1 //===- LoopStrengthReduce.cpp - Strength Reduce IVs in Loops --------------===// 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 transformation analyzes and transforms the induction variables (and 11 // computations derived from them) into forms suitable for efficient execution 12 // on the target. 13 // 14 // This pass performs a strength reduction on array references inside loops that 15 // have as one or more of their components the loop induction variable, it 16 // rewrites expressions to take advantage of scaled-index addressing modes 17 // available on the target, and it performs a variety of other optimizations 18 // related to loop induction variables. 19 // 20 // Terminology note: this code has a lot of handling for "post-increment" or 21 // "post-inc" users. This is not talking about post-increment addressing modes; 22 // it is instead talking about code like this: 23 // 24 // %i = phi [ 0, %entry ], [ %i.next, %latch ] 25 // ... 26 // %i.next = add %i, 1 27 // %c = icmp eq %i.next, %n 28 // 29 // The SCEV for %i is {0,+,1}<%L>. The SCEV for %i.next is {1,+,1}<%L>, however 30 // it's useful to think about these as the same register, with some uses using 31 // the value of the register before the add and some using // it after. In this 32 // example, the icmp is a post-increment user, since it uses %i.next, which is 33 // the value of the induction variable after the increment. The other common 34 // case of post-increment users is users outside the loop. 35 // 36 // TODO: More sophistication in the way Formulae are generated and filtered. 37 // 38 // TODO: Handle multiple loops at a time. 39 // 40 // TODO: Should the addressing mode BaseGV be changed to a ConstantExpr instead 41 // of a GlobalValue? 42 // 43 // TODO: When truncation is free, truncate ICmp users' operands to make it a 44 // smaller encoding (on x86 at least). 45 // 46 // TODO: When a negated register is used by an add (such as in a list of 47 // multiple base registers, or as the increment expression in an addrec), 48 // we may not actually need both reg and (-1 * reg) in registers; the 49 // negation can be implemented by using a sub instead of an add. The 50 // lack of support for taking this into consideration when making 51 // register pressure decisions is partly worked around by the "Special" 52 // use kind. 53 // 54 //===----------------------------------------------------------------------===// 55 56 #define DEBUG_TYPE "loop-reduce" 57 #include "llvm/Transforms/Scalar.h" 58 #include "llvm/ADT/DenseSet.h" 59 #include "llvm/ADT/SetVector.h" 60 #include "llvm/ADT/SmallBitVector.h" 61 #include "llvm/ADT/STLExtras.h" 62 #include "llvm/Analysis/Dominators.h" 63 #include "llvm/Analysis/IVUsers.h" 64 #include "llvm/Analysis/LoopPass.h" 65 #include "llvm/Analysis/ScalarEvolutionExpander.h" 66 #include "llvm/Analysis/TargetTransformInfo.h" 67 #include "llvm/Assembly/Writer.h" 68 #include "llvm/IR/Constants.h" 69 #include "llvm/IR/DerivedTypes.h" 70 #include "llvm/IR/Instructions.h" 71 #include "llvm/IR/IntrinsicInst.h" 72 #include "llvm/Support/CommandLine.h" 73 #include "llvm/Support/Debug.h" 74 #include "llvm/Support/ValueHandle.h" 75 #include "llvm/Support/raw_ostream.h" 76 #include "llvm/Transforms/Utils/BasicBlockUtils.h" 77 #include "llvm/Transforms/Utils/Local.h" 78 #include <algorithm> 79 using namespace llvm; 80 81 /// MaxIVUsers is an arbitrary threshold that provides an early opportunitiy for 82 /// bail out. This threshold is far beyond the number of users that LSR can 83 /// conceivably solve, so it should not affect generated code, but catches the 84 /// worst cases before LSR burns too much compile time and stack space. 85 static const unsigned MaxIVUsers = 200; 86 87 // Temporary flag to cleanup congruent phis after LSR phi expansion. 88 // It's currently disabled until we can determine whether it's truly useful or 89 // not. The flag should be removed after the v3.0 release. 90 // This is now needed for ivchains. 91 static cl::opt<bool> EnablePhiElim( 92 "enable-lsr-phielim", cl::Hidden, cl::init(true), 93 cl::desc("Enable LSR phi elimination")); 94 95 #ifndef NDEBUG 96 // Stress test IV chain generation. 97 static cl::opt<bool> StressIVChain( 98 "stress-ivchain", cl::Hidden, cl::init(false), 99 cl::desc("Stress test LSR IV chains")); 100 #else 101 static bool StressIVChain = false; 102 #endif 103 104 namespace { 105 106 /// RegSortData - This class holds data which is used to order reuse candidates. 107 class RegSortData { 108 public: 109 /// UsedByIndices - This represents the set of LSRUse indices which reference 110 /// a particular register. 111 SmallBitVector UsedByIndices; 112 113 RegSortData() {} 114 115 void print(raw_ostream &OS) const; 116 void dump() const; 117 }; 118 119 } 120 121 void RegSortData::print(raw_ostream &OS) const { 122 OS << "[NumUses=" << UsedByIndices.count() << ']'; 123 } 124 125 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) 126 void RegSortData::dump() const { 127 print(errs()); errs() << '\n'; 128 } 129 #endif 130 131 namespace { 132 133 /// RegUseTracker - Map register candidates to information about how they are 134 /// used. 135 class RegUseTracker { 136 typedef DenseMap<const SCEV *, RegSortData> RegUsesTy; 137 138 RegUsesTy RegUsesMap; 139 SmallVector<const SCEV *, 16> RegSequence; 140 141 public: 142 void CountRegister(const SCEV *Reg, size_t LUIdx); 143 void DropRegister(const SCEV *Reg, size_t LUIdx); 144 void SwapAndDropUse(size_t LUIdx, size_t LastLUIdx); 145 146 bool isRegUsedByUsesOtherThan(const SCEV *Reg, size_t LUIdx) const; 147 148 const SmallBitVector &getUsedByIndices(const SCEV *Reg) const; 149 150 void clear(); 151 152 typedef SmallVectorImpl<const SCEV *>::iterator iterator; 153 typedef SmallVectorImpl<const SCEV *>::const_iterator const_iterator; 154 iterator begin() { return RegSequence.begin(); } 155 iterator end() { return RegSequence.end(); } 156 const_iterator begin() const { return RegSequence.begin(); } 157 const_iterator end() const { return RegSequence.end(); } 158 }; 159 160 } 161 162 void 163 RegUseTracker::CountRegister(const SCEV *Reg, size_t LUIdx) { 164 std::pair<RegUsesTy::iterator, bool> Pair = 165 RegUsesMap.insert(std::make_pair(Reg, RegSortData())); 166 RegSortData &RSD = Pair.first->second; 167 if (Pair.second) 168 RegSequence.push_back(Reg); 169 RSD.UsedByIndices.resize(std::max(RSD.UsedByIndices.size(), LUIdx + 1)); 170 RSD.UsedByIndices.set(LUIdx); 171 } 172 173 void 174 RegUseTracker::DropRegister(const SCEV *Reg, size_t LUIdx) { 175 RegUsesTy::iterator It = RegUsesMap.find(Reg); 176 assert(It != RegUsesMap.end()); 177 RegSortData &RSD = It->second; 178 assert(RSD.UsedByIndices.size() > LUIdx); 179 RSD.UsedByIndices.reset(LUIdx); 180 } 181 182 void 183 RegUseTracker::SwapAndDropUse(size_t LUIdx, size_t LastLUIdx) { 184 assert(LUIdx <= LastLUIdx); 185 186 // Update RegUses. The data structure is not optimized for this purpose; 187 // we must iterate through it and update each of the bit vectors. 188 for (RegUsesTy::iterator I = RegUsesMap.begin(), E = RegUsesMap.end(); 189 I != E; ++I) { 190 SmallBitVector &UsedByIndices = I->second.UsedByIndices; 191 if (LUIdx < UsedByIndices.size()) 192 UsedByIndices[LUIdx] = 193 LastLUIdx < UsedByIndices.size() ? UsedByIndices[LastLUIdx] : 0; 194 UsedByIndices.resize(std::min(UsedByIndices.size(), LastLUIdx)); 195 } 196 } 197 198 bool 199 RegUseTracker::isRegUsedByUsesOtherThan(const SCEV *Reg, size_t LUIdx) const { 200 RegUsesTy::const_iterator I = RegUsesMap.find(Reg); 201 if (I == RegUsesMap.end()) 202 return false; 203 const SmallBitVector &UsedByIndices = I->second.UsedByIndices; 204 int i = UsedByIndices.find_first(); 205 if (i == -1) return false; 206 if ((size_t)i != LUIdx) return true; 207 return UsedByIndices.find_next(i) != -1; 208 } 209 210 const SmallBitVector &RegUseTracker::getUsedByIndices(const SCEV *Reg) const { 211 RegUsesTy::const_iterator I = RegUsesMap.find(Reg); 212 assert(I != RegUsesMap.end() && "Unknown register!"); 213 return I->second.UsedByIndices; 214 } 215 216 void RegUseTracker::clear() { 217 RegUsesMap.clear(); 218 RegSequence.clear(); 219 } 220 221 namespace { 222 223 /// Formula - This class holds information that describes a formula for 224 /// computing satisfying a use. It may include broken-out immediates and scaled 225 /// registers. 226 struct Formula { 227 /// Global base address used for complex addressing. 228 GlobalValue *BaseGV; 229 230 /// Base offset for complex addressing. 231 int64_t BaseOffset; 232 233 /// Whether any complex addressing has a base register. 234 bool HasBaseReg; 235 236 /// The scale of any complex addressing. 237 int64_t Scale; 238 239 /// BaseRegs - The list of "base" registers for this use. When this is 240 /// non-empty, 241 SmallVector<const SCEV *, 4> BaseRegs; 242 243 /// ScaledReg - The 'scaled' register for this use. This should be non-null 244 /// when Scale is not zero. 245 const SCEV *ScaledReg; 246 247 /// UnfoldedOffset - An additional constant offset which added near the 248 /// use. This requires a temporary register, but the offset itself can 249 /// live in an add immediate field rather than a register. 250 int64_t UnfoldedOffset; 251 252 Formula() 253 : BaseGV(0), BaseOffset(0), HasBaseReg(false), Scale(0), ScaledReg(0), 254 UnfoldedOffset(0) {} 255 256 void InitialMatch(const SCEV *S, Loop *L, ScalarEvolution &SE); 257 258 unsigned getNumRegs() const; 259 Type *getType() const; 260 261 void DeleteBaseReg(const SCEV *&S); 262 263 bool referencesReg(const SCEV *S) const; 264 bool hasRegsUsedByUsesOtherThan(size_t LUIdx, 265 const RegUseTracker &RegUses) const; 266 267 void print(raw_ostream &OS) const; 268 void dump() const; 269 }; 270 271 } 272 273 /// DoInitialMatch - Recursion helper for InitialMatch. 274 static void DoInitialMatch(const SCEV *S, Loop *L, 275 SmallVectorImpl<const SCEV *> &Good, 276 SmallVectorImpl<const SCEV *> &Bad, 277 ScalarEvolution &SE) { 278 // Collect expressions which properly dominate the loop header. 279 if (SE.properlyDominates(S, L->getHeader())) { 280 Good.push_back(S); 281 return; 282 } 283 284 // Look at add operands. 285 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) { 286 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end(); 287 I != E; ++I) 288 DoInitialMatch(*I, L, Good, Bad, SE); 289 return; 290 } 291 292 // Look at addrec operands. 293 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) 294 if (!AR->getStart()->isZero()) { 295 DoInitialMatch(AR->getStart(), L, Good, Bad, SE); 296 DoInitialMatch(SE.getAddRecExpr(SE.getConstant(AR->getType(), 0), 297 AR->getStepRecurrence(SE), 298 // FIXME: AR->getNoWrapFlags() 299 AR->getLoop(), SCEV::FlagAnyWrap), 300 L, Good, Bad, SE); 301 return; 302 } 303 304 // Handle a multiplication by -1 (negation) if it didn't fold. 305 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) 306 if (Mul->getOperand(0)->isAllOnesValue()) { 307 SmallVector<const SCEV *, 4> Ops(Mul->op_begin()+1, Mul->op_end()); 308 const SCEV *NewMul = SE.getMulExpr(Ops); 309 310 SmallVector<const SCEV *, 4> MyGood; 311 SmallVector<const SCEV *, 4> MyBad; 312 DoInitialMatch(NewMul, L, MyGood, MyBad, SE); 313 const SCEV *NegOne = SE.getSCEV(ConstantInt::getAllOnesValue( 314 SE.getEffectiveSCEVType(NewMul->getType()))); 315 for (SmallVectorImpl<const SCEV *>::const_iterator I = MyGood.begin(), 316 E = MyGood.end(); I != E; ++I) 317 Good.push_back(SE.getMulExpr(NegOne, *I)); 318 for (SmallVectorImpl<const SCEV *>::const_iterator I = MyBad.begin(), 319 E = MyBad.end(); I != E; ++I) 320 Bad.push_back(SE.getMulExpr(NegOne, *I)); 321 return; 322 } 323 324 // Ok, we can't do anything interesting. Just stuff the whole thing into a 325 // register and hope for the best. 326 Bad.push_back(S); 327 } 328 329 /// InitialMatch - Incorporate loop-variant parts of S into this Formula, 330 /// attempting to keep all loop-invariant and loop-computable values in a 331 /// single base register. 332 void Formula::InitialMatch(const SCEV *S, Loop *L, ScalarEvolution &SE) { 333 SmallVector<const SCEV *, 4> Good; 334 SmallVector<const SCEV *, 4> Bad; 335 DoInitialMatch(S, L, Good, Bad, SE); 336 if (!Good.empty()) { 337 const SCEV *Sum = SE.getAddExpr(Good); 338 if (!Sum->isZero()) 339 BaseRegs.push_back(Sum); 340 HasBaseReg = true; 341 } 342 if (!Bad.empty()) { 343 const SCEV *Sum = SE.getAddExpr(Bad); 344 if (!Sum->isZero()) 345 BaseRegs.push_back(Sum); 346 HasBaseReg = true; 347 } 348 } 349 350 /// getNumRegs - Return the total number of register operands used by this 351 /// formula. This does not include register uses implied by non-constant 352 /// addrec strides. 353 unsigned Formula::getNumRegs() const { 354 return !!ScaledReg + BaseRegs.size(); 355 } 356 357 /// getType - Return the type of this formula, if it has one, or null 358 /// otherwise. This type is meaningless except for the bit size. 359 Type *Formula::getType() const { 360 return !BaseRegs.empty() ? BaseRegs.front()->getType() : 361 ScaledReg ? ScaledReg->getType() : 362 BaseGV ? BaseGV->getType() : 363 0; 364 } 365 366 /// DeleteBaseReg - Delete the given base reg from the BaseRegs list. 367 void Formula::DeleteBaseReg(const SCEV *&S) { 368 if (&S != &BaseRegs.back()) 369 std::swap(S, BaseRegs.back()); 370 BaseRegs.pop_back(); 371 } 372 373 /// referencesReg - Test if this formula references the given register. 374 bool Formula::referencesReg(const SCEV *S) const { 375 return S == ScaledReg || 376 std::find(BaseRegs.begin(), BaseRegs.end(), S) != BaseRegs.end(); 377 } 378 379 /// hasRegsUsedByUsesOtherThan - Test whether this formula uses registers 380 /// which are used by uses other than the use with the given index. 381 bool Formula::hasRegsUsedByUsesOtherThan(size_t LUIdx, 382 const RegUseTracker &RegUses) const { 383 if (ScaledReg) 384 if (RegUses.isRegUsedByUsesOtherThan(ScaledReg, LUIdx)) 385 return true; 386 for (SmallVectorImpl<const SCEV *>::const_iterator I = BaseRegs.begin(), 387 E = BaseRegs.end(); I != E; ++I) 388 if (RegUses.isRegUsedByUsesOtherThan(*I, LUIdx)) 389 return true; 390 return false; 391 } 392 393 void Formula::print(raw_ostream &OS) const { 394 bool First = true; 395 if (BaseGV) { 396 if (!First) OS << " + "; else First = false; 397 WriteAsOperand(OS, BaseGV, /*PrintType=*/false); 398 } 399 if (BaseOffset != 0) { 400 if (!First) OS << " + "; else First = false; 401 OS << BaseOffset; 402 } 403 for (SmallVectorImpl<const SCEV *>::const_iterator I = BaseRegs.begin(), 404 E = BaseRegs.end(); I != E; ++I) { 405 if (!First) OS << " + "; else First = false; 406 OS << "reg(" << **I << ')'; 407 } 408 if (HasBaseReg && BaseRegs.empty()) { 409 if (!First) OS << " + "; else First = false; 410 OS << "**error: HasBaseReg**"; 411 } else if (!HasBaseReg && !BaseRegs.empty()) { 412 if (!First) OS << " + "; else First = false; 413 OS << "**error: !HasBaseReg**"; 414 } 415 if (Scale != 0) { 416 if (!First) OS << " + "; else First = false; 417 OS << Scale << "*reg("; 418 if (ScaledReg) 419 OS << *ScaledReg; 420 else 421 OS << "<unknown>"; 422 OS << ')'; 423 } 424 if (UnfoldedOffset != 0) { 425 if (!First) OS << " + "; else First = false; 426 OS << "imm(" << UnfoldedOffset << ')'; 427 } 428 } 429 430 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) 431 void Formula::dump() const { 432 print(errs()); errs() << '\n'; 433 } 434 #endif 435 436 /// isAddRecSExtable - Return true if the given addrec can be sign-extended 437 /// without changing its value. 438 static bool isAddRecSExtable(const SCEVAddRecExpr *AR, ScalarEvolution &SE) { 439 Type *WideTy = 440 IntegerType::get(SE.getContext(), SE.getTypeSizeInBits(AR->getType()) + 1); 441 return isa<SCEVAddRecExpr>(SE.getSignExtendExpr(AR, WideTy)); 442 } 443 444 /// isAddSExtable - Return true if the given add can be sign-extended 445 /// without changing its value. 446 static bool isAddSExtable(const SCEVAddExpr *A, ScalarEvolution &SE) { 447 Type *WideTy = 448 IntegerType::get(SE.getContext(), SE.getTypeSizeInBits(A->getType()) + 1); 449 return isa<SCEVAddExpr>(SE.getSignExtendExpr(A, WideTy)); 450 } 451 452 /// isMulSExtable - Return true if the given mul can be sign-extended 453 /// without changing its value. 454 static bool isMulSExtable(const SCEVMulExpr *M, ScalarEvolution &SE) { 455 Type *WideTy = 456 IntegerType::get(SE.getContext(), 457 SE.getTypeSizeInBits(M->getType()) * M->getNumOperands()); 458 return isa<SCEVMulExpr>(SE.getSignExtendExpr(M, WideTy)); 459 } 460 461 /// getExactSDiv - Return an expression for LHS /s RHS, if it can be determined 462 /// and if the remainder is known to be zero, or null otherwise. If 463 /// IgnoreSignificantBits is true, expressions like (X * Y) /s Y are simplified 464 /// to Y, ignoring that the multiplication may overflow, which is useful when 465 /// the result will be used in a context where the most significant bits are 466 /// ignored. 467 static const SCEV *getExactSDiv(const SCEV *LHS, const SCEV *RHS, 468 ScalarEvolution &SE, 469 bool IgnoreSignificantBits = false) { 470 // Handle the trivial case, which works for any SCEV type. 471 if (LHS == RHS) 472 return SE.getConstant(LHS->getType(), 1); 473 474 // Handle a few RHS special cases. 475 const SCEVConstant *RC = dyn_cast<SCEVConstant>(RHS); 476 if (RC) { 477 const APInt &RA = RC->getValue()->getValue(); 478 // Handle x /s -1 as x * -1, to give ScalarEvolution a chance to do 479 // some folding. 480 if (RA.isAllOnesValue()) 481 return SE.getMulExpr(LHS, RC); 482 // Handle x /s 1 as x. 483 if (RA == 1) 484 return LHS; 485 } 486 487 // Check for a division of a constant by a constant. 488 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(LHS)) { 489 if (!RC) 490 return 0; 491 const APInt &LA = C->getValue()->getValue(); 492 const APInt &RA = RC->getValue()->getValue(); 493 if (LA.srem(RA) != 0) 494 return 0; 495 return SE.getConstant(LA.sdiv(RA)); 496 } 497 498 // Distribute the sdiv over addrec operands, if the addrec doesn't overflow. 499 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(LHS)) { 500 if (IgnoreSignificantBits || isAddRecSExtable(AR, SE)) { 501 const SCEV *Step = getExactSDiv(AR->getStepRecurrence(SE), RHS, SE, 502 IgnoreSignificantBits); 503 if (!Step) return 0; 504 const SCEV *Start = getExactSDiv(AR->getStart(), RHS, SE, 505 IgnoreSignificantBits); 506 if (!Start) return 0; 507 // FlagNW is independent of the start value, step direction, and is 508 // preserved with smaller magnitude steps. 509 // FIXME: AR->getNoWrapFlags(SCEV::FlagNW) 510 return SE.getAddRecExpr(Start, Step, AR->getLoop(), SCEV::FlagAnyWrap); 511 } 512 return 0; 513 } 514 515 // Distribute the sdiv over add operands, if the add doesn't overflow. 516 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(LHS)) { 517 if (IgnoreSignificantBits || isAddSExtable(Add, SE)) { 518 SmallVector<const SCEV *, 8> Ops; 519 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end(); 520 I != E; ++I) { 521 const SCEV *Op = getExactSDiv(*I, RHS, SE, 522 IgnoreSignificantBits); 523 if (!Op) return 0; 524 Ops.push_back(Op); 525 } 526 return SE.getAddExpr(Ops); 527 } 528 return 0; 529 } 530 531 // Check for a multiply operand that we can pull RHS out of. 532 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(LHS)) { 533 if (IgnoreSignificantBits || isMulSExtable(Mul, SE)) { 534 SmallVector<const SCEV *, 4> Ops; 535 bool Found = false; 536 for (SCEVMulExpr::op_iterator I = Mul->op_begin(), E = Mul->op_end(); 537 I != E; ++I) { 538 const SCEV *S = *I; 539 if (!Found) 540 if (const SCEV *Q = getExactSDiv(S, RHS, SE, 541 IgnoreSignificantBits)) { 542 S = Q; 543 Found = true; 544 } 545 Ops.push_back(S); 546 } 547 return Found ? SE.getMulExpr(Ops) : 0; 548 } 549 return 0; 550 } 551 552 // Otherwise we don't know. 553 return 0; 554 } 555 556 /// ExtractImmediate - If S involves the addition of a constant integer value, 557 /// return that integer value, and mutate S to point to a new SCEV with that 558 /// value excluded. 559 static int64_t ExtractImmediate(const SCEV *&S, ScalarEvolution &SE) { 560 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S)) { 561 if (C->getValue()->getValue().getMinSignedBits() <= 64) { 562 S = SE.getConstant(C->getType(), 0); 563 return C->getValue()->getSExtValue(); 564 } 565 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) { 566 SmallVector<const SCEV *, 8> NewOps(Add->op_begin(), Add->op_end()); 567 int64_t Result = ExtractImmediate(NewOps.front(), SE); 568 if (Result != 0) 569 S = SE.getAddExpr(NewOps); 570 return Result; 571 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) { 572 SmallVector<const SCEV *, 8> NewOps(AR->op_begin(), AR->op_end()); 573 int64_t Result = ExtractImmediate(NewOps.front(), SE); 574 if (Result != 0) 575 S = SE.getAddRecExpr(NewOps, AR->getLoop(), 576 // FIXME: AR->getNoWrapFlags(SCEV::FlagNW) 577 SCEV::FlagAnyWrap); 578 return Result; 579 } 580 return 0; 581 } 582 583 /// ExtractSymbol - If S involves the addition of a GlobalValue address, 584 /// return that symbol, and mutate S to point to a new SCEV with that 585 /// value excluded. 586 static GlobalValue *ExtractSymbol(const SCEV *&S, ScalarEvolution &SE) { 587 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) { 588 if (GlobalValue *GV = dyn_cast<GlobalValue>(U->getValue())) { 589 S = SE.getConstant(GV->getType(), 0); 590 return GV; 591 } 592 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) { 593 SmallVector<const SCEV *, 8> NewOps(Add->op_begin(), Add->op_end()); 594 GlobalValue *Result = ExtractSymbol(NewOps.back(), SE); 595 if (Result) 596 S = SE.getAddExpr(NewOps); 597 return Result; 598 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) { 599 SmallVector<const SCEV *, 8> NewOps(AR->op_begin(), AR->op_end()); 600 GlobalValue *Result = ExtractSymbol(NewOps.front(), SE); 601 if (Result) 602 S = SE.getAddRecExpr(NewOps, AR->getLoop(), 603 // FIXME: AR->getNoWrapFlags(SCEV::FlagNW) 604 SCEV::FlagAnyWrap); 605 return Result; 606 } 607 return 0; 608 } 609 610 /// isAddressUse - Returns true if the specified instruction is using the 611 /// specified value as an address. 612 static bool isAddressUse(Instruction *Inst, Value *OperandVal) { 613 bool isAddress = isa<LoadInst>(Inst); 614 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) { 615 if (SI->getOperand(1) == OperandVal) 616 isAddress = true; 617 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) { 618 // Addressing modes can also be folded into prefetches and a variety 619 // of intrinsics. 