LoopStrengthReduce.cpp revision 8bf295b1bf345dda7f3f5cd12b5c0dafea283e81
1//===- LoopStrengthReduce.cpp - Strength Reduce IVs in Loops --------------===// 2// 3// The LLVM Compiler Infrastructure 4// 5// This file is distributed under the University of Illinois Open Source 6// License. See LICENSE.TXT for details. 7// 8//===----------------------------------------------------------------------===// 9// 10// This transformation analyzes and transforms the induction variables (and 11// computations derived from them) into forms suitable for efficient execution 12// on the target. 13// 14// This pass performs a strength reduction on array references inside loops that 15// have as one or more of their components the loop induction variable, it 16// rewrites expressions to take advantage of scaled-index addressing modes 17// available on the target, and it performs a variety of other optimizations 18// related to loop induction variables. 19// 20// Terminology note: this code has a lot of handling for "post-increment" or 21// "post-inc" users. This is not talking about post-increment addressing modes; 22// it is instead talking about code like this: 23// 24// %i = phi [ 0, %entry ], [ %i.next, %latch ] 25// ... 26// %i.next = add %i, 1 27// %c = icmp eq %i.next, %n 28// 29// The SCEV for %i is {0,+,1}<%L>. The SCEV for %i.next is {1,+,1}<%L>, however 30// it's useful to think about these as the same register, with some uses using 31// the value of the register before the add and some using // it after. In this 32// example, the icmp is a post-increment user, since it uses %i.next, which is 33// the value of the induction variable after the increment. The other common 34// case of post-increment users is users outside the loop. 35// 36// TODO: More sophistication in the way Formulae are generated and filtered. 37// 38// TODO: Handle multiple loops at a time. 39// 40// TODO: Should TargetLowering::AddrMode::BaseGV be changed to a ConstantExpr 41// instead of a GlobalValue? 42// 43// TODO: When truncation is free, truncate ICmp users' operands to make it a 44// smaller encoding (on x86 at least). 45// 46// TODO: When a negated register is used by an add (such as in a list of 47// multiple base registers, or as the increment expression in an addrec), 48// we may not actually need both reg and (-1 * reg) in registers; the 49// negation can be implemented by using a sub instead of an add. The 50// lack of support for taking this into consideration when making 51// register pressure decisions is partly worked around by the "Special" 52// use kind. 53// 54//===----------------------------------------------------------------------===// 55 56#define DEBUG_TYPE "loop-reduce" 57#include "llvm/Transforms/Scalar.h" 58#include "llvm/Constants.h" 59#include "llvm/Instructions.h" 60#include "llvm/IntrinsicInst.h" 61#include "llvm/DerivedTypes.h" 62#include "llvm/Analysis/IVUsers.h" 63#include "llvm/Analysis/Dominators.h" 64#include "llvm/Analysis/LoopPass.h" 65#include "llvm/Analysis/ScalarEvolutionExpander.h" 66#include "llvm/Assembly/Writer.h" 67#include "llvm/Transforms/Utils/BasicBlockUtils.h" 68#include "llvm/Transforms/Utils/Local.h" 69#include "llvm/ADT/SmallBitVector.h" 70#include "llvm/ADT/SetVector.h" 71#include "llvm/ADT/DenseSet.h" 72#include "llvm/Support/Debug.h" 73#include "llvm/Support/CommandLine.h" 74#include "llvm/Support/ValueHandle.h" 75#include "llvm/Support/raw_ostream.h" 76#include "llvm/Target/TargetLowering.h" 77#include <algorithm> 78using namespace llvm; 79 80static cl::opt<bool> EnableNested( 81 "enable-lsr-nested", cl::Hidden, cl::desc("Enable LSR on nested loops")); 82 83static cl::opt<bool> EnableRetry( 84 "enable-lsr-retry", cl::Hidden, cl::desc("Enable LSR retry")); 85 86// Temporary flag to cleanup congruent phis after LSR phi expansion. 87// It's currently disabled until we can determine whether it's truly useful or 88// not. The flag should be removed after the v3.0 release. 89// This is now needed for ivchains. 90static cl::opt<bool> EnablePhiElim( 91 "enable-lsr-phielim", cl::Hidden, cl::init(true), 92 cl::desc("Enable LSR phi elimination")); 93 94namespace { 95 96/// RegSortData - This class holds data which is used to order reuse candidates. 97class RegSortData { 98public: 99 /// UsedByIndices - This represents the set of LSRUse indices which reference 100 /// a particular register. 101 SmallBitVector UsedByIndices; 102 103 RegSortData() {} 104 105 void print(raw_ostream &OS) const; 106 void dump() const; 107}; 108 109} 110 111void RegSortData::print(raw_ostream &OS) const { 112 OS << "[NumUses=" << UsedByIndices.count() << ']'; 113} 114 115void RegSortData::dump() const { 116 print(errs()); errs() << '\n'; 117} 118 119namespace { 120 121/// RegUseTracker - Map register candidates to information about how they are 122/// used. 123class RegUseTracker { 124 typedef DenseMap<const SCEV *, RegSortData> RegUsesTy; 125 126 RegUsesTy RegUsesMap; 127 SmallVector<const SCEV *, 16> RegSequence; 128 129public: 130 void CountRegister(const SCEV *Reg, size_t LUIdx); 131 void DropRegister(const SCEV *Reg, size_t LUIdx); 132 void SwapAndDropUse(size_t LUIdx, size_t LastLUIdx); 133 134 bool isRegUsedByUsesOtherThan(const SCEV *Reg, size_t LUIdx) const; 135 136 const SmallBitVector &getUsedByIndices(const SCEV *Reg) const; 137 138 void clear(); 139 140 typedef SmallVectorImpl<const SCEV *>::iterator iterator; 141 typedef SmallVectorImpl<const SCEV *>::const_iterator const_iterator; 142 iterator begin() { return RegSequence.begin(); } 143 iterator end() { return RegSequence.end(); } 144 const_iterator begin() const { return RegSequence.begin(); } 145 const_iterator end() const { return RegSequence.end(); } 146}; 147 148} 149 150void 151RegUseTracker::CountRegister(const SCEV *Reg, size_t LUIdx) { 152 std::pair<RegUsesTy::iterator, bool> Pair = 153 RegUsesMap.insert(std::make_pair(Reg, RegSortData())); 154 RegSortData &RSD = Pair.first->second; 155 if (Pair.second) 156 RegSequence.push_back(Reg); 157 RSD.UsedByIndices.resize(std::max(RSD.UsedByIndices.size(), LUIdx + 1)); 158 RSD.UsedByIndices.set(LUIdx); 159} 160 161void 162RegUseTracker::DropRegister(const SCEV *Reg, size_t LUIdx) { 163 RegUsesTy::iterator It = RegUsesMap.find(Reg); 164 assert(It != RegUsesMap.end()); 165 RegSortData &RSD = It->second; 166 assert(RSD.UsedByIndices.size() > LUIdx); 167 RSD.UsedByIndices.reset(LUIdx); 168} 169 170void 171RegUseTracker::SwapAndDropUse(size_t LUIdx, size_t LastLUIdx) { 172 assert(LUIdx <= LastLUIdx); 173 174 // Update RegUses. The data structure is not optimized for this purpose; 175 // we must iterate through it and update each of the bit vectors. 176 for (RegUsesTy::iterator I = RegUsesMap.begin(), E = RegUsesMap.end(); 177 I != E; ++I) { 178 SmallBitVector &UsedByIndices = I->second.UsedByIndices; 179 if (LUIdx < UsedByIndices.size()) 180 UsedByIndices[LUIdx] = 181 LastLUIdx < UsedByIndices.size() ? UsedByIndices[LastLUIdx] : 0; 182 UsedByIndices.resize(std::min(UsedByIndices.size(), LastLUIdx)); 183 } 184} 185 186bool 187RegUseTracker::isRegUsedByUsesOtherThan(const SCEV *Reg, size_t LUIdx) const { 188 RegUsesTy::const_iterator I = RegUsesMap.find(Reg); 189 if (I == RegUsesMap.end()) 190 return false; 191 const SmallBitVector &UsedByIndices = I->second.UsedByIndices; 192 int i = UsedByIndices.find_first(); 193 if (i == -1) return false; 194 if ((size_t)i != LUIdx) return true; 195 return UsedByIndices.find_next(i) != -1; 196} 197 198const SmallBitVector &RegUseTracker::getUsedByIndices(const SCEV *Reg) const { 199 RegUsesTy::const_iterator I = RegUsesMap.find(Reg); 200 assert(I != RegUsesMap.end() && "Unknown register!"); 201 return I->second.UsedByIndices; 202} 203 204void RegUseTracker::clear() { 205 RegUsesMap.clear(); 206 RegSequence.clear(); 207} 208 209namespace { 210 211/// Formula - This class holds information that describes a formula for 212/// computing satisfying a use. It may include broken-out immediates and scaled 213/// registers. 214struct Formula { 215 /// AM - This is used to represent complex addressing, as well as other kinds 216 /// of interesting uses. 217 TargetLowering::AddrMode AM; 218 219 /// BaseRegs - The list of "base" registers for this use. When this is 220 /// non-empty, AM.HasBaseReg should be set to true. 221 SmallVector<const SCEV *, 2> BaseRegs; 222 223 /// ScaledReg - The 'scaled' register for this use. This should be non-null 224 /// when AM.Scale is not zero. 225 const SCEV *ScaledReg; 226 227 /// UnfoldedOffset - An additional constant offset which added near the 228 /// use. This requires a temporary register, but the offset itself can 229 /// live in an add immediate field rather than a register. 230 int64_t UnfoldedOffset; 231 232 Formula() : ScaledReg(0), UnfoldedOffset(0) {} 233 234 void InitialMatch(const SCEV *S, Loop *L, ScalarEvolution &SE); 235 236 unsigned getNumRegs() const; 237 Type *getType() const; 238 239 void DeleteBaseReg(const SCEV *&S); 240 241 bool referencesReg(const SCEV *S) const; 242 bool hasRegsUsedByUsesOtherThan(size_t LUIdx, 243 const RegUseTracker &RegUses) const; 244 245 void print(raw_ostream &OS) const; 246 void dump() const; 247}; 248 249} 250 251/// DoInitialMatch - Recursion helper for InitialMatch. 252static void DoInitialMatch(const SCEV *S, Loop *L, 253 SmallVectorImpl<const SCEV *> &Good, 254 SmallVectorImpl<const SCEV *> &Bad, 255 ScalarEvolution &SE) { 256 // Collect expressions which properly dominate the loop header. 257 if (SE.properlyDominates(S, L->getHeader())) { 258 Good.push_back(S); 259 return; 260 } 261 262 // Look at add operands. 263 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) { 264 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end(); 265 I != E; ++I) 266 DoInitialMatch(*I, L, Good, Bad, SE); 267 return; 268 } 269 270 // Look at addrec operands. 271 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) 272 if (!AR->getStart()->isZero()) { 273 DoInitialMatch(AR->getStart(), L, Good, Bad, SE); 274 DoInitialMatch(SE.getAddRecExpr(SE.getConstant(AR->getType(), 0), 275 AR->getStepRecurrence(SE), 276 // FIXME: AR->getNoWrapFlags() 277 AR->getLoop(), SCEV::FlagAnyWrap), 278 L, Good, Bad, SE); 279 return; 280 } 281 282 // Handle a multiplication by -1 (negation) if it didn't fold. 283 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) 284 if (Mul->getOperand(0)->isAllOnesValue()) { 285 SmallVector<const SCEV *, 4> Ops(Mul->op_begin()+1, Mul->op_end()); 286 const SCEV *NewMul = SE.getMulExpr(Ops); 287 288 SmallVector<const SCEV *, 4> MyGood; 289 SmallVector<const SCEV *, 4> MyBad; 290 DoInitialMatch(NewMul, L, MyGood, MyBad, SE); 291 const SCEV *NegOne = SE.getSCEV(ConstantInt::getAllOnesValue( 292 SE.getEffectiveSCEVType(NewMul->getType()))); 293 for (SmallVectorImpl<const SCEV *>::const_iterator I = MyGood.begin(), 294 E = MyGood.end(); I != E; ++I) 295 Good.push_back(SE.getMulExpr(NegOne, *I)); 296 for (SmallVectorImpl<const SCEV *>::const_iterator I = MyBad.begin(), 297 E = MyBad.end(); I != E; ++I) 298 Bad.push_back(SE.getMulExpr(NegOne, *I)); 299 return; 300 } 301 302 // Ok, we can't do anything interesting. Just stuff the whole thing into a 303 // register and hope for the best. 304 Bad.push_back(S); 305} 306 307/// InitialMatch - Incorporate loop-variant parts of S into this Formula, 308/// attempting to keep all loop-invariant and loop-computable values in a 309/// single base register. 310void Formula::InitialMatch(const SCEV *S, Loop *L, ScalarEvolution &SE) { 311 SmallVector<const SCEV *, 4> Good; 312 SmallVector<const SCEV *, 4> Bad; 313 DoInitialMatch(S, L, Good, Bad, SE); 314 if (!Good.empty()) { 315 const SCEV *Sum = SE.getAddExpr(Good); 316 if (!Sum->isZero()) 317 BaseRegs.push_back(Sum); 318 AM.HasBaseReg = true; 319 } 320 if (!Bad.empty()) { 321 const SCEV *Sum = SE.getAddExpr(Bad); 322 if (!Sum->isZero()) 323 BaseRegs.push_back(Sum); 324 AM.HasBaseReg = true; 325 } 326} 327 328/// getNumRegs - Return the total number of register operands used by this 329/// formula. This does not include register uses implied by non-constant 330/// addrec strides. 331unsigned Formula::getNumRegs() const { 332 return !!ScaledReg + BaseRegs.size(); 333} 334 335/// getType - Return the type of this formula, if it has one, or null 336/// otherwise. This type is meaningless except for the bit size. 337Type *Formula::getType() const { 338 return !BaseRegs.empty() ? BaseRegs.front()->getType() : 339 ScaledReg ? ScaledReg->getType() : 340 AM.BaseGV ? AM.BaseGV->getType() : 341 0; 342} 343 344/// DeleteBaseReg - Delete the given base reg from the BaseRegs list. 345void Formula::DeleteBaseReg(const SCEV *&S) { 346 if (&S != &BaseRegs.back()) 347 std::swap(S, BaseRegs.back()); 348 BaseRegs.pop_back(); 349} 350 351/// referencesReg - Test if this formula references the given register. 352bool Formula::referencesReg(const SCEV *S) const { 353 return S == ScaledReg || 354 std::find(BaseRegs.begin(), BaseRegs.end(), S) != BaseRegs.end(); 355} 356 357/// hasRegsUsedByUsesOtherThan - Test whether this formula uses registers 358/// which are used by uses other than the use with the given index. 359bool Formula::hasRegsUsedByUsesOtherThan(size_t LUIdx, 360 const RegUseTracker &RegUses) const { 361 if (ScaledReg) 362 if (RegUses.isRegUsedByUsesOtherThan(ScaledReg, LUIdx)) 363 return true; 364 for (SmallVectorImpl<const SCEV *>::const_iterator I = BaseRegs.begin(), 365 E = BaseRegs.end(); I != E; ++I) 366 if (RegUses.isRegUsedByUsesOtherThan(*I, LUIdx)) 367 return true; 368 return false; 369} 370 371void Formula::print(raw_ostream &OS) const { 372 bool First = true; 373 if (AM.BaseGV) { 374 if (!First) OS << " + "; else First = false; 375 WriteAsOperand(OS, AM.BaseGV, /*PrintType=*/false); 376 } 377 if (AM.BaseOffs != 0) { 378 if (!First) OS << " + "; else First = false; 379 OS << AM.BaseOffs; 380 } 381 for (SmallVectorImpl<const SCEV *>::const_iterator I = BaseRegs.begin(), 382 E = BaseRegs.end(); I != E; ++I) { 383 if (!First) OS << " + "; else First = false; 384 OS << "reg(" << **I << ')'; 385 } 386 if (AM.HasBaseReg && BaseRegs.empty()) { 387 if (!First) OS << " + "; else First = false; 388 OS << "**error: HasBaseReg**"; 389 } else if (!AM.HasBaseReg && !BaseRegs.empty()) { 390 if (!First) OS << " + "; else First = false; 391 OS << "**error: !HasBaseReg**"; 392 } 393 if (AM.Scale != 0) { 394 if (!First) OS << " + "; else First = false; 395 OS << AM.Scale << "*reg("; 396 if (ScaledReg) 397 OS << *ScaledReg; 398 else 399 OS << "<unknown>"; 400 OS << ')'; 401 } 402 if (UnfoldedOffset != 0) { 403 if (!First) OS << " + "; else First = false; 404 OS << "imm(" << UnfoldedOffset << ')'; 405 } 406} 407 408void Formula::dump() const { 409 print(errs()); errs() << '\n'; 410} 411 412/// isAddRecSExtable - Return true if the given addrec can be sign-extended 413/// without changing its value. 414static bool isAddRecSExtable(const SCEVAddRecExpr *AR, ScalarEvolution &SE) { 415 Type *WideTy = 416 IntegerType::get(SE.getContext(), SE.getTypeSizeInBits(AR->getType()) + 1); 417 return isa<SCEVAddRecExpr>(SE.getSignExtendExpr(AR, WideTy)); 418} 419 420/// isAddSExtable - Return true if the given add can be sign-extended 421/// without changing its value. 422static bool isAddSExtable(const SCEVAddExpr *A, ScalarEvolution &SE) { 423 Type *WideTy = 424 IntegerType::get(SE.getContext(), SE.getTypeSizeInBits(A->getType()) + 1); 425 return isa<SCEVAddExpr>(SE.getSignExtendExpr(A, WideTy)); 426} 427 428/// isMulSExtable - Return true if the given mul can be sign-extended 429/// without changing its value. 430static bool isMulSExtable(const SCEVMulExpr *M, ScalarEvolution &SE) { 431 Type *WideTy = 432 IntegerType::get(SE.getContext(), 433 SE.getTypeSizeInBits(M->getType()) * M->getNumOperands()); 434 return isa<SCEVMulExpr>(SE.getSignExtendExpr(M, WideTy)); 435} 436 437/// getExactSDiv - Return an expression for LHS /s RHS, if it can be determined 438/// and if the remainder is known to be zero, or null otherwise. If 439/// IgnoreSignificantBits is true, expressions like (X * Y) /s Y are simplified 440/// to Y, ignoring that the multiplication may overflow, which is useful when 441/// the result will be used in a context where the most significant bits are 442/// ignored. 443static const SCEV *getExactSDiv(const SCEV *LHS, const SCEV *RHS, 444 ScalarEvolution &SE, 445 bool IgnoreSignificantBits = false) { 446 // Handle the trivial case, which works for any SCEV type. 447 if (LHS == RHS) 448 return SE.getConstant(LHS->getType(), 1); 449 450 // Handle a few RHS special cases. 451 const SCEVConstant *RC = dyn_cast<SCEVConstant>(RHS); 452 if (RC) { 453 const APInt &RA = RC->getValue()->getValue(); 454 // Handle x /s -1 as x * -1, to give ScalarEvolution a chance to do 455 // some folding. 456 if (RA.isAllOnesValue()) 457 return SE.getMulExpr(LHS, RC); 458 // Handle x /s 1 as x. 459 if (RA == 1) 460 return LHS; 461 } 462 463 // Check for a division of a constant by a constant. 464 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(LHS)) { 465 if (!RC) 466 return 0; 467 const APInt &LA = C->getValue()->getValue(); 468 const APInt &RA = RC->getValue()->getValue(); 469 if (LA.srem(RA) != 0) 470 return 0; 471 return SE.getConstant(LA.sdiv(RA)); 472 } 473 474 // Distribute the sdiv over addrec operands, if the addrec doesn't overflow. 475 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(LHS)) { 476 if (IgnoreSignificantBits || isAddRecSExtable(AR, SE)) { 477 const SCEV *Step = getExactSDiv(AR->getStepRecurrence(SE), RHS, SE, 478 IgnoreSignificantBits); 479 if (!Step) return 0; 480 const SCEV *Start = getExactSDiv(AR->getStart(), RHS, SE, 481 IgnoreSignificantBits); 482 if (!Start) return 0; 483 // FlagNW is independent of the start value, step direction, and is 484 // preserved with smaller magnitude steps. 485 // FIXME: AR->getNoWrapFlags(SCEV::FlagNW) 486 return SE.getAddRecExpr(Start, Step, AR->getLoop(), SCEV::FlagAnyWrap); 487 } 488 return 0; 489 } 490 491 // Distribute the sdiv over add operands, if the add doesn't overflow. 