) 199 return false; 200 201 // Currently only allow PredBB to have two predecessors, one being BB. 202 // Update BI to branch to BB's only successor instead of BB. 203 BranchInst *PredBB_BI; 204 BasicBlock *NewSucc = nullptr; 205 unsigned FallThruPath; 206 if (PredecessorWithTwoSuccessors) { 207 if (!(PredBB_BI = dyn_cast<BranchInst>(PTI))) 208 return false; 209 BranchInst *BB_JmpI = dyn_cast<BranchInst>(BB->getTerminator()); 210 if (!BB_JmpI || !BB_JmpI->isUnconditional()) 211 return false; 212 NewSucc = BB_JmpI->getSuccessor(0); 213 FallThruPath = PredBB_BI->getSuccessor(0) == BB ? 0 : 1; 214 } 215 216 // Can't merge if there is PHI loop. 217 for (PHINode &PN : BB->phis()) 218 if (llvm::is_contained(PN.incoming_values(), &PN)) 219 return false; 220 221 LLVM_DEBUG(dbgs() << "Merging: " << BB->getName() << " into " 222 << PredBB->getName() << "\n"); 223 224 // Begin by getting rid of unneeded PHIs. 225 SmallVector<AssertingVH<Value>, 4> IncomingValues; 226 if (isa<PHINode>(BB->front())) { 227 for (PHINode &PN : BB->phis()) 228 if (!isa<PHINode>(PN.getIncomingValue(0)) || 229 cast<PHINode>(PN.getIncomingValue(0))->getParent() != BB) 230 IncomingValues.push_back(PN.getIncomingValue(0)); 231 FoldSingleEntryPHINodes(BB, MemDep); 232 } 233 234 if (DT) { 235 assert(!DTU && "cannot use both DT and DTU for updates"); 236 DomTreeNode *PredNode = DT->getNode(PredBB); 237 DomTreeNode *BBNode = DT->getNode(BB); 238 if (PredNode) { 239 assert(BBNode && "PredNode unreachable but BBNode reachable?"); 240 for (DomTreeNode *C : to_vector(BBNode->children())) 241 C->setIDom(PredNode); 242 } 243 } 244 // DTU update: Collect all the edges that exit BB. 245 // These dominator edges will be redirected from Pred. 246 std::vector<DominatorTree::UpdateType> Updates; 247 if (DTU) { 248 assert(!DT && "cannot use both DT and DTU for updates"); 249 // To avoid processing the same predecessor more than once. 250 SmallPtrSet<BasicBlock *, 8> SeenSuccs; 251 SmallPtrSet<BasicBlock *, 2> SuccsOfPredBB(llvm::from_range, 252 successors(PredBB)); 253 Updates.reserve(Updates.size() + 2 * succ_size(BB) + 1); 254 // Add insert edges first. Experimentally, for the particular case of two 255 // blocks that can be merged, with a single successor and single predecessor 256 // respectively, it is beneficial to have all insert updates first. Deleting 257 // edges first may lead to unreachable blocks, followed by inserting edges 258 // making the blocks reachable again. Such DT updates lead to high compile 259 // times. We add inserts before deletes here to reduce compile time. 260 for (BasicBlock *SuccOfBB : successors(BB)) 261 // This successor of BB may already be a PredBB's successor. 262 if (!SuccsOfPredBB.contains(SuccOfBB)) 263 if (SeenSuccs.insert(SuccOfBB).second) 264 Updates.push_back({DominatorTree::Insert, PredBB, SuccOfBB}); 265 SeenSuccs.clear(); 266 for (BasicBlock *SuccOfBB : successors(BB)) 267 if (SeenSuccs.insert(SuccOfBB).second) 268 Updates.push_back({DominatorTree::Delete, BB, SuccOfBB}); 269 Updates.push_back({DominatorTree::Delete, PredBB, BB}); 270 } 271 272 Instruction *STI = BB->getTerminator(); 273 Instruction *Start = &*BB->begin(); 274 // If there's nothing to move, mark the starting instruction as the last 275 // instruction in the block. Terminator instruction is handled separately. 276 if (Start == STI) 277 Start = PTI; 278 279 // Move all definitions in the successor to the predecessor... 280 PredBB->splice(PTI->getIterator(), BB, BB->begin(), STI->getIterator()); 281 282 if (MSSAU) 283 MSSAU->moveAllAfterMergeBlocks(BB, PredBB, Start); 284 285 // Make all PHI nodes that referred to BB now refer to Pred as their 286 // source... 287 BB->replaceAllUsesWith(PredBB); 288 289 if (PredecessorWithTwoSuccessors) { 290 // Delete the unconditional branch from BB. 291 BB->back().eraseFromParent(); 292 293 // Update branch in the predecessor. 294 PredBB_BI->setSuccessor(FallThruPath, NewSucc); 295 } else { 296 // Delete the unconditional branch from the predecessor. 297 PredBB->back().eraseFromParent(); 298 299 // Move terminator instruction. 300 BB->back().moveBeforePreserving(*PredBB, PredBB->end()); 301 302 // Terminator may be a memory accessing instruction too. 303 if (MSSAU) 304 if (MemoryUseOrDef *MUD = cast_or_null<MemoryUseOrDef>( 305 MSSAU->getMemorySSA()->getMemoryAccess(PredBB->getTerminator()))) 306 MSSAU->moveToPlace(MUD, PredBB, MemorySSA::End); 307 } 308 // Add unreachable to now empty BB. 309 new UnreachableInst(BB->getContext(), BB); 310 311 // Inherit predecessors name if it exists. 312 if (!PredBB->hasName()) 313 PredBB->takeName(BB); 314 315 if (LI) 316 LI->removeBlock(BB); 317 318 if (MemDep) 319 MemDep->invalidateCachedPredecessors(); 320 321 if (DTU) 322 DTU->applyUpdates(Updates); 323 324 if (DT) { 325 assert(succ_empty(BB) && 326 "successors should have been transferred to PredBB"); 327 DT->eraseNode(BB); 328 } 329 330 // Finally, erase the old block and update dominator info. 331 DeleteDeadBlock(BB, DTU); 332 333 return true; 334 } 335 336 bool llvm::MergeBlockSuccessorsIntoGivenBlocks( 337 SmallPtrSetImpl<BasicBlock *> &MergeBlocks, Loop *L, DomTreeUpdater *DTU, 338 LoopInfo *LI) { 339 assert(!MergeBlocks.empty() && "MergeBlocks should not be empty"); 340 341 bool BlocksHaveBeenMerged = false; 342 while (!MergeBlocks.empty()) { 343 BasicBlock *BB = *MergeBlocks.begin(); 344 BasicBlock *Dest = BB->getSingleSuccessor(); 345 if (Dest && (!L || L->contains(Dest))) { 346 BasicBlock *Fold = Dest->getUniquePredecessor(); 347 (void)Fold; 348 if (MergeBlockIntoPredecessor(Dest, DTU, LI)) { 349 assert(Fold == BB && 350 "Expecting BB to be unique predecessor of the Dest block"); 351 MergeBlocks.erase(Dest); 352 BlocksHaveBeenMerged = true; 353 } else 354 MergeBlocks.erase(BB); 355 } else 356 MergeBlocks.erase(BB); 357 } 358 return BlocksHaveBeenMerged; 359 } 360 361 /// Remove redundant instructions within sequences of consecutive dbg.value 362 /// instructions. This is done using a backward scan to keep the last dbg.value 363 /// describing a specific variable/fragment. 364 /// 365 /// BackwardScan strategy: 366 /// ---------------------- 367 /// Given a sequence of consecutive DbgValueInst like this 368 /// 369 /// dbg.value ..., "x", FragmentX1 (*) 370 /// dbg.value ..., "y", FragmentY1 371 /// dbg.value ..., "x", FragmentX2 372 /// dbg.value ..., "x", FragmentX1 (**) 373 /// 374 /// then the instruction marked with (*) can be removed (it is guaranteed to be 375 /// obsoleted by the instruction marked with (**) as the latter instruction is 376 /// describing the same variable using the same fragment info). 377 /// 378 /// Possible improvements: 379 /// - Check fully overlapping fragments and not only identical fragments. 380 static bool 381 DbgVariableRecordsRemoveRedundantDbgInstrsUsingBackwardScan(BasicBlock *BB) { 382 SmallVector<DbgVariableRecord *, 8> ToBeRemoved; 383 SmallDenseSet<DebugVariable> VariableSet; 384 for (auto &I : reverse(*BB)) { 385 for (DbgVariableRecord &DR : 386 reverse(filterDbgVars(I.getDbgRecordRange()))) { 387 DbgVariableRecord &DVR = cast<DbgVariableRecord>(DR); 388 389 DebugVariable Key(DVR.getVariable(), DVR.getExpression(), 390 DVR.getDebugLoc()->getInlinedAt()); 391 auto R = VariableSet.insert(Key); 392 // If the same variable fragment is described more than once it is enough 393 // to keep the last one (i.e. the first found since we for reverse 394 // iteration). 395 if (R.second) 396 continue; 397 398 if (DVR.isDbgAssign()) { 399 // Don't delete dbg.assign intrinsics that are linked to instructions. 