SimplifyCFG.cpp revision 3d0a9a371c0ce3a46e845a7bf1f1acb7a1cf523e
1//===- SimplifyCFG.cpp - Code to perform CFG simplification ---------------===// 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// Peephole optimize the CFG. 11// 12//===----------------------------------------------------------------------===// 13 14#define DEBUG_TYPE "simplifycfg" 15#include "llvm/Transforms/Utils/Local.h" 16#include "llvm/Constants.h" 17#include "llvm/Instructions.h" 18#include "llvm/Type.h" 19#include "llvm/DerivedTypes.h" 20#include "llvm/Support/CFG.h" 21#include "llvm/Support/Debug.h" 22#include "llvm/Analysis/ConstantFolding.h" 23#include "llvm/Transforms/Utils/BasicBlockUtils.h" 24#include "llvm/ADT/SmallVector.h" 25#include "llvm/ADT/SmallPtrSet.h" 26#include "llvm/ADT/Statistic.h" 27#include <algorithm> 28#include <functional> 29#include <set> 30#include <map> 31using namespace llvm; 32 33STATISTIC(NumSpeculations, "Number of speculative executed instructions"); 34 35/// SafeToMergeTerminators - Return true if it is safe to merge these two 36/// terminator instructions together. 37/// 38static bool SafeToMergeTerminators(TerminatorInst *SI1, TerminatorInst *SI2) { 39 if (SI1 == SI2) return false; // Can't merge with self! 40 41 // It is not safe to merge these two switch instructions if they have a common 42 // successor, and if that successor has a PHI node, and if *that* PHI node has 43 // conflicting incoming values from the two switch blocks. 44 BasicBlock *SI1BB = SI1->getParent(); 45 BasicBlock *SI2BB = SI2->getParent(); 46 SmallPtrSet<BasicBlock*, 16> SI1Succs(succ_begin(SI1BB), succ_end(SI1BB)); 47 48 for (succ_iterator I = succ_begin(SI2BB), E = succ_end(SI2BB); I != E; ++I) 49 if (SI1Succs.count(*I)) 50 for (BasicBlock::iterator BBI = (*I)->begin(); 51 isa<PHINode>(BBI); ++BBI) { 52 PHINode *PN = cast<PHINode>(BBI); 53 if (PN->getIncomingValueForBlock(SI1BB) != 54 PN->getIncomingValueForBlock(SI2BB)) 55 return false; 56 } 57 58 return true; 59} 60 61/// AddPredecessorToBlock - Update PHI nodes in Succ to indicate that there will 62/// now be entries in it from the 'NewPred' block. The values that will be 63/// flowing into the PHI nodes will be the same as those coming in from 64/// ExistPred, an existing predecessor of Succ. 65static void AddPredecessorToBlock(BasicBlock *Succ, BasicBlock *NewPred, 66 BasicBlock *ExistPred) { 67 assert(std::find(succ_begin(ExistPred), succ_end(ExistPred), Succ) != 68 succ_end(ExistPred) && "ExistPred is not a predecessor of Succ!"); 69 if (!isa<PHINode>(Succ->begin())) return; // Quick exit if nothing to do 70 71 PHINode *PN; 72 for (BasicBlock::iterator I = Succ->begin(); 73 (PN = dyn_cast<PHINode>(I)); ++I) 74 PN->addIncoming(PN->getIncomingValueForBlock(ExistPred), NewPred); 75} 76 77// CanPropagatePredecessorsForPHIs - Return true if we can fold BB, an 78// almost-empty BB ending in an unconditional branch to Succ, into succ. 79// 80// Assumption: Succ is the single successor for BB. 81// 82static bool CanPropagatePredecessorsForPHIs(BasicBlock *BB, BasicBlock *Succ) { 83 assert(*succ_begin(BB) == Succ && "Succ is not successor of BB!"); 84 85 DOUT << "Looking to fold " << BB->getNameStart() << " into " 86 << Succ->getNameStart() << "\n"; 87 // Shortcut, if there is only a single predecessor is must be BB and merging 88 // is always safe 89 if (Succ->getSinglePredecessor()) return true; 90 91 typedef SmallPtrSet<Instruction*, 16> InstrSet; 92 InstrSet BBPHIs; 93 94 // Make a list of all phi nodes in BB 95 BasicBlock::iterator BBI = BB->begin(); 96 while (isa<PHINode>(*BBI)) BBPHIs.insert(BBI++); 97 98 // Make a list of the predecessors of BB 99 typedef SmallPtrSet<BasicBlock*, 16> BlockSet; 100 BlockSet BBPreds(pred_begin(BB), pred_end(BB)); 101 102 // Use that list to make another list of common predecessors of BB and Succ 103 BlockSet CommonPreds; 104 for (pred_iterator PI = pred_begin(Succ), PE = pred_end(Succ); 105 PI != PE; ++PI) 106 if (BBPreds.count(*PI)) 107 CommonPreds.insert(*PI); 108 109 // Shortcut, if there are no common predecessors, merging is always safe 110 if (CommonPreds.empty()) 111 return true; 112 113 // Look at all the phi nodes in Succ, to see if they present a conflict when 114 // merging these blocks 115 for (BasicBlock::iterator I = Succ->begin(); isa<PHINode>(I); ++I) { 116 PHINode *PN = cast<PHINode>(I); 117 118 // If the incoming value from BB is again a PHINode in 119 // BB which has the same incoming value for *PI as PN does, we can 120 // merge the phi nodes and then the blocks can still be merged 121 PHINode *BBPN = dyn_cast<PHINode>(PN->getIncomingValueForBlock(BB)); 122 if (BBPN && BBPN->getParent() == BB) { 123 for (BlockSet::iterator PI = CommonPreds.begin(), PE = CommonPreds.end(); 124 PI != PE; PI++) { 125 if (BBPN->getIncomingValueForBlock(*PI) 126 != PN->getIncomingValueForBlock(*PI)) { 127 DOUT << "Can't fold, phi node " << *PN->getNameStart() << " in " 128 << Succ->getNameStart() << " is conflicting with " 129 << BBPN->getNameStart() << " with regard to common predecessor " 130 << (*PI)->getNameStart() << "\n"; 131 return false; 132 } 133 } 134 // Remove this phinode from the list of phis in BB, since it has been 135 // handled. 136 BBPHIs.erase(BBPN); 137 } else { 138 Value* Val = PN->getIncomingValueForBlock(BB); 139 for (BlockSet::iterator PI = CommonPreds.begin(), PE = CommonPreds.end(); 140 PI != PE; PI++) { 141 // See if the incoming value for the common predecessor is equal to the 142 // one for BB, in which case this phi node will not prevent the merging 143 // of the block. 144 if (Val != PN->getIncomingValueForBlock(*PI)) { 145 DOUT << "Can't fold, phi node " << *PN->getNameStart() << " in " 146 << Succ->getNameStart() << " is conflicting with regard to common " 147 << "predecessor " << (*PI)->getNameStart() << "\n"; 148 return false; 149 } 150 } 151 } 152 } 153 154 // If there are any other phi nodes in BB that don't have a phi node in Succ 155 // to merge with, they must be moved to Succ completely. However, for any 156 // predecessors of Succ, branches will be added to the phi node that just 157 // point to itself. So, for any common predecessors, this must not cause 158 // conflicts. 159 for (InstrSet::iterator I = BBPHIs.begin(), E = BBPHIs.end(); 160 I != E; I++) { 161 PHINode *PN = cast<PHINode>(*I); 162 for (BlockSet::iterator PI = CommonPreds.begin(), PE = CommonPreds.end(); 163 PI != PE; PI++) 164 if (PN->getIncomingValueForBlock(*PI) != PN) { 165 DOUT << "Can't fold, phi node " << *PN->getNameStart() << " in " 166 << BB->getNameStart() << " is conflicting with regard to common " 167 << "predecessor " << (*PI)->getNameStart() << "\n"; 168 return false; 169 } 170 } 171 172 return true; 173} 174 175/// TryToSimplifyUncondBranchFromEmptyBlock - BB contains an unconditional 176/// branch to Succ, and contains no instructions other than PHI nodes and the 177/// branch. If possible, eliminate BB. 178static bool TryToSimplifyUncondBranchFromEmptyBlock(BasicBlock *BB, 179 BasicBlock *Succ) { 180 // Check to see if merging these blocks would cause conflicts for any of the 181 // phi nodes in BB or Succ. If not, we can safely merge. 182 if (!CanPropagatePredecessorsForPHIs(BB, Succ)) return false; 183 184 DOUT << "Killing Trivial BB: \n" << *BB; 185 186 if (isa<PHINode>(Succ->begin())) { 187 // If there is more than one pred of succ, and there are PHI nodes in 188 // the successor, then we need to add incoming edges for the PHI nodes 189 // 190 const SmallVector<BasicBlock*, 16> BBPreds(pred_begin(BB), pred_end(BB)); 191 192 // Loop over all of the PHI nodes in the successor of BB. 193 for (BasicBlock::iterator I = Succ->begin(); isa<PHINode>(I); ++I) { 194 PHINode *PN = cast<PHINode>(I); 195 Value *OldVal = PN->removeIncomingValue(BB, false); 196 assert(OldVal && "No entry in PHI for Pred BB!"); 197 198 // If this incoming value is one of the PHI nodes in BB, the new entries 199 // in the PHI node are the entries from the old PHI. 200 if (isa<PHINode>(OldVal) && cast<PHINode>(OldVal)->getParent() == BB) { 201 PHINode *OldValPN = cast<PHINode>(OldVal); 202 for (unsigned i = 0, e = OldValPN->getNumIncomingValues(); i != e; ++i) 203 // Note that, since we are merging phi nodes and BB and Succ might 204 // have common predecessors, we could end up with a phi node with 205 // identical incoming branches. This will be cleaned up later (and 206 // will trigger asserts if we try to clean it up now, without also 207 // simplifying the corresponding conditional branch). 208 PN->addIncoming(OldValPN->getIncomingValue(i), 209 OldValPN->getIncomingBlock(i)); 210 } else { 211 // Add an incoming value for each of the new incoming values. 212 for (unsigned i = 0, e = BBPreds.size(); i != e; ++i) 213 PN->addIncoming(OldVal, BBPreds[i]); 214 } 215 } 216 } 217 218 if (isa<PHINode>(&BB->front())) { 219 SmallVector<BasicBlock*, 16> 220 OldSuccPreds(pred_begin(Succ), pred_end(Succ)); 221 222 // Move all PHI nodes in BB to Succ if they are alive, otherwise 223 // delete them. 224 while (PHINode *PN = dyn_cast<PHINode>(&BB->front())) 225 if (PN->use_empty()) { 226 // Just remove the dead phi. This happens if Succ's PHIs were the only 227 // users of the PHI nodes. 228 PN->eraseFromParent(); 229 } else { 230 // The instruction is alive, so this means that BB must dominate all 231 // predecessors of Succ (Since all uses of the PN are after its 232 // definition, so in Succ or a block dominated by Succ. If a predecessor 233 // of Succ would not be dominated by BB, PN would violate the def before 234 // use SSA demand). Therefore, we can simply move the phi node to the 235 // next block. 236 Succ->getInstList().splice(Succ->begin(), 237 BB->getInstList(), BB->begin()); 238 239 // We need to add new entries for the PHI node to account for 240 // predecessors of Succ that the PHI node does not take into 241 // account. At this point, since we know that BB dominated succ and all 242 // of its predecessors, this means that we should any newly added 243 // incoming edges should use the PHI node itself as the value for these 244 // edges, because they are loop back edges. 245 for (unsigned i = 0, e = OldSuccPreds.size(); i != e; ++i) 246 if (OldSuccPreds[i] != BB) 247 PN->addIncoming(PN, OldSuccPreds[i]); 248 } 249 } 250 251 // Everything that jumped to BB now goes to Succ. 252 BB->replaceAllUsesWith(Succ); 253 if (!Succ->hasName()) Succ->takeName(BB); 254 BB->eraseFromParent(); // Delete the old basic block. 