JumpThreading.cpp revision 43e2a035309f4e353a8bd5547d10125414597e74
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/Pass.h" 18#include "llvm/ADT/DenseMap.h" 19#include "llvm/ADT/Statistic.h" 20#include "llvm/Transforms/Utils/BasicBlockUtils.h" 21#include "llvm/Transforms/Utils/Local.h" 22#include "llvm/Support/CommandLine.h" 23#include "llvm/Support/Compiler.h" 24#include "llvm/Support/Debug.h" 25using namespace llvm; 26 27STATISTIC(NumThreads, "Number of jumps threaded"); 28STATISTIC(NumFolds, "Number of terminators folded"); 29 30static cl::opt<unsigned> 31Threshold("jump-threading-threshold", 32 cl::desc("Max block size to duplicate for jump threading"), 33 cl::init(6), cl::Hidden); 34 35namespace { 36 /// This pass performs 'jump threading', which looks at blocks that have 37 /// multiple predecessors and multiple successors. If one or more of the 38 /// predecessors of the block can be proven to always jump to one of the 39 /// successors, we forward the edge from the predecessor to the successor by 40 /// duplicating the contents of this block. 41 /// 42 /// An example of when this can occur is code like this: 43 /// 44 /// if () { ... 45 /// X = 4; 46 /// } 47 /// if (X < 3) { 48 /// 49 /// In this case, the unconditional branch at the end of the first if can be 50 /// revectored to the false side of the second if. 51 /// 52 class VISIBILITY_HIDDEN JumpThreading : public FunctionPass { 53 public: 54 static char ID; // Pass identification 55 JumpThreading() : FunctionPass((intptr_t)&ID) {} 56 57 bool runOnFunction(Function &F); 58 bool ThreadBlock(BasicBlock *BB); 59 void ThreadEdge(BasicBlock *BB, BasicBlock *PredBB, BasicBlock *SuccBB); 60 BasicBlock *FactorCommonPHIPreds(PHINode *PN, Constant *CstVal); 61 62 bool ProcessJumpOnPHI(PHINode *PN); 63 bool ProcessBranchOnLogical(Value *V, BasicBlock *BB, bool isAnd); 64 bool ProcessBranchOnCompare(CmpInst *Cmp, BasicBlock *BB); 65 }; 66} 67 68char JumpThreading::ID = 0; 69static RegisterPass<JumpThreading> 70X("jump-threading", "Jump Threading"); 71 72// Public interface to the Jump Threading pass 73FunctionPass *llvm::createJumpThreadingPass() { return new JumpThreading(); } 74 75/// runOnFunction - Top level algorithm. 76/// 77bool JumpThreading::runOnFunction(Function &F) { 78 DOUT << "Jump threading on function '" << F.getNameStart() << "'\n"; 79 80 bool AnotherIteration = true, EverChanged = false; 81 while (AnotherIteration) { 82 AnotherIteration = false; 83 bool Changed = false; 84 for (Function::iterator I = F.begin(), E = F.end(); I != E; ++I) 85 while (ThreadBlock(I)) 86 Changed = true; 87 AnotherIteration = Changed; 88 EverChanged |= Changed; 89 } 90 return EverChanged; 91} 92 93/// FactorCommonPHIPreds - If there are multiple preds with the same incoming 94/// value for the PHI, factor them together so we get one block to thread for 95/// the whole group. 96/// This is important for things like "phi i1 [true, true, false, true, x]" 97/// where we only need to clone the block for the true blocks once. 98/// 99BasicBlock *JumpThreading::FactorCommonPHIPreds(PHINode *PN, Constant *CstVal) { 100 SmallVector<BasicBlock*, 16> CommonPreds; 101 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) 102 if (PN->getIncomingValue(i) == CstVal) 103 CommonPreds.push_back(PN->getIncomingBlock(i)); 104 105 if (CommonPreds.size() == 1) 106 return CommonPreds[0]; 107 108 DOUT << " Factoring out " << CommonPreds.size() 109 << " common predecessors.\n"; 110 return SplitBlockPredecessors(PN->getParent(), 111 &CommonPreds[0], CommonPreds.size(), 112 ".thr_comm", this); 113} 114 115 116/// getJumpThreadDuplicationCost - Return the cost of duplicating this block to 117/// thread across it. 118static unsigned getJumpThreadDuplicationCost(const BasicBlock *BB) { 119 /// Ignore PHI nodes, these will be flattened when duplication happens. 