IfConversion.cpp revision b57d99694b87326a2eea26d76becf67bf5784b49
1//===-- IfConversion.cpp - Machine code if conversion pass. ---------------===// 2// 3// The LLVM Compiler Infrastructure 4// 5// This file is distributed under the University of Illinois Open Source 6// License. See LICENSE.TXT for details. 7// 8//===----------------------------------------------------------------------===// 9// 10// This file implements the machine instruction level if-conversion pass. 11// 12//===----------------------------------------------------------------------===// 13 14#define DEBUG_TYPE "ifcvt" 15#include "llvm/CodeGen/Passes.h" 16#include "BranchFolding.h" 17#include "llvm/ADT/STLExtras.h" 18#include "llvm/ADT/SmallSet.h" 19#include "llvm/ADT/Statistic.h" 20#include "llvm/CodeGen/MachineBranchProbabilityInfo.h" 21#include "llvm/CodeGen/MachineFunctionPass.h" 22#include "llvm/CodeGen/MachineInstrBuilder.h" 23#include "llvm/CodeGen/MachineModuleInfo.h" 24#include "llvm/CodeGen/MachineRegisterInfo.h" 25#include "llvm/MC/MCInstrItineraries.h" 26#include "llvm/Support/CommandLine.h" 27#include "llvm/Support/Debug.h" 28#include "llvm/Support/ErrorHandling.h" 29#include "llvm/Support/raw_ostream.h" 30#include "llvm/Target/TargetInstrInfo.h" 31#include "llvm/Target/TargetLowering.h" 32#include "llvm/Target/TargetMachine.h" 33#include "llvm/Target/TargetRegisterInfo.h" 34using namespace llvm; 35 36// Hidden options for help debugging. 37static cl::opt<int> IfCvtFnStart("ifcvt-fn-start", cl::init(-1), cl::Hidden); 38static cl::opt<int> IfCvtFnStop("ifcvt-fn-stop", cl::init(-1), cl::Hidden); 39static cl::opt<int> IfCvtLimit("ifcvt-limit", cl::init(-1), cl::Hidden); 40static cl::opt<bool> DisableSimple("disable-ifcvt-simple", 41 cl::init(false), cl::Hidden); 42static cl::opt<bool> DisableSimpleF("disable-ifcvt-simple-false", 43 cl::init(false), cl::Hidden); 44static cl::opt<bool> DisableTriangle("disable-ifcvt-triangle", 45 cl::init(false), cl::Hidden); 46static cl::opt<bool> DisableTriangleR("disable-ifcvt-triangle-rev", 47 cl::init(false), cl::Hidden); 48static cl::opt<bool> DisableTriangleF("disable-ifcvt-triangle-false", 49 cl::init(false), cl::Hidden); 50static cl::opt<bool> DisableTriangleFR("disable-ifcvt-triangle-false-rev", 51 cl::init(false), cl::Hidden); 52static cl::opt<bool> DisableDiamond("disable-ifcvt-diamond", 53 cl::init(false), cl::Hidden); 54static cl::opt<bool> IfCvtBranchFold("ifcvt-branch-fold", 55 cl::init(true), cl::Hidden); 56 57STATISTIC(NumSimple, "Number of simple if-conversions performed"); 58STATISTIC(NumSimpleFalse, "Number of simple (F) if-conversions performed"); 59STATISTIC(NumTriangle, "Number of triangle if-conversions performed"); 60STATISTIC(NumTriangleRev, "Number of triangle (R) if-conversions performed"); 61STATISTIC(NumTriangleFalse,"Number of triangle (F) if-conversions performed"); 62STATISTIC(NumTriangleFRev, "Number of triangle (F/R) if-conversions performed"); 63STATISTIC(NumDiamonds, "Number of diamond if-conversions performed"); 64STATISTIC(NumIfConvBBs, "Number of if-converted blocks"); 65STATISTIC(NumDupBBs, "Number of duplicated blocks"); 66STATISTIC(NumUnpred, "Number of true blocks of diamonds unpredicated"); 67 68namespace { 69 class IfConverter : public MachineFunctionPass { 70 enum IfcvtKind { 71 ICNotClassfied, // BB data valid, but not classified. 72 ICSimpleFalse, // Same as ICSimple, but on the false path. 73 ICSimple, // BB is entry of an one split, no rejoin sub-CFG. 74 ICTriangleFRev, // Same as ICTriangleFalse, but false path rev condition. 75 ICTriangleRev, // Same as ICTriangle, but true path rev condition. 76 ICTriangleFalse, // Same as ICTriangle, but on the false path. 77 ICTriangle, // BB is entry of a triangle sub-CFG. 78 ICDiamond // BB is entry of a diamond sub-CFG. 79 }; 80 81 /// BBInfo - One per MachineBasicBlock, this is used to cache the result 82 /// if-conversion feasibility analysis. This includes results from 83 /// TargetInstrInfo::AnalyzeBranch() (i.e. TBB, FBB, and Cond), and its 84 /// classification, and common tail block of its successors (if it's a 85 /// diamond shape), its size, whether it's predicable, and whether any 86 /// instruction can clobber the 'would-be' predicate. 87 /// 88 /// IsDone - True if BB is not to be considered for ifcvt. 89 /// IsBeingAnalyzed - True if BB is currently being analyzed. 90 /// IsAnalyzed - True if BB has been analyzed (info is still valid). 91 /// IsEnqueued - True if BB has been enqueued to be ifcvt'ed. 92 /// IsBrAnalyzable - True if AnalyzeBranch() returns false. 93 /// HasFallThrough - True if BB may fallthrough to the following BB. 94 /// IsUnpredicable - True if BB is known to be unpredicable. 95 /// ClobbersPred - True if BB could modify predicates (e.g. has 96 /// cmp, call, etc.) 97 /// NonPredSize - Number of non-predicated instructions. 98 /// ExtraCost - Extra cost for multi-cycle instructions. 99 /// ExtraCost2 - Some instructions are slower when predicated 100 /// BB - Corresponding MachineBasicBlock. 101 /// TrueBB / FalseBB- See AnalyzeBranch(). 102 /// BrCond - Conditions for end of block conditional branches. 103 /// Predicate - Predicate used in the BB. 104 struct BBInfo { 105 bool IsDone : 1; 106 bool IsBeingAnalyzed : 1; 107 bool IsAnalyzed : 1; 108 bool IsEnqueued : 1; 109 bool IsBrAnalyzable : 1; 110 bool HasFallThrough : 1; 111 bool IsUnpredicable : 1; 112 bool CannotBeCopied : 1; 113 bool ClobbersPred : 1; 114 unsigned NonPredSize; 115 unsigned ExtraCost; 116 unsigned ExtraCost2; 117 MachineBasicBlock *BB; 118 MachineBasicBlock *TrueBB; 119 MachineBasicBlock *FalseBB; 120 SmallVector<MachineOperand, 4> BrCond; 121 SmallVector<MachineOperand, 4> Predicate; 122 BBInfo() : IsDone(false), IsBeingAnalyzed(false), 123 IsAnalyzed(false), IsEnqueued(false), IsBrAnalyzable(false), 124 HasFallThrough(false), IsUnpredicable(false), 125 CannotBeCopied(false), ClobbersPred(false), NonPredSize(0), 126 ExtraCost(0), ExtraCost2(0), BB(0), TrueBB(0), FalseBB(0) {} 127 }; 128 129 /// IfcvtToken - Record information about pending if-conversions to attempt: 130 /// BBI - Corresponding BBInfo. 131 /// Kind - Type of block. See IfcvtKind. 132 /// NeedSubsumption - True if the to-be-predicated BB has already been 133 /// predicated. 134 /// NumDups - Number of instructions that would be duplicated due 135 /// to this if-conversion. (For diamonds, the number of 136 /// identical instructions at the beginnings of both 137 /// paths). 138 /// NumDups2 - For diamonds, the number of identical instructions 139 /// at the ends of both paths. 140 struct IfcvtToken { 141 BBInfo &BBI; 142 IfcvtKind Kind; 143 bool NeedSubsumption; 144 unsigned NumDups; 145 unsigned NumDups2; 146 IfcvtToken(BBInfo &b, IfcvtKind k, bool s, unsigned d, unsigned d2 = 0) 147 : BBI(b), Kind(k), NeedSubsumption(s), NumDups(d), NumDups2(d2) {} 148 }; 149 150 /// BBAnalysis - Results of if-conversion feasibility analysis indexed by 151 /// basic block number. 