JumpThreading.cpp revision 508955156a25a9abc470a29e1760aa176d341cf9
1//===- JumpThreading.cpp - Thread control through conditional blocks ------===// 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 file implements the Jump Threading pass. 11// 12//===----------------------------------------------------------------------===// 13 14#define DEBUG_TYPE "jump-threading" 15#include "llvm/Transforms/Scalar.h" 16#include "llvm/IntrinsicInst.h" 17#include "llvm/LLVMContext.h" 18#include "llvm/Pass.h" 19#include "llvm/Analysis/ConstantFolding.h" 20#include "llvm/Transforms/Utils/BasicBlockUtils.h" 21#include "llvm/Transforms/Utils/Local.h" 22#include "llvm/Target/TargetData.h" 23#include "llvm/ADT/DenseMap.h" 24#include "llvm/ADT/Statistic.h" 25#include "llvm/ADT/STLExtras.h" 26#include "llvm/ADT/SmallPtrSet.h" 27#include "llvm/ADT/SmallSet.h" 28#include "llvm/Support/CommandLine.h" 29#include "llvm/Support/Compiler.h" 30#include "llvm/Support/Debug.h" 31#include "llvm/Support/ValueHandle.h" 32using namespace llvm; 33 34STATISTIC(NumThreads, "Number of jumps threaded"); 35STATISTIC(NumFolds, "Number of terminators folded"); 36 37static cl::opt<unsigned> 38Threshold("jump-threading-threshold", 39 cl::desc("Max block size to duplicate for jump threading"), 40 cl::init(6), cl::Hidden); 41 42namespace { 43 /// This pass performs 'jump threading', which looks at blocks that have 44 /// multiple predecessors and multiple successors. If one or more of the 45 /// predecessors of the block can be proven to always jump to one of the 46 /// successors, we forward the edge from the predecessor to the successor by 47 /// duplicating the contents of this block. 48 /// 49 /// An example of when this can occur is code like this: 50 /// 51 /// if () { ... 52 /// X = 4; 53 /// } 54 /// if (X < 3) { 55 /// 56 /// In this case, the unconditional branch at the end of the first if can be 57 /// revectored to the false side of the second if. 58 /// 59 class VISIBILITY_HIDDEN JumpThreading : public FunctionPass { 60 TargetData *TD; 61#ifdef NDEBUG 62 SmallPtrSet<BasicBlock*, 16> LoopHeaders; 63#else 64 SmallSet<AssertingVH<BasicBlock>, 16> LoopHeaders; 65#endif 66 public: 67 static char ID; // Pass identification 68 JumpThreading() : FunctionPass(&ID) {} 69 70 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 71 AU.addRequired<TargetData>(); 72 } 73 74 bool runOnFunction(Function &F); 75 void FindLoopHeaders(Function &F); 76 77 bool ProcessBlock(BasicBlock *BB); 78 bool ThreadEdge(BasicBlock *BB, BasicBlock *PredBB, BasicBlock *SuccBB, 79 unsigned JumpThreadCost); 80 BasicBlock *FactorCommonPHIPreds(PHINode *PN, Value *Val); 81 bool ProcessBranchOnDuplicateCond(BasicBlock *PredBB, BasicBlock *DestBB); 82 bool ProcessSwitchOnDuplicateCond(BasicBlock *PredBB, BasicBlock *DestBB); 83 84 bool ProcessJumpOnPHI(PHINode *PN); 85 bool ProcessBranchOnLogical(Value *V, BasicBlock *BB, bool isAnd); 86 bool ProcessBranchOnCompare(CmpInst *Cmp, BasicBlock *BB); 87 88 bool SimplifyPartiallyRedundantLoad(LoadInst *LI); 89 }; 90} 91 92char JumpThreading::ID = 0; 93static RegisterPass<JumpThreading> 94X("jump-threading", "Jump Threading"); 95 96// Public interface to the Jump Threading pass 97FunctionPass *llvm::createJumpThreadingPass() { return new JumpThreading(); } 98 99/// runOnFunction - Top level algorithm. 100/// 101bool JumpThreading::runOnFunction(Function &F) { 102 DOUT << "Jump threading on function '" << F.getNameStart() << "'\n"; 103 TD = &getAnalysis<TargetData>(); 104 105 FindLoopHeaders(F); 106 107 bool AnotherIteration = true, EverChanged = false; 108 while (AnotherIteration) { 109 AnotherIteration = false; 110 bool Changed = false; 111 for (Function::iterator I = F.begin(), E = F.end(); I != E;) { 112 BasicBlock *BB = I; 113 while (ProcessBlock(BB)) 114 Changed = true; 115 116 ++I; 117 118 // If the block is trivially dead, zap it. This eliminates the successor 119 // edges which simplifies the CFG. 120 if (pred_begin(BB) == pred_end(BB) && 121 BB != &BB->getParent()->getEntryBlock()) { 122 DOUT << " JT: Deleting dead block '" << BB->getNameStart() 123 << "' with terminator: " << *BB->getTerminator(); 124 LoopHeaders.erase(BB); 125 DeleteDeadBlock(BB); 126 Changed = true; 127 } 128 } 129 AnotherIteration = Changed; 130 EverChanged |= Changed; 131 } 132 133 LoopHeaders.clear(); 134 return EverChanged; 135} 136 137/// FindLoopHeaders - We do not want jump threading to turn proper loop 138/// structures into irreducible loops. Doing this breaks up the loop nesting 139/// hierarchy and pessimizes later transformations. To prevent this from 140/// happening, we first have to find the loop headers. Here we approximate this 141/// by finding targets of backedges in the CFG. 142/// 143/// Note that there definitely are cases when we want to allow threading of 144/// edges across a loop header. For example, threading a jump from outside the 145/// loop (the preheader) to an exit block of the loop is definitely profitable. 