Reassociate.cpp revision 243830
1//===- Reassociate.cpp - Reassociate binary expressions -------------------===// 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 pass reassociates commutative expressions in an order that is designed 11// to promote better constant propagation, GCSE, LICM, PRE, etc. 12// 13// For example: 4 + (x + 5) -> x + (4 + 5) 14// 15// In the implementation of this algorithm, constants are assigned rank = 0, 16// function arguments are rank = 1, and other values are assigned ranks 17// corresponding to the reverse post order traversal of current function 18// (starting at 2), which effectively gives values in deep loops higher rank 19// than values not in loops. 20// 21//===----------------------------------------------------------------------===// 22 23#define DEBUG_TYPE "reassociate" 24#include "llvm/Transforms/Scalar.h" 25#include "llvm/Transforms/Utils/Local.h" 26#include "llvm/Constants.h" 27#include "llvm/DerivedTypes.h" 28#include "llvm/Function.h" 29#include "llvm/IRBuilder.h" 30#include "llvm/Instructions.h" 31#include "llvm/IntrinsicInst.h" 32#include "llvm/Pass.h" 33#include "llvm/ADT/DenseMap.h" 34#include "llvm/ADT/PostOrderIterator.h" 35#include "llvm/ADT/STLExtras.h" 36#include "llvm/ADT/SetVector.h" 37#include "llvm/ADT/Statistic.h" 38#include "llvm/Assembly/Writer.h" 39#include "llvm/Support/CFG.h" 40#include "llvm/Support/Debug.h" 41#include "llvm/Support/ValueHandle.h" 42#include "llvm/Support/raw_ostream.h" 43#include <algorithm> 44using namespace llvm; 45 46STATISTIC(NumChanged, "Number of insts reassociated"); 47STATISTIC(NumAnnihil, "Number of expr tree annihilated"); 48STATISTIC(NumFactor , "Number of multiplies factored"); 49 50namespace { 51 struct ValueEntry { 52 unsigned Rank; 53 Value *Op; 54 ValueEntry(unsigned R, Value *O) : Rank(R), Op(O) {} 55 }; 56 inline bool operator<(const ValueEntry &LHS, const ValueEntry &RHS) { 57 return LHS.Rank > RHS.Rank; // Sort so that highest rank goes to start. 58 } 59} 60 61#ifndef NDEBUG 62/// PrintOps - Print out the expression identified in the Ops list. 63/// 64static void PrintOps(Instruction *I, const SmallVectorImpl<ValueEntry> &Ops) { 65 Module *M = I->getParent()->getParent()->getParent(); 66 dbgs() << Instruction::getOpcodeName(I->getOpcode()) << " " 67 << *Ops[0].Op->getType() << '\t'; 68 for (unsigned i = 0, e = Ops.size(); i != e; ++i) { 69 dbgs() << "[ "; 70 WriteAsOperand(dbgs(), Ops[i].Op, false, M); 71 dbgs() << ", #" << Ops[i].Rank << "] "; 72 } 73} 74#endif 75 76namespace { 77 /// \brief Utility class representing a base and exponent pair which form one 78 /// factor of some product. 79 struct Factor { 80 Value *Base; 81 unsigned Power; 82 83 Factor(Value *Base, unsigned Power) : Base(Base), Power(Power) {} 84 85 /// \brief Sort factors by their Base. 86 struct BaseSorter { 87 bool operator()(const Factor &LHS, const Factor &RHS) { 88 return LHS.Base < RHS.Base; 89 } 90 }; 91 92 /// \brief Compare factors for equal bases. 93 struct BaseEqual { 94 bool operator()(const Factor &LHS, const Factor &RHS) { 95 return LHS.Base == RHS.Base; 96 } 97 }; 98 99 /// \brief Sort factors in descending order by their power. 100 struct PowerDescendingSorter { 101 bool operator()(const Factor &LHS, const Factor &RHS) { 102 return LHS.Power > RHS.Power; 103 } 104 }; 105 106 /// \brief Compare factors for equal powers. 107 struct PowerEqual { 108 bool operator()(const Factor &LHS, const Factor &RHS) { 109 return LHS.Power == RHS.Power; 110 } 111 }; 112 }; 113} 114 115namespace { 116 class Reassociate : public FunctionPass { 117 DenseMap<BasicBlock*, unsigned> RankMap; 118 DenseMap<AssertingVH<Value>, unsigned> ValueRankMap; 119 SetVector<AssertingVH<Instruction> > RedoInsts; 120 bool MadeChange; 121 public: 122 static char ID; // Pass identification, replacement for typeid 123 Reassociate() : FunctionPass(ID) { 124 initializeReassociatePass(*PassRegistry::getPassRegistry()); 125 } 126 127 bool runOnFunction(Function &F); 128 129 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 130 AU.setPreservesCFG(); 131 } 132 private: 133 void BuildRankMap(Function &F); 134 unsigned getRank(Value *V); 135 void ReassociateExpression(BinaryOperator *I); 136 void RewriteExprTree(BinaryOperator *I, SmallVectorImpl<ValueEntry> &Ops); 137 Value *OptimizeExpression(BinaryOperator *I, 138 SmallVectorImpl<ValueEntry> &Ops); 139 Value *OptimizeAdd(Instruction *I, SmallVectorImpl<ValueEntry> &Ops); 140 bool collectMultiplyFactors(SmallVectorImpl<ValueEntry> &Ops, 141 SmallVectorImpl<Factor> &Factors); 142 Value *buildMinimalMultiplyDAG(IRBuilder<> &Builder, 143 SmallVectorImpl<Factor> &Factors); 144 Value *OptimizeMul(BinaryOperator *I, SmallVectorImpl<ValueEntry> &Ops); 145 Value *RemoveFactorFromExpression(Value *V, Value *Factor); 146 void EraseInst(Instruction *I); 147 void OptimizeInst(Instruction *I); 148 }; 149} 150 151char Reassociate::ID = 0; 152INITIALIZE_PASS(Reassociate, "reassociate", 153 "Reassociate expressions", false, false) 154 155// Public interface to the Reassociate pass 156FunctionPass *llvm::createReassociatePass() { return new Reassociate(); } 157 158/// isReassociableOp - Return true if V is an instruction of the specified 159/// opcode and if it only has one use. 160static BinaryOperator *isReassociableOp(Value *V, unsigned Opcode) { 161 if (V->hasOneUse() && isa<Instruction>(V) && 162 cast<Instruction>(V)->getOpcode() == Opcode) 163 return cast<BinaryOperator>(V); 164 return 0; 165} 166 167static bool isUnmovableInstruction(Instruction *I) { 168 if (I->getOpcode() == Instruction::PHI || 169 I->getOpcode() == Instruction::LandingPad || 170 I->getOpcode() == Instruction::Alloca || 171 I->getOpcode() == Instruction::Load || 172 I->getOpcode() == Instruction::Invoke || 173 (I->getOpcode() == Instruction::Call && 174 !isa<DbgInfoIntrinsic>(I)) || 175 I->getOpcode() == Instruction::UDiv || 176 I->getOpcode() == Instruction::SDiv || 177 I->getOpcode() == Instruction::FDiv || 178 I->getOpcode() == Instruction::URem || 179 I->getOpcode() == Instruction::SRem || 180 I->getOpcode() == Instruction::FRem) 181 return true; 182 return false; 183} 184 185void Reassociate::BuildRankMap(Function &F) { 186 unsigned i = 2; 187 188 // Assign distinct ranks to function arguments 189 for (Function::arg_iterator I = F.arg_begin(), E = F.arg_end(); I != E; ++I) 190 ValueRankMap[&*I] = ++i; 191 192 ReversePostOrderTraversal<Function*> RPOT(&F); 193 for (ReversePostOrderTraversal<Function*>::rpo_iterator I = RPOT.begin(), 194 E = RPOT.end(); I != E; ++I) { 195 BasicBlock *BB = *I; 196 unsigned BBRank = RankMap[BB] = ++i << 16; 197 198 // Walk the basic block, adding precomputed ranks for any instructions that 199 // we cannot move. This ensures that the ranks for these instructions are 200 // all different in the block. 201 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) 202 if (isUnmovableInstruction(I)) 203 ValueRankMap[&*I] = ++BBRank; 204 } 205} 206 207unsigned Reassociate::getRank(Value *V) { 208 Instruction *I = dyn_cast<Instruction>(V); 209 if (I == 0) { 210 if (isa<Argument>(V)) return ValueRankMap[V]; // Function argument. 211 return 0; // Otherwise it's a global or constant, rank 0. 212 } 213 214 if (unsigned Rank = ValueRankMap[I]) 215 return Rank; // Rank already known? 216 217 // If this is an expression, return the 1+MAX(rank(LHS), rank(RHS)) so that 218 // we can reassociate expressions for code motion! Since we do not recurse 219 // for PHI nodes, we cannot have infinite recursion here, because there 220 // cannot be loops in the value graph that do not go through PHI nodes. 221 unsigned Rank = 0, MaxRank = RankMap[I->getParent()]; 222 for (unsigned i = 0, e = I->getNumOperands(); 223 i != e && Rank != MaxRank; ++i) 224 Rank = std::max(Rank, getRank(I->getOperand(i))); 225 226 // If this is a not or neg instruction, do not count it for rank. This 227 // assures us that X and ~X will have the same rank. 228 if (!I->getType()->isIntegerTy() || 229 (!BinaryOperator::isNot(I) && !BinaryOperator::isNeg(I))) 230 ++Rank; 231 232 //DEBUG(dbgs() << "Calculated Rank[" << V->getName() << "] = " 233 // << Rank << "\n"); 234 235 return ValueRankMap[I] = Rank; 236} 237 238/// LowerNegateToMultiply - Replace 0-X with X*-1. 