Reassociate.cpp revision d34491f6751ae2f8daf3e857c84bcb5b06fba889
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/Instructions.h" 30#include "llvm/IntrinsicInst.h" 31#include "llvm/Pass.h" 32#include "llvm/Assembly/Writer.h" 33#include "llvm/Support/CFG.h" 34#include "llvm/Support/IRBuilder.h" 35#include "llvm/Support/Debug.h" 36#include "llvm/Support/ValueHandle.h" 37#include "llvm/Support/raw_ostream.h" 38#include "llvm/ADT/DenseMap.h" 39#include "llvm/ADT/PostOrderIterator.h" 40#include "llvm/ADT/SetVector.h" 41#include "llvm/ADT/STLExtras.h" 42#include "llvm/ADT/Statistic.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 Value *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} 341 342/// EvaluateRepeatedConstant - Compute C op C op ... op C where the constant C 343/// is repeated Weight times. 344static Constant *EvaluateRepeatedConstant(unsigned Opcode, Constant *C, 345 APInt Weight) { 346 // For addition the result can be efficiently computed as the product of the 347 // constant and the weight. 348 if (Opcode == Instruction::Add) 349 return ConstantExpr::getMul(C, ConstantInt::get(C->getContext(), Weight)); 350 351 // The weight might be huge, so compute by repeated squaring to ensure that 352 // compile time is proportional to the logarithm of the weight. 353 Constant *Result = 0; 354 Constant *Power = C; // Successively C, C op C, (C op C) op (C op C) etc. 355 // Visit the bits in Weight. 356 while (Weight != 0) { 357 // If the current bit in Weight is non-zero do Result = Result op Power. 358 if (Weight[0]) 359 Result = Result ? ConstantExpr::get(Opcode, Result, Power) : Power; 360 // Move on to the next bit if any more are non-zero. 361 Weight = Weight.lshr(1); 362 if (Weight.isMinValue()) 363 break; 364 // Square the power. 365 Power = ConstantExpr::get(Opcode, Power, Power); 366 } 367 368 assert(Result && "Only positive weights supported!"); 369 return Result; 370} 371 372typedef std::pair<Value*, APInt> RepeatedValue; 373 374/// LinearizeExprTree - Given an associative binary expression, return the leaf 375/// nodes in Ops along with their weights (how many times the leaf occurs). The 376/// original expression is the same as 377/// (Ops[0].first op Ops[0].first op ... Ops[0].first) <- Ops[0].second times 378/// op 379/// (Ops[1].first op Ops[1].first op ... Ops[1].first) <- Ops[1].second times 380/// op 381/// ... 382/// op 383/// (Ops[N].first op Ops[N].first op ... Ops[N].first) <- Ops[N].second times 384/// 385/// Note that the values Ops[0].first, ..., Ops[N].first are all distinct, and 386/// they are all non-constant except possibly for the last one, which if it is 387/// constant will have weight one (Ops[N].second === 1). 388/// 389/// This routine may modify the function, in which case it returns 'true'. The 390/// changes it makes may well be destructive, changing the value computed by 'I' 391/// to something completely different. Thus if the routine returns 'true' then 392/// you MUST either replace I with a new expression computed from the Ops array, 393/// or use RewriteExprTree to put the values back in. 394/// 395/// A leaf node is either not a binary operation of the same kind as the root 396/// node 'I' (i.e. is not a binary operator at all, or is, but with a different 397/// opcode), or is the same kind of binary operator but has a use which either 398/// does not belong to the expression, or does belong to the expression but is 399/// a leaf node. Every leaf node has at least one use that is a non-leaf node 400/// of the expression, while for non-leaf nodes (except for the root 'I') every 401/// use is a non-leaf node of the expression. 402/// 403/// For example: 404/// expression graph node names 405/// 406/// + | I 407/// / \ | 408/// + + | A, B 409/// / \ / \ | 410/// * + * | C, D, E 411/// / \ / \ / \ | 412/// + * | F, G 413/// 414/// The leaf nodes are C, E, F and G. The Ops array will contain (maybe not in 415/// that order) (C, 1), (E, 1), (F, 2), (G, 2). 416/// 417/// The expression is maximal: if some instruction is a binary operator of the 418/// same kind as 'I', and all of its uses are non-leaf nodes of the expression, 419/// then the instruction also belongs to the expression, is not a leaf node of 420/// it, and its operands also belong to the expression (but may be leaf nodes). 421/// 422/// NOTE: This routine will set operands of non-leaf non-root nodes to undef in 423/// order to ensure that every non-root node in the expression has *exactly one* 424/// use by a non-leaf node of the expression. This destruction means that the 425/// caller MUST either replace 'I' with a new expression or use something like 426/// RewriteExprTree to put the values back in if the routine indicates that it 427/// made a change by returning 'true'. 428/// 429/// In the above example either the right operand of A or the left operand of B 430/// will be replaced by undef. If it is B's operand then this gives: 431/// 432/// + | I 433/// / \ | 434/// + + | A, B - operand of B replaced with undef 435/// / \ \ | 436/// * + * | C, D, E 437/// / \ / \ / \ | 438/// + * | F, G 439/// 440/// Note that such undef operands can only be reached by passing through 'I'. 441/// For example, if you visit operands recursively starting from a leaf node 442/// then you will never see such an undef operand unless you get back to 'I', 443/// which requires passing through a phi node. 444/// 445/// Note that this routine may also mutate binary operators of the wrong type 446/// that have all uses inside the expression (i.e. only used by non-leaf nodes 447/// of the expression) if it can turn them into binary operators of the right 448/// type and thus make the expression bigger. 449 450static bool LinearizeExprTree(BinaryOperator *I, 451 SmallVectorImpl<RepeatedValue> &Ops) { 452 DEBUG(dbgs() << "LINEARIZE: " << *I << '\n'); 453 unsigned Bitwidth = I->getType()->getScalarType()->getPrimitiveSizeInBits(); 454 unsigned Opcode = I->getOpcode(); 455 assert(Instruction::isAssociative(Opcode) && 456 Instruction::isCommutative(Opcode) && 457 "Expected an associative and commutative operation!"); 458 // If we see an absorbing element then the entire expression must be equal to 459 // it. For example, if this is a multiplication expression and zero occurs as 460 // an operand somewhere in it then the result of the expression must be zero. 461 Constant *Absorber = ConstantExpr::getBinOpAbsorber(Opcode, I->getType()); 462 463 // Visit all operands of the expression, keeping track of their weight (the 464 // number of paths from the expression root to the operand, or if you like 465 // the number of times that operand occurs in the linearized expression). 