JumpThreading.cpp revision a5ddb59a1319ccd23844c74809a64bc4d88f59d1
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 char JumpThreading::ID = 0; 67 RegisterPass<JumpThreading> X("jump-threading", "Jump Threading"); 68} 69 70// Public interface to the Jump Threading pass 71FunctionPass *llvm::createJumpThreadingPass() { return new JumpThreading(); } 72 73/// runOnFunction - Top level algorithm. 74/// 75bool JumpThreading::runOnFunction(Function &F) { 76 DOUT << "Jump threading on function '" << F.getNameStart() << "'\n"; 77 78 bool AnotherIteration = true, EverChanged = false; 79 while (AnotherIteration) { 80 AnotherIteration = false; 81 bool Changed = false; 82 for (Function::iterator I = F.begin(), E = F.end(); I != E; ++I) 83 while (ThreadBlock(I)) 84 Changed = true; 85 AnotherIteration = Changed; 86 EverChanged |= Changed; 87 } 88 return EverChanged; 89} 90 91/// FactorCommonPHIPreds - If there are multiple preds with the same incoming 92/// value for the PHI, factor them together so we get one block to thread for 93/// the whole group. 94/// This is important for things like "phi i1 [true, true, false, true, x]" 95/// where we only need to clone the block for the true blocks once. 96/// 97BasicBlock *JumpThreading::FactorCommonPHIPreds(PHINode *PN, Constant *CstVal) { 98 SmallVector<BasicBlock*, 16> CommonPreds; 99 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) 100 if (PN->getIncomingValue(i) == CstVal) 101 CommonPreds.push_back(PN->getIncomingBlock(i)); 102 103 if (CommonPreds.size() == 1) 104 return CommonPreds[0]; 105 106 DOUT << " Factoring out " << CommonPreds.size() 107 << " common predecessors.\n"; 108 return SplitBlockPredecessors(PN->getParent(), 109 &CommonPreds[0], CommonPreds.size(), 110 ".thr_comm", this); 111} 112 113 114/// getJumpThreadDuplicationCost - Return the cost of duplicating this block to 115/// thread across it. 116static unsigned getJumpThreadDuplicationCost(const BasicBlock *BB) { 117 BasicBlock::const_iterator I = BB->begin(); 118 /// Ignore PHI nodes, these will be flattened when duplication happens. 119 while (isa<PHINode>(*I)) ++I; 120 121 // Sum up the cost of each instruction until we get to the terminator. Don't 122 // include the terminator because the copy won't include it. 123 unsigned Size = 0; 124 for (; !isa<TerminatorInst>(I); ++I) { 125 // Debugger intrinsics don't incur code size. 126 if (isa<DbgInfoIntrinsic>(I)) continue; 127 128 // If this is a pointer->pointer bitcast, it is free. 129 if (isa<BitCastInst>(I) && isa<PointerType>(I->getType())) 130 continue; 131 132 // All other instructions count for at least one unit. 133 ++Size; 134 135 // Calls are more expensive. If they are non-intrinsic calls, we model them 136 // as having cost of 4. If they are a non-vector intrinsic, we model them 137 // as having cost of 2 total, and if they are a vector intrinsic, we model 138 // them as having cost 1. 139 if (const CallInst *CI = dyn_cast<CallInst>(I)) { 140 if (!isa<IntrinsicInst>(CI)) 141 Size += 3; 142 else if (isa<VectorType>(CI->getType())) 143 Size += 1; 144 } 145 } 146 147 // Threading through a switch statement is particularly profitable. If this 148 // block ends in a switch, decrease its cost to make it more likely to happen. 149 if (isa<SwitchInst>(I)) 150 Size = Size > 6 ? Size-6 : 0; 151 152 return Size; 153} 154 155 156/// ThreadBlock - If there are any predecessors whose control can be threaded 157/// through to a successor, transform them now. 158bool JumpThreading::ThreadBlock(BasicBlock *BB) { 159 // See if this block ends with a branch of switch. If so, see if the 160 // condition is a phi node. If so, and if an entry of the phi node is a 161 // constant, we can thread the block. 162 Value *Condition; 163 if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) { 164 // Can't thread an unconditional jump. 165 if (BI->isUnconditional()) return false; 166 Condition = BI->getCondition(); 167 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator())) 168 Condition = SI->getCondition(); 169 else 170 return false; // Must be an invoke. 