239/// 240static BinaryOperator *LowerNegateToMultiply(Instruction *Neg) { 241 Constant *Cst = Constant::getAllOnesValue(Neg->getType()); 242 243 BinaryOperator *Res = 244 BinaryOperator::CreateMul(Neg->getOperand(1), Cst, "",Neg); 245 Neg->setOperand(1, Constant::getNullValue(Neg->getType())); // Drop use of op. 246 Res->takeName(Neg); 247 Neg->replaceAllUsesWith(Res); 248 Res->setDebugLoc(Neg->getDebugLoc()); 249 return Res; 250} 251 252/// CarmichaelShift - Returns k such that lambda(2^Bitwidth) = 2^k, where lambda 253/// is the Carmichael function. This means that x^(2^k) === 1 mod 2^Bitwidth for 254/// every odd x, i.e. x^(2^k) = 1 for every odd x in Bitwidth-bit arithmetic. 255/// Note that 0 <= k < Bitwidth, and if Bitwidth > 3 then x^(2^k) = 0 for every 256/// even x in Bitwidth-bit arithmetic. 257static unsigned CarmichaelShift(unsigned Bitwidth) { 258 if (Bitwidth < 3) 259 return Bitwidth - 1; 260 return Bitwidth - 2; 261} 262 263/// IncorporateWeight - Add the extra weight 'RHS' to the existing weight 'LHS', 264/// reducing the combined weight using any special properties of the operation. 265/// The existing weight LHS represents the computation X op X op ... op X where 266/// X occurs LHS times. The combined weight represents X op X op ... op X with 267/// X occurring LHS + RHS times. If op is "Xor" for example then the combined 268/// operation is equivalent to X if LHS + RHS is odd, or 0 if LHS + RHS is even; 269/// the routine returns 1 in LHS in the first case, and 0 in LHS in the second. 270static void IncorporateWeight(APInt &LHS, const APInt &RHS, unsigned Opcode) { 271 // If we were working with infinite precision arithmetic then the combined 272 // weight would be LHS + RHS. But we are using finite precision arithmetic, 273 // and the APInt sum LHS + RHS may not be correct if it wraps (it is correct 274 // for nilpotent operations and addition, but not for idempotent operations 275 // and multiplication), so it is important to correctly reduce the combined 276 // weight back into range if wrapping would be wrong. 277 278 // If RHS is zero then the weight didn't change. 279 if (RHS.isMinValue()) 280 return; 281 // If LHS is zero then the combined weight is RHS. 282 if (LHS.isMinValue()) { 283 LHS = RHS; 284 return; 285 } 286 // From this point on we know that neither LHS nor RHS is zero. 287 288 if (Instruction::isIdempotent(Opcode)) { 289 // Idempotent means X op X === X, so any non-zero weight is equivalent to a 290 // weight of 1. Keeping weights at zero or one also means that wrapping is 291 // not a problem. 292 assert(LHS == 1 && RHS == 1 && "Weights not reduced!"); 293 return; // Return a weight of 1. 294 } 295 if (Instruction::isNilpotent(Opcode)) { 296 // Nilpotent means X op X === 0, so reduce weights modulo 2. 297 assert(LHS == 1 && RHS == 1 && "Weights not reduced!"); 298 LHS = 0; // 1 + 1 === 0 modulo 2. 299 return; 300 } 301 if (Opcode == Instruction::Add) { 302 // TODO: Reduce the weight by exploiting nsw/nuw? 303 LHS += RHS; 304 return; 305 } 306 307 assert(Opcode == Instruction::Mul && "Unknown associative operation!"); 308 unsigned Bitwidth = LHS.getBitWidth(); 309 // If CM is the Carmichael number then a weight W satisfying W >= CM+Bitwidth 310 // can be replaced with W-CM. That's because x^W=x^(W-CM) for every Bitwidth 311 // bit number x, since either x is odd in which case x^CM = 1, or x is even in 312 // which case both x^W and x^(W - CM) are zero. By subtracting off multiples 313 // of CM like this weights can always be reduced to the range [0, CM+Bitwidth) 314 // which by a happy accident means that they can always be represented using 315 // Bitwidth bits. 316 // TODO: Reduce the weight by exploiting nsw/nuw? (Could do much better than 317 // the Carmichael number). 318 if (Bitwidth > 3) { 319 /// CM - The value of Carmichael's lambda function. 320 APInt CM = APInt::getOneBitSet(Bitwidth, CarmichaelShift(Bitwidth)); 321 // Any weight W >= Threshold can be replaced with W - CM. 322 APInt Threshold = CM + Bitwidth; 323 assert(LHS.ult(Threshold) && RHS.ult(Threshold) && "Weights not reduced!"); 324 // For Bitwidth 4 or more the following sum does not overflow. 325 LHS += RHS; 326 while (LHS.uge(Threshold)) 327 LHS -= CM; 328 } else { 329 // To avoid problems with overflow do everything the same as above but using 330 // a larger type. 331 unsigned CM = 1U << CarmichaelShift(Bitwidth); 332 unsigned Threshold = CM + Bitwidth; 333 assert(LHS.getZExtValue() < Threshold && RHS.getZExtValue() < Threshold && 334 "Weights not reduced!"); 335 unsigned Total = LHS.getZExtValue() + RHS.getZExtValue(); 336 while (Total >= Threshold) 337 Total -= CM; 338 LHS = Total; 339 } 340} 341 342typedef std::pair<Value*, APInt> RepeatedValue; 343 344/// LinearizeExprTree - Given an associative binary expression, return the leaf 345/// nodes in Ops along with their weights (how many times the leaf occurs). The 346/// original expression is the same as 347/// (Ops[0].first op Ops[0].first op ... Ops[0].first) <- Ops[0].second times 348/// op 349/// (Ops[1].first op Ops[1].first op ... Ops[1].first) <- Ops[1].second times 350/// op 351/// ... 352/// op 353/// (Ops[N].first op Ops[N].first op ... Ops[N].first) <- Ops[N].second times 354/// 355/// Note that the values Ops[0].first, ..., Ops[N].first are all distinct. 356/// 357/// This routine may modify the function, in which case it returns 'true'. The 358/// changes it makes may well be destructive, changing the value computed by 'I' 359/// to something completely different. Thus if the routine returns 'true' then 360/// you MUST either replace I with a new expression computed from the Ops array, 361/// or use RewriteExprTree to put the values back in. 362/// 363/// A leaf node is either not a binary operation of the same kind as the root 364/// node 'I' (i.e. is not a binary operator at all, or is, but with a different 365/// opcode), or is the same kind of binary operator but has a use which either 366/// does not belong to the expression, or does belong to the expression but is 367/// a leaf node. Every leaf node has at least one use that is a non-leaf node 368/// of the expression, while for non-leaf nodes (except for the root 'I') every 369/// use is a non-leaf node of the expression. 370/// 371/// For example: 372/// expression graph node names 373/// 374/// + | I 375/// / \ | 376/// + + | A, B 377/// / \ / \ | 378/// * + * | C, D, E 379/// / \ / \ / \ | 380/// + * | F, G 381/// 382/// The leaf nodes are C, E, F and G. The Ops array will contain (maybe not in 383/// that order) (C, 1), (E, 1), (F, 2), (G, 2). 384/// 385/// The expression is maximal: if some instruction is a binary operator of the 386/// same kind as 'I', and all of its uses are non-leaf nodes of the expression, 387/// then the instruction also belongs to the expression, is not a leaf node of 388/// it, and its operands also belong to the expression (but may be leaf nodes). 389/// 390/// NOTE: This routine will set operands of non-leaf non-root nodes to undef in 391/// order to ensure that every non-root node in the expression has *exactly one* 392/// use by a non-leaf node of the expression. This destruction means that the 393/// caller MUST either replace 'I' with a new expression or use something like 394/// RewriteExprTree to put the values back in if the routine indicates that it 395/// made a change by returning 'true'. 396/// 397/// In the above example either the right operand of A or the left operand of B 398/// will be replaced by undef. If it is B's operand then this gives: 399/// 400/// + | I 401/// / \ | 402/// + + | A, B - operand of B replaced with undef 403/// / \ \ | 404/// * + * | C, D, E 405/// / \ / \ / \ | 406/// + * | F, G 407/// 408/// Note that such undef operands can only be reached by passing through 'I'. 409/// For example, if you visit operands recursively starting from a leaf node 410/// then you will never see such an undef operand unless you get back to 'I', 411/// which requires passing through a phi node. 412/// 413/// Note that this routine may also mutate binary operators of the wrong type 414/// that have all uses inside the expression (i.e. only used by non-leaf nodes 415/// of the expression) if it can turn them into binary operators of the right 416/// type and thus make the expression bigger. 