146/// It is also almost always profitable to thread backedges from within the loop 147/// to exit blocks, and is often profitable to thread backedges to other blocks 148/// within the loop (forming a nested loop). This simple analysis is not rich 149/// enough to track all of these properties and keep it up-to-date as the CFG 150/// mutates, so we don't allow any of these transformations. 151/// 152void JumpThreading::FindLoopHeaders(Function &F) { 153 SmallVector<std::pair<const BasicBlock*,const BasicBlock*>, 32> Edges; 154 FindFunctionBackedges(F, Edges); 155 156 for (unsigned i = 0, e = Edges.size(); i != e; ++i) 157 LoopHeaders.insert(const_cast<BasicBlock*>(Edges[i].second)); 158} 159 160 161/// FactorCommonPHIPreds - If there are multiple preds with the same incoming 162/// value for the PHI, factor them together so we get one block to thread for 163/// the whole group. 164/// This is important for things like "phi i1 [true, true, false, true, x]" 165/// where we only need to clone the block for the true blocks once. 166/// 167BasicBlock *JumpThreading::FactorCommonPHIPreds(PHINode *PN, Value *Val) { 168 SmallVector<BasicBlock*, 16> CommonPreds; 169 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) 170 if (PN->getIncomingValue(i) == Val) 171 CommonPreds.push_back(PN->getIncomingBlock(i)); 172 173 if (CommonPreds.size() == 1) 174 return CommonPreds[0]; 175 176 DOUT << " Factoring out " << CommonPreds.size() 177 << " common predecessors.\n"; 178 return SplitBlockPredecessors(PN->getParent(), 179 &CommonPreds[0], CommonPreds.size(), 180 ".thr_comm", this); 181} 182 183 184/// getJumpThreadDuplicationCost - Return the cost of duplicating this block to 185/// thread across it. 186static unsigned getJumpThreadDuplicationCost(const BasicBlock *BB) { 187 /// Ignore PHI nodes, these will be flattened when duplication happens. 188 BasicBlock::const_iterator I = BB->getFirstNonPHI(); 189 190 // Sum up the cost of each instruction until we get to the terminator. Don't 191 // include the terminator because the copy won't include it. 192 unsigned Size = 0; 193 for (; !isa<TerminatorInst>(I); ++I) { 194 // Debugger intrinsics don't incur code size. 195 if (isa<DbgInfoIntrinsic>(I)) continue; 196 197 // If this is a pointer->pointer bitcast, it is free. 198 if (isa<BitCastInst>(I) && isa<PointerType>(I->getType())) 199 continue; 200 201 // All other instructions count for at least one unit. 202 ++Size; 203 204 // Calls are more expensive. If they are non-intrinsic calls, we model them 205 // as having cost of 4. If they are a non-vector intrinsic, we model them 206 // as having cost of 2 total, and if they are a vector intrinsic, we model 207 // them as having cost 1. 208 if (const CallInst *CI = dyn_cast<CallInst>(I)) { 209 if (!isa<IntrinsicInst>(CI)) 210 Size += 3; 211 else if (!isa<VectorType>(CI->getType())) 212 Size += 1; 213 } 214 } 215 216 // Threading through a switch statement is particularly profitable. If this 217 // block ends in a switch, decrease its cost to make it more likely to happen. 218 if (isa<SwitchInst>(I)) 219 Size = Size > 6 ? Size-6 : 0; 220 221 return Size; 222} 223 224/// ProcessBlock - If there are any predecessors whose control can be threaded 225/// through to a successor, transform them now. 226bool JumpThreading::ProcessBlock(BasicBlock *BB) { 227 // If this block has a single predecessor, and if that pred has a single 228 // successor, merge the blocks. This encourages recursive jump threading 229 // because now the condition in this block can be threaded through 230 // predecessors of our predecessor block. 231 if (BasicBlock *SinglePred = BB->getSinglePredecessor()) 232 if (SinglePred->getTerminator()->getNumSuccessors() == 1 && 233 SinglePred != BB) { 234 // If SinglePred was a loop header, BB becomes one. 235 if (LoopHeaders.erase(SinglePred)) 236 LoopHeaders.insert(BB); 237 238 // Remember if SinglePred was the entry block of the function. If so, we 239 // will need to move BB back to the entry position. 240 bool isEntry = SinglePred == &SinglePred->getParent()->getEntryBlock(); 241 MergeBasicBlockIntoOnlyPred(BB); 242 243 if (isEntry && BB != &BB->getParent()->getEntryBlock()) 244 BB->moveBefore(&BB->getParent()->getEntryBlock()); 245 return true; 246 } 247 248 // See if this block ends with a branch or switch. If so, see if the 249 // condition is a phi node. If so, and if an entry of the phi node is a 250 // constant, we can thread the block. 