LoopStrengthReduce.cpp revision 9243c4f7c54b8d0be22a4c9d411b15f462039d06
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 80// Temporary flag to cleanup congruent phis after LSR phi expansion. 81// It's currently disabled until we can determine whether it's truly useful or 82// not. The flag should be removed after the v3.0 release. 83// This is now needed for ivchains. 84static cl::opt<bool> EnablePhiElim( 85 "enable-lsr-phielim", cl::Hidden, cl::init(true), 86 cl::desc("Enable LSR phi elimination")); 87 88#ifndef NDEBUG 89// Stress test IV chain generation. 90static cl::opt<bool> StressIVChain( 91 "stress-ivchain", cl::Hidden, cl::init(false), 92 cl::desc("Stress test LSR IV chains")); 93#else 94static bool StressIVChain = false; 95#endif 96 97namespace { 98 99/// RegSortData - This class holds data which is used to order reuse candidates. 100class RegSortData { 101public: 102 /// UsedByIndices - This represents the set of LSRUse indices which reference 103 /// a particular register. 104 SmallBitVector UsedByIndices; 105 106 RegSortData() {} 107 108 void print(raw_ostream &OS) const; 109 void dump() const; 110}; 111 112} 113 114void RegSortData::print(raw_ostream &OS) const { 115 OS << "[NumUses=" << UsedByIndices.count() << ']'; 116} 117 118void RegSortData::dump() const { 119 print(errs()); errs() << '\n'; 120} 121 122namespace { 123 124/// RegUseTracker - Map register candidates to information about how they are 125/// used. 126class RegUseTracker { 127 typedef DenseMap<const SCEV *, RegSortData> RegUsesTy; 128 129 RegUsesTy RegUsesMap; 130 SmallVector<const SCEV *, 16> RegSequence; 131 132public: 133 void CountRegister(const SCEV *Reg, size_t LUIdx); 134 void DropRegister(const SCEV *Reg, size_t LUIdx); 135 void SwapAndDropUse(size_t LUIdx, size_t LastLUIdx); 136 137 bool isRegUsedByUsesOtherThan(const SCEV *Reg, size_t LUIdx) const; 138 139 const SmallBitVector &getUsedByIndices(const SCEV *Reg) const; 140 141 void clear(); 142 143 typedef SmallVectorImpl<const SCEV *>::iterator iterator; 144 typedef SmallVectorImpl<const SCEV *>::const_iterator const_iterator; 145 iterator begin() { return RegSequence.begin(); } 146 iterator end() { return RegSequence.end(); } 147 const_iterator begin() const { return RegSequence.begin(); } 148 const_iterator end() const { return RegSequence.end(); } 149}; 150 151} 152 153void 154RegUseTracker::CountRegister(const SCEV *Reg, size_t LUIdx) { 155 std::pair<RegUsesTy::iterator, bool> Pair = 156 RegUsesMap.insert(std::make_pair(Reg, RegSortData())); 157 RegSortData &RSD = Pair.first->second; 158 if (Pair.second) 159 RegSequence.push_back(Reg); 160 RSD.UsedByIndices.resize(std::max(RSD.UsedByIndices.size(), LUIdx + 1)); 161 RSD.UsedByIndices.set(LUIdx); 162} 163 164void 165RegUseTracker::DropRegister(const SCEV *Reg, size_t LUIdx) { 166 RegUsesTy::iterator It = RegUsesMap.find(Reg); 167 assert(It != RegUsesMap.end()); 168 RegSortData &RSD = It->second; 169 assert(RSD.UsedByIndices.size() > LUIdx); 170 RSD.UsedByIndices.reset(LUIdx); 171} 172 173void 174RegUseTracker::SwapAndDropUse(size_t LUIdx, size_t LastLUIdx) { 175 assert(LUIdx <= LastLUIdx); 176 177 // Update RegUses. The data structure is not optimized for this purpose; 178 // we must iterate through it and update each of the bit vectors. 179 for (RegUsesTy::iterator I = RegUsesMap.begin(), E = RegUsesMap.end(); 180 I != E; ++I) { 181 SmallBitVector &UsedByIndices = I->second.UsedByIndices; 182 if (LUIdx < UsedByIndices.size()) 183 UsedByIndices[LUIdx] = 184 LastLUIdx < UsedByIndices.size() ? UsedByIndices[LastLUIdx] : 0; 185 UsedByIndices.resize(std::min(UsedByIndices.size(), LastLUIdx)); 186 } 187} 188 189bool 190RegUseTracker::isRegUsedByUsesOtherThan(const SCEV *Reg, size_t LUIdx) const { 191 RegUsesTy::const_iterator I = RegUsesMap.find(Reg); 192 if (I == RegUsesMap.end()) 193 return false; 194 const SmallBitVector &UsedByIndices = I->second.UsedByIndices; 195 int i = UsedByIndices.find_first(); 196 if (i == -1) return false; 197 if ((size_t)i != LUIdx) return true; 198 return UsedByIndices.find_next(i) != -1; 199} 200 201const SmallBitVector &RegUseTracker::getUsedByIndices(const SCEV *Reg) const { 202 RegUsesTy::const_iterator I = RegUsesMap.find(Reg); 203 assert(I != RegUsesMap.end() && "Unknown register!"); 204 return I->second.UsedByIndices; 205} 206 207void RegUseTracker::clear() { 208 RegUsesMap.clear(); 209 RegSequence.clear(); 210} 211 212namespace { 213 214/// Formula - This class holds information that describes a formula for 215/// computing satisfying a use. It may include broken-out immediates and scaled 216/// registers. 217struct Formula { 218 /// AM - This is used to represent complex addressing, as well as other kinds 219 /// of interesting uses. 220 TargetLowering::AddrMode AM; 221 222 /// BaseRegs - The list of "base" registers for this use. When this is 223 /// non-empty, AM.HasBaseReg should be set to true. 224 SmallVector<const SCEV *, 2> BaseRegs; 225 226 /// ScaledReg - The 'scaled' register for this use. This should be non-null 227 /// when AM.Scale is not zero. 228 const SCEV *ScaledReg; 229 230 /// UnfoldedOffset - An additional constant offset which added near the 231 /// use. This requires a temporary register, but the offset itself can 232 /// live in an add immediate field rather than a register. 233 int64_t UnfoldedOffset; 234 235 Formula() : ScaledReg(0), UnfoldedOffset(0) {} 236 237 void InitialMatch(const SCEV *S, Loop *L, ScalarEvolution &SE); 238 239 unsigned getNumRegs() const; 240 Type *getType() const; 241 242 void DeleteBaseReg(const SCEV *&S); 243 244 bool referencesReg(const SCEV *S) const; 245 bool hasRegsUsedByUsesOtherThan(size_t LUIdx, 246 const RegUseTracker &RegUses) const; 247 248 void print(raw_ostream &OS) const; 249 void dump() const; 250}; 251 252} 253 254/// DoInitialMatch - Recursion helper for InitialMatch. 255static void DoInitialMatch(const SCEV *S, Loop *L, 256 SmallVectorImpl<const SCEV *> &Good, 257 SmallVectorImpl<const SCEV *> &Bad, 258 ScalarEvolution &SE) { 259 // Collect expressions which properly dominate the loop header. 260 if (SE.properlyDominates(S, L->getHeader())) { 261 Good.push_back(S); 262 return; 263 } 264 265 // Look at add operands. 266 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) { 267 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end(); 268 I != E; ++I) 269 DoInitialMatch(*I, L, Good, Bad, SE); 270 return; 271 } 272 273 // Look at addrec operands. 274 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) 275 if (!AR->getStart()->isZero()) { 276 DoInitialMatch(AR->getStart(), L, Good, Bad, SE); 277 DoInitialMatch(SE.getAddRecExpr(SE.getConstant(AR->getType(), 0), 278 AR->getStepRecurrence(SE), 279 // FIXME: AR->getNoWrapFlags() 280 AR->getLoop(), SCEV::FlagAnyWrap), 281 L, Good, Bad, SE); 282 return; 283 } 284 285 // Handle a multiplication by -1 (negation) if it didn't fold. 286 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) 287 if (Mul->getOperand(0)->isAllOnesValue()) { 288 SmallVector<const SCEV *, 4> Ops(Mul->op_begin()+1, Mul->op_end()); 289 const SCEV *NewMul = SE.getMulExpr(Ops); 290 291 SmallVector<const SCEV *, 4> MyGood; 292 SmallVector<const SCEV *, 4> MyBad; 293 DoInitialMatch(NewMul, L, MyGood, MyBad, SE); 294 const SCEV *NegOne = SE.getSCEV(ConstantInt::getAllOnesValue( 295 SE.getEffectiveSCEVType(NewMul->getType()))); 296 for (SmallVectorImpl<const SCEV *>::const_iterator I = MyGood.begin(), 297 E = MyGood.end(); I != E; ++I) 298 Good.push_back(SE.getMulExpr(NegOne, *I)); 299 for (SmallVectorImpl<const SCEV *>::const_iterator I = MyBad.begin(), 300 E = MyBad.end(); I != E; ++I) 301 Bad.push_back(SE.getMulExpr(NegOne, *I)); 302 return; 303 } 304 305 // Ok, we can't do anything interesting. Just stuff the whole thing into a 306 // register and hope for the best. 307 Bad.push_back(S); 308} 309 310/// InitialMatch - Incorporate loop-variant parts of S into this Formula, 311/// attempting to keep all loop-invariant and loop-computable values in a 312/// single base register. 313void Formula::InitialMatch(const SCEV *S, Loop *L, ScalarEvolution &SE) { 314 SmallVector<const SCEV *, 4> Good; 315 SmallVector<const SCEV *, 4> Bad; 316 DoInitialMatch(S, L, Good, Bad, SE); 317 if (!Good.empty()) { 318 const SCEV *Sum = SE.getAddExpr(Good); 319 if (!Sum->isZero()) 320 BaseRegs.push_back(Sum); 321 AM.HasBaseReg = true; 322 } 323 if (!Bad.empty()) { 324 const SCEV *Sum = SE.getAddExpr(Bad); 325 if (!Sum->isZero()) 326 BaseRegs.push_back(Sum); 327 AM.HasBaseReg = true; 328 } 329} 330 331/// getNumRegs - Return the total number of register operands used by this 332/// formula. This does not include register uses implied by non-constant 333/// addrec strides. 334unsigned Formula::getNumRegs() const { 335 return !!ScaledReg + BaseRegs.size(); 336} 337 338/// getType - Return the type of this formula, if it has one, or null 339/// otherwise. This type is meaningless except for the bit size. 340Type *Formula::getType() const { 341 return !BaseRegs.empty() ? BaseRegs.front()->getType() : 342 ScaledReg ? ScaledReg->getType() : 343 AM.BaseGV ? AM.BaseGV->getType() : 344 0; 345} 346 347/// DeleteBaseReg - Delete the given base reg from the BaseRegs list. 348void Formula::DeleteBaseReg(const SCEV *&S) { 349 if (&S != &BaseRegs.back()) 350 std::swap(S, BaseRegs.back()); 351 BaseRegs.pop_back(); 352} 353 354/// referencesReg - Test if this formula references the given register. 355bool Formula::referencesReg(const SCEV *S) const { 356 return S == ScaledReg || 357 std::find(BaseRegs.begin(), BaseRegs.end(), S) != BaseRegs.end(); 358} 359 360/// hasRegsUsedByUsesOtherThan - Test whether this formula uses registers 361/// which are used by uses other than the use with the given index. 362bool Formula::hasRegsUsedByUsesOtherThan(size_t LUIdx, 363 const RegUseTracker &RegUses) const { 364 if (ScaledReg) 365 if (RegUses.isRegUsedByUsesOtherThan(ScaledReg, LUIdx)) 366 return true; 367 for (SmallVectorImpl<const SCEV *>::const_iterator I = BaseRegs.begin(), 368 E = BaseRegs.end(); I != E; ++I) 369 if (RegUses.isRegUsedByUsesOtherThan(*I, LUIdx)) 370 return true; 371 return false; 372} 373 374void Formula::print(raw_ostream &OS) const { 375 bool First = true; 376 if (AM.BaseGV) { 377 if (!First) OS << " + "; else First = false; 378 WriteAsOperand(OS, AM.BaseGV, /*PrintType=*/false); 379 } 380 if (AM.BaseOffs != 0) { 381 if (!First) OS << " + "; else First = false; 382 OS << AM.BaseOffs; 383 } 384 for (SmallVectorImpl<const SCEV *>::const_iterator I = BaseRegs.begin(), 385 E = BaseRegs.end(); I != E; ++I) { 386 if (!First) OS << " + "; else First = false; 387 OS << "reg(" << **I << ')'; 388 } 389 if (AM.HasBaseReg && BaseRegs.empty()) { 390 if (!First) OS << " + "; else First = false; 391 OS << "**error: HasBaseReg**"; 392 } else if (!AM.HasBaseReg && !BaseRegs.empty()) { 393 if (!First) OS << " + "; else First = false; 394 OS << "**error: !HasBaseReg**"; 395 } 396 if (AM.Scale != 0) { 397 if (!First) OS << " + "; else First = false; 398 OS << AM.Scale << "*reg("; 399 if (ScaledReg) 400 OS << *ScaledReg; 401 else 402 OS << "<unknown>"; 403 OS << ')'; 404 } 405 if (UnfoldedOffset != 0) { 406 if (!First) OS << " + "; else First = false; 407 OS << "imm(" << UnfoldedOffset << ')'; 408 } 409} 410 411void Formula::dump() const { 412 print(errs()); errs() << '\n'; 413} 414 415/// isAddRecSExtable - Return true if the given addrec can be sign-extended 416/// without changing its value. 417static bool isAddRecSExtable(const SCEVAddRecExpr *AR, ScalarEvolution &SE) { 418 Type *WideTy = 419 IntegerType::get(SE.getContext(), SE.getTypeSizeInBits(AR->getType()) + 1); 420 return isa<SCEVAddRecExpr>(SE.getSignExtendExpr(AR, WideTy)); 421} 422 423/// isAddSExtable - Return true if the given add can be sign-extended 424/// without changing its value. 425static bool isAddSExtable(const SCEVAddExpr *A, ScalarEvolution &SE) { 426 Type *WideTy = 427 IntegerType::get(SE.getContext(), SE.getTypeSizeInBits(A->getType()) + 1); 428 return isa<SCEVAddExpr>(SE.getSignExtendExpr(A, WideTy)); 429} 430 431/// isMulSExtable - Return true if the given mul can be sign-extended 432/// without changing its value. 433static bool isMulSExtable(const SCEVMulExpr *M, ScalarEvolution &SE) { 434 Type *WideTy = 435 IntegerType::get(SE.getContext(), 436 SE.getTypeSizeInBits(M->getType()) * M->getNumOperands()); 437 return isa<SCEVMulExpr>(SE.getSignExtendExpr(M, WideTy)); 438} 439 440/// getExactSDiv - Return an expression for LHS /s RHS, if it can be determined 441/// and if the remainder is known to be zero, or null otherwise. If 442/// IgnoreSignificantBits is true, expressions like (X * Y) /s Y are simplified 443/// to Y, ignoring that the multiplication may overflow, which is useful when 444/// the result will be used in a context where the most significant bits are 445/// ignored. 446static const SCEV *getExactSDiv(const SCEV *LHS, const SCEV *RHS, 447 ScalarEvolution &SE, 448 bool IgnoreSignificantBits = false) { 449 // Handle the trivial case, which works for any SCEV type. 450 if (LHS == RHS) 451 return SE.getConstant(LHS->getType(), 1); 452 453 // Handle a few RHS special cases. 454 const SCEVConstant *RC = dyn_cast<SCEVConstant>(RHS); 455 if (RC) { 456 const APInt &RA = RC->getValue()->getValue(); 457 // Handle x /s -1 as x * -1, to give ScalarEvolution a chance to do 458 // some folding. 459 if (RA.isAllOnesValue()) 460 return SE.getMulExpr(LHS, RC); 461 // Handle x /s 1 as x. 462 if (RA == 1) 463 return LHS; 464 } 465 466 // Check for a division of a constant by a constant. 467 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(LHS)) { 468 if (!RC) 469 return 0; 470 const APInt &LA = C->getValue()->getValue(); 471 const APInt &RA = RC->getValue()->getValue(); 472 if (LA.srem(RA) != 0) 473 return 0; 474 return SE.getConstant(LA.sdiv(RA)); 475 } 476 477 // Distribute the sdiv over addrec operands, if the addrec doesn't overflow. 478 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(LHS)) { 479 if (IgnoreSignificantBits || isAddRecSExtable(AR, SE)) { 480 const SCEV *Step = getExactSDiv(AR->getStepRecurrence(SE), RHS, SE, 481 IgnoreSignificantBits); 482 if (!Step) return 0; 483 const SCEV *Start = getExactSDiv(AR->getStart(), RHS, SE, 484 IgnoreSignificantBits); 485 if (!Start) return 0; 486 // FlagNW is independent of the start value, step direction, and is 487 // preserved with smaller magnitude steps. 488 // FIXME: AR->getNoWrapFlags(SCEV::FlagNW) 489 return SE.getAddRecExpr(Start, Step, AR->getLoop(), SCEV::FlagAnyWrap); 490 } 491 return 0; 492 } 493 494 // Distribute the sdiv over add operands, if the add doesn't overflow. 495 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(LHS)) { 496 if (IgnoreSignificantBits || isAddSExtable(Add, SE)) { 497 SmallVector<const SCEV *, 8> Ops; 498 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end(); 499 I != E; ++I) { 500 const SCEV *Op = getExactSDiv(*I, RHS, SE, 501 IgnoreSignificantBits); 502 if (!Op) return 0; 503 Ops.push_back(Op); 504 } 505 return SE.getAddExpr(Ops); 506 } 507 return 0; 508 } 509 510 // Check for a multiply operand that we can pull RHS out of. 511 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(LHS)) { 512 if (IgnoreSignificantBits || isMulSExtable(Mul, SE)) { 513 SmallVector<const SCEV *, 4> Ops; 514 bool Found = false; 515 for (SCEVMulExpr::op_iterator I = Mul->op_begin(), E = Mul->op_end(); 516 I != E; ++I) { 517 const SCEV *S = *I; 518 if (!Found) 519 if (const SCEV *Q = getExactSDiv(S, RHS, SE, 520 IgnoreSignificantBits)) { 521 S = Q; 522 Found = true; 523 } 524 Ops.push_back(S); 525 } 526 return Found ? SE.getMulExpr(Ops) : 0; 527 } 528 return 0; 529 } 530 531 // Otherwise we don't know. 532 return 0; 533} 534 535/// ExtractImmediate - If S involves the addition of a constant integer value, 536/// return that integer value, and mutate S to point to a new SCEV with that 537/// value excluded. 538static int64_t ExtractImmediate(const SCEV *&S, ScalarEvolution &SE) { 539 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S)) { 540 if (C->getValue()->getValue().getMinSignedBits() <= 64) { 541 S = SE.getConstant(C->getType(), 0); 542 return C->getValue()->getSExtValue(); 543 } 544 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) { 545 SmallVector<const SCEV *, 8> NewOps(Add->op_begin(), Add->op_end()); 546 int64_t Result = ExtractImmediate(NewOps.front(), SE); 547 if (Result != 0) 548 S = SE.getAddExpr(NewOps); 549 return Result; 550 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) { 551 SmallVector<const SCEV *, 8> NewOps(AR->op_begin(), AR->op_end()); 552 int64_t Result = ExtractImmediate(NewOps.front(), SE); 553 if (Result != 0) 554 S = SE.getAddRecExpr(NewOps, AR->getLoop(), 555 // FIXME: AR->getNoWrapFlags(SCEV::FlagNW) 556 SCEV::FlagAnyWrap); 557 return Result; 558 } 559 return 0; 560} 561 562/// ExtractSymbol - If S involves the addition of a GlobalValue address, 563/// return that symbol, and mutate S to point to a new SCEV with that 564/// value excluded. 565static GlobalValue *ExtractSymbol(const SCEV *&S, ScalarEvolution &SE) { 566 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) { 567 if (GlobalValue *GV = dyn_cast<GlobalValue>(U->getValue())) { 568 S = SE.getConstant(GV->getType(), 0); 569 return GV; 570 } 571 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) { 572 SmallVector<const SCEV *, 8> NewOps(Add->op_begin(), Add->op_end()); 573 GlobalValue *Result = ExtractSymbol(NewOps.back(), SE); 574 if (Result) 575 S = SE.getAddExpr(NewOps); 576 return Result; 577 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) { 578 SmallVector<const SCEV *, 8> NewOps(AR->op_begin(), AR->op_end()); 579 GlobalValue *Result = ExtractSymbol(NewOps.front(), SE); 580 if (Result) 581 S = SE.getAddRecExpr(NewOps, AR->getLoop(), 582 // FIXME: AR->getNoWrapFlags(SCEV::FlagNW) 583 SCEV::FlagAnyWrap); 584 return Result; 585 } 586 return 0; 587} 588 589/// isAddressUse - Returns true if the specified instruction is using the 590/// specified value as an address. 591static bool isAddressUse(Instruction *Inst, Value *OperandVal) { 592 bool isAddress = isa<LoadInst>(Inst); 593 if (StoreInst *SI = dyn_cast<StoreInst>(Inst)) { 594 if (SI->getOperand(1) == OperandVal) 595 isAddress = true; 596 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) { 597 // Addressing modes can also be folded into prefetches and a variety 598 // of intrinsics. 599 switch (II->getIntrinsicID()) { 600 default: break; 601 case Intrinsic::prefetch: 602 case Intrinsic::x86_sse_storeu_ps: 603 case Intrinsic::x86_sse2_storeu_pd: 604 case Intrinsic::x86_sse2_storeu_dq: 605 case Intrinsic::x86_sse2_storel_dq: 606 if (II->getArgOperand(0) == OperandVal) 607 isAddress = true; 608 break; 609 } 610 } 611 return isAddress; 612} 613 614/// getAccessType - Return the type of the memory being accessed. 615static Type *getAccessType(const Instruction *Inst) { 616 Type *AccessTy = Inst->getType(); 617 if (const StoreInst *SI = dyn_cast<StoreInst>(Inst)) 618 AccessTy = SI->getOperand(0)->getType(); 619 else if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst)) { 620 // Addressing modes can also be folded into prefetches and a variety 621 // of intrinsics. 622 switch (II->getIntrinsicID()) { 623 default: break; 624 case Intrinsic::x86_sse_storeu_ps: 625 case Intrinsic::x86_sse2_storeu_pd: 626 case Intrinsic::x86_sse2_storeu_dq: 627 case Intrinsic::x86_sse2_storel_dq: 628 AccessTy = II->getArgOperand(0)->getType(); 629 break; 630 } 631 } 632 633 // All pointers have the same requirements, so canonicalize them to an 634 // arbitrary pointer type to minimize variation. 635 if (PointerType *PTy = dyn_cast<PointerType>(AccessTy)) 636 AccessTy = PointerType::get(IntegerType::get(PTy->getContext(), 1), 637 PTy->getAddressSpace()); 638 639 return AccessTy; 640} 641 642/// isExistingPhi - Return true if this AddRec is already a phi in its loop. 643static bool isExistingPhi(const SCEVAddRecExpr *AR, ScalarEvolution &SE) { 644 for (BasicBlock::iterator I = AR->getLoop()->getHeader()->begin(); 645 PHINode *PN = dyn_cast<PHINode>(I); ++I) { 646 if (SE.isSCEVable(PN->getType()) && 647 (SE.getEffectiveSCEVType(PN->getType()) == 648 SE.getEffectiveSCEVType(AR->getType())) && 649 SE.getSCEV(PN) == AR) 650 return true; 651 } 652 return false; 653} 654 655/// Check if expanding this expression is likely to incur significant cost. This 656/// is tricky because SCEV doesn't track which expressions are actually computed 657/// by the current IR. 658/// 659/// We currently allow expansion of IV increments that involve adds, 660/// multiplication by constants, and AddRecs from existing phis. 661/// 662/// TODO: Allow UDivExpr if we can find an existing IV increment that is an 663/// obvious multiple of the UDivExpr. 664static bool isHighCostExpansion(const SCEV *S, 665 SmallPtrSet<const SCEV*, 8> &Processed, 666 ScalarEvolution &SE) { 667 // Zero/One operand expressions 668 switch (S->getSCEVType()) { 669 case scUnknown: 670 case scConstant: 671 return false; 672 case scTruncate: 673 return isHighCostExpansion(cast<SCEVTruncateExpr>(S)->getOperand(), 674 Processed, SE); 675 case scZeroExtend: 676 return isHighCostExpansion(cast<SCEVZeroExtendExpr>(S)->getOperand(), 677 Processed, SE); 678 case scSignExtend: 679 return isHighCostExpansion(cast<SCEVSignExtendExpr>(S)->getOperand(), 680 Processed, SE); 681 } 682 683 if (!Processed.insert(S)) 684 return false; 685 686 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) { 687 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end(); 688 I != E; ++I) { 689 if (isHighCostExpansion(*I, Processed, SE)) 690 return true; 691 } 692 return false; 693 } 694 695 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) { 696 if (Mul->getNumOperands() == 2) { 697 // Multiplication by a constant is ok 698 if (isa<SCEVConstant>(Mul->getOperand(0))) 699 return isHighCostExpansion(Mul->getOperand(1), Processed, SE); 700 701 // If we have the value of one operand, check if an existing 702 // multiplication already generates this expression. 