466 // For example, if I = X + A, where X = A + B, then I, X and B have weight 1 467 // while A has weight two. 468 469 // Worklist of non-leaf nodes (their operands are in the expression too) along 470 // with their weights, representing a certain number of paths to the operator. 471 // If an operator occurs in the worklist multiple times then we found multiple 472 // ways to get to it. 473 SmallVector<std::pair<BinaryOperator*, APInt>, 8> Worklist; // (Op, Weight) 474 Worklist.push_back(std::make_pair(I, APInt(Bitwidth, 1))); 475 bool MadeChange = false; 476 477 // Leaves of the expression are values that either aren't the right kind of 478 // operation (eg: a constant, or a multiply in an add tree), or are, but have 479 // some uses that are not inside the expression. For example, in I = X + X, 480 // X = A + B, the value X has two uses (by I) that are in the expression. If 481 // X has any other uses, for example in a return instruction, then we consider 482 // X to be a leaf, and won't analyze it further. When we first visit a value, 483 // if it has more than one use then at first we conservatively consider it to 484 // be a leaf. Later, as the expression is explored, we may discover some more 485 // uses of the value from inside the expression. If all uses turn out to be 486 // from within the expression (and the value is a binary operator of the right 487 // kind) then the value is no longer considered to be a leaf, and its operands 488 // are explored. 489 490 // Leaves - Keeps track of the set of putative leaves as well as the number of 491 // paths to each leaf seen so far. 492 typedef DenseMap<Value*, APInt> LeafMap; 493 LeafMap Leaves; // Leaf -> Total weight so far. 494 SmallVector<Value*, 8> LeafOrder; // Ensure deterministic leaf output order. 495 496#ifndef NDEBUG 497 SmallPtrSet<Value*, 8> Visited; // For sanity checking the iteration scheme. 498#endif 499 while (!Worklist.empty()) { 500 std::pair<BinaryOperator*, APInt> P = Worklist.pop_back_val(); 501 I = P.first; // We examine the operands of this binary operator. 502 503 for (unsigned OpIdx = 0; OpIdx < 2; ++OpIdx) { // Visit operands. 504 Value *Op = I->getOperand(OpIdx); 505 APInt Weight = P.second; // Number of paths to this operand. 506 DEBUG(dbgs() << "OPERAND: " << *Op << " (" << Weight << ")\n"); 507 assert(!Op->use_empty() && "No uses, so how did we get to it?!"); 508 509 // If the expression contains an absorbing element then there is no need 510 // to analyze it further: it must evaluate to the absorbing element. 511 if (Op == Absorber && !Weight.isMinValue()) { 512 Ops.push_back(std::make_pair(Absorber, APInt(Bitwidth, 1))); 513 return MadeChange; 514 } 515 516 // If this is a binary operation of the right kind with only one use then 517 // add its operands to the expression. 518 if (BinaryOperator *BO = isReassociableOp(Op, Opcode)) { 519 assert(Visited.insert(Op) && "Not first visit!"); 520 DEBUG(dbgs() << "DIRECT ADD: " << *Op << " (" << Weight << ")\n"); 521 Worklist.push_back(std::make_pair(BO, Weight)); 522 continue; 523 } 524 525 // Appears to be a leaf. Is the operand already in the set of leaves? 526 LeafMap::iterator It = Leaves.find(Op); 527 if (It == Leaves.end()) { 528 // Not in the leaf map. Must be the first time we saw this operand. 529 assert(Visited.insert(Op) && "Not first visit!"); 530 if (!Op->hasOneUse()) { 531 // This value has uses not accounted for by the expression, so it is 532 // not safe to modify. Mark it as being a leaf. 533 DEBUG(dbgs() << "ADD USES LEAF: " << *Op << " (" << Weight << ")\n"); 534 LeafOrder.push_back(Op); 535 Leaves[Op] = Weight; 536 continue; 537 } 538 // No uses outside the expression, try morphing it. 539 } else if (It != Leaves.end()) { 540 // Already in the leaf map. 541 assert(Visited.count(Op) && "In leaf map but not visited!"); 542 543 // Update the number of paths to the leaf. 544 IncorporateWeight(It->second, Weight, Opcode); 545 546 // The leaf already has one use from inside the expression. As we want 547 // exactly one such use, drop this new use of the leaf. 548 assert(!Op->hasOneUse() && "Only one use, but we got here twice!"); 549 I->setOperand(OpIdx, UndefValue::get(I->getType())); 550 MadeChange = true; 551 552 // If the leaf is a binary operation of the right kind and we now see 553 // that its multiple original uses were in fact all by nodes belonging 554 // to the expression, then no longer consider it to be a leaf and add 555 // its operands to the expression. 556 if (BinaryOperator *BO = isReassociableOp(Op, Opcode)) { 557 DEBUG(dbgs() << "UNLEAF: " << *Op << " (" << It->second << ")\n"); 558 Worklist.push_back(std::make_pair(BO, It->second)); 559 Leaves.erase(It); 560 continue; 561 } 562 563 // If we still have uses that are not accounted for by the expression 564 // then it is not safe to modify the value. 565 if (!Op->hasOneUse()) 566 continue; 567 568 // No uses outside the expression, try morphing it. 569 Weight = It->second; 570 Leaves.erase(It); // Since the value may be morphed below. 571 } 572 573 // At this point we have a value which, first of all, is not a binary 574 // expression of the right kind, and secondly, is only used inside the 575 // expression. This means that it can safely be modified. See if we 576 // can usefully morph it into an expression of the right kind. 577 assert((!isa<Instruction>(Op) || 578 cast<Instruction>(Op)->getOpcode() != Opcode) && 579 "Should have been handled above!"); 580 assert(Op->hasOneUse() && "Has uses outside the expression tree!"); 581 582 // If this is a multiply expression, turn any internal negations into 583 // multiplies by -1 so they can be reassociated. 584 BinaryOperator *BO = dyn_cast<BinaryOperator>(Op); 585 if (Opcode == Instruction::Mul && BO && BinaryOperator::isNeg(BO)) { 586 DEBUG(dbgs() << "MORPH LEAF: " << *Op << " (" << Weight << ") TO "); 587 BO = LowerNegateToMultiply(BO); 588 DEBUG(dbgs() << *BO << 'n'); 589 Worklist.push_back(std::make_pair(BO, Weight)); 590 MadeChange = true; 591 continue; 592 } 593 594 // Failed to morph into an expression of the right type. This really is 595 // a leaf. 596 DEBUG(dbgs() << "ADD LEAF: " << *Op << " (" << Weight << ")\n"); 597 assert(!isReassociableOp(Op, Opcode) && "Value was morphed?"); 598 LeafOrder.push_back(Op); 599 Leaves[Op] = Weight; 600 } 601 } 602 603 // The leaves, repeated according to their weights, represent the linearized 604 // form of the expression. 605 Constant *Cst = 0; // Accumulate constants here. 606 for (unsigned i = 0, e = LeafOrder.size(); i != e; ++i) { 607 Value *V = LeafOrder[i]; 608 LeafMap::iterator It = Leaves.find(V); 609 if (It == Leaves.end()) 610 // Node initially thought to be a leaf wasn't. 611 continue; 612 assert(!isReassociableOp(V, Opcode) && "Shouldn't be a leaf!"); 613 APInt Weight = It->second; 614 if (Weight.isMinValue()) 615 // Leaf already output or weight reduction eliminated it. 616 continue; 617 // Ensure the leaf is only output once. 618 It->second = 0; 619 // Glob all constants together into Cst. 620 if (Constant *C = dyn_cast<Constant>(V)) { 621 C = EvaluateRepeatedConstant(Opcode, C, Weight); 622 Cst = Cst ? ConstantExpr::get(Opcode, Cst, C) : C; 623 continue; 624 } 625 // Add non-constant 626 Ops.push_back(std::make_pair(V, Weight)); 627 } 628 629 // Add any constants back into Ops, all globbed together and reduced to having 630 // weight 1 for the convenience of users. 