LoopStrengthReduce.cpp revision 243830
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/AddressingMode.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/Scalar.h" 68#include "llvm/Transforms/Utils/BasicBlockUtils.h" 69#include "llvm/Transforms/Utils/Local.h" 70#include "llvm/ADT/SmallBitVector.h" 71#include "llvm/ADT/SetVector.h" 72#include "llvm/ADT/DenseSet.h" 73#include "llvm/Support/Debug.h" 74#include "llvm/Support/CommandLine.h" 75#include "llvm/Support/ValueHandle.h" 76#include "llvm/Support/raw_ostream.h" 77#include "llvm/Target/TargetLowering.h" 78#include <algorithm> 79using namespace llvm; 80 81/// MaxIVUsers is an arbitrary threshold that provides an early opportunitiy for 82/// bail out. This threshold is far beyond the number of users that LSR can 83/// conceivably solve, so it should not affect generated code, but catches the 84/// worst cases before LSR burns too much compile time and stack space. 85static const unsigned MaxIVUsers = 200; 86 87// Temporary flag to cleanup congruent phis after LSR phi expansion. 88// It's currently disabled until we can determine whether it's truly useful or 89// not. The flag should be removed after the v3.0 release. 90// This is now needed for ivchains. 91static cl::opt<bool> EnablePhiElim( 92 "enable-lsr-phielim", cl::Hidden, cl::init(true), 93 cl::desc("Enable LSR phi elimination")); 94 95#ifndef NDEBUG 96// Stress test IV chain generation. 97static cl::opt<bool> StressIVChain( 98 "stress-ivchain", cl::Hidden, cl::init(false), 99 cl::desc("Stress test LSR IV chains")); 100#else 101static bool StressIVChain = false; 102#endif 103 104namespace { 105 106/// RegSortData - This class holds data which is used to order reuse candidates. 107class RegSortData { 108public: 109 /// UsedByIndices - This represents the set of LSRUse indices which reference 110 /// a particular register. 111 SmallBitVector UsedByIndices; 112 113 RegSortData() {} 114 115 void print(raw_ostream &OS) const; 116 void dump() const; 117}; 118 119} 120 121void RegSortData::print(raw_ostream &OS) const { 122 OS << "[NumUses=" << UsedByIndices.count() << ']'; 123} 124 125#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) 126void RegSortData::dump() const { 127 print(errs()); errs() << '\n'; 128} 129#endif 130 131namespace { 132 133/// RegUseTracker - Map register candidates to information about how they are 134/// used. 135class RegUseTracker { 136 typedef DenseMap<const SCEV *, RegSortData> RegUsesTy; 137 138 RegUsesTy RegUsesMap; 139 SmallVector<const SCEV *, 16> RegSequence; 140 141public: 142 void CountRegister(const SCEV *Reg, size_t LUIdx); 143 void DropRegister(const SCEV *Reg, size_t LUIdx); 144 void SwapAndDropUse(size_t LUIdx, size_t LastLUIdx); 145 146 bool isRegUsedByUsesOtherThan(const SCEV *Reg, size_t LUIdx) const; 147 148 const SmallBitVector &getUsedByIndices(const SCEV *Reg) const; 149 150 void clear(); 151 152 typedef SmallVectorImpl<const SCEV *>::iterator iterator; 153 typedef SmallVectorImpl<const SCEV *>::const_iterator const_iterator; 154 iterator begin() { return RegSequence.begin(); } 155 iterator end() { return RegSequence.end(); } 156 const_iterator begin() const { return RegSequence.begin(); } 157 const_iterator end() const { return RegSequence.end(); } 158}; 159 160} 161 162void 163RegUseTracker::CountRegister(const SCEV *Reg, size_t LUIdx) { 164 std::pair<RegUsesTy::iterator, bool> Pair = 165 RegUsesMap.insert(std::make_pair(Reg, RegSortData())); 166 RegSortData &RSD = Pair.first->second; 167 if (Pair.second) 168 RegSequence.push_back(Reg); 169 RSD.UsedByIndices.resize(std::max(RSD.UsedByIndices.size(), LUIdx + 1)); 170 RSD.UsedByIndices.set(LUIdx); 171} 172 173void 174RegUseTracker::DropRegister(const SCEV *Reg, size_t LUIdx) { 175 RegUsesTy::iterator It = RegUsesMap.find(Reg); 176 assert(It != RegUsesMap.end()); 177 RegSortData &RSD = It->second; 178 assert(RSD.UsedByIndices.size() > LUIdx); 179 RSD.UsedByIndices.reset(LUIdx); 180} 181 182void 183RegUseTracker::SwapAndDropUse(size_t LUIdx, size_t LastLUIdx) { 184 assert(LUIdx <= LastLUIdx); 185 186 // Update RegUses. The data structure is not optimized for this purpose; 187 // we must iterate through it and update each of the bit vectors. 188 for (RegUsesTy::iterator I = RegUsesMap.begin(), E = RegUsesMap.end(); 189 I != E; ++I) { 190 SmallBitVector &UsedByIndices = I->second.UsedByIndices; 191 if (LUIdx < UsedByIndices.size()) 192 UsedByIndices[LUIdx] = 193 LastLUIdx < UsedByIndices.size() ? UsedByIndices[LastLUIdx] : 0; 194 UsedByIndices.resize(std::min(UsedByIndices.size(), LastLUIdx)); 195 } 196} 197 198bool 199RegUseTracker::isRegUsedByUsesOtherThan(const SCEV *Reg, size_t LUIdx) const { 200 RegUsesTy::const_iterator I = RegUsesMap.find(Reg); 201 if (I == RegUsesMap.end()) 202 return false; 203 const SmallBitVector &UsedByIndices = I->second.UsedByIndices; 204 int i = UsedByIndices.find_first(); 205 if (i == -1) return false; 206 if ((size_t)i != LUIdx) return true; 207 return UsedByIndices.find_next(i) != -1; 208} 209 210const SmallBitVector &RegUseTracker::getUsedByIndices(const SCEV *Reg) const { 211 RegUsesTy::const_iterator I = RegUsesMap.find(Reg); 212 assert(I != RegUsesMap.end() && "Unknown register!"); 213 return I->second.UsedByIndices; 214} 215 216void RegUseTracker::clear() { 217 RegUsesMap.clear(); 218 RegSequence.clear(); 219} 220 221namespace { 222 223/// Formula - This class holds information that describes a formula for 224/// computing satisfying a use. It may include broken-out immediates and scaled 225/// registers. 226struct Formula { 227 /// AM - This is used to represent complex addressing, as well as other kinds 228 /// of interesting uses. 229 AddrMode AM; 230 231 /// BaseRegs - The list of "base" registers for this use. When this is 232 /// non-empty, AM.HasBaseReg should be set to true. 233 SmallVector<const SCEV *, 2> BaseRegs; 234 235 /// ScaledReg - The 'scaled' register for this use. This should be non-null 236 /// when AM.Scale is not zero. 237 const SCEV *ScaledReg; 238 239 /// UnfoldedOffset - An additional constant offset which added near the 240 /// use. This requires a temporary register, but the offset itself can 241 /// live in an add immediate field rather than a register. 242 int64_t UnfoldedOffset; 243 244 Formula() : ScaledReg(0), UnfoldedOffset(0) {} 245 246 void InitialMatch(const SCEV *S, Loop *L, ScalarEvolution &SE); 247 248 unsigned getNumRegs() const; 249 Type *getType() const; 250 251 void DeleteBaseReg(const SCEV *&S); 252 253 bool referencesReg(const SCEV *S) const; 254 bool hasRegsUsedByUsesOtherThan(size_t LUIdx, 255 const RegUseTracker &RegUses) const; 256 257 void print(raw_ostream &OS) const; 258 void dump() const; 259}; 260 261} 262 263/// DoInitialMatch - Recursion helper for InitialMatch. 264static void DoInitialMatch(const SCEV *S, Loop *L, 265 SmallVectorImpl<const SCEV *> &Good, 266 SmallVectorImpl<const SCEV *> &Bad, 267 ScalarEvolution &SE) { 268 // Collect expressions which properly dominate the loop header. 269 if (SE.properlyDominates(S, L->getHeader())) { 270 Good.push_back(S); 271 return; 272 } 273 274 // Look at add operands. 275 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) { 276 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end(); 277 I != E; ++I) 278 DoInitialMatch(*I, L, Good, Bad, SE); 279 return; 280 } 281 282 // Look at addrec operands. 283 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) 284 if (!AR->getStart()->isZero()) { 285 DoInitialMatch(AR->getStart(), L, Good, Bad, SE); 286 DoInitialMatch(SE.getAddRecExpr(SE.getConstant(AR->getType(), 0), 287 AR->getStepRecurrence(SE), 288 // FIXME: AR->getNoWrapFlags() 289 AR->getLoop(), SCEV::FlagAnyWrap), 290 L, Good, Bad, SE); 291 return; 292 } 293 294 // Handle a multiplication by -1 (negation) if it didn't fold. 295 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) 296 if (Mul->getOperand(0)->isAllOnesValue()) { 297 SmallVector<const SCEV *, 4> Ops(Mul->op_begin()+1, Mul->op_end()); 298 const SCEV *NewMul = SE.getMulExpr(Ops); 299 300 SmallVector<const SCEV *, 4> MyGood; 301 SmallVector<const SCEV *, 4> MyBad; 302 DoInitialMatch(NewMul, L, MyGood, MyBad, SE); 303 const SCEV *NegOne = SE.getSCEV(ConstantInt::getAllOnesValue( 304 SE.getEffectiveSCEVType(NewMul->getType()))); 305 for (SmallVectorImpl<const SCEV *>::const_iterator I = MyGood.begin(), 306 E = MyGood.end(); I != E; ++I) 307 Good.push_back(SE.getMulExpr(NegOne, *I)); 308 for (SmallVectorImpl<const SCEV *>::const_iterator I = MyBad.begin(), 309 E = MyBad.end(); I != E; ++I) 310 Bad.push_back(SE.getMulExpr(NegOne, *I)); 311 return; 312 } 313 314 // Ok, we can't do anything interesting. Just stuff the whole thing into a 315 // register and hope for the best. 316 Bad.push_back(S); 317} 318 319/// InitialMatch - Incorporate loop-variant parts of S into this Formula, 320/// attempting to keep all loop-invariant and loop-computable values in a 321/// single base register. 322void Formula::InitialMatch(const SCEV *S, Loop *L, ScalarEvolution &SE) { 323 SmallVector<const SCEV *, 4> Good; 324 SmallVector<const SCEV *, 4> Bad; 325 DoInitialMatch(S, L, Good, Bad, SE); 326 if (!Good.empty()) { 327 const SCEV *Sum = SE.getAddExpr(Good); 328 if (!Sum->isZero()) 329 BaseRegs.push_back(Sum); 330 AM.HasBaseReg = true; 331 } 332 if (!Bad.empty()) { 333 const SCEV *Sum = SE.getAddExpr(Bad); 334 if (!Sum->isZero()) 335 BaseRegs.push_back(Sum); 336 AM.HasBaseReg = true; 337 } 338} 339 340/// getNumRegs - Return the total number of register operands used by this 341/// formula. This does not include register uses implied by non-constant 342/// addrec strides. 343unsigned Formula::getNumRegs() const { 344 return !!ScaledReg + BaseRegs.size(); 345} 346 347/// getType - Return the type of this formula, if it has one, or null 348/// otherwise. This type is meaningless except for the bit size. 349Type *Formula::getType() const { 350 return !BaseRegs.empty() ? BaseRegs.front()->getType() : 351 ScaledReg ? ScaledReg->getType() : 352 AM.BaseGV ? AM.BaseGV->getType() : 353 0; 354} 355 356/// DeleteBaseReg - Delete the given base reg from the BaseRegs list. 357void Formula::DeleteBaseReg(const SCEV *&S) { 358 if (&S != &BaseRegs.back()) 359 std::swap(S, BaseRegs.back()); 360 BaseRegs.pop_back(); 361} 362 363/// referencesReg - Test if this formula references the given register. 364bool Formula::referencesReg(const SCEV *S) const { 365 return S == ScaledReg || 366 std::find(BaseRegs.begin(), BaseRegs.end(), S) != BaseRegs.end(); 367} 368 369/// hasRegsUsedByUsesOtherThan - Test whether this formula uses registers 370/// which are used by uses other than the use with the given index. 371bool Formula::hasRegsUsedByUsesOtherThan(size_t LUIdx, 372 const RegUseTracker &RegUses) const { 373 if (ScaledReg) 374 if (RegUses.isRegUsedByUsesOtherThan(ScaledReg, LUIdx)) 375 return true; 376 for (SmallVectorImpl<const SCEV *>::const_iterator I = BaseRegs.begin(), 377 E = BaseRegs.end(); I != E; ++I) 378 if (RegUses.isRegUsedByUsesOtherThan(*I, LUIdx)) 379 return true; 380 return false; 381} 382 383void Formula::print(raw_ostream &OS) const { 384 bool First = true; 385 if (AM.BaseGV) { 386 if (!First) OS << " + "; else First = false; 387 WriteAsOperand(OS, AM.BaseGV, /*PrintType=*/false); 388 } 389 if (AM.BaseOffs != 0) { 390 if (!First) OS << " + "; else First = false; 391 OS << AM.BaseOffs; 392 } 393 for (SmallVectorImpl<const SCEV *>::const_iterator I = BaseRegs.begin(), 394 E = BaseRegs.end(); I != E; ++I) { 395 if (!First) OS << " + "; else First = false; 396 OS << "reg(" << **I << ')'; 397 } 398 if (AM.HasBaseReg && BaseRegs.empty()) { 399 if (!First) OS << " + "; else First = false; 400 OS << "**error: HasBaseReg**"; 401 } else if (!AM.HasBaseReg && !BaseRegs.empty()) { 402 if (!First) OS << " + "; else First = false; 403 OS << "**error: !HasBaseReg**"; 404 } 405 if (AM.Scale != 0) { 406 if (!First) OS << " + "; else First = false; 407 OS << AM.Scale << "*reg("; 408 if (ScaledReg) 409 OS << *ScaledReg; 410 else 411 OS << "<unknown>"; 412 OS << ')'; 413 } 414 if (UnfoldedOffset != 0) { 415 if (!First) OS << " + "; else First = false; 416 OS << "imm(" << UnfoldedOffset << ')'; 417 } 418} 419 420#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) 421void Formula::dump() const { 422 print(errs()); errs() << '\n'; 423} 424#endif 425 426/// isAddRecSExtable - Return true if the given addrec can be sign-extended 427/// without changing its value. 428static bool isAddRecSExtable(const SCEVAddRecExpr *AR, ScalarEvolution &SE) { 429 Type *WideTy = 430 IntegerType::get(SE.getContext(), SE.getTypeSizeInBits(AR->getType()) + 1); 431 return isa<SCEVAddRecExpr>(SE.getSignExtendExpr(AR, WideTy)); 432} 433 434/// isAddSExtable - Return true if the given add can be sign-extended 435/// without changing its value. 436static bool isAddSExtable(const SCEVAddExpr *A, ScalarEvolution &SE) { 437 Type *WideTy = 438 IntegerType::get(SE.getContext(), SE.getTypeSizeInBits(A->getType()) + 1); 439 return isa<SCEVAddExpr>(SE.getSignExtendExpr(A, WideTy)); 440} 441 442/// isMulSExtable - Return true if the given mul can be sign-extended 443/// without changing its value. 444static bool isMulSExtable(const SCEVMulExpr *M, ScalarEvolution &SE) { 445 Type *WideTy = 446 IntegerType::get(SE.getContext(), 447 SE.getTypeSizeInBits(M->getType()) * M->getNumOperands()); 448 return isa<SCEVMulExpr>(SE.getSignExtendExpr(M, WideTy)); 449} 450 451/// getExactSDiv - Return an expression for LHS /s RHS, if it can be determined 452/// and if the remainder is known to be zero, or null otherwise. If 453/// IgnoreSignificantBits is true, expressions like (X * Y) /s Y are simplified 454/// to Y, ignoring that the multiplication may overflow, which is useful when 455/// the result will be used in a context where the most significant bits are 456/// ignored. 457static const SCEV *getExactSDiv(const SCEV *LHS, const SCEV *RHS, 458 ScalarEvolution &SE, 459 bool IgnoreSignificantBits = false) { 460 // Handle the trivial case, which works for any SCEV type. 461 if (LHS == RHS) 462 return SE.getConstant(LHS->getType(), 1); 463 464 // Handle a few RHS special cases. 465 const SCEVConstant *RC = dyn_cast<SCEVConstant>(RHS); 466 if (RC) { 467 const APInt &RA = RC->getValue()->getValue(); 468 // Handle x /s -1 as x * -1, to give ScalarEvolution a chance to do 469 // some folding. 470 if (RA.isAllOnesValue()) 471 return SE.getMulExpr(LHS, RC); 472 // Handle x /s 1 as x. 473 if (RA == 1) 474 return LHS; 475 } 476 477 // Check for a division of a constant by a constant. 478 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(LHS)) { 479 if (!RC) 480 return 0; 481 const APInt &LA = C->getValue()->getValue(); 482 const APInt &RA = RC->getValue()->getValue(); 483 if (LA.srem(RA) != 0) 484 return 0; 485 return SE.getConstant(LA.sdiv(RA)); 486 } 487 488 // Distribute the sdiv over addrec operands, if the addrec doesn't overflow. 489 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(LHS)) { 490 if (IgnoreSignificantBits || isAddRecSExtable(AR, SE)) { 491 const SCEV *Step = getExactSDiv(AR->getStepRecurrence(SE), RHS, SE, 492 IgnoreSignificantBits); 493 if (!Step) return 0; 494 const SCEV *Start = getExactSDiv(AR->getStart(), RHS, SE, 495 IgnoreSignificantBits); 496 if (!Start) return 0; 497 // FlagNW is independent of the start value, step direction, and is 498 // preserved with smaller magnitude steps. 499 // FIXME: AR->getNoWrapFlags(SCEV::FlagNW) 500 return SE.getAddRecExpr(Start, Step, AR->getLoop(), SCEV::FlagAnyWrap); 501 } 502 return 0; 503 } 504 505 // Distribute the sdiv over add operands, if the add doesn't overflow. 506 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(LHS)) { 507 if (IgnoreSignificantBits || isAddSExtable(Add, SE)) { 508 SmallVector<const SCEV *, 8> Ops; 509 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end(); 510 I != E; ++I) { 511 const SCEV *Op = getExactSDiv(*I, RHS, SE, 512 IgnoreSignificantBits); 513 if (!Op) return 0; 514 Ops.push_back(Op); 515 } 516 return SE.getAddExpr(Ops); 517 } 518 return 0; 519 } 520 521 // Check for a multiply operand that we can pull RHS out of. 522 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(LHS)) { 523 if (IgnoreSignificantBits || isMulSExtable(Mul, SE)) { 524 SmallVector<const SCEV *, 4> Ops; 525 bool Found = false; 526 for (SCEVMulExpr::op_iterator I = Mul->op_begin(), E = Mul->op_end(); 527 I != E; ++I) { 528 const SCEV *S = *I; 529 if (!Found) 530 if (const SCEV *Q = getExactSDiv(S, RHS, SE, 531 IgnoreSignificantBits)) { 532 S = Q; 533 Found = true; 534 } 535 Ops.push_back(S); 536 } 537 return Found ? SE.getMulExpr(Ops) : 0; 538 } 539 return 0; 540 } 541 542 // Otherwise we don't know. 543 return 0; 544} 545 546/// ExtractImmediate - If S involves the addition of a constant integer value, 547/// return that integer value, and mutate S to point to a new SCEV with that 548/// value excluded. 549static int64_t ExtractImmediate(const SCEV *&S, ScalarEvolution &SE) { 550 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S)) { 551 if (C->getValue()->getValue().getMinSignedBits() <= 64) { 552 S = SE.getConstant(C->getType(), 0); 553 return C->getValue()->getSExtValue(); 554 } 555 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) { 556 SmallVector<const SCEV *, 8> NewOps(Add->op_begin(), Add->op_end()); 557 int64_t Result = ExtractImmediate(NewOps.front(), SE); 558 if (Result != 0) 559 S = SE.getAddExpr(NewOps); 560 return Result; 561 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) { 562 SmallVector<const SCEV *, 8> NewOps(AR->op_begin(), AR->op_end()); 563 int64_t Result = ExtractImmediate(NewOps.front(), SE); 564 if (Result != 0) 565 S = SE.getAddRecExpr(NewOps, AR->getLoop(), 566 // FIXME: AR->getNoWrapFlags(SCEV::FlagNW) 567 SCEV::FlagAnyWrap); 568 return Result; 569 } 570 return 0; 571} 572 573/// ExtractSymbol - If S involves the addition of a GlobalValue address, 574/// return that symbol, and mutate S to point to a new SCEV with that 575/// value excluded. 576static GlobalValue *ExtractSymbol(const SCEV *&S, ScalarEvolution &SE) { 577 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) { 578 if (GlobalValue *GV = dyn_cast<GlobalValue>(U->getValue())) { 579 S = SE.getConstant(GV->getType(), 0); 580 return GV; 581 } 582 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) { 583 SmallVector<const SCEV *, 8> NewOps(Add->op_begin(), Add->op_end()); 584 GlobalValue *Result = ExtractSymbol(NewOps.back(), SE); 585 if (Result) 586 S = SE.getAddExpr(NewOps); 587 return Result; 588 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) { 589 SmallVector<const SCEV *, 8> NewOps(AR->op_begin(), AR->op_end()); 590 GlobalValue *Result = ExtractSymbol(NewOps.front(), SE); 591 if (Result) 592 S = SE.getAddRecExpr(NewOps, AR->getLoop(), 593 // FIXME: AR->getNoWrapFlags(SCEV::FlagNW) 594 SCEV::FlagAnyWrap); 595 return Result; 596 } 597 return 0; 598} 599 600/// isAddressUse - Returns true if the specified instruction is using the 601/// specified value as an address. 602static bool isAddressUse(Instruction *Inst, Value *OperandVal) { 603 bool isAddress = isa<LoadInst>(Inst); 604 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) { 605 if (SI->getOperand(1) == OperandVal) 606 isAddress = true; 607 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) { 608 // Addressing modes can also be folded into prefetches and a variety 609 // of intrinsics. 610 switch (II->getIntrinsicID()) { 611 default: break; 612 case Intrinsic::prefetch: 613 case Intrinsic::x86_sse_storeu_ps: 614 case Intrinsic::x86_sse2_storeu_pd: 615 case Intrinsic::x86_sse2_storeu_dq: 616 case Intrinsic::x86_sse2_storel_dq: 617 if (II->getArgOperand(0) == OperandVal) 618 isAddress = true; 619 break; 620 } 621 } 622 return isAddress; 623} 624 625/// getAccessType - Return the type of the memory being accessed. 626static Type *getAccessType(const Instruction *Inst) { 627 Type *AccessTy = Inst->getType(); 628 if (const StoreInst *SI = dyn_cast<StoreInst>(Inst)) 629 AccessTy = SI->getOperand(0)->getType(); 630 else if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) { 631 // Addressing modes can also be folded into prefetches and a variety 632 // of intrinsics. 633 switch (II->getIntrinsicID()) { 634 default: break; 635 case Intrinsic::x86_sse_storeu_ps: 636 case Intrinsic::x86_sse2_storeu_pd: 637 case Intrinsic::x86_sse2_storeu_dq: 638 case Intrinsic::x86_sse2_storel_dq: 639 AccessTy = II->getArgOperand(0)->getType(); 640 break; 641 } 642 } 643 644 // All pointers have the same requirements, so canonicalize them to an 645 // arbitrary pointer type to minimize variation. 646 if (PointerType *PTy = dyn_cast<PointerType>(AccessTy)) 647 AccessTy = PointerType::get(IntegerType::get(PTy->getContext(), 1), 648 PTy->getAddressSpace()); 649 650 return AccessTy; 651} 652 653/// isExistingPhi - Return true if this AddRec is already a phi in its loop. 654static bool isExistingPhi(const SCEVAddRecExpr *AR, ScalarEvolution &SE) { 655 for (BasicBlock::iterator I = AR->getLoop()->getHeader()->begin(); 656 PHINode *PN = dyn_cast<PHINode>(I); ++I) { 657 if (SE.isSCEVable(PN->getType()) && 658 (SE.getEffectiveSCEVType(PN->getType()) == 659 SE.getEffectiveSCEVType(AR->getType())) && 660 SE.getSCEV(PN) == AR) 661 return true; 662 } 663 return false; 664} 665 666/// Check if expanding this expression is likely to incur significant cost. This 667/// is tricky because SCEV doesn't track which expressions are actually computed 668/// by the current IR. 669/// 670/// We currently allow expansion of IV increments that involve adds, 671/// multiplication by constants, and AddRecs from existing phis. 672/// 673/// TODO: Allow UDivExpr if we can find an existing IV increment that is an 674/// obvious multiple of the UDivExpr. 