251 Value *Condition; 252 if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) { 253 // Can't thread an unconditional jump. 254 if (BI->isUnconditional()) return false; 255 Condition = BI->getCondition(); 256 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator())) 257 Condition = SI->getCondition(); 258 else 259 return false; // Must be an invoke. 260 261 // If the terminator of this block is branching on a constant, simplify the 262 // terminator to an unconditional branch. This can occur due to threading in 263 // other blocks. 264 if (isa<ConstantInt>(Condition)) { 265 DOUT << " In block '" << BB->getNameStart() 266 << "' folding terminator: " << *BB->getTerminator(); 267 ++NumFolds; 268 ConstantFoldTerminator(BB); 269 return true; 270 } 271 272 // If the terminator is branching on an undef, we can pick any of the 273 // successors to branch to. Since this is arbitrary, we pick the successor 274 // with the fewest predecessors. This should reduce the in-degree of the 275 // others. 276 if (isa<UndefValue>(Condition)) { 277 TerminatorInst *BBTerm = BB->getTerminator(); 278 unsigned MinSucc = 0; 279 BasicBlock *TestBB = BBTerm->getSuccessor(MinSucc); 280 // Compute the successor with the minimum number of predecessors. 281 unsigned MinNumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB)); 282 for (unsigned i = 1, e = BBTerm->getNumSuccessors(); i != e; ++i) { 283 TestBB = BBTerm->getSuccessor(i); 284 unsigned NumPreds = std::distance(pred_begin(TestBB), pred_end(TestBB)); 285 if (NumPreds < MinNumPreds) 286 MinSucc = i; 287 } 288 289 // Fold the branch/switch. 290 for (unsigned i = 0, e = BBTerm->getNumSuccessors(); i != e; ++i) { 291 if (i == MinSucc) continue; 292 BBTerm->getSuccessor(i)->removePredecessor(BB); 293 } 294 295 DOUT << " In block '" << BB->getNameStart() 296 << "' folding undef terminator: " << *BBTerm; 297 BranchInst::Create(BBTerm->getSuccessor(MinSucc), BBTerm); 298 BBTerm->eraseFromParent(); 299 return true; 300 } 301 302 Instruction *CondInst = dyn_cast<Instruction>(Condition); 303 304 // If the condition is an instruction defined in another block, see if a 305 // predecessor has the same condition: 306 // br COND, BBX, BBY 307 // BBX: 308 // br COND, BBZ, BBW 309 if (!Condition->hasOneUse() && // Multiple uses. 310 (CondInst == 0 || CondInst->getParent() != BB)) { // Non-local definition. 311 pred_iterator PI = pred_begin(BB), E = pred_end(BB); 312 if (isa<BranchInst>(BB->getTerminator())) { 313 for (; PI != E; ++PI) 314 if (BranchInst *PBI = dyn_cast<BranchInst>((*PI)->getTerminator())) 315 if (PBI->isConditional() && PBI->getCondition() == Condition && 316 ProcessBranchOnDuplicateCond(*PI, BB)) 317 return true; 318 } else { 319 assert(isa<SwitchInst>(BB->getTerminator()) && "Unknown jump terminator"); 320 for (; PI != E; ++PI) 321 if (SwitchInst *PSI = dyn_cast<SwitchInst>((*PI)->getTerminator())) 322 if (PSI->getCondition() == Condition && 323 ProcessSwitchOnDuplicateCond(*PI, BB)) 324 return true; 325 } 326 } 327 328 // All the rest of our checks depend on the condition being an instruction. 329 if (CondInst == 0) 330 return false; 331 332 // See if this is a phi node in the current block. 333 if (PHINode *PN = dyn_cast<PHINode>(CondInst)) 334 if (PN->getParent() == BB) 335 return ProcessJumpOnPHI(PN); 336 337 // If this is a conditional branch whose condition is and/or of a phi, try to 338 // simplify it. 339 if ((CondInst->getOpcode() == Instruction::And || 340 CondInst->getOpcode() == Instruction::Or) && 341 isa<BranchInst>(BB->getTerminator()) && 342 ProcessBranchOnLogical(CondInst, BB, 343 CondInst->getOpcode() == Instruction::And)) 344 return true; 345 346 if (CmpInst *CondCmp = dyn_cast<CmpInst>(CondInst)) { 347 if (isa<PHINode>(CondCmp->getOperand(0))) { 348 // If we have "br (phi != 42)" and the phi node has any constant values 349 // as operands, we can thread through this block. 350 // 351 // If we have "br (cmp phi, x)" and the phi node contains x such that the 352 // comparison uniquely identifies the branch target, we can thread 353 // through this block. 354 355 if (ProcessBranchOnCompare(CondCmp, BB)) 356 return true; 357 } 358 359 // If we have a comparison, loop over the predecessors to see if there is 360 // a condition with the same value. 361 pred_iterator PI = pred_begin(BB), E = pred_end(BB); 362 for (; PI != E; ++PI) 363 if (BranchInst *PBI = dyn_cast<BranchInst>((*PI)->getTerminator())) 364 if (PBI->isConditional() && *PI != BB) { 365 if (CmpInst *CI = dyn_cast<CmpInst>(PBI->getCondition())) { 366 if (CI->getOperand(0) == CondCmp->getOperand(0) && 367 CI->getOperand(1) == CondCmp->getOperand(1) && 368 CI->getPredicate() == CondCmp->getPredicate()) { 369 // TODO: Could handle things like (x != 4) --> (x == 17) 370 if (ProcessBranchOnDuplicateCond(*PI, BB)) 371 return true; 372 } 373 } 374 } 375 } 376 377 // Check for some cases that are worth simplifying. Right now we want to look 378 // for loads that are used by a switch or by the condition for the branch. If 379 // we see one, check to see if it's partially redundant. If so, insert a PHI 380 // which can then be used to thread the values. 