IfConversion.cpp revision f81b7f6069b27c0a515070dcb392f6828437412f
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/MachineFunctionPass.h" 20#include "llvm/CodeGen/MachineLoopInfo.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/BranchProbability.h" 27#include "llvm/Support/CommandLine.h" 28#include "llvm/Support/Debug.h" 29#include "llvm/Support/ErrorHandling.h" 30#include "llvm/Support/raw_ostream.h" 31#include "llvm/ADT/SmallSet.h" 32#include "llvm/ADT/Statistic.h" 33#include "llvm/ADT/STLExtras.h" 34using namespace llvm; 35 36// Hidden options for help debugging. 37static cl::opt<int> IfCvtFnStart("ifcvt-fn-start", cl::init(-1), cl::Hidden); 38static cl::opt<int> IfCvtFnStop("ifcvt-fn-stop", cl::init(-1), cl::Hidden); 39static cl::opt<int> IfCvtLimit("ifcvt-limit", cl::init(-1), cl::Hidden); 40static cl::opt<bool> DisableSimple("disable-ifcvt-simple", 41 cl::init(false), cl::Hidden); 42static cl::opt<bool> DisableSimpleF("disable-ifcvt-simple-false", 43 cl::init(false), cl::Hidden); 44static cl::opt<bool> DisableTriangle("disable-ifcvt-triangle", 45 cl::init(false), cl::Hidden); 46static cl::opt<bool> DisableTriangleR("disable-ifcvt-triangle-rev", 47 cl::init(false), cl::Hidden); 48static cl::opt<bool> DisableTriangleF("disable-ifcvt-triangle-false", 49 cl::init(false), cl::Hidden); 50static cl::opt<bool> DisableTriangleFR("disable-ifcvt-triangle-false-rev", 51 cl::init(false), cl::Hidden); 52static cl::opt<bool> DisableDiamond("disable-ifcvt-diamond", 53 cl::init(false), cl::Hidden); 54static cl::opt<bool> IfCvtBranchFold("ifcvt-branch-fold", 55 cl::init(true), cl::Hidden); 56 57STATISTIC(NumSimple, "Number of simple if-conversions performed"); 58STATISTIC(NumSimpleFalse, "Number of simple (F) if-conversions performed"); 59STATISTIC(NumTriangle, "Number of triangle if-conversions performed"); 60STATISTIC(NumTriangleRev, "Number of triangle (R) if-conversions performed"); 61STATISTIC(NumTriangleFalse,"Number of triangle (F) if-conversions performed"); 62STATISTIC(NumTriangleFRev, "Number of triangle (F/R) if-conversions performed"); 63STATISTIC(NumDiamonds, "Number of diamond if-conversions performed"); 64STATISTIC(NumIfConvBBs, "Number of if-converted blocks"); 65STATISTIC(NumDupBBs, "Number of duplicated blocks"); 66 67namespace { 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 MachineLoopInfo *MLI; 158 bool MadeChange; 159 int FnNum; 160 public: 161 static char ID; 162 IfConverter() : MachineFunctionPass(ID), FnNum(-1) { 163 initializeIfConverterPass(*PassRegistry::getPassRegistry()); 164 } 165 166 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 167 AU.addRequired<MachineLoopInfo>(); 168 MachineFunctionPass::getAnalysisUsage(AU); 169 } 170 171 virtual bool runOnMachineFunction(MachineFunction &MF); 172 virtual const char *getPassName() const { return "If Converter"; } 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 void CopyAndPredicateBlock(BBInfo &ToBBI, BBInfo &FromBBI, 200 SmallVectorImpl<MachineOperand> &Cond, 201 SmallSet<unsigned, 4> &Redefs, 202 bool IgnoreBr = false); 203 void MergeBlocks(BBInfo &ToBBI, BBInfo &FromBBI, bool AddEdges = true); 204 205 bool MeetIfcvtSizeLimit(MachineBasicBlock &BB, 206 unsigned Cycle, unsigned Extra, 207 const BranchProbability &Prediction) const { 208 return Cycle > 0 && TII->isProfitableToIfCvt(BB, Cycle, Extra, 209 Prediction); 210 } 211 212 bool MeetIfcvtSizeLimit(MachineBasicBlock &TBB, 213 unsigned TCycle, unsigned TExtra, 214 MachineBasicBlock &FBB, 215 unsigned FCycle, unsigned FExtra, 216 const BranchProbability &Prediction) const { 217 return TCycle > 0 && FCycle > 0 && 218 TII->isProfitableToIfCvt(TBB, TCycle, TExtra, FBB, FCycle, FExtra, 219 Prediction); 220 } 221 222 // blockAlwaysFallThrough - Block ends without a terminator. 223 bool blockAlwaysFallThrough(BBInfo &BBI) const { 224 return BBI.IsBrAnalyzable && BBI.TrueBB == NULL; 225 } 226 227 // IfcvtTokenCmp - Used to sort if-conversion candidates. 228 static bool IfcvtTokenCmp(IfcvtToken *C1, IfcvtToken *C2) { 229 int Incr1 = (C1->Kind == ICDiamond) 230 ? -(int)(C1->NumDups + C1->NumDups2) : (int)C1->NumDups; 231 int Incr2 = (C2->Kind == ICDiamond) 232 ? -(int)(C2->NumDups + C2->NumDups2) : (int)C2->NumDups; 233 if (Incr1 > Incr2) 234 return true; 235 else if (Incr1 == Incr2) { 236 // Favors subsumption. 