1 //===- ScalarEvolution.cpp - Scalar Evolution Analysis ----------*- C++ -*-===// 2 // 3 // The LLVM Compiler Infrastructure 4 // 5 // This file was developed by the LLVM research group and is distributed under 6 // the University of Illinois Open Source License. See LICENSE.TXT for details. 7 // 8 //===----------------------------------------------------------------------===// 9 // 10 // This file contains the implementation of the scalar evolution analysis 11 // engine, which is used primarily to analyze expressions involving induction 12 // variables in loops. 13 // 14 // There are several aspects to this library. First is the representation of 15 // scalar expressions, which are represented as subclasses of the SCEV class. 16 // These classes are used to represent certain types of subexpressions that we 17 // can handle. These classes are reference counted, managed by the SCEVHandle 18 // class. We only create one SCEV of a particular shape, so pointer-comparisons 19 // for equality are legal. 20 // 21 // One important aspect of the SCEV objects is that they are never cyclic, even 22 // if there is a cycle in the dataflow for an expression (ie, a PHI node). If 23 // the PHI node is one of the idioms that we can represent (e.g., a polynomial 24 // recurrence) then we represent it directly as a recurrence node, otherwise we 25 // represent it as a SCEVUnknown node. 26 // 27 // In addition to being able to represent expressions of various types, we also 28 // have folders that are used to build the *canonical* representation for a 29 // particular expression. These folders are capable of using a variety of 30 // rewrite rules to simplify the expressions. 31 // 32 // Once the folders are defined, we can implement the more interesting 33 // higher-level code, such as the code that recognizes PHI nodes of various 34 // types, computes the execution count of a loop, etc. 35 // 36 // TODO: We should use these routines and value representations to implement 37 // dependence analysis! 38 // 39 //===----------------------------------------------------------------------===// 40 // 41 // There are several good references for the techniques used in this analysis. 42 // 43 // Chains of recurrences -- a method to expedite the evaluation 44 // of closed-form functions 45 // Olaf Bachmann, Paul S. Wang, Eugene V. Zima 46 // 47 // On computational properties of chains of recurrences 48 // Eugene V. Zima 49 // 50 // Symbolic Evaluation of Chains of Recurrences for Loop Optimization 51 // Robert A. van Engelen 52 // 53 // Efficient Symbolic Analysis for Optimizing Compilers 54 // Robert A. van Engelen 55 // 56 // Using the chains of recurrences algebra for data dependence testing and 57 // induction variable substitution 58 // MS Thesis, Johnie Birch 59 // 60 //===----------------------------------------------------------------------===// 61 62 #include "llvm/Analysis/ScalarEvolutionExpressions.h" 63 #include "llvm/Constants.h" 64 #include "llvm/DerivedTypes.h" 65 #include "llvm/GlobalVariable.h" 66 #include "llvm/Instructions.h" 67 #include "llvm/Analysis/ConstantFolding.h" 68 #include "llvm/Analysis/LoopInfo.h" 69 #include "llvm/Assembly/Writer.h" 70 #include "llvm/Transforms/Scalar.h" 71 #include "llvm/Support/CFG.h" 72 #include "llvm/Support/CommandLine.h" 73 #include "llvm/Support/Compiler.h" 74 #include "llvm/Support/ConstantRange.h" 75 #include "llvm/Support/InstIterator.h" 76 #include "llvm/Support/ManagedStatic.h" 77 #include "llvm/Support/Streams.h" 78 #include "llvm/ADT/Statistic.h" 79 #include <ostream> 80 #include <algorithm> 81 #include <cmath> 82 using namespace llvm; 83 84 namespace { 85 RegisterPass<ScalarEvolution> 86 R("scalar-evolution", "Scalar Evolution Analysis"); 87 88 Statistic 89 NumBruteForceEvaluations("scalar-evolution", 90 "Number of brute force evaluations needed to " 91 "calculate high-order polynomial exit values"); 92 Statistic 93 NumArrayLenItCounts("scalar-evolution", 94 "Number of trip counts computed with array length"); 95 Statistic 96 NumTripCountsComputed("scalar-evolution", 97 "Number of loops with predictable loop counts"); 98 Statistic 99 NumTripCountsNotComputed("scalar-evolution", 100 "Number of loops without predictable loop counts"); 101 Statistic 102 NumBruteForceTripCountsComputed("scalar-evolution", 103 "Number of loops with trip counts computed by force"); 104 105 cl::opt<unsigned> 106 MaxBruteForceIterations("scalar-evolution-max-iterations", cl::ReallyHidden, 107 cl::desc("Maximum number of iterations SCEV will " 108 "symbolically execute a constant derived loop"), 109 cl::init(100)); 110 } 111 112 //===----------------------------------------------------------------------===// 113 // SCEV class definitions 114 //===----------------------------------------------------------------------===// 115 116 //===----------------------------------------------------------------------===// 117 // Implementation of the SCEV class. 118 // 119 SCEV::~SCEV() {} 120 void SCEV::dump() const { 121 print(llvm_cerr); 122 } 123 124 /// getValueRange - Return the tightest constant bounds that this value is 125 /// known to have. This method is only valid on integer SCEV objects. 126 ConstantRange SCEV::getValueRange() const { 127 const Type *Ty = getType(); 128 assert(Ty->isInteger() && "Can't get range for a non-integer SCEV!"); 129 Ty = Ty->getUnsignedVersion(); 130 // Default to a full range if no better information is available. 131 return ConstantRange(getType()); 132 } 133 134 135 SCEVCouldNotCompute::SCEVCouldNotCompute() : SCEV(scCouldNotCompute) {} 136 137 bool SCEVCouldNotCompute::isLoopInvariant(const Loop *L) const { 138 assert(0 && "Attempt to use a SCEVCouldNotCompute object!"); 139 return false; 140 } 141 142 const Type *SCEVCouldNotCompute::getType() const { 143 assert(0 && "Attempt to use a SCEVCouldNotCompute object!"); 144 return 0; 145 } 146 147 bool SCEVCouldNotCompute::hasComputableLoopEvolution(const Loop *L) const { 148 assert(0 && "Attempt to use a SCEVCouldNotCompute object!"); 149 return false; 150 } 151 152 SCEVHandle SCEVCouldNotCompute:: 153 replaceSymbolicValuesWithConcrete(const SCEVHandle &Sym, 154 const SCEVHandle &Conc) const { 155 return this; 156 } 157 158 void SCEVCouldNotCompute::print(std::ostream &OS) const { 159 OS << "***COULDNOTCOMPUTE***"; 160 } 161 162 bool SCEVCouldNotCompute::classof(const SCEV *S) { 163 return S->getSCEVType() == scCouldNotCompute; 164 } 165 166 167 // SCEVConstants - Only allow the creation of one SCEVConstant for any 168 // particular value. Don't use a SCEVHandle here, or else the object will 169 // never be deleted! 170 static ManagedStatic<std::map<ConstantInt*, SCEVConstant*> > SCEVConstants; 171 172 173 SCEVConstant::~SCEVConstant() { 174 SCEVConstants->erase(V); 175 } 176 177 SCEVHandle SCEVConstant::get(ConstantInt *V) { 178 // Make sure that SCEVConstant instances are all unsigned. 179 if (V->getType()->isSigned()) { 180 const Type *NewTy = V->getType()->getUnsignedVersion(); 181 V = cast<ConstantInt>( 182 ConstantExpr::getBitCast(V, NewTy)); 183 } 184 185 SCEVConstant *&R = (*SCEVConstants)[V]; 186 if (R == 0) R = new SCEVConstant(V); 187 return R; 188 } 189 190 ConstantRange SCEVConstant::getValueRange() const { 191 return ConstantRange(V); 192 } 193 194 const Type *SCEVConstant::getType() const { return V->getType(); } 195 196 void SCEVConstant::print(std::ostream &OS) const { 197 WriteAsOperand(OS, V, false); 198 } 199 200 // SCEVTruncates - Only allow the creation of one SCEVTruncateExpr for any 201 // particular input. Don't use a SCEVHandle here, or else the object will 202 // never be deleted! 203 static ManagedStatic<std::map<std::pair<SCEV*, const Type*>, 204 SCEVTruncateExpr*> > SCEVTruncates; 205 206 SCEVTruncateExpr::SCEVTruncateExpr(const SCEVHandle &op, const Type *ty) 207 : SCEV(scTruncate), Op(op), Ty(ty) { 208 assert(Op->getType()->isInteger() && Ty->isInteger() && 209 "Cannot truncate non-integer value!"); 210 assert(Op->getType()->getPrimitiveSize() > Ty->getPrimitiveSize() && 211 "This is not a truncating conversion!"); 212 } 213 214 SCEVTruncateExpr::~SCEVTruncateExpr() { 215 SCEVTruncates->erase(std::make_pair(Op, Ty)); 216 } 217 218 ConstantRange SCEVTruncateExpr::getValueRange() const { 219 return getOperand()->getValueRange().truncate(getType()); 220 } 221 222 void SCEVTruncateExpr::print(std::ostream &OS) const { 223 OS << "(truncate " << *Op << " to " << *Ty << ")"; 224 } 225 226 // SCEVZeroExtends - Only allow the creation of one SCEVZeroExtendExpr for any 227 // particular input. Don't use a SCEVHandle here, or else the object will never 228 // be deleted! 229 static ManagedStatic<std::map<std::pair<SCEV*, const Type*>, 230 SCEVZeroExtendExpr*> > SCEVZeroExtends; 231 232 SCEVZeroExtendExpr::SCEVZeroExtendExpr(const SCEVHandle &op, const Type *ty) 233 : SCEV(scZeroExtend), Op(op), Ty(ty) { 234 assert(Op->getType()->isInteger() && Ty->isInteger() && 235 "Cannot zero extend non-integer value!"); 236 assert(Op->getType()->getPrimitiveSize() < Ty->getPrimitiveSize() && 237 "This is not an extending conversion!"); 238 } 239 240 SCEVZeroExtendExpr::~SCEVZeroExtendExpr() { 241 SCEVZeroExtends->erase(std::make_pair(Op, Ty)); 242 } 243 244 ConstantRange SCEVZeroExtendExpr::getValueRange() const { 245 return getOperand()->getValueRange().zeroExtend(getType()); 246 } 247 248 void SCEVZeroExtendExpr::print(std::ostream &OS) const { 249 OS << "(zeroextend " << *Op << " to " << *Ty << ")"; 250 } 251 252 // SCEVCommExprs - Only allow the creation of one SCEVCommutativeExpr for any 253 // particular input. Don't use a SCEVHandle here, or else the object will never 254 // be deleted! 255 static ManagedStatic<std::map<std::pair<unsigned, std::vector<SCEV*> >, 256 SCEVCommutativeExpr*> > SCEVCommExprs; 257 258 SCEVCommutativeExpr::~SCEVCommutativeExpr() { 259 SCEVCommExprs->erase(std::make_pair(getSCEVType(), 260 std::vector<SCEV*>(Operands.begin(), 261 Operands.end()))); 262 } 263 264 void SCEVCommutativeExpr::print(std::ostream &OS) const { 265 assert(Operands.size() > 1 && "This plus expr shouldn't exist!"); 266 const char *OpStr = getOperationStr(); 267 OS << "(" << *Operands[0]; 268 for (unsigned i = 1, e = Operands.size(); i != e; ++i) 269 OS << OpStr << *Operands[i]; 270 OS << ")"; 271 } 272 273 SCEVHandle SCEVCommutativeExpr:: 274 replaceSymbolicValuesWithConcrete(const SCEVHandle &Sym, 275 const SCEVHandle &Conc) const { 276 for (unsigned i = 0, e = getNumOperands(); i != e; ++i) { 277 SCEVHandle H = getOperand(i)->replaceSymbolicValuesWithConcrete(Sym, Conc); 278 if (H != getOperand(i)) { 279 std::vector<SCEVHandle> NewOps; 280 NewOps.reserve(getNumOperands()); 281 for (unsigned j = 0; j != i; ++j) 282 NewOps.push_back(getOperand(j)); 283 NewOps.push_back(H); 284 for (++i; i != e; ++i) 285 NewOps.push_back(getOperand(i)-> 286 replaceSymbolicValuesWithConcrete(Sym, Conc)); 287 288 if (isa<SCEVAddExpr>(this)) 289 return SCEVAddExpr::get(NewOps); 290 else if (isa<SCEVMulExpr>(this)) 291 return SCEVMulExpr::get(NewOps); 292 else 293 assert(0 && "Unknown commutative expr!"); 294 } 295 } 296 return this; 297 } 298 299 300 // SCEVSDivs - Only allow the creation of one SCEVSDivExpr for any particular 301 // input. Don't use a SCEVHandle here, or else the object will never be 302 // deleted! 303 static ManagedStatic<std::map<std::pair<SCEV*, SCEV*>, 304 SCEVSDivExpr*> > SCEVSDivs; 305 306 SCEVSDivExpr::~SCEVSDivExpr() { 307 SCEVSDivs->erase(std::make_pair(LHS, RHS)); 308 } 309 310 void SCEVSDivExpr::print(std::ostream &OS) const { 311 OS << "(" << *LHS << " /s " << *RHS << ")"; 312 } 313 314 const Type *SCEVSDivExpr::getType() const { 315 const Type *Ty = LHS->getType(); 316 if (Ty->isUnsigned()) Ty = Ty->getSignedVersion(); 317 return Ty; 318 } 319 320 // SCEVAddRecExprs - Only allow the creation of one SCEVAddRecExpr for any 321 // particular input. Don't use a SCEVHandle here, or else the object will never 322 // be deleted! 323 static ManagedStatic<std::map<std::pair<const Loop *, std::vector<SCEV*> >, 324 SCEVAddRecExpr*> > SCEVAddRecExprs; 325 326 SCEVAddRecExpr::~SCEVAddRecExpr() { 327 SCEVAddRecExprs->erase(std::make_pair(L, 328 std::vector<SCEV*>(Operands.begin(), 329 Operands.end()))); 330 } 331 332 SCEVHandle SCEVAddRecExpr:: 333 replaceSymbolicValuesWithConcrete(const SCEVHandle &Sym, 334 const SCEVHandle &Conc) const { 335 for (unsigned i = 0, e = getNumOperands(); i != e; ++i) { 336 SCEVHandle H = getOperand(i)->replaceSymbolicValuesWithConcrete(Sym, Conc); 337 if (H != getOperand(i)) { 338 std::vector<SCEVHandle> NewOps; 339 NewOps.reserve(getNumOperands()); 340 for (unsigned j = 0; j != i; ++j) 341 NewOps.push_back(getOperand(j)); 342 NewOps.push_back(H); 343 for (++i; i != e; ++i) 344 NewOps.push_back(getOperand(i)-> 345 replaceSymbolicValuesWithConcrete(Sym, Conc)); 346 347 return get(NewOps, L); 348 } 349 } 350 return this; 351 } 352 353 354 bool SCEVAddRecExpr::isLoopInvariant(const Loop *QueryLoop) const { 355 // This recurrence is invariant w.r.t to QueryLoop iff QueryLoop doesn't 356 // contain L and if the start is invariant. 