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 "BranchFolding.h" 16 #include "llvm/Function.h" 17 #include "llvm/CodeGen/Passes.h" 18 #include "llvm/CodeGen/MachineModuleInfo.h" 19 #include "llvm/CodeGen/MachineBranchProbabilityInfo.h" 20 #include "llvm/CodeGen/MachineFunctionPass.h" 21 #include "llvm/MC/MCInstrItineraries.h" 22 #include "llvm/Target/TargetInstrInfo.h" 23 #include "llvm/Target/TargetLowering.h" 24 #include "llvm/Target/TargetMachine.h" 25 #include "llvm/Target/TargetRegisterInfo.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/ADT/SmallSet.h" 31 #include "llvm/ADT/Statistic.h" 32 #include "llvm/ADT/STLExtras.h" 33 using namespace llvm; 34 35 // Hidden options for help debugging. 36 static cl::opt<int> IfCvtFnStart("ifcvt-fn-start", cl::init(-1), cl::Hidden); 37 static cl::opt<int> IfCvtFnStop("ifcvt-fn-stop", cl::init(-1), cl::Hidden); 38 static cl::opt<int> IfCvtLimit("ifcvt-limit", cl::init(-1), cl::Hidden); 39 static cl::opt<bool> DisableSimple("disable-ifcvt-simple", 40 cl::init(false), cl::Hidden); 41 static cl::opt<bool> DisableSimpleF("disable-ifcvt-simple-false", 42 cl::init(false), cl::Hidden); 43 static cl::opt<bool> DisableTriangle("disable-ifcvt-triangle", 44 cl::init(false), cl::Hidden); 45 static cl::opt<bool> DisableTriangleR("disable-ifcvt-triangle-rev", 46 cl::init(false), cl::Hidden); 47 static cl::opt<bool> DisableTriangleF("disable-ifcvt-triangle-false", 48 cl::init(false), cl::Hidden); 49 static cl::opt<bool> DisableTriangleFR("disable-ifcvt-triangle-false-rev", 50 cl::init(false), cl::Hidden); 51 static cl::opt<bool> DisableDiamond("disable-ifcvt-diamond", 52 cl::init(false), cl::Hidden); 53 static cl::opt<bool> IfCvtBranchFold("ifcvt-branch-fold", 54 cl::init(true), cl::Hidden); 55 56 STATISTIC(NumSimple, "Number of simple if-conversions performed"); 57 STATISTIC(NumSimpleFalse, "Number of simple (F) if-conversions performed"); 58 STATISTIC(NumTriangle, "Number of triangle if-conversions performed"); 59 STATISTIC(NumTriangleRev, "Number of triangle (R) if-conversions performed"); 60 STATISTIC(NumTriangleFalse,"Number of triangle (F) if-conversions performed"); 61 STATISTIC(NumTriangleFRev, "Number of triangle (F/R) if-conversions performed"); 62 STATISTIC(NumDiamonds, "Number of diamond if-conversions performed"); 63 STATISTIC(NumIfConvBBs, "Number of if-converted blocks"); 64 STATISTIC(NumDupBBs, "Number of duplicated blocks"); 65 STATISTIC(NumUnpred, "Number of true blocks of diamonds unpredicated"); 66 67 namespace { 68 class IfConverter : public MachineFunctionPass { 69 enum IfcvtKind { 70 ICNotClassfied, // BB data valid, but not classified. 71 ICSimpleFalse, // Same as ICSimple, but on the false path. 72 ICSimple, // BB is entry of an one split, no rejoin sub-CFG. 73 ICTriangleFRev, // Same as ICTriangleFalse, but false path rev condition. 74 ICTriangleRev, // Same as ICTriangle, but true path rev condition. 75 ICTriangleFalse, // Same as ICTriangle, but on the false path. 76 ICTriangle, // BB is entry of a triangle sub-CFG. 77 ICDiamond // BB is entry of a diamond sub-CFG. 78 }; 79 80 /// BBInfo - One per MachineBasicBlock, this is used to cache the result 81 /// if-conversion feasibility analysis. This includes results from 82 /// TargetInstrInfo::AnalyzeBranch() (i.e. TBB, FBB, and Cond), and its 83 /// classification, and common tail block of its successors (if it's a 84 /// diamond shape), its size, whether it's predicable, and whether any 85 /// instruction can clobber the 'would-be' predicate. 86 /// 87 /// IsDone - True if BB is not to be considered for ifcvt. 88 /// IsBeingAnalyzed - True if BB is currently being analyzed. 89 /// IsAnalyzed - True if BB has been analyzed (info is still valid). 90 /// IsEnqueued - True if BB has been enqueued to be ifcvt'ed. 91 /// IsBrAnalyzable - True if AnalyzeBranch() returns false. 92 /// HasFallThrough - True if BB may fallthrough to the following BB. 93 /// IsUnpredicable - True if BB is known to be unpredicable. 94 /// ClobbersPred - True if BB could modify predicates (e.g. has 95 /// cmp, call, etc.) 96 /// NonPredSize - Number of non-predicated instructions. 97 /// ExtraCost - Extra cost for multi-cycle instructions. 98 /// ExtraCost2 - Some instructions are slower when predicated 99 /// BB - Corresponding MachineBasicBlock. 100 /// TrueBB / FalseBB- See AnalyzeBranch(). 101 /// BrCond - Conditions for end of block conditional branches. 102 /// Predicate - Predicate used in the BB. 103 struct BBInfo { 104 bool IsDone : 1; 105 bool IsBeingAnalyzed : 1; 106 bool IsAnalyzed : 1; 107 bool IsEnqueued : 1; 108 bool IsBrAnalyzable : 1; 109 bool HasFallThrough : 1; 110 bool IsUnpredicable : 1; 111 bool CannotBeCopied : 1; 112 bool ClobbersPred : 1; 113 unsigned NonPredSize; 114 unsigned ExtraCost; 115 unsigned ExtraCost2; 116 MachineBasicBlock *BB; 117 MachineBasicBlock *TrueBB; 118 MachineBasicBlock *FalseBB; 119 SmallVector<MachineOperand, 4> BrCond; 120 SmallVector<MachineOperand, 4> Predicate; 121 BBInfo() : IsDone(false), IsBeingAnalyzed(false), 122 IsAnalyzed(false), IsEnqueued(false), IsBrAnalyzable(false), 123 HasFallThrough(false), IsUnpredicable(false), 124 CannotBeCopied(false), ClobbersPred(false), NonPredSize(0), 125 ExtraCost(0), ExtraCost2(0), BB(0), TrueBB(0), FalseBB(0) {} 126 }; 127 128 /// IfcvtToken - Record information about pending if-conversions to attempt: 129 /// BBI - Corresponding BBInfo. 130 /// Kind - Type of block. See IfcvtKind. 131 /// NeedSubsumption - True if the to-be-predicated BB has already been 132 /// predicated. 133 /// NumDups - Number of instructions that would be duplicated due 134 /// to this if-conversion. (For diamonds, the number of 135 /// identical instructions at the beginnings of both 136 /// paths). 137 /// NumDups2 - For diamonds, the number of identical instructions 138 /// at the ends of both paths. 139 struct IfcvtToken { 140 BBInfo &BBI; 141 IfcvtKind Kind; 142 bool NeedSubsumption; 143 unsigned NumDups; 144 unsigned NumDups2; 145 IfcvtToken(BBInfo &b, IfcvtKind k, bool s, unsigned d, unsigned d2 = 0) 146 : BBI(b), Kind(k), NeedSubsumption(s), NumDups(d), NumDups2(d2) {} 147 }; 148 149 /// BBAnalysis - Results of if-conversion feasibility analysis indexed by 150 /// basic block number. 