631 Constant *Identity = ConstantExpr::getBinOpIdentity(Opcode, I->getType()); 632 if (Cst && Cst != Identity) { 633 // If combining multiple constants resulted in the absorber then the entire 634 // expression must evaluate to the absorber. 635 if (Cst == Absorber) 636 Ops.clear(); 637 Ops.push_back(std::make_pair(Cst, APInt(Bitwidth, 1))); 638 } 639 640 // For nilpotent operations or addition there may be no operands, for example 641 // because the expression was "X xor X" or consisted of 2^Bitwidth additions: 642 // in both cases the weight reduces to 0 causing the value to be skipped. 643 if (Ops.empty()) { 644 assert(Identity && "Associative operation without identity!"); 645 Ops.push_back(std::make_pair(Identity, APInt(Bitwidth, 1))); 646 } 647 648 return MadeChange; 649} 650 651// RewriteExprTree - Now that the operands for this expression tree are 652// linearized and optimized, emit them in-order. 653void Reassociate::RewriteExprTree(BinaryOperator *I, 654 SmallVectorImpl<ValueEntry> &Ops) { 655 assert(Ops.size() > 1 && "Single values should be used directly!"); 656 657 // Since our optimizations never increase the number of operations, the new 658 // expression can always be written by reusing the existing binary operators 659 // from the original expression tree, without creating any new instructions, 660 // though the rewritten expression may have a completely different topology. 661 // We take care to not change anything if the new expression will be the same 662 // as the original. If more than trivial changes (like commuting operands) 663 // were made then we are obliged to clear out any optional subclass data like 664 // nsw flags. 665 666 /// NodesToRewrite - Nodes from the original expression available for writing 667 /// the new expression into. 668 SmallVector<BinaryOperator*, 8> NodesToRewrite; 669 unsigned Opcode = I->getOpcode(); 670 NodesToRewrite.push_back(I); 671 672 // ExpressionChanged - Non-null if the rewritten expression differs from the 673 // original in some non-trivial way, requiring the clearing of optional flags. 674 // Flags are cleared from the operator in ExpressionChanged up to I inclusive. 675 BinaryOperator *ExpressionChanged = 0; 676 BinaryOperator *Previous; 677 BinaryOperator *Op = 0; 678 for (unsigned i = 0, e = Ops.size(); i != e; ++i) { 679 assert(!NodesToRewrite.empty() && 680 "Optimized expressions has more nodes than original!"); 681 Previous = Op; Op = NodesToRewrite.pop_back_val(); 682 if (ExpressionChanged) 683 // Compactify the tree instructions together with each other to guarantee 684 // that the expression tree is dominated by all of Ops. 685 Op->moveBefore(Previous); 686 687 // The last operation (which comes earliest in the IR) is special as both 688 // operands will come from Ops, rather than just one with the other being 689 // a subexpression. 690 if (i+2 == Ops.size()) { 691 Value *NewLHS = Ops[i].Op; 692 Value *NewRHS = Ops[i+1].Op; 693 Value *OldLHS = Op->getOperand(0); 694 Value *OldRHS = Op->getOperand(1); 695 696 if (NewLHS == OldLHS && NewRHS == OldRHS) 697 // Nothing changed, leave it alone. 698 break; 699 700 if (NewLHS == OldRHS && NewRHS == OldLHS) { 701 // The order of the operands was reversed. Swap them. 702 DEBUG(dbgs() << "RA: " << *Op << '\n'); 703 Op->swapOperands(); 704 DEBUG(dbgs() << "TO: " << *Op << '\n'); 705 MadeChange = true; 706 ++NumChanged; 707 break; 708 } 709 710 // The new operation differs non-trivially from the original. Overwrite 711 // the old operands with the new ones. 712 DEBUG(dbgs() << "RA: " << *Op << '\n'); 713 if (NewLHS != OldLHS) { 714 if (BinaryOperator *BO = isReassociableOp(OldLHS, Opcode)) 715 NodesToRewrite.push_back(BO); 716 Op->setOperand(0, NewLHS); 717 } 718 if (NewRHS != OldRHS) { 719 if (BinaryOperator *BO = isReassociableOp(OldRHS, Opcode)) 720 NodesToRewrite.push_back(BO); 721 Op->setOperand(1, NewRHS); 722 } 723 DEBUG(dbgs() << "TO: " << *Op << '\n'); 724 725 ExpressionChanged = Op; 726 MadeChange = true; 727 ++NumChanged; 728 729 break; 730 } 731 732 // Not the last operation. The left-hand side will be a sub-expression 733 // while the right-hand side will be the current element of Ops. 734 Value *NewRHS = Ops[i].Op; 735 if (NewRHS != Op->getOperand(1)) { 736 DEBUG(dbgs() << "RA: " << *Op << '\n'); 737 if (NewRHS == Op->getOperand(0)) { 738 // The new right-hand side was already present as the left operand. If 739 // we are lucky then swapping the operands will sort out both of them. 740 Op->swapOperands(); 741 } else { 742 // Overwrite with the new right-hand side. 743 if (BinaryOperator *BO = isReassociableOp(Op->getOperand(1), Opcode)) 744 NodesToRewrite.push_back(BO); 745 Op->setOperand(1, NewRHS); 746 ExpressionChanged = Op; 747 } 748 DEBUG(dbgs() << "TO: " << *Op << '\n'); 749 MadeChange = true; 750 ++NumChanged; 751 } 752 753 // Now deal with the left-hand side. If this is already an operation node 754 // from the original expression then just rewrite the rest of the expression 755 // into it. 756 if (BinaryOperator *BO = isReassociableOp(Op->getOperand(0), Opcode)) { 757 NodesToRewrite.push_back(BO); 758 continue; 759 } 760 761 // Otherwise, grab a spare node from the original expression and use that as 762 // the left-hand side. 763 assert(!NodesToRewrite.empty() && 764 "Optimized expressions has more nodes than original!"); 765 DEBUG(dbgs() << "RA: " << *Op << '\n'); 766 Op->setOperand(0, NodesToRewrite.back()); 767 DEBUG(dbgs() << "TO: " << *Op << '\n'); 768 ExpressionChanged = Op; 769 MadeChange = true; 770 ++NumChanged; 771 } 772 773 // If the expression changed non-trivially then clear out all subclass data 774 // starting from the operator specified in ExpressionChanged. 775 if (ExpressionChanged) { 776 do { 777 ExpressionChanged->clearSubclassOptionalData(); 778 if (ExpressionChanged == I) 779 break; 780 ExpressionChanged = cast<BinaryOperator>(*ExpressionChanged->use_begin()); 781 } while (1); 782 } 783 784 // Throw away any left over nodes from the original expression. 