357 return !QueryLoop->contains(L->getHeader()) && 358 getOperand(0)->isLoopInvariant(QueryLoop); 359 } 360 361 362 void SCEVAddRecExpr::print(std::ostream &OS) const { 363 OS << "{" << *Operands[0]; 364 for (unsigned i = 1, e = Operands.size(); i != e; ++i) 365 OS << ",+," << *Operands[i]; 366 OS << "}<" << L->getHeader()->getName() + ">"; 367 } 368 369 // SCEVUnknowns - Only allow the creation of one SCEVUnknown for any particular 370 // value. Don't use a SCEVHandle here, or else the object will never be 371 // deleted! 372 static ManagedStatic<std::map<Value*, SCEVUnknown*> > SCEVUnknowns; 373 374 SCEVUnknown::~SCEVUnknown() { SCEVUnknowns->erase(V); } 375 376 bool SCEVUnknown::isLoopInvariant(const Loop *L) const { 377 // All non-instruction values are loop invariant. All instructions are loop 378 // invariant if they are not contained in the specified loop. 379 if (Instruction *I = dyn_cast<Instruction>(V)) 380 return !L->contains(I->getParent()); 381 return true; 382 } 383 384 const Type *SCEVUnknown::getType() const { 385 return V->getType(); 386 } 387 388 void SCEVUnknown::print(std::ostream &OS) const { 389 WriteAsOperand(OS, V, false); 390 } 391 392 //===----------------------------------------------------------------------===// 393 // SCEV Utilities 394 //===----------------------------------------------------------------------===// 395 396 namespace { 397 /// SCEVComplexityCompare - Return true if the complexity of the LHS is less 398 /// than the complexity of the RHS. This comparator is used to canonicalize 399 /// expressions. 400 struct VISIBILITY_HIDDEN SCEVComplexityCompare { 401 bool operator()(SCEV *LHS, SCEV *RHS) { 402 return LHS->getSCEVType() < RHS->getSCEVType(); 403 } 404 }; 405 } 406 407 /// GroupByComplexity - Given a list of SCEV objects, order them by their 408 /// complexity, and group objects of the same complexity together by value. 409 /// When this routine is finished, we know that any duplicates in the vector are 410 /// consecutive and that complexity is monotonically increasing. 411 /// 412 /// Note that we go take special precautions to ensure that we get determinstic 413 /// results from this routine. In other words, we don't want the results of 414 /// this to depend on where the addresses of various SCEV objects happened to 415 /// land in memory. 416 /// 417 static void GroupByComplexity(std::vector<SCEVHandle> &Ops) { 418 if (Ops.size() < 2) return; // Noop 419 if (Ops.size() == 2) { 420 // This is the common case, which also happens to be trivially simple. 421 // Special case it. 422 if (Ops[0]->getSCEVType() > Ops[1]->getSCEVType()) 423 std::swap(Ops[0], Ops[1]); 424 return; 425 } 426 427 // Do the rough sort by complexity. 428 std::sort(Ops.begin(), Ops.end(), SCEVComplexityCompare()); 429 430 // Now that we are sorted by complexity, group elements of the same 431 // complexity. Note that this is, at worst, N^2, but the vector is likely to 432 // be extremely short in practice. Note that we take this approach because we 433 // do not want to depend on the addresses of the objects we are grouping. 434 for (unsigned i = 0, e = Ops.size(); i != e-2; ++i) { 435 SCEV *S = Ops[i]; 436 unsigned Complexity = S->getSCEVType(); 437 438 // If there are any objects of the same complexity and same value as this 439 // one, group them. 440 for (unsigned j = i+1; j != e && Ops[j]->getSCEVType() == Complexity; ++j) { 441 if (Ops[j] == S) { // Found a duplicate. 442 // Move it to immediately after i'th element. 443 std::swap(Ops[i+1], Ops[j]); 444 ++i; // no need to rescan it. 445 if (i == e-2) return; // Done! 446 } 447 } 448 } 449 } 450 451 452 453 //===----------------------------------------------------------------------===// 454 // Simple SCEV method implementations 455 //===----------------------------------------------------------------------===// 456 457 /// getIntegerSCEV - Given an integer or FP type, create a constant for the 458 /// specified signed integer value and return a SCEV for the constant. 459 SCEVHandle SCEVUnknown::getIntegerSCEV(int Val, const Type *Ty) { 460 Constant *C; 461 if (Val == 0) 462 C = Constant::getNullValue(Ty); 463 else if (Ty->isFloatingPoint()) 464 C = ConstantFP::get(Ty, Val); 465 else if (Ty->isSigned()) 466 C = ConstantInt::get(Ty, Val); 467 else { 468 C = ConstantInt::get(Ty->getSignedVersion(), Val); 469 C = ConstantExpr::getBitCast(C, Ty); 470 } 471 return SCEVUnknown::get(C); 472 } 473 474 /// getTruncateOrZeroExtend - Return a SCEV corresponding to a conversion of the 475 /// input value to the specified type. If the type must be extended, it is zero 476 /// extended. 477 static SCEVHandle getTruncateOrZeroExtend(const SCEVHandle &V, const Type *Ty) { 478 const Type *SrcTy = V->getType(); 479 assert(SrcTy->isInteger() && Ty->isInteger() && 480 "Cannot truncate or zero extend with non-integer arguments!"); 481 if (SrcTy->getPrimitiveSize() == Ty->getPrimitiveSize()) 482 return V; // No conversion 483 if (SrcTy->getPrimitiveSize() > Ty->getPrimitiveSize()) 484 return SCEVTruncateExpr::get(V, Ty); 485 return SCEVZeroExtendExpr::get(V, Ty); 486 } 487 488 /// getNegativeSCEV - Return a SCEV corresponding to -V = -1*V 489 /// 490 SCEVHandle SCEV::getNegativeSCEV(const SCEVHandle &V) { 491 if (SCEVConstant *VC = dyn_cast<SCEVConstant>(V)) 492 return SCEVUnknown::get(ConstantExpr::getNeg(VC->getValue())); 493 494 return SCEVMulExpr::get(V, SCEVUnknown::getIntegerSCEV(-1, V->getType())); 495 } 496 497 /// getMinusSCEV - Return a SCEV corresponding to LHS - RHS. 498 /// 499 SCEVHandle SCEV::getMinusSCEV(const SCEVHandle &LHS, const SCEVHandle &RHS) { 500 // X - Y --> X + -Y 501 return SCEVAddExpr::get(LHS, SCEV::getNegativeSCEV(RHS)); 502 } 503 504 505 /// PartialFact - Compute V!/(V-NumSteps)! 506 static SCEVHandle PartialFact(SCEVHandle V, unsigned NumSteps) { 507 // Handle this case efficiently, it is common to have constant iteration 508 // counts while computing loop exit values. 509 if (SCEVConstant *SC = dyn_cast<SCEVConstant>(V)) { 510 uint64_t Val = SC->getValue()->getZExtValue(); 511 uint64_t Result = 1; 512 for (; NumSteps; --NumSteps) 513 Result *= Val-(NumSteps-1); 514 Constant *Res = ConstantInt::get(Type::ULongTy, Result); 515 return SCEVUnknown::get( 516 ConstantExpr::getTruncOrBitCast(Res, V->getType())); 517 } 518 519 const Type *Ty = V->getType(); 520 if (NumSteps == 0) 521 return SCEVUnknown::getIntegerSCEV(1, Ty); 522 523 SCEVHandle Result = V; 524 for (unsigned i = 1; i != NumSteps; ++i) 525 Result = SCEVMulExpr::get(Result, SCEV::getMinusSCEV(V, 526 SCEVUnknown::getIntegerSCEV(i, Ty))); 527 return Result; 528 } 529 530 531 /// evaluateAtIteration - Return the value of this chain of recurrences at 532 /// the specified iteration number. We can evaluate this recurrence by 533 /// multiplying each element in the chain by the binomial coefficient 534 /// corresponding to it. In other words, we can evaluate {A,+,B,+,C,+,D} as: 535 /// 536 /// A*choose(It, 0) + B*choose(It, 1) + C*choose(It, 2) + D*choose(It, 3) 537 /// 538 /// FIXME/VERIFY: I don't trust that this is correct in the face of overflow. 539 /// Is the binomial equation safe using modular arithmetic?? 540 /// 541 SCEVHandle SCEVAddRecExpr::evaluateAtIteration(SCEVHandle It) const { 542 SCEVHandle Result = getStart(); 543 int Divisor = 1; 544 const Type *Ty = It->getType(); 545 for (unsigned i = 1, e = getNumOperands(); i != e; ++i) { 546 SCEVHandle BC = PartialFact(It, i); 547 Divisor *= i; 548 SCEVHandle Val = SCEVSDivExpr::get(SCEVMulExpr::get(BC, getOperand(i)), 549 SCEVUnknown::getIntegerSCEV(Divisor,Ty)); 550 Result = SCEVAddExpr::get(Result, Val); 551 } 552 return Result; 553 } 554 555 556 //===----------------------------------------------------------------------===// 557 // SCEV Expression folder implementations 558 //===----------------------------------------------------------------------===// 559 560 SCEVHandle SCEVTruncateExpr::get(const SCEVHandle &Op, const Type *Ty) { 561 if (SCEVConstant *SC = dyn_cast<SCEVConstant>(Op)) 562 return SCEVUnknown::get( 563 ConstantExpr::getTrunc(SC->getValue(), Ty)); 564 565 // If the input value is a chrec scev made out of constants, truncate 566 // all of the constants. 567 if (SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(Op)) { 568 std::vector<SCEVHandle> Operands; 569 for (unsigned i = 0, e = AddRec->getNumOperands(); i != e; ++i) 570 // FIXME: This should allow truncation of other expression types! 571 if (isa<SCEVConstant>(AddRec->getOperand(i))) 572 Operands.push_back(get(AddRec->getOperand(i), Ty)); 573 else 574 break; 575 if (Operands.size() == AddRec->getNumOperands()) 576 return SCEVAddRecExpr::get(Operands, AddRec->getLoop()); 577 } 578 579 SCEVTruncateExpr *&Result = (*SCEVTruncates)[std::make_pair(Op, Ty)]; 580 if (Result == 0) Result = new SCEVTruncateExpr(Op, Ty); 581 return Result; 582 } 583 584 SCEVHandle SCEVZeroExtendExpr::get(const SCEVHandle &Op, const Type *Ty) { 585 if (SCEVConstant *SC = dyn_cast<SCEVConstant>(Op)) 586 return SCEVUnknown::get( 587 ConstantExpr::getZeroExtend(SC->getValue(), Ty)); 588 589 // FIXME: If the input value is a chrec scev, and we can prove that the value 590 // did not overflow the old, smaller, value, we can zero extend all of the 591 // operands (often constants). This would allow analysis of something like 592 // this: for (unsigned char X = 0; X < 100; ++X) { int Y = X; } 593 594 SCEVZeroExtendExpr *&Result = (*SCEVZeroExtends)[std::make_pair(Op, Ty)]; 595 if (Result == 0) Result = new SCEVZeroExtendExpr(Op, Ty); 596 return Result; 597 } 598 599 // get - Get a canonical add expression, or something simpler if possible. 600 SCEVHandle SCEVAddExpr::get(std::vector<SCEVHandle> &Ops) { 601 assert(!Ops.empty() && "Cannot get empty add!"); 602 if (Ops.size() == 1) return Ops[0]; 603 604 // Sort by complexity, this groups all similar expression types together. 605 GroupByComplexity(Ops); 606 607 // If there are any constants, fold them together. 608 unsigned Idx = 0; 609 if (SCEVConstant *LHSC = dyn_cast<SCEVConstant>(Ops[0])) { 610 ++Idx; 611 assert(Idx < Ops.size()); 612 while (SCEVConstant *RHSC = dyn_cast<SCEVConstant>(Ops[Idx])) { 613 // We found two constants, fold them together! 614 Constant *Fold = ConstantExpr::getAdd(LHSC->getValue(), RHSC->getValue()); 615 if (ConstantInt *CI = dyn_cast<ConstantInt>(Fold)) { 616 Ops[0] = SCEVConstant::get(CI); 617 Ops.erase(Ops.begin()+1); // Erase the folded element 618 if (Ops.size() == 1) return Ops[0]; 619 LHSC = cast<SCEVConstant>(Ops[0]); 620 } else { 621 // If we couldn't fold the expression, move to the next constant. Note 622 // that this is impossible to happen in practice because we always 623 // constant fold constant ints to constant ints. 624 ++Idx; 625 } 626 } 627 628 // If we are left with a constant zero being added, strip it off. 629 if (cast<SCEVConstant>(Ops[0])->getValue()->isNullValue()) { 630 Ops.erase(Ops.begin()); 631 --Idx; 632 } 633 } 634 635 if (Ops.size() == 1) return Ops[0]; 636 637 // Okay, check to see if the same value occurs in the operand list twice. If 638 // so, merge them together into an multiply expression. Since we sorted the 639 // list, these values are required to be adjacent. 640 const Type *Ty = Ops[0]->getType(); 641 for (unsigned i = 0, e = Ops.size()-1; i != e; ++i) 642 if (Ops[i] == Ops[i+1]) { // X + Y + Y --> X + Y*2 643 // Found a match, merge the two values into a multiply, and add any 644 // remaining values to the result. 645 SCEVHandle Two = SCEVUnknown::getIntegerSCEV(2, Ty); 646 SCEVHandle Mul = SCEVMulExpr::get(Ops[i], Two); 647 if (Ops.size() == 2) 648 return Mul; 649 Ops.erase(Ops.begin()+i, Ops.begin()+i+2); 650 Ops.push_back(Mul); 651 return SCEVAddExpr::get(Ops); 652 } 653 654 // Okay, now we know the first non-constant operand. If there are add 655 // operands they would be next. 656 if (Idx < Ops.size()) { 657 bool DeletedAdd = false; 658 while (SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Ops[Idx])) { 659 // If we have an add, expand the add operands onto the end of the operands 660 // list. 661 Ops.insert(Ops.end(), Add->op_begin(), Add->op_end()); 662 Ops.erase(Ops.begin()+Idx); 663 DeletedAdd = true; 664 } 665 666 // If we deleted at least one add, we added operands to the end of the list, 667 // and they are not necessarily sorted. Recurse to resort and resimplify 668 // any operands we just aquired. 669 if (DeletedAdd) 670 return get(Ops); 671 } 672 673 // Skip over the add expression until we get to a multiply. 