1//===- ScalarEvolution.cpp - Scalar Evolution Analysis ----------*- C++ -*-===// 2// 3// The LLVM Compiler Infrastructure 4// 5// This file is distributed under the University of Illinois Open Source 6// License. See LICENSE.TXT for details. 7// 8//===----------------------------------------------------------------------===// 9// 10// This 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. We only create one SCEV of a particular shape, so 18// pointer-comparisons for equality are legal. 19// 20// One important aspect of the SCEV objects is that they are never cyclic, even 21// if there is a cycle in the dataflow for an expression (ie, a PHI node). If 22// the PHI node is one of the idioms that we can represent (e.g., a polynomial 23// recurrence) then we represent it directly as a recurrence node, otherwise we 24// represent it as a SCEVUnknown node. 25// 26// In addition to being able to represent expressions of various types, we also 27// have folders that are used to build the *canonical* representation for a 28// particular expression. These folders are capable of using a variety of 29// rewrite rules to simplify the expressions. 30// 31// Once the folders are defined, we can implement the more interesting 32// higher-level code, such as the code that recognizes PHI nodes of various 33// types, computes the execution count of a loop, etc. 34// 35// TODO: We should use these routines and value representations to implement 36// dependence analysis! 37// 38//===----------------------------------------------------------------------===// 39// 40// There are several good references for the techniques used in this analysis. 41// 42// Chains of recurrences -- a method to expedite the evaluation 43// of closed-form functions 44// Olaf Bachmann, Paul S. Wang, Eugene V. Zima 45// 46// On computational properties of chains of recurrences 47// Eugene V. Zima 48// 49// Symbolic Evaluation of Chains of Recurrences for Loop Optimization 50// Robert A. van Engelen 51// 52// Efficient Symbolic Analysis for Optimizing Compilers 53// Robert A. van Engelen 54// 55// Using the chains of recurrences algebra for data dependence testing and 56// induction variable substitution 57// MS Thesis, Johnie Birch 58// 59//===----------------------------------------------------------------------===// 60 61#define DEBUG_TYPE "scalar-evolution" 62#include "llvm/Analysis/ScalarEvolutionExpressions.h" 63#include "llvm/Constants.h" 64#include "llvm/DerivedTypes.h" 65#include "llvm/GlobalVariable.h" 66#include "llvm/GlobalAlias.h" 67#include "llvm/Instructions.h" 68#include "llvm/LLVMContext.h" 69#include "llvm/Operator.h" 70#include "llvm/Analysis/ConstantFolding.h" 71#include "llvm/Analysis/Dominators.h" 72#include "llvm/Analysis/InstructionSimplify.h" 73#include "llvm/Analysis/LoopInfo.h" 74#include "llvm/Analysis/ValueTracking.h" 75#include "llvm/Assembly/Writer.h" 76#include "llvm/Target/TargetData.h" 77#include "llvm/Target/TargetLibraryInfo.h" 78#include "llvm/Support/CommandLine.h" 79#include "llvm/Support/ConstantRange.h" 80#include "llvm/Support/Debug.h" 81#include "llvm/Support/ErrorHandling.h" 82#include "llvm/Support/GetElementPtrTypeIterator.h" 83#include "llvm/Support/InstIterator.h" 84#include "llvm/Support/MathExtras.h" 85#include "llvm/Support/raw_ostream.h" 86#include "llvm/ADT/Statistic.h" 87#include "llvm/ADT/STLExtras.h" 88#include "llvm/ADT/SmallPtrSet.h" 89#include <algorithm> 90using namespace llvm; 91 92STATISTIC(NumArrayLenItCounts, 93 "Number of trip counts computed with array length"); 94STATISTIC(NumTripCountsComputed, 95 "Number of loops with predictable loop counts"); 96STATISTIC(NumTripCountsNotComputed, 97 "Number of loops without predictable loop counts"); 98STATISTIC(NumBruteForceTripCountsComputed, 99 "Number of loops with trip counts computed by force"); 100 101static cl::opt<unsigned> 102MaxBruteForceIterations("scalar-evolution-max-iterations", cl::ReallyHidden, 103 cl::desc("Maximum number of iterations SCEV will " 104 "symbolically execute a constant " 105 "derived loop"), 106 cl::init(100)); 107 108INITIALIZE_PASS_BEGIN(ScalarEvolution, "scalar-evolution", 109 "Scalar Evolution Analysis", false, true) 110INITIALIZE_PASS_DEPENDENCY(LoopInfo) 111INITIALIZE_PASS_DEPENDENCY(DominatorTree) 112INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfo) 113INITIALIZE_PASS_END(ScalarEvolution, "scalar-evolution", 114 "Scalar Evolution Analysis", false, true) 115char ScalarEvolution::ID = 0; 116 117//===----------------------------------------------------------------------===// 118// SCEV class definitions 119//===----------------------------------------------------------------------===// 120 121//===----------------------------------------------------------------------===// 122// Implementation of the SCEV class. 123// 124 125void SCEV::dump() const { 126 print(dbgs()); 127 dbgs() << '\n'; 128} 129 130void SCEV::print(raw_ostream &OS) const { 131 switch (getSCEVType()) { 132 case scConstant: 133 WriteAsOperand(OS, cast<SCEVConstant>(this)->getValue(), false); 134 return; 135 case scTruncate: { 136 const SCEVTruncateExpr *Trunc = cast<SCEVTruncateExpr>(this); 137 const SCEV *Op = Trunc->getOperand(); 138 OS << "(trunc " << *Op->getType() << " " << *Op << " to " 139 << *Trunc->getType() << ")"; 140 return; 141 } 142 case scZeroExtend: { 143 const SCEVZeroExtendExpr *ZExt = cast<SCEVZeroExtendExpr>(this); 144 const SCEV *Op = ZExt->getOperand(); 145 OS << "(zext " << *Op->getType() << " " << *Op << " to " 146 << *ZExt->getType() << ")"; 147 return; 148 } 149 case scSignExtend: { 150 const SCEVSignExtendExpr *SExt = cast<SCEVSignExtendExpr>(this); 151 const SCEV *Op = SExt->getOperand(); 152 OS << "(sext " << *Op->getType() << " " << *Op << " to " 153 << *SExt->getType() << ")"; 154 return; 155 } 156 case scAddRecExpr: { 157 const SCEVAddRecExpr *AR = cast<SCEVAddRecExpr>(this); 158 OS << "{" << *AR->getOperand(0); 159 for (unsigned i = 1, e = AR->getNumOperands(); i != e; ++i) 160 OS << ",+," << *AR->getOperand(i); 161 OS << "}<"; 162 if (AR->getNoWrapFlags(FlagNUW)) 163 OS << "nuw><"; 164 if (AR->getNoWrapFlags(FlagNSW)) 165 OS << "nsw><"; 166 if (AR->getNoWrapFlags(FlagNW) && 167 !AR->getNoWrapFlags((NoWrapFlags)(FlagNUW | FlagNSW))) 168 OS << "nw><"; 169 WriteAsOperand(OS, AR->getLoop()->getHeader(), /*PrintType=*/false); 170 OS << ">"; 171 return; 172 } 173 case scAddExpr: 174 case scMulExpr: 175 case scUMaxExpr: 176 case scSMaxExpr: { 177 const SCEVNAryExpr *NAry = cast<SCEVNAryExpr>(this); 178 const char *OpStr = 0; 179 switch (NAry->getSCEVType()) { 180 case scAddExpr: OpStr = " + "; break; 181 case scMulExpr: OpStr = " * "; break; 182 case scUMaxExpr: OpStr = " umax "; break; 183 case scSMaxExpr: OpStr = " smax "; break; 184 } 185 OS << "("; 186 for (SCEVNAryExpr::op_iterator I = NAry->op_begin(), E = NAry->op_end(); 187 I != E; ++I) { 188 OS << **I; 189 if (llvm::next(I) != E) 190 OS << OpStr; 191 } 192 OS << ")"; 193 switch (NAry->getSCEVType()) { 194 case scAddExpr: 195 case scMulExpr: 196 if (NAry->getNoWrapFlags(FlagNUW)) 197 OS << "<nuw>"; 198 if (NAry->getNoWrapFlags(FlagNSW)) 199 OS << "<nsw>"; 200 } 201 return; 202 } 203 case scUDivExpr: { 204 const SCEVUDivExpr *UDiv = cast<SCEVUDivExpr>(this); 205 OS << "(" << *UDiv->getLHS() << " /u " << *UDiv->getRHS() << ")"; 206 return; 207 } 208 case scUnknown: { 209 const SCEVUnknown *U = cast<SCEVUnknown>(this); 210 Type *AllocTy; 211 if (U->isSizeOf(AllocTy)) { 212 OS << "sizeof(" << *AllocTy << ")"; 213 return; 214 } 215 if (U->isAlignOf(AllocTy)) { 216 OS << "alignof(" << *AllocTy << ")"; 217 return; 218 } 219 220 Type *CTy; 221 Constant *FieldNo; 222 if (U->isOffsetOf(CTy, FieldNo)) { 223 OS << "offsetof(" << *CTy << ", "; 224 WriteAsOperand(OS, FieldNo, false); 225 OS << ")"; 226 return; 227 } 228 229 // Otherwise just print it normally. 230 WriteAsOperand(OS, U->getValue(), false); 231 return; 232 } 233 case scCouldNotCompute: 234 OS << "***COULDNOTCOMPUTE***"; 235 return; 236 default: break; 237 } 238 llvm_unreachable("Unknown SCEV kind!"); 239} 240 241Type *SCEV::getType() const { 242 switch (getSCEVType()) { 243 case scConstant: 244 return cast<SCEVConstant>(this)->getType(); 245 case scTruncate: 246 case scZeroExtend: 247 case scSignExtend: 248 return cast<SCEVCastExpr>(this)->getType(); 249 case scAddRecExpr: 250 case scMulExpr: 251 case scUMaxExpr: 252 case scSMaxExpr: 253 return cast<SCEVNAryExpr>(this)->getType(); 254 case scAddExpr: 255 return cast<SCEVAddExpr>(this)->getType(); 256 case scUDivExpr: 257 return cast<SCEVUDivExpr>(this)->getType(); 258 case scUnknown: 259 return cast<SCEVUnknown>(this)->getType(); 260 case scCouldNotCompute: 261 llvm_unreachable("Attempt to use a SCEVCouldNotCompute object!"); 262 default: 263 llvm_unreachable("Unknown SCEV kind!"); 264 } 265} 266 267bool SCEV::isZero() const { 268 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(this)) 269 return SC->getValue()->isZero(); 270 return false; 271} 272 273bool SCEV::isOne() const { 274 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(this)) 275 return SC->getValue()->isOne(); 276 return false; 277} 278 279bool SCEV::isAllOnesValue() const { 280 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(this)) 281 return SC->getValue()->isAllOnesValue(); 282 return false; 283} 284 285/// isNonConstantNegative - Return true if the specified scev is negated, but 286/// not a constant. 287bool SCEV::isNonConstantNegative() const { 288 const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(this); 289 if (!Mul) return false; 290 291 // If there is a constant factor, it will be first. 292 const SCEVConstant *SC = dyn_cast<SCEVConstant>(Mul->getOperand(0)); 293 if (!SC) return false; 294 295 // Return true if the value is negative, this matches things like (-42 * V). 296 return SC->getValue()->getValue().isNegative(); 297} 298 299SCEVCouldNotCompute::SCEVCouldNotCompute() : 300 SCEV(FoldingSetNodeIDRef(), scCouldNotCompute) {} 301 302bool SCEVCouldNotCompute::classof(const SCEV *S) { 303 return S->getSCEVType() == scCouldNotCompute; 304} 305 306const SCEV *ScalarEvolution::getConstant(ConstantInt *V) { 307 FoldingSetNodeID ID; 308 ID.AddInteger(scConstant); 309 ID.AddPointer(V); 310 void *IP = 0; 311 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S; 312 SCEV *S = new (SCEVAllocator) SCEVConstant(ID.Intern(SCEVAllocator), V); 313 UniqueSCEVs.InsertNode(S, IP); 314 return S; 315} 316 317const SCEV *ScalarEvolution::getConstant(const APInt& Val) { 318 return getConstant(ConstantInt::get(getContext(), Val)); 319} 320 321const SCEV * 322ScalarEvolution::getConstant(Type *Ty, uint64_t V, bool isSigned) { 323 IntegerType *ITy = cast<IntegerType>(getEffectiveSCEVType(Ty)); 324 return getConstant(ConstantInt::get(ITy, V, isSigned)); 325} 326 327SCEVCastExpr::SCEVCastExpr(const FoldingSetNodeIDRef ID, 328 unsigned SCEVTy, const SCEV *op, Type *ty) 329 : SCEV(ID, SCEVTy), Op(op), Ty(ty) {} 330 331SCEVTruncateExpr::SCEVTruncateExpr(const FoldingSetNodeIDRef ID, 332 const SCEV *op, Type *ty) 333 : SCEVCastExpr(ID, scTruncate, op, ty) { 334 assert((Op->getType()->isIntegerTy() || Op->getType()->isPointerTy()) && 335 (Ty->isIntegerTy() || Ty->isPointerTy()) && 336 "Cannot truncate non-integer value!"); 337} 338 339SCEVZeroExtendExpr::SCEVZeroExtendExpr(const FoldingSetNodeIDRef ID, 340 const SCEV *op, Type *ty) 341 : SCEVCastExpr(ID, scZeroExtend, op, ty) { 342 assert((Op->getType()->isIntegerTy() || Op->getType()->isPointerTy()) && 343 (Ty->isIntegerTy() || Ty->isPointerTy()) && 344 "Cannot zero extend non-integer value!"); 345} 346 347SCEVSignExtendExpr::SCEVSignExtendExpr(const FoldingSetNodeIDRef ID, 348 const SCEV *op, Type *ty) 349 : SCEVCastExpr(ID, scSignExtend, op, ty) { 350 assert((Op->getType()->isIntegerTy() || Op->getType()->isPointerTy()) && 351 (Ty->isIntegerTy() || Ty->isPointerTy()) && 352 "Cannot sign extend non-integer value!"); 353} 354 355void SCEVUnknown::deleted() { 356 // Clear this SCEVUnknown from various maps. 357 SE->forgetMemoizedResults(this); 358 359 // Remove this SCEVUnknown from the uniquing map. 360 SE->UniqueSCEVs.RemoveNode(this); 361 362 // Release the value. 363 setValPtr(0); 364} 365 366void SCEVUnknown::allUsesReplacedWith(Value *New) { 367 // Clear this SCEVUnknown from various maps. 368 SE->forgetMemoizedResults(this); 369 370 // Remove this SCEVUnknown from the uniquing map. 371 SE->UniqueSCEVs.RemoveNode(this); 372 373 // Update this SCEVUnknown to point to the new value. This is needed 374 // because there may still be outstanding SCEVs which still point to 375 // this SCEVUnknown. 376 setValPtr(New); 377} 378 379bool SCEVUnknown::isSizeOf(Type *&AllocTy) const { 380 if (ConstantExpr *VCE = dyn_cast<ConstantExpr>(getValue())) 381 if (VCE->getOpcode() == Instruction::PtrToInt) 382 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(VCE->getOperand(0))) 383 if (CE->getOpcode() == Instruction::GetElementPtr && 384 CE->getOperand(0)->isNullValue() && 385 CE->getNumOperands() == 2) 386 if (ConstantInt *CI = dyn_cast<ConstantInt>(CE->getOperand(1))) 387 if (CI->isOne()) { 388 AllocTy = cast<PointerType>(CE->getOperand(0)->getType()) 389 ->getElementType(); 390 return true; 391 } 392 393 return false; 394} 395 396bool SCEVUnknown::isAlignOf(Type *&AllocTy) const { 397 if (ConstantExpr *VCE = dyn_cast<ConstantExpr>(getValue())) 398 if (VCE->getOpcode() == Instruction::PtrToInt) 399 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(VCE->getOperand(0))) 400 if (CE->getOpcode() == Instruction::GetElementPtr && 401 CE->getOperand(0)->isNullValue()) { 402 Type *Ty = 403 cast<PointerType>(CE->getOperand(0)->getType())->getElementType(); 404 if (StructType *STy = dyn_cast<StructType>(Ty)) 405 if (!STy->isPacked() && 406 CE->getNumOperands() == 3 && 407 CE->getOperand(1)->isNullValue()) { 408 if (ConstantInt *CI = dyn_cast<ConstantInt>(CE->getOperand(2))) 409 if (CI->isOne() && 410 STy->getNumElements() == 2 && 411 STy->getElementType(0)->isIntegerTy(1)) { 412 AllocTy = STy->getElementType(1); 413 return true; 414 } 415 } 416 } 417 418 return false; 419} 420 421bool SCEVUnknown::isOffsetOf(Type *&CTy, Constant *&FieldNo) const { 422 if (ConstantExpr *VCE = dyn_cast<ConstantExpr>(getValue())) 423 if (VCE->getOpcode() == Instruction::PtrToInt) 424 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(VCE->getOperand(0))) 425 if (CE->getOpcode() == Instruction::GetElementPtr && 426 CE->getNumOperands() == 3 && 427 CE->getOperand(0)->isNullValue() && 428 CE->getOperand(1)->isNullValue()) { 429 Type *Ty = 430 cast<PointerType>(CE->getOperand(0)->getType())->getElementType(); 431 // Ignore vector types here so that ScalarEvolutionExpander doesn't 432 // emit getelementptrs that index into vectors. 433 if (Ty->isStructTy() || Ty->isArrayTy()) { 434 CTy = Ty; 435 FieldNo = CE->getOperand(2); 436 return true; 437 } 438 } 439 440 return false; 441} 442 443//===----------------------------------------------------------------------===// 444// SCEV Utilities 445//===----------------------------------------------------------------------===// 446 447namespace { 448 /// SCEVComplexityCompare - Return true if the complexity of the LHS is less 449 /// than the complexity of the RHS. This comparator is used to canonicalize 450 /// expressions. 451 class SCEVComplexityCompare { 452 const LoopInfo *const LI; 453 public: 454 explicit SCEVComplexityCompare(const LoopInfo *li) : LI(li) {} 455 456 // Return true or false if LHS is less than, or at least RHS, respectively. 457 bool operator()(const SCEV *LHS, const SCEV *RHS) const { 458 return compare(LHS, RHS) < 0; 459 } 460 461 // Return negative, zero, or positive, if LHS is less than, equal to, or 462 // greater than RHS, respectively. A three-way result allows recursive 463 // comparisons to be more efficient. 464 int compare(const SCEV *LHS, const SCEV *RHS) const { 465 // Fast-path: SCEVs are uniqued so we can do a quick equality check. 466 if (LHS == RHS) 467 return 0; 468 469 // Primarily, sort the SCEVs by their getSCEVType(). 470 unsigned LType = LHS->getSCEVType(), RType = RHS->getSCEVType(); 471 if (LType != RType) 472 return (int)LType - (int)RType; 473 474 // Aside from the getSCEVType() ordering, the particular ordering 475 // isn't very important except that it's beneficial to be consistent, 476 // so that (a + b) and (b + a) don't end up as different expressions. 477 switch (LType) { 478 case scUnknown: { 479 const SCEVUnknown *LU = cast<SCEVUnknown>(LHS); 480 const SCEVUnknown *RU = cast<SCEVUnknown>(RHS); 481 482 // Sort SCEVUnknown values with some loose heuristics. TODO: This is 483 // not as complete as it could be. 484 const Value *LV = LU->getValue(), *RV = RU->getValue(); 485 486 // Order pointer values after integer values. This helps SCEVExpander 487 // form GEPs. 488 bool LIsPointer = LV->getType()->isPointerTy(), 489 RIsPointer = RV->getType()->isPointerTy(); 490 if (LIsPointer != RIsPointer) 491 return (int)LIsPointer - (int)RIsPointer; 492 493 // Compare getValueID values. 494 unsigned LID = LV->getValueID(), 495 RID = RV->getValueID(); 496 if (LID != RID) 497 return (int)LID - (int)RID; 498 499 // Sort arguments by their position. 500 if (const Argument *LA = dyn_cast<Argument>(LV)) { 501 const Argument *RA = cast<Argument>(RV); 502 unsigned LArgNo = LA->getArgNo(), RArgNo = RA->getArgNo(); 503 return (int)LArgNo - (int)RArgNo; 504 } 505 506 // For instructions, compare their loop depth, and their operand 507 // count. This is pretty loose. 508 if (const Instruction *LInst = dyn_cast<Instruction>(LV)) { 509 const Instruction *RInst = cast<Instruction>(RV); 510 511 // Compare loop depths. 512 const BasicBlock *LParent = LInst->getParent(), 513 *RParent = RInst->getParent(); 514 if (LParent != RParent) { 515 unsigned LDepth = LI->getLoopDepth(LParent), 516 RDepth = LI->getLoopDepth(RParent); 517 if (LDepth != RDepth) 518 return (int)LDepth - (int)RDepth; 519 } 520 521 // Compare the number of operands. 522 unsigned LNumOps = LInst->getNumOperands(), 523 RNumOps = RInst->getNumOperands(); 524 return (int)LNumOps - (int)RNumOps; 525 } 526 527 return 0; 528 } 529 530 case scConstant: { 531 const SCEVConstant *LC = cast<SCEVConstant>(LHS); 532 const SCEVConstant *RC = cast<SCEVConstant>(RHS); 533 534 // Compare constant values. 535 const APInt &LA = LC->getValue()->getValue(); 536 const APInt &RA = RC->getValue()->getValue(); 537 unsigned LBitWidth = LA.getBitWidth(), RBitWidth = RA.getBitWidth(); 538 if (LBitWidth != RBitWidth) 539 return (int)LBitWidth - (int)RBitWidth; 540 return LA.ult(RA) ? -1 : 1; 541 } 542 543 case scAddRecExpr: { 544 const SCEVAddRecExpr *LA = cast<SCEVAddRecExpr>(LHS); 545 const SCEVAddRecExpr *RA = cast<SCEVAddRecExpr>(RHS); 546 547 // Compare addrec loop depths. 548 const Loop *LLoop = LA->getLoop(), *RLoop = RA->getLoop(); 549 if (LLoop != RLoop) { 550 unsigned LDepth = LLoop->getLoopDepth(), 551 RDepth = RLoop->getLoopDepth(); 552 if (LDepth != RDepth) 553 return (int)LDepth - (int)RDepth; 554 } 555 556 // Addrec complexity grows with operand count. 557 unsigned LNumOps = LA->getNumOperands(), RNumOps = RA->getNumOperands(); 558 if (LNumOps != RNumOps) 559 return (int)LNumOps - (int)RNumOps; 560 561 // Lexicographically compare. 562 for (unsigned i = 0; i != LNumOps; ++i) { 563 long X = compare(LA->getOperand(i), RA->getOperand(i)); 564 if (X != 0) 565 return X; 566 } 567 568 return 0; 569 } 570 571 case scAddExpr: 572 case scMulExpr: 573 case scSMaxExpr: 574 case scUMaxExpr: { 575 const SCEVNAryExpr *LC = cast<SCEVNAryExpr>(LHS); 576 const SCEVNAryExpr *RC = cast<SCEVNAryExpr>(RHS); 577 578 // Lexicographically compare n-ary expressions. 579 unsigned LNumOps = LC->getNumOperands(), RNumOps = RC->getNumOperands(); 580 for (unsigned i = 0; i != LNumOps; ++i) { 581 if (i >= RNumOps) 582 return 1; 583 long X = compare(LC->getOperand(i), RC->getOperand(i)); 584 if (X != 0) 585 return X; 586 } 587 return (int)LNumOps - (int)RNumOps; 588 } 589 590 case scUDivExpr: { 591 const SCEVUDivExpr *LC = cast<SCEVUDivExpr>(LHS); 592 const SCEVUDivExpr *RC = cast<SCEVUDivExpr>(RHS); 593 594 // Lexicographically compare udiv expressions. 595 long X = compare(LC->getLHS(), RC->getLHS()); 596 if (X != 0) 597 return X; 598 return compare(LC->getRHS(), RC->getRHS()); 599 } 600 601 case scTruncate: 602 case scZeroExtend: 603 case scSignExtend: { 604 const SCEVCastExpr *LC = cast<SCEVCastExpr>(LHS); 605 const SCEVCastExpr *RC = cast<SCEVCastExpr>(RHS); 606 607 // Compare cast expressions by operand. 608 return compare(LC->getOperand(), RC->getOperand()); 609 } 610 611 default: 612 llvm_unreachable("Unknown SCEV kind!"); 613 } 614 } 615 }; 616} 617 618/// GroupByComplexity - Given a list of SCEV objects, order them by their 619/// complexity, and group objects of the same complexity together by value. 620/// When this routine is finished, we know that any duplicates in the vector are 621/// consecutive and that complexity is monotonically increasing. 622/// 623/// Note that we go take special precautions to ensure that we get deterministic 624/// results from this routine. In other words, we don't want the results of 625/// this to depend on where the addresses of various SCEV objects happened to 626/// land in memory. 627/// 628static void GroupByComplexity(SmallVectorImpl<const SCEV *> &Ops, 629 LoopInfo *LI) { 630 if (Ops.size() < 2) return; // Noop 631 if (Ops.size() == 2) { 632 // This is the common case, which also happens to be trivially simple. 633 // Special case it. 634 const SCEV *&LHS = Ops[0], *&RHS = Ops[1]; 635 if (SCEVComplexityCompare(LI)(RHS, LHS)) 636 std::swap(LHS, RHS); 637 return; 638 } 639 640 // Do the rough sort by complexity. 641 std::stable_sort(Ops.begin(), Ops.end(), SCEVComplexityCompare(LI)); 642 643 // Now that we are sorted by complexity, group elements of the same 644 // complexity. Note that this is, at worst, N^2, but the vector is likely to 645 // be extremely short in practice. Note that we take this approach because we 646 // do not want to depend on the addresses of the objects we are grouping. 647 for (unsigned i = 0, e = Ops.size(); i != e-2; ++i) { 648 const SCEV *S = Ops[i]; 649 unsigned Complexity = S->getSCEVType(); 650 651 // If there are any objects of the same complexity and same value as this 652 // one, group them. 653 for (unsigned j = i+1; j != e && Ops[j]->getSCEVType() == Complexity; ++j) { 654 if (Ops[j] == S) { // Found a duplicate. 655 // Move it to immediately after i'th element. 656 std::swap(Ops[i+1], Ops[j]); 657 ++i; // no need to rescan it. 658 if (i == e-2) return; // Done! 659 } 660 } 661 } 662} 663 664 665 666//===----------------------------------------------------------------------===// 667// Simple SCEV method implementations 668//===----------------------------------------------------------------------===// 669 670/// BinomialCoefficient - Compute BC(It, K). The result has width W. 671/// Assume, K > 0. 672static const SCEV *BinomialCoefficient(const SCEV *It, unsigned K, 673 ScalarEvolution &SE, 674 Type *ResultTy) { 675 // Handle the simplest case efficiently. 676 if (K == 1) 677 return SE.getTruncateOrZeroExtend(It, ResultTy); 678 679 // We are using the following formula for BC(It, K): 680 // 681 // BC(It, K) = (It * (It - 1) * ... * (It - K + 1)) / K! 682 // 683 // Suppose, W is the bitwidth of the return value. We must be prepared for 684 // overflow. Hence, we must assure that the result of our computation is 685 // equal to the accurate one modulo 2^W. Unfortunately, division isn't 686 // safe in modular arithmetic. 687 // 688 // However, this code doesn't use exactly that formula; the formula it uses 689 // is something like the following, where T is the number of factors of 2 in 690 // K! (i.e. trailing zeros in the binary representation of K!), and ^ is 691 // exponentiation: 692 // 693 // BC(It, K) = (It * (It - 1) * ... * (It - K + 1)) / 2^T / (K! / 2^T) 694 // 695 // This formula is trivially equivalent to the previous formula. However, 696 // this formula can be implemented much more efficiently. The trick is that 697 // K! / 2^T is odd, and exact division by an odd number *is* safe in modular 698 // arithmetic. To do exact division in modular arithmetic, all we have 699 // to do is multiply by the inverse. Therefore, this step can be done at 700 // width W. 701 // 702 // The next issue is how to safely do the division by 2^T. The way this 703 // is done is by doing the multiplication step at a width of at least W + T 704 // bits. This way, the bottom W+T bits of the product are accurate. Then, 705 // when we perform the division by 2^T (which is equivalent to a right shift 706 // by T), the bottom W bits are accurate. Extra bits are okay; they'll get 707 // truncated out after the division by 2^T. 708 // 709 // In comparison to just directly using the first formula, this technique 710 // is much more efficient; using the first formula requires W * K bits, 711 // but this formula less than W + K bits. Also, the first formula requires 712 // a division step, whereas this formula only requires multiplies and shifts. 713 // 714 // It doesn't matter whether the subtraction step is done in the calculation 715 // width or the input iteration count's width; if the subtraction overflows, 716 // the result must be zero anyway. We prefer here to do it in the width of 717 // the induction variable because it helps a lot for certain cases; CodeGen 718 // isn't smart enough to ignore the overflow, which leads to much less 719 // efficient code if the width of the subtraction is wider than the native 720 // register width. 721 // 722 // (It's possible to not widen at all by pulling out factors of 2 before 723 // the multiplication; for example, K=2 can be calculated as 724 // It/2*(It+(It*INT_MIN/INT_MIN)+-1). However, it requires 725 // extra arithmetic, so it's not an obvious win, and it gets 726 // much more complicated for K > 3.) 727 728 // Protection from insane SCEVs; this bound is conservative, 729 // but it probably doesn't matter. 730 if (K > 1000) 731 return SE.getCouldNotCompute(); 732 733 unsigned W = SE.getTypeSizeInBits(ResultTy); 734 735 // Calculate K! / 2^T and T; we divide out the factors of two before 736 // multiplying for calculating K! / 2^T to avoid overflow. 737 // Other overflow doesn't matter because we only care about the bottom 738 // W bits of the result. 739 APInt OddFactorial(W, 1); 740 unsigned T = 1; 741 for (unsigned i = 3; i <= K; ++i) { 742 APInt Mult(W, i); 743 unsigned TwoFactors = Mult.countTrailingZeros(); 744 T += TwoFactors; 745 Mult = Mult.lshr(TwoFactors); 746 OddFactorial *= Mult; 747 } 748 749 // We need at least W + T bits for the multiplication step 750 unsigned CalculationBits = W + T; 751 752 // Calculate 2^T, at width T+W. 753 APInt DivFactor = APInt(CalculationBits, 1).shl(T); 754 755 // Calculate the multiplicative inverse of K! / 2^T; 756 // this multiplication factor will perform the exact division by 757 // K! / 2^T. 758 APInt Mod = APInt::getSignedMinValue(W+1); 759 APInt MultiplyFactor = OddFactorial.zext(W+1); 760 MultiplyFactor = MultiplyFactor.multiplicativeInverse(Mod); 761 MultiplyFactor = MultiplyFactor.trunc(W); 762 763 // Calculate the product, at width T+W 764 IntegerType *CalculationTy = IntegerType::get(SE.getContext(), 765 CalculationBits); 766 const SCEV *Dividend = SE.getTruncateOrZeroExtend(It, CalculationTy); 767 for (unsigned i = 1; i != K; ++i) { 768 const SCEV *S = SE.getMinusSCEV(It, SE.getConstant(It->getType(), i)); 769 Dividend = SE.getMulExpr(Dividend, 770 SE.getTruncateOrZeroExtend(S, CalculationTy)); 771 } 772 773 // Divide by 2^T 774 const SCEV *DivResult = SE.getUDivExpr(Dividend, SE.getConstant(DivFactor)); 775 776 // Truncate the result, and divide by K! / 2^T. 777 778 return SE.getMulExpr(SE.getConstant(MultiplyFactor), 779 SE.getTruncateOrZeroExtend(DivResult, ResultTy)); 780} 781 782/// evaluateAtIteration - Return the value of this chain of recurrences at 783/// the specified iteration number. We can evaluate this recurrence by 784/// multiplying each element in the chain by the binomial coefficient 785/// corresponding to it. In other words, we can evaluate {A,+,B,+,C,+,D} as: 786/// 787/// A*BC(It, 0) + B*BC(It, 1) + C*BC(It, 2) + D*BC(It, 3) 788/// 789/// where BC(It, k) stands for binomial coefficient. 790/// 791const SCEV *SCEVAddRecExpr::evaluateAtIteration(const SCEV *It, 792 ScalarEvolution &SE) const { 793 const SCEV *Result = getStart(); 794 for (unsigned i = 1, e = getNumOperands(); i != e; ++i) { 795 // The computation is correct in the face of overflow provided that the 796 // multiplication is performed _after_ the evaluation of the binomial 797 // coefficient. 798 const SCEV *Coeff = BinomialCoefficient(It, i, SE, getType()); 799 if (isa<SCEVCouldNotCompute>(Coeff)) 800 return Coeff; 801 802 Result = SE.getAddExpr(Result, SE.getMulExpr(getOperand(i), Coeff)); 803 } 804 return Result; 805} 806 807//===----------------------------------------------------------------------===// 808// SCEV Expression folder implementations 809//===----------------------------------------------------------------------===// 810 811const SCEV *ScalarEvolution::getTruncateExpr(const SCEV *Op, 812 Type *Ty) { 813 assert(getTypeSizeInBits(Op->getType()) > getTypeSizeInBits(Ty) && 814 "This is not a truncating conversion!"); 815 assert(isSCEVable(Ty) && 816 "This is not a conversion to a SCEVable type!"); 817 Ty = getEffectiveSCEVType(Ty); 818 819 FoldingSetNodeID ID; 820 ID.AddInteger(scTruncate); 821 ID.AddPointer(Op); 822 ID.AddPointer(Ty); 823 void *IP = 0; 824 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S; 825 826 // Fold if the operand is constant. 827 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(Op)) 828 return getConstant( 829 cast<ConstantInt>(ConstantExpr::getTrunc(SC->getValue(), 830 getEffectiveSCEVType(Ty)))); 831 832 // trunc(trunc(x)) --> trunc(x) 833 if (const SCEVTruncateExpr *ST = dyn_cast<SCEVTruncateExpr>(Op)) 834 return getTruncateExpr(ST->getOperand(), Ty); 835 836 // trunc(sext(x)) --> sext(x) if widening or trunc(x) if narrowing 837 if (const SCEVSignExtendExpr *SS = dyn_cast<SCEVSignExtendExpr>(Op)) 838 return getTruncateOrSignExtend(SS->getOperand(), Ty); 839 840 // trunc(zext(x)) --> zext(x) if widening or trunc(x) if narrowing 841 if (const SCEVZeroExtendExpr *SZ = dyn_cast<SCEVZeroExtendExpr>(Op)) 842 return getTruncateOrZeroExtend(SZ->getOperand(), Ty); 843 844 // trunc(x1+x2+...+xN) --> trunc(x1)+trunc(x2)+...+trunc(xN) if we can 845 // eliminate all the truncates. 846 if (const SCEVAddExpr *SA = dyn_cast<SCEVAddExpr>(Op)) { 847 SmallVector<const SCEV *, 4> Operands; 848 bool hasTrunc = false; 849 for (unsigned i = 0, e = SA->getNumOperands(); i != e && !hasTrunc; ++i) { 850 const SCEV *S = getTruncateExpr(SA->getOperand(i), Ty); 851 hasTrunc = isa<SCEVTruncateExpr>(S); 852 Operands.push_back(S); 853 } 854 if (!hasTrunc) 855 return getAddExpr(Operands); 856 UniqueSCEVs.FindNodeOrInsertPos(ID, IP); // Mutates IP, returns NULL. 857 } 858 859 // trunc(x1*x2*...*xN) --> trunc(x1)*trunc(x2)*...*trunc(xN) if we can 860 // eliminate all the truncates. 861 if (const SCEVMulExpr *SM = dyn_cast<SCEVMulExpr>(Op)) { 862 SmallVector<const SCEV *, 4> Operands; 863 bool hasTrunc = false; 864 for (unsigned i = 0, e = SM->getNumOperands(); i != e && !hasTrunc; ++i) { 865 const SCEV *S = getTruncateExpr(SM->getOperand(i), Ty); 866 hasTrunc = isa<SCEVTruncateExpr>(S); 867 Operands.push_back(S); 868 } 869 if (!hasTrunc) 870 return getMulExpr(Operands); 871 UniqueSCEVs.FindNodeOrInsertPos(ID, IP); // Mutates IP, returns NULL. 872 } 873 874 // If the input value is a chrec scev, truncate the chrec's operands. 875 if (const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(Op)) { 876 SmallVector<const SCEV *, 4> Operands; 877 for (unsigned i = 0, e = AddRec->getNumOperands(); i != e; ++i) 878 Operands.push_back(getTruncateExpr(AddRec->getOperand(i), Ty)); 879 return getAddRecExpr(Operands, AddRec->getLoop(), SCEV::FlagAnyWrap); 880 } 881 882 // As a special case, fold trunc(undef) to undef. We don't want to 883 // know too much about SCEVUnknowns, but this special case is handy 884 // and harmless. 885 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(Op)) 886 if (isa<UndefValue>(U->getValue())) 887 return getSCEV(UndefValue::get(Ty)); 888 889 // The cast wasn't folded; create an explicit cast node. We can reuse 890 // the existing insert position since if we get here, we won't have 891 // made any changes which would invalidate it. 892 SCEV *S = new (SCEVAllocator) SCEVTruncateExpr(ID.Intern(SCEVAllocator), 893 Op, Ty); 894 UniqueSCEVs.InsertNode(S, IP); 895 return S; 896} 897 898const SCEV *ScalarEvolution::getZeroExtendExpr(const SCEV *Op, 899 Type *Ty) { 900 assert(getTypeSizeInBits(Op->getType()) < getTypeSizeInBits(Ty) && 901 "This is not an extending conversion!"); 902 assert(isSCEVable(Ty) && 903 "This is not a conversion to a SCEVable type!"); 904 Ty = getEffectiveSCEVType(Ty); 905 906 // Fold if the operand is constant. 907 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(Op)) 908 return getConstant( 909 cast<ConstantInt>(ConstantExpr::getZExt(SC->getValue(), 910 getEffectiveSCEVType(Ty)))); 911 912 // zext(zext(x)) --> zext(x) 913 if (const SCEVZeroExtendExpr *SZ = dyn_cast<SCEVZeroExtendExpr>(Op)) 914 return getZeroExtendExpr(SZ->getOperand(), Ty); 915 916 // Before doing any expensive analysis, check to see if we've already 917 // computed a SCEV for this Op and Ty. 918 FoldingSetNodeID ID; 919 ID.AddInteger(scZeroExtend); 920 ID.AddPointer(Op); 921 ID.AddPointer(Ty); 922 void *IP = 0; 923 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S; 924 925 // zext(trunc(x)) --> zext(x) or x or trunc(x) 926 if (const SCEVTruncateExpr *ST = dyn_cast<SCEVTruncateExpr>(Op)) { 927 // It's possible the bits taken off by the truncate were all zero bits. If 928 // so, we should be able to simplify this further. 