785 for (unsigned i = 0, e = NodesToRewrite.size(); i != e; ++i) 786 RedoInsts.insert(NodesToRewrite[i]); 787} 788 789/// NegateValue - Insert instructions before the instruction pointed to by BI, 790/// that computes the negative version of the value specified. The negative 791/// version of the value is returned, and BI is left pointing at the instruction 792/// that should be processed next by the reassociation pass. 793static Value *NegateValue(Value *V, Instruction *BI) { 794 if (Constant *C = dyn_cast<Constant>(V)) 795 return ConstantExpr::getNeg(C); 796 797 // We are trying to expose opportunity for reassociation. One of the things 798 // that we want to do to achieve this is to push a negation as deep into an 799 // expression chain as possible, to expose the add instructions. In practice, 800 // this means that we turn this: 801 // X = -(A+12+C+D) into X = -A + -12 + -C + -D = -12 + -A + -C + -D 802 // so that later, a: Y = 12+X could get reassociated with the -12 to eliminate 803 // the constants. We assume that instcombine will clean up the mess later if 804 // we introduce tons of unnecessary negation instructions. 805 // 806 if (BinaryOperator *I = isReassociableOp(V, Instruction::Add)) { 807 // Push the negates through the add. 808 I->setOperand(0, NegateValue(I->getOperand(0), BI)); 809 I->setOperand(1, NegateValue(I->getOperand(1), BI)); 810 811 // We must move the add instruction here, because the neg instructions do 812 // not dominate the old add instruction in general. By moving it, we are 813 // assured that the neg instructions we just inserted dominate the 814 // instruction we are about to insert after them. 815 // 816 I->moveBefore(BI); 817 I->setName(I->getName()+".neg"); 818 return I; 819 } 820 821 // Okay, we need to materialize a negated version of V with an instruction. 822 // Scan the use lists of V to see if we have one already. 823 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E;++UI){ 824 User *U = *UI; 825 if (!BinaryOperator::isNeg(U)) continue; 826 827 // We found one! Now we have to make sure that the definition dominates 828 // this use. We do this by moving it to the entry block (if it is a 829 // non-instruction value) or right after the definition. These negates will 830 // be zapped by reassociate later, so we don't need much finesse here. 831 BinaryOperator *TheNeg = cast<BinaryOperator>(U); 832 833 // Verify that the negate is in this function, V might be a constant expr. 834 if (TheNeg->getParent()->getParent() != BI->getParent()->getParent()) 835 continue; 836 837 BasicBlock::iterator InsertPt; 838 if (Instruction *InstInput = dyn_cast<Instruction>(V)) { 839 if (InvokeInst *II = dyn_cast<InvokeInst>(InstInput)) { 840 InsertPt = II->getNormalDest()->begin(); 841 } else { 842 InsertPt = InstInput; 843 ++InsertPt; 844 } 845 while (isa<PHINode>(InsertPt)) ++InsertPt; 846 } else { 847 InsertPt = TheNeg->getParent()->getParent()->getEntryBlock().begin(); 848 } 849 TheNeg->moveBefore(InsertPt); 850 return TheNeg; 851 } 852 853 // Insert a 'neg' instruction that subtracts the value from zero to get the 854 // negation. 855 return BinaryOperator::CreateNeg(V, V->getName() + ".neg", BI); 856} 857 858/// ShouldBreakUpSubtract - Return true if we should break up this subtract of 859/// X-Y into (X + -Y). 860static bool ShouldBreakUpSubtract(Instruction *Sub) { 861 // If this is a negation, we can't split it up! 862 if (BinaryOperator::isNeg(Sub)) 863 return false; 864 865 // Don't bother to break this up unless either the LHS is an associable add or 866 // subtract or if this is only used by one. 867 if (isReassociableOp(Sub->getOperand(0), Instruction::Add) || 868 isReassociableOp(Sub->getOperand(0), Instruction::Sub)) 869 return true; 870 if (isReassociableOp(Sub->getOperand(1), Instruction::Add) || 871 isReassociableOp(Sub->getOperand(1), Instruction::Sub)) 872 return true; 873 if (Sub->hasOneUse() && 874 (isReassociableOp(Sub->use_back(), Instruction::Add) || 875 isReassociableOp(Sub->use_back(), Instruction::Sub))) 876 return true; 877 878 return false; 879} 880 881/// BreakUpSubtract - If we have (X-Y), and if either X is an add, or if this is 882/// only used by an add, transform this into (X+(0-Y)) to promote better 883/// reassociation. 884static BinaryOperator *BreakUpSubtract(Instruction *Sub) { 885 // Convert a subtract into an add and a neg instruction. This allows sub 886 // instructions to be commuted with other add instructions. 887 // 888 // Calculate the negative value of Operand 1 of the sub instruction, 889 // and set it as the RHS of the add instruction we just made. 890 // 891 Value *NegVal = NegateValue(Sub->getOperand(1), Sub); 892 BinaryOperator *New = 893 BinaryOperator::CreateAdd(Sub->getOperand(0), NegVal, "", Sub); 894 Sub->setOperand(0, Constant::getNullValue(Sub->getType())); // Drop use of op. 895 Sub->setOperand(1, Constant::getNullValue(Sub->getType())); // Drop use of op. 896 New->takeName(Sub); 897 898 // Everyone now refers to the add instruction. 899 Sub->replaceAllUsesWith(New); 900 New->setDebugLoc(Sub->getDebugLoc()); 901 902 DEBUG(dbgs() << "Negated: " << *New << '\n'); 903 return New; 904} 905 906/// ConvertShiftToMul - If this is a shift of a reassociable multiply or is used 907/// by one, change this into a multiply by a constant to assist with further 908/// reassociation. 909static BinaryOperator *ConvertShiftToMul(Instruction *Shl) { 910 Constant *MulCst = ConstantInt::get(Shl->getType(), 1); 911 MulCst = ConstantExpr::getShl(MulCst, cast<Constant>(Shl->getOperand(1))); 912 913 BinaryOperator *Mul = 914 BinaryOperator::CreateMul(Shl->getOperand(0), MulCst, "", Shl); 915 Shl->setOperand(0, UndefValue::get(Shl->getType())); // Drop use of op. 916 Mul->takeName(Shl); 917 Shl->replaceAllUsesWith(Mul); 918 Mul->setDebugLoc(Shl->getDebugLoc()); 919 return Mul; 920} 921 922/// FindInOperandList - Scan backwards and forwards among values with the same 923/// rank as element i to see if X exists. If X does not exist, return i. This 924/// is useful when scanning for 'x' when we see '-x' because they both get the 925/// same rank. 926static unsigned FindInOperandList(SmallVectorImpl<ValueEntry> &Ops, unsigned i, 927 Value *X) { 928 unsigned XRank = Ops[i].Rank; 929 unsigned e = Ops.size(); 930 for (unsigned j = i+1; j != e && Ops[j].Rank == XRank; ++j) 931 if (Ops[j].Op == X) 932 return j; 933 // Scan backwards. 