Reassociate.cpp revision 36b56886974eae4f9c5ebc96befd3e7bfe5de338
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/IR/CFG.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/IR/ValueHandle.h" 38#include "llvm/Pass.h" 39#include "llvm/Support/Debug.h" 40#include "llvm/Support/raw_ostream.h" 41#include "llvm/Transforms/Utils/Local.h" 42#include <algorithm> 43using namespace llvm; 44 45STATISTIC(NumChanged, "Number of insts reassociated"); 46STATISTIC(NumAnnihil, "Number of expr tree annihilated"); 47STATISTIC(NumFactor , "Number of multiplies factored"); 48 49namespace { 50 struct ValueEntry { 51 unsigned Rank; 52 Value *Op; 53 ValueEntry(unsigned R, Value *O) : Rank(R), Op(O) {} 54 }; 55 inline bool operator<(const ValueEntry &LHS, const ValueEntry &RHS) { 56 return LHS.Rank > RHS.Rank; // Sort so that highest rank goes to start. 57 } 58} 59 60#ifndef NDEBUG 61/// PrintOps - Print out the expression identified in the Ops list. 62/// 63static void PrintOps(Instruction *I, const SmallVectorImpl<ValueEntry> &Ops) { 64 Module *M = I->getParent()->getParent()->getParent(); 65 dbgs() << Instruction::getOpcodeName(I->getOpcode()) << " " 66 << *Ops[0].Op->getType() << '\t'; 67 for (unsigned i = 0, e = Ops.size(); i != e; ++i) { 68 dbgs() << "[ "; 69 Ops[i].Op->printAsOperand(dbgs(), false, M); 70 dbgs() << ", #" << Ops[i].Rank << "] "; 71 } 72} 73#endif 74 75namespace { 76 /// \brief Utility class representing a base and exponent pair which form one 77 /// factor of some product. 78 struct Factor { 79 Value *Base; 80 unsigned Power; 81 82 Factor(Value *Base, unsigned Power) : Base(Base), Power(Power) {} 83 84 /// \brief Sort factors by their Base. 85 struct BaseSorter { 86 bool operator()(const Factor &LHS, const Factor &RHS) { 87 return LHS.Base < RHS.Base; 88 } 89 }; 90 91 /// \brief Compare factors for equal bases. 92 struct BaseEqual { 93 bool operator()(const Factor &LHS, const Factor &RHS) { 94 return LHS.Base == RHS.Base; 95 } 96 }; 97 98 /// \brief Sort factors in descending order by their power. 99 struct PowerDescendingSorter { 100 bool operator()(const Factor &LHS, const Factor &RHS) { 101 return LHS.Power > RHS.Power; 102 } 103 }; 104 105 /// \brief Compare factors for equal powers. 106 struct PowerEqual { 107 bool operator()(const Factor &LHS, const Factor &RHS) { 108 return LHS.Power == RHS.Power; 109 } 110 }; 111 }; 112 113 /// Utility class representing a non-constant Xor-operand. We classify 114 /// non-constant Xor-Operands into two categories: 115 /// C1) The operand is in the form "X & C", where C is a constant and C != ~0 116 /// C2) 117 /// C2.1) The operand is in the form of "X | C", where C is a non-zero 118 /// constant. 119 /// C2.2) Any operand E which doesn't fall into C1 and C2.1, we view this 120 /// operand as "E | 0" 121 class XorOpnd { 122 public: 123 XorOpnd(Value *V); 124 125 bool isInvalid() const { return SymbolicPart == 0; } 126 bool isOrExpr() const { return isOr; } 127 Value *getValue() const { return OrigVal; } 128 Value *getSymbolicPart() const { return SymbolicPart; } 129 unsigned getSymbolicRank() const { return SymbolicRank; } 130 const APInt &getConstPart() const { return ConstPart; } 131 132 void Invalidate() { SymbolicPart = OrigVal = 0; } 133 void setSymbolicRank(unsigned R) { SymbolicRank = R; } 134 135 // Sort the XorOpnd-Pointer in ascending order of symbolic-value-rank. 136 // The purpose is twofold: 137 // 1) Cluster together the operands sharing the same symbolic-value. 138 // 2) Operand having smaller symbolic-value-rank is permuted earlier, which 139 // could potentially shorten crital path, and expose more loop-invariants. 140 // Note that values' rank are basically defined in RPO order (FIXME). 141 // So, if Rank(X) < Rank(Y) < Rank(Z), it means X is defined earlier 142 // than Y which is defined earlier than Z. Permute "x | 1", "Y & 2", 143 // "z" in the order of X-Y-Z is better than any other orders. 144 struct PtrSortFunctor { 145 bool operator()(XorOpnd * const &LHS, XorOpnd * const &RHS) { 146 return LHS->getSymbolicRank() < RHS->getSymbolicRank(); 147 } 148 }; 149 private: 150 Value *OrigVal; 151 Value *SymbolicPart; 152 APInt ConstPart; 153 unsigned SymbolicRank; 154 bool isOr; 155 }; 156} 157 158namespace { 159 class Reassociate : public FunctionPass { 160 DenseMap<BasicBlock*, unsigned> RankMap; 161 DenseMap<AssertingVH<Value>, unsigned> ValueRankMap; 162 SetVector<AssertingVH<Instruction> > RedoInsts; 163 bool MadeChange; 164 public: 165 static char ID; // Pass identification, replacement for typeid 166 Reassociate() : FunctionPass(ID) { 167 initializeReassociatePass(*PassRegistry::getPassRegistry()); 168 } 169 170 bool runOnFunction(Function &F) override; 171 172 void getAnalysisUsage(AnalysisUsage &AU) const override { 173 AU.setPreservesCFG(); 174 } 175 private: 176 void BuildRankMap(Function &F); 177 unsigned getRank(Value *V); 178 void ReassociateExpression(BinaryOperator *I); 179 void RewriteExprTree(BinaryOperator *I, SmallVectorImpl<ValueEntry> &Ops); 180 Value *OptimizeExpression(BinaryOperator *I, 181 SmallVectorImpl<ValueEntry> &Ops); 182 Value *OptimizeAdd(Instruction *I, SmallVectorImpl<ValueEntry> &Ops); 183 Value *OptimizeXor(Instruction *I, SmallVectorImpl<ValueEntry> &Ops); 184 bool CombineXorOpnd(Instruction *I, XorOpnd *Opnd1, APInt &ConstOpnd, 185 Value *&Res); 186 bool CombineXorOpnd(Instruction *I, XorOpnd *Opnd1, XorOpnd *Opnd2, 187 APInt &ConstOpnd, Value *&Res); 188 bool collectMultiplyFactors(SmallVectorImpl<ValueEntry> &Ops, 189 SmallVectorImpl<Factor> &Factors); 190 Value *buildMinimalMultiplyDAG(IRBuilder<> &Builder, 191 SmallVectorImpl<Factor> &Factors); 192 Value *OptimizeMul(BinaryOperator *I, SmallVectorImpl<ValueEntry> &Ops); 193 Value *RemoveFactorFromExpression(Value *V, Value *Factor); 194 void EraseInst(Instruction *I); 195 void OptimizeInst(Instruction *I); 196 }; 197} 198 199XorOpnd::XorOpnd(Value *V) { 200 assert(!isa<ConstantInt>(V) && "No ConstantInt"); 201 OrigVal = V; 202 Instruction *I = dyn_cast<Instruction>(V); 203 SymbolicRank = 0; 204 205 if (I && (I->getOpcode() == Instruction::Or || 206 I->getOpcode() == Instruction::And)) { 207 Value *V0 = I->getOperand(0); 208 Value *V1 = I->getOperand(1); 209 if (isa<ConstantInt>(V0)) 210 std::swap(V0, V1); 211 212 if (ConstantInt *C = dyn_cast<ConstantInt>(V1)) { 213 ConstPart = C->getValue(); 214 SymbolicPart = V0; 215 isOr = (I->getOpcode() == Instruction::Or); 216 return; 217 } 218 } 219 220 // view the operand as "V | 0" 221 SymbolicPart = V; 222 ConstPart = APInt::getNullValue(V->getType()->getIntegerBitWidth()); 223 isOr = true; 224} 225 226char Reassociate::ID = 0; 227INITIALIZE_PASS(Reassociate, "reassociate", 228 "Reassociate expressions", false, false) 229 230// Public interface to the Reassociate pass 231FunctionPass *llvm::createReassociatePass() { return new Reassociate(); } 232 233/// isReassociableOp - Return true if V is an instruction of the specified 234/// opcode and if it only has one use. 235static BinaryOperator *isReassociableOp(Value *V, unsigned Opcode) { 236 if (V->hasOneUse() && isa<Instruction>(V) && 237 cast<Instruction>(V)->getOpcode() == Opcode) 238 return cast<BinaryOperator>(V); 239 return 0; 240} 241 242static bool isUnmovableInstruction(Instruction *I) { 243 switch (I->getOpcode()) { 244 case Instruction::PHI: 245 case Instruction::LandingPad: 246 case Instruction::Alloca: 247 case Instruction::Load: 248 case Instruction::Invoke: 249 case Instruction::UDiv: 250 case Instruction::SDiv: 251 case Instruction::FDiv: 252 case Instruction::URem: 253 case Instruction::SRem: 254 case Instruction::FRem: 255 return true; 256 case Instruction::Call: 257 return !isa<DbgInfoIntrinsic>(I); 258 default: 259 return false; 260 } 261} 262 263void Reassociate::BuildRankMap(Function &F) { 264 unsigned i = 2; 265 266 // Assign distinct ranks to function arguments 267 for (Function::arg_iterator I = F.arg_begin(), E = F.arg_end(); I != E; ++I) 268 ValueRankMap[&*I] = ++i; 269 270 ReversePostOrderTraversal<Function*> RPOT(&F); 271 for (ReversePostOrderTraversal<Function*>::rpo_iterator I = RPOT.begin(), 272 E = RPOT.end(); I != E; ++I) { 273 BasicBlock *BB = *I; 274 unsigned BBRank = RankMap[BB] = ++i << 16; 275 276 // Walk the basic block, adding precomputed ranks for any instructions that 277 // we cannot move. This ensures that the ranks for these instructions are 278 // all different in the block. 279 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) 280 if (isUnmovableInstruction(I)) 281 ValueRankMap[&*I] = ++BBRank; 282 } 283} 284 285unsigned Reassociate::getRank(Value *V) { 286 Instruction *I = dyn_cast<Instruction>(V); 287 if (I == 0) { 288 if (isa<Argument>(V)) return ValueRankMap[V]; // Function argument. 289 return 0; // Otherwise it's a global or constant, rank 0. 290 } 291 292 if (unsigned Rank = ValueRankMap[I]) 293 return Rank; // Rank already known? 294 295 // If this is an expression, return the 1+MAX(rank(LHS), rank(RHS)) so that 296 // we can reassociate expressions for code motion! Since we do not recurse 297 // for PHI nodes, we cannot have infinite recursion here, because there 298 // cannot be loops in the value graph that do not go through PHI nodes. 299 unsigned Rank = 0, MaxRank = RankMap[I->getParent()]; 300 for (unsigned i = 0, e = I->getNumOperands(); 301 i != e && Rank != MaxRank; ++i) 302 Rank = std::max(Rank, getRank(I->getOperand(i))); 303 304 // If this is a not or neg instruction, do not count it for rank. This 305 // assures us that X and ~X will have the same rank. 306 if (!I->getType()->isIntegerTy() || 307 (!BinaryOperator::isNot(I) && !BinaryOperator::isNeg(I))) 308 ++Rank; 309 310 //DEBUG(dbgs() << "Calculated Rank[" << V->getName() << "] = " 311 // << Rank << "\n"); 312 313 return ValueRankMap[I] = Rank; 314} 315 316/// LowerNegateToMultiply - Replace 0-X with X*-1. 