255 return true; 256} 257 258/// GetIfCondition - Given a basic block (BB) with two predecessors (and 259/// presumably PHI nodes in it), check to see if the merge at this block is due 260/// to an "if condition". If so, return the boolean condition that determines 261/// which entry into BB will be taken. Also, return by references the block 262/// that will be entered from if the condition is true, and the block that will 263/// be entered if the condition is false. 264/// 265/// 266static Value *GetIfCondition(BasicBlock *BB, 267 BasicBlock *&IfTrue, BasicBlock *&IfFalse) { 268 assert(std::distance(pred_begin(BB), pred_end(BB)) == 2 && 269 "Function can only handle blocks with 2 predecessors!"); 270 BasicBlock *Pred1 = *pred_begin(BB); 271 BasicBlock *Pred2 = *++pred_begin(BB); 272 273 // We can only handle branches. Other control flow will be lowered to 274 // branches if possible anyway. 275 if (!isa<BranchInst>(Pred1->getTerminator()) || 276 !isa<BranchInst>(Pred2->getTerminator())) 277 return 0; 278 BranchInst *Pred1Br = cast<BranchInst>(Pred1->getTerminator()); 279 BranchInst *Pred2Br = cast<BranchInst>(Pred2->getTerminator()); 280 281 // Eliminate code duplication by ensuring that Pred1Br is conditional if 282 // either are. 283 if (Pred2Br->isConditional()) { 284 // If both branches are conditional, we don't have an "if statement". In 285 // reality, we could transform this case, but since the condition will be 286 // required anyway, we stand no chance of eliminating it, so the xform is 287 // probably not profitable. 288 if (Pred1Br->isConditional()) 289 return 0; 290 291 std::swap(Pred1, Pred2); 292 std::swap(Pred1Br, Pred2Br); 293 } 294 295 if (Pred1Br->isConditional()) { 296 // If we found a conditional branch predecessor, make sure that it branches 297 // to BB and Pred2Br. If it doesn't, this isn't an "if statement". 298 if (Pred1Br->getSuccessor(0) == BB && 299 Pred1Br->getSuccessor(1) == Pred2) { 300 IfTrue = Pred1; 301 IfFalse = Pred2; 302 } else if (Pred1Br->getSuccessor(0) == Pred2 && 303 Pred1Br->getSuccessor(1) == BB) { 304 IfTrue = Pred2; 305 IfFalse = Pred1; 306 } else { 307 // We know that one arm of the conditional goes to BB, so the other must 308 // go somewhere unrelated, and this must not be an "if statement". 309 return 0; 310 } 311 312 // The only thing we have to watch out for here is to make sure that Pred2 313 // doesn't have incoming edges from other blocks. If it does, the condition 314 // doesn't dominate BB. 315 if (++pred_begin(Pred2) != pred_end(Pred2)) 316 return 0; 317 318 return Pred1Br->getCondition(); 319 } 320 321 // Ok, if we got here, both predecessors end with an unconditional branch to 322 // BB. Don't panic! If both blocks only have a single (identical) 323 // predecessor, and THAT is a conditional branch, then we're all ok! 324 if (pred_begin(Pred1) == pred_end(Pred1) || 325 ++pred_begin(Pred1) != pred_end(Pred1) || 326 pred_begin(Pred2) == pred_end(Pred2) || 327 ++pred_begin(Pred2) != pred_end(Pred2) || 328 *pred_begin(Pred1) != *pred_begin(Pred2)) 329 return 0; 330 331 // Otherwise, if this is a conditional branch, then we can use it! 332 BasicBlock *CommonPred = *pred_begin(Pred1); 333 if (BranchInst *BI = dyn_cast<BranchInst>(CommonPred->getTerminator())) { 334 assert(BI->isConditional() && "Two successors but not conditional?"); 335 if (BI->getSuccessor(0) == Pred1) { 336 IfTrue = Pred1; 337 IfFalse = Pred2; 338 } else { 339 IfTrue = Pred2; 340 IfFalse = Pred1; 341 } 342 return BI->getCondition(); 343 } 344 return 0; 345} 346 347 348// If we have a merge point of an "if condition" as accepted above, return true 349// if the specified value dominates the block. We don't handle the true 350// generality of domination here, just a special case which works well enough 351// for us. 352// 353// If AggressiveInsts is non-null, and if V does not dominate BB, we check to 354// see if V (which must be an instruction) is cheap to compute and is 355// non-trapping. If both are true, the instruction is inserted into the set and 356// true is returned. 357static bool DominatesMergePoint(Value *V, BasicBlock *BB, 358 std::set<Instruction*> *AggressiveInsts) { 359 Instruction *I = dyn_cast<Instruction>(V); 360 if (!I) { 361 // Non-instructions all dominate instructions, but not all constantexprs 362 // can be executed unconditionally. 363 if (ConstantExpr *C = dyn_cast<ConstantExpr>(V)) 364 if (C->canTrap()) 365 return false; 366 return true; 367 } 368 BasicBlock *PBB = I->getParent(); 369 370 // We don't want to allow weird loops that might have the "if condition" in 371 // the bottom of this block. 372 if (PBB == BB) return false; 373 374 // If this instruction is defined in a block that contains an unconditional 375 // branch to BB, then it must be in the 'conditional' part of the "if 376 // statement". 377 if (BranchInst *BI = dyn_cast<BranchInst>(PBB->getTerminator())) 378 if (BI->isUnconditional() && BI->getSuccessor(0) == BB) { 379 if (!AggressiveInsts) return false; 380 // Okay, it looks like the instruction IS in the "condition". Check to 381 // see if its a cheap instruction to unconditionally compute, and if it 382 // only uses stuff defined outside of the condition. If so, hoist it out. 383 switch (I->getOpcode()) { 384 default: return false; // Cannot hoist this out safely. 385 case Instruction::Load: 386 // We can hoist loads that are non-volatile and obviously cannot trap. 387 if (cast<LoadInst>(I)->isVolatile()) 388 return false; 389 if (!isa<AllocaInst>(I->getOperand(0)) && 390 !isa<Constant>(I->getOperand(0))) 391 return false; 392 393 // Finally, we have to check to make sure there are no instructions 394 // before the load in its basic block, as we are going to hoist the loop 395 // out to its predecessor. 396 if (PBB->begin() != BasicBlock::iterator(I)) 397 return false; 398 break; 399 case Instruction::Add: 400 case Instruction::Sub: 401 case Instruction::And: 402 case Instruction::Or: 403 case Instruction::Xor: 404 case Instruction::Shl: 405 case Instruction::LShr: 406 case Instruction::AShr: 407 case Instruction::ICmp: 408 case Instruction::FCmp: 409 if (I->getOperand(0)->getType()->isFPOrFPVector()) 410 return false; // FP arithmetic might trap. 411 break; // These are all cheap and non-trapping instructions. 412 } 413 414 // Okay, we can only really hoist these out if their operands are not 415 // defined in the conditional region. 416 for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i) 417 if (!DominatesMergePoint(*i, BB, 0)) 418 return false; 419 // Okay, it's safe to do this! Remember this instruction. 420 AggressiveInsts->insert(I); 421 } 422 423 return true; 424} 425 426// GatherConstantSetEQs - Given a potentially 'or'd together collection of 427// icmp_eq instructions that compare a value against a constant, return the 428// value being compared, and stick the constant into the Values vector. 429static Value *GatherConstantSetEQs(Value *V, std::vector<ConstantInt*> &Values){ 430 if (Instruction *Inst = dyn_cast<Instruction>(V)) { 431 if (Inst->getOpcode() == Instruction::ICmp && 432 cast<ICmpInst>(Inst)->getPredicate() == ICmpInst::ICMP_EQ) { 433 if (ConstantInt *C = dyn_cast<ConstantInt>(Inst->getOperand(1))) { 434 Values.push_back(C); 435 return Inst->getOperand(0); 436 } else if (ConstantInt *C = dyn_cast<ConstantInt>(Inst->getOperand(0))) { 437 Values.push_back(C); 438 return Inst->getOperand(1); 439 } 440 } else if (Inst->getOpcode() == Instruction::Or) { 441 if (Value *LHS = GatherConstantSetEQs(Inst->getOperand(0), Values)) 442 if (Value *RHS = GatherConstantSetEQs(Inst->getOperand(1), Values)) 443 if (LHS == RHS) 444 return LHS; 445 } 446 } 447 return 0; 448} 449 450// GatherConstantSetNEs - Given a potentially 'and'd together collection of 451// setne instructions that compare a value against a constant, return the value 452// being compared, and stick the constant into the Values vector. 453static Value *GatherConstantSetNEs(Value *V, std::vector<ConstantInt*> &Values){ 454 if (Instruction *Inst = dyn_cast<Instruction>(V)) { 455 if (Inst->getOpcode() == Instruction::ICmp && 456 cast<ICmpInst>(Inst)->getPredicate() == ICmpInst::ICMP_NE) { 457 if (ConstantInt *C = dyn_cast<ConstantInt>(Inst->getOperand(1))) { 458 Values.push_back(C); 459 return Inst->getOperand(0); 460 } else if (ConstantInt *C = dyn_cast<ConstantInt>(Inst->getOperand(0))) { 461 Values.push_back(C); 462 return Inst->getOperand(1); 463 } 464 } else if (Inst->getOpcode() == Instruction::And) { 465 if (Value *LHS = GatherConstantSetNEs(Inst->getOperand(0), Values)) 466 if (Value *RHS = GatherConstantSetNEs(Inst->getOperand(1), Values)) 467 if (LHS == RHS) 468 return LHS; 469 } 470 } 471 return 0; 472} 473 474 475 476/// GatherValueComparisons - If the specified Cond is an 'and' or 'or' of a 477/// bunch of comparisons of one value against constants, return the value and 478/// the constants being compared. 479static bool GatherValueComparisons(Instruction *Cond, Value *&CompVal, 480 std::vector<ConstantInt*> &Values) { 481 if (Cond->getOpcode() == Instruction::Or) { 482 CompVal = GatherConstantSetEQs(Cond, Values); 483 484 // Return true to indicate that the condition is true if the CompVal is 485 // equal to one of the constants. 486 return true; 487 } else if (Cond->getOpcode() == Instruction::And) { 488 CompVal = GatherConstantSetNEs(Cond, Values); 489 490 // Return false to indicate that the condition is false if the CompVal is 491 // equal to one of the constants. 492 return false; 493 } 494 return false; 495} 496 497/// ErasePossiblyDeadInstructionTree - If the specified instruction is dead and 498/// has no side effects, nuke it. If it uses any instructions that become dead 499/// because the instruction is now gone, nuke them too. 500static void ErasePossiblyDeadInstructionTree(Instruction *I) { 501 if (!isInstructionTriviallyDead(I)) return; 502 503 SmallVector<Instruction*, 16> InstrsToInspect; 504 InstrsToInspect.push_back(I); 505 506 while (!InstrsToInspect.empty()) { 507 I = InstrsToInspect.back(); 508 InstrsToInspect.pop_back(); 509 510 if (!isInstructionTriviallyDead(I)) continue; 511 512 // If I is in the work list multiple times, remove previous instances. 513 for (unsigned i = 0, e = InstrsToInspect.size(); i != e; ++i) 514 if (InstrsToInspect[i] == I) { 515 InstrsToInspect.erase(InstrsToInspect.begin()+i); 516 --i, --e; 517 } 518 519 // Add operands of dead instruction to worklist. 520 for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i) 521 if (Instruction *OpI = dyn_cast<Instruction>(*i)) 522 InstrsToInspect.push_back(OpI); 523 524 // Remove dead instruction. 525 I->eraseFromParent(); 526 } 527} 528 529// isValueEqualityComparison - Return true if the specified terminator checks to 530// see if a value is equal to constant integer value. 