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