317/// 318static BinaryOperator *LowerNegateToMultiply(Instruction *Neg) { 319 Constant *Cst = Constant::getAllOnesValue(Neg->getType()); 320 321 BinaryOperator *Res = 322 BinaryOperator::CreateMul(Neg->getOperand(1), Cst, "",Neg); 323 Neg->setOperand(1, Constant::getNullValue(Neg->getType())); // Drop use of op. 324 Res->takeName(Neg); 325 Neg->replaceAllUsesWith(Res); 326 Res->setDebugLoc(Neg->getDebugLoc()); 327 return Res; 328} 329 330/// CarmichaelShift - Returns k such that lambda(2^Bitwidth) = 2^k, where lambda 331/// is the Carmichael function. This means that x^(2^k) === 1 mod 2^Bitwidth for 332/// every odd x, i.e. x^(2^k) = 1 for every odd x in Bitwidth-bit arithmetic. 333/// Note that 0 <= k < Bitwidth, and if Bitwidth > 3 then x^(2^k) = 0 for every 334/// even x in Bitwidth-bit arithmetic. 335static unsigned CarmichaelShift(unsigned Bitwidth) { 336 if (Bitwidth < 3) 337 return Bitwidth - 1; 338 return Bitwidth - 2; 339} 340 341/// IncorporateWeight - Add the extra weight 'RHS' to the existing weight 'LHS', 342/// reducing the combined weight using any special properties of the operation. 343/// The existing weight LHS represents the computation X op X op ... op X where 344/// X occurs LHS times. The combined weight represents X op X op ... op X with 345/// X occurring LHS + RHS times. If op is "Xor" for example then the combined 346/// operation is equivalent to X if LHS + RHS is odd, or 0 if LHS + RHS is even; 347/// the routine returns 1 in LHS in the first case, and 0 in LHS in the second. 348static void IncorporateWeight(APInt &LHS, const APInt &RHS, unsigned Opcode) { 349 // If we were working with infinite precision arithmetic then the combined 350 // weight would be LHS + RHS. But we are using finite precision arithmetic, 351 // and the APInt sum LHS + RHS may not be correct if it wraps (it is correct 352 // for nilpotent operations and addition, but not for idempotent operations 353 // and multiplication), so it is important to correctly reduce the combined 354 // weight back into range if wrapping would be wrong. 355 356 // If RHS is zero then the weight didn't change. 357 if (RHS.isMinValue()) 358 return; 359 // If LHS is zero then the combined weight is RHS. 360 if (LHS.isMinValue()) { 361 LHS = RHS; 362 return; 363 } 364 // From this point on we know that neither LHS nor RHS is zero. 365 366 if (Instruction::isIdempotent(Opcode)) { 367 // Idempotent means X op X === X, so any non-zero weight is equivalent to a 368 // weight of 1. Keeping weights at zero or one also means that wrapping is 369 // not a problem. 370 assert(LHS == 1 && RHS == 1 && "Weights not reduced!"); 371 return; // Return a weight of 1. 372 } 373 if (Instruction::isNilpotent(Opcode)) { 374 // Nilpotent means X op X === 0, so reduce weights modulo 2. 375 assert(LHS == 1 && RHS == 1 && "Weights not reduced!"); 376 LHS = 0; // 1 + 1 === 0 modulo 2. 377 return; 378 } 379 if (Opcode == Instruction::Add) { 380 // TODO: Reduce the weight by exploiting nsw/nuw? 381 LHS += RHS; 382 return; 383 } 384 385 assert(Opcode == Instruction::Mul && "Unknown associative operation!"); 386 unsigned Bitwidth = LHS.getBitWidth(); 387 // If CM is the Carmichael number then a weight W satisfying W >= CM+Bitwidth 388 // can be replaced with W-CM. That's because x^W=x^(W-CM) for every Bitwidth 389 // bit number x, since either x is odd in which case x^CM = 1, or x is even in 390 // which case both x^W and x^(W - CM) are zero. By subtracting off multiples 391 // of CM like this weights can always be reduced to the range [0, CM+Bitwidth) 392 // which by a happy accident means that they can always be represented using 393 // Bitwidth bits. 394 // TODO: Reduce the weight by exploiting nsw/nuw? (Could do much better than 395 // the Carmichael number). 396 if (Bitwidth > 3) { 397 /// CM - The value of Carmichael's lambda function. 398 APInt CM = APInt::getOneBitSet(Bitwidth, CarmichaelShift(Bitwidth)); 399 // Any weight W >= Threshold can be replaced with W - CM. 400 APInt Threshold = CM + Bitwidth; 401 assert(LHS.ult(Threshold) && RHS.ult(Threshold) && "Weights not reduced!"); 402 // For Bitwidth 4 or more the following sum does not overflow. 403 LHS += RHS; 404 while (LHS.uge(Threshold)) 405 LHS -= CM; 406 } else { 407 // To avoid problems with overflow do everything the same as above but using 408 // a larger type. 409 unsigned CM = 1U << CarmichaelShift(Bitwidth); 410 unsigned Threshold = CM + Bitwidth; 411 assert(LHS.getZExtValue() < Threshold && RHS.getZExtValue() < Threshold && 412 "Weights not reduced!"); 413 unsigned Total = LHS.getZExtValue() + RHS.getZExtValue(); 414 while (Total >= Threshold) 415 Total -= CM; 416 LHS = Total; 417 } 418} 419 420typedef std::pair<Value*, APInt> RepeatedValue; 421 422/// LinearizeExprTree - Given an associative binary expression, return the leaf 423/// nodes in Ops along with their weights (how many times the leaf occurs). The 424/// original expression is the same as 425/// (Ops[0].first op Ops[0].first op ... Ops[0].first) <- Ops[0].second times 426/// op 427/// (Ops[1].first op Ops[1].first op ... Ops[1].first) <- Ops[1].second times 428/// op 429/// ... 430/// op 431/// (Ops[N].first op Ops[N].first op ... Ops[N].first) <- Ops[N].second times 432/// 433/// Note that the values Ops[0].first, ..., Ops[N].first are all distinct. 434/// 435/// This routine may modify the function, in which case it returns 'true'. The 436/// changes it makes may well be destructive, changing the value computed by 'I' 437/// to something completely different. Thus if the routine returns 'true' then 438/// you MUST either replace I with a new expression computed from the Ops array, 439/// or use RewriteExprTree to put the values back in. 440/// 441/// A leaf node is either not a binary operation of the same kind as the root 442/// node 'I' (i.e. is not a binary operator at all, or is, but with a different 443/// opcode), or is the same kind of binary operator but has a use which either 444/// does not belong to the expression, or does belong to the expression but is 445/// a leaf node. Every leaf node has at least one use that is a non-leaf node 446/// of the expression, while for non-leaf nodes (except for the root 'I') every 447/// use is a non-leaf node of the expression. 448/// 449/// For example: 450/// expression graph node names 451/// 452/// + | I 453/// / \ | 454/// + + | A, B 455/// / \ / \ | 456/// * + * | C, D, E 457/// / \ / \ / \ | 458/// + * | F, G 459/// 460/// The leaf nodes are C, E, F and G. The Ops array will contain (maybe not in 461/// that order) (C, 1), (E, 1), (F, 2), (G, 2). 462/// 463/// The expression is maximal: if some instruction is a binary operator of the 464/// same kind as 'I', and all of its uses are non-leaf nodes of the expression, 465/// then the instruction also belongs to the expression, is not a leaf node of 466/// it, and its operands also belong to the expression (but may be leaf nodes). 467/// 468/// NOTE: This routine will set operands of non-leaf non-root nodes to undef in 469/// order to ensure that every non-root node in the expression has *exactly one* 470/// use by a non-leaf node of the expression. This destruction means that the 471/// caller MUST either replace 'I' with a new expression or use something like 472/// RewriteExprTree to put the values back in if the routine indicates that it 473/// made a change by returning 'true'. 474/// 475/// In the above example either the right operand of A or the left operand of B 476/// will be replaced by undef. If it is B's operand then this gives: 477/// 478/// + | I 479/// / \ | 480/// + + | A, B - operand of B replaced with undef 481/// / \ \ | 482/// * + * | C, D, E 483/// / \ / \ / \ | 484/// + * | F, G 485/// 486/// Note that such undef operands can only be reached by passing through 'I'. 487/// For example, if you visit operands recursively starting from a leaf node 488/// then you will never see such an undef operand unless you get back to 'I', 489/// which requires passing through a phi node. 490/// 491/// Note that this routine may also mutate binary operators of the wrong type 492/// that have all uses inside the expression (i.e. only used by non-leaf nodes 493/// of the expression) if it can turn them into binary operators of the right 494/// type and thus make the expression bigger. 495 496static bool LinearizeExprTree(BinaryOperator *I, 497 SmallVectorImpl<RepeatedValue> &Ops) { 498 DEBUG(dbgs() << "LINEARIZE: " << *I << '\n'); 499 unsigned Bitwidth = I->getType()->getScalarType()->getPrimitiveSizeInBits(); 500 unsigned Opcode = I->getOpcode(); 501 assert(Instruction::isAssociative(Opcode) && 502 Instruction::isCommutative(Opcode) && 503 "Expected an associative and commutative operation!"); 504 505 // Visit all operands of the expression, keeping track of their weight (the 506 // number of paths from the expression root to the operand, or if you like 507 // the number of times that operand occurs in the linearized expression). 508 // For example, if I = X + A, where X = A + B, then I, X and B have weight 1 509 // while A has weight two. 510 511 // Worklist of non-leaf nodes (their operands are in the expression too) along 512 // with their weights, representing a certain number of paths to the operator. 