1 //===- LoopStrengthReduce.cpp - Strength Reduce IVs in Loops --------------===// 2 // 3 // The LLVM Compiler Infrastructure 4 // 5 // This file is distributed under the University of Illinois Open Source 6 // License. See LICENSE.TXT for details. 7 // 8 //===----------------------------------------------------------------------===// 9 // 10 // This transformation analyzes and transforms the induction variables (and 11 // computations derived from them) into forms suitable for efficient execution 12 // on the target. 13 // 14 // This pass performs a strength reduction on array references inside loops that 15 // have as one or more of their components the loop induction variable, it 16 // rewrites expressions to take advantage of scaled-index addressing modes 17 // available on the target, and it performs a variety of other optimizations 18 // related to loop induction variables. 19 // 20 // Terminology note: this code has a lot of handling for "post-increment" or 21 // "post-inc" users. This is not talking about post-increment addressing modes; 22 // it is instead talking about code like this: 23 // 24 // %i = phi [ 0, %entry ], [ %i.next, %latch ] 25 // ... 26 // %i.next = add %i, 1 27 // %c = icmp eq %i.next, %n 28 // 29 // The SCEV for %i is {0,+,1}<%L>. The SCEV for %i.next is {1,+,1}<%L>, however 30 // it's useful to think about these as the same register, with some uses using 31 // the value of the register before the add and some using // it after. In this 32 // example, the icmp is a post-increment user, since it uses %i.next, which is 33 // the value of the induction variable after the increment. The other common 34 // case of post-increment users is users outside the loop. 35 // 36 // TODO: More sophistication in the way Formulae are generated and filtered. 37 // 38 // TODO: Handle multiple loops at a time. 39 // 40 // TODO: Should TargetLowering::AddrMode::BaseGV be changed to a ConstantExpr 41 // instead of a GlobalValue? 42 // 43 // TODO: When truncation is free, truncate ICmp users' operands to make it a 44 // smaller encoding (on x86 at least). 45 // 46 // TODO: When a negated register is used by an add (such as in a list of 47 // multiple base registers, or as the increment expression in an addrec), 48 // we may not actually need both reg and (-1 * reg) in registers; the 49 // negation can be implemented by using a sub instead of an add. The 50 // lack of support for taking this into consideration when making 51 // register pressure decisions is partly worked around by the "Special" 52 // use kind. 53 // 54 //===----------------------------------------------------------------------===// 55 56 #define DEBUG_TYPE "loop-reduce" 57 #include "llvm/Transforms/Scalar.h" 58 #include "llvm/Constants.h" 59 #include "llvm/Instructions.h" 60 #include "llvm/IntrinsicInst.h" 61 #include "llvm/DerivedTypes.h" 62 #include "llvm/Analysis/IVUsers.h" 63 #include "llvm/Analysis/Dominators.h" 64 #include "llvm/Analysis/LoopPass.h" 65 #include "llvm/Analysis/ScalarEvolutionExpander.h" 66 #include "llvm/Assembly/Writer.h" 67 #include "llvm/Transforms/Utils/BasicBlockUtils.h" 68 #include "llvm/Transforms/Utils/Local.h" 69 #include "llvm/ADT/SmallBitVector.h" 70 #include "llvm/ADT/SetVector.h" 71 #include "llvm/ADT/DenseSet.h" 72 #include "llvm/Support/Debug.h" 73 #include "llvm/Support/CommandLine.h" 74 #include "llvm/Support/ValueHandle.h" 75 #include "llvm/Support/raw_ostream.h" 76 #include "llvm/Target/TargetLowering.h" 77 #include <algorithm> 78 using namespace llvm; 79 80 /// MaxIVUsers is an arbitrary threshold that provides an early opportunitiy for 81 /// bail out. This threshold is far beyond the number of users that LSR can 82 /// conceivably solve, so it should not affect generated code, but catches the 83 /// worst cases before LSR burns too much compile time and stack space. 84 static const unsigned MaxIVUsers = 200; 85 86 // Temporary flag to cleanup congruent phis after LSR phi expansion. 87 // It's currently disabled until we can determine whether it's truly useful or 88 // not. The flag should be removed after the v3.0 release. 89 // This is now needed for ivchains. 90 static cl::opt<bool> EnablePhiElim( 91 "enable-lsr-phielim", cl::Hidden, cl::init(true), 92 cl::desc("Enable LSR phi elimination")); 93 94 #ifndef NDEBUG 95 // Stress test IV chain generation. 96 static cl::opt<bool> StressIVChain( 97 "stress-ivchain", cl::Hidden, cl::init(false), 98 cl::desc("Stress test LSR IV chains")); 99 #else 100 static bool StressIVChain = false; 101 #endif 102 103 namespace { 104 105 /// RegSortData - This class holds data which is used to order reuse candidates. 106 class RegSortData { 107 public: 108 /// UsedByIndices - This represents the set of LSRUse indices which reference 109 /// a particular register. 110 SmallBitVector UsedByIndices; 111 112 RegSortData() {} 113 114 void print(raw_ostream &OS) const; 115 void dump() const; 116 }; 117 118 } 119 120 void RegSortData::print(raw_ostream &OS) const { 121 OS << "[NumUses=" << UsedByIndices.count() << ']'; 122 } 123 124 void RegSortData::dump() const { 125 print(errs()); errs() << '\n'; 126 } 127 128 namespace { 129 130 /// RegUseTracker - Map register candidates to information about how they are 131 /// used. 132 class RegUseTracker { 133 typedef DenseMap<const SCEV *, RegSortData> RegUsesTy; 134 135 RegUsesTy RegUsesMap; 136 SmallVector<const SCEV *, 16> RegSequence; 137 138 public: 139 void CountRegister(const SCEV *Reg, size_t LUIdx); 140 void DropRegister(const SCEV *Reg, size_t LUIdx); 141 void SwapAndDropUse(size_t LUIdx, size_t LastLUIdx); 142 143 bool isRegUsedByUsesOtherThan(const SCEV *Reg, size_t LUIdx) const; 144 145 const SmallBitVector &getUsedByIndices(const SCEV *Reg) const; 146 147 void clear(); 148 149 typedef SmallVectorImpl<const SCEV *>::iterator iterator; 150 typedef SmallVectorImpl<const SCEV *>::const_iterator const_iterator; 151 iterator begin() { return RegSequence.begin(); } 152 iterator end() { return RegSequence.end(); } 153 const_iterator begin() const { return RegSequence.begin(); } 154 const_iterator end() const { return RegSequence.end(); } 155 }; 156 157 } 158 159 void 160 RegUseTracker::CountRegister(const SCEV *Reg, size_t LUIdx) { 161 std::pair<RegUsesTy::iterator, bool> Pair = 162 RegUsesMap.insert(std::make_pair(Reg, RegSortData())); 163 RegSortData &RSD = Pair.first->second; 164 if (Pair.second) 165 RegSequence.push_back(Reg); 166 RSD.UsedByIndices.resize(std::max(RSD.UsedByIndices.size(), LUIdx + 1)); 167 RSD.UsedByIndices.set(LUIdx); 168 } 169 170 void 171 RegUseTracker::DropRegister(const SCEV *Reg, size_t LUIdx) { 172 RegUsesTy::iterator It = RegUsesMap.find(Reg); 173 assert(It != RegUsesMap.end()); 174 RegSortData &RSD = It->second; 175 assert(RSD.UsedByIndices.size() > LUIdx); 176 RSD.UsedByIndices.reset(LUIdx); 177 } 178 179 void 180 RegUseTracker::SwapAndDropUse(size_t LUIdx, size_t LastLUIdx) { 181 assert(LUIdx <= LastLUIdx); 182 183 // Update RegUses. The data structure is not optimized for this purpose; 184 // we must iterate through it and update each of the bit vectors. 185 for (RegUsesTy::iterator I = RegUsesMap.begin(), E = RegUsesMap.end(); 186 I != E; ++I) { 187 SmallBitVector &UsedByIndices = I->second.UsedByIndices; 188 if (LUIdx < UsedByIndices.size()) 189 UsedByIndices[LUIdx] = 190 LastLUIdx < UsedByIndices.size() ? UsedByIndices[LastLUIdx] : 0; 191 UsedByIndices.resize(std::min(UsedByIndices.size(), LastLUIdx)); 192 } 193 } 194 195 bool 196 RegUseTracker::isRegUsedByUsesOtherThan(const SCEV *Reg, size_t LUIdx) const { 197 RegUsesTy::const_iterator I = RegUsesMap.find(Reg); 198 if (I == RegUsesMap.end()) 199 return false; 200 const SmallBitVector &UsedByIndices = I->second.UsedByIndices; 201 int i = UsedByIndices.find_first(); 202 if (i == -1) return false; 203 if ((size_t)i != LUIdx) return true; 204 return UsedByIndices.find_next(i) != -1; 205 } 206 207 const SmallBitVector &RegUseTracker::getUsedByIndices(const SCEV *Reg) const { 208 RegUsesTy::const_iterator I = RegUsesMap.find(Reg); 209 assert(I != RegUsesMap.end() && "Unknown register!"); 210 return I->second.UsedByIndices; 211 } 212 213 void RegUseTracker::clear() { 214 RegUsesMap.clear(); 215 RegSequence.clear(); 216 } 217 218 namespace { 219 220 /// Formula - This class holds information that describes a formula for 221 /// computing satisfying a use. It may include broken-out immediates and scaled 222 /// registers. 223 struct Formula { 224 /// AM - This is used to represent complex addressing, as well as other kinds 225 /// of interesting uses. 226 TargetLowering::AddrMode AM; 227 228 /// BaseRegs - The list of "base" registers for this use. When this is 229 /// non-empty, AM.HasBaseReg should be set to true. 230 SmallVector<const SCEV *, 2> BaseRegs; 231 232 /// ScaledReg - The 'scaled' register for this use. This should be non-null 233 /// when AM.Scale is not zero. 234 const SCEV *ScaledReg; 235 236 /// UnfoldedOffset - An additional constant offset which added near the 237 /// use. This requires a temporary register, but the offset itself can 238 /// live in an add immediate field rather than a register. 239 int64_t UnfoldedOffset; 240 241 Formula() : ScaledReg(0), UnfoldedOffset(0) {} 242 243 void InitialMatch(const SCEV *S, Loop *L, ScalarEvolution &SE); 244 245 unsigned getNumRegs() const; 246 Type *getType() const; 247 248 void DeleteBaseReg(const SCEV *&S); 249 250 bool referencesReg(const SCEV *S) const; 251 bool hasRegsUsedByUsesOtherThan(size_t LUIdx, 252 const RegUseTracker &RegUses) const; 253 254 void print(raw_ostream &OS) const; 255 void dump() const; 256 }; 257 258 } 259 260 /// DoInitialMatch - Recursion helper for InitialMatch. 261 static void DoInitialMatch(const SCEV *S, Loop *L, 262 SmallVectorImpl<const SCEV *> &Good, 263 SmallVectorImpl<const SCEV *> &Bad, 264 ScalarEvolution &SE) { 265 // Collect expressions which properly dominate the loop header. 266 if (SE.properlyDominates(S, L->getHeader())) { 267 Good.push_back(S); 268 return; 269 } 270 271 // Look at add operands. 272 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) { 273 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end(); 274 I != E; ++I) 275 DoInitialMatch(*I, L, Good, Bad, SE); 276 return; 277 } 278 279 // Look at addrec operands. 280 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) 281 if (!AR->getStart()->isZero()) { 282 DoInitialMatch(AR->getStart(), L, Good, Bad, SE); 283 DoInitialMatch(SE.getAddRecExpr(SE.getConstant(AR->getType(), 0), 284 AR->getStepRecurrence(SE), 285 // FIXME: AR->getNoWrapFlags() 286 AR->getLoop(), SCEV::FlagAnyWrap), 287 L, Good, Bad, SE); 288 return; 289 } 290 291 // Handle a multiplication by -1 (negation) if it didn't fold. 292 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) 293 if (Mul->getOperand(0)->isAllOnesValue()) { 294 SmallVector<const SCEV *, 4> Ops(Mul->op_begin()+1, Mul->op_end()); 295 const SCEV *NewMul = SE.getMulExpr(Ops); 296 297 SmallVector<const SCEV *, 4> MyGood; 298 SmallVector<const SCEV *, 4> MyBad; 299 DoInitialMatch(NewMul, L, MyGood, MyBad, SE); 300 const SCEV *NegOne = SE.getSCEV(ConstantInt::getAllOnesValue( 301 SE.getEffectiveSCEVType(NewMul->getType()))); 302 for (SmallVectorImpl<const SCEV *>::const_iterator I = MyGood.begin(), 303 E = MyGood.end(); I != E; ++I) 304 Good.push_back(SE.getMulExpr(NegOne, *I)); 305 for (SmallVectorImpl<const SCEV *>::const_iterator I = MyBad.begin(), 306 E = MyBad.end(); I != E; ++I) 307 Bad.push_back(SE.getMulExpr(NegOne, *I)); 308 return; 309 } 310 311 // Ok, we can't do anything interesting. Just stuff the whole thing into a 312 // register and hope for the best. 313 Bad.push_back(S); 314 } 315 316 /// InitialMatch - Incorporate loop-variant parts of S into this Formula, 317 /// attempting to keep all loop-invariant and loop-computable values in a 318 /// single base register. 319 void Formula::InitialMatch(const SCEV *S, Loop *L, ScalarEvolution &SE) { 320 SmallVector<const SCEV *, 4> Good; 321 SmallVector<const SCEV *, 4> Bad; 322 DoInitialMatch(S, L, Good, Bad, SE); 323 if (!Good.empty()) { 324 const SCEV *Sum = SE.getAddExpr(Good); 325 if (!Sum->isZero()) 326 BaseRegs.push_back(Sum); 327 AM.HasBaseReg = true; 328 } 329 if (!Bad.empty()) { 330 const SCEV *Sum = SE.getAddExpr(Bad); 331 if (!Sum->isZero()) 332 BaseRegs.push_back(Sum); 333 AM.HasBaseReg = true; 334 } 335 } 336 337 /// getNumRegs - Return the total number of register operands used by this 338 /// formula. This does not include register uses implied by non-constant 339 /// addrec strides. 340 unsigned Formula::getNumRegs() const { 341 return !!ScaledReg + BaseRegs.size(); 342 } 343 344 /// getType - Return the type of this formula, if it has one, or null 345 /// otherwise. This type is meaningless except for the bit size. 346 Type *Formula::getType() const { 347 return !BaseRegs.empty() ? BaseRegs.front()->getType() : 348 ScaledReg ? ScaledReg->getType() : 349 AM.BaseGV ? AM.BaseGV->getType() : 350 0; 351 } 352 353 /// DeleteBaseReg - Delete the given base reg from the BaseRegs list. 354 void Formula::DeleteBaseReg(const SCEV *&S) { 355 if (&S != &BaseRegs.back()) 356 std::swap(S, BaseRegs.back()); 357 BaseRegs.pop_back(); 358 } 359 360 /// referencesReg - Test if this formula references the given register. 361 bool Formula::referencesReg(const SCEV *S) const { 362 return S == ScaledReg || 363 std::find(BaseRegs.begin(), BaseRegs.end(), S) != BaseRegs.end(); 364 } 365 366 /// hasRegsUsedByUsesOtherThan - Test whether this formula uses registers 367 /// which are used by uses other than the use with the given index. 368 bool Formula::hasRegsUsedByUsesOtherThan(size_t LUIdx, 369 const RegUseTracker &RegUses) const { 370 if (ScaledReg) 371 if (RegUses.isRegUsedByUsesOtherThan(ScaledReg, LUIdx)) 372 return true; 373 for (SmallVectorImpl<const SCEV *>::const_iterator I = BaseRegs.begin(), 374 E = BaseRegs.end(); I != E; ++I) 375 if (RegUses.isRegUsedByUsesOtherThan(*I, LUIdx)) 376 return true; 377 return false; 378 } 379 380 void Formula::print(raw_ostream &OS) const { 381 bool First = true; 382 if (AM.BaseGV) { 383 if (!First) OS << " + "; else First = false; 384 WriteAsOperand(OS, AM.BaseGV, /*PrintType=*/false); 385 } 386 if (AM.BaseOffs != 0) { 387 if (!First) OS << " + "; else First = false; 388 OS << AM.BaseOffs; 389 } 390 for (SmallVectorImpl<const SCEV *>::const_iterator I = BaseRegs.begin(), 391 E = BaseRegs.end(); I != E; ++I) { 392 if (!First) OS << " + "; else First = false; 393 OS << "reg(" << **I << ')'; 394 } 395 if (AM.HasBaseReg && BaseRegs.empty()) { 396 if (!First) OS << " + "; else First = false; 397 OS << "**error: HasBaseReg**"; 398 } else if (!AM.HasBaseReg && !BaseRegs.empty()) { 399 if (!First) OS << " + "; else First = false; 400 OS << "**error: !HasBaseReg**"; 401 } 402 if (AM.Scale != 0) { 403 if (!First) OS << " + "; else First = false; 404 OS << AM.Scale << "*reg("; 405 if (ScaledReg) 406 OS << *ScaledReg; 407 else 408 OS << "<unknown>"; 409 OS << ')'; 410 } 411 if (UnfoldedOffset != 0) { 412 if (!First) OS << " + "; else First = false; 413 OS << "imm(" << UnfoldedOffset << ')'; 414 } 415 } 416 417 void Formula::dump() const { 418 print(errs()); errs() << '\n'; 419 } 420 421 /// isAddRecSExtable - Return true if the given addrec can be sign-extended 422 /// without changing its value. 423 static bool isAddRecSExtable(const SCEVAddRecExpr *AR, ScalarEvolution &SE) { 424 Type *WideTy = 425 IntegerType::get(SE.getContext(), SE.getTypeSizeInBits(AR->getType()) + 1); 426 return isa<SCEVAddRecExpr>(SE.getSignExtendExpr(AR, WideTy)); 427 } 428 429 /// isAddSExtable - Return true if the given add can be sign-extended 430 /// without changing its value. 431 static bool isAddSExtable(const SCEVAddExpr *A, ScalarEvolution &SE) { 432 Type *WideTy = 433 IntegerType::get(SE.getContext(), SE.getTypeSizeInBits(A->getType()) + 1); 434 return isa<SCEVAddExpr>(SE.getSignExtendExpr(A, WideTy)); 435 } 436 437 /// isMulSExtable - Return true if the given mul can be sign-extended 438 /// without changing its value. 439 static bool isMulSExtable(const SCEVMulExpr *M, ScalarEvolution &SE) { 440 Type *WideTy = 441 IntegerType::get(SE.getContext(), 442 SE.getTypeSizeInBits(M->getType()) * M->getNumOperands()); 443 return isa<SCEVMulExpr>(SE.getSignExtendExpr(M, WideTy)); 444 } 445 446 /// getExactSDiv - Return an expression for LHS /s RHS, if it can be determined 447 /// and if the remainder is known to be zero, or null otherwise. If 448 /// IgnoreSignificantBits is true, expressions like (X * Y) /s Y are simplified 449 /// to Y, ignoring that the multiplication may overflow, which is useful when 450 /// the result will be used in a context where the most significant bits are 451 /// ignored. 452 static const SCEV *getExactSDiv(const SCEV *LHS, const SCEV *RHS, 453 ScalarEvolution &SE, 454 bool IgnoreSignificantBits = false) { 455 // Handle the trivial case, which works for any SCEV type. 456 if (LHS == RHS) 457 return SE.getConstant(LHS->getType(), 1); 458 459 // Handle a few RHS special cases. 460 const SCEVConstant *RC = dyn_cast<SCEVConstant>(RHS); 461 if (RC) { 462 const APInt &RA = RC->getValue()->getValue(); 463 // Handle x /s -1 as x * -1, to give ScalarEvolution a chance to do 464 // some folding. 465 if (RA.isAllOnesValue()) 466 return SE.getMulExpr(LHS, RC); 467 // Handle x /s 1 as x. 468 if (RA == 1) 469 return LHS; 470 } 471 472 // Check for a division of a constant by a constant. 473 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(LHS)) { 474 if (!RC) 475 return 0; 476 const APInt &LA = C->getValue()->getValue(); 477 const APInt &RA = RC->getValue()->getValue(); 478 if (LA.srem(RA) != 0) 479 return 0; 480 return SE.getConstant(LA.sdiv(RA)); 481 } 482 483 // Distribute the sdiv over addrec operands, if the addrec doesn't overflow. 484 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(LHS)) { 485 if (IgnoreSignificantBits || isAddRecSExtable(AR, SE)) { 486 const SCEV *Step = getExactSDiv(AR->getStepRecurrence(SE), RHS, SE, 487 IgnoreSignificantBits); 488 if (!Step) return 0; 489 const SCEV *Start = getExactSDiv(AR->getStart(), RHS, SE, 490 IgnoreSignificantBits); 491 if (!Start) return 0; 492 // FlagNW is independent of the start value, step direction, and is 493 // preserved with smaller magnitude steps. 494 // FIXME: AR->getNoWrapFlags(SCEV::FlagNW) 495 return SE.getAddRecExpr(Start, Step, AR->getLoop(), SCEV::FlagAnyWrap); 496 } 497 return 0; 498 } 499 500 // Distribute the sdiv over add operands, if the add doesn't overflow. 501 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(LHS)) { 502 if (IgnoreSignificantBits || isAddSExtable(Add, SE)) { 503 SmallVector<const SCEV *, 8> Ops; 504 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end(); 505 I != E; ++I) { 506 const SCEV *Op = getExactSDiv(*I, RHS, SE, 507 IgnoreSignificantBits); 508 if (!Op) return 0; 509 Ops.push_back(Op); 510 } 511 return SE.getAddExpr(Ops); 512 } 513 return 0; 514 } 515 516 // Check for a multiply operand that we can pull RHS out of. 517 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(LHS)) { 518 if (IgnoreSignificantBits || isMulSExtable(Mul, SE)) { 519 SmallVector<const SCEV *, 4> Ops; 520 bool Found = false; 521 for (SCEVMulExpr::op_iterator I = Mul->op_begin(), E = Mul->op_end(); 522 I != E; ++I) { 523 const SCEV *S = *I; 524 if (!Found) 525 if (const SCEV *Q = getExactSDiv(S, RHS, SE, 526 IgnoreSignificantBits)) { 527 S = Q; 528 Found = true; 529 } 530 Ops.push_back(S); 531 } 532 return Found ? SE.getMulExpr(Ops) : 0; 533 } 534 return 0; 535 } 536 537 // Otherwise we don't know. 538 return 0; 539 } 540 541 /// ExtractImmediate - If S involves the addition of a constant integer value, 542 /// return that integer value, and mutate S to point to a new SCEV with that 543 /// value excluded. 544 static int64_t ExtractImmediate(const SCEV *&S, ScalarEvolution &SE) { 545 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S)) { 546 if (C->getValue()->getValue().getMinSignedBits() <= 64) { 547 S = SE.getConstant(C->getType(), 0); 548 return C->getValue()->getSExtValue(); 549 } 550 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) { 551 SmallVector<const SCEV *, 8> NewOps(Add->op_begin(), Add->op_end()); 552 int64_t Result = ExtractImmediate(NewOps.front(), SE); 553 if (Result != 0) 554 S = SE.getAddExpr(NewOps); 555 return Result; 556 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) { 557 SmallVector<const SCEV *, 8> NewOps(AR->op_begin(), AR->op_end()); 558 int64_t Result = ExtractImmediate(NewOps.front(), SE); 559 if (Result != 0) 560 S = SE.getAddRecExpr(NewOps, AR->getLoop(), 561 // FIXME: AR->getNoWrapFlags(SCEV::FlagNW) 562 SCEV::FlagAnyWrap); 563 return Result; 564 } 565 return 0; 566 } 567 568 /// ExtractSymbol - If S involves the addition of a GlobalValue address, 569 /// return that symbol, and mutate S to point to a new SCEV with that 570 /// value excluded. 571 static GlobalValue *ExtractSymbol(const SCEV *&S, ScalarEvolution &SE) { 572 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) { 573 if (GlobalValue *GV = dyn_cast<GlobalValue>(U->getValue())) { 574 S = SE.getConstant(GV->getType(), 0); 575 return GV; 576 } 577 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) { 578 SmallVector<const SCEV *, 8> NewOps(Add->op_begin(), Add->op_end()); 579 GlobalValue *Result = ExtractSymbol(NewOps.back(), SE); 580 if (Result) 581 S = SE.getAddExpr(NewOps); 582 return Result; 583 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) { 584 SmallVector<const SCEV *, 8> NewOps(AR->op_begin(), AR->op_end()); 585 GlobalValue *Result = ExtractSymbol(NewOps.front(), SE); 586 if (Result) 587 S = SE.getAddRecExpr(NewOps, AR->getLoop(), 588 // FIXME: AR->getNoWrapFlags(SCEV::FlagNW) 589 SCEV::FlagAnyWrap); 590 return Result; 591 } 592 return 0; 593 } 594 595 /// isAddressUse - Returns true if the specified instruction is using the 596 /// specified value as an address. 597 static bool isAddressUse(Instruction *Inst, Value *OperandVal) { 598 bool isAddress = isa<LoadInst>(Inst); 599 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) { 600 if (SI->getOperand(1) == OperandVal) 601 isAddress = true; 602 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) { 603 // Addressing modes can also be folded into prefetches and a variety 604 // of intrinsics. 