IfConversion.cpp revision fe60104ac97f3a8736dcfbfdf9547c7b7cc7b951
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/Target/TargetInstrInfo.h" 21#include "llvm/Target/TargetLowering.h" 22#include "llvm/Target/TargetMachine.h" 23#include "llvm/Target/TargetRegisterInfo.h" 24#include "llvm/Support/CommandLine.h" 25#include "llvm/Support/Debug.h" 26#include "llvm/Support/ErrorHandling.h" 27#include "llvm/Support/raw_ostream.h" 28#include "llvm/ADT/DepthFirstIterator.h" 29#include "llvm/ADT/Statistic.h" 30#include "llvm/ADT/STLExtras.h" 31using namespace llvm; 32 33// Hidden options for help debugging. 34static cl::opt<int> IfCvtFnStart("ifcvt-fn-start", cl::init(-1), cl::Hidden); 35static cl::opt<int> IfCvtFnStop("ifcvt-fn-stop", cl::init(-1), cl::Hidden); 36static cl::opt<int> IfCvtLimit("ifcvt-limit", cl::init(-1), cl::Hidden); 37static cl::opt<bool> DisableSimple("disable-ifcvt-simple", 38 cl::init(false), cl::Hidden); 39static cl::opt<bool> DisableSimpleF("disable-ifcvt-simple-false", 40 cl::init(false), cl::Hidden); 41static cl::opt<bool> DisableTriangle("disable-ifcvt-triangle", 42 cl::init(false), cl::Hidden); 43static cl::opt<bool> DisableTriangleR("disable-ifcvt-triangle-rev", 44 cl::init(false), cl::Hidden); 45static cl::opt<bool> DisableTriangleF("disable-ifcvt-triangle-false", 46 cl::init(false), cl::Hidden); 47static cl::opt<bool> DisableTriangleFR("disable-ifcvt-triangle-false-rev", 48 cl::init(false), cl::Hidden); 49static cl::opt<bool> DisableDiamond("disable-ifcvt-diamond", 50 cl::init(false), cl::Hidden); 51static cl::opt<bool> IfCvtBranchFold("ifcvt-branch-fold", 52 cl::init(true), cl::Hidden); 53 54STATISTIC(NumSimple, "Number of simple if-conversions performed"); 55STATISTIC(NumSimpleFalse, "Number of simple (F) if-conversions performed"); 56STATISTIC(NumTriangle, "Number of triangle if-conversions performed"); 57STATISTIC(NumTriangleRev, "Number of triangle (R) if-conversions performed"); 58STATISTIC(NumTriangleFalse,"Number of triangle (F) if-conversions performed"); 59STATISTIC(NumTriangleFRev, "Number of triangle (F/R) if-conversions performed"); 60STATISTIC(NumDiamonds, "Number of diamond if-conversions performed"); 61STATISTIC(NumIfConvBBs, "Number of if-converted blocks"); 62STATISTIC(NumDupBBs, "Number of duplicated blocks"); 63 64namespace { 65 class IfConverter : public MachineFunctionPass { 66 enum IfcvtKind { 67 ICNotClassfied, // BB data valid, but not classified. 68 ICSimpleFalse, // Same as ICSimple, but on the false path. 69 ICSimple, // BB is entry of an one split, no rejoin sub-CFG. 70 ICTriangleFRev, // Same as ICTriangleFalse, but false path rev condition. 71 ICTriangleRev, // Same as ICTriangle, but true path rev condition. 72 ICTriangleFalse, // Same as ICTriangle, but on the false path. 73 ICTriangle, // BB is entry of a triangle sub-CFG. 74 ICDiamond // BB is entry of a diamond sub-CFG. 75 }; 76 77 /// BBInfo - One per MachineBasicBlock, this is used to cache the result 78 /// if-conversion feasibility analysis. This includes results from 79 /// TargetInstrInfo::AnalyzeBranch() (i.e. TBB, FBB, and Cond), and its 80 /// classification, and common tail block of its successors (if it's a 81 /// diamond shape), its size, whether it's predicable, and whether any 82 /// instruction can clobber the 'would-be' predicate. 83 /// 84 /// IsDone - True if BB is not to be considered for ifcvt. 85 /// IsBeingAnalyzed - True if BB is currently being analyzed. 86 /// IsAnalyzed - True if BB has been analyzed (info is still valid). 87 /// IsEnqueued - True if BB has been enqueued to be ifcvt'ed. 88 /// IsBrAnalyzable - True if AnalyzeBranch() returns false. 89 /// HasFallThrough - True if BB may fallthrough to the following BB. 90 /// IsUnpredicable - True if BB is known to be unpredicable. 91 /// ClobbersPred - True if BB could modify predicates (e.g. has 92 /// cmp, call, etc.) 93 /// NonPredSize - Number of non-predicated instructions. 94 /// BB - Corresponding MachineBasicBlock. 95 /// TrueBB / FalseBB- See AnalyzeBranch(). 96 /// BrCond - Conditions for end of block conditional branches. 97 /// Predicate - Predicate used in the BB. 98 struct BBInfo { 99 bool IsDone : 1; 100 bool IsBeingAnalyzed : 1; 101 bool IsAnalyzed : 1; 102 bool IsEnqueued : 1; 103 bool IsBrAnalyzable : 1; 104 bool HasFallThrough : 1; 105 bool IsUnpredicable : 1; 106 bool CannotBeCopied : 1; 107 bool ClobbersPred : 1; 108 unsigned NonPredSize; 109 MachineBasicBlock *BB; 110 MachineBasicBlock *TrueBB; 111 MachineBasicBlock *FalseBB; 112 SmallVector<MachineOperand, 4> BrCond; 113 SmallVector<MachineOperand, 4> Predicate; 114 BBInfo() : IsDone(false), IsBeingAnalyzed(false), 115 IsAnalyzed(false), IsEnqueued(false), IsBrAnalyzable(false), 116 HasFallThrough(false), IsUnpredicable(false), 117 CannotBeCopied(false), ClobbersPred(false), NonPredSize(0), 118 BB(0), TrueBB(0), FalseBB(0) {} 119 }; 120 121 /// IfcvtToken - Record information about pending if-conversions to attempt: 122 /// BBI - Corresponding BBInfo. 123 /// Kind - Type of block. See IfcvtKind. 124 /// NeedSubsumption - True if the to-be-predicated BB has already been 125 /// predicated. 126 /// NumDups - Number of instructions that would be duplicated due 127 /// to this if-conversion. (For diamonds, the number of 128 /// identical instructions at the beginnings of both 129 /// paths). 130 /// NumDups2 - For diamonds, the number of identical instructions 131 /// at the ends of both paths. 132 struct IfcvtToken { 133 BBInfo &BBI; 134 IfcvtKind Kind; 135 bool NeedSubsumption; 136 unsigned NumDups; 137 unsigned NumDups2; 138 IfcvtToken(BBInfo &b, IfcvtKind k, bool s, unsigned d, unsigned d2 = 0) 139 : BBI(b), Kind(k), NeedSubsumption(s), NumDups(d), NumDups2(d2) {} 140 }; 141 142 /// Roots - Basic blocks that do not have successors. These are the starting 143 /// points of Graph traversal. 144 std::vector<MachineBasicBlock*> Roots; 145 146 /// BBAnalysis - Results of if-conversion feasibility analysis indexed by 147 /// basic block number. 