Reassociate.cpp revision 303975
1208963Srdivacky//===- Reassociate.cpp - Reassociate binary expressions -------------------===// 2208963Srdivacky// 3265420Simp// The LLVM Compiler Infrastructure 4231057Sdim// 5208963Srdivacky// This file is distributed under the University of Illinois Open Source 6208963Srdivacky// License. See LICENSE.TXT for details. 7208963Srdivacky// 8210299Sed//===----------------------------------------------------------------------===// 9280031Sdim// 10288943Sdim// This pass reassociates commutative expressions in an order that is designed 11276479Sdim// to promote better constant propagation, GCSE, LICM, PRE, etc. 12210299Sed// 13212904Sdim// For example: 4 + (x + 5) -> x + (4 + 5) 14210299Sed// 15210299Sed// In the implementation of this algorithm, constants are assigned rank = 0, 16218893Sdim// function arguments are rank = 1, and other values are assigned ranks 17261991Sdim// corresponding to the reverse post order traversal of current function 18288943Sdim// (starting at 2), which effectively gives values in deep loops higher rank 19261991Sdim// than values not in loops. 20210299Sed// 21288943Sdim//===----------------------------------------------------------------------===// 22210299Sed 23210299Sed#include "llvm/Transforms/Scalar.h" 24276479Sdim#include "llvm/ADT/DenseMap.h" 25210299Sed#include "llvm/ADT/PostOrderIterator.h" 26288943Sdim#include "llvm/ADT/STLExtras.h" 27218893Sdim#include "llvm/ADT/SetVector.h" 28288943Sdim#include "llvm/ADT/Statistic.h" 29296417Sdim#include "llvm/Analysis/GlobalsModRef.h" 30261991Sdim#include "llvm/Analysis/ValueTracking.h" 31210299Sed#include "llvm/IR/CFG.h" 32210299Sed#include "llvm/IR/Constants.h" 33210299Sed#include "llvm/IR/DerivedTypes.h" 34210299Sed#include "llvm/IR/Function.h" 35234353Sdim#include "llvm/IR/IRBuilder.h" 36288943Sdim#include "llvm/IR/Instructions.h" 37210299Sed#include "llvm/IR/IntrinsicInst.h" 38276479Sdim#include "llvm/IR/ValueHandle.h" 39288943Sdim#include "llvm/Pass.h" 40261991Sdim#include "llvm/Support/Debug.h" 41288943Sdim#include "llvm/Support/raw_ostream.h" 42210299Sed#include "llvm/Transforms/Utils/Local.h" 43234353Sdim#include <algorithm> 44288943Sdimusing namespace llvm; 45210299Sed 46243830Sdim#define DEBUG_TYPE "reassociate" 47210299Sed 48276479SdimSTATISTIC(NumChanged, "Number of insts reassociated"); 49276479SdimSTATISTIC(NumAnnihil, "Number of expr tree annihilated"); 50210299SedSTATISTIC(NumFactor , "Number of multiplies factored"); 51212904Sdim 52288943Sdimnamespace { 53288943Sdim struct ValueEntry { 54261991Sdim unsigned Rank; 55210299Sed Value *Op; 56208963Srdivacky ValueEntry(unsigned R, Value *O) : Rank(R), Op(O) {} 57231057Sdim }; 58231057Sdim inline bool operator<(const ValueEntry &LHS, const ValueEntry &RHS) { 59231057Sdim return LHS.Rank > RHS.Rank; // Sort so that highest rank goes to start. 60231057Sdim } 61231057Sdim} 62296417Sdim 63296417Sdim#ifndef NDEBUG 64208963Srdivacky/// Print out the expression identified in the Ops list. 65208963Srdivacky/// 66static 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 78namespace { 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 147namespace { 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 RecursivelyEraseDeadInsts(Instruction *I, 187 SetVector<AssertingVH<Instruction>> &Insts); 188 void OptimizeInst(Instruction *I); 189 Instruction *canonicalizeNegConstExpr(Instruction *I); 190 }; 191} 192 193XorOpnd::XorOpnd(Value *V) { 194 assert(!isa<ConstantInt>(V) && "No ConstantInt"); 195 OrigVal = V; 196 Instruction *I = dyn_cast<Instruction>(V); 197 SymbolicRank = 0; 198 199 if (I && (I->getOpcode() == Instruction::Or || 200 I->getOpcode() == Instruction::And)) { 201 Value *V0 = I->getOperand(0); 202 Value *V1 = I->getOperand(1); 203 if (isa<ConstantInt>(V0)) 204 std::swap(V0, V1); 205 206 if (ConstantInt *C = dyn_cast<ConstantInt>(V1)) { 207 ConstPart = C->getValue(); 208 SymbolicPart = V0; 209 isOr = (I->getOpcode() == Instruction::Or); 210 return; 211 } 212 } 213 214 // view the operand as "V | 0" 215 SymbolicPart = V; 216 ConstPart = APInt::getNullValue(V->getType()->getIntegerBitWidth()); 217 isOr = true; 218} 219 220char Reassociate::ID = 0; 221INITIALIZE_PASS(Reassociate, "reassociate", 222 "Reassociate expressions", false, false) 223 224// Public interface to the Reassociate pass 225FunctionPass *llvm::createReassociatePass() { return new Reassociate(); } 226 227/// Return true if V is an instruction of the specified opcode and if it 228/// only has one use. 229static BinaryOperator *isReassociableOp(Value *V, unsigned Opcode) { 230 if (V->hasOneUse() && isa<Instruction>(V) && 231 cast<Instruction>(V)->getOpcode() == Opcode && 232 (!isa<FPMathOperator>(V) || 233 cast<Instruction>(V)->hasUnsafeAlgebra())) 234 return cast<BinaryOperator>(V); 235 return nullptr; 236} 237 238static BinaryOperator *isReassociableOp(Value *V, unsigned Opcode1, 239 unsigned Opcode2) { 240 if (V->hasOneUse() && isa<Instruction>(V) && 241 (cast<Instruction>(V)->getOpcode() == Opcode1 || 242 cast<Instruction>(V)->getOpcode() == Opcode2) && 243 (!isa<FPMathOperator>(V) || 244 cast<Instruction>(V)->hasUnsafeAlgebra())) 245 return cast<BinaryOperator>(V); 246 return nullptr; 247} 248 249void Reassociate::BuildRankMap(Function &F) { 250 unsigned i = 2; 251 252 // Assign distinct ranks to function arguments. 253 for (Function::arg_iterator I = F.arg_begin(), E = F.arg_end(); I != E; ++I) { 254 ValueRankMap[&*I] = ++i; 255 DEBUG(dbgs() << "Calculated Rank[" << I->getName() << "] = " << i << "\n"); 256 } 257 258 ReversePostOrderTraversal<Function*> RPOT(&F); 259 for (ReversePostOrderTraversal<Function*>::rpo_iterator I = RPOT.begin(), 260 E = RPOT.end(); I != E; ++I) { 261 BasicBlock *BB = *I; 262 unsigned BBRank = RankMap[BB] = ++i << 16; 263 264 // Walk the basic block, adding precomputed ranks for any instructions that 265 // we cannot move. This ensures that the ranks for these instructions are 266 // all different in the block. 267 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) 268 if (mayBeMemoryDependent(*I)) 269 ValueRankMap[&*I] = ++BBRank; 270 } 271} 272 273unsigned Reassociate::getRank(Value *V) { 274 Instruction *I = dyn_cast<Instruction>(V); 275 if (!I) { 276 if (isa<Argument>(V)) return ValueRankMap[V]; // Function argument. 277 return 0; // Otherwise it's a global or constant, rank 0. 278 } 279 280 if (unsigned Rank = ValueRankMap[I]) 281 return Rank; // Rank already known? 282 283 // If this is an expression, return the 1+MAX(rank(LHS), rank(RHS)) so that 284 // we can reassociate expressions for code motion! Since we do not recurse 285 // for PHI nodes, we cannot have infinite recursion here, because there 286 // cannot be loops in the value graph that do not go through PHI nodes. 