237 if (C1->NeedSubsumption == false && C2->NeedSubsumption == true) 238 return true; 239 else if (C1->NeedSubsumption == C2->NeedSubsumption) { 240 // Favors diamond over triangle, etc. 241 if ((unsigned)C1->Kind < (unsigned)C2->Kind) 242 return true; 243 else if (C1->Kind == C2->Kind) 244 return C1->BBI.BB->getNumber() < C2->BBI.BB->getNumber(); 245 } 246 } 247 return false; 248 } 249 }; 250 251 char IfConverter::ID = 0; 252} 253 254INITIALIZE_PASS_BEGIN(IfConverter, "if-converter", "If Converter", false, false) 255INITIALIZE_PASS_DEPENDENCY(MachineLoopInfo) 256INITIALIZE_PASS_END(IfConverter, "if-converter", "If Converter", false, false) 257 258FunctionPass *llvm::createIfConverterPass() { return new IfConverter(); } 259 260bool IfConverter::runOnMachineFunction(MachineFunction &MF) { 261 TLI = MF.getTarget().getTargetLowering(); 262 TII = MF.getTarget().getInstrInfo(); 263 TRI = MF.getTarget().getRegisterInfo(); 264 MLI = &getAnalysis<MachineLoopInfo>(); 265 InstrItins = MF.getTarget().getInstrItineraryData(); 266 if (!TII) return false; 267 268 // Tail merge tend to expose more if-conversion opportunities. 269 BranchFolder BF(true, false); 270 bool BFChange = BF.OptimizeFunction(MF, TII, 271 MF.getTarget().getRegisterInfo(), 272 getAnalysisIfAvailable<MachineModuleInfo>()); 273 274 DEBUG(dbgs() << "\nIfcvt: function (" << ++FnNum << ") \'" 275 << MF.getFunction()->getName() << "\'"); 276 277 if (FnNum < IfCvtFnStart || (IfCvtFnStop != -1 && FnNum > IfCvtFnStop)) { 278 DEBUG(dbgs() << " skipped\n"); 279 return false; 280 } 281 DEBUG(dbgs() << "\n"); 282 283 MF.RenumberBlocks(); 284 BBAnalysis.resize(MF.getNumBlockIDs()); 285 286 std::vector<IfcvtToken*> Tokens; 287 MadeChange = false; 288 unsigned NumIfCvts = NumSimple + NumSimpleFalse + NumTriangle + 289 NumTriangleRev + NumTriangleFalse + NumTriangleFRev + NumDiamonds; 290 while (IfCvtLimit == -1 || (int)NumIfCvts < IfCvtLimit) { 291 // Do an initial analysis for each basic block and find all the potential 292 // candidates to perform if-conversion. 293 bool Change = false; 294 AnalyzeBlocks(MF, Tokens); 295 while (!Tokens.empty()) { 296 IfcvtToken *Token = Tokens.back(); 297 Tokens.pop_back(); 298 BBInfo &BBI = Token->BBI; 299 IfcvtKind Kind = Token->Kind; 300 unsigned NumDups = Token->NumDups; 301 unsigned NumDups2 = Token->NumDups2; 302 303 delete Token; 304 305 // If the block has been evicted out of the queue or it has already been 306 // marked dead (due to it being predicated), then skip it. 307 if (BBI.IsDone) 308 BBI.IsEnqueued = false; 309 if (!BBI.IsEnqueued) 310 continue; 311 312 BBI.IsEnqueued = false; 313 314 bool RetVal = false; 315 switch (Kind) { 316 default: assert(false && "Unexpected!"); 317 break; 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. 415static 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. 428bool 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. 441static 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. 453bool 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'. 479bool 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. 520bool 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->getDesc().isBranch()) 577 break; 578 } 579 while (FIE != FIB) { 580 --FIE; 581 if (!FIE->getDesc().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. 620void 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 const MCInstrDesc &MCID = I->getDesc(); 655 if (MCID.isNotDuplicable()) 656 BBI.CannotBeCopied = true; 657 658 bool isPredicated = TII->isPredicated(I); 659 bool isCondBr = BBI.IsBrAnalyzable && MCID.isConditionalBranch(); 660 661 if (!isCondBr) { 662 if (!isPredicated) { 663 BBI.NonPredSize++; 664 unsigned ExtraPredCost = 0; 665 unsigned NumCycles = TII->getInstrLatency(InstrItins, &*I, 666 &ExtraPredCost); 667 if (NumCycles > 1) 668 BBI.ExtraCost += NumCycles-1; 669 BBI.ExtraCost2 += ExtraPredCost; 670 } else if (!AlreadyPredicated) { 671 // FIXME: This instruction is already predicated before the 672 // if-conversion pass. It's probably something like a conditional move. 673 // Mark this block unpredicable for now. 674 BBI.IsUnpredicable = true; 675 return; 676 } 677 } 678 679 if (BBI.ClobbersPred && !isPredicated) { 680 // Predicate modification instruction should end the block (except for 681 // already predicated instructions and end of block branches). 682 if (isCondBr) { 683 // A conditional branch is not predicable, but it may be eliminated. 684 continue; 685 } 686 687 // Predicate may have been modified, the subsequent (currently) 688 // unpredicated instructions cannot be correctly predicated. 689 BBI.IsUnpredicable = true; 690 return; 691 } 692 693 // FIXME: Make use of PredDefs? e.g. ADDC, SUBC sets predicates but are 694 // still potentially predicable. 695 std::vector<MachineOperand> PredDefs; 696 if (TII->DefinesPredicate(I, PredDefs)) 697 BBI.ClobbersPred = true; 698 699 if (!TII->isPredicable(I)) { 700 BBI.IsUnpredicable = true; 701 return; 702 } 703 } 704} 705 706/// FeasibilityAnalysis - Determine if the block is a suitable candidate to be 707/// predicated by the specified predicate. 