620 switch (II->getIntrinsicID()) { 621 default: break; 622 case Intrinsic::prefetch: 623 case Intrinsic::x86_sse_storeu_ps: 624 case Intrinsic::x86_sse2_storeu_pd: 625 case Intrinsic::x86_sse2_storeu_dq: 626 case Intrinsic::x86_sse2_storel_dq: 627 if (II->getArgOperand(0) == OperandVal) 628 isAddress = true; 629 break; 630 } 631 } 632 return isAddress; 633 } 634 635 /// getAccessType - Return the type of the memory being accessed. 636 static Type *getAccessType(const Instruction *Inst) { 637 Type *AccessTy = Inst->getType(); 638 if (const StoreInst *SI = dyn_cast<StoreInst>(Inst)) 639 AccessTy = SI->getOperand(0)->getType(); 640 else if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) { 641 // Addressing modes can also be folded into prefetches and a variety 642 // of intrinsics. 643 switch (II->getIntrinsicID()) { 644 default: break; 645 case Intrinsic::x86_sse_storeu_ps: 646 case Intrinsic::x86_sse2_storeu_pd: 647 case Intrinsic::x86_sse2_storeu_dq: 648 case Intrinsic::x86_sse2_storel_dq: 649 AccessTy = II->getArgOperand(0)->getType(); 650 break; 651 } 652 } 653 654 // All pointers have the same requirements, so canonicalize them to an 655 // arbitrary pointer type to minimize variation. 656 if (PointerType *PTy = dyn_cast<PointerType>(AccessTy)) 657 AccessTy = PointerType::get(IntegerType::get(PTy->getContext(), 1), 658 PTy->getAddressSpace()); 659 660 return AccessTy; 661 } 662 663 /// isExistingPhi - Return true if this AddRec is already a phi in its loop. 664 static bool isExistingPhi(const SCEVAddRecExpr *AR, ScalarEvolution &SE) { 665 for (BasicBlock::iterator I = AR->getLoop()->getHeader()->begin(); 666 PHINode *PN = dyn_cast<PHINode>(I); ++I) { 667 if (SE.isSCEVable(PN->getType()) && 668 (SE.getEffectiveSCEVType(PN->getType()) == 669 SE.getEffectiveSCEVType(AR->getType())) && 670 SE.getSCEV(PN) == AR) 671 return true; 672 } 673 return false; 674 } 675 676 /// Check if expanding this expression is likely to incur significant cost. This 677 /// is tricky because SCEV doesn't track which expressions are actually computed 678 /// by the current IR. 679 /// 680 /// We currently allow expansion of IV increments that involve adds, 681 /// multiplication by constants, and AddRecs from existing phis. 682 /// 683 /// TODO: Allow UDivExpr if we can find an existing IV increment that is an 684 /// obvious multiple of the UDivExpr. 685 static bool isHighCostExpansion(const SCEV *S, 686 SmallPtrSet<const SCEV*, 8> &Processed, 687 ScalarEvolution &SE) { 688 // Zero/One operand expressions 689 switch (S->getSCEVType()) { 690 case scUnknown: 691 case scConstant: 692 return false; 693 case scTruncate: 694 return isHighCostExpansion(cast<SCEVTruncateExpr>(S)->getOperand(), 695 Processed, SE); 696 case scZeroExtend: 697 return isHighCostExpansion(cast<SCEVZeroExtendExpr>(S)->getOperand(), 698 Processed, SE); 699 case scSignExtend: 700 return isHighCostExpansion(cast<SCEVSignExtendExpr>(S)->getOperand(), 701 Processed, SE); 702 } 703 704 if (!Processed.insert(S)) 705 return false; 706 707 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) { 708 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end(); 709 I != E; ++I) { 710 if (isHighCostExpansion(*I, Processed, SE)) 711 return true; 712 } 713 return false; 714 } 715 716 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) { 717 if (Mul->getNumOperands() == 2) { 718 // Multiplication by a constant is ok 719 if (isa<SCEVConstant>(Mul->getOperand(0))) 720 return isHighCostExpansion(Mul->getOperand(1), Processed, SE); 721 722 // If we have the value of one operand, check if an existing 723 // multiplication already generates this expression. 724 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(Mul->getOperand(1))) { 725 Value *UVal = U->getValue(); 726 for (Value::use_iterator UI = UVal->use_begin(), UE = UVal->use_end(); 727 UI != UE; ++UI) { 728 // If U is a constant, it may be used by a ConstantExpr. 729 Instruction *User = dyn_cast<Instruction>(*UI); 730 if (User && User->getOpcode() == Instruction::Mul 731 && SE.isSCEVable(User->getType())) { 732 return SE.getSCEV(User) == Mul; 733 } 734 } 735 } 736 } 737 } 738 739 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) { 740 if (isExistingPhi(AR, SE)) 741 return false; 742 } 743 744 // Fow now, consider any other type of expression (div/mul/min/max) high cost. 745 return true; 746 } 747 748 /// DeleteTriviallyDeadInstructions - If any of the instructions is the 749 /// specified set are trivially dead, delete them and see if this makes any of 750 /// their operands subsequently dead. 751 static bool 752 DeleteTriviallyDeadInstructions(SmallVectorImpl<WeakVH> &DeadInsts) { 753 bool Changed = false; 754 755 while (!DeadInsts.empty()) { 756 Value *V = DeadInsts.pop_back_val(); 757 Instruction *I = dyn_cast_or_null<Instruction>(V); 758 759 if (I == 0 || !isInstructionTriviallyDead(I)) 760 continue; 761 762 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI) 763 if (Instruction *U = dyn_cast<Instruction>(*OI)) { 764 *OI = 0; 765 if (U->use_empty()) 766 DeadInsts.push_back(U); 767 } 768 769 I->eraseFromParent(); 770 Changed = true; 771 } 772 773 return Changed; 774 } 775 776 namespace { 777 778 /// Cost - This class is used to measure and compare candidate formulae. 779 class Cost { 780 /// TODO: Some of these could be merged. Also, a lexical ordering 781 /// isn't always optimal. 782 unsigned NumRegs; 783 unsigned AddRecCost; 784 unsigned NumIVMuls; 785 unsigned NumBaseAdds; 786 unsigned ImmCost; 787 unsigned SetupCost; 788 789 public: 790 Cost() 791 : NumRegs(0), AddRecCost(0), NumIVMuls(0), NumBaseAdds(0), ImmCost(0), 792 SetupCost(0) {} 793 794 bool operator<(const Cost &Other) const; 795 796 void Loose(); 797 798 #ifndef NDEBUG 799 // Once any of the metrics loses, they must all remain losers. 800 bool isValid() { 801 return ((NumRegs | AddRecCost | NumIVMuls | NumBaseAdds 802 | ImmCost | SetupCost) != ~0u) 803 || ((NumRegs & AddRecCost & NumIVMuls & NumBaseAdds 804 & ImmCost & SetupCost) == ~0u); 805 } 806 #endif 807 808 bool isLoser() { 809 assert(isValid() && "invalid cost"); 810 return NumRegs == ~0u; 811 } 812 813 void RateFormula(const Formula &F, 814 SmallPtrSet<const SCEV *, 16> &Regs, 815 const DenseSet<const SCEV *> &VisitedRegs, 816 const Loop *L, 817 const SmallVectorImpl<int64_t> &Offsets, 818 ScalarEvolution &SE, DominatorTree &DT, 819 SmallPtrSet<const SCEV *, 16> *LoserRegs = 0); 820 821 void print(raw_ostream &OS) const; 822 void dump() const; 823 824 private: 825 void RateRegister(const SCEV *Reg, 826 SmallPtrSet<const SCEV *, 16> &Regs, 827 const Loop *L, 828 ScalarEvolution &SE, DominatorTree &DT); 829 void RatePrimaryRegister(const SCEV *Reg, 830 SmallPtrSet<const SCEV *, 16> &Regs, 831 const Loop *L, 832 ScalarEvolution &SE, DominatorTree &DT, 833 SmallPtrSet<const SCEV *, 16> *LoserRegs); 834 }; 835 836 } 837 838 /// RateRegister - Tally up interesting quantities from the given register. 839 void Cost::RateRegister(const SCEV *Reg, 840 SmallPtrSet<const SCEV *, 16> &Regs, 841 const Loop *L, 842 ScalarEvolution &SE, DominatorTree &DT) { 843 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Reg)) { 844 // If this is an addrec for another loop, don't second-guess its addrec phi 845 // nodes. LSR isn't currently smart enough to reason about more than one 846 // loop at a time. LSR has already run on inner loops, will not run on outer 847 // loops, and cannot be expected to change sibling loops. 848 if (AR->getLoop() != L) { 849 // If the AddRec exists, consider it's register free and leave it alone. 850 if (isExistingPhi(AR, SE)) 851 return; 852 853 // Otherwise, do not consider this formula at all. 854 Loose(); 855 return; 856 } 857 AddRecCost += 1; /// TODO: This should be a function of the stride. 858 859 // Add the step value register, if it needs one. 860 // TODO: The non-affine case isn't precisely modeled here. 861 if (!AR->isAffine() || !isa<SCEVConstant>(AR->getOperand(1))) { 862 if (!Regs.count(AR->getOperand(1))) { 863 RateRegister(AR->getOperand(1), Regs, L, SE, DT); 864 if (isLoser()) 865 return; 866 } 867 } 868 } 869 ++NumRegs; 870 871 // Rough heuristic; favor registers which don't require extra setup 872 // instructions in the preheader. 873 if (!isa<SCEVUnknown>(Reg) && 874 !isa<SCEVConstant>(Reg) && 875 !(isa<SCEVAddRecExpr>(Reg) && 876 (isa<SCEVUnknown>(cast<SCEVAddRecExpr>(Reg)->getStart()) || 877 isa<SCEVConstant>(cast<SCEVAddRecExpr>(Reg)->getStart())))) 878 ++SetupCost; 879 880 NumIVMuls += isa<SCEVMulExpr>(Reg) && 881 SE.hasComputableLoopEvolution(Reg, L); 882 } 883 884 /// RatePrimaryRegister - Record this register in the set. If we haven't seen it 885 /// before, rate it. Optional LoserRegs provides a way to declare any formula 886 /// that refers to one of those regs an instant loser. 887 void Cost::RatePrimaryRegister(const SCEV *Reg, 888 SmallPtrSet<const SCEV *, 16> &Regs, 889 const Loop *L, 890 ScalarEvolution &SE, DominatorTree &DT, 891 SmallPtrSet<const SCEV *, 16> *LoserRegs) { 892 if (LoserRegs && LoserRegs->count(Reg)) { 893 Loose(); 894 return; 895 } 896 if (Regs.insert(Reg)) { 897 RateRegister(Reg, Regs, L, SE, DT); 898 if (isLoser()) 899 LoserRegs->insert(Reg); 900 } 901 } 902 903 void Cost::RateFormula(const Formula &F, 904 SmallPtrSet<const SCEV *, 16> &Regs, 905 const DenseSet<const SCEV *> &VisitedRegs, 906 const Loop *L, 907 const SmallVectorImpl<int64_t> &Offsets, 908 ScalarEvolution &SE, DominatorTree &DT, 909 SmallPtrSet<const SCEV *, 16> *LoserRegs) { 910 // Tally up the registers. 911 if (const SCEV *ScaledReg = F.ScaledReg) { 912 if (VisitedRegs.count(ScaledReg)) { 913 Loose(); 914 return; 915 } 916 RatePrimaryRegister(ScaledReg, Regs, L, SE, DT, LoserRegs); 917 if (isLoser()) 918 return; 919 } 920 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(), 921 E = F.BaseRegs.end(); I != E; ++I) { 922 const SCEV *BaseReg = *I; 923 if (VisitedRegs.count(BaseReg)) { 924 Loose(); 925 return; 926 } 927 RatePrimaryRegister(BaseReg, Regs, L, SE, DT, LoserRegs); 928 if (isLoser()) 929 return; 930 } 931 932 // Determine how many (unfolded) adds we'll need inside the loop. 933 size_t NumBaseParts = F.BaseRegs.size() + (F.UnfoldedOffset != 0); 934 if (NumBaseParts > 1) 935 NumBaseAdds += NumBaseParts - 1; 936 937 // Tally up the non-zero immediates. 938 for (SmallVectorImpl<int64_t>::const_iterator I = Offsets.begin(), 939 E = Offsets.end(); I != E; ++I) { 940 int64_t Offset = (uint64_t)*I + F.BaseOffset; 941 if (F.BaseGV) 942 ImmCost += 64; // Handle symbolic values conservatively. 943 // TODO: This should probably be the pointer size. 944 else if (Offset != 0) 945 ImmCost += APInt(64, Offset, true).getMinSignedBits(); 946 } 947 assert(isValid() && "invalid cost"); 948 } 949 950 /// Loose - Set this cost to a losing value. 951 void Cost::Loose() { 952 NumRegs = ~0u; 953 AddRecCost = ~0u; 954 NumIVMuls = ~0u; 955 NumBaseAdds = ~0u; 956 ImmCost = ~0u; 957 SetupCost = ~0u; 958 } 959 960 /// operator< - Choose the lower cost. 961 bool Cost::operator<(const Cost &Other) const { 962 if (NumRegs != Other.NumRegs) 963 return NumRegs < Other.NumRegs; 964 if (AddRecCost != Other.AddRecCost) 965 return AddRecCost < Other.AddRecCost; 966 if (NumIVMuls != Other.NumIVMuls) 967 return NumIVMuls < Other.NumIVMuls; 968 if (NumBaseAdds != Other.NumBaseAdds) 969 return NumBaseAdds < Other.NumBaseAdds; 970 if (ImmCost != Other.ImmCost) 971 return ImmCost < Other.ImmCost; 972 if (SetupCost != Other.SetupCost) 973 return SetupCost < Other.SetupCost; 974 return false; 975 } 976 977 void Cost::print(raw_ostream &OS) const { 978 OS << NumRegs << " reg" << (NumRegs == 1 ? "" : "s"); 979 if (AddRecCost != 0) 980 OS << ", with addrec cost " << AddRecCost; 981 if (NumIVMuls != 0) 982 OS << ", plus " << NumIVMuls << " IV mul" << (NumIVMuls == 1 ? "" : "s"); 983 if (NumBaseAdds != 0) 984 OS << ", plus " << NumBaseAdds << " base add" 985 << (NumBaseAdds == 1 ? "" : "s"); 986 if (ImmCost != 0) 987 OS << ", plus " << ImmCost << " imm cost"; 988 if (SetupCost != 0) 989 OS << ", plus " << SetupCost << " setup cost"; 990 } 991 992 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) 993 void Cost::dump() const { 994 print(errs()); errs() << '\n'; 995 } 996 #endif 997 998 namespace { 999 1000 /// LSRFixup - An operand value in an instruction which is to be replaced 1001 /// with some equivalent, possibly strength-reduced, replacement. 1002 struct LSRFixup { 1003 /// UserInst - The instruction which will be updated. 1004 Instruction *UserInst; 1005 1006 /// OperandValToReplace - The operand of the instruction which will 1007 /// be replaced. The operand may be used more than once; every instance 1008 /// will be replaced. 1009 Value *OperandValToReplace; 1010 1011 /// PostIncLoops - If this user is to use the post-incremented value of an 1012 /// induction variable, this variable is non-null and holds the loop 1013 /// associated with the induction variable. 1014 PostIncLoopSet PostIncLoops; 1015 1016 /// LUIdx - The index of the LSRUse describing the expression which 1017 /// this fixup needs, minus an offset (below). 1018 size_t LUIdx; 1019 1020 /// Offset - A constant offset to be added to the LSRUse expression. 1021 /// This allows multiple fixups to share the same LSRUse with different 1022 /// offsets, for example in an unrolled loop. 1023 int64_t Offset; 1024 1025 bool isUseFullyOutsideLoop(const Loop *L) const; 1026 1027 LSRFixup(); 1028 1029 void print(raw_ostream &OS) const; 1030 void dump() const; 1031 }; 1032 1033 } 1034 1035 LSRFixup::LSRFixup() 1036 : UserInst(0), OperandValToReplace(0), LUIdx(~size_t(0)), Offset(0) {} 1037 1038 /// isUseFullyOutsideLoop - Test whether this fixup always uses its 1039 /// value outside of the given loop. 1040 bool LSRFixup::isUseFullyOutsideLoop(const Loop *L) const { 1041 // PHI nodes use their value in their incoming blocks. 1042 if (const PHINode *PN = dyn_cast<PHINode>(UserInst)) { 1043 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) 1044 if (PN->getIncomingValue(i) == OperandValToReplace && 1045 L->contains(PN->getIncomingBlock(i))) 1046 return false; 1047 return true; 1048 } 1049 1050 return !L->contains(UserInst); 1051 } 1052 1053 void LSRFixup::print(raw_ostream &OS) const { 1054 OS << "UserInst="; 1055 // Store is common and interesting enough to be worth special-casing. 1056 if (StoreInst *Store = dyn_cast<StoreInst>(UserInst)) { 1057 OS << "store "; 1058 WriteAsOperand(OS, Store->getOperand(0), /*PrintType=*/false); 1059 } else if (UserInst->getType()->isVoidTy()) 1060 OS << UserInst->getOpcodeName(); 1061 else 1062 WriteAsOperand(OS, UserInst, /*PrintType=*/false); 1063 1064 OS << ", OperandValToReplace="; 1065 WriteAsOperand(OS, OperandValToReplace, /*PrintType=*/false); 1066 1067 for (PostIncLoopSet::const_iterator I = PostIncLoops.begin(), 1068 E = PostIncLoops.end(); I != E; ++I) { 1069 OS << ", PostIncLoop="; 1070 WriteAsOperand(OS, (*I)->getHeader(), /*PrintType=*/false); 1071 } 1072 1073 if (LUIdx != ~size_t(0)) 1074 OS << ", LUIdx=" << LUIdx; 1075 1076 if (Offset != 0) 1077 OS << ", Offset=" << Offset; 1078 } 1079 1080 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) 1081 void LSRFixup::dump() const { 1082 print(errs()); errs() << '\n'; 1083 } 1084 #endif 1085 1086 namespace { 1087 1088 /// UniquifierDenseMapInfo - A DenseMapInfo implementation for holding 1089 /// DenseMaps and DenseSets of sorted SmallVectors of const SCEV*. 1090 struct UniquifierDenseMapInfo { 1091 static SmallVector<const SCEV *, 4> getEmptyKey() { 1092 SmallVector<const SCEV *, 4> V; 1093 V.push_back(reinterpret_cast<const SCEV *>(-1)); 1094 return V; 1095 } 1096 1097 static SmallVector<const SCEV *, 4> getTombstoneKey() { 1098 SmallVector<const SCEV *, 4> V; 1099 V.push_back(reinterpret_cast<const SCEV *>(-2)); 1100 return V; 1101 } 1102 1103 static unsigned getHashValue(const SmallVector<const SCEV *, 4> &V) { 1104 unsigned Result = 0; 1105 for (SmallVectorImpl<const SCEV *>::const_iterator I = V.begin(), 1106 E = V.end(); I != E; ++I) 1107 Result ^= DenseMapInfo<const SCEV *>::getHashValue(*I); 1108 return Result; 1109 } 1110 1111 static bool isEqual(const SmallVector<const SCEV *, 4> &LHS, 1112 const SmallVector<const SCEV *, 4> &RHS) { 1113 return LHS == RHS; 1114 } 1115 }; 1116 1117 /// LSRUse - This class holds the state that LSR keeps for each use in 1118 /// IVUsers, as well as uses invented by LSR itself. It includes information 1119 /// about what kinds of things can be folded into the user, information about 1120 /// the user itself, and information about how the use may be satisfied. 1121 /// TODO: Represent multiple users of the same expression in common? 1122 class LSRUse { 1123 DenseSet<SmallVector<const SCEV *, 4>, UniquifierDenseMapInfo> Uniquifier; 1124 1125 public: 1126 /// KindType - An enum for a kind of use, indicating what types of 1127 /// scaled and immediate operands it might support. 1128 enum KindType { 1129 Basic, ///< A normal use, with no folding. 1130 Special, ///< A special case of basic, allowing -1 scales. 1131 Address, ///< An address use; folding according to TargetLowering 1132 ICmpZero ///< An equality icmp with both operands folded into one. 1133 // TODO: Add a generic icmp too? 1134 }; 1135 1136 KindType Kind; 1137 Type *AccessTy; 1138 1139 SmallVector<int64_t, 8> Offsets; 1140 int64_t MinOffset; 1141 int64_t MaxOffset; 1142 1143 /// AllFixupsOutsideLoop - This records whether all of the fixups using this 1144 /// LSRUse are outside of the loop, in which case some special-case heuristics 1145 /// may be used. 1146 bool AllFixupsOutsideLoop; 1147 1148 /// WidestFixupType - This records the widest use type for any fixup using 1149 /// this LSRUse. FindUseWithSimilarFormula can't consider uses with different 1150 /// max fixup widths to be equivalent, because the narrower one may be relying 1151 /// on the implicit truncation to truncate away bogus bits. 1152 Type *WidestFixupType; 1153 1154 /// Formulae - A list of ways to build a value that can satisfy this user. 1155 /// After the list is populated, one of these is selected heuristically and 1156 /// used to formulate a replacement for OperandValToReplace in UserInst. 1157 SmallVector<Formula, 12> Formulae; 1158 1159 /// Regs - The set of register candidates used by all formulae in this LSRUse. 1160 SmallPtrSet<const SCEV *, 4> Regs; 1161 1162 LSRUse(KindType K, Type *T) : Kind(K), AccessTy(T), 1163 MinOffset(INT64_MAX), 1164 MaxOffset(INT64_MIN), 1165 AllFixupsOutsideLoop(true), 1166 WidestFixupType(0) {} 1167 1168 bool HasFormulaWithSameRegs(const Formula &F) const; 1169 bool InsertFormula(const Formula &F); 1170 void DeleteFormula(Formula &F); 1171 void RecomputeRegs(size_t LUIdx, RegUseTracker &Reguses); 1172 1173 void print(raw_ostream &OS) const; 1174 void dump() const; 1175 }; 1176 1177 } 1178 1179 /// HasFormula - Test whether this use as a formula which has the same 1180 /// registers as the given formula. 1181 bool LSRUse::HasFormulaWithSameRegs(const Formula &F) const { 1182 SmallVector<const SCEV *, 4> Key = F.BaseRegs; 1183 if (F.ScaledReg) Key.push_back(F.ScaledReg); 1184 // Unstable sort by host order ok, because this is only used for uniquifying. 1185 std::sort(Key.begin(), Key.end()); 1186 return Uniquifier.count(Key); 1187 } 1188 1189 /// InsertFormula - If the given formula has not yet been inserted, add it to 1190 /// the list, and return true. Return false otherwise. 1191 bool LSRUse::InsertFormula(const Formula &F) { 1192 SmallVector<const SCEV *, 4> Key = F.BaseRegs; 1193 if (F.ScaledReg) Key.push_back(F.ScaledReg); 1194 // Unstable sort by host order ok, because this is only used for uniquifying. 1195 std::sort(Key.begin(), Key.end()); 1196 1197 if (!Uniquifier.insert(Key).second) 1198 return false; 1199 1200 // Using a register to hold the value of 0 is not profitable. 1201 assert((!F.ScaledReg || !F.ScaledReg->isZero()) && 1202 "Zero allocated in a scaled register!"); 1203 #ifndef NDEBUG 1204 for (SmallVectorImpl<const SCEV *>::const_iterator I = 1205 F.BaseRegs.begin(), E = F.BaseRegs.end(); I != E; ++I) 1206 assert(!(*I)->isZero() && "Zero allocated in a base register!"); 1207 #endif 1208 1209 // Add the formula to the list. 1210 Formulae.push_back(F); 1211 1212 // Record registers now being used by this use. 1213 Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end()); 1214 1215 return true; 1216 } 1217 1218 /// DeleteFormula - Remove the given formula from this use's list. 1219 void LSRUse::DeleteFormula(Formula &F) { 1220 if (&F != &Formulae.back()) 1221 std::swap(F, Formulae.back()); 1222 Formulae.pop_back(); 1223 } 1224 1225 /// RecomputeRegs - Recompute the Regs field, and update RegUses. 1226 void LSRUse::RecomputeRegs(size_t LUIdx, RegUseTracker &RegUses) { 1227 // Now that we've filtered out some formulae, recompute the Regs set. 1228 SmallPtrSet<const SCEV *, 4> OldRegs = Regs; 1229 Regs.clear(); 1230 for (SmallVectorImpl<Formula>::const_iterator I = Formulae.begin(), 1231 E = Formulae.end(); I != E; ++I) { 1232 const Formula &F = *I; 1233 if (F.ScaledReg) Regs.insert(F.ScaledReg); 1234 Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end()); 1235 } 1236 1237 // Update the RegTracker. 1238 for (SmallPtrSet<const SCEV *, 4>::iterator I = OldRegs.begin(), 1239 E = OldRegs.end(); I != E; ++I) 1240 if (!Regs.count(*I)) 1241 RegUses.DropRegister(*I, LUIdx); 1242 } 1243 1244 void LSRUse::print(raw_ostream &OS) const { 1245 OS << "LSR Use: Kind="; 1246 switch (Kind) { 1247 case Basic: OS << "Basic"; break; 1248 case Special: OS << "Special"; break; 1249 case ICmpZero: OS << "ICmpZero"; break; 1250 case Address: 1251 OS << "Address of "; 1252 if (AccessTy->isPointerTy()) 1253 OS << "pointer"; // the full pointer type could be really verbose 1254 else 1255 OS << *AccessTy; 1256 } 1257 1258 OS << ", Offsets={"; 1259 for (SmallVectorImpl<int64_t>::const_iterator I = Offsets.begin(), 1260 E = Offsets.end(); I != E; ++I) { 1261 OS << *I; 1262 if (llvm::next(I) != E) 1263 OS << ','; 1264 } 1265 OS << '}'; 1266 1267 if (AllFixupsOutsideLoop) 1268 OS << ", all-fixups-outside-loop"; 1269 1270 if (WidestFixupType) 1271 OS << ", widest fixup type: " << *WidestFixupType; 1272 } 1273 1274 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) 1275 void LSRUse::dump() const { 1276 print(errs()); errs() << '\n'; 1277 } 1278 #endif 1279 1280 /// isLegalUse - Test whether the use described by AM is "legal", meaning it can 1281 /// be completely folded into the user instruction at isel time. This includes 1282 /// address-mode folding and special icmp tricks. 