Dan Gohman | f17a25c | 2007-07-18 16:29:46 +0000 | [diff] [blame] | 1 | //===-- IfConversion.cpp - Machine code if conversion pass. ---------------===// |
| 2 | // |
| 3 | // The LLVM Compiler Infrastructure |
| 4 | // |
Chris Lattner | 081ce94 | 2007-12-29 20:36:04 +0000 | [diff] [blame] | 5 | // This file is distributed under the University of Illinois Open Source |
| 6 | // License. See LICENSE.TXT for details. |
Dan Gohman | f17a25c | 2007-07-18 16:29:46 +0000 | [diff] [blame] | 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/Function.h" |
| 16 | #include "llvm/CodeGen/Passes.h" |
| 17 | #include "llvm/CodeGen/MachineModuleInfo.h" |
| 18 | #include "llvm/CodeGen/MachineFunctionPass.h" |
| 19 | #include "llvm/Target/TargetInstrInfo.h" |
| 20 | #include "llvm/Target/TargetLowering.h" |
| 21 | #include "llvm/Target/TargetMachine.h" |
| 22 | #include "llvm/Support/CommandLine.h" |
| 23 | #include "llvm/Support/Debug.h" |
| 24 | #include "llvm/ADT/DepthFirstIterator.h" |
| 25 | #include "llvm/ADT/Statistic.h" |
| 26 | #include "llvm/ADT/STLExtras.h" |
| 27 | using namespace llvm; |
| 28 | |
Chris Lattner | 4335bff | 2008-01-07 05:40:58 +0000 | [diff] [blame] | 29 | // Hidden options for help debugging. |
| 30 | static cl::opt<int> IfCvtFnStart("ifcvt-fn-start", cl::init(-1), cl::Hidden); |
| 31 | static cl::opt<int> IfCvtFnStop("ifcvt-fn-stop", cl::init(-1), cl::Hidden); |
| 32 | static cl::opt<int> IfCvtLimit("ifcvt-limit", cl::init(-1), cl::Hidden); |
| 33 | static cl::opt<bool> DisableSimple("disable-ifcvt-simple", |
| 34 | cl::init(false), cl::Hidden); |
| 35 | static cl::opt<bool> DisableSimpleF("disable-ifcvt-simple-false", |
| 36 | cl::init(false), cl::Hidden); |
| 37 | static cl::opt<bool> DisableTriangle("disable-ifcvt-triangle", |
| 38 | cl::init(false), cl::Hidden); |
| 39 | static cl::opt<bool> DisableTriangleR("disable-ifcvt-triangle-rev", |
| 40 | cl::init(false), cl::Hidden); |
| 41 | static cl::opt<bool> DisableTriangleF("disable-ifcvt-triangle-false", |
| 42 | cl::init(false), cl::Hidden); |
| 43 | static cl::opt<bool> DisableTriangleFR("disable-ifcvt-triangle-false-rev", |
| 44 | cl::init(false), cl::Hidden); |
| 45 | static cl::opt<bool> DisableDiamond("disable-ifcvt-diamond", |
| 46 | cl::init(false), cl::Hidden); |
Dan Gohman | f17a25c | 2007-07-18 16:29:46 +0000 | [diff] [blame] | 47 | |
| 48 | STATISTIC(NumSimple, "Number of simple if-conversions performed"); |
| 49 | STATISTIC(NumSimpleFalse, "Number of simple (F) if-conversions performed"); |
| 50 | STATISTIC(NumTriangle, "Number of triangle if-conversions performed"); |
| 51 | STATISTIC(NumTriangleRev, "Number of triangle (R) if-conversions performed"); |
| 52 | STATISTIC(NumTriangleFalse,"Number of triangle (F) if-conversions performed"); |
| 53 | STATISTIC(NumTriangleFRev, "Number of triangle (F/R) if-conversions performed"); |
| 54 | STATISTIC(NumDiamonds, "Number of diamond if-conversions performed"); |
| 55 | STATISTIC(NumIfConvBBs, "Number of if-converted blocks"); |
| 56 | STATISTIC(NumDupBBs, "Number of duplicated blocks"); |
| 57 | |
| 58 | namespace { |
| 59 | class IfConverter : public MachineFunctionPass { |
| 60 | enum IfcvtKind { |
| 61 | ICNotClassfied, // BB data valid, but not classified. |
| 62 | ICSimpleFalse, // Same as ICSimple, but on the false path. |
| 63 | ICSimple, // BB is entry of an one split, no rejoin sub-CFG. |
| 64 | ICTriangleFRev, // Same as ICTriangleFalse, but false path rev condition. |
| 65 | ICTriangleRev, // Same as ICTriangle, but true path rev condition. |
| 66 | ICTriangleFalse, // Same as ICTriangle, but on the false path. |
| 67 | ICTriangle, // BB is entry of a triangle sub-CFG. |
| 68 | ICDiamond // BB is entry of a diamond sub-CFG. |
| 69 | }; |
| 70 | |
| 71 | /// BBInfo - One per MachineBasicBlock, this is used to cache the result |
| 72 | /// if-conversion feasibility analysis. This includes results from |
| 73 | /// TargetInstrInfo::AnalyzeBranch() (i.e. TBB, FBB, and Cond), and its |
| 74 | /// classification, and common tail block of its successors (if it's a |
| 75 | /// diamond shape), its size, whether it's predicable, and whether any |
| 76 | /// instruction can clobber the 'would-be' predicate. |
| 77 | /// |
| 78 | /// IsDone - True if BB is not to be considered for ifcvt. |
| 79 | /// IsBeingAnalyzed - True if BB is currently being analyzed. |
| 80 | /// IsAnalyzed - True if BB has been analyzed (info is still valid). |
| 81 | /// IsEnqueued - True if BB has been enqueued to be ifcvt'ed. |
| 82 | /// IsBrAnalyzable - True if AnalyzeBranch() returns false. |
| 83 | /// HasFallThrough - True if BB may fallthrough to the following BB. |
| 84 | /// IsUnpredicable - True if BB is known to be unpredicable. |
| 85 | /// ClobbersPred - True if BB could modify predicates (e.g. has |
| 86 | /// cmp, call, etc.) |
| 87 | /// NonPredSize - Number of non-predicated instructions. |
| 88 | /// BB - Corresponding MachineBasicBlock. |
| 89 | /// TrueBB / FalseBB- See AnalyzeBranch(). |
| 90 | /// BrCond - Conditions for end of block conditional branches. |
| 91 | /// Predicate - Predicate used in the BB. |
| 92 | struct BBInfo { |
| 93 | bool IsDone : 1; |
| 94 | bool IsBeingAnalyzed : 1; |
| 95 | bool IsAnalyzed : 1; |
| 96 | bool IsEnqueued : 1; |
| 97 | bool IsBrAnalyzable : 1; |
| 98 | bool HasFallThrough : 1; |
| 99 | bool IsUnpredicable : 1; |
| 100 | bool CannotBeCopied : 1; |
| 101 | bool ClobbersPred : 1; |
| 102 | unsigned NonPredSize; |
| 103 | MachineBasicBlock *BB; |
| 104 | MachineBasicBlock *TrueBB; |
| 105 | MachineBasicBlock *FalseBB; |
| 106 | std::vector<MachineOperand> BrCond; |
| 107 | std::vector<MachineOperand> Predicate; |
| 108 | BBInfo() : IsDone(false), IsBeingAnalyzed(false), |
| 109 | IsAnalyzed(false), IsEnqueued(false), IsBrAnalyzable(false), |
| 110 | HasFallThrough(false), IsUnpredicable(false), |
| 111 | CannotBeCopied(false), ClobbersPred(false), NonPredSize(0), |
| 112 | BB(0), TrueBB(0), FalseBB(0) {} |
| 113 | }; |
| 114 | |
| 115 | /// IfcvtToken - Record information about pending if-conversions to attemp: |
| 116 | /// BBI - Corresponding BBInfo. |
| 117 | /// Kind - Type of block. See IfcvtKind. |
| 118 | /// NeedSubsumsion - True if the to be predicated BB has already been |
| 119 | /// predicated. |
| 120 | /// NumDups - Number of instructions that would be duplicated due |
