LoopStrengthReduce.cpp revision 288943
1//===- LoopStrengthReduce.cpp - Strength Reduce IVs in Loops --------------===// 2// 3// The LLVM Compiler Infrastructure 4// 5// This file is distributed under the University of Illinois Open Source 6// License. See LICENSE.TXT for details. 7// 8//===----------------------------------------------------------------------===// 9// 10// This transformation analyzes and transforms the induction variables (and 11// computations derived from them) into forms suitable for efficient execution 12// on the target. 13// 14// This pass performs a strength reduction on array references inside loops that 15// have as one or more of their components the loop induction variable, it 16// rewrites expressions to take advantage of scaled-index addressing modes 17// available on the target, and it performs a variety of other optimizations 18// related to loop induction variables. 19// 20// Terminology note: this code has a lot of handling for "post-increment" or 21// "post-inc" users. This is not talking about post-increment addressing modes; 22// it is instead talking about code like this: 23// 24// %i = phi [ 0, %entry ], [ %i.next, %latch ] 25// ... 26// %i.next = add %i, 1 27// %c = icmp eq %i.next, %n 28// 29// The SCEV for %i is {0,+,1}<%L>. The SCEV for %i.next is {1,+,1}<%L>, however 30// it's useful to think about these as the same register, with some uses using 31// the value of the register before the add and some using it after. In this 32// example, the icmp is a post-increment user, since it uses %i.next, which is 33// the value of the induction variable after the increment. The other common 34// case of post-increment users is users outside the loop. 35// 36// TODO: More sophistication in the way Formulae are generated and filtered. 37// 38// TODO: Handle multiple loops at a time. 39// 40// TODO: Should the addressing mode BaseGV be changed to a ConstantExpr instead 41// of a GlobalValue? 42// 43// TODO: When truncation is free, truncate ICmp users' operands to make it a 44// smaller encoding (on x86 at least). 45// 46// TODO: When a negated register is used by an add (such as in a list of 47// multiple base registers, or as the increment expression in an addrec), 48// we may not actually need both reg and (-1 * reg) in registers; the 49// negation can be implemented by using a sub instead of an add. The 50// lack of support for taking this into consideration when making 51// register pressure decisions is partly worked around by the "Special" 52// use kind. 53// 54//===----------------------------------------------------------------------===// 55 56#include "llvm/Transforms/Scalar.h" 57#include "llvm/ADT/DenseSet.h" 58#include "llvm/ADT/Hashing.h" 59#include "llvm/ADT/STLExtras.h" 60#include "llvm/ADT/SetVector.h" 61#include "llvm/ADT/SmallBitVector.h" 62#include "llvm/Analysis/IVUsers.h" 63#include "llvm/Analysis/LoopPass.h" 64#include "llvm/Analysis/ScalarEvolutionExpander.h" 65#include "llvm/Analysis/TargetTransformInfo.h" 66#include "llvm/IR/Constants.h" 67#include "llvm/IR/DerivedTypes.h" 68#include "llvm/IR/Dominators.h" 69#include "llvm/IR/Instructions.h" 70#include "llvm/IR/IntrinsicInst.h" 71#include "llvm/IR/Module.h" 72#include "llvm/IR/ValueHandle.h" 73#include "llvm/Support/CommandLine.h" 74#include "llvm/Support/Debug.h" 75#include "llvm/Support/raw_ostream.h" 76#include "llvm/Transforms/Utils/BasicBlockUtils.h" 77#include "llvm/Transforms/Utils/Local.h" 78#include <algorithm> 79using namespace llvm; 80 81#define DEBUG_TYPE "loop-reduce" 82 83/// MaxIVUsers is an arbitrary threshold that provides an early opportunitiy for 84/// bail out. This threshold is far beyond the number of users that LSR can 85/// conceivably solve, so it should not affect generated code, but catches the 86/// worst cases before LSR burns too much compile time and stack space. 87static const unsigned MaxIVUsers = 200; 88 89// Temporary flag to cleanup congruent phis after LSR phi expansion. 90// It's currently disabled until we can determine whether it's truly useful or 91// not. The flag should be removed after the v3.0 release. 92// This is now needed for ivchains. 93static cl::opt<bool> EnablePhiElim( 94 "enable-lsr-phielim", cl::Hidden, cl::init(true), 95 cl::desc("Enable LSR phi elimination")); 96 97#ifndef NDEBUG 98// Stress test IV chain generation. 99static cl::opt<bool> StressIVChain( 100 "stress-ivchain", cl::Hidden, cl::init(false), 101 cl::desc("Stress test LSR IV chains")); 102#else 103static bool StressIVChain = false; 104#endif 105 106namespace { 107 108/// RegSortData - This class holds data which is used to order reuse candidates. 109class RegSortData { 110public: 111 /// UsedByIndices - This represents the set of LSRUse indices which reference 112 /// a particular register. 113 SmallBitVector UsedByIndices; 114 115 void print(raw_ostream &OS) const; 116 void dump() const; 117}; 118 119} 120 121void RegSortData::print(raw_ostream &OS) const { 122 OS << "[NumUses=" << UsedByIndices.count() << ']'; 123} 124 125#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) 126void RegSortData::dump() const { 127 print(errs()); errs() << '\n'; 128} 129#endif 130 131namespace { 132 133/// RegUseTracker - Map register candidates to information about how they are 134/// used. 135class RegUseTracker { 136 typedef DenseMap<const SCEV *, RegSortData> RegUsesTy; 137 138 RegUsesTy RegUsesMap; 139 SmallVector<const SCEV *, 16> RegSequence; 140 141public: 142 void CountRegister(const SCEV *Reg, size_t LUIdx); 143 void DropRegister(const SCEV *Reg, size_t LUIdx); 144 void SwapAndDropUse(size_t LUIdx, size_t LastLUIdx); 145 146 bool isRegUsedByUsesOtherThan(const SCEV *Reg, size_t LUIdx) const; 147 148 const SmallBitVector &getUsedByIndices(const SCEV *Reg) const; 149 150 void clear(); 151 152 typedef SmallVectorImpl<const SCEV *>::iterator iterator; 153 typedef SmallVectorImpl<const SCEV *>::const_iterator const_iterator; 154 iterator begin() { return RegSequence.begin(); } 155 iterator end() { return RegSequence.end(); } 156 const_iterator begin() const { return RegSequence.begin(); } 157 const_iterator end() const { return RegSequence.end(); } 158}; 159 160} 161 162void 163RegUseTracker::CountRegister(const SCEV *Reg, size_t LUIdx) { 164 std::pair<RegUsesTy::iterator, bool> Pair = 165 RegUsesMap.insert(std::make_pair(Reg, RegSortData())); 166 RegSortData &RSD = Pair.first->second; 167 if (Pair.second) 168 RegSequence.push_back(Reg); 169 RSD.UsedByIndices.resize(std::max(RSD.UsedByIndices.size(), LUIdx + 1)); 170 RSD.UsedByIndices.set(LUIdx); 171} 172 173void 174RegUseTracker::DropRegister(const SCEV *Reg, size_t LUIdx) { 175 RegUsesTy::iterator It = RegUsesMap.find(Reg); 176 assert(It != RegUsesMap.end()); 177 RegSortData &RSD = It->second; 178 assert(RSD.UsedByIndices.size() > LUIdx); 179 RSD.UsedByIndices.reset(LUIdx); 180} 181 182void 183RegUseTracker::SwapAndDropUse(size_t LUIdx, size_t LastLUIdx) { 184 assert(LUIdx <= LastLUIdx); 185 186 // Update RegUses. The data structure is not optimized for this purpose; 187 // we must iterate through it and update each of the bit vectors. 188 for (auto &Pair : RegUsesMap) { 189 SmallBitVector &UsedByIndices = Pair.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 197bool 198RegUseTracker::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 209const 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 215void RegUseTracker::clear() { 216 RegUsesMap.clear(); 217 RegSequence.clear(); 218} 219 220namespace { 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. 225struct Formula { 226 /// Global base address used for complex addressing. 227 GlobalValue *BaseGV; 228 229 /// Base offset for complex addressing. 230 int64_t BaseOffset; 231 232 /// Whether any complex addressing has a base register. 233 bool HasBaseReg; 234 235 /// The scale of any complex addressing. 236 int64_t Scale; 237 238 /// BaseRegs - The list of "base" registers for this use. When this is 239 /// non-empty. The canonical representation of a formula is 240 /// 1. BaseRegs.size > 1 implies ScaledReg != NULL and 241 /// 2. ScaledReg != NULL implies Scale != 1 || !BaseRegs.empty(). 242 /// #1 enforces that the scaled register is always used when at least two 243 /// registers are needed by the formula: e.g., reg1 + reg2 is reg1 + 1 * reg2. 244 /// #2 enforces that 1 * reg is reg. 245 /// This invariant can be temporarly broken while building a formula. 246 /// However, every formula inserted into the LSRInstance must be in canonical 247 /// form. 248 SmallVector<const SCEV *, 4> BaseRegs; 249 250 /// ScaledReg - The 'scaled' register for this use. This should be non-null 251 /// when Scale is not zero. 252 const SCEV *ScaledReg; 253 254 /// UnfoldedOffset - An additional constant offset which added near the 255 /// use. This requires a temporary register, but the offset itself can 256 /// live in an add immediate field rather than a register. 257 int64_t UnfoldedOffset; 258 259 Formula() 260 : BaseGV(nullptr), BaseOffset(0), HasBaseReg(false), Scale(0), 261 ScaledReg(nullptr), UnfoldedOffset(0) {} 262 263 void InitialMatch(const SCEV *S, Loop *L, ScalarEvolution &SE); 264 265 bool isCanonical() const; 266 267 void Canonicalize(); 268 269 bool Unscale(); 270 271 size_t getNumRegs() const; 272 Type *getType() const; 273 274 void DeleteBaseReg(const SCEV *&S); 275 276 bool referencesReg(const SCEV *S) const; 277 bool hasRegsUsedByUsesOtherThan(size_t LUIdx, 278 const RegUseTracker &RegUses) const; 279 280 void print(raw_ostream &OS) const; 281 void dump() const; 282}; 283 284} 285 286/// DoInitialMatch - Recursion helper for InitialMatch. 287static void DoInitialMatch(const SCEV *S, Loop *L, 288 SmallVectorImpl<const SCEV *> &Good, 289 SmallVectorImpl<const SCEV *> &Bad, 290 ScalarEvolution &SE) { 291 // Collect expressions which properly dominate the loop header. 292 if (SE.properlyDominates(S, L->getHeader())) { 293 Good.push_back(S); 294 return; 295 } 296 297 // Look at add operands. 298 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) { 299 for (const SCEV *S : Add->operands()) 300 DoInitialMatch(S, L, Good, Bad, SE); 301 return; 302 } 303 304 // Look at addrec operands. 305 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) 306 if (!AR->getStart()->isZero()) { 307 DoInitialMatch(AR->getStart(), L, Good, Bad, SE); 308 DoInitialMatch(SE.getAddRecExpr(SE.getConstant(AR->getType(), 0), 309 AR->getStepRecurrence(SE), 310 // FIXME: AR->getNoWrapFlags() 311 AR->getLoop(), SCEV::FlagAnyWrap), 312 L, Good, Bad, SE); 313 return; 314 } 315 316 // Handle a multiplication by -1 (negation) if it didn't fold. 317 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) 318 if (Mul->getOperand(0)->isAllOnesValue()) { 319 SmallVector<const SCEV *, 4> Ops(Mul->op_begin()+1, Mul->op_end()); 320 const SCEV *NewMul = SE.getMulExpr(Ops); 321 322 SmallVector<const SCEV *, 4> MyGood; 323 SmallVector<const SCEV *, 4> MyBad; 324 DoInitialMatch(NewMul, L, MyGood, MyBad, SE); 325 const SCEV *NegOne = SE.getSCEV(ConstantInt::getAllOnesValue( 326 SE.getEffectiveSCEVType(NewMul->getType()))); 327 for (const SCEV *S : MyGood) 328 Good.push_back(SE.getMulExpr(NegOne, S)); 329 for (const SCEV *S : MyBad) 330 Bad.push_back(SE.getMulExpr(NegOne, S)); 331 return; 332 } 333 334 // Ok, we can't do anything interesting. Just stuff the whole thing into a 335 // register and hope for the best. 336 Bad.push_back(S); 337} 338 339/// InitialMatch - Incorporate loop-variant parts of S into this Formula, 340/// attempting to keep all loop-invariant and loop-computable values in a 341/// single base register. 342void Formula::InitialMatch(const SCEV *S, Loop *L, ScalarEvolution &SE) { 343 SmallVector<const SCEV *, 4> Good; 344 SmallVector<const SCEV *, 4> Bad; 345 DoInitialMatch(S, L, Good, Bad, SE); 346 if (!Good.empty()) { 347 const SCEV *Sum = SE.getAddExpr(Good); 348 if (!Sum->isZero()) 349 BaseRegs.push_back(Sum); 350 HasBaseReg = true; 351 } 352 if (!Bad.empty()) { 353 const SCEV *Sum = SE.getAddExpr(Bad); 354 if (!Sum->isZero()) 355 BaseRegs.push_back(Sum); 356 HasBaseReg = true; 357 } 358 Canonicalize(); 359} 360 361/// \brief Check whether or not this formula statisfies the canonical 362/// representation. 363/// \see Formula::BaseRegs. 364bool Formula::isCanonical() const { 365 if (ScaledReg) 366 return Scale != 1 || !BaseRegs.empty(); 367 return BaseRegs.size() <= 1; 368} 369 370/// \brief Helper method to morph a formula into its canonical representation. 371/// \see Formula::BaseRegs. 372/// Every formula having more than one base register, must use the ScaledReg 373/// field. Otherwise, we would have to do special cases everywhere in LSR 374/// to treat reg1 + reg2 + ... the same way as reg1 + 1*reg2 + ... 375/// On the other hand, 1*reg should be canonicalized into reg. 376void Formula::Canonicalize() { 377 if (isCanonical()) 378 return; 379 // So far we did not need this case. This is easy to implement but it is 380 // useless to maintain dead code. Beside it could hurt compile time. 381 assert(!BaseRegs.empty() && "1*reg => reg, should not be needed."); 382 // Keep the invariant sum in BaseRegs and one of the variant sum in ScaledReg. 383 ScaledReg = BaseRegs.back(); 384 BaseRegs.pop_back(); 385 Scale = 1; 386 size_t BaseRegsSize = BaseRegs.size(); 387 size_t Try = 0; 388 // If ScaledReg is an invariant, try to find a variant expression. 389 while (Try < BaseRegsSize && !isa<SCEVAddRecExpr>(ScaledReg)) 390 std::swap(ScaledReg, BaseRegs[Try++]); 391} 392 393/// \brief Get rid of the scale in the formula. 394/// In other words, this method morphes reg1 + 1*reg2 into reg1 + reg2. 395/// \return true if it was possible to get rid of the scale, false otherwise. 396/// \note After this operation the formula may not be in the canonical form. 397bool Formula::Unscale() { 398 if (Scale != 1) 399 return false; 400 Scale = 0; 401 BaseRegs.push_back(ScaledReg); 402 ScaledReg = nullptr; 403 return true; 404} 405 406/// getNumRegs - Return the total number of register operands used by this 407/// formula. This does not include register uses implied by non-constant 408/// addrec strides. 409size_t Formula::getNumRegs() const { 410 return !!ScaledReg + BaseRegs.size(); 411} 412 413/// getType - Return the type of this formula, if it has one, or null 414/// otherwise. This type is meaningless except for the bit size. 415Type *Formula::getType() const { 416 return !BaseRegs.empty() ? BaseRegs.front()->getType() : 417 ScaledReg ? ScaledReg->getType() : 418 BaseGV ? BaseGV->getType() : 419 nullptr; 420} 421 422/// DeleteBaseReg - Delete the given base reg from the BaseRegs list. 423void Formula::DeleteBaseReg(const SCEV *&S) { 424 if (&S != &BaseRegs.back()) 425 std::swap(S, BaseRegs.back()); 426 BaseRegs.pop_back(); 427} 428 429/// referencesReg - Test if this formula references the given register. 430bool Formula::referencesReg(const SCEV *S) const { 431 return S == ScaledReg || 432 std::find(BaseRegs.begin(), BaseRegs.end(), S) != BaseRegs.end(); 433} 434 435/// hasRegsUsedByUsesOtherThan - Test whether this formula uses registers 436/// which are used by uses other than the use with the given index. 437bool Formula::hasRegsUsedByUsesOtherThan(size_t LUIdx, 438 const RegUseTracker &RegUses) const { 439 if (ScaledReg) 440 if (RegUses.isRegUsedByUsesOtherThan(ScaledReg, LUIdx)) 441 return true; 442 for (const SCEV *BaseReg : BaseRegs) 443 if (RegUses.isRegUsedByUsesOtherThan(BaseReg, LUIdx)) 444 return true; 445 return false; 446} 447 448void Formula::print(raw_ostream &OS) const { 449 bool First = true; 450 if (BaseGV) { 451 if (!First) OS << " + "; else First = false; 452 BaseGV->printAsOperand(OS, /*PrintType=*/false); 453 } 454 if (BaseOffset != 0) { 455 if (!First) OS << " + "; else First = false; 456 OS << BaseOffset; 457 } 458 for (const SCEV *BaseReg : BaseRegs) { 459 if (!First) OS << " + "; else First = false; 460 OS << "reg(" << *BaseReg << ')'; 461 } 462 if (HasBaseReg && BaseRegs.empty()) { 463 if (!First) OS << " + "; else First = false; 464 OS << "**error: HasBaseReg**"; 465 } else if (!HasBaseReg && !BaseRegs.empty()) { 466 if (!First) OS << " + "; else First = false; 467 OS << "**error: !HasBaseReg**"; 468 } 469 if (Scale != 0) { 470 if (!First) OS << " + "; else First = false; 471 OS << Scale << "*reg("; 472 if (ScaledReg) 473 OS << *ScaledReg; 474 else 475 OS << "<unknown>"; 476 OS << ')'; 477 } 478 if (UnfoldedOffset != 0) { 479 if (!First) OS << " + "; 480 OS << "imm(" << UnfoldedOffset << ')'; 481 } 482} 483 484#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) 485void Formula::dump() const { 486 print(errs()); errs() << '\n'; 487} 488#endif 489 490/// isAddRecSExtable - Return true if the given addrec can be sign-extended 491/// without changing its value. 492static bool isAddRecSExtable(const SCEVAddRecExpr *AR, ScalarEvolution &SE) { 493 Type *WideTy = 494 IntegerType::get(SE.getContext(), SE.getTypeSizeInBits(AR->getType()) + 1); 495 return isa<SCEVAddRecExpr>(SE.getSignExtendExpr(AR, WideTy)); 496} 497 498/// isAddSExtable - Return true if the given add can be sign-extended 499/// without changing its value. 500static bool isAddSExtable(const SCEVAddExpr *A, ScalarEvolution &SE) { 501 Type *WideTy = 502 IntegerType::get(SE.getContext(), SE.getTypeSizeInBits(A->getType()) + 1); 503 return isa<SCEVAddExpr>(SE.getSignExtendExpr(A, WideTy)); 504} 505 506/// isMulSExtable - Return true if the given mul can be sign-extended 507/// without changing its value. 508static bool isMulSExtable(const SCEVMulExpr *M, ScalarEvolution &SE) { 509 Type *WideTy = 510 IntegerType::get(SE.getContext(), 511 SE.getTypeSizeInBits(M->getType()) * M->getNumOperands()); 512 return isa<SCEVMulExpr>(SE.getSignExtendExpr(M, WideTy)); 513} 514 515/// getExactSDiv - Return an expression for LHS /s RHS, if it can be determined 516/// and if the remainder is known to be zero, or null otherwise. If 517/// IgnoreSignificantBits is true, expressions like (X * Y) /s Y are simplified 518/// to Y, ignoring that the multiplication may overflow, which is useful when 519/// the result will be used in a context where the most significant bits are 520/// ignored. 521static const SCEV *getExactSDiv(const SCEV *LHS, const SCEV *RHS, 522 ScalarEvolution &SE, 523 bool IgnoreSignificantBits = false) { 524 // Handle the trivial case, which works for any SCEV type. 525 if (LHS == RHS) 526 return SE.getConstant(LHS->getType(), 1); 527 528 // Handle a few RHS special cases. 529 const SCEVConstant *RC = dyn_cast<SCEVConstant>(RHS); 530 if (RC) { 531 const APInt &RA = RC->getValue()->getValue(); 532 // Handle x /s -1 as x * -1, to give ScalarEvolution a chance to do 533 // some folding. 534 if (RA.isAllOnesValue()) 535 return SE.getMulExpr(LHS, RC); 536 // Handle x /s 1 as x. 537 if (RA == 1) 538 return LHS; 539 } 540 541 // Check for a division of a constant by a constant. 542 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(LHS)) { 543 if (!RC) 544 return nullptr; 545 const APInt &LA = C->getValue()->getValue(); 546 const APInt &RA = RC->getValue()->getValue(); 547 if (LA.srem(RA) != 0) 548 return nullptr; 549 return SE.getConstant(LA.sdiv(RA)); 550 } 551 552 // Distribute the sdiv over addrec operands, if the addrec doesn't overflow. 553 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(LHS)) { 554 if (IgnoreSignificantBits || isAddRecSExtable(AR, SE)) { 555 const SCEV *Step = getExactSDiv(AR->getStepRecurrence(SE), RHS, SE, 556 IgnoreSignificantBits); 557 if (!Step) return nullptr; 558 const SCEV *Start = getExactSDiv(AR->getStart(), RHS, SE, 559 IgnoreSignificantBits); 560 if (!Start) return nullptr; 561 // FlagNW is independent of the start value, step direction, and is 562 // preserved with smaller magnitude steps. 563 // FIXME: AR->getNoWrapFlags(SCEV::FlagNW) 564 return SE.getAddRecExpr(Start, Step, AR->getLoop(), SCEV::FlagAnyWrap); 565 } 566 return nullptr; 567 } 568 569 // Distribute the sdiv over add operands, if the add doesn't overflow. 570 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(LHS)) { 571 if (IgnoreSignificantBits || isAddSExtable(Add, SE)) { 572 SmallVector<const SCEV *, 8> Ops; 573 for (const SCEV *S : Add->operands()) { 574 const SCEV *Op = getExactSDiv(S, RHS, SE, IgnoreSignificantBits); 575 if (!Op) return nullptr; 576 Ops.push_back(Op); 577 } 578 return SE.getAddExpr(Ops); 579 } 580 return nullptr; 581 } 582 583 // Check for a multiply operand that we can pull RHS out of. 584 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(LHS)) { 585 if (IgnoreSignificantBits || isMulSExtable(Mul, SE)) { 586 SmallVector<const SCEV *, 4> Ops; 587 bool Found = false; 588 for (const SCEV *S : Mul->operands()) { 589 if (!Found) 590 if (const SCEV *Q = getExactSDiv(S, RHS, SE, 591 IgnoreSignificantBits)) { 592 S = Q; 593 Found = true; 594 } 595 Ops.push_back(S); 596 } 597 return Found ? SE.getMulExpr(Ops) : nullptr; 598 } 599 return nullptr; 600 } 601 602 // Otherwise we don't know. 603 return nullptr; 604} 605 606/// ExtractImmediate - If S involves the addition of a constant integer value, 607/// return that integer value, and mutate S to point to a new SCEV with that 608/// value excluded. 609static int64_t ExtractImmediate(const SCEV *&S, ScalarEvolution &SE) { 610 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S)) { 611 if (C->getValue()->getValue().getMinSignedBits() <= 64) { 612 S = SE.getConstant(C->getType(), 0); 613 return C->getValue()->getSExtValue(); 614 } 615 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) { 616 SmallVector<const SCEV *, 8> NewOps(Add->op_begin(), Add->op_end()); 617 int64_t Result = ExtractImmediate(NewOps.front(), SE); 618 if (Result != 0) 619 S = SE.getAddExpr(NewOps); 620 return Result; 621 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) { 622 SmallVector<const SCEV *, 8> NewOps(AR->op_begin(), AR->op_end()); 623 int64_t Result = ExtractImmediate(NewOps.front(), SE); 624 if (Result != 0) 625 S = SE.getAddRecExpr(NewOps, AR->getLoop(), 626 // FIXME: AR->getNoWrapFlags(SCEV::FlagNW) 627 SCEV::FlagAnyWrap); 628 return Result; 629 } 630 return 0; 631} 632 633/// ExtractSymbol - If S involves the addition of a GlobalValue address, 634/// return that symbol, and mutate S to point to a new SCEV with that 635/// value excluded. 636static GlobalValue *ExtractSymbol(const SCEV *&S, ScalarEvolution &SE) { 637 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) { 638 if (GlobalValue *GV = dyn_cast<GlobalValue>(U->getValue())) { 639 S = SE.getConstant(GV->getType(), 0); 640 return GV; 641 } 642 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) { 643 SmallVector<const SCEV *, 8> NewOps(Add->op_begin(), Add->op_end()); 644 GlobalValue *Result = ExtractSymbol(NewOps.back(), SE); 645 if (Result) 646 S = SE.getAddExpr(NewOps); 647 return Result; 648 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) { 649 SmallVector<const SCEV *, 8> NewOps(AR->op_begin(), AR->op_end()); 650 GlobalValue *Result = ExtractSymbol(NewOps.front(), SE); 651 if (Result) 652 S = SE.getAddRecExpr(NewOps, AR->getLoop(), 653 // FIXME: AR->getNoWrapFlags(SCEV::FlagNW) 654 SCEV::FlagAnyWrap); 655 return Result; 656 } 657 return nullptr; 658} 659 660/// isAddressUse - Returns true if the specified instruction is using the 661/// specified value as an address. 