381 // 382 // This is particularly important because reg2mem inserts loads and stores all 383 // over the place, and this blocks jump threading if we don't zap them. 384 Value *SimplifyValue = CondInst; 385 if (CmpInst *CondCmp = dyn_cast<CmpInst>(SimplifyValue)) 386 if (isa<Constant>(CondCmp->getOperand(1))) 387 SimplifyValue = CondCmp->getOperand(0); 388 389 if (LoadInst *LI = dyn_cast<LoadInst>(SimplifyValue)) 390 if (SimplifyPartiallyRedundantLoad(LI)) 391 return true; 392 393 // TODO: If we have: "br (X > 0)" and we have a predecessor where we know 394 // "(X == 4)" thread through this block. 395 396 return false; 397} 398 399/// ProcessBranchOnDuplicateCond - We found a block and a predecessor of that 400/// block that jump on exactly the same condition. This means that we almost 401/// always know the direction of the edge in the DESTBB: 402/// PREDBB: 403/// br COND, DESTBB, BBY 404/// DESTBB: 405/// br COND, BBZ, BBW 406/// 407/// If DESTBB has multiple predecessors, we can't just constant fold the branch 408/// in DESTBB, we have to thread over it. 409bool JumpThreading::ProcessBranchOnDuplicateCond(BasicBlock *PredBB, 410 BasicBlock *BB) { 411 BranchInst *PredBI = cast<BranchInst>(PredBB->getTerminator()); 412 413 // If both successors of PredBB go to DESTBB, we don't know anything. We can 414 // fold the branch to an unconditional one, which allows other recursive 415 // simplifications. 416 bool BranchDir; 417 if (PredBI->getSuccessor(1) != BB) 418 BranchDir = true; 419 else if (PredBI->getSuccessor(0) != BB) 420 BranchDir = false; 421 else { 422 DOUT << " In block '" << PredBB->getNameStart() 423 << "' folding terminator: " << *PredBB->getTerminator(); 424 ++NumFolds; 425 ConstantFoldTerminator(PredBB); 426 return true; 427 } 428 429 BranchInst *DestBI = cast<BranchInst>(BB->getTerminator()); 430 431 // If the dest block has one predecessor, just fix the branch condition to a 432 // constant and fold it. 433 if (BB->getSinglePredecessor()) { 434 DOUT << " In block '" << BB->getNameStart() 435 << "' folding condition to '" << BranchDir << "': " 436 << *BB->getTerminator(); 437 ++NumFolds; 438 DestBI->setCondition(Context->getConstantInt(Type::Int1Ty, BranchDir)); 439 ConstantFoldTerminator(BB); 440 return true; 441 } 442 443 // Otherwise we need to thread from PredBB to DestBB's successor which 444 // involves code duplication. Check to see if it is worth it. 445 unsigned JumpThreadCost = getJumpThreadDuplicationCost(BB); 446 if (JumpThreadCost > Threshold) { 447 DOUT << " Not threading BB '" << BB->getNameStart() 448 << "' - Cost is too high: " << JumpThreadCost << "\n"; 449 return false; 450 } 451 452 // Next, figure out which successor we are threading to. 453 BasicBlock *SuccBB = DestBI->getSuccessor(!BranchDir); 454 455 // Ok, try to thread it! 456 return ThreadEdge(BB, PredBB, SuccBB, JumpThreadCost); 457} 458 459/// ProcessSwitchOnDuplicateCond - We found a block and a predecessor of that 460/// block that switch on exactly the same condition. This means that we almost 461/// always know the direction of the edge in the DESTBB: 462/// PREDBB: 463/// switch COND [... DESTBB, BBY ... ] 464/// DESTBB: 465/// switch COND [... BBZ, BBW ] 466/// 467/// Optimizing switches like this is very important, because simplifycfg builds 468/// switches out of repeated 'if' conditions. 469bool JumpThreading::ProcessSwitchOnDuplicateCond(BasicBlock *PredBB, 470 BasicBlock *DestBB) { 471 // Can't thread edge to self. 472 if (PredBB == DestBB) 473 return false; 474 475 476 SwitchInst *PredSI = cast<SwitchInst>(PredBB->getTerminator()); 477 SwitchInst *DestSI = cast<SwitchInst>(DestBB->getTerminator()); 478 479 // There are a variety of optimizations that we can potentially do on these 480 // blocks: we order them from most to least preferable. 481 482 // If DESTBB *just* contains the switch, then we can forward edges from PREDBB 483 // directly to their destination. This does not introduce *any* code size 484 // growth. Skip debug info first. 485 BasicBlock::iterator BBI = DestBB->begin(); 486 while (isa<DbgInfoIntrinsic>(BBI)) 487 BBI++; 488 489 // FIXME: Thread if it just contains a PHI. 490 if (isa<SwitchInst>(BBI)) { 491 bool MadeChange = false; 492 // Ignore the default edge for now. 493 for (unsigned i = 1, e = DestSI->getNumSuccessors(); i != e; ++i) { 494 ConstantInt *DestVal = DestSI->getCaseValue(i); 495 BasicBlock *DestSucc = DestSI->getSuccessor(i); 496 497 // Okay, DestSI has a case for 'DestVal' that goes to 'DestSucc'. See if 498 // PredSI has an explicit case for it. If so, forward. If it is covered 499 // by the default case, we can't update PredSI. 