IfConversion.cpp revision 36b56886974eae4f9c5ebc96befd3e7bfe5de338
1//===-- IfConversion.cpp - Machine code if conversion pass. ---------------===// 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 machine instruction level if-conversion pass. 11// 12//===----------------------------------------------------------------------===// 13 14#define DEBUG_TYPE "ifcvt" 15#include "llvm/CodeGen/Passes.h" 16#include "BranchFolding.h" 17#include "llvm/ADT/STLExtras.h" 18#include "llvm/ADT/SmallSet.h" 19#include "llvm/ADT/Statistic.h" 20#include "llvm/CodeGen/LivePhysRegs.h" 21#include "llvm/CodeGen/MachineBranchProbabilityInfo.h" 22#include "llvm/CodeGen/MachineFunctionPass.h" 23#include "llvm/CodeGen/MachineInstrBuilder.h" 24#include "llvm/CodeGen/MachineModuleInfo.h" 25#include "llvm/CodeGen/MachineRegisterInfo.h" 26#include "llvm/CodeGen/TargetSchedule.h" 27#include "llvm/MC/MCInstrItineraries.h" 28#include "llvm/Support/CommandLine.h" 29#include "llvm/Support/Debug.h" 30#include "llvm/Support/ErrorHandling.h" 31#include "llvm/Support/raw_ostream.h" 32#include "llvm/Target/TargetInstrInfo.h" 33#include "llvm/Target/TargetLowering.h" 34#include "llvm/Target/TargetMachine.h" 35#include "llvm/Target/TargetRegisterInfo.h" 36#include "llvm/Target/TargetSubtargetInfo.h" 37 38using namespace llvm; 39 40// Hidden options for help debugging. 41static cl::opt<int> IfCvtFnStart("ifcvt-fn-start", cl::init(-1), cl::Hidden); 42static cl::opt<int> IfCvtFnStop("ifcvt-fn-stop", cl::init(-1), cl::Hidden); 43static cl::opt<int> IfCvtLimit("ifcvt-limit", cl::init(-1), cl::Hidden); 44static cl::opt<bool> DisableSimple("disable-ifcvt-simple", 45 cl::init(false), cl::Hidden); 46static cl::opt<bool> DisableSimpleF("disable-ifcvt-simple-false", 47 cl::init(false), cl::Hidden); 48static cl::opt<bool> DisableTriangle("disable-ifcvt-triangle", 49 cl::init(false), cl::Hidden); 50static cl::opt<bool> DisableTriangleR("disable-ifcvt-triangle-rev", 51 cl::init(false), cl::Hidden); 52static cl::opt<bool> DisableTriangleF("disable-ifcvt-triangle-false", 53 cl::init(false), cl::Hidden); 54static cl::opt<bool> DisableTriangleFR("disable-ifcvt-triangle-false-rev", 55 cl::init(false), cl::Hidden); 56static cl::opt<bool> DisableDiamond("disable-ifcvt-diamond", 57 cl::init(false), cl::Hidden); 58static cl::opt<bool> IfCvtBranchFold("ifcvt-branch-fold", 59 cl::init(true), cl::Hidden); 60 61STATISTIC(NumSimple, "Number of simple if-conversions performed"); 62STATISTIC(NumSimpleFalse, "Number of simple (F) if-conversions performed"); 63STATISTIC(NumTriangle, "Number of triangle if-conversions performed"); 64STATISTIC(NumTriangleRev, "Number of triangle (R) if-conversions performed"); 65STATISTIC(NumTriangleFalse,"Number of triangle (F) if-conversions performed"); 66STATISTIC(NumTriangleFRev, "Number of triangle (F/R) if-conversions performed"); 67STATISTIC(NumDiamonds, "Number of diamond if-conversions performed"); 68STATISTIC(NumIfConvBBs, "Number of if-converted blocks"); 69STATISTIC(NumDupBBs, "Number of duplicated blocks"); 70STATISTIC(NumUnpred, "Number of true blocks of diamonds unpredicated"); 71 72namespace { 73 class IfConverter : public MachineFunctionPass { 74 enum IfcvtKind { 75 ICNotClassfied, // BB data valid, but not classified. 76 ICSimpleFalse, // Same as ICSimple, but on the false path. 77 ICSimple, // BB is entry of an one split, no rejoin sub-CFG. 78 ICTriangleFRev, // Same as ICTriangleFalse, but false path rev condition. 79 ICTriangleRev, // Same as ICTriangle, but true path rev condition. 80 ICTriangleFalse, // Same as ICTriangle, but on the false path. 81 ICTriangle, // BB is entry of a triangle sub-CFG. 82 ICDiamond // BB is entry of a diamond sub-CFG. 83 }; 84 85 /// BBInfo - One per MachineBasicBlock, this is used to cache the result 86 /// if-conversion feasibility analysis. This includes results from 87 /// TargetInstrInfo::AnalyzeBranch() (i.e. TBB, FBB, and Cond), and its 88 /// classification, and common tail block of its successors (if it's a 89 /// diamond shape), its size, whether it's predicable, and whether any 90 /// instruction can clobber the 'would-be' predicate. 91 /// 92 /// IsDone - True if BB is not to be considered for ifcvt. 93 /// IsBeingAnalyzed - True if BB is currently being analyzed. 94 /// IsAnalyzed - True if BB has been analyzed (info is still valid). 95 /// IsEnqueued - True if BB has been enqueued to be ifcvt'ed. 96 /// IsBrAnalyzable - True if AnalyzeBranch() returns false. 97 /// HasFallThrough - True if BB may fallthrough to the following BB. 98 /// IsUnpredicable - True if BB is known to be unpredicable. 99 /// ClobbersPred - True if BB could modify predicates (e.g. has 100 /// cmp, call, etc.) 101 /// NonPredSize - Number of non-predicated instructions. 102 /// ExtraCost - Extra cost for multi-cycle instructions. 103 /// ExtraCost2 - Some instructions are slower when predicated 104 /// BB - Corresponding MachineBasicBlock. 105 /// TrueBB / FalseBB- See AnalyzeBranch(). 106 /// BrCond - Conditions for end of block conditional branches. 107 /// Predicate - Predicate used in the BB. 108 struct BBInfo { 109 bool IsDone : 1; 110 bool IsBeingAnalyzed : 1; 111 bool IsAnalyzed : 1; 112 bool IsEnqueued : 1; 113 bool IsBrAnalyzable : 1; 114 bool HasFallThrough : 1; 115 bool IsUnpredicable : 1; 116 bool CannotBeCopied : 1; 117 bool ClobbersPred : 1; 118 unsigned NonPredSize; 119 unsigned ExtraCost; 120 unsigned ExtraCost2; 121 MachineBasicBlock *BB; 122 MachineBasicBlock *TrueBB; 123 MachineBasicBlock *FalseBB; 124 SmallVector<MachineOperand, 4> BrCond; 125 SmallVector<MachineOperand, 4> Predicate; 126 BBInfo() : IsDone(false), IsBeingAnalyzed(false), 127 IsAnalyzed(false), IsEnqueued(false), IsBrAnalyzable(false), 128 HasFallThrough(false), IsUnpredicable(false), 129 CannotBeCopied(false), ClobbersPred(false), NonPredSize(0), 130 ExtraCost(0), ExtraCost2(0), BB(0), TrueBB(0), FalseBB(0) {} 131 }; 132 133 /// IfcvtToken - Record information about pending if-conversions to attempt: 134 /// BBI - Corresponding BBInfo. 135 /// Kind - Type of block. See IfcvtKind. 136 /// NeedSubsumption - True if the to-be-predicated BB has already been 137 /// predicated. 138 /// NumDups - Number of instructions that would be duplicated due 139 /// to this if-conversion. (For diamonds, the number of 140 /// identical instructions at the beginnings of both 141 /// paths). 142 /// NumDups2 - For diamonds, the number of identical instructions 143 /// at the ends of both paths. 144 struct IfcvtToken { 145 BBInfo &BBI; 146 IfcvtKind Kind; 147 bool NeedSubsumption; 148 unsigned NumDups; 149 unsigned NumDups2; 150 IfcvtToken(BBInfo &b, IfcvtKind k, bool s, unsigned d, unsigned d2 = 0) 151 : BBI(b), Kind(k), NeedSubsumption(s), NumDups(d), NumDups2(d2) {} 152 }; 153 154 /// BBAnalysis - Results of if-conversion feasibility analysis indexed by 155 /// basic block number. 