IfConversion.cpp revision 4aad8bdb06a250ddd3d5684a4bb33eb02f5750d5
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/Target/TargetInstrInfo.h" 22#include "llvm/Target/TargetInstrItineraries.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/DepthFirstIterator.h" 31#include "llvm/ADT/Statistic.h" 32#include "llvm/ADT/STLExtras.h" 33using namespace llvm; 34 35// Hidden options for help debugging. 36static cl::opt<int> IfCvtFnStart("ifcvt-fn-start", cl::init(-1), cl::Hidden); 37static cl::opt<int> IfCvtFnStop("ifcvt-fn-stop", cl::init(-1), cl::Hidden); 38static cl::opt<int> IfCvtLimit("ifcvt-limit", cl::init(-1), cl::Hidden); 39static cl::opt<bool> DisableSimple("disable-ifcvt-simple", 40 cl::init(false), cl::Hidden); 41static cl::opt<bool> DisableSimpleF("disable-ifcvt-simple-false", 42 cl::init(false), cl::Hidden); 43static cl::opt<bool> DisableTriangle("disable-ifcvt-triangle", 44 cl::init(false), cl::Hidden); 45static cl::opt<bool> DisableTriangleR("disable-ifcvt-triangle-rev", 46 cl::init(false), cl::Hidden); 47static cl::opt<bool> DisableTriangleF("disable-ifcvt-triangle-false", 48 cl::init(false), cl::Hidden); 49static cl::opt<bool> DisableTriangleFR("disable-ifcvt-triangle-false-rev", 50 cl::init(false), cl::Hidden); 51static cl::opt<bool> DisableDiamond("disable-ifcvt-diamond", 52 cl::init(false), cl::Hidden); 53static cl::opt<bool> IfCvtBranchFold("ifcvt-branch-fold", 54 cl::init(true), cl::Hidden); 55 56STATISTIC(NumSimple, "Number of simple if-conversions performed"); 57STATISTIC(NumSimpleFalse, "Number of simple (F) if-conversions performed"); 58STATISTIC(NumTriangle, "Number of triangle if-conversions performed"); 59STATISTIC(NumTriangleRev, "Number of triangle (R) if-conversions performed"); 60STATISTIC(NumTriangleFalse,"Number of triangle (F) if-conversions performed"); 61STATISTIC(NumTriangleFRev, "Number of triangle (F/R) if-conversions performed"); 62STATISTIC(NumDiamonds, "Number of diamond if-conversions performed"); 63STATISTIC(NumIfConvBBs, "Number of if-converted blocks"); 64STATISTIC(NumDupBBs, "Number of duplicated blocks"); 65 66namespace { 67 class IfConverter : public MachineFunctionPass { 68 enum IfcvtKind { 69 ICNotClassfied, // BB data valid, but not classified. 70 ICSimpleFalse, // Same as ICSimple, but on the false path. 71 ICSimple, // BB is entry of an one split, no rejoin sub-CFG. 72 ICTriangleFRev, // Same as ICTriangleFalse, but false path rev condition. 73 ICTriangleRev, // Same as ICTriangle, but true path rev condition. 74 ICTriangleFalse, // Same as ICTriangle, but on the false path. 75 ICTriangle, // BB is entry of a triangle sub-CFG. 76 ICDiamond // BB is entry of a diamond sub-CFG. 77 }; 78 79 /// BBInfo - One per MachineBasicBlock, this is used to cache the result 80 /// if-conversion feasibility analysis. This includes results from 81 /// TargetInstrInfo::AnalyzeBranch() (i.e. TBB, FBB, and Cond), and its 82 /// classification, and common tail block of its successors (if it's a 83 /// diamond shape), its size, whether it's predicable, and whether any 84 /// instruction can clobber the 'would-be' predicate. 85 /// 86 /// IsDone - True if BB is not to be considered for ifcvt. 87 /// IsBeingAnalyzed - True if BB is currently being analyzed. 88 /// IsAnalyzed - True if BB has been analyzed (info is still valid). 89 /// IsEnqueued - True if BB has been enqueued to be ifcvt'ed. 90 /// IsBrAnalyzable - True if AnalyzeBranch() returns false. 91 /// HasFallThrough - True if BB may fallthrough to the following BB. 92 /// IsUnpredicable - True if BB is known to be unpredicable. 93 /// ClobbersPred - True if BB could modify predicates (e.g. has 94 /// cmp, call, etc.) 95 /// NonPredSize - Number of non-predicated instructions. 96 /// BB - Corresponding MachineBasicBlock. 97 /// TrueBB / FalseBB- See AnalyzeBranch(). 98 /// BrCond - Conditions for end of block conditional branches. 99 /// Predicate - Predicate used in the BB. 100 struct BBInfo { 101 bool IsDone : 1; 102 bool IsBeingAnalyzed : 1; 103 bool IsAnalyzed : 1; 104 bool IsEnqueued : 1; 105 bool IsBrAnalyzable : 1; 106 bool HasFallThrough : 1; 107 bool IsUnpredicable : 1; 108 bool CannotBeCopied : 1; 109 bool ClobbersPred : 1; 110 unsigned NonPredSize; 111 MachineBasicBlock *BB; 112 MachineBasicBlock *TrueBB; 113 MachineBasicBlock *FalseBB; 114 SmallVector<MachineOperand, 4> BrCond; 115 SmallVector<MachineOperand, 4> Predicate; 116 BBInfo() : IsDone(false), IsBeingAnalyzed(false), 117 IsAnalyzed(false), IsEnqueued(false), IsBrAnalyzable(false), 118 HasFallThrough(false), IsUnpredicable(false), 119 CannotBeCopied(false), ClobbersPred(false), NonPredSize(0), 120 BB(0), TrueBB(0), FalseBB(0) {} 121 }; 122 123 /// IfcvtToken - Record information about pending if-conversions to attempt: 124 /// BBI - Corresponding BBInfo. 125 /// Kind - Type of block. See IfcvtKind. 126 /// NeedSubsumption - True if the to-be-predicated BB has already been 127 /// predicated. 128 /// NumDups - Number of instructions that would be duplicated due 129 /// to this if-conversion. (For diamonds, the number of 130 /// identical instructions at the beginnings of both 131 /// paths). 132 /// NumDups2 - For diamonds, the number of identical instructions 133 /// at the ends of both paths. 134 struct IfcvtToken { 135 BBInfo &BBI; 136 IfcvtKind Kind; 137 bool NeedSubsumption; 138 unsigned NumDups; 139 unsigned NumDups2; 140 IfcvtToken(BBInfo &b, IfcvtKind k, bool s, unsigned d, unsigned d2 = 0) 141 : BBI(b), Kind(k), NeedSubsumption(s), NumDups(d), NumDups2(d2) {} 142 }; 143 144 /// Roots - Basic blocks that do not have successors. These are the starting 145 /// points of Graph traversal. 146 std::vector<MachineBasicBlock*> Roots; 147 148 /// BBAnalysis - Results of if-conversion feasibility analysis indexed by 149 /// basic block number. 