287 unsigned Rank = 0, MaxRank = RankMap[I->getParent()]; 288 for (unsigned i = 0, e = I->getNumOperands(); 289 i != e && Rank != MaxRank; ++i) 290 Rank = std::max(Rank, getRank(I->getOperand(i))); 291 292 // If this is a not or neg instruction, do not count it for rank. This 293 // assures us that X and ~X will have the same rank. 294 if (!BinaryOperator::isNot(I) && !BinaryOperator::isNeg(I) && 295 !BinaryOperator::isFNeg(I)) 296 ++Rank; 297 298 DEBUG(dbgs() << "Calculated Rank[" << V->getName() << "] = " << Rank << "\n"); 299 300 return ValueRankMap[I] = Rank; 301} 302 303// Canonicalize constants to RHS. Otherwise, sort the operands by rank. 304void Reassociate::canonicalizeOperands(Instruction *I) { 305 assert(isa<BinaryOperator>(I) && "Expected binary operator."); 306 assert(I->isCommutative() && "Expected commutative operator."); 307 308 Value *LHS = I->getOperand(0); 309 Value *RHS = I->getOperand(1); 310 unsigned LHSRank = getRank(LHS); 311 unsigned RHSRank = getRank(RHS); 312 313 if (isa<Constant>(RHS)) 314 return; 315 316 if (isa<Constant>(LHS) || RHSRank < LHSRank) 317 cast<BinaryOperator>(I)->swapOperands(); 318} 319 320static BinaryOperator *CreateAdd(Value *S1, Value *S2, const Twine &Name, 321 Instruction *InsertBefore, Value *FlagsOp) { 322 if (S1->getType()->isIntOrIntVectorTy()) 323 return BinaryOperator::CreateAdd(S1, S2, Name, InsertBefore); 324 else { 325 BinaryOperator *Res = 326 BinaryOperator::CreateFAdd(S1, S2, Name, InsertBefore); 327 Res->setFastMathFlags(cast<FPMathOperator>(FlagsOp)->getFastMathFlags()); 328 return Res; 329 } 330} 331 332static BinaryOperator *CreateMul(Value *S1, Value *S2, const Twine &Name, 333 Instruction *InsertBefore, Value *FlagsOp) { 334 if (S1->getType()->isIntOrIntVectorTy()) 335 return BinaryOperator::CreateMul(S1, S2, Name, InsertBefore); 336 else { 337 BinaryOperator *Res = 338 BinaryOperator::CreateFMul(S1, S2, Name, InsertBefore); 339 Res->setFastMathFlags(cast<FPMathOperator>(FlagsOp)->getFastMathFlags()); 340 return Res; 341 } 342} 343 344static BinaryOperator *CreateNeg(Value *S1, const Twine &Name, 345 Instruction *InsertBefore, Value *FlagsOp) { 346 if (S1->getType()->isIntOrIntVectorTy()) 347 return BinaryOperator::CreateNeg(S1, Name, InsertBefore); 348 else { 349 BinaryOperator *Res = BinaryOperator::CreateFNeg(S1, Name, InsertBefore); 350 Res->setFastMathFlags(cast<FPMathOperator>(FlagsOp)->getFastMathFlags()); 351 return Res; 352 } 353} 354 355/// Replace 0-X with X*-1. 356static BinaryOperator *LowerNegateToMultiply(Instruction *Neg) { 357 Type *Ty = Neg->getType(); 358 Constant *NegOne = Ty->isIntOrIntVectorTy() ? 359 ConstantInt::getAllOnesValue(Ty) : ConstantFP::get(Ty, -1.0); 360 361 BinaryOperator *Res = CreateMul(Neg->getOperand(1), NegOne, "", Neg, Neg); 362 Neg->setOperand(1, Constant::getNullValue(Ty)); // Drop use of op. 363 Res->takeName(Neg); 364 Neg->replaceAllUsesWith(Res); 365 Res->setDebugLoc(Neg->getDebugLoc()); 366 return Res; 367} 368 369/// Returns k such that lambda(2^Bitwidth) = 2^k, where lambda is the Carmichael 370/// function. This means that x^(2^k) === 1 mod 2^Bitwidth for 371/// every odd x, i.e. x^(2^k) = 1 for every odd x in Bitwidth-bit arithmetic. 372/// Note that 0 <= k < Bitwidth, and if Bitwidth > 3 then x^(2^k) = 0 for every 373/// even x in Bitwidth-bit arithmetic. 374static unsigned CarmichaelShift(unsigned Bitwidth) { 375 if (Bitwidth < 3) 376 return Bitwidth - 1; 377 return Bitwidth - 2; 378} 379 380/// Add the extra weight 'RHS' to the existing weight 'LHS', 381/// reducing the combined weight using any special properties of the operation. 382/// The existing weight LHS represents the computation X op X op ... op X where 383/// X occurs LHS times. The combined weight represents X op X op ... op X with 384/// X occurring LHS + RHS times. If op is "Xor" for example then the combined 385/// operation is equivalent to X if LHS + RHS is odd, or 0 if LHS + RHS is even; 386/// the routine returns 1 in LHS in the first case, and 0 in LHS in the second. 387static void IncorporateWeight(APInt &LHS, const APInt &RHS, unsigned Opcode) { 388 // If we were working with infinite precision arithmetic then the combined 389 // weight would be LHS + RHS. But we are using finite precision arithmetic, 390 // and the APInt sum LHS + RHS may not be correct if it wraps (it is correct 391 // for nilpotent operations and addition, but not for idempotent operations 392 // and multiplication), so it is important to correctly reduce the combined 393 // weight back into range if wrapping would be wrong. 394 395 // If RHS is zero then the weight didn't change. 396 if (RHS.isMinValue()) 397 return; 398 // If LHS is zero then the combined weight is RHS. 399 if (LHS.isMinValue()) { 400 LHS = RHS; 401 return; 402 } 403 // From this point on we know that neither LHS nor RHS is zero. 404 405 if (Instruction::isIdempotent(Opcode)) { 406 // Idempotent means X op X === X, so any non-zero weight is equivalent to a 407 // weight of 1. Keeping weights at zero or one also means that wrapping is 408 // not a problem. 409 assert(LHS == 1 && RHS == 1 && "Weights not reduced!"); 410 return; // Return a weight of 1. 411 } 412 if (Instruction::isNilpotent(Opcode)) { 413 // Nilpotent means X op X === 0, so reduce weights modulo 2. 414 assert(LHS == 1 && RHS == 1 && "Weights not reduced!"); 415 LHS = 0; // 1 + 1 === 0 modulo 2. 416 return; 417 } 418 if (Opcode == Instruction::Add || Opcode == Instruction::FAdd) { 419 // TODO: Reduce the weight by exploiting nsw/nuw? 420 LHS += RHS; 421 return; 422 } 423 424 assert((Opcode == Instruction::Mul || Opcode == Instruction::FMul) && 425 "Unknown associative operation!"); 426 unsigned Bitwidth = LHS.getBitWidth(); 427 // If CM is the Carmichael number then a weight W satisfying W >= CM+Bitwidth 428 // can be replaced with W-CM. That's because x^W=x^(W-CM) for every Bitwidth 429 // bit number x, since either x is odd in which case x^CM = 1, or x is even in 430 // which case both x^W and x^(W - CM) are zero. By subtracting off multiples 431 // of CM like this weights can always be reduced to the range [0, CM+Bitwidth) 432 // which by a happy accident means that they can always be represented using 433 // Bitwidth bits. 434 // TODO: Reduce the weight by exploiting nsw/nuw? (Could do much better than 435 // the Carmichael number). 436 if (Bitwidth > 3) { 437 /// CM - The value of Carmichael's lambda function. 438 APInt CM = APInt::getOneBitSet(Bitwidth, CarmichaelShift(Bitwidth)); 439 // Any weight W >= Threshold can be replaced with W - CM. 440 APInt Threshold = CM + Bitwidth; 441 assert(LHS.ult(Threshold) && RHS.ult(Threshold) && "Weights not reduced!"); 442 // For Bitwidth 4 or more the following sum does not overflow. 443 LHS += RHS; 444 while (LHS.uge(Threshold)) 445 LHS -= CM; 446 } else { 447 // To avoid problems with overflow do everything the same as above but using 448 // a larger type. 449 unsigned CM = 1U << CarmichaelShift(Bitwidth); 450 unsigned Threshold = CM + Bitwidth; 451 assert(LHS.getZExtValue() < Threshold && RHS.getZExtValue() < Threshold && 452 "Weights not reduced!"); 453 unsigned Total = LHS.getZExtValue() + RHS.getZExtValue(); 454 while (Total >= Threshold) 455 Total -= CM; 456 LHS = Total; 457 } 458} 459 460typedef std::pair<Value*, APInt> RepeatedValue; 461 462/// Given an associative binary expression, return the leaf 463/// nodes in Ops along with their weights (how many times the leaf occurs). The 464/// original expression is the same as 465/// (Ops[0].first op Ops[0].first op ... Ops[0].first) <- Ops[0].second times 466/// op 467/// (Ops[1].first op Ops[1].first op ... Ops[1].first) <- Ops[1].second times 468/// op 469/// ... 470/// op 471/// (Ops[N].first op Ops[N].first op ... Ops[N].first) <- Ops[N].second times 472/// 473/// Note that the values Ops[0].first, ..., Ops[N].first are all distinct. 474/// 475/// This routine may modify the function, in which case it returns 'true'. The 476/// changes it makes may well be destructive, changing the value computed by 'I' 477/// to something completely different. Thus if the routine returns 'true' then 478/// you MUST either replace I with a new expression computed from the Ops array, 479/// or use RewriteExprTree to put the values back in. 480/// 481/// A leaf node is either not a binary operation of the same kind as the root 482/// node 'I' (i.e. is not a binary operator at all, or is, but with a different 483/// opcode), or is the same kind of binary operator but has a use which either 484/// does not belong to the expression, or does belong to the expression but is 485/// a leaf node. Every leaf node has at least one use that is a non-leaf node 486/// of the expression, while for non-leaf nodes (except for the root 'I') every 487/// use is a non-leaf node of the expression. 