400 if (!at::getAssignmentInsts(&DVR).empty()) 401 continue; 402 // Unlinked dbg.assign intrinsics can be treated like dbg.values. 403 } 404 405 ToBeRemoved.push_back(&DVR); 406 } 407 // Sequence with consecutive dbg.value instrs ended. Clear the map to 408 // restart identifying redundant instructions if case we find another 409 // dbg.value sequence. 410 VariableSet.clear(); 411 } 412 413 for (auto &DVR : ToBeRemoved) 414 DVR->eraseFromParent(); 415 416 return !ToBeRemoved.empty(); 417 } 418 419 static bool removeRedundantDbgInstrsUsingBackwardScan(BasicBlock *BB) { 420 return DbgVariableRecordsRemoveRedundantDbgInstrsUsingBackwardScan(BB); 421 } 422 423 /// Remove redundant dbg.value instructions using a forward scan. This can 424 /// remove a dbg.value instruction that is redundant due to indicating that a 425 /// variable has the same value as already being indicated by an earlier 426 /// dbg.value. 427 /// 428 /// ForwardScan strategy: 429 /// --------------------- 430 /// Given two identical dbg.value instructions, separated by a block of 431 /// instructions that isn't describing the same variable, like this 432 /// 433 /// dbg.value X1, "x", FragmentX1 (**) 434 /// <block of instructions, none being "dbg.value ..., "x", ..."> 435 /// dbg.value X1, "x", FragmentX1 (*) 436 /// 437 /// then the instruction marked with (*) can be removed. Variable "x" is already 438 /// described as being mapped to the SSA value X1. 439 /// 440 /// Possible improvements: 441 /// - Keep track of non-overlapping fragments. 442 static bool 443 DbgVariableRecordsRemoveRedundantDbgInstrsUsingForwardScan(BasicBlock *BB) { 444 SmallVector<DbgVariableRecord *, 8> ToBeRemoved; 445 SmallDenseMap<DebugVariable, 446 std::pair<SmallVector<Value *, 4>, DIExpression *>, 4> 447 VariableMap; 448 for (auto &I : *BB) { 449 for (DbgVariableRecord &DVR : filterDbgVars(I.getDbgRecordRange())) { 450 if (DVR.getType() == DbgVariableRecord::LocationType::Declare) 451 continue; 452 DebugVariable Key(DVR.getVariable(), std::nullopt, 453 DVR.getDebugLoc()->getInlinedAt()); 454 auto [VMI, Inserted] = VariableMap.try_emplace(Key); 455 // A dbg.assign with no linked instructions can be treated like a 456 // dbg.value (i.e. can be deleted). 457 bool IsDbgValueKind = 458 (!DVR.isDbgAssign() || at::getAssignmentInsts(&DVR).empty()); 459 460 // Update the map if we found a new value/expression describing the 461 // variable, or if the variable wasn't mapped already. 462 SmallVector<Value *, 4> Values(DVR.location_ops()); 463 if (Inserted || VMI->second.first != Values || 464 VMI->second.second != DVR.getExpression()) { 465 if (IsDbgValueKind) 466 VMI->second = {Values, DVR.getExpression()}; 467 else 468 VMI->second = {Values, nullptr}; 469 continue; 470 } 471 // Don't delete dbg.assign intrinsics that are linked to instructions. 472 if (!IsDbgValueKind) 473 continue; 474 // Found an identical mapping. Remember the instruction for later removal. 475 ToBeRemoved.push_back(&DVR); 476 } 477 } 478 479 for (auto *DVR : ToBeRemoved) 480 DVR->eraseFromParent(); 481 482 return !ToBeRemoved.empty(); 483 } 484 485 static bool 486 DbgVariableRecordsRemoveUndefDbgAssignsFromEntryBlock(BasicBlock *BB) { 487 assert(BB->isEntryBlock() && "expected entry block"); 488 SmallVector<DbgVariableRecord *, 8> ToBeRemoved; 489 DenseSet<DebugVariable> SeenDefForAggregate; 490 // Returns the DebugVariable for DVI with no fragment info. 491 auto GetAggregateVariable = [](const DbgVariableRecord &DVR) { 492 return DebugVariable(DVR.getVariable(), std::nullopt, 493 DVR.getDebugLoc().getInlinedAt()); 494 }; 495 496 // Remove undef dbg.assign intrinsics that are encountered before 497 // any non-undef intrinsics from the entry block. 498 for (auto &I : *BB) { 499 for (DbgVariableRecord &DVR : filterDbgVars(I.getDbgRecordRange())) { 500 if (!DVR.isDbgValue() && !DVR.isDbgAssign()) 501 continue; 502 bool IsDbgValueKind = 503 (DVR.isDbgValue() || at::getAssignmentInsts(&DVR).empty()); 504 DebugVariable Aggregate = GetAggregateVariable(DVR); 505 if (!SeenDefForAggregate.contains(Aggregate)) { 506 bool IsKill = DVR.isKillLocation() && IsDbgValueKind; 507 if (!IsKill) { 508 SeenDefForAggregate.insert(Aggregate); 509 } else if (DVR.isDbgAssign()) { 510 ToBeRemoved.push_back(&DVR); 511 } 512 } 513 } 514 } 515 516 for (DbgVariableRecord *DVR : ToBeRemoved) 517 DVR->eraseFromParent(); 518 519 return !ToBeRemoved.empty(); 520 } 521 522 static bool removeRedundantDbgInstrsUsingForwardScan(BasicBlock *BB) { 523 return DbgVariableRecordsRemoveRedundantDbgInstrsUsingForwardScan(BB); 524 } 525 526 /// Remove redundant undef dbg.assign intrinsic from an entry block using a 527 /// forward scan. 528 /// Strategy: 529 /// --------------------- 530 /// Scanning forward, delete dbg.assign intrinsics iff they are undef, not 531 /// linked to an intrinsic, and don't share an aggregate variable with a debug 532 /// intrinsic that didn't meet the criteria. In other words, undef dbg.assigns 533 /// that come before non-undef debug intrinsics for the variable are 534 /// deleted. Given: 535 /// 536 /// dbg.assign undef, "x", FragmentX1 (*) 537 /// <block of instructions, none being "dbg.value ..., "x", ..."> 538 /// dbg.value %V, "x", FragmentX2 539 /// <block of instructions, none being "dbg.value ..., "x", ..."> 540 /// dbg.assign undef, "x", FragmentX1 541 /// 542 /// then (only) the instruction marked with (*) can be removed. 543 /// Possible improvements: 544 /// - Keep track of non-overlapping fragments. 545 static bool removeUndefDbgAssignsFromEntryBlock(BasicBlock *BB) { 546 return DbgVariableRecordsRemoveUndefDbgAssignsFromEntryBlock(BB); 547 } 548 549 bool llvm::RemoveRedundantDbgInstrs(BasicBlock *BB) { 550 bool MadeChanges = false; 551 // By using the "backward scan" strategy before the "forward scan" strategy we 552 // can remove both dbg.value (2) and (3) in a situation like this: 553 // 554 // (1) dbg.value V1, "x", DIExpression() 555 // ... 556 // (2) dbg.value V2, "x", DIExpression() 557 // (3) dbg.value V1, "x", DIExpression() 558 // 559 // The backward scan will remove (2), it is made obsolete by (3). After 560 // getting (2) out of the way, the foward scan will remove (3) since "x" 561 // already is described as having the value V1 at (1). 562 MadeChanges |= removeRedundantDbgInstrsUsingBackwardScan(BB); 563 if (BB->isEntryBlock() && 564 isAssignmentTrackingEnabled(*BB->getParent()->getParent())) 565 MadeChanges |= removeUndefDbgAssignsFromEntryBlock(BB); 566 MadeChanges |= removeRedundantDbgInstrsUsingForwardScan(BB); 567 568 if (MadeChanges) 569 LLVM_DEBUG(dbgs() << "Removed redundant dbg instrs from: " 570 << BB->getName() << "\n"); 571 return MadeChanges; 572 } 573 574 void llvm::ReplaceInstWithValue(BasicBlock::iterator &BI, Value *V) { 575 Instruction &I = *BI; 576 // Replaces all of the uses of the instruction with uses of the value 577 I.replaceAllUsesWith(V); 578 579 // Make sure to propagate a name if there is one already. 580 if (I.hasName() && !V->hasName()) 581 V->takeName(&I); 582 583 // Delete the unnecessary instruction now... 584 BI = BI->eraseFromParent(); 585 } 586 587 void llvm::ReplaceInstWithInst(BasicBlock *BB, BasicBlock::iterator &BI, 588 Instruction *I) { 589 assert(I->getParent() == nullptr && 590 "ReplaceInstWithInst: Instruction already inserted into basic block!"); 591 592 // Copy debug location to newly added instruction, if it wasn't already set 593 // by the caller. 