674 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scMulExpr) 675 ++Idx; 676 677 // If we are adding something to a multiply expression, make sure the 678 // something is not already an operand of the multiply. If so, merge it into 679 // the multiply. 680 for (; Idx < Ops.size() && isa<SCEVMulExpr>(Ops[Idx]); ++Idx) { 681 SCEVMulExpr *Mul = cast<SCEVMulExpr>(Ops[Idx]); 682 for (unsigned MulOp = 0, e = Mul->getNumOperands(); MulOp != e; ++MulOp) { 683 SCEV *MulOpSCEV = Mul->getOperand(MulOp); 684 for (unsigned AddOp = 0, e = Ops.size(); AddOp != e; ++AddOp) 685 if (MulOpSCEV == Ops[AddOp] && !isa<SCEVConstant>(MulOpSCEV)) { 686 // Fold W + X + (X * Y * Z) --> W + (X * ((Y*Z)+1)) 687 SCEVHandle InnerMul = Mul->getOperand(MulOp == 0); 688 if (Mul->getNumOperands() != 2) { 689 // If the multiply has more than two operands, we must get the 690 // Y*Z term. 691 std::vector<SCEVHandle> MulOps(Mul->op_begin(), Mul->op_end()); 692 MulOps.erase(MulOps.begin()+MulOp); 693 InnerMul = SCEVMulExpr::get(MulOps); 694 } 695 SCEVHandle One = SCEVUnknown::getIntegerSCEV(1, Ty); 696 SCEVHandle AddOne = SCEVAddExpr::get(InnerMul, One); 697 SCEVHandle OuterMul = SCEVMulExpr::get(AddOne, Ops[AddOp]); 698 if (Ops.size() == 2) return OuterMul; 699 if (AddOp < Idx) { 700 Ops.erase(Ops.begin()+AddOp); 701 Ops.erase(Ops.begin()+Idx-1); 702 } else { 703 Ops.erase(Ops.begin()+Idx); 704 Ops.erase(Ops.begin()+AddOp-1); 705 } 706 Ops.push_back(OuterMul); 707 return SCEVAddExpr::get(Ops); 708 } 709 710 // Check this multiply against other multiplies being added together. 711 for (unsigned OtherMulIdx = Idx+1; 712 OtherMulIdx < Ops.size() && isa<SCEVMulExpr>(Ops[OtherMulIdx]); 713 ++OtherMulIdx) { 714 SCEVMulExpr *OtherMul = cast<SCEVMulExpr>(Ops[OtherMulIdx]); 715 // If MulOp occurs in OtherMul, we can fold the two multiplies 716 // together. 717 for (unsigned OMulOp = 0, e = OtherMul->getNumOperands(); 718 OMulOp != e; ++OMulOp) 719 if (OtherMul->getOperand(OMulOp) == MulOpSCEV) { 720 // Fold X + (A*B*C) + (A*D*E) --> X + (A*(B*C+D*E)) 721 SCEVHandle InnerMul1 = Mul->getOperand(MulOp == 0); 722 if (Mul->getNumOperands() != 2) { 723 std::vector<SCEVHandle> MulOps(Mul->op_begin(), Mul->op_end()); 724 MulOps.erase(MulOps.begin()+MulOp); 725 InnerMul1 = SCEVMulExpr::get(MulOps); 726 } 727 SCEVHandle InnerMul2 = OtherMul->getOperand(OMulOp == 0); 728 if (OtherMul->getNumOperands() != 2) { 729 std::vector<SCEVHandle> MulOps(OtherMul->op_begin(), 730 OtherMul->op_end()); 731 MulOps.erase(MulOps.begin()+OMulOp); 732 InnerMul2 = SCEVMulExpr::get(MulOps); 733 } 734 SCEVHandle InnerMulSum = SCEVAddExpr::get(InnerMul1,InnerMul2); 735 SCEVHandle OuterMul = SCEVMulExpr::get(MulOpSCEV, InnerMulSum); 736 if (Ops.size() == 2) return OuterMul; 737 Ops.erase(Ops.begin()+Idx); 738 Ops.erase(Ops.begin()+OtherMulIdx-1); 739 Ops.push_back(OuterMul); 740 return SCEVAddExpr::get(Ops); 741 } 742 } 743 } 744 } 745 746 // If there are any add recurrences in the operands list, see if any other 747 // added values are loop invariant. If so, we can fold them into the 748 // recurrence. 749 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scAddRecExpr) 750 ++Idx; 751 752 // Scan over all recurrences, trying to fold loop invariants into them. 753 for (; Idx < Ops.size() && isa<SCEVAddRecExpr>(Ops[Idx]); ++Idx) { 754 // Scan all of the other operands to this add and add them to the vector if 755 // they are loop invariant w.r.t. the recurrence. 756 std::vector<SCEVHandle> LIOps; 757 SCEVAddRecExpr *AddRec = cast<SCEVAddRecExpr>(Ops[Idx]); 758 for (unsigned i = 0, e = Ops.size(); i != e; ++i) 759 if (Ops[i]->isLoopInvariant(AddRec->getLoop())) { 760 LIOps.push_back(Ops[i]); 761 Ops.erase(Ops.begin()+i); 762 --i; --e; 763 } 764 765 // If we found some loop invariants, fold them into the recurrence. 766 if (!LIOps.empty()) { 767 // NLI + LI + { Start,+,Step} --> NLI + { LI+Start,+,Step } 768 LIOps.push_back(AddRec->getStart()); 769 770 std::vector<SCEVHandle> AddRecOps(AddRec->op_begin(), AddRec->op_end()); 771 AddRecOps[0] = SCEVAddExpr::get(LIOps); 772 773 SCEVHandle NewRec = SCEVAddRecExpr::get(AddRecOps, AddRec->getLoop()); 774 // If all of the other operands were loop invariant, we are done. 775 if (Ops.size() == 1) return NewRec; 776 777 // Otherwise, add the folded AddRec by the non-liv parts. 778 for (unsigned i = 0;; ++i) 779 if (Ops[i] == AddRec) { 780 Ops[i] = NewRec; 781 break; 782 } 783 return SCEVAddExpr::get(Ops); 784 } 785 786 // Okay, if there weren't any loop invariants to be folded, check to see if 787 // there are multiple AddRec's with the same loop induction variable being 788 // added together. If so, we can fold them. 789 for (unsigned OtherIdx = Idx+1; 790 OtherIdx < Ops.size() && isa<SCEVAddRecExpr>(Ops[OtherIdx]);++OtherIdx) 791 if (OtherIdx != Idx) { 792 SCEVAddRecExpr *OtherAddRec = cast<SCEVAddRecExpr>(Ops[OtherIdx]); 793 if (AddRec->getLoop() == OtherAddRec->getLoop()) { 794 // Other + {A,+,B} + {C,+,D} --> Other + {A+C,+,B+D} 795 std::vector<SCEVHandle> NewOps(AddRec->op_begin(), AddRec->op_end()); 796 for (unsigned i = 0, e = OtherAddRec->getNumOperands(); i != e; ++i) { 797 if (i >= NewOps.size()) { 798 NewOps.insert(NewOps.end(), OtherAddRec->op_begin()+i, 799 OtherAddRec->op_end()); 800 break; 801 } 802 NewOps[i] = SCEVAddExpr::get(NewOps[i], OtherAddRec->getOperand(i)); 803 } 804 SCEVHandle NewAddRec = SCEVAddRecExpr::get(NewOps, AddRec->getLoop()); 805 806 if (Ops.size() == 2) return NewAddRec; 807 808 Ops.erase(Ops.begin()+Idx); 809 Ops.erase(Ops.begin()+OtherIdx-1); 810 Ops.push_back(NewAddRec); 811 return SCEVAddExpr::get(Ops); 812 } 813 } 814 815 // Otherwise couldn't fold anything into this recurrence. Move onto the 816 // next one. 817 } 818 819 // Okay, it looks like we really DO need an add expr. Check to see if we 820 // already have one, otherwise create a new one. 821 std::vector<SCEV*> SCEVOps(Ops.begin(), Ops.end()); 822 SCEVCommutativeExpr *&Result = (*SCEVCommExprs)[std::make_pair(scAddExpr, 823 SCEVOps)]; 824 if (Result == 0) Result = new SCEVAddExpr(Ops); 825 return Result; 826 } 827 828 829 SCEVHandle SCEVMulExpr::get(std::vector<SCEVHandle> &Ops) { 830 assert(!Ops.empty() && "Cannot get empty mul!"); 831 832 // Sort by complexity, this groups all similar expression types together. 833 GroupByComplexity(Ops); 834 835 // If there are any constants, fold them together. 836 unsigned Idx = 0; 837 if (SCEVConstant *LHSC = dyn_cast<SCEVConstant>(Ops[0])) { 838 839 // C1*(C2+V) -> C1*C2 + C1*V 840 if (Ops.size() == 2) 841 if (SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Ops[1])) 842 if (Add->getNumOperands() == 2 && 843 isa<SCEVConstant>(Add->getOperand(0))) 844 return SCEVAddExpr::get(SCEVMulExpr::get(LHSC, Add->getOperand(0)), 845 SCEVMulExpr::get(LHSC, Add->getOperand(1))); 846 847 848 ++Idx; 849 while (SCEVConstant *RHSC = dyn_cast<SCEVConstant>(Ops[Idx])) { 850 // We found two constants, fold them together! 851 Constant *Fold = ConstantExpr::getMul(LHSC->getValue(), RHSC->getValue()); 852 if (ConstantInt *CI = dyn_cast<ConstantInt>(Fold)) { 853 Ops[0] = SCEVConstant::get(CI); 854 Ops.erase(Ops.begin()+1); // Erase the folded element 855 if (Ops.size() == 1) return Ops[0]; 856 LHSC = cast<SCEVConstant>(Ops[0]); 857 } else { 858 // If we couldn't fold the expression, move to the next constant. Note 859 // that this is impossible to happen in practice because we always 860 // constant fold constant ints to constant ints. 861 ++Idx; 862 } 863 } 864 865 // If we are left with a constant one being multiplied, strip it off. 866 if (cast<SCEVConstant>(Ops[0])->getValue()->equalsInt(1)) { 867 Ops.erase(Ops.begin()); 868 --Idx; 869 } else if (cast<SCEVConstant>(Ops[0])->getValue()->isNullValue()) { 870 // If we have a multiply of zero, it will always be zero. 871 return Ops[0]; 872 } 873 } 874 875 // Skip over the add expression until we get to a multiply. 876 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scMulExpr) 877 ++Idx; 878 879 if (Ops.size() == 1) 880 return Ops[0]; 881 882 // If there are mul operands inline them all into this expression. 883 if (Idx < Ops.size()) { 884 bool DeletedMul = false; 885 while (SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(Ops[Idx])) { 886 // If we have an mul, expand the mul operands onto the end of the operands 887 // list. 888 Ops.insert(Ops.end(), Mul->op_begin(), Mul->op_end()); 889 Ops.erase(Ops.begin()+Idx); 890 DeletedMul = true; 891 } 892 893 // If we deleted at least one mul, we added operands to the end of the list, 894 // and they are not necessarily sorted. Recurse to resort and resimplify 895 // any operands we just aquired. 896 if (DeletedMul) 897 return get(Ops); 898 } 899 900 // If there are any add recurrences in the operands list, see if any other 901 // added values are loop invariant. If so, we can fold them into the 902 // recurrence. 903 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scAddRecExpr) 904 ++Idx; 905 906 // Scan over all recurrences, trying to fold loop invariants into them. 907 for (; Idx < Ops.size() && isa<SCEVAddRecExpr>(Ops[Idx]); ++Idx) { 908 // Scan all of the other operands to this mul and add them to the vector if 909 // they are loop invariant w.r.t. the recurrence. 910 std::vector<SCEVHandle> LIOps; 911 SCEVAddRecExpr *AddRec = cast<SCEVAddRecExpr>(Ops[Idx]); 912 for (unsigned i = 0, e = Ops.size(); i != e; ++i) 913 if (Ops[i]->isLoopInvariant(AddRec->getLoop())) { 914 LIOps.push_back(Ops[i]); 915 Ops.erase(Ops.begin()+i); 916 --i; --e; 917 } 918 919 // If we found some loop invariants, fold them into the recurrence. 920 if (!LIOps.empty()) { 921 // NLI * LI * { Start,+,Step} --> NLI * { LI*Start,+,LI*Step } 922 std::vector<SCEVHandle> NewOps; 923 NewOps.reserve(AddRec->getNumOperands()); 924 if (LIOps.size() == 1) { 925 SCEV *Scale = LIOps[0]; 926 for (unsigned i = 0, e = AddRec->getNumOperands(); i != e; ++i) 927 NewOps.push_back(SCEVMulExpr::get(Scale, AddRec->getOperand(i))); 928 } else { 929 for (unsigned i = 0, e = AddRec->getNumOperands(); i != e; ++i) { 930 std::vector<SCEVHandle> MulOps(LIOps); 931 MulOps.push_back(AddRec->getOperand(i)); 932 NewOps.push_back(SCEVMulExpr::get(MulOps)); 933 } 934 } 935 936 SCEVHandle NewRec = SCEVAddRecExpr::get(NewOps, AddRec->getLoop()); 937 938 // If all of the other operands were loop invariant, we are done. 939 if (Ops.size() == 1) return NewRec; 940 941 // Otherwise, multiply the folded AddRec by the non-liv parts. 942 for (unsigned i = 0;; ++i) 943 if (Ops[i] == AddRec) { 944 Ops[i] = NewRec; 945 break; 946 } 947 return SCEVMulExpr::get(Ops); 948 } 949 950 // Okay, if there weren't any loop invariants to be folded, check to see if 951 // there are multiple AddRec's with the same loop induction variable being 952 // multiplied together. If so, we can fold them. 953 for (unsigned OtherIdx = Idx+1; 954 OtherIdx < Ops.size() && isa<SCEVAddRecExpr>(Ops[OtherIdx]);++OtherIdx) 955 if (OtherIdx != Idx) { 956 SCEVAddRecExpr *OtherAddRec = cast<SCEVAddRecExpr>(Ops[OtherIdx]); 957 if (AddRec->getLoop() == OtherAddRec->getLoop()) { 958 // F * G --> {A,+,B} * {C,+,D} --> {A*C,+,F*D + G*B + B*D} 959 SCEVAddRecExpr *F = AddRec, *G = OtherAddRec; 960 SCEVHandle NewStart = SCEVMulExpr::get(F->getStart(), 961 G->getStart()); 962 SCEVHandle B = F->getStepRecurrence(); 963 SCEVHandle D = G->getStepRecurrence(); 964 SCEVHandle NewStep = SCEVAddExpr::get(SCEVMulExpr::get(F, D), 965 SCEVMulExpr::get(G, B), 966 SCEVMulExpr::get(B, D)); 967 SCEVHandle NewAddRec = SCEVAddRecExpr::get(NewStart, NewStep, 968 F->getLoop()); 969 if (Ops.size() == 2) return NewAddRec; 970 971 Ops.erase(Ops.begin()+Idx); 972 Ops.erase(Ops.begin()+OtherIdx-1); 973 Ops.push_back(NewAddRec); 974 return SCEVMulExpr::get(Ops); 975 } 976 } 977 978 // Otherwise couldn't fold anything into this recurrence. Move onto the 979 // next one. 980 } 981 982 // Okay, it looks like we really DO need an mul expr. Check to see if we 983 // already have one, otherwise create a new one. 984 std::vector<SCEV*> SCEVOps(Ops.begin(), Ops.end()); 985 SCEVCommutativeExpr *&Result = (*SCEVCommExprs)[std::make_pair(scMulExpr, 986 SCEVOps)]; 987 if (Result == 0) 988 Result = new SCEVMulExpr(Ops); 989 return Result; 990 } 991 992 SCEVHandle SCEVSDivExpr::get(const SCEVHandle &LHS, const SCEVHandle &RHS) { 993 if (SCEVConstant *RHSC = dyn_cast<SCEVConstant>(RHS)) { 994 if (RHSC->getValue()->equalsInt(1)) 995 return LHS; // X sdiv 1 --> x 996 if (RHSC->getValue()->isAllOnesValue()) 997 return SCEV::getNegativeSCEV(LHS); // X sdiv -1 --> -x 998 999 if (SCEVConstant *LHSC = dyn_cast<SCEVConstant>(LHS)) { 1000 Constant *LHSCV = LHSC->getValue(); 1001 Constant *RHSCV = RHSC->getValue(); 1002 return SCEVUnknown::get(ConstantExpr::getSDiv(LHSCV, RHSCV)); 1003 } 1004 } 1005 1006 // FIXME: implement folding of (X*4)/4 when we know X*4 doesn't overflow. 