675static bool isHighCostExpansion(const SCEV *S, 676 SmallPtrSet<const SCEV*, 8> &Processed, 677 ScalarEvolution &SE) { 678 // Zero/One operand expressions 679 switch (S->getSCEVType()) { 680 case scUnknown: 681 case scConstant: 682 return false; 683 case scTruncate: 684 return isHighCostExpansion(cast<SCEVTruncateExpr>(S)->getOperand(), 685 Processed, SE); 686 case scZeroExtend: 687 return isHighCostExpansion(cast<SCEVZeroExtendExpr>(S)->getOperand(), 688 Processed, SE); 689 case scSignExtend: 690 return isHighCostExpansion(cast<SCEVSignExtendExpr>(S)->getOperand(), 691 Processed, SE); 692 } 693 694 if (!Processed.insert(S)) 695 return false; 696 697 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) { 698 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end(); 699 I != E; ++I) { 700 if (isHighCostExpansion(*I, Processed, SE)) 701 return true; 702 } 703 return false; 704 } 705 706 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) { 707 if (Mul->getNumOperands() == 2) { 708 // Multiplication by a constant is ok 709 if (isa<SCEVConstant>(Mul->getOperand(0))) 710 return isHighCostExpansion(Mul->getOperand(1), Processed, SE); 711 712 // If we have the value of one operand, check if an existing 713 // multiplication already generates this expression. 714 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(Mul->getOperand(1))) { 715 Value *UVal = U->getValue(); 716 for (Value::use_iterator UI = UVal->use_begin(), UE = UVal->use_end(); 717 UI != UE; ++UI) { 718 // If U is a constant, it may be used by a ConstantExpr. 719 Instruction *User = dyn_cast<Instruction>(*UI); 720 if (User && User->getOpcode() == Instruction::Mul 721 && SE.isSCEVable(User->getType())) { 722 return SE.getSCEV(User) == Mul; 723 } 724 } 725 } 726 } 727 } 728 729 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) { 730 if (isExistingPhi(AR, SE)) 731 return false; 732 } 733 734 // Fow now, consider any other type of expression (div/mul/min/max) high cost. 735 return true; 736} 737 738/// DeleteTriviallyDeadInstructions - If any of the instructions is the 739/// specified set are trivially dead, delete them and see if this makes any of 740/// their operands subsequently dead. 741static bool 742DeleteTriviallyDeadInstructions(SmallVectorImpl<WeakVH> &DeadInsts) { 743 bool Changed = false; 744 745 while (!DeadInsts.empty()) { 746 Value *V = DeadInsts.pop_back_val(); 747 Instruction *I = dyn_cast_or_null<Instruction>(V); 748 749 if (I == 0 || !isInstructionTriviallyDead(I)) 750 continue; 751 752 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI) 753 if (Instruction *U = dyn_cast<Instruction>(*OI)) { 754 *OI = 0; 755 if (U->use_empty()) 756 DeadInsts.push_back(U); 757 } 758 759 I->eraseFromParent(); 760 Changed = true; 761 } 762 763 return Changed; 764} 765 766namespace { 767 768/// Cost - This class is used to measure and compare candidate formulae. 769class Cost { 770 /// TODO: Some of these could be merged. Also, a lexical ordering 771 /// isn't always optimal. 772 unsigned NumRegs; 773 unsigned AddRecCost; 774 unsigned NumIVMuls; 775 unsigned NumBaseAdds; 776 unsigned ImmCost; 777 unsigned SetupCost; 778 779public: 780 Cost() 781 : NumRegs(0), AddRecCost(0), NumIVMuls(0), NumBaseAdds(0), ImmCost(0), 782 SetupCost(0) {} 783 784 bool operator<(const Cost &Other) const; 785 786 void Loose(); 787 788#ifndef NDEBUG 789 // Once any of the metrics loses, they must all remain losers. 790 bool isValid() { 791 return ((NumRegs | AddRecCost | NumIVMuls | NumBaseAdds 792 | ImmCost | SetupCost) != ~0u) 793 || ((NumRegs & AddRecCost & NumIVMuls & NumBaseAdds 794 & ImmCost & SetupCost) == ~0u); 795 } 796#endif 797 798 bool isLoser() { 799 assert(isValid() && "invalid cost"); 800 return NumRegs == ~0u; 801 } 802 803 void 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 SmallPtrSet<const SCEV *, 16> *LoserRegs = 0); 810 811 void print(raw_ostream &OS) const; 812 void dump() const; 813 814private: 815 void RateRegister(const SCEV *Reg, 816 SmallPtrSet<const SCEV *, 16> &Regs, 817 const Loop *L, 818 ScalarEvolution &SE, DominatorTree &DT); 819 void RatePrimaryRegister(const SCEV *Reg, 820 SmallPtrSet<const SCEV *, 16> &Regs, 821 const Loop *L, 822 ScalarEvolution &SE, DominatorTree &DT, 823 SmallPtrSet<const SCEV *, 16> *LoserRegs); 824}; 825 826} 827 828/// RateRegister - Tally up interesting quantities from the given register. 829void Cost::RateRegister(const SCEV *Reg, 830 SmallPtrSet<const SCEV *, 16> &Regs, 831 const Loop *L, 832 ScalarEvolution &SE, DominatorTree &DT) { 833 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Reg)) { 834 // If this is an addrec for another loop, don't second-guess its addrec phi 835 // nodes. LSR isn't currently smart enough to reason about more than one 836 // loop at a time. LSR has already run on inner loops, will not run on outer 837 // loops, and cannot be expected to change sibling loops. 838 if (AR->getLoop() != L) { 839 // If the AddRec exists, consider it's register free and leave it alone. 840 if (isExistingPhi(AR, SE)) 841 return; 842 843 // Otherwise, do not consider this formula at all. 844 Loose(); 845 return; 846 } 847 AddRecCost += 1; /// TODO: This should be a function of the stride. 848 849 // Add the step value register, if it needs one. 850 // TODO: The non-affine case isn't precisely modeled here. 851 if (!AR->isAffine() || !isa<SCEVConstant>(AR->getOperand(1))) { 852 if (!Regs.count(AR->getOperand(1))) { 853 RateRegister(AR->getOperand(1), Regs, L, SE, DT); 854 if (isLoser()) 855 return; 856 } 857 } 858 } 859 ++NumRegs; 860 861 // Rough heuristic; favor registers which don't require extra setup 862 // instructions in the preheader. 863 if (!isa<SCEVUnknown>(Reg) && 864 !isa<SCEVConstant>(Reg) && 865 !(isa<SCEVAddRecExpr>(Reg) && 866 (isa<SCEVUnknown>(cast<SCEVAddRecExpr>(Reg)->getStart()) || 867 isa<SCEVConstant>(cast<SCEVAddRecExpr>(Reg)->getStart())))) 868 ++SetupCost; 869 870 NumIVMuls += isa<SCEVMulExpr>(Reg) && 871 SE.hasComputableLoopEvolution(Reg, L); 872} 873 874/// RatePrimaryRegister - Record this register in the set. If we haven't seen it 875/// before, rate it. Optional LoserRegs provides a way to declare any formula 876/// that refers to one of those regs an instant loser. 877void Cost::RatePrimaryRegister(const SCEV *Reg, 878 SmallPtrSet<const SCEV *, 16> &Regs, 879 const Loop *L, 880 ScalarEvolution &SE, DominatorTree &DT, 881 SmallPtrSet<const SCEV *, 16> *LoserRegs) { 882 if (LoserRegs && LoserRegs->count(Reg)) { 883 Loose(); 884 return; 885 } 886 if (Regs.insert(Reg)) { 887 RateRegister(Reg, Regs, L, SE, DT); 888 if (isLoser()) 889 LoserRegs->insert(Reg); 890 } 891} 892 893void Cost::RateFormula(const Formula &F, 894 SmallPtrSet<const SCEV *, 16> &Regs, 895 const DenseSet<const SCEV *> &VisitedRegs, 896 const Loop *L, 897 const SmallVectorImpl<int64_t> &Offsets, 898 ScalarEvolution &SE, DominatorTree &DT, 899 SmallPtrSet<const SCEV *, 16> *LoserRegs) { 900 // Tally up the registers. 901 if (const SCEV *ScaledReg = F.ScaledReg) { 902 if (VisitedRegs.count(ScaledReg)) { 903 Loose(); 904 return; 905 } 906 RatePrimaryRegister(ScaledReg, Regs, L, SE, DT, LoserRegs); 907 if (isLoser()) 908 return; 909 } 910 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(), 911 E = F.BaseRegs.end(); I != E; ++I) { 912 const SCEV *BaseReg = *I; 913 if (VisitedRegs.count(BaseReg)) { 914 Loose(); 915 return; 916 } 917 RatePrimaryRegister(BaseReg, Regs, L, SE, DT, LoserRegs); 918 if (isLoser()) 919 return; 920 } 921 922 // Determine how many (unfolded) adds we'll need inside the loop. 923 size_t NumBaseParts = F.BaseRegs.size() + (F.UnfoldedOffset != 0); 924 if (NumBaseParts > 1) 925 NumBaseAdds += NumBaseParts - 1; 926 927 // Tally up the non-zero immediates. 928 for (SmallVectorImpl<int64_t>::const_iterator I = Offsets.begin(), 929 E = Offsets.end(); I != E; ++I) { 930 int64_t Offset = (uint64_t)*I + F.AM.BaseOffs; 931 if (F.AM.BaseGV) 932 ImmCost += 64; // Handle symbolic values conservatively. 933 // TODO: This should probably be the pointer size. 934 else if (Offset != 0) 935 ImmCost += APInt(64, Offset, true).getMinSignedBits(); 936 } 937 assert(isValid() && "invalid cost"); 938} 939 940/// Loose - Set this cost to a losing value. 941void Cost::Loose() { 942 NumRegs = ~0u; 943 AddRecCost = ~0u; 944 NumIVMuls = ~0u; 945 NumBaseAdds = ~0u; 946 ImmCost = ~0u; 947 SetupCost = ~0u; 948} 949 950/// operator< - Choose the lower cost. 951bool Cost::operator<(const Cost &Other) const { 952 if (NumRegs != Other.NumRegs) 953 return NumRegs < Other.NumRegs; 954 if (AddRecCost != Other.AddRecCost) 955 return AddRecCost < Other.AddRecCost; 956 if (NumIVMuls != Other.NumIVMuls) 957 return NumIVMuls < Other.NumIVMuls; 958 if (NumBaseAdds != Other.NumBaseAdds) 959 return NumBaseAdds < Other.NumBaseAdds; 960 if (ImmCost != Other.ImmCost) 961 return ImmCost < Other.ImmCost; 962 if (SetupCost != Other.SetupCost) 963 return SetupCost < Other.SetupCost; 964 return false; 965} 966 967void Cost::print(raw_ostream &OS) const { 968 OS << NumRegs << " reg" << (NumRegs == 1 ? "" : "s"); 969 if (AddRecCost != 0) 970 OS << ", with addrec cost " << AddRecCost; 971 if (NumIVMuls != 0) 972 OS << ", plus " << NumIVMuls << " IV mul" << (NumIVMuls == 1 ? "" : "s"); 973 if (NumBaseAdds != 0) 974 OS << ", plus " << NumBaseAdds << " base add" 975 << (NumBaseAdds == 1 ? "" : "s"); 976 if (ImmCost != 0) 977 OS << ", plus " << ImmCost << " imm cost"; 978 if (SetupCost != 0) 979 OS << ", plus " << SetupCost << " setup cost"; 980} 981 982#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) 983void Cost::dump() const { 984 print(errs()); errs() << '\n'; 985} 986#endif 987 988namespace { 989 990/// LSRFixup - An operand value in an instruction which is to be replaced 991/// with some equivalent, possibly strength-reduced, replacement. 992struct LSRFixup { 993 /// UserInst - The instruction which will be updated. 994 Instruction *UserInst; 995 996 /// OperandValToReplace - The operand of the instruction which will 997 /// be replaced. The operand may be used more than once; every instance 998 /// will be replaced. 999 Value *OperandValToReplace; 1000 1001 /// PostIncLoops - If this user is to use the post-incremented value of an 1002 /// induction variable, this variable is non-null and holds the loop 1003 /// associated with the induction variable. 1004 PostIncLoopSet PostIncLoops; 1005 1006 /// LUIdx - The index of the LSRUse describing the expression which 1007 /// this fixup needs, minus an offset (below). 1008 size_t LUIdx; 1009 1010 /// Offset - A constant offset to be added to the LSRUse expression. 1011 /// This allows multiple fixups to share the same LSRUse with different 1012 /// offsets, for example in an unrolled loop. 1013 int64_t Offset; 1014 1015 bool isUseFullyOutsideLoop(const Loop *L) const; 1016 1017 LSRFixup(); 1018 1019 void print(raw_ostream &OS) const; 1020 void dump() const; 1021}; 1022 1023} 1024 1025LSRFixup::LSRFixup() 1026 : UserInst(0), OperandValToReplace(0), LUIdx(~size_t(0)), Offset(0) {} 1027 1028/// isUseFullyOutsideLoop - Test whether this fixup always uses its 1029/// value outside of the given loop. 1030bool LSRFixup::isUseFullyOutsideLoop(const Loop *L) const { 1031 // PHI nodes use their value in their incoming blocks. 1032 if (const PHINode *PN = dyn_cast<PHINode>(UserInst)) { 1033 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) 1034 if (PN->getIncomingValue(i) == OperandValToReplace && 1035 L->contains(PN->getIncomingBlock(i))) 1036 return false; 1037 return true; 1038 } 1039 1040 return !L->contains(UserInst); 1041} 1042 1043void LSRFixup::print(raw_ostream &OS) const { 1044 OS << "UserInst="; 1045 // Store is common and interesting enough to be worth special-casing. 1046 if (StoreInst *Store = dyn_cast<StoreInst>(UserInst)) { 1047 OS << "store "; 1048 WriteAsOperand(OS, Store->getOperand(0), /*PrintType=*/false); 1049 } else if (UserInst->getType()->isVoidTy()) 1050 OS << UserInst->getOpcodeName(); 1051 else 1052 WriteAsOperand(OS, UserInst, /*PrintType=*/false); 1053 1054 OS << ", OperandValToReplace="; 1055 WriteAsOperand(OS, OperandValToReplace, /*PrintType=*/false); 1056 1057 for (PostIncLoopSet::const_iterator I = PostIncLoops.begin(), 1058 E = PostIncLoops.end(); I != E; ++I) { 1059 OS << ", PostIncLoop="; 1060 WriteAsOperand(OS, (*I)->getHeader(), /*PrintType=*/false); 1061 } 1062 1063 if (LUIdx != ~size_t(0)) 1064 OS << ", LUIdx=" << LUIdx; 1065 1066 if (Offset != 0) 1067 OS << ", Offset=" << Offset; 1068} 1069 1070#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) 1071void LSRFixup::dump() const { 1072 print(errs()); errs() << '\n'; 1073} 1074#endif 1075 1076namespace { 1077 1078/// UniquifierDenseMapInfo - A DenseMapInfo implementation for holding 1079/// DenseMaps and DenseSets of sorted SmallVectors of const SCEV*. 1080struct UniquifierDenseMapInfo { 1081 static SmallVector<const SCEV *, 2> getEmptyKey() { 1082 SmallVector<const SCEV *, 2> V; 1083 V.push_back(reinterpret_cast<const SCEV *>(-1)); 1084 return V; 1085 } 1086 1087 static SmallVector<const SCEV *, 2> getTombstoneKey() { 1088 SmallVector<const SCEV *, 2> V; 1089 V.push_back(reinterpret_cast<const SCEV *>(-2)); 1090 return V; 1091 } 1092 1093 static unsigned getHashValue(const SmallVector<const SCEV *, 2> &V) { 1094 unsigned Result = 0; 1095 for (SmallVectorImpl<const SCEV *>::const_iterator I = V.begin(), 1096 E = V.end(); I != E; ++I) 1097 Result ^= DenseMapInfo<const SCEV *>::getHashValue(*I); 1098 return Result; 1099 } 1100 1101 static bool isEqual(const SmallVector<const SCEV *, 2> &LHS, 1102 const SmallVector<const SCEV *, 2> &RHS) { 1103 return LHS == RHS; 1104 } 1105}; 1106 1107/// LSRUse - This class holds the state that LSR keeps for each use in 1108/// IVUsers, as well as uses invented by LSR itself. It includes information 1109/// about what kinds of things can be folded into the user, information about 1110/// the user itself, and information about how the use may be satisfied. 1111/// TODO: Represent multiple users of the same expression in common? 1112class LSRUse { 1113 DenseSet<SmallVector<const SCEV *, 2>, UniquifierDenseMapInfo> Uniquifier; 1114 1115public: 1116 /// KindType - An enum for a kind of use, indicating what types of 1117 /// scaled and immediate operands it might support. 1118 enum KindType { 1119 Basic, ///< A normal use, with no folding. 1120 Special, ///< A special case of basic, allowing -1 scales. 1121 Address, ///< An address use; folding according to TargetLowering 1122 ICmpZero ///< An equality icmp with both operands folded into one. 1123 // TODO: Add a generic icmp too? 1124 }; 1125 1126 KindType Kind; 1127 Type *AccessTy; 1128 1129 SmallVector<int64_t, 8> Offsets; 1130 int64_t MinOffset; 1131 int64_t MaxOffset; 1132 1133 /// AllFixupsOutsideLoop - This records whether all of the fixups using this 1134 /// LSRUse are outside of the loop, in which case some special-case heuristics 1135 /// may be used. 1136 bool AllFixupsOutsideLoop; 1137 1138 /// WidestFixupType - This records the widest use type for any fixup using 1139 /// this LSRUse. FindUseWithSimilarFormula can't consider uses with different 1140 /// max fixup widths to be equivalent, because the narrower one may be relying 1141 /// on the implicit truncation to truncate away bogus bits. 1142 Type *WidestFixupType; 1143 1144 /// Formulae - A list of ways to build a value that can satisfy this user. 1145 /// After the list is populated, one of these is selected heuristically and 1146 /// used to formulate a replacement for OperandValToReplace in UserInst. 1147 SmallVector<Formula, 12> Formulae; 1148 1149 /// Regs - The set of register candidates used by all formulae in this LSRUse. 1150 SmallPtrSet<const SCEV *, 4> Regs; 1151 1152 LSRUse(KindType K, Type *T) : Kind(K), AccessTy(T), 1153 MinOffset(INT64_MAX), 1154 MaxOffset(INT64_MIN), 1155 AllFixupsOutsideLoop(true), 1156 WidestFixupType(0) {} 1157 1158 bool HasFormulaWithSameRegs(const Formula &F) const; 1159 bool InsertFormula(const Formula &F); 1160 void DeleteFormula(Formula &F); 1161 void RecomputeRegs(size_t LUIdx, RegUseTracker &Reguses); 1162 1163 void print(raw_ostream &OS) const; 1164 void dump() const; 1165}; 1166 1167} 1168 1169/// HasFormula - Test whether this use as a formula which has the same 1170/// registers as the given formula. 1171bool LSRUse::HasFormulaWithSameRegs(const Formula &F) const { 1172 SmallVector<const SCEV *, 2> Key = F.BaseRegs; 1173 if (F.ScaledReg) Key.push_back(F.ScaledReg); 1174 // Unstable sort by host order ok, because this is only used for uniquifying. 1175 std::sort(Key.begin(), Key.end()); 1176 return Uniquifier.count(Key); 1177} 1178 1179/// InsertFormula - If the given formula has not yet been inserted, add it to 1180/// the list, and return true. Return false otherwise. 1181bool LSRUse::InsertFormula(const Formula &F) { 1182 SmallVector<const SCEV *, 2> Key = F.BaseRegs; 1183 if (F.ScaledReg) Key.push_back(F.ScaledReg); 1184 // Unstable sort by host order ok, because this is only used for uniquifying. 1185 std::sort(Key.begin(), Key.end()); 1186 1187 if (!Uniquifier.insert(Key).second) 1188 return false; 1189 1190 // Using a register to hold the value of 0 is not profitable. 1191 assert((!F.ScaledReg || !F.ScaledReg->isZero()) && 1192 "Zero allocated in a scaled register!"); 1193#ifndef NDEBUG 1194 for (SmallVectorImpl<const SCEV *>::const_iterator I = 1195 F.BaseRegs.begin(), E = F.BaseRegs.end(); I != E; ++I) 1196 assert(!(*I)->isZero() && "Zero allocated in a base register!"); 1197#endif 1198 1199 // Add the formula to the list. 1200 Formulae.push_back(F); 1201 1202 // Record registers now being used by this use. 1203 Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end()); 1204 1205 return true; 1206} 1207 1208/// DeleteFormula - Remove the given formula from this use's list. 1209void LSRUse::DeleteFormula(Formula &F) { 1210 if (&F != &Formulae.back()) 1211 std::swap(F, Formulae.back()); 1212 Formulae.pop_back(); 1213} 1214 1215/// RecomputeRegs - Recompute the Regs field, and update RegUses. 1216void LSRUse::RecomputeRegs(size_t LUIdx, RegUseTracker &RegUses) { 1217 // Now that we've filtered out some formulae, recompute the Regs set. 1218 SmallPtrSet<const SCEV *, 4> OldRegs = Regs; 1219 Regs.clear(); 1220 for (SmallVectorImpl<Formula>::const_iterator I = Formulae.begin(), 1221 E = Formulae.end(); I != E; ++I) { 1222 const Formula &F = *I; 1223 if (F.ScaledReg) Regs.insert(F.ScaledReg); 1224 Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end()); 1225 } 1226 1227 // Update the RegTracker. 1228 for (SmallPtrSet<const SCEV *, 4>::iterator I = OldRegs.begin(), 1229 E = OldRegs.end(); I != E; ++I) 1230 if (!Regs.count(*I)) 1231 RegUses.DropRegister(*I, LUIdx); 1232} 1233 1234void LSRUse::print(raw_ostream &OS) const { 1235 OS << "LSR Use: Kind="; 1236 switch (Kind) { 1237 case Basic: OS << "Basic"; break; 1238 case Special: OS << "Special"; break; 1239 case ICmpZero: OS << "ICmpZero"; break; 1240 case Address: 1241 OS << "Address of "; 1242 if (AccessTy->isPointerTy()) 1243 OS << "pointer"; // the full pointer type could be really verbose 1244 else 1245 OS << *AccessTy; 1246 } 1247 1248 OS << ", Offsets={"; 1249 for (SmallVectorImpl<int64_t>::const_iterator I = Offsets.begin(), 1250 E = Offsets.end(); I != E; ++I) { 1251 OS << *I; 1252 if (llvm::next(I) != E) 1253 OS << ','; 1254 } 1255 OS << '}'; 1256 1257 if (AllFixupsOutsideLoop) 1258 OS << ", all-fixups-outside-loop"; 1259 1260 if (WidestFixupType) 1261 OS << ", widest fixup type: " << *WidestFixupType; 1262} 1263 1264#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) 1265void LSRUse::dump() const { 1266 print(errs()); errs() << '\n'; 1267} 1268#endif 1269 1270/// isLegalUse - Test whether the use described by AM is "legal", meaning it can 1271/// be completely folded into the user instruction at isel time. This includes 1272/// address-mode folding and special icmp tricks. 1273static bool isLegalUse(const AddrMode &AM, 1274 LSRUse::KindType Kind, Type *AccessTy, 1275 const TargetLowering *TLI) { 1276 switch (Kind) { 1277 case LSRUse::Address: 1278 // If we have low-level target information, ask the target if it can 1279 // completely fold this address. 1280 if (TLI) return TLI->isLegalAddressingMode(AM, AccessTy); 1281 1282 // Otherwise, just guess that reg+reg addressing is legal. 1283 return !AM.BaseGV && AM.BaseOffs == 0 && AM.Scale <= 1; 1284 1285 case LSRUse::ICmpZero: 1286 // There's not even a target hook for querying whether it would be legal to 1287 // fold a GV into an ICmp. 1288 if (AM.BaseGV) 1289 return false; 1290 1291 // ICmp only has two operands; don't allow more than two non-trivial parts. 1292 if (AM.Scale != 0 && AM.HasBaseReg && AM.BaseOffs != 0) 1293 return false; 1294 1295 // ICmp only supports no scale or a -1 scale, as we can "fold" a -1 scale by 1296 // putting the scaled register in the other operand of the icmp. 1297 if (AM.Scale != 0 && AM.Scale != -1) 1298 return false; 1299 1300 // If we have low-level target information, ask the target if it can fold an 1301 // integer immediate on an icmp. 1302 if (AM.BaseOffs != 0) { 1303 if (!TLI) 1304 return false; 1305 // We have one of: 1306 // ICmpZero BaseReg + Offset => ICmp BaseReg, -Offset 1307 // ICmpZero -1*ScaleReg + Offset => ICmp ScaleReg, Offset 1308 // Offs is the ICmp immediate. 1309 int64_t Offs = AM.BaseOffs; 1310 if (AM.Scale == 0) 1311 Offs = -(uint64_t)Offs; // The cast does the right thing with INT64_MIN. 1312 return TLI->isLegalICmpImmediate(Offs); 1313 } 1314 1315 // ICmpZero BaseReg + -1*ScaleReg => ICmp BaseReg, ScaleReg 1316 return true; 1317 1318 case LSRUse::Basic: 1319 // Only handle single-register values. 1320 return !AM.BaseGV && AM.Scale == 0 && AM.BaseOffs == 0; 1321 1322 case LSRUse::Special: 1323 // Special case Basic to handle -1 scales. 1324 return !AM.BaseGV && (AM.Scale == 0 || AM.Scale == -1) && AM.BaseOffs == 0; 1325 } 1326 1327 llvm_unreachable("Invalid LSRUse Kind!"); 1328} 1329 1330static bool isLegalUse(AddrMode AM, 1331 int64_t MinOffset, int64_t MaxOffset, 1332 LSRUse::KindType Kind, Type *AccessTy, 1333 const TargetLowering *TLI) { 1334 // Check for overflow. 