934 for (unsigned j = i-1; j != ~0U && Ops[j].Rank == XRank; --j) 935 if (Ops[j].Op == X) 936 return j; 937 return i; 938} 939 940/// EmitAddTreeOfValues - Emit a tree of add instructions, summing Ops together 941/// and returning the result. Insert the tree before I. 942static Value *EmitAddTreeOfValues(Instruction *I, 943 SmallVectorImpl<WeakVH> &Ops){ 944 if (Ops.size() == 1) return Ops.back(); 945 946 Value *V1 = Ops.back(); 947 Ops.pop_back(); 948 Value *V2 = EmitAddTreeOfValues(I, Ops); 949 return BinaryOperator::CreateAdd(V2, V1, "tmp", I); 950} 951 952/// RemoveFactorFromExpression - If V is an expression tree that is a 953/// multiplication sequence, and if this sequence contains a multiply by Factor, 954/// remove Factor from the tree and return the new tree. 955Value *Reassociate::RemoveFactorFromExpression(Value *V, Value *Factor) { 956 BinaryOperator *BO = isReassociableOp(V, Instruction::Mul); 957 if (!BO) return 0; 958 959 SmallVector<RepeatedValue, 8> Tree; 960 MadeChange |= LinearizeExprTree(BO, Tree); 961 SmallVector<ValueEntry, 8> Factors; 962 Factors.reserve(Tree.size()); 963 for (unsigned i = 0, e = Tree.size(); i != e; ++i) { 964 RepeatedValue E = Tree[i]; 965 Factors.append(E.second.getZExtValue(), 966 ValueEntry(getRank(E.first), E.first)); 967 } 968 969 bool FoundFactor = false; 970 bool NeedsNegate = false; 971 for (unsigned i = 0, e = Factors.size(); i != e; ++i) { 972 if (Factors[i].Op == Factor) { 973 FoundFactor = true; 974 Factors.erase(Factors.begin()+i); 975 break; 976 } 977 978 // If this is a negative version of this factor, remove it. 979 if (ConstantInt *FC1 = dyn_cast<ConstantInt>(Factor)) 980 if (ConstantInt *FC2 = dyn_cast<ConstantInt>(Factors[i].Op)) 981 if (FC1->getValue() == -FC2->getValue()) { 982 FoundFactor = NeedsNegate = true; 983 Factors.erase(Factors.begin()+i); 984 break; 985 } 986 } 987 988 if (!FoundFactor) { 989 // Make sure to restore the operands to the expression tree. 990 RewriteExprTree(BO, Factors); 991 return 0; 992 } 993 994 BasicBlock::iterator InsertPt = BO; ++InsertPt; 995 996 // If this was just a single multiply, remove the multiply and return the only 997 // remaining operand. 998 if (Factors.size() == 1) { 999 RedoInsts.insert(BO); 1000 V = Factors[0].Op; 1001 } else { 1002 RewriteExprTree(BO, Factors); 1003 V = BO; 1004 } 1005 1006 if (NeedsNegate) 1007 V = BinaryOperator::CreateNeg(V, "neg", InsertPt); 1008 1009 return V; 1010} 1011 1012/// FindSingleUseMultiplyFactors - If V is a single-use multiply, recursively 1013/// add its operands as factors, otherwise add V to the list of factors. 1014/// 1015/// Ops is the top-level list of add operands we're trying to factor. 1016static void FindSingleUseMultiplyFactors(Value *V, 1017 SmallVectorImpl<Value*> &Factors, 1018 const SmallVectorImpl<ValueEntry> &Ops) { 1019 BinaryOperator *BO = isReassociableOp(V, Instruction::Mul); 1020 if (!BO) { 1021 Factors.push_back(V); 1022 return; 1023 } 1024 1025 // Otherwise, add the LHS and RHS to the list of factors. 1026 FindSingleUseMultiplyFactors(BO->getOperand(1), Factors, Ops); 1027 FindSingleUseMultiplyFactors(BO->getOperand(0), Factors, Ops); 1028} 1029 1030/// OptimizeAndOrXor - Optimize a series of operands to an 'and', 'or', or 'xor' 1031/// instruction. This optimizes based on identities. If it can be reduced to 1032/// a single Value, it is returned, otherwise the Ops list is mutated as 1033/// necessary. 1034static Value *OptimizeAndOrXor(unsigned Opcode, 1035 SmallVectorImpl<ValueEntry> &Ops) { 1036 // Scan the operand lists looking for X and ~X pairs, along with X,X pairs. 1037 // If we find any, we can simplify the expression. X&~X == 0, X|~X == -1. 1038 for (unsigned i = 0, e = Ops.size(); i != e; ++i) { 1039 // First, check for X and ~X in the operand list. 1040 assert(i < Ops.size()); 1041 if (BinaryOperator::isNot(Ops[i].Op)) { // Cannot occur for ^. 1042 Value *X = BinaryOperator::getNotArgument(Ops[i].Op); 1043 unsigned FoundX = FindInOperandList(Ops, i, X); 1044 if (FoundX != i) { 1045 if (Opcode == Instruction::And) // ...&X&~X = 0 1046 return Constant::getNullValue(X->getType()); 1047 1048 if (Opcode == Instruction::Or) // ...|X|~X = -1 1049 return Constant::getAllOnesValue(X->getType()); 1050 } 1051 } 1052 1053 // Next, check for duplicate pairs of values, which we assume are next to 1054 // each other, due to our sorting criteria. 1055 assert(i < Ops.size()); 1056 if (i+1 != Ops.size() && Ops[i+1].Op == Ops[i].Op) { 1057 if (Opcode == Instruction::And || Opcode == Instruction::Or) { 1058 // Drop duplicate values for And and Or. 1059 Ops.erase(Ops.begin()+i); 1060 --i; --e; 1061 ++NumAnnihil; 1062 continue; 1063 } 1064 1065 // Drop pairs of values for Xor. 1066 assert(Opcode == Instruction::Xor); 1067 if (e == 2) 1068 return Constant::getNullValue(Ops[0].Op->getType()); 1069 1070 // Y ^ X^X -> Y 1071 Ops.erase(Ops.begin()+i, Ops.begin()+i+2); 1072 i -= 1; e -= 2; 1073 ++NumAnnihil; 1074 } 1075 } 1076 return 0; 1077} 1078 1079/// OptimizeAdd - Optimize a series of operands to an 'add' instruction. This 1080/// optimizes based on identities. If it can be reduced to a single Value, it 1081/// is returned, otherwise the Ops list is mutated as necessary. 1082Value *Reassociate::OptimizeAdd(Instruction *I, 1083 SmallVectorImpl<ValueEntry> &Ops) { 1084 // Scan the operand lists looking for X and -X pairs. If we find any, we 1085 // can simplify the expression. X+-X == 0. While we're at it, scan for any 1086 // duplicates. We want to canonicalize Y+Y+Y+Z -> 3*Y+Z. 1087 // 1088 // TODO: We could handle "X + ~X" -> "-1" if we wanted, since "-X = ~X+1". 1089 // 1090 for (unsigned i = 0, e = Ops.size(); i != e; ++i) { 1091 Value *TheOp = Ops[i].Op; 1092 // Check to see if we've seen this operand before. If so, we factor all 1093 // instances of the operand together. Due to our sorting criteria, we know 1094 // that these need to be next to each other in the vector. 1095 if (i+1 != Ops.size() && Ops[i+1].Op == TheOp) { 1096 // Rescan the list, remove all instances of this operand from the expr. 1097 unsigned NumFound = 0; 1098 do { 1099 Ops.erase(Ops.begin()+i); 1100 ++NumFound; 1101 } while (i != Ops.size() && Ops[i].Op == TheOp); 1102 1103 DEBUG(errs() << "\nFACTORING [" << NumFound << "]: " << *TheOp << '\n'); 1104 ++NumFactor; 1105 1106 // Insert a new multiply. 1107 Value *Mul = ConstantInt::get(cast<IntegerType>(I->getType()), NumFound); 1108 Mul = BinaryOperator::CreateMul(TheOp, Mul, "factor", I); 1109 1110 // Now that we have inserted a multiply, optimize it. This allows us to 1111 // handle cases that require multiple factoring steps, such as this: 1112 // (X*2) + (X*2) + (X*2) -> (X*2)*3 -> X*6 1113 RedoInsts.insert(cast<Instruction>(Mul)); 1114 1115 // If every add operand was a duplicate, return the multiply. 1116 if (Ops.empty()) 1117 return Mul; 1118 1119 // Otherwise, we had some input that didn't have the dupe, such as 1120 // "A + A + B" -> "A*2 + B". Add the new multiply to the list of 1121 // things being added by this operation. 1122 Ops.insert(Ops.begin(), ValueEntry(getRank(Mul), Mul)); 1123 1124 --i; 1125 e = Ops.size(); 1126 continue; 1127 } 1128 1129 // Check for X and -X in the operand list. 1130 if (!BinaryOperator::isNeg(TheOp)) 1131 continue; 1132 1133 Value *X = BinaryOperator::getNegArgument(TheOp); 1134 unsigned FoundX = FindInOperandList(Ops, i, X); 1135 if (FoundX == i) 1136 continue; 1137 1138 // Remove X and -X from the operand list. 1139 if (Ops.size() == 2) 1140 return Constant::getNullValue(X->getType()); 1141 1142 Ops.erase(Ops.begin()+i); 1143 if (i < FoundX) 1144 --FoundX; 1145 else 1146 --i; // Need to back up an extra one. 1147 Ops.erase(Ops.begin()+FoundX); 1148 ++NumAnnihil; 1149 --i; // Revisit element. 