120 BasicBlock::const_iterator I = BB->getFirstNonPHI(); 121 122 // Sum up the cost of each instruction until we get to the terminator. Don't 123 // include the terminator because the copy won't include it. 124 unsigned Size = 0; 125 for (; !isa<TerminatorInst>(I); ++I) { 126 // Debugger intrinsics don't incur code size. 127 if (isa<DbgInfoIntrinsic>(I)) continue; 128 129 // If this is a pointer->pointer bitcast, it is free. 130 if (isa<BitCastInst>(I) && isa<PointerType>(I->getType())) 131 continue; 132 133 // All other instructions count for at least one unit. 134 ++Size; 135 136 // Calls are more expensive. If they are non-intrinsic calls, we model them 137 // as having cost of 4. If they are a non-vector intrinsic, we model them 138 // as having cost of 2 total, and if they are a vector intrinsic, we model 139 // them as having cost 1. 140 if (const CallInst *CI = dyn_cast<CallInst>(I)) { 141 if (!isa<IntrinsicInst>(CI)) 142 Size += 3; 143 else if (isa<VectorType>(CI->getType())) 144 Size += 1; 145 } 146 } 147 148 // Threading through a switch statement is particularly profitable. If this 149 // block ends in a switch, decrease its cost to make it more likely to happen. 150 if (isa<SwitchInst>(I)) 151 Size = Size > 6 ? Size-6 : 0; 152 153 return Size; 154} 155 156 157/// ThreadBlock - If there are any predecessors whose control can be threaded 158/// through to a successor, transform them now. 159bool JumpThreading::ThreadBlock(BasicBlock *BB) { 160 // See if this block ends with a branch or switch. If so, see if the 161 // condition is a phi node. If so, and if an entry of the phi node is a 162 // constant, we can thread the block. 163 Value *Condition; 164 if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) { 165 // Can't thread an unconditional jump. 166 if (BI->isUnconditional()) return false; 167 Condition = BI->getCondition(); 168 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator())) 169 Condition = SI->getCondition(); 170 else 171 return false; // Must be an invoke. 172 173 // If the terminator of this block is branching on a constant, simplify the 174 // terminator to an unconditional branch. This can occur due to threading in 175 // other blocks. 176 if (isa<ConstantInt>(Condition)) { 177 DOUT << " In block '" << BB->getNameStart() 178 << "' folding terminator: " << *BB->getTerminator(); 179 ++NumFolds; 180 ConstantFoldTerminator(BB); 181 return true; 182 } 183 184 // If there is only a single predecessor of this block, nothing to fold. 185 if (BB->getSinglePredecessor()) 186 return false; 187 188 // See if this is a phi node in the current block. 189 PHINode *PN = dyn_cast<PHINode>(Condition); 190 if (PN && PN->getParent() == BB) 191 return ProcessJumpOnPHI(PN); 192 193 // If this is a conditional branch whose condition is and/or of a phi, try to 194 // simplify it. 195 if (BinaryOperator *CondI = dyn_cast<BinaryOperator>(Condition)) { 196 if ((CondI->getOpcode() == Instruction::And || 197 CondI->getOpcode() == Instruction::Or) && 198 isa<BranchInst>(BB->getTerminator()) && 199 ProcessBranchOnLogical(CondI, BB, 200 CondI->getOpcode() == Instruction::And)) 201 return true; 202 } 203 204 // If we have "br (phi != 42)" and the phi node has any constant values as 205 // operands, we can thread through this block. 206 if (CmpInst *CondCmp = dyn_cast<CmpInst>(Condition)) 207 if (isa<PHINode>(CondCmp->getOperand(0)) && 208 isa<Constant>(CondCmp->getOperand(1)) && 209 ProcessBranchOnCompare(CondCmp, BB)) 210 return true; 211 212 return false; 213} 214 215/// ProcessJumpOnPHI - We have a conditional branch of switch on a PHI node in 216/// the current block. See if there are any simplifications we can do based on 217/// inputs to the phi node. 218/// 219bool JumpThreading::ProcessJumpOnPHI(PHINode *PN) { 220 // See if the phi node has any constant values. If so, we can determine where 221 // the corresponding predecessor will branch. 