148 std::vector<BBInfo> BBAnalysis; 149 150 const TargetLowering *TLI; 151 const TargetInstrInfo *TII; 152 const TargetRegisterInfo *TRI; 153 bool MadeChange; 154 int FnNum; 155 public: 156 static char ID; 157 IfConverter() : MachineFunctionPass(&ID), FnNum(-1) {} 158 159 virtual bool runOnMachineFunction(MachineFunction &MF); 160 virtual const char *getPassName() const { return "If Converter"; } 161 162 private: 163 bool ReverseBranchCondition(BBInfo &BBI); 164 bool ValidSimple(BBInfo &TrueBBI, unsigned &Dups) const; 165 bool ValidTriangle(BBInfo &TrueBBI, BBInfo &FalseBBI, 166 bool FalseBranch, unsigned &Dups) const; 167 bool ValidDiamond(BBInfo &TrueBBI, BBInfo &FalseBBI, 168 unsigned &Dups1, unsigned &Dups2) const; 169 void ScanInstructions(BBInfo &BBI); 170 BBInfo &AnalyzeBlock(MachineBasicBlock *BB, 171 std::vector<IfcvtToken*> &Tokens); 172 bool FeasibilityAnalysis(BBInfo &BBI, SmallVectorImpl<MachineOperand> &Cond, 173 bool isTriangle = false, bool RevBranch = false); 174 void AnalyzeBlocks(MachineFunction &MF, std::vector<IfcvtToken*> &Tokens); 175 void InvalidatePreds(MachineBasicBlock *BB); 176 void RemoveExtraEdges(BBInfo &BBI); 177 bool IfConvertSimple(BBInfo &BBI, IfcvtKind Kind); 178 bool IfConvertTriangle(BBInfo &BBI, IfcvtKind Kind); 179 bool IfConvertDiamond(BBInfo &BBI, IfcvtKind Kind, 180 unsigned NumDups1, unsigned NumDups2); 181 void PredicateBlock(BBInfo &BBI, 182 MachineBasicBlock::iterator E, 183 SmallVectorImpl<MachineOperand> &Cond, 184 SmallSet<unsigned, 4> &Redefs); 185 void CopyAndPredicateBlock(BBInfo &ToBBI, BBInfo &FromBBI, 186 SmallVectorImpl<MachineOperand> &Cond, 187 SmallSet<unsigned, 4> &Redefs, 188 bool IgnoreBr = false); 189 void MergeBlocks(BBInfo &ToBBI, BBInfo &FromBBI); 190 191 bool MeetIfcvtSizeLimit(unsigned Size) const { 192 return Size > 0 && Size <= TLI->getIfCvtBlockSizeLimit(); 193 } 194 195 // blockAlwaysFallThrough - Block ends without a terminator. 196 bool blockAlwaysFallThrough(BBInfo &BBI) const { 197 return BBI.IsBrAnalyzable && BBI.TrueBB == NULL; 198 } 199 200 // IfcvtTokenCmp - Used to sort if-conversion candidates. 201 static bool IfcvtTokenCmp(IfcvtToken *C1, IfcvtToken *C2) { 202 int Incr1 = (C1->Kind == ICDiamond) 203 ? -(int)(C1->NumDups + C1->NumDups2) : (int)C1->NumDups; 204 int Incr2 = (C2->Kind == ICDiamond) 205 ? -(int)(C2->NumDups + C2->NumDups2) : (int)C2->NumDups; 206 if (Incr1 > Incr2) 207 return true; 208 else if (Incr1 == Incr2) { 209 // Favors subsumption. 210 if (C1->NeedSubsumption == false && C2->NeedSubsumption == true) 211 return true; 212 else if (C1->NeedSubsumption == C2->NeedSubsumption) { 213 // Favors diamond over triangle, etc. 214 if ((unsigned)C1->Kind < (unsigned)C2->Kind) 215 return true; 216 else if (C1->Kind == C2->Kind) 217 return C1->BBI.BB->getNumber() < C2->BBI.BB->getNumber(); 218 } 219 } 220 return false; 221 } 222 }; 223 224 char IfConverter::ID = 0; 225} 226 227static RegisterPass<IfConverter> 228X("if-converter", "If Converter"); 229 230FunctionPass *llvm::createIfConverterPass() { return new IfConverter(); } 231 232bool IfConverter::runOnMachineFunction(MachineFunction &MF) { 233 TLI = MF.getTarget().getTargetLowering(); 234 TII = MF.getTarget().getInstrInfo(); 235 TRI = MF.getTarget().getRegisterInfo(); 236 if (!TII) return false; 237 238 // Tail merge tend to expose more if-conversion opportunities. 239 BranchFolder BF(true); 240 bool BFChange = BF.OptimizeFunction(MF, TII, 241 MF.getTarget().getRegisterInfo(), 242 getAnalysisIfAvailable<MachineModuleInfo>()); 243 244 DEBUG(dbgs() << "\nIfcvt: function (" << ++FnNum << ") \'" 245 << MF.getFunction()->getName() << "\'"); 246 247 if (FnNum < IfCvtFnStart || (IfCvtFnStop != -1 && FnNum > IfCvtFnStop)) { 248 DEBUG(dbgs() << " skipped\n"); 249 return false; 250 } 251 DEBUG(dbgs() << "\n"); 252 253 MF.RenumberBlocks(); 254 BBAnalysis.resize(MF.getNumBlockIDs()); 255 256 // Look for root nodes, i.e. blocks without successors. 257 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; ++I) 258 if (I->succ_empty()) 259 Roots.push_back(I); 260 261 std::vector<IfcvtToken*> Tokens; 262 MadeChange = false; 263 unsigned NumIfCvts = NumSimple + NumSimpleFalse + NumTriangle + 264 NumTriangleRev + NumTriangleFalse + NumTriangleFRev + NumDiamonds; 265 while (IfCvtLimit == -1 || (int)NumIfCvts < IfCvtLimit) { 266 // Do an initial analysis for each basic block and find all the potential 267 // candidates to perform if-conversion. 268 bool Change = false; 269 AnalyzeBlocks(MF, Tokens); 270 while (!Tokens.empty()) { 271 IfcvtToken *Token = Tokens.back(); 272 Tokens.pop_back(); 273 BBInfo &BBI = Token->BBI; 274 IfcvtKind Kind = Token->Kind; 275 unsigned NumDups = Token->NumDups; 276 unsigned NumDups2 = Token->NumDups2; 277 278 delete Token; 279 280 // If the block has been evicted out of the queue or it has already been 281 // marked dead (due to it being predicated), then skip it. 282 if (BBI.IsDone) 283 BBI.IsEnqueued = false; 284 if (!BBI.IsEnqueued) 285 continue; 286 287 BBI.IsEnqueued = false; 288 289 bool RetVal = false; 290 switch (Kind) { 291 default: assert(false && "Unexpected!"); 292 break; 293 case ICSimple: 294 case ICSimpleFalse: { 295 bool isFalse = Kind == ICSimpleFalse; 296 if ((isFalse && DisableSimpleF) || (!isFalse && DisableSimple)) break; 297 DEBUG(dbgs() << "Ifcvt (Simple" << (Kind == ICSimpleFalse ? 