152 std::vector<BBInfo> BBAnalysis; 153 154 const TargetLoweringBase *TLI; 155 const TargetInstrInfo *TII; 156 const TargetRegisterInfo *TRI; 157 const InstrItineraryData *InstrItins; 158 const MachineBranchProbabilityInfo *MBPI; 159 MachineRegisterInfo *MRI; 160 161 bool PreRegAlloc; 162 bool MadeChange; 163 int FnNum; 164 public: 165 static char ID; 166 IfConverter() : MachineFunctionPass(ID), FnNum(-1) { 167 initializeIfConverterPass(*PassRegistry::getPassRegistry()); 168 } 169 170 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 171 AU.addRequired<MachineBranchProbabilityInfo>(); 172 MachineFunctionPass::getAnalysisUsage(AU); 173 } 174 175 virtual bool runOnMachineFunction(MachineFunction &MF); 176 177 private: 178 bool ReverseBranchCondition(BBInfo &BBI); 179 bool ValidSimple(BBInfo &TrueBBI, unsigned &Dups, 180 const BranchProbability &Prediction) const; 181 bool ValidTriangle(BBInfo &TrueBBI, BBInfo &FalseBBI, 182 bool FalseBranch, unsigned &Dups, 183 const BranchProbability &Prediction) const; 184 bool ValidDiamond(BBInfo &TrueBBI, BBInfo &FalseBBI, 185 unsigned &Dups1, unsigned &Dups2) const; 186 void ScanInstructions(BBInfo &BBI); 187 BBInfo &AnalyzeBlock(MachineBasicBlock *BB, 188 std::vector<IfcvtToken*> &Tokens); 189 bool FeasibilityAnalysis(BBInfo &BBI, SmallVectorImpl<MachineOperand> &Cond, 190 bool isTriangle = false, bool RevBranch = false); 191 void AnalyzeBlocks(MachineFunction &MF, std::vector<IfcvtToken*> &Tokens); 192 void InvalidatePreds(MachineBasicBlock *BB); 193 void RemoveExtraEdges(BBInfo &BBI); 194 bool IfConvertSimple(BBInfo &BBI, IfcvtKind Kind); 195 bool IfConvertTriangle(BBInfo &BBI, IfcvtKind Kind); 196 bool IfConvertDiamond(BBInfo &BBI, IfcvtKind Kind, 197 unsigned NumDups1, unsigned NumDups2); 198 void PredicateBlock(BBInfo &BBI, 199 MachineBasicBlock::iterator E, 200 SmallVectorImpl<MachineOperand> &Cond, 201 SmallSet<unsigned, 4> &Redefs, 202 SmallSet<unsigned, 4> *LaterRedefs = 0); 203 void CopyAndPredicateBlock(BBInfo &ToBBI, BBInfo &FromBBI, 204 SmallVectorImpl<MachineOperand> &Cond, 205 SmallSet<unsigned, 4> &Redefs, 206 bool IgnoreBr = false); 207 void MergeBlocks(BBInfo &ToBBI, BBInfo &FromBBI, bool AddEdges = true); 208 209 bool MeetIfcvtSizeLimit(MachineBasicBlock &BB, 210 unsigned Cycle, unsigned Extra, 211 const BranchProbability &Prediction) const { 212 return Cycle > 0 && TII->isProfitableToIfCvt(BB, Cycle, Extra, 213 Prediction); 214 } 215 216 bool MeetIfcvtSizeLimit(MachineBasicBlock &TBB, 217 unsigned TCycle, unsigned TExtra, 218 MachineBasicBlock &FBB, 219 unsigned FCycle, unsigned FExtra, 220 const BranchProbability &Prediction) const { 221 return TCycle > 0 && FCycle > 0 && 222 TII->isProfitableToIfCvt(TBB, TCycle, TExtra, FBB, FCycle, FExtra, 223 Prediction); 224 } 225 226 // blockAlwaysFallThrough - Block ends without a terminator. 227 bool blockAlwaysFallThrough(BBInfo &BBI) const { 228 return BBI.IsBrAnalyzable && BBI.TrueBB == NULL; 229 } 230 231 // IfcvtTokenCmp - Used to sort if-conversion candidates. 232 static bool IfcvtTokenCmp(IfcvtToken *C1, IfcvtToken *C2) { 233 int Incr1 = (C1->Kind == ICDiamond) 234 ? -(int)(C1->NumDups + C1->NumDups2) : (int)C1->NumDups; 235 int Incr2 = (C2->Kind == ICDiamond) 236 ? -(int)(C2->NumDups + C2->NumDups2) : (int)C2->NumDups; 237 if (Incr1 > Incr2) 238 return true; 239 else if (Incr1 == Incr2) { 240 // Favors subsumption. 241 if (C1->NeedSubsumption == false && C2->NeedSubsumption == true) 242 return true; 243 else if (C1->NeedSubsumption == C2->NeedSubsumption) { 244 // Favors diamond over triangle, etc. 245 if ((unsigned)C1->Kind < (unsigned)C2->Kind) 246 return true; 247 else if (C1->Kind == C2->Kind) 248 return C1->BBI.BB->getNumber() < C2->BBI.BB->getNumber(); 249 } 250 } 251 return false; 252 } 253 }; 254 255 char IfConverter::ID = 0; 256} 257 258char &llvm::IfConverterID = IfConverter::ID; 259 260INITIALIZE_PASS_BEGIN(IfConverter, "if-converter", "If Converter", false, false) 261INITIALIZE_PASS_DEPENDENCY(MachineBranchProbabilityInfo) 262INITIALIZE_PASS_END(IfConverter, "if-converter", "If Converter", false, false) 263 264bool IfConverter::runOnMachineFunction(MachineFunction &MF) { 265 TLI = MF.getTarget().getTargetLowering(); 266 TII = MF.getTarget().getInstrInfo(); 267 TRI = MF.getTarget().getRegisterInfo(); 268 MBPI = &getAnalysis<MachineBranchProbabilityInfo>(); 269 MRI = &MF.getRegInfo(); 270 InstrItins = MF.getTarget().getInstrItineraryData(); 271 if (!TII) return false; 272 273 PreRegAlloc = MRI->isSSA(); 274 275 bool BFChange = false; 276 if (!PreRegAlloc) { 277 // Tail merge tend to expose more if-conversion opportunities. 278 BranchFolder BF(true, false); 279 BFChange = BF.OptimizeFunction(MF, TII, 280 MF.getTarget().getRegisterInfo(), 281 getAnalysisIfAvailable<MachineModuleInfo>()); 282 } 283 284 DEBUG(dbgs() << "\nIfcvt: function (" << ++FnNum << ") \'" 285 << MF.getName() << "\'"); 286 287 if (FnNum < IfCvtFnStart || (IfCvtFnStop != -1 && FnNum > IfCvtFnStop)) { 288 DEBUG(dbgs() << " skipped\n"); 289 return false; 290 } 291 DEBUG(dbgs() << "\n"); 292 293 MF.RenumberBlocks(); 294 BBAnalysis.resize(MF.getNumBlockIDs()); 295 296 std::vector<IfcvtToken*> Tokens; 297 MadeChange = false; 298 unsigned NumIfCvts = NumSimple + NumSimpleFalse + NumTriangle + 299 NumTriangleRev + NumTriangleFalse + NumTriangleFRev + NumDiamonds; 300 while (IfCvtLimit == -1 || (int)NumIfCvts < IfCvtLimit) { 301 // Do an initial analysis for each basic block and find all the potential 302 // candidates to perform if-conversion. 303 bool Change = false; 304 AnalyzeBlocks(MF, Tokens); 305 while (!Tokens.empty()) { 306 IfcvtToken *Token = Tokens.back(); 307 Tokens.pop_back(); 308 BBInfo &BBI = Token->BBI; 309 IfcvtKind Kind = Token->Kind; 310 unsigned NumDups = Token->NumDups; 311 unsigned NumDups2 = Token->NumDups2; 312 313 delete Token; 314 315 // If the block has been evicted out of the queue or it has already been 316 // marked dead (due to it being predicated), then skip it. 317 if (BBI.IsDone) 318 BBI.IsEnqueued = false; 319 if (!BBI.IsEnqueued) 320 continue; 321 322 BBI.IsEnqueued = false; 323 324 bool RetVal = false; 325 switch (Kind) { 326 default: llvm_unreachable("Unexpected!"); 327 case ICSimple: 328 case ICSimpleFalse: { 329 bool isFalse = Kind == ICSimpleFalse; 330 if ((isFalse && DisableSimpleF) || (!isFalse && DisableSimple)) break; 331 DEBUG(dbgs() << "Ifcvt (Simple" << (Kind == ICSimpleFalse ? 