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 #include "llvm/Transforms/Scalar.h" 24 #include "llvm/ADT/DenseMap.h" 25 #include "llvm/ADT/PostOrderIterator.h" 26 #include "llvm/ADT/STLExtras.h" 27 #include "llvm/ADT/SetVector.h" 28 #include "llvm/ADT/Statistic.h" 29 #include "llvm/Analysis/GlobalsModRef.h" 30 #include "llvm/Analysis/ValueTracking.h" 31 #include "llvm/IR/CFG.h" 32 #include "llvm/IR/Constants.h" 33 #include "llvm/IR/DerivedTypes.h" 34 #include "llvm/IR/Function.h" 35 #include "llvm/IR/IRBuilder.h" 36 #include "llvm/IR/Instructions.h" 37 #include "llvm/IR/IntrinsicInst.h" 38 #include "llvm/IR/ValueHandle.h" 39 #include "llvm/Pass.h" 40 #include "llvm/Support/Debug.h" 41 #include "llvm/Support/raw_ostream.h" 42 #include "llvm/Transforms/Utils/Local.h" 43 #include <algorithm> 44 using namespace llvm; 45 46 #define DEBUG_TYPE "reassociate" 47 48 STATISTIC(NumChanged, "Number of insts reassociated"); 49 STATISTIC(NumAnnihil, "Number of expr tree annihilated"); 50 STATISTIC(NumFactor , "Number of multiplies factored"); 51 52 namespace { 53 struct ValueEntry { 54 unsigned Rank; 55 Value *Op; 56 ValueEntry(unsigned R, Value *O) : Rank(R), Op(O) {} 57 }; 58 inline bool operator<(const ValueEntry &LHS, const ValueEntry &RHS) { 59 return LHS.Rank > RHS.Rank; // Sort so that highest rank goes to start. 60 } 61 } 62 63 #ifndef NDEBUG 64 /// Print out the expression identified in the Ops list. 65 /// 66 static void PrintOps(Instruction *I, const SmallVectorImpl<ValueEntry> &Ops) { 67 Module *M = I->getModule(); 68 dbgs() << Instruction::getOpcodeName(I->getOpcode()) << " " 69 << *Ops[0].Op->getType() << '\t'; 70 for (unsigned i = 0, e = Ops.size(); i != e; ++i) { 71 dbgs() << "[ "; 72 Ops[i].Op->printAsOperand(dbgs(), false, M); 73 dbgs() << ", #" << Ops[i].Rank << "] "; 74 } 75 } 76 #endif 77 78 namespace { 79 /// \brief Utility class representing a base and exponent pair which form one 80 /// factor of some product. 81 struct Factor { 82 Value *Base; 83 unsigned Power; 84 85 Factor(Value *Base, unsigned Power) : Base(Base), Power(Power) {} 86 87 /// \brief Sort factors in descending order by their power. 88 struct PowerDescendingSorter { 89 bool operator()(const Factor &LHS, const Factor &RHS) { 90 return LHS.Power > RHS.Power; 91 } 92 }; 93 94 /// \brief Compare factors for equal powers. 95 struct PowerEqual { 96 bool operator()(const Factor &LHS, const Factor &RHS) { 97 return LHS.Power == RHS.Power; 98 } 99 }; 100 }; 101 102 /// Utility class representing a non-constant Xor-operand. We classify 103 /// non-constant Xor-Operands into two categories: 104 /// C1) The operand is in the form "X & C", where C is a constant and C != ~0 105 /// C2) 106 /// C2.1) The operand is in the form of "X | C", where C is a non-zero 107 /// constant. 108 /// C2.2) Any operand E which doesn't fall into C1 and C2.1, we view this 109 /// operand as "E | 0" 110 class XorOpnd { 111 public: 112 XorOpnd(Value *V); 113 114 bool isInvalid() const { return SymbolicPart == nullptr; } 115 bool isOrExpr() const { return isOr; } 116 Value *getValue() const { return OrigVal; } 117 Value *getSymbolicPart() const { return SymbolicPart; } 118 unsigned getSymbolicRank() const { return SymbolicRank; } 119 const APInt &getConstPart() const { return ConstPart; } 120 121 void Invalidate() { SymbolicPart = OrigVal = nullptr; } 122 void setSymbolicRank(unsigned R) { SymbolicRank = R; } 123 124 // Sort the XorOpnd-Pointer in ascending order of symbolic-value-rank. 125 // The purpose is twofold: 126 // 1) Cluster together the operands sharing the same symbolic-value. 127 // 2) Operand having smaller symbolic-value-rank is permuted earlier, which 128 // could potentially shorten crital path, and expose more loop-invariants. 129 // Note that values' rank are basically defined in RPO order (FIXME). 130 // So, if Rank(X) < Rank(Y) < Rank(Z), it means X is defined earlier 131 // than Y which is defined earlier than Z. Permute "x | 1", "Y & 2", 132 // "z" in the order of X-Y-Z is better than any other orders. 133 struct PtrSortFunctor { 134 bool operator()(XorOpnd * const &LHS, XorOpnd * const &RHS) { 135 return LHS->getSymbolicRank() < RHS->getSymbolicRank(); 136 } 137 }; 138 private: 139 Value *OrigVal; 140 Value *SymbolicPart; 141 APInt ConstPart; 142 unsigned SymbolicRank; 143 bool isOr; 144 }; 145 } 146 147 namespace { 148 class Reassociate : public FunctionPass { 149 DenseMap<BasicBlock*, unsigned> RankMap; 150 DenseMap<AssertingVH<Value>, unsigned> ValueRankMap; 151 SetVector<AssertingVH<Instruction> > RedoInsts; 152 bool MadeChange; 153 public: 154 static char ID; // Pass identification, replacement for typeid 155 Reassociate() : FunctionPass(ID) { 156 initializeReassociatePass(*PassRegistry::getPassRegistry()); 157 } 158 159 bool runOnFunction(Function &F) override; 160 161 void getAnalysisUsage(AnalysisUsage &AU) const override { 162 AU.setPreservesCFG(); 163 AU.addPreserved<GlobalsAAWrapperPass>(); 164 } 165 private: 166 void BuildRankMap(Function &F); 167 unsigned getRank(Value *V); 168 void canonicalizeOperands(Instruction *I); 169 void ReassociateExpression(BinaryOperator *I); 170 void RewriteExprTree(BinaryOperator *I, SmallVectorImpl<ValueEntry> &Ops); 171 Value *OptimizeExpression(BinaryOperator *I, 172 SmallVectorImpl<ValueEntry> &Ops); 173 Value *OptimizeAdd(Instruction *I, SmallVectorImpl<ValueEntry> &Ops); 174 Value *OptimizeXor(Instruction *I, SmallVectorImpl<ValueEntry> &Ops); 175 bool CombineXorOpnd(Instruction *I, XorOpnd *Opnd1, APInt &ConstOpnd, 176 Value *&Res); 177 bool CombineXorOpnd(Instruction *I, XorOpnd *Opnd1, XorOpnd *Opnd2, 178 APInt &ConstOpnd, Value *&Res); 179 bool collectMultiplyFactors(SmallVectorImpl<ValueEntry> &Ops, 180 SmallVectorImpl<Factor> &Factors); 181 Value *buildMinimalMultiplyDAG(IRBuilder<> &Builder, 182 SmallVectorImpl<Factor> &Factors); 183 Value *OptimizeMul(BinaryOperator *I, SmallVectorImpl<ValueEntry> &Ops); 184 Value *RemoveFactorFromExpression(Value *V, Value *Factor); 185 void EraseInst(Instruction *I); 186 void OptimizeInst(Instruction *I); 187 Instruction *canonicalizeNegConstExpr(Instruction *I); 188 }; 189 } 190 191 XorOpnd::XorOpnd(Value *V) { 192 assert(!isa<ConstantInt>(V) && "No ConstantInt"); 193 OrigVal = V; 194 Instruction *I = dyn_cast<Instruction>(V); 195 SymbolicRank = 0; 196 197 if (I && (I->getOpcode() == Instruction::Or || 198 I->getOpcode() == Instruction::And)) { 199 Value *V0 = I->getOperand(0); 200 Value *V1 = I->getOperand(1); 201 if (isa<ConstantInt>(V0)) 202 std::swap(V0, V1); 203 204 if (ConstantInt *C = dyn_cast<ConstantInt>(V1)) { 205 ConstPart = C->getValue(); 206 SymbolicPart = V0; 207 isOr = (I->getOpcode() == Instruction::Or); 208 return; 209 } 210 } 211 212 // view the operand as "V | 0" 213 SymbolicPart = V; 214 ConstPart = APInt::getNullValue(V->getType()->getIntegerBitWidth()); 215 isOr = true; 216 } 217 218 char Reassociate::ID = 0; 219 INITIALIZE_PASS(Reassociate, "reassociate", 220 "Reassociate expressions", false, false) 221 222 // Public interface to the Reassociate pass 223 FunctionPass *llvm::createReassociatePass() { return new Reassociate(); } 224 225 /// Return true if V is an instruction of the specified opcode and if it 226 /// only has one use. 227 static BinaryOperator *isReassociableOp(Value *V, unsigned Opcode) { 228 if (V->hasOneUse() && isa<Instruction>(V) && 229 cast<Instruction>(V)->getOpcode() == Opcode && 230 (!isa<FPMathOperator>(V) || 231 cast<Instruction>(V)->hasUnsafeAlgebra())) 232 return cast<BinaryOperator>(V); 233 return nullptr; 234 } 235 236 static BinaryOperator *isReassociableOp(Value *V, unsigned Opcode1, 237 unsigned Opcode2) { 238 if (V->hasOneUse() && isa<Instruction>(V) && 239 (cast<Instruction>(V)->getOpcode() == Opcode1 || 240 cast<Instruction>(V)->getOpcode() == Opcode2) && 241 (!isa<FPMathOperator>(V) || 242 cast<Instruction>(V)->hasUnsafeAlgebra())) 243 return cast<BinaryOperator>(V); 244 return nullptr; 245 } 246 247 void Reassociate::BuildRankMap(Function &F) { 248 unsigned i = 2; 249 250 // Assign distinct ranks to function arguments. 251 for (Function::arg_iterator I = F.arg_begin(), E = F.arg_end(); I != E; ++I) { 252 ValueRankMap[&*I] = ++i; 253 DEBUG(dbgs() << "Calculated Rank[" << I->getName() << "] = " << i << "\n"); 254 } 255 256 ReversePostOrderTraversal<Function*> RPOT(&F); 257 for (ReversePostOrderTraversal<Function*>::rpo_iterator I = RPOT.begin(), 258 E = RPOT.end(); I != E; ++I) { 259 BasicBlock *BB = *I; 260 unsigned BBRank = RankMap[BB] = ++i << 16; 261 262 // Walk the basic block, adding precomputed ranks for any instructions that 263 // we cannot move. This ensures that the ranks for these instructions are 264 // all different in the block. 265 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) 266 if (mayBeMemoryDependent(*I)) 267 ValueRankMap[&*I] = ++BBRank; 268 } 269 } 270 271 unsigned Reassociate::getRank(Value *V) { 272 Instruction *I = dyn_cast<Instruction>(V); 273 if (!I) { 274 if (isa<Argument>(V)) return ValueRankMap[V]; // Function argument. 275 return 0; // Otherwise it's a global or constant, rank 0. 276 } 277 278 if (unsigned Rank = ValueRankMap[I]) 279 return Rank; // Rank already known? 280 281 // If this is an expression, return the 1+MAX(rank(LHS), rank(RHS)) so that 282 // we can reassociate expressions for code motion! Since we do not recurse 283 // for PHI nodes, we cannot have infinite recursion here, because there 284 // cannot be loops in the value graph that do not go through PHI nodes. 285 unsigned Rank = 0, MaxRank = RankMap[I->getParent()]; 286 for (unsigned i = 0, e = I->getNumOperands(); 287 i != e && Rank != MaxRank; ++i) 288 Rank = std::max(Rank, getRank(I->getOperand(i))); 289 290 // If this is a not or neg instruction, do not count it for rank. This 291 // assures us that X and ~X will have the same rank. 292 if (!BinaryOperator::isNot(I) && !BinaryOperator::isNeg(I) && 293 !BinaryOperator::isFNeg(I)) 294 ++Rank; 295 296 DEBUG(dbgs() << "Calculated Rank[" << V->getName() << "] = " << Rank << "\n"); 297 298 return ValueRankMap[I] = Rank; 299 } 300 301 // Canonicalize constants to RHS. Otherwise, sort the operands by rank. 302 void Reassociate::canonicalizeOperands(Instruction *I) { 303 assert(isa<BinaryOperator>(I) && "Expected binary operator."); 304 assert(I->isCommutative() && "Expected commutative operator."); 305 306 Value *LHS = I->getOperand(0); 307 Value *RHS = I->getOperand(1); 308 unsigned LHSRank = getRank(LHS); 309 unsigned RHSRank = getRank(RHS); 310 311 if (isa<Constant>(RHS)) 312 return; 313 314 if (isa<Constant>(LHS) || RHSRank < LHSRank) 315 cast<BinaryOperator>(I)->swapOperands(); 316 } 317 318 static BinaryOperator *CreateAdd(Value *S1, Value *S2, const Twine &Name, 319 Instruction *InsertBefore, Value *FlagsOp) { 320 if (S1->getType()->isIntOrIntVectorTy()) 321 return BinaryOperator::CreateAdd(S1, S2, Name, InsertBefore); 322 else { 323 BinaryOperator *Res = 324 BinaryOperator::CreateFAdd(S1, S2, Name, InsertBefore); 325 Res->setFastMathFlags(cast<FPMathOperator>(FlagsOp)->getFastMathFlags()); 326 return Res; 327 } 328 } 329 330 static BinaryOperator *CreateMul(Value *S1, Value *S2, const Twine &Name, 331 Instruction *InsertBefore, Value *FlagsOp) { 332 if (S1->getType()->isIntOrIntVectorTy()) 333 return BinaryOperator::CreateMul(S1, S2, Name, InsertBefore); 334 else { 335 BinaryOperator *Res = 336 BinaryOperator::CreateFMul(S1, S2, Name, InsertBefore); 337 Res->setFastMathFlags(cast<FPMathOperator>(FlagsOp)->getFastMathFlags()); 338 return Res; 339 } 340 } 341 342 static BinaryOperator *CreateNeg(Value *S1, const Twine &Name, 343 Instruction *InsertBefore, Value *FlagsOp) { 344 if (S1->getType()->isIntOrIntVectorTy()) 345 return BinaryOperator::CreateNeg(S1, Name, InsertBefore); 346 else { 347 BinaryOperator *Res = BinaryOperator::CreateFNeg(S1, Name, InsertBefore); 348 Res->setFastMathFlags(cast<FPMathOperator>(FlagsOp)->getFastMathFlags()); 349 return Res; 350 } 351 } 352 353 /// Replace 0-X with X*-1. 