488/// 489/// For example: 490/// expression graph node names 491/// 492/// + | I 493/// / \ | 494/// + + | A, B 495/// / \ / \ | 496/// * + * | C, D, E 497/// / \ / \ / \ | 498/// + * | F, G 499/// 500/// The leaf nodes are C, E, F and G. The Ops array will contain (maybe not in 501/// that order) (C, 1), (E, 1), (F, 2), (G, 2). 502/// 503/// The expression is maximal: if some instruction is a binary operator of the 504/// same kind as 'I', and all of its uses are non-leaf nodes of the expression, 505/// then the instruction also belongs to the expression, is not a leaf node of 506/// it, and its operands also belong to the expression (but may be leaf nodes). 507/// 508/// NOTE: This routine will set operands of non-leaf non-root nodes to undef in 509/// order to ensure that every non-root node in the expression has *exactly one* 510/// use by a non-leaf node of the expression. This destruction means that the 511/// caller MUST either replace 'I' with a new expression or use something like 512/// RewriteExprTree to put the values back in if the routine indicates that it 513/// made a change by returning 'true'. 514/// 515/// In the above example either the right operand of A or the left operand of B 516/// will be replaced by undef. If it is B's operand then this gives: 517/// 518/// + | I 519/// / \ | 520/// + + | A, B - operand of B replaced with undef 521/// / \ \ | 522/// * + * | C, D, E 523/// / \ / \ / \ | 524/// + * | F, G 525/// 526/// Note that such undef operands can only be reached by passing through 'I'. 527/// For example, if you visit operands recursively starting from a leaf node 528/// then you will never see such an undef operand unless you get back to 'I', 529/// which requires passing through a phi node. 530/// 531/// Note that this routine may also mutate binary operators of the wrong type 532/// that have all uses inside the expression (i.e. only used by non-leaf nodes 533/// of the expression) if it can turn them into binary operators of the right 534/// type and thus make the expression bigger. 535 536static bool LinearizeExprTree(BinaryOperator *I, 537 SmallVectorImpl<RepeatedValue> &Ops) { 538 DEBUG(dbgs() << "LINEARIZE: " << *I << '\n'); 539 unsigned Bitwidth = I->getType()->getScalarType()->getPrimitiveSizeInBits(); 540 unsigned Opcode = I->getOpcode(); 541 assert(I->isAssociative() && I->isCommutative() && 542 "Expected an associative and commutative operation!"); 543 544 // Visit all operands of the expression, keeping track of their weight (the 545 // number of paths from the expression root to the operand, or if you like 546 // the number of times that operand occurs in the linearized expression). 547 // For example, if I = X + A, where X = A + B, then I, X and B have weight 1 548 // while A has weight two. 549 550 // Worklist of non-leaf nodes (their operands are in the expression too) along 551 // with their weights, representing a certain number of paths to the operator. 552 // If an operator occurs in the worklist multiple times then we found multiple 553 // ways to get to it. 554 SmallVector<std::pair<BinaryOperator*, APInt>, 8> Worklist; // (Op, Weight) 555 Worklist.push_back(std::make_pair(I, APInt(Bitwidth, 1))); 556 bool Changed = false; 557 558 // Leaves of the expression are values that either aren't the right kind of 559 // operation (eg: a constant, or a multiply in an add tree), or are, but have 560 // some uses that are not inside the expression. For example, in I = X + X, 561 // X = A + B, the value X has two uses (by I) that are in the expression. If 562 // X has any other uses, for example in a return instruction, then we consider 563 // X to be a leaf, and won't analyze it further. When we first visit a value, 564 // if it has more than one use then at first we conservatively consider it to 565 // be a leaf. Later, as the expression is explored, we may discover some more 566 // uses of the value from inside the expression. If all uses turn out to be 567 // from within the expression (and the value is a binary operator of the right 568 // kind) then the value is no longer considered to be a leaf, and its operands 569 // are explored. 570 571 // Leaves - Keeps track of the set of putative leaves as well as the number of 572 // paths to each leaf seen so far. 573 typedef DenseMap<Value*, APInt> LeafMap; 574 LeafMap Leaves; // Leaf -> Total weight so far. 575 SmallVector<Value*, 8> LeafOrder; // Ensure deterministic leaf output order. 576 577#ifndef NDEBUG 578 SmallPtrSet<Value*, 8> Visited; // For sanity checking the iteration scheme. 579#endif 580 while (!Worklist.empty()) { 581 std::pair<BinaryOperator*, APInt> P = Worklist.pop_back_val(); 582 I = P.first; // We examine the operands of this binary operator. 583 584 for (unsigned OpIdx = 0; OpIdx < 2; ++OpIdx) { // Visit operands. 585 Value *Op = I->getOperand(OpIdx); 586 APInt Weight = P.second; // Number of paths to this operand. 587 DEBUG(dbgs() << "OPERAND: " << *Op << " (" << Weight << ")\n"); 588 assert(!Op->use_empty() && "No uses, so how did we get to it?!"); 589 590 // If this is a binary operation of the right kind with only one use then 591 // add its operands to the expression. 592 if (BinaryOperator *BO = isReassociableOp(Op, Opcode)) { 593 assert(Visited.insert(Op).second && "Not first visit!"); 594 DEBUG(dbgs() << "DIRECT ADD: " << *Op << " (" << Weight << ")\n"); 595 Worklist.push_back(std::make_pair(BO, Weight)); 596 continue; 597 } 598 599 // Appears to be a leaf. Is the operand already in the set of leaves? 600 LeafMap::iterator It = Leaves.find(Op); 601 if (It == Leaves.end()) { 602 // Not in the leaf map. Must be the first time we saw this operand. 603 assert(Visited.insert(Op).second && "Not first visit!"); 604 if (!Op->hasOneUse()) { 605 // This value has uses not accounted for by the expression, so it is 606 // not safe to modify. Mark it as being a leaf. 607 DEBUG(dbgs() << "ADD USES LEAF: " << *Op << " (" << Weight << ")\n"); 608 LeafOrder.push_back(Op); 609 Leaves[Op] = Weight; 610 continue; 611 } 612 // No uses outside the expression, try morphing it. 613 } else if (It != Leaves.end()) { 614 // Already in the leaf map. 615 assert(Visited.count(Op) && "In leaf map but not visited!"); 616 617 // Update the number of paths to the leaf. 618 IncorporateWeight(It->second, Weight, Opcode); 619 620#if 0 // TODO: Re-enable once PR13021 is fixed. 621 // The leaf already has one use from inside the expression. As we want 622 // exactly one such use, drop this new use of the leaf. 623 assert(!Op->hasOneUse() && "Only one use, but we got here twice!"); 624 I->setOperand(OpIdx, UndefValue::get(I->getType())); 625 Changed = true; 626 627 // If the leaf is a binary operation of the right kind and we now see 628 // that its multiple original uses were in fact all by nodes belonging 629 // to the expression, then no longer consider it to be a leaf and add 630 // its operands to the expression. 631 if (BinaryOperator *BO = isReassociableOp(Op, Opcode)) { 632 DEBUG(dbgs() << "UNLEAF: " << *Op << " (" << It->second << ")\n"); 633 Worklist.push_back(std::make_pair(BO, It->second)); 634 Leaves.erase(It); 635 continue; 636 } 637#endif 638 639 // If we still have uses that are not accounted for by the expression 640 // then it is not safe to modify the value. 