417 418static bool LinearizeExprTree(BinaryOperator *I, 419 SmallVectorImpl<RepeatedValue> &Ops) { 420 DEBUG(dbgs() << "LINEARIZE: " << *I << '\n'); 421 unsigned Bitwidth = I->getType()->getScalarType()->getPrimitiveSizeInBits(); 422 unsigned Opcode = I->getOpcode(); 423 assert(Instruction::isAssociative(Opcode) && 424 Instruction::isCommutative(Opcode) && 425 "Expected an associative and commutative operation!"); 426 // If we see an absorbing element then the entire expression must be equal to 427 // it. For example, if this is a multiplication expression and zero occurs as 428 // an operand somewhere in it then the result of the expression must be zero. 429 Constant *Absorber = ConstantExpr::getBinOpAbsorber(Opcode, I->getType()); 430 431 // Visit all operands of the expression, keeping track of their weight (the 432 // number of paths from the expression root to the operand, or if you like 433 // the number of times that operand occurs in the linearized expression). 434 // For example, if I = X + A, where X = A + B, then I, X and B have weight 1 435 // while A has weight two. 436 437 // Worklist of non-leaf nodes (their operands are in the expression too) along 438 // with their weights, representing a certain number of paths to the operator. 439 // If an operator occurs in the worklist multiple times then we found multiple 440 // ways to get to it. 441 SmallVector<std::pair<BinaryOperator*, APInt>, 8> Worklist; // (Op, Weight) 442 Worklist.push_back(std::make_pair(I, APInt(Bitwidth, 1))); 443 bool MadeChange = false; 444 445 // Leaves of the expression are values that either aren't the right kind of 446 // operation (eg: a constant, or a multiply in an add tree), or are, but have 447 // some uses that are not inside the expression. For example, in I = X + X, 448 // X = A + B, the value X has two uses (by I) that are in the expression. If 449 // X has any other uses, for example in a return instruction, then we consider 450 // X to be a leaf, and won't analyze it further. When we first visit a value, 451 // if it has more than one use then at first we conservatively consider it to 452 // be a leaf. Later, as the expression is explored, we may discover some more 453 // uses of the value from inside the expression. If all uses turn out to be 454 // from within the expression (and the value is a binary operator of the right 455 // kind) then the value is no longer considered to be a leaf, and its operands 456 // are explored. 457 458 // Leaves - Keeps track of the set of putative leaves as well as the number of 459 // paths to each leaf seen so far. 460 typedef DenseMap<Value*, APInt> LeafMap; 461 LeafMap Leaves; // Leaf -> Total weight so far. 462 SmallVector<Value*, 8> LeafOrder; // Ensure deterministic leaf output order. 463 464#ifndef NDEBUG 465 SmallPtrSet<Value*, 8> Visited; // For sanity checking the iteration scheme. 466#endif 467 while (!Worklist.empty()) { 468 std::pair<BinaryOperator*, APInt> P = Worklist.pop_back_val(); 469 I = P.first; // We examine the operands of this binary operator. 470 471 for (unsigned OpIdx = 0; OpIdx < 2; ++OpIdx) { // Visit operands. 472 Value *Op = I->getOperand(OpIdx); 473 APInt Weight = P.second; // Number of paths to this operand. 474 DEBUG(dbgs() << "OPERAND: " << *Op << " (" << Weight << ")\n"); 475 assert(!Op->use_empty() && "No uses, so how did we get to it?!"); 476 477 // If the expression contains an absorbing element then there is no need 478 // to analyze it further: it must evaluate to the absorbing element. 479 if (Op == Absorber && !Weight.isMinValue()) { 480 Ops.push_back(std::make_pair(Absorber, APInt(Bitwidth, 1))); 481 return MadeChange; 482 } 483 484 // If this is a binary operation of the right kind with only one use then 485 // add its operands to the expression. 486 if (BinaryOperator *BO = isReassociableOp(Op, Opcode)) { 487 assert(Visited.insert(Op) && "Not first visit!"); 488 DEBUG(dbgs() << "DIRECT ADD: " << *Op << " (" << Weight << ")\n"); 489 Worklist.push_back(std::make_pair(BO, Weight)); 490 continue; 491 } 492 493 // Appears to be a leaf. Is the operand already in the set of leaves? 494 LeafMap::iterator It = Leaves.find(Op); 495 if (It == Leaves.end()) { 496 // Not in the leaf map. Must be the first time we saw this operand. 497 assert(Visited.insert(Op) && "Not first visit!"); 498 if (!Op->hasOneUse()) { 499 // This value has uses not accounted for by the expression, so it is 500 // not safe to modify. Mark it as being a leaf. 501 DEBUG(dbgs() << "ADD USES LEAF: " << *Op << " (" << Weight << ")\n"); 502 LeafOrder.push_back(Op); 503 Leaves[Op] = Weight; 504 continue; 505 } 506 // No uses outside the expression, try morphing it. 507 } else if (It != Leaves.end()) { 508 // Already in the leaf map. 509 assert(Visited.count(Op) && "In leaf map but not visited!"); 510 511 // Update the number of paths to the leaf. 512 IncorporateWeight(It->second, Weight, Opcode); 513 514#if 0 // TODO: Re-enable once PR13021 is fixed. 515 // The leaf already has one use from inside the expression. As we want 516 // exactly one such use, drop this new use of the leaf. 517 assert(!Op->hasOneUse() && "Only one use, but we got here twice!"); 518 I->setOperand(OpIdx, UndefValue::get(I->getType())); 519 MadeChange = true; 520 521 // If the leaf is a binary operation of the right kind and we now see 522 // that its multiple original uses were in fact all by nodes belonging 523 // to the expression, then no longer consider it to be a leaf and add 524 // its operands to the expression. 525 if (BinaryOperator *BO = isReassociableOp(Op, Opcode)) { 526 DEBUG(dbgs() << "UNLEAF: " << *Op << " (" << It->second << ")\n"); 527 Worklist.push_back(std::make_pair(BO, It->second)); 528 Leaves.erase(It); 529 continue; 530 } 531#endif 532 533 // If we still have uses that are not accounted for by the expression 534 // then it is not safe to modify the value. 535 if (!Op->hasOneUse()) 536 continue; 537 538 // No uses outside the expression, try morphing it. 539 Weight = It->second; 540 Leaves.erase(It); // Since the value may be morphed below. 541 } 542 543 // At this point we have a value which, first of all, is not a binary 544 // expression of the right kind, and secondly, is only used inside the 545 // expression. This means that it can safely be modified. See if we 546 // can usefully morph it into an expression of the right kind. 547 assert((!isa<Instruction>(Op) || 548 cast<Instruction>(Op)->getOpcode() != Opcode) && 549 "Should have been handled above!"); 550 assert(Op->hasOneUse() && "Has uses outside the expression tree!"); 551 552 // If this is a multiply expression, turn any internal negations into 553 // multiplies by -1 so they can be reassociated. 554 BinaryOperator *BO = dyn_cast<BinaryOperator>(Op); 555 if (Opcode == Instruction::Mul && BO && BinaryOperator::isNeg(BO)) { 556 DEBUG(dbgs() << "MORPH LEAF: " << *Op << " (" << Weight << ") TO "); 557 BO = LowerNegateToMultiply(BO); 558 DEBUG(dbgs() << *BO << 'n'); 559 Worklist.push_back(std::make_pair(BO, Weight)); 560 MadeChange = true; 561 continue; 562 } 563 564 // Failed to morph into an expression of the right type. This really is 565 // a leaf. 566 DEBUG(dbgs() << "ADD LEAF: " << *Op << " (" << Weight << ")\n"); 567 assert(!isReassociableOp(Op, Opcode) && "Value was morphed?"); 568 LeafOrder.push_back(Op); 569 Leaves[Op] = Weight; 570 } 571 } 572 573 // The leaves, repeated according to their weights, represent the linearized 574 // form of the expression. 575 for (unsigned i = 0, e = LeafOrder.size(); i != e; ++i) { 576 Value *V = LeafOrder[i]; 577 LeafMap::iterator It = Leaves.find(V); 578 if (It == Leaves.end()) 579 // Node initially thought to be a leaf wasn't. 580 continue; 581 assert(!isReassociableOp(V, Opcode) && "Shouldn't be a leaf!"); 582 APInt Weight = It->second; 583 if (Weight.isMinValue()) 584 // Leaf already output or weight reduction eliminated it. 585 continue; 586 // Ensure the leaf is only output once. 587 It->second = 0; 588 Ops.push_back(std::make_pair(V, Weight)); 589 } 590 591 // For nilpotent operations or addition there may be no operands, for example 592 // because the expression was "X xor X" or consisted of 2^Bitwidth additions: 593 // in both cases the weight reduces to 0 causing the value to be skipped. 594 if (Ops.empty()) { 595 Constant *Identity = ConstantExpr::getBinOpIdentity(Opcode, I->getType()); 596 assert(Identity && "Associative operation without identity!"); 597 Ops.push_back(std::make_pair(Identity, APInt(Bitwidth, 1))); 598 } 599 600 return MadeChange; 601} 602 603// RewriteExprTree - Now that the operands for this expression tree are 604// linearized and optimized, emit them in-order. 