531static Value *isValueEqualityComparison(TerminatorInst *TI) { 532 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) { 533 // Do not permit merging of large switch instructions into their 534 // predecessors unless there is only one predecessor. 535 if (SI->getNumSuccessors() * std::distance(pred_begin(SI->getParent()), 536 pred_end(SI->getParent())) > 128) 537 return 0; 538 539 return SI->getCondition(); 540 } 541 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) 542 if (BI->isConditional() && BI->getCondition()->hasOneUse()) 543 if (ICmpInst *ICI = dyn_cast<ICmpInst>(BI->getCondition())) 544 if ((ICI->getPredicate() == ICmpInst::ICMP_EQ || 545 ICI->getPredicate() == ICmpInst::ICMP_NE) && 546 isa<ConstantInt>(ICI->getOperand(1))) 547 return ICI->getOperand(0); 548 return 0; 549} 550 551// Given a value comparison instruction, decode all of the 'cases' that it 552// represents and return the 'default' block. 553static BasicBlock * 554GetValueEqualityComparisonCases(TerminatorInst *TI, 555 std::vector<std::pair<ConstantInt*, 556 BasicBlock*> > &Cases) { 557 if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) { 558 Cases.reserve(SI->getNumCases()); 559 for (unsigned i = 1, e = SI->getNumCases(); i != e; ++i) 560 Cases.push_back(std::make_pair(SI->getCaseValue(i), SI->getSuccessor(i))); 561 return SI->getDefaultDest(); 562 } 563 564 BranchInst *BI = cast<BranchInst>(TI); 565 ICmpInst *ICI = cast<ICmpInst>(BI->getCondition()); 566 Cases.push_back(std::make_pair(cast<ConstantInt>(ICI->getOperand(1)), 567 BI->getSuccessor(ICI->getPredicate() == 568 ICmpInst::ICMP_NE))); 569 return BI->getSuccessor(ICI->getPredicate() == ICmpInst::ICMP_EQ); 570} 571 572 573// EliminateBlockCases - Given a vector of bb/value pairs, remove any entries 574// in the list that match the specified block. 575static void EliminateBlockCases(BasicBlock *BB, 576 std::vector<std::pair<ConstantInt*, BasicBlock*> > &Cases) { 577 for (unsigned i = 0, e = Cases.size(); i != e; ++i) 578 if (Cases[i].second == BB) { 579 Cases.erase(Cases.begin()+i); 580 --i; --e; 581 } 582} 583 584// ValuesOverlap - Return true if there are any keys in C1 that exist in C2 as 585// well. 586static bool 587ValuesOverlap(std::vector<std::pair<ConstantInt*, BasicBlock*> > &C1, 588 std::vector<std::pair<ConstantInt*, BasicBlock*> > &C2) { 589 std::vector<std::pair<ConstantInt*, BasicBlock*> > *V1 = &C1, *V2 = &C2; 590 591 // Make V1 be smaller than V2. 592 if (V1->size() > V2->size()) 593 std::swap(V1, V2); 594 595 if (V1->size() == 0) return false; 596 if (V1->size() == 1) { 597 // Just scan V2. 598 ConstantInt *TheVal = (*V1)[0].first; 599 for (unsigned i = 0, e = V2->size(); i != e; ++i) 600 if (TheVal == (*V2)[i].first) 601 return true; 602 } 603 604 // Otherwise, just sort both lists and compare element by element. 605 std::sort(V1->begin(), V1->end()); 606 std::sort(V2->begin(), V2->end()); 607 unsigned i1 = 0, i2 = 0, e1 = V1->size(), e2 = V2->size(); 608 while (i1 != e1 && i2 != e2) { 609 if ((*V1)[i1].first == (*V2)[i2].first) 610 return true; 611 if ((*V1)[i1].first < (*V2)[i2].first) 612 ++i1; 613 else 614 ++i2; 615 } 616 return false; 617} 618 619// SimplifyEqualityComparisonWithOnlyPredecessor - If TI is known to be a 620// terminator instruction and its block is known to only have a single 621// predecessor block, check to see if that predecessor is also a value 622// comparison with the same value, and if that comparison determines the outcome 623// of this comparison. If so, simplify TI. This does a very limited form of 624// jump threading. 625static bool SimplifyEqualityComparisonWithOnlyPredecessor(TerminatorInst *TI, 626 BasicBlock *Pred) { 627 Value *PredVal = isValueEqualityComparison(Pred->getTerminator()); 628 if (!PredVal) return false; // Not a value comparison in predecessor. 629 630 Value *ThisVal = isValueEqualityComparison(TI); 631 assert(ThisVal && "This isn't a value comparison!!"); 632 if (ThisVal != PredVal) return false; // Different predicates. 633 634 // Find out information about when control will move from Pred to TI's block. 635 std::vector<std::pair<ConstantInt*, BasicBlock*> > PredCases; 636 BasicBlock *PredDef = GetValueEqualityComparisonCases(Pred->getTerminator(), 637 PredCases); 638 EliminateBlockCases(PredDef, PredCases); // Remove default from cases. 639 640 // Find information about how control leaves this block. 641 std::vector<std::pair<ConstantInt*, BasicBlock*> > ThisCases; 642 BasicBlock *ThisDef = GetValueEqualityComparisonCases(TI, ThisCases); 643 EliminateBlockCases(ThisDef, ThisCases); // Remove default from cases. 644 645 // If TI's block is the default block from Pred's comparison, potentially 646 // simplify TI based on this knowledge. 647 if (PredDef == TI->getParent()) { 648 // If we are here, we know that the value is none of those cases listed in 649 // PredCases. If there are any cases in ThisCases that are in PredCases, we 650 // can simplify TI. 651 if (ValuesOverlap(PredCases, ThisCases)) { 652 if (BranchInst *BTI = dyn_cast<BranchInst>(TI)) { 653 // Okay, one of the successors of this condbr is dead. Convert it to a 654 // uncond br. 655 assert(ThisCases.size() == 1 && "Branch can only have one case!"); 656 Value *Cond = BTI->getCondition(); 657 // Insert the new branch. 658 Instruction *NI = BranchInst::Create(ThisDef, TI); 659 660 // Remove PHI node entries for the dead edge. 661 ThisCases[0].second->removePredecessor(TI->getParent()); 662 663 DOUT << "Threading pred instr: " << *Pred->getTerminator() 664 << "Through successor TI: " << *TI << "Leaving: " << *NI << "\n"; 665 666 TI->eraseFromParent(); // Nuke the old one. 667 // If condition is now dead, nuke it. 668 if (Instruction *CondI = dyn_cast<Instruction>(Cond)) 669 ErasePossiblyDeadInstructionTree(CondI); 670 return true; 671 672 } else { 673 SwitchInst *SI = cast<SwitchInst>(TI); 674 // Okay, TI has cases that are statically dead, prune them away. 675 SmallPtrSet<Constant*, 16> DeadCases; 676 for (unsigned i = 0, e = PredCases.size(); i != e; ++i) 677 DeadCases.insert(PredCases[i].first); 678 679 DOUT << "Threading pred instr: " << *Pred->getTerminator() 680 << "Through successor TI: " << *TI; 681 682 for (unsigned i = SI->getNumCases()-1; i != 0; --i) 683 if (DeadCases.count(SI->getCaseValue(i))) { 684 SI->getSuccessor(i)->removePredecessor(TI->getParent()); 685 SI->removeCase(i); 686 } 687 688 DOUT << "Leaving: " << *TI << "\n"; 689 return true; 690 } 691 } 692 693 } else { 694 // Otherwise, TI's block must correspond to some matched value. Find out 695 // which value (or set of values) this is. 696 ConstantInt *TIV = 0; 697 BasicBlock *TIBB = TI->getParent(); 698 for (unsigned i = 0, e = PredCases.size(); i != e; ++i) 699 if (PredCases[i].second == TIBB) { 700 if (TIV == 0) 701 TIV = PredCases[i].first; 702 else 703 return false; // Cannot handle multiple values coming to this block. 704 } 705 assert(TIV && "No edge from pred to succ?"); 706 707 // Okay, we found the one constant that our value can be if we get into TI's 708 // BB. Find out which successor will unconditionally be branched to. 709 BasicBlock *TheRealDest = 0; 710 for (unsigned i = 0, e = ThisCases.size(); i != e; ++i) 711 if (ThisCases[i].first == TIV) { 712 TheRealDest = ThisCases[i].second; 713 break; 714 } 715 716 // If not handled by any explicit cases, it is handled by the default case. 717 if (TheRealDest == 0) TheRealDest = ThisDef; 718 719 // Remove PHI node entries for dead edges. 720 BasicBlock *CheckEdge = TheRealDest; 721 for (succ_iterator SI = succ_begin(TIBB), e = succ_end(TIBB); SI != e; ++SI) 722 if (*SI != CheckEdge) 723 (*SI)->removePredecessor(TIBB); 724 else 725 CheckEdge = 0; 726 727 // Insert the new branch. 728 Instruction *NI = BranchInst::Create(TheRealDest, TI); 729 730 DOUT << "Threading pred instr: " << *Pred->getTerminator() 731 << "Through successor TI: " << *TI << "Leaving: " << *NI << "\n"; 732 Instruction *Cond = 0; 733 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) 734 Cond = dyn_cast<Instruction>(BI->getCondition()); 735 TI->eraseFromParent(); // Nuke the old one. 736 737 if (Cond) ErasePossiblyDeadInstructionTree(Cond); 738 return true; 739 } 740 return false; 741} 742 743// FoldValueComparisonIntoPredecessors - The specified terminator is a value 744// equality comparison instruction (either a switch or a branch on "X == c"). 745// See if any of the predecessors of the terminator block are value comparisons 746// on the same value. If so, and if safe to do so, fold them together. 747static bool FoldValueComparisonIntoPredecessors(TerminatorInst *TI) { 748 BasicBlock *BB = TI->getParent(); 749 Value *CV = isValueEqualityComparison(TI); // CondVal 750 assert(CV && "Not a comparison?"); 751 bool Changed = false; 752 753 SmallVector<BasicBlock*, 16> Preds(pred_begin(BB), pred_end(BB)); 754 while (!Preds.empty()) { 755 BasicBlock *Pred = Preds.back(); 756 Preds.pop_back(); 757 758 // See if the predecessor is a comparison with the same value. 759 TerminatorInst *PTI = Pred->getTerminator(); 760 Value *PCV = isValueEqualityComparison(PTI); // PredCondVal 761 762 if (PCV == CV && SafeToMergeTerminators(TI, PTI)) { 763 // Figure out which 'cases' to copy from SI to PSI. 764 std::vector<std::pair<ConstantInt*, BasicBlock*> > BBCases; 765 BasicBlock *BBDefault = GetValueEqualityComparisonCases(TI, BBCases); 766 767 std::vector<std::pair<ConstantInt*, BasicBlock*> > PredCases; 768 BasicBlock *PredDefault = GetValueEqualityComparisonCases(PTI, PredCases); 769 770 // Based on whether the default edge from PTI goes to BB or not, fill in 771 // PredCases and PredDefault with the new switch cases we would like to 772 // build. 773 SmallVector<BasicBlock*, 8> NewSuccessors; 774 775 if (PredDefault == BB) { 776 // If this is the default destination from PTI, only the edges in TI 777 // that don't occur in PTI, or that branch to BB will be activated. 778 std::set<ConstantInt*> PTIHandled; 779 for (unsigned i = 0, e = PredCases.size(); i != e; ++i) 780 if (PredCases[i].second != BB) 781 PTIHandled.insert(PredCases[i].first); 782 else { 783 // The default destination is BB, we don't need explicit targets. 784 std::swap(PredCases[i], PredCases.back()); 785 PredCases.pop_back(); 786 --i; --e; 787 } 788 789 // Reconstruct the new switch statement we will be building. 790 if (PredDefault != BBDefault) { 791 PredDefault->removePredecessor(Pred); 792 PredDefault = BBDefault; 793 NewSuccessors.push_back(BBDefault); 794 } 795 for (unsigned i = 0, e = BBCases.size(); i != e; ++i) 796 if (!PTIHandled.count(BBCases[i].first) && 797 BBCases[i].second != BBDefault) { 798 PredCases.push_back(BBCases[i]); 799 NewSuccessors.push_back(BBCases[i].second); 800 } 801 802 } else { 803 // If this is not the default destination from PSI, only the edges 804 // in SI that occur in PSI with a destination of BB will be 805 // activated. 