929 const SCEV *X = ST->getOperand(); 930 ConstantRange CR = getUnsignedRange(X); 931 unsigned TruncBits = getTypeSizeInBits(ST->getType()); 932 unsigned NewBits = getTypeSizeInBits(Ty); 933 if (CR.truncate(TruncBits).zeroExtend(NewBits).contains( 934 CR.zextOrTrunc(NewBits))) 935 return getTruncateOrZeroExtend(X, Ty); 936 } 937 938 // If the input value is a chrec scev, and we can prove that the value 939 // did not overflow the old, smaller, value, we can zero extend all of the 940 // operands (often constants). This allows analysis of something like 941 // this: for (unsigned char X = 0; X < 100; ++X) { int Y = X; } 942 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Op)) 943 if (AR->isAffine()) { 944 const SCEV *Start = AR->getStart(); 945 const SCEV *Step = AR->getStepRecurrence(*this); 946 unsigned BitWidth = getTypeSizeInBits(AR->getType()); 947 const Loop *L = AR->getLoop(); 948 949 // If we have special knowledge that this addrec won't overflow, 950 // we don't need to do any further analysis. 951 if (AR->getNoWrapFlags(SCEV::FlagNUW)) 952 return getAddRecExpr(getZeroExtendExpr(Start, Ty), 953 getZeroExtendExpr(Step, Ty), 954 L, AR->getNoWrapFlags()); 955 956 // Check whether the backedge-taken count is SCEVCouldNotCompute. 957 // Note that this serves two purposes: It filters out loops that are 958 // simply not analyzable, and it covers the case where this code is 959 // being called from within backedge-taken count analysis, such that 960 // attempting to ask for the backedge-taken count would likely result 961 // in infinite recursion. In the later case, the analysis code will 962 // cope with a conservative value, and it will take care to purge 963 // that value once it has finished. 964 const SCEV *MaxBECount = getMaxBackedgeTakenCount(L); 965 if (!isa<SCEVCouldNotCompute>(MaxBECount)) { 966 // Manually compute the final value for AR, checking for 967 // overflow. 968 969 // Check whether the backedge-taken count can be losslessly casted to 970 // the addrec's type. The count is always unsigned. 971 const SCEV *CastedMaxBECount = 972 getTruncateOrZeroExtend(MaxBECount, Start->getType()); 973 const SCEV *RecastedMaxBECount = 974 getTruncateOrZeroExtend(CastedMaxBECount, MaxBECount->getType()); 975 if (MaxBECount == RecastedMaxBECount) { 976 Type *WideTy = IntegerType::get(getContext(), BitWidth * 2); 977 // Check whether Start+Step*MaxBECount has no unsigned overflow. 978 const SCEV *ZMul = getMulExpr(CastedMaxBECount, Step); 979 const SCEV *Add = getAddExpr(Start, ZMul); 980 const SCEV *OperandExtendedAdd = 981 getAddExpr(getZeroExtendExpr(Start, WideTy), 982 getMulExpr(getZeroExtendExpr(CastedMaxBECount, WideTy), 983 getZeroExtendExpr(Step, WideTy))); 984 if (getZeroExtendExpr(Add, WideTy) == OperandExtendedAdd) { 985 // Cache knowledge of AR NUW, which is propagated to this AddRec. 986 const_cast<SCEVAddRecExpr *>(AR)->setNoWrapFlags(SCEV::FlagNUW); 987 // Return the expression with the addrec on the outside. 988 return getAddRecExpr(getZeroExtendExpr(Start, Ty), 989 getZeroExtendExpr(Step, Ty), 990 L, AR->getNoWrapFlags()); 991 } 992 // Similar to above, only this time treat the step value as signed. 993 // This covers loops that count down. 994 const SCEV *SMul = getMulExpr(CastedMaxBECount, Step); 995 Add = getAddExpr(Start, SMul); 996 OperandExtendedAdd = 997 getAddExpr(getZeroExtendExpr(Start, WideTy), 998 getMulExpr(getZeroExtendExpr(CastedMaxBECount, WideTy), 999 getSignExtendExpr(Step, WideTy))); 1000 if (getZeroExtendExpr(Add, WideTy) == OperandExtendedAdd) { 1001 // Cache knowledge of AR NW, which is propagated to this AddRec. 1002 // Negative step causes unsigned wrap, but it still can't self-wrap. 1003 const_cast<SCEVAddRecExpr *>(AR)->setNoWrapFlags(SCEV::FlagNW); 1004 // Return the expression with the addrec on the outside. 1005 return getAddRecExpr(getZeroExtendExpr(Start, Ty), 1006 getSignExtendExpr(Step, Ty), 1007 L, AR->getNoWrapFlags()); 1008 } 1009 } 1010 1011 // If the backedge is guarded by a comparison with the pre-inc value 1012 // the addrec is safe. Also, if the entry is guarded by a comparison 1013 // with the start value and the backedge is guarded by a comparison 1014 // with the post-inc value, the addrec is safe. 1015 if (isKnownPositive(Step)) { 1016 const SCEV *N = getConstant(APInt::getMinValue(BitWidth) - 1017 getUnsignedRange(Step).getUnsignedMax()); 1018 if (isLoopBackedgeGuardedByCond(L, ICmpInst::ICMP_ULT, AR, N) || 1019 (isLoopEntryGuardedByCond(L, ICmpInst::ICMP_ULT, Start, N) && 1020 isLoopBackedgeGuardedByCond(L, ICmpInst::ICMP_ULT, 1021 AR->getPostIncExpr(*this), N))) { 1022 // Cache knowledge of AR NUW, which is propagated to this AddRec. 1023 const_cast<SCEVAddRecExpr *>(AR)->setNoWrapFlags(SCEV::FlagNUW); 1024 // Return the expression with the addrec on the outside. 1025 return getAddRecExpr(getZeroExtendExpr(Start, Ty), 1026 getZeroExtendExpr(Step, Ty), 1027 L, AR->getNoWrapFlags()); 1028 } 1029 } else if (isKnownNegative(Step)) { 1030 const SCEV *N = getConstant(APInt::getMaxValue(BitWidth) - 1031 getSignedRange(Step).getSignedMin()); 1032 if (isLoopBackedgeGuardedByCond(L, ICmpInst::ICMP_UGT, AR, N) || 1033 (isLoopEntryGuardedByCond(L, ICmpInst::ICMP_UGT, Start, N) && 1034 isLoopBackedgeGuardedByCond(L, ICmpInst::ICMP_UGT, 1035 AR->getPostIncExpr(*this), N))) { 1036 // Cache knowledge of AR NW, which is propagated to this AddRec. 1037 // Negative step causes unsigned wrap, but it still can't self-wrap. 1038 const_cast<SCEVAddRecExpr *>(AR)->setNoWrapFlags(SCEV::FlagNW); 1039 // Return the expression with the addrec on the outside. 1040 return getAddRecExpr(getZeroExtendExpr(Start, Ty), 1041 getSignExtendExpr(Step, Ty), 1042 L, AR->getNoWrapFlags()); 1043 } 1044 } 1045 } 1046 } 1047 1048 // The cast wasn't folded; create an explicit cast node. 1049 // Recompute the insert position, as it may have been invalidated. 1050 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S; 1051 SCEV *S = new (SCEVAllocator) SCEVZeroExtendExpr(ID.Intern(SCEVAllocator), 1052 Op, Ty); 1053 UniqueSCEVs.InsertNode(S, IP); 1054 return S; 1055} 1056 1057// Get the limit of a recurrence such that incrementing by Step cannot cause 1058// signed overflow as long as the value of the recurrence within the loop does 1059// not exceed this limit before incrementing. 1060static const SCEV *getOverflowLimitForStep(const SCEV *Step, 1061 ICmpInst::Predicate *Pred, 1062 ScalarEvolution *SE) { 1063 unsigned BitWidth = SE->getTypeSizeInBits(Step->getType()); 1064 if (SE->isKnownPositive(Step)) { 1065 *Pred = ICmpInst::ICMP_SLT; 1066 return SE->getConstant(APInt::getSignedMinValue(BitWidth) - 1067 SE->getSignedRange(Step).getSignedMax()); 1068 } 1069 if (SE->isKnownNegative(Step)) { 1070 *Pred = ICmpInst::ICMP_SGT; 1071 return SE->getConstant(APInt::getSignedMaxValue(BitWidth) - 1072 SE->getSignedRange(Step).getSignedMin()); 1073 } 1074 return 0; 1075} 1076 1077// The recurrence AR has been shown to have no signed wrap. Typically, if we can 1078// prove NSW for AR, then we can just as easily prove NSW for its preincrement 1079// or postincrement sibling. This allows normalizing a sign extended AddRec as 1080// such: {sext(Step + Start),+,Step} => {(Step + sext(Start),+,Step} As a 1081// result, the expression "Step + sext(PreIncAR)" is congruent with 1082// "sext(PostIncAR)" 1083static const SCEV *getPreStartForSignExtend(const SCEVAddRecExpr *AR, 1084 Type *Ty, 1085 ScalarEvolution *SE) { 1086 const Loop *L = AR->getLoop(); 1087 const SCEV *Start = AR->getStart(); 1088 const SCEV *Step = AR->getStepRecurrence(*SE); 1089 1090 // Check for a simple looking step prior to loop entry. 1091 const SCEVAddExpr *SA = dyn_cast<SCEVAddExpr>(Start); 1092 if (!SA) 1093 return 0; 1094 1095 // Create an AddExpr for "PreStart" after subtracting Step. Full SCEV 1096 // subtraction is expensive. For this purpose, perform a quick and dirty 1097 // difference, by checking for Step in the operand list. 1098 SmallVector<const SCEV *, 4> DiffOps; 1099 for (SCEVAddExpr::op_iterator I = SA->op_begin(), E = SA->op_end(); 1100 I != E; ++I) { 1101 if (*I != Step) 1102 DiffOps.push_back(*I); 1103 } 1104 if (DiffOps.size() == SA->getNumOperands()) 1105 return 0; 1106 1107 // This is a postinc AR. Check for overflow on the preinc recurrence using the 1108 // same three conditions that getSignExtendedExpr checks. 1109 1110 // 1. NSW flags on the step increment. 1111 const SCEV *PreStart = SE->getAddExpr(DiffOps, SA->getNoWrapFlags()); 1112 const SCEVAddRecExpr *PreAR = dyn_cast<SCEVAddRecExpr>( 1113 SE->getAddRecExpr(PreStart, Step, L, SCEV::FlagAnyWrap)); 1114 1115 if (PreAR && PreAR->getNoWrapFlags(SCEV::FlagNSW)) 1116 return PreStart; 1117 1118 // 2. Direct overflow check on the step operation's expression. 1119 unsigned BitWidth = SE->getTypeSizeInBits(AR->getType()); 1120 Type *WideTy = IntegerType::get(SE->getContext(), BitWidth * 2); 1121 const SCEV *OperandExtendedStart = 1122 SE->getAddExpr(SE->getSignExtendExpr(PreStart, WideTy), 1123 SE->getSignExtendExpr(Step, WideTy)); 1124 if (SE->getSignExtendExpr(Start, WideTy) == OperandExtendedStart) { 1125 // Cache knowledge of PreAR NSW. 1126 if (PreAR) 1127 const_cast<SCEVAddRecExpr *>(PreAR)->setNoWrapFlags(SCEV::FlagNSW); 1128 // FIXME: this optimization needs a unit test 1129 DEBUG(dbgs() << "SCEV: untested prestart overflow check\n"); 1130 return PreStart; 1131 } 1132 1133 // 3. Loop precondition. 1134 ICmpInst::Predicate Pred; 1135 const SCEV *OverflowLimit = getOverflowLimitForStep(Step, &Pred, SE); 1136 1137 if (OverflowLimit && 1138 SE->isLoopEntryGuardedByCond(L, Pred, PreStart, OverflowLimit)) { 1139 return PreStart; 1140 } 1141 return 0; 1142} 1143 1144// Get the normalized sign-extended expression for this AddRec's Start. 1145static const SCEV *getSignExtendAddRecStart(const SCEVAddRecExpr *AR, 1146 Type *Ty, 1147 ScalarEvolution *SE) { 1148 const SCEV *PreStart = getPreStartForSignExtend(AR, Ty, SE); 1149 if (!PreStart) 1150 return SE->getSignExtendExpr(AR->getStart(), Ty); 1151 1152 return SE->getAddExpr(SE->getSignExtendExpr(AR->getStepRecurrence(*SE), Ty), 1153 SE->getSignExtendExpr(PreStart, Ty)); 1154} 1155 1156const SCEV *ScalarEvolution::getSignExtendExpr(const SCEV *Op, 1157 Type *Ty) { 1158 assert(getTypeSizeInBits(Op->getType()) < getTypeSizeInBits(Ty) && 1159 "This is not an extending conversion!"); 1160 assert(isSCEVable(Ty) && 1161 "This is not a conversion to a SCEVable type!"); 1162 Ty = getEffectiveSCEVType(Ty); 1163 1164 // Fold if the operand is constant. 1165 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(Op)) 1166 return getConstant( 1167 cast<ConstantInt>(ConstantExpr::getSExt(SC->getValue(), 1168 getEffectiveSCEVType(Ty)))); 1169 1170 // sext(sext(x)) --> sext(x) 1171 if (const SCEVSignExtendExpr *SS = dyn_cast<SCEVSignExtendExpr>(Op)) 1172 return getSignExtendExpr(SS->getOperand(), Ty); 1173 1174 // sext(zext(x)) --> zext(x) 1175 if (const SCEVZeroExtendExpr *SZ = dyn_cast<SCEVZeroExtendExpr>(Op)) 1176 return getZeroExtendExpr(SZ->getOperand(), Ty); 1177 1178 // Before doing any expensive analysis, check to see if we've already 1179 // computed a SCEV for this Op and Ty. 1180 FoldingSetNodeID ID; 1181 ID.AddInteger(scSignExtend); 1182 ID.AddPointer(Op); 1183 ID.AddPointer(Ty); 1184 void *IP = 0; 1185 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S; 1186 1187 // If the input value is provably positive, build a zext instead. 1188 if (isKnownNonNegative(Op)) 1189 return getZeroExtendExpr(Op, Ty); 1190 1191 // sext(trunc(x)) --> sext(x) or x or trunc(x) 1192 if (const SCEVTruncateExpr *ST = dyn_cast<SCEVTruncateExpr>(Op)) { 1193 // It's possible the bits taken off by the truncate were all sign bits. If 1194 // so, we should be able to simplify this further. 1195 const SCEV *X = ST->getOperand(); 1196 ConstantRange CR = getSignedRange(X); 1197 unsigned TruncBits = getTypeSizeInBits(ST->getType()); 1198 unsigned NewBits = getTypeSizeInBits(Ty); 1199 if (CR.truncate(TruncBits).signExtend(NewBits).contains( 1200 CR.sextOrTrunc(NewBits))) 1201 return getTruncateOrSignExtend(X, Ty); 1202 } 1203 1204 // If the input value is a chrec scev, and we can prove that the value 1205 // did not overflow the old, smaller, value, we can sign extend all of the 1206 // operands (often constants). This allows analysis of something like 1207 // this: for (signed char X = 0; X < 100; ++X) { int Y = X; } 1208 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Op)) 1209 if (AR->isAffine()) { 1210 const SCEV *Start = AR->getStart(); 1211 const SCEV *Step = AR->getStepRecurrence(*this); 1212 unsigned BitWidth = getTypeSizeInBits(AR->getType()); 1213 const Loop *L = AR->getLoop(); 1214 1215 // If we have special knowledge that this addrec won't overflow, 1216 // we don't need to do any further analysis. 1217 if (AR->getNoWrapFlags(SCEV::FlagNSW)) 1218 return getAddRecExpr(getSignExtendAddRecStart(AR, Ty, this), 1219 getSignExtendExpr(Step, Ty), 1220 L, SCEV::FlagNSW); 1221 1222 // Check whether the backedge-taken count is SCEVCouldNotCompute. 1223 // Note that this serves two purposes: It filters out loops that are 1224 // simply not analyzable, and it covers the case where this code is 1225 // being called from within backedge-taken count analysis, such that 1226 // attempting to ask for the backedge-taken count would likely result 1227 // in infinite recursion. In the later case, the analysis code will 1228 // cope with a conservative value, and it will take care to purge 1229 // that value once it has finished. 1230 const SCEV *MaxBECount = getMaxBackedgeTakenCount(L); 1231 if (!isa<SCEVCouldNotCompute>(MaxBECount)) { 1232 // Manually compute the final value for AR, checking for 1233 // overflow. 1234 1235 // Check whether the backedge-taken count can be losslessly casted to 1236 // the addrec's type. The count is always unsigned. 1237 const SCEV *CastedMaxBECount = 1238 getTruncateOrZeroExtend(MaxBECount, Start->getType()); 1239 const SCEV *RecastedMaxBECount = 1240 getTruncateOrZeroExtend(CastedMaxBECount, MaxBECount->getType()); 1241 if (MaxBECount == RecastedMaxBECount) { 1242 Type *WideTy = IntegerType::get(getContext(), BitWidth * 2); 1243 // Check whether Start+Step*MaxBECount has no signed overflow. 1244 const SCEV *SMul = getMulExpr(CastedMaxBECount, Step); 1245 const SCEV *Add = getAddExpr(Start, SMul); 1246 const SCEV *OperandExtendedAdd = 1247 getAddExpr(getSignExtendExpr(Start, WideTy), 1248 getMulExpr(getZeroExtendExpr(CastedMaxBECount, WideTy), 1249 getSignExtendExpr(Step, WideTy))); 1250 if (getSignExtendExpr(Add, WideTy) == OperandExtendedAdd) { 1251 // Cache knowledge of AR NSW, which is propagated to this AddRec. 1252 const_cast<SCEVAddRecExpr *>(AR)->setNoWrapFlags(SCEV::FlagNSW); 1253 // Return the expression with the addrec on the outside. 1254 return getAddRecExpr(getSignExtendAddRecStart(AR, Ty, this), 1255 getSignExtendExpr(Step, Ty), 1256 L, AR->getNoWrapFlags()); 1257 } 1258 // Similar to above, only this time treat the step value as unsigned. 1259 // This covers loops that count up with an unsigned step. 1260 const SCEV *UMul = getMulExpr(CastedMaxBECount, Step); 1261 Add = getAddExpr(Start, UMul); 1262 OperandExtendedAdd = 1263 getAddExpr(getSignExtendExpr(Start, WideTy), 1264 getMulExpr(getZeroExtendExpr(CastedMaxBECount, WideTy), 1265 getZeroExtendExpr(Step, WideTy))); 1266 if (getSignExtendExpr(Add, WideTy) == OperandExtendedAdd) { 1267 // Cache knowledge of AR NSW, which is propagated to this AddRec. 1268 const_cast<SCEVAddRecExpr *>(AR)->setNoWrapFlags(SCEV::FlagNSW); 1269 // Return the expression with the addrec on the outside. 1270 return getAddRecExpr(getSignExtendAddRecStart(AR, Ty, this), 1271 getZeroExtendExpr(Step, Ty), 1272 L, AR->getNoWrapFlags()); 1273 } 1274 } 1275 1276 // If the backedge is guarded by a comparison with the pre-inc value 1277 // the addrec is safe. Also, if the entry is guarded by a comparison 1278 // with the start value and the backedge is guarded by a comparison 1279 // with the post-inc value, the addrec is safe. 1280 ICmpInst::Predicate Pred; 1281 const SCEV *OverflowLimit = getOverflowLimitForStep(Step, &Pred, this); 1282 if (OverflowLimit && 1283 (isLoopBackedgeGuardedByCond(L, Pred, AR, OverflowLimit) || 1284 (isLoopEntryGuardedByCond(L, Pred, Start, OverflowLimit) && 1285 isLoopBackedgeGuardedByCond(L, Pred, AR->getPostIncExpr(*this), 1286 OverflowLimit)))) { 1287 // Cache knowledge of AR NSW, then propagate NSW to the wide AddRec. 1288 const_cast<SCEVAddRecExpr *>(AR)->setNoWrapFlags(SCEV::FlagNSW); 1289 return getAddRecExpr(getSignExtendAddRecStart(AR, Ty, this), 1290 getSignExtendExpr(Step, Ty), 1291 L, AR->getNoWrapFlags()); 1292 } 1293 } 1294 } 1295 1296 // The cast wasn't folded; create an explicit cast node. 1297 // Recompute the insert position, as it may have been invalidated. 1298 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S; 1299 SCEV *S = new (SCEVAllocator) SCEVSignExtendExpr(ID.Intern(SCEVAllocator), 1300 Op, Ty); 1301 UniqueSCEVs.InsertNode(S, IP); 1302 return S; 1303} 1304 1305/// getAnyExtendExpr - Return a SCEV for the given operand extended with 1306/// unspecified bits out to the given type. 1307/// 1308const SCEV *ScalarEvolution::getAnyExtendExpr(const SCEV *Op, 1309 Type *Ty) { 1310 assert(getTypeSizeInBits(Op->getType()) < getTypeSizeInBits(Ty) && 1311 "This is not an extending conversion!"); 1312 assert(isSCEVable(Ty) && 1313 "This is not a conversion to a SCEVable type!"); 1314 Ty = getEffectiveSCEVType(Ty); 1315 1316 // Sign-extend negative constants. 1317 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(Op)) 1318 if (SC->getValue()->getValue().isNegative()) 1319 return getSignExtendExpr(Op, Ty); 1320 1321 // Peel off a truncate cast. 1322 if (const SCEVTruncateExpr *T = dyn_cast<SCEVTruncateExpr>(Op)) { 1323 const SCEV *NewOp = T->getOperand(); 1324 if (getTypeSizeInBits(NewOp->getType()) < getTypeSizeInBits(Ty)) 1325 return getAnyExtendExpr(NewOp, Ty); 1326 return getTruncateOrNoop(NewOp, Ty); 1327 } 1328 1329 // Next try a zext cast. If the cast is folded, use it. 1330 const SCEV *ZExt = getZeroExtendExpr(Op, Ty); 1331 if (!isa<SCEVZeroExtendExpr>(ZExt)) 1332 return ZExt; 1333 1334 // Next try a sext cast. If the cast is folded, use it. 1335 const SCEV *SExt = getSignExtendExpr(Op, Ty); 1336 if (!isa<SCEVSignExtendExpr>(SExt)) 1337 return SExt; 1338 1339 // Force the cast to be folded into the operands of an addrec. 1340 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Op)) { 1341 SmallVector<const SCEV *, 4> Ops; 1342 for (SCEVAddRecExpr::op_iterator I = AR->op_begin(), E = AR->op_end(); 1343 I != E; ++I) 1344 Ops.push_back(getAnyExtendExpr(*I, Ty)); 1345 return getAddRecExpr(Ops, AR->getLoop(), SCEV::FlagNW); 1346 } 1347 1348 // As a special case, fold anyext(undef) to undef. We don't want to 1349 // know too much about SCEVUnknowns, but this special case is handy 1350 // and harmless. 1351 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(Op)) 1352 if (isa<UndefValue>(U->getValue())) 1353 return getSCEV(UndefValue::get(Ty)); 1354 1355 // If the expression is obviously signed, use the sext cast value. 1356 if (isa<SCEVSMaxExpr>(Op)) 1357 return SExt; 1358 1359 // Absent any other information, use the zext cast value. 1360 return ZExt; 1361} 1362 1363/// CollectAddOperandsWithScales - Process the given Ops list, which is 1364/// a list of operands to be added under the given scale, update the given 1365/// map. This is a helper function for getAddRecExpr. As an example of 1366/// what it does, given a sequence of operands that would form an add 1367/// expression like this: 1368/// 1369/// m + n + 13 + (A * (o + p + (B * q + m + 29))) + r + (-1 * r) 1370/// 1371/// where A and B are constants, update the map with these values: 1372/// 1373/// (m, 1+A*B), (n, 1), (o, A), (p, A), (q, A*B), (r, 0) 1374/// 1375/// and add 13 + A*B*29 to AccumulatedConstant. 1376/// This will allow getAddRecExpr to produce this: 1377/// 1378/// 13+A*B*29 + n + (m * (1+A*B)) + ((o + p) * A) + (q * A*B) 1379/// 1380/// This form often exposes folding opportunities that are hidden in 1381/// the original operand list. 1382/// 1383/// Return true iff it appears that any interesting folding opportunities 1384/// may be exposed. This helps getAddRecExpr short-circuit extra work in 1385/// the common case where no interesting opportunities are present, and 1386/// is also used as a check to avoid infinite recursion. 1387/// 1388static bool 1389CollectAddOperandsWithScales(DenseMap<const SCEV *, APInt> &M, 1390 SmallVector<const SCEV *, 8> &NewOps, 1391 APInt &AccumulatedConstant, 1392 const SCEV *const *Ops, size_t NumOperands, 1393 const APInt &Scale, 1394 ScalarEvolution &SE) { 1395 bool Interesting = false; 1396 1397 // Iterate over the add operands. They are sorted, with constants first. 1398 unsigned i = 0; 1399 while (const SCEVConstant *C = dyn_cast<SCEVConstant>(Ops[i])) { 1400 ++i; 1401 // Pull a buried constant out to the outside. 1402 if (Scale != 1 || AccumulatedConstant != 0 || C->getValue()->isZero()) 1403 Interesting = true; 1404 AccumulatedConstant += Scale * C->getValue()->getValue(); 1405 } 1406 1407 // Next comes everything else. We're especially interested in multiplies 1408 // here, but they're in the middle, so just visit the rest with one loop. 1409 for (; i != NumOperands; ++i) { 1410 const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(Ops[i]); 1411 if (Mul && isa<SCEVConstant>(Mul->getOperand(0))) { 1412 APInt NewScale = 1413 Scale * cast<SCEVConstant>(Mul->getOperand(0))->getValue()->getValue(); 1414 if (Mul->getNumOperands() == 2 && isa<SCEVAddExpr>(Mul->getOperand(1))) { 1415 // A multiplication of a constant with another add; recurse. 1416 const SCEVAddExpr *Add = cast<SCEVAddExpr>(Mul->getOperand(1)); 1417 Interesting |= 1418 CollectAddOperandsWithScales(M, NewOps, AccumulatedConstant, 1419 Add->op_begin(), Add->getNumOperands(), 1420 NewScale, SE); 1421 } else { 1422 // A multiplication of a constant with some other value. Update 1423 // the map. 1424 SmallVector<const SCEV *, 4> MulOps(Mul->op_begin()+1, Mul->op_end()); 1425 const SCEV *Key = SE.getMulExpr(MulOps); 1426 std::pair<DenseMap<const SCEV *, APInt>::iterator, bool> Pair = 1427 M.insert(std::make_pair(Key, NewScale)); 1428 if (Pair.second) { 1429 NewOps.push_back(Pair.first->first); 1430 } else { 1431 Pair.first->second += NewScale; 1432 // The map already had an entry for this value, which may indicate 1433 // a folding opportunity. 1434 Interesting = true; 1435 } 1436 } 1437 } else { 1438 // An ordinary operand. Update the map. 1439 std::pair<DenseMap<const SCEV *, APInt>::iterator, bool> Pair = 1440 M.insert(std::make_pair(Ops[i], Scale)); 1441 if (Pair.second) { 1442 NewOps.push_back(Pair.first->first); 1443 } else { 1444 Pair.first->second += Scale; 1445 // The map already had an entry for this value, which may indicate 1446 // a folding opportunity. 1447 Interesting = true; 1448 } 1449 } 1450 } 1451 1452 return Interesting; 1453} 1454 1455namespace { 1456 struct APIntCompare { 1457 bool operator()(const APInt &LHS, const APInt &RHS) const { 1458 return LHS.ult(RHS); 1459 } 1460 }; 1461} 1462 1463/// getAddExpr - Get a canonical add expression, or something simpler if 1464/// possible. 1465const SCEV *ScalarEvolution::getAddExpr(SmallVectorImpl<const SCEV *> &Ops, 1466 SCEV::NoWrapFlags Flags) { 1467 assert(!(Flags & ~(SCEV::FlagNUW | SCEV::FlagNSW)) && 1468 "only nuw or nsw allowed"); 1469 assert(!Ops.empty() && "Cannot get empty add!"); 1470 if (Ops.size() == 1) return Ops[0]; 1471#ifndef NDEBUG 1472 Type *ETy = getEffectiveSCEVType(Ops[0]->getType()); 1473 for (unsigned i = 1, e = Ops.size(); i != e; ++i) 1474 assert(getEffectiveSCEVType(Ops[i]->getType()) == ETy && 1475 "SCEVAddExpr operand types don't match!"); 1476#endif 1477 1478 // If FlagNSW is true and all the operands are non-negative, infer FlagNUW. 1479 // And vice-versa. 1480 int SignOrUnsignMask = SCEV::FlagNUW | SCEV::FlagNSW; 1481 SCEV::NoWrapFlags SignOrUnsignWrap = maskFlags(Flags, SignOrUnsignMask); 1482 if (SignOrUnsignWrap && (SignOrUnsignWrap != SignOrUnsignMask)) { 1483 bool All = true; 1484 for (SmallVectorImpl<const SCEV *>::const_iterator I = Ops.begin(), 1485 E = Ops.end(); I != E; ++I) 1486 if (!isKnownNonNegative(*I)) { 1487 All = false; 1488 break; 1489 } 1490 if (All) Flags = setFlags(Flags, (SCEV::NoWrapFlags)SignOrUnsignMask); 1491 } 1492 1493 // Sort by complexity, this groups all similar expression types together. 1494 GroupByComplexity(Ops, LI); 1495 1496 // If there are any constants, fold them together. 1497 unsigned Idx = 0; 1498 if (const SCEVConstant *LHSC = dyn_cast<SCEVConstant>(Ops[0])) { 1499 ++Idx; 1500 assert(Idx < Ops.size()); 1501 while (const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(Ops[Idx])) { 1502 // We found two constants, fold them together! 1503 Ops[0] = getConstant(LHSC->getValue()->getValue() + 1504 RHSC->getValue()->getValue()); 1505 if (Ops.size() == 2) return Ops[0]; 1506 Ops.erase(Ops.begin()+1); // Erase the folded element 1507 LHSC = cast<SCEVConstant>(Ops[0]); 1508 } 1509 1510 // If we are left with a constant zero being added, strip it off. 1511 if (LHSC->getValue()->isZero()) { 1512 Ops.erase(Ops.begin()); 1513 --Idx; 1514 } 1515 1516 if (Ops.size() == 1) return Ops[0]; 1517 } 1518 1519 // Okay, check to see if the same value occurs in the operand list more than 1520 // once. If so, merge them together into an multiply expression. Since we 1521 // sorted the list, these values are required to be adjacent. 1522 Type *Ty = Ops[0]->getType(); 1523 bool FoundMatch = false; 1524 for (unsigned i = 0, e = Ops.size(); i != e-1; ++i) 1525 if (Ops[i] == Ops[i+1]) { // X + Y + Y --> X + Y*2 1526 // Scan ahead to count how many equal operands there are. 1527 unsigned Count = 2; 1528 while (i+Count != e && Ops[i+Count] == Ops[i]) 1529 ++Count; 1530 // Merge the values into a multiply. 1531 const SCEV *Scale = getConstant(Ty, Count); 1532 const SCEV *Mul = getMulExpr(Scale, Ops[i]); 1533 if (Ops.size() == Count) 1534 return Mul; 1535 Ops[i] = Mul; 1536 Ops.erase(Ops.begin()+i+1, Ops.begin()+i+Count); 1537 --i; e -= Count - 1; 1538 FoundMatch = true; 1539 } 1540 if (FoundMatch) 1541 return getAddExpr(Ops, Flags); 1542 1543 // Check for truncates. If all the operands are truncated from the same 1544 // type, see if factoring out the truncate would permit the result to be 1545 // folded. eg., trunc(x) + m*trunc(n) --> trunc(x + trunc(m)*n) 1546 // if the contents of the resulting outer trunc fold to something simple. 1547 for (; Idx < Ops.size() && isa<SCEVTruncateExpr>(Ops[Idx]); ++Idx) { 1548 const SCEVTruncateExpr *Trunc = cast<SCEVTruncateExpr>(Ops[Idx]); 1549 Type *DstType = Trunc->getType(); 1550 Type *SrcType = Trunc->getOperand()->getType(); 1551 SmallVector<const SCEV *, 8> LargeOps; 1552 bool Ok = true; 1553 // Check all the operands to see if they can be represented in the 1554 // source type of the truncate. 1555 for (unsigned i = 0, e = Ops.size(); i != e; ++i) { 1556 if (const SCEVTruncateExpr *T = dyn_cast<SCEVTruncateExpr>(Ops[i])) { 1557 if (T->getOperand()->getType() != SrcType) { 1558 Ok = false; 1559 break; 1560 } 1561 LargeOps.push_back(T->getOperand()); 1562 } else if (const SCEVConstant *C = dyn_cast<SCEVConstant>(Ops[i])) { 1563 LargeOps.push_back(getAnyExtendExpr(C, SrcType)); 1564 } else if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(Ops[i])) { 1565 SmallVector<const SCEV *, 8> LargeMulOps; 1566 for (unsigned j = 0, f = M->getNumOperands(); j != f && Ok; ++j) { 1567 if (const SCEVTruncateExpr *T = 1568 dyn_cast<SCEVTruncateExpr>(M->getOperand(j))) { 1569 if (T->getOperand()->getType() != SrcType) { 1570 Ok = false; 1571 break; 1572 } 1573 LargeMulOps.push_back(T->getOperand()); 1574 } else if (const SCEVConstant *C = 1575 dyn_cast<SCEVConstant>(M->getOperand(j))) { 1576 LargeMulOps.push_back(getAnyExtendExpr(C, SrcType)); 1577 } else { 1578 Ok = false; 1579 break; 1580 } 1581 } 1582 if (Ok) 1583 LargeOps.push_back(getMulExpr(LargeMulOps)); 1584 } else { 1585 Ok = false; 1586 break; 1587 } 1588 } 1589 if (Ok) { 1590 // Evaluate the expression in the larger type. 1591 const SCEV *Fold = getAddExpr(LargeOps, Flags); 1592 // If it folds to something simple, use it. Otherwise, don't. 1593 if (isa<SCEVConstant>(Fold) || isa<SCEVUnknown>(Fold)) 1594 return getTruncateExpr(Fold, DstType); 1595 } 1596 } 1597 1598 // Skip past any other cast SCEVs. 1599 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scAddExpr) 1600 ++Idx; 1601 1602 // If there are add operands they would be next. 1603 if (Idx < Ops.size()) { 1604 bool DeletedAdd = false; 1605 while (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Ops[Idx])) { 1606 // If we have an add, expand the add operands onto the end of the operands 1607 // list. 1608 Ops.erase(Ops.begin()+Idx); 1609 Ops.append(Add->op_begin(), Add->op_end()); 1610 DeletedAdd = true; 1611 } 1612 1613 // If we deleted at least one add, we added operands to the end of the list, 1614 // and they are not necessarily sorted. Recurse to resort and resimplify 1615 // any operands we just acquired. 1616 if (DeletedAdd) 1617 return getAddExpr(Ops); 1618 } 1619 1620 // Skip over the add expression until we get to a multiply. 1621 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scMulExpr) 1622 ++Idx; 1623 1624 // Check to see if there are any folding opportunities present with 1625 // operands multiplied by constant values. 1626 if (Idx < Ops.size() && isa<SCEVMulExpr>(Ops[Idx])) { 1627 uint64_t BitWidth = getTypeSizeInBits(Ty); 1628 DenseMap<const SCEV *, APInt> M; 1629 SmallVector<const SCEV *, 8> NewOps; 1630 APInt AccumulatedConstant(BitWidth, 0); 1631 if (CollectAddOperandsWithScales(M, NewOps, AccumulatedConstant, 1632 Ops.data(), Ops.size(), 1633 APInt(BitWidth, 1), *this)) { 1634 // Some interesting folding opportunity is present, so its worthwhile to 1635 // re-generate the operands list. Group the operands by constant scale, 1636 // to avoid multiplying by the same constant scale multiple times. 1637 std::map<APInt, SmallVector<const SCEV *, 4>, APIntCompare> MulOpLists; 1638 for (SmallVector<const SCEV *, 8>::const_iterator I = NewOps.begin(), 1639 E = NewOps.end(); I != E; ++I) 1640 MulOpLists[M.find(*I)->second].push_back(*I); 1641 // Re-generate the operands list. 1642 Ops.clear(); 1643 if (AccumulatedConstant != 0) 1644 Ops.push_back(getConstant(AccumulatedConstant)); 1645 for (std::map<APInt, SmallVector<const SCEV *, 4>, APIntCompare>::iterator 1646 I = MulOpLists.begin(), E = MulOpLists.end(); I != E; ++I) 1647 if (I->first != 0) 1648 Ops.push_back(getMulExpr(getConstant(I->first), 1649 getAddExpr(I->second))); 1650 if (Ops.empty()) 1651 return getConstant(Ty, 0); 1652 if (Ops.size() == 1) 1653 return Ops[0]; 1654 return getAddExpr(Ops); 1655 } 1656 } 1657 1658 // If we are adding something to a multiply expression, make sure the 1659 // something is not already an operand of the multiply. If so, merge it into 1660 // the multiply. 1661 for (; Idx < Ops.size() && isa<SCEVMulExpr>(Ops[Idx]); ++Idx) { 1662 const SCEVMulExpr *Mul = cast<SCEVMulExpr>(Ops[Idx]); 1663 for (unsigned MulOp = 0, e = Mul->getNumOperands(); MulOp != e; ++MulOp) { 1664 const SCEV *MulOpSCEV = Mul->getOperand(MulOp); 1665 if (isa<SCEVConstant>(MulOpSCEV)) 1666 continue; 1667 for (unsigned AddOp = 0, e = Ops.size(); AddOp != e; ++AddOp) 1668 if (MulOpSCEV == Ops[AddOp]) { 1669 // Fold W + X + (X * Y * Z) --> W + (X * ((Y*Z)+1)) 1670 const SCEV *InnerMul = Mul->getOperand(MulOp == 0); 1671 if (Mul->getNumOperands() != 2) { 1672 // If the multiply has more than two operands, we must get the 1673 // Y*Z term. 1674 SmallVector<const SCEV *, 4> MulOps(Mul->op_begin(), 1675 Mul->op_begin()+MulOp); 1676 MulOps.append(Mul->op_begin()+MulOp+1, Mul->op_end()); 1677 InnerMul = getMulExpr(MulOps); 1678 } 1679 const SCEV *One = getConstant(Ty, 1); 1680 const SCEV *AddOne = getAddExpr(One, InnerMul); 1681 const SCEV *OuterMul = getMulExpr(AddOne, MulOpSCEV); 1682 if (Ops.size() == 2) return OuterMul; 1683 if (AddOp < Idx) { 1684 Ops.erase(Ops.begin()+AddOp); 1685 Ops.erase(Ops.begin()+Idx-1); 1686 } else { 1687 Ops.erase(Ops.begin()+Idx); 1688 Ops.erase(Ops.begin()+AddOp-1); 1689 } 1690 Ops.push_back(OuterMul); 1691 return getAddExpr(Ops); 1692 } 1693 1694 // Check this multiply against other multiplies being added together. 1695 for (unsigned OtherMulIdx = Idx+1; 1696 OtherMulIdx < Ops.size() && isa<SCEVMulExpr>(Ops[OtherMulIdx]); 1697 ++OtherMulIdx) { 1698 const SCEVMulExpr *OtherMul = cast<SCEVMulExpr>(Ops[OtherMulIdx]); 1699 // If MulOp occurs in OtherMul, we can fold the two multiplies 1700 // together. 1701 for (unsigned OMulOp = 0, e = OtherMul->getNumOperands(); 1702 OMulOp != e; ++OMulOp) 1703 if (OtherMul->getOperand(OMulOp) == MulOpSCEV) { 1704 // Fold X + (A*B*C) + (A*D*E) --> X + (A*(B*C+D*E)) 1705 const SCEV *InnerMul1 = Mul->getOperand(MulOp == 0); 1706 if (Mul->getNumOperands() != 2) { 1707 SmallVector<const SCEV *, 4> MulOps(Mul->op_begin(), 1708 Mul->op_begin()+MulOp); 1709 MulOps.append(Mul->op_begin()+MulOp+1, Mul->op_end()); 1710 InnerMul1 = getMulExpr(MulOps); 1711 } 1712 const SCEV *InnerMul2 = OtherMul->getOperand(OMulOp == 0); 1713 if (OtherMul->getNumOperands() != 2) { 1714 SmallVector<const SCEV *, 4> MulOps(OtherMul->op_begin(), 1715 OtherMul->op_begin()+OMulOp); 1716 MulOps.append(OtherMul->op_begin()+OMulOp+1, OtherMul->op_end()); 1717 InnerMul2 = getMulExpr(MulOps); 1718 } 1719 const SCEV *InnerMulSum = getAddExpr(InnerMul1,InnerMul2); 1720 const SCEV *OuterMul = getMulExpr(MulOpSCEV, InnerMulSum); 1721 if (Ops.size() == 2) return OuterMul; 1722 Ops.erase(Ops.begin()+Idx); 1723 Ops.erase(Ops.begin()+OtherMulIdx-1); 1724 Ops.push_back(OuterMul); 1725 return getAddExpr(Ops); 1726 } 1727 } 1728 } 1729 } 1730 1731 // If there are any add recurrences in the operands list, see if any other 1732 // added values are loop invariant. If so, we can fold them into the 1733 // recurrence. 1734 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scAddRecExpr) 1735 ++Idx; 1736 1737 // Scan over all recurrences, trying to fold loop invariants into them. 1738 for (; Idx < Ops.size() && isa<SCEVAddRecExpr>(Ops[Idx]); ++Idx) { 1739 // Scan all of the other operands to this add and add them to the vector if 1740 // they are loop invariant w.r.t. the recurrence. 