| 121 | /// to this if-conversion. (For diamonds, the number of |
| 122 | /// identical instructions at the beginnings of both |
| 123 | /// paths). |
| 124 | /// NumDups2 - For diamonds, the number of identical instructions |
| 125 | /// at the ends of both paths. |
| 126 | struct IfcvtToken { |
| 127 | BBInfo &BBI; |
| 128 | IfcvtKind Kind; |
| 129 | bool NeedSubsumsion; |
| 130 | unsigned NumDups; |
| 131 | unsigned NumDups2; |
| 132 | IfcvtToken(BBInfo &b, IfcvtKind k, bool s, unsigned d, unsigned d2 = 0) |
| 133 | : BBI(b), Kind(k), NeedSubsumsion(s), NumDups(d), NumDups2(d2) {} |
| 134 | }; |
| 135 | |
| 136 | /// Roots - Basic blocks that do not have successors. These are the starting |
| 137 | /// points of Graph traversal. |
| 138 | std::vector<MachineBasicBlock*> Roots; |
| 139 | |
| 140 | /// BBAnalysis - Results of if-conversion feasibility analysis indexed by |
| 141 | /// basic block number. |
| 142 | std::vector<BBInfo> BBAnalysis; |
| 143 | |
| 144 | const TargetLowering *TLI; |
| 145 | const TargetInstrInfo *TII; |
| 146 | bool MadeChange; |
| 147 | public: |
| 148 | static char ID; |
| 149 | IfConverter() : MachineFunctionPass((intptr_t)&ID) {} |
| 150 | |
| 151 | virtual bool runOnMachineFunction(MachineFunction &MF); |
| 152 | virtual const char *getPassName() const { return "If converter"; } |
| 153 | |
| 154 | private: |
| 155 | bool ReverseBranchCondition(BBInfo &BBI); |
| 156 | bool ValidSimple(BBInfo &TrueBBI, unsigned &Dups) const; |
| 157 | bool ValidTriangle(BBInfo &TrueBBI, BBInfo &FalseBBI, |
| 158 | bool FalseBranch, unsigned &Dups) const; |
| 159 | bool ValidDiamond(BBInfo &TrueBBI, BBInfo &FalseBBI, |
| 160 | unsigned &Dups1, unsigned &Dups2) const; |
| 161 | void ScanInstructions(BBInfo &BBI); |
| 162 | BBInfo &AnalyzeBlock(MachineBasicBlock *BB, |
| 163 | std::vector<IfcvtToken*> &Tokens); |
| 164 | bool FeasibilityAnalysis(BBInfo &BBI, std::vector<MachineOperand> &Cond, |
| 165 | bool isTriangle = false, bool RevBranch = false); |
| 166 | bool AnalyzeBlocks(MachineFunction &MF, |
| 167 | std::vector<IfcvtToken*> &Tokens); |
| 168 | void InvalidatePreds(MachineBasicBlock *BB); |
| 169 | void RemoveExtraEdges(BBInfo &BBI); |
| 170 | bool IfConvertSimple(BBInfo &BBI, IfcvtKind Kind); |
| 171 | bool IfConvertTriangle(BBInfo &BBI, IfcvtKind Kind); |
| 172 | bool IfConvertDiamond(BBInfo &BBI, IfcvtKind Kind, |
| 173 | unsigned NumDups1, unsigned NumDups2); |
| 174 | void PredicateBlock(BBInfo &BBI, |
| 175 | MachineBasicBlock::iterator E, |
| 176 | std::vector<MachineOperand> &Cond); |
| 177 | void CopyAndPredicateBlock(BBInfo &ToBBI, BBInfo &FromBBI, |
| 178 | std::vector<MachineOperand> &Cond, |
| 179 | bool IgnoreBr = false); |
| 180 | void MergeBlocks(BBInfo &ToBBI, BBInfo &FromBBI); |
| 181 | |
| 182 | bool MeetIfcvtSizeLimit(unsigned Size) const { |
| 183 | return Size > 0 && Size <= TLI->getIfCvtBlockSizeLimit(); |
| 184 | } |
| 185 | |
| 186 | // blockAlwaysFallThrough - Block ends without a terminator. |
| 187 | bool blockAlwaysFallThrough(BBInfo &BBI) const { |
| 188 | return BBI.IsBrAnalyzable && BBI.TrueBB == NULL; |
| 189 | } |
| 190 | |
| 191 | // IfcvtTokenCmp - Used to sort if-conversion candidates. |
| 192 | static bool IfcvtTokenCmp(IfcvtToken *C1, IfcvtToken *C2) { |
| 193 | int Incr1 = (C1->Kind == ICDiamond) |
| 194 | ? -(int)(C1->NumDups + C1->NumDups2) : (int)C1->NumDups; |
| 195 | int Incr2 = (C2->Kind == ICDiamond) |
| 196 | ? -(int)(C2->NumDups + C2->NumDups2) : (int)C2->NumDups; |
| 197 | if (Incr1 > Incr2) |
| 198 | return true; |
| 199 | else if (Incr1 == Incr2) { |
| 200 | // Favors subsumsion. |
| 201 | if (C1->NeedSubsumsion == false && C2->NeedSubsumsion == true) |
| 202 | return true; |
| 203 | else if (C1->NeedSubsumsion == C2->NeedSubsumsion) { |
| 204 | // Favors diamond over triangle, etc. |
| 205 | if ((unsigned)C1->Kind < (unsigned)C2->Kind) |
| 206 | return true; |
| 207 | else if (C1->Kind == C2->Kind) |
| 208 | return C1->BBI.BB->getNumber() < C2->BBI.BB->getNumber(); |
| 209 | } |
| 210 | } |
| 211 | return false; |
| 212 | } |
| 213 | }; |
| 214 | |
| 215 | char IfConverter::ID = 0; |
| 216 | } |
| 217 | |
| 218 | FunctionPass *llvm::createIfConverterPass() { return new IfConverter(); } |
| 219 | |
| 220 | bool IfConverter::runOnMachineFunction(MachineFunction &MF) { |
| 221 | TLI = MF.getTarget().getTargetLowering(); |
| 222 | TII = MF.getTarget().getInstrInfo(); |
| 223 | if (!TII) return false; |
| 224 | |
| 225 | static int FnNum = -1; |
| 226 | DOUT << "\nIfcvt: function (" << ++FnNum << ") \'" |
| 227 | << MF.getFunction()->getName() << "\'"; |
| 228 | |
| 229 | if (FnNum < IfCvtFnStart || (IfCvtFnStop != -1 && FnNum > IfCvtFnStop)) { |
| 230 | DOUT << " skipped\n"; |
| 231 | return false; |
| 232 | } |
| 233 | DOUT << "\n"; |
| 234 | |
| 235 | MF.RenumberBlocks(); |
| 236 | BBAnalysis.resize(MF.getNumBlockIDs()); |
| 237 | |
| 238 | // Look for root nodes, i.e. blocks without successors. |
| 239 | for (MachineFunction::iterator I = MF.begin(), E = MF.end(); I != E; ++I) |
| 240 | if (I->succ_size() == 0) |
| 241 | Roots.push_back(I); |
| 242 | |
| 243 | std::vector<IfcvtToken*> Tokens; |
| 244 | MadeChange = false; |
| 245 | unsigned NumIfCvts = NumSimple + NumSimpleFalse + NumTriangle + |
| 246 | NumTriangleRev + NumTriangleFalse + NumTriangleFRev + NumDiamonds; |
| 247 | while (IfCvtLimit == -1 || (int)NumIfCvts < IfCvtLimit) { |
| 248 | // Do an intial analysis for each basic block and finding all the potential |
| 249 | // candidates to perform if-convesion. |
| 250 | bool Change = AnalyzeBlocks(MF, Tokens); |
| 251 | while (!Tokens.empty()) { |
| 252 | IfcvtToken *Token = Tokens.back(); |
| 253 | Tokens.pop_back(); |
| 254 | BBInfo &BBI = Token->BBI; |
| 255 | IfcvtKind Kind = Token->Kind; |
| 256 | |
| 257 | // If the block has been evicted out of the queue or it has already been |
| 258 | // marked dead (due to it being predicated), then skip it. |
| 259 | if (BBI.IsDone) |
| 260 | BBI.IsEnqueued = false; |
| 261 | if (!BBI.IsEnqueued) |
| 262 | continue; |
| 263 | |
| 264 | BBI.IsEnqueued = false; |
| 265 | |
| 266 | bool RetVal = false; |
| 267 | switch (Kind) { |
| 268 | default: assert(false && "Unexpected!"); |
| 269 | break; |
| 270 | case ICSimple: |
| 271 | case ICSimpleFalse: { |
| 272 | bool isFalse = Kind == ICSimpleFalse; |