605void Reassociate::RewriteExprTree(BinaryOperator *I, 606 SmallVectorImpl<ValueEntry> &Ops) { 607 assert(Ops.size() > 1 && "Single values should be used directly!"); 608 609 // Since our optimizations should never increase the number of operations, the 610 // new expression can usually be written reusing the existing binary operators 611 // from the original expression tree, without creating any new instructions, 612 // though the rewritten expression may have a completely different topology. 613 // We take care to not change anything if the new expression will be the same 614 // as the original. If more than trivial changes (like commuting operands) 615 // were made then we are obliged to clear out any optional subclass data like 616 // nsw flags. 617 618 /// NodesToRewrite - Nodes from the original expression available for writing 619 /// the new expression into. 620 SmallVector<BinaryOperator*, 8> NodesToRewrite; 621 unsigned Opcode = I->getOpcode(); 622 BinaryOperator *Op = I; 623 624 /// NotRewritable - The operands being written will be the leaves of the new 625 /// expression and must not be used as inner nodes (via NodesToRewrite) by 626 /// mistake. Inner nodes are always reassociable, and usually leaves are not 627 /// (if they were they would have been incorporated into the expression and so 628 /// would not be leaves), so most of the time there is no danger of this. But 629 /// in rare cases a leaf may become reassociable if an optimization kills uses 630 /// of it, or it may momentarily become reassociable during rewriting (below) 631 /// due it being removed as an operand of one of its uses. Ensure that misuse 632 /// of leaf nodes as inner nodes cannot occur by remembering all of the future 633 /// leaves and refusing to reuse any of them as inner nodes. 634 SmallPtrSet<Value*, 8> NotRewritable; 635 for (unsigned i = 0, e = Ops.size(); i != e; ++i) 636 NotRewritable.insert(Ops[i].Op); 637 638 // ExpressionChanged - Non-null if the rewritten expression differs from the 639 // original in some non-trivial way, requiring the clearing of optional flags. 640 // Flags are cleared from the operator in ExpressionChanged up to I inclusive. 641 BinaryOperator *ExpressionChanged = 0; 642 for (unsigned i = 0; ; ++i) { 643 // The last operation (which comes earliest in the IR) is special as both 644 // operands will come from Ops, rather than just one with the other being 645 // a subexpression. 646 if (i+2 == Ops.size()) { 647 Value *NewLHS = Ops[i].Op; 648 Value *NewRHS = Ops[i+1].Op; 649 Value *OldLHS = Op->getOperand(0); 650 Value *OldRHS = Op->getOperand(1); 651 652 if (NewLHS == OldLHS && NewRHS == OldRHS) 653 // Nothing changed, leave it alone. 654 break; 655 656 if (NewLHS == OldRHS && NewRHS == OldLHS) { 657 // The order of the operands was reversed. Swap them. 658 DEBUG(dbgs() << "RA: " << *Op << '\n'); 659 Op->swapOperands(); 660 DEBUG(dbgs() << "TO: " << *Op << '\n'); 661 MadeChange = true; 662 ++NumChanged; 663 break; 664 } 665 666 // The new operation differs non-trivially from the original. Overwrite 667 // the old operands with the new ones. 668 DEBUG(dbgs() << "RA: " << *Op << '\n'); 669 if (NewLHS != OldLHS) { 670 BinaryOperator *BO = isReassociableOp(OldLHS, Opcode); 671 if (BO && !NotRewritable.count(BO)) 672 NodesToRewrite.push_back(BO); 673 Op->setOperand(0, NewLHS); 674 } 675 if (NewRHS != OldRHS) { 676 BinaryOperator *BO = isReassociableOp(OldRHS, Opcode); 677 if (BO && !NotRewritable.count(BO)) 678 NodesToRewrite.push_back(BO); 679 Op->setOperand(1, NewRHS); 680 } 681 DEBUG(dbgs() << "TO: " << *Op << '\n'); 682 683 ExpressionChanged = Op; 684 MadeChange = true; 685 ++NumChanged; 686 687 break; 688 } 689 690 // Not the last operation. The left-hand side will be a sub-expression 691 // while the right-hand side will be the current element of Ops. 692 Value *NewRHS = Ops[i].Op; 693 if (NewRHS != Op->getOperand(1)) { 694 DEBUG(dbgs() << "RA: " << *Op << '\n'); 695 if (NewRHS == Op->getOperand(0)) { 696 // The new right-hand side was already present as the left operand. If 697 // we are lucky then swapping the operands will sort out both of them. 698 Op->swapOperands(); 699 } else { 700 // Overwrite with the new right-hand side. 701 BinaryOperator *BO = isReassociableOp(Op->getOperand(1), Opcode); 702 if (BO && !NotRewritable.count(BO)) 703 NodesToRewrite.push_back(BO); 704 Op->setOperand(1, NewRHS); 705 ExpressionChanged = Op; 706 } 707 DEBUG(dbgs() << "TO: " << *Op << '\n'); 708 MadeChange = true; 709 ++NumChanged; 710 } 711 712 // Now deal with the left-hand side. If this is already an operation node 713 // from the original expression then just rewrite the rest of the expression 714 // into it. 715 BinaryOperator *BO = isReassociableOp(Op->getOperand(0), Opcode); 716 if (BO && !NotRewritable.count(BO)) { 717 Op = BO; 718 continue; 719 } 720 721 // Otherwise, grab a spare node from the original expression and use that as 722 // the left-hand side. If there are no nodes left then the optimizers made 723 // an expression with more nodes than the original! This usually means that 724 // they did something stupid but it might mean that the problem was just too 725 // hard (finding the mimimal number of multiplications needed to realize a 726 // multiplication expression is NP-complete). Whatever the reason, smart or 727 // stupid, create a new node if there are none left. 728 BinaryOperator *NewOp; 729 if (NodesToRewrite.empty()) { 730 Constant *Undef = UndefValue::get(I->getType()); 731 NewOp = BinaryOperator::Create(Instruction::BinaryOps(Opcode), 732 Undef, Undef, "", I); 733 } else { 734 NewOp = NodesToRewrite.pop_back_val(); 735 } 736 737 DEBUG(dbgs() << "RA: " << *Op << '\n'); 738 Op->setOperand(0, NewOp); 739 DEBUG(dbgs() << "TO: " << *Op << '\n'); 740 ExpressionChanged = Op; 741 MadeChange = true; 742 ++NumChanged; 743 Op = NewOp; 744 } 745 746 // If the expression changed non-trivially then clear out all subclass data 747 // starting from the operator specified in ExpressionChanged, and compactify 748 // the operators to just before the expression root to guarantee that the 749 // expression tree is dominated by all of Ops. 750 if (ExpressionChanged) 751 do { 752 ExpressionChanged->clearSubclassOptionalData(); 753 if (ExpressionChanged == I) 754 break; 755 ExpressionChanged->moveBefore(I); 756 ExpressionChanged = cast<BinaryOperator>(*ExpressionChanged->use_begin()); 757 } while (1); 758 759 // Throw away any left over nodes from the original expression. 760 for (unsigned i = 0, e = NodesToRewrite.size(); i != e; ++i) 761 RedoInsts.insert(NodesToRewrite[i]); 762} 763 764/// NegateValue - Insert instructions before the instruction pointed to by BI, 765/// that computes the negative version of the value specified. The negative 766/// version of the value is returned, and BI is left pointing at the instruction 767/// that should be processed next by the reassociation pass. 768static Value *NegateValue(Value *V, Instruction *BI) { 769 if (Constant *C = dyn_cast<Constant>(V)) 770 return ConstantExpr::getNeg(C); 771 772 // We are trying to expose opportunity for reassociation. One of the things 773 // that we want to do to achieve this is to push a negation as deep into an 774 // expression chain as possible, to expose the add instructions. In practice, 775 // this means that we turn this: 776 // X = -(A+12+C+D) into X = -A + -12 + -C + -D = -12 + -A + -C + -D 777 // so that later, a: Y = 12+X could get reassociated with the -12 to eliminate 778 // the constants. We assume that instcombine will clean up the mess later if 779 // we introduce tons of unnecessary negation instructions. 780 // 781 if (BinaryOperator *I = isReassociableOp(V, Instruction::Add)) { 782 // Push the negates through the add. 783 I->setOperand(0, NegateValue(I->getOperand(0), BI)); 784 I->setOperand(1, NegateValue(I->getOperand(1), BI)); 785 786 // We must move the add instruction here, because the neg instructions do 787 // not dominate the old add instruction in general. By moving it, we are 788 // assured that the neg instructions we just inserted dominate the 789 // instruction we are about to insert after them. 790 // 791 I->moveBefore(BI); 792 I->setName(I->getName()+".neg"); 793 return I; 794 } 795 796 // Okay, we need to materialize a negated version of V with an instruction. 797 // Scan the use lists of V to see if we have one already. 