1335 if (((int64_t)((uint64_t)AM.BaseOffs + MinOffset) > AM.BaseOffs) != 1336 (MinOffset > 0)) 1337 return false; 1338 AM.BaseOffs = (uint64_t)AM.BaseOffs + MinOffset; 1339 if (isLegalUse(AM, Kind, AccessTy, TLI)) { 1340 AM.BaseOffs = (uint64_t)AM.BaseOffs - MinOffset; 1341 // Check for overflow. 1342 if (((int64_t)((uint64_t)AM.BaseOffs + MaxOffset) > AM.BaseOffs) != 1343 (MaxOffset > 0)) 1344 return false; 1345 AM.BaseOffs = (uint64_t)AM.BaseOffs + MaxOffset; 1346 return isLegalUse(AM, Kind, AccessTy, TLI); 1347 } 1348 return false; 1349} 1350 1351static bool isAlwaysFoldable(int64_t BaseOffs, 1352 GlobalValue *BaseGV, 1353 bool HasBaseReg, 1354 LSRUse::KindType Kind, Type *AccessTy, 1355 const TargetLowering *TLI) { 1356 // Fast-path: zero is always foldable. 1357 if (BaseOffs == 0 && !BaseGV) return true; 1358 1359 // Conservatively, create an address with an immediate and a 1360 // base and a scale. 1361 AddrMode AM; 1362 AM.BaseOffs = BaseOffs; 1363 AM.BaseGV = BaseGV; 1364 AM.HasBaseReg = HasBaseReg; 1365 AM.Scale = Kind == LSRUse::ICmpZero ? -1 : 1; 1366 1367 // Canonicalize a scale of 1 to a base register if the formula doesn't 1368 // already have a base register. 1369 if (!AM.HasBaseReg && AM.Scale == 1) { 1370 AM.Scale = 0; 1371 AM.HasBaseReg = true; 1372 } 1373 1374 return isLegalUse(AM, Kind, AccessTy, TLI); 1375} 1376 1377static bool isAlwaysFoldable(const SCEV *S, 1378 int64_t MinOffset, int64_t MaxOffset, 1379 bool HasBaseReg, 1380 LSRUse::KindType Kind, Type *AccessTy, 1381 const TargetLowering *TLI, 1382 ScalarEvolution &SE) { 1383 // Fast-path: zero is always foldable. 1384 if (S->isZero()) return true; 1385 1386 // Conservatively, create an address with an immediate and a 1387 // base and a scale. 1388 int64_t BaseOffs = ExtractImmediate(S, SE); 1389 GlobalValue *BaseGV = ExtractSymbol(S, SE); 1390 1391 // If there's anything else involved, it's not foldable. 1392 if (!S->isZero()) return false; 1393 1394 // Fast-path: zero is always foldable. 1395 if (BaseOffs == 0 && !BaseGV) return true; 1396 1397 // Conservatively, create an address with an immediate and a 1398 // base and a scale. 1399 AddrMode AM; 1400 AM.BaseOffs = BaseOffs; 1401 AM.BaseGV = BaseGV; 1402 AM.HasBaseReg = HasBaseReg; 1403 AM.Scale = Kind == LSRUse::ICmpZero ? -1 : 1; 1404 1405 return isLegalUse(AM, MinOffset, MaxOffset, Kind, AccessTy, TLI); 1406} 1407 1408namespace { 1409 1410/// UseMapDenseMapInfo - A DenseMapInfo implementation for holding 1411/// DenseMaps and DenseSets of pairs of const SCEV* and LSRUse::Kind. 1412struct UseMapDenseMapInfo { 1413 static std::pair<const SCEV *, LSRUse::KindType> getEmptyKey() { 1414 return std::make_pair(reinterpret_cast<const SCEV *>(-1), LSRUse::Basic); 1415 } 1416 1417 static std::pair<const SCEV *, LSRUse::KindType> getTombstoneKey() { 1418 return std::make_pair(reinterpret_cast<const SCEV *>(-2), LSRUse::Basic); 1419 } 1420 1421 static unsigned 1422 getHashValue(const std::pair<const SCEV *, LSRUse::KindType> &V) { 1423 unsigned Result = DenseMapInfo<const SCEV *>::getHashValue(V.first); 1424 Result ^= DenseMapInfo<unsigned>::getHashValue(unsigned(V.second)); 1425 return Result; 1426 } 1427 1428 static bool isEqual(const std::pair<const SCEV *, LSRUse::KindType> &LHS, 1429 const std::pair<const SCEV *, LSRUse::KindType> &RHS) { 1430 return LHS == RHS; 1431 } 1432}; 1433 1434/// IVInc - An individual increment in a Chain of IV increments. 1435/// Relate an IV user to an expression that computes the IV it uses from the IV 1436/// used by the previous link in the Chain. 1437/// 1438/// For the head of a chain, IncExpr holds the absolute SCEV expression for the 1439/// original IVOperand. The head of the chain's IVOperand is only valid during 1440/// chain collection, before LSR replaces IV users. During chain generation, 1441/// IncExpr can be used to find the new IVOperand that computes the same 1442/// expression. 1443struct IVInc { 1444 Instruction *UserInst; 1445 Value* IVOperand; 1446 const SCEV *IncExpr; 1447 1448 IVInc(Instruction *U, Value *O, const SCEV *E): 1449 UserInst(U), IVOperand(O), IncExpr(E) {} 1450}; 1451 1452// IVChain - The list of IV increments in program order. 1453// We typically add the head of a chain without finding subsequent links. 1454struct IVChain { 1455 SmallVector<IVInc,1> Incs; 1456 const SCEV *ExprBase; 1457 1458 IVChain() : ExprBase(0) {} 1459 1460 IVChain(const IVInc &Head, const SCEV *Base) 1461 : Incs(1, Head), ExprBase(Base) {} 1462 1463 typedef SmallVectorImpl<IVInc>::const_iterator const_iterator; 1464 1465 // begin - return the first increment in the chain. 1466 const_iterator begin() const { 1467 assert(!Incs.empty()); 1468 return llvm::next(Incs.begin()); 1469 } 1470 const_iterator end() const { 1471 return Incs.end(); 1472 } 1473 1474 // hasIncs - Returns true if this chain contains any increments. 1475 bool hasIncs() const { return Incs.size() >= 2; } 1476 1477 // add - Add an IVInc to the end of this chain. 1478 void add(const IVInc &X) { Incs.push_back(X); } 1479 1480 // tailUserInst - Returns the last UserInst in the chain. 1481 Instruction *tailUserInst() const { return Incs.back().UserInst; } 1482 1483 // isProfitableIncrement - Returns true if IncExpr can be profitably added to 1484 // this chain. 1485 bool isProfitableIncrement(const SCEV *OperExpr, 1486 const SCEV *IncExpr, 1487 ScalarEvolution&); 1488}; 1489 1490/// ChainUsers - Helper for CollectChains to track multiple IV increment uses. 1491/// Distinguish between FarUsers that definitely cross IV increments and 1492/// NearUsers that may be used between IV increments. 1493struct ChainUsers { 1494 SmallPtrSet<Instruction*, 4> FarUsers; 1495 SmallPtrSet<Instruction*, 4> NearUsers; 1496}; 1497 1498/// LSRInstance - This class holds state for the main loop strength reduction 1499/// logic. 1500class LSRInstance { 1501 IVUsers &IU; 1502 ScalarEvolution &SE; 1503 DominatorTree &DT; 1504 LoopInfo &LI; 1505 const TargetLowering *const TLI; 1506 Loop *const L; 1507 bool Changed; 1508 1509 /// IVIncInsertPos - This is the insert position that the current loop's 1510 /// induction variable increment should be placed. In simple loops, this is 1511 /// the latch block's terminator. But in more complicated cases, this is a 1512 /// position which will dominate all the in-loop post-increment users. 1513 Instruction *IVIncInsertPos; 1514 1515 /// Factors - Interesting factors between use strides. 1516 SmallSetVector<int64_t, 8> Factors; 1517 1518 /// Types - Interesting use types, to facilitate truncation reuse. 1519 SmallSetVector<Type *, 4> Types; 1520 1521 /// Fixups - The list of operands which are to be replaced. 1522 SmallVector<LSRFixup, 16> Fixups; 1523 1524 /// Uses - The list of interesting uses. 1525 SmallVector<LSRUse, 16> Uses; 1526 1527 /// RegUses - Track which uses use which register candidates. 1528 RegUseTracker RegUses; 1529 1530 // Limit the number of chains to avoid quadratic behavior. We don't expect to 1531 // have more than a few IV increment chains in a loop. Missing a Chain falls 1532 // back to normal LSR behavior for those uses. 1533 static const unsigned MaxChains = 8; 1534 1535 /// IVChainVec - IV users can form a chain of IV increments. 1536 SmallVector<IVChain, MaxChains> IVChainVec; 1537 1538 /// IVIncSet - IV users that belong to profitable IVChains. 1539 SmallPtrSet<Use*, MaxChains> IVIncSet; 1540 1541 void OptimizeShadowIV(); 1542 bool FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse); 1543 ICmpInst *OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse); 1544 void OptimizeLoopTermCond(); 1545 1546 void ChainInstruction(Instruction *UserInst, Instruction *IVOper, 1547 SmallVectorImpl<ChainUsers> &ChainUsersVec); 1548 void FinalizeChain(IVChain &Chain); 1549 void CollectChains(); 1550 void GenerateIVChain(const IVChain &Chain, SCEVExpander &Rewriter, 1551 SmallVectorImpl<WeakVH> &DeadInsts); 1552 1553 void CollectInterestingTypesAndFactors(); 1554 void CollectFixupsAndInitialFormulae(); 1555 1556 LSRFixup &getNewFixup() { 1557 Fixups.push_back(LSRFixup()); 1558 return Fixups.back(); 1559 } 1560 1561 // Support for sharing of LSRUses between LSRFixups. 1562 typedef DenseMap<std::pair<const SCEV *, LSRUse::KindType>, 1563 size_t, 1564 UseMapDenseMapInfo> UseMapTy; 1565 UseMapTy UseMap; 1566 1567 bool reconcileNewOffset(LSRUse &LU, int64_t NewOffset, bool HasBaseReg, 1568 LSRUse::KindType Kind, Type *AccessTy); 1569 1570 std::pair<size_t, int64_t> getUse(const SCEV *&Expr, 1571 LSRUse::KindType Kind, 1572 Type *AccessTy); 1573 1574 void DeleteUse(LSRUse &LU, size_t LUIdx); 1575 1576 LSRUse *FindUseWithSimilarFormula(const Formula &F, const LSRUse &OrigLU); 1577 1578 void InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx); 1579 void InsertSupplementalFormula(const SCEV *S, LSRUse &LU, size_t LUIdx); 1580 void CountRegisters(const Formula &F, size_t LUIdx); 1581 bool InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F); 1582 1583 void CollectLoopInvariantFixupsAndFormulae(); 1584 1585 void GenerateReassociations(LSRUse &LU, unsigned LUIdx, Formula Base, 1586 unsigned Depth = 0); 1587 void GenerateCombinations(LSRUse &LU, unsigned LUIdx, Formula Base); 1588 void GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx, Formula Base); 1589 void GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx, Formula Base); 1590 void GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx, Formula Base); 1591 void GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base); 1592 void GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base); 1593 void GenerateCrossUseConstantOffsets(); 1594 void GenerateAllReuseFormulae(); 1595 1596 void FilterOutUndesirableDedicatedRegisters(); 1597 1598 size_t EstimateSearchSpaceComplexity() const; 1599 void NarrowSearchSpaceByDetectingSupersets(); 1600 void NarrowSearchSpaceByCollapsingUnrolledCode(); 1601 void NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters(); 1602 void NarrowSearchSpaceByPickingWinnerRegs(); 1603 void NarrowSearchSpaceUsingHeuristics(); 1604 1605 void SolveRecurse(SmallVectorImpl<const Formula *> &Solution, 1606 Cost &SolutionCost, 1607 SmallVectorImpl<const Formula *> &Workspace, 1608 const Cost &CurCost, 1609 const SmallPtrSet<const SCEV *, 16> &CurRegs, 1610 DenseSet<const SCEV *> &VisitedRegs) const; 1611 void Solve(SmallVectorImpl<const Formula *> &Solution) const; 1612 1613 BasicBlock::iterator 1614 HoistInsertPosition(BasicBlock::iterator IP, 1615 const SmallVectorImpl<Instruction *> &Inputs) const; 1616 BasicBlock::iterator 1617 AdjustInsertPositionForExpand(BasicBlock::iterator IP, 1618 const LSRFixup &LF, 1619 const LSRUse &LU, 1620 SCEVExpander &Rewriter) const; 1621 1622 Value *Expand(const LSRFixup &LF, 1623 const Formula &F, 1624 BasicBlock::iterator IP, 1625 SCEVExpander &Rewriter, 1626 SmallVectorImpl<WeakVH> &DeadInsts) const; 1627 void RewriteForPHI(PHINode *PN, const LSRFixup &LF, 1628 const Formula &F, 1629 SCEVExpander &Rewriter, 1630 SmallVectorImpl<WeakVH> &DeadInsts, 1631 Pass *P) const; 1632 void Rewrite(const LSRFixup &LF, 1633 const Formula &F, 1634 SCEVExpander &Rewriter, 1635 SmallVectorImpl<WeakVH> &DeadInsts, 1636 Pass *P) const; 1637 void ImplementSolution(const SmallVectorImpl<const Formula *> &Solution, 1638 Pass *P); 1639 1640public: 1641 LSRInstance(const TargetLowering *tli, Loop *l, Pass *P); 1642 1643 bool getChanged() const { return Changed; } 1644 1645 void print_factors_and_types(raw_ostream &OS) const; 1646 void print_fixups(raw_ostream &OS) const; 1647 void print_uses(raw_ostream &OS) const; 1648 void print(raw_ostream &OS) const; 1649 void dump() const; 1650}; 1651 1652} 1653 1654/// OptimizeShadowIV - If IV is used in a int-to-float cast 1655/// inside the loop then try to eliminate the cast operation. 1656void LSRInstance::OptimizeShadowIV() { 1657 const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L); 1658 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount)) 1659 return; 1660 1661 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); 1662 UI != E; /* empty */) { 1663 IVUsers::const_iterator CandidateUI = UI; 1664 ++UI; 1665 Instruction *ShadowUse = CandidateUI->getUser(); 1666 Type *DestTy = NULL; 1667 bool IsSigned = false; 1668 1669 /* If shadow use is a int->float cast then insert a second IV 1670 to eliminate this cast. 1671 1672 for (unsigned i = 0; i < n; ++i) 1673 foo((double)i); 1674 1675 is transformed into 1676 1677 double d = 0.0; 1678 for (unsigned i = 0; i < n; ++i, ++d) 1679 foo(d); 1680 */ 1681 if (UIToFPInst *UCast = dyn_cast<UIToFPInst>(CandidateUI->getUser())) { 1682 IsSigned = false; 1683 DestTy = UCast->getDestTy(); 1684 } 1685 else if (SIToFPInst *SCast = dyn_cast<SIToFPInst>(CandidateUI->getUser())) { 1686 IsSigned = true; 1687 DestTy = SCast->getDestTy(); 1688 } 1689 if (!DestTy) continue; 1690 1691 if (TLI) { 1692 // If target does not support DestTy natively then do not apply 1693 // this transformation. 1694 EVT DVT = TLI->getValueType(DestTy); 1695 if (!TLI->isTypeLegal(DVT)) continue; 1696 } 1697 1698 PHINode *PH = dyn_cast<PHINode>(ShadowUse->getOperand(0)); 1699 if (!PH) continue; 1700 if (PH->getNumIncomingValues() != 2) continue; 1701 1702 Type *SrcTy = PH->getType(); 1703 int Mantissa = DestTy->getFPMantissaWidth(); 1704 if (Mantissa == -1) continue; 1705 if ((int)SE.getTypeSizeInBits(SrcTy) > Mantissa) 1706 continue; 1707 1708 unsigned Entry, Latch; 1709 if (PH->getIncomingBlock(0) == L->getLoopPreheader()) { 1710 Entry = 0; 1711 Latch = 1; 1712 } else { 1713 Entry = 1; 1714 Latch = 0; 1715 } 1716 1717 ConstantInt *Init = dyn_cast<ConstantInt>(PH->getIncomingValue(Entry)); 1718 if (!Init) continue; 1719 Constant *NewInit = ConstantFP::get(DestTy, IsSigned ? 1720 (double)Init->getSExtValue() : 1721 (double)Init->getZExtValue()); 1722 1723 BinaryOperator *Incr = 1724 dyn_cast<BinaryOperator>(PH->getIncomingValue(Latch)); 1725 if (!Incr) continue; 1726 if (Incr->getOpcode() != Instruction::Add 1727 && Incr->getOpcode() != Instruction::Sub) 1728 continue; 1729 1730 /* Initialize new IV, double d = 0.0 in above example. */ 1731 ConstantInt *C = NULL; 1732 if (Incr->getOperand(0) == PH) 1733 C = dyn_cast<ConstantInt>(Incr->getOperand(1)); 1734 else if (Incr->getOperand(1) == PH) 1735 C = dyn_cast<ConstantInt>(Incr->getOperand(0)); 1736 else 1737 continue; 1738 1739 if (!C) continue; 1740 1741 // Ignore negative constants, as the code below doesn't handle them 1742 // correctly. TODO: Remove this restriction. 1743 if (!C->getValue().isStrictlyPositive()) continue; 1744 1745 /* Add new PHINode. */ 1746 PHINode *NewPH = PHINode::Create(DestTy, 2, "IV.S.", PH); 1747 1748 /* create new increment. '++d' in above example. */ 1749 Constant *CFP = ConstantFP::get(DestTy, C->getZExtValue()); 1750 BinaryOperator *NewIncr = 1751 BinaryOperator::Create(Incr->getOpcode() == Instruction::Add ? 1752 Instruction::FAdd : Instruction::FSub, 1753 NewPH, CFP, "IV.S.next.", Incr); 1754 1755 NewPH->addIncoming(NewInit, PH->getIncomingBlock(Entry)); 1756 NewPH->addIncoming(NewIncr, PH->getIncomingBlock(Latch)); 1757 1758 /* Remove cast operation */ 1759 ShadowUse->replaceAllUsesWith(NewPH); 1760 ShadowUse->eraseFromParent(); 1761 Changed = true; 1762 break; 1763 } 1764} 1765 1766/// FindIVUserForCond - If Cond has an operand that is an expression of an IV, 1767/// set the IV user and stride information and return true, otherwise return 1768/// false. 1769bool LSRInstance::FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse) { 1770 for (IVUsers::iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI) 1771 if (UI->getUser() == Cond) { 1772 // NOTE: we could handle setcc instructions with multiple uses here, but 1773 // InstCombine does it as well for simple uses, it's not clear that it 1774 // occurs enough in real life to handle. 1775 CondUse = UI; 1776 return true; 1777 } 1778 return false; 1779} 1780 1781/// OptimizeMax - Rewrite the loop's terminating condition if it uses 1782/// a max computation. 1783/// 1784/// This is a narrow solution to a specific, but acute, problem. For loops 1785/// like this: 1786/// 1787/// i = 0; 1788/// do { 1789/// p[i] = 0.0; 1790/// } while (++i < n); 1791/// 1792/// the trip count isn't just 'n', because 'n' might not be positive. And 1793/// unfortunately this can come up even for loops where the user didn't use 1794/// a C do-while loop. For example, seemingly well-behaved top-test loops 1795/// will commonly be lowered like this: 1796// 1797/// if (n > 0) { 1798/// i = 0; 1799/// do { 1800/// p[i] = 0.0; 1801/// } while (++i < n); 1802/// } 1803/// 1804/// and then it's possible for subsequent optimization to obscure the if 1805/// test in such a way that indvars can't find it. 1806/// 1807/// When indvars can't find the if test in loops like this, it creates a 1808/// max expression, which allows it to give the loop a canonical 1809/// induction variable: 1810/// 1811/// i = 0; 1812/// max = n < 1 ? 1 : n; 1813/// do { 1814/// p[i] = 0.0; 1815/// } while (++i != max); 1816/// 1817/// Canonical induction variables are necessary because the loop passes 1818/// are designed around them. The most obvious example of this is the 1819/// LoopInfo analysis, which doesn't remember trip count values. It 1820/// expects to be able to rediscover the trip count each time it is 1821/// needed, and it does this using a simple analysis that only succeeds if 1822/// the loop has a canonical induction variable. 1823/// 1824/// However, when it comes time to generate code, the maximum operation 1825/// can be quite costly, especially if it's inside of an outer loop. 1826/// 1827/// This function solves this problem by detecting this type of loop and 1828/// rewriting their conditions from ICMP_NE back to ICMP_SLT, and deleting 1829/// the instructions for the maximum computation. 1830/// 1831ICmpInst *LSRInstance::OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse) { 1832 // Check that the loop matches the pattern we're looking for. 1833 if (Cond->getPredicate() != CmpInst::ICMP_EQ && 1834 Cond->getPredicate() != CmpInst::ICMP_NE) 1835 return Cond; 1836 1837 SelectInst *Sel = dyn_cast<SelectInst>(Cond->getOperand(1)); 1838 if (!Sel || !Sel->hasOneUse()) return Cond; 1839 1840 const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L); 1841 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount)) 1842 return Cond; 1843 const SCEV *One = SE.getConstant(BackedgeTakenCount->getType(), 1); 1844 1845 // Add one to the backedge-taken count to get the trip count. 1846 const SCEV *IterationCount = SE.getAddExpr(One, BackedgeTakenCount); 1847 if (IterationCount != SE.getSCEV(Sel)) return Cond; 1848 1849 // Check for a max calculation that matches the pattern. There's no check 1850 // for ICMP_ULE here because the comparison would be with zero, which 1851 // isn't interesting. 1852 CmpInst::Predicate Pred = ICmpInst::BAD_ICMP_PREDICATE; 1853 const SCEVNAryExpr *Max = 0; 1854 if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(BackedgeTakenCount)) { 1855 Pred = ICmpInst::ICMP_SLE; 1856 Max = S; 1857 } else if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(IterationCount)) { 1858 Pred = ICmpInst::ICMP_SLT; 1859 Max = S; 1860 } else if (const SCEVUMaxExpr *U = dyn_cast<SCEVUMaxExpr>(IterationCount)) { 1861 Pred = ICmpInst::ICMP_ULT; 1862 Max = U; 1863 } else { 1864 // No match; bail. 1865 return Cond; 1866 } 1867 1868 // To handle a max with more than two operands, this optimization would 1869 // require additional checking and setup. 1870 if (Max->getNumOperands() != 2) 1871 return Cond; 1872 1873 const SCEV *MaxLHS = Max->getOperand(0); 1874 const SCEV *MaxRHS = Max->getOperand(1); 1875 1876 // ScalarEvolution canonicalizes constants to the left. For < and >, look 1877 // for a comparison with 1. For <= and >=, a comparison with zero. 1878 if (!MaxLHS || 1879 (ICmpInst::isTrueWhenEqual(Pred) ? !MaxLHS->isZero() : (MaxLHS != One))) 1880 return Cond; 1881 1882 // Check the relevant induction variable for conformance to 1883 // the pattern. 1884 const SCEV *IV = SE.getSCEV(Cond->getOperand(0)); 1885 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(IV); 1886 if (!AR || !AR->isAffine() || 1887 AR->getStart() != One || 1888 AR->getStepRecurrence(SE) != One) 1889 return Cond; 1890 1891 assert(AR->getLoop() == L && 1892 "Loop condition operand is an addrec in a different loop!"); 1893 1894 // Check the right operand of the select, and remember it, as it will 1895 // be used in the new comparison instruction. 1896 Value *NewRHS = 0; 1897 if (ICmpInst::isTrueWhenEqual(Pred)) { 1898 // Look for n+1, and grab n. 1899 if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(1))) 1900 if (isa<ConstantInt>(BO->getOperand(1)) && 1901 cast<ConstantInt>(BO->getOperand(1))->isOne() && 1902 SE.getSCEV(BO->getOperand(0)) == MaxRHS) 1903 NewRHS = BO->getOperand(0); 1904 if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(2))) 1905 if (isa<ConstantInt>(BO->getOperand(1)) && 1906 cast<ConstantInt>(BO->getOperand(1))->isOne() && 1907 SE.getSCEV(BO->getOperand(0)) == MaxRHS) 1908 NewRHS = BO->getOperand(0); 1909 if (!NewRHS) 1910 return Cond; 1911 } else if (SE.getSCEV(Sel->getOperand(1)) == MaxRHS) 1912 NewRHS = Sel->getOperand(1); 1913 else if (SE.getSCEV(Sel->getOperand(2)) == MaxRHS) 1914 NewRHS = Sel->getOperand(2); 1915 else if (const SCEVUnknown *SU = dyn_cast<SCEVUnknown>(MaxRHS)) 1916 NewRHS = SU->getValue(); 1917 else 1918 // Max doesn't match expected pattern. 1919 return Cond; 1920 1921 // Determine the new comparison opcode. It may be signed or unsigned, 1922 // and the original comparison may be either equality or inequality. 1923 if (Cond->getPredicate() == CmpInst::ICMP_EQ) 1924 Pred = CmpInst::getInversePredicate(Pred); 1925 1926 // Ok, everything looks ok to change the condition into an SLT or SGE and 1927 // delete the max calculation. 1928 ICmpInst *NewCond = 1929 new ICmpInst(Cond, Pred, Cond->getOperand(0), NewRHS, "scmp"); 1930 1931 // Delete the max calculation instructions. 