298 " false" : "") 299 << "): BB#" << BBI.BB->getNumber() << " (" 300 << ((Kind == ICSimpleFalse) 301 ? BBI.FalseBB->getNumber() 302 : BBI.TrueBB->getNumber()) << ") "); 303 RetVal = IfConvertSimple(BBI, Kind); 304 DEBUG(dbgs() << (RetVal ? "succeeded!" : "failed!") << "\n"); 305 if (RetVal) { 306 if (isFalse) ++NumSimpleFalse; 307 else ++NumSimple; 308 } 309 break; 310 } 311 case ICTriangle: 312 case ICTriangleRev: 313 case ICTriangleFalse: 314 case ICTriangleFRev: { 315 bool isFalse = Kind == ICTriangleFalse; 316 bool isRev = (Kind == ICTriangleRev || Kind == ICTriangleFRev); 317 if (DisableTriangle && !isFalse && !isRev) break; 318 if (DisableTriangleR && !isFalse && isRev) break; 319 if (DisableTriangleF && isFalse && !isRev) break; 320 if (DisableTriangleFR && isFalse && isRev) break; 321 DEBUG(dbgs() << "Ifcvt (Triangle"); 322 if (isFalse) 323 DEBUG(dbgs() << " false"); 324 if (isRev) 325 DEBUG(dbgs() << " rev"); 326 DEBUG(dbgs() << "): BB#" << BBI.BB->getNumber() << " (T:" 327 << BBI.TrueBB->getNumber() << ",F:" 328 << BBI.FalseBB->getNumber() << ") "); 329 RetVal = IfConvertTriangle(BBI, Kind); 330 DEBUG(dbgs() << (RetVal ? "succeeded!" : "failed!") << "\n"); 331 if (RetVal) { 332 if (isFalse) { 333 if (isRev) ++NumTriangleFRev; 334 else ++NumTriangleFalse; 335 } else { 336 if (isRev) ++NumTriangleRev; 337 else ++NumTriangle; 338 } 339 } 340 break; 341 } 342 case ICDiamond: { 343 if (DisableDiamond) break; 344 DEBUG(dbgs() << "Ifcvt (Diamond): BB#" << BBI.BB->getNumber() << " (T:" 345 << BBI.TrueBB->getNumber() << ",F:" 346 << BBI.FalseBB->getNumber() << ") "); 347 RetVal = IfConvertDiamond(BBI, Kind, NumDups, NumDups2); 348 DEBUG(dbgs() << (RetVal ? "succeeded!" : "failed!") << "\n"); 349 if (RetVal) ++NumDiamonds; 350 break; 351 } 352 } 353 354 Change |= RetVal; 355 356 NumIfCvts = NumSimple + NumSimpleFalse + NumTriangle + NumTriangleRev + 357 NumTriangleFalse + NumTriangleFRev + NumDiamonds; 358 if (IfCvtLimit != -1 && (int)NumIfCvts >= IfCvtLimit) 359 break; 360 } 361 362 if (!Change) 363 break; 364 MadeChange |= Change; 365 } 366 367 // Delete tokens in case of early exit. 368 while (!Tokens.empty()) { 369 IfcvtToken *Token = Tokens.back(); 370 Tokens.pop_back(); 371 delete Token; 372 } 373 374 Tokens.clear(); 375 Roots.clear(); 376 BBAnalysis.clear(); 377 378 if (MadeChange && IfCvtBranchFold) { 379 BranchFolder BF(false); 380 BF.OptimizeFunction(MF, TII, 381 MF.getTarget().getRegisterInfo(), 382 getAnalysisIfAvailable<MachineModuleInfo>()); 383 } 384 385 MadeChange |= BFChange; 386 return MadeChange; 387} 388 389/// findFalseBlock - BB has a fallthrough. Find its 'false' successor given 390/// its 'true' successor. 391static MachineBasicBlock *findFalseBlock(MachineBasicBlock *BB, 392 MachineBasicBlock *TrueBB) { 393 for (MachineBasicBlock::succ_iterator SI = BB->succ_begin(), 394 E = BB->succ_end(); SI != E; ++SI) { 395 MachineBasicBlock *SuccBB = *SI; 396 if (SuccBB != TrueBB) 397 return SuccBB; 398 } 399 return NULL; 400} 401 402/// ReverseBranchCondition - Reverse the condition of the end of the block 403/// branch. Swap block's 'true' and 'false' successors. 404bool IfConverter::ReverseBranchCondition(BBInfo &BBI) { 405 DebugLoc dl; // FIXME: this is nowhere 406 if (!TII->ReverseBranchCondition(BBI.BrCond)) { 407 TII->RemoveBranch(*BBI.BB); 408 TII->InsertBranch(*BBI.BB, BBI.FalseBB, BBI.TrueBB, BBI.BrCond, dl); 409 std::swap(BBI.TrueBB, BBI.FalseBB); 410 return true; 411 } 412 return false; 413} 414 415/// getNextBlock - Returns the next block in the function blocks ordering. If 416/// it is the end, returns NULL. 417static inline MachineBasicBlock *getNextBlock(MachineBasicBlock *BB) { 418 MachineFunction::iterator I = BB; 419 MachineFunction::iterator E = BB->getParent()->end(); 420 if (++I == E) 421 return NULL; 422 return I; 423} 424 425/// ValidSimple - Returns true if the 'true' block (along with its 426/// predecessor) forms a valid simple shape for ifcvt. It also returns the 427/// number of instructions that the ifcvt would need to duplicate if performed 428/// in Dups. 429bool IfConverter::ValidSimple(BBInfo &TrueBBI, unsigned &Dups) const { 430 Dups = 0; 431 if (TrueBBI.IsBeingAnalyzed || TrueBBI.IsDone) 432 return false; 433 434 if (TrueBBI.IsBrAnalyzable) 435 return false; 436 437 if (TrueBBI.BB->pred_size() > 1) { 438 if (TrueBBI.CannotBeCopied || 439 TrueBBI.NonPredSize > TLI->getIfCvtDupBlockSizeLimit()) 440 return false; 441 Dups = TrueBBI.NonPredSize; 442 } 443 444 return true; 445} 446 447/// ValidTriangle - Returns true if the 'true' and 'false' blocks (along 448/// with their common predecessor) forms a valid triangle shape for ifcvt. 449/// If 'FalseBranch' is true, it checks if 'true' block's false branch 450/// branches to the 'false' block rather than the other way around. It also 451/// returns the number of instructions that the ifcvt would need to duplicate 452/// if performed in 'Dups'. 453bool IfConverter::ValidTriangle(BBInfo &TrueBBI, BBInfo &FalseBBI, 454 bool FalseBranch, unsigned &Dups) const { 455 Dups = 0; 456 if (TrueBBI.IsBeingAnalyzed || TrueBBI.IsDone) 457 return false; 458 459 if (TrueBBI.BB->pred_size() > 1) { 460 if (TrueBBI.CannotBeCopied) 461 return false; 462 463 unsigned Size = TrueBBI.NonPredSize; 464 if (TrueBBI.IsBrAnalyzable) { 465 if (TrueBBI.TrueBB && TrueBBI.BrCond.empty()) 466 // Ends with an unconditional branch. It will be removed. 467 --Size; 468 else { 469 MachineBasicBlock *FExit = FalseBranch 470 ? TrueBBI.TrueBB : TrueBBI.FalseBB; 471 if (FExit) 472 // Require a conditional branch 473 ++Size; 474 } 475 } 476 if (Size > TLI->getIfCvtDupBlockSizeLimit()) 477 return false; 478 Dups = Size; 479 } 480 481 MachineBasicBlock *TExit = FalseBranch ? TrueBBI.FalseBB : TrueBBI.TrueBB; 482 if (!TExit && blockAlwaysFallThrough(TrueBBI)) { 483 MachineFunction::iterator I = TrueBBI.BB; 484 if (++I == TrueBBI.BB->getParent()->end()) 485 return false; 486 TExit = I; 487 } 488 return TExit && TExit == FalseBBI.BB; 489} 490 491static 492MachineBasicBlock::iterator firstNonBranchInst(MachineBasicBlock *BB, 493 const TargetInstrInfo *TII) { 494 MachineBasicBlock::iterator I = BB->end(); 495 while (I != BB->begin()) { 496 --I; 497 if (!I->getDesc().isBranch()) 498 break; 499 } 500 return I; 501} 502 503/// ValidDiamond - Returns true if the 'true' and 'false' blocks (along 504/// with their common predecessor) forms a valid diamond shape for ifcvt. 