Local.cpp revision 3cb8509b680209c79632d7445709452f28213057
1//===-- Local.cpp - Functions to perform local transformations ------------===// 2// 3// The LLVM Compiler Infrastructure 4// 5// This file is distributed under the University of Illinois Open Source 6// License. See LICENSE.TXT for details. 7// 8//===----------------------------------------------------------------------===// 9// 10// This family of functions perform various local transformations to the 11// program. 12// 13//===----------------------------------------------------------------------===// 14 15#include "llvm/Transforms/Utils/Local.h" 16#include "llvm/Constants.h" 17#include "llvm/GlobalAlias.h" 18#include "llvm/GlobalVariable.h" 19#include "llvm/DerivedTypes.h" 20#include "llvm/Instructions.h" 21#include "llvm/Intrinsics.h" 22#include "llvm/IntrinsicInst.h" 23#include "llvm/ADT/DenseMap.h" 24#include "llvm/ADT/SmallPtrSet.h" 25#include "llvm/Analysis/ConstantFolding.h" 26#include "llvm/Analysis/InstructionSimplify.h" 27#include "llvm/Analysis/ProfileInfo.h" 28#include "llvm/Target/TargetData.h" 29#include "llvm/Support/CFG.h" 30#include "llvm/Support/Debug.h" 31#include "llvm/Support/GetElementPtrTypeIterator.h" 32#include "llvm/Support/MathExtras.h" 33#include "llvm/Support/ValueHandle.h" 34#include "llvm/Support/raw_ostream.h" 35using namespace llvm; 36 37//===----------------------------------------------------------------------===// 38// Local analysis. 39// 40 41/// getUnderlyingObjectWithOffset - Strip off up to MaxLookup GEPs and 42/// bitcasts to get back to the underlying object being addressed, keeping 43/// track of the offset in bytes from the GEPs relative to the result. 44/// This is closely related to Value::getUnderlyingObject but is located 45/// here to avoid making VMCore depend on TargetData. 46static Value *getUnderlyingObjectWithOffset(Value *V, const TargetData *TD, 47 uint64_t &ByteOffset, 48 unsigned MaxLookup = 6) { 49 if (!isa<PointerType>(V->getType())) 50 return V; 51 for (unsigned Count = 0; MaxLookup == 0 || Count < MaxLookup; ++Count) { 52 if (GEPOperator *GEP = dyn_cast<GEPOperator>(V)) { 53 if (!GEP->hasAllConstantIndices()) 54 return V; 55 SmallVector<Value*, 8> Indices(GEP->op_begin() + 1, GEP->op_end()); 56 ByteOffset += TD->getIndexedOffset(GEP->getPointerOperandType(), 57 &Indices[0], Indices.size()); 58 V = GEP->getPointerOperand(); 59 } else if (Operator::getOpcode(V) == Instruction::BitCast) { 60 V = cast<Operator>(V)->getOperand(0); 61 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) { 62 if (GA->mayBeOverridden()) 63 return V; 64 V = GA->getAliasee(); 65 } else { 66 return V; 67 } 68 assert(isa<PointerType>(V->getType()) && "Unexpected operand type!"); 69 } 70 return V; 71} 72 73/// isSafeToLoadUnconditionally - Return true if we know that executing a load 74/// from this value cannot trap. If it is not obviously safe to load from the 75/// specified pointer, we do a quick local scan of the basic block containing 76/// ScanFrom, to determine if the address is already accessed. 77bool llvm::isSafeToLoadUnconditionally(Value *V, Instruction *ScanFrom, 78 unsigned Align, const TargetData *TD) { 79 uint64_t ByteOffset = 0; 80 Value *Base = V; 81 if (TD) 82 Base = getUnderlyingObjectWithOffset(V, TD, ByteOffset); 83 84 const Type *BaseType = 0; 85 unsigned BaseAlign = 0; 86 if (const AllocaInst *AI = dyn_cast<AllocaInst>(Base)) { 87 // An alloca is safe to load from as load as it is suitably aligned. 