605 switch (II->getIntrinsicID()) { 606 default: break; 607 case Intrinsic::prefetch: 608 case Intrinsic::x86_sse_storeu_ps: 609 case Intrinsic::x86_sse2_storeu_pd: 610 case Intrinsic::x86_sse2_storeu_dq: 611 case Intrinsic::x86_sse2_storel_dq: 612 if (II->getArgOperand(0) == OperandVal) 613 isAddress = true; 614 break; 615 } 616 } 617 return isAddress; 618 } 619 620 /// getAccessType - Return the type of the memory being accessed. 621 static Type *getAccessType(const Instruction *Inst) { 622 Type *AccessTy = Inst->getType(); 623 if (const StoreInst *SI = dyn_cast<StoreInst>(Inst)) 624 AccessTy = SI->getOperand(0)->getType(); 625 else if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) { 626 // Addressing modes can also be folded into prefetches and a variety 627 // of intrinsics. 628 switch (II->getIntrinsicID()) { 629 default: break; 630 case Intrinsic::x86_sse_storeu_ps: 631 case Intrinsic::x86_sse2_storeu_pd: 632 case Intrinsic::x86_sse2_storeu_dq: 633 case Intrinsic::x86_sse2_storel_dq: 634 AccessTy = II->getArgOperand(0)->getType(); 635 break; 636 } 637 } 638 639 // All pointers have the same requirements, so canonicalize them to an 640 // arbitrary pointer type to minimize variation. 641 if (PointerType *PTy = dyn_cast<PointerType>(AccessTy)) 642 AccessTy = PointerType::get(IntegerType::get(PTy->getContext(), 1), 643 PTy->getAddressSpace()); 644 645 return AccessTy; 646 } 647 648 /// isExistingPhi - Return true if this AddRec is already a phi in its loop. 649 static bool isExistingPhi(const SCEVAddRecExpr *AR, ScalarEvolution &SE) { 650 for (BasicBlock::iterator I = AR->getLoop()->getHeader()->begin(); 651 PHINode *PN = dyn_cast<PHINode>(I); ++I) { 652 if (SE.isSCEVable(PN->getType()) && 653 (SE.getEffectiveSCEVType(PN->getType()) == 654 SE.getEffectiveSCEVType(AR->getType())) && 655 SE.getSCEV(PN) == AR) 656 return true; 657 } 658 return false; 659 } 660 661 /// Check if expanding this expression is likely to incur significant cost. This 662 /// is tricky because SCEV doesn't track which expressions are actually computed 663 /// by the current IR. 664 /// 665 /// We currently allow expansion of IV increments that involve adds, 666 /// multiplication by constants, and AddRecs from existing phis. 667 /// 668 /// TODO: Allow UDivExpr if we can find an existing IV increment that is an 669 /// obvious multiple of the UDivExpr. 670 static bool isHighCostExpansion(const SCEV *S, 671 SmallPtrSet<const SCEV*, 8> &Processed, 672 ScalarEvolution &SE) { 673 // Zero/One operand expressions 674 switch (S->getSCEVType()) { 675 case scUnknown: 676 case scConstant: 677 return false; 678 case scTruncate: 679 return isHighCostExpansion(cast<SCEVTruncateExpr>(S)->getOperand(), 680 Processed, SE); 681 case scZeroExtend: 682 return isHighCostExpansion(cast<SCEVZeroExtendExpr>(S)->getOperand(), 683 Processed, SE); 684 case scSignExtend: 685 return isHighCostExpansion(cast<SCEVSignExtendExpr>(S)->getOperand(), 686 Processed, SE); 687 } 688 689 if (!Processed.insert(S)) 690 return false; 691 692 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) { 693 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end(); 694 I != E; ++I) { 695 if (isHighCostExpansion(*I, Processed, SE)) 696 return true; 697 } 698 return false; 699 } 700 701 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) { 702 if (Mul->getNumOperands() == 2) { 703 // Multiplication by a constant is ok 704 if (isa<SCEVConstant>(Mul->getOperand(0))) 705 return isHighCostExpansion(Mul->getOperand(1), Processed, SE); 706 707 // If we have the value of one operand, check if an existing 708 // multiplication already generates this expression. 709 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(Mul->getOperand(1))) { 710 Value *UVal = U->getValue(); 711 for (Value::use_iterator UI = UVal->use_begin(), UE = UVal->use_end(); 712 UI != UE; ++UI) { 713 // If U is a constant, it may be used by a ConstantExpr. 714 Instruction *User = dyn_cast<Instruction>(*UI); 715 if (User && User->getOpcode() == Instruction::Mul 716 && SE.isSCEVable(User->getType())) { 717 return SE.getSCEV(User) == Mul; 718 } 719 } 720 } 721 } 722 } 723 724 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) { 725 if (isExistingPhi(AR, SE)) 726 return false; 727 } 728 729 // Fow now, consider any other type of expression (div/mul/min/max) high cost. 730 return true; 731 } 732 733 /// DeleteTriviallyDeadInstructions - If any of the instructions is the 734 /// specified set are trivially dead, delete them and see if this makes any of 735 /// their operands subsequently dead. 736 static bool 737 DeleteTriviallyDeadInstructions(SmallVectorImpl<WeakVH> &DeadInsts) { 738 bool Changed = false; 739 740 while (!DeadInsts.empty()) { 741 Instruction *I = dyn_cast_or_null<Instruction>(&*DeadInsts.pop_back_val()); 742 743 if (I == 0 || !isInstructionTriviallyDead(I)) 744 continue; 745 746 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI) 747 if (Instruction *U = dyn_cast<Instruction>(*OI)) { 748 *OI = 0; 749 if (U->use_empty()) 750 DeadInsts.push_back(U); 751 } 752 753 I->eraseFromParent(); 754 Changed = true; 755 } 756 757 return Changed; 758 } 759 760 namespace { 761 762 /// Cost - This class is used to measure and compare candidate formulae. 763 class Cost { 764 /// TODO: Some of these could be merged. Also, a lexical ordering 765 /// isn't always optimal. 766 unsigned NumRegs; 767 unsigned AddRecCost; 768 unsigned NumIVMuls; 769 unsigned NumBaseAdds; 770 unsigned ImmCost; 771 unsigned SetupCost; 772 773 public: 774 Cost() 775 : NumRegs(0), AddRecCost(0), NumIVMuls(0), NumBaseAdds(0), ImmCost(0), 776 SetupCost(0) {} 777 778 bool operator<(const Cost &Other) const; 779 780 void Loose(); 781 782 #ifndef NDEBUG 783 // Once any of the metrics loses, they must all remain losers. 784 bool isValid() { 785 return ((NumRegs | AddRecCost | NumIVMuls | NumBaseAdds 786 | ImmCost | SetupCost) != ~0u) 787 || ((NumRegs & AddRecCost & NumIVMuls & NumBaseAdds 788 & ImmCost & SetupCost) == ~0u); 789 } 790 #endif 791 792 bool isLoser() { 793 assert(isValid() && "invalid cost"); 794 return NumRegs == ~0u; 795 } 796 797 void RateFormula(const Formula &F, 798 SmallPtrSet<const SCEV *, 16> &Regs, 799 const DenseSet<const SCEV *> &VisitedRegs, 800 const Loop *L, 801 const SmallVectorImpl<int64_t> &Offsets, 802 ScalarEvolution &SE, DominatorTree &DT, 803 SmallPtrSet<const SCEV *, 16> *LoserRegs = 0); 804 805 void print(raw_ostream &OS) const; 806 void dump() const; 807 808 private: 809 void RateRegister(const SCEV *Reg, 810 SmallPtrSet<const SCEV *, 16> &Regs, 811 const Loop *L, 812 ScalarEvolution &SE, DominatorTree &DT); 813 void RatePrimaryRegister(const SCEV *Reg, 814 SmallPtrSet<const SCEV *, 16> &Regs, 815 const Loop *L, 816 ScalarEvolution &SE, DominatorTree &DT, 817 SmallPtrSet<const SCEV *, 16> *LoserRegs); 818 }; 819 820 } 821 822 /// RateRegister - Tally up interesting quantities from the given register. 823 void Cost::RateRegister(const SCEV *Reg, 824 SmallPtrSet<const SCEV *, 16> &Regs, 825 const Loop *L, 826 ScalarEvolution &SE, DominatorTree &DT) { 827 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Reg)) { 828 // If this is an addrec for another loop, don't second-guess its addrec phi 829 // nodes. LSR isn't currently smart enough to reason about more than one 830 // loop at a time. LSR has already run on inner loops, will not run on outer 831 // loops, and cannot be expected to change sibling loops. 832 if (AR->getLoop() != L) { 833 // If the AddRec exists, consider it's register free and leave it alone. 834 if (isExistingPhi(AR, SE)) 835 return; 836 837 // Otherwise, do not consider this formula at all. 838 Loose(); 839 return; 840 } 841 AddRecCost += 1; /// TODO: This should be a function of the stride. 842 843 // Add the step value register, if it needs one. 844 // TODO: The non-affine case isn't precisely modeled here. 845 if (!AR->isAffine() || !isa<SCEVConstant>(AR->getOperand(1))) { 846 if (!Regs.count(AR->getOperand(1))) { 847 RateRegister(AR->getOperand(1), Regs, L, SE, DT); 848 if (isLoser()) 849 return; 850 } 851 } 852 } 853 ++NumRegs; 854 855 // Rough heuristic; favor registers which don't require extra setup 856 // instructions in the preheader. 857 if (!isa<SCEVUnknown>(Reg) && 858 !isa<SCEVConstant>(Reg) && 859 !(isa<SCEVAddRecExpr>(Reg) && 860 (isa<SCEVUnknown>(cast<SCEVAddRecExpr>(Reg)->getStart()) || 861 isa<SCEVConstant>(cast<SCEVAddRecExpr>(Reg)->getStart())))) 862 ++SetupCost; 863 864 NumIVMuls += isa<SCEVMulExpr>(Reg) && 865 SE.hasComputableLoopEvolution(Reg, L); 866 } 867 868 /// RatePrimaryRegister - Record this register in the set. If we haven't seen it 869 /// before, rate it. Optional LoserRegs provides a way to declare any formula 870 /// that refers to one of those regs an instant loser. 871 void Cost::RatePrimaryRegister(const SCEV *Reg, 872 SmallPtrSet<const SCEV *, 16> &Regs, 873 const Loop *L, 874 ScalarEvolution &SE, DominatorTree &DT, 875 SmallPtrSet<const SCEV *, 16> *LoserRegs) { 876 if (LoserRegs && LoserRegs->count(Reg)) { 877 Loose(); 878 return; 879 } 880 if (Regs.insert(Reg)) { 881 RateRegister(Reg, Regs, L, SE, DT); 882 if (isLoser()) 883 LoserRegs->insert(Reg); 884 } 885 } 886 887 void Cost::RateFormula(const Formula &F, 888 SmallPtrSet<const SCEV *, 16> &Regs, 889 const DenseSet<const SCEV *> &VisitedRegs, 890 const Loop *L, 891 const SmallVectorImpl<int64_t> &Offsets, 892 ScalarEvolution &SE, DominatorTree &DT, 893 SmallPtrSet<const SCEV *, 16> *LoserRegs) { 894 // Tally up the registers. 895 if (const SCEV *ScaledReg = F.ScaledReg) { 896 if (VisitedRegs.count(ScaledReg)) { 897 Loose(); 898 return; 899 } 900 RatePrimaryRegister(ScaledReg, Regs, L, SE, DT, LoserRegs); 901 if (isLoser()) 902 return; 903 } 904 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(), 905 E = F.BaseRegs.end(); I != E; ++I) { 906 const SCEV *BaseReg = *I; 907 if (VisitedRegs.count(BaseReg)) { 908 Loose(); 909 return; 910 } 911 RatePrimaryRegister(BaseReg, Regs, L, SE, DT, LoserRegs); 912 if (isLoser()) 913 return; 914 } 915 916 // Determine how many (unfolded) adds we'll need inside the loop. 917 size_t NumBaseParts = F.BaseRegs.size() + (F.UnfoldedOffset != 0); 918 if (NumBaseParts > 1) 919 NumBaseAdds += NumBaseParts - 1; 920 921 // Tally up the non-zero immediates. 922 for (SmallVectorImpl<int64_t>::const_iterator I = Offsets.begin(), 923 E = Offsets.end(); I != E; ++I) { 924 int64_t Offset = (uint64_t)*I + F.AM.BaseOffs; 925 if (F.AM.BaseGV) 926 ImmCost += 64; // Handle symbolic values conservatively. 927 // TODO: This should probably be the pointer size. 928 else if (Offset != 0) 929 ImmCost += APInt(64, Offset, true).getMinSignedBits(); 930 } 931 assert(isValid() && "invalid cost"); 932 } 933 934 /// Loose - Set this cost to a losing value. 935 void Cost::Loose() { 936 NumRegs = ~0u; 937 AddRecCost = ~0u; 938 NumIVMuls = ~0u; 939 NumBaseAdds = ~0u; 940 ImmCost = ~0u; 941 SetupCost = ~0u; 942 } 943 944 /// operator< - Choose the lower cost. 945 bool Cost::operator<(const Cost &Other) const { 946 if (NumRegs != Other.NumRegs) 947 return NumRegs < Other.NumRegs; 948 if (AddRecCost != Other.AddRecCost) 949 return AddRecCost < Other.AddRecCost; 950 if (NumIVMuls != Other.NumIVMuls) 951 return NumIVMuls < Other.NumIVMuls; 952 if (NumBaseAdds != Other.NumBaseAdds) 953 return NumBaseAdds < Other.NumBaseAdds; 954 if (ImmCost != Other.ImmCost) 955 return ImmCost < Other.ImmCost; 956 if (SetupCost != Other.SetupCost) 957 return SetupCost < Other.SetupCost; 958 return false; 959 } 960 961 void Cost::print(raw_ostream &OS) const { 962 OS << NumRegs << " reg" << (NumRegs == 1 ? "" : "s"); 963 if (AddRecCost != 0) 964 OS << ", with addrec cost " << AddRecCost; 965 if (NumIVMuls != 0) 966 OS << ", plus " << NumIVMuls << " IV mul" << (NumIVMuls == 1 ? "" : "s"); 967 if (NumBaseAdds != 0) 968 OS << ", plus " << NumBaseAdds << " base add" 969 << (NumBaseAdds == 1 ? "" : "s"); 970 if (ImmCost != 0) 971 OS << ", plus " << ImmCost << " imm cost"; 972 if (SetupCost != 0) 973 OS << ", plus " << SetupCost << " setup cost"; 974 } 975 976 void Cost::dump() const { 977 print(errs()); errs() << '\n'; 978 } 979 980 namespace { 981 982 /// LSRFixup - An operand value in an instruction which is to be replaced 983 /// with some equivalent, possibly strength-reduced, replacement. 984 struct LSRFixup { 985 /// UserInst - The instruction which will be updated. 986 Instruction *UserInst; 987 988 /// OperandValToReplace - The operand of the instruction which will 989 /// be replaced. The operand may be used more than once; every instance 990 /// will be replaced. 991 Value *OperandValToReplace; 992 993 /// PostIncLoops - If this user is to use the post-incremented value of an 994 /// induction variable, this variable is non-null and holds the loop 995 /// associated with the induction variable. 996 PostIncLoopSet PostIncLoops; 997 998 /// LUIdx - The index of the LSRUse describing the expression which 999 /// this fixup needs, minus an offset (below). 1000 size_t LUIdx; 1001 1002 /// Offset - A constant offset to be added to the LSRUse expression. 1003 /// This allows multiple fixups to share the same LSRUse with different 1004 /// offsets, for example in an unrolled loop. 1005 int64_t Offset; 1006 1007 bool isUseFullyOutsideLoop(const Loop *L) const; 1008 1009 LSRFixup(); 1010 1011 void print(raw_ostream &OS) const; 1012 void dump() const; 1013 }; 1014 1015 } 1016 1017 LSRFixup::LSRFixup() 1018 : UserInst(0), OperandValToReplace(0), LUIdx(~size_t(0)), Offset(0) {} 1019 1020 /// isUseFullyOutsideLoop - Test whether this fixup always uses its 1021 /// value outside of the given loop. 1022 bool LSRFixup::isUseFullyOutsideLoop(const Loop *L) const { 1023 // PHI nodes use their value in their incoming blocks. 1024 if (const PHINode *PN = dyn_cast<PHINode>(UserInst)) { 1025 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) 1026 if (PN->getIncomingValue(i) == OperandValToReplace && 1027 L->contains(PN->getIncomingBlock(i))) 1028 return false; 1029 return true; 1030 } 1031 1032 return !L->contains(UserInst); 1033 } 1034 1035 void LSRFixup::print(raw_ostream &OS) const { 1036 OS << "UserInst="; 1037 // Store is common and interesting enough to be worth special-casing. 1038 if (StoreInst *Store = dyn_cast<StoreInst>(UserInst)) { 1039 OS << "store "; 1040 WriteAsOperand(OS, Store->getOperand(0), /*PrintType=*/false); 1041 } else if (UserInst->getType()->isVoidTy()) 1042 OS << UserInst->getOpcodeName(); 1043 else 1044 WriteAsOperand(OS, UserInst, /*PrintType=*/false); 1045 1046 OS << ", OperandValToReplace="; 1047 WriteAsOperand(OS, OperandValToReplace, /*PrintType=*/false); 1048 1049 for (PostIncLoopSet::const_iterator I = PostIncLoops.begin(), 1050 E = PostIncLoops.end(); I != E; ++I) { 1051 OS << ", PostIncLoop="; 1052 WriteAsOperand(OS, (*I)->getHeader(), /*PrintType=*/false); 1053 } 1054 1055 if (LUIdx != ~size_t(0)) 1056 OS << ", LUIdx=" << LUIdx; 1057 1058 if (Offset != 0) 1059 OS << ", Offset=" << Offset; 1060 } 1061 1062 void LSRFixup::dump() const { 1063 print(errs()); errs() << '\n'; 1064 } 1065 1066 namespace { 1067 1068 /// UniquifierDenseMapInfo - A DenseMapInfo implementation for holding 1069 /// DenseMaps and DenseSets of sorted SmallVectors of const SCEV*. 1070 struct UniquifierDenseMapInfo { 1071 static SmallVector<const SCEV *, 2> getEmptyKey() { 1072 SmallVector<const SCEV *, 2> V; 1073 V.push_back(reinterpret_cast<const SCEV *>(-1)); 1074 return V; 1075 } 1076 1077 static SmallVector<const SCEV *, 2> getTombstoneKey() { 1078 SmallVector<const SCEV *, 2> V; 1079 V.push_back(reinterpret_cast<const SCEV *>(-2)); 1080 return V; 1081 } 1082 1083 static unsigned getHashValue(const SmallVector<const SCEV *, 2> &V) { 1084 unsigned Result = 0; 1085 for (SmallVectorImpl<const SCEV *>::const_iterator I = V.begin(), 1086 E = V.end(); I != E; ++I) 1087 Result ^= DenseMapInfo<const SCEV *>::getHashValue(*I); 1088 return Result; 1089 } 1090 1091 static bool isEqual(const SmallVector<const SCEV *, 2> &LHS, 1092 const SmallVector<const SCEV *, 2> &RHS) { 1093 return LHS == RHS; 1094 } 1095 }; 1096 1097 /// LSRUse - This class holds the state that LSR keeps for each use in 1098 /// IVUsers, as well as uses invented by LSR itself. It includes information 1099 /// about what kinds of things can be folded into the user, information about 1100 /// the user itself, and information about how the use may be satisfied. 1101 /// TODO: Represent multiple users of the same expression in common? 1102 class LSRUse { 1103 DenseSet<SmallVector<const SCEV *, 2>, UniquifierDenseMapInfo> Uniquifier; 1104 1105 public: 1106 /// KindType - An enum for a kind of use, indicating what types of 1107 /// scaled and immediate operands it might support. 1108 enum KindType { 1109 Basic, ///< A normal use, with no folding. 1110 Special, ///< A special case of basic, allowing -1 scales. 1111 Address, ///< An address use; folding according to TargetLowering 1112 ICmpZero ///< An equality icmp with both operands folded into one. 1113 // TODO: Add a generic icmp too? 1114 }; 1115 1116 KindType Kind; 1117 Type *AccessTy; 1118 1119 SmallVector<int64_t, 8> Offsets; 1120 int64_t MinOffset; 1121 int64_t MaxOffset; 1122 1123 /// AllFixupsOutsideLoop - This records whether all of the fixups using this 1124 /// LSRUse are outside of the loop, in which case some special-case heuristics 1125 /// may be used. 1126 bool AllFixupsOutsideLoop; 1127 1128 /// WidestFixupType - This records the widest use type for any fixup using 1129 /// this LSRUse. FindUseWithSimilarFormula can't consider uses with different 1130 /// max fixup widths to be equivalent, because the narrower one may be relying 1131 /// on the implicit truncation to truncate away bogus bits. 1132 Type *WidestFixupType; 1133 1134 /// Formulae - A list of ways to build a value that can satisfy this user. 1135 /// After the list is populated, one of these is selected heuristically and 1136 /// used to formulate a replacement for OperandValToReplace in UserInst. 1137 SmallVector<Formula, 12> Formulae; 1138 1139 /// Regs - The set of register candidates used by all formulae in this LSRUse. 1140 SmallPtrSet<const SCEV *, 4> Regs; 1141 1142 LSRUse(KindType K, Type *T) : Kind(K), AccessTy(T), 1143 MinOffset(INT64_MAX), 1144 MaxOffset(INT64_MIN), 1145 AllFixupsOutsideLoop(true), 1146 WidestFixupType(0) {} 1147 1148 bool HasFormulaWithSameRegs(const Formula &F) const; 1149 bool InsertFormula(const Formula &F); 1150 void DeleteFormula(Formula &F); 1151 void RecomputeRegs(size_t LUIdx, RegUseTracker &Reguses); 1152 1153 void print(raw_ostream &OS) const; 1154 void dump() const; 1155 }; 1156 1157 } 1158 1159 /// HasFormula - Test whether this use as a formula which has the same 1160 /// registers as the given formula. 1161 bool LSRUse::HasFormulaWithSameRegs(const Formula &F) const { 1162 SmallVector<const SCEV *, 2> Key = F.BaseRegs; 1163 if (F.ScaledReg) Key.push_back(F.ScaledReg); 1164 // Unstable sort by host order ok, because this is only used for uniquifying. 1165 std::sort(Key.begin(), Key.end()); 1166 return Uniquifier.count(Key); 1167 } 1168 1169 /// InsertFormula - If the given formula has not yet been inserted, add it to 1170 /// the list, and return true. Return false otherwise. 1171 bool LSRUse::InsertFormula(const Formula &F) { 1172 SmallVector<const SCEV *, 2> Key = F.BaseRegs; 1173 if (F.ScaledReg) Key.push_back(F.ScaledReg); 1174 // Unstable sort by host order ok, because this is only used for uniquifying. 1175 std::sort(Key.begin(), Key.end()); 1176 1177 if (!Uniquifier.insert(Key).second) 1178 return false; 1179 1180 // Using a register to hold the value of 0 is not profitable. 1181 assert((!F.ScaledReg || !F.ScaledReg->isZero()) && 1182 "Zero allocated in a scaled register!"); 1183 #ifndef NDEBUG 1184 for (SmallVectorImpl<const SCEV *>::const_iterator I = 1185 F.BaseRegs.begin(), E = F.BaseRegs.end(); I != E; ++I) 1186 assert(!(*I)->isZero() && "Zero allocated in a base register!"); 1187 #endif 1188 1189 // Add the formula to the list. 1190 Formulae.push_back(F); 1191 1192 // Record registers now being used by this use. 1193 Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end()); 1194 1195 return true; 1196 } 1197 1198 /// DeleteFormula - Remove the given formula from this use's list. 1199 void LSRUse::DeleteFormula(Formula &F) { 1200 if (&F != &Formulae.back()) 1201 std::swap(F, Formulae.back()); 1202 Formulae.pop_back(); 1203 } 1204 1205 /// RecomputeRegs - Recompute the Regs field, and update RegUses. 1206 void LSRUse::RecomputeRegs(size_t LUIdx, RegUseTracker &RegUses) { 1207 // Now that we've filtered out some formulae, recompute the Regs set. 1208 SmallPtrSet<const SCEV *, 4> OldRegs = Regs; 1209 Regs.clear(); 1210 for (SmallVectorImpl<Formula>::const_iterator I = Formulae.begin(), 1211 E = Formulae.end(); I != E; ++I) { 1212 const Formula &F = *I; 1213 if (F.ScaledReg) Regs.insert(F.ScaledReg); 1214 Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end()); 1215 } 1216 1217 // Update the RegTracker. 1218 for (SmallPtrSet<const SCEV *, 4>::iterator I = OldRegs.begin(), 1219 E = OldRegs.end(); I != E; ++I) 1220 if (!Regs.count(*I)) 1221 RegUses.DropRegister(*I, LUIdx); 1222 } 1223 1224 void LSRUse::print(raw_ostream &OS) const { 1225 OS << "LSR Use: Kind="; 1226 switch (Kind) { 1227 case Basic: OS << "Basic"; break; 1228 case Special: OS << "Special"; break; 1229 case ICmpZero: OS << "ICmpZero"; break; 1230 case Address: 1231 OS << "Address of "; 1232 if (AccessTy->isPointerTy()) 1233 OS << "pointer"; // the full pointer type could be really verbose 1234 else 1235 OS << *AccessTy; 1236 } 1237 1238 OS << ", Offsets={"; 1239 for (SmallVectorImpl<int64_t>::const_iterator I = Offsets.begin(), 1240 E = Offsets.end(); I != E; ++I) { 1241 OS << *I; 1242 if (llvm::next(I) != E) 1243 OS << ','; 1244 } 1245 OS << '}'; 1246 1247 if (AllFixupsOutsideLoop) 1248 OS << ", all-fixups-outside-loop"; 1249 1250 if (WidestFixupType) 1251 OS << ", widest fixup type: " << *WidestFixupType; 1252 } 1253 1254 void LSRUse::dump() const { 1255 print(errs()); errs() << '\n'; 1256 } 1257 1258 /// isLegalUse - Test whether the use described by AM is "legal", meaning it can 1259 /// be completely folded into the user instruction at isel time. This includes 1260 /// address-mode folding and special icmp tricks. 