594 if (!I->getDebugLoc()) 595 I->setDebugLoc(BI->getDebugLoc()); 596 597 // Insert the new instruction into the basic block... 598 BasicBlock::iterator New = I->insertInto(BB, BI); 599 600 // Replace all uses of the old instruction, and delete it. 601 ReplaceInstWithValue(BI, I); 602 603 // Move BI back to point to the newly inserted instruction 604 BI = New; 605 } 606 607 bool llvm::IsBlockFollowedByDeoptOrUnreachable(const BasicBlock *BB) { 608 // Remember visited blocks to avoid infinite loop 609 SmallPtrSet<const BasicBlock *, 8> VisitedBlocks; 610 unsigned Depth = 0; 611 while (BB && Depth++ < MaxDeoptOrUnreachableSuccessorCheckDepth && 612 VisitedBlocks.insert(BB).second) { 613 if (isa<UnreachableInst>(BB->getTerminator()) || 614 BB->getTerminatingDeoptimizeCall()) 615 return true; 616 BB = BB->getUniqueSuccessor(); 617 } 618 return false; 619 } 620 621 void llvm::ReplaceInstWithInst(Instruction *From, Instruction *To) { 622 BasicBlock::iterator BI(From); 623 ReplaceInstWithInst(From->getParent(), BI, To); 624 } 625 626 BasicBlock *llvm::SplitEdge(BasicBlock *BB, BasicBlock *Succ, DominatorTree *DT, 627 LoopInfo *LI, MemorySSAUpdater *MSSAU, 628 const Twine &BBName) { 629 unsigned SuccNum = GetSuccessorNumber(BB, Succ); 630 631 Instruction *LatchTerm = BB->getTerminator(); 632 633 CriticalEdgeSplittingOptions Options = 634 CriticalEdgeSplittingOptions(DT, LI, MSSAU).setPreserveLCSSA(); 635 636 if ((isCriticalEdge(LatchTerm, SuccNum, Options.MergeIdenticalEdges))) { 637 // If this is a critical edge, let SplitKnownCriticalEdge do it. 638 return SplitKnownCriticalEdge(LatchTerm, SuccNum, Options, BBName); 639 } 640 641 // If the edge isn't critical, then BB has a single successor or Succ has a 642 // single pred. Split the block. 643 if (BasicBlock *SP = Succ->getSinglePredecessor()) { 644 // If the successor only has a single pred, split the top of the successor 645 // block. 646 assert(SP == BB && "CFG broken"); 647 (void)SP; 648 return SplitBlock(Succ, &Succ->front(), DT, LI, MSSAU, BBName, 649 /*Before=*/true); 650 } 651 652 // Otherwise, if BB has a single successor, split it at the bottom of the 653 // block. 654 assert(BB->getTerminator()->getNumSuccessors() == 1 && 655 "Should have a single succ!"); 656 return SplitBlock(BB, BB->getTerminator(), DT, LI, MSSAU, BBName); 657 } 658 659 void llvm::setUnwindEdgeTo(Instruction *TI, BasicBlock *Succ) { 660 if (auto *II = dyn_cast<InvokeInst>(TI)) 661 II->setUnwindDest(Succ); 662 else if (auto *CS = dyn_cast<CatchSwitchInst>(TI)) 663 CS->setUnwindDest(Succ); 664 else if (auto *CR = dyn_cast<CleanupReturnInst>(TI)) 665 CR->setUnwindDest(Succ); 666 else 667 llvm_unreachable("unexpected terminator instruction"); 668 } 669 670 void llvm::updatePhiNodes(BasicBlock *DestBB, BasicBlock *OldPred, 671 BasicBlock *NewPred, PHINode *Until) { 672 int BBIdx = 0; 673 for (PHINode &PN : DestBB->phis()) { 674 // We manually update the LandingPadReplacement PHINode and it is the last 675 // PHI Node. So, if we find it, we are done. 676 if (Until == &PN) 677 break; 678 679 // Reuse the previous value of BBIdx if it lines up. In cases where we 680 // have multiple phi nodes with *lots* of predecessors, this is a speed 681 // win because we don't have to scan the PHI looking for TIBB. This 682 // happens because the BB list of PHI nodes are usually in the same 683 // order. 684 if (PN.getIncomingBlock(BBIdx) != OldPred) 685 BBIdx = PN.getBasicBlockIndex(OldPred); 686 687 assert(BBIdx != -1 && "Invalid PHI Index!"); 688 PN.setIncomingBlock(BBIdx, NewPred); 689 } 690 } 691 692 BasicBlock *llvm::ehAwareSplitEdge(BasicBlock *BB, BasicBlock *Succ, 693 LandingPadInst *OriginalPad, 694 PHINode *LandingPadReplacement, 695 const CriticalEdgeSplittingOptions &Options, 696 const Twine &BBName) { 697 698 auto PadInst = Succ->getFirstNonPHIIt(); 699 if (!LandingPadReplacement && !PadInst->isEHPad()) 700 return SplitEdge(BB, Succ, Options.DT, Options.LI, Options.MSSAU, BBName); 701 702 auto *LI = Options.LI; 703 SmallVector<BasicBlock *, 4> LoopPreds; 704 // Check if extra modifications will be required to preserve loop-simplify 705 // form after splitting. If it would require splitting blocks with IndirectBr 706 // terminators, bail out if preserving loop-simplify form is requested. 707 if (Options.PreserveLoopSimplify && LI) { 708 if (Loop *BBLoop = LI->getLoopFor(BB)) { 709 710 // The only way that we can break LoopSimplify form by splitting a 711 // critical edge is when there exists some edge from BBLoop to Succ *and* 712 // the only edge into Succ from outside of BBLoop is that of NewBB after 713 // the split. If the first isn't true, then LoopSimplify still holds, 714 // NewBB is the new exit block and it has no non-loop predecessors. If the 715 // second isn't true, then Succ was not in LoopSimplify form prior to 716 // the split as it had a non-loop predecessor. In both of these cases, 717 // the predecessor must be directly in BBLoop, not in a subloop, or again 718 // LoopSimplify doesn't hold. 719 for (BasicBlock *P : predecessors(Succ)) { 720 if (P == BB) 721 continue; // The new block is known. 722 if (LI->getLoopFor(P) != BBLoop) { 723 // Loop is not in LoopSimplify form, no need to re simplify after 724 // splitting edge. 725 LoopPreds.clear(); 726 break; 727 } 728 LoopPreds.push_back(P); 729 } 730 // Loop-simplify form can be preserved, if we can split all in-loop 731 // predecessors. 732 if (any_of(LoopPreds, [](BasicBlock *Pred) { 733 return isa<IndirectBrInst>(Pred->getTerminator()); 734 })) { 735 return nullptr; 736 } 737 } 738 } 739 740 auto *NewBB = 741 BasicBlock::Create(BB->getContext(), BBName, BB->getParent(), Succ); 742 setUnwindEdgeTo(BB->getTerminator(), NewBB); 743 updatePhiNodes(Succ, BB, NewBB, LandingPadReplacement); 744 745 if (LandingPadReplacement) { 746 auto *NewLP = OriginalPad->clone(); 747 auto *Terminator = BranchInst::Create(Succ, NewBB); 748 NewLP->insertBefore(Terminator->getIterator()); 749 LandingPadReplacement->addIncoming(NewLP, NewBB); 750 } else { 751 Value *ParentPad = nullptr; 752 if (auto *FuncletPad = dyn_cast<FuncletPadInst>(PadInst)) 753 ParentPad = FuncletPad->getParentPad(); 754 else if (auto *CatchSwitch = dyn_cast<CatchSwitchInst>(PadInst)) 755 ParentPad = CatchSwitch->getParentPad(); 756 else if (auto *CleanupPad = dyn_cast<CleanupPadInst>(PadInst)) 757 ParentPad = CleanupPad->getParentPad(); 758 else if (auto *LandingPad = dyn_cast<LandingPadInst>(PadInst)) 759 ParentPad = LandingPad->getParent(); 760 else 761 llvm_unreachable("handling for other EHPads not implemented yet"); 762 763 auto *NewCleanupPad = CleanupPadInst::Create(ParentPad, {}, BBName, NewBB); 764 CleanupReturnInst::Create(NewCleanupPad, Succ, NewBB); 765 } 766 767 auto *DT = Options.DT; 768 auto *MSSAU = Options.MSSAU; 769 if (!DT && !LI) 770 return NewBB; 771 772 if (DT) { 773 DomTreeUpdater DTU(DT, DomTreeUpdater::UpdateStrategy::Lazy); 774 SmallVector<DominatorTree::UpdateType, 3> Updates; 775 776 Updates.push_back({DominatorTree::Insert, BB, NewBB}); 777 Updates.push_back({DominatorTree::Insert, NewBB, Succ}); 778 Updates.push_back({DominatorTree::Delete, BB, Succ}); 779 780 DTU.applyUpdates(Updates); 781 DTU.flush(); 782 783 if (MSSAU) { 784 MSSAU->applyUpdates(Updates, *DT); 785 if (VerifyMemorySSA) 786 MSSAU->getMemorySSA()->verifyMemorySSA(); 787 } 788 } 789 790 if (LI) { 791 if (Loop *BBLoop = LI->getLoopFor(BB)) { 792 // If one or the other blocks were not in a loop, the new block is not 793 // either, and thus LI doesn't need to be updated. 