662static bool isAddressUse(Instruction *Inst, Value *OperandVal) { 663 bool isAddress = isa<LoadInst>(Inst); 664 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) { 665 if (SI->getOperand(1) == OperandVal) 666 isAddress = true; 667 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) { 668 // Addressing modes can also be folded into prefetches and a variety 669 // of intrinsics. 670 switch (II->getIntrinsicID()) { 671 default: break; 672 case Intrinsic::prefetch: 673 case Intrinsic::x86_sse_storeu_ps: 674 case Intrinsic::x86_sse2_storeu_pd: 675 case Intrinsic::x86_sse2_storeu_dq: 676 case Intrinsic::x86_sse2_storel_dq: 677 if (II->getArgOperand(0) == OperandVal) 678 isAddress = true; 679 break; 680 } 681 } 682 return isAddress; 683} 684 685/// getAccessType - Return the type of the memory being accessed. 686static Type *getAccessType(const Instruction *Inst) { 687 Type *AccessTy = Inst->getType(); 688 if (const StoreInst *SI = dyn_cast<StoreInst>(Inst)) 689 AccessTy = SI->getOperand(0)->getType(); 690 else if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) { 691 // Addressing modes can also be folded into prefetches and a variety 692 // of intrinsics. 693 switch (II->getIntrinsicID()) { 694 default: break; 695 case Intrinsic::x86_sse_storeu_ps: 696 case Intrinsic::x86_sse2_storeu_pd: 697 case Intrinsic::x86_sse2_storeu_dq: 698 case Intrinsic::x86_sse2_storel_dq: 699 AccessTy = II->getArgOperand(0)->getType(); 700 break; 701 } 702 } 703 704 // All pointers have the same requirements, so canonicalize them to an 705 // arbitrary pointer type to minimize variation. 706 if (PointerType *PTy = dyn_cast<PointerType>(AccessTy)) 707 AccessTy = PointerType::get(IntegerType::get(PTy->getContext(), 1), 708 PTy->getAddressSpace()); 709 710 return AccessTy; 711} 712 713/// isExistingPhi - Return true if this AddRec is already a phi in its loop. 714static bool isExistingPhi(const SCEVAddRecExpr *AR, ScalarEvolution &SE) { 715 for (BasicBlock::iterator I = AR->getLoop()->getHeader()->begin(); 716 PHINode *PN = dyn_cast<PHINode>(I); ++I) { 717 if (SE.isSCEVable(PN->getType()) && 718 (SE.getEffectiveSCEVType(PN->getType()) == 719 SE.getEffectiveSCEVType(AR->getType())) && 720 SE.getSCEV(PN) == AR) 721 return true; 722 } 723 return false; 724} 725 726/// Check if expanding this expression is likely to incur significant cost. This 727/// is tricky because SCEV doesn't track which expressions are actually computed 728/// by the current IR. 729/// 730/// We currently allow expansion of IV increments that involve adds, 731/// multiplication by constants, and AddRecs from existing phis. 732/// 733/// TODO: Allow UDivExpr if we can find an existing IV increment that is an 734/// obvious multiple of the UDivExpr. 735static bool isHighCostExpansion(const SCEV *S, 736 SmallPtrSetImpl<const SCEV*> &Processed, 737 ScalarEvolution &SE) { 738 // Zero/One operand expressions 739 switch (S->getSCEVType()) { 740 case scUnknown: 741 case scConstant: 742 return false; 743 case scTruncate: 744 return isHighCostExpansion(cast<SCEVTruncateExpr>(S)->getOperand(), 745 Processed, SE); 746 case scZeroExtend: 747 return isHighCostExpansion(cast<SCEVZeroExtendExpr>(S)->getOperand(), 748 Processed, SE); 749 case scSignExtend: 750 return isHighCostExpansion(cast<SCEVSignExtendExpr>(S)->getOperand(), 751 Processed, SE); 752 } 753 754 if (!Processed.insert(S).second) 755 return false; 756 757 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) { 758 for (const SCEV *S : Add->operands()) { 759 if (isHighCostExpansion(S, Processed, SE)) 760 return true; 761 } 762 return false; 763 } 764 765 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) { 766 if (Mul->getNumOperands() == 2) { 767 // Multiplication by a constant is ok 768 if (isa<SCEVConstant>(Mul->getOperand(0))) 769 return isHighCostExpansion(Mul->getOperand(1), Processed, SE); 770 771 // If we have the value of one operand, check if an existing 772 // multiplication already generates this expression. 773 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(Mul->getOperand(1))) { 774 Value *UVal = U->getValue(); 775 for (User *UR : UVal->users()) { 776 // If U is a constant, it may be used by a ConstantExpr. 777 Instruction *UI = dyn_cast<Instruction>(UR); 778 if (UI && UI->getOpcode() == Instruction::Mul && 779 SE.isSCEVable(UI->getType())) { 780 return SE.getSCEV(UI) == Mul; 781 } 782 } 783 } 784 } 785 } 786 787 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) { 788 if (isExistingPhi(AR, SE)) 789 return false; 790 } 791 792 // Fow now, consider any other type of expression (div/mul/min/max) high cost. 793 return true; 794} 795 796/// DeleteTriviallyDeadInstructions - If any of the instructions is the 797/// specified set are trivially dead, delete them and see if this makes any of 798/// their operands subsequently dead. 799static bool 800DeleteTriviallyDeadInstructions(SmallVectorImpl<WeakVH> &DeadInsts) { 801 bool Changed = false; 802 803 while (!DeadInsts.empty()) { 804 Value *V = DeadInsts.pop_back_val(); 805 Instruction *I = dyn_cast_or_null<Instruction>(V); 806 807 if (!I || !isInstructionTriviallyDead(I)) 808 continue; 809 810 for (Use &O : I->operands()) 811 if (Instruction *U = dyn_cast<Instruction>(O)) { 812 O = nullptr; 813 if (U->use_empty()) 814 DeadInsts.emplace_back(U); 815 } 816 817 I->eraseFromParent(); 818 Changed = true; 819 } 820 821 return Changed; 822} 823 824namespace { 825class LSRUse; 826} 827 828/// \brief Check if the addressing mode defined by \p F is completely 829/// folded in \p LU at isel time. 830/// This includes address-mode folding and special icmp tricks. 831/// This function returns true if \p LU can accommodate what \p F 832/// defines and up to 1 base + 1 scaled + offset. 833/// In other words, if \p F has several base registers, this function may 834/// still return true. Therefore, users still need to account for 835/// additional base registers and/or unfolded offsets to derive an 836/// accurate cost model. 837static bool isAMCompletelyFolded(const TargetTransformInfo &TTI, 838 const LSRUse &LU, const Formula &F); 839// Get the cost of the scaling factor used in F for LU. 840static unsigned getScalingFactorCost(const TargetTransformInfo &TTI, 841 const LSRUse &LU, const Formula &F); 842 843namespace { 844 845/// Cost - This class is used to measure and compare candidate formulae. 846class Cost { 847 /// TODO: Some of these could be merged. Also, a lexical ordering 848 /// isn't always optimal. 849 unsigned NumRegs; 850 unsigned AddRecCost; 851 unsigned NumIVMuls; 852 unsigned NumBaseAdds; 853 unsigned ImmCost; 854 unsigned SetupCost; 855 unsigned ScaleCost; 856 857public: 858 Cost() 859 : NumRegs(0), AddRecCost(0), NumIVMuls(0), NumBaseAdds(0), ImmCost(0), 860 SetupCost(0), ScaleCost(0) {} 861 862 bool operator<(const Cost &Other) const; 863 864 void Lose(); 865 866#ifndef NDEBUG 867 // Once any of the metrics loses, they must all remain losers. 868 bool isValid() { 869 return ((NumRegs | AddRecCost | NumIVMuls | NumBaseAdds 870 | ImmCost | SetupCost | ScaleCost) != ~0u) 871 || ((NumRegs & AddRecCost & NumIVMuls & NumBaseAdds 872 & ImmCost & SetupCost & ScaleCost) == ~0u); 873 } 874#endif 875 876 bool isLoser() { 877 assert(isValid() && "invalid cost"); 878 return NumRegs == ~0u; 879 } 880 881 void RateFormula(const TargetTransformInfo &TTI, 882 const Formula &F, 883 SmallPtrSetImpl<const SCEV *> &Regs, 884 const DenseSet<const SCEV *> &VisitedRegs, 885 const Loop *L, 886 const SmallVectorImpl<int64_t> &Offsets, 887 ScalarEvolution &SE, DominatorTree &DT, 888 const LSRUse &LU, 889 SmallPtrSetImpl<const SCEV *> *LoserRegs = nullptr); 890 891 void print(raw_ostream &OS) const; 892 void dump() const; 893 894private: 895 void RateRegister(const SCEV *Reg, 896 SmallPtrSetImpl<const SCEV *> &Regs, 897 const Loop *L, 898 ScalarEvolution &SE, DominatorTree &DT); 899 void RatePrimaryRegister(const SCEV *Reg, 900 SmallPtrSetImpl<const SCEV *> &Regs, 901 const Loop *L, 902 ScalarEvolution &SE, DominatorTree &DT, 903 SmallPtrSetImpl<const SCEV *> *LoserRegs); 904}; 905 906} 907 908/// RateRegister - Tally up interesting quantities from the given register. 909void Cost::RateRegister(const SCEV *Reg, 910 SmallPtrSetImpl<const SCEV *> &Regs, 911 const Loop *L, 912 ScalarEvolution &SE, DominatorTree &DT) { 913 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Reg)) { 914 // If this is an addrec for another loop, don't second-guess its addrec phi 915 // nodes. LSR isn't currently smart enough to reason about more than one 916 // loop at a time. LSR has already run on inner loops, will not run on outer 917 // loops, and cannot be expected to change sibling loops. 918 if (AR->getLoop() != L) { 919 // If the AddRec exists, consider it's register free and leave it alone. 920 if (isExistingPhi(AR, SE)) 921 return; 922 923 // Otherwise, do not consider this formula at all. 924 Lose(); 925 return; 926 } 927 AddRecCost += 1; /// TODO: This should be a function of the stride. 928 929 // Add the step value register, if it needs one. 930 // TODO: The non-affine case isn't precisely modeled here. 931 if (!AR->isAffine() || !isa<SCEVConstant>(AR->getOperand(1))) { 932 if (!Regs.count(AR->getOperand(1))) { 933 RateRegister(AR->getOperand(1), Regs, L, SE, DT); 934 if (isLoser()) 935 return; 936 } 937 } 938 } 939 ++NumRegs; 940 941 // Rough heuristic; favor registers which don't require extra setup 942 // instructions in the preheader. 943 if (!isa<SCEVUnknown>(Reg) && 944 !isa<SCEVConstant>(Reg) && 945 !(isa<SCEVAddRecExpr>(Reg) && 946 (isa<SCEVUnknown>(cast<SCEVAddRecExpr>(Reg)->getStart()) || 947 isa<SCEVConstant>(cast<SCEVAddRecExpr>(Reg)->getStart())))) 948 ++SetupCost; 949 950 NumIVMuls += isa<SCEVMulExpr>(Reg) && 951 SE.hasComputableLoopEvolution(Reg, L); 952} 953 954/// RatePrimaryRegister - Record this register in the set. If we haven't seen it 955/// before, rate it. Optional LoserRegs provides a way to declare any formula 956/// that refers to one of those regs an instant loser. 957void Cost::RatePrimaryRegister(const SCEV *Reg, 958 SmallPtrSetImpl<const SCEV *> &Regs, 959 const Loop *L, 960 ScalarEvolution &SE, DominatorTree &DT, 961 SmallPtrSetImpl<const SCEV *> *LoserRegs) { 962 if (LoserRegs && LoserRegs->count(Reg)) { 963 Lose(); 964 return; 965 } 966 if (Regs.insert(Reg).second) { 967 RateRegister(Reg, Regs, L, SE, DT); 968 if (LoserRegs && isLoser()) 969 LoserRegs->insert(Reg); 970 } 971} 972 973void Cost::RateFormula(const TargetTransformInfo &TTI, 974 const Formula &F, 975 SmallPtrSetImpl<const SCEV *> &Regs, 976 const DenseSet<const SCEV *> &VisitedRegs, 977 const Loop *L, 978 const SmallVectorImpl<int64_t> &Offsets, 979 ScalarEvolution &SE, DominatorTree &DT, 980 const LSRUse &LU, 981 SmallPtrSetImpl<const SCEV *> *LoserRegs) { 982 assert(F.isCanonical() && "Cost is accurate only for canonical formula"); 983 // Tally up the registers. 984 if (const SCEV *ScaledReg = F.ScaledReg) { 985 if (VisitedRegs.count(ScaledReg)) { 986 Lose(); 987 return; 988 } 989 RatePrimaryRegister(ScaledReg, Regs, L, SE, DT, LoserRegs); 990 if (isLoser()) 991 return; 992 } 993 for (const SCEV *BaseReg : F.BaseRegs) { 994 if (VisitedRegs.count(BaseReg)) { 995 Lose(); 996 return; 997 } 998 RatePrimaryRegister(BaseReg, Regs, L, SE, DT, LoserRegs); 999 if (isLoser()) 1000 return; 1001 } 1002 1003 // Determine how many (unfolded) adds we'll need inside the loop. 1004 size_t NumBaseParts = F.getNumRegs(); 1005 if (NumBaseParts > 1) 1006 // Do not count the base and a possible second register if the target 1007 // allows to fold 2 registers. 1008 NumBaseAdds += 1009 NumBaseParts - (1 + (F.Scale && isAMCompletelyFolded(TTI, LU, F))); 1010 NumBaseAdds += (F.UnfoldedOffset != 0); 1011 1012 // Accumulate non-free scaling amounts. 1013 ScaleCost += getScalingFactorCost(TTI, LU, F); 1014 1015 // Tally up the non-zero immediates. 1016 for (int64_t O : Offsets) { 1017 int64_t Offset = (uint64_t)O + F.BaseOffset; 1018 if (F.BaseGV) 1019 ImmCost += 64; // Handle symbolic values conservatively. 1020 // TODO: This should probably be the pointer size. 1021 else if (Offset != 0) 1022 ImmCost += APInt(64, Offset, true).getMinSignedBits(); 1023 } 1024 assert(isValid() && "invalid cost"); 1025} 1026 1027/// Lose - Set this cost to a losing value. 1028void Cost::Lose() { 1029 NumRegs = ~0u; 1030 AddRecCost = ~0u; 1031 NumIVMuls = ~0u; 1032 NumBaseAdds = ~0u; 1033 ImmCost = ~0u; 1034 SetupCost = ~0u; 1035 ScaleCost = ~0u; 1036} 1037 1038/// operator< - Choose the lower cost. 1039bool Cost::operator<(const Cost &Other) const { 1040 return std::tie(NumRegs, AddRecCost, NumIVMuls, NumBaseAdds, ScaleCost, 1041 ImmCost, SetupCost) < 1042 std::tie(Other.NumRegs, Other.AddRecCost, Other.NumIVMuls, 1043 Other.NumBaseAdds, Other.ScaleCost, Other.ImmCost, 1044 Other.SetupCost); 1045} 1046 1047void Cost::print(raw_ostream &OS) const { 1048 OS << NumRegs << " reg" << (NumRegs == 1 ? "" : "s"); 1049 if (AddRecCost != 0) 1050 OS << ", with addrec cost " << AddRecCost; 1051 if (NumIVMuls != 0) 1052 OS << ", plus " << NumIVMuls << " IV mul" << (NumIVMuls == 1 ? "" : "s"); 1053 if (NumBaseAdds != 0) 1054 OS << ", plus " << NumBaseAdds << " base add" 1055 << (NumBaseAdds == 1 ? "" : "s"); 1056 if (ScaleCost != 0) 1057 OS << ", plus " << ScaleCost << " scale cost"; 1058 if (ImmCost != 0) 1059 OS << ", plus " << ImmCost << " imm cost"; 1060 if (SetupCost != 0) 1061 OS << ", plus " << SetupCost << " setup cost"; 1062} 1063 1064#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) 1065void Cost::dump() const { 1066 print(errs()); errs() << '\n'; 1067} 1068#endif 1069 1070namespace { 1071 1072/// LSRFixup - An operand value in an instruction which is to be replaced 1073/// with some equivalent, possibly strength-reduced, replacement. 1074struct LSRFixup { 1075 /// UserInst - The instruction which will be updated. 1076 Instruction *UserInst; 1077 1078 /// OperandValToReplace - The operand of the instruction which will 1079 /// be replaced. The operand may be used more than once; every instance 1080 /// will be replaced. 1081 Value *OperandValToReplace; 1082 1083 /// PostIncLoops - If this user is to use the post-incremented value of an 1084 /// induction variable, this variable is non-null and holds the loop 1085 /// associated with the induction variable. 1086 PostIncLoopSet PostIncLoops; 1087 1088 /// LUIdx - The index of the LSRUse describing the expression which 1089 /// this fixup needs, minus an offset (below). 1090 size_t LUIdx; 1091 1092 /// Offset - A constant offset to be added to the LSRUse expression. 1093 /// This allows multiple fixups to share the same LSRUse with different 1094 /// offsets, for example in an unrolled loop. 1095 int64_t Offset; 1096 1097 bool isUseFullyOutsideLoop(const Loop *L) const; 1098 1099 LSRFixup(); 1100 1101 void print(raw_ostream &OS) const; 1102 void dump() const; 1103}; 1104 1105} 1106 1107LSRFixup::LSRFixup() 1108 : UserInst(nullptr), OperandValToReplace(nullptr), LUIdx(~size_t(0)), 1109 Offset(0) {} 1110 1111/// isUseFullyOutsideLoop - Test whether this fixup always uses its 1112/// value outside of the given loop. 1113bool LSRFixup::isUseFullyOutsideLoop(const Loop *L) const { 1114 // PHI nodes use their value in their incoming blocks. 1115 if (const PHINode *PN = dyn_cast<PHINode>(UserInst)) { 1116 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) 1117 if (PN->getIncomingValue(i) == OperandValToReplace && 1118 L->contains(PN->getIncomingBlock(i))) 1119 return false; 1120 return true; 1121 } 1122 1123 return !L->contains(UserInst); 1124} 1125 1126void LSRFixup::print(raw_ostream &OS) const { 1127 OS << "UserInst="; 1128 // Store is common and interesting enough to be worth special-casing. 1129 if (StoreInst *Store = dyn_cast<StoreInst>(UserInst)) { 1130 OS << "store "; 1131 Store->getOperand(0)->printAsOperand(OS, /*PrintType=*/false); 1132 } else if (UserInst->getType()->isVoidTy()) 1133 OS << UserInst->getOpcodeName(); 1134 else 1135 UserInst->printAsOperand(OS, /*PrintType=*/false); 1136 1137 OS << ", OperandValToReplace="; 1138 OperandValToReplace->printAsOperand(OS, /*PrintType=*/false); 1139 1140 for (const Loop *PIL : PostIncLoops) { 1141 OS << ", PostIncLoop="; 1142 PIL->getHeader()->printAsOperand(OS, /*PrintType=*/false); 1143 } 1144 1145 if (LUIdx != ~size_t(0)) 1146 OS << ", LUIdx=" << LUIdx; 1147 1148 if (Offset != 0) 1149 OS << ", Offset=" << Offset; 1150} 1151 1152#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) 1153void LSRFixup::dump() const { 1154 print(errs()); errs() << '\n'; 1155} 1156#endif 1157 1158namespace { 1159 1160/// UniquifierDenseMapInfo - A DenseMapInfo implementation for holding 1161/// DenseMaps and DenseSets of sorted SmallVectors of const SCEV*. 1162struct UniquifierDenseMapInfo { 1163 static SmallVector<const SCEV *, 4> getEmptyKey() { 1164 SmallVector<const SCEV *, 4> V; 1165 V.push_back(reinterpret_cast<const SCEV *>(-1)); 1166 return V; 1167 } 1168 1169 static SmallVector<const SCEV *, 4> getTombstoneKey() { 1170 SmallVector<const SCEV *, 4> V; 1171 V.push_back(reinterpret_cast<const SCEV *>(-2)); 1172 return V; 1173 } 1174 1175 static unsigned getHashValue(const SmallVector<const SCEV *, 4> &V) { 1176 return static_cast<unsigned>(hash_combine_range(V.begin(), V.end())); 1177 } 1178 1179 static bool isEqual(const SmallVector<const SCEV *, 4> &LHS, 1180 const SmallVector<const SCEV *, 4> &RHS) { 1181 return LHS == RHS; 1182 } 1183}; 1184 1185/// LSRUse - This class holds the state that LSR keeps for each use in 1186/// IVUsers, as well as uses invented by LSR itself. It includes information 1187/// about what kinds of things can be folded into the user, information about 1188/// the user itself, and information about how the use may be satisfied. 1189/// TODO: Represent multiple users of the same expression in common? 1190class LSRUse { 1191 DenseSet<SmallVector<const SCEV *, 4>, UniquifierDenseMapInfo> Uniquifier; 1192 1193public: 1194 /// KindType - An enum for a kind of use, indicating what types of 1195 /// scaled and immediate operands it might support. 1196 enum KindType { 1197 Basic, ///< A normal use, with no folding. 1198 Special, ///< A special case of basic, allowing -1 scales. 1199 Address, ///< An address use; folding according to TargetLowering 1200 ICmpZero ///< An equality icmp with both operands folded into one. 1201 // TODO: Add a generic icmp too? 1202 }; 1203 1204 typedef PointerIntPair<const SCEV *, 2, KindType> SCEVUseKindPair; 1205 1206 KindType Kind; 1207 Type *AccessTy; 1208 1209 SmallVector<int64_t, 8> Offsets; 1210 int64_t MinOffset; 1211 int64_t MaxOffset; 1212 1213 /// AllFixupsOutsideLoop - This records whether all of the fixups using this 1214 /// LSRUse are outside of the loop, in which case some special-case heuristics 1215 /// may be used. 1216 bool AllFixupsOutsideLoop; 1217 1218 /// RigidFormula is set to true to guarantee that this use will be associated 1219 /// with a single formula--the one that initially matched. Some SCEV 1220 /// expressions cannot be expanded. This allows LSR to consider the registers 1221 /// used by those expressions without the need to expand them later after 1222 /// changing the formula. 1223 bool RigidFormula; 1224 1225 /// WidestFixupType - This records the widest use type for any fixup using 1226 /// this LSRUse. FindUseWithSimilarFormula can't consider uses with different 1227 /// max fixup widths to be equivalent, because the narrower one may be relying 1228 /// on the implicit truncation to truncate away bogus bits. 1229 Type *WidestFixupType; 1230 1231 /// Formulae - A list of ways to build a value that can satisfy this user. 1232 /// After the list is populated, one of these is selected heuristically and 1233 /// used to formulate a replacement for OperandValToReplace in UserInst. 1234 SmallVector<Formula, 12> Formulae; 1235 1236 /// Regs - The set of register candidates used by all formulae in this LSRUse. 1237 SmallPtrSet<const SCEV *, 4> Regs; 1238 1239 LSRUse(KindType K, Type *T) : Kind(K), AccessTy(T), 1240 MinOffset(INT64_MAX), 1241 MaxOffset(INT64_MIN), 1242 AllFixupsOutsideLoop(true), 1243 RigidFormula(false), 1244 WidestFixupType(nullptr) {} 1245 1246 bool HasFormulaWithSameRegs(const Formula &F) const; 1247 bool InsertFormula(const Formula &F); 1248 void DeleteFormula(Formula &F); 1249 void RecomputeRegs(size_t LUIdx, RegUseTracker &Reguses); 1250 1251 void print(raw_ostream &OS) const; 1252 void dump() const; 1253}; 1254 1255} 1256 1257/// HasFormula - Test whether this use as a formula which has the same 1258/// registers as the given formula. 1259bool LSRUse::HasFormulaWithSameRegs(const Formula &F) const { 1260 SmallVector<const SCEV *, 4> Key = F.BaseRegs; 1261 if (F.ScaledReg) Key.push_back(F.ScaledReg); 1262 // Unstable sort by host order ok, because this is only used for uniquifying. 1263 std::sort(Key.begin(), Key.end()); 1264 return Uniquifier.count(Key); 1265} 1266 1267/// InsertFormula - If the given formula has not yet been inserted, add it to 1268/// the list, and return true. Return false otherwise. 1269/// The formula must be in canonical form. 1270bool LSRUse::InsertFormula(const Formula &F) { 1271 assert(F.isCanonical() && "Invalid canonical representation"); 1272 1273 if (!Formulae.empty() && RigidFormula) 1274 return false; 1275 1276 SmallVector<const SCEV *, 4> Key = F.BaseRegs; 1277 if (F.ScaledReg) Key.push_back(F.ScaledReg); 1278 // Unstable sort by host order ok, because this is only used for uniquifying. 1279 std::sort(Key.begin(), Key.end()); 1280 1281 if (!Uniquifier.insert(Key).second) 1282 return false; 1283 1284 // Using a register to hold the value of 0 is not profitable. 1285 assert((!F.ScaledReg || !F.ScaledReg->isZero()) && 1286 "Zero allocated in a scaled register!"); 1287#ifndef NDEBUG 1288 for (const SCEV *BaseReg : F.BaseRegs) 1289 assert(!BaseReg->isZero() && "Zero allocated in a base register!"); 1290#endif 1291 1292 // Add the formula to the list. 1293 Formulae.push_back(F); 1294 1295 // Record registers now being used by this use. 1296 Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end()); 1297 if (F.ScaledReg) 1298 Regs.insert(F.ScaledReg); 1299 1300 return true; 1301} 1302 1303/// DeleteFormula - Remove the given formula from this use's list. 1304void LSRUse::DeleteFormula(Formula &F) { 1305 if (&F != &Formulae.back()) 1306 std::swap(F, Formulae.back()); 1307 Formulae.pop_back(); 1308} 1309 1310/// RecomputeRegs - Recompute the Regs field, and update RegUses. 1311void LSRUse::RecomputeRegs(size_t LUIdx, RegUseTracker &RegUses) { 1312 // Now that we've filtered out some formulae, recompute the Regs set. 1313 SmallPtrSet<const SCEV *, 4> OldRegs = std::move(Regs); 1314 Regs.clear(); 1315 for (const Formula &F : Formulae) { 1316 if (F.ScaledReg) Regs.insert(F.ScaledReg); 1317 Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end()); 1318 } 1319 1320 // Update the RegTracker. 