505bool IfConverter::ValidDiamond(BBInfo &TrueBBI, BBInfo &FalseBBI, 506 unsigned &Dups1, unsigned &Dups2) const { 507 Dups1 = Dups2 = 0; 508 if (TrueBBI.IsBeingAnalyzed || TrueBBI.IsDone || 509 FalseBBI.IsBeingAnalyzed || FalseBBI.IsDone) 510 return false; 511 512 MachineBasicBlock *TT = TrueBBI.TrueBB; 513 MachineBasicBlock *FT = FalseBBI.TrueBB; 514 515 if (!TT && blockAlwaysFallThrough(TrueBBI)) 516 TT = getNextBlock(TrueBBI.BB); 517 if (!FT && blockAlwaysFallThrough(FalseBBI)) 518 FT = getNextBlock(FalseBBI.BB); 519 if (TT != FT) 520 return false; 521 if (TT == NULL && (TrueBBI.IsBrAnalyzable || FalseBBI.IsBrAnalyzable)) 522 return false; 523 if (TrueBBI.BB->pred_size() > 1 || FalseBBI.BB->pred_size() > 1) 524 return false; 525 526 // FIXME: Allow true block to have an early exit? 527 if (TrueBBI.FalseBB || FalseBBI.FalseBB || 528 (TrueBBI.ClobbersPred && FalseBBI.ClobbersPred)) 529 return false; 530 531 MachineBasicBlock::iterator TI = TrueBBI.BB->begin(); 532 MachineBasicBlock::iterator FI = FalseBBI.BB->begin(); 533 MachineBasicBlock::iterator TIE = TrueBBI.BB->end(); 534 MachineBasicBlock::iterator FIE = FalseBBI.BB->end(); 535 // Skip dbg_value instructions 536 while (TI != TIE && TI->isDebugValue()) 537 ++TI; 538 while (FI != FIE && FI->isDebugValue()) 539 ++FI; 540 while (TI != TIE && FI != FIE) { 541 // Skip dbg_value instructions. These do not count. 542 if (TI->isDebugValue()) { 543 while (TI != TIE && TI->isDebugValue()) 544 ++TI; 545 if (TI == TIE) 546 break; 547 } 548 if (FI->isDebugValue()) { 549 while (FI != FIE && FI->isDebugValue()) 550 ++FI; 551 if (FI == FIE) 552 break; 553 } 554 if (!TI->isIdenticalTo(FI)) 555 break; 556 ++Dups1; 557 ++TI; 558 ++FI; 559 } 560 561 TI = firstNonBranchInst(TrueBBI.BB, TII); 562 FI = firstNonBranchInst(FalseBBI.BB, TII); 563 MachineBasicBlock::iterator TIB = TrueBBI.BB->begin(); 564 MachineBasicBlock::iterator FIB = FalseBBI.BB->begin(); 565 // Skip dbg_value instructions at end of the bb's. 566 while (TI != TIB && TI->isDebugValue()) 567 --TI; 568 while (FI != FIB && FI->isDebugValue()) 569 --FI; 570 while (TI != TIB && FI != FIB) { 571 // Skip dbg_value instructions. These do not count. 572 if (TI->isDebugValue()) { 573 while (TI != TIB && TI->isDebugValue()) 574 --TI; 575 if (TI == TIB) 576 break; 577 } 578 if (FI->isDebugValue()) { 579 while (FI != FIB && FI->isDebugValue()) 580 --FI; 581 if (FI == FIB) 582 break; 583 } 584 if (!TI->isIdenticalTo(FI)) 585 break; 586 ++Dups2; 587 --TI; 588 --FI; 589 } 590 591 return true; 592} 593 594/// ScanInstructions - Scan all the instructions in the block to determine if 595/// the block is predicable. In most cases, that means all the instructions 596/// in the block are isPredicable(). Also checks if the block contains any 597/// instruction which can clobber a predicate (e.g. condition code register). 598/// If so, the block is not predicable unless it's the last instruction. 599void IfConverter::ScanInstructions(BBInfo &BBI) { 600 if (BBI.IsDone) 601 return; 602 603 bool AlreadyPredicated = BBI.Predicate.size() > 0; 604 // First analyze the end of BB branches. 605 BBI.TrueBB = BBI.FalseBB = NULL; 606 BBI.BrCond.clear(); 607 BBI.IsBrAnalyzable = 608 !TII->AnalyzeBranch(*BBI.BB, BBI.TrueBB, BBI.FalseBB, BBI.BrCond); 609 BBI.HasFallThrough = BBI.IsBrAnalyzable && BBI.FalseBB == NULL; 610 611 if (BBI.BrCond.size()) { 612 // No false branch. This BB must end with a conditional branch and a 613 // fallthrough. 614 if (!BBI.FalseBB) 615 BBI.FalseBB = findFalseBlock(BBI.BB, BBI.TrueBB); 616 if (!BBI.FalseBB) { 617 // Malformed bcc? True and false blocks are the same? 618 BBI.IsUnpredicable = true; 619 return; 620 } 621 } 622 623 // Then scan all the instructions. 624 BBI.NonPredSize = 0; 625 BBI.ClobbersPred = false; 626 for (MachineBasicBlock::iterator I = BBI.BB->begin(), E = BBI.BB->end(); 627 I != E; ++I) { 628 if (I->isDebugValue()) 629 continue; 630 631 const TargetInstrDesc &TID = I->getDesc(); 632 if (TID.isNotDuplicable()) 633 BBI.CannotBeCopied = true; 634 635 bool isPredicated = TII->isPredicated(I); 636 bool isCondBr = BBI.IsBrAnalyzable && TID.isConditionalBranch(); 637 638 if (!isCondBr) { 639 if (!isPredicated) 640 BBI.NonPredSize++; 641 else if (!AlreadyPredicated) { 642 // FIXME: This instruction is already predicated before the 643 // if-conversion pass. It's probably something like a conditional move. 644 // Mark this block unpredicable for now. 645 BBI.IsUnpredicable = true; 646 return; 647 } 648 } 649 650 if (BBI.ClobbersPred && !isPredicated) { 651 // Predicate modification instruction should end the block (except for 652 // already predicated instructions and end of block branches). 653 if (isCondBr) { 654 // A conditional branch is not predicable, but it may be eliminated. 655 continue; 656 } 657 658 // Predicate may have been modified, the subsequent (currently) 659 // unpredicated instructions cannot be correctly predicated. 660 BBI.IsUnpredicable = true; 661 return; 662 } 663 664 // FIXME: Make use of PredDefs? e.g. ADDC, SUBC sets predicates but are 665 // still potentially predicable. 666 std::vector<MachineOperand> PredDefs; 667 if (TII->DefinesPredicate(I, PredDefs)) 668 BBI.ClobbersPred = true; 669 670 if (!TII->isPredicable(I)) { 671 BBI.IsUnpredicable = true; 672 return; 673 } 674 } 675} 676 677/// FeasibilityAnalysis - Determine if the block is a suitable candidate to be 678/// predicated by the specified predicate. 679bool IfConverter::FeasibilityAnalysis(BBInfo &BBI, 680 SmallVectorImpl<MachineOperand> &Pred, 681 bool isTriangle, bool RevBranch) { 682 // If the block is dead or unpredicable, then it cannot be predicated. 683 if (BBI.IsDone || BBI.IsUnpredicable) 684 return false; 685 686 // If it is already predicated, check if its predicate subsumes the new 687 // predicate. 688 if (BBI.Predicate.size() && !TII->SubsumesPredicate(BBI.Predicate, Pred)) 689 return false; 690 691 if (BBI.BrCond.size()) { 692 if (!isTriangle) 693 return false; 694 695 // Test predicate subsumption. 696 SmallVector<MachineOperand, 4> RevPred(Pred.begin(), Pred.end()); 697 SmallVector<MachineOperand, 4> Cond(BBI.BrCond.begin(), BBI.BrCond.end()); 698 if (RevBranch) { 699 if (TII->ReverseBranchCondition(Cond)) 700 return false; 701 } 702 if (TII->ReverseBranchCondition(RevPred) || 703 !TII->SubsumesPredicate(Cond, RevPred)) 704 return false; 705 } 706 707 return true; 708} 709 710/// AnalyzeBlock - Analyze the structure of the sub-CFG starting from 711/// the specified block. Record its successors and whether it looks like an 712/// if-conversion candidate. 