156 std::vector<BBInfo> BBAnalysis; 157 TargetSchedModel SchedModel; 158 159 const TargetLoweringBase *TLI; 160 const TargetInstrInfo *TII; 161 const TargetRegisterInfo *TRI; 162 const MachineBranchProbabilityInfo *MBPI; 163 MachineRegisterInfo *MRI; 164 165 LivePhysRegs Redefs; 166 LivePhysRegs DontKill; 167 168 bool PreRegAlloc; 169 bool MadeChange; 170 int FnNum; 171 public: 172 static char ID; 173 IfConverter() : MachineFunctionPass(ID), FnNum(-1) { 174 initializeIfConverterPass(*PassRegistry::getPassRegistry()); 175 } 176 177 void getAnalysisUsage(AnalysisUsage &AU) const override { 178 AU.addRequired<MachineBranchProbabilityInfo>(); 179 MachineFunctionPass::getAnalysisUsage(AU); 180 } 181 182 bool runOnMachineFunction(MachineFunction &MF) override; 183 184 private: 185 bool ReverseBranchCondition(BBInfo &BBI); 186 bool ValidSimple(BBInfo &TrueBBI, unsigned &Dups, 187 const BranchProbability &Prediction) const; 188 bool ValidTriangle(BBInfo &TrueBBI, BBInfo &FalseBBI, 189 bool FalseBranch, unsigned &Dups, 190 const BranchProbability &Prediction) const; 191 bool ValidDiamond(BBInfo &TrueBBI, BBInfo &FalseBBI, 192 unsigned &Dups1, unsigned &Dups2) const; 193 void ScanInstructions(BBInfo &BBI); 194 BBInfo &AnalyzeBlock(MachineBasicBlock *BB, 195 std::vector<IfcvtToken*> &Tokens); 196 bool FeasibilityAnalysis(BBInfo &BBI, SmallVectorImpl<MachineOperand> &Cond, 197 bool isTriangle = false, bool RevBranch = false); 198 void AnalyzeBlocks(MachineFunction &MF, std::vector<IfcvtToken*> &Tokens); 199 void InvalidatePreds(MachineBasicBlock *BB); 200 void RemoveExtraEdges(BBInfo &BBI); 201 bool IfConvertSimple(BBInfo &BBI, IfcvtKind Kind); 202 bool IfConvertTriangle(BBInfo &BBI, IfcvtKind Kind); 203 bool IfConvertDiamond(BBInfo &BBI, IfcvtKind Kind, 204 unsigned NumDups1, unsigned NumDups2); 205 void PredicateBlock(BBInfo &BBI, 206 MachineBasicBlock::iterator E, 207 SmallVectorImpl<MachineOperand> &Cond, 208 SmallSet<unsigned, 4> *LaterRedefs = 0); 209 void CopyAndPredicateBlock(BBInfo &ToBBI, BBInfo &FromBBI, 210 SmallVectorImpl<MachineOperand> &Cond, 211 bool IgnoreBr = false); 212 void MergeBlocks(BBInfo &ToBBI, BBInfo &FromBBI, bool AddEdges = true); 213 214 bool MeetIfcvtSizeLimit(MachineBasicBlock &BB, 215 unsigned Cycle, unsigned Extra, 216 const BranchProbability &Prediction) const { 217 return Cycle > 0 && TII->isProfitableToIfCvt(BB, Cycle, Extra, 218 Prediction); 219 } 220 221 bool MeetIfcvtSizeLimit(MachineBasicBlock &TBB, 222 unsigned TCycle, unsigned TExtra, 223 MachineBasicBlock &FBB, 224 unsigned FCycle, unsigned FExtra, 225 const BranchProbability &Prediction) const { 226 return TCycle > 0 && FCycle > 0 && 227 TII->isProfitableToIfCvt(TBB, TCycle, TExtra, FBB, FCycle, FExtra, 228 Prediction); 229 } 230 231 // blockAlwaysFallThrough - Block ends without a terminator. 232 bool blockAlwaysFallThrough(BBInfo &BBI) const { 233 return BBI.IsBrAnalyzable && BBI.TrueBB == NULL; 234 } 235 236 // IfcvtTokenCmp - Used to sort if-conversion candidates. 237 static bool IfcvtTokenCmp(IfcvtToken *C1, IfcvtToken *C2) { 238 int Incr1 = (C1->Kind == ICDiamond) 239 ? -(int)(C1->NumDups + C1->NumDups2) : (int)C1->NumDups; 240 int Incr2 = (C2->Kind == ICDiamond) 241 ? -(int)(C2->NumDups + C2->NumDups2) : (int)C2->NumDups; 242 if (Incr1 > Incr2) 243 return true; 244 else if (Incr1 == Incr2) { 245 // Favors subsumption. 246 if (C1->NeedSubsumption == false && C2->NeedSubsumption == true) 247 return true; 248 else if (C1->NeedSubsumption == C2->NeedSubsumption) { 249 // Favors diamond over triangle, etc. 250 if ((unsigned)C1->Kind < (unsigned)C2->Kind) 251 return true; 252 else if (C1->Kind == C2->Kind) 253 return C1->BBI.BB->getNumber() < C2->BBI.BB->getNumber(); 254 } 255 } 256 return false; 257 } 258 }; 259 260 char IfConverter::ID = 0; 261} 262 263char &llvm::IfConverterID = IfConverter::ID; 264 265INITIALIZE_PASS_BEGIN(IfConverter, "if-converter", "If Converter", false, false) 266INITIALIZE_PASS_DEPENDENCY(MachineBranchProbabilityInfo) 267INITIALIZE_PASS_END(IfConverter, "if-converter", "If Converter", false, false) 268 269bool IfConverter::runOnMachineFunction(MachineFunction &MF) { 270 TLI = MF.getTarget().getTargetLowering(); 271 TII = MF.getTarget().getInstrInfo(); 272 TRI = MF.getTarget().getRegisterInfo(); 273 MBPI = &getAnalysis<MachineBranchProbabilityInfo>(); 274 MRI = &MF.getRegInfo(); 275 276 const TargetSubtargetInfo &ST = 277 MF.getTarget().getSubtarget<TargetSubtargetInfo>(); 278 SchedModel.init(*ST.getSchedModel(), &ST, TII); 279 280 if (!TII) return false; 281 282 PreRegAlloc = MRI->isSSA(); 283 284 bool BFChange = false; 285 if (!PreRegAlloc) { 286 // Tail merge tend to expose more if-conversion opportunities. 287 BranchFolder BF(true, false); 288 BFChange = BF.OptimizeFunction(MF, TII, 289 MF.getTarget().getRegisterInfo(), 290 getAnalysisIfAvailable<MachineModuleInfo>()); 291 } 292 293 DEBUG(dbgs() << "\nIfcvt: function (" << ++FnNum << ") \'" 294 << MF.getName() << "\'"); 295 296 if (FnNum < IfCvtFnStart || (IfCvtFnStop != -1 && FnNum > IfCvtFnStop)) { 297 DEBUG(dbgs() << " skipped\n"); 298 return false; 299 } 300 DEBUG(dbgs() << "\n"); 301 302 MF.RenumberBlocks(); 303 BBAnalysis.resize(MF.getNumBlockIDs()); 304 305 std::vector<IfcvtToken*> Tokens; 306 MadeChange = false; 307 unsigned NumIfCvts = NumSimple + NumSimpleFalse + NumTriangle + 308 NumTriangleRev + NumTriangleFalse + NumTriangleFRev + NumDiamonds; 309 while (IfCvtLimit == -1 || (int)NumIfCvts < IfCvtLimit) { 310 // Do an initial analysis for each basic block and find all the potential 311 // candidates to perform if-conversion. 312 bool Change = false; 313 AnalyzeBlocks(MF, Tokens); 314 while (!Tokens.empty()) { 315 IfcvtToken *Token = Tokens.back(); 316 Tokens.pop_back(); 317 BBInfo &BBI = Token->BBI; 318 IfcvtKind Kind = Token->Kind; 319 unsigned NumDups = Token->NumDups; 320 unsigned NumDups2 = Token->NumDups2; 321 322 delete Token; 323 324 // If the block has been evicted out of the queue or it has already been 325 // marked dead (due to it being predicated), then skip it. 326 if (BBI.IsDone) 327 BBI.IsEnqueued = false; 328 if (!BBI.IsEnqueued) 329 continue; 330 331 BBI.IsEnqueued = false; 332 333 bool RetVal = false; 334 switch (Kind) { 335 default: llvm_unreachable("Unexpected!"); 336 case ICSimple: 337 case ICSimpleFalse: { 338 bool isFalse = Kind == ICSimpleFalse; 339 if ((isFalse && DisableSimpleF) || (!isFalse && DisableSimple)) break; 340 DEBUG(dbgs() << "Ifcvt (Simple" << (Kind == ICSimpleFalse ? 341 " false" : "") 342 << "): BB#" << BBI.BB->getNumber() << " (" 343 << ((Kind == ICSimpleFalse) 344 ? BBI.FalseBB->getNumber() 345 : BBI.TrueBB->getNumber()) << ") "); 346 RetVal = IfConvertSimple(BBI, Kind); 347 DEBUG(dbgs() << (RetVal ? "succeeded!" : "failed!") << "\n"); 348 if (RetVal) { 349 if (isFalse) ++NumSimpleFalse; 350 else ++NumSimple; 351 } 352 break; 353 } 354 case ICTriangle: 355 case ICTriangleRev: 356 case ICTriangleFalse: 357 case ICTriangleFRev: { 358 bool isFalse = Kind == ICTriangleFalse; 359 bool isRev = (Kind == ICTriangleRev || Kind == ICTriangleFRev); 360 if (DisableTriangle && !isFalse && !isRev) break; 361 if (DisableTriangleR && !isFalse && isRev) break; 362 if (DisableTriangleF && isFalse && !isRev) break; 363 if (DisableTriangleFR && isFalse && isRev) break; 364 DEBUG(dbgs() << "Ifcvt (Triangle"); 365 if (isFalse) 366 DEBUG(dbgs() << " false"); 367 if (isRev) 368 DEBUG(dbgs() << " rev"); 369 DEBUG(dbgs() << "): BB#" << BBI.BB->getNumber() << " (T:" 370 << BBI.TrueBB->getNumber() << ",F:" 371 << BBI.FalseBB->getNumber() << ") "); 372 RetVal = IfConvertTriangle(BBI, Kind); 373 DEBUG(dbgs() << (RetVal ? "succeeded!" : "failed!") << "\n"); 374 if (RetVal) { 375 if (isFalse) { 376 if (isRev) ++NumTriangleFRev; 377 else ++NumTriangleFalse; 378 } else { 379 if (isRev) ++NumTriangleRev; 380 else ++NumTriangle; 381 } 382 } 383 break; 384 } 385 case ICDiamond: { 386 if (DisableDiamond) break; 387 DEBUG(dbgs() << "Ifcvt (Diamond): BB#" << BBI.BB->getNumber() << " (T:" 