1321 for (const SCEV *S : OldRegs) 1322 if (!Regs.count(S)) 1323 RegUses.DropRegister(S, LUIdx); 1324} 1325 1326void LSRUse::print(raw_ostream &OS) const { 1327 OS << "LSR Use: Kind="; 1328 switch (Kind) { 1329 case Basic: OS << "Basic"; break; 1330 case Special: OS << "Special"; break; 1331 case ICmpZero: OS << "ICmpZero"; break; 1332 case Address: 1333 OS << "Address of "; 1334 if (AccessTy->isPointerTy()) 1335 OS << "pointer"; // the full pointer type could be really verbose 1336 else 1337 OS << *AccessTy; 1338 } 1339 1340 OS << ", Offsets={"; 1341 bool NeedComma = false; 1342 for (int64_t O : Offsets) { 1343 if (NeedComma) OS << ','; 1344 OS << O; 1345 NeedComma = true; 1346 } 1347 OS << '}'; 1348 1349 if (AllFixupsOutsideLoop) 1350 OS << ", all-fixups-outside-loop"; 1351 1352 if (WidestFixupType) 1353 OS << ", widest fixup type: " << *WidestFixupType; 1354} 1355 1356#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) 1357void LSRUse::dump() const { 1358 print(errs()); errs() << '\n'; 1359} 1360#endif 1361 1362static bool isAMCompletelyFolded(const TargetTransformInfo &TTI, 1363 LSRUse::KindType Kind, Type *AccessTy, 1364 GlobalValue *BaseGV, int64_t BaseOffset, 1365 bool HasBaseReg, int64_t Scale) { 1366 switch (Kind) { 1367 case LSRUse::Address: 1368 return TTI.isLegalAddressingMode(AccessTy, BaseGV, BaseOffset, HasBaseReg, Scale); 1369 1370 case LSRUse::ICmpZero: 1371 // There's not even a target hook for querying whether it would be legal to 1372 // fold a GV into an ICmp. 1373 if (BaseGV) 1374 return false; 1375 1376 // ICmp only has two operands; don't allow more than two non-trivial parts. 1377 if (Scale != 0 && HasBaseReg && BaseOffset != 0) 1378 return false; 1379 1380 // ICmp only supports no scale or a -1 scale, as we can "fold" a -1 scale by 1381 // putting the scaled register in the other operand of the icmp. 1382 if (Scale != 0 && Scale != -1) 1383 return false; 1384 1385 // If we have low-level target information, ask the target if it can fold an 1386 // integer immediate on an icmp. 1387 if (BaseOffset != 0) { 1388 // We have one of: 1389 // ICmpZero BaseReg + BaseOffset => ICmp BaseReg, -BaseOffset 1390 // ICmpZero -1*ScaleReg + BaseOffset => ICmp ScaleReg, BaseOffset 1391 // Offs is the ICmp immediate. 1392 if (Scale == 0) 1393 // The cast does the right thing with INT64_MIN. 1394 BaseOffset = -(uint64_t)BaseOffset; 1395 return TTI.isLegalICmpImmediate(BaseOffset); 1396 } 1397 1398 // ICmpZero BaseReg + -1*ScaleReg => ICmp BaseReg, ScaleReg 1399 return true; 1400 1401 case LSRUse::Basic: 1402 // Only handle single-register values. 1403 return !BaseGV && Scale == 0 && BaseOffset == 0; 1404 1405 case LSRUse::Special: 1406 // Special case Basic to handle -1 scales. 1407 return !BaseGV && (Scale == 0 || Scale == -1) && BaseOffset == 0; 1408 } 1409 1410 llvm_unreachable("Invalid LSRUse Kind!"); 1411} 1412 1413static bool isAMCompletelyFolded(const TargetTransformInfo &TTI, 1414 int64_t MinOffset, int64_t MaxOffset, 1415 LSRUse::KindType Kind, Type *AccessTy, 1416 GlobalValue *BaseGV, int64_t BaseOffset, 1417 bool HasBaseReg, int64_t Scale) { 1418 // Check for overflow. 1419 if (((int64_t)((uint64_t)BaseOffset + MinOffset) > BaseOffset) != 1420 (MinOffset > 0)) 1421 return false; 1422 MinOffset = (uint64_t)BaseOffset + MinOffset; 1423 if (((int64_t)((uint64_t)BaseOffset + MaxOffset) > BaseOffset) != 1424 (MaxOffset > 0)) 1425 return false; 1426 MaxOffset = (uint64_t)BaseOffset + MaxOffset; 1427 1428 return isAMCompletelyFolded(TTI, Kind, AccessTy, BaseGV, MinOffset, 1429 HasBaseReg, Scale) && 1430 isAMCompletelyFolded(TTI, Kind, AccessTy, BaseGV, MaxOffset, 1431 HasBaseReg, Scale); 1432} 1433 1434static bool isAMCompletelyFolded(const TargetTransformInfo &TTI, 1435 int64_t MinOffset, int64_t MaxOffset, 1436 LSRUse::KindType Kind, Type *AccessTy, 1437 const Formula &F) { 1438 // For the purpose of isAMCompletelyFolded either having a canonical formula 1439 // or a scale not equal to zero is correct. 1440 // Problems may arise from non canonical formulae having a scale == 0. 1441 // Strictly speaking it would best to just rely on canonical formulae. 1442 // However, when we generate the scaled formulae, we first check that the 1443 // scaling factor is profitable before computing the actual ScaledReg for 1444 // compile time sake. 1445 assert((F.isCanonical() || F.Scale != 0)); 1446 return isAMCompletelyFolded(TTI, MinOffset, MaxOffset, Kind, AccessTy, 1447 F.BaseGV, F.BaseOffset, F.HasBaseReg, F.Scale); 1448} 1449 1450/// isLegalUse - Test whether we know how to expand the current formula. 1451static bool isLegalUse(const TargetTransformInfo &TTI, int64_t MinOffset, 1452 int64_t MaxOffset, LSRUse::KindType Kind, Type *AccessTy, 1453 GlobalValue *BaseGV, int64_t BaseOffset, bool HasBaseReg, 1454 int64_t Scale) { 1455 // We know how to expand completely foldable formulae. 1456 return isAMCompletelyFolded(TTI, MinOffset, MaxOffset, Kind, AccessTy, BaseGV, 1457 BaseOffset, HasBaseReg, Scale) || 1458 // Or formulae that use a base register produced by a sum of base 1459 // registers. 1460 (Scale == 1 && 1461 isAMCompletelyFolded(TTI, MinOffset, MaxOffset, Kind, AccessTy, 1462 BaseGV, BaseOffset, true, 0)); 1463} 1464 1465static bool isLegalUse(const TargetTransformInfo &TTI, int64_t MinOffset, 1466 int64_t MaxOffset, LSRUse::KindType Kind, Type *AccessTy, 1467 const Formula &F) { 1468 return isLegalUse(TTI, MinOffset, MaxOffset, Kind, AccessTy, F.BaseGV, 1469 F.BaseOffset, F.HasBaseReg, F.Scale); 1470} 1471 1472static bool isAMCompletelyFolded(const TargetTransformInfo &TTI, 1473 const LSRUse &LU, const Formula &F) { 1474 return isAMCompletelyFolded(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, 1475 LU.AccessTy, F.BaseGV, F.BaseOffset, F.HasBaseReg, 1476 F.Scale); 1477} 1478 1479static unsigned getScalingFactorCost(const TargetTransformInfo &TTI, 1480 const LSRUse &LU, const Formula &F) { 1481 if (!F.Scale) 1482 return 0; 1483 1484 // If the use is not completely folded in that instruction, we will have to 1485 // pay an extra cost only for scale != 1. 1486 if (!isAMCompletelyFolded(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, 1487 LU.AccessTy, F)) 1488 return F.Scale != 1; 1489 1490 switch (LU.Kind) { 1491 case LSRUse::Address: { 1492 // Check the scaling factor cost with both the min and max offsets. 1493 int ScaleCostMinOffset = 1494 TTI.getScalingFactorCost(LU.AccessTy, F.BaseGV, 1495 F.BaseOffset + LU.MinOffset, 1496 F.HasBaseReg, F.Scale); 1497 int ScaleCostMaxOffset = 1498 TTI.getScalingFactorCost(LU.AccessTy, F.BaseGV, 1499 F.BaseOffset + LU.MaxOffset, 1500 F.HasBaseReg, F.Scale); 1501 1502 assert(ScaleCostMinOffset >= 0 && ScaleCostMaxOffset >= 0 && 1503 "Legal addressing mode has an illegal cost!"); 1504 return std::max(ScaleCostMinOffset, ScaleCostMaxOffset); 1505 } 1506 case LSRUse::ICmpZero: 1507 case LSRUse::Basic: 1508 case LSRUse::Special: 1509 // The use is completely folded, i.e., everything is folded into the 1510 // instruction. 1511 return 0; 1512 } 1513 1514 llvm_unreachable("Invalid LSRUse Kind!"); 1515} 1516 1517static bool isAlwaysFoldable(const TargetTransformInfo &TTI, 1518 LSRUse::KindType Kind, Type *AccessTy, 1519 GlobalValue *BaseGV, int64_t BaseOffset, 1520 bool HasBaseReg) { 1521 // Fast-path: zero is always foldable. 1522 if (BaseOffset == 0 && !BaseGV) return true; 1523 1524 // Conservatively, create an address with an immediate and a 1525 // base and a scale. 1526 int64_t Scale = Kind == LSRUse::ICmpZero ? -1 : 1; 1527 1528 // Canonicalize a scale of 1 to a base register if the formula doesn't 1529 // already have a base register. 1530 if (!HasBaseReg && Scale == 1) { 1531 Scale = 0; 1532 HasBaseReg = true; 1533 } 1534 1535 return isAMCompletelyFolded(TTI, Kind, AccessTy, BaseGV, BaseOffset, 1536 HasBaseReg, Scale); 1537} 1538 1539static bool isAlwaysFoldable(const TargetTransformInfo &TTI, 1540 ScalarEvolution &SE, int64_t MinOffset, 1541 int64_t MaxOffset, LSRUse::KindType Kind, 1542 Type *AccessTy, const SCEV *S, bool HasBaseReg) { 1543 // Fast-path: zero is always foldable. 1544 if (S->isZero()) return true; 1545 1546 // Conservatively, create an address with an immediate and a 1547 // base and a scale. 1548 int64_t BaseOffset = ExtractImmediate(S, SE); 1549 GlobalValue *BaseGV = ExtractSymbol(S, SE); 1550 1551 // If there's anything else involved, it's not foldable. 1552 if (!S->isZero()) return false; 1553 1554 // Fast-path: zero is always foldable. 1555 if (BaseOffset == 0 && !BaseGV) return true; 1556 1557 // Conservatively, create an address with an immediate and a 1558 // base and a scale. 1559 int64_t Scale = Kind == LSRUse::ICmpZero ? -1 : 1; 1560 1561 return isAMCompletelyFolded(TTI, MinOffset, MaxOffset, Kind, AccessTy, BaseGV, 1562 BaseOffset, HasBaseReg, Scale); 1563} 1564 1565namespace { 1566 1567/// IVInc - An individual increment in a Chain of IV increments. 1568/// Relate an IV user to an expression that computes the IV it uses from the IV 1569/// used by the previous link in the Chain. 1570/// 1571/// For the head of a chain, IncExpr holds the absolute SCEV expression for the 1572/// original IVOperand. The head of the chain's IVOperand is only valid during 1573/// chain collection, before LSR replaces IV users. During chain generation, 1574/// IncExpr can be used to find the new IVOperand that computes the same 1575/// expression. 1576struct IVInc { 1577 Instruction *UserInst; 1578 Value* IVOperand; 1579 const SCEV *IncExpr; 1580 1581 IVInc(Instruction *U, Value *O, const SCEV *E): 1582 UserInst(U), IVOperand(O), IncExpr(E) {} 1583}; 1584 1585// IVChain - The list of IV increments in program order. 1586// We typically add the head of a chain without finding subsequent links. 1587struct IVChain { 1588 SmallVector<IVInc,1> Incs; 1589 const SCEV *ExprBase; 1590 1591 IVChain() : ExprBase(nullptr) {} 1592 1593 IVChain(const IVInc &Head, const SCEV *Base) 1594 : Incs(1, Head), ExprBase(Base) {} 1595 1596 typedef SmallVectorImpl<IVInc>::const_iterator const_iterator; 1597 1598 // begin - return the first increment in the chain. 1599 const_iterator begin() const { 1600 assert(!Incs.empty()); 1601 return std::next(Incs.begin()); 1602 } 1603 const_iterator end() const { 1604 return Incs.end(); 1605 } 1606 1607 // hasIncs - Returns true if this chain contains any increments. 1608 bool hasIncs() const { return Incs.size() >= 2; } 1609 1610 // add - Add an IVInc to the end of this chain. 1611 void add(const IVInc &X) { Incs.push_back(X); } 1612 1613 // tailUserInst - Returns the last UserInst in the chain. 1614 Instruction *tailUserInst() const { return Incs.back().UserInst; } 1615 1616 // isProfitableIncrement - Returns true if IncExpr can be profitably added to 1617 // this chain. 1618 bool isProfitableIncrement(const SCEV *OperExpr, 1619 const SCEV *IncExpr, 1620 ScalarEvolution&); 1621}; 1622 1623/// ChainUsers - Helper for CollectChains to track multiple IV increment uses. 1624/// Distinguish between FarUsers that definitely cross IV increments and 1625/// NearUsers that may be used between IV increments. 1626struct ChainUsers { 1627 SmallPtrSet<Instruction*, 4> FarUsers; 1628 SmallPtrSet<Instruction*, 4> NearUsers; 1629}; 1630 1631/// LSRInstance - This class holds state for the main loop strength reduction 1632/// logic. 1633class LSRInstance { 1634 IVUsers &IU; 1635 ScalarEvolution &SE; 1636 DominatorTree &DT; 1637 LoopInfo &LI; 1638 const TargetTransformInfo &TTI; 1639 Loop *const L; 1640 bool Changed; 1641 1642 /// IVIncInsertPos - This is the insert position that the current loop's 1643 /// induction variable increment should be placed. In simple loops, this is 1644 /// the latch block's terminator. But in more complicated cases, this is a 1645 /// position which will dominate all the in-loop post-increment users. 1646 Instruction *IVIncInsertPos; 1647 1648 /// Factors - Interesting factors between use strides. 1649 SmallSetVector<int64_t, 8> Factors; 1650 1651 /// Types - Interesting use types, to facilitate truncation reuse. 1652 SmallSetVector<Type *, 4> Types; 1653 1654 /// Fixups - The list of operands which are to be replaced. 1655 SmallVector<LSRFixup, 16> Fixups; 1656 1657 /// Uses - The list of interesting uses. 1658 SmallVector<LSRUse, 16> Uses; 1659 1660 /// RegUses - Track which uses use which register candidates. 1661 RegUseTracker RegUses; 1662 1663 // Limit the number of chains to avoid quadratic behavior. We don't expect to 1664 // have more than a few IV increment chains in a loop. Missing a Chain falls 1665 // back to normal LSR behavior for those uses. 1666 static const unsigned MaxChains = 8; 1667 1668 /// IVChainVec - IV users can form a chain of IV increments. 1669 SmallVector<IVChain, MaxChains> IVChainVec; 1670 1671 /// IVIncSet - IV users that belong to profitable IVChains. 1672 SmallPtrSet<Use*, MaxChains> IVIncSet; 1673 1674 void OptimizeShadowIV(); 1675 bool FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse); 1676 ICmpInst *OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse); 1677 void OptimizeLoopTermCond(); 1678 1679 void ChainInstruction(Instruction *UserInst, Instruction *IVOper, 1680 SmallVectorImpl<ChainUsers> &ChainUsersVec); 1681 void FinalizeChain(IVChain &Chain); 1682 void CollectChains(); 1683 void GenerateIVChain(const IVChain &Chain, SCEVExpander &Rewriter, 1684 SmallVectorImpl<WeakVH> &DeadInsts); 1685 1686 void CollectInterestingTypesAndFactors(); 1687 void CollectFixupsAndInitialFormulae(); 1688 1689 LSRFixup &getNewFixup() { 1690 Fixups.push_back(LSRFixup()); 1691 return Fixups.back(); 1692 } 1693 1694 // Support for sharing of LSRUses between LSRFixups. 1695 typedef DenseMap<LSRUse::SCEVUseKindPair, size_t> UseMapTy; 1696 UseMapTy UseMap; 1697 1698 bool reconcileNewOffset(LSRUse &LU, int64_t NewOffset, bool HasBaseReg, 1699 LSRUse::KindType Kind, Type *AccessTy); 1700 1701 std::pair<size_t, int64_t> getUse(const SCEV *&Expr, 1702 LSRUse::KindType Kind, 1703 Type *AccessTy); 1704 1705 void DeleteUse(LSRUse &LU, size_t LUIdx); 1706 1707 LSRUse *FindUseWithSimilarFormula(const Formula &F, const LSRUse &OrigLU); 1708 1709 void InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx); 1710 void InsertSupplementalFormula(const SCEV *S, LSRUse &LU, size_t LUIdx); 1711 void CountRegisters(const Formula &F, size_t LUIdx); 1712 bool InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F); 1713 1714 void CollectLoopInvariantFixupsAndFormulae(); 1715 1716 void GenerateReassociations(LSRUse &LU, unsigned LUIdx, Formula Base, 1717 unsigned Depth = 0); 1718 1719 void GenerateReassociationsImpl(LSRUse &LU, unsigned LUIdx, 1720 const Formula &Base, unsigned Depth, 1721 size_t Idx, bool IsScaledReg = false); 1722 void GenerateCombinations(LSRUse &LU, unsigned LUIdx, Formula Base); 1723 void GenerateSymbolicOffsetsImpl(LSRUse &LU, unsigned LUIdx, 1724 const Formula &Base, size_t Idx, 1725 bool IsScaledReg = false); 1726 void GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx, Formula Base); 1727 void GenerateConstantOffsetsImpl(LSRUse &LU, unsigned LUIdx, 1728 const Formula &Base, 1729 const SmallVectorImpl<int64_t> &Worklist, 1730 size_t Idx, bool IsScaledReg = false); 1731 void GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx, Formula Base); 1732 void GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx, Formula Base); 1733 void GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base); 1734 void GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base); 1735 void GenerateCrossUseConstantOffsets(); 1736 void GenerateAllReuseFormulae(); 1737 1738 void FilterOutUndesirableDedicatedRegisters(); 1739 1740 size_t EstimateSearchSpaceComplexity() const; 1741 void NarrowSearchSpaceByDetectingSupersets(); 1742 void NarrowSearchSpaceByCollapsingUnrolledCode(); 1743 void NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters(); 1744 void NarrowSearchSpaceByPickingWinnerRegs(); 1745 void NarrowSearchSpaceUsingHeuristics(); 1746 1747 void SolveRecurse(SmallVectorImpl<const Formula *> &Solution, 1748 Cost &SolutionCost, 1749 SmallVectorImpl<const Formula *> &Workspace, 1750 const Cost &CurCost, 1751 const SmallPtrSet<const SCEV *, 16> &CurRegs, 1752 DenseSet<const SCEV *> &VisitedRegs) const; 1753 void Solve(SmallVectorImpl<const Formula *> &Solution) const; 1754 1755 BasicBlock::iterator 1756 HoistInsertPosition(BasicBlock::iterator IP, 1757 const SmallVectorImpl<Instruction *> &Inputs) const; 1758 BasicBlock::iterator 1759 AdjustInsertPositionForExpand(BasicBlock::iterator IP, 1760 const LSRFixup &LF, 1761 const LSRUse &LU, 1762 SCEVExpander &Rewriter) const; 1763 1764 Value *Expand(const LSRFixup &LF, 1765 const Formula &F, 1766 BasicBlock::iterator IP, 1767 SCEVExpander &Rewriter, 1768 SmallVectorImpl<WeakVH> &DeadInsts) const; 1769 void RewriteForPHI(PHINode *PN, const LSRFixup &LF, 1770 const Formula &F, 1771 SCEVExpander &Rewriter, 1772 SmallVectorImpl<WeakVH> &DeadInsts, 1773 Pass *P) const; 1774 void Rewrite(const LSRFixup &LF, 1775 const Formula &F, 1776 SCEVExpander &Rewriter, 1777 SmallVectorImpl<WeakVH> &DeadInsts, 1778 Pass *P) const; 1779 void ImplementSolution(const SmallVectorImpl<const Formula *> &Solution, 1780 Pass *P); 1781 1782public: 1783 LSRInstance(Loop *L, Pass *P); 1784 1785 bool getChanged() const { return Changed; } 1786 1787 void print_factors_and_types(raw_ostream &OS) const; 1788 void print_fixups(raw_ostream &OS) const; 1789 void print_uses(raw_ostream &OS) const; 1790 void print(raw_ostream &OS) const; 1791 void dump() const; 1792}; 1793 1794} 1795 1796/// OptimizeShadowIV - If IV is used in a int-to-float cast 1797/// inside the loop then try to eliminate the cast operation. 1798void LSRInstance::OptimizeShadowIV() { 1799 const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L); 1800 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount)) 1801 return; 1802 1803 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); 1804 UI != E; /* empty */) { 1805 IVUsers::const_iterator CandidateUI = UI; 1806 ++UI; 1807 Instruction *ShadowUse = CandidateUI->getUser(); 1808 Type *DestTy = nullptr; 1809 bool IsSigned = false; 1810 1811 /* If shadow use is a int->float cast then insert a second IV 1812 to eliminate this cast. 1813 1814 for (unsigned i = 0; i < n; ++i) 1815 foo((double)i); 1816 1817 is transformed into 1818 1819 double d = 0.0; 1820 for (unsigned i = 0; i < n; ++i, ++d) 1821 foo(d); 1822 */ 1823 if (UIToFPInst *UCast = dyn_cast<UIToFPInst>(CandidateUI->getUser())) { 1824 IsSigned = false; 1825 DestTy = UCast->getDestTy(); 1826 } 1827 else if (SIToFPInst *SCast = dyn_cast<SIToFPInst>(CandidateUI->getUser())) { 1828 IsSigned = true; 1829 DestTy = SCast->getDestTy(); 1830 } 1831 if (!DestTy) continue; 1832 1833 // If target does not support DestTy natively then do not apply 1834 // this transformation. 1835 if (!TTI.isTypeLegal(DestTy)) continue; 1836 1837 PHINode *PH = dyn_cast<PHINode>(ShadowUse->getOperand(0)); 1838 if (!PH) continue; 1839 if (PH->getNumIncomingValues() != 2) continue; 1840 1841 Type *SrcTy = PH->getType(); 1842 int Mantissa = DestTy->getFPMantissaWidth(); 1843 if (Mantissa == -1) continue; 1844 if ((int)SE.getTypeSizeInBits(SrcTy) > Mantissa) 1845 continue; 1846 1847 unsigned Entry, Latch; 1848 if (PH->getIncomingBlock(0) == L->getLoopPreheader()) { 1849 Entry = 0; 1850 Latch = 1; 1851 } else { 1852 Entry = 1; 1853 Latch = 0; 1854 } 1855 1856 ConstantInt *Init = dyn_cast<ConstantInt>(PH->getIncomingValue(Entry)); 1857 if (!Init) continue; 1858 Constant *NewInit = ConstantFP::get(DestTy, IsSigned ? 1859 (double)Init->getSExtValue() : 1860 (double)Init->getZExtValue()); 1861 1862 BinaryOperator *Incr = 1863 dyn_cast<BinaryOperator>(PH->getIncomingValue(Latch)); 1864 if (!Incr) continue; 1865 if (Incr->getOpcode() != Instruction::Add 1866 && Incr->getOpcode() != Instruction::Sub) 1867 continue; 1868 1869 /* Initialize new IV, double d = 0.0 in above example. */ 1870 ConstantInt *C = nullptr; 1871 if (Incr->getOperand(0) == PH) 1872 C = dyn_cast<ConstantInt>(Incr->getOperand(1)); 1873 else if (Incr->getOperand(1) == PH) 1874 C = dyn_cast<ConstantInt>(Incr->getOperand(0)); 1875 else 1876 continue; 1877 1878 if (!C) continue; 1879 1880 // Ignore negative constants, as the code below doesn't handle them 1881 // correctly. TODO: Remove this restriction. 1882 if (!C->getValue().isStrictlyPositive()) continue; 1883 1884 /* Add new PHINode. */ 1885 PHINode *NewPH = PHINode::Create(DestTy, 2, "IV.S.", PH); 1886 1887 /* create new increment. '++d' in above example. */ 1888 Constant *CFP = ConstantFP::get(DestTy, C->getZExtValue()); 1889 BinaryOperator *NewIncr = 1890 BinaryOperator::Create(Incr->getOpcode() == Instruction::Add ? 1891 Instruction::FAdd : Instruction::FSub, 1892 NewPH, CFP, "IV.S.next.", Incr); 1893 1894 NewPH->addIncoming(NewInit, PH->getIncomingBlock(Entry)); 1895 NewPH->addIncoming(NewIncr, PH->getIncomingBlock(Latch)); 1896 1897 /* Remove cast operation */ 1898 ShadowUse->replaceAllUsesWith(NewPH); 1899 ShadowUse->eraseFromParent(); 1900 Changed = true; 1901 break; 1902 } 1903} 1904 1905/// FindIVUserForCond - If Cond has an operand that is an expression of an IV, 1906/// set the IV user and stride information and return true, otherwise return 1907/// false. 1908bool LSRInstance::FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse) { 1909 for (IVStrideUse &U : IU) 1910 if (U.getUser() == Cond) { 1911 // NOTE: we could handle setcc instructions with multiple uses here, but 1912 // InstCombine does it as well for simple uses, it's not clear that it 1913 // occurs enough in real life to handle. 1914 CondUse = &U; 1915 return true; 1916 } 1917 return false; 1918} 1919 1920/// OptimizeMax - Rewrite the loop's terminating condition if it uses 1921/// a max computation. 1922/// 1923/// This is a narrow solution to a specific, but acute, problem. For loops 1924/// like this: 1925/// 1926/// i = 0; 1927/// do { 1928/// p[i] = 0.0; 1929/// } while (++i < n); 1930/// 1931/// the trip count isn't just 'n', because 'n' might not be positive. And 1932/// unfortunately this can come up even for loops where the user didn't use 1933/// a C do-while loop. For example, seemingly well-behaved top-test loops 1934/// will commonly be lowered like this: 1935// 1936/// if (n > 0) { 1937/// i = 0; 1938/// do { 1939/// p[i] = 0.0; 1940/// } while (++i < n); 1941/// } 1942/// 1943/// and then it's possible for subsequent optimization to obscure the if 1944/// test in such a way that indvars can't find it. 1945/// 1946/// When indvars can't find the if test in loops like this, it creates a 1947/// max expression, which allows it to give the loop a canonical 1948/// induction variable: 1949/// 1950/// i = 0; 1951/// max = n < 1 ? 1 : n; 1952/// do { 1953/// p[i] = 0.0; 1954/// } while (++i != max); 1955/// 1956/// Canonical induction variables are necessary because the loop passes 1957/// are designed around them. The most obvious example of this is the 1958/// LoopInfo analysis, which doesn't remember trip count values. It 1959/// expects to be able to rediscover the trip count each time it is 1960/// needed, and it does this using a simple analysis that only succeeds if 1961/// the loop has a canonical induction variable. 1962/// 1963/// However, when it comes time to generate code, the maximum operation 1964/// can be quite costly, especially if it's inside of an outer loop. 1965/// 1966/// This function solves this problem by detecting this type of loop and 1967/// rewriting their conditions from ICMP_NE back to ICMP_SLT, and deleting 1968/// the instructions for the maximum computation. 