794 if (Loop *SuccLoop = LI->getLoopFor(Succ)) { 795 if (BBLoop == SuccLoop) { 796 // Both in the same loop, the NewBB joins loop. 797 SuccLoop->addBasicBlockToLoop(NewBB, *LI); 798 } else if (BBLoop->contains(SuccLoop)) { 799 // Edge from an outer loop to an inner loop. Add to the outer loop. 800 BBLoop->addBasicBlockToLoop(NewBB, *LI); 801 } else if (SuccLoop->contains(BBLoop)) { 802 // Edge from an inner loop to an outer loop. Add to the outer loop. 803 SuccLoop->addBasicBlockToLoop(NewBB, *LI); 804 } else { 805 // Edge from two loops with no containment relation. Because these 806 // are natural loops, we know that the destination block must be the 807 // header of its loop (adding a branch into a loop elsewhere would 808 // create an irreducible loop). 809 assert(SuccLoop->getHeader() == Succ && 810 "Should not create irreducible loops!"); 811 if (Loop *P = SuccLoop->getParentLoop()) 812 P->addBasicBlockToLoop(NewBB, *LI); 813 } 814 } 815 816 // If BB is in a loop and Succ is outside of that loop, we may need to 817 // update LoopSimplify form and LCSSA form. 818 if (!BBLoop->contains(Succ)) { 819 assert(!BBLoop->contains(NewBB) && 820 "Split point for loop exit is contained in loop!"); 821 822 // Update LCSSA form in the newly created exit block. 823 if (Options.PreserveLCSSA) { 824 createPHIsForSplitLoopExit(BB, NewBB, Succ); 825 } 826 827 if (!LoopPreds.empty()) { 828 BasicBlock *NewExitBB = SplitBlockPredecessors( 829 Succ, LoopPreds, "split", DT, LI, MSSAU, Options.PreserveLCSSA); 830 if (Options.PreserveLCSSA) 831 createPHIsForSplitLoopExit(LoopPreds, NewExitBB, Succ); 832 } 833 } 834 } 835 } 836 837 return NewBB; 838 } 839 840 void llvm::createPHIsForSplitLoopExit(ArrayRef<BasicBlock *> Preds, 841 BasicBlock *SplitBB, BasicBlock *DestBB) { 842 // SplitBB shouldn't have anything non-trivial in it yet. 843 assert((&*SplitBB->getFirstNonPHIIt() == SplitBB->getTerminator() || 844 SplitBB->isLandingPad()) && 845 "SplitBB has non-PHI nodes!"); 846 847 // For each PHI in the destination block. 848 for (PHINode &PN : DestBB->phis()) { 849 int Idx = PN.getBasicBlockIndex(SplitBB); 850 assert(Idx >= 0 && "Invalid Block Index"); 851 Value *V = PN.getIncomingValue(Idx); 852 853 // If the input is a PHI which already satisfies LCSSA, don't create 854 // a new one. 855 if (const PHINode *VP = dyn_cast<PHINode>(V)) 856 if (VP->getParent() == SplitBB) 857 continue; 858 859 // Otherwise a new PHI is needed. Create one and populate it. 860 PHINode *NewPN = PHINode::Create(PN.getType(), Preds.size(), "split"); 861 BasicBlock::iterator InsertPos = 862 SplitBB->isLandingPad() ? SplitBB->begin() 863 : SplitBB->getTerminator()->getIterator(); 864 NewPN->insertBefore(InsertPos); 865 for (BasicBlock *BB : Preds) 866 NewPN->addIncoming(V, BB); 867 868 // Update the original PHI. 869 PN.setIncomingValue(Idx, NewPN); 870 } 871 } 872 873 unsigned 874 llvm::SplitAllCriticalEdges(Function &F, 875 const CriticalEdgeSplittingOptions &Options) { 876 unsigned NumBroken = 0; 877 for (BasicBlock &BB : F) { 878 Instruction *TI = BB.getTerminator(); 879 if (TI->getNumSuccessors() > 1 && !isa<IndirectBrInst>(TI)) 880 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i) 881 if (SplitCriticalEdge(TI, i, Options)) 882 ++NumBroken; 883 } 884 return NumBroken; 885 } 886 887 static BasicBlock *SplitBlockImpl(BasicBlock *Old, BasicBlock::iterator SplitPt, 888 DomTreeUpdater *DTU, DominatorTree *DT, 889 LoopInfo *LI, MemorySSAUpdater *MSSAU, 890 const Twine &BBName, bool Before) { 891 if (Before) { 892 DomTreeUpdater LocalDTU(DT, DomTreeUpdater::UpdateStrategy::Lazy); 893 return splitBlockBefore(Old, SplitPt, 894 DTU ? DTU : (DT ? &LocalDTU : nullptr), LI, MSSAU, 895 BBName); 896 } 897 BasicBlock::iterator SplitIt = SplitPt; 898 while (isa<PHINode>(SplitIt) || SplitIt->isEHPad()) { 899 ++SplitIt; 900 assert(SplitIt != SplitPt->getParent()->end()); 901 } 902 std::string Name = BBName.str(); 903 BasicBlock *New = Old->splitBasicBlock( 904 SplitIt, Name.empty() ? Old->getName() + ".split" : Name); 905 906 // The new block lives in whichever loop the old one did. This preserves 907 // LCSSA as well, because we force the split point to be after any PHI nodes. 908 if (LI) 909 if (Loop *L = LI->getLoopFor(Old)) 910 L->addBasicBlockToLoop(New, *LI); 911 912 if (DTU) { 913 SmallVector<DominatorTree::UpdateType, 8> Updates; 914 // Old dominates New. New node dominates all other nodes dominated by Old. 915 SmallPtrSet<BasicBlock *, 8> UniqueSuccessorsOfOld; 916 Updates.push_back({DominatorTree::Insert, Old, New}); 917 Updates.reserve(Updates.size() + 2 * succ_size(New)); 918 for (BasicBlock *SuccessorOfOld : successors(New)) 919 if (UniqueSuccessorsOfOld.insert(SuccessorOfOld).second) { 920 Updates.push_back({DominatorTree::Insert, New, SuccessorOfOld}); 921 Updates.push_back({DominatorTree::Delete, Old, SuccessorOfOld}); 922 } 923 924 DTU->applyUpdates(Updates); 925 } else if (DT) 926 // Old dominates New. New node dominates all other nodes dominated by Old. 927 if (DomTreeNode *OldNode = DT->getNode(Old)) { 928 std::vector<DomTreeNode *> Children(OldNode->begin(), OldNode->end()); 929 930 DomTreeNode *NewNode = DT->addNewBlock(New, Old); 931 for (DomTreeNode *I : Children) 932 DT->changeImmediateDominator(I, NewNode); 933 } 934 935 // Move MemoryAccesses still tracked in Old, but part of New now. 936 // Update accesses in successor blocks accordingly. 937 if (MSSAU) 938 MSSAU->moveAllAfterSpliceBlocks(Old, New, &*(New->begin())); 939 940 return New; 941 } 942 943 BasicBlock *llvm::SplitBlock(BasicBlock *Old, BasicBlock::iterator SplitPt, 944 DominatorTree *DT, LoopInfo *LI, 945 MemorySSAUpdater *MSSAU, const Twine &BBName, 946 bool Before) { 947 return SplitBlockImpl(Old, SplitPt, /*DTU=*/nullptr, DT, LI, MSSAU, BBName, 948 Before); 949 } 950 BasicBlock *llvm::SplitBlock(BasicBlock *Old, BasicBlock::iterator SplitPt, 951 DomTreeUpdater *DTU, LoopInfo *LI, 952 MemorySSAUpdater *MSSAU, const Twine &BBName, 953 bool Before) { 954 return SplitBlockImpl(Old, SplitPt, DTU, /*DT=*/nullptr, LI, MSSAU, BBName, 955 Before); 956 } 957 958 BasicBlock *llvm::splitBlockBefore(BasicBlock *Old, BasicBlock::iterator SplitPt, 959 DomTreeUpdater *DTU, LoopInfo *LI, 960 MemorySSAUpdater *MSSAU, 961 const Twine &BBName) { 962 963 BasicBlock::iterator SplitIt = SplitPt; 964 while (isa<PHINode>(SplitIt) || SplitIt->isEHPad()) 965 ++SplitIt; 966 std::string Name = BBName.str(); 967 BasicBlock *New = Old->splitBasicBlock( 968 SplitIt, Name.empty() ? Old->getName() + ".split" : Name, 969 /* Before=*/true); 970 971 // The new block lives in whichever loop the old one did. This preserves 972 // LCSSA as well, because we force the split point to be after any PHI nodes. 973 if (LI) 974 if (Loop *L = LI->getLoopFor(Old)) 975 L->addBasicBlockToLoop(New, *LI); 976 977 if (DTU) { 978 SmallVector<DominatorTree::UpdateType, 8> DTUpdates; 979 // New dominates Old. The predecessor nodes of the Old node dominate 980 // New node. 981 SmallPtrSet<BasicBlock *, 8> UniquePredecessorsOfOld; 982 DTUpdates.push_back({DominatorTree::Insert, New, Old}); 983 DTUpdates.reserve(DTUpdates.size() + 2 * pred_size(New)); 984 for (BasicBlock *PredecessorOfOld : predecessors(New)) 985 if (UniquePredecessorsOfOld.insert(PredecessorOfOld).second) { 986 DTUpdates.push_back({DominatorTree::Insert, PredecessorOfOld, New}); 987 DTUpdates.push_back({DominatorTree::Delete, PredecessorOfOld, Old}); 988 } 989 990 DTU->applyUpdates(DTUpdates); 991 992 // Move MemoryAccesses still tracked in Old, but part of New now. 993 // Update accesses in successor blocks accordingly. 