354 static BinaryOperator *LowerNegateToMultiply(Instruction *Neg) { 355 Type *Ty = Neg->getType(); 356 Constant *NegOne = Ty->isIntOrIntVectorTy() ? 357 ConstantInt::getAllOnesValue(Ty) : ConstantFP::get(Ty, -1.0); 358 359 BinaryOperator *Res = CreateMul(Neg->getOperand(1), NegOne, "", Neg, Neg); 360 Neg->setOperand(1, Constant::getNullValue(Ty)); // Drop use of op. 361 Res->takeName(Neg); 362 Neg->replaceAllUsesWith(Res); 363 Res->setDebugLoc(Neg->getDebugLoc()); 364 return Res; 365 } 366 367 /// Returns k such that lambda(2^Bitwidth) = 2^k, where lambda is the Carmichael 368 /// function. This means that x^(2^k) === 1 mod 2^Bitwidth for 369 /// every odd x, i.e. x^(2^k) = 1 for every odd x in Bitwidth-bit arithmetic. 370 /// Note that 0 <= k < Bitwidth, and if Bitwidth > 3 then x^(2^k) = 0 for every 371 /// even x in Bitwidth-bit arithmetic. 372 static unsigned CarmichaelShift(unsigned Bitwidth) { 373 if (Bitwidth < 3) 374 return Bitwidth - 1; 375 return Bitwidth - 2; 376 } 377 378 /// Add the extra weight 'RHS' to the existing weight 'LHS', 379 /// reducing the combined weight using any special properties of the operation. 380 /// The existing weight LHS represents the computation X op X op ... op X where 381 /// X occurs LHS times. The combined weight represents X op X op ... op X with 382 /// X occurring LHS + RHS times. If op is "Xor" for example then the combined 383 /// operation is equivalent to X if LHS + RHS is odd, or 0 if LHS + RHS is even; 384 /// the routine returns 1 in LHS in the first case, and 0 in LHS in the second. 385 static void IncorporateWeight(APInt &LHS, const APInt &RHS, unsigned Opcode) { 386 // If we were working with infinite precision arithmetic then the combined 387 // weight would be LHS + RHS. But we are using finite precision arithmetic, 388 // and the APInt sum LHS + RHS may not be correct if it wraps (it is correct 389 // for nilpotent operations and addition, but not for idempotent operations 390 // and multiplication), so it is important to correctly reduce the combined 391 // weight back into range if wrapping would be wrong. 392 393 // If RHS is zero then the weight didn't change. 394 if (RHS.isMinValue()) 395 return; 396 // If LHS is zero then the combined weight is RHS. 397 if (LHS.isMinValue()) { 398 LHS = RHS; 399 return; 400 } 401 // From this point on we know that neither LHS nor RHS is zero. 402 403 if (Instruction::isIdempotent(Opcode)) { 404 // Idempotent means X op X === X, so any non-zero weight is equivalent to a 405 // weight of 1. Keeping weights at zero or one also means that wrapping is 406 // not a problem. 407 assert(LHS == 1 && RHS == 1 && "Weights not reduced!"); 408 return; // Return a weight of 1. 409 } 410 if (Instruction::isNilpotent(Opcode)) { 411 // Nilpotent means X op X === 0, so reduce weights modulo 2. 412 assert(LHS == 1 && RHS == 1 && "Weights not reduced!"); 413 LHS = 0; // 1 + 1 === 0 modulo 2. 414 return; 415 } 416 if (Opcode == Instruction::Add || Opcode == Instruction::FAdd) { 417 // TODO: Reduce the weight by exploiting nsw/nuw? 418 LHS += RHS; 419 return; 420 } 421 422 assert((Opcode == Instruction::Mul || Opcode == Instruction::FMul) && 423 "Unknown associative operation!"); 424 unsigned Bitwidth = LHS.getBitWidth(); 425 // If CM is the Carmichael number then a weight W satisfying W >= CM+Bitwidth 426 // can be replaced with W-CM. That's because x^W=x^(W-CM) for every Bitwidth 427 // bit number x, since either x is odd in which case x^CM = 1, or x is even in 428 // which case both x^W and x^(W - CM) are zero. By subtracting off multiples 429 // of CM like this weights can always be reduced to the range [0, CM+Bitwidth) 430 // which by a happy accident means that they can always be represented using 431 // Bitwidth bits. 432 // TODO: Reduce the weight by exploiting nsw/nuw? (Could do much better than 433 // the Carmichael number). 434 if (Bitwidth > 3) { 435 /// CM - The value of Carmichael's lambda function. 436 APInt CM = APInt::getOneBitSet(Bitwidth, CarmichaelShift(Bitwidth)); 437 // Any weight W >= Threshold can be replaced with W - CM. 438 APInt Threshold = CM + Bitwidth; 439 assert(LHS.ult(Threshold) && RHS.ult(Threshold) && "Weights not reduced!"); 440 // For Bitwidth 4 or more the following sum does not overflow. 441 LHS += RHS; 442 while (LHS.uge(Threshold)) 443 LHS -= CM; 444 } else { 445 // To avoid problems with overflow do everything the same as above but using 446 // a larger type. 447 unsigned CM = 1U << CarmichaelShift(Bitwidth); 448 unsigned Threshold = CM + Bitwidth; 449 assert(LHS.getZExtValue() < Threshold && RHS.getZExtValue() < Threshold && 450 "Weights not reduced!"); 451 unsigned Total = LHS.getZExtValue() + RHS.getZExtValue(); 452 while (Total >= Threshold) 453 Total -= CM; 454 LHS = Total; 455 } 456 } 457 458 typedef std::pair<Value*, APInt> RepeatedValue; 459 460 /// Given an associative binary expression, return the leaf 461 /// nodes in Ops along with their weights (how many times the leaf occurs). The 462 /// original expression is the same as 463 /// (Ops[0].first op Ops[0].first op ... Ops[0].first) <- Ops[0].second times 464 /// op 465 /// (Ops[1].first op Ops[1].first op ... Ops[1].first) <- Ops[1].second times 466 /// op 467 /// ... 468 /// op 469 /// (Ops[N].first op Ops[N].first op ... Ops[N].first) <- Ops[N].second times 470 /// 471 /// Note that the values Ops[0].first, ..., Ops[N].first are all distinct. 472 /// 473 /// This routine may modify the function, in which case it returns 'true'. The 474 /// changes it makes may well be destructive, changing the value computed by 'I' 475 /// to something completely different. Thus if the routine returns 'true' then 476 /// you MUST either replace I with a new expression computed from the Ops array, 477 /// or use RewriteExprTree to put the values back in. 478 /// 479 /// A leaf node is either not a binary operation of the same kind as the root 480 /// node 'I' (i.e. is not a binary operator at all, or is, but with a different 481 /// opcode), or is the same kind of binary operator but has a use which either 482 /// does not belong to the expression, or does belong to the expression but is 483 /// a leaf node. Every leaf node has at least one use that is a non-leaf node 484 /// of the expression, while for non-leaf nodes (except for the root 'I') every 485 /// use is a non-leaf node of the expression. 486 /// 487 /// For example: 488 /// expression graph node names 489 /// 490 /// + | I 491 /// / \ | 492 /// + + | A, B 493 /// / \ / \ | 494 /// * + * | C, D, E 495 /// / \ / \ / \ | 496 /// + * | F, G 497 /// 498 /// The leaf nodes are C, E, F and G. The Ops array will contain (maybe not in 499 /// that order) (C, 1), (E, 1), (F, 2), (G, 2). 500 /// 501 /// The expression is maximal: if some instruction is a binary operator of the 502 /// same kind as 'I', and all of its uses are non-leaf nodes of the expression, 503 /// then the instruction also belongs to the expression, is not a leaf node of 504 /// it, and its operands also belong to the expression (but may be leaf nodes). 505 /// 506 /// NOTE: This routine will set operands of non-leaf non-root nodes to undef in 507 /// order to ensure that every non-root node in the expression has *exactly one* 508 /// use by a non-leaf node of the expression. This destruction means that the 509 /// caller MUST either replace 'I' with a new expression or use something like 510 /// RewriteExprTree to put the values back in if the routine indicates that it 511 /// made a change by returning 'true'. 512 /// 513 /// In the above example either the right operand of A or the left operand of B 514 /// will be replaced by undef. If it is B's operand then this gives: 515 /// 516 /// + | I 517 /// / \ | 518 /// + + | A, B - operand of B replaced with undef 519 /// / \ \ | 520 /// * + * | C, D, E 521 /// / \ / \ / \ | 522 /// + * | F, G 523 /// 524 /// Note that such undef operands can only be reached by passing through 'I'. 525 /// For example, if you visit operands recursively starting from a leaf node 526 /// then you will never see such an undef operand unless you get back to 'I', 527 /// which requires passing through a phi node. 528 /// 529 /// Note that this routine may also mutate binary operators of the wrong type 530 /// that have all uses inside the expression (i.e. only used by non-leaf nodes 531 /// of the expression) if it can turn them into binary operators of the right 532 /// type and thus make the expression bigger. 533 534 static bool LinearizeExprTree(BinaryOperator *I, 535 SmallVectorImpl<RepeatedValue> &Ops) { 536 DEBUG(dbgs() << "LINEARIZE: " << *I << '\n'); 537 unsigned Bitwidth = I->getType()->getScalarType()->getPrimitiveSizeInBits(); 538 unsigned Opcode = I->getOpcode(); 539 assert(I->isAssociative() && I->isCommutative() && 540 "Expected an associative and commutative operation!"); 541 542 // Visit all operands of the expression, keeping track of their weight (the 543 // number of paths from the expression root to the operand, or if you like 544 // the number of times that operand occurs in the linearized expression). 545 // For example, if I = X + A, where X = A + B, then I, X and B have weight 1 546 // while A has weight two. 547 548 // Worklist of non-leaf nodes (their operands are in the expression too) along 549 // with their weights, representing a certain number of paths to the operator. 