1283 static bool isLegalUse(const TargetTransformInfo &TTI, LSRUse::KindType Kind, 1284 Type *AccessTy, GlobalValue *BaseGV, int64_t BaseOffset, 1285 bool HasBaseReg, int64_t Scale) { 1286 switch (Kind) { 1287 case LSRUse::Address: 1288 return TTI.isLegalAddressingMode(AccessTy, BaseGV, BaseOffset, HasBaseReg, Scale); 1289 1290 // Otherwise, just guess that reg+reg addressing is legal. 1291 //return ; 1292 1293 case LSRUse::ICmpZero: 1294 // There's not even a target hook for querying whether it would be legal to 1295 // fold a GV into an ICmp. 1296 if (BaseGV) 1297 return false; 1298 1299 // ICmp only has two operands; don't allow more than two non-trivial parts. 1300 if (Scale != 0 && HasBaseReg && BaseOffset != 0) 1301 return false; 1302 1303 // ICmp only supports no scale or a -1 scale, as we can "fold" a -1 scale by 1304 // putting the scaled register in the other operand of the icmp. 1305 if (Scale != 0 && Scale != -1) 1306 return false; 1307 1308 // If we have low-level target information, ask the target if it can fold an 1309 // integer immediate on an icmp. 1310 if (BaseOffset != 0) { 1311 // We have one of: 1312 // ICmpZero BaseReg + BaseOffset => ICmp BaseReg, -BaseOffset 1313 // ICmpZero -1*ScaleReg + BaseOffset => ICmp ScaleReg, BaseOffset 1314 // Offs is the ICmp immediate. 1315 if (Scale == 0) 1316 // The cast does the right thing with INT64_MIN. 1317 BaseOffset = -(uint64_t)BaseOffset; 1318 return TTI.isLegalICmpImmediate(BaseOffset); 1319 } 1320 1321 // ICmpZero BaseReg + -1*ScaleReg => ICmp BaseReg, ScaleReg 1322 return true; 1323 1324 case LSRUse::Basic: 1325 // Only handle single-register values. 1326 return !BaseGV && Scale == 0 && BaseOffset == 0; 1327 1328 case LSRUse::Special: 1329 // Special case Basic to handle -1 scales. 1330 return !BaseGV && (Scale == 0 || Scale == -1) && BaseOffset == 0; 1331 } 1332 1333 llvm_unreachable("Invalid LSRUse Kind!"); 1334 } 1335 1336 static bool isLegalUse(const TargetTransformInfo &TTI, int64_t MinOffset, 1337 int64_t MaxOffset, LSRUse::KindType Kind, Type *AccessTy, 1338 GlobalValue *BaseGV, int64_t BaseOffset, bool HasBaseReg, 1339 int64_t Scale) { 1340 // Check for overflow. 1341 if (((int64_t)((uint64_t)BaseOffset + MinOffset) > BaseOffset) != 1342 (MinOffset > 0)) 1343 return false; 1344 MinOffset = (uint64_t)BaseOffset + MinOffset; 1345 if (((int64_t)((uint64_t)BaseOffset + MaxOffset) > BaseOffset) != 1346 (MaxOffset > 0)) 1347 return false; 1348 MaxOffset = (uint64_t)BaseOffset + MaxOffset; 1349 1350 return isLegalUse(TTI, Kind, AccessTy, BaseGV, MinOffset, HasBaseReg, 1351 Scale) && 1352 isLegalUse(TTI, Kind, AccessTy, BaseGV, MaxOffset, HasBaseReg, Scale); 1353 } 1354 1355 static bool isLegalUse(const TargetTransformInfo &TTI, int64_t MinOffset, 1356 int64_t MaxOffset, LSRUse::KindType Kind, Type *AccessTy, 1357 const Formula &F) { 1358 return isLegalUse(TTI, MinOffset, MaxOffset, Kind, AccessTy, F.BaseGV, 1359 F.BaseOffset, F.HasBaseReg, F.Scale); 1360 } 1361 1362 static bool isAlwaysFoldable(const TargetTransformInfo &TTI, 1363 LSRUse::KindType Kind, Type *AccessTy, 1364 GlobalValue *BaseGV, int64_t BaseOffset, 1365 bool HasBaseReg) { 1366 // Fast-path: zero is always foldable. 1367 if (BaseOffset == 0 && !BaseGV) return true; 1368 1369 // Conservatively, create an address with an immediate and a 1370 // base and a scale. 1371 int64_t Scale = Kind == LSRUse::ICmpZero ? -1 : 1; 1372 1373 // Canonicalize a scale of 1 to a base register if the formula doesn't 1374 // already have a base register. 1375 if (!HasBaseReg && Scale == 1) { 1376 Scale = 0; 1377 HasBaseReg = true; 1378 } 1379 1380 return isLegalUse(TTI, Kind, AccessTy, BaseGV, BaseOffset, HasBaseReg, Scale); 1381 } 1382 1383 static bool isAlwaysFoldable(const TargetTransformInfo &TTI, 1384 ScalarEvolution &SE, int64_t MinOffset, 1385 int64_t MaxOffset, LSRUse::KindType Kind, 1386 Type *AccessTy, const SCEV *S, bool HasBaseReg) { 1387 // Fast-path: zero is always foldable. 1388 if (S->isZero()) return true; 1389 1390 // Conservatively, create an address with an immediate and a 1391 // base and a scale. 1392 int64_t BaseOffset = ExtractImmediate(S, SE); 1393 GlobalValue *BaseGV = ExtractSymbol(S, SE); 1394 1395 // If there's anything else involved, it's not foldable. 1396 if (!S->isZero()) return false; 1397 1398 // Fast-path: zero is always foldable. 1399 if (BaseOffset == 0 && !BaseGV) return true; 1400 1401 // Conservatively, create an address with an immediate and a 1402 // base and a scale. 1403 int64_t Scale = Kind == LSRUse::ICmpZero ? -1 : 1; 1404 1405 return isLegalUse(TTI, MinOffset, MaxOffset, Kind, AccessTy, BaseGV, 1406 BaseOffset, HasBaseReg, Scale); 1407 } 1408 1409 namespace { 1410 1411 /// UseMapDenseMapInfo - A DenseMapInfo implementation for holding 1412 /// DenseMaps and DenseSets of pairs of const SCEV* and LSRUse::Kind. 1413 struct UseMapDenseMapInfo { 1414 static std::pair<const SCEV *, LSRUse::KindType> getEmptyKey() { 1415 return std::make_pair(reinterpret_cast<const SCEV *>(-1), LSRUse::Basic); 1416 } 1417 1418 static std::pair<const SCEV *, LSRUse::KindType> getTombstoneKey() { 1419 return std::make_pair(reinterpret_cast<const SCEV *>(-2), LSRUse::Basic); 1420 } 1421 1422 static unsigned 1423 getHashValue(const std::pair<const SCEV *, LSRUse::KindType> &V) { 1424 unsigned Result = DenseMapInfo<const SCEV *>::getHashValue(V.first); 1425 Result ^= DenseMapInfo<unsigned>::getHashValue(unsigned(V.second)); 1426 return Result; 1427 } 1428 1429 static bool isEqual(const std::pair<const SCEV *, LSRUse::KindType> &LHS, 1430 const std::pair<const SCEV *, LSRUse::KindType> &RHS) { 1431 return LHS == RHS; 1432 } 1433 }; 1434 1435 /// IVInc - An individual increment in a Chain of IV increments. 1436 /// Relate an IV user to an expression that computes the IV it uses from the IV 1437 /// used by the previous link in the Chain. 1438 /// 1439 /// For the head of a chain, IncExpr holds the absolute SCEV expression for the 1440 /// original IVOperand. The head of the chain's IVOperand is only valid during 1441 /// chain collection, before LSR replaces IV users. During chain generation, 1442 /// IncExpr can be used to find the new IVOperand that computes the same 1443 /// expression. 1444 struct IVInc { 1445 Instruction *UserInst; 1446 Value* IVOperand; 1447 const SCEV *IncExpr; 1448 1449 IVInc(Instruction *U, Value *O, const SCEV *E): 1450 UserInst(U), IVOperand(O), IncExpr(E) {} 1451 }; 1452 1453 // IVChain - The list of IV increments in program order. 1454 // We typically add the head of a chain without finding subsequent links. 1455 struct IVChain { 1456 SmallVector<IVInc,1> Incs; 1457 const SCEV *ExprBase; 1458 1459 IVChain() : ExprBase(0) {} 1460 1461 IVChain(const IVInc &Head, const SCEV *Base) 1462 : Incs(1, Head), ExprBase(Base) {} 1463 1464 typedef SmallVectorImpl<IVInc>::const_iterator const_iterator; 1465 1466 // begin - return the first increment in the chain. 1467 const_iterator begin() const { 1468 assert(!Incs.empty()); 1469 return llvm::next(Incs.begin()); 1470 } 1471 const_iterator end() const { 1472 return Incs.end(); 1473 } 1474 1475 // hasIncs - Returns true if this chain contains any increments. 1476 bool hasIncs() const { return Incs.size() >= 2; } 1477 1478 // add - Add an IVInc to the end of this chain. 1479 void add(const IVInc &X) { Incs.push_back(X); } 1480 1481 // tailUserInst - Returns the last UserInst in the chain. 1482 Instruction *tailUserInst() const { return Incs.back().UserInst; } 1483 1484 // isProfitableIncrement - Returns true if IncExpr can be profitably added to 1485 // this chain. 1486 bool isProfitableIncrement(const SCEV *OperExpr, 1487 const SCEV *IncExpr, 1488 ScalarEvolution&); 1489 }; 1490 1491 /// ChainUsers - Helper for CollectChains to track multiple IV increment uses. 1492 /// Distinguish between FarUsers that definitely cross IV increments and 1493 /// NearUsers that may be used between IV increments. 1494 struct ChainUsers { 1495 SmallPtrSet<Instruction*, 4> FarUsers; 1496 SmallPtrSet<Instruction*, 4> NearUsers; 1497 }; 1498 1499 /// LSRInstance - This class holds state for the main loop strength reduction 1500 /// logic. 1501 class LSRInstance { 1502 IVUsers &IU; 1503 ScalarEvolution &SE; 1504 DominatorTree &DT; 1505 LoopInfo &LI; 1506 const TargetTransformInfo &TTI; 1507 Loop *const L; 1508 bool Changed; 1509 1510 /// IVIncInsertPos - This is the insert position that the current loop's 1511 /// induction variable increment should be placed. In simple loops, this is 1512 /// the latch block's terminator. But in more complicated cases, this is a 1513 /// position which will dominate all the in-loop post-increment users. 1514 Instruction *IVIncInsertPos; 1515 1516 /// Factors - Interesting factors between use strides. 1517 SmallSetVector<int64_t, 8> Factors; 1518 1519 /// Types - Interesting use types, to facilitate truncation reuse. 1520 SmallSetVector<Type *, 4> Types; 1521 1522 /// Fixups - The list of operands which are to be replaced. 1523 SmallVector<LSRFixup, 16> Fixups; 1524 1525 /// Uses - The list of interesting uses. 1526 SmallVector<LSRUse, 16> Uses; 1527 1528 /// RegUses - Track which uses use which register candidates. 1529 RegUseTracker RegUses; 1530 1531 // Limit the number of chains to avoid quadratic behavior. We don't expect to 1532 // have more than a few IV increment chains in a loop. Missing a Chain falls 1533 // back to normal LSR behavior for those uses. 1534 static const unsigned MaxChains = 8; 1535 1536 /// IVChainVec - IV users can form a chain of IV increments. 1537 SmallVector<IVChain, MaxChains> IVChainVec; 1538 1539 /// IVIncSet - IV users that belong to profitable IVChains. 1540 SmallPtrSet<Use*, MaxChains> IVIncSet; 1541 1542 void OptimizeShadowIV(); 1543 bool FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse); 1544 ICmpInst *OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse); 1545 void OptimizeLoopTermCond(); 1546 1547 void ChainInstruction(Instruction *UserInst, Instruction *IVOper, 1548 SmallVectorImpl<ChainUsers> &ChainUsersVec); 1549 void FinalizeChain(IVChain &Chain); 1550 void CollectChains(); 1551 void GenerateIVChain(const IVChain &Chain, SCEVExpander &Rewriter, 1552 SmallVectorImpl<WeakVH> &DeadInsts); 1553 1554 void CollectInterestingTypesAndFactors(); 1555 void CollectFixupsAndInitialFormulae(); 1556 1557 LSRFixup &getNewFixup() { 1558 Fixups.push_back(LSRFixup()); 1559 return Fixups.back(); 1560 } 1561 1562 // Support for sharing of LSRUses between LSRFixups. 1563 typedef DenseMap<std::pair<const SCEV *, LSRUse::KindType>, 1564 size_t, 1565 UseMapDenseMapInfo> UseMapTy; 1566 UseMapTy UseMap; 1567 1568 bool reconcileNewOffset(LSRUse &LU, int64_t NewOffset, bool HasBaseReg, 1569 LSRUse::KindType Kind, Type *AccessTy); 1570 1571 std::pair<size_t, int64_t> getUse(const SCEV *&Expr, 1572 LSRUse::KindType Kind, 1573 Type *AccessTy); 1574 1575 void DeleteUse(LSRUse &LU, size_t LUIdx); 1576 1577 LSRUse *FindUseWithSimilarFormula(const Formula &F, const LSRUse &OrigLU); 1578 1579 void InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx); 1580 void InsertSupplementalFormula(const SCEV *S, LSRUse &LU, size_t LUIdx); 1581 void CountRegisters(const Formula &F, size_t LUIdx); 1582 bool InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F); 1583 1584 void CollectLoopInvariantFixupsAndFormulae(); 1585 1586 void GenerateReassociations(LSRUse &LU, unsigned LUIdx, Formula Base, 1587 unsigned Depth = 0); 1588 void GenerateCombinations(LSRUse &LU, unsigned LUIdx, Formula Base); 1589 void GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx, Formula Base); 1590 void GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx, Formula Base); 1591 void GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx, Formula Base); 1592 void GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base); 1593 void GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base); 1594 void GenerateCrossUseConstantOffsets(); 1595 void GenerateAllReuseFormulae(); 1596 1597 void FilterOutUndesirableDedicatedRegisters(); 1598 1599 size_t EstimateSearchSpaceComplexity() const; 1600 void NarrowSearchSpaceByDetectingSupersets(); 1601 void NarrowSearchSpaceByCollapsingUnrolledCode(); 1602 void NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters(); 1603 void NarrowSearchSpaceByPickingWinnerRegs(); 1604 void NarrowSearchSpaceUsingHeuristics(); 1605 1606 void SolveRecurse(SmallVectorImpl<const Formula *> &Solution, 1607 Cost &SolutionCost, 1608 SmallVectorImpl<const Formula *> &Workspace, 1609 const Cost &CurCost, 1610 const SmallPtrSet<const SCEV *, 16> &CurRegs, 1611 DenseSet<const SCEV *> &VisitedRegs) const; 1612 void Solve(SmallVectorImpl<const Formula *> &Solution) const; 1613 1614 BasicBlock::iterator 1615 HoistInsertPosition(BasicBlock::iterator IP, 1616 const SmallVectorImpl<Instruction *> &Inputs) const; 1617 BasicBlock::iterator 1618 AdjustInsertPositionForExpand(BasicBlock::iterator IP, 1619 const LSRFixup &LF, 1620 const LSRUse &LU, 1621 SCEVExpander &Rewriter) const; 1622 1623 Value *Expand(const LSRFixup &LF, 1624 const Formula &F, 1625 BasicBlock::iterator IP, 1626 SCEVExpander &Rewriter, 1627 SmallVectorImpl<WeakVH> &DeadInsts) const; 1628 void RewriteForPHI(PHINode *PN, const LSRFixup &LF, 1629 const Formula &F, 1630 SCEVExpander &Rewriter, 1631 SmallVectorImpl<WeakVH> &DeadInsts, 1632 Pass *P) const; 1633 void Rewrite(const LSRFixup &LF, 1634 const Formula &F, 1635 SCEVExpander &Rewriter, 1636 SmallVectorImpl<WeakVH> &DeadInsts, 1637 Pass *P) const; 1638 void ImplementSolution(const SmallVectorImpl<const Formula *> &Solution, 1639 Pass *P); 1640 1641 public: 1642 LSRInstance(Loop *L, Pass *P); 1643 1644 bool getChanged() const { return Changed; } 1645 1646 void print_factors_and_types(raw_ostream &OS) const; 1647 void print_fixups(raw_ostream &OS) const; 1648 void print_uses(raw_ostream &OS) const; 1649 void print(raw_ostream &OS) const; 1650 void dump() const; 1651 }; 1652 1653 } 1654 1655 /// OptimizeShadowIV - If IV is used in a int-to-float cast 1656 /// inside the loop then try to eliminate the cast operation. 1657 void LSRInstance::OptimizeShadowIV() { 1658 const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L); 1659 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount)) 1660 return; 1661 1662 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); 1663 UI != E; /* empty */) { 1664 IVUsers::const_iterator CandidateUI = UI; 1665 ++UI; 1666 Instruction *ShadowUse = CandidateUI->getUser(); 1667 Type *DestTy = NULL; 1668 bool IsSigned = false; 1669 1670 /* If shadow use is a int->float cast then insert a second IV 1671 to eliminate this cast. 1672 1673 for (unsigned i = 0; i < n; ++i) 1674 foo((double)i); 1675 1676 is transformed into 1677 1678 double d = 0.0; 1679 for (unsigned i = 0; i < n; ++i, ++d) 1680 foo(d); 1681 */ 1682 if (UIToFPInst *UCast = dyn_cast<UIToFPInst>(CandidateUI->getUser())) { 1683 IsSigned = false; 1684 DestTy = UCast->getDestTy(); 1685 } 1686 else if (SIToFPInst *SCast = dyn_cast<SIToFPInst>(CandidateUI->getUser())) { 1687 IsSigned = true; 1688 DestTy = SCast->getDestTy(); 1689 } 1690 if (!DestTy) continue; 1691 1692 // If target does not support DestTy natively then do not apply 1693 // this transformation. 1694 if (!TTI.isTypeLegal(DestTy)) continue; 1695 1696 PHINode *PH = dyn_cast<PHINode>(ShadowUse->getOperand(0)); 1697 if (!PH) continue; 1698 if (PH->getNumIncomingValues() != 2) continue; 1699 1700 Type *SrcTy = PH->getType(); 1701 int Mantissa = DestTy->getFPMantissaWidth(); 1702 if (Mantissa == -1) continue; 1703 if ((int)SE.getTypeSizeInBits(SrcTy) > Mantissa) 1704 continue; 1705 1706 unsigned Entry, Latch; 1707 if (PH->getIncomingBlock(0) == L->getLoopPreheader()) { 1708 Entry = 0; 1709 Latch = 1; 1710 } else { 1711 Entry = 1; 1712 Latch = 0; 1713 } 1714 1715 ConstantInt *Init = dyn_cast<ConstantInt>(PH->getIncomingValue(Entry)); 1716 if (!Init) continue; 1717 Constant *NewInit = ConstantFP::get(DestTy, IsSigned ? 1718 (double)Init->getSExtValue() : 1719 (double)Init->getZExtValue()); 1720 1721 BinaryOperator *Incr = 1722 dyn_cast<BinaryOperator>(PH->getIncomingValue(Latch)); 1723 if (!Incr) continue; 1724 if (Incr->getOpcode() != Instruction::Add 1725 && Incr->getOpcode() != Instruction::Sub) 1726 continue; 1727 1728 /* Initialize new IV, double d = 0.0 in above example. */ 1729 ConstantInt *C = NULL; 1730 if (Incr->getOperand(0) == PH) 1731 C = dyn_cast<ConstantInt>(Incr->getOperand(1)); 1732 else if (Incr->getOperand(1) == PH) 1733 C = dyn_cast<ConstantInt>(Incr->getOperand(0)); 1734 else 1735 continue; 1736 1737 if (!C) continue; 1738 1739 // Ignore negative constants, as the code below doesn't handle them 1740 // correctly. TODO: Remove this restriction. 1741 if (!C->getValue().isStrictlyPositive()) continue; 1742 1743 /* Add new PHINode. */ 1744 PHINode *NewPH = PHINode::Create(DestTy, 2, "IV.S.", PH); 1745 1746 /* create new increment. '++d' in above example. */ 1747 Constant *CFP = ConstantFP::get(DestTy, C->getZExtValue()); 1748 BinaryOperator *NewIncr = 1749 BinaryOperator::Create(Incr->getOpcode() == Instruction::Add ? 1750 Instruction::FAdd : Instruction::FSub, 1751 NewPH, CFP, "IV.S.next.", Incr); 1752 1753 NewPH->addIncoming(NewInit, PH->getIncomingBlock(Entry)); 1754 NewPH->addIncoming(NewIncr, PH->getIncomingBlock(Latch)); 1755 1756 /* Remove cast operation */ 1757 ShadowUse->replaceAllUsesWith(NewPH); 1758 ShadowUse->eraseFromParent(); 1759 Changed = true; 1760 break; 1761 } 1762 } 1763 1764 /// FindIVUserForCond - If Cond has an operand that is an expression of an IV, 1765 /// set the IV user and stride information and return true, otherwise return 1766 /// false. 1767 bool LSRInstance::FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse) { 1768 for (IVUsers::iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI) 1769 if (UI->getUser() == Cond) { 1770 // NOTE: we could handle setcc instructions with multiple uses here, but 1771 // InstCombine does it as well for simple uses, it's not clear that it 1772 // occurs enough in real life to handle. 1773 CondUse = UI; 1774 return true; 1775 } 1776 return false; 1777 } 1778 1779 /// OptimizeMax - Rewrite the loop's terminating condition if it uses 1780 /// a max computation. 1781 /// 1782 /// This is a narrow solution to a specific, but acute, problem. For loops 1783 /// like this: 1784 /// 1785 /// i = 0; 1786 /// do { 1787 /// p[i] = 0.0; 1788 /// } while (++i < n); 1789 /// 1790 /// the trip count isn't just 'n', because 'n' might not be positive. And 1791 /// unfortunately this can come up even for loops where the user didn't use 1792 /// a C do-while loop. For example, seemingly well-behaved top-test loops 1793 /// will commonly be lowered like this: 1794 // 1795 /// if (n > 0) { 1796 /// i = 0; 1797 /// do { 1798 /// p[i] = 0.0; 1799 /// } while (++i < n); 1800 /// } 1801 /// 1802 /// and then it's possible for subsequent optimization to obscure the if 1803 /// test in such a way that indvars can't find it. 1804 /// 1805 /// When indvars can't find the if test in loops like this, it creates a 1806 /// max expression, which allows it to give the loop a canonical 1807 /// induction variable: 1808 /// 1809 /// i = 0; 1810 /// max = n < 1 ? 1 : n; 1811 /// do { 1812 /// p[i] = 0.0; 1813 /// } while (++i != max); 1814 /// 1815 /// Canonical induction variables are necessary because the loop passes 1816 /// are designed around them. The most obvious example of this is the 1817 /// LoopInfo analysis, which doesn't remember trip count values. It 1818 /// expects to be able to rediscover the trip count each time it is 1819 /// needed, and it does this using a simple analysis that only succeeds if 1820 /// the loop has a canonical induction variable. 1821 /// 1822 /// However, when it comes time to generate code, the maximum operation 1823 /// can be quite costly, especially if it's inside of an outer loop. 1824 /// 1825 /// This function solves this problem by detecting this type of loop and 1826 /// rewriting their conditions from ICMP_NE back to ICMP_SLT, and deleting 1827 /// the instructions for the maximum computation. 1828 /// 1829 ICmpInst *LSRInstance::OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse) { 1830 // Check that the loop matches the pattern we're looking for. 1831 if (Cond->getPredicate() != CmpInst::ICMP_EQ && 1832 Cond->getPredicate() != CmpInst::ICMP_NE) 1833 return Cond; 1834 1835 SelectInst *Sel = dyn_cast<SelectInst>(Cond->getOperand(1)); 1836 if (!Sel || !Sel->hasOneUse()) return Cond; 1837 1838 const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L); 1839 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount)) 1840 return Cond; 1841 const SCEV *One = SE.getConstant(BackedgeTakenCount->getType(), 1); 1842 1843 // Add one to the backedge-taken count to get the trip count. 1844 const SCEV *IterationCount = SE.getAddExpr(One, BackedgeTakenCount); 1845 if (IterationCount != SE.getSCEV(Sel)) return Cond; 1846 1847 // Check for a max calculation that matches the pattern. There's no check 1848 // for ICMP_ULE here because the comparison would be with zero, which 1849 // isn't interesting. 1850 CmpInst::Predicate Pred = ICmpInst::BAD_ICMP_PREDICATE; 1851 const SCEVNAryExpr *Max = 0; 1852 if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(BackedgeTakenCount)) { 1853 Pred = ICmpInst::ICMP_SLE; 1854 Max = S; 1855 } else if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(IterationCount)) { 1856 Pred = ICmpInst::ICMP_SLT; 1857 Max = S; 1858 } else if (const SCEVUMaxExpr *U = dyn_cast<SCEVUMaxExpr>(IterationCount)) { 1859 Pred = ICmpInst::ICMP_ULT; 1860 Max = U; 1861 } else { 1862 // No match; bail. 1863 return Cond; 1864 } 1865 1866 // To handle a max with more than two operands, this optimization would 1867 // require additional checking and setup. 1868 if (Max->getNumOperands() != 2) 1869 return Cond; 1870 1871 const SCEV *MaxLHS = Max->getOperand(0); 1872 const SCEV *MaxRHS = Max->getOperand(1); 1873 1874 // ScalarEvolution canonicalizes constants to the left. For < and >, look 1875 // for a comparison with 1. For <= and >=, a comparison with zero. 1876 if (!MaxLHS || 1877 (ICmpInst::isTrueWhenEqual(Pred) ? !MaxLHS->isZero() : (MaxLHS != One))) 1878 return Cond; 1879 1880 // Check the relevant induction variable for conformance to 1881 // the pattern. 