1261 static bool isLegalUse(const TargetLowering::AddrMode &AM, 1262 LSRUse::KindType Kind, Type *AccessTy, 1263 const TargetLowering *TLI) { 1264 switch (Kind) { 1265 case LSRUse::Address: 1266 // If we have low-level target information, ask the target if it can 1267 // completely fold this address. 1268 if (TLI) return TLI->isLegalAddressingMode(AM, AccessTy); 1269 1270 // Otherwise, just guess that reg+reg addressing is legal. 1271 return !AM.BaseGV && AM.BaseOffs == 0 && AM.Scale <= 1; 1272 1273 case LSRUse::ICmpZero: 1274 // There's not even a target hook for querying whether it would be legal to 1275 // fold a GV into an ICmp. 1276 if (AM.BaseGV) 1277 return false; 1278 1279 // ICmp only has two operands; don't allow more than two non-trivial parts. 1280 if (AM.Scale != 0 && AM.HasBaseReg && AM.BaseOffs != 0) 1281 return false; 1282 1283 // ICmp only supports no scale or a -1 scale, as we can "fold" a -1 scale by 1284 // putting the scaled register in the other operand of the icmp. 1285 if (AM.Scale != 0 && AM.Scale != -1) 1286 return false; 1287 1288 // If we have low-level target information, ask the target if it can fold an 1289 // integer immediate on an icmp. 1290 if (AM.BaseOffs != 0) { 1291 if (!TLI) 1292 return false; 1293 // We have one of: 1294 // ICmpZero BaseReg + Offset => ICmp BaseReg, -Offset 1295 // ICmpZero -1*ScaleReg + Offset => ICmp ScaleReg, Offset 1296 // Offs is the ICmp immediate. 1297 int64_t Offs = AM.BaseOffs; 1298 if (AM.Scale == 0) 1299 Offs = -(uint64_t)Offs; // The cast does the right thing with INT64_MIN. 1300 return TLI->isLegalICmpImmediate(Offs); 1301 } 1302 1303 // ICmpZero BaseReg + -1*ScaleReg => ICmp BaseReg, ScaleReg 1304 return true; 1305 1306 case LSRUse::Basic: 1307 // Only handle single-register values. 1308 return !AM.BaseGV && AM.Scale == 0 && AM.BaseOffs == 0; 1309 1310 case LSRUse::Special: 1311 // Only handle -1 scales, or no scale. 1312 return AM.Scale == 0 || AM.Scale == -1; 1313 } 1314 1315 llvm_unreachable("Invalid LSRUse Kind!"); 1316 } 1317 1318 static bool isLegalUse(TargetLowering::AddrMode AM, 1319 int64_t MinOffset, int64_t MaxOffset, 1320 LSRUse::KindType Kind, Type *AccessTy, 1321 const TargetLowering *TLI) { 1322 // Check for overflow. 1323 if (((int64_t)((uint64_t)AM.BaseOffs + MinOffset) > AM.BaseOffs) != 1324 (MinOffset > 0)) 1325 return false; 1326 AM.BaseOffs = (uint64_t)AM.BaseOffs + MinOffset; 1327 if (isLegalUse(AM, Kind, AccessTy, TLI)) { 1328 AM.BaseOffs = (uint64_t)AM.BaseOffs - MinOffset; 1329 // Check for overflow. 1330 if (((int64_t)((uint64_t)AM.BaseOffs + MaxOffset) > AM.BaseOffs) != 1331 (MaxOffset > 0)) 1332 return false; 1333 AM.BaseOffs = (uint64_t)AM.BaseOffs + MaxOffset; 1334 return isLegalUse(AM, Kind, AccessTy, TLI); 1335 } 1336 return false; 1337 } 1338 1339 static bool isAlwaysFoldable(int64_t BaseOffs, 1340 GlobalValue *BaseGV, 1341 bool HasBaseReg, 1342 LSRUse::KindType Kind, Type *AccessTy, 1343 const TargetLowering *TLI) { 1344 // Fast-path: zero is always foldable. 1345 if (BaseOffs == 0 && !BaseGV) return true; 1346 1347 // Conservatively, create an address with an immediate and a 1348 // base and a scale. 1349 TargetLowering::AddrMode AM; 1350 AM.BaseOffs = BaseOffs; 1351 AM.BaseGV = BaseGV; 1352 AM.HasBaseReg = HasBaseReg; 1353 AM.Scale = Kind == LSRUse::ICmpZero ? -1 : 1; 1354 1355 // Canonicalize a scale of 1 to a base register if the formula doesn't 1356 // already have a base register. 1357 if (!AM.HasBaseReg && AM.Scale == 1) { 1358 AM.Scale = 0; 1359 AM.HasBaseReg = true; 1360 } 1361 1362 return isLegalUse(AM, Kind, AccessTy, TLI); 1363 } 1364 1365 static bool isAlwaysFoldable(const SCEV *S, 1366 int64_t MinOffset, int64_t MaxOffset, 1367 bool HasBaseReg, 1368 LSRUse::KindType Kind, Type *AccessTy, 1369 const TargetLowering *TLI, 1370 ScalarEvolution &SE) { 1371 // Fast-path: zero is always foldable. 1372 if (S->isZero()) return true; 1373 1374 // Conservatively, create an address with an immediate and a 1375 // base and a scale. 1376 int64_t BaseOffs = ExtractImmediate(S, SE); 1377 GlobalValue *BaseGV = ExtractSymbol(S, SE); 1378 1379 // If there's anything else involved, it's not foldable. 1380 if (!S->isZero()) return false; 1381 1382 // Fast-path: zero is always foldable. 1383 if (BaseOffs == 0 && !BaseGV) return true; 1384 1385 // Conservatively, create an address with an immediate and a 1386 // base and a scale. 1387 TargetLowering::AddrMode AM; 1388 AM.BaseOffs = BaseOffs; 1389 AM.BaseGV = BaseGV; 1390 AM.HasBaseReg = HasBaseReg; 1391 AM.Scale = Kind == LSRUse::ICmpZero ? -1 : 1; 1392 1393 return isLegalUse(AM, MinOffset, MaxOffset, Kind, AccessTy, TLI); 1394 } 1395 1396 namespace { 1397 1398 /// UseMapDenseMapInfo - A DenseMapInfo implementation for holding 1399 /// DenseMaps and DenseSets of pairs of const SCEV* and LSRUse::Kind. 1400 struct UseMapDenseMapInfo { 1401 static std::pair<const SCEV *, LSRUse::KindType> getEmptyKey() { 1402 return std::make_pair(reinterpret_cast<const SCEV *>(-1), LSRUse::Basic); 1403 } 1404 1405 static std::pair<const SCEV *, LSRUse::KindType> getTombstoneKey() { 1406 return std::make_pair(reinterpret_cast<const SCEV *>(-2), LSRUse::Basic); 1407 } 1408 1409 static unsigned 1410 getHashValue(const std::pair<const SCEV *, LSRUse::KindType> &V) { 1411 unsigned Result = DenseMapInfo<const SCEV *>::getHashValue(V.first); 1412 Result ^= DenseMapInfo<unsigned>::getHashValue(unsigned(V.second)); 1413 return Result; 1414 } 1415 1416 static bool isEqual(const std::pair<const SCEV *, LSRUse::KindType> &LHS, 1417 const std::pair<const SCEV *, LSRUse::KindType> &RHS) { 1418 return LHS == RHS; 1419 } 1420 }; 1421 1422 /// IVInc - An individual increment in a Chain of IV increments. 1423 /// Relate an IV user to an expression that computes the IV it uses from the IV 1424 /// used by the previous link in the Chain. 1425 /// 1426 /// For the head of a chain, IncExpr holds the absolute SCEV expression for the 1427 /// original IVOperand. The head of the chain's IVOperand is only valid during 1428 /// chain collection, before LSR replaces IV users. During chain generation, 1429 /// IncExpr can be used to find the new IVOperand that computes the same 1430 /// expression. 1431 struct IVInc { 1432 Instruction *UserInst; 1433 Value* IVOperand; 1434 const SCEV *IncExpr; 1435 1436 IVInc(Instruction *U, Value *O, const SCEV *E): 1437 UserInst(U), IVOperand(O), IncExpr(E) {} 1438 }; 1439 1440 // IVChain - The list of IV increments in program order. 1441 // We typically add the head of a chain without finding subsequent links. 1442 typedef SmallVector<IVInc,1> IVChain; 1443 1444 /// ChainUsers - Helper for CollectChains to track multiple IV increment uses. 1445 /// Distinguish between FarUsers that definitely cross IV increments and 1446 /// NearUsers that may be used between IV increments. 1447 struct ChainUsers { 1448 SmallPtrSet<Instruction*, 4> FarUsers; 1449 SmallPtrSet<Instruction*, 4> NearUsers; 1450 }; 1451 1452 /// LSRInstance - This class holds state for the main loop strength reduction 1453 /// logic. 1454 class LSRInstance { 1455 IVUsers &IU; 1456 ScalarEvolution &SE; 1457 DominatorTree &DT; 1458 LoopInfo &LI; 1459 const TargetLowering *const TLI; 1460 Loop *const L; 1461 bool Changed; 1462 1463 /// IVIncInsertPos - This is the insert position that the current loop's 1464 /// induction variable increment should be placed. In simple loops, this is 1465 /// the latch block's terminator. But in more complicated cases, this is a 1466 /// position which will dominate all the in-loop post-increment users. 1467 Instruction *IVIncInsertPos; 1468 1469 /// Factors - Interesting factors between use strides. 1470 SmallSetVector<int64_t, 8> Factors; 1471 1472 /// Types - Interesting use types, to facilitate truncation reuse. 1473 SmallSetVector<Type *, 4> Types; 1474 1475 /// Fixups - The list of operands which are to be replaced. 1476 SmallVector<LSRFixup, 16> Fixups; 1477 1478 /// Uses - The list of interesting uses. 1479 SmallVector<LSRUse, 16> Uses; 1480 1481 /// RegUses - Track which uses use which register candidates. 1482 RegUseTracker RegUses; 1483 1484 // Limit the number of chains to avoid quadratic behavior. We don't expect to 1485 // have more than a few IV increment chains in a loop. Missing a Chain falls 1486 // back to normal LSR behavior for those uses. 1487 static const unsigned MaxChains = 8; 1488 1489 /// IVChainVec - IV users can form a chain of IV increments. 1490 SmallVector<IVChain, MaxChains> IVChainVec; 1491 1492 /// IVIncSet - IV users that belong to profitable IVChains. 1493 SmallPtrSet<Use*, MaxChains> IVIncSet; 1494 1495 void OptimizeShadowIV(); 1496 bool FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse); 1497 ICmpInst *OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse); 1498 void OptimizeLoopTermCond(); 1499 1500 void ChainInstruction(Instruction *UserInst, Instruction *IVOper, 1501 SmallVectorImpl<ChainUsers> &ChainUsersVec); 1502 void FinalizeChain(IVChain &Chain); 1503 void CollectChains(); 1504 void GenerateIVChain(const IVChain &Chain, SCEVExpander &Rewriter, 1505 SmallVectorImpl<WeakVH> &DeadInsts); 1506 1507 void CollectInterestingTypesAndFactors(); 1508 void CollectFixupsAndInitialFormulae(); 1509 1510 LSRFixup &getNewFixup() { 1511 Fixups.push_back(LSRFixup()); 1512 return Fixups.back(); 1513 } 1514 1515 // Support for sharing of LSRUses between LSRFixups. 1516 typedef DenseMap<std::pair<const SCEV *, LSRUse::KindType>, 1517 size_t, 1518 UseMapDenseMapInfo> UseMapTy; 1519 UseMapTy UseMap; 1520 1521 bool reconcileNewOffset(LSRUse &LU, int64_t NewOffset, bool HasBaseReg, 1522 LSRUse::KindType Kind, Type *AccessTy); 1523 1524 std::pair<size_t, int64_t> getUse(const SCEV *&Expr, 1525 LSRUse::KindType Kind, 1526 Type *AccessTy); 1527 1528 void DeleteUse(LSRUse &LU, size_t LUIdx); 1529 1530 LSRUse *FindUseWithSimilarFormula(const Formula &F, const LSRUse &OrigLU); 1531 1532 void InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx); 1533 void InsertSupplementalFormula(const SCEV *S, LSRUse &LU, size_t LUIdx); 1534 void CountRegisters(const Formula &F, size_t LUIdx); 1535 bool InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F); 1536 1537 void CollectLoopInvariantFixupsAndFormulae(); 1538 1539 void GenerateReassociations(LSRUse &LU, unsigned LUIdx, Formula Base, 1540 unsigned Depth = 0); 1541 void GenerateCombinations(LSRUse &LU, unsigned LUIdx, Formula Base); 1542 void GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx, Formula Base); 1543 void GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx, Formula Base); 1544 void GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx, Formula Base); 1545 void GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base); 1546 void GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base); 1547 void GenerateCrossUseConstantOffsets(); 1548 void GenerateAllReuseFormulae(); 1549 1550 void FilterOutUndesirableDedicatedRegisters(); 1551 1552 size_t EstimateSearchSpaceComplexity() const; 1553 void NarrowSearchSpaceByDetectingSupersets(); 1554 void NarrowSearchSpaceByCollapsingUnrolledCode(); 1555 void NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters(); 1556 void NarrowSearchSpaceByPickingWinnerRegs(); 1557 void NarrowSearchSpaceUsingHeuristics(); 1558 1559 void SolveRecurse(SmallVectorImpl<const Formula *> &Solution, 1560 Cost &SolutionCost, 1561 SmallVectorImpl<const Formula *> &Workspace, 1562 const Cost &CurCost, 1563 const SmallPtrSet<const SCEV *, 16> &CurRegs, 1564 DenseSet<const SCEV *> &VisitedRegs) const; 1565 void Solve(SmallVectorImpl<const Formula *> &Solution) const; 1566 1567 BasicBlock::iterator 1568 HoistInsertPosition(BasicBlock::iterator IP, 1569 const SmallVectorImpl<Instruction *> &Inputs) const; 1570 BasicBlock::iterator 1571 AdjustInsertPositionForExpand(BasicBlock::iterator IP, 1572 const LSRFixup &LF, 1573 const LSRUse &LU, 1574 SCEVExpander &Rewriter) const; 1575 1576 Value *Expand(const LSRFixup &LF, 1577 const Formula &F, 1578 BasicBlock::iterator IP, 1579 SCEVExpander &Rewriter, 1580 SmallVectorImpl<WeakVH> &DeadInsts) const; 1581 void RewriteForPHI(PHINode *PN, const LSRFixup &LF, 1582 const Formula &F, 1583 SCEVExpander &Rewriter, 1584 SmallVectorImpl<WeakVH> &DeadInsts, 1585 Pass *P) const; 1586 void Rewrite(const LSRFixup &LF, 1587 const Formula &F, 1588 SCEVExpander &Rewriter, 1589 SmallVectorImpl<WeakVH> &DeadInsts, 1590 Pass *P) const; 1591 void ImplementSolution(const SmallVectorImpl<const Formula *> &Solution, 1592 Pass *P); 1593 1594 public: 1595 LSRInstance(const TargetLowering *tli, Loop *l, Pass *P); 1596 1597 bool getChanged() const { return Changed; } 1598 1599 void print_factors_and_types(raw_ostream &OS) const; 1600 void print_fixups(raw_ostream &OS) const; 1601 void print_uses(raw_ostream &OS) const; 1602 void print(raw_ostream &OS) const; 1603 void dump() const; 1604 }; 1605 1606 } 1607 1608 /// OptimizeShadowIV - If IV is used in a int-to-float cast 1609 /// inside the loop then try to eliminate the cast operation. 1610 void LSRInstance::OptimizeShadowIV() { 1611 const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L); 1612 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount)) 1613 return; 1614 1615 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); 1616 UI != E; /* empty */) { 1617 IVUsers::const_iterator CandidateUI = UI; 1618 ++UI; 1619 Instruction *ShadowUse = CandidateUI->getUser(); 1620 Type *DestTy = NULL; 1621 bool IsSigned = false; 1622 1623 /* If shadow use is a int->float cast then insert a second IV 1624 to eliminate this cast. 1625 1626 for (unsigned i = 0; i < n; ++i) 1627 foo((double)i); 1628 1629 is transformed into 1630 1631 double d = 0.0; 1632 for (unsigned i = 0; i < n; ++i, ++d) 1633 foo(d); 1634 */ 1635 if (UIToFPInst *UCast = dyn_cast<UIToFPInst>(CandidateUI->getUser())) { 1636 IsSigned = false; 1637 DestTy = UCast->getDestTy(); 1638 } 1639 else if (SIToFPInst *SCast = dyn_cast<SIToFPInst>(CandidateUI->getUser())) { 1640 IsSigned = true; 1641 DestTy = SCast->getDestTy(); 1642 } 1643 if (!DestTy) continue; 1644 1645 if (TLI) { 1646 // If target does not support DestTy natively then do not apply 1647 // this transformation. 1648 EVT DVT = TLI->getValueType(DestTy); 1649 if (!TLI->isTypeLegal(DVT)) continue; 1650 } 1651 1652 PHINode *PH = dyn_cast<PHINode>(ShadowUse->getOperand(0)); 1653 if (!PH) continue; 1654 if (PH->getNumIncomingValues() != 2) continue; 1655 1656 Type *SrcTy = PH->getType(); 1657 int Mantissa = DestTy->getFPMantissaWidth(); 1658 if (Mantissa == -1) continue; 1659 if ((int)SE.getTypeSizeInBits(SrcTy) > Mantissa) 1660 continue; 1661 1662 unsigned Entry, Latch; 1663 if (PH->getIncomingBlock(0) == L->getLoopPreheader()) { 1664 Entry = 0; 1665 Latch = 1; 1666 } else { 1667 Entry = 1; 1668 Latch = 0; 1669 } 1670 1671 ConstantInt *Init = dyn_cast<ConstantInt>(PH->getIncomingValue(Entry)); 1672 if (!Init) continue; 1673 Constant *NewInit = ConstantFP::get(DestTy, IsSigned ? 1674 (double)Init->getSExtValue() : 1675 (double)Init->getZExtValue()); 1676 1677 BinaryOperator *Incr = 1678 dyn_cast<BinaryOperator>(PH->getIncomingValue(Latch)); 1679 if (!Incr) continue; 1680 if (Incr->getOpcode() != Instruction::Add 1681 && Incr->getOpcode() != Instruction::Sub) 1682 continue; 1683 1684 /* Initialize new IV, double d = 0.0 in above example. */ 1685 ConstantInt *C = NULL; 1686 if (Incr->getOperand(0) == PH) 1687 C = dyn_cast<ConstantInt>(Incr->getOperand(1)); 1688 else if (Incr->getOperand(1) == PH) 1689 C = dyn_cast<ConstantInt>(Incr->getOperand(0)); 1690 else 1691 continue; 1692 1693 if (!C) continue; 1694 1695 // Ignore negative constants, as the code below doesn't handle them 1696 // correctly. TODO: Remove this restriction. 1697 if (!C->getValue().isStrictlyPositive()) continue; 1698 1699 /* Add new PHINode. */ 1700 PHINode *NewPH = PHINode::Create(DestTy, 2, "IV.S.", PH); 1701 1702 /* create new increment. '++d' in above example. */ 1703 Constant *CFP = ConstantFP::get(DestTy, C->getZExtValue()); 1704 BinaryOperator *NewIncr = 1705 BinaryOperator::Create(Incr->getOpcode() == Instruction::Add ? 1706 Instruction::FAdd : Instruction::FSub, 1707 NewPH, CFP, "IV.S.next.", Incr); 1708 1709 NewPH->addIncoming(NewInit, PH->getIncomingBlock(Entry)); 1710 NewPH->addIncoming(NewIncr, PH->getIncomingBlock(Latch)); 1711 1712 /* Remove cast operation */ 1713 ShadowUse->replaceAllUsesWith(NewPH); 1714 ShadowUse->eraseFromParent(); 1715 Changed = true; 1716 break; 1717 } 1718 } 1719 1720 /// FindIVUserForCond - If Cond has an operand that is an expression of an IV, 1721 /// set the IV user and stride information and return true, otherwise return 1722 /// false. 1723 bool LSRInstance::FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse) { 1724 for (IVUsers::iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI) 1725 if (UI->getUser() == Cond) { 1726 // NOTE: we could handle setcc instructions with multiple uses here, but 1727 // InstCombine does it as well for simple uses, it's not clear that it 1728 // occurs enough in real life to handle. 1729 CondUse = UI; 1730 return true; 1731 } 1732 return false; 1733 } 1734 1735 /// OptimizeMax - Rewrite the loop's terminating condition if it uses 1736 /// a max computation. 1737 /// 1738 /// This is a narrow solution to a specific, but acute, problem. For loops 1739 /// like this: 1740 /// 1741 /// i = 0; 1742 /// do { 1743 /// p[i] = 0.0; 1744 /// } while (++i < n); 1745 /// 1746 /// the trip count isn't just 'n', because 'n' might not be positive. And 1747 /// unfortunately this can come up even for loops where the user didn't use 1748 /// a C do-while loop. For example, seemingly well-behaved top-test loops 1749 /// will commonly be lowered like this: 1750 // 1751 /// if (n > 0) { 1752 /// i = 0; 1753 /// do { 1754 /// p[i] = 0.0; 1755 /// } while (++i < n); 1756 /// } 1757 /// 1758 /// and then it's possible for subsequent optimization to obscure the if 1759 /// test in such a way that indvars can't find it. 1760 /// 1761 /// When indvars can't find the if test in loops like this, it creates a 1762 /// max expression, which allows it to give the loop a canonical 1763 /// induction variable: 1764 /// 1765 /// i = 0; 1766 /// max = n < 1 ? 1 : n; 1767 /// do { 1768 /// p[i] = 0.0; 1769 /// } while (++i != max); 1770 /// 1771 /// Canonical induction variables are necessary because the loop passes 1772 /// are designed around them. The most obvious example of this is the 1773 /// LoopInfo analysis, which doesn't remember trip count values. It 1774 /// expects to be able to rediscover the trip count each time it is 1775 /// needed, and it does this using a simple analysis that only succeeds if 1776 /// the loop has a canonical induction variable. 1777 /// 1778 /// However, when it comes time to generate code, the maximum operation 1779 /// can be quite costly, especially if it's inside of an outer loop. 1780 /// 1781 /// This function solves this problem by detecting this type of loop and 1782 /// rewriting their conditions from ICMP_NE back to ICMP_SLT, and deleting 1783 /// the instructions for the maximum computation. 1784 /// 1785 ICmpInst *LSRInstance::OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse) { 1786 // Check that the loop matches the pattern we're looking for. 1787 if (Cond->getPredicate() != CmpInst::ICMP_EQ && 1788 Cond->getPredicate() != CmpInst::ICMP_NE) 1789 return Cond; 1790 1791 SelectInst *Sel = dyn_cast<SelectInst>(Cond->getOperand(1)); 1792 if (!Sel || !Sel->hasOneUse()) return Cond; 1793 1794 const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L); 1795 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount)) 1796 return Cond; 1797 const SCEV *One = SE.getConstant(BackedgeTakenCount->getType(), 1); 1798 1799 // Add one to the backedge-taken count to get the trip count. 1800 const SCEV *IterationCount = SE.getAddExpr(One, BackedgeTakenCount); 1801 if (IterationCount != SE.getSCEV(Sel)) return Cond; 1802 1803 // Check for a max calculation that matches the pattern. There's no check 1804 // for ICMP_ULE here because the comparison would be with zero, which 1805 // isn't interesting. 1806 CmpInst::Predicate Pred = ICmpInst::BAD_ICMP_PREDICATE; 1807 const SCEVNAryExpr *Max = 0; 1808 if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(BackedgeTakenCount)) { 1809 Pred = ICmpInst::ICMP_SLE; 1810 Max = S; 1811 } else if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(IterationCount)) { 1812 Pred = ICmpInst::ICMP_SLT; 1813 Max = S; 1814 } else if (const SCEVUMaxExpr *U = dyn_cast<SCEVUMaxExpr>(IterationCount)) { 1815 Pred = ICmpInst::ICMP_ULT; 1816 Max = U; 1817 } else { 1818 // No match; bail. 1819 return Cond; 1820 } 1821 1822 // To handle a max with more than two operands, this optimization would 1823 // require additional checking and setup. 1824 if (Max->getNumOperands() != 2) 1825 return Cond; 1826 1827 const SCEV *MaxLHS = Max->getOperand(0); 1828 const SCEV *MaxRHS = Max->getOperand(1); 1829 1830 // ScalarEvolution canonicalizes constants to the left. For < and >, look 1831 // for a comparison with 1. For <= and >=, a comparison with zero. 1832 if (!MaxLHS || 1833 (ICmpInst::isTrueWhenEqual(Pred) ? !MaxLHS->isZero() : (MaxLHS != One))) 1834 return Cond; 1835 1836 // Check the relevant induction variable for conformance to 1837 // the pattern. 1838 const SCEV *IV = SE.getSCEV(Cond->getOperand(0)); 1839 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(IV); 1840 if (!AR || !AR->isAffine() || 1841 AR->getStart() != One || 1842 AR->getStepRecurrence(SE) != One) 1843 return Cond; 1844 1845 assert(AR->getLoop() == L && 1846 "Loop condition operand is an addrec in a different loop!"); 1847 1848 // Check the right operand of the select, and remember it, as it will 1849 // be used in the new comparison instruction. 