513 // If an operator occurs in the worklist multiple times then we found multiple 514 // ways to get to it. 515 SmallVector<std::pair<BinaryOperator*, APInt>, 8> Worklist; // (Op, Weight) 516 Worklist.push_back(std::make_pair(I, APInt(Bitwidth, 1))); 517 bool MadeChange = false; 518 519 // Leaves of the expression are values that either aren't the right kind of 520 // operation (eg: a constant, or a multiply in an add tree), or are, but have 521 // some uses that are not inside the expression. For example, in I = X + X, 522 // X = A + B, the value X has two uses (by I) that are in the expression. If 523 // X has any other uses, for example in a return instruction, then we consider 524 // X to be a leaf, and won't analyze it further. When we first visit a value, 525 // if it has more than one use then at first we conservatively consider it to 526 // be a leaf. Later, as the expression is explored, we may discover some more 527 // uses of the value from inside the expression. If all uses turn out to be 528 // from within the expression (and the value is a binary operator of the right 529 // kind) then the value is no longer considered to be a leaf, and its operands 530 // are explored. 531 532 // Leaves - Keeps track of the set of putative leaves as well as the number of 533 // paths to each leaf seen so far. 534 typedef DenseMap<Value*, APInt> LeafMap; 535 LeafMap Leaves; // Leaf -> Total weight so far. 536 SmallVector<Value*, 8> LeafOrder; // Ensure deterministic leaf output order. 537 538#ifndef NDEBUG 539 SmallPtrSet<Value*, 8> Visited; // For sanity checking the iteration scheme. 540#endif 541 while (!Worklist.empty()) { 542 std::pair<BinaryOperator*, APInt> P = Worklist.pop_back_val(); 543 I = P.first; // We examine the operands of this binary operator. 544 545 for (unsigned OpIdx = 0; OpIdx < 2; ++OpIdx) { // Visit operands. 546 Value *Op = I->getOperand(OpIdx); 547 APInt Weight = P.second; // Number of paths to this operand. 548 DEBUG(dbgs() << "OPERAND: " << *Op << " (" << Weight << ")\n"); 549 assert(!Op->use_empty() && "No uses, so how did we get to it?!"); 550 551 // If this is a binary operation of the right kind with only one use then 552 // add its operands to the expression. 553 if (BinaryOperator *BO = isReassociableOp(Op, Opcode)) { 554 assert(Visited.insert(Op) && "Not first visit!"); 555 DEBUG(dbgs() << "DIRECT ADD: " << *Op << " (" << Weight << ")\n"); 556 Worklist.push_back(std::make_pair(BO, Weight)); 557 continue; 558 } 559 560 // Appears to be a leaf. Is the operand already in the set of leaves? 561 LeafMap::iterator It = Leaves.find(Op); 562 if (It == Leaves.end()) { 563 // Not in the leaf map. Must be the first time we saw this operand. 564 assert(Visited.insert(Op) && "Not first visit!"); 565 if (!Op->hasOneUse()) { 566 // This value has uses not accounted for by the expression, so it is 567 // not safe to modify. Mark it as being a leaf. 568 DEBUG(dbgs() << "ADD USES LEAF: " << *Op << " (" << Weight << ")\n"); 569 LeafOrder.push_back(Op); 570 Leaves[Op] = Weight; 571 continue; 572 } 573 // No uses outside the expression, try morphing it. 574 } else if (It != Leaves.end()) { 575 // Already in the leaf map. 576 assert(Visited.count(Op) && "In leaf map but not visited!"); 577 578 // Update the number of paths to the leaf. 579 IncorporateWeight(It->second, Weight, Opcode); 580 581#if 0 // TODO: Re-enable once PR13021 is fixed. 582 // The leaf already has one use from inside the expression. As we want 583 // exactly one such use, drop this new use of the leaf. 584 assert(!Op->hasOneUse() && "Only one use, but we got here twice!"); 585 I->setOperand(OpIdx, UndefValue::get(I->getType())); 586 MadeChange = true; 587 588 // If the leaf is a binary operation of the right kind and we now see 589 // that its multiple original uses were in fact all by nodes belonging 590 // to the expression, then no longer consider it to be a leaf and add 591 // its operands to the expression. 592 if (BinaryOperator *BO = isReassociableOp(Op, Opcode)) { 593 DEBUG(dbgs() << "UNLEAF: " << *Op << " (" << It->second << ")\n"); 594 Worklist.push_back(std::make_pair(BO, It->second)); 595 Leaves.erase(It); 596 continue; 597 } 598#endif 599 600 // If we still have uses that are not accounted for by the expression 601 // then it is not safe to modify the value. 602 if (!Op->hasOneUse()) 603 continue; 604 605 // No uses outside the expression, try morphing it. 606 Weight = It->second; 607 Leaves.erase(It); // Since the value may be morphed below. 608 } 609 610 // At this point we have a value which, first of all, is not a binary 611 // expression of the right kind, and secondly, is only used inside the 612 // expression. This means that it can safely be modified. See if we 613 // can usefully morph it into an expression of the right kind. 614 assert((!isa<Instruction>(Op) || 615 cast<Instruction>(Op)->getOpcode() != Opcode) && 616 "Should have been handled above!"); 617 assert(Op->hasOneUse() && "Has uses outside the expression tree!"); 618 619 // If this is a multiply expression, turn any internal negations into 620 // multiplies by -1 so they can be reassociated. 621 BinaryOperator *BO = dyn_cast<BinaryOperator>(Op); 622 if (Opcode == Instruction::Mul && BO && BinaryOperator::isNeg(BO)) { 623 DEBUG(dbgs() << "MORPH LEAF: " << *Op << " (" << Weight << ") TO "); 624 BO = LowerNegateToMultiply(BO); 625 DEBUG(dbgs() << *BO << 'n'); 626 Worklist.push_back(std::make_pair(BO, Weight)); 627 MadeChange = true; 628 continue; 629 } 630 631 // Failed to morph into an expression of the right type. This really is 632 // a leaf. 633 DEBUG(dbgs() << "ADD LEAF: " << *Op << " (" << Weight << ")\n"); 634 assert(!isReassociableOp(Op, Opcode) && "Value was morphed?"); 635 LeafOrder.push_back(Op); 636 Leaves[Op] = Weight; 637 } 638 } 639 640 // The leaves, repeated according to their weights, represent the linearized 641 // form of the expression. 642 for (unsigned i = 0, e = LeafOrder.size(); i != e; ++i) { 643 Value *V = LeafOrder[i]; 644 LeafMap::iterator It = Leaves.find(V); 645 if (It == Leaves.end()) 646 // Node initially thought to be a leaf wasn't. 647 continue; 648 assert(!isReassociableOp(V, Opcode) && "Shouldn't be a leaf!"); 649 APInt Weight = It->second; 650 if (Weight.isMinValue()) 651 // Leaf already output or weight reduction eliminated it. 652 continue; 653 // Ensure the leaf is only output once. 654 It->second = 0; 655 Ops.push_back(std::make_pair(V, Weight)); 656 } 657 658 // For nilpotent operations or addition there may be no operands, for example 659 // because the expression was "X xor X" or consisted of 2^Bitwidth additions: 660 // in both cases the weight reduces to 0 causing the value to be skipped. 661 if (Ops.empty()) { 662 Constant *Identity = ConstantExpr::getBinOpIdentity(Opcode, I->getType()); 663 assert(Identity && "Associative operation without identity!"); 664 Ops.push_back(std::make_pair(Identity, APInt(Bitwidth, 1))); 665 } 666 667 return MadeChange; 668} 669 670// RewriteExprTree - Now that the operands for this expression tree are 671// linearized and optimized, emit them in-order. 672void Reassociate::RewriteExprTree(BinaryOperator *I, 673 SmallVectorImpl<ValueEntry> &Ops) { 674 assert(Ops.size() > 1 && "Single values should be used directly!"); 675 676 // Since our optimizations should never increase the number of operations, the 677 // new expression can usually be written reusing the existing binary operators 678 // from the original expression tree, without creating any new instructions, 679 // though the rewritten expression may have a completely different topology. 680 // We take care to not change anything if the new expression will be the same 681 // as the original. If more than trivial changes (like commuting operands) 682 // were made then we are obliged to clear out any optional subclass data like 683 // nsw flags. 684 685 /// NodesToRewrite - Nodes from the original expression available for writing 686 /// the new expression into. 687 SmallVector<BinaryOperator*, 8> NodesToRewrite; 688 unsigned Opcode = I->getOpcode(); 689 BinaryOperator *Op = I; 690 691 /// NotRewritable - The operands being written will be the leaves of the new 692 /// expression and must not be used as inner nodes (via NodesToRewrite) by 693 /// mistake. Inner nodes are always reassociable, and usually leaves are not 694 /// (if they were they would have been incorporated into the expression and so 695 /// would not be leaves), so most of the time there is no danger of this. But 696 /// in rare cases a leaf may become reassociable if an optimization kills uses 697 /// of it, or it may momentarily become reassociable during rewriting (below) 698 /// due it being removed as an operand of one of its uses. Ensure that misuse 699 /// of leaf nodes as inner nodes cannot occur by remembering all of the future 700 /// leaves and refusing to reuse any of them as inner nodes. 701 SmallPtrSet<Value*, 8> NotRewritable; 702 for (unsigned i = 0, e = Ops.size(); i != e; ++i) 703 NotRewritable.insert(Ops[i].Op); 704 705 // ExpressionChanged - Non-null if the rewritten expression differs from the 706 // original in some non-trivial way, requiring the clearing of optional flags. 707 // Flags are cleared from the operator in ExpressionChanged up to I inclusive. 708 BinaryOperator *ExpressionChanged = 0; 709 for (unsigned i = 0; ; ++i) { 710 // The last operation (which comes earliest in the IR) is special as both 711 // operands will come from Ops, rather than just one with the other being 712 // a subexpression. 