| 273 | if ((isFalse && DisableSimpleF) || (!isFalse && DisableSimple)) break; |
| 274 | DOUT << "Ifcvt (Simple" << (Kind == ICSimpleFalse ? " false" :"") |
| 275 | << "): BB#" << BBI.BB->getNumber() << " (" |
| 276 | << ((Kind == ICSimpleFalse) |
| 277 | ? BBI.FalseBB->getNumber() |
| 278 | : BBI.TrueBB->getNumber()) << ") "; |
| 279 | RetVal = IfConvertSimple(BBI, Kind); |
| 280 | DOUT << (RetVal ? "succeeded!" : "failed!") << "\n"; |
| 281 | if (RetVal) |
| 282 | if (isFalse) NumSimpleFalse++; |
| 283 | else NumSimple++; |
| 284 | break; |
| 285 | } |
| 286 | case ICTriangle: |
| 287 | case ICTriangleRev: |
| 288 | case ICTriangleFalse: |
| 289 | case ICTriangleFRev: { |
| 290 | bool isFalse = Kind == ICTriangleFalse; |
| 291 | bool isRev = (Kind == ICTriangleRev || Kind == ICTriangleFRev); |
| 292 | if (DisableTriangle && !isFalse && !isRev) break; |
| 293 | if (DisableTriangleR && !isFalse && isRev) break; |
| 294 | if (DisableTriangleF && isFalse && !isRev) break; |
| 295 | if (DisableTriangleFR && isFalse && isRev) break; |
| 296 | DOUT << "Ifcvt (Triangle"; |
| 297 | if (isFalse) |
| 298 | DOUT << " false"; |
| 299 | if (isRev) |
| 300 | DOUT << " rev"; |
| 301 | DOUT << "): BB#" << BBI.BB->getNumber() << " (T:" |
| 302 | << BBI.TrueBB->getNumber() << ",F:" |
| 303 | << BBI.FalseBB->getNumber() << ") "; |
| 304 | RetVal = IfConvertTriangle(BBI, Kind); |
| 305 | DOUT << (RetVal ? "succeeded!" : "failed!") << "\n"; |
| 306 | if (RetVal) { |
| 307 | if (isFalse) { |
| 308 | if (isRev) NumTriangleFRev++; |
| 309 | else NumTriangleFalse++; |
| 310 | } else { |
| 311 | if (isRev) NumTriangleRev++; |
| 312 | else NumTriangle++; |
| 313 | } |
| 314 | } |
| 315 | break; |
| 316 | } |
| 317 | case ICDiamond: { |
| 318 | if (DisableDiamond) break; |
| 319 | DOUT << "Ifcvt (Diamond): BB#" << BBI.BB->getNumber() << " (T:" |
| 320 | << BBI.TrueBB->getNumber() << ",F:" |
| 321 | << BBI.FalseBB->getNumber() << ") "; |
| 322 | RetVal = IfConvertDiamond(BBI, Kind, Token->NumDups, Token->NumDups2); |
| 323 | DOUT << (RetVal ? "succeeded!" : "failed!") << "\n"; |
| 324 | if (RetVal) NumDiamonds++; |
| 325 | break; |
| 326 | } |
| 327 | } |
| 328 | |
| 329 | Change |= RetVal; |
| 330 | |
| 331 | NumIfCvts = NumSimple + NumSimpleFalse + NumTriangle + NumTriangleRev + |
| 332 | NumTriangleFalse + NumTriangleFRev + NumDiamonds; |
| 333 | if (IfCvtLimit != -1 && (int)NumIfCvts >= IfCvtLimit) |
| 334 | break; |
| 335 | } |
| 336 | |
| 337 | if (!Change) |
| 338 | break; |
| 339 | MadeChange |= Change; |
| 340 | } |
| 341 | |
| 342 | // Delete tokens in case of early exit. |
| 343 | while (!Tokens.empty()) { |
| 344 | IfcvtToken *Token = Tokens.back(); |
| 345 | Tokens.pop_back(); |
| 346 | delete Token; |
| 347 | } |
| 348 | |
| 349 | Tokens.clear(); |
| 350 | Roots.clear(); |
| 351 | BBAnalysis.clear(); |
| 352 | |
| 353 | return MadeChange; |
| 354 | } |
| 355 | |
| 356 | /// findFalseBlock - BB has a fallthrough. Find its 'false' successor given |
| 357 | /// its 'true' successor. |
| 358 | static MachineBasicBlock *findFalseBlock(MachineBasicBlock *BB, |
| 359 | MachineBasicBlock *TrueBB) { |
| 360 | for (MachineBasicBlock::succ_iterator SI = BB->succ_begin(), |
| 361 | E = BB->succ_end(); SI != E; ++SI) { |
| 362 | MachineBasicBlock *SuccBB = *SI; |
| 363 | if (SuccBB != TrueBB) |
| 364 | return SuccBB; |
| 365 | } |
| 366 | return NULL; |
| 367 | } |
| 368 | |
| 369 | /// ReverseBranchCondition - Reverse the condition of the end of the block |
| 370 | /// branchs. Swap block's 'true' and 'false' successors. |
| 371 | bool IfConverter::ReverseBranchCondition(BBInfo &BBI) { |
| 372 | if (!TII->ReverseBranchCondition(BBI.BrCond)) { |
| 373 | TII->RemoveBranch(*BBI.BB); |
| 374 | TII->InsertBranch(*BBI.BB, BBI.FalseBB, BBI.TrueBB, BBI.BrCond); |
| 375 | std::swap(BBI.TrueBB, BBI.FalseBB); |
| 376 | return true; |
| 377 | } |
| 378 | return false; |
| 379 | } |
| 380 | |
| 381 | /// getNextBlock - Returns the next block in the function blocks ordering. If |
| 382 | /// it is the end, returns NULL. |
| 383 | static inline MachineBasicBlock *getNextBlock(MachineBasicBlock *BB) { |
| 384 | MachineFunction::iterator I = BB; |
| 385 | MachineFunction::iterator E = BB->getParent()->end(); |
| 386 | if (++I == E) |
| 387 | return NULL; |
| 388 | return I; |
| 389 | } |
| 390 | |
| 391 | /// ValidSimple - Returns true if the 'true' block (along with its |
| 392 | /// predecessor) forms a valid simple shape for ifcvt. It also returns the |
| 393 | /// number of instructions that the ifcvt would need to duplicate if performed |
| 394 | /// in Dups. |
| 395 | bool IfConverter::ValidSimple(BBInfo &TrueBBI, unsigned &Dups) const { |
| 396 | Dups = 0; |
| 397 | if (TrueBBI.IsBeingAnalyzed || TrueBBI.IsDone) |
| 398 | return false; |
| 399 | |
| 400 | if (TrueBBI.IsBrAnalyzable) |
| 401 | return false; |
| 402 | |
| 403 | if (TrueBBI.BB->pred_size() > 1) { |
| 404 | if (TrueBBI.CannotBeCopied || |
| 405 | TrueBBI.NonPredSize > TLI->getIfCvtDupBlockSizeLimit()) |
| 406 | return false; |
| 407 | Dups = TrueBBI.NonPredSize; |
| 408 | } |
| 409 | |
| 410 | return true; |
| 411 | } |
| 412 | |
| 413 | /// ValidTriangle - Returns true if the 'true' and 'false' blocks (along |
| 414 | /// with their common predecessor) forms a valid triangle shape for ifcvt. |
| 415 | /// If 'FalseBranch' is true, it checks if 'true' block's false branch |
| 416 | /// branches to the false branch rather than the other way around. It also |
| 417 | /// returns the number of instructions that the ifcvt would need to duplicate |
| 418 | /// if performed in 'Dups'. |
| 419 | bool IfConverter::ValidTriangle(BBInfo &TrueBBI, BBInfo &FalseBBI, |
| 420 | bool FalseBranch, unsigned &Dups) const { |
| 421 | Dups = 0; |
| 422 | if (TrueBBI.IsBeingAnalyzed || TrueBBI.IsDone) |
| 423 | return false; |
| 424 | |
| 425 | if (TrueBBI.BB->pred_size() > 1) { |
| 426 | if (TrueBBI.CannotBeCopied) |
| 427 | return false; |
| 428 | |
| 429 | unsigned Size = TrueBBI.NonPredSize; |
| 430 | if (TrueBBI.IsBrAnalyzable) { |
| 431 | if (TrueBBI.TrueBB && TrueBBI.BrCond.size() == 0) |
| 432 | // End with an unconditional branch. It will be removed. |
| 433 | --Size; |
| 434 | else { |
| 435 | MachineBasicBlock *FExit = FalseBranch |
| 436 | ? TrueBBI.TrueBB : TrueBBI.FalseBB; |
| 437 | if (FExit) |
| 438 | // Require a conditional branch |