1850 Value *NewRHS = 0; 1851 if (ICmpInst::isTrueWhenEqual(Pred)) { 1852 // Look for n+1, and grab n. 1853 if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(1))) 1854 if (isa<ConstantInt>(BO->getOperand(1)) && 1855 cast<ConstantInt>(BO->getOperand(1))->isOne() && 1856 SE.getSCEV(BO->getOperand(0)) == MaxRHS) 1857 NewRHS = BO->getOperand(0); 1858 if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(2))) 1859 if (isa<ConstantInt>(BO->getOperand(1)) && 1860 cast<ConstantInt>(BO->getOperand(1))->isOne() && 1861 SE.getSCEV(BO->getOperand(0)) == MaxRHS) 1862 NewRHS = BO->getOperand(0); 1863 if (!NewRHS) 1864 return Cond; 1865 } else if (SE.getSCEV(Sel->getOperand(1)) == MaxRHS) 1866 NewRHS = Sel->getOperand(1); 1867 else if (SE.getSCEV(Sel->getOperand(2)) == MaxRHS) 1868 NewRHS = Sel->getOperand(2); 1869 else if (const SCEVUnknown *SU = dyn_cast<SCEVUnknown>(MaxRHS)) 1870 NewRHS = SU->getValue(); 1871 else 1872 // Max doesn't match expected pattern. 1873 return Cond; 1874 1875 // Determine the new comparison opcode. It may be signed or unsigned, 1876 // and the original comparison may be either equality or inequality. 1877 if (Cond->getPredicate() == CmpInst::ICMP_EQ) 1878 Pred = CmpInst::getInversePredicate(Pred); 1879 1880 // Ok, everything looks ok to change the condition into an SLT or SGE and 1881 // delete the max calculation. 1882 ICmpInst *NewCond = 1883 new ICmpInst(Cond, Pred, Cond->getOperand(0), NewRHS, "scmp"); 1884 1885 // Delete the max calculation instructions. 1886 Cond->replaceAllUsesWith(NewCond); 1887 CondUse->setUser(NewCond); 1888 Instruction *Cmp = cast<Instruction>(Sel->getOperand(0)); 1889 Cond->eraseFromParent(); 1890 Sel->eraseFromParent(); 1891 if (Cmp->use_empty()) 1892 Cmp->eraseFromParent(); 1893 return NewCond; 1894 } 1895 1896 /// OptimizeLoopTermCond - Change loop terminating condition to use the 1897 /// postinc iv when possible. 1898 void 1899 LSRInstance::OptimizeLoopTermCond() { 1900 SmallPtrSet<Instruction *, 4> PostIncs; 1901 1902 BasicBlock *LatchBlock = L->getLoopLatch(); 1903 SmallVector<BasicBlock*, 8> ExitingBlocks; 1904 L->getExitingBlocks(ExitingBlocks); 1905 1906 for (unsigned i = 0, e = ExitingBlocks.size(); i != e; ++i) { 1907 BasicBlock *ExitingBlock = ExitingBlocks[i]; 1908 1909 // Get the terminating condition for the loop if possible. If we 1910 // can, we want to change it to use a post-incremented version of its 1911 // induction variable, to allow coalescing the live ranges for the IV into 1912 // one register value. 1913 1914 BranchInst *TermBr = dyn_cast<BranchInst>(ExitingBlock->getTerminator()); 1915 if (!TermBr) 1916 continue; 1917 // FIXME: Overly conservative, termination condition could be an 'or' etc.. 1918 if (TermBr->isUnconditional() || !isa<ICmpInst>(TermBr->getCondition())) 1919 continue; 1920 1921 // Search IVUsesByStride to find Cond's IVUse if there is one. 1922 IVStrideUse *CondUse = 0; 1923 ICmpInst *Cond = cast<ICmpInst>(TermBr->getCondition()); 1924 if (!FindIVUserForCond(Cond, CondUse)) 1925 continue; 1926 1927 // If the trip count is computed in terms of a max (due to ScalarEvolution 1928 // being unable to find a sufficient guard, for example), change the loop 1929 // comparison to use SLT or ULT instead of NE. 1930 // One consequence of doing this now is that it disrupts the count-down 1931 // optimization. That's not always a bad thing though, because in such 1932 // cases it may still be worthwhile to avoid a max. 1933 Cond = OptimizeMax(Cond, CondUse); 1934 1935 // If this exiting block dominates the latch block, it may also use 1936 // the post-inc value if it won't be shared with other uses. 1937 // Check for dominance. 1938 if (!DT.dominates(ExitingBlock, LatchBlock)) 1939 continue; 1940 1941 // Conservatively avoid trying to use the post-inc value in non-latch 1942 // exits if there may be pre-inc users in intervening blocks. 1943 if (LatchBlock != ExitingBlock) 1944 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI) 1945 // Test if the use is reachable from the exiting block. This dominator 1946 // query is a conservative approximation of reachability. 1947 if (&*UI != CondUse && 1948 !DT.properlyDominates(UI->getUser()->getParent(), ExitingBlock)) { 1949 // Conservatively assume there may be reuse if the quotient of their 1950 // strides could be a legal scale. 1951 const SCEV *A = IU.getStride(*CondUse, L); 1952 const SCEV *B = IU.getStride(*UI, L); 1953 if (!A || !B) continue; 1954 if (SE.getTypeSizeInBits(A->getType()) != 1955 SE.getTypeSizeInBits(B->getType())) { 1956 if (SE.getTypeSizeInBits(A->getType()) > 1957 SE.getTypeSizeInBits(B->getType())) 1958 B = SE.getSignExtendExpr(B, A->getType()); 1959 else 1960 A = SE.getSignExtendExpr(A, B->getType()); 1961 } 1962 if (const SCEVConstant *D = 1963 dyn_cast_or_null<SCEVConstant>(getExactSDiv(B, A, SE))) { 1964 const ConstantInt *C = D->getValue(); 1965 // Stride of one or negative one can have reuse with non-addresses. 1966 if (C->isOne() || C->isAllOnesValue()) 1967 goto decline_post_inc; 1968 // Avoid weird situations. 1969 if (C->getValue().getMinSignedBits() >= 64 || 1970 C->getValue().isMinSignedValue()) 1971 goto decline_post_inc; 1972 // Without TLI, assume that any stride might be valid, and so any 1973 // use might be shared. 1974 if (!TLI) 1975 goto decline_post_inc; 1976 // Check for possible scaled-address reuse. 1977 Type *AccessTy = getAccessType(UI->getUser()); 1978 TargetLowering::AddrMode AM; 1979 AM.Scale = C->getSExtValue(); 1980 if (TLI->isLegalAddressingMode(AM, AccessTy)) 1981 goto decline_post_inc; 1982 AM.Scale = -AM.Scale; 1983 if (TLI->isLegalAddressingMode(AM, AccessTy)) 1984 goto decline_post_inc; 1985 } 1986 } 1987 1988 DEBUG(dbgs() << " Change loop exiting icmp to use postinc iv: " 1989 << *Cond << '\n'); 1990 1991 // It's possible for the setcc instruction to be anywhere in the loop, and 1992 // possible for it to have multiple users. If it is not immediately before 1993 // the exiting block branch, move it. 1994 if (&*++BasicBlock::iterator(Cond) != TermBr) { 1995 if (Cond->hasOneUse()) { 1996 Cond->moveBefore(TermBr); 1997 } else { 1998 // Clone the terminating condition and insert into the loopend. 1999 ICmpInst *OldCond = Cond; 2000 Cond = cast<ICmpInst>(Cond->clone()); 2001 Cond->setName(L->getHeader()->getName() + ".termcond"); 2002 ExitingBlock->getInstList().insert(TermBr, Cond); 2003 2004 // Clone the IVUse, as the old use still exists! 2005 CondUse = &IU.AddUser(Cond, CondUse->getOperandValToReplace()); 2006 TermBr->replaceUsesOfWith(OldCond, Cond); 2007 } 2008 } 2009 2010 // If we get to here, we know that we can transform the setcc instruction to 2011 // use the post-incremented version of the IV, allowing us to coalesce the 2012 // live ranges for the IV correctly. 2013 CondUse->transformToPostInc(L); 2014 Changed = true; 2015 2016 PostIncs.insert(Cond); 2017 decline_post_inc:; 2018 } 2019 2020 // Determine an insertion point for the loop induction variable increment. It 2021 // must dominate all the post-inc comparisons we just set up, and it must 2022 // dominate the loop latch edge. 2023 IVIncInsertPos = L->getLoopLatch()->getTerminator(); 2024 for (SmallPtrSet<Instruction *, 4>::const_iterator I = PostIncs.begin(), 2025 E = PostIncs.end(); I != E; ++I) { 2026 BasicBlock *BB = 2027 DT.findNearestCommonDominator(IVIncInsertPos->getParent(), 2028 (*I)->getParent()); 2029 if (BB == (*I)->getParent()) 2030 IVIncInsertPos = *I; 2031 else if (BB != IVIncInsertPos->getParent()) 2032 IVIncInsertPos = BB->getTerminator(); 2033 } 2034 } 2035 2036 /// reconcileNewOffset - Determine if the given use can accommodate a fixup 2037 /// at the given offset and other details. If so, update the use and 2038 /// return true. 2039 bool 2040 LSRInstance::reconcileNewOffset(LSRUse &LU, int64_t NewOffset, bool HasBaseReg, 2041 LSRUse::KindType Kind, Type *AccessTy) { 2042 int64_t NewMinOffset = LU.MinOffset; 2043 int64_t NewMaxOffset = LU.MaxOffset; 2044 Type *NewAccessTy = AccessTy; 2045 2046 // Check for a mismatched kind. It's tempting to collapse mismatched kinds to 2047 // something conservative, however this can pessimize in the case that one of 2048 // the uses will have all its uses outside the loop, for example. 2049 if (LU.Kind != Kind) 2050 return false; 2051 // Conservatively assume HasBaseReg is true for now. 2052 if (NewOffset < LU.MinOffset) { 2053 if (!isAlwaysFoldable(LU.MaxOffset - NewOffset, 0, HasBaseReg, 2054 Kind, AccessTy, TLI)) 2055 return false; 2056 NewMinOffset = NewOffset; 2057 } else if (NewOffset > LU.MaxOffset) { 2058 if (!isAlwaysFoldable(NewOffset - LU.MinOffset, 0, HasBaseReg, 2059 Kind, AccessTy, TLI)) 2060 return false; 2061 NewMaxOffset = NewOffset; 2062 } 2063 // Check for a mismatched access type, and fall back conservatively as needed. 2064 // TODO: Be less conservative when the type is similar and can use the same 2065 // addressing modes. 2066 if (Kind == LSRUse::Address && AccessTy != LU.AccessTy) 2067 NewAccessTy = Type::getVoidTy(AccessTy->getContext()); 2068 2069 // Update the use. 2070 LU.MinOffset = NewMinOffset; 2071 LU.MaxOffset = NewMaxOffset; 2072 LU.AccessTy = NewAccessTy; 2073 if (NewOffset != LU.Offsets.back()) 2074 LU.Offsets.push_back(NewOffset); 2075 return true; 2076 } 2077 2078 /// getUse - Return an LSRUse index and an offset value for a fixup which 2079 /// needs the given expression, with the given kind and optional access type. 2080 /// Either reuse an existing use or create a new one, as needed. 2081 std::pair<size_t, int64_t> 2082 LSRInstance::getUse(const SCEV *&Expr, 2083 LSRUse::KindType Kind, Type *AccessTy) { 2084 const SCEV *Copy = Expr; 2085 int64_t Offset = ExtractImmediate(Expr, SE); 2086 2087 // Basic uses can't accept any offset, for example. 2088 if (!isAlwaysFoldable(Offset, 0, /*HasBaseReg=*/true, Kind, AccessTy, TLI)) { 2089 Expr = Copy; 2090 Offset = 0; 2091 } 2092 2093 std::pair<UseMapTy::iterator, bool> P = 2094 UseMap.insert(std::make_pair(std::make_pair(Expr, Kind), 0)); 2095 if (!P.second) { 2096 // A use already existed with this base. 2097 size_t LUIdx = P.first->second; 2098 LSRUse &LU = Uses[LUIdx]; 2099 if (reconcileNewOffset(LU, Offset, /*HasBaseReg=*/true, Kind, AccessTy)) 2100 // Reuse this use. 2101 return std::make_pair(LUIdx, Offset); 2102 } 2103 2104 // Create a new use. 2105 size_t LUIdx = Uses.size(); 2106 P.first->second = LUIdx; 2107 Uses.push_back(LSRUse(Kind, AccessTy)); 2108 LSRUse &LU = Uses[LUIdx]; 2109 2110 // We don't need to track redundant offsets, but we don't need to go out 2111 // of our way here to avoid them. 2112 if (LU.Offsets.empty() || Offset != LU.Offsets.back()) 2113 LU.Offsets.push_back(Offset); 2114 2115 LU.MinOffset = Offset; 2116 LU.MaxOffset = Offset; 2117 return std::make_pair(LUIdx, Offset); 2118 } 2119 2120 /// DeleteUse - Delete the given use from the Uses list. 2121 void LSRInstance::DeleteUse(LSRUse &LU, size_t LUIdx) { 2122 if (&LU != &Uses.back()) 2123 std::swap(LU, Uses.back()); 2124 Uses.pop_back(); 2125 2126 // Update RegUses. 2127 RegUses.SwapAndDropUse(LUIdx, Uses.size()); 2128 } 2129 2130 /// FindUseWithFormula - Look for a use distinct from OrigLU which is has 2131 /// a formula that has the same registers as the given formula. 2132 LSRUse * 2133 LSRInstance::FindUseWithSimilarFormula(const Formula &OrigF, 2134 const LSRUse &OrigLU) { 2135 // Search all uses for the formula. This could be more clever. 2136 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) { 2137 LSRUse &LU = Uses[LUIdx]; 2138 // Check whether this use is close enough to OrigLU, to see whether it's 2139 // worthwhile looking through its formulae. 2140 // Ignore ICmpZero uses because they may contain formulae generated by 2141 // GenerateICmpZeroScales, in which case adding fixup offsets may 2142 // be invalid. 2143 if (&LU != &OrigLU && 2144 LU.Kind != LSRUse::ICmpZero && 2145 LU.Kind == OrigLU.Kind && OrigLU.AccessTy == LU.AccessTy && 2146 LU.WidestFixupType == OrigLU.WidestFixupType && 2147 LU.HasFormulaWithSameRegs(OrigF)) { 2148 // Scan through this use's formulae. 2149 for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(), 2150 E = LU.Formulae.end(); I != E; ++I) { 2151 const Formula &F = *I; 2152 // Check to see if this formula has the same registers and symbols 2153 // as OrigF. 2154 if (F.BaseRegs == OrigF.BaseRegs && 2155 F.ScaledReg == OrigF.ScaledReg && 2156 F.AM.BaseGV == OrigF.AM.BaseGV && 2157 F.AM.Scale == OrigF.AM.Scale && 2158 F.UnfoldedOffset == OrigF.UnfoldedOffset) { 2159 if (F.AM.BaseOffs == 0) 2160 return &LU; 2161 // This is the formula where all the registers and symbols matched; 2162 // there aren't going to be any others. Since we declined it, we 2163 // can skip the rest of the formulae and procede to the next LSRUse. 2164 break; 2165 } 2166 } 2167 } 2168 } 2169 2170 // Nothing looked good. 2171 return 0; 2172 } 2173 2174 void LSRInstance::CollectInterestingTypesAndFactors() { 2175 SmallSetVector<const SCEV *, 4> Strides; 2176 2177 // Collect interesting types and strides. 2178 SmallVector<const SCEV *, 4> Worklist; 2179 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI) { 2180 const SCEV *Expr = IU.getExpr(*UI); 2181 2182 // Collect interesting types. 2183 Types.insert(SE.getEffectiveSCEVType(Expr->getType())); 2184 2185 // Add strides for mentioned loops. 2186 Worklist.push_back(Expr); 2187 do { 2188 const SCEV *S = Worklist.pop_back_val(); 2189 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) { 2190 if (AR->getLoop() == L) 2191 Strides.insert(AR->getStepRecurrence(SE)); 2192 Worklist.push_back(AR->getStart()); 2193 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) { 2194 Worklist.append(Add->op_begin(), Add->op_end()); 2195 } 2196 } while (!Worklist.empty()); 2197 } 2198 2199 // Compute interesting factors from the set of interesting strides. 2200 for (SmallSetVector<const SCEV *, 4>::const_iterator 2201 I = Strides.begin(), E = Strides.end(); I != E; ++I) 2202 for (SmallSetVector<const SCEV *, 4>::const_iterator NewStrideIter = 2203 llvm::next(I); NewStrideIter != E; ++NewStrideIter) { 2204 const SCEV *OldStride = *I; 2205 const SCEV *NewStride = *NewStrideIter; 2206 2207 if (SE.getTypeSizeInBits(OldStride->getType()) != 2208 SE.getTypeSizeInBits(NewStride->getType())) { 2209 if (SE.getTypeSizeInBits(OldStride->getType()) > 2210 SE.getTypeSizeInBits(NewStride->getType())) 2211 NewStride = SE.getSignExtendExpr(NewStride, OldStride->getType()); 2212 else 2213 OldStride = SE.getSignExtendExpr(OldStride, NewStride->getType()); 2214 } 2215 if (const SCEVConstant *Factor = 2216 dyn_cast_or_null<SCEVConstant>(getExactSDiv(NewStride, OldStride, 2217 SE, true))) { 2218 if (Factor->getValue()->getValue().getMinSignedBits() <= 64) 2219 Factors.insert(Factor->getValue()->getValue().getSExtValue()); 2220 } else if (const SCEVConstant *Factor = 2221 dyn_cast_or_null<SCEVConstant>(getExactSDiv(OldStride, 2222 NewStride, 2223 SE, true))) { 2224 if (Factor->getValue()->getValue().getMinSignedBits() <= 64) 2225 Factors.insert(Factor->getValue()->getValue().getSExtValue()); 2226 } 2227 } 2228 2229 // If all uses use the same type, don't bother looking for truncation-based 2230 // reuse. 2231 if (Types.size() == 1) 2232 Types.clear(); 2233 2234 DEBUG(print_factors_and_types(dbgs())); 2235 } 2236 2237 /// findIVOperand - Helper for CollectChains that finds an IV operand (computed 2238 /// by an AddRec in this loop) within [OI,OE) or returns OE. If IVUsers mapped 2239 /// Instructions to IVStrideUses, we could partially skip this. 2240 static User::op_iterator 2241 findIVOperand(User::op_iterator OI, User::op_iterator OE, 2242 Loop *L, ScalarEvolution &SE) { 2243 for(; OI != OE; ++OI) { 2244 if (Instruction *Oper = dyn_cast<Instruction>(*OI)) { 2245 if (!SE.isSCEVable(Oper->getType())) 2246 continue; 2247 2248 if (const SCEVAddRecExpr *AR = 2249 dyn_cast<SCEVAddRecExpr>(SE.getSCEV(Oper))) { 2250 if (AR->getLoop() == L) 2251 break; 2252 } 2253 } 2254 } 2255 return OI; 2256 } 2257 2258 /// getWideOperand - IVChain logic must consistenctly peek base TruncInst 2259 /// operands, so wrap it in a convenient helper. 2260 static Value *getWideOperand(Value *Oper) { 2261 if (TruncInst *Trunc = dyn_cast<TruncInst>(Oper)) 2262 return Trunc->getOperand(0); 2263 return Oper; 2264 } 2265 2266 /// isCompatibleIVType - Return true if we allow an IV chain to include both 2267 /// types. 2268 static bool isCompatibleIVType(Value *LVal, Value *RVal) { 2269 Type *LType = LVal->getType(); 2270 Type *RType = RVal->getType(); 2271 return (LType == RType) || (LType->isPointerTy() && RType->isPointerTy()); 2272 } 2273 2274 /// getExprBase - Return an approximation of this SCEV expression's "base", or 2275 /// NULL for any constant. Returning the expression itself is 2276 /// conservative. Returning a deeper subexpression is more precise and valid as 2277 /// long as it isn't less complex than another subexpression. For expressions 2278 /// involving multiple unscaled values, we need to return the pointer-type 2279 /// SCEVUnknown. This avoids forming chains across objects, such as: 2280 /// PrevOper==a[i], IVOper==b[i], IVInc==b-a. 2281 /// 2282 /// Since SCEVUnknown is the rightmost type, and pointers are the rightmost 2283 /// SCEVUnknown, we simply return the rightmost SCEV operand. 2284 static const SCEV *getExprBase(const SCEV *S) { 2285 switch (S->getSCEVType()) { 2286 default: // uncluding scUnknown. 2287 return S; 2288 case scConstant: 2289 return 0; 2290 case scTruncate: 2291 return getExprBase(cast<SCEVTruncateExpr>(S)->getOperand()); 2292 case scZeroExtend: 2293 return getExprBase(cast<SCEVZeroExtendExpr>(S)->getOperand()); 2294 case scSignExtend: 2295 return getExprBase(cast<SCEVSignExtendExpr>(S)->getOperand()); 2296 case scAddExpr: { 2297 // Skip over scaled operands (scMulExpr) to follow add operands as long as 2298 // there's nothing more complex. 2299 // FIXME: not sure if we want to recognize negation. 2300 const SCEVAddExpr *Add = cast<SCEVAddExpr>(S); 2301 for (std::reverse_iterator<SCEVAddExpr::op_iterator> I(Add->op_end()), 2302 E(Add->op_begin()); I != E; ++I) { 2303 const SCEV *SubExpr = *I; 2304 if (SubExpr->getSCEVType() == scAddExpr) 2305 return getExprBase(SubExpr); 2306 2307 if (SubExpr->getSCEVType() != scMulExpr) 2308 return SubExpr; 2309 } 2310 return S; // all operands are scaled, be conservative. 2311 } 2312 case scAddRecExpr: 2313 return getExprBase(cast<SCEVAddRecExpr>(S)->getStart()); 2314 } 2315 } 2316 2317 /// Return true if the chain increment is profitable to expand into a loop 2318 /// invariant value, which may require its own register. A profitable chain 2319 /// increment will be an offset relative to the same base. We allow such offsets 2320 /// to potentially be used as chain increment as long as it's not obviously 2321 /// expensive to expand using real instructions. 2322 static const SCEV * 2323 getProfitableChainIncrement(Value *NextIV, Value *PrevIV, 2324 const IVChain &Chain, Loop *L, 2325 ScalarEvolution &SE, const TargetLowering *TLI) { 2326 // Prune the solution space aggressively by checking that both IV operands 2327 // are expressions that operate on the same unscaled SCEVUnknown. This 2328 // "base" will be canceled by the subsequent getMinusSCEV call. Checking first 2329 // avoids creating extra SCEV expressions. 2330 const SCEV *OperExpr = SE.getSCEV(NextIV); 2331 const SCEV *PrevExpr = SE.getSCEV(PrevIV); 2332 if (getExprBase(OperExpr) != getExprBase(PrevExpr) && !StressIVChain) 2333 return 0; 2334 2335 const SCEV *IncExpr = SE.getMinusSCEV(OperExpr, PrevExpr); 2336 if (!SE.isLoopInvariant(IncExpr, L)) 2337 return 0; 2338 2339 // We are not able to expand an increment unless it is loop invariant, 2340 // however, the following checks are purely for profitability. 2341 if (StressIVChain) 2342 return IncExpr; 2343 2344 // Do not replace a constant offset from IV head with a nonconstant IV 2345 // increment. 2346 if (!isa<SCEVConstant>(IncExpr)) { 2347 const SCEV *HeadExpr = SE.getSCEV(getWideOperand(Chain[0].IVOperand)); 2348 if (isa<SCEVConstant>(SE.getMinusSCEV(OperExpr, HeadExpr))) 2349 return 0; 2350 } 2351 2352 SmallPtrSet<const SCEV*, 8> Processed; 2353 if (isHighCostExpansion(IncExpr, Processed, SE)) 2354 return 0; 2355 2356 return IncExpr; 2357 } 2358 2359 /// Return true if the number of registers needed for the chain is estimated to 2360 /// be less than the number required for the individual IV users. First prohibit 2361 /// any IV users that keep the IV live across increments (the Users set should 2362 /// be empty). Next count the number and type of increments in the chain. 2363 /// 2364 /// Chaining IVs can lead to considerable code bloat if ISEL doesn't 2365 /// effectively use postinc addressing modes. Only consider it profitable it the 2366 /// increments can be computed in fewer registers when chained. 2367 /// 2368 /// TODO: Consider IVInc free if it's already used in another chains. 2369 static bool 2370 isProfitableChain(IVChain &Chain, SmallPtrSet<Instruction*, 4> &Users, 2371 ScalarEvolution &SE, const TargetLowering *TLI) { 2372 if (StressIVChain) 2373 return true; 2374 2375 if (Chain.size() <= 2) 2376 return false; 2377 2378 if (!Users.empty()) { 2379 DEBUG(dbgs() << "Chain: " << *Chain[0].UserInst << " users:\n"; 2380 for (SmallPtrSet<Instruction*, 4>::const_iterator I = Users.begin(), 2381 E = Users.end(); I != E; ++I) { 2382 dbgs() << " " << **I << "\n"; 2383 }); 2384 return false; 2385 } 2386 assert(!Chain.empty() && "empty IV chains are not allowed"); 2387 2388 // The chain itself may require a register, so intialize cost to 1. 2389 int cost = 1; 2390 2391 // A complete chain likely eliminates the need for keeping the original IV in 2392 // a register. LSR does not currently know how to form a complete chain unless 2393 // the header phi already exists. 