1007 1008 SCEVSDivExpr *&Result = (*SCEVSDivs)[std::make_pair(LHS, RHS)]; 1009 if (Result == 0) Result = new SCEVSDivExpr(LHS, RHS); 1010 return Result; 1011 } 1012 1013 1014 /// SCEVAddRecExpr::get - Get a add recurrence expression for the 1015 /// specified loop. Simplify the expression as much as possible. 1016 SCEVHandle SCEVAddRecExpr::get(const SCEVHandle &Start, 1017 const SCEVHandle &Step, const Loop *L) { 1018 std::vector<SCEVHandle> Operands; 1019 Operands.push_back(Start); 1020 if (SCEVAddRecExpr *StepChrec = dyn_cast<SCEVAddRecExpr>(Step)) 1021 if (StepChrec->getLoop() == L) { 1022 Operands.insert(Operands.end(), StepChrec->op_begin(), 1023 StepChrec->op_end()); 1024 return get(Operands, L); 1025 } 1026 1027 Operands.push_back(Step); 1028 return get(Operands, L); 1029 } 1030 1031 /// SCEVAddRecExpr::get - Get a add recurrence expression for the 1032 /// specified loop. Simplify the expression as much as possible. 1033 SCEVHandle SCEVAddRecExpr::get(std::vector<SCEVHandle> &Operands, 1034 const Loop *L) { 1035 if (Operands.size() == 1) return Operands[0]; 1036 1037 if (SCEVConstant *StepC = dyn_cast<SCEVConstant>(Operands.back())) 1038 if (StepC->getValue()->isNullValue()) { 1039 Operands.pop_back(); 1040 return get(Operands, L); // { X,+,0 } --> X 1041 } 1042 1043 SCEVAddRecExpr *&Result = 1044 (*SCEVAddRecExprs)[std::make_pair(L, std::vector<SCEV*>(Operands.begin(), 1045 Operands.end()))]; 1046 if (Result == 0) Result = new SCEVAddRecExpr(Operands, L); 1047 return Result; 1048 } 1049 1050 SCEVHandle SCEVUnknown::get(Value *V) { 1051 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) 1052 return SCEVConstant::get(CI); 1053 SCEVUnknown *&Result = (*SCEVUnknowns)[V]; 1054 if (Result == 0) Result = new SCEVUnknown(V); 1055 return Result; 1056 } 1057 1058 1059 //===----------------------------------------------------------------------===// 1060 // ScalarEvolutionsImpl Definition and Implementation 1061 //===----------------------------------------------------------------------===// 1062 // 1063 /// ScalarEvolutionsImpl - This class implements the main driver for the scalar 1064 /// evolution code. 1065 /// 1066 namespace { 1067 struct VISIBILITY_HIDDEN ScalarEvolutionsImpl { 1068 /// F - The function we are analyzing. 1069 /// 1070 Function &F; 1071 1072 /// LI - The loop information for the function we are currently analyzing. 1073 /// 1074 LoopInfo &LI; 1075 1076 /// UnknownValue - This SCEV is used to represent unknown trip counts and 1077 /// things. 1078 SCEVHandle UnknownValue; 1079 1080 /// Scalars - This is a cache of the scalars we have analyzed so far. 1081 /// 1082 std::map<Value*, SCEVHandle> Scalars; 1083 1084 /// IterationCounts - Cache the iteration count of the loops for this 1085 /// function as they are computed. 1086 std::map<const Loop*, SCEVHandle> IterationCounts; 1087 1088 /// ConstantEvolutionLoopExitValue - This map contains entries for all of 1089 /// the PHI instructions that we attempt to compute constant evolutions for. 1090 /// This allows us to avoid potentially expensive recomputation of these 1091 /// properties. An instruction maps to null if we are unable to compute its 1092 /// exit value. 1093 std::map<PHINode*, Constant*> ConstantEvolutionLoopExitValue; 1094 1095 public: 1096 ScalarEvolutionsImpl(Function &f, LoopInfo &li) 1097 : F(f), LI(li), UnknownValue(new SCEVCouldNotCompute()) {} 1098 1099 /// getSCEV - Return an existing SCEV if it exists, otherwise analyze the 1100 /// expression and create a new one. 1101 SCEVHandle getSCEV(Value *V); 1102 1103 /// hasSCEV - Return true if the SCEV for this value has already been 1104 /// computed. 1105 bool hasSCEV(Value *V) const { 1106 return Scalars.count(V); 1107 } 1108 1109 /// setSCEV - Insert the specified SCEV into the map of current SCEVs for 1110 /// the specified value. 1111 void setSCEV(Value *V, const SCEVHandle &H) { 1112 bool isNew = Scalars.insert(std::make_pair(V, H)).second; 1113 assert(isNew && "This entry already existed!"); 1114 } 1115 1116 1117 /// getSCEVAtScope - Compute the value of the specified expression within 1118 /// the indicated loop (which may be null to indicate in no loop). If the 1119 /// expression cannot be evaluated, return UnknownValue itself. 1120 SCEVHandle getSCEVAtScope(SCEV *V, const Loop *L); 1121 1122 1123 /// hasLoopInvariantIterationCount - Return true if the specified loop has 1124 /// an analyzable loop-invariant iteration count. 1125 bool hasLoopInvariantIterationCount(const Loop *L); 1126 1127 /// getIterationCount - If the specified loop has a predictable iteration 1128 /// count, return it. Note that it is not valid to call this method on a 1129 /// loop without a loop-invariant iteration count. 1130 SCEVHandle getIterationCount(const Loop *L); 1131 1132 /// deleteInstructionFromRecords - This method should be called by the 1133 /// client before it removes an instruction from the program, to make sure 1134 /// that no dangling references are left around. 1135 void deleteInstructionFromRecords(Instruction *I); 1136 1137 private: 1138 /// createSCEV - We know that there is no SCEV for the specified value. 1139 /// Analyze the expression. 1140 SCEVHandle createSCEV(Value *V); 1141 1142 /// createNodeForPHI - Provide the special handling we need to analyze PHI 1143 /// SCEVs. 1144 SCEVHandle createNodeForPHI(PHINode *PN); 1145 1146 /// ReplaceSymbolicValueWithConcrete - This looks up the computed SCEV value 1147 /// for the specified instruction and replaces any references to the 1148 /// symbolic value SymName with the specified value. This is used during 1149 /// PHI resolution. 1150 void ReplaceSymbolicValueWithConcrete(Instruction *I, 1151 const SCEVHandle &SymName, 1152 const SCEVHandle &NewVal); 1153 1154 /// ComputeIterationCount - Compute the number of times the specified loop 1155 /// will iterate. 1156 SCEVHandle ComputeIterationCount(const Loop *L); 1157 1158 /// ComputeLoadConstantCompareIterationCount - Given an exit condition of 1159 /// 'setcc load X, cst', try to se if we can compute the trip count. 1160 SCEVHandle ComputeLoadConstantCompareIterationCount(LoadInst *LI, 1161 Constant *RHS, 1162 const Loop *L, 1163 unsigned SetCCOpcode); 1164 1165 /// ComputeIterationCountExhaustively - If the trip is known to execute a 1166 /// constant number of times (the condition evolves only from constants), 1167 /// try to evaluate a few iterations of the loop until we get the exit 1168 /// condition gets a value of ExitWhen (true or false). If we cannot 1169 /// evaluate the trip count of the loop, return UnknownValue. 1170 SCEVHandle ComputeIterationCountExhaustively(const Loop *L, Value *Cond, 1171 bool ExitWhen); 1172 1173 /// HowFarToZero - Return the number of times a backedge comparing the 1174 /// specified value to zero will execute. If not computable, return 1175 /// UnknownValue. 1176 SCEVHandle HowFarToZero(SCEV *V, const Loop *L); 1177 1178 /// HowFarToNonZero - Return the number of times a backedge checking the 1179 /// specified value for nonzero will execute. If not computable, return 1180 /// UnknownValue. 1181 SCEVHandle HowFarToNonZero(SCEV *V, const Loop *L); 1182 1183 /// HowManyLessThans - Return the number of times a backedge containing the 1184 /// specified less-than comparison will execute. If not computable, return 1185 /// UnknownValue. 1186 SCEVHandle HowManyLessThans(SCEV *LHS, SCEV *RHS, const Loop *L); 1187 1188 /// getConstantEvolutionLoopExitValue - If we know that the specified Phi is 1189 /// in the header of its containing loop, we know the loop executes a 1190 /// constant number of times, and the PHI node is just a recurrence 1191 /// involving constants, fold it. 1192 Constant *getConstantEvolutionLoopExitValue(PHINode *PN, uint64_t Its, 1193 const Loop *L); 1194 }; 1195 } 1196 1197 //===----------------------------------------------------------------------===// 1198 // Basic SCEV Analysis and PHI Idiom Recognition Code 1199 // 1200 1201 /// deleteInstructionFromRecords - This method should be called by the 1202 /// client before it removes an instruction from the program, to make sure 1203 /// that no dangling references are left around. 1204 void ScalarEvolutionsImpl::deleteInstructionFromRecords(Instruction *I) { 1205 Scalars.erase(I); 1206 if (PHINode *PN = dyn_cast<PHINode>(I)) 1207 ConstantEvolutionLoopExitValue.erase(PN); 1208 } 1209 1210 1211 /// getSCEV - Return an existing SCEV if it exists, otherwise analyze the 1212 /// expression and create a new one. 1213 SCEVHandle ScalarEvolutionsImpl::getSCEV(Value *V) { 1214 assert(V->getType() != Type::VoidTy && "Can't analyze void expressions!"); 1215 1216 std::map<Value*, SCEVHandle>::iterator I = Scalars.find(V); 1217 if (I != Scalars.end()) return I->second; 1218 SCEVHandle S = createSCEV(V); 1219 Scalars.insert(std::make_pair(V, S)); 1220 return S; 1221 } 1222 1223 /// ReplaceSymbolicValueWithConcrete - This looks up the computed SCEV value for 1224 /// the specified instruction and replaces any references to the symbolic value 1225 /// SymName with the specified value. This is used during PHI resolution. 1226 void ScalarEvolutionsImpl:: 1227 ReplaceSymbolicValueWithConcrete(Instruction *I, const SCEVHandle &SymName, 1228 const SCEVHandle &NewVal) { 1229 std::map<Value*, SCEVHandle>::iterator SI = Scalars.find(I); 1230 if (SI == Scalars.end()) return; 1231 1232 SCEVHandle NV = 1233 SI->second->replaceSymbolicValuesWithConcrete(SymName, NewVal); 1234 if (NV == SI->second) return; // No change. 1235 1236 SI->second = NV; // Update the scalars map! 1237 1238 // Any instruction values that use this instruction might also need to be 1239 // updated! 1240 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); 1241 UI != E; ++UI) 1242 ReplaceSymbolicValueWithConcrete(cast<Instruction>(*UI), SymName, NewVal); 1243 } 1244 1245 /// createNodeForPHI - PHI nodes have two cases. Either the PHI node exists in 1246 /// a loop header, making it a potential recurrence, or it doesn't. 1247 /// 1248 SCEVHandle ScalarEvolutionsImpl::createNodeForPHI(PHINode *PN) { 1249 if (PN->getNumIncomingValues() == 2) // The loops have been canonicalized. 1250 if (const Loop *L = LI.getLoopFor(PN->getParent())) 1251 if (L->getHeader() == PN->getParent()) { 1252 // If it lives in the loop header, it has two incoming values, one 1253 // from outside the loop, and one from inside. 1254 unsigned IncomingEdge = L->contains(PN->getIncomingBlock(0)); 1255 unsigned BackEdge = IncomingEdge^1; 1256 1257 // While we are analyzing this PHI node, handle its value symbolically. 1258 SCEVHandle SymbolicName = SCEVUnknown::get(PN); 1259 assert(Scalars.find(PN) == Scalars.end() && 1260 "PHI node already processed?"); 1261 Scalars.insert(std::make_pair(PN, SymbolicName)); 1262 1263 // Using this symbolic name for the PHI, analyze the value coming around 1264 // the back-edge. 1265 SCEVHandle BEValue = getSCEV(PN->getIncomingValue(BackEdge)); 1266 1267 // NOTE: If BEValue is loop invariant, we know that the PHI node just 1268 // has a special value for the first iteration of the loop. 1269 1270 // If the value coming around the backedge is an add with the symbolic 1271 // value we just inserted, then we found a simple induction variable! 1272 if (SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(BEValue)) { 1273 // If there is a single occurrence of the symbolic value, replace it 1274 // with a recurrence. 1275 unsigned FoundIndex = Add->getNumOperands(); 1276 for (unsigned i = 0, e = Add->getNumOperands(); i != e; ++i) 1277 if (Add->getOperand(i) == SymbolicName) 1278 if (FoundIndex == e) { 1279 FoundIndex = i; 1280 break; 1281 } 1282 1283 if (FoundIndex != Add->getNumOperands()) { 1284 // Create an add with everything but the specified operand. 1285 std::vector<SCEVHandle> Ops; 1286 for (unsigned i = 0, e = Add->getNumOperands(); i != e; ++i) 1287 if (i != FoundIndex) 1288 Ops.push_back(Add->getOperand(i)); 1289 SCEVHandle Accum = SCEVAddExpr::get(Ops); 1290 1291 // This is not a valid addrec if the step amount is varying each 1292 // loop iteration, but is not itself an addrec in this loop. 1293 if (Accum->isLoopInvariant(L) || 1294 (isa<SCEVAddRecExpr>(Accum) && 1295 cast<SCEVAddRecExpr>(Accum)->getLoop() == L)) { 1296 SCEVHandle StartVal = getSCEV(PN->getIncomingValue(IncomingEdge)); 1297 SCEVHandle PHISCEV = SCEVAddRecExpr::get(StartVal, Accum, L); 1298 1299 // Okay, for the entire analysis of this edge we assumed the PHI 1300 // to be symbolic. We now need to go back and update all of the 1301 // entries for the scalars that use the PHI (except for the PHI 1302 // itself) to use the new analyzed value instead of the "symbolic" 1303 // value. 