798 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E;++UI){ 799 User *U = *UI; 800 if (!BinaryOperator::isNeg(U)) continue; 801 802 // We found one! Now we have to make sure that the definition dominates 803 // this use. We do this by moving it to the entry block (if it is a 804 // non-instruction value) or right after the definition. These negates will 805 // be zapped by reassociate later, so we don't need much finesse here. 806 BinaryOperator *TheNeg = cast<BinaryOperator>(U); 807 808 // Verify that the negate is in this function, V might be a constant expr. 809 if (TheNeg->getParent()->getParent() != BI->getParent()->getParent()) 810 continue; 811 812 BasicBlock::iterator InsertPt; 813 if (Instruction *InstInput = dyn_cast<Instruction>(V)) { 814 if (InvokeInst *II = dyn_cast<InvokeInst>(InstInput)) { 815 InsertPt = II->getNormalDest()->begin(); 816 } else { 817 InsertPt = InstInput; 818 ++InsertPt; 819 } 820 while (isa<PHINode>(InsertPt)) ++InsertPt; 821 } else { 822 InsertPt = TheNeg->getParent()->getParent()->getEntryBlock().begin(); 823 } 824 TheNeg->moveBefore(InsertPt); 825 return TheNeg; 826 } 827 828 // Insert a 'neg' instruction that subtracts the value from zero to get the 829 // negation. 830 return BinaryOperator::CreateNeg(V, V->getName() + ".neg", BI); 831} 832 833/// ShouldBreakUpSubtract - Return true if we should break up this subtract of 834/// X-Y into (X + -Y). 835static bool ShouldBreakUpSubtract(Instruction *Sub) { 836 // If this is a negation, we can't split it up! 837 if (BinaryOperator::isNeg(Sub)) 838 return false; 839 840 // Don't bother to break this up unless either the LHS is an associable add or 841 // subtract or if this is only used by one. 842 if (isReassociableOp(Sub->getOperand(0), Instruction::Add) || 843 isReassociableOp(Sub->getOperand(0), Instruction::Sub)) 844 return true; 845 if (isReassociableOp(Sub->getOperand(1), Instruction::Add) || 846 isReassociableOp(Sub->getOperand(1), Instruction::Sub)) 847 return true; 848 if (Sub->hasOneUse() && 849 (isReassociableOp(Sub->use_back(), Instruction::Add) || 850 isReassociableOp(Sub->use_back(), Instruction::Sub))) 851 return true; 852 853 return false; 854} 855 856/// BreakUpSubtract - If we have (X-Y), and if either X is an add, or if this is 857/// only used by an add, transform this into (X+(0-Y)) to promote better 858/// reassociation. 859static BinaryOperator *BreakUpSubtract(Instruction *Sub) { 860 // Convert a subtract into an add and a neg instruction. This allows sub 861 // instructions to be commuted with other add instructions. 862 // 863 // Calculate the negative value of Operand 1 of the sub instruction, 864 // and set it as the RHS of the add instruction we just made. 865 // 866 Value *NegVal = NegateValue(Sub->getOperand(1), Sub); 867 BinaryOperator *New = 868 BinaryOperator::CreateAdd(Sub->getOperand(0), NegVal, "", Sub); 869 Sub->setOperand(0, Constant::getNullValue(Sub->getType())); // Drop use of op. 870 Sub->setOperand(1, Constant::getNullValue(Sub->getType())); // Drop use of op. 871 New->takeName(Sub); 872 873 // Everyone now refers to the add instruction. 874 Sub->replaceAllUsesWith(New); 875 New->setDebugLoc(Sub->getDebugLoc()); 876 877 DEBUG(dbgs() << "Negated: " << *New << '\n'); 878 return New; 879} 880 881/// ConvertShiftToMul - If this is a shift of a reassociable multiply or is used 882/// by one, change this into a multiply by a constant to assist with further 883/// reassociation. 884static BinaryOperator *ConvertShiftToMul(Instruction *Shl) { 885 Constant *MulCst = ConstantInt::get(Shl->getType(), 1); 886 MulCst = ConstantExpr::getShl(MulCst, cast<Constant>(Shl->getOperand(1))); 887 888 BinaryOperator *Mul = 889 BinaryOperator::CreateMul(Shl->getOperand(0), MulCst, "", Shl); 890 Shl->setOperand(0, UndefValue::get(Shl->getType())); // Drop use of op. 891 Mul->takeName(Shl); 892 Shl->replaceAllUsesWith(Mul); 893 Mul->setDebugLoc(Shl->getDebugLoc()); 894 return Mul; 895} 896 897/// FindInOperandList - Scan backwards and forwards among values with the same 898/// rank as element i to see if X exists. If X does not exist, return i. This 899/// is useful when scanning for 'x' when we see '-x' because they both get the 900/// same rank. 901static unsigned FindInOperandList(SmallVectorImpl<ValueEntry> &Ops, unsigned i, 902 Value *X) { 903 unsigned XRank = Ops[i].Rank; 904 unsigned e = Ops.size(); 905 for (unsigned j = i+1; j != e && Ops[j].Rank == XRank; ++j) 906 if (Ops[j].Op == X) 907 return j; 908 // Scan backwards. 909 for (unsigned j = i-1; j != ~0U && Ops[j].Rank == XRank; --j) 910 if (Ops[j].Op == X) 911 return j; 912 return i; 913} 914 915/// EmitAddTreeOfValues - Emit a tree of add instructions, summing Ops together 916/// and returning the result. Insert the tree before I. 917static Value *EmitAddTreeOfValues(Instruction *I, 918 SmallVectorImpl<WeakVH> &Ops){ 919 if (Ops.size() == 1) return Ops.back(); 920 921 Value *V1 = Ops.back(); 922 Ops.pop_back(); 923 Value *V2 = EmitAddTreeOfValues(I, Ops); 924 return BinaryOperator::CreateAdd(V2, V1, "tmp", I); 925} 926 927/// RemoveFactorFromExpression - If V is an expression tree that is a 928/// multiplication sequence, and if this sequence contains a multiply by Factor, 929/// remove Factor from the tree and return the new tree. 930Value *Reassociate::RemoveFactorFromExpression(Value *V, Value *Factor) { 931 BinaryOperator *BO = isReassociableOp(V, Instruction::Mul); 932 if (!BO) return 0; 933 934 SmallVector<RepeatedValue, 8> Tree; 935 MadeChange |= LinearizeExprTree(BO, Tree); 936 SmallVector<ValueEntry, 8> Factors; 937 Factors.reserve(Tree.size()); 938 for (unsigned i = 0, e = Tree.size(); i != e; ++i) { 939 RepeatedValue E = Tree[i]; 940 Factors.append(E.second.getZExtValue(), 941 ValueEntry(getRank(E.first), E.first)); 942 } 943 944 bool FoundFactor = false; 945 bool NeedsNegate = false; 946 for (unsigned i = 0, e = Factors.size(); i != e; ++i) { 947 if (Factors[i].Op == Factor) { 948 FoundFactor = true; 949 Factors.erase(Factors.begin()+i); 950 break; 951 } 952 953 // If this is a negative version of this factor, remove it. 954 if (ConstantInt *FC1 = dyn_cast<ConstantInt>(Factor)) 955 if (ConstantInt *FC2 = dyn_cast<ConstantInt>(Factors[i].Op)) 956 if (FC1->getValue() == -FC2->getValue()) { 957 FoundFactor = NeedsNegate = true; 958 Factors.erase(Factors.begin()+i); 959 break; 960 } 961 } 962 963 if (!FoundFactor) { 964 // Make sure to restore the operands to the expression tree. 965 RewriteExprTree(BO, Factors); 966 return 0; 967 } 968 969 BasicBlock::iterator InsertPt = BO; ++InsertPt; 970 971 // If this was just a single multiply, remove the multiply and return the only 972 // remaining operand. 973 if (Factors.size() == 1) { 974 RedoInsts.insert(BO); 975 V = Factors[0].Op; 976 } else { 977 RewriteExprTree(BO, Factors); 978 V = BO; 979 } 980 981 if (NeedsNegate) 982 V = BinaryOperator::CreateNeg(V, "neg", InsertPt); 983 984 return V; 985} 986 987/// FindSingleUseMultiplyFactors - If V is a single-use multiply, recursively 988/// add its operands as factors, otherwise add V to the list of factors. 989/// 990/// Ops is the top-level list of add operands we're trying to factor. 991static void FindSingleUseMultiplyFactors(Value *V, 992 SmallVectorImpl<Value*> &Factors, 993 const SmallVectorImpl<ValueEntry> &Ops) { 994 BinaryOperator *BO = isReassociableOp(V, Instruction::Mul); 995 if (!BO) { 996 Factors.push_back(V); 997 return; 998 } 999 1000 // Otherwise, add the LHS and RHS to the list of factors. 1001 FindSingleUseMultiplyFactors(BO->getOperand(1), Factors, Ops); 1002 FindSingleUseMultiplyFactors(BO->getOperand(0), Factors, Ops); 1003} 1004 1005/// OptimizeAndOrXor - Optimize a series of operands to an 'and', 'or', or 'xor' 1006/// instruction. This optimizes based on identities. If it can be reduced to 1007/// a single Value, it is returned, otherwise the Ops list is mutated as 1008/// necessary. 1009static Value *OptimizeAndOrXor(unsigned Opcode, 1010 SmallVectorImpl<ValueEntry> &Ops) { 1011 // Scan the operand lists looking for X and ~X pairs, along with X,X pairs. 1012 // If we find any, we can simplify the expression. X&~X == 0, X|~X == -1. 1013 for (unsigned i = 0, e = Ops.size(); i != e; ++i) { 1014 // First, check for X and ~X in the operand list. 1015 assert(i < Ops.size()); 1016 if (BinaryOperator::isNot(Ops[i].Op)) { // Cannot occur for ^. 