1150 e -= 2; // Removed two elements. 1151 } 1152 1153 // Scan the operand list, checking to see if there are any common factors 1154 // between operands. Consider something like A*A+A*B*C+D. We would like to 1155 // reassociate this to A*(A+B*C)+D, which reduces the number of multiplies. 1156 // To efficiently find this, we count the number of times a factor occurs 1157 // for any ADD operands that are MULs. 1158 DenseMap<Value*, unsigned> FactorOccurrences; 1159 1160 // Keep track of each multiply we see, to avoid triggering on (X*4)+(X*4) 1161 // where they are actually the same multiply. 1162 unsigned MaxOcc = 0; 1163 Value *MaxOccVal = 0; 1164 for (unsigned i = 0, e = Ops.size(); i != e; ++i) { 1165 BinaryOperator *BOp = isReassociableOp(Ops[i].Op, Instruction::Mul); 1166 if (!BOp) 1167 continue; 1168 1169 // Compute all of the factors of this added value. 1170 SmallVector<Value*, 8> Factors; 1171 FindSingleUseMultiplyFactors(BOp, Factors, Ops); 1172 assert(Factors.size() > 1 && "Bad linearize!"); 1173 1174 // Add one to FactorOccurrences for each unique factor in this op. 1175 SmallPtrSet<Value*, 8> Duplicates; 1176 for (unsigned i = 0, e = Factors.size(); i != e; ++i) { 1177 Value *Factor = Factors[i]; 1178 if (!Duplicates.insert(Factor)) continue; 1179 1180 unsigned Occ = ++FactorOccurrences[Factor]; 1181 if (Occ > MaxOcc) { MaxOcc = Occ; MaxOccVal = Factor; } 1182 1183 // If Factor is a negative constant, add the negated value as a factor 1184 // because we can percolate the negate out. Watch for minint, which 1185 // cannot be positivified. 1186 if (ConstantInt *CI = dyn_cast<ConstantInt>(Factor)) 1187 if (CI->isNegative() && !CI->isMinValue(true)) { 1188 Factor = ConstantInt::get(CI->getContext(), -CI->getValue()); 1189 assert(!Duplicates.count(Factor) && 1190 "Shouldn't have two constant factors, missed a canonicalize"); 1191 1192 unsigned Occ = ++FactorOccurrences[Factor]; 1193 if (Occ > MaxOcc) { MaxOcc = Occ; MaxOccVal = Factor; } 1194 } 1195 } 1196 } 1197 1198 // If any factor occurred more than one time, we can pull it out. 1199 if (MaxOcc > 1) { 1200 DEBUG(errs() << "\nFACTORING [" << MaxOcc << "]: " << *MaxOccVal << '\n'); 1201 ++NumFactor; 1202 1203 // Create a new instruction that uses the MaxOccVal twice. If we don't do 1204 // this, we could otherwise run into situations where removing a factor 1205 // from an expression will drop a use of maxocc, and this can cause 1206 // RemoveFactorFromExpression on successive values to behave differently. 1207 Instruction *DummyInst = BinaryOperator::CreateAdd(MaxOccVal, MaxOccVal); 1208 SmallVector<WeakVH, 4> NewMulOps; 1209 for (unsigned i = 0; i != Ops.size(); ++i) { 1210 // Only try to remove factors from expressions we're allowed to. 1211 BinaryOperator *BOp = isReassociableOp(Ops[i].Op, Instruction::Mul); 1212 if (!BOp) 1213 continue; 1214 1215 if (Value *V = RemoveFactorFromExpression(Ops[i].Op, MaxOccVal)) { 1216 // The factorized operand may occur several times. Convert them all in 1217 // one fell swoop. 1218 for (unsigned j = Ops.size(); j != i;) { 1219 --j; 1220 if (Ops[j].Op == Ops[i].Op) { 1221 NewMulOps.push_back(V); 1222 Ops.erase(Ops.begin()+j); 1223 } 1224 } 1225 --i; 1226 } 1227 } 1228 1229 // No need for extra uses anymore. 1230 delete DummyInst; 1231 1232 unsigned NumAddedValues = NewMulOps.size(); 1233 Value *V = EmitAddTreeOfValues(I, NewMulOps); 1234 1235 // Now that we have inserted the add tree, optimize it. This allows us to 1236 // handle cases that require multiple factoring steps, such as this: 1237 // A*A*B + A*A*C --> A*(A*B+A*C) --> A*(A*(B+C)) 1238 assert(NumAddedValues > 1 && "Each occurrence should contribute a value"); 1239 (void)NumAddedValues; 1240 if (Instruction *VI = dyn_cast<Instruction>(V)) 1241 RedoInsts.insert(VI); 1242 1243 // Create the multiply. 1244 Instruction *V2 = BinaryOperator::CreateMul(V, MaxOccVal, "tmp", I); 1245 1246 // Rerun associate on the multiply in case the inner expression turned into 1247 // a multiply. We want to make sure that we keep things in canonical form. 1248 RedoInsts.insert(V2); 1249 1250 // If every add operand included the factor (e.g. "A*B + A*C"), then the 1251 // entire result expression is just the multiply "A*(B+C)". 1252 if (Ops.empty()) 1253 return V2; 1254 1255 // Otherwise, we had some input that didn't have the factor, such as 1256 // "A*B + A*C + D" -> "A*(B+C) + D". Add the new multiply to the list of 1257 // things being added by this operation. 1258 Ops.insert(Ops.begin(), ValueEntry(getRank(V2), V2)); 1259 } 1260 1261 return 0; 1262} 1263 1264namespace { 1265 /// \brief Predicate tests whether a ValueEntry's op is in a map. 1266 struct IsValueInMap { 1267 const DenseMap<Value *, unsigned> ⤅ 1268 1269 IsValueInMap(const DenseMap<Value *, unsigned> &Map) : Map(Map) {} 1270 1271 bool operator()(const ValueEntry &Entry) { 1272 return Map.find(Entry.Op) != Map.end(); 1273 } 1274 }; 1275} 1276 1277/// \brief Build up a vector of value/power pairs factoring a product. 1278/// 1279/// Given a series of multiplication operands, build a vector of factors and 1280/// the powers each is raised to when forming the final product. Sort them in 1281/// the order of descending power. 1282/// 1283/// (x*x) -> [(x, 2)] 1284/// ((x*x)*x) -> [(x, 3)] 1285/// ((((x*y)*x)*y)*x) -> [(x, 3), (y, 2)] 1286/// 1287/// \returns Whether any factors have a power greater than one. 1288bool Reassociate::collectMultiplyFactors(SmallVectorImpl<ValueEntry> &Ops, 1289 SmallVectorImpl<Factor> &Factors) { 1290 // FIXME: Have Ops be (ValueEntry, Multiplicity) pairs, simplifying this. 1291 // Compute the sum of powers of simplifiable factors. 1292 unsigned FactorPowerSum = 0; 1293 for (unsigned Idx = 1, Size = Ops.size(); Idx < Size; ++Idx) { 1294 Value *Op = Ops[Idx-1].Op; 1295 1296 // Count the number of occurrences of this value. 1297 unsigned Count = 1; 1298 for (; Idx < Size && Ops[Idx].Op == Op; ++Idx) 1299 ++Count; 1300 // Track for simplification all factors which occur 2 or more times. 1301 if (Count > 1) 1302 FactorPowerSum += Count; 1303 } 1304 1305 // We can only simplify factors if the sum of the powers of our simplifiable 1306 // factors is 4 or higher. When that is the case, we will *always* have 1307 // a simplification. This is an important invariant to prevent cyclicly 1308 // trying to simplify already minimal formations. 1309 if (FactorPowerSum < 4) 1310 return false; 1311 1312 // Now gather the simplifiable factors, removing them from Ops. 1313 FactorPowerSum = 0; 1314 for (unsigned Idx = 1; Idx < Ops.size(); ++Idx) { 1315 Value *Op = Ops[Idx-1].Op; 1316 1317 // Count the number of occurrences of this value. 