| 439 | ++Size; |
| 440 | } |
| 441 | } |
| 442 | if (Size > TLI->getIfCvtDupBlockSizeLimit()) |
| 443 | return false; |
| 444 | Dups = Size; |
| 445 | } |
| 446 | |
| 447 | MachineBasicBlock *TExit = FalseBranch ? TrueBBI.FalseBB : TrueBBI.TrueBB; |
| 448 | if (!TExit && blockAlwaysFallThrough(TrueBBI)) { |
| 449 | MachineFunction::iterator I = TrueBBI.BB; |
| 450 | if (++I == TrueBBI.BB->getParent()->end()) |
| 451 | return false; |
| 452 | TExit = I; |
| 453 | } |
| 454 | return TExit && TExit == FalseBBI.BB; |
| 455 | } |
| 456 | |
| 457 | static |
| 458 | MachineBasicBlock::iterator firstNonBranchInst(MachineBasicBlock *BB, |
| 459 | const TargetInstrInfo *TII) { |
| 460 | MachineBasicBlock::iterator I = BB->end(); |
| 461 | while (I != BB->begin()) { |
| 462 | --I; |
Chris Lattner | 5b93037 | 2008-01-07 07:27:27 +0000 | [diff] [blame^] | 463 | if (!I->getDesc().isBranch()) |
Dan Gohman | f17a25c | 2007-07-18 16:29:46 +0000 | [diff] [blame] | 464 | break; |
| 465 | } |
| 466 | return I; |
| 467 | } |
| 468 | |
| 469 | /// ValidDiamond - Returns true if the 'true' and 'false' blocks (along |
| 470 | /// with their common predecessor) forms a valid diamond shape for ifcvt. |
| 471 | bool IfConverter::ValidDiamond(BBInfo &TrueBBI, BBInfo &FalseBBI, |
| 472 | unsigned &Dups1, unsigned &Dups2) const { |
| 473 | Dups1 = Dups2 = 0; |
| 474 | if (TrueBBI.IsBeingAnalyzed || TrueBBI.IsDone || |
| 475 | FalseBBI.IsBeingAnalyzed || FalseBBI.IsDone) |
| 476 | return false; |
| 477 | |
| 478 | MachineBasicBlock *TT = TrueBBI.TrueBB; |
| 479 | MachineBasicBlock *FT = FalseBBI.TrueBB; |
| 480 | |
| 481 | if (!TT && blockAlwaysFallThrough(TrueBBI)) |
| 482 | TT = getNextBlock(TrueBBI.BB); |
| 483 | if (!FT && blockAlwaysFallThrough(FalseBBI)) |
| 484 | FT = getNextBlock(FalseBBI.BB); |
| 485 | if (TT != FT) |
| 486 | return false; |
| 487 | if (TT == NULL && (TrueBBI.IsBrAnalyzable || FalseBBI.IsBrAnalyzable)) |
| 488 | return false; |
| 489 | if (TrueBBI.BB->pred_size() > 1 || FalseBBI.BB->pred_size() > 1) |
| 490 | return false; |
| 491 | |
| 492 | // FIXME: Allow true block to have an early exit? |
| 493 | if (TrueBBI.FalseBB || FalseBBI.FalseBB || |
| 494 | (TrueBBI.ClobbersPred && FalseBBI.ClobbersPred)) |
| 495 | return false; |
| 496 | |
| 497 | MachineBasicBlock::iterator TI = TrueBBI.BB->begin(); |
| 498 | MachineBasicBlock::iterator FI = FalseBBI.BB->begin(); |
| 499 | while (TI != TrueBBI.BB->end() && FI != FalseBBI.BB->end()) { |
| 500 | if (!TI->isIdenticalTo(FI)) |
| 501 | break; |
| 502 | ++Dups1; |
| 503 | ++TI; |
| 504 | ++FI; |
| 505 | } |
| 506 | |
| 507 | TI = firstNonBranchInst(TrueBBI.BB, TII); |
| 508 | FI = firstNonBranchInst(FalseBBI.BB, TII); |
| 509 | while (TI != TrueBBI.BB->begin() && FI != FalseBBI.BB->begin()) { |
| 510 | if (!TI->isIdenticalTo(FI)) |
| 511 | break; |
| 512 | ++Dups2; |
| 513 | --TI; |
| 514 | --FI; |
| 515 | } |
| 516 | |
| 517 | return true; |
| 518 | } |
| 519 | |
| 520 | /// ScanInstructions - Scan all the instructions in the block to determine if |
| 521 | /// the block is predicable. In most cases, that means all the instructions |
Chris Lattner | 6232760 | 2008-01-07 01:56:04 +0000 | [diff] [blame] | 522 | /// in the block are isPredicable(). Also checks if the block contains any |
Dan Gohman | f17a25c | 2007-07-18 16:29:46 +0000 | [diff] [blame] | 523 | /// instruction which can clobber a predicate (e.g. condition code register). |
| 524 | /// If so, the block is not predicable unless it's the last instruction. |
| 525 | void IfConverter::ScanInstructions(BBInfo &BBI) { |
| 526 | if (BBI.IsDone) |
| 527 | return; |
| 528 | |
| 529 | bool AlreadyPredicated = BBI.Predicate.size() > 0; |
| 530 | // First analyze the end of BB branches. |
| 531 | BBI.TrueBB = BBI.FalseBB = NULL; |
| 532 | BBI.BrCond.clear(); |
| 533 | BBI.IsBrAnalyzable = |
| 534 | !TII->AnalyzeBranch(*BBI.BB, BBI.TrueBB, BBI.FalseBB, BBI.BrCond); |
| 535 | BBI.HasFallThrough = BBI.IsBrAnalyzable && BBI.FalseBB == NULL; |
| 536 | |
| 537 | if (BBI.BrCond.size()) { |
| 538 | // No false branch. This BB must end with a conditional branch and a |
| 539 | // fallthrough. |
| 540 | if (!BBI.FalseBB) |
| 541 | BBI.FalseBB = findFalseBlock(BBI.BB, BBI.TrueBB); |
| 542 | assert(BBI.FalseBB && "Expected to find the fallthrough block!"); |
| 543 | } |
| 544 | |
| 545 | // Then scan all the instructions. |
| 546 | BBI.NonPredSize = 0; |
| 547 | BBI.ClobbersPred = false; |
| 548 | bool SeenCondBr = false; |
| 549 | for (MachineBasicBlock::iterator I = BBI.BB->begin(), E = BBI.BB->end(); |
| 550 | I != E; ++I) { |
Chris Lattner | 5b93037 | 2008-01-07 07:27:27 +0000 | [diff] [blame^] | 551 | const TargetInstrDesc &TID = I->getDesc(); |
| 552 | if (TID.isNotDuplicable()) |
Dan Gohman | f17a25c | 2007-07-18 16:29:46 +0000 | [diff] [blame] | 553 | BBI.CannotBeCopied = true; |
| 554 | |
| 555 | bool isPredicated = TII->isPredicated(I); |
Chris Lattner | 5b93037 | 2008-01-07 07:27:27 +0000 | [diff] [blame^] | 556 | bool isCondBr = BBI.IsBrAnalyzable && TID.isConditionalBranch(); |
Dan Gohman | f17a25c | 2007-07-18 16:29:46 +0000 | [diff] [blame] | 557 | |
| 558 | if (!isCondBr) { |
| 559 | if (!isPredicated) |
| 560 | BBI.NonPredSize++; |
| 561 | else if (!AlreadyPredicated) { |
| 562 | // FIXME: This instruction is already predicated before the |
| 563 | // if-conversion pass. It's probably something like a conditional move. |
| 564 | // Mark this block unpredicable for now. |
| 565 | BBI.IsUnpredicable = true; |
| 566 | return; |
| 567 | } |
| 568 | |
| 569 | } |
| 570 | |
| 571 | if (BBI.ClobbersPred && !isPredicated) { |
| 572 | // Predicate modification instruction should end the block (except for |
| 573 | // already predicated instructions and end of block branches). |
| 574 | if (isCondBr) { |
| 575 | SeenCondBr = true; |
| 576 | |
| 577 | // Conditional branches is not predicable. But it may be eliminated. |
| 578 | continue; |
| 579 | } |
| 580 | |
| 581 | // Predicate may have been modified, the subsequent (currently) |
| 582 | // unpredicated instructions cannot be correctly predicated. |
| 583 | BBI.IsUnpredicable = true; |
| 584 | return; |
| 585 | } |
| 586 | |
| 587 | // FIXME: Make use of PredDefs? e.g. ADDC, SUBC sets predicates but are |
| 588 | // still potentially predicable. |
| 589 | std::vector<MachineOperand> PredDefs; |