151 std::vector<BBInfo> BBAnalysis; 152 153 const TargetLowering *TLI; 154 const TargetInstrInfo *TII; 155 const TargetRegisterInfo *TRI; 156 const InstrItineraryData *InstrItins; 157 const MachineBranchProbabilityInfo *MBPI; 158 159 bool MadeChange; 160 int FnNum; 161 public: 162 static char ID; 163 IfConverter() : MachineFunctionPass(ID), FnNum(-1) { 164 initializeIfConverterPass(*PassRegistry::getPassRegistry()); 165 } 166 167 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 168 AU.addRequired<MachineBranchProbabilityInfo>(); 169 MachineFunctionPass::getAnalysisUsage(AU); 170 } 171 172 virtual bool runOnMachineFunction(MachineFunction &MF); 173 174 private: 175 bool ReverseBranchCondition(BBInfo &BBI); 176 bool ValidSimple(BBInfo &TrueBBI, unsigned &Dups, 177 const BranchProbability &Prediction) const; 178 bool ValidTriangle(BBInfo &TrueBBI, BBInfo &FalseBBI, 179 bool FalseBranch, unsigned &Dups, 180 const BranchProbability &Prediction) const; 181 bool ValidDiamond(BBInfo &TrueBBI, BBInfo &FalseBBI, 182 unsigned &Dups1, unsigned &Dups2) const; 183 void ScanInstructions(BBInfo &BBI); 184 BBInfo &AnalyzeBlock(MachineBasicBlock *BB, 185 std::vector<IfcvtToken*> &Tokens); 186 bool FeasibilityAnalysis(BBInfo &BBI, SmallVectorImpl<MachineOperand> &Cond, 187 bool isTriangle = false, bool RevBranch = false); 188 void AnalyzeBlocks(MachineFunction &MF, std::vector<IfcvtToken*> &Tokens); 189 void InvalidatePreds(MachineBasicBlock *BB); 190 void RemoveExtraEdges(BBInfo &BBI); 191 bool IfConvertSimple(BBInfo &BBI, IfcvtKind Kind); 192 bool IfConvertTriangle(BBInfo &BBI, IfcvtKind Kind); 193 bool IfConvertDiamond(BBInfo &BBI, IfcvtKind Kind, 194 unsigned NumDups1, unsigned NumDups2); 195 void PredicateBlock(BBInfo &BBI, 196 MachineBasicBlock::iterator E, 197 SmallVectorImpl<MachineOperand> &Cond, 198 SmallSet<unsigned, 4> &Redefs, 199 SmallSet<unsigned, 4> *LaterRedefs = 0); 200 void CopyAndPredicateBlock(BBInfo &ToBBI, BBInfo &FromBBI, 201 SmallVectorImpl<MachineOperand> &Cond, 202 SmallSet<unsigned, 4> &Redefs, 203 bool IgnoreBr = false); 204 void MergeBlocks(BBInfo &ToBBI, BBInfo &FromBBI, bool AddEdges = true); 205 206 bool MeetIfcvtSizeLimit(MachineBasicBlock &BB, 207 unsigned Cycle, unsigned Extra, 208 const BranchProbability &Prediction) const { 209 return Cycle > 0 && TII->isProfitableToIfCvt(BB, Cycle, Extra, 210 Prediction); 211 } 212 213 bool MeetIfcvtSizeLimit(MachineBasicBlock &TBB, 214 unsigned TCycle, unsigned TExtra, 215 MachineBasicBlock &FBB, 216 unsigned FCycle, unsigned FExtra, 217 const BranchProbability &Prediction) const { 218 return TCycle > 0 && FCycle > 0 && 219 TII->isProfitableToIfCvt(TBB, TCycle, TExtra, FBB, FCycle, FExtra, 220 Prediction); 221 } 222 223 // blockAlwaysFallThrough - Block ends without a terminator. 224 bool blockAlwaysFallThrough(BBInfo &BBI) const { 225 return BBI.IsBrAnalyzable && BBI.TrueBB == NULL; 226 } 227 228 // IfcvtTokenCmp - Used to sort if-conversion candidates. 229 static bool IfcvtTokenCmp(IfcvtToken *C1, IfcvtToken *C2) { 230 int Incr1 = (C1->Kind == ICDiamond) 231 ? -(int)(C1->NumDups + C1->NumDups2) : (int)C1->NumDups; 232 int Incr2 = (C2->Kind == ICDiamond) 233 ? -(int)(C2->NumDups + C2->NumDups2) : (int)C2->NumDups; 234 if (Incr1 > Incr2) 235 return true; 236 else if (Incr1 == Incr2) { 237 // Favors subsumption. 238 if (C1->NeedSubsumption == false && C2->NeedSubsumption == true) 239 return true; 240 else if (C1->NeedSubsumption == C2->NeedSubsumption) { 241 // Favors diamond over triangle, etc. 242 if ((unsigned)C1->Kind < (unsigned)C2->Kind) 243 return true; 244 else if (C1->Kind == C2->Kind) 245 return C1->BBI.BB->getNumber() < C2->BBI.BB->getNumber(); 246 } 247 } 248 return false; 249 } 250 }; 251 252 char IfConverter::ID = 0; 253 } 254 255 char &llvm::IfConverterID = IfConverter::ID; 256 257 INITIALIZE_PASS_BEGIN(IfConverter, "if-converter", "If Converter", false, false) 258 INITIALIZE_PASS_DEPENDENCY(MachineBranchProbabilityInfo) 259 INITIALIZE_PASS_END(IfConverter, "if-converter", "If Converter", false, false) 260 261 bool IfConverter::runOnMachineFunction(MachineFunction &MF) { 262 TLI = MF.getTarget().getTargetLowering(); 263 TII = MF.getTarget().getInstrInfo(); 264 TRI = MF.getTarget().getRegisterInfo(); 265 MBPI = &getAnalysis<MachineBranchProbabilityInfo>(); 266 InstrItins = MF.getTarget().getInstrItineraryData(); 267 if (!TII) return false; 268 269 // Tail merge tend to expose more if-conversion opportunities. 270 BranchFolder BF(true, false); 271 bool BFChange = BF.OptimizeFunction(MF, TII, 272 MF.getTarget().getRegisterInfo(), 273 getAnalysisIfAvailable<MachineModuleInfo>()); 274 275 DEBUG(dbgs() << "\nIfcvt: function (" << ++FnNum << ") \'" 276 << MF.getFunction()->getName() << "\'"); 277 278 if (FnNum < IfCvtFnStart || (IfCvtFnStop != -1 && FnNum > IfCvtFnStop)) { 279 DEBUG(dbgs() << " skipped\n"); 280 return false; 281 } 282 DEBUG(dbgs() << "\n"); 283 284 MF.RenumberBlocks(); 285 BBAnalysis.resize(MF.getNumBlockIDs()); 286 287 std::vector<IfcvtToken*> Tokens; 288 MadeChange = false; 289 unsigned NumIfCvts = NumSimple + NumSimpleFalse + NumTriangle + 290 NumTriangleRev + NumTriangleFalse + NumTriangleFRev + NumDiamonds; 291 while (IfCvtLimit == -1 || (int)NumIfCvts < IfCvtLimit) { 292 // Do an initial analysis for each basic block and find all the potential 293 // candidates to perform if-conversion. 294 bool Change = false; 295 AnalyzeBlocks(MF, Tokens); 296 while (!Tokens.empty()) { 297 IfcvtToken *Token = Tokens.back(); 298 Tokens.pop_back(); 299 BBInfo &BBI = Token->BBI; 300 IfcvtKind Kind = Token->Kind; 301 unsigned NumDups = Token->NumDups; 302 unsigned NumDups2 = Token->NumDups2; 303 304 delete Token; 305 306 // If the block has been evicted out of the queue or it has already been 307 // marked dead (due to it being predicated), then skip it. 308 if (BBI.IsDone) 309 BBI.IsEnqueued = false; 310 if (!BBI.IsEnqueued) 311 continue; 312 313 BBI.IsEnqueued = false; 314 315 bool RetVal = false; 316 switch (Kind) { 317 default: llvm_unreachable("Unexpected!"); 318 case ICSimple: 319 case ICSimpleFalse: { 320 bool isFalse = Kind == ICSimpleFalse; 321 if ((isFalse && DisableSimpleF) || (!isFalse && DisableSimple)) break; 322 DEBUG(dbgs() << "Ifcvt (Simple" << (Kind == ICSimpleFalse ? 323 " false" : "") 324 << "): BB#" << BBI.BB->getNumber() << " (" 325 << ((Kind == ICSimpleFalse) 326 ? BBI.FalseBB->getNumber() 327 : BBI.TrueBB->getNumber()) << ") "); 328 RetVal = IfConvertSimple(BBI, Kind); 329 DEBUG(dbgs() << (RetVal ? "succeeded!" : "failed!") << "\n"); 330 if (RetVal) { 331 if (isFalse) ++NumSimpleFalse; 332 else ++NumSimple; 333 } 334 break; 335 } 336 case ICTriangle: 337 case ICTriangleRev: 338 case ICTriangleFalse: 339 case ICTriangleFRev: { 340 bool isFalse = Kind == ICTriangleFalse; 341 bool isRev = (Kind == ICTriangleRev || Kind == ICTriangleFRev); 342 if (DisableTriangle && !isFalse && !isRev) break; 343 if (DisableTriangleR && !isFalse && isRev) break; 344 if (DisableTriangleF && isFalse && !isRev) break; 345 if (DisableTriangleFR && isFalse && isRev) break; 346 DEBUG(dbgs() << "Ifcvt (Triangle"); 347 if (isFalse) 348 DEBUG(dbgs() << " false"); 349 if (isRev) 350 DEBUG(dbgs() << " rev"); 351 DEBUG(dbgs() << "): BB#" << BBI.BB->getNumber() << " (T:" 352 << BBI.TrueBB->getNumber() << ",F:" 353 << BBI.FalseBB->getNumber() << ") "); 354 RetVal = IfConvertTriangle(BBI, Kind); 355 DEBUG(dbgs() << (RetVal ? "succeeded!" : "failed!") << "\n"); 356 if (RetVal) { 357 if (isFalse) { 358 if (isRev) ++NumTriangleFRev; 359 else ++NumTriangleFalse; 360 } else { 361 if (isRev) ++NumTriangleRev; 362 else ++NumTriangle; 363 } 364 } 365 break; 366 } 367 case ICDiamond: { 368 if (DisableDiamond) break; 369 DEBUG(dbgs() << "Ifcvt (Diamond): BB#" << BBI.BB->getNumber() << " (T:" 370 << BBI.TrueBB->getNumber() << ",F:" 371 << BBI.FalseBB->getNumber() << ") "); 372 RetVal = IfConvertDiamond(BBI, Kind, NumDups, NumDups2); 373 DEBUG(dbgs() << (RetVal ? "succeeded!" : "failed!") << "\n"); 374 if (RetVal) ++NumDiamonds; 375 break; 376 } 377 } 378 379 Change |= RetVal; 380 381 NumIfCvts = NumSimple + NumSimpleFalse + NumTriangle + NumTriangleRev + 382 NumTriangleFalse + NumTriangleFRev + NumDiamonds; 383 if (IfCvtLimit != -1 && (int)NumIfCvts >= IfCvtLimit) 384 break; 385 } 386 387 if (!Change) 388 break; 389 MadeChange |= Change; 390 } 391 392 // Delete tokens in case of early exit. 393 while (!Tokens.empty()) { 394 IfcvtToken *Token = Tokens.back(); 395 Tokens.pop_back(); 396 delete Token; 397 } 398 399 Tokens.clear(); 400 BBAnalysis.clear(); 401 402 if (MadeChange && IfCvtBranchFold) { 403 BranchFolder BF(false, false); 404 BF.OptimizeFunction(MF, TII, 405 MF.getTarget().getRegisterInfo(), 406 getAnalysisIfAvailable<MachineModuleInfo>()); 407 } 408 409 MadeChange |= BFChange; 410 return MadeChange; 411 } 412 413 /// findFalseBlock - BB has a fallthrough. Find its 'false' successor given 414 /// its 'true' successor. 415 static MachineBasicBlock *findFalseBlock(MachineBasicBlock *BB, 416 MachineBasicBlock *TrueBB) { 417 for (MachineBasicBlock::succ_iterator SI = BB->succ_begin(), 418 E = BB->succ_end(); SI != E; ++SI) { 419 MachineBasicBlock *SuccBB = *SI; 420 if (SuccBB != TrueBB) 421 return SuccBB; 422 } 423 return NULL; 424 } 425 426 /// ReverseBranchCondition - Reverse the condition of the end of the block 427 /// branch. Swap block's 'true' and 'false' successors. 428 bool IfConverter::ReverseBranchCondition(BBInfo &BBI) { 429 DebugLoc dl; // FIXME: this is nowhere 430 if (!TII->ReverseBranchCondition(BBI.BrCond)) { 431 TII->RemoveBranch(*BBI.BB); 432 TII->InsertBranch(*BBI.BB, BBI.FalseBB, BBI.TrueBB, BBI.BrCond, dl); 433 std::swap(BBI.TrueBB, BBI.FalseBB); 434 return true; 435 } 436 return false; 437 } 438 439 /// getNextBlock - Returns the next block in the function blocks ordering. If 440 /// it is the end, returns NULL. 441 static inline MachineBasicBlock *getNextBlock(MachineBasicBlock *BB) { 442 MachineFunction::iterator I = BB; 443 MachineFunction::iterator E = BB->getParent()->end(); 444 if (++I == E) 445 return NULL; 446 return I; 447 } 448 449 /// ValidSimple - Returns true if the 'true' block (along with its 450 /// predecessor) forms a valid simple shape for ifcvt. It also returns the 451 /// number of instructions that the ifcvt would need to duplicate if performed 452 /// in Dups. 453 bool IfConverter::ValidSimple(BBInfo &TrueBBI, unsigned &Dups, 454 const BranchProbability &Prediction) const { 455 Dups = 0; 456 if (TrueBBI.IsBeingAnalyzed || TrueBBI.IsDone) 457 return false; 458 459 if (TrueBBI.IsBrAnalyzable) 460 return false; 461 462 if (TrueBBI.BB->pred_size() > 1) { 463 if (TrueBBI.CannotBeCopied || 464 !TII->isProfitableToDupForIfCvt(*TrueBBI.BB, TrueBBI.NonPredSize, 465 Prediction)) 466 return false; 467 Dups = TrueBBI.NonPredSize; 468 } 469 470 return true; 471 } 472 473 /// ValidTriangle - Returns true if the 'true' and 'false' blocks (along 474 /// with their common predecessor) forms a valid triangle shape for ifcvt. 475 /// If 'FalseBranch' is true, it checks if 'true' block's false branch 476 /// branches to the 'false' block rather than the other way around. It also 477 /// returns the number of instructions that the ifcvt would need to duplicate 478 /// if performed in 'Dups'. 479 bool IfConverter::ValidTriangle(BBInfo &TrueBBI, BBInfo &FalseBBI, 480 bool FalseBranch, unsigned &Dups, 481 const BranchProbability &Prediction) const { 482 Dups = 0; 483 if (TrueBBI.IsBeingAnalyzed || TrueBBI.IsDone) 484 return false; 485 486 if (TrueBBI.BB->pred_size() > 1) { 487 if (TrueBBI.CannotBeCopied) 488 return false; 489 490 unsigned Size = TrueBBI.NonPredSize; 491 if (TrueBBI.IsBrAnalyzable) { 492 if (TrueBBI.TrueBB && TrueBBI.BrCond.empty()) 493 // Ends with an unconditional branch. It will be removed. 494 --Size; 495 else { 496 MachineBasicBlock *FExit = FalseBranch 497 ? TrueBBI.TrueBB : TrueBBI.FalseBB; 498 if (FExit) 499 // Require a conditional branch 500 ++Size; 501 } 502 } 503 if (!TII->isProfitableToDupForIfCvt(*TrueBBI.BB, Size, Prediction)) 504 return false; 505 Dups = Size; 506 } 507 508 MachineBasicBlock *TExit = FalseBranch ? TrueBBI.FalseBB : TrueBBI.TrueBB; 509 if (!TExit && blockAlwaysFallThrough(TrueBBI)) { 510 MachineFunction::iterator I = TrueBBI.BB; 511 if (++I == TrueBBI.BB->getParent()->end()) 512 return false; 513 TExit = I; 514 } 515 return TExit && TExit == FalseBBI.BB; 516 } 517 518 /// ValidDiamond - Returns true if the 'true' and 'false' blocks (along 519 /// with their common predecessor) forms a valid diamond shape for ifcvt. 520 bool IfConverter::ValidDiamond(BBInfo &TrueBBI, BBInfo &FalseBBI, 521 unsigned &Dups1, unsigned &Dups2) const { 522 Dups1 = Dups2 = 0; 523 if (TrueBBI.IsBeingAnalyzed || TrueBBI.IsDone || 524 FalseBBI.IsBeingAnalyzed || FalseBBI.IsDone) 525 return false; 526 527 MachineBasicBlock *TT = TrueBBI.TrueBB; 528 MachineBasicBlock *FT = FalseBBI.TrueBB; 529 530 if (!TT && blockAlwaysFallThrough(TrueBBI)) 531 TT = getNextBlock(TrueBBI.BB); 532 if (!FT && blockAlwaysFallThrough(FalseBBI)) 533 FT = getNextBlock(FalseBBI.BB); 534 if (TT != FT) 535 return false; 536 if (TT == NULL && (TrueBBI.IsBrAnalyzable || FalseBBI.IsBrAnalyzable)) 537 return false; 538 if (TrueBBI.BB->pred_size() > 1 || FalseBBI.BB->pred_size() > 1) 539 return false; 540 541 // FIXME: Allow true block to have an early exit? 542 if (TrueBBI.FalseBB || FalseBBI.FalseBB || 543 (TrueBBI.ClobbersPred && FalseBBI.ClobbersPred)) 544 return false; 545 546 // Count duplicate instructions at the beginning of the true and false blocks. 547 MachineBasicBlock::iterator TIB = TrueBBI.BB->begin(); 548 MachineBasicBlock::iterator FIB = FalseBBI.BB->begin(); 549 MachineBasicBlock::iterator TIE = TrueBBI.BB->end(); 550 MachineBasicBlock::iterator FIE = FalseBBI.BB->end(); 551 while (TIB != TIE && FIB != FIE) { 552 // Skip dbg_value instructions. These do not count. 553 if (TIB->isDebugValue()) { 554 while (TIB != TIE && TIB->isDebugValue()) 555 ++TIB; 556 if (TIB == TIE) 557 break; 558 } 559 if (FIB->isDebugValue()) { 560 while (FIB != FIE && FIB->isDebugValue()) 561 ++FIB; 562 if (FIB == FIE) 563 break; 564 } 565 if (!TIB->isIdenticalTo(FIB)) 566 break; 567 ++Dups1; 568 ++TIB; 569 ++FIB; 570 } 571 572 // Now, in preparation for counting duplicate instructions at the ends of the 573 // blocks, move the end iterators up past any branch instructions. 