550 // If an operator occurs in the worklist multiple times then we found multiple 551 // ways to get to it. 552 SmallVector<std::pair<BinaryOperator*, APInt>, 8> Worklist; // (Op, Weight) 553 Worklist.push_back(std::make_pair(I, APInt(Bitwidth, 1))); 554 bool Changed = false; 555 556 // Leaves of the expression are values that either aren't the right kind of 557 // operation (eg: a constant, or a multiply in an add tree), or are, but have 558 // some uses that are not inside the expression. For example, in I = X + X, 559 // X = A + B, the value X has two uses (by I) that are in the expression. If 560 // X has any other uses, for example in a return instruction, then we consider 561 // X to be a leaf, and won't analyze it further. When we first visit a value, 562 // if it has more than one use then at first we conservatively consider it to 563 // be a leaf. Later, as the expression is explored, we may discover some more 564 // uses of the value from inside the expression. If all uses turn out to be 565 // from within the expression (and the value is a binary operator of the right 566 // kind) then the value is no longer considered to be a leaf, and its operands 567 // are explored. 568 569 // Leaves - Keeps track of the set of putative leaves as well as the number of 570 // paths to each leaf seen so far. 571 typedef DenseMap<Value*, APInt> LeafMap; 572 LeafMap Leaves; // Leaf -> Total weight so far. 573 SmallVector<Value*, 8> LeafOrder; // Ensure deterministic leaf output order. 574 575 #ifndef NDEBUG 576 SmallPtrSet<Value*, 8> Visited; // For sanity checking the iteration scheme. 577 #endif 578 while (!Worklist.empty()) { 579 std::pair<BinaryOperator*, APInt> P = Worklist.pop_back_val(); 580 I = P.first; // We examine the operands of this binary operator. 581 582 for (unsigned OpIdx = 0; OpIdx < 2; ++OpIdx) { // Visit operands. 583 Value *Op = I->getOperand(OpIdx); 584 APInt Weight = P.second; // Number of paths to this operand. 585 DEBUG(dbgs() << "OPERAND: " << *Op << " (" << Weight << ")\n"); 586 assert(!Op->use_empty() && "No uses, so how did we get to it?!"); 587 588 // If this is a binary operation of the right kind with only one use then 589 // add its operands to the expression. 590 if (BinaryOperator *BO = isReassociableOp(Op, Opcode)) { 591 assert(Visited.insert(Op).second && "Not first visit!"); 592 DEBUG(dbgs() << "DIRECT ADD: " << *Op << " (" << Weight << ")\n"); 593 Worklist.push_back(std::make_pair(BO, Weight)); 594 continue; 595 } 596 597 // Appears to be a leaf. Is the operand already in the set of leaves? 598 LeafMap::iterator It = Leaves.find(Op); 599 if (It == Leaves.end()) { 600 // Not in the leaf map. Must be the first time we saw this operand. 601 assert(Visited.insert(Op).second && "Not first visit!"); 602 if (!Op->hasOneUse()) { 603 // This value has uses not accounted for by the expression, so it is 604 // not safe to modify. Mark it as being a leaf. 605 DEBUG(dbgs() << "ADD USES LEAF: " << *Op << " (" << Weight << ")\n"); 606 LeafOrder.push_back(Op); 607 Leaves[Op] = Weight; 608 continue; 609 } 610 // No uses outside the expression, try morphing it. 611 } else if (It != Leaves.end()) { 612 // Already in the leaf map. 613 assert(Visited.count(Op) && "In leaf map but not visited!"); 614 615 // Update the number of paths to the leaf. 616 IncorporateWeight(It->second, Weight, Opcode); 617 618 #if 0 // TODO: Re-enable once PR13021 is fixed. 619 // The leaf already has one use from inside the expression. As we want 620 // exactly one such use, drop this new use of the leaf. 621 assert(!Op->hasOneUse() && "Only one use, but we got here twice!"); 622 I->setOperand(OpIdx, UndefValue::get(I->getType())); 623 Changed = true; 624 625 // If the leaf is a binary operation of the right kind and we now see 626 // that its multiple original uses were in fact all by nodes belonging 627 // to the expression, then no longer consider it to be a leaf and add 628 // its operands to the expression. 629 if (BinaryOperator *BO = isReassociableOp(Op, Opcode)) { 630 DEBUG(dbgs() << "UNLEAF: " << *Op << " (" << It->second << ")\n"); 631 Worklist.push_back(std::make_pair(BO, It->second)); 632 Leaves.erase(It); 633 continue; 634 } 635 #endif 636 637 // If we still have uses that are not accounted for by the expression 638 // then it is not safe to modify the value. 639 if (!Op->hasOneUse()) 640 continue; 641 642 // No uses outside the expression, try morphing it. 643 Weight = It->second; 644 Leaves.erase(It); // Since the value may be morphed below. 645 } 646 647 // At this point we have a value which, first of all, is not a binary 648 // expression of the right kind, and secondly, is only used inside the 649 // expression. This means that it can safely be modified. See if we 650 // can usefully morph it into an expression of the right kind. 651 assert((!isa<Instruction>(Op) || 652 cast<Instruction>(Op)->getOpcode() != Opcode 653 || (isa<FPMathOperator>(Op) && 654 !cast<Instruction>(Op)->hasUnsafeAlgebra())) && 655 "Should have been handled above!"); 656 assert(Op->hasOneUse() && "Has uses outside the expression tree!"); 657 658 // If this is a multiply expression, turn any internal negations into 659 // multiplies by -1 so they can be reassociated. 660 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op)) 661 if ((Opcode == Instruction::Mul && BinaryOperator::isNeg(BO)) || 662 (Opcode == Instruction::FMul && BinaryOperator::isFNeg(BO))) { 663 DEBUG(dbgs() << "MORPH LEAF: " << *Op << " (" << Weight << ") TO "); 664 BO = LowerNegateToMultiply(BO); 665 DEBUG(dbgs() << *BO << '\n'); 666 Worklist.push_back(std::make_pair(BO, Weight)); 667 Changed = true; 668 continue; 669 } 670 671 // Failed to morph into an expression of the right type. This really is 672 // a leaf. 673 DEBUG(dbgs() << "ADD LEAF: " << *Op << " (" << Weight << ")\n"); 674 assert(!isReassociableOp(Op, Opcode) && "Value was morphed?"); 675 LeafOrder.push_back(Op); 676 Leaves[Op] = Weight; 677 } 678 } 679 680 // The leaves, repeated according to their weights, represent the linearized 681 // form of the expression. 682 for (unsigned i = 0, e = LeafOrder.size(); i != e; ++i) { 683 Value *V = LeafOrder[i]; 684 LeafMap::iterator It = Leaves.find(V); 685 if (It == Leaves.end()) 686 // Node initially thought to be a leaf wasn't. 687 continue; 688 assert(!isReassociableOp(V, Opcode) && "Shouldn't be a leaf!"); 689 APInt Weight = It->second; 690 if (Weight.isMinValue()) 691 // Leaf already output or weight reduction eliminated it. 692 continue; 693 // Ensure the leaf is only output once. 694 It->second = 0; 695 Ops.push_back(std::make_pair(V, Weight)); 696 } 697 698 // For nilpotent operations or addition there may be no operands, for example 699 // because the expression was "X xor X" or consisted of 2^Bitwidth additions: 700 // in both cases the weight reduces to 0 causing the value to be skipped. 701 if (Ops.empty()) { 702 Constant *Identity = ConstantExpr::getBinOpIdentity(Opcode, I->getType()); 703 assert(Identity && "Associative operation without identity!"); 704 Ops.emplace_back(Identity, APInt(Bitwidth, 1)); 705 } 706 707 return Changed; 708 } 709 710 /// Now that the operands for this expression tree are 711 /// linearized and optimized, emit them in-order. 712 void Reassociate::RewriteExprTree(BinaryOperator *I, 713 SmallVectorImpl<ValueEntry> &Ops) { 714 assert(Ops.size() > 1 && "Single values should be used directly!"); 715 716 // Since our optimizations should never increase the number of operations, the 717 // new expression can usually be written reusing the existing binary operators 718 // from the original expression tree, without creating any new instructions, 719 // though the rewritten expression may have a completely different topology. 720 // We take care to not change anything if the new expression will be the same 721 // as the original. If more than trivial changes (like commuting operands) 722 // were made then we are obliged to clear out any optional subclass data like 723 // nsw flags. 724 725 /// NodesToRewrite - Nodes from the original expression available for writing 726 /// the new expression into. 727 SmallVector<BinaryOperator*, 8> NodesToRewrite; 728 unsigned Opcode = I->getOpcode(); 729 BinaryOperator *Op = I; 730 731 /// NotRewritable - The operands being written will be the leaves of the new 732 /// expression and must not be used as inner nodes (via NodesToRewrite) by 733 /// mistake. Inner nodes are always reassociable, and usually leaves are not 734 /// (if they were they would have been incorporated into the expression and so 735 /// would not be leaves), so most of the time there is no danger of this. But 736 /// in rare cases a leaf may become reassociable if an optimization kills uses 737 /// of it, or it may momentarily become reassociable during rewriting (below) 738 /// due it being removed as an operand of one of its uses. Ensure that misuse 739 /// of leaf nodes as inner nodes cannot occur by remembering all of the future 740 /// leaves and refusing to reuse any of them as inner nodes. 741 SmallPtrSet<Value*, 8> NotRewritable; 742 for (unsigned i = 0, e = Ops.size(); i != e; ++i) 743 NotRewritable.insert(Ops[i].Op); 744 745 // ExpressionChanged - Non-null if the rewritten expression differs from the 746 // original in some non-trivial way, requiring the clearing of optional flags. 747 // Flags are cleared from the operator in ExpressionChanged up to I inclusive. 748 BinaryOperator *ExpressionChanged = nullptr; 749 for (unsigned i = 0; ; ++i) { 750 // The last operation (which comes earliest in the IR) is special as both 751 // operands will come from Ops, rather than just one with the other being 752 // a subexpression. 753 if (i+2 == Ops.size()) { 754 Value *NewLHS = Ops[i].Op; 755 Value *NewRHS = Ops[i+1].Op; 756 Value *OldLHS = Op->getOperand(0); 757 Value *OldRHS = Op->getOperand(1); 758 759 if (NewLHS == OldLHS && NewRHS == OldRHS) 760 // Nothing changed, leave it alone. 