806 std::set<ConstantInt*> PTIHandled; 807 for (unsigned i = 0, e = PredCases.size(); i != e; ++i) 808 if (PredCases[i].second == BB) { 809 PTIHandled.insert(PredCases[i].first); 810 std::swap(PredCases[i], PredCases.back()); 811 PredCases.pop_back(); 812 --i; --e; 813 } 814 815 // Okay, now we know which constants were sent to BB from the 816 // predecessor. Figure out where they will all go now. 817 for (unsigned i = 0, e = BBCases.size(); i != e; ++i) 818 if (PTIHandled.count(BBCases[i].first)) { 819 // If this is one we are capable of getting... 820 PredCases.push_back(BBCases[i]); 821 NewSuccessors.push_back(BBCases[i].second); 822 PTIHandled.erase(BBCases[i].first);// This constant is taken care of 823 } 824 825 // If there are any constants vectored to BB that TI doesn't handle, 826 // they must go to the default destination of TI. 827 for (std::set<ConstantInt*>::iterator I = PTIHandled.begin(), 828 E = PTIHandled.end(); I != E; ++I) { 829 PredCases.push_back(std::make_pair(*I, BBDefault)); 830 NewSuccessors.push_back(BBDefault); 831 } 832 } 833 834 // Okay, at this point, we know which new successor Pred will get. Make 835 // sure we update the number of entries in the PHI nodes for these 836 // successors. 837 for (unsigned i = 0, e = NewSuccessors.size(); i != e; ++i) 838 AddPredecessorToBlock(NewSuccessors[i], Pred, BB); 839 840 // Now that the successors are updated, create the new Switch instruction. 841 SwitchInst *NewSI = SwitchInst::Create(CV, PredDefault, 842 PredCases.size(), PTI); 843 for (unsigned i = 0, e = PredCases.size(); i != e; ++i) 844 NewSI->addCase(PredCases[i].first, PredCases[i].second); 845 846 Instruction *DeadCond = 0; 847 if (BranchInst *BI = dyn_cast<BranchInst>(PTI)) 848 // If PTI is a branch, remember the condition. 849 DeadCond = dyn_cast<Instruction>(BI->getCondition()); 850 Pred->getInstList().erase(PTI); 851 852 // If the condition is dead now, remove the instruction tree. 853 if (DeadCond) ErasePossiblyDeadInstructionTree(DeadCond); 854 855 // Okay, last check. If BB is still a successor of PSI, then we must 856 // have an infinite loop case. If so, add an infinitely looping block 857 // to handle the case to preserve the behavior of the code. 858 BasicBlock *InfLoopBlock = 0; 859 for (unsigned i = 0, e = NewSI->getNumSuccessors(); i != e; ++i) 860 if (NewSI->getSuccessor(i) == BB) { 861 if (InfLoopBlock == 0) { 862 // Insert it at the end of the function, because it's either code, 863 // or it won't matter if it's hot. :) 864 InfLoopBlock = BasicBlock::Create("infloop", BB->getParent()); 865 BranchInst::Create(InfLoopBlock, InfLoopBlock); 866 } 867 NewSI->setSuccessor(i, InfLoopBlock); 868 } 869 870 Changed = true; 871 } 872 } 873 return Changed; 874} 875 876/// HoistThenElseCodeToIf - Given a conditional branch that goes to BB1 and 877/// BB2, hoist any common code in the two blocks up into the branch block. The 878/// caller of this function guarantees that BI's block dominates BB1 and BB2. 879static bool HoistThenElseCodeToIf(BranchInst *BI) { 880 // This does very trivial matching, with limited scanning, to find identical 881 // instructions in the two blocks. In particular, we don't want to get into 882 // O(M*N) situations here where M and N are the sizes of BB1 and BB2. As 883 // such, we currently just scan for obviously identical instructions in an 884 // identical order. 885 BasicBlock *BB1 = BI->getSuccessor(0); // The true destination. 886 BasicBlock *BB2 = BI->getSuccessor(1); // The false destination 887 888 Instruction *I1 = BB1->begin(), *I2 = BB2->begin(); 889 if (I1->getOpcode() != I2->getOpcode() || isa<PHINode>(I1) || 890 isa<InvokeInst>(I1) || !I1->isIdenticalTo(I2)) 891 return false; 892 893 // If we get here, we can hoist at least one instruction. 894 BasicBlock *BIParent = BI->getParent(); 895 896 do { 897 // If we are hoisting the terminator instruction, don't move one (making a 898 // broken BB), instead clone it, and remove BI. 899 if (isa<TerminatorInst>(I1)) 900 goto HoistTerminator; 901 902 // For a normal instruction, we just move one to right before the branch, 903 // then replace all uses of the other with the first. Finally, we remove 904 // the now redundant second instruction. 905 BIParent->getInstList().splice(BI, BB1->getInstList(), I1); 906 if (!I2->use_empty()) 907 I2->replaceAllUsesWith(I1); 908 BB2->getInstList().erase(I2); 909 910 I1 = BB1->begin(); 911 I2 = BB2->begin(); 912 } while (I1->getOpcode() == I2->getOpcode() && I1->isIdenticalTo(I2)); 913 914 return true; 915 916HoistTerminator: 917 // Okay, it is safe to hoist the terminator. 918 Instruction *NT = I1->clone(); 919 BIParent->getInstList().insert(BI, NT); 920 if (NT->getType() != Type::VoidTy) { 921 I1->replaceAllUsesWith(NT); 922 I2->replaceAllUsesWith(NT); 923 NT->takeName(I1); 924 } 925 926 // Hoisting one of the terminators from our successor is a great thing. 927 // Unfortunately, the successors of the if/else blocks may have PHI nodes in 928 // them. If they do, all PHI entries for BB1/BB2 must agree for all PHI 929 // nodes, so we insert select instruction to compute the final result. 930 std::map<std::pair<Value*,Value*>, SelectInst*> InsertedSelects; 931 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI) { 932 PHINode *PN; 933 for (BasicBlock::iterator BBI = SI->begin(); 934 (PN = dyn_cast<PHINode>(BBI)); ++BBI) { 935 Value *BB1V = PN->getIncomingValueForBlock(BB1); 936 Value *BB2V = PN->getIncomingValueForBlock(BB2); 937 if (BB1V != BB2V) { 938 // These values do not agree. Insert a select instruction before NT 939 // that determines the right value. 940 SelectInst *&SI = InsertedSelects[std::make_pair(BB1V, BB2V)]; 941 if (SI == 0) 942 SI = SelectInst::Create(BI->getCondition(), BB1V, BB2V, 943 BB1V->getName()+"."+BB2V->getName(), NT); 944 // Make the PHI node use the select for all incoming values for BB1/BB2 945 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) 946 if (PN->getIncomingBlock(i) == BB1 || PN->getIncomingBlock(i) == BB2) 947 PN->setIncomingValue(i, SI); 948 } 949 } 950 } 951 952 // Update any PHI nodes in our new successors. 953 for (succ_iterator SI = succ_begin(BB1), E = succ_end(BB1); SI != E; ++SI) 954 AddPredecessorToBlock(*SI, BIParent, BB1); 955 956 BI->eraseFromParent(); 957 return true; 958} 959 960/// SpeculativelyExecuteBB - Given a conditional branch that goes to BB1 961/// and an BB2 and the only successor of BB1 is BB2, hoist simple code 962/// (for now, restricted to a single instruction that's side effect free) from 963/// the BB1 into the branch block to speculatively execute it. 964static bool SpeculativelyExecuteBB(BranchInst *BI, BasicBlock *BB1) { 965 // Only speculatively execution a single instruction (not counting the 966 // terminator) for now. 967 BasicBlock::iterator BBI = BB1->begin(); 968 ++BBI; // must have at least a terminator 969 if (BBI == BB1->end()) return false; // only one inst 970 ++BBI; 971 if (BBI != BB1->end()) return false; // more than 2 insts. 972 973 // Be conservative for now. FP select instruction can often be expensive. 974 Value *BrCond = BI->getCondition(); 975 if (isa<Instruction>(BrCond) && 976 cast<Instruction>(BrCond)->getOpcode() == Instruction::FCmp) 977 return false; 978 979 // If BB1 is actually on the false edge of the conditional branch, remember 980 // to swap the select operands later. 981 bool Invert = false; 982 if (BB1 != BI->getSuccessor(0)) { 983 assert(BB1 == BI->getSuccessor(1) && "No edge from 'if' block?"); 984 Invert = true; 985 } 986 987 // Turn 988 // BB: 989 // %t1 = icmp 990 // br i1 %t1, label %BB1, label %BB2 991 // BB1: 992 // %t3 = add %t2, c 993 // br label BB2 994 // BB2: 995 // => 996 // BB: 997 // %t1 = icmp 998 // %t4 = add %t2, c 999 // %t3 = select i1 %t1, %t2, %t3 1000 Instruction *I = BB1->begin(); 1001 switch (I->getOpcode()) { 1002 default: return false; // Not safe / profitable to hoist. 1003 case Instruction::Add: 1004 case Instruction::Sub: 1005 case Instruction::And: 1006 case Instruction::Or: 1007 case Instruction::Xor: 1008 case Instruction::Shl: 1009 case Instruction::LShr: 1010 case Instruction::AShr: 1011 if (!I->getOperand(0)->getType()->isInteger()) 1012 // FP arithmetic might trap. Not worth doing for vector ops. 1013 return false; 1014 break; // These are all cheap and non-trapping instructions. 1015 } 1016 1017 // Can we speculatively execute the instruction? And what is the value 1018 // if the condition is false? Consider the phi uses, if the incoming value 1019 // from the "if" block are all the same V, then V is the value of the 1020 // select if the condition is false. 1021 BasicBlock *BIParent = BI->getParent(); 1022 SmallVector<PHINode*, 4> PHIUses; 1023 Value *FalseV = NULL; 1024 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); 1025 UI != E; ++UI) { 1026 PHINode *PN = dyn_cast<PHINode>(UI); 1027 if (!PN) 1028 continue; 1029 PHIUses.push_back(PN); 1030 Value *PHIV = PN->getIncomingValueForBlock(BIParent); 1031 if (!FalseV) 1032 FalseV = PHIV; 1033 else if (FalseV != PHIV) 1034 return false; // Don't know the value when condition is false. 1035 } 1036 if (!FalseV) // Can this happen? 1037 return false; 1038 1039 // Do not hoist the instruction if any of its operands are defined but not 1040 // used in this BB. The transformation will prevent the operand from 1041 // being sunk into the use block. 1042 for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i) { 1043 Instruction *OpI = dyn_cast<Instruction>(*i); 1044 if (OpI && OpI->getParent() == BIParent && 1045 !OpI->isUsedInBasicBlock(BIParent)) 1046 return false; 1047 } 1048 1049 // If we get here, we can hoist the instruction. Try to place it 1050 // before the icmp instruction preceeding the conditional branch. 1051 BasicBlock::iterator InsertPos = BI; 1052 if (InsertPos != BIParent->begin()) 1053 --InsertPos; 1054 if (InsertPos == BrCond) { 1055 SmallPtrSet<Instruction *, 4> BB1Insns; 1056 for(BasicBlock::iterator BB1I = BB1->begin(), BB1E = BB1->end(); 1057 BB1I != BB1E; ++BB1I) 1058 BB1Insns.insert(BB1I); 1059 for(Value::use_iterator UI = BrCond->use_begin(), UE = BrCond->use_end(); 1060 UI != UE; ++UI) { 1061 Instruction *Use = cast<Instruction>(*UI); 1062 if (BB1Insns.count(Use)) { 1063 // If BrCond uses the instruction that place it just before 1064 // branch instruction. 1065 InsertPos = BI; 1066 break; 1067 } 1068 } 1069 } else 1070 InsertPos = BI; 1071 BIParent->getInstList().splice(InsertPos, BB1->getInstList(), I); 1072 1073 // Create a select whose true value is the speculatively executed value and 1074 // false value is the previously determined FalseV. 