708bool IfConverter::FeasibilityAnalysis(BBInfo &BBI, 709 SmallVectorImpl<MachineOperand> &Pred, 710 bool isTriangle, bool RevBranch) { 711 // If the block is dead or unpredicable, then it cannot be predicated. 712 if (BBI.IsDone || BBI.IsUnpredicable) 713 return false; 714 715 // If it is already predicated, check if its predicate subsumes the new 716 // predicate. 717 if (BBI.Predicate.size() && !TII->SubsumesPredicate(BBI.Predicate, Pred)) 718 return false; 719 720 if (BBI.BrCond.size()) { 721 if (!isTriangle) 722 return false; 723 724 // Test predicate subsumption. 725 SmallVector<MachineOperand, 4> RevPred(Pred.begin(), Pred.end()); 726 SmallVector<MachineOperand, 4> Cond(BBI.BrCond.begin(), BBI.BrCond.end()); 727 if (RevBranch) { 728 if (TII->ReverseBranchCondition(Cond)) 729 return false; 730 } 731 if (TII->ReverseBranchCondition(RevPred) || 732 !TII->SubsumesPredicate(Cond, RevPred)) 733 return false; 734 } 735 736 return true; 737} 738 739/// AnalyzeBlock - Analyze the structure of the sub-CFG starting from 740/// the specified block. Record its successors and whether it looks like an 741/// if-conversion candidate. 742IfConverter::BBInfo &IfConverter::AnalyzeBlock(MachineBasicBlock *BB, 743 std::vector<IfcvtToken*> &Tokens) { 744 BBInfo &BBI = BBAnalysis[BB->getNumber()]; 745 746 if (BBI.IsAnalyzed || BBI.IsBeingAnalyzed) 747 return BBI; 748 749 BBI.BB = BB; 750 BBI.IsBeingAnalyzed = true; 751 752 ScanInstructions(BBI); 753 754 // Unanalyzable or ends with fallthrough or unconditional branch, or if is not 755 // considered for ifcvt anymore. 756 if (!BBI.IsBrAnalyzable || BBI.BrCond.empty() || BBI.IsDone) { 757 BBI.IsBeingAnalyzed = false; 758 BBI.IsAnalyzed = true; 759 return BBI; 760 } 761 762 // Do not ifcvt if either path is a back edge to the entry block. 763 if (BBI.TrueBB == BB || BBI.FalseBB == BB) { 764 BBI.IsBeingAnalyzed = false; 765 BBI.IsAnalyzed = true; 766 return BBI; 767 } 768 769 // Do not ifcvt if true and false fallthrough blocks are the same. 770 if (!BBI.FalseBB) { 771 BBI.IsBeingAnalyzed = false; 772 BBI.IsAnalyzed = true; 773 return BBI; 774 } 775 776 BBInfo &TrueBBI = AnalyzeBlock(BBI.TrueBB, Tokens); 777 BBInfo &FalseBBI = AnalyzeBlock(BBI.FalseBB, Tokens); 778 779 if (TrueBBI.IsDone && FalseBBI.IsDone) { 780 BBI.IsBeingAnalyzed = false; 781 BBI.IsAnalyzed = true; 782 return BBI; 783 } 784 785 SmallVector<MachineOperand, 4> RevCond(BBI.BrCond.begin(), BBI.BrCond.end()); 786 bool CanRevCond = !TII->ReverseBranchCondition(RevCond); 787 788 unsigned Dups = 0; 789 unsigned Dups2 = 0; 790 bool TNeedSub = TrueBBI.Predicate.size() > 0; 791 bool FNeedSub = FalseBBI.Predicate.size() > 0; 792 bool Enqueued = false; 793 794 // Try to predict the branch, using loop info to guide us. 795 // General heuristics are: 796 // - backedge -> 90% taken 797 // - early exit -> 20% taken 798 // - branch predictor confidence -> 90% 799 BranchProbability Prediction(5, 10); 800 MachineLoop *Loop = MLI->getLoopFor(BB); 801 if (Loop) { 802 if (TrueBBI.BB == Loop->getHeader()) 803 Prediction = BranchProbability(9, 10); 804 else if (FalseBBI.BB == Loop->getHeader()) 805 Prediction = BranchProbability(1, 10); 806 807 MachineLoop *TrueLoop = MLI->getLoopFor(TrueBBI.BB); 808 MachineLoop *FalseLoop = MLI->getLoopFor(FalseBBI.BB); 809 if (!TrueLoop || TrueLoop->getParentLoop() == Loop) 810 Prediction = BranchProbability(2, 10); 811 else if (!FalseLoop || FalseLoop->getParentLoop() == Loop) 812 Prediction = BranchProbability(8, 10); 813 } 814 815 if (CanRevCond && ValidDiamond(TrueBBI, FalseBBI, Dups, Dups2) && 816 MeetIfcvtSizeLimit(*TrueBBI.BB, (TrueBBI.NonPredSize - (Dups + Dups2) + 817 TrueBBI.ExtraCost), TrueBBI.ExtraCost2, 818 *FalseBBI.BB, (FalseBBI.NonPredSize - (Dups + Dups2) + 819 FalseBBI.ExtraCost),FalseBBI.ExtraCost2, 820 Prediction) && 821 FeasibilityAnalysis(TrueBBI, BBI.BrCond) && 822 FeasibilityAnalysis(FalseBBI, RevCond)) { 823 // Diamond: 824 // EBB 825 // / \_ 826 // | | 827 // TBB FBB 828 // \ / 829 // TailBB 830 // Note TailBB can be empty. 831 Tokens.push_back(new IfcvtToken(BBI, ICDiamond, TNeedSub|FNeedSub, Dups, 832 Dups2)); 833 Enqueued = true; 834 } 835 836 if (ValidTriangle(TrueBBI, FalseBBI, false, Dups, Prediction) && 837 MeetIfcvtSizeLimit(*TrueBBI.BB, TrueBBI.NonPredSize + TrueBBI.ExtraCost, 838 TrueBBI.ExtraCost2, Prediction) && 839 FeasibilityAnalysis(TrueBBI, BBI.BrCond, true)) { 840 // Triangle: 841 // EBB 842 // | \_ 843 // | | 844 // | TBB 845 // | / 846 // FBB 847 Tokens.push_back(new IfcvtToken(BBI, ICTriangle, TNeedSub, Dups)); 848 Enqueued = true; 849 } 850 851 if (ValidTriangle(TrueBBI, FalseBBI, true, Dups, Prediction) && 852 MeetIfcvtSizeLimit(*TrueBBI.BB, TrueBBI.NonPredSize + TrueBBI.ExtraCost, 853 TrueBBI.ExtraCost2, Prediction) && 