761 break; 762 763 if (NewLHS == OldRHS && NewRHS == OldLHS) { 764 // The order of the operands was reversed. Swap them. 765 DEBUG(dbgs() << "RA: " << *Op << '\n'); 766 Op->swapOperands(); 767 DEBUG(dbgs() << "TO: " << *Op << '\n'); 768 MadeChange = true; 769 ++NumChanged; 770 break; 771 } 772 773 // The new operation differs non-trivially from the original. Overwrite 774 // the old operands with the new ones. 775 DEBUG(dbgs() << "RA: " << *Op << '\n'); 776 if (NewLHS != OldLHS) { 777 BinaryOperator *BO = isReassociableOp(OldLHS, Opcode); 778 if (BO && !NotRewritable.count(BO)) 779 NodesToRewrite.push_back(BO); 780 Op->setOperand(0, NewLHS); 781 } 782 if (NewRHS != OldRHS) { 783 BinaryOperator *BO = isReassociableOp(OldRHS, Opcode); 784 if (BO && !NotRewritable.count(BO)) 785 NodesToRewrite.push_back(BO); 786 Op->setOperand(1, NewRHS); 787 } 788 DEBUG(dbgs() << "TO: " << *Op << '\n'); 789 790 ExpressionChanged = Op; 791 MadeChange = true; 792 ++NumChanged; 793 794 break; 795 } 796 797 // Not the last operation. The left-hand side will be a sub-expression 798 // while the right-hand side will be the current element of Ops. 799 Value *NewRHS = Ops[i].Op; 800 if (NewRHS != Op->getOperand(1)) { 801 DEBUG(dbgs() << "RA: " << *Op << '\n'); 802 if (NewRHS == Op->getOperand(0)) { 803 // The new right-hand side was already present as the left operand. If 804 // we are lucky then swapping the operands will sort out both of them. 805 Op->swapOperands(); 806 } else { 807 // Overwrite with the new right-hand side. 808 BinaryOperator *BO = isReassociableOp(Op->getOperand(1), Opcode); 809 if (BO && !NotRewritable.count(BO)) 810 NodesToRewrite.push_back(BO); 811 Op->setOperand(1, NewRHS); 812 ExpressionChanged = Op; 813 } 814 DEBUG(dbgs() << "TO: " << *Op << '\n'); 815 MadeChange = true; 816 ++NumChanged; 817 } 818 819 // Now deal with the left-hand side. If this is already an operation node 820 // from the original expression then just rewrite the rest of the expression 821 // into it. 822 BinaryOperator *BO = isReassociableOp(Op->getOperand(0), Opcode); 823 if (BO && !NotRewritable.count(BO)) { 824 Op = BO; 825 continue; 826 } 827 828 // Otherwise, grab a spare node from the original expression and use that as 829 // the left-hand side. If there are no nodes left then the optimizers made 830 // an expression with more nodes than the original! This usually means that 831 // they did something stupid but it might mean that the problem was just too 832 // hard (finding the mimimal number of multiplications needed to realize a 833 // multiplication expression is NP-complete). Whatever the reason, smart or 834 // stupid, create a new node if there are none left. 835 BinaryOperator *NewOp; 836 if (NodesToRewrite.empty()) { 837 Constant *Undef = UndefValue::get(I->getType()); 838 NewOp = BinaryOperator::Create(Instruction::BinaryOps(Opcode), 839 Undef, Undef, "", I); 840 if (NewOp->getType()->isFPOrFPVectorTy()) 841 NewOp->setFastMathFlags(I->getFastMathFlags()); 842 } else { 843 NewOp = NodesToRewrite.pop_back_val(); 844 } 845 846 DEBUG(dbgs() << "RA: " << *Op << '\n'); 847 Op->setOperand(0, NewOp); 848 DEBUG(dbgs() << "TO: " << *Op << '\n'); 849 ExpressionChanged = Op; 850 MadeChange = true; 851 ++NumChanged; 852 Op = NewOp; 853 } 854 855 // If the expression changed non-trivially then clear out all subclass data 856 // starting from the operator specified in ExpressionChanged, and compactify 857 // the operators to just before the expression root to guarantee that the 858 // expression tree is dominated by all of Ops. 859 if (ExpressionChanged) 860 do { 861 // Preserve FastMathFlags. 862 if (isa<FPMathOperator>(I)) { 863 FastMathFlags Flags = I->getFastMathFlags(); 864 ExpressionChanged->clearSubclassOptionalData(); 865 ExpressionChanged->setFastMathFlags(Flags); 866 } else 867 ExpressionChanged->clearSubclassOptionalData(); 868 869 if (ExpressionChanged == I) 870 break; 871 ExpressionChanged->moveBefore(I); 872 ExpressionChanged = cast<BinaryOperator>(*ExpressionChanged->user_begin()); 873 } while (1); 874 875 // Throw away any left over nodes from the original expression. 876 for (unsigned i = 0, e = NodesToRewrite.size(); i != e; ++i) 877 RedoInsts.insert(NodesToRewrite[i]); 878 } 879 880 /// Insert instructions before the instruction pointed to by BI, 881 /// that computes the negative version of the value specified. The negative 882 /// version of the value is returned, and BI is left pointing at the instruction 883 /// that should be processed next by the reassociation pass. 884 /// Also add intermediate instructions to the redo list that are modified while 885 /// pushing the negates through adds. These will be revisited to see if 886 /// additional opportunities have been exposed. 887 static Value *NegateValue(Value *V, Instruction *BI, 888 SetVector<AssertingVH<Instruction>> &ToRedo) { 889 if (Constant *C = dyn_cast<Constant>(V)) { 890 if (C->getType()->isFPOrFPVectorTy()) { 891 return ConstantExpr::getFNeg(C); 892 } 893 return ConstantExpr::getNeg(C); 894 } 895 896 897 // We are trying to expose opportunity for reassociation. One of the things 898 // that we want to do to achieve this is to push a negation as deep into an 899 // expression chain as possible, to expose the add instructions. In practice, 900 // this means that we turn this: 901 // X = -(A+12+C+D) into X = -A + -12 + -C + -D = -12 + -A + -C + -D 902 // so that later, a: Y = 12+X could get reassociated with the -12 to eliminate 903 // the constants. We assume that instcombine will clean up the mess later if 904 // we introduce tons of unnecessary negation instructions. 905 // 906 if (BinaryOperator *I = 907 isReassociableOp(V, Instruction::Add, Instruction::FAdd)) { 908 // Push the negates through the add. 909 I->setOperand(0, NegateValue(I->getOperand(0), BI, ToRedo)); 910 I->setOperand(1, NegateValue(I->getOperand(1), BI, ToRedo)); 911 if (I->getOpcode() == Instruction::Add) { 912 I->setHasNoUnsignedWrap(false); 913 I->setHasNoSignedWrap(false); 914 } 915 916 // We must move the add instruction here, because the neg instructions do 917 // not dominate the old add instruction in general. By moving it, we are 918 // assured that the neg instructions we just inserted dominate the 919 // instruction we are about to insert after them. 920 // 921 I->moveBefore(BI); 922 I->setName(I->getName()+".neg"); 923 924 // Add the intermediate negates to the redo list as processing them later 925 // could expose more reassociating opportunities. 926 ToRedo.insert(I); 927 return I; 928 } 929 930 // Okay, we need to materialize a negated version of V with an instruction. 931 // Scan the use lists of V to see if we have one already. 932 for (User *U : V->users()) { 933 if (!BinaryOperator::isNeg(U) && !BinaryOperator::isFNeg(U)) 934 continue; 935 936 // We found one! Now we have to make sure that the definition dominates 937 // this use. We do this by moving it to the entry block (if it is a 938 // non-instruction value) or right after the definition. These negates will 939 // be zapped by reassociate later, so we don't need much finesse here. 940 BinaryOperator *TheNeg = cast<BinaryOperator>(U); 941 942 // Verify that the negate is in this function, V might be a constant expr. 943 if (TheNeg->getParent()->getParent() != BI->getParent()->getParent()) 944 continue; 945 946 BasicBlock::iterator InsertPt; 947 if (Instruction *InstInput = dyn_cast<Instruction>(V)) { 948 if (InvokeInst *II = dyn_cast<InvokeInst>(InstInput)) { 949 InsertPt = II->getNormalDest()->begin(); 950 } else { 951 InsertPt = ++InstInput->getIterator(); 952 } 953 while (isa<PHINode>(InsertPt)) ++InsertPt; 954 } else { 955 InsertPt = TheNeg->getParent()->getParent()->getEntryBlock().begin(); 956 } 957 TheNeg->moveBefore(&*InsertPt); 958 if (TheNeg->getOpcode() == Instruction::Sub) { 959 TheNeg->setHasNoUnsignedWrap(false); 960 TheNeg->setHasNoSignedWrap(false); 961 } else { 962 TheNeg->andIRFlags(BI); 963 } 964 ToRedo.insert(TheNeg); 965 return TheNeg; 966 } 967 968 // Insert a 'neg' instruction that subtracts the value from zero to get the 969 // negation. 970 BinaryOperator *NewNeg = CreateNeg(V, V->getName() + ".neg", BI, BI); 971 ToRedo.insert(NewNeg); 972 return NewNeg; 973 } 974 975 /// Return true if we should break up this subtract of X-Y into (X + -Y). 976 static bool ShouldBreakUpSubtract(Instruction *Sub) { 977 // If this is a negation, we can't split it up! 978 if (BinaryOperator::isNeg(Sub) || BinaryOperator::isFNeg(Sub)) 979 return false; 980 981 // Don't breakup X - undef. 982 if (isa<UndefValue>(Sub->getOperand(1))) 983 return false; 984 985 // Don't bother to break this up unless either the LHS is an associable add or 986 // subtract or if this is only used by one. 987 Value *V0 = Sub->getOperand(0); 988 if (isReassociableOp(V0, Instruction::Add, Instruction::FAdd) || 989 isReassociableOp(V0, Instruction::Sub, Instruction::FSub)) 990 return true; 991 Value *V1 = Sub->getOperand(1); 992 if (isReassociableOp(V1, Instruction::Add, Instruction::FAdd) || 993 isReassociableOp(V1, Instruction::Sub, Instruction::FSub)) 994 return true; 995 Value *VB = Sub->user_back(); 996 if (Sub->hasOneUse() && 997 (isReassociableOp(VB, Instruction::Add, Instruction::FAdd) || 998 isReassociableOp(VB, Instruction::Sub, Instruction::FSub))) 999 return true; 1000 1001 return false; 1002 } 1003 1004 /// If we have (X-Y), and if either X is an add, or if this is only used by an 1005 /// add, transform this into (X+(0-Y)) to promote better reassociation. 1006 static BinaryOperator * 1007 BreakUpSubtract(Instruction *Sub, SetVector<AssertingVH<Instruction>> &ToRedo) { 1008 // Convert a subtract into an add and a neg instruction. This allows sub 1009 // instructions to be commuted with other add instructions. 1010 // 1011 // Calculate the negative value of Operand 1 of the sub instruction, 1012 // and set it as the RHS of the add instruction we just made. 1013 // 1014 Value *NegVal = NegateValue(Sub->getOperand(1), Sub, ToRedo); 1015 BinaryOperator *New = CreateAdd(Sub->getOperand(0), NegVal, "", Sub, Sub); 1016 Sub->setOperand(0, Constant::getNullValue(Sub->getType())); // Drop use of op. 1017 Sub->setOperand(1, Constant::getNullValue(Sub->getType())); // Drop use of op. 1018 New->takeName(Sub); 1019 1020 // Everyone now refers to the add instruction. 1021 Sub->replaceAllUsesWith(New); 1022 New->setDebugLoc(Sub->getDebugLoc()); 1023 1024 DEBUG(dbgs() << "Negated: " << *New << '\n'); 1025 return New; 1026 } 1027 1028 /// If this is a shift of a reassociable multiply or is used by one, change 1029 /// this into a multiply by a constant to assist with further reassociation. 1030 static BinaryOperator *ConvertShiftToMul(Instruction *Shl) { 1031 Constant *MulCst = ConstantInt::get(Shl->getType(), 1); 1032 MulCst = ConstantExpr::getShl(MulCst, cast<Constant>(Shl->getOperand(1))); 1033 1034 BinaryOperator *Mul = 1035 BinaryOperator::CreateMul(Shl->getOperand(0), MulCst, "", Shl); 1036 Shl->setOperand(0, UndefValue::get(Shl->getType())); // Drop use of op. 1037 Mul->takeName(Shl); 1038 1039 // Everyone now refers to the mul instruction. 1040 Shl->replaceAllUsesWith(Mul); 1041 Mul->setDebugLoc(Shl->getDebugLoc()); 1042 1043 // We can safely preserve the nuw flag in all cases. It's also safe to turn a 1044 // nuw nsw shl into a nuw nsw mul. However, nsw in isolation requires special 1045 // handling. 1046 bool NSW = cast<BinaryOperator>(Shl)->hasNoSignedWrap(); 1047 bool NUW = cast<BinaryOperator>(Shl)->hasNoUnsignedWrap(); 1048 if (NSW && NUW) 1049 Mul->setHasNoSignedWrap(true); 1050 Mul->setHasNoUnsignedWrap(NUW); 1051 return Mul; 1052 } 1053 1054 /// Scan backwards and forwards among values with the same rank as element i 1055 /// to see if X exists. If X does not exist, return i. This is useful when 1056 /// scanning for 'x' when we see '-x' because they both get the same rank. 1057 static unsigned FindInOperandList(SmallVectorImpl<ValueEntry> &Ops, unsigned i, 1058 Value *X) { 1059 unsigned XRank = Ops[i].Rank; 1060 unsigned e = Ops.size(); 1061 for (unsigned j = i+1; j != e && Ops[j].Rank == XRank; ++j) { 1062 if (Ops[j].Op == X) 1063 return j; 1064 if (Instruction *I1 = dyn_cast<Instruction>(Ops[j].Op)) 1065 if (Instruction *I2 = dyn_cast<Instruction>(X)) 1066 if (I1->isIdenticalTo(I2)) 1067 return j; 1068 } 1069 // Scan backwards. 