150 std::vector<BBInfo> BBAnalysis; 151 152 const TargetLowering *TLI; 153 const TargetInstrInfo *TII; 154 const TargetRegisterInfo *TRI; 155 const InstrItineraryData *InstrItins; 156 const MachineLoopInfo *MLI; 157 bool MadeChange; 158 int FnNum; 159 public: 160 static char ID; 161 IfConverter() : MachineFunctionPass(ID), FnNum(-1) {} 162 163 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 164 AU.addRequired<MachineLoopInfo>(); 165 MachineFunctionPass::getAnalysisUsage(AU); 166 } 167 168 virtual bool runOnMachineFunction(MachineFunction &MF); 169 virtual const char *getPassName() const { return "If Converter"; } 170 171 private: 172 bool ReverseBranchCondition(BBInfo &BBI); 173 bool ValidSimple(BBInfo &TrueBBI, unsigned &Dups, float Prediction) const; 174 bool ValidTriangle(BBInfo &TrueBBI, BBInfo &FalseBBI, 175 bool FalseBranch, unsigned &Dups) const; 176 bool ValidDiamond(BBInfo &TrueBBI, BBInfo &FalseBBI, 177 unsigned &Dups1, unsigned &Dups2) const; 178 void ScanInstructions(BBInfo &BBI); 179 BBInfo &AnalyzeBlock(MachineBasicBlock *BB, 180 std::vector<IfcvtToken*> &Tokens); 181 bool FeasibilityAnalysis(BBInfo &BBI, SmallVectorImpl<MachineOperand> &Cond, 182 bool isTriangle = false, bool RevBranch = false); 183 void AnalyzeBlocks(MachineFunction &MF, std::vector<IfcvtToken*> &Tokens); 184 void InvalidatePreds(MachineBasicBlock *BB); 185 void RemoveExtraEdges(BBInfo &BBI); 186 bool IfConvertSimple(BBInfo &BBI, IfcvtKind Kind); 187 bool IfConvertTriangle(BBInfo &BBI, IfcvtKind Kind); 188 bool IfConvertDiamond(BBInfo &BBI, IfcvtKind Kind, 189 unsigned NumDups1, unsigned NumDups2); 190 void PredicateBlock(BBInfo &BBI, 191 MachineBasicBlock::iterator E, 192 SmallVectorImpl<MachineOperand> &Cond, 193 SmallSet<unsigned, 4> &Redefs); 194 void CopyAndPredicateBlock(BBInfo &ToBBI, BBInfo &FromBBI, 195 SmallVectorImpl<MachineOperand> &Cond, 196 SmallSet<unsigned, 4> &Redefs, 197 bool IgnoreBr = false); 198 void MergeBlocks(BBInfo &ToBBI, BBInfo &FromBBI, bool AddEdges = true); 199 200 bool MeetIfcvtSizeLimit(MachineBasicBlock &BB, unsigned Size, 201 float Prediction) const { 202 return Size > 0 && TII->isProfitableToIfCvt(BB, Size, Prediction); 203 } 204 205 bool MeetIfcvtSizeLimit(MachineBasicBlock &TBB, unsigned TSize, 206 MachineBasicBlock &FBB, unsigned FSize, 207 float Prediction) const { 208 return TSize > 0 && FSize > 0 && 209 TII->isProfitableToIfCvt(TBB, TSize, FBB, FSize, Prediction); 210 } 211 212 // blockAlwaysFallThrough - Block ends without a terminator. 213 bool blockAlwaysFallThrough(BBInfo &BBI) const { 214 return BBI.IsBrAnalyzable && BBI.TrueBB == NULL; 215 } 216 217 // IfcvtTokenCmp - Used to sort if-conversion candidates. 218 static bool IfcvtTokenCmp(IfcvtToken *C1, IfcvtToken *C2) { 219 int Incr1 = (C1->Kind == ICDiamond) 220 ? -(int)(C1->NumDups + C1->NumDups2) : (int)C1->NumDups; 221 int Incr2 = (C2->Kind == ICDiamond) 222 ? -(int)(C2->NumDups + C2->NumDups2) : (int)C2->NumDups; 223 if (Incr1 > Incr2) 224 return true; 225 else if (Incr1 == Incr2) { 226 // Favors subsumption. 227 if (C1->NeedSubsumption == false && C2->NeedSubsumption == true) 228 return true; 229 else if (C1->NeedSubsumption == C2->NeedSubsumption) { 230 // Favors diamond over triangle, etc. 231 if ((unsigned)C1->Kind < (unsigned)C2->Kind) 232 return true; 233 else if (C1->Kind == C2->Kind) 234 return C1->BBI.BB->getNumber() < C2->BBI.BB->getNumber(); 235 } 236 } 237 return false; 238 } 239 }; 240 241 char IfConverter::ID = 0; 242} 243 244INITIALIZE_PASS(IfConverter, "if-converter", "If Converter", false, false); 245 246FunctionPass *llvm::createIfConverterPass() { return new IfConverter(); } 247 248bool IfConverter::runOnMachineFunction(MachineFunction &MF) { 249 TLI = MF.getTarget().getTargetLowering(); 250 TII = MF.getTarget().getInstrInfo(); 251 TRI = MF.getTarget().getRegisterInfo(); 252 MLI = &getAnalysis<MachineLoopInfo>(); 253 InstrItins = MF.getTarget().getInstrItineraryData(); 254 if (!TII) return false; 255 256 // Tail merge tend to expose more if-conversion opportunities. 257 BranchFolder BF(true); 258 bool BFChange = BF.OptimizeFunction(MF, TII, 259 MF.getTarget().getRegisterInfo(), 260 getAnalysisIfAvailable<MachineModuleInfo>()); 261 262 DEBUG(dbgs() << "\nIfcvt: function (" << ++FnNum << ") \'" 263 << MF.getFunction()->getName() << "\'"); 264 265 if (FnNum < IfCvtFnStart || (IfCvtFnStop != -1 && FnNum > IfCvtFnStop)) { 266 DEBUG(dbgs() << " skipped\n"); 267 return false; 268 } 269 DEBUG(dbgs() << "\n"); 270 271 MF.RenumberBlocks(); 272 BBAnalysis.resize(MF.getNumBlockIDs()); 273 274 // Look for root nodes, i.e. blocks without successors. 275 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; ++I) 276 if (I->succ_empty()) 277 Roots.push_back(I); 278 279 std::vector<IfcvtToken*> Tokens; 280 MadeChange = false; 281 unsigned NumIfCvts = NumSimple + NumSimpleFalse + NumTriangle + 282 NumTriangleRev + NumTriangleFalse + NumTriangleFRev + NumDiamonds; 283 while (IfCvtLimit == -1 || (int)NumIfCvts < IfCvtLimit) { 284 // Do an initial analysis for each basic block and find all the potential 285 // candidates to perform if-conversion. 286 bool Change = false; 287 AnalyzeBlocks(MF, Tokens); 288 while (!Tokens.empty()) { 289 IfcvtToken *Token = Tokens.back(); 290 Tokens.pop_back(); 291 BBInfo &BBI = Token->BBI; 292 IfcvtKind Kind = Token->Kind; 293 unsigned NumDups = Token->NumDups; 294 unsigned NumDups2 = Token->NumDups2; 295 296 delete Token; 297 298 // If the block has been evicted out of the queue or it has already been 299 // marked dead (due to it being predicated), then skip it. 300 if (BBI.IsDone) 301 BBI.IsEnqueued = false; 302 if (!BBI.IsEnqueued) 303 continue; 304 305 BBI.IsEnqueued = false; 306 307 bool RetVal = false; 308 switch (Kind) { 309 default: assert(false && "Unexpected!"); 310 break; 311 case ICSimple: 312 case ICSimpleFalse: { 313 bool isFalse = Kind == ICSimpleFalse; 314 if ((isFalse && DisableSimpleF) || (!isFalse && DisableSimple)) break; 315 DEBUG(dbgs() << "Ifcvt (Simple" << (Kind == ICSimpleFalse ? 316 " false" : "") 317 << "): BB#" << BBI.BB->getNumber() << " (" 318 << ((Kind == ICSimpleFalse) 319 ? BBI.FalseBB->getNumber() 320 : BBI.TrueBB->getNumber()) << ") "); 321 RetVal = IfConvertSimple(BBI, Kind); 322 DEBUG(dbgs() << (RetVal ? "succeeded!" : "failed!") << "\n"); 323 if (RetVal) { 324 if (isFalse) ++NumSimpleFalse; 325 else ++NumSimple; 326 } 327 break; 328 } 329 case ICTriangle: 330 case ICTriangleRev: 331 case ICTriangleFalse: 332 case ICTriangleFRev: { 333 bool isFalse = Kind == ICTriangleFalse; 334 bool isRev = (Kind == ICTriangleRev || Kind == ICTriangleFRev); 335 if (DisableTriangle && !isFalse && !isRev) break; 336 if (DisableTriangleR && !isFalse && isRev) break; 337 if (DisableTriangleF && isFalse && !isRev) break; 338 if (DisableTriangleFR && isFalse && isRev) break; 339 DEBUG(dbgs() << "Ifcvt (Triangle"); 340 if (isFalse) 341 DEBUG(dbgs() << " false"); 342 if (isRev) 343 DEBUG(dbgs() << " rev"); 344 DEBUG(dbgs() << "): BB#" << BBI.BB->getNumber() << " (T:" 345 << BBI.TrueBB->getNumber() << ",F:" 