1741 SmallVector<const SCEV *, 8> LIOps; 1742 const SCEVAddRecExpr *AddRec = cast<SCEVAddRecExpr>(Ops[Idx]); 1743 const Loop *AddRecLoop = AddRec->getLoop(); 1744 for (unsigned i = 0, e = Ops.size(); i != e; ++i) 1745 if (isLoopInvariant(Ops[i], AddRecLoop)) { 1746 LIOps.push_back(Ops[i]); 1747 Ops.erase(Ops.begin()+i); 1748 --i; --e; 1749 } 1750 1751 // If we found some loop invariants, fold them into the recurrence. 1752 if (!LIOps.empty()) { 1753 // NLI + LI + {Start,+,Step} --> NLI + {LI+Start,+,Step} 1754 LIOps.push_back(AddRec->getStart()); 1755 1756 SmallVector<const SCEV *, 4> AddRecOps(AddRec->op_begin(), 1757 AddRec->op_end()); 1758 AddRecOps[0] = getAddExpr(LIOps); 1759 1760 // Build the new addrec. Propagate the NUW and NSW flags if both the 1761 // outer add and the inner addrec are guaranteed to have no overflow. 1762 // Always propagate NW. 1763 Flags = AddRec->getNoWrapFlags(setFlags(Flags, SCEV::FlagNW)); 1764 const SCEV *NewRec = getAddRecExpr(AddRecOps, AddRecLoop, Flags); 1765 1766 // If all of the other operands were loop invariant, we are done. 1767 if (Ops.size() == 1) return NewRec; 1768 1769 // Otherwise, add the folded AddRec by the non-invariant parts. 1770 for (unsigned i = 0;; ++i) 1771 if (Ops[i] == AddRec) { 1772 Ops[i] = NewRec; 1773 break; 1774 } 1775 return getAddExpr(Ops); 1776 } 1777 1778 // Okay, if there weren't any loop invariants to be folded, check to see if 1779 // there are multiple AddRec's with the same loop induction variable being 1780 // added together. If so, we can fold them. 1781 for (unsigned OtherIdx = Idx+1; 1782 OtherIdx < Ops.size() && isa<SCEVAddRecExpr>(Ops[OtherIdx]); 1783 ++OtherIdx) 1784 if (AddRecLoop == cast<SCEVAddRecExpr>(Ops[OtherIdx])->getLoop()) { 1785 // Other + {A,+,B}<L> + {C,+,D}<L> --> Other + {A+C,+,B+D}<L> 1786 SmallVector<const SCEV *, 4> AddRecOps(AddRec->op_begin(), 1787 AddRec->op_end()); 1788 for (; OtherIdx != Ops.size() && isa<SCEVAddRecExpr>(Ops[OtherIdx]); 1789 ++OtherIdx) 1790 if (const SCEVAddRecExpr *OtherAddRec = 1791 dyn_cast<SCEVAddRecExpr>(Ops[OtherIdx])) 1792 if (OtherAddRec->getLoop() == AddRecLoop) { 1793 for (unsigned i = 0, e = OtherAddRec->getNumOperands(); 1794 i != e; ++i) { 1795 if (i >= AddRecOps.size()) { 1796 AddRecOps.append(OtherAddRec->op_begin()+i, 1797 OtherAddRec->op_end()); 1798 break; 1799 } 1800 AddRecOps[i] = getAddExpr(AddRecOps[i], 1801 OtherAddRec->getOperand(i)); 1802 } 1803 Ops.erase(Ops.begin() + OtherIdx); --OtherIdx; 1804 } 1805 // Step size has changed, so we cannot guarantee no self-wraparound. 1806 Ops[Idx] = getAddRecExpr(AddRecOps, AddRecLoop, SCEV::FlagAnyWrap); 1807 return getAddExpr(Ops); 1808 } 1809 1810 // Otherwise couldn't fold anything into this recurrence. Move onto the 1811 // next one. 1812 } 1813 1814 // Okay, it looks like we really DO need an add expr. Check to see if we 1815 // already have one, otherwise create a new one. 1816 FoldingSetNodeID ID; 1817 ID.AddInteger(scAddExpr); 1818 for (unsigned i = 0, e = Ops.size(); i != e; ++i) 1819 ID.AddPointer(Ops[i]); 1820 void *IP = 0; 1821 SCEVAddExpr *S = 1822 static_cast<SCEVAddExpr *>(UniqueSCEVs.FindNodeOrInsertPos(ID, IP)); 1823 if (!S) { 1824 const SCEV **O = SCEVAllocator.Allocate<const SCEV *>(Ops.size()); 1825 std::uninitialized_copy(Ops.begin(), Ops.end(), O); 1826 S = new (SCEVAllocator) SCEVAddExpr(ID.Intern(SCEVAllocator), 1827 O, Ops.size()); 1828 UniqueSCEVs.InsertNode(S, IP); 1829 } 1830 S->setNoWrapFlags(Flags); 1831 return S; 1832} 1833 1834static uint64_t umul_ov(uint64_t i, uint64_t j, bool &Overflow) { 1835 uint64_t k = i*j; 1836 if (j > 1 && k / j != i) Overflow = true; 1837 return k; 1838} 1839 1840/// Compute the result of "n choose k", the binomial coefficient. If an 1841/// intermediate computation overflows, Overflow will be set and the return will 1842/// be garbage. Overflow is not cleared on absense of overflow. 1843static uint64_t Choose(uint64_t n, uint64_t k, bool &Overflow) { 1844 // We use the multiplicative formula: 1845 // n(n-1)(n-2)...(n-(k-1)) / k(k-1)(k-2)...1 . 1846 // At each iteration, we take the n-th term of the numeral and divide by the 1847 // (k-n)th term of the denominator. This division will always produce an 1848 // integral result, and helps reduce the chance of overflow in the 1849 // intermediate computations. However, we can still overflow even when the 1850 // final result would fit. 1851 1852 if (n == 0 || n == k) return 1; 1853 if (k > n) return 0; 1854 1855 if (k > n/2) 1856 k = n-k; 1857 1858 uint64_t r = 1; 1859 for (uint64_t i = 1; i <= k; ++i) { 1860 r = umul_ov(r, n-(i-1), Overflow); 1861 r /= i; 1862 } 1863 return r; 1864} 1865 1866/// getMulExpr - Get a canonical multiply expression, or something simpler if 1867/// possible. 1868const SCEV *ScalarEvolution::getMulExpr(SmallVectorImpl<const SCEV *> &Ops, 1869 SCEV::NoWrapFlags Flags) { 1870 assert(Flags == maskFlags(Flags, SCEV::FlagNUW | SCEV::FlagNSW) && 1871 "only nuw or nsw allowed"); 1872 assert(!Ops.empty() && "Cannot get empty mul!"); 1873 if (Ops.size() == 1) return Ops[0]; 1874#ifndef NDEBUG 1875 Type *ETy = getEffectiveSCEVType(Ops[0]->getType()); 1876 for (unsigned i = 1, e = Ops.size(); i != e; ++i) 1877 assert(getEffectiveSCEVType(Ops[i]->getType()) == ETy && 1878 "SCEVMulExpr operand types don't match!"); 1879#endif 1880 1881 // If FlagNSW is true and all the operands are non-negative, infer FlagNUW. 1882 // And vice-versa. 1883 int SignOrUnsignMask = SCEV::FlagNUW | SCEV::FlagNSW; 1884 SCEV::NoWrapFlags SignOrUnsignWrap = maskFlags(Flags, SignOrUnsignMask); 1885 if (SignOrUnsignWrap && (SignOrUnsignWrap != SignOrUnsignMask)) { 1886 bool All = true; 1887 for (SmallVectorImpl<const SCEV *>::const_iterator I = Ops.begin(), 1888 E = Ops.end(); I != E; ++I) 1889 if (!isKnownNonNegative(*I)) { 1890 All = false; 1891 break; 1892 } 1893 if (All) Flags = setFlags(Flags, (SCEV::NoWrapFlags)SignOrUnsignMask); 1894 } 1895 1896 // Sort by complexity, this groups all similar expression types together. 1897 GroupByComplexity(Ops, LI); 1898 1899 // If there are any constants, fold them together. 1900 unsigned Idx = 0; 1901 if (const SCEVConstant *LHSC = dyn_cast<SCEVConstant>(Ops[0])) { 1902 1903 // C1*(C2+V) -> C1*C2 + C1*V 1904 if (Ops.size() == 2) 1905 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Ops[1])) 1906 if (Add->getNumOperands() == 2 && 1907 isa<SCEVConstant>(Add->getOperand(0))) 1908 return getAddExpr(getMulExpr(LHSC, Add->getOperand(0)), 1909 getMulExpr(LHSC, Add->getOperand(1))); 1910 1911 ++Idx; 1912 while (const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(Ops[Idx])) { 1913 // We found two constants, fold them together! 1914 ConstantInt *Fold = ConstantInt::get(getContext(), 1915 LHSC->getValue()->getValue() * 1916 RHSC->getValue()->getValue()); 1917 Ops[0] = getConstant(Fold); 1918 Ops.erase(Ops.begin()+1); // Erase the folded element 1919 if (Ops.size() == 1) return Ops[0]; 1920 LHSC = cast<SCEVConstant>(Ops[0]); 1921 } 1922 1923 // If we are left with a constant one being multiplied, strip it off. 1924 if (cast<SCEVConstant>(Ops[0])->getValue()->equalsInt(1)) { 1925 Ops.erase(Ops.begin()); 1926 --Idx; 1927 } else if (cast<SCEVConstant>(Ops[0])->getValue()->isZero()) { 1928 // If we have a multiply of zero, it will always be zero. 1929 return Ops[0]; 1930 } else if (Ops[0]->isAllOnesValue()) { 1931 // If we have a mul by -1 of an add, try distributing the -1 among the 1932 // add operands. 1933 if (Ops.size() == 2) { 1934 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(Ops[1])) { 1935 SmallVector<const SCEV *, 4> NewOps; 1936 bool AnyFolded = false; 1937 for (SCEVAddRecExpr::op_iterator I = Add->op_begin(), 1938 E = Add->op_end(); I != E; ++I) { 1939 const SCEV *Mul = getMulExpr(Ops[0], *I); 1940 if (!isa<SCEVMulExpr>(Mul)) AnyFolded = true; 1941 NewOps.push_back(Mul); 1942 } 1943 if (AnyFolded) 1944 return getAddExpr(NewOps); 1945 } 1946 else if (const SCEVAddRecExpr * 1947 AddRec = dyn_cast<SCEVAddRecExpr>(Ops[1])) { 1948 // Negation preserves a recurrence's no self-wrap property. 1949 SmallVector<const SCEV *, 4> Operands; 1950 for (SCEVAddRecExpr::op_iterator I = AddRec->op_begin(), 1951 E = AddRec->op_end(); I != E; ++I) { 1952 Operands.push_back(getMulExpr(Ops[0], *I)); 1953 } 1954 return getAddRecExpr(Operands, AddRec->getLoop(), 1955 AddRec->getNoWrapFlags(SCEV::FlagNW)); 1956 } 1957 } 1958 } 1959 1960 if (Ops.size() == 1) 1961 return Ops[0]; 1962 } 1963 1964 // Skip over the add expression until we get to a multiply. 1965 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scMulExpr) 1966 ++Idx; 1967 1968 // If there are mul operands inline them all into this expression. 1969 if (Idx < Ops.size()) { 1970 bool DeletedMul = false; 1971 while (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(Ops[Idx])) { 1972 // If we have an mul, expand the mul operands onto the end of the operands 1973 // list. 1974 Ops.erase(Ops.begin()+Idx); 1975 Ops.append(Mul->op_begin(), Mul->op_end()); 1976 DeletedMul = true; 1977 } 1978 1979 // If we deleted at least one mul, we added operands to the end of the list, 1980 // and they are not necessarily sorted. Recurse to resort and resimplify 1981 // any operands we just acquired. 1982 if (DeletedMul) 1983 return getMulExpr(Ops); 1984 } 1985 1986 // If there are any add recurrences in the operands list, see if any other 1987 // added values are loop invariant. If so, we can fold them into the 1988 // recurrence. 1989 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scAddRecExpr) 1990 ++Idx; 1991 1992 // Scan over all recurrences, trying to fold loop invariants into them. 1993 for (; Idx < Ops.size() && isa<SCEVAddRecExpr>(Ops[Idx]); ++Idx) { 1994 // Scan all of the other operands to this mul and add them to the vector if 1995 // they are loop invariant w.r.t. the recurrence. 1996 SmallVector<const SCEV *, 8> LIOps; 1997 const SCEVAddRecExpr *AddRec = cast<SCEVAddRecExpr>(Ops[Idx]); 1998 const Loop *AddRecLoop = AddRec->getLoop(); 1999 for (unsigned i = 0, e = Ops.size(); i != e; ++i) 2000 if (isLoopInvariant(Ops[i], AddRecLoop)) { 2001 LIOps.push_back(Ops[i]); 2002 Ops.erase(Ops.begin()+i); 2003 --i; --e; 2004 } 2005 2006 // If we found some loop invariants, fold them into the recurrence. 2007 if (!LIOps.empty()) { 2008 // NLI * LI * {Start,+,Step} --> NLI * {LI*Start,+,LI*Step} 2009 SmallVector<const SCEV *, 4> NewOps; 2010 NewOps.reserve(AddRec->getNumOperands()); 2011 const SCEV *Scale = getMulExpr(LIOps); 2012 for (unsigned i = 0, e = AddRec->getNumOperands(); i != e; ++i) 2013 NewOps.push_back(getMulExpr(Scale, AddRec->getOperand(i))); 2014 2015 // Build the new addrec. Propagate the NUW and NSW flags if both the 2016 // outer mul and the inner addrec are guaranteed to have no overflow. 2017 // 2018 // No self-wrap cannot be guaranteed after changing the step size, but 2019 // will be inferred if either NUW or NSW is true. 2020 Flags = AddRec->getNoWrapFlags(clearFlags(Flags, SCEV::FlagNW)); 2021 const SCEV *NewRec = getAddRecExpr(NewOps, AddRecLoop, Flags); 2022 2023 // If all of the other operands were loop invariant, we are done. 2024 if (Ops.size() == 1) return NewRec; 2025 2026 // Otherwise, multiply the folded AddRec by the non-invariant parts. 2027 for (unsigned i = 0;; ++i) 2028 if (Ops[i] == AddRec) { 2029 Ops[i] = NewRec; 2030 break; 2031 } 2032 return getMulExpr(Ops); 2033 } 2034 2035 // Okay, if there weren't any loop invariants to be folded, check to see if 2036 // there are multiple AddRec's with the same loop induction variable being 2037 // multiplied together. If so, we can fold them. 2038 for (unsigned OtherIdx = Idx+1; 2039 OtherIdx < Ops.size() && isa<SCEVAddRecExpr>(Ops[OtherIdx]); 2040 ++OtherIdx) { 2041 if (AddRecLoop == cast<SCEVAddRecExpr>(Ops[OtherIdx])->getLoop()) { 2042 // {A1,+,A2,+,...,+,An}<L> * {B1,+,B2,+,...,+,Bn}<L> 2043 // = {x=1 in [ sum y=x..2x [ sum z=max(y-x, y-n)..min(x,n) [ 2044 // choose(x, 2x)*choose(2x-y, x-z)*A_{y-z}*B_z 2045 // ]]],+,...up to x=2n}. 2046 // Note that the arguments to choose() are always integers with values 2047 // known at compile time, never SCEV objects. 2048 // 2049 // The implementation avoids pointless extra computations when the two 2050 // addrec's are of different length (mathematically, it's equivalent to 2051 // an infinite stream of zeros on the right). 2052 bool OpsModified = false; 2053 for (; OtherIdx != Ops.size() && isa<SCEVAddRecExpr>(Ops[OtherIdx]); 2054 ++OtherIdx) 2055 if (const SCEVAddRecExpr *OtherAddRec = 2056 dyn_cast<SCEVAddRecExpr>(Ops[OtherIdx])) 2057 if (OtherAddRec->getLoop() == AddRecLoop) { 2058 bool Overflow = false; 2059 Type *Ty = AddRec->getType(); 2060 bool LargerThan64Bits = getTypeSizeInBits(Ty) > 64; 2061 SmallVector<const SCEV*, 7> AddRecOps; 2062 for (int x = 0, xe = AddRec->getNumOperands() + 2063 OtherAddRec->getNumOperands() - 1; 2064 x != xe && !Overflow; ++x) { 2065 const SCEV *Term = getConstant(Ty, 0); 2066 for (int y = x, ye = 2*x+1; y != ye && !Overflow; ++y) { 2067 uint64_t Coeff1 = Choose(x, 2*x - y, Overflow); 2068 for (int z = std::max(y-x, y-(int)AddRec->getNumOperands()+1), 2069 ze = std::min(x+1, (int)OtherAddRec->getNumOperands()); 2070 z < ze && !Overflow; ++z) { 2071 uint64_t Coeff2 = Choose(2*x - y, x-z, Overflow); 2072 uint64_t Coeff; 2073 if (LargerThan64Bits) 2074 Coeff = umul_ov(Coeff1, Coeff2, Overflow); 2075 else 2076 Coeff = Coeff1*Coeff2; 2077 const SCEV *CoeffTerm = getConstant(Ty, Coeff); 2078 const SCEV *Term1 = AddRec->getOperand(y-z); 2079 const SCEV *Term2 = OtherAddRec->getOperand(z); 2080 Term = getAddExpr(Term, getMulExpr(CoeffTerm, Term1,Term2)); 2081 } 2082 } 2083 AddRecOps.push_back(Term); 2084 } 2085 if (!Overflow) { 2086 const SCEV *NewAddRec = getAddRecExpr(AddRecOps, 2087 AddRec->getLoop(), 2088 SCEV::FlagAnyWrap); 2089 if (Ops.size() == 2) return NewAddRec; 2090 Ops[Idx] = AddRec = cast<SCEVAddRecExpr>(NewAddRec); 2091 Ops.erase(Ops.begin() + OtherIdx); --OtherIdx; 2092 OpsModified = true; 2093 } 2094 } 2095 if (OpsModified) 2096 return getMulExpr(Ops); 2097 } 2098 } 2099 2100 // Otherwise couldn't fold anything into this recurrence. Move onto the 2101 // next one. 2102 } 2103 2104 // Okay, it looks like we really DO need an mul expr. Check to see if we 2105 // already have one, otherwise create a new one. 2106 FoldingSetNodeID ID; 2107 ID.AddInteger(scMulExpr); 2108 for (unsigned i = 0, e = Ops.size(); i != e; ++i) 2109 ID.AddPointer(Ops[i]); 2110 void *IP = 0; 2111 SCEVMulExpr *S = 2112 static_cast<SCEVMulExpr *>(UniqueSCEVs.FindNodeOrInsertPos(ID, IP)); 2113 if (!S) { 2114 const SCEV **O = SCEVAllocator.Allocate<const SCEV *>(Ops.size()); 2115 std::uninitialized_copy(Ops.begin(), Ops.end(), O); 2116 S = new (SCEVAllocator) SCEVMulExpr(ID.Intern(SCEVAllocator), 2117 O, Ops.size()); 2118 UniqueSCEVs.InsertNode(S, IP); 2119 } 2120 S->setNoWrapFlags(Flags); 2121 return S; 2122} 2123 2124/// getUDivExpr - Get a canonical unsigned division expression, or something 2125/// simpler if possible. 2126const SCEV *ScalarEvolution::getUDivExpr(const SCEV *LHS, 2127 const SCEV *RHS) { 2128 assert(getEffectiveSCEVType(LHS->getType()) == 2129 getEffectiveSCEVType(RHS->getType()) && 2130 "SCEVUDivExpr operand types don't match!"); 2131 2132 if (const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(RHS)) { 2133 if (RHSC->getValue()->equalsInt(1)) 2134 return LHS; // X udiv 1 --> x 2135 // If the denominator is zero, the result of the udiv is undefined. Don't 2136 // try to analyze it, because the resolution chosen here may differ from 2137 // the resolution chosen in other parts of the compiler. 2138 if (!RHSC->getValue()->isZero()) { 2139 // Determine if the division can be folded into the operands of 2140 // its operands. 2141 // TODO: Generalize this to non-constants by using known-bits information. 2142 Type *Ty = LHS->getType(); 2143 unsigned LZ = RHSC->getValue()->getValue().countLeadingZeros(); 2144 unsigned MaxShiftAmt = getTypeSizeInBits(Ty) - LZ - 1; 2145 // For non-power-of-two values, effectively round the value up to the 2146 // nearest power of two. 2147 if (!RHSC->getValue()->getValue().isPowerOf2()) 2148 ++MaxShiftAmt; 2149 IntegerType *ExtTy = 2150 IntegerType::get(getContext(), getTypeSizeInBits(Ty) + MaxShiftAmt); 2151 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(LHS)) 2152 if (const SCEVConstant *Step = 2153 dyn_cast<SCEVConstant>(AR->getStepRecurrence(*this))) { 2154 // {X,+,N}/C --> {X/C,+,N/C} if safe and N/C can be folded. 2155 const APInt &StepInt = Step->getValue()->getValue(); 2156 const APInt &DivInt = RHSC->getValue()->getValue(); 2157 if (!StepInt.urem(DivInt) && 2158 getZeroExtendExpr(AR, ExtTy) == 2159 getAddRecExpr(getZeroExtendExpr(AR->getStart(), ExtTy), 2160 getZeroExtendExpr(Step, ExtTy), 2161 AR->getLoop(), SCEV::FlagAnyWrap)) { 2162 SmallVector<const SCEV *, 4> Operands; 2163 for (unsigned i = 0, e = AR->getNumOperands(); i != e; ++i) 2164 Operands.push_back(getUDivExpr(AR->getOperand(i), RHS)); 2165 return getAddRecExpr(Operands, AR->getLoop(), 2166 SCEV::FlagNW); 2167 } 2168 /// Get a canonical UDivExpr for a recurrence. 2169 /// {X,+,N}/C => {Y,+,N}/C where Y=X-(X%N). Safe when C%N=0. 2170 // We can currently only fold X%N if X is constant. 2171 const SCEVConstant *StartC = dyn_cast<SCEVConstant>(AR->getStart()); 2172 if (StartC && !DivInt.urem(StepInt) && 2173 getZeroExtendExpr(AR, ExtTy) == 2174 getAddRecExpr(getZeroExtendExpr(AR->getStart(), ExtTy), 2175 getZeroExtendExpr(Step, ExtTy), 2176 AR->getLoop(), SCEV::FlagAnyWrap)) { 2177 const APInt &StartInt = StartC->getValue()->getValue(); 2178 const APInt &StartRem = StartInt.urem(StepInt); 2179 if (StartRem != 0) 2180 LHS = getAddRecExpr(getConstant(StartInt - StartRem), Step, 2181 AR->getLoop(), SCEV::FlagNW); 2182 } 2183 } 2184 // (A*B)/C --> A*(B/C) if safe and B/C can be folded. 2185 if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(LHS)) { 2186 SmallVector<const SCEV *, 4> Operands; 2187 for (unsigned i = 0, e = M->getNumOperands(); i != e; ++i) 2188 Operands.push_back(getZeroExtendExpr(M->getOperand(i), ExtTy)); 2189 if (getZeroExtendExpr(M, ExtTy) == getMulExpr(Operands)) 2190 // Find an operand that's safely divisible. 2191 for (unsigned i = 0, e = M->getNumOperands(); i != e; ++i) { 2192 const SCEV *Op = M->getOperand(i); 2193 const SCEV *Div = getUDivExpr(Op, RHSC); 2194 if (!isa<SCEVUDivExpr>(Div) && getMulExpr(Div, RHSC) == Op) { 2195 Operands = SmallVector<const SCEV *, 4>(M->op_begin(), 2196 M->op_end()); 2197 Operands[i] = Div; 2198 return getMulExpr(Operands); 2199 } 2200 } 2201 } 2202 // (A+B)/C --> (A/C + B/C) if safe and A/C and B/C can be folded. 2203 if (const SCEVAddExpr *A = dyn_cast<SCEVAddExpr>(LHS)) { 2204 SmallVector<const SCEV *, 4> Operands; 2205 for (unsigned i = 0, e = A->getNumOperands(); i != e; ++i) 2206 Operands.push_back(getZeroExtendExpr(A->getOperand(i), ExtTy)); 2207 if (getZeroExtendExpr(A, ExtTy) == getAddExpr(Operands)) { 2208 Operands.clear(); 2209 for (unsigned i = 0, e = A->getNumOperands(); i != e; ++i) { 2210 const SCEV *Op = getUDivExpr(A->getOperand(i), RHS); 2211 if (isa<SCEVUDivExpr>(Op) || 2212 getMulExpr(Op, RHS) != A->getOperand(i)) 2213 break; 2214 Operands.push_back(Op); 2215 } 2216 if (Operands.size() == A->getNumOperands()) 2217 return getAddExpr(Operands); 2218 } 2219 } 2220 2221 // Fold if both operands are constant. 2222 if (const SCEVConstant *LHSC = dyn_cast<SCEVConstant>(LHS)) { 2223 Constant *LHSCV = LHSC->getValue(); 2224 Constant *RHSCV = RHSC->getValue(); 2225 return getConstant(cast<ConstantInt>(ConstantExpr::getUDiv(LHSCV, 2226 RHSCV))); 2227 } 2228 } 2229 } 2230 2231 FoldingSetNodeID ID; 2232 ID.AddInteger(scUDivExpr); 2233 ID.AddPointer(LHS); 2234 ID.AddPointer(RHS); 2235 void *IP = 0; 2236 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S; 2237 SCEV *S = new (SCEVAllocator) SCEVUDivExpr(ID.Intern(SCEVAllocator), 2238 LHS, RHS); 2239 UniqueSCEVs.InsertNode(S, IP); 2240 return S; 2241} 2242 2243 2244/// getAddRecExpr - Get an add recurrence expression for the specified loop. 2245/// Simplify the expression as much as possible. 2246const SCEV *ScalarEvolution::getAddRecExpr(const SCEV *Start, const SCEV *Step, 2247 const Loop *L, 2248 SCEV::NoWrapFlags Flags) { 2249 SmallVector<const SCEV *, 4> Operands; 2250 Operands.push_back(Start); 2251 if (const SCEVAddRecExpr *StepChrec = dyn_cast<SCEVAddRecExpr>(Step)) 2252 if (StepChrec->getLoop() == L) { 2253 Operands.append(StepChrec->op_begin(), StepChrec->op_end()); 2254 return getAddRecExpr(Operands, L, maskFlags(Flags, SCEV::FlagNW)); 2255 } 2256 2257 Operands.push_back(Step); 2258 return getAddRecExpr(Operands, L, Flags); 2259} 2260 2261/// getAddRecExpr - Get an add recurrence expression for the specified loop. 2262/// Simplify the expression as much as possible. 2263const SCEV * 2264ScalarEvolution::getAddRecExpr(SmallVectorImpl<const SCEV *> &Operands, 2265 const Loop *L, SCEV::NoWrapFlags Flags) { 2266 if (Operands.size() == 1) return Operands[0]; 2267#ifndef NDEBUG 2268 Type *ETy = getEffectiveSCEVType(Operands[0]->getType()); 2269 for (unsigned i = 1, e = Operands.size(); i != e; ++i) 2270 assert(getEffectiveSCEVType(Operands[i]->getType()) == ETy && 2271 "SCEVAddRecExpr operand types don't match!"); 2272 for (unsigned i = 0, e = Operands.size(); i != e; ++i) 2273 assert(isLoopInvariant(Operands[i], L) && 2274 "SCEVAddRecExpr operand is not loop-invariant!"); 2275#endif 2276 2277 if (Operands.back()->isZero()) { 2278 Operands.pop_back(); 2279 return getAddRecExpr(Operands, L, SCEV::FlagAnyWrap); // {X,+,0} --> X 2280 } 2281 2282 // It's tempting to want to call getMaxBackedgeTakenCount count here and 2283 // use that information to infer NUW and NSW flags. However, computing a 2284 // BE count requires calling getAddRecExpr, so we may not yet have a 2285 // meaningful BE count at this point (and if we don't, we'd be stuck 2286 // with a SCEVCouldNotCompute as the cached BE count). 2287 2288 // If FlagNSW is true and all the operands are non-negative, infer FlagNUW. 2289 // And vice-versa. 2290 int SignOrUnsignMask = SCEV::FlagNUW | SCEV::FlagNSW; 2291 SCEV::NoWrapFlags SignOrUnsignWrap = maskFlags(Flags, SignOrUnsignMask); 2292 if (SignOrUnsignWrap && (SignOrUnsignWrap != SignOrUnsignMask)) { 2293 bool All = true; 2294 for (SmallVectorImpl<const SCEV *>::const_iterator I = Operands.begin(), 2295 E = Operands.end(); I != E; ++I) 2296 if (!isKnownNonNegative(*I)) { 2297 All = false; 2298 break; 2299 } 2300 if (All) Flags = setFlags(Flags, (SCEV::NoWrapFlags)SignOrUnsignMask); 2301 } 2302 2303 // Canonicalize nested AddRecs in by nesting them in order of loop depth. 2304 if (const SCEVAddRecExpr *NestedAR = dyn_cast<SCEVAddRecExpr>(Operands[0])) { 2305 const Loop *NestedLoop = NestedAR->getLoop(); 2306 if (L->contains(NestedLoop) ? 2307 (L->getLoopDepth() < NestedLoop->getLoopDepth()) : 2308 (!NestedLoop->contains(L) && 2309 DT->dominates(L->getHeader(), NestedLoop->getHeader()))) { 2310 SmallVector<const SCEV *, 4> NestedOperands(NestedAR->op_begin(), 2311 NestedAR->op_end()); 2312 Operands[0] = NestedAR->getStart(); 2313 // AddRecs require their operands be loop-invariant with respect to their 2314 // loops. Don't perform this transformation if it would break this 2315 // requirement. 2316 bool AllInvariant = true; 2317 for (unsigned i = 0, e = Operands.size(); i != e; ++i) 2318 if (!isLoopInvariant(Operands[i], L)) { 2319 AllInvariant = false; 2320 break; 2321 } 2322 if (AllInvariant) { 2323 // Create a recurrence for the outer loop with the same step size. 2324 // 2325 // The outer recurrence keeps its NW flag but only keeps NUW/NSW if the 2326 // inner recurrence has the same property. 2327 SCEV::NoWrapFlags OuterFlags = 2328 maskFlags(Flags, SCEV::FlagNW | NestedAR->getNoWrapFlags()); 2329 2330 NestedOperands[0] = getAddRecExpr(Operands, L, OuterFlags); 2331 AllInvariant = true; 2332 for (unsigned i = 0, e = NestedOperands.size(); i != e; ++i) 2333 if (!isLoopInvariant(NestedOperands[i], NestedLoop)) { 2334 AllInvariant = false; 2335 break; 2336 } 2337 if (AllInvariant) { 2338 // Ok, both add recurrences are valid after the transformation. 2339 // 2340 // The inner recurrence keeps its NW flag but only keeps NUW/NSW if 2341 // the outer recurrence has the same property. 2342 SCEV::NoWrapFlags InnerFlags = 2343 maskFlags(NestedAR->getNoWrapFlags(), SCEV::FlagNW | Flags); 2344 return getAddRecExpr(NestedOperands, NestedLoop, InnerFlags); 2345 } 2346 } 2347 // Reset Operands to its original state. 2348 Operands[0] = NestedAR; 2349 } 2350 } 2351 2352 // Okay, it looks like we really DO need an addrec expr. Check to see if we 2353 // already have one, otherwise create a new one. 2354 FoldingSetNodeID ID; 2355 ID.AddInteger(scAddRecExpr); 2356 for (unsigned i = 0, e = Operands.size(); i != e; ++i) 2357 ID.AddPointer(Operands[i]); 2358 ID.AddPointer(L); 2359 void *IP = 0; 2360 SCEVAddRecExpr *S = 2361 static_cast<SCEVAddRecExpr *>(UniqueSCEVs.FindNodeOrInsertPos(ID, IP)); 2362 if (!S) { 2363 const SCEV **O = SCEVAllocator.Allocate<const SCEV *>(Operands.size()); 2364 std::uninitialized_copy(Operands.begin(), Operands.end(), O); 2365 S = new (SCEVAllocator) SCEVAddRecExpr(ID.Intern(SCEVAllocator), 2366 O, Operands.size(), L); 2367 UniqueSCEVs.InsertNode(S, IP); 2368 } 2369 S->setNoWrapFlags(Flags); 2370 return S; 2371} 2372 2373const SCEV *ScalarEvolution::getSMaxExpr(const SCEV *LHS, 2374 const SCEV *RHS) { 2375 SmallVector<const SCEV *, 2> Ops; 2376 Ops.push_back(LHS); 2377 Ops.push_back(RHS); 2378 return getSMaxExpr(Ops); 2379} 2380 2381const SCEV * 2382ScalarEvolution::getSMaxExpr(SmallVectorImpl<const SCEV *> &Ops) { 2383 assert(!Ops.empty() && "Cannot get empty smax!"); 2384 if (Ops.size() == 1) return Ops[0]; 2385#ifndef NDEBUG 2386 Type *ETy = getEffectiveSCEVType(Ops[0]->getType()); 2387 for (unsigned i = 1, e = Ops.size(); i != e; ++i) 2388 assert(getEffectiveSCEVType(Ops[i]->getType()) == ETy && 2389 "SCEVSMaxExpr operand types don't match!"); 2390#endif 2391 2392 // Sort by complexity, this groups all similar expression types together. 2393 GroupByComplexity(Ops, LI); 2394 2395 // If there are any constants, fold them together. 2396 unsigned Idx = 0; 2397 if (const SCEVConstant *LHSC = dyn_cast<SCEVConstant>(Ops[0])) { 2398 ++Idx; 2399 assert(Idx < Ops.size()); 2400 while (const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(Ops[Idx])) { 2401 // We found two constants, fold them together! 2402 ConstantInt *Fold = ConstantInt::get(getContext(), 2403 APIntOps::smax(LHSC->getValue()->getValue(), 2404 RHSC->getValue()->getValue())); 2405 Ops[0] = getConstant(Fold); 2406 Ops.erase(Ops.begin()+1); // Erase the folded element 2407 if (Ops.size() == 1) return Ops[0]; 2408 LHSC = cast<SCEVConstant>(Ops[0]); 2409 } 2410 2411 // If we are left with a constant minimum-int, strip it off. 2412 if (cast<SCEVConstant>(Ops[0])->getValue()->isMinValue(true)) { 2413 Ops.erase(Ops.begin()); 2414 --Idx; 2415 } else if (cast<SCEVConstant>(Ops[0])->getValue()->isMaxValue(true)) { 2416 // If we have an smax with a constant maximum-int, it will always be 2417 // maximum-int. 2418 return Ops[0]; 2419 } 2420 2421 if (Ops.size() == 1) return Ops[0]; 2422 } 2423 2424 // Find the first SMax 2425 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scSMaxExpr) 2426 ++Idx; 2427 2428 // Check to see if one of the operands is an SMax. If so, expand its operands 2429 // onto our operand list, and recurse to simplify. 2430 if (Idx < Ops.size()) { 2431 bool DeletedSMax = false; 2432 while (const SCEVSMaxExpr *SMax = dyn_cast<SCEVSMaxExpr>(Ops[Idx])) { 2433 Ops.erase(Ops.begin()+Idx); 2434 Ops.append(SMax->op_begin(), SMax->op_end()); 2435 DeletedSMax = true; 2436 } 2437 2438 if (DeletedSMax) 2439 return getSMaxExpr(Ops); 2440 } 2441 2442 // Okay, check to see if the same value occurs in the operand list twice. If 2443 // so, delete one. Since we sorted the list, these values are required to 2444 // be adjacent. 2445 for (unsigned i = 0, e = Ops.size()-1; i != e; ++i) 2446 // X smax Y smax Y --> X smax Y 2447 // X smax Y --> X, if X is always greater than Y 2448 if (Ops[i] == Ops[i+1] || 2449 isKnownPredicate(ICmpInst::ICMP_SGE, Ops[i], Ops[i+1])) { 2450 Ops.erase(Ops.begin()+i+1, Ops.begin()+i+2); 2451 --i; --e; 2452 } else if (isKnownPredicate(ICmpInst::ICMP_SLE, Ops[i], Ops[i+1])) { 2453 Ops.erase(Ops.begin()+i, Ops.begin()+i+1); 2454 --i; --e; 2455 } 2456 2457 if (Ops.size() == 1) return Ops[0]; 2458 2459 assert(!Ops.empty() && "Reduced smax down to nothing!"); 2460 2461 // Okay, it looks like we really DO need an smax expr. Check to see if we 2462 // already have one, otherwise create a new one. 2463 FoldingSetNodeID ID; 2464 ID.AddInteger(scSMaxExpr); 2465 for (unsigned i = 0, e = Ops.size(); i != e; ++i) 2466 ID.AddPointer(Ops[i]); 2467 void *IP = 0; 2468 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S; 2469 const SCEV **O = SCEVAllocator.Allocate<const SCEV *>(Ops.size()); 2470 std::uninitialized_copy(Ops.begin(), Ops.end(), O); 2471 SCEV *S = new (SCEVAllocator) SCEVSMaxExpr(ID.Intern(SCEVAllocator), 2472 O, Ops.size()); 2473 UniqueSCEVs.InsertNode(S, IP); 2474 return S; 2475} 2476 2477const SCEV *ScalarEvolution::getUMaxExpr(const SCEV *LHS, 2478 const SCEV *RHS) { 2479 SmallVector<const SCEV *, 2> Ops; 2480 Ops.push_back(LHS); 2481 Ops.push_back(RHS); 2482 return getUMaxExpr(Ops); 2483} 2484 2485const SCEV * 2486ScalarEvolution::getUMaxExpr(SmallVectorImpl<const SCEV *> &Ops) { 2487 assert(!Ops.empty() && "Cannot get empty umax!"); 2488 if (Ops.size() == 1) return Ops[0]; 2489#ifndef NDEBUG 2490 Type *ETy = getEffectiveSCEVType(Ops[0]->getType()); 2491 for (unsigned i = 1, e = Ops.size(); i != e; ++i) 2492 assert(getEffectiveSCEVType(Ops[i]->getType()) == ETy && 2493 "SCEVUMaxExpr operand types don't match!"); 2494#endif 2495 2496 // Sort by complexity, this groups all similar expression types together. 2497 GroupByComplexity(Ops, LI); 2498 2499 // If there are any constants, fold them together. 2500 unsigned Idx = 0; 2501 if (const SCEVConstant *LHSC = dyn_cast<SCEVConstant>(Ops[0])) { 2502 ++Idx; 2503 assert(Idx < Ops.size()); 2504 while (const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(Ops[Idx])) { 2505 // We found two constants, fold them together! 2506 ConstantInt *Fold = ConstantInt::get(getContext(), 2507 APIntOps::umax(LHSC->getValue()->getValue(), 2508 RHSC->getValue()->getValue())); 2509 Ops[0] = getConstant(Fold); 2510 Ops.erase(Ops.begin()+1); // Erase the folded element 2511 if (Ops.size() == 1) return Ops[0]; 2512 LHSC = cast<SCEVConstant>(Ops[0]); 2513 } 2514 2515 // If we are left with a constant minimum-int, strip it off. 2516 if (cast<SCEVConstant>(Ops[0])->getValue()->isMinValue(false)) { 2517 Ops.erase(Ops.begin()); 2518 --Idx; 2519 } else if (cast<SCEVConstant>(Ops[0])->getValue()->isMaxValue(false)) { 2520 // If we have an umax with a constant maximum-int, it will always be 2521 // maximum-int. 2522 return Ops[0]; 2523 } 2524 2525 if (Ops.size() == 1) return Ops[0]; 2526 } 2527 2528 // Find the first UMax 2529 while (Idx < Ops.size() && Ops[Idx]->getSCEVType() < scUMaxExpr) 2530 ++Idx; 2531 2532 // Check to see if one of the operands is a UMax. If so, expand its operands 2533 // onto our operand list, and recurse to simplify. 2534 if (Idx < Ops.size()) { 2535 bool DeletedUMax = false; 2536 while (const SCEVUMaxExpr *UMax = dyn_cast<SCEVUMaxExpr>(Ops[Idx])) { 2537 Ops.erase(Ops.begin()+Idx); 2538 Ops.append(UMax->op_begin(), UMax->op_end()); 2539 DeletedUMax = true; 2540 } 2541 2542 if (DeletedUMax) 2543 return getUMaxExpr(Ops); 2544 } 2545 2546 // Okay, check to see if the same value occurs in the operand list twice. If 2547 // so, delete one. Since we sorted the list, these values are required to 2548 // be adjacent. 2549 for (unsigned i = 0, e = Ops.size()-1; i != e; ++i) 2550 // X umax Y umax Y --> X umax Y 2551 // X umax Y --> X, if X is always greater than Y 2552 if (Ops[i] == Ops[i+1] || 2553 isKnownPredicate(ICmpInst::ICMP_UGE, Ops[i], Ops[i+1])) { 2554 Ops.erase(Ops.begin()+i+1, Ops.begin()+i+2); 2555 --i; --e; 2556 } else if (isKnownPredicate(ICmpInst::ICMP_ULE, Ops[i], Ops[i+1])) { 2557 Ops.erase(Ops.begin()+i, Ops.begin()+i+1); 2558 --i; --e; 2559 } 2560 2561 if (Ops.size() == 1) return Ops[0]; 2562 2563 assert(!Ops.empty() && "Reduced umax down to nothing!"); 2564 2565 // Okay, it looks like we really DO need a umax expr. Check to see if we 2566 // already have one, otherwise create a new one. 2567 FoldingSetNodeID ID; 2568 ID.AddInteger(scUMaxExpr); 2569 for (unsigned i = 0, e = Ops.size(); i != e; ++i) 2570 ID.AddPointer(Ops[i]); 2571 void *IP = 0; 2572 if (const SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) return S; 2573 const SCEV **O = SCEVAllocator.Allocate<const SCEV *>(Ops.size()); 2574 std::uninitialized_copy(Ops.begin(), Ops.end(), O); 2575 SCEV *S = new (SCEVAllocator) SCEVUMaxExpr(ID.Intern(SCEVAllocator), 2576 O, Ops.size()); 2577 UniqueSCEVs.InsertNode(S, IP); 2578 return S; 2579} 2580 2581const SCEV *ScalarEvolution::getSMinExpr(const SCEV *LHS, 2582 const SCEV *RHS) { 2583 // ~smax(~x, ~y) == smin(x, y). 2584 return getNotSCEV(getSMaxExpr(getNotSCEV(LHS), getNotSCEV(RHS))); 2585} 2586 2587const SCEV *ScalarEvolution::getUMinExpr(const SCEV *LHS, 2588 const SCEV *RHS) { 2589 // ~umax(~x, ~y) == umin(x, y) 2590 return getNotSCEV(getUMaxExpr(getNotSCEV(LHS), getNotSCEV(RHS))); 2591} 2592 2593const SCEV *ScalarEvolution::getSizeOfExpr(Type *AllocTy) { 2594 // If we have TargetData, we can bypass creating a target-independent 2595 // constant expression and then folding it back into a ConstantInt. 2596 // This is just a compile-time optimization. 2597 if (TD) 2598 return getConstant(TD->getIntPtrType(getContext()), 2599 TD->getTypeAllocSize(AllocTy)); 2600 2601 Constant *C = ConstantExpr::getSizeOf(AllocTy); 2602 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) 2603 if (Constant *Folded = ConstantFoldConstantExpression(CE, TD, TLI)) 2604 C = Folded; 2605 Type *Ty = getEffectiveSCEVType(PointerType::getUnqual(AllocTy)); 2606 return getTruncateOrZeroExtend(getSCEV(C), Ty); 2607} 2608 2609const SCEV *ScalarEvolution::getAlignOfExpr(Type *AllocTy) { 2610 Constant *C = ConstantExpr::getAlignOf(AllocTy); 2611 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) 2612 if (Constant *Folded = ConstantFoldConstantExpression(CE, TD, TLI)) 2613 C = Folded; 2614 Type *Ty = getEffectiveSCEVType(PointerType::getUnqual(AllocTy)); 2615 return getTruncateOrZeroExtend(getSCEV(C), Ty); 2616} 2617 2618const SCEV *ScalarEvolution::getOffsetOfExpr(StructType *STy, 2619 unsigned FieldNo) { 2620 // If we have TargetData, we can bypass creating a target-independent 2621 // constant expression and then folding it back into a ConstantInt. 