703 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(Mul->getOperand(1))) { 704 Value *UVal = U->getValue(); 705 for (Value::use_iterator UI = UVal->use_begin(), UE = UVal->use_end(); 706 UI != UE; ++UI) { 707 // If U is a constant, it may be used by a ConstantExpr. 708 Instruction *User = dyn_cast<Instruction>(*UI); 709 if (User && User->getOpcode() == Instruction::Mul 710 && SE.isSCEVable(User->getType())) { 711 return SE.getSCEV(User) == Mul; 712 } 713 } 714 } 715 } 716 } 717 718 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) { 719 if (isExistingPhi(AR, SE)) 720 return false; 721 } 722 723 // Fow now, consider any other type of expression (div/mul/min/max) high cost. 724 return true; 725} 726 727/// DeleteTriviallyDeadInstructions - If any of the instructions is the 728/// specified set are trivially dead, delete them and see if this makes any of 729/// their operands subsequently dead. 730static bool 731DeleteTriviallyDeadInstructions(SmallVectorImpl<WeakVH> &DeadInsts) { 732 bool Changed = false; 733 734 while (!DeadInsts.empty()) { 735 Instruction *I = dyn_cast_or_null<Instruction>(&*DeadInsts.pop_back_val()); 736 737 if (I == 0 || !isInstructionTriviallyDead(I)) 738 continue; 739 740 for (User::op_iterator OI = I->op_begin(), E = I->op_end(); OI != E; ++OI) 741 if (Instruction *U = dyn_cast<Instruction>(*OI)) { 742 *OI = 0; 743 if (U->use_empty()) 744 DeadInsts.push_back(U); 745 } 746 747 I->eraseFromParent(); 748 Changed = true; 749 } 750 751 return Changed; 752} 753 754namespace { 755 756/// Cost - This class is used to measure and compare candidate formulae. 757class Cost { 758 /// TODO: Some of these could be merged. Also, a lexical ordering 759 /// isn't always optimal. 760 unsigned NumRegs; 761 unsigned AddRecCost; 762 unsigned NumIVMuls; 763 unsigned NumBaseAdds; 764 unsigned ImmCost; 765 unsigned SetupCost; 766 767public: 768 Cost() 769 : NumRegs(0), AddRecCost(0), NumIVMuls(0), NumBaseAdds(0), ImmCost(0), 770 SetupCost(0) {} 771 772 bool operator<(const Cost &Other) const; 773 774 void Loose(); 775 776#ifndef NDEBUG 777 // Once any of the metrics loses, they must all remain losers. 778 bool isValid() { 779 return ((NumRegs | AddRecCost | NumIVMuls | NumBaseAdds 780 | ImmCost | SetupCost) != ~0u) 781 || ((NumRegs & AddRecCost & NumIVMuls & NumBaseAdds 782 & ImmCost & SetupCost) == ~0u); 783 } 784#endif 785 786 bool isLoser() { 787 assert(isValid() && "invalid cost"); 788 return NumRegs == ~0u; 789 } 790 791 void RateFormula(const Formula &F, 792 SmallPtrSet<const SCEV *, 16> &Regs, 793 const DenseSet<const SCEV *> &VisitedRegs, 794 const Loop *L, 795 const SmallVectorImpl<int64_t> &Offsets, 796 ScalarEvolution &SE, DominatorTree &DT, 797 SmallPtrSet<const SCEV *, 16> *LoserRegs = 0); 798 799 void print(raw_ostream &OS) const; 800 void dump() const; 801 802private: 803 void RateRegister(const SCEV *Reg, 804 SmallPtrSet<const SCEV *, 16> &Regs, 805 const Loop *L, 806 ScalarEvolution &SE, DominatorTree &DT); 807 void RatePrimaryRegister(const SCEV *Reg, 808 SmallPtrSet<const SCEV *, 16> &Regs, 809 const Loop *L, 810 ScalarEvolution &SE, DominatorTree &DT, 811 SmallPtrSet<const SCEV *, 16> *LoserRegs); 812}; 813 814} 815 816/// RateRegister - Tally up interesting quantities from the given register. 817void Cost::RateRegister(const SCEV *Reg, 818 SmallPtrSet<const SCEV *, 16> &Regs, 819 const Loop *L, 820 ScalarEvolution &SE, DominatorTree &DT) { 821 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Reg)) { 822 // If this is an addrec for another loop, don't second-guess its addrec phi 823 // nodes. LSR isn't currently smart enough to reason about more than one 824 // loop at a time. LSR has already run on inner loops, will not run on outer 825 // loops, and cannot be expected to change sibling loops. 826 if (AR->getLoop() != L) { 827 // If the AddRec exists, consider it's register free and leave it alone. 828 if (isExistingPhi(AR, SE)) 829 return; 830 831 // Otherwise, do not consider this formula at all. 832 Loose(); 833 return; 834 } 835 AddRecCost += 1; /// TODO: This should be a function of the stride. 836 837 // Add the step value register, if it needs one. 838 // TODO: The non-affine case isn't precisely modeled here. 839 if (!AR->isAffine() || !isa<SCEVConstant>(AR->getOperand(1))) { 840 if (!Regs.count(AR->getOperand(1))) { 841 RateRegister(AR->getOperand(1), Regs, L, SE, DT); 842 if (isLoser()) 843 return; 844 } 845 } 846 } 847 ++NumRegs; 848 849 // Rough heuristic; favor registers which don't require extra setup 850 // instructions in the preheader. 851 if (!isa<SCEVUnknown>(Reg) && 852 !isa<SCEVConstant>(Reg) && 853 !(isa<SCEVAddRecExpr>(Reg) && 854 (isa<SCEVUnknown>(cast<SCEVAddRecExpr>(Reg)->getStart()) || 855 isa<SCEVConstant>(cast<SCEVAddRecExpr>(Reg)->getStart())))) 856 ++SetupCost; 857 858 NumIVMuls += isa<SCEVMulExpr>(Reg) && 859 SE.hasComputableLoopEvolution(Reg, L); 860} 861 862/// RatePrimaryRegister - Record this register in the set. If we haven't seen it 863/// before, rate it. Optional LoserRegs provides a way to declare any formula 864/// that refers to one of those regs an instant loser. 865void Cost::RatePrimaryRegister(const SCEV *Reg, 866 SmallPtrSet<const SCEV *, 16> &Regs, 867 const Loop *L, 868 ScalarEvolution &SE, DominatorTree &DT, 869 SmallPtrSet<const SCEV *, 16> *LoserRegs) { 870 if (LoserRegs && LoserRegs->count(Reg)) { 871 Loose(); 872 return; 873 } 874 if (Regs.insert(Reg)) { 875 RateRegister(Reg, Regs, L, SE, DT); 876 if (isLoser()) 877 LoserRegs->insert(Reg); 878 } 879} 880 881void Cost::RateFormula(const Formula &F, 882 SmallPtrSet<const SCEV *, 16> &Regs, 883 const DenseSet<const SCEV *> &VisitedRegs, 884 const Loop *L, 885 const SmallVectorImpl<int64_t> &Offsets, 886 ScalarEvolution &SE, DominatorTree &DT, 887 SmallPtrSet<const SCEV *, 16> *LoserRegs) { 888 // Tally up the registers. 889 if (const SCEV *ScaledReg = F.ScaledReg) { 890 if (VisitedRegs.count(ScaledReg)) { 891 Loose(); 892 return; 893 } 894 RatePrimaryRegister(ScaledReg, Regs, L, SE, DT, LoserRegs); 895 if (isLoser()) 896 return; 897 } 898 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(), 899 E = F.BaseRegs.end(); I != E; ++I) { 900 const SCEV *BaseReg = *I; 901 if (VisitedRegs.count(BaseReg)) { 902 Loose(); 903 return; 904 } 905 RatePrimaryRegister(BaseReg, Regs, L, SE, DT, LoserRegs); 906 if (isLoser()) 907 return; 908 } 909 910 // Determine how many (unfolded) adds we'll need inside the loop. 911 size_t NumBaseParts = F.BaseRegs.size() + (F.UnfoldedOffset != 0); 912 if (NumBaseParts > 1) 913 NumBaseAdds += NumBaseParts - 1; 914 915 // Tally up the non-zero immediates. 916 for (SmallVectorImpl<int64_t>::const_iterator I = Offsets.begin(), 917 E = Offsets.end(); I != E; ++I) { 918 int64_t Offset = (uint64_t)*I + F.AM.BaseOffs; 919 if (F.AM.BaseGV) 920 ImmCost += 64; // Handle symbolic values conservatively. 921 // TODO: This should probably be the pointer size. 922 else if (Offset != 0) 923 ImmCost += APInt(64, Offset, true).getMinSignedBits(); 924 } 925 assert(isValid() && "invalid cost"); 926} 927 928/// Loose - Set this cost to a losing value. 929void Cost::Loose() { 930 NumRegs = ~0u; 931 AddRecCost = ~0u; 932 NumIVMuls = ~0u; 933 NumBaseAdds = ~0u; 934 ImmCost = ~0u; 935 SetupCost = ~0u; 936} 937 938/// operator< - Choose the lower cost. 939bool Cost::operator<(const Cost &Other) const { 940 if (NumRegs != Other.NumRegs) 941 return NumRegs < Other.NumRegs; 942 if (AddRecCost != Other.AddRecCost) 943 return AddRecCost < Other.AddRecCost; 944 if (NumIVMuls != Other.NumIVMuls) 945 return NumIVMuls < Other.NumIVMuls; 946 if (NumBaseAdds != Other.NumBaseAdds) 947 return NumBaseAdds < Other.NumBaseAdds; 948 if (ImmCost != Other.ImmCost) 949 return ImmCost < Other.ImmCost; 950 if (SetupCost != Other.SetupCost) 951 return SetupCost < Other.SetupCost; 952 return false; 953} 954 955void Cost::print(raw_ostream &OS) const { 956 OS << NumRegs << " reg" << (NumRegs == 1 ? "" : "s"); 957 if (AddRecCost != 0) 958 OS << ", with addrec cost " << AddRecCost; 959 if (NumIVMuls != 0) 960 OS << ", plus " << NumIVMuls << " IV mul" << (NumIVMuls == 1 ? "" : "s"); 961 if (NumBaseAdds != 0) 962 OS << ", plus " << NumBaseAdds << " base add" 963 << (NumBaseAdds == 1 ? "" : "s"); 964 if (ImmCost != 0) 965 OS << ", plus " << ImmCost << " imm cost"; 966 if (SetupCost != 0) 967 OS << ", plus " << SetupCost << " setup cost"; 968} 969 970void Cost::dump() const { 971 print(errs()); errs() << '\n'; 972} 973 974namespace { 975 976/// LSRFixup - An operand value in an instruction which is to be replaced 977/// with some equivalent, possibly strength-reduced, replacement. 978struct LSRFixup { 979 /// UserInst - The instruction which will be updated. 980 Instruction *UserInst; 981 982 /// OperandValToReplace - The operand of the instruction which will 983 /// be replaced. The operand may be used more than once; every instance 984 /// will be replaced. 985 Value *OperandValToReplace; 986 987 /// PostIncLoops - If this user is to use the post-incremented value of an 988 /// induction variable, this variable is non-null and holds the loop 989 /// associated with the induction variable. 990 PostIncLoopSet PostIncLoops; 991 992 /// LUIdx - The index of the LSRUse describing the expression which 993 /// this fixup needs, minus an offset (below). 994 size_t LUIdx; 995 996 /// Offset - A constant offset to be added to the LSRUse expression. 997 /// This allows multiple fixups to share the same LSRUse with different 998 /// offsets, for example in an unrolled loop. 999 int64_t Offset; 1000 1001 bool isUseFullyOutsideLoop(const Loop *L) const; 1002 1003 LSRFixup(); 1004 1005 void print(raw_ostream &OS) const; 1006 void dump() const; 1007}; 1008 1009} 1010 1011LSRFixup::LSRFixup() 1012 : UserInst(0), OperandValToReplace(0), LUIdx(~size_t(0)), Offset(0) {} 1013 1014/// isUseFullyOutsideLoop - Test whether this fixup always uses its 1015/// value outside of the given loop. 1016bool LSRFixup::isUseFullyOutsideLoop(const Loop *L) const { 1017 // PHI nodes use their value in their incoming blocks. 1018 if (const PHINode *PN = dyn_cast<PHINode>(UserInst)) { 1019 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) 1020 if (PN->getIncomingValue(i) == OperandValToReplace && 1021 L->contains(PN->getIncomingBlock(i))) 1022 return false; 1023 return true; 1024 } 1025 1026 return !L->contains(UserInst); 1027} 1028 1029void LSRFixup::print(raw_ostream &OS) const { 1030 OS << "UserInst="; 1031 // Store is common and interesting enough to be worth special-casing. 1032 if (StoreInst *Store = dyn_cast<StoreInst>(UserInst)) { 1033 OS << "store "; 1034 WriteAsOperand(OS, Store->getOperand(0), /*PrintType=*/false); 1035 } else if (UserInst->getType()->isVoidTy()) 1036 OS << UserInst->getOpcodeName(); 1037 else 1038 WriteAsOperand(OS, UserInst, /*PrintType=*/false); 1039 1040 OS << ", OperandValToReplace="; 1041 WriteAsOperand(OS, OperandValToReplace, /*PrintType=*/false); 1042 1043 for (PostIncLoopSet::const_iterator I = PostIncLoops.begin(), 1044 E = PostIncLoops.end(); I != E; ++I) { 1045 OS << ", PostIncLoop="; 1046 WriteAsOperand(OS, (*I)->getHeader(), /*PrintType=*/false); 1047 } 1048 1049 if (LUIdx != ~size_t(0)) 1050 OS << ", LUIdx=" << LUIdx; 1051 1052 if (Offset != 0) 1053 OS << ", Offset=" << Offset; 1054} 1055 1056void LSRFixup::dump() const { 1057 print(errs()); errs() << '\n'; 1058} 1059 1060namespace { 1061 1062/// UniquifierDenseMapInfo - A DenseMapInfo implementation for holding 1063/// DenseMaps and DenseSets of sorted SmallVectors of const SCEV*. 1064struct UniquifierDenseMapInfo { 1065 static SmallVector<const SCEV *, 2> getEmptyKey() { 1066 SmallVector<const SCEV *, 2> V; 1067 V.push_back(reinterpret_cast<const SCEV *>(-1)); 1068 return V; 1069 } 1070 1071 static SmallVector<const SCEV *, 2> getTombstoneKey() { 1072 SmallVector<const SCEV *, 2> V; 1073 V.push_back(reinterpret_cast<const SCEV *>(-2)); 1074 return V; 1075 } 1076 1077 static unsigned getHashValue(const SmallVector<const SCEV *, 2> &V) { 1078 unsigned Result = 0; 1079 for (SmallVectorImpl<const SCEV *>::const_iterator I = V.begin(), 1080 E = V.end(); I != E; ++I) 1081 Result ^= DenseMapInfo<const SCEV *>::getHashValue(*I); 1082 return Result; 1083 } 1084 1085 static bool isEqual(const SmallVector<const SCEV *, 2> &LHS, 1086 const SmallVector<const SCEV *, 2> &RHS) { 1087 return LHS == RHS; 1088 } 1089}; 1090 1091/// LSRUse - This class holds the state that LSR keeps for each use in 1092/// IVUsers, as well as uses invented by LSR itself. It includes information 1093/// about what kinds of things can be folded into the user, information about 1094/// the user itself, and information about how the use may be satisfied. 1095/// TODO: Represent multiple users of the same expression in common? 1096class LSRUse { 1097 DenseSet<SmallVector<const SCEV *, 2>, UniquifierDenseMapInfo> Uniquifier; 1098 1099public: 1100 /// KindType - An enum for a kind of use, indicating what types of 1101 /// scaled and immediate operands it might support. 1102 enum KindType { 1103 Basic, ///< A normal use, with no folding. 1104 Special, ///< A special case of basic, allowing -1 scales. 1105 Address, ///< An address use; folding according to TargetLowering 1106 ICmpZero ///< An equality icmp with both operands folded into one. 1107 // TODO: Add a generic icmp too? 1108 }; 1109 1110 KindType Kind; 1111 Type *AccessTy; 1112 1113 SmallVector<int64_t, 8> Offsets; 1114 int64_t MinOffset; 1115 int64_t MaxOffset; 1116 1117 /// AllFixupsOutsideLoop - This records whether all of the fixups using this 1118 /// LSRUse are outside of the loop, in which case some special-case heuristics 1119 /// may be used. 1120 bool AllFixupsOutsideLoop; 1121 1122 /// WidestFixupType - This records the widest use type for any fixup using 1123 /// this LSRUse. FindUseWithSimilarFormula can't consider uses with different 1124 /// max fixup widths to be equivalent, because the narrower one may be relying 1125 /// on the implicit truncation to truncate away bogus bits. 1126 Type *WidestFixupType; 1127 1128 /// Formulae - A list of ways to build a value that can satisfy this user. 1129 /// After the list is populated, one of these is selected heuristically and 1130 /// used to formulate a replacement for OperandValToReplace in UserInst. 1131 SmallVector<Formula, 12> Formulae; 1132 1133 /// Regs - The set of register candidates used by all formulae in this LSRUse. 1134 SmallPtrSet<const SCEV *, 4> Regs; 1135 1136 LSRUse(KindType K, Type *T) : Kind(K), AccessTy(T), 1137 MinOffset(INT64_MAX), 1138 MaxOffset(INT64_MIN), 1139 AllFixupsOutsideLoop(true), 1140 WidestFixupType(0) {} 1141 1142 bool HasFormulaWithSameRegs(const Formula &F) const; 1143 bool InsertFormula(const Formula &F); 1144 void DeleteFormula(Formula &F); 1145 void RecomputeRegs(size_t LUIdx, RegUseTracker &Reguses); 1146 1147 void print(raw_ostream &OS) const; 1148 void dump() const; 1149}; 1150 1151} 1152 1153/// HasFormula - Test whether this use as a formula which has the same 1154/// registers as the given formula. 1155bool LSRUse::HasFormulaWithSameRegs(const Formula &F) const { 1156 SmallVector<const SCEV *, 2> Key = F.BaseRegs; 1157 if (F.ScaledReg) Key.push_back(F.ScaledReg); 1158 // Unstable sort by host order ok, because this is only used for uniquifying. 1159 std::sort(Key.begin(), Key.end()); 1160 return Uniquifier.count(Key); 1161} 1162 1163/// InsertFormula - If the given formula has not yet been inserted, add it to 1164/// the list, and return true. Return false otherwise. 1165bool LSRUse::InsertFormula(const Formula &F) { 1166 SmallVector<const SCEV *, 2> Key = F.BaseRegs; 1167 if (F.ScaledReg) Key.push_back(F.ScaledReg); 1168 // Unstable sort by host order ok, because this is only used for uniquifying. 1169 std::sort(Key.begin(), Key.end()); 1170 1171 if (!Uniquifier.insert(Key).second) 1172 return false; 1173 1174 // Using a register to hold the value of 0 is not profitable. 1175 assert((!F.ScaledReg || !F.ScaledReg->isZero()) && 1176 "Zero allocated in a scaled register!"); 1177#ifndef NDEBUG 1178 for (SmallVectorImpl<const SCEV *>::const_iterator I = 1179 F.BaseRegs.begin(), E = F.BaseRegs.end(); I != E; ++I) 1180 assert(!(*I)->isZero() && "Zero allocated in a base register!"); 1181#endif 1182 1183 // Add the formula to the list. 1184 Formulae.push_back(F); 1185 1186 // Record registers now being used by this use. 1187 Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end()); 1188 1189 return true; 1190} 1191 1192/// DeleteFormula - Remove the given formula from this use's list. 1193void LSRUse::DeleteFormula(Formula &F) { 1194 if (&F != &Formulae.back()) 1195 std::swap(F, Formulae.back()); 1196 Formulae.pop_back(); 1197} 1198 1199/// RecomputeRegs - Recompute the Regs field, and update RegUses. 1200void LSRUse::RecomputeRegs(size_t LUIdx, RegUseTracker &RegUses) { 1201 // Now that we've filtered out some formulae, recompute the Regs set. 1202 SmallPtrSet<const SCEV *, 4> OldRegs = Regs; 1203 Regs.clear(); 1204 for (SmallVectorImpl<Formula>::const_iterator I = Formulae.begin(), 1205 E = Formulae.end(); I != E; ++I) { 1206 const Formula &F = *I; 1207 if (F.ScaledReg) Regs.insert(F.ScaledReg); 1208 Regs.insert(F.BaseRegs.begin(), F.BaseRegs.end()); 1209 } 1210 1211 // Update the RegTracker. 1212 for (SmallPtrSet<const SCEV *, 4>::iterator I = OldRegs.begin(), 1213 E = OldRegs.end(); I != E; ++I) 1214 if (!Regs.count(*I)) 1215 RegUses.DropRegister(*I, LUIdx); 1216} 1217 1218void LSRUse::print(raw_ostream &OS) const { 1219 OS << "LSR Use: Kind="; 1220 switch (Kind) { 1221 case Basic: OS << "Basic"; break; 1222 case Special: OS << "Special"; break; 1223 case ICmpZero: OS << "ICmpZero"; break; 1224 case Address: 1225 OS << "Address of "; 1226 if (AccessTy->isPointerTy()) 1227 OS << "pointer"; // the full pointer type could be really verbose 1228 else 1229 OS << *AccessTy; 1230 } 1231 1232 OS << ", Offsets={"; 1233 for (SmallVectorImpl<int64_t>::const_iterator I = Offsets.begin(), 1234 E = Offsets.end(); I != E; ++I) { 1235 OS << *I; 1236 if (llvm::next(I) != E) 1237 OS << ','; 1238 } 1239 OS << '}'; 1240 1241 if (AllFixupsOutsideLoop) 1242 OS << ", all-fixups-outside-loop"; 1243 1244 if (WidestFixupType) 1245 OS << ", widest fixup type: " << *WidestFixupType; 1246} 1247 1248void LSRUse::dump() const { 1249 print(errs()); errs() << '\n'; 1250} 1251 1252/// isLegalUse - Test whether the use described by AM is "legal", meaning it can 1253/// be completely folded into the user instruction at isel time. This includes 1254/// address-mode folding and special icmp tricks. 1255static bool isLegalUse(const TargetLowering::AddrMode &AM, 1256 LSRUse::KindType Kind, Type *AccessTy, 1257 const TargetLowering *TLI) { 1258 switch (Kind) { 1259 case LSRUse::Address: 1260 // If we have low-level target information, ask the target if it can 1261 // completely fold this address. 1262 if (TLI) return TLI->isLegalAddressingMode(AM, AccessTy); 1263 1264 // Otherwise, just guess that reg+reg addressing is legal. 1265 return !AM.BaseGV && AM.BaseOffs == 0 && AM.Scale <= 1; 1266 1267 case LSRUse::ICmpZero: 1268 // There's not even a target hook for querying whether it would be legal to 1269 // fold a GV into an ICmp. 1270 if (AM.BaseGV) 1271 return false; 1272 1273 // ICmp only has two operands; don't allow more than two non-trivial parts. 1274 if (AM.Scale != 0 && AM.HasBaseReg && AM.BaseOffs != 0) 1275 return false; 1276 1277 // ICmp only supports no scale or a -1 scale, as we can "fold" a -1 scale by 1278 // putting the scaled register in the other operand of the icmp. 1279 if (AM.Scale != 0 && AM.Scale != -1) 1280 return false; 1281 1282 // If we have low-level target information, ask the target if it can fold an 1283 // integer immediate on an icmp. 1284 if (AM.BaseOffs != 0) { 1285 if (!TLI) 1286 return false; 1287 // We have one of: 1288 // ICmpZero BaseReg + Offset => ICmp BaseReg, -Offset 1289 // ICmpZero -1*ScaleReg + Offset => ICmp ScaleReg, Offset 1290 // Offs is the ICmp immediate. 1291 int64_t Offs = AM.BaseOffs; 1292 if (AM.Scale == 0) 1293 Offs = -(uint64_t)Offs; // The cast does the right thing with INT64_MIN. 1294 return TLI->isLegalICmpImmediate(Offs); 1295 } 1296 1297 // ICmpZero BaseReg + -1*ScaleReg => ICmp BaseReg, ScaleReg 1298 return true; 1299 1300 case LSRUse::Basic: 1301 // Only handle single-register values. 1302 return !AM.BaseGV && AM.Scale == 0 && AM.BaseOffs == 0; 1303 1304 case LSRUse::Special: 1305 // Only handle -1 scales, or no scale. 1306 return AM.Scale == 0 || AM.Scale == -1; 1307 } 1308 1309 llvm_unreachable("Invalid LSRUse Kind!"); 1310} 1311 1312static bool isLegalUse(TargetLowering::AddrMode AM, 1313 int64_t MinOffset, int64_t MaxOffset, 1314 LSRUse::KindType Kind, Type *AccessTy, 1315 const TargetLowering *TLI) { 1316 // Check for overflow. 1317 if (((int64_t)((uint64_t)AM.BaseOffs + MinOffset) > AM.BaseOffs) != 1318 (MinOffset > 0)) 1319 return false; 1320 AM.BaseOffs = (uint64_t)AM.BaseOffs + MinOffset; 1321 if (isLegalUse(AM, Kind, AccessTy, TLI)) { 1322 AM.BaseOffs = (uint64_t)AM.BaseOffs - MinOffset; 1323 // Check for overflow. 