1075 SelectInst *SI; 1076 if (Invert) 1077 SI = SelectInst::Create(BrCond, FalseV, I, 1078 FalseV->getName() + "." + I->getName(), BI); 1079 else 1080 SI = SelectInst::Create(BrCond, I, FalseV, 1081 I->getName() + "." + FalseV->getName(), BI); 1082 1083 // Make the PHI node use the select for all incoming values for "then" and 1084 // "if" blocks. 1085 for (unsigned i = 0, e = PHIUses.size(); i != e; ++i) { 1086 PHINode *PN = PHIUses[i]; 1087 for (unsigned j = 0, ee = PN->getNumIncomingValues(); j != ee; ++j) 1088 if (PN->getIncomingBlock(j) == BB1 || 1089 PN->getIncomingBlock(j) == BIParent) 1090 PN->setIncomingValue(j, SI); 1091 } 1092 1093 ++NumSpeculations; 1094 return true; 1095} 1096 1097/// BlockIsSimpleEnoughToThreadThrough - Return true if we can thread a branch 1098/// across this block. 1099static bool BlockIsSimpleEnoughToThreadThrough(BasicBlock *BB) { 1100 BranchInst *BI = cast<BranchInst>(BB->getTerminator()); 1101 unsigned Size = 0; 1102 1103 // If this basic block contains anything other than a PHI (which controls the 1104 // branch) and branch itself, bail out. FIXME: improve this in the future. 1105 for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI, ++Size) { 1106 if (Size > 10) return false; // Don't clone large BB's. 1107 1108 // We can only support instructions that are do not define values that are 1109 // live outside of the current basic block. 1110 for (Value::use_iterator UI = BBI->use_begin(), E = BBI->use_end(); 1111 UI != E; ++UI) { 1112 Instruction *U = cast<Instruction>(*UI); 1113 if (U->getParent() != BB || isa<PHINode>(U)) return false; 1114 } 1115 1116 // Looks ok, continue checking. 1117 } 1118 1119 return true; 1120} 1121 1122/// FoldCondBranchOnPHI - If we have a conditional branch on a PHI node value 1123/// that is defined in the same block as the branch and if any PHI entries are 1124/// constants, thread edges corresponding to that entry to be branches to their 1125/// ultimate destination. 1126static bool FoldCondBranchOnPHI(BranchInst *BI) { 1127 BasicBlock *BB = BI->getParent(); 1128 PHINode *PN = dyn_cast<PHINode>(BI->getCondition()); 1129 // NOTE: we currently cannot transform this case if the PHI node is used 1130 // outside of the block. 1131 if (!PN || PN->getParent() != BB || !PN->hasOneUse()) 1132 return false; 1133 1134 // Degenerate case of a single entry PHI. 1135 if (PN->getNumIncomingValues() == 1) { 1136 if (PN->getIncomingValue(0) != PN) 1137 PN->replaceAllUsesWith(PN->getIncomingValue(0)); 1138 else 1139 PN->replaceAllUsesWith(UndefValue::get(PN->getType())); 1140 PN->eraseFromParent(); 1141 return true; 1142 } 1143 1144 // Now we know that this block has multiple preds and two succs. 1145 if (!BlockIsSimpleEnoughToThreadThrough(BB)) return false; 1146 1147 // Okay, this is a simple enough basic block. See if any phi values are 1148 // constants. 1149 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 1150 ConstantInt *CB; 1151 if ((CB = dyn_cast<ConstantInt>(PN->getIncomingValue(i))) && 1152 CB->getType() == Type::Int1Ty) { 1153 // Okay, we now know that all edges from PredBB should be revectored to 1154 // branch to RealDest. 1155 BasicBlock *PredBB = PN->getIncomingBlock(i); 1156 BasicBlock *RealDest = BI->getSuccessor(!CB->getZExtValue()); 1157 1158 if (RealDest == BB) continue; // Skip self loops. 1159 1160 // The dest block might have PHI nodes, other predecessors and other 1161 // difficult cases. Instead of being smart about this, just insert a new 1162 // block that jumps to the destination block, effectively splitting 1163 // the edge we are about to create. 1164 BasicBlock *EdgeBB = BasicBlock::Create(RealDest->getName()+".critedge", 1165 RealDest->getParent(), RealDest); 1166 BranchInst::Create(RealDest, EdgeBB); 1167 PHINode *PN; 1168 for (BasicBlock::iterator BBI = RealDest->begin(); 1169 (PN = dyn_cast<PHINode>(BBI)); ++BBI) { 1170 Value *V = PN->getIncomingValueForBlock(BB); 1171 PN->addIncoming(V, EdgeBB); 1172 } 1173 1174 // BB may have instructions that are being threaded over. Clone these 1175 // instructions into EdgeBB. We know that there will be no uses of the 1176 // cloned instructions outside of EdgeBB. 1177 BasicBlock::iterator InsertPt = EdgeBB->begin(); 1178 std::map<Value*, Value*> TranslateMap; // Track translated values. 1179 for (BasicBlock::iterator BBI = BB->begin(); &*BBI != BI; ++BBI) { 1180 if (PHINode *PN = dyn_cast<PHINode>(BBI)) { 1181 TranslateMap[PN] = PN->getIncomingValueForBlock(PredBB); 1182 } else { 1183 // Clone the instruction. 1184 Instruction *N = BBI->clone(); 1185 if (BBI->hasName()) N->setName(BBI->getName()+".c"); 1186 1187 // Update operands due to translation. 1188 for (User::op_iterator i = N->op_begin(), e = N->op_end(); 1189 i != e; ++i) { 1190 std::map<Value*, Value*>::iterator PI = 1191 TranslateMap.find(*i); 1192 if (PI != TranslateMap.end()) 1193 *i = PI->second; 1194 } 1195 1196 // Check for trivial simplification. 1197 if (Constant *C = ConstantFoldInstruction(N)) { 1198 TranslateMap[BBI] = C; 1199 delete N; // Constant folded away, don't need actual inst 1200 } else { 1201 // Insert the new instruction into its new home. 1202 EdgeBB->getInstList().insert(InsertPt, N); 1203 if (!BBI->use_empty()) 1204 TranslateMap[BBI] = N; 1205 } 1206 } 1207 } 1208 1209 // Loop over all of the edges from PredBB to BB, changing them to branch 1210 // to EdgeBB instead. 1211 TerminatorInst *PredBBTI = PredBB->getTerminator(); 1212 for (unsigned i = 0, e = PredBBTI->getNumSuccessors(); i != e; ++i) 1213 if (PredBBTI->getSuccessor(i) == BB) { 1214 BB->removePredecessor(PredBB); 1215 PredBBTI->setSuccessor(i, EdgeBB); 1216 } 1217 1218 // Recurse, simplifying any other constants. 1219 return FoldCondBranchOnPHI(BI) | true; 1220 } 1221 } 1222 1223 return false; 1224} 1225 1226/// FoldTwoEntryPHINode - Given a BB that starts with the specified two-entry 1227/// PHI node, see if we can eliminate it. 1228static bool FoldTwoEntryPHINode(PHINode *PN) { 1229 // Ok, this is a two entry PHI node. Check to see if this is a simple "if 1230 // statement", which has a very simple dominance structure. Basically, we 1231 // are trying to find the condition that is being branched on, which 1232 // subsequently causes this merge to happen. We really want control 1233 // dependence information for this check, but simplifycfg can't keep it up 1234 // to date, and this catches most of the cases we care about anyway. 1235 // 1236 BasicBlock *BB = PN->getParent(); 1237 BasicBlock *IfTrue, *IfFalse; 1238 Value *IfCond = GetIfCondition(BB, IfTrue, IfFalse); 1239 if (!IfCond) return false; 1240 1241 // Okay, we found that we can merge this two-entry phi node into a select. 1242 // Doing so would require us to fold *all* two entry phi nodes in this block. 1243 // At some point this becomes non-profitable (particularly if the target 1244 // doesn't support cmov's). Only do this transformation if there are two or 1245 // fewer PHI nodes in this block. 1246 unsigned NumPhis = 0; 1247 for (BasicBlock::iterator I = BB->begin(); isa<PHINode>(I); ++NumPhis, ++I) 1248 if (NumPhis > 2) 1249 return false; 1250 1251 DOUT << "FOUND IF CONDITION! " << *IfCond << " T: " 1252 << IfTrue->getName() << " F: " << IfFalse->getName() << "\n"; 1253 1254 // Loop over the PHI's seeing if we can promote them all to select 1255 // instructions. While we are at it, keep track of the instructions 1256 // that need to be moved to the dominating block. 1257 std::set<Instruction*> AggressiveInsts; 1258 1259 BasicBlock::iterator AfterPHIIt = BB->begin(); 1260 while (isa<PHINode>(AfterPHIIt)) { 1261 PHINode *PN = cast<PHINode>(AfterPHIIt++); 1262 if (PN->getIncomingValue(0) == PN->getIncomingValue(1)) { 1263 if (PN->getIncomingValue(0) != PN) 1264 PN->replaceAllUsesWith(PN->getIncomingValue(0)); 1265 else 1266 PN->replaceAllUsesWith(UndefValue::get(PN->getType())); 1267 } else if (!DominatesMergePoint(PN->getIncomingValue(0), BB, 1268 &AggressiveInsts) || 1269 !DominatesMergePoint(PN->getIncomingValue(1), BB, 1270 &AggressiveInsts)) { 1271 return false; 1272 } 1273 } 1274 1275 // If we all PHI nodes are promotable, check to make sure that all 1276 // instructions in the predecessor blocks can be promoted as well. If 1277 // not, we won't be able to get rid of the control flow, so it's not 1278 // worth promoting to select instructions. 1279 BasicBlock *DomBlock = 0, *IfBlock1 = 0, *IfBlock2 = 0; 1280 PN = cast<PHINode>(BB->begin()); 1281 BasicBlock *Pred = PN->getIncomingBlock(0); 1282 if (cast<BranchInst>(Pred->getTerminator())->isUnconditional()) { 1283 IfBlock1 = Pred; 1284 DomBlock = *pred_begin(Pred); 1285 for (BasicBlock::iterator I = Pred->begin(); 1286 !isa<TerminatorInst>(I); ++I) 1287 if (!AggressiveInsts.count(I)) { 1288 // This is not an aggressive instruction that we can promote. 1289 // Because of this, we won't be able to get rid of the control 1290 // flow, so the xform is not worth it. 1291 return false; 1292 } 1293 } 1294 1295 Pred = PN->getIncomingBlock(1); 1296 if (cast<BranchInst>(Pred->getTerminator())->isUnconditional()) { 1297 IfBlock2 = Pred; 1298 DomBlock = *pred_begin(Pred); 1299 for (BasicBlock::iterator I = Pred->begin(); 1300 !isa<TerminatorInst>(I); ++I) 1301 if (!AggressiveInsts.count(I)) { 1302 // This is not an aggressive instruction that we can promote. 1303 // Because of this, we won't be able to get rid of the control 1304 // flow, so the xform is not worth it. 1305 return false; 1306 } 1307 } 1308 1309 // If we can still promote the PHI nodes after this gauntlet of tests, 1310 // do all of the PHI's now. 1311 1312 // Move all 'aggressive' instructions, which are defined in the 1313 // conditional parts of the if's up to the dominating block. 1314 if (IfBlock1) { 1315 DomBlock->getInstList().splice(DomBlock->getTerminator(), 1316 IfBlock1->getInstList(), 1317 IfBlock1->begin(), 1318 IfBlock1->getTerminator()); 1319 } 1320 if (IfBlock2) { 1321 DomBlock->getInstList().splice(DomBlock->getTerminator(), 1322 IfBlock2->getInstList(), 1323 IfBlock2->begin(), 1324 IfBlock2->getTerminator()); 1325 } 1326 1327 while (PHINode *PN = dyn_cast<PHINode>(BB->begin())) { 1328 // Change the PHI node into a select instruction. 1329 Value *TrueVal = 1330 PN->getIncomingValue(PN->getIncomingBlock(0) == IfFalse); 1331 Value *FalseVal = 1332 PN->getIncomingValue(PN->getIncomingBlock(0) == IfTrue); 1333 1334 Value *NV = SelectInst::Create(IfCond, TrueVal, FalseVal, "", AfterPHIIt); 1335 PN->replaceAllUsesWith(NV); 1336 NV->takeName(PN); 1337 1338 BB->getInstList().erase(PN); 1339 } 1340 return true; 1341} 1342 1343/// SimplifyCondBranchToTwoReturns - If we found a conditional branch that goes 1344/// to two returning blocks, try to merge them together into one return, 1345/// introducing a select if the return values disagree. 