1969/// 1970ICmpInst *LSRInstance::OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse) { 1971 // Check that the loop matches the pattern we're looking for. 1972 if (Cond->getPredicate() != CmpInst::ICMP_EQ && 1973 Cond->getPredicate() != CmpInst::ICMP_NE) 1974 return Cond; 1975 1976 SelectInst *Sel = dyn_cast<SelectInst>(Cond->getOperand(1)); 1977 if (!Sel || !Sel->hasOneUse()) return Cond; 1978 1979 const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L); 1980 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount)) 1981 return Cond; 1982 const SCEV *One = SE.getConstant(BackedgeTakenCount->getType(), 1); 1983 1984 // Add one to the backedge-taken count to get the trip count. 1985 const SCEV *IterationCount = SE.getAddExpr(One, BackedgeTakenCount); 1986 if (IterationCount != SE.getSCEV(Sel)) return Cond; 1987 1988 // Check for a max calculation that matches the pattern. There's no check 1989 // for ICMP_ULE here because the comparison would be with zero, which 1990 // isn't interesting. 1991 CmpInst::Predicate Pred = ICmpInst::BAD_ICMP_PREDICATE; 1992 const SCEVNAryExpr *Max = nullptr; 1993 if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(BackedgeTakenCount)) { 1994 Pred = ICmpInst::ICMP_SLE; 1995 Max = S; 1996 } else if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(IterationCount)) { 1997 Pred = ICmpInst::ICMP_SLT; 1998 Max = S; 1999 } else if (const SCEVUMaxExpr *U = dyn_cast<SCEVUMaxExpr>(IterationCount)) { 2000 Pred = ICmpInst::ICMP_ULT; 2001 Max = U; 2002 } else { 2003 // No match; bail. 2004 return Cond; 2005 } 2006 2007 // To handle a max with more than two operands, this optimization would 2008 // require additional checking and setup. 2009 if (Max->getNumOperands() != 2) 2010 return Cond; 2011 2012 const SCEV *MaxLHS = Max->getOperand(0); 2013 const SCEV *MaxRHS = Max->getOperand(1); 2014 2015 // ScalarEvolution canonicalizes constants to the left. For < and >, look 2016 // for a comparison with 1. For <= and >=, a comparison with zero. 2017 if (!MaxLHS || 2018 (ICmpInst::isTrueWhenEqual(Pred) ? !MaxLHS->isZero() : (MaxLHS != One))) 2019 return Cond; 2020 2021 // Check the relevant induction variable for conformance to 2022 // the pattern. 2023 const SCEV *IV = SE.getSCEV(Cond->getOperand(0)); 2024 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(IV); 2025 if (!AR || !AR->isAffine() || 2026 AR->getStart() != One || 2027 AR->getStepRecurrence(SE) != One) 2028 return Cond; 2029 2030 assert(AR->getLoop() == L && 2031 "Loop condition operand is an addrec in a different loop!"); 2032 2033 // Check the right operand of the select, and remember it, as it will 2034 // be used in the new comparison instruction. 2035 Value *NewRHS = nullptr; 2036 if (ICmpInst::isTrueWhenEqual(Pred)) { 2037 // Look for n+1, and grab n. 2038 if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(1))) 2039 if (ConstantInt *BO1 = dyn_cast<ConstantInt>(BO->getOperand(1))) 2040 if (BO1->isOne() && SE.getSCEV(BO->getOperand(0)) == MaxRHS) 2041 NewRHS = BO->getOperand(0); 2042 if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(2))) 2043 if (ConstantInt *BO1 = dyn_cast<ConstantInt>(BO->getOperand(1))) 2044 if (BO1->isOne() && SE.getSCEV(BO->getOperand(0)) == MaxRHS) 2045 NewRHS = BO->getOperand(0); 2046 if (!NewRHS) 2047 return Cond; 2048 } else if (SE.getSCEV(Sel->getOperand(1)) == MaxRHS) 2049 NewRHS = Sel->getOperand(1); 2050 else if (SE.getSCEV(Sel->getOperand(2)) == MaxRHS) 2051 NewRHS = Sel->getOperand(2); 2052 else if (const SCEVUnknown *SU = dyn_cast<SCEVUnknown>(MaxRHS)) 2053 NewRHS = SU->getValue(); 2054 else 2055 // Max doesn't match expected pattern. 2056 return Cond; 2057 2058 // Determine the new comparison opcode. It may be signed or unsigned, 2059 // and the original comparison may be either equality or inequality. 2060 if (Cond->getPredicate() == CmpInst::ICMP_EQ) 2061 Pred = CmpInst::getInversePredicate(Pred); 2062 2063 // Ok, everything looks ok to change the condition into an SLT or SGE and 2064 // delete the max calculation. 2065 ICmpInst *NewCond = 2066 new ICmpInst(Cond, Pred, Cond->getOperand(0), NewRHS, "scmp"); 2067 2068 // Delete the max calculation instructions. 2069 Cond->replaceAllUsesWith(NewCond); 2070 CondUse->setUser(NewCond); 2071 Instruction *Cmp = cast<Instruction>(Sel->getOperand(0)); 2072 Cond->eraseFromParent(); 2073 Sel->eraseFromParent(); 2074 if (Cmp->use_empty()) 2075 Cmp->eraseFromParent(); 2076 return NewCond; 2077} 2078 2079/// OptimizeLoopTermCond - Change loop terminating condition to use the 2080/// postinc iv when possible. 2081void 2082LSRInstance::OptimizeLoopTermCond() { 2083 SmallPtrSet<Instruction *, 4> PostIncs; 2084 2085 BasicBlock *LatchBlock = L->getLoopLatch(); 2086 SmallVector<BasicBlock*, 8> ExitingBlocks; 2087 L->getExitingBlocks(ExitingBlocks); 2088 2089 for (BasicBlock *ExitingBlock : ExitingBlocks) { 2090 2091 // Get the terminating condition for the loop if possible. If we 2092 // can, we want to change it to use a post-incremented version of its 2093 // induction variable, to allow coalescing the live ranges for the IV into 2094 // one register value. 2095 2096 BranchInst *TermBr = dyn_cast<BranchInst>(ExitingBlock->getTerminator()); 2097 if (!TermBr) 2098 continue; 2099 // FIXME: Overly conservative, termination condition could be an 'or' etc.. 2100 if (TermBr->isUnconditional() || !isa<ICmpInst>(TermBr->getCondition())) 2101 continue; 2102 2103 // Search IVUsesByStride to find Cond's IVUse if there is one. 2104 IVStrideUse *CondUse = nullptr; 2105 ICmpInst *Cond = cast<ICmpInst>(TermBr->getCondition()); 2106 if (!FindIVUserForCond(Cond, CondUse)) 2107 continue; 2108 2109 // If the trip count is computed in terms of a max (due to ScalarEvolution 2110 // being unable to find a sufficient guard, for example), change the loop 2111 // comparison to use SLT or ULT instead of NE. 2112 // One consequence of doing this now is that it disrupts the count-down 2113 // optimization. That's not always a bad thing though, because in such 2114 // cases it may still be worthwhile to avoid a max. 2115 Cond = OptimizeMax(Cond, CondUse); 2116 2117 // If this exiting block dominates the latch block, it may also use 2118 // the post-inc value if it won't be shared with other uses. 2119 // Check for dominance. 2120 if (!DT.dominates(ExitingBlock, LatchBlock)) 2121 continue; 2122 2123 // Conservatively avoid trying to use the post-inc value in non-latch 2124 // exits if there may be pre-inc users in intervening blocks. 2125 if (LatchBlock != ExitingBlock) 2126 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI) 2127 // Test if the use is reachable from the exiting block. This dominator 2128 // query is a conservative approximation of reachability. 2129 if (&*UI != CondUse && 2130 !DT.properlyDominates(UI->getUser()->getParent(), ExitingBlock)) { 2131 // Conservatively assume there may be reuse if the quotient of their 2132 // strides could be a legal scale. 2133 const SCEV *A = IU.getStride(*CondUse, L); 2134 const SCEV *B = IU.getStride(*UI, L); 2135 if (!A || !B) continue; 2136 if (SE.getTypeSizeInBits(A->getType()) != 2137 SE.getTypeSizeInBits(B->getType())) { 2138 if (SE.getTypeSizeInBits(A->getType()) > 2139 SE.getTypeSizeInBits(B->getType())) 2140 B = SE.getSignExtendExpr(B, A->getType()); 2141 else 2142 A = SE.getSignExtendExpr(A, B->getType()); 2143 } 2144 if (const SCEVConstant *D = 2145 dyn_cast_or_null<SCEVConstant>(getExactSDiv(B, A, SE))) { 2146 const ConstantInt *C = D->getValue(); 2147 // Stride of one or negative one can have reuse with non-addresses. 2148 if (C->isOne() || C->isAllOnesValue()) 2149 goto decline_post_inc; 2150 // Avoid weird situations. 2151 if (C->getValue().getMinSignedBits() >= 64 || 2152 C->getValue().isMinSignedValue()) 2153 goto decline_post_inc; 2154 // Check for possible scaled-address reuse. 2155 Type *AccessTy = getAccessType(UI->getUser()); 2156 int64_t Scale = C->getSExtValue(); 2157 if (TTI.isLegalAddressingMode(AccessTy, /*BaseGV=*/ nullptr, 2158 /*BaseOffset=*/ 0, 2159 /*HasBaseReg=*/ false, Scale)) 2160 goto decline_post_inc; 2161 Scale = -Scale; 2162 if (TTI.isLegalAddressingMode(AccessTy, /*BaseGV=*/ nullptr, 2163 /*BaseOffset=*/ 0, 2164 /*HasBaseReg=*/ false, Scale)) 2165 goto decline_post_inc; 2166 } 2167 } 2168 2169 DEBUG(dbgs() << " Change loop exiting icmp to use postinc iv: " 2170 << *Cond << '\n'); 2171 2172 // It's possible for the setcc instruction to be anywhere in the loop, and 2173 // possible for it to have multiple users. If it is not immediately before 2174 // the exiting block branch, move it. 2175 if (&*++BasicBlock::iterator(Cond) != TermBr) { 2176 if (Cond->hasOneUse()) { 2177 Cond->moveBefore(TermBr); 2178 } else { 2179 // Clone the terminating condition and insert into the loopend. 2180 ICmpInst *OldCond = Cond; 2181 Cond = cast<ICmpInst>(Cond->clone()); 2182 Cond->setName(L->getHeader()->getName() + ".termcond"); 2183 ExitingBlock->getInstList().insert(TermBr, Cond); 2184 2185 // Clone the IVUse, as the old use still exists! 2186 CondUse = &IU.AddUser(Cond, CondUse->getOperandValToReplace()); 2187 TermBr->replaceUsesOfWith(OldCond, Cond); 2188 } 2189 } 2190 2191 // If we get to here, we know that we can transform the setcc instruction to 2192 // use the post-incremented version of the IV, allowing us to coalesce the 2193 // live ranges for the IV correctly. 2194 CondUse->transformToPostInc(L); 2195 Changed = true; 2196 2197 PostIncs.insert(Cond); 2198 decline_post_inc:; 2199 } 2200 2201 // Determine an insertion point for the loop induction variable increment. It 2202 // must dominate all the post-inc comparisons we just set up, and it must 2203 // dominate the loop latch edge. 2204 IVIncInsertPos = L->getLoopLatch()->getTerminator(); 2205 for (Instruction *Inst : PostIncs) { 2206 BasicBlock *BB = 2207 DT.findNearestCommonDominator(IVIncInsertPos->getParent(), 2208 Inst->getParent()); 2209 if (BB == Inst->getParent()) 2210 IVIncInsertPos = Inst; 2211 else if (BB != IVIncInsertPos->getParent()) 2212 IVIncInsertPos = BB->getTerminator(); 2213 } 2214} 2215 2216/// reconcileNewOffset - Determine if the given use can accommodate a fixup 2217/// at the given offset and other details. If so, update the use and 2218/// return true. 2219bool 2220LSRInstance::reconcileNewOffset(LSRUse &LU, int64_t NewOffset, bool HasBaseReg, 2221 LSRUse::KindType Kind, Type *AccessTy) { 2222 int64_t NewMinOffset = LU.MinOffset; 2223 int64_t NewMaxOffset = LU.MaxOffset; 2224 Type *NewAccessTy = AccessTy; 2225 2226 // Check for a mismatched kind. It's tempting to collapse mismatched kinds to 2227 // something conservative, however this can pessimize in the case that one of 2228 // the uses will have all its uses outside the loop, for example. 2229 if (LU.Kind != Kind) 2230 return false; 2231 2232 // Check for a mismatched access type, and fall back conservatively as needed. 2233 // TODO: Be less conservative when the type is similar and can use the same 2234 // addressing modes. 2235 if (Kind == LSRUse::Address && AccessTy != LU.AccessTy) 2236 NewAccessTy = Type::getVoidTy(AccessTy->getContext()); 2237 2238 // Conservatively assume HasBaseReg is true for now. 2239 if (NewOffset < LU.MinOffset) { 2240 if (!isAlwaysFoldable(TTI, Kind, NewAccessTy, /*BaseGV=*/nullptr, 2241 LU.MaxOffset - NewOffset, HasBaseReg)) 2242 return false; 2243 NewMinOffset = NewOffset; 2244 } else if (NewOffset > LU.MaxOffset) { 2245 if (!isAlwaysFoldable(TTI, Kind, NewAccessTy, /*BaseGV=*/nullptr, 2246 NewOffset - LU.MinOffset, HasBaseReg)) 2247 return false; 2248 NewMaxOffset = NewOffset; 2249 } 2250 2251 // Update the use. 2252 LU.MinOffset = NewMinOffset; 2253 LU.MaxOffset = NewMaxOffset; 2254 LU.AccessTy = NewAccessTy; 2255 if (NewOffset != LU.Offsets.back()) 2256 LU.Offsets.push_back(NewOffset); 2257 return true; 2258} 2259 2260/// getUse - Return an LSRUse index and an offset value for a fixup which 2261/// needs the given expression, with the given kind and optional access type. 2262/// Either reuse an existing use or create a new one, as needed. 2263std::pair<size_t, int64_t> 2264LSRInstance::getUse(const SCEV *&Expr, 2265 LSRUse::KindType Kind, Type *AccessTy) { 2266 const SCEV *Copy = Expr; 2267 int64_t Offset = ExtractImmediate(Expr, SE); 2268 2269 // Basic uses can't accept any offset, for example. 2270 if (!isAlwaysFoldable(TTI, Kind, AccessTy, /*BaseGV=*/ nullptr, 2271 Offset, /*HasBaseReg=*/ true)) { 2272 Expr = Copy; 2273 Offset = 0; 2274 } 2275 2276 std::pair<UseMapTy::iterator, bool> P = 2277 UseMap.insert(std::make_pair(LSRUse::SCEVUseKindPair(Expr, Kind), 0)); 2278 if (!P.second) { 2279 // A use already existed with this base. 2280 size_t LUIdx = P.first->second; 2281 LSRUse &LU = Uses[LUIdx]; 2282 if (reconcileNewOffset(LU, Offset, /*HasBaseReg=*/true, Kind, AccessTy)) 2283 // Reuse this use. 2284 return std::make_pair(LUIdx, Offset); 2285 } 2286 2287 // Create a new use. 2288 size_t LUIdx = Uses.size(); 2289 P.first->second = LUIdx; 2290 Uses.push_back(LSRUse(Kind, AccessTy)); 2291 LSRUse &LU = Uses[LUIdx]; 2292 2293 // We don't need to track redundant offsets, but we don't need to go out 2294 // of our way here to avoid them. 2295 if (LU.Offsets.empty() || Offset != LU.Offsets.back()) 2296 LU.Offsets.push_back(Offset); 2297 2298 LU.MinOffset = Offset; 2299 LU.MaxOffset = Offset; 2300 return std::make_pair(LUIdx, Offset); 2301} 2302 2303/// DeleteUse - Delete the given use from the Uses list. 2304void LSRInstance::DeleteUse(LSRUse &LU, size_t LUIdx) { 2305 if (&LU != &Uses.back()) 2306 std::swap(LU, Uses.back()); 2307 Uses.pop_back(); 2308 2309 // Update RegUses. 2310 RegUses.SwapAndDropUse(LUIdx, Uses.size()); 2311} 2312 2313/// FindUseWithFormula - Look for a use distinct from OrigLU which is has 2314/// a formula that has the same registers as the given formula. 2315LSRUse * 2316LSRInstance::FindUseWithSimilarFormula(const Formula &OrigF, 2317 const LSRUse &OrigLU) { 2318 // Search all uses for the formula. This could be more clever. 2319 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) { 2320 LSRUse &LU = Uses[LUIdx]; 2321 // Check whether this use is close enough to OrigLU, to see whether it's 2322 // worthwhile looking through its formulae. 2323 // Ignore ICmpZero uses because they may contain formulae generated by 2324 // GenerateICmpZeroScales, in which case adding fixup offsets may 2325 // be invalid. 2326 if (&LU != &OrigLU && 2327 LU.Kind != LSRUse::ICmpZero && 2328 LU.Kind == OrigLU.Kind && OrigLU.AccessTy == LU.AccessTy && 2329 LU.WidestFixupType == OrigLU.WidestFixupType && 2330 LU.HasFormulaWithSameRegs(OrigF)) { 2331 // Scan through this use's formulae. 2332 for (const Formula &F : LU.Formulae) { 2333 // Check to see if this formula has the same registers and symbols 2334 // as OrigF. 2335 if (F.BaseRegs == OrigF.BaseRegs && 2336 F.ScaledReg == OrigF.ScaledReg && 2337 F.BaseGV == OrigF.BaseGV && 2338 F.Scale == OrigF.Scale && 2339 F.UnfoldedOffset == OrigF.UnfoldedOffset) { 2340 if (F.BaseOffset == 0) 2341 return &LU; 2342 // This is the formula where all the registers and symbols matched; 2343 // there aren't going to be any others. Since we declined it, we 2344 // can skip the rest of the formulae and proceed to the next LSRUse. 2345 break; 2346 } 2347 } 2348 } 2349 } 2350 2351 // Nothing looked good. 2352 return nullptr; 2353} 2354 2355void LSRInstance::CollectInterestingTypesAndFactors() { 2356 SmallSetVector<const SCEV *, 4> Strides; 2357 2358 // Collect interesting types and strides. 2359 SmallVector<const SCEV *, 4> Worklist; 2360 for (const IVStrideUse &U : IU) { 2361 const SCEV *Expr = IU.getExpr(U); 2362 2363 // Collect interesting types. 2364 Types.insert(SE.getEffectiveSCEVType(Expr->getType())); 2365 2366 // Add strides for mentioned loops. 2367 Worklist.push_back(Expr); 2368 do { 2369 const SCEV *S = Worklist.pop_back_val(); 2370 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) { 2371 if (AR->getLoop() == L) 2372 Strides.insert(AR->getStepRecurrence(SE)); 2373 Worklist.push_back(AR->getStart()); 2374 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) { 2375 Worklist.append(Add->op_begin(), Add->op_end()); 2376 } 2377 } while (!Worklist.empty()); 2378 } 2379 2380 // Compute interesting factors from the set of interesting strides. 2381 for (SmallSetVector<const SCEV *, 4>::const_iterator 2382 I = Strides.begin(), E = Strides.end(); I != E; ++I) 2383 for (SmallSetVector<const SCEV *, 4>::const_iterator NewStrideIter = 2384 std::next(I); NewStrideIter != E; ++NewStrideIter) { 2385 const SCEV *OldStride = *I; 2386 const SCEV *NewStride = *NewStrideIter; 2387 2388 if (SE.getTypeSizeInBits(OldStride->getType()) != 2389 SE.getTypeSizeInBits(NewStride->getType())) { 2390 if (SE.getTypeSizeInBits(OldStride->getType()) > 2391 SE.getTypeSizeInBits(NewStride->getType())) 2392 NewStride = SE.getSignExtendExpr(NewStride, OldStride->getType()); 2393 else 2394 OldStride = SE.getSignExtendExpr(OldStride, NewStride->getType()); 2395 } 2396 if (const SCEVConstant *Factor = 2397 dyn_cast_or_null<SCEVConstant>(getExactSDiv(NewStride, OldStride, 2398 SE, true))) { 2399 if (Factor->getValue()->getValue().getMinSignedBits() <= 64) 2400 Factors.insert(Factor->getValue()->getValue().getSExtValue()); 2401 } else if (const SCEVConstant *Factor = 2402 dyn_cast_or_null<SCEVConstant>(getExactSDiv(OldStride, 2403 NewStride, 2404 SE, true))) { 2405 if (Factor->getValue()->getValue().getMinSignedBits() <= 64) 2406 Factors.insert(Factor->getValue()->getValue().getSExtValue()); 2407 } 2408 } 2409 2410 // If all uses use the same type, don't bother looking for truncation-based 2411 // reuse. 2412 if (Types.size() == 1) 2413 Types.clear(); 2414 2415 DEBUG(print_factors_and_types(dbgs())); 2416} 2417 2418/// findIVOperand - Helper for CollectChains that finds an IV operand (computed 2419/// by an AddRec in this loop) within [OI,OE) or returns OE. If IVUsers mapped 2420/// Instructions to IVStrideUses, we could partially skip this. 2421static User::op_iterator 2422findIVOperand(User::op_iterator OI, User::op_iterator OE, 2423 Loop *L, ScalarEvolution &SE) { 2424 for(; OI != OE; ++OI) { 2425 if (Instruction *Oper = dyn_cast<Instruction>(*OI)) { 2426 if (!SE.isSCEVable(Oper->getType())) 2427 continue; 2428 2429 if (const SCEVAddRecExpr *AR = 2430 dyn_cast<SCEVAddRecExpr>(SE.getSCEV(Oper))) { 2431 if (AR->getLoop() == L) 2432 break; 2433 } 2434 } 2435 } 2436 return OI; 2437} 2438 2439/// getWideOperand - IVChain logic must consistenctly peek base TruncInst 2440/// operands, so wrap it in a convenient helper. 2441static Value *getWideOperand(Value *Oper) { 2442 if (TruncInst *Trunc = dyn_cast<TruncInst>(Oper)) 2443 return Trunc->getOperand(0); 2444 return Oper; 2445} 2446 2447/// isCompatibleIVType - Return true if we allow an IV chain to include both 2448/// types. 2449static bool isCompatibleIVType(Value *LVal, Value *RVal) { 2450 Type *LType = LVal->getType(); 2451 Type *RType = RVal->getType(); 2452 return (LType == RType) || (LType->isPointerTy() && RType->isPointerTy()); 2453} 2454 2455/// getExprBase - Return an approximation of this SCEV expression's "base", or 2456/// NULL for any constant. Returning the expression itself is 2457/// conservative. Returning a deeper subexpression is more precise and valid as 2458/// long as it isn't less complex than another subexpression. For expressions 2459/// involving multiple unscaled values, we need to return the pointer-type 2460/// SCEVUnknown. This avoids forming chains across objects, such as: 2461/// PrevOper==a[i], IVOper==b[i], IVInc==b-a. 2462/// 2463/// Since SCEVUnknown is the rightmost type, and pointers are the rightmost 2464/// SCEVUnknown, we simply return the rightmost SCEV operand. 2465static const SCEV *getExprBase(const SCEV *S) { 2466 switch (S->getSCEVType()) { 2467 default: // uncluding scUnknown. 2468 return S; 2469 case scConstant: 2470 return nullptr; 2471 case scTruncate: 2472 return getExprBase(cast<SCEVTruncateExpr>(S)->getOperand()); 2473 case scZeroExtend: 2474 return getExprBase(cast<SCEVZeroExtendExpr>(S)->getOperand()); 2475 case scSignExtend: 2476 return getExprBase(cast<SCEVSignExtendExpr>(S)->getOperand()); 2477 case scAddExpr: { 2478 // Skip over scaled operands (scMulExpr) to follow add operands as long as 2479 // there's nothing more complex. 2480 // FIXME: not sure if we want to recognize negation. 2481 const SCEVAddExpr *Add = cast<SCEVAddExpr>(S); 2482 for (std::reverse_iterator<SCEVAddExpr::op_iterator> I(Add->op_end()), 2483 E(Add->op_begin()); I != E; ++I) { 2484 const SCEV *SubExpr = *I; 2485 if (SubExpr->getSCEVType() == scAddExpr) 2486 return getExprBase(SubExpr); 2487 2488 if (SubExpr->getSCEVType() != scMulExpr) 2489 return SubExpr; 2490 } 2491 return S; // all operands are scaled, be conservative. 2492 } 2493 case scAddRecExpr: 2494 return getExprBase(cast<SCEVAddRecExpr>(S)->getStart()); 2495 } 2496} 2497 2498/// Return true if the chain increment is profitable to expand into a loop 2499/// invariant value, which may require its own register. A profitable chain 2500/// increment will be an offset relative to the same base. We allow such offsets 2501/// to potentially be used as chain increment as long as it's not obviously 2502/// expensive to expand using real instructions. 2503bool IVChain::isProfitableIncrement(const SCEV *OperExpr, 2504 const SCEV *IncExpr, 2505 ScalarEvolution &SE) { 2506 // Aggressively form chains when -stress-ivchain. 2507 if (StressIVChain) 2508 return true; 2509 2510 // Do not replace a constant offset from IV head with a nonconstant IV 2511 // increment. 2512 if (!isa<SCEVConstant>(IncExpr)) { 2513 const SCEV *HeadExpr = SE.getSCEV(getWideOperand(Incs[0].IVOperand)); 2514 if (isa<SCEVConstant>(SE.getMinusSCEV(OperExpr, HeadExpr))) 2515 return 0; 2516 } 2517 2518 SmallPtrSet<const SCEV*, 8> Processed; 2519 return !isHighCostExpansion(IncExpr, Processed, SE); 2520} 2521 2522/// Return true if the number of registers needed for the chain is estimated to 2523/// be less than the number required for the individual IV users. First prohibit 2524/// any IV users that keep the IV live across increments (the Users set should 2525/// be empty). Next count the number and type of increments in the chain. 2526/// 2527/// Chaining IVs can lead to considerable code bloat if ISEL doesn't 2528/// effectively use postinc addressing modes. Only consider it profitable it the 2529/// increments can be computed in fewer registers when chained. 2530/// 2531/// TODO: Consider IVInc free if it's already used in another chains. 2532static bool 2533isProfitableChain(IVChain &Chain, SmallPtrSetImpl<Instruction*> &Users, 2534 ScalarEvolution &SE, const TargetTransformInfo &TTI) { 2535 if (StressIVChain) 2536 return true; 2537 2538 if (!Chain.hasIncs()) 2539 return false; 2540 2541 if (!Users.empty()) { 2542 DEBUG(dbgs() << "Chain: " << *Chain.Incs[0].UserInst << " users:\n"; 2543 for (Instruction *Inst : Users) { 2544 dbgs() << " " << *Inst << "\n"; 2545 }); 2546 return false; 2547 } 2548 assert(!Chain.Incs.empty() && "empty IV chains are not allowed"); 2549 2550 // The chain itself may require a register, so intialize cost to 1. 