346 << BBI.FalseBB->getNumber() << ") "); 347 RetVal = IfConvertTriangle(BBI, Kind); 348 DEBUG(dbgs() << (RetVal ? "succeeded!" : "failed!") << "\n"); 349 if (RetVal) { 350 if (isFalse) { 351 if (isRev) ++NumTriangleFRev; 352 else ++NumTriangleFalse; 353 } else { 354 if (isRev) ++NumTriangleRev; 355 else ++NumTriangle; 356 } 357 } 358 break; 359 } 360 case ICDiamond: { 361 if (DisableDiamond) break; 362 DEBUG(dbgs() << "Ifcvt (Diamond): BB#" << BBI.BB->getNumber() << " (T:" 363 << BBI.TrueBB->getNumber() << ",F:" 364 << BBI.FalseBB->getNumber() << ") "); 365 RetVal = IfConvertDiamond(BBI, Kind, NumDups, NumDups2); 366 DEBUG(dbgs() << (RetVal ? "succeeded!" : "failed!") << "\n"); 367 if (RetVal) ++NumDiamonds; 368 break; 369 } 370 } 371 372 Change |= RetVal; 373 374 NumIfCvts = NumSimple + NumSimpleFalse + NumTriangle + NumTriangleRev + 375 NumTriangleFalse + NumTriangleFRev + NumDiamonds; 376 if (IfCvtLimit != -1 && (int)NumIfCvts >= IfCvtLimit) 377 break; 378 } 379 380 if (!Change) 381 break; 382 MadeChange |= Change; 383 } 384 385 // Delete tokens in case of early exit. 386 while (!Tokens.empty()) { 387 IfcvtToken *Token = Tokens.back(); 388 Tokens.pop_back(); 389 delete Token; 390 } 391 392 Tokens.clear(); 393 Roots.clear(); 394 BBAnalysis.clear(); 395 396 if (MadeChange && IfCvtBranchFold) { 397 BranchFolder BF(false); 398 BF.OptimizeFunction(MF, TII, 399 MF.getTarget().getRegisterInfo(), 400 getAnalysisIfAvailable<MachineModuleInfo>()); 401 } 402 403 MadeChange |= BFChange; 404 return MadeChange; 405} 406 407/// findFalseBlock - BB has a fallthrough. Find its 'false' successor given 408/// its 'true' successor. 409static MachineBasicBlock *findFalseBlock(MachineBasicBlock *BB, 410 MachineBasicBlock *TrueBB) { 411 for (MachineBasicBlock::succ_iterator SI = BB->succ_begin(), 412 E = BB->succ_end(); SI != E; ++SI) { 413 MachineBasicBlock *SuccBB = *SI; 414 if (SuccBB != TrueBB) 415 return SuccBB; 416 } 417 return NULL; 418} 419 420/// ReverseBranchCondition - Reverse the condition of the end of the block 421/// branch. Swap block's 'true' and 'false' successors. 422bool IfConverter::ReverseBranchCondition(BBInfo &BBI) { 423 DebugLoc dl; // FIXME: this is nowhere 424 if (!TII->ReverseBranchCondition(BBI.BrCond)) { 425 TII->RemoveBranch(*BBI.BB); 426 TII->InsertBranch(*BBI.BB, BBI.FalseBB, BBI.TrueBB, BBI.BrCond, dl); 427 std::swap(BBI.TrueBB, BBI.FalseBB); 428 return true; 429 } 430 return false; 431} 432 433/// getNextBlock - Returns the next block in the function blocks ordering. If 434/// it is the end, returns NULL. 435static inline MachineBasicBlock *getNextBlock(MachineBasicBlock *BB) { 436 MachineFunction::iterator I = BB; 437 MachineFunction::iterator E = BB->getParent()->end(); 438 if (++I == E) 439 return NULL; 440 return I; 441} 442 443/// ValidSimple - Returns true if the 'true' block (along with its 444/// predecessor) forms a valid simple shape for ifcvt. It also returns the 445/// number of instructions that the ifcvt would need to duplicate if performed 446/// in Dups. 447bool IfConverter::ValidSimple(BBInfo &TrueBBI, unsigned &Dups, 448 float Prediction) const { 449 Dups = 0; 450 if (TrueBBI.IsBeingAnalyzed || TrueBBI.IsDone) 451 return false; 452 453 if (TrueBBI.IsBrAnalyzable) 454 return false; 455 456 if (TrueBBI.BB->pred_size() > 1) { 457 if (TrueBBI.CannotBeCopied || 458 !TII->isProfitableToDupForIfCvt(*TrueBBI.BB, TrueBBI.NonPredSize, 459 Prediction)) 460 return false; 461 Dups = TrueBBI.NonPredSize; 462 } 463 464 return true; 465} 466 467/// ValidTriangle - Returns true if the 'true' and 'false' blocks (along 468/// with their common predecessor) forms a valid triangle shape for ifcvt. 469/// If 'FalseBranch' is true, it checks if 'true' block's false branch 470/// branches to the 'false' block rather than the other way around. It also 471/// returns the number of instructions that the ifcvt would need to duplicate 472/// if performed in 'Dups'. 473bool IfConverter::ValidTriangle(BBInfo &TrueBBI, BBInfo &FalseBBI, 474 bool FalseBranch, unsigned &Dups) const { 475 Dups = 0; 476 if (TrueBBI.IsBeingAnalyzed || TrueBBI.IsDone) 477 return false; 478 479 if (TrueBBI.BB->pred_size() > 1) { 480 if (TrueBBI.CannotBeCopied) 481 return false; 482 483 unsigned Size = TrueBBI.NonPredSize; 484 if (TrueBBI.IsBrAnalyzable) { 485 if (TrueBBI.TrueBB && TrueBBI.BrCond.empty()) 486 // Ends with an unconditional branch. It will be removed. 487 --Size; 488 else { 489 MachineBasicBlock *FExit = FalseBranch 490 ? TrueBBI.TrueBB : TrueBBI.FalseBB; 491 if (FExit) 492 // Require a conditional branch 493 ++Size; 494 } 495 } 496 if (!TII->isProfitableToDupForIfCvt(*TrueBBI.BB, Size, 0.5)) 497 return false; 498 Dups = Size; 499 } 500 501 MachineBasicBlock *TExit = FalseBranch ? TrueBBI.FalseBB : TrueBBI.TrueBB; 502 if (!TExit && blockAlwaysFallThrough(TrueBBI)) { 503 MachineFunction::iterator I = TrueBBI.BB; 504 if (++I == TrueBBI.BB->getParent()->end()) 505 return false; 506 TExit = I; 507 } 508 return TExit && TExit == FalseBBI.BB; 509} 510 511static 512MachineBasicBlock::iterator firstNonBranchInst(MachineBasicBlock *BB, 513 const TargetInstrInfo *TII) { 514 MachineBasicBlock::iterator I = BB->end(); 515 while (I != BB->begin()) { 516 --I; 517 if (!I->getDesc().isBranch()) 518 break; 519 } 520 return I; 521} 522 523/// ValidDiamond - Returns true if the 'true' and 'false' blocks (along 524/// with their common predecessor) forms a valid diamond shape for ifcvt. 525bool IfConverter::ValidDiamond(BBInfo &TrueBBI, BBInfo &FalseBBI, 526 unsigned &Dups1, unsigned &Dups2) const { 527 Dups1 = Dups2 = 0; 528 if (TrueBBI.IsBeingAnalyzed || TrueBBI.IsDone || 529 FalseBBI.IsBeingAnalyzed || FalseBBI.IsDone) 530 return false; 531 532 MachineBasicBlock *TT = TrueBBI.TrueBB; 533 MachineBasicBlock *FT = FalseBBI.TrueBB; 534 535 if (!TT && blockAlwaysFallThrough(TrueBBI)) 536 TT = getNextBlock(TrueBBI.BB); 537 if (!FT && blockAlwaysFallThrough(FalseBBI)) 538 FT = getNextBlock(FalseBBI.BB); 539 if (TT != FT) 540 return false; 541 if (TT == NULL && (TrueBBI.IsBrAnalyzable || FalseBBI.IsBrAnalyzable)) 542 return false; 543 if (TrueBBI.BB->pred_size() > 1 || FalseBBI.BB->pred_size() > 1) 544 return false; 545 546 // FIXME: Allow true block to have an early exit? 547 if (TrueBBI.FalseBB || FalseBBI.FalseBB || 548 (TrueBBI.ClobbersPred && FalseBBI.ClobbersPred)) 549 return false; 550 551 MachineBasicBlock::iterator TI = TrueBBI.BB->begin(); 552 MachineBasicBlock::iterator FI = FalseBBI.BB->begin(); 553 MachineBasicBlock::iterator TIE = TrueBBI.BB->end(); 554 MachineBasicBlock::iterator FIE = FalseBBI.BB->end(); 555 // Skip dbg_value instructions 556 while (TI != TIE && TI->isDebugValue()) 557 ++TI; 558 while (FI != FIE && FI->isDebugValue()) 559 ++FI; 560 while (TI != TIE && FI != FIE) { 561 // Skip dbg_value instructions. These do not count. 