388 << BBI.TrueBB->getNumber() << ",F:" 389 << BBI.FalseBB->getNumber() << ") "); 390 RetVal = IfConvertDiamond(BBI, Kind, NumDups, NumDups2); 391 DEBUG(dbgs() << (RetVal ? "succeeded!" : "failed!") << "\n"); 392 if (RetVal) ++NumDiamonds; 393 break; 394 } 395 } 396 397 Change |= RetVal; 398 399 NumIfCvts = NumSimple + NumSimpleFalse + NumTriangle + NumTriangleRev + 400 NumTriangleFalse + NumTriangleFRev + NumDiamonds; 401 if (IfCvtLimit != -1 && (int)NumIfCvts >= IfCvtLimit) 402 break; 403 } 404 405 if (!Change) 406 break; 407 MadeChange |= Change; 408 } 409 410 // Delete tokens in case of early exit. 411 while (!Tokens.empty()) { 412 IfcvtToken *Token = Tokens.back(); 413 Tokens.pop_back(); 414 delete Token; 415 } 416 417 Tokens.clear(); 418 BBAnalysis.clear(); 419 420 if (MadeChange && IfCvtBranchFold) { 421 BranchFolder BF(false, false); 422 BF.OptimizeFunction(MF, TII, 423 MF.getTarget().getRegisterInfo(), 424 getAnalysisIfAvailable<MachineModuleInfo>()); 425 } 426 427 MadeChange |= BFChange; 428 return MadeChange; 429} 430 431/// findFalseBlock - BB has a fallthrough. Find its 'false' successor given 432/// its 'true' successor. 433static MachineBasicBlock *findFalseBlock(MachineBasicBlock *BB, 434 MachineBasicBlock *TrueBB) { 435 for (MachineBasicBlock::succ_iterator SI = BB->succ_begin(), 436 E = BB->succ_end(); SI != E; ++SI) { 437 MachineBasicBlock *SuccBB = *SI; 438 if (SuccBB != TrueBB) 439 return SuccBB; 440 } 441 return NULL; 442} 443 444/// ReverseBranchCondition - Reverse the condition of the end of the block 445/// branch. Swap block's 'true' and 'false' successors. 446bool IfConverter::ReverseBranchCondition(BBInfo &BBI) { 447 DebugLoc dl; // FIXME: this is nowhere 448 if (!TII->ReverseBranchCondition(BBI.BrCond)) { 449 TII->RemoveBranch(*BBI.BB); 450 TII->InsertBranch(*BBI.BB, BBI.FalseBB, BBI.TrueBB, BBI.BrCond, dl); 451 std::swap(BBI.TrueBB, BBI.FalseBB); 452 return true; 453 } 454 return false; 455} 456 457/// getNextBlock - Returns the next block in the function blocks ordering. If 458/// it is the end, returns NULL. 459static inline MachineBasicBlock *getNextBlock(MachineBasicBlock *BB) { 460 MachineFunction::iterator I = BB; 461 MachineFunction::iterator E = BB->getParent()->end(); 462 if (++I == E) 463 return NULL; 464 return I; 465} 466 467/// ValidSimple - Returns true if the 'true' block (along with its 468/// predecessor) forms a valid simple shape for ifcvt. It also returns the 469/// number of instructions that the ifcvt would need to duplicate if performed 470/// in Dups. 471bool IfConverter::ValidSimple(BBInfo &TrueBBI, unsigned &Dups, 472 const BranchProbability &Prediction) const { 473 Dups = 0; 474 if (TrueBBI.IsBeingAnalyzed || TrueBBI.IsDone) 475 return false; 476 477 if (TrueBBI.IsBrAnalyzable) 478 return false; 479 480 if (TrueBBI.BB->pred_size() > 1) { 481 if (TrueBBI.CannotBeCopied || 482 !TII->isProfitableToDupForIfCvt(*TrueBBI.BB, TrueBBI.NonPredSize, 483 Prediction)) 484 return false; 485 Dups = TrueBBI.NonPredSize; 486 } 487 488 return true; 489} 490 491/// ValidTriangle - Returns true if the 'true' and 'false' blocks (along 492/// with their common predecessor) forms a valid triangle shape for ifcvt. 493/// If 'FalseBranch' is true, it checks if 'true' block's false branch 494/// branches to the 'false' block rather than the other way around. It also 495/// returns the number of instructions that the ifcvt would need to duplicate 496/// if performed in 'Dups'. 497bool IfConverter::ValidTriangle(BBInfo &TrueBBI, BBInfo &FalseBBI, 498 bool FalseBranch, unsigned &Dups, 499 const BranchProbability &Prediction) const { 500 Dups = 0; 501 if (TrueBBI.IsBeingAnalyzed || TrueBBI.IsDone) 502 return false; 503 504 if (TrueBBI.BB->pred_size() > 1) { 505 if (TrueBBI.CannotBeCopied) 506 return false; 507 508 unsigned Size = TrueBBI.NonPredSize; 509 if (TrueBBI.IsBrAnalyzable) { 510 if (TrueBBI.TrueBB && TrueBBI.BrCond.empty()) 511 // Ends with an unconditional branch. It will be removed. 512 --Size; 513 else { 514 MachineBasicBlock *FExit = FalseBranch 515 ? TrueBBI.TrueBB : TrueBBI.FalseBB; 516 if (FExit) 517 // Require a conditional branch 518 ++Size; 519 } 520 } 521 if (!TII->isProfitableToDupForIfCvt(*TrueBBI.BB, Size, Prediction)) 522 return false; 523 Dups = Size; 524 } 525 526 MachineBasicBlock *TExit = FalseBranch ? TrueBBI.FalseBB : TrueBBI.TrueBB; 527 if (!TExit && blockAlwaysFallThrough(TrueBBI)) { 528 MachineFunction::iterator I = TrueBBI.BB; 529 if (++I == TrueBBI.BB->getParent()->end()) 530 return false; 531 TExit = I; 532 } 533 return TExit && TExit == FalseBBI.BB; 534} 535 536/// ValidDiamond - Returns true if the 'true' and 'false' blocks (along 537/// with their common predecessor) forms a valid diamond shape for ifcvt. 538bool IfConverter::ValidDiamond(BBInfo &TrueBBI, BBInfo &FalseBBI, 539 unsigned &Dups1, unsigned &Dups2) const { 540 Dups1 = Dups2 = 0; 541 if (TrueBBI.IsBeingAnalyzed || TrueBBI.IsDone || 542 FalseBBI.IsBeingAnalyzed || FalseBBI.IsDone) 543 return false; 544 545 MachineBasicBlock *TT = TrueBBI.TrueBB; 546 MachineBasicBlock *FT = FalseBBI.TrueBB; 547 548 if (!TT && blockAlwaysFallThrough(TrueBBI)) 549 TT = getNextBlock(TrueBBI.BB); 550 if (!FT && blockAlwaysFallThrough(FalseBBI)) 551 FT = getNextBlock(FalseBBI.BB); 552 if (TT != FT) 553 return false; 554 if (TT == NULL && (TrueBBI.IsBrAnalyzable || FalseBBI.IsBrAnalyzable)) 555 return false; 556 if (TrueBBI.BB->pred_size() > 1 || FalseBBI.BB->pred_size() > 1) 557 return false; 558 559 // FIXME: Allow true block to have an early exit? 560 if (TrueBBI.FalseBB || FalseBBI.FalseBB || 561 (TrueBBI.ClobbersPred && FalseBBI.ClobbersPred)) 562 return false; 563 564 // Count duplicate instructions at the beginning of the true and false blocks. 565 MachineBasicBlock::iterator TIB = TrueBBI.BB->begin(); 566 MachineBasicBlock::iterator FIB = FalseBBI.BB->begin(); 567 MachineBasicBlock::iterator TIE = TrueBBI.BB->end(); 568 MachineBasicBlock::iterator FIE = FalseBBI.BB->end(); 569 while (TIB != TIE && FIB != FIE) { 570 // Skip dbg_value instructions. These do not count. 571 if (TIB->isDebugValue()) { 572 while (TIB != TIE && TIB->isDebugValue()) 573 ++TIB; 574 if (TIB == TIE) 575 break; 576 } 577 if (FIB->isDebugValue()) { 578 while (FIB != FIE && FIB->isDebugValue()) 579 ++FIB; 580 if (FIB == FIE) 581 break; 582 } 583 if (!TIB->isIdenticalTo(FIB)) 584 break; 585 ++Dups1; 586 ++TIB; 587 ++FIB; 588 } 589 590 // Now, in preparation for counting duplicate instructions at the ends of the 591 // blocks, move the end iterators up past any branch instructions. 592 while (TIE != TIB) { 593 --TIE; 594 if (!TIE->isBranch()) 595 break; 596 } 597 while (FIE != FIB) { 598 --FIE; 599 if (!FIE->isBranch()) 600 break; 601 } 602 603 // If Dups1 includes all of a block, then don't count duplicate 604 // instructions at the end of the blocks. 605 if (TIB == TIE || FIB == FIE) 606 return true; 607 608 // Count duplicate instructions at the ends of the blocks. 609 while (TIE != TIB && FIE != FIB) { 610 // Skip dbg_value instructions. These do not count. 611 if (TIE->isDebugValue()) { 612 while (TIE != TIB && TIE->isDebugValue()) 613 --TIE; 614 if (TIE == TIB) 615 break; 616 } 617 if (FIE->isDebugValue()) { 618 while (FIE != FIB && FIE->isDebugValue()) 619 --FIE; 620 if (FIE == FIB) 621 break; 622 } 623 if (!TIE->isIdenticalTo(FIE)) 624 break; 625 ++Dups2; 626 --TIE; 627 --FIE; 628 } 629 630 return true; 631} 632 633/// ScanInstructions - Scan all the instructions in the block to determine if 634/// the block is predicable. In most cases, that means all the instructions 635/// in the block are isPredicable(). Also checks if the block contains any 636/// instruction which can clobber a predicate (e.g. condition code register). 