1070 for (unsigned j = i-1; j != ~0U && Ops[j].Rank == XRank; --j) { 1071 if (Ops[j].Op == X) 1072 return j; 1073 if (Instruction *I1 = dyn_cast<Instruction>(Ops[j].Op)) 1074 if (Instruction *I2 = dyn_cast<Instruction>(X)) 1075 if (I1->isIdenticalTo(I2)) 1076 return j; 1077 } 1078 return i; 1079 } 1080 1081 /// Emit a tree of add instructions, summing Ops together 1082 /// and returning the result. Insert the tree before I. 1083 static Value *EmitAddTreeOfValues(Instruction *I, 1084 SmallVectorImpl<WeakVH> &Ops){ 1085 if (Ops.size() == 1) return Ops.back(); 1086 1087 Value *V1 = Ops.back(); 1088 Ops.pop_back(); 1089 Value *V2 = EmitAddTreeOfValues(I, Ops); 1090 return CreateAdd(V2, V1, "tmp", I, I); 1091 } 1092 1093 /// If V is an expression tree that is a multiplication sequence, 1094 /// and if this sequence contains a multiply by Factor, 1095 /// remove Factor from the tree and return the new tree. 1096 Value *Reassociate::RemoveFactorFromExpression(Value *V, Value *Factor) { 1097 BinaryOperator *BO = isReassociableOp(V, Instruction::Mul, Instruction::FMul); 1098 if (!BO) 1099 return nullptr; 1100 1101 SmallVector<RepeatedValue, 8> Tree; 1102 MadeChange |= LinearizeExprTree(BO, Tree); 1103 SmallVector<ValueEntry, 8> Factors; 1104 Factors.reserve(Tree.size()); 1105 for (unsigned i = 0, e = Tree.size(); i != e; ++i) { 1106 RepeatedValue E = Tree[i]; 1107 Factors.append(E.second.getZExtValue(), 1108 ValueEntry(getRank(E.first), E.first)); 1109 } 1110 1111 bool FoundFactor = false; 1112 bool NeedsNegate = false; 1113 for (unsigned i = 0, e = Factors.size(); i != e; ++i) { 1114 if (Factors[i].Op == Factor) { 1115 FoundFactor = true; 1116 Factors.erase(Factors.begin()+i); 1117 break; 1118 } 1119 1120 // If this is a negative version of this factor, remove it. 1121 if (ConstantInt *FC1 = dyn_cast<ConstantInt>(Factor)) { 1122 if (ConstantInt *FC2 = dyn_cast<ConstantInt>(Factors[i].Op)) 1123 if (FC1->getValue() == -FC2->getValue()) { 1124 FoundFactor = NeedsNegate = true; 1125 Factors.erase(Factors.begin()+i); 1126 break; 1127 } 1128 } else if (ConstantFP *FC1 = dyn_cast<ConstantFP>(Factor)) { 1129 if (ConstantFP *FC2 = dyn_cast<ConstantFP>(Factors[i].Op)) { 1130 APFloat F1(FC1->getValueAPF()); 1131 APFloat F2(FC2->getValueAPF()); 1132 F2.changeSign(); 1133 if (F1.compare(F2) == APFloat::cmpEqual) { 1134 FoundFactor = NeedsNegate = true; 1135 Factors.erase(Factors.begin() + i); 1136 break; 1137 } 1138 } 1139 } 1140 } 1141 1142 if (!FoundFactor) { 1143 // Make sure to restore the operands to the expression tree. 1144 RewriteExprTree(BO, Factors); 1145 return nullptr; 1146 } 1147 1148 BasicBlock::iterator InsertPt = ++BO->getIterator(); 1149 1150 // If this was just a single multiply, remove the multiply and return the only 1151 // remaining operand. 1152 if (Factors.size() == 1) { 1153 RedoInsts.insert(BO); 1154 V = Factors[0].Op; 1155 } else { 1156 RewriteExprTree(BO, Factors); 1157 V = BO; 1158 } 1159 1160 if (NeedsNegate) 1161 V = CreateNeg(V, "neg", &*InsertPt, BO); 1162 1163 return V; 1164 } 1165 1166 /// If V is a single-use multiply, recursively add its operands as factors, 1167 /// otherwise add V to the list of factors. 1168 /// 1169 /// Ops is the top-level list of add operands we're trying to factor. 1170 static void FindSingleUseMultiplyFactors(Value *V, 1171 SmallVectorImpl<Value*> &Factors, 1172 const SmallVectorImpl<ValueEntry> &Ops) { 1173 BinaryOperator *BO = isReassociableOp(V, Instruction::Mul, Instruction::FMul); 1174 if (!BO) { 1175 Factors.push_back(V); 1176 return; 1177 } 1178 1179 // Otherwise, add the LHS and RHS to the list of factors. 1180 FindSingleUseMultiplyFactors(BO->getOperand(1), Factors, Ops); 1181 FindSingleUseMultiplyFactors(BO->getOperand(0), Factors, Ops); 1182 } 1183 1184 /// Optimize a series of operands to an 'and', 'or', or 'xor' instruction. 1185 /// This optimizes based on identities. If it can be reduced to a single Value, 1186 /// it is returned, otherwise the Ops list is mutated as necessary. 1187 static Value *OptimizeAndOrXor(unsigned Opcode, 1188 SmallVectorImpl<ValueEntry> &Ops) { 1189 // Scan the operand lists looking for X and ~X pairs, along with X,X pairs. 1190 // If we find any, we can simplify the expression. X&~X == 0, X|~X == -1. 1191 for (unsigned i = 0, e = Ops.size(); i != e; ++i) { 1192 // First, check for X and ~X in the operand list. 1193 assert(i < Ops.size()); 1194 if (BinaryOperator::isNot(Ops[i].Op)) { // Cannot occur for ^. 1195 Value *X = BinaryOperator::getNotArgument(Ops[i].Op); 1196 unsigned FoundX = FindInOperandList(Ops, i, X); 1197 if (FoundX != i) { 1198 if (Opcode == Instruction::And) // ...&X&~X = 0 1199 return Constant::getNullValue(X->getType()); 1200 1201 if (Opcode == Instruction::Or) // ...|X|~X = -1 1202 return Constant::getAllOnesValue(X->getType()); 1203 } 1204 } 1205 1206 // Next, check for duplicate pairs of values, which we assume are next to 1207 // each other, due to our sorting criteria. 1208 assert(i < Ops.size()); 1209 if (i+1 != Ops.size() && Ops[i+1].Op == Ops[i].Op) { 1210 if (Opcode == Instruction::And || Opcode == Instruction::Or) { 1211 // Drop duplicate values for And and Or. 1212 Ops.erase(Ops.begin()+i); 1213 --i; --e; 1214 ++NumAnnihil; 1215 continue; 1216 } 1217 1218 // Drop pairs of values for Xor. 1219 assert(Opcode == Instruction::Xor); 1220 if (e == 2) 1221 return Constant::getNullValue(Ops[0].Op->getType()); 1222 1223 // Y ^ X^X -> Y 1224 Ops.erase(Ops.begin()+i, Ops.begin()+i+2); 1225 i -= 1; e -= 2; 1226 ++NumAnnihil; 1227 } 1228 } 1229 return nullptr; 1230 } 1231 1232 /// Helper function of CombineXorOpnd(). It creates a bitwise-and 1233 /// instruction with the given two operands, and return the resulting 1234 /// instruction. There are two special cases: 1) if the constant operand is 0, 1235 /// it will return NULL. 2) if the constant is ~0, the symbolic operand will 1236 /// be returned. 1237 static Value *createAndInstr(Instruction *InsertBefore, Value *Opnd, 1238 const APInt &ConstOpnd) { 1239 if (ConstOpnd != 0) { 1240 if (!ConstOpnd.isAllOnesValue()) { 1241 LLVMContext &Ctx = Opnd->getType()->getContext(); 1242 Instruction *I; 1243 I = BinaryOperator::CreateAnd(Opnd, ConstantInt::get(Ctx, ConstOpnd), 1244 "and.ra", InsertBefore); 1245 I->setDebugLoc(InsertBefore->getDebugLoc()); 1246 return I; 1247 } 1248 return Opnd; 1249 } 1250 return nullptr; 1251 } 1252 1253 // Helper function of OptimizeXor(). It tries to simplify "Opnd1 ^ ConstOpnd" 1254 // into "R ^ C", where C would be 0, and R is a symbolic value. 1255 // 1256 // If it was successful, true is returned, and the "R" and "C" is returned 1257 // via "Res" and "ConstOpnd", respectively; otherwise, false is returned, 1258 // and both "Res" and "ConstOpnd" remain unchanged. 1259 // 1260 bool Reassociate::CombineXorOpnd(Instruction *I, XorOpnd *Opnd1, 1261 APInt &ConstOpnd, Value *&Res) { 1262 // Xor-Rule 1: (x | c1) ^ c2 = (x | c1) ^ (c1 ^ c1) ^ c2 1263 // = ((x | c1) ^ c1) ^ (c1 ^ c2) 1264 // = (x & ~c1) ^ (c1 ^ c2) 1265 // It is useful only when c1 == c2. 1266 if (Opnd1->isOrExpr() && Opnd1->getConstPart() != 0) { 1267 if (!Opnd1->getValue()->hasOneUse()) 1268 return false; 1269 1270 const APInt &C1 = Opnd1->getConstPart(); 1271 if (C1 != ConstOpnd) 1272 return false; 1273 1274 Value *X = Opnd1->getSymbolicPart(); 1275 Res = createAndInstr(I, X, ~C1); 1276 // ConstOpnd was C2, now C1 ^ C2. 1277 ConstOpnd ^= C1; 1278 1279 if (Instruction *T = dyn_cast<Instruction>(Opnd1->getValue())) 1280 RedoInsts.insert(T); 1281 return true; 1282 } 1283 return false; 1284 } 1285 1286 1287 // Helper function of OptimizeXor(). It tries to simplify 1288 // "Opnd1 ^ Opnd2 ^ ConstOpnd" into "R ^ C", where C would be 0, and R is a 1289 // symbolic value. 1290 // 1291 // If it was successful, true is returned, and the "R" and "C" is returned 1292 // via "Res" and "ConstOpnd", respectively (If the entire expression is 1293 // evaluated to a constant, the Res is set to NULL); otherwise, false is 1294 // returned, and both "Res" and "ConstOpnd" remain unchanged. 1295 bool Reassociate::CombineXorOpnd(Instruction *I, XorOpnd *Opnd1, XorOpnd *Opnd2, 1296 APInt &ConstOpnd, Value *&Res) { 1297 Value *X = Opnd1->getSymbolicPart(); 1298 if (X != Opnd2->getSymbolicPart()) 1299 return false; 1300 1301 // This many instruction become dead.(At least "Opnd1 ^ Opnd2" will die.) 1302 int DeadInstNum = 1; 1303 if (Opnd1->getValue()->hasOneUse()) 1304 DeadInstNum++; 1305 if (Opnd2->getValue()->hasOneUse()) 1306 DeadInstNum++; 1307 1308 // Xor-Rule 2: 1309 // (x | c1) ^ (x & c2) 1310 // = (x|c1) ^ (x&c2) ^ (c1 ^ c1) = ((x|c1) ^ c1) ^ (x & c2) ^ c1 1311 // = (x & ~c1) ^ (x & c2) ^ c1 // Xor-Rule 1 1312 // = (x & c3) ^ c1, where c3 = ~c1 ^ c2 // Xor-rule 3 1313 // 1314 if (Opnd1->isOrExpr() != Opnd2->isOrExpr()) { 1315 if (Opnd2->isOrExpr()) 1316 std::swap(Opnd1, Opnd2); 1317 1318 const APInt &C1 = Opnd1->getConstPart(); 1319 const APInt &C2 = Opnd2->getConstPart(); 1320 APInt C3((~C1) ^ C2); 1321 1322 // Do not increase code size! 1323 if (C3 != 0 && !C3.isAllOnesValue()) { 1324 int NewInstNum = ConstOpnd != 0 ? 1 : 2; 1325 if (NewInstNum > DeadInstNum) 1326 return false; 1327 } 1328 1329 Res = createAndInstr(I, X, C3); 1330 ConstOpnd ^= C1; 1331 1332 } else if (Opnd1->isOrExpr()) { 1333 // Xor-Rule 3: (x | c1) ^ (x | c2) = (x & c3) ^ c3 where c3 = c1 ^ c2 1334 // 1335 const APInt &C1 = Opnd1->getConstPart(); 1336 const APInt &C2 = Opnd2->getConstPart(); 1337 APInt C3 = C1 ^ C2; 1338 1339 // Do not increase code size 1340 if (C3 != 0 && !C3.isAllOnesValue()) { 1341 int NewInstNum = ConstOpnd != 0 ? 1 : 2; 1342 if (NewInstNum > DeadInstNum) 1343 return false; 1344 } 1345 1346 Res = createAndInstr(I, X, C3); 1347 ConstOpnd ^= C3; 1348 } else { 1349 // Xor-Rule 4: (x & c1) ^ (x & c2) = (x & (c1^c2)) 1350 // 1351 const APInt &C1 = Opnd1->getConstPart(); 1352 const APInt &C2 = Opnd2->getConstPart(); 1353 APInt C3 = C1 ^ C2; 1354 Res = createAndInstr(I, X, C3); 1355 } 1356 1357 // Put the original operands in the Redo list; hope they will be deleted 1358 // as dead code. 1359 if (Instruction *T = dyn_cast<Instruction>(Opnd1->getValue())) 1360 RedoInsts.insert(T); 1361 if (Instruction *T = dyn_cast<Instruction>(Opnd2->getValue())) 1362 RedoInsts.insert(T); 1363 1364 return true; 1365 } 1366 1367 /// Optimize a series of operands to an 'xor' instruction. If it can be reduced 1368 /// to a single Value, it is returned, otherwise the Ops list is mutated as 1369 /// necessary. 1370 Value *Reassociate::OptimizeXor(Instruction *I, 1371 SmallVectorImpl<ValueEntry> &Ops) { 1372 if (Value *V = OptimizeAndOrXor(Instruction::Xor, Ops)) 1373 return V; 1374 1375 if (Ops.size() == 1) 1376 return nullptr; 1377 1378 SmallVector<XorOpnd, 8> Opnds; 1379 SmallVector<XorOpnd*, 8> OpndPtrs; 1380 Type *Ty = Ops[0].Op->getType(); 1381 APInt ConstOpnd(Ty->getIntegerBitWidth(), 0); 1382 1383 // Step 1: Convert ValueEntry to XorOpnd 1384 for (unsigned i = 0, e = Ops.size(); i != e; ++i) { 1385 Value *V = Ops[i].Op; 1386 if (!isa<ConstantInt>(V)) { 1387 XorOpnd O(V); 1388 O.setSymbolicRank(getRank(O.getSymbolicPart())); 1389 Opnds.push_back(O); 1390 } else 1391 ConstOpnd ^= cast<ConstantInt>(V)->getValue(); 1392 } 1393 1394 // NOTE: From this point on, do *NOT* add/delete element to/from "Opnds". 