88 BaseType = AI->getAllocatedType(); 89 BaseAlign = AI->getAlignment(); 90 } else if (const GlobalValue *GV = dyn_cast<GlobalValue>(Base)) { 91 // Global variables are safe to load from but their size cannot be 92 // guaranteed if they are overridden. 93 if (!isa<GlobalAlias>(GV) && !GV->mayBeOverridden()) { 94 BaseType = GV->getType()->getElementType(); 95 BaseAlign = GV->getAlignment(); 96 } 97 } 98 99 if (BaseType && BaseType->isSized()) { 100 if (TD && BaseAlign == 0) 101 BaseAlign = TD->getPrefTypeAlignment(BaseType); 102 103 if (Align <= BaseAlign) { 104 if (!TD) 105 return true; // Loading directly from an alloca or global is OK. 106 107 // Check if the load is within the bounds of the underlying object. 108 const PointerType *AddrTy = cast<PointerType>(V->getType()); 109 uint64_t LoadSize = TD->getTypeStoreSize(AddrTy->getElementType()); 110 if (ByteOffset + LoadSize <= TD->getTypeAllocSize(BaseType) && 111 (Align == 0 || (ByteOffset % Align) == 0)) 112 return true; 113 } 114 } 115 116 // Otherwise, be a little bit aggressive by scanning the local block where we 117 // want to check to see if the pointer is already being loaded or stored 118 // from/to. If so, the previous load or store would have already trapped, 119 // so there is no harm doing an extra load (also, CSE will later eliminate 120 // the load entirely). 121 BasicBlock::iterator BBI = ScanFrom, E = ScanFrom->getParent()->begin(); 122 123 while (BBI != E) { 124 --BBI; 125 126 // If we see a free or a call which may write to memory (i.e. which might do 127 // a free) the pointer could be marked invalid. 128 if (isa<CallInst>(BBI) && BBI->mayWriteToMemory() && 129 !isa<DbgInfoIntrinsic>(BBI)) 130 return false; 131 132 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) { 133 if (LI->getOperand(0) == V) return true; 134 } else if (StoreInst *SI = dyn_cast<StoreInst>(BBI)) { 135 if (SI->getOperand(1) == V) return true; 136 } 137 } 138 return false; 139} 140 141 142//===----------------------------------------------------------------------===// 143// Local constant propagation. 144// 145 146// ConstantFoldTerminator - If a terminator instruction is predicated on a 147// constant value, convert it into an unconditional branch to the constant 148// destination. 149// 150bool llvm::ConstantFoldTerminator(BasicBlock *BB) { 151 TerminatorInst *T = BB->getTerminator(); 152 153 // Branch - See if we are conditional jumping on constant 154 if (BranchInst *BI = dyn_cast<BranchInst>(T)) { 155 if (BI->isUnconditional()) return false; // Can't optimize uncond branch 156 BasicBlock *Dest1 = BI->getSuccessor(0); 157 BasicBlock *Dest2 = BI->getSuccessor(1); 158 159 if (ConstantInt *Cond = dyn_cast<ConstantInt>(BI->getCondition())) { 160 // Are we branching on constant? 161 // YES. Change to unconditional branch... 162 BasicBlock *Destination = Cond->getZExtValue() ? Dest1 : Dest2; 163 BasicBlock *OldDest = Cond->getZExtValue() ? Dest2 : Dest1; 164 165 //cerr << "Function: " << T->getParent()->getParent() 166 // << "\nRemoving branch from " << T->getParent() 167 // << "\n\nTo: " << OldDest << endl; 168 169 // Let the basic block know that we are letting go of it. Based on this, 170 // it will adjust it's PHI nodes. 171 assert(BI->getParent() && "Terminator not inserted in block!"); 172 OldDest->removePredecessor(BI->getParent()); 173 174 // Set the unconditional destination, and change the insn to be an 175 // unconditional branch. 176 BI->setUnconditionalDest(Destination); 177 return true; 178 } 179 180 if (Dest2 == Dest1) { // Conditional branch to same location? 