2394 if (isa<PHINode>(Chain.back().UserInst) 2395 && SE.getSCEV(Chain.back().UserInst) == Chain[0].IncExpr) { 2396 --cost; 2397 } 2398 const SCEV *LastIncExpr = 0; 2399 unsigned NumConstIncrements = 0; 2400 unsigned NumVarIncrements = 0; 2401 unsigned NumReusedIncrements = 0; 2402 for (IVChain::const_iterator I = llvm::next(Chain.begin()), E = Chain.end(); 2403 I != E; ++I) { 2404 2405 if (I->IncExpr->isZero()) 2406 continue; 2407 2408 // Incrementing by zero or some constant is neutral. We assume constants can 2409 // be folded into an addressing mode or an add's immediate operand. 2410 if (isa<SCEVConstant>(I->IncExpr)) { 2411 ++NumConstIncrements; 2412 continue; 2413 } 2414 2415 if (I->IncExpr == LastIncExpr) 2416 ++NumReusedIncrements; 2417 else 2418 ++NumVarIncrements; 2419 2420 LastIncExpr = I->IncExpr; 2421 } 2422 // An IV chain with a single increment is handled by LSR's postinc 2423 // uses. However, a chain with multiple increments requires keeping the IV's 2424 // value live longer than it needs to be if chained. 2425 if (NumConstIncrements > 1) 2426 --cost; 2427 2428 // Materializing increment expressions in the preheader that didn't exist in 2429 // the original code may cost a register. For example, sign-extended array 2430 // indices can produce ridiculous increments like this: 2431 // IV + ((sext i32 (2 * %s) to i64) + (-1 * (sext i32 %s to i64))) 2432 cost += NumVarIncrements; 2433 2434 // Reusing variable increments likely saves a register to hold the multiple of 2435 // the stride. 2436 cost -= NumReusedIncrements; 2437 2438 DEBUG(dbgs() << "Chain: " << *Chain[0].UserInst << " Cost: " << cost << "\n"); 2439 2440 return cost < 0; 2441 } 2442 2443 /// ChainInstruction - Add this IV user to an existing chain or make it the head 2444 /// of a new chain. 2445 void LSRInstance::ChainInstruction(Instruction *UserInst, Instruction *IVOper, 2446 SmallVectorImpl<ChainUsers> &ChainUsersVec) { 2447 // When IVs are used as types of varying widths, they are generally converted 2448 // to a wider type with some uses remaining narrow under a (free) trunc. 2449 Value *NextIV = getWideOperand(IVOper); 2450 2451 // Visit all existing chains. Check if its IVOper can be computed as a 2452 // profitable loop invariant increment from the last link in the Chain. 2453 unsigned ChainIdx = 0, NChains = IVChainVec.size(); 2454 const SCEV *LastIncExpr = 0; 2455 for (; ChainIdx < NChains; ++ChainIdx) { 2456 Value *PrevIV = getWideOperand(IVChainVec[ChainIdx].back().IVOperand); 2457 if (!isCompatibleIVType(PrevIV, NextIV)) 2458 continue; 2459 2460 // A phi node terminates a chain. 2461 if (isa<PHINode>(UserInst) 2462 && isa<PHINode>(IVChainVec[ChainIdx].back().UserInst)) 2463 continue; 2464 2465 if (const SCEV *IncExpr = 2466 getProfitableChainIncrement(NextIV, PrevIV, IVChainVec[ChainIdx], 2467 L, SE, TLI)) { 2468 LastIncExpr = IncExpr; 2469 break; 2470 } 2471 } 2472 // If we haven't found a chain, create a new one, unless we hit the max. Don't 2473 // bother for phi nodes, because they must be last in the chain. 2474 if (ChainIdx == NChains) { 2475 if (isa<PHINode>(UserInst)) 2476 return; 2477 if (NChains >= MaxChains && !StressIVChain) { 2478 DEBUG(dbgs() << "IV Chain Limit\n"); 2479 return; 2480 } 2481 LastIncExpr = SE.getSCEV(NextIV); 2482 // IVUsers may have skipped over sign/zero extensions. We don't currently 2483 // attempt to form chains involving extensions unless they can be hoisted 2484 // into this loop's AddRec. 2485 if (!isa<SCEVAddRecExpr>(LastIncExpr)) 2486 return; 2487 ++NChains; 2488 IVChainVec.resize(NChains); 2489 ChainUsersVec.resize(NChains); 2490 DEBUG(dbgs() << "IV Head: (" << *UserInst << ") IV=" << *LastIncExpr 2491 << "\n"); 2492 } 2493 else 2494 DEBUG(dbgs() << "IV Inc: (" << *UserInst << ") IV+" << *LastIncExpr 2495 << "\n"); 2496 2497 // Add this IV user to the end of the chain. 2498 IVChainVec[ChainIdx].push_back(IVInc(UserInst, IVOper, LastIncExpr)); 2499 2500 SmallPtrSet<Instruction*,4> &NearUsers = ChainUsersVec[ChainIdx].NearUsers; 2501 // This chain's NearUsers become FarUsers. 2502 if (!LastIncExpr->isZero()) { 2503 ChainUsersVec[ChainIdx].FarUsers.insert(NearUsers.begin(), 2504 NearUsers.end()); 2505 NearUsers.clear(); 2506 } 2507 2508 // All other uses of IVOperand become near uses of the chain. 2509 // We currently ignore intermediate values within SCEV expressions, assuming 2510 // they will eventually be used be the current chain, or can be computed 2511 // from one of the chain increments. To be more precise we could 2512 // transitively follow its user and only add leaf IV users to the set. 2513 for (Value::use_iterator UseIter = IVOper->use_begin(), 2514 UseEnd = IVOper->use_end(); UseIter != UseEnd; ++UseIter) { 2515 Instruction *OtherUse = dyn_cast<Instruction>(*UseIter); 2516 if (!OtherUse || OtherUse == UserInst) 2517 continue; 2518 if (SE.isSCEVable(OtherUse->getType()) 2519 && !isa<SCEVUnknown>(SE.getSCEV(OtherUse)) 2520 && IU.isIVUserOrOperand(OtherUse)) { 2521 continue; 2522 } 2523 NearUsers.insert(OtherUse); 2524 } 2525 2526 // Since this user is part of the chain, it's no longer considered a use 2527 // of the chain. 2528 ChainUsersVec[ChainIdx].FarUsers.erase(UserInst); 2529 } 2530 2531 /// CollectChains - Populate the vector of Chains. 2532 /// 2533 /// This decreases ILP at the architecture level. Targets with ample registers, 2534 /// multiple memory ports, and no register renaming probably don't want 2535 /// this. However, such targets should probably disable LSR altogether. 2536 /// 2537 /// The job of LSR is to make a reasonable choice of induction variables across 2538 /// the loop. Subsequent passes can easily "unchain" computation exposing more 2539 /// ILP *within the loop* if the target wants it. 2540 /// 2541 /// Finding the best IV chain is potentially a scheduling problem. Since LSR 2542 /// will not reorder memory operations, it will recognize this as a chain, but 2543 /// will generate redundant IV increments. Ideally this would be corrected later 2544 /// by a smart scheduler: 2545 /// = A[i] 2546 /// = A[i+x] 2547 /// A[i] = 2548 /// A[i+x] = 2549 /// 2550 /// TODO: Walk the entire domtree within this loop, not just the path to the 2551 /// loop latch. This will discover chains on side paths, but requires 2552 /// maintaining multiple copies of the Chains state. 2553 void LSRInstance::CollectChains() { 2554 SmallVector<ChainUsers, 8> ChainUsersVec; 2555 2556 SmallVector<BasicBlock *,8> LatchPath; 2557 BasicBlock *LoopHeader = L->getHeader(); 2558 for (DomTreeNode *Rung = DT.getNode(L->getLoopLatch()); 2559 Rung->getBlock() != LoopHeader; Rung = Rung->getIDom()) { 2560 LatchPath.push_back(Rung->getBlock()); 2561 } 2562 LatchPath.push_back(LoopHeader); 2563 2564 // Walk the instruction stream from the loop header to the loop latch. 2565 for (SmallVectorImpl<BasicBlock *>::reverse_iterator 2566 BBIter = LatchPath.rbegin(), BBEnd = LatchPath.rend(); 2567 BBIter != BBEnd; ++BBIter) { 2568 for (BasicBlock::iterator I = (*BBIter)->begin(), E = (*BBIter)->end(); 2569 I != E; ++I) { 2570 // Skip instructions that weren't seen by IVUsers analysis. 2571 if (isa<PHINode>(I) || !IU.isIVUserOrOperand(I)) 2572 continue; 2573 2574 // Ignore users that are part of a SCEV expression. This way we only 2575 // consider leaf IV Users. This effectively rediscovers a portion of 2576 // IVUsers analysis but in program order this time. 2577 if (SE.isSCEVable(I->getType()) && !isa<SCEVUnknown>(SE.getSCEV(I))) 2578 continue; 2579 2580 // Remove this instruction from any NearUsers set it may be in. 2581 for (unsigned ChainIdx = 0, NChains = IVChainVec.size(); 2582 ChainIdx < NChains; ++ChainIdx) { 2583 ChainUsersVec[ChainIdx].NearUsers.erase(I); 2584 } 2585 // Search for operands that can be chained. 2586 SmallPtrSet<Instruction*, 4> UniqueOperands; 2587 User::op_iterator IVOpEnd = I->op_end(); 2588 User::op_iterator IVOpIter = findIVOperand(I->op_begin(), IVOpEnd, L, SE); 2589 while (IVOpIter != IVOpEnd) { 2590 Instruction *IVOpInst = cast<Instruction>(*IVOpIter); 2591 if (UniqueOperands.insert(IVOpInst)) 2592 ChainInstruction(I, IVOpInst, ChainUsersVec); 2593 IVOpIter = findIVOperand(llvm::next(IVOpIter), IVOpEnd, L, SE); 2594 } 2595 } // Continue walking down the instructions. 2596 } // Continue walking down the domtree. 2597 // Visit phi backedges to determine if the chain can generate the IV postinc. 2598 for (BasicBlock::iterator I = L->getHeader()->begin(); 2599 PHINode *PN = dyn_cast<PHINode>(I); ++I) { 2600 if (!SE.isSCEVable(PN->getType())) 2601 continue; 2602 2603 Instruction *IncV = 2604 dyn_cast<Instruction>(PN->getIncomingValueForBlock(L->getLoopLatch())); 2605 if (IncV) 2606 ChainInstruction(PN, IncV, ChainUsersVec); 2607 } 2608 // Remove any unprofitable chains. 2609 unsigned ChainIdx = 0; 2610 for (unsigned UsersIdx = 0, NChains = IVChainVec.size(); 2611 UsersIdx < NChains; ++UsersIdx) { 2612 if (!isProfitableChain(IVChainVec[UsersIdx], 2613 ChainUsersVec[UsersIdx].FarUsers, SE, TLI)) 2614 continue; 2615 // Preserve the chain at UsesIdx. 2616 if (ChainIdx != UsersIdx) 2617 IVChainVec[ChainIdx] = IVChainVec[UsersIdx]; 2618 FinalizeChain(IVChainVec[ChainIdx]); 2619 ++ChainIdx; 2620 } 2621 IVChainVec.resize(ChainIdx); 2622 } 2623 2624 void LSRInstance::FinalizeChain(IVChain &Chain) { 2625 assert(!Chain.empty() && "empty IV chains are not allowed"); 2626 DEBUG(dbgs() << "Final Chain: " << *Chain[0].UserInst << "\n"); 2627 2628 for (IVChain::const_iterator I = llvm::next(Chain.begin()), E = Chain.end(); 2629 I != E; ++I) { 2630 DEBUG(dbgs() << " Inc: " << *I->UserInst << "\n"); 2631 User::op_iterator UseI = 2632 std::find(I->UserInst->op_begin(), I->UserInst->op_end(), I->IVOperand); 2633 assert(UseI != I->UserInst->op_end() && "cannot find IV operand"); 2634 IVIncSet.insert(UseI); 2635 } 2636 } 2637 2638 /// Return true if the IVInc can be folded into an addressing mode. 2639 static bool canFoldIVIncExpr(const SCEV *IncExpr, Instruction *UserInst, 2640 Value *Operand, const TargetLowering *TLI) { 2641 const SCEVConstant *IncConst = dyn_cast<SCEVConstant>(IncExpr); 2642 if (!IncConst || !isAddressUse(UserInst, Operand)) 2643 return false; 2644 2645 if (IncConst->getValue()->getValue().getMinSignedBits() > 64) 2646 return false; 2647 2648 int64_t IncOffset = IncConst->getValue()->getSExtValue(); 2649 if (!isAlwaysFoldable(IncOffset, /*BaseGV=*/0, /*HaseBaseReg=*/false, 2650 LSRUse::Address, getAccessType(UserInst), TLI)) 2651 return false; 2652 2653 return true; 2654 } 2655 2656 /// GenerateIVChains - Generate an add or subtract for each IVInc in a chain to 2657 /// materialize the IV user's operand from the previous IV user's operand. 2658 void LSRInstance::GenerateIVChain(const IVChain &Chain, SCEVExpander &Rewriter, 2659 SmallVectorImpl<WeakVH> &DeadInsts) { 2660 // Find the new IVOperand for the head of the chain. It may have been replaced 2661 // by LSR. 2662 const IVInc &Head = Chain[0]; 2663 User::op_iterator IVOpEnd = Head.UserInst->op_end(); 2664 User::op_iterator IVOpIter = findIVOperand(Head.UserInst->op_begin(), 2665 IVOpEnd, L, SE); 2666 Value *IVSrc = 0; 2667 while (IVOpIter != IVOpEnd) { 2668 IVSrc = getWideOperand(*IVOpIter); 2669 2670 // If this operand computes the expression that the chain needs, we may use 2671 // it. (Check this after setting IVSrc which is used below.) 2672 // 2673 // Note that if Head.IncExpr is wider than IVSrc, then this phi is too 2674 // narrow for the chain, so we can no longer use it. We do allow using a 2675 // wider phi, assuming the LSR checked for free truncation. In that case we 2676 // should already have a truncate on this operand such that 2677 // getSCEV(IVSrc) == IncExpr. 2678 if (SE.getSCEV(*IVOpIter) == Head.IncExpr 2679 || SE.getSCEV(IVSrc) == Head.IncExpr) { 2680 break; 2681 } 2682 IVOpIter = findIVOperand(llvm::next(IVOpIter), IVOpEnd, L, SE); 2683 } 2684 if (IVOpIter == IVOpEnd) { 2685 // Gracefully give up on this chain. 2686 DEBUG(dbgs() << "Concealed chain head: " << *Head.UserInst << "\n"); 2687 return; 2688 } 2689 2690 DEBUG(dbgs() << "Generate chain at: " << *IVSrc << "\n"); 2691 Type *IVTy = IVSrc->getType(); 2692 Type *IntTy = SE.getEffectiveSCEVType(IVTy); 2693 const SCEV *LeftOverExpr = 0; 2694 for (IVChain::const_iterator IncI = llvm::next(Chain.begin()), 2695 IncE = Chain.end(); IncI != IncE; ++IncI) { 2696 2697 Instruction *InsertPt = IncI->UserInst; 2698 if (isa<PHINode>(InsertPt)) 2699 InsertPt = L->getLoopLatch()->getTerminator(); 2700 2701 // IVOper will replace the current IV User's operand. IVSrc is the IV 2702 // value currently held in a register. 2703 Value *IVOper = IVSrc; 2704 if (!IncI->IncExpr->isZero()) { 2705 // IncExpr was the result of subtraction of two narrow values, so must 2706 // be signed. 2707 const SCEV *IncExpr = SE.getNoopOrSignExtend(IncI->IncExpr, IntTy); 2708 LeftOverExpr = LeftOverExpr ? 2709 SE.getAddExpr(LeftOverExpr, IncExpr) : IncExpr; 2710 } 2711 if (LeftOverExpr && !LeftOverExpr->isZero()) { 2712 // Expand the IV increment. 2713 Rewriter.clearPostInc(); 2714 Value *IncV = Rewriter.expandCodeFor(LeftOverExpr, IntTy, InsertPt); 2715 const SCEV *IVOperExpr = SE.getAddExpr(SE.getUnknown(IVSrc), 2716 SE.getUnknown(IncV)); 2717 IVOper = Rewriter.expandCodeFor(IVOperExpr, IVTy, InsertPt); 2718 2719 // If an IV increment can't be folded, use it as the next IV value. 2720 if (!canFoldIVIncExpr(LeftOverExpr, IncI->UserInst, IncI->IVOperand, 2721 TLI)) { 2722 assert(IVTy == IVOper->getType() && "inconsistent IV increment type"); 2723 IVSrc = IVOper; 2724 LeftOverExpr = 0; 2725 } 2726 } 2727 Type *OperTy = IncI->IVOperand->getType(); 2728 if (IVTy != OperTy) { 2729 assert(SE.getTypeSizeInBits(IVTy) >= SE.getTypeSizeInBits(OperTy) && 2730 "cannot extend a chained IV"); 2731 IRBuilder<> Builder(InsertPt); 2732 IVOper = Builder.CreateTruncOrBitCast(IVOper, OperTy, "lsr.chain"); 2733 } 2734 IncI->UserInst->replaceUsesOfWith(IncI->IVOperand, IVOper); 2735 DeadInsts.push_back(IncI->IVOperand); 2736 } 2737 // If LSR created a new, wider phi, we may also replace its postinc. We only 2738 // do this if we also found a wide value for the head of the chain. 2739 if (isa<PHINode>(Chain.back().UserInst)) { 2740 for (BasicBlock::iterator I = L->getHeader()->begin(); 2741 PHINode *Phi = dyn_cast<PHINode>(I); ++I) { 2742 if (!isCompatibleIVType(Phi, IVSrc)) 2743 continue; 2744 Instruction *PostIncV = dyn_cast<Instruction>( 2745 Phi->getIncomingValueForBlock(L->getLoopLatch())); 2746 if (!PostIncV || (SE.getSCEV(PostIncV) != SE.getSCEV(IVSrc))) 2747 continue; 2748 Value *IVOper = IVSrc; 2749 Type *PostIncTy = PostIncV->getType(); 2750 if (IVTy != PostIncTy) { 2751 assert(PostIncTy->isPointerTy() && "mixing int/ptr IV types"); 2752 IRBuilder<> Builder(L->getLoopLatch()->getTerminator()); 2753 Builder.SetCurrentDebugLocation(PostIncV->getDebugLoc()); 2754 IVOper = Builder.CreatePointerCast(IVSrc, PostIncTy, "lsr.chain"); 2755 } 2756 Phi->replaceUsesOfWith(PostIncV, IVOper); 2757 DeadInsts.push_back(PostIncV); 2758 } 2759 } 2760 } 2761 2762 void LSRInstance::CollectFixupsAndInitialFormulae() { 2763 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI) { 2764 Instruction *UserInst = UI->getUser(); 2765 // Skip IV users that are part of profitable IV Chains. 2766 User::op_iterator UseI = std::find(UserInst->op_begin(), UserInst->op_end(), 2767 UI->getOperandValToReplace()); 2768 assert(UseI != UserInst->op_end() && "cannot find IV operand"); 2769 if (IVIncSet.count(UseI)) 2770 continue; 2771 2772 // Record the uses. 2773 LSRFixup &LF = getNewFixup(); 2774 LF.UserInst = UserInst; 2775 LF.OperandValToReplace = UI->getOperandValToReplace(); 2776 LF.PostIncLoops = UI->getPostIncLoops(); 2777 2778 LSRUse::KindType Kind = LSRUse::Basic; 2779 Type *AccessTy = 0; 2780 if (isAddressUse(LF.UserInst, LF.OperandValToReplace)) { 2781 Kind = LSRUse::Address; 2782 AccessTy = getAccessType(LF.UserInst); 2783 } 2784 2785 const SCEV *S = IU.getExpr(*UI); 2786 2787 // Equality (== and !=) ICmps are special. We can rewrite (i == N) as 2788 // (N - i == 0), and this allows (N - i) to be the expression that we work 2789 // with rather than just N or i, so we can consider the register 2790 // requirements for both N and i at the same time. Limiting this code to 2791 // equality icmps is not a problem because all interesting loops use 2792 // equality icmps, thanks to IndVarSimplify. 2793 if (ICmpInst *CI = dyn_cast<ICmpInst>(LF.UserInst)) 2794 if (CI->isEquality()) { 2795 // Swap the operands if needed to put the OperandValToReplace on the 2796 // left, for consistency. 2797 Value *NV = CI->getOperand(1); 2798 if (NV == LF.OperandValToReplace) { 2799 CI->setOperand(1, CI->getOperand(0)); 2800 CI->setOperand(0, NV); 2801 NV = CI->getOperand(1); 2802 Changed = true; 2803 } 2804 2805 // x == y --> x - y == 0 2806 const SCEV *N = SE.getSCEV(NV); 2807 if (SE.isLoopInvariant(N, L)) { 2808 // S is normalized, so normalize N before folding it into S 2809 // to keep the result normalized. 2810 N = TransformForPostIncUse(Normalize, N, CI, 0, 2811 LF.PostIncLoops, SE, DT); 2812 Kind = LSRUse::ICmpZero; 2813 S = SE.getMinusSCEV(N, S); 2814 } 2815 2816 // -1 and the negations of all interesting strides (except the negation 2817 // of -1) are now also interesting. 2818 for (size_t i = 0, e = Factors.size(); i != e; ++i) 2819 if (Factors[i] != -1) 2820 Factors.insert(-(uint64_t)Factors[i]); 2821 Factors.insert(-1); 2822 } 2823 2824 // Set up the initial formula for this use. 2825 std::pair<size_t, int64_t> P = getUse(S, Kind, AccessTy); 2826 LF.LUIdx = P.first; 2827 LF.Offset = P.second; 2828 LSRUse &LU = Uses[LF.LUIdx]; 2829 LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L); 2830 if (!LU.WidestFixupType || 2831 SE.getTypeSizeInBits(LU.WidestFixupType) < 2832 SE.getTypeSizeInBits(LF.OperandValToReplace->getType())) 2833 LU.WidestFixupType = LF.OperandValToReplace->getType(); 2834 2835 // If this is the first use of this LSRUse, give it a formula. 2836 if (LU.Formulae.empty()) { 2837 InsertInitialFormula(S, LU, LF.LUIdx); 2838 CountRegisters(LU.Formulae.back(), LF.LUIdx); 2839 } 2840 } 2841 2842 DEBUG(print_fixups(dbgs())); 2843 } 2844 2845 /// InsertInitialFormula - Insert a formula for the given expression into 2846 /// the given use, separating out loop-variant portions from loop-invariant 2847 /// and loop-computable portions. 2848 void 2849 LSRInstance::InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx) { 2850 Formula F; 2851 F.InitialMatch(S, L, SE); 2852 bool Inserted = InsertFormula(LU, LUIdx, F); 2853 assert(Inserted && "Initial formula already exists!"); (void)Inserted; 2854 } 2855 2856 /// InsertSupplementalFormula - Insert a simple single-register formula for 2857 /// the given expression into the given use. 2858 void 2859 LSRInstance::InsertSupplementalFormula(const SCEV *S, 2860 LSRUse &LU, size_t LUIdx) { 2861 Formula F; 2862 F.BaseRegs.push_back(S); 2863 F.AM.HasBaseReg = true; 2864 bool Inserted = InsertFormula(LU, LUIdx, F); 2865 assert(Inserted && "Supplemental formula already exists!"); (void)Inserted; 2866 } 2867 2868 /// CountRegisters - Note which registers are used by the given formula, 2869 /// updating RegUses. 2870 void LSRInstance::CountRegisters(const Formula &F, size_t LUIdx) { 2871 if (F.ScaledReg) 2872 RegUses.CountRegister(F.ScaledReg, LUIdx); 2873 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(), 2874 E = F.BaseRegs.end(); I != E; ++I) 2875 RegUses.CountRegister(*I, LUIdx); 2876 } 2877 2878 /// InsertFormula - If the given formula has not yet been inserted, add it to 2879 /// the list, and return true. Return false otherwise. 2880 bool LSRInstance::InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F) { 2881 if (!LU.InsertFormula(F)) 2882 return false; 2883 2884 CountRegisters(F, LUIdx); 2885 return true; 2886 } 2887 2888 /// CollectLoopInvariantFixupsAndFormulae - Check for other uses of 2889 /// loop-invariant values which we're tracking. These other uses will pin these 2890 /// values in registers, making them less profitable for elimination. 2891 /// TODO: This currently misses non-constant addrec step registers. 2892 /// TODO: Should this give more weight to users inside the loop? 2893 void 2894 LSRInstance::CollectLoopInvariantFixupsAndFormulae() { 2895 SmallVector<const SCEV *, 8> Worklist(RegUses.begin(), RegUses.end()); 2896 SmallPtrSet<const SCEV *, 8> Inserted; 2897 2898 while (!Worklist.empty()) { 2899 const SCEV *S = Worklist.pop_back_val(); 2900 2901 if (const SCEVNAryExpr *N = dyn_cast<SCEVNAryExpr>(S)) 2902 Worklist.append(N->op_begin(), N->op_end()); 2903 else if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(S)) 2904 Worklist.push_back(C->getOperand()); 2905 else if (const SCEVUDivExpr *D = dyn_cast<SCEVUDivExpr>(S)) { 2906 Worklist.push_back(D->getLHS()); 2907 Worklist.push_back(D->getRHS()); 2908 } else if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) { 2909 if (!Inserted.insert(U)) continue; 2910 const Value *V = U->getValue(); 2911 if (const Instruction *Inst = dyn_cast<Instruction>(V)) { 2912 // Look for instructions defined outside the loop. 2913 if (L->contains(Inst)) continue; 2914 } else if (isa<UndefValue>(V)) 2915 // Undef doesn't have a live range, so it doesn't matter. 2916 continue; 2917 for (Value::const_use_iterator UI = V->use_begin(), UE = V->use_end(); 2918 UI != UE; ++UI) { 2919 const Instruction *UserInst = dyn_cast<Instruction>(*UI); 2920 // Ignore non-instructions. 2921 if (!UserInst) 2922 continue; 2923 // Ignore instructions in other functions (as can happen with 2924 // Constants). 2925 if (UserInst->getParent()->getParent() != L->getHeader()->getParent()) 2926 continue; 2927 // Ignore instructions not dominated by the loop. 2928 const BasicBlock *UseBB = !isa<PHINode>(UserInst) ? 2929 UserInst->getParent() : 2930 cast<PHINode>(UserInst)->getIncomingBlock( 2931 PHINode::getIncomingValueNumForOperand(UI.getOperandNo())); 2932 if (!DT.dominates(L->getHeader(), UseBB)) 2933 continue; 2934 // Ignore uses which are part of other SCEV expressions, to avoid 2935 // analyzing them multiple times. 2936 if (SE.isSCEVable(UserInst->getType())) { 2937 const SCEV *UserS = SE.getSCEV(const_cast<Instruction *>(UserInst)); 2938 // If the user is a no-op, look through to its uses. 2939 if (!isa<SCEVUnknown>(UserS)) 2940 continue; 2941 if (UserS == U) { 2942 Worklist.push_back( 2943 SE.getUnknown(const_cast<Instruction *>(UserInst))); 2944 continue; 2945 } 2946 } 2947 // Ignore icmp instructions which are already being analyzed. 