1395 // It would otherwise invalidate the "Opnds"'s iterator, and hence invalidate 1396 // the "OpndPtrs" as well. For the similar reason, do not fuse this loop 1397 // with the previous loop --- the iterator of the "Opnds" may be invalidated 1398 // when new elements are added to the vector. 1399 for (unsigned i = 0, e = Opnds.size(); i != e; ++i) 1400 OpndPtrs.push_back(&Opnds[i]); 1401 1402 // Step 2: Sort the Xor-Operands in a way such that the operands containing 1403 // the same symbolic value cluster together. For instance, the input operand 1404 // sequence ("x | 123", "y & 456", "x & 789") will be sorted into: 1405 // ("x | 123", "x & 789", "y & 456"). 1406 std::stable_sort(OpndPtrs.begin(), OpndPtrs.end(), XorOpnd::PtrSortFunctor()); 1407 1408 // Step 3: Combine adjacent operands 1409 XorOpnd *PrevOpnd = nullptr; 1410 bool Changed = false; 1411 for (unsigned i = 0, e = Opnds.size(); i < e; i++) { 1412 XorOpnd *CurrOpnd = OpndPtrs[i]; 1413 // The combined value 1414 Value *CV; 1415 1416 // Step 3.1: Try simplifying "CurrOpnd ^ ConstOpnd" 1417 if (ConstOpnd != 0 && CombineXorOpnd(I, CurrOpnd, ConstOpnd, CV)) { 1418 Changed = true; 1419 if (CV) 1420 *CurrOpnd = XorOpnd(CV); 1421 else { 1422 CurrOpnd->Invalidate(); 1423 continue; 1424 } 1425 } 1426 1427 if (!PrevOpnd || CurrOpnd->getSymbolicPart() != PrevOpnd->getSymbolicPart()) { 1428 PrevOpnd = CurrOpnd; 1429 continue; 1430 } 1431 1432 // step 3.2: When previous and current operands share the same symbolic 1433 // value, try to simplify "PrevOpnd ^ CurrOpnd ^ ConstOpnd" 1434 // 1435 if (CombineXorOpnd(I, CurrOpnd, PrevOpnd, ConstOpnd, CV)) { 1436 // Remove previous operand 1437 PrevOpnd->Invalidate(); 1438 if (CV) { 1439 *CurrOpnd = XorOpnd(CV); 1440 PrevOpnd = CurrOpnd; 1441 } else { 1442 CurrOpnd->Invalidate(); 1443 PrevOpnd = nullptr; 1444 } 1445 Changed = true; 1446 } 1447 } 1448 1449 // Step 4: Reassemble the Ops 1450 if (Changed) { 1451 Ops.clear(); 1452 for (unsigned int i = 0, e = Opnds.size(); i < e; i++) { 1453 XorOpnd &O = Opnds[i]; 1454 if (O.isInvalid()) 1455 continue; 1456 ValueEntry VE(getRank(O.getValue()), O.getValue()); 1457 Ops.push_back(VE); 1458 } 1459 if (ConstOpnd != 0) { 1460 Value *C = ConstantInt::get(Ty->getContext(), ConstOpnd); 1461 ValueEntry VE(getRank(C), C); 1462 Ops.push_back(VE); 1463 } 1464 int Sz = Ops.size(); 1465 if (Sz == 1) 1466 return Ops.back().Op; 1467 else if (Sz == 0) { 1468 assert(ConstOpnd == 0); 1469 return ConstantInt::get(Ty->getContext(), ConstOpnd); 1470 } 1471 } 1472 1473 return nullptr; 1474 } 1475 1476 /// Optimize a series of operands to an 'add' instruction. This 1477 /// optimizes based on identities. If it can be reduced to a single Value, it 1478 /// is returned, otherwise the Ops list is mutated as necessary. 1479 Value *Reassociate::OptimizeAdd(Instruction *I, 1480 SmallVectorImpl<ValueEntry> &Ops) { 1481 // Scan the operand lists looking for X and -X pairs. If we find any, we 1482 // can simplify expressions like X+-X == 0 and X+~X ==-1. While we're at it, 1483 // scan for any 1484 // duplicates. We want to canonicalize Y+Y+Y+Z -> 3*Y+Z. 1485 1486 for (unsigned i = 0, e = Ops.size(); i != e; ++i) { 1487 Value *TheOp = Ops[i].Op; 1488 // Check to see if we've seen this operand before. If so, we factor all 1489 // instances of the operand together. Due to our sorting criteria, we know 1490 // that these need to be next to each other in the vector. 1491 if (i+1 != Ops.size() && Ops[i+1].Op == TheOp) { 1492 // Rescan the list, remove all instances of this operand from the expr. 1493 unsigned NumFound = 0; 1494 do { 1495 Ops.erase(Ops.begin()+i); 1496 ++NumFound; 1497 } while (i != Ops.size() && Ops[i].Op == TheOp); 1498 1499 DEBUG(dbgs() << "\nFACTORING [" << NumFound << "]: " << *TheOp << '\n'); 1500 ++NumFactor; 1501 1502 // Insert a new multiply. 1503 Type *Ty = TheOp->getType(); 1504 Constant *C = Ty->isIntOrIntVectorTy() ? 1505 ConstantInt::get(Ty, NumFound) : ConstantFP::get(Ty, NumFound); 1506 Instruction *Mul = CreateMul(TheOp, C, "factor", I, I); 1507 1508 // Now that we have inserted a multiply, optimize it. This allows us to 1509 // handle cases that require multiple factoring steps, such as this: 1510 // (X*2) + (X*2) + (X*2) -> (X*2)*3 -> X*6 1511 RedoInsts.insert(Mul); 1512 1513 // If every add operand was a duplicate, return the multiply. 1514 if (Ops.empty()) 1515 return Mul; 1516 1517 // Otherwise, we had some input that didn't have the dupe, such as 1518 // "A + A + B" -> "A*2 + B". Add the new multiply to the list of 1519 // things being added by this operation. 1520 Ops.insert(Ops.begin(), ValueEntry(getRank(Mul), Mul)); 1521 1522 --i; 1523 e = Ops.size(); 1524 continue; 1525 } 1526 1527 // Check for X and -X or X and ~X in the operand list. 1528 if (!BinaryOperator::isNeg(TheOp) && !BinaryOperator::isFNeg(TheOp) && 1529 !BinaryOperator::isNot(TheOp)) 1530 continue; 1531 1532 Value *X = nullptr; 1533 if (BinaryOperator::isNeg(TheOp) || BinaryOperator::isFNeg(TheOp)) 1534 X = BinaryOperator::getNegArgument(TheOp); 1535 else if (BinaryOperator::isNot(TheOp)) 1536 X = BinaryOperator::getNotArgument(TheOp); 1537 1538 unsigned FoundX = FindInOperandList(Ops, i, X); 1539 if (FoundX == i) 1540 continue; 1541 1542 // Remove X and -X from the operand list. 1543 if (Ops.size() == 2 && 1544 (BinaryOperator::isNeg(TheOp) || BinaryOperator::isFNeg(TheOp))) 1545 return Constant::getNullValue(X->getType()); 1546 1547 // Remove X and ~X from the operand list. 1548 if (Ops.size() == 2 && BinaryOperator::isNot(TheOp)) 1549 return Constant::getAllOnesValue(X->getType()); 1550 1551 Ops.erase(Ops.begin()+i); 1552 if (i < FoundX) 1553 --FoundX; 1554 else 1555 --i; // Need to back up an extra one. 1556 Ops.erase(Ops.begin()+FoundX); 1557 ++NumAnnihil; 1558 --i; // Revisit element. 1559 e -= 2; // Removed two elements. 1560 1561 // if X and ~X we append -1 to the operand list. 1562 if (BinaryOperator::isNot(TheOp)) { 1563 Value *V = Constant::getAllOnesValue(X->getType()); 1564 Ops.insert(Ops.end(), ValueEntry(getRank(V), V)); 1565 e += 1; 1566 } 1567 } 1568 1569 // Scan the operand list, checking to see if there are any common factors 1570 // between operands. Consider something like A*A+A*B*C+D. We would like to 1571 // reassociate this to A*(A+B*C)+D, which reduces the number of multiplies. 1572 // To efficiently find this, we count the number of times a factor occurs 1573 // for any ADD operands that are MULs. 1574 DenseMap<Value*, unsigned> FactorOccurrences; 1575 1576 // Keep track of each multiply we see, to avoid triggering on (X*4)+(X*4) 1577 // where they are actually the same multiply. 1578 unsigned MaxOcc = 0; 1579 Value *MaxOccVal = nullptr; 1580 for (unsigned i = 0, e = Ops.size(); i != e; ++i) { 1581 BinaryOperator *BOp = 1582 isReassociableOp(Ops[i].Op, Instruction::Mul, Instruction::FMul); 1583 if (!BOp) 1584 continue; 1585 1586 // Compute all of the factors of this added value. 1587 SmallVector<Value*, 8> Factors; 1588 FindSingleUseMultiplyFactors(BOp, Factors, Ops); 1589 assert(Factors.size() > 1 && "Bad linearize!"); 1590 1591 // Add one to FactorOccurrences for each unique factor in this op. 1592 SmallPtrSet<Value*, 8> Duplicates; 1593 for (unsigned i = 0, e = Factors.size(); i != e; ++i) { 1594 Value *Factor = Factors[i]; 1595 if (!Duplicates.insert(Factor).second) 1596 continue; 1597 1598 unsigned Occ = ++FactorOccurrences[Factor]; 1599 if (Occ > MaxOcc) { 1600 MaxOcc = Occ; 1601 MaxOccVal = Factor; 1602 } 1603 1604 // If Factor is a negative constant, add the negated value as a factor 1605 // because we can percolate the negate out. Watch for minint, which 1606 // cannot be positivified. 1607 if (ConstantInt *CI = dyn_cast<ConstantInt>(Factor)) { 1608 if (CI->isNegative() && !CI->isMinValue(true)) { 1609 Factor = ConstantInt::get(CI->getContext(), -CI->getValue()); 1610 assert(!Duplicates.count(Factor) && 1611 "Shouldn't have two constant factors, missed a canonicalize"); 1612 unsigned Occ = ++FactorOccurrences[Factor]; 1613 if (Occ > MaxOcc) { 1614 MaxOcc = Occ; 1615 MaxOccVal = Factor; 1616 } 1617 } 1618 } else if (ConstantFP *CF = dyn_cast<ConstantFP>(Factor)) { 1619 if (CF->isNegative()) { 1620 APFloat F(CF->getValueAPF()); 1621 F.changeSign(); 1622 Factor = ConstantFP::get(CF->getContext(), F); 1623 assert(!Duplicates.count(Factor) && 1624 "Shouldn't have two constant factors, missed a canonicalize"); 1625 unsigned Occ = ++FactorOccurrences[Factor]; 1626 if (Occ > MaxOcc) { 1627 MaxOcc = Occ; 1628 MaxOccVal = Factor; 1629 } 1630 } 1631 } 1632 } 1633 } 1634 1635 // If any factor occurred more than one time, we can pull it out. 1636 if (MaxOcc > 1) { 1637 DEBUG(dbgs() << "\nFACTORING [" << MaxOcc << "]: " << *MaxOccVal << '\n'); 1638 ++NumFactor; 1639 1640 // Create a new instruction that uses the MaxOccVal twice. If we don't do 1641 // this, we could otherwise run into situations where removing a factor 1642 // from an expression will drop a use of maxocc, and this can cause 1643 // RemoveFactorFromExpression on successive values to behave differently. 1644 Instruction *DummyInst = 1645 I->getType()->isIntOrIntVectorTy() 1646 ? BinaryOperator::CreateAdd(MaxOccVal, MaxOccVal) 1647 : BinaryOperator::CreateFAdd(MaxOccVal, MaxOccVal); 1648 1649 SmallVector<WeakVH, 4> NewMulOps; 1650 for (unsigned i = 0; i != Ops.size(); ++i) { 1651 // Only try to remove factors from expressions we're allowed to. 1652 BinaryOperator *BOp = 1653 isReassociableOp(Ops[i].Op, Instruction::Mul, Instruction::FMul); 1654 if (!BOp) 1655 continue; 1656 1657 if (Value *V = RemoveFactorFromExpression(Ops[i].Op, MaxOccVal)) { 1658 // The factorized operand may occur several times. Convert them all in 1659 // one fell swoop. 1660 for (unsigned j = Ops.size(); j != i;) { 1661 --j; 1662 if (Ops[j].Op == Ops[i].Op) { 1663 NewMulOps.push_back(V); 1664 Ops.erase(Ops.begin()+j); 1665 } 1666 } 1667 --i; 1668 } 1669 } 1670 1671 // No need for extra uses anymore. 1672 delete DummyInst; 1673 1674 unsigned NumAddedValues = NewMulOps.size(); 1675 Value *V = EmitAddTreeOfValues(I, NewMulOps); 1676 1677 // Now that we have inserted the add tree, optimize it. This allows us to 1678 // handle cases that require multiple factoring steps, such as this: 1679 // A*A*B + A*A*C --> A*(A*B+A*C) --> A*(A*(B+C)) 1680 assert(NumAddedValues > 1 && "Each occurrence should contribute a value"); 1681 (void)NumAddedValues; 1682 if (Instruction *VI = dyn_cast<Instruction>(V)) 1683 RedoInsts.insert(VI); 1684 1685 // Create the multiply. 