1346static bool SimplifyCondBranchToTwoReturns(BranchInst *BI) { 1347 assert(BI->isConditional() && "Must be a conditional branch"); 1348 BasicBlock *TrueSucc = BI->getSuccessor(0); 1349 BasicBlock *FalseSucc = BI->getSuccessor(1); 1350 ReturnInst *TrueRet = cast<ReturnInst>(TrueSucc->getTerminator()); 1351 ReturnInst *FalseRet = cast<ReturnInst>(FalseSucc->getTerminator()); 1352 1353 // Check to ensure both blocks are empty (just a return) or optionally empty 1354 // with PHI nodes. If there are other instructions, merging would cause extra 1355 // computation on one path or the other. 1356 BasicBlock::iterator BBI = TrueRet; 1357 if (BBI != TrueSucc->begin() && !isa<PHINode>(--BBI)) 1358 return false; // Not empty with optional phi nodes. 1359 BBI = FalseRet; 1360 if (BBI != FalseSucc->begin() && !isa<PHINode>(--BBI)) 1361 return false; // Not empty with optional phi nodes. 1362 1363 // Okay, we found a branch that is going to two return nodes. If 1364 // there is no return value for this function, just change the 1365 // branch into a return. 1366 if (FalseRet->getNumOperands() == 0) { 1367 TrueSucc->removePredecessor(BI->getParent()); 1368 FalseSucc->removePredecessor(BI->getParent()); 1369 ReturnInst::Create(0, BI); 1370 BI->eraseFromParent(); 1371 return true; 1372 } 1373 1374 // Otherwise, figure out what the true and false return values are 1375 // so we can insert a new select instruction. 1376 Value *TrueValue = TrueRet->getReturnValue(); 1377 Value *FalseValue = FalseRet->getReturnValue(); 1378 1379 // Unwrap any PHI nodes in the return blocks. 1380 if (PHINode *TVPN = dyn_cast_or_null<PHINode>(TrueValue)) 1381 if (TVPN->getParent() == TrueSucc) 1382 TrueValue = TVPN->getIncomingValueForBlock(BI->getParent()); 1383 if (PHINode *FVPN = dyn_cast_or_null<PHINode>(FalseValue)) 1384 if (FVPN->getParent() == FalseSucc) 1385 FalseValue = FVPN->getIncomingValueForBlock(BI->getParent()); 1386 1387 // In order for this transformation to be safe, we must be able to 1388 // unconditionally execute both operands to the return. This is 1389 // normally the case, but we could have a potentially-trapping 1390 // constant expression that prevents this transformation from being 1391 // safe. 1392 if (ConstantExpr *TCV = dyn_cast_or_null<ConstantExpr>(TrueValue)) 1393 if (TCV->canTrap()) 1394 return false; 1395 if (ConstantExpr *FCV = dyn_cast_or_null<ConstantExpr>(FalseValue)) 1396 if (FCV->canTrap()) 1397 return false; 1398 1399 // Okay, we collected all the mapped values and checked them for sanity, and 1400 // defined to really do this transformation. First, update the CFG. 1401 TrueSucc->removePredecessor(BI->getParent()); 1402 FalseSucc->removePredecessor(BI->getParent()); 1403 1404 // Insert select instructions where needed. 1405 Value *BrCond = BI->getCondition(); 1406 if (TrueValue) { 1407 // Insert a select if the results differ. 1408 if (TrueValue == FalseValue || isa<UndefValue>(FalseValue)) { 1409 } else if (isa<UndefValue>(TrueValue)) { 1410 TrueValue = FalseValue; 1411 } else { 1412 TrueValue = SelectInst::Create(BrCond, TrueValue, 1413 FalseValue, "retval", BI); 1414 } 1415 } 1416 1417 Value *RI = !TrueValue ? 1418 ReturnInst::Create(BI) : 1419 ReturnInst::Create(TrueValue, BI); 1420 1421 DOUT << "\nCHANGING BRANCH TO TWO RETURNS INTO SELECT:" 1422 << "\n " << *BI << "NewRet = " << *RI 1423 << "TRUEBLOCK: " << *TrueSucc << "FALSEBLOCK: "<< *FalseSucc; 1424 1425 BI->eraseFromParent(); 1426 1427 if (Instruction *BrCondI = dyn_cast<Instruction>(BrCond)) 1428 ErasePossiblyDeadInstructionTree(BrCondI); 1429 return true; 1430} 1431 1432/// FoldBranchToCommonDest - If this basic block is ONLY a setcc and a branch, 1433/// and if a predecessor branches to us and one of our successors, fold the 1434/// setcc into the predecessor and use logical operations to pick the right 1435/// destination. 1436static bool FoldBranchToCommonDest(BranchInst *BI) { 1437 BasicBlock *BB = BI->getParent(); 1438 Instruction *Cond = dyn_cast<Instruction>(BI->getCondition()); 1439 if (Cond == 0) return false; 1440 1441 1442 // Only allow this if the condition is a simple instruction that can be 1443 // executed unconditionally. It must be in the same block as the branch, and 1444 // must be at the front of the block. 1445 if ((!isa<CmpInst>(Cond) && !isa<BinaryOperator>(Cond)) || 1446 Cond->getParent() != BB || &BB->front() != Cond || !Cond->hasOneUse()) 1447 return false; 1448 1449 // Make sure the instruction after the condition is the cond branch. 1450 BasicBlock::iterator CondIt = Cond; ++CondIt; 1451 if (&*CondIt != BI) 1452 return false; 1453 1454 // Finally, don't infinitely unroll conditional loops. 1455 BasicBlock *TrueDest = BI->getSuccessor(0); 1456 BasicBlock *FalseDest = BI->getSuccessor(1); 1457 if (TrueDest == BB || FalseDest == BB) 1458 return false; 1459 1460 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) { 1461 BasicBlock *PredBlock = *PI; 1462 BranchInst *PBI = dyn_cast<BranchInst>(PredBlock->getTerminator()); 1463 // Check that we have two conditional branches. If there is a PHI node in 1464 // the common successor, verify that the same value flows in from both 1465 // blocks. 1466 if (PBI == 0 || PBI->isUnconditional() || 1467 !SafeToMergeTerminators(BI, PBI)) 1468 continue; 1469 1470 Instruction::BinaryOps Opc; 1471 bool InvertPredCond = false; 1472 1473 if (PBI->getSuccessor(0) == TrueDest) 1474 Opc = Instruction::Or; 1475 else if (PBI->getSuccessor(1) == FalseDest) 1476 Opc = Instruction::And; 1477 else if (PBI->getSuccessor(0) == FalseDest) 1478 Opc = Instruction::And, InvertPredCond = true; 1479 else if (PBI->getSuccessor(1) == TrueDest) 1480 Opc = Instruction::Or, InvertPredCond = true; 1481 else 1482 continue; 1483 1484 // If we need to invert the condition in the pred block to match, do so now. 1485 if (InvertPredCond) { 1486 Value *NewCond = 1487 BinaryOperator::CreateNot(PBI->getCondition(), 1488 PBI->getCondition()->getName()+".not", PBI); 1489 PBI->setCondition(NewCond); 1490 BasicBlock *OldTrue = PBI->getSuccessor(0); 1491 BasicBlock *OldFalse = PBI->getSuccessor(1); 1492 PBI->setSuccessor(0, OldFalse); 1493 PBI->setSuccessor(1, OldTrue); 1494 } 1495 1496 // Clone Cond into the predecessor basic block, and or/and the 1497 // two conditions together. 1498 Instruction *New = Cond->clone(); 1499 PredBlock->getInstList().insert(PBI, New); 1500 New->takeName(Cond); 1501 Cond->setName(New->getName()+".old"); 1502 1503 Value *NewCond = BinaryOperator::Create(Opc, PBI->getCondition(), 1504 New, "or.cond", PBI); 1505 PBI->setCondition(NewCond); 1506 if (PBI->getSuccessor(0) == BB) { 1507 AddPredecessorToBlock(TrueDest, PredBlock, BB); 1508 PBI->setSuccessor(0, TrueDest); 1509 } 1510 if (PBI->getSuccessor(1) == BB) { 1511 AddPredecessorToBlock(FalseDest, PredBlock, BB); 1512 PBI->setSuccessor(1, FalseDest); 1513 } 1514 return true; 1515 } 1516 return false; 1517} 1518 1519/// SimplifyCondBranchToCondBranch - If we have a conditional branch as a 1520/// predecessor of another block, this function tries to simplify it. We know 1521/// that PBI and BI are both conditional branches, and BI is in one of the 1522/// successor blocks of PBI - PBI branches to BI. 1523static bool SimplifyCondBranchToCondBranch(BranchInst *PBI, BranchInst *BI) { 1524 assert(PBI->isConditional() && BI->isConditional()); 1525 BasicBlock *BB = BI->getParent(); 1526 1527 // If this block ends with a branch instruction, and if there is a 1528 // predecessor that ends on a branch of the same condition, make 1529 // this conditional branch redundant. 1530 if (PBI->getCondition() == BI->getCondition() && 1531 PBI->getSuccessor(0) != PBI->getSuccessor(1)) { 1532 // Okay, the outcome of this conditional branch is statically 1533 // knowable. If this block had a single pred, handle specially. 1534 if (BB->getSinglePredecessor()) { 1535 // Turn this into a branch on constant. 1536 bool CondIsTrue = PBI->getSuccessor(0) == BB; 1537 BI->setCondition(ConstantInt::get(Type::Int1Ty, CondIsTrue)); 1538 return true; // Nuke the branch on constant. 1539 } 1540 1541 // Otherwise, if there are multiple predecessors, insert a PHI that merges 1542 // in the constant and simplify the block result. Subsequent passes of 1543 // simplifycfg will thread the block. 1544 if (BlockIsSimpleEnoughToThreadThrough(BB)) { 1545 PHINode *NewPN = PHINode::Create(Type::Int1Ty, 1546 BI->getCondition()->getName() + ".pr", 1547 BB->begin()); 1548 // Okay, we're going to insert the PHI node. Since PBI is not the only 1549 // predecessor, compute the PHI'd conditional value for all of the preds. 1550 // Any predecessor where the condition is not computable we keep symbolic. 1551 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) 1552 if ((PBI = dyn_cast<BranchInst>((*PI)->getTerminator())) && 1553 PBI != BI && PBI->isConditional() && 1554 PBI->getCondition() == BI->getCondition() && 1555 PBI->getSuccessor(0) != PBI->getSuccessor(1)) { 1556 bool CondIsTrue = PBI->getSuccessor(0) == BB; 1557 NewPN->addIncoming(ConstantInt::get(Type::Int1Ty, 1558 CondIsTrue), *PI); 1559 } else { 1560 NewPN->addIncoming(BI->getCondition(), *PI); 1561 } 1562 1563 BI->setCondition(NewPN); 1564 return true; 1565 } 1566 } 1567 1568 // If this is a conditional branch in an empty block, and if any 1569 // predecessors is a conditional branch to one of our destinations, 1570 // fold the conditions into logical ops and one cond br. 1571 if (&BB->front() != BI) 1572 return false; 1573 1574 int PBIOp, BIOp; 1575 if (PBI->getSuccessor(0) == BI->getSuccessor(0)) 1576 PBIOp = BIOp = 0; 1577 else if (PBI->getSuccessor(0) == BI->getSuccessor(1)) 1578 PBIOp = 0, BIOp = 1; 1579 else if (PBI->getSuccessor(1) == BI->getSuccessor(0)) 1580 PBIOp = 1, BIOp = 0; 1581 else if (PBI->getSuccessor(1) == BI->getSuccessor(1)) 1582 PBIOp = BIOp = 1; 1583 else 1584 return false; 1585 1586 // Check to make sure that the other destination of this branch 1587 // isn't BB itself. If so, this is an infinite loop that will 1588 // keep getting unwound. 1589 if (PBI->getSuccessor(PBIOp) == BB) 1590 return false; 1591 1592 // Do not perform this transformation if it would require 1593 // insertion of a large number of select instructions. For targets 1594 // without predication/cmovs, this is a big pessimization. 1595 BasicBlock *CommonDest = PBI->getSuccessor(PBIOp); 1596 1597 unsigned NumPhis = 0; 1598 for (BasicBlock::iterator II = CommonDest->begin(); 1599 isa<PHINode>(II); ++II, ++NumPhis) 1600 if (NumPhis > 2) // Disable this xform. 