1882 const SCEV *IV = SE.getSCEV(Cond->getOperand(0)); 1883 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(IV); 1884 if (!AR || !AR->isAffine() || 1885 AR->getStart() != One || 1886 AR->getStepRecurrence(SE) != One) 1887 return Cond; 1888 1889 assert(AR->getLoop() == L && 1890 "Loop condition operand is an addrec in a different loop!"); 1891 1892 // Check the right operand of the select, and remember it, as it will 1893 // be used in the new comparison instruction. 1894 Value *NewRHS = 0; 1895 if (ICmpInst::isTrueWhenEqual(Pred)) { 1896 // Look for n+1, and grab n. 1897 if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(1))) 1898 if (isa<ConstantInt>(BO->getOperand(1)) && 1899 cast<ConstantInt>(BO->getOperand(1))->isOne() && 1900 SE.getSCEV(BO->getOperand(0)) == MaxRHS) 1901 NewRHS = BO->getOperand(0); 1902 if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(2))) 1903 if (isa<ConstantInt>(BO->getOperand(1)) && 1904 cast<ConstantInt>(BO->getOperand(1))->isOne() && 1905 SE.getSCEV(BO->getOperand(0)) == MaxRHS) 1906 NewRHS = BO->getOperand(0); 1907 if (!NewRHS) 1908 return Cond; 1909 } else if (SE.getSCEV(Sel->getOperand(1)) == MaxRHS) 1910 NewRHS = Sel->getOperand(1); 1911 else if (SE.getSCEV(Sel->getOperand(2)) == MaxRHS) 1912 NewRHS = Sel->getOperand(2); 1913 else if (const SCEVUnknown *SU = dyn_cast<SCEVUnknown>(MaxRHS)) 1914 NewRHS = SU->getValue(); 1915 else 1916 // Max doesn't match expected pattern. 1917 return Cond; 1918 1919 // Determine the new comparison opcode. It may be signed or unsigned, 1920 // and the original comparison may be either equality or inequality. 1921 if (Cond->getPredicate() == CmpInst::ICMP_EQ) 1922 Pred = CmpInst::getInversePredicate(Pred); 1923 1924 // Ok, everything looks ok to change the condition into an SLT or SGE and 1925 // delete the max calculation. 1926 ICmpInst *NewCond = 1927 new ICmpInst(Cond, Pred, Cond->getOperand(0), NewRHS, "scmp"); 1928 1929 // Delete the max calculation instructions. 1930 Cond->replaceAllUsesWith(NewCond); 1931 CondUse->setUser(NewCond); 1932 Instruction *Cmp = cast<Instruction>(Sel->getOperand(0)); 1933 Cond->eraseFromParent(); 1934 Sel->eraseFromParent(); 1935 if (Cmp->use_empty()) 1936 Cmp->eraseFromParent(); 1937 return NewCond; 1938 } 1939 1940 /// OptimizeLoopTermCond - Change loop terminating condition to use the 1941 /// postinc iv when possible. 1942 void 1943 LSRInstance::OptimizeLoopTermCond() { 1944 SmallPtrSet<Instruction *, 4> PostIncs; 1945 1946 BasicBlock *LatchBlock = L->getLoopLatch(); 1947 SmallVector<BasicBlock*, 8> ExitingBlocks; 1948 L->getExitingBlocks(ExitingBlocks); 1949 1950 for (unsigned i = 0, e = ExitingBlocks.size(); i != e; ++i) { 1951 BasicBlock *ExitingBlock = ExitingBlocks[i]; 1952 1953 // Get the terminating condition for the loop if possible. If we 1954 // can, we want to change it to use a post-incremented version of its 1955 // induction variable, to allow coalescing the live ranges for the IV into 1956 // one register value. 1957 1958 BranchInst *TermBr = dyn_cast<BranchInst>(ExitingBlock->getTerminator()); 1959 if (!TermBr) 1960 continue; 1961 // FIXME: Overly conservative, termination condition could be an 'or' etc.. 1962 if (TermBr->isUnconditional() || !isa<ICmpInst>(TermBr->getCondition())) 1963 continue; 1964 1965 // Search IVUsesByStride to find Cond's IVUse if there is one. 1966 IVStrideUse *CondUse = 0; 1967 ICmpInst *Cond = cast<ICmpInst>(TermBr->getCondition()); 1968 if (!FindIVUserForCond(Cond, CondUse)) 1969 continue; 1970 1971 // If the trip count is computed in terms of a max (due to ScalarEvolution 1972 // being unable to find a sufficient guard, for example), change the loop 1973 // comparison to use SLT or ULT instead of NE. 1974 // One consequence of doing this now is that it disrupts the count-down 1975 // optimization. That's not always a bad thing though, because in such 1976 // cases it may still be worthwhile to avoid a max. 1977 Cond = OptimizeMax(Cond, CondUse); 1978 1979 // If this exiting block dominates the latch block, it may also use 1980 // the post-inc value if it won't be shared with other uses. 1981 // Check for dominance. 1982 if (!DT.dominates(ExitingBlock, LatchBlock)) 1983 continue; 1984 1985 // Conservatively avoid trying to use the post-inc value in non-latch 1986 // exits if there may be pre-inc users in intervening blocks. 1987 if (LatchBlock != ExitingBlock) 1988 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI) 1989 // Test if the use is reachable from the exiting block. This dominator 1990 // query is a conservative approximation of reachability. 1991 if (&*UI != CondUse && 1992 !DT.properlyDominates(UI->getUser()->getParent(), ExitingBlock)) { 1993 // Conservatively assume there may be reuse if the quotient of their 1994 // strides could be a legal scale. 1995 const SCEV *A = IU.getStride(*CondUse, L); 1996 const SCEV *B = IU.getStride(*UI, L); 1997 if (!A || !B) continue; 1998 if (SE.getTypeSizeInBits(A->getType()) != 1999 SE.getTypeSizeInBits(B->getType())) { 2000 if (SE.getTypeSizeInBits(A->getType()) > 2001 SE.getTypeSizeInBits(B->getType())) 2002 B = SE.getSignExtendExpr(B, A->getType()); 2003 else 2004 A = SE.getSignExtendExpr(A, B->getType()); 2005 } 2006 if (const SCEVConstant *D = 2007 dyn_cast_or_null<SCEVConstant>(getExactSDiv(B, A, SE))) { 2008 const ConstantInt *C = D->getValue(); 2009 // Stride of one or negative one can have reuse with non-addresses. 2010 if (C->isOne() || C->isAllOnesValue()) 2011 goto decline_post_inc; 2012 // Avoid weird situations. 2013 if (C->getValue().getMinSignedBits() >= 64 || 2014 C->getValue().isMinSignedValue()) 2015 goto decline_post_inc; 2016 // Check for possible scaled-address reuse. 2017 Type *AccessTy = getAccessType(UI->getUser()); 2018 int64_t Scale = C->getSExtValue(); 2019 if (TTI.isLegalAddressingMode(AccessTy, /*BaseGV=*/ 0, 2020 /*BaseOffset=*/ 0, 2021 /*HasBaseReg=*/ false, Scale)) 2022 goto decline_post_inc; 2023 Scale = -Scale; 2024 if (TTI.isLegalAddressingMode(AccessTy, /*BaseGV=*/ 0, 2025 /*BaseOffset=*/ 0, 2026 /*HasBaseReg=*/ false, Scale)) 2027 goto decline_post_inc; 2028 } 2029 } 2030 2031 DEBUG(dbgs() << " Change loop exiting icmp to use postinc iv: " 2032 << *Cond << '\n'); 2033 2034 // It's possible for the setcc instruction to be anywhere in the loop, and 2035 // possible for it to have multiple users. If it is not immediately before 2036 // the exiting block branch, move it. 2037 if (&*++BasicBlock::iterator(Cond) != TermBr) { 2038 if (Cond->hasOneUse()) { 2039 Cond->moveBefore(TermBr); 2040 } else { 2041 // Clone the terminating condition and insert into the loopend. 2042 ICmpInst *OldCond = Cond; 2043 Cond = cast<ICmpInst>(Cond->clone()); 2044 Cond->setName(L->getHeader()->getName() + ".termcond"); 2045 ExitingBlock->getInstList().insert(TermBr, Cond); 2046 2047 // Clone the IVUse, as the old use still exists! 2048 CondUse = &IU.AddUser(Cond, CondUse->getOperandValToReplace()); 2049 TermBr->replaceUsesOfWith(OldCond, Cond); 2050 } 2051 } 2052 2053 // If we get to here, we know that we can transform the setcc instruction to 2054 // use the post-incremented version of the IV, allowing us to coalesce the 2055 // live ranges for the IV correctly. 2056 CondUse->transformToPostInc(L); 2057 Changed = true; 2058 2059 PostIncs.insert(Cond); 2060 decline_post_inc:; 2061 } 2062 2063 // Determine an insertion point for the loop induction variable increment. It 2064 // must dominate all the post-inc comparisons we just set up, and it must 2065 // dominate the loop latch edge. 2066 IVIncInsertPos = L->getLoopLatch()->getTerminator(); 2067 for (SmallPtrSet<Instruction *, 4>::const_iterator I = PostIncs.begin(), 2068 E = PostIncs.end(); I != E; ++I) { 2069 BasicBlock *BB = 2070 DT.findNearestCommonDominator(IVIncInsertPos->getParent(), 2071 (*I)->getParent()); 2072 if (BB == (*I)->getParent()) 2073 IVIncInsertPos = *I; 2074 else if (BB != IVIncInsertPos->getParent()) 2075 IVIncInsertPos = BB->getTerminator(); 2076 } 2077 } 2078 2079 /// reconcileNewOffset - Determine if the given use can accommodate a fixup 2080 /// at the given offset and other details. If so, update the use and 2081 /// return true. 2082 bool 2083 LSRInstance::reconcileNewOffset(LSRUse &LU, int64_t NewOffset, bool HasBaseReg, 2084 LSRUse::KindType Kind, Type *AccessTy) { 2085 int64_t NewMinOffset = LU.MinOffset; 2086 int64_t NewMaxOffset = LU.MaxOffset; 2087 Type *NewAccessTy = AccessTy; 2088 2089 // Check for a mismatched kind. It's tempting to collapse mismatched kinds to 2090 // something conservative, however this can pessimize in the case that one of 2091 // the uses will have all its uses outside the loop, for example. 2092 if (LU.Kind != Kind) 2093 return false; 2094 // Conservatively assume HasBaseReg is true for now. 2095 if (NewOffset < LU.MinOffset) { 2096 if (!isAlwaysFoldable(TTI, Kind, AccessTy, /*BaseGV=*/ 0, 2097 LU.MaxOffset - NewOffset, HasBaseReg)) 2098 return false; 2099 NewMinOffset = NewOffset; 2100 } else if (NewOffset > LU.MaxOffset) { 2101 if (!isAlwaysFoldable(TTI, Kind, AccessTy, /*BaseGV=*/ 0, 2102 NewOffset - LU.MinOffset, HasBaseReg)) 2103 return false; 2104 NewMaxOffset = NewOffset; 2105 } 2106 // Check for a mismatched access type, and fall back conservatively as needed. 2107 // TODO: Be less conservative when the type is similar and can use the same 2108 // addressing modes. 2109 if (Kind == LSRUse::Address && AccessTy != LU.AccessTy) 2110 NewAccessTy = Type::getVoidTy(AccessTy->getContext()); 2111 2112 // Update the use. 2113 LU.MinOffset = NewMinOffset; 2114 LU.MaxOffset = NewMaxOffset; 2115 LU.AccessTy = NewAccessTy; 2116 if (NewOffset != LU.Offsets.back()) 2117 LU.Offsets.push_back(NewOffset); 2118 return true; 2119 } 2120 2121 /// getUse - Return an LSRUse index and an offset value for a fixup which 2122 /// needs the given expression, with the given kind and optional access type. 2123 /// Either reuse an existing use or create a new one, as needed. 2124 std::pair<size_t, int64_t> 2125 LSRInstance::getUse(const SCEV *&Expr, 2126 LSRUse::KindType Kind, Type *AccessTy) { 2127 const SCEV *Copy = Expr; 2128 int64_t Offset = ExtractImmediate(Expr, SE); 2129 2130 // Basic uses can't accept any offset, for example. 2131 if (!isAlwaysFoldable(TTI, Kind, AccessTy, /*BaseGV=*/ 0, 2132 Offset, /*HasBaseReg=*/ true)) { 2133 Expr = Copy; 2134 Offset = 0; 2135 } 2136 2137 std::pair<UseMapTy::iterator, bool> P = 2138 UseMap.insert(std::make_pair(std::make_pair(Expr, Kind), 0)); 2139 if (!P.second) { 2140 // A use already existed with this base. 2141 size_t LUIdx = P.first->second; 2142 LSRUse &LU = Uses[LUIdx]; 2143 if (reconcileNewOffset(LU, Offset, /*HasBaseReg=*/true, Kind, AccessTy)) 2144 // Reuse this use. 2145 return std::make_pair(LUIdx, Offset); 2146 } 2147 2148 // Create a new use. 2149 size_t LUIdx = Uses.size(); 2150 P.first->second = LUIdx; 2151 Uses.push_back(LSRUse(Kind, AccessTy)); 2152 LSRUse &LU = Uses[LUIdx]; 2153 2154 // We don't need to track redundant offsets, but we don't need to go out 2155 // of our way here to avoid them. 2156 if (LU.Offsets.empty() || Offset != LU.Offsets.back()) 2157 LU.Offsets.push_back(Offset); 2158 2159 LU.MinOffset = Offset; 2160 LU.MaxOffset = Offset; 2161 return std::make_pair(LUIdx, Offset); 2162 } 2163 2164 /// DeleteUse - Delete the given use from the Uses list. 2165 void LSRInstance::DeleteUse(LSRUse &LU, size_t LUIdx) { 2166 if (&LU != &Uses.back()) 2167 std::swap(LU, Uses.back()); 2168 Uses.pop_back(); 2169 2170 // Update RegUses. 2171 RegUses.SwapAndDropUse(LUIdx, Uses.size()); 2172 } 2173 2174 /// FindUseWithFormula - Look for a use distinct from OrigLU which is has 2175 /// a formula that has the same registers as the given formula. 2176 LSRUse * 2177 LSRInstance::FindUseWithSimilarFormula(const Formula &OrigF, 2178 const LSRUse &OrigLU) { 2179 // Search all uses for the formula. This could be more clever. 2180 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) { 2181 LSRUse &LU = Uses[LUIdx]; 2182 // Check whether this use is close enough to OrigLU, to see whether it's 2183 // worthwhile looking through its formulae. 2184 // Ignore ICmpZero uses because they may contain formulae generated by 2185 // GenerateICmpZeroScales, in which case adding fixup offsets may 2186 // be invalid. 2187 if (&LU != &OrigLU && 2188 LU.Kind != LSRUse::ICmpZero && 2189 LU.Kind == OrigLU.Kind && OrigLU.AccessTy == LU.AccessTy && 2190 LU.WidestFixupType == OrigLU.WidestFixupType && 2191 LU.HasFormulaWithSameRegs(OrigF)) { 2192 // Scan through this use's formulae. 2193 for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(), 2194 E = LU.Formulae.end(); I != E; ++I) { 2195 const Formula &F = *I; 2196 // Check to see if this formula has the same registers and symbols 2197 // as OrigF. 2198 if (F.BaseRegs == OrigF.BaseRegs && 2199 F.ScaledReg == OrigF.ScaledReg && 2200 F.BaseGV == OrigF.BaseGV && 2201 F.Scale == OrigF.Scale && 2202 F.UnfoldedOffset == OrigF.UnfoldedOffset) { 2203 if (F.BaseOffset == 0) 2204 return &LU; 2205 // This is the formula where all the registers and symbols matched; 2206 // there aren't going to be any others. Since we declined it, we 2207 // can skip the rest of the formulae and proceed to the next LSRUse. 2208 break; 2209 } 2210 } 2211 } 2212 } 2213 2214 // Nothing looked good. 2215 return 0; 2216 } 2217 2218 void LSRInstance::CollectInterestingTypesAndFactors() { 2219 SmallSetVector<const SCEV *, 4> Strides; 2220 2221 // Collect interesting types and strides. 2222 SmallVector<const SCEV *, 4> Worklist; 2223 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI) { 2224 const SCEV *Expr = IU.getExpr(*UI); 2225 2226 // Collect interesting types. 2227 Types.insert(SE.getEffectiveSCEVType(Expr->getType())); 2228 2229 // Add strides for mentioned loops. 2230 Worklist.push_back(Expr); 2231 do { 2232 const SCEV *S = Worklist.pop_back_val(); 2233 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) { 2234 if (AR->getLoop() == L) 2235 Strides.insert(AR->getStepRecurrence(SE)); 2236 Worklist.push_back(AR->getStart()); 2237 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) { 2238 Worklist.append(Add->op_begin(), Add->op_end()); 2239 } 2240 } while (!Worklist.empty()); 2241 } 2242 2243 // Compute interesting factors from the set of interesting strides. 2244 for (SmallSetVector<const SCEV *, 4>::const_iterator 2245 I = Strides.begin(), E = Strides.end(); I != E; ++I) 2246 for (SmallSetVector<const SCEV *, 4>::const_iterator NewStrideIter = 2247 llvm::next(I); NewStrideIter != E; ++NewStrideIter) { 2248 const SCEV *OldStride = *I; 2249 const SCEV *NewStride = *NewStrideIter; 2250 2251 if (SE.getTypeSizeInBits(OldStride->getType()) != 2252 SE.getTypeSizeInBits(NewStride->getType())) { 2253 if (SE.getTypeSizeInBits(OldStride->getType()) > 2254 SE.getTypeSizeInBits(NewStride->getType())) 2255 NewStride = SE.getSignExtendExpr(NewStride, OldStride->getType()); 2256 else 2257 OldStride = SE.getSignExtendExpr(OldStride, NewStride->getType()); 2258 } 2259 if (const SCEVConstant *Factor = 2260 dyn_cast_or_null<SCEVConstant>(getExactSDiv(NewStride, OldStride, 2261 SE, true))) { 2262 if (Factor->getValue()->getValue().getMinSignedBits() <= 64) 2263 Factors.insert(Factor->getValue()->getValue().getSExtValue()); 2264 } else if (const SCEVConstant *Factor = 2265 dyn_cast_or_null<SCEVConstant>(getExactSDiv(OldStride, 2266 NewStride, 2267 SE, true))) { 2268 if (Factor->getValue()->getValue().getMinSignedBits() <= 64) 2269 Factors.insert(Factor->getValue()->getValue().getSExtValue()); 2270 } 2271 } 2272 2273 // If all uses use the same type, don't bother looking for truncation-based 2274 // reuse. 2275 if (Types.size() == 1) 2276 Types.clear(); 2277 2278 DEBUG(print_factors_and_types(dbgs())); 2279 } 2280 2281 /// findIVOperand - Helper for CollectChains that finds an IV operand (computed 2282 /// by an AddRec in this loop) within [OI,OE) or returns OE. If IVUsers mapped 2283 /// Instructions to IVStrideUses, we could partially skip this. 2284 static User::op_iterator 2285 findIVOperand(User::op_iterator OI, User::op_iterator OE, 2286 Loop *L, ScalarEvolution &SE) { 2287 for(; OI != OE; ++OI) { 2288 if (Instruction *Oper = dyn_cast<Instruction>(*OI)) { 2289 if (!SE.isSCEVable(Oper->getType())) 2290 continue; 2291 2292 if (const SCEVAddRecExpr *AR = 2293 dyn_cast<SCEVAddRecExpr>(SE.getSCEV(Oper))) { 2294 if (AR->getLoop() == L) 2295 break; 2296 } 2297 } 2298 } 2299 return OI; 2300 } 2301 2302 /// getWideOperand - IVChain logic must consistenctly peek base TruncInst 2303 /// operands, so wrap it in a convenient helper. 2304 static Value *getWideOperand(Value *Oper) { 2305 if (TruncInst *Trunc = dyn_cast<TruncInst>(Oper)) 2306 return Trunc->getOperand(0); 2307 return Oper; 2308 } 2309 2310 /// isCompatibleIVType - Return true if we allow an IV chain to include both 2311 /// types. 2312 static bool isCompatibleIVType(Value *LVal, Value *RVal) { 2313 Type *LType = LVal->getType(); 2314 Type *RType = RVal->getType(); 2315 return (LType == RType) || (LType->isPointerTy() && RType->isPointerTy()); 2316 } 2317 2318 /// getExprBase - Return an approximation of this SCEV expression's "base", or 2319 /// NULL for any constant. Returning the expression itself is 2320 /// conservative. Returning a deeper subexpression is more precise and valid as 2321 /// long as it isn't less complex than another subexpression. For expressions 2322 /// involving multiple unscaled values, we need to return the pointer-type 2323 /// SCEVUnknown. This avoids forming chains across objects, such as: 2324 /// PrevOper==a[i], IVOper==b[i], IVInc==b-a. 2325 /// 2326 /// Since SCEVUnknown is the rightmost type, and pointers are the rightmost 2327 /// SCEVUnknown, we simply return the rightmost SCEV operand. 2328 static const SCEV *getExprBase(const SCEV *S) { 2329 switch (S->getSCEVType()) { 2330 default: // uncluding scUnknown. 2331 return S; 2332 case scConstant: 2333 return 0; 2334 case scTruncate: 2335 return getExprBase(cast<SCEVTruncateExpr>(S)->getOperand()); 2336 case scZeroExtend: 2337 return getExprBase(cast<SCEVZeroExtendExpr>(S)->getOperand()); 2338 case scSignExtend: 2339 return getExprBase(cast<SCEVSignExtendExpr>(S)->getOperand()); 2340 case scAddExpr: { 2341 // Skip over scaled operands (scMulExpr) to follow add operands as long as 2342 // there's nothing more complex. 2343 // FIXME: not sure if we want to recognize negation. 2344 const SCEVAddExpr *Add = cast<SCEVAddExpr>(S); 2345 for (std::reverse_iterator<SCEVAddExpr::op_iterator> I(Add->op_end()), 2346 E(Add->op_begin()); I != E; ++I) { 2347 const SCEV *SubExpr = *I; 2348 if (SubExpr->getSCEVType() == scAddExpr) 2349 return getExprBase(SubExpr); 2350 2351 if (SubExpr->getSCEVType() != scMulExpr) 2352 return SubExpr; 2353 } 2354 return S; // all operands are scaled, be conservative. 2355 } 2356 case scAddRecExpr: 2357 return getExprBase(cast<SCEVAddRecExpr>(S)->getStart()); 2358 } 2359 } 2360 2361 /// Return true if the chain increment is profitable to expand into a loop 2362 /// invariant value, which may require its own register. A profitable chain 2363 /// increment will be an offset relative to the same base. We allow such offsets 2364 /// to potentially be used as chain increment as long as it's not obviously 2365 /// expensive to expand using real instructions. 2366 bool IVChain::isProfitableIncrement(const SCEV *OperExpr, 2367 const SCEV *IncExpr, 2368 ScalarEvolution &SE) { 2369 // Aggressively form chains when -stress-ivchain. 2370 if (StressIVChain) 2371 return true; 2372 2373 // Do not replace a constant offset from IV head with a nonconstant IV 2374 // increment. 2375 if (!isa<SCEVConstant>(IncExpr)) { 2376 const SCEV *HeadExpr = SE.getSCEV(getWideOperand(Incs[0].IVOperand)); 2377 if (isa<SCEVConstant>(SE.getMinusSCEV(OperExpr, HeadExpr))) 2378 return 0; 2379 } 2380 2381 SmallPtrSet<const SCEV*, 8> Processed; 2382 return !isHighCostExpansion(IncExpr, Processed, SE); 2383 } 2384 2385 /// Return true if the number of registers needed for the chain is estimated to 2386 /// be less than the number required for the individual IV users. First prohibit 2387 /// any IV users that keep the IV live across increments (the Users set should 2388 /// be empty). Next count the number and type of increments in the chain. 2389 /// 2390 /// Chaining IVs can lead to considerable code bloat if ISEL doesn't 2391 /// effectively use postinc addressing modes. Only consider it profitable it the 2392 /// increments can be computed in fewer registers when chained. 2393 /// 2394 /// TODO: Consider IVInc free if it's already used in another chains. 2395 static bool 2396 isProfitableChain(IVChain &Chain, SmallPtrSet<Instruction*, 4> &Users, 2397 ScalarEvolution &SE, const TargetTransformInfo &TTI) { 2398 if (StressIVChain) 2399 return true; 2400 2401 if (!Chain.hasIncs()) 2402 return false; 2403 2404 if (!Users.empty()) { 2405 DEBUG(dbgs() << "Chain: " << *Chain.Incs[0].UserInst << " users:\n"; 2406 for (SmallPtrSet<Instruction*, 4>::const_iterator I = Users.begin(), 2407 E = Users.end(); I != E; ++I) { 2408 dbgs() << " " << **I << "\n"; 2409 }); 2410 return false; 2411 } 2412 assert(!Chain.Incs.empty() && "empty IV chains are not allowed"); 2413 2414 // The chain itself may require a register, so intialize cost to 1. 2415 int cost = 1; 2416 2417 // A complete chain likely eliminates the need for keeping the original IV in 2418 // a register. LSR does not currently know how to form a complete chain unless 2419 // the header phi already exists. 2420 if (isa<PHINode>(Chain.tailUserInst()) 2421 && SE.getSCEV(Chain.tailUserInst()) == Chain.Incs[0].IncExpr) { 2422 --cost; 2423 } 2424 const SCEV *LastIncExpr = 0; 2425 unsigned NumConstIncrements = 0; 2426 unsigned NumVarIncrements = 0; 2427 unsigned NumReusedIncrements = 0; 2428 for (IVChain::const_iterator I = Chain.begin(), E = Chain.end(); 2429 I != E; ++I) { 2430 2431 if (I->IncExpr->isZero()) 2432 continue; 2433 2434 // Incrementing by zero or some constant is neutral. We assume constants can 2435 // be folded into an addressing mode or an add's immediate operand. 2436 if (isa<SCEVConstant>(I->IncExpr)) { 2437 ++NumConstIncrements; 2438 continue; 2439 } 2440 2441 if (I->IncExpr == LastIncExpr) 2442 ++NumReusedIncrements; 2443 else 2444 ++NumVarIncrements; 2445 2446 LastIncExpr = I->IncExpr; 2447 } 2448 // An IV chain with a single increment is handled by LSR's postinc 2449 // uses. However, a chain with multiple increments requires keeping the IV's 2450 // value live longer than it needs to be if chained. 