332 " false" : "") 333 << "): BB#" << BBI.BB->getNumber() << " (" 334 << ((Kind == ICSimpleFalse) 335 ? BBI.FalseBB->getNumber() 336 : BBI.TrueBB->getNumber()) << ") "); 337 RetVal = IfConvertSimple(BBI, Kind); 338 DEBUG(dbgs() << (RetVal ? "succeeded!" : "failed!") << "\n"); 339 if (RetVal) { 340 if (isFalse) ++NumSimpleFalse; 341 else ++NumSimple; 342 } 343 break; 344 } 345 case ICTriangle: 346 case ICTriangleRev: 347 case ICTriangleFalse: 348 case ICTriangleFRev: { 349 bool isFalse = Kind == ICTriangleFalse; 350 bool isRev = (Kind == ICTriangleRev || Kind == ICTriangleFRev); 351 if (DisableTriangle && !isFalse && !isRev) break; 352 if (DisableTriangleR && !isFalse && isRev) break; 353 if (DisableTriangleF && isFalse && !isRev) break; 354 if (DisableTriangleFR && isFalse && isRev) break; 355 DEBUG(dbgs() << "Ifcvt (Triangle"); 356 if (isFalse) 357 DEBUG(dbgs() << " false"); 358 if (isRev) 359 DEBUG(dbgs() << " rev"); 360 DEBUG(dbgs() << "): BB#" << BBI.BB->getNumber() << " (T:" 361 << BBI.TrueBB->getNumber() << ",F:" 362 << BBI.FalseBB->getNumber() << ") "); 363 RetVal = IfConvertTriangle(BBI, Kind); 364 DEBUG(dbgs() << (RetVal ? "succeeded!" : "failed!") << "\n"); 365 if (RetVal) { 366 if (isFalse) { 367 if (isRev) ++NumTriangleFRev; 368 else ++NumTriangleFalse; 369 } else { 370 if (isRev) ++NumTriangleRev; 371 else ++NumTriangle; 372 } 373 } 374 break; 375 } 376 case ICDiamond: { 377 if (DisableDiamond) break; 378 DEBUG(dbgs() << "Ifcvt (Diamond): BB#" << BBI.BB->getNumber() << " (T:" 379 << BBI.TrueBB->getNumber() << ",F:" 380 << BBI.FalseBB->getNumber() << ") "); 381 RetVal = IfConvertDiamond(BBI, Kind, NumDups, NumDups2); 382 DEBUG(dbgs() << (RetVal ? "succeeded!" : "failed!") << "\n"); 383 if (RetVal) ++NumDiamonds; 384 break; 385 } 386 } 387 388 Change |= RetVal; 389 390 NumIfCvts = NumSimple + NumSimpleFalse + NumTriangle + NumTriangleRev + 391 NumTriangleFalse + NumTriangleFRev + NumDiamonds; 392 if (IfCvtLimit != -1 && (int)NumIfCvts >= IfCvtLimit) 393 break; 394 } 395 396 if (!Change) 397 break; 398 MadeChange |= Change; 399 } 400 401 // Delete tokens in case of early exit. 402 while (!Tokens.empty()) { 403 IfcvtToken *Token = Tokens.back(); 404 Tokens.pop_back(); 405 delete Token; 406 } 407 408 Tokens.clear(); 409 BBAnalysis.clear(); 410 411 if (MadeChange && IfCvtBranchFold) { 412 BranchFolder BF(false, false); 413 BF.OptimizeFunction(MF, TII, 414 MF.getTarget().getRegisterInfo(), 415 getAnalysisIfAvailable<MachineModuleInfo>()); 416 } 417 418 MadeChange |= BFChange; 419 return MadeChange; 420} 421 422/// findFalseBlock - BB has a fallthrough. Find its 'false' successor given 423/// its 'true' successor. 424static MachineBasicBlock *findFalseBlock(MachineBasicBlock *BB, 425 MachineBasicBlock *TrueBB) { 426 for (MachineBasicBlock::succ_iterator SI = BB->succ_begin(), 427 E = BB->succ_end(); SI != E; ++SI) { 428 MachineBasicBlock *SuccBB = *SI; 429 if (SuccBB != TrueBB) 430 return SuccBB; 431 } 432 return NULL; 433} 434 435/// ReverseBranchCondition - Reverse the condition of the end of the block 436/// branch. Swap block's 'true' and 'false' successors. 437bool IfConverter::ReverseBranchCondition(BBInfo &BBI) { 438 DebugLoc dl; // FIXME: this is nowhere 439 if (!TII->ReverseBranchCondition(BBI.BrCond)) { 440 TII->RemoveBranch(*BBI.BB); 441 TII->InsertBranch(*BBI.BB, BBI.FalseBB, BBI.TrueBB, BBI.BrCond, dl); 442 std::swap(BBI.TrueBB, BBI.FalseBB); 443 return true; 444 } 445 return false; 446} 447 448/// getNextBlock - Returns the next block in the function blocks ordering. If 449/// it is the end, returns NULL. 450static inline MachineBasicBlock *getNextBlock(MachineBasicBlock *BB) { 451 MachineFunction::iterator I = BB; 452 MachineFunction::iterator E = BB->getParent()->end(); 453 if (++I == E) 454 return NULL; 455 return I; 456} 457 458/// ValidSimple - Returns true if the 'true' block (along with its 459/// predecessor) forms a valid simple shape for ifcvt. It also returns the 460/// number of instructions that the ifcvt would need to duplicate if performed 461/// in Dups. 462bool IfConverter::ValidSimple(BBInfo &TrueBBI, unsigned &Dups, 463 const BranchProbability &Prediction) const { 464 Dups = 0; 465 if (TrueBBI.IsBeingAnalyzed || TrueBBI.IsDone) 466 return false; 467 468 if (TrueBBI.IsBrAnalyzable) 469 return false; 470 471 if (TrueBBI.BB->pred_size() > 1) { 472 if (TrueBBI.CannotBeCopied || 473 !TII->isProfitableToDupForIfCvt(*TrueBBI.BB, TrueBBI.NonPredSize, 474 Prediction)) 475 return false; 476 Dups = TrueBBI.NonPredSize; 477 } 478 479 return true; 480} 481 482/// ValidTriangle - Returns true if the 'true' and 'false' blocks (along 483/// with their common predecessor) forms a valid triangle shape for ifcvt. 484/// If 'FalseBranch' is true, it checks if 'true' block's false branch 485/// branches to the 'false' block rather than the other way around. It also 486/// returns the number of instructions that the ifcvt would need to duplicate 487/// if performed in 'Dups'. 488bool IfConverter::ValidTriangle(BBInfo &TrueBBI, BBInfo &FalseBBI, 489 bool FalseBranch, unsigned &Dups, 490 const BranchProbability &Prediction) const { 491 Dups = 0; 492 if (TrueBBI.IsBeingAnalyzed || TrueBBI.IsDone) 493 return false; 494 495 if (TrueBBI.BB->pred_size() > 1) { 496 if (TrueBBI.CannotBeCopied) 497 return false; 498 499 unsigned Size = TrueBBI.NonPredSize; 500 if (TrueBBI.IsBrAnalyzable) { 501 if (TrueBBI.TrueBB && TrueBBI.BrCond.empty()) 502 // Ends with an unconditional branch. It will be removed. 503 --Size; 504 else { 505 MachineBasicBlock *FExit = FalseBranch 506 ? TrueBBI.TrueBB : TrueBBI.FalseBB; 507 if (FExit) 508 // Require a conditional branch 509 ++Size; 510 } 511 } 512 if (!TII->isProfitableToDupForIfCvt(*TrueBBI.BB, Size, Prediction)) 513 return false; 514 Dups = Size; 515 } 516 517 MachineBasicBlock *TExit = FalseBranch ? TrueBBI.FalseBB : TrueBBI.TrueBB; 518 if (!TExit && blockAlwaysFallThrough(TrueBBI)) { 519 MachineFunction::iterator I = TrueBBI.BB; 520 if (++I == TrueBBI.BB->getParent()->end()) 521 return false; 522 TExit = I; 523 } 524 return TExit && TExit == FalseBBI.BB; 525} 526 527/// ValidDiamond - Returns true if the 'true' and 'false' blocks (along 528/// with their common predecessor) forms a valid diamond shape for ifcvt. 529bool IfConverter::ValidDiamond(BBInfo &TrueBBI, BBInfo &FalseBBI, 530 unsigned &Dups1, unsigned &Dups2) const { 531 Dups1 = Dups2 = 0; 532 if (TrueBBI.IsBeingAnalyzed || TrueBBI.IsDone || 533 FalseBBI.IsBeingAnalyzed || FalseBBI.IsDone) 534 return false; 535 536 MachineBasicBlock *TT = TrueBBI.TrueBB; 537 MachineBasicBlock *FT = FalseBBI.TrueBB; 538 539 if (!TT && blockAlwaysFallThrough(TrueBBI)) 540 TT = getNextBlock(TrueBBI.BB); 541 if (!FT && blockAlwaysFallThrough(FalseBBI)) 542 FT = getNextBlock(FalseBBI.BB); 543 if (TT != FT) 544 return false; 545 if (TT == NULL && (TrueBBI.IsBrAnalyzable || FalseBBI.IsBrAnalyzable)) 546 return false; 547 if (TrueBBI.BB->pred_size() > 1 || FalseBBI.BB->pred_size() > 1) 548 return false; 549 550 // FIXME: Allow true block to have an early exit? 551 if (TrueBBI.FalseBB || FalseBBI.FalseBB || 552 (TrueBBI.ClobbersPred && FalseBBI.ClobbersPred)) 553 return false; 554 555 // Count duplicate instructions at the beginning of the true and false blocks. 556 MachineBasicBlock::iterator TIB = TrueBBI.BB->begin(); 557 MachineBasicBlock::iterator FIB = FalseBBI.BB->begin(); 558 MachineBasicBlock::iterator TIE = TrueBBI.BB->end(); 559 MachineBasicBlock::iterator FIE = FalseBBI.BB->end(); 560 while (TIB != TIE && FIB != FIE) { 561 // Skip dbg_value instructions. These do not count. 