637/// If so, the block is not predicable unless it's the last instruction. 638void IfConverter::ScanInstructions(BBInfo &BBI) { 639 if (BBI.IsDone) 640 return; 641 642 bool AlreadyPredicated = !BBI.Predicate.empty(); 643 // First analyze the end of BB branches. 644 BBI.TrueBB = BBI.FalseBB = NULL; 645 BBI.BrCond.clear(); 646 BBI.IsBrAnalyzable = 647 !TII->AnalyzeBranch(*BBI.BB, BBI.TrueBB, BBI.FalseBB, BBI.BrCond); 648 BBI.HasFallThrough = BBI.IsBrAnalyzable && BBI.FalseBB == NULL; 649 650 if (BBI.BrCond.size()) { 651 // No false branch. This BB must end with a conditional branch and a 652 // fallthrough. 653 if (!BBI.FalseBB) 654 BBI.FalseBB = findFalseBlock(BBI.BB, BBI.TrueBB); 655 if (!BBI.FalseBB) { 656 // Malformed bcc? True and false blocks are the same? 657 BBI.IsUnpredicable = true; 658 return; 659 } 660 } 661 662 // Then scan all the instructions. 663 BBI.NonPredSize = 0; 664 BBI.ExtraCost = 0; 665 BBI.ExtraCost2 = 0; 666 BBI.ClobbersPred = false; 667 for (MachineBasicBlock::iterator I = BBI.BB->begin(), E = BBI.BB->end(); 668 I != E; ++I) { 669 if (I->isDebugValue()) 670 continue; 671 672 if (I->isNotDuplicable()) 673 BBI.CannotBeCopied = true; 674 675 bool isPredicated = TII->isPredicated(I); 676 bool isCondBr = BBI.IsBrAnalyzable && I->isConditionalBranch(); 677 678 // A conditional branch is not predicable, but it may be eliminated. 679 if (isCondBr) 680 continue; 681 682 if (!isPredicated) { 683 BBI.NonPredSize++; 684 unsigned ExtraPredCost = TII->getPredicationCost(&*I); 685 unsigned NumCycles = SchedModel.computeInstrLatency(&*I, false); 686 if (NumCycles > 1) 687 BBI.ExtraCost += NumCycles-1; 688 BBI.ExtraCost2 += ExtraPredCost; 689 } else if (!AlreadyPredicated) { 690 // FIXME: This instruction is already predicated before the 691 // if-conversion pass. It's probably something like a conditional move. 692 // Mark this block unpredicable for now. 693 BBI.IsUnpredicable = true; 694 return; 695 } 696 697 if (BBI.ClobbersPred && !isPredicated) { 698 // Predicate modification instruction should end the block (except for 699 // already predicated instructions and end of block branches). 700 // Predicate may have been modified, the subsequent (currently) 701 // unpredicated instructions cannot be correctly predicated. 702 BBI.IsUnpredicable = true; 703 return; 704 } 705 706 // FIXME: Make use of PredDefs? e.g. ADDC, SUBC sets predicates but are 707 // still potentially predicable. 708 std::vector<MachineOperand> PredDefs; 709 if (TII->DefinesPredicate(I, PredDefs)) 710 BBI.ClobbersPred = true; 711 712 if (!TII->isPredicable(I)) { 713 BBI.IsUnpredicable = true; 714 return; 715 } 716 } 717} 718 719/// FeasibilityAnalysis - Determine if the block is a suitable candidate to be 720/// predicated by the specified predicate. 721bool IfConverter::FeasibilityAnalysis(BBInfo &BBI, 722 SmallVectorImpl<MachineOperand> &Pred, 723 bool isTriangle, bool RevBranch) { 724 // If the block is dead or unpredicable, then it cannot be predicated. 725 if (BBI.IsDone || BBI.IsUnpredicable) 726 return false; 727 728 // If it is already predicated, check if the new predicate subsumes 729 // its predicate. 730 if (BBI.Predicate.size() && !TII->SubsumesPredicate(Pred, BBI.Predicate)) 731 return false; 732 733 if (BBI.BrCond.size()) { 734 if (!isTriangle) 735 return false; 736 737 // Test predicate subsumption. 738 SmallVector<MachineOperand, 4> RevPred(Pred.begin(), Pred.end()); 739 SmallVector<MachineOperand, 4> Cond(BBI.BrCond.begin(), BBI.BrCond.end()); 740 if (RevBranch) { 741 if (TII->ReverseBranchCondition(Cond)) 742 return false; 743 } 744 if (TII->ReverseBranchCondition(RevPred) || 745 !TII->SubsumesPredicate(Cond, RevPred)) 746 return false; 747 } 748 749 return true; 750} 751 752/// AnalyzeBlock - Analyze the structure of the sub-CFG starting from 753/// the specified block. Record its successors and whether it looks like an 754/// if-conversion candidate. 755IfConverter::BBInfo &IfConverter::AnalyzeBlock(MachineBasicBlock *BB, 756 std::vector<IfcvtToken*> &Tokens) { 757 BBInfo &BBI = BBAnalysis[BB->getNumber()]; 758 759 if (BBI.IsAnalyzed || BBI.IsBeingAnalyzed) 760 return BBI; 761 762 BBI.BB = BB; 763 BBI.IsBeingAnalyzed = true; 764 765 ScanInstructions(BBI); 766 767 // Unanalyzable or ends with fallthrough or unconditional branch, or if is not 768 // considered for ifcvt anymore. 769 if (!BBI.IsBrAnalyzable || BBI.BrCond.empty() || BBI.IsDone) { 770 BBI.IsBeingAnalyzed = false; 771 BBI.IsAnalyzed = true; 772 return BBI; 773 } 774 775 // Do not ifcvt if either path is a back edge to the entry block. 776 if (BBI.TrueBB == BB || BBI.FalseBB == BB) { 777 BBI.IsBeingAnalyzed = false; 778 BBI.IsAnalyzed = true; 779 return BBI; 780 } 781 782 // Do not ifcvt if true and false fallthrough blocks are the same. 783 if (!BBI.FalseBB) { 784 BBI.IsBeingAnalyzed = false; 785 BBI.IsAnalyzed = true; 786 return BBI; 787 } 788 789 BBInfo &TrueBBI = AnalyzeBlock(BBI.TrueBB, Tokens); 790 BBInfo &FalseBBI = AnalyzeBlock(BBI.FalseBB, Tokens); 791 792 if (TrueBBI.IsDone && FalseBBI.IsDone) { 793 BBI.IsBeingAnalyzed = false; 794 BBI.IsAnalyzed = true; 795 return BBI; 796 } 797 798 SmallVector<MachineOperand, 4> RevCond(BBI.BrCond.begin(), BBI.BrCond.end()); 799 bool CanRevCond = !TII->ReverseBranchCondition(RevCond); 800 801 unsigned Dups = 0; 802 unsigned Dups2 = 0; 803 bool TNeedSub = !TrueBBI.Predicate.empty(); 804 bool FNeedSub = !FalseBBI.Predicate.empty(); 805 bool Enqueued = false; 806 807 BranchProbability Prediction = MBPI->getEdgeProbability(BB, TrueBBI.BB); 808 809 if (CanRevCond && ValidDiamond(TrueBBI, FalseBBI, Dups, Dups2) && 810 MeetIfcvtSizeLimit(*TrueBBI.BB, (TrueBBI.NonPredSize - (Dups + Dups2) + 811 TrueBBI.ExtraCost), TrueBBI.ExtraCost2, 812 *FalseBBI.BB, (FalseBBI.NonPredSize - (Dups + Dups2) + 813 FalseBBI.ExtraCost),FalseBBI.ExtraCost2, 814 Prediction) && 815 FeasibilityAnalysis(TrueBBI, BBI.BrCond) && 816 FeasibilityAnalysis(FalseBBI, RevCond)) { 817 // Diamond: 818 // EBB 819 // / \_ 820 // | | 821 // TBB FBB 822 // \ / 823 // TailBB 824 // Note TailBB can be empty. 