2551 int cost = 1; 2552 2553 // A complete chain likely eliminates the need for keeping the original IV in 2554 // a register. LSR does not currently know how to form a complete chain unless 2555 // the header phi already exists. 2556 if (isa<PHINode>(Chain.tailUserInst()) 2557 && SE.getSCEV(Chain.tailUserInst()) == Chain.Incs[0].IncExpr) { 2558 --cost; 2559 } 2560 const SCEV *LastIncExpr = nullptr; 2561 unsigned NumConstIncrements = 0; 2562 unsigned NumVarIncrements = 0; 2563 unsigned NumReusedIncrements = 0; 2564 for (const IVInc &Inc : Chain) { 2565 if (Inc.IncExpr->isZero()) 2566 continue; 2567 2568 // Incrementing by zero or some constant is neutral. We assume constants can 2569 // be folded into an addressing mode or an add's immediate operand. 2570 if (isa<SCEVConstant>(Inc.IncExpr)) { 2571 ++NumConstIncrements; 2572 continue; 2573 } 2574 2575 if (Inc.IncExpr == LastIncExpr) 2576 ++NumReusedIncrements; 2577 else 2578 ++NumVarIncrements; 2579 2580 LastIncExpr = Inc.IncExpr; 2581 } 2582 // An IV chain with a single increment is handled by LSR's postinc 2583 // uses. However, a chain with multiple increments requires keeping the IV's 2584 // value live longer than it needs to be if chained. 2585 if (NumConstIncrements > 1) 2586 --cost; 2587 2588 // Materializing increment expressions in the preheader that didn't exist in 2589 // the original code may cost a register. For example, sign-extended array 2590 // indices can produce ridiculous increments like this: 2591 // IV + ((sext i32 (2 * %s) to i64) + (-1 * (sext i32 %s to i64))) 2592 cost += NumVarIncrements; 2593 2594 // Reusing variable increments likely saves a register to hold the multiple of 2595 // the stride. 2596 cost -= NumReusedIncrements; 2597 2598 DEBUG(dbgs() << "Chain: " << *Chain.Incs[0].UserInst << " Cost: " << cost 2599 << "\n"); 2600 2601 return cost < 0; 2602} 2603 2604/// ChainInstruction - Add this IV user to an existing chain or make it the head 2605/// of a new chain. 2606void LSRInstance::ChainInstruction(Instruction *UserInst, Instruction *IVOper, 2607 SmallVectorImpl<ChainUsers> &ChainUsersVec) { 2608 // When IVs are used as types of varying widths, they are generally converted 2609 // to a wider type with some uses remaining narrow under a (free) trunc. 2610 Value *const NextIV = getWideOperand(IVOper); 2611 const SCEV *const OperExpr = SE.getSCEV(NextIV); 2612 const SCEV *const OperExprBase = getExprBase(OperExpr); 2613 2614 // Visit all existing chains. Check if its IVOper can be computed as a 2615 // profitable loop invariant increment from the last link in the Chain. 2616 unsigned ChainIdx = 0, NChains = IVChainVec.size(); 2617 const SCEV *LastIncExpr = nullptr; 2618 for (; ChainIdx < NChains; ++ChainIdx) { 2619 IVChain &Chain = IVChainVec[ChainIdx]; 2620 2621 // Prune the solution space aggressively by checking that both IV operands 2622 // are expressions that operate on the same unscaled SCEVUnknown. This 2623 // "base" will be canceled by the subsequent getMinusSCEV call. Checking 2624 // first avoids creating extra SCEV expressions. 2625 if (!StressIVChain && Chain.ExprBase != OperExprBase) 2626 continue; 2627 2628 Value *PrevIV = getWideOperand(Chain.Incs.back().IVOperand); 2629 if (!isCompatibleIVType(PrevIV, NextIV)) 2630 continue; 2631 2632 // A phi node terminates a chain. 2633 if (isa<PHINode>(UserInst) && isa<PHINode>(Chain.tailUserInst())) 2634 continue; 2635 2636 // The increment must be loop-invariant so it can be kept in a register. 2637 const SCEV *PrevExpr = SE.getSCEV(PrevIV); 2638 const SCEV *IncExpr = SE.getMinusSCEV(OperExpr, PrevExpr); 2639 if (!SE.isLoopInvariant(IncExpr, L)) 2640 continue; 2641 2642 if (Chain.isProfitableIncrement(OperExpr, IncExpr, SE)) { 2643 LastIncExpr = IncExpr; 2644 break; 2645 } 2646 } 2647 // If we haven't found a chain, create a new one, unless we hit the max. Don't 2648 // bother for phi nodes, because they must be last in the chain. 2649 if (ChainIdx == NChains) { 2650 if (isa<PHINode>(UserInst)) 2651 return; 2652 if (NChains >= MaxChains && !StressIVChain) { 2653 DEBUG(dbgs() << "IV Chain Limit\n"); 2654 return; 2655 } 2656 LastIncExpr = OperExpr; 2657 // IVUsers may have skipped over sign/zero extensions. We don't currently 2658 // attempt to form chains involving extensions unless they can be hoisted 2659 // into this loop's AddRec. 2660 if (!isa<SCEVAddRecExpr>(LastIncExpr)) 2661 return; 2662 ++NChains; 2663 IVChainVec.push_back(IVChain(IVInc(UserInst, IVOper, LastIncExpr), 2664 OperExprBase)); 2665 ChainUsersVec.resize(NChains); 2666 DEBUG(dbgs() << "IV Chain#" << ChainIdx << " Head: (" << *UserInst 2667 << ") IV=" << *LastIncExpr << "\n"); 2668 } else { 2669 DEBUG(dbgs() << "IV Chain#" << ChainIdx << " Inc: (" << *UserInst 2670 << ") IV+" << *LastIncExpr << "\n"); 2671 // Add this IV user to the end of the chain. 2672 IVChainVec[ChainIdx].add(IVInc(UserInst, IVOper, LastIncExpr)); 2673 } 2674 IVChain &Chain = IVChainVec[ChainIdx]; 2675 2676 SmallPtrSet<Instruction*,4> &NearUsers = ChainUsersVec[ChainIdx].NearUsers; 2677 // This chain's NearUsers become FarUsers. 2678 if (!LastIncExpr->isZero()) { 2679 ChainUsersVec[ChainIdx].FarUsers.insert(NearUsers.begin(), 2680 NearUsers.end()); 2681 NearUsers.clear(); 2682 } 2683 2684 // All other uses of IVOperand become near uses of the chain. 2685 // We currently ignore intermediate values within SCEV expressions, assuming 2686 // they will eventually be used be the current chain, or can be computed 2687 // from one of the chain increments. To be more precise we could 2688 // transitively follow its user and only add leaf IV users to the set. 2689 for (User *U : IVOper->users()) { 2690 Instruction *OtherUse = dyn_cast<Instruction>(U); 2691 if (!OtherUse) 2692 continue; 2693 // Uses in the chain will no longer be uses if the chain is formed. 2694 // Include the head of the chain in this iteration (not Chain.begin()). 2695 IVChain::const_iterator IncIter = Chain.Incs.begin(); 2696 IVChain::const_iterator IncEnd = Chain.Incs.end(); 2697 for( ; IncIter != IncEnd; ++IncIter) { 2698 if (IncIter->UserInst == OtherUse) 2699 break; 2700 } 2701 if (IncIter != IncEnd) 2702 continue; 2703 2704 if (SE.isSCEVable(OtherUse->getType()) 2705 && !isa<SCEVUnknown>(SE.getSCEV(OtherUse)) 2706 && IU.isIVUserOrOperand(OtherUse)) { 2707 continue; 2708 } 2709 NearUsers.insert(OtherUse); 2710 } 2711 2712 // Since this user is part of the chain, it's no longer considered a use 2713 // of the chain. 2714 ChainUsersVec[ChainIdx].FarUsers.erase(UserInst); 2715} 2716 2717/// CollectChains - Populate the vector of Chains. 2718/// 2719/// This decreases ILP at the architecture level. Targets with ample registers, 2720/// multiple memory ports, and no register renaming probably don't want 2721/// this. However, such targets should probably disable LSR altogether. 2722/// 2723/// The job of LSR is to make a reasonable choice of induction variables across 2724/// the loop. Subsequent passes can easily "unchain" computation exposing more 2725/// ILP *within the loop* if the target wants it. 2726/// 2727/// Finding the best IV chain is potentially a scheduling problem. Since LSR 2728/// will not reorder memory operations, it will recognize this as a chain, but 2729/// will generate redundant IV increments. Ideally this would be corrected later 2730/// by a smart scheduler: 2731/// = A[i] 2732/// = A[i+x] 2733/// A[i] = 2734/// A[i+x] = 2735/// 2736/// TODO: Walk the entire domtree within this loop, not just the path to the 2737/// loop latch. This will discover chains on side paths, but requires 2738/// maintaining multiple copies of the Chains state. 2739void LSRInstance::CollectChains() { 2740 DEBUG(dbgs() << "Collecting IV Chains.\n"); 2741 SmallVector<ChainUsers, 8> ChainUsersVec; 2742 2743 SmallVector<BasicBlock *,8> LatchPath; 2744 BasicBlock *LoopHeader = L->getHeader(); 2745 for (DomTreeNode *Rung = DT.getNode(L->getLoopLatch()); 2746 Rung->getBlock() != LoopHeader; Rung = Rung->getIDom()) { 2747 LatchPath.push_back(Rung->getBlock()); 2748 } 2749 LatchPath.push_back(LoopHeader); 2750 2751 // Walk the instruction stream from the loop header to the loop latch. 2752 for (SmallVectorImpl<BasicBlock *>::reverse_iterator 2753 BBIter = LatchPath.rbegin(), BBEnd = LatchPath.rend(); 2754 BBIter != BBEnd; ++BBIter) { 2755 for (BasicBlock::iterator I = (*BBIter)->begin(), E = (*BBIter)->end(); 2756 I != E; ++I) { 2757 // Skip instructions that weren't seen by IVUsers analysis. 2758 if (isa<PHINode>(I) || !IU.isIVUserOrOperand(I)) 2759 continue; 2760 2761 // Ignore users that are part of a SCEV expression. This way we only 2762 // consider leaf IV Users. This effectively rediscovers a portion of 2763 // IVUsers analysis but in program order this time. 2764 if (SE.isSCEVable(I->getType()) && !isa<SCEVUnknown>(SE.getSCEV(I))) 2765 continue; 2766 2767 // Remove this instruction from any NearUsers set it may be in. 2768 for (unsigned ChainIdx = 0, NChains = IVChainVec.size(); 2769 ChainIdx < NChains; ++ChainIdx) { 2770 ChainUsersVec[ChainIdx].NearUsers.erase(I); 2771 } 2772 // Search for operands that can be chained. 2773 SmallPtrSet<Instruction*, 4> UniqueOperands; 2774 User::op_iterator IVOpEnd = I->op_end(); 2775 User::op_iterator IVOpIter = findIVOperand(I->op_begin(), IVOpEnd, L, SE); 2776 while (IVOpIter != IVOpEnd) { 2777 Instruction *IVOpInst = cast<Instruction>(*IVOpIter); 2778 if (UniqueOperands.insert(IVOpInst).second) 2779 ChainInstruction(I, IVOpInst, ChainUsersVec); 2780 IVOpIter = findIVOperand(std::next(IVOpIter), IVOpEnd, L, SE); 2781 } 2782 } // Continue walking down the instructions. 2783 } // Continue walking down the domtree. 2784 // Visit phi backedges to determine if the chain can generate the IV postinc. 2785 for (BasicBlock::iterator I = L->getHeader()->begin(); 2786 PHINode *PN = dyn_cast<PHINode>(I); ++I) { 2787 if (!SE.isSCEVable(PN->getType())) 2788 continue; 2789 2790 Instruction *IncV = 2791 dyn_cast<Instruction>(PN->getIncomingValueForBlock(L->getLoopLatch())); 2792 if (IncV) 2793 ChainInstruction(PN, IncV, ChainUsersVec); 2794 } 2795 // Remove any unprofitable chains. 2796 unsigned ChainIdx = 0; 2797 for (unsigned UsersIdx = 0, NChains = IVChainVec.size(); 2798 UsersIdx < NChains; ++UsersIdx) { 2799 if (!isProfitableChain(IVChainVec[UsersIdx], 2800 ChainUsersVec[UsersIdx].FarUsers, SE, TTI)) 2801 continue; 2802 // Preserve the chain at UsesIdx. 2803 if (ChainIdx != UsersIdx) 2804 IVChainVec[ChainIdx] = IVChainVec[UsersIdx]; 2805 FinalizeChain(IVChainVec[ChainIdx]); 2806 ++ChainIdx; 2807 } 2808 IVChainVec.resize(ChainIdx); 2809} 2810 2811void LSRInstance::FinalizeChain(IVChain &Chain) { 2812 assert(!Chain.Incs.empty() && "empty IV chains are not allowed"); 2813 DEBUG(dbgs() << "Final Chain: " << *Chain.Incs[0].UserInst << "\n"); 2814 2815 for (const IVInc &Inc : Chain) { 2816 DEBUG(dbgs() << " Inc: " << Inc.UserInst << "\n"); 2817 auto UseI = std::find(Inc.UserInst->op_begin(), Inc.UserInst->op_end(), 2818 Inc.IVOperand); 2819 assert(UseI != Inc.UserInst->op_end() && "cannot find IV operand"); 2820 IVIncSet.insert(UseI); 2821 } 2822} 2823 2824/// Return true if the IVInc can be folded into an addressing mode. 2825static bool canFoldIVIncExpr(const SCEV *IncExpr, Instruction *UserInst, 2826 Value *Operand, const TargetTransformInfo &TTI) { 2827 const SCEVConstant *IncConst = dyn_cast<SCEVConstant>(IncExpr); 2828 if (!IncConst || !isAddressUse(UserInst, Operand)) 2829 return false; 2830 2831 if (IncConst->getValue()->getValue().getMinSignedBits() > 64) 2832 return false; 2833 2834 int64_t IncOffset = IncConst->getValue()->getSExtValue(); 2835 if (!isAlwaysFoldable(TTI, LSRUse::Address, 2836 getAccessType(UserInst), /*BaseGV=*/ nullptr, 2837 IncOffset, /*HaseBaseReg=*/ false)) 2838 return false; 2839 2840 return true; 2841} 2842 2843/// GenerateIVChains - Generate an add or subtract for each IVInc in a chain to 2844/// materialize the IV user's operand from the previous IV user's operand. 2845void LSRInstance::GenerateIVChain(const IVChain &Chain, SCEVExpander &Rewriter, 2846 SmallVectorImpl<WeakVH> &DeadInsts) { 2847 // Find the new IVOperand for the head of the chain. It may have been replaced 2848 // by LSR. 2849 const IVInc &Head = Chain.Incs[0]; 2850 User::op_iterator IVOpEnd = Head.UserInst->op_end(); 2851 // findIVOperand returns IVOpEnd if it can no longer find a valid IV user. 2852 User::op_iterator IVOpIter = findIVOperand(Head.UserInst->op_begin(), 2853 IVOpEnd, L, SE); 2854 Value *IVSrc = nullptr; 2855 while (IVOpIter != IVOpEnd) { 2856 IVSrc = getWideOperand(*IVOpIter); 2857 2858 // If this operand computes the expression that the chain needs, we may use 2859 // it. (Check this after setting IVSrc which is used below.) 2860 // 2861 // Note that if Head.IncExpr is wider than IVSrc, then this phi is too 2862 // narrow for the chain, so we can no longer use it. We do allow using a 2863 // wider phi, assuming the LSR checked for free truncation. In that case we 2864 // should already have a truncate on this operand such that 2865 // getSCEV(IVSrc) == IncExpr. 2866 if (SE.getSCEV(*IVOpIter) == Head.IncExpr 2867 || SE.getSCEV(IVSrc) == Head.IncExpr) { 2868 break; 2869 } 2870 IVOpIter = findIVOperand(std::next(IVOpIter), IVOpEnd, L, SE); 2871 } 2872 if (IVOpIter == IVOpEnd) { 2873 // Gracefully give up on this chain. 2874 DEBUG(dbgs() << "Concealed chain head: " << *Head.UserInst << "\n"); 2875 return; 2876 } 2877 2878 DEBUG(dbgs() << "Generate chain at: " << *IVSrc << "\n"); 2879 Type *IVTy = IVSrc->getType(); 2880 Type *IntTy = SE.getEffectiveSCEVType(IVTy); 2881 const SCEV *LeftOverExpr = nullptr; 2882 for (const IVInc &Inc : Chain) { 2883 Instruction *InsertPt = Inc.UserInst; 2884 if (isa<PHINode>(InsertPt)) 2885 InsertPt = L->getLoopLatch()->getTerminator(); 2886 2887 // IVOper will replace the current IV User's operand. IVSrc is the IV 2888 // value currently held in a register. 2889 Value *IVOper = IVSrc; 2890 if (!Inc.IncExpr->isZero()) { 2891 // IncExpr was the result of subtraction of two narrow values, so must 2892 // be signed. 2893 const SCEV *IncExpr = SE.getNoopOrSignExtend(Inc.IncExpr, IntTy); 2894 LeftOverExpr = LeftOverExpr ? 2895 SE.getAddExpr(LeftOverExpr, IncExpr) : IncExpr; 2896 } 2897 if (LeftOverExpr && !LeftOverExpr->isZero()) { 2898 // Expand the IV increment. 2899 Rewriter.clearPostInc(); 2900 Value *IncV = Rewriter.expandCodeFor(LeftOverExpr, IntTy, InsertPt); 2901 const SCEV *IVOperExpr = SE.getAddExpr(SE.getUnknown(IVSrc), 2902 SE.getUnknown(IncV)); 2903 IVOper = Rewriter.expandCodeFor(IVOperExpr, IVTy, InsertPt); 2904 2905 // If an IV increment can't be folded, use it as the next IV value. 2906 if (!canFoldIVIncExpr(LeftOverExpr, Inc.UserInst, Inc.IVOperand, TTI)) { 2907 assert(IVTy == IVOper->getType() && "inconsistent IV increment type"); 2908 IVSrc = IVOper; 2909 LeftOverExpr = nullptr; 2910 } 2911 } 2912 Type *OperTy = Inc.IVOperand->getType(); 2913 if (IVTy != OperTy) { 2914 assert(SE.getTypeSizeInBits(IVTy) >= SE.getTypeSizeInBits(OperTy) && 2915 "cannot extend a chained IV"); 2916 IRBuilder<> Builder(InsertPt); 2917 IVOper = Builder.CreateTruncOrBitCast(IVOper, OperTy, "lsr.chain"); 2918 } 2919 Inc.UserInst->replaceUsesOfWith(Inc.IVOperand, IVOper); 2920 DeadInsts.emplace_back(Inc.IVOperand); 2921 } 2922 // If LSR created a new, wider phi, we may also replace its postinc. We only 2923 // do this if we also found a wide value for the head of the chain. 2924 if (isa<PHINode>(Chain.tailUserInst())) { 2925 for (BasicBlock::iterator I = L->getHeader()->begin(); 2926 PHINode *Phi = dyn_cast<PHINode>(I); ++I) { 2927 if (!isCompatibleIVType(Phi, IVSrc)) 2928 continue; 2929 Instruction *PostIncV = dyn_cast<Instruction>( 2930 Phi->getIncomingValueForBlock(L->getLoopLatch())); 2931 if (!PostIncV || (SE.getSCEV(PostIncV) != SE.getSCEV(IVSrc))) 2932 continue; 2933 Value *IVOper = IVSrc; 2934 Type *PostIncTy = PostIncV->getType(); 2935 if (IVTy != PostIncTy) { 2936 assert(PostIncTy->isPointerTy() && "mixing int/ptr IV types"); 2937 IRBuilder<> Builder(L->getLoopLatch()->getTerminator()); 2938 Builder.SetCurrentDebugLocation(PostIncV->getDebugLoc()); 2939 IVOper = Builder.CreatePointerCast(IVSrc, PostIncTy, "lsr.chain"); 2940 } 2941 Phi->replaceUsesOfWith(PostIncV, IVOper); 2942 DeadInsts.emplace_back(PostIncV); 2943 } 2944 } 2945} 2946 2947void LSRInstance::CollectFixupsAndInitialFormulae() { 2948 for (const IVStrideUse &U : IU) { 2949 Instruction *UserInst = U.getUser(); 2950 // Skip IV users that are part of profitable IV Chains. 2951 User::op_iterator UseI = std::find(UserInst->op_begin(), UserInst->op_end(), 2952 U.getOperandValToReplace()); 2953 assert(UseI != UserInst->op_end() && "cannot find IV operand"); 2954 if (IVIncSet.count(UseI)) 2955 continue; 2956 2957 // Record the uses. 2958 LSRFixup &LF = getNewFixup(); 2959 LF.UserInst = UserInst; 2960 LF.OperandValToReplace = U.getOperandValToReplace(); 2961 LF.PostIncLoops = U.getPostIncLoops(); 2962 2963 LSRUse::KindType Kind = LSRUse::Basic; 2964 Type *AccessTy = nullptr; 2965 if (isAddressUse(LF.UserInst, LF.OperandValToReplace)) { 2966 Kind = LSRUse::Address; 2967 AccessTy = getAccessType(LF.UserInst); 2968 } 2969 2970 const SCEV *S = IU.getExpr(U); 2971 2972 // Equality (== and !=) ICmps are special. We can rewrite (i == N) as 2973 // (N - i == 0), and this allows (N - i) to be the expression that we work 2974 // with rather than just N or i, so we can consider the register 2975 // requirements for both N and i at the same time. Limiting this code to 2976 // equality icmps is not a problem because all interesting loops use 2977 // equality icmps, thanks to IndVarSimplify. 2978 if (ICmpInst *CI = dyn_cast<ICmpInst>(LF.UserInst)) 2979 if (CI->isEquality()) { 2980 // Swap the operands if needed to put the OperandValToReplace on the 2981 // left, for consistency. 2982 Value *NV = CI->getOperand(1); 2983 if (NV == LF.OperandValToReplace) { 2984 CI->setOperand(1, CI->getOperand(0)); 2985 CI->setOperand(0, NV); 2986 NV = CI->getOperand(1); 2987 Changed = true; 2988 } 2989 2990 // x == y --> x - y == 0 2991 const SCEV *N = SE.getSCEV(NV); 2992 if (SE.isLoopInvariant(N, L) && isSafeToExpand(N, SE)) { 2993 // S is normalized, so normalize N before folding it into S 2994 // to keep the result normalized. 2995 N = TransformForPostIncUse(Normalize, N, CI, nullptr, 2996 LF.PostIncLoops, SE, DT); 2997 Kind = LSRUse::ICmpZero; 2998 S = SE.getMinusSCEV(N, S); 2999 } 3000 3001 // -1 and the negations of all interesting strides (except the negation 3002 // of -1) are now also interesting. 3003 for (size_t i = 0, e = Factors.size(); i != e; ++i) 3004 if (Factors[i] != -1) 3005 Factors.insert(-(uint64_t)Factors[i]); 3006 Factors.insert(-1); 3007 } 3008 3009 // Set up the initial formula for this use. 3010 std::pair<size_t, int64_t> P = getUse(S, Kind, AccessTy); 3011 LF.LUIdx = P.first; 3012 LF.Offset = P.second; 3013 LSRUse &LU = Uses[LF.LUIdx]; 3014 LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L); 3015 if (!LU.WidestFixupType || 3016 SE.getTypeSizeInBits(LU.WidestFixupType) < 3017 SE.getTypeSizeInBits(LF.OperandValToReplace->getType())) 3018 LU.WidestFixupType = LF.OperandValToReplace->getType(); 3019 3020 // If this is the first use of this LSRUse, give it a formula. 3021 if (LU.Formulae.empty()) { 3022 InsertInitialFormula(S, LU, LF.LUIdx); 3023 CountRegisters(LU.Formulae.back(), LF.LUIdx); 3024 } 3025 } 3026 3027 DEBUG(print_fixups(dbgs())); 3028} 3029 3030/// InsertInitialFormula - Insert a formula for the given expression into 3031/// the given use, separating out loop-variant portions from loop-invariant 3032/// and loop-computable portions. 3033void 3034LSRInstance::InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx) { 3035 // Mark uses whose expressions cannot be expanded. 3036 if (!isSafeToExpand(S, SE)) 3037 LU.RigidFormula = true; 3038 3039 Formula F; 3040 F.InitialMatch(S, L, SE); 3041 bool Inserted = InsertFormula(LU, LUIdx, F); 3042 assert(Inserted && "Initial formula already exists!"); (void)Inserted; 3043} 3044 3045/// InsertSupplementalFormula - Insert a simple single-register formula for 3046/// the given expression into the given use. 3047void 3048LSRInstance::InsertSupplementalFormula(const SCEV *S, 3049 LSRUse &LU, size_t LUIdx) { 3050 Formula F; 3051 F.BaseRegs.push_back(S); 3052 F.HasBaseReg = true; 3053 bool Inserted = InsertFormula(LU, LUIdx, F); 3054 assert(Inserted && "Supplemental formula already exists!"); (void)Inserted; 3055} 3056 3057/// CountRegisters - Note which registers are used by the given formula, 3058/// updating RegUses. 3059void LSRInstance::CountRegisters(const Formula &F, size_t LUIdx) { 3060 if (F.ScaledReg) 3061 RegUses.CountRegister(F.ScaledReg, LUIdx); 3062 for (const SCEV *BaseReg : F.BaseRegs) 3063 RegUses.CountRegister(BaseReg, LUIdx); 3064} 3065 3066/// InsertFormula - If the given formula has not yet been inserted, add it to 3067/// the list, and return true. Return false otherwise. 3068bool LSRInstance::InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F) { 3069 // Do not insert formula that we will not be able to expand. 3070 assert(isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy, F) && 3071 "Formula is illegal"); 3072 if (!LU.InsertFormula(F)) 3073 return false; 3074 3075 CountRegisters(F, LUIdx); 3076 return true; 3077} 3078 3079/// CollectLoopInvariantFixupsAndFormulae - Check for other uses of 3080/// loop-invariant values which we're tracking. These other uses will pin these 3081/// values in registers, making them less profitable for elimination. 3082/// TODO: This currently misses non-constant addrec step registers. 3083/// TODO: Should this give more weight to users inside the loop? 3084void 3085LSRInstance::CollectLoopInvariantFixupsAndFormulae() { 3086 SmallVector<const SCEV *, 8> Worklist(RegUses.begin(), RegUses.end()); 3087 SmallPtrSet<const SCEV *, 32> Visited; 3088 3089 while (!Worklist.empty()) { 3090 const SCEV *S = Worklist.pop_back_val(); 3091 3092 // Don't process the same SCEV twice 3093 if (!Visited.insert(S).second) 3094 continue; 3095 3096 if (const SCEVNAryExpr *N = dyn_cast<SCEVNAryExpr>(S)) 3097 Worklist.append(N->op_begin(), N->op_end()); 3098 else if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(S)) 3099 Worklist.push_back(C->getOperand()); 3100 else if (const SCEVUDivExpr *D = dyn_cast<SCEVUDivExpr>(S)) { 3101 Worklist.push_back(D->getLHS()); 3102 Worklist.push_back(D->getRHS()); 3103 } else if (const SCEVUnknown *US = dyn_cast<SCEVUnknown>(S)) { 3104 const Value *V = US->getValue(); 3105 if (const Instruction *Inst = dyn_cast<Instruction>(V)) { 3106 // Look for instructions defined outside the loop. 3107 if (L->contains(Inst)) continue; 3108 } else if (isa<UndefValue>(V)) 3109 // Undef doesn't have a live range, so it doesn't matter. 3110 continue; 3111 for (const Use &U : V->uses()) { 3112 const Instruction *UserInst = dyn_cast<Instruction>(U.getUser()); 3113 // Ignore non-instructions. 3114 if (!UserInst) 3115 continue; 3116 // Ignore instructions in other functions (as can happen with 3117 // Constants). 3118 if (UserInst->getParent()->getParent() != L->getHeader()->getParent()) 3119 continue; 3120 // Ignore instructions not dominated by the loop. 3121 const BasicBlock *UseBB = !isa<PHINode>(UserInst) ? 