1601 return false; 1602 1603 // Finally, if everything is ok, fold the branches to logical ops. 1604 BasicBlock *OtherDest = BI->getSuccessor(BIOp ^ 1); 1605 1606 DOUT << "FOLDING BRs:" << *PBI->getParent() 1607 << "AND: " << *BI->getParent(); 1608 1609 1610 // If OtherDest *is* BB, then BB is a basic block with a single conditional 1611 // branch in it, where one edge (OtherDest) goes back to itself but the other 1612 // exits. We don't *know* that the program avoids the infinite loop 1613 // (even though that seems likely). If we do this xform naively, we'll end up 1614 // recursively unpeeling the loop. Since we know that (after the xform is 1615 // done) that the block *is* infinite if reached, we just make it an obviously 1616 // infinite loop with no cond branch. 1617 if (OtherDest == BB) { 1618 // Insert it at the end of the function, because it's either code, 1619 // or it won't matter if it's hot. :) 1620 BasicBlock *InfLoopBlock = BasicBlock::Create("infloop", BB->getParent()); 1621 BranchInst::Create(InfLoopBlock, InfLoopBlock); 1622 OtherDest = InfLoopBlock; 1623 } 1624 1625 DOUT << *PBI->getParent()->getParent(); 1626 1627 // BI may have other predecessors. Because of this, we leave 1628 // it alone, but modify PBI. 1629 1630 // Make sure we get to CommonDest on True&True directions. 1631 Value *PBICond = PBI->getCondition(); 1632 if (PBIOp) 1633 PBICond = BinaryOperator::CreateNot(PBICond, 1634 PBICond->getName()+".not", 1635 PBI); 1636 Value *BICond = BI->getCondition(); 1637 if (BIOp) 1638 BICond = BinaryOperator::CreateNot(BICond, 1639 BICond->getName()+".not", 1640 PBI); 1641 // Merge the conditions. 1642 Value *Cond = BinaryOperator::CreateOr(PBICond, BICond, "brmerge", PBI); 1643 1644 // Modify PBI to branch on the new condition to the new dests. 1645 PBI->setCondition(Cond); 1646 PBI->setSuccessor(0, CommonDest); 1647 PBI->setSuccessor(1, OtherDest); 1648 1649 // OtherDest may have phi nodes. If so, add an entry from PBI's 1650 // block that are identical to the entries for BI's block. 1651 PHINode *PN; 1652 for (BasicBlock::iterator II = OtherDest->begin(); 1653 (PN = dyn_cast<PHINode>(II)); ++II) { 1654 Value *V = PN->getIncomingValueForBlock(BB); 1655 PN->addIncoming(V, PBI->getParent()); 1656 } 1657 1658 // We know that the CommonDest already had an edge from PBI to 1659 // it. If it has PHIs though, the PHIs may have different 1660 // entries for BB and PBI's BB. If so, insert a select to make 1661 // them agree. 1662 for (BasicBlock::iterator II = CommonDest->begin(); 1663 (PN = dyn_cast<PHINode>(II)); ++II) { 1664 Value *BIV = PN->getIncomingValueForBlock(BB); 1665 unsigned PBBIdx = PN->getBasicBlockIndex(PBI->getParent()); 1666 Value *PBIV = PN->getIncomingValue(PBBIdx); 1667 if (BIV != PBIV) { 1668 // Insert a select in PBI to pick the right value. 1669 Value *NV = SelectInst::Create(PBICond, PBIV, BIV, 1670 PBIV->getName()+".mux", PBI); 1671 PN->setIncomingValue(PBBIdx, NV); 1672 } 1673 } 1674 1675 DOUT << "INTO: " << *PBI->getParent(); 1676 1677 DOUT << *PBI->getParent()->getParent(); 1678 1679 // This basic block is probably dead. We know it has at least 1680 // one fewer predecessor. 1681 return true; 1682} 1683 1684 1685namespace { 1686 /// ConstantIntOrdering - This class implements a stable ordering of constant 1687 /// integers that does not depend on their address. This is important for 1688 /// applications that sort ConstantInt's to ensure uniqueness. 1689 struct ConstantIntOrdering { 1690 bool operator()(const ConstantInt *LHS, const ConstantInt *RHS) const { 1691 return LHS->getValue().ult(RHS->getValue()); 1692 } 1693 }; 1694} 1695 1696// SimplifyCFG - This function is used to do simplification of a CFG. For 1697// example, it adjusts branches to branches to eliminate the extra hop, it 1698// eliminates unreachable basic blocks, and does other "peephole" optimization 1699// of the CFG. It returns true if a modification was made. 1700// 1701// WARNING: The entry node of a function may not be simplified. 1702// 1703bool llvm::SimplifyCFG(BasicBlock *BB) { 1704 bool Changed = false; 1705 Function *M = BB->getParent(); 1706 1707 assert(BB && BB->getParent() && "Block not embedded in function!"); 1708 assert(BB->getTerminator() && "Degenerate basic block encountered!"); 1709 assert(&BB->getParent()->getEntryBlock() != BB && 1710 "Can't Simplify entry block!"); 1711 1712 // Remove basic blocks that have no predecessors... which are unreachable. 1713 if ((pred_begin(BB) == pred_end(BB)) || 1714 (*pred_begin(BB) == BB && ++pred_begin(BB) == pred_end(BB))) { 1715 DOUT << "Removing BB: \n" << *BB; 1716 1717 // Loop through all of our successors and make sure they know that one 1718 // of their predecessors is going away. 1719 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI) 1720 SI->removePredecessor(BB); 1721 1722 while (!BB->empty()) { 1723 Instruction &I = BB->back(); 1724 // If this instruction is used, replace uses with an arbitrary 1725 // value. Because control flow can't get here, we don't care 1726 // what we replace the value with. Note that since this block is 1727 // unreachable, and all values contained within it must dominate their 1728 // uses, that all uses will eventually be removed. 1729 if (!I.use_empty()) 1730 // Make all users of this instruction use undef instead 1731 I.replaceAllUsesWith(UndefValue::get(I.getType())); 1732 1733 // Remove the instruction from the basic block 1734 BB->getInstList().pop_back(); 1735 } 1736 M->getBasicBlockList().erase(BB); 1737 return true; 1738 } 1739 1740 // Check to see if we can constant propagate this terminator instruction 1741 // away... 1742 Changed |= ConstantFoldTerminator(BB); 1743 1744 // If there is a trivial two-entry PHI node in this basic block, and we can 1745 // eliminate it, do so now. 1746 if (PHINode *PN = dyn_cast<PHINode>(BB->begin())) 1747 if (PN->getNumIncomingValues() == 2) 1748 Changed |= FoldTwoEntryPHINode(PN); 1749 1750 // If this is a returning block with only PHI nodes in it, fold the return 1751 // instruction into any unconditional branch predecessors. 1752 // 1753 // If any predecessor is a conditional branch that just selects among 1754 // different return values, fold the replace the branch/return with a select 1755 // and return. 1756 if (ReturnInst *RI = dyn_cast<ReturnInst>(BB->getTerminator())) { 1757 BasicBlock::iterator BBI = BB->getTerminator(); 1758 if (BBI == BB->begin() || isa<PHINode>(--BBI)) { 1759 // Find predecessors that end with branches. 1760 SmallVector<BasicBlock*, 8> UncondBranchPreds; 1761 SmallVector<BranchInst*, 8> CondBranchPreds; 1762 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) { 1763 TerminatorInst *PTI = (*PI)->getTerminator(); 1764 if (BranchInst *BI = dyn_cast<BranchInst>(PTI)) { 1765 if (BI->isUnconditional()) 1766 UncondBranchPreds.push_back(*PI); 1767 else 1768 CondBranchPreds.push_back(BI); 1769 } 1770 } 1771 1772 // If we found some, do the transformation! 1773 if (!UncondBranchPreds.empty()) { 1774 while (!UncondBranchPreds.empty()) { 1775 BasicBlock *Pred = UncondBranchPreds.back(); 1776 DOUT << "FOLDING: " << *BB 1777 << "INTO UNCOND BRANCH PRED: " << *Pred; 1778 UncondBranchPreds.pop_back(); 1779 Instruction *UncondBranch = Pred->getTerminator(); 1780 // Clone the return and add it to the end of the predecessor. 1781 Instruction *NewRet = RI->clone(); 1782 Pred->getInstList().push_back(NewRet); 1783 1784 // If the return instruction returns a value, and if the value was a 1785 // PHI node in "BB", propagate the right value into the return. 1786 for (User::op_iterator i = NewRet->op_begin(), e = NewRet->op_end(); 1787 i != e; ++i) 1788 if (PHINode *PN = dyn_cast<PHINode>(*i)) 1789 if (PN->getParent() == BB) 1790 *i = PN->getIncomingValueForBlock(Pred); 1791 1792 // Update any PHI nodes in the returning block to realize that we no 1793 // longer branch to them. 1794 BB->removePredecessor(Pred); 1795 Pred->getInstList().erase(UncondBranch); 1796 } 1797 1798 // If we eliminated all predecessors of the block, delete the block now. 1799 if (pred_begin(BB) == pred_end(BB)) 1800 // We know there are no successors, so just nuke the block. 1801 M->getBasicBlockList().erase(BB); 1802 1803 return true; 1804 } 1805 1806 // Check out all of the conditional branches going to this return 1807 // instruction. If any of them just select between returns, change the 1808 // branch itself into a select/return pair. 1809 while (!CondBranchPreds.empty()) { 1810 BranchInst *BI = CondBranchPreds.back(); 1811 CondBranchPreds.pop_back(); 1812 1813 // Check to see if the non-BB successor is also a return block. 1814 if (isa<ReturnInst>(BI->getSuccessor(0)->getTerminator()) && 1815 isa<ReturnInst>(BI->getSuccessor(1)->getTerminator()) && 1816 SimplifyCondBranchToTwoReturns(BI)) 1817 return true; 1818 } 1819 } 1820 } else if (isa<UnwindInst>(BB->begin())) { 1821 // Check to see if the first instruction in this block is just an unwind. 1822 // If so, replace any invoke instructions which use this as an exception 1823 // destination with call instructions, and any unconditional branch 1824 // predecessor with an unwind. 1825 // 1826 SmallVector<BasicBlock*, 8> Preds(pred_begin(BB), pred_end(BB)); 1827 while (!Preds.empty()) { 1828 BasicBlock *Pred = Preds.back(); 1829 if (BranchInst *BI = dyn_cast<BranchInst>(Pred->getTerminator())) { 1830 if (BI->isUnconditional()) { 1831 Pred->getInstList().pop_back(); // nuke uncond branch 1832 new UnwindInst(Pred); // Use unwind. 1833 Changed = true; 1834 } 1835 } else if (InvokeInst *II = dyn_cast<InvokeInst>(Pred->getTerminator())) 1836 if (II->getUnwindDest() == BB) { 1837 // Insert a new branch instruction before the invoke, because this 1838 // is now a fall through... 1839 BranchInst *BI = BranchInst::Create(II->getNormalDest(), II); 1840 Pred->getInstList().remove(II); // Take out of symbol table 1841 1842 // Insert the call now... 1843 SmallVector<Value*,8> Args(II->op_begin()+3, II->op_end()); 1844 CallInst *CI = CallInst::Create(II->getCalledValue(), 1845 Args.begin(), Args.end(), 1846 II->getName(), BI); 1847 CI->setCallingConv(II->getCallingConv()); 1848 CI->setParamAttrs(II->getParamAttrs()); 1849 // If the invoke produced a value, the Call now does instead 1850 II->replaceAllUsesWith(CI); 1851 delete II; 1852 Changed = true; 1853 } 1854 1855 Preds.pop_back(); 1856 } 1857 1858 // If this block is now dead, remove it. 1859 if (pred_begin(BB) == pred_end(BB)) { 1860 // We know there are no successors, so just nuke the block. 1861 M->getBasicBlockList().erase(BB); 1862 return true; 1863 } 1864 1865 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator())) { 1866 if (isValueEqualityComparison(SI)) { 1867 // If we only have one predecessor, and if it is a branch on this value, 1868 // see if that predecessor totally determines the outcome of this switch. 