713IfConverter::BBInfo &IfConverter::AnalyzeBlock(MachineBasicBlock *BB, 714 std::vector<IfcvtToken*> &Tokens) { 715 BBInfo &BBI = BBAnalysis[BB->getNumber()]; 716 717 if (BBI.IsAnalyzed || BBI.IsBeingAnalyzed) 718 return BBI; 719 720 BBI.BB = BB; 721 BBI.IsBeingAnalyzed = true; 722 723 ScanInstructions(BBI); 724 725 // Unanalyzable or ends with fallthrough or unconditional branch. 726 if (!BBI.IsBrAnalyzable || BBI.BrCond.empty()) { 727 BBI.IsBeingAnalyzed = false; 728 BBI.IsAnalyzed = true; 729 return BBI; 730 } 731 732 // Do not ifcvt if either path is a back edge to the entry block. 733 if (BBI.TrueBB == BB || BBI.FalseBB == BB) { 734 BBI.IsBeingAnalyzed = false; 735 BBI.IsAnalyzed = true; 736 return BBI; 737 } 738 739 // Do not ifcvt if true and false fallthrough blocks are the same. 740 if (!BBI.FalseBB) { 741 BBI.IsBeingAnalyzed = false; 742 BBI.IsAnalyzed = true; 743 return BBI; 744 } 745 746 BBInfo &TrueBBI = AnalyzeBlock(BBI.TrueBB, Tokens); 747 BBInfo &FalseBBI = AnalyzeBlock(BBI.FalseBB, Tokens); 748 749 if (TrueBBI.IsDone && FalseBBI.IsDone) { 750 BBI.IsBeingAnalyzed = false; 751 BBI.IsAnalyzed = true; 752 return BBI; 753 } 754 755 SmallVector<MachineOperand, 4> RevCond(BBI.BrCond.begin(), BBI.BrCond.end()); 756 bool CanRevCond = !TII->ReverseBranchCondition(RevCond); 757 758 unsigned Dups = 0; 759 unsigned Dups2 = 0; 760 bool TNeedSub = TrueBBI.Predicate.size() > 0; 761 bool FNeedSub = FalseBBI.Predicate.size() > 0; 762 bool Enqueued = false; 763 if (CanRevCond && ValidDiamond(TrueBBI, FalseBBI, Dups, Dups2) && 764 MeetIfcvtSizeLimit(TrueBBI.NonPredSize - (Dups + Dups2)) && 765 MeetIfcvtSizeLimit(FalseBBI.NonPredSize - (Dups + Dups2)) && 766 FeasibilityAnalysis(TrueBBI, BBI.BrCond) && 767 FeasibilityAnalysis(FalseBBI, RevCond)) { 768 // Diamond: 769 // EBB 770 // / \_ 771 // | | 772 // TBB FBB 773 // \ / 774 // TailBB 775 // Note TailBB can be empty. 776 Tokens.push_back(new IfcvtToken(BBI, ICDiamond, TNeedSub|FNeedSub, Dups, 777 Dups2)); 778 Enqueued = true; 779 } 780 781 if (ValidTriangle(TrueBBI, FalseBBI, false, Dups) && 782 MeetIfcvtSizeLimit(TrueBBI.NonPredSize) && 783 FeasibilityAnalysis(TrueBBI, BBI.BrCond, true)) { 784 // Triangle: 785 // EBB 786 // | \_ 787 // | | 788 // | TBB 789 // | / 790 // FBB 791 Tokens.push_back(new IfcvtToken(BBI, ICTriangle, TNeedSub, Dups)); 792 Enqueued = true; 793 } 794 795 if (ValidTriangle(TrueBBI, FalseBBI, true, Dups) && 796 MeetIfcvtSizeLimit(TrueBBI.NonPredSize) && 797 FeasibilityAnalysis(TrueBBI, BBI.BrCond, true, true)) { 798 Tokens.push_back(new IfcvtToken(BBI, ICTriangleRev, TNeedSub, Dups)); 799 Enqueued = true; 800 } 801 802 if (ValidSimple(TrueBBI, Dups) && 803 MeetIfcvtSizeLimit(TrueBBI.NonPredSize) && 804 FeasibilityAnalysis(TrueBBI, BBI.BrCond)) { 805 // Simple (split, no rejoin): 806 // EBB 807 // | \_ 808 // | | 809 // | TBB---> exit 810 // | 811 // FBB 812 Tokens.push_back(new IfcvtToken(BBI, ICSimple, TNeedSub, Dups)); 813 Enqueued = true; 814 } 815 816 if (CanRevCond) { 817 // Try the other path... 818 if (ValidTriangle(FalseBBI, TrueBBI, false, Dups) && 819 MeetIfcvtSizeLimit(FalseBBI.NonPredSize) && 820 FeasibilityAnalysis(FalseBBI, RevCond, true)) { 821 Tokens.push_back(new IfcvtToken(BBI, ICTriangleFalse, FNeedSub, Dups)); 822 Enqueued = true; 823 } 824 825 if (ValidTriangle(FalseBBI, TrueBBI, true, Dups) && 826 MeetIfcvtSizeLimit(FalseBBI.NonPredSize) && 827 FeasibilityAnalysis(FalseBBI, RevCond, true, true)) { 828 Tokens.push_back(new IfcvtToken(BBI, ICTriangleFRev, FNeedSub, Dups)); 829 Enqueued = true; 830 } 831 832 if (ValidSimple(FalseBBI, Dups) && 833 MeetIfcvtSizeLimit(FalseBBI.NonPredSize) && 834 FeasibilityAnalysis(FalseBBI, RevCond)) { 835 Tokens.push_back(new IfcvtToken(BBI, ICSimpleFalse, FNeedSub, Dups)); 836 Enqueued = true; 837 } 838 } 839 840 BBI.IsEnqueued = Enqueued; 841 BBI.IsBeingAnalyzed = false; 842 BBI.IsAnalyzed = true; 843 return BBI; 844} 845 846/// AnalyzeBlocks - Analyze all blocks and find entries for all if-conversion 847/// candidates. 848void IfConverter::AnalyzeBlocks(MachineFunction &MF, 849 std::vector<IfcvtToken*> &Tokens) { 850 std::set<MachineBasicBlock*> Visited; 851 for (unsigned i = 0, e = Roots.size(); i != e; ++i) { 852 for (idf_ext_iterator<MachineBasicBlock*> I=idf_ext_begin(Roots[i],Visited), 853 E = idf_ext_end(Roots[i], Visited); I != E; ++I) { 854 MachineBasicBlock *BB = *I; 855 AnalyzeBlock(BB, Tokens); 856 } 857 } 858 859 // Sort to favor more complex ifcvt scheme. 860 std::stable_sort(Tokens.begin(), Tokens.end(), IfcvtTokenCmp); 861} 862 863/// canFallThroughTo - Returns true either if ToBB is the next block after BB or 864/// that all the intervening blocks are empty (given BB can fall through to its 865/// next block). 866static bool canFallThroughTo(MachineBasicBlock *BB, MachineBasicBlock *ToBB) { 867 MachineFunction::iterator PI = BB; 868 MachineFunction::iterator I = llvm::next(PI); 869 MachineFunction::iterator TI = ToBB; 870 MachineFunction::iterator E = BB->getParent()->end(); 871 while (I != TI) { 872 // Check isSuccessor to avoid case where the next block is empty, but 873 // it's not a successor. 874 if (I == E || !I->empty() || !PI->isSuccessor(I)) 875 return false; 876 PI = I++; 877 } 878 return true; 879} 880 881/// InvalidatePreds - Invalidate predecessor BB info so it would be re-analyzed 882/// to determine if it can be if-converted. If predecessor is already enqueued, 883/// dequeue it! 884void IfConverter::InvalidatePreds(MachineBasicBlock *BB) { 885 for (MachineBasicBlock::pred_iterator PI = BB->pred_begin(), 886 E = BB->pred_end(); PI != E; ++PI) { 887 BBInfo &PBBI = BBAnalysis[(*PI)->getNumber()]; 888 if (PBBI.IsDone || PBBI.BB == BB) 889 continue; 890 PBBI.IsAnalyzed = false; 891 PBBI.IsEnqueued = false; 892 } 893} 894 895/// InsertUncondBranch - Inserts an unconditional branch from BB to ToBB. 896/// 897static void InsertUncondBranch(MachineBasicBlock *BB, MachineBasicBlock *ToBB, 898 const TargetInstrInfo *TII) { 899 DebugLoc dl; // FIXME: this is nowhere 900 SmallVector<MachineOperand, 0> NoCond; 901 TII->InsertBranch(*BB, ToBB, NULL, NoCond, dl); 902} 903 904/// RemoveExtraEdges - Remove true / false edges if either / both are no longer 905/// successors. 