854 FeasibilityAnalysis(TrueBBI, BBI.BrCond, true, true)) { 855 Tokens.push_back(new IfcvtToken(BBI, ICTriangleRev, TNeedSub, Dups)); 856 Enqueued = true; 857 } 858 859 if (ValidSimple(TrueBBI, Dups, Prediction) && 860 MeetIfcvtSizeLimit(*TrueBBI.BB, TrueBBI.NonPredSize + TrueBBI.ExtraCost, 861 TrueBBI.ExtraCost2, Prediction) && 862 FeasibilityAnalysis(TrueBBI, BBI.BrCond)) { 863 // Simple (split, no rejoin): 864 // EBB 865 // | \_ 866 // | | 867 // | TBB---> exit 868 // | 869 // FBB 870 Tokens.push_back(new IfcvtToken(BBI, ICSimple, TNeedSub, Dups)); 871 Enqueued = true; 872 } 873 874 if (CanRevCond) { 875 // Try the other path... 876 if (ValidTriangle(FalseBBI, TrueBBI, false, Dups, 877 Prediction.getCompl()) && 878 MeetIfcvtSizeLimit(*FalseBBI.BB, 879 FalseBBI.NonPredSize + FalseBBI.ExtraCost, 880 FalseBBI.ExtraCost2, Prediction.getCompl()) && 881 FeasibilityAnalysis(FalseBBI, RevCond, true)) { 882 Tokens.push_back(new IfcvtToken(BBI, ICTriangleFalse, FNeedSub, Dups)); 883 Enqueued = true; 884 } 885 886 if (ValidTriangle(FalseBBI, TrueBBI, true, Dups, 887 Prediction.getCompl()) && 888 MeetIfcvtSizeLimit(*FalseBBI.BB, 889 FalseBBI.NonPredSize + FalseBBI.ExtraCost, 890 FalseBBI.ExtraCost2, Prediction.getCompl()) && 891 FeasibilityAnalysis(FalseBBI, RevCond, true, true)) { 892 Tokens.push_back(new IfcvtToken(BBI, ICTriangleFRev, FNeedSub, Dups)); 893 Enqueued = true; 894 } 895 896 if (ValidSimple(FalseBBI, Dups, Prediction.getCompl()) && 897 MeetIfcvtSizeLimit(*FalseBBI.BB, 898 FalseBBI.NonPredSize + FalseBBI.ExtraCost, 899 FalseBBI.ExtraCost2, Prediction.getCompl()) && 900 FeasibilityAnalysis(FalseBBI, RevCond)) { 901 Tokens.push_back(new IfcvtToken(BBI, ICSimpleFalse, FNeedSub, Dups)); 902 Enqueued = true; 903 } 904 } 905 906 BBI.IsEnqueued = Enqueued; 907 BBI.IsBeingAnalyzed = false; 908 BBI.IsAnalyzed = true; 909 return BBI; 910} 911 912/// AnalyzeBlocks - Analyze all blocks and find entries for all if-conversion 913/// candidates. 914void IfConverter::AnalyzeBlocks(MachineFunction &MF, 915 std::vector<IfcvtToken*> &Tokens) { 916 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; ++I) { 917 MachineBasicBlock *BB = I; 918 AnalyzeBlock(BB, Tokens); 919 } 920 921 // Sort to favor more complex ifcvt scheme. 922 std::stable_sort(Tokens.begin(), Tokens.end(), IfcvtTokenCmp); 923} 924 925/// canFallThroughTo - Returns true either if ToBB is the next block after BB or 926/// that all the intervening blocks are empty (given BB can fall through to its 927/// next block). 928static bool canFallThroughTo(MachineBasicBlock *BB, MachineBasicBlock *ToBB) { 929 MachineFunction::iterator PI = BB; 930 MachineFunction::iterator I = llvm::next(PI); 931 MachineFunction::iterator TI = ToBB; 932 MachineFunction::iterator E = BB->getParent()->end(); 933 while (I != TI) { 934 // Check isSuccessor to avoid case where the next block is empty, but 935 // it's not a successor. 936 if (I == E || !I->empty() || !PI->isSuccessor(I)) 937 return false; 938 PI = I++; 939 } 940 return true; 941} 942 943/// InvalidatePreds - Invalidate predecessor BB info so it would be re-analyzed 944/// to determine if it can be if-converted. If predecessor is already enqueued, 945/// dequeue it! 946void IfConverter::InvalidatePreds(MachineBasicBlock *BB) { 947 for (MachineBasicBlock::pred_iterator PI = BB->pred_begin(), 948 E = BB->pred_end(); PI != E; ++PI) { 949 BBInfo &PBBI = BBAnalysis[(*PI)->getNumber()]; 950 if (PBBI.IsDone || PBBI.BB == BB) 951 continue; 952 PBBI.IsAnalyzed = false; 953 PBBI.IsEnqueued = false; 954 } 955} 956 957/// InsertUncondBranch - Inserts an unconditional branch from BB to ToBB. 958/// 959static void InsertUncondBranch(MachineBasicBlock *BB, MachineBasicBlock *ToBB, 960 const TargetInstrInfo *TII) { 961 DebugLoc dl; // FIXME: this is nowhere 962 SmallVector<MachineOperand, 0> NoCond; 963 TII->InsertBranch(*BB, ToBB, NULL, NoCond, dl); 964} 965 966/// RemoveExtraEdges - Remove true / false edges if either / both are no longer 967/// successors. 968void IfConverter::RemoveExtraEdges(BBInfo &BBI) { 969 MachineBasicBlock *TBB = NULL, *FBB = NULL; 970 SmallVector<MachineOperand, 4> Cond; 971 if (!TII->AnalyzeBranch(*BBI.BB, TBB, FBB, Cond)) 972 BBI.BB->CorrectExtraCFGEdges(TBB, FBB, !Cond.empty()); 973} 974 975/// InitPredRedefs / UpdatePredRedefs - Defs by predicated instructions are 976/// modeled as read + write (sort like two-address instructions). These 977/// routines track register liveness and add implicit uses to if-converted 978/// instructions to conform to the model. 