2622 // This is just a compile-time optimization. 2623 if (TD) 2624 return getConstant(TD->getIntPtrType(getContext()), 2625 TD->getStructLayout(STy)->getElementOffset(FieldNo)); 2626 2627 Constant *C = ConstantExpr::getOffsetOf(STy, FieldNo); 2628 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) 2629 if (Constant *Folded = ConstantFoldConstantExpression(CE, TD, TLI)) 2630 C = Folded; 2631 Type *Ty = getEffectiveSCEVType(PointerType::getUnqual(STy)); 2632 return getTruncateOrZeroExtend(getSCEV(C), Ty); 2633} 2634 2635const SCEV *ScalarEvolution::getOffsetOfExpr(Type *CTy, 2636 Constant *FieldNo) { 2637 Constant *C = ConstantExpr::getOffsetOf(CTy, FieldNo); 2638 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) 2639 if (Constant *Folded = ConstantFoldConstantExpression(CE, TD, TLI)) 2640 C = Folded; 2641 Type *Ty = getEffectiveSCEVType(PointerType::getUnqual(CTy)); 2642 return getTruncateOrZeroExtend(getSCEV(C), Ty); 2643} 2644 2645const SCEV *ScalarEvolution::getUnknown(Value *V) { 2646 // Don't attempt to do anything other than create a SCEVUnknown object 2647 // here. createSCEV only calls getUnknown after checking for all other 2648 // interesting possibilities, and any other code that calls getUnknown 2649 // is doing so in order to hide a value from SCEV canonicalization. 2650 2651 FoldingSetNodeID ID; 2652 ID.AddInteger(scUnknown); 2653 ID.AddPointer(V); 2654 void *IP = 0; 2655 if (SCEV *S = UniqueSCEVs.FindNodeOrInsertPos(ID, IP)) { 2656 assert(cast<SCEVUnknown>(S)->getValue() == V && 2657 "Stale SCEVUnknown in uniquing map!"); 2658 return S; 2659 } 2660 SCEV *S = new (SCEVAllocator) SCEVUnknown(ID.Intern(SCEVAllocator), V, this, 2661 FirstUnknown); 2662 FirstUnknown = cast<SCEVUnknown>(S); 2663 UniqueSCEVs.InsertNode(S, IP); 2664 return S; 2665} 2666 2667//===----------------------------------------------------------------------===// 2668// Basic SCEV Analysis and PHI Idiom Recognition Code 2669// 2670 2671/// isSCEVable - Test if values of the given type are analyzable within 2672/// the SCEV framework. This primarily includes integer types, and it 2673/// can optionally include pointer types if the ScalarEvolution class 2674/// has access to target-specific information. 2675bool ScalarEvolution::isSCEVable(Type *Ty) const { 2676 // Integers and pointers are always SCEVable. 2677 return Ty->isIntegerTy() || Ty->isPointerTy(); 2678} 2679 2680/// getTypeSizeInBits - Return the size in bits of the specified type, 2681/// for which isSCEVable must return true. 2682uint64_t ScalarEvolution::getTypeSizeInBits(Type *Ty) const { 2683 assert(isSCEVable(Ty) && "Type is not SCEVable!"); 2684 2685 // If we have a TargetData, use it! 2686 if (TD) 2687 return TD->getTypeSizeInBits(Ty); 2688 2689 // Integer types have fixed sizes. 2690 if (Ty->isIntegerTy()) 2691 return Ty->getPrimitiveSizeInBits(); 2692 2693 // The only other support type is pointer. Without TargetData, conservatively 2694 // assume pointers are 64-bit. 2695 assert(Ty->isPointerTy() && "isSCEVable permitted a non-SCEVable type!"); 2696 return 64; 2697} 2698 2699/// getEffectiveSCEVType - Return a type with the same bitwidth as 2700/// the given type and which represents how SCEV will treat the given 2701/// type, for which isSCEVable must return true. For pointer types, 2702/// this is the pointer-sized integer type. 2703Type *ScalarEvolution::getEffectiveSCEVType(Type *Ty) const { 2704 assert(isSCEVable(Ty) && "Type is not SCEVable!"); 2705 2706 if (Ty->isIntegerTy()) 2707 return Ty; 2708 2709 // The only other support type is pointer. 2710 assert(Ty->isPointerTy() && "Unexpected non-pointer non-integer type!"); 2711 if (TD) return TD->getIntPtrType(getContext()); 2712 2713 // Without TargetData, conservatively assume pointers are 64-bit. 2714 return Type::getInt64Ty(getContext()); 2715} 2716 2717const SCEV *ScalarEvolution::getCouldNotCompute() { 2718 return &CouldNotCompute; 2719} 2720 2721/// getSCEV - Return an existing SCEV if it exists, otherwise analyze the 2722/// expression and create a new one. 2723const SCEV *ScalarEvolution::getSCEV(Value *V) { 2724 assert(isSCEVable(V->getType()) && "Value is not SCEVable!"); 2725 2726 ValueExprMapType::const_iterator I = ValueExprMap.find(V); 2727 if (I != ValueExprMap.end()) return I->second; 2728 const SCEV *S = createSCEV(V); 2729 2730 // The process of creating a SCEV for V may have caused other SCEVs 2731 // to have been created, so it's necessary to insert the new entry 2732 // from scratch, rather than trying to remember the insert position 2733 // above. 2734 ValueExprMap.insert(std::make_pair(SCEVCallbackVH(V, this), S)); 2735 return S; 2736} 2737 2738/// getNegativeSCEV - Return a SCEV corresponding to -V = -1*V 2739/// 2740const SCEV *ScalarEvolution::getNegativeSCEV(const SCEV *V) { 2741 if (const SCEVConstant *VC = dyn_cast<SCEVConstant>(V)) 2742 return getConstant( 2743 cast<ConstantInt>(ConstantExpr::getNeg(VC->getValue()))); 2744 2745 Type *Ty = V->getType(); 2746 Ty = getEffectiveSCEVType(Ty); 2747 return getMulExpr(V, 2748 getConstant(cast<ConstantInt>(Constant::getAllOnesValue(Ty)))); 2749} 2750 2751/// getNotSCEV - Return a SCEV corresponding to ~V = -1-V 2752const SCEV *ScalarEvolution::getNotSCEV(const SCEV *V) { 2753 if (const SCEVConstant *VC = dyn_cast<SCEVConstant>(V)) 2754 return getConstant( 2755 cast<ConstantInt>(ConstantExpr::getNot(VC->getValue()))); 2756 2757 Type *Ty = V->getType(); 2758 Ty = getEffectiveSCEVType(Ty); 2759 const SCEV *AllOnes = 2760 getConstant(cast<ConstantInt>(Constant::getAllOnesValue(Ty))); 2761 return getMinusSCEV(AllOnes, V); 2762} 2763 2764/// getMinusSCEV - Return LHS-RHS. Minus is represented in SCEV as A+B*-1. 2765const SCEV *ScalarEvolution::getMinusSCEV(const SCEV *LHS, const SCEV *RHS, 2766 SCEV::NoWrapFlags Flags) { 2767 assert(!maskFlags(Flags, SCEV::FlagNUW) && "subtraction does not have NUW"); 2768 2769 // Fast path: X - X --> 0. 2770 if (LHS == RHS) 2771 return getConstant(LHS->getType(), 0); 2772 2773 // X - Y --> X + -Y 2774 return getAddExpr(LHS, getNegativeSCEV(RHS), Flags); 2775} 2776 2777/// getTruncateOrZeroExtend - Return a SCEV corresponding to a conversion of the 2778/// input value to the specified type. If the type must be extended, it is zero 2779/// extended. 2780const SCEV * 2781ScalarEvolution::getTruncateOrZeroExtend(const SCEV *V, Type *Ty) { 2782 Type *SrcTy = V->getType(); 2783 assert((SrcTy->isIntegerTy() || SrcTy->isPointerTy()) && 2784 (Ty->isIntegerTy() || Ty->isPointerTy()) && 2785 "Cannot truncate or zero extend with non-integer arguments!"); 2786 if (getTypeSizeInBits(SrcTy) == getTypeSizeInBits(Ty)) 2787 return V; // No conversion 2788 if (getTypeSizeInBits(SrcTy) > getTypeSizeInBits(Ty)) 2789 return getTruncateExpr(V, Ty); 2790 return getZeroExtendExpr(V, Ty); 2791} 2792 2793/// getTruncateOrSignExtend - Return a SCEV corresponding to a conversion of the 2794/// input value to the specified type. If the type must be extended, it is sign 2795/// extended. 2796const SCEV * 2797ScalarEvolution::getTruncateOrSignExtend(const SCEV *V, 2798 Type *Ty) { 2799 Type *SrcTy = V->getType(); 2800 assert((SrcTy->isIntegerTy() || SrcTy->isPointerTy()) && 2801 (Ty->isIntegerTy() || Ty->isPointerTy()) && 2802 "Cannot truncate or zero extend with non-integer arguments!"); 2803 if (getTypeSizeInBits(SrcTy) == getTypeSizeInBits(Ty)) 2804 return V; // No conversion 2805 if (getTypeSizeInBits(SrcTy) > getTypeSizeInBits(Ty)) 2806 return getTruncateExpr(V, Ty); 2807 return getSignExtendExpr(V, Ty); 2808} 2809 2810/// getNoopOrZeroExtend - Return a SCEV corresponding to a conversion of the 2811/// input value to the specified type. If the type must be extended, it is zero 2812/// extended. The conversion must not be narrowing. 2813const SCEV * 2814ScalarEvolution::getNoopOrZeroExtend(const SCEV *V, Type *Ty) { 2815 Type *SrcTy = V->getType(); 2816 assert((SrcTy->isIntegerTy() || SrcTy->isPointerTy()) && 2817 (Ty->isIntegerTy() || Ty->isPointerTy()) && 2818 "Cannot noop or zero extend with non-integer arguments!"); 2819 assert(getTypeSizeInBits(SrcTy) <= getTypeSizeInBits(Ty) && 2820 "getNoopOrZeroExtend cannot truncate!"); 2821 if (getTypeSizeInBits(SrcTy) == getTypeSizeInBits(Ty)) 2822 return V; // No conversion 2823 return getZeroExtendExpr(V, Ty); 2824} 2825 2826/// getNoopOrSignExtend - Return a SCEV corresponding to a conversion of the 2827/// input value to the specified type. If the type must be extended, it is sign 2828/// extended. The conversion must not be narrowing. 2829const SCEV * 2830ScalarEvolution::getNoopOrSignExtend(const SCEV *V, Type *Ty) { 2831 Type *SrcTy = V->getType(); 2832 assert((SrcTy->isIntegerTy() || SrcTy->isPointerTy()) && 2833 (Ty->isIntegerTy() || Ty->isPointerTy()) && 2834 "Cannot noop or sign extend with non-integer arguments!"); 2835 assert(getTypeSizeInBits(SrcTy) <= getTypeSizeInBits(Ty) && 2836 "getNoopOrSignExtend cannot truncate!"); 2837 if (getTypeSizeInBits(SrcTy) == getTypeSizeInBits(Ty)) 2838 return V; // No conversion 2839 return getSignExtendExpr(V, Ty); 2840} 2841 2842/// getNoopOrAnyExtend - Return a SCEV corresponding to a conversion of 2843/// the input value to the specified type. If the type must be extended, 2844/// it is extended with unspecified bits. The conversion must not be 2845/// narrowing. 2846const SCEV * 2847ScalarEvolution::getNoopOrAnyExtend(const SCEV *V, Type *Ty) { 2848 Type *SrcTy = V->getType(); 2849 assert((SrcTy->isIntegerTy() || SrcTy->isPointerTy()) && 2850 (Ty->isIntegerTy() || Ty->isPointerTy()) && 2851 "Cannot noop or any extend with non-integer arguments!"); 2852 assert(getTypeSizeInBits(SrcTy) <= getTypeSizeInBits(Ty) && 2853 "getNoopOrAnyExtend cannot truncate!"); 2854 if (getTypeSizeInBits(SrcTy) == getTypeSizeInBits(Ty)) 2855 return V; // No conversion 2856 return getAnyExtendExpr(V, Ty); 2857} 2858 2859/// getTruncateOrNoop - Return a SCEV corresponding to a conversion of the 2860/// input value to the specified type. The conversion must not be widening. 2861const SCEV * 2862ScalarEvolution::getTruncateOrNoop(const SCEV *V, Type *Ty) { 2863 Type *SrcTy = V->getType(); 2864 assert((SrcTy->isIntegerTy() || SrcTy->isPointerTy()) && 2865 (Ty->isIntegerTy() || Ty->isPointerTy()) && 2866 "Cannot truncate or noop with non-integer arguments!"); 2867 assert(getTypeSizeInBits(SrcTy) >= getTypeSizeInBits(Ty) && 2868 "getTruncateOrNoop cannot extend!"); 2869 if (getTypeSizeInBits(SrcTy) == getTypeSizeInBits(Ty)) 2870 return V; // No conversion 2871 return getTruncateExpr(V, Ty); 2872} 2873 2874/// getUMaxFromMismatchedTypes - Promote the operands to the wider of 2875/// the types using zero-extension, and then perform a umax operation 2876/// with them. 2877const SCEV *ScalarEvolution::getUMaxFromMismatchedTypes(const SCEV *LHS, 2878 const SCEV *RHS) { 2879 const SCEV *PromotedLHS = LHS; 2880 const SCEV *PromotedRHS = RHS; 2881 2882 if (getTypeSizeInBits(LHS->getType()) > getTypeSizeInBits(RHS->getType())) 2883 PromotedRHS = getZeroExtendExpr(RHS, LHS->getType()); 2884 else 2885 PromotedLHS = getNoopOrZeroExtend(LHS, RHS->getType()); 2886 2887 return getUMaxExpr(PromotedLHS, PromotedRHS); 2888} 2889 2890/// getUMinFromMismatchedTypes - Promote the operands to the wider of 2891/// the types using zero-extension, and then perform a umin operation 2892/// with them. 2893const SCEV *ScalarEvolution::getUMinFromMismatchedTypes(const SCEV *LHS, 2894 const SCEV *RHS) { 2895 const SCEV *PromotedLHS = LHS; 2896 const SCEV *PromotedRHS = RHS; 2897 2898 if (getTypeSizeInBits(LHS->getType()) > getTypeSizeInBits(RHS->getType())) 2899 PromotedRHS = getZeroExtendExpr(RHS, LHS->getType()); 2900 else 2901 PromotedLHS = getNoopOrZeroExtend(LHS, RHS->getType()); 2902 2903 return getUMinExpr(PromotedLHS, PromotedRHS); 2904} 2905 2906/// getPointerBase - Transitively follow the chain of pointer-type operands 2907/// until reaching a SCEV that does not have a single pointer operand. This 2908/// returns a SCEVUnknown pointer for well-formed pointer-type expressions, 2909/// but corner cases do exist. 2910const SCEV *ScalarEvolution::getPointerBase(const SCEV *V) { 2911 // A pointer operand may evaluate to a nonpointer expression, such as null. 2912 if (!V->getType()->isPointerTy()) 2913 return V; 2914 2915 if (const SCEVCastExpr *Cast = dyn_cast<SCEVCastExpr>(V)) { 2916 return getPointerBase(Cast->getOperand()); 2917 } 2918 else if (const SCEVNAryExpr *NAry = dyn_cast<SCEVNAryExpr>(V)) { 2919 const SCEV *PtrOp = 0; 2920 for (SCEVNAryExpr::op_iterator I = NAry->op_begin(), E = NAry->op_end(); 2921 I != E; ++I) { 2922 if ((*I)->getType()->isPointerTy()) { 2923 // Cannot find the base of an expression with multiple pointer operands. 2924 if (PtrOp) 2925 return V; 2926 PtrOp = *I; 2927 } 2928 } 2929 if (!PtrOp) 2930 return V; 2931 return getPointerBase(PtrOp); 2932 } 2933 return V; 2934} 2935 2936/// PushDefUseChildren - Push users of the given Instruction 2937/// onto the given Worklist. 2938static void 2939PushDefUseChildren(Instruction *I, 2940 SmallVectorImpl<Instruction *> &Worklist) { 2941 // Push the def-use children onto the Worklist stack. 2942 for (Value::use_iterator UI = I->use_begin(), UE = I->use_end(); 2943 UI != UE; ++UI) 2944 Worklist.push_back(cast<Instruction>(*UI)); 2945} 2946 2947/// ForgetSymbolicValue - This looks up computed SCEV values for all 2948/// instructions that depend on the given instruction and removes them from 2949/// the ValueExprMapType map if they reference SymName. This is used during PHI 2950/// resolution. 2951void 2952ScalarEvolution::ForgetSymbolicName(Instruction *PN, const SCEV *SymName) { 2953 SmallVector<Instruction *, 16> Worklist; 2954 PushDefUseChildren(PN, Worklist); 2955 2956 SmallPtrSet<Instruction *, 8> Visited; 2957 Visited.insert(PN); 2958 while (!Worklist.empty()) { 2959 Instruction *I = Worklist.pop_back_val(); 2960 if (!Visited.insert(I)) continue; 2961 2962 ValueExprMapType::iterator It = 2963 ValueExprMap.find(static_cast<Value *>(I)); 2964 if (It != ValueExprMap.end()) { 2965 const SCEV *Old = It->second; 2966 2967 // Short-circuit the def-use traversal if the symbolic name 2968 // ceases to appear in expressions. 2969 if (Old != SymName && !hasOperand(Old, SymName)) 2970 continue; 2971 2972 // SCEVUnknown for a PHI either means that it has an unrecognized 2973 // structure, it's a PHI that's in the progress of being computed 2974 // by createNodeForPHI, or it's a single-value PHI. In the first case, 2975 // additional loop trip count information isn't going to change anything. 2976 // In the second case, createNodeForPHI will perform the necessary 2977 // updates on its own when it gets to that point. In the third, we do 2978 // want to forget the SCEVUnknown. 2979 if (!isa<PHINode>(I) || 2980 !isa<SCEVUnknown>(Old) || 2981 (I != PN && Old == SymName)) { 2982 forgetMemoizedResults(Old); 2983 ValueExprMap.erase(It); 2984 } 2985 } 2986 2987 PushDefUseChildren(I, Worklist); 2988 } 2989} 2990 2991/// createNodeForPHI - PHI nodes have two cases. Either the PHI node exists in 2992/// a loop header, making it a potential recurrence, or it doesn't. 2993/// 2994const SCEV *ScalarEvolution::createNodeForPHI(PHINode *PN) { 2995 if (const Loop *L = LI->getLoopFor(PN->getParent())) 2996 if (L->getHeader() == PN->getParent()) { 2997 // The loop may have multiple entrances or multiple exits; we can analyze 2998 // this phi as an addrec if it has a unique entry value and a unique 2999 // backedge value. 3000 Value *BEValueV = 0, *StartValueV = 0; 3001 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) { 3002 Value *V = PN->getIncomingValue(i); 3003 if (L->contains(PN->getIncomingBlock(i))) { 3004 if (!BEValueV) { 3005 BEValueV = V; 3006 } else if (BEValueV != V) { 3007 BEValueV = 0; 3008 break; 3009 } 3010 } else if (!StartValueV) { 3011 StartValueV = V; 3012 } else if (StartValueV != V) { 3013 StartValueV = 0; 3014 break; 3015 } 3016 } 3017 if (BEValueV && StartValueV) { 3018 // While we are analyzing this PHI node, handle its value symbolically. 3019 const SCEV *SymbolicName = getUnknown(PN); 3020 assert(ValueExprMap.find(PN) == ValueExprMap.end() && 3021 "PHI node already processed?"); 3022 ValueExprMap.insert(std::make_pair(SCEVCallbackVH(PN, this), SymbolicName)); 3023 3024 // Using this symbolic name for the PHI, analyze the value coming around 3025 // the back-edge. 3026 const SCEV *BEValue = getSCEV(BEValueV); 3027 3028 // NOTE: If BEValue is loop invariant, we know that the PHI node just 3029 // has a special value for the first iteration of the loop. 3030 3031 // If the value coming around the backedge is an add with the symbolic 3032 // value we just inserted, then we found a simple induction variable! 3033 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(BEValue)) { 3034 // If there is a single occurrence of the symbolic value, replace it 3035 // with a recurrence. 3036 unsigned FoundIndex = Add->getNumOperands(); 3037 for (unsigned i = 0, e = Add->getNumOperands(); i != e; ++i) 3038 if (Add->getOperand(i) == SymbolicName) 3039 if (FoundIndex == e) { 3040 FoundIndex = i; 3041 break; 3042 } 3043 3044 if (FoundIndex != Add->getNumOperands()) { 3045 // Create an add with everything but the specified operand. 3046 SmallVector<const SCEV *, 8> Ops; 3047 for (unsigned i = 0, e = Add->getNumOperands(); i != e; ++i) 3048 if (i != FoundIndex) 3049 Ops.push_back(Add->getOperand(i)); 3050 const SCEV *Accum = getAddExpr(Ops); 3051 3052 // This is not a valid addrec if the step amount is varying each 3053 // loop iteration, but is not itself an addrec in this loop. 3054 if (isLoopInvariant(Accum, L) || 3055 (isa<SCEVAddRecExpr>(Accum) && 3056 cast<SCEVAddRecExpr>(Accum)->getLoop() == L)) { 3057 SCEV::NoWrapFlags Flags = SCEV::FlagAnyWrap; 3058 3059 // If the increment doesn't overflow, then neither the addrec nor 3060 // the post-increment will overflow. 3061 if (const AddOperator *OBO = dyn_cast<AddOperator>(BEValueV)) { 3062 if (OBO->hasNoUnsignedWrap()) 3063 Flags = setFlags(Flags, SCEV::FlagNUW); 3064 if (OBO->hasNoSignedWrap()) 3065 Flags = setFlags(Flags, SCEV::FlagNSW); 3066 } else if (const GEPOperator *GEP = 3067 dyn_cast<GEPOperator>(BEValueV)) { 3068 // If the increment is an inbounds GEP, then we know the address 3069 // space cannot be wrapped around. We cannot make any guarantee 3070 // about signed or unsigned overflow because pointers are 3071 // unsigned but we may have a negative index from the base 3072 // pointer. 3073 if (GEP->isInBounds()) 3074 Flags = setFlags(Flags, SCEV::FlagNW); 3075 } 3076 3077 const SCEV *StartVal = getSCEV(StartValueV); 3078 const SCEV *PHISCEV = getAddRecExpr(StartVal, Accum, L, Flags); 3079 3080 // Since the no-wrap flags are on the increment, they apply to the 3081 // post-incremented value as well. 3082 if (isLoopInvariant(Accum, L)) 3083 (void)getAddRecExpr(getAddExpr(StartVal, Accum), 3084 Accum, L, Flags); 3085 3086 // Okay, for the entire analysis of this edge we assumed the PHI 3087 // to be symbolic. We now need to go back and purge all of the 3088 // entries for the scalars that use the symbolic expression. 3089 ForgetSymbolicName(PN, SymbolicName); 3090 ValueExprMap[SCEVCallbackVH(PN, this)] = PHISCEV; 3091 return PHISCEV; 3092 } 3093 } 3094 } else if (const SCEVAddRecExpr *AddRec = 3095 dyn_cast<SCEVAddRecExpr>(BEValue)) { 3096 // Otherwise, this could be a loop like this: 3097 // i = 0; for (j = 1; ..; ++j) { .... i = j; } 3098 // In this case, j = {1,+,1} and BEValue is j. 3099 // Because the other in-value of i (0) fits the evolution of BEValue 3100 // i really is an addrec evolution. 3101 if (AddRec->getLoop() == L && AddRec->isAffine()) { 3102 const SCEV *StartVal = getSCEV(StartValueV); 3103 3104 // If StartVal = j.start - j.stride, we can use StartVal as the 3105 // initial step of the addrec evolution. 3106 if (StartVal == getMinusSCEV(AddRec->getOperand(0), 3107 AddRec->getOperand(1))) { 3108 // FIXME: For constant StartVal, we should be able to infer 3109 // no-wrap flags. 3110 const SCEV *PHISCEV = 3111 getAddRecExpr(StartVal, AddRec->getOperand(1), L, 3112 SCEV::FlagAnyWrap); 3113 3114 // Okay, for the entire analysis of this edge we assumed the PHI 3115 // to be symbolic. We now need to go back and purge all of the 3116 // entries for the scalars that use the symbolic expression. 3117 ForgetSymbolicName(PN, SymbolicName); 3118 ValueExprMap[SCEVCallbackVH(PN, this)] = PHISCEV; 3119 return PHISCEV; 3120 } 3121 } 3122 } 3123 } 3124 } 3125 3126 // If the PHI has a single incoming value, follow that value, unless the 3127 // PHI's incoming blocks are in a different loop, in which case doing so 3128 // risks breaking LCSSA form. Instcombine would normally zap these, but 3129 // it doesn't have DominatorTree information, so it may miss cases. 3130 if (Value *V = SimplifyInstruction(PN, TD, TLI, DT)) 3131 if (LI->replacementPreservesLCSSAForm(PN, V)) 3132 return getSCEV(V); 3133 3134 // If it's not a loop phi, we can't handle it yet. 3135 return getUnknown(PN); 3136} 3137 3138/// createNodeForGEP - Expand GEP instructions into add and multiply 3139/// operations. This allows them to be analyzed by regular SCEV code. 3140/// 3141const SCEV *ScalarEvolution::createNodeForGEP(GEPOperator *GEP) { 3142 3143 // Don't blindly transfer the inbounds flag from the GEP instruction to the 3144 // Add expression, because the Instruction may be guarded by control flow 3145 // and the no-overflow bits may not be valid for the expression in any 3146 // context. 3147 bool isInBounds = GEP->isInBounds(); 3148 3149 Type *IntPtrTy = getEffectiveSCEVType(GEP->getType()); 3150 Value *Base = GEP->getOperand(0); 3151 // Don't attempt to analyze GEPs over unsized objects. 3152 if (!cast<PointerType>(Base->getType())->getElementType()->isSized()) 3153 return getUnknown(GEP); 3154 const SCEV *TotalOffset = getConstant(IntPtrTy, 0); 3155 gep_type_iterator GTI = gep_type_begin(GEP); 3156 for (GetElementPtrInst::op_iterator I = llvm::next(GEP->op_begin()), 3157 E = GEP->op_end(); 3158 I != E; ++I) { 3159 Value *Index = *I; 3160 // Compute the (potentially symbolic) offset in bytes for this index. 3161 if (StructType *STy = dyn_cast<StructType>(*GTI++)) { 3162 // For a struct, add the member offset. 3163 unsigned FieldNo = cast<ConstantInt>(Index)->getZExtValue(); 3164 const SCEV *FieldOffset = getOffsetOfExpr(STy, FieldNo); 3165 3166 // Add the field offset to the running total offset. 3167 TotalOffset = getAddExpr(TotalOffset, FieldOffset); 3168 } else { 3169 // For an array, add the element offset, explicitly scaled. 3170 const SCEV *ElementSize = getSizeOfExpr(*GTI); 3171 const SCEV *IndexS = getSCEV(Index); 3172 // Getelementptr indices are signed. 3173 IndexS = getTruncateOrSignExtend(IndexS, IntPtrTy); 3174 3175 // Multiply the index by the element size to compute the element offset. 3176 const SCEV *LocalOffset = getMulExpr(IndexS, ElementSize, 3177 isInBounds ? SCEV::FlagNSW : 3178 SCEV::FlagAnyWrap); 3179 3180 // Add the element offset to the running total offset. 3181 TotalOffset = getAddExpr(TotalOffset, LocalOffset); 3182 } 3183 } 3184 3185 // Get the SCEV for the GEP base. 3186 const SCEV *BaseS = getSCEV(Base); 3187 3188 // Add the total offset from all the GEP indices to the base. 3189 return getAddExpr(BaseS, TotalOffset, 3190 isInBounds ? SCEV::FlagNSW : SCEV::FlagAnyWrap); 3191} 3192 3193/// GetMinTrailingZeros - Determine the minimum number of zero bits that S is 3194/// guaranteed to end in (at every loop iteration). It is, at the same time, 3195/// the minimum number of times S is divisible by 2. For example, given {4,+,8} 3196/// it returns 2. If S is guaranteed to be 0, it returns the bitwidth of S. 3197uint32_t 3198ScalarEvolution::GetMinTrailingZeros(const SCEV *S) { 3199 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S)) 3200 return C->getValue()->getValue().countTrailingZeros(); 3201 3202 if (const SCEVTruncateExpr *T = dyn_cast<SCEVTruncateExpr>(S)) 3203 return std::min(GetMinTrailingZeros(T->getOperand()), 3204 (uint32_t)getTypeSizeInBits(T->getType())); 3205 3206 if (const SCEVZeroExtendExpr *E = dyn_cast<SCEVZeroExtendExpr>(S)) { 3207 uint32_t OpRes = GetMinTrailingZeros(E->getOperand()); 3208 return OpRes == getTypeSizeInBits(E->getOperand()->getType()) ? 3209 getTypeSizeInBits(E->getType()) : OpRes; 3210 } 3211 3212 if (const SCEVSignExtendExpr *E = dyn_cast<SCEVSignExtendExpr>(S)) { 3213 uint32_t OpRes = GetMinTrailingZeros(E->getOperand()); 3214 return OpRes == getTypeSizeInBits(E->getOperand()->getType()) ? 3215 getTypeSizeInBits(E->getType()) : OpRes; 3216 } 3217 3218 if (const SCEVAddExpr *A = dyn_cast<SCEVAddExpr>(S)) { 3219 // The result is the min of all operands results. 3220 uint32_t MinOpRes = GetMinTrailingZeros(A->getOperand(0)); 3221 for (unsigned i = 1, e = A->getNumOperands(); MinOpRes && i != e; ++i) 3222 MinOpRes = std::min(MinOpRes, GetMinTrailingZeros(A->getOperand(i))); 3223 return MinOpRes; 3224 } 3225 3226 if (const SCEVMulExpr *M = dyn_cast<SCEVMulExpr>(S)) { 3227 // The result is the sum of all operands results. 3228 uint32_t SumOpRes = GetMinTrailingZeros(M->getOperand(0)); 3229 uint32_t BitWidth = getTypeSizeInBits(M->getType()); 3230 for (unsigned i = 1, e = M->getNumOperands(); 3231 SumOpRes != BitWidth && i != e; ++i) 3232 SumOpRes = std::min(SumOpRes + GetMinTrailingZeros(M->getOperand(i)), 3233 BitWidth); 3234 return SumOpRes; 3235 } 3236 3237 if (const SCEVAddRecExpr *A = dyn_cast<SCEVAddRecExpr>(S)) { 3238 // The result is the min of all operands results. 3239 uint32_t MinOpRes = GetMinTrailingZeros(A->getOperand(0)); 3240 for (unsigned i = 1, e = A->getNumOperands(); MinOpRes && i != e; ++i) 3241 MinOpRes = std::min(MinOpRes, GetMinTrailingZeros(A->getOperand(i))); 3242 return MinOpRes; 3243 } 3244 3245 if (const SCEVSMaxExpr *M = dyn_cast<SCEVSMaxExpr>(S)) { 3246 // The result is the min of all operands results. 3247 uint32_t MinOpRes = GetMinTrailingZeros(M->getOperand(0)); 3248 for (unsigned i = 1, e = M->getNumOperands(); MinOpRes && i != e; ++i) 3249 MinOpRes = std::min(MinOpRes, GetMinTrailingZeros(M->getOperand(i))); 3250 return MinOpRes; 3251 } 3252 3253 if (const SCEVUMaxExpr *M = dyn_cast<SCEVUMaxExpr>(S)) { 3254 // The result is the min of all operands results. 3255 uint32_t MinOpRes = GetMinTrailingZeros(M->getOperand(0)); 3256 for (unsigned i = 1, e = M->getNumOperands(); MinOpRes && i != e; ++i) 3257 MinOpRes = std::min(MinOpRes, GetMinTrailingZeros(M->getOperand(i))); 3258 return MinOpRes; 3259 } 3260 3261 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) { 3262 // For a SCEVUnknown, ask ValueTracking. 3263 unsigned BitWidth = getTypeSizeInBits(U->getType()); 3264 APInt Zeros(BitWidth, 0), Ones(BitWidth, 0); 3265 ComputeMaskedBits(U->getValue(), Zeros, Ones); 3266 return Zeros.countTrailingOnes(); 3267 } 3268 3269 // SCEVUDivExpr 3270 return 0; 3271} 3272 3273/// getUnsignedRange - Determine the unsigned range for a particular SCEV. 3274/// 3275ConstantRange 3276ScalarEvolution::getUnsignedRange(const SCEV *S) { 3277 // See if we've computed this range already. 3278 DenseMap<const SCEV *, ConstantRange>::iterator I = UnsignedRanges.find(S); 3279 if (I != UnsignedRanges.end()) 3280 return I->second; 3281 3282 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S)) 3283 return setUnsignedRange(C, ConstantRange(C->getValue()->getValue())); 3284 3285 unsigned BitWidth = getTypeSizeInBits(S->getType()); 3286 ConstantRange ConservativeResult(BitWidth, /*isFullSet=*/true); 3287 3288 // If the value has known zeros, the maximum unsigned value will have those 3289 // known zeros as well. 3290 uint32_t TZ = GetMinTrailingZeros(S); 3291 if (TZ != 0) 3292 ConservativeResult = 3293 ConstantRange(APInt::getMinValue(BitWidth), 3294 APInt::getMaxValue(BitWidth).lshr(TZ).shl(TZ) + 1); 3295 3296 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) { 3297 ConstantRange X = getUnsignedRange(Add->getOperand(0)); 3298 for (unsigned i = 1, e = Add->getNumOperands(); i != e; ++i) 3299 X = X.add(getUnsignedRange(Add->getOperand(i))); 3300 return setUnsignedRange(Add, ConservativeResult.intersectWith(X)); 3301 } 3302 3303 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) { 3304 ConstantRange X = getUnsignedRange(Mul->getOperand(0)); 3305 for (unsigned i = 1, e = Mul->getNumOperands(); i != e; ++i) 3306 X = X.multiply(getUnsignedRange(Mul->getOperand(i))); 3307 return setUnsignedRange(Mul, ConservativeResult.intersectWith(X)); 3308 } 3309 3310 if (const SCEVSMaxExpr *SMax = dyn_cast<SCEVSMaxExpr>(S)) { 3311 ConstantRange X = getUnsignedRange(SMax->getOperand(0)); 3312 for (unsigned i = 1, e = SMax->getNumOperands(); i != e; ++i) 3313 X = X.smax(getUnsignedRange(SMax->getOperand(i))); 3314 return setUnsignedRange(SMax, ConservativeResult.intersectWith(X)); 3315 } 3316 3317 if (const SCEVUMaxExpr *UMax = dyn_cast<SCEVUMaxExpr>(S)) { 3318 ConstantRange X = getUnsignedRange(UMax->getOperand(0)); 3319 for (unsigned i = 1, e = UMax->getNumOperands(); i != e; ++i) 3320 X = X.umax(getUnsignedRange(UMax->getOperand(i))); 3321 return setUnsignedRange(UMax, ConservativeResult.intersectWith(X)); 3322 } 3323 3324 if (const SCEVUDivExpr *UDiv = dyn_cast<SCEVUDivExpr>(S)) { 3325 ConstantRange X = getUnsignedRange(UDiv->getLHS()); 3326 ConstantRange Y = getUnsignedRange(UDiv->getRHS()); 3327 return setUnsignedRange(UDiv, ConservativeResult.intersectWith(X.udiv(Y))); 3328 } 3329 3330 if (const SCEVZeroExtendExpr *ZExt = dyn_cast<SCEVZeroExtendExpr>(S)) { 3331 ConstantRange X = getUnsignedRange(ZExt->getOperand()); 3332 return setUnsignedRange(ZExt, 3333 ConservativeResult.intersectWith(X.zeroExtend(BitWidth))); 3334 } 3335 3336 if (const SCEVSignExtendExpr *SExt = dyn_cast<SCEVSignExtendExpr>(S)) { 3337 ConstantRange X = getUnsignedRange(SExt->getOperand()); 3338 return setUnsignedRange(SExt, 3339 ConservativeResult.intersectWith(X.signExtend(BitWidth))); 3340 } 3341 3342 if (const SCEVTruncateExpr *Trunc = dyn_cast<SCEVTruncateExpr>(S)) { 3343 ConstantRange X = getUnsignedRange(Trunc->getOperand()); 3344 return setUnsignedRange(Trunc, 3345 ConservativeResult.intersectWith(X.truncate(BitWidth))); 3346 } 3347 3348 if (const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(S)) { 3349 // If there's no unsigned wrap, the value will never be less than its 3350 // initial value. 3351 if (AddRec->getNoWrapFlags(SCEV::FlagNUW)) 3352 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(AddRec->getStart())) 3353 if (!C->getValue()->isZero()) 3354 ConservativeResult = 3355 ConservativeResult.intersectWith( 3356 ConstantRange(C->getValue()->getValue(), APInt(BitWidth, 0))); 3357 3358 // TODO: non-affine addrec 3359 if (AddRec->isAffine()) { 3360 Type *Ty = AddRec->getType(); 3361 const SCEV *MaxBECount = getMaxBackedgeTakenCount(AddRec->getLoop()); 3362 if (!isa<SCEVCouldNotCompute>(MaxBECount) && 3363 getTypeSizeInBits(MaxBECount->getType()) <= BitWidth) { 3364 MaxBECount = getNoopOrZeroExtend(MaxBECount, Ty); 3365 3366 const SCEV *Start = AddRec->getStart(); 3367 const SCEV *Step = AddRec->getStepRecurrence(*this); 3368 3369 ConstantRange StartRange = getUnsignedRange(Start); 3370 ConstantRange StepRange = getSignedRange(Step); 3371 ConstantRange MaxBECountRange = getUnsignedRange(MaxBECount); 3372 ConstantRange EndRange = 3373 StartRange.add(MaxBECountRange.multiply(StepRange)); 3374 3375 // Check for overflow. This must be done with ConstantRange arithmetic 3376 // because we could be called from within the ScalarEvolution overflow 3377 // checking code. 3378 ConstantRange ExtStartRange = StartRange.zextOrTrunc(BitWidth*2+1); 3379 ConstantRange ExtStepRange = StepRange.sextOrTrunc(BitWidth*2+1); 3380 ConstantRange ExtMaxBECountRange = 3381 MaxBECountRange.zextOrTrunc(BitWidth*2+1); 3382 ConstantRange ExtEndRange = EndRange.zextOrTrunc(BitWidth*2+1); 3383 if (ExtStartRange.add(ExtMaxBECountRange.multiply(ExtStepRange)) != 3384 ExtEndRange) 3385 return setUnsignedRange(AddRec, ConservativeResult); 3386 3387 APInt Min = APIntOps::umin(StartRange.getUnsignedMin(), 3388 EndRange.getUnsignedMin()); 3389 APInt Max = APIntOps::umax(StartRange.getUnsignedMax(), 3390 EndRange.getUnsignedMax()); 3391 if (Min.isMinValue() && Max.isMaxValue()) 3392 return setUnsignedRange(AddRec, ConservativeResult); 3393 return setUnsignedRange(AddRec, 3394 ConservativeResult.intersectWith(ConstantRange(Min, Max+1))); 3395 } 3396 } 3397 3398 return setUnsignedRange(AddRec, ConservativeResult); 3399 } 3400 3401 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) { 3402 // For a SCEVUnknown, ask ValueTracking. 3403 APInt Zeros(BitWidth, 0), Ones(BitWidth, 0); 3404 ComputeMaskedBits(U->getValue(), Zeros, Ones, TD); 3405 if (Ones == ~Zeros + 1) 3406 return setUnsignedRange(U, ConservativeResult); 3407 return setUnsignedRange(U, 3408 ConservativeResult.intersectWith(ConstantRange(Ones, ~Zeros + 1))); 3409 } 3410 3411 return setUnsignedRange(S, ConservativeResult); 3412} 3413 3414/// getSignedRange - Determine the signed range for a particular SCEV. 3415/// 3416ConstantRange 3417ScalarEvolution::getSignedRange(const SCEV *S) { 3418 // See if we've computed this range already. 