3122 UserInst->getParent() : 3123 cast<PHINode>(UserInst)->getIncomingBlock( 3124 PHINode::getIncomingValueNumForOperand(U.getOperandNo())); 3125 if (!DT.dominates(L->getHeader(), UseBB)) 3126 continue; 3127 // Ignore uses which are part of other SCEV expressions, to avoid 3128 // analyzing them multiple times. 3129 if (SE.isSCEVable(UserInst->getType())) { 3130 const SCEV *UserS = SE.getSCEV(const_cast<Instruction *>(UserInst)); 3131 // If the user is a no-op, look through to its uses. 3132 if (!isa<SCEVUnknown>(UserS)) 3133 continue; 3134 if (UserS == US) { 3135 Worklist.push_back( 3136 SE.getUnknown(const_cast<Instruction *>(UserInst))); 3137 continue; 3138 } 3139 } 3140 // Ignore icmp instructions which are already being analyzed. 3141 if (const ICmpInst *ICI = dyn_cast<ICmpInst>(UserInst)) { 3142 unsigned OtherIdx = !U.getOperandNo(); 3143 Value *OtherOp = const_cast<Value *>(ICI->getOperand(OtherIdx)); 3144 if (SE.hasComputableLoopEvolution(SE.getSCEV(OtherOp), L)) 3145 continue; 3146 } 3147 3148 LSRFixup &LF = getNewFixup(); 3149 LF.UserInst = const_cast<Instruction *>(UserInst); 3150 LF.OperandValToReplace = U; 3151 std::pair<size_t, int64_t> P = getUse(S, LSRUse::Basic, nullptr); 3152 LF.LUIdx = P.first; 3153 LF.Offset = P.second; 3154 LSRUse &LU = Uses[LF.LUIdx]; 3155 LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L); 3156 if (!LU.WidestFixupType || 3157 SE.getTypeSizeInBits(LU.WidestFixupType) < 3158 SE.getTypeSizeInBits(LF.OperandValToReplace->getType())) 3159 LU.WidestFixupType = LF.OperandValToReplace->getType(); 3160 InsertSupplementalFormula(US, LU, LF.LUIdx); 3161 CountRegisters(LU.Formulae.back(), Uses.size() - 1); 3162 break; 3163 } 3164 } 3165 } 3166} 3167 3168/// CollectSubexprs - Split S into subexpressions which can be pulled out into 3169/// separate registers. If C is non-null, multiply each subexpression by C. 3170/// 3171/// Return remainder expression after factoring the subexpressions captured by 3172/// Ops. If Ops is complete, return NULL. 3173static const SCEV *CollectSubexprs(const SCEV *S, const SCEVConstant *C, 3174 SmallVectorImpl<const SCEV *> &Ops, 3175 const Loop *L, 3176 ScalarEvolution &SE, 3177 unsigned Depth = 0) { 3178 // Arbitrarily cap recursion to protect compile time. 3179 if (Depth >= 3) 3180 return S; 3181 3182 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) { 3183 // Break out add operands. 3184 for (const SCEV *S : Add->operands()) { 3185 const SCEV *Remainder = CollectSubexprs(S, C, Ops, L, SE, Depth+1); 3186 if (Remainder) 3187 Ops.push_back(C ? SE.getMulExpr(C, Remainder) : Remainder); 3188 } 3189 return nullptr; 3190 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) { 3191 // Split a non-zero base out of an addrec. 3192 if (AR->getStart()->isZero()) 3193 return S; 3194 3195 const SCEV *Remainder = CollectSubexprs(AR->getStart(), 3196 C, Ops, L, SE, Depth+1); 3197 // Split the non-zero AddRec unless it is part of a nested recurrence that 3198 // does not pertain to this loop. 3199 if (Remainder && (AR->getLoop() == L || !isa<SCEVAddRecExpr>(Remainder))) { 3200 Ops.push_back(C ? SE.getMulExpr(C, Remainder) : Remainder); 3201 Remainder = nullptr; 3202 } 3203 if (Remainder != AR->getStart()) { 3204 if (!Remainder) 3205 Remainder = SE.getConstant(AR->getType(), 0); 3206 return SE.getAddRecExpr(Remainder, 3207 AR->getStepRecurrence(SE), 3208 AR->getLoop(), 3209 //FIXME: AR->getNoWrapFlags(SCEV::FlagNW) 3210 SCEV::FlagAnyWrap); 3211 } 3212 } else if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) { 3213 // Break (C * (a + b + c)) into C*a + C*b + C*c. 3214 if (Mul->getNumOperands() != 2) 3215 return S; 3216 if (const SCEVConstant *Op0 = 3217 dyn_cast<SCEVConstant>(Mul->getOperand(0))) { 3218 C = C ? cast<SCEVConstant>(SE.getMulExpr(C, Op0)) : Op0; 3219 const SCEV *Remainder = 3220 CollectSubexprs(Mul->getOperand(1), C, Ops, L, SE, Depth+1); 3221 if (Remainder) 3222 Ops.push_back(SE.getMulExpr(C, Remainder)); 3223 return nullptr; 3224 } 3225 } 3226 return S; 3227} 3228 3229/// \brief Helper function for LSRInstance::GenerateReassociations. 3230void LSRInstance::GenerateReassociationsImpl(LSRUse &LU, unsigned LUIdx, 3231 const Formula &Base, 3232 unsigned Depth, size_t Idx, 3233 bool IsScaledReg) { 3234 const SCEV *BaseReg = IsScaledReg ? Base.ScaledReg : Base.BaseRegs[Idx]; 3235 SmallVector<const SCEV *, 8> AddOps; 3236 const SCEV *Remainder = CollectSubexprs(BaseReg, nullptr, AddOps, L, SE); 3237 if (Remainder) 3238 AddOps.push_back(Remainder); 3239 3240 if (AddOps.size() == 1) 3241 return; 3242 3243 for (SmallVectorImpl<const SCEV *>::const_iterator J = AddOps.begin(), 3244 JE = AddOps.end(); 3245 J != JE; ++J) { 3246 3247 // Loop-variant "unknown" values are uninteresting; we won't be able to 3248 // do anything meaningful with them. 3249 if (isa<SCEVUnknown>(*J) && !SE.isLoopInvariant(*J, L)) 3250 continue; 3251 3252 // Don't pull a constant into a register if the constant could be folded 3253 // into an immediate field. 3254 if (isAlwaysFoldable(TTI, SE, LU.MinOffset, LU.MaxOffset, LU.Kind, 3255 LU.AccessTy, *J, Base.getNumRegs() > 1)) 3256 continue; 3257 3258 // Collect all operands except *J. 3259 SmallVector<const SCEV *, 8> InnerAddOps( 3260 ((const SmallVector<const SCEV *, 8> &)AddOps).begin(), J); 3261 InnerAddOps.append(std::next(J), 3262 ((const SmallVector<const SCEV *, 8> &)AddOps).end()); 3263 3264 // Don't leave just a constant behind in a register if the constant could 3265 // be folded into an immediate field. 3266 if (InnerAddOps.size() == 1 && 3267 isAlwaysFoldable(TTI, SE, LU.MinOffset, LU.MaxOffset, LU.Kind, 3268 LU.AccessTy, InnerAddOps[0], Base.getNumRegs() > 1)) 3269 continue; 3270 3271 const SCEV *InnerSum = SE.getAddExpr(InnerAddOps); 3272 if (InnerSum->isZero()) 3273 continue; 3274 Formula F = Base; 3275 3276 // Add the remaining pieces of the add back into the new formula. 3277 const SCEVConstant *InnerSumSC = dyn_cast<SCEVConstant>(InnerSum); 3278 if (InnerSumSC && SE.getTypeSizeInBits(InnerSumSC->getType()) <= 64 && 3279 TTI.isLegalAddImmediate((uint64_t)F.UnfoldedOffset + 3280 InnerSumSC->getValue()->getZExtValue())) { 3281 F.UnfoldedOffset = 3282 (uint64_t)F.UnfoldedOffset + InnerSumSC->getValue()->getZExtValue(); 3283 if (IsScaledReg) 3284 F.ScaledReg = nullptr; 3285 else 3286 F.BaseRegs.erase(F.BaseRegs.begin() + Idx); 3287 } else if (IsScaledReg) 3288 F.ScaledReg = InnerSum; 3289 else 3290 F.BaseRegs[Idx] = InnerSum; 3291 3292 // Add J as its own register, or an unfolded immediate. 3293 const SCEVConstant *SC = dyn_cast<SCEVConstant>(*J); 3294 if (SC && SE.getTypeSizeInBits(SC->getType()) <= 64 && 3295 TTI.isLegalAddImmediate((uint64_t)F.UnfoldedOffset + 3296 SC->getValue()->getZExtValue())) 3297 F.UnfoldedOffset = 3298 (uint64_t)F.UnfoldedOffset + SC->getValue()->getZExtValue(); 3299 else 3300 F.BaseRegs.push_back(*J); 3301 // We may have changed the number of register in base regs, adjust the 3302 // formula accordingly. 3303 F.Canonicalize(); 3304 3305 if (InsertFormula(LU, LUIdx, F)) 3306 // If that formula hadn't been seen before, recurse to find more like 3307 // it. 3308 GenerateReassociations(LU, LUIdx, LU.Formulae.back(), Depth + 1); 3309 } 3310} 3311 3312/// GenerateReassociations - Split out subexpressions from adds and the bases of 3313/// addrecs. 3314void LSRInstance::GenerateReassociations(LSRUse &LU, unsigned LUIdx, 3315 Formula Base, unsigned Depth) { 3316 assert(Base.isCanonical() && "Input must be in the canonical form"); 3317 // Arbitrarily cap recursion to protect compile time. 3318 if (Depth >= 3) 3319 return; 3320 3321 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) 3322 GenerateReassociationsImpl(LU, LUIdx, Base, Depth, i); 3323 3324 if (Base.Scale == 1) 3325 GenerateReassociationsImpl(LU, LUIdx, Base, Depth, 3326 /* Idx */ -1, /* IsScaledReg */ true); 3327} 3328 3329/// GenerateCombinations - Generate a formula consisting of all of the 3330/// loop-dominating registers added into a single register. 3331void LSRInstance::GenerateCombinations(LSRUse &LU, unsigned LUIdx, 3332 Formula Base) { 3333 // This method is only interesting on a plurality of registers. 3334 if (Base.BaseRegs.size() + (Base.Scale == 1) <= 1) 3335 return; 3336 3337 // Flatten the representation, i.e., reg1 + 1*reg2 => reg1 + reg2, before 3338 // processing the formula. 3339 Base.Unscale(); 3340 Formula F = Base; 3341 F.BaseRegs.clear(); 3342 SmallVector<const SCEV *, 4> Ops; 3343 for (const SCEV *BaseReg : Base.BaseRegs) { 3344 if (SE.properlyDominates(BaseReg, L->getHeader()) && 3345 !SE.hasComputableLoopEvolution(BaseReg, L)) 3346 Ops.push_back(BaseReg); 3347 else 3348 F.BaseRegs.push_back(BaseReg); 3349 } 3350 if (Ops.size() > 1) { 3351 const SCEV *Sum = SE.getAddExpr(Ops); 3352 // TODO: If Sum is zero, it probably means ScalarEvolution missed an 3353 // opportunity to fold something. For now, just ignore such cases 3354 // rather than proceed with zero in a register. 3355 if (!Sum->isZero()) { 3356 F.BaseRegs.push_back(Sum); 3357 F.Canonicalize(); 3358 (void)InsertFormula(LU, LUIdx, F); 3359 } 3360 } 3361} 3362 3363/// \brief Helper function for LSRInstance::GenerateSymbolicOffsets. 3364void LSRInstance::GenerateSymbolicOffsetsImpl(LSRUse &LU, unsigned LUIdx, 3365 const Formula &Base, size_t Idx, 3366 bool IsScaledReg) { 3367 const SCEV *G = IsScaledReg ? Base.ScaledReg : Base.BaseRegs[Idx]; 3368 GlobalValue *GV = ExtractSymbol(G, SE); 3369 if (G->isZero() || !GV) 3370 return; 3371 Formula F = Base; 3372 F.BaseGV = GV; 3373 if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy, F)) 3374 return; 3375 if (IsScaledReg) 3376 F.ScaledReg = G; 3377 else 3378 F.BaseRegs[Idx] = G; 3379 (void)InsertFormula(LU, LUIdx, F); 3380} 3381 3382/// GenerateSymbolicOffsets - Generate reuse formulae using symbolic offsets. 3383void LSRInstance::GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx, 3384 Formula Base) { 3385 // We can't add a symbolic offset if the address already contains one. 3386 if (Base.BaseGV) return; 3387 3388 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) 3389 GenerateSymbolicOffsetsImpl(LU, LUIdx, Base, i); 3390 if (Base.Scale == 1) 3391 GenerateSymbolicOffsetsImpl(LU, LUIdx, Base, /* Idx */ -1, 3392 /* IsScaledReg */ true); 3393} 3394 3395/// \brief Helper function for LSRInstance::GenerateConstantOffsets. 3396void LSRInstance::GenerateConstantOffsetsImpl( 3397 LSRUse &LU, unsigned LUIdx, const Formula &Base, 3398 const SmallVectorImpl<int64_t> &Worklist, size_t Idx, bool IsScaledReg) { 3399 const SCEV *G = IsScaledReg ? Base.ScaledReg : Base.BaseRegs[Idx]; 3400 for (int64_t Offset : Worklist) { 3401 Formula F = Base; 3402 F.BaseOffset = (uint64_t)Base.BaseOffset - Offset; 3403 if (isLegalUse(TTI, LU.MinOffset - Offset, LU.MaxOffset - Offset, LU.Kind, 3404 LU.AccessTy, F)) { 3405 // Add the offset to the base register. 3406 const SCEV *NewG = SE.getAddExpr(SE.getConstant(G->getType(), Offset), G); 3407 // If it cancelled out, drop the base register, otherwise update it. 3408 if (NewG->isZero()) { 3409 if (IsScaledReg) { 3410 F.Scale = 0; 3411 F.ScaledReg = nullptr; 3412 } else 3413 F.DeleteBaseReg(F.BaseRegs[Idx]); 3414 F.Canonicalize(); 3415 } else if (IsScaledReg) 3416 F.ScaledReg = NewG; 3417 else 3418 F.BaseRegs[Idx] = NewG; 3419 3420 (void)InsertFormula(LU, LUIdx, F); 3421 } 3422 } 3423 3424 int64_t Imm = ExtractImmediate(G, SE); 3425 if (G->isZero() || Imm == 0) 3426 return; 3427 Formula F = Base; 3428 F.BaseOffset = (uint64_t)F.BaseOffset + Imm; 3429 if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy, F)) 3430 return; 3431 if (IsScaledReg) 3432 F.ScaledReg = G; 3433 else 3434 F.BaseRegs[Idx] = G; 3435 (void)InsertFormula(LU, LUIdx, F); 3436} 3437 3438/// GenerateConstantOffsets - Generate reuse formulae using symbolic offsets. 3439void LSRInstance::GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx, 3440 Formula Base) { 3441 // TODO: For now, just add the min and max offset, because it usually isn't 3442 // worthwhile looking at everything inbetween. 3443 SmallVector<int64_t, 2> Worklist; 3444 Worklist.push_back(LU.MinOffset); 3445 if (LU.MaxOffset != LU.MinOffset) 3446 Worklist.push_back(LU.MaxOffset); 3447 3448 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) 3449 GenerateConstantOffsetsImpl(LU, LUIdx, Base, Worklist, i); 3450 if (Base.Scale == 1) 3451 GenerateConstantOffsetsImpl(LU, LUIdx, Base, Worklist, /* Idx */ -1, 3452 /* IsScaledReg */ true); 3453} 3454 3455/// GenerateICmpZeroScales - For ICmpZero, check to see if we can scale up 3456/// the comparison. For example, x == y -> x*c == y*c. 3457void LSRInstance::GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx, 3458 Formula Base) { 3459 if (LU.Kind != LSRUse::ICmpZero) return; 3460 3461 // Determine the integer type for the base formula. 3462 Type *IntTy = Base.getType(); 3463 if (!IntTy) return; 3464 if (SE.getTypeSizeInBits(IntTy) > 64) return; 3465 3466 // Don't do this if there is more than one offset. 3467 if (LU.MinOffset != LU.MaxOffset) return; 3468 3469 assert(!Base.BaseGV && "ICmpZero use is not legal!"); 3470 3471 // Check each interesting stride. 3472 for (int64_t Factor : Factors) { 3473 // Check that the multiplication doesn't overflow. 3474 if (Base.BaseOffset == INT64_MIN && Factor == -1) 3475 continue; 3476 int64_t NewBaseOffset = (uint64_t)Base.BaseOffset * Factor; 3477 if (NewBaseOffset / Factor != Base.BaseOffset) 3478 continue; 3479 // If the offset will be truncated at this use, check that it is in bounds. 3480 if (!IntTy->isPointerTy() && 3481 !ConstantInt::isValueValidForType(IntTy, NewBaseOffset)) 3482 continue; 3483 3484 // Check that multiplying with the use offset doesn't overflow. 3485 int64_t Offset = LU.MinOffset; 3486 if (Offset == INT64_MIN && Factor == -1) 3487 continue; 3488 Offset = (uint64_t)Offset * Factor; 3489 if (Offset / Factor != LU.MinOffset) 3490 continue; 3491 // If the offset will be truncated at this use, check that it is in bounds. 3492 if (!IntTy->isPointerTy() && 3493 !ConstantInt::isValueValidForType(IntTy, Offset)) 3494 continue; 3495 3496 Formula F = Base; 3497 F.BaseOffset = NewBaseOffset; 3498 3499 // Check that this scale is legal. 3500 if (!isLegalUse(TTI, Offset, Offset, LU.Kind, LU.AccessTy, F)) 3501 continue; 3502 3503 // Compensate for the use having MinOffset built into it. 3504 F.BaseOffset = (uint64_t)F.BaseOffset + Offset - LU.MinOffset; 3505 3506 const SCEV *FactorS = SE.getConstant(IntTy, Factor); 3507 3508 // Check that multiplying with each base register doesn't overflow. 3509 for (size_t i = 0, e = F.BaseRegs.size(); i != e; ++i) { 3510 F.BaseRegs[i] = SE.getMulExpr(F.BaseRegs[i], FactorS); 3511 if (getExactSDiv(F.BaseRegs[i], FactorS, SE) != Base.BaseRegs[i]) 3512 goto next; 3513 } 3514 3515 // Check that multiplying with the scaled register doesn't overflow. 3516 if (F.ScaledReg) { 3517 F.ScaledReg = SE.getMulExpr(F.ScaledReg, FactorS); 3518 if (getExactSDiv(F.ScaledReg, FactorS, SE) != Base.ScaledReg) 3519 continue; 3520 } 3521 3522 // Check that multiplying with the unfolded offset doesn't overflow. 3523 if (F.UnfoldedOffset != 0) { 3524 if (F.UnfoldedOffset == INT64_MIN && Factor == -1) 3525 continue; 3526 F.UnfoldedOffset = (uint64_t)F.UnfoldedOffset * Factor; 3527 if (F.UnfoldedOffset / Factor != Base.UnfoldedOffset) 3528 continue; 3529 // If the offset will be truncated, check that it is in bounds. 3530 if (!IntTy->isPointerTy() && 3531 !ConstantInt::isValueValidForType(IntTy, F.UnfoldedOffset)) 3532 continue; 3533 } 3534 3535 // If we make it here and it's legal, add it. 3536 (void)InsertFormula(LU, LUIdx, F); 3537 next:; 3538 } 3539} 3540 3541/// GenerateScales - Generate stride factor reuse formulae by making use of 3542/// scaled-offset address modes, for example. 3543void LSRInstance::GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base) { 3544 // Determine the integer type for the base formula. 3545 Type *IntTy = Base.getType(); 3546 if (!IntTy) return; 3547 3548 // If this Formula already has a scaled register, we can't add another one. 3549 // Try to unscale the formula to generate a better scale. 3550 if (Base.Scale != 0 && !Base.Unscale()) 3551 return; 3552 3553 assert(Base.Scale == 0 && "Unscale did not did its job!"); 3554 3555 // Check each interesting stride. 3556 for (int64_t Factor : Factors) { 3557 Base.Scale = Factor; 3558 Base.HasBaseReg = Base.BaseRegs.size() > 1; 3559 // Check whether this scale is going to be legal. 3560 if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy, 3561 Base)) { 3562 // As a special-case, handle special out-of-loop Basic users specially. 3563 // TODO: Reconsider this special case. 3564 if (LU.Kind == LSRUse::Basic && 3565 isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LSRUse::Special, 3566 LU.AccessTy, Base) && 3567 LU.AllFixupsOutsideLoop) 3568 LU.Kind = LSRUse::Special; 3569 else 3570 continue; 3571 } 3572 // For an ICmpZero, negating a solitary base register won't lead to 3573 // new solutions. 3574 if (LU.Kind == LSRUse::ICmpZero && 3575 !Base.HasBaseReg && Base.BaseOffset == 0 && !Base.BaseGV) 3576 continue; 3577 // For each addrec base reg, apply the scale, if possible. 3578 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) 3579 if (const SCEVAddRecExpr *AR = 3580 dyn_cast<SCEVAddRecExpr>(Base.BaseRegs[i])) { 3581 const SCEV *FactorS = SE.getConstant(IntTy, Factor); 3582 if (FactorS->isZero()) 3583 continue; 3584 // Divide out the factor, ignoring high bits, since we'll be 3585 // scaling the value back up in the end. 3586 if (const SCEV *Quotient = getExactSDiv(AR, FactorS, SE, true)) { 3587 // TODO: This could be optimized to avoid all the copying. 3588 Formula F = Base; 3589 F.ScaledReg = Quotient; 3590 F.DeleteBaseReg(F.BaseRegs[i]); 3591 // The canonical representation of 1*reg is reg, which is already in 3592 // Base. In that case, do not try to insert the formula, it will be 3593 // rejected anyway. 3594 if (F.Scale == 1 && F.BaseRegs.empty()) 3595 continue; 3596 (void)InsertFormula(LU, LUIdx, F); 3597 } 3598 } 3599 } 3600} 3601 3602/// GenerateTruncates - Generate reuse formulae from different IV types. 3603void LSRInstance::GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base) { 3604 // Don't bother truncating symbolic values. 3605 if (Base.BaseGV) return; 3606 3607 // Determine the integer type for the base formula. 3608 Type *DstTy = Base.getType(); 3609 if (!DstTy) return; 3610 DstTy = SE.getEffectiveSCEVType(DstTy); 3611 3612 for (Type *SrcTy : Types) { 3613 if (SrcTy != DstTy && TTI.isTruncateFree(SrcTy, DstTy)) { 3614 Formula F = Base; 3615 3616 if (F.ScaledReg) F.ScaledReg = SE.getAnyExtendExpr(F.ScaledReg, SrcTy); 3617 for (const SCEV *&BaseReg : F.BaseRegs) 3618 BaseReg = SE.getAnyExtendExpr(BaseReg, SrcTy); 3619 3620 // TODO: This assumes we've done basic processing on all uses and 3621 // have an idea what the register usage is. 3622 if (!F.hasRegsUsedByUsesOtherThan(LUIdx, RegUses)) 3623 continue; 3624 3625 (void)InsertFormula(LU, LUIdx, F); 3626 } 3627 } 3628} 3629 3630namespace { 3631 3632/// WorkItem - Helper class for GenerateCrossUseConstantOffsets. It's used to 3633/// defer modifications so that the search phase doesn't have to worry about 3634/// the data structures moving underneath it. 3635struct WorkItem { 3636 size_t LUIdx; 3637 int64_t Imm; 3638 const SCEV *OrigReg; 3639 3640 WorkItem(size_t LI, int64_t I, const SCEV *R) 3641 : LUIdx(LI), Imm(I), OrigReg(R) {} 3642 3643 void print(raw_ostream &OS) const; 3644 void dump() const; 3645}; 3646 3647} 3648 3649void WorkItem::print(raw_ostream &OS) const { 3650 OS << "in formulae referencing " << *OrigReg << " in use " << LUIdx 3651 << " , add offset " << Imm; 3652} 3653 3654#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) 3655void WorkItem::dump() const { 3656 print(errs()); errs() << '\n'; 3657} 3658#endif 3659 3660/// GenerateCrossUseConstantOffsets - Look for registers which are a constant 3661/// distance apart and try to form reuse opportunities between them. 3662void LSRInstance::GenerateCrossUseConstantOffsets() { 3663 // Group the registers by their value without any added constant offset. 3664 typedef std::map<int64_t, const SCEV *> ImmMapTy; 3665 DenseMap<const SCEV *, ImmMapTy> Map; 3666 DenseMap<const SCEV *, SmallBitVector> UsedByIndicesMap; 3667 SmallVector<const SCEV *, 8> Sequence; 3668 for (const SCEV *Use : RegUses) { 3669 const SCEV *Reg = Use; // Make a copy for ExtractImmediate to modify. 3670 int64_t Imm = ExtractImmediate(Reg, SE); 3671 auto Pair = Map.insert(std::make_pair(Reg, ImmMapTy())); 3672 if (Pair.second) 3673 Sequence.push_back(Reg); 3674 Pair.first->second.insert(std::make_pair(Imm, Use)); 3675 UsedByIndicesMap[Reg] |= RegUses.getUsedByIndices(Use); 3676 } 3677 3678 // Now examine each set of registers with the same base value. Build up 3679 // a list of work to do and do the work in a separate step so that we're 3680 // not adding formulae and register counts while we're searching. 3681 SmallVector<WorkItem, 32> WorkItems; 3682 SmallSet<std::pair<size_t, int64_t>, 32> UniqueItems; 3683 for (const SCEV *Reg : Sequence) { 3684 const ImmMapTy &Imms = Map.find(Reg)->second; 3685 3686 // It's not worthwhile looking for reuse if there's only one offset. 3687 if (Imms.size() == 1) 3688 continue; 3689 3690 DEBUG(dbgs() << "Generating cross-use offsets for " << *Reg << ':'; 3691 for (const auto &Entry : Imms) 3692 dbgs() << ' ' << Entry.first; 3693 dbgs() << '\n'); 3694 3695 // Examine each offset. 3696 for (ImmMapTy::const_iterator J = Imms.begin(), JE = Imms.end(); 3697 J != JE; ++J) { 3698 const SCEV *OrigReg = J->second; 3699 3700 int64_t JImm = J->first; 3701 const SmallBitVector &UsedByIndices = RegUses.getUsedByIndices(OrigReg); 3702 3703 if (!isa<SCEVConstant>(OrigReg) && 3704 UsedByIndicesMap[Reg].count() == 1) { 3705 DEBUG(dbgs() << "Skipping cross-use reuse for " << *OrigReg << '\n'); 3706 continue; 3707 } 3708 3709 // Conservatively examine offsets between this orig reg a few selected 3710 // other orig regs. 3711 ImmMapTy::const_iterator OtherImms[] = { 3712 Imms.begin(), std::prev(Imms.end()), 3713 Imms.lower_bound((Imms.begin()->first + std::prev(Imms.end())->first) / 3714 2) 3715 }; 3716 for (size_t i = 0, e = array_lengthof(OtherImms); i != e; ++i) { 3717 ImmMapTy::const_iterator M = OtherImms[i]; 3718 if (M == J || M == JE) continue; 3719 3720 // Compute the difference between the two. 3721 int64_t Imm = (uint64_t)JImm - M->first; 3722 for (int LUIdx = UsedByIndices.find_first(); LUIdx != -1; 3723 LUIdx = UsedByIndices.find_next(LUIdx)) 3724 // Make a memo of this use, offset, and register tuple. 3725 if (UniqueItems.insert(std::make_pair(LUIdx, Imm)).second) 3726 WorkItems.push_back(WorkItem(LUIdx, Imm, OrigReg)); 3727 } 3728 } 3729 } 3730 3731 Map.clear(); 3732 Sequence.clear(); 3733 UsedByIndicesMap.clear(); 3734 UniqueItems.clear(); 3735 3736 // Now iterate through the worklist and add new formulae. 3737 for (const WorkItem &WI : WorkItems) { 3738 size_t LUIdx = WI.LUIdx; 3739 LSRUse &LU = Uses[LUIdx]; 3740 int64_t Imm = WI.Imm; 3741 const SCEV *OrigReg = WI.OrigReg; 3742 3743 Type *IntTy = SE.getEffectiveSCEVType(OrigReg->getType()); 3744 const SCEV *NegImmS = SE.getSCEV(ConstantInt::get(IntTy, -(uint64_t)Imm)); 3745 unsigned BitWidth = SE.getTypeSizeInBits(IntTy); 3746 3747 // TODO: Use a more targeted data structure. 