825 Tokens.push_back(new IfcvtToken(BBI, ICDiamond, TNeedSub|FNeedSub, Dups, 826 Dups2)); 827 Enqueued = true; 828 } 829 830 if (ValidTriangle(TrueBBI, FalseBBI, false, Dups, Prediction) && 831 MeetIfcvtSizeLimit(*TrueBBI.BB, TrueBBI.NonPredSize + TrueBBI.ExtraCost, 832 TrueBBI.ExtraCost2, Prediction) && 833 FeasibilityAnalysis(TrueBBI, BBI.BrCond, true)) { 834 // Triangle: 835 // EBB 836 // | \_ 837 // | | 838 // | TBB 839 // | / 840 // FBB 841 Tokens.push_back(new IfcvtToken(BBI, ICTriangle, TNeedSub, Dups)); 842 Enqueued = true; 843 } 844 845 if (ValidTriangle(TrueBBI, FalseBBI, true, Dups, Prediction) && 846 MeetIfcvtSizeLimit(*TrueBBI.BB, TrueBBI.NonPredSize + TrueBBI.ExtraCost, 847 TrueBBI.ExtraCost2, Prediction) && 848 FeasibilityAnalysis(TrueBBI, BBI.BrCond, true, true)) { 849 Tokens.push_back(new IfcvtToken(BBI, ICTriangleRev, TNeedSub, Dups)); 850 Enqueued = true; 851 } 852 853 if (ValidSimple(TrueBBI, Dups, Prediction) && 854 MeetIfcvtSizeLimit(*TrueBBI.BB, TrueBBI.NonPredSize + TrueBBI.ExtraCost, 855 TrueBBI.ExtraCost2, Prediction) && 856 FeasibilityAnalysis(TrueBBI, BBI.BrCond)) { 857 // Simple (split, no rejoin): 858 // EBB 859 // | \_ 860 // | | 861 // | TBB---> exit 862 // | 863 // FBB 864 Tokens.push_back(new IfcvtToken(BBI, ICSimple, TNeedSub, Dups)); 865 Enqueued = true; 866 } 867 868 if (CanRevCond) { 869 // Try the other path... 870 if (ValidTriangle(FalseBBI, TrueBBI, false, Dups, 871 Prediction.getCompl()) && 872 MeetIfcvtSizeLimit(*FalseBBI.BB, 873 FalseBBI.NonPredSize + FalseBBI.ExtraCost, 874 FalseBBI.ExtraCost2, Prediction.getCompl()) && 875 FeasibilityAnalysis(FalseBBI, RevCond, true)) { 876 Tokens.push_back(new IfcvtToken(BBI, ICTriangleFalse, FNeedSub, Dups)); 877 Enqueued = true; 878 } 879 880 if (ValidTriangle(FalseBBI, TrueBBI, true, Dups, 881 Prediction.getCompl()) && 882 MeetIfcvtSizeLimit(*FalseBBI.BB, 883 FalseBBI.NonPredSize + FalseBBI.ExtraCost, 884 FalseBBI.ExtraCost2, Prediction.getCompl()) && 885 FeasibilityAnalysis(FalseBBI, RevCond, true, true)) { 886 Tokens.push_back(new IfcvtToken(BBI, ICTriangleFRev, FNeedSub, Dups)); 887 Enqueued = true; 888 } 889 890 if (ValidSimple(FalseBBI, Dups, Prediction.getCompl()) && 891 MeetIfcvtSizeLimit(*FalseBBI.BB, 892 FalseBBI.NonPredSize + FalseBBI.ExtraCost, 893 FalseBBI.ExtraCost2, Prediction.getCompl()) && 894 FeasibilityAnalysis(FalseBBI, RevCond)) { 895 Tokens.push_back(new IfcvtToken(BBI, ICSimpleFalse, FNeedSub, Dups)); 896 Enqueued = true; 897 } 898 } 899 900 BBI.IsEnqueued = Enqueued; 901 BBI.IsBeingAnalyzed = false; 902 BBI.IsAnalyzed = true; 903 return BBI; 904} 905 906/// AnalyzeBlocks - Analyze all blocks and find entries for all if-conversion 907/// candidates. 908void IfConverter::AnalyzeBlocks(MachineFunction &MF, 909 std::vector<IfcvtToken*> &Tokens) { 910 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; ++I) { 911 MachineBasicBlock *BB = I; 912 AnalyzeBlock(BB, Tokens); 913 } 914 915 // Sort to favor more complex ifcvt scheme. 916 std::stable_sort(Tokens.begin(), Tokens.end(), IfcvtTokenCmp); 917} 918 919/// canFallThroughTo - Returns true either if ToBB is the next block after BB or 920/// that all the intervening blocks are empty (given BB can fall through to its 921/// next block). 922static bool canFallThroughTo(MachineBasicBlock *BB, MachineBasicBlock *ToBB) { 923 MachineFunction::iterator PI = BB; 924 MachineFunction::iterator I = std::next(PI); 925 MachineFunction::iterator TI = ToBB; 926 MachineFunction::iterator E = BB->getParent()->end(); 927 while (I != TI) { 928 // Check isSuccessor to avoid case where the next block is empty, but 929 // it's not a successor. 930 if (I == E || !I->empty() || !PI->isSuccessor(I)) 931 return false; 932 PI = I++; 933 } 934 return true; 935} 936 937/// InvalidatePreds - Invalidate predecessor BB info so it would be re-analyzed 938/// to determine if it can be if-converted. If predecessor is already enqueued, 939/// dequeue it! 940void IfConverter::InvalidatePreds(MachineBasicBlock *BB) { 941 for (MachineBasicBlock::pred_iterator PI = BB->pred_begin(), 942 E = BB->pred_end(); PI != E; ++PI) { 943 BBInfo &PBBI = BBAnalysis[(*PI)->getNumber()]; 944 if (PBBI.IsDone || PBBI.BB == BB) 945 continue; 946 PBBI.IsAnalyzed = false; 947 PBBI.IsEnqueued = false; 948 } 949} 950 951/// InsertUncondBranch - Inserts an unconditional branch from BB to ToBB. 952/// 953static void InsertUncondBranch(MachineBasicBlock *BB, MachineBasicBlock *ToBB, 954 const TargetInstrInfo *TII) { 955 DebugLoc dl; // FIXME: this is nowhere 956 SmallVector<MachineOperand, 0> NoCond; 957 TII->InsertBranch(*BB, ToBB, NULL, NoCond, dl); 958} 959 960/// RemoveExtraEdges - Remove true / false edges if either / both are no longer 961/// successors. 962void IfConverter::RemoveExtraEdges(BBInfo &BBI) { 963 MachineBasicBlock *TBB = NULL, *FBB = NULL; 964 SmallVector<MachineOperand, 4> Cond; 965 if (!TII->AnalyzeBranch(*BBI.BB, TBB, FBB, Cond)) 966 BBI.BB->CorrectExtraCFGEdges(TBB, FBB, !Cond.empty()); 967} 968 969/// Behaves like LiveRegUnits::StepForward() but also adds implicit uses to all 970/// values defined in MI which are not live/used by MI. 971static void UpdatePredRedefs(MachineInstr *MI, LivePhysRegs &Redefs) { 972 for (ConstMIBundleOperands Ops(MI); Ops.isValid(); ++Ops) { 973 if (!Ops->isReg() || !Ops->isKill()) 974 continue; 975 unsigned Reg = Ops->getReg(); 976 if (Reg == 0) 977 continue; 978 Redefs.removeReg(Reg); 979 } 980 for (MIBundleOperands Ops(MI); Ops.isValid(); ++Ops) { 981 if (!Ops->isReg() || !Ops->isDef()) 982 continue; 983 unsigned Reg = Ops->getReg(); 984 if (Reg == 0 || Redefs.contains(Reg)) 985 continue; 986 Redefs.addReg(Reg); 987 988 MachineOperand &Op = *Ops; 989 MachineInstr *MI = Op.getParent(); 990 MachineInstrBuilder MIB(*MI->getParent()->getParent(), MI); 991 MIB.addReg(Reg, RegState::Implicit | RegState::Undef); 992 } 993} 994 995/** 996 * Remove kill flags from operands with a registers in the @p DontKill set. 997 */ 998static void RemoveKills(MachineInstr &MI, const LivePhysRegs &DontKill) { 999 for (MIBundleOperands O(&MI); O.isValid(); ++O) { 1000 if (!O->isReg() || !O->isKill()) 1001 continue; 1002 if (DontKill.contains(O->getReg())) 1003 O->setIsKill(false); 1004 } 1005} 1006 1007/** 1008 * Walks a range of machine instructions and removes kill flags for registers 1009 * in the @p DontKill set. 1010 */ 1011static void RemoveKills(MachineBasicBlock::iterator I, 1012 MachineBasicBlock::iterator E, 1013 const LivePhysRegs &DontKill, 1014 const MCRegisterInfo &MCRI) { 1015 for ( ; I != E; ++I) 1016 RemoveKills(*I, DontKill); 1017} 1018 1019/// IfConvertSimple - If convert a simple (split, no rejoin) sub-CFG. 1020/// 1021bool IfConverter::IfConvertSimple(BBInfo &BBI, IfcvtKind Kind) { 1022 BBInfo &TrueBBI = BBAnalysis[BBI.TrueBB->getNumber()]; 1023 BBInfo &FalseBBI = BBAnalysis[BBI.FalseBB->getNumber()]; 1024 BBInfo *CvtBBI = &TrueBBI; 1025 BBInfo *NextBBI = &FalseBBI; 1026 1027 SmallVector<MachineOperand, 4> Cond(BBI.BrCond.begin(), BBI.BrCond.end()); 1028 if (Kind == ICSimpleFalse) 1029 std::swap(CvtBBI, NextBBI); 1030 1031 if (CvtBBI->IsDone || 1032 (CvtBBI->CannotBeCopied && CvtBBI->BB->pred_size() > 1)) { 1033 // Something has changed. It's no longer safe to predicate this block. 1034 BBI.IsAnalyzed = false; 1035 CvtBBI->IsAnalyzed = false; 1036 return false; 1037 } 1038 1039 if (CvtBBI->BB->hasAddressTaken()) 1040 // Conservatively abort if-conversion if BB's address is taken. 1041 return false; 1042 1043 if (Kind == ICSimpleFalse) 1044 if (TII->ReverseBranchCondition(Cond)) 1045 llvm_unreachable("Unable to reverse branch condition!"); 1046 1047 // Initialize liveins to the first BB. These are potentiall redefined by 1048 // predicated instructions. 1049 Redefs.init(TRI); 1050 Redefs.addLiveIns(CvtBBI->BB); 1051 Redefs.addLiveIns(NextBBI->BB); 1052 1053 // Compute a set of registers which must not be killed by instructions in 1054 // BB1: This is everything live-in to BB2. 1055 DontKill.init(TRI); 1056 DontKill.addLiveIns(NextBBI->BB); 1057 1058 if (CvtBBI->BB->pred_size() > 1) { 1059 BBI.NonPredSize -= TII->RemoveBranch(*BBI.BB); 1060 // Copy instructions in the true block, predicate them, and add them to 1061 // the entry block. 1062 CopyAndPredicateBlock(BBI, *CvtBBI, Cond); 1063 1064 // RemoveExtraEdges won't work if the block has an unanalyzable branch, so 1065 // explicitly remove CvtBBI as a successor. 