574 while (TIE != TIB) { 575 --TIE; 576 if (!TIE->isBranch()) 577 break; 578 } 579 while (FIE != FIB) { 580 --FIE; 581 if (!FIE->isBranch()) 582 break; 583 } 584 585 // If Dups1 includes all of a block, then don't count duplicate 586 // instructions at the end of the blocks. 587 if (TIB == TIE || FIB == FIE) 588 return true; 589 590 // Count duplicate instructions at the ends of the blocks. 591 while (TIE != TIB && FIE != FIB) { 592 // Skip dbg_value instructions. These do not count. 593 if (TIE->isDebugValue()) { 594 while (TIE != TIB && TIE->isDebugValue()) 595 --TIE; 596 if (TIE == TIB) 597 break; 598 } 599 if (FIE->isDebugValue()) { 600 while (FIE != FIB && FIE->isDebugValue()) 601 --FIE; 602 if (FIE == FIB) 603 break; 604 } 605 if (!TIE->isIdenticalTo(FIE)) 606 break; 607 ++Dups2; 608 --TIE; 609 --FIE; 610 } 611 612 return true; 613 } 614 615 /// ScanInstructions - Scan all the instructions in the block to determine if 616 /// the block is predicable. In most cases, that means all the instructions 617 /// in the block are isPredicable(). Also checks if the block contains any 618 /// instruction which can clobber a predicate (e.g. condition code register). 619 /// If so, the block is not predicable unless it's the last instruction. 620 void IfConverter::ScanInstructions(BBInfo &BBI) { 621 if (BBI.IsDone) 622 return; 623 624 bool AlreadyPredicated = BBI.Predicate.size() > 0; 625 // First analyze the end of BB branches. 626 BBI.TrueBB = BBI.FalseBB = NULL; 627 BBI.BrCond.clear(); 628 BBI.IsBrAnalyzable = 629 !TII->AnalyzeBranch(*BBI.BB, BBI.TrueBB, BBI.FalseBB, BBI.BrCond); 630 BBI.HasFallThrough = BBI.IsBrAnalyzable && BBI.FalseBB == NULL; 631 632 if (BBI.BrCond.size()) { 633 // No false branch. This BB must end with a conditional branch and a 634 // fallthrough. 635 if (!BBI.FalseBB) 636 BBI.FalseBB = findFalseBlock(BBI.BB, BBI.TrueBB); 637 if (!BBI.FalseBB) { 638 // Malformed bcc? True and false blocks are the same? 639 BBI.IsUnpredicable = true; 640 return; 641 } 642 } 643 644 // Then scan all the instructions. 645 BBI.NonPredSize = 0; 646 BBI.ExtraCost = 0; 647 BBI.ExtraCost2 = 0; 648 BBI.ClobbersPred = false; 649 for (MachineBasicBlock::iterator I = BBI.BB->begin(), E = BBI.BB->end(); 650 I != E; ++I) { 651 if (I->isDebugValue()) 652 continue; 653 654 if (I->isNotDuplicable()) 655 BBI.CannotBeCopied = true; 656 657 bool isPredicated = TII->isPredicated(I); 658 bool isCondBr = BBI.IsBrAnalyzable && I->isConditionalBranch(); 659 660 if (!isCondBr) { 661 if (!isPredicated) { 662 BBI.NonPredSize++; 663 unsigned ExtraPredCost = 0; 664 unsigned NumCycles = TII->getInstrLatency(InstrItins, &*I, 665 &ExtraPredCost); 666 if (NumCycles > 1) 667 BBI.ExtraCost += NumCycles-1; 668 BBI.ExtraCost2 += ExtraPredCost; 669 } else if (!AlreadyPredicated) { 670 // FIXME: This instruction is already predicated before the 671 // if-conversion pass. It's probably something like a conditional move. 672 // Mark this block unpredicable for now. 673 BBI.IsUnpredicable = true; 674 return; 675 } 676 } 677 678 if (BBI.ClobbersPred && !isPredicated) { 679 // Predicate modification instruction should end the block (except for 680 // already predicated instructions and end of block branches). 681 if (isCondBr) { 682 // A conditional branch is not predicable, but it may be eliminated. 683 continue; 684 } 685 686 // Predicate may have been modified, the subsequent (currently) 687 // unpredicated instructions cannot be correctly predicated. 688 BBI.IsUnpredicable = true; 689 return; 690 } 691 692 // FIXME: Make use of PredDefs? e.g. ADDC, SUBC sets predicates but are 693 // still potentially predicable. 694 std::vector<MachineOperand> PredDefs; 695 if (TII->DefinesPredicate(I, PredDefs)) 696 BBI.ClobbersPred = true; 697 698 if (!TII->isPredicable(I)) { 699 BBI.IsUnpredicable = true; 700 return; 701 } 702 } 703 } 704 705 /// FeasibilityAnalysis - Determine if the block is a suitable candidate to be 706 /// predicated by the specified predicate. 707 bool IfConverter::FeasibilityAnalysis(BBInfo &BBI, 708 SmallVectorImpl<MachineOperand> &Pred, 709 bool isTriangle, bool RevBranch) { 710 // If the block is dead or unpredicable, then it cannot be predicated. 711 if (BBI.IsDone || BBI.IsUnpredicable) 712 return false; 713 714 // If it is already predicated, check if its predicate subsumes the new 715 // predicate. 716 if (BBI.Predicate.size() && !TII->SubsumesPredicate(BBI.Predicate, Pred)) 717 return false; 718 719 if (BBI.BrCond.size()) { 720 if (!isTriangle) 721 return false; 722 723 // Test predicate subsumption. 724 SmallVector<MachineOperand, 4> RevPred(Pred.begin(), Pred.end()); 725 SmallVector<MachineOperand, 4> Cond(BBI.BrCond.begin(), BBI.BrCond.end()); 726 if (RevBranch) { 727 if (TII->ReverseBranchCondition(Cond)) 728 return false; 729 } 730 if (TII->ReverseBranchCondition(RevPred) || 731 !TII->SubsumesPredicate(Cond, RevPred)) 732 return false; 733 } 734 735 return true; 736 } 737 738 /// AnalyzeBlock - Analyze the structure of the sub-CFG starting from 739 /// the specified block. Record its successors and whether it looks like an 740 /// if-conversion candidate. 741 IfConverter::BBInfo &IfConverter::AnalyzeBlock(MachineBasicBlock *BB, 742 std::vector<IfcvtToken*> &Tokens) { 743 BBInfo &BBI = BBAnalysis[BB->getNumber()]; 744 745 if (BBI.IsAnalyzed || BBI.IsBeingAnalyzed) 746 return BBI; 747 748 BBI.BB = BB; 749 BBI.IsBeingAnalyzed = true; 750 751 ScanInstructions(BBI); 752 753 // Unanalyzable or ends with fallthrough or unconditional branch, or if is not 754 // considered for ifcvt anymore. 755 if (!BBI.IsBrAnalyzable || BBI.BrCond.empty() || BBI.IsDone) { 756 BBI.IsBeingAnalyzed = false; 757 BBI.IsAnalyzed = true; 758 return BBI; 759 } 760 761 // Do not ifcvt if either path is a back edge to the entry block. 762 if (BBI.TrueBB == BB || BBI.FalseBB == BB) { 763 BBI.IsBeingAnalyzed = false; 764 BBI.IsAnalyzed = true; 765 return BBI; 766 } 767 768 // Do not ifcvt if true and false fallthrough blocks are the same. 769 if (!BBI.FalseBB) { 770 BBI.IsBeingAnalyzed = false; 771 BBI.IsAnalyzed = true; 772 return BBI; 773 } 774 775 BBInfo &TrueBBI = AnalyzeBlock(BBI.TrueBB, Tokens); 776 BBInfo &FalseBBI = AnalyzeBlock(BBI.FalseBB, Tokens); 777 778 if (TrueBBI.IsDone && FalseBBI.IsDone) { 779 BBI.IsBeingAnalyzed = false; 780 BBI.IsAnalyzed = true; 781 return BBI; 782 } 783 784 SmallVector<MachineOperand, 4> RevCond(BBI.BrCond.begin(), BBI.BrCond.end()); 785 bool CanRevCond = !TII->ReverseBranchCondition(RevCond); 786 787 unsigned Dups = 0; 788 unsigned Dups2 = 0; 789 bool TNeedSub = TrueBBI.Predicate.size() > 0; 790 bool FNeedSub = FalseBBI.Predicate.size() > 0; 791 bool Enqueued = false; 792 793 BranchProbability Prediction = MBPI->getEdgeProbability(BB, TrueBBI.BB); 794 795 if (CanRevCond && ValidDiamond(TrueBBI, FalseBBI, Dups, Dups2) && 796 MeetIfcvtSizeLimit(*TrueBBI.BB, (TrueBBI.NonPredSize - (Dups + Dups2) + 797 TrueBBI.ExtraCost), TrueBBI.ExtraCost2, 798 *FalseBBI.BB, (FalseBBI.NonPredSize - (Dups + Dups2) + 799 FalseBBI.ExtraCost),FalseBBI.ExtraCost2, 800 Prediction) && 801 FeasibilityAnalysis(TrueBBI, BBI.BrCond) && 802 FeasibilityAnalysis(FalseBBI, RevCond)) { 803 // Diamond: 804 // EBB 805 // / \_ 806 // | | 807 // TBB FBB 808 // \ / 809 // TailBB 810 // Note TailBB can be empty. 