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