1324 if (((int64_t)((uint64_t)AM.BaseOffs + MaxOffset) > AM.BaseOffs) != 1325 (MaxOffset > 0)) 1326 return false; 1327 AM.BaseOffs = (uint64_t)AM.BaseOffs + MaxOffset; 1328 return isLegalUse(AM, Kind, AccessTy, TLI); 1329 } 1330 return false; 1331} 1332 1333static bool isAlwaysFoldable(int64_t BaseOffs, 1334 GlobalValue *BaseGV, 1335 bool HasBaseReg, 1336 LSRUse::KindType Kind, Type *AccessTy, 1337 const TargetLowering *TLI) { 1338 // Fast-path: zero is always foldable. 1339 if (BaseOffs == 0 && !BaseGV) return true; 1340 1341 // Conservatively, create an address with an immediate and a 1342 // base and a scale. 1343 TargetLowering::AddrMode AM; 1344 AM.BaseOffs = BaseOffs; 1345 AM.BaseGV = BaseGV; 1346 AM.HasBaseReg = HasBaseReg; 1347 AM.Scale = Kind == LSRUse::ICmpZero ? -1 : 1; 1348 1349 // Canonicalize a scale of 1 to a base register if the formula doesn't 1350 // already have a base register. 1351 if (!AM.HasBaseReg && AM.Scale == 1) { 1352 AM.Scale = 0; 1353 AM.HasBaseReg = true; 1354 } 1355 1356 return isLegalUse(AM, Kind, AccessTy, TLI); 1357} 1358 1359static bool isAlwaysFoldable(const SCEV *S, 1360 int64_t MinOffset, int64_t MaxOffset, 1361 bool HasBaseReg, 1362 LSRUse::KindType Kind, Type *AccessTy, 1363 const TargetLowering *TLI, 1364 ScalarEvolution &SE) { 1365 // Fast-path: zero is always foldable. 1366 if (S->isZero()) return true; 1367 1368 // Conservatively, create an address with an immediate and a 1369 // base and a scale. 1370 int64_t BaseOffs = ExtractImmediate(S, SE); 1371 GlobalValue *BaseGV = ExtractSymbol(S, SE); 1372 1373 // If there's anything else involved, it's not foldable. 1374 if (!S->isZero()) return false; 1375 1376 // Fast-path: zero is always foldable. 1377 if (BaseOffs == 0 && !BaseGV) return true; 1378 1379 // Conservatively, create an address with an immediate and a 1380 // base and a scale. 1381 TargetLowering::AddrMode AM; 1382 AM.BaseOffs = BaseOffs; 1383 AM.BaseGV = BaseGV; 1384 AM.HasBaseReg = HasBaseReg; 1385 AM.Scale = Kind == LSRUse::ICmpZero ? -1 : 1; 1386 1387 return isLegalUse(AM, MinOffset, MaxOffset, Kind, AccessTy, TLI); 1388} 1389 1390namespace { 1391 1392/// UseMapDenseMapInfo - A DenseMapInfo implementation for holding 1393/// DenseMaps and DenseSets of pairs of const SCEV* and LSRUse::Kind. 1394struct UseMapDenseMapInfo { 1395 static std::pair<const SCEV *, LSRUse::KindType> getEmptyKey() { 1396 return std::make_pair(reinterpret_cast<const SCEV *>(-1), LSRUse::Basic); 1397 } 1398 1399 static std::pair<const SCEV *, LSRUse::KindType> getTombstoneKey() { 1400 return std::make_pair(reinterpret_cast<const SCEV *>(-2), LSRUse::Basic); 1401 } 1402 1403 static unsigned 1404 getHashValue(const std::pair<const SCEV *, LSRUse::KindType> &V) { 1405 unsigned Result = DenseMapInfo<const SCEV *>::getHashValue(V.first); 1406 Result ^= DenseMapInfo<unsigned>::getHashValue(unsigned(V.second)); 1407 return Result; 1408 } 1409 1410 static bool isEqual(const std::pair<const SCEV *, LSRUse::KindType> &LHS, 1411 const std::pair<const SCEV *, LSRUse::KindType> &RHS) { 1412 return LHS == RHS; 1413 } 1414}; 1415 1416/// IVInc - An individual increment in a Chain of IV increments. 1417/// Relate an IV user to an expression that computes the IV it uses from the IV 1418/// used by the previous link in the Chain. 1419/// 1420/// For the head of a chain, IncExpr holds the absolute SCEV expression for the 1421/// original IVOperand. The head of the chain's IVOperand is only valid during 1422/// chain collection, before LSR replaces IV users. During chain generation, 1423/// IncExpr can be used to find the new IVOperand that computes the same 1424/// expression. 1425struct IVInc { 1426 Instruction *UserInst; 1427 Value* IVOperand; 1428 const SCEV *IncExpr; 1429 1430 IVInc(Instruction *U, Value *O, const SCEV *E): 1431 UserInst(U), IVOperand(O), IncExpr(E) {} 1432}; 1433 1434// IVChain - The list of IV increments in program order. 1435// We typically add the head of a chain without finding subsequent links. 1436typedef SmallVector<IVInc,1> IVChain; 1437 1438/// ChainUsers - Helper for CollectChains to track multiple IV increment uses. 1439/// Distinguish between FarUsers that definitely cross IV increments and 1440/// NearUsers that may be used between IV increments. 1441struct ChainUsers { 1442 SmallPtrSet<Instruction*, 4> FarUsers; 1443 SmallPtrSet<Instruction*, 4> NearUsers; 1444}; 1445 1446/// LSRInstance - This class holds state for the main loop strength reduction 1447/// logic. 1448class LSRInstance { 1449 IVUsers &IU; 1450 ScalarEvolution &SE; 1451 DominatorTree &DT; 1452 LoopInfo &LI; 1453 const TargetLowering *const TLI; 1454 Loop *const L; 1455 bool Changed; 1456 1457 /// IVIncInsertPos - This is the insert position that the current loop's 1458 /// induction variable increment should be placed. In simple loops, this is 1459 /// the latch block's terminator. But in more complicated cases, this is a 1460 /// position which will dominate all the in-loop post-increment users. 1461 Instruction *IVIncInsertPos; 1462 1463 /// Factors - Interesting factors between use strides. 1464 SmallSetVector<int64_t, 8> Factors; 1465 1466 /// Types - Interesting use types, to facilitate truncation reuse. 1467 SmallSetVector<Type *, 4> Types; 1468 1469 /// Fixups - The list of operands which are to be replaced. 1470 SmallVector<LSRFixup, 16> Fixups; 1471 1472 /// Uses - The list of interesting uses. 1473 SmallVector<LSRUse, 16> Uses; 1474 1475 /// RegUses - Track which uses use which register candidates. 1476 RegUseTracker RegUses; 1477 1478 // Limit the number of chains to avoid quadratic behavior. We don't expect to 1479 // have more than a few IV increment chains in a loop. Missing a Chain falls 1480 // back to normal LSR behavior for those uses. 1481 static const unsigned MaxChains = 8; 1482 1483 /// IVChainVec - IV users can form a chain of IV increments. 1484 SmallVector<IVChain, MaxChains> IVChainVec; 1485 1486 /// IVIncSet - IV users that belong to profitable IVChains. 1487 SmallPtrSet<Use*, MaxChains> IVIncSet; 1488 1489 void OptimizeShadowIV(); 1490 bool FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse); 1491 ICmpInst *OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse); 1492 void OptimizeLoopTermCond(); 1493 1494 void ChainInstruction(Instruction *UserInst, Instruction *IVOper, 1495 SmallVectorImpl<ChainUsers> &ChainUsersVec); 1496 void FinalizeChain(IVChain &Chain); 1497 void CollectChains(); 1498 void GenerateIVChain(const IVChain &Chain, SCEVExpander &Rewriter, 1499 SmallVectorImpl<WeakVH> &DeadInsts); 1500 1501 void CollectInterestingTypesAndFactors(); 1502 void CollectFixupsAndInitialFormulae(); 1503 1504 LSRFixup &getNewFixup() { 1505 Fixups.push_back(LSRFixup()); 1506 return Fixups.back(); 1507 } 1508 1509 // Support for sharing of LSRUses between LSRFixups. 1510 typedef DenseMap<std::pair<const SCEV *, LSRUse::KindType>, 1511 size_t, 1512 UseMapDenseMapInfo> UseMapTy; 1513 UseMapTy UseMap; 1514 1515 bool reconcileNewOffset(LSRUse &LU, int64_t NewOffset, bool HasBaseReg, 1516 LSRUse::KindType Kind, Type *AccessTy); 1517 1518 std::pair<size_t, int64_t> getUse(const SCEV *&Expr, 1519 LSRUse::KindType Kind, 1520 Type *AccessTy); 1521 1522 void DeleteUse(LSRUse &LU, size_t LUIdx); 1523 1524 LSRUse *FindUseWithSimilarFormula(const Formula &F, const LSRUse &OrigLU); 1525 1526 void InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx); 1527 void InsertSupplementalFormula(const SCEV *S, LSRUse &LU, size_t LUIdx); 1528 void CountRegisters(const Formula &F, size_t LUIdx); 1529 bool InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F); 1530 1531 void CollectLoopInvariantFixupsAndFormulae(); 1532 1533 void GenerateReassociations(LSRUse &LU, unsigned LUIdx, Formula Base, 1534 unsigned Depth = 0); 1535 void GenerateCombinations(LSRUse &LU, unsigned LUIdx, Formula Base); 1536 void GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx, Formula Base); 1537 void GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx, Formula Base); 1538 void GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx, Formula Base); 1539 void GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base); 1540 void GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base); 1541 void GenerateCrossUseConstantOffsets(); 1542 void GenerateAllReuseFormulae(); 1543 1544 void FilterOutUndesirableDedicatedRegisters(); 1545 1546 size_t EstimateSearchSpaceComplexity() const; 1547 void NarrowSearchSpaceByDetectingSupersets(); 1548 void NarrowSearchSpaceByCollapsingUnrolledCode(); 1549 void NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters(); 1550 void NarrowSearchSpaceByPickingWinnerRegs(); 1551 void NarrowSearchSpaceUsingHeuristics(); 1552 1553 void SolveRecurse(SmallVectorImpl<const Formula *> &Solution, 1554 Cost &SolutionCost, 1555 SmallVectorImpl<const Formula *> &Workspace, 1556 const Cost &CurCost, 1557 const SmallPtrSet<const SCEV *, 16> &CurRegs, 1558 DenseSet<const SCEV *> &VisitedRegs) const; 1559 void Solve(SmallVectorImpl<const Formula *> &Solution) const; 1560 1561 BasicBlock::iterator 1562 HoistInsertPosition(BasicBlock::iterator IP, 1563 const SmallVectorImpl<Instruction *> &Inputs) const; 1564 BasicBlock::iterator 1565 AdjustInsertPositionForExpand(BasicBlock::iterator IP, 1566 const LSRFixup &LF, 1567 const LSRUse &LU, 1568 SCEVExpander &Rewriter) const; 1569 1570 Value *Expand(const LSRFixup &LF, 1571 const Formula &F, 1572 BasicBlock::iterator IP, 1573 SCEVExpander &Rewriter, 1574 SmallVectorImpl<WeakVH> &DeadInsts) const; 1575 void RewriteForPHI(PHINode *PN, const LSRFixup &LF, 1576 const Formula &F, 1577 SCEVExpander &Rewriter, 1578 SmallVectorImpl<WeakVH> &DeadInsts, 1579 Pass *P) const; 1580 void Rewrite(const LSRFixup &LF, 1581 const Formula &F, 1582 SCEVExpander &Rewriter, 1583 SmallVectorImpl<WeakVH> &DeadInsts, 1584 Pass *P) const; 1585 void ImplementSolution(const SmallVectorImpl<const Formula *> &Solution, 1586 Pass *P); 1587 1588public: 1589 LSRInstance(const TargetLowering *tli, Loop *l, Pass *P); 1590 1591 bool getChanged() const { return Changed; } 1592 1593 void print_factors_and_types(raw_ostream &OS) const; 1594 void print_fixups(raw_ostream &OS) const; 1595 void print_uses(raw_ostream &OS) const; 1596 void print(raw_ostream &OS) const; 1597 void dump() const; 1598}; 1599 1600} 1601 1602/// OptimizeShadowIV - If IV is used in a int-to-float cast 1603/// inside the loop then try to eliminate the cast operation. 1604void LSRInstance::OptimizeShadowIV() { 1605 const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L); 1606 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount)) 1607 return; 1608 1609 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); 1610 UI != E; /* empty */) { 1611 IVUsers::const_iterator CandidateUI = UI; 1612 ++UI; 1613 Instruction *ShadowUse = CandidateUI->getUser(); 1614 Type *DestTy = NULL; 1615 bool IsSigned = false; 1616 1617 /* If shadow use is a int->float cast then insert a second IV 1618 to eliminate this cast. 1619 1620 for (unsigned i = 0; i < n; ++i) 1621 foo((double)i); 1622 1623 is transformed into 1624 1625 double d = 0.0; 1626 for (unsigned i = 0; i < n; ++i, ++d) 1627 foo(d); 1628 */ 1629 if (UIToFPInst *UCast = dyn_cast<UIToFPInst>(CandidateUI->getUser())) { 1630 IsSigned = false; 1631 DestTy = UCast->getDestTy(); 1632 } 1633 else if (SIToFPInst *SCast = dyn_cast<SIToFPInst>(CandidateUI->getUser())) { 1634 IsSigned = true; 1635 DestTy = SCast->getDestTy(); 1636 } 1637 if (!DestTy) continue; 1638 1639 if (TLI) { 1640 // If target does not support DestTy natively then do not apply 1641 // this transformation. 1642 EVT DVT = TLI->getValueType(DestTy); 1643 if (!TLI->isTypeLegal(DVT)) continue; 1644 } 1645 1646 PHINode *PH = dyn_cast<PHINode>(ShadowUse->getOperand(0)); 1647 if (!PH) continue; 1648 if (PH->getNumIncomingValues() != 2) continue; 1649 1650 Type *SrcTy = PH->getType(); 1651 int Mantissa = DestTy->getFPMantissaWidth(); 1652 if (Mantissa == -1) continue; 1653 if ((int)SE.getTypeSizeInBits(SrcTy) > Mantissa) 1654 continue; 1655 1656 unsigned Entry, Latch; 1657 if (PH->getIncomingBlock(0) == L->getLoopPreheader()) { 1658 Entry = 0; 1659 Latch = 1; 1660 } else { 1661 Entry = 1; 1662 Latch = 0; 1663 } 1664 1665 ConstantInt *Init = dyn_cast<ConstantInt>(PH->getIncomingValue(Entry)); 1666 if (!Init) continue; 1667 Constant *NewInit = ConstantFP::get(DestTy, IsSigned ? 1668 (double)Init->getSExtValue() : 1669 (double)Init->getZExtValue()); 1670 1671 BinaryOperator *Incr = 1672 dyn_cast<BinaryOperator>(PH->getIncomingValue(Latch)); 1673 if (!Incr) continue; 1674 if (Incr->getOpcode() != Instruction::Add 1675 && Incr->getOpcode() != Instruction::Sub) 1676 continue; 1677 1678 /* Initialize new IV, double d = 0.0 in above example. */ 1679 ConstantInt *C = NULL; 1680 if (Incr->getOperand(0) == PH) 1681 C = dyn_cast<ConstantInt>(Incr->getOperand(1)); 1682 else if (Incr->getOperand(1) == PH) 1683 C = dyn_cast<ConstantInt>(Incr->getOperand(0)); 1684 else 1685 continue; 1686 1687 if (!C) continue; 1688 1689 // Ignore negative constants, as the code below doesn't handle them 1690 // correctly. TODO: Remove this restriction. 1691 if (!C->getValue().isStrictlyPositive()) continue; 1692 1693 /* Add new PHINode. */ 1694 PHINode *NewPH = PHINode::Create(DestTy, 2, "IV.S.", PH); 1695 1696 /* create new increment. '++d' in above example. */ 1697 Constant *CFP = ConstantFP::get(DestTy, C->getZExtValue()); 1698 BinaryOperator *NewIncr = 1699 BinaryOperator::Create(Incr->getOpcode() == Instruction::Add ? 1700 Instruction::FAdd : Instruction::FSub, 1701 NewPH, CFP, "IV.S.next.", Incr); 1702 1703 NewPH->addIncoming(NewInit, PH->getIncomingBlock(Entry)); 1704 NewPH->addIncoming(NewIncr, PH->getIncomingBlock(Latch)); 1705 1706 /* Remove cast operation */ 1707 ShadowUse->replaceAllUsesWith(NewPH); 1708 ShadowUse->eraseFromParent(); 1709 Changed = true; 1710 break; 1711 } 1712} 1713 1714/// FindIVUserForCond - If Cond has an operand that is an expression of an IV, 1715/// set the IV user and stride information and return true, otherwise return 1716/// false. 1717bool LSRInstance::FindIVUserForCond(ICmpInst *Cond, IVStrideUse *&CondUse) { 1718 for (IVUsers::iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI) 1719 if (UI->getUser() == Cond) { 1720 // NOTE: we could handle setcc instructions with multiple uses here, but 1721 // InstCombine does it as well for simple uses, it's not clear that it 1722 // occurs enough in real life to handle. 1723 CondUse = UI; 1724 return true; 1725 } 1726 return false; 1727} 1728 1729/// OptimizeMax - Rewrite the loop's terminating condition if it uses 1730/// a max computation. 1731/// 1732/// This is a narrow solution to a specific, but acute, problem. For loops 1733/// like this: 1734/// 1735/// i = 0; 1736/// do { 1737/// p[i] = 0.0; 1738/// } while (++i < n); 1739/// 1740/// the trip count isn't just 'n', because 'n' might not be positive. And 1741/// unfortunately this can come up even for loops where the user didn't use 1742/// a C do-while loop. For example, seemingly well-behaved top-test loops 1743/// will commonly be lowered like this: 1744// 1745/// if (n > 0) { 1746/// i = 0; 1747/// do { 1748/// p[i] = 0.0; 1749/// } while (++i < n); 1750/// } 1751/// 1752/// and then it's possible for subsequent optimization to obscure the if 1753/// test in such a way that indvars can't find it. 1754/// 1755/// When indvars can't find the if test in loops like this, it creates a 1756/// max expression, which allows it to give the loop a canonical 1757/// induction variable: 1758/// 1759/// i = 0; 1760/// max = n < 1 ? 1 : n; 1761/// do { 1762/// p[i] = 0.0; 1763/// } while (++i != max); 1764/// 1765/// Canonical induction variables are necessary because the loop passes 1766/// are designed around them. The most obvious example of this is the 1767/// LoopInfo analysis, which doesn't remember trip count values. It 1768/// expects to be able to rediscover the trip count each time it is 1769/// needed, and it does this using a simple analysis that only succeeds if 1770/// the loop has a canonical induction variable. 1771/// 1772/// However, when it comes time to generate code, the maximum operation 1773/// can be quite costly, especially if it's inside of an outer loop. 1774/// 1775/// This function solves this problem by detecting this type of loop and 1776/// rewriting their conditions from ICMP_NE back to ICMP_SLT, and deleting 1777/// the instructions for the maximum computation. 1778/// 1779ICmpInst *LSRInstance::OptimizeMax(ICmpInst *Cond, IVStrideUse* &CondUse) { 1780 // Check that the loop matches the pattern we're looking for. 1781 if (Cond->getPredicate() != CmpInst::ICMP_EQ && 1782 Cond->getPredicate() != CmpInst::ICMP_NE) 1783 return Cond; 1784 1785 SelectInst *Sel = dyn_cast<SelectInst>(Cond->getOperand(1)); 1786 if (!Sel || !Sel->hasOneUse()) return Cond; 1787 1788 const SCEV *BackedgeTakenCount = SE.getBackedgeTakenCount(L); 1789 if (isa<SCEVCouldNotCompute>(BackedgeTakenCount)) 1790 return Cond; 1791 const SCEV *One = SE.getConstant(BackedgeTakenCount->getType(), 1); 1792 1793 // Add one to the backedge-taken count to get the trip count. 1794 const SCEV *IterationCount = SE.getAddExpr(One, BackedgeTakenCount); 1795 if (IterationCount != SE.getSCEV(Sel)) return Cond; 1796 1797 // Check for a max calculation that matches the pattern. There's no check 1798 // for ICMP_ULE here because the comparison would be with zero, which 1799 // isn't interesting. 1800 CmpInst::Predicate Pred = ICmpInst::BAD_ICMP_PREDICATE; 1801 const SCEVNAryExpr *Max = 0; 1802 if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(BackedgeTakenCount)) { 1803 Pred = ICmpInst::ICMP_SLE; 1804 Max = S; 1805 } else if (const SCEVSMaxExpr *S = dyn_cast<SCEVSMaxExpr>(IterationCount)) { 1806 Pred = ICmpInst::ICMP_SLT; 1807 Max = S; 1808 } else if (const SCEVUMaxExpr *U = dyn_cast<SCEVUMaxExpr>(IterationCount)) { 1809 Pred = ICmpInst::ICMP_ULT; 1810 Max = U; 1811 } else { 1812 // No match; bail. 1813 return Cond; 1814 } 1815 1816 // To handle a max with more than two operands, this optimization would 1817 // require additional checking and setup. 1818 if (Max->getNumOperands() != 2) 1819 return Cond; 1820 1821 const SCEV *MaxLHS = Max->getOperand(0); 1822 const SCEV *MaxRHS = Max->getOperand(1); 1823 1824 // ScalarEvolution canonicalizes constants to the left. For < and >, look 1825 // for a comparison with 1. For <= and >=, a comparison with zero. 1826 if (!MaxLHS || 1827 (ICmpInst::isTrueWhenEqual(Pred) ? !MaxLHS->isZero() : (MaxLHS != One))) 1828 return Cond; 1829 1830 // Check the relevant induction variable for conformance to 1831 // the pattern. 1832 const SCEV *IV = SE.getSCEV(Cond->getOperand(0)); 1833 const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(IV); 1834 if (!AR || !AR->isAffine() || 1835 AR->getStart() != One || 1836 AR->getStepRecurrence(SE) != One) 1837 return Cond; 1838 1839 assert(AR->getLoop() == L && 1840 "Loop condition operand is an addrec in a different loop!"); 1841 1842 // Check the right operand of the select, and remember it, as it will 1843 // be used in the new comparison instruction. 1844 Value *NewRHS = 0; 1845 if (ICmpInst::isTrueWhenEqual(Pred)) { 1846 // Look for n+1, and grab n. 1847 if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(1))) 1848 if (isa<ConstantInt>(BO->getOperand(1)) && 1849 cast<ConstantInt>(BO->getOperand(1))->isOne() && 1850 SE.getSCEV(BO->getOperand(0)) == MaxRHS) 1851 NewRHS = BO->getOperand(0); 1852 if (AddOperator *BO = dyn_cast<AddOperator>(Sel->getOperand(2))) 1853 if (isa<ConstantInt>(BO->getOperand(1)) && 1854 cast<ConstantInt>(BO->getOperand(1))->isOne() && 1855 SE.getSCEV(BO->getOperand(0)) == MaxRHS) 1856 NewRHS = BO->getOperand(0); 1857 if (!NewRHS) 1858 return Cond; 1859 } else if (SE.getSCEV(Sel->getOperand(1)) == MaxRHS) 1860 NewRHS = Sel->getOperand(1); 1861 else if (SE.getSCEV(Sel->getOperand(2)) == MaxRHS) 1862 NewRHS = Sel->getOperand(2); 1863 else if (const SCEVUnknown *SU = dyn_cast<SCEVUnknown>(MaxRHS)) 1864 NewRHS = SU->getValue(); 1865 else 1866 // Max doesn't match expected pattern. 1867 return Cond; 1868 1869 // Determine the new comparison opcode. It may be signed or unsigned, 1870 // and the original comparison may be either equality or inequality. 1871 if (Cond->getPredicate() == CmpInst::ICMP_EQ) 1872 Pred = CmpInst::getInversePredicate(Pred); 1873 1874 // Ok, everything looks ok to change the condition into an SLT or SGE and 1875 // delete the max calculation. 1876 ICmpInst *NewCond = 1877 new ICmpInst(Cond, Pred, Cond->getOperand(0), NewRHS, "scmp"); 1878 1879 // Delete the max calculation instructions. 1880 Cond->replaceAllUsesWith(NewCond); 1881 CondUse->setUser(NewCond); 1882 Instruction *Cmp = cast<Instruction>(Sel->getOperand(0)); 1883 Cond->eraseFromParent(); 1884 Sel->eraseFromParent(); 1885 if (Cmp->use_empty()) 1886 Cmp->eraseFromParent(); 1887 return NewCond; 1888} 1889 1890/// OptimizeLoopTermCond - Change loop terminating condition to use the 1891/// postinc iv when possible. 1892void 1893LSRInstance::OptimizeLoopTermCond() { 1894 SmallPtrSet<Instruction *, 4> PostIncs; 1895 1896 BasicBlock *LatchBlock = L->getLoopLatch(); 1897 SmallVector<BasicBlock*, 8> ExitingBlocks; 1898 L->getExitingBlocks(ExitingBlocks); 1899 1900 for (unsigned i = 0, e = ExitingBlocks.size(); i != e; ++i) { 1901 BasicBlock *ExitingBlock = ExitingBlocks[i]; 1902 1903 // Get the terminating condition for the loop if possible. If we 1904 // can, we want to change it to use a post-incremented version of its 1905 // induction variable, to allow coalescing the live ranges for the IV into 1906 // one register value. 