181 // This branch matches something like this: 182 // br bool %cond, label %Dest, label %Dest 183 // and changes it into: br label %Dest 184 185 // Let the basic block know that we are letting go of one copy of it. 186 assert(BI->getParent() && "Terminator not inserted in block!"); 187 Dest1->removePredecessor(BI->getParent()); 188 189 // Change a conditional branch to unconditional. 190 BI->setUnconditionalDest(Dest1); 191 return true; 192 } 193 return false; 194 } 195 196 if (SwitchInst *SI = dyn_cast<SwitchInst>(T)) { 197 // If we are switching on a constant, we can convert the switch into a 198 // single branch instruction! 199 ConstantInt *CI = dyn_cast<ConstantInt>(SI->getCondition()); 200 BasicBlock *TheOnlyDest = SI->getSuccessor(0); // The default dest 201 BasicBlock *DefaultDest = TheOnlyDest; 202 assert(TheOnlyDest == SI->getDefaultDest() && 203 "Default destination is not successor #0?"); 204 205 // Figure out which case it goes to. 206 for (unsigned i = 1, e = SI->getNumSuccessors(); i != e; ++i) { 207 // Found case matching a constant operand? 208 if (SI->getSuccessorValue(i) == CI) { 209 TheOnlyDest = SI->getSuccessor(i); 210 break; 211 } 212 213 // Check to see if this branch is going to the same place as the default 214 // dest. If so, eliminate it as an explicit compare. 215 if (SI->getSuccessor(i) == DefaultDest) { 216 // Remove this entry. 217 DefaultDest->removePredecessor(SI->getParent()); 218 SI->removeCase(i); 219 --i; --e; // Don't skip an entry... 220 continue; 221 } 222 223 // Otherwise, check to see if the switch only branches to one destination. 224 // We do this by reseting "TheOnlyDest" to null when we find two non-equal 225 // destinations. 226 if (SI->getSuccessor(i) != TheOnlyDest) TheOnlyDest = 0; 227 } 228 229 if (CI && !TheOnlyDest) { 230 // Branching on a constant, but not any of the cases, go to the default 231 // successor. 232 TheOnlyDest = SI->getDefaultDest(); 233 } 234 235 // If we found a single destination that we can fold the switch into, do so 236 // now. 237 if (TheOnlyDest) { 238 // Insert the new branch. 239 BranchInst::Create(TheOnlyDest, SI); 240 BasicBlock *BB = SI->getParent(); 241 242 // Remove entries from PHI nodes which we no longer branch to... 243 for (unsigned i = 0, e = SI->getNumSuccessors(); i != e; ++i) { 244 // Found case matching a constant operand? 245 BasicBlock *Succ = SI->getSuccessor(i); 246 if (Succ == TheOnlyDest) 247 TheOnlyDest = 0; // Don't modify the first branch to TheOnlyDest 248 else 249 Succ->removePredecessor(BB); 250 } 251 252 // Delete the old switch. 253 BB->getInstList().erase(SI); 254 return true; 255 } 256 257 if (SI->getNumSuccessors() == 2) { 258 // Otherwise, we can fold this switch into a conditional branch 259 // instruction if it has only one non-default destination. 260 Value *Cond = new ICmpInst(SI, ICmpInst::ICMP_EQ, SI->getCondition(), 261 SI->getSuccessorValue(1), "cond"); 262 // Insert the new branch. 263 BranchInst::Create(SI->getSuccessor(1), SI->getSuccessor(0), Cond, SI); 264 265 // Delete the old switch. 266 SI->eraseFromParent(); 267 return true; 268 } 269 return false; 270 } 271 272 if (IndirectBrInst *IBI = dyn_cast<IndirectBrInst>(T)) { 273 // indirectbr blockaddress(@F, @BB) -> br label @BB 274 if (BlockAddress *BA = 275 dyn_cast<BlockAddress>(IBI->getAddress()->stripPointerCasts())) { 276 BasicBlock *TheOnlyDest = BA->getBasicBlock(); 277 // Insert the new branch. 