1304 ReplaceSymbolicValueWithConcrete(PN, SymbolicName, PHISCEV); 1305 return PHISCEV; 1306 } 1307 } 1308 } else if (SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(BEValue)) { 1309 // Otherwise, this could be a loop like this: 1310 // i = 0; for (j = 1; ..; ++j) { .... i = j; } 1311 // In this case, j = {1,+,1} and BEValue is j. 1312 // Because the other in-value of i (0) fits the evolution of BEValue 1313 // i really is an addrec evolution. 1314 if (AddRec->getLoop() == L && AddRec->isAffine()) { 1315 SCEVHandle StartVal = getSCEV(PN->getIncomingValue(IncomingEdge)); 1316 1317 // If StartVal = j.start - j.stride, we can use StartVal as the 1318 // initial step of the addrec evolution. 1319 if (StartVal == SCEV::getMinusSCEV(AddRec->getOperand(0), 1320 AddRec->getOperand(1))) { 1321 SCEVHandle PHISCEV = 1322 SCEVAddRecExpr::get(StartVal, AddRec->getOperand(1), L); 1323 1324 // Okay, for the entire analysis of this edge we assumed the PHI 1325 // to be symbolic. We now need to go back and update all of the 1326 // entries for the scalars that use the PHI (except for the PHI 1327 // itself) to use the new analyzed value instead of the "symbolic" 1328 // value. 1329 ReplaceSymbolicValueWithConcrete(PN, SymbolicName, PHISCEV); 1330 return PHISCEV; 1331 } 1332 } 1333 } 1334 1335 return SymbolicName; 1336 } 1337 1338 // If it's not a loop phi, we can't handle it yet. 1339 return SCEVUnknown::get(PN); 1340 } 1341 1342 1343 /// createSCEV - We know that there is no SCEV for the specified value. 1344 /// Analyze the expression. 1345 /// 1346 SCEVHandle ScalarEvolutionsImpl::createSCEV(Value *V) { 1347 if (Instruction *I = dyn_cast<Instruction>(V)) { 1348 switch (I->getOpcode()) { 1349 case Instruction::Add: 1350 return SCEVAddExpr::get(getSCEV(I->getOperand(0)), 1351 getSCEV(I->getOperand(1))); 1352 case Instruction::Mul: 1353 return SCEVMulExpr::get(getSCEV(I->getOperand(0)), 1354 getSCEV(I->getOperand(1))); 1355 case Instruction::SDiv: 1356 return SCEVSDivExpr::get(getSCEV(I->getOperand(0)), 1357 getSCEV(I->getOperand(1))); 1358 break; 1359 1360 case Instruction::Sub: 1361 return SCEV::getMinusSCEV(getSCEV(I->getOperand(0)), 1362 getSCEV(I->getOperand(1))); 1363 1364 case Instruction::Shl: 1365 // Turn shift left of a constant amount into a multiply. 1366 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) { 1367 Constant *X = ConstantInt::get(V->getType(), 1); 1368 X = ConstantExpr::getShl(X, SA); 1369 return SCEVMulExpr::get(getSCEV(I->getOperand(0)), getSCEV(X)); 1370 } 1371 break; 1372 1373 case Instruction::Trunc: 1374 // We must prevent boolean types such as setne, etc. from entering here 1375 // because we don't want to pass SCEVUnknown to the TruncateExpr. 1376 if (I->getType()->isInteger() && I->getOperand(0)->getType()->isInteger()) 1377 return SCEVTruncateExpr::get(getSCEV(I->getOperand(0)), 1378 I->getType()->getUnsignedVersion()); 1379 break; 1380 1381 case Instruction::ZExt: 1382 // We must prevent boolean types such as setne, etc. from entering here 1383 // because we don't want to pass SCEVUnknown to the ZExtExpr. 1384 if (I->getType()->isInteger() && I->getOperand(0)->getType()->isInteger()) 1385 return SCEVZeroExtendExpr::get(getSCEV(I->getOperand(0)), 1386 I->getType()->getUnsignedVersion()); 1387 break; 1388 1389 case Instruction::BitCast: 1390 // BitCasts are no-op casts so we just eliminate the cast. 1391 return getSCEV(I->getOperand(0)); 1392 1393 case Instruction::PHI: 1394 return createNodeForPHI(cast<PHINode>(I)); 1395 1396 default: // We cannot analyze this expression. 1397 break; 1398 } 1399 } 1400 1401 return SCEVUnknown::get(V); 1402 } 1403 1404 1405 1406 //===----------------------------------------------------------------------===// 1407 // Iteration Count Computation Code 1408 // 1409 1410 /// getIterationCount - If the specified loop has a predictable iteration 1411 /// count, return it. Note that it is not valid to call this method on a 1412 /// loop without a loop-invariant iteration count. 1413 SCEVHandle ScalarEvolutionsImpl::getIterationCount(const Loop *L) { 1414 std::map<const Loop*, SCEVHandle>::iterator I = IterationCounts.find(L); 1415 if (I == IterationCounts.end()) { 1416 SCEVHandle ItCount = ComputeIterationCount(L); 1417 I = IterationCounts.insert(std::make_pair(L, ItCount)).first; 1418 if (ItCount != UnknownValue) { 1419 assert(ItCount->isLoopInvariant(L) && 1420 "Computed trip count isn't loop invariant for loop!"); 1421 ++NumTripCountsComputed; 1422 } else if (isa<PHINode>(L->getHeader()->begin())) { 1423 // Only count loops that have phi nodes as not being computable. 1424 ++NumTripCountsNotComputed; 1425 } 1426 } 1427 return I->second; 1428 } 1429 1430 /// ComputeIterationCount - Compute the number of times the specified loop 1431 /// will iterate. 1432 SCEVHandle ScalarEvolutionsImpl::ComputeIterationCount(const Loop *L) { 1433 // If the loop has a non-one exit block count, we can't analyze it. 1434 std::vector<BasicBlock*> ExitBlocks; 1435 L->getExitBlocks(ExitBlocks); 1436 if (ExitBlocks.size() != 1) return UnknownValue; 1437 1438 // Okay, there is one exit block. Try to find the condition that causes the 1439 // loop to be exited. 1440 BasicBlock *ExitBlock = ExitBlocks[0]; 1441 1442 BasicBlock *ExitingBlock = 0; 1443 for (pred_iterator PI = pred_begin(ExitBlock), E = pred_end(ExitBlock); 1444 PI != E; ++PI) 1445 if (L->contains(*PI)) { 1446 if (ExitingBlock == 0) 1447 ExitingBlock = *PI; 1448 else 1449 return UnknownValue; // More than one block exiting! 1450 } 1451 assert(ExitingBlock && "No exits from loop, something is broken!"); 1452 1453 // Okay, we've computed the exiting block. See what condition causes us to 1454 // exit. 1455 // 1456 // FIXME: we should be able to handle switch instructions (with a single exit) 1457 // FIXME: We should handle cast of int to bool as well 1458 BranchInst *ExitBr = dyn_cast<BranchInst>(ExitingBlock->getTerminator()); 1459 if (ExitBr == 0) return UnknownValue; 1460 assert(ExitBr->isConditional() && "If unconditional, it can't be in loop!"); 1461 SetCondInst *ExitCond = dyn_cast<SetCondInst>(ExitBr->getCondition()); 1462 if (ExitCond == 0) // Not a setcc 1463 return ComputeIterationCountExhaustively(L, ExitBr->getCondition(), 1464 ExitBr->getSuccessor(0) == ExitBlock); 1465 1466 // If the condition was exit on true, convert the condition to exit on false. 1467 Instruction::BinaryOps Cond; 1468 if (ExitBr->getSuccessor(1) == ExitBlock) 1469 Cond = ExitCond->getOpcode(); 1470 else 1471 Cond = ExitCond->getInverseCondition(); 1472 1473 // Handle common loops like: for (X = "string"; *X; ++X) 1474 if (LoadInst *LI = dyn_cast<LoadInst>(ExitCond->getOperand(0))) 1475 if (Constant *RHS = dyn_cast<Constant>(ExitCond->getOperand(1))) { 1476 SCEVHandle ItCnt = 1477 ComputeLoadConstantCompareIterationCount(LI, RHS, L, Cond); 1478 if (!isa<SCEVCouldNotCompute>(ItCnt)) return ItCnt; 1479 } 1480 1481 SCEVHandle LHS = getSCEV(ExitCond->getOperand(0)); 1482 SCEVHandle RHS = getSCEV(ExitCond->getOperand(1)); 1483 1484 // Try to evaluate any dependencies out of the loop. 1485 SCEVHandle Tmp = getSCEVAtScope(LHS, L); 1486 if (!isa<SCEVCouldNotCompute>(Tmp)) LHS = Tmp; 1487 Tmp = getSCEVAtScope(RHS, L); 1488 if (!isa<SCEVCouldNotCompute>(Tmp)) RHS = Tmp; 1489 1490 // At this point, we would like to compute how many iterations of the loop the 1491 // predicate will return true for these inputs. 1492 if (isa<SCEVConstant>(LHS) && !isa<SCEVConstant>(RHS)) { 1493 // If there is a constant, force it into the RHS. 1494 std::swap(LHS, RHS); 1495 Cond = SetCondInst::getSwappedCondition(Cond); 1496 } 1497 1498 // FIXME: think about handling pointer comparisons! i.e.: 1499 // while (P != P+100) ++P; 1500 1501 // If we have a comparison of a chrec against a constant, try to use value 1502 // ranges to answer this query. 1503 if (SCEVConstant *RHSC = dyn_cast<SCEVConstant>(RHS)) 1504 if (SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(LHS)) 1505 if (AddRec->getLoop() == L) { 1506 // Form the comparison range using the constant of the correct type so 1507 // that the ConstantRange class knows to do a signed or unsigned 1508 // comparison. 1509 ConstantInt *CompVal = RHSC->getValue(); 1510 const Type *RealTy = ExitCond->getOperand(0)->getType(); 1511 CompVal = dyn_cast<ConstantInt>(ConstantExpr::getCast(CompVal, RealTy)); 1512 if (CompVal) { 1513 // Form the constant range. 1514 ConstantRange CompRange(Cond, CompVal); 1515 1516 // Now that we have it, if it's signed, convert it to an unsigned 1517 // range. 1518 if (CompRange.getLower()->getType()->isSigned()) { 1519 const Type *NewTy = RHSC->getValue()->getType(); 1520 Constant *NewL = ConstantExpr::getCast(CompRange.getLower(), NewTy); 1521 Constant *NewU = ConstantExpr::getCast(CompRange.getUpper(), NewTy); 1522 CompRange = ConstantRange(NewL, NewU); 1523 } 1524 1525 SCEVHandle Ret = AddRec->getNumIterationsInRange(CompRange); 1526 if (!isa<SCEVCouldNotCompute>(Ret)) return Ret; 1527 } 1528 } 1529 1530 switch (Cond) { 1531 case Instruction::SetNE: // while (X != Y) 1532 // Convert to: while (X-Y != 0) 1533 if (LHS->getType()->isInteger()) { 1534 SCEVHandle TC = HowFarToZero(SCEV::getMinusSCEV(LHS, RHS), L); 1535 if (!isa<SCEVCouldNotCompute>(TC)) return TC; 1536 } 1537 break; 1538 case Instruction::SetEQ: 1539 // Convert to: while (X-Y == 0) // while (X == Y) 1540 if (LHS->getType()->isInteger()) { 1541 SCEVHandle TC = HowFarToNonZero(SCEV::getMinusSCEV(LHS, RHS), L); 1542 if (!isa<SCEVCouldNotCompute>(TC)) return TC; 1543 } 1544 break; 1545 case Instruction::SetLT: 1546 if (LHS->getType()->isInteger() && 1547 ExitCond->getOperand(0)->getType()->isSigned()) { 1548 SCEVHandle TC = HowManyLessThans(LHS, RHS, L); 1549 if (!isa<SCEVCouldNotCompute>(TC)) return TC; 1550 } 1551 break; 1552 case Instruction::SetGT: 1553 if (LHS->getType()->isInteger() && 1554 ExitCond->getOperand(0)->getType()->isSigned()) { 1555 SCEVHandle TC = HowManyLessThans(RHS, LHS, L); 1556 if (!isa<SCEVCouldNotCompute>(TC)) return TC; 1557 } 1558 break; 1559 default: 1560 #if 0 1561 llvm_cerr << "ComputeIterationCount "; 1562 if (ExitCond->getOperand(0)->getType()->isUnsigned()) 1563 llvm_cerr << "[unsigned] "; 1564 llvm_cerr << *LHS << " " 1565 << Instruction::getOpcodeName(Cond) << " " << *RHS << "\n"; 1566 #endif 1567 break; 1568 } 1569 1570 return ComputeIterationCountExhaustively(L, ExitCond, 1571 ExitBr->getSuccessor(0) == ExitBlock); 1572 } 1573 1574 static ConstantInt * 1575 EvaluateConstantChrecAtConstant(const SCEVAddRecExpr *AddRec, Constant *C) { 1576 SCEVHandle InVal = SCEVConstant::get(cast<ConstantInt>(C)); 1577 SCEVHandle Val = AddRec->evaluateAtIteration(InVal); 1578 assert(isa<SCEVConstant>(Val) && 1579 "Evaluation of SCEV at constant didn't fold correctly?"); 1580 return cast<SCEVConstant>(Val)->getValue(); 1581 } 1582 1583 /// GetAddressedElementFromGlobal - Given a global variable with an initializer 1584 /// and a GEP expression (missing the pointer index) indexing into it, return 1585 /// the addressed element of the initializer or null if the index expression is 1586 /// invalid. 1587 static Constant * 1588 GetAddressedElementFromGlobal(GlobalVariable *GV, 1589 const std::vector<ConstantInt*> &Indices) { 1590 Constant *Init = GV->getInitializer(); 1591 for (unsigned i = 0, e = Indices.size(); i != e; ++i) { 1592 uint64_t Idx = Indices[i]->getZExtValue(); 1593 if (ConstantStruct *CS = dyn_cast<ConstantStruct>(Init)) { 1594 assert(Idx < CS->getNumOperands() && "Bad struct index!"); 1595 Init = cast<Constant>(CS->getOperand(Idx)); 1596 } else if (ConstantArray *CA = dyn_cast<ConstantArray>(Init)) { 1597 if (Idx >= CA->getNumOperands()) return 0; // Bogus program 1598 Init = cast<Constant>(CA->getOperand(Idx)); 1599 } else if (isa<ConstantAggregateZero>(Init)) { 1600 if (const StructType *STy = dyn_cast<StructType>(Init->getType())) { 1601 assert(Idx < STy->getNumElements() && "Bad struct index!"); 1602 Init = Constant::getNullValue(STy->getElementType(Idx)); 1603 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Init->getType())) { 1604 if (Idx >= ATy->getNumElements()) return 0; // Bogus program 1605 Init = Constant::getNullValue(ATy->getElementType()); 1606 } else { 1607 assert(0 && "Unknown constant aggregate type!"); 1608 } 1609 return 0; 1610 } else { 1611 return 0; // Unknown initializer type 1612 } 1613 } 1614 return Init; 1615 } 1616 1617 /// ComputeLoadConstantCompareIterationCount - Given an exit condition of 1618 /// 'setcc load X, cst', try to se if we can compute the trip count. 1619 SCEVHandle ScalarEvolutionsImpl:: 1620 ComputeLoadConstantCompareIterationCount(LoadInst *LI, Constant *RHS, 1621 const Loop *L, unsigned SetCCOpcode) { 1622 if (LI->isVolatile()) return UnknownValue; 1623 1624 // Check to see if the loaded pointer is a getelementptr of a global. 1625 GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(LI->getOperand(0)); 1626 if (!GEP) return UnknownValue; 1627 1628 // Make sure that it is really a constant global we are gepping, with an 1629 // initializer, and make sure the first IDX is really 0. 