1907 1908 BranchInst *TermBr = dyn_cast<BranchInst>(ExitingBlock->getTerminator()); 1909 if (!TermBr) 1910 continue; 1911 // FIXME: Overly conservative, termination condition could be an 'or' etc.. 1912 if (TermBr->isUnconditional() || !isa<ICmpInst>(TermBr->getCondition())) 1913 continue; 1914 1915 // Search IVUsesByStride to find Cond's IVUse if there is one. 1916 IVStrideUse *CondUse = 0; 1917 ICmpInst *Cond = cast<ICmpInst>(TermBr->getCondition()); 1918 if (!FindIVUserForCond(Cond, CondUse)) 1919 continue; 1920 1921 // If the trip count is computed in terms of a max (due to ScalarEvolution 1922 // being unable to find a sufficient guard, for example), change the loop 1923 // comparison to use SLT or ULT instead of NE. 1924 // One consequence of doing this now is that it disrupts the count-down 1925 // optimization. That's not always a bad thing though, because in such 1926 // cases it may still be worthwhile to avoid a max. 1927 Cond = OptimizeMax(Cond, CondUse); 1928 1929 // If this exiting block dominates the latch block, it may also use 1930 // the post-inc value if it won't be shared with other uses. 1931 // Check for dominance. 1932 if (!DT.dominates(ExitingBlock, LatchBlock)) 1933 continue; 1934 1935 // Conservatively avoid trying to use the post-inc value in non-latch 1936 // exits if there may be pre-inc users in intervening blocks. 1937 if (LatchBlock != ExitingBlock) 1938 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI) 1939 // Test if the use is reachable from the exiting block. This dominator 1940 // query is a conservative approximation of reachability. 1941 if (&*UI != CondUse && 1942 !DT.properlyDominates(UI->getUser()->getParent(), ExitingBlock)) { 1943 // Conservatively assume there may be reuse if the quotient of their 1944 // strides could be a legal scale. 1945 const SCEV *A = IU.getStride(*CondUse, L); 1946 const SCEV *B = IU.getStride(*UI, L); 1947 if (!A || !B) continue; 1948 if (SE.getTypeSizeInBits(A->getType()) != 1949 SE.getTypeSizeInBits(B->getType())) { 1950 if (SE.getTypeSizeInBits(A->getType()) > 1951 SE.getTypeSizeInBits(B->getType())) 1952 B = SE.getSignExtendExpr(B, A->getType()); 1953 else 1954 A = SE.getSignExtendExpr(A, B->getType()); 1955 } 1956 if (const SCEVConstant *D = 1957 dyn_cast_or_null<SCEVConstant>(getExactSDiv(B, A, SE))) { 1958 const ConstantInt *C = D->getValue(); 1959 // Stride of one or negative one can have reuse with non-addresses. 1960 if (C->isOne() || C->isAllOnesValue()) 1961 goto decline_post_inc; 1962 // Avoid weird situations. 1963 if (C->getValue().getMinSignedBits() >= 64 || 1964 C->getValue().isMinSignedValue()) 1965 goto decline_post_inc; 1966 // Without TLI, assume that any stride might be valid, and so any 1967 // use might be shared. 1968 if (!TLI) 1969 goto decline_post_inc; 1970 // Check for possible scaled-address reuse. 1971 Type *AccessTy = getAccessType(UI->getUser()); 1972 TargetLowering::AddrMode AM; 1973 AM.Scale = C->getSExtValue(); 1974 if (TLI->isLegalAddressingMode(AM, AccessTy)) 1975 goto decline_post_inc; 1976 AM.Scale = -AM.Scale; 1977 if (TLI->isLegalAddressingMode(AM, AccessTy)) 1978 goto decline_post_inc; 1979 } 1980 } 1981 1982 DEBUG(dbgs() << " Change loop exiting icmp to use postinc iv: " 1983 << *Cond << '\n'); 1984 1985 // It's possible for the setcc instruction to be anywhere in the loop, and 1986 // possible for it to have multiple users. If it is not immediately before 1987 // the exiting block branch, move it. 1988 if (&*++BasicBlock::iterator(Cond) != TermBr) { 1989 if (Cond->hasOneUse()) { 1990 Cond->moveBefore(TermBr); 1991 } else { 1992 // Clone the terminating condition and insert into the loopend. 1993 ICmpInst *OldCond = Cond; 1994 Cond = cast<ICmpInst>(Cond->clone()); 1995 Cond->setName(L->getHeader()->getName() + ".termcond"); 1996 ExitingBlock->getInstList().insert(TermBr, Cond); 1997 1998 // Clone the IVUse, as the old use still exists! 1999 CondUse = &IU.AddUser(Cond, CondUse->getOperandValToReplace()); 2000 TermBr->replaceUsesOfWith(OldCond, Cond); 2001 } 2002 } 2003 2004 // If we get to here, we know that we can transform the setcc instruction to 2005 // use the post-incremented version of the IV, allowing us to coalesce the 2006 // live ranges for the IV correctly. 2007 CondUse->transformToPostInc(L); 2008 Changed = true; 2009 2010 PostIncs.insert(Cond); 2011 decline_post_inc:; 2012 } 2013 2014 // Determine an insertion point for the loop induction variable increment. It 2015 // must dominate all the post-inc comparisons we just set up, and it must 2016 // dominate the loop latch edge. 2017 IVIncInsertPos = L->getLoopLatch()->getTerminator(); 2018 for (SmallPtrSet<Instruction *, 4>::const_iterator I = PostIncs.begin(), 2019 E = PostIncs.end(); I != E; ++I) { 2020 BasicBlock *BB = 2021 DT.findNearestCommonDominator(IVIncInsertPos->getParent(), 2022 (*I)->getParent()); 2023 if (BB == (*I)->getParent()) 2024 IVIncInsertPos = *I; 2025 else if (BB != IVIncInsertPos->getParent()) 2026 IVIncInsertPos = BB->getTerminator(); 2027 } 2028} 2029 2030/// reconcileNewOffset - Determine if the given use can accommodate a fixup 2031/// at the given offset and other details. If so, update the use and 2032/// return true. 2033bool 2034LSRInstance::reconcileNewOffset(LSRUse &LU, int64_t NewOffset, bool HasBaseReg, 2035 LSRUse::KindType Kind, Type *AccessTy) { 2036 int64_t NewMinOffset = LU.MinOffset; 2037 int64_t NewMaxOffset = LU.MaxOffset; 2038 Type *NewAccessTy = AccessTy; 2039 2040 // Check for a mismatched kind. It's tempting to collapse mismatched kinds to 2041 // something conservative, however this can pessimize in the case that one of 2042 // the uses will have all its uses outside the loop, for example. 2043 if (LU.Kind != Kind) 2044 return false; 2045 // Conservatively assume HasBaseReg is true for now. 2046 if (NewOffset < LU.MinOffset) { 2047 if (!isAlwaysFoldable(LU.MaxOffset - NewOffset, 0, HasBaseReg, 2048 Kind, AccessTy, TLI)) 2049 return false; 2050 NewMinOffset = NewOffset; 2051 } else if (NewOffset > LU.MaxOffset) { 2052 if (!isAlwaysFoldable(NewOffset - LU.MinOffset, 0, HasBaseReg, 2053 Kind, AccessTy, TLI)) 2054 return false; 2055 NewMaxOffset = NewOffset; 2056 } 2057 // Check for a mismatched access type, and fall back conservatively as needed. 2058 // TODO: Be less conservative when the type is similar and can use the same 2059 // addressing modes. 2060 if (Kind == LSRUse::Address && AccessTy != LU.AccessTy) 2061 NewAccessTy = Type::getVoidTy(AccessTy->getContext()); 2062 2063 // Update the use. 2064 LU.MinOffset = NewMinOffset; 2065 LU.MaxOffset = NewMaxOffset; 2066 LU.AccessTy = NewAccessTy; 2067 if (NewOffset != LU.Offsets.back()) 2068 LU.Offsets.push_back(NewOffset); 2069 return true; 2070} 2071 2072/// getUse - Return an LSRUse index and an offset value for a fixup which 2073/// needs the given expression, with the given kind and optional access type. 2074/// Either reuse an existing use or create a new one, as needed. 2075std::pair<size_t, int64_t> 2076LSRInstance::getUse(const SCEV *&Expr, 2077 LSRUse::KindType Kind, Type *AccessTy) { 2078 const SCEV *Copy = Expr; 2079 int64_t Offset = ExtractImmediate(Expr, SE); 2080 2081 // Basic uses can't accept any offset, for example. 2082 if (!isAlwaysFoldable(Offset, 0, /*HasBaseReg=*/true, Kind, AccessTy, TLI)) { 2083 Expr = Copy; 2084 Offset = 0; 2085 } 2086 2087 std::pair<UseMapTy::iterator, bool> P = 2088 UseMap.insert(std::make_pair(std::make_pair(Expr, Kind), 0)); 2089 if (!P.second) { 2090 // A use already existed with this base. 2091 size_t LUIdx = P.first->second; 2092 LSRUse &LU = Uses[LUIdx]; 2093 if (reconcileNewOffset(LU, Offset, /*HasBaseReg=*/true, Kind, AccessTy)) 2094 // Reuse this use. 2095 return std::make_pair(LUIdx, Offset); 2096 } 2097 2098 // Create a new use. 2099 size_t LUIdx = Uses.size(); 2100 P.first->second = LUIdx; 2101 Uses.push_back(LSRUse(Kind, AccessTy)); 2102 LSRUse &LU = Uses[LUIdx]; 2103 2104 // We don't need to track redundant offsets, but we don't need to go out 2105 // of our way here to avoid them. 2106 if (LU.Offsets.empty() || Offset != LU.Offsets.back()) 2107 LU.Offsets.push_back(Offset); 2108 2109 LU.MinOffset = Offset; 2110 LU.MaxOffset = Offset; 2111 return std::make_pair(LUIdx, Offset); 2112} 2113 2114/// DeleteUse - Delete the given use from the Uses list. 2115void LSRInstance::DeleteUse(LSRUse &LU, size_t LUIdx) { 2116 if (&LU != &Uses.back()) 2117 std::swap(LU, Uses.back()); 2118 Uses.pop_back(); 2119 2120 // Update RegUses. 2121 RegUses.SwapAndDropUse(LUIdx, Uses.size()); 2122} 2123 2124/// FindUseWithFormula - Look for a use distinct from OrigLU which is has 2125/// a formula that has the same registers as the given formula. 2126LSRUse * 2127LSRInstance::FindUseWithSimilarFormula(const Formula &OrigF, 2128 const LSRUse &OrigLU) { 2129 // Search all uses for the formula. This could be more clever. 2130 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) { 2131 LSRUse &LU = Uses[LUIdx]; 2132 // Check whether this use is close enough to OrigLU, to see whether it's 2133 // worthwhile looking through its formulae. 2134 // Ignore ICmpZero uses because they may contain formulae generated by 2135 // GenerateICmpZeroScales, in which case adding fixup offsets may 2136 // be invalid. 2137 if (&LU != &OrigLU && 2138 LU.Kind != LSRUse::ICmpZero && 2139 LU.Kind == OrigLU.Kind && OrigLU.AccessTy == LU.AccessTy && 2140 LU.WidestFixupType == OrigLU.WidestFixupType && 2141 LU.HasFormulaWithSameRegs(OrigF)) { 2142 // Scan through this use's formulae. 2143 for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(), 2144 E = LU.Formulae.end(); I != E; ++I) { 2145 const Formula &F = *I; 2146 // Check to see if this formula has the same registers and symbols 2147 // as OrigF. 2148 if (F.BaseRegs == OrigF.BaseRegs && 2149 F.ScaledReg == OrigF.ScaledReg && 2150 F.AM.BaseGV == OrigF.AM.BaseGV && 2151 F.AM.Scale == OrigF.AM.Scale && 2152 F.UnfoldedOffset == OrigF.UnfoldedOffset) { 2153 if (F.AM.BaseOffs == 0) 2154 return &LU; 2155 // This is the formula where all the registers and symbols matched; 2156 // there aren't going to be any others. Since we declined it, we 2157 // can skip the rest of the formulae and procede to the next LSRUse. 2158 break; 2159 } 2160 } 2161 } 2162 } 2163 2164 // Nothing looked good. 2165 return 0; 2166} 2167 2168void LSRInstance::CollectInterestingTypesAndFactors() { 2169 SmallSetVector<const SCEV *, 4> Strides; 2170 2171 // Collect interesting types and strides. 2172 SmallVector<const SCEV *, 4> Worklist; 2173 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI) { 2174 const SCEV *Expr = IU.getExpr(*UI); 2175 2176 // Collect interesting types. 2177 Types.insert(SE.getEffectiveSCEVType(Expr->getType())); 2178 2179 // Add strides for mentioned loops. 2180 Worklist.push_back(Expr); 2181 do { 2182 const SCEV *S = Worklist.pop_back_val(); 2183 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) { 2184 if (AR->getLoop() == L) 2185 Strides.insert(AR->getStepRecurrence(SE)); 2186 Worklist.push_back(AR->getStart()); 2187 } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) { 2188 Worklist.append(Add->op_begin(), Add->op_end()); 2189 } 2190 } while (!Worklist.empty()); 2191 } 2192 2193 // Compute interesting factors from the set of interesting strides. 2194 for (SmallSetVector<const SCEV *, 4>::const_iterator 2195 I = Strides.begin(), E = Strides.end(); I != E; ++I) 2196 for (SmallSetVector<const SCEV *, 4>::const_iterator NewStrideIter = 2197 llvm::next(I); NewStrideIter != E; ++NewStrideIter) { 2198 const SCEV *OldStride = *I; 2199 const SCEV *NewStride = *NewStrideIter; 2200 2201 if (SE.getTypeSizeInBits(OldStride->getType()) != 2202 SE.getTypeSizeInBits(NewStride->getType())) { 2203 if (SE.getTypeSizeInBits(OldStride->getType()) > 2204 SE.getTypeSizeInBits(NewStride->getType())) 2205 NewStride = SE.getSignExtendExpr(NewStride, OldStride->getType()); 2206 else 2207 OldStride = SE.getSignExtendExpr(OldStride, NewStride->getType()); 2208 } 2209 if (const SCEVConstant *Factor = 2210 dyn_cast_or_null<SCEVConstant>(getExactSDiv(NewStride, OldStride, 2211 SE, true))) { 2212 if (Factor->getValue()->getValue().getMinSignedBits() <= 64) 2213 Factors.insert(Factor->getValue()->getValue().getSExtValue()); 2214 } else if (const SCEVConstant *Factor = 2215 dyn_cast_or_null<SCEVConstant>(getExactSDiv(OldStride, 2216 NewStride, 2217 SE, true))) { 2218 if (Factor->getValue()->getValue().getMinSignedBits() <= 64) 2219 Factors.insert(Factor->getValue()->getValue().getSExtValue()); 2220 } 2221 } 2222 2223 // If all uses use the same type, don't bother looking for truncation-based 2224 // reuse. 2225 if (Types.size() == 1) 2226 Types.clear(); 2227 2228 DEBUG(print_factors_and_types(dbgs())); 2229} 2230 2231/// findIVOperand - Helper for CollectChains that finds an IV operand (computed 2232/// by an AddRec in this loop) within [OI,OE) or returns OE. If IVUsers mapped 2233/// Instructions to IVStrideUses, we could partially skip this. 2234static User::op_iterator 2235findIVOperand(User::op_iterator OI, User::op_iterator OE, 2236 Loop *L, ScalarEvolution &SE) { 2237 for(; OI != OE; ++OI) { 2238 if (Instruction *Oper = dyn_cast<Instruction>(*OI)) { 2239 if (!SE.isSCEVable(Oper->getType())) 2240 continue; 2241 2242 if (const SCEVAddRecExpr *AR = 2243 dyn_cast<SCEVAddRecExpr>(SE.getSCEV(Oper))) { 2244 if (AR->getLoop() == L) 2245 break; 2246 } 2247 } 2248 } 2249 return OI; 2250} 2251 2252/// getWideOperand - IVChain logic must consistenctly peek base TruncInst 2253/// operands, so wrap it in a convenient helper. 2254static Value *getWideOperand(Value *Oper) { 2255 if (TruncInst *Trunc = dyn_cast<TruncInst>(Oper)) 2256 return Trunc->getOperand(0); 2257 return Oper; 2258} 2259 2260/// isCompatibleIVType - Return true if we allow an IV chain to include both 2261/// types. 2262static bool isCompatibleIVType(Value *LVal, Value *RVal) { 2263 Type *LType = LVal->getType(); 2264 Type *RType = RVal->getType(); 2265 return (LType == RType) || (LType->isPointerTy() && RType->isPointerTy()); 2266} 2267 2268/// getExprBase - Return an approximation of this SCEV expression's "base", or 2269/// NULL for any constant. Returning the expression itself is 2270/// conservative. Returning a deeper subexpression is more precise and valid as 2271/// long as it isn't less complex than another subexpression. For expressions 2272/// involving multiple unscaled values, we need to return the pointer-type 2273/// SCEVUnknown. This avoids forming chains across objects, such as: 2274/// PrevOper==a[i], IVOper==b[i], IVInc==b-a. 2275/// 2276/// Since SCEVUnknown is the rightmost type, and pointers are the rightmost 2277/// SCEVUnknown, we simply return the rightmost SCEV operand. 2278static const SCEV *getExprBase(const SCEV *S) { 2279 switch (S->getSCEVType()) { 2280 default: // uncluding scUnknown. 2281 return S; 2282 case scConstant: 2283 return 0; 2284 case scTruncate: 2285 return getExprBase(cast<SCEVTruncateExpr>(S)->getOperand()); 2286 case scZeroExtend: 2287 return getExprBase(cast<SCEVZeroExtendExpr>(S)->getOperand()); 2288 case scSignExtend: 2289 return getExprBase(cast<SCEVSignExtendExpr>(S)->getOperand()); 2290 case scAddExpr: { 2291 // Skip over scaled operands (scMulExpr) to follow add operands as long as 2292 // there's nothing more complex. 2293 // FIXME: not sure if we want to recognize negation. 2294 const SCEVAddExpr *Add = cast<SCEVAddExpr>(S); 2295 for (std::reverse_iterator<SCEVAddExpr::op_iterator> I(Add->op_end()), 2296 E(Add->op_begin()); I != E; ++I) { 2297 const SCEV *SubExpr = *I; 2298 if (SubExpr->getSCEVType() == scAddExpr) 2299 return getExprBase(SubExpr); 2300 2301 if (SubExpr->getSCEVType() != scMulExpr) 2302 return SubExpr; 2303 } 2304 return S; // all operands are scaled, be conservative. 2305 } 2306 case scAddRecExpr: 2307 return getExprBase(cast<SCEVAddRecExpr>(S)->getStart()); 2308 } 2309} 2310 2311/// Return true if the chain increment is profitable to expand into a loop 2312/// invariant value, which may require its own register. A profitable chain 2313/// increment will be an offset relative to the same base. We allow such offsets 2314/// to potentially be used as chain increment as long as it's not obviously 2315/// expensive to expand using real instructions. 2316static const SCEV * 2317getProfitableChainIncrement(Value *NextIV, Value *PrevIV, 2318 const IVChain &Chain, Loop *L, 2319 ScalarEvolution &SE, const TargetLowering *TLI) { 2320 // Prune the solution space aggressively by checking that both IV operands 2321 // are expressions that operate on the same unscaled SCEVUnknown. This 2322 // "base" will be canceled by the subsequent getMinusSCEV call. Checking first 2323 // avoids creating extra SCEV expressions. 2324 const SCEV *OperExpr = SE.getSCEV(NextIV); 2325 const SCEV *PrevExpr = SE.getSCEV(PrevIV); 2326 if (getExprBase(OperExpr) != getExprBase(PrevExpr) && !StressIVChain) 2327 return 0; 2328 2329 const SCEV *IncExpr = SE.getMinusSCEV(OperExpr, PrevExpr); 2330 if (!SE.isLoopInvariant(IncExpr, L)) 2331 return 0; 2332 2333 // We are not able to expand an increment unless it is loop invariant, 2334 // however, the following checks are purely for profitability. 2335 if (StressIVChain) 2336 return IncExpr; 2337 2338 // Do not replace a constant offset from IV head with a nonconstant IV 2339 // increment. 2340 if (!isa<SCEVConstant>(IncExpr)) { 2341 const SCEV *HeadExpr = SE.getSCEV(getWideOperand(Chain[0].IVOperand)); 2342 if (isa<SCEVConstant>(SE.getMinusSCEV(OperExpr, HeadExpr))) 2343 return 0; 2344 } 2345 2346 SmallPtrSet<const SCEV*, 8> Processed; 2347 if (isHighCostExpansion(IncExpr, Processed, SE)) 2348 return 0; 2349 2350 return IncExpr; 2351} 2352 2353/// Return true if the number of registers needed for the chain is estimated to 2354/// be less than the number required for the individual IV users. First prohibit 2355/// any IV users that keep the IV live across increments (the Users set should 2356/// be empty). Next count the number and type of increments in the chain. 2357/// 2358/// Chaining IVs can lead to considerable code bloat if ISEL doesn't 2359/// effectively use postinc addressing modes. Only consider it profitable it the 2360/// increments can be computed in fewer registers when chained. 2361/// 2362/// TODO: Consider IVInc free if it's already used in another chains. 2363static bool 2364isProfitableChain(IVChain &Chain, SmallPtrSet<Instruction*, 4> &Users, 2365 ScalarEvolution &SE, const TargetLowering *TLI) { 2366 if (StressIVChain) 2367 return true; 2368 2369 if (Chain.size() <= 2) 2370 return false; 2371 2372 if (!Users.empty()) { 2373 DEBUG(dbgs() << "Chain: " << *Chain[0].UserInst << " users:\n"; 2374 for (SmallPtrSet<Instruction*, 4>::const_iterator I = Users.begin(), 2375 E = Users.end(); I != E; ++I) { 2376 dbgs() << " " << **I << "\n"; 2377 }); 2378 return false; 2379 } 2380 assert(!Chain.empty() && "empty IV chains are not allowed"); 2381 2382 // The chain itself may require a register, so intialize cost to 1. 2383 int cost = 1; 2384 2385 // A complete chain likely eliminates the need for keeping the original IV in 2386 // a register. LSR does not currently know how to form a complete chain unless 2387 // the header phi already exists. 2388 if (isa<PHINode>(Chain.back().UserInst) 2389 && SE.getSCEV(Chain.back().UserInst) == Chain[0].IncExpr) { 2390 --cost; 2391 } 2392 const SCEV *LastIncExpr = 0; 2393 unsigned NumConstIncrements = 0; 2394 unsigned NumVarIncrements = 0; 2395 unsigned NumReusedIncrements = 0; 2396 for (IVChain::const_iterator I = llvm::next(Chain.begin()), E = Chain.end(); 2397 I != E; ++I) { 2398 2399 if (I->IncExpr->isZero()) 2400 continue; 2401 2402 // Incrementing by zero or some constant is neutral. We assume constants can 2403 // be folded into an addressing mode or an add's immediate operand. 2404 if (isa<SCEVConstant>(I->IncExpr)) { 2405 ++NumConstIncrements; 2406 continue; 2407 } 2408 2409 if (I->IncExpr == LastIncExpr) 2410 ++NumReusedIncrements; 2411 else 2412 ++NumVarIncrements; 2413 2414 LastIncExpr = I->IncExpr; 2415 } 2416 // An IV chain with a single increment is handled by LSR's postinc 2417 // uses. However, a chain with multiple increments requires keeping the IV's 2418 // value live longer than it needs to be if chained. 2419 if (NumConstIncrements > 1) 2420 --cost; 2421 2422 // Materializing increment expressions in the preheader that didn't exist in 2423 // the original code may cost a register. For example, sign-extended array 2424 // indices can produce ridiculous increments like this: 2425 // IV + ((sext i32 (2 * %s) to i64) + (-1 * (sext i32 %s to i64))) 2426 cost += NumVarIncrements; 2427 2428 // Reusing variable increments likely saves a register to hold the multiple of 2429 // the stride. 2430 cost -= NumReusedIncrements; 2431 2432 DEBUG(dbgs() << "Chain: " << *Chain[0].UserInst << " Cost: " << cost << "\n"); 2433 2434 return cost < 0; 2435} 2436 2437/// ChainInstruction - Add this IV user to an existing chain or make it the head 2438/// of a new chain. 2439void LSRInstance::ChainInstruction(Instruction *UserInst, Instruction *IVOper, 2440 SmallVectorImpl<ChainUsers> &ChainUsersVec) { 2441 // When IVs are used as types of varying widths, they are generally converted 2442 // to a wider type with some uses remaining narrow under a (free) trunc. 