562 if (TIB->isDebugValue()) { 563 while (TIB != TIE && TIB->isDebugValue()) 564 ++TIB; 565 if (TIB == TIE) 566 break; 567 } 568 if (FIB->isDebugValue()) { 569 while (FIB != FIE && FIB->isDebugValue()) 570 ++FIB; 571 if (FIB == FIE) 572 break; 573 } 574 if (!TIB->isIdenticalTo(FIB)) 575 break; 576 ++Dups1; 577 ++TIB; 578 ++FIB; 579 } 580 581 // Now, in preparation for counting duplicate instructions at the ends of the 582 // blocks, move the end iterators up past any branch instructions. 583 while (TIE != TIB) { 584 --TIE; 585 if (!TIE->isBranch()) 586 break; 587 } 588 while (FIE != FIB) { 589 --FIE; 590 if (!FIE->isBranch()) 591 break; 592 } 593 594 // If Dups1 includes all of a block, then don't count duplicate 595 // instructions at the end of the blocks. 596 if (TIB == TIE || FIB == FIE) 597 return true; 598 599 // Count duplicate instructions at the ends of the blocks. 600 while (TIE != TIB && FIE != FIB) { 601 // Skip dbg_value instructions. These do not count. 602 if (TIE->isDebugValue()) { 603 while (TIE != TIB && TIE->isDebugValue()) 604 --TIE; 605 if (TIE == TIB) 606 break; 607 } 608 if (FIE->isDebugValue()) { 609 while (FIE != FIB && FIE->isDebugValue()) 610 --FIE; 611 if (FIE == FIB) 612 break; 613 } 614 if (!TIE->isIdenticalTo(FIE)) 615 break; 616 ++Dups2; 617 --TIE; 618 --FIE; 619 } 620 621 return true; 622} 623 624/// ScanInstructions - Scan all the instructions in the block to determine if 625/// the block is predicable. In most cases, that means all the instructions 626/// in the block are isPredicable(). Also checks if the block contains any 627/// instruction which can clobber a predicate (e.g. condition code register). 628/// If so, the block is not predicable unless it's the last instruction. 629void IfConverter::ScanInstructions(BBInfo &BBI) { 630 if (BBI.IsDone) 631 return; 632 633 bool AlreadyPredicated = !BBI.Predicate.empty(); 634 // First analyze the end of BB branches. 635 BBI.TrueBB = BBI.FalseBB = NULL; 636 BBI.BrCond.clear(); 637 BBI.IsBrAnalyzable = 638 !TII->AnalyzeBranch(*BBI.BB, BBI.TrueBB, BBI.FalseBB, BBI.BrCond); 639 BBI.HasFallThrough = BBI.IsBrAnalyzable && BBI.FalseBB == NULL; 640 641 if (BBI.BrCond.size()) { 642 // No false branch. This BB must end with a conditional branch and a 643 // fallthrough. 644 if (!BBI.FalseBB) 645 BBI.FalseBB = findFalseBlock(BBI.BB, BBI.TrueBB); 646 if (!BBI.FalseBB) { 647 // Malformed bcc? True and false blocks are the same? 648 BBI.IsUnpredicable = true; 649 return; 650 } 651 } 652 653 // Then scan all the instructions. 654 BBI.NonPredSize = 0; 655 BBI.ExtraCost = 0; 656 BBI.ExtraCost2 = 0; 657 BBI.ClobbersPred = false; 658 for (MachineBasicBlock::iterator I = BBI.BB->begin(), E = BBI.BB->end(); 659 I != E; ++I) { 660 if (I->isDebugValue()) 661 continue; 662 663 if (I->isNotDuplicable()) 664 BBI.CannotBeCopied = true; 665 666 bool isPredicated = TII->isPredicated(I); 667 bool isCondBr = BBI.IsBrAnalyzable && I->isConditionalBranch(); 668 669 // A conditional branch is not predicable, but it may be eliminated. 670 if (isCondBr) 671 continue; 672 673 if (!isPredicated) { 674 BBI.NonPredSize++; 675 unsigned ExtraPredCost = 0; 676 unsigned NumCycles = TII->getInstrLatency(InstrItins, &*I, 677 &ExtraPredCost); 678 if (NumCycles > 1) 679 BBI.ExtraCost += NumCycles-1; 680 BBI.ExtraCost2 += ExtraPredCost; 681 } else if (!AlreadyPredicated) { 682 // FIXME: This instruction is already predicated before the 683 // if-conversion pass. It's probably something like a conditional move. 684 // Mark this block unpredicable for now. 685 BBI.IsUnpredicable = true; 686 return; 687 } 688 689 if (BBI.ClobbersPred && !isPredicated) { 690 // Predicate modification instruction should end the block (except for 691 // already predicated instructions and end of block branches). 692 // Predicate may have been modified, the subsequent (currently) 693 // unpredicated instructions cannot be correctly predicated. 694 BBI.IsUnpredicable = true; 695 return; 696 } 697 698 // FIXME: Make use of PredDefs? e.g. ADDC, SUBC sets predicates but are 699 // still potentially predicable. 700 std::vector<MachineOperand> PredDefs; 701 if (TII->DefinesPredicate(I, PredDefs)) 702 BBI.ClobbersPred = true; 703 704 if (!TII->isPredicable(I)) { 705 BBI.IsUnpredicable = true; 706 return; 707 } 708 } 709} 710 711/// FeasibilityAnalysis - Determine if the block is a suitable candidate to be 712/// predicated by the specified predicate. 713bool IfConverter::FeasibilityAnalysis(BBInfo &BBI, 714 SmallVectorImpl<MachineOperand> &Pred, 715 bool isTriangle, bool RevBranch) { 716 // If the block is dead or unpredicable, then it cannot be predicated. 717 if (BBI.IsDone || BBI.IsUnpredicable) 718 return false; 719 720 // If it is already predicated, check if the new predicate subsumes 721 // its predicate. 722 if (BBI.Predicate.size() && !TII->SubsumesPredicate(Pred, BBI.Predicate)) 723 return false; 724 725 if (BBI.BrCond.size()) { 726 if (!isTriangle) 727 return false; 728 729 // Test predicate subsumption. 730 SmallVector<MachineOperand, 4> RevPred(Pred.begin(), Pred.end()); 731 SmallVector<MachineOperand, 4> Cond(BBI.BrCond.begin(), BBI.BrCond.end()); 732 if (RevBranch) { 733 if (TII->ReverseBranchCondition(Cond)) 734 return false; 735 } 736 if (TII->ReverseBranchCondition(RevPred) || 737 !TII->SubsumesPredicate(Cond, RevPred)) 738 return false; 739 } 740 741 return true; 742} 743 744/// AnalyzeBlock - Analyze the structure of the sub-CFG starting from 745/// the specified block. Record its successors and whether it looks like an 746/// if-conversion candidate. 747IfConverter::BBInfo &IfConverter::AnalyzeBlock(MachineBasicBlock *BB, 748 std::vector<IfcvtToken*> &Tokens) { 749 BBInfo &BBI = BBAnalysis[BB->getNumber()]; 750 751 if (BBI.IsAnalyzed || BBI.IsBeingAnalyzed) 752 return BBI; 753 754 BBI.BB = BB; 755 BBI.IsBeingAnalyzed = true; 756 757 ScanInstructions(BBI); 758 759 // Unanalyzable or ends with fallthrough or unconditional branch, or if is not 760 // considered for ifcvt anymore. 761 if (!BBI.IsBrAnalyzable || BBI.BrCond.empty() || BBI.IsDone) { 762 BBI.IsBeingAnalyzed = false; 763 BBI.IsAnalyzed = true; 764 return BBI; 765 } 766 767 // Do not ifcvt if either path is a back edge to the entry block. 768 if (BBI.TrueBB == BB || BBI.FalseBB == BB) { 769 BBI.IsBeingAnalyzed = false; 770 BBI.IsAnalyzed = true; 771 return BBI; 772 } 773 774 // Do not ifcvt if true and false fallthrough blocks are the same. 775 if (!BBI.FalseBB) { 776 BBI.IsBeingAnalyzed = false; 777 BBI.IsAnalyzed = true; 778 return BBI; 779 } 780 781 BBInfo &TrueBBI = AnalyzeBlock(BBI.TrueBB, Tokens); 782 BBInfo &FalseBBI = AnalyzeBlock(BBI.FalseBB, Tokens); 783 784 if (TrueBBI.IsDone && FalseBBI.IsDone) { 785 BBI.IsBeingAnalyzed = false; 786 BBI.IsAnalyzed = true; 787 return BBI; 788 } 789 790 SmallVector<MachineOperand, 4> RevCond(BBI.BrCond.begin(), BBI.BrCond.end()); 791 bool CanRevCond = !TII->ReverseBranchCondition(RevCond); 792 793 unsigned Dups = 0; 794 unsigned Dups2 = 0; 795 bool TNeedSub = !TrueBBI.Predicate.empty(); 796 bool FNeedSub = !FalseBBI.Predicate.empty(); 797 bool Enqueued = false; 798 799 BranchProbability Prediction = MBPI->getEdgeProbability(BB, TrueBBI.BB); 800 801 if (CanRevCond && ValidDiamond(TrueBBI, FalseBBI, Dups, Dups2) && 802 MeetIfcvtSizeLimit(*TrueBBI.BB, (TrueBBI.NonPredSize - (Dups + Dups2) + 803 TrueBBI.ExtraCost), TrueBBI.ExtraCost2, 804 *FalseBBI.BB, (FalseBBI.NonPredSize - (Dups + Dups2) + 805 FalseBBI.ExtraCost),FalseBBI.ExtraCost2, 806 Prediction) && 807 FeasibilityAnalysis(TrueBBI, BBI.BrCond) && 808 FeasibilityAnalysis(FalseBBI, RevCond)) { 809 // Diamond: 810 // EBB 811 // / \_ 812 // | | 813 // TBB FBB 814 // \ / 815 // TailBB 816 // Note TailBB can be empty. 