1869 if (BasicBlock *OnlyPred = BB->getSinglePredecessor()) 1870 if (SimplifyEqualityComparisonWithOnlyPredecessor(SI, OnlyPred)) 1871 return SimplifyCFG(BB) || 1; 1872 1873 // If the block only contains the switch, see if we can fold the block 1874 // away into any preds. 1875 if (SI == &BB->front()) 1876 if (FoldValueComparisonIntoPredecessors(SI)) 1877 return SimplifyCFG(BB) || 1; 1878 } 1879 } else if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) { 1880 if (BI->isUnconditional()) { 1881 BasicBlock::iterator BBI = BB->getFirstNonPHI(); 1882 1883 BasicBlock *Succ = BI->getSuccessor(0); 1884 if (BBI->isTerminator() && // Terminator is the only non-phi instruction! 1885 Succ != BB) // Don't hurt infinite loops! 1886 if (TryToSimplifyUncondBranchFromEmptyBlock(BB, Succ)) 1887 return true; 1888 1889 } else { // Conditional branch 1890 if (isValueEqualityComparison(BI)) { 1891 // If we only have one predecessor, and if it is a branch on this value, 1892 // see if that predecessor totally determines the outcome of this 1893 // switch. 1894 if (BasicBlock *OnlyPred = BB->getSinglePredecessor()) 1895 if (SimplifyEqualityComparisonWithOnlyPredecessor(BI, OnlyPred)) 1896 return SimplifyCFG(BB) || 1; 1897 1898 // This block must be empty, except for the setcond inst, if it exists. 1899 BasicBlock::iterator I = BB->begin(); 1900 if (&*I == BI || 1901 (&*I == cast<Instruction>(BI->getCondition()) && 1902 &*++I == BI)) 1903 if (FoldValueComparisonIntoPredecessors(BI)) 1904 return SimplifyCFG(BB) | true; 1905 } 1906 1907 // If this is a branch on a phi node in the current block, thread control 1908 // through this block if any PHI node entries are constants. 1909 if (PHINode *PN = dyn_cast<PHINode>(BI->getCondition())) 1910 if (PN->getParent() == BI->getParent()) 1911 if (FoldCondBranchOnPHI(BI)) 1912 return SimplifyCFG(BB) | true; 1913 1914 // If this basic block is ONLY a setcc and a branch, and if a predecessor 1915 // branches to us and one of our successors, fold the setcc into the 1916 // predecessor and use logical operations to pick the right destination. 1917 if (FoldBranchToCommonDest(BI)) 1918 return SimplifyCFG(BB) | 1; 1919 1920 1921 // Scan predecessor blocks for conditional branches. 1922 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) 1923 if (BranchInst *PBI = dyn_cast<BranchInst>((*PI)->getTerminator())) 1924 if (PBI != BI && PBI->isConditional()) 1925 if (SimplifyCondBranchToCondBranch(PBI, BI)) 1926 return SimplifyCFG(BB) | true; 1927 } 1928 } else if (isa<UnreachableInst>(BB->getTerminator())) { 1929 // If there are any instructions immediately before the unreachable that can 1930 // be removed, do so. 1931 Instruction *Unreachable = BB->getTerminator(); 1932 while (Unreachable != BB->begin()) { 1933 BasicBlock::iterator BBI = Unreachable; 1934 --BBI; 1935 if (isa<CallInst>(BBI)) break; 1936 // Delete this instruction 1937 BB->getInstList().erase(BBI); 1938 Changed = true; 1939 } 1940 1941 // If the unreachable instruction is the first in the block, take a gander 1942 // at all of the predecessors of this instruction, and simplify them. 1943 if (&BB->front() == Unreachable) { 1944 SmallVector<BasicBlock*, 8> Preds(pred_begin(BB), pred_end(BB)); 1945 for (unsigned i = 0, e = Preds.size(); i != e; ++i) { 1946 TerminatorInst *TI = Preds[i]->getTerminator(); 1947 1948 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) { 1949 if (BI->isUnconditional()) { 1950 if (BI->getSuccessor(0) == BB) { 1951 new UnreachableInst(TI); 1952 TI->eraseFromParent(); 1953 Changed = true; 1954 } 1955 } else { 1956 if (BI->getSuccessor(0) == BB) { 1957 BranchInst::Create(BI->getSuccessor(1), BI); 1958 BI->eraseFromParent(); 1959 } else if (BI->getSuccessor(1) == BB) { 1960 BranchInst::Create(BI->getSuccessor(0), BI); 1961 BI->eraseFromParent(); 1962 Changed = true; 1963 } 1964 } 1965 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) { 1966 for (unsigned i = 1, e = SI->getNumCases(); i != e; ++i) 1967 if (SI->getSuccessor(i) == BB) { 1968 BB->removePredecessor(SI->getParent()); 1969 SI->removeCase(i); 1970 --i; --e; 1971 Changed = true; 1972 } 1973 // If the default value is unreachable, figure out the most popular 1974 // destination and make it the default. 1975 if (SI->getSuccessor(0) == BB) { 1976 std::map<BasicBlock*, unsigned> Popularity; 1977 for (unsigned i = 1, e = SI->getNumCases(); i != e; ++i) 1978 Popularity[SI->getSuccessor(i)]++; 1979 1980 // Find the most popular block. 1981 unsigned MaxPop = 0; 1982 BasicBlock *MaxBlock = 0; 1983 for (std::map<BasicBlock*, unsigned>::iterator 1984 I = Popularity.begin(), E = Popularity.end(); I != E; ++I) { 1985 if (I->second > MaxPop) { 1986 MaxPop = I->second; 1987 MaxBlock = I->first; 1988 } 1989 } 1990 if (MaxBlock) { 1991 // Make this the new default, allowing us to delete any explicit 1992 // edges to it. 1993 SI->setSuccessor(0, MaxBlock); 1994 Changed = true; 1995 1996 // If MaxBlock has phinodes in it, remove MaxPop-1 entries from 1997 // it. 1998 if (isa<PHINode>(MaxBlock->begin())) 1999 for (unsigned i = 0; i != MaxPop-1; ++i) 2000 MaxBlock->removePredecessor(SI->getParent()); 2001 2002 for (unsigned i = 1, e = SI->getNumCases(); i != e; ++i) 2003 if (SI->getSuccessor(i) == MaxBlock) { 2004 SI->removeCase(i); 2005 --i; --e; 2006 } 2007 } 2008 } 2009 } else if (InvokeInst *II = dyn_cast<InvokeInst>(TI)) { 2010 if (II->getUnwindDest() == BB) { 2011 // Convert the invoke to a call instruction. This would be a good 2012 // place to note that the call does not throw though. 2013 BranchInst *BI = BranchInst::Create(II->getNormalDest(), II); 2014 II->removeFromParent(); // Take out of symbol table 2015 2016 // Insert the call now... 2017 SmallVector<Value*, 8> Args(II->op_begin()+3, II->op_end()); 2018 CallInst *CI = CallInst::Create(II->getCalledValue(), 2019 Args.begin(), Args.end(), 2020 II->getName(), BI); 2021 CI->setCallingConv(II->getCallingConv()); 2022 CI->setParamAttrs(II->getParamAttrs()); 2023 // If the invoke produced a value, the Call does now instead. 2024 II->replaceAllUsesWith(CI); 2025 delete II; 2026 Changed = true; 2027 } 2028 } 2029 } 2030 2031 // If this block is now dead, remove it. 2032 if (pred_begin(BB) == pred_end(BB)) { 2033 // We know there are no successors, so just nuke the block. 2034 M->getBasicBlockList().erase(BB); 2035 return true; 2036 } 2037 } 2038 } 2039 2040 // Merge basic blocks into their predecessor if there is only one distinct 2041 // pred, and if there is only one distinct successor of the predecessor, and 2042 // if there are no PHI nodes. 2043 // 2044 if (MergeBlockIntoPredecessor(BB)) 2045 return true; 2046 2047 // Otherwise, if this block only has a single predecessor, and if that block 2048 // is a conditional branch, see if we can hoist any code from this block up 2049 // into our predecessor. 2050 pred_iterator PI(pred_begin(BB)), PE(pred_end(BB)); 2051 BasicBlock *OnlyPred = *PI++; 2052 for (; PI != PE; ++PI) // Search all predecessors, see if they are all same 2053 if (*PI != OnlyPred) { 2054 OnlyPred = 0; // There are multiple different predecessors... 2055 break; 2056 } 2057 2058 if (OnlyPred) 2059 if (BranchInst *BI = dyn_cast<BranchInst>(OnlyPred->getTerminator())) 2060 if (BI->isConditional()) { 2061 // Get the other block. 2062 BasicBlock *OtherBB = BI->getSuccessor(BI->getSuccessor(0) == BB); 2063 PI = pred_begin(OtherBB); 2064 ++PI; 2065 2066 if (PI == pred_end(OtherBB)) { 2067 // We have a conditional branch to two blocks that are only reachable 2068 // from the condbr. We know that the condbr dominates the two blocks, 2069 // so see if there is any identical code in the "then" and "else" 2070 // blocks. If so, we can hoist it up to the branching block. 2071 Changed |= HoistThenElseCodeToIf(BI); 2072 } else { 2073 BasicBlock* OnlySucc = NULL; 2074 for (succ_iterator SI = succ_begin(BB), SE = succ_end(BB); 2075 SI != SE; ++SI) { 2076 if (!OnlySucc) 2077 OnlySucc = *SI; 2078 else if (*SI != OnlySucc) { 2079 OnlySucc = 0; // There are multiple distinct successors! 2080 break; 2081 } 2082 } 2083 2084 if (OnlySucc == OtherBB) { 2085 // If BB's only successor is the other successor of the predecessor, 2086 // i.e. a triangle, see if we can hoist any code from this block up 2087 // to the "if" block. 2088 Changed |= SpeculativelyExecuteBB(BI, BB); 2089 } 2090 } 2091 } 2092 2093 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) 2094 if (BranchInst *BI = dyn_cast<BranchInst>((*PI)->getTerminator())) 2095 // Change br (X == 0 | X == 1), T, F into a switch instruction. 2096 if (BI->isConditional() && isa<Instruction>(BI->getCondition())) { 2097 Instruction *Cond = cast<Instruction>(BI->getCondition()); 2098 // If this is a bunch of seteq's or'd together, or if it's a bunch of 2099 // 'setne's and'ed together, collect them. 2100 Value *CompVal = 0; 2101 std::vector<ConstantInt*> Values; 2102 bool TrueWhenEqual = GatherValueComparisons(Cond, CompVal, Values); 2103 if (CompVal && CompVal->getType()->isInteger()) { 2104 // There might be duplicate constants in the list, which the switch 2105 // instruction can't handle, remove them now. 2106 std::sort(Values.begin(), Values.end(), ConstantIntOrdering()); 2107 Values.erase(std::unique(Values.begin(), Values.end()), Values.end()); 2108 2109 // Figure out which block is which destination. 2110 BasicBlock *DefaultBB = BI->getSuccessor(1); 2111 BasicBlock *EdgeBB = BI->getSuccessor(0); 2112 if (!TrueWhenEqual) std::swap(DefaultBB, EdgeBB); 2113 2114 // Create the new switch instruction now. 2115 SwitchInst *New = SwitchInst::Create(CompVal, DefaultBB, 2116 Values.size(), BI); 2117 2118 // Add all of the 'cases' to the switch instruction. 2119 for (unsigned i = 0, e = Values.size(); i != e; ++i) 2120 New->addCase(Values[i], EdgeBB); 2121 2122 // We added edges from PI to the EdgeBB. As such, if there were any 2123 // PHI nodes in EdgeBB, they need entries to be added corresponding to 2124 // the number of edges added. 2125 for (BasicBlock::iterator BBI = EdgeBB->begin(); 2126 isa<PHINode>(BBI); ++BBI) { 2127 PHINode *PN = cast<PHINode>(BBI); 2128 Value *InVal = PN->getIncomingValueForBlock(*PI); 2129 for (unsigned i = 0, e = Values.size()-1; i != e; ++i) 2130 PN->addIncoming(InVal, *PI); 2131 } 2132 2133 // Erase the old branch instruction. 2134 (*PI)->getInstList().erase(BI); 2135 2136 // Erase the potentially condition tree that was used to computed the 2137 // branch condition. 2138 ErasePossiblyDeadInstructionTree(Cond); 2139 return true; 2140 } 2141 } 2142 2143 return Changed; 2144} 2145