1932 Cond->replaceAllUsesWith(NewCond); 1933 CondUse->setUser(NewCond); 1934 Instruction *Cmp = cast<Instruction>(Sel->getOperand(0)); 1935 Cond->eraseFromParent(); 1936 Sel->eraseFromParent(); 1937 if (Cmp->use_empty()) 1938 Cmp->eraseFromParent(); 1939 return NewCond; 1940} 1941 1942/// OptimizeLoopTermCond - Change loop terminating condition to use the 1943/// postinc iv when possible. 1944void 1945LSRInstance::OptimizeLoopTermCond() { 1946 SmallPtrSet<Instruction *, 4> PostIncs; 1947 1948 BasicBlock *LatchBlock = L->getLoopLatch(); 1949 SmallVector<BasicBlock*, 8> ExitingBlocks; 1950 L->getExitingBlocks(ExitingBlocks); 1951 1952 for (unsigned i = 0, e = ExitingBlocks.size(); i != e; ++i) { 1953 BasicBlock *ExitingBlock = ExitingBlocks[i]; 1954 1955 // Get the terminating condition for the loop if possible. If we 1956 // can, we want to change it to use a post-incremented version of its 1957 // induction variable, to allow coalescing the live ranges for the IV into 1958 // one register value. 1959 1960 BranchInst *TermBr = dyn_cast<BranchInst>(ExitingBlock->getTerminator()); 1961 if (!TermBr) 1962 continue; 1963 // FIXME: Overly conservative, termination condition could be an 'or' etc.. 1964 if (TermBr->isUnconditional() || !isa<ICmpInst>(TermBr->getCondition())) 1965 continue; 1966 1967 // Search IVUsesByStride to find Cond's IVUse if there is one. 1968 IVStrideUse *CondUse = 0; 1969 ICmpInst *Cond = cast<ICmpInst>(TermBr->getCondition()); 1970 if (!FindIVUserForCond(Cond, CondUse)) 1971 continue; 1972 1973 // If the trip count is computed in terms of a max (due to ScalarEvolution 1974 // being unable to find a sufficient guard, for example), change the loop 1975 // comparison to use SLT or ULT instead of NE. 1976 // One consequence of doing this now is that it disrupts the count-down 1977 // optimization. That's not always a bad thing though, because in such 1978 // cases it may still be worthwhile to avoid a max. 1979 Cond = OptimizeMax(Cond, CondUse); 1980 1981 // If this exiting block dominates the latch block, it may also use 1982 // the post-inc value if it won't be shared with other uses. 1983 // Check for dominance. 1984 if (!DT.dominates(ExitingBlock, LatchBlock)) 1985 continue; 1986 1987 // Conservatively avoid trying to use the post-inc value in non-latch 1988 // exits if there may be pre-inc users in intervening blocks. 1989 if (LatchBlock != ExitingBlock) 1990 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI) 1991 // Test if the use is reachable from the exiting block. This dominator 1992 // query is a conservative approximation of reachability. 1993 if (&*UI != CondUse && 1994 !DT.properlyDominates(UI->getUser()->getParent(), ExitingBlock)) { 1995 // Conservatively assume there may be reuse if the quotient of their 1996 // strides could be a legal scale. 1997 const SCEV *A = IU.getStride(*CondUse, L); 1998 const SCEV *B = IU.getStride(*UI, L); 1999 if (!A || !B) continue; 2000 if (SE.getTypeSizeInBits(A->getType()) != 2001 SE.getTypeSizeInBits(B->getType())) { 2002 if (SE.getTypeSizeInBits(A->getType()) > 2003 SE.getTypeSizeInBits(B->getType())) 2004 B = SE.getSignExtendExpr(B, A->getType()); 2005 else 2006 A = SE.getSignExtendExpr(A, B->getType()); 2007 } 2008 if (const SCEVConstant *D = 2009 dyn_cast_or_null<SCEVConstant>(getExactSDiv(B, A, SE))) { 2010 const ConstantInt *C = D->getValue(); 2011 // Stride of one or negative one can have reuse with non-addresses. 2012 if (C->isOne() || C->isAllOnesValue()) 2013 goto decline_post_inc; 2014 // Avoid weird situations. 2015 if (C->getValue().getMinSignedBits() >= 64 || 2016 C->getValue().isMinSignedValue()) 2017 goto decline_post_inc; 2018 // Without TLI, assume that any stride might be valid, and so any 2019 // use might be shared. 2020 if (!TLI) 2021 goto decline_post_inc; 2022 // Check for possible scaled-address reuse. 2023 Type *AccessTy = getAccessType(UI->getUser()); 2024 AddrMode AM; 2025 AM.Scale = C->getSExtValue(); 2026 if (TLI->isLegalAddressingMode(AM, AccessTy)) 2027 goto decline_post_inc; 2028 AM.Scale = -AM.Scale; 2029 if (TLI->isLegalAddressingMode(AM, AccessTy)) 2030 goto decline_post_inc; 2031 } 2032 } 2033 2034 DEBUG(dbgs() << " Change loop exiting icmp to use postinc iv: " 2035 << *Cond << '\n'); 2036 2037 // It's possible for the setcc instruction to be anywhere in the loop, and 2038 // possible for it to have multiple users. If it is not immediately before 2039 // the exiting block branch, move it. 2040 if (&*++BasicBlock::iterator(Cond) != TermBr) { 2041 if (Cond->hasOneUse()) { 2042 Cond->moveBefore(TermBr); 2043 } else { 2044 // Clone the terminating condition and insert into the loopend. 2045 ICmpInst *OldCond = Cond; 2046 Cond = cast<ICmpInst>(Cond->clone()); 2047 Cond->setName(L->getHeader()->getName() + ".termcond"); 2048 ExitingBlock->getInstList().insert(TermBr, Cond); 2049 2050 // Clone the IVUse, as the old use still exists! 2051 CondUse = &IU.AddUser(Cond, CondUse->getOperandValToReplace()); 2052 TermBr->replaceUsesOfWith(OldCond, Cond); 2053 } 2054 } 2055 2056 // If we get to here, we know that we can transform the setcc instruction to 2057 // use the post-incremented version of the IV, allowing us to coalesce the 2058 // live ranges for the IV correctly. 2059 CondUse->transformToPostInc(L); 2060 Changed = true; 2061 2062 PostIncs.insert(Cond); 2063 decline_post_inc:; 2064 } 2065 2066 // Determine an insertion point for the loop induction variable increment. It 2067 // must dominate all the post-inc comparisons we just set up, and it must 2068 // dominate the loop latch edge. 2069 IVIncInsertPos = L->getLoopLatch()->getTerminator(); 2070 for (SmallPtrSet<Instruction *, 4>::const_iterator I = PostIncs.begin(), 2071 E = PostIncs.end(); I != E; ++I) { 2072 BasicBlock *BB = 2073 DT.findNearestCommonDominator(IVIncInsertPos->getParent(), 2074 (*I)->getParent()); 2075 if (BB == (*I)->getParent()) 2076 IVIncInsertPos = *I; 2077 else if (BB != IVIncInsertPos->getParent()) 2078 IVIncInsertPos = BB->getTerminator(); 2079 } 2080} 2081 2082/// reconcileNewOffset - Determine if the given use can accommodate a fixup 2083/// at the given offset and other details. If so, update the use and 2084/// return true. 2085bool 2086LSRInstance::reconcileNewOffset(LSRUse &LU, int64_t NewOffset, bool HasBaseReg, 2087 LSRUse::KindType Kind, Type *AccessTy) { 2088 int64_t NewMinOffset = LU.MinOffset; 2089 int64_t NewMaxOffset = LU.MaxOffset; 2090 Type *NewAccessTy = AccessTy; 2091 2092 // Check for a mismatched kind. It's tempting to collapse mismatched kinds to 2093 // something conservative, however this can pessimize in the case that one of 2094 // the uses will have all its uses outside the loop, for example. 2095 if (LU.Kind != Kind) 2096 return false; 2097 // Conservatively assume HasBaseReg is true for now. 2098 if (NewOffset < LU.MinOffset) { 2099 if (!isAlwaysFoldable(LU.MaxOffset - NewOffset, 0, HasBaseReg, 2100 Kind, AccessTy, TLI)) 2101 return false; 2102 NewMinOffset = NewOffset; 2103 } else if (NewOffset > LU.MaxOffset) { 2104 if (!isAlwaysFoldable(NewOffset - LU.MinOffset, 0, HasBaseReg, 2105 Kind, AccessTy, TLI)) 2106 return false; 2107 NewMaxOffset = NewOffset; 2108 } 2109 // Check for a mismatched access type, and fall back conservatively as needed. 2110 // TODO: Be less conservative when the type is similar and can use the same 2111 // addressing modes. 2112 if (Kind == LSRUse::Address && AccessTy != LU.AccessTy) 2113 NewAccessTy = Type::getVoidTy(AccessTy->getContext()); 2114 2115 // Update the use. 2116 LU.MinOffset = NewMinOffset; 2117 LU.MaxOffset = NewMaxOffset; 2118 LU.AccessTy = NewAccessTy; 2119 if (NewOffset != LU.Offsets.back()) 2120 LU.Offsets.push_back(NewOffset); 2121 return true; 2122} 2123 2124/// getUse - Return an LSRUse index and an offset value for a fixup which 2125/// needs the given expression, with the given kind and optional access type. 2126/// Either reuse an existing use or create a new one, as needed. 2127std::pair<size_t, int64_t> 2128LSRInstance::getUse(const SCEV *&Expr, 2129 LSRUse::KindType Kind, Type *AccessTy) { 2130 const SCEV *Copy = Expr; 2131 int64_t Offset = ExtractImmediate(Expr, SE); 2132 2133 // Basic uses can't accept any offset, for example. 2134 if (!isAlwaysFoldable(Offset, 0, /*HasBaseReg=*/true, Kind, AccessTy, TLI)) { 2135 Expr = Copy; 2136 Offset = 0; 2137 } 2138 2139 std::pair<UseMapTy::iterator, bool> P = 2140 UseMap.insert(std::make_pair(std::make_pair(Expr, Kind), 0)); 2141 if (!P.second) { 2142 // A use already existed with this base. 2143 size_t LUIdx = P.first->second; 2144 LSRUse &LU = Uses[LUIdx]; 2145 if (reconcileNewOffset(LU, Offset, /*HasBaseReg=*/true, Kind, AccessTy)) 2146 // Reuse this use. 2147 return std::make_pair(LUIdx, Offset); 2148 } 2149 2150 // Create a new use. 2151 size_t LUIdx = Uses.size(); 2152 P.first->second = LUIdx; 2153 Uses.push_back(LSRUse(Kind, AccessTy)); 2154 LSRUse &LU = Uses[LUIdx]; 2155 2156 // We don't need to track redundant offsets, but we don't need to go out 2157 // of our way here to avoid them. 2158 if (LU.Offsets.empty() || Offset != LU.Offsets.back()) 2159 LU.Offsets.push_back(Offset); 2160 2161 LU.MinOffset = Offset; 2162 LU.MaxOffset = Offset; 2163 return std::make_pair(LUIdx, Offset); 2164} 2165 2166/// DeleteUse - Delete the given use from the Uses list. 2167void LSRInstance::DeleteUse(LSRUse &LU, size_t LUIdx) { 2168 if (&LU != &Uses.back()) 2169 std::swap(LU, Uses.back()); 2170 Uses.pop_back(); 2171 2172 // Update RegUses. 2173 RegUses.SwapAndDropUse(LUIdx, Uses.size()); 2174} 2175 2176/// FindUseWithFormula - Look for a use distinct from OrigLU which is has 2177/// a formula that has the same registers as the given formula. 2178LSRUse * 2179LSRInstance::FindUseWithSimilarFormula(const Formula &OrigF, 2180 const LSRUse &OrigLU) { 2181 // Search all uses for the formula. This could be more clever. 2182 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) { 2183 LSRUse &LU = Uses[LUIdx]; 2184 // Check whether this use is close enough to OrigLU, to see whether it's 2185 // worthwhile looking through its formulae. 2186 // Ignore ICmpZero uses because they may contain formulae generated by 2187 // GenerateICmpZeroScales, in which case adding fixup offsets may 2188 // be invalid. 2189 if (&LU != &OrigLU && 2190 LU.Kind != LSRUse::ICmpZero && 2191 LU.Kind == OrigLU.Kind && OrigLU.AccessTy == LU.AccessTy && 2192 LU.WidestFixupType == OrigLU.WidestFixupType && 2193 LU.HasFormulaWithSameRegs(OrigF)) { 2194 // Scan through this use's formulae. 2195 for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(), 2196 E = LU.Formulae.end(); I != E; ++I) { 2197 const Formula &F = *I; 2198 // Check to see if this formula has the same registers and symbols 2199 // as OrigF. 2200 if (F.BaseRegs == OrigF.BaseRegs && 2201 F.ScaledReg == OrigF.ScaledReg && 2202 F.AM.BaseGV == OrigF.AM.BaseGV && 2203 F.AM.Scale == OrigF.AM.Scale && 2204 F.UnfoldedOffset == OrigF.UnfoldedOffset) { 2205 if (F.AM.BaseOffs == 0) 2206 return &LU; 2207 // This is the formula where all the registers and symbols matched; 2208 // there aren't going to be any others. Since we declined it, we 2209 // can skip the rest of the formulae and proceed to the next LSRUse. 2210 break; 2211 } 2212 } 2213 } 2214 } 2215 2216 // Nothing looked good. 2217 return 0; 2218} 2219 2220void LSRInstance::CollectInterestingTypesAndFactors() { 2221 SmallSetVector<const SCEV *, 4> Strides; 2222 2223 // Collect interesting types and strides. 2224 SmallVector<const SCEV *, 4> Worklist; 2225 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI) { 2226 const SCEV *Expr = IU.getExpr(*UI); 2227 2228 // Collect interesting types. 2229 Types.insert(SE.getEffectiveSCEVType(Expr->getType())); 2230 2231 // Add strides for mentioned loops. 2232 Worklist.push_back(Expr); 2233 do { 2234 const SCEV *S = Worklist.pop_back_val(); 2235 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) { 2236 if (AR->getLoop() == L) 2237 Strides.insert(AR->getStepRecurrence(SE)); 2238 Worklist.push_back(AR->getStart()); 2239 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) { 2240 Worklist.append(Add->op_begin(), Add->op_end()); 2241 } 2242 } while (!Worklist.empty()); 2243 } 2244 2245 // Compute interesting factors from the set of interesting strides. 2246 for (SmallSetVector<const SCEV *, 4>::const_iterator 2247 I = Strides.begin(), E = Strides.end(); I != E; ++I) 2248 for (SmallSetVector<const SCEV *, 4>::const_iterator NewStrideIter = 2249 llvm::next(I); NewStrideIter != E; ++NewStrideIter) { 2250 const SCEV *OldStride = *I; 2251 const SCEV *NewStride = *NewStrideIter; 2252 2253 if (SE.getTypeSizeInBits(OldStride->getType()) != 2254 SE.getTypeSizeInBits(NewStride->getType())) { 2255 if (SE.getTypeSizeInBits(OldStride->getType()) > 2256 SE.getTypeSizeInBits(NewStride->getType())) 2257 NewStride = SE.getSignExtendExpr(NewStride, OldStride->getType()); 2258 else 2259 OldStride = SE.getSignExtendExpr(OldStride, NewStride->getType()); 2260 } 2261 if (const SCEVConstant *Factor = 2262 dyn_cast_or_null<SCEVConstant>(getExactSDiv(NewStride, OldStride, 2263 SE, true))) { 2264 if (Factor->getValue()->getValue().getMinSignedBits() <= 64) 2265 Factors.insert(Factor->getValue()->getValue().getSExtValue()); 2266 } else if (const SCEVConstant *Factor = 2267 dyn_cast_or_null<SCEVConstant>(getExactSDiv(OldStride, 2268 NewStride, 2269 SE, true))) { 2270 if (Factor->getValue()->getValue().getMinSignedBits() <= 64) 2271 Factors.insert(Factor->getValue()->getValue().getSExtValue()); 2272 } 2273 } 2274 2275 // If all uses use the same type, don't bother looking for truncation-based 2276 // reuse. 2277 if (Types.size() == 1) 2278 Types.clear(); 2279 2280 DEBUG(print_factors_and_types(dbgs())); 2281} 2282 2283/// findIVOperand - Helper for CollectChains that finds an IV operand (computed 2284/// by an AddRec in this loop) within [OI,OE) or returns OE. If IVUsers mapped 2285/// Instructions to IVStrideUses, we could partially skip this. 2286static User::op_iterator 2287findIVOperand(User::op_iterator OI, User::op_iterator OE, 2288 Loop *L, ScalarEvolution &SE) { 2289 for(; OI != OE; ++OI) { 2290 if (Instruction *Oper = dyn_cast<Instruction>(*OI)) { 2291 if (!SE.isSCEVable(Oper->getType())) 2292 continue; 2293 2294 if (const SCEVAddRecExpr *AR = 2295 dyn_cast<SCEVAddRecExpr>(SE.getSCEV(Oper))) { 2296 if (AR->getLoop() == L) 2297 break; 2298 } 2299 } 2300 } 2301 return OI; 2302} 2303 2304/// getWideOperand - IVChain logic must consistenctly peek base TruncInst 2305/// operands, so wrap it in a convenient helper. 2306static Value *getWideOperand(Value *Oper) { 2307 if (TruncInst *Trunc = dyn_cast<TruncInst>(Oper)) 2308 return Trunc->getOperand(0); 2309 return Oper; 2310} 2311 2312/// isCompatibleIVType - Return true if we allow an IV chain to include both 2313/// types. 2314static bool isCompatibleIVType(Value *LVal, Value *RVal) { 2315 Type *LType = LVal->getType(); 2316 Type *RType = RVal->getType(); 2317 return (LType == RType) || (LType->isPointerTy() && RType->isPointerTy()); 2318} 2319 2320/// getExprBase - Return an approximation of this SCEV expression's "base", or 2321/// NULL for any constant. Returning the expression itself is 2322/// conservative. Returning a deeper subexpression is more precise and valid as 2323/// long as it isn't less complex than another subexpression. For expressions 2324/// involving multiple unscaled values, we need to return the pointer-type 2325/// SCEVUnknown. This avoids forming chains across objects, such as: 2326/// PrevOper==a[i], IVOper==b[i], IVInc==b-a. 2327/// 2328/// Since SCEVUnknown is the rightmost type, and pointers are the rightmost 2329/// SCEVUnknown, we simply return the rightmost SCEV operand. 2330static const SCEV *getExprBase(const SCEV *S) { 2331 switch (S->getSCEVType()) { 2332 default: // uncluding scUnknown. 2333 return S; 2334 case scConstant: 2335 return 0; 2336 case scTruncate: 2337 return getExprBase(cast<SCEVTruncateExpr>(S)->getOperand()); 2338 case scZeroExtend: 2339 return getExprBase(cast<SCEVZeroExtendExpr>(S)->getOperand()); 2340 case scSignExtend: 2341 return getExprBase(cast<SCEVSignExtendExpr>(S)->getOperand()); 2342 case scAddExpr: { 2343 // Skip over scaled operands (scMulExpr) to follow add operands as long as 2344 // there's nothing more complex. 2345 // FIXME: not sure if we want to recognize negation. 2346 const SCEVAddExpr *Add = cast<SCEVAddExpr>(S); 2347 for (std::reverse_iterator<SCEVAddExpr::op_iterator> I(Add->op_end()), 2348 E(Add->op_begin()); I != E; ++I) { 2349 const SCEV *SubExpr = *I; 2350 if (SubExpr->getSCEVType() == scAddExpr) 2351 return getExprBase(SubExpr); 2352 2353 if (SubExpr->getSCEVType() != scMulExpr) 2354 return SubExpr; 2355 } 2356 return S; // all operands are scaled, be conservative. 2357 } 2358 case scAddRecExpr: 2359 return getExprBase(cast<SCEVAddRecExpr>(S)->getStart()); 2360 } 2361} 2362 2363/// Return true if the chain increment is profitable to expand into a loop 2364/// invariant value, which may require its own register. A profitable chain 2365/// increment will be an offset relative to the same base. We allow such offsets 2366/// to potentially be used as chain increment as long as it's not obviously 2367/// expensive to expand using real instructions. 2368bool IVChain::isProfitableIncrement(const SCEV *OperExpr, 2369 const SCEV *IncExpr, 2370 ScalarEvolution &SE) { 2371 // Aggressively form chains when -stress-ivchain. 2372 if (StressIVChain) 2373 return true; 2374 2375 // Do not replace a constant offset from IV head with a nonconstant IV 2376 // increment. 2377 if (!isa<SCEVConstant>(IncExpr)) { 2378 const SCEV *HeadExpr = SE.getSCEV(getWideOperand(Incs[0].IVOperand)); 2379 if (isa<SCEVConstant>(SE.getMinusSCEV(OperExpr, HeadExpr))) 2380 return 0; 2381 } 2382 2383 SmallPtrSet<const SCEV*, 8> Processed; 2384 return !isHighCostExpansion(IncExpr, Processed, SE); 2385} 2386 2387/// Return true if the number of registers needed for the chain is estimated to 2388/// be less than the number required for the individual IV users. First prohibit 2389/// any IV users that keep the IV live across increments (the Users set should 2390/// be empty). Next count the number and type of increments in the chain. 2391/// 2392/// Chaining IVs can lead to considerable code bloat if ISEL doesn't 2393/// effectively use postinc addressing modes. Only consider it profitable it the 2394/// increments can be computed in fewer registers when chained. 2395/// 2396/// TODO: Consider IVInc free if it's already used in another chains. 2397static bool 2398isProfitableChain(IVChain &Chain, SmallPtrSet<Instruction*, 4> &Users, 2399 ScalarEvolution &SE, const TargetLowering *TLI) { 2400 if (StressIVChain) 2401 return true; 2402 2403 if (!Chain.hasIncs()) 2404 return false; 2405 2406 if (!Users.empty()) { 2407 DEBUG(dbgs() << "Chain: " << *Chain.Incs[0].UserInst << " users:\n"; 2408 for (SmallPtrSet<Instruction*, 4>::const_iterator I = Users.begin(), 2409 E = Users.end(); I != E; ++I) { 2410 dbgs() << " " << **I << "\n"; 2411 }); 2412 return false; 2413 } 2414 assert(!Chain.Incs.empty() && "empty IV chains are not allowed"); 2415 2416 // The chain itself may require a register, so intialize cost to 1. 2417 int cost = 1; 2418 2419 // A complete chain likely eliminates the need for keeping the original IV in 2420 // a register. LSR does not currently know how to form a complete chain unless 2421 // the header phi already exists. 2422 if (isa<PHINode>(Chain.tailUserInst()) 2423 && SE.getSCEV(Chain.tailUserInst()) == Chain.Incs[0].IncExpr) { 2424 --cost; 2425 } 2426 const SCEV *LastIncExpr = 0; 2427 unsigned NumConstIncrements = 0; 2428 unsigned NumVarIncrements = 0; 2429 unsigned NumReusedIncrements = 0; 2430 for (IVChain::const_iterator I = Chain.begin(), E = Chain.end(); 2431 I != E; ++I) { 2432 2433 if (I->IncExpr->isZero()) 2434 continue; 2435 2436 // Incrementing by zero or some constant is neutral. We assume constants can 2437 // be folded into an addressing mode or an add's immediate operand. 2438 if (isa<SCEVConstant>(I->IncExpr)) { 2439 ++NumConstIncrements; 2440 continue; 2441 } 2442 2443 if (I->IncExpr == LastIncExpr) 2444 ++NumReusedIncrements; 2445 else 2446 ++NumVarIncrements; 2447 2448 LastIncExpr = I->IncExpr; 2449 } 2450 // An IV chain with a single increment is handled by LSR's postinc 2451 // uses. However, a chain with multiple increments requires keeping the IV's 2452 // value live longer than it needs to be if chained. 2453 if (NumConstIncrements > 1) 2454 --cost; 2455 2456 // Materializing increment expressions in the preheader that didn't exist in 2457 // the original code may cost a register. For example, sign-extended array 2458 // indices can produce ridiculous increments like this: 2459 // IV + ((sext i32 (2 * %s) to i64) + (-1 * (sext i32 %s to i64))) 2460 cost += NumVarIncrements; 2461 2462 // Reusing variable increments likely saves a register to hold the multiple of 2463 // the stride. 2464 cost -= NumReusedIncrements; 2465 2466 DEBUG(dbgs() << "Chain: " << *Chain.Incs[0].UserInst << " Cost: " << cost 2467 << "\n"); 2468 2469 return cost < 0; 2470} 2471 2472/// ChainInstruction - Add this IV user to an existing chain or make it the head 2473/// of a new chain. 2474void LSRInstance::ChainInstruction(Instruction *UserInst, Instruction *IVOper, 2475 SmallVectorImpl<ChainUsers> &ChainUsersVec) { 2476 // When IVs are used as types of varying widths, they are generally converted 2477 // to a wider type with some uses remaining narrow under a (free) trunc. 2478 Value *const NextIV = getWideOperand(IVOper); 2479 const SCEV *const OperExpr = SE.getSCEV(NextIV); 2480 const SCEV *const OperExprBase = getExprBase(OperExpr); 2481 2482 // Visit all existing chains. Check if its IVOper can be computed as a 2483 // profitable loop invariant increment from the last link in the Chain. 2484 unsigned ChainIdx = 0, NChains = IVChainVec.size(); 2485 const SCEV *LastIncExpr = 0; 2486 for (; ChainIdx < NChains; ++ChainIdx) { 2487 IVChain &Chain = IVChainVec[ChainIdx]; 2488 2489 // Prune the solution space aggressively by checking that both IV operands 2490 // are expressions that operate on the same unscaled SCEVUnknown. This 2491 // "base" will be canceled by the subsequent getMinusSCEV call. Checking 2492 // first avoids creating extra SCEV expressions. 