1630 GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)); 1631 if (!GV || !GV->isConstant() || !GV->hasInitializer() || 1632 GEP->getNumOperands() < 3 || !isa<Constant>(GEP->getOperand(1)) || 1633 !cast<Constant>(GEP->getOperand(1))->isNullValue()) 1634 return UnknownValue; 1635 1636 // Okay, we allow one non-constant index into the GEP instruction. 1637 Value *VarIdx = 0; 1638 std::vector<ConstantInt*> Indexes; 1639 unsigned VarIdxNum = 0; 1640 for (unsigned i = 2, e = GEP->getNumOperands(); i != e; ++i) 1641 if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i))) { 1642 Indexes.push_back(CI); 1643 } else if (!isa<ConstantInt>(GEP->getOperand(i))) { 1644 if (VarIdx) return UnknownValue; // Multiple non-constant idx's. 1645 VarIdx = GEP->getOperand(i); 1646 VarIdxNum = i-2; 1647 Indexes.push_back(0); 1648 } 1649 1650 // Okay, we know we have a (load (gep GV, 0, X)) comparison with a constant. 1651 // Check to see if X is a loop variant variable value now. 1652 SCEVHandle Idx = getSCEV(VarIdx); 1653 SCEVHandle Tmp = getSCEVAtScope(Idx, L); 1654 if (!isa<SCEVCouldNotCompute>(Tmp)) Idx = Tmp; 1655 1656 // We can only recognize very limited forms of loop index expressions, in 1657 // particular, only affine AddRec's like {C1,+,C2}. 1658 SCEVAddRecExpr *IdxExpr = dyn_cast<SCEVAddRecExpr>(Idx); 1659 if (!IdxExpr || !IdxExpr->isAffine() || IdxExpr->isLoopInvariant(L) || 1660 !isa<SCEVConstant>(IdxExpr->getOperand(0)) || 1661 !isa<SCEVConstant>(IdxExpr->getOperand(1))) 1662 return UnknownValue; 1663 1664 unsigned MaxSteps = MaxBruteForceIterations; 1665 for (unsigned IterationNum = 0; IterationNum != MaxSteps; ++IterationNum) { 1666 ConstantInt *ItCst = 1667 ConstantInt::get(IdxExpr->getType()->getUnsignedVersion(), IterationNum); 1668 ConstantInt *Val = EvaluateConstantChrecAtConstant(IdxExpr, ItCst); 1669 1670 // Form the GEP offset. 1671 Indexes[VarIdxNum] = Val; 1672 1673 Constant *Result = GetAddressedElementFromGlobal(GV, Indexes); 1674 if (Result == 0) break; // Cannot compute! 1675 1676 // Evaluate the condition for this iteration. 1677 Result = ConstantExpr::get(SetCCOpcode, Result, RHS); 1678 if (!isa<ConstantBool>(Result)) break; // Couldn't decide for sure 1679 if (cast<ConstantBool>(Result)->getValue() == false) { 1680 #if 0 1681 llvm_cerr << "\n***\n*** Computed loop count " << *ItCst 1682 << "\n*** From global " << *GV << "*** BB: " << *L->getHeader() 1683 << "***\n"; 1684 #endif 1685 ++NumArrayLenItCounts; 1686 return SCEVConstant::get(ItCst); // Found terminating iteration! 1687 } 1688 } 1689 return UnknownValue; 1690 } 1691 1692 1693 /// CanConstantFold - Return true if we can constant fold an instruction of the 1694 /// specified type, assuming that all operands were constants. 1695 static bool CanConstantFold(const Instruction *I) { 1696 if (isa<BinaryOperator>(I) || isa<ShiftInst>(I) || 1697 isa<SelectInst>(I) || isa<CastInst>(I) || isa<GetElementPtrInst>(I)) 1698 return true; 1699 1700 if (const CallInst *CI = dyn_cast<CallInst>(I)) 1701 if (const Function *F = CI->getCalledFunction()) 1702 return canConstantFoldCallTo((Function*)F); // FIXME: elim cast 1703 return false; 1704 } 1705 1706 /// ConstantFold - Constant fold an instruction of the specified type with the 1707 /// specified constant operands. This function may modify the operands vector. 1708 static Constant *ConstantFold(const Instruction *I, 1709 std::vector<Constant*> &Operands) { 1710 if (isa<BinaryOperator>(I) || isa<ShiftInst>(I)) 1711 return ConstantExpr::get(I->getOpcode(), Operands[0], Operands[1]); 1712 1713 if (isa<CastInst>(I)) 1714 return ConstantExpr::getCast(I->getOpcode(), Operands[0], I->getType()); 1715 1716 switch (I->getOpcode()) { 1717 case Instruction::Select: 1718 return ConstantExpr::getSelect(Operands[0], Operands[1], Operands[2]); 1719 case Instruction::Call: 1720 if (Function *GV = dyn_cast<Function>(Operands[0])) { 1721 Operands.erase(Operands.begin()); 1722 return ConstantFoldCall(cast<Function>(GV), Operands); 1723 } 1724 return 0; 1725 case Instruction::GetElementPtr: 1726 Constant *Base = Operands[0]; 1727 Operands.erase(Operands.begin()); 1728 return ConstantExpr::getGetElementPtr(Base, Operands); 1729 } 1730 return 0; 1731 } 1732 1733 1734 /// getConstantEvolvingPHI - Given an LLVM value and a loop, return a PHI node 1735 /// in the loop that V is derived from. We allow arbitrary operations along the 1736 /// way, but the operands of an operation must either be constants or a value 1737 /// derived from a constant PHI. If this expression does not fit with these 1738 /// constraints, return null. 1739 static PHINode *getConstantEvolvingPHI(Value *V, const Loop *L) { 1740 // If this is not an instruction, or if this is an instruction outside of the 1741 // loop, it can't be derived from a loop PHI. 1742 Instruction *I = dyn_cast<Instruction>(V); 1743 if (I == 0 || !L->contains(I->getParent())) return 0; 1744 1745 if (PHINode *PN = dyn_cast<PHINode>(I)) 1746 if (L->getHeader() == I->getParent()) 1747 return PN; 1748 else 1749 // We don't currently keep track of the control flow needed to evaluate 1750 // PHIs, so we cannot handle PHIs inside of loops. 1751 return 0; 1752 1753 // If we won't be able to constant fold this expression even if the operands 1754 // are constants, return early. 1755 if (!CanConstantFold(I)) return 0; 1756 1757 // Otherwise, we can evaluate this instruction if all of its operands are 1758 // constant or derived from a PHI node themselves. 1759 PHINode *PHI = 0; 1760 for (unsigned Op = 0, e = I->getNumOperands(); Op != e; ++Op) 1761 if (!(isa<Constant>(I->getOperand(Op)) || 1762 isa<GlobalValue>(I->getOperand(Op)))) { 1763 PHINode *P = getConstantEvolvingPHI(I->getOperand(Op), L); 1764 if (P == 0) return 0; // Not evolving from PHI 1765 if (PHI == 0) 1766 PHI = P; 1767 else if (PHI != P) 1768 return 0; // Evolving from multiple different PHIs. 1769 } 1770 1771 // This is a expression evolving from a constant PHI! 1772 return PHI; 1773 } 1774 1775 /// EvaluateExpression - Given an expression that passes the 1776 /// getConstantEvolvingPHI predicate, evaluate its value assuming the PHI node 1777 /// in the loop has the value PHIVal. If we can't fold this expression for some 1778 /// reason, return null. 1779 static Constant *EvaluateExpression(Value *V, Constant *PHIVal) { 1780 if (isa<PHINode>(V)) return PHIVal; 1781 if (GlobalValue *GV = dyn_cast<GlobalValue>(V)) 1782 return GV; 1783 if (Constant *C = dyn_cast<Constant>(V)) return C; 1784 Instruction *I = cast<Instruction>(V); 1785 1786 std::vector<Constant*> Operands; 1787 Operands.resize(I->getNumOperands()); 1788 1789 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) { 1790 Operands[i] = EvaluateExpression(I->getOperand(i), PHIVal); 1791 if (Operands[i] == 0) return 0; 1792 } 1793 1794 return ConstantFold(I, Operands); 1795 } 1796 1797 /// getConstantEvolutionLoopExitValue - If we know that the specified Phi is 1798 /// in the header of its containing loop, we know the loop executes a 1799 /// constant number of times, and the PHI node is just a recurrence 1800 /// involving constants, fold it. 1801 Constant *ScalarEvolutionsImpl:: 1802 getConstantEvolutionLoopExitValue(PHINode *PN, uint64_t Its, const Loop *L) { 1803 std::map<PHINode*, Constant*>::iterator I = 1804 ConstantEvolutionLoopExitValue.find(PN); 1805 if (I != ConstantEvolutionLoopExitValue.end()) 1806 return I->second; 1807 1808 if (Its > MaxBruteForceIterations) 1809 return ConstantEvolutionLoopExitValue[PN] = 0; // Not going to evaluate it. 1810 1811 Constant *&RetVal = ConstantEvolutionLoopExitValue[PN]; 1812 1813 // Since the loop is canonicalized, the PHI node must have two entries. One 1814 // entry must be a constant (coming in from outside of the loop), and the 1815 // second must be derived from the same PHI. 1816 bool SecondIsBackedge = L->contains(PN->getIncomingBlock(1)); 1817 Constant *StartCST = 1818 dyn_cast<Constant>(PN->getIncomingValue(!SecondIsBackedge)); 1819 if (StartCST == 0) 1820 return RetVal = 0; // Must be a constant. 1821 1822 Value *BEValue = PN->getIncomingValue(SecondIsBackedge); 1823 PHINode *PN2 = getConstantEvolvingPHI(BEValue, L); 1824 if (PN2 != PN) 1825 return RetVal = 0; // Not derived from same PHI. 1826 1827 // Execute the loop symbolically to determine the exit value. 1828 unsigned IterationNum = 0; 1829 unsigned NumIterations = Its; 1830 if (NumIterations != Its) 1831 return RetVal = 0; // More than 2^32 iterations?? 1832 1833 for (Constant *PHIVal = StartCST; ; ++IterationNum) { 1834 if (IterationNum == NumIterations) 1835 return RetVal = PHIVal; // Got exit value! 1836 1837 // Compute the value of the PHI node for the next iteration. 1838 Constant *NextPHI = EvaluateExpression(BEValue, PHIVal); 1839 if (NextPHI == PHIVal) 1840 return RetVal = NextPHI; // Stopped evolving! 1841 if (NextPHI == 0) 1842 return 0; // Couldn't evaluate! 1843 PHIVal = NextPHI; 1844 } 1845 } 1846 1847 /// ComputeIterationCountExhaustively - If the trip is known to execute a 1848 /// constant number of times (the condition evolves only from constants), 1849 /// try to evaluate a few iterations of the loop until we get the exit 1850 /// condition gets a value of ExitWhen (true or false). If we cannot 1851 /// evaluate the trip count of the loop, return UnknownValue. 1852 SCEVHandle ScalarEvolutionsImpl:: 1853 ComputeIterationCountExhaustively(const Loop *L, Value *Cond, bool ExitWhen) { 1854 PHINode *PN = getConstantEvolvingPHI(Cond, L); 1855 if (PN == 0) return UnknownValue; 1856 1857 // Since the loop is canonicalized, the PHI node must have two entries. One 1858 // entry must be a constant (coming in from outside of the loop), and the 1859 // second must be derived from the same PHI. 1860 bool SecondIsBackedge = L->contains(PN->getIncomingBlock(1)); 1861 Constant *StartCST = 1862 dyn_cast<Constant>(PN->getIncomingValue(!SecondIsBackedge)); 1863 if (StartCST == 0) return UnknownValue; // Must be a constant. 1864 1865 Value *BEValue = PN->getIncomingValue(SecondIsBackedge); 1866 PHINode *PN2 = getConstantEvolvingPHI(BEValue, L); 1867 if (PN2 != PN) return UnknownValue; // Not derived from same PHI. 1868 1869 // Okay, we find a PHI node that defines the trip count of this loop. Execute 1870 // the loop symbolically to determine when the condition gets a value of 1871 // "ExitWhen". 1872 unsigned IterationNum = 0; 1873 unsigned MaxIterations = MaxBruteForceIterations; // Limit analysis. 1874 for (Constant *PHIVal = StartCST; 1875 IterationNum != MaxIterations; ++IterationNum) { 1876 ConstantBool *CondVal = 1877 dyn_cast_or_null<ConstantBool>(EvaluateExpression(Cond, PHIVal)); 1878 if (!CondVal) return UnknownValue; // Couldn't symbolically evaluate. 1879 1880 if (CondVal->getValue() == ExitWhen) { 1881 ConstantEvolutionLoopExitValue[PN] = PHIVal; 1882 ++NumBruteForceTripCountsComputed; 1883 return SCEVConstant::get(ConstantInt::get(Type::UIntTy, IterationNum)); 1884 } 1885 1886 // Compute the value of the PHI node for the next iteration. 1887 Constant *NextPHI = EvaluateExpression(BEValue, PHIVal); 1888 if (NextPHI == 0 || NextPHI == PHIVal) 1889 return UnknownValue; // Couldn't evaluate or not making progress... 1890 PHIVal = NextPHI; 1891 } 1892 1893 // Too many iterations were needed to evaluate. 1894 return UnknownValue; 1895 } 1896 1897 /// getSCEVAtScope - Compute the value of the specified expression within the 1898 /// indicated loop (which may be null to indicate in no loop). If the 1899 /// expression cannot be evaluated, return UnknownValue. 1900 SCEVHandle ScalarEvolutionsImpl::getSCEVAtScope(SCEV *V, const Loop *L) { 1901 // FIXME: this should be turned into a virtual method on SCEV! 1902 1903 if (isa<SCEVConstant>(V)) return V; 1904 1905 // If this instruction is evolves from a constant-evolving PHI, compute the 1906 // exit value from the loop without using SCEVs. 1907 if (SCEVUnknown *SU = dyn_cast<SCEVUnknown>(V)) { 1908 if (Instruction *I = dyn_cast<Instruction>(SU->getValue())) { 1909 const Loop *LI = this->LI[I->getParent()]; 1910 if (LI && LI->getParentLoop() == L) // Looking for loop exit value. 1911 if (PHINode *PN = dyn_cast<PHINode>(I)) 1912 if (PN->getParent() == LI->getHeader()) { 1913 // Okay, there is no closed form solution for the PHI node. Check 1914 // to see if the loop that contains it has a known iteration count. 1915 // If so, we may be able to force computation of the exit value. 1916 SCEVHandle IterationCount = getIterationCount(LI); 1917 if (SCEVConstant *ICC = dyn_cast<SCEVConstant>(IterationCount)) { 1918 // Okay, we know how many times the containing loop executes. If 1919 // this is a constant evolving PHI node, get the final value at 1920 // the specified iteration number. 