562 if (TI->isDebugValue()) { 563 while (TI != TIE && TI->isDebugValue()) 564 ++TI; 565 if (TI == TIE) 566 break; 567 } 568 if (FI->isDebugValue()) { 569 while (FI != FIE && FI->isDebugValue()) 570 ++FI; 571 if (FI == FIE) 572 break; 573 } 574 if (!TI->isIdenticalTo(FI)) 575 break; 576 ++Dups1; 577 ++TI; 578 ++FI; 579 } 580 581 TI = firstNonBranchInst(TrueBBI.BB, TII); 582 FI = firstNonBranchInst(FalseBBI.BB, TII); 583 MachineBasicBlock::iterator TIB = TrueBBI.BB->begin(); 584 MachineBasicBlock::iterator FIB = FalseBBI.BB->begin(); 585 // Skip dbg_value instructions at end of the bb's. 586 while (TI != TIB && TI->isDebugValue()) 587 --TI; 588 while (FI != FIB && FI->isDebugValue()) 589 --FI; 590 while (TI != TIB && FI != FIB) { 591 // Skip dbg_value instructions. These do not count. 592 if (TI->isDebugValue()) { 593 while (TI != TIB && TI->isDebugValue()) 594 --TI; 595 if (TI == TIB) 596 break; 597 } 598 if (FI->isDebugValue()) { 599 while (FI != FIB && FI->isDebugValue()) 600 --FI; 601 if (FI == FIB) 602 break; 603 } 604 if (!TI->isIdenticalTo(FI)) 605 break; 606 ++Dups2; 607 --TI; 608 --FI; 609 } 610 611 return true; 612} 613 614/// ScanInstructions - Scan all the instructions in the block to determine if 615/// the block is predicable. In most cases, that means all the instructions 616/// in the block are isPredicable(). Also checks if the block contains any 617/// instruction which can clobber a predicate (e.g. condition code register). 618/// If so, the block is not predicable unless it's the last instruction. 619void IfConverter::ScanInstructions(BBInfo &BBI) { 620 if (BBI.IsDone) 621 return; 622 623 bool AlreadyPredicated = BBI.Predicate.size() > 0; 624 // First analyze the end of BB branches. 625 BBI.TrueBB = BBI.FalseBB = NULL; 626 BBI.BrCond.clear(); 627 BBI.IsBrAnalyzable = 628 !TII->AnalyzeBranch(*BBI.BB, BBI.TrueBB, BBI.FalseBB, BBI.BrCond); 629 BBI.HasFallThrough = BBI.IsBrAnalyzable && BBI.FalseBB == NULL; 630 631 if (BBI.BrCond.size()) { 632 // No false branch. This BB must end with a conditional branch and a 633 // fallthrough. 634 if (!BBI.FalseBB) 635 BBI.FalseBB = findFalseBlock(BBI.BB, BBI.TrueBB); 636 if (!BBI.FalseBB) { 637 // Malformed bcc? True and false blocks are the same? 638 BBI.IsUnpredicable = true; 639 return; 640 } 641 } 642 643 // Then scan all the instructions. 644 BBI.NonPredSize = 0; 645 BBI.ClobbersPred = false; 646 for (MachineBasicBlock::iterator I = BBI.BB->begin(), E = BBI.BB->end(); 647 I != E; ++I) { 648 if (I->isDebugValue()) 649 continue; 650 651 const TargetInstrDesc &TID = I->getDesc(); 652 if (TID.isNotDuplicable()) 653 BBI.CannotBeCopied = true; 654 655 bool isPredicated = TII->isPredicated(I); 656 bool isCondBr = BBI.IsBrAnalyzable && TID.isConditionalBranch(); 657 658 if (!isCondBr) { 659 if (!isPredicated) { 660 unsigned NumOps = TII->getNumMicroOps(&*I, InstrItins); 661 BBI.NonPredSize += NumOps; 662 } else if (!AlreadyPredicated) { 663 // FIXME: This instruction is already predicated before the 664 // if-conversion pass. It's probably something like a conditional move. 665 // Mark this block unpredicable for now. 666 BBI.IsUnpredicable = true; 667 return; 668 } 669 } 670 671 if (BBI.ClobbersPred && !isPredicated) { 672 // Predicate modification instruction should end the block (except for 673 // already predicated instructions and end of block branches). 674 if (isCondBr) { 675 // A conditional branch is not predicable, but it may be eliminated. 676 continue; 677 } 678 679 // Predicate may have been modified, the subsequent (currently) 680 // unpredicated instructions cannot be correctly predicated. 681 BBI.IsUnpredicable = true; 682 return; 683 } 684 685 // FIXME: Make use of PredDefs? e.g. ADDC, SUBC sets predicates but are 686 // still potentially predicable. 687 std::vector<MachineOperand> PredDefs; 688 if (TII->DefinesPredicate(I, PredDefs)) 689 BBI.ClobbersPred = true; 690 691 if (!TII->isPredicable(I)) { 692 BBI.IsUnpredicable = true; 693 return; 694 } 695 } 696} 697 698/// FeasibilityAnalysis - Determine if the block is a suitable candidate to be 699/// predicated by the specified predicate. 700bool IfConverter::FeasibilityAnalysis(BBInfo &BBI, 701 SmallVectorImpl<MachineOperand> &Pred, 702 bool isTriangle, bool RevBranch) { 703 // If the block is dead or unpredicable, then it cannot be predicated. 704 if (BBI.IsDone || BBI.IsUnpredicable) 705 return false; 706 707 // If it is already predicated, check if its predicate subsumes the new 708 // predicate. 709 if (BBI.Predicate.size() && !TII->SubsumesPredicate(BBI.Predicate, Pred)) 710 return false; 711 712 if (BBI.BrCond.size()) { 713 if (!isTriangle) 714 return false; 715 716 // Test predicate subsumption. 717 SmallVector<MachineOperand, 4> RevPred(Pred.begin(), Pred.end()); 718 SmallVector<MachineOperand, 4> Cond(BBI.BrCond.begin(), BBI.BrCond.end()); 719 if (RevBranch) { 720 if (TII->ReverseBranchCondition(Cond)) 721 return false; 722 } 723 if (TII->ReverseBranchCondition(RevPred) || 724 !TII->SubsumesPredicate(Cond, RevPred)) 725 return false; 726 } 727 728 return true; 729} 730 731/// AnalyzeBlock - Analyze the structure of the sub-CFG starting from 732/// the specified block. Record its successors and whether it looks like an 733/// if-conversion candidate. 734IfConverter::BBInfo &IfConverter::AnalyzeBlock(MachineBasicBlock *BB, 735 std::vector<IfcvtToken*> &Tokens) { 736 BBInfo &BBI = BBAnalysis[BB->getNumber()]; 737 738 if (BBI.IsAnalyzed || BBI.IsBeingAnalyzed) 739 return BBI; 740 741 BBI.BB = BB; 742 BBI.IsBeingAnalyzed = true; 743 744 ScanInstructions(BBI); 745 746 // Unanalyzable or ends with fallthrough or unconditional branch. 747 if (!BBI.IsBrAnalyzable || BBI.BrCond.empty()) { 748 BBI.IsBeingAnalyzed = false; 749 BBI.IsAnalyzed = true; 750 return BBI; 751 } 752 753 // Do not ifcvt if either path is a back edge to the entry block. 