2443 Value *NextIV = getWideOperand(IVOper); 2444 2445 // Visit all existing chains. Check if its IVOper can be computed as a 2446 // profitable loop invariant increment from the last link in the Chain. 2447 unsigned ChainIdx = 0, NChains = IVChainVec.size(); 2448 const SCEV *LastIncExpr = 0; 2449 for (; ChainIdx < NChains; ++ChainIdx) { 2450 Value *PrevIV = getWideOperand(IVChainVec[ChainIdx].back().IVOperand); 2451 if (!isCompatibleIVType(PrevIV, NextIV)) 2452 continue; 2453 2454 // A phi node terminates a chain. 2455 if (isa<PHINode>(UserInst) 2456 && isa<PHINode>(IVChainVec[ChainIdx].back().UserInst)) 2457 continue; 2458 2459 if (const SCEV *IncExpr = 2460 getProfitableChainIncrement(NextIV, PrevIV, IVChainVec[ChainIdx], 2461 L, SE, TLI)) { 2462 LastIncExpr = IncExpr; 2463 break; 2464 } 2465 } 2466 // If we haven't found a chain, create a new one, unless we hit the max. Don't 2467 // bother for phi nodes, because they must be last in the chain. 2468 if (ChainIdx == NChains) { 2469 if (isa<PHINode>(UserInst)) 2470 return; 2471 if (NChains >= MaxChains && !StressIVChain) { 2472 DEBUG(dbgs() << "IV Chain Limit\n"); 2473 return; 2474 } 2475 LastIncExpr = SE.getSCEV(NextIV); 2476 // IVUsers may have skipped over sign/zero extensions. We don't currently 2477 // attempt to form chains involving extensions unless they can be hoisted 2478 // into this loop's AddRec. 2479 if (!isa<SCEVAddRecExpr>(LastIncExpr)) 2480 return; 2481 ++NChains; 2482 IVChainVec.resize(NChains); 2483 ChainUsersVec.resize(NChains); 2484 DEBUG(dbgs() << "IV Head: (" << *UserInst << ") IV=" << *LastIncExpr 2485 << "\n"); 2486 } 2487 else 2488 DEBUG(dbgs() << "IV Inc: (" << *UserInst << ") IV+" << *LastIncExpr 2489 << "\n"); 2490 2491 // Add this IV user to the end of the chain. 2492 IVChainVec[ChainIdx].push_back(IVInc(UserInst, IVOper, LastIncExpr)); 2493 2494 SmallPtrSet<Instruction*,4> &NearUsers = ChainUsersVec[ChainIdx].NearUsers; 2495 // This chain's NearUsers become FarUsers. 2496 if (!LastIncExpr->isZero()) { 2497 ChainUsersVec[ChainIdx].FarUsers.insert(NearUsers.begin(), 2498 NearUsers.end()); 2499 NearUsers.clear(); 2500 } 2501 2502 // All other uses of IVOperand become near uses of the chain. 2503 // We currently ignore intermediate values within SCEV expressions, assuming 2504 // they will eventually be used be the current chain, or can be computed 2505 // from one of the chain increments. To be more precise we could 2506 // transitively follow its user and only add leaf IV users to the set. 2507 for (Value::use_iterator UseIter = IVOper->use_begin(), 2508 UseEnd = IVOper->use_end(); UseIter != UseEnd; ++UseIter) { 2509 Instruction *OtherUse = dyn_cast<Instruction>(*UseIter); 2510 if (!OtherUse || OtherUse == UserInst) 2511 continue; 2512 if (SE.isSCEVable(OtherUse->getType()) 2513 && !isa<SCEVUnknown>(SE.getSCEV(OtherUse)) 2514 && IU.isIVUserOrOperand(OtherUse)) { 2515 continue; 2516 } 2517 NearUsers.insert(OtherUse); 2518 } 2519 2520 // Since this user is part of the chain, it's no longer considered a use 2521 // of the chain. 2522 ChainUsersVec[ChainIdx].FarUsers.erase(UserInst); 2523} 2524 2525/// CollectChains - Populate the vector of Chains. 2526/// 2527/// This decreases ILP at the architecture level. Targets with ample registers, 2528/// multiple memory ports, and no register renaming probably don't want 2529/// this. However, such targets should probably disable LSR altogether. 2530/// 2531/// The job of LSR is to make a reasonable choice of induction variables across 2532/// the loop. Subsequent passes can easily "unchain" computation exposing more 2533/// ILP *within the loop* if the target wants it. 2534/// 2535/// Finding the best IV chain is potentially a scheduling problem. Since LSR 2536/// will not reorder memory operations, it will recognize this as a chain, but 2537/// will generate redundant IV increments. Ideally this would be corrected later 2538/// by a smart scheduler: 2539/// = A[i] 2540/// = A[i+x] 2541/// A[i] = 2542/// A[i+x] = 2543/// 2544/// TODO: Walk the entire domtree within this loop, not just the path to the 2545/// loop latch. This will discover chains on side paths, but requires 2546/// maintaining multiple copies of the Chains state. 2547void LSRInstance::CollectChains() { 2548 SmallVector<ChainUsers, 8> ChainUsersVec; 2549 2550 SmallVector<BasicBlock *,8> LatchPath; 2551 BasicBlock *LoopHeader = L->getHeader(); 2552 for (DomTreeNode *Rung = DT.getNode(L->getLoopLatch()); 2553 Rung->getBlock() != LoopHeader; Rung = Rung->getIDom()) { 2554 LatchPath.push_back(Rung->getBlock()); 2555 } 2556 LatchPath.push_back(LoopHeader); 2557 2558 // Walk the instruction stream from the loop header to the loop latch. 2559 for (SmallVectorImpl<BasicBlock *>::reverse_iterator 2560 BBIter = LatchPath.rbegin(), BBEnd = LatchPath.rend(); 2561 BBIter != BBEnd; ++BBIter) { 2562 for (BasicBlock::iterator I = (*BBIter)->begin(), E = (*BBIter)->end(); 2563 I != E; ++I) { 2564 // Skip instructions that weren't seen by IVUsers analysis. 2565 if (isa<PHINode>(I) || !IU.isIVUserOrOperand(I)) 2566 continue; 2567 2568 // Ignore users that are part of a SCEV expression. This way we only 2569 // consider leaf IV Users. This effectively rediscovers a portion of 2570 // IVUsers analysis but in program order this time. 2571 if (SE.isSCEVable(I->getType()) && !isa<SCEVUnknown>(SE.getSCEV(I))) 2572 continue; 2573 2574 // Remove this instruction from any NearUsers set it may be in. 2575 for (unsigned ChainIdx = 0, NChains = IVChainVec.size(); 2576 ChainIdx < NChains; ++ChainIdx) { 2577 ChainUsersVec[ChainIdx].NearUsers.erase(I); 2578 } 2579 // Search for operands that can be chained. 2580 SmallPtrSet<Instruction*, 4> UniqueOperands; 2581 User::op_iterator IVOpEnd = I->op_end(); 2582 User::op_iterator IVOpIter = findIVOperand(I->op_begin(), IVOpEnd, L, SE); 2583 while (IVOpIter != IVOpEnd) { 2584 Instruction *IVOpInst = cast<Instruction>(*IVOpIter); 2585 if (UniqueOperands.insert(IVOpInst)) 2586 ChainInstruction(I, IVOpInst, ChainUsersVec); 2587 IVOpIter = findIVOperand(llvm::next(IVOpIter), IVOpEnd, L, SE); 2588 } 2589 } // Continue walking down the instructions. 2590 } // Continue walking down the domtree. 2591 // Visit phi backedges to determine if the chain can generate the IV postinc. 2592 for (BasicBlock::iterator I = L->getHeader()->begin(); 2593 PHINode *PN = dyn_cast<PHINode>(I); ++I) { 2594 if (!SE.isSCEVable(PN->getType())) 2595 continue; 2596 2597 Instruction *IncV = 2598 dyn_cast<Instruction>(PN->getIncomingValueForBlock(L->getLoopLatch())); 2599 if (IncV) 2600 ChainInstruction(PN, IncV, ChainUsersVec); 2601 } 2602 // Remove any unprofitable chains. 2603 unsigned ChainIdx = 0; 2604 for (unsigned UsersIdx = 0, NChains = IVChainVec.size(); 2605 UsersIdx < NChains; ++UsersIdx) { 2606 if (!isProfitableChain(IVChainVec[UsersIdx], 2607 ChainUsersVec[UsersIdx].FarUsers, SE, TLI)) 2608 continue; 2609 // Preserve the chain at UsesIdx. 2610 if (ChainIdx != UsersIdx) 2611 IVChainVec[ChainIdx] = IVChainVec[UsersIdx]; 2612 FinalizeChain(IVChainVec[ChainIdx]); 2613 ++ChainIdx; 2614 } 2615 IVChainVec.resize(ChainIdx); 2616} 2617 2618void LSRInstance::FinalizeChain(IVChain &Chain) { 2619 assert(!Chain.empty() && "empty IV chains are not allowed"); 2620 DEBUG(dbgs() << "Final Chain: " << *Chain[0].UserInst << "\n"); 2621 2622 for (IVChain::const_iterator I = llvm::next(Chain.begin()), E = Chain.end(); 2623 I != E; ++I) { 2624 DEBUG(dbgs() << " Inc: " << *I->UserInst << "\n"); 2625 User::op_iterator UseI = 2626 std::find(I->UserInst->op_begin(), I->UserInst->op_end(), I->IVOperand); 2627 assert(UseI != I->UserInst->op_end() && "cannot find IV operand"); 2628 IVIncSet.insert(UseI); 2629 } 2630} 2631 2632/// Return true if the IVInc can be folded into an addressing mode. 2633static bool canFoldIVIncExpr(const SCEV *IncExpr, Instruction *UserInst, 2634 Value *Operand, const TargetLowering *TLI) { 2635 const SCEVConstant *IncConst = dyn_cast<SCEVConstant>(IncExpr); 2636 if (!IncConst || !isAddressUse(UserInst, Operand)) 2637 return false; 2638 2639 if (IncConst->getValue()->getValue().getMinSignedBits() > 64) 2640 return false; 2641 2642 int64_t IncOffset = IncConst->getValue()->getSExtValue(); 2643 if (!isAlwaysFoldable(IncOffset, /*BaseGV=*/0, /*HaseBaseReg=*/false, 2644 LSRUse::Address, getAccessType(UserInst), TLI)) 2645 return false; 2646 2647 return true; 2648} 2649 2650/// GenerateIVChains - Generate an add or subtract for each IVInc in a chain to 2651/// materialize the IV user's operand from the previous IV user's operand. 2652void LSRInstance::GenerateIVChain(const IVChain &Chain, SCEVExpander &Rewriter, 2653 SmallVectorImpl<WeakVH> &DeadInsts) { 2654 // Find the new IVOperand for the head of the chain. It may have been replaced 2655 // by LSR. 2656 const IVInc &Head = Chain[0]; 2657 User::op_iterator IVOpEnd = Head.UserInst->op_end(); 2658 User::op_iterator IVOpIter = findIVOperand(Head.UserInst->op_begin(), 2659 IVOpEnd, L, SE); 2660 Value *IVSrc = 0; 2661 while (IVOpIter != IVOpEnd) { 2662 IVSrc = getWideOperand(*IVOpIter); 2663 2664 // If this operand computes the expression that the chain needs, we may use 2665 // it. (Check this after setting IVSrc which is used below.) 2666 // 2667 // Note that if Head.IncExpr is wider than IVSrc, then this phi is too 2668 // narrow for the chain, so we can no longer use it. We do allow using a 2669 // wider phi, assuming the LSR checked for free truncation. In that case we 2670 // should already have a truncate on this operand such that 2671 // getSCEV(IVSrc) == IncExpr. 2672 if (SE.getSCEV(*IVOpIter) == Head.IncExpr 2673 || SE.getSCEV(IVSrc) == Head.IncExpr) { 2674 break; 2675 } 2676 IVOpIter = findIVOperand(llvm::next(IVOpIter), IVOpEnd, L, SE); 2677 } 2678 if (IVOpIter == IVOpEnd) { 2679 // Gracefully give up on this chain. 2680 DEBUG(dbgs() << "Concealed chain head: " << *Head.UserInst << "\n"); 2681 return; 2682 } 2683 2684 DEBUG(dbgs() << "Generate chain at: " << *IVSrc << "\n"); 2685 Type *IVTy = IVSrc->getType(); 2686 Type *IntTy = SE.getEffectiveSCEVType(IVTy); 2687 const SCEV *LeftOverExpr = 0; 2688 for (IVChain::const_iterator IncI = llvm::next(Chain.begin()), 2689 IncE = Chain.end(); IncI != IncE; ++IncI) { 2690 2691 Instruction *InsertPt = IncI->UserInst; 2692 if (isa<PHINode>(InsertPt)) 2693 InsertPt = L->getLoopLatch()->getTerminator(); 2694 2695 // IVOper will replace the current IV User's operand. IVSrc is the IV 2696 // value currently held in a register. 2697 Value *IVOper = IVSrc; 2698 if (!IncI->IncExpr->isZero()) { 2699 // IncExpr was the result of subtraction of two narrow values, so must 2700 // be signed. 2701 const SCEV *IncExpr = SE.getNoopOrSignExtend(IncI->IncExpr, IntTy); 2702 LeftOverExpr = LeftOverExpr ? 2703 SE.getAddExpr(LeftOverExpr, IncExpr) : IncExpr; 2704 } 2705 if (LeftOverExpr && !LeftOverExpr->isZero()) { 2706 // Expand the IV increment. 2707 Rewriter.clearPostInc(); 2708 Value *IncV = Rewriter.expandCodeFor(LeftOverExpr, IntTy, InsertPt); 2709 const SCEV *IVOperExpr = SE.getAddExpr(SE.getUnknown(IVSrc), 2710 SE.getUnknown(IncV)); 2711 IVOper = Rewriter.expandCodeFor(IVOperExpr, IVTy, InsertPt); 2712 2713 // If an IV increment can't be folded, use it as the next IV value. 2714 if (!canFoldIVIncExpr(LeftOverExpr, IncI->UserInst, IncI->IVOperand, 2715 TLI)) { 2716 assert(IVTy == IVOper->getType() && "inconsistent IV increment type"); 2717 IVSrc = IVOper; 2718 LeftOverExpr = 0; 2719 } 2720 } 2721 Type *OperTy = IncI->IVOperand->getType(); 2722 if (IVTy != OperTy) { 2723 assert(SE.getTypeSizeInBits(IVTy) >= SE.getTypeSizeInBits(OperTy) && 2724 "cannot extend a chained IV"); 2725 IRBuilder<> Builder(InsertPt); 2726 IVOper = Builder.CreateTruncOrBitCast(IVOper, OperTy, "lsr.chain"); 2727 } 2728 IncI->UserInst->replaceUsesOfWith(IncI->IVOperand, IVOper); 2729 DeadInsts.push_back(IncI->IVOperand); 2730 } 2731 // If LSR created a new, wider phi, we may also replace its postinc. We only 2732 // do this if we also found a wide value for the head of the chain. 2733 if (isa<PHINode>(Chain.back().UserInst)) { 2734 for (BasicBlock::iterator I = L->getHeader()->begin(); 2735 PHINode *Phi = dyn_cast<PHINode>(I); ++I) { 2736 if (!isCompatibleIVType(Phi, IVSrc)) 2737 continue; 2738 Instruction *PostIncV = dyn_cast<Instruction>( 2739 Phi->getIncomingValueForBlock(L->getLoopLatch())); 2740 if (!PostIncV || (SE.getSCEV(PostIncV) != SE.getSCEV(IVSrc))) 2741 continue; 2742 Value *IVOper = IVSrc; 2743 Type *PostIncTy = PostIncV->getType(); 2744 if (IVTy != PostIncTy) { 2745 assert(PostIncTy->isPointerTy() && "mixing int/ptr IV types"); 2746 IRBuilder<> Builder(L->getLoopLatch()->getTerminator()); 2747 Builder.SetCurrentDebugLocation(PostIncV->getDebugLoc()); 2748 IVOper = Builder.CreatePointerCast(IVSrc, PostIncTy, "lsr.chain"); 2749 } 2750 Phi->replaceUsesOfWith(PostIncV, IVOper); 2751 DeadInsts.push_back(PostIncV); 2752 } 2753 } 2754} 2755 2756void LSRInstance::CollectFixupsAndInitialFormulae() { 2757 for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI) { 2758 Instruction *UserInst = UI->getUser(); 2759 // Skip IV users that are part of profitable IV Chains. 2760 User::op_iterator UseI = std::find(UserInst->op_begin(), UserInst->op_end(), 2761 UI->getOperandValToReplace()); 2762 assert(UseI != UserInst->op_end() && "cannot find IV operand"); 2763 if (IVIncSet.count(UseI)) 2764 continue; 2765 2766 // Record the uses. 2767 LSRFixup &LF = getNewFixup(); 2768 LF.UserInst = UserInst; 2769 LF.OperandValToReplace = UI->getOperandValToReplace(); 2770 LF.PostIncLoops = UI->getPostIncLoops(); 2771 2772 LSRUse::KindType Kind = LSRUse::Basic; 2773 Type *AccessTy = 0; 2774 if (isAddressUse(LF.UserInst, LF.OperandValToReplace)) { 2775 Kind = LSRUse::Address; 2776 AccessTy = getAccessType(LF.UserInst); 2777 } 2778 2779 const SCEV *S = IU.getExpr(*UI); 2780 2781 // Equality (== and !=) ICmps are special. We can rewrite (i == N) as 2782 // (N - i == 0), and this allows (N - i) to be the expression that we work 2783 // with rather than just N or i, so we can consider the register 2784 // requirements for both N and i at the same time. Limiting this code to 2785 // equality icmps is not a problem because all interesting loops use 2786 // equality icmps, thanks to IndVarSimplify. 2787 if (ICmpInst *CI = dyn_cast<ICmpInst>(LF.UserInst)) 2788 if (CI->isEquality()) { 2789 // Swap the operands if needed to put the OperandValToReplace on the 2790 // left, for consistency. 2791 Value *NV = CI->getOperand(1); 2792 if (NV == LF.OperandValToReplace) { 2793 CI->setOperand(1, CI->getOperand(0)); 2794 CI->setOperand(0, NV); 2795 NV = CI->getOperand(1); 2796 Changed = true; 2797 } 2798 2799 // x == y --> x - y == 0 2800 const SCEV *N = SE.getSCEV(NV); 2801 if (SE.isLoopInvariant(N, L)) { 2802 // S is normalized, so normalize N before folding it into S 2803 // to keep the result normalized. 2804 N = TransformForPostIncUse(Normalize, N, CI, 0, 2805 LF.PostIncLoops, SE, DT); 2806 Kind = LSRUse::ICmpZero; 2807 S = SE.getMinusSCEV(N, S); 2808 } 2809 2810 // -1 and the negations of all interesting strides (except the negation 2811 // of -1) are now also interesting. 2812 for (size_t i = 0, e = Factors.size(); i != e; ++i) 2813 if (Factors[i] != -1) 2814 Factors.insert(-(uint64_t)Factors[i]); 2815 Factors.insert(-1); 2816 } 2817 2818 // Set up the initial formula for this use. 2819 std::pair<size_t, int64_t> P = getUse(S, Kind, AccessTy); 2820 LF.LUIdx = P.first; 2821 LF.Offset = P.second; 2822 LSRUse &LU = Uses[LF.LUIdx]; 2823 LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L); 2824 if (!LU.WidestFixupType || 2825 SE.getTypeSizeInBits(LU.WidestFixupType) < 2826 SE.getTypeSizeInBits(LF.OperandValToReplace->getType())) 2827 LU.WidestFixupType = LF.OperandValToReplace->getType(); 2828 2829 // If this is the first use of this LSRUse, give it a formula. 2830 if (LU.Formulae.empty()) { 2831 InsertInitialFormula(S, LU, LF.LUIdx); 2832 CountRegisters(LU.Formulae.back(), LF.LUIdx); 2833 } 2834 } 2835 2836 DEBUG(print_fixups(dbgs())); 2837} 2838 2839/// InsertInitialFormula - Insert a formula for the given expression into 2840/// the given use, separating out loop-variant portions from loop-invariant 2841/// and loop-computable portions. 2842void 2843LSRInstance::InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx) { 2844 Formula F; 2845 F.InitialMatch(S, L, SE); 2846 bool Inserted = InsertFormula(LU, LUIdx, F); 2847 assert(Inserted && "Initial formula already exists!"); (void)Inserted; 2848} 2849 2850/// InsertSupplementalFormula - Insert a simple single-register formula for 2851/// the given expression into the given use. 2852void 2853LSRInstance::InsertSupplementalFormula(const SCEV *S, 2854 LSRUse &LU, size_t LUIdx) { 2855 Formula F; 2856 F.BaseRegs.push_back(S); 2857 F.AM.HasBaseReg = true; 2858 bool Inserted = InsertFormula(LU, LUIdx, F); 2859 assert(Inserted && "Supplemental formula already exists!"); (void)Inserted; 2860} 2861 2862/// CountRegisters - Note which registers are used by the given formula, 2863/// updating RegUses. 2864void LSRInstance::CountRegisters(const Formula &F, size_t LUIdx) { 2865 if (F.ScaledReg) 2866 RegUses.CountRegister(F.ScaledReg, LUIdx); 2867 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(), 2868 E = F.BaseRegs.end(); I != E; ++I) 2869 RegUses.CountRegister(*I, LUIdx); 2870} 2871 2872/// InsertFormula - If the given formula has not yet been inserted, add it to 2873/// the list, and return true. Return false otherwise. 2874bool LSRInstance::InsertFormula(LSRUse &LU, unsigned LUIdx, const Formula &F) { 2875 if (!LU.InsertFormula(F)) 2876 return false; 2877 2878 CountRegisters(F, LUIdx); 2879 return true; 2880} 2881 2882/// CollectLoopInvariantFixupsAndFormulae - Check for other uses of 2883/// loop-invariant values which we're tracking. These other uses will pin these 2884/// values in registers, making them less profitable for elimination. 2885/// TODO: This currently misses non-constant addrec step registers. 2886/// TODO: Should this give more weight to users inside the loop? 2887void 2888LSRInstance::CollectLoopInvariantFixupsAndFormulae() { 2889 SmallVector<const SCEV *, 8> Worklist(RegUses.begin(), RegUses.end()); 2890 SmallPtrSet<const SCEV *, 8> Inserted; 2891 2892 while (!Worklist.empty()) { 2893 const SCEV *S = Worklist.pop_back_val(); 2894 2895 if (const SCEVNAryExpr *N = dyn_cast<SCEVNAryExpr>(S)) 2896 Worklist.append(N->op_begin(), N->op_end()); 2897 else if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(S)) 2898 Worklist.push_back(C->getOperand()); 2899 else if (const SCEVUDivExpr *D = dyn_cast<SCEVUDivExpr>(S)) { 2900 Worklist.push_back(D->getLHS()); 2901 Worklist.push_back(D->getRHS()); 2902 } else if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) { 2903 if (!Inserted.insert(U)) continue; 2904 const Value *V = U->getValue(); 2905 if (const Instruction *Inst = dyn_cast<Instruction>(V)) { 2906 // Look for instructions defined outside the loop. 2907 if (L->contains(Inst)) continue; 2908 } else if (isa<UndefValue>(V)) 2909 // Undef doesn't have a live range, so it doesn't matter. 2910 continue; 2911 for (Value::const_use_iterator UI = V->use_begin(), UE = V->use_end(); 2912 UI != UE; ++UI) { 2913 const Instruction *UserInst = dyn_cast<Instruction>(*UI); 2914 // Ignore non-instructions. 2915 if (!UserInst) 2916 continue; 2917 // Ignore instructions in other functions (as can happen with 2918 // Constants). 2919 if (UserInst->getParent()->getParent() != L->getHeader()->getParent()) 2920 continue; 2921 // Ignore instructions not dominated by the loop. 2922 const BasicBlock *UseBB = !isa<PHINode>(UserInst) ? 2923 UserInst->getParent() : 2924 cast<PHINode>(UserInst)->getIncomingBlock( 2925 PHINode::getIncomingValueNumForOperand(UI.getOperandNo())); 2926 if (!DT.dominates(L->getHeader(), UseBB)) 2927 continue; 2928 // Ignore uses which are part of other SCEV expressions, to avoid 2929 // analyzing them multiple times. 2930 if (SE.isSCEVable(UserInst->getType())) { 2931 const SCEV *UserS = SE.getSCEV(const_cast<Instruction *>(UserInst)); 2932 // If the user is a no-op, look through to its uses. 2933 if (!isa<SCEVUnknown>(UserS)) 2934 continue; 2935 if (UserS == U) { 2936 Worklist.push_back( 2937 SE.getUnknown(const_cast<Instruction *>(UserInst))); 2938 continue; 2939 } 2940 } 2941 // Ignore icmp instructions which are already being analyzed. 2942 if (const ICmpInst *ICI = dyn_cast<ICmpInst>(UserInst)) { 2943 unsigned OtherIdx = !UI.getOperandNo(); 2944 Value *OtherOp = const_cast<Value *>(ICI->getOperand(OtherIdx)); 2945 if (SE.hasComputableLoopEvolution(SE.getSCEV(OtherOp), L)) 2946 continue; 2947 } 2948 2949 LSRFixup &LF = getNewFixup(); 2950 LF.UserInst = const_cast<Instruction *>(UserInst); 2951 LF.OperandValToReplace = UI.getUse(); 2952 std::pair<size_t, int64_t> P = getUse(S, LSRUse::Basic, 0); 2953 LF.LUIdx = P.first; 2954 LF.Offset = P.second; 2955 LSRUse &LU = Uses[LF.LUIdx]; 2956 LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L); 2957 if (!LU.WidestFixupType || 2958 SE.getTypeSizeInBits(LU.WidestFixupType) < 2959 SE.getTypeSizeInBits(LF.OperandValToReplace->getType())) 2960 LU.WidestFixupType = LF.OperandValToReplace->getType(); 2961 InsertSupplementalFormula(U, LU, LF.LUIdx); 2962 CountRegisters(LU.Formulae.back(), Uses.size() - 1); 2963 break; 2964 } 2965 } 2966 } 2967} 2968 2969/// CollectSubexprs - Split S into subexpressions which can be pulled out into 2970/// separate registers. If C is non-null, multiply each subexpression by C. 2971static void CollectSubexprs(const SCEV *S, const SCEVConstant *C, 2972 SmallVectorImpl<const SCEV *> &Ops, 2973 const Loop *L, 2974 ScalarEvolution &SE) { 2975 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) { 2976 // Break out add operands. 