3748 for (size_t L = 0, LE = LU.Formulae.size(); L != LE; ++L) { 3749 Formula F = LU.Formulae[L]; 3750 // FIXME: The code for the scaled and unscaled registers looks 3751 // very similar but slightly different. Investigate if they 3752 // could be merged. That way, we would not have to unscale the 3753 // Formula. 3754 F.Unscale(); 3755 // Use the immediate in the scaled register. 3756 if (F.ScaledReg == OrigReg) { 3757 int64_t Offset = (uint64_t)F.BaseOffset + Imm * (uint64_t)F.Scale; 3758 // Don't create 50 + reg(-50). 3759 if (F.referencesReg(SE.getSCEV( 3760 ConstantInt::get(IntTy, -(uint64_t)Offset)))) 3761 continue; 3762 Formula NewF = F; 3763 NewF.BaseOffset = Offset; 3764 if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy, 3765 NewF)) 3766 continue; 3767 NewF.ScaledReg = SE.getAddExpr(NegImmS, NewF.ScaledReg); 3768 3769 // If the new scale is a constant in a register, and adding the constant 3770 // value to the immediate would produce a value closer to zero than the 3771 // immediate itself, then the formula isn't worthwhile. 3772 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(NewF.ScaledReg)) 3773 if (C->getValue()->isNegative() != 3774 (NewF.BaseOffset < 0) && 3775 (C->getValue()->getValue().abs() * APInt(BitWidth, F.Scale)) 3776 .ule(std::abs(NewF.BaseOffset))) 3777 continue; 3778 3779 // OK, looks good. 3780 NewF.Canonicalize(); 3781 (void)InsertFormula(LU, LUIdx, NewF); 3782 } else { 3783 // Use the immediate in a base register. 3784 for (size_t N = 0, NE = F.BaseRegs.size(); N != NE; ++N) { 3785 const SCEV *BaseReg = F.BaseRegs[N]; 3786 if (BaseReg != OrigReg) 3787 continue; 3788 Formula NewF = F; 3789 NewF.BaseOffset = (uint64_t)NewF.BaseOffset + Imm; 3790 if (!isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, 3791 LU.Kind, LU.AccessTy, NewF)) { 3792 if (!TTI.isLegalAddImmediate((uint64_t)NewF.UnfoldedOffset + Imm)) 3793 continue; 3794 NewF = F; 3795 NewF.UnfoldedOffset = (uint64_t)NewF.UnfoldedOffset + Imm; 3796 } 3797 NewF.BaseRegs[N] = SE.getAddExpr(NegImmS, BaseReg); 3798 3799 // If the new formula has a constant in a register, and adding the 3800 // constant value to the immediate would produce a value closer to 3801 // zero than the immediate itself, then the formula isn't worthwhile. 3802 for (const SCEV *NewReg : NewF.BaseRegs) 3803 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(NewReg)) 3804 if ((C->getValue()->getValue() + NewF.BaseOffset).abs().slt( 3805 std::abs(NewF.BaseOffset)) && 3806 (C->getValue()->getValue() + 3807 NewF.BaseOffset).countTrailingZeros() >= 3808 countTrailingZeros<uint64_t>(NewF.BaseOffset)) 3809 goto skip_formula; 3810 3811 // Ok, looks good. 3812 NewF.Canonicalize(); 3813 (void)InsertFormula(LU, LUIdx, NewF); 3814 break; 3815 skip_formula:; 3816 } 3817 } 3818 } 3819 } 3820} 3821 3822/// GenerateAllReuseFormulae - Generate formulae for each use. 3823void 3824LSRInstance::GenerateAllReuseFormulae() { 3825 // This is split into multiple loops so that hasRegsUsedByUsesOtherThan 3826 // queries are more precise. 3827 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) { 3828 LSRUse &LU = Uses[LUIdx]; 3829 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i) 3830 GenerateReassociations(LU, LUIdx, LU.Formulae[i]); 3831 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i) 3832 GenerateCombinations(LU, LUIdx, LU.Formulae[i]); 3833 } 3834 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) { 3835 LSRUse &LU = Uses[LUIdx]; 3836 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i) 3837 GenerateSymbolicOffsets(LU, LUIdx, LU.Formulae[i]); 3838 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i) 3839 GenerateConstantOffsets(LU, LUIdx, LU.Formulae[i]); 3840 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i) 3841 GenerateICmpZeroScales(LU, LUIdx, LU.Formulae[i]); 3842 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i) 3843 GenerateScales(LU, LUIdx, LU.Formulae[i]); 3844 } 3845 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) { 3846 LSRUse &LU = Uses[LUIdx]; 3847 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i) 3848 GenerateTruncates(LU, LUIdx, LU.Formulae[i]); 3849 } 3850 3851 GenerateCrossUseConstantOffsets(); 3852 3853 DEBUG(dbgs() << "\n" 3854 "After generating reuse formulae:\n"; 3855 print_uses(dbgs())); 3856} 3857 3858/// If there are multiple formulae with the same set of registers used 3859/// by other uses, pick the best one and delete the others. 3860void LSRInstance::FilterOutUndesirableDedicatedRegisters() { 3861 DenseSet<const SCEV *> VisitedRegs; 3862 SmallPtrSet<const SCEV *, 16> Regs; 3863 SmallPtrSet<const SCEV *, 16> LoserRegs; 3864#ifndef NDEBUG 3865 bool ChangedFormulae = false; 3866#endif 3867 3868 // Collect the best formula for each unique set of shared registers. This 3869 // is reset for each use. 3870 typedef DenseMap<SmallVector<const SCEV *, 4>, size_t, UniquifierDenseMapInfo> 3871 BestFormulaeTy; 3872 BestFormulaeTy BestFormulae; 3873 3874 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) { 3875 LSRUse &LU = Uses[LUIdx]; 3876 DEBUG(dbgs() << "Filtering for use "; LU.print(dbgs()); dbgs() << '\n'); 3877 3878 bool Any = false; 3879 for (size_t FIdx = 0, NumForms = LU.Formulae.size(); 3880 FIdx != NumForms; ++FIdx) { 3881 Formula &F = LU.Formulae[FIdx]; 3882 3883 // Some formulas are instant losers. For example, they may depend on 3884 // nonexistent AddRecs from other loops. These need to be filtered 3885 // immediately, otherwise heuristics could choose them over others leading 3886 // to an unsatisfactory solution. Passing LoserRegs into RateFormula here 3887 // avoids the need to recompute this information across formulae using the 3888 // same bad AddRec. Passing LoserRegs is also essential unless we remove 3889 // the corresponding bad register from the Regs set. 3890 Cost CostF; 3891 Regs.clear(); 3892 CostF.RateFormula(TTI, F, Regs, VisitedRegs, L, LU.Offsets, SE, DT, LU, 3893 &LoserRegs); 3894 if (CostF.isLoser()) { 3895 // During initial formula generation, undesirable formulae are generated 3896 // by uses within other loops that have some non-trivial address mode or 3897 // use the postinc form of the IV. LSR needs to provide these formulae 3898 // as the basis of rediscovering the desired formula that uses an AddRec 3899 // corresponding to the existing phi. Once all formulae have been 3900 // generated, these initial losers may be pruned. 3901 DEBUG(dbgs() << " Filtering loser "; F.print(dbgs()); 3902 dbgs() << "\n"); 3903 } 3904 else { 3905 SmallVector<const SCEV *, 4> Key; 3906 for (const SCEV *Reg : F.BaseRegs) { 3907 if (RegUses.isRegUsedByUsesOtherThan(Reg, LUIdx)) 3908 Key.push_back(Reg); 3909 } 3910 if (F.ScaledReg && 3911 RegUses.isRegUsedByUsesOtherThan(F.ScaledReg, LUIdx)) 3912 Key.push_back(F.ScaledReg); 3913 // Unstable sort by host order ok, because this is only used for 3914 // uniquifying. 3915 std::sort(Key.begin(), Key.end()); 3916 3917 std::pair<BestFormulaeTy::const_iterator, bool> P = 3918 BestFormulae.insert(std::make_pair(Key, FIdx)); 3919 if (P.second) 3920 continue; 3921 3922 Formula &Best = LU.Formulae[P.first->second]; 3923 3924 Cost CostBest; 3925 Regs.clear(); 3926 CostBest.RateFormula(TTI, Best, Regs, VisitedRegs, L, LU.Offsets, SE, 3927 DT, LU); 3928 if (CostF < CostBest) 3929 std::swap(F, Best); 3930 DEBUG(dbgs() << " Filtering out formula "; F.print(dbgs()); 3931 dbgs() << "\n" 3932 " in favor of formula "; Best.print(dbgs()); 3933 dbgs() << '\n'); 3934 } 3935#ifndef NDEBUG 3936 ChangedFormulae = true; 3937#endif 3938 LU.DeleteFormula(F); 3939 --FIdx; 3940 --NumForms; 3941 Any = true; 3942 } 3943 3944 // Now that we've filtered out some formulae, recompute the Regs set. 3945 if (Any) 3946 LU.RecomputeRegs(LUIdx, RegUses); 3947 3948 // Reset this to prepare for the next use. 3949 BestFormulae.clear(); 3950 } 3951 3952 DEBUG(if (ChangedFormulae) { 3953 dbgs() << "\n" 3954 "After filtering out undesirable candidates:\n"; 3955 print_uses(dbgs()); 3956 }); 3957} 3958 3959// This is a rough guess that seems to work fairly well. 3960static const size_t ComplexityLimit = UINT16_MAX; 3961 3962/// EstimateSearchSpaceComplexity - Estimate the worst-case number of 3963/// solutions the solver might have to consider. It almost never considers 3964/// this many solutions because it prune the search space, but the pruning 3965/// isn't always sufficient. 3966size_t LSRInstance::EstimateSearchSpaceComplexity() const { 3967 size_t Power = 1; 3968 for (const LSRUse &LU : Uses) { 3969 size_t FSize = LU.Formulae.size(); 3970 if (FSize >= ComplexityLimit) { 3971 Power = ComplexityLimit; 3972 break; 3973 } 3974 Power *= FSize; 3975 if (Power >= ComplexityLimit) 3976 break; 3977 } 3978 return Power; 3979} 3980 3981/// NarrowSearchSpaceByDetectingSupersets - When one formula uses a superset 3982/// of the registers of another formula, it won't help reduce register 3983/// pressure (though it may not necessarily hurt register pressure); remove 3984/// it to simplify the system. 3985void LSRInstance::NarrowSearchSpaceByDetectingSupersets() { 3986 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) { 3987 DEBUG(dbgs() << "The search space is too complex.\n"); 3988 3989 DEBUG(dbgs() << "Narrowing the search space by eliminating formulae " 3990 "which use a superset of registers used by other " 3991 "formulae.\n"); 3992 3993 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) { 3994 LSRUse &LU = Uses[LUIdx]; 3995 bool Any = false; 3996 for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) { 3997 Formula &F = LU.Formulae[i]; 3998 // Look for a formula with a constant or GV in a register. If the use 3999 // also has a formula with that same value in an immediate field, 4000 // delete the one that uses a register. 4001 for (SmallVectorImpl<const SCEV *>::const_iterator 4002 I = F.BaseRegs.begin(), E = F.BaseRegs.end(); I != E; ++I) { 4003 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(*I)) { 4004 Formula NewF = F; 4005 NewF.BaseOffset += C->getValue()->getSExtValue(); 4006 NewF.BaseRegs.erase(NewF.BaseRegs.begin() + 4007 (I - F.BaseRegs.begin())); 4008 if (LU.HasFormulaWithSameRegs(NewF)) { 4009 DEBUG(dbgs() << " Deleting "; F.print(dbgs()); dbgs() << '\n'); 4010 LU.DeleteFormula(F); 4011 --i; 4012 --e; 4013 Any = true; 4014 break; 4015 } 4016 } else if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(*I)) { 4017 if (GlobalValue *GV = dyn_cast<GlobalValue>(U->getValue())) 4018 if (!F.BaseGV) { 4019 Formula NewF = F; 4020 NewF.BaseGV = GV; 4021 NewF.BaseRegs.erase(NewF.BaseRegs.begin() + 4022 (I - F.BaseRegs.begin())); 4023 if (LU.HasFormulaWithSameRegs(NewF)) { 4024 DEBUG(dbgs() << " Deleting "; F.print(dbgs()); 4025 dbgs() << '\n'); 4026 LU.DeleteFormula(F); 4027 --i; 4028 --e; 4029 Any = true; 4030 break; 4031 } 4032 } 4033 } 4034 } 4035 } 4036 if (Any) 4037 LU.RecomputeRegs(LUIdx, RegUses); 4038 } 4039 4040 DEBUG(dbgs() << "After pre-selection:\n"; 4041 print_uses(dbgs())); 4042 } 4043} 4044 4045/// NarrowSearchSpaceByCollapsingUnrolledCode - When there are many registers 4046/// for expressions like A, A+1, A+2, etc., allocate a single register for 4047/// them. 4048void LSRInstance::NarrowSearchSpaceByCollapsingUnrolledCode() { 4049 if (EstimateSearchSpaceComplexity() < ComplexityLimit) 4050 return; 4051 4052 DEBUG(dbgs() << "The search space is too complex.\n" 4053 "Narrowing the search space by assuming that uses separated " 4054 "by a constant offset will use the same registers.\n"); 4055 4056 // This is especially useful for unrolled loops. 4057 4058 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) { 4059 LSRUse &LU = Uses[LUIdx]; 4060 for (const Formula &F : LU.Formulae) { 4061 if (F.BaseOffset == 0 || (F.Scale != 0 && F.Scale != 1)) 4062 continue; 4063 4064 LSRUse *LUThatHas = FindUseWithSimilarFormula(F, LU); 4065 if (!LUThatHas) 4066 continue; 4067 4068 if (!reconcileNewOffset(*LUThatHas, F.BaseOffset, /*HasBaseReg=*/ false, 4069 LU.Kind, LU.AccessTy)) 4070 continue; 4071 4072 DEBUG(dbgs() << " Deleting use "; LU.print(dbgs()); dbgs() << '\n'); 4073 4074 LUThatHas->AllFixupsOutsideLoop &= LU.AllFixupsOutsideLoop; 4075 4076 // Update the relocs to reference the new use. 4077 for (LSRFixup &Fixup : Fixups) { 4078 if (Fixup.LUIdx == LUIdx) { 4079 Fixup.LUIdx = LUThatHas - &Uses.front(); 4080 Fixup.Offset += F.BaseOffset; 4081 // Add the new offset to LUThatHas' offset list. 4082 if (LUThatHas->Offsets.back() != Fixup.Offset) { 4083 LUThatHas->Offsets.push_back(Fixup.Offset); 4084 if (Fixup.Offset > LUThatHas->MaxOffset) 4085 LUThatHas->MaxOffset = Fixup.Offset; 4086 if (Fixup.Offset < LUThatHas->MinOffset) 4087 LUThatHas->MinOffset = Fixup.Offset; 4088 } 4089 DEBUG(dbgs() << "New fixup has offset " << Fixup.Offset << '\n'); 4090 } 4091 if (Fixup.LUIdx == NumUses-1) 4092 Fixup.LUIdx = LUIdx; 4093 } 4094 4095 // Delete formulae from the new use which are no longer legal. 4096 bool Any = false; 4097 for (size_t i = 0, e = LUThatHas->Formulae.size(); i != e; ++i) { 4098 Formula &F = LUThatHas->Formulae[i]; 4099 if (!isLegalUse(TTI, LUThatHas->MinOffset, LUThatHas->MaxOffset, 4100 LUThatHas->Kind, LUThatHas->AccessTy, F)) { 4101 DEBUG(dbgs() << " Deleting "; F.print(dbgs()); 4102 dbgs() << '\n'); 4103 LUThatHas->DeleteFormula(F); 4104 --i; 4105 --e; 4106 Any = true; 4107 } 4108 } 4109 4110 if (Any) 4111 LUThatHas->RecomputeRegs(LUThatHas - &Uses.front(), RegUses); 4112 4113 // Delete the old use. 4114 DeleteUse(LU, LUIdx); 4115 --LUIdx; 4116 --NumUses; 4117 break; 4118 } 4119 } 4120 4121 DEBUG(dbgs() << "After pre-selection:\n"; print_uses(dbgs())); 4122} 4123 4124/// NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters - Call 4125/// FilterOutUndesirableDedicatedRegisters again, if necessary, now that 4126/// we've done more filtering, as it may be able to find more formulae to 4127/// eliminate. 4128void LSRInstance::NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters(){ 4129 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) { 4130 DEBUG(dbgs() << "The search space is too complex.\n"); 4131 4132 DEBUG(dbgs() << "Narrowing the search space by re-filtering out " 4133 "undesirable dedicated registers.\n"); 4134 4135 FilterOutUndesirableDedicatedRegisters(); 4136 4137 DEBUG(dbgs() << "After pre-selection:\n"; 4138 print_uses(dbgs())); 4139 } 4140} 4141 4142/// NarrowSearchSpaceByPickingWinnerRegs - Pick a register which seems likely 4143/// to be profitable, and then in any use which has any reference to that 4144/// register, delete all formulae which do not reference that register. 4145void LSRInstance::NarrowSearchSpaceByPickingWinnerRegs() { 4146 // With all other options exhausted, loop until the system is simple 4147 // enough to handle. 4148 SmallPtrSet<const SCEV *, 4> Taken; 4149 while (EstimateSearchSpaceComplexity() >= ComplexityLimit) { 4150 // Ok, we have too many of formulae on our hands to conveniently handle. 4151 // Use a rough heuristic to thin out the list. 4152 DEBUG(dbgs() << "The search space is too complex.\n"); 4153 4154 // Pick the register which is used by the most LSRUses, which is likely 4155 // to be a good reuse register candidate. 4156 const SCEV *Best = nullptr; 4157 unsigned BestNum = 0; 4158 for (const SCEV *Reg : RegUses) { 4159 if (Taken.count(Reg)) 4160 continue; 4161 if (!Best) 4162 Best = Reg; 4163 else { 4164 unsigned Count = RegUses.getUsedByIndices(Reg).count(); 4165 if (Count > BestNum) { 4166 Best = Reg; 4167 BestNum = Count; 4168 } 4169 } 4170 } 4171 4172 DEBUG(dbgs() << "Narrowing the search space by assuming " << *Best 4173 << " will yield profitable reuse.\n"); 4174 Taken.insert(Best); 4175 4176 // In any use with formulae which references this register, delete formulae 4177 // which don't reference it. 4178 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) { 4179 LSRUse &LU = Uses[LUIdx]; 4180 if (!LU.Regs.count(Best)) continue; 4181 4182 bool Any = false; 4183 for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) { 4184 Formula &F = LU.Formulae[i]; 4185 if (!F.referencesReg(Best)) { 4186 DEBUG(dbgs() << " Deleting "; F.print(dbgs()); dbgs() << '\n'); 4187 LU.DeleteFormula(F); 4188 --e; 4189 --i; 4190 Any = true; 4191 assert(e != 0 && "Use has no formulae left! Is Regs inconsistent?"); 4192 continue; 4193 } 4194 } 4195 4196 if (Any) 4197 LU.RecomputeRegs(LUIdx, RegUses); 4198 } 4199 4200 DEBUG(dbgs() << "After pre-selection:\n"; 4201 print_uses(dbgs())); 4202 } 4203} 4204 4205/// NarrowSearchSpaceUsingHeuristics - If there are an extraordinary number of 4206/// formulae to choose from, use some rough heuristics to prune down the number 4207/// of formulae. This keeps the main solver from taking an extraordinary amount 4208/// of time in some worst-case scenarios. 4209void LSRInstance::NarrowSearchSpaceUsingHeuristics() { 4210 NarrowSearchSpaceByDetectingSupersets(); 4211 NarrowSearchSpaceByCollapsingUnrolledCode(); 4212 NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters(); 4213 NarrowSearchSpaceByPickingWinnerRegs(); 4214} 4215 4216/// SolveRecurse - This is the recursive solver. 4217void LSRInstance::SolveRecurse(SmallVectorImpl<const Formula *> &Solution, 4218 Cost &SolutionCost, 4219 SmallVectorImpl<const Formula *> &Workspace, 4220 const Cost &CurCost, 4221 const SmallPtrSet<const SCEV *, 16> &CurRegs, 4222 DenseSet<const SCEV *> &VisitedRegs) const { 4223 // Some ideas: 4224 // - prune more: 4225 // - use more aggressive filtering 4226 // - sort the formula so that the most profitable solutions are found first 4227 // - sort the uses too 4228 // - search faster: 4229 // - don't compute a cost, and then compare. compare while computing a cost 4230 // and bail early. 4231 // - track register sets with SmallBitVector 4232 4233 const LSRUse &LU = Uses[Workspace.size()]; 4234 4235 // If this use references any register that's already a part of the 4236 // in-progress solution, consider it a requirement that a formula must 4237 // reference that register in order to be considered. This prunes out 4238 // unprofitable searching. 4239 SmallSetVector<const SCEV *, 4> ReqRegs; 4240 for (const SCEV *S : CurRegs) 4241 if (LU.Regs.count(S)) 4242 ReqRegs.insert(S); 4243 4244 SmallPtrSet<const SCEV *, 16> NewRegs; 4245 Cost NewCost; 4246 for (const Formula &F : LU.Formulae) { 4247 // Ignore formulae which may not be ideal in terms of register reuse of 4248 // ReqRegs. The formula should use all required registers before 4249 // introducing new ones. 4250 int NumReqRegsToFind = std::min(F.getNumRegs(), ReqRegs.size()); 4251 for (const SCEV *Reg : ReqRegs) { 4252 if ((F.ScaledReg && F.ScaledReg == Reg) || 4253 std::find(F.BaseRegs.begin(), F.BaseRegs.end(), Reg) != 4254 F.BaseRegs.end()) { 4255 --NumReqRegsToFind; 4256 if (NumReqRegsToFind == 0) 4257 break; 4258 } 4259 } 4260 if (NumReqRegsToFind != 0) { 4261 // If none of the formulae satisfied the required registers, then we could 4262 // clear ReqRegs and try again. Currently, we simply give up in this case. 4263 continue; 4264 } 4265 4266 // Evaluate the cost of the current formula. If it's already worse than 4267 // the current best, prune the search at that point. 4268 NewCost = CurCost; 4269 NewRegs = CurRegs; 4270 NewCost.RateFormula(TTI, F, NewRegs, VisitedRegs, L, LU.Offsets, SE, DT, 4271 LU); 4272 if (NewCost < SolutionCost) { 4273 Workspace.push_back(&F); 4274 if (Workspace.size() != Uses.size()) { 4275 SolveRecurse(Solution, SolutionCost, Workspace, NewCost, 4276 NewRegs, VisitedRegs); 4277 if (F.getNumRegs() == 1 && Workspace.size() == 1) 4278 VisitedRegs.insert(F.ScaledReg ? F.ScaledReg : F.BaseRegs[0]); 4279 } else { 4280 DEBUG(dbgs() << "New best at "; NewCost.print(dbgs()); 4281 dbgs() << ".\n Regs:"; 4282 for (const SCEV *S : NewRegs) 4283 dbgs() << ' ' << *S; 4284 dbgs() << '\n'); 4285 4286 SolutionCost = NewCost; 4287 Solution = Workspace; 4288 } 4289 Workspace.pop_back(); 4290 } 4291 } 4292} 4293 4294/// Solve - Choose one formula from each use. Return the results in the given 4295/// Solution vector. 4296void LSRInstance::Solve(SmallVectorImpl<const Formula *> &Solution) const { 4297 SmallVector<const Formula *, 8> Workspace; 4298 Cost SolutionCost; 4299 SolutionCost.Lose(); 4300 Cost CurCost; 4301 SmallPtrSet<const SCEV *, 16> CurRegs; 4302 DenseSet<const SCEV *> VisitedRegs; 4303 Workspace.reserve(Uses.size()); 4304 4305 // SolveRecurse does all the work. 4306 SolveRecurse(Solution, SolutionCost, Workspace, CurCost, 4307 CurRegs, VisitedRegs); 4308 if (Solution.empty()) { 4309 DEBUG(dbgs() << "\nNo Satisfactory Solution\n"); 4310 return; 4311 } 4312 4313 // Ok, we've now made all our decisions. 4314 DEBUG(dbgs() << "\n" 4315 "The chosen solution requires "; SolutionCost.print(dbgs()); 4316 dbgs() << ":\n"; 4317 for (size_t i = 0, e = Uses.size(); i != e; ++i) { 4318 dbgs() << " "; 4319 Uses[i].print(dbgs()); 4320 dbgs() << "\n" 4321 " "; 4322 Solution[i]->print(dbgs()); 4323 dbgs() << '\n'; 4324 }); 4325 4326 assert(Solution.size() == Uses.size() && "Malformed solution!"); 4327} 4328 4329/// HoistInsertPosition - Helper for AdjustInsertPositionForExpand. Climb up 4330/// the dominator tree far as we can go while still being dominated by the 4331/// input positions. This helps canonicalize the insert position, which 4332/// encourages sharing. 4333BasicBlock::iterator 4334LSRInstance::HoistInsertPosition(BasicBlock::iterator IP, 4335 const SmallVectorImpl<Instruction *> &Inputs) 4336 const { 4337 for (;;) { 4338 const Loop *IPLoop = LI.getLoopFor(IP->getParent()); 4339 unsigned IPLoopDepth = IPLoop ? IPLoop->getLoopDepth() : 0; 4340 4341 BasicBlock *IDom; 4342 for (DomTreeNode *Rung = DT.getNode(IP->getParent()); ; ) { 4343 if (!Rung) return IP; 4344 Rung = Rung->getIDom(); 4345 if (!Rung) return IP; 4346 IDom = Rung->getBlock(); 4347 4348 // Don't climb into a loop though. 4349 const Loop *IDomLoop = LI.getLoopFor(IDom); 4350 unsigned IDomDepth = IDomLoop ? IDomLoop->getLoopDepth() : 0; 4351 if (IDomDepth <= IPLoopDepth && 4352 (IDomDepth != IPLoopDepth || IDomLoop == IPLoop)) 4353 break; 4354 } 4355 4356 bool AllDominate = true; 4357 Instruction *BetterPos = nullptr; 4358 Instruction *Tentative = IDom->getTerminator(); 4359 for (Instruction *Inst : Inputs) { 4360 if (Inst == Tentative || !DT.dominates(Inst, Tentative)) { 4361 AllDominate = false; 4362 break; 4363 } 4364 // Attempt to find an insert position in the middle of the block, 4365 // instead of at the end, so that it can be used for other expansions. 4366 if (IDom == Inst->getParent() && 4367 (!BetterPos || !DT.dominates(Inst, BetterPos))) 4368 BetterPos = std::next(BasicBlock::iterator(Inst)); 4369 } 4370 if (!AllDominate) 4371 break; 4372 if (BetterPos) 4373 IP = BetterPos; 4374 else 4375 IP = Tentative; 4376 } 4377 4378 return IP; 4379} 4380 4381/// AdjustInsertPositionForExpand - Determine an input position which will be 4382/// dominated by the operands and which will dominate the result. 