500 unsigned PredCase = PredSI->findCaseValue(DestVal); 501 if (PredCase == 0) continue; 502 503 // If PredSI doesn't go to DestBB on this value, then it won't reach the 504 // case on this condition. 505 if (PredSI->getSuccessor(PredCase) != DestBB && 506 DestSI->getSuccessor(i) != DestBB) 507 continue; 508 509 // Otherwise, we're safe to make the change. Make sure that the edge from 510 // DestSI to DestSucc is not critical and has no PHI nodes. 511 DOUT << "FORWARDING EDGE " << *DestVal << " FROM: " << *PredSI; 512 DOUT << "THROUGH: " << *DestSI; 513 514 // If the destination has PHI nodes, just split the edge for updating 515 // simplicity. 516 if (isa<PHINode>(DestSucc->begin()) && !DestSucc->getSinglePredecessor()){ 517 SplitCriticalEdge(DestSI, i, this); 518 DestSucc = DestSI->getSuccessor(i); 519 } 520 FoldSingleEntryPHINodes(DestSucc); 521 PredSI->setSuccessor(PredCase, DestSucc); 522 MadeChange = true; 523 } 524 525 if (MadeChange) 526 return true; 527 } 528 529 return false; 530} 531 532 533/// SimplifyPartiallyRedundantLoad - If LI is an obviously partially redundant 534/// load instruction, eliminate it by replacing it with a PHI node. This is an 535/// important optimization that encourages jump threading, and needs to be run 536/// interlaced with other jump threading tasks. 537bool JumpThreading::SimplifyPartiallyRedundantLoad(LoadInst *LI) { 538 // Don't hack volatile loads. 539 if (LI->isVolatile()) return false; 540 541 // If the load is defined in a block with exactly one predecessor, it can't be 542 // partially redundant. 543 BasicBlock *LoadBB = LI->getParent(); 544 if (LoadBB->getSinglePredecessor()) 545 return false; 546 547 Value *LoadedPtr = LI->getOperand(0); 548 549 // If the loaded operand is defined in the LoadBB, it can't be available. 550 // FIXME: Could do PHI translation, that would be fun :) 551 if (Instruction *PtrOp = dyn_cast<Instruction>(LoadedPtr)) 552 if (PtrOp->getParent() == LoadBB) 553 return false; 554 555 // Scan a few instructions up from the load, to see if it is obviously live at 556 // the entry to its block. 557 BasicBlock::iterator BBIt = LI; 558 559 if (Value *AvailableVal = FindAvailableLoadedValue(LoadedPtr, LoadBB, 560 BBIt, 6)) { 561 // If the value if the load is locally available within the block, just use 562 // it. This frequently occurs for reg2mem'd allocas. 563 //cerr << "LOAD ELIMINATED:\n" << *BBIt << *LI << "\n"; 564 565 // If the returned value is the load itself, replace with an undef. This can 566 // only happen in dead loops. 567 if (AvailableVal == LI) AvailableVal = Context->getUndef(LI->getType()); 568 LI->replaceAllUsesWith(AvailableVal); 569 LI->eraseFromParent(); 570 return true; 571 } 572 573 // Otherwise, if we scanned the whole block and got to the top of the block, 574 // we know the block is locally transparent to the load. If not, something 575 // might clobber its value. 576 if (BBIt != LoadBB->begin()) 577 return false; 578 579 580 SmallPtrSet<BasicBlock*, 8> PredsScanned; 581 typedef SmallVector<std::pair<BasicBlock*, Value*>, 8> AvailablePredsTy; 582 AvailablePredsTy AvailablePreds; 583 BasicBlock *OneUnavailablePred = 0; 584 585 // If we got here, the loaded value is transparent through to the start of the 586 // block. Check to see if it is available in any of the predecessor blocks. 587 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB); 588 PI != PE; ++PI) { 589 BasicBlock *PredBB = *PI; 590 591 // If we already scanned this predecessor, skip it. 592 if (!PredsScanned.insert(PredBB)) 593 continue; 594 595 // Scan the predecessor to see if the value is available in the pred. 596 BBIt = PredBB->end(); 597 Value *PredAvailable = FindAvailableLoadedValue(LoadedPtr, PredBB, BBIt, 6); 598 if (!PredAvailable) { 599 OneUnavailablePred = PredBB; 600 continue; 601 } 602 603 // If so, this load is partially redundant. Remember this info so that we 604 // can create a PHI node. 605 AvailablePreds.push_back(std::make_pair(PredBB, PredAvailable)); 606 } 607 608 // If the loaded value isn't available in any predecessor, it isn't partially 609 // redundant. 610 if (AvailablePreds.empty()) return false; 611 612 // Okay, the loaded value is available in at least one (and maybe all!) 613 // predecessors. If the value is unavailable in more than one unique 614 // predecessor, we want to insert a merge block for those common predecessors. 