278 BranchInst::Create(TheOnlyDest, IBI); 279 280 for (unsigned i = 0, e = IBI->getNumDestinations(); i != e; ++i) { 281 if (IBI->getDestination(i) == TheOnlyDest) 282 TheOnlyDest = 0; 283 else 284 IBI->getDestination(i)->removePredecessor(IBI->getParent()); 285 } 286 IBI->eraseFromParent(); 287 288 // If we didn't find our destination in the IBI successor list, then we 289 // have undefined behavior. Replace the unconditional branch with an 290 // 'unreachable' instruction. 291 if (TheOnlyDest) { 292 BB->getTerminator()->eraseFromParent(); 293 new UnreachableInst(BB->getContext(), BB); 294 } 295 296 return true; 297 } 298 } 299 300 return false; 301} 302 303 304//===----------------------------------------------------------------------===// 305// Local dead code elimination. 306// 307 308/// isInstructionTriviallyDead - Return true if the result produced by the 309/// instruction is not used, and the instruction has no side effects. 310/// 311bool llvm::isInstructionTriviallyDead(Instruction *I) { 312 if (!I->use_empty() || isa<TerminatorInst>(I)) return false; 313 314 // We don't want debug info removed by anything this general. 315 if (isa<DbgInfoIntrinsic>(I)) return false; 316 317 // Likewise for memory use markers. 318 if (isa<MemoryUseIntrinsic>(I)) return false; 319 320 if (!I->mayHaveSideEffects()) return true; 321 322 // Special case intrinsics that "may have side effects" but can be deleted 323 // when dead. 324 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) 325 // Safe to delete llvm.stacksave if dead. 326 if (II->getIntrinsicID() == Intrinsic::stacksave) 327 return true; 328 return false; 329} 330 331/// RecursivelyDeleteTriviallyDeadInstructions - If the specified value is a 332/// trivially dead instruction, delete it. If that makes any of its operands 333/// trivially dead, delete them too, recursively. Return true if any 334/// instructions were deleted. 335bool llvm::RecursivelyDeleteTriviallyDeadInstructions(Value *V) { 336 Instruction *I = dyn_cast<Instruction>(V); 337 if (!I || !I->use_empty() || !isInstructionTriviallyDead(I)) 338 return false; 339 340 SmallVector<Instruction*, 16> DeadInsts; 341 DeadInsts.push_back(I); 342 343 do { 344 I = DeadInsts.pop_back_val(); 345 346 // Null out all of the instruction's operands to see if any operand becomes 347 // dead as we go. 348 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) { 349 Value *OpV = I->getOperand(i); 350 I->setOperand(i, 0); 351 352 if (!OpV->use_empty()) continue; 353 354 // If the operand is an instruction that became dead as we nulled out the 355 // operand, and if it is 'trivially' dead, delete it in a future loop 356 // iteration. 357 if (Instruction *OpI = dyn_cast<Instruction>(OpV)) 358 if (isInstructionTriviallyDead(OpI)) 359 DeadInsts.push_back(OpI); 360 } 361 362 I->eraseFromParent(); 363 } while (!DeadInsts.empty()); 364 365 return true; 366} 367 368/// RecursivelyDeleteDeadPHINode - If the specified value is an effectively 369/// dead PHI node, due to being a def-use chain of single-use nodes that 370/// either forms a cycle or is terminated by a trivially dead instruction, 371/// delete it. If that makes any of its operands trivially dead, delete them 372/// too, recursively. Return true if the PHI node is actually deleted. 373bool 374llvm::RecursivelyDeleteDeadPHINode(PHINode *PN) { 375 // We can remove a PHI if it is on a cycle in the def-use graph 376 // where each node in the cycle has degree one, i.e. only one use, 377 // and is an instruction with no side effects. 378 if (!PN->hasOneUse()) 379 return false; 380 381 bool Changed = false; 382 SmallPtrSet<PHINode *, 4> PHIs; 383 PHIs.insert(PN); 384 for (Instruction *J = cast<Instruction>(*PN->use_begin()); 385 J->hasOneUse() && !J->mayHaveSideEffects(); 386 J = cast<Instruction>(*J->use_begin())) 387 // If we find a PHI more than once, we're on a cycle that 388 // won't prove fruitful. 