754 if (BBI.TrueBB == BB || BBI.FalseBB == BB) { 755 BBI.IsBeingAnalyzed = false; 756 BBI.IsAnalyzed = true; 757 return BBI; 758 } 759 760 // Do not ifcvt if true and false fallthrough blocks are the same. 761 if (!BBI.FalseBB) { 762 BBI.IsBeingAnalyzed = false; 763 BBI.IsAnalyzed = true; 764 return BBI; 765 } 766 767 BBInfo &TrueBBI = AnalyzeBlock(BBI.TrueBB, Tokens); 768 BBInfo &FalseBBI = AnalyzeBlock(BBI.FalseBB, Tokens); 769 770 if (TrueBBI.IsDone && FalseBBI.IsDone) { 771 BBI.IsBeingAnalyzed = false; 772 BBI.IsAnalyzed = true; 773 return BBI; 774 } 775 776 SmallVector<MachineOperand, 4> RevCond(BBI.BrCond.begin(), BBI.BrCond.end()); 777 bool CanRevCond = !TII->ReverseBranchCondition(RevCond); 778 779 unsigned Dups = 0; 780 unsigned Dups2 = 0; 781 bool TNeedSub = TrueBBI.Predicate.size() > 0; 782 bool FNeedSub = FalseBBI.Predicate.size() > 0; 783 bool Enqueued = false; 784 785 // Try to predict the branch, using loop info to guide us. 786 // General heuristics are: 787 // - backedge -> 90% taken 788 // - early exit -> 20% taken 789 float Prediction = 0.5f; 790 MachineLoop *Loop = MLI->getLoopFor(BB); 791 if (Loop) { 792 if (TrueBBI.BB == Loop->getHeader()) 793 Prediction = 0.9f; 794 else if (FalseBBI.BB == Loop->getHeader()) 795 Prediction = 0.1f; 796 797 MachineLoop *TrueLoop = MLI->getLoopFor(TrueBBI.BB); 798 MachineLoop *FalseLoop = MLI->getLoopFor(FalseBBI.BB); 799 if (!TrueLoop || TrueLoop->getParentLoop() == Loop) 800 Prediction = 0.2f; 801 else if (!FalseLoop || FalseLoop->getParentLoop() == Loop) 802 Prediction = 0.8f; 803 } 804 805 if (CanRevCond && ValidDiamond(TrueBBI, FalseBBI, Dups, Dups2) && 806 MeetIfcvtSizeLimit(*TrueBBI.BB, TrueBBI.NonPredSize - (Dups + Dups2), 807 *FalseBBI.BB, FalseBBI.NonPredSize - (Dups + Dups2), 808 Prediction) && 809 FeasibilityAnalysis(TrueBBI, BBI.BrCond) && 810 FeasibilityAnalysis(FalseBBI, RevCond)) { 811 // Diamond: 812 // EBB 813 // / \_ 814 // | | 815 // TBB FBB 816 // \ / 817 // TailBB 818 // Note TailBB can be empty. 819 Tokens.push_back(new IfcvtToken(BBI, ICDiamond, TNeedSub|FNeedSub, Dups, 820 Dups2)); 821 Enqueued = true; 822 } 823 824 if (ValidTriangle(TrueBBI, FalseBBI, false, Dups) && 825 MeetIfcvtSizeLimit(*TrueBBI.BB, TrueBBI.NonPredSize, Prediction) && 826 FeasibilityAnalysis(TrueBBI, BBI.BrCond, true)) { 827 // Triangle: 828 // EBB 829 // | \_ 830 // | | 831 // | TBB 832 // | / 833 // FBB 834 Tokens.push_back(new IfcvtToken(BBI, ICTriangle, TNeedSub, Dups)); 835 Enqueued = true; 836 } 837 838 if (ValidTriangle(TrueBBI, FalseBBI, true, Dups) && 839 MeetIfcvtSizeLimit(*TrueBBI.BB, TrueBBI.NonPredSize, Prediction) && 840 FeasibilityAnalysis(TrueBBI, BBI.BrCond, true, true)) { 841 Tokens.push_back(new IfcvtToken(BBI, ICTriangleRev, TNeedSub, Dups)); 842 Enqueued = true; 843 } 844 845 if (ValidSimple(TrueBBI, Dups, Prediction) && 846 MeetIfcvtSizeLimit(*TrueBBI.BB, TrueBBI.NonPredSize, Prediction) && 847 FeasibilityAnalysis(TrueBBI, BBI.BrCond)) { 848 // Simple (split, no rejoin): 849 // EBB 850 // | \_ 851 // | | 852 // | TBB---> exit 853 // | 854 // FBB 855 Tokens.push_back(new IfcvtToken(BBI, ICSimple, TNeedSub, Dups)); 856 Enqueued = true; 857 } 858 859 if (CanRevCond) { 860 // Try the other path... 861 if (ValidTriangle(FalseBBI, TrueBBI, false, Dups) && 862 MeetIfcvtSizeLimit(*FalseBBI.BB, FalseBBI.NonPredSize,1.0-Prediction) && 863 FeasibilityAnalysis(FalseBBI, RevCond, true)) { 864 Tokens.push_back(new IfcvtToken(BBI, ICTriangleFalse, FNeedSub, Dups)); 865 Enqueued = true; 866 } 867 868 if (ValidTriangle(FalseBBI, TrueBBI, true, Dups) && 869 MeetIfcvtSizeLimit(*FalseBBI.BB, FalseBBI.NonPredSize,1.0-Prediction) && 870 FeasibilityAnalysis(FalseBBI, RevCond, true, true)) { 871 Tokens.push_back(new IfcvtToken(BBI, ICTriangleFRev, FNeedSub, Dups)); 872 Enqueued = true; 873 } 874 875 if (ValidSimple(FalseBBI, Dups, 1.0-Prediction) && 876 MeetIfcvtSizeLimit(*FalseBBI.BB, FalseBBI.NonPredSize,1.0-Prediction) && 877 FeasibilityAnalysis(FalseBBI, RevCond)) { 878 Tokens.push_back(new IfcvtToken(BBI, ICSimpleFalse, FNeedSub, Dups)); 879 Enqueued = true; 880 } 881 } 882 883 BBI.IsEnqueued = Enqueued; 884 BBI.IsBeingAnalyzed = false; 885 BBI.IsAnalyzed = true; 886 return BBI; 887} 888 889/// AnalyzeBlocks - Analyze all blocks and find entries for all if-conversion 890/// candidates. 891void IfConverter::AnalyzeBlocks(MachineFunction &MF, 892 std::vector<IfcvtToken*> &Tokens) { 893 std::set<MachineBasicBlock*> Visited; 894 for (unsigned i = 0, e = Roots.size(); i != e; ++i) { 895 for (idf_ext_iterator<MachineBasicBlock*> I=idf_ext_begin(Roots[i],Visited), 896 E = idf_ext_end(Roots[i], Visited); I != E; ++I) { 897 MachineBasicBlock *BB = *I; 898 AnalyzeBlock(BB, Tokens); 899 } 900 } 901 902 // Sort to favor more complex ifcvt scheme. 903 std::stable_sort(Tokens.begin(), Tokens.end(), IfcvtTokenCmp); 904} 905 906/// canFallThroughTo - Returns true either if ToBB is the next block after BB or 907/// that all the intervening blocks are empty (given BB can fall through to its 908/// next block). 909static bool canFallThroughTo(MachineBasicBlock *BB, MachineBasicBlock *ToBB) { 910 MachineFunction::iterator PI = BB; 911 MachineFunction::iterator I = llvm::next(PI); 912 MachineFunction::iterator TI = ToBB; 913 MachineFunction::iterator E = BB->getParent()->end(); 914 while (I != TI) { 915 // Check isSuccessor to avoid case where the next block is empty, but 916 // it's not a successor. 917 if (I == E || !I->empty() || !PI->isSuccessor(I)) 918 return false; 919 PI = I++; 920 } 921 return true; 922} 923 924/// InvalidatePreds - Invalidate predecessor BB info so it would be re-analyzed 925/// to determine if it can be if-converted. If predecessor is already enqueued, 926/// dequeue it! 927void IfConverter::InvalidatePreds(MachineBasicBlock *BB) { 928 for (MachineBasicBlock::pred_iterator PI = BB->pred_begin(), 929 E = BB->pred_end(); PI != E; ++PI) { 930 BBInfo &PBBI = BBAnalysis[(*PI)->getNumber()]; 931 if (PBBI.IsDone || PBBI.BB == BB) 932 continue; 933 PBBI.IsAnalyzed = false; 934 PBBI.IsEnqueued = false; 935 } 936} 937 938/// InsertUncondBranch - Inserts an unconditional branch from BB to ToBB. 939/// 940static void InsertUncondBranch(MachineBasicBlock *BB, MachineBasicBlock *ToBB, 941 const TargetInstrInfo *TII) { 942 DebugLoc dl; // FIXME: this is nowhere 943 SmallVector<MachineOperand, 0> NoCond; 944 TII->InsertBranch(*BB, ToBB, NULL, NoCond, dl); 945} 946 947/// RemoveExtraEdges - Remove true / false edges if either / both are no longer 948/// successors. 