641 if (!Op->hasOneUse()) 642 continue; 643 644 // No uses outside the expression, try morphing it. 645 Weight = It->second; 646 Leaves.erase(It); // Since the value may be morphed below. 647 } 648 649 // At this point we have a value which, first of all, is not a binary 650 // expression of the right kind, and secondly, is only used inside the 651 // expression. This means that it can safely be modified. See if we 652 // can usefully morph it into an expression of the right kind. 653 assert((!isa<Instruction>(Op) || 654 cast<Instruction>(Op)->getOpcode() != Opcode 655 || (isa<FPMathOperator>(Op) && 656 !cast<Instruction>(Op)->hasUnsafeAlgebra())) && 657 "Should have been handled above!"); 658 assert(Op->hasOneUse() && "Has uses outside the expression tree!"); 659 660 // If this is a multiply expression, turn any internal negations into 661 // multiplies by -1 so they can be reassociated. 662 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op)) 663 if ((Opcode == Instruction::Mul && BinaryOperator::isNeg(BO)) || 664 (Opcode == Instruction::FMul && BinaryOperator::isFNeg(BO))) { 665 DEBUG(dbgs() << "MORPH LEAF: " << *Op << " (" << Weight << ") TO "); 666 BO = LowerNegateToMultiply(BO); 667 DEBUG(dbgs() << *BO << '\n'); 668 Worklist.push_back(std::make_pair(BO, Weight)); 669 Changed = true; 670 continue; 671 } 672 673 // Failed to morph into an expression of the right type. This really is 674 // a leaf. 675 DEBUG(dbgs() << "ADD LEAF: " << *Op << " (" << Weight << ")\n"); 676 assert(!isReassociableOp(Op, Opcode) && "Value was morphed?"); 677 LeafOrder.push_back(Op); 678 Leaves[Op] = Weight; 679 } 680 } 681 682 // The leaves, repeated according to their weights, represent the linearized 683 // form of the expression. 684 for (unsigned i = 0, e = LeafOrder.size(); i != e; ++i) { 685 Value *V = LeafOrder[i]; 686 LeafMap::iterator It = Leaves.find(V); 687 if (It == Leaves.end()) 688 // Node initially thought to be a leaf wasn't. 689 continue; 690 assert(!isReassociableOp(V, Opcode) && "Shouldn't be a leaf!"); 691 APInt Weight = It->second; 692 if (Weight.isMinValue()) 693 // Leaf already output or weight reduction eliminated it. 694 continue; 695 // Ensure the leaf is only output once. 696 It->second = 0; 697 Ops.push_back(std::make_pair(V, Weight)); 698 } 699 700 // For nilpotent operations or addition there may be no operands, for example 701 // because the expression was "X xor X" or consisted of 2^Bitwidth additions: 702 // in both cases the weight reduces to 0 causing the value to be skipped. 703 if (Ops.empty()) { 704 Constant *Identity = ConstantExpr::getBinOpIdentity(Opcode, I->getType()); 705 assert(Identity && "Associative operation without identity!"); 706 Ops.emplace_back(Identity, APInt(Bitwidth, 1)); 707 } 708 709 return Changed; 710} 711 712/// Now that the operands for this expression tree are 713/// linearized and optimized, emit them in-order. 714void Reassociate::RewriteExprTree(BinaryOperator *I, 715 SmallVectorImpl<ValueEntry> &Ops) { 716 assert(Ops.size() > 1 && "Single values should be used directly!"); 717 718 // Since our optimizations should never increase the number of operations, the 719 // new expression can usually be written reusing the existing binary operators 720 // from the original expression tree, without creating any new instructions, 721 // though the rewritten expression may have a completely different topology. 722 // We take care to not change anything if the new expression will be the same 723 // as the original. If more than trivial changes (like commuting operands) 724 // were made then we are obliged to clear out any optional subclass data like 725 // nsw flags. 726 727 /// NodesToRewrite - Nodes from the original expression available for writing 728 /// the new expression into. 729 SmallVector<BinaryOperator*, 8> NodesToRewrite; 730 unsigned Opcode = I->getOpcode(); 731 BinaryOperator *Op = I; 732 733 /// NotRewritable - The operands being written will be the leaves of the new 734 /// expression and must not be used as inner nodes (via NodesToRewrite) by 735 /// mistake. Inner nodes are always reassociable, and usually leaves are not 736 /// (if they were they would have been incorporated into the expression and so 737 /// would not be leaves), so most of the time there is no danger of this. But 738 /// in rare cases a leaf may become reassociable if an optimization kills uses 739 /// of it, or it may momentarily become reassociable during rewriting (below) 740 /// due it being removed as an operand of one of its uses. Ensure that misuse 741 /// of leaf nodes as inner nodes cannot occur by remembering all of the future 742 /// leaves and refusing to reuse any of them as inner nodes. 743 SmallPtrSet<Value*, 8> NotRewritable; 744 for (unsigned i = 0, e = Ops.size(); i != e; ++i) 745 NotRewritable.insert(Ops[i].Op); 746 747 // ExpressionChanged - Non-null if the rewritten expression differs from the 748 // original in some non-trivial way, requiring the clearing of optional flags. 749 // Flags are cleared from the operator in ExpressionChanged up to I inclusive. 750 BinaryOperator *ExpressionChanged = nullptr; 751 for (unsigned i = 0; ; ++i) { 752 // The last operation (which comes earliest in the IR) is special as both 753 // operands will come from Ops, rather than just one with the other being 754 // a subexpression. 755 if (i+2 == Ops.size()) { 756 Value *NewLHS = Ops[i].Op; 757 Value *NewRHS = Ops[i+1].Op; 758 Value *OldLHS = Op->getOperand(0); 759 Value *OldRHS = Op->getOperand(1); 760 761 if (NewLHS == OldLHS && NewRHS == OldRHS) 762 // Nothing changed, leave it alone. 763 break; 764 765 if (NewLHS == OldRHS && NewRHS == OldLHS) { 766 // The order of the operands was reversed. Swap them. 767 DEBUG(dbgs() << "RA: " << *Op << '\n'); 768 Op->swapOperands(); 769 DEBUG(dbgs() << "TO: " << *Op << '\n'); 770 MadeChange = true; 771 ++NumChanged; 772 break; 773 } 774 775 // The new operation differs non-trivially from the original. Overwrite 776 // the old operands with the new ones. 777 DEBUG(dbgs() << "RA: " << *Op << '\n'); 778 if (NewLHS != OldLHS) { 779 BinaryOperator *BO = isReassociableOp(OldLHS, Opcode); 780 if (BO && !NotRewritable.count(BO)) 781 NodesToRewrite.push_back(BO); 782 Op->setOperand(0, NewLHS); 783 } 784 if (NewRHS != OldRHS) { 785 BinaryOperator *BO = isReassociableOp(OldRHS, Opcode); 786 if (BO && !NotRewritable.count(BO)) 787 NodesToRewrite.push_back(BO); 788 Op->setOperand(1, NewRHS); 789 } 790 DEBUG(dbgs() << "TO: " << *Op << '\n'); 791 792 ExpressionChanged = Op; 793 MadeChange = true; 794 ++NumChanged; 795 796 break; 797 } 798 799 // Not the last operation. The left-hand side will be a sub-expression 800 // while the right-hand side will be the current element of Ops. 801 Value *NewRHS = Ops[i].Op; 802 if (NewRHS != Op->getOperand(1)) { 803 DEBUG(dbgs() << "RA: " << *Op << '\n'); 804 if (NewRHS == Op->getOperand(0)) { 805 // The new right-hand side was already present as the left operand. If 806 // we are lucky then swapping the operands will sort out both of them. 807 Op->swapOperands(); 808 } else { 809 // Overwrite with the new right-hand side. 810 BinaryOperator *BO = isReassociableOp(Op->getOperand(1), Opcode); 811 if (BO && !NotRewritable.count(BO)) 812 NodesToRewrite.push_back(BO); 813 Op->setOperand(1, NewRHS); 814 ExpressionChanged = Op; 815 } 816 DEBUG(dbgs() << "TO: " << *Op << '\n'); 817 MadeChange = true; 818 ++NumChanged; 819 } 820 821 // Now deal with the left-hand side. If