817 Tokens.push_back(new IfcvtToken(BBI, ICDiamond, TNeedSub|FNeedSub, Dups, 818 Dups2)); 819 Enqueued = true; 820 } 821 822 if (ValidTriangle(TrueBBI, FalseBBI, false, Dups, Prediction) && 823 MeetIfcvtSizeLimit(*TrueBBI.BB, TrueBBI.NonPredSize + TrueBBI.ExtraCost, 824 TrueBBI.ExtraCost2, Prediction) && 825 FeasibilityAnalysis(TrueBBI, BBI.BrCond, true)) { 826 // Triangle: 827 // EBB 828 // | \_ 829 // | | 830 // | TBB 831 // | / 832 // FBB 833 Tokens.push_back(new IfcvtToken(BBI, ICTriangle, TNeedSub, Dups)); 834 Enqueued = true; 835 } 836 837 if (ValidTriangle(TrueBBI, FalseBBI, true, Dups, Prediction) && 838 MeetIfcvtSizeLimit(*TrueBBI.BB, TrueBBI.NonPredSize + TrueBBI.ExtraCost, 839 TrueBBI.ExtraCost2, 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 + TrueBBI.ExtraCost, 847 TrueBBI.ExtraCost2, Prediction) && 848 FeasibilityAnalysis(TrueBBI, BBI.BrCond)) { 849 // Simple (split, no rejoin): 850 // EBB 851 // | \_ 852 // | | 853 // | TBB---> exit 854 // | 855 // FBB 856 Tokens.push_back(new IfcvtToken(BBI, ICSimple, TNeedSub, Dups)); 857 Enqueued = true; 858 } 859 860 if (CanRevCond) { 861 // Try the other path... 862 if (ValidTriangle(FalseBBI, TrueBBI, false, Dups, 863 Prediction.getCompl()) && 864 MeetIfcvtSizeLimit(*FalseBBI.BB, 865 FalseBBI.NonPredSize + FalseBBI.ExtraCost, 866 FalseBBI.ExtraCost2, Prediction.getCompl()) && 867 FeasibilityAnalysis(FalseBBI, RevCond, true)) { 868 Tokens.push_back(new IfcvtToken(BBI, ICTriangleFalse, FNeedSub, Dups)); 869 Enqueued = true; 870 } 871 872 if (ValidTriangle(FalseBBI, TrueBBI, true, Dups, 873 Prediction.getCompl()) && 874 MeetIfcvtSizeLimit(*FalseBBI.BB, 875 FalseBBI.NonPredSize + FalseBBI.ExtraCost, 876 FalseBBI.ExtraCost2, Prediction.getCompl()) && 877 FeasibilityAnalysis(FalseBBI, RevCond, true, true)) { 878 Tokens.push_back(new IfcvtToken(BBI, ICTriangleFRev, FNeedSub, Dups)); 879 Enqueued = true; 880 } 881 882 if (ValidSimple(FalseBBI, Dups, Prediction.getCompl()) && 883 MeetIfcvtSizeLimit(*FalseBBI.BB, 884 FalseBBI.NonPredSize + FalseBBI.ExtraCost, 885 FalseBBI.ExtraCost2, Prediction.getCompl()) && 886 FeasibilityAnalysis(FalseBBI, RevCond)) { 887 Tokens.push_back(new IfcvtToken(BBI, ICSimpleFalse, FNeedSub, Dups)); 888 Enqueued = true; 889 } 890 } 891 892 BBI.IsEnqueued = Enqueued; 893 BBI.IsBeingAnalyzed = false; 894 BBI.IsAnalyzed = true; 895 return BBI; 896} 897 898/// AnalyzeBlocks - Analyze all blocks and find entries for all if-conversion 899/// candidates. 900void IfConverter::AnalyzeBlocks(MachineFunction &MF, 901 std::vector<IfcvtToken*> &Tokens) { 902 for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; ++I) { 903 MachineBasicBlock *BB = I; 904 AnalyzeBlock(BB, Tokens); 905 } 906 907 // Sort to favor more complex ifcvt scheme. 908 std::stable_sort(Tokens.begin(), Tokens.end(), IfcvtTokenCmp); 909} 910 911/// canFallThroughTo - Returns true either if ToBB is the next block after BB or 912/// that all the intervening blocks are empty (given BB can fall through to its 913/// next block). 914static bool canFallThroughTo(MachineBasicBlock *BB, MachineBasicBlock *ToBB) { 915 MachineFunction::iterator PI = BB; 916 MachineFunction::iterator I = llvm::next(PI); 917 MachineFunction::iterator TI = ToBB; 918 MachineFunction::iterator E = BB->getParent()->end(); 919 while (I != TI) { 920 // Check isSuccessor to avoid case where the next block is empty, but 921 // it's not a successor. 922 if (I == E || !I->empty() || !PI->isSuccessor(I)) 923 return false; 924 PI = I++; 925 } 926 return true; 927} 928 929/// InvalidatePreds - Invalidate predecessor BB info so it would be re-analyzed 930/// to determine if it can be if-converted. If predecessor is already enqueued, 931/// dequeue it! 932void IfConverter::InvalidatePreds(MachineBasicBlock *BB) { 933 for (MachineBasicBlock::pred_iterator PI = BB->pred_begin(), 934 E = BB->pred_end(); PI != E; ++PI) { 935 BBInfo &PBBI = BBAnalysis[(*PI)->getNumber()]; 936 if (PBBI.IsDone || PBBI.BB == BB) 937 continue; 938 PBBI.IsAnalyzed = false; 939 PBBI.IsEnqueued = false; 940 } 941} 942 943/// InsertUncondBranch - Inserts an unconditional branch from BB to ToBB. 944/// 945static void InsertUncondBranch(MachineBasicBlock *BB, MachineBasicBlock *ToBB, 946 const TargetInstrInfo *TII) { 947 DebugLoc dl; // FIXME: this is nowhere 948 SmallVector<MachineOperand, 0> NoCond; 949 TII->InsertBranch(*BB, ToBB, NULL, NoCond, dl); 950} 951 952/// RemoveExtraEdges - Remove true / false edges if either / both are no longer 953/// successors. 954void IfConverter::RemoveExtraEdges(BBInfo &BBI) { 955 MachineBasicBlock *TBB = NULL, *FBB = NULL; 956 SmallVector<MachineOperand, 4> Cond; 957 if (!TII->AnalyzeBranch(*BBI.BB, TBB, FBB, Cond)) 958 BBI.BB->CorrectExtraCFGEdges(TBB, FBB, !Cond.empty()); 959} 960 961/// InitPredRedefs / UpdatePredRedefs - Defs by predicated instructions are 962/// modeled as read + write (sort like two-address instructions). These 963/// routines track register liveness and add implicit uses to if-converted 964/// instructions to conform to the model. 965static void InitPredRedefs(MachineBasicBlock *BB, SmallSet<unsigned,4> &Redefs, 966 const TargetRegisterInfo *TRI) { 967 for (MachineBasicBlock::livein_iterator I = BB->livein_begin(), 968 E = BB->livein_end(); I != E; ++I) { 969 unsigned Reg = *I; 970 for (MCSubRegIterator SubRegs(Reg, TRI, /*IncludeSelf=*/true); 971 SubRegs.isValid(); ++SubRegs) 972 Redefs.insert(*SubRegs); 973 } 974} 975 976static void UpdatePredRedefs(MachineInstr *MI, SmallSet<unsigned,4> &Redefs, 977 const TargetRegisterInfo *TRI, 978 bool AddImpUse = false) { 979 SmallVector<unsigned, 4> Defs; 980 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) { 981 const MachineOperand &MO = MI->getOperand(i); 982 if (!MO.isReg()) 983 continue; 984 unsigned Reg = MO.getReg(); 985 if (!Reg) 986 continue; 987 if (MO.isDef()) 988 Defs.push_back(Reg); 989 else if (MO.isKill()) { 990 for (MCSubRegIterator SubRegs(Reg, TRI, /*IncludeSelf=*/true); 991 SubRegs.isValid(); ++SubRegs) 992 Redefs.erase(*SubRegs); 993 } 994 } 995 MachineInstrBuilder MIB(*MI->getParent()->getParent(), MI); 996 for (unsigned i = 0, e = Defs.size(); i != e; ++i) { 997 unsigned Reg = Defs[i]; 998 if (!Redefs.insert(Reg)) { 999 if (AddImpUse) 1000 // Treat predicated update as read + write. 