492 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(LHS)) { 493 if (IgnoreSignificantBits || isAddSExtable(Add, SE)) { 494 SmallVector<const SCEV *, 8> Ops; 495 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end(); 496 I != E; ++I) { 497 const SCEV *Op = getExactSDiv(*I, RHS, SE, 498 IgnoreSignificantBits); 499 if (!Op) return 0; 500 Ops.push_back(Op); 501 } 502 return SE.getAddExpr(Ops); 503 } 504 return 0; 505 } 506 507 // Check for a multiply operand that we can pull RHS out of. 508 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(LHS)) { 509 if (IgnoreSignificantBits || isMulSExtable(Mul, SE)) { 510 SmallVector<const SCEV *, 4> Ops; 511 bool Found = false; 512 for (SCEVMulExpr::op_iterator I = Mul->op_begin(), E = Mul->op_end(); 513 I != E; ++I) { 514 const SCEV *S = *I; 515 if (!Found) 516 if (const SCEV *Q = getExactSDiv(S, RHS, SE, 517 IgnoreSignificantBits)) { 518 S = Q; 519 Found = true; 520 } 521 Ops.push_back(S); 522 } 523 return Found ? SE.getMulExpr(Ops) : 0; 524 } 525 return 0; 526 } 527 528 // Otherwise we don't know. 529 return 0; 530} 531 532/// ExtractImmediate - If S involves the addition of a constant integer value, 533/// return that integer value, and mutate S to point to a new SCEV with that 534/// value excluded. 535static int64_t ExtractImmediate(const SCEV *&S, ScalarEvolution &SE) { 536 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S)) { 537 if (C->getValue()->getValue().getMinSignedBits() <= 64) { 538 S = SE.getConstant(C->getType(), 0); 539 return C->getValue()->getSExtValue(); 540 } 541 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) { 542 SmallVector<const SCEV *, 8> NewOps(Add->op_begin(), Add->op_end()); 543 int64_t Result = ExtractImmediate(NewOps.front(), SE); 544 if (Result != 0) 545 S = SE.getAddExpr(NewOps); 546 return Result; 547 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) { 548 SmallVector<const SCEV *, 8> NewOps(AR->op_begin(), AR->op_end()); 549 int64_t Result = ExtractImmediate(NewOps.front(), SE); 550 if (Result != 0) 551 S = SE.getAddRecExpr(NewOps, AR->getLoop(), 552 // FIXME: AR->getNoWrapFlags(SCEV::FlagNW) 553 SCEV::FlagAnyWrap); 554 return Result; 555 } 556 return 0; 557} 558 559/// ExtractSymbol - If S involves the addition of a GlobalValue address, 560/// return that symbol, and mutate S to point to a new SCEV with that 561/// value excluded. 562static GlobalValue *ExtractSymbol(const SCEV *&S, ScalarEvolution &SE) { 563 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) { 564 if (GlobalValue *GV = dyn_cast<GlobalValue>(U->getValue())) { 565 S = SE.getConstant(GV->getType(), 0); 566 return GV; 567 } 568 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) { 569 SmallVector<const SCEV *, 8> NewOps(Add->op_begin(), Add->op_end()); 570 GlobalValue *Result = ExtractSymbol(NewOps.back(), SE); 571 if (Result) 572 S = SE.getAddExpr(NewOps); 573 return Result; 574 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) { 575 SmallVector<const SCEV *, 8> NewOps(AR->op_begin(), AR->op_end()); 576 GlobalValue *Result = ExtractSymbol(NewOps.front(), SE); 577 if (Result) 578 S = SE.getAddRecExpr(NewOps, AR->getLoop(), 579 // FIXME: AR->getNoWrapFlags(SCEV::FlagNW) 580 SCEV::FlagAnyWrap); 581 return Result; 582 } 583 return 0; 584} 585 586/// isAddressUse - Returns true if the specified instruction is using the 587/// specified value as an address. 588static bool isAddressUse(Instruction *Inst, Value *OperandVal) { 589 bool isAddress = isa<LoadInst>(Inst); 590 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) { 591 if (SI->getOperand(1) == OperandVal) 592 isAddress = true; 593 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) { 594 // Addressing modes can also be folded into prefetches and a variety 595 // of intrinsics. 596 switch (II->getIntrinsicID()) { 597 default: break; 598 case Intrinsic::prefetch: 599 case Intrinsic::x86_sse_storeu_ps: 600 case Intrinsic::x86_sse2_storeu_pd: 601 case Intrinsic::x86_sse2_storeu_dq: 602 case Intrinsic::x86_sse2_storel_dq: 603 if (II->getArgOperand(0) == OperandVal) 604 isAddress = true; 605 break; 606 } 607 } 608 return isAddress; 609} 610 611/// getAccessType - Return the type of the memory being accessed. 612static Type *getAccessType(const Instruction *Inst) { 613 Type *AccessTy = Inst->getType(); 614 if (const StoreInst *SI = dyn_cast<StoreInst>(Inst)) 615 AccessTy = SI->getOperand(0)->getType(); 616 else if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) { 617 // Addressing modes can also be folded into prefetches and a variety 618 // of intrinsics. 619 switch (II->getIntrinsicID()) { 620 default: break; 621 case Intrinsic::x86_sse_storeu_ps: 622 case Intrinsic::x86_sse2_storeu_pd: 623 case Intrinsic::x86_sse2_storeu_dq: 624 case Intrinsic::x86_sse2_storel_dq: 625 AccessTy = II->getArgOperand(0)->getType(); 626 break; 627 } 628 } 629 630 // All pointers have the same requirements, so canonicalize them to an 631 // arbitrary pointer type to minimize variation. 632 if (PointerType *PTy = dyn_cast<PointerType>(AccessTy)) 633 AccessTy = PointerType::get(IntegerType::get(PTy->getContext(), 1), 634 PTy->getAddressSpace()); 635 636 return AccessTy; 637} 638 639/// isExistingPhi - Return true if this AddRec is already a phi in its loop. 640static bool isExistingPhi(const SCEVAddRecExpr *AR, ScalarEvolution &SE) { 641 for (BasicBlock::iterator I = AR->getLoop()->getHeader()->begin(); 642 PHINode *PN = dyn_cast<PHINode>(I); ++I) { 643 if (SE.isSCEVable(PN->getType()) && 644 (SE.getEffectiveSCEVType(PN->getType()) == 645 SE.getEffectiveSCEVType(AR->getType())) && 646 SE.getSCEV(PN) == AR) 647 return true; 648 } 649 return false; 650} 651 652/// DeleteTriviallyDeadInstructions - If any of the instructions is the 653/// specified set are trivially dead, delete them and see if this makes any of 654/// their operands subsequently dead. 655static bool 656DeleteTriviallyDeadInstructions(SmallVectorImpl<WeakVH> &DeadInsts) { 657 bool Changed = false; 658 659 while (!DeadInsts.empty()) { 660 Instruction *I = dyn_cast_or_null<Instruction>(&*DeadInsts.pop_back_val()); 661 662 if (I == 0 || !isInstructionTriviallyDead(I)) 663 continue; 664 665 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI) 666 if (Instruction *U = dyn_cast<Instruction>(*OI)) { 667 *OI = 0; 668 if (U->use_empty()) 669 DeadInsts.push_back(U); 670 } 671 672 I->eraseFromParent(); 673 Changed = true; 674 } 675 676 return Changed; 677} 678 679namespace { 680 681/// Cost - This class is used to measure and compare candidate formulae. 682class Cost { 683 /// TODO: Some of these could be merged. Also, a lexical ordering 684 /// isn't always optimal. 685 unsigned NumRegs; 686 unsigned AddRecCost; 687 unsigned NumIVMuls; 688 unsigned NumBaseAdds; 689 unsigned ImmCost; 690 unsigned SetupCost; 691 692public: 693 Cost() 694 : NumRegs(0), AddRecCost(0), NumIVMuls(0), NumBaseAdds(0), ImmCost(0), 695 SetupCost(0) {} 696 697 bool operator<(const Cost &Other) const; 698 699 void Loose(); 700 701#ifndef NDEBUG 702 // Once any of the metrics loses, they must all remain losers. 703 bool isValid() { 704 return ((NumRegs | AddRecCost | NumIVMuls | NumBaseAdds 705 | ImmCost | SetupCost) != ~0u) 706 || ((NumRegs & AddRecCost & NumIVMuls & NumBaseAdds 707 & ImmCost & SetupCost) == ~0u); 708 } 709#endif 710 711 bool isLoser() { 712 assert(isValid() && "invalid cost"); 713 return NumRegs == ~0u; 714 } 715 716 void RateFormula(const Formula &F, 717 SmallPtrSet<const SCEV *, 16> &Regs, 718 const DenseSet<const SCEV *> &VisitedRegs, 719 const Loop *L, 720 const SmallVectorImpl<int64_t> &Offsets, 721 ScalarEvolution &SE, DominatorTree &DT, 722 SmallPtrSet<const SCEV *, 16> *LoserRegs = 0); 723 724 void print(raw_ostream &OS) const; 725 void dump() const; 726 727private: 728 void RateRegister(const SCEV *Reg, 729 SmallPtrSet<const SCEV *, 16> &Regs, 730 const Loop *L, 731 ScalarEvolution &SE, DominatorTree &DT); 732 void RatePrimaryRegister(const SCEV *Reg, 733 SmallPtrSet<const SCEV *, 16> &Regs, 734 const Loop *L, 735 ScalarEvolution &SE, DominatorTree &DT, 736 SmallPtrSet<const SCEV *, 16> *LoserRegs); 737}; 738 739} 740 741/// RateRegister - Tally up interesting quantities from the given register. 742void Cost::RateRegister(const SCEV *Reg, 743 SmallPtrSet<const SCEV *, 16> &Regs, 744 const Loop *L, 745 ScalarEvolution &SE, DominatorTree &DT) { 746 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Reg)) { 747 if (AR->getLoop() == L) 748 AddRecCost += 1; /// TODO: This should be a function of the stride. 749 750 // If this is an addrec for another loop, don't second-guess its addrec phi 751 // nodes. LSR isn't currently smart enough to reason about more than one 752 // loop at a time. LSR has either already run on inner loops, will not run 753 // on other loops, and cannot be expected to change sibling loops. If the 754 // AddRec exists, consider it's register free and leave it alone. Otherwise, 755 // do not consider this formula at all. 756 else if (!EnableNested || L->contains(AR->getLoop()) || 757 (!AR->getLoop()->contains(L) && 758 DT.dominates(L->getHeader(), AR->getLoop()->getHeader()))) { 759 if (isExistingPhi(AR, SE)) 760 return; 761 762 // For !EnableNested, never rewrite IVs in other loops. 763 if (!EnableNested) { 764 Loose(); 765 return; 766 } 767 // If this isn't one of the addrecs that the loop already has, it 768 // would require a costly new phi and add. TODO: This isn't 769 // precisely modeled right now. 770 ++NumBaseAdds; 771 if (!Regs.count(AR->getStart())) { 772 RateRegister(AR->getStart(), Regs, L, SE, DT); 773 if (isLoser()) 774 return; 775 } 776 } 777 778 // Add the step value register, if it needs one. 779 // TODO: The non-affine case isn't precisely modeled here. 780 if (!AR->isAffine() || !isa<SCEVConstant>(AR->getOperand(1))) { 781 if (!Regs.count(AR->getOperand(1))) { 782 RateRegister(AR->getOperand(1), Regs, L, SE, DT); 783 if (isLoser()) 784 return; 785 } 786 } 787 } 788 ++NumRegs; 789 790 // Rough heuristic; favor registers which don't require extra setup 791 // instructions in the preheader. 792 if (!isa<SCEVUnknown>(Reg) && 793 !isa<SCEVConstant>(Reg) && 794 !(isa<SCEVAddRecExpr>(Reg) && 795 (isa<SCEVUnknown>(cast<SCEVAddRecExpr>(Reg)->getStart()) || 796 isa<SCEVConstant>(cast<SCEVAddRecExpr>(Reg)->getStart())))) 797 ++SetupCost; 798 799 NumIVMuls += isa<SCEVMulExpr>(Reg) && 800 SE.hasComputableLoopEvolution(Reg, L); 801} 802 803/// RatePrimaryRegister - Record this register in the set. If we haven't seen it 804/// before, rate it. Optional LoserRegs provides a way to declare any formula 805/// that refers to one of those regs an instant loser. 806void Cost::RatePrimaryRegister(const SCEV *Reg, 807 SmallPtrSet<const SCEV *, 16> &Regs, 808 const Loop *L, 809 ScalarEvolution &SE, DominatorTree &DT, 810 SmallPtrSet<const SCEV *, 16> *LoserRegs) { 811 if (LoserRegs && LoserRegs->count(Reg)) { 812 Loose(); 813 return; 814 } 815 if (Regs.insert(Reg)) { 816 RateRegister(Reg, Regs, L, SE, DT); 817 if (isLoser()) 818 LoserRegs->insert(Reg); 819 } 820} 821 822void Cost::RateFormula(const Formula &F, 823 SmallPtrSet<const SCEV *, 16> &Regs, 824 const DenseSet<const SCEV *> &VisitedRegs, 825 const Loop *L, 826 const SmallVectorImpl<int64_t> &Offsets, 827 ScalarEvolution &SE, DominatorTree &DT, 828 SmallPtrSet<const SCEV *, 16> *LoserRegs) { 829 // Tally up the registers. 830 if (const SCEV *ScaledReg = F.ScaledReg) { 831 if (VisitedRegs.count(ScaledReg)) { 832 Loose(); 833 return; 834 } 835 RatePrimaryRegister(ScaledReg, Regs, L, SE, DT, LoserRegs); 836 if (isLoser()) 837 return; 838 } 839 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(), 840 E = F.BaseRegs.end(); I != E; ++I) { 841 const SCEV *BaseReg = *I; 842 if (VisitedRegs.count(BaseReg)) { 843 Loose(); 844 return; 845 } 846 RatePrimaryRegister(BaseReg, Regs, L, SE, DT, LoserRegs); 847 if (isLoser()) 848 return; 849 } 850 851 // Determine how many (unfolded) adds we'll need inside the loop. 852 size_t NumBaseParts = F.BaseRegs.size() + (F.UnfoldedOffset != 0); 853 if (NumBaseParts > 1) 854 NumBaseAdds += NumBaseParts - 1; 855 856 // Tally up the non-zero immediates. 857 for (SmallVectorImpl<int64_t>::const_iterator I = Offsets.begin(), 858 E = Offsets.end(); I != E; ++I) { 859 int64_t Offset = (uint64_t)*I + F.AM.BaseOffs; 860 if (F.AM.BaseGV) 861 ImmCost += 64; // Handle symbolic values conservatively. 862 // TODO: This should probably be the pointer size. 863 else if (Offset != 0) 864 ImmCost += APInt(64, Offset, true).getMinSignedBits(); 865 } 866 assert(isValid() && "invalid cost"); 867} 868 869/// Loose - Set this cost to a losing value. 870void Cost::Loose() { 871 NumRegs = ~0u; 872 AddRecCost = ~0u; 873 NumIVMuls = ~0u; 874 NumBaseAdds = ~0u; 875 ImmCost = ~0u; 876 SetupCost = ~0u; 877} 878 879/// operator< - Choose the lower cost. 880bool Cost::operator<(const Cost &Other) const { 881 if (NumRegs != Other.NumRegs) 882 return NumRegs < Other.NumRegs; 883 if (AddRecCost != Other.AddRecCost) 884 return AddRecCost < Other.AddRecCost; 885 if (NumIVMuls != Other.NumIVMuls) 886 return NumIVMuls < Other.NumIVMuls; 887 if (NumBaseAdds != Other.NumBaseAdds) 888 return NumBaseAdds < Other.NumBaseAdds; 889 if (ImmCost != Other.ImmCost) 890 return ImmCost < Other.ImmCost; 891 if (SetupCost != Other.SetupCost) 892 return SetupCost < Other.SetupCost; 893 return false; 894} 895 896void Cost::print(raw_ostream &OS) const { 897 OS << NumRegs << " reg" << (NumRegs == 1 ? "" : "s"); 898 if (AddRecCost != 0) 899 OS << ", with addrec cost " << AddRecCost; 900 if (NumIVMuls != 0) 901 OS << ", plus " << NumIVMuls << " IV mul" << (NumIVMuls == 1 ? "" : "s"); 902 if (NumBaseAdds != 0) 903 OS << ", plus " << NumBaseAdds << " base add" 904 << (NumBaseAdds == 1 ? "" : "s"); 905 if (ImmCost != 0) 906 OS << ", plus " << ImmCost << " imm cost"; 907 if (SetupCost != 0) 908 OS << ", plus " << SetupCost << " setup cost"; 909} 910 911void Cost::dump() const { 912 print(errs()); errs() << '\n'; 913} 914 915namespace { 916 917/// LSRFixup - An operand value in an instruction which is to be replaced 918/// with some equivalent, possibly strength-reduced, replacement. 919struct LSRFixup { 920 /// UserInst - The instruction which will be updated. 921 Instruction *UserInst; 922 923 /// OperandValToReplace - The operand of the instruction which will 924 /// be replaced. The operand may be used more than once; every instance 925 /// will be replaced. 926 Value *OperandValToReplace; 927 928 /// PostIncLoops - If this user is to use the post-incremented value of an 929 /// induction variable, this variable is non-null and holds the loop 930 /// associated with the induction variable. 931 PostIncLoopSet PostIncLoops; 932 933 /// LUIdx - The index of the LSRUse describing the expression which 934 /// this fixup needs, minus an offset (below). 935 size_t LUIdx; 936 937 /// Offset - A constant offset to be added to the LSRUse expression. 938 /// This allows multiple fixups to share the same LSRUse with different 939 /// offsets, for example in an unrolled loop. 940 int64_t Offset; 941 942 bool isUseFullyOutsideLoop(const Loop *L) const; 943 944 LSRFixup(); 945 946 void print(raw_ostream &OS) const; 947 void dump() const; 948}; 949 950} 951 952LSRFixup::LSRFixup() 953 : UserInst(0), OperandValToReplace(0), LUIdx(~size_t(0)), Offset(0) {} 954 955/// isUseFullyOutsideLoop - Test whether this fixup always uses its 956/// value outside of the given loop. 957bool LSRFixup::isUseFullyOutsideLoop(const Loop *L) const { 958 // PHI nodes use their value in their incoming blocks. 959 if (const PHINode *PN = dyn_cast<PHINode>(UserInst)) { 960 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) 961 if (PN->getIncomingValue(i) == OperandValToReplace && 962 L->contains(PN->getIncomingBlock(i))) 963 return false; 964 return true; 965 } 966 967 return !L->contains(UserInst); 968} 969 970void LSRFixup::print(raw_ostream &OS) const { 971 OS << "UserInst="; 972 // Store is common and interesting enough to be worth special-casing. 973 if (StoreInst *Store = dyn_cast<StoreInst>(UserInst)) { 974 OS << "store "; 975 WriteAsOperand(OS, Store->getOperand(0), /*PrintType=*/false); 976 } else if (UserInst->getType()->isVoidTy()) 977 OS << UserInst->getOpcodeName(); 978 else 979 WriteAsOperand(OS, UserInst, /*PrintType=*/false); 980 981 OS << ", OperandValToReplace="; 982 WriteAsOperand(OS, OperandValToReplace, /*PrintType=*/false); 983 984 for (PostIncLoopSet::const_iterator I = PostIncLoops.begin(), 985 E = PostIncLoops.end(); I != E; ++I) { 986 OS << ", PostIncLoop="; 987 WriteAsOperand(OS, (*I)->getHeader(), /*PrintType=*/false); 988 } 989 990 if (LUIdx != ~size_t(0)) 991 OS << ", LUIdx=" << LUIdx; 992 993 if (Offset != 0) 994 OS << ", Offset=" << Offset; 995} 996 997void LSRFixup::dump() const { 998 print(errs()); errs() << '\n'; 999} 1000 1001namespace { 1002 1003/// UniquifierDenseMapInfo - A DenseMapInfo implementation for holding 1004/// DenseMaps and DenseSets of sorted SmallVectors of const SCEV*. 1005struct UniquifierDenseMapInfo { 1006 static SmallVector<const SCEV *, 2> getEmptyKey() { 1007 SmallVector<const SCEV *, 2> V; 1008 V.push_back(reinterpret_cast<const SCEV *>(-1)); 1009 return V; 1010 } 1011 1012 static SmallVector<const SCEV *, 2> getTombstoneKey() { 1013 SmallVector<const SCEV *, 2> V; 1014 V.push_back(reinterpret_cast<const SCEV *>(-2)); 1015 return V; 1016 } 1017 1018 static unsigned getHashValue(const SmallVector<const SCEV *, 2> &V) { 1019 unsigned Result = 0; 1020 for (SmallVectorImpl<const SCEV *>::const_iterator I = V.begin(), 1021 E = V.end(); I != E; ++I) 1022 Result ^= DenseMapInfo<const SCEV *>::getHashValue(*I); 1023 return Result; 1024 } 1025 1026 static bool isEqual(const SmallVector<const SCEV *, 2> &LHS, 1027 const SmallVector<const SCEV *, 2> &RHS) { 1028 return LHS == RHS; 1029 } 1030}; 1031 1032/// LSRUse - This class holds the state that LSR keeps for each use in 1033/// IVUsers, as well as uses invented by LSR itself. It includes information 1034/// about what kinds of things can be folded into the user, information about 1035/// the user itself, and information about how the use may be satisfied. 1036/// TODO: Represent multiple users of the same expression in common? 1037class LSRUse { 1038 DenseSet<SmallVector<const SCEV *, 2>, UniquifierDenseMapInfo> Uniquifier; 1039 1040public: 1041 /// KindType - An enum for a kind of use, indicating what types of 1042 /// scaled and immediate operands it might support. 1043 enum KindType { 1044 Basic, ///< A normal use, with no folding. 1045 Special, ///< A special case of basic, allowing -1 scales. 