713 if (i+2 == Ops.size()) { 714 Value *NewLHS = Ops[i].Op; 715 Value *NewRHS = Ops[i+1].Op; 716 Value *OldLHS = Op->getOperand(0); 717 Value *OldRHS = Op->getOperand(1); 718 719 if (NewLHS == OldLHS && NewRHS == OldRHS) 720 // Nothing changed, leave it alone. 721 break; 722 723 if (NewLHS == OldRHS && NewRHS == OldLHS) { 724 // The order of the operands was reversed. Swap them. 725 DEBUG(dbgs() << "RA: " << *Op << '\n'); 726 Op->swapOperands(); 727 DEBUG(dbgs() << "TO: " << *Op << '\n'); 728 MadeChange = true; 729 ++NumChanged; 730 break; 731 } 732 733 // The new operation differs non-trivially from the original. Overwrite 734 // the old operands with the new ones. 735 DEBUG(dbgs() << "RA: " << *Op << '\n'); 736 if (NewLHS != OldLHS) { 737 BinaryOperator *BO = isReassociableOp(OldLHS, Opcode); 738 if (BO && !NotRewritable.count(BO)) 739 NodesToRewrite.push_back(BO); 740 Op->setOperand(0, NewLHS); 741 } 742 if (NewRHS != OldRHS) { 743 BinaryOperator *BO = isReassociableOp(OldRHS, Opcode); 744 if (BO && !NotRewritable.count(BO)) 745 NodesToRewrite.push_back(BO); 746 Op->setOperand(1, NewRHS); 747 } 748 DEBUG(dbgs() << "TO: " << *Op << '\n'); 749 750 ExpressionChanged = Op; 751 MadeChange = true; 752 ++NumChanged; 753 754 break; 755 } 756 757 // Not the last operation. The left-hand side will be a sub-expression 758 // while the right-hand side will be the current element of Ops. 759 Value *NewRHS = Ops[i].Op; 760 if (NewRHS != Op->getOperand(1)) { 761 DEBUG(dbgs() << "RA: " << *Op << '\n'); 762 if (NewRHS == Op->getOperand(0)) { 763 // The new right-hand side was already present as the left operand. If 764 // we are lucky then swapping the operands will sort out both of them. 765 Op->swapOperands(); 766 } else { 767 // Overwrite with the new right-hand side. 768 BinaryOperator *BO = isReassociableOp(Op->getOperand(1), Opcode); 769 if (BO && !NotRewritable.count(BO)) 770 NodesToRewrite.push_back(BO); 771 Op->setOperand(1, NewRHS); 772 ExpressionChanged = Op; 773 } 774 DEBUG(dbgs() << "TO: " << *Op << '\n'); 775 MadeChange = true; 776 ++NumChanged; 777 } 778 779 // Now deal with the left-hand side. If this is already an operation node 780 // from the original expression then just rewrite the rest of the expression 781 // into it. 782 BinaryOperator *BO = isReassociableOp(Op->getOperand(0), Opcode); 783 if (BO && !NotRewritable.count(BO)) { 784 Op = BO; 785 continue; 786 } 787 788 // Otherwise, grab a spare node from the original expression and use that as 789 // the left-hand side. If there are no nodes left then the optimizers made 790 // an expression with more nodes than the original! This usually means that 791 // they did something stupid but it might mean that the problem was just too 792 // hard (finding the mimimal number of multiplications needed to realize a 793 // multiplication expression is NP-complete). Whatever the reason, smart or 794 // stupid, create a new node if there are none left. 795 BinaryOperator *NewOp; 796 if (NodesToRewrite.empty()) { 797 Constant *Undef = UndefValue::get(I->getType()); 798 NewOp = BinaryOperator::Create(Instruction::BinaryOps(Opcode), 799 Undef, Undef, "", I); 800 } else { 801 NewOp = NodesToRewrite.pop_back_val(); 802 } 803 804 DEBUG(dbgs() << "RA: " << *Op << '\n'); 805 Op->setOperand(0, NewOp); 806 DEBUG(dbgs() << "TO: " << *Op << '\n'); 807 ExpressionChanged = Op; 808 MadeChange = true; 809 ++NumChanged; 810 Op = NewOp; 811 } 812 813 // If the expression changed non-trivially then clear out all subclass data 814 // starting from the operator specified in ExpressionChanged, and compactify 815 // the operators to just before the expression root to guarantee that the 816 // expression tree is dominated by all of Ops. 817 if (ExpressionChanged) 818 do { 819 ExpressionChanged->clearSubclassOptionalData(); 820 if (ExpressionChanged == I) 821 break; 822 ExpressionChanged->moveBefore(I); 823 ExpressionChanged = cast<BinaryOperator>(*ExpressionChanged->user_begin()); 824 } while (1); 825 826 // Throw away any left over nodes from the original expression. 827 for (unsigned i = 0, e = NodesToRewrite.size(); i != e; ++i) 828 RedoInsts.insert(NodesToRewrite[i]); 829} 830 831/// NegateValue - Insert instructions before the instruction pointed to by BI, 832/// that computes the negative version of the value specified. The negative 833/// version of the value is returned, and BI is left pointing at the instruction 834/// that should be processed next by the reassociation pass. 835static Value *NegateValue(Value *V, Instruction *BI) { 836 if (Constant *C = dyn_cast<Constant>(V)) 837 return ConstantExpr::getNeg(C); 838 839 // We are trying to expose opportunity for reassociation. One of the things 840 // that we want to do to achieve this is to push a negation as deep into an 841 // expression chain as possible, to expose the add instructions. In practice, 842 // this means that we turn this: 843 // X = -(A+12+C+D) into X = -A + -12 + -C + -D = -12 + -A + -C + -D 844 // so that later, a: Y = 12+X could get reassociated with the -12 to eliminate 845 // the constants. We assume that instcombine will clean up the mess later if 846 // we introduce tons of unnecessary negation instructions. 847 // 848 if (BinaryOperator *I = isReassociableOp(V, Instruction::Add)) { 849 // Push the negates through the add. 850 I->setOperand(0, NegateValue(I->getOperand(0), BI)); 851 I->setOperand(1, NegateValue(I->getOperand(1), BI)); 852 853 // We must move the add instruction here, because the neg instructions do 854 // not dominate the old add instruction in general. By moving it, we are 855 // assured that the neg instructions we just inserted dominate the 856 // instruction we are about to insert after them. 857 // 858 I->moveBefore(BI); 859 I->setName(I->getName()+".neg"); 860 return I; 861 } 862 863 // Okay, we need to materialize a negated version of V with an instruction. 864 // Scan the use lists of V to see if we have one already. 865 for (User *U : V->users()) { 866 if (!BinaryOperator::isNeg(U)) continue; 867 868 // We found one! Now we have to make sure that the definition dominates 869 // this use. We do this by moving it to the entry block (if it is a 870 // non-instruction value) or right after the definition. These negates will 871 // be zapped by reassociate later, so we don't need much finesse here. 872 BinaryOperator *TheNeg = cast<BinaryOperator>(U); 873 874 // Verify that the negate is in this function, V might be a constant expr. 875 if (TheNeg->getParent()->getParent() != BI->getParent()->getParent()) 876 continue; 877 878 BasicBlock::iterator InsertPt; 879 if (Instruction *InstInput = dyn_cast<Instruction>(V)) { 880 if (InvokeInst *II = dyn_cast<InvokeInst>(InstInput)) { 881 InsertPt = II->getNormalDest()->begin(); 882 } else { 883 InsertPt = InstInput; 884 ++InsertPt; 885 } 886 while (isa<PHINode>(InsertPt)) ++InsertPt; 887 } else { 888 InsertPt = TheNeg->getParent()->getParent()->getEntryBlock().begin(); 889 } 890 TheNeg->moveBefore(InsertPt); 891 return TheNeg; 892 } 893 894 // Insert a 'neg' instruction that subtracts the value from zero to get the 895 // negation. 896 return BinaryOperator::CreateNeg(V, V->getName() + ".neg", BI); 897} 898 899/// ShouldBreakUpSubtract - Return true if we should break up this subtract of 900/// X-Y into (X + -Y). 901static bool ShouldBreakUpSubtract(Instruction *Sub) { 902 // If this is a negation, we can't split it up! 903 if (BinaryOperator::isNeg(Sub)) 904 return false; 905 906 // Don't bother to break this up unless either the LHS is an associable add or 907 // subtract or if this is only used by one. 908 if (isReassociableOp(Sub->getOperand(0), Instruction::Add) || 909 isReassociableOp(Sub->getOperand(0), Instruction::Sub)) 910 return true; 911 if (isReassociableOp(Sub->getOperand(1), Instruction::Add) || 912 isReassociableOp(Sub->getOperand(1), Instruction::Sub)) 913 return true; 914 if (Sub->hasOneUse() && 915 (isReassociableOp(Sub->user_back(), Instruction::Add) || 916 isReassociableOp(Sub->user_back(), Instruction::Sub))) 917 return true; 918 919 return false; 920} 921 922/// BreakUpSubtract - If we have (X-Y), and if either X is an add, or if this is 923/// only used by an add, transform this into (X+(0-Y)) to promote better 924/// reassociation. 925static BinaryOperator *BreakUpSubtract(Instruction *Sub) { 926 // Convert a subtract into an add and a neg instruction. This allows sub 927 // instructions to be commuted with other add instructions. 928 // 929 // Calculate the negative value of Operand 1 of the sub instruction, 930 // and set it as the RHS of the add instruction we just made. 931 // 932 Value *NegVal = NegateValue(Sub->getOperand(1), Sub); 933 BinaryOperator *New = 934 BinaryOperator::CreateAdd(Sub->getOperand(0), NegVal, "", Sub); 935 Sub->setOperand(0, Constant::getNullValue(Sub->getType())); // Drop use of op. 936 Sub->setOperand(1, Constant::getNullValue(Sub->getType())); // Drop use of op. 937 New->takeName(Sub); 938 939 // Everyone now refers to the add instruction. 940 Sub->replaceAllUsesWith(New); 941 New->setDebugLoc(Sub->getDebugLoc()); 942 943 DEBUG(dbgs() << "Negated: " << *New << '\n'); 944 return New; 945} 946 947/// ConvertShiftToMul - If this is a shift of a reassociable multiply or is used 948/// by one, change this into a multiply by a constant to assist with further 949/// reassociation. 