1066 BBI.BB->removeSuccessor(CvtBBI->BB); 1067 } else { 1068 RemoveKills(CvtBBI->BB->begin(), CvtBBI->BB->end(), DontKill, *TRI); 1069 PredicateBlock(*CvtBBI, CvtBBI->BB->end(), Cond); 1070 1071 // Merge converted block into entry block. 1072 BBI.NonPredSize -= TII->RemoveBranch(*BBI.BB); 1073 MergeBlocks(BBI, *CvtBBI); 1074 } 1075 1076 bool IterIfcvt = true; 1077 if (!canFallThroughTo(BBI.BB, NextBBI->BB)) { 1078 InsertUncondBranch(BBI.BB, NextBBI->BB, TII); 1079 BBI.HasFallThrough = false; 1080 // Now ifcvt'd block will look like this: 1081 // BB: 1082 // ... 1083 // t, f = cmp 1084 // if t op 1085 // b BBf 1086 // 1087 // We cannot further ifcvt this block because the unconditional branch 1088 // will have to be predicated on the new condition, that will not be 1089 // available if cmp executes. 1090 IterIfcvt = false; 1091 } 1092 1093 RemoveExtraEdges(BBI); 1094 1095 // Update block info. BB can be iteratively if-converted. 1096 if (!IterIfcvt) 1097 BBI.IsDone = true; 1098 InvalidatePreds(BBI.BB); 1099 CvtBBI->IsDone = true; 1100 1101 // FIXME: Must maintain LiveIns. 1102 return true; 1103} 1104 1105/// Scale down weights to fit into uint32_t. NewTrue is the new weight 1106/// for successor TrueBB, and NewFalse is the new weight for successor 1107/// FalseBB. 1108static void ScaleWeights(uint64_t NewTrue, uint64_t NewFalse, 1109 MachineBasicBlock *MBB, 1110 const MachineBasicBlock *TrueBB, 1111 const MachineBasicBlock *FalseBB, 1112 const MachineBranchProbabilityInfo *MBPI) { 1113 uint64_t NewMax = (NewTrue > NewFalse) ? NewTrue : NewFalse; 1114 uint32_t Scale = (NewMax / UINT32_MAX) + 1; 1115 for (MachineBasicBlock::succ_iterator SI = MBB->succ_begin(), 1116 SE = MBB->succ_end(); 1117 SI != SE; ++SI) { 1118 if (*SI == TrueBB) 1119 MBB->setSuccWeight(SI, (uint32_t)(NewTrue / Scale)); 1120 else if (*SI == FalseBB) 1121 MBB->setSuccWeight(SI, (uint32_t)(NewFalse / Scale)); 1122 else 1123 MBB->setSuccWeight(SI, MBPI->getEdgeWeight(MBB, SI) / Scale); 1124 } 1125} 1126 1127/// IfConvertTriangle - If convert a triangle sub-CFG. 1128/// 1129bool IfConverter::IfConvertTriangle(BBInfo &BBI, IfcvtKind Kind) { 1130 BBInfo &TrueBBI = BBAnalysis[BBI.TrueBB->getNumber()]; 1131 BBInfo &FalseBBI = BBAnalysis[BBI.FalseBB->getNumber()]; 1132 BBInfo *CvtBBI = &TrueBBI; 1133 BBInfo *NextBBI = &FalseBBI; 1134 DebugLoc dl; // FIXME: this is nowhere 1135 1136 SmallVector<MachineOperand, 4> Cond(BBI.BrCond.begin(), BBI.BrCond.end()); 1137 if (Kind == ICTriangleFalse || Kind == ICTriangleFRev) 1138 std::swap(CvtBBI, NextBBI); 1139 1140 if (CvtBBI->IsDone || 1141 (CvtBBI->CannotBeCopied && CvtBBI->BB->pred_size() > 1)) { 1142 // Something has changed. It's no longer safe to predicate this block. 1143 BBI.IsAnalyzed = false; 1144 CvtBBI->IsAnalyzed = false; 1145 return false; 1146 } 1147 1148 if (CvtBBI->BB->hasAddressTaken()) 1149 // Conservatively abort if-conversion if BB's address is taken. 1150 return false; 1151 1152 if (Kind == ICTriangleFalse || Kind == ICTriangleFRev) 1153 if (TII->ReverseBranchCondition(Cond)) 1154 llvm_unreachable("Unable to reverse branch condition!"); 1155 1156 if (Kind == ICTriangleRev || Kind == ICTriangleFRev) { 1157 if (ReverseBranchCondition(*CvtBBI)) { 1158 // BB has been changed, modify its predecessors (except for this 1159 // one) so they don't get ifcvt'ed based on bad intel. 1160 for (MachineBasicBlock::pred_iterator PI = CvtBBI->BB->pred_begin(), 1161 E = CvtBBI->BB->pred_end(); PI != E; ++PI) { 1162 MachineBasicBlock *PBB = *PI; 1163 if (PBB == BBI.BB) 1164 continue; 1165 BBInfo &PBBI = BBAnalysis[PBB->getNumber()]; 1166 if (PBBI.IsEnqueued) { 1167 PBBI.IsAnalyzed = false; 1168 PBBI.IsEnqueued = false; 1169 } 1170 } 1171 } 1172 } 1173 1174 // Initialize liveins to the first BB. These are potentially redefined by 1175 // predicated instructions. 1176 Redefs.init(TRI); 1177 Redefs.addLiveIns(CvtBBI->BB); 1178 Redefs.addLiveIns(NextBBI->BB); 1179 1180 DontKill.clear(); 1181 1182 bool HasEarlyExit = CvtBBI->FalseBB != NULL; 1183 uint64_t CvtNext = 0, CvtFalse = 0, BBNext = 0, BBCvt = 0, SumWeight = 0; 1184 uint32_t WeightScale = 0; 1185 if (HasEarlyExit) { 1186 // Get weights before modifying CvtBBI->BB and BBI.BB. 1187 CvtNext = MBPI->getEdgeWeight(CvtBBI->BB, NextBBI->BB); 1188 CvtFalse = MBPI->getEdgeWeight(CvtBBI->BB, CvtBBI->FalseBB); 1189 BBNext = MBPI->getEdgeWeight(BBI.BB, NextBBI->BB); 1190 BBCvt = MBPI->getEdgeWeight(BBI.BB, CvtBBI->BB); 1191 SumWeight = MBPI->getSumForBlock(CvtBBI->BB, WeightScale); 1192 } 1193 if (CvtBBI->BB->pred_size() > 1) { 1194 BBI.NonPredSize -= TII->RemoveBranch(*BBI.BB); 1195 // Copy instructions in the true block, predicate them, and add them to 1196 // the entry block. 1197 CopyAndPredicateBlock(BBI, *CvtBBI, Cond, true); 1198 1199 // RemoveExtraEdges won't work if the block has an unanalyzable branch, so 1200 // explicitly remove CvtBBI as a successor. 1201 BBI.BB->removeSuccessor(CvtBBI->BB); 1202 } else { 1203 // Predicate the 'true' block after removing its branch. 1204 CvtBBI->NonPredSize -= TII->RemoveBranch(*CvtBBI->BB); 1205 PredicateBlock(*CvtBBI, CvtBBI->BB->end(), Cond); 1206 1207 // Now merge the entry of the triangle with the true block. 1208 BBI.NonPredSize -= TII->RemoveBranch(*BBI.BB); 1209 MergeBlocks(BBI, *CvtBBI, false); 1210 } 1211 1212 // If 'true' block has a 'false' successor, add an exit branch to it. 1213 if (HasEarlyExit) { 1214 SmallVector<MachineOperand, 4> RevCond(CvtBBI->BrCond.begin(), 1215 CvtBBI->BrCond.end()); 1216 if (TII->ReverseBranchCondition(RevCond)) 1217 llvm_unreachable("Unable to reverse branch condition!"); 1218 TII->InsertBranch(*BBI.BB, CvtBBI->FalseBB, NULL, RevCond, dl); 1219 BBI.BB->addSuccessor(CvtBBI->FalseBB); 1220 // Update the edge weight for both CvtBBI->FalseBB and NextBBI. 1221 // New_Weight(BBI.BB, NextBBI->BB) = 1222 // Weight(BBI.BB, NextBBI->BB) * getSumForBlock(CvtBBI->BB) + 1223 // Weight(BBI.BB, CvtBBI->BB) * Weight(CvtBBI->BB, NextBBI->BB) 1224 // New_Weight(BBI.BB, CvtBBI->FalseBB) = 1225 // Weight(BBI.BB, CvtBBI->BB) * Weight(CvtBBI->BB, CvtBBI->FalseBB) 1226 1227 uint64_t NewNext = BBNext * SumWeight + (BBCvt * CvtNext) / WeightScale; 1228 uint64_t NewFalse = (BBCvt * CvtFalse) / WeightScale; 1229 // We need to scale down all weights of BBI.BB to fit uint32_t. 1230 // Here BBI.BB is connected to CvtBBI->FalseBB and will fall through to 1231 // the next block. 1232 ScaleWeights(NewNext, NewFalse, BBI.BB, getNextBlock(BBI.BB), 1233 CvtBBI->FalseBB, MBPI); 1234 } 1235 1236 // Merge in the 'false' block if the 'false' block has no other 1237 // predecessors. Otherwise, add an unconditional branch to 'false'. 1238 bool FalseBBDead = false; 1239 bool IterIfcvt = true; 1240 bool isFallThrough = canFallThroughTo(BBI.BB, NextBBI->BB); 1241 if (!isFallThrough) { 1242 // Only merge them if the true block does not fallthrough to the false 1243 // block. By not merging them, we make it possible to iteratively 1244 // ifcvt the blocks. 1245 if (!HasEarlyExit && 1246 NextBBI->BB->pred_size() == 1 && !NextBBI->HasFallThrough && 1247 !NextBBI->BB->hasAddressTaken()) { 1248 MergeBlocks(BBI, *NextBBI); 1249 FalseBBDead = true; 1250 } else { 1251 InsertUncondBranch(BBI.BB, NextBBI->BB, TII); 1252 BBI.HasFallThrough = false; 1253 } 1254 // Mixed predicated and unpredicated code. This cannot be iteratively 1255 // predicated. 1256 IterIfcvt = false; 1257 } 1258 1259 RemoveExtraEdges(BBI); 1260 1261 // Update block info. BB can be iteratively if-converted. 