1017 Value *X = BinaryOperator::getNotArgument(Ops[i].Op); 1018 unsigned FoundX = FindInOperandList(Ops, i, X); 1019 if (FoundX != i) { 1020 if (Opcode == Instruction::And) // ...&X&~X = 0 1021 return Constant::getNullValue(X->getType()); 1022 1023 if (Opcode == Instruction::Or) // ...|X|~X = -1 1024 return Constant::getAllOnesValue(X->getType()); 1025 } 1026 } 1027 1028 // Next, check for duplicate pairs of values, which we assume are next to 1029 // each other, due to our sorting criteria. 1030 assert(i < Ops.size()); 1031 if (i+1 != Ops.size() && Ops[i+1].Op == Ops[i].Op) { 1032 if (Opcode == Instruction::And || Opcode == Instruction::Or) { 1033 // Drop duplicate values for And and Or. 1034 Ops.erase(Ops.begin()+i); 1035 --i; --e; 1036 ++NumAnnihil; 1037 continue; 1038 } 1039 1040 // Drop pairs of values for Xor. 1041 assert(Opcode == Instruction::Xor); 1042 if (e == 2) 1043 return Constant::getNullValue(Ops[0].Op->getType()); 1044 1045 // Y ^ X^X -> Y 1046 Ops.erase(Ops.begin()+i, Ops.begin()+i+2); 1047 i -= 1; e -= 2; 1048 ++NumAnnihil; 1049 } 1050 } 1051 return 0; 1052} 1053 1054/// OptimizeAdd - Optimize a series of operands to an 'add' instruction. This 1055/// optimizes based on identities. If it can be reduced to a single Value, it 1056/// is returned, otherwise the Ops list is mutated as necessary. 1057Value *Reassociate::OptimizeAdd(Instruction *I, 1058 SmallVectorImpl<ValueEntry> &Ops) { 1059 // Scan the operand lists looking for X and -X pairs. If we find any, we 1060 // can simplify the expression. X+-X == 0. While we're at it, scan for any 1061 // duplicates. We want to canonicalize Y+Y+Y+Z -> 3*Y+Z. 1062 // 1063 // TODO: We could handle "X + ~X" -> "-1" if we wanted, since "-X = ~X+1". 1064 // 1065 for (unsigned i = 0, e = Ops.size(); i != e; ++i) { 1066 Value *TheOp = Ops[i].Op; 1067 // Check to see if we've seen this operand before. If so, we factor all 1068 // instances of the operand together. Due to our sorting criteria, we know 1069 // that these need to be next to each other in the vector. 1070 if (i+1 != Ops.size() && Ops[i+1].Op == TheOp) { 1071 // Rescan the list, remove all instances of this operand from the expr. 1072 unsigned NumFound = 0; 1073 do { 1074 Ops.erase(Ops.begin()+i); 1075 ++NumFound; 1076 } while (i != Ops.size() && Ops[i].Op == TheOp); 1077 1078 DEBUG(errs() << "\nFACTORING [" << NumFound << "]: " << *TheOp << '\n'); 1079 ++NumFactor; 1080 1081 // Insert a new multiply. 1082 Value *Mul = ConstantInt::get(cast<IntegerType>(I->getType()), NumFound); 1083 Mul = BinaryOperator::CreateMul(TheOp, Mul, "factor", I); 1084 1085 // Now that we have inserted a multiply, optimize it. This allows us to 1086 // handle cases that require multiple factoring steps, such as this: 1087 // (X*2) + (X*2) + (X*2) -> (X*2)*3 -> X*6 1088 RedoInsts.insert(cast<Instruction>(Mul)); 1089 1090 // If every add operand was a duplicate, return the multiply. 1091 if (Ops.empty()) 1092 return Mul; 1093 1094 // Otherwise, we had some input that didn't have the dupe, such as 1095 // "A + A + B" -> "A*2 + B". Add the new multiply to the list of 1096 // things being added by this operation. 1097 Ops.insert(Ops.begin(), ValueEntry(getRank(Mul), Mul)); 1098 1099 --i; 1100 e = Ops.size(); 1101 continue; 1102 } 1103 1104 // Check for X and -X in the operand list. 1105 if (!BinaryOperator::isNeg(TheOp)) 1106 continue; 1107 1108 Value *X = BinaryOperator::getNegArgument(TheOp); 1109 unsigned FoundX = FindInOperandList(Ops, i, X); 1110 if (FoundX == i) 1111 continue; 1112 1113 // Remove X and -X from the operand list. 1114 if (Ops.size() == 2) 1115 return Constant::getNullValue(X->getType()); 1116 1117 Ops.erase(Ops.begin()+i); 1118 if (i < FoundX) 1119 --FoundX; 1120 else 1121 --i; // Need to back up an extra one. 1122 Ops.erase(Ops.begin()+FoundX); 1123 ++NumAnnihil; 1124 --i; // Revisit element. 1125 e -= 2; // Removed two elements. 1126 } 1127 1128 // Scan the operand list, checking to see if there are any common factors 1129 // between operands. Consider something like A*A+A*B*C+D. We would like to 1130 // reassociate this to A*(A+B*C)+D, which reduces the number of multiplies. 1131 // To efficiently find this, we count the number of times a factor occurs 1132 // for any ADD operands that are MULs. 1133 DenseMap<Value*, unsigned> FactorOccurrences; 1134 1135 // Keep track of each multiply we see, to avoid triggering on (X*4)+(X*4) 1136 // where they are actually the same multiply. 1137 unsigned MaxOcc = 0; 1138 Value *MaxOccVal = 0; 1139 for (unsigned i = 0, e = Ops.size(); i != e; ++i) { 1140 BinaryOperator *BOp = isReassociableOp(Ops[i].Op, Instruction::Mul); 1141 if (!BOp) 1142 continue; 1143 1144 // Compute all of the factors of this added value. 1145 SmallVector<Value*, 8> Factors; 1146 FindSingleUseMultiplyFactors(BOp, Factors, Ops); 1147 assert(Factors.size() > 1 && "Bad linearize!"); 1148 1149 // Add one to FactorOccurrences for each unique factor in this op. 1150 SmallPtrSet<Value*, 8> Duplicates; 1151 for (unsigned i = 0, e = Factors.size(); i != e; ++i) { 1152 Value *Factor = Factors[i]; 1153 if (!Duplicates.insert(Factor)) continue; 1154 1155 unsigned Occ = ++FactorOccurrences[Factor]; 1156 if (Occ > MaxOcc) { MaxOcc = Occ; MaxOccVal = Factor; } 1157 1158 // If Factor is a negative constant, add the negated value as a factor 1159 // because we can percolate the negate out. Watch for minint, which 1160 // cannot be positivified. 1161 if (ConstantInt *CI = dyn_cast<ConstantInt>(Factor)) 1162 if (CI->isNegative() && !CI->isMinValue(true)) { 1163 Factor = ConstantInt::get(CI->getContext(), -CI->getValue()); 1164 assert(!Duplicates.count(Factor) && 1165 "Shouldn't have two constant factors, missed a canonicalize"); 1166 1167 unsigned Occ = ++FactorOccurrences[Factor]; 1168 if (Occ > MaxOcc) { MaxOcc = Occ; MaxOccVal = Factor; } 1169 } 1170 } 1171 } 1172 1173 // If any factor occurred more than one time, we can pull it out. 1174 if (MaxOcc > 1) { 1175 DEBUG(errs() << "\nFACTORING [" << MaxOcc << "]: " << *MaxOccVal << '\n'); 1176 ++NumFactor; 1177 1178 // Create a new instruction that uses the MaxOccVal twice. If we don't do 1179 // this, we could otherwise run into situations where removing a factor 1180 // from an expression will drop a use of maxocc, and this can cause 1181 // RemoveFactorFromExpression on successive values to behave differently. 1182 Instruction *DummyInst = BinaryOperator::CreateAdd(MaxOccVal, MaxOccVal); 1183 SmallVector<WeakVH, 4> NewMulOps; 1184 for (unsigned i = 0; i != Ops.size(); ++i) { 1185 // Only try to remove factors from expressions we're allowed to. 1186 BinaryOperator *BOp = isReassociableOp(Ops[i].Op, Instruction::Mul); 1187 if (!BOp) 1188 continue; 1189 1190 if (Value *V = RemoveFactorFromExpression(Ops[i].Op, MaxOccVal)) { 1191 // The factorized operand may occur several times. Convert them all in 1192 // one fell swoop. 1193 for (unsigned j = Ops.size(); j != i;) { 1194 --j; 1195 if (Ops[j].Op == Ops[i].Op) { 1196 NewMulOps.push_back(V); 1197 Ops.erase(Ops.begin()+j); 1198 } 1199 } 1200 --i; 1201 } 1202 } 1203 1204 // No need for extra uses anymore. 1205 delete DummyInst; 1206 1207 unsigned NumAddedValues = NewMulOps.size(); 1208 Value *V = EmitAddTreeOfValues(I, NewMulOps); 1209 1210 // Now that we have inserted the add tree, optimize it. This allows us to 1211 // handle cases that require multiple factoring steps, such as this: 1212 // A*A*B + A*A*C --> A*(A*B+A*C) --> A*(A*(B+C)) 1213 assert(NumAddedValues > 1 && "Each occurrence should contribute a value"); 1214 (void)NumAddedValues; 1215 if (Instruction *VI = dyn_cast<Instruction>(V)) 1216 RedoInsts.insert(VI); 1217 1218 // Create the multiply. 1219 Instruction *V2 = BinaryOperator::CreateMul(V, MaxOccVal, "tmp", I); 1220 1221 // Rerun associate on the multiply in case the inner expression turned into 1222 // a multiply. We want to make sure that we keep things in canonical form. 1223 RedoInsts.insert(V2); 1224 1225 // If every add operand included the factor (e.g. "A*B + A*C"), then the 1226 // entire result expression is just the multiply "A*(B+C)". 1227 if (Ops.empty()) 1228 return V2; 1229 1230 // Otherwise, we had some input that didn't have the factor, such as 1231 // "A*B + A*C + D" -> "A*(B+C) + D". Add the new multiply to the list of 1232 // things being added by this operation. 