171 172 // If the terminator of this block is branching on a constant, simplify the 173 // terminator to an unconditional branch. This can occur due to threading in 174 // other blocks. 175 if (isa<ConstantInt>(Condition)) { 176 DOUT << " In block '" << BB->getNameStart() 177 << "' folding terminator: " << *BB->getTerminator(); 178 ++NumFolds; 179 ConstantFoldTerminator(BB); 180 return true; 181 } 182 183 // If there is only a single predecessor of this block, nothing to fold. 184 if (BB->getSinglePredecessor()) 185 return false; 186 187 // See if this is a phi node in the current block. 188 PHINode *PN = dyn_cast<PHINode>(Condition); 189 if (PN && PN->getParent() == BB) 190 return ProcessJumpOnPHI(PN); 191 192 // If this is a conditional branch whose condition is and/or of a phi, try to 193 // simplify it. 194 if (BinaryOperator *CondI = dyn_cast<BinaryOperator>(Condition)) { 195 if ((CondI->getOpcode() == Instruction::And || 196 CondI->getOpcode() == Instruction::Or) && 197 isa<BranchInst>(BB->getTerminator()) && 198 ProcessBranchOnLogical(CondI, BB, 199 CondI->getOpcode() == Instruction::And)) 200 return true; 201 } 202 203 // If we have "br (phi != 42)" and the phi node has any constant values as 204 // operands, we can thread through this block. 205 if (CmpInst *CondCmp = dyn_cast<CmpInst>(Condition)) 206 if (isa<PHINode>(CondCmp->getOperand(0)) && 207 isa<Constant>(CondCmp->getOperand(1)) && 208 ProcessBranchOnCompare(CondCmp, BB)) 209 return true; 210 211 return false; 212} 213 214/// ProcessJumpOnPHI - We have a conditional branch of switch on a PHI node in 215/// the current block. See if there are any simplifications we can do based on 216/// inputs to the phi node. 217/// 218bool JumpThreading::ProcessJumpOnPHI(PHINode *PN) { 219 // See if the phi node has any constant values. If so, we can determine where 220 // the corresponding predecessor will branch. 221 ConstantInt *PredCst = 0; 222 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) 223 if ((PredCst = dyn_cast<ConstantInt>(PN->getIncomingValue(i)))) 224 break; 225 226 // If no incoming value has a constant, we don't know the destination of any 227 // predecessors. 228 if (PredCst == 0) 229 return false; 230 231 // See if the cost of duplicating this block is low enough. 232 BasicBlock *BB = PN->getParent(); 233 unsigned JumpThreadCost = getJumpThreadDuplicationCost(BB); 234 if (JumpThreadCost > Threshold) { 235 DOUT << " Not threading BB '" << BB->getNameStart() 236 << "' - Cost is too high: " << JumpThreadCost << "\n"; 237 return false; 238 } 239 240 // If so, we can actually do this threading. Merge any common predecessors 241 // that will act the same. 242 BasicBlock *PredBB = FactorCommonPHIPreds(PN, PredCst); 243 244 // Next, figure out which successor we are threading to. 245 BasicBlock *SuccBB; 246 if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) 247 SuccBB = BI->getSuccessor(PredCst == ConstantInt::getFalse()); 248 else { 249 SwitchInst *SI = cast<SwitchInst>(BB->getTerminator()); 250 SuccBB = SI->getSuccessor(SI->findCaseValue(PredCst)); 251 } 252 253 // And finally, do it! 254 DOUT << " Threading edge from '" << PredBB->getNameStart() << "' to '" 255 << SuccBB->getNameStart() << "' with cost: " << JumpThreadCost 256 << ", across block:\n " 257 << *BB << "\n"; 258 259 ThreadEdge(BB, PredBB, SuccBB); 260 ++NumThreads; 261 return true; 262} 263 264/// ProcessJumpOnLogicalPHI - PN's basic block contains a conditional branch 265/// whose condition is an AND/OR where one side is PN. If PN has constant 266/// operands that permit us to evaluate the condition for some operand, thread 267/// through the block. For example with: 268/// br (and X, phi(Y, Z, false)) 269/// the predecessor corresponding to the 'false' will always jump to the false 270/// destination of the branch. 