1001 MIB.addReg(Reg, RegState::Implicit | RegState::Undef); 1002 } else { 1003 for (MCSubRegIterator SubRegs(Reg, TRI); SubRegs.isValid(); ++SubRegs) 1004 Redefs.insert(*SubRegs); 1005 } 1006 } 1007} 1008 1009static void UpdatePredRedefs(MachineBasicBlock::iterator I, 1010 MachineBasicBlock::iterator E, 1011 SmallSet<unsigned,4> &Redefs, 1012 const TargetRegisterInfo *TRI) { 1013 while (I != E) { 1014 UpdatePredRedefs(I, Redefs, TRI); 1015 ++I; 1016 } 1017} 1018 1019/// IfConvertSimple - If convert a simple (split, no rejoin) sub-CFG. 1020/// 1021bool IfConverter::IfConvertSimple(BBInfo &BBI, IfcvtKind Kind) { 1022 BBInfo &TrueBBI = BBAnalysis[BBI.TrueBB->getNumber()]; 1023 BBInfo &FalseBBI = BBAnalysis[BBI.FalseBB->getNumber()]; 1024 BBInfo *CvtBBI = &TrueBBI; 1025 BBInfo *NextBBI = &FalseBBI; 1026 1027 SmallVector<MachineOperand, 4> Cond(BBI.BrCond.begin(), BBI.BrCond.end()); 1028 if (Kind == ICSimpleFalse) 1029 std::swap(CvtBBI, NextBBI); 1030 1031 if (CvtBBI->IsDone || 1032 (CvtBBI->CannotBeCopied && CvtBBI->BB->pred_size() > 1)) { 1033 // Something has changed. It's no longer safe to predicate this block. 1034 BBI.IsAnalyzed = false; 1035 CvtBBI->IsAnalyzed = false; 1036 return false; 1037 } 1038 1039 if (CvtBBI->BB->hasAddressTaken()) 1040 // Conservatively abort if-conversion if BB's address is taken. 1041 return false; 1042 1043 if (Kind == ICSimpleFalse) 1044 if (TII->ReverseBranchCondition(Cond)) 1045 llvm_unreachable("Unable to reverse branch condition!"); 1046 1047 // Initialize liveins to the first BB. These are potentiall redefined by 1048 // predicated instructions. 1049 SmallSet<unsigned, 4> Redefs; 1050 InitPredRedefs(CvtBBI->BB, Redefs, TRI); 1051 InitPredRedefs(NextBBI->BB, Redefs, TRI); 1052 1053 if (CvtBBI->BB->pred_size() > 1) { 1054 BBI.NonPredSize -= TII->RemoveBranch(*BBI.BB); 1055 // Copy instructions in the true block, predicate them, and add them to 1056 // the entry block. 1057 CopyAndPredicateBlock(BBI, *CvtBBI, Cond, Redefs); 1058 1059 // RemoveExtraEdges won't work if the block has an unanalyzable branch, so 1060 // explicitly remove CvtBBI as a successor. 1061 BBI.BB->removeSuccessor(CvtBBI->BB); 1062 } else { 1063 PredicateBlock(*CvtBBI, CvtBBI->BB->end(), Cond, Redefs); 1064 1065 // Merge converted block into entry block. 1066 BBI.NonPredSize -= TII->RemoveBranch(*BBI.BB); 1067 MergeBlocks(BBI, *CvtBBI); 1068 } 1069 1070 bool IterIfcvt = true; 1071 if (!canFallThroughTo(BBI.BB, NextBBI->BB)) { 1072 InsertUncondBranch(BBI.BB, NextBBI->BB, TII); 1073 BBI.HasFallThrough = false; 1074 // Now ifcvt'd block will look like this: 1075 // BB: 1076 // ... 1077 // t, f = cmp 1078 // if t op 1079 // b BBf 1080 // 1081 // We cannot further ifcvt this block because the unconditional branch 1082 // will have to be predicated on the new condition, that will not be 1083 // available if cmp executes. 1084 IterIfcvt = false; 1085 } 1086 1087 RemoveExtraEdges(BBI); 1088 1089 // Update block info. BB can be iteratively if-converted. 1090 if (!IterIfcvt) 1091 BBI.IsDone = true; 1092 InvalidatePreds(BBI.BB); 1093 CvtBBI->IsDone = true; 1094 1095 // FIXME: Must maintain LiveIns. 1096 return true; 1097} 1098 1099/// IfConvertTriangle - If convert a triangle sub-CFG. 1100/// 1101bool IfConverter::IfConvertTriangle(BBInfo &BBI, IfcvtKind Kind) { 1102 BBInfo &TrueBBI = BBAnalysis[BBI.TrueBB->getNumber()]; 1103 BBInfo &FalseBBI = BBAnalysis[BBI.FalseBB->getNumber()]; 1104 BBInfo *CvtBBI = &TrueBBI; 1105 BBInfo *NextBBI = &FalseBBI; 1106 DebugLoc dl; // FIXME: this is nowhere 1107 1108 SmallVector<MachineOperand, 4> Cond(BBI.BrCond.begin(), BBI.BrCond.end()); 1109 if (Kind == ICTriangleFalse || Kind == ICTriangleFRev) 1110 std::swap(CvtBBI, NextBBI); 1111 1112 if (CvtBBI->IsDone || 1113 (CvtBBI->CannotBeCopied && CvtBBI->BB->pred_size() > 1)) { 1114 // Something has changed. It's no longer safe to predicate this block. 1115 BBI.IsAnalyzed = false; 1116 CvtBBI->IsAnalyzed = false; 1117 return false; 1118 } 1119 1120 if (CvtBBI->BB->hasAddressTaken()) 1121 // Conservatively abort if-conversion if BB's address is taken. 1122 return false; 1123 1124 if (Kind == ICTriangleFalse || Kind == ICTriangleFRev) 1125 if (TII->ReverseBranchCondition(Cond)) 1126 llvm_unreachable("Unable to reverse branch condition!"); 1127 1128 if (Kind == ICTriangleRev || Kind == ICTriangleFRev) { 1129 if (ReverseBranchCondition(*CvtBBI)) { 1130 // BB has been changed, modify its predecessors (except for this 1131 // one) so they don't get ifcvt'ed based on bad intel. 1132 for (MachineBasicBlock::pred_iterator PI = CvtBBI->BB->pred_begin(), 1133 E = CvtBBI->BB->pred_end(); PI != E; ++PI) { 1134 MachineBasicBlock *PBB = *PI; 1135 if (PBB == BBI.BB) 1136 continue; 1137 BBInfo &PBBI = BBAnalysis[PBB->getNumber()]; 1138 if (PBBI.IsEnqueued) { 1139 PBBI.IsAnalyzed = false; 1140 PBBI.IsEnqueued = false; 1141 } 1142 } 1143 } 1144 } 1145 1146 // Initialize liveins to the first BB. These are potentially redefined by 1147 // predicated instructions. 1148 SmallSet<unsigned, 4> Redefs; 1149 InitPredRedefs(CvtBBI->BB, Redefs, TRI); 1150 InitPredRedefs(NextBBI->BB, Redefs, TRI); 1151 1152 bool HasEarlyExit = CvtBBI->FalseBB != NULL; 1153 if (CvtBBI->BB->pred_size() > 1) { 1154 BBI.NonPredSize -= TII->RemoveBranch(*BBI.BB); 1155 // Copy instructions in the true block, predicate them, and add them to 1156 // the entry block. 1157 CopyAndPredicateBlock(BBI, *CvtBBI, Cond, Redefs, true); 1158 1159 // RemoveExtraEdges won't work if the block has an unanalyzable branch, so 1160 // explicitly remove CvtBBI as a successor. 1161 BBI.BB->removeSuccessor(CvtBBI->BB); 1162 } else { 1163 // Predicate the 'true' block after removing its branch. 1164 CvtBBI->NonPredSize -= TII->RemoveBranch(*CvtBBI->BB); 1165 PredicateBlock(*CvtBBI, CvtBBI->BB->end(), Cond, Redefs); 1166 1167 // Now merge the entry of the triangle with the true block. 1168 BBI.NonPredSize -= TII->RemoveBranch(*BBI.BB); 1169 MergeBlocks(BBI, *CvtBBI, false); 1170 } 1171 1172 // If 'true' block has a 'false' successor, add an exit branch to it. 1173 if (HasEarlyExit) { 1174 SmallVector<MachineOperand, 4> RevCond(CvtBBI->BrCond.begin(), 1175 CvtBBI->BrCond.end()); 1176 if (TII->ReverseBranchCondition(RevCond)) 1177 llvm_unreachable("Unable to reverse branch condition!"); 1178 TII->InsertBranch(*BBI.BB, CvtBBI->FalseBB, NULL, RevCond, dl); 1179 BBI.BB->addSuccessor(CvtBBI->FalseBB); 1180 } 1181 1182 // Merge in the 'false' block if the 'false' block has no other 1183 // predecessors. Otherwise, add an unconditional branch to 'false'. 1184 bool FalseBBDead = false; 1185 bool IterIfcvt = true; 1186 bool isFallThrough = canFallThroughTo(BBI.BB, NextBBI->BB); 1187 if (!isFallThrough) { 1188 // Only merge them if the true block does not fallthrough to the false 1189 // block. By not merging them, we make it possible to iteratively 1190 // ifcvt the blocks. 1191 if (!HasEarlyExit && 1192 NextBBI->BB->pred_size() == 1 && !NextBBI->HasFallThrough && 1193 !NextBBI->BB->hasAddressTaken()) { 1194 MergeBlocks(BBI, *NextBBI); 1195 FalseBBDead = true; 1196 } else { 1197 InsertUncondBranch(BBI.BB, NextBBI->BB, TII); 1198 BBI.HasFallThrough = false; 1199 } 1200 // Mixed predicated and unpredicated code. This cannot be iteratively 1201 // predicated. 1202 IterIfcvt = false; 1203 } 1204 1205 RemoveExtraEdges(BBI); 1206 1207 // Update block info. BB can be iteratively if-converted. 