979static void InitPredRedefs(MachineBasicBlock *BB, SmallSet<unsigned,4> &Redefs, 980 const TargetRegisterInfo *TRI) { 981 for (MachineBasicBlock::livein_iterator I = BB->livein_begin(), 982 E = BB->livein_end(); I != E; ++I) { 983 unsigned Reg = *I; 984 Redefs.insert(Reg); 985 for (const unsigned *Subreg = TRI->getSubRegisters(Reg); 986 *Subreg; ++Subreg) 987 Redefs.insert(*Subreg); 988 } 989} 990 991static void UpdatePredRedefs(MachineInstr *MI, SmallSet<unsigned,4> &Redefs, 992 const TargetRegisterInfo *TRI, 993 bool AddImpUse = false) { 994 SmallVector<unsigned, 4> Defs; 995 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) { 996 const MachineOperand &MO = MI->getOperand(i); 997 if (!MO.isReg()) 998 continue; 999 unsigned Reg = MO.getReg(); 1000 if (!Reg) 1001 continue; 1002 if (MO.isDef()) 1003 Defs.push_back(Reg); 1004 else if (MO.isKill()) { 1005 Redefs.erase(Reg); 1006 for (const unsigned *SR = TRI->getSubRegisters(Reg); *SR; ++SR) 1007 Redefs.erase(*SR); 1008 } 1009 } 1010 for (unsigned i = 0, e = Defs.size(); i != e; ++i) { 1011 unsigned Reg = Defs[i]; 1012 if (Redefs.count(Reg)) { 1013 if (AddImpUse) 1014 // Treat predicated update as read + write. 1015 MI->addOperand(MachineOperand::CreateReg(Reg, false/*IsDef*/, 1016 true/*IsImp*/,false/*IsKill*/)); 1017 } else { 1018 Redefs.insert(Reg); 1019 for (const unsigned *SR = TRI->getSubRegisters(Reg); *SR; ++SR) 1020 Redefs.insert(*SR); 1021 } 1022 } 1023} 1024 1025static void UpdatePredRedefs(MachineBasicBlock::iterator I, 1026 MachineBasicBlock::iterator E, 1027 SmallSet<unsigned,4> &Redefs, 1028 const TargetRegisterInfo *TRI) { 1029 while (I != E) { 1030 UpdatePredRedefs(I, Redefs, TRI); 1031 ++I; 1032 } 1033} 1034 1035/// IfConvertSimple - If convert a simple (split, no rejoin) sub-CFG. 1036/// 1037bool IfConverter::IfConvertSimple(BBInfo &BBI, IfcvtKind Kind) { 1038 BBInfo &TrueBBI = BBAnalysis[BBI.TrueBB->getNumber()]; 1039 BBInfo &FalseBBI = BBAnalysis[BBI.FalseBB->getNumber()]; 1040 BBInfo *CvtBBI = &TrueBBI; 1041 BBInfo *NextBBI = &FalseBBI; 1042 1043 SmallVector<MachineOperand, 4> Cond(BBI.BrCond.begin(), BBI.BrCond.end()); 1044 if (Kind == ICSimpleFalse) 1045 std::swap(CvtBBI, NextBBI); 1046 1047 if (CvtBBI->IsDone || 1048 (CvtBBI->CannotBeCopied && CvtBBI->BB->pred_size() > 1)) { 1049 // Something has changed. It's no longer safe to predicate this block. 1050 BBI.IsAnalyzed = false; 1051 CvtBBI->IsAnalyzed = false; 1052 return false; 1053 } 1054 1055 if (Kind == ICSimpleFalse) 1056 if (TII->ReverseBranchCondition(Cond)) 1057 assert(false && "Unable to reverse branch condition!"); 1058 1059 // Initialize liveins to the first BB. These are potentiall redefined by 1060 // predicated instructions. 1061 SmallSet<unsigned, 4> Redefs; 1062 InitPredRedefs(CvtBBI->BB, Redefs, TRI); 1063 InitPredRedefs(NextBBI->BB, Redefs, TRI); 1064 1065 if (CvtBBI->BB->pred_size() > 1) { 1066 BBI.NonPredSize -= TII->RemoveBranch(*BBI.BB); 1067 // Copy instructions in the true block, predicate them, and add them to 1068 // the entry block. 1069 CopyAndPredicateBlock(BBI, *CvtBBI, Cond, Redefs); 1070 } else { 1071 PredicateBlock(*CvtBBI, CvtBBI->BB->end(), Cond, Redefs); 1072 1073 // Merge converted block into entry block. 1074 BBI.NonPredSize -= TII->RemoveBranch(*BBI.BB); 1075 MergeBlocks(BBI, *CvtBBI); 1076 } 1077 1078 bool IterIfcvt = true; 1079 if (!canFallThroughTo(BBI.BB, NextBBI->BB)) { 1080 InsertUncondBranch(BBI.BB, NextBBI->BB, TII); 1081 BBI.HasFallThrough = false; 1082 // Now ifcvt'd block will look like this: 1083 // BB: 1084 // ... 1085 // t, f = cmp 1086 // if t op 1087 // b BBf 1088 // 1089 // We cannot further ifcvt this block because the unconditional branch 1090 // will have to be predicated on the new condition, that will not be 1091 // available if cmp executes. 1092 IterIfcvt = false; 1093 } 1094 1095 RemoveExtraEdges(BBI); 1096 1097 // Update block info. BB can be iteratively if-converted. 1098 if (!IterIfcvt) 1099 BBI.IsDone = true; 1100 InvalidatePreds(BBI.BB); 1101 CvtBBI->IsDone = true; 1102 1103 // FIXME: Must maintain LiveIns. 1104 return true; 1105} 1106 1107/// IfConvertTriangle - If convert a triangle sub-CFG. 1108/// 1109bool IfConverter::IfConvertTriangle(BBInfo &BBI, IfcvtKind Kind) { 1110 BBInfo &TrueBBI = BBAnalysis[BBI.TrueBB->getNumber()]; 1111 BBInfo &FalseBBI = BBAnalysis[BBI.FalseBB->getNumber()]; 1112 BBInfo *CvtBBI = &TrueBBI; 1113 BBInfo *NextBBI = &FalseBBI; 1114 DebugLoc dl; // FIXME: this is nowhere 1115 1116 SmallVector<MachineOperand, 4> Cond(BBI.BrCond.begin(), BBI.BrCond.end()); 1117 if (Kind == ICTriangleFalse || Kind == ICTriangleFRev) 1118 std::swap(CvtBBI, NextBBI); 1119 1120 if (CvtBBI->IsDone || 1121 (CvtBBI->CannotBeCopied && CvtBBI->BB->pred_size() > 1)) { 1122 // Something has changed. It's no longer safe to predicate this block. 