811 Tokens.push_back(new IfcvtToken(BBI, ICDiamond, TNeedSub|FNeedSub, Dups, 812 Dups2)); 813 Enqueued = true; 814 } 815 816 if (ValidTriangle(TrueBBI, FalseBBI, false, Dups, Prediction) && 817 MeetIfcvtSizeLimit(*TrueBBI.BB, TrueBBI.NonPredSize + TrueBBI.ExtraCost, 818 TrueBBI.ExtraCost2, Prediction) && 819 FeasibilityAnalysis(TrueBBI, BBI.BrCond, true)) { 820 // Triangle: 821 // EBB 822 // | \_ 823 // | | 824 // | TBB 825 // | / 826 // FBB 827 Tokens.push_back(new IfcvtToken(BBI, ICTriangle, TNeedSub, Dups)); 828 Enqueued = true; 829 } 830 831 if (ValidTriangle(TrueBBI, FalseBBI, true, Dups, Prediction) && 832 MeetIfcvtSizeLimit(*TrueBBI.BB, TrueBBI.NonPredSize + TrueBBI.ExtraCost, 833 TrueBBI.ExtraCost2, Prediction) && 834 FeasibilityAnalysis(TrueBBI, BBI.BrCond, true, true)) { 835 Tokens.push_back(new IfcvtToken(BBI, ICTriangleRev, TNeedSub, Dups)); 836 Enqueued = true; 837 } 838 839 if (ValidSimple(TrueBBI, Dups, Prediction) && 840 MeetIfcvtSizeLimit(*TrueBBI.BB, TrueBBI.NonPredSize + TrueBBI.ExtraCost, 841 TrueBBI.ExtraCost2, Prediction) && 842 FeasibilityAnalysis(TrueBBI, BBI.BrCond)) { 843 // Simple (split, no rejoin): 844 // EBB 845 // | \_ 846 // | | 847 // | TBB---> exit 848 // | 849 // FBB 850 Tokens.push_back(new IfcvtToken(BBI, ICSimple, TNeedSub, Dups)); 851 Enqueued = true; 852 } 853 854 if (CanRevCond) { 855 // Try the other path... 856 if (ValidTriangle(FalseBBI, TrueBBI, false, Dups, 857 Prediction.getCompl()) && 858 MeetIfcvtSizeLimit(*FalseBBI.BB, 859 FalseBBI.NonPredSize + FalseBBI.ExtraCost, 860 FalseBBI.ExtraCost2, Prediction.getCompl()) && 861 FeasibilityAnalysis(FalseBBI, RevCond, true)) { 862 Tokens.push_back(new IfcvtToken(BBI, ICTriangleFalse, FNeedSub, Dups)); 863 Enqueued = true; 864 } 865 866 if (ValidTriangle(FalseBBI, TrueBBI, true, Dups, 867 Prediction.getCompl()) && 868 MeetIfcvtSizeLimit(*FalseBBI.BB, 869 FalseBBI.NonPredSize + FalseBBI.ExtraCost, 870 FalseBBI.ExtraCost2, Prediction.getCompl()) && 871 FeasibilityAnalysis(FalseBBI, RevCond, true, true)) { 872 Tokens.push_back(new IfcvtToken(BBI, ICTriangleFRev, FNeedSub, Dups)); 873 Enqueued = true; 874 } 875 876 if (ValidSimple(FalseBBI, Dups, Prediction.getCompl()) && 877 MeetIfcvtSizeLimit(*FalseBBI.BB, 878 FalseBBI.NonPredSize + FalseBBI.ExtraCost, 879 FalseBBI.ExtraCost2, Prediction.getCompl()) && 880 FeasibilityAnalysis(FalseBBI, RevCond)) { 881 Tokens.push_back(new IfcvtToken(BBI, ICSimpleFalse, FNeedSub, Dups)); 882 Enqueued = true; 883 } 884 } 885 886 BBI.IsEnqueued = Enqueued; 887 BBI.IsBeingAnalyzed = false; 888 BBI.IsAnalyzed = true; 889 return BBI; 890 } 891 892 /// AnalyzeBlocks - Analyze all blocks and find entries for all if-conversion 893 /// candidates. 894 void IfConverter::AnalyzeBlocks(MachineFunction &MF, 895 std::vector<IfcvtToken*> &Tokens) { 896 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; ++I) { 897 MachineBasicBlock *BB = I; 898 AnalyzeBlock(BB, Tokens); 899 } 900 901 // Sort to favor more complex ifcvt scheme. 902 std::stable_sort(Tokens.begin(), Tokens.end(), IfcvtTokenCmp); 903 } 904 905 /// canFallThroughTo - Returns true either if ToBB is the next block after BB or 906 /// that all the intervening blocks are empty (given BB can fall through to its 907 /// next block). 908 static bool canFallThroughTo(MachineBasicBlock *BB, MachineBasicBlock *ToBB) { 909 MachineFunction::iterator PI = BB; 910 MachineFunction::iterator I = llvm::next(PI); 911 MachineFunction::iterator TI = ToBB; 912 MachineFunction::iterator E = BB->getParent()->end(); 913 while (I != TI) { 914 // Check isSuccessor to avoid case where the next block is empty, but 915 // it's not a successor. 916 if (I == E || !I->empty() || !PI->isSuccessor(I)) 917 return false; 918 PI = I++; 919 } 920 return true; 921 } 922 923 /// InvalidatePreds - Invalidate predecessor BB info so it would be re-analyzed 924 /// to determine if it can be if-converted. If predecessor is already enqueued, 925 /// dequeue it! 926 void IfConverter::InvalidatePreds(MachineBasicBlock *BB) { 927 for (MachineBasicBlock::pred_iterator PI = BB->pred_begin(), 928 E = BB->pred_end(); PI != E; ++PI) { 929 BBInfo &PBBI = BBAnalysis[(*PI)->getNumber()]; 930 if (PBBI.IsDone || PBBI.BB == BB) 931 continue; 932 PBBI.IsAnalyzed = false; 933 PBBI.IsEnqueued = false; 934 } 935 } 936 937 /// InsertUncondBranch - Inserts an unconditional branch from BB to ToBB. 938 /// 939 static void InsertUncondBranch(MachineBasicBlock *BB, MachineBasicBlock *ToBB, 940 const TargetInstrInfo *TII) { 941 DebugLoc dl; // FIXME: this is nowhere 942 SmallVector<MachineOperand, 0> NoCond; 943 TII->InsertBranch(*BB, ToBB, NULL, NoCond, dl); 944 } 945 946 /// RemoveExtraEdges - Remove true / false edges if either / both are no longer 947 /// successors. 948 void IfConverter::RemoveExtraEdges(BBInfo &BBI) { 949 MachineBasicBlock *TBB = NULL, *FBB = NULL; 950 SmallVector<MachineOperand, 4> Cond; 951 if (!TII->AnalyzeBranch(*BBI.BB, TBB, FBB, Cond)) 952 BBI.BB->CorrectExtraCFGEdges(TBB, FBB, !Cond.empty()); 953 } 954 955 /// InitPredRedefs / UpdatePredRedefs - Defs by predicated instructions are 956 /// modeled as read + write (sort like two-address instructions). These 957 /// routines track register liveness and add implicit uses to if-converted 958 /// instructions to conform to the model. 959 static void InitPredRedefs(MachineBasicBlock *BB, SmallSet<unsigned,4> &Redefs, 960 const TargetRegisterInfo *TRI) { 961 for (MachineBasicBlock::livein_iterator I = BB->livein_begin(), 962 E = BB->livein_end(); I != E; ++I) { 963 unsigned Reg = *I; 964 Redefs.insert(Reg); 965 for (const uint16_t *Subreg = TRI->getSubRegisters(Reg); 966 *Subreg; ++Subreg) 967 Redefs.insert(*Subreg); 968 } 969 } 970 971 static void UpdatePredRedefs(MachineInstr *MI, SmallSet<unsigned,4> &Redefs, 972 const TargetRegisterInfo *TRI, 973 bool AddImpUse = false) { 974 SmallVector<unsigned, 4> Defs; 975 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) { 976 const MachineOperand &MO = MI->getOperand(i); 977 if (!MO.isReg()) 978 continue; 979 unsigned Reg = MO.getReg(); 980 if (!Reg) 981 continue; 982 if (MO.isDef()) 983 Defs.push_back(Reg); 984 else if (MO.isKill()) { 985 Redefs.erase(Reg); 986 for (const uint16_t *SR = TRI->getSubRegisters(Reg); *SR; ++SR) 987 Redefs.erase(*SR); 988 } 989 } 990 for (unsigned i = 0, e = Defs.size(); i != e; ++i) { 991 unsigned Reg = Defs[i]; 992 if (Redefs.count(Reg)) { 993 if (AddImpUse) 994 // Treat predicated update as read + write. 