950static BinaryOperator *ConvertShiftToMul(Instruction *Shl) { 951 Constant *MulCst = ConstantInt::get(Shl->getType(), 1); 952 MulCst = ConstantExpr::getShl(MulCst, cast<Constant>(Shl->getOperand(1))); 953 954 BinaryOperator *Mul = 955 BinaryOperator::CreateMul(Shl->getOperand(0), MulCst, "", Shl); 956 Shl->setOperand(0, UndefValue::get(Shl->getType())); // Drop use of op. 957 Mul->takeName(Shl); 958 Shl->replaceAllUsesWith(Mul); 959 Mul->setDebugLoc(Shl->getDebugLoc()); 960 return Mul; 961} 962 963/// FindInOperandList - Scan backwards and forwards among values with the same 964/// rank as element i to see if X exists. If X does not exist, return i. This 965/// is useful when scanning for 'x' when we see '-x' because they both get the 966/// same rank. 967static unsigned FindInOperandList(SmallVectorImpl<ValueEntry> &Ops, unsigned i, 968 Value *X) { 969 unsigned XRank = Ops[i].Rank; 970 unsigned e = Ops.size(); 971 for (unsigned j = i+1; j != e && Ops[j].Rank == XRank; ++j) 972 if (Ops[j].Op == X) 973 return j; 974 // Scan backwards. 975 for (unsigned j = i-1; j != ~0U && Ops[j].Rank == XRank; --j) 976 if (Ops[j].Op == X) 977 return j; 978 return i; 979} 980 981/// EmitAddTreeOfValues - Emit a tree of add instructions, summing Ops together 982/// and returning the result. Insert the tree before I. 983static Value *EmitAddTreeOfValues(Instruction *I, 984 SmallVectorImpl<WeakVH> &Ops){ 985 if (Ops.size() == 1) return Ops.back(); 986 987 Value *V1 = Ops.back(); 988 Ops.pop_back(); 989 Value *V2 = EmitAddTreeOfValues(I, Ops); 990 return BinaryOperator::CreateAdd(V2, V1, "tmp", I); 991} 992 993/// RemoveFactorFromExpression - If V is an expression tree that is a 994/// multiplication sequence, and if this sequence contains a multiply by Factor, 995/// remove Factor from the tree and return the new tree. 996Value *Reassociate::RemoveFactorFromExpression(Value *V, Value *Factor) { 997 BinaryOperator *BO = isReassociableOp(V, Instruction::Mul); 998 if (!BO) return 0; 999 1000 SmallVector<RepeatedValue, 8> Tree; 1001 MadeChange |= LinearizeExprTree(BO, Tree); 1002 SmallVector<ValueEntry, 8> Factors; 1003 Factors.reserve(Tree.size()); 1004 for (unsigned i = 0, e = Tree.size(); i != e; ++i) { 1005 RepeatedValue E = Tree[i]; 1006 Factors.append(E.second.getZExtValue(), 1007 ValueEntry(getRank(E.first), E.first)); 1008 } 1009 1010 bool FoundFactor = false; 1011 bool NeedsNegate = false; 1012 for (unsigned i = 0, e = Factors.size(); i != e; ++i) { 1013 if (Factors[i].Op == Factor) { 1014 FoundFactor = true; 1015 Factors.erase(Factors.begin()+i); 1016 break; 1017 } 1018 1019 // If this is a negative version of this factor, remove it. 1020 if (ConstantInt *FC1 = dyn_cast<ConstantInt>(Factor)) 1021 if (ConstantInt *FC2 = dyn_cast<ConstantInt>(Factors[i].Op)) 1022 if (FC1->getValue() == -FC2->getValue()) { 1023 FoundFactor = NeedsNegate = true; 1024 Factors.erase(Factors.begin()+i); 1025 break; 1026 } 1027 } 1028 1029 if (!FoundFactor) { 1030 // Make sure to restore the operands to the expression tree. 1031 RewriteExprTree(BO, Factors); 1032 return 0; 1033 } 1034 1035 BasicBlock::iterator InsertPt = BO; ++InsertPt; 1036 1037 // If this was just a single multiply, remove the multiply and return the only 1038 // remaining operand. 1039 if (Factors.size() == 1) { 1040 RedoInsts.insert(BO); 1041 V = Factors[0].Op; 1042 } else { 1043 RewriteExprTree(BO, Factors); 1044 V = BO; 1045 } 1046 1047 if (NeedsNegate) 1048 V = BinaryOperator::CreateNeg(V, "neg", InsertPt); 1049 1050 return V; 1051} 1052 1053/// FindSingleUseMultiplyFactors - If V is a single-use multiply, recursively 1054/// add its operands as factors, otherwise add V to the list of factors. 1055/// 1056/// Ops is the top-level list of add operands we're trying to factor. 1057static void FindSingleUseMultiplyFactors(Value *V, 1058 SmallVectorImpl<Value*> &Factors, 1059 const SmallVectorImpl<ValueEntry> &Ops) { 1060 BinaryOperator *BO = isReassociableOp(V, Instruction::Mul); 1061 if (!BO) { 1062 Factors.push_back(V); 1063 return; 1064 } 1065 1066 // Otherwise, add the LHS and RHS to the list of factors. 1067 FindSingleUseMultiplyFactors(BO->getOperand(1), Factors, Ops); 1068 FindSingleUseMultiplyFactors(BO->getOperand(0), Factors, Ops); 1069} 1070 1071/// OptimizeAndOrXor - Optimize a series of operands to an 'and', 'or', or 'xor' 1072/// instruction. This optimizes based on identities. If it can be reduced to 1073/// a single Value, it is returned, otherwise the Ops list is mutated as 1074/// necessary. 1075static Value *OptimizeAndOrXor(unsigned Opcode, 1076 SmallVectorImpl<ValueEntry> &Ops) { 1077 // Scan the operand lists looking for X and ~X pairs, along with X,X pairs. 1078 // If we find any, we can simplify the expression. X&~X == 0, X|~X == -1. 1079 for (unsigned i = 0, e = Ops.size(); i != e; ++i) { 1080 // First, check for X and ~X in the operand list. 1081 assert(i < Ops.size()); 1082 if (BinaryOperator::isNot(Ops[i].Op)) { // Cannot occur for ^. 1083 Value *X = BinaryOperator::getNotArgument(Ops[i].Op); 1084 unsigned FoundX = FindInOperandList(Ops, i, X); 1085 if (FoundX != i) { 1086 if (Opcode == Instruction::And) // ...&X&~X = 0 1087 return Constant::getNullValue(X->getType()); 1088 1089 if (Opcode == Instruction::Or) // ...|X|~X = -1 1090 return Constant::getAllOnesValue(X->getType()); 1091 } 1092 } 1093 1094 // Next, check for duplicate pairs of values, which we assume are next to 1095 // each other, due to our sorting criteria. 1096 assert(i < Ops.size()); 1097 if (i+1 != Ops.size() && Ops[i+1].Op == Ops[i].Op) { 1098 if (Opcode == Instruction::And || Opcode == Instruction::Or) { 1099 // Drop duplicate values for And and Or. 1100 Ops.erase(Ops.begin()+i); 1101 --i; --e; 1102 ++NumAnnihil; 1103 continue; 1104 } 1105 1106 // Drop pairs of values for Xor. 1107 assert(Opcode == Instruction::Xor); 1108 if (e == 2) 1109 return Constant::getNullValue(Ops[0].Op->getType()); 1110 1111 // Y ^ X^X -> Y 1112 Ops.erase(Ops.begin()+i, Ops.begin()+i+2); 1113 i -= 1; e -= 2; 1114 ++NumAnnihil; 1115 } 1116 } 1117 return 0; 1118} 1119 1120/// Helper funciton of CombineXorOpnd(). It creates a bitwise-and 1121/// instruction with the given two operands, and return the resulting 1122/// instruction. There are two special cases: 1) if the constant operand is 0, 1123/// it will return NULL. 2) if the constant is ~0, the symbolic operand will 1124/// be returned. 1125static Value *createAndInstr(Instruction *InsertBefore, Value *Opnd, 1126 const APInt &ConstOpnd) { 1127 if (ConstOpnd != 0) { 1128 if (!ConstOpnd.isAllOnesValue()) { 1129 LLVMContext &Ctx = Opnd->getType()->getContext(); 1130 Instruction *I; 1131 I = BinaryOperator::CreateAnd(Opnd, ConstantInt::get(Ctx, ConstOpnd), 1132 "and.ra", InsertBefore); 1133 I->setDebugLoc(InsertBefore->getDebugLoc()); 1134 return I; 1135 } 1136 return Opnd; 1137 } 1138 return 0; 1139} 1140 1141// Helper function of OptimizeXor(). It tries to simplify "Opnd1 ^ ConstOpnd" 1142// into "R ^ C", where C would be 0, and R is a symbolic value. 1143// 1144// If it was successful, true is returned, and the "R" and "C" is returned 1145// via "Res" and "ConstOpnd", respectively; otherwise, false is returned, 1146// and both "Res" and "ConstOpnd" remain unchanged. 1147// 1148bool Reassociate::CombineXorOpnd(Instruction *I, XorOpnd *Opnd1, 1149 APInt &ConstOpnd, Value *&Res) { 1150 // Xor-Rule 1: (x | c1) ^ c2 = (x | c1) ^ (c1 ^ c1) ^ c2 1151 // = ((x | c1) ^ c1) ^ (c1 ^ c2) 1152 // = (x & ~c1) ^ (c1 ^ c2) 1153 // It is useful only when c1 == c2. 1154 if (Opnd1->isOrExpr() && Opnd1->getConstPart() != 0) { 1155 if (!Opnd1->getValue()->hasOneUse()) 1156 return false; 1157 1158 const APInt &C1 = Opnd1->getConstPart(); 1159 if (C1 != ConstOpnd) 1160 return false; 1161 1162 Value *X = Opnd1->getSymbolicPart(); 1163 Res = createAndInstr(I, X, ~C1); 1164 // ConstOpnd was C2, now C1 ^ C2. 1165 ConstOpnd ^= C1; 1166 1167 if (Instruction *T = dyn_cast<Instruction>(Opnd1->getValue())) 1168 RedoInsts.insert(T); 1169 return true; 1170 } 1171 return false; 1172} 1173 1174 1175// Helper function of OptimizeXor(). It tries to simplify 1176// "Opnd1 ^ Opnd2 ^ ConstOpnd" into "R ^ C", where C would be 0, and R is a 1177// symbolic value. 1178// 1179// If it was successful, true is returned, and the "R" and "C" is returned 1180// via "Res" and "ConstOpnd", respectively (If the entire expression is 1181// evaluated to a constant, the Res is set to NULL); otherwise, false is 1182// returned, and both "Res" and "ConstOpnd" remain unchanged. 1183bool Reassociate::CombineXorOpnd(Instruction *I, XorOpnd *Opnd1, XorOpnd *Opnd2, 1184 APInt &ConstOpnd, Value *&Res) { 1185 Value *X = Opnd1->getSymbolicPart(); 1186 if (X != Opnd2->getSymbolicPart()) 1187 return false; 1188 1189 // This many instruction become dead.(At least "Opnd1 ^ Opnd2" will die.) 1190 int DeadInstNum = 1; 1191 if (Opnd1->getValue()->hasOneUse()) 1192 DeadInstNum++; 1193 if (Opnd2->getValue()->hasOneUse()) 1194 DeadInstNum++; 1195 1196 // Xor-Rule 2: 1197 // (x | c1) ^ (x & c2) 1198 // = (x|c1) ^ (x&c2) ^ (c1 ^ c1) = ((x|c1) ^ c1) ^ (x & c2) ^ c1 1199 // = (x & ~c1) ^ (x & c2) ^ c1 // Xor-Rule 1 1200 // = (x & c3) ^ c1, where c3 = ~c1 ^ c2 // Xor-rule 3 1201 // 1202 if (Opnd1->isOrExpr() != Opnd2->isOrExpr()) { 1203 if (Opnd2->isOrExpr()) 1204 std::swap(Opnd1, Opnd2); 1205 1206 const APInt &C1 = Opnd1->getConstPart(); 1207 const APInt &C2 = Opnd2->getConstPart(); 1208 APInt C3((~C1) ^ C2); 1209 1210 // Do not increase code size! 1211 if (C3 != 0 && !C3.isAllOnesValue()) { 1212 int NewInstNum = ConstOpnd != 0 ? 