1046 Address, ///< An address use; folding according to TargetLowering 1047 ICmpZero ///< An equality icmp with both operands folded into one. 1048 // TODO: Add a generic icmp too? 1049 }; 1050 1051 KindType Kind; 1052 Type *AccessTy; 1053 1054 SmallVector<int64_t, 8> Offsets; 1055 int64_t MinOffset; 1056 int64_t MaxOffset; 1057 1058 /// AllFixupsOutsideLoop - This records whether all of the fixups using this 1059 /// LSRUse are outside of the loop, in which case some special-case heuristics 1060 /// may be used. 1061 bool AllFixupsOutsideLoop; 1062 1063 /// WidestFixupType - This records the widest use type for any fixup using 1064 /// this LSRUse. FindUseWithSimilarFormula can't consider uses with different 1065 /// max fixup widths to be equivalent, because the narrower one may be relying 1066 /// on the implicit truncation to truncate away bogus bits. 1067 Type *WidestFixupType; 1068 1069 /// Formulae - A list of ways to build a value that can satisfy this user. 1070 /// After the list is populated, one of these is selected heuristically and 1071 /// used to formulate a replacement for OperandValToReplace in UserInst. 1072 SmallVector<Formula, 12> Formulae; 1073 1074 /// Regs - The set of register candidates used by all formulae in this LSRUse. 1075 SmallPtrSet<const SCEV *, 4> Regs; 1076 1077 LSRUse(KindType K, Type *T) : Kind(K), AccessTy(T), 1078 MinOffset(INT64_MAX), 1079 MaxOffset(INT64_MIN), 1080 AllFixupsOutsideLoop(true), 1081 WidestFixupType(0) {} 1082 1083 bool HasFormulaWithSameRegs(const Formula &F) const; 1084 bool InsertFormula(const Formula &F); 1085 void DeleteFormula(Formula &F); 1086 void RecomputeRegs(size_t LUIdx, RegUseTracker &Reguses); 1087 1088 void print(raw_ostream &OS) const; 1089 void dump() const; 1090}; 1091 1092} 1093 1094/// HasFormula - Test whether this use as a formula which has the same 1095/// registers as the given formula. 1096bool LSRUse::HasFormulaWithSameRegs(const Formula &F) const { 1097 SmallVector<const SCEV *, 2> Key = F.BaseRegs; 1098 if (F.ScaledReg) Key.push_back(F.ScaledReg); 1099 // Unstable sort by host order ok, because this is only used for uniquifying. 1100 std::sort(Key.begin(), Key.end()); 1101 return Uniquifier.count(Key); 1102} 1103 1104/// InsertFormula - If the given formula has not yet been inserted, add it to 1105/// the list, and return true. Return false otherwise. 1106bool LSRUse::InsertFormula(const Formula &F) { 1107 SmallVector<const SCEV *, 2> Key = F.BaseRegs; 1108 if (F.ScaledReg) Key.push_back(F.ScaledReg); 1109 // Unstable sort by host order ok, because this is only used for uniquifying. 1110 std::sort(Key.begin(), Key.end()); 1111 1112 if (!Uniquifier.insert(Key).second) 1113 return false; 1114 1115 // Using a register to hold the value of 0 is not profitable. 1116 assert((!F.ScaledReg || !F.ScaledReg->isZero()) && 1117 "Zero allocated in a scaled register!"); 1118#ifndef NDEBUG 1119 for (SmallVectorImpl<const SCEV *>::const_iterator I = 1120 F.BaseRegs.begin(), E = F.BaseRegs.end(); I != E; ++I) 1121 assert(!(*I)->isZero() && "Zero allocated in a base register!"); 1122#endif 1123 1124 // Add the formula to the list. 1125 Formulae.push_back(F); 1126 1127 // Record registers now being used by this use. 1128 Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end()); 1129 1130 return true; 1131} 1132 1133/// DeleteFormula - Remove the given formula from this use's list. 1134void LSRUse::DeleteFormula(Formula &F) { 1135 if (&F != &Formulae.back()) 1136 std::swap(F, Formulae.back()); 1137 Formulae.pop_back(); 1138} 1139 1140/// RecomputeRegs - Recompute the Regs field, and update RegUses. 1141void LSRUse::RecomputeRegs(size_t LUIdx, RegUseTracker &RegUses) { 1142 // Now that we've filtered out some formulae, recompute the Regs set. 1143 SmallPtrSet<const SCEV *, 4> OldRegs = Regs; 1144 Regs.clear(); 1145 for (SmallVectorImpl<Formula>::const_iterator I = Formulae.begin(), 1146 E = Formulae.end(); I != E; ++I) { 1147 const Formula &F = *I; 1148 if (F.ScaledReg) Regs.insert(F.ScaledReg); 1149 Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end()); 1150 } 1151 1152 // Update the RegTracker. 1153 for (SmallPtrSet<const SCEV *, 4>::iterator I = OldRegs.begin(), 1154 E = OldRegs.end(); I != E; ++I) 1155 if (!Regs.count(*I)) 1156 RegUses.DropRegister(*I, LUIdx); 1157} 1158 1159void LSRUse::print(raw_ostream &OS) const { 1160 OS << "LSR Use: Kind="; 1161 switch (Kind) { 1162 case Basic: OS << "Basic"; break; 1163 case Special: OS << "Special"; break; 1164 case ICmpZero: OS << "ICmpZero"; break; 1165 case Address: 1166 OS << "Address of "; 1167 if (AccessTy->isPointerTy()) 1168 OS << "pointer"; // the full pointer type could be really verbose 1169 else 1170 OS << *AccessTy; 1171 } 1172 1173 OS << ", Offsets={"; 1174 for (SmallVectorImpl<int64_t>::const_iterator I = Offsets.begin(), 1175 E = Offsets.end(); I != E; ++I) { 1176 OS << *I; 1177 if (llvm::next(I) != E) 1178 OS << ','; 1179 } 1180 OS << '}'; 1181 1182 if (AllFixupsOutsideLoop) 1183 OS << ", all-fixups-outside-loop"; 1184 1185 if (WidestFixupType) 1186 OS << ", widest fixup type: " << *WidestFixupType; 1187} 1188 1189void LSRUse::dump() const { 1190 print(errs()); errs() << '\n'; 1191} 1192 1193/// isLegalUse - Test whether the use described by AM is "legal", meaning it can 1194/// be completely folded into the user instruction at isel time. This includes 1195/// address-mode folding and special icmp tricks. 1196static bool isLegalUse(const TargetLowering::AddrMode &AM, 1197 LSRUse::KindType Kind, Type *AccessTy, 1198 const TargetLowering *TLI) { 1199 switch (Kind) { 1200 case LSRUse::Address: 1201 // If we have low-level target information, ask the target if it can 1202 // completely fold this address. 1203 if (TLI) return TLI->isLegalAddressingMode(AM, AccessTy); 1204 1205 // Otherwise, just guess that reg+reg addressing is legal. 1206 return !AM.BaseGV && AM.BaseOffs == 0 && AM.Scale <= 1; 1207 1208 case LSRUse::ICmpZero: 1209 // There's not even a target hook for querying whether it would be legal to 1210 // fold a GV into an ICmp. 1211 if (AM.BaseGV) 1212 return false; 1213 1214 // ICmp only has two operands; don't allow more than two non-trivial parts. 1215 if (AM.Scale != 0 && AM.HasBaseReg && AM.BaseOffs != 0) 1216 return false; 1217 1218 // ICmp only supports no scale or a -1 scale, as we can "fold" a -1 scale by 1219 // putting the scaled register in the other operand of the icmp. 1220 if (AM.Scale != 0 && AM.Scale != -1) 1221 return false; 1222 1223 // If we have low-level target information, ask the target if it can fold an 1224 // integer immediate on an icmp. 1225 if (AM.BaseOffs != 0) { 1226 if (TLI) return TLI->isLegalICmpImmediate(-(uint64_t)AM.BaseOffs); 1227 return false; 1228 } 1229 1230 return true; 1231 1232 case LSRUse::Basic: 1233 // Only handle single-register values. 1234 return !AM.BaseGV && AM.Scale == 0 && AM.BaseOffs == 0; 1235 1236 case LSRUse::Special: 1237 // Only handle -1 scales, or no scale. 1238 return AM.Scale == 0 || AM.Scale == -1; 1239 } 1240 1241 return false; 1242} 1243 1244static bool isLegalUse(TargetLowering::AddrMode AM, 1245 int64_t MinOffset, int64_t MaxOffset, 1246 LSRUse::KindType Kind, Type *AccessTy, 1247 const TargetLowering *TLI) { 1248 // Check for overflow. 1249 if (((int64_t)((uint64_t)AM.BaseOffs + MinOffset) > AM.BaseOffs) != 1250 (MinOffset > 0)) 1251 return false; 1252 AM.BaseOffs = (uint64_t)AM.BaseOffs + MinOffset; 1253 if (isLegalUse(AM, Kind, AccessTy, TLI)) { 1254 AM.BaseOffs = (uint64_t)AM.BaseOffs - MinOffset; 1255 // Check for overflow. 1256 if (((int64_t)((uint64_t)AM.BaseOffs + MaxOffset) > AM.BaseOffs) != 1257 (MaxOffset > 0)) 1258 return false; 1259 AM.BaseOffs = (uint64_t)AM.BaseOffs + MaxOffset; 1260 return isLegalUse(AM, Kind, AccessTy, TLI); 1261 } 1262 return false; 1263} 1264 1265static bool isAlwaysFoldable(int64_t BaseOffs, 1266 GlobalValue *BaseGV, 1267 bool HasBaseReg, 1268 LSRUse::KindType Kind, Type *AccessTy, 1269 const TargetLowering *TLI) { 1270 // Fast-path: zero is always foldable. 1271 if (BaseOffs == 0 && !BaseGV) return true; 1272 1273 // Conservatively, create an address with an immediate and a 1274 // base and a scale. 1275 TargetLowering::AddrMode AM; 1276 AM.BaseOffs = BaseOffs; 1277 AM.BaseGV = BaseGV; 1278 AM.HasBaseReg = HasBaseReg; 1279 AM.Scale = Kind == LSRUse::ICmpZero ? -1 : 1; 1280 1281 // Canonicalize a scale of 1 to a base register if the formula doesn't 1282 // already have a base register. 1283 if (!AM.HasBaseReg && AM.Scale == 1) { 1284 AM.Scale = 0; 1285 AM.HasBaseReg = true; 1286 } 1287 1288 return isLegalUse(AM, Kind, AccessTy, TLI); 1289} 1290 1291static bool isAlwaysFoldable(const SCEV *S, 1292 int64_t MinOffset, int64_t MaxOffset, 1293 bool HasBaseReg, 1294 LSRUse::KindType Kind, Type *AccessTy, 1295 const TargetLowering *TLI, 1296 ScalarEvolution &SE) { 1297 // Fast-path: zero is always foldable. 1298 if (S->isZero()) return true; 1299 1300 // Conservatively, create an address with an immediate and a 1301 // base and a scale. 1302 int64_t BaseOffs = ExtractImmediate(S, SE); 1303 GlobalValue *BaseGV = ExtractSymbol(S, SE); 1304 1305 // If there's anything else involved, it's not foldable. 1306 if (!S->isZero()) return false; 1307 1308 // Fast-path: zero is always foldable. 1309 if (BaseOffs == 0 && !BaseGV) return true; 1310 1311 // Conservatively, create an address with an immediate and a 1312 // base and a scale. 1313 TargetLowering::AddrMode AM; 1314 AM.BaseOffs = BaseOffs; 1315 AM.BaseGV = BaseGV; 1316 AM.HasBaseReg = HasBaseReg; 1317 AM.Scale = Kind == LSRUse::ICmpZero ? -1 : 1; 1318 1319 return isLegalUse(AM, MinOffset, MaxOffset, Kind, AccessTy, TLI); 1320} 1321 1322namespace { 1323 1324/// UseMapDenseMapInfo - A DenseMapInfo implementation for holding 1325/// DenseMaps and DenseSets of pairs of const SCEV* and LSRUse::Kind. 1326struct UseMapDenseMapInfo { 1327 static std::pair<const SCEV *, LSRUse::KindType> getEmptyKey() { 1328 return std::make_pair(reinterpret_cast<const SCEV *>(-1), LSRUse::Basic); 1329 } 1330 1331 static std::pair<const SCEV *, LSRUse::KindType> getTombstoneKey() { 1332 return std::make_pair(reinterpret_cast<const SCEV *>(-2), LSRUse::Basic); 1333 } 1334 1335 static unsigned 1336 getHashValue(const std::pair<const SCEV *, LSRUse::KindType> &V) { 1337 unsigned Result = DenseMapInfo<const SCEV *>::getHashValue(V.first); 1338 Result ^= DenseMapInfo<unsigned>::getHashValue(unsigned(V.second)); 1339 return Result; 1340 } 1341 1342 static bool isEqual(const std::pair<const SCEV *, LSRUse::KindType> &LHS, 1343 const std::pair<const SCEV *, LSRUse::KindType> &RHS) { 1344 return LHS == RHS; 1345 } 1346}; 1347 1348/// LSRInstance - This class holds state for the main loop strength reduction 1349/// logic. 1350class LSRInstance { 1351 IVUsers &IU; 1352 ScalarEvolution &SE; 1353 DominatorTree &DT; 1354 LoopInfo &LI; 1355 const TargetLowering *const TLI; 1356 Loop *const L; 1357 bool Changed; 1358 1359 /// IVIncInsertPos - This is the insert position that the current loop's 1360 /// induction variable increment should be placed. In simple loops, this is 1361 /// the latch block's terminator. But in more complicated cases, this is a 1362 /// position which will dominate all the in-loop post-increment users. 1363 Instruction *IVIncInsertPos; 1364 1365 /// Factors - Interesting factors between use strides. 1366 SmallSetVector<int64_t, 8> Factors; 1367 1368 /// Types - Interesting use types, to facilitate truncation reuse. 1369 SmallSetVector<Type *, 4> Types; 1370 1371 /// Fixups - The list of operands which are to be replaced. 1372 SmallVector<LSRFixup, 16> Fixups; 1373 1374 /// Uses - The list of interesting uses. 1375 SmallVector<LSRUse, 16> Uses; 1376 1377 /// RegUses - Track which uses use which register candidates. 1378 RegUseTracker RegUses; 1379 1380 void OptimizeShadowIV(); 1381 bool FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse); 1382 ICmpInst *OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse); 1383 void OptimizeLoopTermCond(); 1384 1385 void CollectInterestingTypesAndFactors(); 1386 void CollectFixupsAndInitialFormulae(); 1387 1388 LSRFixup &getNewFixup() { 1389 Fixups.push_back(LSRFixup()); 1390 return Fixups.back(); 1391 } 1392 1393 // Support for sharing of LSRUses between LSRFixups. 1394 typedef DenseMap<std::pair<const SCEV *, LSRUse::KindType>, 1395 size_t, 1396 UseMapDenseMapInfo> UseMapTy; 1397 UseMapTy UseMap; 1398 1399 bool reconcileNewOffset(LSRUse &LU, int64_t NewOffset, bool HasBaseReg, 1400 LSRUse::KindType Kind, Type *AccessTy); 1401 1402 std::pair<size_t, int64_t> getUse(const SCEV *&Expr, 1403 LSRUse::KindType Kind, 1404 Type *AccessTy); 1405 1406 void DeleteUse(LSRUse &LU, size_t LUIdx); 1407 1408 LSRUse *FindUseWithSimilarFormula(const Formula &F, const LSRUse &OrigLU); 1409 1410 void InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx); 1411 void InsertSupplementalFormula(const SCEV *S, LSRUse &LU, size_t LUIdx); 1412 void CountRegisters(const Formula &F, size_t LUIdx); 1413 bool InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F); 1414 1415 void CollectLoopInvariantFixupsAndFormulae(); 1416 1417 void GenerateReassociations(LSRUse &LU, unsigned LUIdx, Formula Base, 1418 unsigned Depth = 0); 1419 void GenerateCombinations(LSRUse &LU, unsigned LUIdx, Formula Base); 1420 void GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx, Formula Base); 1421 void GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx, Formula Base); 1422 void GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx, Formula Base); 1423 void GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base); 1424 void GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base); 1425 void GenerateCrossUseConstantOffsets(); 1426 void GenerateAllReuseFormulae(); 1427 1428 void FilterOutUndesirableDedicatedRegisters(); 1429 1430 size_t EstimateSearchSpaceComplexity() const; 1431 void NarrowSearchSpaceByDetectingSupersets(); 1432 void NarrowSearchSpaceByCollapsingUnrolledCode(); 1433 void NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters(); 1434 void NarrowSearchSpaceByPickingWinnerRegs(); 1435 void NarrowSearchSpaceUsingHeuristics(); 1436 1437 void SolveRecurse(SmallVectorImpl<const Formula *> &Solution, 1438 Cost &SolutionCost, 1439 SmallVectorImpl<const Formula *> &Workspace, 1440 const Cost &CurCost, 1441 const SmallPtrSet<const SCEV *, 16> &CurRegs, 1442 DenseSet<const SCEV *> &VisitedRegs) const; 1443 void Solve(SmallVectorImpl<const Formula *> &Solution) const; 1444 1445 BasicBlock::iterator 1446 HoistInsertPosition(BasicBlock::iterator IP, 1447 const SmallVectorImpl<Instruction *> &Inputs) const; 1448 BasicBlock::iterator AdjustInsertPositionForExpand(BasicBlock::iterator IP, 1449 const LSRFixup &LF, 1450 const LSRUse &LU) const; 1451 1452 Value *Expand(const LSRFixup &LF, 1453 const Formula &F, 1454 BasicBlock::iterator IP, 1455 SCEVExpander &Rewriter, 1456 SmallVectorImpl<WeakVH> &DeadInsts) const; 1457 void RewriteForPHI(PHINode *PN, const LSRFixup &LF, 1458 const Formula &F, 1459 SCEVExpander &Rewriter, 1460 SmallVectorImpl<WeakVH> &DeadInsts, 1461 Pass *P) const; 1462 void Rewrite(const LSRFixup &LF, 1463 const Formula &F, 1464 SCEVExpander &Rewriter, 1465 SmallVectorImpl<WeakVH> &DeadInsts, 1466 Pass *P) const; 1467 void ImplementSolution(const SmallVectorImpl<const Formula *> &Solution, 1468 Pass *P); 1469 1470public: 1471 LSRInstance(const TargetLowering *tli, Loop *l, Pass *P); 1472 1473 bool getChanged() const { return Changed; } 1474 1475 void print_factors_and_types(raw_ostream &OS) const; 1476 void print_fixups(raw_ostream &OS) const; 1477 void print_uses(raw_ostream &OS) const; 1478 void print(raw_ostream &OS) const; 1479 void dump() const; 1480}; 1481 1482} 1483 1484/// OptimizeShadowIV - If IV is used in a int-to-float cast 1485/// inside the loop then try to eliminate the cast operation. 1486void LSRInstance::OptimizeShadowIV() { 1487 const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L); 1488 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount)) 1489 return; 1490 1491 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); 1492 UI != E; /* empty */) { 1493 IVUsers::const_iterator CandidateUI = UI; 1494 ++UI; 1495 Instruction *ShadowUse = CandidateUI->getUser(); 1496 Type *DestTy = NULL; 1497 bool IsSigned = false; 1498 1499 /* If shadow use is a int->float cast then insert a second IV 1500 to eliminate this cast. 1501 1502 for (unsigned i = 0; i < n; ++i) 1503 foo((double)i); 1504 1505 is transformed into 1506 1507 double d = 0.0; 1508 for (unsigned i = 0; i < n; ++i, ++d) 1509 foo(d); 1510 */ 1511 if (UIToFPInst *UCast = dyn_cast<UIToFPInst>(CandidateUI->getUser())) { 1512 IsSigned = false; 1513 DestTy = UCast->getDestTy(); 1514 } 1515 else if (SIToFPInst *SCast = dyn_cast<SIToFPInst>(CandidateUI->getUser())) { 1516 IsSigned = true; 1517 DestTy = SCast->getDestTy(); 1518 } 1519 if (!DestTy) continue; 1520 1521 if (TLI) { 1522 // If target does not support DestTy natively then do not apply 1523 // this transformation. 1524 EVT DVT = TLI->getValueType(DestTy); 1525 if (!TLI->isTypeLegal(DVT)) continue; 1526 } 1527 1528 PHINode *PH = dyn_cast<PHINode>(ShadowUse->getOperand(0)); 1529 if (!PH) continue; 1530 if (PH->getNumIncomingValues() != 2) continue; 1531 1532 Type *SrcTy = PH->getType(); 1533 int Mantissa = DestTy->getFPMantissaWidth(); 1534 if (Mantissa == -1) continue; 1535 if ((int)SE.getTypeSizeInBits(SrcTy) > Mantissa) 1536 continue; 1537 1538 unsigned Entry, Latch; 1539 if (PH->getIncomingBlock(0) == L->getLoopPreheader()) { 1540 Entry = 0; 1541 Latch = 1; 1542 } else { 1543 Entry = 1; 1544 Latch = 0; 1545 } 1546 1547 ConstantInt *Init = dyn_cast<ConstantInt>(PH->getIncomingValue(Entry)); 1548 if (!Init) continue; 1549 Constant *NewInit = ConstantFP::get(DestTy, IsSigned ? 1550 (double)Init->getSExtValue() : 1551 (double)Init->getZExtValue()); 1552 1553 BinaryOperator *Incr = 1554 dyn_cast<BinaryOperator>(PH->getIncomingValue(Latch)); 1555 if (!Incr) continue; 1556 if (Incr->getOpcode() != Instruction::Add 1557 && Incr->getOpcode() != Instruction::Sub) 1558 continue; 1559 1560 /* Initialize new IV, double d = 0.0 in above example. */ 1561 ConstantInt *C = NULL; 1562 if (Incr->getOperand(0) == PH) 1563 C = dyn_cast<ConstantInt>(Incr->getOperand(1)); 1564 else if (Incr->getOperand(1) == PH) 1565 C = dyn_cast<ConstantInt>(Incr->getOperand(0)); 1566 else 1567 continue; 1568 1569 if (!C) continue; 1570 1571 // Ignore negative constants, as the code below doesn't handle them 1572 // correctly. TODO: Remove this restriction. 1573 if (!C->getValue().isStrictlyPositive()) continue; 1574 1575 /* Add new PHINode. */ 1576 PHINode *NewPH = PHINode::Create(DestTy, 2, "IV.S.", PH); 1577 1578 /* create new increment. '++d' in above example. */ 1579 Constant *CFP = ConstantFP::get(DestTy, C->getZExtValue()); 1580 BinaryOperator *NewIncr = 1581 BinaryOperator::Create(Incr->getOpcode() == Instruction::Add ? 