615 // This ensures that we only have to insert one reload, thus not increasing 616 // code size. 617 BasicBlock *UnavailablePred = 0; 618 619 // If there is exactly one predecessor where the value is unavailable, the 620 // already computed 'OneUnavailablePred' block is it. If it ends in an 621 // unconditional branch, we know that it isn't a critical edge. 622 if (PredsScanned.size() == AvailablePreds.size()+1 && 623 OneUnavailablePred->getTerminator()->getNumSuccessors() == 1) { 624 UnavailablePred = OneUnavailablePred; 625 } else if (PredsScanned.size() != AvailablePreds.size()) { 626 // Otherwise, we had multiple unavailable predecessors or we had a critical 627 // edge from the one. 628 SmallVector<BasicBlock*, 8> PredsToSplit; 629 SmallPtrSet<BasicBlock*, 8> AvailablePredSet; 630 631 for (unsigned i = 0, e = AvailablePreds.size(); i != e; ++i) 632 AvailablePredSet.insert(AvailablePreds[i].first); 633 634 // Add all the unavailable predecessors to the PredsToSplit list. 635 for (pred_iterator PI = pred_begin(LoadBB), PE = pred_end(LoadBB); 636 PI != PE; ++PI) 637 if (!AvailablePredSet.count(*PI)) 638 PredsToSplit.push_back(*PI); 639 640 // Split them out to their own block. 641 UnavailablePred = 642 SplitBlockPredecessors(LoadBB, &PredsToSplit[0], PredsToSplit.size(), 643 "thread-split", this); 644 } 645 646 // If the value isn't available in all predecessors, then there will be 647 // exactly one where it isn't available. Insert a load on that edge and add 648 // it to the AvailablePreds list. 649 if (UnavailablePred) { 650 assert(UnavailablePred->getTerminator()->getNumSuccessors() == 1 && 651 "Can't handle critical edge here!"); 652 Value *NewVal = new LoadInst(LoadedPtr, LI->getName()+".pr", 653 UnavailablePred->getTerminator()); 654 AvailablePreds.push_back(std::make_pair(UnavailablePred, NewVal)); 655 } 656 657 // Now we know that each predecessor of this block has a value in 658 // AvailablePreds, sort them for efficient access as we're walking the preds. 659 array_pod_sort(AvailablePreds.begin(), AvailablePreds.end()); 660 661 // Create a PHI node at the start of the block for the PRE'd load value. 662 PHINode *PN = PHINode::Create(LI->getType(), "", LoadBB->begin()); 663 PN->takeName(LI); 664 665 // Insert new entries into the PHI for each predecessor. A single block may 666 // have multiple entries here. 667 for (pred_iterator PI = pred_begin(LoadBB), E = pred_end(LoadBB); PI != E; 668 ++PI) { 669 AvailablePredsTy::iterator I = 670 std::lower_bound(AvailablePreds.begin(), AvailablePreds.end(), 671 std::make_pair(*PI, (Value*)0)); 672 673 assert(I != AvailablePreds.end() && I->first == *PI && 674 "Didn't find entry for predecessor!"); 675 676 PN->addIncoming(I->second, I->first); 677 } 678 679 //cerr << "PRE: " << *LI << *PN << "\n"; 680 681 LI->replaceAllUsesWith(PN); 682 LI->eraseFromParent(); 683 684 return true; 685} 686 687 688/// ProcessJumpOnPHI - We have a conditional branch of switch on a PHI node in 689/// the current block. See if there are any simplifications we can do based on 690/// inputs to the phi node. 691/// 692bool JumpThreading::ProcessJumpOnPHI(PHINode *PN) { 693 // See if the phi node has any constant values. If so, we can determine where 694 // the corresponding predecessor will branch. 695 ConstantInt *PredCst = 0; 696 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) 697 if ((PredCst = dyn_cast<ConstantInt>(PN->getIncomingValue(i)))) 698 break; 699 700 // If no incoming value has a constant, we don't know the destination of any 701 // predecessors. 702 if (PredCst == 0) 703 return false; 704 705 // See if the cost of duplicating this block is low enough. 706 BasicBlock *BB = PN->getParent(); 707 unsigned JumpThreadCost = getJumpThreadDuplicationCost(BB); 708 if (JumpThreadCost > Threshold) { 709 DOUT << " Not threading BB '" << BB->getNameStart() 710 << "' - Cost is too high: " << JumpThreadCost << "\n"; 711 return false; 712 } 713 714 // If so, we can actually do this threading. Merge any common predecessors 715 // that will act the same. 716 BasicBlock *PredBB = FactorCommonPHIPreds(PN, PredCst); 717 718 // Next, figure out which successor we are threading to. 719 BasicBlock *SuccBB; 720 if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) 721 SuccBB = BI->getSuccessor(PredCst == Context->getConstantIntFalse()); 722 else { 723 SwitchInst *SI = cast<SwitchInst>(BB->getTerminator()); 724 SuccBB = SI->getSuccessor(SI->findCaseValue(PredCst)); 725 } 726 727 // Ok, try to thread it! 728 return ThreadEdge(BB, PredBB, SuccBB, JumpThreadCost); 729} 730 731/// ProcessJumpOnLogicalPHI - PN's basic block contains a conditional branch 732/// whose condition is an AND/OR where one side is PN. If PN has constant 733/// operands that permit us to evaluate the condition for some operand, thread 734/// through the block. For example with: 735/// br (and X, phi(Y, Z, false)) 736/// the predecessor corresponding to the 'false' will always jump to the false 737/// destination of the branch. 