906void IfConverter::RemoveExtraEdges(BBInfo &BBI) { 907 MachineBasicBlock *TBB = NULL, *FBB = NULL; 908 SmallVector<MachineOperand, 4> Cond; 909 if (!TII->AnalyzeBranch(*BBI.BB, TBB, FBB, Cond)) 910 BBI.BB->CorrectExtraCFGEdges(TBB, FBB, !Cond.empty()); 911} 912 913/// InitPredRedefs / UpdatePredRedefs - Defs by predicated instructions are 914/// modeled as read + write (sort like two-address instructions). These 915/// routines track register liveness and add implicit uses to if-converted 916/// instructions to conform to the model. 917static void InitPredRedefs(MachineBasicBlock *BB, SmallSet<unsigned,4> &Redefs, 918 const TargetRegisterInfo *TRI) { 919 for (MachineBasicBlock::livein_iterator I = BB->livein_begin(), 920 E = BB->livein_end(); I != E; ++I) { 921 unsigned Reg = *I; 922 Redefs.insert(Reg); 923 for (const unsigned *Subreg = TRI->getSubRegisters(Reg); 924 *Subreg; ++Subreg) 925 Redefs.insert(*Subreg); 926 } 927} 928 929static void UpdatePredRedefs(MachineInstr *MI, SmallSet<unsigned,4> &Redefs, 930 const TargetRegisterInfo *TRI, 931 bool AddImpUse = false) { 932 SmallVector<unsigned, 4> Defs; 933 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) { 934 const MachineOperand &MO = MI->getOperand(i); 935 if (!MO.isReg()) 936 continue; 937 unsigned Reg = MO.getReg(); 938 if (!Reg) 939 continue; 940 if (MO.isDef()) 941 Defs.push_back(Reg); 942 else if (MO.isKill()) { 943 Redefs.erase(Reg); 944 for (const unsigned *SR = TRI->getSubRegisters(Reg); *SR; ++SR) 945 Redefs.erase(*SR); 946 } 947 } 948 for (unsigned i = 0, e = Defs.size(); i != e; ++i) { 949 unsigned Reg = Defs[i]; 950 if (Redefs.count(Reg)) { 951 if (AddImpUse) 952 // Treat predicated update as read + write. 953 MI->addOperand(MachineOperand::CreateReg(Reg, false/*IsDef*/, 954 true/*IsImp*/,false/*IsKill*/)); 955 } else { 956 Redefs.insert(Reg); 957 for (const unsigned *SR = TRI->getSubRegisters(Reg); *SR; ++SR) 958 Redefs.insert(*SR); 959 } 960 } 961} 962 963static void UpdatePredRedefs(MachineBasicBlock::iterator I, 964 MachineBasicBlock::iterator E, 965 SmallSet<unsigned,4> &Redefs, 966 const TargetRegisterInfo *TRI) { 967 while (I != E) { 968 UpdatePredRedefs(I, Redefs, TRI); 969 ++I; 970 } 971} 972 973/// IfConvertSimple - If convert a simple (split, no rejoin) sub-CFG. 974/// 975bool IfConverter::IfConvertSimple(BBInfo &BBI, IfcvtKind Kind) { 976 BBInfo &TrueBBI = BBAnalysis[BBI.TrueBB->getNumber()]; 977 BBInfo &FalseBBI = BBAnalysis[BBI.FalseBB->getNumber()]; 978 BBInfo *CvtBBI = &TrueBBI; 979 BBInfo *NextBBI = &FalseBBI; 980 981 SmallVector<MachineOperand, 4> Cond(BBI.BrCond.begin(), BBI.BrCond.end()); 982 if (Kind == ICSimpleFalse) 983 std::swap(CvtBBI, NextBBI); 984 985 if (CvtBBI->IsDone || 986 (CvtBBI->CannotBeCopied && CvtBBI->BB->pred_size() > 1)) { 987 // Something has changed. It's no longer safe to predicate this block. 988 BBI.IsAnalyzed = false; 989 CvtBBI->IsAnalyzed = false; 990 return false; 991 } 992 993 if (Kind == ICSimpleFalse) 994 if (TII->ReverseBranchCondition(Cond)) 995 assert(false && "Unable to reverse branch condition!"); 996 997 // Initialize liveins to the first BB. These are potentiall redefined by 998 // predicated instructions. 999 SmallSet<unsigned, 4> Redefs; 1000 InitPredRedefs(CvtBBI->BB, Redefs, TRI); 1001 InitPredRedefs(NextBBI->BB, Redefs, TRI); 1002 1003 if (CvtBBI->BB->pred_size() > 1) { 1004 BBI.NonPredSize -= TII->RemoveBranch(*BBI.BB); 1005 // Copy instructions in the true block, predicate them, and add them to 1006 // the entry block. 1007 CopyAndPredicateBlock(BBI, *CvtBBI, Cond, Redefs); 1008 } else { 1009 PredicateBlock(*CvtBBI, CvtBBI->BB->end(), Cond, Redefs); 1010 1011 // Merge converted block into entry block. 1012 BBI.NonPredSize -= TII->RemoveBranch(*BBI.BB); 1013 MergeBlocks(BBI, *CvtBBI); 1014 } 1015 1016 bool IterIfcvt = true; 1017 if (!canFallThroughTo(BBI.BB, NextBBI->BB)) { 1018 InsertUncondBranch(BBI.BB, NextBBI->BB, TII); 1019 BBI.HasFallThrough = false; 1020 // Now ifcvt'd block will look like this: 1021 // BB: 1022 // ... 1023 // t, f = cmp 1024 // if t op 1025 // b BBf 1026 // 1027 // We cannot further ifcvt this block because the unconditional branch 1028 // will have to be predicated on the new condition, that will not be 1029 // available if cmp executes. 1030 IterIfcvt = false; 1031 } 1032 1033 RemoveExtraEdges(BBI); 1034 1035 // Update block info. BB can be iteratively if-converted. 1036 if (!IterIfcvt) 1037 BBI.IsDone = true; 1038 InvalidatePreds(BBI.BB); 1039 CvtBBI->IsDone = true; 1040 1041 // FIXME: Must maintain LiveIns. 1042 return true; 1043} 1044 1045/// IfConvertTriangle - If convert a triangle sub-CFG. 1046/// 1047bool IfConverter::IfConvertTriangle(BBInfo &BBI, IfcvtKind Kind) { 1048 BBInfo &TrueBBI = BBAnalysis[BBI.TrueBB->getNumber()]; 1049 BBInfo &FalseBBI = BBAnalysis[BBI.FalseBB->getNumber()]; 1050 BBInfo *CvtBBI = &TrueBBI; 1051 BBInfo *NextBBI = &FalseBBI; 1052 DebugLoc dl; // FIXME: this is nowhere 1053 1054 SmallVector<MachineOperand, 4> Cond(BBI.BrCond.begin(), BBI.BrCond.end()); 1055 if (Kind == ICTriangleFalse || Kind == ICTriangleFRev) 1056 std::swap(CvtBBI, NextBBI); 1057 1058 if (CvtBBI->IsDone || 1059 (CvtBBI->CannotBeCopied && CvtBBI->BB->pred_size() > 1)) { 1060 // Something has changed. It's no longer safe to predicate this block. 1061 BBI.IsAnalyzed = false; 1062 CvtBBI->IsAnalyzed = false; 1063 return false; 1064 } 1065 1066 if (Kind == ICTriangleFalse || Kind == ICTriangleFRev) 1067 if (TII->ReverseBranchCondition(Cond)) 1068 assert(false && "Unable to reverse branch condition!"); 1069 1070 if (Kind == ICTriangleRev || Kind == ICTriangleFRev) { 1071 if (ReverseBranchCondition(*CvtBBI)) { 1072 // BB has been changed, modify its predecessors (except for this 1073 // one) so they don't get ifcvt'ed based on bad intel. 1074 for (MachineBasicBlock::pred_iterator PI = CvtBBI->BB->pred_begin(), 1075 E = CvtBBI->BB->pred_end(); PI != E; ++PI) { 1076 MachineBasicBlock *PBB = *PI; 1077 if (PBB == BBI.BB) 1078 continue; 1079 BBInfo &PBBI = BBAnalysis[PBB->getNumber()]; 1080 if (PBBI.IsEnqueued) { 1081 PBBI.IsAnalyzed = false; 1082 PBBI.IsEnqueued = false; 1083 } 1084 } 1085 } 1086 } 1087 1088 // Initialize liveins to the first BB. These are potentially redefined by 1089 // predicated instructions. 