1582 Instruction::FAdd : Instruction::FSub, 1583 NewPH, CFP, "IV.S.next.", Incr); 1584 1585 NewPH->addIncoming(NewInit, PH->getIncomingBlock(Entry)); 1586 NewPH->addIncoming(NewIncr, PH->getIncomingBlock(Latch)); 1587 1588 /* Remove cast operation */ 1589 ShadowUse->replaceAllUsesWith(NewPH); 1590 ShadowUse->eraseFromParent(); 1591 Changed = true; 1592 break; 1593 } 1594} 1595 1596/// FindIVUserForCond - If Cond has an operand that is an expression of an IV, 1597/// set the IV user and stride information and return true, otherwise return 1598/// false. 1599bool LSRInstance::FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse) { 1600 for (IVUsers::iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI) 1601 if (UI->getUser() == Cond) { 1602 // NOTE: we could handle setcc instructions with multiple uses here, but 1603 // InstCombine does it as well for simple uses, it's not clear that it 1604 // occurs enough in real life to handle. 1605 CondUse = UI; 1606 return true; 1607 } 1608 return false; 1609} 1610 1611/// OptimizeMax - Rewrite the loop's terminating condition if it uses 1612/// a max computation. 1613/// 1614/// This is a narrow solution to a specific, but acute, problem. For loops 1615/// like this: 1616/// 1617/// i = 0; 1618/// do { 1619/// p[i] = 0.0; 1620/// } while (++i < n); 1621/// 1622/// the trip count isn't just 'n', because 'n' might not be positive. And 1623/// unfortunately this can come up even for loops where the user didn't use 1624/// a C do-while loop. For example, seemingly well-behaved top-test loops 1625/// will commonly be lowered like this: 1626// 1627/// if (n > 0) { 1628/// i = 0; 1629/// do { 1630/// p[i] = 0.0; 1631/// } while (++i < n); 1632/// } 1633/// 1634/// and then it's possible for subsequent optimization to obscure the if 1635/// test in such a way that indvars can't find it. 1636/// 1637/// When indvars can't find the if test in loops like this, it creates a 1638/// max expression, which allows it to give the loop a canonical 1639/// induction variable: 1640/// 1641/// i = 0; 1642/// max = n < 1 ? 1 : n; 1643/// do { 1644/// p[i] = 0.0; 1645/// } while (++i != max); 1646/// 1647/// Canonical induction variables are necessary because the loop passes 1648/// are designed around them. The most obvious example of this is the 1649/// LoopInfo analysis, which doesn't remember trip count values. It 1650/// expects to be able to rediscover the trip count each time it is 1651/// needed, and it does this using a simple analysis that only succeeds if 1652/// the loop has a canonical induction variable. 1653/// 1654/// However, when it comes time to generate code, the maximum operation 1655/// can be quite costly, especially if it's inside of an outer loop. 1656/// 1657/// This function solves this problem by detecting this type of loop and 1658/// rewriting their conditions from ICMP_NE back to ICMP_SLT, and deleting 1659/// the instructions for the maximum computation. 1660/// 1661ICmpInst *LSRInstance::OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse) { 1662 // Check that the loop matches the pattern we're looking for. 1663 if (Cond->getPredicate() != CmpInst::ICMP_EQ && 1664 Cond->getPredicate() != CmpInst::ICMP_NE) 1665 return Cond; 1666 1667 SelectInst *Sel = dyn_cast<SelectInst>(Cond->getOperand(1)); 1668 if (!Sel || !Sel->hasOneUse()) return Cond; 1669 1670 const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L); 1671 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount)) 1672 return Cond; 1673 const SCEV *One = SE.getConstant(BackedgeTakenCount->getType(), 1); 1674 1675 // Add one to the backedge-taken count to get the trip count. 1676 const SCEV *IterationCount = SE.getAddExpr(One, BackedgeTakenCount); 1677 if (IterationCount != SE.getSCEV(Sel)) return Cond; 1678 1679 // Check for a max calculation that matches the pattern. There's no check 1680 // for ICMP_ULE here because the comparison would be with zero, which 1681 // isn't interesting. 1682 CmpInst::Predicate Pred = ICmpInst::BAD_ICMP_PREDICATE; 1683 const SCEVNAryExpr *Max = 0; 1684 if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(BackedgeTakenCount)) { 1685 Pred = ICmpInst::ICMP_SLE; 1686 Max = S; 1687 } else if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(IterationCount)) { 1688 Pred = ICmpInst::ICMP_SLT; 1689 Max = S; 1690 } else if (const SCEVUMaxExpr *U = dyn_cast<SCEVUMaxExpr>(IterationCount)) { 1691 Pred = ICmpInst::ICMP_ULT; 1692 Max = U; 1693 } else { 1694 // No match; bail. 1695 return Cond; 1696 } 1697 1698 // To handle a max with more than two operands, this optimization would 1699 // require additional checking and setup. 1700 if (Max->getNumOperands() != 2) 1701 return Cond; 1702 1703 const SCEV *MaxLHS = Max->getOperand(0); 1704 const SCEV *MaxRHS = Max->getOperand(1); 1705 1706 // ScalarEvolution canonicalizes constants to the left. For < and >, look 1707 // for a comparison with 1. For <= and >=, a comparison with zero. 1708 if (!MaxLHS || 1709 (ICmpInst::isTrueWhenEqual(Pred) ? !MaxLHS->isZero() : (MaxLHS != One))) 1710 return Cond; 1711 1712 // Check the relevant induction variable for conformance to 1713 // the pattern. 1714 const SCEV *IV = SE.getSCEV(Cond->getOperand(0)); 1715 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(IV); 1716 if (!AR || !AR->isAffine() || 1717 AR->getStart() != One || 1718 AR->getStepRecurrence(SE) != One) 1719 return Cond; 1720 1721 assert(AR->getLoop() == L && 1722 "Loop condition operand is an addrec in a different loop!"); 1723 1724 // Check the right operand of the select, and remember it, as it will 1725 // be used in the new comparison instruction. 1726 Value *NewRHS = 0; 1727 if (ICmpInst::isTrueWhenEqual(Pred)) { 1728 // Look for n+1, and grab n. 1729 if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(1))) 1730 if (isa<ConstantInt>(BO->getOperand(1)) && 1731 cast<ConstantInt>(BO->getOperand(1))->isOne() && 1732 SE.getSCEV(BO->getOperand(0)) == MaxRHS) 1733 NewRHS = BO->getOperand(0); 1734 if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(2))) 1735 if (isa<ConstantInt>(BO->getOperand(1)) && 1736 cast<ConstantInt>(BO->getOperand(1))->isOne() && 1737 SE.getSCEV(BO->getOperand(0)) == MaxRHS) 1738 NewRHS = BO->getOperand(0); 1739 if (!NewRHS) 1740 return Cond; 1741 } else if (SE.getSCEV(Sel->getOperand(1)) == MaxRHS) 1742 NewRHS = Sel->getOperand(1); 1743 else if (SE.getSCEV(Sel->getOperand(2)) == MaxRHS) 1744 NewRHS = Sel->getOperand(2); 1745 else if (const SCEVUnknown *SU = dyn_cast<SCEVUnknown>(MaxRHS)) 1746 NewRHS = SU->getValue(); 1747 else 1748 // Max doesn't match expected pattern. 1749 return Cond; 1750 1751 // Determine the new comparison opcode. It may be signed or unsigned, 1752 // and the original comparison may be either equality or inequality. 1753 if (Cond->getPredicate() == CmpInst::ICMP_EQ) 1754 Pred = CmpInst::getInversePredicate(Pred); 1755 1756 // Ok, everything looks ok to change the condition into an SLT or SGE and 1757 // delete the max calculation. 1758 ICmpInst *NewCond = 1759 new ICmpInst(Cond, Pred, Cond->getOperand(0), NewRHS, "scmp"); 1760 1761 // Delete the max calculation instructions. 1762 Cond->replaceAllUsesWith(NewCond); 1763 CondUse->setUser(NewCond); 1764 Instruction *Cmp = cast<Instruction>(Sel->getOperand(0)); 1765 Cond->eraseFromParent(); 1766 Sel->eraseFromParent(); 1767 if (Cmp->use_empty()) 1768 Cmp->eraseFromParent(); 1769 return NewCond; 1770} 1771 1772/// OptimizeLoopTermCond - Change loop terminating condition to use the 1773/// postinc iv when possible. 1774void 1775LSRInstance::OptimizeLoopTermCond() { 1776 SmallPtrSet<Instruction *, 4> PostIncs; 1777 1778 BasicBlock *LatchBlock = L->getLoopLatch(); 1779 SmallVector<BasicBlock*, 8> ExitingBlocks; 1780 L->getExitingBlocks(ExitingBlocks); 1781 1782 for (unsigned i = 0, e = ExitingBlocks.size(); i != e; ++i) { 1783 BasicBlock *ExitingBlock = ExitingBlocks[i]; 1784 1785 // Get the terminating condition for the loop if possible. If we 1786 // can, we want to change it to use a post-incremented version of its 1787 // induction variable, to allow coalescing the live ranges for the IV into 1788 // one register value. 1789 1790 BranchInst *TermBr = dyn_cast<BranchInst>(ExitingBlock->getTerminator()); 1791 if (!TermBr) 1792 continue; 1793 // FIXME: Overly conservative, termination condition could be an 'or' etc.. 1794 if (TermBr->isUnconditional() || !isa<ICmpInst>(TermBr->getCondition())) 1795 continue; 1796 1797 // Search IVUsesByStride to find Cond's IVUse if there is one. 1798 IVStrideUse *CondUse = 0; 1799 ICmpInst *Cond = cast<ICmpInst>(TermBr->getCondition()); 1800 if (!FindIVUserForCond(Cond, CondUse)) 1801 continue; 1802 1803 // If the trip count is computed in terms of a max (due to ScalarEvolution 1804 // being unable to find a sufficient guard, for example), change the loop 1805 // comparison to use SLT or ULT instead of NE. 1806 // One consequence of doing this now is that it disrupts the count-down 1807 // optimization. That's not always a bad thing though, because in such 1808 // cases it may still be worthwhile to avoid a max. 1809 Cond = OptimizeMax(Cond, CondUse); 1810 1811 // If this exiting block dominates the latch block, it may also use 1812 // the post-inc value if it won't be shared with other uses. 1813 // Check for dominance. 1814 if (!DT.dominates(ExitingBlock, LatchBlock)) 1815 continue; 1816 1817 // Conservatively avoid trying to use the post-inc value in non-latch 1818 // exits if there may be pre-inc users in intervening blocks. 1819 if (LatchBlock != ExitingBlock) 1820 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI) 1821 // Test if the use is reachable from the exiting block. This dominator 1822 // query is a conservative approximation of reachability. 1823 if (&*UI != CondUse && 1824 !DT.properlyDominates(UI->getUser()->getParent(), ExitingBlock)) { 1825 // Conservatively assume there may be reuse if the quotient of their 1826 // strides could be a legal scale. 1827 const SCEV *A = IU.getStride(*CondUse, L); 1828 const SCEV *B = IU.getStride(*UI, L); 1829 if (!A || !B) continue; 1830 if (SE.getTypeSizeInBits(A->getType()) != 1831 SE.getTypeSizeInBits(B->getType())) { 1832 if (SE.getTypeSizeInBits(A->getType()) > 1833 SE.getTypeSizeInBits(B->getType())) 1834 B = SE.getSignExtendExpr(B, A->getType()); 1835 else 1836 A = SE.getSignExtendExpr(A, B->getType()); 1837 } 1838 if (const SCEVConstant *D = 1839 dyn_cast_or_null<SCEVConstant>(getExactSDiv(B, A, SE))) { 1840 const ConstantInt *C = D->getValue(); 1841 // Stride of one or negative one can have reuse with non-addresses. 1842 if (C->isOne() || C->isAllOnesValue()) 1843 goto decline_post_inc; 1844 // Avoid weird situations. 1845 if (C->getValue().getMinSignedBits() >= 64 || 1846 C->getValue().isMinSignedValue()) 1847 goto decline_post_inc; 1848 // Without TLI, assume that any stride might be valid, and so any 1849 // use might be shared. 1850 if (!TLI) 1851 goto decline_post_inc; 1852 // Check for possible scaled-address reuse. 1853 Type *AccessTy = getAccessType(UI->getUser()); 1854 TargetLowering::AddrMode AM; 1855 AM.Scale = C->getSExtValue(); 1856 if (TLI->isLegalAddressingMode(AM, AccessTy)) 1857 goto decline_post_inc; 1858 AM.Scale = -AM.Scale; 1859 if (TLI->isLegalAddressingMode(AM, AccessTy)) 1860 goto decline_post_inc; 1861 } 1862 } 1863 1864 DEBUG(dbgs() << " Change loop exiting icmp to use postinc iv: " 1865 << *Cond << '\n'); 1866 1867 // It's possible for the setcc instruction to be anywhere in the loop, and 1868 // possible for it to have multiple users. If it is not immediately before 1869 // the exiting block branch, move it. 1870 if (&*++BasicBlock::iterator(Cond) != TermBr) { 1871 if (Cond->hasOneUse()) { 1872 Cond->moveBefore(TermBr); 1873 } else { 1874 // Clone the terminating condition and insert into the loopend. 1875 ICmpInst *OldCond = Cond; 1876 Cond = cast<ICmpInst>(Cond->clone()); 1877 Cond->setName(L->getHeader()->getName() + ".termcond"); 1878 ExitingBlock->getInstList().insert(TermBr, Cond); 1879 1880 // Clone the IVUse, as the old use still exists! 1881 CondUse = &IU.AddUser(Cond, CondUse->getOperandValToReplace()); 1882 TermBr->replaceUsesOfWith(OldCond, Cond); 1883 } 1884 } 1885 1886 // If we get to here, we know that we can transform the setcc instruction to 1887 // use the post-incremented version of the IV, allowing us to coalesce the 1888 // live ranges for the IV correctly. 1889 CondUse->transformToPostInc(L); 1890 Changed = true; 1891 1892 PostIncs.insert(Cond); 1893 decline_post_inc:; 1894 } 1895 1896 // Determine an insertion point for the loop induction variable increment. It 1897 // must dominate all the post-inc comparisons we just set up, and it must 1898 // dominate the loop latch edge. 1899 IVIncInsertPos = L->getLoopLatch()->getTerminator(); 1900 for (SmallPtrSet<Instruction *, 4>::const_iterator I = PostIncs.begin(), 1901 E = PostIncs.end(); I != E; ++I) { 1902 BasicBlock *BB = 1903 DT.findNearestCommonDominator(IVIncInsertPos->getParent(), 1904 (*I)->getParent()); 1905 if (BB == (*I)->getParent()) 1906 IVIncInsertPos = *I; 1907 else if (BB != IVIncInsertPos->getParent()) 1908 IVIncInsertPos = BB->getTerminator(); 1909 } 1910} 1911 1912/// reconcileNewOffset - Determine if the given use can accommodate a fixup 1913/// at the given offset and other details. If so, update the use and 1914/// return true. 1915bool 1916LSRInstance::reconcileNewOffset(LSRUse &LU, int64_t NewOffset, bool HasBaseReg, 1917 LSRUse::KindType Kind, Type *AccessTy) { 1918 int64_t NewMinOffset = LU.MinOffset; 1919 int64_t NewMaxOffset = LU.MaxOffset; 1920 Type *NewAccessTy = AccessTy; 1921 1922 // Check for a mismatched kind. It's tempting to collapse mismatched kinds to 1923 // something conservative, however this can pessimize in the case that one of 1924 // the uses will have all its uses outside the loop, for example. 1925 if (LU.Kind != Kind) 1926 return false; 1927 // Conservatively assume HasBaseReg is true for now. 1928 if (NewOffset < LU.MinOffset) { 1929 if (!isAlwaysFoldable(LU.MaxOffset - NewOffset, 0, HasBaseReg, 1930 Kind, AccessTy, TLI)) 1931 return false; 1932 NewMinOffset = NewOffset; 1933 } else if (NewOffset > LU.MaxOffset) { 1934 if (!isAlwaysFoldable(NewOffset - LU.MinOffset, 0, HasBaseReg, 1935 Kind, AccessTy, TLI)) 1936 return false; 1937 NewMaxOffset = NewOffset; 1938 } 1939 // Check for a mismatched access type, and fall back conservatively as needed. 1940 // TODO: Be less conservative when the type is similar and can use the same 1941 // addressing modes. 1942 if (Kind == LSRUse::Address && AccessTy != LU.AccessTy) 1943 NewAccessTy = Type::getVoidTy(AccessTy->getContext()); 1944 1945 // Update the use. 1946 LU.MinOffset = NewMinOffset; 1947 LU.MaxOffset = NewMaxOffset; 1948 LU.AccessTy = NewAccessTy; 1949 if (NewOffset != LU.Offsets.back()) 1950 LU.Offsets.push_back(NewOffset); 1951 return true; 1952} 1953 1954/// getUse - Return an LSRUse index and an offset value for a fixup which 1955/// needs the given expression, with the given kind and optional access type. 1956/// Either reuse an existing use or create a new one, as needed. 1957std::pair<size_t, int64_t> 1958LSRInstance::getUse(const SCEV *&Expr, 1959 LSRUse::KindType Kind, Type *AccessTy) { 1960 const SCEV *Copy = Expr; 1961 int64_t Offset = ExtractImmediate(Expr, SE); 1962 1963 // Basic uses can't accept any offset, for example. 1964 if (!isAlwaysFoldable(Offset, 0, /*HasBaseReg=*/true, Kind, AccessTy, TLI)) { 1965 Expr = Copy; 1966 Offset = 0; 1967 } 1968 1969 std::pair<UseMapTy::iterator, bool> P = 1970 UseMap.insert(std::make_pair(std::make_pair(Expr, Kind), 0)); 1971 if (!P.second) { 1972 // A use already existed with this base. 1973 size_t LUIdx = P.first->second; 1974 LSRUse &LU = Uses[LUIdx]; 1975 if (reconcileNewOffset(LU, Offset, /*HasBaseReg=*/true, Kind, AccessTy)) 1976 // Reuse this use. 1977 return std::make_pair(LUIdx, Offset); 1978 } 1979 1980 // Create a new use. 1981 size_t LUIdx = Uses.size(); 1982 P.first->second = LUIdx; 1983 Uses.push_back(LSRUse(Kind, AccessTy)); 1984 LSRUse &LU = Uses[LUIdx]; 1985 1986 // We don't need to track redundant offsets, but we don't need to go out 1987 // of our way here to avoid them. 1988 if (LU.Offsets.empty() || Offset != LU.Offsets.back()) 1989 LU.Offsets.push_back(Offset); 1990 1991 LU.MinOffset = Offset; 1992 LU.MaxOffset = Offset; 1993 return std::make_pair(LUIdx, Offset); 1994} 1995 1996/// DeleteUse - Delete the given use from the Uses list. 1997void LSRInstance::DeleteUse(LSRUse &LU, size_t LUIdx) { 1998 if (&LU != &Uses.back()) 1999 std::swap(LU, Uses.back()); 2000 Uses.pop_back(); 2001 2002 // Update RegUses. 2003 RegUses.SwapAndDropUse(LUIdx, Uses.size()); 2004} 2005 2006/// FindUseWithFormula - Look for a use distinct from OrigLU which is has 2007/// a formula that has the same registers as the given formula. 2008LSRUse * 2009LSRInstance::FindUseWithSimilarFormula(const Formula &OrigF, 2010 const LSRUse &OrigLU) { 2011 // Search all uses for the formula. This could be more clever. 2012 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) { 2013 LSRUse &LU = Uses[LUIdx]; 2014 // Check whether this use is close enough to OrigLU, to see whether it's 2015 // worthwhile looking through its formulae. 2016 // Ignore ICmpZero uses because they may contain formulae generated by 2017 // GenerateICmpZeroScales, in which case adding fixup offsets may 2018 // be invalid. 2019 if (&LU != &OrigLU && 2020 LU.Kind != LSRUse::ICmpZero && 2021 LU.Kind == OrigLU.Kind && OrigLU.AccessTy == LU.AccessTy && 2022 LU.WidestFixupType == OrigLU.WidestFixupType && 2023 LU.HasFormulaWithSameRegs(OrigF)) { 2024 // Scan through this use's formulae. 2025 for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(), 2026 E = LU.Formulae.end(); I != E; ++I) { 2027 const Formula &F = *I; 2028 // Check to see if this formula has the same registers and symbols 2029 // as OrigF. 2030 if (F.BaseRegs == OrigF.BaseRegs && 2031 F.ScaledReg == OrigF.ScaledReg && 2032 F.AM.BaseGV == OrigF.AM.BaseGV && 2033 F.AM.Scale == OrigF.AM.Scale && 2034 F.UnfoldedOffset == OrigF.UnfoldedOffset) { 2035 if (F.AM.BaseOffs == 0) 2036 return &LU; 2037 // This is the formula where all the registers and symbols matched; 2038 // there aren't going to be any others. Since we declined it, we 2039 // can skip the rest of the formulae and procede to the next LSRUse. 