994 if (MSSAU) { 995 MSSAU->applyUpdates(DTUpdates, DTU->getDomTree()); 996 if (VerifyMemorySSA) 997 MSSAU->getMemorySSA()->verifyMemorySSA(); 998 } 999 } 1000 return New; 1001 } 1002 1003 /// Update DominatorTree, LoopInfo, and LCCSA analysis information. 1004 /// Invalidates DFS Numbering when DTU or DT is provided. 1005 static void UpdateAnalysisInformation(BasicBlock *OldBB, BasicBlock *NewBB, 1006 ArrayRef<BasicBlock *> Preds, 1007 DomTreeUpdater *DTU, DominatorTree *DT, 1008 LoopInfo *LI, MemorySSAUpdater *MSSAU, 1009 bool PreserveLCSSA, bool &HasLoopExit) { 1010 // Update dominator tree if available. 1011 if (DTU) { 1012 // Recalculation of DomTree is needed when updating a forward DomTree and 1013 // the Entry BB is replaced. 1014 if (NewBB->isEntryBlock() && DTU->hasDomTree()) { 1015 // The entry block was removed and there is no external interface for 1016 // the dominator tree to be notified of this change. In this corner-case 1017 // we recalculate the entire tree. 1018 DTU->recalculate(*NewBB->getParent()); 1019 } else { 1020 // Split block expects NewBB to have a non-empty set of predecessors. 1021 SmallVector<DominatorTree::UpdateType, 8> Updates; 1022 SmallPtrSet<BasicBlock *, 8> UniquePreds; 1023 Updates.push_back({DominatorTree::Insert, NewBB, OldBB}); 1024 Updates.reserve(Updates.size() + 2 * Preds.size()); 1025 for (auto *Pred : Preds) 1026 if (UniquePreds.insert(Pred).second) { 1027 Updates.push_back({DominatorTree::Insert, Pred, NewBB}); 1028 Updates.push_back({DominatorTree::Delete, Pred, OldBB}); 1029 } 1030 DTU->applyUpdates(Updates); 1031 } 1032 } else if (DT) { 1033 if (OldBB == DT->getRootNode()->getBlock()) { 1034 assert(NewBB->isEntryBlock()); 1035 DT->setNewRoot(NewBB); 1036 } else { 1037 // Split block expects NewBB to have a non-empty set of predecessors. 1038 DT->splitBlock(NewBB); 1039 } 1040 } 1041 1042 // Update MemoryPhis after split if MemorySSA is available 1043 if (MSSAU) 1044 MSSAU->wireOldPredecessorsToNewImmediatePredecessor(OldBB, NewBB, Preds); 1045 1046 // The rest of the logic is only relevant for updating the loop structures. 1047 if (!LI) 1048 return; 1049 1050 if (DTU && DTU->hasDomTree()) 1051 DT = &DTU->getDomTree(); 1052 assert(DT && "DT should be available to update LoopInfo!"); 1053 Loop *L = LI->getLoopFor(OldBB); 1054 1055 // If we need to preserve loop analyses, collect some information about how 1056 // this split will affect loops. 1057 bool IsLoopEntry = !!L; 1058 bool SplitMakesNewLoopHeader = false; 1059 for (BasicBlock *Pred : Preds) { 1060 // Preds that are not reachable from entry should not be used to identify if 1061 // OldBB is a loop entry or if SplitMakesNewLoopHeader. Unreachable blocks 1062 // are not within any loops, so we incorrectly mark SplitMakesNewLoopHeader 1063 // as true and make the NewBB the header of some loop. This breaks LI. 1064 if (!DT->isReachableFromEntry(Pred)) 1065 continue; 1066 // If we need to preserve LCSSA, determine if any of the preds is a loop 1067 // exit. 1068 if (PreserveLCSSA) 1069 if (Loop *PL = LI->getLoopFor(Pred)) 1070 if (!PL->contains(OldBB)) 1071 HasLoopExit = true; 1072 1073 // If we need to preserve LoopInfo, note whether any of the preds crosses 1074 // an interesting loop boundary. 1075 if (!L) 1076 continue; 1077 if (L->contains(Pred)) 1078 IsLoopEntry = false; 1079 else 1080 SplitMakesNewLoopHeader = true; 1081 } 1082 1083 // Unless we have a loop for OldBB, nothing else to do here. 1084 if (!L) 1085 return; 1086 1087 if (IsLoopEntry) { 1088 // Add the new block to the nearest enclosing loop (and not an adjacent 1089 // loop). To find this, examine each of the predecessors and determine which 1090 // loops enclose them, and select the most-nested loop which contains the 1091 // loop containing the block being split. 1092 Loop *InnermostPredLoop = nullptr; 1093 for (BasicBlock *Pred : Preds) { 1094 if (Loop *PredLoop = LI->getLoopFor(Pred)) { 1095 // Seek a loop which actually contains the block being split (to avoid 1096 // adjacent loops). 1097 while (PredLoop && !PredLoop->contains(OldBB)) 1098 PredLoop = PredLoop->getParentLoop(); 1099 1100 // Select the most-nested of these loops which contains the block. 1101 if (PredLoop && PredLoop->contains(OldBB) && 1102 (!InnermostPredLoop || 1103 InnermostPredLoop->getLoopDepth() < PredLoop->getLoopDepth())) 1104 InnermostPredLoop = PredLoop; 1105 } 1106 } 1107 1108 if (InnermostPredLoop) 1109 InnermostPredLoop->addBasicBlockToLoop(NewBB, *LI); 1110 } else { 1111 L->addBasicBlockToLoop(NewBB, *LI); 1112 if (SplitMakesNewLoopHeader) 1113 L->moveToHeader(NewBB); 1114 } 1115 } 1116 1117 /// Update the PHI nodes in OrigBB to include the values coming from NewBB. 1118 /// This also updates AliasAnalysis, if available. 1119 static void UpdatePHINodes(BasicBlock *OrigBB, BasicBlock *NewBB, 1120 ArrayRef<BasicBlock *> Preds, BranchInst *BI, 1121 bool HasLoopExit) { 1122 // Otherwise, create a new PHI node in NewBB for each PHI node in OrigBB. 1123 SmallPtrSet<BasicBlock *, 16> PredSet(llvm::from_range, Preds); 1124 for (BasicBlock::iterator I = OrigBB->begin(); isa<PHINode>(I); ) { 1125 PHINode *PN = cast<PHINode>(I++); 1126 1127 // Check to see if all of the values coming in are the same. If so, we 1128 // don't need to create a new PHI node, unless it's needed for LCSSA. 1129 Value *InVal = nullptr; 1130 if (!HasLoopExit) { 1131 InVal = PN->getIncomingValueForBlock(Preds[0]); 1132 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 1133 if (!PredSet.count(PN->getIncomingBlock(i))) 1134 continue; 1135 if (!InVal) 1136 InVal = PN->getIncomingValue(i); 1137 else if (InVal != PN->getIncomingValue(i)) { 1138 InVal = nullptr; 1139 break; 1140 } 1141 } 1142 } 1143 1144 if (InVal) { 1145 // If all incoming values for the new PHI would be the same, just don't 1146 // make a new PHI. Instead, just remove the incoming values from the old 1147 // PHI. 1148 PN->removeIncomingValueIf( 1149 [&](unsigned Idx) { 1150 return PredSet.contains(PN->getIncomingBlock(Idx)); 1151 }, 1152 /* DeletePHIIfEmpty */ false); 1153 1154 // Add an incoming value to the PHI node in the loop for the preheader 1155 // edge. 1156 PN->addIncoming(InVal, NewBB); 1157 continue; 1158 } 1159 1160 // If the values coming into the block are not the same, we need a new 1161 // PHI. 1162 // Create the new PHI node, insert it into NewBB at the end of the block 1163 PHINode *NewPHI = 1164 PHINode::Create(PN->getType(), Preds.size(), PN->getName() + ".ph", BI->getIterator()); 1165 1166 // NOTE! This loop walks backwards for a reason! First off, this minimizes 1167 // the cost of removal if we end up removing a large number of values, and 1168 // second off, this ensures that the indices for the incoming values aren't 1169 // invalidated when we remove one. 1170 for (int64_t i = PN->getNumIncomingValues() - 1; i >= 0; --i) { 1171 BasicBlock *IncomingBB = PN->getIncomingBlock(i); 1172 if (PredSet.count(IncomingBB)) { 1173 Value *V = PN->removeIncomingValue(i, false); 1174 NewPHI->addIncoming(V, IncomingBB); 1175 } 1176 } 1177 1178 PN->addIncoming(NewPHI, NewBB); 1179 } 1180 } 1181 1182 static void SplitLandingPadPredecessorsImpl( 1183 BasicBlock *OrigBB, ArrayRef<BasicBlock *> Preds, const char *Suffix1, 1184 const char *Suffix2, SmallVectorImpl<BasicBlock *> &NewBBs, 1185 DomTreeUpdater *DTU, DominatorTree *DT, LoopInfo *LI, 1186 MemorySSAUpdater *MSSAU, bool PreserveLCSSA); 1187 1188 static BasicBlock * 1189 SplitBlockPredecessorsImpl(BasicBlock *BB, ArrayRef<BasicBlock *> Preds, 1190 const char *Suffix, DomTreeUpdater *DTU, 1191 DominatorTree *DT, LoopInfo *LI, 1192 MemorySSAUpdater *MSSAU, bool PreserveLCSSA) { 1193 // Do not attempt to split that which cannot be split. 