389 if (PHINode *JP = dyn_cast<PHINode>(J)) 390 if (!PHIs.insert(cast<PHINode>(JP))) { 391 // Break the cycle and delete the PHI and its operands. 392 JP->replaceAllUsesWith(UndefValue::get(JP->getType())); 393 (void)RecursivelyDeleteTriviallyDeadInstructions(JP); 394 Changed = true; 395 break; 396 } 397 return Changed; 398} 399 400/// SimplifyInstructionsInBlock - Scan the specified basic block and try to 401/// simplify any instructions in it and recursively delete dead instructions. 402/// 403/// This returns true if it changed the code, note that it can delete 404/// instructions in other blocks as well in this block. 405bool llvm::SimplifyInstructionsInBlock(BasicBlock *BB, const TargetData *TD) { 406 bool MadeChange = false; 407 for (BasicBlock::iterator BI = BB->begin(), E = BB->end(); BI != E; ) { 408 Instruction *Inst = BI++; 409 410 if (Value *V = SimplifyInstruction(Inst, TD)) { 411 WeakVH BIHandle(BI); 412 ReplaceAndSimplifyAllUses(Inst, V, TD); 413 MadeChange = true; 414 if (BIHandle == 0) 415 BI = BB->begin(); 416 continue; 417 } 418 419 MadeChange |= RecursivelyDeleteTriviallyDeadInstructions(Inst); 420 } 421 return MadeChange; 422} 423 424//===----------------------------------------------------------------------===// 425// Control Flow Graph Restructuring. 426// 427 428 429/// RemovePredecessorAndSimplify - Like BasicBlock::removePredecessor, this 430/// method is called when we're about to delete Pred as a predecessor of BB. If 431/// BB contains any PHI nodes, this drops the entries in the PHI nodes for Pred. 432/// 433/// Unlike the removePredecessor method, this attempts to simplify uses of PHI 434/// nodes that collapse into identity values. For example, if we have: 435/// x = phi(1, 0, 0, 0) 436/// y = and x, z 437/// 438/// .. and delete the predecessor corresponding to the '1', this will attempt to 439/// recursively fold the and to 0. 440void llvm::RemovePredecessorAndSimplify(BasicBlock *BB, BasicBlock *Pred, 441 TargetData *TD) { 442 // This only adjusts blocks with PHI nodes. 443 if (!isa<PHINode>(BB->begin())) 444 return; 445 446 // Remove the entries for Pred from the PHI nodes in BB, but do not simplify 447 // them down. This will leave us with single entry phi nodes and other phis 448 // that can be removed. 449 BB->removePredecessor(Pred, true); 450 451 WeakVH PhiIt = &BB->front(); 452 while (PHINode *PN = dyn_cast<PHINode>(PhiIt)) { 453 PhiIt = &*++BasicBlock::iterator(cast<Instruction>(PhiIt)); 454 455 Value *PNV = PN->hasConstantValue(); 456 if (PNV == 0) continue; 457 458 // If we're able to simplify the phi to a single value, substitute the new 459 // value into all of its uses. 460 assert(PNV != PN && "hasConstantValue broken"); 461 462 ReplaceAndSimplifyAllUses(PN, PNV, TD); 463 464 // If recursive simplification ended up deleting the next PHI node we would 465 // iterate to, then our iterator is invalid, restart scanning from the top 466 // of the block. 467 if (PhiIt == 0) PhiIt = &BB->front(); 468 } 469} 470 471 472/// MergeBasicBlockIntoOnlyPred - DestBB is a block with one predecessor and its 473/// predecessor is known to have one successor (DestBB!). Eliminate the edge 474/// between them, moving the instructions in the predecessor into DestBB and 475/// deleting the predecessor block. 