1262 if (!IterIfcvt) 1263 BBI.IsDone = true; 1264 InvalidatePreds(BBI.BB); 1265 CvtBBI->IsDone = true; 1266 if (FalseBBDead) 1267 NextBBI->IsDone = true; 1268 1269 // FIXME: Must maintain LiveIns. 1270 return true; 1271} 1272 1273/// IfConvertDiamond - If convert a diamond sub-CFG. 1274/// 1275bool IfConverter::IfConvertDiamond(BBInfo &BBI, IfcvtKind Kind, 1276 unsigned NumDups1, unsigned NumDups2) { 1277 BBInfo &TrueBBI = BBAnalysis[BBI.TrueBB->getNumber()]; 1278 BBInfo &FalseBBI = BBAnalysis[BBI.FalseBB->getNumber()]; 1279 MachineBasicBlock *TailBB = TrueBBI.TrueBB; 1280 // True block must fall through or end with an unanalyzable terminator. 1281 if (!TailBB) { 1282 if (blockAlwaysFallThrough(TrueBBI)) 1283 TailBB = FalseBBI.TrueBB; 1284 assert((TailBB || !TrueBBI.IsBrAnalyzable) && "Unexpected!"); 1285 } 1286 1287 if (TrueBBI.IsDone || FalseBBI.IsDone || 1288 TrueBBI.BB->pred_size() > 1 || 1289 FalseBBI.BB->pred_size() > 1) { 1290 // Something has changed. It's no longer safe to predicate these blocks. 1291 BBI.IsAnalyzed = false; 1292 TrueBBI.IsAnalyzed = false; 1293 FalseBBI.IsAnalyzed = false; 1294 return false; 1295 } 1296 1297 if (TrueBBI.BB->hasAddressTaken() || FalseBBI.BB->hasAddressTaken()) 1298 // Conservatively abort if-conversion if either BB has its address taken. 1299 return false; 1300 1301 // Put the predicated instructions from the 'true' block before the 1302 // instructions from the 'false' block, unless the true block would clobber 1303 // the predicate, in which case, do the opposite. 1304 BBInfo *BBI1 = &TrueBBI; 1305 BBInfo *BBI2 = &FalseBBI; 1306 SmallVector<MachineOperand, 4> RevCond(BBI.BrCond.begin(), BBI.BrCond.end()); 1307 if (TII->ReverseBranchCondition(RevCond)) 1308 llvm_unreachable("Unable to reverse branch condition!"); 1309 SmallVector<MachineOperand, 4> *Cond1 = &BBI.BrCond; 1310 SmallVector<MachineOperand, 4> *Cond2 = &RevCond; 1311 1312 // Figure out the more profitable ordering. 1313 bool DoSwap = false; 1314 if (TrueBBI.ClobbersPred && !FalseBBI.ClobbersPred) 1315 DoSwap = true; 1316 else if (TrueBBI.ClobbersPred == FalseBBI.ClobbersPred) { 1317 if (TrueBBI.NonPredSize > FalseBBI.NonPredSize) 1318 DoSwap = true; 1319 } 1320 if (DoSwap) { 1321 std::swap(BBI1, BBI2); 1322 std::swap(Cond1, Cond2); 1323 } 1324 1325 // Remove the conditional branch from entry to the blocks. 1326 BBI.NonPredSize -= TII->RemoveBranch(*BBI.BB); 1327 1328 // Initialize liveins to the first BB. These are potentially redefined by 1329 // predicated instructions. 1330 Redefs.init(TRI); 1331 Redefs.addLiveIns(BBI1->BB); 1332 1333 // Remove the duplicated instructions at the beginnings of both paths. 1334 MachineBasicBlock::iterator DI1 = BBI1->BB->begin(); 1335 MachineBasicBlock::iterator DI2 = BBI2->BB->begin(); 1336 MachineBasicBlock::iterator DIE1 = BBI1->BB->end(); 1337 MachineBasicBlock::iterator DIE2 = BBI2->BB->end(); 1338 // Skip dbg_value instructions 1339 while (DI1 != DIE1 && DI1->isDebugValue()) 1340 ++DI1; 1341 while (DI2 != DIE2 && DI2->isDebugValue()) 1342 ++DI2; 1343 BBI1->NonPredSize -= NumDups1; 1344 BBI2->NonPredSize -= NumDups1; 1345 1346 // Skip past the dups on each side separately since there may be 1347 // differing dbg_value entries. 1348 for (unsigned i = 0; i < NumDups1; ++DI1) { 1349 if (!DI1->isDebugValue()) 1350 ++i; 1351 } 1352 while (NumDups1 != 0) { 1353 ++DI2; 1354 if (!DI2->isDebugValue()) 1355 --NumDups1; 1356 } 1357 1358 // Compute a set of registers which must not be killed by instructions in BB1: 1359 // This is everything used+live in BB2 after the duplicated instructions. We 1360 // can compute this set by simulating liveness backwards from the end of BB2. 1361 DontKill.init(TRI); 1362 for (MachineBasicBlock::reverse_iterator I = BBI2->BB->rbegin(), 1363 E = MachineBasicBlock::reverse_iterator(DI2); I != E; ++I) { 1364 DontKill.stepBackward(*I); 1365 } 1366 1367 for (MachineBasicBlock::const_iterator I = BBI1->BB->begin(), E = DI1; I != E; 1368 ++I) { 1369 Redefs.stepForward(*I); 1370 } 1371 BBI.BB->splice(BBI.BB->end(), BBI1->BB, BBI1->BB->begin(), DI1); 1372 BBI2->BB->erase(BBI2->BB->begin(), DI2); 1373 1374 // Remove branch from 'true' block and remove duplicated instructions. 1375 BBI1->NonPredSize -= TII->RemoveBranch(*BBI1->BB); 1376 DI1 = BBI1->BB->end(); 1377 for (unsigned i = 0; i != NumDups2; ) { 1378 // NumDups2 only counted non-dbg_value instructions, so this won't 1379 // run off the head of the list. 1380 assert (DI1 != BBI1->BB->begin()); 1381 --DI1; 1382 // skip dbg_value instructions 1383 if (!DI1->isDebugValue()) 1384 ++i; 1385 } 1386 BBI1->BB->erase(DI1, BBI1->BB->end()); 1387 1388 // Kill flags in the true block for registers living into the false block 1389 // must be removed. 1390 RemoveKills(BBI1->BB->begin(), BBI1->BB->end(), DontKill, *TRI); 1391 1392 // Remove 'false' block branch and find the last instruction to predicate. 1393 BBI2->NonPredSize -= TII->RemoveBranch(*BBI2->BB); 1394 DI2 = BBI2->BB->end(); 1395 while (NumDups2 != 0) { 1396 // NumDups2 only counted non-dbg_value instructions, so this won't 1397 // run off the head of the list. 1398 assert (DI2 != BBI2->BB->begin()); 1399 --DI2; 1400 // skip dbg_value instructions 1401 if (!DI2->isDebugValue()) 1402 --NumDups2; 1403 } 1404 1405 // Remember which registers would later be defined by the false block. 1406 // This allows us not to predicate instructions in the true block that would 1407 // later be re-defined. That is, rather than 1408 // subeq r0, r1, #1 1409 // addne r0, r1, #1 1410 // generate: 1411 // sub r0, r1, #1 1412 // addne r0, r1, #1 1413 SmallSet<unsigned, 4> RedefsByFalse; 1414 SmallSet<unsigned, 4> ExtUses; 1415 if (TII->isProfitableToUnpredicate(*BBI1->BB, *BBI2->BB)) { 1416 for (MachineBasicBlock::iterator FI = BBI2->BB->begin(); FI != DI2; ++FI) { 1417 if (FI->isDebugValue()) 1418 continue; 1419 SmallVector<unsigned, 4> Defs; 1420 for (unsigned i = 0, e = FI->getNumOperands(); i != e; ++i) { 1421 const MachineOperand &MO = FI->getOperand(i); 1422 if (!MO.isReg()) 1423 continue; 1424 unsigned Reg = MO.getReg(); 1425 if (!Reg) 1426 continue; 1427 if (MO.isDef()) { 1428 Defs.push_back(Reg); 1429 } else if (!RedefsByFalse.count(Reg)) { 1430 // These are defined before ctrl flow reach the 'false' instructions. 1431 // They cannot be modified by the 'true' instructions. 1432 for (MCSubRegIterator SubRegs(Reg, TRI, /*IncludeSelf=*/true); 1433 SubRegs.isValid(); ++SubRegs) 1434 ExtUses.insert(*SubRegs); 1435 } 1436 } 1437 1438 for (unsigned i = 0, e = Defs.size(); i != e; ++i) { 1439 unsigned Reg = Defs[i]; 1440 if (!ExtUses.count(Reg)) { 1441 for (MCSubRegIterator SubRegs(Reg, TRI, /*IncludeSelf=*/true); 1442 SubRegs.isValid(); ++SubRegs) 1443 RedefsByFalse.insert(*SubRegs); 1444 } 1445 } 1446 } 1447 } 1448 1449 // Predicate the 'true' block. 1450 PredicateBlock(*BBI1, BBI1->BB->end(), *Cond1, &RedefsByFalse); 1451 1452 // Predicate the 'false' block. 1453 PredicateBlock(*BBI2, DI2, *Cond2); 1454 1455 // Merge the true block into the entry of the diamond. 1456 MergeBlocks(BBI, *BBI1, TailBB == 0); 1457 MergeBlocks(BBI, *BBI2, TailBB == 0); 1458 1459 // If the if-converted block falls through or unconditionally branches into 1460 // the tail block, and the tail block does not have other predecessors, then 1461 // fold the tail block in as well. Otherwise, unless it falls through to the 1462 // tail, add a unconditional branch to it. 1463 if (TailBB) { 1464 BBInfo &TailBBI = BBAnalysis[TailBB->getNumber()]; 1465 bool CanMergeTail = !TailBBI.HasFallThrough && 1466 !TailBBI.BB->hasAddressTaken(); 1467 // There may still be a fall-through edge from BBI1 or BBI2 to TailBB; 1468 // check if there are any other predecessors besides those. 