2493 if (!StressIVChain && Chain.ExprBase != OperExprBase) 2494 continue; 2495 2496 Value *PrevIV = getWideOperand(Chain.Incs.back().IVOperand); 2497 if (!isCompatibleIVType(PrevIV, NextIV)) 2498 continue; 2499 2500 // A phi node terminates a chain. 2501 if (isa<PHINode>(UserInst) && isa<PHINode>(Chain.tailUserInst())) 2502 continue; 2503 2504 // The increment must be loop-invariant so it can be kept in a register. 2505 const SCEV *PrevExpr = SE.getSCEV(PrevIV); 2506 const SCEV *IncExpr = SE.getMinusSCEV(OperExpr, PrevExpr); 2507 if (!SE.isLoopInvariant(IncExpr, L)) 2508 continue; 2509 2510 if (Chain.isProfitableIncrement(OperExpr, IncExpr, SE)) { 2511 LastIncExpr = IncExpr; 2512 break; 2513 } 2514 } 2515 // If we haven't found a chain, create a new one, unless we hit the max. Don't 2516 // bother for phi nodes, because they must be last in the chain. 2517 if (ChainIdx == NChains) { 2518 if (isa<PHINode>(UserInst)) 2519 return; 2520 if (NChains >= MaxChains && !StressIVChain) { 2521 DEBUG(dbgs() << "IV Chain Limit\n"); 2522 return; 2523 } 2524 LastIncExpr = OperExpr; 2525 // IVUsers may have skipped over sign/zero extensions. We don't currently 2526 // attempt to form chains involving extensions unless they can be hoisted 2527 // into this loop's AddRec. 2528 if (!isa<SCEVAddRecExpr>(LastIncExpr)) 2529 return; 2530 ++NChains; 2531 IVChainVec.push_back(IVChain(IVInc(UserInst, IVOper, LastIncExpr), 2532 OperExprBase)); 2533 ChainUsersVec.resize(NChains); 2534 DEBUG(dbgs() << "IV Chain#" << ChainIdx << " Head: (" << *UserInst 2535 << ") IV=" << *LastIncExpr << "\n"); 2536 } else { 2537 DEBUG(dbgs() << "IV Chain#" << ChainIdx << " Inc: (" << *UserInst 2538 << ") IV+" << *LastIncExpr << "\n"); 2539 // Add this IV user to the end of the chain. 2540 IVChainVec[ChainIdx].add(IVInc(UserInst, IVOper, LastIncExpr)); 2541 } 2542 2543 SmallPtrSet<Instruction*,4> &NearUsers = ChainUsersVec[ChainIdx].NearUsers; 2544 // This chain's NearUsers become FarUsers. 2545 if (!LastIncExpr->isZero()) { 2546 ChainUsersVec[ChainIdx].FarUsers.insert(NearUsers.begin(), 2547 NearUsers.end()); 2548 NearUsers.clear(); 2549 } 2550 2551 // All other uses of IVOperand become near uses of the chain. 2552 // We currently ignore intermediate values within SCEV expressions, assuming 2553 // they will eventually be used be the current chain, or can be computed 2554 // from one of the chain increments. To be more precise we could 2555 // transitively follow its user and only add leaf IV users to the set. 2556 for (Value::use_iterator UseIter = IVOper->use_begin(), 2557 UseEnd = IVOper->use_end(); UseIter != UseEnd; ++UseIter) { 2558 Instruction *OtherUse = dyn_cast<Instruction>(*UseIter); 2559 if (!OtherUse || OtherUse == UserInst) 2560 continue; 2561 if (SE.isSCEVable(OtherUse->getType()) 2562 && !isa<SCEVUnknown>(SE.getSCEV(OtherUse)) 2563 && IU.isIVUserOrOperand(OtherUse)) { 2564 continue; 2565 } 2566 NearUsers.insert(OtherUse); 2567 } 2568 2569 // Since this user is part of the chain, it's no longer considered a use 2570 // of the chain. 2571 ChainUsersVec[ChainIdx].FarUsers.erase(UserInst); 2572} 2573 2574/// CollectChains - Populate the vector of Chains. 2575/// 2576/// This decreases ILP at the architecture level. Targets with ample registers, 2577/// multiple memory ports, and no register renaming probably don't want 2578/// this. However, such targets should probably disable LSR altogether. 2579/// 2580/// The job of LSR is to make a reasonable choice of induction variables across 2581/// the loop. Subsequent passes can easily "unchain" computation exposing more 2582/// ILP *within the loop* if the target wants it. 2583/// 2584/// Finding the best IV chain is potentially a scheduling problem. Since LSR 2585/// will not reorder memory operations, it will recognize this as a chain, but 2586/// will generate redundant IV increments. Ideally this would be corrected later 2587/// by a smart scheduler: 2588/// = A[i] 2589/// = A[i+x] 2590/// A[i] = 2591/// A[i+x] = 2592/// 2593/// TODO: Walk the entire domtree within this loop, not just the path to the 2594/// loop latch. This will discover chains on side paths, but requires 2595/// maintaining multiple copies of the Chains state. 2596void LSRInstance::CollectChains() { 2597 DEBUG(dbgs() << "Collecting IV Chains.\n"); 2598 SmallVector<ChainUsers, 8> ChainUsersVec; 2599 2600 SmallVector<BasicBlock *,8> LatchPath; 2601 BasicBlock *LoopHeader = L->getHeader(); 2602 for (DomTreeNode *Rung = DT.getNode(L->getLoopLatch()); 2603 Rung->getBlock() != LoopHeader; Rung = Rung->getIDom()) { 2604 LatchPath.push_back(Rung->getBlock()); 2605 } 2606 LatchPath.push_back(LoopHeader); 2607 2608 // Walk the instruction stream from the loop header to the loop latch. 2609 for (SmallVectorImpl<BasicBlock *>::reverse_iterator 2610 BBIter = LatchPath.rbegin(), BBEnd = LatchPath.rend(); 2611 BBIter != BBEnd; ++BBIter) { 2612 for (BasicBlock::iterator I = (*BBIter)->begin(), E = (*BBIter)->end(); 2613 I != E; ++I) { 2614 // Skip instructions that weren't seen by IVUsers analysis. 2615 if (isa<PHINode>(I) || !IU.isIVUserOrOperand(I)) 2616 continue; 2617 2618 // Ignore users that are part of a SCEV expression. This way we only 2619 // consider leaf IV Users. This effectively rediscovers a portion of 2620 // IVUsers analysis but in program order this time. 2621 if (SE.isSCEVable(I->getType()) && !isa<SCEVUnknown>(SE.getSCEV(I))) 2622 continue; 2623 2624 // Remove this instruction from any NearUsers set it may be in. 2625 for (unsigned ChainIdx = 0, NChains = IVChainVec.size(); 2626 ChainIdx < NChains; ++ChainIdx) { 2627 ChainUsersVec[ChainIdx].NearUsers.erase(I); 2628 } 2629 // Search for operands that can be chained. 2630 SmallPtrSet<Instruction*, 4> UniqueOperands; 2631 User::op_iterator IVOpEnd = I->op_end(); 2632 User::op_iterator IVOpIter = findIVOperand(I->op_begin(), IVOpEnd, L, SE); 2633 while (IVOpIter != IVOpEnd) { 2634 Instruction *IVOpInst = cast<Instruction>(*IVOpIter); 2635 if (UniqueOperands.insert(IVOpInst)) 2636 ChainInstruction(I, IVOpInst, ChainUsersVec); 2637 IVOpIter = findIVOperand(llvm::next(IVOpIter), IVOpEnd, L, SE); 2638 } 2639 } // Continue walking down the instructions. 2640 } // Continue walking down the domtree. 2641 // Visit phi backedges to determine if the chain can generate the IV postinc. 2642 for (BasicBlock::iterator I = L->getHeader()->begin(); 2643 PHINode *PN = dyn_cast<PHINode>(I); ++I) { 2644 if (!SE.isSCEVable(PN->getType())) 2645 continue; 2646 2647 Instruction *IncV = 2648 dyn_cast<Instruction>(PN->getIncomingValueForBlock(L->getLoopLatch())); 2649 if (IncV) 2650 ChainInstruction(PN, IncV, ChainUsersVec); 2651 } 2652 // Remove any unprofitable chains. 2653 unsigned ChainIdx = 0; 2654 for (unsigned UsersIdx = 0, NChains = IVChainVec.size(); 2655 UsersIdx < NChains; ++UsersIdx) { 2656 if (!isProfitableChain(IVChainVec[UsersIdx], 2657 ChainUsersVec[UsersIdx].FarUsers, SE, TLI)) 2658 continue; 2659 // Preserve the chain at UsesIdx. 2660 if (ChainIdx != UsersIdx) 2661 IVChainVec[ChainIdx] = IVChainVec[UsersIdx]; 2662 FinalizeChain(IVChainVec[ChainIdx]); 2663 ++ChainIdx; 2664 } 2665 IVChainVec.resize(ChainIdx); 2666} 2667 2668void LSRInstance::FinalizeChain(IVChain &Chain) { 2669 assert(!Chain.Incs.empty() && "empty IV chains are not allowed"); 2670 DEBUG(dbgs() << "Final Chain: " << *Chain.Incs[0].UserInst << "\n"); 2671 2672 for (IVChain::const_iterator I = Chain.begin(), E = Chain.end(); 2673 I != E; ++I) { 2674 DEBUG(dbgs() << " Inc: " << *I->UserInst << "\n"); 2675 User::op_iterator UseI = 2676 std::find(I->UserInst->op_begin(), I->UserInst->op_end(), I->IVOperand); 2677 assert(UseI != I->UserInst->op_end() && "cannot find IV operand"); 2678 IVIncSet.insert(UseI); 2679 } 2680} 2681 2682/// Return true if the IVInc can be folded into an addressing mode. 2683static bool canFoldIVIncExpr(const SCEV *IncExpr, Instruction *UserInst, 2684 Value *Operand, const TargetLowering *TLI) { 2685 const SCEVConstant *IncConst = dyn_cast<SCEVConstant>(IncExpr); 2686 if (!IncConst || !isAddressUse(UserInst, Operand)) 2687 return false; 2688 2689 if (IncConst->getValue()->getValue().getMinSignedBits() > 64) 2690 return false; 2691 2692 int64_t IncOffset = IncConst->getValue()->getSExtValue(); 2693 if (!isAlwaysFoldable(IncOffset, /*BaseGV=*/0, /*HaseBaseReg=*/false, 2694 LSRUse::Address, getAccessType(UserInst), TLI)) 2695 return false; 2696 2697 return true; 2698} 2699 2700/// GenerateIVChains - Generate an add or subtract for each IVInc in a chain to 2701/// materialize the IV user's operand from the previous IV user's operand. 2702void LSRInstance::GenerateIVChain(const IVChain &Chain, SCEVExpander &Rewriter, 2703 SmallVectorImpl<WeakVH> &DeadInsts) { 2704 // Find the new IVOperand for the head of the chain. It may have been replaced 2705 // by LSR. 2706 const IVInc &Head = Chain.Incs[0]; 2707 User::op_iterator IVOpEnd = Head.UserInst->op_end(); 2708 User::op_iterator IVOpIter = findIVOperand(Head.UserInst->op_begin(), 2709 IVOpEnd, L, SE); 2710 Value *IVSrc = 0; 2711 while (IVOpIter != IVOpEnd) { 2712 IVSrc = getWideOperand(*IVOpIter); 2713 2714 // If this operand computes the expression that the chain needs, we may use 2715 // it. (Check this after setting IVSrc which is used below.) 2716 // 2717 // Note that if Head.IncExpr is wider than IVSrc, then this phi is too 2718 // narrow for the chain, so we can no longer use it. We do allow using a 2719 // wider phi, assuming the LSR checked for free truncation. In that case we 2720 // should already have a truncate on this operand such that 2721 // getSCEV(IVSrc) == IncExpr. 2722 if (SE.getSCEV(*IVOpIter) == Head.IncExpr 2723 || SE.getSCEV(IVSrc) == Head.IncExpr) { 2724 break; 2725 } 2726 IVOpIter = findIVOperand(llvm::next(IVOpIter), IVOpEnd, L, SE); 2727 } 2728 if (IVOpIter == IVOpEnd) { 2729 // Gracefully give up on this chain. 2730 DEBUG(dbgs() << "Concealed chain head: " << *Head.UserInst << "\n"); 2731 return; 2732 } 2733 2734 DEBUG(dbgs() << "Generate chain at: " << *IVSrc << "\n"); 2735 Type *IVTy = IVSrc->getType(); 2736 Type *IntTy = SE.getEffectiveSCEVType(IVTy); 2737 const SCEV *LeftOverExpr = 0; 2738 for (IVChain::const_iterator IncI = Chain.begin(), 2739 IncE = Chain.end(); IncI != IncE; ++IncI) { 2740 2741 Instruction *InsertPt = IncI->UserInst; 2742 if (isa<PHINode>(InsertPt)) 2743 InsertPt = L->getLoopLatch()->getTerminator(); 2744 2745 // IVOper will replace the current IV User's operand. IVSrc is the IV 2746 // value currently held in a register. 2747 Value *IVOper = IVSrc; 2748 if (!IncI->IncExpr->isZero()) { 2749 // IncExpr was the result of subtraction of two narrow values, so must 2750 // be signed. 2751 const SCEV *IncExpr = SE.getNoopOrSignExtend(IncI->IncExpr, IntTy); 2752 LeftOverExpr = LeftOverExpr ? 2753 SE.getAddExpr(LeftOverExpr, IncExpr) : IncExpr; 2754 } 2755 if (LeftOverExpr && !LeftOverExpr->isZero()) { 2756 // Expand the IV increment. 2757 Rewriter.clearPostInc(); 2758 Value *IncV = Rewriter.expandCodeFor(LeftOverExpr, IntTy, InsertPt); 2759 const SCEV *IVOperExpr = SE.getAddExpr(SE.getUnknown(IVSrc), 2760 SE.getUnknown(IncV)); 2761 IVOper = Rewriter.expandCodeFor(IVOperExpr, IVTy, InsertPt); 2762 2763 // If an IV increment can't be folded, use it as the next IV value. 2764 if (!canFoldIVIncExpr(LeftOverExpr, IncI->UserInst, IncI->IVOperand, 2765 TLI)) { 2766 assert(IVTy == IVOper->getType() && "inconsistent IV increment type"); 2767 IVSrc = IVOper; 2768 LeftOverExpr = 0; 2769 } 2770 } 2771 Type *OperTy = IncI->IVOperand->getType(); 2772 if (IVTy != OperTy) { 2773 assert(SE.getTypeSizeInBits(IVTy) >= SE.getTypeSizeInBits(OperTy) && 2774 "cannot extend a chained IV"); 2775 IRBuilder<> Builder(InsertPt); 2776 IVOper = Builder.CreateTruncOrBitCast(IVOper, OperTy, "lsr.chain"); 2777 } 2778 IncI->UserInst->replaceUsesOfWith(IncI->IVOperand, IVOper); 2779 DeadInsts.push_back(IncI->IVOperand); 2780 } 2781 // If LSR created a new, wider phi, we may also replace its postinc. We only 2782 // do this if we also found a wide value for the head of the chain. 2783 if (isa<PHINode>(Chain.tailUserInst())) { 2784 for (BasicBlock::iterator I = L->getHeader()->begin(); 2785 PHINode *Phi = dyn_cast<PHINode>(I); ++I) { 2786 if (!isCompatibleIVType(Phi, IVSrc)) 2787 continue; 2788 Instruction *PostIncV = dyn_cast<Instruction>( 2789 Phi->getIncomingValueForBlock(L->getLoopLatch())); 2790 if (!PostIncV || (SE.getSCEV(PostIncV) != SE.getSCEV(IVSrc))) 2791 continue; 2792 Value *IVOper = IVSrc; 2793 Type *PostIncTy = PostIncV->getType(); 2794 if (IVTy != PostIncTy) { 2795 assert(PostIncTy->isPointerTy() && "mixing int/ptr IV types"); 2796 IRBuilder<> Builder(L->getLoopLatch()->getTerminator()); 2797 Builder.SetCurrentDebugLocation(PostIncV->getDebugLoc()); 2798 IVOper = Builder.CreatePointerCast(IVSrc, PostIncTy, "lsr.chain"); 2799 } 2800 Phi->replaceUsesOfWith(PostIncV, IVOper); 2801 DeadInsts.push_back(PostIncV); 2802 } 2803 } 2804} 2805 2806void LSRInstance::CollectFixupsAndInitialFormulae() { 2807 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI) { 2808 Instruction *UserInst = UI->getUser(); 2809 // Skip IV users that are part of profitable IV Chains. 2810 User::op_iterator UseI = std::find(UserInst->op_begin(), UserInst->op_end(), 2811 UI->getOperandValToReplace()); 2812 assert(UseI != UserInst->op_end() && "cannot find IV operand"); 2813 if (IVIncSet.count(UseI)) 2814 continue; 2815 2816 // Record the uses. 2817 LSRFixup &LF = getNewFixup(); 2818 LF.UserInst = UserInst; 2819 LF.OperandValToReplace = UI->getOperandValToReplace(); 2820 LF.PostIncLoops = UI->getPostIncLoops(); 2821 2822 LSRUse::KindType Kind = LSRUse::Basic; 2823 Type *AccessTy = 0; 2824 if (isAddressUse(LF.UserInst, LF.OperandValToReplace)) { 2825 Kind = LSRUse::Address; 2826 AccessTy = getAccessType(LF.UserInst); 2827 } 2828 2829 const SCEV *S = IU.getExpr(*UI); 2830 2831 // Equality (== and !=) ICmps are special. We can rewrite (i == N) as 2832 // (N - i == 0), and this allows (N - i) to be the expression that we work 2833 // with rather than just N or i, so we can consider the register 2834 // requirements for both N and i at the same time. Limiting this code to 2835 // equality icmps is not a problem because all interesting loops use 2836 // equality icmps, thanks to IndVarSimplify. 2837 if (ICmpInst *CI = dyn_cast<ICmpInst>(LF.UserInst)) 2838 if (CI->isEquality()) { 2839 // Swap the operands if needed to put the OperandValToReplace on the 2840 // left, for consistency. 2841 Value *NV = CI->getOperand(1); 2842 if (NV == LF.OperandValToReplace) { 2843 CI->setOperand(1, CI->getOperand(0)); 2844 CI->setOperand(0, NV); 2845 NV = CI->getOperand(1); 2846 Changed = true; 2847 } 2848 2849 // x == y --> x - y == 0 2850 const SCEV *N = SE.getSCEV(NV); 2851 if (SE.isLoopInvariant(N, L) && isSafeToExpand(N)) { 2852 // S is normalized, so normalize N before folding it into S 2853 // to keep the result normalized. 2854 N = TransformForPostIncUse(Normalize, N, CI, 0, 2855 LF.PostIncLoops, SE, DT); 2856 Kind = LSRUse::ICmpZero; 2857 S = SE.getMinusSCEV(N, S); 2858 } 2859 2860 // -1 and the negations of all interesting strides (except the negation 2861 // of -1) are now also interesting. 2862 for (size_t i = 0, e = Factors.size(); i != e; ++i) 2863 if (Factors[i] != -1) 2864 Factors.insert(-(uint64_t)Factors[i]); 2865 Factors.insert(-1); 2866 } 2867 2868 // Set up the initial formula for this use. 2869 std::pair<size_t, int64_t> P = getUse(S, Kind, AccessTy); 2870 LF.LUIdx = P.first; 2871 LF.Offset = P.second; 2872 LSRUse &LU = Uses[LF.LUIdx]; 2873 LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L); 2874 if (!LU.WidestFixupType || 2875 SE.getTypeSizeInBits(LU.WidestFixupType) < 2876 SE.getTypeSizeInBits(LF.OperandValToReplace->getType())) 2877 LU.WidestFixupType = LF.OperandValToReplace->getType(); 2878 2879 // If this is the first use of this LSRUse, give it a formula. 2880 if (LU.Formulae.empty()) { 2881 InsertInitialFormula(S, LU, LF.LUIdx); 2882 CountRegisters(LU.Formulae.back(), LF.LUIdx); 2883 } 2884 } 2885 2886 DEBUG(print_fixups(dbgs())); 2887} 2888 2889/// InsertInitialFormula - Insert a formula for the given expression into 2890/// the given use, separating out loop-variant portions from loop-invariant 2891/// and loop-computable portions. 2892void 2893LSRInstance::InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx) { 2894 Formula F; 2895 F.InitialMatch(S, L, SE); 2896 bool Inserted = InsertFormula(LU, LUIdx, F); 2897 assert(Inserted && "Initial formula already exists!"); (void)Inserted; 2898} 2899 2900/// InsertSupplementalFormula - Insert a simple single-register formula for 2901/// the given expression into the given use. 2902void 2903LSRInstance::InsertSupplementalFormula(const SCEV *S, 2904 LSRUse &LU, size_t LUIdx) { 2905 Formula F; 2906 F.BaseRegs.push_back(S); 2907 F.AM.HasBaseReg = true; 2908 bool Inserted = InsertFormula(LU, LUIdx, F); 2909 assert(Inserted && "Supplemental formula already exists!"); (void)Inserted; 2910} 2911 2912/// CountRegisters - Note which registers are used by the given formula, 2913/// updating RegUses. 2914void LSRInstance::CountRegisters(const Formula &F, size_t LUIdx) { 2915 if (F.ScaledReg) 2916 RegUses.CountRegister(F.ScaledReg, LUIdx); 2917 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(), 2918 E = F.BaseRegs.end(); I != E; ++I) 2919 RegUses.CountRegister(*I, LUIdx); 2920} 2921 2922/// InsertFormula - If the given formula has not yet been inserted, add it to 2923/// the list, and return true. Return false otherwise. 2924bool LSRInstance::InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F) { 2925 if (!LU.InsertFormula(F)) 2926 return false; 2927 2928 CountRegisters(F, LUIdx); 2929 return true; 2930} 2931 2932/// CollectLoopInvariantFixupsAndFormulae - Check for other uses of 2933/// loop-invariant values which we're tracking. These other uses will pin these 2934/// values in registers, making them less profitable for elimination. 2935/// TODO: This currently misses non-constant addrec step registers. 2936/// TODO: Should this give more weight to users inside the loop? 2937void 2938LSRInstance::CollectLoopInvariantFixupsAndFormulae() { 2939 SmallVector<const SCEV *, 8> Worklist(RegUses.begin(), RegUses.end()); 2940 SmallPtrSet<const SCEV *, 8> Inserted; 2941 2942 while (!Worklist.empty()) { 2943 const SCEV *S = Worklist.pop_back_val(); 2944 2945 if (const SCEVNAryExpr *N = dyn_cast<SCEVNAryExpr>(S)) 2946 Worklist.append(N->op_begin(), N->op_end()); 2947 else if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(S)) 2948 Worklist.push_back(C->getOperand()); 2949 else if (const SCEVUDivExpr *D = dyn_cast<SCEVUDivExpr>(S)) { 2950 Worklist.push_back(D->getLHS()); 2951 Worklist.push_back(D->getRHS()); 2952 } else if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) { 2953 if (!Inserted.insert(U)) continue; 2954 const Value *V = U->getValue(); 2955 if (const Instruction *Inst = dyn_cast<Instruction>(V)) { 2956 // Look for instructions defined outside the loop. 2957 if (L->contains(Inst)) continue; 2958 } else if (isa<UndefValue>(V)) 2959 // Undef doesn't have a live range, so it doesn't matter. 2960 continue; 2961 for (Value::const_use_iterator UI = V->use_begin(), UE = V->use_end(); 2962 UI != UE; ++UI) { 2963 const Instruction *UserInst = dyn_cast<Instruction>(*UI); 2964 // Ignore non-instructions. 2965 if (!UserInst) 2966 continue; 2967 // Ignore instructions in other functions (as can happen with 2968 // Constants). 2969 if (UserInst->getParent()->getParent() != L->getHeader()->getParent()) 2970 continue; 2971 // Ignore instructions not dominated by the loop. 2972 const BasicBlock *UseBB = !isa<PHINode>(UserInst) ? 2973 UserInst->getParent() : 2974 cast<PHINode>(UserInst)->getIncomingBlock( 2975 PHINode::getIncomingValueNumForOperand(UI.getOperandNo())); 2976 if (!DT.dominates(L->getHeader(), UseBB)) 2977 continue; 2978 // Ignore uses which are part of other SCEV expressions, to avoid 2979 // analyzing them multiple times. 2980 if (SE.isSCEVable(UserInst->getType())) { 2981 const SCEV *UserS = SE.getSCEV(const_cast<Instruction *>(UserInst)); 2982 // If the user is a no-op, look through to its uses. 2983 if (!isa<SCEVUnknown>(UserS)) 2984 continue; 2985 if (UserS == U) { 2986 Worklist.push_back( 2987 SE.getUnknown(const_cast<Instruction *>(UserInst))); 2988 continue; 2989 } 2990 } 2991 // Ignore icmp instructions which are already being analyzed. 2992 if (const ICmpInst *ICI = dyn_cast<ICmpInst>(UserInst)) { 2993 unsigned OtherIdx = !UI.getOperandNo(); 2994 Value *OtherOp = const_cast<Value *>(ICI->getOperand(OtherIdx)); 2995 if (SE.hasComputableLoopEvolution(SE.getSCEV(OtherOp), L)) 2996 continue; 2997 } 2998 2999 LSRFixup &LF = getNewFixup(); 3000 LF.UserInst = const_cast<Instruction *>(UserInst); 3001 LF.OperandValToReplace = UI.getUse(); 3002 std::pair<size_t, int64_t> P = getUse(S, LSRUse::Basic, 0); 3003 LF.LUIdx = P.first; 3004 LF.Offset = P.second; 3005 LSRUse &LU = Uses[LF.LUIdx]; 3006 LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L); 3007 if (!LU.WidestFixupType || 3008 SE.getTypeSizeInBits(LU.WidestFixupType) < 3009 SE.getTypeSizeInBits(LF.OperandValToReplace->getType())) 3010 LU.WidestFixupType = LF.OperandValToReplace->getType(); 3011 InsertSupplementalFormula(U, LU, LF.LUIdx); 3012 CountRegisters(LU.Formulae.back(), Uses.size() - 1); 3013 break; 3014 } 3015 } 3016 } 3017} 3018 3019/// CollectSubexprs - Split S into subexpressions which can be pulled out into 3020/// separate registers. If C is non-null, multiply each subexpression by C. 3021/// 3022/// Return remainder expression after factoring the subexpressions captured by 3023/// Ops. If Ops is complete, return NULL. 3024static const SCEV *CollectSubexprs(const SCEV *S, const SCEVConstant *C, 3025 SmallVectorImpl<const SCEV *> &Ops, 3026 const Loop *L, 3027 ScalarEvolution &SE, 3028 unsigned Depth = 0) { 3029 // Arbitrarily cap recursion to protect compile time. 3030 if (Depth >= 3) 3031 return S; 3032 3033 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) { 3034 // Break out add operands. 