2040 break; 2041 } 2042 } 2043 } 2044 } 2045 2046 // Nothing looked good. 2047 return 0; 2048} 2049 2050void LSRInstance::CollectInterestingTypesAndFactors() { 2051 SmallSetVector<const SCEV *, 4> Strides; 2052 2053 // Collect interesting types and strides. 2054 SmallVector<const SCEV *, 4> Worklist; 2055 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI) { 2056 const SCEV *Expr = IU.getExpr(*UI); 2057 2058 // Collect interesting types. 2059 Types.insert(SE.getEffectiveSCEVType(Expr->getType())); 2060 2061 // Add strides for mentioned loops. 2062 Worklist.push_back(Expr); 2063 do { 2064 const SCEV *S = Worklist.pop_back_val(); 2065 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) { 2066 if (EnableNested || AR->getLoop() == L) 2067 Strides.insert(AR->getStepRecurrence(SE)); 2068 Worklist.push_back(AR->getStart()); 2069 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) { 2070 Worklist.append(Add->op_begin(), Add->op_end()); 2071 } 2072 } while (!Worklist.empty()); 2073 } 2074 2075 // Compute interesting factors from the set of interesting strides. 2076 for (SmallSetVector<const SCEV *, 4>::const_iterator 2077 I = Strides.begin(), E = Strides.end(); I != E; ++I) 2078 for (SmallSetVector<const SCEV *, 4>::const_iterator NewStrideIter = 2079 llvm::next(I); NewStrideIter != E; ++NewStrideIter) { 2080 const SCEV *OldStride = *I; 2081 const SCEV *NewStride = *NewStrideIter; 2082 2083 if (SE.getTypeSizeInBits(OldStride->getType()) != 2084 SE.getTypeSizeInBits(NewStride->getType())) { 2085 if (SE.getTypeSizeInBits(OldStride->getType()) > 2086 SE.getTypeSizeInBits(NewStride->getType())) 2087 NewStride = SE.getSignExtendExpr(NewStride, OldStride->getType()); 2088 else 2089 OldStride = SE.getSignExtendExpr(OldStride, NewStride->getType()); 2090 } 2091 if (const SCEVConstant *Factor = 2092 dyn_cast_or_null<SCEVConstant>(getExactSDiv(NewStride, OldStride, 2093 SE, true))) { 2094 if (Factor->getValue()->getValue().getMinSignedBits() <= 64) 2095 Factors.insert(Factor->getValue()->getValue().getSExtValue()); 2096 } else if (const SCEVConstant *Factor = 2097 dyn_cast_or_null<SCEVConstant>(getExactSDiv(OldStride, 2098 NewStride, 2099 SE, true))) { 2100 if (Factor->getValue()->getValue().getMinSignedBits() <= 64) 2101 Factors.insert(Factor->getValue()->getValue().getSExtValue()); 2102 } 2103 } 2104 2105 // If all uses use the same type, don't bother looking for truncation-based 2106 // reuse. 2107 if (Types.size() == 1) 2108 Types.clear(); 2109 2110 DEBUG(print_factors_and_types(dbgs())); 2111} 2112 2113void LSRInstance::CollectFixupsAndInitialFormulae() { 2114 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI) { 2115 // Record the uses. 2116 LSRFixup &LF = getNewFixup(); 2117 LF.UserInst = UI->getUser(); 2118 LF.OperandValToReplace = UI->getOperandValToReplace(); 2119 LF.PostIncLoops = UI->getPostIncLoops(); 2120 2121 LSRUse::KindType Kind = LSRUse::Basic; 2122 Type *AccessTy = 0; 2123 if (isAddressUse(LF.UserInst, LF.OperandValToReplace)) { 2124 Kind = LSRUse::Address; 2125 AccessTy = getAccessType(LF.UserInst); 2126 } 2127 2128 const SCEV *S = IU.getExpr(*UI); 2129 2130 // Equality (== and !=) ICmps are special. We can rewrite (i == N) as 2131 // (N - i == 0), and this allows (N - i) to be the expression that we work 2132 // with rather than just N or i, so we can consider the register 2133 // requirements for both N and i at the same time. Limiting this code to 2134 // equality icmps is not a problem because all interesting loops use 2135 // equality icmps, thanks to IndVarSimplify. 2136 if (ICmpInst *CI = dyn_cast<ICmpInst>(LF.UserInst)) 2137 if (CI->isEquality()) { 2138 // Swap the operands if needed to put the OperandValToReplace on the 2139 // left, for consistency. 2140 Value *NV = CI->getOperand(1); 2141 if (NV == LF.OperandValToReplace) { 2142 CI->setOperand(1, CI->getOperand(0)); 2143 CI->setOperand(0, NV); 2144 NV = CI->getOperand(1); 2145 Changed = true; 2146 } 2147 2148 // x == y --> x - y == 0 2149 const SCEV *N = SE.getSCEV(NV); 2150 if (SE.isLoopInvariant(N, L)) { 2151 // S is normalized, so normalize N before folding it into S 2152 // to keep the result normalized. 2153 N = TransformForPostIncUse(Normalize, N, CI, 0, 2154 LF.PostIncLoops, SE, DT); 2155 Kind = LSRUse::ICmpZero; 2156 S = SE.getMinusSCEV(N, S); 2157 } 2158 2159 // -1 and the negations of all interesting strides (except the negation 2160 // of -1) are now also interesting. 2161 for (size_t i = 0, e = Factors.size(); i != e; ++i) 2162 if (Factors[i] != -1) 2163 Factors.insert(-(uint64_t)Factors[i]); 2164 Factors.insert(-1); 2165 } 2166 2167 // Set up the initial formula for this use. 2168 std::pair<size_t, int64_t> P = getUse(S, Kind, AccessTy); 2169 LF.LUIdx = P.first; 2170 LF.Offset = P.second; 2171 LSRUse &LU = Uses[LF.LUIdx]; 2172 LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L); 2173 if (!LU.WidestFixupType || 2174 SE.getTypeSizeInBits(LU.WidestFixupType) < 2175 SE.getTypeSizeInBits(LF.OperandValToReplace->getType())) 2176 LU.WidestFixupType = LF.OperandValToReplace->getType(); 2177 2178 // If this is the first use of this LSRUse, give it a formula. 2179 if (LU.Formulae.empty()) { 2180 InsertInitialFormula(S, LU, LF.LUIdx); 2181 CountRegisters(LU.Formulae.back(), LF.LUIdx); 2182 } 2183 } 2184 2185 DEBUG(print_fixups(dbgs())); 2186} 2187 2188/// InsertInitialFormula - Insert a formula for the given expression into 2189/// the given use, separating out loop-variant portions from loop-invariant 2190/// and loop-computable portions. 2191void 2192LSRInstance::InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx) { 2193 Formula F; 2194 F.InitialMatch(S, L, SE); 2195 bool Inserted = InsertFormula(LU, LUIdx, F); 2196 assert(Inserted && "Initial formula already exists!"); (void)Inserted; 2197} 2198 2199/// InsertSupplementalFormula - Insert a simple single-register formula for 2200/// the given expression into the given use. 2201void 2202LSRInstance::InsertSupplementalFormula(const SCEV *S, 2203 LSRUse &LU, size_t LUIdx) { 2204 Formula F; 2205 F.BaseRegs.push_back(S); 2206 F.AM.HasBaseReg = true; 2207 bool Inserted = InsertFormula(LU, LUIdx, F); 2208 assert(Inserted && "Supplemental formula already exists!"); (void)Inserted; 2209} 2210 2211/// CountRegisters - Note which registers are used by the given formula, 2212/// updating RegUses. 2213void LSRInstance::CountRegisters(const Formula &F, size_t LUIdx) { 2214 if (F.ScaledReg) 2215 RegUses.CountRegister(F.ScaledReg, LUIdx); 2216 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(), 2217 E = F.BaseRegs.end(); I != E; ++I) 2218 RegUses.CountRegister(*I, LUIdx); 2219} 2220 2221/// InsertFormula - If the given formula has not yet been inserted, add it to 2222/// the list, and return true. Return false otherwise. 2223bool LSRInstance::InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F) { 2224 if (!LU.InsertFormula(F)) 2225 return false; 2226 2227 CountRegisters(F, LUIdx); 2228 return true; 2229} 2230 2231/// CollectLoopInvariantFixupsAndFormulae - Check for other uses of 2232/// loop-invariant values which we're tracking. These other uses will pin these 2233/// values in registers, making them less profitable for elimination. 2234/// TODO: This currently misses non-constant addrec step registers. 2235/// TODO: Should this give more weight to users inside the loop? 2236void 2237LSRInstance::CollectLoopInvariantFixupsAndFormulae() { 2238 SmallVector<const SCEV *, 8> Worklist(RegUses.begin(), RegUses.end()); 2239 SmallPtrSet<const SCEV *, 8> Inserted; 2240 2241 while (!Worklist.empty()) { 2242 const SCEV *S = Worklist.pop_back_val(); 2243 2244 if (const SCEVNAryExpr *N = dyn_cast<SCEVNAryExpr>(S)) 2245 Worklist.append(N->op_begin(), N->op_end()); 2246 else if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(S)) 2247 Worklist.push_back(C->getOperand()); 2248 else if (const SCEVUDivExpr *D = dyn_cast<SCEVUDivExpr>(S)) { 2249 Worklist.push_back(D->getLHS()); 2250 Worklist.push_back(D->getRHS()); 2251 } else if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) { 2252 if (!Inserted.insert(U)) continue; 2253 const Value *V = U->getValue(); 2254 if (const Instruction *Inst = dyn_cast<Instruction>(V)) { 2255 // Look for instructions defined outside the loop. 2256 if (L->contains(Inst)) continue; 2257 } else if (isa<UndefValue>(V)) 2258 // Undef doesn't have a live range, so it doesn't matter. 2259 continue; 2260 for (Value::const_use_iterator UI = V->use_begin(), UE = V->use_end(); 2261 UI != UE; ++UI) { 2262 const Instruction *UserInst = dyn_cast<Instruction>(*UI); 2263 // Ignore non-instructions. 2264 if (!UserInst) 2265 continue; 2266 // Ignore instructions in other functions (as can happen with 2267 // Constants). 2268 if (UserInst->getParent()->getParent() != L->getHeader()->getParent()) 2269 continue; 2270 // Ignore instructions not dominated by the loop. 2271 const BasicBlock *UseBB = !isa<PHINode>(UserInst) ? 2272 UserInst->getParent() : 2273 cast<PHINode>(UserInst)->getIncomingBlock( 2274 PHINode::getIncomingValueNumForOperand(UI.getOperandNo())); 2275 if (!DT.dominates(L->getHeader(), UseBB)) 2276 continue; 2277 // Ignore uses which are part of other SCEV expressions, to avoid 2278 // analyzing them multiple times. 2279 if (SE.isSCEVable(UserInst->getType())) { 2280 const SCEV *UserS = SE.getSCEV(const_cast<Instruction *>(UserInst)); 2281 // If the user is a no-op, look through to its uses. 2282 if (!isa<SCEVUnknown>(UserS)) 2283 continue; 2284 if (UserS == U) { 2285 Worklist.push_back( 2286 SE.getUnknown(const_cast<Instruction *>(UserInst))); 2287 continue; 2288 } 2289 } 2290 // Ignore icmp instructions which are already being analyzed. 2291 if (const ICmpInst *ICI = dyn_cast<ICmpInst>(UserInst)) { 2292 unsigned OtherIdx = !UI.getOperandNo(); 2293 Value *OtherOp = const_cast<Value *>(ICI->getOperand(OtherIdx)); 2294 if (SE.hasComputableLoopEvolution(SE.getSCEV(OtherOp), L)) 2295 continue; 2296 } 2297 2298 LSRFixup &LF = getNewFixup(); 2299 LF.UserInst = const_cast<Instruction *>(UserInst); 2300 LF.OperandValToReplace = UI.getUse(); 2301 std::pair<size_t, int64_t> P = getUse(S, LSRUse::Basic, 0); 2302 LF.LUIdx = P.first; 2303 LF.Offset = P.second; 2304 LSRUse &LU = Uses[LF.LUIdx]; 2305 LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L); 2306 if (!LU.WidestFixupType || 2307 SE.getTypeSizeInBits(LU.WidestFixupType) < 2308 SE.getTypeSizeInBits(LF.OperandValToReplace->getType())) 2309 LU.WidestFixupType = LF.OperandValToReplace->getType(); 2310 InsertSupplementalFormula(U, LU, LF.LUIdx); 2311 CountRegisters(LU.Formulae.back(), Uses.size() - 1); 2312 break; 2313 } 2314 } 2315 } 2316} 2317 2318/// CollectSubexprs - Split S into subexpressions which can be pulled out into 2319/// separate registers. If C is non-null, multiply each subexpression by C. 2320static void CollectSubexprs(const SCEV *S, const SCEVConstant *C, 2321 SmallVectorImpl<const SCEV *> &Ops, 2322 const Loop *L, 2323 ScalarEvolution &SE) { 2324 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) { 2325 // Break out add operands. 2326 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end(); 2327 I != E; ++I) 2328 CollectSubexprs(*I, C, Ops, L, SE); 2329 return; 2330 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) { 2331 // Split a non-zero base out of an addrec. 2332 if (!AR->getStart()->isZero()) { 2333 CollectSubexprs(SE.getAddRecExpr(SE.getConstant(AR->getType(), 0), 2334 AR->getStepRecurrence(SE), 2335 AR->getLoop(), 2336 //FIXME: AR->getNoWrapFlags(SCEV::FlagNW) 2337 SCEV::FlagAnyWrap), 2338 C, Ops, L, SE); 2339 CollectSubexprs(AR->getStart(), C, Ops, L, SE); 2340 return; 2341 } 2342 } else if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) { 2343 // Break (C * (a + b + c)) into C*a + C*b + C*c. 2344 if (Mul->getNumOperands() == 2) 2345 if (const SCEVConstant *Op0 = 2346 dyn_cast<SCEVConstant>(Mul->getOperand(0))) { 2347 CollectSubexprs(Mul->getOperand(1), 2348 C ? cast<SCEVConstant>(SE.getMulExpr(C, Op0)) : Op0, 2349 Ops, L, SE); 2350 return; 2351 } 2352 } 2353 2354 // Otherwise use the value itself, optionally with a scale applied. 2355 Ops.push_back(C ? SE.getMulExpr(C, S) : S); 2356} 2357 2358/// GenerateReassociations - Split out subexpressions from adds and the bases of 2359/// addrecs. 2360void LSRInstance::GenerateReassociations(LSRUse &LU, unsigned LUIdx, 2361 Formula Base, 2362 unsigned Depth) { 2363 // Arbitrarily cap recursion to protect compile time. 2364 if (Depth >= 3) return; 2365 2366 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) { 2367 const SCEV *BaseReg = Base.BaseRegs[i]; 2368 2369 SmallVector<const SCEV *, 8> AddOps; 2370 CollectSubexprs(BaseReg, 0, AddOps, L, SE); 2371 2372 if (AddOps.size() == 1) continue; 2373 2374 for (SmallVectorImpl<const SCEV *>::const_iterator J = AddOps.begin(), 2375 JE = AddOps.end(); J != JE; ++J) { 2376 2377 // Loop-variant "unknown" values are uninteresting; we won't be able to 2378 // do anything meaningful with them. 2379 if (isa<SCEVUnknown>(*J) && !SE.isLoopInvariant(*J, L)) 2380 continue; 2381 2382 // Don't pull a constant into a register if the constant could be folded 2383 // into an immediate field. 2384 if (isAlwaysFoldable(*J, LU.MinOffset, LU.MaxOffset, 2385 Base.getNumRegs() > 1, 2386 LU.Kind, LU.AccessTy, TLI, SE)) 2387 continue; 2388 2389 // Collect all operands except *J. 2390 SmallVector<const SCEV *, 8> InnerAddOps 2391 (((const SmallVector<const SCEV *, 8> &)AddOps).begin(), J); 2392 InnerAddOps.append 2393 (llvm::next(J), ((const SmallVector<const SCEV *, 8> &)AddOps).end()); 2394 2395 // Don't leave just a constant behind in a register if the constant could 2396 // be folded into an immediate field. 2397 if (InnerAddOps.size() == 1 && 2398 isAlwaysFoldable(InnerAddOps[0], LU.MinOffset, LU.MaxOffset, 2399 Base.getNumRegs() > 1, 2400 LU.Kind, LU.AccessTy, TLI, SE)) 2401 continue; 2402 2403 const SCEV *InnerSum = SE.getAddExpr(InnerAddOps); 2404 if (InnerSum->isZero()) 2405 continue; 2406 Formula F = Base; 2407 2408 // Add the remaining pieces of the add back into the new formula. 2409 const SCEVConstant *InnerSumSC = dyn_cast<SCEVConstant>(InnerSum); 2410 if (TLI && InnerSumSC && 2411 SE.getTypeSizeInBits(InnerSumSC->getType()) <= 64 && 2412 TLI->isLegalAddImmediate((uint64_t)F.UnfoldedOffset + 2413 InnerSumSC->getValue()->getZExtValue())) { 2414 F.UnfoldedOffset = (uint64_t)F.UnfoldedOffset + 2415 InnerSumSC->getValue()->getZExtValue(); 2416 F.BaseRegs.erase(F.BaseRegs.begin() + i); 2417 } else 2418 F.BaseRegs[i] = InnerSum; 2419 2420 // Add J as its own register, or an unfolded immediate. 2421 const SCEVConstant *SC = dyn_cast<SCEVConstant>(*J); 2422 if (TLI && SC && SE.getTypeSizeInBits(SC->getType()) <= 64 && 2423 TLI->isLegalAddImmediate((uint64_t)F.UnfoldedOffset + 2424 SC->getValue()->getZExtValue())) 2425 F.UnfoldedOffset = (uint64_t)F.UnfoldedOffset + 2426 SC->getValue()->getZExtValue(); 2427 else 2428 F.BaseRegs.push_back(*J); 2429 2430 if (InsertFormula(LU, LUIdx, F)) 2431 // If that formula hadn't been seen before, recurse to find more like 2432 // it. 2433 GenerateReassociations(LU, LUIdx, LU.Formulae.back(), Depth+1); 2434 } 2435 } 2436} 2437 2438/// GenerateCombinations - Generate a formula consisting of all of the 2439/// loop-dominating registers added into a single register. 2440void LSRInstance::GenerateCombinations(LSRUse &LU, unsigned LUIdx, 2441 Formula Base) { 2442 // This method is only interesting on a plurality of registers. 2443 if (Base.BaseRegs.size() <= 1) return; 2444 2445 Formula F = Base; 2446 F.BaseRegs.clear(); 2447 SmallVector<const SCEV *, 4> Ops; 2448 for (SmallVectorImpl<const SCEV *>::const_iterator 2449 I = Base.BaseRegs.begin(), E = Base.BaseRegs.end(); I != E; ++I) { 2450 const SCEV *BaseReg = *I; 2451 if (SE.properlyDominates(BaseReg, L->getHeader()) && 2452 !SE.hasComputableLoopEvolution(BaseReg, L)) 2453 Ops.push_back(BaseReg); 2454 else 2455 F.BaseRegs.push_back(BaseReg); 2456 } 2457 if (Ops.size() > 1) { 2458 const SCEV *Sum = SE.getAddExpr(Ops); 2459 // TODO: If Sum is zero, it probably means ScalarEvolution missed an 2460 // opportunity to fold something. For now, just ignore such cases 2461 // rather than proceed with zero in a register. 2462 if (!Sum->isZero()) { 2463 F.BaseRegs.push_back(Sum); 2464 (void)InsertFormula(LU, LUIdx, F); 2465 } 2466 } 2467} 2468 2469/// GenerateSymbolicOffsets - Generate reuse formulae using symbolic offsets. 2470void LSRInstance::GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx, 2471 Formula Base) { 2472 // We can't add a symbolic offset if the address already contains one. 2473 if (Base.AM.BaseGV) return; 2474 2475 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) { 2476 const SCEV *G = Base.BaseRegs[i]; 2477 GlobalValue *GV = ExtractSymbol(G, SE); 2478 if (G->isZero() || !GV) 2479 continue; 2480 Formula F = Base; 2481 F.AM.BaseGV = GV; 2482 if (!isLegalUse(F.AM, LU.MinOffset, LU.MaxOffset, 2483 LU.Kind, LU.AccessTy, TLI)) 2484 continue; 2485 F.BaseRegs[i] = G; 2486 (void)InsertFormula(LU, LUIdx, F); 2487 } 2488} 2489 2490/// GenerateConstantOffsets - Generate reuse formulae using symbolic offsets. 2491void LSRInstance::GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx, 2492 Formula Base) { 2493 // TODO: For now, just add the min and max offset, because it usually isn't 2494 // worthwhile looking at everything inbetween. 2495 SmallVector<int64_t, 2> Worklist; 2496 Worklist.push_back(LU.MinOffset); 2497 if (LU.MaxOffset != LU.MinOffset) 2498 Worklist.push_back(LU.MaxOffset); 2499 2500 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) { 2501 const SCEV *G = Base.BaseRegs[i]; 2502 2503 for (SmallVectorImpl<int64_t>::const_iterator I = Worklist.begin(), 2504 E = Worklist.end(); I != E; ++I) { 2505 Formula F = Base; 2506 F.AM.BaseOffs = (uint64_t)Base.AM.BaseOffs - *I; 2507 if (isLegalUse(F.AM, LU.MinOffset - *I, LU.MaxOffset - *I, 2508 LU.Kind, LU.AccessTy, TLI)) { 2509 // Add the offset to the base register. 