| 590 | if (TII->DefinesPredicate(I, PredDefs)) |
| 591 | BBI.ClobbersPred = true; |
| 592 | |
Chris Lattner | 5b93037 | 2008-01-07 07:27:27 +0000 | [diff] [blame^] | 593 | if (!TID.isPredicable()) { |
Dan Gohman | f17a25c | 2007-07-18 16:29:46 +0000 | [diff] [blame] | 594 | BBI.IsUnpredicable = true; |
| 595 | return; |
| 596 | } |
| 597 | } |
| 598 | } |
| 599 | |
| 600 | /// FeasibilityAnalysis - Determine if the block is a suitable candidate to be |
| 601 | /// predicated by the specified predicate. |
| 602 | bool IfConverter::FeasibilityAnalysis(BBInfo &BBI, |
| 603 | std::vector<MachineOperand> &Pred, |
| 604 | bool isTriangle, bool RevBranch) { |
| 605 | // If the block is dead or unpredicable, then it cannot be predicated. |
| 606 | if (BBI.IsDone || BBI.IsUnpredicable) |
| 607 | return false; |
| 608 | |
| 609 | // If it is already predicated, check if its predicate subsumes the new |
| 610 | // predicate. |
| 611 | if (BBI.Predicate.size() && !TII->SubsumesPredicate(BBI.Predicate, Pred)) |
| 612 | return false; |
| 613 | |
| 614 | if (BBI.BrCond.size()) { |
| 615 | if (!isTriangle) |
| 616 | return false; |
| 617 | |
| 618 | // Test predicate subsumsion. |
| 619 | std::vector<MachineOperand> RevPred(Pred); |
| 620 | std::vector<MachineOperand> Cond(BBI.BrCond); |
| 621 | if (RevBranch) { |
| 622 | if (TII->ReverseBranchCondition(Cond)) |
| 623 | return false; |
| 624 | } |
| 625 | if (TII->ReverseBranchCondition(RevPred) || |
| 626 | !TII->SubsumesPredicate(Cond, RevPred)) |
| 627 | return false; |
| 628 | } |
| 629 | |
| 630 | return true; |
| 631 | } |
| 632 | |
| 633 | /// AnalyzeBlock - Analyze the structure of the sub-CFG starting from |
| 634 | /// the specified block. Record its successors and whether it looks like an |
| 635 | /// if-conversion candidate. |
| 636 | IfConverter::BBInfo &IfConverter::AnalyzeBlock(MachineBasicBlock *BB, |
| 637 | std::vector<IfcvtToken*> &Tokens) { |
| 638 | BBInfo &BBI = BBAnalysis[BB->getNumber()]; |
| 639 | |
| 640 | if (BBI.IsAnalyzed || BBI.IsBeingAnalyzed) |
| 641 | return BBI; |
| 642 | |
| 643 | BBI.BB = BB; |
| 644 | BBI.IsBeingAnalyzed = true; |
| 645 | |
| 646 | ScanInstructions(BBI); |
| 647 | |
| 648 | // Unanalyable or ends with fallthrough or unconditional branch. |
| 649 | if (!BBI.IsBrAnalyzable || BBI.BrCond.size() == 0) { |
| 650 | BBI.IsBeingAnalyzed = false; |
| 651 | BBI.IsAnalyzed = true; |
| 652 | return BBI; |
| 653 | } |
| 654 | |
| 655 | // Do not ifcvt if either path is a back edge to the entry block. |
| 656 | if (BBI.TrueBB == BB || BBI.FalseBB == BB) { |
| 657 | BBI.IsBeingAnalyzed = false; |
| 658 | BBI.IsAnalyzed = true; |
| 659 | return BBI; |
| 660 | } |
| 661 | |
| 662 | BBInfo &TrueBBI = AnalyzeBlock(BBI.TrueBB, Tokens); |
| 663 | BBInfo &FalseBBI = AnalyzeBlock(BBI.FalseBB, Tokens); |
| 664 | |
| 665 | if (TrueBBI.IsDone && FalseBBI.IsDone) { |
| 666 | BBI.IsBeingAnalyzed = false; |
| 667 | BBI.IsAnalyzed = true; |
| 668 | return BBI; |
| 669 | } |
| 670 | |
| 671 | std::vector<MachineOperand> RevCond(BBI.BrCond); |
| 672 | bool CanRevCond = !TII->ReverseBranchCondition(RevCond); |
| 673 | |
| 674 | unsigned Dups = 0; |
| 675 | unsigned Dups2 = 0; |
| 676 | bool TNeedSub = TrueBBI.Predicate.size() > 0; |
| 677 | bool FNeedSub = FalseBBI.Predicate.size() > 0; |
| 678 | bool Enqueued = false; |
| 679 | if (CanRevCond && ValidDiamond(TrueBBI, FalseBBI, Dups, Dups2) && |
| 680 | MeetIfcvtSizeLimit(TrueBBI.NonPredSize - (Dups + Dups2)) && |
| 681 | MeetIfcvtSizeLimit(FalseBBI.NonPredSize - (Dups + Dups2)) && |
| 682 | FeasibilityAnalysis(TrueBBI, BBI.BrCond) && |
| 683 | FeasibilityAnalysis(FalseBBI, RevCond)) { |
| 684 | // Diamond: |
| 685 | // EBB |
| 686 | // / \_ |
| 687 | // | | |
| 688 | // TBB FBB |
| 689 | // \ / |
| 690 | // TailBB |
| 691 | // Note TailBB can be empty. |
| 692 | Tokens.push_back(new IfcvtToken(BBI, ICDiamond, TNeedSub|FNeedSub, Dups, |
| 693 | Dups2)); |
| 694 | Enqueued = true; |
| 695 | } |
| 696 | |
| 697 | if (ValidTriangle(TrueBBI, FalseBBI, false, Dups) && |
| 698 | MeetIfcvtSizeLimit(TrueBBI.NonPredSize) && |
| 699 | FeasibilityAnalysis(TrueBBI, BBI.BrCond, true)) { |
| 700 | // Triangle: |
| 701 | // EBB |
| 702 | // | \_ |
| 703 | // | | |
| 704 | // | TBB |
| 705 | // | / |
| 706 | // FBB |
| 707 | Tokens.push_back(new IfcvtToken(BBI, ICTriangle, TNeedSub, Dups)); |
| 708 | Enqueued = true; |
| 709 | } |
| 710 | |
| 711 | if (ValidTriangle(TrueBBI, FalseBBI, true, Dups) && |
| 712 | MeetIfcvtSizeLimit(TrueBBI.NonPredSize) && |
| 713 | FeasibilityAnalysis(TrueBBI, BBI.BrCond, true, true)) { |
| 714 | Tokens.push_back(new IfcvtToken(BBI, ICTriangleRev, TNeedSub, Dups)); |
| 715 | Enqueued = true; |
| 716 | } |
| 717 | |
| 718 | if (ValidSimple(TrueBBI, Dups) && |
| 719 | MeetIfcvtSizeLimit(TrueBBI.NonPredSize) && |
| 720 | FeasibilityAnalysis(TrueBBI, BBI.BrCond)) { |
| 721 | // Simple (split, no rejoin): |
| 722 | // EBB |
| 723 | // | \_ |
| 724 | // | | |
| 725 | // | TBB---> exit |
| 726 | // | |
| 727 | // FBB |
| 728 | Tokens.push_back(new IfcvtToken(BBI, ICSimple, TNeedSub, Dups)); |
| 729 | Enqueued = true; |
| 730 | } |
| 731 | |
| 732 | if (CanRevCond) { |
| 733 | // Try the other path... |
| 734 | if (ValidTriangle(FalseBBI, TrueBBI, false, Dups) && |
| 735 | MeetIfcvtSizeLimit(FalseBBI.NonPredSize) && |
| 736 | FeasibilityAnalysis(FalseBBI, RevCond, true)) { |
| 737 | Tokens.push_back(new IfcvtToken(BBI, ICTriangleFalse, FNeedSub, Dups)); |
| 738 | Enqueued = true; |
| 739 | } |
| 740 | |
| 741 | if (ValidTriangle(FalseBBI, TrueBBI, true, Dups) && |
| 742 | MeetIfcvtSizeLimit(FalseBBI.NonPredSize) && |
| 743 | FeasibilityAnalysis(FalseBBI, RevCond, true, true)) { |
| 744 | Tokens.push_back(new IfcvtToken(BBI, ICTriangleFRev, FNeedSub, Dups)); |
| 745 | Enqueued = true; |
| 746 | } |
| 747 | |
| 748 | if (ValidSimple(FalseBBI, Dups) && |
| 749 | MeetIfcvtSizeLimit(FalseBBI.NonPredSize) && |
| 750 | FeasibilityAnalysis(FalseBBI, RevCond)) { |
| 751 | Tokens.push_back(new IfcvtToken(BBI, ICSimpleFalse, FNeedSub, Dups)); |
| 752 | Enqueued = true; |
| 753 | } |
| 754 | } |
| 755 | |
| 756 | BBI.IsEnqueued = Enqueued; |
| 757 | BBI.IsBeingAnalyzed = false; |
| 758 | BBI.IsAnalyzed = true; |
| 759 | return BBI; |
| 760 | } |
| 761 | |
| 762 | /// AnalyzeBlocks - Analyze all blocks and find entries for all if-conversion |