1318 unsigned Count = 1; 1319 for (; Idx < Ops.size() && Ops[Idx].Op == Op; ++Idx) 1320 ++Count; 1321 if (Count == 1) 1322 continue; 1323 // Move an even number of occurrences to Factors. 1324 Count &= ~1U; 1325 Idx -= Count; 1326 FactorPowerSum += Count; 1327 Factors.push_back(Factor(Op, Count)); 1328 Ops.erase(Ops.begin()+Idx, Ops.begin()+Idx+Count); 1329 } 1330 1331 // None of the adjustments above should have reduced the sum of factor powers 1332 // below our mininum of '4'. 1333 assert(FactorPowerSum >= 4); 1334 1335 std::sort(Factors.begin(), Factors.end(), Factor::PowerDescendingSorter()); 1336 return true; 1337} 1338 1339/// \brief Build a tree of multiplies, computing the product of Ops. 1340static Value *buildMultiplyTree(IRBuilder<> &Builder, 1341 SmallVectorImpl<Value*> &Ops) { 1342 if (Ops.size() == 1) 1343 return Ops.back(); 1344 1345 Value *LHS = Ops.pop_back_val(); 1346 do { 1347 LHS = Builder.CreateMul(LHS, Ops.pop_back_val()); 1348 } while (!Ops.empty()); 1349 1350 return LHS; 1351} 1352 1353/// \brief Build a minimal multiplication DAG for (a^x)*(b^y)*(c^z)*... 1354/// 1355/// Given a vector of values raised to various powers, where no two values are 1356/// equal and the powers are sorted in decreasing order, compute the minimal 1357/// DAG of multiplies to compute the final product, and return that product 1358/// value. 1359Value *Reassociate::buildMinimalMultiplyDAG(IRBuilder<> &Builder, 1360 SmallVectorImpl<Factor> &Factors) { 1361 assert(Factors[0].Power); 1362 SmallVector<Value *, 4> OuterProduct; 1363 for (unsigned LastIdx = 0, Idx = 1, Size = Factors.size(); 1364 Idx < Size && Factors[Idx].Power > 0; ++Idx) { 1365 if (Factors[Idx].Power != Factors[LastIdx].Power) { 1366 LastIdx = Idx; 1367 continue; 1368 } 1369 1370 // We want to multiply across all the factors with the same power so that 1371 // we can raise them to that power as a single entity. Build a mini tree 1372 // for that. 1373 SmallVector<Value *, 4> InnerProduct; 1374 InnerProduct.push_back(Factors[LastIdx].Base); 1375 do { 1376 InnerProduct.push_back(Factors[Idx].Base); 1377 ++Idx; 1378 } while (Idx < Size && Factors[Idx].Power == Factors[LastIdx].Power); 1379 1380 // Reset the base value of the first factor to the new expression tree. 1381 // We'll remove all the factors with the same power in a second pass. 1382 Value *M = Factors[LastIdx].Base = buildMultiplyTree(Builder, InnerProduct); 1383 if (Instruction *MI = dyn_cast<Instruction>(M)) 1384 RedoInsts.insert(MI); 1385 1386 LastIdx = Idx; 1387 } 1388 // Unique factors with equal powers -- we've folded them into the first one's 1389 // base. 1390 Factors.erase(std::unique(Factors.begin(), Factors.end(), 1391 Factor::PowerEqual()), 1392 Factors.end()); 1393 1394 // Iteratively collect the base of each factor with an add power into the 1395 // outer product, and halve each power in preparation for squaring the 1396 // expression. 1397 for (unsigned Idx = 0, Size = Factors.size(); Idx != Size; ++Idx) { 1398 if (Factors[Idx].Power & 1) 1399 OuterProduct.push_back(Factors[Idx].Base); 1400 Factors[Idx].Power >>= 1; 1401 } 1402 if (Factors[0].Power) { 1403 Value *SquareRoot = buildMinimalMultiplyDAG(Builder, Factors); 1404 OuterProduct.push_back(SquareRoot); 1405 OuterProduct.push_back(SquareRoot); 1406 } 1407 if (OuterProduct.size() == 1) 1408 return OuterProduct.front(); 1409 1410 Value *V = buildMultiplyTree(Builder, OuterProduct); 1411 return V; 1412} 1413 1414Value *Reassociate::OptimizeMul(BinaryOperator *I, 1415 SmallVectorImpl<ValueEntry> &Ops) { 1416 // We can only optimize the multiplies when there is a chain of more than 1417 // three, such that a balanced tree might require fewer total multiplies. 1418 if (Ops.size() < 4) 1419 return 0; 1420 1421 // Try to turn linear trees of multiplies without other uses of the 1422 // intermediate stages into minimal multiply DAGs with perfect sub-expression 1423 // re-use. 1424 SmallVector<Factor, 4> Factors; 1425 if (!collectMultiplyFactors(Ops, Factors)) 1426 return 0; // All distinct factors, so nothing left for us to do. 1427 1428 IRBuilder<> Builder(I); 1429 Value *V = buildMinimalMultiplyDAG(Builder, Factors); 1430 if (Ops.empty()) 1431 return V; 1432 1433 ValueEntry NewEntry = ValueEntry(getRank(V), V); 1434 Ops.insert(std::lower_bound(Ops.begin(), Ops.end(), NewEntry), NewEntry); 1435 return 0; 1436} 1437 1438Value *Reassociate::OptimizeExpression(BinaryOperator *I, 1439 SmallVectorImpl<ValueEntry> &Ops) { 1440 // Now that we have the linearized expression tree, try to optimize it. 1441 // Start by folding any constants that we found. 1442 if (Ops.size() == 1) return Ops[0].Op; 1443 1444 unsigned Opcode = I->getOpcode(); 1445 1446 // Handle destructive annihilation due to identities between elements in the 1447 // argument list here. 1448 unsigned NumOps = Ops.size(); 1449 switch (Opcode) { 1450 default: break; 1451 case Instruction::And: 1452 case Instruction::Or: 1453 case Instruction::Xor: 1454 if (Value *Result = OptimizeAndOrXor(Opcode, Ops)) 1455 return Result; 1456 break; 1457 1458 case Instruction::Add: 1459 if (Value *Result = OptimizeAdd(I, Ops)) 1460 return Result; 1461 break; 1462 1463 case Instruction::Mul: 1464 if (Value *Result = OptimizeMul(I, Ops)) 1465 return Result; 1466 break; 1467 } 1468 1469 if (Ops.size() != NumOps) 1470 return OptimizeExpression(I, Ops); 1471 return 0; 1472} 1473 1474/// EraseInst - Zap the given instruction, adding interesting operands to the 1475/// work list. 1476void Reassociate::EraseInst(Instruction *I) { 1477 assert(isInstructionTriviallyDead(I) && "Trivially dead instructions only!"); 1478 SmallVector<Value*, 8> Ops(I->op_begin(), I->op_end()); 1479 // Erase the dead instruction. 1480 ValueRankMap.erase(I); 1481 I->eraseFromParent(); 1482 // Optimize its operands. 1483 for (unsigned i = 0, e = Ops.size(); i != e; ++i) 1484 if (Instruction *Op = dyn_cast<Instruction>(Ops[i])) { 1485 // If this is a node in an expression tree, climb to the expression root 1486 // and add that since that's where optimization actually happens. 1487 unsigned Opcode = Op->getOpcode(); 1488 while (Op->hasOneUse() && Op->use_back()->getOpcode() == Opcode) 1489 Op = Op->use_back(); 1490 RedoInsts.insert(Op); 1491 } 1492} 1493 1494/// OptimizeInst - Inspect and optimize the given instruction. Note that erasing 1495/// instructions is not allowed. 1496void Reassociate::OptimizeInst(Instruction *I) { 1497 // Only consider operations that we understand. 1498 if (!isa<BinaryOperator>(I)) 1499 return; 1500 1501 if (I->getOpcode() == Instruction::Shl && 1502 isa<ConstantInt>(I->getOperand(1))) 1503 // If an operand of this shift is a reassociable multiply, or if the shift 1504 // is used by a reassociable multiply or add, turn into a multiply. 