738/// 739bool JumpThreading::ProcessBranchOnLogical(Value *V, BasicBlock *BB, 740 bool isAnd) { 741 // If this is a binary operator tree of the same AND/OR opcode, check the 742 // LHS/RHS. 743 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(V)) 744 if ((isAnd && BO->getOpcode() == Instruction::And) || 745 (!isAnd && BO->getOpcode() == Instruction::Or)) { 746 if (ProcessBranchOnLogical(BO->getOperand(0), BB, isAnd)) 747 return true; 748 if (ProcessBranchOnLogical(BO->getOperand(1), BB, isAnd)) 749 return true; 750 } 751 752 // If this isn't a PHI node, we can't handle it. 753 PHINode *PN = dyn_cast<PHINode>(V); 754 if (!PN || PN->getParent() != BB) return false; 755 756 // We can only do the simplification for phi nodes of 'false' with AND or 757 // 'true' with OR. See if we have any entries in the phi for this. 758 unsigned PredNo = ~0U; 759 ConstantInt *PredCst = Context->getConstantInt(Type::Int1Ty, !isAnd); 760 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 761 if (PN->getIncomingValue(i) == PredCst) { 762 PredNo = i; 763 break; 764 } 765 } 766 767 // If no match, bail out. 768 if (PredNo == ~0U) 769 return false; 770 771 // See if the cost of duplicating this block is low enough. 772 unsigned JumpThreadCost = getJumpThreadDuplicationCost(BB); 773 if (JumpThreadCost > Threshold) { 774 DOUT << " Not threading BB '" << BB->getNameStart() 775 << "' - Cost is too high: " << JumpThreadCost << "\n"; 776 return false; 777 } 778 779 // If so, we can actually do this threading. Merge any common predecessors 780 // that will act the same. 781 BasicBlock *PredBB = FactorCommonPHIPreds(PN, PredCst); 782 783 // Next, figure out which successor we are threading to. If this was an AND, 784 // the constant must be FALSE, and we must be targeting the 'false' block. 785 // If this is an OR, the constant must be TRUE, and we must be targeting the 786 // 'true' block. 787 BasicBlock *SuccBB = BB->getTerminator()->getSuccessor(isAnd); 788 789 // Ok, try to thread it! 790 return ThreadEdge(BB, PredBB, SuccBB, JumpThreadCost); 791} 792 793/// GetResultOfComparison - Given an icmp/fcmp predicate and the left and right 794/// hand sides of the compare instruction, try to determine the result. If the 795/// result can not be determined, a null pointer is returned. 796static Constant *GetResultOfComparison(CmpInst::Predicate pred, 797 Value *LHS, Value *RHS, 798 LLVMContext* Context) { 799 if (Constant *CLHS = dyn_cast<Constant>(LHS)) 800 if (Constant *CRHS = dyn_cast<Constant>(RHS)) 801 return Context->getConstantExprCompare(pred, CLHS, CRHS); 802 803 if (LHS == RHS) 804 if (isa<IntegerType>(LHS->getType()) || isa<PointerType>(LHS->getType())) 805 return ICmpInst::isTrueWhenEqual(pred) ? 806 Context->getConstantIntTrue() : Context->getConstantIntFalse(); 807 808 return 0; 809} 810 811/// ProcessBranchOnCompare - We found a branch on a comparison between a phi 812/// node and a value. If we can identify when the comparison is true between 813/// the phi inputs and the value, we can fold the compare for that edge and 814/// thread through it. 815bool JumpThreading::ProcessBranchOnCompare(CmpInst *Cmp, BasicBlock *BB) { 816 PHINode *PN = cast<PHINode>(Cmp->getOperand(0)); 817 Value *RHS = Cmp->getOperand(1); 818 819 // If the phi isn't in the current block, an incoming edge to this block 820 // doesn't control the destination. 821 if (PN->getParent() != BB) 822 return false; 823 824 // We can do this simplification if any comparisons fold to true or false. 825 // See if any do. 826 Value *PredVal = 0; 827 bool TrueDirection = false; 828 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 829 PredVal = PN->getIncomingValue(i); 830 831 Constant *Res = GetResultOfComparison(Cmp->getPredicate(), PredVal, 832 RHS, Context); 833 if (!Res) { 834 PredVal = 0; 835 continue; 836 } 837 838 // If this folded to a constant expr, we can't do anything. 839 if (ConstantInt *ResC = dyn_cast<ConstantInt>(Res)) { 840 TrueDirection = ResC->getZExtValue(); 841 break; 842 } 843 // If this folded to undef, just go the false way. 844 if (isa<UndefValue>(Res)) { 845 TrueDirection = false; 846 break; 847 } 848 849 // Otherwise, we can't fold this input. 850 PredVal = 0; 851 } 852 853 // If no match, bail out. 854 if (PredVal == 0) 855 return false; 856 857 // See if the cost of duplicating this block is low enough. 858 unsigned JumpThreadCost = getJumpThreadDuplicationCost(BB); 859 if (JumpThreadCost > Threshold) { 860 DOUT << " Not threading BB '" << BB->getNameStart() 861 << "' - Cost is too high: " << JumpThreadCost << "\n"; 862 return false; 863 } 864 865 // If so, we can actually do this threading. Merge any common predecessors 866 // that will act the same. 867 BasicBlock *PredBB = FactorCommonPHIPreds(PN, PredVal); 868 869 // Next, get our successor. 870 BasicBlock *SuccBB = BB->getTerminator()->getSuccessor(!TrueDirection); 871 872 // Ok, try to thread it! 873 return ThreadEdge(BB, PredBB, SuccBB, JumpThreadCost); 874} 875 876 877/// ThreadEdge - We have decided that it is safe and profitable to thread an 878/// edge from PredBB to SuccBB across BB. Transform the IR to reflect this 879/// change. 