| 763 | /// candidates. It returns true if any CFG restructuring is done to expose more |
| 764 | /// if-conversion opportunities. |
| 765 | bool IfConverter::AnalyzeBlocks(MachineFunction &MF, |
| 766 | std::vector<IfcvtToken*> &Tokens) { |
| 767 | bool Change = false; |
| 768 | std::set<MachineBasicBlock*> Visited; |
| 769 | for (unsigned i = 0, e = Roots.size(); i != e; ++i) { |
| 770 | for (idf_ext_iterator<MachineBasicBlock*> I=idf_ext_begin(Roots[i],Visited), |
| 771 | E = idf_ext_end(Roots[i], Visited); I != E; ++I) { |
| 772 | MachineBasicBlock *BB = *I; |
| 773 | AnalyzeBlock(BB, Tokens); |
| 774 | } |
| 775 | } |
| 776 | |
| 777 | // Sort to favor more complex ifcvt scheme. |
| 778 | std::stable_sort(Tokens.begin(), Tokens.end(), IfcvtTokenCmp); |
| 779 | |
| 780 | return Change; |
| 781 | } |
| 782 | |
| 783 | /// canFallThroughTo - Returns true either if ToBB is the next block after BB or |
| 784 | /// that all the intervening blocks are empty (given BB can fall through to its |
| 785 | /// next block). |
| 786 | static bool canFallThroughTo(MachineBasicBlock *BB, MachineBasicBlock *ToBB) { |
| 787 | MachineFunction::iterator I = BB; |
| 788 | MachineFunction::iterator TI = ToBB; |
| 789 | MachineFunction::iterator E = BB->getParent()->end(); |
| 790 | while (++I != TI) |
| 791 | if (I == E || !I->empty()) |
| 792 | return false; |
| 793 | return true; |
| 794 | } |
| 795 | |
| 796 | /// InvalidatePreds - Invalidate predecessor BB info so it would be re-analyzed |
| 797 | /// to determine if it can be if-converted. If predecessor is already enqueued, |
| 798 | /// dequeue it! |
| 799 | void IfConverter::InvalidatePreds(MachineBasicBlock *BB) { |
| 800 | for (MachineBasicBlock::pred_iterator PI = BB->pred_begin(), |
| 801 | E = BB->pred_end(); PI != E; ++PI) { |
| 802 | BBInfo &PBBI = BBAnalysis[(*PI)->getNumber()]; |
| 803 | if (PBBI.IsDone || PBBI.BB == BB) |
| 804 | continue; |
| 805 | PBBI.IsAnalyzed = false; |
| 806 | PBBI.IsEnqueued = false; |
| 807 | } |
| 808 | } |
| 809 | |
| 810 | /// InsertUncondBranch - Inserts an unconditional branch from BB to ToBB. |
| 811 | /// |
| 812 | static void InsertUncondBranch(MachineBasicBlock *BB, MachineBasicBlock *ToBB, |
| 813 | const TargetInstrInfo *TII) { |
| 814 | std::vector<MachineOperand> NoCond; |
| 815 | TII->InsertBranch(*BB, ToBB, NULL, NoCond); |
| 816 | } |
| 817 | |
| 818 | /// RemoveExtraEdges - Remove true / false edges if either / both are no longer |
| 819 | /// successors. |
| 820 | void IfConverter::RemoveExtraEdges(BBInfo &BBI) { |
| 821 | MachineBasicBlock *TBB = NULL, *FBB = NULL; |
| 822 | std::vector<MachineOperand> Cond; |
| 823 | if (!TII->AnalyzeBranch(*BBI.BB, TBB, FBB, Cond)) |
| 824 | BBI.BB->CorrectExtraCFGEdges(TBB, FBB, !Cond.empty()); |
| 825 | } |
| 826 | |
| 827 | /// IfConvertSimple - If convert a simple (split, no rejoin) sub-CFG. |
| 828 | /// |
| 829 | bool IfConverter::IfConvertSimple(BBInfo &BBI, IfcvtKind Kind) { |
| 830 | BBInfo &TrueBBI = BBAnalysis[BBI.TrueBB->getNumber()]; |
| 831 | BBInfo &FalseBBI = BBAnalysis[BBI.FalseBB->getNumber()]; |
| 832 | BBInfo *CvtBBI = &TrueBBI; |
| 833 | BBInfo *NextBBI = &FalseBBI; |
| 834 | |
| 835 | std::vector<MachineOperand> Cond(BBI.BrCond); |
| 836 | if (Kind == ICSimpleFalse) |
| 837 | std::swap(CvtBBI, NextBBI); |
| 838 | |
| 839 | if (CvtBBI->IsDone || |
| 840 | (CvtBBI->CannotBeCopied && CvtBBI->BB->pred_size() > 1)) { |
| 841 | // Something has changed. It's no longer safe to predicate this block. |
| 842 | BBI.IsAnalyzed = false; |
| 843 | CvtBBI->IsAnalyzed = false; |
| 844 | return false; |
| 845 | } |
| 846 | |
| 847 | if (Kind == ICSimpleFalse) |
| 848 | TII->ReverseBranchCondition(Cond); |
| 849 | |
| 850 | if (CvtBBI->BB->pred_size() > 1) { |
| 851 | BBI.NonPredSize -= TII->RemoveBranch(*BBI.BB); |
| 852 | // Copy instructions in the true block, predicate them add them to |
| 853 | // the entry block. |
| 854 | CopyAndPredicateBlock(BBI, *CvtBBI, Cond); |
| 855 | } else { |
| 856 | PredicateBlock(*CvtBBI, CvtBBI->BB->end(), Cond); |
| 857 | |
| 858 | // Merge converted block into entry block. |
| 859 | BBI.NonPredSize -= TII->RemoveBranch(*BBI.BB); |
| 860 | MergeBlocks(BBI, *CvtBBI); |
| 861 | } |
| 862 | |
| 863 | bool IterIfcvt = true; |
| 864 | if (!canFallThroughTo(BBI.BB, NextBBI->BB)) { |
| 865 | InsertUncondBranch(BBI.BB, NextBBI->BB, TII); |
| 866 | BBI.HasFallThrough = false; |
| 867 | // Now ifcvt'd block will look like this: |
| 868 | // BB: |
| 869 | // ... |
| 870 | // t, f = cmp |
| 871 | // if t op |
| 872 | // b BBf |
| 873 | // |
| 874 | // We cannot further ifcvt this block because the unconditional branch |
| 875 | // will have to be predicated on the new condition, that will not be |
| 876 | // available if cmp executes. |
| 877 | IterIfcvt = false; |
| 878 | } |
| 879 | |
| 880 | RemoveExtraEdges(BBI); |
| 881 | |
| 882 | // Update block info. BB can be iteratively if-converted. |
| 883 | if (!IterIfcvt) |
| 884 | BBI.IsDone = true; |
| 885 | InvalidatePreds(BBI.BB); |
| 886 | CvtBBI->IsDone = true; |
| 887 | |
| 888 | // FIXME: Must maintain LiveIns. |
| 889 | return true; |
| 890 | } |
| 891 | |
| 892 | /// IfConvertTriangle - If convert a triangle sub-CFG. |
| 893 | /// |
| 894 | bool IfConverter::IfConvertTriangle(BBInfo &BBI, IfcvtKind Kind) { |
| 895 | BBInfo &TrueBBI = BBAnalysis[BBI.TrueBB->getNumber()]; |
| 896 | BBInfo &FalseBBI = BBAnalysis[BBI.FalseBB->getNumber()]; |
| 897 | BBInfo *CvtBBI = &TrueBBI; |
| 898 | BBInfo *NextBBI = &FalseBBI; |
| 899 | |
| 900 | std::vector<MachineOperand> Cond(BBI.BrCond); |
| 901 | if (Kind == ICTriangleFalse || Kind == ICTriangleFRev) |
| 902 | std::swap(CvtBBI, NextBBI); |
| 903 | |
| 904 | if (CvtBBI->IsDone || |
| 905 | (CvtBBI->CannotBeCopied && CvtBBI->BB->pred_size() > 1)) { |
| 906 | // Something has changed. It's no longer safe to predicate this block. |
| 907 | BBI.IsAnalyzed = false; |
| 908 | CvtBBI->IsAnalyzed = false; |
| 909 | return false; |
| 910 | } |
| 911 | |
| 912 | if (Kind == ICTriangleFalse || Kind == ICTriangleFRev) |
| 913 | TII->ReverseBranchCondition(Cond); |
| 914 | |
| 915 | if (Kind == ICTriangleRev || Kind == ICTriangleFRev) { |
| 916 | ReverseBranchCondition(*CvtBBI); |
| 917 | // BB has been changed, modify its predecessors (except for this |
| 918 | // one) so they don't get ifcvt'ed based on bad intel. |