3419 DenseMap<const SCEV *, ConstantRange>::iterator I = SignedRanges.find(S); 3420 if (I != SignedRanges.end()) 3421 return I->second; 3422 3423 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(S)) 3424 return setSignedRange(C, ConstantRange(C->getValue()->getValue())); 3425 3426 unsigned BitWidth = getTypeSizeInBits(S->getType()); 3427 ConstantRange ConservativeResult(BitWidth, /*isFullSet=*/true); 3428 3429 // If the value has known zeros, the maximum signed value will have those 3430 // known zeros as well. 3431 uint32_t TZ = GetMinTrailingZeros(S); 3432 if (TZ != 0) 3433 ConservativeResult = 3434 ConstantRange(APInt::getSignedMinValue(BitWidth), 3435 APInt::getSignedMaxValue(BitWidth).ashr(TZ).shl(TZ) + 1); 3436 3437 if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) { 3438 ConstantRange X = getSignedRange(Add->getOperand(0)); 3439 for (unsigned i = 1, e = Add->getNumOperands(); i != e; ++i) 3440 X = X.add(getSignedRange(Add->getOperand(i))); 3441 return setSignedRange(Add, ConservativeResult.intersectWith(X)); 3442 } 3443 3444 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S)) { 3445 ConstantRange X = getSignedRange(Mul->getOperand(0)); 3446 for (unsigned i = 1, e = Mul->getNumOperands(); i != e; ++i) 3447 X = X.multiply(getSignedRange(Mul->getOperand(i))); 3448 return setSignedRange(Mul, ConservativeResult.intersectWith(X)); 3449 } 3450 3451 if (const SCEVSMaxExpr *SMax = dyn_cast<SCEVSMaxExpr>(S)) { 3452 ConstantRange X = getSignedRange(SMax->getOperand(0)); 3453 for (unsigned i = 1, e = SMax->getNumOperands(); i != e; ++i) 3454 X = X.smax(getSignedRange(SMax->getOperand(i))); 3455 return setSignedRange(SMax, ConservativeResult.intersectWith(X)); 3456 } 3457 3458 if (const SCEVUMaxExpr *UMax = dyn_cast<SCEVUMaxExpr>(S)) { 3459 ConstantRange X = getSignedRange(UMax->getOperand(0)); 3460 for (unsigned i = 1, e = UMax->getNumOperands(); i != e; ++i) 3461 X = X.umax(getSignedRange(UMax->getOperand(i))); 3462 return setSignedRange(UMax, ConservativeResult.intersectWith(X)); 3463 } 3464 3465 if (const SCEVUDivExpr *UDiv = dyn_cast<SCEVUDivExpr>(S)) { 3466 ConstantRange X = getSignedRange(UDiv->getLHS()); 3467 ConstantRange Y = getSignedRange(UDiv->getRHS()); 3468 return setSignedRange(UDiv, ConservativeResult.intersectWith(X.udiv(Y))); 3469 } 3470 3471 if (const SCEVZeroExtendExpr *ZExt = dyn_cast<SCEVZeroExtendExpr>(S)) { 3472 ConstantRange X = getSignedRange(ZExt->getOperand()); 3473 return setSignedRange(ZExt, 3474 ConservativeResult.intersectWith(X.zeroExtend(BitWidth))); 3475 } 3476 3477 if (const SCEVSignExtendExpr *SExt = dyn_cast<SCEVSignExtendExpr>(S)) { 3478 ConstantRange X = getSignedRange(SExt->getOperand()); 3479 return setSignedRange(SExt, 3480 ConservativeResult.intersectWith(X.signExtend(BitWidth))); 3481 } 3482 3483 if (const SCEVTruncateExpr *Trunc = dyn_cast<SCEVTruncateExpr>(S)) { 3484 ConstantRange X = getSignedRange(Trunc->getOperand()); 3485 return setSignedRange(Trunc, 3486 ConservativeResult.intersectWith(X.truncate(BitWidth))); 3487 } 3488 3489 if (const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(S)) { 3490 // If there's no signed wrap, and all the operands have the same sign or 3491 // zero, the value won't ever change sign. 3492 if (AddRec->getNoWrapFlags(SCEV::FlagNSW)) { 3493 bool AllNonNeg = true; 3494 bool AllNonPos = true; 3495 for (unsigned i = 0, e = AddRec->getNumOperands(); i != e; ++i) { 3496 if (!isKnownNonNegative(AddRec->getOperand(i))) AllNonNeg = false; 3497 if (!isKnownNonPositive(AddRec->getOperand(i))) AllNonPos = false; 3498 } 3499 if (AllNonNeg) 3500 ConservativeResult = ConservativeResult.intersectWith( 3501 ConstantRange(APInt(BitWidth, 0), 3502 APInt::getSignedMinValue(BitWidth))); 3503 else if (AllNonPos) 3504 ConservativeResult = ConservativeResult.intersectWith( 3505 ConstantRange(APInt::getSignedMinValue(BitWidth), 3506 APInt(BitWidth, 1))); 3507 } 3508 3509 // TODO: non-affine addrec 3510 if (AddRec->isAffine()) { 3511 Type *Ty = AddRec->getType(); 3512 const SCEV *MaxBECount = getMaxBackedgeTakenCount(AddRec->getLoop()); 3513 if (!isa<SCEVCouldNotCompute>(MaxBECount) && 3514 getTypeSizeInBits(MaxBECount->getType()) <= BitWidth) { 3515 MaxBECount = getNoopOrZeroExtend(MaxBECount, Ty); 3516 3517 const SCEV *Start = AddRec->getStart(); 3518 const SCEV *Step = AddRec->getStepRecurrence(*this); 3519 3520 ConstantRange StartRange = getSignedRange(Start); 3521 ConstantRange StepRange = getSignedRange(Step); 3522 ConstantRange MaxBECountRange = getUnsignedRange(MaxBECount); 3523 ConstantRange EndRange = 3524 StartRange.add(MaxBECountRange.multiply(StepRange)); 3525 3526 // Check for overflow. This must be done with ConstantRange arithmetic 3527 // because we could be called from within the ScalarEvolution overflow 3528 // checking code. 3529 ConstantRange ExtStartRange = StartRange.sextOrTrunc(BitWidth*2+1); 3530 ConstantRange ExtStepRange = StepRange.sextOrTrunc(BitWidth*2+1); 3531 ConstantRange ExtMaxBECountRange = 3532 MaxBECountRange.zextOrTrunc(BitWidth*2+1); 3533 ConstantRange ExtEndRange = EndRange.sextOrTrunc(BitWidth*2+1); 3534 if (ExtStartRange.add(ExtMaxBECountRange.multiply(ExtStepRange)) != 3535 ExtEndRange) 3536 return setSignedRange(AddRec, ConservativeResult); 3537 3538 APInt Min = APIntOps::smin(StartRange.getSignedMin(), 3539 EndRange.getSignedMin()); 3540 APInt Max = APIntOps::smax(StartRange.getSignedMax(), 3541 EndRange.getSignedMax()); 3542 if (Min.isMinSignedValue() && Max.isMaxSignedValue()) 3543 return setSignedRange(AddRec, ConservativeResult); 3544 return setSignedRange(AddRec, 3545 ConservativeResult.intersectWith(ConstantRange(Min, Max+1))); 3546 } 3547 } 3548 3549 return setSignedRange(AddRec, ConservativeResult); 3550 } 3551 3552 if (const SCEVUnknown *U = dyn_cast<SCEVUnknown>(S)) { 3553 // For a SCEVUnknown, ask ValueTracking. 3554 if (!U->getValue()->getType()->isIntegerTy() && !TD) 3555 return setSignedRange(U, ConservativeResult); 3556 unsigned NS = ComputeNumSignBits(U->getValue(), TD); 3557 if (NS == 1) 3558 return setSignedRange(U, ConservativeResult); 3559 return setSignedRange(U, ConservativeResult.intersectWith( 3560 ConstantRange(APInt::getSignedMinValue(BitWidth).ashr(NS - 1), 3561 APInt::getSignedMaxValue(BitWidth).ashr(NS - 1)+1))); 3562 } 3563 3564 return setSignedRange(S, ConservativeResult); 3565} 3566 3567/// createSCEV - We know that there is no SCEV for the specified value. 3568/// Analyze the expression. 3569/// 3570const SCEV *ScalarEvolution::createSCEV(Value *V) { 3571 if (!isSCEVable(V->getType())) 3572 return getUnknown(V); 3573 3574 unsigned Opcode = Instruction::UserOp1; 3575 if (Instruction *I = dyn_cast<Instruction>(V)) { 3576 Opcode = I->getOpcode(); 3577 3578 // Don't attempt to analyze instructions in blocks that aren't 3579 // reachable. Such instructions don't matter, and they aren't required 3580 // to obey basic rules for definitions dominating uses which this 3581 // analysis depends on. 3582 if (!DT->isReachableFromEntry(I->getParent())) 3583 return getUnknown(V); 3584 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) 3585 Opcode = CE->getOpcode(); 3586 else if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) 3587 return getConstant(CI); 3588 else if (isa<ConstantPointerNull>(V)) 3589 return getConstant(V->getType(), 0); 3590 else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) 3591 return GA->mayBeOverridden() ? getUnknown(V) : getSCEV(GA->getAliasee()); 3592 else 3593 return getUnknown(V); 3594 3595 Operator *U = cast<Operator>(V); 3596 switch (Opcode) { 3597 case Instruction::Add: { 3598 // The simple thing to do would be to just call getSCEV on both operands 3599 // and call getAddExpr with the result. However if we're looking at a 3600 // bunch of things all added together, this can be quite inefficient, 3601 // because it leads to N-1 getAddExpr calls for N ultimate operands. 3602 // Instead, gather up all the operands and make a single getAddExpr call. 3603 // LLVM IR canonical form means we need only traverse the left operands. 3604 // 3605 // Don't apply this instruction's NSW or NUW flags to the new 3606 // expression. The instruction may be guarded by control flow that the 3607 // no-wrap behavior depends on. Non-control-equivalent instructions can be 3608 // mapped to the same SCEV expression, and it would be incorrect to transfer 3609 // NSW/NUW semantics to those operations. 3610 SmallVector<const SCEV *, 4> AddOps; 3611 AddOps.push_back(getSCEV(U->getOperand(1))); 3612 for (Value *Op = U->getOperand(0); ; Op = U->getOperand(0)) { 3613 unsigned Opcode = Op->getValueID() - Value::InstructionVal; 3614 if (Opcode != Instruction::Add && Opcode != Instruction::Sub) 3615 break; 3616 U = cast<Operator>(Op); 3617 const SCEV *Op1 = getSCEV(U->getOperand(1)); 3618 if (Opcode == Instruction::Sub) 3619 AddOps.push_back(getNegativeSCEV(Op1)); 3620 else 3621 AddOps.push_back(Op1); 3622 } 3623 AddOps.push_back(getSCEV(U->getOperand(0))); 3624 return getAddExpr(AddOps); 3625 } 3626 case Instruction::Mul: { 3627 // Don't transfer NSW/NUW for the same reason as AddExpr. 3628 SmallVector<const SCEV *, 4> MulOps; 3629 MulOps.push_back(getSCEV(U->getOperand(1))); 3630 for (Value *Op = U->getOperand(0); 3631 Op->getValueID() == Instruction::Mul + Value::InstructionVal; 3632 Op = U->getOperand(0)) { 3633 U = cast<Operator>(Op); 3634 MulOps.push_back(getSCEV(U->getOperand(1))); 3635 } 3636 MulOps.push_back(getSCEV(U->getOperand(0))); 3637 return getMulExpr(MulOps); 3638 } 3639 case Instruction::UDiv: 3640 return getUDivExpr(getSCEV(U->getOperand(0)), 3641 getSCEV(U->getOperand(1))); 3642 case Instruction::Sub: 3643 return getMinusSCEV(getSCEV(U->getOperand(0)), 3644 getSCEV(U->getOperand(1))); 3645 case Instruction::And: 3646 // For an expression like x&255 that merely masks off the high bits, 3647 // use zext(trunc(x)) as the SCEV expression. 3648 if (ConstantInt *CI = dyn_cast<ConstantInt>(U->getOperand(1))) { 3649 if (CI->isNullValue()) 3650 return getSCEV(U->getOperand(1)); 3651 if (CI->isAllOnesValue()) 3652 return getSCEV(U->getOperand(0)); 3653 const APInt &A = CI->getValue(); 3654 3655 // Instcombine's ShrinkDemandedConstant may strip bits out of 3656 // constants, obscuring what would otherwise be a low-bits mask. 3657 // Use ComputeMaskedBits to compute what ShrinkDemandedConstant 3658 // knew about to reconstruct a low-bits mask value. 3659 unsigned LZ = A.countLeadingZeros(); 3660 unsigned BitWidth = A.getBitWidth(); 3661 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0); 3662 ComputeMaskedBits(U->getOperand(0), KnownZero, KnownOne, TD); 3663 3664 APInt EffectiveMask = APInt::getLowBitsSet(BitWidth, BitWidth - LZ); 3665 3666 if (LZ != 0 && !((~A & ~KnownZero) & EffectiveMask)) 3667 return 3668 getZeroExtendExpr(getTruncateExpr(getSCEV(U->getOperand(0)), 3669 IntegerType::get(getContext(), BitWidth - LZ)), 3670 U->getType()); 3671 } 3672 break; 3673 3674 case Instruction::Or: 3675 // If the RHS of the Or is a constant, we may have something like: 3676 // X*4+1 which got turned into X*4|1. Handle this as an Add so loop 3677 // optimizations will transparently handle this case. 3678 // 3679 // In order for this transformation to be safe, the LHS must be of the 3680 // form X*(2^n) and the Or constant must be less than 2^n. 3681 if (ConstantInt *CI = dyn_cast<ConstantInt>(U->getOperand(1))) { 3682 const SCEV *LHS = getSCEV(U->getOperand(0)); 3683 const APInt &CIVal = CI->getValue(); 3684 if (GetMinTrailingZeros(LHS) >= 3685 (CIVal.getBitWidth() - CIVal.countLeadingZeros())) { 3686 // Build a plain add SCEV. 3687 const SCEV *S = getAddExpr(LHS, getSCEV(CI)); 3688 // If the LHS of the add was an addrec and it has no-wrap flags, 3689 // transfer the no-wrap flags, since an or won't introduce a wrap. 3690 if (const SCEVAddRecExpr *NewAR = dyn_cast<SCEVAddRecExpr>(S)) { 3691 const SCEVAddRecExpr *OldAR = cast<SCEVAddRecExpr>(LHS); 3692 const_cast<SCEVAddRecExpr *>(NewAR)->setNoWrapFlags( 3693 OldAR->getNoWrapFlags()); 3694 } 3695 return S; 3696 } 3697 } 3698 break; 3699 case Instruction::Xor: 3700 if (ConstantInt *CI = dyn_cast<ConstantInt>(U->getOperand(1))) { 3701 // If the RHS of the xor is a signbit, then this is just an add. 3702 // Instcombine turns add of signbit into xor as a strength reduction step. 3703 if (CI->getValue().isSignBit()) 3704 return getAddExpr(getSCEV(U->getOperand(0)), 3705 getSCEV(U->getOperand(1))); 3706 3707 // If the RHS of xor is -1, then this is a not operation. 3708 if (CI->isAllOnesValue()) 3709 return getNotSCEV(getSCEV(U->getOperand(0))); 3710 3711 // Model xor(and(x, C), C) as and(~x, C), if C is a low-bits mask. 3712 // This is a variant of the check for xor with -1, and it handles 3713 // the case where instcombine has trimmed non-demanded bits out 3714 // of an xor with -1. 3715 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(U->getOperand(0))) 3716 if (ConstantInt *LCI = dyn_cast<ConstantInt>(BO->getOperand(1))) 3717 if (BO->getOpcode() == Instruction::And && 3718 LCI->getValue() == CI->getValue()) 3719 if (const SCEVZeroExtendExpr *Z = 3720 dyn_cast<SCEVZeroExtendExpr>(getSCEV(U->getOperand(0)))) { 3721 Type *UTy = U->getType(); 3722 const SCEV *Z0 = Z->getOperand(); 3723 Type *Z0Ty = Z0->getType(); 3724 unsigned Z0TySize = getTypeSizeInBits(Z0Ty); 3725 3726 // If C is a low-bits mask, the zero extend is serving to 3727 // mask off the high bits. Complement the operand and 3728 // re-apply the zext. 3729 if (APIntOps::isMask(Z0TySize, CI->getValue())) 3730 return getZeroExtendExpr(getNotSCEV(Z0), UTy); 3731 3732 // If C is a single bit, it may be in the sign-bit position 3733 // before the zero-extend. In this case, represent the xor 3734 // using an add, which is equivalent, and re-apply the zext. 3735 APInt Trunc = CI->getValue().trunc(Z0TySize); 3736 if (Trunc.zext(getTypeSizeInBits(UTy)) == CI->getValue() && 3737 Trunc.isSignBit()) 3738 return getZeroExtendExpr(getAddExpr(Z0, getConstant(Trunc)), 3739 UTy); 3740 } 3741 } 3742 break; 3743 3744 case Instruction::Shl: 3745 // Turn shift left of a constant amount into a multiply. 3746 if (ConstantInt *SA = dyn_cast<ConstantInt>(U->getOperand(1))) { 3747 uint32_t BitWidth = cast<IntegerType>(U->getType())->getBitWidth(); 3748 3749 // If the shift count is not less than the bitwidth, the result of 3750 // the shift is undefined. Don't try to analyze it, because the 3751 // resolution chosen here may differ from the resolution chosen in 3752 // other parts of the compiler. 3753 if (SA->getValue().uge(BitWidth)) 3754 break; 3755 3756 Constant *X = ConstantInt::get(getContext(), 3757 APInt(BitWidth, 1).shl(SA->getZExtValue())); 3758 return getMulExpr(getSCEV(U->getOperand(0)), getSCEV(X)); 3759 } 3760 break; 3761 3762 case Instruction::LShr: 3763 // Turn logical shift right of a constant into a unsigned divide. 3764 if (ConstantInt *SA = dyn_cast<ConstantInt>(U->getOperand(1))) { 3765 uint32_t BitWidth = cast<IntegerType>(U->getType())->getBitWidth(); 3766 3767 // If the shift count is not less than the bitwidth, the result of 3768 // the shift is undefined. Don't try to analyze it, because the 3769 // resolution chosen here may differ from the resolution chosen in 3770 // other parts of the compiler. 3771 if (SA->getValue().uge(BitWidth)) 3772 break; 3773 3774 Constant *X = ConstantInt::get(getContext(), 3775 APInt(BitWidth, 1).shl(SA->getZExtValue())); 3776 return getUDivExpr(getSCEV(U->getOperand(0)), getSCEV(X)); 3777 } 3778 break; 3779 3780 case Instruction::AShr: 3781 // For a two-shift sext-inreg, use sext(trunc(x)) as the SCEV expression. 3782 if (ConstantInt *CI = dyn_cast<ConstantInt>(U->getOperand(1))) 3783 if (Operator *L = dyn_cast<Operator>(U->getOperand(0))) 3784 if (L->getOpcode() == Instruction::Shl && 3785 L->getOperand(1) == U->getOperand(1)) { 3786 uint64_t BitWidth = getTypeSizeInBits(U->getType()); 3787 3788 // If the shift count is not less than the bitwidth, the result of 3789 // the shift is undefined. Don't try to analyze it, because the 3790 // resolution chosen here may differ from the resolution chosen in 3791 // other parts of the compiler. 3792 if (CI->getValue().uge(BitWidth)) 3793 break; 3794 3795 uint64_t Amt = BitWidth - CI->getZExtValue(); 3796 if (Amt == BitWidth) 3797 return getSCEV(L->getOperand(0)); // shift by zero --> noop 3798 return 3799 getSignExtendExpr(getTruncateExpr(getSCEV(L->getOperand(0)), 3800 IntegerType::get(getContext(), 3801 Amt)), 3802 U->getType()); 3803 } 3804 break; 3805 3806 case Instruction::Trunc: 3807 return getTruncateExpr(getSCEV(U->getOperand(0)), U->getType()); 3808 3809 case Instruction::ZExt: 3810 return getZeroExtendExpr(getSCEV(U->getOperand(0)), U->getType()); 3811 3812 case Instruction::SExt: 3813 return getSignExtendExpr(getSCEV(U->getOperand(0)), U->getType()); 3814 3815 case Instruction::BitCast: 3816 // BitCasts are no-op casts so we just eliminate the cast. 3817 if (isSCEVable(U->getType()) && isSCEVable(U->getOperand(0)->getType())) 3818 return getSCEV(U->getOperand(0)); 3819 break; 3820 3821 // It's tempting to handle inttoptr and ptrtoint as no-ops, however this can 3822 // lead to pointer expressions which cannot safely be expanded to GEPs, 3823 // because ScalarEvolution doesn't respect the GEP aliasing rules when 3824 // simplifying integer expressions. 3825 3826 case Instruction::GetElementPtr: 3827 return createNodeForGEP(cast<GEPOperator>(U)); 3828 3829 case Instruction::PHI: 3830 return createNodeForPHI(cast<PHINode>(U)); 3831 3832 case Instruction::Select: 3833 // This could be a smax or umax that was lowered earlier. 3834 // Try to recover it. 3835 if (ICmpInst *ICI = dyn_cast<ICmpInst>(U->getOperand(0))) { 3836 Value *LHS = ICI->getOperand(0); 3837 Value *RHS = ICI->getOperand(1); 3838 switch (ICI->getPredicate()) { 3839 case ICmpInst::ICMP_SLT: 3840 case ICmpInst::ICMP_SLE: 3841 std::swap(LHS, RHS); 3842 // fall through 3843 case ICmpInst::ICMP_SGT: 3844 case ICmpInst::ICMP_SGE: 3845 // a >s b ? a+x : b+x -> smax(a, b)+x 3846 // a >s b ? b+x : a+x -> smin(a, b)+x 3847 if (LHS->getType() == U->getType()) { 3848 const SCEV *LS = getSCEV(LHS); 3849 const SCEV *RS = getSCEV(RHS); 3850 const SCEV *LA = getSCEV(U->getOperand(1)); 3851 const SCEV *RA = getSCEV(U->getOperand(2)); 3852 const SCEV *LDiff = getMinusSCEV(LA, LS); 3853 const SCEV *RDiff = getMinusSCEV(RA, RS); 3854 if (LDiff == RDiff) 3855 return getAddExpr(getSMaxExpr(LS, RS), LDiff); 3856 LDiff = getMinusSCEV(LA, RS); 3857 RDiff = getMinusSCEV(RA, LS); 3858 if (LDiff == RDiff) 3859 return getAddExpr(getSMinExpr(LS, RS), LDiff); 3860 } 3861 break; 3862 case ICmpInst::ICMP_ULT: 3863 case ICmpInst::ICMP_ULE: 3864 std::swap(LHS, RHS); 3865 // fall through 3866 case ICmpInst::ICMP_UGT: 3867 case ICmpInst::ICMP_UGE: 3868 // a >u b ? a+x : b+x -> umax(a, b)+x 3869 // a >u b ? b+x : a+x -> umin(a, b)+x 3870 if (LHS->getType() == U->getType()) { 3871 const SCEV *LS = getSCEV(LHS); 3872 const SCEV *RS = getSCEV(RHS); 3873 const SCEV *LA = getSCEV(U->getOperand(1)); 3874 const SCEV *RA = getSCEV(U->getOperand(2)); 3875 const SCEV *LDiff = getMinusSCEV(LA, LS); 3876 const SCEV *RDiff = getMinusSCEV(RA, RS); 3877 if (LDiff == RDiff) 3878 return getAddExpr(getUMaxExpr(LS, RS), LDiff); 3879 LDiff = getMinusSCEV(LA, RS); 3880 RDiff = getMinusSCEV(RA, LS); 3881 if (LDiff == RDiff) 3882 return getAddExpr(getUMinExpr(LS, RS), LDiff); 3883 } 3884 break; 3885 case ICmpInst::ICMP_NE: 3886 // n != 0 ? n+x : 1+x -> umax(n, 1)+x 3887 if (LHS->getType() == U->getType() && 3888 isa<ConstantInt>(RHS) && 3889 cast<ConstantInt>(RHS)->isZero()) { 3890 const SCEV *One = getConstant(LHS->getType(), 1); 3891 const SCEV *LS = getSCEV(LHS); 3892 const SCEV *LA = getSCEV(U->getOperand(1)); 3893 const SCEV *RA = getSCEV(U->getOperand(2)); 3894 const SCEV *LDiff = getMinusSCEV(LA, LS); 3895 const SCEV *RDiff = getMinusSCEV(RA, One); 3896 if (LDiff == RDiff) 3897 return getAddExpr(getUMaxExpr(One, LS), LDiff); 3898 } 3899 break; 3900 case ICmpInst::ICMP_EQ: 3901 // n == 0 ? 1+x : n+x -> umax(n, 1)+x 3902 if (LHS->getType() == U->getType() && 3903 isa<ConstantInt>(RHS) && 3904 cast<ConstantInt>(RHS)->isZero()) { 3905 const SCEV *One = getConstant(LHS->getType(), 1); 3906 const SCEV *LS = getSCEV(LHS); 3907 const SCEV *LA = getSCEV(U->getOperand(1)); 3908 const SCEV *RA = getSCEV(U->getOperand(2)); 3909 const SCEV *LDiff = getMinusSCEV(LA, One); 3910 const SCEV *RDiff = getMinusSCEV(RA, LS); 3911 if (LDiff == RDiff) 3912 return getAddExpr(getUMaxExpr(One, LS), LDiff); 3913 } 3914 break; 3915 default: 3916 break; 3917 } 3918 } 3919 3920 default: // We cannot analyze this expression. 3921 break; 3922 } 3923 3924 return getUnknown(V); 3925} 3926 3927 3928 3929//===----------------------------------------------------------------------===// 3930// Iteration Count Computation Code 3931// 3932 3933/// getSmallConstantTripCount - Returns the maximum trip count of this loop as a 3934/// normal unsigned value. Returns 0 if the trip count is unknown or not 3935/// constant. Will also return 0 if the maximum trip count is very large (>= 3936/// 2^32). 3937/// 3938/// This "trip count" assumes that control exits via ExitingBlock. More 3939/// precisely, it is the number of times that control may reach ExitingBlock 3940/// before taking the branch. For loops with multiple exits, it may not be the 3941/// number times that the loop header executes because the loop may exit 3942/// prematurely via another branch. 3943unsigned ScalarEvolution:: 3944getSmallConstantTripCount(Loop *L, BasicBlock *ExitingBlock) { 3945 const SCEVConstant *ExitCount = 3946 dyn_cast<SCEVConstant>(getExitCount(L, ExitingBlock)); 3947 if (!ExitCount) 3948 return 0; 3949 3950 ConstantInt *ExitConst = ExitCount->getValue(); 3951 3952 // Guard against huge trip counts. 3953 if (ExitConst->getValue().getActiveBits() > 32) 3954 return 0; 3955 3956 // In case of integer overflow, this returns 0, which is correct. 3957 return ((unsigned)ExitConst->getZExtValue()) + 1; 3958} 3959 3960/// getSmallConstantTripMultiple - Returns the largest constant divisor of the 3961/// trip count of this loop as a normal unsigned value, if possible. This 3962/// means that the actual trip count is always a multiple of the returned 3963/// value (don't forget the trip count could very well be zero as well!). 3964/// 3965/// Returns 1 if the trip count is unknown or not guaranteed to be the 3966/// multiple of a constant (which is also the case if the trip count is simply 3967/// constant, use getSmallConstantTripCount for that case), Will also return 1 3968/// if the trip count is very large (>= 2^32). 3969/// 3970/// As explained in the comments for getSmallConstantTripCount, this assumes 3971/// that control exits the loop via ExitingBlock. 3972unsigned ScalarEvolution:: 3973getSmallConstantTripMultiple(Loop *L, BasicBlock *ExitingBlock) { 3974 const SCEV *ExitCount = getExitCount(L, ExitingBlock); 3975 if (ExitCount == getCouldNotCompute()) 3976 return 1; 3977 3978 // Get the trip count from the BE count by adding 1. 3979 const SCEV *TCMul = getAddExpr(ExitCount, 3980 getConstant(ExitCount->getType(), 1)); 3981 // FIXME: SCEV distributes multiplication as V1*C1 + V2*C1. We could attempt 3982 // to factor simple cases. 3983 if (const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(TCMul)) 3984 TCMul = Mul->getOperand(0); 3985 3986 const SCEVConstant *MulC = dyn_cast<SCEVConstant>(TCMul); 3987 if (!MulC) 3988 return 1; 3989 3990 ConstantInt *Result = MulC->getValue(); 3991 3992 // Guard against huge trip counts. 3993 if (!Result || Result->getValue().getActiveBits() > 32) 3994 return 1; 3995 3996 return (unsigned)Result->getZExtValue(); 3997} 3998 3999// getExitCount - Get the expression for the number of loop iterations for which 4000// this loop is guaranteed not to exit via ExitintBlock. Otherwise return 4001// SCEVCouldNotCompute. 4002const SCEV *ScalarEvolution::getExitCount(Loop *L, BasicBlock *ExitingBlock) { 4003 return getBackedgeTakenInfo(L).getExact(ExitingBlock, this); 4004} 4005 4006/// getBackedgeTakenCount - If the specified loop has a predictable 4007/// backedge-taken count, return it, otherwise return a SCEVCouldNotCompute 4008/// object. The backedge-taken count is the number of times the loop header 4009/// will be branched to from within the loop. This is one less than the 4010/// trip count of the loop, since it doesn't count the first iteration, 4011/// when the header is branched to from outside the loop. 4012/// 4013/// Note that it is not valid to call this method on a loop without a 4014/// loop-invariant backedge-taken count (see 4015/// hasLoopInvariantBackedgeTakenCount). 4016/// 4017const SCEV *ScalarEvolution::getBackedgeTakenCount(const Loop *L) { 4018 return getBackedgeTakenInfo(L).getExact(this); 4019} 4020 4021/// getMaxBackedgeTakenCount - Similar to getBackedgeTakenCount, except 4022/// return the least SCEV value that is known never to be less than the 4023/// actual backedge taken count. 4024const SCEV *ScalarEvolution::getMaxBackedgeTakenCount(const Loop *L) { 4025 return getBackedgeTakenInfo(L).getMax(this); 4026} 4027 4028/// PushLoopPHIs - Push PHI nodes in the header of the given loop 4029/// onto the given Worklist. 4030static void 4031PushLoopPHIs(const Loop *L, SmallVectorImpl<Instruction *> &Worklist) { 4032 BasicBlock *Header = L->getHeader(); 4033 4034 // Push all Loop-header PHIs onto the Worklist stack. 4035 for (BasicBlock::iterator I = Header->begin(); 4036 PHINode *PN = dyn_cast<PHINode>(I); ++I) 4037 Worklist.push_back(PN); 4038} 4039 4040const ScalarEvolution::BackedgeTakenInfo & 4041ScalarEvolution::getBackedgeTakenInfo(const Loop *L) { 4042 // Initially insert an invalid entry for this loop. If the insertion 4043 // succeeds, proceed to actually compute a backedge-taken count and 4044 // update the value. The temporary CouldNotCompute value tells SCEV 4045 // code elsewhere that it shouldn't attempt to request a new 4046 // backedge-taken count, which could result in infinite recursion. 4047 std::pair<DenseMap<const Loop *, BackedgeTakenInfo>::iterator, bool> Pair = 4048 BackedgeTakenCounts.insert(std::make_pair(L, BackedgeTakenInfo())); 4049 if (!Pair.second) 4050 return Pair.first->second; 4051 4052 // ComputeBackedgeTakenCount may allocate memory for its result. Inserting it 4053 // into the BackedgeTakenCounts map transfers ownership. Otherwise, the result 4054 // must be cleared in this scope. 4055 BackedgeTakenInfo Result = ComputeBackedgeTakenCount(L); 4056 4057 if (Result.getExact(this) != getCouldNotCompute()) { 4058 assert(isLoopInvariant(Result.getExact(this), L) && 4059 isLoopInvariant(Result.getMax(this), L) && 4060 "Computed backedge-taken count isn't loop invariant for loop!"); 4061 ++NumTripCountsComputed; 4062 } 4063 else if (Result.getMax(this) == getCouldNotCompute() && 4064 isa<PHINode>(L->getHeader()->begin())) { 4065 // Only count loops that have phi nodes as not being computable. 4066 ++NumTripCountsNotComputed; 4067 } 4068 4069 // Now that we know more about the trip count for this loop, forget any 4070 // existing SCEV values for PHI nodes in this loop since they are only 4071 // conservative estimates made without the benefit of trip count 4072 // information. This is similar to the code in forgetLoop, except that 4073 // it handles SCEVUnknown PHI nodes specially. 4074 if (Result.hasAnyInfo()) { 4075 SmallVector<Instruction *, 16> Worklist; 4076 PushLoopPHIs(L, Worklist); 4077 4078 SmallPtrSet<Instruction *, 8> Visited; 4079 while (!Worklist.empty()) { 4080 Instruction *I = Worklist.pop_back_val(); 4081 if (!Visited.insert(I)) continue; 4082 4083 ValueExprMapType::iterator It = 4084 ValueExprMap.find(static_cast<Value *>(I)); 4085 if (It != ValueExprMap.end()) { 4086 const SCEV *Old = It->second; 4087 4088 // SCEVUnknown for a PHI either means that it has an unrecognized 4089 // structure, or it's a PHI that's in the progress of being computed 4090 // by createNodeForPHI. In the former case, additional loop trip 4091 // count information isn't going to change anything. In the later 4092 // case, createNodeForPHI will perform the necessary updates on its 4093 // own when it gets to that point. 4094 if (!isa<PHINode>(I) || !isa<SCEVUnknown>(Old)) { 4095 forgetMemoizedResults(Old); 4096 ValueExprMap.erase(It); 4097 } 4098 if (PHINode *PN = dyn_cast<PHINode>(I)) 4099 ConstantEvolutionLoopExitValue.erase(PN); 4100 } 4101 4102 PushDefUseChildren(I, Worklist); 4103 } 4104 } 4105 4106 // Re-lookup the insert position, since the call to 4107 // ComputeBackedgeTakenCount above could result in a 4108 // recusive call to getBackedgeTakenInfo (on a different 4109 // loop), which would invalidate the iterator computed 4110 // earlier. 4111 return BackedgeTakenCounts.find(L)->second = Result; 4112} 4113 4114/// forgetLoop - This method should be called by the client when it has 4115/// changed a loop in a way that may effect ScalarEvolution's ability to 4116/// compute a trip count, or if the loop is deleted. 4117void ScalarEvolution::forgetLoop(const Loop *L) { 4118 // Drop any stored trip count value. 4119 DenseMap<const Loop*, BackedgeTakenInfo>::iterator BTCPos = 4120 BackedgeTakenCounts.find(L); 4121 if (BTCPos != BackedgeTakenCounts.end()) { 4122 BTCPos->second.clear(); 4123 BackedgeTakenCounts.erase(BTCPos); 4124 } 4125 4126 // Drop information about expressions based on loop-header PHIs. 4127 SmallVector<Instruction *, 16> Worklist; 4128 PushLoopPHIs(L, Worklist); 4129 4130 SmallPtrSet<Instruction *, 8> Visited; 4131 while (!Worklist.empty()) { 4132 Instruction *I = Worklist.pop_back_val(); 4133 if (!Visited.insert(I)) continue; 4134 4135 ValueExprMapType::iterator It = ValueExprMap.find(static_cast<Value *>(I)); 4136 if (It != ValueExprMap.end()) { 4137 forgetMemoizedResults(It->second); 4138 ValueExprMap.erase(It); 4139 if (PHINode *PN = dyn_cast<PHINode>(I)) 4140 ConstantEvolutionLoopExitValue.erase(PN); 4141 } 4142 4143 PushDefUseChildren(I, Worklist); 4144 } 4145 4146 // Forget all contained loops too, to avoid dangling entries in the 4147 // ValuesAtScopes map. 4148 for (Loop::iterator I = L->begin(), E = L->end(); I != E; ++I) 4149 forgetLoop(*I); 4150} 4151 4152/// forgetValue - This method should be called by the client when it has 4153/// changed a value in a way that may effect its value, or which may 4154/// disconnect it from a def-use chain linking it to a loop. 4155void ScalarEvolution::forgetValue(Value *V) { 4156 Instruction *I = dyn_cast<Instruction>(V); 4157 if (!I) return; 4158 4159 // Drop information about expressions based on loop-header PHIs. 4160 SmallVector<Instruction *, 16> Worklist; 4161 Worklist.push_back(I); 4162 4163 SmallPtrSet<Instruction *, 8> Visited; 4164 while (!Worklist.empty()) { 4165 I = Worklist.pop_back_val(); 4166 if (!Visited.insert(I)) continue; 4167 4168 ValueExprMapType::iterator It = ValueExprMap.find(static_cast<Value *>(I)); 4169 if (It != ValueExprMap.end()) { 4170 forgetMemoizedResults(It->second); 4171 ValueExprMap.erase(It); 4172 if (PHINode *PN = dyn_cast<PHINode>(I)) 4173 ConstantEvolutionLoopExitValue.erase(PN); 4174 } 4175 4176 PushDefUseChildren(I, Worklist); 4177 } 4178} 4179 4180/// getExact - Get the exact loop backedge taken count considering all loop 4181/// exits. A computable result can only be return for loops with a single exit. 4182/// Returning the minimum taken count among all exits is incorrect because one 4183/// of the loop's exit limit's may have been skipped. HowFarToZero assumes that 4184/// the limit of each loop test is never skipped. This is a valid assumption as 4185/// long as the loop exits via that test. For precise results, it is the 4186/// caller's responsibility to specify the relevant loop exit using 4187/// getExact(ExitingBlock, SE). 4188const SCEV * 4189ScalarEvolution::BackedgeTakenInfo::getExact(ScalarEvolution *SE) const { 4190 // If any exits were not computable, the loop is not computable. 4191 if (!ExitNotTaken.isCompleteList()) return SE->getCouldNotCompute(); 4192 4193 // We need exactly one computable exit. 4194 if (!ExitNotTaken.ExitingBlock) return SE->getCouldNotCompute(); 4195 assert(ExitNotTaken.ExactNotTaken && "uninitialized not-taken info"); 4196 4197 const SCEV *BECount = 0; 4198 for (const ExitNotTakenInfo *ENT = &ExitNotTaken; 4199 ENT != 0; ENT = ENT->getNextExit()) { 4200 4201 assert(ENT->ExactNotTaken != SE->getCouldNotCompute() && "bad exit SCEV"); 4202 4203 if (!BECount) 4204 BECount = ENT->ExactNotTaken; 4205 else if (BECount != ENT->ExactNotTaken) 4206 return SE->getCouldNotCompute(); 4207 } 4208 assert(BECount && "Invalid not taken count for loop exit"); 4209 return BECount; 4210} 4211 4212/// getExact - Get the exact not taken count for this loop exit. 4213const SCEV * 4214ScalarEvolution::BackedgeTakenInfo::getExact(BasicBlock *ExitingBlock, 4215 ScalarEvolution *SE) const { 4216 for (const ExitNotTakenInfo *ENT = &ExitNotTaken; 4217 ENT != 0; ENT = ENT->getNextExit()) { 4218 4219 if (ENT->ExitingBlock == ExitingBlock) 4220 return ENT->ExactNotTaken; 4221 } 4222 return SE->getCouldNotCompute(); 4223} 4224 4225/// getMax - Get the max backedge taken count for the loop. 4226const SCEV * 4227ScalarEvolution::BackedgeTakenInfo::getMax(ScalarEvolution *SE) const { 4228 return Max ? Max : SE->getCouldNotCompute(); 4229} 4230 4231/// Allocate memory for BackedgeTakenInfo and copy the not-taken count of each 4232/// computable exit into a persistent ExitNotTakenInfo array. 4233ScalarEvolution::BackedgeTakenInfo::BackedgeTakenInfo( 4234 SmallVectorImpl< std::pair<BasicBlock *, const SCEV *> > &ExitCounts, 4235 bool Complete, const SCEV *MaxCount) : Max(MaxCount) { 4236 4237 if (!Complete) 4238 ExitNotTaken.setIncomplete(); 4239 4240 unsigned NumExits = ExitCounts.size(); 4241 if (NumExits == 0) return; 4242 4243 ExitNotTaken.ExitingBlock = ExitCounts[0].first; 4244 ExitNotTaken.ExactNotTaken = ExitCounts[0].second; 4245 if (NumExits == 1) return; 4246 4247 // Handle the rare case of multiple computable exits. 4248 ExitNotTakenInfo *ENT = new ExitNotTakenInfo[NumExits-1]; 4249 4250 ExitNotTakenInfo *PrevENT = &ExitNotTaken; 4251 for (unsigned i = 1; i < NumExits; ++i, PrevENT = ENT, ++ENT) { 4252 PrevENT->setNextExit(ENT); 4253 ENT->ExitingBlock = ExitCounts[i].first; 4254 ENT->ExactNotTaken = ExitCounts[i].second; 4255 } 4256} 4257 4258/// clear - Invalidate this result and free the ExitNotTakenInfo array. 4259void ScalarEvolution::BackedgeTakenInfo::clear() { 4260 ExitNotTaken.ExitingBlock = 0; 4261 ExitNotTaken.ExactNotTaken = 0; 4262 delete[] ExitNotTaken.getNextExit(); 4263} 4264 4265/// ComputeBackedgeTakenCount - Compute the number of times the backedge 4266/// of the specified loop will execute. 