1 : 2; 1213 if (NewInstNum > DeadInstNum) 1214 return false; 1215 } 1216 1217 Res = createAndInstr(I, X, C3); 1218 ConstOpnd ^= C1; 1219 1220 } else if (Opnd1->isOrExpr()) { 1221 // Xor-Rule 3: (x | c1) ^ (x | c2) = (x & c3) ^ c3 where c3 = c1 ^ c2 1222 // 1223 const APInt &C1 = Opnd1->getConstPart(); 1224 const APInt &C2 = Opnd2->getConstPart(); 1225 APInt C3 = C1 ^ C2; 1226 1227 // Do not increase code size 1228 if (C3 != 0 && !C3.isAllOnesValue()) { 1229 int NewInstNum = ConstOpnd != 0 ? 1 : 2; 1230 if (NewInstNum > DeadInstNum) 1231 return false; 1232 } 1233 1234 Res = createAndInstr(I, X, C3); 1235 ConstOpnd ^= C3; 1236 } else { 1237 // Xor-Rule 4: (x & c1) ^ (x & c2) = (x & (c1^c2)) 1238 // 1239 const APInt &C1 = Opnd1->getConstPart(); 1240 const APInt &C2 = Opnd2->getConstPart(); 1241 APInt C3 = C1 ^ C2; 1242 Res = createAndInstr(I, X, C3); 1243 } 1244 1245 // Put the original operands in the Redo list; hope they will be deleted 1246 // as dead code. 1247 if (Instruction *T = dyn_cast<Instruction>(Opnd1->getValue())) 1248 RedoInsts.insert(T); 1249 if (Instruction *T = dyn_cast<Instruction>(Opnd2->getValue())) 1250 RedoInsts.insert(T); 1251 1252 return true; 1253} 1254 1255/// Optimize a series of operands to an 'xor' instruction. If it can be reduced 1256/// to a single Value, it is returned, otherwise the Ops list is mutated as 1257/// necessary. 1258Value *Reassociate::OptimizeXor(Instruction *I, 1259 SmallVectorImpl<ValueEntry> &Ops) { 1260 if (Value *V = OptimizeAndOrXor(Instruction::Xor, Ops)) 1261 return V; 1262 1263 if (Ops.size() == 1) 1264 return 0; 1265 1266 SmallVector<XorOpnd, 8> Opnds; 1267 SmallVector<XorOpnd*, 8> OpndPtrs; 1268 Type *Ty = Ops[0].Op->getType(); 1269 APInt ConstOpnd(Ty->getIntegerBitWidth(), 0); 1270 1271 // Step 1: Convert ValueEntry to XorOpnd 1272 for (unsigned i = 0, e = Ops.size(); i != e; ++i) { 1273 Value *V = Ops[i].Op; 1274 if (!isa<ConstantInt>(V)) { 1275 XorOpnd O(V); 1276 O.setSymbolicRank(getRank(O.getSymbolicPart())); 1277 Opnds.push_back(O); 1278 } else 1279 ConstOpnd ^= cast<ConstantInt>(V)->getValue(); 1280 } 1281 1282 // NOTE: From this point on, do *NOT* add/delete element to/from "Opnds". 1283 // It would otherwise invalidate the "Opnds"'s iterator, and hence invalidate 1284 // the "OpndPtrs" as well. For the similar reason, do not fuse this loop 1285 // with the previous loop --- the iterator of the "Opnds" may be invalidated 1286 // when new elements are added to the vector. 1287 for (unsigned i = 0, e = Opnds.size(); i != e; ++i) 1288 OpndPtrs.push_back(&Opnds[i]); 1289 1290 // Step 2: Sort the Xor-Operands in a way such that the operands containing 1291 // the same symbolic value cluster together. For instance, the input operand 1292 // sequence ("x | 123", "y & 456", "x & 789") will be sorted into: 1293 // ("x | 123", "x & 789", "y & 456"). 1294 std::stable_sort(OpndPtrs.begin(), OpndPtrs.end(), XorOpnd::PtrSortFunctor()); 1295 1296 // Step 3: Combine adjacent operands 1297 XorOpnd *PrevOpnd = 0; 1298 bool Changed = false; 1299 for (unsigned i = 0, e = Opnds.size(); i < e; i++) { 1300 XorOpnd *CurrOpnd = OpndPtrs[i]; 1301 // The combined value 1302 Value *CV; 1303 1304 // Step 3.1: Try simplifying "CurrOpnd ^ ConstOpnd" 1305 if (ConstOpnd != 0 && CombineXorOpnd(I, CurrOpnd, ConstOpnd, CV)) { 1306 Changed = true; 1307 if (CV) 1308 *CurrOpnd = XorOpnd(CV); 1309 else { 1310 CurrOpnd->Invalidate(); 1311 continue; 1312 } 1313 } 1314 1315 if (!PrevOpnd || CurrOpnd->getSymbolicPart() != PrevOpnd->getSymbolicPart()) { 1316 PrevOpnd = CurrOpnd; 1317 continue; 1318 } 1319 1320 // step 3.2: When previous and current operands share the same symbolic 1321 // value, try to simplify "PrevOpnd ^ CurrOpnd ^ ConstOpnd" 1322 // 1323 if (CombineXorOpnd(I, CurrOpnd, PrevOpnd, ConstOpnd, CV)) { 1324 // Remove previous operand 1325 PrevOpnd->Invalidate(); 1326 if (CV) { 1327 *CurrOpnd = XorOpnd(CV); 1328 PrevOpnd = CurrOpnd; 1329 } else { 1330 CurrOpnd->Invalidate(); 1331 PrevOpnd = 0; 1332 } 1333 Changed = true; 1334 } 1335 } 1336 1337 // Step 4: Reassemble the Ops 1338 if (Changed) { 1339 Ops.clear(); 1340 for (unsigned int i = 0, e = Opnds.size(); i < e; i++) { 1341 XorOpnd &O = Opnds[i]; 1342 if (O.isInvalid()) 1343 continue; 1344 ValueEntry VE(getRank(O.getValue()), O.getValue()); 1345 Ops.push_back(VE); 1346 } 1347 if (ConstOpnd != 0) { 1348 Value *C = ConstantInt::get(Ty->getContext(), ConstOpnd); 1349 ValueEntry VE(getRank(C), C); 1350 Ops.push_back(VE); 1351 } 1352 int Sz = Ops.size(); 1353 if (Sz == 1) 1354 return Ops.back().Op; 1355 else if (Sz == 0) { 1356 assert(ConstOpnd == 0); 1357 return ConstantInt::get(Ty->getContext(), ConstOpnd); 1358 } 1359 } 1360 1361 return 0; 1362} 1363 1364/// OptimizeAdd - Optimize a series of operands to an 'add' instruction. This 1365/// optimizes based on identities. If it can be reduced to a single Value, it 1366/// is returned, otherwise the Ops list is mutated as necessary. 1367Value *Reassociate::OptimizeAdd(Instruction *I, 1368 SmallVectorImpl<ValueEntry> &Ops) { 1369 // Scan the operand lists looking for X and -X pairs. If we find any, we 1370 // can simplify the expression. X+-X == 0. While we're at it, scan for any 1371 // duplicates. We want to canonicalize Y+Y+Y+Z -> 3*Y+Z. 1372 // 1373 // TODO: We could handle "X + ~X" -> "-1" if we wanted, since "-X = ~X+1". 1374 // 1375 for (unsigned i = 0, e = Ops.size(); i != e; ++i) { 1376 Value *TheOp = Ops[i].Op; 1377 // Check to see if we've seen this operand before. If so, we factor all 1378 // instances of the operand together. Due to our sorting criteria, we know 1379 // that these need to be next to each other in the vector. 1380 if (i+1 != Ops.size() && Ops[i+1].Op == TheOp) { 1381 // Rescan the list, remove all instances of this operand from the expr. 1382 unsigned NumFound = 0; 1383 do { 1384 Ops.erase(Ops.begin()+i); 1385 ++NumFound; 1386 } while (i != Ops.size() && Ops[i].Op == TheOp); 1387 1388 DEBUG(errs() << "\nFACTORING [" << NumFound << "]: " << *TheOp << '\n'); 1389 ++NumFactor; 1390 1391 // Insert a new multiply. 1392 Value *Mul = ConstantInt::get(cast<IntegerType>(I->getType()), NumFound); 1393 Mul = BinaryOperator::CreateMul(TheOp, Mul, "factor", I); 1394 1395 // Now that we have inserted a multiply, optimize it. This allows us to 1396 // handle cases that require multiple factoring steps, such as this: 1397 // (X*2) + (X*2) + (X*2) -> (X*2)*3 -> X*6 1398 RedoInsts.insert(cast<Instruction>(Mul)); 1399 1400 // If every add operand was a duplicate, return the multiply. 1401 if (Ops.empty()) 1402 return Mul; 1403 1404 // Otherwise, we had some input that didn't have the dupe, such as 1405 // "A + A + B" -> "A*2 + B". Add the new multiply to the list of 1406 // things being added by this operation. 1407 Ops.insert(Ops.begin(), ValueEntry(getRank(Mul), Mul)); 1408 1409 --i; 1410 e = Ops.size(); 1411 continue; 1412 } 1413 1414 // Check for X and -X in the operand list. 1415 if (!BinaryOperator::isNeg(TheOp)) 1416 continue; 1417 1418 Value *X = BinaryOperator::getNegArgument(TheOp); 1419 unsigned FoundX = FindInOperandList(Ops, i, X); 1420 if (FoundX == i) 1421 continue; 1422 1423 // Remove X and -X from the operand list. 1424 if (Ops.size() == 2) 1425 return Constant::getNullValue(X->getType()); 1426 1427 Ops.erase(Ops.begin()+i); 1428 if (i < FoundX) 1429 --FoundX; 1430 else 1431 --i; // Need to back up an extra one. 1432 Ops.erase(Ops.begin()+FoundX); 1433 ++NumAnnihil; 1434 --i; // Revisit element. 1435 e -= 2; // Removed two elements. 1436 } 1437 1438 // Scan the operand list, checking to see if there are any common factors 1439 // between operands. Consider something like A*A+A*B*C+D. We would like to 1440 // reassociate this to A*(A+B*C)+D, which reduces the number of multiplies. 1441 // To efficiently find this, we count the number of times a factor occurs 1442 // for any ADD operands that are MULs. 1443 DenseMap<Value*, unsigned> FactorOccurrences; 1444 1445 // Keep track of each multiply we see, to avoid triggering on (X*4)+(X*4) 1446 // where they are actually the same multiply. 1447 unsigned MaxOcc = 0; 1448 Value *MaxOccVal = 0; 1449 for (unsigned i = 0, e = Ops.size(); i != e; ++i) { 1450 BinaryOperator *BOp = isReassociableOp(Ops[i].Op, Instruction::Mul); 1451 if (!BOp) 1452 continue; 1453 1454 // Compute all of the factors of this added value. 1455 SmallVector<Value*, 8> Factors; 1456 FindSingleUseMultiplyFactors(BOp, Factors, Ops); 1457 assert(Factors.size() > 1 && "Bad linearize!"); 1458 1459 // Add one to FactorOccurrences for each unique factor in this op. 1460 SmallPtrSet<Value*, 8> Duplicates; 1461 for (unsigned i = 0, e = Factors.size(); i != e; ++i) { 1462 Value *Factor = Factors[i]; 1463 if (!Duplicates.insert(Factor)) continue; 1464 1465 unsigned Occ = ++FactorOccurrences[Factor]; 1466 if (Occ > MaxOcc) { MaxOcc = Occ; MaxOccVal = Factor; } 1467 1468 // If Factor is a negative constant, add the negated value as a factor 1469 // because we can percolate the negate out. Watch for minint, which 1470 // cannot be positivified. 1471 if (ConstantInt *CI = dyn_cast<ConstantInt>(Factor)) 1472 if (CI->isNegative() && !CI->isMinValue(true)) { 1473 Factor = ConstantInt::get(CI->getContext(), -CI->getValue()); 1474 assert(!Duplicates.count(Factor) && 1475 "Shouldn't have two constant factors, missed a canonicalize"); 1476 1477 unsigned Occ = ++FactorOccurrences[Factor]; 1478 if (Occ > MaxOcc) { MaxOcc = Occ; MaxOccVal = Factor; } 1479 } 1480 } 1481 } 1482 1483 // If any factor occurred more than one time, we can pull it out. 1484 if (MaxOcc > 1) { 1485 DEBUG(errs() << "\nFACTORING [" << MaxOcc << "]: " << *MaxOccVal << '\n'); 1486 ++NumFactor; 1487 1488 // Create a new instruction that uses the MaxOccVal twice. If we don't do 1489 // this, we could otherwise run into situations where removing a factor 1490 // from an expression will drop a use of maxocc, and this can cause 1491 // RemoveFactorFromExpression on successive values to behave differently. 