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