995 MI->addOperand(MachineOperand::CreateReg(Reg, false/*IsDef*/, 996 true/*IsImp*/,false/*IsKill*/)); 997 } else { 998 Redefs.insert(Reg); 999 for (const uint16_t *SR = TRI->getSubRegisters(Reg); *SR; ++SR) 1000 Redefs.insert(*SR); 1001 } 1002 } 1003 } 1004 1005 static void UpdatePredRedefs(MachineBasicBlock::iterator I, 1006 MachineBasicBlock::iterator E, 1007 SmallSet<unsigned,4> &Redefs, 1008 const TargetRegisterInfo *TRI) { 1009 while (I != E) { 1010 UpdatePredRedefs(I, Redefs, TRI); 1011 ++I; 1012 } 1013 } 1014 1015 /// IfConvertSimple - If convert a simple (split, no rejoin) sub-CFG. 1016 /// 1017 bool IfConverter::IfConvertSimple(BBInfo &BBI, IfcvtKind Kind) { 1018 BBInfo &TrueBBI = BBAnalysis[BBI.TrueBB->getNumber()]; 1019 BBInfo &FalseBBI = BBAnalysis[BBI.FalseBB->getNumber()]; 1020 BBInfo *CvtBBI = &TrueBBI; 1021 BBInfo *NextBBI = &FalseBBI; 1022 1023 SmallVector<MachineOperand, 4> Cond(BBI.BrCond.begin(), BBI.BrCond.end()); 1024 if (Kind == ICSimpleFalse) 1025 std::swap(CvtBBI, NextBBI); 1026 1027 if (CvtBBI->IsDone || 1028 (CvtBBI->CannotBeCopied && CvtBBI->BB->pred_size() > 1)) { 1029 // Something has changed. It's no longer safe to predicate this block. 1030 BBI.IsAnalyzed = false; 1031 CvtBBI->IsAnalyzed = false; 1032 return false; 1033 } 1034 1035 if (Kind == ICSimpleFalse) 1036 if (TII->ReverseBranchCondition(Cond)) 1037 llvm_unreachable("Unable to reverse branch condition!"); 1038 1039 // Initialize liveins to the first BB. These are potentiall redefined by 1040 // predicated instructions. 1041 SmallSet<unsigned, 4> Redefs; 1042 InitPredRedefs(CvtBBI->BB, Redefs, TRI); 1043 InitPredRedefs(NextBBI->BB, Redefs, TRI); 1044 1045 if (CvtBBI->BB->pred_size() > 1) { 1046 BBI.NonPredSize -= TII->RemoveBranch(*BBI.BB); 1047 // Copy instructions in the true block, predicate them, and add them to 1048 // the entry block. 1049 CopyAndPredicateBlock(BBI, *CvtBBI, Cond, Redefs); 1050 } else { 1051 PredicateBlock(*CvtBBI, CvtBBI->BB->end(), Cond, Redefs); 1052 1053 // Merge converted block into entry block. 1054 BBI.NonPredSize -= TII->RemoveBranch(*BBI.BB); 1055 MergeBlocks(BBI, *CvtBBI); 1056 } 1057 1058 bool IterIfcvt = true; 1059 if (!canFallThroughTo(BBI.BB, NextBBI->BB)) { 1060 InsertUncondBranch(BBI.BB, NextBBI->BB, TII); 1061 BBI.HasFallThrough = false; 1062 // Now ifcvt'd block will look like this: 1063 // BB: 1064 // ... 1065 // t, f = cmp 1066 // if t op 1067 // b BBf 1068 // 1069 // We cannot further ifcvt this block because the unconditional branch 1070 // will have to be predicated on the new condition, that will not be 1071 // available if cmp executes. 1072 IterIfcvt = false; 1073 } 1074 1075 RemoveExtraEdges(BBI); 1076 1077 // Update block info. BB can be iteratively if-converted. 1078 if (!IterIfcvt) 1079 BBI.IsDone = true; 1080 InvalidatePreds(BBI.BB); 1081 CvtBBI->IsDone = true; 1082 1083 // FIXME: Must maintain LiveIns. 1084 return true; 1085 } 1086 1087 /// IfConvertTriangle - If convert a triangle sub-CFG. 1088 /// 1089 bool IfConverter::IfConvertTriangle(BBInfo &BBI, IfcvtKind Kind) { 1090 BBInfo &TrueBBI = BBAnalysis[BBI.TrueBB->getNumber()]; 1091 BBInfo &FalseBBI = BBAnalysis[BBI.FalseBB->getNumber()]; 1092 BBInfo *CvtBBI = &TrueBBI; 1093 BBInfo *NextBBI = &FalseBBI; 1094 DebugLoc dl; // FIXME: this is nowhere 1095 1096 SmallVector<MachineOperand, 4> Cond(BBI.BrCond.begin(), BBI.BrCond.end()); 1097 if (Kind == ICTriangleFalse || Kind == ICTriangleFRev) 1098 std::swap(CvtBBI, NextBBI); 1099 1100 if (CvtBBI->IsDone || 1101 (CvtBBI->CannotBeCopied && CvtBBI->BB->pred_size() > 1)) { 1102 // Something has changed. It's no longer safe to predicate this block. 1103 BBI.IsAnalyzed = false; 1104 CvtBBI->IsAnalyzed = false; 1105 return false; 1106 } 1107 1108 if (Kind == ICTriangleFalse || Kind == ICTriangleFRev) 1109 if (TII->ReverseBranchCondition(Cond)) 1110 llvm_unreachable("Unable to reverse branch condition!"); 1111 1112 if (Kind == ICTriangleRev || Kind == ICTriangleFRev) { 1113 if (ReverseBranchCondition(*CvtBBI)) { 1114 // BB has been changed, modify its predecessors (except for this 1115 // one) so they don't get ifcvt'ed based on bad intel. 1116 for (MachineBasicBlock::pred_iterator PI = CvtBBI->BB->pred_begin(), 1117 E = CvtBBI->BB->pred_end(); PI != E; ++PI) { 1118 MachineBasicBlock *PBB = *PI; 1119 if (PBB == BBI.BB) 1120 continue; 1121 BBInfo &PBBI = BBAnalysis[PBB->getNumber()]; 1122 if (PBBI.IsEnqueued) { 1123 PBBI.IsAnalyzed = false; 1124 PBBI.IsEnqueued = false; 1125 } 1126 } 1127 } 1128 } 1129 1130 // Initialize liveins to the first BB. These are potentially redefined by 1131 // predicated instructions. 1132 SmallSet<unsigned, 4> Redefs; 1133 InitPredRedefs(CvtBBI->BB, Redefs, TRI); 1134 InitPredRedefs(NextBBI->BB, Redefs, TRI); 1135 1136 bool HasEarlyExit = CvtBBI->FalseBB != NULL; 1137 if (CvtBBI->BB->pred_size() > 1) { 1138 BBI.NonPredSize -= TII->RemoveBranch(*BBI.BB); 1139 // Copy instructions in the true block, predicate them, and add them to 1140 // the entry block. 1141 CopyAndPredicateBlock(BBI, *CvtBBI, Cond, Redefs, true); 1142 } else { 1143 // Predicate the 'true' block after removing its branch. 1144 CvtBBI->NonPredSize -= TII->RemoveBranch(*CvtBBI->BB); 1145 PredicateBlock(*CvtBBI, CvtBBI->BB->end(), Cond, Redefs); 1146 1147 // Now merge the entry of the triangle with the true block. 1148 BBI.NonPredSize -= TII->RemoveBranch(*BBI.BB); 1149 MergeBlocks(BBI, *CvtBBI, false); 1150 } 1151 1152 // If 'true' block has a 'false' successor, add an exit branch to it. 1153 if (HasEarlyExit) { 1154 SmallVector<MachineOperand, 4> RevCond(CvtBBI->BrCond.begin(), 1155 CvtBBI->BrCond.end()); 1156 if (TII->ReverseBranchCondition(RevCond)) 1157 llvm_unreachable("Unable to reverse branch condition!"); 1158 TII->InsertBranch(*BBI.BB, CvtBBI->FalseBB, NULL, RevCond, dl); 1159 BBI.BB->addSuccessor(CvtBBI->FalseBB); 1160 } 1161 1162 // Merge in the 'false' block if the 'false' block has no other 1163 // predecessors. Otherwise, add an unconditional branch to 'false'. 1164 bool FalseBBDead = false; 1165 bool IterIfcvt = true; 1166 bool isFallThrough = canFallThroughTo(BBI.BB, NextBBI->BB); 1167 if (!isFallThrough) { 1168 // Only merge them if the true block does not fallthrough to the false 1169 // block. By not merging them, we make it possible to iteratively 1170 // ifcvt the blocks. 1171 if (!HasEarlyExit && 1172 NextBBI->BB->pred_size() == 1 && !NextBBI->HasFallThrough) { 1173 MergeBlocks(BBI, *NextBBI); 1174 FalseBBDead = true; 1175 } else { 1176 InsertUncondBranch(BBI.BB, NextBBI->BB, TII); 1177 BBI.HasFallThrough = false; 1178 } 1179 // Mixed predicated and unpredicated code. This cannot be iteratively 1180 // predicated. 1181 IterIfcvt = false; 1182 } 1183 1184 RemoveExtraEdges(BBI); 1185 1186 // Update block info. BB can be iteratively if-converted. 