1921 Constant *RV = getConstantEvolutionLoopExitValue(PN, 1922 ICC->getValue()->getZExtValue(), 1923 LI); 1924 if (RV) return SCEVUnknown::get(RV); 1925 } 1926 } 1927 1928 // Okay, this is an expression that we cannot symbolically evaluate 1929 // into a SCEV. Check to see if it's possible to symbolically evaluate 1930 // the arguments into constants, and if so, try to constant propagate the 1931 // result. This is particularly useful for computing loop exit values. 1932 if (CanConstantFold(I)) { 1933 std::vector<Constant*> Operands; 1934 Operands.reserve(I->getNumOperands()); 1935 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) { 1936 Value *Op = I->getOperand(i); 1937 if (Constant *C = dyn_cast<Constant>(Op)) { 1938 Operands.push_back(C); 1939 } else { 1940 SCEVHandle OpV = getSCEVAtScope(getSCEV(Op), L); 1941 if (SCEVConstant *SC = dyn_cast<SCEVConstant>(OpV)) 1942 Operands.push_back(ConstantExpr::getCast(SC->getValue(), 1943 Op->getType())); 1944 else if (SCEVUnknown *SU = dyn_cast<SCEVUnknown>(OpV)) { 1945 if (Constant *C = dyn_cast<Constant>(SU->getValue())) 1946 Operands.push_back(ConstantExpr::getCast(C, Op->getType())); 1947 else 1948 return V; 1949 } else { 1950 return V; 1951 } 1952 } 1953 } 1954 return SCEVUnknown::get(ConstantFold(I, Operands)); 1955 } 1956 } 1957 1958 // This is some other type of SCEVUnknown, just return it. 1959 return V; 1960 } 1961 1962 if (SCEVCommutativeExpr *Comm = dyn_cast<SCEVCommutativeExpr>(V)) { 1963 // Avoid performing the look-up in the common case where the specified 1964 // expression has no loop-variant portions. 1965 for (unsigned i = 0, e = Comm->getNumOperands(); i != e; ++i) { 1966 SCEVHandle OpAtScope = getSCEVAtScope(Comm->getOperand(i), L); 1967 if (OpAtScope != Comm->getOperand(i)) { 1968 if (OpAtScope == UnknownValue) return UnknownValue; 1969 // Okay, at least one of these operands is loop variant but might be 1970 // foldable. Build a new instance of the folded commutative expression. 1971 std::vector<SCEVHandle> NewOps(Comm->op_begin(), Comm->op_begin()+i); 1972 NewOps.push_back(OpAtScope); 1973 1974 for (++i; i != e; ++i) { 1975 OpAtScope = getSCEVAtScope(Comm->getOperand(i), L); 1976 if (OpAtScope == UnknownValue) return UnknownValue; 1977 NewOps.push_back(OpAtScope); 1978 } 1979 if (isa<SCEVAddExpr>(Comm)) 1980 return SCEVAddExpr::get(NewOps); 1981 assert(isa<SCEVMulExpr>(Comm) && "Only know about add and mul!"); 1982 return SCEVMulExpr::get(NewOps); 1983 } 1984 } 1985 // If we got here, all operands are loop invariant. 1986 return Comm; 1987 } 1988 1989 if (SCEVSDivExpr *Div = dyn_cast<SCEVSDivExpr>(V)) { 1990 SCEVHandle LHS = getSCEVAtScope(Div->getLHS(), L); 1991 if (LHS == UnknownValue) return LHS; 1992 SCEVHandle RHS = getSCEVAtScope(Div->getRHS(), L); 1993 if (RHS == UnknownValue) return RHS; 1994 if (LHS == Div->getLHS() && RHS == Div->getRHS()) 1995 return Div; // must be loop invariant 1996 return SCEVSDivExpr::get(LHS, RHS); 1997 } 1998 1999 // If this is a loop recurrence for a loop that does not contain L, then we 2000 // are dealing with the final value computed by the loop. 2001 if (SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(V)) { 2002 if (!L || !AddRec->getLoop()->contains(L->getHeader())) { 2003 // To evaluate this recurrence, we need to know how many times the AddRec 2004 // loop iterates. Compute this now. 2005 SCEVHandle IterationCount = getIterationCount(AddRec->getLoop()); 2006 if (IterationCount == UnknownValue) return UnknownValue; 2007 IterationCount = getTruncateOrZeroExtend(IterationCount, 2008 AddRec->getType()); 2009 2010 // If the value is affine, simplify the expression evaluation to just 2011 // Start + Step*IterationCount. 2012 if (AddRec->isAffine()) 2013 return SCEVAddExpr::get(AddRec->getStart(), 2014 SCEVMulExpr::get(IterationCount, 2015 AddRec->getOperand(1))); 2016 2017 // Otherwise, evaluate it the hard way. 2018 return AddRec->evaluateAtIteration(IterationCount); 2019 } 2020 return UnknownValue; 2021 } 2022 2023 //assert(0 && "Unknown SCEV type!"); 2024 return UnknownValue; 2025 } 2026 2027 2028 /// SolveQuadraticEquation - Find the roots of the quadratic equation for the 2029 /// given quadratic chrec {L,+,M,+,N}. This returns either the two roots (which 2030 /// might be the same) or two SCEVCouldNotCompute objects. 2031 /// 2032 static std::pair<SCEVHandle,SCEVHandle> 2033 SolveQuadraticEquation(const SCEVAddRecExpr *AddRec) { 2034 assert(AddRec->getNumOperands() == 3 && "This is not a quadratic chrec!"); 2035 SCEVConstant *L = dyn_cast<SCEVConstant>(AddRec->getOperand(0)); 2036 SCEVConstant *M = dyn_cast<SCEVConstant>(AddRec->getOperand(1)); 2037 SCEVConstant *N = dyn_cast<SCEVConstant>(AddRec->getOperand(2)); 2038 2039 // We currently can only solve this if the coefficients are constants. 2040 if (!L || !M || !N) { 2041 SCEV *CNC = new SCEVCouldNotCompute(); 2042 return std::make_pair(CNC, CNC); 2043 } 2044 2045 Constant *C = L->getValue(); 2046 Constant *Two = ConstantInt::get(C->getType(), 2); 2047 2048 // Convert from chrec coefficients to polynomial coefficients AX^2+BX+C 2049 // The B coefficient is M-N/2 2050 Constant *B = ConstantExpr::getSub(M->getValue(), 2051 ConstantExpr::getSDiv(N->getValue(), 2052 Two)); 2053 // The A coefficient is N/2 2054 Constant *A = ConstantExpr::getSDiv(N->getValue(), Two); 2055 2056 // Compute the B^2-4ac term. 2057 Constant *SqrtTerm = 2058 ConstantExpr::getMul(ConstantInt::get(C->getType(), 4), 2059 ConstantExpr::getMul(A, C)); 2060 SqrtTerm = ConstantExpr::getSub(ConstantExpr::getMul(B, B), SqrtTerm); 2061 2062 // Compute floor(sqrt(B^2-4ac)) 2063 ConstantInt *SqrtVal = 2064 cast<ConstantInt>(ConstantExpr::getCast(SqrtTerm, 2065 SqrtTerm->getType()->getUnsignedVersion())); 2066 uint64_t SqrtValV = SqrtVal->getZExtValue(); 2067 uint64_t SqrtValV2 = (uint64_t)sqrt((double)SqrtValV); 2068 // The square root might not be precise for arbitrary 64-bit integer 2069 // values. Do some sanity checks to ensure it's correct. 2070 if (SqrtValV2*SqrtValV2 > SqrtValV || 2071 (SqrtValV2+1)*(SqrtValV2+1) <= SqrtValV) { 2072 SCEV *CNC = new SCEVCouldNotCompute(); 2073 return std::make_pair(CNC, CNC); 2074 } 2075 2076 SqrtVal = ConstantInt::get(Type::ULongTy, SqrtValV2); 2077 SqrtTerm = ConstantExpr::getCast(SqrtVal, SqrtTerm->getType()); 2078 2079 Constant *NegB = ConstantExpr::getNeg(B); 2080 Constant *TwoA = ConstantExpr::getMul(A, Two); 2081 2082 // The divisions must be performed as signed divisions. 2083 const Type *SignedTy = NegB->getType()->getSignedVersion(); 2084 NegB = ConstantExpr::getCast(NegB, SignedTy); 2085 TwoA = ConstantExpr::getCast(TwoA, SignedTy); 2086 SqrtTerm = ConstantExpr::getCast(SqrtTerm, SignedTy); 2087 2088 Constant *Solution1 = 2089 ConstantExpr::getSDiv(ConstantExpr::getAdd(NegB, SqrtTerm), TwoA); 2090 Constant *Solution2 = 2091 ConstantExpr::getSDiv(ConstantExpr::getSub(NegB, SqrtTerm), TwoA); 2092 return std::make_pair(SCEVUnknown::get(Solution1), 2093 SCEVUnknown::get(Solution2)); 2094 } 2095 2096 /// HowFarToZero - Return the number of times a backedge comparing the specified 2097 /// value to zero will execute. If not computable, return UnknownValue 2098 SCEVHandle ScalarEvolutionsImpl::HowFarToZero(SCEV *V, const Loop *L) { 2099 // If the value is a constant 2100 if (SCEVConstant *C = dyn_cast<SCEVConstant>(V)) { 2101 // If the value is already zero, the branch will execute zero times. 2102 if (C->getValue()->isNullValue()) return C; 2103 return UnknownValue; // Otherwise it will loop infinitely. 2104 } 2105 2106 SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(V); 2107 if (!AddRec || AddRec->getLoop() != L) 2108 return UnknownValue; 2109 2110 if (AddRec->isAffine()) { 2111 // If this is an affine expression the execution count of this branch is 2112 // equal to: 2113 // 2114 // (0 - Start/Step) iff Start % Step == 0 2115 // 2116 // Get the initial value for the loop. 2117 SCEVHandle Start = getSCEVAtScope(AddRec->getStart(), L->getParentLoop()); 2118 if (isa<SCEVCouldNotCompute>(Start)) return UnknownValue; 2119 SCEVHandle Step = AddRec->getOperand(1); 2120 2121 Step = getSCEVAtScope(Step, L->getParentLoop()); 2122 2123 // Figure out if Start % Step == 0. 2124 // FIXME: We should add DivExpr and RemExpr operations to our AST. 2125 if (SCEVConstant *StepC = dyn_cast<SCEVConstant>(Step)) { 2126 if (StepC->getValue()->equalsInt(1)) // N % 1 == 0 2127 return SCEV::getNegativeSCEV(Start); // 0 - Start/1 == -Start 2128 if (StepC->getValue()->isAllOnesValue()) // N % -1 == 0 2129 return Start; // 0 - Start/-1 == Start 2130 2131 // Check to see if Start is divisible by SC with no remainder. 2132 if (SCEVConstant *StartC = dyn_cast<SCEVConstant>(Start)) { 2133 ConstantInt *StartCC = StartC->getValue(); 2134 Constant *StartNegC = ConstantExpr::getNeg(StartCC); 2135 Constant *Rem = ConstantExpr::getSRem(StartNegC, StepC->getValue()); 2136 if (Rem->isNullValue()) { 2137 Constant *Result =ConstantExpr::getSDiv(StartNegC,StepC->getValue()); 2138 return SCEVUnknown::get(Result); 2139 } 2140 } 2141 } 2142 } else if (AddRec->isQuadratic() && AddRec->getType()->isInteger()) { 2143 // If this is a quadratic (3-term) AddRec {L,+,M,+,N}, find the roots of 2144 // the quadratic equation to solve it. 2145 std::pair<SCEVHandle,SCEVHandle> Roots = SolveQuadraticEquation(AddRec); 2146 SCEVConstant *R1 = dyn_cast<SCEVConstant>(Roots.first); 2147 SCEVConstant *R2 = dyn_cast<SCEVConstant>(Roots.second); 2148 if (R1) { 2149 #if 0 2150 llvm_cerr << "HFTZ: " << *V << " - sol#1: " << *R1 2151 << " sol#2: " << *R2 << "\n"; 2152 #endif 2153 // Pick the smallest positive root value. 2154 assert(R1->getType()->isUnsigned()&&"Didn't canonicalize to unsigned?"); 2155 if (ConstantBool *CB = 2156 dyn_cast<ConstantBool>(ConstantExpr::getSetLT(R1->getValue(), 2157 R2->getValue()))) { 2158 if (CB->getValue() == false) 2159 std::swap(R1, R2); // R1 is the minimum root now. 2160 2161 // We can only use this value if the chrec ends up with an exact zero 2162 // value at this index. When solving for "X*X != 5", for example, we 2163 // should not accept a root of 2. 2164 SCEVHandle Val = AddRec->evaluateAtIteration(R1); 2165 if (SCEVConstant *EvalVal = dyn_cast<SCEVConstant>(Val)) 2166 if (EvalVal->getValue()->isNullValue()) 2167 return R1; // We found a quadratic root! 2168 } 2169 } 2170 } 2171 2172 return UnknownValue; 2173 } 2174 2175 /// HowFarToNonZero - Return the number of times a backedge checking the 2176 /// specified value for nonzero will execute. If not computable, return 2177 /// UnknownValue 2178 SCEVHandle ScalarEvolutionsImpl::HowFarToNonZero(SCEV *V, const Loop *L) { 2179 // Loops that look like: while (X == 0) are very strange indeed. We don't 2180 // handle them yet except for the trivial case. This could be expanded in the 2181 // future as needed. 2182 2183 // If the value is a constant, check to see if it is known to be non-zero 2184 // already. If so, the backedge will execute zero times. 2185 if (SCEVConstant *C = dyn_cast<SCEVConstant>(V)) { 2186 Constant *Zero = Constant::getNullValue(C->getValue()->getType()); 2187 Constant *NonZero = ConstantExpr::getSetNE(C->getValue(), Zero); 2188 if (NonZero == ConstantBool::getTrue()) 2189 return getSCEV(Zero); 2190 return UnknownValue; // Otherwise it will loop infinitely. 2191 } 2192 2193 // We could implement others, but I really doubt anyone writes loops like 2194 // this, and if they did, they would already be constant folded. 2195 return UnknownValue; 2196 } 2197 2198 /// HowManyLessThans - Return the number of times a backedge containing the 2199 /// specified less-than comparison will execute. If not computable, return 2200 /// UnknownValue. 2201 SCEVHandle ScalarEvolutionsImpl:: 2202 HowManyLessThans(SCEV *LHS, SCEV *RHS, const Loop *L) { 2203 // Only handle: "ADDREC < LoopInvariant". 2204 if (!RHS->isLoopInvariant(L)) return UnknownValue; 2205 2206 SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(LHS); 2207 if (!AddRec || AddRec->getLoop() != L) 2208 return UnknownValue; 2209 2210 if (AddRec->isAffine()) { 2211 // FORNOW: We only support unit strides. 2212 SCEVHandle One = SCEVUnknown::getIntegerSCEV(1, RHS->getType()); 2213 if (AddRec->getOperand(1) != One) 2214 return UnknownValue; 2215 2216 // The number of iterations for "[n,+,1] < m", is m-n. However, we don't 2217 // know that m is >= n on input to the loop. If it is, the condition return 2218 // true zero times. What we really should return, for full generality, is 2219 // SMAX(0, m-n). Since we cannot check this, we will instead check for a 2220 // canonical loop form: most do-loops will have a check that dominates the 2221 // loop, that only enters the loop if [n-1]<m. If we can find this check, 2222 // we know that the SMAX will evaluate to m-n, because we know that m >= n. 2223 2224 // Search for the check. 