1492 Instruction *DummyInst = BinaryOperator::CreateAdd(MaxOccVal, MaxOccVal); 1493 SmallVector<WeakVH, 4> NewMulOps; 1494 for (unsigned i = 0; i != Ops.size(); ++i) { 1495 // Only try to remove factors from expressions we're allowed to. 1496 BinaryOperator *BOp = isReassociableOp(Ops[i].Op, Instruction::Mul); 1497 if (!BOp) 1498 continue; 1499 1500 if (Value *V = RemoveFactorFromExpression(Ops[i].Op, MaxOccVal)) { 1501 // The factorized operand may occur several times. Convert them all in 1502 // one fell swoop. 1503 for (unsigned j = Ops.size(); j != i;) { 1504 --j; 1505 if (Ops[j].Op == Ops[i].Op) { 1506 NewMulOps.push_back(V); 1507 Ops.erase(Ops.begin()+j); 1508 } 1509 } 1510 --i; 1511 } 1512 } 1513 1514 // No need for extra uses anymore. 1515 delete DummyInst; 1516 1517 unsigned NumAddedValues = NewMulOps.size(); 1518 Value *V = EmitAddTreeOfValues(I, NewMulOps); 1519 1520 // Now that we have inserted the add tree, optimize it. This allows us to 1521 // handle cases that require multiple factoring steps, such as this: 1522 // A*A*B + A*A*C --> A*(A*B+A*C) --> A*(A*(B+C)) 1523 assert(NumAddedValues > 1 && "Each occurrence should contribute a value"); 1524 (void)NumAddedValues; 1525 if (Instruction *VI = dyn_cast<Instruction>(V)) 1526 RedoInsts.insert(VI); 1527 1528 // Create the multiply. 1529 Instruction *V2 = BinaryOperator::CreateMul(V, MaxOccVal, "tmp", I); 1530 1531 // Rerun associate on the multiply in case the inner expression turned into 1532 // a multiply. We want to make sure that we keep things in canonical form. 1533 RedoInsts.insert(V2); 1534 1535 // If every add operand included the factor (e.g. "A*B + A*C"), then the 1536 // entire result expression is just the multiply "A*(B+C)". 1537 if (Ops.empty()) 1538 return V2; 1539 1540 // Otherwise, we had some input that didn't have the factor, such as 1541 // "A*B + A*C + D" -> "A*(B+C) + D". Add the new multiply to the list of 1542 // things being added by this operation. 1543 Ops.insert(Ops.begin(), ValueEntry(getRank(V2), V2)); 1544 } 1545 1546 return 0; 1547} 1548 1549/// \brief Build up a vector of value/power pairs factoring a product. 1550/// 1551/// Given a series of multiplication operands, build a vector of factors and 1552/// the powers each is raised to when forming the final product. Sort them in 1553/// the order of descending power. 1554/// 1555/// (x*x) -> [(x, 2)] 1556/// ((x*x)*x) -> [(x, 3)] 1557/// ((((x*y)*x)*y)*x) -> [(x, 3), (y, 2)] 1558/// 1559/// \returns Whether any factors have a power greater than one. 1560bool Reassociate::collectMultiplyFactors(SmallVectorImpl<ValueEntry> &Ops, 1561 SmallVectorImpl<Factor> &Factors) { 1562 // FIXME: Have Ops be (ValueEntry, Multiplicity) pairs, simplifying this. 1563 // Compute the sum of powers of simplifiable factors. 1564 unsigned FactorPowerSum = 0; 1565 for (unsigned Idx = 1, Size = Ops.size(); Idx < Size; ++Idx) { 1566 Value *Op = Ops[Idx-1].Op; 1567 1568 // Count the number of occurrences of this value. 1569 unsigned Count = 1; 1570 for (; Idx < Size && Ops[Idx].Op == Op; ++Idx) 1571 ++Count; 1572 // Track for simplification all factors which occur 2 or more times. 1573 if (Count > 1) 1574 FactorPowerSum += Count; 1575 } 1576 1577 // We can only simplify factors if the sum of the powers of our simplifiable 1578 // factors is 4 or higher. When that is the case, we will *always* have 1579 // a simplification. This is an important invariant to prevent cyclicly 1580 // trying to simplify already minimal formations. 1581 if (FactorPowerSum < 4) 1582 return false; 1583 1584 // Now gather the simplifiable factors, removing them from Ops. 1585 FactorPowerSum = 0; 1586 for (unsigned Idx = 1; Idx < Ops.size(); ++Idx) { 1587 Value *Op = Ops[Idx-1].Op; 1588 1589 // Count the number of occurrences of this value. 1590 unsigned Count = 1; 1591 for (; Idx < Ops.size() && Ops[Idx].Op == Op; ++Idx) 1592 ++Count; 1593 if (Count == 1) 1594 continue; 1595 // Move an even number of occurrences to Factors. 1596 Count &= ~1U; 1597 Idx -= Count; 1598 FactorPowerSum += Count; 1599 Factors.push_back(Factor(Op, Count)); 1600 Ops.erase(Ops.begin()+Idx, Ops.begin()+Idx+Count); 1601 } 1602 1603 // None of the adjustments above should have reduced the sum of factor powers 1604 // below our mininum of '4'. 1605 assert(FactorPowerSum >= 4); 1606 1607 std::stable_sort(Factors.begin(), Factors.end(), Factor::PowerDescendingSorter()); 1608 return true; 1609} 1610 1611/// \brief Build a tree of multiplies, computing the product of Ops. 1612static Value *buildMultiplyTree(IRBuilder<> &Builder, 1613 SmallVectorImpl<Value*> &Ops) { 1614 if (Ops.size() == 1) 1615 return Ops.back(); 1616 1617 Value *LHS = Ops.pop_back_val(); 1618 do { 1619 LHS = Builder.CreateMul(LHS, Ops.pop_back_val()); 1620 } while (!Ops.empty()); 1621 1622 return LHS; 1623} 1624 1625/// \brief Build a minimal multiplication DAG for (a^x)*(b^y)*(c^z)*... 1626/// 1627/// Given a vector of values raised to various powers, where no two values are 1628/// equal and the powers are sorted in decreasing order, compute the minimal 1629/// DAG of multiplies to compute the final product, and return that product 1630/// value. 1631Value *Reassociate::buildMinimalMultiplyDAG(IRBuilder<> &Builder, 1632 SmallVectorImpl<Factor> &Factors) { 1633 assert(Factors[0].Power); 1634 SmallVector<Value *, 4> OuterProduct; 1635 for (unsigned LastIdx = 0, Idx = 1, Size = Factors.size(); 1636 Idx < Size && Factors[Idx].Power > 0; ++Idx) { 1637 if (Factors[Idx].Power != Factors[LastIdx].Power) { 1638 LastIdx = Idx; 1639 continue; 1640 } 1641 1642 // We want to multiply across all the factors with the same power so that 1643 // we can raise them to that power as a single entity. Build a mini tree 1644 // for that. 1645 SmallVector<Value *, 4> InnerProduct; 1646 InnerProduct.push_back(Factors[LastIdx].Base); 1647 do { 1648 InnerProduct.push_back(Factors[Idx].Base); 1649 ++Idx; 1650 } while (Idx < Size && Factors[Idx].Power == Factors[LastIdx].Power); 1651 1652 // Reset the base value of the first factor to the new expression tree. 1653 // We'll remove all the factors with the same power in a second pass. 1654 Value *M = Factors[LastIdx].Base = buildMultiplyTree(Builder, InnerProduct); 1655 if (Instruction *MI = dyn_cast<Instruction>(M)) 1656 RedoInsts.insert(MI); 1657 1658 LastIdx = Idx; 1659 } 1660 // Unique factors with equal powers -- we've folded them into the first one's 1661 // base. 1662 Factors.erase(std::unique(Factors.begin(), Factors.end(), 1663 Factor::PowerEqual()), 1664 Factors.end()); 1665 1666 // Iteratively collect the base of each factor with an add power into the 1667 // outer product, and halve each power in preparation for squaring the 1668 // expression. 1669 for (unsigned Idx = 0, Size = Factors.size(); Idx != Size; ++Idx) { 1670 if (Factors[Idx].Power & 1) 1671 OuterProduct.push_back(Factors[Idx].Base); 1672 Factors[Idx].Power >>= 1; 1673 } 1674 if (Factors[0].Power) { 1675 Value *SquareRoot = buildMinimalMultiplyDAG(Builder, Factors); 1676 OuterProduct.push_back(SquareRoot); 1677 OuterProduct.push_back(SquareRoot); 1678 } 1679 if (OuterProduct.size() == 1) 1680 return OuterProduct.front(); 1681 1682 Value *V = buildMultiplyTree(Builder, OuterProduct); 1683 return V; 1684} 1685 1686Value *Reassociate::OptimizeMul(BinaryOperator *I, 1687 SmallVectorImpl<ValueEntry> &Ops) { 1688 // We can only optimize the multiplies when there is a chain of more than 1689 // three, such that a balanced tree might require fewer total multiplies. 1690 if (Ops.size() < 4) 1691 return 0; 1692 1693 // Try to turn linear trees of multiplies without other uses of the 1694 // intermediate stages into minimal multiply DAGs with perfect sub-expression 1695 // re-use. 1696 SmallVector<Factor, 4> Factors; 1697 if (!collectMultiplyFactors(Ops, Factors)) 1698 return 0; // All distinct factors, so nothing left for us to do. 1699 1700 IRBuilder<> Builder(I); 1701 Value *V = buildMinimalMultiplyDAG(Builder, Factors); 1702 if (Ops.empty()) 1703 return V; 1704 1705 ValueEntry NewEntry = ValueEntry(getRank(V), V); 1706 Ops.insert(std::lower_bound(Ops.begin(), Ops.end(), NewEntry), NewEntry); 1707 return 0; 1708} 1709 1710Value *Reassociate::OptimizeExpression(BinaryOperator *I, 1711 SmallVectorImpl<ValueEntry> &Ops) { 1712 // Now that we have the linearized expression tree, try to optimize it. 1713 // Start by folding any constants that we found. 