| 919 | for (MachineBasicBlock::pred_iterator PI = CvtBBI->BB->pred_begin(), |
| 920 | E = CvtBBI->BB->pred_end(); PI != E; ++PI) { |
| 921 | MachineBasicBlock *PBB = *PI; |
| 922 | if (PBB == BBI.BB) |
| 923 | continue; |
| 924 | BBInfo &PBBI = BBAnalysis[PBB->getNumber()]; |
| 925 | if (PBBI.IsEnqueued) { |
| 926 | PBBI.IsAnalyzed = false; |
| 927 | PBBI.IsEnqueued = false; |
| 928 | } |
| 929 | } |
| 930 | } |
| 931 | |
| 932 | bool HasEarlyExit = CvtBBI->FalseBB != NULL; |
| 933 | bool DupBB = CvtBBI->BB->pred_size() > 1; |
| 934 | if (DupBB) { |
| 935 | BBI.NonPredSize -= TII->RemoveBranch(*BBI.BB); |
| 936 | // Copy instructions in the true block, predicate them add them to |
| 937 | // the entry block. |
| 938 | CopyAndPredicateBlock(BBI, *CvtBBI, Cond, true); |
| 939 | } else { |
| 940 | // Predicate the 'true' block after removing its branch. |
| 941 | CvtBBI->NonPredSize -= TII->RemoveBranch(*CvtBBI->BB); |
| 942 | PredicateBlock(*CvtBBI, CvtBBI->BB->end(), Cond); |
| 943 | } |
| 944 | |
| 945 | if (!DupBB) { |
| 946 | // Now merge the entry of the triangle with the true block. |
| 947 | BBI.NonPredSize -= TII->RemoveBranch(*BBI.BB); |
| 948 | MergeBlocks(BBI, *CvtBBI); |
| 949 | } |
| 950 | |
| 951 | // If 'true' block has a 'false' successor, add an exit branch to it. |
| 952 | if (HasEarlyExit) { |
| 953 | std::vector<MachineOperand> RevCond(CvtBBI->BrCond); |
| 954 | if (TII->ReverseBranchCondition(RevCond)) |
| 955 | assert(false && "Unable to reverse branch condition!"); |
| 956 | TII->InsertBranch(*BBI.BB, CvtBBI->FalseBB, NULL, RevCond); |
| 957 | BBI.BB->addSuccessor(CvtBBI->FalseBB); |
| 958 | } |
| 959 | |
| 960 | // Merge in the 'false' block if the 'false' block has no other |
| 961 | // predecessors. Otherwise, add a unconditional branch from to 'false'. |
| 962 | bool FalseBBDead = false; |
| 963 | bool IterIfcvt = true; |
| 964 | bool isFallThrough = canFallThroughTo(BBI.BB, NextBBI->BB); |
| 965 | if (!isFallThrough) { |
| 966 | // Only merge them if the true block does not fallthrough to the false |
| 967 | // block. By not merging them, we make it possible to iteratively |
| 968 | // ifcvt the blocks. |
| 969 | if (!HasEarlyExit && |
| 970 | NextBBI->BB->pred_size() == 1 && !NextBBI->HasFallThrough) { |
| 971 | MergeBlocks(BBI, *NextBBI); |
| 972 | FalseBBDead = true; |
| 973 | } else { |
| 974 | InsertUncondBranch(BBI.BB, NextBBI->BB, TII); |
| 975 | BBI.HasFallThrough = false; |
| 976 | } |
| 977 | // Mixed predicated and unpredicated code. This cannot be iteratively |
| 978 | // predicated. |
| 979 | IterIfcvt = false; |
| 980 | } |
| 981 | |
| 982 | RemoveExtraEdges(BBI); |
| 983 | |
| 984 | // Update block info. BB can be iteratively if-converted. |
| 985 | if (!IterIfcvt) |
| 986 | BBI.IsDone = true; |
| 987 | InvalidatePreds(BBI.BB); |
| 988 | CvtBBI->IsDone = true; |
| 989 | if (FalseBBDead) |
| 990 | NextBBI->IsDone = true; |
| 991 | |
| 992 | // FIXME: Must maintain LiveIns. |
| 993 | return true; |
| 994 | } |
| 995 | |
| 996 | /// IfConvertDiamond - If convert a diamond sub-CFG. |
| 997 | /// |
| 998 | bool IfConverter::IfConvertDiamond(BBInfo &BBI, IfcvtKind Kind, |
| 999 | unsigned NumDups1, unsigned NumDups2) { |
| 1000 | BBInfo &TrueBBI = BBAnalysis[BBI.TrueBB->getNumber()]; |
| 1001 | BBInfo &FalseBBI = BBAnalysis[BBI.FalseBB->getNumber()]; |
| 1002 | MachineBasicBlock *TailBB = TrueBBI.TrueBB; |
| 1003 | // True block must fall through or ended with unanalyzable terminator. |
| 1004 | if (!TailBB) { |
| 1005 | if (blockAlwaysFallThrough(TrueBBI)) |
| 1006 | TailBB = FalseBBI.TrueBB; |
| 1007 | assert((TailBB || !TrueBBI.IsBrAnalyzable) && "Unexpected!"); |
| 1008 | } |
| 1009 | |
| 1010 | if (TrueBBI.IsDone || FalseBBI.IsDone || |
| 1011 | TrueBBI.BB->pred_size() > 1 || |
| 1012 | FalseBBI.BB->pred_size() > 1) { |
| 1013 | // Something has changed. It's no longer safe to predicate these blocks. |
| 1014 | BBI.IsAnalyzed = false; |
| 1015 | TrueBBI.IsAnalyzed = false; |
| 1016 | FalseBBI.IsAnalyzed = false; |
| 1017 | return false; |
| 1018 | } |
| 1019 | |
| 1020 | // Merge the 'true' and 'false' blocks by copying the instructions |
| 1021 | // from the 'false' block to the 'true' block. That is, unless the true |
| 1022 | // block would clobber the predicate, in that case, do the opposite. |
| 1023 | BBInfo *BBI1 = &TrueBBI; |
| 1024 | BBInfo *BBI2 = &FalseBBI; |
| 1025 | std::vector<MachineOperand> RevCond(BBI.BrCond); |
| 1026 | TII->ReverseBranchCondition(RevCond); |
| 1027 | std::vector<MachineOperand> *Cond1 = &BBI.BrCond; |
| 1028 | std::vector<MachineOperand> *Cond2 = &RevCond; |
| 1029 | |
| 1030 | // Figure out the more profitable ordering. |
| 1031 | bool DoSwap = false; |
| 1032 | if (TrueBBI.ClobbersPred && !FalseBBI.ClobbersPred) |
| 1033 | DoSwap = true; |
| 1034 | else if (TrueBBI.ClobbersPred == FalseBBI.ClobbersPred) { |
| 1035 | if (TrueBBI.NonPredSize > FalseBBI.NonPredSize) |
| 1036 | DoSwap = true; |
| 1037 | } |
| 1038 | if (DoSwap) { |
| 1039 | std::swap(BBI1, BBI2); |
| 1040 | std::swap(Cond1, Cond2); |
| 1041 | } |
| 1042 | |
| 1043 | // Remove the conditional branch from entry to the blocks. |
| 1044 | BBI.NonPredSize -= TII->RemoveBranch(*BBI.BB); |
| 1045 | |
| 1046 | // Remove the duplicated instructions at the beginnings of both paths. |
| 1047 | MachineBasicBlock::iterator DI1 = BBI1->BB->begin(); |
| 1048 | MachineBasicBlock::iterator DI2 = BBI2->BB->begin(); |
| 1049 | BBI1->NonPredSize -= NumDups1; |
| 1050 | BBI2->NonPredSize -= NumDups1; |
| 1051 | while (NumDups1 != 0) { |
| 1052 | ++DI1; |
| 1053 | ++DI2; |
| 1054 | --NumDups1; |
| 1055 | } |
| 1056 | BBI.BB->splice(BBI.BB->end(), BBI1->BB, BBI1->BB->begin(), DI1); |
| 1057 | BBI2->BB->erase(BBI2->BB->begin(), DI2); |
| 1058 | |
| 1059 | // Predicate the 'true' block after removing its branch. |
| 1060 | BBI1->NonPredSize -= TII->RemoveBranch(*BBI1->BB); |
| 1061 | DI1 = BBI1->BB->end(); |
| 1062 | for (unsigned i = 0; i != NumDups2; ++i) |
| 1063 | --DI1; |
| 1064 | BBI1->BB->erase(DI1, BBI1->BB->end()); |
| 1065 | PredicateBlock(*BBI1, BBI1->BB->end(), *Cond1); |
| 1066 | |
| 1067 | // Predicate the 'false' block. |
| 1068 | BBI2->NonPredSize -= TII->RemoveBranch(*BBI2->BB); |
| 1069 | DI2 = BBI2->BB->end(); |
| 1070 | while (NumDups2 != 0) { |
| 1071 | --DI2; |
| 1072 | --NumDups2; |