1187 if (!IterIfcvt) 1188 BBI.IsDone = true; 1189 InvalidatePreds(BBI.BB); 1190 CvtBBI->IsDone = true; 1191 if (FalseBBDead) 1192 NextBBI->IsDone = true; 1193 1194 // FIXME: Must maintain LiveIns. 1195 return true; 1196 } 1197 1198 /// IfConvertDiamond - If convert a diamond sub-CFG. 1199 /// 1200 bool IfConverter::IfConvertDiamond(BBInfo &BBI, IfcvtKind Kind, 1201 unsigned NumDups1, unsigned NumDups2) { 1202 BBInfo &TrueBBI = BBAnalysis[BBI.TrueBB->getNumber()]; 1203 BBInfo &FalseBBI = BBAnalysis[BBI.FalseBB->getNumber()]; 1204 MachineBasicBlock *TailBB = TrueBBI.TrueBB; 1205 // True block must fall through or end with an unanalyzable terminator. 1206 if (!TailBB) { 1207 if (blockAlwaysFallThrough(TrueBBI)) 1208 TailBB = FalseBBI.TrueBB; 1209 assert((TailBB || !TrueBBI.IsBrAnalyzable) && "Unexpected!"); 1210 } 1211 1212 if (TrueBBI.IsDone || FalseBBI.IsDone || 1213 TrueBBI.BB->pred_size() > 1 || 1214 FalseBBI.BB->pred_size() > 1) { 1215 // Something has changed. It's no longer safe to predicate these blocks. 1216 BBI.IsAnalyzed = false; 1217 TrueBBI.IsAnalyzed = false; 1218 FalseBBI.IsAnalyzed = false; 1219 return false; 1220 } 1221 1222 // Put the predicated instructions from the 'true' block before the 1223 // instructions from the 'false' block, unless the true block would clobber 1224 // the predicate, in which case, do the opposite. 1225 BBInfo *BBI1 = &TrueBBI; 1226 BBInfo *BBI2 = &FalseBBI; 1227 SmallVector<MachineOperand, 4> RevCond(BBI.BrCond.begin(), BBI.BrCond.end()); 1228 if (TII->ReverseBranchCondition(RevCond)) 1229 llvm_unreachable("Unable to reverse branch condition!"); 1230 SmallVector<MachineOperand, 4> *Cond1 = &BBI.BrCond; 1231 SmallVector<MachineOperand, 4> *Cond2 = &RevCond; 1232 1233 // Figure out the more profitable ordering. 1234 bool DoSwap = false; 1235 if (TrueBBI.ClobbersPred && !FalseBBI.ClobbersPred) 1236 DoSwap = true; 1237 else if (TrueBBI.ClobbersPred == FalseBBI.ClobbersPred) { 1238 if (TrueBBI.NonPredSize > FalseBBI.NonPredSize) 1239 DoSwap = true; 1240 } 1241 if (DoSwap) { 1242 std::swap(BBI1, BBI2); 1243 std::swap(Cond1, Cond2); 1244 } 1245 1246 // Remove the conditional branch from entry to the blocks. 1247 BBI.NonPredSize -= TII->RemoveBranch(*BBI.BB); 1248 1249 // Initialize liveins to the first BB. These are potentially redefined by 1250 // predicated instructions. 1251 SmallSet<unsigned, 4> Redefs; 1252 InitPredRedefs(BBI1->BB, Redefs, TRI); 1253 1254 // Remove the duplicated instructions at the beginnings of both paths. 1255 MachineBasicBlock::iterator DI1 = BBI1->BB->begin(); 1256 MachineBasicBlock::iterator DI2 = BBI2->BB->begin(); 1257 MachineBasicBlock::iterator DIE1 = BBI1->BB->end(); 1258 MachineBasicBlock::iterator DIE2 = BBI2->BB->end(); 1259 // Skip dbg_value instructions 1260 while (DI1 != DIE1 && DI1->isDebugValue()) 1261 ++DI1; 1262 while (DI2 != DIE2 && DI2->isDebugValue()) 1263 ++DI2; 1264 BBI1->NonPredSize -= NumDups1; 1265 BBI2->NonPredSize -= NumDups1; 1266 1267 // Skip past the dups on each side separately since there may be 1268 // differing dbg_value entries. 1269 for (unsigned i = 0; i < NumDups1; ++DI1) { 1270 if (!DI1->isDebugValue()) 1271 ++i; 1272 } 1273 while (NumDups1 != 0) { 1274 ++DI2; 1275 if (!DI2->isDebugValue()) 1276 --NumDups1; 1277 } 1278 1279 UpdatePredRedefs(BBI1->BB->begin(), DI1, Redefs, TRI); 1280 BBI.BB->splice(BBI.BB->end(), BBI1->BB, BBI1->BB->begin(), DI1); 1281 BBI2->BB->erase(BBI2->BB->begin(), DI2); 1282 1283 // Remove branch from 'true' block and remove duplicated instructions. 1284 BBI1->NonPredSize -= TII->RemoveBranch(*BBI1->BB); 1285 DI1 = BBI1->BB->end(); 1286 for (unsigned i = 0; i != NumDups2; ) { 1287 // NumDups2 only counted non-dbg_value instructions, so this won't 1288 // run off the head of the list. 1289 assert (DI1 != BBI1->BB->begin()); 1290 --DI1; 1291 // skip dbg_value instructions 1292 if (!DI1->isDebugValue()) 1293 ++i; 1294 } 1295 BBI1->BB->erase(DI1, BBI1->BB->end()); 1296 1297 // Remove 'false' block branch and find the last instruction to predicate. 1298 BBI2->NonPredSize -= TII->RemoveBranch(*BBI2->BB); 1299 DI2 = BBI2->BB->end(); 1300 while (NumDups2 != 0) { 1301 // NumDups2 only counted non-dbg_value instructions, so this won't 1302 // run off the head of the list. 1303 assert (DI2 != BBI2->BB->begin()); 1304 --DI2; 1305 // skip dbg_value instructions 1306 if (!DI2->isDebugValue()) 1307 --NumDups2; 1308 } 1309 1310 // Remember which registers would later be defined by the false block. 1311 // This allows us not to predicate instructions in the true block that would 1312 // later be re-defined. That is, rather than 1313 // subeq r0, r1, #1 1314 // addne r0, r1, #1 1315 // generate: 1316 // sub r0, r1, #1 1317 // addne r0, r1, #1 1318 SmallSet<unsigned, 4> RedefsByFalse; 1319 SmallSet<unsigned, 4> ExtUses; 1320 if (TII->isProfitableToUnpredicate(*BBI1->BB, *BBI2->BB)) { 1321 for (MachineBasicBlock::iterator FI = BBI2->BB->begin(); FI != DI2; ++FI) { 1322 if (FI->isDebugValue()) 1323 continue; 1324 SmallVector<unsigned, 4> Defs; 1325 for (unsigned i = 0, e = FI->getNumOperands(); i != e; ++i) { 1326 const MachineOperand &MO = FI->getOperand(i); 1327 if (!MO.isReg()) 1328 continue; 1329 unsigned Reg = MO.getReg(); 1330 if (!Reg) 1331 continue; 1332 if (MO.isDef()) { 1333 Defs.push_back(Reg); 1334 } else if (!RedefsByFalse.count(Reg)) { 1335 // These are defined before ctrl flow reach the 'false' instructions. 1336 // They cannot be modified by the 'true' instructions. 1337 ExtUses.insert(Reg); 1338 for (const uint16_t *SR = TRI->getSubRegisters(Reg); *SR; ++SR) 1339 ExtUses.insert(*SR); 1340 } 1341 } 1342 1343 for (unsigned i = 0, e = Defs.size(); i != e; ++i) { 1344 unsigned Reg = Defs[i]; 1345 if (!ExtUses.count(Reg)) { 1346 RedefsByFalse.insert(Reg); 1347 for (const uint16_t *SR = TRI->getSubRegisters(Reg); *SR; ++SR) 1348 RedefsByFalse.insert(*SR); 1349 } 1350 } 1351 } 1352 } 1353 1354 // Predicate the 'true' block. 1355 PredicateBlock(*BBI1, BBI1->BB->end(), *Cond1, Redefs, &RedefsByFalse); 1356 1357 // Predicate the 'false' block. 1358 PredicateBlock(*BBI2, DI2, *Cond2, Redefs); 1359 1360 // Merge the true block into the entry of the diamond. 1361 MergeBlocks(BBI, *BBI1, TailBB == 0); 1362 MergeBlocks(BBI, *BBI2, TailBB == 0); 1363 1364 // If the if-converted block falls through or unconditionally branches into 1365 // the tail block, and the tail block does not have other predecessors, then 1366 // fold the tail block in as well. Otherwise, unless it falls through to the 1367 // tail, add a unconditional branch to it. 1368 if (TailBB) { 1369 BBInfo &TailBBI = BBAnalysis[TailBB->getNumber()]; 1370 bool CanMergeTail = !TailBBI.HasFallThrough; 1371 // There may still be a fall-through edge from BBI1 or BBI2 to TailBB; 1372 // check if there are any other predecessors besides those. 