1123 BBI.IsAnalyzed = false; 1124 CvtBBI->IsAnalyzed = false; 1125 return false; 1126 } 1127 1128 if (Kind == ICTriangleFalse || Kind == ICTriangleFRev) 1129 if (TII->ReverseBranchCondition(Cond)) 1130 assert(false && "Unable to reverse branch condition!"); 1131 1132 if (Kind == ICTriangleRev || Kind == ICTriangleFRev) { 1133 if (ReverseBranchCondition(*CvtBBI)) { 1134 // BB has been changed, modify its predecessors (except for this 1135 // one) so they don't get ifcvt'ed based on bad intel. 1136 for (MachineBasicBlock::pred_iterator PI = CvtBBI->BB->pred_begin(), 1137 E = CvtBBI->BB->pred_end(); PI != E; ++PI) { 1138 MachineBasicBlock *PBB = *PI; 1139 if (PBB == BBI.BB) 1140 continue; 1141 BBInfo &PBBI = BBAnalysis[PBB->getNumber()]; 1142 if (PBBI.IsEnqueued) { 1143 PBBI.IsAnalyzed = false; 1144 PBBI.IsEnqueued = false; 1145 } 1146 } 1147 } 1148 } 1149 1150 // Initialize liveins to the first BB. These are potentially redefined by 1151 // predicated instructions. 1152 SmallSet<unsigned, 4> Redefs; 1153 InitPredRedefs(CvtBBI->BB, Redefs, TRI); 1154 InitPredRedefs(NextBBI->BB, Redefs, TRI); 1155 1156 bool HasEarlyExit = CvtBBI->FalseBB != NULL; 1157 if (CvtBBI->BB->pred_size() > 1) { 1158 BBI.NonPredSize -= TII->RemoveBranch(*BBI.BB); 1159 // Copy instructions in the true block, predicate them, and add them to 1160 // the entry block. 1161 CopyAndPredicateBlock(BBI, *CvtBBI, Cond, Redefs, true); 1162 } else { 1163 // Predicate the 'true' block after removing its branch. 1164 CvtBBI->NonPredSize -= TII->RemoveBranch(*CvtBBI->BB); 1165 PredicateBlock(*CvtBBI, CvtBBI->BB->end(), Cond, Redefs); 1166 1167 // Now merge the entry of the triangle with the true block. 1168 BBI.NonPredSize -= TII->RemoveBranch(*BBI.BB); 1169 MergeBlocks(BBI, *CvtBBI, false); 1170 } 1171 1172 // If 'true' block has a 'false' successor, add an exit branch to it. 1173 if (HasEarlyExit) { 1174 SmallVector<MachineOperand, 4> RevCond(CvtBBI->BrCond.begin(), 1175 CvtBBI->BrCond.end()); 1176 if (TII->ReverseBranchCondition(RevCond)) 1177 assert(false && "Unable to reverse branch condition!"); 1178 TII->InsertBranch(*BBI.BB, CvtBBI->FalseBB, NULL, RevCond, dl); 1179 BBI.BB->addSuccessor(CvtBBI->FalseBB); 1180 } 1181 1182 // Merge in the 'false' block if the 'false' block has no other 1183 // predecessors. Otherwise, add an unconditional branch to 'false'. 1184 bool FalseBBDead = false; 1185 bool IterIfcvt = true; 1186 bool isFallThrough = canFallThroughTo(BBI.BB, NextBBI->BB); 1187 if (!isFallThrough) { 1188 // Only merge them if the true block does not fallthrough to the false 1189 // block. By not merging them, we make it possible to iteratively 1190 // ifcvt the blocks. 1191 if (!HasEarlyExit && 1192 NextBBI->BB->pred_size() == 1 && !NextBBI->HasFallThrough) { 1193 MergeBlocks(BBI, *NextBBI); 1194 FalseBBDead = true; 1195 } else { 1196 InsertUncondBranch(BBI.BB, NextBBI->BB, TII); 1197 BBI.HasFallThrough = false; 1198 } 1199 // Mixed predicated and unpredicated code. This cannot be iteratively 1200 // predicated. 1201 IterIfcvt = false; 1202 } 1203 1204 RemoveExtraEdges(BBI); 1205 1206 // Update block info. BB can be iteratively if-converted. 1207 if (!IterIfcvt) 1208 BBI.IsDone = true; 1209 InvalidatePreds(BBI.BB); 1210 CvtBBI->IsDone = true; 1211 if (FalseBBDead) 1212 NextBBI->IsDone = true; 1213 1214 // FIXME: Must maintain LiveIns. 1215 return true; 1216} 1217 1218/// IfConvertDiamond - If convert a diamond sub-CFG. 1219/// 1220bool IfConverter::IfConvertDiamond(BBInfo &BBI, IfcvtKind Kind, 1221 unsigned NumDups1, unsigned NumDups2) { 1222 BBInfo &TrueBBI = BBAnalysis[BBI.TrueBB->getNumber()]; 1223 BBInfo &FalseBBI = BBAnalysis[BBI.FalseBB->getNumber()]; 1224 MachineBasicBlock *TailBB = TrueBBI.TrueBB; 1225 // True block must fall through or end with an unanalyzable terminator. 1226 if (!TailBB) { 1227 if (blockAlwaysFallThrough(TrueBBI)) 1228 TailBB = FalseBBI.TrueBB; 1229 assert((TailBB || !TrueBBI.IsBrAnalyzable) && "Unexpected!"); 1230 } 1231 1232 if (TrueBBI.IsDone || FalseBBI.IsDone || 1233 TrueBBI.BB->pred_size() > 1 || 1234 FalseBBI.BB->pred_size() > 1) { 1235 // Something has changed. It's no longer safe to predicate these blocks. 1236 BBI.IsAnalyzed = false; 1237 TrueBBI.IsAnalyzed = false; 1238 FalseBBI.IsAnalyzed = false; 1239 return false; 1240 } 1241 1242 // Put the predicated instructions from the 'true' block before the 1243 // instructions from the 'false' block, unless the true block would clobber 1244 // the predicate, in which case, do the opposite. 1245 BBInfo *BBI1 = &TrueBBI; 1246 BBInfo *BBI2 = &FalseBBI; 1247 SmallVector<MachineOperand, 4> RevCond(BBI.BrCond.begin(), BBI.BrCond.end()); 1248 if (TII->ReverseBranchCondition(RevCond)) 1249 assert(false && "Unable to reverse branch condition!"); 1250 SmallVector<MachineOperand, 4> *Cond1 = &BBI.BrCond; 1251 SmallVector<MachineOperand, 4> *Cond2 = &RevCond; 1252 1253 // Figure out the more profitable ordering. 