949void IfConverter::RemoveExtraEdges(BBInfo &BBI) { 950 MachineBasicBlock *TBB = NULL, *FBB = NULL; 951 SmallVector<MachineOperand, 4> Cond; 952 if (!TII->AnalyzeBranch(*BBI.BB, TBB, FBB, Cond)) 953 BBI.BB->CorrectExtraCFGEdges(TBB, FBB, !Cond.empty()); 954} 955 956/// InitPredRedefs / UpdatePredRedefs - Defs by predicated instructions are 957/// modeled as read + write (sort like two-address instructions). These 958/// routines track register liveness and add implicit uses to if-converted 959/// instructions to conform to the model. 960static void InitPredRedefs(MachineBasicBlock *BB, SmallSet<unsigned,4> &Redefs, 961 const TargetRegisterInfo *TRI) { 962 for (MachineBasicBlock::livein_iterator I = BB->livein_begin(), 963 E = BB->livein_end(); I != E; ++I) { 964 unsigned Reg = *I; 965 Redefs.insert(Reg); 966 for (const unsigned *Subreg = TRI->getSubRegisters(Reg); 967 *Subreg; ++Subreg) 968 Redefs.insert(*Subreg); 969 } 970} 971 972static void UpdatePredRedefs(MachineInstr *MI, SmallSet<unsigned,4> &Redefs, 973 const TargetRegisterInfo *TRI, 974 bool AddImpUse = false) { 975 SmallVector<unsigned, 4> Defs; 976 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) { 977 const MachineOperand &MO = MI->getOperand(i); 978 if (!MO.isReg()) 979 continue; 980 unsigned Reg = MO.getReg(); 981 if (!Reg) 982 continue; 983 if (MO.isDef()) 984 Defs.push_back(Reg); 985 else if (MO.isKill()) { 986 Redefs.erase(Reg); 987 for (const unsigned *SR = TRI->getSubRegisters(Reg); *SR; ++SR) 988 Redefs.erase(*SR); 989 } 990 } 991 for (unsigned i = 0, e = Defs.size(); i != e; ++i) { 992 unsigned Reg = Defs[i]; 993 if (Redefs.count(Reg)) { 994 if (AddImpUse) 995 // Treat predicated update as read + write. 996 MI->addOperand(MachineOperand::CreateReg(Reg, false/*IsDef*/, 997 true/*IsImp*/,false/*IsKill*/)); 998 } else { 999 Redefs.insert(Reg); 1000 for (const unsigned *SR = TRI->getSubRegisters(Reg); *SR; ++SR) 1001 Redefs.insert(*SR); 1002 } 1003 } 1004} 1005 1006static void UpdatePredRedefs(MachineBasicBlock::iterator I, 1007 MachineBasicBlock::iterator E, 1008 SmallSet<unsigned,4> &Redefs, 1009 const TargetRegisterInfo *TRI) { 1010 while (I != E) { 1011 UpdatePredRedefs(I, Redefs, TRI); 1012 ++I; 1013 } 1014} 1015 1016/// IfConvertSimple - If convert a simple (split, no rejoin) sub-CFG. 1017/// 1018bool IfConverter::IfConvertSimple(BBInfo &BBI, IfcvtKind Kind) { 1019 BBInfo &TrueBBI = BBAnalysis[BBI.TrueBB->getNumber()]; 1020 BBInfo &FalseBBI = BBAnalysis[BBI.FalseBB->getNumber()]; 1021 BBInfo *CvtBBI = &TrueBBI; 1022 BBInfo *NextBBI = &FalseBBI; 1023 1024 SmallVector<MachineOperand, 4> Cond(BBI.BrCond.begin(), BBI.BrCond.end()); 1025 if (Kind == ICSimpleFalse) 1026 std::swap(CvtBBI, NextBBI); 1027 1028 if (CvtBBI->IsDone || 1029 (CvtBBI->CannotBeCopied && CvtBBI->BB->pred_size() > 1)) { 1030 // Something has changed. It's no longer safe to predicate this block. 1031 BBI.IsAnalyzed = false; 1032 CvtBBI->IsAnalyzed = false; 1033 return false; 1034 } 1035 1036 if (Kind == ICSimpleFalse) 1037 if (TII->ReverseBranchCondition(Cond)) 1038 assert(false && "Unable to reverse branch condition!"); 1039 1040 // Initialize liveins to the first BB. These are potentiall redefined by 1041 // predicated instructions. 1042 SmallSet<unsigned, 4> Redefs; 1043 InitPredRedefs(CvtBBI->BB, Redefs, TRI); 1044 InitPredRedefs(NextBBI->BB, Redefs, TRI); 1045 1046 if (CvtBBI->BB->pred_size() > 1) { 1047 BBI.NonPredSize -= TII->RemoveBranch(*BBI.BB); 1048 // Copy instructions in the true block, predicate them, and add them to 1049 // the entry block. 1050 CopyAndPredicateBlock(BBI, *CvtBBI, Cond, Redefs); 1051 } else { 1052 PredicateBlock(*CvtBBI, CvtBBI->BB->end(), Cond, Redefs); 1053 1054 // Merge converted block into entry block. 1055 BBI.NonPredSize -= TII->RemoveBranch(*BBI.BB); 1056 MergeBlocks(BBI, *CvtBBI); 1057 } 1058 1059 bool IterIfcvt = true; 1060 if (!canFallThroughTo(BBI.BB, NextBBI->BB)) { 1061 InsertUncondBranch(BBI.BB, NextBBI->BB, TII); 1062 BBI.HasFallThrough = false; 1063 // Now ifcvt'd block will look like this: 1064 // BB: 1065 // ... 1066 // t, f = cmp 1067 // if t op 1068 // b BBf 1069 // 1070 // We cannot further ifcvt this block because the unconditional branch 1071 // will have to be predicated on the new condition, that will not be 1072 // available if cmp executes. 1073 IterIfcvt = false; 1074 } 1075 1076 RemoveExtraEdges(BBI); 1077 1078 // Update block info. BB can be iteratively if-converted. 1079 if (!IterIfcvt) 1080 BBI.IsDone = true; 1081 InvalidatePreds(BBI.BB); 1082 CvtBBI->IsDone = true; 1083 1084 // FIXME: Must maintain LiveIns. 1085 return true; 1086} 1087 1088/// IfConvertTriangle - If convert a triangle sub-CFG. 1089/// 1090bool IfConverter::IfConvertTriangle(BBInfo &BBI, IfcvtKind Kind) { 1091 BBInfo &TrueBBI = BBAnalysis[BBI.TrueBB->getNumber()]; 1092 BBInfo &FalseBBI = BBAnalysis[BBI.FalseBB->getNumber()]; 1093 BBInfo *CvtBBI = &TrueBBI; 1094 BBInfo *NextBBI = &FalseBBI; 1095 DebugLoc dl; // FIXME: this is nowhere 1096 1097 SmallVector<MachineOperand, 4> Cond(BBI.BrCond.begin(), BBI.BrCond.end()); 1098 if (Kind == ICTriangleFalse || Kind == ICTriangleFRev) 1099 std::swap(CvtBBI, NextBBI); 1100 1101 if (CvtBBI->IsDone || 1102 (CvtBBI->CannotBeCopied && CvtBBI->BB->pred_size() > 1)) { 1103 // Something has changed. It's no longer safe to predicate this block. 1104 BBI.IsAnalyzed = false; 1105 CvtBBI->IsAnalyzed = false; 1106 return false; 1107 } 1108 1109 if (Kind == ICTriangleFalse || Kind == ICTriangleFRev) 1110 if (TII->ReverseBranchCondition(Cond)) 1111 assert(false && "Unable to reverse branch condition!"); 1112 1113 if (Kind == ICTriangleRev || Kind == ICTriangleFRev) { 1114 if (ReverseBranchCondition(*CvtBBI)) { 1115 // BB has been changed, modify its predecessors (except for this 1116 // one) so they don't get ifcvt'ed based on bad intel. 1117 for (MachineBasicBlock::pred_iterator PI = CvtBBI->BB->pred_begin(), 1118 E = CvtBBI->BB->pred_end(); PI != E; ++PI) { 1119 MachineBasicBlock *PBB = *PI; 1120 if (PBB == BBI.BB) 1121 continue; 1122 BBInfo &PBBI = BBAnalysis[PBB->getNumber()]; 1123 if (PBBI.IsEnqueued) { 1124 PBBI.IsAnalyzed = false; 1125 PBBI.IsEnqueued = false; 1126 } 1127 } 1128 } 1129 } 1130 1131 // Initialize liveins to the first BB. These are potentially redefined by 1132 // predicated instructions. 