2948 if (const ICmpInst *ICI = dyn_cast<ICmpInst>(UserInst)) { 2949 unsigned OtherIdx = !UI.getOperandNo(); 2950 Value *OtherOp = const_cast<Value *>(ICI->getOperand(OtherIdx)); 2951 if (SE.hasComputableLoopEvolution(SE.getSCEV(OtherOp), L)) 2952 continue; 2953 } 2954 2955 LSRFixup &LF = getNewFixup(); 2956 LF.UserInst = const_cast<Instruction *>(UserInst); 2957 LF.OperandValToReplace = UI.getUse(); 2958 std::pair<size_t, int64_t> P = getUse(S, LSRUse::Basic, 0); 2959 LF.LUIdx = P.first; 2960 LF.Offset = P.second; 2961 LSRUse &LU = Uses[LF.LUIdx]; 2962 LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L); 2963 if (!LU.WidestFixupType || 2964 SE.getTypeSizeInBits(LU.WidestFixupType) < 2965 SE.getTypeSizeInBits(LF.OperandValToReplace->getType())) 2966 LU.WidestFixupType = LF.OperandValToReplace->getType(); 2967 InsertSupplementalFormula(U, LU, LF.LUIdx); 2968 CountRegisters(LU.Formulae.back(), Uses.size() - 1); 2969 break; 2970 } 2971 } 2972 } 2973 } 2974 2975 /// CollectSubexprs - Split S into subexpressions which can be pulled out into 2976 /// separate registers. If C is non-null, multiply each subexpression by C. 2977 static void CollectSubexprs(const SCEV *S, const SCEVConstant *C, 2978 SmallVectorImpl<const SCEV *> &Ops, 2979 const Loop *L, 2980 ScalarEvolution &SE) { 2981 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) { 2982 // Break out add operands. 2983 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end(); 2984 I != E; ++I) 2985 CollectSubexprs(*I, C, Ops, L, SE); 2986 return; 2987 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) { 2988 // Split a non-zero base out of an addrec. 2989 if (!AR->getStart()->isZero()) { 2990 CollectSubexprs(SE.getAddRecExpr(SE.getConstant(AR->getType(), 0), 2991 AR->getStepRecurrence(SE), 2992 AR->getLoop(), 2993 //FIXME: AR->getNoWrapFlags(SCEV::FlagNW) 2994 SCEV::FlagAnyWrap), 2995 C, Ops, L, SE); 2996 CollectSubexprs(AR->getStart(), C, Ops, L, SE); 2997 return; 2998 } 2999 } else if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) { 3000 // Break (C * (a + b + c)) into C*a + C*b + C*c. 3001 if (Mul->getNumOperands() == 2) 3002 if (const SCEVConstant *Op0 = 3003 dyn_cast<SCEVConstant>(Mul->getOperand(0))) { 3004 CollectSubexprs(Mul->getOperand(1), 3005 C ? cast<SCEVConstant>(SE.getMulExpr(C, Op0)) : Op0, 3006 Ops, L, SE); 3007 return; 3008 } 3009 } 3010 3011 // Otherwise use the value itself, optionally with a scale applied. 3012 Ops.push_back(C ? SE.getMulExpr(C, S) : S); 3013 } 3014 3015 /// GenerateReassociations - Split out subexpressions from adds and the bases of 3016 /// addrecs. 3017 void LSRInstance::GenerateReassociations(LSRUse &LU, unsigned LUIdx, 3018 Formula Base, 3019 unsigned Depth) { 3020 // Arbitrarily cap recursion to protect compile time. 3021 if (Depth >= 3) return; 3022 3023 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) { 3024 const SCEV *BaseReg = Base.BaseRegs[i]; 3025 3026 SmallVector<const SCEV *, 8> AddOps; 3027 CollectSubexprs(BaseReg, 0, AddOps, L, SE); 3028 3029 if (AddOps.size() == 1) continue; 3030 3031 for (SmallVectorImpl<const SCEV *>::const_iterator J = AddOps.begin(), 3032 JE = AddOps.end(); J != JE; ++J) { 3033 3034 // Loop-variant "unknown" values are uninteresting; we won't be able to 3035 // do anything meaningful with them. 3036 if (isa<SCEVUnknown>(*J) && !SE.isLoopInvariant(*J, L)) 3037 continue; 3038 3039 // Don't pull a constant into a register if the constant could be folded 3040 // into an immediate field. 3041 if (isAlwaysFoldable(*J, LU.MinOffset, LU.MaxOffset, 3042 Base.getNumRegs() > 1, 3043 LU.Kind, LU.AccessTy, TLI, SE)) 3044 continue; 3045 3046 // Collect all operands except *J. 3047 SmallVector<const SCEV *, 8> InnerAddOps 3048 (((const SmallVector<const SCEV *, 8> &)AddOps).begin(), J); 3049 InnerAddOps.append 3050 (llvm::next(J), ((const SmallVector<const SCEV *, 8> &)AddOps).end()); 3051 3052 // Don't leave just a constant behind in a register if the constant could 3053 // be folded into an immediate field. 3054 if (InnerAddOps.size() == 1 && 3055 isAlwaysFoldable(InnerAddOps[0], LU.MinOffset, LU.MaxOffset, 3056 Base.getNumRegs() > 1, 3057 LU.Kind, LU.AccessTy, TLI, SE)) 3058 continue; 3059 3060 const SCEV *InnerSum = SE.getAddExpr(InnerAddOps); 3061 if (InnerSum->isZero()) 3062 continue; 3063 Formula F = Base; 3064 3065 // Add the remaining pieces of the add back into the new formula. 3066 const SCEVConstant *InnerSumSC = dyn_cast<SCEVConstant>(InnerSum); 3067 if (TLI && InnerSumSC && 3068 SE.getTypeSizeInBits(InnerSumSC->getType()) <= 64 && 3069 TLI->isLegalAddImmediate((uint64_t)F.UnfoldedOffset + 3070 InnerSumSC->getValue()->getZExtValue())) { 3071 F.UnfoldedOffset = (uint64_t)F.UnfoldedOffset + 3072 InnerSumSC->getValue()->getZExtValue(); 3073 F.BaseRegs.erase(F.BaseRegs.begin() + i); 3074 } else 3075 F.BaseRegs[i] = InnerSum; 3076 3077 // Add J as its own register, or an unfolded immediate. 3078 const SCEVConstant *SC = dyn_cast<SCEVConstant>(*J); 3079 if (TLI && SC && SE.getTypeSizeInBits(SC->getType()) <= 64 && 3080 TLI->isLegalAddImmediate((uint64_t)F.UnfoldedOffset + 3081 SC->getValue()->getZExtValue())) 3082 F.UnfoldedOffset = (uint64_t)F.UnfoldedOffset + 3083 SC->getValue()->getZExtValue(); 3084 else 3085 F.BaseRegs.push_back(*J); 3086 3087 if (InsertFormula(LU, LUIdx, F)) 3088 // If that formula hadn't been seen before, recurse to find more like 3089 // it. 3090 GenerateReassociations(LU, LUIdx, LU.Formulae.back(), Depth+1); 3091 } 3092 } 3093 } 3094 3095 /// GenerateCombinations - Generate a formula consisting of all of the 3096 /// loop-dominating registers added into a single register. 3097 void LSRInstance::GenerateCombinations(LSRUse &LU, unsigned LUIdx, 3098 Formula Base) { 3099 // This method is only interesting on a plurality of registers. 3100 if (Base.BaseRegs.size() <= 1) return; 3101 3102 Formula F = Base; 3103 F.BaseRegs.clear(); 3104 SmallVector<const SCEV *, 4> Ops; 3105 for (SmallVectorImpl<const SCEV *>::const_iterator 3106 I = Base.BaseRegs.begin(), E = Base.BaseRegs.end(); I != E; ++I) { 3107 const SCEV *BaseReg = *I; 3108 if (SE.properlyDominates(BaseReg, L->getHeader()) && 3109 !SE.hasComputableLoopEvolution(BaseReg, L)) 3110 Ops.push_back(BaseReg); 3111 else 3112 F.BaseRegs.push_back(BaseReg); 3113 } 3114 if (Ops.size() > 1) { 3115 const SCEV *Sum = SE.getAddExpr(Ops); 3116 // TODO: If Sum is zero, it probably means ScalarEvolution missed an 3117 // opportunity to fold something. For now, just ignore such cases 3118 // rather than proceed with zero in a register. 3119 if (!Sum->isZero()) { 3120 F.BaseRegs.push_back(Sum); 3121 (void)InsertFormula(LU, LUIdx, F); 3122 } 3123 } 3124 } 3125 3126 /// GenerateSymbolicOffsets - Generate reuse formulae using symbolic offsets. 3127 void LSRInstance::GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx, 3128 Formula Base) { 3129 // We can't add a symbolic offset if the address already contains one. 3130 if (Base.AM.BaseGV) return; 3131 3132 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) { 3133 const SCEV *G = Base.BaseRegs[i]; 3134 GlobalValue *GV = ExtractSymbol(G, SE); 3135 if (G->isZero() || !GV) 3136 continue; 3137 Formula F = Base; 3138 F.AM.BaseGV = GV; 3139 if (!isLegalUse(F.AM, LU.MinOffset, LU.MaxOffset, 3140 LU.Kind, LU.AccessTy, TLI)) 3141 continue; 3142 F.BaseRegs[i] = G; 3143 (void)InsertFormula(LU, LUIdx, F); 3144 } 3145 } 3146 3147 /// GenerateConstantOffsets - Generate reuse formulae using symbolic offsets. 3148 void LSRInstance::GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx, 3149 Formula Base) { 3150 // TODO: For now, just add the min and max offset, because it usually isn't 3151 // worthwhile looking at everything inbetween. 3152 SmallVector<int64_t, 2> Worklist; 3153 Worklist.push_back(LU.MinOffset); 3154 if (LU.MaxOffset != LU.MinOffset) 3155 Worklist.push_back(LU.MaxOffset); 3156 3157 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) { 3158 const SCEV *G = Base.BaseRegs[i]; 3159 3160 for (SmallVectorImpl<int64_t>::const_iterator I = Worklist.begin(), 3161 E = Worklist.end(); I != E; ++I) { 3162 Formula F = Base; 3163 F.AM.BaseOffs = (uint64_t)Base.AM.BaseOffs - *I; 3164 if (isLegalUse(F.AM, LU.MinOffset - *I, LU.MaxOffset - *I, 3165 LU.Kind, LU.AccessTy, TLI)) { 3166 // Add the offset to the base register. 3167 const SCEV *NewG = SE.getAddExpr(SE.getConstant(G->getType(), *I), G); 3168 // If it cancelled out, drop the base register, otherwise update it. 3169 if (NewG->isZero()) { 3170 std::swap(F.BaseRegs[i], F.BaseRegs.back()); 3171 F.BaseRegs.pop_back(); 3172 } else 3173 F.BaseRegs[i] = NewG; 3174 3175 (void)InsertFormula(LU, LUIdx, F); 3176 } 3177 } 3178 3179 int64_t Imm = ExtractImmediate(G, SE); 3180 if (G->isZero() || Imm == 0) 3181 continue; 3182 Formula F = Base; 3183 F.AM.BaseOffs = (uint64_t)F.AM.BaseOffs + Imm; 3184 if (!isLegalUse(F.AM, LU.MinOffset, LU.MaxOffset, 3185 LU.Kind, LU.AccessTy, TLI)) 3186 continue; 3187 F.BaseRegs[i] = G; 3188 (void)InsertFormula(LU, LUIdx, F); 3189 } 3190 } 3191 3192 /// GenerateICmpZeroScales - For ICmpZero, check to see if we can scale up 3193 /// the comparison. For example, x == y -> x*c == y*c. 3194 void LSRInstance::GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx, 3195 Formula Base) { 3196 if (LU.Kind != LSRUse::ICmpZero) return; 3197 3198 // Determine the integer type for the base formula. 3199 Type *IntTy = Base.getType(); 3200 if (!IntTy) return; 3201 if (SE.getTypeSizeInBits(IntTy) > 64) return; 3202 3203 // Don't do this if there is more than one offset. 3204 if (LU.MinOffset != LU.MaxOffset) return; 3205 3206 assert(!Base.AM.BaseGV && "ICmpZero use is not legal!"); 3207 3208 // Check each interesting stride. 3209 for (SmallSetVector<int64_t, 8>::const_iterator 3210 I = Factors.begin(), E = Factors.end(); I != E; ++I) { 3211 int64_t Factor = *I; 3212 3213 // Check that the multiplication doesn't overflow. 3214 if (Base.AM.BaseOffs == INT64_MIN && Factor == -1) 3215 continue; 3216 int64_t NewBaseOffs = (uint64_t)Base.AM.BaseOffs * Factor; 3217 if (NewBaseOffs / Factor != Base.AM.BaseOffs) 3218 continue; 3219 3220 // Check that multiplying with the use offset doesn't overflow. 3221 int64_t Offset = LU.MinOffset; 3222 if (Offset == INT64_MIN && Factor == -1) 3223 continue; 3224 Offset = (uint64_t)Offset * Factor; 3225 if (Offset / Factor != LU.MinOffset) 3226 continue; 3227 3228 Formula F = Base; 3229 F.AM.BaseOffs = NewBaseOffs; 3230 3231 // Check that this scale is legal. 3232 if (!isLegalUse(F.AM, Offset, Offset, LU.Kind, LU.AccessTy, TLI)) 3233 continue; 3234 3235 // Compensate for the use having MinOffset built into it. 3236 F.AM.BaseOffs = (uint64_t)F.AM.BaseOffs + Offset - LU.MinOffset; 3237 3238 const SCEV *FactorS = SE.getConstant(IntTy, Factor); 3239 3240 // Check that multiplying with each base register doesn't overflow. 3241 for (size_t i = 0, e = F.BaseRegs.size(); i != e; ++i) { 3242 F.BaseRegs[i] = SE.getMulExpr(F.BaseRegs[i], FactorS); 3243 if (getExactSDiv(F.BaseRegs[i], FactorS, SE) != Base.BaseRegs[i]) 3244 goto next; 3245 } 3246 3247 // Check that multiplying with the scaled register doesn't overflow. 3248 if (F.ScaledReg) { 3249 F.ScaledReg = SE.getMulExpr(F.ScaledReg, FactorS); 3250 if (getExactSDiv(F.ScaledReg, FactorS, SE) != Base.ScaledReg) 3251 continue; 3252 } 3253 3254 // Check that multiplying with the unfolded offset doesn't overflow. 3255 if (F.UnfoldedOffset != 0) { 3256 if (F.UnfoldedOffset == INT64_MIN && Factor == -1) 3257 continue; 3258 F.UnfoldedOffset = (uint64_t)F.UnfoldedOffset * Factor; 3259 if (F.UnfoldedOffset / Factor != Base.UnfoldedOffset) 3260 continue; 3261 } 3262 3263 // If we make it here and it's legal, add it. 3264 (void)InsertFormula(LU, LUIdx, F); 3265 next:; 3266 } 3267 } 3268 3269 /// GenerateScales - Generate stride factor reuse formulae by making use of 3270 /// scaled-offset address modes, for example. 3271 void LSRInstance::GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base) { 3272 // Determine the integer type for the base formula. 3273 Type *IntTy = Base.getType(); 3274 if (!IntTy) return; 3275 3276 // If this Formula already has a scaled register, we can't add another one. 3277 if (Base.AM.Scale != 0) return; 3278 3279 // Check each interesting stride. 3280 for (SmallSetVector<int64_t, 8>::const_iterator 3281 I = Factors.begin(), E = Factors.end(); I != E; ++I) { 3282 int64_t Factor = *I; 3283 3284 Base.AM.Scale = Factor; 3285 Base.AM.HasBaseReg = Base.BaseRegs.size() > 1; 3286 // Check whether this scale is going to be legal. 3287 if (!isLegalUse(Base.AM, LU.MinOffset, LU.MaxOffset, 3288 LU.Kind, LU.AccessTy, TLI)) { 3289 // As a special-case, handle special out-of-loop Basic users specially. 3290 // TODO: Reconsider this special case. 3291 if (LU.Kind == LSRUse::Basic && 3292 isLegalUse(Base.AM, LU.MinOffset, LU.MaxOffset, 3293 LSRUse::Special, LU.AccessTy, TLI) && 3294 LU.AllFixupsOutsideLoop) 3295 LU.Kind = LSRUse::Special; 3296 else 3297 continue; 3298 } 3299 // For an ICmpZero, negating a solitary base register won't lead to 3300 // new solutions. 3301 if (LU.Kind == LSRUse::ICmpZero && 3302 !Base.AM.HasBaseReg && Base.AM.BaseOffs == 0 && !Base.AM.BaseGV) 3303 continue; 3304 // For each addrec base reg, apply the scale, if possible. 3305 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) 3306 if (const SCEVAddRecExpr *AR = 3307 dyn_cast<SCEVAddRecExpr>(Base.BaseRegs[i])) { 3308 const SCEV *FactorS = SE.getConstant(IntTy, Factor); 3309 if (FactorS->isZero()) 3310 continue; 3311 // Divide out the factor, ignoring high bits, since we'll be 3312 // scaling the value back up in the end. 3313 if (const SCEV *Quotient = getExactSDiv(AR, FactorS, SE, true)) { 3314 // TODO: This could be optimized to avoid all the copying. 3315 Formula F = Base; 3316 F.ScaledReg = Quotient; 3317 F.DeleteBaseReg(F.BaseRegs[i]); 3318 (void)InsertFormula(LU, LUIdx, F); 3319 } 3320 } 3321 } 3322 } 3323 3324 /// GenerateTruncates - Generate reuse formulae from different IV types. 3325 void LSRInstance::GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base) { 3326 // This requires TargetLowering to tell us which truncates are free. 3327 if (!TLI) return; 3328 3329 // Don't bother truncating symbolic values. 3330 if (Base.AM.BaseGV) return; 3331 3332 // Determine the integer type for the base formula. 3333 Type *DstTy = Base.getType(); 3334 if (!DstTy) return; 3335 DstTy = SE.getEffectiveSCEVType(DstTy); 3336 3337 for (SmallSetVector<Type *, 4>::const_iterator 3338 I = Types.begin(), E = Types.end(); I != E; ++I) { 3339 Type *SrcTy = *I; 3340 if (SrcTy != DstTy && TLI->isTruncateFree(SrcTy, DstTy)) { 3341 Formula F = Base; 3342 3343 if (F.ScaledReg) F.ScaledReg = SE.getAnyExtendExpr(F.ScaledReg, *I); 3344 for (SmallVectorImpl<const SCEV *>::iterator J = F.BaseRegs.begin(), 3345 JE = F.BaseRegs.end(); J != JE; ++J) 3346 *J = SE.getAnyExtendExpr(*J, SrcTy); 3347 3348 // TODO: This assumes we've done basic processing on all uses and 3349 // have an idea what the register usage is. 3350 if (!F.hasRegsUsedByUsesOtherThan(LUIdx, RegUses)) 3351 continue; 3352 3353 (void)InsertFormula(LU, LUIdx, F); 3354 } 3355 } 3356 } 3357 3358 namespace { 3359 3360 /// WorkItem - Helper class for GenerateCrossUseConstantOffsets. It's used to 3361 /// defer modifications so that the search phase doesn't have to worry about 3362 /// the data structures moving underneath it. 3363 struct WorkItem { 3364 size_t LUIdx; 3365 int64_t Imm; 3366 const SCEV *OrigReg; 3367 3368 WorkItem(size_t LI, int64_t I, const SCEV *R) 3369 : LUIdx(LI), Imm(I), OrigReg(R) {} 3370 3371 void print(raw_ostream &OS) const; 3372 void dump() const; 3373 }; 3374 3375 } 3376 3377 void WorkItem::print(raw_ostream &OS) const { 3378 OS << "in formulae referencing " << *OrigReg << " in use " << LUIdx 3379 << " , add offset " << Imm; 3380 } 3381 3382 void WorkItem::dump() const { 3383 print(errs()); errs() << '\n'; 3384 } 3385 3386 /// GenerateCrossUseConstantOffsets - Look for registers which are a constant 3387 /// distance apart and try to form reuse opportunities between them. 3388 void LSRInstance::GenerateCrossUseConstantOffsets() { 3389 // Group the registers by their value without any added constant offset. 3390 typedef std::map<int64_t, const SCEV *> ImmMapTy; 3391 typedef DenseMap<const SCEV *, ImmMapTy> RegMapTy; 3392 RegMapTy Map; 3393 DenseMap<const SCEV *, SmallBitVector> UsedByIndicesMap; 3394 SmallVector<const SCEV *, 8> Sequence; 3395 for (RegUseTracker::const_iterator I = RegUses.begin(), E = RegUses.end(); 3396 I != E; ++I) { 3397 const SCEV *Reg = *I; 3398 int64_t Imm = ExtractImmediate(Reg, SE); 3399 std::pair<RegMapTy::iterator, bool> Pair = 3400 Map.insert(std::make_pair(Reg, ImmMapTy())); 3401 if (Pair.second) 3402 Sequence.push_back(Reg); 3403 Pair.first->second.insert(std::make_pair(Imm, *I)); 3404 UsedByIndicesMap[Reg] |= RegUses.getUsedByIndices(*I); 3405 } 3406 3407 // Now examine each set of registers with the same base value. Build up 3408 // a list of work to do and do the work in a separate step so that we're 3409 // not adding formulae and register counts while we're searching. 3410 SmallVector<WorkItem, 32> WorkItems; 3411 SmallSet<std::pair<size_t, int64_t>, 32> UniqueItems; 3412 for (SmallVectorImpl<const SCEV *>::const_iterator I = Sequence.begin(), 3413 E = Sequence.end(); I != E; ++I) { 3414 const SCEV *Reg = *I; 3415 const ImmMapTy &Imms = Map.find(Reg)->second; 3416 3417 // It's not worthwhile looking for reuse if there's only one offset. 3418 if (Imms.size() == 1) 3419 continue; 3420 3421 DEBUG(dbgs() << "Generating cross-use offsets for " << *Reg << ':'; 3422 for (ImmMapTy::const_iterator J = Imms.begin(), JE = Imms.end(); 3423 J != JE; ++J) 3424 dbgs() << ' ' << J->first; 3425 dbgs() << '\n'); 3426 3427 // Examine each offset. 3428 for (ImmMapTy::const_iterator J = Imms.begin(), JE = Imms.end(); 3429 J != JE; ++J) { 3430 const SCEV *OrigReg = J->second; 3431 3432 int64_t JImm = J->first; 3433 const SmallBitVector &UsedByIndices = RegUses.getUsedByIndices(OrigReg); 3434 3435 if (!isa<SCEVConstant>(OrigReg) && 3436 UsedByIndicesMap[Reg].count() == 1) { 3437 DEBUG(dbgs() << "Skipping cross-use reuse for " << *OrigReg << '\n'); 3438 continue; 3439 } 3440 3441 // Conservatively examine offsets between this orig reg a few selected 3442 // other orig regs. 3443 ImmMapTy::const_iterator OtherImms[] = { 3444 Imms.begin(), prior(Imms.end()), 3445 Imms.lower_bound((Imms.begin()->first + prior(Imms.end())->first) / 2) 3446 }; 3447 for (size_t i = 0, e = array_lengthof(OtherImms); i != e; ++i) { 3448 ImmMapTy::const_iterator M = OtherImms[i]; 3449 if (M == J || M == JE) continue; 3450 3451 // Compute the difference between the two. 3452 int64_t Imm = (uint64_t)JImm - M->first; 3453 for (int LUIdx = UsedByIndices.find_first(); LUIdx != -1; 3454 LUIdx = UsedByIndices.find_next(LUIdx)) 3455 // Make a memo of this use, offset, and register tuple. 3456 if (UniqueItems.insert(std::make_pair(LUIdx, Imm))) 3457 WorkItems.push_back(WorkItem(LUIdx, Imm, OrigReg)); 3458 } 3459 } 3460 } 3461 3462 Map.clear(); 3463 Sequence.clear(); 3464 UsedByIndicesMap.clear(); 3465 UniqueItems.clear(); 3466 3467 // Now iterate through the worklist and add new formulae. 3468 for (SmallVectorImpl<WorkItem>::const_iterator I = WorkItems.begin(), 3469 E = WorkItems.end(); I != E; ++I) { 3470 const WorkItem &WI = *I; 3471 size_t LUIdx = WI.LUIdx; 3472 LSRUse &LU = Uses[LUIdx]; 3473 int64_t Imm = WI.Imm; 3474 const SCEV *OrigReg = WI.OrigReg; 3475 3476 Type *IntTy = SE.getEffectiveSCEVType(OrigReg->getType()); 3477 const SCEV *NegImmS = SE.getSCEV(ConstantInt::get(IntTy, -(uint64_t)Imm)); 3478 unsigned BitWidth = SE.getTypeSizeInBits(IntTy); 3479 3480 // TODO: Use a more targeted data structure. 3481 for (size_t L = 0, LE = LU.Formulae.size(); L != LE; ++L) { 3482 const Formula &F = LU.Formulae[L]; 3483 // Use the immediate in the scaled register. 3484 if (F.ScaledReg == OrigReg) { 3485 int64_t Offs = (uint64_t)F.AM.BaseOffs + 3486 Imm * (uint64_t)F.AM.Scale; 3487 // Don't create 50 + reg(-50). 3488 if (F.referencesReg(SE.getSCEV( 3489 ConstantInt::get(IntTy, -(uint64_t)Offs)))) 3490 continue; 3491 Formula NewF = F; 3492 NewF.AM.BaseOffs = Offs; 3493 if (!isLegalUse(NewF.AM, LU.MinOffset, LU.MaxOffset, 3494 LU.Kind, LU.AccessTy, TLI)) 3495 continue; 3496 NewF.ScaledReg = SE.getAddExpr(NegImmS, NewF.ScaledReg); 3497 3498 // If the new scale is a constant in a register, and adding the constant 3499 // value to the immediate would produce a value closer to zero than the 3500 // immediate itself, then the formula isn't worthwhile. 3501 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(NewF.ScaledReg)) 3502 if (C->getValue()->isNegative() != 3503 (NewF.AM.BaseOffs < 0) && 3504 (C->getValue()->getValue().abs() * APInt(BitWidth, F.AM.Scale)) 3505 .ule(abs64(NewF.AM.BaseOffs))) 3506 continue; 3507 3508 // OK, looks good. 3509 (void)InsertFormula(LU, LUIdx, NewF); 3510 } else { 3511 // Use the immediate in a base register. 3512 for (size_t N = 0, NE = F.BaseRegs.size(); N != NE; ++N) { 3513 const SCEV *BaseReg = F.BaseRegs[N]; 3514 if (BaseReg != OrigReg) 3515 continue; 3516 Formula NewF = F; 3517 NewF.AM.BaseOffs = (uint64_t)NewF.AM.BaseOffs + Imm; 3518 if (!isLegalUse(NewF.AM, LU.MinOffset, LU.MaxOffset, 3519 LU.Kind, LU.AccessTy, TLI)) { 3520 if (!TLI || 3521 !TLI->isLegalAddImmediate((uint64_t)NewF.UnfoldedOffset + Imm)) 3522 continue; 3523 NewF = F; 3524 NewF.UnfoldedOffset = (uint64_t)NewF.UnfoldedOffset + Imm; 3525 } 3526 NewF.BaseRegs[N] = SE.getAddExpr(NegImmS, BaseReg); 3527 3528 // If the new formula has a constant in a register, and adding the 3529 // constant value to the immediate would produce a value closer to 3530 // zero than the immediate itself, then the formula isn't worthwhile. 