222 ConstantInt *PredCst = 0; 223 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) 224 if ((PredCst = dyn_cast<ConstantInt>(PN->getIncomingValue(i)))) 225 break; 226 227 // If no incoming value has a constant, we don't know the destination of any 228 // predecessors. 229 if (PredCst == 0) 230 return false; 231 232 // See if the cost of duplicating this block is low enough. 233 BasicBlock *BB = PN->getParent(); 234 unsigned JumpThreadCost = getJumpThreadDuplicationCost(BB); 235 if (JumpThreadCost > Threshold) { 236 DOUT << " Not threading BB '" << BB->getNameStart() 237 << "' - Cost is too high: " << JumpThreadCost << "\n"; 238 return false; 239 } 240 241 // If so, we can actually do this threading. Merge any common predecessors 242 // that will act the same. 243 BasicBlock *PredBB = FactorCommonPHIPreds(PN, PredCst); 244 245 // Next, figure out which successor we are threading to. 246 BasicBlock *SuccBB; 247 if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) 248 SuccBB = BI->getSuccessor(PredCst == ConstantInt::getFalse()); 249 else { 250 SwitchInst *SI = cast<SwitchInst>(BB->getTerminator()); 251 SuccBB = SI->getSuccessor(SI->findCaseValue(PredCst)); 252 } 253 254 // If threading to the same block as we come from, we would infinite loop. 255 if (SuccBB == BB) { 256 DOUT << " Not threading BB '" << BB->getNameStart() 257 << "' - would thread to self!\n"; 258 return false; 259 } 260 261 // And finally, do it! 262 DOUT << " Threading edge from '" << PredBB->getNameStart() << "' to '" 263 << SuccBB->getNameStart() << "' with cost: " << JumpThreadCost 264 << ", across block:\n " 265 << *BB << "\n"; 266 267 ThreadEdge(BB, PredBB, SuccBB); 268 ++NumThreads; 269 return true; 270} 271 272/// ProcessJumpOnLogicalPHI - PN's basic block contains a conditional branch 273/// whose condition is an AND/OR where one side is PN. If PN has constant 274/// operands that permit us to evaluate the condition for some operand, thread 275/// through the block. For example with: 276/// br (and X, phi(Y, Z, false)) 277/// the predecessor corresponding to the 'false' will always jump to the false 278/// destination of the branch. 279/// 280bool JumpThreading::ProcessBranchOnLogical(Value *V, BasicBlock *BB, 281 bool isAnd) { 282 // If this is a binary operator tree of the same AND/OR opcode, check the 283 // LHS/RHS. 284 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(V)) 285 if ((isAnd && BO->getOpcode() == Instruction::And) || 286 (!isAnd && BO->getOpcode() == Instruction::Or)) { 287 if (ProcessBranchOnLogical(BO->getOperand(0), BB, isAnd)) 288 return true; 289 if (ProcessBranchOnLogical(BO->getOperand(1), BB, isAnd)) 290 return true; 291 } 292 293 // If this isn't a PHI node, we can't handle it. 294 PHINode *PN = dyn_cast<PHINode>(V); 295 if (!PN || PN->getParent() != BB) return false; 296 297 // We can only do the simplification for phi nodes of 'false' with AND or 298 // 'true' with OR. See if we have any entries in the phi for this. 299 unsigned PredNo = ~0U; 300 ConstantInt *PredCst = ConstantInt::get(Type::Int1Ty, !isAnd); 301 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 302 if (PN->getIncomingValue(i) == PredCst) { 303 PredNo = i; 304 break; 305 } 306 } 307 308 // If no match, bail out. 309 if (PredNo == ~0U) 310 return false; 311 312 // See if the cost of duplicating this block is low enough. 