1133 SmallSet<unsigned, 4> Redefs; 1134 InitPredRedefs(CvtBBI->BB, Redefs, TRI); 1135 InitPredRedefs(NextBBI->BB, Redefs, TRI); 1136 1137 bool HasEarlyExit = CvtBBI->FalseBB != NULL; 1138 if (CvtBBI->BB->pred_size() > 1) { 1139 BBI.NonPredSize -= TII->RemoveBranch(*BBI.BB); 1140 // Copy instructions in the true block, predicate them, and add them to 1141 // the entry block. 1142 CopyAndPredicateBlock(BBI, *CvtBBI, Cond, Redefs, true); 1143 } else { 1144 // Predicate the 'true' block after removing its branch. 1145 CvtBBI->NonPredSize -= TII->RemoveBranch(*CvtBBI->BB); 1146 PredicateBlock(*CvtBBI, CvtBBI->BB->end(), Cond, Redefs); 1147 1148 // Now merge the entry of the triangle with the true block. 1149 BBI.NonPredSize -= TII->RemoveBranch(*BBI.BB); 1150 MergeBlocks(BBI, *CvtBBI, false); 1151 } 1152 1153 // If 'true' block has a 'false' successor, add an exit branch to it. 1154 if (HasEarlyExit) { 1155 SmallVector<MachineOperand, 4> RevCond(CvtBBI->BrCond.begin(), 1156 CvtBBI->BrCond.end()); 1157 if (TII->ReverseBranchCondition(RevCond)) 1158 assert(false && "Unable to reverse branch condition!"); 1159 TII->InsertBranch(*BBI.BB, CvtBBI->FalseBB, NULL, RevCond, dl); 1160 BBI.BB->addSuccessor(CvtBBI->FalseBB); 1161 } 1162 1163 // Merge in the 'false' block if the 'false' block has no other 1164 // predecessors. Otherwise, add an unconditional branch to 'false'. 1165 bool FalseBBDead = false; 1166 bool IterIfcvt = true; 1167 bool isFallThrough = canFallThroughTo(BBI.BB, NextBBI->BB); 1168 if (!isFallThrough) { 1169 // Only merge them if the true block does not fallthrough to the false 1170 // block. By not merging them, we make it possible to iteratively 1171 // ifcvt the blocks. 1172 if (!HasEarlyExit && 1173 NextBBI->BB->pred_size() == 1 && !NextBBI->HasFallThrough) { 1174 MergeBlocks(BBI, *NextBBI); 1175 FalseBBDead = true; 1176 } else { 1177 InsertUncondBranch(BBI.BB, NextBBI->BB, TII); 1178 BBI.HasFallThrough = false; 1179 } 1180 // Mixed predicated and unpredicated code. This cannot be iteratively 1181 // predicated. 1182 IterIfcvt = false; 1183 } 1184 1185 RemoveExtraEdges(BBI); 1186 1187 // Update block info. BB can be iteratively if-converted. 1188 if (!IterIfcvt) 1189 BBI.IsDone = true; 1190 InvalidatePreds(BBI.BB); 1191 CvtBBI->IsDone = true; 1192 if (FalseBBDead) 1193 NextBBI->IsDone = true; 1194 1195 // FIXME: Must maintain LiveIns. 1196 return true; 1197} 1198 1199/// IfConvertDiamond - If convert a diamond sub-CFG. 1200/// 1201bool IfConverter::IfConvertDiamond(BBInfo &BBI, IfcvtKind Kind, 1202 unsigned NumDups1, unsigned NumDups2) { 1203 BBInfo &TrueBBI = BBAnalysis[BBI.TrueBB->getNumber()]; 1204 BBInfo &FalseBBI = BBAnalysis[BBI.FalseBB->getNumber()]; 1205 MachineBasicBlock *TailBB = TrueBBI.TrueBB; 1206 // True block must fall through or end with an unanalyzable terminator. 1207 if (!TailBB) { 1208 if (blockAlwaysFallThrough(TrueBBI)) 1209 TailBB = FalseBBI.TrueBB; 1210 assert((TailBB || !TrueBBI.IsBrAnalyzable) && "Unexpected!"); 1211 } 1212 1213 if (TrueBBI.IsDone || FalseBBI.IsDone || 1214 TrueBBI.BB->pred_size() > 1 || 1215 FalseBBI.BB->pred_size() > 1) { 1216 // Something has changed. It's no longer safe to predicate these blocks. 1217 BBI.IsAnalyzed = false; 1218 TrueBBI.IsAnalyzed = false; 1219 FalseBBI.IsAnalyzed = false; 1220 return false; 1221 } 1222 1223 // Put the predicated instructions from the 'true' block before the 1224 // instructions from the 'false' block, unless the true block would clobber 1225 // the predicate, in which case, do the opposite. 1226 BBInfo *BBI1 = &TrueBBI; 1227 BBInfo *BBI2 = &FalseBBI; 1228 SmallVector<MachineOperand, 4> RevCond(BBI.BrCond.begin(), BBI.BrCond.end()); 1229 if (TII->ReverseBranchCondition(RevCond)) 1230 assert(false && "Unable to reverse branch condition!"); 1231 SmallVector<MachineOperand, 4> *Cond1 = &BBI.BrCond; 1232 SmallVector<MachineOperand, 4> *Cond2 = &RevCond; 1233 1234 // Figure out the more profitable ordering. 1235 bool DoSwap = false; 1236 if (TrueBBI.ClobbersPred && !FalseBBI.ClobbersPred) 1237 DoSwap = true; 1238 else if (TrueBBI.ClobbersPred == FalseBBI.ClobbersPred) { 1239 if (TrueBBI.NonPredSize > FalseBBI.NonPredSize) 1240 DoSwap = true; 1241 } 1242 if (DoSwap) { 1243 std::swap(BBI1, BBI2); 1244 std::swap(Cond1, Cond2); 1245 } 1246 1247 // Remove the conditional branch from entry to the blocks. 1248 BBI.NonPredSize -= TII->RemoveBranch(*BBI.BB); 1249 1250 // Initialize liveins to the first BB. These are potentially redefined by 1251 // predicated instructions. 1252 SmallSet<unsigned, 4> Redefs; 1253 InitPredRedefs(BBI1->BB, Redefs, TRI); 1254 1255 // Remove the duplicated instructions at the beginnings of both paths. 1256 MachineBasicBlock::iterator DI1 = BBI1->BB->begin(); 1257 MachineBasicBlock::iterator DI2 = BBI2->BB->begin(); 1258 MachineBasicBlock::iterator DIE1 = BBI1->BB->end(); 1259 MachineBasicBlock::iterator DIE2 = BBI2->BB->end(); 1260 // Skip dbg_value instructions 1261 while (DI1 != DIE1 && DI1->isDebugValue()) 1262 ++DI1; 1263 while (DI2 != DIE2 && DI2->isDebugValue()) 1264 ++DI2; 1265 BBI1->NonPredSize -= NumDups1; 1266 BBI2->NonPredSize -= NumDups1; 1267 1268 // Skip past the dups on each side separately since there may be 1269 // differing dbg_value entries. 1270 for (unsigned i = 0; i < NumDups1; ++DI1) { 1271 if (!DI1->isDebugValue()) 1272 ++i; 1273 } 1274 while (NumDups1 != 0) { 1275 ++DI2; 1276 if (!DI2->isDebugValue()) 1277 --NumDups1; 1278 } 1279 1280 UpdatePredRedefs(BBI1->BB->begin(), DI1, Redefs, TRI); 1281 BBI.BB->splice(BBI.BB->end(), BBI1->BB, BBI1->BB->begin(), DI1); 1282 BBI2->BB->erase(BBI2->BB->begin(), DI2); 1283 1284 // Predicate the 'true' block after removing its branch. 1285 BBI1->NonPredSize -= TII->RemoveBranch(*BBI1->BB); 1286 DI1 = BBI1->BB->end(); 1287 for (unsigned i = 0; i != NumDups2; ) { 1288 // NumDups2 only counted non-dbg_value instructions, so this won't 1289 // run off the head of the list. 1290 assert (DI1 != BBI1->BB->begin()); 1291 --DI1; 1292 // skip dbg_value instructions 1293 if (!DI1->isDebugValue()) 1294 ++i; 1295 } 1296 BBI1->BB->erase(DI1, BBI1->BB->end()); 1297 PredicateBlock(*BBI1, BBI1->BB->end(), *Cond1, Redefs); 1298 1299 // Predicate the 'false' block. 1300 BBI2->NonPredSize -= TII->RemoveBranch(*BBI2->BB); 1301 DI2 = BBI2->BB->end(); 1302 while (NumDups2 != 0) { 1303 // NumDups2 only counted non-dbg_value instructions, so this won't 1304 // run off the head of the list. 1305 assert (DI2 != BBI2->BB->begin()); 1306 --DI2; 1307 // skip dbg_value instructions 1308 if (!DI2->isDebugValue()) 1309 --NumDups2; 1310 } 1311 PredicateBlock(*BBI2, DI2, *Cond2, Redefs); 1312 1313 // Merge the true block into the entry of the diamond. 