2977 for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end(); 2978 I != E; ++I) 2979 CollectSubexprs(*I, C, Ops, L, SE); 2980 return; 2981 } else if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) { 2982 // Split a non-zero base out of an addrec. 2983 if (!AR->getStart()->isZero()) { 2984 CollectSubexprs(SE.getAddRecExpr(SE.getConstant(AR->getType(), 0), 2985 AR->getStepRecurrence(SE), 2986 AR->getLoop(), 2987 //FIXME: AR->getNoWrapFlags(SCEV::FlagNW) 2988 SCEV::FlagAnyWrap), 2989 C, Ops, L, SE); 2990 CollectSubexprs(AR->getStart(), C, Ops, L, SE); 2991 return; 2992 } 2993 } else if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) { 2994 // Break (C * (a + b + c)) into C*a + C*b + C*c. 2995 if (Mul->getNumOperands() == 2) 2996 if (const SCEVConstant *Op0 = 2997 dyn_cast<SCEVConstant>(Mul->getOperand(0))) { 2998 CollectSubexprs(Mul->getOperand(1), 2999 C ? cast<SCEVConstant>(SE.getMulExpr(C, Op0)) : Op0, 3000 Ops, L, SE); 3001 return; 3002 } 3003 } 3004 3005 // Otherwise use the value itself, optionally with a scale applied. 3006 Ops.push_back(C ? SE.getMulExpr(C, S) : S); 3007} 3008 3009/// GenerateReassociations - Split out subexpressions from adds and the bases of 3010/// addrecs. 3011void LSRInstance::GenerateReassociations(LSRUse &LU, unsigned LUIdx, 3012 Formula Base, 3013 unsigned Depth) { 3014 // Arbitrarily cap recursion to protect compile time. 3015 if (Depth >= 3) return; 3016 3017 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) { 3018 const SCEV *BaseReg = Base.BaseRegs[i]; 3019 3020 SmallVector<const SCEV *, 8> AddOps; 3021 CollectSubexprs(BaseReg, 0, AddOps, L, SE); 3022 3023 if (AddOps.size() == 1) continue; 3024 3025 for (SmallVectorImpl<const SCEV *>::const_iterator J = AddOps.begin(), 3026 JE = AddOps.end(); J != JE; ++J) { 3027 3028 // Loop-variant "unknown" values are uninteresting; we won't be able to 3029 // do anything meaningful with them. 3030 if (isa<SCEVUnknown>(*J) && !SE.isLoopInvariant(*J, L)) 3031 continue; 3032 3033 // Don't pull a constant into a register if the constant could be folded 3034 // into an immediate field. 3035 if (isAlwaysFoldable(*J, LU.MinOffset, LU.MaxOffset, 3036 Base.getNumRegs() > 1, 3037 LU.Kind, LU.AccessTy, TLI, SE)) 3038 continue; 3039 3040 // Collect all operands except *J. 3041 SmallVector<const SCEV *, 8> InnerAddOps 3042 (((const SmallVector<const SCEV *, 8> &)AddOps).begin(), J); 3043 InnerAddOps.append 3044 (llvm::next(J), ((const SmallVector<const SCEV *, 8> &)AddOps).end()); 3045 3046 // Don't leave just a constant behind in a register if the constant could 3047 // be folded into an immediate field. 3048 if (InnerAddOps.size() == 1 && 3049 isAlwaysFoldable(InnerAddOps[0], LU.MinOffset, LU.MaxOffset, 3050 Base.getNumRegs() > 1, 3051 LU.Kind, LU.AccessTy, TLI, SE)) 3052 continue; 3053 3054 const SCEV *InnerSum = SE.getAddExpr(InnerAddOps); 3055 if (InnerSum->isZero()) 3056 continue; 3057 Formula F = Base; 3058 3059 // Add the remaining pieces of the add back into the new formula. 3060 const SCEVConstant *InnerSumSC = dyn_cast<SCEVConstant>(InnerSum); 3061 if (TLI && InnerSumSC && 3062 SE.getTypeSizeInBits(InnerSumSC->getType()) <= 64 && 3063 TLI->isLegalAddImmediate((uint64_t)F.UnfoldedOffset + 3064 InnerSumSC->getValue()->getZExtValue())) { 3065 F.UnfoldedOffset = (uint64_t)F.UnfoldedOffset + 3066 InnerSumSC->getValue()->getZExtValue(); 3067 F.BaseRegs.erase(F.BaseRegs.begin() + i); 3068 } else 3069 F.BaseRegs[i] = InnerSum; 3070 3071 // Add J as its own register, or an unfolded immediate. 3072 const SCEVConstant *SC = dyn_cast<SCEVConstant>(*J); 3073 if (TLI && SC && SE.getTypeSizeInBits(SC->getType()) <= 64 && 3074 TLI->isLegalAddImmediate((uint64_t)F.UnfoldedOffset + 3075 SC->getValue()->getZExtValue())) 3076 F.UnfoldedOffset = (uint64_t)F.UnfoldedOffset + 3077 SC->getValue()->getZExtValue(); 3078 else 3079 F.BaseRegs.push_back(*J); 3080 3081 if (InsertFormula(LU, LUIdx, F)) 3082 // If that formula hadn't been seen before, recurse to find more like 3083 // it. 3084 GenerateReassociations(LU, LUIdx, LU.Formulae.back(), Depth+1); 3085 } 3086 } 3087} 3088 3089/// GenerateCombinations - Generate a formula consisting of all of the 3090/// loop-dominating registers added into a single register. 3091void LSRInstance::GenerateCombinations(LSRUse &LU, unsigned LUIdx, 3092 Formula Base) { 3093 // This method is only interesting on a plurality of registers. 3094 if (Base.BaseRegs.size() <= 1) return; 3095 3096 Formula F = Base; 3097 F.BaseRegs.clear(); 3098 SmallVector<const SCEV *, 4> Ops; 3099 for (SmallVectorImpl<const SCEV *>::const_iterator 3100 I = Base.BaseRegs.begin(), E = Base.BaseRegs.end(); I != E; ++I) { 3101 const SCEV *BaseReg = *I; 3102 if (SE.properlyDominates(BaseReg, L->getHeader()) && 3103 !SE.hasComputableLoopEvolution(BaseReg, L)) 3104 Ops.push_back(BaseReg); 3105 else 3106 F.BaseRegs.push_back(BaseReg); 3107 } 3108 if (Ops.size() > 1) { 3109 const SCEV *Sum = SE.getAddExpr(Ops); 3110 // TODO: If Sum is zero, it probably means ScalarEvolution missed an 3111 // opportunity to fold something. For now, just ignore such cases 3112 // rather than proceed with zero in a register. 3113 if (!Sum->isZero()) { 3114 F.BaseRegs.push_back(Sum); 3115 (void)InsertFormula(LU, LUIdx, F); 3116 } 3117 } 3118} 3119 3120/// GenerateSymbolicOffsets - Generate reuse formulae using symbolic offsets. 3121void LSRInstance::GenerateSymbolicOffsets(LSRUse &LU, unsigned LUIdx, 3122 Formula Base) { 3123 // We can't add a symbolic offset if the address already contains one. 3124 if (Base.AM.BaseGV) return; 3125 3126 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) { 3127 const SCEV *G = Base.BaseRegs[i]; 3128 GlobalValue *GV = ExtractSymbol(G, SE); 3129 if (G->isZero() || !GV) 3130 continue; 3131 Formula F = Base; 3132 F.AM.BaseGV = GV; 3133 if (!isLegalUse(F.AM, LU.MinOffset, LU.MaxOffset, 3134 LU.Kind, LU.AccessTy, TLI)) 3135 continue; 3136 F.BaseRegs[i] = G; 3137 (void)InsertFormula(LU, LUIdx, F); 3138 } 3139} 3140 3141/// GenerateConstantOffsets - Generate reuse formulae using symbolic offsets. 3142void LSRInstance::GenerateConstantOffsets(LSRUse &LU, unsigned LUIdx, 3143 Formula Base) { 3144 // TODO: For now, just add the min and max offset, because it usually isn't 3145 // worthwhile looking at everything inbetween. 3146 SmallVector<int64_t, 2> Worklist; 3147 Worklist.push_back(LU.MinOffset); 3148 if (LU.MaxOffset != LU.MinOffset) 3149 Worklist.push_back(LU.MaxOffset); 3150 3151 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) { 3152 const SCEV *G = Base.BaseRegs[i]; 3153 3154 for (SmallVectorImpl<int64_t>::const_iterator I = Worklist.begin(), 3155 E = Worklist.end(); I != E; ++I) { 3156 Formula F = Base; 3157 F.AM.BaseOffs = (uint64_t)Base.AM.BaseOffs - *I; 3158 if (isLegalUse(F.AM, LU.MinOffset - *I, LU.MaxOffset - *I, 3159 LU.Kind, LU.AccessTy, TLI)) { 3160 // Add the offset to the base register. 3161 const SCEV *NewG = SE.getAddExpr(SE.getConstant(G->getType(), *I), G); 3162 // If it cancelled out, drop the base register, otherwise update it. 3163 if (NewG->isZero()) { 3164 std::swap(F.BaseRegs[i], F.BaseRegs.back()); 3165 F.BaseRegs.pop_back(); 3166 } else 3167 F.BaseRegs[i] = NewG; 3168 3169 (void)InsertFormula(LU, LUIdx, F); 3170 } 3171 } 3172 3173 int64_t Imm = ExtractImmediate(G, SE); 3174 if (G->isZero() || Imm == 0) 3175 continue; 3176 Formula F = Base; 3177 F.AM.BaseOffs = (uint64_t)F.AM.BaseOffs + Imm; 3178 if (!isLegalUse(F.AM, LU.MinOffset, LU.MaxOffset, 3179 LU.Kind, LU.AccessTy, TLI)) 3180 continue; 3181 F.BaseRegs[i] = G; 3182 (void)InsertFormula(LU, LUIdx, F); 3183 } 3184} 3185 3186/// GenerateICmpZeroScales - For ICmpZero, check to see if we can scale up 3187/// the comparison. For example, x == y -> x*c == y*c. 3188void LSRInstance::GenerateICmpZeroScales(LSRUse &LU, unsigned LUIdx, 3189 Formula Base) { 3190 if (LU.Kind != LSRUse::ICmpZero) return; 3191 3192 // Determine the integer type for the base formula. 3193 Type *IntTy = Base.getType(); 3194 if (!IntTy) return; 3195 if (SE.getTypeSizeInBits(IntTy) > 64) return; 3196 3197 // Don't do this if there is more than one offset. 3198 if (LU.MinOffset != LU.MaxOffset) return; 3199 3200 assert(!Base.AM.BaseGV && "ICmpZero use is not legal!"); 3201 3202 // Check each interesting stride. 3203 for (SmallSetVector<int64_t, 8>::const_iterator 3204 I = Factors.begin(), E = Factors.end(); I != E; ++I) { 3205 int64_t Factor = *I; 3206 3207 // Check that the multiplication doesn't overflow. 3208 if (Base.AM.BaseOffs == INT64_MIN && Factor == -1) 3209 continue; 3210 int64_t NewBaseOffs = (uint64_t)Base.AM.BaseOffs * Factor; 3211 if (NewBaseOffs / Factor != Base.AM.BaseOffs) 3212 continue; 3213 3214 // Check that multiplying with the use offset doesn't overflow. 3215 int64_t Offset = LU.MinOffset; 3216 if (Offset == INT64_MIN && Factor == -1) 3217 continue; 3218 Offset = (uint64_t)Offset * Factor; 3219 if (Offset / Factor != LU.MinOffset) 3220 continue; 3221 3222 Formula F = Base; 3223 F.AM.BaseOffs = NewBaseOffs; 3224 3225 // Check that this scale is legal. 3226 if (!isLegalUse(F.AM, Offset, Offset, LU.Kind, LU.AccessTy, TLI)) 3227 continue; 3228 3229 // Compensate for the use having MinOffset built into it. 3230 F.AM.BaseOffs = (uint64_t)F.AM.BaseOffs + Offset - LU.MinOffset; 3231 3232 const SCEV *FactorS = SE.getConstant(IntTy, Factor); 3233 3234 // Check that multiplying with each base register doesn't overflow. 3235 for (size_t i = 0, e = F.BaseRegs.size(); i != e; ++i) { 3236 F.BaseRegs[i] = SE.getMulExpr(F.BaseRegs[i], FactorS); 3237 if (getExactSDiv(F.BaseRegs[i], FactorS, SE) != Base.BaseRegs[i]) 3238 goto next; 3239 } 3240 3241 // Check that multiplying with the scaled register doesn't overflow. 3242 if (F.ScaledReg) { 3243 F.ScaledReg = SE.getMulExpr(F.ScaledReg, FactorS); 3244 if (getExactSDiv(F.ScaledReg, FactorS, SE) != Base.ScaledReg) 3245 continue; 3246 } 3247 3248 // Check that multiplying with the unfolded offset doesn't overflow. 3249 if (F.UnfoldedOffset != 0) { 3250 if (F.UnfoldedOffset == INT64_MIN && Factor == -1) 3251 continue; 3252 F.UnfoldedOffset = (uint64_t)F.UnfoldedOffset * Factor; 3253 if (F.UnfoldedOffset / Factor != Base.UnfoldedOffset) 3254 continue; 3255 } 3256 3257 // If we make it here and it's legal, add it. 3258 (void)InsertFormula(LU, LUIdx, F); 3259 next:; 3260 } 3261} 3262 3263/// GenerateScales - Generate stride factor reuse formulae by making use of 3264/// scaled-offset address modes, for example. 3265void LSRInstance::GenerateScales(LSRUse &LU, unsigned LUIdx, Formula Base) { 3266 // Determine the integer type for the base formula. 3267 Type *IntTy = Base.getType(); 3268 if (!IntTy) return; 3269 3270 // If this Formula already has a scaled register, we can't add another one. 3271 if (Base.AM.Scale != 0) return; 3272 3273 // Check each interesting stride. 3274 for (SmallSetVector<int64_t, 8>::const_iterator 3275 I = Factors.begin(), E = Factors.end(); I != E; ++I) { 3276 int64_t Factor = *I; 3277 3278 Base.AM.Scale = Factor; 3279 Base.AM.HasBaseReg = Base.BaseRegs.size() > 1; 3280 // Check whether this scale is going to be legal. 3281 if (!isLegalUse(Base.AM, LU.MinOffset, LU.MaxOffset, 3282 LU.Kind, LU.AccessTy, TLI)) { 3283 // As a special-case, handle special out-of-loop Basic users specially. 3284 // TODO: Reconsider this special case. 3285 if (LU.Kind == LSRUse::Basic && 3286 isLegalUse(Base.AM, LU.MinOffset, LU.MaxOffset, 3287 LSRUse::Special, LU.AccessTy, TLI) && 3288 LU.AllFixupsOutsideLoop) 3289 LU.Kind = LSRUse::Special; 3290 else 3291 continue; 3292 } 3293 // For an ICmpZero, negating a solitary base register won't lead to 3294 // new solutions. 3295 if (LU.Kind == LSRUse::ICmpZero && 3296 !Base.AM.HasBaseReg && Base.AM.BaseOffs == 0 && !Base.AM.BaseGV) 3297 continue; 3298 // For each addrec base reg, apply the scale, if possible. 3299 for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) 3300 if (const SCEVAddRecExpr *AR = 3301 dyn_cast<SCEVAddRecExpr>(Base.BaseRegs[i])) { 3302 const SCEV *FactorS = SE.getConstant(IntTy, Factor); 3303 if (FactorS->isZero()) 3304 continue; 3305 // Divide out the factor, ignoring high bits, since we'll be 3306 // scaling the value back up in the end. 3307 if (const SCEV *Quotient = getExactSDiv(AR, FactorS, SE, true)) { 3308 // TODO: This could be optimized to avoid all the copying. 3309 Formula F = Base; 3310 F.ScaledReg = Quotient; 3311 F.DeleteBaseReg(F.BaseRegs[i]); 3312 (void)InsertFormula(LU, LUIdx, F); 3313 } 3314 } 3315 } 3316} 3317 3318/// GenerateTruncates - Generate reuse formulae from different IV types. 3319void LSRInstance::GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base) { 3320 // This requires TargetLowering to tell us which truncates are free. 3321 if (!TLI) return; 3322 3323 // Don't bother truncating symbolic values. 3324 if (Base.AM.BaseGV) return; 3325 3326 // Determine the integer type for the base formula. 3327 Type *DstTy = Base.getType(); 3328 if (!DstTy) return; 3329 DstTy = SE.getEffectiveSCEVType(DstTy); 3330 3331 for (SmallSetVector<Type *, 4>::const_iterator 3332 I = Types.begin(), E = Types.end(); I != E; ++I) { 3333 Type *SrcTy = *I; 3334 if (SrcTy != DstTy && TLI->isTruncateFree(SrcTy, DstTy)) { 3335 Formula F = Base; 3336 3337 if (F.ScaledReg) F.ScaledReg = SE.getAnyExtendExpr(F.ScaledReg, *I); 3338 for (SmallVectorImpl<const SCEV *>::iterator J = F.BaseRegs.begin(), 3339 JE = F.BaseRegs.end(); J != JE; ++J) 3340 *J = SE.getAnyExtendExpr(*J, SrcTy); 3341 3342 // TODO: This assumes we've done basic processing on all uses and 3343 // have an idea what the register usage is. 3344 if (!F.hasRegsUsedByUsesOtherThan(LUIdx, RegUses)) 3345 continue; 3346 3347 (void)InsertFormula(LU, LUIdx, F); 3348 } 3349 } 3350} 3351 3352namespace { 3353 3354/// WorkItem - Helper class for GenerateCrossUseConstantOffsets. It's used to 3355/// defer modifications so that the search phase doesn't have to worry about 3356/// the data structures moving underneath it. 3357struct WorkItem { 3358 size_t LUIdx; 3359 int64_t Imm; 3360 const SCEV *OrigReg; 3361 3362 WorkItem(size_t LI, int64_t I, const SCEV *R) 3363 : LUIdx(LI), Imm(I), OrigReg(R) {} 3364 3365 void print(raw_ostream &OS) const; 3366 void dump() const; 3367}; 3368 3369} 3370 3371void WorkItem::print(raw_ostream &OS) const { 3372 OS << "in formulae referencing " << *OrigReg << " in use " << LUIdx 3373 << " , add offset " << Imm; 3374} 3375 3376void WorkItem::dump() const { 3377 print(errs()); errs() << '\n'; 3378} 3379 3380/// GenerateCrossUseConstantOffsets - Look for registers which are a constant 3381/// distance apart and try to form reuse opportunities between them. 3382void LSRInstance::GenerateCrossUseConstantOffsets() { 3383 // Group the registers by their value without any added constant offset. 3384 typedef std::map<int64_t, const SCEV *> ImmMapTy; 3385 typedef DenseMap<const SCEV *, ImmMapTy> RegMapTy; 3386 RegMapTy Map; 3387 DenseMap<const SCEV *, SmallBitVector> UsedByIndicesMap; 3388 SmallVector<const SCEV *, 8> Sequence; 3389 for (RegUseTracker::const_iterator I = RegUses.begin(), E = RegUses.end(); 3390 I != E; ++I) { 3391 const SCEV *Reg = *I; 3392 int64_t Imm = ExtractImmediate(Reg, SE); 3393 std::pair<RegMapTy::iterator, bool> Pair = 3394 Map.insert(std::make_pair(Reg, ImmMapTy())); 3395 if (Pair.second) 3396 Sequence.push_back(Reg); 3397 Pair.first->second.insert(std::make_pair(Imm, *I)); 3398 UsedByIndicesMap[Reg] |= RegUses.getUsedByIndices(*I); 3399 } 3400 3401 // Now examine each set of registers with the same base value. Build up 3402 // a list of work to do and do the work in a separate step so that we're 3403 // not adding formulae and register counts while we're searching. 3404 SmallVector<WorkItem, 32> WorkItems; 3405 SmallSet<std::pair<size_t, int64_t>, 32> UniqueItems; 3406 for (SmallVectorImpl<const SCEV *>::const_iterator I = Sequence.begin(), 3407 E = Sequence.end(); I != E; ++I) { 3408 const SCEV *Reg = *I; 3409 const ImmMapTy &Imms = Map.find(Reg)->second; 3410 3411 // It's not worthwhile looking for reuse if there's only one offset. 3412 if (Imms.size() == 1) 3413 continue; 3414 3415 DEBUG(dbgs() << "Generating cross-use offsets for " << *Reg << ':'; 3416 for (ImmMapTy::const_iterator J = Imms.begin(), JE = Imms.end(); 3417 J != JE; ++J) 3418 dbgs() << ' ' << J->first; 3419 dbgs() << '\n'); 3420 3421 // Examine each offset. 3422 for (ImmMapTy::const_iterator J = Imms.begin(), JE = Imms.end(); 3423 J != JE; ++J) { 3424 const SCEV *OrigReg = J->second; 3425 3426 int64_t JImm = J->first; 3427 const SmallBitVector &UsedByIndices = RegUses.getUsedByIndices(OrigReg); 3428 3429 if (!isa<SCEVConstant>(OrigReg) && 3430 UsedByIndicesMap[Reg].count() == 1) { 3431 DEBUG(dbgs() << "Skipping cross-use reuse for " << *OrigReg << '\n'); 3432 continue; 3433 } 3434 3435 // Conservatively examine offsets between this orig reg a few selected 3436 // other orig regs. 3437 ImmMapTy::const_iterator OtherImms[] = { 3438 Imms.begin(), prior(Imms.end()), 3439 Imms.lower_bound((Imms.begin()->first + prior(Imms.end())->first) / 2) 3440 }; 3441 for (size_t i = 0, e = array_lengthof(OtherImms); i != e; ++i) { 3442 ImmMapTy::const_iterator M = OtherImms[i]; 3443 if (M == J || M == JE) continue; 3444 3445 // Compute the difference between the two. 3446 int64_t Imm = (uint64_t)JImm - M->first; 3447 for (int LUIdx = UsedByIndices.find_first(); LUIdx != -1; 3448 LUIdx = UsedByIndices.find_next(LUIdx)) 3449 // Make a memo of this use, offset, and register tuple. 3450 if (UniqueItems.insert(std::make_pair(LUIdx, Imm))) 3451 WorkItems.push_back(WorkItem(LUIdx, Imm, OrigReg)); 3452 } 3453 } 3454 } 3455 3456 Map.clear(); 3457 Sequence.clear(); 3458 UsedByIndicesMap.clear(); 3459 UniqueItems.clear(); 3460 3461 // Now iterate through the worklist and add new formulae. 3462 for (SmallVectorImpl<WorkItem>::const_iterator I = WorkItems.begin(), 3463 E = WorkItems.end(); I != E; ++I) { 3464 const WorkItem &WI = *I; 3465 size_t LUIdx = WI.LUIdx; 3466 LSRUse &LU = Uses[LUIdx]; 3467 int64_t Imm = WI.Imm; 3468 const SCEV *OrigReg = WI.OrigReg; 3469 3470 Type *IntTy = SE.getEffectiveSCEVType(OrigReg->getType()); 3471 const SCEV *NegImmS = SE.getSCEV(ConstantInt::get(IntTy, -(uint64_t)Imm)); 3472 unsigned BitWidth = SE.getTypeSizeInBits(IntTy); 3473 3474 // TODO: Use a more targeted data structure. 3475 for (size_t L = 0, LE = LU.Formulae.size(); L != LE; ++L) { 3476 const Formula &F = LU.Formulae[L]; 3477 // Use the immediate in the scaled register. 3478 if (F.ScaledReg == OrigReg) { 3479 int64_t Offs = (uint64_t)F.AM.BaseOffs + 3480 Imm * (uint64_t)F.AM.Scale; 3481 // Don't create 50 + reg(-50). 3482 if (F.referencesReg(SE.getSCEV( 3483 ConstantInt::get(IntTy, -(uint64_t)Offs)))) 3484 continue; 3485 Formula NewF = F; 3486 NewF.AM.BaseOffs = Offs; 3487 if (!isLegalUse(NewF.AM, LU.MinOffset, LU.MaxOffset, 3488 LU.Kind, LU.AccessTy, TLI)) 3489 continue; 3490 NewF.ScaledReg = SE.getAddExpr(NegImmS, NewF.ScaledReg); 3491 3492 // If the new scale is a constant in a register, and adding the constant 3493 // value to the immediate would produce a value closer to zero than the 3494 // immediate itself, then the formula isn't worthwhile. 3495 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(NewF.ScaledReg)) 3496 if (C->getValue()->isNegative() != 3497 (NewF.AM.BaseOffs < 0) && 3498 (C->getValue()->getValue().abs() * APInt(BitWidth, F.AM.Scale)) 3499 .ule(abs64(NewF.AM.BaseOffs))) 3500 continue; 3501 3502 // OK, looks good. 3503 (void)InsertFormula(LU, LUIdx, NewF); 3504 } else { 3505 // Use the immediate in a base register. 3506 for (size_t N = 0, NE = F.BaseRegs.size(); N != NE; ++N) { 3507 const SCEV *BaseReg = F.BaseRegs[N]; 3508 if (BaseReg != OrigReg) 3509 continue; 3510 Formula NewF = F; 3511 NewF.AM.BaseOffs = (uint64_t)NewF.AM.BaseOffs + Imm; 3512 if (!isLegalUse(NewF.AM, LU.MinOffset, LU.MaxOffset, 3513 LU.Kind, LU.AccessTy, TLI)) { 3514 if (!TLI || 3515 !TLI->isLegalAddImmediate((uint64_t)NewF.UnfoldedOffset + Imm)) 3516 continue; 3517 NewF = F; 3518 NewF.UnfoldedOffset = (uint64_t)NewF.UnfoldedOffset + Imm; 3519 } 3520 NewF.BaseRegs[N] = SE.getAddExpr(NegImmS, BaseReg); 3521 3522 // If the new formula has a constant in a register, and adding the 3523 // constant value to the immediate would produce a value closer to 3524 // zero than the immediate itself, then the formula isn't worthwhile. 3525 for (SmallVectorImpl<const SCEV *>::const_iterator 3526 J = NewF.BaseRegs.begin(), JE = NewF.BaseRegs.end(); 3527 J != JE; ++J) 3528 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(*J)) 3529 if ((C->getValue()->getValue() + NewF.AM.BaseOffs).abs().slt( 3530 abs64(NewF.AM.BaseOffs)) && 3531 (C->getValue()->getValue() + 3532 NewF.AM.BaseOffs).countTrailingZeros() >= 3533 CountTrailingZeros_64(NewF.AM.BaseOffs)) 3534 goto skip_formula; 3535 3536 // Ok, looks good. 3537 (void)InsertFormula(LU, LUIdx, NewF); 3538 break; 3539 skip_formula:; 3540 } 3541 } 3542 } 3543 } 3544} 3545 3546/// GenerateAllReuseFormulae - Generate formulae for each use. 3547void 3548LSRInstance::GenerateAllReuseFormulae() { 3549 // This is split into multiple loops so that hasRegsUsedByUsesOtherThan 3550 // queries are more precise. 