880bool JumpThreading::ThreadEdge(BasicBlock *BB, BasicBlock *PredBB, 881 BasicBlock *SuccBB, unsigned JumpThreadCost) { 882 883 // If threading to the same block as we come from, we would infinite loop. 884 if (SuccBB == BB) { 885 DOUT << " Not threading across BB '" << BB->getNameStart() 886 << "' - would thread to self!\n"; 887 return false; 888 } 889 890 // If threading this would thread across a loop header, don't thread the edge. 891 // See the comments above FindLoopHeaders for justifications and caveats. 892 if (LoopHeaders.count(BB)) { 893 DOUT << " Not threading from '" << PredBB->getNameStart() 894 << "' across loop header BB '" << BB->getNameStart() 895 << "' to dest BB '" << SuccBB->getNameStart() 896 << "' - it might create an irreducible loop!\n"; 897 return false; 898 } 899 900 // And finally, do it! 901 DOUT << " Threading edge from '" << PredBB->getNameStart() << "' to '" 902 << SuccBB->getNameStart() << "' with cost: " << JumpThreadCost 903 << ", across block:\n " 904 << *BB << "\n"; 905 906 // Jump Threading can not update SSA properties correctly if the values 907 // defined in the duplicated block are used outside of the block itself. For 908 // this reason, we spill all values that are used outside of BB to the stack. 909 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) { 910 if (!I->isUsedOutsideOfBlock(BB)) 911 continue; 912 913 // We found a use of I outside of BB. Create a new stack slot to 914 // break this inter-block usage pattern. 915 DemoteRegToStack(*I); 916 } 917 918 // We are going to have to map operands from the original BB block to the new 919 // copy of the block 'NewBB'. If there are PHI nodes in BB, evaluate them to 920 // account for entry from PredBB. 921 DenseMap<Instruction*, Value*> ValueMapping; 922 923 BasicBlock *NewBB = 924 BasicBlock::Create(BB->getName()+".thread", BB->getParent(), BB); 925 NewBB->moveAfter(PredBB); 926 927 BasicBlock::iterator BI = BB->begin(); 928 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI) 929 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB); 930 931 // Clone the non-phi instructions of BB into NewBB, keeping track of the 932 // mapping and using it to remap operands in the cloned instructions. 933 for (; !isa<TerminatorInst>(BI); ++BI) { 934 Instruction *New = BI->clone(); 935 New->setName(BI->getNameStart()); 936 NewBB->getInstList().push_back(New); 937 ValueMapping[BI] = New; 938 939 // Remap operands to patch up intra-block references. 940 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i) 941 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) { 942 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst); 943 if (I != ValueMapping.end()) 944 New->setOperand(i, I->second); 945 } 946 } 947 948 // We didn't copy the terminator from BB over to NewBB, because there is now 949 // an unconditional jump to SuccBB. Insert the unconditional jump. 950 BranchInst::Create(SuccBB, NewBB); 951 952 // Check to see if SuccBB has PHI nodes. If so, we need to add entries to the 953 // PHI nodes for NewBB now. 954 for (BasicBlock::iterator PNI = SuccBB->begin(); isa<PHINode>(PNI); ++PNI) { 955 PHINode *PN = cast<PHINode>(PNI); 956 // Ok, we have a PHI node. Figure out what the incoming value was for the 957 // DestBlock. 958 Value *IV = PN->getIncomingValueForBlock(BB); 959 960 // Remap the value if necessary. 961 if (Instruction *Inst = dyn_cast<Instruction>(IV)) { 962 DenseMap<Instruction*, Value*>::iterator I = ValueMapping.find(Inst); 963 if (I != ValueMapping.end()) 964 IV = I->second; 965 } 966 PN->addIncoming(IV, NewBB); 967 } 968 969 // Ok, NewBB is good to go. Update the terminator of PredBB to jump to 970 // NewBB instead of BB. This eliminates predecessors from BB, which requires 971 // us to simplify any PHI nodes in BB. 972 TerminatorInst *PredTerm = PredBB->getTerminator(); 973 for (unsigned i = 0, e = PredTerm->getNumSuccessors(); i != e; ++i) 974 if (PredTerm->getSuccessor(i) == BB) { 975 BB->removePredecessor(PredBB); 976 PredTerm->setSuccessor(i, NewBB); 977 } 978 979 // At this point, the IR is fully up to date and consistent. Do a quick scan 980 // over the new instructions and zap any that are constants or dead. This 981 // frequently happens because of phi translation. 982 BI = NewBB->begin(); 983 for (BasicBlock::iterator E = NewBB->end(); BI != E; ) { 984 Instruction *Inst = BI++; 985 if (Constant *C = ConstantFoldInstruction(Inst, BB->getContext(), TD)) { 986 Inst->replaceAllUsesWith(C); 987 Inst->eraseFromParent(); 988 continue; 989 } 990 991 RecursivelyDeleteTriviallyDeadInstructions(Inst); 992 } 993 994 // Threaded an edge! 995 ++NumThreads; 996 return true; 997} 998