1373 unsigned NumPreds = TailBB->pred_size(); 1374 if (NumPreds > 1) 1375 CanMergeTail = false; 1376 else if (NumPreds == 1 && CanMergeTail) { 1377 MachineBasicBlock::pred_iterator PI = TailBB->pred_begin(); 1378 if (*PI != BBI1->BB && *PI != BBI2->BB) 1379 CanMergeTail = false; 1380 } 1381 if (CanMergeTail) { 1382 MergeBlocks(BBI, TailBBI); 1383 TailBBI.IsDone = true; 1384 } else { 1385 BBI.BB->addSuccessor(TailBB); 1386 InsertUncondBranch(BBI.BB, TailBB, TII); 1387 BBI.HasFallThrough = false; 1388 } 1389 } 1390 1391 // RemoveExtraEdges won't work if the block has an unanalyzable branch, 1392 // which can happen here if TailBB is unanalyzable and is merged, so 1393 // explicitly remove BBI1 and BBI2 as successors. 1394 BBI.BB->removeSuccessor(BBI1->BB); 1395 BBI.BB->removeSuccessor(BBI2->BB); 1396 RemoveExtraEdges(BBI); 1397 1398 // Update block info. 1399 BBI.IsDone = TrueBBI.IsDone = FalseBBI.IsDone = true; 1400 InvalidatePreds(BBI.BB); 1401 1402 // FIXME: Must maintain LiveIns. 1403 return true; 1404 } 1405 1406 static bool MaySpeculate(const MachineInstr *MI, 1407 SmallSet<unsigned, 4> &LaterRedefs, 1408 const TargetInstrInfo *TII) { 1409 bool SawStore = true; 1410 if (!MI->isSafeToMove(TII, 0, SawStore)) 1411 return false; 1412 1413 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) { 1414 const MachineOperand &MO = MI->getOperand(i); 1415 if (!MO.isReg()) 1416 continue; 1417 unsigned Reg = MO.getReg(); 1418 if (!Reg) 1419 continue; 1420 if (MO.isDef() && !LaterRedefs.count(Reg)) 1421 return false; 1422 } 1423 1424 return true; 1425 } 1426 1427 /// PredicateBlock - Predicate instructions from the start of the block to the 1428 /// specified end with the specified condition. 1429 void IfConverter::PredicateBlock(BBInfo &BBI, 1430 MachineBasicBlock::iterator E, 1431 SmallVectorImpl<MachineOperand> &Cond, 1432 SmallSet<unsigned, 4> &Redefs, 1433 SmallSet<unsigned, 4> *LaterRedefs) { 1434 bool AnyUnpred = false; 1435 bool MaySpec = LaterRedefs != 0; 1436 for (MachineBasicBlock::iterator I = BBI.BB->begin(); I != E; ++I) { 1437 if (I->isDebugValue() || TII->isPredicated(I)) 1438 continue; 1439 // It may be possible not to predicate an instruction if it's the 'true' 1440 // side of a diamond and the 'false' side may re-define the instruction's 1441 // defs. 1442 if (MaySpec && MaySpeculate(I, *LaterRedefs, TII)) { 1443 AnyUnpred = true; 1444 continue; 1445 } 1446 // If any instruction is predicated, then every instruction after it must 1447 // be predicated. 1448 MaySpec = false; 1449 if (!TII->PredicateInstruction(I, Cond)) { 1450 #ifndef NDEBUG 1451 dbgs() << "Unable to predicate " << *I << "!\n"; 1452 #endif 1453 llvm_unreachable(0); 1454 } 1455 1456 // If the predicated instruction now redefines a register as the result of 1457 // if-conversion, add an implicit kill. 1458 UpdatePredRedefs(I, Redefs, TRI, true); 1459 } 1460 1461 std::copy(Cond.begin(), Cond.end(), std::back_inserter(BBI.Predicate)); 1462 1463 BBI.IsAnalyzed = false; 1464 BBI.NonPredSize = 0; 1465 1466 ++NumIfConvBBs; 1467 if (AnyUnpred) 1468 ++NumUnpred; 1469 } 1470 1471 /// CopyAndPredicateBlock - Copy and predicate instructions from source BB to 1472 /// the destination block. Skip end of block branches if IgnoreBr is true. 1473 void IfConverter::CopyAndPredicateBlock(BBInfo &ToBBI, BBInfo &FromBBI, 1474 SmallVectorImpl<MachineOperand> &Cond, 1475 SmallSet<unsigned, 4> &Redefs, 1476 bool IgnoreBr) { 1477 MachineFunction &MF = *ToBBI.BB->getParent(); 1478 1479 for (MachineBasicBlock::iterator I = FromBBI.BB->begin(), 1480 E = FromBBI.BB->end(); I != E; ++I) { 1481 // Do not copy the end of the block branches. 1482 if (IgnoreBr && I->isBranch()) 1483 break; 1484 1485 MachineInstr *MI = MF.CloneMachineInstr(I); 1486 ToBBI.BB->insert(ToBBI.BB->end(), MI); 1487 ToBBI.NonPredSize++; 1488 unsigned ExtraPredCost = 0; 1489 unsigned NumCycles = TII->getInstrLatency(InstrItins, &*I, &ExtraPredCost); 1490 if (NumCycles > 1) 1491 ToBBI.ExtraCost += NumCycles-1; 1492 ToBBI.ExtraCost2 += ExtraPredCost; 1493 1494 if (!TII->isPredicated(I) && !MI->isDebugValue()) { 1495 if (!TII->PredicateInstruction(MI, Cond)) { 1496 #ifndef NDEBUG 1497 dbgs() << "Unable to predicate " << *I << "!\n"; 1498 #endif 1499 llvm_unreachable(0); 1500 } 1501 } 1502 1503 // If the predicated instruction now redefines a register as the result of 1504 // if-conversion, add an implicit kill. 1505 UpdatePredRedefs(MI, Redefs, TRI, true); 1506 } 1507 1508 if (!IgnoreBr) { 1509 std::vector<MachineBasicBlock *> Succs(FromBBI.BB->succ_begin(), 1510 FromBBI.BB->succ_end()); 1511 MachineBasicBlock *NBB = getNextBlock(FromBBI.BB); 1512 MachineBasicBlock *FallThrough = FromBBI.HasFallThrough ? NBB : NULL; 1513 1514 for (unsigned i = 0, e = Succs.size(); i != e; ++i) { 1515 MachineBasicBlock *Succ = Succs[i]; 1516 // Fallthrough edge can't be transferred. 1517 if (Succ == FallThrough) 1518 continue; 1519 ToBBI.BB->addSuccessor(Succ); 1520 } 1521 } 1522 1523 std::copy(FromBBI.Predicate.begin(), FromBBI.Predicate.end(), 1524 std::back_inserter(ToBBI.Predicate)); 1525 std::copy(Cond.begin(), Cond.end(), std::back_inserter(ToBBI.Predicate)); 1526 1527 ToBBI.ClobbersPred |= FromBBI.ClobbersPred; 1528 ToBBI.IsAnalyzed = false; 1529 1530 ++NumDupBBs; 1531 } 1532 1533 /// MergeBlocks - Move all instructions from FromBB to the end of ToBB. 1534 /// This will leave FromBB as an empty block, so remove all of its 1535 /// successor edges except for the fall-through edge. If AddEdges is true, 1536 /// i.e., when FromBBI's branch is being moved, add those successor edges to 1537 /// ToBBI. 1538 void IfConverter::MergeBlocks(BBInfo &ToBBI, BBInfo &FromBBI, bool AddEdges) { 1539 ToBBI.BB->splice(ToBBI.BB->end(), 1540 FromBBI.BB, FromBBI.BB->begin(), FromBBI.BB->end()); 1541 1542 std::vector<MachineBasicBlock *> Succs(FromBBI.BB->succ_begin(), 1543 FromBBI.BB->succ_end()); 1544 MachineBasicBlock *NBB = getNextBlock(FromBBI.BB); 1545 MachineBasicBlock *FallThrough = FromBBI.HasFallThrough ? NBB : NULL; 1546 1547 for (unsigned i = 0, e = Succs.size(); i != e; ++i) { 1548 MachineBasicBlock *Succ = Succs[i]; 1549 // Fallthrough edge can't be transferred. 1550 if (Succ == FallThrough) 1551 continue; 1552 FromBBI.BB->removeSuccessor(Succ); 1553 if (AddEdges) 1554 ToBBI.BB->addSuccessor(Succ); 1555 } 1556 1557 // Now FromBBI always falls through to the next block! 1558 if (NBB && !FromBBI.BB->isSuccessor(NBB)) 1559 FromBBI.BB->addSuccessor(NBB); 1560 1561 std::copy(FromBBI.Predicate.begin(), FromBBI.Predicate.end(), 1562 std::back_inserter(ToBBI.Predicate)); 1563 FromBBI.Predicate.clear(); 1564 1565 ToBBI.NonPredSize += FromBBI.NonPredSize; 1566 ToBBI.ExtraCost += FromBBI.ExtraCost; 1567 ToBBI.ExtraCost2 += FromBBI.ExtraCost2; 1568 FromBBI.NonPredSize = 0; 1569 FromBBI.ExtraCost = 0; 1570 FromBBI.ExtraCost2 = 0; 1571 1572 ToBBI.ClobbersPred |= FromBBI.ClobbersPred; 1573 ToBBI.HasFallThrough = FromBBI.HasFallThrough; 1574 ToBBI.IsAnalyzed = false; 1575 FromBBI.IsAnalyzed = false; 1576 } 1577