313 unsigned JumpThreadCost = getJumpThreadDuplicationCost(BB); 314 if (JumpThreadCost > Threshold) { 315 DOUT << " Not threading BB '" << BB->getNameStart() 316 << "' - Cost is too high: " << JumpThreadCost << "\n"; 317 return false; 318 } 319 320 // If so, we can actually do this threading. Merge any common predecessors 321 // that will act the same. 322 BasicBlock *PredBB = FactorCommonPHIPreds(PN, PredCst); 323 324 // Next, figure out which successor we are threading to. If this was an AND, 325 // the constant must be FALSE, and we must be targeting the 'false' block. 326 // If this is an OR, the constant must be TRUE, and we must be targeting the 327 // 'true' block. 328 BasicBlock *SuccBB = BB->getTerminator()->getSuccessor(isAnd); 329 330 // If threading to the same block as we come from, we would infinite loop. 331 if (SuccBB == BB) { 332 DOUT << " Not threading BB '" << BB->getNameStart() 333 << "' - would thread to self!\n"; 334 return false; 335 } 336 337 // And finally, do it! 338 DOUT << " Threading edge through bool from '" << PredBB->getNameStart() 339 << "' to '" << SuccBB->getNameStart() << "' with cost: " 340 << JumpThreadCost << ", across block:\n " 341 << *BB << "\n"; 342 343 ThreadEdge(BB, PredBB, SuccBB); 344 ++NumThreads; 345 return true; 346} 347 348/// ProcessBranchOnCompare - We found a branch on a comparison between a phi 349/// node and a constant. If the PHI node contains any constants as inputs, we 350/// can fold the compare for that edge and thread through it. 351bool JumpThreading::ProcessBranchOnCompare(CmpInst *Cmp, BasicBlock *BB) { 352 PHINode *PN = cast<PHINode>(Cmp->getOperand(0)); 353 Constant *RHS = cast<Constant>(Cmp->getOperand(1)); 354 355 // If the phi isn't in the current block, an incoming edge to this block 356 // doesn't control the destination. 357 if (PN->getParent() != BB) 358 return false; 359 360 // We can do this simplification if any comparisons fold to true or false. 361 // See if any do. 362 Constant *PredCst = 0; 363 bool TrueDirection = false; 364 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 365 PredCst = dyn_cast<Constant>(PN->getIncomingValue(i)); 366 if (PredCst == 0) continue; 367 368 Constant *Res; 369 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Cmp)) 370 Res = ConstantExpr::getICmp(ICI->getPredicate(), PredCst, RHS); 371 else 372 Res = ConstantExpr::getFCmp(cast<FCmpInst>(Cmp)->getPredicate(), 373 PredCst, RHS); 374 // If this folded to a constant expr, we can't do anything. 375 if (ConstantInt *ResC = dyn_cast<ConstantInt>(Res)) { 376 TrueDirection = ResC->getZExtValue(); 377 break; 378 } 379 // If this folded to undef, just go the false way. 380 if (isa<UndefValue>(Res)) { 381 TrueDirection = false; 382 break; 383 } 384 385 // Otherwise, we can't fold this input. 386 PredCst = 0; 387 } 388 389 // If no match, bail out. 390 if (PredCst == 0) 391 return false; 392 393 // See if the cost of duplicating this block is low enough. 394 unsigned JumpThreadCost = getJumpThreadDuplicationCost(BB); 395 if (JumpThreadCost > Threshold) { 396 DOUT << " Not threading BB '" << BB->getNameStart() 397 << "' - Cost is too high: " << JumpThreadCost << "\n"; 398 return false; 399 } 400 401 // If so, we can actually do this threading. Merge any common predecessors 402 // that will act the same. 403 BasicBlock *PredBB = FactorCommonPHIPreds(PN, PredCst); 404 405 // Next, get our successor. 406 BasicBlock *SuccBB = BB->getTerminator()->getSuccessor(!TrueDirection); 407 408 // If threading to the same block as we come from, we would infinite loop. 409 if (SuccBB == BB) { 410 DOUT << " Not threading BB '" << BB->getNameStart() 411 << "' - would thread to self!\n"; 412 return false; 413 } 414 415 416 // And finally, do it! 417 DOUT << " Threading edge through bool from '" << PredBB->getNameStart() 418 << "' to '" << SuccBB->getNameStart() << "' with cost: " 419 << JumpThreadCost << ", across block:\n " 420 << *BB << "\n"; 421 422 ThreadEdge(BB, PredBB, SuccBB); 423 ++NumThreads; 424 return true; 425} 426 427 428/// ThreadEdge - We have decided that it is safe and profitable to thread an 429/// edge from PredBB to SuccBB across BB. Transform the IR to reflect this 430/// change. 