1233 Ops.insert(Ops.begin(), ValueEntry(getRank(V2), V2)); 1234 } 1235 1236 return 0; 1237} 1238 1239namespace { 1240 /// \brief Predicate tests whether a ValueEntry's op is in a map. 1241 struct IsValueInMap { 1242 const DenseMap<Value *, unsigned> ⤅ 1243 1244 IsValueInMap(const DenseMap<Value *, unsigned> &Map) : Map(Map) {} 1245 1246 bool operator()(const ValueEntry &Entry) { 1247 return Map.find(Entry.Op) != Map.end(); 1248 } 1249 }; 1250} 1251 1252/// \brief Build up a vector of value/power pairs factoring a product. 1253/// 1254/// Given a series of multiplication operands, build a vector of factors and 1255/// the powers each is raised to when forming the final product. Sort them in 1256/// the order of descending power. 1257/// 1258/// (x*x) -> [(x, 2)] 1259/// ((x*x)*x) -> [(x, 3)] 1260/// ((((x*y)*x)*y)*x) -> [(x, 3), (y, 2)] 1261/// 1262/// \returns Whether any factors have a power greater than one. 1263bool Reassociate::collectMultiplyFactors(SmallVectorImpl<ValueEntry> &Ops, 1264 SmallVectorImpl<Factor> &Factors) { 1265 // FIXME: Have Ops be (ValueEntry, Multiplicity) pairs, simplifying this. 1266 // Compute the sum of powers of simplifiable factors. 1267 unsigned FactorPowerSum = 0; 1268 for (unsigned Idx = 1, Size = Ops.size(); Idx < Size; ++Idx) { 1269 Value *Op = Ops[Idx-1].Op; 1270 1271 // Count the number of occurrences of this value. 1272 unsigned Count = 1; 1273 for (; Idx < Size && Ops[Idx].Op == Op; ++Idx) 1274 ++Count; 1275 // Track for simplification all factors which occur 2 or more times. 1276 if (Count > 1) 1277 FactorPowerSum += Count; 1278 } 1279 1280 // We can only simplify factors if the sum of the powers of our simplifiable 1281 // factors is 4 or higher. When that is the case, we will *always* have 1282 // a simplification. This is an important invariant to prevent cyclicly 1283 // trying to simplify already minimal formations. 1284 if (FactorPowerSum < 4) 1285 return false; 1286 1287 // Now gather the simplifiable factors, removing them from Ops. 1288 FactorPowerSum = 0; 1289 for (unsigned Idx = 1; Idx < Ops.size(); ++Idx) { 1290 Value *Op = Ops[Idx-1].Op; 1291 1292 // Count the number of occurrences of this value. 1293 unsigned Count = 1; 1294 for (; Idx < Ops.size() && Ops[Idx].Op == Op; ++Idx) 1295 ++Count; 1296 if (Count == 1) 1297 continue; 1298 // Move an even number of occurrences to Factors. 1299 Count &= ~1U; 1300 Idx -= Count; 1301 FactorPowerSum += Count; 1302 Factors.push_back(Factor(Op, Count)); 1303 Ops.erase(Ops.begin()+Idx, Ops.begin()+Idx+Count); 1304 } 1305 1306 // None of the adjustments above should have reduced the sum of factor powers 1307 // below our mininum of '4'. 1308 assert(FactorPowerSum >= 4); 1309 1310 std::sort(Factors.begin(), Factors.end(), Factor::PowerDescendingSorter()); 1311 return true; 1312} 1313 1314/// \brief Build a tree of multiplies, computing the product of Ops. 1315static Value *buildMultiplyTree(IRBuilder<> &Builder, 1316 SmallVectorImpl<Value*> &Ops) { 1317 if (Ops.size() == 1) 1318 return Ops.back(); 1319 1320 Value *LHS = Ops.pop_back_val(); 1321 do { 1322 LHS = Builder.CreateMul(LHS, Ops.pop_back_val()); 1323 } while (!Ops.empty()); 1324 1325 return LHS; 1326} 1327 1328/// \brief Build a minimal multiplication DAG for (a^x)*(b^y)*(c^z)*... 1329/// 1330/// Given a vector of values raised to various powers, where no two values are 1331/// equal and the powers are sorted in decreasing order, compute the minimal 1332/// DAG of multiplies to compute the final product, and return that product 1333/// value. 1334Value *Reassociate::buildMinimalMultiplyDAG(IRBuilder<> &Builder, 1335 SmallVectorImpl<Factor> &Factors) { 1336 assert(Factors[0].Power); 1337 SmallVector<Value *, 4> OuterProduct; 1338 for (unsigned LastIdx = 0, Idx = 1, Size = Factors.size(); 1339 Idx < Size && Factors[Idx].Power > 0; ++Idx) { 1340 if (Factors[Idx].Power != Factors[LastIdx].Power) { 1341 LastIdx = Idx; 1342 continue; 1343 } 1344 1345 // We want to multiply across all the factors with the same power so that 1346 // we can raise them to that power as a single entity. Build a mini tree 1347 // for that. 1348 SmallVector<Value *, 4> InnerProduct; 1349 InnerProduct.push_back(Factors[LastIdx].Base); 1350 do { 1351 InnerProduct.push_back(Factors[Idx].Base); 1352 ++Idx; 1353 } while (Idx < Size && Factors[Idx].Power == Factors[LastIdx].Power); 1354 1355 // Reset the base value of the first factor to the new expression tree. 1356 // We'll remove all the factors with the same power in a second pass. 1357 Value *M = Factors[LastIdx].Base = buildMultiplyTree(Builder, InnerProduct); 1358 if (Instruction *MI = dyn_cast<Instruction>(M)) 1359 RedoInsts.insert(MI); 1360 1361 LastIdx = Idx; 1362 } 1363 // Unique factors with equal powers -- we've folded them into the first one's 1364 // base. 1365 Factors.erase(std::unique(Factors.begin(), Factors.end(), 1366 Factor::PowerEqual()), 1367 Factors.end()); 1368 1369 // Iteratively collect the base of each factor with an add power into the 1370 // outer product, and halve each power in preparation for squaring the 1371 // expression. 1372 for (unsigned Idx = 0, Size = Factors.size(); Idx != Size; ++Idx) { 1373 if (Factors[Idx].Power & 1) 1374 OuterProduct.push_back(Factors[Idx].Base); 1375 Factors[Idx].Power >>= 1; 1376 } 1377 if (Factors[0].Power) { 1378 Value *SquareRoot = buildMinimalMultiplyDAG(Builder, Factors); 1379 OuterProduct.push_back(SquareRoot); 1380 OuterProduct.push_back(SquareRoot); 1381 } 1382 if (OuterProduct.size() == 1) 1383 return OuterProduct.front(); 1384 1385 Value *V = buildMultiplyTree(Builder, OuterProduct); 1386 return V; 1387} 1388 1389Value *Reassociate::OptimizeMul(BinaryOperator *I, 1390 SmallVectorImpl<ValueEntry> &Ops) { 1391 // We can only optimize the multiplies when there is a chain of more than 1392 // three, such that a balanced tree might require fewer total multiplies. 1393 if (Ops.size() < 4) 1394 return 0; 1395 1396 // Try to turn linear trees of multiplies without other uses of the 1397 // intermediate stages into minimal multiply DAGs with perfect sub-expression 1398 // re-use. 1399 SmallVector<Factor, 4> Factors; 1400 if (!collectMultiplyFactors(Ops, Factors)) 1401 return 0; // All distinct factors, so nothing left for us to do. 1402 1403 IRBuilder<> Builder(I); 1404 Value *V = buildMinimalMultiplyDAG(Builder, Factors); 1405 if (Ops.empty()) 1406 return V; 1407 1408 ValueEntry NewEntry = ValueEntry(getRank(V), V); 1409 Ops.insert(std::lower_bound(Ops.begin(), Ops.end(), NewEntry), NewEntry); 1410 return 0; 1411} 1412 1413Value *Reassociate::OptimizeExpression(BinaryOperator *I, 1414 SmallVectorImpl<ValueEntry> &Ops) { 1415 // Now that we have the linearized expression tree, try to optimize it. 1416 // Start by folding any constants that we found. 1417 Constant *Cst = 0; 1418 unsigned Opcode = I->getOpcode(); 1419 while (!Ops.empty() && isa<Constant>(Ops.back().Op)) { 1420 Constant *C = cast<Constant>(Ops.pop_back_val().Op); 1421 Cst = Cst ? ConstantExpr::get(Opcode, C, Cst) : C; 1422 } 1423 // If there was nothing but constants then we are done. 1424 if (Ops.empty()) 1425 return Cst; 1426 1427 // Put the combined constant back at the end of the operand list, except if 1428 // there is no point. For example, an add of 0 gets dropped here, while a 1429 // multiplication by zero turns the whole expression into zero. 1430 if (Cst && Cst != ConstantExpr::getBinOpIdentity(Opcode, I->getType())) { 1431 if (Cst == ConstantExpr::getBinOpAbsorber(Opcode, I->getType())) 1432 return Cst; 1433 Ops.push_back(ValueEntry(0, Cst)); 1434 } 1435 1436 if (Ops.size() == 1) return Ops[0].Op; 1437 1438 // Handle destructive annihilation due to identities between elements in the 1439 // argument list here. 1440 unsigned NumOps = Ops.size(); 1441 switch (Opcode) { 1442 default: break; 1443 case Instruction::And: 1444 case Instruction::Or: 1445 case Instruction::Xor: 1446 if (Value *Result = OptimizeAndOrXor(Opcode, Ops)) 1447 return Result; 1448 break; 1449 1450 case Instruction::Add: 1451 if (Value *Result = OptimizeAdd(I, Ops)) 1452 return Result; 1453 break; 1454 1455 case Instruction::Mul: 1456 if (Value *Result = OptimizeMul(I, Ops)) 1457 return Result; 1458 break; 1459 } 1460 1461 if (Ops.size() != NumOps) 1462 return OptimizeExpression(I, Ops); 1463 return 0; 1464} 1465 1466/// EraseInst - Zap the given instruction, adding interesting operands to the 1467/// work list. 1468void Reassociate::EraseInst(Instruction *I) { 1469 assert(isInstructionTriviallyDead(I) && "Trivially dead instructions only!"); 1470 SmallVector<Value*, 8> Ops(I->op_begin(), I->op_end()); 1471 // Erase the dead instruction. 