1686 Instruction *V2 = CreateMul(V, MaxOccVal, "tmp", I, I); 1687 1688 // Rerun associate on the multiply in case the inner expression turned into 1689 // a multiply. We want to make sure that we keep things in canonical form. 1690 RedoInsts.insert(V2); 1691 1692 // If every add operand included the factor (e.g. "A*B + A*C"), then the 1693 // entire result expression is just the multiply "A*(B+C)". 1694 if (Ops.empty()) 1695 return V2; 1696 1697 // Otherwise, we had some input that didn't have the factor, such as 1698 // "A*B + A*C + D" -> "A*(B+C) + D". Add the new multiply to the list of 1699 // things being added by this operation. 1700 Ops.insert(Ops.begin(), ValueEntry(getRank(V2), V2)); 1701 } 1702 1703 return nullptr; 1704 } 1705 1706 /// \brief Build up a vector of value/power pairs factoring a product. 1707 /// 1708 /// Given a series of multiplication operands, build a vector of factors and 1709 /// the powers each is raised to when forming the final product. Sort them in 1710 /// the order of descending power. 1711 /// 1712 /// (x*x) -> [(x, 2)] 1713 /// ((x*x)*x) -> [(x, 3)] 1714 /// ((((x*y)*x)*y)*x) -> [(x, 3), (y, 2)] 1715 /// 1716 /// \returns Whether any factors have a power greater than one. 1717 bool Reassociate::collectMultiplyFactors(SmallVectorImpl<ValueEntry> &Ops, 1718 SmallVectorImpl<Factor> &Factors) { 1719 // FIXME: Have Ops be (ValueEntry, Multiplicity) pairs, simplifying this. 1720 // Compute the sum of powers of simplifiable factors. 1721 unsigned FactorPowerSum = 0; 1722 for (unsigned Idx = 1, Size = Ops.size(); Idx < Size; ++Idx) { 1723 Value *Op = Ops[Idx-1].Op; 1724 1725 // Count the number of occurrences of this value. 1726 unsigned Count = 1; 1727 for (; Idx < Size && Ops[Idx].Op == Op; ++Idx) 1728 ++Count; 1729 // Track for simplification all factors which occur 2 or more times. 1730 if (Count > 1) 1731 FactorPowerSum += Count; 1732 } 1733 1734 // We can only simplify factors if the sum of the powers of our simplifiable 1735 // factors is 4 or higher. When that is the case, we will *always* have 1736 // a simplification. This is an important invariant to prevent cyclicly 1737 // trying to simplify already minimal formations. 1738 if (FactorPowerSum < 4) 1739 return false; 1740 1741 // Now gather the simplifiable factors, removing them from Ops. 1742 FactorPowerSum = 0; 1743 for (unsigned Idx = 1; Idx < Ops.size(); ++Idx) { 1744 Value *Op = Ops[Idx-1].Op; 1745 1746 // Count the number of occurrences of this value. 1747 unsigned Count = 1; 1748 for (; Idx < Ops.size() && Ops[Idx].Op == Op; ++Idx) 1749 ++Count; 1750 if (Count == 1) 1751 continue; 1752 // Move an even number of occurrences to Factors. 1753 Count &= ~1U; 1754 Idx -= Count; 1755 FactorPowerSum += Count; 1756 Factors.push_back(Factor(Op, Count)); 1757 Ops.erase(Ops.begin()+Idx, Ops.begin()+Idx+Count); 1758 } 1759 1760 // None of the adjustments above should have reduced the sum of factor powers 1761 // below our mininum of '4'. 1762 assert(FactorPowerSum >= 4); 1763 1764 std::stable_sort(Factors.begin(), Factors.end(), Factor::PowerDescendingSorter()); 1765 return true; 1766 } 1767 1768 /// \brief Build a tree of multiplies, computing the product of Ops. 1769 static Value *buildMultiplyTree(IRBuilder<> &Builder, 1770 SmallVectorImpl<Value*> &Ops) { 1771 if (Ops.size() == 1) 1772 return Ops.back(); 1773 1774 Value *LHS = Ops.pop_back_val(); 1775 do { 1776 if (LHS->getType()->isIntOrIntVectorTy()) 1777 LHS = Builder.CreateMul(LHS, Ops.pop_back_val()); 1778 else 1779 LHS = Builder.CreateFMul(LHS, Ops.pop_back_val()); 1780 } while (!Ops.empty()); 1781 1782 return LHS; 1783 } 1784 1785 /// \brief Build a minimal multiplication DAG for (a^x)*(b^y)*(c^z)*... 1786 /// 1787 /// Given a vector of values raised to various powers, where no two values are 1788 /// equal and the powers are sorted in decreasing order, compute the minimal 1789 /// DAG of multiplies to compute the final product, and return that product 1790 /// value. 1791 Value *Reassociate::buildMinimalMultiplyDAG(IRBuilder<> &Builder, 1792 SmallVectorImpl<Factor> &Factors) { 1793 assert(Factors[0].Power); 1794 SmallVector<Value *, 4> OuterProduct; 1795 for (unsigned LastIdx = 0, Idx = 1, Size = Factors.size(); 1796 Idx < Size && Factors[Idx].Power > 0; ++Idx) { 1797 if (Factors[Idx].Power != Factors[LastIdx].Power) { 1798 LastIdx = Idx; 1799 continue; 1800 } 1801 1802 // We want to multiply across all the factors with the same power so that 1803 // we can raise them to that power as a single entity. Build a mini tree 1804 // for that. 1805 SmallVector<Value *, 4> InnerProduct; 1806 InnerProduct.push_back(Factors[LastIdx].Base); 1807 do { 1808 InnerProduct.push_back(Factors[Idx].Base); 1809 ++Idx; 1810 } while (Idx < Size && Factors[Idx].Power == Factors[LastIdx].Power); 1811 1812 // Reset the base value of the first factor to the new expression tree. 1813 // We'll remove all the factors with the same power in a second pass. 1814 Value *M = Factors[LastIdx].Base = buildMultiplyTree(Builder, InnerProduct); 1815 if (Instruction *MI = dyn_cast<Instruction>(M)) 1816 RedoInsts.insert(MI); 1817 1818 LastIdx = Idx; 1819 } 1820 // Unique factors with equal powers -- we've folded them into the first one's 1821 // base. 1822 Factors.erase(std::unique(Factors.begin(), Factors.end(), 1823 Factor::PowerEqual()), 1824 Factors.end()); 1825 1826 // Iteratively collect the base of each factor with an add power into the 1827 // outer product, and halve each power in preparation for squaring the 1828 // expression. 1829 for (unsigned Idx = 0, Size = Factors.size(); Idx != Size; ++Idx) { 1830 if (Factors[Idx].Power & 1) 1831 OuterProduct.push_back(Factors[Idx].Base); 1832 Factors[Idx].Power >>= 1; 1833 } 1834 if (Factors[0].Power) { 1835 Value *SquareRoot = buildMinimalMultiplyDAG(Builder, Factors); 1836 OuterProduct.push_back(SquareRoot); 1837 OuterProduct.push_back(SquareRoot); 1838 } 1839 if (OuterProduct.size() == 1) 1840 return OuterProduct.front(); 1841 1842 Value *V = buildMultiplyTree(Builder, OuterProduct); 1843 return V; 1844 } 1845 1846 Value *Reassociate::OptimizeMul(BinaryOperator *I, 1847 SmallVectorImpl<ValueEntry> &Ops) { 1848 // We can only optimize the multiplies when there is a chain of more than 1849 // three, such that a balanced tree might require fewer total multiplies. 1850 if (Ops.size() < 4) 1851 return nullptr; 1852 1853 // Try to turn linear trees of multiplies without other uses of the 1854 // intermediate stages into minimal multiply DAGs with perfect sub-expression 1855 // re-use. 1856 SmallVector<Factor, 4> Factors; 1857 if (!collectMultiplyFactors(Ops, Factors)) 1858 return nullptr; // All distinct factors, so nothing left for us to do. 1859 1860 IRBuilder<> Builder(I); 1861 Value *V = buildMinimalMultiplyDAG(Builder, Factors); 1862 if (Ops.empty()) 1863 return V; 1864 1865 ValueEntry NewEntry = ValueEntry(getRank(V), V); 1866 Ops.insert(std::lower_bound(Ops.begin(), Ops.end(), NewEntry), NewEntry); 1867 return nullptr; 1868 } 1869 1870 Value *Reassociate::OptimizeExpression(BinaryOperator *I, 1871 SmallVectorImpl<ValueEntry> &Ops) { 1872 // Now that we have the linearized expression tree, try to optimize it. 1873 // Start by folding any constants that we found. 1874 Constant *Cst = nullptr; 1875 unsigned Opcode = I->getOpcode(); 1876 while (!Ops.empty() && isa<Constant>(Ops.back().Op)) { 1877 Constant *C = cast<Constant>(Ops.pop_back_val().Op); 1878 Cst = Cst ? ConstantExpr::get(Opcode, C, Cst) : C; 1879 } 1880 // If there was nothing but constants then we are done. 1881 if (Ops.empty()) 1882 return Cst; 1883 1884 // Put the combined constant back at the end of the operand list, except if 1885 // there is no point. For example, an add of 0 gets dropped here, while a 1886 // multiplication by zero turns the whole expression into zero. 1887 if (Cst && Cst != ConstantExpr::getBinOpIdentity(Opcode, I->getType())) { 1888 if (Cst == ConstantExpr::getBinOpAbsorber(Opcode, I->getType())) 1889 return Cst; 1890 Ops.push_back(ValueEntry(0, Cst)); 1891 } 1892 1893 if (Ops.size() == 1) return Ops[0].Op; 1894 1895 // Handle destructive annihilation due to identities between elements in the 1896 // argument list here. 1897 unsigned NumOps = Ops.size(); 1898 switch (Opcode) { 1899 default: break; 1900 case Instruction::And: 1901 case Instruction::Or: 1902 if (Value *Result = OptimizeAndOrXor(Opcode, Ops)) 1903 return Result; 1904 break; 1905 1906 case Instruction::Xor: 1907 if (Value *Result = OptimizeXor(I, Ops)) 1908 return Result; 1909 break; 1910 1911 case Instruction::Add: 1912 case Instruction::FAdd: 1913 if (Value *Result = OptimizeAdd(I, Ops)) 1914 return Result; 1915 break; 1916 1917 case Instruction::Mul: 1918 case Instruction::FMul: 1919 if (Value *Result = OptimizeMul(I, Ops)) 1920 return Result; 1921 break; 1922 } 1923 1924 if (Ops.size() != NumOps) 1925 return OptimizeExpression(I, Ops); 1926 return nullptr; 1927 } 1928 1929 /// Zap the given instruction, adding interesting operands to the work list. 1930 void Reassociate::EraseInst(Instruction *I) { 1931 assert(isInstructionTriviallyDead(I) && "Trivially dead instructions only!"); 1932 SmallVector<Value*, 8> Ops(I->op_begin(), I->op_end()); 1933 // Erase the dead instruction. 1934 ValueRankMap.erase(I); 1935 RedoInsts.remove(I); 1936 I->eraseFromParent(); 1937 // Optimize its operands. 1938 SmallPtrSet<Instruction *, 8> Visited; // Detect self-referential nodes. 1939 for (unsigned i = 0, e = Ops.size(); i != e; ++i) 1940 if (Instruction *Op = dyn_cast<Instruction>(Ops[i])) { 1941 // If this is a node in an expression tree, climb to the expression root 1942 // and add that since that's where optimization actually happens. 1943 unsigned Opcode = Op->getOpcode(); 1944 while (Op->hasOneUse() && Op->user_back()->getOpcode() == Opcode && 1945 Visited.insert(Op).second) 1946 Op = Op->user_back(); 1947 RedoInsts.insert(Op); 1948 } 1949 } 1950 1951 // Canonicalize expressions of the following form: 1952 // x + (-Constant * y) -> x - (Constant * y) 1953 // x - (-Constant * y) -> x + (Constant * y) 1954 Instruction *Reassociate::canonicalizeNegConstExpr(Instruction *I) { 1955 if (!I->hasOneUse() || I->getType()->isVectorTy()) 1956 return nullptr; 1957 1958 // Must be a fmul or fdiv instruction. 1959 unsigned Opcode = I->getOpcode(); 1960 if (Opcode != Instruction::FMul && Opcode != Instruction::FDiv) 1961 return nullptr; 1962 1963 auto *C0 = dyn_cast<ConstantFP>(I->getOperand(0)); 1964 auto *C1 = dyn_cast<ConstantFP>(I->getOperand(1)); 1965 1966 // Both operands are constant, let it get constant folded away. 1967 if (C0 && C1) 1968 return nullptr; 1969 1970 ConstantFP *CF = C0 ? C0 : C1; 1971 1972 // Must have one constant operand. 1973 if (!CF) 1974 return nullptr; 1975 1976 // Must be a negative ConstantFP. 