476/// 477void llvm::MergeBasicBlockIntoOnlyPred(BasicBlock *DestBB, Pass *P) { 478 // If BB has single-entry PHI nodes, fold them. 479 while (PHINode *PN = dyn_cast<PHINode>(DestBB->begin())) { 480 Value *NewVal = PN->getIncomingValue(0); 481 // Replace self referencing PHI with undef, it must be dead. 482 if (NewVal == PN) NewVal = UndefValue::get(PN->getType()); 483 PN->replaceAllUsesWith(NewVal); 484 PN->eraseFromParent(); 485 } 486 487 BasicBlock *PredBB = DestBB->getSinglePredecessor(); 488 assert(PredBB && "Block doesn't have a single predecessor!"); 489 490 // Splice all the instructions from PredBB to DestBB. 491 PredBB->getTerminator()->eraseFromParent(); 492 DestBB->getInstList().splice(DestBB->begin(), PredBB->getInstList()); 493 494 // Anything that branched to PredBB now branches to DestBB. 495 PredBB->replaceAllUsesWith(DestBB); 496 497 if (P) { 498 ProfileInfo *PI = P->getAnalysisIfAvailable<ProfileInfo>(); 499 if (PI) { 500 PI->replaceAllUses(PredBB, DestBB); 501 PI->removeEdge(ProfileInfo::getEdge(PredBB, DestBB)); 502 } 503 } 504 // Nuke BB. 505 PredBB->eraseFromParent(); 506} 507 508/// CanPropagatePredecessorsForPHIs - Return true if we can fold BB, an 509/// almost-empty BB ending in an unconditional branch to Succ, into succ. 510/// 511/// Assumption: Succ is the single successor for BB. 512/// 513static bool CanPropagatePredecessorsForPHIs(BasicBlock *BB, BasicBlock *Succ) { 514 assert(*succ_begin(BB) == Succ && "Succ is not successor of BB!"); 515 516 DEBUG(dbgs() << "Looking to fold " << BB->getName() << " into " 517 << Succ->getName() << "\n"); 518 // Shortcut, if there is only a single predecessor it must be BB and merging 519 // is always safe 520 if (Succ->getSinglePredecessor()) return true; 521 522 // Make a list of the predecessors of BB 523 typedef SmallPtrSet<BasicBlock*, 16> BlockSet; 524 BlockSet BBPreds(pred_begin(BB), pred_end(BB)); 525 526 // Use that list to make another list of common predecessors of BB and Succ 527 BlockSet CommonPreds; 528 for (pred_iterator PI = pred_begin(Succ), PE = pred_end(Succ); 529 PI != PE; ++PI) 530 if (BBPreds.count(*PI)) 531 CommonPreds.insert(*PI); 532 533 // Shortcut, if there are no common predecessors, merging is always safe 534 if (CommonPreds.empty()) 535 return true; 536 537 // Look at all the phi nodes in Succ, to see if they present a conflict when 538 // merging these blocks 539 for (BasicBlock::iterator I = Succ->begin(); isa<PHINode>(I); ++I) { 540 PHINode *PN = cast<PHINode>(I); 541 542 // If the incoming value from BB is again a PHINode in 543 // BB which has the same incoming value for *PI as PN does, we can 544 // merge the phi nodes and then the blocks can still be merged 545 PHINode *BBPN = dyn_cast<PHINode>(PN->getIncomingValueForBlock(BB)); 546 if (BBPN && BBPN->getParent() == BB) { 547 for (BlockSet::iterator PI = CommonPreds.begin(), PE = CommonPreds.end(); 548 PI != PE; PI++) { 549 if (BBPN->getIncomingValueForBlock(*PI) 550 != PN->getIncomingValueForBlock(*PI)) { 551 DEBUG(dbgs() << "Can't fold, phi node " << PN->getName() << " in " 552 << Succ->getName() << " is conflicting with " 553 << BBPN->getName() << " with regard to common predecessor " 554 << (*PI)->getName() << "\n"); 555 return false; 556 } 557 } 558 } else { 559 Value* Val = PN->getIncomingValueForBlock(BB); 560 for (BlockSet::iterator PI = CommonPreds.begin(), PE = CommonPreds.end(); 561 PI != PE; PI++) { 562 // See if the incoming value for the common predecessor is equal to the 563 // one for BB, in which case this phi node will not prevent the merging 564 // of the block. 