3035 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end(); 3036 I != E; ++I) { 3037 const SCEV *Remainder = CollectSubexprs(*I, C, Ops, L, SE, Depth+1); 3038 if (Remainder) 3039 Ops.push_back(C ? SE.getMulExpr(C, Remainder) : Remainder); 3040 } 3041 return NULL; 3042 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) { 3043 // Split a non-zero base out of an addrec. 3044 if (AR->getStart()->isZero()) 3045 return S; 3046 3047 const SCEV *Remainder = CollectSubexprs(AR->getStart(), 3048 C, Ops, L, SE, Depth+1); 3049 // Split the non-zero AddRec unless it is part of a nested recurrence that 3050 // does not pertain to this loop. 3051 if (Remainder && (AR->getLoop() == L || !isa<SCEVAddRecExpr>(Remainder))) { 3052 Ops.push_back(C ? SE.getMulExpr(C, Remainder) : Remainder); 3053 Remainder = NULL; 3054 } 3055 if (Remainder != AR->getStart()) { 3056 if (!Remainder) 3057 Remainder = SE.getConstant(AR->getType(), 0); 3058 return SE.getAddRecExpr(Remainder, 3059 AR->getStepRecurrence(SE), 3060 AR->getLoop(), 3061 //FIXME: AR->getNoWrapFlags(SCEV::FlagNW) 3062 SCEV::FlagAnyWrap); 3063 } 3064 } else if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) { 3065 // Break (C * (a + b + c)) into C*a + C*b + C*c. 3066 if (Mul->getNumOperands() != 2) 3067 return S; 3068 if (const SCEVConstant *Op0 = 3069 dyn_cast<SCEVConstant>(Mul->getOperand(0))) { 3070 C = C ? cast<SCEVConstant>(SE.getMulExpr(C, Op0)) : Op0; 3071 const SCEV *Remainder = 3072 CollectSubexprs(Mul->getOperand(1), C, Ops, L, SE, Depth+1); 3073 if (Remainder) 3074 Ops.push_back(SE.getMulExpr(C, Remainder)); 3075 return NULL; 3076 } 3077 } 3078 return S; 3079} 3080 3081/// GenerateReassociations - Split out subexpressions from adds and the bases of 3082/// addrecs. 3083void LSRInstance::GenerateReassociations(LSRUse &LU, unsigned LUIdx, 3084 Formula Base, 3085 unsigned Depth) { 3086 // Arbitrarily cap recursion to protect compile time. 3087 if (Depth >= 3) return; 3088 3089 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) { 3090 const SCEV *BaseReg = Base.BaseRegs[i]; 3091 3092 SmallVector<const SCEV *, 8> AddOps; 3093 const SCEV *Remainder = CollectSubexprs(BaseReg, 0, AddOps, L, SE); 3094 if (Remainder) 3095 AddOps.push_back(Remainder); 3096 3097 if (AddOps.size() == 1) continue; 3098 3099 for (SmallVectorImpl<const SCEV *>::const_iterator J = AddOps.begin(), 3100 JE = AddOps.end(); J != JE; ++J) { 3101 3102 // Loop-variant "unknown" values are uninteresting; we won't be able to 3103 // do anything meaningful with them. 3104 if (isa<SCEVUnknown>(*J) && !SE.isLoopInvariant(*J, L)) 3105 continue; 3106 3107 // Don't pull a constant into a register if the constant could be folded 3108 // into an immediate field. 3109 if (isAlwaysFoldable(*J, LU.MinOffset, LU.MaxOffset, 3110 Base.getNumRegs() > 1, 3111 LU.Kind, LU.AccessTy, TLI, SE)) 3112 continue; 3113 3114 // Collect all operands except *J. 3115 SmallVector<const SCEV *, 8> InnerAddOps 3116 (((const SmallVector<const SCEV *, 8> &)AddOps).begin(), J); 3117 InnerAddOps.append 3118 (llvm::next(J), ((const SmallVector<const SCEV *, 8> &)AddOps).end()); 3119 3120 // Don't leave just a constant behind in a register if the constant could 3121 // be folded into an immediate field. 3122 if (InnerAddOps.size() == 1 && 3123 isAlwaysFoldable(InnerAddOps[0], LU.MinOffset, LU.MaxOffset, 3124 Base.getNumRegs() > 1, 3125 LU.Kind, LU.AccessTy, TLI, SE)) 3126 continue; 3127 3128 const SCEV *InnerSum = SE.getAddExpr(InnerAddOps); 3129 if (InnerSum->isZero()) 3130 continue; 3131 Formula F = Base; 3132 3133 // Add the remaining pieces of the add back into the new formula. 3134 const SCEVConstant *InnerSumSC = dyn_cast<SCEVConstant>(InnerSum); 3135 if (TLI && InnerSumSC && 3136 SE.getTypeSizeInBits(InnerSumSC->getType()) <= 64 && 3137 TLI->isLegalAddImmediate((uint64_t)F.UnfoldedOffset + 3138 InnerSumSC->getValue()->getZExtValue())) { 3139 F.UnfoldedOffset = (uint64_t)F.UnfoldedOffset + 3140 InnerSumSC->getValue()->getZExtValue(); 3141 F.BaseRegs.erase(F.BaseRegs.begin() + i); 3142 } else 3143 F.BaseRegs[i] = InnerSum; 3144 3145 // Add J as its own register, or an unfolded immediate. 3146 const SCEVConstant *SC = dyn_cast<SCEVConstant>(*J); 3147 if (TLI && SC && SE.getTypeSizeInBits(SC->getType()) <= 64 && 3148 TLI->isLegalAddImmediate((uint64_t)F.UnfoldedOffset + 3149 SC->getValue()->getZExtValue())) 3150 F.UnfoldedOffset = (uint64_t)F.UnfoldedOffset + 3151 SC->getValue()->getZExtValue(); 3152 else 3153 F.BaseRegs.push_back(*J); 3154 3155 if (InsertFormula(LU, LUIdx, F)) 3156 // If that formula hadn't been seen before, recurse to find more like 3157 // it. 3158 GenerateReassociations(LU, LUIdx, LU.Formulae.back(), Depth+1); 3159 } 3160 } 3161} 3162 3163/// GenerateCombinations - Generate a formula consisting of all of the 3164/// loop-dominating registers added into a single register. 3165void LSRInstance::GenerateCombinations(LSRUse &LU, unsigned LUIdx, 3166 Formula Base) { 3167 // This method is only interesting on a plurality of registers. 3168 if (Base.BaseRegs.size() <= 1) return; 3169 3170 Formula F = Base; 3171 F.BaseRegs.clear(); 3172 SmallVector<const SCEV *, 4> Ops; 3173 for (SmallVectorImpl<const SCEV *>::const_iterator 3174 I = Base.BaseRegs.begin(), E = Base.BaseRegs.end(); I != E; ++I) { 3175 const SCEV *BaseReg = *I; 3176 if (SE.properlyDominates(BaseReg, L->getHeader()) && 3177 !SE.hasComputableLoopEvolution(BaseReg, L)) 3178 Ops.push_back(BaseReg); 3179 else 3180 F.BaseRegs.push_back(BaseReg); 3181 } 3182 if (Ops.size() > 1) { 3183 const SCEV *Sum = SE.getAddExpr(Ops); 3184 // TODO: If Sum is zero, it probably means ScalarEvolution missed an 3185 // opportunity to fold something. For now, just ignore such cases 3186 // rather than proceed with zero in a register. 3187 if (!Sum->isZero()) { 3188 F.BaseRegs.push_back(Sum); 3189 (void)InsertFormula(LU, LUIdx, F); 3190 } 3191 } 3192} 3193 3194/// GenerateSymbolicOffsets - Generate reuse formulae using symbolic offsets. 3195void LSRInstance::GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx, 3196 Formula Base) { 3197 // We can't add a symbolic offset if the address already contains one. 3198 if (Base.AM.BaseGV) return; 3199 3200 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) { 3201 const SCEV *G = Base.BaseRegs[i]; 3202 GlobalValue *GV = ExtractSymbol(G, SE); 3203 if (G->isZero() || !GV) 3204 continue; 3205 Formula F = Base; 3206 F.AM.BaseGV = GV; 3207 if (!isLegalUse(F.AM, LU.MinOffset, LU.MaxOffset, 3208 LU.Kind, LU.AccessTy, TLI)) 3209 continue; 3210 F.BaseRegs[i] = G; 3211 (void)InsertFormula(LU, LUIdx, F); 3212 } 3213} 3214 3215/// GenerateConstantOffsets - Generate reuse formulae using symbolic offsets. 3216void LSRInstance::GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx, 3217 Formula Base) { 3218 // TODO: For now, just add the min and max offset, because it usually isn't 3219 // worthwhile looking at everything inbetween. 3220 SmallVector<int64_t, 2> Worklist; 3221 Worklist.push_back(LU.MinOffset); 3222 if (LU.MaxOffset != LU.MinOffset) 3223 Worklist.push_back(LU.MaxOffset); 3224 3225 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) { 3226 const SCEV *G = Base.BaseRegs[i]; 3227 3228 for (SmallVectorImpl<int64_t>::const_iterator I = Worklist.begin(), 3229 E = Worklist.end(); I != E; ++I) { 3230 Formula F = Base; 3231 F.AM.BaseOffs = (uint64_t)Base.AM.BaseOffs - *I; 3232 if (isLegalUse(F.AM, LU.MinOffset - *I, LU.MaxOffset - *I, 3233 LU.Kind, LU.AccessTy, TLI)) { 3234 // Add the offset to the base register. 3235 const SCEV *NewG = SE.getAddExpr(SE.getConstant(G->getType(), *I), G); 3236 // If it cancelled out, drop the base register, otherwise update it. 3237 if (NewG->isZero()) { 3238 std::swap(F.BaseRegs[i], F.BaseRegs.back()); 3239 F.BaseRegs.pop_back(); 3240 } else 3241 F.BaseRegs[i] = NewG; 3242 3243 (void)InsertFormula(LU, LUIdx, F); 3244 } 3245 } 3246 3247 int64_t Imm = ExtractImmediate(G, SE); 3248 if (G->isZero() || Imm == 0) 3249 continue; 3250 Formula F = Base; 3251 F.AM.BaseOffs = (uint64_t)F.AM.BaseOffs + Imm; 3252 if (!isLegalUse(F.AM, LU.MinOffset, LU.MaxOffset, 3253 LU.Kind, LU.AccessTy, TLI)) 3254 continue; 3255 F.BaseRegs[i] = G; 3256 (void)InsertFormula(LU, LUIdx, F); 3257 } 3258} 3259 3260/// GenerateICmpZeroScales - For ICmpZero, check to see if we can scale up 3261/// the comparison. For example, x == y -> x*c == y*c. 3262void LSRInstance::GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx, 3263 Formula Base) { 3264 if (LU.Kind != LSRUse::ICmpZero) return; 3265 3266 // Determine the integer type for the base formula. 3267 Type *IntTy = Base.getType(); 3268 if (!IntTy) return; 3269 if (SE.getTypeSizeInBits(IntTy) > 64) return; 3270 3271 // Don't do this if there is more than one offset. 3272 if (LU.MinOffset != LU.MaxOffset) return; 3273 3274 assert(!Base.AM.BaseGV && "ICmpZero use is not legal!"); 3275 3276 // Check each interesting stride. 3277 for (SmallSetVector<int64_t, 8>::const_iterator 3278 I = Factors.begin(), E = Factors.end(); I != E; ++I) { 3279 int64_t Factor = *I; 3280 3281 // Check that the multiplication doesn't overflow. 3282 if (Base.AM.BaseOffs == INT64_MIN && Factor == -1) 3283 continue; 3284 int64_t NewBaseOffs = (uint64_t)Base.AM.BaseOffs * Factor; 3285 if (NewBaseOffs / Factor != Base.AM.BaseOffs) 3286 continue; 3287 3288 // Check that multiplying with the use offset doesn't overflow. 3289 int64_t Offset = LU.MinOffset; 3290 if (Offset == INT64_MIN && Factor == -1) 3291 continue; 3292 Offset = (uint64_t)Offset * Factor; 3293 if (Offset / Factor != LU.MinOffset) 3294 continue; 3295 3296 Formula F = Base; 3297 F.AM.BaseOffs = NewBaseOffs; 3298 3299 // Check that this scale is legal. 3300 if (!isLegalUse(F.AM, Offset, Offset, LU.Kind, LU.AccessTy, TLI)) 3301 continue; 3302 3303 // Compensate for the use having MinOffset built into it. 3304 F.AM.BaseOffs = (uint64_t)F.AM.BaseOffs + Offset - LU.MinOffset; 3305 3306 const SCEV *FactorS = SE.getConstant(IntTy, Factor); 3307 3308 // Check that multiplying with each base register doesn't overflow. 3309 for (size_t i = 0, e = F.BaseRegs.size(); i != e; ++i) { 3310 F.BaseRegs[i] = SE.getMulExpr(F.BaseRegs[i], FactorS); 3311 if (getExactSDiv(F.BaseRegs[i], FactorS, SE) != Base.BaseRegs[i]) 3312 goto next; 3313 } 3314 3315 // Check that multiplying with the scaled register doesn't overflow. 3316 if (F.ScaledReg) { 3317 F.ScaledReg = SE.getMulExpr(F.ScaledReg, FactorS); 3318 if (getExactSDiv(F.ScaledReg, FactorS, SE) != Base.ScaledReg) 3319 continue; 3320 } 3321 3322 // Check that multiplying with the unfolded offset doesn't overflow. 3323 if (F.UnfoldedOffset != 0) { 3324 if (F.UnfoldedOffset == INT64_MIN && Factor == -1) 3325 continue; 3326 F.UnfoldedOffset = (uint64_t)F.UnfoldedOffset * Factor; 3327 if (F.UnfoldedOffset / Factor != Base.UnfoldedOffset) 3328 continue; 3329 } 3330 3331 // If we make it here and it's legal, add it. 3332 (void)InsertFormula(LU, LUIdx, F); 3333 next:; 3334 } 3335} 3336 3337/// GenerateScales - Generate stride factor reuse formulae by making use of 3338/// scaled-offset address modes, for example. 3339void LSRInstance::GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base) { 3340 // Determine the integer type for the base formula. 3341 Type *IntTy = Base.getType(); 3342 if (!IntTy) return; 3343 3344 // If this Formula already has a scaled register, we can't add another one. 3345 if (Base.AM.Scale != 0) return; 3346 3347 // Check each interesting stride. 3348 for (SmallSetVector<int64_t, 8>::const_iterator 3349 I = Factors.begin(), E = Factors.end(); I != E; ++I) { 3350 int64_t Factor = *I; 3351 3352 Base.AM.Scale = Factor; 3353 Base.AM.HasBaseReg = Base.BaseRegs.size() > 1; 3354 // Check whether this scale is going to be legal. 3355 if (!isLegalUse(Base.AM, LU.MinOffset, LU.MaxOffset, 3356 LU.Kind, LU.AccessTy, TLI)) { 3357 // As a special-case, handle special out-of-loop Basic users specially. 3358 // TODO: Reconsider this special case. 3359 if (LU.Kind == LSRUse::Basic && 3360 isLegalUse(Base.AM, LU.MinOffset, LU.MaxOffset, 3361 LSRUse::Special, LU.AccessTy, TLI) && 3362 LU.AllFixupsOutsideLoop) 3363 LU.Kind = LSRUse::Special; 3364 else 3365 continue; 3366 } 3367 // For an ICmpZero, negating a solitary base register won't lead to 3368 // new solutions. 3369 if (LU.Kind == LSRUse::ICmpZero && 3370 !Base.AM.HasBaseReg && Base.AM.BaseOffs == 0 && !Base.AM.BaseGV) 3371 continue; 3372 // For each addrec base reg, apply the scale, if possible. 3373 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) 3374 if (const SCEVAddRecExpr *AR = 3375 dyn_cast<SCEVAddRecExpr>(Base.BaseRegs[i])) { 3376 const SCEV *FactorS = SE.getConstant(IntTy, Factor); 3377 if (FactorS->isZero()) 3378 continue; 3379 // Divide out the factor, ignoring high bits, since we'll be 3380 // scaling the value back up in the end. 3381 if (const SCEV *Quotient = getExactSDiv(AR, FactorS, SE, true)) { 3382 // TODO: This could be optimized to avoid all the copying. 3383 Formula F = Base; 3384 F.ScaledReg = Quotient; 3385 F.DeleteBaseReg(F.BaseRegs[i]); 3386 (void)InsertFormula(LU, LUIdx, F); 3387 } 3388 } 3389 } 3390} 3391 3392/// GenerateTruncates - Generate reuse formulae from different IV types. 3393void LSRInstance::GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base) { 3394 // This requires TargetLowering to tell us which truncates are free. 3395 if (!TLI) return; 3396 3397 // Don't bother truncating symbolic values. 3398 if (Base.AM.BaseGV) return; 3399 3400 // Determine the integer type for the base formula. 3401 Type *DstTy = Base.getType(); 3402 if (!DstTy) return; 3403 DstTy = SE.getEffectiveSCEVType(DstTy); 3404 3405 for (SmallSetVector<Type *, 4>::const_iterator 3406 I = Types.begin(), E = Types.end(); I != E; ++I) { 3407 Type *SrcTy = *I; 3408 if (SrcTy != DstTy && TLI->isTruncateFree(SrcTy, DstTy)) { 3409 Formula F = Base; 3410 3411 if (F.ScaledReg) F.ScaledReg = SE.getAnyExtendExpr(F.ScaledReg, *I); 3412 for (SmallVectorImpl<const SCEV *>::iterator J = F.BaseRegs.begin(), 3413 JE = F.BaseRegs.end(); J != JE; ++J) 3414 *J = SE.getAnyExtendExpr(*J, SrcTy); 3415 3416 // TODO: This assumes we've done basic processing on all uses and 3417 // have an idea what the register usage is. 3418 if (!F.hasRegsUsedByUsesOtherThan(LUIdx, RegUses)) 3419 continue; 3420 3421 (void)InsertFormula(LU, LUIdx, F); 3422 } 3423 } 3424} 3425 3426namespace { 3427 3428/// WorkItem - Helper class for GenerateCrossUseConstantOffsets. It's used to 3429/// defer modifications so that the search phase doesn't have to worry about 3430/// the data structures moving underneath it. 3431struct WorkItem { 3432 size_t LUIdx; 3433 int64_t Imm; 3434 const SCEV *OrigReg; 3435 3436 WorkItem(size_t LI, int64_t I, const SCEV *R) 3437 : LUIdx(LI), Imm(I), OrigReg(R) {} 3438 3439 void print(raw_ostream &OS) const; 3440 void dump() const; 3441}; 3442 3443} 3444 3445void WorkItem::print(raw_ostream &OS) const { 3446 OS << "in formulae referencing " << *OrigReg << " in use " << LUIdx 3447 << " , add offset " << Imm; 3448} 3449 3450#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) 3451void WorkItem::dump() const { 3452 print(errs()); errs() << '\n'; 3453} 3454#endif 3455 3456/// GenerateCrossUseConstantOffsets - Look for registers which are a constant 3457/// distance apart and try to form reuse opportunities between them. 3458void LSRInstance::GenerateCrossUseConstantOffsets() { 3459 // Group the registers by their value without any added constant offset. 3460 typedef std::map<int64_t, const SCEV *> ImmMapTy; 3461 typedef DenseMap<const SCEV *, ImmMapTy> RegMapTy; 3462 RegMapTy Map; 3463 DenseMap<const SCEV *, SmallBitVector> UsedByIndicesMap; 3464 SmallVector<const SCEV *, 8> Sequence; 3465 for (RegUseTracker::const_iterator I = RegUses.begin(), E = RegUses.end(); 3466 I != E; ++I) { 3467 const SCEV *Reg = *I; 3468 int64_t Imm = ExtractImmediate(Reg, SE); 3469 std::pair<RegMapTy::iterator, bool> Pair = 3470 Map.insert(std::make_pair(Reg, ImmMapTy())); 3471 if (Pair.second) 3472 Sequence.push_back(Reg); 3473 Pair.first->second.insert(std::make_pair(Imm, *I)); 3474 UsedByIndicesMap[Reg] |= RegUses.getUsedByIndices(*I); 3475 } 3476 3477 // Now examine each set of registers with the same base value. Build up 3478 // a list of work to do and do the work in a separate step so that we're 3479 // not adding formulae and register counts while we're searching. 3480 SmallVector<WorkItem, 32> WorkItems; 3481 SmallSet<std::pair<size_t, int64_t>, 32> UniqueItems; 3482 for (SmallVectorImpl<const SCEV *>::const_iterator I = Sequence.begin(), 3483 E = Sequence.end(); I != E; ++I) { 3484 const SCEV *Reg = *I; 3485 const ImmMapTy &Imms = Map.find(Reg)->second; 3486 3487 // It's not worthwhile looking for reuse if there's only one offset. 3488 if (Imms.size() == 1) 3489 continue; 3490 3491 DEBUG(dbgs() << "Generating cross-use offsets for " << *Reg << ':'; 3492 for (ImmMapTy::const_iterator J = Imms.begin(), JE = Imms.end(); 3493 J != JE; ++J) 3494 dbgs() << ' ' << J->first; 3495 dbgs() << '\n'); 3496 3497 // Examine each offset. 3498 for (ImmMapTy::const_iterator J = Imms.begin(), JE = Imms.end(); 3499 J != JE; ++J) { 3500 const SCEV *OrigReg = J->second; 3501 3502 int64_t JImm = J->first; 3503 const SmallBitVector &UsedByIndices = RegUses.getUsedByIndices(OrigReg); 3504 3505 if (!isa<SCEVConstant>(OrigReg) && 3506 UsedByIndicesMap[Reg].count() == 1) { 3507 DEBUG(dbgs() << "Skipping cross-use reuse for " << *OrigReg << '\n'); 3508 continue; 3509 } 3510 3511 // Conservatively examine offsets between this orig reg a few selected 3512 // other orig regs. 3513 ImmMapTy::const_iterator OtherImms[] = { 3514 Imms.begin(), prior(Imms.end()), 3515 Imms.lower_bound((Imms.begin()->first + prior(Imms.end())->first) / 2) 3516 }; 3517 for (size_t i = 0, e = array_lengthof(OtherImms); i != e; ++i) { 3518 ImmMapTy::const_iterator M = OtherImms[i]; 3519 if (M == J || M == JE) continue; 3520 3521 // Compute the difference between the two. 3522 int64_t Imm = (uint64_t)JImm - M->first; 3523 for (int LUIdx = UsedByIndices.find_first(); LUIdx != -1; 3524 LUIdx = UsedByIndices.find_next(LUIdx)) 3525 // Make a memo of this use, offset, and register tuple. 3526 if (UniqueItems.insert(std::make_pair(LUIdx, Imm))) 3527 WorkItems.push_back(WorkItem(LUIdx, Imm, OrigReg)); 3528 } 3529 } 3530 } 3531 3532 Map.clear(); 3533 Sequence.clear(); 3534 UsedByIndicesMap.clear(); 3535 UniqueItems.clear(); 3536 3537 // Now iterate through the worklist and add new formulae. 3538 for (SmallVectorImpl<WorkItem>::const_iterator I = WorkItems.begin(), 3539 E = WorkItems.end(); I != E; ++I) { 3540 const WorkItem &WI = *I; 3541 size_t LUIdx = WI.LUIdx; 3542 LSRUse &LU = Uses[LUIdx]; 3543 int64_t Imm = WI.Imm; 3544 const SCEV *OrigReg = WI.OrigReg; 3545 3546 Type *IntTy = SE.getEffectiveSCEVType(OrigReg->getType()); 3547 const SCEV *NegImmS = SE.getSCEV(ConstantInt::get(IntTy, -(uint64_t)Imm)); 3548 unsigned BitWidth = SE.getTypeSizeInBits(IntTy); 3549 3550 // TODO: Use a more targeted data structure. 3551 for (size_t L = 0, LE = LU.Formulae.size(); L != LE; ++L) { 3552 const Formula &F = LU.Formulae[L]; 3553 // Use the immediate in the scaled register. 3554 if (F.ScaledReg == OrigReg) { 3555 int64_t Offs = (uint64_t)F.AM.BaseOffs + 3556 Imm * (uint64_t)F.AM.Scale; 3557 // Don't create 50 + reg(-50). 3558 if (F.referencesReg(SE.getSCEV( 3559 ConstantInt::get(IntTy, -(uint64_t)Offs)))) 3560 continue; 3561 Formula NewF = F; 3562 NewF.AM.BaseOffs = Offs; 3563 if (!isLegalUse(NewF.AM, LU.MinOffset, LU.MaxOffset, 3564 LU.Kind, LU.AccessTy, TLI)) 3565 continue; 3566 NewF.ScaledReg = SE.getAddExpr(NegImmS, NewF.ScaledReg); 3567 3568 // If the new scale is a constant in a register, and adding the constant 3569 // value to the immediate would produce a value closer to zero than the 3570 // immediate itself, then the formula isn't worthwhile. 3571 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(NewF.ScaledReg)) 3572 if (C->getValue()->isNegative() != 3573 (NewF.AM.BaseOffs < 0) && 3574 (C->getValue()->getValue().abs() * APInt(BitWidth, F.AM.Scale)) 3575 .ule(abs64(NewF.AM.BaseOffs))) 3576 continue; 3577 3578 // OK, looks good. 3579 (void)InsertFormula(LU, LUIdx, NewF); 3580 } else { 3581 // Use the immediate in a base register. 3582 for (size_t N = 0, NE = F.BaseRegs.size(); N != NE; ++N) { 3583 const SCEV *BaseReg = F.BaseRegs[N]; 3584 if (BaseReg != OrigReg) 3585 continue; 3586 Formula NewF = F; 3587 NewF.AM.BaseOffs = (uint64_t)NewF.AM.BaseOffs + Imm; 3588 if (!isLegalUse(NewF.AM, LU.MinOffset, LU.MaxOffset, 3589 LU.Kind, LU.AccessTy, TLI)) { 3590 if (!TLI || 3591 !TLI->isLegalAddImmediate((uint64_t)NewF.UnfoldedOffset + Imm)) 3592 continue; 3593 NewF = F; 3594 NewF.UnfoldedOffset = (uint64_t)NewF.UnfoldedOffset + Imm; 3595 } 3596 NewF.BaseRegs[N] = SE.getAddExpr(NegImmS, BaseReg); 3597 3598 // If the new formula has a constant in a register, and adding the 3599 // constant value to the immediate would produce a value closer to 3600 // zero than the immediate itself, then the formula isn't worthwhile. 3601 for (SmallVectorImpl<const SCEV *>::const_iterator 3602 J = NewF.BaseRegs.begin(), JE = NewF.BaseRegs.end(); 3603 J != JE; ++J) 3604 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(*J)) 3605 if ((C->getValue()->getValue() + NewF.AM.BaseOffs).abs().slt( 3606 abs64(NewF.AM.BaseOffs)) && 3607 (C->getValue()->getValue() + 3608 NewF.AM.BaseOffs).countTrailingZeros() >= 3609 CountTrailingZeros_64(NewF.AM.BaseOffs)) 3610 goto skip_formula; 3611 3612 // Ok, looks good. 3613 (void)InsertFormula(LU, LUIdx, NewF); 3614 break; 3615 skip_formula:; 3616 } 3617 } 3618 } 3619 } 3620} 3621 3622/// GenerateAllReuseFormulae - Generate formulae for each use. 3623void 3624LSRInstance::GenerateAllReuseFormulae() { 3625 // This is split into multiple loops so that hasRegsUsedByUsesOtherThan 3626 // queries are more precise. 