271/// 272bool JumpThreading::ProcessBranchOnLogical(Value *V, BasicBlock *BB, 273 bool isAnd) { 274 // If this is a binary operator tree of the same AND/OR opcode, check the 275 // LHS/RHS. 276 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(V)) 277 if (isAnd && BO->getOpcode() == Instruction::And || 278 !isAnd && BO->getOpcode() == Instruction::Or) { 279 if (ProcessBranchOnLogical(BO->getOperand(0), BB, isAnd)) 280 return true; 281 if (ProcessBranchOnLogical(BO->getOperand(1), BB, isAnd)) 282 return true; 283 } 284 285 // If this isn't a PHI node, we can't handle it. 286 PHINode *PN = dyn_cast<PHINode>(V); 287 if (!PN || PN->getParent() != BB) return false; 288 289 // We can only do the simplification for phi nodes of 'false' with AND or 290 // 'true' with OR. See if we have any entries in the phi for this. 291 unsigned PredNo = ~0U; 292 ConstantInt *PredCst = ConstantInt::get(Type::Int1Ty, !isAnd); 293 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 294 if (PN->getIncomingValue(i) == PredCst) { 295 PredNo = i; 296 break; 297 } 298 } 299 300 // If no match, bail out. 301 if (PredNo == ~0U) 302 return false; 303 304 // See if the cost of duplicating this block is low enough. 305 unsigned JumpThreadCost = getJumpThreadDuplicationCost(BB); 306 if (JumpThreadCost > Threshold) { 307 DOUT << " Not threading BB '" << BB->getNameStart() 308 << "' - Cost is too high: " << JumpThreadCost << "\n"; 309 return false; 310 } 311 312 // If so, we can actually do this threading. Merge any common predecessors 313 // that will act the same. 314 BasicBlock *PredBB = FactorCommonPHIPreds(PN, PredCst); 315 316 // Next, figure out which successor we are threading to. If this was an AND, 317 // the constant must be FALSE, and we must be targeting the 'false' block. 318 // If this is an OR, the constant must be TRUE, and we must be targeting the 319 // 'true' block. 320 BasicBlock *SuccBB = BB->getTerminator()->getSuccessor(isAnd); 321 322 // And finally, do it! 323 DOUT << " Threading edge through bool from '" << PredBB->getNameStart() 324 << "' to '" << SuccBB->getNameStart() << "' with cost: " 325 << JumpThreadCost << ", across block:\n " 326 << *BB << "\n"; 327 328 ThreadEdge(BB, PredBB, SuccBB); 329 ++NumThreads; 330 return true; 331} 332 333/// ProcessBranchOnCompare - We found a branch on a comparison between a phi 334/// node and a constant. If the PHI node contains any constants as inputs, we 335/// can fold the compare for that edge and thread through it. 336bool JumpThreading::ProcessBranchOnCompare(CmpInst *Cmp, BasicBlock *BB) { 337 PHINode *PN = cast<PHINode>(Cmp->getOperand(0)); 338 Constant *RHS = cast<Constant>(Cmp->getOperand(1)); 339 340 // If the phi isn't in the current block, an incoming edge to this block 341 // doesn't control the destination. 342 if (PN->getParent() != BB) 343 return false; 344 345 // We can do this simplification if any comparisons fold to true or false. 346 // See if any do. 347 Constant *PredCst = 0; 348 bool TrueDirection = false; 349 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 350 PredCst = dyn_cast<Constant>(PN->getIncomingValue(i)); 351 if (PredCst == 0) continue; 352 353 Constant *Res; 354 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Cmp)) 355 Res = ConstantExpr::getICmp(ICI->getPredicate(), PredCst, RHS); 356 else 357 Res = ConstantExpr::getFCmp(cast<FCmpInst>(Cmp)->getPredicate(), 358 PredCst, RHS); 359 // If this folded to a constant expr, we can't do anything. 360 if (ConstantInt *ResC = dyn_cast<ConstantInt>(Res)) { 361 TrueDirection = ResC->getZExtValue(); 362 break; 363 } 364 // If this folded to undef, just go the false way. 365 if (isa<UndefValue>(Res)) { 366 TrueDirection = false; 367 break; 368 } 369 370 // Otherwise, we can't fold this input. 371 PredCst = 0; 372 } 373 374 // If no match, bail out. 375 if (PredCst == 0) 376 return false; 377 378 // See if the cost of duplicating this block is low enough. 