1472 ValueRankMap.erase(I); 1473 RedoInsts.remove(I); 1474 I->eraseFromParent(); 1475 // Optimize its operands. 1476 SmallPtrSet<Instruction *, 8> Visited; // Detect self-referential nodes. 1477 for (unsigned i = 0, e = Ops.size(); i != e; ++i) 1478 if (Instruction *Op = dyn_cast<Instruction>(Ops[i])) { 1479 // If this is a node in an expression tree, climb to the expression root 1480 // and add that since that's where optimization actually happens. 1481 unsigned Opcode = Op->getOpcode(); 1482 while (Op->hasOneUse() && Op->use_back()->getOpcode() == Opcode && 1483 Visited.insert(Op)) 1484 Op = Op->use_back(); 1485 RedoInsts.insert(Op); 1486 } 1487} 1488 1489/// OptimizeInst - Inspect and optimize the given instruction. Note that erasing 1490/// instructions is not allowed. 1491void Reassociate::OptimizeInst(Instruction *I) { 1492 // Only consider operations that we understand. 1493 if (!isa<BinaryOperator>(I)) 1494 return; 1495 1496 if (I->getOpcode() == Instruction::Shl && 1497 isa<ConstantInt>(I->getOperand(1))) 1498 // If an operand of this shift is a reassociable multiply, or if the shift 1499 // is used by a reassociable multiply or add, turn into a multiply. 1500 if (isReassociableOp(I->getOperand(0), Instruction::Mul) || 1501 (I->hasOneUse() && 1502 (isReassociableOp(I->use_back(), Instruction::Mul) || 1503 isReassociableOp(I->use_back(), Instruction::Add)))) { 1504 Instruction *NI = ConvertShiftToMul(I); 1505 RedoInsts.insert(I); 1506 MadeChange = true; 1507 I = NI; 1508 } 1509 1510 // Floating point binary operators are not associative, but we can still 1511 // commute (some) of them, to canonicalize the order of their operands. 1512 // This can potentially expose more CSE opportunities, and makes writing 1513 // other transformations simpler. 1514 if ((I->getType()->isFloatingPointTy() || I->getType()->isVectorTy())) { 1515 // FAdd and FMul can be commuted. 1516 if (I->getOpcode() != Instruction::FMul && 1517 I->getOpcode() != Instruction::FAdd) 1518 return; 1519 1520 Value *LHS = I->getOperand(0); 1521 Value *RHS = I->getOperand(1); 1522 unsigned LHSRank = getRank(LHS); 1523 unsigned RHSRank = getRank(RHS); 1524 1525 // Sort the operands by rank. 1526 if (RHSRank < LHSRank) { 1527 I->setOperand(0, RHS); 1528 I->setOperand(1, LHS); 1529 } 1530 1531 return; 1532 } 1533 1534 // Do not reassociate boolean (i1) expressions. We want to preserve the 1535 // original order of evaluation for short-circuited comparisons that 1536 // SimplifyCFG has folded to AND/OR expressions. If the expression 1537 // is not further optimized, it is likely to be transformed back to a 1538 // short-circuited form for code gen, and the source order may have been 1539 // optimized for the most likely conditions. 1540 if (I->getType()->isIntegerTy(1)) 1541 return; 1542 1543 // If this is a subtract instruction which is not already in negate form, 1544 // see if we can convert it to X+-Y. 1545 if (I->getOpcode() == Instruction::Sub) { 1546 if (ShouldBreakUpSubtract(I)) { 1547 Instruction *NI = BreakUpSubtract(I); 1548 RedoInsts.insert(I); 1549 MadeChange = true; 1550 I = NI; 1551 } else if (BinaryOperator::isNeg(I)) { 1552 // Otherwise, this is a negation. See if the operand is a multiply tree 1553 // and if this is not an inner node of a multiply tree. 1554 if (isReassociableOp(I->getOperand(1), Instruction::Mul) && 1555 (!I->hasOneUse() || 1556 !isReassociableOp(I->use_back(), Instruction::Mul))) { 1557 Instruction *NI = LowerNegateToMultiply(I); 1558 RedoInsts.insert(I); 1559 MadeChange = true; 1560 I = NI; 1561 } 1562 } 1563 } 1564 1565 // If this instruction is an associative binary operator, process it. 1566 if (!I->isAssociative()) return; 1567 BinaryOperator *BO = cast<BinaryOperator>(I); 1568 1569 // If this is an interior node of a reassociable tree, ignore it until we 1570 // get to the root of the tree, to avoid N^2 analysis. 1571 unsigned Opcode = BO->getOpcode(); 1572 if (BO->hasOneUse() && BO->use_back()->getOpcode() == Opcode) 1573 return; 1574 1575 // If this is an add tree that is used by a sub instruction, ignore it 1576 // until we process the subtract. 1577 if (BO->hasOneUse() && BO->getOpcode() == Instruction::Add && 1578 cast<Instruction>(BO->use_back())->getOpcode() == Instruction::Sub) 1579 return; 1580 1581 ReassociateExpression(BO); 1582} 1583 1584void Reassociate::ReassociateExpression(BinaryOperator *I) { 1585 1586 // First, walk the expression tree, linearizing the tree, collecting the 1587 // operand information. 1588 SmallVector<RepeatedValue, 8> Tree; 1589 MadeChange |= LinearizeExprTree(I, Tree); 1590 SmallVector<ValueEntry, 8> Ops; 1591 Ops.reserve(Tree.size()); 1592 for (unsigned i = 0, e = Tree.size(); i != e; ++i) { 1593 RepeatedValue E = Tree[i]; 1594 Ops.append(E.second.getZExtValue(), 1595 ValueEntry(getRank(E.first), E.first)); 1596 } 1597 1598 DEBUG(dbgs() << "RAIn:\t"; PrintOps(I, Ops); dbgs() << '\n'); 1599 1600 // Now that we have linearized the tree to a list and have gathered all of 1601 // the operands and their ranks, sort the operands by their rank. Use a 1602 // stable_sort so that values with equal ranks will have their relative 1603 // positions maintained (and so the compiler is deterministic). Note that 1604 // this sorts so that the highest ranking values end up at the beginning of 1605 // the vector. 1606 std::stable_sort(Ops.begin(), Ops.end()); 1607 1608 // OptimizeExpression - Now that we have the expression tree in a convenient 1609 // sorted form, optimize it globally if possible. 1610 if (Value *V = OptimizeExpression(I, Ops)) { 1611 if (V == I) 1612 // Self-referential expression in unreachable code. 1613 return; 1614 // This expression tree simplified to something that isn't a tree, 1615 // eliminate it. 1616 DEBUG(dbgs() << "Reassoc to scalar: " << *V << '\n'); 1617 I->replaceAllUsesWith(V); 1618 if (Instruction *VI = dyn_cast<Instruction>(V)) 1619 VI->setDebugLoc(I->getDebugLoc()); 1620 RedoInsts.insert(I); 1621 ++NumAnnihil; 1622 return; 1623 } 1624 1625 // We want to sink immediates as deeply as possible except in the case where 1626 // this is a multiply tree used only by an add, and the immediate is a -1. 1627 // In this case we reassociate to put the negation on the outside so that we 1628 // can fold the negation into the add: (-X)*Y + Z -> Z-X*Y 1629 if (I->getOpcode() == Instruction::Mul && I->hasOneUse() && 1630 cast<Instruction>(I->use_back())->getOpcode() == Instruction::Add && 1631 isa<ConstantInt>(Ops.back().Op) && 1632 cast<ConstantInt>(Ops.back().Op)->isAllOnesValue()) { 1633 ValueEntry Tmp = Ops.pop_back_val(); 1634 Ops.insert(Ops.begin(), Tmp); 1635 } 1636 1637 DEBUG(dbgs() << "RAOut:\t"; PrintOps(I, Ops); dbgs() << '\n'); 1638 1639 if (Ops.size() == 1) { 1640 if (Ops[0].Op == I) 1641 // Self-referential expression in unreachable code. 1642 return; 1643 1644 // This expression tree simplified to something that isn't a tree, 1645 // eliminate it. 1646 I->replaceAllUsesWith(Ops[0].Op); 1647 if (Instruction *OI = dyn_cast<Instruction>(Ops[0].Op)) 1648 OI->setDebugLoc(I->getDebugLoc()); 1649 RedoInsts.insert(I); 1650 return; 1651 } 1652 1653 // Now that we ordered and optimized the expressions, splat them back into 1654 // the expression tree, removing any unneeded nodes. 1655 RewriteExprTree(I, Ops); 1656} 1657 1658bool Reassociate::runOnFunction(Function &F) { 1659 // Calculate the rank map for F 1660 BuildRankMap(F); 1661 1662 MadeChange = false; 1663 for (Function::iterator BI = F.begin(), BE = F.end(); BI != BE; ++BI) { 1664 // Optimize every instruction in the basic block. 1665 for (BasicBlock::iterator II = BI->begin(), IE = BI->end(); II != IE; ) 1666 if (isInstructionTriviallyDead(II)) { 1667 EraseInst(II++); 1668 } else { 1669 OptimizeInst(II); 1670 assert(II->getParent() == BI && "Moved to a different block!"); 1671 ++II; 1672 } 1673 1674 // If this produced extra instructions to optimize, handle them now. 1675 while (!RedoInsts.empty()) { 1676 Instruction *I = RedoInsts.pop_back_val(); 1677 if (isInstructionTriviallyDead(I)) 1678 EraseInst(I); 1679 else 1680 OptimizeInst(I); 1681 } 1682 } 1683 1684 // We are done with the rank map. 1685 RankMap.clear(); 1686 ValueRankMap.clear(); 1687 1688 return MadeChange; 1689} 1690