1469 unsigned NumPreds = TailBB->pred_size(); 1470 if (NumPreds > 1) 1471 CanMergeTail = false; 1472 else if (NumPreds == 1 && CanMergeTail) { 1473 MachineBasicBlock::pred_iterator PI = TailBB->pred_begin(); 1474 if (*PI != BBI1->BB && *PI != BBI2->BB) 1475 CanMergeTail = false; 1476 } 1477 if (CanMergeTail) { 1478 MergeBlocks(BBI, TailBBI); 1479 TailBBI.IsDone = true; 1480 } else { 1481 BBI.BB->addSuccessor(TailBB); 1482 InsertUncondBranch(BBI.BB, TailBB, TII); 1483 BBI.HasFallThrough = false; 1484 } 1485 } 1486 1487 // RemoveExtraEdges won't work if the block has an unanalyzable branch, 1488 // which can happen here if TailBB is unanalyzable and is merged, so 1489 // explicitly remove BBI1 and BBI2 as successors. 1490 BBI.BB->removeSuccessor(BBI1->BB); 1491 BBI.BB->removeSuccessor(BBI2->BB); 1492 RemoveExtraEdges(BBI); 1493 1494 // Update block info. 1495 BBI.IsDone = TrueBBI.IsDone = FalseBBI.IsDone = true; 1496 InvalidatePreds(BBI.BB); 1497 1498 // FIXME: Must maintain LiveIns. 1499 return true; 1500} 1501 1502static bool MaySpeculate(const MachineInstr *MI, 1503 SmallSet<unsigned, 4> &LaterRedefs, 1504 const TargetInstrInfo *TII) { 1505 bool SawStore = true; 1506 if (!MI->isSafeToMove(TII, 0, SawStore)) 1507 return false; 1508 1509 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) { 1510 const MachineOperand &MO = MI->getOperand(i); 1511 if (!MO.isReg()) 1512 continue; 1513 unsigned Reg = MO.getReg(); 1514 if (!Reg) 1515 continue; 1516 if (MO.isDef() && !LaterRedefs.count(Reg)) 1517 return false; 1518 } 1519 1520 return true; 1521} 1522 1523/// PredicateBlock - Predicate instructions from the start of the block to the 1524/// specified end with the specified condition. 1525void IfConverter::PredicateBlock(BBInfo &BBI, 1526 MachineBasicBlock::iterator E, 1527 SmallVectorImpl<MachineOperand> &Cond, 1528 SmallSet<unsigned, 4> *LaterRedefs) { 1529 bool AnyUnpred = false; 1530 bool MaySpec = LaterRedefs != 0; 1531 for (MachineBasicBlock::iterator I = BBI.BB->begin(); I != E; ++I) { 1532 if (I->isDebugValue() || TII->isPredicated(I)) 1533 continue; 1534 // It may be possible not to predicate an instruction if it's the 'true' 1535 // side of a diamond and the 'false' side may re-define the instruction's 1536 // defs. 1537 if (MaySpec && MaySpeculate(I, *LaterRedefs, TII)) { 1538 AnyUnpred = true; 1539 continue; 1540 } 1541 // If any instruction is predicated, then every instruction after it must 1542 // be predicated. 1543 MaySpec = false; 1544 if (!TII->PredicateInstruction(I, Cond)) { 1545#ifndef NDEBUG 1546 dbgs() << "Unable to predicate " << *I << "!\n"; 1547#endif 1548 llvm_unreachable(0); 1549 } 1550 1551 // If the predicated instruction now redefines a register as the result of 1552 // if-conversion, add an implicit kill. 1553 UpdatePredRedefs(I, Redefs); 1554 } 1555 1556 std::copy(Cond.begin(), Cond.end(), std::back_inserter(BBI.Predicate)); 1557 1558 BBI.IsAnalyzed = false; 1559 BBI.NonPredSize = 0; 1560 1561 ++NumIfConvBBs; 1562 if (AnyUnpred) 1563 ++NumUnpred; 1564} 1565 1566/// CopyAndPredicateBlock - Copy and predicate instructions from source BB to 1567/// the destination block. Skip end of block branches if IgnoreBr is true. 1568void IfConverter::CopyAndPredicateBlock(BBInfo &ToBBI, BBInfo &FromBBI, 1569 SmallVectorImpl<MachineOperand> &Cond, 1570 bool IgnoreBr) { 1571 MachineFunction &MF = *ToBBI.BB->getParent(); 1572 1573 for (MachineBasicBlock::iterator I = FromBBI.BB->begin(), 1574 E = FromBBI.BB->end(); I != E; ++I) { 1575 // Do not copy the end of the block branches. 1576 if (IgnoreBr && I->isBranch()) 1577 break; 1578 1579 MachineInstr *MI = MF.CloneMachineInstr(I); 1580 ToBBI.BB->insert(ToBBI.BB->end(), MI); 1581 ToBBI.NonPredSize++; 1582 unsigned ExtraPredCost = TII->getPredicationCost(&*I); 1583 unsigned NumCycles = SchedModel.computeInstrLatency(&*I, false); 1584 if (NumCycles > 1) 1585 ToBBI.ExtraCost += NumCycles-1; 1586 ToBBI.ExtraCost2 += ExtraPredCost; 1587 1588 if (!TII->isPredicated(I) && !MI->isDebugValue()) { 1589 if (!TII->PredicateInstruction(MI, Cond)) { 1590#ifndef NDEBUG 1591 dbgs() << "Unable to predicate " << *I << "!\n"; 1592#endif 1593 llvm_unreachable(0); 1594 } 1595 } 1596 1597 // If the predicated instruction now redefines a register as the result of 1598 // if-conversion, add an implicit kill. 1599 UpdatePredRedefs(MI, Redefs); 1600 1601 // Some kill flags may not be correct anymore. 1602 if (!DontKill.empty()) 1603 RemoveKills(*MI, DontKill); 1604 } 1605 1606 if (!IgnoreBr) { 1607 std::vector<MachineBasicBlock *> Succs(FromBBI.BB->succ_begin(), 1608 FromBBI.BB->succ_end()); 1609 MachineBasicBlock *NBB = getNextBlock(FromBBI.BB); 1610 MachineBasicBlock *FallThrough = FromBBI.HasFallThrough ? NBB : NULL; 1611 1612 for (unsigned i = 0, e = Succs.size(); i != e; ++i) { 1613 MachineBasicBlock *Succ = Succs[i]; 1614 // Fallthrough edge can't be transferred. 1615 if (Succ == FallThrough) 1616 continue; 1617 ToBBI.BB->addSuccessor(Succ); 1618 } 1619 } 1620 1621 std::copy(FromBBI.Predicate.begin(), FromBBI.Predicate.end(), 1622 std::back_inserter(ToBBI.Predicate)); 1623 std::copy(Cond.begin(), Cond.end(), std::back_inserter(ToBBI.Predicate)); 1624 1625 ToBBI.ClobbersPred |= FromBBI.ClobbersPred; 1626 ToBBI.IsAnalyzed = false; 1627 1628 ++NumDupBBs; 1629} 1630 1631/// MergeBlocks - Move all instructions from FromBB to the end of ToBB. 1632/// This will leave FromBB as an empty block, so remove all of its 1633/// successor edges except for the fall-through edge. If AddEdges is true, 1634/// i.e., when FromBBI's branch is being moved, add those successor edges to 1635/// ToBBI. 1636void IfConverter::MergeBlocks(BBInfo &ToBBI, BBInfo &FromBBI, bool AddEdges) { 1637 assert(!FromBBI.BB->hasAddressTaken() && 1638 "Removing a BB whose address is taken!"); 1639 1640 ToBBI.BB->splice(ToBBI.BB->end(), 1641 FromBBI.BB, FromBBI.BB->begin(), FromBBI.BB->end()); 1642 1643 std::vector<MachineBasicBlock *> Succs(FromBBI.BB->succ_begin(), 1644 FromBBI.BB->succ_end()); 1645 MachineBasicBlock *NBB = getNextBlock(FromBBI.BB); 1646 MachineBasicBlock *FallThrough = FromBBI.HasFallThrough ? NBB : NULL; 1647 1648 for (unsigned i = 0, e = Succs.size(); i != e; ++i) { 1649 MachineBasicBlock *Succ = Succs[i]; 1650 // Fallthrough edge can't be transferred. 1651 if (Succ == FallThrough) 1652 continue; 1653 FromBBI.BB->removeSuccessor(Succ); 1654 if (AddEdges && !ToBBI.BB->isSuccessor(Succ)) 1655 ToBBI.BB->addSuccessor(Succ); 1656 } 1657 1658 // Now FromBBI always falls through to the next block! 1659 if (NBB && !FromBBI.BB->isSuccessor(NBB)) 1660 FromBBI.BB->addSuccessor(NBB); 1661 1662 std::copy(FromBBI.Predicate.begin(), FromBBI.Predicate.end(), 1663 std::back_inserter(ToBBI.Predicate)); 1664 FromBBI.Predicate.clear(); 1665 1666 ToBBI.NonPredSize += FromBBI.NonPredSize; 1667 ToBBI.ExtraCost += FromBBI.ExtraCost; 1668 ToBBI.ExtraCost2 += FromBBI.ExtraCost2; 1669 FromBBI.NonPredSize = 0; 1670 FromBBI.ExtraCost = 0; 1671 FromBBI.ExtraCost2 = 0; 1672 1673 ToBBI.ClobbersPred |= FromBBI.ClobbersPred; 1674 ToBBI.HasFallThrough = FromBBI.HasFallThrough; 1675 ToBBI.IsAnalyzed = false; 1676 FromBBI.IsAnalyzed = false; 1677} 1678