1194 if (!BB->canSplitPredecessors()) 1195 return nullptr; 1196 1197 // For the landingpads we need to act a bit differently. 1198 // Delegate this work to the SplitLandingPadPredecessors. 1199 if (BB->isLandingPad()) { 1200 SmallVector<BasicBlock*, 2> NewBBs; 1201 std::string NewName = std::string(Suffix) + ".split-lp"; 1202 1203 SplitLandingPadPredecessorsImpl(BB, Preds, Suffix, NewName.c_str(), NewBBs, 1204 DTU, DT, LI, MSSAU, PreserveLCSSA); 1205 return NewBBs[0]; 1206 } 1207 1208 // Create new basic block, insert right before the original block. 1209 BasicBlock *NewBB = BasicBlock::Create( 1210 BB->getContext(), BB->getName() + Suffix, BB->getParent(), BB); 1211 1212 // The new block unconditionally branches to the old block. 1213 BranchInst *BI = BranchInst::Create(BB, NewBB); 1214 1215 Loop *L = nullptr; 1216 BasicBlock *OldLatch = nullptr; 1217 // Splitting the predecessors of a loop header creates a preheader block. 1218 if (LI && LI->isLoopHeader(BB)) { 1219 L = LI->getLoopFor(BB); 1220 // Using the loop start line number prevents debuggers stepping into the 1221 // loop body for this instruction. 1222 BI->setDebugLoc(L->getStartLoc()); 1223 1224 // If BB is the header of the Loop, it is possible that the loop is 1225 // modified, such that the current latch does not remain the latch of the 1226 // loop. If that is the case, the loop metadata from the current latch needs 1227 // to be applied to the new latch. 1228 OldLatch = L->getLoopLatch(); 1229 } else 1230 BI->setDebugLoc(BB->getFirstNonPHIOrDbg()->getDebugLoc()); 1231 1232 // Move the edges from Preds to point to NewBB instead of BB. 1233 for (BasicBlock *Pred : Preds) { 1234 // This is slightly more strict than necessary; the minimum requirement 1235 // is that there be no more than one indirectbr branching to BB. And 1236 // all BlockAddress uses would need to be updated. 1237 assert(!isa<IndirectBrInst>(Pred->getTerminator()) && 1238 "Cannot split an edge from an IndirectBrInst"); 1239 Pred->getTerminator()->replaceSuccessorWith(BB, NewBB); 1240 } 1241 1242 // Insert a new PHI node into NewBB for every PHI node in BB and that new PHI 1243 // node becomes an incoming value for BB's phi node. However, if the Preds 1244 // list is empty, we need to insert dummy entries into the PHI nodes in BB to 1245 // account for the newly created predecessor. 1246 if (Preds.empty()) { 1247 // Insert dummy values as the incoming value. 1248 for (BasicBlock::iterator I = BB->begin(); isa<PHINode>(I); ++I) 1249 cast<PHINode>(I)->addIncoming(PoisonValue::get(I->getType()), NewBB); 1250 } 1251 1252 // Update DominatorTree, LoopInfo, and LCCSA analysis information. 1253 bool HasLoopExit = false; 1254 UpdateAnalysisInformation(BB, NewBB, Preds, DTU, DT, LI, MSSAU, PreserveLCSSA, 1255 HasLoopExit); 1256 1257 if (!Preds.empty()) { 1258 // Update the PHI nodes in BB with the values coming from NewBB. 1259 UpdatePHINodes(BB, NewBB, Preds, BI, HasLoopExit); 1260 } 1261 1262 if (OldLatch) { 1263 BasicBlock *NewLatch = L->getLoopLatch(); 1264 if (NewLatch != OldLatch) { 1265 MDNode *MD = OldLatch->getTerminator()->getMetadata(LLVMContext::MD_loop); 1266 NewLatch->getTerminator()->setMetadata(LLVMContext::MD_loop, MD); 1267 // It's still possible that OldLatch is the latch of another inner loop, 1268 // in which case we do not remove the metadata. 1269 Loop *IL = LI->getLoopFor(OldLatch); 1270 if (IL && IL->getLoopLatch() != OldLatch) 1271 OldLatch->getTerminator()->setMetadata(LLVMContext::MD_loop, nullptr); 1272 } 1273 } 1274 1275 return NewBB; 1276 } 1277 1278 BasicBlock *llvm::SplitBlockPredecessors(BasicBlock *BB, 1279 ArrayRef<BasicBlock *> Preds, 1280 const char *Suffix, DominatorTree *DT, 1281 LoopInfo *LI, MemorySSAUpdater *MSSAU, 1282 bool PreserveLCSSA) { 1283 return SplitBlockPredecessorsImpl(BB, Preds, Suffix, /*DTU=*/nullptr, DT, LI, 1284 MSSAU, PreserveLCSSA); 1285 } 1286 BasicBlock *llvm::SplitBlockPredecessors(BasicBlock *BB, 1287 ArrayRef<BasicBlock *> Preds, 1288 const char *Suffix, 1289 DomTreeUpdater *DTU, LoopInfo *LI, 1290 MemorySSAUpdater *MSSAU, 1291 bool PreserveLCSSA) { 1292 return SplitBlockPredecessorsImpl(BB, Preds, Suffix, DTU, 1293 /*DT=*/nullptr, LI, MSSAU, PreserveLCSSA); 1294 } 1295 1296 static void SplitLandingPadPredecessorsImpl( 1297 BasicBlock *OrigBB, ArrayRef<BasicBlock *> Preds, const char *Suffix1, 1298 const char *Suffix2, SmallVectorImpl<BasicBlock *> &NewBBs, 1299 DomTreeUpdater *DTU, DominatorTree *DT, LoopInfo *LI, 1300 MemorySSAUpdater *MSSAU, bool PreserveLCSSA) { 1301 assert(OrigBB->isLandingPad() && "Trying to split a non-landing pad!"); 1302 1303 // Create a new basic block for OrigBB's predecessors listed in Preds. Insert 1304 // it right before the original block. 1305 BasicBlock *NewBB1 = BasicBlock::Create(OrigBB->getContext(), 1306 OrigBB->getName() + Suffix1, 1307 OrigBB->getParent(), OrigBB); 1308 NewBBs.push_back(NewBB1); 1309 1310 // The new block unconditionally branches to the old block. 1311 BranchInst *BI1 = BranchInst::Create(OrigBB, NewBB1); 1312 BI1->setDebugLoc(OrigBB->getFirstNonPHIIt()->getDebugLoc()); 1313 1314 // Move the edges from Preds to point to NewBB1 instead of OrigBB. 1315 for (BasicBlock *Pred : Preds) { 1316 // This is slightly more strict than necessary; the minimum requirement 1317 // is that there be no more than one indirectbr branching to BB. And 1318 // all BlockAddress uses would need to be updated. 1319 assert(!isa<IndirectBrInst>(Pred->getTerminator()) && 1320 "Cannot split an edge from an IndirectBrInst"); 1321 Pred->getTerminator()->replaceUsesOfWith(OrigBB, NewBB1); 1322 } 1323 1324 bool HasLoopExit = false; 1325 UpdateAnalysisInformation(OrigBB, NewBB1, Preds, DTU, DT, LI, MSSAU, 1326 PreserveLCSSA, HasLoopExit); 1327 1328 // Update the PHI nodes in OrigBB with the values coming from NewBB1. 1329 UpdatePHINodes(OrigBB, NewBB1, Preds, BI1, HasLoopExit); 1330 1331 // Move the remaining edges from OrigBB to point to NewBB2. 1332 SmallVector<BasicBlock*, 8> NewBB2Preds; 1333 for (pred_iterator i = pred_begin(OrigBB), e = pred_end(OrigBB); 1334 i != e; ) { 1335 BasicBlock *Pred = *i++; 1336 if (Pred == NewBB1) continue; 1337 assert(!isa<IndirectBrInst>(Pred->getTerminator()) && 1338 "Cannot split an edge from an IndirectBrInst"); 1339 NewBB2Preds.push_back(Pred); 1340 e = pred_end(OrigBB); 1341 } 1342 1343 BasicBlock *NewBB2 = nullptr; 1344 if (!NewBB2Preds.empty()) { 1345 // Create another basic block for the rest of OrigBB's predecessors. 1346 NewBB2 = BasicBlock::Create(OrigBB->getContext(), 1347 OrigBB->getName() + Suffix2, 1348 OrigBB->getParent(), OrigBB); 1349 NewBBs.push_back(NewBB2); 1350 1351 // The new block unconditionally branches to the old block. 1352 BranchInst *BI2 = BranchInst::Create(OrigBB, NewBB2); 1353 BI2->setDebugLoc(OrigBB->getFirstNonPHIIt()->getDebugLoc()); 1354 1355 // Move the remaining edges from OrigBB to point to NewBB2. 1356 for (BasicBlock *NewBB2Pred : NewBB2Preds) 1357 NewBB2Pred->getTerminator()->replaceUsesOfWith(OrigBB, NewBB2); 1358 1359 // Update DominatorTree, LoopInfo, and LCCSA analysis information. 