4267ScalarEvolution::BackedgeTakenInfo 4268ScalarEvolution::ComputeBackedgeTakenCount(const Loop *L) { 4269 SmallVector<BasicBlock *, 8> ExitingBlocks; 4270 L->getExitingBlocks(ExitingBlocks); 4271 4272 // Examine all exits and pick the most conservative values. 4273 const SCEV *MaxBECount = getCouldNotCompute(); 4274 bool CouldComputeBECount = true; 4275 SmallVector<std::pair<BasicBlock *, const SCEV *>, 4> ExitCounts; 4276 for (unsigned i = 0, e = ExitingBlocks.size(); i != e; ++i) { 4277 ExitLimit EL = ComputeExitLimit(L, ExitingBlocks[i]); 4278 if (EL.Exact == getCouldNotCompute()) 4279 // We couldn't compute an exact value for this exit, so 4280 // we won't be able to compute an exact value for the loop. 4281 CouldComputeBECount = false; 4282 else 4283 ExitCounts.push_back(std::make_pair(ExitingBlocks[i], EL.Exact)); 4284 4285 if (MaxBECount == getCouldNotCompute()) 4286 MaxBECount = EL.Max; 4287 else if (EL.Max != getCouldNotCompute()) { 4288 // We cannot take the "min" MaxBECount, because non-unit stride loops may 4289 // skip some loop tests. Taking the max over the exits is sufficiently 4290 // conservative. TODO: We could do better taking into consideration 4291 // that (1) the loop has unit stride (2) the last loop test is 4292 // less-than/greater-than (3) any loop test is less-than/greater-than AND 4293 // falls-through some constant times less then the other tests. 4294 MaxBECount = getUMaxFromMismatchedTypes(MaxBECount, EL.Max); 4295 } 4296 } 4297 4298 return BackedgeTakenInfo(ExitCounts, CouldComputeBECount, MaxBECount); 4299} 4300 4301/// ComputeExitLimit - Compute the number of times the backedge of the specified 4302/// loop will execute if it exits via the specified block. 4303ScalarEvolution::ExitLimit 4304ScalarEvolution::ComputeExitLimit(const Loop *L, BasicBlock *ExitingBlock) { 4305 4306 // Okay, we've chosen an exiting block. See what condition causes us to 4307 // exit at this block. 4308 // 4309 // FIXME: we should be able to handle switch instructions (with a single exit) 4310 BranchInst *ExitBr = dyn_cast<BranchInst>(ExitingBlock->getTerminator()); 4311 if (ExitBr == 0) return getCouldNotCompute(); 4312 assert(ExitBr->isConditional() && "If unconditional, it can't be in loop!"); 4313 4314 // At this point, we know we have a conditional branch that determines whether 4315 // the loop is exited. However, we don't know if the branch is executed each 4316 // time through the loop. If not, then the execution count of the branch will 4317 // not be equal to the trip count of the loop. 4318 // 4319 // Currently we check for this by checking to see if the Exit branch goes to 4320 // the loop header. If so, we know it will always execute the same number of 4321 // times as the loop. We also handle the case where the exit block *is* the 4322 // loop header. This is common for un-rotated loops. 4323 // 4324 // If both of those tests fail, walk up the unique predecessor chain to the 4325 // header, stopping if there is an edge that doesn't exit the loop. If the 4326 // header is reached, the execution count of the branch will be equal to the 4327 // trip count of the loop. 4328 // 4329 // More extensive analysis could be done to handle more cases here. 4330 // 4331 if (ExitBr->getSuccessor(0) != L->getHeader() && 4332 ExitBr->getSuccessor(1) != L->getHeader() && 4333 ExitBr->getParent() != L->getHeader()) { 4334 // The simple checks failed, try climbing the unique predecessor chain 4335 // up to the header. 4336 bool Ok = false; 4337 for (BasicBlock *BB = ExitBr->getParent(); BB; ) { 4338 BasicBlock *Pred = BB->getUniquePredecessor(); 4339 if (!Pred) 4340 return getCouldNotCompute(); 4341 TerminatorInst *PredTerm = Pred->getTerminator(); 4342 for (unsigned i = 0, e = PredTerm->getNumSuccessors(); i != e; ++i) { 4343 BasicBlock *PredSucc = PredTerm->getSuccessor(i); 4344 if (PredSucc == BB) 4345 continue; 4346 // If the predecessor has a successor that isn't BB and isn't 4347 // outside the loop, assume the worst. 4348 if (L->contains(PredSucc)) 4349 return getCouldNotCompute(); 4350 } 4351 if (Pred == L->getHeader()) { 4352 Ok = true; 4353 break; 4354 } 4355 BB = Pred; 4356 } 4357 if (!Ok) 4358 return getCouldNotCompute(); 4359 } 4360 4361 // Proceed to the next level to examine the exit condition expression. 4362 return ComputeExitLimitFromCond(L, ExitBr->getCondition(), 4363 ExitBr->getSuccessor(0), 4364 ExitBr->getSuccessor(1)); 4365} 4366 4367/// ComputeExitLimitFromCond - Compute the number of times the 4368/// backedge of the specified loop will execute if its exit condition 4369/// were a conditional branch of ExitCond, TBB, and FBB. 4370ScalarEvolution::ExitLimit 4371ScalarEvolution::ComputeExitLimitFromCond(const Loop *L, 4372 Value *ExitCond, 4373 BasicBlock *TBB, 4374 BasicBlock *FBB) { 4375 // Check if the controlling expression for this loop is an And or Or. 4376 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(ExitCond)) { 4377 if (BO->getOpcode() == Instruction::And) { 4378 // Recurse on the operands of the and. 4379 ExitLimit EL0 = ComputeExitLimitFromCond(L, BO->getOperand(0), TBB, FBB); 4380 ExitLimit EL1 = ComputeExitLimitFromCond(L, BO->getOperand(1), TBB, FBB); 4381 const SCEV *BECount = getCouldNotCompute(); 4382 const SCEV *MaxBECount = getCouldNotCompute(); 4383 if (L->contains(TBB)) { 4384 // Both conditions must be true for the loop to continue executing. 4385 // Choose the less conservative count. 4386 if (EL0.Exact == getCouldNotCompute() || 4387 EL1.Exact == getCouldNotCompute()) 4388 BECount = getCouldNotCompute(); 4389 else 4390 BECount = getUMinFromMismatchedTypes(EL0.Exact, EL1.Exact); 4391 if (EL0.Max == getCouldNotCompute()) 4392 MaxBECount = EL1.Max; 4393 else if (EL1.Max == getCouldNotCompute()) 4394 MaxBECount = EL0.Max; 4395 else 4396 MaxBECount = getUMinFromMismatchedTypes(EL0.Max, EL1.Max); 4397 } else { 4398 // Both conditions must be true at the same time for the loop to exit. 4399 // For now, be conservative. 4400 assert(L->contains(FBB) && "Loop block has no successor in loop!"); 4401 if (EL0.Max == EL1.Max) 4402 MaxBECount = EL0.Max; 4403 if (EL0.Exact == EL1.Exact) 4404 BECount = EL0.Exact; 4405 } 4406 4407 return ExitLimit(BECount, MaxBECount); 4408 } 4409 if (BO->getOpcode() == Instruction::Or) { 4410 // Recurse on the operands of the or. 4411 ExitLimit EL0 = ComputeExitLimitFromCond(L, BO->getOperand(0), TBB, FBB); 4412 ExitLimit EL1 = ComputeExitLimitFromCond(L, BO->getOperand(1), TBB, FBB); 4413 const SCEV *BECount = getCouldNotCompute(); 4414 const SCEV *MaxBECount = getCouldNotCompute(); 4415 if (L->contains(FBB)) { 4416 // Both conditions must be false for the loop to continue executing. 4417 // Choose the less conservative count. 4418 if (EL0.Exact == getCouldNotCompute() || 4419 EL1.Exact == getCouldNotCompute()) 4420 BECount = getCouldNotCompute(); 4421 else 4422 BECount = getUMinFromMismatchedTypes(EL0.Exact, EL1.Exact); 4423 if (EL0.Max == getCouldNotCompute()) 4424 MaxBECount = EL1.Max; 4425 else if (EL1.Max == getCouldNotCompute()) 4426 MaxBECount = EL0.Max; 4427 else 4428 MaxBECount = getUMinFromMismatchedTypes(EL0.Max, EL1.Max); 4429 } else { 4430 // Both conditions must be false at the same time for the loop to exit. 4431 // For now, be conservative. 4432 assert(L->contains(TBB) && "Loop block has no successor in loop!"); 4433 if (EL0.Max == EL1.Max) 4434 MaxBECount = EL0.Max; 4435 if (EL0.Exact == EL1.Exact) 4436 BECount = EL0.Exact; 4437 } 4438 4439 return ExitLimit(BECount, MaxBECount); 4440 } 4441 } 4442 4443 // With an icmp, it may be feasible to compute an exact backedge-taken count. 4444 // Proceed to the next level to examine the icmp. 4445 if (ICmpInst *ExitCondICmp = dyn_cast<ICmpInst>(ExitCond)) 4446 return ComputeExitLimitFromICmp(L, ExitCondICmp, TBB, FBB); 4447 4448 // Check for a constant condition. These are normally stripped out by 4449 // SimplifyCFG, but ScalarEvolution may be used by a pass which wishes to 4450 // preserve the CFG and is temporarily leaving constant conditions 4451 // in place. 4452 if (ConstantInt *CI = dyn_cast<ConstantInt>(ExitCond)) { 4453 if (L->contains(FBB) == !CI->getZExtValue()) 4454 // The backedge is always taken. 4455 return getCouldNotCompute(); 4456 else 4457 // The backedge is never taken. 4458 return getConstant(CI->getType(), 0); 4459 } 4460 4461 // If it's not an integer or pointer comparison then compute it the hard way. 4462 return ComputeExitCountExhaustively(L, ExitCond, !L->contains(TBB)); 4463} 4464 4465/// ComputeExitLimitFromICmp - Compute the number of times the 4466/// backedge of the specified loop will execute if its exit condition 4467/// were a conditional branch of the ICmpInst ExitCond, TBB, and FBB. 4468ScalarEvolution::ExitLimit 4469ScalarEvolution::ComputeExitLimitFromICmp(const Loop *L, 4470 ICmpInst *ExitCond, 4471 BasicBlock *TBB, 4472 BasicBlock *FBB) { 4473 4474 // If the condition was exit on true, convert the condition to exit on false 4475 ICmpInst::Predicate Cond; 4476 if (!L->contains(FBB)) 4477 Cond = ExitCond->getPredicate(); 4478 else 4479 Cond = ExitCond->getInversePredicate(); 4480 4481 // Handle common loops like: for (X = "string"; *X; ++X) 4482 if (LoadInst *LI = dyn_cast<LoadInst>(ExitCond->getOperand(0))) 4483 if (Constant *RHS = dyn_cast<Constant>(ExitCond->getOperand(1))) { 4484 ExitLimit ItCnt = 4485 ComputeLoadConstantCompareExitLimit(LI, RHS, L, Cond); 4486 if (ItCnt.hasAnyInfo()) 4487 return ItCnt; 4488 } 4489 4490 const SCEV *LHS = getSCEV(ExitCond->getOperand(0)); 4491 const SCEV *RHS = getSCEV(ExitCond->getOperand(1)); 4492 4493 // Try to evaluate any dependencies out of the loop. 4494 LHS = getSCEVAtScope(LHS, L); 4495 RHS = getSCEVAtScope(RHS, L); 4496 4497 // At this point, we would like to compute how many iterations of the 4498 // loop the predicate will return true for these inputs. 4499 if (isLoopInvariant(LHS, L) && !isLoopInvariant(RHS, L)) { 4500 // If there is a loop-invariant, force it into the RHS. 4501 std::swap(LHS, RHS); 4502 Cond = ICmpInst::getSwappedPredicate(Cond); 4503 } 4504 4505 // Simplify the operands before analyzing them. 4506 (void)SimplifyICmpOperands(Cond, LHS, RHS); 4507 4508 // If we have a comparison of a chrec against a constant, try to use value 4509 // ranges to answer this query. 4510 if (const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(RHS)) 4511 if (const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(LHS)) 4512 if (AddRec->getLoop() == L) { 4513 // Form the constant range. 4514 ConstantRange CompRange( 4515 ICmpInst::makeConstantRange(Cond, RHSC->getValue()->getValue())); 4516 4517 const SCEV *Ret = AddRec->getNumIterationsInRange(CompRange, *this); 4518 if (!isa<SCEVCouldNotCompute>(Ret)) return Ret; 4519 } 4520 4521 switch (Cond) { 4522 case ICmpInst::ICMP_NE: { // while (X != Y) 4523 // Convert to: while (X-Y != 0) 4524 ExitLimit EL = HowFarToZero(getMinusSCEV(LHS, RHS), L); 4525 if (EL.hasAnyInfo()) return EL; 4526 break; 4527 } 4528 case ICmpInst::ICMP_EQ: { // while (X == Y) 4529 // Convert to: while (X-Y == 0) 4530 ExitLimit EL = HowFarToNonZero(getMinusSCEV(LHS, RHS), L); 4531 if (EL.hasAnyInfo()) return EL; 4532 break; 4533 } 4534 case ICmpInst::ICMP_SLT: { 4535 ExitLimit EL = HowManyLessThans(LHS, RHS, L, true); 4536 if (EL.hasAnyInfo()) return EL; 4537 break; 4538 } 4539 case ICmpInst::ICMP_SGT: { 4540 ExitLimit EL = HowManyLessThans(getNotSCEV(LHS), 4541 getNotSCEV(RHS), L, true); 4542 if (EL.hasAnyInfo()) return EL; 4543 break; 4544 } 4545 case ICmpInst::ICMP_ULT: { 4546 ExitLimit EL = HowManyLessThans(LHS, RHS, L, false); 4547 if (EL.hasAnyInfo()) return EL; 4548 break; 4549 } 4550 case ICmpInst::ICMP_UGT: { 4551 ExitLimit EL = HowManyLessThans(getNotSCEV(LHS), 4552 getNotSCEV(RHS), L, false); 4553 if (EL.hasAnyInfo()) return EL; 4554 break; 4555 } 4556 default: 4557#if 0 4558 dbgs() << "ComputeBackedgeTakenCount "; 4559 if (ExitCond->getOperand(0)->getType()->isUnsigned()) 4560 dbgs() << "[unsigned] "; 4561 dbgs() << *LHS << " " 4562 << Instruction::getOpcodeName(Instruction::ICmp) 4563 << " " << *RHS << "\n"; 4564#endif 4565 break; 4566 } 4567 return ComputeExitCountExhaustively(L, ExitCond, !L->contains(TBB)); 4568} 4569 4570static ConstantInt * 4571EvaluateConstantChrecAtConstant(const SCEVAddRecExpr *AddRec, ConstantInt *C, 4572 ScalarEvolution &SE) { 4573 const SCEV *InVal = SE.getConstant(C); 4574 const SCEV *Val = AddRec->evaluateAtIteration(InVal, SE); 4575 assert(isa<SCEVConstant>(Val) && 4576 "Evaluation of SCEV at constant didn't fold correctly?"); 4577 return cast<SCEVConstant>(Val)->getValue(); 4578} 4579 4580/// ComputeLoadConstantCompareExitLimit - Given an exit condition of 4581/// 'icmp op load X, cst', try to see if we can compute the backedge 4582/// execution count. 4583ScalarEvolution::ExitLimit 4584ScalarEvolution::ComputeLoadConstantCompareExitLimit( 4585 LoadInst *LI, 4586 Constant *RHS, 4587 const Loop *L, 4588 ICmpInst::Predicate predicate) { 4589 4590 if (LI->isVolatile()) return getCouldNotCompute(); 4591 4592 // Check to see if the loaded pointer is a getelementptr of a global. 4593 // TODO: Use SCEV instead of manually grubbing with GEPs. 4594 GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(LI->getOperand(0)); 4595 if (!GEP) return getCouldNotCompute(); 4596 4597 // Make sure that it is really a constant global we are gepping, with an 4598 // initializer, and make sure the first IDX is really 0. 4599 GlobalVariable *GV = dyn_cast<GlobalVariable>(GEP->getOperand(0)); 4600 if (!GV || !GV->isConstant() || !GV->hasDefinitiveInitializer() || 4601 GEP->getNumOperands() < 3 || !isa<Constant>(GEP->getOperand(1)) || 4602 !cast<Constant>(GEP->getOperand(1))->isNullValue()) 4603 return getCouldNotCompute(); 4604 4605 // Okay, we allow one non-constant index into the GEP instruction. 4606 Value *VarIdx = 0; 4607 std::vector<Constant*> Indexes; 4608 unsigned VarIdxNum = 0; 4609 for (unsigned i = 2, e = GEP->getNumOperands(); i != e; ++i) 4610 if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i))) { 4611 Indexes.push_back(CI); 4612 } else if (!isa<ConstantInt>(GEP->getOperand(i))) { 4613 if (VarIdx) return getCouldNotCompute(); // Multiple non-constant idx's. 4614 VarIdx = GEP->getOperand(i); 4615 VarIdxNum = i-2; 4616 Indexes.push_back(0); 4617 } 4618 4619 // Loop-invariant loads may be a byproduct of loop optimization. Skip them. 4620 if (!VarIdx) 4621 return getCouldNotCompute(); 4622 4623 // Okay, we know we have a (load (gep GV, 0, X)) comparison with a constant. 4624 // Check to see if X is a loop variant variable value now. 4625 const SCEV *Idx = getSCEV(VarIdx); 4626 Idx = getSCEVAtScope(Idx, L); 4627 4628 // We can only recognize very limited forms of loop index expressions, in 4629 // particular, only affine AddRec's like {C1,+,C2}. 4630 const SCEVAddRecExpr *IdxExpr = dyn_cast<SCEVAddRecExpr>(Idx); 4631 if (!IdxExpr || !IdxExpr->isAffine() || isLoopInvariant(IdxExpr, L) || 4632 !isa<SCEVConstant>(IdxExpr->getOperand(0)) || 4633 !isa<SCEVConstant>(IdxExpr->getOperand(1))) 4634 return getCouldNotCompute(); 4635 4636 unsigned MaxSteps = MaxBruteForceIterations; 4637 for (unsigned IterationNum = 0; IterationNum != MaxSteps; ++IterationNum) { 4638 ConstantInt *ItCst = ConstantInt::get( 4639 cast<IntegerType>(IdxExpr->getType()), IterationNum); 4640 ConstantInt *Val = EvaluateConstantChrecAtConstant(IdxExpr, ItCst, *this); 4641 4642 // Form the GEP offset. 4643 Indexes[VarIdxNum] = Val; 4644 4645 Constant *Result = ConstantFoldLoadThroughGEPIndices(GV->getInitializer(), 4646 Indexes); 4647 if (Result == 0) break; // Cannot compute! 4648 4649 // Evaluate the condition for this iteration. 4650 Result = ConstantExpr::getICmp(predicate, Result, RHS); 4651 if (!isa<ConstantInt>(Result)) break; // Couldn't decide for sure 4652 if (cast<ConstantInt>(Result)->getValue().isMinValue()) { 4653#if 0 4654 dbgs() << "\n***\n*** Computed loop count " << *ItCst 4655 << "\n*** From global " << *GV << "*** BB: " << *L->getHeader() 4656 << "***\n"; 4657#endif 4658 ++NumArrayLenItCounts; 4659 return getConstant(ItCst); // Found terminating iteration! 4660 } 4661 } 4662 return getCouldNotCompute(); 4663} 4664 4665 4666/// CanConstantFold - Return true if we can constant fold an instruction of the 4667/// specified type, assuming that all operands were constants. 4668static bool CanConstantFold(const Instruction *I) { 4669 if (isa<BinaryOperator>(I) || isa<CmpInst>(I) || 4670 isa<SelectInst>(I) || isa<CastInst>(I) || isa<GetElementPtrInst>(I) || 4671 isa<LoadInst>(I)) 4672 return true; 4673 4674 if (const CallInst *CI = dyn_cast<CallInst>(I)) 4675 if (const Function *F = CI->getCalledFunction()) 4676 return canConstantFoldCallTo(F); 4677 return false; 4678} 4679 4680/// Determine whether this instruction can constant evolve within this loop 4681/// assuming its operands can all constant evolve. 4682static bool canConstantEvolve(Instruction *I, const Loop *L) { 4683 // An instruction outside of the loop can't be derived from a loop PHI. 4684 if (!L->contains(I)) return false; 4685 4686 if (isa<PHINode>(I)) { 4687 if (L->getHeader() == I->getParent()) 4688 return true; 4689 else 4690 // We don't currently keep track of the control flow needed to evaluate 4691 // PHIs, so we cannot handle PHIs inside of loops. 4692 return false; 4693 } 4694 4695 // If we won't be able to constant fold this expression even if the operands 4696 // are constants, bail early. 4697 return CanConstantFold(I); 4698} 4699 4700/// getConstantEvolvingPHIOperands - Implement getConstantEvolvingPHI by 4701/// recursing through each instruction operand until reaching a loop header phi. 4702static PHINode * 4703getConstantEvolvingPHIOperands(Instruction *UseInst, const Loop *L, 4704 DenseMap<Instruction *, PHINode *> &PHIMap) { 4705 4706 // Otherwise, we can evaluate this instruction if all of its operands are 4707 // constant or derived from a PHI node themselves. 4708 PHINode *PHI = 0; 4709 for (Instruction::op_iterator OpI = UseInst->op_begin(), 4710 OpE = UseInst->op_end(); OpI != OpE; ++OpI) { 4711 4712 if (isa<Constant>(*OpI)) continue; 4713 4714 Instruction *OpInst = dyn_cast<Instruction>(*OpI); 4715 if (!OpInst || !canConstantEvolve(OpInst, L)) return 0; 4716 4717 PHINode *P = dyn_cast<PHINode>(OpInst); 4718 if (!P) 4719 // If this operand is already visited, reuse the prior result. 4720 // We may have P != PHI if this is the deepest point at which the 4721 // inconsistent paths meet. 4722 P = PHIMap.lookup(OpInst); 4723 if (!P) { 4724 // Recurse and memoize the results, whether a phi is found or not. 4725 // This recursive call invalidates pointers into PHIMap. 4726 P = getConstantEvolvingPHIOperands(OpInst, L, PHIMap); 4727 PHIMap[OpInst] = P; 4728 } 4729 if (P == 0) return 0; // Not evolving from PHI 4730 if (PHI && PHI != P) return 0; // Evolving from multiple different PHIs. 4731 PHI = P; 4732 } 4733 // This is a expression evolving from a constant PHI! 4734 return PHI; 4735} 4736 4737/// getConstantEvolvingPHI - Given an LLVM value and a loop, return a PHI node 4738/// in the loop that V is derived from. We allow arbitrary operations along the 4739/// way, but the operands of an operation must either be constants or a value 4740/// derived from a constant PHI. If this expression does not fit with these 4741/// constraints, return null. 4742static PHINode *getConstantEvolvingPHI(Value *V, const Loop *L) { 4743 Instruction *I = dyn_cast<Instruction>(V); 4744 if (I == 0 || !canConstantEvolve(I, L)) return 0; 4745 4746 if (PHINode *PN = dyn_cast<PHINode>(I)) { 4747 return PN; 4748 } 4749 4750 // Record non-constant instructions contained by the loop. 4751 DenseMap<Instruction *, PHINode *> PHIMap; 4752 return getConstantEvolvingPHIOperands(I, L, PHIMap); 4753} 4754 4755/// EvaluateExpression - Given an expression that passes the 4756/// getConstantEvolvingPHI predicate, evaluate its value assuming the PHI node 4757/// in the loop has the value PHIVal. If we can't fold this expression for some 4758/// reason, return null. 4759static Constant *EvaluateExpression(Value *V, const Loop *L, 4760 DenseMap<Instruction *, Constant *> &Vals, 4761 const TargetData *TD, 4762 const TargetLibraryInfo *TLI) { 4763 // Convenient constant check, but redundant for recursive calls. 4764 if (Constant *C = dyn_cast<Constant>(V)) return C; 4765 Instruction *I = dyn_cast<Instruction>(V); 4766 if (!I) return 0; 4767 4768 if (Constant *C = Vals.lookup(I)) return C; 4769 4770 // An instruction inside the loop depends on a value outside the loop that we 4771 // weren't given a mapping for, or a value such as a call inside the loop. 4772 if (!canConstantEvolve(I, L)) return 0; 4773 4774 // An unmapped PHI can be due to a branch or another loop inside this loop, 4775 // or due to this not being the initial iteration through a loop where we 4776 // couldn't compute the evolution of this particular PHI last time. 4777 if (isa<PHINode>(I)) return 0; 4778 4779 std::vector<Constant*> Operands(I->getNumOperands()); 4780 4781 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) { 4782 Instruction *Operand = dyn_cast<Instruction>(I->getOperand(i)); 4783 if (!Operand) { 4784 Operands[i] = dyn_cast<Constant>(I->getOperand(i)); 4785 if (!Operands[i]) return 0; 4786 continue; 4787 } 4788 Constant *C = EvaluateExpression(Operand, L, Vals, TD, TLI); 4789 Vals[Operand] = C; 4790 if (!C) return 0; 4791 Operands[i] = C; 4792 } 4793 4794 if (CmpInst *CI = dyn_cast<CmpInst>(I)) 4795 return ConstantFoldCompareInstOperands(CI->getPredicate(), Operands[0], 4796 Operands[1], TD, TLI); 4797 if (LoadInst *LI = dyn_cast<LoadInst>(I)) { 4798 if (!LI->isVolatile()) 4799 return ConstantFoldLoadFromConstPtr(Operands[0], TD); 4800 } 4801 return ConstantFoldInstOperands(I->getOpcode(), I->getType(), Operands, TD, 4802 TLI); 4803} 4804 4805/// getConstantEvolutionLoopExitValue - If we know that the specified Phi is 4806/// in the header of its containing loop, we know the loop executes a 4807/// constant number of times, and the PHI node is just a recurrence 4808/// involving constants, fold it. 4809Constant * 4810ScalarEvolution::getConstantEvolutionLoopExitValue(PHINode *PN, 4811 const APInt &BEs, 4812 const Loop *L) { 4813 DenseMap<PHINode*, Constant*>::const_iterator I = 4814 ConstantEvolutionLoopExitValue.find(PN); 4815 if (I != ConstantEvolutionLoopExitValue.end()) 4816 return I->second; 4817 4818 if (BEs.ugt(MaxBruteForceIterations)) 4819 return ConstantEvolutionLoopExitValue[PN] = 0; // Not going to evaluate it. 4820 4821 Constant *&RetVal = ConstantEvolutionLoopExitValue[PN]; 4822 4823 DenseMap<Instruction *, Constant *> CurrentIterVals; 4824 BasicBlock *Header = L->getHeader(); 4825 assert(PN->getParent() == Header && "Can't evaluate PHI not in loop header!"); 4826 4827 // Since the loop is canonicalized, the PHI node must have two entries. One 4828 // entry must be a constant (coming in from outside of the loop), and the 4829 // second must be derived from the same PHI. 4830 bool SecondIsBackedge = L->contains(PN->getIncomingBlock(1)); 4831 PHINode *PHI = 0; 4832 for (BasicBlock::iterator I = Header->begin(); 4833 (PHI = dyn_cast<PHINode>(I)); ++I) { 4834 Constant *StartCST = 4835 dyn_cast<Constant>(PHI->getIncomingValue(!SecondIsBackedge)); 4836 if (StartCST == 0) continue; 4837 CurrentIterVals[PHI] = StartCST; 4838 } 4839 if (!CurrentIterVals.count(PN)) 4840 return RetVal = 0; 4841 4842 Value *BEValue = PN->getIncomingValue(SecondIsBackedge); 4843 4844 // Execute the loop symbolically to determine the exit value. 4845 if (BEs.getActiveBits() >= 32) 4846 return RetVal = 0; // More than 2^32-1 iterations?? Not doing it! 4847 4848 unsigned NumIterations = BEs.getZExtValue(); // must be in range 4849 unsigned IterationNum = 0; 4850 for (; ; ++IterationNum) { 4851 if (IterationNum == NumIterations) 4852 return RetVal = CurrentIterVals[PN]; // Got exit value! 4853 4854 // Compute the value of the PHIs for the next iteration. 4855 // EvaluateExpression adds non-phi values to the CurrentIterVals map. 4856 DenseMap<Instruction *, Constant *> NextIterVals; 4857 Constant *NextPHI = EvaluateExpression(BEValue, L, CurrentIterVals, TD, 4858 TLI); 4859 if (NextPHI == 0) 4860 return 0; // Couldn't evaluate! 4861 NextIterVals[PN] = NextPHI; 4862 4863 bool StoppedEvolving = NextPHI == CurrentIterVals[PN]; 4864 4865 // Also evaluate the other PHI nodes. However, we don't get to stop if we 4866 // cease to be able to evaluate one of them or if they stop evolving, 4867 // because that doesn't necessarily prevent us from computing PN. 4868 SmallVector<std::pair<PHINode *, Constant *>, 8> PHIsToCompute; 4869 for (DenseMap<Instruction *, Constant *>::const_iterator 4870 I = CurrentIterVals.begin(), E = CurrentIterVals.end(); I != E; ++I){ 4871 PHINode *PHI = dyn_cast<PHINode>(I->first); 4872 if (!PHI || PHI == PN || PHI->getParent() != Header) continue; 4873 PHIsToCompute.push_back(std::make_pair(PHI, I->second)); 4874 } 4875 // We use two distinct loops because EvaluateExpression may invalidate any 4876 // iterators into CurrentIterVals. 4877 for (SmallVectorImpl<std::pair<PHINode *, Constant*> >::const_iterator 4878 I = PHIsToCompute.begin(), E = PHIsToCompute.end(); I != E; ++I) { 4879 PHINode *PHI = I->first; 4880 Constant *&NextPHI = NextIterVals[PHI]; 4881 if (!NextPHI) { // Not already computed. 4882 Value *BEValue = PHI->getIncomingValue(SecondIsBackedge); 4883 NextPHI = EvaluateExpression(BEValue, L, CurrentIterVals, TD, TLI); 4884 } 4885 if (NextPHI != I->second) 4886 StoppedEvolving = false; 4887 } 4888 4889 // If all entries in CurrentIterVals == NextIterVals then we can stop 4890 // iterating, the loop can't continue to change. 4891 if (StoppedEvolving) 4892 return RetVal = CurrentIterVals[PN]; 4893 4894 CurrentIterVals.swap(NextIterVals); 4895 } 4896} 4897 4898/// ComputeExitCountExhaustively - If the loop is known to execute a 4899/// constant number of times (the condition evolves only from constants), 4900/// try to evaluate a few iterations of the loop until we get the exit 4901/// condition gets a value of ExitWhen (true or false). If we cannot 4902/// evaluate the trip count of the loop, return getCouldNotCompute(). 4903const SCEV *ScalarEvolution::ComputeExitCountExhaustively(const Loop *L, 4904 Value *Cond, 4905 bool ExitWhen) { 4906 PHINode *PN = getConstantEvolvingPHI(Cond, L); 4907 if (PN == 0) return getCouldNotCompute(); 4908 4909 // If the loop is canonicalized, the PHI will have exactly two entries. 4910 // That's the only form we support here. 4911 if (PN->getNumIncomingValues() != 2) return getCouldNotCompute(); 4912 4913 DenseMap<Instruction *, Constant *> CurrentIterVals; 4914 BasicBlock *Header = L->getHeader(); 4915 assert(PN->getParent() == Header && "Can't evaluate PHI not in loop header!"); 4916 4917 // One entry must be a constant (coming in from outside of the loop), and the 4918 // second must be derived from the same PHI. 4919 bool SecondIsBackedge = L->contains(PN->getIncomingBlock(1)); 4920 PHINode *PHI = 0; 4921 for (BasicBlock::iterator I = Header->begin(); 4922 (PHI = dyn_cast<PHINode>(I)); ++I) { 4923 Constant *StartCST = 4924 dyn_cast<Constant>(PHI->getIncomingValue(!SecondIsBackedge)); 4925 if (StartCST == 0) continue; 4926 CurrentIterVals[PHI] = StartCST; 4927 } 4928 if (!CurrentIterVals.count(PN)) 4929 return getCouldNotCompute(); 4930 4931 // Okay, we find a PHI node that defines the trip count of this loop. Execute 4932 // the loop symbolically to determine when the condition gets a value of 4933 // "ExitWhen". 4934 4935 unsigned MaxIterations = MaxBruteForceIterations; // Limit analysis. 4936 for (unsigned IterationNum = 0; IterationNum != MaxIterations;++IterationNum){ 4937 ConstantInt *CondVal = 4938 dyn_cast_or_null<ConstantInt>(EvaluateExpression(Cond, L, CurrentIterVals, 4939 TD, TLI)); 4940 4941 // Couldn't symbolically evaluate. 4942 if (!CondVal) return getCouldNotCompute(); 4943 4944 if (CondVal->getValue() == uint64_t(ExitWhen)) { 4945 ++NumBruteForceTripCountsComputed; 4946 return getConstant(Type::getInt32Ty(getContext()), IterationNum); 4947 } 4948 4949 // Update all the PHI nodes for the next iteration. 4950 DenseMap<Instruction *, Constant *> NextIterVals; 4951 4952 // Create a list of which PHIs we need to compute. We want to do this before 4953 // calling EvaluateExpression on them because that may invalidate iterators 4954 // into CurrentIterVals. 4955 SmallVector<PHINode *, 8> PHIsToCompute; 4956 for (DenseMap<Instruction *, Constant *>::const_iterator 4957 I = CurrentIterVals.begin(), E = CurrentIterVals.end(); I != E; ++I){ 4958 PHINode *PHI = dyn_cast<PHINode>(I->first); 4959 if (!PHI || PHI->getParent() != Header) continue; 4960 PHIsToCompute.push_back(PHI); 4961 } 4962 for (SmallVectorImpl<PHINode *>::const_iterator I = PHIsToCompute.begin(), 4963 E = PHIsToCompute.end(); I != E; ++I) { 4964 PHINode *PHI = *I; 4965 Constant *&NextPHI = NextIterVals[PHI]; 4966 if (NextPHI) continue; // Already computed! 4967 4968 Value *BEValue = PHI->getIncomingValue(SecondIsBackedge); 4969 NextPHI = EvaluateExpression(BEValue, L, CurrentIterVals, TD, TLI); 4970 } 4971 CurrentIterVals.swap(NextIterVals); 4972 } 4973 4974 // Too many iterations were needed to evaluate. 4975 return getCouldNotCompute(); 4976} 4977 4978/// getSCEVAtScope - Return a SCEV expression for the specified value 4979/// at the specified scope in the program. The L value specifies a loop 4980/// nest to evaluate the expression at, where null is the top-level or a 4981/// specified loop is immediately inside of the loop. 4982/// 4983/// This method can be used to compute the exit value for a variable defined 4984/// in a loop by querying what the value will hold in the parent loop. 4985/// 4986/// In the case that a relevant loop exit value cannot be computed, the 4987/// original value V is returned. 4988const SCEV *ScalarEvolution::getSCEVAtScope(const SCEV *V, const Loop *L) { 4989 // Check to see if we've folded this expression at this loop before. 4990 std::map<const Loop *, const SCEV *> &Values = ValuesAtScopes[V]; 4991 std::pair<std::map<const Loop *, const SCEV *>::iterator, bool> Pair = 4992 Values.insert(std::make_pair(L, static_cast<const SCEV *>(0))); 4993 if (!Pair.second) 4994 return Pair.first->second ? Pair.first->second : V; 4995 4996 // Otherwise compute it. 4997 const SCEV *C = computeSCEVAtScope(V, L); 4998 ValuesAtScopes[V][L] = C; 4999 return C; 5000} 5001 5002/// This builds up a Constant using the ConstantExpr interface. That way, we 5003/// will return Constants for objects which aren't represented by a 5004/// SCEVConstant, because SCEVConstant is restricted to ConstantInt. 5005/// Returns NULL if the SCEV isn't representable as a Constant. 5006static Constant *BuildConstantFromSCEV(const SCEV *V) { 5007 switch (V->getSCEVType()) { 5008 default: // TODO: smax, umax. 5009 case scCouldNotCompute: 5010 case scAddRecExpr: 5011 break; 5012 case scConstant: 5013 return cast<SCEVConstant>(V)->getValue(); 5014 case scUnknown: 5015 return dyn_cast<Constant>(cast<SCEVUnknown>(V)->getValue()); 5016 case scSignExtend: { 5017 const SCEVSignExtendExpr *SS = cast<SCEVSignExtendExpr>(V); 5018 if (Constant *CastOp = BuildConstantFromSCEV(SS->getOperand())) 5019 return ConstantExpr::getSExt(CastOp, SS->getType()); 5020 break; 5021 } 5022 case scZeroExtend: { 5023 const SCEVZeroExtendExpr *SZ = cast<SCEVZeroExtendExpr>(V); 5024 if (Constant *CastOp = BuildConstantFromSCEV(SZ->getOperand())) 5025 return ConstantExpr::getZExt(CastOp, SZ->getType()); 5026 break; 5027 } 5028 case scTruncate: { 5029 const SCEVTruncateExpr *ST = cast<SCEVTruncateExpr>(V); 5030 if (Constant *CastOp = BuildConstantFromSCEV(ST->getOperand())) 5031 return ConstantExpr::getTrunc(CastOp, ST->getType()); 5032 break; 5033 } 5034 case scAddExpr: { 5035 const SCEVAddExpr *SA = cast<SCEVAddExpr>(V); 5036 if (Constant *C = BuildConstantFromSCEV(SA->getOperand(0))) { 5037 if (C->getType()->isPointerTy()) 5038 C = ConstantExpr::getBitCast(C, Type::getInt8PtrTy(C->getContext())); 5039 for (unsigned i = 1, e = SA->getNumOperands(); i != e; ++i) { 5040 Constant *C2 = BuildConstantFromSCEV(SA->getOperand(i)); 5041 if (!C2) return 0; 5042 5043 // First pointer! 5044 if (!C->getType()->isPointerTy() && C2->getType()->isPointerTy()) { 5045 std::swap(C, C2); 5046 // The offsets have been converted to bytes. We can add bytes to an 5047 // i8* by GEP with the byte count in the first index. 5048 C = ConstantExpr::getBitCast(C,Type::getInt8PtrTy(C->getContext())); 5049 } 5050 5051 // Don't bother trying to sum two pointers. We probably can't 5052 // statically compute a load that results from it anyway. 5053 if (C2->getType()->isPointerTy()) 5054 return 0; 5055 5056 if (C->getType()->isPointerTy()) { 5057 if (cast<PointerType>(C->getType())->getElementType()->isStructTy()) 5058 C2 = ConstantExpr::getIntegerCast( 5059 C2, Type::getInt32Ty(C->getContext()), true); 5060 C = ConstantExpr::getGetElementPtr(C, C2); 5061 } else 5062 C = ConstantExpr::getAdd(C, C2); 5063 } 5064 return C; 5065 } 5066 break; 5067 } 5068 case scMulExpr: { 5069 const SCEVMulExpr *SM = cast<SCEVMulExpr>(V); 5070 if (Constant *C = BuildConstantFromSCEV(SM->getOperand(0))) { 5071 // Don't bother with pointers at all. 5072 if (C->getType()->isPointerTy()) return 0; 5073 for (unsigned i = 1, e = SM->getNumOperands(); i != e; ++i) { 5074 Constant *C2 = BuildConstantFromSCEV(SM->getOperand(i)); 5075 if (!C2 || C2->getType()->isPointerTy()) return 0; 5076 C = ConstantExpr::getMul(C, C2); 5077 } 5078 return C; 5079 } 5080 break; 5081 } 5082 case scUDivExpr: { 5083 const SCEVUDivExpr *SU = cast<SCEVUDivExpr>(V); 5084 if (Constant *LHS = BuildConstantFromSCEV(SU->getLHS())) 5085 if (Constant *RHS = BuildConstantFromSCEV(SU->getRHS())) 5086 if (LHS->getType() == RHS->getType()) 5087 return ConstantExpr::getUDiv(LHS, RHS); 5088 break; 5089 } 5090 } 5091 return 0; 5092} 5093 5094const SCEV *ScalarEvolution::computeSCEVAtScope(const SCEV *V, const Loop *L) { 5095 if (isa<SCEVConstant>(V)) return V; 5096 5097 // If this instruction is evolved from a constant-evolving PHI, compute the 5098 // exit value from the loop without using SCEVs. 5099 if (const SCEVUnknown *SU = dyn_cast<SCEVUnknown>(V)) { 5100 if (Instruction *I = dyn_cast<Instruction>(SU->getValue())) { 5101 const Loop *LI = (*this->LI)[I->getParent()]; 5102 if (LI && LI->getParentLoop() == L) // Looking for loop exit value. 