2225 BasicBlock *Preheader = L->getLoopPreheader(); 2226 BasicBlock *PreheaderDest = L->getHeader(); 2227 if (Preheader == 0) return UnknownValue; 2228 2229 BranchInst *LoopEntryPredicate = 2230 dyn_cast<BranchInst>(Preheader->getTerminator()); 2231 if (!LoopEntryPredicate) return UnknownValue; 2232 2233 // This might be a critical edge broken out. If the loop preheader ends in 2234 // an unconditional branch to the loop, check to see if the preheader has a 2235 // single predecessor, and if so, look for its terminator. 2236 while (LoopEntryPredicate->isUnconditional()) { 2237 PreheaderDest = Preheader; 2238 Preheader = Preheader->getSinglePredecessor(); 2239 if (!Preheader) return UnknownValue; // Multiple preds. 2240 2241 LoopEntryPredicate = 2242 dyn_cast<BranchInst>(Preheader->getTerminator()); 2243 if (!LoopEntryPredicate) return UnknownValue; 2244 } 2245 2246 // Now that we found a conditional branch that dominates the loop, check to 2247 // see if it is the comparison we are looking for. 2248 SetCondInst *SCI =dyn_cast<SetCondInst>(LoopEntryPredicate->getCondition()); 2249 if (!SCI) return UnknownValue; 2250 Value *PreCondLHS = SCI->getOperand(0); 2251 Value *PreCondRHS = SCI->getOperand(1); 2252 Instruction::BinaryOps Cond; 2253 if (LoopEntryPredicate->getSuccessor(0) == PreheaderDest) 2254 Cond = SCI->getOpcode(); 2255 else 2256 Cond = SCI->getInverseCondition(); 2257 2258 switch (Cond) { 2259 case Instruction::SetGT: 2260 std::swap(PreCondLHS, PreCondRHS); 2261 Cond = Instruction::SetLT; 2262 // Fall Through. 2263 case Instruction::SetLT: 2264 if (PreCondLHS->getType()->isInteger() && 2265 PreCondLHS->getType()->isSigned()) { 2266 if (RHS != getSCEV(PreCondRHS)) 2267 return UnknownValue; // Not a comparison against 'm'. 2268 2269 if (SCEV::getMinusSCEV(AddRec->getOperand(0), One) 2270 != getSCEV(PreCondLHS)) 2271 return UnknownValue; // Not a comparison against 'n-1'. 2272 break; 2273 } else { 2274 return UnknownValue; 2275 } 2276 default: break; 2277 } 2278 2279 //llvm_cerr << "Computed Loop Trip Count as: " << 2280 // *SCEV::getMinusSCEV(RHS, AddRec->getOperand(0)) << "\n"; 2281 return SCEV::getMinusSCEV(RHS, AddRec->getOperand(0)); 2282 } 2283 2284 return UnknownValue; 2285 } 2286 2287 /// getNumIterationsInRange - Return the number of iterations of this loop that 2288 /// produce values in the specified constant range. Another way of looking at 2289 /// this is that it returns the first iteration number where the value is not in 2290 /// the condition, thus computing the exit count. If the iteration count can't 2291 /// be computed, an instance of SCEVCouldNotCompute is returned. 2292 SCEVHandle SCEVAddRecExpr::getNumIterationsInRange(ConstantRange Range) const { 2293 if (Range.isFullSet()) // Infinite loop. 2294 return new SCEVCouldNotCompute(); 2295 2296 // If the start is a non-zero constant, shift the range to simplify things. 2297 if (SCEVConstant *SC = dyn_cast<SCEVConstant>(getStart())) 2298 if (!SC->getValue()->isNullValue()) { 2299 std::vector<SCEVHandle> Operands(op_begin(), op_end()); 2300 Operands[0] = SCEVUnknown::getIntegerSCEV(0, SC->getType()); 2301 SCEVHandle Shifted = SCEVAddRecExpr::get(Operands, getLoop()); 2302 if (SCEVAddRecExpr *ShiftedAddRec = dyn_cast<SCEVAddRecExpr>(Shifted)) 2303 return ShiftedAddRec->getNumIterationsInRange( 2304 Range.subtract(SC->getValue())); 2305 // This is strange and shouldn't happen. 2306 return new SCEVCouldNotCompute(); 2307 } 2308 2309 // The only time we can solve this is when we have all constant indices. 2310 // Otherwise, we cannot determine the overflow conditions. 2311 for (unsigned i = 0, e = getNumOperands(); i != e; ++i) 2312 if (!isa<SCEVConstant>(getOperand(i))) 2313 return new SCEVCouldNotCompute(); 2314 2315 2316 // Okay at this point we know that all elements of the chrec are constants and 2317 // that the start element is zero. 2318 2319 // First check to see if the range contains zero. If not, the first 2320 // iteration exits. 2321 ConstantInt *Zero = ConstantInt::get(getType(), 0); 2322 if (!Range.contains(Zero)) return SCEVConstant::get(Zero); 2323 2324 if (isAffine()) { 2325 // If this is an affine expression then we have this situation: 2326 // Solve {0,+,A} in Range === Ax in Range 2327 2328 // Since we know that zero is in the range, we know that the upper value of 2329 // the range must be the first possible exit value. Also note that we 2330 // already checked for a full range. 2331 ConstantInt *Upper = cast<ConstantInt>(Range.getUpper()); 2332 ConstantInt *A = cast<SCEVConstant>(getOperand(1))->getValue(); 2333 ConstantInt *One = ConstantInt::get(getType(), 1); 2334 2335 // The exit value should be (Upper+A-1)/A. 2336 Constant *ExitValue = Upper; 2337 if (A != One) { 2338 ExitValue = ConstantExpr::getSub(ConstantExpr::getAdd(Upper, A), One); 2339 ExitValue = ConstantExpr::getSDiv(ExitValue, A); 2340 } 2341 assert(isa<ConstantInt>(ExitValue) && 2342 "Constant folding of integers not implemented?"); 2343 2344 // Evaluate at the exit value. If we really did fall out of the valid 2345 // range, then we computed our trip count, otherwise wrap around or other 2346 // things must have happened. 2347 ConstantInt *Val = EvaluateConstantChrecAtConstant(this, ExitValue); 2348 if (Range.contains(Val)) 2349 return new SCEVCouldNotCompute(); // Something strange happened 2350 2351 // Ensure that the previous value is in the range. This is a sanity check. 2352 assert(Range.contains(EvaluateConstantChrecAtConstant(this, 2353 ConstantExpr::getSub(ExitValue, One))) && 2354 "Linear scev computation is off in a bad way!"); 2355 return SCEVConstant::get(cast<ConstantInt>(ExitValue)); 2356 } else if (isQuadratic()) { 2357 // If this is a quadratic (3-term) AddRec {L,+,M,+,N}, find the roots of the 2358 // quadratic equation to solve it. To do this, we must frame our problem in 2359 // terms of figuring out when zero is crossed, instead of when 2360 // Range.getUpper() is crossed. 2361 std::vector<SCEVHandle> NewOps(op_begin(), op_end()); 2362 NewOps[0] = SCEV::getNegativeSCEV(SCEVUnknown::get(Range.getUpper())); 2363 SCEVHandle NewAddRec = SCEVAddRecExpr::get(NewOps, getLoop()); 2364 2365 // Next, solve the constructed addrec 2366 std::pair<SCEVHandle,SCEVHandle> Roots = 2367 SolveQuadraticEquation(cast<SCEVAddRecExpr>(NewAddRec)); 2368 SCEVConstant *R1 = dyn_cast<SCEVConstant>(Roots.first); 2369 SCEVConstant *R2 = dyn_cast<SCEVConstant>(Roots.second); 2370 if (R1) { 2371 // Pick the smallest positive root value. 2372 assert(R1->getType()->isUnsigned() && "Didn't canonicalize to unsigned?"); 2373 if (ConstantBool *CB = 2374 dyn_cast<ConstantBool>(ConstantExpr::getSetLT(R1->getValue(), 2375 R2->getValue()))) { 2376 if (CB->getValue() == false) 2377 std::swap(R1, R2); // R1 is the minimum root now. 2378 2379 // Make sure the root is not off by one. The returned iteration should 2380 // not be in the range, but the previous one should be. When solving 2381 // for "X*X < 5", for example, we should not return a root of 2. 2382 ConstantInt *R1Val = EvaluateConstantChrecAtConstant(this, 2383 R1->getValue()); 2384 if (Range.contains(R1Val)) { 2385 // The next iteration must be out of the range... 2386 Constant *NextVal = 2387 ConstantExpr::getAdd(R1->getValue(), 2388 ConstantInt::get(R1->getType(), 1)); 2389 2390 R1Val = EvaluateConstantChrecAtConstant(this, NextVal); 2391 if (!Range.contains(R1Val)) 2392 return SCEVUnknown::get(NextVal); 2393 return new SCEVCouldNotCompute(); // Something strange happened 2394 } 2395 2396 // If R1 was not in the range, then it is a good return value. Make 2397 // sure that R1-1 WAS in the range though, just in case. 2398 Constant *NextVal = 2399 ConstantExpr::getSub(R1->getValue(), 2400 ConstantInt::get(R1->getType(), 1)); 2401 R1Val = EvaluateConstantChrecAtConstant(this, NextVal); 2402 if (Range.contains(R1Val)) 2403 return R1; 2404 return new SCEVCouldNotCompute(); // Something strange happened 2405 } 2406 } 2407 } 2408 2409 // Fallback, if this is a general polynomial, figure out the progression 2410 // through brute force: evaluate until we find an iteration that fails the 2411 // test. This is likely to be slow, but getting an accurate trip count is 2412 // incredibly important, we will be able to simplify the exit test a lot, and 2413 // we are almost guaranteed to get a trip count in this case. 2414 ConstantInt *TestVal = ConstantInt::get(getType(), 0); 2415 ConstantInt *One = ConstantInt::get(getType(), 1); 2416 ConstantInt *EndVal = TestVal; // Stop when we wrap around. 2417 do { 2418 ++NumBruteForceEvaluations; 2419 SCEVHandle Val = evaluateAtIteration(SCEVConstant::get(TestVal)); 2420 if (!isa<SCEVConstant>(Val)) // This shouldn't happen. 2421 return new SCEVCouldNotCompute(); 2422 2423 // Check to see if we found the value! 2424 if (!Range.contains(cast<SCEVConstant>(Val)->getValue())) 2425 return SCEVConstant::get(TestVal); 2426 2427 // Increment to test the next index. 2428 TestVal = cast<ConstantInt>(ConstantExpr::getAdd(TestVal, One)); 2429 } while (TestVal != EndVal); 2430 2431 return new SCEVCouldNotCompute(); 2432 } 2433 2434 2435 2436 //===----------------------------------------------------------------------===// 2437 // ScalarEvolution Class Implementation 2438 //===----------------------------------------------------------------------===// 2439 2440 bool ScalarEvolution::runOnFunction(Function &F) { 2441 Impl = new ScalarEvolutionsImpl(F, getAnalysis<LoopInfo>()); 2442 return false; 2443 } 2444 2445 void ScalarEvolution::releaseMemory() { 2446 delete (ScalarEvolutionsImpl*)Impl; 2447 Impl = 0; 2448 } 2449 2450 void ScalarEvolution::getAnalysisUsage(AnalysisUsage &AU) const { 2451 AU.setPreservesAll(); 2452 AU.addRequiredTransitive<LoopInfo>(); 2453 } 2454 2455 SCEVHandle ScalarEvolution::getSCEV(Value *V) const { 2456 return ((ScalarEvolutionsImpl*)Impl)->getSCEV(V); 2457 } 2458 2459 /// hasSCEV - Return true if the SCEV for this value has already been 2460 /// computed. 2461 bool ScalarEvolution::hasSCEV(Value *V) const { 2462 return ((ScalarEvolutionsImpl*)Impl)->hasSCEV(V); 2463 } 2464 2465 2466 /// setSCEV - Insert the specified SCEV into the map of current SCEVs for 2467 /// the specified value. 2468 void ScalarEvolution::setSCEV(Value *V, const SCEVHandle &H) { 2469 ((ScalarEvolutionsImpl*)Impl)->setSCEV(V, H); 2470 } 2471 2472 2473 SCEVHandle ScalarEvolution::getIterationCount(const Loop *L) const { 2474 return ((ScalarEvolutionsImpl*)Impl)->getIterationCount(L); 2475 } 2476 2477 bool ScalarEvolution::hasLoopInvariantIterationCount(const Loop *L) const { 2478 return !isa<SCEVCouldNotCompute>(getIterationCount(L)); 2479 } 2480 2481 SCEVHandle ScalarEvolution::getSCEVAtScope(Value *V, const Loop *L) const { 2482 return ((ScalarEvolutionsImpl*)Impl)->getSCEVAtScope(getSCEV(V), L); 2483 } 2484 2485 void ScalarEvolution::deleteInstructionFromRecords(Instruction *I) const { 2486 return ((ScalarEvolutionsImpl*)Impl)->deleteInstructionFromRecords(I); 2487 } 2488 2489 static void PrintLoopInfo(std::ostream &OS, const ScalarEvolution *SE, 2490 const Loop *L) { 2491 // Print all inner loops first 2492 for (Loop::iterator I = L->begin(), E = L->end(); I != E; ++I) 2493 PrintLoopInfo(OS, SE, *I); 2494 2495 llvm_cerr << "Loop " << L->getHeader()->getName() << ": "; 2496 2497 std::vector<BasicBlock*> ExitBlocks; 2498 L->getExitBlocks(ExitBlocks); 2499 if (ExitBlocks.size() != 1) 2500 llvm_cerr << "<multiple exits> "; 2501 2502 if (SE->hasLoopInvariantIterationCount(L)) { 2503 llvm_cerr << *SE->getIterationCount(L) << " iterations! "; 2504 } else { 2505 llvm_cerr << "Unpredictable iteration count. "; 2506 } 2507 2508 llvm_cerr << "\n"; 2509 } 2510 2511 void ScalarEvolution::print(std::ostream &OS, const Module* ) const { 2512 Function &F = ((ScalarEvolutionsImpl*)Impl)->F; 2513 LoopInfo &LI = ((ScalarEvolutionsImpl*)Impl)->LI; 2514 2515 OS << "Classifying expressions for: " << F.getName() << "\n"; 2516 for (inst_iterator I = inst_begin(F), E = inst_end(F); I != E; ++I) 2517 if (I->getType()->isInteger()) { 2518 OS << *I; 2519 OS << " --> "; 2520 SCEVHandle SV = getSCEV(&*I); 2521 SV->print(OS); 2522 OS << "\t\t"; 2523 2524 if ((*I).getType()->isIntegral()) { 2525 ConstantRange Bounds = SV->getValueRange(); 2526 if (!Bounds.isFullSet()) 2527 OS << "Bounds: " << Bounds << " "; 2528 } 2529 2530 if (const Loop *L = LI.getLoopFor((*I).getParent())) { 2531 OS << "Exits: "; 2532 SCEVHandle ExitValue = getSCEVAtScope(&*I, L->getParentLoop()); 2533 if (isa<SCEVCouldNotCompute>(ExitValue)) { 2534 OS << "<<Unknown>>"; 2535 } else { 2536 OS << *ExitValue; 2537 } 2538 } 2539 2540 2541 OS << "\n"; 2542 } 2543 2544 OS << "Determining loop execution counts for: " << F.getName() << "\n"; 2545 for (LoopInfo::iterator I = LI.begin(), E = LI.end(); I != E; ++I) 2546 PrintLoopInfo(OS, this, *I); 2547 } 2548 2549