1360 HasLoopExit = false; 1361 UpdateAnalysisInformation(OrigBB, NewBB2, NewBB2Preds, DTU, DT, LI, MSSAU, 1362 PreserveLCSSA, HasLoopExit); 1363 1364 // Update the PHI nodes in OrigBB with the values coming from NewBB2. 1365 UpdatePHINodes(OrigBB, NewBB2, NewBB2Preds, BI2, HasLoopExit); 1366 } 1367 1368 LandingPadInst *LPad = OrigBB->getLandingPadInst(); 1369 Instruction *Clone1 = LPad->clone(); 1370 Clone1->setName(Twine("lpad") + Suffix1); 1371 Clone1->insertInto(NewBB1, NewBB1->getFirstInsertionPt()); 1372 1373 if (NewBB2) { 1374 Instruction *Clone2 = LPad->clone(); 1375 Clone2->setName(Twine("lpad") + Suffix2); 1376 Clone2->insertInto(NewBB2, NewBB2->getFirstInsertionPt()); 1377 1378 // Create a PHI node for the two cloned landingpad instructions only 1379 // if the original landingpad instruction has some uses. 1380 if (!LPad->use_empty()) { 1381 assert(!LPad->getType()->isTokenTy() && 1382 "Split cannot be applied if LPad is token type. Otherwise an " 1383 "invalid PHINode of token type would be created."); 1384 PHINode *PN = PHINode::Create(LPad->getType(), 2, "lpad.phi", LPad->getIterator()); 1385 PN->addIncoming(Clone1, NewBB1); 1386 PN->addIncoming(Clone2, NewBB2); 1387 LPad->replaceAllUsesWith(PN); 1388 } 1389 LPad->eraseFromParent(); 1390 } else { 1391 // There is no second clone. Just replace the landing pad with the first 1392 // clone. 1393 LPad->replaceAllUsesWith(Clone1); 1394 LPad->eraseFromParent(); 1395 } 1396 } 1397 1398 void llvm::SplitLandingPadPredecessors(BasicBlock *OrigBB, 1399 ArrayRef<BasicBlock *> Preds, 1400 const char *Suffix1, const char *Suffix2, 1401 SmallVectorImpl<BasicBlock *> &NewBBs, 1402 DomTreeUpdater *DTU, LoopInfo *LI, 1403 MemorySSAUpdater *MSSAU, 1404 bool PreserveLCSSA) { 1405 return SplitLandingPadPredecessorsImpl(OrigBB, Preds, Suffix1, Suffix2, 1406 NewBBs, DTU, /*DT=*/nullptr, LI, MSSAU, 1407 PreserveLCSSA); 1408 } 1409 1410 ReturnInst *llvm::FoldReturnIntoUncondBranch(ReturnInst *RI, BasicBlock *BB, 1411 BasicBlock *Pred, 1412 DomTreeUpdater *DTU) { 1413 Instruction *UncondBranch = Pred->getTerminator(); 1414 // Clone the return and add it to the end of the predecessor. 1415 Instruction *NewRet = RI->clone(); 1416 NewRet->insertInto(Pred, Pred->end()); 1417 1418 // If the return instruction returns a value, and if the value was a 1419 // PHI node in "BB", propagate the right value into the return. 1420 for (Use &Op : NewRet->operands()) { 1421 Value *V = Op; 1422 Instruction *NewBC = nullptr; 1423 if (BitCastInst *BCI = dyn_cast<BitCastInst>(V)) { 1424 // Return value might be bitcasted. Clone and insert it before the 1425 // return instruction. 1426 V = BCI->getOperand(0); 1427 NewBC = BCI->clone(); 1428 NewBC->insertInto(Pred, NewRet->getIterator()); 1429 Op = NewBC; 1430 } 1431 1432 Instruction *NewEV = nullptr; 1433 if (ExtractValueInst *EVI = dyn_cast<ExtractValueInst>(V)) { 1434 V = EVI->getOperand(0); 1435 NewEV = EVI->clone(); 1436 if (NewBC) { 1437 NewBC->setOperand(0, NewEV); 1438 NewEV->insertInto(Pred, NewBC->getIterator()); 1439 } else { 1440 NewEV->insertInto(Pred, NewRet->getIterator()); 1441 Op = NewEV; 1442 } 1443 } 1444 1445 if (PHINode *PN = dyn_cast<PHINode>(V)) { 1446 if (PN->getParent() == BB) { 1447 if (NewEV) { 1448 NewEV->setOperand(0, PN->getIncomingValueForBlock(Pred)); 1449 } else if (NewBC) 1450 NewBC->setOperand(0, PN->getIncomingValueForBlock(Pred)); 1451 else 1452 Op = PN->getIncomingValueForBlock(Pred); 1453 } 1454 } 1455 } 1456 1457 // Update any PHI nodes in the returning block to realize that we no 1458 // longer branch to them. 1459 BB->removePredecessor(Pred); 1460 UncondBranch->eraseFromParent(); 1461 1462 if (DTU) 1463 DTU->applyUpdates({{DominatorTree::Delete, Pred, BB}}); 1464 1465 return cast<ReturnInst>(NewRet); 1466 } 1467 1468 Instruction *llvm::SplitBlockAndInsertIfThen(Value *Cond, 1469 BasicBlock::iterator SplitBefore, 1470 bool Unreachable, 1471 MDNode *BranchWeights, 1472 DomTreeUpdater *DTU, LoopInfo *LI, 1473 BasicBlock *ThenBlock) { 1474 SplitBlockAndInsertIfThenElse( 1475 Cond, SplitBefore, &ThenBlock, /* ElseBlock */ nullptr, 1476 /* UnreachableThen */ Unreachable, 1477 /* UnreachableElse */ false, BranchWeights, DTU, LI); 1478 return ThenBlock->getTerminator(); 1479 } 1480 1481 Instruction *llvm::SplitBlockAndInsertIfElse(Value *Cond, 1482 BasicBlock::iterator SplitBefore, 1483 bool Unreachable, 1484 MDNode *BranchWeights, 1485 DomTreeUpdater *DTU, LoopInfo *LI, 1486 BasicBlock *ElseBlock) { 1487 SplitBlockAndInsertIfThenElse( 1488 Cond, SplitBefore, /* ThenBlock */ nullptr, &ElseBlock, 1489 /* UnreachableThen */ false, 1490 /* UnreachableElse */ Unreachable, BranchWeights, DTU, LI); 1491 return ElseBlock->getTerminator(); 1492 } 1493 1494 void llvm::SplitBlockAndInsertIfThenElse(Value *Cond, BasicBlock::iterator SplitBefore, 1495 Instruction **ThenTerm, 1496 Instruction **ElseTerm, 1497 MDNode *BranchWeights, 1498 DomTreeUpdater *DTU, LoopInfo *LI) { 1499 BasicBlock *ThenBlock = nullptr; 1500 BasicBlock *ElseBlock = nullptr; 1501 SplitBlockAndInsertIfThenElse( 1502 Cond, SplitBefore, &ThenBlock, &ElseBlock, /* UnreachableThen */ false, 1503 /* UnreachableElse */ false, BranchWeights, DTU, LI); 1504 1505 *ThenTerm = ThenBlock->getTerminator(); 1506 *ElseTerm = ElseBlock->getTerminator(); 1507 } 1508 1509 void llvm::SplitBlockAndInsertIfThenElse( 1510 Value *Cond, BasicBlock::iterator SplitBefore, BasicBlock **ThenBlock, 1511 BasicBlock **ElseBlock, bool UnreachableThen, bool UnreachableElse, 1512 MDNode *BranchWeights, DomTreeUpdater *DTU, LoopInfo *LI) { 1513 assert((ThenBlock || ElseBlock) && 1514 "At least one branch block must be created"); 1515 assert((!UnreachableThen || !UnreachableElse) && 1516 "Split block tail must be reachable"); 1517 1518 SmallVector<DominatorTree::UpdateType, 8> Updates; 1519 SmallPtrSet<BasicBlock *, 8> UniqueOrigSuccessors; 1520 BasicBlock *Head = SplitBefore->getParent(); 1521 if (DTU) { 1522 UniqueOrigSuccessors.insert_range(successors(Head)); 1523 Updates.reserve(4 + 2 * UniqueOrigSuccessors.size()); 1524 } 1525 1526 LLVMContext &C = Head->getContext(); 1527 BasicBlock *Tail = Head->splitBasicBlock(SplitBefore); 1528 BasicBlock *TrueBlock = Tail; 1529 BasicBlock *FalseBlock = Tail; 1530 bool ThenToTailEdge = false; 1531 bool ElseToTailEdge = false; 1532 1533 // Encapsulate the logic around creation/insertion/etc of a new block. 1534 auto handleBlock = [&](BasicBlock **PBB, bool Unreachable, BasicBlock *&BB, 1535 bool &ToTailEdge) { 1536 if (PBB == nullptr) 1537 return; // Do not create/insert a block. 1538 1539 if (*PBB) 1540 BB = *PBB; // Caller supplied block, use it. 1541 else { 1542 // Create a new block. 1543 BB = BasicBlock::Create(C, "", Head->getParent(), Tail); 1544 if (Unreachable) 1545 (void)new UnreachableInst(C, BB); 1546 else { 1547 (void)BranchInst::Create(Tail, BB); 1548 ToTailEdge = true; 1549 } 1550 BB->getTerminator()->setDebugLoc(SplitBefore->getDebugLoc()); 1551 // Pass the new block back to the caller. 