this is already an operation node 822 // from the original expression then just rewrite the rest of the expression 823 // into it. 824 BinaryOperator *BO = isReassociableOp(Op->getOperand(0), Opcode); 825 if (BO && !NotRewritable.count(BO)) { 826 Op = BO; 827 continue; 828 } 829 830 // Otherwise, grab a spare node from the original expression and use that as 831 // the left-hand side. If there are no nodes left then the optimizers made 832 // an expression with more nodes than the original! This usually means that 833 // they did something stupid but it might mean that the problem was just too 834 // hard (finding the mimimal number of multiplications needed to realize a 835 // multiplication expression is NP-complete). Whatever the reason, smart or 836 // stupid, create a new node if there are none left. 837 BinaryOperator *NewOp; 838 if (NodesToRewrite.empty()) { 839 Constant *Undef = UndefValue::get(I->getType()); 840 NewOp = BinaryOperator::Create(Instruction::BinaryOps(Opcode), 841 Undef, Undef, "", I); 842 if (NewOp->getType()->isFPOrFPVectorTy()) 843 NewOp->setFastMathFlags(I->getFastMathFlags()); 844 } else { 845 NewOp = NodesToRewrite.pop_back_val(); 846 } 847 848 DEBUG(dbgs() << "RA: " << *Op << '\n'); 849 Op->setOperand(0, NewOp); 850 DEBUG(dbgs() << "TO: " << *Op << '\n'); 851 ExpressionChanged = Op; 852 MadeChange = true; 853 ++NumChanged; 854 Op = NewOp; 855 } 856 857 // If the expression changed non-trivially then clear out all subclass data 858 // starting from the operator specified in ExpressionChanged, and compactify 859 // the operators to just before the expression root to guarantee that the 860 // expression tree is dominated by all of Ops. 861 if (ExpressionChanged) 862 do { 863 // Preserve FastMathFlags. 864 if (isa<FPMathOperator>(I)) { 865 FastMathFlags Flags = I->getFastMathFlags(); 866 ExpressionChanged->clearSubclassOptionalData(); 867 ExpressionChanged->setFastMathFlags(Flags); 868 } else 869 ExpressionChanged->clearSubclassOptionalData(); 870 871 if (ExpressionChanged == I) 872 break; 873 ExpressionChanged->moveBefore(I); 874 ExpressionChanged = cast<BinaryOperator>(*ExpressionChanged->user_begin()); 875 } while (1); 876 877 // Throw away any left over nodes from the original expression. 878 for (unsigned i = 0, e = NodesToRewrite.size(); i != e; ++i) 879 RedoInsts.insert(NodesToRewrite[i]); 880} 881 882/// Insert instructions before the instruction pointed to by BI, 883/// that computes the negative version of the value specified. The negative 884/// version of the value is returned, and BI is left pointing at the instruction 885/// that should be processed next by the reassociation pass. 886/// Also add intermediate instructions to the redo list that are modified while 887/// pushing the negates through adds. These will be revisited to see if 888/// additional opportunities have been exposed. 889static Value *NegateValue(Value *V, Instruction *BI, 890 SetVector<AssertingVH<Instruction>> &ToRedo) { 891 if (Constant *C = dyn_cast<Constant>(V)) { 892 if (C->getType()->isFPOrFPVectorTy()) { 893 return ConstantExpr::getFNeg(C); 894 } 895 return ConstantExpr::getNeg(C); 896 } 897 898 899 // We are trying to expose opportunity for reassociation. One of the things 900 // that we want to do to achieve this is to push a negation as deep into an 901 // expression chain as possible, to expose the add instructions. In practice, 902 // this means that we turn this: 903 // X = -(A+12+C+D) into X = -A + -12 + -C + -D = -12 + -A + -C + -D 904 // so that later, a: Y = 12+X could get reassociated with the -12 to eliminate 905 // the constants. We assume that instcombine will clean up the mess later if 906 // we introduce tons of unnecessary negation instructions. 907 // 908 if (BinaryOperator *I = 909 isReassociableOp(V, Instruction::Add, Instruction::FAdd)) { 910 // Push the negates through the add. 911 I->setOperand(0, NegateValue(I->getOperand(0), BI, ToRedo)); 912 I->setOperand(1, NegateValue(I->getOperand(1), BI, ToRedo)); 913 if (I->getOpcode() == Instruction::Add) { 914 I->setHasNoUnsignedWrap(false); 915 I->setHasNoSignedWrap(false); 916 } 917 918 // We must move the add instruction here, because the neg instructions do 919 // not dominate the old add instruction in general. By moving it, we are 920 // assured that the neg instructions we just inserted dominate the 921 // instruction we are about to insert after them. 922 // 923 I->moveBefore(BI); 924 I->setName(I->getName()+".neg"); 925 926 // Add the intermediate negates to the redo list as processing them later 927 // could expose more reassociating opportunities. 928 ToRedo.insert(I); 929 return I; 930 } 931 932 // Okay, we need to materialize a negated version of V with an instruction. 933 // Scan the use lists of V to see if we have one already. 934 for (User *U : V->users()) { 935 if (!BinaryOperator::isNeg(U) && !BinaryOperator::isFNeg(U)) 936 continue; 937 938 // We found one! Now we have to make sure that the definition dominates 939 // this use. We do this by moving it to the entry block (if it is a 940 // non-instruction value) or right after the definition. These negates will 941 // be zapped by reassociate later, so we don't need much finesse here. 942 BinaryOperator *TheNeg = cast<BinaryOperator>(U); 943 944 // Verify that the negate is in this function, V might be a constant expr. 945 if (TheNeg->getParent()->getParent() != BI->getParent()->getParent()) 946 continue; 947 948 BasicBlock::iterator InsertPt; 949 if (Instruction *InstInput = dyn_cast<Instruction>(V)) { 950 if (InvokeInst *II = dyn_cast<InvokeInst>(InstInput)) { 951 InsertPt = II->getNormalDest()->begin(); 952 } else { 953 InsertPt = ++InstInput->getIterator(); 954 } 955 while (isa<PHINode>(InsertPt)) ++InsertPt; 956 } else { 957 InsertPt = TheNeg->getParent()->getParent()->getEntryBlock().begin(); 958 } 959 TheNeg->moveBefore(&*InsertPt); 960 if (TheNeg->getOpcode() == Instruction::Sub) { 961 TheNeg->setHasNoUnsignedWrap(false); 962 TheNeg->setHasNoSignedWrap(false); 963 } else { 964 TheNeg->andIRFlags(BI); 965 } 966 ToRedo.insert(TheNeg); 967 return TheNeg; 968 } 969 970 // Insert a 'neg' instruction that subtracts the value from zero to get the 971 // negation. 972 BinaryOperator *NewNeg = CreateNeg(V, V->getName() + ".neg", BI, BI); 973 ToRedo.insert(NewNeg); 974 return NewNeg; 975} 976 977/// Return true if we should break up this subtract of X-Y into (X + -Y). 978static bool ShouldBreakUpSubtract(Instruction *Sub) { 979 // If this is a negation, we can't split it up! 980 if (BinaryOperator::isNeg(Sub) || BinaryOperator::isFNeg(Sub)) 981 return false; 982 983 // Don't breakup X - undef. 984 if (isa<UndefValue>(Sub->getOperand(1))) 985 return false; 986 987 // Don't bother to break this up unless either the LHS is an associable add or 988 // subtract or if this is only used by one. 989 Value *V0 = Sub->getOperand(0); 990 if (isReassociableOp(V0, Instruction::Add, Instruction::FAdd) || 991 isReassociableOp(V0, Instruction::Sub, Instruction::FSub)) 992 return true; 993 Value *V1 = Sub->getOperand(1); 994 if (isReassociableOp(V1, Instruction::Add, Instruction::FAdd) || 995 isReassociableOp(V1, Instruction::Sub, Instruction::FSub)) 996 return true; 997 Value *VB = Sub->user_back(); 998 if (Sub->hasOneUse() && 999 (isReassociableOp(VB, Instruction::Add, Instruction::FAdd) || 1000 isReassociableOp(VB, Instruction::Sub, Instruction::FSub))) 1001 return true; 1002 1003 return false; 1004} 1005 1006/// If we have (X-Y), and if either X is an add, or if this is only used by an 1007/// add, transform this into (X+(0-Y)) to promote better reassociation. 