2451 if (NumConstIncrements > 1) 2452 --cost; 2453 2454 // Materializing increment expressions in the preheader that didn't exist in 2455 // the original code may cost a register. For example, sign-extended array 2456 // indices can produce ridiculous increments like this: 2457 // IV + ((sext i32 (2 * %s) to i64) + (-1 * (sext i32 %s to i64))) 2458 cost += NumVarIncrements; 2459 2460 // Reusing variable increments likely saves a register to hold the multiple of 2461 // the stride. 2462 cost -= NumReusedIncrements; 2463 2464 DEBUG(dbgs() << "Chain: " << *Chain.Incs[0].UserInst << " Cost: " << cost 2465 << "\n"); 2466 2467 return cost < 0; 2468 } 2469 2470 /// ChainInstruction - Add this IV user to an existing chain or make it the head 2471 /// of a new chain. 2472 void LSRInstance::ChainInstruction(Instruction *UserInst, Instruction *IVOper, 2473 SmallVectorImpl<ChainUsers> &ChainUsersVec) { 2474 // When IVs are used as types of varying widths, they are generally converted 2475 // to a wider type with some uses remaining narrow under a (free) trunc. 2476 Value *const NextIV = getWideOperand(IVOper); 2477 const SCEV *const OperExpr = SE.getSCEV(NextIV); 2478 const SCEV *const OperExprBase = getExprBase(OperExpr); 2479 2480 // Visit all existing chains. Check if its IVOper can be computed as a 2481 // profitable loop invariant increment from the last link in the Chain. 2482 unsigned ChainIdx = 0, NChains = IVChainVec.size(); 2483 const SCEV *LastIncExpr = 0; 2484 for (; ChainIdx < NChains; ++ChainIdx) { 2485 IVChain &Chain = IVChainVec[ChainIdx]; 2486 2487 // Prune the solution space aggressively by checking that both IV operands 2488 // are expressions that operate on the same unscaled SCEVUnknown. This 2489 // "base" will be canceled by the subsequent getMinusSCEV call. Checking 2490 // first avoids creating extra SCEV expressions. 2491 if (!StressIVChain && Chain.ExprBase != OperExprBase) 2492 continue; 2493 2494 Value *PrevIV = getWideOperand(Chain.Incs.back().IVOperand); 2495 if (!isCompatibleIVType(PrevIV, NextIV)) 2496 continue; 2497 2498 // A phi node terminates a chain. 2499 if (isa<PHINode>(UserInst) && isa<PHINode>(Chain.tailUserInst())) 2500 continue; 2501 2502 // The increment must be loop-invariant so it can be kept in a register. 2503 const SCEV *PrevExpr = SE.getSCEV(PrevIV); 2504 const SCEV *IncExpr = SE.getMinusSCEV(OperExpr, PrevExpr); 2505 if (!SE.isLoopInvariant(IncExpr, L)) 2506 continue; 2507 2508 if (Chain.isProfitableIncrement(OperExpr, IncExpr, SE)) { 2509 LastIncExpr = IncExpr; 2510 break; 2511 } 2512 } 2513 // If we haven't found a chain, create a new one, unless we hit the max. Don't 2514 // bother for phi nodes, because they must be last in the chain. 2515 if (ChainIdx == NChains) { 2516 if (isa<PHINode>(UserInst)) 2517 return; 2518 if (NChains >= MaxChains && !StressIVChain) { 2519 DEBUG(dbgs() << "IV Chain Limit\n"); 2520 return; 2521 } 2522 LastIncExpr = OperExpr; 2523 // IVUsers may have skipped over sign/zero extensions. We don't currently 2524 // attempt to form chains involving extensions unless they can be hoisted 2525 // into this loop's AddRec. 2526 if (!isa<SCEVAddRecExpr>(LastIncExpr)) 2527 return; 2528 ++NChains; 2529 IVChainVec.push_back(IVChain(IVInc(UserInst, IVOper, LastIncExpr), 2530 OperExprBase)); 2531 ChainUsersVec.resize(NChains); 2532 DEBUG(dbgs() << "IV Chain#" << ChainIdx << " Head: (" << *UserInst 2533 << ") IV=" << *LastIncExpr << "\n"); 2534 } else { 2535 DEBUG(dbgs() << "IV Chain#" << ChainIdx << " Inc: (" << *UserInst 2536 << ") IV+" << *LastIncExpr << "\n"); 2537 // Add this IV user to the end of the chain. 2538 IVChainVec[ChainIdx].add(IVInc(UserInst, IVOper, LastIncExpr)); 2539 } 2540 IVChain &Chain = IVChainVec[ChainIdx]; 2541 2542 SmallPtrSet<Instruction*,4> &NearUsers = ChainUsersVec[ChainIdx].NearUsers; 2543 // This chain's NearUsers become FarUsers. 2544 if (!LastIncExpr->isZero()) { 2545 ChainUsersVec[ChainIdx].FarUsers.insert(NearUsers.begin(), 2546 NearUsers.end()); 2547 NearUsers.clear(); 2548 } 2549 2550 // All other uses of IVOperand become near uses of the chain. 2551 // We currently ignore intermediate values within SCEV expressions, assuming 2552 // they will eventually be used be the current chain, or can be computed 2553 // from one of the chain increments. To be more precise we could 2554 // transitively follow its user and only add leaf IV users to the set. 2555 for (Value::use_iterator UseIter = IVOper->use_begin(), 2556 UseEnd = IVOper->use_end(); UseIter != UseEnd; ++UseIter) { 2557 Instruction *OtherUse = dyn_cast<Instruction>(*UseIter); 2558 if (!OtherUse) 2559 continue; 2560 // Uses in the chain will no longer be uses if the chain is formed. 2561 // Include the head of the chain in this iteration (not Chain.begin()). 2562 IVChain::const_iterator IncIter = Chain.Incs.begin(); 2563 IVChain::const_iterator IncEnd = Chain.Incs.end(); 2564 for( ; IncIter != IncEnd; ++IncIter) { 2565 if (IncIter->UserInst == OtherUse) 2566 break; 2567 } 2568 if (IncIter != IncEnd) 2569 continue; 2570 2571 if (SE.isSCEVable(OtherUse->getType()) 2572 && !isa<SCEVUnknown>(SE.getSCEV(OtherUse)) 2573 && IU.isIVUserOrOperand(OtherUse)) { 2574 continue; 2575 } 2576 NearUsers.insert(OtherUse); 2577 } 2578 2579 // Since this user is part of the chain, it's no longer considered a use 2580 // of the chain. 2581 ChainUsersVec[ChainIdx].FarUsers.erase(UserInst); 2582 } 2583 2584 /// CollectChains - Populate the vector of Chains. 2585 /// 2586 /// This decreases ILP at the architecture level. Targets with ample registers, 2587 /// multiple memory ports, and no register renaming probably don't want 2588 /// this. However, such targets should probably disable LSR altogether. 2589 /// 2590 /// The job of LSR is to make a reasonable choice of induction variables across 2591 /// the loop. Subsequent passes can easily "unchain" computation exposing more 2592 /// ILP *within the loop* if the target wants it. 2593 /// 2594 /// Finding the best IV chain is potentially a scheduling problem. Since LSR 2595 /// will not reorder memory operations, it will recognize this as a chain, but 2596 /// will generate redundant IV increments. Ideally this would be corrected later 2597 /// by a smart scheduler: 2598 /// = A[i] 2599 /// = A[i+x] 2600 /// A[i] = 2601 /// A[i+x] = 2602 /// 2603 /// TODO: Walk the entire domtree within this loop, not just the path to the 2604 /// loop latch. This will discover chains on side paths, but requires 2605 /// maintaining multiple copies of the Chains state. 2606 void LSRInstance::CollectChains() { 2607 DEBUG(dbgs() << "Collecting IV Chains.\n"); 2608 SmallVector<ChainUsers, 8> ChainUsersVec; 2609 2610 SmallVector<BasicBlock *,8> LatchPath; 2611 BasicBlock *LoopHeader = L->getHeader(); 2612 for (DomTreeNode *Rung = DT.getNode(L->getLoopLatch()); 2613 Rung->getBlock() != LoopHeader; Rung = Rung->getIDom()) { 2614 LatchPath.push_back(Rung->getBlock()); 2615 } 2616 LatchPath.push_back(LoopHeader); 2617 2618 // Walk the instruction stream from the loop header to the loop latch. 2619 for (SmallVectorImpl<BasicBlock *>::reverse_iterator 2620 BBIter = LatchPath.rbegin(), BBEnd = LatchPath.rend(); 2621 BBIter != BBEnd; ++BBIter) { 2622 for (BasicBlock::iterator I = (*BBIter)->begin(), E = (*BBIter)->end(); 2623 I != E; ++I) { 2624 // Skip instructions that weren't seen by IVUsers analysis. 2625 if (isa<PHINode>(I) || !IU.isIVUserOrOperand(I)) 2626 continue; 2627 2628 // Ignore users that are part of a SCEV expression. This way we only 2629 // consider leaf IV Users. This effectively rediscovers a portion of 2630 // IVUsers analysis but in program order this time. 2631 if (SE.isSCEVable(I->getType()) && !isa<SCEVUnknown>(SE.getSCEV(I))) 2632 continue; 2633 2634 // Remove this instruction from any NearUsers set it may be in. 2635 for (unsigned ChainIdx = 0, NChains = IVChainVec.size(); 2636 ChainIdx < NChains; ++ChainIdx) { 2637 ChainUsersVec[ChainIdx].NearUsers.erase(I); 2638 } 2639 // Search for operands that can be chained. 2640 SmallPtrSet<Instruction*, 4> UniqueOperands; 2641 User::op_iterator IVOpEnd = I->op_end(); 2642 User::op_iterator IVOpIter = findIVOperand(I->op_begin(), IVOpEnd, L, SE); 2643 while (IVOpIter != IVOpEnd) { 2644 Instruction *IVOpInst = cast<Instruction>(*IVOpIter); 2645 if (UniqueOperands.insert(IVOpInst)) 2646 ChainInstruction(I, IVOpInst, ChainUsersVec); 2647 IVOpIter = findIVOperand(llvm::next(IVOpIter), IVOpEnd, L, SE); 2648 } 2649 } // Continue walking down the instructions. 2650 } // Continue walking down the domtree. 2651 // Visit phi backedges to determine if the chain can generate the IV postinc. 2652 for (BasicBlock::iterator I = L->getHeader()->begin(); 2653 PHINode *PN = dyn_cast<PHINode>(I); ++I) { 2654 if (!SE.isSCEVable(PN->getType())) 2655 continue; 2656 2657 Instruction *IncV = 2658 dyn_cast<Instruction>(PN->getIncomingValueForBlock(L->getLoopLatch())); 2659 if (IncV) 2660 ChainInstruction(PN, IncV, ChainUsersVec); 2661 } 2662 // Remove any unprofitable chains. 2663 unsigned ChainIdx = 0; 2664 for (unsigned UsersIdx = 0, NChains = IVChainVec.size(); 2665 UsersIdx < NChains; ++UsersIdx) { 2666 if (!isProfitableChain(IVChainVec[UsersIdx], 2667 ChainUsersVec[UsersIdx].FarUsers, SE, TTI)) 2668 continue; 2669 // Preserve the chain at UsesIdx. 2670 if (ChainIdx != UsersIdx) 2671 IVChainVec[ChainIdx] = IVChainVec[UsersIdx]; 2672 FinalizeChain(IVChainVec[ChainIdx]); 2673 ++ChainIdx; 2674 } 2675 IVChainVec.resize(ChainIdx); 2676 } 2677 2678 void LSRInstance::FinalizeChain(IVChain &Chain) { 2679 assert(!Chain.Incs.empty() && "empty IV chains are not allowed"); 2680 DEBUG(dbgs() << "Final Chain: " << *Chain.Incs[0].UserInst << "\n"); 2681 2682 for (IVChain::const_iterator I = Chain.begin(), E = Chain.end(); 2683 I != E; ++I) { 2684 DEBUG(dbgs() << " Inc: " << *I->UserInst << "\n"); 2685 User::op_iterator UseI = 2686 std::find(I->UserInst->op_begin(), I->UserInst->op_end(), I->IVOperand); 2687 assert(UseI != I->UserInst->op_end() && "cannot find IV operand"); 2688 IVIncSet.insert(UseI); 2689 } 2690 } 2691 2692 /// Return true if the IVInc can be folded into an addressing mode. 2693 static bool canFoldIVIncExpr(const SCEV *IncExpr, Instruction *UserInst, 2694 Value *Operand, const TargetTransformInfo &TTI) { 2695 const SCEVConstant *IncConst = dyn_cast<SCEVConstant>(IncExpr); 2696 if (!IncConst || !isAddressUse(UserInst, Operand)) 2697 return false; 2698 2699 if (IncConst->getValue()->getValue().getMinSignedBits() > 64) 2700 return false; 2701 2702 int64_t IncOffset = IncConst->getValue()->getSExtValue(); 2703 if (!isAlwaysFoldable(TTI, LSRUse::Address, 2704 getAccessType(UserInst), /*BaseGV=*/ 0, 2705 IncOffset, /*HaseBaseReg=*/ false)) 2706 return false; 2707 2708 return true; 2709 } 2710 2711 /// GenerateIVChains - Generate an add or subtract for each IVInc in a chain to 2712 /// materialize the IV user's operand from the previous IV user's operand. 2713 void LSRInstance::GenerateIVChain(const IVChain &Chain, SCEVExpander &Rewriter, 2714 SmallVectorImpl<WeakVH> &DeadInsts) { 2715 // Find the new IVOperand for the head of the chain. It may have been replaced 2716 // by LSR. 2717 const IVInc &Head = Chain.Incs[0]; 2718 User::op_iterator IVOpEnd = Head.UserInst->op_end(); 2719 User::op_iterator IVOpIter = findIVOperand(Head.UserInst->op_begin(), 2720 IVOpEnd, L, SE); 2721 Value *IVSrc = 0; 2722 while (IVOpIter != IVOpEnd) { 2723 IVSrc = getWideOperand(*IVOpIter); 2724 2725 // If this operand computes the expression that the chain needs, we may use 2726 // it. (Check this after setting IVSrc which is used below.) 2727 // 2728 // Note that if Head.IncExpr is wider than IVSrc, then this phi is too 2729 // narrow for the chain, so we can no longer use it. We do allow using a 2730 // wider phi, assuming the LSR checked for free truncation. In that case we 2731 // should already have a truncate on this operand such that 2732 // getSCEV(IVSrc) == IncExpr. 2733 if (SE.getSCEV(*IVOpIter) == Head.IncExpr 2734 || SE.getSCEV(IVSrc) == Head.IncExpr) { 2735 break; 2736 } 2737 IVOpIter = findIVOperand(llvm::next(IVOpIter), IVOpEnd, L, SE); 2738 } 2739 if (IVOpIter == IVOpEnd) { 2740 // Gracefully give up on this chain. 2741 DEBUG(dbgs() << "Concealed chain head: " << *Head.UserInst << "\n"); 2742 return; 2743 } 2744 2745 DEBUG(dbgs() << "Generate chain at: " << *IVSrc << "\n"); 2746 Type *IVTy = IVSrc->getType(); 2747 Type *IntTy = SE.getEffectiveSCEVType(IVTy); 2748 const SCEV *LeftOverExpr = 0; 2749 for (IVChain::const_iterator IncI = Chain.begin(), 2750 IncE = Chain.end(); IncI != IncE; ++IncI) { 2751 2752 Instruction *InsertPt = IncI->UserInst; 2753 if (isa<PHINode>(InsertPt)) 2754 InsertPt = L->getLoopLatch()->getTerminator(); 2755 2756 // IVOper will replace the current IV User's operand. IVSrc is the IV 2757 // value currently held in a register. 2758 Value *IVOper = IVSrc; 2759 if (!IncI->IncExpr->isZero()) { 2760 // IncExpr was the result of subtraction of two narrow values, so must 2761 // be signed. 2762 const SCEV *IncExpr = SE.getNoopOrSignExtend(IncI->IncExpr, IntTy); 2763 LeftOverExpr = LeftOverExpr ? 2764 SE.getAddExpr(LeftOverExpr, IncExpr) : IncExpr; 2765 } 2766 if (LeftOverExpr && !LeftOverExpr->isZero()) { 2767 // Expand the IV increment. 2768 Rewriter.clearPostInc(); 2769 Value *IncV = Rewriter.expandCodeFor(LeftOverExpr, IntTy, InsertPt); 2770 const SCEV *IVOperExpr = SE.getAddExpr(SE.getUnknown(IVSrc), 2771 SE.getUnknown(IncV)); 2772 IVOper = Rewriter.expandCodeFor(IVOperExpr, IVTy, InsertPt); 2773 2774 // If an IV increment can't be folded, use it as the next IV value. 2775 if (!canFoldIVIncExpr(LeftOverExpr, IncI->UserInst, IncI->IVOperand, 2776 TTI)) { 2777 assert(IVTy == IVOper->getType() && "inconsistent IV increment type"); 2778 IVSrc = IVOper; 2779 LeftOverExpr = 0; 2780 } 2781 } 2782 Type *OperTy = IncI->IVOperand->getType(); 2783 if (IVTy != OperTy) { 2784 assert(SE.getTypeSizeInBits(IVTy) >= SE.getTypeSizeInBits(OperTy) && 2785 "cannot extend a chained IV"); 2786 IRBuilder<> Builder(InsertPt); 2787 IVOper = Builder.CreateTruncOrBitCast(IVOper, OperTy, "lsr.chain"); 2788 } 2789 IncI->UserInst->replaceUsesOfWith(IncI->IVOperand, IVOper); 2790 DeadInsts.push_back(IncI->IVOperand); 2791 } 2792 // If LSR created a new, wider phi, we may also replace its postinc. We only 2793 // do this if we also found a wide value for the head of the chain. 2794 if (isa<PHINode>(Chain.tailUserInst())) { 2795 for (BasicBlock::iterator I = L->getHeader()->begin(); 2796 PHINode *Phi = dyn_cast<PHINode>(I); ++I) { 2797 if (!isCompatibleIVType(Phi, IVSrc)) 2798 continue; 2799 Instruction *PostIncV = dyn_cast<Instruction>( 2800 Phi->getIncomingValueForBlock(L->getLoopLatch())); 2801 if (!PostIncV || (SE.getSCEV(PostIncV) != SE.getSCEV(IVSrc))) 2802 continue; 2803 Value *IVOper = IVSrc; 2804 Type *PostIncTy = PostIncV->getType(); 2805 if (IVTy != PostIncTy) { 2806 assert(PostIncTy->isPointerTy() && "mixing int/ptr IV types"); 2807 IRBuilder<> Builder(L->getLoopLatch()->getTerminator()); 2808 Builder.SetCurrentDebugLocation(PostIncV->getDebugLoc()); 2809 IVOper = Builder.CreatePointerCast(IVSrc, PostIncTy, "lsr.chain"); 2810 } 2811 Phi->replaceUsesOfWith(PostIncV, IVOper); 2812 DeadInsts.push_back(PostIncV); 2813 } 2814 } 2815 } 2816 2817 void LSRInstance::CollectFixupsAndInitialFormulae() { 2818 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI) { 2819 Instruction *UserInst = UI->getUser(); 2820 // Skip IV users that are part of profitable IV Chains. 2821 User::op_iterator UseI = std::find(UserInst->op_begin(), UserInst->op_end(), 2822 UI->getOperandValToReplace()); 2823 assert(UseI != UserInst->op_end() && "cannot find IV operand"); 2824 if (IVIncSet.count(UseI)) 2825 continue; 2826 2827 // Record the uses. 2828 LSRFixup &LF = getNewFixup(); 2829 LF.UserInst = UserInst; 2830 LF.OperandValToReplace = UI->getOperandValToReplace(); 2831 LF.PostIncLoops = UI->getPostIncLoops(); 2832 2833 LSRUse::KindType Kind = LSRUse::Basic; 2834 Type *AccessTy = 0; 2835 if (isAddressUse(LF.UserInst, LF.OperandValToReplace)) { 2836 Kind = LSRUse::Address; 2837 AccessTy = getAccessType(LF.UserInst); 2838 } 2839 2840 const SCEV *S = IU.getExpr(*UI); 2841 2842 // Equality (== and !=) ICmps are special. We can rewrite (i == N) as 2843 // (N - i == 0), and this allows (N - i) to be the expression that we work 2844 // with rather than just N or i, so we can consider the register 2845 // requirements for both N and i at the same time. Limiting this code to 2846 // equality icmps is not a problem because all interesting loops use 2847 // equality icmps, thanks to IndVarSimplify. 2848 if (ICmpInst *CI = dyn_cast<ICmpInst>(LF.UserInst)) 2849 if (CI->isEquality()) { 2850 // Swap the operands if needed to put the OperandValToReplace on the 2851 // left, for consistency. 2852 Value *NV = CI->getOperand(1); 2853 if (NV == LF.OperandValToReplace) { 2854 CI->setOperand(1, CI->getOperand(0)); 2855 CI->setOperand(0, NV); 2856 NV = CI->getOperand(1); 2857 Changed = true; 2858 } 2859 2860 // x == y --> x - y == 0 2861 const SCEV *N = SE.getSCEV(NV); 2862 if (SE.isLoopInvariant(N, L) && isSafeToExpand(N)) { 2863 // S is normalized, so normalize N before folding it into S 2864 // to keep the result normalized. 2865 N = TransformForPostIncUse(Normalize, N, CI, 0, 2866 LF.PostIncLoops, SE, DT); 2867 Kind = LSRUse::ICmpZero; 2868 S = SE.getMinusSCEV(N, S); 2869 } 2870 2871 // -1 and the negations of all interesting strides (except the negation 2872 // of -1) are now also interesting. 2873 for (size_t i = 0, e = Factors.size(); i != e; ++i) 2874 if (Factors[i] != -1) 2875 Factors.insert(-(uint64_t)Factors[i]); 2876 Factors.insert(-1); 2877 } 2878 2879 // Set up the initial formula for this use. 2880 std::pair<size_t, int64_t> P = getUse(S, Kind, AccessTy); 2881 LF.LUIdx = P.first; 2882 LF.Offset = P.second; 2883 LSRUse &LU = Uses[LF.LUIdx]; 2884 LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L); 2885 if (!LU.WidestFixupType || 2886 SE.getTypeSizeInBits(LU.WidestFixupType) < 2887 SE.getTypeSizeInBits(LF.OperandValToReplace->getType())) 2888 LU.WidestFixupType = LF.OperandValToReplace->getType(); 2889 2890 // If this is the first use of this LSRUse, give it a formula. 2891 if (LU.Formulae.empty()) { 2892 InsertInitialFormula(S, LU, LF.LUIdx); 2893 CountRegisters(LU.Formulae.back(), LF.LUIdx); 2894 } 2895 } 2896 2897 DEBUG(print_fixups(dbgs())); 2898 } 2899 2900 /// InsertInitialFormula - Insert a formula for the given expression into 2901 /// the given use, separating out loop-variant portions from loop-invariant 2902 /// and loop-computable portions. 2903 void 2904 LSRInstance::InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx) { 2905 Formula F; 2906 F.InitialMatch(S, L, SE); 2907 bool Inserted = InsertFormula(LU, LUIdx, F); 2908 assert(Inserted && "Initial formula already exists!"); (void)Inserted; 2909 } 2910 2911 /// InsertSupplementalFormula - Insert a simple single-register formula for 2912 /// the given expression into the given use. 2913 void 2914 LSRInstance::InsertSupplementalFormula(const SCEV *S, 2915 LSRUse &LU, size_t LUIdx) { 2916 Formula F; 2917 F.BaseRegs.push_back(S); 2918 F.HasBaseReg = true; 2919 bool Inserted = InsertFormula(LU, LUIdx, F); 2920 assert(Inserted && "Supplemental formula already exists!"); (void)Inserted; 2921 } 2922 2923 /// CountRegisters - Note which registers are used by the given formula, 2924 /// updating RegUses. 2925 void LSRInstance::CountRegisters(const Formula &F, size_t LUIdx) { 2926 if (F.ScaledReg) 2927 RegUses.CountRegister(F.ScaledReg, LUIdx); 2928 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(), 2929 E = F.BaseRegs.end(); I != E; ++I) 2930 RegUses.CountRegister(*I, LUIdx); 2931 } 2932 2933 /// InsertFormula - If the given formula has not yet been inserted, add it to 2934 /// the list, and return true. Return false otherwise. 2935 bool LSRInstance::InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F) { 2936 if (!LU.InsertFormula(F)) 2937 return false; 2938 2939 CountRegisters(F, LUIdx); 2940 return true; 2941 } 2942 2943 /// CollectLoopInvariantFixupsAndFormulae - Check for other uses of 2944 /// loop-invariant values which we're tracking. These other uses will pin these 2945 /// values in registers, making them less profitable for elimination. 2946 /// TODO: This currently misses non-constant addrec step registers. 2947 /// TODO: Should this give more weight to users inside the loop? 2948 void 2949 LSRInstance::CollectLoopInvariantFixupsAndFormulae() { 2950 SmallVector<const SCEV *, 8> Worklist(RegUses.begin(), RegUses.end()); 2951 SmallPtrSet<const SCEV *, 8> Inserted; 2952 2953 while (!Worklist.empty()) { 2954 const SCEV *S = Worklist.pop_back_val(); 2955 2956 if (const SCEVNAryExpr *N = dyn_cast<SCEVNAryExpr>(S)) 2957 Worklist.append(N->op_begin(), N->op_end()); 2958 else if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(S)) 2959 Worklist.push_back(C->getOperand()); 2960 else if (const SCEVUDivExpr *D = dyn_cast<SCEVUDivExpr>(S)) { 2961 Worklist.push_back(D->getLHS()); 2962 Worklist.push_back(D->getRHS()); 2963 } else if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) { 2964 if (!Inserted.insert(U)) continue; 2965 const Value *V = U->getValue(); 2966 if (const Instruction *Inst = dyn_cast<Instruction>(V)) { 2967 // Look for instructions defined outside the loop. 2968 if (L->contains(Inst)) continue; 2969 } else if (isa<UndefValue>(V)) 2970 // Undef doesn't have a live range, so it doesn't matter. 2971 continue; 2972 for (Value::const_use_iterator UI = V->use_begin(), UE = V->use_end(); 2973 UI != UE; ++UI) { 2974 const Instruction *UserInst = dyn_cast<Instruction>(*UI); 2975 // Ignore non-instructions. 2976 if (!UserInst) 2977 continue; 2978 // Ignore instructions in other functions (as can happen with 2979 // Constants). 2980 if (UserInst->getParent()->getParent() != L->getHeader()->getParent()) 2981 continue; 2982 // Ignore instructions not dominated by the loop. 2983 const BasicBlock *UseBB = !isa<PHINode>(UserInst) ? 2984 UserInst->getParent() : 2985 cast<PHINode>(UserInst)->getIncomingBlock( 2986 PHINode::getIncomingValueNumForOperand(UI.getOperandNo())); 2987 if (!DT.dominates(L->getHeader(), UseBB)) 2988 continue; 2989 // Ignore uses which are part of other SCEV expressions, to avoid 2990 // analyzing them multiple times. 2991 if (SE.isSCEVable(UserInst->getType())) { 2992 const SCEV *UserS = SE.getSCEV(const_cast<Instruction *>(UserInst)); 2993 // If the user is a no-op, look through to its uses. 2994 if (!isa<SCEVUnknown>(UserS)) 2995 continue; 2996 if (UserS == U) { 2997 Worklist.push_back( 2998 SE.getUnknown(const_cast<Instruction *>(UserInst))); 2999 continue; 3000 } 3001 } 3002 // Ignore icmp instructions which are already being analyzed. 