5103 if (PHINode *PN = dyn_cast<PHINode>(I)) 5104 if (PN->getParent() == LI->getHeader()) { 5105 // Okay, there is no closed form solution for the PHI node. Check 5106 // to see if the loop that contains it has a known backedge-taken 5107 // count. If so, we may be able to force computation of the exit 5108 // value. 5109 const SCEV *BackedgeTakenCount = getBackedgeTakenCount(LI); 5110 if (const SCEVConstant *BTCC = 5111 dyn_cast<SCEVConstant>(BackedgeTakenCount)) { 5112 // Okay, we know how many times the containing loop executes. If 5113 // this is a constant evolving PHI node, get the final value at 5114 // the specified iteration number. 5115 Constant *RV = getConstantEvolutionLoopExitValue(PN, 5116 BTCC->getValue()->getValue(), 5117 LI); 5118 if (RV) return getSCEV(RV); 5119 } 5120 } 5121 5122 // Okay, this is an expression that we cannot symbolically evaluate 5123 // into a SCEV. Check to see if it's possible to symbolically evaluate 5124 // the arguments into constants, and if so, try to constant propagate the 5125 // result. This is particularly useful for computing loop exit values. 5126 if (CanConstantFold(I)) { 5127 SmallVector<Constant *, 4> Operands; 5128 bool MadeImprovement = false; 5129 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) { 5130 Value *Op = I->getOperand(i); 5131 if (Constant *C = dyn_cast<Constant>(Op)) { 5132 Operands.push_back(C); 5133 continue; 5134 } 5135 5136 // If any of the operands is non-constant and if they are 5137 // non-integer and non-pointer, don't even try to analyze them 5138 // with scev techniques. 5139 if (!isSCEVable(Op->getType())) 5140 return V; 5141 5142 const SCEV *OrigV = getSCEV(Op); 5143 const SCEV *OpV = getSCEVAtScope(OrigV, L); 5144 MadeImprovement |= OrigV != OpV; 5145 5146 Constant *C = BuildConstantFromSCEV(OpV); 5147 if (!C) return V; 5148 if (C->getType() != Op->getType()) 5149 C = ConstantExpr::getCast(CastInst::getCastOpcode(C, false, 5150 Op->getType(), 5151 false), 5152 C, Op->getType()); 5153 Operands.push_back(C); 5154 } 5155 5156 // Check to see if getSCEVAtScope actually made an improvement. 5157 if (MadeImprovement) { 5158 Constant *C = 0; 5159 if (const CmpInst *CI = dyn_cast<CmpInst>(I)) 5160 C = ConstantFoldCompareInstOperands(CI->getPredicate(), 5161 Operands[0], Operands[1], TD, 5162 TLI); 5163 else if (const LoadInst *LI = dyn_cast<LoadInst>(I)) { 5164 if (!LI->isVolatile()) 5165 C = ConstantFoldLoadFromConstPtr(Operands[0], TD); 5166 } else 5167 C = ConstantFoldInstOperands(I->getOpcode(), I->getType(), 5168 Operands, TD, TLI); 5169 if (!C) return V; 5170 return getSCEV(C); 5171 } 5172 } 5173 } 5174 5175 // This is some other type of SCEVUnknown, just return it. 5176 return V; 5177 } 5178 5179 if (const SCEVCommutativeExpr *Comm = dyn_cast<SCEVCommutativeExpr>(V)) { 5180 // Avoid performing the look-up in the common case where the specified 5181 // expression has no loop-variant portions. 5182 for (unsigned i = 0, e = Comm->getNumOperands(); i != e; ++i) { 5183 const SCEV *OpAtScope = getSCEVAtScope(Comm->getOperand(i), L); 5184 if (OpAtScope != Comm->getOperand(i)) { 5185 // Okay, at least one of these operands is loop variant but might be 5186 // foldable. Build a new instance of the folded commutative expression. 5187 SmallVector<const SCEV *, 8> NewOps(Comm->op_begin(), 5188 Comm->op_begin()+i); 5189 NewOps.push_back(OpAtScope); 5190 5191 for (++i; i != e; ++i) { 5192 OpAtScope = getSCEVAtScope(Comm->getOperand(i), L); 5193 NewOps.push_back(OpAtScope); 5194 } 5195 if (isa<SCEVAddExpr>(Comm)) 5196 return getAddExpr(NewOps); 5197 if (isa<SCEVMulExpr>(Comm)) 5198 return getMulExpr(NewOps); 5199 if (isa<SCEVSMaxExpr>(Comm)) 5200 return getSMaxExpr(NewOps); 5201 if (isa<SCEVUMaxExpr>(Comm)) 5202 return getUMaxExpr(NewOps); 5203 llvm_unreachable("Unknown commutative SCEV type!"); 5204 } 5205 } 5206 // If we got here, all operands are loop invariant. 5207 return Comm; 5208 } 5209 5210 if (const SCEVUDivExpr *Div = dyn_cast<SCEVUDivExpr>(V)) { 5211 const SCEV *LHS = getSCEVAtScope(Div->getLHS(), L); 5212 const SCEV *RHS = getSCEVAtScope(Div->getRHS(), L); 5213 if (LHS == Div->getLHS() && RHS == Div->getRHS()) 5214 return Div; // must be loop invariant 5215 return getUDivExpr(LHS, RHS); 5216 } 5217 5218 // If this is a loop recurrence for a loop that does not contain L, then we 5219 // are dealing with the final value computed by the loop. 5220 if (const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(V)) { 5221 // First, attempt to evaluate each operand. 5222 // Avoid performing the look-up in the common case where the specified 5223 // expression has no loop-variant portions. 5224 for (unsigned i = 0, e = AddRec->getNumOperands(); i != e; ++i) { 5225 const SCEV *OpAtScope = getSCEVAtScope(AddRec->getOperand(i), L); 5226 if (OpAtScope == AddRec->getOperand(i)) 5227 continue; 5228 5229 // Okay, at least one of these operands is loop variant but might be 5230 // foldable. Build a new instance of the folded commutative expression. 5231 SmallVector<const SCEV *, 8> NewOps(AddRec->op_begin(), 5232 AddRec->op_begin()+i); 5233 NewOps.push_back(OpAtScope); 5234 for (++i; i != e; ++i) 5235 NewOps.push_back(getSCEVAtScope(AddRec->getOperand(i), L)); 5236 5237 const SCEV *FoldedRec = 5238 getAddRecExpr(NewOps, AddRec->getLoop(), 5239 AddRec->getNoWrapFlags(SCEV::FlagNW)); 5240 AddRec = dyn_cast<SCEVAddRecExpr>(FoldedRec); 5241 // The addrec may be folded to a nonrecurrence, for example, if the 5242 // induction variable is multiplied by zero after constant folding. Go 5243 // ahead and return the folded value. 5244 if (!AddRec) 5245 return FoldedRec; 5246 break; 5247 } 5248 5249 // If the scope is outside the addrec's loop, evaluate it by using the 5250 // loop exit value of the addrec. 5251 if (!AddRec->getLoop()->contains(L)) { 5252 // To evaluate this recurrence, we need to know how many times the AddRec 5253 // loop iterates. Compute this now. 5254 const SCEV *BackedgeTakenCount = getBackedgeTakenCount(AddRec->getLoop()); 5255 if (BackedgeTakenCount == getCouldNotCompute()) return AddRec; 5256 5257 // Then, evaluate the AddRec. 5258 return AddRec->evaluateAtIteration(BackedgeTakenCount, *this); 5259 } 5260 5261 return AddRec; 5262 } 5263 5264 if (const SCEVZeroExtendExpr *Cast = dyn_cast<SCEVZeroExtendExpr>(V)) { 5265 const SCEV *Op = getSCEVAtScope(Cast->getOperand(), L); 5266 if (Op == Cast->getOperand()) 5267 return Cast; // must be loop invariant 5268 return getZeroExtendExpr(Op, Cast->getType()); 5269 } 5270 5271 if (const SCEVSignExtendExpr *Cast = dyn_cast<SCEVSignExtendExpr>(V)) { 5272 const SCEV *Op = getSCEVAtScope(Cast->getOperand(), L); 5273 if (Op == Cast->getOperand()) 5274 return Cast; // must be loop invariant 5275 return getSignExtendExpr(Op, Cast->getType()); 5276 } 5277 5278 if (const SCEVTruncateExpr *Cast = dyn_cast<SCEVTruncateExpr>(V)) { 5279 const SCEV *Op = getSCEVAtScope(Cast->getOperand(), L); 5280 if (Op == Cast->getOperand()) 5281 return Cast; // must be loop invariant 5282 return getTruncateExpr(Op, Cast->getType()); 5283 } 5284 5285 llvm_unreachable("Unknown SCEV type!"); 5286} 5287 5288/// getSCEVAtScope - This is a convenience function which does 5289/// getSCEVAtScope(getSCEV(V), L). 5290const SCEV *ScalarEvolution::getSCEVAtScope(Value *V, const Loop *L) { 5291 return getSCEVAtScope(getSCEV(V), L); 5292} 5293 5294/// SolveLinEquationWithOverflow - Finds the minimum unsigned root of the 5295/// following equation: 5296/// 5297/// A * X = B (mod N) 5298/// 5299/// where N = 2^BW and BW is the common bit width of A and B. The signedness of 5300/// A and B isn't important. 5301/// 5302/// If the equation does not have a solution, SCEVCouldNotCompute is returned. 5303static const SCEV *SolveLinEquationWithOverflow(const APInt &A, const APInt &B, 5304 ScalarEvolution &SE) { 5305 uint32_t BW = A.getBitWidth(); 5306 assert(BW == B.getBitWidth() && "Bit widths must be the same."); 5307 assert(A != 0 && "A must be non-zero."); 5308 5309 // 1. D = gcd(A, N) 5310 // 5311 // The gcd of A and N may have only one prime factor: 2. The number of 5312 // trailing zeros in A is its multiplicity 5313 uint32_t Mult2 = A.countTrailingZeros(); 5314 // D = 2^Mult2 5315 5316 // 2. Check if B is divisible by D. 5317 // 5318 // B is divisible by D if and only if the multiplicity of prime factor 2 for B 5319 // is not less than multiplicity of this prime factor for D. 5320 if (B.countTrailingZeros() < Mult2) 5321 return SE.getCouldNotCompute(); 5322 5323 // 3. Compute I: the multiplicative inverse of (A / D) in arithmetic 5324 // modulo (N / D). 5325 // 5326 // (N / D) may need BW+1 bits in its representation. Hence, we'll use this 5327 // bit width during computations. 5328 APInt AD = A.lshr(Mult2).zext(BW + 1); // AD = A / D 5329 APInt Mod(BW + 1, 0); 5330 Mod.setBit(BW - Mult2); // Mod = N / D 5331 APInt I = AD.multiplicativeInverse(Mod); 5332 5333 // 4. Compute the minimum unsigned root of the equation: 5334 // I * (B / D) mod (N / D) 5335 APInt Result = (I * B.lshr(Mult2).zext(BW + 1)).urem(Mod); 5336 5337 // The result is guaranteed to be less than 2^BW so we may truncate it to BW 5338 // bits. 5339 return SE.getConstant(Result.trunc(BW)); 5340} 5341 5342/// SolveQuadraticEquation - Find the roots of the quadratic equation for the 5343/// given quadratic chrec {L,+,M,+,N}. This returns either the two roots (which 5344/// might be the same) or two SCEVCouldNotCompute objects. 5345/// 5346static std::pair<const SCEV *,const SCEV *> 5347SolveQuadraticEquation(const SCEVAddRecExpr *AddRec, ScalarEvolution &SE) { 5348 assert(AddRec->getNumOperands() == 3 && "This is not a quadratic chrec!"); 5349 const SCEVConstant *LC = dyn_cast<SCEVConstant>(AddRec->getOperand(0)); 5350 const SCEVConstant *MC = dyn_cast<SCEVConstant>(AddRec->getOperand(1)); 5351 const SCEVConstant *NC = dyn_cast<SCEVConstant>(AddRec->getOperand(2)); 5352 5353 // We currently can only solve this if the coefficients are constants. 5354 if (!LC || !MC || !NC) { 5355 const SCEV *CNC = SE.getCouldNotCompute(); 5356 return std::make_pair(CNC, CNC); 5357 } 5358 5359 uint32_t BitWidth = LC->getValue()->getValue().getBitWidth(); 5360 const APInt &L = LC->getValue()->getValue(); 5361 const APInt &M = MC->getValue()->getValue(); 5362 const APInt &N = NC->getValue()->getValue(); 5363 APInt Two(BitWidth, 2); 5364 APInt Four(BitWidth, 4); 5365 5366 { 5367 using namespace APIntOps; 5368 const APInt& C = L; 5369 // Convert from chrec coefficients to polynomial coefficients AX^2+BX+C 5370 // The B coefficient is M-N/2 5371 APInt B(M); 5372 B -= sdiv(N,Two); 5373 5374 // The A coefficient is N/2 5375 APInt A(N.sdiv(Two)); 5376 5377 // Compute the B^2-4ac term. 5378 APInt SqrtTerm(B); 5379 SqrtTerm *= B; 5380 SqrtTerm -= Four * (A * C); 5381 5382 // Compute sqrt(B^2-4ac). This is guaranteed to be the nearest 5383 // integer value or else APInt::sqrt() will assert. 5384 APInt SqrtVal(SqrtTerm.sqrt()); 5385 5386 // Compute the two solutions for the quadratic formula. 5387 // The divisions must be performed as signed divisions. 5388 APInt NegB(-B); 5389 APInt TwoA(A << 1); 5390 if (TwoA.isMinValue()) { 5391 const SCEV *CNC = SE.getCouldNotCompute(); 5392 return std::make_pair(CNC, CNC); 5393 } 5394 5395 LLVMContext &Context = SE.getContext(); 5396 5397 ConstantInt *Solution1 = 5398 ConstantInt::get(Context, (NegB + SqrtVal).sdiv(TwoA)); 5399 ConstantInt *Solution2 = 5400 ConstantInt::get(Context, (NegB - SqrtVal).sdiv(TwoA)); 5401 5402 return std::make_pair(SE.getConstant(Solution1), 5403 SE.getConstant(Solution2)); 5404 } // end APIntOps namespace 5405} 5406 5407/// HowFarToZero - Return the number of times a backedge comparing the specified 5408/// value to zero will execute. If not computable, return CouldNotCompute. 5409/// 5410/// This is only used for loops with a "x != y" exit test. The exit condition is 5411/// now expressed as a single expression, V = x-y. So the exit test is 5412/// effectively V != 0. We know and take advantage of the fact that this 5413/// expression only being used in a comparison by zero context. 5414ScalarEvolution::ExitLimit 5415ScalarEvolution::HowFarToZero(const SCEV *V, const Loop *L) { 5416 // If the value is a constant 5417 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(V)) { 5418 // If the value is already zero, the branch will execute zero times. 5419 if (C->getValue()->isZero()) return C; 5420 return getCouldNotCompute(); // Otherwise it will loop infinitely. 5421 } 5422 5423 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(V); 5424 if (!AddRec || AddRec->getLoop() != L) 5425 return getCouldNotCompute(); 5426 5427 // If this is a quadratic (3-term) AddRec {L,+,M,+,N}, find the roots of 5428 // the quadratic equation to solve it. 5429 if (AddRec->isQuadratic() && AddRec->getType()->isIntegerTy()) { 5430 std::pair<const SCEV *,const SCEV *> Roots = 5431 SolveQuadraticEquation(AddRec, *this); 5432 const SCEVConstant *R1 = dyn_cast<SCEVConstant>(Roots.first); 5433 const SCEVConstant *R2 = dyn_cast<SCEVConstant>(Roots.second); 5434 if (R1 && R2) { 5435#if 0 5436 dbgs() << "HFTZ: " << *V << " - sol#1: " << *R1 5437 << " sol#2: " << *R2 << "\n"; 5438#endif 5439 // Pick the smallest positive root value. 5440 if (ConstantInt *CB = 5441 dyn_cast<ConstantInt>(ConstantExpr::getICmp(CmpInst::ICMP_ULT, 5442 R1->getValue(), 5443 R2->getValue()))) { 5444 if (CB->getZExtValue() == false) 5445 std::swap(R1, R2); // R1 is the minimum root now. 5446 5447 // We can only use this value if the chrec ends up with an exact zero 5448 // value at this index. When solving for "X*X != 5", for example, we 5449 // should not accept a root of 2. 5450 const SCEV *Val = AddRec->evaluateAtIteration(R1, *this); 5451 if (Val->isZero()) 5452 return R1; // We found a quadratic root! 5453 } 5454 } 5455 return getCouldNotCompute(); 5456 } 5457 5458 // Otherwise we can only handle this if it is affine. 5459 if (!AddRec->isAffine()) 5460 return getCouldNotCompute(); 5461 5462 // If this is an affine expression, the execution count of this branch is 5463 // the minimum unsigned root of the following equation: 5464 // 5465 // Start + Step*N = 0 (mod 2^BW) 5466 // 5467 // equivalent to: 5468 // 5469 // Step*N = -Start (mod 2^BW) 5470 // 5471 // where BW is the common bit width of Start and Step. 5472 5473 // Get the initial value for the loop. 5474 const SCEV *Start = getSCEVAtScope(AddRec->getStart(), L->getParentLoop()); 5475 const SCEV *Step = getSCEVAtScope(AddRec->getOperand(1), L->getParentLoop()); 5476 5477 // For now we handle only constant steps. 5478 // 5479 // TODO: Handle a nonconstant Step given AddRec<NUW>. If the 5480 // AddRec is NUW, then (in an unsigned sense) it cannot be counting up to wrap 5481 // to 0, it must be counting down to equal 0. Consequently, N = Start / -Step. 5482 // We have not yet seen any such cases. 5483 const SCEVConstant *StepC = dyn_cast<SCEVConstant>(Step); 5484 if (StepC == 0) 5485 return getCouldNotCompute(); 5486 5487 // For positive steps (counting up until unsigned overflow): 5488 // N = -Start/Step (as unsigned) 5489 // For negative steps (counting down to zero): 5490 // N = Start/-Step 5491 // First compute the unsigned distance from zero in the direction of Step. 5492 bool CountDown = StepC->getValue()->getValue().isNegative(); 5493 const SCEV *Distance = CountDown ? Start : getNegativeSCEV(Start); 5494 5495 // Handle unitary steps, which cannot wraparound. 5496 // 1*N = -Start; -1*N = Start (mod 2^BW), so: 5497 // N = Distance (as unsigned) 5498 if (StepC->getValue()->equalsInt(1) || StepC->getValue()->isAllOnesValue()) { 5499 ConstantRange CR = getUnsignedRange(Start); 5500 const SCEV *MaxBECount; 5501 if (!CountDown && CR.getUnsignedMin().isMinValue()) 5502 // When counting up, the worst starting value is 1, not 0. 5503 MaxBECount = CR.getUnsignedMax().isMinValue() 5504 ? getConstant(APInt::getMinValue(CR.getBitWidth())) 5505 : getConstant(APInt::getMaxValue(CR.getBitWidth())); 5506 else 5507 MaxBECount = getConstant(CountDown ? CR.getUnsignedMax() 5508 : -CR.getUnsignedMin()); 5509 return ExitLimit(Distance, MaxBECount); 5510 } 5511 5512 // If the recurrence is known not to wraparound, unsigned divide computes the 5513 // back edge count. We know that the value will either become zero (and thus 5514 // the loop terminates), that the loop will terminate through some other exit 5515 // condition first, or that the loop has undefined behavior. This means 5516 // we can't "miss" the exit value, even with nonunit stride. 5517 // 5518 // FIXME: Prove that loops always exhibits *acceptable* undefined 5519 // behavior. Loops must exhibit defined behavior until a wrapped value is 5520 // actually used. So the trip count computed by udiv could be smaller than the 5521 // number of well-defined iterations. 5522 if (AddRec->getNoWrapFlags(SCEV::FlagNW)) { 5523 // FIXME: We really want an "isexact" bit for udiv. 5524 return getUDivExpr(Distance, CountDown ? getNegativeSCEV(Step) : Step); 5525 } 5526 // Then, try to solve the above equation provided that Start is constant. 5527 if (const SCEVConstant *StartC = dyn_cast<SCEVConstant>(Start)) 5528 return SolveLinEquationWithOverflow(StepC->getValue()->getValue(), 5529 -StartC->getValue()->getValue(), 5530 *this); 5531 return getCouldNotCompute(); 5532} 5533 5534/// HowFarToNonZero - Return the number of times a backedge checking the 5535/// specified value for nonzero will execute. If not computable, return 5536/// CouldNotCompute 5537ScalarEvolution::ExitLimit 5538ScalarEvolution::HowFarToNonZero(const SCEV *V, const Loop *L) { 5539 // Loops that look like: while (X == 0) are very strange indeed. We don't 5540 // handle them yet except for the trivial case. This could be expanded in the 5541 // future as needed. 5542 5543 // If the value is a constant, check to see if it is known to be non-zero 5544 // already. If so, the backedge will execute zero times. 5545 if (const SCEVConstant *C = dyn_cast<SCEVConstant>(V)) { 5546 if (!C->getValue()->isNullValue()) 5547 return getConstant(C->getType(), 0); 5548 return getCouldNotCompute(); // Otherwise it will loop infinitely. 5549 } 5550 5551 // We could implement others, but I really doubt anyone writes loops like 5552 // this, and if they did, they would already be constant folded. 5553 return getCouldNotCompute(); 5554} 5555 5556/// getPredecessorWithUniqueSuccessorForBB - Return a predecessor of BB 5557/// (which may not be an immediate predecessor) which has exactly one 5558/// successor from which BB is reachable, or null if no such block is 5559/// found. 5560/// 5561std::pair<BasicBlock *, BasicBlock *> 5562ScalarEvolution::getPredecessorWithUniqueSuccessorForBB(BasicBlock *BB) { 5563 // If the block has a unique predecessor, then there is no path from the 5564 // predecessor to the block that does not go through the direct edge 5565 // from the predecessor to the block. 5566 if (BasicBlock *Pred = BB->getSinglePredecessor()) 5567 return std::make_pair(Pred, BB); 5568 5569 // A loop's header is defined to be a block that dominates the loop. 5570 // If the header has a unique predecessor outside the loop, it must be 5571 // a block that has exactly one successor that can reach the loop. 5572 if (Loop *L = LI->getLoopFor(BB)) 5573 return std::make_pair(L->getLoopPredecessor(), L->getHeader()); 5574 5575 return std::pair<BasicBlock *, BasicBlock *>(); 5576} 5577 5578/// HasSameValue - SCEV structural equivalence is usually sufficient for 5579/// testing whether two expressions are equal, however for the purposes of 5580/// looking for a condition guarding a loop, it can be useful to be a little 5581/// more general, since a front-end may have replicated the controlling 5582/// expression. 5583/// 5584static bool HasSameValue(const SCEV *A, const SCEV *B) { 5585 // Quick check to see if they are the same SCEV. 5586 if (A == B) return true; 5587 5588 // Otherwise, if they're both SCEVUnknown, it's possible that they hold 5589 // two different instructions with the same value. Check for this case. 5590 if (const SCEVUnknown *AU = dyn_cast<SCEVUnknown>(A)) 5591 if (const SCEVUnknown *BU = dyn_cast<SCEVUnknown>(B)) 5592 if (const Instruction *AI = dyn_cast<Instruction>(AU->getValue())) 5593 if (const Instruction *BI = dyn_cast<Instruction>(BU->getValue())) 5594 if (AI->isIdenticalTo(BI) && !AI->mayReadFromMemory()) 5595 return true; 5596 5597 // Otherwise assume they may have a different value. 5598 return false; 5599} 5600 5601/// SimplifyICmpOperands - Simplify LHS and RHS in a comparison with 5602/// predicate Pred. Return true iff any changes were made. 5603/// 5604bool ScalarEvolution::SimplifyICmpOperands(ICmpInst::Predicate &Pred, 5605 const SCEV *&LHS, const SCEV *&RHS) { 5606 bool Changed = false; 5607 5608 // Canonicalize a constant to the right side. 5609 if (const SCEVConstant *LHSC = dyn_cast<SCEVConstant>(LHS)) { 5610 // Check for both operands constant. 5611 if (const SCEVConstant *RHSC = dyn_cast<SCEVConstant>(RHS)) { 5612 if (ConstantExpr::getICmp(Pred, 5613 LHSC->getValue(), 5614 RHSC->getValue())->isNullValue()) 5615 goto trivially_false; 5616 else 5617 goto trivially_true; 5618 } 5619 // Otherwise swap the operands to put the constant on the right. 5620 std::swap(LHS, RHS); 5621 Pred = ICmpInst::getSwappedPredicate(Pred); 5622 Changed = true; 5623 } 5624 5625 // If we're comparing an addrec with a value which is loop-invariant in the 5626 // addrec's loop, put the addrec on the left. Also make a dominance check, 5627 // as both operands could be addrecs loop-invariant in each other's loop. 5628 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(RHS)) { 5629 const Loop *L = AR->getLoop(); 5630 if (isLoopInvariant(LHS, L) && properlyDominates(LHS, L->getHeader())) { 5631 std::swap(LHS, RHS); 5632 Pred = ICmpInst::getSwappedPredicate(Pred); 5633 Changed = true; 5634 } 5635 } 5636 5637 // If there's a constant operand, canonicalize comparisons with boundary 5638 // cases, and canonicalize *-or-equal comparisons to regular comparisons. 5639 if (const SCEVConstant *RC = dyn_cast<SCEVConstant>(RHS)) { 5640 const APInt &RA = RC->getValue()->getValue(); 5641 switch (Pred) { 5642 default: llvm_unreachable("Unexpected ICmpInst::Predicate value!"); 5643 case ICmpInst::ICMP_EQ: 5644 case ICmpInst::ICMP_NE: 5645 break; 5646 case ICmpInst::ICMP_UGE: 5647 if ((RA - 1).isMinValue()) { 5648 Pred = ICmpInst::ICMP_NE; 5649 RHS = getConstant(RA - 1); 5650 Changed = true; 5651 break; 5652 } 5653 if (RA.isMaxValue()) { 5654 Pred = ICmpInst::ICMP_EQ; 5655 Changed = true; 5656 break; 5657 } 5658 if (RA.isMinValue()) goto trivially_true; 5659 5660 Pred = ICmpInst::ICMP_UGT; 5661 RHS = getConstant(RA - 1); 5662 Changed = true; 5663 break; 5664 case ICmpInst::ICMP_ULE: 5665 if ((RA + 1).isMaxValue()) { 5666 Pred = ICmpInst::ICMP_NE; 5667 RHS = getConstant(RA + 1); 5668 Changed = true; 5669 break; 5670 } 5671 if (RA.isMinValue()) { 5672 Pred = ICmpInst::ICMP_EQ; 5673 Changed = true; 5674 break; 5675 } 5676 if (RA.isMaxValue()) goto trivially_true; 5677 5678 Pred = ICmpInst::ICMP_ULT; 5679 RHS = getConstant(RA + 1); 5680 Changed = true; 5681 break; 5682 case ICmpInst::ICMP_SGE: 5683 if ((RA - 1).isMinSignedValue()) { 5684 Pred = ICmpInst::ICMP_NE; 5685 RHS = getConstant(RA - 1); 5686 Changed = true; 5687 break; 5688 } 5689 if (RA.isMaxSignedValue()) { 5690 Pred = ICmpInst::ICMP_EQ; 5691 Changed = true; 5692 break; 5693 } 5694 if (RA.isMinSignedValue()) goto trivially_true; 5695 5696 Pred = ICmpInst::ICMP_SGT; 5697 RHS = getConstant(RA - 1); 5698 Changed = true; 5699 break; 5700 case ICmpInst::ICMP_SLE: 5701 if ((RA + 1).isMaxSignedValue()) { 5702 Pred = ICmpInst::ICMP_NE; 5703 RHS = getConstant(RA + 1); 5704 Changed = true; 5705 break; 5706 } 5707 if (RA.isMinSignedValue()) { 5708 Pred = ICmpInst::ICMP_EQ; 5709 Changed = true; 5710 break; 5711 } 5712 if (RA.isMaxSignedValue()) goto trivially_true; 5713 5714 Pred = ICmpInst::ICMP_SLT; 5715 RHS = getConstant(RA + 1); 5716 Changed = true; 5717 break; 5718 case ICmpInst::ICMP_UGT: 5719 if (RA.isMinValue()) { 5720 Pred = ICmpInst::ICMP_NE; 5721 Changed = true; 5722 break; 5723 } 5724 if ((RA + 1).isMaxValue()) { 5725 Pred = ICmpInst::ICMP_EQ; 5726 RHS = getConstant(RA + 1); 5727 Changed = true; 5728 break; 5729 } 5730 if (RA.isMaxValue()) goto trivially_false; 5731 break; 5732 case ICmpInst::ICMP_ULT: 5733 if (RA.isMaxValue()) { 5734 Pred = ICmpInst::ICMP_NE; 5735 Changed = true; 5736 break; 5737 } 5738 if ((RA - 1).isMinValue()) { 5739 Pred = ICmpInst::ICMP_EQ; 5740 RHS = getConstant(RA - 1); 5741 Changed = true; 5742 break; 5743 } 5744 if (RA.isMinValue()) goto trivially_false; 5745 break; 5746 case ICmpInst::ICMP_SGT: 5747 if (RA.isMinSignedValue()) { 5748 Pred = ICmpInst::ICMP_NE; 5749 Changed = true; 5750 break; 5751 } 5752 if ((RA + 1).isMaxSignedValue()) { 5753 Pred = ICmpInst::ICMP_EQ; 5754 RHS = getConstant(RA + 1); 5755 Changed = true; 5756 break; 5757 } 5758 if (RA.isMaxSignedValue()) goto trivially_false; 5759 break; 5760 case ICmpInst::ICMP_SLT: 5761 if (RA.isMaxSignedValue()) { 5762 Pred = ICmpInst::ICMP_NE; 5763 Changed = true; 5764 break; 5765 } 5766 if ((RA - 1).isMinSignedValue()) { 5767 Pred = ICmpInst::ICMP_EQ; 5768 RHS = getConstant(RA - 1); 5769 Changed = true; 5770 break; 5771 } 5772 if (RA.isMinSignedValue()) goto trivially_false; 5773 break; 5774 } 5775 } 5776 5777 // Check for obvious equality. 5778 if (HasSameValue(LHS, RHS)) { 5779 if (ICmpInst::isTrueWhenEqual(Pred)) 5780 goto trivially_true; 5781 if (ICmpInst::isFalseWhenEqual(Pred)) 5782 goto trivially_false; 5783 } 5784 5785 // If possible, canonicalize GE/LE comparisons to GT/LT comparisons, by 5786 // adding or subtracting 1 from one of the operands. 5787 switch (Pred) { 5788 case ICmpInst::ICMP_SLE: 5789 if (!getSignedRange(RHS).getSignedMax().isMaxSignedValue()) { 5790 RHS = getAddExpr(getConstant(RHS->getType(), 1, true), RHS, 5791 SCEV::FlagNSW); 5792 Pred = ICmpInst::ICMP_SLT; 5793 Changed = true; 5794 } else if (!getSignedRange(LHS).getSignedMin().isMinSignedValue()) { 5795 LHS = getAddExpr(getConstant(RHS->getType(), (uint64_t)-1, true), LHS, 5796 SCEV::FlagNSW); 5797 Pred = ICmpInst::ICMP_SLT; 5798 Changed = true; 5799 } 5800 break; 5801 case ICmpInst::ICMP_SGE: 5802 if (!getSignedRange(RHS).getSignedMin().isMinSignedValue()) { 5803 RHS = getAddExpr(getConstant(RHS->getType(), (uint64_t)-1, true), RHS, 5804 SCEV::FlagNSW); 5805 Pred = ICmpInst::ICMP_SGT; 5806 Changed = true; 5807 } else if (!getSignedRange(LHS).getSignedMax().isMaxSignedValue()) { 5808 LHS = getAddExpr(getConstant(RHS->getType(), 1, true), LHS, 5809 SCEV::FlagNSW); 5810 Pred = ICmpInst::ICMP_SGT; 5811 Changed = true; 5812 } 5813 break; 5814 case ICmpInst::ICMP_ULE: 5815 if (!getUnsignedRange(RHS).getUnsignedMax().isMaxValue()) { 5816 RHS = getAddExpr(getConstant(RHS->getType(), 1, true), RHS, 5817 SCEV::FlagNUW); 5818 Pred = ICmpInst::ICMP_ULT; 5819 Changed = true; 5820 } else if (!getUnsignedRange(LHS).getUnsignedMin().isMinValue()) { 5821 LHS = getAddExpr(getConstant(RHS->getType(), (uint64_t)-1, true), LHS, 5822 SCEV::FlagNUW); 5823 Pred = ICmpInst::ICMP_ULT; 5824 Changed = true; 5825 } 5826 break; 5827 case ICmpInst::ICMP_UGE: 5828 if (!getUnsignedRange(RHS).getUnsignedMin().isMinValue()) { 5829 RHS = getAddExpr(getConstant(RHS->getType(), (uint64_t)-1, true), RHS, 5830 SCEV::FlagNUW); 5831 Pred = ICmpInst::ICMP_UGT; 5832 Changed = true; 5833 } else if (!getUnsignedRange(LHS).getUnsignedMax().isMaxValue()) { 5834 LHS = getAddExpr(getConstant(RHS->getType(), 1, true), LHS, 5835 SCEV::FlagNUW); 5836 Pred = ICmpInst::ICMP_UGT; 5837 Changed = true; 5838 } 5839 break; 5840 default: 5841 break; 5842 } 5843 5844 // TODO: More simplifications are possible here. 5845 5846 return Changed; 5847 5848trivially_true: 5849 // Return 0 == 0. 5850 LHS = RHS = getConstant(ConstantInt::getFalse(getContext())); 5851 Pred = ICmpInst::ICMP_EQ; 5852 return true; 5853 5854trivially_false: 5855 // Return 0 != 0. 5856 LHS = RHS = getConstant(ConstantInt::getFalse(getContext())); 5857 Pred = ICmpInst::ICMP_NE; 5858 return true; 5859} 5860 5861bool ScalarEvolution::isKnownNegative(const SCEV *S) { 5862 return getSignedRange(S).getSignedMax().isNegative(); 5863} 5864 5865bool ScalarEvolution::isKnownPositive(const SCEV *S) { 5866 return getSignedRange(S).getSignedMin().isStrictlyPositive(); 5867} 5868 5869bool ScalarEvolution::isKnownNonNegative(const SCEV *S) { 5870 return !getSignedRange(S).getSignedMin().isNegative(); 5871} 5872 5873bool ScalarEvolution::isKnownNonPositive(const SCEV *S) { 5874 return !getSignedRange(S).getSignedMax().isStrictlyPositive(); 5875} 5876 5877bool ScalarEvolution::isKnownNonZero(const SCEV *S) { 5878 return isKnownNegative(S) || isKnownPositive(S); 5879} 5880 5881bool ScalarEvolution::isKnownPredicate(ICmpInst::Predicate Pred, 5882 const SCEV *LHS, const SCEV *RHS) { 5883 // Canonicalize the inputs first. 5884 (void)SimplifyICmpOperands(Pred, LHS, RHS); 5885 5886 // If LHS or RHS is an addrec, check to see if the condition is true in 5887 // every iteration of the loop. 5888 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(LHS)) 5889 if (isLoopEntryGuardedByCond( 5890 AR->getLoop(), Pred, AR->getStart(), RHS) && 5891 isLoopBackedgeGuardedByCond( 5892 AR->getLoop(), Pred, AR->getPostIncExpr(*this), RHS)) 5893 return true; 5894 if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(RHS)) 5895 if (isLoopEntryGuardedByCond( 5896 AR->getLoop(), Pred, LHS, AR->getStart()) && 5897 isLoopBackedgeGuardedByCond( 5898 AR->getLoop(), Pred, LHS, AR->getPostIncExpr(*this))) 5899 return true; 5900 5901 // Otherwise see what can be done with known constant ranges. 5902 return isKnownPredicateWithRanges(Pred, LHS, RHS); 5903} 5904 5905bool 5906ScalarEvolution::isKnownPredicateWithRanges(ICmpInst::Predicate Pred, 5907 const SCEV *LHS, const SCEV *RHS) { 5908 if (HasSameValue(LHS, RHS)) 5909 return ICmpInst::isTrueWhenEqual(Pred); 5910 5911 // This code is split out from isKnownPredicate because it is called from 5912 // within isLoopEntryGuardedByCond. 5913 switch (Pred) { 5914 default: 5915 llvm_unreachable("Unexpected ICmpInst::Predicate value!"); 5916 case ICmpInst::ICMP_SGT: 5917 Pred = ICmpInst::ICMP_SLT; 5918 std::swap(LHS, RHS); 5919 case ICmpInst::ICMP_SLT: { 5920 ConstantRange LHSRange = getSignedRange(LHS); 5921 ConstantRange RHSRange = getSignedRange(RHS); 5922 if (LHSRange.getSignedMax().slt(RHSRange.getSignedMin())) 5923 return true; 5924 if (LHSRange.getSignedMin().sge(RHSRange.getSignedMax())) 5925 return false; 5926 break; 5927 } 5928 case ICmpInst::ICMP_SGE: 5929 Pred = ICmpInst::ICMP_SLE; 5930 std::swap(LHS, RHS); 5931 case ICmpInst::ICMP_SLE: { 5932 ConstantRange LHSRange = getSignedRange(LHS); 5933 ConstantRange RHSRange = getSignedRange(RHS); 5934 if (LHSRange.getSignedMax().sle(RHSRange.getSignedMin())) 5935 return true; 5936 if (LHSRange.getSignedMin().sgt(RHSRange.getSignedMax())) 5937 return false; 5938 break; 5939 } 5940 case ICmpInst::ICMP_UGT: 5941 Pred = ICmpInst::ICMP_ULT; 5942 std::swap(LHS, RHS); 5943 case ICmpInst::ICMP_ULT: { 5944 ConstantRange LHSRange = getUnsignedRange(LHS); 5945 ConstantRange RHSRange = getUnsignedRange(RHS); 5946 if (LHSRange.getUnsignedMax().ult(RHSRange.getUnsignedMin())) 5947 return true; 5948 if (LHSRange.getUnsignedMin().uge(RHSRange.getUnsignedMax())) 5949 return false; 5950 break; 5951 } 5952 case ICmpInst::ICMP_UGE: 5953 Pred = ICmpInst::ICMP_ULE; 5954 std::swap(LHS, RHS); 5955 case ICmpInst::ICMP_ULE: { 5956 ConstantRange LHSRange = getUnsignedRange(LHS); 5957 ConstantRange RHSRange = getUnsignedRange(RHS); 5958 if (LHSRange.getUnsignedMax().ule(RHSRange.getUnsignedMin())) 5959 return true; 5960 if (LHSRange.getUnsignedMin().ugt(RHSRange.getUnsignedMax())) 5961 return false; 5962 break; 5963 } 5964 case ICmpInst::ICMP_NE: { 5965 if (getUnsignedRange(LHS).intersectWith(getUnsignedRange(RHS)).isEmptySet()) 5966 return true; 5967 if (getSignedRange(LHS).intersectWith(getSignedRange(RHS)).isEmptySet()) 5968 return true; 5969 5970 const SCEV *Diff = getMinusSCEV(LHS, RHS); 5971 if (isKnownNonZero(Diff)) 5972 return true; 5973 break; 5974 } 5975 case ICmpInst::ICMP_EQ: 5976 // The check at the top of the function catches the case where 5977 // the values are known to be equal. 5978 break; 5979 } 5980 return false; 5981} 5982 5983/// isLoopBackedgeGuardedByCond - Test whether the backedge of the loop is 5984/// protected by a conditional between LHS and RHS. This is used to 5985/// to eliminate casts. 5986bool 5987ScalarEvolution::isLoopBackedgeGuardedByCond(const Loop *L, 5988 ICmpInst::Predicate Pred, 5989 const SCEV *LHS, const SCEV *RHS) { 5990 // Interpret a null as meaning no loop, where there is obviously no guard 5991 // (interprocedural conditions notwithstanding). 5992 if (!L) return true; 5993 5994 BasicBlock *Latch = L->getLoopLatch(); 5995 if (!Latch) 5996 return false; 5997 5998 BranchInst *LoopContinuePredicate = 5999 dyn_cast<BranchInst>(Latch->getTerminator()); 6000 if (!LoopContinuePredicate || 6001 LoopContinuePredicate->isUnconditional()) 6002 return false; 6003 6004 return isImpliedCond(Pred, LHS, RHS, 6005 LoopContinuePredicate->getCondition(), 6006 LoopContinuePredicate->getSuccessor(0) != L->getHeader()); 6007} 6008 6009/// isLoopEntryGuardedByCond - Test whether entry to the loop is protected 6010/// by a conditional between LHS and RHS. This is used to help avoid max 6011/// expressions in loop trip counts, and to eliminate casts. 6012bool 6013ScalarEvolution::isLoopEntryGuardedByCond(const Loop *L, 6014 ICmpInst::Predicate Pred, 6015 const SCEV *LHS, const SCEV *RHS) { 6016 // Interpret a null as meaning no loop, where there is obviously no guard 6017 // (interprocedural conditions notwithstanding). 6018 if (!L) return false; 6019 6020 // Starting at the loop predecessor, climb up the predecessor chain, as long 6021 // as there are predecessors that can be found that have unique successors 6022 // leading to the original header. 