379 unsigned JumpThreadCost = getJumpThreadDuplicationCost(BB); 380 if (JumpThreadCost > Threshold) { 381 DOUT << " Not threading BB '" << BB->getNameStart() 382 << "' - Cost is too high: " << JumpThreadCost << "\n"; 383 return false; 384 } 385 386 // If so, we can actually do this threading. Merge any common predecessors 387 // that will act the same. 388 BasicBlock *PredBB = FactorCommonPHIPreds(PN, PredCst); 389 390 // Next, get our successor. 391 BasicBlock *SuccBB = BB->getTerminator()->getSuccessor(!TrueDirection); 392 393 // And finally, do it! 394 DOUT << " Threading edge through bool from '" << PredBB->getNameStart() 395 << "' to '" << SuccBB->getNameStart() << "' with cost: " 396 << JumpThreadCost << ", across block:\n " 397 << *BB << "\n"; 398 399 ThreadEdge(BB, PredBB, SuccBB); 400 ++NumThreads; 401 return true; 402} 403 404 405/// ThreadEdge - We have decided that it is safe and profitable to thread an 406/// edge from PredBB to SuccBB across BB. Transform the IR to reflect this 407/// change. 408void JumpThreading::ThreadEdge(BasicBlock *BB, BasicBlock *PredBB, 409 BasicBlock *SuccBB) { 410 411 // Jump Threading can not update SSA properties correctly if the values 412 // defined in the duplicated block are used outside of the block itself. For 413 // this reason, we spill all values that are used outside of BB to the stack. 414 for (BasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) 415 if (I->isUsedOutsideOfBlock(BB)) { 416 // We found a use of I outside of BB. Create a new stack slot to 417 // break this inter-block usage pattern. 418 DemoteRegToStack(*I); 419 } 420 421 // We are going to have to map operands from the original BB block to the new 422 // copy of the block 'NewBB'. If there are PHI nodes in BB, evaluate them to 423 // account for entry from PredBB. 424 DenseMap<Instruction*, Value*> ValueMapping; 425 426 BasicBlock *NewBB = 427 BasicBlock::Create(BB->getName()+".thread", BB->getParent(), BB); 428 NewBB->moveAfter(PredBB); 429 430 BasicBlock::iterator BI = BB->begin(); 431 for (; PHINode *PN = dyn_cast<PHINode>(BI); ++BI) 432 ValueMapping[PN] = PN->getIncomingValueForBlock(PredBB); 433 434 // Clone the non-phi instructions of BB into NewBB, keeping track of the 435 // mapping and using it to remap operands in the cloned instructions. 436 for (; !isa<TerminatorInst>(BI); ++BI) { 437 Instruction *New = BI->clone(); 438 New->setName(BI->getNameStart()); 439 NewBB->getInstList().push_back(New); 440 ValueMapping[BI] = New; 441 442 // Remap operands to patch up intra-block references. 443 for (unsigned i = 0, e = New->getNumOperands(); i != e; ++i) 444 if (Instruction *Inst = dyn_cast<Instruction>(New->getOperand(i))) 445 if (Value *Remapped = ValueMapping[Inst]) 446 New->setOperand(i, Remapped); 447 } 448 449 // We didn't copy the terminator from BB over to NewBB, because there is now 450 // an unconditional jump to SuccBB. Insert the unconditional jump. 451 BranchInst::Create(SuccBB, NewBB); 452 453 // Check to see if SuccBB has PHI nodes. If so, we need to add entries to the 454 // PHI nodes for NewBB now. 455 for (BasicBlock::iterator PNI = SuccBB->begin(); isa<PHINode>(PNI); ++PNI) { 456 PHINode *PN = cast<PHINode>(PNI); 457 // Ok, we have a PHI node. Figure out what the incoming value was for the 458 // DestBlock. 459 Value *IV = PN->getIncomingValueForBlock(BB); 460 461 // Remap the value if necessary. 462 if (Instruction *Inst = dyn_cast<Instruction>(IV)) 463 if (Value *MappedIV = ValueMapping[Inst]) 464 IV = MappedIV; 465 PN->addIncoming(IV, NewBB); 466 } 467 468 // Finally, NewBB is good to go. Update the terminator of PredBB to jump to 469 // NewBB instead of BB. This eliminates predecessors from BB, which requires 470 // us to simplify any PHI nodes in BB. 471 TerminatorInst *PredTerm = PredBB->getTerminator(); 472 for (unsigned i = 0, e = PredTerm->getNumSuccessors(); i != e; ++i) 473 if (PredTerm->getSuccessor(i) == BB) { 474 BB->removePredecessor(PredBB); 475 PredTerm->setSuccessor(i, NewBB); 476 } 477} 478