1254 bool DoSwap = false; 1255 if (TrueBBI.ClobbersPred && !FalseBBI.ClobbersPred) 1256 DoSwap = true; 1257 else if (TrueBBI.ClobbersPred == FalseBBI.ClobbersPred) { 1258 if (TrueBBI.NonPredSize > FalseBBI.NonPredSize) 1259 DoSwap = true; 1260 } 1261 if (DoSwap) { 1262 std::swap(BBI1, BBI2); 1263 std::swap(Cond1, Cond2); 1264 } 1265 1266 // Remove the conditional branch from entry to the blocks. 1267 BBI.NonPredSize -= TII->RemoveBranch(*BBI.BB); 1268 1269 // Initialize liveins to the first BB. These are potentially redefined by 1270 // predicated instructions. 1271 SmallSet<unsigned, 4> Redefs; 1272 InitPredRedefs(BBI1->BB, Redefs, TRI); 1273 1274 // Remove the duplicated instructions at the beginnings of both paths. 1275 MachineBasicBlock::iterator DI1 = BBI1->BB->begin(); 1276 MachineBasicBlock::iterator DI2 = BBI2->BB->begin(); 1277 MachineBasicBlock::iterator DIE1 = BBI1->BB->end(); 1278 MachineBasicBlock::iterator DIE2 = BBI2->BB->end(); 1279 // Skip dbg_value instructions 1280 while (DI1 != DIE1 && DI1->isDebugValue()) 1281 ++DI1; 1282 while (DI2 != DIE2 && DI2->isDebugValue()) 1283 ++DI2; 1284 BBI1->NonPredSize -= NumDups1; 1285 BBI2->NonPredSize -= NumDups1; 1286 1287 // Skip past the dups on each side separately since there may be 1288 // differing dbg_value entries. 1289 for (unsigned i = 0; i < NumDups1; ++DI1) { 1290 if (!DI1->isDebugValue()) 1291 ++i; 1292 } 1293 while (NumDups1 != 0) { 1294 ++DI2; 1295 if (!DI2->isDebugValue()) 1296 --NumDups1; 1297 } 1298 1299 UpdatePredRedefs(BBI1->BB->begin(), DI1, Redefs, TRI); 1300 BBI.BB->splice(BBI.BB->end(), BBI1->BB, BBI1->BB->begin(), DI1); 1301 BBI2->BB->erase(BBI2->BB->begin(), DI2); 1302 1303 // Predicate the 'true' block after removing its branch. 1304 BBI1->NonPredSize -= TII->RemoveBranch(*BBI1->BB); 1305 DI1 = BBI1->BB->end(); 1306 for (unsigned i = 0; i != NumDups2; ) { 1307 // NumDups2 only counted non-dbg_value instructions, so this won't 1308 // run off the head of the list. 1309 assert (DI1 != BBI1->BB->begin()); 1310 --DI1; 1311 // skip dbg_value instructions 1312 if (!DI1->isDebugValue()) 1313 ++i; 1314 } 1315 BBI1->BB->erase(DI1, BBI1->BB->end()); 1316 PredicateBlock(*BBI1, BBI1->BB->end(), *Cond1, Redefs); 1317 1318 // Predicate the 'false' block. 1319 BBI2->NonPredSize -= TII->RemoveBranch(*BBI2->BB); 1320 DI2 = BBI2->BB->end(); 1321 while (NumDups2 != 0) { 1322 // NumDups2 only counted non-dbg_value instructions, so this won't 1323 // run off the head of the list. 1324 assert (DI2 != BBI2->BB->begin()); 1325 --DI2; 1326 // skip dbg_value instructions 1327 if (!DI2->isDebugValue()) 1328 --NumDups2; 1329 } 1330 PredicateBlock(*BBI2, DI2, *Cond2, Redefs); 1331 1332 // Merge the true block into the entry of the diamond. 1333 MergeBlocks(BBI, *BBI1, TailBB == 0); 1334 MergeBlocks(BBI, *BBI2, TailBB == 0); 1335 1336 // If the if-converted block falls through or unconditionally branches into 1337 // the tail block, and the tail block does not have other predecessors, then 1338 // fold the tail block in as well. Otherwise, unless it falls through to the 1339 // tail, add a unconditional branch to it. 1340 if (TailBB) { 1341 BBInfo TailBBI = BBAnalysis[TailBB->getNumber()]; 1342 bool CanMergeTail = !TailBBI.HasFallThrough; 1343 // There may still be a fall-through edge from BBI1 or BBI2 to TailBB; 1344 // check if there are any other predecessors besides those. 1345 unsigned NumPreds = TailBB->pred_size(); 1346 if (NumPreds > 1) 1347 CanMergeTail = false; 1348 else if (NumPreds == 1 && CanMergeTail) { 1349 MachineBasicBlock::pred_iterator PI = TailBB->pred_begin(); 1350 if (*PI != BBI1->BB && *PI != BBI2->BB) 1351 CanMergeTail = false; 1352 } 1353 if (CanMergeTail) { 1354 MergeBlocks(BBI, TailBBI); 1355 TailBBI.IsDone = true; 1356 } else { 1357 BBI.BB->addSuccessor(TailBB); 1358 InsertUncondBranch(BBI.BB, TailBB, TII); 1359 BBI.HasFallThrough = false; 1360 } 1361 } 1362 1363 // RemoveExtraEdges won't work if the block has an unanalyzable branch, 1364 // which can happen here if TailBB is unanalyzable and is merged, so 1365 // explicitly remove BBI1 and BBI2 as successors. 1366 BBI.BB->removeSuccessor(BBI1->BB); 1367 BBI.BB->removeSuccessor(BBI2->BB); 1368 RemoveExtraEdges(BBI); 1369 1370 // Update block info. 1371 BBI.IsDone = TrueBBI.IsDone = FalseBBI.IsDone = true; 1372 InvalidatePreds(BBI.BB); 1373 1374 // FIXME: Must maintain LiveIns. 1375 return true; 1376} 1377 1378/// PredicateBlock - Predicate instructions from the start of the block to the 1379/// specified end with the specified condition. 