2510 const SCEV *NewG = SE.getAddExpr(SE.getConstant(G->getType(), *I), G); 2511 // If it cancelled out, drop the base register, otherwise update it. 2512 if (NewG->isZero()) { 2513 std::swap(F.BaseRegs[i], F.BaseRegs.back()); 2514 F.BaseRegs.pop_back(); 2515 } else 2516 F.BaseRegs[i] = NewG; 2517 2518 (void)InsertFormula(LU, LUIdx, F); 2519 } 2520 } 2521 2522 int64_t Imm = ExtractImmediate(G, SE); 2523 if (G->isZero() || Imm == 0) 2524 continue; 2525 Formula F = Base; 2526 F.AM.BaseOffs = (uint64_t)F.AM.BaseOffs + Imm; 2527 if (!isLegalUse(F.AM, LU.MinOffset, LU.MaxOffset, 2528 LU.Kind, LU.AccessTy, TLI)) 2529 continue; 2530 F.BaseRegs[i] = G; 2531 (void)InsertFormula(LU, LUIdx, F); 2532 } 2533} 2534 2535/// GenerateICmpZeroScales - For ICmpZero, check to see if we can scale up 2536/// the comparison. For example, x == y -> x*c == y*c. 2537void LSRInstance::GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx, 2538 Formula Base) { 2539 if (LU.Kind != LSRUse::ICmpZero) return; 2540 2541 // Determine the integer type for the base formula. 2542 Type *IntTy = Base.getType(); 2543 if (!IntTy) return; 2544 if (SE.getTypeSizeInBits(IntTy) > 64) return; 2545 2546 // Don't do this if there is more than one offset. 2547 if (LU.MinOffset != LU.MaxOffset) return; 2548 2549 assert(!Base.AM.BaseGV && "ICmpZero use is not legal!"); 2550 2551 // Check each interesting stride. 2552 for (SmallSetVector<int64_t, 8>::const_iterator 2553 I = Factors.begin(), E = Factors.end(); I != E; ++I) { 2554 int64_t Factor = *I; 2555 2556 // Check that the multiplication doesn't overflow. 2557 if (Base.AM.BaseOffs == INT64_MIN && Factor == -1) 2558 continue; 2559 int64_t NewBaseOffs = (uint64_t)Base.AM.BaseOffs * Factor; 2560 if (NewBaseOffs / Factor != Base.AM.BaseOffs) 2561 continue; 2562 2563 // Check that multiplying with the use offset doesn't overflow. 2564 int64_t Offset = LU.MinOffset; 2565 if (Offset == INT64_MIN && Factor == -1) 2566 continue; 2567 Offset = (uint64_t)Offset * Factor; 2568 if (Offset / Factor != LU.MinOffset) 2569 continue; 2570 2571 Formula F = Base; 2572 F.AM.BaseOffs = NewBaseOffs; 2573 2574 // Check that this scale is legal. 2575 if (!isLegalUse(F.AM, Offset, Offset, LU.Kind, LU.AccessTy, TLI)) 2576 continue; 2577 2578 // Compensate for the use having MinOffset built into it. 2579 F.AM.BaseOffs = (uint64_t)F.AM.BaseOffs + Offset - LU.MinOffset; 2580 2581 const SCEV *FactorS = SE.getConstant(IntTy, Factor); 2582 2583 // Check that multiplying with each base register doesn't overflow. 2584 for (size_t i = 0, e = F.BaseRegs.size(); i != e; ++i) { 2585 F.BaseRegs[i] = SE.getMulExpr(F.BaseRegs[i], FactorS); 2586 if (getExactSDiv(F.BaseRegs[i], FactorS, SE) != Base.BaseRegs[i]) 2587 goto next; 2588 } 2589 2590 // Check that multiplying with the scaled register doesn't overflow. 2591 if (F.ScaledReg) { 2592 F.ScaledReg = SE.getMulExpr(F.ScaledReg, FactorS); 2593 if (getExactSDiv(F.ScaledReg, FactorS, SE) != Base.ScaledReg) 2594 continue; 2595 } 2596 2597 // Check that multiplying with the unfolded offset doesn't overflow. 2598 if (F.UnfoldedOffset != 0) { 2599 if (F.UnfoldedOffset == INT64_MIN && Factor == -1) 2600 continue; 2601 F.UnfoldedOffset = (uint64_t)F.UnfoldedOffset * Factor; 2602 if (F.UnfoldedOffset / Factor != Base.UnfoldedOffset) 2603 continue; 2604 } 2605 2606 // If we make it here and it's legal, add it. 2607 (void)InsertFormula(LU, LUIdx, F); 2608 next:; 2609 } 2610} 2611 2612/// GenerateScales - Generate stride factor reuse formulae by making use of 2613/// scaled-offset address modes, for example. 2614void LSRInstance::GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base) { 2615 // Determine the integer type for the base formula. 2616 Type *IntTy = Base.getType(); 2617 if (!IntTy) return; 2618 2619 // If this Formula already has a scaled register, we can't add another one. 2620 if (Base.AM.Scale != 0) return; 2621 2622 // Check each interesting stride. 2623 for (SmallSetVector<int64_t, 8>::const_iterator 2624 I = Factors.begin(), E = Factors.end(); I != E; ++I) { 2625 int64_t Factor = *I; 2626 2627 Base.AM.Scale = Factor; 2628 Base.AM.HasBaseReg = Base.BaseRegs.size() > 1; 2629 // Check whether this scale is going to be legal. 2630 if (!isLegalUse(Base.AM, LU.MinOffset, LU.MaxOffset, 2631 LU.Kind, LU.AccessTy, TLI)) { 2632 // As a special-case, handle special out-of-loop Basic users specially. 2633 // TODO: Reconsider this special case. 2634 if (LU.Kind == LSRUse::Basic && 2635 isLegalUse(Base.AM, LU.MinOffset, LU.MaxOffset, 2636 LSRUse::Special, LU.AccessTy, TLI) && 2637 LU.AllFixupsOutsideLoop) 2638 LU.Kind = LSRUse::Special; 2639 else 2640 continue; 2641 } 2642 // For an ICmpZero, negating a solitary base register won't lead to 2643 // new solutions. 2644 if (LU.Kind == LSRUse::ICmpZero && 2645 !Base.AM.HasBaseReg && Base.AM.BaseOffs == 0 && !Base.AM.BaseGV) 2646 continue; 2647 // For each addrec base reg, apply the scale, if possible. 2648 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) 2649 if (const SCEVAddRecExpr *AR = 2650 dyn_cast<SCEVAddRecExpr>(Base.BaseRegs[i])) { 2651 const SCEV *FactorS = SE.getConstant(IntTy, Factor); 2652 if (FactorS->isZero()) 2653 continue; 2654 // Divide out the factor, ignoring high bits, since we'll be 2655 // scaling the value back up in the end. 2656 if (const SCEV *Quotient = getExactSDiv(AR, FactorS, SE, true)) { 2657 // TODO: This could be optimized to avoid all the copying. 2658 Formula F = Base; 2659 F.ScaledReg = Quotient; 2660 F.DeleteBaseReg(F.BaseRegs[i]); 2661 (void)InsertFormula(LU, LUIdx, F); 2662 } 2663 } 2664 } 2665} 2666 2667/// GenerateTruncates - Generate reuse formulae from different IV types. 2668void LSRInstance::GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base) { 2669 // This requires TargetLowering to tell us which truncates are free. 2670 if (!TLI) return; 2671 2672 // Don't bother truncating symbolic values. 2673 if (Base.AM.BaseGV) return; 2674 2675 // Determine the integer type for the base formula. 2676 Type *DstTy = Base.getType(); 2677 if (!DstTy) return; 2678 DstTy = SE.getEffectiveSCEVType(DstTy); 2679 2680 for (SmallSetVector<Type *, 4>::const_iterator 2681 I = Types.begin(), E = Types.end(); I != E; ++I) { 2682 Type *SrcTy = *I; 2683 if (SrcTy != DstTy && TLI->isTruncateFree(SrcTy, DstTy)) { 2684 Formula F = Base; 2685 2686 if (F.ScaledReg) F.ScaledReg = SE.getAnyExtendExpr(F.ScaledReg, *I); 2687 for (SmallVectorImpl<const SCEV *>::iterator J = F.BaseRegs.begin(), 2688 JE = F.BaseRegs.end(); J != JE; ++J) 2689 *J = SE.getAnyExtendExpr(*J, SrcTy); 2690 2691 // TODO: This assumes we've done basic processing on all uses and 2692 // have an idea what the register usage is. 2693 if (!F.hasRegsUsedByUsesOtherThan(LUIdx, RegUses)) 2694 continue; 2695 2696 (void)InsertFormula(LU, LUIdx, F); 2697 } 2698 } 2699} 2700 2701namespace { 2702 2703/// WorkItem - Helper class for GenerateCrossUseConstantOffsets. It's used to 2704/// defer modifications so that the search phase doesn't have to worry about 2705/// the data structures moving underneath it. 2706struct WorkItem { 2707 size_t LUIdx; 2708 int64_t Imm; 2709 const SCEV *OrigReg; 2710 2711 WorkItem(size_t LI, int64_t I, const SCEV *R) 2712 : LUIdx(LI), Imm(I), OrigReg(R) {} 2713 2714 void print(raw_ostream &OS) const; 2715 void dump() const; 2716}; 2717 2718} 2719 2720void WorkItem::print(raw_ostream &OS) const { 2721 OS << "in formulae referencing " << *OrigReg << " in use " << LUIdx 2722 << " , add offset " << Imm; 2723} 2724 2725void WorkItem::dump() const { 2726 print(errs()); errs() << '\n'; 2727} 2728 2729/// GenerateCrossUseConstantOffsets - Look for registers which are a constant 2730/// distance apart and try to form reuse opportunities between them. 2731void LSRInstance::GenerateCrossUseConstantOffsets() { 2732 // Group the registers by their value without any added constant offset. 2733 typedef std::map<int64_t, const SCEV *> ImmMapTy; 2734 typedef DenseMap<const SCEV *, ImmMapTy> RegMapTy; 2735 RegMapTy Map; 2736 DenseMap<const SCEV *, SmallBitVector> UsedByIndicesMap; 2737 SmallVector<const SCEV *, 8> Sequence; 2738 for (RegUseTracker::const_iterator I = RegUses.begin(), E = RegUses.end(); 2739 I != E; ++I) { 2740 const SCEV *Reg = *I; 2741 int64_t Imm = ExtractImmediate(Reg, SE); 2742 std::pair<RegMapTy::iterator, bool> Pair = 2743 Map.insert(std::make_pair(Reg, ImmMapTy())); 2744 if (Pair.second) 2745 Sequence.push_back(Reg); 2746 Pair.first->second.insert(std::make_pair(Imm, *I)); 2747 UsedByIndicesMap[Reg] |= RegUses.getUsedByIndices(*I); 2748 } 2749 2750 // Now examine each set of registers with the same base value. Build up 2751 // a list of work to do and do the work in a separate step so that we're 2752 // not adding formulae and register counts while we're searching. 2753 SmallVector<WorkItem, 32> WorkItems; 2754 SmallSet<std::pair<size_t, int64_t>, 32> UniqueItems; 2755 for (SmallVectorImpl<const SCEV *>::const_iterator I = Sequence.begin(), 2756 E = Sequence.end(); I != E; ++I) { 2757 const SCEV *Reg = *I; 2758 const ImmMapTy &Imms = Map.find(Reg)->second; 2759 2760 // It's not worthwhile looking for reuse if there's only one offset. 2761 if (Imms.size() == 1) 2762 continue; 2763 2764 DEBUG(dbgs() << "Generating cross-use offsets for " << *Reg << ':'; 2765 for (ImmMapTy::const_iterator J = Imms.begin(), JE = Imms.end(); 2766 J != JE; ++J) 2767 dbgs() << ' ' << J->first; 2768 dbgs() << '\n'); 2769 2770 // Examine each offset. 2771 for (ImmMapTy::const_iterator J = Imms.begin(), JE = Imms.end(); 2772 J != JE; ++J) { 2773 const SCEV *OrigReg = J->second; 2774 2775 int64_t JImm = J->first; 2776 const SmallBitVector &UsedByIndices = RegUses.getUsedByIndices(OrigReg); 2777 2778 if (!isa<SCEVConstant>(OrigReg) && 2779 UsedByIndicesMap[Reg].count() == 1) { 2780 DEBUG(dbgs() << "Skipping cross-use reuse for " << *OrigReg << '\n'); 2781 continue; 2782 } 2783 2784 // Conservatively examine offsets between this orig reg a few selected 2785 // other orig regs. 2786 ImmMapTy::const_iterator OtherImms[] = { 2787 Imms.begin(), prior(Imms.end()), 2788 Imms.lower_bound((Imms.begin()->first + prior(Imms.end())->first) / 2) 2789 }; 2790 for (size_t i = 0, e = array_lengthof(OtherImms); i != e; ++i) { 2791 ImmMapTy::const_iterator M = OtherImms[i]; 2792 if (M == J || M == JE) continue; 2793 2794 // Compute the difference between the two. 2795 int64_t Imm = (uint64_t)JImm - M->first; 2796 for (int LUIdx = UsedByIndices.find_first(); LUIdx != -1; 2797 LUIdx = UsedByIndices.find_next(LUIdx)) 2798 // Make a memo of this use, offset, and register tuple. 2799 if (UniqueItems.insert(std::make_pair(LUIdx, Imm))) 2800 WorkItems.push_back(WorkItem(LUIdx, Imm, OrigReg)); 2801 } 2802 } 2803 } 2804 2805 Map.clear(); 2806 Sequence.clear(); 2807 UsedByIndicesMap.clear(); 2808 UniqueItems.clear(); 2809 2810 // Now iterate through the worklist and add new formulae. 2811 for (SmallVectorImpl<WorkItem>::const_iterator I = WorkItems.begin(), 2812 E = WorkItems.end(); I != E; ++I) { 2813 const WorkItem &WI = *I; 2814 size_t LUIdx = WI.LUIdx; 2815 LSRUse &LU = Uses[LUIdx]; 2816 int64_t Imm = WI.Imm; 2817 const SCEV *OrigReg = WI.OrigReg; 2818 2819 Type *IntTy = SE.getEffectiveSCEVType(OrigReg->getType()); 2820 const SCEV *NegImmS = SE.getSCEV(ConstantInt::get(IntTy, -(uint64_t)Imm)); 2821 unsigned BitWidth = SE.getTypeSizeInBits(IntTy); 2822 2823 // TODO: Use a more targeted data structure. 2824 for (size_t L = 0, LE = LU.Formulae.size(); L != LE; ++L) { 2825 const Formula &F = LU.Formulae[L]; 2826 // Use the immediate in the scaled register. 2827 if (F.ScaledReg == OrigReg) { 2828 int64_t Offs = (uint64_t)F.AM.BaseOffs + 2829 Imm * (uint64_t)F.AM.Scale; 2830 // Don't create 50 + reg(-50). 2831 if (F.referencesReg(SE.getSCEV( 2832 ConstantInt::get(IntTy, -(uint64_t)Offs)))) 2833 continue; 2834 Formula NewF = F; 2835 NewF.AM.BaseOffs = Offs; 2836 if (!isLegalUse(NewF.AM, LU.MinOffset, LU.MaxOffset, 2837 LU.Kind, LU.AccessTy, TLI)) 2838 continue; 2839 NewF.ScaledReg = SE.getAddExpr(NegImmS, NewF.ScaledReg); 2840 2841 // If the new scale is a constant in a register, and adding the constant 2842 // value to the immediate would produce a value closer to zero than the 2843 // immediate itself, then the formula isn't worthwhile. 2844 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(NewF.ScaledReg)) 2845 if (C->getValue()->isNegative() != 2846 (NewF.AM.BaseOffs < 0) && 2847 (C->getValue()->getValue().abs() * APInt(BitWidth, F.AM.Scale)) 2848 .ule(abs64(NewF.AM.BaseOffs))) 2849 continue; 2850 2851 // OK, looks good. 2852 (void)InsertFormula(LU, LUIdx, NewF); 2853 } else { 2854 // Use the immediate in a base register. 2855 for (size_t N = 0, NE = F.BaseRegs.size(); N != NE; ++N) { 2856 const SCEV *BaseReg = F.BaseRegs[N]; 2857 if (BaseReg != OrigReg) 2858 continue; 2859 Formula NewF = F; 2860 NewF.AM.BaseOffs = (uint64_t)NewF.AM.BaseOffs + Imm; 2861 if (!isLegalUse(NewF.AM, LU.MinOffset, LU.MaxOffset, 2862 LU.Kind, LU.AccessTy, TLI)) { 2863 if (!TLI || 2864 !TLI->isLegalAddImmediate((uint64_t)NewF.UnfoldedOffset + Imm)) 2865 continue; 2866 NewF = F; 2867 NewF.UnfoldedOffset = (uint64_t)NewF.UnfoldedOffset + Imm; 2868 } 2869 NewF.BaseRegs[N] = SE.getAddExpr(NegImmS, BaseReg); 2870 2871 // If the new formula has a constant in a register, and adding the 2872 // constant value to the immediate would produce a value closer to 2873 // zero than the immediate itself, then the formula isn't worthwhile. 2874 for (SmallVectorImpl<const SCEV *>::const_iterator 2875 J = NewF.BaseRegs.begin(), JE = NewF.BaseRegs.end(); 2876 J != JE; ++J) 2877 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(*J)) 2878 if ((C->getValue()->getValue() + NewF.AM.BaseOffs).abs().slt( 2879 abs64(NewF.AM.BaseOffs)) && 2880 (C->getValue()->getValue() + 2881 NewF.AM.BaseOffs).countTrailingZeros() >= 2882 CountTrailingZeros_64(NewF.AM.BaseOffs)) 2883 goto skip_formula; 2884 2885 // Ok, looks good. 2886 (void)InsertFormula(LU, LUIdx, NewF); 2887 break; 2888 skip_formula:; 2889 } 2890 } 2891 } 2892 } 2893} 2894 2895/// GenerateAllReuseFormulae - Generate formulae for each use. 2896void 2897LSRInstance::GenerateAllReuseFormulae() { 2898 // This is split into multiple loops so that hasRegsUsedByUsesOtherThan 2899 // queries are more precise. 2900 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) { 2901 LSRUse &LU = Uses[LUIdx]; 2902 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i) 2903 GenerateReassociations(LU, LUIdx, LU.Formulae[i]); 2904 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i) 2905 GenerateCombinations(LU, LUIdx, LU.Formulae[i]); 2906 } 2907 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) { 2908 LSRUse &LU = Uses[LUIdx]; 2909 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i) 2910 GenerateSymbolicOffsets(LU, LUIdx, LU.Formulae[i]); 2911 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i) 2912 GenerateConstantOffsets(LU, LUIdx, LU.Formulae[i]); 2913 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i) 2914 GenerateICmpZeroScales(LU, LUIdx, LU.Formulae[i]); 2915 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i) 2916 GenerateScales(LU, LUIdx, LU.Formulae[i]); 2917 } 2918 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) { 2919 LSRUse &LU = Uses[LUIdx]; 2920 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i) 2921 GenerateTruncates(LU, LUIdx, LU.Formulae[i]); 2922 } 2923 2924 GenerateCrossUseConstantOffsets(); 2925 2926 DEBUG(dbgs() << "\n" 2927 "After generating reuse formulae:\n"; 2928 print_uses(dbgs())); 2929} 2930 2931/// If there are multiple formulae with the same set of registers used 2932/// by other uses, pick the best one and delete the others. 2933void LSRInstance::FilterOutUndesirableDedicatedRegisters() { 2934 DenseSet<const SCEV *> VisitedRegs; 2935 SmallPtrSet<const SCEV *, 16> Regs; 2936 SmallPtrSet<const SCEV *, 16> LoserRegs; 2937#ifndef NDEBUG 2938 bool ChangedFormulae = false; 2939#endif 2940 2941 // Collect the best formula for each unique set of shared registers. This 2942 // is reset for each use. 2943 typedef DenseMap<SmallVector<const SCEV *, 2>, size_t, UniquifierDenseMapInfo> 2944 BestFormulaeTy; 2945 BestFormulaeTy BestFormulae; 2946 2947 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) { 2948 LSRUse &LU = Uses[LUIdx]; 2949 DEBUG(dbgs() << "Filtering for use "; LU.print(dbgs()); dbgs() << '\n'); 2950 2951 bool Any = false; 2952 for (size_t FIdx = 0, NumForms = LU.Formulae.size(); 2953 FIdx != NumForms; ++FIdx) { 2954 Formula &F = LU.Formulae[FIdx]; 2955 2956 // Some formulas are instant losers. For example, they may depend on 2957 // nonexistent AddRecs from other loops. These need to be filtered 2958 // immediately, otherwise heuristics could choose them over others leading 2959 // to an unsatisfactory solution. Passing LoserRegs into RateFormula here 2960 // avoids the need to recompute this information across formulae using the 2961 // same bad AddRec. Passing LoserRegs is also essential unless we remove 2962 // the corresponding bad register from the Regs set. 2963 Cost CostF; 2964 Regs.clear(); 2965 CostF.RateFormula(F, Regs, VisitedRegs, L, LU.Offsets, SE, DT, 2966 &LoserRegs); 2967 if (CostF.isLoser()) { 2968 // During initial formula generation, undesirable formulae are generated 2969 // by uses within other loops that have some non-trivial address mode or 2970 // use the postinc form of the IV. LSR needs to provide these formulae 2971 // as the basis of rediscovering the desired formula that uses an AddRec 2972 // corresponding to the existing phi. Once all formulae have been 2973 // generated, these initial losers may be pruned. 2974 DEBUG(dbgs() << " Filtering loser "; F.print(dbgs()); 2975 dbgs() << "\n"); 2976 } 2977 else { 2978 SmallVector<const SCEV *, 2> Key; 2979 for (SmallVectorImpl<const SCEV *>::const_iterator J = F.BaseRegs.begin(), 2980 JE = F.BaseRegs.end(); J != JE; ++J) { 2981 const SCEV *Reg = *J; 2982 if (RegUses.isRegUsedByUsesOtherThan(Reg, LUIdx)) 2983 Key.push_back(Reg); 2984 } 2985 if (F.ScaledReg && 2986 RegUses.isRegUsedByUsesOtherThan(F.ScaledReg, LUIdx)) 2987 Key.push_back(F.ScaledReg); 2988 // Unstable sort by host order ok, because this is only used for 2989 // uniquifying. 2990 std::sort(Key.begin(), Key.end()); 2991 2992 std::pair<BestFormulaeTy::const_iterator, bool> P = 2993 BestFormulae.insert(std::make_pair(Key, FIdx)); 2994 if (P.second) 2995 continue; 2996 2997 Formula &Best = LU.Formulae[P.first->second]; 2998 2999 Cost CostBest; 3000 Regs.clear(); 3001 CostBest.RateFormula(Best, Regs, VisitedRegs, L, LU.Offsets, SE, DT); 3002 if (CostF < CostBest) 3003 std::swap(F, Best); 3004 DEBUG(dbgs() << " Filtering out formula "; F.print(dbgs()); 3005 dbgs() << "\n" 3006 " in favor of formula "; Best.print(dbgs()); 3007 dbgs() << '\n'); 3008 } 3009#ifndef NDEBUG 3010 ChangedFormulae = true; 3011#endif 3012 LU.DeleteFormula(F); 3013 --FIdx; 3014 --NumForms; 3015 Any = true; 3016 } 3017 3018 // Now that we've filtered out some formulae, recompute the Regs set. 3019 if (Any) 3020 LU.RecomputeRegs(LUIdx, RegUses); 3021 3022 // Reset this to prepare for the next use. 3023 BestFormulae.clear(); 3024 } 3025 3026 DEBUG(if (ChangedFormulae) { 3027 dbgs() << "\n" 3028 "After filtering out undesirable candidates:\n"; 3029 print_uses(dbgs()); 3030 }); 3031} 3032 3033// This is a rough guess that seems to work fairly well. 3034static const size_t ComplexityLimit = UINT16_MAX; 3035 3036/// EstimateSearchSpaceComplexity - Estimate the worst-case number of 3037/// solutions the solver might have to consider. It almost never considers 3038/// this many solutions because it prune the search space, but the pruning 3039/// isn't always sufficient. 