1208 if (!IterIfcvt) 1209 BBI.IsDone = true; 1210 InvalidatePreds(BBI.BB); 1211 CvtBBI->IsDone = true; 1212 if (FalseBBDead) 1213 NextBBI->IsDone = true; 1214 1215 // FIXME: Must maintain LiveIns. 1216 return true; 1217} 1218 1219/// IfConvertDiamond - If convert a diamond sub-CFG. 1220/// 1221bool IfConverter::IfConvertDiamond(BBInfo &BBI, IfcvtKind Kind, 1222 unsigned NumDups1, unsigned NumDups2) { 1223 BBInfo &TrueBBI = BBAnalysis[BBI.TrueBB->getNumber()]; 1224 BBInfo &FalseBBI = BBAnalysis[BBI.FalseBB->getNumber()]; 1225 MachineBasicBlock *TailBB = TrueBBI.TrueBB; 1226 // True block must fall through or end with an unanalyzable terminator. 1227 if (!TailBB) { 1228 if (blockAlwaysFallThrough(TrueBBI)) 1229 TailBB = FalseBBI.TrueBB; 1230 assert((TailBB || !TrueBBI.IsBrAnalyzable) && "Unexpected!"); 1231 } 1232 1233 if (TrueBBI.IsDone || FalseBBI.IsDone || 1234 TrueBBI.BB->pred_size() > 1 || 1235 FalseBBI.BB->pred_size() > 1) { 1236 // Something has changed. It's no longer safe to predicate these blocks. 1237 BBI.IsAnalyzed = false; 1238 TrueBBI.IsAnalyzed = false; 1239 FalseBBI.IsAnalyzed = false; 1240 return false; 1241 } 1242 1243 if (TrueBBI.BB->hasAddressTaken() || FalseBBI.BB->hasAddressTaken()) 1244 // Conservatively abort if-conversion if either BB has its address taken. 1245 return false; 1246 1247 // Put the predicated instructions from the 'true' block before the 1248 // instructions from the 'false' block, unless the true block would clobber 1249 // the predicate, in which case, do the opposite. 1250 BBInfo *BBI1 = &TrueBBI; 1251 BBInfo *BBI2 = &FalseBBI; 1252 SmallVector<MachineOperand, 4> RevCond(BBI.BrCond.begin(), BBI.BrCond.end()); 1253 if (TII->ReverseBranchCondition(RevCond)) 1254 llvm_unreachable("Unable to reverse branch condition!"); 1255 SmallVector<MachineOperand, 4> *Cond1 = &BBI.BrCond; 1256 SmallVector<MachineOperand, 4> *Cond2 = &RevCond; 1257 1258 // Figure out the more profitable ordering. 1259 bool DoSwap = false; 1260 if (TrueBBI.ClobbersPred && !FalseBBI.ClobbersPred) 1261 DoSwap = true; 1262 else if (TrueBBI.ClobbersPred == FalseBBI.ClobbersPred) { 1263 if (TrueBBI.NonPredSize > FalseBBI.NonPredSize) 1264 DoSwap = true; 1265 } 1266 if (DoSwap) { 1267 std::swap(BBI1, BBI2); 1268 std::swap(Cond1, Cond2); 1269 } 1270 1271 // Remove the conditional branch from entry to the blocks. 1272 BBI.NonPredSize -= TII->RemoveBranch(*BBI.BB); 1273 1274 // Initialize liveins to the first BB. These are potentially redefined by 1275 // predicated instructions. 1276 SmallSet<unsigned, 4> Redefs; 1277 InitPredRedefs(BBI1->BB, Redefs, TRI); 1278 1279 // Remove the duplicated instructions at the beginnings of both paths. 1280 MachineBasicBlock::iterator DI1 = BBI1->BB->begin(); 1281 MachineBasicBlock::iterator DI2 = BBI2->BB->begin(); 1282 MachineBasicBlock::iterator DIE1 = BBI1->BB->end(); 1283 MachineBasicBlock::iterator DIE2 = BBI2->BB->end(); 1284 // Skip dbg_value instructions 1285 while (DI1 != DIE1 && DI1->isDebugValue()) 1286 ++DI1; 1287 while (DI2 != DIE2 && DI2->isDebugValue()) 1288 ++DI2; 1289 BBI1->NonPredSize -= NumDups1; 1290 BBI2->NonPredSize -= NumDups1; 1291 1292 // Skip past the dups on each side separately since there may be 1293 // differing dbg_value entries. 1294 for (unsigned i = 0; i < NumDups1; ++DI1) { 1295 if (!DI1->isDebugValue()) 1296 ++i; 1297 } 1298 while (NumDups1 != 0) { 1299 ++DI2; 1300 if (!DI2->isDebugValue()) 1301 --NumDups1; 1302 } 1303 1304 UpdatePredRedefs(BBI1->BB->begin(), DI1, Redefs, TRI); 1305 BBI.BB->splice(BBI.BB->end(), BBI1->BB, BBI1->BB->begin(), DI1); 1306 BBI2->BB->erase(BBI2->BB->begin(), DI2); 1307 1308 // Remove branch from 'true' block and remove duplicated instructions. 1309 BBI1->NonPredSize -= TII->RemoveBranch(*BBI1->BB); 1310 DI1 = BBI1->BB->end(); 1311 for (unsigned i = 0; i != NumDups2; ) { 1312 // NumDups2 only counted non-dbg_value instructions, so this won't 1313 // run off the head of the list. 1314 assert (DI1 != BBI1->BB->begin()); 1315 --DI1; 1316 // skip dbg_value instructions 1317 if (!DI1->isDebugValue()) 1318 ++i; 1319 } 1320 BBI1->BB->erase(DI1, BBI1->BB->end()); 1321 1322 // Remove 'false' block branch and find the last instruction to predicate. 1323 BBI2->NonPredSize -= TII->RemoveBranch(*BBI2->BB); 1324 DI2 = BBI2->BB->end(); 1325 while (NumDups2 != 0) { 1326 // NumDups2 only counted non-dbg_value instructions, so this won't 1327 // run off the head of the list. 1328 assert (DI2 != BBI2->BB->begin()); 1329 --DI2; 1330 // skip dbg_value instructions 1331 if (!DI2->isDebugValue()) 1332 --NumDups2; 1333 } 1334 1335 // Remember which registers would later be defined by the false block. 1336 // This allows us not to predicate instructions in the true block that would 1337 // later be re-defined. That is, rather than 1338 // subeq r0, r1, #1 1339 // addne r0, r1, #1 1340 // generate: 1341 // sub r0, r1, #1 1342 // addne r0, r1, #1 1343 SmallSet<unsigned, 4> RedefsByFalse; 1344 SmallSet<unsigned, 4> ExtUses; 1345 if (TII->isProfitableToUnpredicate(*BBI1->BB, *BBI2->BB)) { 1346 for (MachineBasicBlock::iterator FI = BBI2->BB->begin(); FI != DI2; ++FI) { 1347 if (FI->isDebugValue()) 1348 continue; 1349 SmallVector<unsigned, 4> Defs; 1350 for (unsigned i = 0, e = FI->getNumOperands(); i != e; ++i) { 1351 const MachineOperand &MO = FI->getOperand(i); 1352 if (!MO.isReg()) 1353 continue; 1354 unsigned Reg = MO.getReg(); 1355 if (!Reg) 1356 continue; 1357 if (MO.isDef()) { 1358 Defs.push_back(Reg); 1359 } else if (!RedefsByFalse.count(Reg)) { 1360 // These are defined before ctrl flow reach the 'false' instructions. 1361 // They cannot be modified by the 'true' instructions. 1362 for (MCSubRegIterator SubRegs(Reg, TRI, /*IncludeSelf=*/true); 1363 SubRegs.isValid(); ++SubRegs) 1364 ExtUses.insert(*SubRegs); 1365 } 1366 } 1367 1368 for (unsigned i = 0, e = Defs.size(); i != e; ++i) { 1369 unsigned Reg = Defs[i]; 1370 if (!ExtUses.count(Reg)) { 1371 for (MCSubRegIterator SubRegs(Reg, TRI, /*IncludeSelf=*/true); 1372 SubRegs.isValid(); ++SubRegs) 1373 RedefsByFalse.insert(*SubRegs); 1374 } 1375 } 1376 } 1377 } 1378 1379 // Predicate the 'true' block. 1380 PredicateBlock(*BBI1, BBI1->BB->end(), *Cond1, Redefs, &RedefsByFalse); 1381 1382 // Predicate the 'false' block. 1383 PredicateBlock(*BBI2, DI2, *Cond2, Redefs); 1384 1385 // Merge the true block into the entry of the diamond. 1386 MergeBlocks(BBI, *BBI1, TailBB == 0); 1387 MergeBlocks(BBI, *BBI2, TailBB == 0); 1388 1389 // If the if-converted block falls through or unconditionally branches into 1390 // the tail block, and the tail block does not have other predecessors, then 1391 // fold the tail block in as well. Otherwise, unless it falls through to the 1392 // tail, add a unconditional branch to it. 1393 if (TailBB) { 1394 BBInfo &TailBBI = BBAnalysis[TailBB->getNumber()]; 1395 bool CanMergeTail = !TailBBI.HasFallThrough && 1396 !TailBBI.BB->hasAddressTaken(); 1397 // There may still be a fall-through edge from BBI1 or BBI2 to TailBB; 1398 // check if there are any other predecessors besides those. 