1090 SmallSet<unsigned, 4> Redefs; 1091 InitPredRedefs(CvtBBI->BB, Redefs, TRI); 1092 InitPredRedefs(NextBBI->BB, Redefs, TRI); 1093 1094 bool HasEarlyExit = CvtBBI->FalseBB != NULL; 1095 bool DupBB = CvtBBI->BB->pred_size() > 1; 1096 if (DupBB) { 1097 BBI.NonPredSize -= TII->RemoveBranch(*BBI.BB); 1098 // Copy instructions in the true block, predicate them, and add them to 1099 // the entry block. 1100 CopyAndPredicateBlock(BBI, *CvtBBI, Cond, Redefs, true); 1101 } else { 1102 // Predicate the 'true' block after removing its branch. 1103 CvtBBI->NonPredSize -= TII->RemoveBranch(*CvtBBI->BB); 1104 PredicateBlock(*CvtBBI, CvtBBI->BB->end(), Cond, Redefs); 1105 1106 // Now merge the entry of the triangle with the true block. 1107 BBI.NonPredSize -= TII->RemoveBranch(*BBI.BB); 1108 MergeBlocks(BBI, *CvtBBI); 1109 } 1110 1111 // If 'true' block has a 'false' successor, add an exit branch to it. 1112 if (HasEarlyExit) { 1113 SmallVector<MachineOperand, 4> RevCond(CvtBBI->BrCond.begin(), 1114 CvtBBI->BrCond.end()); 1115 if (TII->ReverseBranchCondition(RevCond)) 1116 assert(false && "Unable to reverse branch condition!"); 1117 TII->InsertBranch(*BBI.BB, CvtBBI->FalseBB, NULL, RevCond, dl); 1118 BBI.BB->addSuccessor(CvtBBI->FalseBB); 1119 } 1120 1121 // Merge in the 'false' block if the 'false' block has no other 1122 // predecessors. Otherwise, add an unconditional branch to 'false'. 1123 bool FalseBBDead = false; 1124 bool IterIfcvt = true; 1125 bool isFallThrough = canFallThroughTo(BBI.BB, NextBBI->BB); 1126 if (!isFallThrough) { 1127 // Only merge them if the true block does not fallthrough to the false 1128 // block. By not merging them, we make it possible to iteratively 1129 // ifcvt the blocks. 1130 if (!HasEarlyExit && 1131 NextBBI->BB->pred_size() == 1 && !NextBBI->HasFallThrough) { 1132 MergeBlocks(BBI, *NextBBI); 1133 FalseBBDead = true; 1134 } else { 1135 InsertUncondBranch(BBI.BB, NextBBI->BB, TII); 1136 BBI.HasFallThrough = false; 1137 } 1138 // Mixed predicated and unpredicated code. This cannot be iteratively 1139 // predicated. 1140 IterIfcvt = false; 1141 } 1142 1143 RemoveExtraEdges(BBI); 1144 1145 // Update block info. BB can be iteratively if-converted. 1146 if (!IterIfcvt) 1147 BBI.IsDone = true; 1148 InvalidatePreds(BBI.BB); 1149 CvtBBI->IsDone = true; 1150 if (FalseBBDead) 1151 NextBBI->IsDone = true; 1152 1153 // FIXME: Must maintain LiveIns. 1154 return true; 1155} 1156 1157/// IfConvertDiamond - If convert a diamond sub-CFG. 1158/// 1159bool IfConverter::IfConvertDiamond(BBInfo &BBI, IfcvtKind Kind, 1160 unsigned NumDups1, unsigned NumDups2) { 1161 BBInfo &TrueBBI = BBAnalysis[BBI.TrueBB->getNumber()]; 1162 BBInfo &FalseBBI = BBAnalysis[BBI.FalseBB->getNumber()]; 1163 MachineBasicBlock *TailBB = TrueBBI.TrueBB; 1164 // True block must fall through or end with an unanalyzable terminator. 1165 if (!TailBB) { 1166 if (blockAlwaysFallThrough(TrueBBI)) 1167 TailBB = FalseBBI.TrueBB; 1168 assert((TailBB || !TrueBBI.IsBrAnalyzable) && "Unexpected!"); 1169 } 1170 1171 if (TrueBBI.IsDone || FalseBBI.IsDone || 1172 TrueBBI.BB->pred_size() > 1 || 1173 FalseBBI.BB->pred_size() > 1) { 1174 // Something has changed. It's no longer safe to predicate these blocks. 1175 BBI.IsAnalyzed = false; 1176 TrueBBI.IsAnalyzed = false; 1177 FalseBBI.IsAnalyzed = false; 1178 return false; 1179 } 1180 1181 // Merge the 'true' and 'false' blocks by copying the instructions 1182 // from the 'false' block to the 'true' block. That is, unless the true 1183 // block would clobber the predicate, in that case, do the opposite. 1184 BBInfo *BBI1 = &TrueBBI; 1185 BBInfo *BBI2 = &FalseBBI; 1186 SmallVector<MachineOperand, 4> RevCond(BBI.BrCond.begin(), BBI.BrCond.end()); 1187 if (TII->ReverseBranchCondition(RevCond)) 1188 assert(false && "Unable to reverse branch condition!"); 1189 SmallVector<MachineOperand, 4> *Cond1 = &BBI.BrCond; 1190 SmallVector<MachineOperand, 4> *Cond2 = &RevCond; 1191 1192 // Figure out the more profitable ordering. 1193 bool DoSwap = false; 1194 if (TrueBBI.ClobbersPred && !FalseBBI.ClobbersPred) 1195 DoSwap = true; 1196 else if (TrueBBI.ClobbersPred == FalseBBI.ClobbersPred) { 1197 if (TrueBBI.NonPredSize > FalseBBI.NonPredSize) 1198 DoSwap = true; 1199 } 1200 if (DoSwap) { 1201 std::swap(BBI1, BBI2); 1202 std::swap(Cond1, Cond2); 1203 } 1204 1205 // Remove the conditional branch from entry to the blocks. 1206 BBI.NonPredSize -= TII->RemoveBranch(*BBI.BB); 1207 1208 // Initialize liveins to the first BB. These are potentiall redefined by 1209 // predicated instructions. 1210 SmallSet<unsigned, 4> Redefs; 1211 InitPredRedefs(BBI1->BB, Redefs, TRI); 1212 1213 // Remove the duplicated instructions at the beginnings of both paths. 1214 MachineBasicBlock::iterator DI1 = BBI1->BB->begin(); 1215 MachineBasicBlock::iterator DI2 = BBI2->BB->begin(); 1216 MachineBasicBlock::iterator DIE1 = BBI1->BB->end(); 1217 MachineBasicBlock::iterator DIE2 = BBI2->BB->end(); 1218 // Skip dbg_value instructions 1219 while (DI1 != DIE1 && DI1->isDebugValue()) 1220 ++DI1; 1221 while (DI2 != DIE2 && DI2->isDebugValue()) 1222 ++DI2; 1223 BBI1->NonPredSize -= NumDups1; 1224 BBI2->NonPredSize -= NumDups1; 1225 while (NumDups1 != 0) { 1226 ++DI1; 1227 ++DI2; 1228 --NumDups1; 1229 } 1230 1231 UpdatePredRedefs(BBI1->BB->begin(), DI1, Redefs, TRI); 1232 BBI.BB->splice(BBI.BB->end(), BBI1->BB, BBI1->BB->begin(), DI1); 1233 BBI2->BB->erase(BBI2->BB->begin(), DI2); 1234 1235 // Predicate the 'true' block after removing its branch. 1236 BBI1->NonPredSize -= TII->RemoveBranch(*BBI1->BB); 1237 DI1 = BBI1->BB->end(); 1238 for (unsigned i = 0; i != NumDups2; ) { 1239 // NumDups2 only counted non-dbg_value instructions, so this won't 1240 // run off the head of the list. 1241 assert (DI1 != BBI1->BB->begin()); 1242 --DI1; 1243 // skip dbg_value instructions 1244 if (!DI1->isDebugValue()) 1245 ++i; 1246 } 1247 BBI1->BB->erase(DI1, BBI1->BB->end()); 1248 PredicateBlock(*BBI1, BBI1->BB->end(), *Cond1, Redefs); 1249 1250 // Predicate the 'false' block. 