3040size_t LSRInstance::EstimateSearchSpaceComplexity() const { 3041 size_t Power = 1; 3042 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(), 3043 E = Uses.end(); I != E; ++I) { 3044 size_t FSize = I->Formulae.size(); 3045 if (FSize >= ComplexityLimit) { 3046 Power = ComplexityLimit; 3047 break; 3048 } 3049 Power *= FSize; 3050 if (Power >= ComplexityLimit) 3051 break; 3052 } 3053 return Power; 3054} 3055 3056/// NarrowSearchSpaceByDetectingSupersets - When one formula uses a superset 3057/// of the registers of another formula, it won't help reduce register 3058/// pressure (though it may not necessarily hurt register pressure); remove 3059/// it to simplify the system. 3060void LSRInstance::NarrowSearchSpaceByDetectingSupersets() { 3061 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) { 3062 DEBUG(dbgs() << "The search space is too complex.\n"); 3063 3064 DEBUG(dbgs() << "Narrowing the search space by eliminating formulae " 3065 "which use a superset of registers used by other " 3066 "formulae.\n"); 3067 3068 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) { 3069 LSRUse &LU = Uses[LUIdx]; 3070 bool Any = false; 3071 for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) { 3072 Formula &F = LU.Formulae[i]; 3073 // Look for a formula with a constant or GV in a register. If the use 3074 // also has a formula with that same value in an immediate field, 3075 // delete the one that uses a register. 3076 for (SmallVectorImpl<const SCEV *>::const_iterator 3077 I = F.BaseRegs.begin(), E = F.BaseRegs.end(); I != E; ++I) { 3078 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(*I)) { 3079 Formula NewF = F; 3080 NewF.AM.BaseOffs += C->getValue()->getSExtValue(); 3081 NewF.BaseRegs.erase(NewF.BaseRegs.begin() + 3082 (I - F.BaseRegs.begin())); 3083 if (LU.HasFormulaWithSameRegs(NewF)) { 3084 DEBUG(dbgs() << " Deleting "; F.print(dbgs()); dbgs() << '\n'); 3085 LU.DeleteFormula(F); 3086 --i; 3087 --e; 3088 Any = true; 3089 break; 3090 } 3091 } else if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(*I)) { 3092 if (GlobalValue *GV = dyn_cast<GlobalValue>(U->getValue())) 3093 if (!F.AM.BaseGV) { 3094 Formula NewF = F; 3095 NewF.AM.BaseGV = GV; 3096 NewF.BaseRegs.erase(NewF.BaseRegs.begin() + 3097 (I - F.BaseRegs.begin())); 3098 if (LU.HasFormulaWithSameRegs(NewF)) { 3099 DEBUG(dbgs() << " Deleting "; F.print(dbgs()); 3100 dbgs() << '\n'); 3101 LU.DeleteFormula(F); 3102 --i; 3103 --e; 3104 Any = true; 3105 break; 3106 } 3107 } 3108 } 3109 } 3110 } 3111 if (Any) 3112 LU.RecomputeRegs(LUIdx, RegUses); 3113 } 3114 3115 DEBUG(dbgs() << "After pre-selection:\n"; 3116 print_uses(dbgs())); 3117 } 3118} 3119 3120/// NarrowSearchSpaceByCollapsingUnrolledCode - When there are many registers 3121/// for expressions like A, A+1, A+2, etc., allocate a single register for 3122/// them. 3123void LSRInstance::NarrowSearchSpaceByCollapsingUnrolledCode() { 3124 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) { 3125 DEBUG(dbgs() << "The search space is too complex.\n"); 3126 3127 DEBUG(dbgs() << "Narrowing the search space by assuming that uses " 3128 "separated by a constant offset will use the same " 3129 "registers.\n"); 3130 3131 // This is especially useful for unrolled loops. 3132 3133 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) { 3134 LSRUse &LU = Uses[LUIdx]; 3135 for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(), 3136 E = LU.Formulae.end(); I != E; ++I) { 3137 const Formula &F = *I; 3138 if (F.AM.BaseOffs != 0 && F.AM.Scale == 0) { 3139 if (LSRUse *LUThatHas = FindUseWithSimilarFormula(F, LU)) { 3140 if (reconcileNewOffset(*LUThatHas, F.AM.BaseOffs, 3141 /*HasBaseReg=*/false, 3142 LU.Kind, LU.AccessTy)) { 3143 DEBUG(dbgs() << " Deleting use "; LU.print(dbgs()); 3144 dbgs() << '\n'); 3145 3146 LUThatHas->AllFixupsOutsideLoop &= LU.AllFixupsOutsideLoop; 3147 3148 // Update the relocs to reference the new use. 3149 for (SmallVectorImpl<LSRFixup>::iterator I = Fixups.begin(), 3150 E = Fixups.end(); I != E; ++I) { 3151 LSRFixup &Fixup = *I; 3152 if (Fixup.LUIdx == LUIdx) { 3153 Fixup.LUIdx = LUThatHas - &Uses.front(); 3154 Fixup.Offset += F.AM.BaseOffs; 3155 // Add the new offset to LUThatHas' offset list. 3156 if (LUThatHas->Offsets.back() != Fixup.Offset) { 3157 LUThatHas->Offsets.push_back(Fixup.Offset); 3158 if (Fixup.Offset > LUThatHas->MaxOffset) 3159 LUThatHas->MaxOffset = Fixup.Offset; 3160 if (Fixup.Offset < LUThatHas->MinOffset) 3161 LUThatHas->MinOffset = Fixup.Offset; 3162 } 3163 DEBUG(dbgs() << "New fixup has offset " 3164 << Fixup.Offset << '\n'); 3165 } 3166 if (Fixup.LUIdx == NumUses-1) 3167 Fixup.LUIdx = LUIdx; 3168 } 3169 3170 // Delete formulae from the new use which are no longer legal. 3171 bool Any = false; 3172 for (size_t i = 0, e = LUThatHas->Formulae.size(); i != e; ++i) { 3173 Formula &F = LUThatHas->Formulae[i]; 3174 if (!isLegalUse(F.AM, 3175 LUThatHas->MinOffset, LUThatHas->MaxOffset, 3176 LUThatHas->Kind, LUThatHas->AccessTy, TLI)) { 3177 DEBUG(dbgs() << " Deleting "; F.print(dbgs()); 3178 dbgs() << '\n'); 3179 LUThatHas->DeleteFormula(F); 3180 --i; 3181 --e; 3182 Any = true; 3183 } 3184 } 3185 if (Any) 3186 LUThatHas->RecomputeRegs(LUThatHas - &Uses.front(), RegUses); 3187 3188 // Delete the old use. 3189 DeleteUse(LU, LUIdx); 3190 --LUIdx; 3191 --NumUses; 3192 break; 3193 } 3194 } 3195 } 3196 } 3197 } 3198 3199 DEBUG(dbgs() << "After pre-selection:\n"; 3200 print_uses(dbgs())); 3201 } 3202} 3203 3204/// NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters - Call 3205/// FilterOutUndesirableDedicatedRegisters again, if necessary, now that 3206/// we've done more filtering, as it may be able to find more formulae to 3207/// eliminate. 3208void LSRInstance::NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters(){ 3209 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) { 3210 DEBUG(dbgs() << "The search space is too complex.\n"); 3211 3212 DEBUG(dbgs() << "Narrowing the search space by re-filtering out " 3213 "undesirable dedicated registers.\n"); 3214 3215 FilterOutUndesirableDedicatedRegisters(); 3216 3217 DEBUG(dbgs() << "After pre-selection:\n"; 3218 print_uses(dbgs())); 3219 } 3220} 3221 3222/// NarrowSearchSpaceByPickingWinnerRegs - Pick a register which seems likely 3223/// to be profitable, and then in any use which has any reference to that 3224/// register, delete all formulae which do not reference that register. 3225void LSRInstance::NarrowSearchSpaceByPickingWinnerRegs() { 3226 // With all other options exhausted, loop until the system is simple 3227 // enough to handle. 3228 SmallPtrSet<const SCEV *, 4> Taken; 3229 while (EstimateSearchSpaceComplexity() >= ComplexityLimit) { 3230 // Ok, we have too many of formulae on our hands to conveniently handle. 3231 // Use a rough heuristic to thin out the list. 3232 DEBUG(dbgs() << "The search space is too complex.\n"); 3233 3234 // Pick the register which is used by the most LSRUses, which is likely 3235 // to be a good reuse register candidate. 3236 const SCEV *Best = 0; 3237 unsigned BestNum = 0; 3238 for (RegUseTracker::const_iterator I = RegUses.begin(), E = RegUses.end(); 3239 I != E; ++I) { 3240 const SCEV *Reg = *I; 3241 if (Taken.count(Reg)) 3242 continue; 3243 if (!Best) 3244 Best = Reg; 3245 else { 3246 unsigned Count = RegUses.getUsedByIndices(Reg).count(); 3247 if (Count > BestNum) { 3248 Best = Reg; 3249 BestNum = Count; 3250 } 3251 } 3252 } 3253 3254 DEBUG(dbgs() << "Narrowing the search space by assuming " << *Best 3255 << " will yield profitable reuse.\n"); 3256 Taken.insert(Best); 3257 3258 // In any use with formulae which references this register, delete formulae 3259 // which don't reference it. 3260 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) { 3261 LSRUse &LU = Uses[LUIdx]; 3262 if (!LU.Regs.count(Best)) continue; 3263 3264 bool Any = false; 3265 for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) { 3266 Formula &F = LU.Formulae[i]; 3267 if (!F.referencesReg(Best)) { 3268 DEBUG(dbgs() << " Deleting "; F.print(dbgs()); dbgs() << '\n'); 3269 LU.DeleteFormula(F); 3270 --e; 3271 --i; 3272 Any = true; 3273 assert(e != 0 && "Use has no formulae left! Is Regs inconsistent?"); 3274 continue; 3275 } 3276 } 3277 3278 if (Any) 3279 LU.RecomputeRegs(LUIdx, RegUses); 3280 } 3281 3282 DEBUG(dbgs() << "After pre-selection:\n"; 3283 print_uses(dbgs())); 3284 } 3285} 3286 3287/// NarrowSearchSpaceUsingHeuristics - If there are an extraordinary number of 3288/// formulae to choose from, use some rough heuristics to prune down the number 3289/// of formulae. This keeps the main solver from taking an extraordinary amount 3290/// of time in some worst-case scenarios. 3291void LSRInstance::NarrowSearchSpaceUsingHeuristics() { 3292 NarrowSearchSpaceByDetectingSupersets(); 3293 NarrowSearchSpaceByCollapsingUnrolledCode(); 3294 NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters(); 3295 NarrowSearchSpaceByPickingWinnerRegs(); 3296} 3297 3298/// SolveRecurse - This is the recursive solver. 3299void LSRInstance::SolveRecurse(SmallVectorImpl<const Formula *> &Solution, 3300 Cost &SolutionCost, 3301 SmallVectorImpl<const Formula *> &Workspace, 3302 const Cost &CurCost, 3303 const SmallPtrSet<const SCEV *, 16> &CurRegs, 3304 DenseSet<const SCEV *> &VisitedRegs) const { 3305 // Some ideas: 3306 // - prune more: 3307 // - use more aggressive filtering 3308 // - sort the formula so that the most profitable solutions are found first 3309 // - sort the uses too 3310 // - search faster: 3311 // - don't compute a cost, and then compare. compare while computing a cost 3312 // and bail early. 3313 // - track register sets with SmallBitVector 3314 3315 const LSRUse &LU = Uses[Workspace.size()]; 3316 3317 // If this use references any register that's already a part of the 3318 // in-progress solution, consider it a requirement that a formula must 3319 // reference that register in order to be considered. This prunes out 3320 // unprofitable searching. 3321 SmallSetVector<const SCEV *, 4> ReqRegs; 3322 for (SmallPtrSet<const SCEV *, 16>::const_iterator I = CurRegs.begin(), 3323 E = CurRegs.end(); I != E; ++I) 3324 if (LU.Regs.count(*I)) 3325 ReqRegs.insert(*I); 3326 3327 bool AnySatisfiedReqRegs = false; 3328 SmallPtrSet<const SCEV *, 16> NewRegs; 3329 Cost NewCost; 3330retry: 3331 for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(), 3332 E = LU.Formulae.end(); I != E; ++I) { 3333 const Formula &F = *I; 3334 3335 // Ignore formulae which do not use any of the required registers. 3336 for (SmallSetVector<const SCEV *, 4>::const_iterator J = ReqRegs.begin(), 3337 JE = ReqRegs.end(); J != JE; ++J) { 3338 const SCEV *Reg = *J; 3339 if ((!F.ScaledReg || F.ScaledReg != Reg) && 3340 std::find(F.BaseRegs.begin(), F.BaseRegs.end(), Reg) == 3341 F.BaseRegs.end()) 3342 goto skip; 3343 } 3344 AnySatisfiedReqRegs = true; 3345 3346 // Evaluate the cost of the current formula. If it's already worse than 3347 // the current best, prune the search at that point. 3348 NewCost = CurCost; 3349 NewRegs = CurRegs; 3350 NewCost.RateFormula(F, NewRegs, VisitedRegs, L, LU.Offsets, SE, DT); 3351 if (NewCost < SolutionCost) { 3352 Workspace.push_back(&F); 3353 if (Workspace.size() != Uses.size()) { 3354 SolveRecurse(Solution, SolutionCost, Workspace, NewCost, 3355 NewRegs, VisitedRegs); 3356 if (F.getNumRegs() == 1 && Workspace.size() == 1) 3357 VisitedRegs.insert(F.ScaledReg ? F.ScaledReg : F.BaseRegs[0]); 3358 } else { 3359 DEBUG(dbgs() << "New best at "; NewCost.print(dbgs()); 3360 dbgs() << ".\n Regs:"; 3361 for (SmallPtrSet<const SCEV *, 16>::const_iterator 3362 I = NewRegs.begin(), E = NewRegs.end(); I != E; ++I) 3363 dbgs() << ' ' << **I; 3364 dbgs() << '\n'); 3365 3366 SolutionCost = NewCost; 3367 Solution = Workspace; 3368 } 3369 Workspace.pop_back(); 3370 } 3371 skip:; 3372 } 3373 3374 if (!EnableRetry && !AnySatisfiedReqRegs) 3375 return; 3376 3377 // If none of the formulae had all of the required registers, relax the 3378 // constraint so that we don't exclude all formulae. 3379 if (!AnySatisfiedReqRegs) { 3380 assert(!ReqRegs.empty() && "Solver failed even without required registers"); 3381 ReqRegs.clear(); 3382 goto retry; 3383 } 3384} 3385 3386/// Solve - Choose one formula from each use. Return the results in the given 3387/// Solution vector. 3388void LSRInstance::Solve(SmallVectorImpl<const Formula *> &Solution) const { 3389 SmallVector<const Formula *, 8> Workspace; 3390 Cost SolutionCost; 3391 SolutionCost.Loose(); 3392 Cost CurCost; 3393 SmallPtrSet<const SCEV *, 16> CurRegs; 3394 DenseSet<const SCEV *> VisitedRegs; 3395 Workspace.reserve(Uses.size()); 3396 3397 // SolveRecurse does all the work. 3398 SolveRecurse(Solution, SolutionCost, Workspace, CurCost, 3399 CurRegs, VisitedRegs); 3400 if (Solution.empty()) { 3401 DEBUG(dbgs() << "\nNo Satisfactory Solution\n"); 3402 return; 3403 } 3404 3405 // Ok, we've now made all our decisions. 3406 DEBUG(dbgs() << "\n" 3407 "The chosen solution requires "; SolutionCost.print(dbgs()); 3408 dbgs() << ":\n"; 3409 for (size_t i = 0, e = Uses.size(); i != e; ++i) { 3410 dbgs() << " "; 3411 Uses[i].print(dbgs()); 3412 dbgs() << "\n" 3413 " "; 3414 Solution[i]->print(dbgs()); 3415 dbgs() << '\n'; 3416 }); 3417 3418 assert(Solution.size() == Uses.size() && "Malformed solution!"); 3419} 3420 3421/// HoistInsertPosition - Helper for AdjustInsertPositionForExpand. Climb up 3422/// the dominator tree far as we can go while still being dominated by the 3423/// input positions. This helps canonicalize the insert position, which 3424/// encourages sharing. 3425BasicBlock::iterator 3426LSRInstance::HoistInsertPosition(BasicBlock::iterator IP, 3427 const SmallVectorImpl<Instruction *> &Inputs) 3428 const { 3429 for (;;) { 3430 const Loop *IPLoop = LI.getLoopFor(IP->getParent()); 3431 unsigned IPLoopDepth = IPLoop ? IPLoop->getLoopDepth() : 0; 3432 3433 BasicBlock *IDom; 3434 for (DomTreeNode *Rung = DT.getNode(IP->getParent()); ; ) { 3435 if (!Rung) return IP; 3436 Rung = Rung->getIDom(); 3437 if (!Rung) return IP; 3438 IDom = Rung->getBlock(); 3439 3440 // Don't climb into a loop though. 3441 const Loop *IDomLoop = LI.getLoopFor(IDom); 3442 unsigned IDomDepth = IDomLoop ? IDomLoop->getLoopDepth() : 0; 3443 if (IDomDepth <= IPLoopDepth && 3444 (IDomDepth != IPLoopDepth || IDomLoop == IPLoop)) 3445 break; 3446 } 3447 3448 bool AllDominate = true; 3449 Instruction *BetterPos = 0; 3450 Instruction *Tentative = IDom->getTerminator(); 3451 for (SmallVectorImpl<Instruction *>::const_iterator I = Inputs.begin(), 3452 E = Inputs.end(); I != E; ++I) { 3453 Instruction *Inst = *I; 3454 if (Inst == Tentative || !DT.dominates(Inst, Tentative)) { 3455 AllDominate = false; 3456 break; 3457 } 3458 // Attempt to find an insert position in the middle of the block, 3459 // instead of at the end, so that it can be used for other expansions. 3460 if (IDom == Inst->getParent() && 3461 (!BetterPos || DT.dominates(BetterPos, Inst))) 3462 BetterPos = llvm::next(BasicBlock::iterator(Inst)); 3463 } 3464 if (!AllDominate) 3465 break; 3466 if (BetterPos) 3467 IP = BetterPos; 3468 else 3469 IP = Tentative; 3470 } 3471 3472 return IP; 3473} 3474 3475/// AdjustInsertPositionForExpand - Determine an input position which will be 3476/// dominated by the operands and which will dominate the result. 3477BasicBlock::iterator 3478LSRInstance::AdjustInsertPositionForExpand(BasicBlock::iterator IP, 3479 const LSRFixup &LF, 3480 const LSRUse &LU) const { 3481 // Collect some instructions which must be dominated by the 3482 // expanding replacement. These must be dominated by any operands that 3483 // will be required in the expansion. 3484 SmallVector<Instruction *, 4> Inputs; 3485 if (Instruction *I = dyn_cast<Instruction>(LF.OperandValToReplace)) 3486 Inputs.push_back(I); 3487 if (LU.Kind == LSRUse::ICmpZero) 3488 if (Instruction *I = 3489 dyn_cast<Instruction>(cast<ICmpInst>(LF.UserInst)->getOperand(1))) 3490 Inputs.push_back(I); 3491 if (LF.PostIncLoops.count(L)) { 3492 if (LF.isUseFullyOutsideLoop(L)) 3493 Inputs.push_back(L->getLoopLatch()->getTerminator()); 3494 else 3495 Inputs.push_back(IVIncInsertPos); 3496 } 3497 // The expansion must also be dominated by the increment positions of any 3498 // loops it for which it is using post-inc mode. 3499 for (PostIncLoopSet::const_iterator I = LF.PostIncLoops.begin(), 3500 E = LF.PostIncLoops.end(); I != E; ++I) { 3501 const Loop *PIL = *I; 3502 if (PIL == L) continue; 3503 3504 // Be dominated by the loop exit. 3505 SmallVector<BasicBlock *, 4> ExitingBlocks; 3506 PIL->getExitingBlocks(ExitingBlocks); 3507 if (!ExitingBlocks.empty()) { 3508 BasicBlock *BB = ExitingBlocks[0]; 3509 for (unsigned i = 1, e = ExitingBlocks.size(); i != e; ++i) 3510 BB = DT.findNearestCommonDominator(BB, ExitingBlocks[i]); 3511 Inputs.push_back(BB->getTerminator()); 3512 } 3513 } 3514 3515 // Then, climb up the immediate dominator tree as far as we can go while 3516 // still being dominated by the input positions. 3517 IP = HoistInsertPosition(IP, Inputs); 3518 3519 // Don't insert instructions before PHI nodes. 3520 while (isa<PHINode>(IP)) ++IP; 3521 3522 // Ignore landingpad instructions. 3523 while (isa<LandingPadInst>(IP)) ++IP; 3524 3525 // Ignore debug intrinsics. 3526 while (isa<DbgInfoIntrinsic>(IP)) ++IP; 3527 3528 return IP; 3529} 3530 3531/// Expand - Emit instructions for the leading candidate expression for this 3532/// LSRUse (this is called "expanding"). 3533Value *LSRInstance::Expand(const LSRFixup &LF, 3534 const Formula &F, 3535 BasicBlock::iterator IP, 3536 SCEVExpander &Rewriter, 3537 SmallVectorImpl<WeakVH> &DeadInsts) const { 3538 const LSRUse &LU = Uses[LF.LUIdx]; 3539 3540 // Determine an input position which will be dominated by the operands and 3541 // which will dominate the result. 