1505 if (isReassociableOp(I->getOperand(0), Instruction::Mul) || 1506 (I->hasOneUse() && 1507 (isReassociableOp(I->use_back(), Instruction::Mul) || 1508 isReassociableOp(I->use_back(), Instruction::Add)))) { 1509 Instruction *NI = ConvertShiftToMul(I); 1510 RedoInsts.insert(I); 1511 MadeChange = true; 1512 I = NI; 1513 } 1514 1515 // Floating point binary operators are not associative, but we can still 1516 // commute (some) of them, to canonicalize the order of their operands. 1517 // This can potentially expose more CSE opportunities, and makes writing 1518 // other transformations simpler. 1519 if ((I->getType()->isFloatingPointTy() || I->getType()->isVectorTy())) { 1520 // FAdd and FMul can be commuted. 1521 if (I->getOpcode() != Instruction::FMul && 1522 I->getOpcode() != Instruction::FAdd) 1523 return; 1524 1525 Value *LHS = I->getOperand(0); 1526 Value *RHS = I->getOperand(1); 1527 unsigned LHSRank = getRank(LHS); 1528 unsigned RHSRank = getRank(RHS); 1529 1530 // Sort the operands by rank. 1531 if (RHSRank < LHSRank) { 1532 I->setOperand(0, RHS); 1533 I->setOperand(1, LHS); 1534 } 1535 1536 return; 1537 } 1538 1539 // Do not reassociate boolean (i1) expressions. We want to preserve the 1540 // original order of evaluation for short-circuited comparisons that 1541 // SimplifyCFG has folded to AND/OR expressions. If the expression 1542 // is not further optimized, it is likely to be transformed back to a 1543 // short-circuited form for code gen, and the source order may have been 1544 // optimized for the most likely conditions. 1545 if (I->getType()->isIntegerTy(1)) 1546 return; 1547 1548 // If this is a subtract instruction which is not already in negate form, 1549 // see if we can convert it to X+-Y. 1550 if (I->getOpcode() == Instruction::Sub) { 1551 if (ShouldBreakUpSubtract(I)) { 1552 Instruction *NI = BreakUpSubtract(I); 1553 RedoInsts.insert(I); 1554 MadeChange = true; 1555 I = NI; 1556 } else if (BinaryOperator::isNeg(I)) { 1557 // Otherwise, this is a negation. See if the operand is a multiply tree 1558 // and if this is not an inner node of a multiply tree. 1559 if (isReassociableOp(I->getOperand(1), Instruction::Mul) && 1560 (!I->hasOneUse() || 1561 !isReassociableOp(I->use_back(), Instruction::Mul))) { 1562 Instruction *NI = LowerNegateToMultiply(I); 1563 RedoInsts.insert(I); 1564 MadeChange = true; 1565 I = NI; 1566 } 1567 } 1568 } 1569 1570 // If this instruction is an associative binary operator, process it. 1571 if (!I->isAssociative()) return; 1572 BinaryOperator *BO = cast<BinaryOperator>(I); 1573 1574 // If this is an interior node of a reassociable tree, ignore it until we 1575 // get to the root of the tree, to avoid N^2 analysis. 1576 if (BO->hasOneUse() && BO->use_back()->getOpcode() == BO->getOpcode()) 1577 return; 1578 1579 // If this is an add tree that is used by a sub instruction, ignore it 1580 // until we process the subtract. 1581 if (BO->hasOneUse() && BO->getOpcode() == Instruction::Add && 1582 cast<Instruction>(BO->use_back())->getOpcode() == Instruction::Sub) 1583 return; 1584 1585 ReassociateExpression(BO); 1586} 1587 1588Value *Reassociate::ReassociateExpression(BinaryOperator *I) { 1589 1590 // First, walk the expression tree, linearizing the tree, collecting the 1591 // operand information. 1592 SmallVector<RepeatedValue, 8> Tree; 1593 MadeChange |= LinearizeExprTree(I, Tree); 1594 SmallVector<ValueEntry, 8> Ops; 1595 Ops.reserve(Tree.size()); 1596 for (unsigned i = 0, e = Tree.size(); i != e; ++i) { 1597 RepeatedValue E = Tree[i]; 1598 Ops.append(E.second.getZExtValue(), 1599 ValueEntry(getRank(E.first), E.first)); 1600 } 1601 1602 DEBUG(dbgs() << "RAIn:\t"; PrintOps(I, Ops); dbgs() << '\n'); 1603 1604 // Now that we have linearized the tree to a list and have gathered all of 1605 // the operands and their ranks, sort the operands by their rank. Use a 1606 // stable_sort so that values with equal ranks will have their relative 1607 // positions maintained (and so the compiler is deterministic). Note that 1608 // this sorts so that the highest ranking values end up at the beginning of 1609 // the vector. 1610 std::stable_sort(Ops.begin(), Ops.end()); 1611 1612 // OptimizeExpression - Now that we have the expression tree in a convenient 1613 // sorted form, optimize it globally if possible. 1614 if (Value *V = OptimizeExpression(I, Ops)) { 1615 // This expression tree simplified to something that isn't a tree, 1616 // eliminate it. 1617 DEBUG(dbgs() << "Reassoc to scalar: " << *V << '\n'); 1618 I->replaceAllUsesWith(V); 1619 if (Instruction *VI = dyn_cast<Instruction>(V)) 1620 VI->setDebugLoc(I->getDebugLoc()); 1621 RedoInsts.insert(I); 1622 ++NumAnnihil; 1623 return V; 1624 } 1625 1626 // We want to sink immediates as deeply as possible except in the case where 1627 // this is a multiply tree used only by an add, and the immediate is a -1. 1628 // In this case we reassociate to put the negation on the outside so that we 1629 // can fold the negation into the add: (-X)*Y + Z -> Z-X*Y 1630 if (I->getOpcode() == Instruction::Mul && I->hasOneUse() && 1631 cast<Instruction>(I->use_back())->getOpcode() == Instruction::Add && 1632 isa<ConstantInt>(Ops.back().Op) && 1633 cast<ConstantInt>(Ops.back().Op)->isAllOnesValue()) { 1634 ValueEntry Tmp = Ops.pop_back_val(); 1635 Ops.insert(Ops.begin(), Tmp); 1636 } 1637 1638 DEBUG(dbgs() << "RAOut:\t"; PrintOps(I, Ops); dbgs() << '\n'); 1639 1640 if (Ops.size() == 1) { 1641 // This expression tree simplified to something that isn't a tree, 1642 // eliminate it. 1643 I->replaceAllUsesWith(Ops[0].Op); 1644 if (Instruction *OI = dyn_cast<Instruction>(Ops[0].Op)) 1645 OI->setDebugLoc(I->getDebugLoc()); 1646 RedoInsts.insert(I); 1647 return Ops[0].Op; 1648 } 1649 1650 // Now that we ordered and optimized the expressions, splat them back into 1651 // the expression tree, removing any unneeded nodes. 1652 RewriteExprTree(I, Ops); 1653 return I; 1654} 1655 1656bool Reassociate::runOnFunction(Function &F) { 1657 // Calculate the rank map for F 1658 BuildRankMap(F); 1659 1660 MadeChange = false; 1661 for (Function::iterator BI = F.begin(), BE = F.end(); BI != BE; ++BI) { 1662 // Optimize every instruction in the basic block. 1663 for (BasicBlock::iterator II = BI->begin(), IE = BI->end(); II != IE; ) 1664 if (isInstructionTriviallyDead(II)) { 1665 EraseInst(II++); 1666 } else { 1667 OptimizeInst(II); 1668 assert(II->getParent() == BI && "Moved to a different block!"); 1669 ++II; 1670 } 1671 1672 // If this produced extra instructions to optimize, handle them now. 1673 while (!RedoInsts.empty()) { 1674 Instruction *I = RedoInsts.pop_back_val(); 1675 if (isInstructionTriviallyDead(I)) 1676 EraseInst(I); 1677 else 1678 OptimizeInst(I); 1679 } 1680 } 1681 1682 // We are done with the rank map. 1683 RankMap.clear(); 1684 ValueRankMap.clear(); 1685 1686 return MadeChange; 1687} 1688