1314 MergeBlocks(BBI, *BBI1, TailBB == 0); 1315 MergeBlocks(BBI, *BBI2, TailBB == 0); 1316 1317 // If the if-converted block falls through or unconditionally branches into 1318 // the tail block, and the tail block does not have other predecessors, then 1319 // fold the tail block in as well. Otherwise, unless it falls through to the 1320 // tail, add a unconditional branch to it. 1321 if (TailBB) { 1322 BBInfo TailBBI = BBAnalysis[TailBB->getNumber()]; 1323 bool CanMergeTail = !TailBBI.HasFallThrough; 1324 // There may still be a fall-through edge from BBI1 or BBI2 to TailBB; 1325 // check if there are any other predecessors besides those. 1326 unsigned NumPreds = TailBB->pred_size(); 1327 if (NumPreds > 1) 1328 CanMergeTail = false; 1329 else if (NumPreds == 1 && CanMergeTail) { 1330 MachineBasicBlock::pred_iterator PI = TailBB->pred_begin(); 1331 if (*PI != BBI1->BB && *PI != BBI2->BB) 1332 CanMergeTail = false; 1333 } 1334 if (CanMergeTail) { 1335 MergeBlocks(BBI, TailBBI); 1336 TailBBI.IsDone = true; 1337 } else { 1338 BBI.BB->addSuccessor(TailBB); 1339 InsertUncondBranch(BBI.BB, TailBB, TII); 1340 BBI.HasFallThrough = false; 1341 } 1342 } 1343 1344 // RemoveExtraEdges won't work if the block has an unanalyzable branch, 1345 // which can happen here if TailBB is unanalyzable and is merged, so 1346 // explicitly remove BBI1 and BBI2 as successors. 1347 BBI.BB->removeSuccessor(BBI1->BB); 1348 BBI.BB->removeSuccessor(BBI2->BB); 1349 RemoveExtraEdges(BBI); 1350 1351 // Update block info. 1352 BBI.IsDone = TrueBBI.IsDone = FalseBBI.IsDone = true; 1353 InvalidatePreds(BBI.BB); 1354 1355 // FIXME: Must maintain LiveIns. 1356 return true; 1357} 1358 1359/// PredicateBlock - Predicate instructions from the start of the block to the 1360/// specified end with the specified condition. 1361void IfConverter::PredicateBlock(BBInfo &BBI, 1362 MachineBasicBlock::iterator E, 1363 SmallVectorImpl<MachineOperand> &Cond, 1364 SmallSet<unsigned, 4> &Redefs) { 1365 for (MachineBasicBlock::iterator I = BBI.BB->begin(); I != E; ++I) { 1366 if (I->isDebugValue() || TII->isPredicated(I)) 1367 continue; 1368 if (!TII->PredicateInstruction(I, Cond)) { 1369#ifndef NDEBUG 1370 dbgs() << "Unable to predicate " << *I << "!\n"; 1371#endif 1372 llvm_unreachable(0); 1373 } 1374 1375 // If the predicated instruction now redefines a register as the result of 1376 // if-conversion, add an implicit kill. 1377 UpdatePredRedefs(I, Redefs, TRI, true); 1378 } 1379 1380 std::copy(Cond.begin(), Cond.end(), std::back_inserter(BBI.Predicate)); 1381 1382 BBI.IsAnalyzed = false; 1383 BBI.NonPredSize = 0; 1384 1385 ++NumIfConvBBs; 1386} 1387 1388/// CopyAndPredicateBlock - Copy and predicate instructions from source BB to 1389/// the destination block. Skip end of block branches if IgnoreBr is true. 1390void IfConverter::CopyAndPredicateBlock(BBInfo &ToBBI, BBInfo &FromBBI, 1391 SmallVectorImpl<MachineOperand> &Cond, 1392 SmallSet<unsigned, 4> &Redefs, 1393 bool IgnoreBr) { 1394 MachineFunction &MF = *ToBBI.BB->getParent(); 1395 1396 for (MachineBasicBlock::iterator I = FromBBI.BB->begin(), 1397 E = FromBBI.BB->end(); I != E; ++I) { 1398 const TargetInstrDesc &TID = I->getDesc(); 1399 // Do not copy the end of the block branches. 1400 if (IgnoreBr && TID.isBranch()) 1401 break; 1402 1403 MachineInstr *MI = MF.CloneMachineInstr(I); 1404 ToBBI.BB->insert(ToBBI.BB->end(), MI); 1405 unsigned NumOps = TII->getNumMicroOps(MI, InstrItins); 1406 ToBBI.NonPredSize += NumOps; 1407 1408 if (!TII->isPredicated(I) && !MI->isDebugValue()) { 1409 if (!TII->PredicateInstruction(MI, Cond)) { 1410#ifndef NDEBUG 1411 dbgs() << "Unable to predicate " << *I << "!\n"; 1412#endif 1413 llvm_unreachable(0); 1414 } 1415 } 1416 1417 // If the predicated instruction now redefines a register as the result of 1418 // if-conversion, add an implicit kill. 1419 UpdatePredRedefs(MI, Redefs, TRI, true); 1420 } 1421 1422 if (!IgnoreBr) { 1423 std::vector<MachineBasicBlock *> Succs(FromBBI.BB->succ_begin(), 1424 FromBBI.BB->succ_end()); 1425 MachineBasicBlock *NBB = getNextBlock(FromBBI.BB); 1426 MachineBasicBlock *FallThrough = FromBBI.HasFallThrough ? NBB : NULL; 1427 1428 for (unsigned i = 0, e = Succs.size(); i != e; ++i) { 1429 MachineBasicBlock *Succ = Succs[i]; 1430 // Fallthrough edge can't be transferred. 1431 if (Succ == FallThrough) 1432 continue; 1433 ToBBI.BB->addSuccessor(Succ); 1434 } 1435 } 1436 1437 std::copy(FromBBI.Predicate.begin(), FromBBI.Predicate.end(), 1438 std::back_inserter(ToBBI.Predicate)); 1439 std::copy(Cond.begin(), Cond.end(), std::back_inserter(ToBBI.Predicate)); 1440 1441 ToBBI.ClobbersPred |= FromBBI.ClobbersPred; 1442 ToBBI.IsAnalyzed = false; 1443 1444 ++NumDupBBs; 1445} 1446 1447/// MergeBlocks - Move all instructions from FromBB to the end of ToBB. 1448/// This will leave FromBB as an empty block, so remove all of its 1449/// successor edges except for the fall-through edge. If AddEdges is true, 1450/// i.e., when FromBBI's branch is being moved, add those successor edges to 1451/// ToBBI. 1452void IfConverter::MergeBlocks(BBInfo &ToBBI, BBInfo &FromBBI, bool AddEdges) { 1453 ToBBI.BB->splice(ToBBI.BB->end(), 1454 FromBBI.BB, FromBBI.BB->begin(), FromBBI.BB->end()); 1455 1456 std::vector<MachineBasicBlock *> Succs(FromBBI.BB->succ_begin(), 1457 FromBBI.BB->succ_end()); 1458 MachineBasicBlock *NBB = getNextBlock(FromBBI.BB); 1459 MachineBasicBlock *FallThrough = FromBBI.HasFallThrough ? NBB : NULL; 1460 1461 for (unsigned i = 0, e = Succs.size(); i != e; ++i) { 1462 MachineBasicBlock *Succ = Succs[i]; 1463 // Fallthrough edge can't be transferred. 1464 if (Succ == FallThrough) 1465 continue; 1466 FromBBI.BB->removeSuccessor(Succ); 1467 if (AddEdges) 1468 ToBBI.BB->addSuccessor(Succ); 1469 } 1470 1471 // Now FromBBI always falls through to the next block! 1472 if (NBB && !FromBBI.BB->isSuccessor(NBB)) 1473 FromBBI.BB->addSuccessor(NBB); 1474 1475 std::copy(FromBBI.Predicate.begin(), FromBBI.Predicate.end(), 1476 std::back_inserter(ToBBI.Predicate)); 1477 FromBBI.Predicate.clear(); 1478 1479 ToBBI.NonPredSize += FromBBI.NonPredSize; 1480 FromBBI.NonPredSize = 0; 1481 1482 ToBBI.ClobbersPred |= FromBBI.ClobbersPred; 1483 ToBBI.HasFallThrough = FromBBI.HasFallThrough; 1484 ToBBI.IsAnalyzed = false; 1485 FromBBI.IsAnalyzed = false; 1486} 1487