3542 IP = AdjustInsertPositionForExpand(IP, LF, LU); 3543 3544 // Inform the Rewriter if we have a post-increment use, so that it can 3545 // perform an advantageous expansion. 3546 Rewriter.setPostInc(LF.PostIncLoops); 3547 3548 // This is the type that the user actually needs. 3549 Type *OpTy = LF.OperandValToReplace->getType(); 3550 // This will be the type that we'll initially expand to. 3551 Type *Ty = F.getType(); 3552 if (!Ty) 3553 // No type known; just expand directly to the ultimate type. 3554 Ty = OpTy; 3555 else if (SE.getEffectiveSCEVType(Ty) == SE.getEffectiveSCEVType(OpTy)) 3556 // Expand directly to the ultimate type if it's the right size. 3557 Ty = OpTy; 3558 // This is the type to do integer arithmetic in. 3559 Type *IntTy = SE.getEffectiveSCEVType(Ty); 3560 3561 // Build up a list of operands to add together to form the full base. 3562 SmallVector<const SCEV *, 8> Ops; 3563 3564 // Expand the BaseRegs portion. 3565 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(), 3566 E = F.BaseRegs.end(); I != E; ++I) { 3567 const SCEV *Reg = *I; 3568 assert(!Reg->isZero() && "Zero allocated in a base register!"); 3569 3570 // If we're expanding for a post-inc user, make the post-inc adjustment. 3571 PostIncLoopSet &Loops = const_cast<PostIncLoopSet &>(LF.PostIncLoops); 3572 Reg = TransformForPostIncUse(Denormalize, Reg, 3573 LF.UserInst, LF.OperandValToReplace, 3574 Loops, SE, DT); 3575 3576 Ops.push_back(SE.getUnknown(Rewriter.expandCodeFor(Reg, 0, IP))); 3577 } 3578 3579 // Flush the operand list to suppress SCEVExpander hoisting. 3580 if (!Ops.empty()) { 3581 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP); 3582 Ops.clear(); 3583 Ops.push_back(SE.getUnknown(FullV)); 3584 } 3585 3586 // Expand the ScaledReg portion. 3587 Value *ICmpScaledV = 0; 3588 if (F.AM.Scale != 0) { 3589 const SCEV *ScaledS = F.ScaledReg; 3590 3591 // If we're expanding for a post-inc user, make the post-inc adjustment. 3592 PostIncLoopSet &Loops = const_cast<PostIncLoopSet &>(LF.PostIncLoops); 3593 ScaledS = TransformForPostIncUse(Denormalize, ScaledS, 3594 LF.UserInst, LF.OperandValToReplace, 3595 Loops, SE, DT); 3596 3597 if (LU.Kind == LSRUse::ICmpZero) { 3598 // An interesting way of "folding" with an icmp is to use a negated 3599 // scale, which we'll implement by inserting it into the other operand 3600 // of the icmp. 3601 assert(F.AM.Scale == -1 && 3602 "The only scale supported by ICmpZero uses is -1!"); 3603 ICmpScaledV = Rewriter.expandCodeFor(ScaledS, 0, IP); 3604 } else { 3605 // Otherwise just expand the scaled register and an explicit scale, 3606 // which is expected to be matched as part of the address. 3607 ScaledS = SE.getUnknown(Rewriter.expandCodeFor(ScaledS, 0, IP)); 3608 ScaledS = SE.getMulExpr(ScaledS, 3609 SE.getConstant(ScaledS->getType(), F.AM.Scale)); 3610 Ops.push_back(ScaledS); 3611 3612 // Flush the operand list to suppress SCEVExpander hoisting. 3613 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP); 3614 Ops.clear(); 3615 Ops.push_back(SE.getUnknown(FullV)); 3616 } 3617 } 3618 3619 // Expand the GV portion. 3620 if (F.AM.BaseGV) { 3621 Ops.push_back(SE.getUnknown(F.AM.BaseGV)); 3622 3623 // Flush the operand list to suppress SCEVExpander hoisting. 3624 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP); 3625 Ops.clear(); 3626 Ops.push_back(SE.getUnknown(FullV)); 3627 } 3628 3629 // Expand the immediate portion. 3630 int64_t Offset = (uint64_t)F.AM.BaseOffs + LF.Offset; 3631 if (Offset != 0) { 3632 if (LU.Kind == LSRUse::ICmpZero) { 3633 // The other interesting way of "folding" with an ICmpZero is to use a 3634 // negated immediate. 3635 if (!ICmpScaledV) 3636 ICmpScaledV = ConstantInt::get(IntTy, -(uint64_t)Offset); 3637 else { 3638 Ops.push_back(SE.getUnknown(ICmpScaledV)); 3639 ICmpScaledV = ConstantInt::get(IntTy, Offset); 3640 } 3641 } else { 3642 // Just add the immediate values. These again are expected to be matched 3643 // as part of the address. 3644 Ops.push_back(SE.getUnknown(ConstantInt::getSigned(IntTy, Offset))); 3645 } 3646 } 3647 3648 // Expand the unfolded offset portion. 3649 int64_t UnfoldedOffset = F.UnfoldedOffset; 3650 if (UnfoldedOffset != 0) { 3651 // Just add the immediate values. 3652 Ops.push_back(SE.getUnknown(ConstantInt::getSigned(IntTy, 3653 UnfoldedOffset))); 3654 } 3655 3656 // Emit instructions summing all the operands. 3657 const SCEV *FullS = Ops.empty() ? 3658 SE.getConstant(IntTy, 0) : 3659 SE.getAddExpr(Ops); 3660 Value *FullV = Rewriter.expandCodeFor(FullS, Ty, IP); 3661 3662 // We're done expanding now, so reset the rewriter. 3663 Rewriter.clearPostInc(); 3664 3665 // An ICmpZero Formula represents an ICmp which we're handling as a 3666 // comparison against zero. Now that we've expanded an expression for that 3667 // form, update the ICmp's other operand. 3668 if (LU.Kind == LSRUse::ICmpZero) { 3669 ICmpInst *CI = cast<ICmpInst>(LF.UserInst); 3670 DeadInsts.push_back(CI->getOperand(1)); 3671 assert(!F.AM.BaseGV && "ICmp does not support folding a global value and " 3672 "a scale at the same time!"); 3673 if (F.AM.Scale == -1) { 3674 if (ICmpScaledV->getType() != OpTy) { 3675 Instruction *Cast = 3676 CastInst::Create(CastInst::getCastOpcode(ICmpScaledV, false, 3677 OpTy, false), 3678 ICmpScaledV, OpTy, "tmp", CI); 3679 ICmpScaledV = Cast; 3680 } 3681 CI->setOperand(1, ICmpScaledV); 3682 } else { 3683 assert(F.AM.Scale == 0 && 3684 "ICmp does not support folding a global value and " 3685 "a scale at the same time!"); 3686 Constant *C = ConstantInt::getSigned(SE.getEffectiveSCEVType(OpTy), 3687 -(uint64_t)Offset); 3688 if (C->getType() != OpTy) 3689 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false, 3690 OpTy, false), 3691 C, OpTy); 3692 3693 CI->setOperand(1, C); 3694 } 3695 } 3696 3697 return FullV; 3698} 3699 3700/// RewriteForPHI - Helper for Rewrite. PHI nodes are special because the use 3701/// of their operands effectively happens in their predecessor blocks, so the 3702/// expression may need to be expanded in multiple places. 3703void LSRInstance::RewriteForPHI(PHINode *PN, 3704 const LSRFixup &LF, 3705 const Formula &F, 3706 SCEVExpander &Rewriter, 3707 SmallVectorImpl<WeakVH> &DeadInsts, 3708 Pass *P) const { 3709 DenseMap<BasicBlock *, Value *> Inserted; 3710 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) 3711 if (PN->getIncomingValue(i) == LF.OperandValToReplace) { 3712 BasicBlock *BB = PN->getIncomingBlock(i); 3713 3714 // If this is a critical edge, split the edge so that we do not insert 3715 // the code on all predecessor/successor paths. We do this unless this 3716 // is the canonical backedge for this loop, which complicates post-inc 3717 // users. 3718 if (e != 1 && BB->getTerminator()->getNumSuccessors() > 1 && 3719 !isa<IndirectBrInst>(BB->getTerminator())) { 3720 BasicBlock *Parent = PN->getParent(); 3721 Loop *PNLoop = LI.getLoopFor(Parent); 3722 if (!PNLoop || Parent != PNLoop->getHeader()) { 3723 // Split the critical edge. 3724 BasicBlock *NewBB = 0; 3725 if (!Parent->isLandingPad()) { 3726 NewBB = SplitCriticalEdge(BB, Parent, P, 3727 /*MergeIdenticalEdges=*/true, 3728 /*DontDeleteUselessPhis=*/true); 3729 } else { 3730 SmallVector<BasicBlock*, 2> NewBBs; 3731 SplitLandingPadPredecessors(Parent, BB, "", "", P, NewBBs); 3732 NewBB = NewBBs[0]; 3733 } 3734 3735 // If PN is outside of the loop and BB is in the loop, we want to 3736 // move the block to be immediately before the PHI block, not 3737 // immediately after BB. 3738 if (L->contains(BB) && !L->contains(PN)) 3739 NewBB->moveBefore(PN->getParent()); 3740 3741 // Splitting the edge can reduce the number of PHI entries we have. 3742 e = PN->getNumIncomingValues(); 3743 BB = NewBB; 3744 i = PN->getBasicBlockIndex(BB); 3745 } 3746 } 3747 3748 std::pair<DenseMap<BasicBlock *, Value *>::iterator, bool> Pair = 3749 Inserted.insert(std::make_pair(BB, static_cast<Value *>(0))); 3750 if (!Pair.second) 3751 PN->setIncomingValue(i, Pair.first->second); 3752 else { 3753 Value *FullV = Expand(LF, F, BB->getTerminator(), Rewriter, DeadInsts); 3754 3755 // If this is reuse-by-noop-cast, insert the noop cast. 3756 Type *OpTy = LF.OperandValToReplace->getType(); 3757 if (FullV->getType() != OpTy) 3758 FullV = 3759 CastInst::Create(CastInst::getCastOpcode(FullV, false, 3760 OpTy, false), 3761 FullV, LF.OperandValToReplace->getType(), 3762 "tmp", BB->getTerminator()); 3763 3764 PN->setIncomingValue(i, FullV); 3765 Pair.first->second = FullV; 3766 } 3767 } 3768} 3769 3770/// Rewrite - Emit instructions for the leading candidate expression for this 3771/// LSRUse (this is called "expanding"), and update the UserInst to reference 3772/// the newly expanded value. 3773void LSRInstance::Rewrite(const LSRFixup &LF, 3774 const Formula &F, 3775 SCEVExpander &Rewriter, 3776 SmallVectorImpl<WeakVH> &DeadInsts, 3777 Pass *P) const { 3778 // First, find an insertion point that dominates UserInst. For PHI nodes, 3779 // find the nearest block which dominates all the relevant uses. 3780 if (PHINode *PN = dyn_cast<PHINode>(LF.UserInst)) { 3781 RewriteForPHI(PN, LF, F, Rewriter, DeadInsts, P); 3782 } else { 3783 Value *FullV = Expand(LF, F, LF.UserInst, Rewriter, DeadInsts); 3784 3785 // If this is reuse-by-noop-cast, insert the noop cast. 3786 Type *OpTy = LF.OperandValToReplace->getType(); 3787 if (FullV->getType() != OpTy) { 3788 Instruction *Cast = 3789 CastInst::Create(CastInst::getCastOpcode(FullV, false, OpTy, false), 3790 FullV, OpTy, "tmp", LF.UserInst); 3791 FullV = Cast; 3792 } 3793 3794 // Update the user. ICmpZero is handled specially here (for now) because 3795 // Expand may have updated one of the operands of the icmp already, and 3796 // its new value may happen to be equal to LF.OperandValToReplace, in 3797 // which case doing replaceUsesOfWith leads to replacing both operands 3798 // with the same value. TODO: Reorganize this. 3799 if (Uses[LF.LUIdx].Kind == LSRUse::ICmpZero) 3800 LF.UserInst->setOperand(0, FullV); 3801 else 3802 LF.UserInst->replaceUsesOfWith(LF.OperandValToReplace, FullV); 3803 } 3804 3805 DeadInsts.push_back(LF.OperandValToReplace); 3806} 3807 3808/// ImplementSolution - Rewrite all the fixup locations with new values, 3809/// following the chosen solution. 3810void 3811LSRInstance::ImplementSolution(const SmallVectorImpl<const Formula *> &Solution, 3812 Pass *P) { 3813 // Keep track of instructions we may have made dead, so that 3814 // we can remove them after we are done working. 3815 SmallVector<WeakVH, 16> DeadInsts; 3816 3817 SCEVExpander Rewriter(SE, "lsr"); 3818#ifndef NDEBUG 3819 Rewriter.setDebugType(DEBUG_TYPE); 3820#endif 3821 Rewriter.disableCanonicalMode(); 3822 Rewriter.enableLSRMode(); 3823 Rewriter.setIVIncInsertPos(L, IVIncInsertPos); 3824 3825 // Expand the new value definitions and update the users. 3826 for (SmallVectorImpl<LSRFixup>::const_iterator I = Fixups.begin(), 3827 E = Fixups.end(); I != E; ++I) { 3828 const LSRFixup &Fixup = *I; 3829 3830 Rewrite(Fixup, *Solution[Fixup.LUIdx], Rewriter, DeadInsts, P); 3831 3832 Changed = true; 3833 } 3834 3835 // Clean up after ourselves. This must be done before deleting any 3836 // instructions. 3837 Rewriter.clear(); 3838 3839 Changed |= DeleteTriviallyDeadInstructions(DeadInsts); 3840} 3841 3842LSRInstance::LSRInstance(const TargetLowering *tli, Loop *l, Pass *P) 3843 : IU(P->getAnalysis<IVUsers>()), 3844 SE(P->getAnalysis<ScalarEvolution>()), 3845 DT(P->getAnalysis<DominatorTree>()), 3846 LI(P->getAnalysis<LoopInfo>()), 3847 TLI(tli), L(l), Changed(false), IVIncInsertPos(0) { 3848 3849 // If LoopSimplify form is not available, stay out of trouble. 3850 if (!L->isLoopSimplifyForm()) 3851 return; 3852 3853 // All outer loops must have preheaders, or SCEVExpander may not be able to 3854 // materialize an AddRecExpr whose Start is an outer AddRecExpr. 3855 for (const Loop *OuterLoop = L; (OuterLoop = OuterLoop->getParentLoop());) { 3856 if (!OuterLoop->getLoopPreheader()) 3857 return; 3858 } 3859 // If there's no interesting work to be done, bail early. 3860 if (IU.empty()) return; 3861 3862 DEBUG(dbgs() << "\nLSR on loop "; 3863 WriteAsOperand(dbgs(), L->getHeader(), /*PrintType=*/false); 3864 dbgs() << ":\n"); 3865 3866 // First, perform some low-level loop optimizations. 3867 OptimizeShadowIV(); 3868 OptimizeLoopTermCond(); 3869 3870 // If loop preparation eliminates all interesting IV users, bail. 3871 if (IU.empty()) return; 3872 3873 // Skip nested loops until we can model them better with formulae. 3874 if (!EnableNested && !L->empty()) { 3875 DEBUG(dbgs() << "LSR skipping outer loop " << *L << "\n"); 3876 return; 3877 } 3878 3879 // Start collecting data and preparing for the solver. 3880 CollectInterestingTypesAndFactors(); 3881 CollectFixupsAndInitialFormulae(); 3882 CollectLoopInvariantFixupsAndFormulae(); 3883 3884 DEBUG(dbgs() << "LSR found " << Uses.size() << " uses:\n"; 3885 print_uses(dbgs())); 3886 3887 // Now use the reuse data to generate a bunch of interesting ways 3888 // to formulate the values needed for the uses. 3889 GenerateAllReuseFormulae(); 3890 3891 FilterOutUndesirableDedicatedRegisters(); 3892 NarrowSearchSpaceUsingHeuristics(); 3893 3894 SmallVector<const Formula *, 8> Solution; 3895 Solve(Solution); 3896 3897 // Release memory that is no longer needed. 3898 Factors.clear(); 3899 Types.clear(); 3900 RegUses.clear(); 3901 3902 if (Solution.empty()) 3903 return; 3904 3905#ifndef NDEBUG 3906 // Formulae should be legal. 3907 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(), 3908 E = Uses.end(); I != E; ++I) { 3909 const LSRUse &LU = *I; 3910 for (SmallVectorImpl<Formula>::const_iterator J = LU.Formulae.begin(), 3911 JE = LU.Formulae.end(); J != JE; ++J) 3912 assert(isLegalUse(J->AM, LU.MinOffset, LU.MaxOffset, 3913 LU.Kind, LU.AccessTy, TLI) && 3914 "Illegal formula generated!"); 3915 }; 3916#endif 3917 3918 // Now that we've decided what we want, make it so. 3919 ImplementSolution(Solution, P); 3920} 3921 3922void LSRInstance::print_factors_and_types(raw_ostream &OS) const { 3923 if (Factors.empty() && Types.empty()) return; 3924 3925 OS << "LSR has identified the following interesting factors and types: "; 3926 bool First = true; 3927 3928 for (SmallSetVector<int64_t, 8>::const_iterator 3929 I = Factors.begin(), E = Factors.end(); I != E; ++I) { 3930 if (!First) OS << ", "; 3931 First = false; 3932 OS << '*' << *I; 3933 } 3934 3935 for (SmallSetVector<Type *, 4>::const_iterator 3936 I = Types.begin(), E = Types.end(); I != E; ++I) { 3937 if (!First) OS << ", "; 3938 First = false; 3939 OS << '(' << **I << ')'; 3940 } 3941 OS << '\n'; 3942} 3943 3944void LSRInstance::print_fixups(raw_ostream &OS) const { 3945 OS << "LSR is examining the following fixup sites:\n"; 3946 for (SmallVectorImpl<LSRFixup>::const_iterator I = Fixups.begin(), 3947 E = Fixups.end(); I != E; ++I) { 3948 dbgs() << " "; 3949 I->print(OS); 3950 OS << '\n'; 3951 } 3952} 3953 3954void LSRInstance::print_uses(raw_ostream &OS) const { 3955 OS << "LSR is examining the following uses:\n"; 3956 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(), 3957 E = Uses.end(); I != E; ++I) { 3958 const LSRUse &LU = *I; 3959 dbgs() << " "; 3960 LU.print(OS); 3961 OS << '\n'; 3962 for (SmallVectorImpl<Formula>::const_iterator J = LU.Formulae.begin(), 3963 JE = LU.Formulae.end(); J != JE; ++J) { 3964 OS << " "; 3965 J->print(OS); 3966 OS << '\n'; 3967 } 3968 } 3969} 3970 3971void LSRInstance::print(raw_ostream &OS) const { 3972 print_factors_and_types(OS); 3973 print_fixups(OS); 3974 print_uses(OS); 3975} 3976 3977void LSRInstance::dump() const { 3978 print(errs()); errs() << '\n'; 3979} 3980 3981namespace { 3982 3983class LoopStrengthReduce : public LoopPass { 3984 /// TLI - Keep a pointer of a TargetLowering to consult for determining 3985 /// transformation profitability. 3986 const TargetLowering *const TLI; 3987 3988public: 3989 static char ID; // Pass ID, replacement for typeid 3990 explicit LoopStrengthReduce(const TargetLowering *tli = 0); 3991 3992private: 3993 bool runOnLoop(Loop *L, LPPassManager &LPM); 3994 void getAnalysisUsage(AnalysisUsage &AU) const; 3995}; 3996 3997} 3998 3999char LoopStrengthReduce::ID = 0; 4000INITIALIZE_PASS_BEGIN(LoopStrengthReduce, "loop-reduce", 4001 "Loop Strength Reduction", false, false) 4002INITIALIZE_PASS_DEPENDENCY(DominatorTree) 4003INITIALIZE_PASS_DEPENDENCY(ScalarEvolution) 4004INITIALIZE_PASS_DEPENDENCY(IVUsers) 4005INITIALIZE_PASS_DEPENDENCY(LoopInfo) 4006INITIALIZE_PASS_DEPENDENCY(LoopSimplify) 4007INITIALIZE_PASS_END(LoopStrengthReduce, "loop-reduce", 4008 "Loop Strength Reduction", false, false) 4009 4010 4011Pass *llvm::createLoopStrengthReducePass(const TargetLowering *TLI) { 4012 return new LoopStrengthReduce(TLI); 4013} 4014 4015LoopStrengthReduce::LoopStrengthReduce(const TargetLowering *tli) 4016 : LoopPass(ID), TLI(tli) { 4017 initializeLoopStrengthReducePass(*PassRegistry::getPassRegistry()); 4018 } 4019 4020void LoopStrengthReduce::getAnalysisUsage(AnalysisUsage &AU) const { 4021 // We split critical edges, so we change the CFG. However, we do update 4022 // many analyses if they are around. 4023 AU.addPreservedID(LoopSimplifyID); 4024 4025 AU.addRequired<LoopInfo>(); 4026 AU.addPreserved<LoopInfo>(); 4027 AU.addRequiredID(LoopSimplifyID); 4028 AU.addRequired<DominatorTree>(); 4029 AU.addPreserved<DominatorTree>(); 4030 AU.addRequired<ScalarEvolution>(); 4031 AU.addPreserved<ScalarEvolution>(); 4032 // Requiring LoopSimplify a second time here prevents IVUsers from running 4033 // twice, since LoopSimplify was invalidated by running ScalarEvolution. 4034 AU.addRequiredID(LoopSimplifyID); 4035 AU.addRequired<IVUsers>(); 4036 AU.addPreserved<IVUsers>(); 4037} 4038 4039bool LoopStrengthReduce::runOnLoop(Loop *L, LPPassManager & /*LPM*/) { 4040 bool Changed = false; 4041 4042 // Run the main LSR transformation. 4043 Changed |= LSRInstance(TLI, L, this).getChanged(); 4044 4045 // Remove any extra phis created by processing inner loops. 4046 Changed |= DeleteDeadPHIs(L->getHeader()); 4047 if (EnablePhiElim) { 4048 SmallVector<WeakVH, 16> DeadInsts; 4049 SCEVExpander Rewriter(getAnalysis<ScalarEvolution>(), "lsr"); 4050#ifndef NDEBUG 4051 Rewriter.setDebugType(DEBUG_TYPE); 4052#endif 4053 unsigned numFolded = Rewriter. 4054 replaceCongruentIVs(L, &getAnalysis<DominatorTree>(), DeadInsts, TLI); 4055 if (numFolded) { 4056 Changed = true; 4057 DeleteTriviallyDeadInstructions(DeadInsts); 4058 DeleteDeadPHIs(L->getHeader()); 4059 } 4060 } 4061 return Changed; 4062} 4063