6023 for (std::pair<BasicBlock *, BasicBlock *> 6024 Pair(L->getLoopPredecessor(), L->getHeader()); 6025 Pair.first; 6026 Pair = getPredecessorWithUniqueSuccessorForBB(Pair.first)) { 6027 6028 BranchInst *LoopEntryPredicate = 6029 dyn_cast<BranchInst>(Pair.first->getTerminator()); 6030 if (!LoopEntryPredicate || 6031 LoopEntryPredicate->isUnconditional()) 6032 continue; 6033 6034 if (isImpliedCond(Pred, LHS, RHS, 6035 LoopEntryPredicate->getCondition(), 6036 LoopEntryPredicate->getSuccessor(0) != Pair.second)) 6037 return true; 6038 } 6039 6040 return false; 6041} 6042 6043/// isImpliedCond - Test whether the condition described by Pred, LHS, 6044/// and RHS is true whenever the given Cond value evaluates to true. 6045bool ScalarEvolution::isImpliedCond(ICmpInst::Predicate Pred, 6046 const SCEV *LHS, const SCEV *RHS, 6047 Value *FoundCondValue, 6048 bool Inverse) { 6049 // Recursively handle And and Or conditions. 6050 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(FoundCondValue)) { 6051 if (BO->getOpcode() == Instruction::And) { 6052 if (!Inverse) 6053 return isImpliedCond(Pred, LHS, RHS, BO->getOperand(0), Inverse) || 6054 isImpliedCond(Pred, LHS, RHS, BO->getOperand(1), Inverse); 6055 } else if (BO->getOpcode() == Instruction::Or) { 6056 if (Inverse) 6057 return isImpliedCond(Pred, LHS, RHS, BO->getOperand(0), Inverse) || 6058 isImpliedCond(Pred, LHS, RHS, BO->getOperand(1), Inverse); 6059 } 6060 } 6061 6062 ICmpInst *ICI = dyn_cast<ICmpInst>(FoundCondValue); 6063 if (!ICI) return false; 6064 6065 // Bail if the ICmp's operands' types are wider than the needed type 6066 // before attempting to call getSCEV on them. This avoids infinite 6067 // recursion, since the analysis of widening casts can require loop 6068 // exit condition information for overflow checking, which would 6069 // lead back here. 6070 if (getTypeSizeInBits(LHS->getType()) < 6071 getTypeSizeInBits(ICI->getOperand(0)->getType())) 6072 return false; 6073 6074 // Now that we found a conditional branch that dominates the loop, check to 6075 // see if it is the comparison we are looking for. 6076 ICmpInst::Predicate FoundPred; 6077 if (Inverse) 6078 FoundPred = ICI->getInversePredicate(); 6079 else 6080 FoundPred = ICI->getPredicate(); 6081 6082 const SCEV *FoundLHS = getSCEV(ICI->getOperand(0)); 6083 const SCEV *FoundRHS = getSCEV(ICI->getOperand(1)); 6084 6085 // Balance the types. The case where FoundLHS' type is wider than 6086 // LHS' type is checked for above. 6087 if (getTypeSizeInBits(LHS->getType()) > 6088 getTypeSizeInBits(FoundLHS->getType())) { 6089 if (CmpInst::isSigned(Pred)) { 6090 FoundLHS = getSignExtendExpr(FoundLHS, LHS->getType()); 6091 FoundRHS = getSignExtendExpr(FoundRHS, LHS->getType()); 6092 } else { 6093 FoundLHS = getZeroExtendExpr(FoundLHS, LHS->getType()); 6094 FoundRHS = getZeroExtendExpr(FoundRHS, LHS->getType()); 6095 } 6096 } 6097 6098 // Canonicalize the query to match the way instcombine will have 6099 // canonicalized the comparison. 6100 if (SimplifyICmpOperands(Pred, LHS, RHS)) 6101 if (LHS == RHS) 6102 return CmpInst::isTrueWhenEqual(Pred); 6103 if (SimplifyICmpOperands(FoundPred, FoundLHS, FoundRHS)) 6104 if (FoundLHS == FoundRHS) 6105 return CmpInst::isFalseWhenEqual(Pred); 6106 6107 // Check to see if we can make the LHS or RHS match. 6108 if (LHS == FoundRHS || RHS == FoundLHS) { 6109 if (isa<SCEVConstant>(RHS)) { 6110 std::swap(FoundLHS, FoundRHS); 6111 FoundPred = ICmpInst::getSwappedPredicate(FoundPred); 6112 } else { 6113 std::swap(LHS, RHS); 6114 Pred = ICmpInst::getSwappedPredicate(Pred); 6115 } 6116 } 6117 6118 // Check whether the found predicate is the same as the desired predicate. 6119 if (FoundPred == Pred) 6120 return isImpliedCondOperands(Pred, LHS, RHS, FoundLHS, FoundRHS); 6121 6122 // Check whether swapping the found predicate makes it the same as the 6123 // desired predicate. 6124 if (ICmpInst::getSwappedPredicate(FoundPred) == Pred) { 6125 if (isa<SCEVConstant>(RHS)) 6126 return isImpliedCondOperands(Pred, LHS, RHS, FoundRHS, FoundLHS); 6127 else 6128 return isImpliedCondOperands(ICmpInst::getSwappedPredicate(Pred), 6129 RHS, LHS, FoundLHS, FoundRHS); 6130 } 6131 6132 // Check whether the actual condition is beyond sufficient. 6133 if (FoundPred == ICmpInst::ICMP_EQ) 6134 if (ICmpInst::isTrueWhenEqual(Pred)) 6135 if (isImpliedCondOperands(Pred, LHS, RHS, FoundLHS, FoundRHS)) 6136 return true; 6137 if (Pred == ICmpInst::ICMP_NE) 6138 if (!ICmpInst::isTrueWhenEqual(FoundPred)) 6139 if (isImpliedCondOperands(FoundPred, LHS, RHS, FoundLHS, FoundRHS)) 6140 return true; 6141 6142 // Otherwise assume the worst. 6143 return false; 6144} 6145 6146/// isImpliedCondOperands - Test whether the condition described by Pred, 6147/// LHS, and RHS is true whenever the condition described by Pred, FoundLHS, 6148/// and FoundRHS is true. 6149bool ScalarEvolution::isImpliedCondOperands(ICmpInst::Predicate Pred, 6150 const SCEV *LHS, const SCEV *RHS, 6151 const SCEV *FoundLHS, 6152 const SCEV *FoundRHS) { 6153 return isImpliedCondOperandsHelper(Pred, LHS, RHS, 6154 FoundLHS, FoundRHS) || 6155 // ~x < ~y --> x > y 6156 isImpliedCondOperandsHelper(Pred, LHS, RHS, 6157 getNotSCEV(FoundRHS), 6158 getNotSCEV(FoundLHS)); 6159} 6160 6161/// isImpliedCondOperandsHelper - Test whether the condition described by 6162/// Pred, LHS, and RHS is true whenever the condition described by Pred, 6163/// FoundLHS, and FoundRHS is true. 6164bool 6165ScalarEvolution::isImpliedCondOperandsHelper(ICmpInst::Predicate Pred, 6166 const SCEV *LHS, const SCEV *RHS, 6167 const SCEV *FoundLHS, 6168 const SCEV *FoundRHS) { 6169 switch (Pred) { 6170 default: llvm_unreachable("Unexpected ICmpInst::Predicate value!"); 6171 case ICmpInst::ICMP_EQ: 6172 case ICmpInst::ICMP_NE: 6173 if (HasSameValue(LHS, FoundLHS) && HasSameValue(RHS, FoundRHS)) 6174 return true; 6175 break; 6176 case ICmpInst::ICMP_SLT: 6177 case ICmpInst::ICMP_SLE: 6178 if (isKnownPredicateWithRanges(ICmpInst::ICMP_SLE, LHS, FoundLHS) && 6179 isKnownPredicateWithRanges(ICmpInst::ICMP_SGE, RHS, FoundRHS)) 6180 return true; 6181 break; 6182 case ICmpInst::ICMP_SGT: 6183 case ICmpInst::ICMP_SGE: 6184 if (isKnownPredicateWithRanges(ICmpInst::ICMP_SGE, LHS, FoundLHS) && 6185 isKnownPredicateWithRanges(ICmpInst::ICMP_SLE, RHS, FoundRHS)) 6186 return true; 6187 break; 6188 case ICmpInst::ICMP_ULT: 6189 case ICmpInst::ICMP_ULE: 6190 if (isKnownPredicateWithRanges(ICmpInst::ICMP_ULE, LHS, FoundLHS) && 6191 isKnownPredicateWithRanges(ICmpInst::ICMP_UGE, RHS, FoundRHS)) 6192 return true; 6193 break; 6194 case ICmpInst::ICMP_UGT: 6195 case ICmpInst::ICMP_UGE: 6196 if (isKnownPredicateWithRanges(ICmpInst::ICMP_UGE, LHS, FoundLHS) && 6197 isKnownPredicateWithRanges(ICmpInst::ICMP_ULE, RHS, FoundRHS)) 6198 return true; 6199 break; 6200 } 6201 6202 return false; 6203} 6204 6205/// getBECount - Subtract the end and start values and divide by the step, 6206/// rounding up, to get the number of times the backedge is executed. Return 6207/// CouldNotCompute if an intermediate computation overflows. 6208const SCEV *ScalarEvolution::getBECount(const SCEV *Start, 6209 const SCEV *End, 6210 const SCEV *Step, 6211 bool NoWrap) { 6212 assert(!isKnownNegative(Step) && 6213 "This code doesn't handle negative strides yet!"); 6214 6215 Type *Ty = Start->getType(); 6216 6217 // When Start == End, we have an exact BECount == 0. Short-circuit this case 6218 // here because SCEV may not be able to determine that the unsigned division 6219 // after rounding is zero. 6220 if (Start == End) 6221 return getConstant(Ty, 0); 6222 6223 const SCEV *NegOne = getConstant(Ty, (uint64_t)-1); 6224 const SCEV *Diff = getMinusSCEV(End, Start); 6225 const SCEV *RoundUp = getAddExpr(Step, NegOne); 6226 6227 // Add an adjustment to the difference between End and Start so that 6228 // the division will effectively round up. 6229 const SCEV *Add = getAddExpr(Diff, RoundUp); 6230 6231 if (!NoWrap) { 6232 // Check Add for unsigned overflow. 6233 // TODO: More sophisticated things could be done here. 6234 Type *WideTy = IntegerType::get(getContext(), 6235 getTypeSizeInBits(Ty) + 1); 6236 const SCEV *EDiff = getZeroExtendExpr(Diff, WideTy); 6237 const SCEV *ERoundUp = getZeroExtendExpr(RoundUp, WideTy); 6238 const SCEV *OperandExtendedAdd = getAddExpr(EDiff, ERoundUp); 6239 if (getZeroExtendExpr(Add, WideTy) != OperandExtendedAdd) 6240 return getCouldNotCompute(); 6241 } 6242 6243 return getUDivExpr(Add, Step); 6244} 6245 6246/// HowManyLessThans - Return the number of times a backedge containing the 6247/// specified less-than comparison will execute. If not computable, return 6248/// CouldNotCompute. 6249ScalarEvolution::ExitLimit 6250ScalarEvolution::HowManyLessThans(const SCEV *LHS, const SCEV *RHS, 6251 const Loop *L, bool isSigned) { 6252 // Only handle: "ADDREC < LoopInvariant". 6253 if (!isLoopInvariant(RHS, L)) return getCouldNotCompute(); 6254 6255 const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(LHS); 6256 if (!AddRec || AddRec->getLoop() != L) 6257 return getCouldNotCompute(); 6258 6259 // Check to see if we have a flag which makes analysis easy. 6260 bool NoWrap = isSigned ? 6261 AddRec->getNoWrapFlags((SCEV::NoWrapFlags)(SCEV::FlagNSW | SCEV::FlagNW)) : 6262 AddRec->getNoWrapFlags((SCEV::NoWrapFlags)(SCEV::FlagNUW | SCEV::FlagNW)); 6263 6264 if (AddRec->isAffine()) { 6265 unsigned BitWidth = getTypeSizeInBits(AddRec->getType()); 6266 const SCEV *Step = AddRec->getStepRecurrence(*this); 6267 6268 if (Step->isZero()) 6269 return getCouldNotCompute(); 6270 if (Step->isOne()) { 6271 // With unit stride, the iteration never steps past the limit value. 6272 } else if (isKnownPositive(Step)) { 6273 // Test whether a positive iteration can step past the limit 6274 // value and past the maximum value for its type in a single step. 6275 // Note that it's not sufficient to check NoWrap here, because even 6276 // though the value after a wrap is undefined, it's not undefined 6277 // behavior, so if wrap does occur, the loop could either terminate or 6278 // loop infinitely, but in either case, the loop is guaranteed to 6279 // iterate at least until the iteration where the wrapping occurs. 6280 const SCEV *One = getConstant(Step->getType(), 1); 6281 if (isSigned) { 6282 APInt Max = APInt::getSignedMaxValue(BitWidth); 6283 if ((Max - getSignedRange(getMinusSCEV(Step, One)).getSignedMax()) 6284 .slt(getSignedRange(RHS).getSignedMax())) 6285 return getCouldNotCompute(); 6286 } else { 6287 APInt Max = APInt::getMaxValue(BitWidth); 6288 if ((Max - getUnsignedRange(getMinusSCEV(Step, One)).getUnsignedMax()) 6289 .ult(getUnsignedRange(RHS).getUnsignedMax())) 6290 return getCouldNotCompute(); 6291 } 6292 } else 6293 // TODO: Handle negative strides here and below. 6294 return getCouldNotCompute(); 6295 6296 // We know the LHS is of the form {n,+,s} and the RHS is some loop-invariant 6297 // m. So, we count the number of iterations in which {n,+,s} < m is true. 6298 // Note that we cannot simply return max(m-n,0)/s because it's not safe to 6299 // treat m-n as signed nor unsigned due to overflow possibility. 6300 6301 // First, we get the value of the LHS in the first iteration: n 6302 const SCEV *Start = AddRec->getOperand(0); 6303 6304 // Determine the minimum constant start value. 6305 const SCEV *MinStart = getConstant(isSigned ? 6306 getSignedRange(Start).getSignedMin() : 6307 getUnsignedRange(Start).getUnsignedMin()); 6308 6309 // If we know that the condition is true in order to enter the loop, 6310 // then we know that it will run exactly (m-n)/s times. Otherwise, we 6311 // only know that it will execute (max(m,n)-n)/s times. In both cases, 6312 // the division must round up. 6313 const SCEV *End = RHS; 6314 if (!isLoopEntryGuardedByCond(L, 6315 isSigned ? ICmpInst::ICMP_SLT : 6316 ICmpInst::ICMP_ULT, 6317 getMinusSCEV(Start, Step), RHS)) 6318 End = isSigned ? getSMaxExpr(RHS, Start) 6319 : getUMaxExpr(RHS, Start); 6320 6321 // Determine the maximum constant end value. 6322 const SCEV *MaxEnd = getConstant(isSigned ? 6323 getSignedRange(End).getSignedMax() : 6324 getUnsignedRange(End).getUnsignedMax()); 6325 6326 // If MaxEnd is within a step of the maximum integer value in its type, 6327 // adjust it down to the minimum value which would produce the same effect. 6328 // This allows the subsequent ceiling division of (N+(step-1))/step to 6329 // compute the correct value. 6330 const SCEV *StepMinusOne = getMinusSCEV(Step, 6331 getConstant(Step->getType(), 1)); 6332 MaxEnd = isSigned ? 6333 getSMinExpr(MaxEnd, 6334 getMinusSCEV(getConstant(APInt::getSignedMaxValue(BitWidth)), 6335 StepMinusOne)) : 6336 getUMinExpr(MaxEnd, 6337 getMinusSCEV(getConstant(APInt::getMaxValue(BitWidth)), 6338 StepMinusOne)); 6339 6340 // Finally, we subtract these two values and divide, rounding up, to get 6341 // the number of times the backedge is executed. 6342 const SCEV *BECount = getBECount(Start, End, Step, NoWrap); 6343 6344 // The maximum backedge count is similar, except using the minimum start 6345 // value and the maximum end value. 6346 // If we already have an exact constant BECount, use it instead. 6347 const SCEV *MaxBECount = isa<SCEVConstant>(BECount) ? BECount 6348 : getBECount(MinStart, MaxEnd, Step, NoWrap); 6349 6350 // If the stride is nonconstant, and NoWrap == true, then 6351 // getBECount(MinStart, MaxEnd) may not compute. This would result in an 6352 // exact BECount and invalid MaxBECount, which should be avoided to catch 6353 // more optimization opportunities. 6354 if (isa<SCEVCouldNotCompute>(MaxBECount)) 6355 MaxBECount = BECount; 6356 6357 return ExitLimit(BECount, MaxBECount); 6358 } 6359 6360 return getCouldNotCompute(); 6361} 6362 6363/// getNumIterationsInRange - Return the number of iterations of this loop that 6364/// produce values in the specified constant range. Another way of looking at 6365/// this is that it returns the first iteration number where the value is not in 6366/// the condition, thus computing the exit count. If the iteration count can't 6367/// be computed, an instance of SCEVCouldNotCompute is returned. 6368const SCEV *SCEVAddRecExpr::getNumIterationsInRange(ConstantRange Range, 6369 ScalarEvolution &SE) const { 6370 if (Range.isFullSet()) // Infinite loop. 6371 return SE.getCouldNotCompute(); 6372 6373 // If the start is a non-zero constant, shift the range to simplify things. 6374 if (const SCEVConstant *SC = dyn_cast<SCEVConstant>(getStart())) 6375 if (!SC->getValue()->isZero()) { 6376 SmallVector<const SCEV *, 4> Operands(op_begin(), op_end()); 6377 Operands[0] = SE.getConstant(SC->getType(), 0); 6378 const SCEV *Shifted = SE.getAddRecExpr(Operands, getLoop(), 6379 getNoWrapFlags(FlagNW)); 6380 if (const SCEVAddRecExpr *ShiftedAddRec = 6381 dyn_cast<SCEVAddRecExpr>(Shifted)) 6382 return ShiftedAddRec->getNumIterationsInRange( 6383 Range.subtract(SC->getValue()->getValue()), SE); 6384 // This is strange and shouldn't happen. 6385 return SE.getCouldNotCompute(); 6386 } 6387 6388 // The only time we can solve this is when we have all constant indices. 6389 // Otherwise, we cannot determine the overflow conditions. 6390 for (unsigned i = 0, e = getNumOperands(); i != e; ++i) 6391 if (!isa<SCEVConstant>(getOperand(i))) 6392 return SE.getCouldNotCompute(); 6393 6394 6395 // Okay at this point we know that all elements of the chrec are constants and 6396 // that the start element is zero. 6397 6398 // First check to see if the range contains zero. If not, the first 6399 // iteration exits. 6400 unsigned BitWidth = SE.getTypeSizeInBits(getType()); 6401 if (!Range.contains(APInt(BitWidth, 0))) 6402 return SE.getConstant(getType(), 0); 6403 6404 if (isAffine()) { 6405 // If this is an affine expression then we have this situation: 6406 // Solve {0,+,A} in Range === Ax in Range 6407 6408 // We know that zero is in the range. If A is positive then we know that 6409 // the upper value of the range must be the first possible exit value. 6410 // If A is negative then the lower of the range is the last possible loop 6411 // value. Also note that we already checked for a full range. 6412 APInt One(BitWidth,1); 6413 APInt A = cast<SCEVConstant>(getOperand(1))->getValue()->getValue(); 6414 APInt End = A.sge(One) ? (Range.getUpper() - One) : Range.getLower(); 6415 6416 // The exit value should be (End+A)/A. 6417 APInt ExitVal = (End + A).udiv(A); 6418 ConstantInt *ExitValue = ConstantInt::get(SE.getContext(), ExitVal); 6419 6420 // Evaluate at the exit value. If we really did fall out of the valid 6421 // range, then we computed our trip count, otherwise wrap around or other 6422 // things must have happened. 6423 ConstantInt *Val = EvaluateConstantChrecAtConstant(this, ExitValue, SE); 6424 if (Range.contains(Val->getValue())) 6425 return SE.getCouldNotCompute(); // Something strange happened 6426 6427 // Ensure that the previous value is in the range. This is a sanity check. 6428 assert(Range.contains( 6429 EvaluateConstantChrecAtConstant(this, 6430 ConstantInt::get(SE.getContext(), ExitVal - One), SE)->getValue()) && 6431 "Linear scev computation is off in a bad way!"); 6432 return SE.getConstant(ExitValue); 6433 } else if (isQuadratic()) { 6434 // If this is a quadratic (3-term) AddRec {L,+,M,+,N}, find the roots of the 6435 // quadratic equation to solve it. To do this, we must frame our problem in 6436 // terms of figuring out when zero is crossed, instead of when 6437 // Range.getUpper() is crossed. 6438 SmallVector<const SCEV *, 4> NewOps(op_begin(), op_end()); 6439 NewOps[0] = SE.getNegativeSCEV(SE.getConstant(Range.getUpper())); 6440 const SCEV *NewAddRec = SE.getAddRecExpr(NewOps, getLoop(), 6441 // getNoWrapFlags(FlagNW) 6442 FlagAnyWrap); 6443 6444 // Next, solve the constructed addrec 6445 std::pair<const SCEV *,const SCEV *> Roots = 6446 SolveQuadraticEquation(cast<SCEVAddRecExpr>(NewAddRec), SE); 6447 const SCEVConstant *R1 = dyn_cast<SCEVConstant>(Roots.first); 6448 const SCEVConstant *R2 = dyn_cast<SCEVConstant>(Roots.second); 6449 if (R1) { 6450 // Pick the smallest positive root value. 6451 if (ConstantInt *CB = 6452 dyn_cast<ConstantInt>(ConstantExpr::getICmp(ICmpInst::ICMP_ULT, 6453 R1->getValue(), R2->getValue()))) { 6454 if (CB->getZExtValue() == false) 6455 std::swap(R1, R2); // R1 is the minimum root now. 6456 6457 // Make sure the root is not off by one. The returned iteration should 6458 // not be in the range, but the previous one should be. When solving 6459 // for "X*X < 5", for example, we should not return a root of 2. 6460 ConstantInt *R1Val = EvaluateConstantChrecAtConstant(this, 6461 R1->getValue(), 6462 SE); 6463 if (Range.contains(R1Val->getValue())) { 6464 // The next iteration must be out of the range... 6465 ConstantInt *NextVal = 6466 ConstantInt::get(SE.getContext(), R1->getValue()->getValue()+1); 6467 6468 R1Val = EvaluateConstantChrecAtConstant(this, NextVal, SE); 6469 if (!Range.contains(R1Val->getValue())) 6470 return SE.getConstant(NextVal); 6471 return SE.getCouldNotCompute(); // Something strange happened 6472 } 6473 6474 // If R1 was not in the range, then it is a good return value. Make 6475 // sure that R1-1 WAS in the range though, just in case. 6476 ConstantInt *NextVal = 6477 ConstantInt::get(SE.getContext(), R1->getValue()->getValue()-1); 6478 R1Val = EvaluateConstantChrecAtConstant(this, NextVal, SE); 6479 if (Range.contains(R1Val->getValue())) 6480 return R1; 6481 return SE.getCouldNotCompute(); // Something strange happened 6482 } 6483 } 6484 } 6485 6486 return SE.getCouldNotCompute(); 6487} 6488 6489 6490 6491//===----------------------------------------------------------------------===// 6492// SCEVCallbackVH Class Implementation 6493//===----------------------------------------------------------------------===// 6494 6495void ScalarEvolution::SCEVCallbackVH::deleted() { 6496 assert(SE && "SCEVCallbackVH called with a null ScalarEvolution!"); 6497 if (PHINode *PN = dyn_cast<PHINode>(getValPtr())) 6498 SE->ConstantEvolutionLoopExitValue.erase(PN); 6499 SE->ValueExprMap.erase(getValPtr()); 6500 // this now dangles! 6501} 6502 6503void ScalarEvolution::SCEVCallbackVH::allUsesReplacedWith(Value *V) { 6504 assert(SE && "SCEVCallbackVH called with a null ScalarEvolution!"); 6505 6506 // Forget all the expressions associated with users of the old value, 6507 // so that future queries will recompute the expressions using the new 6508 // value. 6509 Value *Old = getValPtr(); 6510 SmallVector<User *, 16> Worklist; 6511 SmallPtrSet<User *, 8> Visited; 6512 for (Value::use_iterator UI = Old->use_begin(), UE = Old->use_end(); 6513 UI != UE; ++UI) 6514 Worklist.push_back(*UI); 6515 while (!Worklist.empty()) { 6516 User *U = Worklist.pop_back_val(); 6517 // Deleting the Old value will cause this to dangle. Postpone 6518 // that until everything else is done. 6519 if (U == Old) 6520 continue; 6521 if (!Visited.insert(U)) 6522 continue; 6523 if (PHINode *PN = dyn_cast<PHINode>(U)) 6524 SE->ConstantEvolutionLoopExitValue.erase(PN); 6525 SE->ValueExprMap.erase(U); 6526 for (Value::use_iterator UI = U->use_begin(), UE = U->use_end(); 6527 UI != UE; ++UI) 6528 Worklist.push_back(*UI); 6529 } 6530 // Delete the Old value. 6531 if (PHINode *PN = dyn_cast<PHINode>(Old)) 6532 SE->ConstantEvolutionLoopExitValue.erase(PN); 6533 SE->ValueExprMap.erase(Old); 6534 // this now dangles! 6535} 6536 6537ScalarEvolution::SCEVCallbackVH::SCEVCallbackVH(Value *V, ScalarEvolution *se) 6538 : CallbackVH(V), SE(se) {} 6539 6540//===----------------------------------------------------------------------===// 6541// ScalarEvolution Class Implementation 6542//===----------------------------------------------------------------------===// 6543 6544ScalarEvolution::ScalarEvolution() 6545 : FunctionPass(ID), FirstUnknown(0) { 6546 initializeScalarEvolutionPass(*PassRegistry::getPassRegistry()); 6547} 6548 6549bool ScalarEvolution::runOnFunction(Function &F) { 6550 this->F = &F; 6551 LI = &getAnalysis<LoopInfo>(); 6552 TD = getAnalysisIfAvailable<TargetData>(); 6553 TLI = &getAnalysis<TargetLibraryInfo>(); 6554 DT = &getAnalysis<DominatorTree>(); 6555 return false; 6556} 6557 6558void ScalarEvolution::releaseMemory() { 6559 // Iterate through all the SCEVUnknown instances and call their 6560 // destructors, so that they release their references to their values. 6561 for (SCEVUnknown *U = FirstUnknown; U; U = U->Next) 6562 U->~SCEVUnknown(); 6563 FirstUnknown = 0; 6564 6565 ValueExprMap.clear(); 6566 6567 // Free any extra memory created for ExitNotTakenInfo in the unlikely event 6568 // that a loop had multiple computable exits. 6569 for (DenseMap<const Loop*, BackedgeTakenInfo>::iterator I = 6570 BackedgeTakenCounts.begin(), E = BackedgeTakenCounts.end(); 6571 I != E; ++I) { 6572 I->second.clear(); 6573 } 6574 6575 BackedgeTakenCounts.clear(); 6576 ConstantEvolutionLoopExitValue.clear(); 6577 ValuesAtScopes.clear(); 6578 LoopDispositions.clear(); 6579 BlockDispositions.clear(); 6580 UnsignedRanges.clear(); 6581 SignedRanges.clear(); 6582 UniqueSCEVs.clear(); 6583 SCEVAllocator.Reset(); 6584} 6585 6586void ScalarEvolution::getAnalysisUsage(AnalysisUsage &AU) const { 6587 AU.setPreservesAll(); 6588 AU.addRequiredTransitive<LoopInfo>(); 6589 AU.addRequiredTransitive<DominatorTree>(); 6590 AU.addRequired<TargetLibraryInfo>(); 6591} 6592 6593bool ScalarEvolution::hasLoopInvariantBackedgeTakenCount(const Loop *L) { 6594 return !isa<SCEVCouldNotCompute>(getBackedgeTakenCount(L)); 6595} 6596 6597static void PrintLoopInfo(raw_ostream &OS, ScalarEvolution *SE, 6598 const Loop *L) { 6599 // Print all inner loops first 6600 for (Loop::iterator I = L->begin(), E = L->end(); I != E; ++I) 6601 PrintLoopInfo(OS, SE, *I); 6602 6603 OS << "Loop "; 6604 WriteAsOperand(OS, L->getHeader(), /*PrintType=*/false); 6605 OS << ": "; 6606 6607 SmallVector<BasicBlock *, 8> ExitBlocks; 6608 L->getExitBlocks(ExitBlocks); 6609 if (ExitBlocks.size() != 1) 6610 OS << "<multiple exits> "; 6611 6612 if (SE->hasLoopInvariantBackedgeTakenCount(L)) { 6613 OS << "backedge-taken count is " << *SE->getBackedgeTakenCount(L); 6614 } else { 6615 OS << "Unpredictable backedge-taken count. "; 6616 } 6617 6618 OS << "\n" 6619 "Loop "; 6620 WriteAsOperand(OS, L->getHeader(), /*PrintType=*/false); 6621 OS << ": "; 6622 6623 if (!isa<SCEVCouldNotCompute>(SE->getMaxBackedgeTakenCount(L))) { 6624 OS << "max backedge-taken count is " << *SE->getMaxBackedgeTakenCount(L); 6625 } else { 6626 OS << "Unpredictable max backedge-taken count. "; 6627 } 6628 6629 OS << "\n"; 6630} 6631 6632void ScalarEvolution::print(raw_ostream &OS, const Module *) const { 6633 // ScalarEvolution's implementation of the print method is to print 6634 // out SCEV values of all instructions that are interesting. Doing 6635 // this potentially causes it to create new SCEV objects though, 6636 // which technically conflicts with the const qualifier. This isn't 6637 // observable from outside the class though, so casting away the 6638 // const isn't dangerous. 6639 ScalarEvolution &SE = *const_cast<ScalarEvolution *>(this); 6640 6641 OS << "Classifying expressions for: "; 6642 WriteAsOperand(OS, F, /*PrintType=*/false); 6643 OS << "\n"; 6644 for (inst_iterator I = inst_begin(F), E = inst_end(F); I != E; ++I) 6645 if (isSCEVable(I->getType()) && !isa<CmpInst>(*I)) { 6646 OS << *I << '\n'; 6647 OS << " --> "; 6648 const SCEV *SV = SE.getSCEV(&*I); 6649 SV->print(OS); 6650 6651 const Loop *L = LI->getLoopFor((*I).getParent()); 6652 6653 const SCEV *AtUse = SE.getSCEVAtScope(SV, L); 6654 if (AtUse != SV) { 6655 OS << " --> "; 6656 AtUse->print(OS); 6657 } 6658 6659 if (L) { 6660 OS << "\t\t" "Exits: "; 6661 const SCEV *ExitValue = SE.getSCEVAtScope(SV, L->getParentLoop()); 6662 if (!SE.isLoopInvariant(ExitValue, L)) { 6663 OS << "<<Unknown>>"; 6664 } else { 6665 OS << *ExitValue; 6666 } 6667 } 6668 6669 OS << "\n"; 6670 } 6671 6672 OS << "Determining loop execution counts for: "; 6673 WriteAsOperand(OS, F, /*PrintType=*/false); 6674 OS << "\n"; 6675 for (LoopInfo::iterator I = LI->begin(), E = LI->end(); I != E; ++I) 6676 PrintLoopInfo(OS, &SE, *I); 6677} 6678 6679ScalarEvolution::LoopDisposition 6680ScalarEvolution::getLoopDisposition(const SCEV *S, const Loop *L) { 6681 std::map<const Loop *, LoopDisposition> &Values = LoopDispositions[S]; 6682 std::pair<std::map<const Loop *, LoopDisposition>::iterator, bool> Pair = 6683 Values.insert(std::make_pair(L, LoopVariant)); 6684 if (!Pair.second) 6685 return Pair.first->second; 6686 6687 LoopDisposition D = computeLoopDisposition(S, L); 6688 return LoopDispositions[S][L] = D; 6689} 6690 6691ScalarEvolution::LoopDisposition 6692ScalarEvolution::computeLoopDisposition(const SCEV *S, const Loop *L) { 6693 switch (S->getSCEVType()) { 6694 case scConstant: 6695 return LoopInvariant; 6696 case scTruncate: 6697 case scZeroExtend: 6698 case scSignExtend: 6699 return getLoopDisposition(cast<SCEVCastExpr>(S)->getOperand(), L); 6700 case scAddRecExpr: { 6701 const SCEVAddRecExpr *AR = cast<SCEVAddRecExpr>(S); 6702 6703 // If L is the addrec's loop, it's computable. 6704 if (AR->getLoop() == L) 6705 return LoopComputable; 6706 6707 // Add recurrences are never invariant in the function-body (null loop). 6708 if (!L) 6709 return LoopVariant; 6710 6711 // This recurrence is variant w.r.t. L if L contains AR's loop. 6712 if (L->contains(AR->getLoop())) 6713 return LoopVariant; 6714 6715 // This recurrence is invariant w.r.t. L if AR's loop contains L. 6716 if (AR->getLoop()->contains(L)) 6717 return LoopInvariant; 6718 6719 // This recurrence is variant w.r.t. L if any of its operands 6720 // are variant. 6721 for (SCEVAddRecExpr::op_iterator I = AR->op_begin(), E = AR->op_end(); 6722 I != E; ++I) 6723 if (!isLoopInvariant(*I, L)) 6724 return LoopVariant; 6725 6726 // Otherwise it's loop-invariant. 6727 return LoopInvariant; 6728 } 6729 case scAddExpr: 6730 case scMulExpr: 6731 case scUMaxExpr: 6732 case scSMaxExpr: { 6733 const SCEVNAryExpr *NAry = cast<SCEVNAryExpr>(S); 6734 bool HasVarying = false; 6735 for (SCEVNAryExpr::op_iterator I = NAry->op_begin(), E = NAry->op_end(); 6736 I != E; ++I) { 6737 LoopDisposition D = getLoopDisposition(*I, L); 6738 if (D == LoopVariant) 6739 return LoopVariant; 6740 if (D == LoopComputable) 6741 HasVarying = true; 6742 } 6743 return HasVarying ? LoopComputable : LoopInvariant; 6744 } 6745 case scUDivExpr: { 6746 const SCEVUDivExpr *UDiv = cast<SCEVUDivExpr>(S); 6747 LoopDisposition LD = getLoopDisposition(UDiv->getLHS(), L); 6748 if (LD == LoopVariant) 6749 return LoopVariant; 6750 LoopDisposition RD = getLoopDisposition(UDiv->getRHS(), L); 6751 if (RD == LoopVariant) 6752 return LoopVariant; 6753 return (LD == LoopInvariant && RD == LoopInvariant) ? 6754 LoopInvariant : LoopComputable; 6755 } 6756 case scUnknown: 6757 // All non-instruction values are loop invariant. All instructions are loop 6758 // invariant if they are not contained in the specified loop. 6759 // Instructions are never considered invariant in the function body 6760 // (null loop) because they are defined within the "loop". 6761 if (Instruction *I = dyn_cast<Instruction>(cast<SCEVUnknown>(S)->getValue())) 6762 return (L && !L->contains(I)) ? LoopInvariant : LoopVariant; 6763 return LoopInvariant; 6764 case scCouldNotCompute: 6765 llvm_unreachable("Attempt to use a SCEVCouldNotCompute object!"); 6766 default: llvm_unreachable("Unknown SCEV kind!"); 6767 } 6768} 6769 6770bool ScalarEvolution::isLoopInvariant(const SCEV *S, const Loop *L) { 6771 return getLoopDisposition(S, L) == LoopInvariant; 6772} 6773 6774bool ScalarEvolution::hasComputableLoopEvolution(const SCEV *S, const Loop *L) { 6775 return getLoopDisposition(S, L) == LoopComputable; 6776} 6777 6778ScalarEvolution::BlockDisposition 6779ScalarEvolution::getBlockDisposition(const SCEV *S, const BasicBlock *BB) { 6780 std::map<const BasicBlock *, BlockDisposition> &Values = BlockDispositions[S]; 6781 std::pair<std::map<const BasicBlock *, BlockDisposition>::iterator, bool> 6782 Pair = Values.insert(std::make_pair(BB, DoesNotDominateBlock)); 6783 if (!Pair.second) 6784 return Pair.first->second; 6785 6786 BlockDisposition D = computeBlockDisposition(S, BB); 6787 return BlockDispositions[S][BB] = D; 6788} 6789 6790ScalarEvolution::BlockDisposition 6791ScalarEvolution::computeBlockDisposition(const SCEV *S, const BasicBlock *BB) { 6792 switch (S->getSCEVType()) { 6793 case scConstant: 6794 return ProperlyDominatesBlock; 6795 case scTruncate: 6796 case scZeroExtend: 6797 case scSignExtend: 6798 return getBlockDisposition(cast<SCEVCastExpr>(S)->getOperand(), BB); 6799 case scAddRecExpr: { 6800 // This uses a "dominates" query instead of "properly dominates" query 6801 // to test for proper dominance too, because the instruction which 6802 // produces the addrec's value is a PHI, and a PHI effectively properly 6803 // dominates its entire containing block. 6804 const SCEVAddRecExpr *AR = cast<SCEVAddRecExpr>(S); 6805 if (!DT->dominates(AR->getLoop()->getHeader(), BB)) 6806 return DoesNotDominateBlock; 6807 } 6808 // FALL THROUGH into SCEVNAryExpr handling. 6809 case scAddExpr: 6810 case scMulExpr: 6811 case scUMaxExpr: 6812 case scSMaxExpr: { 6813 const SCEVNAryExpr *NAry = cast<SCEVNAryExpr>(S); 6814 bool Proper = true; 6815 for (SCEVNAryExpr::op_iterator I = NAry->op_begin(), E = NAry->op_end(); 6816 I != E; ++I) { 6817 BlockDisposition D = getBlockDisposition(*I, BB); 6818 if (D == DoesNotDominateBlock) 6819 return DoesNotDominateBlock; 6820 if (D == DominatesBlock) 6821 Proper = false; 6822 } 6823 return Proper ? ProperlyDominatesBlock : DominatesBlock; 6824 } 6825 case scUDivExpr: { 6826 const SCEVUDivExpr *UDiv = cast<SCEVUDivExpr>(S); 6827 const SCEV *LHS = UDiv->getLHS(), *RHS = UDiv->getRHS(); 6828 BlockDisposition LD = getBlockDisposition(LHS, BB); 6829 if (LD == DoesNotDominateBlock) 6830 return DoesNotDominateBlock; 6831 BlockDisposition RD = getBlockDisposition(RHS, BB); 6832 if (RD == DoesNotDominateBlock) 6833 return DoesNotDominateBlock; 6834 return (LD == ProperlyDominatesBlock && RD == ProperlyDominatesBlock) ? 6835 ProperlyDominatesBlock : DominatesBlock; 6836 } 6837 case scUnknown: 6838 if (Instruction *I = 6839 dyn_cast<Instruction>(cast<SCEVUnknown>(S)->getValue())) { 6840 if (I->getParent() == BB) 6841 return DominatesBlock; 6842 if (DT->properlyDominates(I->getParent(), BB)) 6843 return ProperlyDominatesBlock; 6844 return DoesNotDominateBlock; 6845 } 6846 return ProperlyDominatesBlock; 6847 case scCouldNotCompute: 6848 llvm_unreachable("Attempt to use a SCEVCouldNotCompute object!"); 6849 default: 6850 llvm_unreachable("Unknown SCEV kind!"); 6851 } 6852} 6853 6854bool ScalarEvolution::dominates(const SCEV *S, const BasicBlock *BB) { 6855 return getBlockDisposition(S, BB) >= DominatesBlock; 6856} 6857 6858bool ScalarEvolution::properlyDominates(const SCEV *S, const BasicBlock *BB) { 6859 return getBlockDisposition(S, BB) == ProperlyDominatesBlock; 6860} 6861 6862bool ScalarEvolution::hasOperand(const SCEV *S, const SCEV *Op) const { 6863 switch (S->getSCEVType()) { 6864 case scConstant: 6865 return false; 6866 case scTruncate: 6867 case scZeroExtend: 6868 case scSignExtend: { 6869 const SCEVCastExpr *Cast = cast<SCEVCastExpr>(S); 6870 const SCEV *CastOp = Cast->getOperand(); 6871 return Op == CastOp || hasOperand(CastOp, Op); 6872 } 6873 case scAddRecExpr: 6874 case scAddExpr: 6875 case scMulExpr: 6876 case scUMaxExpr: 6877 case scSMaxExpr: { 6878 const SCEVNAryExpr *NAry = cast<SCEVNAryExpr>(S); 6879 for (SCEVNAryExpr::op_iterator I = NAry->op_begin(), E = NAry->op_end(); 6880 I != E; ++I) { 6881 const SCEV *NAryOp = *I; 6882 if (NAryOp == Op || hasOperand(NAryOp, Op)) 6883 return true; 6884 } 6885 return false; 6886 } 6887 case scUDivExpr: { 6888 const SCEVUDivExpr *UDiv = cast<SCEVUDivExpr>(S); 6889 const SCEV *LHS = UDiv->getLHS(), *RHS = UDiv->getRHS(); 6890 return LHS == Op || hasOperand(LHS, Op) || 6891 RHS == Op || hasOperand(RHS, Op); 6892 } 6893 case scUnknown: 6894 return false; 6895 case scCouldNotCompute: 6896 llvm_unreachable("Attempt to use a SCEVCouldNotCompute object!"); 6897 default: 6898 llvm_unreachable("Unknown SCEV kind!"); 6899 } 6900} 6901 6902void ScalarEvolution::forgetMemoizedResults(const SCEV *S) { 6903 ValuesAtScopes.erase(S); 6904 LoopDispositions.erase(S); 6905 BlockDispositions.erase(S); 6906 UnsignedRanges.erase(S); 6907 SignedRanges.erase(S); 6908} 6909