3627 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) { 3628 LSRUse &LU = Uses[LUIdx]; 3629 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i) 3630 GenerateReassociations(LU, LUIdx, LU.Formulae[i]); 3631 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i) 3632 GenerateCombinations(LU, LUIdx, LU.Formulae[i]); 3633 } 3634 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) { 3635 LSRUse &LU = Uses[LUIdx]; 3636 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i) 3637 GenerateSymbolicOffsets(LU, LUIdx, LU.Formulae[i]); 3638 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i) 3639 GenerateConstantOffsets(LU, LUIdx, LU.Formulae[i]); 3640 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i) 3641 GenerateICmpZeroScales(LU, LUIdx, LU.Formulae[i]); 3642 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i) 3643 GenerateScales(LU, LUIdx, LU.Formulae[i]); 3644 } 3645 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) { 3646 LSRUse &LU = Uses[LUIdx]; 3647 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i) 3648 GenerateTruncates(LU, LUIdx, LU.Formulae[i]); 3649 } 3650 3651 GenerateCrossUseConstantOffsets(); 3652 3653 DEBUG(dbgs() << "\n" 3654 "After generating reuse formulae:\n"; 3655 print_uses(dbgs())); 3656} 3657 3658/// If there are multiple formulae with the same set of registers used 3659/// by other uses, pick the best one and delete the others. 3660void LSRInstance::FilterOutUndesirableDedicatedRegisters() { 3661 DenseSet<const SCEV *> VisitedRegs; 3662 SmallPtrSet<const SCEV *, 16> Regs; 3663 SmallPtrSet<const SCEV *, 16> LoserRegs; 3664#ifndef NDEBUG 3665 bool ChangedFormulae = false; 3666#endif 3667 3668 // Collect the best formula for each unique set of shared registers. This 3669 // is reset for each use. 3670 typedef DenseMap<SmallVector<const SCEV *, 2>, size_t, UniquifierDenseMapInfo> 3671 BestFormulaeTy; 3672 BestFormulaeTy BestFormulae; 3673 3674 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) { 3675 LSRUse &LU = Uses[LUIdx]; 3676 DEBUG(dbgs() << "Filtering for use "; LU.print(dbgs()); dbgs() << '\n'); 3677 3678 bool Any = false; 3679 for (size_t FIdx = 0, NumForms = LU.Formulae.size(); 3680 FIdx != NumForms; ++FIdx) { 3681 Formula &F = LU.Formulae[FIdx]; 3682 3683 // Some formulas are instant losers. For example, they may depend on 3684 // nonexistent AddRecs from other loops. These need to be filtered 3685 // immediately, otherwise heuristics could choose them over others leading 3686 // to an unsatisfactory solution. Passing LoserRegs into RateFormula here 3687 // avoids the need to recompute this information across formulae using the 3688 // same bad AddRec. Passing LoserRegs is also essential unless we remove 3689 // the corresponding bad register from the Regs set. 3690 Cost CostF; 3691 Regs.clear(); 3692 CostF.RateFormula(F, Regs, VisitedRegs, L, LU.Offsets, SE, DT, 3693 &LoserRegs); 3694 if (CostF.isLoser()) { 3695 // During initial formula generation, undesirable formulae are generated 3696 // by uses within other loops that have some non-trivial address mode or 3697 // use the postinc form of the IV. LSR needs to provide these formulae 3698 // as the basis of rediscovering the desired formula that uses an AddRec 3699 // corresponding to the existing phi. Once all formulae have been 3700 // generated, these initial losers may be pruned. 3701 DEBUG(dbgs() << " Filtering loser "; F.print(dbgs()); 3702 dbgs() << "\n"); 3703 } 3704 else { 3705 SmallVector<const SCEV *, 2> Key; 3706 for (SmallVectorImpl<const SCEV *>::const_iterator J = F.BaseRegs.begin(), 3707 JE = F.BaseRegs.end(); J != JE; ++J) { 3708 const SCEV *Reg = *J; 3709 if (RegUses.isRegUsedByUsesOtherThan(Reg, LUIdx)) 3710 Key.push_back(Reg); 3711 } 3712 if (F.ScaledReg && 3713 RegUses.isRegUsedByUsesOtherThan(F.ScaledReg, LUIdx)) 3714 Key.push_back(F.ScaledReg); 3715 // Unstable sort by host order ok, because this is only used for 3716 // uniquifying. 3717 std::sort(Key.begin(), Key.end()); 3718 3719 std::pair<BestFormulaeTy::const_iterator, bool> P = 3720 BestFormulae.insert(std::make_pair(Key, FIdx)); 3721 if (P.second) 3722 continue; 3723 3724 Formula &Best = LU.Formulae[P.first->second]; 3725 3726 Cost CostBest; 3727 Regs.clear(); 3728 CostBest.RateFormula(Best, Regs, VisitedRegs, L, LU.Offsets, SE, DT); 3729 if (CostF < CostBest) 3730 std::swap(F, Best); 3731 DEBUG(dbgs() << " Filtering out formula "; F.print(dbgs()); 3732 dbgs() << "\n" 3733 " in favor of formula "; Best.print(dbgs()); 3734 dbgs() << '\n'); 3735 } 3736#ifndef NDEBUG 3737 ChangedFormulae = true; 3738#endif 3739 LU.DeleteFormula(F); 3740 --FIdx; 3741 --NumForms; 3742 Any = true; 3743 } 3744 3745 // Now that we've filtered out some formulae, recompute the Regs set. 3746 if (Any) 3747 LU.RecomputeRegs(LUIdx, RegUses); 3748 3749 // Reset this to prepare for the next use. 3750 BestFormulae.clear(); 3751 } 3752 3753 DEBUG(if (ChangedFormulae) { 3754 dbgs() << "\n" 3755 "After filtering out undesirable candidates:\n"; 3756 print_uses(dbgs()); 3757 }); 3758} 3759 3760// This is a rough guess that seems to work fairly well. 3761static const size_t ComplexityLimit = UINT16_MAX; 3762 3763/// EstimateSearchSpaceComplexity - Estimate the worst-case number of 3764/// solutions the solver might have to consider. It almost never considers 3765/// this many solutions because it prune the search space, but the pruning 3766/// isn't always sufficient. 3767size_t LSRInstance::EstimateSearchSpaceComplexity() const { 3768 size_t Power = 1; 3769 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(), 3770 E = Uses.end(); I != E; ++I) { 3771 size_t FSize = I->Formulae.size(); 3772 if (FSize >= ComplexityLimit) { 3773 Power = ComplexityLimit; 3774 break; 3775 } 3776 Power *= FSize; 3777 if (Power >= ComplexityLimit) 3778 break; 3779 } 3780 return Power; 3781} 3782 3783/// NarrowSearchSpaceByDetectingSupersets - When one formula uses a superset 3784/// of the registers of another formula, it won't help reduce register 3785/// pressure (though it may not necessarily hurt register pressure); remove 3786/// it to simplify the system. 3787void LSRInstance::NarrowSearchSpaceByDetectingSupersets() { 3788 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) { 3789 DEBUG(dbgs() << "The search space is too complex.\n"); 3790 3791 DEBUG(dbgs() << "Narrowing the search space by eliminating formulae " 3792 "which use a superset of registers used by other " 3793 "formulae.\n"); 3794 3795 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) { 3796 LSRUse &LU = Uses[LUIdx]; 3797 bool Any = false; 3798 for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) { 3799 Formula &F = LU.Formulae[i]; 3800 // Look for a formula with a constant or GV in a register. If the use 3801 // also has a formula with that same value in an immediate field, 3802 // delete the one that uses a register. 3803 for (SmallVectorImpl<const SCEV *>::const_iterator 3804 I = F.BaseRegs.begin(), E = F.BaseRegs.end(); I != E; ++I) { 3805 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(*I)) { 3806 Formula NewF = F; 3807 NewF.AM.BaseOffs += C->getValue()->getSExtValue(); 3808 NewF.BaseRegs.erase(NewF.BaseRegs.begin() + 3809 (I - F.BaseRegs.begin())); 3810 if (LU.HasFormulaWithSameRegs(NewF)) { 3811 DEBUG(dbgs() << " Deleting "; F.print(dbgs()); dbgs() << '\n'); 3812 LU.DeleteFormula(F); 3813 --i; 3814 --e; 3815 Any = true; 3816 break; 3817 } 3818 } else if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(*I)) { 3819 if (GlobalValue *GV = dyn_cast<GlobalValue>(U->getValue())) 3820 if (!F.AM.BaseGV) { 3821 Formula NewF = F; 3822 NewF.AM.BaseGV = GV; 3823 NewF.BaseRegs.erase(NewF.BaseRegs.begin() + 3824 (I - F.BaseRegs.begin())); 3825 if (LU.HasFormulaWithSameRegs(NewF)) { 3826 DEBUG(dbgs() << " Deleting "; F.print(dbgs()); 3827 dbgs() << '\n'); 3828 LU.DeleteFormula(F); 3829 --i; 3830 --e; 3831 Any = true; 3832 break; 3833 } 3834 } 3835 } 3836 } 3837 } 3838 if (Any) 3839 LU.RecomputeRegs(LUIdx, RegUses); 3840 } 3841 3842 DEBUG(dbgs() << "After pre-selection:\n"; 3843 print_uses(dbgs())); 3844 } 3845} 3846 3847/// NarrowSearchSpaceByCollapsingUnrolledCode - When there are many registers 3848/// for expressions like A, A+1, A+2, etc., allocate a single register for 3849/// them. 3850void LSRInstance::NarrowSearchSpaceByCollapsingUnrolledCode() { 3851 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) { 3852 DEBUG(dbgs() << "The search space is too complex.\n"); 3853 3854 DEBUG(dbgs() << "Narrowing the search space by assuming that uses " 3855 "separated by a constant offset will use the same " 3856 "registers.\n"); 3857 3858 // This is especially useful for unrolled loops. 3859 3860 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) { 3861 LSRUse &LU = Uses[LUIdx]; 3862 for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(), 3863 E = LU.Formulae.end(); I != E; ++I) { 3864 const Formula &F = *I; 3865 if (F.AM.BaseOffs != 0 && F.AM.Scale == 0) { 3866 if (LSRUse *LUThatHas = FindUseWithSimilarFormula(F, LU)) { 3867 if (reconcileNewOffset(*LUThatHas, F.AM.BaseOffs, 3868 /*HasBaseReg=*/false, 3869 LU.Kind, LU.AccessTy)) { 3870 DEBUG(dbgs() << " Deleting use "; LU.print(dbgs()); 3871 dbgs() << '\n'); 3872 3873 LUThatHas->AllFixupsOutsideLoop &= LU.AllFixupsOutsideLoop; 3874 3875 // Update the relocs to reference the new use. 3876 for (SmallVectorImpl<LSRFixup>::iterator I = Fixups.begin(), 3877 E = Fixups.end(); I != E; ++I) { 3878 LSRFixup &Fixup = *I; 3879 if (Fixup.LUIdx == LUIdx) { 3880 Fixup.LUIdx = LUThatHas - &Uses.front(); 3881 Fixup.Offset += F.AM.BaseOffs; 3882 // Add the new offset to LUThatHas' offset list. 3883 if (LUThatHas->Offsets.back() != Fixup.Offset) { 3884 LUThatHas->Offsets.push_back(Fixup.Offset); 3885 if (Fixup.Offset > LUThatHas->MaxOffset) 3886 LUThatHas->MaxOffset = Fixup.Offset; 3887 if (Fixup.Offset < LUThatHas->MinOffset) 3888 LUThatHas->MinOffset = Fixup.Offset; 3889 } 3890 DEBUG(dbgs() << "New fixup has offset " 3891 << Fixup.Offset << '\n'); 3892 } 3893 if (Fixup.LUIdx == NumUses-1) 3894 Fixup.LUIdx = LUIdx; 3895 } 3896 3897 // Delete formulae from the new use which are no longer legal. 3898 bool Any = false; 3899 for (size_t i = 0, e = LUThatHas->Formulae.size(); i != e; ++i) { 3900 Formula &F = LUThatHas->Formulae[i]; 3901 if (!isLegalUse(F.AM, 3902 LUThatHas->MinOffset, LUThatHas->MaxOffset, 3903 LUThatHas->Kind, LUThatHas->AccessTy, TLI)) { 3904 DEBUG(dbgs() << " Deleting "; F.print(dbgs()); 3905 dbgs() << '\n'); 3906 LUThatHas->DeleteFormula(F); 3907 --i; 3908 --e; 3909 Any = true; 3910 } 3911 } 3912 if (Any) 3913 LUThatHas->RecomputeRegs(LUThatHas - &Uses.front(), RegUses); 3914 3915 // Delete the old use. 3916 DeleteUse(LU, LUIdx); 3917 --LUIdx; 3918 --NumUses; 3919 break; 3920 } 3921 } 3922 } 3923 } 3924 } 3925 3926 DEBUG(dbgs() << "After pre-selection:\n"; 3927 print_uses(dbgs())); 3928 } 3929} 3930 3931/// NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters - Call 3932/// FilterOutUndesirableDedicatedRegisters again, if necessary, now that 3933/// we've done more filtering, as it may be able to find more formulae to 3934/// eliminate. 3935void LSRInstance::NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters(){ 3936 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) { 3937 DEBUG(dbgs() << "The search space is too complex.\n"); 3938 3939 DEBUG(dbgs() << "Narrowing the search space by re-filtering out " 3940 "undesirable dedicated registers.\n"); 3941 3942 FilterOutUndesirableDedicatedRegisters(); 3943 3944 DEBUG(dbgs() << "After pre-selection:\n"; 3945 print_uses(dbgs())); 3946 } 3947} 3948 3949/// NarrowSearchSpaceByPickingWinnerRegs - Pick a register which seems likely 3950/// to be profitable, and then in any use which has any reference to that 3951/// register, delete all formulae which do not reference that register. 3952void LSRInstance::NarrowSearchSpaceByPickingWinnerRegs() { 3953 // With all other options exhausted, loop until the system is simple 3954 // enough to handle. 3955 SmallPtrSet<const SCEV *, 4> Taken; 3956 while (EstimateSearchSpaceComplexity() >= ComplexityLimit) { 3957 // Ok, we have too many of formulae on our hands to conveniently handle. 3958 // Use a rough heuristic to thin out the list. 3959 DEBUG(dbgs() << "The search space is too complex.\n"); 3960 3961 // Pick the register which is used by the most LSRUses, which is likely 3962 // to be a good reuse register candidate. 3963 const SCEV *Best = 0; 3964 unsigned BestNum = 0; 3965 for (RegUseTracker::const_iterator I = RegUses.begin(), E = RegUses.end(); 3966 I != E; ++I) { 3967 const SCEV *Reg = *I; 3968 if (Taken.count(Reg)) 3969 continue; 3970 if (!Best) 3971 Best = Reg; 3972 else { 3973 unsigned Count = RegUses.getUsedByIndices(Reg).count(); 3974 if (Count > BestNum) { 3975 Best = Reg; 3976 BestNum = Count; 3977 } 3978 } 3979 } 3980 3981 DEBUG(dbgs() << "Narrowing the search space by assuming " << *Best 3982 << " will yield profitable reuse.\n"); 3983 Taken.insert(Best); 3984 3985 // In any use with formulae which references this register, delete formulae 3986 // which don't reference it. 3987 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) { 3988 LSRUse &LU = Uses[LUIdx]; 3989 if (!LU.Regs.count(Best)) continue; 3990 3991 bool Any = false; 3992 for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) { 3993 Formula &F = LU.Formulae[i]; 3994 if (!F.referencesReg(Best)) { 3995 DEBUG(dbgs() << " Deleting "; F.print(dbgs()); dbgs() << '\n'); 3996 LU.DeleteFormula(F); 3997 --e; 3998 --i; 3999 Any = true; 4000 assert(e != 0 && "Use has no formulae left! Is Regs inconsistent?"); 4001 continue; 4002 } 4003 } 4004 4005 if (Any) 4006 LU.RecomputeRegs(LUIdx, RegUses); 4007 } 4008 4009 DEBUG(dbgs() << "After pre-selection:\n"; 4010 print_uses(dbgs())); 4011 } 4012} 4013 4014/// NarrowSearchSpaceUsingHeuristics - If there are an extraordinary number of 4015/// formulae to choose from, use some rough heuristics to prune down the number 4016/// of formulae. This keeps the main solver from taking an extraordinary amount 4017/// of time in some worst-case scenarios. 4018void LSRInstance::NarrowSearchSpaceUsingHeuristics() { 4019 NarrowSearchSpaceByDetectingSupersets(); 4020 NarrowSearchSpaceByCollapsingUnrolledCode(); 4021 NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters(); 4022 NarrowSearchSpaceByPickingWinnerRegs(); 4023} 4024 4025/// SolveRecurse - This is the recursive solver. 4026void LSRInstance::SolveRecurse(SmallVectorImpl<const Formula *> &Solution, 4027 Cost &SolutionCost, 4028 SmallVectorImpl<const Formula *> &Workspace, 4029 const Cost &CurCost, 4030 const SmallPtrSet<const SCEV *, 16> &CurRegs, 4031 DenseSet<const SCEV *> &VisitedRegs) const { 4032 // Some ideas: 4033 // - prune more: 4034 // - use more aggressive filtering 4035 // - sort the formula so that the most profitable solutions are found first 4036 // - sort the uses too 4037 // - search faster: 4038 // - don't compute a cost, and then compare. compare while computing a cost 4039 // and bail early. 4040 // - track register sets with SmallBitVector 4041 4042 const LSRUse &LU = Uses[Workspace.size()]; 4043 4044 // If this use references any register that's already a part of the 4045 // in-progress solution, consider it a requirement that a formula must 4046 // reference that register in order to be considered. This prunes out 4047 // unprofitable searching. 4048 SmallSetVector<const SCEV *, 4> ReqRegs; 4049 for (SmallPtrSet<const SCEV *, 16>::const_iterator I = CurRegs.begin(), 4050 E = CurRegs.end(); I != E; ++I) 4051 if (LU.Regs.count(*I)) 4052 ReqRegs.insert(*I); 4053 4054 SmallPtrSet<const SCEV *, 16> NewRegs; 4055 Cost NewCost; 4056 for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(), 4057 E = LU.Formulae.end(); I != E; ++I) { 4058 const Formula &F = *I; 4059 4060 // Ignore formulae which do not use any of the required registers. 4061 bool SatisfiedReqReg = true; 4062 for (SmallSetVector<const SCEV *, 4>::const_iterator J = ReqRegs.begin(), 4063 JE = ReqRegs.end(); J != JE; ++J) { 4064 const SCEV *Reg = *J; 4065 if ((!F.ScaledReg || F.ScaledReg != Reg) && 4066 std::find(F.BaseRegs.begin(), F.BaseRegs.end(), Reg) == 4067 F.BaseRegs.end()) { 4068 SatisfiedReqReg = false; 4069 break; 4070 } 4071 } 4072 if (!SatisfiedReqReg) { 4073 // If none of the formulae satisfied the required registers, then we could 4074 // clear ReqRegs and try again. Currently, we simply give up in this case. 4075 continue; 4076 } 4077 4078 // Evaluate the cost of the current formula. If it's already worse than 4079 // the current best, prune the search at that point. 4080 NewCost = CurCost; 4081 NewRegs = CurRegs; 4082 NewCost.RateFormula(F, NewRegs, VisitedRegs, L, LU.Offsets, SE, DT); 4083 if (NewCost < SolutionCost) { 4084 Workspace.push_back(&F); 4085 if (Workspace.size() != Uses.size()) { 4086 SolveRecurse(Solution, SolutionCost, Workspace, NewCost, 4087 NewRegs, VisitedRegs); 4088 if (F.getNumRegs() == 1 && Workspace.size() == 1) 4089 VisitedRegs.insert(F.ScaledReg ? F.ScaledReg : F.BaseRegs[0]); 4090 } else { 4091 DEBUG(dbgs() << "New best at "; NewCost.print(dbgs()); 4092 dbgs() << ".\n Regs:"; 4093 for (SmallPtrSet<const SCEV *, 16>::const_iterator 4094 I = NewRegs.begin(), E = NewRegs.end(); I != E; ++I) 4095 dbgs() << ' ' << **I; 4096 dbgs() << '\n'); 4097 4098 SolutionCost = NewCost; 4099 Solution = Workspace; 4100 } 4101 Workspace.pop_back(); 4102 } 4103 } 4104} 4105 4106/// Solve - Choose one formula from each use. Return the results in the given 4107/// Solution vector. 4108void LSRInstance::Solve(SmallVectorImpl<const Formula *> &Solution) const { 4109 SmallVector<const Formula *, 8> Workspace; 4110 Cost SolutionCost; 4111 SolutionCost.Loose(); 4112 Cost CurCost; 4113 SmallPtrSet<const SCEV *, 16> CurRegs; 4114 DenseSet<const SCEV *> VisitedRegs; 4115 Workspace.reserve(Uses.size()); 4116 4117 // SolveRecurse does all the work. 4118 SolveRecurse(Solution, SolutionCost, Workspace, CurCost, 4119 CurRegs, VisitedRegs); 4120 if (Solution.empty()) { 4121 DEBUG(dbgs() << "\nNo Satisfactory Solution\n"); 4122 return; 4123 } 4124 4125 // Ok, we've now made all our decisions. 4126 DEBUG(dbgs() << "\n" 4127 "The chosen solution requires "; SolutionCost.print(dbgs()); 4128 dbgs() << ":\n"; 4129 for (size_t i = 0, e = Uses.size(); i != e; ++i) { 4130 dbgs() << " "; 4131 Uses[i].print(dbgs()); 4132 dbgs() << "\n" 4133 " "; 4134 Solution[i]->print(dbgs()); 4135 dbgs() << '\n'; 4136 }); 4137 4138 assert(Solution.size() == Uses.size() && "Malformed solution!"); 4139} 4140 4141/// HoistInsertPosition - Helper for AdjustInsertPositionForExpand. Climb up 4142/// the dominator tree far as we can go while still being dominated by the 4143/// input positions. This helps canonicalize the insert position, which 4144/// encourages sharing. 4145BasicBlock::iterator 4146LSRInstance::HoistInsertPosition(BasicBlock::iterator IP, 4147 const SmallVectorImpl<Instruction *> &Inputs) 4148 const { 4149 for (;;) { 4150 const Loop *IPLoop = LI.getLoopFor(IP->getParent()); 4151 unsigned IPLoopDepth = IPLoop ? IPLoop->getLoopDepth() : 0; 4152 4153 BasicBlock *IDom; 4154 for (DomTreeNode *Rung = DT.getNode(IP->getParent()); ; ) { 4155 if (!Rung) return IP; 4156 Rung = Rung->getIDom(); 4157 if (!Rung) return IP; 4158 IDom = Rung->getBlock(); 4159 4160 // Don't climb into a loop though. 4161 const Loop *IDomLoop = LI.getLoopFor(IDom); 4162 unsigned IDomDepth = IDomLoop ? IDomLoop->getLoopDepth() : 0; 4163 if (IDomDepth <= IPLoopDepth && 4164 (IDomDepth != IPLoopDepth || IDomLoop == IPLoop)) 4165 break; 4166 } 4167 4168 bool AllDominate = true; 4169 Instruction *BetterPos = 0; 4170 Instruction *Tentative = IDom->getTerminator(); 4171 for (SmallVectorImpl<Instruction *>::const_iterator I = Inputs.begin(), 4172 E = Inputs.end(); I != E; ++I) { 4173 Instruction *Inst = *I; 4174 if (Inst == Tentative || !DT.dominates(Inst, Tentative)) { 4175 AllDominate = false; 4176 break; 4177 } 4178 // Attempt to find an insert position in the middle of the block, 4179 // instead of at the end, so that it can be used for other expansions. 4180 if (IDom == Inst->getParent() && 4181 (!BetterPos || !DT.dominates(Inst, BetterPos))) 4182 BetterPos = llvm::next(BasicBlock::iterator(Inst)); 4183 } 4184 if (!AllDominate) 4185 break; 4186 if (BetterPos) 4187 IP = BetterPos; 4188 else 4189 IP = Tentative; 4190 } 4191 4192 return IP; 4193} 4194 4195/// AdjustInsertPositionForExpand - Determine an input position which will be 4196/// dominated by the operands and which will dominate the result. 4197BasicBlock::iterator 4198LSRInstance::AdjustInsertPositionForExpand(BasicBlock::iterator LowestIP, 4199 const LSRFixup &LF, 4200 const LSRUse &LU, 4201 SCEVExpander &Rewriter) const { 4202 // Collect some instructions which must be dominated by the 4203 // expanding replacement. These must be dominated by any operands that 4204 // will be required in the expansion. 4205 SmallVector<Instruction *, 4> Inputs; 4206 if (Instruction *I = dyn_cast<Instruction>(LF.OperandValToReplace)) 4207 Inputs.push_back(I); 4208 if (LU.Kind == LSRUse::ICmpZero) 4209 if (Instruction *I = 4210 dyn_cast<Instruction>(cast<ICmpInst>(LF.UserInst)->getOperand(1))) 4211 Inputs.push_back(I); 4212 if (LF.PostIncLoops.count(L)) { 4213 if (LF.isUseFullyOutsideLoop(L)) 4214 Inputs.push_back(L->getLoopLatch()->getTerminator()); 4215 else 4216 Inputs.push_back(IVIncInsertPos); 4217 } 4218 // The expansion must also be dominated by the increment positions of any 4219 // loops it for which it is using post-inc mode. 4220 for (PostIncLoopSet::const_iterator I = LF.PostIncLoops.begin(), 4221 E = LF.PostIncLoops.end(); I != E; ++I) { 4222 const Loop *PIL = *I; 4223 if (PIL == L) continue; 4224 4225 // Be dominated by the loop exit. 4226 SmallVector<BasicBlock *, 4> ExitingBlocks; 4227 PIL->getExitingBlocks(ExitingBlocks); 4228 if (!ExitingBlocks.empty()) { 4229 BasicBlock *BB = ExitingBlocks[0]; 4230 for (unsigned i = 1, e = ExitingBlocks.size(); i != e; ++i) 4231 BB = DT.findNearestCommonDominator(BB, ExitingBlocks[i]); 4232 Inputs.push_back(BB->getTerminator()); 4233 } 4234 } 4235 4236 assert(!isa<PHINode>(LowestIP) && !isa<LandingPadInst>(LowestIP) 4237 && !isa<DbgInfoIntrinsic>(LowestIP) && 4238 "Insertion point must be a normal instruction"); 4239 4240 // Then, climb up the immediate dominator tree as far as we can go while 4241 // still being dominated by the input positions. 