1380void IfConverter::PredicateBlock(BBInfo &BBI, 1381 MachineBasicBlock::iterator E, 1382 SmallVectorImpl<MachineOperand> &Cond, 1383 SmallSet<unsigned, 4> &Redefs) { 1384 for (MachineBasicBlock::iterator I = BBI.BB->begin(); I != E; ++I) { 1385 if (I->isDebugValue() || TII->isPredicated(I)) 1386 continue; 1387 if (!TII->PredicateInstruction(I, Cond)) { 1388#ifndef NDEBUG 1389 dbgs() << "Unable to predicate " << *I << "!\n"; 1390#endif 1391 llvm_unreachable(0); 1392 } 1393 1394 // If the predicated instruction now redefines a register as the result of 1395 // if-conversion, add an implicit kill. 1396 UpdatePredRedefs(I, Redefs, TRI, true); 1397 } 1398 1399 std::copy(Cond.begin(), Cond.end(), std::back_inserter(BBI.Predicate)); 1400 1401 BBI.IsAnalyzed = false; 1402 BBI.NonPredSize = 0; 1403 1404 ++NumIfConvBBs; 1405} 1406 1407/// CopyAndPredicateBlock - Copy and predicate instructions from source BB to 1408/// the destination block. Skip end of block branches if IgnoreBr is true. 1409void IfConverter::CopyAndPredicateBlock(BBInfo &ToBBI, BBInfo &FromBBI, 1410 SmallVectorImpl<MachineOperand> &Cond, 1411 SmallSet<unsigned, 4> &Redefs, 1412 bool IgnoreBr) { 1413 MachineFunction &MF = *ToBBI.BB->getParent(); 1414 1415 for (MachineBasicBlock::iterator I = FromBBI.BB->begin(), 1416 E = FromBBI.BB->end(); I != E; ++I) { 1417 const MCInstrDesc &MCID = I->getDesc(); 1418 // Do not copy the end of the block branches. 1419 if (IgnoreBr && MCID.isBranch()) 1420 break; 1421 1422 MachineInstr *MI = MF.CloneMachineInstr(I); 1423 ToBBI.BB->insert(ToBBI.BB->end(), MI); 1424 ToBBI.NonPredSize++; 1425 unsigned ExtraPredCost = 0; 1426 unsigned NumCycles = TII->getInstrLatency(InstrItins, &*I, &ExtraPredCost); 1427 if (NumCycles > 1) 1428 ToBBI.ExtraCost += NumCycles-1; 1429 ToBBI.ExtraCost2 += ExtraPredCost; 1430 1431 if (!TII->isPredicated(I) && !MI->isDebugValue()) { 1432 if (!TII->PredicateInstruction(MI, Cond)) { 1433#ifndef NDEBUG 1434 dbgs() << "Unable to predicate " << *I << "!\n"; 1435#endif 1436 llvm_unreachable(0); 1437 } 1438 } 1439 1440 // If the predicated instruction now redefines a register as the result of 1441 // if-conversion, add an implicit kill. 1442 UpdatePredRedefs(MI, Redefs, TRI, true); 1443 } 1444 1445 if (!IgnoreBr) { 1446 std::vector<MachineBasicBlock *> Succs(FromBBI.BB->succ_begin(), 1447 FromBBI.BB->succ_end()); 1448 MachineBasicBlock *NBB = getNextBlock(FromBBI.BB); 1449 MachineBasicBlock *FallThrough = FromBBI.HasFallThrough ? NBB : NULL; 1450 1451 for (unsigned i = 0, e = Succs.size(); i != e; ++i) { 1452 MachineBasicBlock *Succ = Succs[i]; 1453 // Fallthrough edge can't be transferred. 1454 if (Succ == FallThrough) 1455 continue; 1456 ToBBI.BB->addSuccessor(Succ); 1457 } 1458 } 1459 1460 std::copy(FromBBI.Predicate.begin(), FromBBI.Predicate.end(), 1461 std::back_inserter(ToBBI.Predicate)); 1462 std::copy(Cond.begin(), Cond.end(), std::back_inserter(ToBBI.Predicate)); 1463 1464 ToBBI.ClobbersPred |= FromBBI.ClobbersPred; 1465 ToBBI.IsAnalyzed = false; 1466 1467 ++NumDupBBs; 1468} 1469 1470/// MergeBlocks - Move all instructions from FromBB to the end of ToBB. 1471/// This will leave FromBB as an empty block, so remove all of its 1472/// successor edges except for the fall-through edge. If AddEdges is true, 1473/// i.e., when FromBBI's branch is being moved, add those successor edges to 1474/// ToBBI. 1475void IfConverter::MergeBlocks(BBInfo &ToBBI, BBInfo &FromBBI, bool AddEdges) { 1476 ToBBI.BB->splice(ToBBI.BB->end(), 1477 FromBBI.BB, FromBBI.BB->begin(), FromBBI.BB->end()); 1478 1479 std::vector<MachineBasicBlock *> Succs(FromBBI.BB->succ_begin(), 1480 FromBBI.BB->succ_end()); 1481 MachineBasicBlock *NBB = getNextBlock(FromBBI.BB); 1482 MachineBasicBlock *FallThrough = FromBBI.HasFallThrough ? NBB : NULL; 1483 1484 for (unsigned i = 0, e = Succs.size(); i != e; ++i) { 1485 MachineBasicBlock *Succ = Succs[i]; 1486 // Fallthrough edge can't be transferred. 1487 if (Succ == FallThrough) 1488 continue; 1489 FromBBI.BB->removeSuccessor(Succ); 1490 if (AddEdges) 1491 ToBBI.BB->addSuccessor(Succ); 1492 } 1493 1494 // Now FromBBI always falls through to the next block! 1495 if (NBB && !FromBBI.BB->isSuccessor(NBB)) 1496 FromBBI.BB->addSuccessor(NBB); 1497 1498 std::copy(FromBBI.Predicate.begin(), FromBBI.Predicate.end(), 1499 std::back_inserter(ToBBI.Predicate)); 1500 FromBBI.Predicate.clear(); 1501 1502 ToBBI.NonPredSize += FromBBI.NonPredSize; 1503 ToBBI.ExtraCost += FromBBI.ExtraCost; 1504 ToBBI.ExtraCost2 += FromBBI.ExtraCost2; 1505 FromBBI.NonPredSize = 0; 1506 FromBBI.ExtraCost = 0; 1507 FromBBI.ExtraCost2 = 0; 1508 1509 ToBBI.ClobbersPred |= FromBBI.ClobbersPred; 1510 ToBBI.HasFallThrough = FromBBI.HasFallThrough; 1511 ToBBI.IsAnalyzed = false; 1512 FromBBI.IsAnalyzed = false; 1513} 1514