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