3003 if (const ICmpInst *ICI = dyn_cast<ICmpInst>(UserInst)) { 3004 unsigned OtherIdx = !UI.getOperandNo(); 3005 Value *OtherOp = const_cast<Value *>(ICI->getOperand(OtherIdx)); 3006 if (SE.hasComputableLoopEvolution(SE.getSCEV(OtherOp), L)) 3007 continue; 3008 } 3009 3010 LSRFixup &LF = getNewFixup(); 3011 LF.UserInst = const_cast<Instruction *>(UserInst); 3012 LF.OperandValToReplace = UI.getUse(); 3013 std::pair<size_t, int64_t> P = getUse(S, LSRUse::Basic, 0); 3014 LF.LUIdx = P.first; 3015 LF.Offset = P.second; 3016 LSRUse &LU = Uses[LF.LUIdx]; 3017 LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L); 3018 if (!LU.WidestFixupType || 3019 SE.getTypeSizeInBits(LU.WidestFixupType) < 3020 SE.getTypeSizeInBits(LF.OperandValToReplace->getType())) 3021 LU.WidestFixupType = LF.OperandValToReplace->getType(); 3022 InsertSupplementalFormula(U, LU, LF.LUIdx); 3023 CountRegisters(LU.Formulae.back(), Uses.size() - 1); 3024 break; 3025 } 3026 } 3027 } 3028 } 3029 3030 /// CollectSubexprs - Split S into subexpressions which can be pulled out into 3031 /// separate registers. If C is non-null, multiply each subexpression by C. 3032 /// 3033 /// Return remainder expression after factoring the subexpressions captured by 3034 /// Ops. If Ops is complete, return NULL. 3035 static const SCEV *CollectSubexprs(const SCEV *S, const SCEVConstant *C, 3036 SmallVectorImpl<const SCEV *> &Ops, 3037 const Loop *L, 3038 ScalarEvolution &SE, 3039 unsigned Depth = 0) { 3040 // Arbitrarily cap recursion to protect compile time. 3041 if (Depth >= 3) 3042 return S; 3043 3044 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) { 3045 // Break out add operands. 3046 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end(); 3047 I != E; ++I) { 3048 const SCEV *Remainder = CollectSubexprs(*I, C, Ops, L, SE, Depth+1); 3049 if (Remainder) 3050 Ops.push_back(C ? SE.getMulExpr(C, Remainder) : Remainder); 3051 } 3052 return NULL; 3053 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) { 3054 // Split a non-zero base out of an addrec. 3055 if (AR->getStart()->isZero()) 3056 return S; 3057 3058 const SCEV *Remainder = CollectSubexprs(AR->getStart(), 3059 C, Ops, L, SE, Depth+1); 3060 // Split the non-zero AddRec unless it is part of a nested recurrence that 3061 // does not pertain to this loop. 3062 if (Remainder && (AR->getLoop() == L || !isa<SCEVAddRecExpr>(Remainder))) { 3063 Ops.push_back(C ? SE.getMulExpr(C, Remainder) : Remainder); 3064 Remainder = NULL; 3065 } 3066 if (Remainder != AR->getStart()) { 3067 if (!Remainder) 3068 Remainder = SE.getConstant(AR->getType(), 0); 3069 return SE.getAddRecExpr(Remainder, 3070 AR->getStepRecurrence(SE), 3071 AR->getLoop(), 3072 //FIXME: AR->getNoWrapFlags(SCEV::FlagNW) 3073 SCEV::FlagAnyWrap); 3074 } 3075 } else if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) { 3076 // Break (C * (a + b + c)) into C*a + C*b + C*c. 3077 if (Mul->getNumOperands() != 2) 3078 return S; 3079 if (const SCEVConstant *Op0 = 3080 dyn_cast<SCEVConstant>(Mul->getOperand(0))) { 3081 C = C ? cast<SCEVConstant>(SE.getMulExpr(C, Op0)) : Op0; 3082 const SCEV *Remainder = 3083 CollectSubexprs(Mul->getOperand(1), C, Ops, L, SE, Depth+1); 3084 if (Remainder) 3085 Ops.push_back(SE.getMulExpr(C, Remainder)); 3086 return NULL; 3087 } 3088 } 3089 return S; 3090 } 3091 3092 /// GenerateReassociations - Split out subexpressions from adds and the bases of 3093 /// addrecs. 3094 void LSRInstance::GenerateReassociations(LSRUse &LU, unsigned LUIdx, 3095 Formula Base, 3096 unsigned Depth) { 3097 // Arbitrarily cap recursion to protect compile time. 3098 if (Depth >= 3) return; 3099 3100 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) { 3101 const SCEV *BaseReg = Base.BaseRegs[i]; 3102 3103 SmallVector<const SCEV *, 8> AddOps; 3104 const SCEV *Remainder = CollectSubexprs(BaseReg, 0, AddOps, L, SE); 3105 if (Remainder) 3106 AddOps.push_back(Remainder); 3107 3108 if (AddOps.size() == 1) continue; 3109 3110 for (SmallVectorImpl<const SCEV *>::const_iterator J = AddOps.begin(), 3111 JE = AddOps.end(); J != JE; ++J) { 3112 3113 // Loop-variant "unknown" values are uninteresting; we won't be able to 3114 // do anything meaningful with them. 3115 if (isa<SCEVUnknown>(*J) && !SE.isLoopInvariant(*J, L)) 3116 continue; 3117 3118 // Don't pull a constant into a register if the constant could be folded 3119 // into an immediate field. 3120 if (isAlwaysFoldable(TTI, SE, LU.MinOffset, LU.MaxOffset, LU.Kind, 3121 LU.AccessTy, *J, Base.getNumRegs() > 1)) 3122 continue; 3123 3124 // Collect all operands except *J. 3125 SmallVector<const SCEV *, 8> InnerAddOps 3126 (((const SmallVector<const SCEV *, 8> &)AddOps).begin(), J); 3127 InnerAddOps.append 3128 (llvm::next(J), ((const SmallVector<const SCEV *, 8> &)AddOps).end()); 3129 3130 // Don't leave just a constant behind in a register if the constant could 3131 // be folded into an immediate field. 3132 if (InnerAddOps.size() == 1 && 3133 isAlwaysFoldable(TTI, SE, LU.MinOffset, LU.MaxOffset, LU.Kind, 3134 LU.AccessTy, InnerAddOps[0], Base.getNumRegs() > 1)) 3135 continue; 3136 3137 const SCEV *InnerSum = SE.getAddExpr(InnerAddOps); 3138 if (InnerSum->isZero()) 3139 continue; 3140 Formula F = Base; 3141 3142 // Add the remaining pieces of the add back into the new formula. 3143 const SCEVConstant *InnerSumSC = dyn_cast<SCEVConstant>(InnerSum); 3144 if (InnerSumSC && 3145 SE.getTypeSizeInBits(InnerSumSC->getType()) <= 64 && 3146 TTI.isLegalAddImmediate((uint64_t)F.UnfoldedOffset + 3147 InnerSumSC->getValue()->getZExtValue())) { 3148 F.UnfoldedOffset = (uint64_t)F.UnfoldedOffset + 3149 InnerSumSC->getValue()->getZExtValue(); 3150 F.BaseRegs.erase(F.BaseRegs.begin() + i); 3151 } else 3152 F.BaseRegs[i] = InnerSum; 3153 3154 // Add J as its own register, or an unfolded immediate. 3155 const SCEVConstant *SC = dyn_cast<SCEVConstant>(*J); 3156 if (SC && SE.getTypeSizeInBits(SC->getType()) <= 64 && 3157 TTI.isLegalAddImmediate((uint64_t)F.UnfoldedOffset + 3158 SC->getValue()->getZExtValue())) 3159 F.UnfoldedOffset = (uint64_t)F.UnfoldedOffset + 3160 SC->getValue()->getZExtValue(); 3161 else 3162 F.BaseRegs.push_back(*J); 3163 3164 if (InsertFormula(LU, LUIdx, F)) 3165 // If that formula hadn't been seen before, recurse to find more like 3166 // it. 3167 GenerateReassociations(LU, LUIdx, LU.Formulae.back(), Depth+1); 3168 } 3169 } 3170 } 3171 3172 /// GenerateCombinations - Generate a formula consisting of all of the 3173 /// loop-dominating registers added into a single register. 3174 void LSRInstance::GenerateCombinations(LSRUse &LU, unsigned LUIdx, 3175 Formula Base) { 3176 // This method is only interesting on a plurality of registers. 3177 if (Base.BaseRegs.size() <= 1) return; 3178 3179 Formula F = Base; 3180 F.BaseRegs.clear(); 3181 SmallVector<const SCEV *, 4> Ops; 3182 for (SmallVectorImpl<const SCEV *>::const_iterator 3183 I = Base.BaseRegs.begin(), E = Base.BaseRegs.end(); I != E; ++I) { 3184 const SCEV *BaseReg = *I; 3185 if (SE.properlyDominates(BaseReg, L->getHeader()) && 3186 !SE.hasComputableLoopEvolution(BaseReg, L)) 3187 Ops.push_back(BaseReg); 3188 else 3189 F.BaseRegs.push_back(BaseReg); 3190 } 3191 if (Ops.size() > 1) { 3192 const SCEV *Sum = SE.getAddExpr(Ops); 3193 // TODO: If Sum is zero, it probably means ScalarEvolution missed an 3194 // opportunity to fold something. For now, just ignore such cases 3195 // rather than proceed with zero in a register. 3196 if (!Sum->isZero()) { 3197 F.BaseRegs.push_back(Sum); 3198 (void)InsertFormula(LU, LUIdx, F); 3199 } 3200 } 3201 } 3202 3203 /// GenerateSymbolicOffsets - Generate reuse formulae using symbolic offsets. 3204 void LSRInstance::GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx, 3205 Formula Base) { 3206 // We can't add a symbolic offset if the address already contains one. 3207 if (Base.BaseGV) return; 3208 3209 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) { 3210 const SCEV *G = Base.BaseRegs[i]; 3211 GlobalValue *GV = ExtractSymbol(G, SE); 3212 if (G->isZero() || !GV) 3213 continue; 3214 Formula F = Base; 3215 F.BaseGV = GV; 3216 if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy, F)) 3217 continue; 3218 F.BaseRegs[i] = G; 3219 (void)InsertFormula(LU, LUIdx, F); 3220 } 3221 } 3222 3223 /// GenerateConstantOffsets - Generate reuse formulae using symbolic offsets. 3224 void LSRInstance::GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx, 3225 Formula Base) { 3226 // TODO: For now, just add the min and max offset, because it usually isn't 3227 // worthwhile looking at everything inbetween. 3228 SmallVector<int64_t, 2> Worklist; 3229 Worklist.push_back(LU.MinOffset); 3230 if (LU.MaxOffset != LU.MinOffset) 3231 Worklist.push_back(LU.MaxOffset); 3232 3233 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) { 3234 const SCEV *G = Base.BaseRegs[i]; 3235 3236 for (SmallVectorImpl<int64_t>::const_iterator I = Worklist.begin(), 3237 E = Worklist.end(); I != E; ++I) { 3238 Formula F = Base; 3239 F.BaseOffset = (uint64_t)Base.BaseOffset - *I; 3240 if (isLegalUse(TTI, LU.MinOffset - *I, LU.MaxOffset - *I, LU.Kind, 3241 LU.AccessTy, F)) { 3242 // Add the offset to the base register. 3243 const SCEV *NewG = SE.getAddExpr(SE.getConstant(G->getType(), *I), G); 3244 // If it cancelled out, drop the base register, otherwise update it. 3245 if (NewG->isZero()) { 3246 std::swap(F.BaseRegs[i], F.BaseRegs.back()); 3247 F.BaseRegs.pop_back(); 3248 } else 3249 F.BaseRegs[i] = NewG; 3250 3251 (void)InsertFormula(LU, LUIdx, F); 3252 } 3253 } 3254 3255 int64_t Imm = ExtractImmediate(G, SE); 3256 if (G->isZero() || Imm == 0) 3257 continue; 3258 Formula F = Base; 3259 F.BaseOffset = (uint64_t)F.BaseOffset + Imm; 3260 if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy, F)) 3261 continue; 3262 F.BaseRegs[i] = G; 3263 (void)InsertFormula(LU, LUIdx, F); 3264 } 3265 } 3266 3267 /// GenerateICmpZeroScales - For ICmpZero, check to see if we can scale up 3268 /// the comparison. For example, x == y -> x*c == y*c. 3269 void LSRInstance::GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx, 3270 Formula Base) { 3271 if (LU.Kind != LSRUse::ICmpZero) return; 3272 3273 // Determine the integer type for the base formula. 3274 Type *IntTy = Base.getType(); 3275 if (!IntTy) return; 3276 if (SE.getTypeSizeInBits(IntTy) > 64) return; 3277 3278 // Don't do this if there is more than one offset. 3279 if (LU.MinOffset != LU.MaxOffset) return; 3280 3281 assert(!Base.BaseGV && "ICmpZero use is not legal!"); 3282 3283 // Check each interesting stride. 3284 for (SmallSetVector<int64_t, 8>::const_iterator 3285 I = Factors.begin(), E = Factors.end(); I != E; ++I) { 3286 int64_t Factor = *I; 3287 3288 // Check that the multiplication doesn't overflow. 3289 if (Base.BaseOffset == INT64_MIN && Factor == -1) 3290 continue; 3291 int64_t NewBaseOffset = (uint64_t)Base.BaseOffset * Factor; 3292 if (NewBaseOffset / Factor != Base.BaseOffset) 3293 continue; 3294 3295 // Check that multiplying with the use offset doesn't overflow. 3296 int64_t Offset = LU.MinOffset; 3297 if (Offset == INT64_MIN && Factor == -1) 3298 continue; 3299 Offset = (uint64_t)Offset * Factor; 3300 if (Offset / Factor != LU.MinOffset) 3301 continue; 3302 3303 Formula F = Base; 3304 F.BaseOffset = NewBaseOffset; 3305 3306 // Check that this scale is legal. 3307 if (!isLegalUse(TTI, Offset, Offset, LU.Kind, LU.AccessTy, F)) 3308 continue; 3309 3310 // Compensate for the use having MinOffset built into it. 3311 F.BaseOffset = (uint64_t)F.BaseOffset + Offset - LU.MinOffset; 3312 3313 const SCEV *FactorS = SE.getConstant(IntTy, Factor); 3314 3315 // Check that multiplying with each base register doesn't overflow. 3316 for (size_t i = 0, e = F.BaseRegs.size(); i != e; ++i) { 3317 F.BaseRegs[i] = SE.getMulExpr(F.BaseRegs[i], FactorS); 3318 if (getExactSDiv(F.BaseRegs[i], FactorS, SE) != Base.BaseRegs[i]) 3319 goto next; 3320 } 3321 3322 // Check that multiplying with the scaled register doesn't overflow. 3323 if (F.ScaledReg) { 3324 F.ScaledReg = SE.getMulExpr(F.ScaledReg, FactorS); 3325 if (getExactSDiv(F.ScaledReg, FactorS, SE) != Base.ScaledReg) 3326 continue; 3327 } 3328 3329 // Check that multiplying with the unfolded offset doesn't overflow. 3330 if (F.UnfoldedOffset != 0) { 3331 if (F.UnfoldedOffset == INT64_MIN && Factor == -1) 3332 continue; 3333 F.UnfoldedOffset = (uint64_t)F.UnfoldedOffset * Factor; 3334 if (F.UnfoldedOffset / Factor != Base.UnfoldedOffset) 3335 continue; 3336 } 3337 3338 // If we make it here and it's legal, add it. 3339 (void)InsertFormula(LU, LUIdx, F); 3340 next:; 3341 } 3342 } 3343 3344 /// GenerateScales - Generate stride factor reuse formulae by making use of 3345 /// scaled-offset address modes, for example. 3346 void LSRInstance::GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base) { 3347 // Determine the integer type for the base formula. 3348 Type *IntTy = Base.getType(); 3349 if (!IntTy) return; 3350 3351 // If this Formula already has a scaled register, we can't add another one. 3352 if (Base.Scale != 0) return; 3353 3354 // Check each interesting stride. 3355 for (SmallSetVector<int64_t, 8>::const_iterator 3356 I = Factors.begin(), E = Factors.end(); I != E; ++I) { 3357 int64_t Factor = *I; 3358 3359 Base.Scale = Factor; 3360 Base.HasBaseReg = Base.BaseRegs.size() > 1; 3361 // Check whether this scale is going to be legal. 3362 if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy, 3363 Base)) { 3364 // As a special-case, handle special out-of-loop Basic users specially. 3365 // TODO: Reconsider this special case. 3366 if (LU.Kind == LSRUse::Basic && 3367 isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LSRUse::Special, 3368 LU.AccessTy, Base) && 3369 LU.AllFixupsOutsideLoop) 3370 LU.Kind = LSRUse::Special; 3371 else 3372 continue; 3373 } 3374 // For an ICmpZero, negating a solitary base register won't lead to 3375 // new solutions. 3376 if (LU.Kind == LSRUse::ICmpZero && 3377 !Base.HasBaseReg && Base.BaseOffset == 0 && !Base.BaseGV) 3378 continue; 3379 // For each addrec base reg, apply the scale, if possible. 3380 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) 3381 if (const SCEVAddRecExpr *AR = 3382 dyn_cast<SCEVAddRecExpr>(Base.BaseRegs[i])) { 3383 const SCEV *FactorS = SE.getConstant(IntTy, Factor); 3384 if (FactorS->isZero()) 3385 continue; 3386 // Divide out the factor, ignoring high bits, since we'll be 3387 // scaling the value back up in the end. 3388 if (const SCEV *Quotient = getExactSDiv(AR, FactorS, SE, true)) { 3389 // TODO: This could be optimized to avoid all the copying. 3390 Formula F = Base; 3391 F.ScaledReg = Quotient; 3392 F.DeleteBaseReg(F.BaseRegs[i]); 3393 (void)InsertFormula(LU, LUIdx, F); 3394 } 3395 } 3396 } 3397 } 3398 3399 /// GenerateTruncates - Generate reuse formulae from different IV types. 3400 void LSRInstance::GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base) { 3401 // Don't bother truncating symbolic values. 3402 if (Base.BaseGV) return; 3403 3404 // Determine the integer type for the base formula. 3405 Type *DstTy = Base.getType(); 3406 if (!DstTy) return; 3407 DstTy = SE.getEffectiveSCEVType(DstTy); 3408 3409 for (SmallSetVector<Type *, 4>::const_iterator 3410 I = Types.begin(), E = Types.end(); I != E; ++I) { 3411 Type *SrcTy = *I; 3412 if (SrcTy != DstTy && TTI.isTruncateFree(SrcTy, DstTy)) { 3413 Formula F = Base; 3414 3415 if (F.ScaledReg) F.ScaledReg = SE.getAnyExtendExpr(F.ScaledReg, *I); 3416 for (SmallVectorImpl<const SCEV *>::iterator J = F.BaseRegs.begin(), 3417 JE = F.BaseRegs.end(); J != JE; ++J) 3418 *J = SE.getAnyExtendExpr(*J, SrcTy); 3419 3420 // TODO: This assumes we've done basic processing on all uses and 3421 // have an idea what the register usage is. 3422 if (!F.hasRegsUsedByUsesOtherThan(LUIdx, RegUses)) 3423 continue; 3424 3425 (void)InsertFormula(LU, LUIdx, F); 3426 } 3427 } 3428 } 3429 3430 namespace { 3431 3432 /// WorkItem - Helper class for GenerateCrossUseConstantOffsets. It's used to 3433 /// defer modifications so that the search phase doesn't have to worry about 3434 /// the data structures moving underneath it. 3435 struct WorkItem { 3436 size_t LUIdx; 3437 int64_t Imm; 3438 const SCEV *OrigReg; 3439 3440 WorkItem(size_t LI, int64_t I, const SCEV *R) 3441 : LUIdx(LI), Imm(I), OrigReg(R) {} 3442 3443 void print(raw_ostream &OS) const; 3444 void dump() const; 3445 }; 3446 3447 } 3448 3449 void WorkItem::print(raw_ostream &OS) const { 3450 OS << "in formulae referencing " << *OrigReg << " in use " << LUIdx 3451 << " , add offset " << Imm; 3452 } 3453 3454 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) 3455 void WorkItem::dump() const { 3456 print(errs()); errs() << '\n'; 3457 } 3458 #endif 3459 3460 /// GenerateCrossUseConstantOffsets - Look for registers which are a constant 3461 /// distance apart and try to form reuse opportunities between them. 3462 void LSRInstance::GenerateCrossUseConstantOffsets() { 3463 // Group the registers by their value without any added constant offset. 3464 typedef std::map<int64_t, const SCEV *> ImmMapTy; 3465 typedef DenseMap<const SCEV *, ImmMapTy> RegMapTy; 3466 RegMapTy Map; 3467 DenseMap<const SCEV *, SmallBitVector> UsedByIndicesMap; 3468 SmallVector<const SCEV *, 8> Sequence; 3469 for (RegUseTracker::const_iterator I = RegUses.begin(), E = RegUses.end(); 3470 I != E; ++I) { 3471 const SCEV *Reg = *I; 3472 int64_t Imm = ExtractImmediate(Reg, SE); 3473 std::pair<RegMapTy::iterator, bool> Pair = 3474 Map.insert(std::make_pair(Reg, ImmMapTy())); 3475 if (Pair.second) 3476 Sequence.push_back(Reg); 3477 Pair.first->second.insert(std::make_pair(Imm, *I)); 3478 UsedByIndicesMap[Reg] |= RegUses.getUsedByIndices(*I); 3479 } 3480 3481 // Now examine each set of registers with the same base value. Build up 3482 // a list of work to do and do the work in a separate step so that we're 3483 // not adding formulae and register counts while we're searching. 3484 SmallVector<WorkItem, 32> WorkItems; 3485 SmallSet<std::pair<size_t, int64_t>, 32> UniqueItems; 3486 for (SmallVectorImpl<const SCEV *>::const_iterator I = Sequence.begin(), 3487 E = Sequence.end(); I != E; ++I) { 3488 const SCEV *Reg = *I; 3489 const ImmMapTy &Imms = Map.find(Reg)->second; 3490 3491 // It's not worthwhile looking for reuse if there's only one offset. 3492 if (Imms.size() == 1) 3493 continue; 3494 3495 DEBUG(dbgs() << "Generating cross-use offsets for " << *Reg << ':'; 3496 for (ImmMapTy::const_iterator J = Imms.begin(), JE = Imms.end(); 3497 J != JE; ++J) 3498 dbgs() << ' ' << J->first; 3499 dbgs() << '\n'); 3500 3501 // Examine each offset. 3502 for (ImmMapTy::const_iterator J = Imms.begin(), JE = Imms.end(); 3503 J != JE; ++J) { 3504 const SCEV *OrigReg = J->second; 3505 3506 int64_t JImm = J->first; 3507 const SmallBitVector &UsedByIndices = RegUses.getUsedByIndices(OrigReg); 3508 3509 if (!isa<SCEVConstant>(OrigReg) && 3510 UsedByIndicesMap[Reg].count() == 1) { 3511 DEBUG(dbgs() << "Skipping cross-use reuse for " << *OrigReg << '\n'); 3512 continue; 3513 } 3514 3515 // Conservatively examine offsets between this orig reg a few selected 3516 // other orig regs. 3517 ImmMapTy::const_iterator OtherImms[] = { 3518 Imms.begin(), prior(Imms.end()), 3519 Imms.lower_bound((Imms.begin()->first + prior(Imms.end())->first) / 2) 3520 }; 3521 for (size_t i = 0, e = array_lengthof(OtherImms); i != e; ++i) { 3522 ImmMapTy::const_iterator M = OtherImms[i]; 3523 if (M == J || M == JE) continue; 3524 3525 // Compute the difference between the two. 3526 int64_t Imm = (uint64_t)JImm - M->first; 3527 for (int LUIdx = UsedByIndices.find_first(); LUIdx != -1; 3528 LUIdx = UsedByIndices.find_next(LUIdx)) 3529 // Make a memo of this use, offset, and register tuple. 3530 if (UniqueItems.insert(std::make_pair(LUIdx, Imm))) 3531 WorkItems.push_back(WorkItem(LUIdx, Imm, OrigReg)); 3532 } 3533 } 3534 } 3535 3536 Map.clear(); 3537 Sequence.clear(); 3538 UsedByIndicesMap.clear(); 3539 UniqueItems.clear(); 3540 3541 // Now iterate through the worklist and add new formulae. 3542 for (SmallVectorImpl<WorkItem>::const_iterator I = WorkItems.begin(), 3543 E = WorkItems.end(); I != E; ++I) { 3544 const WorkItem &WI = *I; 3545 size_t LUIdx = WI.LUIdx; 3546 LSRUse &LU = Uses[LUIdx]; 3547 int64_t Imm = WI.Imm; 3548 const SCEV *OrigReg = WI.OrigReg; 3549 3550 Type *IntTy = SE.getEffectiveSCEVType(OrigReg->getType()); 3551 const SCEV *NegImmS = SE.getSCEV(ConstantInt::get(IntTy, -(uint64_t)Imm)); 3552 unsigned BitWidth = SE.getTypeSizeInBits(IntTy); 3553 3554 // TODO: Use a more targeted data structure. 3555 for (size_t L = 0, LE = LU.Formulae.size(); L != LE; ++L) { 3556 const Formula &F = LU.Formulae[L]; 3557 // Use the immediate in the scaled register. 3558 if (F.ScaledReg == OrigReg) { 3559 int64_t Offset = (uint64_t)F.BaseOffset + Imm * (uint64_t)F.Scale; 3560 // Don't create 50 + reg(-50). 3561 if (F.referencesReg(SE.getSCEV( 3562 ConstantInt::get(IntTy, -(uint64_t)Offset)))) 3563 continue; 3564 Formula NewF = F; 3565 NewF.BaseOffset = Offset; 3566 if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy, 3567 NewF)) 3568 continue; 3569 NewF.ScaledReg = SE.getAddExpr(NegImmS, NewF.ScaledReg); 3570 3571 // If the new scale is a constant in a register, and adding the constant 3572 // value to the immediate would produce a value closer to zero than the 3573 // immediate itself, then the formula isn't worthwhile. 3574 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(NewF.ScaledReg)) 3575 if (C->getValue()->isNegative() != 3576 (NewF.BaseOffset < 0) && 3577 (C->getValue()->getValue().abs() * APInt(BitWidth, F.Scale)) 3578 .ule(abs64(NewF.BaseOffset))) 3579 continue; 3580 3581 // OK, looks good. 3582 (void)InsertFormula(LU, LUIdx, NewF); 3583 } else { 3584 // Use the immediate in a base register. 3585 for (size_t N = 0, NE = F.BaseRegs.size(); N != NE; ++N) { 3586 const SCEV *BaseReg = F.BaseRegs[N]; 3587 if (BaseReg != OrigReg) 3588 continue; 3589 Formula NewF = F; 3590 NewF.BaseOffset = (uint64_t)NewF.BaseOffset + Imm; 3591 if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, 3592 LU.Kind, LU.AccessTy, NewF)) { 3593 if (!TTI.isLegalAddImmediate((uint64_t)NewF.UnfoldedOffset + Imm)) 3594 continue; 3595 NewF = F; 3596 NewF.UnfoldedOffset = (uint64_t)NewF.UnfoldedOffset + Imm; 3597 } 3598 NewF.BaseRegs[N] = SE.getAddExpr(NegImmS, BaseReg); 3599 3600 // If the new formula has a constant in a register, and adding the 3601 // constant value to the immediate would produce a value closer to 3602 // zero than the immediate itself, then the formula isn't worthwhile. 