565 if (Val != PN->getIncomingValueForBlock(*PI)) { 566 DEBUG(dbgs() << "Can't fold, phi node " << PN->getName() << " in " 567 << Succ->getName() << " is conflicting with regard to common " 568 << "predecessor " << (*PI)->getName() << "\n"); 569 return false; 570 } 571 } 572 } 573 } 574 575 return true; 576} 577 578/// TryToSimplifyUncondBranchFromEmptyBlock - BB is known to contain an 579/// unconditional branch, and contains no instructions other than PHI nodes, 580/// potential debug intrinsics and the branch. If possible, eliminate BB by 581/// rewriting all the predecessors to branch to the successor block and return 582/// true. If we can't transform, return false. 583bool llvm::TryToSimplifyUncondBranchFromEmptyBlock(BasicBlock *BB) { 584 // We can't eliminate infinite loops. 585 BasicBlock *Succ = cast<BranchInst>(BB->getTerminator())->getSuccessor(0); 586 if (BB == Succ) return false; 587 588 // Check to see if merging these blocks would cause conflicts for any of the 589 // phi nodes in BB or Succ. If not, we can safely merge. 590 if (!CanPropagatePredecessorsForPHIs(BB, Succ)) return false; 591 592 // Check for cases where Succ has multiple predecessors and a PHI node in BB 593 // has uses which will not disappear when the PHI nodes are merged. It is 594 // possible to handle such cases, but difficult: it requires checking whether 595 // BB dominates Succ, which is non-trivial to calculate in the case where 596 // Succ has multiple predecessors. Also, it requires checking whether 597 // constructing the necessary self-referential PHI node doesn't intoduce any 598 // conflicts; this isn't too difficult, but the previous code for doing this 599 // was incorrect. 600 // 601 // Note that if this check finds a live use, BB dominates Succ, so BB is 602 // something like a loop pre-header (or rarely, a part of an irreducible CFG); 603 // folding the branch isn't profitable in that case anyway. 604 if (!Succ->getSinglePredecessor()) { 605 BasicBlock::iterator BBI = BB->begin(); 606 while (isa<PHINode>(*BBI)) { 607 for (Value::use_iterator UI = BBI->use_begin(), E = BBI->use_end(); 608 UI != E; ++UI) { 609 if (PHINode* PN = dyn_cast<PHINode>(*UI)) { 610 if (PN->getIncomingBlock(UI) != BB) 611 return false; 612 } else { 613 return false; 614 } 615 } 616 ++BBI; 617 } 618 } 619 620 DEBUG(dbgs() << "Killing Trivial BB: \n" << *BB); 621 622 if (isa<PHINode>(Succ->begin())) { 623 // If there is more than one pred of succ, and there are PHI nodes in 624 // the successor, then we need to add incoming edges for the PHI nodes 625 // 626 const SmallVector<BasicBlock*, 16> BBPreds(pred_begin(BB), pred_end(BB)); 627 628 // Loop over all of the PHI nodes in the successor of BB. 629 for (BasicBlock::iterator I = Succ->begin(); isa<PHINode>(I); ++I) { 630 PHINode *PN = cast<PHINode>(I); 631 Value *OldVal = PN->removeIncomingValue(BB, false); 632 assert(OldVal && "No entry in PHI for Pred BB!"); 633 634 // If this incoming value is one of the PHI nodes in BB, the new entries 635 // in the PHI node are the entries from the old PHI. 636 if (isa<PHINode>(OldVal) && cast<PHINode>(OldVal)->getParent() == BB) { 637 PHINode *OldValPN = cast<PHINode>(OldVal); 638 for (unsigned i = 0, e = OldValPN->getNumIncomingValues(); i != e; ++i) 639 // Note that, since we are merging phi nodes and BB and Succ might 640 // have common predecessors, we could end up with a phi node with 641 // identical incoming branches. This will be cleaned up later (and 642 // will trigger asserts if we try to clean it up now, without also 643 // simplifying the corresponding conditional branch). 