4242 BasicBlock::iterator IP = HoistInsertPosition(LowestIP, Inputs); 4243 4244 // Don't insert instructions before PHI nodes. 4245 while (isa<PHINode>(IP)) ++IP; 4246 4247 // Ignore landingpad instructions. 4248 while (isa<LandingPadInst>(IP)) ++IP; 4249 4250 // Ignore debug intrinsics. 4251 while (isa<DbgInfoIntrinsic>(IP)) ++IP; 4252 4253 // Set IP below instructions recently inserted by SCEVExpander. This keeps the 4254 // IP consistent across expansions and allows the previously inserted 4255 // instructions to be reused by subsequent expansion. 4256 while (Rewriter.isInsertedInstruction(IP) && IP != LowestIP) ++IP; 4257 4258 return IP; 4259} 4260 4261/// Expand - Emit instructions for the leading candidate expression for this 4262/// LSRUse (this is called "expanding"). 4263Value *LSRInstance::Expand(const LSRFixup &LF, 4264 const Formula &F, 4265 BasicBlock::iterator IP, 4266 SCEVExpander &Rewriter, 4267 SmallVectorImpl<WeakVH> &DeadInsts) const { 4268 const LSRUse &LU = Uses[LF.LUIdx]; 4269 4270 // Determine an input position which will be dominated by the operands and 4271 // which will dominate the result. 4272 IP = AdjustInsertPositionForExpand(IP, LF, LU, Rewriter); 4273 4274 // Inform the Rewriter if we have a post-increment use, so that it can 4275 // perform an advantageous expansion. 4276 Rewriter.setPostInc(LF.PostIncLoops); 4277 4278 // This is the type that the user actually needs. 4279 Type *OpTy = LF.OperandValToReplace->getType(); 4280 // This will be the type that we'll initially expand to. 4281 Type *Ty = F.getType(); 4282 if (!Ty) 4283 // No type known; just expand directly to the ultimate type. 4284 Ty = OpTy; 4285 else if (SE.getEffectiveSCEVType(Ty) == SE.getEffectiveSCEVType(OpTy)) 4286 // Expand directly to the ultimate type if it's the right size. 4287 Ty = OpTy; 4288 // This is the type to do integer arithmetic in. 4289 Type *IntTy = SE.getEffectiveSCEVType(Ty); 4290 4291 // Build up a list of operands to add together to form the full base. 4292 SmallVector<const SCEV *, 8> Ops; 4293 4294 // Expand the BaseRegs portion. 4295 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(), 4296 E = F.BaseRegs.end(); I != E; ++I) { 4297 const SCEV *Reg = *I; 4298 assert(!Reg->isZero() && "Zero allocated in a base register!"); 4299 4300 // If we're expanding for a post-inc user, make the post-inc adjustment. 4301 PostIncLoopSet &Loops = const_cast<PostIncLoopSet &>(LF.PostIncLoops); 4302 Reg = TransformForPostIncUse(Denormalize, Reg, 4303 LF.UserInst, LF.OperandValToReplace, 4304 Loops, SE, DT); 4305 4306 Ops.push_back(SE.getUnknown(Rewriter.expandCodeFor(Reg, 0, IP))); 4307 } 4308 4309 // Expand the ScaledReg portion. 4310 Value *ICmpScaledV = 0; 4311 if (F.AM.Scale != 0) { 4312 const SCEV *ScaledS = F.ScaledReg; 4313 4314 // If we're expanding for a post-inc user, make the post-inc adjustment. 4315 PostIncLoopSet &Loops = const_cast<PostIncLoopSet &>(LF.PostIncLoops); 4316 ScaledS = TransformForPostIncUse(Denormalize, ScaledS, 4317 LF.UserInst, LF.OperandValToReplace, 4318 Loops, SE, DT); 4319 4320 if (LU.Kind == LSRUse::ICmpZero) { 4321 // An interesting way of "folding" with an icmp is to use a negated 4322 // scale, which we'll implement by inserting it into the other operand 4323 // of the icmp. 4324 assert(F.AM.Scale == -1 && 4325 "The only scale supported by ICmpZero uses is -1!"); 4326 ICmpScaledV = Rewriter.expandCodeFor(ScaledS, 0, IP); 4327 } else { 4328 // Otherwise just expand the scaled register and an explicit scale, 4329 // which is expected to be matched as part of the address. 4330 4331 // Flush the operand list to suppress SCEVExpander hoisting address modes. 4332 if (!Ops.empty() && LU.Kind == LSRUse::Address) { 4333 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP); 4334 Ops.clear(); 4335 Ops.push_back(SE.getUnknown(FullV)); 4336 } 4337 ScaledS = SE.getUnknown(Rewriter.expandCodeFor(ScaledS, 0, IP)); 4338 ScaledS = SE.getMulExpr(ScaledS, 4339 SE.getConstant(ScaledS->getType(), F.AM.Scale)); 4340 Ops.push_back(ScaledS); 4341 } 4342 } 4343 4344 // Expand the GV portion. 4345 if (F.AM.BaseGV) { 4346 // Flush the operand list to suppress SCEVExpander hoisting. 4347 if (!Ops.empty()) { 4348 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP); 4349 Ops.clear(); 4350 Ops.push_back(SE.getUnknown(FullV)); 4351 } 4352 Ops.push_back(SE.getUnknown(F.AM.BaseGV)); 4353 } 4354 4355 // Flush the operand list to suppress SCEVExpander hoisting of both folded and 4356 // unfolded offsets. LSR assumes they both live next to their uses. 4357 if (!Ops.empty()) { 4358 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP); 4359 Ops.clear(); 4360 Ops.push_back(SE.getUnknown(FullV)); 4361 } 4362 4363 // Expand the immediate portion. 4364 int64_t Offset = (uint64_t)F.AM.BaseOffs + LF.Offset; 4365 if (Offset != 0) { 4366 if (LU.Kind == LSRUse::ICmpZero) { 4367 // The other interesting way of "folding" with an ICmpZero is to use a 4368 // negated immediate. 4369 if (!ICmpScaledV) 4370 ICmpScaledV = ConstantInt::get(IntTy, -(uint64_t)Offset); 4371 else { 4372 Ops.push_back(SE.getUnknown(ICmpScaledV)); 4373 ICmpScaledV = ConstantInt::get(IntTy, Offset); 4374 } 4375 } else { 4376 // Just add the immediate values. These again are expected to be matched 4377 // as part of the address. 4378 Ops.push_back(SE.getUnknown(ConstantInt::getSigned(IntTy, Offset))); 4379 } 4380 } 4381 4382 // Expand the unfolded offset portion. 4383 int64_t UnfoldedOffset = F.UnfoldedOffset; 4384 if (UnfoldedOffset != 0) { 4385 // Just add the immediate values. 4386 Ops.push_back(SE.getUnknown(ConstantInt::getSigned(IntTy, 4387 UnfoldedOffset))); 4388 } 4389 4390 // Emit instructions summing all the operands. 4391 const SCEV *FullS = Ops.empty() ? 4392 SE.getConstant(IntTy, 0) : 4393 SE.getAddExpr(Ops); 4394 Value *FullV = Rewriter.expandCodeFor(FullS, Ty, IP); 4395 4396 // We're done expanding now, so reset the rewriter. 4397 Rewriter.clearPostInc(); 4398 4399 // An ICmpZero Formula represents an ICmp which we're handling as a 4400 // comparison against zero. Now that we've expanded an expression for that 4401 // form, update the ICmp's other operand. 4402 if (LU.Kind == LSRUse::ICmpZero) { 4403 ICmpInst *CI = cast<ICmpInst>(LF.UserInst); 4404 DeadInsts.push_back(CI->getOperand(1)); 4405 assert(!F.AM.BaseGV && "ICmp does not support folding a global value and " 4406 "a scale at the same time!"); 4407 if (F.AM.Scale == -1) { 4408 if (ICmpScaledV->getType() != OpTy) { 4409 Instruction *Cast = 4410 CastInst::Create(CastInst::getCastOpcode(ICmpScaledV, false, 4411 OpTy, false), 4412 ICmpScaledV, OpTy, "tmp", CI); 4413 ICmpScaledV = Cast; 4414 } 4415 CI->setOperand(1, ICmpScaledV); 4416 } else { 4417 assert(F.AM.Scale == 0 && 4418 "ICmp does not support folding a global value and " 4419 "a scale at the same time!"); 4420 Constant *C = ConstantInt::getSigned(SE.getEffectiveSCEVType(OpTy), 4421 -(uint64_t)Offset); 4422 if (C->getType() != OpTy) 4423 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false, 4424 OpTy, false), 4425 C, OpTy); 4426 4427 CI->setOperand(1, C); 4428 } 4429 } 4430 4431 return FullV; 4432} 4433 4434/// RewriteForPHI - Helper for Rewrite. PHI nodes are special because the use 4435/// of their operands effectively happens in their predecessor blocks, so the 4436/// expression may need to be expanded in multiple places. 4437void LSRInstance::RewriteForPHI(PHINode *PN, 4438 const LSRFixup &LF, 4439 const Formula &F, 4440 SCEVExpander &Rewriter, 4441 SmallVectorImpl<WeakVH> &DeadInsts, 4442 Pass *P) const { 4443 DenseMap<BasicBlock *, Value *> Inserted; 4444 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) 4445 if (PN->getIncomingValue(i) == LF.OperandValToReplace) { 4446 BasicBlock *BB = PN->getIncomingBlock(i); 4447 4448 // If this is a critical edge, split the edge so that we do not insert 4449 // the code on all predecessor/successor paths. We do this unless this 4450 // is the canonical backedge for this loop, which complicates post-inc 4451 // users. 4452 if (e != 1 && BB->getTerminator()->getNumSuccessors() > 1 && 4453 !isa<IndirectBrInst>(BB->getTerminator())) { 4454 BasicBlock *Parent = PN->getParent(); 4455 Loop *PNLoop = LI.getLoopFor(Parent); 4456 if (!PNLoop || Parent != PNLoop->getHeader()) { 4457 // Split the critical edge. 4458 BasicBlock *NewBB = 0; 4459 if (!Parent->isLandingPad()) { 4460 NewBB = SplitCriticalEdge(BB, Parent, P, 4461 /*MergeIdenticalEdges=*/true, 4462 /*DontDeleteUselessPhis=*/true); 4463 } else { 4464 SmallVector<BasicBlock*, 2> NewBBs; 4465 SplitLandingPadPredecessors(Parent, BB, "", "", P, NewBBs); 4466 NewBB = NewBBs[0]; 4467 } 4468 // If NewBB==NULL, then SplitCriticalEdge refused to split because all 4469 // phi predecessors are identical. The simple thing to do is skip 4470 // splitting in this case rather than complicate the API. 4471 if (NewBB) { 4472 // If PN is outside of the loop and BB is in the loop, we want to 4473 // move the block to be immediately before the PHI block, not 4474 // immediately after BB. 4475 if (L->contains(BB) && !L->contains(PN)) 4476 NewBB->moveBefore(PN->getParent()); 4477 4478 // Splitting the edge can reduce the number of PHI entries we have. 4479 e = PN->getNumIncomingValues(); 4480 BB = NewBB; 4481 i = PN->getBasicBlockIndex(BB); 4482 } 4483 } 4484 } 4485 4486 std::pair<DenseMap<BasicBlock *, Value *>::iterator, bool> Pair = 4487 Inserted.insert(std::make_pair(BB, static_cast<Value *>(0))); 4488 if (!Pair.second) 4489 PN->setIncomingValue(i, Pair.first->second); 4490 else { 4491 Value *FullV = Expand(LF, F, BB->getTerminator(), Rewriter, DeadInsts); 4492 4493 // If this is reuse-by-noop-cast, insert the noop cast. 4494 Type *OpTy = LF.OperandValToReplace->getType(); 4495 if (FullV->getType() != OpTy) 4496 FullV = 4497 CastInst::Create(CastInst::getCastOpcode(FullV, false, 4498 OpTy, false), 4499 FullV, LF.OperandValToReplace->getType(), 4500 "tmp", BB->getTerminator()); 4501 4502 PN->setIncomingValue(i, FullV); 4503 Pair.first->second = FullV; 4504 } 4505 } 4506} 4507 4508/// Rewrite - Emit instructions for the leading candidate expression for this 4509/// LSRUse (this is called "expanding"), and update the UserInst to reference 4510/// the newly expanded value. 4511void LSRInstance::Rewrite(const LSRFixup &LF, 4512 const Formula &F, 4513 SCEVExpander &Rewriter, 4514 SmallVectorImpl<WeakVH> &DeadInsts, 4515 Pass *P) const { 4516 // First, find an insertion point that dominates UserInst. For PHI nodes, 4517 // find the nearest block which dominates all the relevant uses. 4518 if (PHINode *PN = dyn_cast<PHINode>(LF.UserInst)) { 4519 RewriteForPHI(PN, LF, F, Rewriter, DeadInsts, P); 4520 } else { 4521 Value *FullV = Expand(LF, F, LF.UserInst, Rewriter, DeadInsts); 4522 4523 // If this is reuse-by-noop-cast, insert the noop cast. 4524 Type *OpTy = LF.OperandValToReplace->getType(); 4525 if (FullV->getType() != OpTy) { 4526 Instruction *Cast = 4527 CastInst::Create(CastInst::getCastOpcode(FullV, false, OpTy, false), 4528 FullV, OpTy, "tmp", LF.UserInst); 4529 FullV = Cast; 4530 } 4531 4532 // Update the user. ICmpZero is handled specially here (for now) because 4533 // Expand may have updated one of the operands of the icmp already, and 4534 // its new value may happen to be equal to LF.OperandValToReplace, in 4535 // which case doing replaceUsesOfWith leads to replacing both operands 4536 // with the same value. TODO: Reorganize this. 4537 if (Uses[LF.LUIdx].Kind == LSRUse::ICmpZero) 4538 LF.UserInst->setOperand(0, FullV); 4539 else 4540 LF.UserInst->replaceUsesOfWith(LF.OperandValToReplace, FullV); 4541 } 4542 4543 DeadInsts.push_back(LF.OperandValToReplace); 4544} 4545 4546/// ImplementSolution - Rewrite all the fixup locations with new values, 4547/// following the chosen solution. 4548void 4549LSRInstance::ImplementSolution(const SmallVectorImpl<const Formula *> &Solution, 4550 Pass *P) { 4551 // Keep track of instructions we may have made dead, so that 4552 // we can remove them after we are done working. 4553 SmallVector<WeakVH, 16> DeadInsts; 4554 4555 SCEVExpander Rewriter(SE, "lsr"); 4556#ifndef NDEBUG 4557 Rewriter.setDebugType(DEBUG_TYPE); 4558#endif 4559 Rewriter.disableCanonicalMode(); 4560 Rewriter.enableLSRMode(); 4561 Rewriter.setIVIncInsertPos(L, IVIncInsertPos); 4562 4563 // Mark phi nodes that terminate chains so the expander tries to reuse them. 4564 for (SmallVectorImpl<IVChain>::const_iterator ChainI = IVChainVec.begin(), 4565 ChainE = IVChainVec.end(); ChainI != ChainE; ++ChainI) { 4566 if (PHINode *PN = dyn_cast<PHINode>(ChainI->tailUserInst())) 4567 Rewriter.setChainedPhi(PN); 4568 } 4569 4570 // Expand the new value definitions and update the users. 4571 for (SmallVectorImpl<LSRFixup>::const_iterator I = Fixups.begin(), 4572 E = Fixups.end(); I != E; ++I) { 4573 const LSRFixup &Fixup = *I; 4574 4575 Rewrite(Fixup, *Solution[Fixup.LUIdx], Rewriter, DeadInsts, P); 4576 4577 Changed = true; 4578 } 4579 4580 for (SmallVectorImpl<IVChain>::const_iterator ChainI = IVChainVec.begin(), 4581 ChainE = IVChainVec.end(); ChainI != ChainE; ++ChainI) { 4582 GenerateIVChain(*ChainI, Rewriter, DeadInsts); 4583 Changed = true; 4584 } 4585 // Clean up after ourselves. This must be done before deleting any 4586 // instructions. 4587 Rewriter.clear(); 4588 4589 Changed |= DeleteTriviallyDeadInstructions(DeadInsts); 4590} 4591 4592LSRInstance::LSRInstance(const TargetLowering *tli, Loop *l, Pass *P) 4593 : IU(P->getAnalysis<IVUsers>()), 4594 SE(P->getAnalysis<ScalarEvolution>()), 4595 DT(P->getAnalysis<DominatorTree>()), 4596 LI(P->getAnalysis<LoopInfo>()), 4597 TLI(tli), L(l), Changed(false), IVIncInsertPos(0) { 4598 4599 // If LoopSimplify form is not available, stay out of trouble. 4600 if (!L->isLoopSimplifyForm()) 4601 return; 4602 4603 // If there's no interesting work to be done, bail early. 4604 if (IU.empty()) return; 4605 4606 // If there's too much analysis to be done, bail early. We won't be able to 4607 // model the problem anyway. 4608 unsigned NumUsers = 0; 4609 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI) { 4610 if (++NumUsers > MaxIVUsers) { 4611 DEBUG(dbgs() << "LSR skipping loop, too many IV Users in " << *L 4612 << "\n"); 4613 return; 4614 } 4615 } 4616 4617#ifndef NDEBUG 4618 // All dominating loops must have preheaders, or SCEVExpander may not be able 4619 // to materialize an AddRecExpr whose Start is an outer AddRecExpr. 4620 // 4621 // IVUsers analysis should only create users that are dominated by simple loop 4622 // headers. Since this loop should dominate all of its users, its user list 4623 // should be empty if this loop itself is not within a simple loop nest. 4624 for (DomTreeNode *Rung = DT.getNode(L->getLoopPreheader()); 4625 Rung; Rung = Rung->getIDom()) { 4626 BasicBlock *BB = Rung->getBlock(); 4627 const Loop *DomLoop = LI.getLoopFor(BB); 4628 if (DomLoop && DomLoop->getHeader() == BB) { 4629 assert(DomLoop->getLoopPreheader() && "LSR needs a simplified loop nest"); 4630 } 4631 } 4632#endif // DEBUG 4633 4634 DEBUG(dbgs() << "\nLSR on loop "; 4635 WriteAsOperand(dbgs(), L->getHeader(), /*PrintType=*/false); 4636 dbgs() << ":\n"); 4637 4638 // First, perform some low-level loop optimizations. 4639 OptimizeShadowIV(); 4640 OptimizeLoopTermCond(); 4641 4642 // If loop preparation eliminates all interesting IV users, bail. 4643 if (IU.empty()) return; 4644 4645 // Skip nested loops until we can model them better with formulae. 4646 if (!L->empty()) { 4647 DEBUG(dbgs() << "LSR skipping outer loop " << *L << "\n"); 4648 return; 4649 } 4650 4651 // Start collecting data and preparing for the solver. 4652 CollectChains(); 4653 CollectInterestingTypesAndFactors(); 4654 CollectFixupsAndInitialFormulae(); 4655 CollectLoopInvariantFixupsAndFormulae(); 4656 4657 assert(!Uses.empty() && "IVUsers reported at least one use"); 4658 DEBUG(dbgs() << "LSR found " << Uses.size() << " uses:\n"; 4659 print_uses(dbgs())); 4660 4661 // Now use the reuse data to generate a bunch of interesting ways 4662 // to formulate the values needed for the uses. 4663 GenerateAllReuseFormulae(); 4664 4665 FilterOutUndesirableDedicatedRegisters(); 4666 NarrowSearchSpaceUsingHeuristics(); 4667 4668 SmallVector<const Formula *, 8> Solution; 4669 Solve(Solution); 4670 4671 // Release memory that is no longer needed. 4672 Factors.clear(); 4673 Types.clear(); 4674 RegUses.clear(); 4675 4676 if (Solution.empty()) 4677 return; 4678 4679#ifndef NDEBUG 4680 // Formulae should be legal. 4681 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(), 4682 E = Uses.end(); I != E; ++I) { 4683 const LSRUse &LU = *I; 4684 for (SmallVectorImpl<Formula>::const_iterator J = LU.Formulae.begin(), 4685 JE = LU.Formulae.end(); J != JE; ++J) 4686 assert(isLegalUse(J->AM, LU.MinOffset, LU.MaxOffset, 4687 LU.Kind, LU.AccessTy, TLI) && 4688 "Illegal formula generated!"); 4689 }; 4690#endif 4691 4692 // Now that we've decided what we want, make it so. 4693 ImplementSolution(Solution, P); 4694} 4695 4696void LSRInstance::print_factors_and_types(raw_ostream &OS) const { 4697 if (Factors.empty() && Types.empty()) return; 4698 4699 OS << "LSR has identified the following interesting factors and types: "; 4700 bool First = true; 4701 4702 for (SmallSetVector<int64_t, 8>::const_iterator 4703 I = Factors.begin(), E = Factors.end(); I != E; ++I) { 4704 if (!First) OS << ", "; 4705 First = false; 4706 OS << '*' << *I; 4707 } 4708 4709 for (SmallSetVector<Type *, 4>::const_iterator 4710 I = Types.begin(), E = Types.end(); I != E; ++I) { 4711 if (!First) OS << ", "; 4712 First = false; 4713 OS << '(' << **I << ')'; 4714 } 4715 OS << '\n'; 4716} 4717 4718void LSRInstance::print_fixups(raw_ostream &OS) const { 4719 OS << "LSR is examining the following fixup sites:\n"; 4720 for (SmallVectorImpl<LSRFixup>::const_iterator I = Fixups.begin(), 4721 E = Fixups.end(); I != E; ++I) { 4722 dbgs() << " "; 4723 I->print(OS); 4724 OS << '\n'; 4725 } 4726} 4727 4728void LSRInstance::print_uses(raw_ostream &OS) const { 4729 OS << "LSR is examining the following uses:\n"; 4730 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(), 4731 E = Uses.end(); I != E; ++I) { 4732 const LSRUse &LU = *I; 4733 dbgs() << " "; 4734 LU.print(OS); 4735 OS << '\n'; 4736 for (SmallVectorImpl<Formula>::const_iterator J = LU.Formulae.begin(), 4737 JE = LU.Formulae.end(); J != JE; ++J) { 4738 OS << " "; 4739 J->print(OS); 4740 OS << '\n'; 4741 } 4742 } 4743} 4744 4745void LSRInstance::print(raw_ostream &OS) const { 4746 print_factors_and_types(OS); 4747 print_fixups(OS); 4748 print_uses(OS); 4749} 4750 4751#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) 4752void LSRInstance::dump() const { 4753 print(errs()); errs() << '\n'; 4754} 4755#endif 4756 4757namespace { 4758 4759class LoopStrengthReduce : public LoopPass { 4760 /// TLI - Keep a pointer of a TargetLowering to consult for determining 4761 /// transformation profitability. 4762 const TargetLowering *const TLI; 4763 4764public: 4765 static char ID; // Pass ID, replacement for typeid 4766 explicit LoopStrengthReduce(const TargetLowering *tli = 0); 4767 4768private: 4769 bool runOnLoop(Loop *L, LPPassManager &LPM); 4770 void getAnalysisUsage(AnalysisUsage &AU) const; 4771}; 4772 4773} 4774 4775char LoopStrengthReduce::ID = 0; 4776INITIALIZE_PASS_BEGIN(LoopStrengthReduce, "loop-reduce", 4777 "Loop Strength Reduction", false, false) 4778INITIALIZE_PASS_DEPENDENCY(DominatorTree) 4779INITIALIZE_PASS_DEPENDENCY(ScalarEvolution) 4780INITIALIZE_PASS_DEPENDENCY(IVUsers) 4781INITIALIZE_PASS_DEPENDENCY(LoopInfo) 4782INITIALIZE_PASS_DEPENDENCY(LoopSimplify) 4783INITIALIZE_PASS_END(LoopStrengthReduce, "loop-reduce", 4784 "Loop Strength Reduction", false, false) 4785 4786 4787Pass *llvm::createLoopStrengthReducePass(const TargetLowering *TLI) { 4788 return new LoopStrengthReduce(TLI); 4789} 4790 4791LoopStrengthReduce::LoopStrengthReduce(const TargetLowering *tli) 4792 : LoopPass(ID), TLI(tli) { 4793 initializeLoopStrengthReducePass(*PassRegistry::getPassRegistry()); 4794 } 4795 4796void LoopStrengthReduce::getAnalysisUsage(AnalysisUsage &AU) const { 4797 // We split critical edges, so we change the CFG. However, we do update 4798 // many analyses if they are around. 4799 AU.addPreservedID(LoopSimplifyID); 4800 4801 AU.addRequired<LoopInfo>(); 4802 AU.addPreserved<LoopInfo>(); 4803 AU.addRequiredID(LoopSimplifyID); 4804 AU.addRequired<DominatorTree>(); 4805 AU.addPreserved<DominatorTree>(); 4806 AU.addRequired<ScalarEvolution>(); 4807 AU.addPreserved<ScalarEvolution>(); 4808 // Requiring LoopSimplify a second time here prevents IVUsers from running 4809 // twice, since LoopSimplify was invalidated by running ScalarEvolution. 4810 AU.addRequiredID(LoopSimplifyID); 4811 AU.addRequired<IVUsers>(); 4812 AU.addPreserved<IVUsers>(); 4813} 4814 4815bool LoopStrengthReduce::runOnLoop(Loop *L, LPPassManager & /*LPM*/) { 4816 bool Changed = false; 4817 4818 // Run the main LSR transformation. 4819 Changed |= LSRInstance(TLI, L, this).getChanged(); 4820 4821 // Remove any extra phis created by processing inner loops. 4822 Changed |= DeleteDeadPHIs(L->getHeader()); 4823 if (EnablePhiElim) { 4824 SmallVector<WeakVH, 16> DeadInsts; 4825 SCEVExpander Rewriter(getAnalysis<ScalarEvolution>(), "lsr"); 4826#ifndef NDEBUG 4827 Rewriter.setDebugType(DEBUG_TYPE); 4828#endif 4829 unsigned numFolded = Rewriter. 4830 replaceCongruentIVs(L, &getAnalysis<DominatorTree>(), DeadInsts, TLI); 4831 if (numFolded) { 4832 Changed = true; 4833 DeleteTriviallyDeadInstructions(DeadInsts); 4834 DeleteDeadPHIs(L->getHeader()); 4835 } 4836 } 4837 return Changed; 4838} 4839