1008static BinaryOperator * 1009BreakUpSubtract(Instruction *Sub, SetVector<AssertingVH<Instruction>> &ToRedo) { 1010 // Convert a subtract into an add and a neg instruction. This allows sub 1011 // instructions to be commuted with other add instructions. 1012 // 1013 // Calculate the negative value of Operand 1 of the sub instruction, 1014 // and set it as the RHS of the add instruction we just made. 1015 // 1016 Value *NegVal = NegateValue(Sub->getOperand(1), Sub, ToRedo); 1017 BinaryOperator *New = CreateAdd(Sub->getOperand(0), NegVal, "", Sub, Sub); 1018 Sub->setOperand(0, Constant::getNullValue(Sub->getType())); // Drop use of op. 1019 Sub->setOperand(1, Constant::getNullValue(Sub->getType())); // Drop use of op. 1020 New->takeName(Sub); 1021 1022 // Everyone now refers to the add instruction. 1023 Sub->replaceAllUsesWith(New); 1024 New->setDebugLoc(Sub->getDebugLoc()); 1025 1026 DEBUG(dbgs() << "Negated: " << *New << '\n'); 1027 return New; 1028} 1029 1030/// If this is a shift of a reassociable multiply or is used by one, change 1031/// this into a multiply by a constant to assist with further reassociation. 1032static BinaryOperator *ConvertShiftToMul(Instruction *Shl) { 1033 Constant *MulCst = ConstantInt::get(Shl->getType(), 1); 1034 MulCst = ConstantExpr::getShl(MulCst, cast<Constant>(Shl->getOperand(1))); 1035 1036 BinaryOperator *Mul = 1037 BinaryOperator::CreateMul(Shl->getOperand(0), MulCst, "", Shl); 1038 Shl->setOperand(0, UndefValue::get(Shl->getType())); // Drop use of op. 1039 Mul->takeName(Shl); 1040 1041 // Everyone now refers to the mul instruction. 1042 Shl->replaceAllUsesWith(Mul); 1043 Mul->setDebugLoc(Shl->getDebugLoc()); 1044 1045 // We can safely preserve the nuw flag in all cases. It's also safe to turn a 1046 // nuw nsw shl into a nuw nsw mul. However, nsw in isolation requires special 1047 // handling. 1048 bool NSW = cast<BinaryOperator>(Shl)->hasNoSignedWrap(); 1049 bool NUW = cast<BinaryOperator>(Shl)->hasNoUnsignedWrap(); 1050 if (NSW && NUW) 1051 Mul->setHasNoSignedWrap(true); 1052 Mul->setHasNoUnsignedWrap(NUW); 1053 return Mul; 1054} 1055 1056/// Scan backwards and forwards among values with the same rank as element i 1057/// to see if X exists. If X does not exist, return i. This is useful when 1058/// scanning for 'x' when we see '-x' because they both get the same rank. 1059static unsigned FindInOperandList(SmallVectorImpl<ValueEntry> &Ops, unsigned i, 1060 Value *X) { 1061 unsigned XRank = Ops[i].Rank; 1062 unsigned e = Ops.size(); 1063 for (unsigned j = i+1; j != e && Ops[j].Rank == XRank; ++j) { 1064 if (Ops[j].Op == X) 1065 return j; 1066 if (Instruction *I1 = dyn_cast<Instruction>(Ops[j].Op)) 1067 if (Instruction *I2 = dyn_cast<Instruction>(X)) 1068 if (I1->isIdenticalTo(I2)) 1069 return j; 1070 } 1071 // Scan backwards. 1072 for (unsigned j = i-1; j != ~0U && Ops[j].Rank == XRank; --j) { 1073 if (Ops[j].Op == X) 1074 return j; 1075 if (Instruction *I1 = dyn_cast<Instruction>(Ops[j].Op)) 1076 if (Instruction *I2 = dyn_cast<Instruction>(X)) 1077 if (I1->isIdenticalTo(I2)) 1078 return j; 1079 } 1080 return i; 1081} 1082 1083/// Emit a tree of add instructions, summing Ops together 1084/// and returning the result. Insert the tree before I. 1085static Value *EmitAddTreeOfValues(Instruction *I, 1086 SmallVectorImpl<WeakVH> &Ops){ 1087 if (Ops.size() == 1) return Ops.back(); 1088 1089 Value *V1 = Ops.back(); 1090 Ops.pop_back(); 1091 Value *V2 = EmitAddTreeOfValues(I, Ops); 1092 return CreateAdd(V2, V1, "tmp", I, I); 1093} 1094 1095/// If V is an expression tree that is a multiplication sequence, 1096/// and if this sequence contains a multiply by Factor, 1097/// remove Factor from the tree and return the new tree. 1098Value *Reassociate::RemoveFactorFromExpression(Value *V, Value *Factor) { 1099 BinaryOperator *BO = isReassociableOp(V, Instruction::Mul, Instruction::FMul); 1100 if (!BO) 1101 return nullptr; 1102 1103 SmallVector<RepeatedValue, 8> Tree; 1104 MadeChange |= LinearizeExprTree(BO, Tree); 1105 SmallVector<ValueEntry, 8> Factors; 1106 Factors.reserve(Tree.size()); 1107 for (unsigned i = 0, e = Tree.size(); i != e; ++i) { 1108 RepeatedValue E = Tree[i]; 1109 Factors.append(E.second.getZExtValue(), 1110 ValueEntry(getRank(E.first), E.first)); 1111 } 1112 1113 bool FoundFactor = false; 1114 bool NeedsNegate = false; 1115 for (unsigned i = 0, e = Factors.size(); i != e; ++i) { 1116 if (Factors[i].Op == Factor) { 1117 FoundFactor = true; 1118 Factors.erase(Factors.begin()+i); 1119 break; 1120 } 1121 1122 // If this is a negative version of this factor, remove it. 1123 if (ConstantInt *FC1 = dyn_cast<ConstantInt>(Factor)) { 1124 if (ConstantInt *FC2 = dyn_cast<ConstantInt>(Factors[i].Op)) 1125 if (FC1->getValue() == -FC2->getValue()) { 1126 FoundFactor = NeedsNegate = true; 1127 Factors.erase(Factors.begin()+i); 1128 break; 1129 } 1130 } else if (ConstantFP *FC1 = dyn_cast<ConstantFP>(Factor)) { 1131 if (ConstantFP *FC2 = dyn_cast<ConstantFP>(Factors[i].Op)) { 1132 APFloat F1(FC1->getValueAPF()); 1133 APFloat F2(FC2->getValueAPF()); 1134 F2.changeSign(); 1135 if (F1.compare(F2) == APFloat::cmpEqual) { 1136 FoundFactor = NeedsNegate = true; 1137 Factors.erase(Factors.begin() + i); 1138 break; 1139 } 1140 } 1141 } 1142 } 1143 1144 if (!FoundFactor) { 1145 // Make sure to restore the operands to the expression tree. 1146 RewriteExprTree(BO, Factors); 1147 return nullptr; 1148 } 1149 1150 BasicBlock::iterator InsertPt = ++BO->getIterator(); 1151 1152 // If this was just a single multiply, remove the multiply and return the only 1153 // remaining operand. 1154 if (Factors.size() == 1) { 1155 RedoInsts.insert(BO); 1156 V = Factors[0].Op; 1157 } else { 1158 RewriteExprTree(BO, Factors); 1159 V = BO; 1160 } 1161 1162 if (NeedsNegate) 1163 V = CreateNeg(V, "neg", &*InsertPt, BO); 1164 1165 return V; 1166} 1167 1168/// If V is a single-use multiply, recursively add its operands as factors, 1169/// otherwise add V to the list of factors. 1170/// 1171/// Ops is the top-level list of add operands we're trying to factor. 1172static void FindSingleUseMultiplyFactors(Value *V, 1173 SmallVectorImpl<Value*> &Factors, 1174 const SmallVectorImpl<ValueEntry> &Ops) { 1175 BinaryOperator *BO = isReassociableOp(V, Instruction::Mul, Instruction::FMul); 1176 if (!BO) { 1177 Factors.push_back(V); 1178 return; 1179 } 1180 1181 // Otherwise, add the LHS and RHS to the list of factors. 1182 FindSingleUseMultiplyFactors(BO->getOperand(1), Factors, Ops); 1183 FindSingleUseMultiplyFactors(BO->getOperand(0), Factors, Ops); 1184} 1185 1186/// Optimize a series of operands to an 'and', 'or', or 'xor' instruction. 1187/// This optimizes based on identities. If it can be reduced to a single Value, 1188/// it is returned, otherwise the Ops list is mutated as necessary. 1189static Value *OptimizeAndOrXor(unsigned Opcode, 1190 SmallVectorImpl<ValueEntry> &Ops) { 1191 // Scan the operand lists looking for X and ~X pairs, along with X,X pairs. 1192 // If we find any, we can simplify the expression. X&~X == 0, X|~X == -1. 1193 for (unsigned i = 0, e = Ops.size(); i != e; ++i) { 1194 // First, check for X and ~X in the operand list. 1195 assert(i < Ops.size()); 1196 if (BinaryOperator::isNot(Ops[i].Op)) { // Cannot occur for ^. 1197 Value *X = BinaryOperator::getNotArgument(Ops[i].Op); 1198 unsigned FoundX = FindInOperandList(Ops, i, X); 1199 if (FoundX != i) { 1200 if (Opcode == Instruction::And) // ...&X&~X = 0 1201 return Constant::getNullValue(X->getType()); 1202 1203 if (Opcode == Instruction::Or) // ...