1552 *PBB = BB; 1553 } 1554 }; 1555 1556 handleBlock(ThenBlock, UnreachableThen, TrueBlock, ThenToTailEdge); 1557 handleBlock(ElseBlock, UnreachableElse, FalseBlock, ElseToTailEdge); 1558 1559 Instruction *HeadOldTerm = Head->getTerminator(); 1560 BranchInst *HeadNewTerm = 1561 BranchInst::Create(/*ifTrue*/ TrueBlock, /*ifFalse*/ FalseBlock, Cond); 1562 HeadNewTerm->setMetadata(LLVMContext::MD_prof, BranchWeights); 1563 ReplaceInstWithInst(HeadOldTerm, HeadNewTerm); 1564 1565 if (DTU) { 1566 Updates.emplace_back(DominatorTree::Insert, Head, TrueBlock); 1567 Updates.emplace_back(DominatorTree::Insert, Head, FalseBlock); 1568 if (ThenToTailEdge) 1569 Updates.emplace_back(DominatorTree::Insert, TrueBlock, Tail); 1570 if (ElseToTailEdge) 1571 Updates.emplace_back(DominatorTree::Insert, FalseBlock, Tail); 1572 for (BasicBlock *UniqueOrigSuccessor : UniqueOrigSuccessors) 1573 Updates.emplace_back(DominatorTree::Insert, Tail, UniqueOrigSuccessor); 1574 for (BasicBlock *UniqueOrigSuccessor : UniqueOrigSuccessors) 1575 Updates.emplace_back(DominatorTree::Delete, Head, UniqueOrigSuccessor); 1576 DTU->applyUpdates(Updates); 1577 } 1578 1579 if (LI) { 1580 if (Loop *L = LI->getLoopFor(Head); L) { 1581 if (ThenToTailEdge) 1582 L->addBasicBlockToLoop(TrueBlock, *LI); 1583 if (ElseToTailEdge) 1584 L->addBasicBlockToLoop(FalseBlock, *LI); 1585 L->addBasicBlockToLoop(Tail, *LI); 1586 } 1587 } 1588 } 1589 1590 std::pair<Instruction *, Value *> 1591 llvm::SplitBlockAndInsertSimpleForLoop(Value *End, 1592 BasicBlock::iterator SplitBefore) { 1593 BasicBlock *LoopPred = SplitBefore->getParent(); 1594 BasicBlock *LoopBody = SplitBlock(SplitBefore->getParent(), SplitBefore); 1595 BasicBlock *LoopExit = SplitBlock(SplitBefore->getParent(), SplitBefore); 1596 1597 auto *Ty = End->getType(); 1598 auto &DL = SplitBefore->getDataLayout(); 1599 const unsigned Bitwidth = DL.getTypeSizeInBits(Ty); 1600 1601 IRBuilder<> Builder(LoopBody->getTerminator()); 1602 auto *IV = Builder.CreatePHI(Ty, 2, "iv"); 1603 auto *IVNext = 1604 Builder.CreateAdd(IV, ConstantInt::get(Ty, 1), IV->getName() + ".next", 1605 /*HasNUW=*/true, /*HasNSW=*/Bitwidth != 2); 1606 auto *IVCheck = Builder.CreateICmpEQ(IVNext, End, 1607 IV->getName() + ".check"); 1608 Builder.CreateCondBr(IVCheck, LoopExit, LoopBody); 1609 LoopBody->getTerminator()->eraseFromParent(); 1610 1611 // Populate the IV PHI. 1612 IV->addIncoming(ConstantInt::get(Ty, 0), LoopPred); 1613 IV->addIncoming(IVNext, LoopBody); 1614 1615 return std::make_pair(&*LoopBody->getFirstNonPHIIt(), IV); 1616 } 1617 1618 void llvm::SplitBlockAndInsertForEachLane( 1619 ElementCount EC, Type *IndexTy, BasicBlock::iterator InsertBefore, 1620 std::function<void(IRBuilderBase &, Value *)> Func) { 1621 1622 IRBuilder<> IRB(InsertBefore->getParent(), InsertBefore); 1623 1624 if (EC.isScalable()) { 1625 Value *NumElements = IRB.CreateElementCount(IndexTy, EC); 1626 1627 auto [BodyIP, Index] = 1628 SplitBlockAndInsertSimpleForLoop(NumElements, InsertBefore); 1629 1630 IRB.SetInsertPoint(BodyIP); 1631 Func(IRB, Index); 1632 return; 1633 } 1634 1635 unsigned Num = EC.getFixedValue(); 1636 for (unsigned Idx = 0; Idx < Num; ++Idx) { 1637 IRB.SetInsertPoint(InsertBefore); 1638 Func(IRB, ConstantInt::get(IndexTy, Idx)); 1639 } 1640 } 1641 1642 void llvm::SplitBlockAndInsertForEachLane( 1643 Value *EVL, BasicBlock::iterator InsertBefore, 1644 std::function<void(IRBuilderBase &, Value *)> Func) { 1645 1646 IRBuilder<> IRB(InsertBefore->getParent(), InsertBefore); 1647 Type *Ty = EVL->getType(); 1648 1649 if (!isa<ConstantInt>(EVL)) { 1650 auto [BodyIP, Index] = SplitBlockAndInsertSimpleForLoop(EVL, InsertBefore); 1651 IRB.SetInsertPoint(BodyIP); 1652 Func(IRB, Index); 1653 return; 1654 } 1655 1656 unsigned Num = cast<ConstantInt>(EVL)->getZExtValue(); 1657 for (unsigned Idx = 0; Idx < Num; ++Idx) { 1658 IRB.SetInsertPoint(InsertBefore); 1659 Func(IRB, ConstantInt::get(Ty, Idx)); 1660 } 1661 } 1662 1663 BranchInst *llvm::GetIfCondition(BasicBlock *BB, BasicBlock *&IfTrue, 1664 BasicBlock *&IfFalse) { 1665 PHINode *SomePHI = dyn_cast<PHINode>(BB->begin()); 1666 BasicBlock *Pred1 = nullptr; 1667 BasicBlock *Pred2 = nullptr; 1668 1669 if (SomePHI) { 1670 if (SomePHI->getNumIncomingValues() != 2) 1671 return nullptr; 1672 Pred1 = SomePHI->getIncomingBlock(0); 1673 Pred2 = SomePHI->getIncomingBlock(1); 1674 } else { 1675 pred_iterator PI = pred_begin(BB), PE = pred_end(BB); 1676 if (PI == PE) // No predecessor 1677 return nullptr; 1678 Pred1 = *PI++; 1679 if (PI == PE) // Only one predecessor 1680 return nullptr; 1681 Pred2 = *PI++; 1682 if (PI != PE) // More than two predecessors 1683 return nullptr; 1684 } 1685 1686 // We can only handle branches. Other control flow will be lowered to 1687 // branches if possible anyway. 1688 BranchInst *Pred1Br = dyn_cast<BranchInst>(Pred1->getTerminator()); 1689 BranchInst *Pred2Br = dyn_cast<BranchInst>(Pred2->getTerminator()); 1690 if (!Pred1Br || !Pred2Br) 1691 return nullptr; 1692 1693 // Eliminate code duplication by ensuring that Pred1Br is conditional if 1694 // either are. 1695 if (Pred2Br->isConditional()) { 1696 // If both branches are conditional, we don't have an "if statement". In 1697 // reality, we could transform this case, but since the condition will be 1698 // required anyway, we stand no chance of eliminating it, so the xform is 1699 // probably not profitable. 1700 if (Pred1Br->isConditional()) 1701 return nullptr; 1702 1703 std::swap(Pred1, Pred2); 1704 std::swap(Pred1Br, Pred2Br); 1705 } 1706 1707 if (Pred1Br->isConditional()) { 1708 // The only thing we have to watch out for here is to make sure that Pred2 1709 // doesn't have incoming edges from other blocks. If it does, the condition 1710 // doesn't dominate BB. 1711 if (!Pred2->getSinglePredecessor()) 1712 return nullptr; 1713 1714 // If we found a conditional branch predecessor, make sure that it branches 1715 // to BB and Pred2Br. If it doesn't, this isn't an "if statement". 1716 if (Pred1Br->getSuccessor(0) == BB && 1717 Pred1Br->getSuccessor(1) == Pred2) { 1718 IfTrue = Pred1; 1719 IfFalse = Pred2; 1720 } else if (Pred1Br->getSuccessor(0) == Pred2 && 1721 Pred1Br->getSuccessor(1) == BB) { 1722 IfTrue = Pred2; 1723 IfFalse = Pred1; 1724 } else { 1725 // We know that one arm of the conditional goes to BB, so the other must 1726 // go somewhere unrelated, and this must not be an "if statement". 1727 return nullptr; 1728 } 1729 1730 return Pred1Br; 1731 } 1732 1733 // Ok, if we got here, both predecessors end with an unconditional branch to 1734 // BB. Don't panic! If both blocks only have a single (identical) 1735 // predecessor, and THAT is a conditional branch, then we're all ok! 1736 BasicBlock *CommonPred = Pred1->getSinglePredecessor(); 1737 if (CommonPred == nullptr || CommonPred != Pred2->getSinglePredecessor()) 1738 return nullptr; 1739 1740 // Otherwise, if this is a conditional branch, then we can use it! 1741 BranchInst *BI = dyn_cast<BranchInst>(CommonPred->getTerminator()); 1742 if (!BI) return nullptr; 1743 1744 assert(BI->isConditional() && "Two successors but not conditional?"); 1745 if (BI->getSuccessor(0) == Pred1) { 1746 IfTrue = Pred1; 1747 IfFalse = Pred2; 1748 } else { 1749 IfTrue = Pred2; 1750 IfFalse = Pred1; 1751 } 1752 return BI; 1753 } 1754 1755 void llvm::InvertBranch(BranchInst *PBI, IRBuilderBase &Builder) { 1756 Value *NewCond = PBI->getCondition(); 1757 // If this is a "cmp" instruction, only used for branching (and nowhere 1758 // else), then we can simply invert the predicate. 1759 if (NewCond->hasOneUse() && isa<CmpInst>(NewCond)) { 1760 CmpInst *CI = cast<CmpInst>(NewCond); 1761 CI->setPredicate(CI->getInversePredicate()); 1762 } else 1763 NewCond = Builder.CreateNot(NewCond, NewCond->getName() + ".not"); 1764 1765 PBI->setCondition(NewCond); 1766 PBI->swapSuccessors(); 1767 } 1768 1769 bool llvm::hasOnlySimpleTerminator(const Function &F) { 1770 for (auto &BB : F) { 1771 auto *Term = BB.getTerminator(); 1772 if (!(isa<ReturnInst>(Term) || isa<UnreachableInst>(Term) || 1773 isa<BranchInst>(Term))) 1774 return false; 1775 } 1776 return true; 1777 } 1778