4383BasicBlock::iterator 4384LSRInstance::AdjustInsertPositionForExpand(BasicBlock::iterator LowestIP, 4385 const LSRFixup &LF, 4386 const LSRUse &LU, 4387 SCEVExpander &Rewriter) const { 4388 // Collect some instructions which must be dominated by the 4389 // expanding replacement. These must be dominated by any operands that 4390 // will be required in the expansion. 4391 SmallVector<Instruction *, 4> Inputs; 4392 if (Instruction *I = dyn_cast<Instruction>(LF.OperandValToReplace)) 4393 Inputs.push_back(I); 4394 if (LU.Kind == LSRUse::ICmpZero) 4395 if (Instruction *I = 4396 dyn_cast<Instruction>(cast<ICmpInst>(LF.UserInst)->getOperand(1))) 4397 Inputs.push_back(I); 4398 if (LF.PostIncLoops.count(L)) { 4399 if (LF.isUseFullyOutsideLoop(L)) 4400 Inputs.push_back(L->getLoopLatch()->getTerminator()); 4401 else 4402 Inputs.push_back(IVIncInsertPos); 4403 } 4404 // The expansion must also be dominated by the increment positions of any 4405 // loops it for which it is using post-inc mode. 4406 for (const Loop *PIL : LF.PostIncLoops) { 4407 if (PIL == L) continue; 4408 4409 // Be dominated by the loop exit. 4410 SmallVector<BasicBlock *, 4> ExitingBlocks; 4411 PIL->getExitingBlocks(ExitingBlocks); 4412 if (!ExitingBlocks.empty()) { 4413 BasicBlock *BB = ExitingBlocks[0]; 4414 for (unsigned i = 1, e = ExitingBlocks.size(); i != e; ++i) 4415 BB = DT.findNearestCommonDominator(BB, ExitingBlocks[i]); 4416 Inputs.push_back(BB->getTerminator()); 4417 } 4418 } 4419 4420 assert(!isa<PHINode>(LowestIP) && !isa<LandingPadInst>(LowestIP) 4421 && !isa<DbgInfoIntrinsic>(LowestIP) && 4422 "Insertion point must be a normal instruction"); 4423 4424 // Then, climb up the immediate dominator tree as far as we can go while 4425 // still being dominated by the input positions. 4426 BasicBlock::iterator IP = HoistInsertPosition(LowestIP, Inputs); 4427 4428 // Don't insert instructions before PHI nodes. 4429 while (isa<PHINode>(IP)) ++IP; 4430 4431 // Ignore landingpad instructions. 4432 while (isa<LandingPadInst>(IP)) ++IP; 4433 4434 // Ignore debug intrinsics. 4435 while (isa<DbgInfoIntrinsic>(IP)) ++IP; 4436 4437 // Set IP below instructions recently inserted by SCEVExpander. This keeps the 4438 // IP consistent across expansions and allows the previously inserted 4439 // instructions to be reused by subsequent expansion. 4440 while (Rewriter.isInsertedInstruction(IP) && IP != LowestIP) ++IP; 4441 4442 return IP; 4443} 4444 4445/// Expand - Emit instructions for the leading candidate expression for this 4446/// LSRUse (this is called "expanding"). 4447Value *LSRInstance::Expand(const LSRFixup &LF, 4448 const Formula &F, 4449 BasicBlock::iterator IP, 4450 SCEVExpander &Rewriter, 4451 SmallVectorImpl<WeakVH> &DeadInsts) const { 4452 const LSRUse &LU = Uses[LF.LUIdx]; 4453 if (LU.RigidFormula) 4454 return LF.OperandValToReplace; 4455 4456 // Determine an input position which will be dominated by the operands and 4457 // which will dominate the result. 4458 IP = AdjustInsertPositionForExpand(IP, LF, LU, Rewriter); 4459 4460 // Inform the Rewriter if we have a post-increment use, so that it can 4461 // perform an advantageous expansion. 4462 Rewriter.setPostInc(LF.PostIncLoops); 4463 4464 // This is the type that the user actually needs. 4465 Type *OpTy = LF.OperandValToReplace->getType(); 4466 // This will be the type that we'll initially expand to. 4467 Type *Ty = F.getType(); 4468 if (!Ty) 4469 // No type known; just expand directly to the ultimate type. 4470 Ty = OpTy; 4471 else if (SE.getEffectiveSCEVType(Ty) == SE.getEffectiveSCEVType(OpTy)) 4472 // Expand directly to the ultimate type if it's the right size. 4473 Ty = OpTy; 4474 // This is the type to do integer arithmetic in. 4475 Type *IntTy = SE.getEffectiveSCEVType(Ty); 4476 4477 // Build up a list of operands to add together to form the full base. 4478 SmallVector<const SCEV *, 8> Ops; 4479 4480 // Expand the BaseRegs portion. 4481 for (const SCEV *Reg : F.BaseRegs) { 4482 assert(!Reg->isZero() && "Zero allocated in a base register!"); 4483 4484 // If we're expanding for a post-inc user, make the post-inc adjustment. 4485 PostIncLoopSet &Loops = const_cast<PostIncLoopSet &>(LF.PostIncLoops); 4486 Reg = TransformForPostIncUse(Denormalize, Reg, 4487 LF.UserInst, LF.OperandValToReplace, 4488 Loops, SE, DT); 4489 4490 Ops.push_back(SE.getUnknown(Rewriter.expandCodeFor(Reg, nullptr, IP))); 4491 } 4492 4493 // Expand the ScaledReg portion. 4494 Value *ICmpScaledV = nullptr; 4495 if (F.Scale != 0) { 4496 const SCEV *ScaledS = F.ScaledReg; 4497 4498 // If we're expanding for a post-inc user, make the post-inc adjustment. 4499 PostIncLoopSet &Loops = const_cast<PostIncLoopSet &>(LF.PostIncLoops); 4500 ScaledS = TransformForPostIncUse(Denormalize, ScaledS, 4501 LF.UserInst, LF.OperandValToReplace, 4502 Loops, SE, DT); 4503 4504 if (LU.Kind == LSRUse::ICmpZero) { 4505 // Expand ScaleReg as if it was part of the base regs. 4506 if (F.Scale == 1) 4507 Ops.push_back( 4508 SE.getUnknown(Rewriter.expandCodeFor(ScaledS, nullptr, IP))); 4509 else { 4510 // An interesting way of "folding" with an icmp is to use a negated 4511 // scale, which we'll implement by inserting it into the other operand 4512 // of the icmp. 4513 assert(F.Scale == -1 && 4514 "The only scale supported by ICmpZero uses is -1!"); 4515 ICmpScaledV = Rewriter.expandCodeFor(ScaledS, nullptr, IP); 4516 } 4517 } else { 4518 // Otherwise just expand the scaled register and an explicit scale, 4519 // which is expected to be matched as part of the address. 4520 4521 // Flush the operand list to suppress SCEVExpander hoisting address modes. 4522 // Unless the addressing mode will not be folded. 4523 if (!Ops.empty() && LU.Kind == LSRUse::Address && 4524 isAMCompletelyFolded(TTI, LU, F)) { 4525 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP); 4526 Ops.clear(); 4527 Ops.push_back(SE.getUnknown(FullV)); 4528 } 4529 ScaledS = SE.getUnknown(Rewriter.expandCodeFor(ScaledS, nullptr, IP)); 4530 if (F.Scale != 1) 4531 ScaledS = 4532 SE.getMulExpr(ScaledS, SE.getConstant(ScaledS->getType(), F.Scale)); 4533 Ops.push_back(ScaledS); 4534 } 4535 } 4536 4537 // Expand the GV portion. 4538 if (F.BaseGV) { 4539 // Flush the operand list to suppress SCEVExpander hoisting. 4540 if (!Ops.empty()) { 4541 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP); 4542 Ops.clear(); 4543 Ops.push_back(SE.getUnknown(FullV)); 4544 } 4545 Ops.push_back(SE.getUnknown(F.BaseGV)); 4546 } 4547 4548 // Flush the operand list to suppress SCEVExpander hoisting of both folded and 4549 // unfolded offsets. LSR assumes they both live next to their uses. 4550 if (!Ops.empty()) { 4551 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP); 4552 Ops.clear(); 4553 Ops.push_back(SE.getUnknown(FullV)); 4554 } 4555 4556 // Expand the immediate portion. 4557 int64_t Offset = (uint64_t)F.BaseOffset + LF.Offset; 4558 if (Offset != 0) { 4559 if (LU.Kind == LSRUse::ICmpZero) { 4560 // The other interesting way of "folding" with an ICmpZero is to use a 4561 // negated immediate. 4562 if (!ICmpScaledV) 4563 ICmpScaledV = ConstantInt::get(IntTy, -(uint64_t)Offset); 4564 else { 4565 Ops.push_back(SE.getUnknown(ICmpScaledV)); 4566 ICmpScaledV = ConstantInt::get(IntTy, Offset); 4567 } 4568 } else { 4569 // Just add the immediate values. These again are expected to be matched 4570 // as part of the address. 4571 Ops.push_back(SE.getUnknown(ConstantInt::getSigned(IntTy, Offset))); 4572 } 4573 } 4574 4575 // Expand the unfolded offset portion. 4576 int64_t UnfoldedOffset = F.UnfoldedOffset; 4577 if (UnfoldedOffset != 0) { 4578 // Just add the immediate values. 4579 Ops.push_back(SE.getUnknown(ConstantInt::getSigned(IntTy, 4580 UnfoldedOffset))); 4581 } 4582 4583 // Emit instructions summing all the operands. 4584 const SCEV *FullS = Ops.empty() ? 4585 SE.getConstant(IntTy, 0) : 4586 SE.getAddExpr(Ops); 4587 Value *FullV = Rewriter.expandCodeFor(FullS, Ty, IP); 4588 4589 // We're done expanding now, so reset the rewriter. 4590 Rewriter.clearPostInc(); 4591 4592 // An ICmpZero Formula represents an ICmp which we're handling as a 4593 // comparison against zero. Now that we've expanded an expression for that 4594 // form, update the ICmp's other operand. 4595 if (LU.Kind == LSRUse::ICmpZero) { 4596 ICmpInst *CI = cast<ICmpInst>(LF.UserInst); 4597 DeadInsts.emplace_back(CI->getOperand(1)); 4598 assert(!F.BaseGV && "ICmp does not support folding a global value and " 4599 "a scale at the same time!"); 4600 if (F.Scale == -1) { 4601 if (ICmpScaledV->getType() != OpTy) { 4602 Instruction *Cast = 4603 CastInst::Create(CastInst::getCastOpcode(ICmpScaledV, false, 4604 OpTy, false), 4605 ICmpScaledV, OpTy, "tmp", CI); 4606 ICmpScaledV = Cast; 4607 } 4608 CI->setOperand(1, ICmpScaledV); 4609 } else { 4610 // A scale of 1 means that the scale has been expanded as part of the 4611 // base regs. 4612 assert((F.Scale == 0 || F.Scale == 1) && 4613 "ICmp does not support folding a global value and " 4614 "a scale at the same time!"); 4615 Constant *C = ConstantInt::getSigned(SE.getEffectiveSCEVType(OpTy), 4616 -(uint64_t)Offset); 4617 if (C->getType() != OpTy) 4618 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false, 4619 OpTy, false), 4620 C, OpTy); 4621 4622 CI->setOperand(1, C); 4623 } 4624 } 4625 4626 return FullV; 4627} 4628 4629/// RewriteForPHI - Helper for Rewrite. PHI nodes are special because the use 4630/// of their operands effectively happens in their predecessor blocks, so the 4631/// expression may need to be expanded in multiple places. 4632void LSRInstance::RewriteForPHI(PHINode *PN, 4633 const LSRFixup &LF, 4634 const Formula &F, 4635 SCEVExpander &Rewriter, 4636 SmallVectorImpl<WeakVH> &DeadInsts, 4637 Pass *P) const { 4638 DenseMap<BasicBlock *, Value *> Inserted; 4639 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) 4640 if (PN->getIncomingValue(i) == LF.OperandValToReplace) { 4641 BasicBlock *BB = PN->getIncomingBlock(i); 4642 4643 // If this is a critical edge, split the edge so that we do not insert 4644 // the code on all predecessor/successor paths. We do this unless this 4645 // is the canonical backedge for this loop, which complicates post-inc 4646 // users. 4647 if (e != 1 && BB->getTerminator()->getNumSuccessors() > 1 && 4648 !isa<IndirectBrInst>(BB->getTerminator())) { 4649 BasicBlock *Parent = PN->getParent(); 4650 Loop *PNLoop = LI.getLoopFor(Parent); 4651 if (!PNLoop || Parent != PNLoop->getHeader()) { 4652 // Split the critical edge. 4653 BasicBlock *NewBB = nullptr; 4654 if (!Parent->isLandingPad()) { 4655 NewBB = SplitCriticalEdge(BB, Parent, 4656 CriticalEdgeSplittingOptions(&DT, &LI) 4657 .setMergeIdenticalEdges() 4658 .setDontDeleteUselessPHIs()); 4659 } else { 4660 SmallVector<BasicBlock*, 2> NewBBs; 4661 SplitLandingPadPredecessors(Parent, BB, "", "", NewBBs, 4662 /*AliasAnalysis*/ nullptr, &DT, &LI); 4663 NewBB = NewBBs[0]; 4664 } 4665 // If NewBB==NULL, then SplitCriticalEdge refused to split because all 4666 // phi predecessors are identical. The simple thing to do is skip 4667 // splitting in this case rather than complicate the API. 4668 if (NewBB) { 4669 // If PN is outside of the loop and BB is in the loop, we want to 4670 // move the block to be immediately before the PHI block, not 4671 // immediately after BB. 4672 if (L->contains(BB) && !L->contains(PN)) 4673 NewBB->moveBefore(PN->getParent()); 4674 4675 // Splitting the edge can reduce the number of PHI entries we have. 4676 e = PN->getNumIncomingValues(); 4677 BB = NewBB; 4678 i = PN->getBasicBlockIndex(BB); 4679 } 4680 } 4681 } 4682 4683 std::pair<DenseMap<BasicBlock *, Value *>::iterator, bool> Pair = 4684 Inserted.insert(std::make_pair(BB, static_cast<Value *>(nullptr))); 4685 if (!Pair.second) 4686 PN->setIncomingValue(i, Pair.first->second); 4687 else { 4688 Value *FullV = Expand(LF, F, BB->getTerminator(), Rewriter, DeadInsts); 4689 4690 // If this is reuse-by-noop-cast, insert the noop cast. 4691 Type *OpTy = LF.OperandValToReplace->getType(); 4692 if (FullV->getType() != OpTy) 4693 FullV = 4694 CastInst::Create(CastInst::getCastOpcode(FullV, false, 4695 OpTy, false), 4696 FullV, LF.OperandValToReplace->getType(), 4697 "tmp", BB->getTerminator()); 4698 4699 PN->setIncomingValue(i, FullV); 4700 Pair.first->second = FullV; 4701 } 4702 } 4703} 4704 4705/// Rewrite - Emit instructions for the leading candidate expression for this 4706/// LSRUse (this is called "expanding"), and update the UserInst to reference 4707/// the newly expanded value. 4708void LSRInstance::Rewrite(const LSRFixup &LF, 4709 const Formula &F, 4710 SCEVExpander &Rewriter, 4711 SmallVectorImpl<WeakVH> &DeadInsts, 4712 Pass *P) const { 4713 // First, find an insertion point that dominates UserInst. For PHI nodes, 4714 // find the nearest block which dominates all the relevant uses. 4715 if (PHINode *PN = dyn_cast<PHINode>(LF.UserInst)) { 4716 RewriteForPHI(PN, LF, F, Rewriter, DeadInsts, P); 4717 } else { 4718 Value *FullV = Expand(LF, F, LF.UserInst, Rewriter, DeadInsts); 4719 4720 // If this is reuse-by-noop-cast, insert the noop cast. 4721 Type *OpTy = LF.OperandValToReplace->getType(); 4722 if (FullV->getType() != OpTy) { 4723 Instruction *Cast = 4724 CastInst::Create(CastInst::getCastOpcode(FullV, false, OpTy, false), 4725 FullV, OpTy, "tmp", LF.UserInst); 4726 FullV = Cast; 4727 } 4728 4729 // Update the user. ICmpZero is handled specially here (for now) because 4730 // Expand may have updated one of the operands of the icmp already, and 4731 // its new value may happen to be equal to LF.OperandValToReplace, in 4732 // which case doing replaceUsesOfWith leads to replacing both operands 4733 // with the same value. TODO: Reorganize this. 4734 if (Uses[LF.LUIdx].Kind == LSRUse::ICmpZero) 4735 LF.UserInst->setOperand(0, FullV); 4736 else 4737 LF.UserInst->replaceUsesOfWith(LF.OperandValToReplace, FullV); 4738 } 4739 4740 DeadInsts.emplace_back(LF.OperandValToReplace); 4741} 4742 4743/// ImplementSolution - Rewrite all the fixup locations with new values, 4744/// following the chosen solution. 4745void 4746LSRInstance::ImplementSolution(const SmallVectorImpl<const Formula *> &Solution, 4747 Pass *P) { 4748 // Keep track of instructions we may have made dead, so that 4749 // we can remove them after we are done working. 4750 SmallVector<WeakVH, 16> DeadInsts; 4751 4752 SCEVExpander Rewriter(SE, L->getHeader()->getModule()->getDataLayout(), 4753 "lsr"); 4754#ifndef NDEBUG 4755 Rewriter.setDebugType(DEBUG_TYPE); 4756#endif 4757 Rewriter.disableCanonicalMode(); 4758 Rewriter.enableLSRMode(); 4759 Rewriter.setIVIncInsertPos(L, IVIncInsertPos); 4760 4761 // Mark phi nodes that terminate chains so the expander tries to reuse them. 4762 for (const IVChain &Chain : IVChainVec) { 4763 if (PHINode *PN = dyn_cast<PHINode>(Chain.tailUserInst())) 4764 Rewriter.setChainedPhi(PN); 4765 } 4766 4767 // Expand the new value definitions and update the users. 4768 for (const LSRFixup &Fixup : Fixups) { 4769 Rewrite(Fixup, *Solution[Fixup.LUIdx], Rewriter, DeadInsts, P); 4770 4771 Changed = true; 4772 } 4773 4774 for (const IVChain &Chain : IVChainVec) { 4775 GenerateIVChain(Chain, Rewriter, DeadInsts); 4776 Changed = true; 4777 } 4778 // Clean up after ourselves. This must be done before deleting any 4779 // instructions. 4780 Rewriter.clear(); 4781 4782 Changed |= DeleteTriviallyDeadInstructions(DeadInsts); 4783} 4784 4785LSRInstance::LSRInstance(Loop *L, Pass *P) 4786 : IU(P->getAnalysis<IVUsers>()), SE(P->getAnalysis<ScalarEvolution>()), 4787 DT(P->getAnalysis<DominatorTreeWrapperPass>().getDomTree()), 4788 LI(P->getAnalysis<LoopInfoWrapperPass>().getLoopInfo()), 4789 TTI(P->getAnalysis<TargetTransformInfoWrapperPass>().getTTI( 4790 *L->getHeader()->getParent())), 4791 L(L), Changed(false), IVIncInsertPos(nullptr) { 4792 // If LoopSimplify form is not available, stay out of trouble. 4793 if (!L->isLoopSimplifyForm()) 4794 return; 4795 4796 // If there's no interesting work to be done, bail early. 4797 if (IU.empty()) return; 4798 4799 // If there's too much analysis to be done, bail early. We won't be able to 4800 // model the problem anyway. 4801 unsigned NumUsers = 0; 4802 for (const IVStrideUse &U : IU) { 4803 if (++NumUsers > MaxIVUsers) { 4804 (void)U; 4805 DEBUG(dbgs() << "LSR skipping loop, too many IV Users in " << U << "\n"); 4806 return; 4807 } 4808 } 4809 4810#ifndef NDEBUG 4811 // All dominating loops must have preheaders, or SCEVExpander may not be able 4812 // to materialize an AddRecExpr whose Start is an outer AddRecExpr. 4813 // 4814 // IVUsers analysis should only create users that are dominated by simple loop 4815 // headers. Since this loop should dominate all of its users, its user list 4816 // should be empty if this loop itself is not within a simple loop nest. 4817 for (DomTreeNode *Rung = DT.getNode(L->getLoopPreheader()); 4818 Rung; Rung = Rung->getIDom()) { 4819 BasicBlock *BB = Rung->getBlock(); 4820 const Loop *DomLoop = LI.getLoopFor(BB); 4821 if (DomLoop && DomLoop->getHeader() == BB) { 4822 assert(DomLoop->getLoopPreheader() && "LSR needs a simplified loop nest"); 4823 } 4824 } 4825#endif // DEBUG 4826 4827 DEBUG(dbgs() << "\nLSR on loop "; 4828 L->getHeader()->printAsOperand(dbgs(), /*PrintType=*/false); 4829 dbgs() << ":\n"); 4830 4831 // First, perform some low-level loop optimizations. 4832 OptimizeShadowIV(); 4833 OptimizeLoopTermCond(); 4834 4835 // If loop preparation eliminates all interesting IV users, bail. 4836 if (IU.empty()) return; 4837 4838 // Skip nested loops until we can model them better with formulae. 4839 if (!L->empty()) { 4840 DEBUG(dbgs() << "LSR skipping outer loop " << *L << "\n"); 4841 return; 4842 } 4843 4844 // Start collecting data and preparing for the solver. 4845 CollectChains(); 4846 CollectInterestingTypesAndFactors(); 4847 CollectFixupsAndInitialFormulae(); 4848 CollectLoopInvariantFixupsAndFormulae(); 4849 4850 assert(!Uses.empty() && "IVUsers reported at least one use"); 4851 DEBUG(dbgs() << "LSR found " << Uses.size() << " uses:\n"; 4852 print_uses(dbgs())); 4853 4854 // Now use the reuse data to generate a bunch of interesting ways 4855 // to formulate the values needed for the uses. 4856 GenerateAllReuseFormulae(); 4857 4858 FilterOutUndesirableDedicatedRegisters(); 4859 NarrowSearchSpaceUsingHeuristics(); 4860 4861 SmallVector<const Formula *, 8> Solution; 4862 Solve(Solution); 4863 4864 // Release memory that is no longer needed. 4865 Factors.clear(); 4866 Types.clear(); 4867 RegUses.clear(); 4868 4869 if (Solution.empty()) 4870 return; 4871 4872#ifndef NDEBUG 4873 // Formulae should be legal. 4874 for (const LSRUse &LU : Uses) { 4875 for (const Formula &F : LU.Formulae) 4876 assert(isLegalUse(TTI, LU.MinOffset, LU.MaxOffset, LU.Kind, LU.AccessTy, 4877 F) && "Illegal formula generated!"); 4878 }; 4879#endif 4880 4881 // Now that we've decided what we want, make it so. 4882 ImplementSolution(Solution, P); 4883} 4884 4885void LSRInstance::print_factors_and_types(raw_ostream &OS) const { 4886 if (Factors.empty() && Types.empty()) return; 4887 4888 OS << "LSR has identified the following interesting factors and types: "; 4889 bool First = true; 4890 4891 for (int64_t Factor : Factors) { 4892 if (!First) OS << ", "; 4893 First = false; 4894 OS << '*' << Factor; 4895 } 4896 4897 for (Type *Ty : Types) { 4898 if (!First) OS << ", "; 4899 First = false; 4900 OS << '(' << *Ty << ')'; 4901 } 4902 OS << '\n'; 4903} 4904 4905void LSRInstance::print_fixups(raw_ostream &OS) const { 4906 OS << "LSR is examining the following fixup sites:\n"; 4907 for (const LSRFixup &LF : Fixups) { 4908 dbgs() << " "; 4909 LF.print(OS); 4910 OS << '\n'; 4911 } 4912} 4913 4914void LSRInstance::print_uses(raw_ostream &OS) const { 4915 OS << "LSR is examining the following uses:\n"; 4916 for (const LSRUse &LU : Uses) { 4917 dbgs() << " "; 4918 LU.print(OS); 4919 OS << '\n'; 4920 for (const Formula &F : LU.Formulae) { 4921 OS << " "; 4922 F.print(OS); 4923 OS << '\n'; 4924 } 4925 } 4926} 4927 4928void LSRInstance::print(raw_ostream &OS) const { 4929 print_factors_and_types(OS); 4930 print_fixups(OS); 4931 print_uses(OS); 4932} 4933 4934#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) 4935void LSRInstance::dump() const { 4936 print(errs()); errs() << '\n'; 4937} 4938#endif 4939 4940namespace { 4941 4942class LoopStrengthReduce : public LoopPass { 4943public: 4944 static char ID; // Pass ID, replacement for typeid 4945 LoopStrengthReduce(); 4946 4947private: 4948 bool runOnLoop(Loop *L, LPPassManager &LPM) override; 4949 void getAnalysisUsage(AnalysisUsage &AU) const override; 4950}; 4951 4952} 4953 4954char LoopStrengthReduce::ID = 0; 4955INITIALIZE_PASS_BEGIN(LoopStrengthReduce, "loop-reduce", 4956 "Loop Strength Reduction", false, false) 4957INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass) 4958INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass) 4959INITIALIZE_PASS_DEPENDENCY(ScalarEvolution) 4960INITIALIZE_PASS_DEPENDENCY(IVUsers) 4961INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass) 4962INITIALIZE_PASS_DEPENDENCY(LoopSimplify) 4963INITIALIZE_PASS_END(LoopStrengthReduce, "loop-reduce", 4964 "Loop Strength Reduction", false, false) 4965 4966 4967Pass *llvm::createLoopStrengthReducePass() { 4968 return new LoopStrengthReduce(); 4969} 4970 4971LoopStrengthReduce::LoopStrengthReduce() : LoopPass(ID) { 4972 initializeLoopStrengthReducePass(*PassRegistry::getPassRegistry()); 4973} 4974 4975void LoopStrengthReduce::getAnalysisUsage(AnalysisUsage &AU) const { 4976 // We split critical edges, so we change the CFG. However, we do update 4977 // many analyses if they are around. 4978 AU.addPreservedID(LoopSimplifyID); 4979 4980 AU.addRequired<LoopInfoWrapperPass>(); 4981 AU.addPreserved<LoopInfoWrapperPass>(); 4982 AU.addRequiredID(LoopSimplifyID); 4983 AU.addRequired<DominatorTreeWrapperPass>(); 4984 AU.addPreserved<DominatorTreeWrapperPass>(); 4985 AU.addRequired<ScalarEvolution>(); 4986 AU.addPreserved<ScalarEvolution>(); 4987 // Requiring LoopSimplify a second time here prevents IVUsers from running 4988 // twice, since LoopSimplify was invalidated by running ScalarEvolution. 4989 AU.addRequiredID(LoopSimplifyID); 4990 AU.addRequired<IVUsers>(); 4991 AU.addPreserved<IVUsers>(); 4992 AU.addRequired<TargetTransformInfoWrapperPass>(); 4993} 4994 4995bool LoopStrengthReduce::runOnLoop(Loop *L, LPPassManager & /*LPM*/) { 4996 if (skipOptnoneFunction(L)) 4997 return false; 4998 4999 bool Changed = false; 5000 5001 // Run the main LSR transformation. 5002 Changed |= LSRInstance(L, this).getChanged(); 5003 5004 // Remove any extra phis created by processing inner loops. 5005 Changed |= DeleteDeadPHIs(L->getHeader()); 5006 if (EnablePhiElim && L->isLoopSimplifyForm()) { 5007 SmallVector<WeakVH, 16> DeadInsts; 5008 const DataLayout &DL = L->getHeader()->getModule()->getDataLayout(); 5009 SCEVExpander Rewriter(getAnalysis<ScalarEvolution>(), DL, "lsr"); 5010#ifndef NDEBUG 5011 Rewriter.setDebugType(DEBUG_TYPE); 5012#endif 5013 unsigned numFolded = Rewriter.replaceCongruentIVs( 5014 L, &getAnalysis<DominatorTreeWrapperPass>().getDomTree(), DeadInsts, 5015 &getAnalysis<TargetTransformInfoWrapperPass>().getTTI( 5016 *L->getHeader()->getParent())); 5017 if (numFolded) { 5018 Changed = true; 5019 DeleteTriviallyDeadInstructions(DeadInsts); 5020 DeleteDeadPHIs(L->getHeader()); 5021 } 5022 } 5023 return Changed; 5024} 5025