1977 if (!CF->isNegative()) 1978 return nullptr; 1979 1980 // User must be a binary operator with one or more uses. 1981 Instruction *User = I->user_back(); 1982 if (!isa<BinaryOperator>(User) || !User->hasNUsesOrMore(1)) 1983 return nullptr; 1984 1985 unsigned UserOpcode = User->getOpcode(); 1986 if (UserOpcode != Instruction::FAdd && UserOpcode != Instruction::FSub) 1987 return nullptr; 1988 1989 // Subtraction is not commutative. Explicitly, the following transform is 1990 // not valid: (-Constant * y) - x -> x + (Constant * y) 1991 if (!User->isCommutative() && User->getOperand(1) != I) 1992 return nullptr; 1993 1994 // Change the sign of the constant. 1995 APFloat Val = CF->getValueAPF(); 1996 Val.changeSign(); 1997 I->setOperand(C0 ? 0 : 1, ConstantFP::get(CF->getContext(), Val)); 1998 1999 // Canonicalize I to RHS to simplify the next bit of logic. E.g., 2000 // ((-Const*y) + x) -> (x + (-Const*y)). 2001 if (User->getOperand(0) == I && User->isCommutative()) 2002 cast<BinaryOperator>(User)->swapOperands(); 2003 2004 Value *Op0 = User->getOperand(0); 2005 Value *Op1 = User->getOperand(1); 2006 BinaryOperator *NI; 2007 switch (UserOpcode) { 2008 default: 2009 llvm_unreachable("Unexpected Opcode!"); 2010 case Instruction::FAdd: 2011 NI = BinaryOperator::CreateFSub(Op0, Op1); 2012 NI->setFastMathFlags(cast<FPMathOperator>(User)->getFastMathFlags()); 2013 break; 2014 case Instruction::FSub: 2015 NI = BinaryOperator::CreateFAdd(Op0, Op1); 2016 NI->setFastMathFlags(cast<FPMathOperator>(User)->getFastMathFlags()); 2017 break; 2018 } 2019 2020 NI->insertBefore(User); 2021 NI->setName(User->getName()); 2022 User->replaceAllUsesWith(NI); 2023 NI->setDebugLoc(I->getDebugLoc()); 2024 RedoInsts.insert(I); 2025 MadeChange = true; 2026 return NI; 2027 } 2028 2029 /// Inspect and optimize the given instruction. Note that erasing 2030 /// instructions is not allowed. 2031 void Reassociate::OptimizeInst(Instruction *I) { 2032 // Only consider operations that we understand. 2033 if (!isa<BinaryOperator>(I)) 2034 return; 2035 2036 if (I->getOpcode() == Instruction::Shl && isa<ConstantInt>(I->getOperand(1))) 2037 // If an operand of this shift is a reassociable multiply, or if the shift 2038 // is used by a reassociable multiply or add, turn into a multiply. 2039 if (isReassociableOp(I->getOperand(0), Instruction::Mul) || 2040 (I->hasOneUse() && 2041 (isReassociableOp(I->user_back(), Instruction::Mul) || 2042 isReassociableOp(I->user_back(), Instruction::Add)))) { 2043 Instruction *NI = ConvertShiftToMul(I); 2044 RedoInsts.insert(I); 2045 MadeChange = true; 2046 I = NI; 2047 } 2048 2049 // Canonicalize negative constants out of expressions. 2050 if (Instruction *Res = canonicalizeNegConstExpr(I)) 2051 I = Res; 2052 2053 // Commute binary operators, to canonicalize the order of their operands. 2054 // This can potentially expose more CSE opportunities, and makes writing other 2055 // transformations simpler. 2056 if (I->isCommutative()) 2057 canonicalizeOperands(I); 2058 2059 // TODO: We should optimize vector Xor instructions, but they are 2060 // currently unsupported. 2061 if (I->getType()->isVectorTy() && I->getOpcode() == Instruction::Xor) 2062 return; 2063 2064 // Don't optimize floating point instructions that don't have unsafe algebra. 2065 if (I->getType()->isFPOrFPVectorTy() && !I->hasUnsafeAlgebra()) 2066 return; 2067 2068 // Do not reassociate boolean (i1) expressions. We want to preserve the 2069 // original order of evaluation for short-circuited comparisons that 2070 // SimplifyCFG has folded to AND/OR expressions. If the expression 2071 // is not further optimized, it is likely to be transformed back to a 2072 // short-circuited form for code gen, and the source order may have been 2073 // optimized for the most likely conditions. 2074 if (I->getType()->isIntegerTy(1)) 2075 return; 2076 2077 // If this is a subtract instruction which is not already in negate form, 2078 // see if we can convert it to X+-Y. 2079 if (I->getOpcode() == Instruction::Sub) { 2080 if (ShouldBreakUpSubtract(I)) { 2081 Instruction *NI = BreakUpSubtract(I, RedoInsts); 2082 RedoInsts.insert(I); 2083 MadeChange = true; 2084 I = NI; 2085 } else if (BinaryOperator::isNeg(I)) { 2086 // Otherwise, this is a negation. See if the operand is a multiply tree 2087 // and if this is not an inner node of a multiply tree. 2088 if (isReassociableOp(I->getOperand(1), Instruction::Mul) && 2089 (!I->hasOneUse() || 2090 !isReassociableOp(I->user_back(), Instruction::Mul))) { 2091 Instruction *NI = LowerNegateToMultiply(I); 2092 // If the negate was simplified, revisit the users to see if we can 2093 // reassociate further. 2094 for (User *U : NI->users()) { 2095 if (BinaryOperator *Tmp = dyn_cast<BinaryOperator>(U)) 2096 RedoInsts.insert(Tmp); 2097 } 2098 RedoInsts.insert(I); 2099 MadeChange = true; 2100 I = NI; 2101 } 2102 } 2103 } else if (I->getOpcode() == Instruction::FSub) { 2104 if (ShouldBreakUpSubtract(I)) { 2105 Instruction *NI = BreakUpSubtract(I, RedoInsts); 2106 RedoInsts.insert(I); 2107 MadeChange = true; 2108 I = NI; 2109 } else if (BinaryOperator::isFNeg(I)) { 2110 // Otherwise, this is a negation. See if the operand is a multiply tree 2111 // and if this is not an inner node of a multiply tree. 2112 if (isReassociableOp(I->getOperand(1), Instruction::FMul) && 2113 (!I->hasOneUse() || 2114 !isReassociableOp(I->user_back(), Instruction::FMul))) { 2115 // If the negate was simplified, revisit the users to see if we can 2116 // reassociate further. 2117 Instruction *NI = LowerNegateToMultiply(I); 2118 for (User *U : NI->users()) { 2119 if (BinaryOperator *Tmp = dyn_cast<BinaryOperator>(U)) 2120 RedoInsts.insert(Tmp); 2121 } 2122 RedoInsts.insert(I); 2123 MadeChange = true; 2124 I = NI; 2125 } 2126 } 2127 } 2128 2129 // If this instruction is an associative binary operator, process it. 2130 if (!I->isAssociative()) return; 2131 BinaryOperator *BO = cast<BinaryOperator>(I); 2132 2133 // If this is an interior node of a reassociable tree, ignore it until we 2134 // get to the root of the tree, to avoid N^2 analysis. 2135 unsigned Opcode = BO->getOpcode(); 2136 if (BO->hasOneUse() && BO->user_back()->getOpcode() == Opcode) { 2137 // During the initial run we will get to the root of the tree. 2138 // But if we get here while we are redoing instructions, there is no 2139 // guarantee that the root will be visited. So Redo later 2140 if (BO->user_back() != BO) 2141 RedoInsts.insert(BO->user_back()); 2142 return; 2143 } 2144 2145 // If this is an add tree that is used by a sub instruction, ignore it 2146 // until we process the subtract. 2147 if (BO->hasOneUse() && BO->getOpcode() == Instruction::Add && 2148 cast<Instruction>(BO->user_back())->getOpcode() == Instruction::Sub) 2149 return; 2150 if (BO->hasOneUse() && BO->getOpcode() == Instruction::FAdd && 2151 cast<Instruction>(BO->user_back())->getOpcode() == Instruction::FSub) 2152 return; 2153 2154 ReassociateExpression(BO); 2155 } 2156 2157 void Reassociate::ReassociateExpression(BinaryOperator *I) { 2158 // First, walk the expression tree, linearizing the tree, collecting the 2159 // operand information. 2160 SmallVector<RepeatedValue, 8> Tree; 2161 MadeChange |= LinearizeExprTree(I, Tree); 2162 SmallVector<ValueEntry, 8> Ops; 2163 Ops.reserve(Tree.size()); 2164 for (unsigned i = 0, e = Tree.size(); i != e; ++i) { 2165 RepeatedValue E = Tree[i]; 2166 Ops.append(E.second.getZExtValue(), 2167 ValueEntry(getRank(E.first), E.first)); 2168 } 2169 2170 DEBUG(dbgs() << "RAIn:\t"; PrintOps(I, Ops); dbgs() << '\n'); 2171 2172 // Now that we have linearized the tree to a list and have gathered all of 2173 // the operands and their ranks, sort the operands by their rank. Use a 2174 // stable_sort so that values with equal ranks will have their relative 2175 // positions maintained (and so the compiler is deterministic). Note that 2176 // this sorts so that the highest ranking values end up at the beginning of 2177 // the vector. 2178 std::stable_sort(Ops.begin(), Ops.end()); 2179 2180 // Now that we have the expression tree in a convenient 2181 // sorted form, optimize it globally if possible. 2182 if (Value *V = OptimizeExpression(I, Ops)) { 2183 if (V == I) 2184 // Self-referential expression in unreachable code. 2185 return; 2186 // This expression tree simplified to something that isn't a tree, 2187 // eliminate it. 2188 DEBUG(dbgs() << "Reassoc to scalar: " << *V << '\n'); 2189 I->replaceAllUsesWith(V); 2190 if (Instruction *VI = dyn_cast<Instruction>(V)) 2191 VI->setDebugLoc(I->getDebugLoc()); 2192 RedoInsts.insert(I); 2193 ++NumAnnihil; 2194 return; 2195 } 2196 2197 // We want to sink immediates as deeply as possible except in the case where 2198 // this is a multiply tree used only by an add, and the immediate is a -1. 2199 // In this case we reassociate to put the negation on the outside so that we 2200 // can fold the negation into the add: (-X)*Y + Z -> Z-X*Y 2201 if (I->hasOneUse()) { 2202 if (I->getOpcode() == Instruction::Mul && 2203 cast<Instruction>(I->user_back())->getOpcode() == Instruction::Add && 2204 isa<ConstantInt>(Ops.back().Op) && 2205 cast<ConstantInt>(Ops.back().Op)->isAllOnesValue()) { 2206 ValueEntry Tmp = Ops.pop_back_val(); 2207 Ops.insert(Ops.begin(), Tmp); 2208 } else if (I->getOpcode() == Instruction::FMul && 2209 cast<Instruction>(I->user_back())->getOpcode() == 2210 Instruction::FAdd && 2211 isa<ConstantFP>(Ops.back().Op) && 2212 cast<ConstantFP>(Ops.back().Op)->isExactlyValue(-1.0)) { 2213 ValueEntry Tmp = Ops.pop_back_val(); 2214 Ops.insert(Ops.begin(), Tmp); 2215 } 2216 } 2217 2218 DEBUG(dbgs() << "RAOut:\t"; PrintOps(I, Ops); dbgs() << '\n'); 2219 2220 if (Ops.size() == 1) { 2221 if (Ops[0].Op == I) 2222 // Self-referential expression in unreachable code. 2223 return; 2224 2225 // This expression tree simplified to something that isn't a tree, 2226 // eliminate it. 2227 I->replaceAllUsesWith(Ops[0].Op); 2228 if (Instruction *OI = dyn_cast<Instruction>(Ops[0].Op)) 2229 OI->setDebugLoc(I->getDebugLoc()); 2230 RedoInsts.insert(I); 2231 return; 2232 } 2233 2234 // Now that we ordered and optimized the expressions, splat them back into 2235 // the expression tree, removing any unneeded nodes. 2236 RewriteExprTree(I, Ops); 2237 } 2238 2239 bool Reassociate::runOnFunction(Function &F) { 2240 if (skipOptnoneFunction(F)) 2241 return false; 2242 2243 // Calculate the rank map for F 2244 BuildRankMap(F); 2245 2246 MadeChange = false; 2247 for (Function::iterator BI = F.begin(), BE = F.end(); BI != BE; ++BI) { 2248 // Optimize every instruction in the basic block. 2249 for (BasicBlock::iterator II = BI->begin(), IE = BI->end(); II != IE; ) 2250 if (isInstructionTriviallyDead(&*II)) { 2251 EraseInst(&*II++); 2252 } else { 2253 OptimizeInst(&*II); 2254 assert(II->getParent() == BI && "Moved to a different block!"); 2255 ++II; 2256 } 2257 2258 // If this produced extra instructions to optimize, handle them now. 2259 while (!RedoInsts.empty()) { 2260 Instruction *I = RedoInsts.pop_back_val(); 2261 if (isInstructionTriviallyDead(I)) 2262 EraseInst(I); 2263 else 2264 OptimizeInst(I); 2265 } 2266 } 2267 2268 // We are done with the rank map. 2269 RankMap.clear(); 2270 ValueRankMap.clear(); 2271 2272 return MadeChange; 2273 } 2274