3603 for (SmallVectorImpl<const SCEV *>::const_iterator 3604 J = NewF.BaseRegs.begin(), JE = NewF.BaseRegs.end(); 3605 J != JE; ++J) 3606 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(*J)) 3607 if ((C->getValue()->getValue() + NewF.BaseOffset).abs().slt( 3608 abs64(NewF.BaseOffset)) && 3609 (C->getValue()->getValue() + 3610 NewF.BaseOffset).countTrailingZeros() >= 3611 CountTrailingZeros_64(NewF.BaseOffset)) 3612 goto skip_formula; 3613 3614 // Ok, looks good. 3615 (void)InsertFormula(LU, LUIdx, NewF); 3616 break; 3617 skip_formula:; 3618 } 3619 } 3620 } 3621 } 3622 } 3623 3624 /// GenerateAllReuseFormulae - Generate formulae for each use. 3625 void 3626 LSRInstance::GenerateAllReuseFormulae() { 3627 // This is split into multiple loops so that hasRegsUsedByUsesOtherThan 3628 // queries are more precise. 3629 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) { 3630 LSRUse &LU = Uses[LUIdx]; 3631 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i) 3632 GenerateReassociations(LU, LUIdx, LU.Formulae[i]); 3633 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i) 3634 GenerateCombinations(LU, LUIdx, LU.Formulae[i]); 3635 } 3636 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) { 3637 LSRUse &LU = Uses[LUIdx]; 3638 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i) 3639 GenerateSymbolicOffsets(LU, LUIdx, LU.Formulae[i]); 3640 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i) 3641 GenerateConstantOffsets(LU, LUIdx, LU.Formulae[i]); 3642 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i) 3643 GenerateICmpZeroScales(LU, LUIdx, LU.Formulae[i]); 3644 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i) 3645 GenerateScales(LU, LUIdx, LU.Formulae[i]); 3646 } 3647 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) { 3648 LSRUse &LU = Uses[LUIdx]; 3649 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i) 3650 GenerateTruncates(LU, LUIdx, LU.Formulae[i]); 3651 } 3652 3653 GenerateCrossUseConstantOffsets(); 3654 3655 DEBUG(dbgs() << "\n" 3656 "After generating reuse formulae:\n"; 3657 print_uses(dbgs())); 3658 } 3659 3660 /// If there are multiple formulae with the same set of registers used 3661 /// by other uses, pick the best one and delete the others. 3662 void LSRInstance::FilterOutUndesirableDedicatedRegisters() { 3663 DenseSet<const SCEV *> VisitedRegs; 3664 SmallPtrSet<const SCEV *, 16> Regs; 3665 SmallPtrSet<const SCEV *, 16> LoserRegs; 3666 #ifndef NDEBUG 3667 bool ChangedFormulae = false; 3668 #endif 3669 3670 // Collect the best formula for each unique set of shared registers. This 3671 // is reset for each use. 3672 typedef DenseMap<SmallVector<const SCEV *, 4>, size_t, UniquifierDenseMapInfo> 3673 BestFormulaeTy; 3674 BestFormulaeTy BestFormulae; 3675 3676 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) { 3677 LSRUse &LU = Uses[LUIdx]; 3678 DEBUG(dbgs() << "Filtering for use "; LU.print(dbgs()); dbgs() << '\n'); 3679 3680 bool Any = false; 3681 for (size_t FIdx = 0, NumForms = LU.Formulae.size(); 3682 FIdx != NumForms; ++FIdx) { 3683 Formula &F = LU.Formulae[FIdx]; 3684 3685 // Some formulas are instant losers. For example, they may depend on 3686 // nonexistent AddRecs from other loops. These need to be filtered 3687 // immediately, otherwise heuristics could choose them over others leading 3688 // to an unsatisfactory solution. Passing LoserRegs into RateFormula here 3689 // avoids the need to recompute this information across formulae using the 3690 // same bad AddRec. Passing LoserRegs is also essential unless we remove 3691 // the corresponding bad register from the Regs set. 3692 Cost CostF; 3693 Regs.clear(); 3694 CostF.RateFormula(F, Regs, VisitedRegs, L, LU.Offsets, SE, DT, 3695 &LoserRegs); 3696 if (CostF.isLoser()) { 3697 // During initial formula generation, undesirable formulae are generated 3698 // by uses within other loops that have some non-trivial address mode or 3699 // use the postinc form of the IV. LSR needs to provide these formulae 3700 // as the basis of rediscovering the desired formula that uses an AddRec 3701 // corresponding to the existing phi. Once all formulae have been 3702 // generated, these initial losers may be pruned. 3703 DEBUG(dbgs() << " Filtering loser "; F.print(dbgs()); 3704 dbgs() << "\n"); 3705 } 3706 else { 3707 SmallVector<const SCEV *, 4> Key; 3708 for (SmallVectorImpl<const SCEV *>::const_iterator J = F.BaseRegs.begin(), 3709 JE = F.BaseRegs.end(); J != JE; ++J) { 3710 const SCEV *Reg = *J; 3711 if (RegUses.isRegUsedByUsesOtherThan(Reg, LUIdx)) 3712 Key.push_back(Reg); 3713 } 3714 if (F.ScaledReg && 3715 RegUses.isRegUsedByUsesOtherThan(F.ScaledReg, LUIdx)) 3716 Key.push_back(F.ScaledReg); 3717 // Unstable sort by host order ok, because this is only used for 3718 // uniquifying. 3719 std::sort(Key.begin(), Key.end()); 3720 3721 std::pair<BestFormulaeTy::const_iterator, bool> P = 3722 BestFormulae.insert(std::make_pair(Key, FIdx)); 3723 if (P.second) 3724 continue; 3725 3726 Formula &Best = LU.Formulae[P.first->second]; 3727 3728 Cost CostBest; 3729 Regs.clear(); 3730 CostBest.RateFormula(Best, Regs, VisitedRegs, L, LU.Offsets, SE, DT); 3731 if (CostF < CostBest) 3732 std::swap(F, Best); 3733 DEBUG(dbgs() << " Filtering out formula "; F.print(dbgs()); 3734 dbgs() << "\n" 3735 " in favor of formula "; Best.print(dbgs()); 3736 dbgs() << '\n'); 3737 } 3738 #ifndef NDEBUG 3739 ChangedFormulae = true; 3740 #endif 3741 LU.DeleteFormula(F); 3742 --FIdx; 3743 --NumForms; 3744 Any = true; 3745 } 3746 3747 // Now that we've filtered out some formulae, recompute the Regs set. 3748 if (Any) 3749 LU.RecomputeRegs(LUIdx, RegUses); 3750 3751 // Reset this to prepare for the next use. 3752 BestFormulae.clear(); 3753 } 3754 3755 DEBUG(if (ChangedFormulae) { 3756 dbgs() << "\n" 3757 "After filtering out undesirable candidates:\n"; 3758 print_uses(dbgs()); 3759 }); 3760 } 3761 3762 // This is a rough guess that seems to work fairly well. 3763 static const size_t ComplexityLimit = UINT16_MAX; 3764 3765 /// EstimateSearchSpaceComplexity - Estimate the worst-case number of 3766 /// solutions the solver might have to consider. It almost never considers 3767 /// this many solutions because it prune the search space, but the pruning 3768 /// isn't always sufficient. 3769 size_t LSRInstance::EstimateSearchSpaceComplexity() const { 3770 size_t Power = 1; 3771 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(), 3772 E = Uses.end(); I != E; ++I) { 3773 size_t FSize = I->Formulae.size(); 3774 if (FSize >= ComplexityLimit) { 3775 Power = ComplexityLimit; 3776 break; 3777 } 3778 Power *= FSize; 3779 if (Power >= ComplexityLimit) 3780 break; 3781 } 3782 return Power; 3783 } 3784 3785 /// NarrowSearchSpaceByDetectingSupersets - When one formula uses a superset 3786 /// of the registers of another formula, it won't help reduce register 3787 /// pressure (though it may not necessarily hurt register pressure); remove 3788 /// it to simplify the system. 3789 void LSRInstance::NarrowSearchSpaceByDetectingSupersets() { 3790 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) { 3791 DEBUG(dbgs() << "The search space is too complex.\n"); 3792 3793 DEBUG(dbgs() << "Narrowing the search space by eliminating formulae " 3794 "which use a superset of registers used by other " 3795 "formulae.\n"); 3796 3797 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) { 3798 LSRUse &LU = Uses[LUIdx]; 3799 bool Any = false; 3800 for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) { 3801 Formula &F = LU.Formulae[i]; 3802 // Look for a formula with a constant or GV in a register. If the use 3803 // also has a formula with that same value in an immediate field, 3804 // delete the one that uses a register. 3805 for (SmallVectorImpl<const SCEV *>::const_iterator 3806 I = F.BaseRegs.begin(), E = F.BaseRegs.end(); I != E; ++I) { 3807 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(*I)) { 3808 Formula NewF = F; 3809 NewF.BaseOffset += C->getValue()->getSExtValue(); 3810 NewF.BaseRegs.erase(NewF.BaseRegs.begin() + 3811 (I - F.BaseRegs.begin())); 3812 if (LU.HasFormulaWithSameRegs(NewF)) { 3813 DEBUG(dbgs() << " Deleting "; F.print(dbgs()); dbgs() << '\n'); 3814 LU.DeleteFormula(F); 3815 --i; 3816 --e; 3817 Any = true; 3818 break; 3819 } 3820 } else if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(*I)) { 3821 if (GlobalValue *GV = dyn_cast<GlobalValue>(U->getValue())) 3822 if (!F.BaseGV) { 3823 Formula NewF = F; 3824 NewF.BaseGV = GV; 3825 NewF.BaseRegs.erase(NewF.BaseRegs.begin() + 3826 (I - F.BaseRegs.begin())); 3827 if (LU.HasFormulaWithSameRegs(NewF)) { 3828 DEBUG(dbgs() << " Deleting "; F.print(dbgs()); 3829 dbgs() << '\n'); 3830 LU.DeleteFormula(F); 3831 --i; 3832 --e; 3833 Any = true; 3834 break; 3835 } 3836 } 3837 } 3838 } 3839 } 3840 if (Any) 3841 LU.RecomputeRegs(LUIdx, RegUses); 3842 } 3843 3844 DEBUG(dbgs() << "After pre-selection:\n"; 3845 print_uses(dbgs())); 3846 } 3847 } 3848 3849 /// NarrowSearchSpaceByCollapsingUnrolledCode - When there are many registers 3850 /// for expressions like A, A+1, A+2, etc., allocate a single register for 3851 /// them. 3852 void LSRInstance::NarrowSearchSpaceByCollapsingUnrolledCode() { 3853 if (EstimateSearchSpaceComplexity() < ComplexityLimit) 3854 return; 3855 3856 DEBUG(dbgs() << "The search space is too complex.\n" 3857 "Narrowing the search space by assuming that uses separated " 3858 "by a constant offset will use the same registers.\n"); 3859 3860 // This is especially useful for unrolled loops. 3861 3862 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) { 3863 LSRUse &LU = Uses[LUIdx]; 3864 for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(), 3865 E = LU.Formulae.end(); I != E; ++I) { 3866 const Formula &F = *I; 3867 if (F.BaseOffset == 0 || F.Scale != 0) 3868 continue; 3869 3870 LSRUse *LUThatHas = FindUseWithSimilarFormula(F, LU); 3871 if (!LUThatHas) 3872 continue; 3873 3874 if (!reconcileNewOffset(*LUThatHas, F.BaseOffset, /*HasBaseReg=*/ false, 3875 LU.Kind, LU.AccessTy)) 3876 continue; 3877 3878 DEBUG(dbgs() << " Deleting use "; LU.print(dbgs()); dbgs() << '\n'); 3879 3880 LUThatHas->AllFixupsOutsideLoop &= LU.AllFixupsOutsideLoop; 3881 3882 // Update the relocs to reference the new use. 3883 for (SmallVectorImpl<LSRFixup>::iterator I = Fixups.begin(), 3884 E = Fixups.end(); I != E; ++I) { 3885 LSRFixup &Fixup = *I; 3886 if (Fixup.LUIdx == LUIdx) { 3887 Fixup.LUIdx = LUThatHas - &Uses.front(); 3888 Fixup.Offset += F.BaseOffset; 3889 // Add the new offset to LUThatHas' offset list. 3890 if (LUThatHas->Offsets.back() != Fixup.Offset) { 3891 LUThatHas->Offsets.push_back(Fixup.Offset); 3892 if (Fixup.Offset > LUThatHas->MaxOffset) 3893 LUThatHas->MaxOffset = Fixup.Offset; 3894 if (Fixup.Offset < LUThatHas->MinOffset) 3895 LUThatHas->MinOffset = Fixup.Offset; 3896 } 3897 DEBUG(dbgs() << "New fixup has offset " << Fixup.Offset << '\n'); 3898 } 3899 if (Fixup.LUIdx == NumUses-1) 3900 Fixup.LUIdx = LUIdx; 3901 } 3902 3903 // Delete formulae from the new use which are no longer legal. 3904 bool Any = false; 3905 for (size_t i = 0, e = LUThatHas->Formulae.size(); i != e; ++i) { 3906 Formula &F = LUThatHas->Formulae[i]; 3907 if (!isLegalUse(TTI, LUThatHas->MinOffset, LUThatHas->MaxOffset, 3908 LUThatHas->Kind, LUThatHas->AccessTy, F)) { 3909 DEBUG(dbgs() << " Deleting "; F.print(dbgs()); 3910 dbgs() << '\n'); 3911 LUThatHas->DeleteFormula(F); 3912 --i; 3913 --e; 3914 Any = true; 3915 } 3916 } 3917 3918 if (Any) 3919 LUThatHas->RecomputeRegs(LUThatHas - &Uses.front(), RegUses); 3920 3921 // Delete the old use. 3922 DeleteUse(LU, LUIdx); 3923 --LUIdx; 3924 --NumUses; 3925 break; 3926 } 3927 } 3928 3929 DEBUG(dbgs() << "After pre-selection:\n"; print_uses(dbgs())); 3930 } 3931 3932 /// NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters - Call 3933 /// FilterOutUndesirableDedicatedRegisters again, if necessary, now that 3934 /// we've done more filtering, as it may be able to find more formulae to 3935 /// eliminate. 3936 void LSRInstance::NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters(){ 3937 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) { 3938 DEBUG(dbgs() << "The search space is too complex.\n"); 3939 3940 DEBUG(dbgs() << "Narrowing the search space by re-filtering out " 3941 "undesirable dedicated registers.\n"); 3942 3943 FilterOutUndesirableDedicatedRegisters(); 3944 3945 DEBUG(dbgs() << "After pre-selection:\n"; 3946 print_uses(dbgs())); 3947 } 3948 } 3949 3950 /// NarrowSearchSpaceByPickingWinnerRegs - Pick a register which seems likely 3951 /// to be profitable, and then in any use which has any reference to that 3952 /// register, delete all formulae which do not reference that register. 3953 void LSRInstance::NarrowSearchSpaceByPickingWinnerRegs() { 3954 // With all other options exhausted, loop until the system is simple 3955 // enough to handle. 3956 SmallPtrSet<const SCEV *, 4> Taken; 3957 while (EstimateSearchSpaceComplexity() >= ComplexityLimit) { 3958 // Ok, we have too many of formulae on our hands to conveniently handle. 3959 // Use a rough heuristic to thin out the list. 3960 DEBUG(dbgs() << "The search space is too complex.\n"); 3961 3962 // Pick the register which is used by the most LSRUses, which is likely 3963 // to be a good reuse register candidate. 3964 const SCEV *Best = 0; 3965 unsigned BestNum = 0; 3966 for (RegUseTracker::const_iterator I = RegUses.begin(), E = RegUses.end(); 3967 I != E; ++I) { 3968 const SCEV *Reg = *I; 3969 if (Taken.count(Reg)) 3970 continue; 3971 if (!Best) 3972 Best = Reg; 3973 else { 3974 unsigned Count = RegUses.getUsedByIndices(Reg).count(); 3975 if (Count > BestNum) { 3976 Best = Reg; 3977 BestNum = Count; 3978 } 3979 } 3980 } 3981 3982 DEBUG(dbgs() << "Narrowing the search space by assuming " << *Best 3983 << " will yield profitable reuse.\n"); 3984 Taken.insert(Best); 3985 3986 // In any use with formulae which references this register, delete formulae 3987 // which don't reference it. 3988 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) { 3989 LSRUse &LU = Uses[LUIdx]; 3990 if (!LU.Regs.count(Best)) continue; 3991 3992 bool Any = false; 3993 for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) { 3994 Formula &F = LU.Formulae[i]; 3995 if (!F.referencesReg(Best)) { 3996 DEBUG(dbgs() << " Deleting "; F.print(dbgs()); dbgs() << '\n'); 3997 LU.DeleteFormula(F); 3998 --e; 3999 --i; 4000 Any = true; 4001 assert(e != 0 && "Use has no formulae left! Is Regs inconsistent?"); 4002 continue; 4003 } 4004 } 4005 4006 if (Any) 4007 LU.RecomputeRegs(LUIdx, RegUses); 4008 } 4009 4010 DEBUG(dbgs() << "After pre-selection:\n"; 4011 print_uses(dbgs())); 4012 } 4013 } 4014 4015 /// NarrowSearchSpaceUsingHeuristics - If there are an extraordinary number of 4016 /// formulae to choose from, use some rough heuristics to prune down the number 4017 /// of formulae. This keeps the main solver from taking an extraordinary amount 4018 /// of time in some worst-case scenarios. 4019 void LSRInstance::NarrowSearchSpaceUsingHeuristics() { 4020 NarrowSearchSpaceByDetectingSupersets(); 4021 NarrowSearchSpaceByCollapsingUnrolledCode(); 4022 NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters(); 4023 NarrowSearchSpaceByPickingWinnerRegs(); 4024 } 4025 4026 /// SolveRecurse - This is the recursive solver. 4027 void LSRInstance::SolveRecurse(SmallVectorImpl<const Formula *> &Solution, 4028 Cost &SolutionCost, 4029 SmallVectorImpl<const Formula *> &Workspace, 4030 const Cost &CurCost, 4031 const SmallPtrSet<const SCEV *, 16> &CurRegs, 4032 DenseSet<const SCEV *> &VisitedRegs) const { 4033 // Some ideas: 4034 // - prune more: 4035 // - use more aggressive filtering 4036 // - sort the formula so that the most profitable solutions are found first 4037 // - sort the uses too 4038 // - search faster: 4039 // - don't compute a cost, and then compare. compare while computing a cost 4040 // and bail early. 4041 // - track register sets with SmallBitVector 4042 4043 const LSRUse &LU = Uses[Workspace.size()]; 4044 4045 // If this use references any register that's already a part of the 4046 // in-progress solution, consider it a requirement that a formula must 4047 // reference that register in order to be considered. This prunes out 4048 // unprofitable searching. 4049 SmallSetVector<const SCEV *, 4> ReqRegs; 4050 for (SmallPtrSet<const SCEV *, 16>::const_iterator I = CurRegs.begin(), 4051 E = CurRegs.end(); I != E; ++I) 4052 if (LU.Regs.count(*I)) 4053 ReqRegs.insert(*I); 4054 4055 SmallPtrSet<const SCEV *, 16> NewRegs; 4056 Cost NewCost; 4057 for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(), 4058 E = LU.Formulae.end(); I != E; ++I) { 4059 const Formula &F = *I; 4060 4061 // Ignore formulae which do not use any of the required registers. 4062 bool SatisfiedReqReg = true; 4063 for (SmallSetVector<const SCEV *, 4>::const_iterator J = ReqRegs.begin(), 4064 JE = ReqRegs.end(); J != JE; ++J) { 4065 const SCEV *Reg = *J; 4066 if ((!F.ScaledReg || F.ScaledReg != Reg) && 4067 std::find(F.BaseRegs.begin(), F.BaseRegs.end(), Reg) == 4068 F.BaseRegs.end()) { 4069 SatisfiedReqReg = false; 4070 break; 4071 } 4072 } 4073 if (!SatisfiedReqReg) { 4074 // If none of the formulae satisfied the required registers, then we could 4075 // clear ReqRegs and try again. Currently, we simply give up in this case. 4076 continue; 4077 } 4078 4079 // Evaluate the cost of the current formula. If it's already worse than 4080 // the current best, prune the search at that point. 4081 NewCost = CurCost; 4082 NewRegs = CurRegs; 4083 NewCost.RateFormula(F, NewRegs, VisitedRegs, L, LU.Offsets, SE, DT); 4084 if (NewCost < SolutionCost) { 4085 Workspace.push_back(&F); 4086 if (Workspace.size() != Uses.size()) { 4087 SolveRecurse(Solution, SolutionCost, Workspace, NewCost, 4088 NewRegs, VisitedRegs); 4089 if (F.getNumRegs() == 1 && Workspace.size() == 1) 4090 VisitedRegs.insert(F.ScaledReg ? F.ScaledReg : F.BaseRegs[0]); 4091 } else { 4092 DEBUG(dbgs() << "New best at "; NewCost.print(dbgs()); 4093 dbgs() << ".\n Regs:"; 4094 for (SmallPtrSet<const SCEV *, 16>::const_iterator 4095 I = NewRegs.begin(), E = NewRegs.end(); I != E; ++I) 4096 dbgs() << ' ' << **I; 4097 dbgs() << '\n'); 4098 4099 SolutionCost = NewCost; 4100 Solution = Workspace; 4101 } 4102 Workspace.pop_back(); 4103 } 4104 } 4105 } 4106 4107 /// Solve - Choose one formula from each use. Return the results in the given 4108 /// Solution vector. 4109 void LSRInstance::Solve(SmallVectorImpl<const Formula *> &Solution) const { 4110 SmallVector<const Formula *, 8> Workspace; 4111 Cost SolutionCost; 4112 SolutionCost.Loose(); 4113 Cost CurCost; 4114 SmallPtrSet<const SCEV *, 16> CurRegs; 4115 DenseSet<const SCEV *> VisitedRegs; 4116 Workspace.reserve(Uses.size()); 4117 4118 // SolveRecurse does all the work. 4119 SolveRecurse(Solution, SolutionCost, Workspace, CurCost, 4120 CurRegs, VisitedRegs); 4121 if (Solution.empty()) { 4122 DEBUG(dbgs() << "\nNo Satisfactory Solution\n"); 4123 return; 4124 } 4125 4126 // Ok, we've now made all our decisions. 4127 DEBUG(dbgs() << "\n" 4128 "The chosen solution requires "; SolutionCost.print(dbgs()); 4129 dbgs() << ":\n"; 4130 for (size_t i = 0, e = Uses.size(); i != e; ++i) { 4131 dbgs() << " "; 4132 Uses[i].print(dbgs()); 4133 dbgs() << "\n" 4134 " "; 4135 Solution[i]->print(dbgs()); 4136 dbgs() << '\n'; 4137 }); 4138 4139 assert(Solution.size() == Uses.size() && "Malformed solution!"); 4140 } 4141 4142 /// HoistInsertPosition - Helper for AdjustInsertPositionForExpand. Climb up 4143 /// the dominator tree far as we can go while still being dominated by the 4144 /// input positions. This helps canonicalize the insert position, which 4145 /// encourages sharing. 4146 BasicBlock::iterator 4147 LSRInstance::HoistInsertPosition(BasicBlock::iterator IP, 4148 const SmallVectorImpl<Instruction *> &Inputs) 4149 const { 4150 for (;;) { 4151 const Loop *IPLoop = LI.getLoopFor(IP->getParent()); 4152 unsigned IPLoopDepth = IPLoop ? IPLoop->getLoopDepth() : 0; 4153 4154 BasicBlock *IDom; 4155 for (DomTreeNode *Rung = DT.getNode(IP->getParent()); ; ) { 4156 if (!Rung) return IP; 4157 Rung = Rung->getIDom(); 4158 if (!Rung) return IP; 4159 IDom = Rung->getBlock(); 4160 4161 // Don't climb into a loop though. 4162 const Loop *IDomLoop = LI.getLoopFor(IDom); 4163 unsigned IDomDepth = IDomLoop ? IDomLoop->getLoopDepth() : 0; 4164 if (IDomDepth <= IPLoopDepth && 4165 (IDomDepth != IPLoopDepth || IDomLoop == IPLoop)) 4166 break; 4167 } 4168 4169 bool AllDominate = true; 4170 Instruction *BetterPos = 0; 4171 Instruction *Tentative = IDom->getTerminator(); 4172 for (SmallVectorImpl<Instruction *>::const_iterator I = Inputs.begin(), 4173 E = Inputs.end(); I != E; ++I) { 4174 Instruction *Inst = *I; 4175 if (Inst == Tentative || !DT.dominates(Inst, Tentative)) { 4176 AllDominate = false; 4177 break; 4178 } 4179 // Attempt to find an insert position in the middle of the block, 4180 // instead of at the end, so that it can be used for other expansions. 4181 if (IDom == Inst->getParent() && 4182 (!BetterPos || !DT.dominates(Inst, BetterPos))) 4183 BetterPos = llvm::next(BasicBlock::iterator(Inst)); 4184 } 4185 if (!AllDominate) 4186 break; 4187 if (BetterPos) 4188 IP = BetterPos; 4189 else 4190 IP = Tentative; 4191 } 4192 4193 return IP; 4194 } 4195 4196 /// AdjustInsertPositionForExpand - Determine an input position which will be 4197 /// dominated by the operands and which will dominate the result. 4198 BasicBlock::iterator 4199 LSRInstance::AdjustInsertPositionForExpand(BasicBlock::iterator LowestIP, 4200 const LSRFixup &LF, 4201 const LSRUse &LU, 4202 SCEVExpander &Rewriter) const { 4203 // Collect some instructions which must be dominated by the 4204 // expanding replacement. These must be dominated by any operands that 4205 // will be required in the expansion. 