3531 for (SmallVectorImpl<const SCEV *>::const_iterator 3532 J = NewF.BaseRegs.begin(), JE = NewF.BaseRegs.end(); 3533 J != JE; ++J) 3534 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(*J)) 3535 if ((C->getValue()->getValue() + NewF.AM.BaseOffs).abs().slt( 3536 abs64(NewF.AM.BaseOffs)) && 3537 (C->getValue()->getValue() + 3538 NewF.AM.BaseOffs).countTrailingZeros() >= 3539 CountTrailingZeros_64(NewF.AM.BaseOffs)) 3540 goto skip_formula; 3541 3542 // Ok, looks good. 3543 (void)InsertFormula(LU, LUIdx, NewF); 3544 break; 3545 skip_formula:; 3546 } 3547 } 3548 } 3549 } 3550 } 3551 3552 /// GenerateAllReuseFormulae - Generate formulae for each use. 3553 void 3554 LSRInstance::GenerateAllReuseFormulae() { 3555 // This is split into multiple loops so that hasRegsUsedByUsesOtherThan 3556 // queries are more precise. 3557 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) { 3558 LSRUse &LU = Uses[LUIdx]; 3559 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i) 3560 GenerateReassociations(LU, LUIdx, LU.Formulae[i]); 3561 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i) 3562 GenerateCombinations(LU, LUIdx, LU.Formulae[i]); 3563 } 3564 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) { 3565 LSRUse &LU = Uses[LUIdx]; 3566 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i) 3567 GenerateSymbolicOffsets(LU, LUIdx, LU.Formulae[i]); 3568 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i) 3569 GenerateConstantOffsets(LU, LUIdx, LU.Formulae[i]); 3570 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i) 3571 GenerateICmpZeroScales(LU, LUIdx, LU.Formulae[i]); 3572 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i) 3573 GenerateScales(LU, LUIdx, LU.Formulae[i]); 3574 } 3575 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) { 3576 LSRUse &LU = Uses[LUIdx]; 3577 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i) 3578 GenerateTruncates(LU, LUIdx, LU.Formulae[i]); 3579 } 3580 3581 GenerateCrossUseConstantOffsets(); 3582 3583 DEBUG(dbgs() << "\n" 3584 "After generating reuse formulae:\n"; 3585 print_uses(dbgs())); 3586 } 3587 3588 /// If there are multiple formulae with the same set of registers used 3589 /// by other uses, pick the best one and delete the others. 3590 void LSRInstance::FilterOutUndesirableDedicatedRegisters() { 3591 DenseSet<const SCEV *> VisitedRegs; 3592 SmallPtrSet<const SCEV *, 16> Regs; 3593 SmallPtrSet<const SCEV *, 16> LoserRegs; 3594 #ifndef NDEBUG 3595 bool ChangedFormulae = false; 3596 #endif 3597 3598 // Collect the best formula for each unique set of shared registers. This 3599 // is reset for each use. 3600 typedef DenseMap<SmallVector<const SCEV *, 2>, size_t, UniquifierDenseMapInfo> 3601 BestFormulaeTy; 3602 BestFormulaeTy BestFormulae; 3603 3604 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) { 3605 LSRUse &LU = Uses[LUIdx]; 3606 DEBUG(dbgs() << "Filtering for use "; LU.print(dbgs()); dbgs() << '\n'); 3607 3608 bool Any = false; 3609 for (size_t FIdx = 0, NumForms = LU.Formulae.size(); 3610 FIdx != NumForms; ++FIdx) { 3611 Formula &F = LU.Formulae[FIdx]; 3612 3613 // Some formulas are instant losers. For example, they may depend on 3614 // nonexistent AddRecs from other loops. These need to be filtered 3615 // immediately, otherwise heuristics could choose them over others leading 3616 // to an unsatisfactory solution. Passing LoserRegs into RateFormula here 3617 // avoids the need to recompute this information across formulae using the 3618 // same bad AddRec. Passing LoserRegs is also essential unless we remove 3619 // the corresponding bad register from the Regs set. 3620 Cost CostF; 3621 Regs.clear(); 3622 CostF.RateFormula(F, Regs, VisitedRegs, L, LU.Offsets, SE, DT, 3623 &LoserRegs); 3624 if (CostF.isLoser()) { 3625 // During initial formula generation, undesirable formulae are generated 3626 // by uses within other loops that have some non-trivial address mode or 3627 // use the postinc form of the IV. LSR needs to provide these formulae 3628 // as the basis of rediscovering the desired formula that uses an AddRec 3629 // corresponding to the existing phi. Once all formulae have been 3630 // generated, these initial losers may be pruned. 3631 DEBUG(dbgs() << " Filtering loser "; F.print(dbgs()); 3632 dbgs() << "\n"); 3633 } 3634 else { 3635 SmallVector<const SCEV *, 2> Key; 3636 for (SmallVectorImpl<const SCEV *>::const_iterator J = F.BaseRegs.begin(), 3637 JE = F.BaseRegs.end(); J != JE; ++J) { 3638 const SCEV *Reg = *J; 3639 if (RegUses.isRegUsedByUsesOtherThan(Reg, LUIdx)) 3640 Key.push_back(Reg); 3641 } 3642 if (F.ScaledReg && 3643 RegUses.isRegUsedByUsesOtherThan(F.ScaledReg, LUIdx)) 3644 Key.push_back(F.ScaledReg); 3645 // Unstable sort by host order ok, because this is only used for 3646 // uniquifying. 3647 std::sort(Key.begin(), Key.end()); 3648 3649 std::pair<BestFormulaeTy::const_iterator, bool> P = 3650 BestFormulae.insert(std::make_pair(Key, FIdx)); 3651 if (P.second) 3652 continue; 3653 3654 Formula &Best = LU.Formulae[P.first->second]; 3655 3656 Cost CostBest; 3657 Regs.clear(); 3658 CostBest.RateFormula(Best, Regs, VisitedRegs, L, LU.Offsets, SE, DT); 3659 if (CostF < CostBest) 3660 std::swap(F, Best); 3661 DEBUG(dbgs() << " Filtering out formula "; F.print(dbgs()); 3662 dbgs() << "\n" 3663 " in favor of formula "; Best.print(dbgs()); 3664 dbgs() << '\n'); 3665 } 3666 #ifndef NDEBUG 3667 ChangedFormulae = true; 3668 #endif 3669 LU.DeleteFormula(F); 3670 --FIdx; 3671 --NumForms; 3672 Any = true; 3673 } 3674 3675 // Now that we've filtered out some formulae, recompute the Regs set. 3676 if (Any) 3677 LU.RecomputeRegs(LUIdx, RegUses); 3678 3679 // Reset this to prepare for the next use. 3680 BestFormulae.clear(); 3681 } 3682 3683 DEBUG(if (ChangedFormulae) { 3684 dbgs() << "\n" 3685 "After filtering out undesirable candidates:\n"; 3686 print_uses(dbgs()); 3687 }); 3688 } 3689 3690 // This is a rough guess that seems to work fairly well. 3691 static const size_t ComplexityLimit = UINT16_MAX; 3692 3693 /// EstimateSearchSpaceComplexity - Estimate the worst-case number of 3694 /// solutions the solver might have to consider. It almost never considers 3695 /// this many solutions because it prune the search space, but the pruning 3696 /// isn't always sufficient. 3697 size_t LSRInstance::EstimateSearchSpaceComplexity() const { 3698 size_t Power = 1; 3699 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(), 3700 E = Uses.end(); I != E; ++I) { 3701 size_t FSize = I->Formulae.size(); 3702 if (FSize >= ComplexityLimit) { 3703 Power = ComplexityLimit; 3704 break; 3705 } 3706 Power *= FSize; 3707 if (Power >= ComplexityLimit) 3708 break; 3709 } 3710 return Power; 3711 } 3712 3713 /// NarrowSearchSpaceByDetectingSupersets - When one formula uses a superset 3714 /// of the registers of another formula, it won't help reduce register 3715 /// pressure (though it may not necessarily hurt register pressure); remove 3716 /// it to simplify the system. 3717 void LSRInstance::NarrowSearchSpaceByDetectingSupersets() { 3718 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) { 3719 DEBUG(dbgs() << "The search space is too complex.\n"); 3720 3721 DEBUG(dbgs() << "Narrowing the search space by eliminating formulae " 3722 "which use a superset of registers used by other " 3723 "formulae.\n"); 3724 3725 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) { 3726 LSRUse &LU = Uses[LUIdx]; 3727 bool Any = false; 3728 for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) { 3729 Formula &F = LU.Formulae[i]; 3730 // Look for a formula with a constant or GV in a register. If the use 3731 // also has a formula with that same value in an immediate field, 3732 // delete the one that uses a register. 3733 for (SmallVectorImpl<const SCEV *>::const_iterator 3734 I = F.BaseRegs.begin(), E = F.BaseRegs.end(); I != E; ++I) { 3735 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(*I)) { 3736 Formula NewF = F; 3737 NewF.AM.BaseOffs += C->getValue()->getSExtValue(); 3738 NewF.BaseRegs.erase(NewF.BaseRegs.begin() + 3739 (I - F.BaseRegs.begin())); 3740 if (LU.HasFormulaWithSameRegs(NewF)) { 3741 DEBUG(dbgs() << " Deleting "; F.print(dbgs()); dbgs() << '\n'); 3742 LU.DeleteFormula(F); 3743 --i; 3744 --e; 3745 Any = true; 3746 break; 3747 } 3748 } else if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(*I)) { 3749 if (GlobalValue *GV = dyn_cast<GlobalValue>(U->getValue())) 3750 if (!F.AM.BaseGV) { 3751 Formula NewF = F; 3752 NewF.AM.BaseGV = GV; 3753 NewF.BaseRegs.erase(NewF.BaseRegs.begin() + 3754 (I - F.BaseRegs.begin())); 3755 if (LU.HasFormulaWithSameRegs(NewF)) { 3756 DEBUG(dbgs() << " Deleting "; F.print(dbgs()); 3757 dbgs() << '\n'); 3758 LU.DeleteFormula(F); 3759 --i; 3760 --e; 3761 Any = true; 3762 break; 3763 } 3764 } 3765 } 3766 } 3767 } 3768 if (Any) 3769 LU.RecomputeRegs(LUIdx, RegUses); 3770 } 3771 3772 DEBUG(dbgs() << "After pre-selection:\n"; 3773 print_uses(dbgs())); 3774 } 3775 } 3776 3777 /// NarrowSearchSpaceByCollapsingUnrolledCode - When there are many registers 3778 /// for expressions like A, A+1, A+2, etc., allocate a single register for 3779 /// them. 3780 void LSRInstance::NarrowSearchSpaceByCollapsingUnrolledCode() { 3781 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) { 3782 DEBUG(dbgs() << "The search space is too complex.\n"); 3783 3784 DEBUG(dbgs() << "Narrowing the search space by assuming that uses " 3785 "separated by a constant offset will use the same " 3786 "registers.\n"); 3787 3788 // This is especially useful for unrolled loops. 3789 3790 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) { 3791 LSRUse &LU = Uses[LUIdx]; 3792 for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(), 3793 E = LU.Formulae.end(); I != E; ++I) { 3794 const Formula &F = *I; 3795 if (F.AM.BaseOffs != 0 && F.AM.Scale == 0) { 3796 if (LSRUse *LUThatHas = FindUseWithSimilarFormula(F, LU)) { 3797 if (reconcileNewOffset(*LUThatHas, F.AM.BaseOffs, 3798 /*HasBaseReg=*/false, 3799 LU.Kind, LU.AccessTy)) { 3800 DEBUG(dbgs() << " Deleting use "; LU.print(dbgs()); 3801 dbgs() << '\n'); 3802 3803 LUThatHas->AllFixupsOutsideLoop &= LU.AllFixupsOutsideLoop; 3804 3805 // Update the relocs to reference the new use. 3806 for (SmallVectorImpl<LSRFixup>::iterator I = Fixups.begin(), 3807 E = Fixups.end(); I != E; ++I) { 3808 LSRFixup &Fixup = *I; 3809 if (Fixup.LUIdx == LUIdx) { 3810 Fixup.LUIdx = LUThatHas - &Uses.front(); 3811 Fixup.Offset += F.AM.BaseOffs; 3812 // Add the new offset to LUThatHas' offset list. 3813 if (LUThatHas->Offsets.back() != Fixup.Offset) { 3814 LUThatHas->Offsets.push_back(Fixup.Offset); 3815 if (Fixup.Offset > LUThatHas->MaxOffset) 3816 LUThatHas->MaxOffset = Fixup.Offset; 3817 if (Fixup.Offset < LUThatHas->MinOffset) 3818 LUThatHas->MinOffset = Fixup.Offset; 3819 } 3820 DEBUG(dbgs() << "New fixup has offset " 3821 << Fixup.Offset << '\n'); 3822 } 3823 if (Fixup.LUIdx == NumUses-1) 3824 Fixup.LUIdx = LUIdx; 3825 } 3826 3827 // Delete formulae from the new use which are no longer legal. 3828 bool Any = false; 3829 for (size_t i = 0, e = LUThatHas->Formulae.size(); i != e; ++i) { 3830 Formula &F = LUThatHas->Formulae[i]; 3831 if (!isLegalUse(F.AM, 3832 LUThatHas->MinOffset, LUThatHas->MaxOffset, 3833 LUThatHas->Kind, LUThatHas->AccessTy, TLI)) { 3834 DEBUG(dbgs() << " Deleting "; F.print(dbgs()); 3835 dbgs() << '\n'); 3836 LUThatHas->DeleteFormula(F); 3837 --i; 3838 --e; 3839 Any = true; 3840 } 3841 } 3842 if (Any) 3843 LUThatHas->RecomputeRegs(LUThatHas - &Uses.front(), RegUses); 3844 3845 // Delete the old use. 3846 DeleteUse(LU, LUIdx); 3847 --LUIdx; 3848 --NumUses; 3849 break; 3850 } 3851 } 3852 } 3853 } 3854 } 3855 3856 DEBUG(dbgs() << "After pre-selection:\n"; 3857 print_uses(dbgs())); 3858 } 3859 } 3860 3861 /// NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters - Call 3862 /// FilterOutUndesirableDedicatedRegisters again, if necessary, now that 3863 /// we've done more filtering, as it may be able to find more formulae to 3864 /// eliminate. 3865 void LSRInstance::NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters(){ 3866 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) { 3867 DEBUG(dbgs() << "The search space is too complex.\n"); 3868 3869 DEBUG(dbgs() << "Narrowing the search space by re-filtering out " 3870 "undesirable dedicated registers.\n"); 3871 3872 FilterOutUndesirableDedicatedRegisters(); 3873 3874 DEBUG(dbgs() << "After pre-selection:\n"; 3875 print_uses(dbgs())); 3876 } 3877 } 3878 3879 /// NarrowSearchSpaceByPickingWinnerRegs - Pick a register which seems likely 3880 /// to be profitable, and then in any use which has any reference to that 3881 /// register, delete all formulae which do not reference that register. 3882 void LSRInstance::NarrowSearchSpaceByPickingWinnerRegs() { 3883 // With all other options exhausted, loop until the system is simple 3884 // enough to handle. 3885 SmallPtrSet<const SCEV *, 4> Taken; 3886 while (EstimateSearchSpaceComplexity() >= ComplexityLimit) { 3887 // Ok, we have too many of formulae on our hands to conveniently handle. 3888 // Use a rough heuristic to thin out the list. 3889 DEBUG(dbgs() << "The search space is too complex.\n"); 3890 3891 // Pick the register which is used by the most LSRUses, which is likely 3892 // to be a good reuse register candidate. 3893 const SCEV *Best = 0; 3894 unsigned BestNum = 0; 3895 for (RegUseTracker::const_iterator I = RegUses.begin(), E = RegUses.end(); 3896 I != E; ++I) { 3897 const SCEV *Reg = *I; 3898 if (Taken.count(Reg)) 3899 continue; 3900 if (!Best) 3901 Best = Reg; 3902 else { 3903 unsigned Count = RegUses.getUsedByIndices(Reg).count(); 3904 if (Count > BestNum) { 3905 Best = Reg; 3906 BestNum = Count; 3907 } 3908 } 3909 } 3910 3911 DEBUG(dbgs() << "Narrowing the search space by assuming " << *Best 3912 << " will yield profitable reuse.\n"); 3913 Taken.insert(Best); 3914 3915 // In any use with formulae which references this register, delete formulae 3916 // which don't reference it. 3917 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) { 3918 LSRUse &LU = Uses[LUIdx]; 3919 if (!LU.Regs.count(Best)) continue; 3920 3921 bool Any = false; 3922 for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) { 3923 Formula &F = LU.Formulae[i]; 3924 if (!F.referencesReg(Best)) { 3925 DEBUG(dbgs() << " Deleting "; F.print(dbgs()); dbgs() << '\n'); 3926 LU.DeleteFormula(F); 3927 --e; 3928 --i; 3929 Any = true; 3930 assert(e != 0 && "Use has no formulae left! Is Regs inconsistent?"); 3931 continue; 3932 } 3933 } 3934 3935 if (Any) 3936 LU.RecomputeRegs(LUIdx, RegUses); 3937 } 3938 3939 DEBUG(dbgs() << "After pre-selection:\n"; 3940 print_uses(dbgs())); 3941 } 3942 } 3943 3944 /// NarrowSearchSpaceUsingHeuristics - If there are an extraordinary number of 3945 /// formulae to choose from, use some rough heuristics to prune down the number 3946 /// of formulae. This keeps the main solver from taking an extraordinary amount 3947 /// of time in some worst-case scenarios. 3948 void LSRInstance::NarrowSearchSpaceUsingHeuristics() { 3949 NarrowSearchSpaceByDetectingSupersets(); 3950 NarrowSearchSpaceByCollapsingUnrolledCode(); 3951 NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters(); 3952 NarrowSearchSpaceByPickingWinnerRegs(); 3953 } 3954 3955 /// SolveRecurse - This is the recursive solver. 3956 void LSRInstance::SolveRecurse(SmallVectorImpl<const Formula *> &Solution, 3957 Cost &SolutionCost, 3958 SmallVectorImpl<const Formula *> &Workspace, 3959 const Cost &CurCost, 3960 const SmallPtrSet<const SCEV *, 16> &CurRegs, 3961 DenseSet<const SCEV *> &VisitedRegs) const { 3962 // Some ideas: 3963 // - prune more: 3964 // - use more aggressive filtering 3965 // - sort the formula so that the most profitable solutions are found first 3966 // - sort the uses too 3967 // - search faster: 3968 // - don't compute a cost, and then compare. compare while computing a cost 3969 // and bail early. 3970 // - track register sets with SmallBitVector 3971 3972 const LSRUse &LU = Uses[Workspace.size()]; 3973 3974 // If this use references any register that's already a part of the 3975 // in-progress solution, consider it a requirement that a formula must 3976 // reference that register in order to be considered. This prunes out 3977 // unprofitable searching. 3978 SmallSetVector<const SCEV *, 4> ReqRegs; 3979 for (SmallPtrSet<const SCEV *, 16>::const_iterator I = CurRegs.begin(), 3980 E = CurRegs.end(); I != E; ++I) 3981 if (LU.Regs.count(*I)) 3982 ReqRegs.insert(*I); 3983 3984 SmallPtrSet<const SCEV *, 16> NewRegs; 3985 Cost NewCost; 3986 for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(), 3987 E = LU.Formulae.end(); I != E; ++I) { 3988 const Formula &F = *I; 3989 3990 // Ignore formulae which do not use any of the required registers. 3991 bool SatisfiedReqReg = true; 3992 for (SmallSetVector<const SCEV *, 4>::const_iterator J = ReqRegs.begin(), 3993 JE = ReqRegs.end(); J != JE; ++J) { 3994 const SCEV *Reg = *J; 3995 if ((!F.ScaledReg || F.ScaledReg != Reg) && 3996 std::find(F.BaseRegs.begin(), F.BaseRegs.end(), Reg) == 3997 F.BaseRegs.end()) { 3998 SatisfiedReqReg = false; 3999 break; 4000 } 4001 } 4002 if (!SatisfiedReqReg) { 4003 // If none of the formulae satisfied the required registers, then we could 4004 // clear ReqRegs and try again. Currently, we simply give up in this case. 4005 continue; 4006 } 4007 4008 // Evaluate the cost of the current formula. If it's already worse than 4009 // the current best, prune the search at that point. 4010 NewCost = CurCost; 4011 NewRegs = CurRegs; 4012 NewCost.RateFormula(F, NewRegs, VisitedRegs, L, LU.Offsets, SE, DT); 4013 if (NewCost < SolutionCost) { 4014 Workspace.push_back(&F); 4015 if (Workspace.size() != Uses.size()) { 4016 SolveRecurse(Solution, SolutionCost, Workspace, NewCost, 4017 NewRegs, VisitedRegs); 4018 if (F.getNumRegs() == 1 && Workspace.size() == 1) 4019 VisitedRegs.insert(F.ScaledReg ? F.ScaledReg : F.BaseRegs[0]); 4020 } else { 4021 DEBUG(dbgs() << "New best at "; NewCost.print(dbgs()); 4022 dbgs() << ".\n Regs:"; 4023 for (SmallPtrSet<const SCEV *, 16>::const_iterator 4024 I = NewRegs.begin(), E = NewRegs.end(); I != E; ++I) 4025 dbgs() << ' ' << **I; 4026 dbgs() << '\n'); 4027 4028 SolutionCost = NewCost; 4029 Solution = Workspace; 4030 } 4031 Workspace.pop_back(); 4032 } 4033 } 4034 } 4035 4036 /// Solve - Choose one formula from each use. Return the results in the given 4037 /// Solution vector. 4038 void LSRInstance::Solve(SmallVectorImpl<const Formula *> &Solution) const { 4039 SmallVector<const Formula *, 8> Workspace; 4040 Cost SolutionCost; 4041 SolutionCost.Loose(); 4042 Cost CurCost; 4043 SmallPtrSet<const SCEV *, 16> CurRegs; 4044 DenseSet<const SCEV *> VisitedRegs; 4045 Workspace.reserve(Uses.size()); 4046 4047 // SolveRecurse does all the work. 4048 SolveRecurse(Solution, SolutionCost, Workspace, CurCost, 4049 CurRegs, VisitedRegs); 4050 if (Solution.empty()) { 4051 DEBUG(dbgs() << "\nNo Satisfactory Solution\n"); 4052 return; 4053 } 4054 4055 // Ok, we've now made all our decisions. 4056 DEBUG(dbgs() << "\n" 4057 "The chosen solution requires "; SolutionCost.print(dbgs()); 4058 dbgs() << ":\n"; 4059 for (size_t i = 0, e = Uses.size(); i != e; ++i) { 4060 dbgs() << " "; 4061 Uses[i].print(dbgs()); 4062 dbgs() << "\n" 4063 " "; 4064 Solution[i]->print(dbgs()); 4065 dbgs() << '\n'; 4066 }); 4067 4068 assert(Solution.size() == Uses.size() && "Malformed solution!"); 4069 } 4070 4071 /// HoistInsertPosition - Helper for AdjustInsertPositionForExpand. Climb up 4072 /// the dominator tree far as we can go while still being dominated by the 4073 /// input positions. This helps canonicalize the insert position, which 4074 /// encourages sharing. 4075 BasicBlock::iterator 4076 LSRInstance::HoistInsertPosition(BasicBlock::iterator IP, 4077 const SmallVectorImpl<Instruction *> &Inputs) 4078 const { 4079 for (;;) { 4080 const Loop *IPLoop = LI.getLoopFor(IP->getParent()); 4081 unsigned IPLoopDepth = IPLoop ? IPLoop->getLoopDepth() : 0; 4082 4083 BasicBlock *IDom; 4084 for (DomTreeNode *Rung = DT.getNode(IP->getParent()); ; ) { 4085 if (!Rung) return IP; 4086 Rung = Rung->getIDom(); 4087 if (!Rung) return IP; 4088 IDom = Rung->getBlock(); 4089 4090 // Don't climb into a loop though. 4091 const Loop *IDomLoop = LI.getLoopFor(IDom); 4092 unsigned IDomDepth = IDomLoop ? IDomLoop->getLoopDepth() : 0; 4093 if (IDomDepth <= IPLoopDepth && 4094 (IDomDepth != IPLoopDepth || IDomLoop == IPLoop)) 4095 break; 4096 } 4097 4098 bool AllDominate = true; 4099 Instruction *BetterPos = 0; 4100 Instruction *Tentative = IDom->getTerminator(); 4101 for (SmallVectorImpl<Instruction *>::const_iterator I = Inputs.begin(), 4102 E = Inputs.end(); I != E; ++I) { 4103 Instruction *Inst = *I; 4104 if (Inst == Tentative || !DT.dominates(Inst, Tentative)) { 4105 AllDominate = false; 4106 break; 4107 } 4108 // Attempt to find an insert position in the middle of the block, 4109 // instead of at the end, so that it can be used for other expansions. 4110 if (IDom == Inst->getParent() && 4111 (!BetterPos || DT.dominates(BetterPos, Inst))) 4112 BetterPos = llvm::next(BasicBlock::iterator(Inst)); 4113 } 4114 if (!AllDominate) 4115 break; 4116 if (BetterPos) 4117 IP = BetterPos; 4118 else 4119 IP = Tentative; 4120 } 4121 4122 return IP; 4123 } 4124 4125 /// AdjustInsertPositionForExpand - Determine an input position which will be 4126 /// dominated by the operands and which will dominate the result. 4127 BasicBlock::iterator 4128 LSRInstance::AdjustInsertPositionForExpand(BasicBlock::iterator LowestIP, 4129 const LSRFixup &LF, 4130 const LSRUse &LU, 4131 SCEVExpander &Rewriter) const { 4132 // Collect some instructions which must be dominated by the 4133 // expanding replacement. These must be dominated by any operands that 4134 // will be required in the expansion. 