| 1073 | } |
| 1074 | PredicateBlock(*BBI2, DI2, *Cond2); |
| 1075 | |
| 1076 | // Merge the true block into the entry of the diamond. |
| 1077 | MergeBlocks(BBI, *BBI1); |
| 1078 | MergeBlocks(BBI, *BBI2); |
| 1079 | |
| 1080 | // If the if-converted block fallthrough or unconditionally branch into the |
| 1081 | // tail block, and the tail block does not have other predecessors, then |
| 1082 | // fold the tail block in as well. Otherwise, unless it falls through to the |
| 1083 | // tail, add a unconditional branch to it. |
| 1084 | if (TailBB) { |
| 1085 | BBInfo TailBBI = BBAnalysis[TailBB->getNumber()]; |
| 1086 | if (TailBB->pred_size() == 1 && !TailBBI.HasFallThrough) { |
| 1087 | BBI.NonPredSize -= TII->RemoveBranch(*BBI.BB); |
| 1088 | MergeBlocks(BBI, TailBBI); |
| 1089 | TailBBI.IsDone = true; |
| 1090 | } else { |
| 1091 | InsertUncondBranch(BBI.BB, TailBB, TII); |
| 1092 | BBI.HasFallThrough = false; |
| 1093 | } |
| 1094 | } |
| 1095 | |
| 1096 | RemoveExtraEdges(BBI); |
| 1097 | |
| 1098 | // Update block info. |
| 1099 | BBI.IsDone = TrueBBI.IsDone = FalseBBI.IsDone = true; |
| 1100 | InvalidatePreds(BBI.BB); |
| 1101 | |
| 1102 | // FIXME: Must maintain LiveIns. |
| 1103 | return true; |
| 1104 | } |
| 1105 | |
| 1106 | /// PredicateBlock - Predicate instructions from the start of the block to the |
| 1107 | /// specified end with the specified condition. |
| 1108 | void IfConverter::PredicateBlock(BBInfo &BBI, |
| 1109 | MachineBasicBlock::iterator E, |
| 1110 | std::vector<MachineOperand> &Cond) { |
| 1111 | for (MachineBasicBlock::iterator I = BBI.BB->begin(); I != E; ++I) { |
| 1112 | if (TII->isPredicated(I)) |
| 1113 | continue; |
| 1114 | if (!TII->PredicateInstruction(I, Cond)) { |
| 1115 | cerr << "Unable to predicate " << *I << "!\n"; |
| 1116 | abort(); |
| 1117 | } |
| 1118 | } |
| 1119 | |
| 1120 | std::copy(Cond.begin(), Cond.end(), std::back_inserter(BBI.Predicate)); |
| 1121 | |
| 1122 | BBI.IsAnalyzed = false; |
| 1123 | BBI.NonPredSize = 0; |
| 1124 | |
| 1125 | NumIfConvBBs++; |
| 1126 | } |
| 1127 | |
| 1128 | /// CopyAndPredicateBlock - Copy and predicate instructions from source BB to |
| 1129 | /// the destination block. Skip end of block branches if IgnoreBr is true. |
| 1130 | void IfConverter::CopyAndPredicateBlock(BBInfo &ToBBI, BBInfo &FromBBI, |
| 1131 | std::vector<MachineOperand> &Cond, |
| 1132 | bool IgnoreBr) { |
| 1133 | for (MachineBasicBlock::iterator I = FromBBI.BB->begin(), |
| 1134 | E = FromBBI.BB->end(); I != E; ++I) { |
Chris Lattner | 5b93037 | 2008-01-07 07:27:27 +0000 | [diff] [blame^] | 1135 | const TargetInstrDesc &TID = I->getDesc(); |
Dan Gohman | f17a25c | 2007-07-18 16:29:46 +0000 | [diff] [blame] | 1136 | bool isPredicated = TII->isPredicated(I); |
| 1137 | // Do not copy the end of the block branches. |
Chris Lattner | 5b93037 | 2008-01-07 07:27:27 +0000 | [diff] [blame^] | 1138 | if (IgnoreBr && !isPredicated && TID.isBranch()) |
Dan Gohman | f17a25c | 2007-07-18 16:29:46 +0000 | [diff] [blame] | 1139 | break; |
| 1140 | |
| 1141 | MachineInstr *MI = I->clone(); |
| 1142 | ToBBI.BB->insert(ToBBI.BB->end(), MI); |
| 1143 | ToBBI.NonPredSize++; |
| 1144 | |
| 1145 | if (!isPredicated) |
| 1146 | if (!TII->PredicateInstruction(MI, Cond)) { |
| 1147 | cerr << "Unable to predicate " << *MI << "!\n"; |
| 1148 | abort(); |
| 1149 | } |
| 1150 | } |
| 1151 | |
| 1152 | std::vector<MachineBasicBlock *> Succs(FromBBI.BB->succ_begin(), |
| 1153 | FromBBI.BB->succ_end()); |
| 1154 | MachineBasicBlock *NBB = getNextBlock(FromBBI.BB); |
| 1155 | MachineBasicBlock *FallThrough = FromBBI.HasFallThrough ? NBB : NULL; |
| 1156 | |
| 1157 | for (unsigned i = 0, e = Succs.size(); i != e; ++i) { |
| 1158 | MachineBasicBlock *Succ = Succs[i]; |
| 1159 | // Fallthrough edge can't be transferred. |
| 1160 | if (Succ == FallThrough) |
| 1161 | continue; |
| 1162 | if (!ToBBI.BB->isSuccessor(Succ)) |
| 1163 | ToBBI.BB->addSuccessor(Succ); |
| 1164 | } |
| 1165 | |
| 1166 | std::copy(FromBBI.Predicate.begin(), FromBBI.Predicate.end(), |
| 1167 | std::back_inserter(ToBBI.Predicate)); |
| 1168 | std::copy(Cond.begin(), Cond.end(), std::back_inserter(ToBBI.Predicate)); |
| 1169 | |
| 1170 | ToBBI.ClobbersPred |= FromBBI.ClobbersPred; |
| 1171 | ToBBI.IsAnalyzed = false; |
| 1172 | |
| 1173 | NumDupBBs++; |
| 1174 | } |
| 1175 | |
| 1176 | /// MergeBlocks - Move all instructions from FromBB to the end of ToBB. |
| 1177 | /// |
| 1178 | void IfConverter::MergeBlocks(BBInfo &ToBBI, BBInfo &FromBBI) { |
| 1179 | ToBBI.BB->splice(ToBBI.BB->end(), |
| 1180 | FromBBI.BB, FromBBI.BB->begin(), FromBBI.BB->end()); |
| 1181 | |
| 1182 | // Redirect all branches to FromBB to ToBB. |
| 1183 | std::vector<MachineBasicBlock *> Preds(FromBBI.BB->pred_begin(), |
| 1184 | FromBBI.BB->pred_end()); |
| 1185 | for (unsigned i = 0, e = Preds.size(); i != e; ++i) { |
| 1186 | MachineBasicBlock *Pred = Preds[i]; |
| 1187 | if (Pred == ToBBI.BB) |
| 1188 | continue; |
| 1189 | Pred->ReplaceUsesOfBlockWith(FromBBI.BB, ToBBI.BB); |
| 1190 | } |
| 1191 | |
| 1192 | std::vector<MachineBasicBlock *> Succs(FromBBI.BB->succ_begin(), |
| 1193 | FromBBI.BB->succ_end()); |
| 1194 | MachineBasicBlock *NBB = getNextBlock(FromBBI.BB); |
| 1195 | MachineBasicBlock *FallThrough = FromBBI.HasFallThrough ? NBB : NULL; |
| 1196 | |
| 1197 | for (unsigned i = 0, e = Succs.size(); i != e; ++i) { |
| 1198 | MachineBasicBlock *Succ = Succs[i]; |
| 1199 | // Fallthrough edge can't be transferred. |
| 1200 | if (Succ == FallThrough) |
| 1201 | continue; |
| 1202 | FromBBI.BB->removeSuccessor(Succ); |
| 1203 | if (!ToBBI.BB->isSuccessor(Succ)) |
| 1204 | ToBBI.BB->addSuccessor(Succ); |
| 1205 | } |
| 1206 | |
| 1207 | // Now FromBBI always fall through to the next block! |
| 1208 | if (NBB && !FromBBI.BB->isSuccessor(NBB)) |
| 1209 | FromBBI.BB->addSuccessor(NBB); |
| 1210 | |
| 1211 | std::copy(FromBBI.Predicate.begin(), FromBBI.Predicate.end(), |
| 1212 | std::back_inserter(ToBBI.Predicate)); |
| 1213 | FromBBI.Predicate.clear(); |
| 1214 | |
| 1215 | ToBBI.NonPredSize += FromBBI.NonPredSize; |
| 1216 | FromBBI.NonPredSize = 0; |
| 1217 | |
| 1218 | ToBBI.ClobbersPred |= FromBBI.ClobbersPred; |
| 1219 | ToBBI.HasFallThrough = FromBBI.HasFallThrough; |
| 1220 | ToBBI.IsAnalyzed = false; |
| 1221 | FromBBI.IsAnalyzed = false; |
| 1222 | } |