|X|~X = -1 1204 return Constant::getAllOnesValue(X->getType()); 1205 } 1206 } 1207 1208 // Next, check for duplicate pairs of values, which we assume are next to 1209 // each other, due to our sorting criteria. 1210 assert(i < Ops.size()); 1211 if (i+1 != Ops.size() && Ops[i+1].Op == Ops[i].Op) { 1212 if (Opcode == Instruction::And || Opcode == Instruction::Or) { 1213 // Drop duplicate values for And and Or. 1214 Ops.erase(Ops.begin()+i); 1215 --i; --e; 1216 ++NumAnnihil; 1217 continue; 1218 } 1219 1220 // Drop pairs of values for Xor. 1221 assert(Opcode == Instruction::Xor); 1222 if (e == 2) 1223 return Constant::getNullValue(Ops[0].Op->getType()); 1224 1225 // Y ^ X^X -> Y 1226 Ops.erase(Ops.begin()+i, Ops.begin()+i+2); 1227 i -= 1; e -= 2; 1228 ++NumAnnihil; 1229 } 1230 } 1231 return nullptr; 1232} 1233 1234/// Helper function of CombineXorOpnd(). It creates a bitwise-and 1235/// instruction with the given two operands, and return the resulting 1236/// instruction. There are two special cases: 1) if the constant operand is 0, 1237/// it will return NULL. 2) if the constant is ~0, the symbolic operand will 1238/// be returned. 1239static Value *createAndInstr(Instruction *InsertBefore, Value *Opnd, 1240 const APInt &ConstOpnd) { 1241 if (ConstOpnd != 0) { 1242 if (!ConstOpnd.isAllOnesValue()) { 1243 LLVMContext &Ctx = Opnd->getType()->getContext(); 1244 Instruction *I; 1245 I = BinaryOperator::CreateAnd(Opnd, ConstantInt::get(Ctx, ConstOpnd), 1246 "and.ra", InsertBefore); 1247 I->setDebugLoc(InsertBefore->getDebugLoc()); 1248 return I; 1249 } 1250 return Opnd; 1251 } 1252 return nullptr; 1253} 1254 1255// Helper function of OptimizeXor(). It tries to simplify "Opnd1 ^ ConstOpnd" 1256// into "R ^ C", where C would be 0, and R is a symbolic value. 1257// 1258// If it was successful, true is returned, and the "R" and "C" is returned 1259// via "Res" and "ConstOpnd", respectively; otherwise, false is returned, 1260// and both "Res" and "ConstOpnd" remain unchanged. 1261// 1262bool Reassociate::CombineXorOpnd(Instruction *I, XorOpnd *Opnd1, 1263 APInt &ConstOpnd, Value *&Res) { 1264 // Xor-Rule 1: (x | c1) ^ c2 = (x | c1) ^ (c1 ^ c1) ^ c2 1265 // = ((x | c1) ^ c1) ^ (c1 ^ c2) 1266 // = (x & ~c1) ^ (c1 ^ c2) 1267 // It is useful only when c1 == c2. 1268 if (Opnd1->isOrExpr() && Opnd1->getConstPart() != 0) { 1269 if (!Opnd1->getValue()->hasOneUse()) 1270 return false; 1271 1272 const APInt &C1 = Opnd1->getConstPart(); 1273 if (C1 != ConstOpnd) 1274 return false; 1275 1276 Value *X = Opnd1->getSymbolicPart(); 1277 Res = createAndInstr(I, X, ~C1); 1278 // ConstOpnd was C2, now C1 ^ C2. 1279 ConstOpnd ^= C1; 1280 1281 if (Instruction *T = dyn_cast<Instruction>(Opnd1->getValue())) 1282 RedoInsts.insert(T); 1283 return true; 1284 } 1285 return false; 1286} 1287 1288 1289// Helper function of OptimizeXor(). It tries to simplify 1290// "Opnd1 ^ Opnd2 ^ ConstOpnd" into "R ^ C", where C would be 0, and R is a 1291// symbolic value. 1292// 1293// If it was successful, true is returned, and the "R" and "C" is returned 1294// via "Res" and "ConstOpnd", respectively (If the entire expression is 1295// evaluated to a constant, the Res is set to NULL); otherwise, false is 1296// returned, and both "Res" and "ConstOpnd" remain unchanged. 1297bool Reassociate::CombineXorOpnd(Instruction *I, XorOpnd *Opnd1, XorOpnd *Opnd2, 1298 APInt &ConstOpnd, Value *&Res) { 1299 Value *X = Opnd1->getSymbolicPart(); 1300 if (X != Opnd2->getSymbolicPart()) 1301 return false; 1302 1303 // This many instruction become dead.(At least "Opnd1 ^ Opnd2" will die.) 1304 int DeadInstNum = 1; 1305 if (Opnd1->getValue()->hasOneUse()) 1306 DeadInstNum++; 1307 if (Opnd2->getValue()->hasOneUse()) 1308 DeadInstNum++; 1309 1310 // Xor-Rule 2: 1311 // (x | c1) ^ (x & c2) 1312 // = (x|c1) ^ (x&c2) ^ (c1 ^ c1) = ((x|c1) ^ c1) ^ (x & c2) ^ c1 1313 // = (x & ~c1) ^ (x & c2) ^ c1 // Xor-Rule 1 1314 // = (x & c3) ^ c1, where c3 = ~c1 ^ c2 // Xor-rule 3 1315 // 1316 if (Opnd1->isOrExpr() != Opnd2->isOrExpr()) { 1317 if (Opnd2->isOrExpr()) 1318 std::swap(Opnd1, Opnd2); 1319 1320 const APInt &C1 = Opnd1->getConstPart(); 1321 const APInt &C2 = Opnd2->getConstPart(); 1322 APInt C3((~C1) ^ C2); 1323 1324 // Do not increase code size! 1325 if (C3 != 0 && !C3.isAllOnesValue()) { 1326 int NewInstNum = ConstOpnd != 0 ? 1 : 2; 1327 if (NewInstNum > DeadInstNum) 1328 return false; 1329 } 1330 1331 Res = createAndInstr(I, X, C3); 1332 ConstOpnd ^= C1; 1333 1334 } else if (Opnd1->isOrExpr()) { 1335 // Xor-Rule 3: (x | c1) ^ (x | c2) = (x & c3) ^ c3 where c3 = c1 ^ c2 1336 // 1337 const APInt &C1 = Opnd1->getConstPart(); 1338 const APInt &C2 = Opnd2->getConstPart(); 1339 APInt C3 = C1 ^ C2; 1340 1341 // Do not increase code size 1342 if (C3 != 0 && !C3.isAllOnesValue()) { 1343 int NewInstNum = ConstOpnd != 0 ? 1 : 2; 1344 if (NewInstNum > DeadInstNum) 1345 return false; 1346 } 1347 1348 Res = createAndInstr(I, X, C3); 1349 ConstOpnd ^= C3; 1350 } else { 1351 // Xor-Rule 4: (x & c1) ^ (x & c2) = (x & (c1^c2)) 1352 // 1353 const APInt &C1 = Opnd1->getConstPart(); 1354 const APInt &C2 = Opnd2->getConstPart(); 1355 APInt C3 = C1 ^ C2; 1356 Res = createAndInstr(I, X, C3); 1357 } 1358 1359 // Put the original operands in the Redo list; hope they will be deleted 1360 // as dead code. 1361 if (Instruction *T = dyn_cast<Instruction>(Opnd1->getValue())) 1362 RedoInsts.insert(T); 1363 if (Instruction *T = dyn_cast<Instruction>(Opnd2->getValue())) 1364 RedoInsts.insert(T); 1365 1366 return true; 1367} 1368 1369/// Optimize a series of operands to an 'xor' instruction. If it can be reduced 1370/// to a single Value, it is returned, otherwise the Ops list is mutated as 1371/// necessary. 1372Value *Reassociate::OptimizeXor(Instruction *I, 1373 SmallVectorImpl<ValueEntry> &Ops) { 1374 if (Value *V = OptimizeAndOrXor(Instruction::Xor, Ops)) 1375 return V; 1376 1377 if (Ops.size() == 1) 1378 return nullptr; 1379 1380 SmallVector<XorOpnd, 8> Opnds; 1381 SmallVector<XorOpnd*, 8> OpndPtrs; 1382 Type *Ty = Ops[0].Op->getType(); 1383 APInt ConstOpnd(Ty->getIntegerBitWidth(), 0); 1384 1385 // Step 1: Convert ValueEntry to XorOpnd 1386 for (unsigned i = 0, e = Ops.size(); i != e; ++i) { 1387 Value *V = Ops[i].Op; 1388 if (!isa<ConstantInt>(V)) { 1389 XorOpnd O(V); 1390 O.setSymbolicRank(getRank(O.getSymbolicPart())); 1391 Opnds.push_back(O); 1392 } else 1393 ConstOpnd ^= cast<ConstantInt>(V)->getValue(); 1394 } 1395 1396 // NOTE: From this point on, do *NOT* add/delete element to/from "Opnds". 1397 // It would otherwise invalidate the "Opnds"'s iterator, and hence invalidate 1398 // the "OpndPtrs" as well. For the similar reason, do not fuse this loop 1399 // with the previous loop --- the iterator of the "Opnds" may be invalidated 1400 // when new elements are added to the vector. 1401 for (unsigned i = 0, e = Opnds.size(); i != e; ++i) 1402 OpndPtrs.push_back(&Opnds[i]); 1403 1404 // Step 2: Sort the Xor-Operands in a way such that the operands containing 1405 // the same symbolic value cluster together. For instance, the input operand 1406 // sequence ("x | 123", "y & 456", "x & 789") will be sorted into: 1407 // ("x | 123", "x & 789", "y & 456"). 