4135 SmallVector<Instruction *, 4> Inputs; 4136 if (Instruction *I = dyn_cast<Instruction>(LF.OperandValToReplace)) 4137 Inputs.push_back(I); 4138 if (LU.Kind == LSRUse::ICmpZero) 4139 if (Instruction *I = 4140 dyn_cast<Instruction>(cast<ICmpInst>(LF.UserInst)->getOperand(1))) 4141 Inputs.push_back(I); 4142 if (LF.PostIncLoops.count(L)) { 4143 if (LF.isUseFullyOutsideLoop(L)) 4144 Inputs.push_back(L->getLoopLatch()->getTerminator()); 4145 else 4146 Inputs.push_back(IVIncInsertPos); 4147 } 4148 // The expansion must also be dominated by the increment positions of any 4149 // loops it for which it is using post-inc mode. 4150 for (PostIncLoopSet::const_iterator I = LF.PostIncLoops.begin(), 4151 E = LF.PostIncLoops.end(); I != E; ++I) { 4152 const Loop *PIL = *I; 4153 if (PIL == L) continue; 4154 4155 // Be dominated by the loop exit. 4156 SmallVector<BasicBlock *, 4> ExitingBlocks; 4157 PIL->getExitingBlocks(ExitingBlocks); 4158 if (!ExitingBlocks.empty()) { 4159 BasicBlock *BB = ExitingBlocks[0]; 4160 for (unsigned i = 1, e = ExitingBlocks.size(); i != e; ++i) 4161 BB = DT.findNearestCommonDominator(BB, ExitingBlocks[i]); 4162 Inputs.push_back(BB->getTerminator()); 4163 } 4164 } 4165 4166 assert(!isa<PHINode>(LowestIP) && !isa<LandingPadInst>(LowestIP) 4167 && !isa<DbgInfoIntrinsic>(LowestIP) && 4168 "Insertion point must be a normal instruction"); 4169 4170 // Then, climb up the immediate dominator tree as far as we can go while 4171 // still being dominated by the input positions. 4172 BasicBlock::iterator IP = HoistInsertPosition(LowestIP, Inputs); 4173 4174 // Don't insert instructions before PHI nodes. 4175 while (isa<PHINode>(IP)) ++IP; 4176 4177 // Ignore landingpad instructions. 4178 while (isa<LandingPadInst>(IP)) ++IP; 4179 4180 // Ignore debug intrinsics. 4181 while (isa<DbgInfoIntrinsic>(IP)) ++IP; 4182 4183 // Set IP below instructions recently inserted by SCEVExpander. This keeps the 4184 // IP consistent across expansions and allows the previously inserted 4185 // instructions to be reused by subsequent expansion. 4186 while (Rewriter.isInsertedInstruction(IP) && IP != LowestIP) ++IP; 4187 4188 return IP; 4189 } 4190 4191 /// Expand - Emit instructions for the leading candidate expression for this 4192 /// LSRUse (this is called "expanding"). 4193 Value *LSRInstance::Expand(const LSRFixup &LF, 4194 const Formula &F, 4195 BasicBlock::iterator IP, 4196 SCEVExpander &Rewriter, 4197 SmallVectorImpl<WeakVH> &DeadInsts) const { 4198 const LSRUse &LU = Uses[LF.LUIdx]; 4199 4200 // Determine an input position which will be dominated by the operands and 4201 // which will dominate the result. 4202 IP = AdjustInsertPositionForExpand(IP, LF, LU, Rewriter); 4203 4204 // Inform the Rewriter if we have a post-increment use, so that it can 4205 // perform an advantageous expansion. 4206 Rewriter.setPostInc(LF.PostIncLoops); 4207 4208 // This is the type that the user actually needs. 4209 Type *OpTy = LF.OperandValToReplace->getType(); 4210 // This will be the type that we'll initially expand to. 4211 Type *Ty = F.getType(); 4212 if (!Ty) 4213 // No type known; just expand directly to the ultimate type. 4214 Ty = OpTy; 4215 else if (SE.getEffectiveSCEVType(Ty) == SE.getEffectiveSCEVType(OpTy)) 4216 // Expand directly to the ultimate type if it's the right size. 4217 Ty = OpTy; 4218 // This is the type to do integer arithmetic in. 4219 Type *IntTy = SE.getEffectiveSCEVType(Ty); 4220 4221 // Build up a list of operands to add together to form the full base. 4222 SmallVector<const SCEV *, 8> Ops; 4223 4224 // Expand the BaseRegs portion. 4225 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(), 4226 E = F.BaseRegs.end(); I != E; ++I) { 4227 const SCEV *Reg = *I; 4228 assert(!Reg->isZero() && "Zero allocated in a base register!"); 4229 4230 // If we're expanding for a post-inc user, make the post-inc adjustment. 4231 PostIncLoopSet &Loops = const_cast<PostIncLoopSet &>(LF.PostIncLoops); 4232 Reg = TransformForPostIncUse(Denormalize, Reg, 4233 LF.UserInst, LF.OperandValToReplace, 4234 Loops, SE, DT); 4235 4236 Ops.push_back(SE.getUnknown(Rewriter.expandCodeFor(Reg, 0, IP))); 4237 } 4238 4239 // Flush the operand list to suppress SCEVExpander hoisting. 4240 if (!Ops.empty()) { 4241 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP); 4242 Ops.clear(); 4243 Ops.push_back(SE.getUnknown(FullV)); 4244 } 4245 4246 // Expand the ScaledReg portion. 4247 Value *ICmpScaledV = 0; 4248 if (F.AM.Scale != 0) { 4249 const SCEV *ScaledS = F.ScaledReg; 4250 4251 // If we're expanding for a post-inc user, make the post-inc adjustment. 4252 PostIncLoopSet &Loops = const_cast<PostIncLoopSet &>(LF.PostIncLoops); 4253 ScaledS = TransformForPostIncUse(Denormalize, ScaledS, 4254 LF.UserInst, LF.OperandValToReplace, 4255 Loops, SE, DT); 4256 4257 if (LU.Kind == LSRUse::ICmpZero) { 4258 // An interesting way of "folding" with an icmp is to use a negated 4259 // scale, which we'll implement by inserting it into the other operand 4260 // of the icmp. 4261 assert(F.AM.Scale == -1 && 4262 "The only scale supported by ICmpZero uses is -1!"); 4263 ICmpScaledV = Rewriter.expandCodeFor(ScaledS, 0, IP); 4264 } else { 4265 // Otherwise just expand the scaled register and an explicit scale, 4266 // which is expected to be matched as part of the address. 4267 ScaledS = SE.getUnknown(Rewriter.expandCodeFor(ScaledS, 0, IP)); 4268 ScaledS = SE.getMulExpr(ScaledS, 4269 SE.getConstant(ScaledS->getType(), F.AM.Scale)); 4270 Ops.push_back(ScaledS); 4271 4272 // Flush the operand list to suppress SCEVExpander hoisting. 4273 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP); 4274 Ops.clear(); 4275 Ops.push_back(SE.getUnknown(FullV)); 4276 } 4277 } 4278 4279 // Expand the GV portion. 4280 if (F.AM.BaseGV) { 4281 Ops.push_back(SE.getUnknown(F.AM.BaseGV)); 4282 4283 // Flush the operand list to suppress SCEVExpander hoisting. 4284 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP); 4285 Ops.clear(); 4286 Ops.push_back(SE.getUnknown(FullV)); 4287 } 4288 4289 // Expand the immediate portion. 4290 int64_t Offset = (uint64_t)F.AM.BaseOffs + LF.Offset; 4291 if (Offset != 0) { 4292 if (LU.Kind == LSRUse::ICmpZero) { 4293 // The other interesting way of "folding" with an ICmpZero is to use a 4294 // negated immediate. 4295 if (!ICmpScaledV) 4296 ICmpScaledV = ConstantInt::get(IntTy, -(uint64_t)Offset); 4297 else { 4298 Ops.push_back(SE.getUnknown(ICmpScaledV)); 4299 ICmpScaledV = ConstantInt::get(IntTy, Offset); 4300 } 4301 } else { 4302 // Just add the immediate values. These again are expected to be matched 4303 // as part of the address. 4304 Ops.push_back(SE.getUnknown(ConstantInt::getSigned(IntTy, Offset))); 4305 } 4306 } 4307 4308 // Expand the unfolded offset portion. 4309 int64_t UnfoldedOffset = F.UnfoldedOffset; 4310 if (UnfoldedOffset != 0) { 4311 // Just add the immediate values. 4312 Ops.push_back(SE.getUnknown(ConstantInt::getSigned(IntTy, 4313 UnfoldedOffset))); 4314 } 4315 4316 // Emit instructions summing all the operands. 4317 const SCEV *FullS = Ops.empty() ? 4318 SE.getConstant(IntTy, 0) : 4319 SE.getAddExpr(Ops); 4320 Value *FullV = Rewriter.expandCodeFor(FullS, Ty, IP); 4321 4322 // We're done expanding now, so reset the rewriter. 4323 Rewriter.clearPostInc(); 4324 4325 // An ICmpZero Formula represents an ICmp which we're handling as a 4326 // comparison against zero. Now that we've expanded an expression for that 4327 // form, update the ICmp's other operand. 4328 if (LU.Kind == LSRUse::ICmpZero) { 4329 ICmpInst *CI = cast<ICmpInst>(LF.UserInst); 4330 DeadInsts.push_back(CI->getOperand(1)); 4331 assert(!F.AM.BaseGV && "ICmp does not support folding a global value and " 4332 "a scale at the same time!"); 4333 if (F.AM.Scale == -1) { 4334 if (ICmpScaledV->getType() != OpTy) { 4335 Instruction *Cast = 4336 CastInst::Create(CastInst::getCastOpcode(ICmpScaledV, false, 4337 OpTy, false), 4338 ICmpScaledV, OpTy, "tmp", CI); 4339 ICmpScaledV = Cast; 4340 } 4341 CI->setOperand(1, ICmpScaledV); 4342 } else { 4343 assert(F.AM.Scale == 0 && 4344 "ICmp does not support folding a global value and " 4345 "a scale at the same time!"); 4346 Constant *C = ConstantInt::getSigned(SE.getEffectiveSCEVType(OpTy), 4347 -(uint64_t)Offset); 4348 if (C->getType() != OpTy) 4349 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false, 4350 OpTy, false), 4351 C, OpTy); 4352 4353 CI->setOperand(1, C); 4354 } 4355 } 4356 4357 return FullV; 4358 } 4359 4360 /// RewriteForPHI - Helper for Rewrite. PHI nodes are special because the use 4361 /// of their operands effectively happens in their predecessor blocks, so the 4362 /// expression may need to be expanded in multiple places. 4363 void LSRInstance::RewriteForPHI(PHINode *PN, 4364 const LSRFixup &LF, 4365 const Formula &F, 4366 SCEVExpander &Rewriter, 4367 SmallVectorImpl<WeakVH> &DeadInsts, 4368 Pass *P) const { 4369 DenseMap<BasicBlock *, Value *> Inserted; 4370 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) 4371 if (PN->getIncomingValue(i) == LF.OperandValToReplace) { 4372 BasicBlock *BB = PN->getIncomingBlock(i); 4373 4374 // If this is a critical edge, split the edge so that we do not insert 4375 // the code on all predecessor/successor paths. We do this unless this 4376 // is the canonical backedge for this loop, which complicates post-inc 4377 // users. 4378 if (e != 1 && BB->getTerminator()->getNumSuccessors() > 1 && 4379 !isa<IndirectBrInst>(BB->getTerminator())) { 4380 BasicBlock *Parent = PN->getParent(); 4381 Loop *PNLoop = LI.getLoopFor(Parent); 4382 if (!PNLoop || Parent != PNLoop->getHeader()) { 4383 // Split the critical edge. 4384 BasicBlock *NewBB = 0; 4385 if (!Parent->isLandingPad()) { 4386 NewBB = SplitCriticalEdge(BB, Parent, P, 4387 /*MergeIdenticalEdges=*/true, 4388 /*DontDeleteUselessPhis=*/true); 4389 } else { 4390 SmallVector<BasicBlock*, 2> NewBBs; 4391 SplitLandingPadPredecessors(Parent, BB, "", "", P, NewBBs); 4392 NewBB = NewBBs[0]; 4393 } 4394 4395 // If PN is outside of the loop and BB is in the loop, we want to 4396 // move the block to be immediately before the PHI block, not 4397 // immediately after BB. 4398 if (L->contains(BB) && !L->contains(PN)) 4399 NewBB->moveBefore(PN->getParent()); 4400 4401 // Splitting the edge can reduce the number of PHI entries we have. 4402 e = PN->getNumIncomingValues(); 4403 BB = NewBB; 4404 i = PN->getBasicBlockIndex(BB); 4405 } 4406 } 4407 4408 std::pair<DenseMap<BasicBlock *, Value *>::iterator, bool> Pair = 4409 Inserted.insert(std::make_pair(BB, static_cast<Value *>(0))); 4410 if (!Pair.second) 4411 PN->setIncomingValue(i, Pair.first->second); 4412 else { 4413 Value *FullV = Expand(LF, F, BB->getTerminator(), Rewriter, DeadInsts); 4414 4415 // If this is reuse-by-noop-cast, insert the noop cast. 4416 Type *OpTy = LF.OperandValToReplace->getType(); 4417 if (FullV->getType() != OpTy) 4418 FullV = 4419 CastInst::Create(CastInst::getCastOpcode(FullV, false, 4420 OpTy, false), 4421 FullV, LF.OperandValToReplace->getType(), 4422 "tmp", BB->getTerminator()); 4423 4424 PN->setIncomingValue(i, FullV); 4425 Pair.first->second = FullV; 4426 } 4427 } 4428 } 4429 4430 /// Rewrite - Emit instructions for the leading candidate expression for this 4431 /// LSRUse (this is called "expanding"), and update the UserInst to reference 4432 /// the newly expanded value. 4433 void LSRInstance::Rewrite(const LSRFixup &LF, 4434 const Formula &F, 4435 SCEVExpander &Rewriter, 4436 SmallVectorImpl<WeakVH> &DeadInsts, 4437 Pass *P) const { 4438 // First, find an insertion point that dominates UserInst. For PHI nodes, 4439 // find the nearest block which dominates all the relevant uses. 4440 if (PHINode *PN = dyn_cast<PHINode>(LF.UserInst)) { 4441 RewriteForPHI(PN, LF, F, Rewriter, DeadInsts, P); 4442 } else { 4443 Value *FullV = Expand(LF, F, LF.UserInst, Rewriter, DeadInsts); 4444 4445 // If this is reuse-by-noop-cast, insert the noop cast. 4446 Type *OpTy = LF.OperandValToReplace->getType(); 4447 if (FullV->getType() != OpTy) { 4448 Instruction *Cast = 4449 CastInst::Create(CastInst::getCastOpcode(FullV, false, OpTy, false), 4450 FullV, OpTy, "tmp", LF.UserInst); 4451 FullV = Cast; 4452 } 4453 4454 // Update the user. ICmpZero is handled specially here (for now) because 4455 // Expand may have updated one of the operands of the icmp already, and 4456 // its new value may happen to be equal to LF.OperandValToReplace, in 4457 // which case doing replaceUsesOfWith leads to replacing both operands 4458 // with the same value. TODO: Reorganize this. 4459 if (Uses[LF.LUIdx].Kind == LSRUse::ICmpZero) 4460 LF.UserInst->setOperand(0, FullV); 4461 else 4462 LF.UserInst->replaceUsesOfWith(LF.OperandValToReplace, FullV); 4463 } 4464 4465 DeadInsts.push_back(LF.OperandValToReplace); 4466 } 4467 4468 /// ImplementSolution - Rewrite all the fixup locations with new values, 4469 /// following the chosen solution. 4470 void 4471 LSRInstance::ImplementSolution(const SmallVectorImpl<const Formula *> &Solution, 4472 Pass *P) { 4473 // Keep track of instructions we may have made dead, so that 4474 // we can remove them after we are done working. 4475 SmallVector<WeakVH, 16> DeadInsts; 4476 4477 SCEVExpander Rewriter(SE, "lsr"); 4478 #ifndef NDEBUG 4479 Rewriter.setDebugType(DEBUG_TYPE); 4480 #endif 4481 Rewriter.disableCanonicalMode(); 4482 Rewriter.enableLSRMode(); 4483 Rewriter.setIVIncInsertPos(L, IVIncInsertPos); 4484 4485 // Mark phi nodes that terminate chains so the expander tries to reuse them. 4486 for (SmallVectorImpl<IVChain>::const_iterator ChainI = IVChainVec.begin(), 4487 ChainE = IVChainVec.end(); ChainI != ChainE; ++ChainI) { 4488 if (PHINode *PN = dyn_cast<PHINode>(ChainI->back().UserInst)) 4489 Rewriter.setChainedPhi(PN); 4490 } 4491 4492 // Expand the new value definitions and update the users. 4493 for (SmallVectorImpl<LSRFixup>::const_iterator I = Fixups.begin(), 4494 E = Fixups.end(); I != E; ++I) { 4495 const LSRFixup &Fixup = *I; 4496 4497 Rewrite(Fixup, *Solution[Fixup.LUIdx], Rewriter, DeadInsts, P); 4498 4499 Changed = true; 4500 } 4501 4502 for (SmallVectorImpl<IVChain>::const_iterator ChainI = IVChainVec.begin(), 4503 ChainE = IVChainVec.end(); ChainI != ChainE; ++ChainI) { 4504 GenerateIVChain(*ChainI, Rewriter, DeadInsts); 4505 Changed = true; 4506 } 4507 // Clean up after ourselves. This must be done before deleting any 4508 // instructions. 4509 Rewriter.clear(); 4510 4511 Changed |= DeleteTriviallyDeadInstructions(DeadInsts); 4512 } 4513 4514 LSRInstance::LSRInstance(const TargetLowering *tli, Loop *l, Pass *P) 4515 : IU(P->getAnalysis<IVUsers>()), 4516 SE(P->getAnalysis<ScalarEvolution>()), 4517 DT(P->getAnalysis<DominatorTree>()), 4518 LI(P->getAnalysis<LoopInfo>()), 4519 TLI(tli), L(l), Changed(false), IVIncInsertPos(0) { 4520 4521 // If LoopSimplify form is not available, stay out of trouble. 4522 if (!L->isLoopSimplifyForm()) 4523 return; 4524 4525 // If there's no interesting work to be done, bail early. 4526 if (IU.empty()) return; 4527 4528 // If there's too much analysis to be done, bail early. We won't be able to 4529 // model the problem anyway. 4530 unsigned NumUsers = 0; 4531 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI) { 4532 if (++NumUsers > MaxIVUsers) { 4533 DEBUG(dbgs() << "LSR skipping loop, too many IV Users in " << *L 4534 << "\n"); 4535 return; 4536 } 4537 } 4538 4539 #ifndef NDEBUG 4540 // All dominating loops must have preheaders, or SCEVExpander may not be able 4541 // to materialize an AddRecExpr whose Start is an outer AddRecExpr. 4542 // 4543 // IVUsers analysis should only create users that are dominated by simple loop 4544 // headers. Since this loop should dominate all of its users, its user list 4545 // should be empty if this loop itself is not within a simple loop nest. 4546 for (DomTreeNode *Rung = DT.getNode(L->getLoopPreheader()); 4547 Rung; Rung = Rung->getIDom()) { 4548 BasicBlock *BB = Rung->getBlock(); 4549 const Loop *DomLoop = LI.getLoopFor(BB); 4550 if (DomLoop && DomLoop->getHeader() == BB) { 4551 assert(DomLoop->getLoopPreheader() && "LSR needs a simplified loop nest"); 4552 } 4553 } 4554 #endif // DEBUG 4555 4556 DEBUG(dbgs() << "\nLSR on loop "; 4557 WriteAsOperand(dbgs(), L->getHeader(), /*PrintType=*/false); 4558 dbgs() << ":\n"); 4559 4560 // First, perform some low-level loop optimizations. 4561 OptimizeShadowIV(); 4562 OptimizeLoopTermCond(); 4563 4564 // If loop preparation eliminates all interesting IV users, bail. 4565 if (IU.empty()) return; 4566 4567 // Skip nested loops until we can model them better with formulae. 4568 if (!L->empty()) { 4569 DEBUG(dbgs() << "LSR skipping outer loop " << *L << "\n"); 4570 return; 4571 } 4572 4573 // Start collecting data and preparing for the solver. 4574 CollectChains(); 4575 CollectInterestingTypesAndFactors(); 4576 CollectFixupsAndInitialFormulae(); 4577 CollectLoopInvariantFixupsAndFormulae(); 4578 4579 assert(!Uses.empty() && "IVUsers reported at least one use"); 4580 DEBUG(dbgs() << "LSR found " << Uses.size() << " uses:\n"; 4581 print_uses(dbgs())); 4582 4583 // Now use the reuse data to generate a bunch of interesting ways 4584 // to formulate the values needed for the uses. 4585 GenerateAllReuseFormulae(); 4586 4587 FilterOutUndesirableDedicatedRegisters(); 4588 NarrowSearchSpaceUsingHeuristics(); 4589 4590 SmallVector<const Formula *, 8> Solution; 4591 Solve(Solution); 4592 4593 // Release memory that is no longer needed. 4594 Factors.clear(); 4595 Types.clear(); 4596 RegUses.clear(); 4597 4598 if (Solution.empty()) 4599 return; 4600 4601 #ifndef NDEBUG 4602 // Formulae should be legal. 4603 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(), 4604 E = Uses.end(); I != E; ++I) { 4605 const LSRUse &LU = *I; 4606 for (SmallVectorImpl<Formula>::const_iterator J = LU.Formulae.begin(), 4607 JE = LU.Formulae.end(); J != JE; ++J) 4608 assert(isLegalUse(J->AM, LU.MinOffset, LU.MaxOffset, 4609 LU.Kind, LU.AccessTy, TLI) && 4610 "Illegal formula generated!"); 4611 }; 4612 #endif 4613 4614 // Now that we've decided what we want, make it so. 4615 ImplementSolution(Solution, P); 4616 } 4617 4618 void LSRInstance::print_factors_and_types(raw_ostream &OS) const { 4619 if (Factors.empty() && Types.empty()) return; 4620 4621 OS << "LSR has identified the following interesting factors and types: "; 4622 bool First = true; 4623 4624 for (SmallSetVector<int64_t, 8>::const_iterator 4625 I = Factors.begin(), E = Factors.end(); I != E; ++I) { 4626 if (!First) OS << ", "; 4627 First = false; 4628 OS << '*' << *I; 4629 } 4630 4631 for (SmallSetVector<Type *, 4>::const_iterator 4632 I = Types.begin(), E = Types.end(); I != E; ++I) { 4633 if (!First) OS << ", "; 4634 First = false; 4635 OS << '(' << **I << ')'; 4636 } 4637 OS << '\n'; 4638 } 4639 4640 void LSRInstance::print_fixups(raw_ostream &OS) const { 4641 OS << "LSR is examining the following fixup sites:\n"; 4642 for (SmallVectorImpl<LSRFixup>::const_iterator I = Fixups.begin(), 4643 E = Fixups.end(); I != E; ++I) { 4644 dbgs() << " "; 4645 I->print(OS); 4646 OS << '\n'; 4647 } 4648 } 4649 4650 void LSRInstance::print_uses(raw_ostream &OS) const { 4651 OS << "LSR is examining the following uses:\n"; 4652 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(), 4653 E = Uses.end(); I != E; ++I) { 4654 const LSRUse &LU = *I; 4655 dbgs() << " "; 4656 LU.print(OS); 4657 OS << '\n'; 4658 for (SmallVectorImpl<Formula>::const_iterator J = LU.Formulae.begin(), 4659 JE = LU.Formulae.end(); J != JE; ++J) { 4660 OS << " "; 4661 J->print(OS); 4662 OS << '\n'; 4663 } 4664 } 4665 } 4666 4667 void LSRInstance::print(raw_ostream &OS) const { 4668 print_factors_and_types(OS); 4669 print_fixups(OS); 4670 print_uses(OS); 4671 } 4672 4673 void LSRInstance::dump() const { 4674 print(errs()); errs() << '\n'; 4675 } 4676 4677 namespace { 4678 4679 class LoopStrengthReduce : public LoopPass { 4680 /// TLI - Keep a pointer of a TargetLowering to consult for determining 4681 /// transformation profitability. 4682 const TargetLowering *const TLI; 4683 4684 public: 4685 static char ID; // Pass ID, replacement for typeid 4686 explicit LoopStrengthReduce(const TargetLowering *tli = 0); 4687 4688 private: 4689 bool runOnLoop(Loop *L, LPPassManager &LPM); 4690 void getAnalysisUsage(AnalysisUsage &AU) const; 4691 }; 4692 4693 } 4694 4695 char LoopStrengthReduce::ID = 0; 4696 INITIALIZE_PASS_BEGIN(LoopStrengthReduce, "loop-reduce", 4697 "Loop Strength Reduction", false, false) 4698 INITIALIZE_PASS_DEPENDENCY(DominatorTree) 4699 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution) 4700 INITIALIZE_PASS_DEPENDENCY(IVUsers) 4701 INITIALIZE_PASS_DEPENDENCY(LoopInfo) 4702 INITIALIZE_PASS_DEPENDENCY(LoopSimplify) 4703 INITIALIZE_PASS_END(LoopStrengthReduce, "loop-reduce", 4704 "Loop Strength Reduction", false, false) 4705 4706 4707 Pass *llvm::createLoopStrengthReducePass(const TargetLowering *TLI) { 4708 return new LoopStrengthReduce(TLI); 4709 } 4710 4711 LoopStrengthReduce::LoopStrengthReduce(const TargetLowering *tli) 4712 : LoopPass(ID), TLI(tli) { 4713 initializeLoopStrengthReducePass(*PassRegistry::getPassRegistry()); 4714 } 4715 4716 void LoopStrengthReduce::getAnalysisUsage(AnalysisUsage &AU) const { 4717 // We split critical edges, so we change the CFG. However, we do update 4718 // many analyses if they are around. 4719 AU.addPreservedID(LoopSimplifyID); 4720 4721 AU.addRequired<LoopInfo>(); 4722 AU.addPreserved<LoopInfo>(); 4723 AU.addRequiredID(LoopSimplifyID); 4724 AU.addRequired<DominatorTree>(); 4725 AU.addPreserved<DominatorTree>(); 4726 AU.addRequired<ScalarEvolution>(); 4727 AU.addPreserved<ScalarEvolution>(); 4728 // Requiring LoopSimplify a second time here prevents IVUsers from running 4729 // twice, since LoopSimplify was invalidated by running ScalarEvolution. 4730 AU.addRequiredID(LoopSimplifyID); 4731 AU.addRequired<IVUsers>(); 4732 AU.addPreserved<IVUsers>(); 4733 } 4734 4735 bool LoopStrengthReduce::runOnLoop(Loop *L, LPPassManager & /*LPM*/) { 4736 bool Changed = false; 4737 4738 // Run the main LSR transformation. 4739 Changed |= LSRInstance(TLI, L, this).getChanged(); 4740 4741 // Remove any extra phis created by processing inner loops. 4742 Changed |= DeleteDeadPHIs(L->getHeader()); 4743 if (EnablePhiElim) { 4744 SmallVector<WeakVH, 16> DeadInsts; 4745 SCEVExpander Rewriter(getAnalysis<ScalarEvolution>(), "lsr"); 4746 #ifndef NDEBUG 4747 Rewriter.setDebugType(DEBUG_TYPE); 4748 #endif 4749 unsigned numFolded = Rewriter. 4750 replaceCongruentIVs(L, &getAnalysis<DominatorTree>(), DeadInsts, TLI); 4751 if (numFolded) { 4752 Changed = true; 4753 DeleteTriviallyDeadInstructions(DeadInsts); 4754 DeleteDeadPHIs(L->getHeader()); 4755 } 4756 } 4757 return Changed; 4758 } 4759