4206 SmallVector<Instruction *, 4> Inputs; 4207 if (Instruction *I = dyn_cast<Instruction>(LF.OperandValToReplace)) 4208 Inputs.push_back(I); 4209 if (LU.Kind == LSRUse::ICmpZero) 4210 if (Instruction *I = 4211 dyn_cast<Instruction>(cast<ICmpInst>(LF.UserInst)->getOperand(1))) 4212 Inputs.push_back(I); 4213 if (LF.PostIncLoops.count(L)) { 4214 if (LF.isUseFullyOutsideLoop(L)) 4215 Inputs.push_back(L->getLoopLatch()->getTerminator()); 4216 else 4217 Inputs.push_back(IVIncInsertPos); 4218 } 4219 // The expansion must also be dominated by the increment positions of any 4220 // loops it for which it is using post-inc mode. 4221 for (PostIncLoopSet::const_iterator I = LF.PostIncLoops.begin(), 4222 E = LF.PostIncLoops.end(); I != E; ++I) { 4223 const Loop *PIL = *I; 4224 if (PIL == L) continue; 4225 4226 // Be dominated by the loop exit. 4227 SmallVector<BasicBlock *, 4> ExitingBlocks; 4228 PIL->getExitingBlocks(ExitingBlocks); 4229 if (!ExitingBlocks.empty()) { 4230 BasicBlock *BB = ExitingBlocks[0]; 4231 for (unsigned i = 1, e = ExitingBlocks.size(); i != e; ++i) 4232 BB = DT.findNearestCommonDominator(BB, ExitingBlocks[i]); 4233 Inputs.push_back(BB->getTerminator()); 4234 } 4235 } 4236 4237 assert(!isa<PHINode>(LowestIP) && !isa<LandingPadInst>(LowestIP) 4238 && !isa<DbgInfoIntrinsic>(LowestIP) && 4239 "Insertion point must be a normal instruction"); 4240 4241 // Then, climb up the immediate dominator tree as far as we can go while 4242 // still being dominated by the input positions. 4243 BasicBlock::iterator IP = HoistInsertPosition(LowestIP, Inputs); 4244 4245 // Don't insert instructions before PHI nodes. 4246 while (isa<PHINode>(IP)) ++IP; 4247 4248 // Ignore landingpad instructions. 4249 while (isa<LandingPadInst>(IP)) ++IP; 4250 4251 // Ignore debug intrinsics. 4252 while (isa<DbgInfoIntrinsic>(IP)) ++IP; 4253 4254 // Set IP below instructions recently inserted by SCEVExpander. This keeps the 4255 // IP consistent across expansions and allows the previously inserted 4256 // instructions to be reused by subsequent expansion. 4257 while (Rewriter.isInsertedInstruction(IP) && IP != LowestIP) ++IP; 4258 4259 return IP; 4260 } 4261 4262 /// Expand - Emit instructions for the leading candidate expression for this 4263 /// LSRUse (this is called "expanding"). 4264 Value *LSRInstance::Expand(const LSRFixup &LF, 4265 const Formula &F, 4266 BasicBlock::iterator IP, 4267 SCEVExpander &Rewriter, 4268 SmallVectorImpl<WeakVH> &DeadInsts) const { 4269 const LSRUse &LU = Uses[LF.LUIdx]; 4270 4271 // Determine an input position which will be dominated by the operands and 4272 // which will dominate the result. 4273 IP = AdjustInsertPositionForExpand(IP, LF, LU, Rewriter); 4274 4275 // Inform the Rewriter if we have a post-increment use, so that it can 4276 // perform an advantageous expansion. 4277 Rewriter.setPostInc(LF.PostIncLoops); 4278 4279 // This is the type that the user actually needs. 4280 Type *OpTy = LF.OperandValToReplace->getType(); 4281 // This will be the type that we'll initially expand to. 4282 Type *Ty = F.getType(); 4283 if (!Ty) 4284 // No type known; just expand directly to the ultimate type. 4285 Ty = OpTy; 4286 else if (SE.getEffectiveSCEVType(Ty) == SE.getEffectiveSCEVType(OpTy)) 4287 // Expand directly to the ultimate type if it's the right size. 4288 Ty = OpTy; 4289 // This is the type to do integer arithmetic in. 4290 Type *IntTy = SE.getEffectiveSCEVType(Ty); 4291 4292 // Build up a list of operands to add together to form the full base. 4293 SmallVector<const SCEV *, 8> Ops; 4294 4295 // Expand the BaseRegs portion. 4296 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(), 4297 E = F.BaseRegs.end(); I != E; ++I) { 4298 const SCEV *Reg = *I; 4299 assert(!Reg->isZero() && "Zero allocated in a base register!"); 4300 4301 // If we're expanding for a post-inc user, make the post-inc adjustment. 4302 PostIncLoopSet &Loops = const_cast<PostIncLoopSet &>(LF.PostIncLoops); 4303 Reg = TransformForPostIncUse(Denormalize, Reg, 4304 LF.UserInst, LF.OperandValToReplace, 4305 Loops, SE, DT); 4306 4307 Ops.push_back(SE.getUnknown(Rewriter.expandCodeFor(Reg, 0, IP))); 4308 } 4309 4310 // Expand the ScaledReg portion. 4311 Value *ICmpScaledV = 0; 4312 if (F.Scale != 0) { 4313 const SCEV *ScaledS = F.ScaledReg; 4314 4315 // If we're expanding for a post-inc user, make the post-inc adjustment. 4316 PostIncLoopSet &Loops = const_cast<PostIncLoopSet &>(LF.PostIncLoops); 4317 ScaledS = TransformForPostIncUse(Denormalize, ScaledS, 4318 LF.UserInst, LF.OperandValToReplace, 4319 Loops, SE, DT); 4320 4321 if (LU.Kind == LSRUse::ICmpZero) { 4322 // An interesting way of "folding" with an icmp is to use a negated 4323 // scale, which we'll implement by inserting it into the other operand 4324 // of the icmp. 4325 assert(F.Scale == -1 && 4326 "The only scale supported by ICmpZero uses is -1!"); 4327 ICmpScaledV = Rewriter.expandCodeFor(ScaledS, 0, IP); 4328 } else { 4329 // Otherwise just expand the scaled register and an explicit scale, 4330 // which is expected to be matched as part of the address. 4331 4332 // Flush the operand list to suppress SCEVExpander hoisting address modes. 4333 if (!Ops.empty() && LU.Kind == LSRUse::Address) { 4334 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP); 4335 Ops.clear(); 4336 Ops.push_back(SE.getUnknown(FullV)); 4337 } 4338 ScaledS = SE.getUnknown(Rewriter.expandCodeFor(ScaledS, 0, IP)); 4339 ScaledS = SE.getMulExpr(ScaledS, 4340 SE.getConstant(ScaledS->getType(), F.Scale)); 4341 Ops.push_back(ScaledS); 4342 } 4343 } 4344 4345 // Expand the GV portion. 4346 if (F.BaseGV) { 4347 // Flush the operand list to suppress SCEVExpander hoisting. 4348 if (!Ops.empty()) { 4349 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP); 4350 Ops.clear(); 4351 Ops.push_back(SE.getUnknown(FullV)); 4352 } 4353 Ops.push_back(SE.getUnknown(F.BaseGV)); 4354 } 4355 4356 // Flush the operand list to suppress SCEVExpander hoisting of both folded and 4357 // unfolded offsets. LSR assumes they both live next to their uses. 4358 if (!Ops.empty()) { 4359 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP); 4360 Ops.clear(); 4361 Ops.push_back(SE.getUnknown(FullV)); 4362 } 4363 4364 // Expand the immediate portion. 4365 int64_t Offset = (uint64_t)F.BaseOffset + LF.Offset; 4366 if (Offset != 0) { 4367 if (LU.Kind == LSRUse::ICmpZero) { 4368 // The other interesting way of "folding" with an ICmpZero is to use a 4369 // negated immediate. 4370 if (!ICmpScaledV) 4371 ICmpScaledV = ConstantInt::get(IntTy, -(uint64_t)Offset); 4372 else { 4373 Ops.push_back(SE.getUnknown(ICmpScaledV)); 4374 ICmpScaledV = ConstantInt::get(IntTy, Offset); 4375 } 4376 } else { 4377 // Just add the immediate values. These again are expected to be matched 4378 // as part of the address. 4379 Ops.push_back(SE.getUnknown(ConstantInt::getSigned(IntTy, Offset))); 4380 } 4381 } 4382 4383 // Expand the unfolded offset portion. 4384 int64_t UnfoldedOffset = F.UnfoldedOffset; 4385 if (UnfoldedOffset != 0) { 4386 // Just add the immediate values. 4387 Ops.push_back(SE.getUnknown(ConstantInt::getSigned(IntTy, 4388 UnfoldedOffset))); 4389 } 4390 4391 // Emit instructions summing all the operands. 4392 const SCEV *FullS = Ops.empty() ? 4393 SE.getConstant(IntTy, 0) : 4394 SE.getAddExpr(Ops); 4395 Value *FullV = Rewriter.expandCodeFor(FullS, Ty, IP); 4396 4397 // We're done expanding now, so reset the rewriter. 4398 Rewriter.clearPostInc(); 4399 4400 // An ICmpZero Formula represents an ICmp which we're handling as a 4401 // comparison against zero. Now that we've expanded an expression for that 4402 // form, update the ICmp's other operand. 4403 if (LU.Kind == LSRUse::ICmpZero) { 4404 ICmpInst *CI = cast<ICmpInst>(LF.UserInst); 4405 DeadInsts.push_back(CI->getOperand(1)); 4406 assert(!F.BaseGV && "ICmp does not support folding a global value and " 4407 "a scale at the same time!"); 4408 if (F.Scale == -1) { 4409 if (ICmpScaledV->getType() != OpTy) { 4410 Instruction *Cast = 4411 CastInst::Create(CastInst::getCastOpcode(ICmpScaledV, false, 4412 OpTy, false), 4413 ICmpScaledV, OpTy, "tmp", CI); 4414 ICmpScaledV = Cast; 4415 } 4416 CI->setOperand(1, ICmpScaledV); 4417 } else { 4418 assert(F.Scale == 0 && 4419 "ICmp does not support folding a global value and " 4420 "a scale at the same time!"); 4421 Constant *C = ConstantInt::getSigned(SE.getEffectiveSCEVType(OpTy), 4422 -(uint64_t)Offset); 4423 if (C->getType() != OpTy) 4424 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false, 4425 OpTy, false), 4426 C, OpTy); 4427 4428 CI->setOperand(1, C); 4429 } 4430 } 4431 4432 return FullV; 4433 } 4434 4435 /// RewriteForPHI - Helper for Rewrite. PHI nodes are special because the use 4436 /// of their operands effectively happens in their predecessor blocks, so the 4437 /// expression may need to be expanded in multiple places. 4438 void LSRInstance::RewriteForPHI(PHINode *PN, 4439 const LSRFixup &LF, 4440 const Formula &F, 4441 SCEVExpander &Rewriter, 4442 SmallVectorImpl<WeakVH> &DeadInsts, 4443 Pass *P) const { 4444 DenseMap<BasicBlock *, Value *> Inserted; 4445 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) 4446 if (PN->getIncomingValue(i) == LF.OperandValToReplace) { 4447 BasicBlock *BB = PN->getIncomingBlock(i); 4448 4449 // If this is a critical edge, split the edge so that we do not insert 4450 // the code on all predecessor/successor paths. We do this unless this 4451 // is the canonical backedge for this loop, which complicates post-inc 4452 // users. 4453 if (e != 1 && BB->getTerminator()->getNumSuccessors() > 1 && 4454 !isa<IndirectBrInst>(BB->getTerminator())) { 4455 BasicBlock *Parent = PN->getParent(); 4456 Loop *PNLoop = LI.getLoopFor(Parent); 4457 if (!PNLoop || Parent != PNLoop->getHeader()) { 4458 // Split the critical edge. 4459 BasicBlock *NewBB = 0; 4460 if (!Parent->isLandingPad()) { 4461 NewBB = SplitCriticalEdge(BB, Parent, P, 4462 /*MergeIdenticalEdges=*/true, 4463 /*DontDeleteUselessPhis=*/true); 4464 } else { 4465 SmallVector<BasicBlock*, 2> NewBBs; 4466 SplitLandingPadPredecessors(Parent, BB, "", "", P, NewBBs); 4467 NewBB = NewBBs[0]; 4468 } 4469 // If NewBB==NULL, then SplitCriticalEdge refused to split because all 4470 // phi predecessors are identical. The simple thing to do is skip 4471 // splitting in this case rather than complicate the API. 4472 if (NewBB) { 4473 // If PN is outside of the loop and BB is in the loop, we want to 4474 // move the block to be immediately before the PHI block, not 4475 // immediately after BB. 4476 if (L->contains(BB) && !L->contains(PN)) 4477 NewBB->moveBefore(PN->getParent()); 4478 4479 // Splitting the edge can reduce the number of PHI entries we have. 4480 e = PN->getNumIncomingValues(); 4481 BB = NewBB; 4482 i = PN->getBasicBlockIndex(BB); 4483 } 4484 } 4485 } 4486 4487 std::pair<DenseMap<BasicBlock *, Value *>::iterator, bool> Pair = 4488 Inserted.insert(std::make_pair(BB, static_cast<Value *>(0))); 4489 if (!Pair.second) 4490 PN->setIncomingValue(i, Pair.first->second); 4491 else { 4492 Value *FullV = Expand(LF, F, BB->getTerminator(), Rewriter, DeadInsts); 4493 4494 // If this is reuse-by-noop-cast, insert the noop cast. 4495 Type *OpTy = LF.OperandValToReplace->getType(); 4496 if (FullV->getType() != OpTy) 4497 FullV = 4498 CastInst::Create(CastInst::getCastOpcode(FullV, false, 4499 OpTy, false), 4500 FullV, LF.OperandValToReplace->getType(), 4501 "tmp", BB->getTerminator()); 4502 4503 PN->setIncomingValue(i, FullV); 4504 Pair.first->second = FullV; 4505 } 4506 } 4507 } 4508 4509 /// Rewrite - Emit instructions for the leading candidate expression for this 4510 /// LSRUse (this is called "expanding"), and update the UserInst to reference 4511 /// the newly expanded value. 4512 void LSRInstance::Rewrite(const LSRFixup &LF, 4513 const Formula &F, 4514 SCEVExpander &Rewriter, 4515 SmallVectorImpl<WeakVH> &DeadInsts, 4516 Pass *P) const { 4517 // First, find an insertion point that dominates UserInst. For PHI nodes, 4518 // find the nearest block which dominates all the relevant uses. 4519 if (PHINode *PN = dyn_cast<PHINode>(LF.UserInst)) { 4520 RewriteForPHI(PN, LF, F, Rewriter, DeadInsts, P); 4521 } else { 4522 Value *FullV = Expand(LF, F, LF.UserInst, Rewriter, DeadInsts); 4523 4524 // If this is reuse-by-noop-cast, insert the noop cast. 4525 Type *OpTy = LF.OperandValToReplace->getType(); 4526 if (FullV->getType() != OpTy) { 4527 Instruction *Cast = 4528 CastInst::Create(CastInst::getCastOpcode(FullV, false, OpTy, false), 4529 FullV, OpTy, "tmp", LF.UserInst); 4530 FullV = Cast; 4531 } 4532 4533 // Update the user. ICmpZero is handled specially here (for now) because 4534 // Expand may have updated one of the operands of the icmp already, and 4535 // its new value may happen to be equal to LF.OperandValToReplace, in 4536 // which case doing replaceUsesOfWith leads to replacing both operands 4537 // with the same value. TODO: Reorganize this. 4538 if (Uses[LF.LUIdx].Kind == LSRUse::ICmpZero) 4539 LF.UserInst->setOperand(0, FullV); 4540 else 4541 LF.UserInst->replaceUsesOfWith(LF.OperandValToReplace, FullV); 4542 } 4543 4544 DeadInsts.push_back(LF.OperandValToReplace); 4545 } 4546 4547 /// ImplementSolution - Rewrite all the fixup locations with new values, 4548 /// following the chosen solution. 4549 void 4550 LSRInstance::ImplementSolution(const SmallVectorImpl<const Formula *> &Solution, 4551 Pass *P) { 4552 // Keep track of instructions we may have made dead, so that 4553 // we can remove them after we are done working. 4554 SmallVector<WeakVH, 16> DeadInsts; 4555 4556 SCEVExpander Rewriter(SE, "lsr"); 4557 #ifndef NDEBUG 4558 Rewriter.setDebugType(DEBUG_TYPE); 4559 #endif 4560 Rewriter.disableCanonicalMode(); 4561 Rewriter.enableLSRMode(); 4562 Rewriter.setIVIncInsertPos(L, IVIncInsertPos); 4563 4564 // Mark phi nodes that terminate chains so the expander tries to reuse them. 4565 for (SmallVectorImpl<IVChain>::const_iterator ChainI = IVChainVec.begin(), 4566 ChainE = IVChainVec.end(); ChainI != ChainE; ++ChainI) { 4567 if (PHINode *PN = dyn_cast<PHINode>(ChainI->tailUserInst())) 4568 Rewriter.setChainedPhi(PN); 4569 } 4570 4571 // Expand the new value definitions and update the users. 4572 for (SmallVectorImpl<LSRFixup>::const_iterator I = Fixups.begin(), 4573 E = Fixups.end(); I != E; ++I) { 4574 const LSRFixup &Fixup = *I; 4575 4576 Rewrite(Fixup, *Solution[Fixup.LUIdx], Rewriter, DeadInsts, P); 4577 4578 Changed = true; 4579 } 4580 4581 for (SmallVectorImpl<IVChain>::const_iterator ChainI = IVChainVec.begin(), 4582 ChainE = IVChainVec.end(); ChainI != ChainE; ++ChainI) { 4583 GenerateIVChain(*ChainI, Rewriter, DeadInsts); 4584 Changed = true; 4585 } 4586 // Clean up after ourselves. This must be done before deleting any 4587 // instructions. 4588 Rewriter.clear(); 4589 4590 Changed |= DeleteTriviallyDeadInstructions(DeadInsts); 4591 } 4592 4593 LSRInstance::LSRInstance(Loop *L, Pass *P) 4594 : IU(P->getAnalysis<IVUsers>()), SE(P->getAnalysis<ScalarEvolution>()), 4595 DT(P->getAnalysis<DominatorTree>()), LI(P->getAnalysis<LoopInfo>()), 4596 TTI(P->getAnalysis<TargetTransformInfo>()), L(L), Changed(false), 4597 IVIncInsertPos(0) { 4598 // If LoopSimplify form is not available, stay out of trouble. 4599 if (!L->isLoopSimplifyForm()) 4600 return; 4601 4602 // If there's no interesting work to be done, bail early. 4603 if (IU.empty()) return; 4604 4605 // If there's too much analysis to be done, bail early. We won't be able to 4606 // model the problem anyway. 4607 unsigned NumUsers = 0; 4608 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI) { 4609 if (++NumUsers > MaxIVUsers) { 4610 DEBUG(dbgs() << "LSR skipping loop, too many IV Users in " << *L 4611 << "\n"); 4612 return; 4613 } 4614 } 4615 4616 #ifndef NDEBUG 4617 // All dominating loops must have preheaders, or SCEVExpander may not be able 4618 // to materialize an AddRecExpr whose Start is an outer AddRecExpr. 4619 // 4620 // IVUsers analysis should only create users that are dominated by simple loop 4621 // headers. Since this loop should dominate all of its users, its user list 4622 // should be empty if this loop itself is not within a simple loop nest. 4623 for (DomTreeNode *Rung = DT.getNode(L->getLoopPreheader()); 4624 Rung; Rung = Rung->getIDom()) { 4625 BasicBlock *BB = Rung->getBlock(); 4626 const Loop *DomLoop = LI.getLoopFor(BB); 4627 if (DomLoop && DomLoop->getHeader() == BB) { 4628 assert(DomLoop->getLoopPreheader() && "LSR needs a simplified loop nest"); 4629 } 4630 } 4631 #endif // DEBUG 4632 4633 DEBUG(dbgs() << "\nLSR on loop "; 4634 WriteAsOperand(dbgs(), L->getHeader(), /*PrintType=*/false); 4635 dbgs() << ":\n"); 4636 4637 // First, perform some low-level loop optimizations. 4638 OptimizeShadowIV(); 4639 OptimizeLoopTermCond(); 4640 4641 // If loop preparation eliminates all interesting IV users, bail. 4642 if (IU.empty()) return; 4643 4644 // Skip nested loops until we can model them better with formulae. 4645 if (!L->empty()) { 4646 DEBUG(dbgs() << "LSR skipping outer loop " << *L << "\n"); 4647 return; 4648 } 4649 4650 // Start collecting data and preparing for the solver. 4651 CollectChains(); 4652 CollectInterestingTypesAndFactors(); 4653 CollectFixupsAndInitialFormulae(); 4654 CollectLoopInvariantFixupsAndFormulae(); 4655 4656 assert(!Uses.empty() && "IVUsers reported at least one use"); 4657 DEBUG(dbgs() << "LSR found " << Uses.size() << " uses:\n"; 4658 print_uses(dbgs())); 4659 4660 // Now use the reuse data to generate a bunch of interesting ways 4661 // to formulate the values needed for the uses. 4662 GenerateAllReuseFormulae(); 4663 4664 FilterOutUndesirableDedicatedRegisters(); 4665 NarrowSearchSpaceUsingHeuristics(); 4666 4667 SmallVector<const Formula *, 8> Solution; 4668 Solve(Solution); 4669 4670 // Release memory that is no longer needed. 4671 Factors.clear(); 4672 Types.clear(); 4673 RegUses.clear(); 4674 4675 if (Solution.empty()) 4676 return; 4677 4678 #ifndef NDEBUG 4679 // Formulae should be legal. 4680 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(), E = Uses.end(); 4681 I != E; ++I) { 4682 const LSRUse &LU = *I; 4683 for (SmallVectorImpl<Formula>::const_iterator J = LU.Formulae.begin(), 4684 JE = LU.Formulae.end(); 4685 J != JE; ++J) 4686 assert(isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy, 4687 *J) && "Illegal formula generated!"); 4688 }; 4689 #endif 4690 4691 // Now that we've decided what we want, make it so. 4692 ImplementSolution(Solution, P); 4693 } 4694 4695 void LSRInstance::print_factors_and_types(raw_ostream &OS) const { 4696 if (Factors.empty() && Types.empty()) return; 4697 4698 OS << "LSR has identified the following interesting factors and types: "; 4699 bool First = true; 4700 4701 for (SmallSetVector<int64_t, 8>::const_iterator 4702 I = Factors.begin(), E = Factors.end(); I != E; ++I) { 4703 if (!First) OS << ", "; 4704 First = false; 4705 OS << '*' << *I; 4706 } 4707 4708 for (SmallSetVector<Type *, 4>::const_iterator 4709 I = Types.begin(), E = Types.end(); I != E; ++I) { 4710 if (!First) OS << ", "; 4711 First = false; 4712 OS << '(' << **I << ')'; 4713 } 4714 OS << '\n'; 4715 } 4716 4717 void LSRInstance::print_fixups(raw_ostream &OS) const { 4718 OS << "LSR is examining the following fixup sites:\n"; 4719 for (SmallVectorImpl<LSRFixup>::const_iterator I = Fixups.begin(), 4720 E = Fixups.end(); I != E; ++I) { 4721 dbgs() << " "; 4722 I->print(OS); 4723 OS << '\n'; 4724 } 4725 } 4726 4727 void LSRInstance::print_uses(raw_ostream &OS) const { 4728 OS << "LSR is examining the following uses:\n"; 4729 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(), 4730 E = Uses.end(); I != E; ++I) { 4731 const LSRUse &LU = *I; 4732 dbgs() << " "; 4733 LU.print(OS); 4734 OS << '\n'; 4735 for (SmallVectorImpl<Formula>::const_iterator J = LU.Formulae.begin(), 4736 JE = LU.Formulae.end(); J != JE; ++J) { 4737 OS << " "; 4738 J->print(OS); 4739 OS << '\n'; 4740 } 4741 } 4742 } 4743 4744 void LSRInstance::print(raw_ostream &OS) const { 4745 print_factors_and_types(OS); 4746 print_fixups(OS); 4747 print_uses(OS); 4748 } 4749 4750 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) 4751 void LSRInstance::dump() const { 4752 print(errs()); errs() << '\n'; 4753 } 4754 #endif 4755 4756 namespace { 4757 4758 class LoopStrengthReduce : public LoopPass { 4759 public: 4760 static char ID; // Pass ID, replacement for typeid 4761 LoopStrengthReduce(); 4762 4763 private: 4764 bool runOnLoop(Loop *L, LPPassManager &LPM); 4765 void getAnalysisUsage(AnalysisUsage &AU) const; 4766 }; 4767 4768 } 4769 4770 char LoopStrengthReduce::ID = 0; 4771 INITIALIZE_PASS_BEGIN(LoopStrengthReduce, "loop-reduce", 4772 "Loop Strength Reduction", false, false) 4773 INITIALIZE_AG_DEPENDENCY(TargetTransformInfo) 4774 INITIALIZE_PASS_DEPENDENCY(DominatorTree) 4775 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution) 4776 INITIALIZE_PASS_DEPENDENCY(IVUsers) 4777 INITIALIZE_PASS_DEPENDENCY(LoopInfo) 4778 INITIALIZE_PASS_DEPENDENCY(LoopSimplify) 4779 INITIALIZE_PASS_END(LoopStrengthReduce, "loop-reduce", 4780 "Loop Strength Reduction", false, false) 4781 4782 4783 Pass *llvm::createLoopStrengthReducePass() { 4784 return new LoopStrengthReduce(); 4785 } 4786 4787 LoopStrengthReduce::LoopStrengthReduce() : LoopPass(ID) { 4788 initializeLoopStrengthReducePass(*PassRegistry::getPassRegistry()); 4789 } 4790 4791 void LoopStrengthReduce::getAnalysisUsage(AnalysisUsage &AU) const { 4792 // We split critical edges, so we change the CFG. However, we do update 4793 // many analyses if they are around. 4794 AU.addPreservedID(LoopSimplifyID); 4795 4796 AU.addRequired<LoopInfo>(); 4797 AU.addPreserved<LoopInfo>(); 4798 AU.addRequiredID(LoopSimplifyID); 4799 AU.addRequired<DominatorTree>(); 4800 AU.addPreserved<DominatorTree>(); 4801 AU.addRequired<ScalarEvolution>(); 4802 AU.addPreserved<ScalarEvolution>(); 4803 // Requiring LoopSimplify a second time here prevents IVUsers from running 4804 // twice, since LoopSimplify was invalidated by running ScalarEvolution. 4805 AU.addRequiredID(LoopSimplifyID); 4806 AU.addRequired<IVUsers>(); 4807 AU.addPreserved<IVUsers>(); 4808 AU.addRequired<TargetTransformInfo>(); 4809 } 4810 4811 bool LoopStrengthReduce::runOnLoop(Loop *L, LPPassManager & /*LPM*/) { 4812 bool Changed = false; 4813 4814 // Run the main LSR transformation. 4815 Changed |= LSRInstance(L, this).getChanged(); 4816 4817 // Remove any extra phis created by processing inner loops. 4818 Changed |= DeleteDeadPHIs(L->getHeader()); 4819 if (EnablePhiElim && L->isLoopSimplifyForm()) { 4820 SmallVector<WeakVH, 16> DeadInsts; 4821 SCEVExpander Rewriter(getAnalysis<ScalarEvolution>(), "lsr"); 4822 #ifndef NDEBUG 4823 Rewriter.setDebugType(DEBUG_TYPE); 4824 #endif 4825 unsigned numFolded = 4826 Rewriter.replaceCongruentIVs(L, &getAnalysis<DominatorTree>(), 4827 DeadInsts, 4828 &getAnalysis<TargetTransformInfo>()); 4829 if (numFolded) { 4830 Changed = true; 4831 DeleteTriviallyDeadInstructions(DeadInsts); 4832 DeleteDeadPHIs(L->getHeader()); 4833 } 4834 } 4835 return Changed; 4836 } 4837