431void JumpThreading::ThreadEdge(BasicBlock *BB, BasicBlock *PredBB, 432 BasicBlock *SuccBB) { 433 434 // Jump Threading can not update SSA properties correctly if the values 435 // defined in the duplicated block are used outside of the block itself. For 436 // this reason, we spill all values that are used outside of BB to the stack. 437 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) { 438 if (!I->isUsedOutsideOfBlock(BB)) 439 continue; 440 441 // We found a use of I outside of BB. Create a new stack slot to 442 // break this inter-block usage pattern. 443 if (!isa<StructType>(I->getType())) { 444 DemoteRegToStack(*I); 445 continue; 446 } 447 448 // Alternatively, I must be a call or invoke that returns multiple retvals. 449 // We can't use 'DemoteRegToStack' because that will create loads and 450 // stores of aggregates which is not valid yet. If I is a call, we can just 451 // pull all the getresult instructions up to this block. If I is an invoke, 452 // we are out of luck. 453 BasicBlock::iterator IP = I; ++IP; 454 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); 455 UI != E; ++UI) 456 cast<GetResultInst>(UI)->moveBefore(IP); 457 } 458 459 // We are going to have to map operands from the original BB block to the new 460 // copy of the block 'NewBB'. If there are PHI nodes in BB, evaluate them to 461 // account for entry from PredBB. 462 DenseMap<Instruction*, Value*> ValueMapping; 463 464 BasicBlock *NewBB = 465 BasicBlock::Create(BB->getName()+".thread", BB->getParent(), BB); 466 NewBB->moveAfter(PredBB); 467 468 BasicBlock::iterator BI = BB->begin(); 469 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI) 470 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB); 471 472 // Clone the non-phi instructions of BB into NewBB, keeping track of the 473 // mapping and using it to remap operands in the cloned instructions. 474 for (; !isa<TerminatorInst>(BI); ++BI) { 475 Instruction *New = BI->clone(); 476 New->setName(BI->getNameStart()); 477 NewBB->getInstList().push_back(New); 478 ValueMapping[BI] = New; 479 480 // Remap operands to patch up intra-block references. 481 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i) 482 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) 483 if (Value *Remapped = ValueMapping[Inst]) 484 New->setOperand(i, Remapped); 485 } 486 487 // We didn't copy the terminator from BB over to NewBB, because there is now 488 // an unconditional jump to SuccBB. Insert the unconditional jump. 489 BranchInst::Create(SuccBB, NewBB); 490 491 // Check to see if SuccBB has PHI nodes. If so, we need to add entries to the 492 // PHI nodes for NewBB now. 493 for (BasicBlock::iterator PNI = SuccBB->begin(); isa<PHINode>(PNI); ++PNI) { 494 PHINode *PN = cast<PHINode>(PNI); 495 // Ok, we have a PHI node. Figure out what the incoming value was for the 496 // DestBlock. 497 Value *IV = PN->getIncomingValueForBlock(BB); 498 499 // Remap the value if necessary. 500 if (Instruction *Inst = dyn_cast<Instruction>(IV)) 501 if (Value *MappedIV = ValueMapping[Inst]) 502 IV = MappedIV; 503 PN->addIncoming(IV, NewBB); 504 } 505 506 // Finally, NewBB is good to go. Update the terminator of PredBB to jump to 507 // NewBB instead of BB. This eliminates predecessors from BB, which requires 508 // us to simplify any PHI nodes in BB. 509 TerminatorInst *PredTerm = PredBB->getTerminator(); 510 for (unsigned i = 0, e = PredTerm->getNumSuccessors(); i != e; ++i) 511 if (PredTerm->getSuccessor(i) == BB) { 512 BB->removePredecessor(PredBB); 513 PredTerm->setSuccessor(i, NewBB); 514 } 515} 516