1399 unsigned NumPreds = TailBB->pred_size(); 1400 if (NumPreds > 1) 1401 CanMergeTail = false; 1402 else if (NumPreds == 1 && CanMergeTail) { 1403 MachineBasicBlock::pred_iterator PI = TailBB->pred_begin(); 1404 if (*PI != BBI1->BB && *PI != BBI2->BB) 1405 CanMergeTail = false; 1406 } 1407 if (CanMergeTail) { 1408 MergeBlocks(BBI, TailBBI); 1409 TailBBI.IsDone = true; 1410 } else { 1411 BBI.BB->addSuccessor(TailBB); 1412 InsertUncondBranch(BBI.BB, TailBB, TII); 1413 BBI.HasFallThrough = false; 1414 } 1415 } 1416 1417 // RemoveExtraEdges won't work if the block has an unanalyzable branch, 1418 // which can happen here if TailBB is unanalyzable and is merged, so 1419 // explicitly remove BBI1 and BBI2 as successors. 1420 BBI.BB->removeSuccessor(BBI1->BB); 1421 BBI.BB->removeSuccessor(BBI2->BB); 1422 RemoveExtraEdges(BBI); 1423 1424 // Update block info. 1425 BBI.IsDone = TrueBBI.IsDone = FalseBBI.IsDone = true; 1426 InvalidatePreds(BBI.BB); 1427 1428 // FIXME: Must maintain LiveIns. 1429 return true; 1430} 1431 1432static bool MaySpeculate(const MachineInstr *MI, 1433 SmallSet<unsigned, 4> &LaterRedefs, 1434 const TargetInstrInfo *TII) { 1435 bool SawStore = true; 1436 if (!MI->isSafeToMove(TII, 0, SawStore)) 1437 return false; 1438 1439 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) { 1440 const MachineOperand &MO = MI->getOperand(i); 1441 if (!MO.isReg()) 1442 continue; 1443 unsigned Reg = MO.getReg(); 1444 if (!Reg) 1445 continue; 1446 if (MO.isDef() && !LaterRedefs.count(Reg)) 1447 return false; 1448 } 1449 1450 return true; 1451} 1452 1453/// PredicateBlock - Predicate instructions from the start of the block to the 1454/// specified end with the specified condition. 1455void IfConverter::PredicateBlock(BBInfo &BBI, 1456 MachineBasicBlock::iterator E, 1457 SmallVectorImpl<MachineOperand> &Cond, 1458 SmallSet<unsigned, 4> &Redefs, 1459 SmallSet<unsigned, 4> *LaterRedefs) { 1460 bool AnyUnpred = false; 1461 bool MaySpec = LaterRedefs != 0; 1462 for (MachineBasicBlock::iterator I = BBI.BB->begin(); I != E; ++I) { 1463 if (I->isDebugValue() || TII->isPredicated(I)) 1464 continue; 1465 // It may be possible not to predicate an instruction if it's the 'true' 1466 // side of a diamond and the 'false' side may re-define the instruction's 1467 // defs. 1468 if (MaySpec && MaySpeculate(I, *LaterRedefs, TII)) { 1469 AnyUnpred = true; 1470 continue; 1471 } 1472 // If any instruction is predicated, then every instruction after it must 1473 // be predicated. 1474 MaySpec = false; 1475 if (!TII->PredicateInstruction(I, Cond)) { 1476#ifndef NDEBUG 1477 dbgs() << "Unable to predicate " << *I << "!\n"; 1478#endif 1479 llvm_unreachable(0); 1480 } 1481 1482 // If the predicated instruction now redefines a register as the result of 1483 // if-conversion, add an implicit kill. 1484 UpdatePredRedefs(I, Redefs, TRI, true); 1485 } 1486 1487 std::copy(Cond.begin(), Cond.end(), std::back_inserter(BBI.Predicate)); 1488 1489 BBI.IsAnalyzed = false; 1490 BBI.NonPredSize = 0; 1491 1492 ++NumIfConvBBs; 1493 if (AnyUnpred) 1494 ++NumUnpred; 1495} 1496 1497/// CopyAndPredicateBlock - Copy and predicate instructions from source BB to 1498/// the destination block. Skip end of block branches if IgnoreBr is true. 1499void IfConverter::CopyAndPredicateBlock(BBInfo &ToBBI, BBInfo &FromBBI, 1500 SmallVectorImpl<MachineOperand> &Cond, 1501 SmallSet<unsigned, 4> &Redefs, 1502 bool IgnoreBr) { 1503 MachineFunction &MF = *ToBBI.BB->getParent(); 1504 1505 for (MachineBasicBlock::iterator I = FromBBI.BB->begin(), 1506 E = FromBBI.BB->end(); I != E; ++I) { 1507 // Do not copy the end of the block branches. 1508 if (IgnoreBr && I->isBranch()) 1509 break; 1510 1511 MachineInstr *MI = MF.CloneMachineInstr(I); 1512 ToBBI.BB->insert(ToBBI.BB->end(), MI); 1513 ToBBI.NonPredSize++; 1514 unsigned ExtraPredCost = 0; 1515 unsigned NumCycles = TII->getInstrLatency(InstrItins, &*I, &ExtraPredCost); 1516 if (NumCycles > 1) 1517 ToBBI.ExtraCost += NumCycles-1; 1518 ToBBI.ExtraCost2 += ExtraPredCost; 1519 1520 if (!TII->isPredicated(I) && !MI->isDebugValue()) { 1521 if (!TII->PredicateInstruction(MI, Cond)) { 1522#ifndef NDEBUG 1523 dbgs() << "Unable to predicate " << *I << "!\n"; 1524#endif 1525 llvm_unreachable(0); 1526 } 1527 } 1528 1529 // If the predicated instruction now redefines a register as the result of 1530 // if-conversion, add an implicit kill. 1531 UpdatePredRedefs(MI, Redefs, TRI, true); 1532 } 1533 1534 if (!IgnoreBr) { 1535 std::vector<MachineBasicBlock *> Succs(FromBBI.BB->succ_begin(), 1536 FromBBI.BB->succ_end()); 1537 MachineBasicBlock *NBB = getNextBlock(FromBBI.BB); 1538 MachineBasicBlock *FallThrough = FromBBI.HasFallThrough ? NBB : NULL; 1539 1540 for (unsigned i = 0, e = Succs.size(); i != e; ++i) { 1541 MachineBasicBlock *Succ = Succs[i]; 1542 // Fallthrough edge can't be transferred. 1543 if (Succ == FallThrough) 1544 continue; 1545 ToBBI.BB->addSuccessor(Succ); 1546 } 1547 } 1548 1549 std::copy(FromBBI.Predicate.begin(), FromBBI.Predicate.end(), 1550 std::back_inserter(ToBBI.Predicate)); 1551 std::copy(Cond.begin(), Cond.end(), std::back_inserter(ToBBI.Predicate)); 1552 1553 ToBBI.ClobbersPred |= FromBBI.ClobbersPred; 1554 ToBBI.IsAnalyzed = false; 1555 1556 ++NumDupBBs; 1557} 1558 1559/// MergeBlocks - Move all instructions from FromBB to the end of ToBB. 1560/// This will leave FromBB as an empty block, so remove all of its 1561/// successor edges except for the fall-through edge. If AddEdges is true, 1562/// i.e., when FromBBI's branch is being moved, add those successor edges to 1563/// ToBBI. 1564void IfConverter::MergeBlocks(BBInfo &ToBBI, BBInfo &FromBBI, bool AddEdges) { 1565 assert(!FromBBI.BB->hasAddressTaken() && 1566 "Removing a BB whose address is taken!"); 1567 1568 ToBBI.BB->splice(ToBBI.BB->end(), 1569 FromBBI.BB, FromBBI.BB->begin(), FromBBI.BB->end()); 1570 1571 std::vector<MachineBasicBlock *> Succs(FromBBI.BB->succ_begin(), 1572 FromBBI.BB->succ_end()); 1573 MachineBasicBlock *NBB = getNextBlock(FromBBI.BB); 1574 MachineBasicBlock *FallThrough = FromBBI.HasFallThrough ? NBB : NULL; 1575 1576 for (unsigned i = 0, e = Succs.size(); i != e; ++i) { 1577 MachineBasicBlock *Succ = Succs[i]; 1578 // Fallthrough edge can't be transferred. 1579 if (Succ == FallThrough) 1580 continue; 1581 FromBBI.BB->removeSuccessor(Succ); 1582 if (AddEdges && !ToBBI.BB->isSuccessor(Succ)) 1583 ToBBI.BB->addSuccessor(Succ); 1584 } 1585 1586 // Now FromBBI always falls through to the next block! 1587 if (NBB && !FromBBI.BB->isSuccessor(NBB)) 1588 FromBBI.BB->addSuccessor(NBB); 1589 1590 std::copy(FromBBI.Predicate.begin(), FromBBI.Predicate.end(), 1591 std::back_inserter(ToBBI.Predicate)); 1592 FromBBI.Predicate.clear(); 1593 1594 ToBBI.NonPredSize += FromBBI.NonPredSize; 1595 ToBBI.ExtraCost += FromBBI.ExtraCost; 1596 ToBBI.ExtraCost2 += FromBBI.ExtraCost2; 1597 FromBBI.NonPredSize = 0; 1598 FromBBI.ExtraCost = 0; 1599 FromBBI.ExtraCost2 = 0; 1600 1601 ToBBI.ClobbersPred |= FromBBI.ClobbersPred; 1602 ToBBI.HasFallThrough = FromBBI.HasFallThrough; 1603 ToBBI.IsAnalyzed = false; 1604 FromBBI.IsAnalyzed = false; 1605} 1606