1714 Constant *Cst = 0; 1715 unsigned Opcode = I->getOpcode(); 1716 while (!Ops.empty() && isa<Constant>(Ops.back().Op)) { 1717 Constant *C = cast<Constant>(Ops.pop_back_val().Op); 1718 Cst = Cst ? ConstantExpr::get(Opcode, C, Cst) : C; 1719 } 1720 // If there was nothing but constants then we are done. 1721 if (Ops.empty()) 1722 return Cst; 1723 1724 // Put the combined constant back at the end of the operand list, except if 1725 // there is no point. For example, an add of 0 gets dropped here, while a 1726 // multiplication by zero turns the whole expression into zero. 1727 if (Cst && Cst != ConstantExpr::getBinOpIdentity(Opcode, I->getType())) { 1728 if (Cst == ConstantExpr::getBinOpAbsorber(Opcode, I->getType())) 1729 return Cst; 1730 Ops.push_back(ValueEntry(0, Cst)); 1731 } 1732 1733 if (Ops.size() == 1) return Ops[0].Op; 1734 1735 // Handle destructive annihilation due to identities between elements in the 1736 // argument list here. 1737 unsigned NumOps = Ops.size(); 1738 switch (Opcode) { 1739 default: break; 1740 case Instruction::And: 1741 case Instruction::Or: 1742 if (Value *Result = OptimizeAndOrXor(Opcode, Ops)) 1743 return Result; 1744 break; 1745 1746 case Instruction::Xor: 1747 if (Value *Result = OptimizeXor(I, Ops)) 1748 return Result; 1749 break; 1750 1751 case Instruction::Add: 1752 if (Value *Result = OptimizeAdd(I, Ops)) 1753 return Result; 1754 break; 1755 1756 case Instruction::Mul: 1757 if (Value *Result = OptimizeMul(I, Ops)) 1758 return Result; 1759 break; 1760 } 1761 1762 if (Ops.size() != NumOps) 1763 return OptimizeExpression(I, Ops); 1764 return 0; 1765} 1766 1767/// EraseInst - Zap the given instruction, adding interesting operands to the 1768/// work list. 1769void Reassociate::EraseInst(Instruction *I) { 1770 assert(isInstructionTriviallyDead(I) && "Trivially dead instructions only!"); 1771 SmallVector<Value*, 8> Ops(I->op_begin(), I->op_end()); 1772 // Erase the dead instruction. 1773 ValueRankMap.erase(I); 1774 RedoInsts.remove(I); 1775 I->eraseFromParent(); 1776 // Optimize its operands. 1777 SmallPtrSet<Instruction *, 8> Visited; // Detect self-referential nodes. 1778 for (unsigned i = 0, e = Ops.size(); i != e; ++i) 1779 if (Instruction *Op = dyn_cast<Instruction>(Ops[i])) { 1780 // If this is a node in an expression tree, climb to the expression root 1781 // and add that since that's where optimization actually happens. 1782 unsigned Opcode = Op->getOpcode(); 1783 while (Op->hasOneUse() && Op->user_back()->getOpcode() == Opcode && 1784 Visited.insert(Op)) 1785 Op = Op->user_back(); 1786 RedoInsts.insert(Op); 1787 } 1788} 1789 1790/// OptimizeInst - Inspect and optimize the given instruction. Note that erasing 1791/// instructions is not allowed. 1792void Reassociate::OptimizeInst(Instruction *I) { 1793 // Only consider operations that we understand. 1794 if (!isa<BinaryOperator>(I)) 1795 return; 1796 1797 if (I->getOpcode() == Instruction::Shl && 1798 isa<ConstantInt>(I->getOperand(1))) 1799 // If an operand of this shift is a reassociable multiply, or if the shift 1800 // is used by a reassociable multiply or add, turn into a multiply. 1801 if (isReassociableOp(I->getOperand(0), Instruction::Mul) || 1802 (I->hasOneUse() && 1803 (isReassociableOp(I->user_back(), Instruction::Mul) || 1804 isReassociableOp(I->user_back(), Instruction::Add)))) { 1805 Instruction *NI = ConvertShiftToMul(I); 1806 RedoInsts.insert(I); 1807 MadeChange = true; 1808 I = NI; 1809 } 1810 1811 // Floating point binary operators are not associative, but we can still 1812 // commute (some) of them, to canonicalize the order of their operands. 1813 // This can potentially expose more CSE opportunities, and makes writing 1814 // other transformations simpler. 1815 if ((I->getType()->isFloatingPointTy() || I->getType()->isVectorTy())) { 1816 // FAdd and FMul can be commuted. 1817 if (I->getOpcode() != Instruction::FMul && 1818 I->getOpcode() != Instruction::FAdd) 1819 return; 1820 1821 Value *LHS = I->getOperand(0); 1822 Value *RHS = I->getOperand(1); 1823 unsigned LHSRank = getRank(LHS); 1824 unsigned RHSRank = getRank(RHS); 1825 1826 // Sort the operands by rank. 1827 if (RHSRank < LHSRank) { 1828 I->setOperand(0, RHS); 1829 I->setOperand(1, LHS); 1830 } 1831 1832 return; 1833 } 1834 1835 // Do not reassociate boolean (i1) expressions. We want to preserve the 1836 // original order of evaluation for short-circuited comparisons that 1837 // SimplifyCFG has folded to AND/OR expressions. If the expression 1838 // is not further optimized, it is likely to be transformed back to a 1839 // short-circuited form for code gen, and the source order may have been 1840 // optimized for the most likely conditions. 1841 if (I->getType()->isIntegerTy(1)) 1842 return; 1843 1844 // If this is a subtract instruction which is not already in negate form, 1845 // see if we can convert it to X+-Y. 1846 if (I->getOpcode() == Instruction::Sub) { 1847 if (ShouldBreakUpSubtract(I)) { 1848 Instruction *NI = BreakUpSubtract(I); 1849 RedoInsts.insert(I); 1850 MadeChange = true; 1851 I = NI; 1852 } else if (BinaryOperator::isNeg(I)) { 1853 // Otherwise, this is a negation. See if the operand is a multiply tree 1854 // and if this is not an inner node of a multiply tree. 1855 if (isReassociableOp(I->getOperand(1), Instruction::Mul) && 1856 (!I->hasOneUse() || 1857 !isReassociableOp(I->user_back(), Instruction::Mul))) { 1858 Instruction *NI = LowerNegateToMultiply(I); 1859 RedoInsts.insert(I); 1860 MadeChange = true; 1861 I = NI; 1862 } 1863 } 1864 } 1865 1866 // If this instruction is an associative binary operator, process it. 1867 if (!I->isAssociative()) return; 1868 BinaryOperator *BO = cast<BinaryOperator>(I); 1869 1870 // If this is an interior node of a reassociable tree, ignore it until we 1871 // get to the root of the tree, to avoid N^2 analysis. 1872 unsigned Opcode = BO->getOpcode(); 1873 if (BO->hasOneUse() && BO->user_back()->getOpcode() == Opcode) 1874 return; 1875 1876 // If this is an add tree that is used by a sub instruction, ignore it 1877 // until we process the subtract. 1878 if (BO->hasOneUse() && BO->getOpcode() == Instruction::Add && 1879 cast<Instruction>(BO->user_back())->getOpcode() == Instruction::Sub) 1880 return; 1881 1882 ReassociateExpression(BO); 1883} 1884 1885void Reassociate::ReassociateExpression(BinaryOperator *I) { 1886 1887 // First, walk the expression tree, linearizing the tree, collecting the 1888 // operand information. 1889 SmallVector<RepeatedValue, 8> Tree; 1890 MadeChange |= LinearizeExprTree(I, Tree); 1891 SmallVector<ValueEntry, 8> Ops; 1892 Ops.reserve(Tree.size()); 1893 for (unsigned i = 0, e = Tree.size(); i != e; ++i) { 1894 RepeatedValue E = Tree[i]; 1895 Ops.append(E.second.getZExtValue(), 1896 ValueEntry(getRank(E.first), E.first)); 1897 } 1898 1899 DEBUG(dbgs() << "RAIn:\t"; PrintOps(I, Ops); dbgs() << '\n'); 1900 1901 // Now that we have linearized the tree to a list and have gathered all of 1902 // the operands and their ranks, sort the operands by their rank. Use a 1903 // stable_sort so that values with equal ranks will have their relative 1904 // positions maintained (and so the compiler is deterministic). Note that 1905 // this sorts so that the highest ranking values end up at the beginning of 1906 // the vector. 1907 std::stable_sort(Ops.begin(), Ops.end()); 1908 1909 // OptimizeExpression - Now that we have the expression tree in a convenient 1910 // sorted form, optimize it globally if possible. 1911 if (Value *V = OptimizeExpression(I, Ops)) { 1912 if (V == I) 1913 // Self-referential expression in unreachable code. 1914 return; 1915 // This expression tree simplified to something that isn't a tree, 1916 // eliminate it. 1917 DEBUG(dbgs() << "Reassoc to scalar: " << *V << '\n'); 1918 I->replaceAllUsesWith(V); 1919 if (Instruction *VI = dyn_cast<Instruction>(V)) 1920 VI->setDebugLoc(I->getDebugLoc()); 1921 RedoInsts.insert(I); 1922 ++NumAnnihil; 1923 return; 1924 } 1925 1926 // We want to sink immediates as deeply as possible except in the case where 1927 // this is a multiply tree used only by an add, and the immediate is a -1. 1928 // In this case we reassociate to put the negation on the outside so that we 1929 // can fold the negation into the add: (-X)*Y + Z -> Z-X*Y 1930 if (I->getOpcode() == Instruction::Mul && I->hasOneUse() && 1931 cast<Instruction>(I->user_back())->getOpcode() == Instruction::Add && 1932 isa<ConstantInt>(Ops.back().Op) && 1933 cast<ConstantInt>(Ops.back().Op)->isAllOnesValue()) { 1934 ValueEntry Tmp = Ops.pop_back_val(); 1935 Ops.insert(Ops.begin(), Tmp); 1936 } 1937 1938 DEBUG(dbgs() << "RAOut:\t"; PrintOps(I, Ops); dbgs() << '\n'); 1939 1940 if (Ops.size() == 1) { 1941 if (Ops[0].Op == I) 1942 // Self-referential expression in unreachable code. 1943 return; 1944 1945 // This expression tree simplified to something that isn't a tree, 1946 // eliminate it. 1947 I->replaceAllUsesWith(Ops[0].Op); 1948 if (Instruction *OI = dyn_cast<Instruction>(Ops[0].Op)) 1949 OI->setDebugLoc(I->getDebugLoc()); 1950 RedoInsts.insert(I); 1951 return; 1952 } 1953 1954 // Now that we ordered and optimized the expressions, splat them back into 1955 // the expression tree, removing any unneeded nodes. 1956 RewriteExprTree(I, Ops); 1957} 1958 1959bool Reassociate::runOnFunction(Function &F) { 1960 if (skipOptnoneFunction(F)) 1961 return false; 1962 1963 // Calculate the rank map for F 1964 BuildRankMap(F); 1965 1966 MadeChange = false; 1967 for (Function::iterator BI = F.begin(), BE = F.end(); BI != BE; ++BI) { 1968 // Optimize every instruction in the basic block. 1969 for (BasicBlock::iterator II = BI->begin(), IE = BI->end(); II != IE; ) 1970 if (isInstructionTriviallyDead(II)) { 1971 EraseInst(II++); 1972 } else { 1973 OptimizeInst(II); 1974 assert(II->getParent() == BI && "Moved to a different block!"); 1975 ++II; 1976 } 1977 1978 // If this produced extra instructions to optimize, handle them now. 1979 while (!RedoInsts.empty()) { 1980 Instruction *I = RedoInsts.pop_back_val(); 1981 if (isInstructionTriviallyDead(I)) 1982 EraseInst(I); 1983 else 1984 OptimizeInst(I); 1985 } 1986 } 1987 1988 // We are done with the rank map. 1989 RankMap.clear(); 1990 ValueRankMap.clear(); 1991 1992 return MadeChange; 1993} 1994