1251 BBI2->NonPredSize -= TII->RemoveBranch(*BBI2->BB); 1252 DI2 = BBI2->BB->end(); 1253 while (NumDups2 != 0) { 1254 // NumDups2 only counted non-dbg_value instructions, so this won't 1255 // run off the head of the list. 1256 assert (DI2 != BBI2->BB->begin()); 1257 --DI2; 1258 // skip dbg_value instructions 1259 if (!DI2->isDebugValue()) 1260 --NumDups2; 1261 } 1262 PredicateBlock(*BBI2, DI2, *Cond2, Redefs); 1263 1264 // Merge the true block into the entry of the diamond. 1265 MergeBlocks(BBI, *BBI1); 1266 MergeBlocks(BBI, *BBI2); 1267 1268 // If the if-converted block falls through or unconditionally branches into 1269 // the tail block, and the tail block does not have other predecessors, then 1270 // fold the tail block in as well. Otherwise, unless it falls through to the 1271 // tail, add a unconditional branch to it. 1272 if (TailBB) { 1273 BBInfo TailBBI = BBAnalysis[TailBB->getNumber()]; 1274 if (TailBB->pred_size() == 1 && !TailBBI.HasFallThrough) { 1275 BBI.NonPredSize -= TII->RemoveBranch(*BBI.BB); 1276 MergeBlocks(BBI, TailBBI); 1277 TailBBI.IsDone = true; 1278 } else { 1279 InsertUncondBranch(BBI.BB, TailBB, TII); 1280 BBI.HasFallThrough = false; 1281 } 1282 } 1283 1284 RemoveExtraEdges(BBI); 1285 1286 // Update block info. 1287 BBI.IsDone = TrueBBI.IsDone = FalseBBI.IsDone = true; 1288 InvalidatePreds(BBI.BB); 1289 1290 // FIXME: Must maintain LiveIns. 1291 return true; 1292} 1293 1294/// PredicateBlock - Predicate instructions from the start of the block to the 1295/// specified end with the specified condition. 1296void IfConverter::PredicateBlock(BBInfo &BBI, 1297 MachineBasicBlock::iterator E, 1298 SmallVectorImpl<MachineOperand> &Cond, 1299 SmallSet<unsigned, 4> &Redefs) { 1300 for (MachineBasicBlock::iterator I = BBI.BB->begin(); I != E; ++I) { 1301 if (I->isDebugValue() || TII->isPredicated(I)) 1302 continue; 1303 if (!TII->PredicateInstruction(I, Cond)) { 1304#ifndef NDEBUG 1305 dbgs() << "Unable to predicate " << *I << "!\n"; 1306#endif 1307 llvm_unreachable(0); 1308 } 1309 1310 // If the predicated instruction now redefines a register as the result of 1311 // if-conversion, add an implicit kill. 1312 UpdatePredRedefs(I, Redefs, TRI, true); 1313 } 1314 1315 std::copy(Cond.begin(), Cond.end(), std::back_inserter(BBI.Predicate)); 1316 1317 BBI.IsAnalyzed = false; 1318 BBI.NonPredSize = 0; 1319 1320 ++NumIfConvBBs; 1321} 1322 1323/// CopyAndPredicateBlock - Copy and predicate instructions from source BB to 1324/// the destination block. Skip end of block branches if IgnoreBr is true. 1325void IfConverter::CopyAndPredicateBlock(BBInfo &ToBBI, BBInfo &FromBBI, 1326 SmallVectorImpl<MachineOperand> &Cond, 1327 SmallSet<unsigned, 4> &Redefs, 1328 bool IgnoreBr) { 1329 MachineFunction &MF = *ToBBI.BB->getParent(); 1330 1331 for (MachineBasicBlock::iterator I = FromBBI.BB->begin(), 1332 E = FromBBI.BB->end(); I != E; ++I) { 1333 const TargetInstrDesc &TID = I->getDesc(); 1334 // Do not copy the end of the block branches. 1335 if (IgnoreBr && TID.isBranch()) 1336 break; 1337 1338 MachineInstr *MI = MF.CloneMachineInstr(I); 1339 ToBBI.BB->insert(ToBBI.BB->end(), MI); 1340 ToBBI.NonPredSize++; 1341 1342 if (!TII->isPredicated(I) && !MI->isDebugValue()) { 1343 if (!TII->PredicateInstruction(MI, Cond)) { 1344#ifndef NDEBUG 1345 dbgs() << "Unable to predicate " << *I << "!\n"; 1346#endif 1347 llvm_unreachable(0); 1348 } 1349 } 1350 1351 // If the predicated instruction now redefines a register as the result of 1352 // if-conversion, add an implicit kill. 1353 UpdatePredRedefs(MI, Redefs, TRI, true); 1354 } 1355 1356 std::vector<MachineBasicBlock *> Succs(FromBBI.BB->succ_begin(), 1357 FromBBI.BB->succ_end()); 1358 MachineBasicBlock *NBB = getNextBlock(FromBBI.BB); 1359 MachineBasicBlock *FallThrough = FromBBI.HasFallThrough ? NBB : NULL; 1360 1361 for (unsigned i = 0, e = Succs.size(); i != e; ++i) { 1362 MachineBasicBlock *Succ = Succs[i]; 1363 // Fallthrough edge can't be transferred. 1364 if (Succ == FallThrough) 1365 continue; 1366 ToBBI.BB->addSuccessor(Succ); 1367 } 1368 1369 std::copy(FromBBI.Predicate.begin(), FromBBI.Predicate.end(), 1370 std::back_inserter(ToBBI.Predicate)); 1371 std::copy(Cond.begin(), Cond.end(), std::back_inserter(ToBBI.Predicate)); 1372 1373 ToBBI.ClobbersPred |= FromBBI.ClobbersPred; 1374 ToBBI.IsAnalyzed = false; 1375 1376 ++NumDupBBs; 1377} 1378 1379/// MergeBlocks - Move all instructions from FromBB to the end of ToBB. 1380/// 1381void IfConverter::MergeBlocks(BBInfo &ToBBI, BBInfo &FromBBI) { 1382 ToBBI.BB->splice(ToBBI.BB->end(), 1383 FromBBI.BB, FromBBI.BB->begin(), FromBBI.BB->end()); 1384 1385 // Redirect all branches to FromBB to ToBB. 1386 std::vector<MachineBasicBlock *> Preds(FromBBI.BB->pred_begin(), 1387 FromBBI.BB->pred_end()); 1388 for (unsigned i = 0, e = Preds.size(); i != e; ++i) { 1389 MachineBasicBlock *Pred = Preds[i]; 1390 if (Pred == ToBBI.BB) 1391 continue; 1392 Pred->ReplaceUsesOfBlockWith(FromBBI.BB, ToBBI.BB); 1393 } 1394 1395 std::vector<MachineBasicBlock *> Succs(FromBBI.BB->succ_begin(), 1396 FromBBI.BB->succ_end()); 1397 MachineBasicBlock *NBB = getNextBlock(FromBBI.BB); 1398 MachineBasicBlock *FallThrough = FromBBI.HasFallThrough ? NBB : NULL; 1399 1400 for (unsigned i = 0, e = Succs.size(); i != e; ++i) { 1401 MachineBasicBlock *Succ = Succs[i]; 1402 // Fallthrough edge can't be transferred. 1403 if (Succ == FallThrough) 1404 continue; 1405 FromBBI.BB->removeSuccessor(Succ); 1406 ToBBI.BB->addSuccessor(Succ); 1407 } 1408 1409 // Now FromBBI always falls through to the next block! 1410 if (NBB && !FromBBI.BB->isSuccessor(NBB)) 1411 FromBBI.BB->addSuccessor(NBB); 1412 1413 std::copy(FromBBI.Predicate.begin(), FromBBI.Predicate.end(), 1414 std::back_inserter(ToBBI.Predicate)); 1415 FromBBI.Predicate.clear(); 1416 1417 ToBBI.NonPredSize += FromBBI.NonPredSize; 1418 FromBBI.NonPredSize = 0; 1419 1420 ToBBI.ClobbersPred |= FromBBI.ClobbersPred; 1421 ToBBI.HasFallThrough = FromBBI.HasFallThrough; 1422 ToBBI.IsAnalyzed = false; 1423 FromBBI.IsAnalyzed = false; 1424} 1425