644 PN->addIncoming(OldValPN->getIncomingValue(i), 645 OldValPN->getIncomingBlock(i)); 646 } else { 647 // Add an incoming value for each of the new incoming values. 648 for (unsigned i = 0, e = BBPreds.size(); i != e; ++i) 649 PN->addIncoming(OldVal, BBPreds[i]); 650 } 651 } 652 } 653 654 while (PHINode *PN = dyn_cast<PHINode>(&BB->front())) { 655 if (Succ->getSinglePredecessor()) { 656 // BB is the only predecessor of Succ, so Succ will end up with exactly 657 // the same predecessors BB had. 658 Succ->getInstList().splice(Succ->begin(), 659 BB->getInstList(), BB->begin()); 660 } else { 661 // We explicitly check for such uses in CanPropagatePredecessorsForPHIs. 662 assert(PN->use_empty() && "There shouldn't be any uses here!"); 663 PN->eraseFromParent(); 664 } 665 } 666 667 // Everything that jumped to BB now goes to Succ. 668 BB->replaceAllUsesWith(Succ); 669 if (!Succ->hasName()) Succ->takeName(BB); 670 BB->eraseFromParent(); // Delete the old basic block. 671 return true; 672} 673 674/// EliminateDuplicatePHINodes - Check for and eliminate duplicate PHI 675/// nodes in this block. This doesn't try to be clever about PHI nodes 676/// which differ only in the order of the incoming values, but instcombine 677/// orders them so it usually won't matter. 678/// 679bool llvm::EliminateDuplicatePHINodes(BasicBlock *BB) { 680 bool Changed = false; 681 682 // This implementation doesn't currently consider undef operands 683 // specially. Theroetically, two phis which are identical except for 684 // one having an undef where the other doesn't could be collapsed. 685 686 // Map from PHI hash values to PHI nodes. If multiple PHIs have 687 // the same hash value, the element is the first PHI in the 688 // linked list in CollisionMap. 689 DenseMap<uintptr_t, PHINode *> HashMap; 690 691 // Maintain linked lists of PHI nodes with common hash values. 692 DenseMap<PHINode *, PHINode *> CollisionMap; 693 694 // Examine each PHI. 695 for (BasicBlock::iterator I = BB->begin(); 696 PHINode *PN = dyn_cast<PHINode>(I++); ) { 697 // Compute a hash value on the operands. Instcombine will likely have sorted 698 // them, which helps expose duplicates, but we have to check all the 699 // operands to be safe in case instcombine hasn't run. 700 uintptr_t Hash = 0; 701 for (User::op_iterator I = PN->op_begin(), E = PN->op_end(); I != E; ++I) { 702 // This hash algorithm is quite weak as hash functions go, but it seems 703 // to do a good enough job for this particular purpose, and is very quick. 704 Hash ^= reinterpret_cast<uintptr_t>(static_cast<Value *>(*I)); 705 Hash = (Hash << 7) | (Hash >> (sizeof(uintptr_t) * CHAR_BIT - 7)); 706 } 707 // If we've never seen this hash value before, it's a unique PHI. 708 std::pair<DenseMap<uintptr_t, PHINode *>::iterator, bool> Pair = 709 HashMap.insert(std::make_pair(Hash, PN)); 710 if (Pair.second) continue; 711 // Otherwise it's either a duplicate or a hash collision. 712 for (PHINode *OtherPN = Pair.first->second; ; ) { 713 if (OtherPN->isIdenticalTo(PN)) { 714 // A duplicate. Replace this PHI with its duplicate. 715 PN->replaceAllUsesWith(OtherPN); 716 PN->eraseFromParent(); 717 Changed = true; 718 break; 719 } 720 // A non-duplicate hash collision. 721 DenseMap<PHINode *, PHINode *>::iterator I = CollisionMap.find(OtherPN); 722 if (I == CollisionMap.end()) { 723 // Set this PHI to be the head of the linked list of colliding PHIs. 724 PHINode *Old = Pair.first->second; 725 Pair.first->second = PN; 726 CollisionMap[PN] = Old; 727 break; 728 } 729 // Procede to the next PHI in the list. 730 OtherPN = I->second; 731 } 732 } 733 734 return Changed; 735} 736