3551 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) { 3552 LSRUse &LU = Uses[LUIdx]; 3553 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i) 3554 GenerateReassociations(LU, LUIdx, LU.Formulae[i]); 3555 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i) 3556 GenerateCombinations(LU, LUIdx, LU.Formulae[i]); 3557 } 3558 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) { 3559 LSRUse &LU = Uses[LUIdx]; 3560 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i) 3561 GenerateSymbolicOffsets(LU, LUIdx, LU.Formulae[i]); 3562 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i) 3563 GenerateConstantOffsets(LU, LUIdx, LU.Formulae[i]); 3564 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i) 3565 GenerateICmpZeroScales(LU, LUIdx, LU.Formulae[i]); 3566 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i) 3567 GenerateScales(LU, LUIdx, LU.Formulae[i]); 3568 } 3569 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) { 3570 LSRUse &LU = Uses[LUIdx]; 3571 for (size_t i = 0, f = LU.Formulae.size(); i != f; ++i) 3572 GenerateTruncates(LU, LUIdx, LU.Formulae[i]); 3573 } 3574 3575 GenerateCrossUseConstantOffsets(); 3576 3577 DEBUG(dbgs() << "\n" 3578 "After generating reuse formulae:\n"; 3579 print_uses(dbgs())); 3580} 3581 3582/// If there are multiple formulae with the same set of registers used 3583/// by other uses, pick the best one and delete the others. 3584void LSRInstance::FilterOutUndesirableDedicatedRegisters() { 3585 DenseSet<const SCEV *> VisitedRegs; 3586 SmallPtrSet<const SCEV *, 16> Regs; 3587 SmallPtrSet<const SCEV *, 16> LoserRegs; 3588#ifndef NDEBUG 3589 bool ChangedFormulae = false; 3590#endif 3591 3592 // Collect the best formula for each unique set of shared registers. This 3593 // is reset for each use. 3594 typedef DenseMap<SmallVector<const SCEV *, 2>, size_t, UniquifierDenseMapInfo> 3595 BestFormulaeTy; 3596 BestFormulaeTy BestFormulae; 3597 3598 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) { 3599 LSRUse &LU = Uses[LUIdx]; 3600 DEBUG(dbgs() << "Filtering for use "; LU.print(dbgs()); dbgs() << '\n'); 3601 3602 bool Any = false; 3603 for (size_t FIdx = 0, NumForms = LU.Formulae.size(); 3604 FIdx != NumForms; ++FIdx) { 3605 Formula &F = LU.Formulae[FIdx]; 3606 3607 // Some formulas are instant losers. For example, they may depend on 3608 // nonexistent AddRecs from other loops. These need to be filtered 3609 // immediately, otherwise heuristics could choose them over others leading 3610 // to an unsatisfactory solution. Passing LoserRegs into RateFormula here 3611 // avoids the need to recompute this information across formulae using the 3612 // same bad AddRec. Passing LoserRegs is also essential unless we remove 3613 // the corresponding bad register from the Regs set. 3614 Cost CostF; 3615 Regs.clear(); 3616 CostF.RateFormula(F, Regs, VisitedRegs, L, LU.Offsets, SE, DT, 3617 &LoserRegs); 3618 if (CostF.isLoser()) { 3619 // During initial formula generation, undesirable formulae are generated 3620 // by uses within other loops that have some non-trivial address mode or 3621 // use the postinc form of the IV. LSR needs to provide these formulae 3622 // as the basis of rediscovering the desired formula that uses an AddRec 3623 // corresponding to the existing phi. Once all formulae have been 3624 // generated, these initial losers may be pruned. 3625 DEBUG(dbgs() << " Filtering loser "; F.print(dbgs()); 3626 dbgs() << "\n"); 3627 } 3628 else { 3629 SmallVector<const SCEV *, 2> Key; 3630 for (SmallVectorImpl<const SCEV *>::const_iterator J = F.BaseRegs.begin(), 3631 JE = F.BaseRegs.end(); J != JE; ++J) { 3632 const SCEV *Reg = *J; 3633 if (RegUses.isRegUsedByUsesOtherThan(Reg, LUIdx)) 3634 Key.push_back(Reg); 3635 } 3636 if (F.ScaledReg && 3637 RegUses.isRegUsedByUsesOtherThan(F.ScaledReg, LUIdx)) 3638 Key.push_back(F.ScaledReg); 3639 // Unstable sort by host order ok, because this is only used for 3640 // uniquifying. 3641 std::sort(Key.begin(), Key.end()); 3642 3643 std::pair<BestFormulaeTy::const_iterator, bool> P = 3644 BestFormulae.insert(std::make_pair(Key, FIdx)); 3645 if (P.second) 3646 continue; 3647 3648 Formula &Best = LU.Formulae[P.first->second]; 3649 3650 Cost CostBest; 3651 Regs.clear(); 3652 CostBest.RateFormula(Best, Regs, VisitedRegs, L, LU.Offsets, SE, DT); 3653 if (CostF < CostBest) 3654 std::swap(F, Best); 3655 DEBUG(dbgs() << " Filtering out formula "; F.print(dbgs()); 3656 dbgs() << "\n" 3657 " in favor of formula "; Best.print(dbgs()); 3658 dbgs() << '\n'); 3659 } 3660#ifndef NDEBUG 3661 ChangedFormulae = true; 3662#endif 3663 LU.DeleteFormula(F); 3664 --FIdx; 3665 --NumForms; 3666 Any = true; 3667 } 3668 3669 // Now that we've filtered out some formulae, recompute the Regs set. 3670 if (Any) 3671 LU.RecomputeRegs(LUIdx, RegUses); 3672 3673 // Reset this to prepare for the next use. 3674 BestFormulae.clear(); 3675 } 3676 3677 DEBUG(if (ChangedFormulae) { 3678 dbgs() << "\n" 3679 "After filtering out undesirable candidates:\n"; 3680 print_uses(dbgs()); 3681 }); 3682} 3683 3684// This is a rough guess that seems to work fairly well. 3685static const size_t ComplexityLimit = UINT16_MAX; 3686 3687/// EstimateSearchSpaceComplexity - Estimate the worst-case number of 3688/// solutions the solver might have to consider. It almost never considers 3689/// this many solutions because it prune the search space, but the pruning 3690/// isn't always sufficient. 3691size_t LSRInstance::EstimateSearchSpaceComplexity() const { 3692 size_t Power = 1; 3693 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(), 3694 E = Uses.end(); I != E; ++I) { 3695 size_t FSize = I->Formulae.size(); 3696 if (FSize >= ComplexityLimit) { 3697 Power = ComplexityLimit; 3698 break; 3699 } 3700 Power *= FSize; 3701 if (Power >= ComplexityLimit) 3702 break; 3703 } 3704 return Power; 3705} 3706 3707/// NarrowSearchSpaceByDetectingSupersets - When one formula uses a superset 3708/// of the registers of another formula, it won't help reduce register 3709/// pressure (though it may not necessarily hurt register pressure); remove 3710/// it to simplify the system. 3711void LSRInstance::NarrowSearchSpaceByDetectingSupersets() { 3712 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) { 3713 DEBUG(dbgs() << "The search space is too complex.\n"); 3714 3715 DEBUG(dbgs() << "Narrowing the search space by eliminating formulae " 3716 "which use a superset of registers used by other " 3717 "formulae.\n"); 3718 3719 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) { 3720 LSRUse &LU = Uses[LUIdx]; 3721 bool Any = false; 3722 for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) { 3723 Formula &F = LU.Formulae[i]; 3724 // Look for a formula with a constant or GV in a register. If the use 3725 // also has a formula with that same value in an immediate field, 3726 // delete the one that uses a register. 3727 for (SmallVectorImpl<const SCEV *>::const_iterator 3728 I = F.BaseRegs.begin(), E = F.BaseRegs.end(); I != E; ++I) { 3729 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(*I)) { 3730 Formula NewF = F; 3731 NewF.AM.BaseOffs += C->getValue()->getSExtValue(); 3732 NewF.BaseRegs.erase(NewF.BaseRegs.begin() + 3733 (I - F.BaseRegs.begin())); 3734 if (LU.HasFormulaWithSameRegs(NewF)) { 3735 DEBUG(dbgs() << " Deleting "; F.print(dbgs()); dbgs() << '\n'); 3736 LU.DeleteFormula(F); 3737 --i; 3738 --e; 3739 Any = true; 3740 break; 3741 } 3742 } else if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(*I)) { 3743 if (GlobalValue *GV = dyn_cast<GlobalValue>(U->getValue())) 3744 if (!F.AM.BaseGV) { 3745 Formula NewF = F; 3746 NewF.AM.BaseGV = GV; 3747 NewF.BaseRegs.erase(NewF.BaseRegs.begin() + 3748 (I - F.BaseRegs.begin())); 3749 if (LU.HasFormulaWithSameRegs(NewF)) { 3750 DEBUG(dbgs() << " Deleting "; F.print(dbgs()); 3751 dbgs() << '\n'); 3752 LU.DeleteFormula(F); 3753 --i; 3754 --e; 3755 Any = true; 3756 break; 3757 } 3758 } 3759 } 3760 } 3761 } 3762 if (Any) 3763 LU.RecomputeRegs(LUIdx, RegUses); 3764 } 3765 3766 DEBUG(dbgs() << "After pre-selection:\n"; 3767 print_uses(dbgs())); 3768 } 3769} 3770 3771/// NarrowSearchSpaceByCollapsingUnrolledCode - When there are many registers 3772/// for expressions like A, A+1, A+2, etc., allocate a single register for 3773/// them. 3774void LSRInstance::NarrowSearchSpaceByCollapsingUnrolledCode() { 3775 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) { 3776 DEBUG(dbgs() << "The search space is too complex.\n"); 3777 3778 DEBUG(dbgs() << "Narrowing the search space by assuming that uses " 3779 "separated by a constant offset will use the same " 3780 "registers.\n"); 3781 3782 // This is especially useful for unrolled loops. 3783 3784 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) { 3785 LSRUse &LU = Uses[LUIdx]; 3786 for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(), 3787 E = LU.Formulae.end(); I != E; ++I) { 3788 const Formula &F = *I; 3789 if (F.AM.BaseOffs != 0 && F.AM.Scale == 0) { 3790 if (LSRUse *LUThatHas = FindUseWithSimilarFormula(F, LU)) { 3791 if (reconcileNewOffset(*LUThatHas, F.AM.BaseOffs, 3792 /*HasBaseReg=*/false, 3793 LU.Kind, LU.AccessTy)) { 3794 DEBUG(dbgs() << " Deleting use "; LU.print(dbgs()); 3795 dbgs() << '\n'); 3796 3797 LUThatHas->AllFixupsOutsideLoop &= LU.AllFixupsOutsideLoop; 3798 3799 // Update the relocs to reference the new use. 3800 for (SmallVectorImpl<LSRFixup>::iterator I = Fixups.begin(), 3801 E = Fixups.end(); I != E; ++I) { 3802 LSRFixup &Fixup = *I; 3803 if (Fixup.LUIdx == LUIdx) { 3804 Fixup.LUIdx = LUThatHas - &Uses.front(); 3805 Fixup.Offset += F.AM.BaseOffs; 3806 // Add the new offset to LUThatHas' offset list. 3807 if (LUThatHas->Offsets.back() != Fixup.Offset) { 3808 LUThatHas->Offsets.push_back(Fixup.Offset); 3809 if (Fixup.Offset > LUThatHas->MaxOffset) 3810 LUThatHas->MaxOffset = Fixup.Offset; 3811 if (Fixup.Offset < LUThatHas->MinOffset) 3812 LUThatHas->MinOffset = Fixup.Offset; 3813 } 3814 DEBUG(dbgs() << "New fixup has offset " 3815 << Fixup.Offset << '\n'); 3816 } 3817 if (Fixup.LUIdx == NumUses-1) 3818 Fixup.LUIdx = LUIdx; 3819 } 3820 3821 // Delete formulae from the new use which are no longer legal. 3822 bool Any = false; 3823 for (size_t i = 0, e = LUThatHas->Formulae.size(); i != e; ++i) { 3824 Formula &F = LUThatHas->Formulae[i]; 3825 if (!isLegalUse(F.AM, 3826 LUThatHas->MinOffset, LUThatHas->MaxOffset, 3827 LUThatHas->Kind, LUThatHas->AccessTy, TLI)) { 3828 DEBUG(dbgs() << " Deleting "; F.print(dbgs()); 3829 dbgs() << '\n'); 3830 LUThatHas->DeleteFormula(F); 3831 --i; 3832 --e; 3833 Any = true; 3834 } 3835 } 3836 if (Any) 3837 LUThatHas->RecomputeRegs(LUThatHas - &Uses.front(), RegUses); 3838 3839 // Delete the old use. 3840 DeleteUse(LU, LUIdx); 3841 --LUIdx; 3842 --NumUses; 3843 break; 3844 } 3845 } 3846 } 3847 } 3848 } 3849 3850 DEBUG(dbgs() << "After pre-selection:\n"; 3851 print_uses(dbgs())); 3852 } 3853} 3854 3855/// NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters - Call 3856/// FilterOutUndesirableDedicatedRegisters again, if necessary, now that 3857/// we've done more filtering, as it may be able to find more formulae to 3858/// eliminate. 3859void LSRInstance::NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters(){ 3860 if (EstimateSearchSpaceComplexity() >= ComplexityLimit) { 3861 DEBUG(dbgs() << "The search space is too complex.\n"); 3862 3863 DEBUG(dbgs() << "Narrowing the search space by re-filtering out " 3864 "undesirable dedicated registers.\n"); 3865 3866 FilterOutUndesirableDedicatedRegisters(); 3867 3868 DEBUG(dbgs() << "After pre-selection:\n"; 3869 print_uses(dbgs())); 3870 } 3871} 3872 3873/// NarrowSearchSpaceByPickingWinnerRegs - Pick a register which seems likely 3874/// to be profitable, and then in any use which has any reference to that 3875/// register, delete all formulae which do not reference that register. 3876void LSRInstance::NarrowSearchSpaceByPickingWinnerRegs() { 3877 // With all other options exhausted, loop until the system is simple 3878 // enough to handle. 3879 SmallPtrSet<const SCEV *, 4> Taken; 3880 while (EstimateSearchSpaceComplexity() >= ComplexityLimit) { 3881 // Ok, we have too many of formulae on our hands to conveniently handle. 3882 // Use a rough heuristic to thin out the list. 3883 DEBUG(dbgs() << "The search space is too complex.\n"); 3884 3885 // Pick the register which is used by the most LSRUses, which is likely 3886 // to be a good reuse register candidate. 3887 const SCEV *Best = 0; 3888 unsigned BestNum = 0; 3889 for (RegUseTracker::const_iterator I = RegUses.begin(), E = RegUses.end(); 3890 I != E; ++I) { 3891 const SCEV *Reg = *I; 3892 if (Taken.count(Reg)) 3893 continue; 3894 if (!Best) 3895 Best = Reg; 3896 else { 3897 unsigned Count = RegUses.getUsedByIndices(Reg).count(); 3898 if (Count > BestNum) { 3899 Best = Reg; 3900 BestNum = Count; 3901 } 3902 } 3903 } 3904 3905 DEBUG(dbgs() << "Narrowing the search space by assuming " << *Best 3906 << " will yield profitable reuse.\n"); 3907 Taken.insert(Best); 3908 3909 // In any use with formulae which references this register, delete formulae 3910 // which don't reference it. 3911 for (size_t LUIdx = 0, NumUses = Uses.size(); LUIdx != NumUses; ++LUIdx) { 3912 LSRUse &LU = Uses[LUIdx]; 3913 if (!LU.Regs.count(Best)) continue; 3914 3915 bool Any = false; 3916 for (size_t i = 0, e = LU.Formulae.size(); i != e; ++i) { 3917 Formula &F = LU.Formulae[i]; 3918 if (!F.referencesReg(Best)) { 3919 DEBUG(dbgs() << " Deleting "; F.print(dbgs()); dbgs() << '\n'); 3920 LU.DeleteFormula(F); 3921 --e; 3922 --i; 3923 Any = true; 3924 assert(e != 0 && "Use has no formulae left! Is Regs inconsistent?"); 3925 continue; 3926 } 3927 } 3928 3929 if (Any) 3930 LU.RecomputeRegs(LUIdx, RegUses); 3931 } 3932 3933 DEBUG(dbgs() << "After pre-selection:\n"; 3934 print_uses(dbgs())); 3935 } 3936} 3937 3938/// NarrowSearchSpaceUsingHeuristics - If there are an extraordinary number of 3939/// formulae to choose from, use some rough heuristics to prune down the number 3940/// of formulae. This keeps the main solver from taking an extraordinary amount 3941/// of time in some worst-case scenarios. 3942void LSRInstance::NarrowSearchSpaceUsingHeuristics() { 3943 NarrowSearchSpaceByDetectingSupersets(); 3944 NarrowSearchSpaceByCollapsingUnrolledCode(); 3945 NarrowSearchSpaceByRefilteringUndesirableDedicatedRegisters(); 3946 NarrowSearchSpaceByPickingWinnerRegs(); 3947} 3948 3949/// SolveRecurse - This is the recursive solver. 3950void LSRInstance::SolveRecurse(SmallVectorImpl<const Formula *> &Solution, 3951 Cost &SolutionCost, 3952 SmallVectorImpl<const Formula *> &Workspace, 3953 const Cost &CurCost, 3954 const SmallPtrSet<const SCEV *, 16> &CurRegs, 3955 DenseSet<const SCEV *> &VisitedRegs) const { 3956 // Some ideas: 3957 // - prune more: 3958 // - use more aggressive filtering 3959 // - sort the formula so that the most profitable solutions are found first 3960 // - sort the uses too 3961 // - search faster: 3962 // - don't compute a cost, and then compare. compare while computing a cost 3963 // and bail early. 3964 // - track register sets with SmallBitVector 3965 3966 const LSRUse &LU = Uses[Workspace.size()]; 3967 3968 // If this use references any register that's already a part of the 3969 // in-progress solution, consider it a requirement that a formula must 3970 // reference that register in order to be considered. This prunes out 3971 // unprofitable searching. 3972 SmallSetVector<const SCEV *, 4> ReqRegs; 3973 for (SmallPtrSet<const SCEV *, 16>::const_iterator I = CurRegs.begin(), 3974 E = CurRegs.end(); I != E; ++I) 3975 if (LU.Regs.count(*I)) 3976 ReqRegs.insert(*I); 3977 3978 SmallPtrSet<const SCEV *, 16> NewRegs; 3979 Cost NewCost; 3980 for (SmallVectorImpl<Formula>::const_iterator I = LU.Formulae.begin(), 3981 E = LU.Formulae.end(); I != E; ++I) { 3982 const Formula &F = *I; 3983 3984 // Ignore formulae which do not use any of the required registers. 3985 bool SatisfiedReqReg = true; 3986 for (SmallSetVector<const SCEV *, 4>::const_iterator J = ReqRegs.begin(), 3987 JE = ReqRegs.end(); J != JE; ++J) { 3988 const SCEV *Reg = *J; 3989 if ((!F.ScaledReg || F.ScaledReg != Reg) && 3990 std::find(F.BaseRegs.begin(), F.BaseRegs.end(), Reg) == 3991 F.BaseRegs.end()) { 3992 SatisfiedReqReg = false; 3993 break; 3994 } 3995 } 3996 if (!SatisfiedReqReg) { 3997 // If none of the formulae satisfied the required registers, then we could 3998 // clear ReqRegs and try again. Currently, we simply give up in this case. 3999 continue; 4000 } 4001 4002 // Evaluate the cost of the current formula. If it's already worse than 4003 // the current best, prune the search at that point. 4004 NewCost = CurCost; 4005 NewRegs = CurRegs; 4006 NewCost.RateFormula(F, NewRegs, VisitedRegs, L, LU.Offsets, SE, DT); 4007 if (NewCost < SolutionCost) { 4008 Workspace.push_back(&F); 4009 if (Workspace.size() != Uses.size()) { 4010 SolveRecurse(Solution, SolutionCost, Workspace, NewCost, 4011 NewRegs, VisitedRegs); 4012 if (F.getNumRegs() == 1 && Workspace.size() == 1) 4013 VisitedRegs.insert(F.ScaledReg ? F.ScaledReg : F.BaseRegs[0]); 4014 } else { 4015 DEBUG(dbgs() << "New best at "; NewCost.print(dbgs()); 4016 dbgs() << ".\n Regs:"; 4017 for (SmallPtrSet<const SCEV *, 16>::const_iterator 4018 I = NewRegs.begin(), E = NewRegs.end(); I != E; ++I) 4019 dbgs() << ' ' << **I; 4020 dbgs() << '\n'); 4021 4022 SolutionCost = NewCost; 4023 Solution = Workspace; 4024 } 4025 Workspace.pop_back(); 4026 } 4027 } 4028} 4029 4030/// Solve - Choose one formula from each use. Return the results in the given 4031/// Solution vector. 4032void LSRInstance::Solve(SmallVectorImpl<const Formula *> &Solution) const { 4033 SmallVector<const Formula *, 8> Workspace; 4034 Cost SolutionCost; 4035 SolutionCost.Loose(); 4036 Cost CurCost; 4037 SmallPtrSet<const SCEV *, 16> CurRegs; 4038 DenseSet<const SCEV *> VisitedRegs; 4039 Workspace.reserve(Uses.size()); 4040 4041 // SolveRecurse does all the work. 4042 SolveRecurse(Solution, SolutionCost, Workspace, CurCost, 4043 CurRegs, VisitedRegs); 4044 if (Solution.empty()) { 4045 DEBUG(dbgs() << "\nNo Satisfactory Solution\n"); 4046 return; 4047 } 4048 4049 // Ok, we've now made all our decisions. 4050 DEBUG(dbgs() << "\n" 4051 "The chosen solution requires "; SolutionCost.print(dbgs()); 4052 dbgs() << ":\n"; 4053 for (size_t i = 0, e = Uses.size(); i != e; ++i) { 4054 dbgs() << " "; 4055 Uses[i].print(dbgs()); 4056 dbgs() << "\n" 4057 " "; 4058 Solution[i]->print(dbgs()); 4059 dbgs() << '\n'; 4060 }); 4061 4062 assert(Solution.size() == Uses.size() && "Malformed solution!"); 4063} 4064 4065/// HoistInsertPosition - Helper for AdjustInsertPositionForExpand. Climb up 4066/// the dominator tree far as we can go while still being dominated by the 4067/// input positions. This helps canonicalize the insert position, which 4068/// encourages sharing. 4069BasicBlock::iterator 4070LSRInstance::HoistInsertPosition(BasicBlock::iterator IP, 4071 const SmallVectorImpl<Instruction *> &Inputs) 4072 const { 4073 for (;;) { 4074 const Loop *IPLoop = LI.getLoopFor(IP->getParent()); 4075 unsigned IPLoopDepth = IPLoop ? IPLoop->getLoopDepth() : 0; 4076 4077 BasicBlock *IDom; 4078 for (DomTreeNode *Rung = DT.getNode(IP->getParent()); ; ) { 4079 if (!Rung) return IP; 4080 Rung = Rung->getIDom(); 4081 if (!Rung) return IP; 4082 IDom = Rung->getBlock(); 4083 4084 // Don't climb into a loop though. 4085 const Loop *IDomLoop = LI.getLoopFor(IDom); 4086 unsigned IDomDepth = IDomLoop ? IDomLoop->getLoopDepth() : 0; 4087 if (IDomDepth <= IPLoopDepth && 4088 (IDomDepth != IPLoopDepth || IDomLoop == IPLoop)) 4089 break; 4090 } 4091 4092 bool AllDominate = true; 4093 Instruction *BetterPos = 0; 4094 Instruction *Tentative = IDom->getTerminator(); 4095 for (SmallVectorImpl<Instruction *>::const_iterator I = Inputs.begin(), 4096 E = Inputs.end(); I != E; ++I) { 4097 Instruction *Inst = *I; 4098 if (Inst == Tentative || !DT.dominates(Inst, Tentative)) { 4099 AllDominate = false; 4100 break; 4101 } 4102 // Attempt to find an insert position in the middle of the block, 4103 // instead of at the end, so that it can be used for other expansions. 4104 if (IDom == Inst->getParent() && 4105 (!BetterPos || DT.dominates(BetterPos, Inst))) 4106 BetterPos = llvm::next(BasicBlock::iterator(Inst)); 4107 } 4108 if (!AllDominate) 4109 break; 4110 if (BetterPos) 4111 IP = BetterPos; 4112 else 4113 IP = Tentative; 4114 } 4115 4116 return IP; 4117} 4118 4119/// AdjustInsertPositionForExpand - Determine an input position which will be 4120/// dominated by the operands and which will dominate the result. 4121BasicBlock::iterator 4122LSRInstance::AdjustInsertPositionForExpand(BasicBlock::iterator LowestIP, 4123 const LSRFixup &LF, 4124 const LSRUse &LU, 4125 SCEVExpander &Rewriter) const { 4126 // Collect some instructions which must be dominated by the 4127 // expanding replacement. These must be dominated by any operands that 4128 // will be required in the expansion. 4129 SmallVector<Instruction *, 4> Inputs; 4130 if (Instruction *I = dyn_cast<Instruction>(LF.OperandValToReplace)) 4131 Inputs.push_back(I); 4132 if (LU.Kind == LSRUse::ICmpZero) 4133 if (Instruction *I = 4134 dyn_cast<Instruction>(cast<ICmpInst>(LF.UserInst)->getOperand(1))) 4135 Inputs.push_back(I); 4136 if (LF.PostIncLoops.count(L)) { 4137 if (LF.isUseFullyOutsideLoop(L)) 4138 Inputs.push_back(L->getLoopLatch()->getTerminator()); 4139 else 4140 Inputs.push_back(IVIncInsertPos); 4141 } 4142 // The expansion must also be dominated by the increment positions of any 4143 // loops it for which it is using post-inc mode. 4144 for (PostIncLoopSet::const_iterator I = LF.PostIncLoops.begin(), 4145 E = LF.PostIncLoops.end(); I != E; ++I) { 4146 const Loop *PIL = *I; 4147 if (PIL == L) continue; 4148 4149 // Be dominated by the loop exit. 4150 SmallVector<BasicBlock *, 4> ExitingBlocks; 4151 PIL->getExitingBlocks(ExitingBlocks); 4152 if (!ExitingBlocks.empty()) { 4153 BasicBlock *BB = ExitingBlocks[0]; 4154 for (unsigned i = 1, e = ExitingBlocks.size(); i != e; ++i) 4155 BB = DT.findNearestCommonDominator(BB, ExitingBlocks[i]); 4156 Inputs.push_back(BB->getTerminator()); 4157 } 4158 } 4159 4160 assert(!isa<PHINode>(LowestIP) && !isa<LandingPadInst>(LowestIP) 4161 && !isa<DbgInfoIntrinsic>(LowestIP) && 4162 "Insertion point must be a normal instruction"); 4163 4164 // Then, climb up the immediate dominator tree as far as we can go while 4165 // still being dominated by the input positions. 4166 BasicBlock::iterator IP = HoistInsertPosition(LowestIP, Inputs); 4167 4168 // Don't insert instructions before PHI nodes. 4169 while (isa<PHINode>(IP)) ++IP; 4170 4171 // Ignore landingpad instructions. 4172 while (isa<LandingPadInst>(IP)) ++IP; 4173 4174 // Ignore debug intrinsics. 4175 while (isa<DbgInfoIntrinsic>(IP)) ++IP; 4176 4177 // Set IP below instructions recently inserted by SCEVExpander. This keeps the 4178 // IP consistent across expansions and allows the previously inserted 4179 // instructions to be reused by subsequent expansion. 4180 while (Rewriter.isInsertedInstruction(IP) && IP != LowestIP) ++IP; 4181 4182 return IP; 4183} 4184 4185/// Expand - Emit instructions for the leading candidate expression for this 4186/// LSRUse (this is called "expanding"). 4187Value *LSRInstance::Expand(const LSRFixup &LF, 4188 const Formula &F, 4189 BasicBlock::iterator IP, 4190 SCEVExpander &Rewriter, 4191 SmallVectorImpl<WeakVH> &DeadInsts) const { 4192 const LSRUse &LU = Uses[LF.LUIdx]; 4193 4194 // Determine an input position which will be dominated by the operands and 4195 // which will dominate the result. 4196 IP = AdjustInsertPositionForExpand(IP, LF, LU, Rewriter); 4197 4198 // Inform the Rewriter if we have a post-increment use, so that it can 4199 // perform an advantageous expansion. 4200 Rewriter.setPostInc(LF.PostIncLoops); 4201 4202 // This is the type that the user actually needs. 4203 Type *OpTy = LF.OperandValToReplace->getType(); 4204 // This will be the type that we'll initially expand to. 4205 Type *Ty = F.getType(); 4206 if (!Ty) 4207 // No type known; just expand directly to the ultimate type. 4208 Ty = OpTy; 4209 else if (SE.getEffectiveSCEVType(Ty) == SE.getEffectiveSCEVType(OpTy)) 4210 // Expand directly to the ultimate type if it's the right size. 4211 Ty = OpTy; 4212 // This is the type to do integer arithmetic in. 4213 Type *IntTy = SE.getEffectiveSCEVType(Ty); 4214 4215 // Build up a list of operands to add together to form the full base. 4216 SmallVector<const SCEV *, 8> Ops; 4217 4218 // Expand the BaseRegs portion. 4219 for (SmallVectorImpl<const SCEV *>::const_iterator I = F.BaseRegs.begin(), 4220 E = F.BaseRegs.end(); I != E; ++I) { 4221 const SCEV *Reg = *I; 4222 assert(!Reg->isZero() && "Zero allocated in a base register!"); 4223 4224 // If we're expanding for a post-inc user, make the post-inc adjustment. 4225 PostIncLoopSet &Loops = const_cast<PostIncLoopSet &>(LF.PostIncLoops); 4226 Reg = TransformForPostIncUse(Denormalize, Reg, 4227 LF.UserInst, LF.OperandValToReplace, 4228 Loops, SE, DT); 4229 4230 Ops.push_back(SE.getUnknown(Rewriter.expandCodeFor(Reg, 0, IP))); 4231 } 4232 4233 // Flush the operand list to suppress SCEVExpander hoisting. 4234 if (!Ops.empty()) { 4235 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP); 4236 Ops.clear(); 4237 Ops.push_back(SE.getUnknown(FullV)); 4238 } 4239 4240 // Expand the ScaledReg portion. 4241 Value *ICmpScaledV = 0; 4242 if (F.AM.Scale != 0) { 4243 const SCEV *ScaledS = F.ScaledReg; 4244 4245 // If we're expanding for a post-inc user, make the post-inc adjustment. 4246 PostIncLoopSet &Loops = const_cast<PostIncLoopSet &>(LF.PostIncLoops); 4247 ScaledS = TransformForPostIncUse(Denormalize, ScaledS, 4248 LF.UserInst, LF.OperandValToReplace, 4249 Loops, SE, DT); 4250 4251 if (LU.Kind == LSRUse::ICmpZero) { 4252 // An interesting way of "folding" with an icmp is to use a negated 4253 // scale, which we'll implement by inserting it into the other operand 4254 // of the icmp. 4255 assert(F.AM.Scale == -1 && 4256 "The only scale supported by ICmpZero uses is -1!"); 4257 ICmpScaledV = Rewriter.expandCodeFor(ScaledS, 0, IP); 4258 } else { 4259 // Otherwise just expand the scaled register and an explicit scale, 4260 // which is expected to be matched as part of the address. 4261 ScaledS = SE.getUnknown(Rewriter.expandCodeFor(ScaledS, 0, IP)); 4262 ScaledS = SE.getMulExpr(ScaledS, 4263 SE.getConstant(ScaledS->getType(), F.AM.Scale)); 4264 Ops.push_back(ScaledS); 4265 4266 // Flush the operand list to suppress SCEVExpander hoisting. 4267 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP); 4268 Ops.clear(); 4269 Ops.push_back(SE.getUnknown(FullV)); 4270 } 4271 } 4272 4273 // Expand the GV portion. 4274 if (F.AM.BaseGV) { 4275 Ops.push_back(SE.getUnknown(F.AM.BaseGV)); 4276 4277 // Flush the operand list to suppress SCEVExpander hoisting. 4278 Value *FullV = Rewriter.expandCodeFor(SE.getAddExpr(Ops), Ty, IP); 4279 Ops.clear(); 4280 Ops.push_back(SE.getUnknown(FullV)); 4281 } 4282 4283 // Expand the immediate portion. 4284 int64_t Offset = (uint64_t)F.AM.BaseOffs + LF.Offset; 4285 if (Offset != 0) { 4286 if (LU.Kind == LSRUse::ICmpZero) { 4287 // The other interesting way of "folding" with an ICmpZero is to use a 4288 // negated immediate. 4289 if (!ICmpScaledV) 4290 ICmpScaledV = ConstantInt::get(IntTy, -(uint64_t)Offset); 4291 else { 4292 Ops.push_back(SE.getUnknown(ICmpScaledV)); 4293 ICmpScaledV = ConstantInt::get(IntTy, Offset); 4294 } 4295 } else { 4296 // Just add the immediate values. These again are expected to be matched 4297 // as part of the address. 4298 Ops.push_back(SE.getUnknown(ConstantInt::getSigned(IntTy, Offset))); 4299 } 4300 } 4301 4302 // Expand the unfolded offset portion. 4303 int64_t UnfoldedOffset = F.UnfoldedOffset; 4304 if (UnfoldedOffset != 0) { 4305 // Just add the immediate values. 4306 Ops.push_back(SE.getUnknown(ConstantInt::getSigned(IntTy, 4307 UnfoldedOffset))); 4308 } 4309 4310 // Emit instructions summing all the operands. 4311 const SCEV *FullS = Ops.empty() ? 4312 SE.getConstant(IntTy, 0) : 4313 SE.getAddExpr(Ops); 4314 Value *FullV = Rewriter.expandCodeFor(FullS, Ty, IP); 4315 4316 // We're done expanding now, so reset the rewriter. 4317 Rewriter.clearPostInc(); 4318 4319 // An ICmpZero Formula represents an ICmp which we're handling as a 4320 // comparison against zero. Now that we've expanded an expression for that 4321 // form, update the ICmp's other operand. 4322 if (LU.Kind == LSRUse::ICmpZero) { 4323 ICmpInst *CI = cast<ICmpInst>(LF.UserInst); 4324 DeadInsts.push_back(CI->getOperand(1)); 4325 assert(!F.AM.BaseGV && "ICmp does not support folding a global value and " 4326 "a scale at the same time!"); 4327 if (F.AM.Scale == -1) { 4328 if (ICmpScaledV->getType() != OpTy) { 4329 Instruction *Cast = 4330 CastInst::Create(CastInst::getCastOpcode(ICmpScaledV, false, 4331 OpTy, false), 4332 ICmpScaledV, OpTy, "tmp", CI); 4333 ICmpScaledV = Cast; 4334 } 4335 CI->setOperand(1, ICmpScaledV); 4336 } else { 4337 assert(F.AM.Scale == 0 && 4338 "ICmp does not support folding a global value and " 4339 "a scale at the same time!"); 4340 Constant *C = ConstantInt::getSigned(SE.getEffectiveSCEVType(OpTy), 4341 -(uint64_t)Offset); 4342 if (C->getType() != OpTy) 4343 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false, 4344 OpTy, false), 4345 C, OpTy); 4346 4347 CI->setOperand(1, C); 4348 } 4349 } 4350 4351 return FullV; 4352} 4353 4354/// RewriteForPHI - Helper for Rewrite. PHI nodes are special because the use 4355/// of their operands effectively happens in their predecessor blocks, so the 4356/// expression may need to be expanded in multiple places. 4357void LSRInstance::RewriteForPHI(PHINode *PN, 4358 const LSRFixup &LF, 4359 const Formula &F, 4360 SCEVExpander &Rewriter, 4361 SmallVectorImpl<WeakVH> &DeadInsts, 4362 Pass *P) const { 4363 DenseMap<BasicBlock *, Value *> Inserted; 4364 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) 4365 if (PN->getIncomingValue(i) == LF.OperandValToReplace) { 4366 BasicBlock *BB = PN->getIncomingBlock(i); 4367 4368 // If this is a critical edge, split the edge so that we do not insert 4369 // the code on all predecessor/successor paths. We do this unless this 4370 // is the canonical backedge for this loop, which complicates post-inc 4371 // users. 4372 if (e != 1 && BB->getTerminator()->getNumSuccessors() > 1 && 4373 !isa<IndirectBrInst>(BB->getTerminator())) { 4374 BasicBlock *Parent = PN->getParent(); 4375 Loop *PNLoop = LI.getLoopFor(Parent); 4376 if (!PNLoop || Parent != PNLoop->getHeader()) { 4377 // Split the critical edge. 4378 BasicBlock *NewBB = 0; 4379 if (!Parent->isLandingPad()) { 4380 NewBB = SplitCriticalEdge(BB, Parent, P, 4381 /*MergeIdenticalEdges=*/true, 4382 /*DontDeleteUselessPhis=*/true); 4383 } else { 4384 SmallVector<BasicBlock*, 2> NewBBs; 4385 SplitLandingPadPredecessors(Parent, BB, "", "", P, NewBBs); 4386 NewBB = NewBBs[0]; 4387 } 4388 4389 // If PN is outside of the loop and BB is in the loop, we want to 4390 // move the block to be immediately before the PHI block, not 4391 // immediately after BB. 4392 if (L->contains(BB) && !L->contains(PN)) 4393 NewBB->moveBefore(PN->getParent()); 4394 4395 // Splitting the edge can reduce the number of PHI entries we have. 4396 e = PN->getNumIncomingValues(); 4397 BB = NewBB; 4398 i = PN->getBasicBlockIndex(BB); 4399 } 4400 } 4401 4402 std::pair<DenseMap<BasicBlock *, Value *>::iterator, bool> Pair = 4403 Inserted.insert(std::make_pair(BB, static_cast<Value *>(0))); 4404 if (!Pair.second) 4405 PN->setIncomingValue(i, Pair.first->second); 4406 else { 4407 Value *FullV = Expand(LF, F, BB->getTerminator(), Rewriter, DeadInsts); 4408 4409 // If this is reuse-by-noop-cast, insert the noop cast. 4410 Type *OpTy = LF.OperandValToReplace->getType(); 4411 if (FullV->getType() != OpTy) 4412 FullV = 4413 CastInst::Create(CastInst::getCastOpcode(FullV, false, 4414 OpTy, false), 4415 FullV, LF.OperandValToReplace->getType(), 4416 "tmp", BB->getTerminator()); 4417 4418 PN->setIncomingValue(i, FullV); 4419 Pair.first->second = FullV; 4420 } 4421 } 4422} 4423 4424/// Rewrite - Emit instructions for the leading candidate expression for this 4425/// LSRUse (this is called "expanding"), and update the UserInst to reference 4426/// the newly expanded value. 4427void LSRInstance::Rewrite(const LSRFixup &LF, 4428 const Formula &F, 4429 SCEVExpander &Rewriter, 4430 SmallVectorImpl<WeakVH> &DeadInsts, 4431 Pass *P) const { 4432 // First, find an insertion point that dominates UserInst. For PHI nodes, 4433 // find the nearest block which dominates all the relevant uses. 4434 if (PHINode *PN = dyn_cast<PHINode>(LF.UserInst)) { 4435 RewriteForPHI(PN, LF, F, Rewriter, DeadInsts, P); 4436 } else { 4437 Value *FullV = Expand(LF, F, LF.UserInst, Rewriter, DeadInsts); 4438 4439 // If this is reuse-by-noop-cast, insert the noop cast. 4440 Type *OpTy = LF.OperandValToReplace->getType(); 4441 if (FullV->getType() != OpTy) { 4442 Instruction *Cast = 4443 CastInst::Create(CastInst::getCastOpcode(FullV, false, OpTy, false), 4444 FullV, OpTy, "tmp", LF.UserInst); 4445 FullV = Cast; 4446 } 4447 4448 // Update the user. ICmpZero is handled specially here (for now) because 4449 // Expand may have updated one of the operands of the icmp already, and 4450 // its new value may happen to be equal to LF.OperandValToReplace, in 4451 // which case doing replaceUsesOfWith leads to replacing both operands 4452 // with the same value. TODO: Reorganize this. 4453 if (Uses[LF.LUIdx].Kind == LSRUse::ICmpZero) 4454 LF.UserInst->setOperand(0, FullV); 4455 else 4456 LF.UserInst->replaceUsesOfWith(LF.OperandValToReplace, FullV); 4457 } 4458 4459 DeadInsts.push_back(LF.OperandValToReplace); 4460} 4461 4462/// ImplementSolution - Rewrite all the fixup locations with new values, 4463/// following the chosen solution. 4464void 4465LSRInstance::ImplementSolution(const SmallVectorImpl<const Formula *> &Solution, 4466 Pass *P) { 4467 // Keep track of instructions we may have made dead, so that 4468 // we can remove them after we are done working. 4469 SmallVector<WeakVH, 16> DeadInsts; 4470 4471 SCEVExpander Rewriter(SE, "lsr"); 4472#ifndef NDEBUG 4473 Rewriter.setDebugType(DEBUG_TYPE); 4474#endif 4475 Rewriter.disableCanonicalMode(); 4476 Rewriter.enableLSRMode(); 4477 Rewriter.setIVIncInsertPos(L, IVIncInsertPos); 4478 4479 // Mark phi nodes that terminate chains so the expander tries to reuse them. 4480 for (SmallVectorImpl<IVChain>::const_iterator ChainI = IVChainVec.begin(), 4481 ChainE = IVChainVec.end(); ChainI != ChainE; ++ChainI) { 4482 if (PHINode *PN = dyn_cast<PHINode>(ChainI->back().UserInst)) 4483 Rewriter.setChainedPhi(PN); 4484 } 4485 4486 // Expand the new value definitions and update the users. 4487 for (SmallVectorImpl<LSRFixup>::const_iterator I = Fixups.begin(), 4488 E = Fixups.end(); I != E; ++I) { 4489 const LSRFixup &Fixup = *I; 4490 4491 Rewrite(Fixup, *Solution[Fixup.LUIdx], Rewriter, DeadInsts, P); 4492 4493 Changed = true; 4494 } 4495 4496 for (SmallVectorImpl<IVChain>::const_iterator ChainI = IVChainVec.begin(), 4497 ChainE = IVChainVec.end(); ChainI != ChainE; ++ChainI) { 4498 GenerateIVChain(*ChainI, Rewriter, DeadInsts); 4499 Changed = true; 4500 } 4501 // Clean up after ourselves. This must be done before deleting any 4502 // instructions. 4503 Rewriter.clear(); 4504 4505 Changed |= DeleteTriviallyDeadInstructions(DeadInsts); 4506} 4507 4508LSRInstance::LSRInstance(const TargetLowering *tli, Loop *l, Pass *P) 4509 : IU(P->getAnalysis<IVUsers>()), 4510 SE(P->getAnalysis<ScalarEvolution>()), 4511 DT(P->getAnalysis<DominatorTree>()), 4512 LI(P->getAnalysis<LoopInfo>()), 4513 TLI(tli), L(l), Changed(false), IVIncInsertPos(0) { 4514 4515 // If LoopSimplify form is not available, stay out of trouble. 4516 if (!L->isLoopSimplifyForm()) 4517 return; 4518 4519 // If there's no interesting work to be done, bail early. 4520 if (IU.empty()) return; 4521 4522#ifndef NDEBUG 4523 // All dominating loops must have preheaders, or SCEVExpander may not be able 4524 // to materialize an AddRecExpr whose Start is an outer AddRecExpr. 4525 // 4526 // IVUsers analysis should only create users that are dominated by simple loop 4527 // headers. Since this loop should dominate all of its users, its user list 4528 // should be empty if this loop itself is not within a simple loop nest. 4529 for (DomTreeNode *Rung = DT.getNode(L->getLoopPreheader()); 4530 Rung; Rung = Rung->getIDom()) { 4531 BasicBlock *BB = Rung->getBlock(); 4532 const Loop *DomLoop = LI.getLoopFor(BB); 4533 if (DomLoop && DomLoop->getHeader() == BB) { 4534 assert(DomLoop->getLoopPreheader() && "LSR needs a simplified loop nest"); 4535 } 4536 } 4537#endif // DEBUG 4538 4539 DEBUG(dbgs() << "\nLSR on loop "; 4540 WriteAsOperand(dbgs(), L->getHeader(), /*PrintType=*/false); 4541 dbgs() << ":\n"); 4542 4543 // First, perform some low-level loop optimizations. 4544 OptimizeShadowIV(); 4545 OptimizeLoopTermCond(); 4546 4547 // If loop preparation eliminates all interesting IV users, bail. 4548 if (IU.empty()) return; 4549 4550 // Skip nested loops until we can model them better with formulae. 4551 if (!L->empty()) { 4552 DEBUG(dbgs() << "LSR skipping outer loop " << *L << "\n"); 4553 return; 4554 } 4555 4556 // Start collecting data and preparing for the solver. 4557 CollectChains(); 4558 CollectInterestingTypesAndFactors(); 4559 CollectFixupsAndInitialFormulae(); 4560 CollectLoopInvariantFixupsAndFormulae(); 4561 4562 assert(!Uses.empty() && "IVUsers reported at least one use"); 4563 DEBUG(dbgs() << "LSR found " << Uses.size() << " uses:\n"; 4564 print_uses(dbgs())); 4565 4566 // Now use the reuse data to generate a bunch of interesting ways 4567 // to formulate the values needed for the uses. 4568 GenerateAllReuseFormulae(); 4569 4570 FilterOutUndesirableDedicatedRegisters(); 4571 NarrowSearchSpaceUsingHeuristics(); 4572 4573 SmallVector<const Formula *, 8> Solution; 4574 Solve(Solution); 4575 4576 // Release memory that is no longer needed. 4577 Factors.clear(); 4578 Types.clear(); 4579 RegUses.clear(); 4580 4581 if (Solution.empty()) 4582 return; 4583 4584#ifndef NDEBUG 4585 // Formulae should be legal. 4586 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(), 4587 E = Uses.end(); I != E; ++I) { 4588 const LSRUse &LU = *I; 4589 for (SmallVectorImpl<Formula>::const_iterator J = LU.Formulae.begin(), 4590 JE = LU.Formulae.end(); J != JE; ++J) 4591 assert(isLegalUse(J->AM, LU.MinOffset, LU.MaxOffset, 4592 LU.Kind, LU.AccessTy, TLI) && 4593 "Illegal formula generated!"); 4594 }; 4595#endif 4596 4597 // Now that we've decided what we want, make it so. 4598 ImplementSolution(Solution, P); 4599} 4600 4601void LSRInstance::print_factors_and_types(raw_ostream &OS) const { 4602 if (Factors.empty() && Types.empty()) return; 4603 4604 OS << "LSR has identified the following interesting factors and types: "; 4605 bool First = true; 4606 4607 for (SmallSetVector<int64_t, 8>::const_iterator 4608 I = Factors.begin(), E = Factors.end(); I != E; ++I) { 4609 if (!First) OS << ", "; 4610 First = false; 4611 OS << '*' << *I; 4612 } 4613 4614 for (SmallSetVector<Type *, 4>::const_iterator 4615 I = Types.begin(), E = Types.end(); I != E; ++I) { 4616 if (!First) OS << ", "; 4617 First = false; 4618 OS << '(' << **I << ')'; 4619 } 4620 OS << '\n'; 4621} 4622 4623void LSRInstance::print_fixups(raw_ostream &OS) const { 4624 OS << "LSR is examining the following fixup sites:\n"; 4625 for (SmallVectorImpl<LSRFixup>::const_iterator I = Fixups.begin(), 4626 E = Fixups.end(); I != E; ++I) { 4627 dbgs() << " "; 4628 I->print(OS); 4629 OS << '\n'; 4630 } 4631} 4632 4633void LSRInstance::print_uses(raw_ostream &OS) const { 4634 OS << "LSR is examining the following uses:\n"; 4635 for (SmallVectorImpl<LSRUse>::const_iterator I = Uses.begin(), 4636 E = Uses.end(); I != E; ++I) { 4637 const LSRUse &LU = *I; 4638 dbgs() << " "; 4639 LU.print(OS); 4640 OS << '\n'; 4641 for (SmallVectorImpl<Formula>::const_iterator J = LU.Formulae.begin(), 4642 JE = LU.Formulae.end(); J != JE; ++J) { 4643 OS << " "; 4644 J->print(OS); 4645 OS << '\n'; 4646 } 4647 } 4648} 4649 4650void LSRInstance::print(raw_ostream &OS) const { 4651 print_factors_and_types(OS); 4652 print_fixups(OS); 4653 print_uses(OS); 4654} 4655 4656void LSRInstance::dump() const { 4657 print(errs()); errs() << '\n'; 4658} 4659 4660namespace { 4661 4662class LoopStrengthReduce : public LoopPass { 4663 /// TLI - Keep a pointer of a TargetLowering to consult for determining 4664 /// transformation profitability. 4665 const TargetLowering *const TLI; 4666 4667public: 4668 static char ID; // Pass ID, replacement for typeid 4669 explicit LoopStrengthReduce(const TargetLowering *tli = 0); 4670 4671private: 4672 bool runOnLoop(Loop *L, LPPassManager &LPM); 4673 void getAnalysisUsage(AnalysisUsage &AU) const; 4674}; 4675 4676} 4677 4678char LoopStrengthReduce::ID = 0; 4679INITIALIZE_PASS_BEGIN(LoopStrengthReduce, "loop-reduce", 4680 "Loop Strength Reduction", false, false) 4681INITIALIZE_PASS_DEPENDENCY(DominatorTree) 4682INITIALIZE_PASS_DEPENDENCY(ScalarEvolution) 4683INITIALIZE_PASS_DEPENDENCY(IVUsers) 4684INITIALIZE_PASS_DEPENDENCY(LoopInfo) 4685INITIALIZE_PASS_DEPENDENCY(LoopSimplify) 4686INITIALIZE_PASS_END(LoopStrengthReduce, "loop-reduce", 4687 "Loop Strength Reduction", false, false) 4688 4689 4690Pass *llvm::createLoopStrengthReducePass(const TargetLowering *TLI) { 4691 return new LoopStrengthReduce(TLI); 4692} 4693 4694LoopStrengthReduce::LoopStrengthReduce(const TargetLowering *tli) 4695 : LoopPass(ID), TLI(tli) { 4696 initializeLoopStrengthReducePass(*PassRegistry::getPassRegistry()); 4697 } 4698 4699void LoopStrengthReduce::getAnalysisUsage(AnalysisUsage &AU) const { 4700 // We split critical edges, so we change the CFG. However, we do update 4701 // many analyses if they are around. 4702 AU.addPreservedID(LoopSimplifyID); 4703 4704 AU.addRequired<LoopInfo>(); 4705 AU.addPreserved<LoopInfo>(); 4706 AU.addRequiredID(LoopSimplifyID); 4707 AU.addRequired<DominatorTree>(); 4708 AU.addPreserved<DominatorTree>(); 4709 AU.addRequired<ScalarEvolution>(); 4710 AU.addPreserved<ScalarEvolution>(); 4711 // Requiring LoopSimplify a second time here prevents IVUsers from running 4712 // twice, since LoopSimplify was invalidated by running ScalarEvolution. 4713 AU.addRequiredID(LoopSimplifyID); 4714 AU.addRequired<IVUsers>(); 4715 AU.addPreserved<IVUsers>(); 4716} 4717 4718bool LoopStrengthReduce::runOnLoop(Loop *L, LPPassManager & /*LPM*/) { 4719 bool Changed = false; 4720 4721 // Run the main LSR transformation. 4722 Changed |= LSRInstance(TLI, L, this).getChanged(); 4723 4724 // Remove any extra phis created by processing inner loops. 4725 Changed |= DeleteDeadPHIs(L->getHeader()); 4726 if (EnablePhiElim) { 4727 SmallVector<WeakVH, 16> DeadInsts; 4728 SCEVExpander Rewriter(getAnalysis<ScalarEvolution>(), "lsr"); 4729#ifndef NDEBUG 4730 Rewriter.setDebugType(DEBUG_TYPE); 4731#endif 4732 unsigned numFolded = Rewriter. 4733 replaceCongruentIVs(L, &getAnalysis<DominatorTree>(), DeadInsts, TLI); 4734 if (numFolded) { 4735 Changed = true; 4736 DeleteTriviallyDeadInstructions(DeadInsts); 4737 DeleteDeadPHIs(L->getHeader()); 4738 } 4739 } 4740 return Changed; 4741} 4742