1408 std::stable_sort(OpndPtrs.begin(), OpndPtrs.end(), XorOpnd::PtrSortFunctor()); 1409 1410 // Step 3: Combine adjacent operands 1411 XorOpnd *PrevOpnd = nullptr; 1412 bool Changed = false; 1413 for (unsigned i = 0, e = Opnds.size(); i < e; i++) { 1414 XorOpnd *CurrOpnd = OpndPtrs[i]; 1415 // The combined value 1416 Value *CV; 1417 1418 // Step 3.1: Try simplifying "CurrOpnd ^ ConstOpnd" 1419 if (ConstOpnd != 0 && CombineXorOpnd(I, CurrOpnd, ConstOpnd, CV)) { 1420 Changed = true; 1421 if (CV) 1422 *CurrOpnd = XorOpnd(CV); 1423 else { 1424 CurrOpnd->Invalidate(); 1425 continue; 1426 } 1427 } 1428 1429 if (!PrevOpnd || CurrOpnd->getSymbolicPart() != PrevOpnd->getSymbolicPart()) { 1430 PrevOpnd = CurrOpnd; 1431 continue; 1432 } 1433 1434 // step 3.2: When previous and current operands share the same symbolic 1435 // value, try to simplify "PrevOpnd ^ CurrOpnd ^ ConstOpnd" 1436 // 1437 if (CombineXorOpnd(I, CurrOpnd, PrevOpnd, ConstOpnd, CV)) { 1438 // Remove previous operand 1439 PrevOpnd->Invalidate(); 1440 if (CV) { 1441 *CurrOpnd = XorOpnd(CV); 1442 PrevOpnd = CurrOpnd; 1443 } else { 1444 CurrOpnd->Invalidate(); 1445 PrevOpnd = nullptr; 1446 } 1447 Changed = true; 1448 } 1449 } 1450 1451 // Step 4: Reassemble the Ops 1452 if (Changed) { 1453 Ops.clear(); 1454 for (unsigned int i = 0, e = Opnds.size(); i < e; i++) { 1455 XorOpnd &O = Opnds[i]; 1456 if (O.isInvalid()) 1457 continue; 1458 ValueEntry VE(getRank(O.getValue()), O.getValue()); 1459 Ops.push_back(VE); 1460 } 1461 if (ConstOpnd != 0) { 1462 Value *C = ConstantInt::get(Ty->getContext(), ConstOpnd); 1463 ValueEntry VE(getRank(C), C); 1464 Ops.push_back(VE); 1465 } 1466 int Sz = Ops.size(); 1467 if (Sz == 1) 1468 return Ops.back().Op; 1469 else if (Sz == 0) { 1470 assert(ConstOpnd == 0); 1471 return ConstantInt::get(Ty->getContext(), ConstOpnd); 1472 } 1473 } 1474 1475 return nullptr; 1476} 1477 1478/// Optimize a series of operands to an 'add' instruction. This 1479/// optimizes based on identities. If it can be reduced to a single Value, it 1480/// is returned, otherwise the Ops list is mutated as necessary. 1481Value *Reassociate::OptimizeAdd(Instruction *I, 1482 SmallVectorImpl<ValueEntry> &Ops) { 1483 // Scan the operand lists looking for X and -X pairs. If we find any, we 1484 // can simplify expressions like X+-X == 0 and X+~X ==-1. While we're at it, 1485 // scan for any 1486 // duplicates. We want to canonicalize Y+Y+Y+Z -> 3*Y+Z. 1487 1488 for (unsigned i = 0, e = Ops.size(); i != e; ++i) { 1489 Value *TheOp = Ops[i].Op; 1490 // Check to see if we've seen this operand before. If so, we factor all 1491 // instances of the operand together. Due to our sorting criteria, we know 1492 // that these need to be next to each other in the vector. 1493 if (i+1 != Ops.size() && Ops[i+1].Op == TheOp) { 1494 // Rescan the list, remove all instances of this operand from the expr. 1495 unsigned NumFound = 0; 1496 do { 1497 Ops.erase(Ops.begin()+i); 1498 ++NumFound; 1499 } while (i != Ops.size() && Ops[i].Op == TheOp); 1500 1501 DEBUG(dbgs() << "\nFACTORING [" << NumFound << "]: " << *TheOp << '\n'); 1502 ++NumFactor; 1503 1504 // Insert a new multiply. 1505 Type *Ty = TheOp->getType(); 1506 Constant *C = Ty->isIntOrIntVectorTy() ? 1507 ConstantInt::get(Ty, NumFound) : ConstantFP::get(Ty, NumFound); 1508 Instruction *Mul = CreateMul(TheOp, C, "factor", I, I); 1509 1510 // Now that we have inserted a multiply, optimize it. This allows us to 1511 // handle cases that require multiple factoring steps, such as this: 1512 // (X*2) + (X*2) + (X*2) -> (X*2)*3 -> X*6 1513 RedoInsts.insert(Mul); 1514 1515 // If every add operand was a duplicate, return the multiply. 1516 if (Ops.empty()) 1517 return Mul; 1518 1519 // Otherwise, we had some input that didn't have the dupe, such as 1520 // "A + A + B" -> "A*2 + B". Add the new multiply to the list of 1521 // things being added by this operation. 1522 Ops.insert(Ops.begin(), ValueEntry(getRank(Mul), Mul)); 1523 1524 --i; 1525 e = Ops.size(); 1526 continue; 1527 } 1528 1529 // Check for X and -X or X and ~X in the operand list. 1530 if (!BinaryOperator::isNeg(TheOp) && !BinaryOperator::isFNeg(TheOp) && 1531 !BinaryOperator::isNot(TheOp)) 1532 continue; 1533 1534 Value *X = nullptr; 1535 if (BinaryOperator::isNeg(TheOp) || BinaryOperator::isFNeg(TheOp)) 1536 X = BinaryOperator::getNegArgument(TheOp); 1537 else if (BinaryOperator::isNot(TheOp)) 1538 X = BinaryOperator::getNotArgument(TheOp); 1539 1540 unsigned FoundX = FindInOperandList(Ops, i, X); 1541 if (FoundX == i) 1542 continue; 1543 1544 // Remove X and -X from the operand list. 1545 if (Ops.size() == 2 && 1546 (BinaryOperator::isNeg(TheOp) || BinaryOperator::isFNeg(TheOp))) 1547 return Constant::getNullValue(X->getType()); 1548 1549 // Remove X and ~X from the operand list. 1550 if (Ops.size() == 2 && BinaryOperator::isNot(TheOp)) 1551 return Constant::getAllOnesValue(X->getType()); 1552 1553 Ops.erase(Ops.begin()+i); 1554 if (i < FoundX) 1555 --FoundX; 1556 else 1557 --i; // Need to back up an extra one. 1558 Ops.erase(Ops.begin()+FoundX); 1559 ++NumAnnihil; 1560 --i; // Revisit element. 1561 e -= 2; // Removed two elements. 1562 1563 // if X and ~X we append -1 to the operand list. 1564 if (BinaryOperator::isNot(TheOp)) { 1565 Value *V = Constant::getAllOnesValue(X->getType()); 1566 Ops.insert(Ops.end(), ValueEntry(getRank(V), V)); 1567 e += 1; 1568 } 1569 } 1570 1571 // Scan the operand list, checking to see if there are any common factors 1572 // between operands. Consider something like A*A+A*B*C+D. We would like to 1573 // reassociate this to A*(A+B*C)+D, which reduces the number of multiplies. 1574 // To efficiently find this, we count the number of times a factor occurs 1575 // for any ADD operands that are MULs. 1576 DenseMap<Value*, unsigned> FactorOccurrences; 1577 1578 // Keep track of each multiply we see, to avoid triggering on (X*4)+(X*4) 1579 // where they are actually the same multiply. 1580 unsigned MaxOcc = 0; 1581 Value *MaxOccVal = nullptr; 1582 for (unsigned i = 0, e = Ops.size(); i != e; ++i) { 1583 BinaryOperator *BOp = 1584 isReassociableOp(Ops[i].Op, Instruction::Mul, Instruction::FMul); 1585 if (!BOp) 1586 continue; 1587 1588 // Compute all of the factors of this added value. 1589 SmallVector<Value*, 8> Factors; 1590 FindSingleUseMultiplyFactors(BOp, Factors, Ops); 1591 assert(Factors.size() > 1 && "Bad linearize!"); 1592 1593 // Add one to FactorOccurrences for each unique factor in this op. 1594 SmallPtrSet<Value*, 8> Duplicates; 1595 for (unsigned i = 0, e = Factors.size(); i != e; ++i) { 1596 Value *Factor = Factors[i]; 1597 if (!Duplicates.insert(Factor).second) 1598 continue; 1599 1600 unsigned Occ = ++FactorOccurrences[Factor]; 1601 if (Occ > MaxOcc) { 1602 MaxOcc = Occ; 1603 MaxOccVal = Factor; 1604 } 1605 1606 // If Factor is a negative constant, add the negated value as a factor 1607 // because we can percolate the negate out. Watch for minint, which 1608 // cannot be positivified. 1609 if (ConstantInt *CI = dyn_cast<ConstantInt>(Factor)) { 1610 if (CI->isNegative() && !CI->isMinValue(true)) { 1611 Factor = ConstantInt::get(CI->getContext(), -CI->getValue()); 1612 assert(!Duplicates.count(Factor) && 1613 "Shouldn't have two constant factors, missed a canonicalize"); 1614 unsigned Occ = ++FactorOccurrences[Factor]; 1615 if (Occ > MaxOcc) { 1616 MaxOcc = Occ; 1617 MaxOccVal = Factor; 1618 } 1619 } 1620 } else if (ConstantFP *CF = dyn_cast<ConstantFP>(Factor)) { 1621 if (CF->isNegative()) { 1622 APFloat F(CF->getValueAPF()); 1623 F.changeSign(); 1624 Factor = ConstantFP::get(CF->getContext(), F); 1625 assert(!Duplicates.count(Factor) && 1626 "Shouldn't have two constant factors, missed a canonicalize"); 1627 unsigned Occ = ++FactorOccurrences[Factor]; 1628 if (Occ > MaxOcc) { 1629 MaxOcc = Occ; 1630 MaxOccVal = Factor; 1631 } 1632 } 1633 } 1634 } 1635 } 1636 1637 // If any factor occurred more than one time, we can pull it out. 1638 if (MaxOcc > 1) { 1639 DEBUG(dbgs() << "\nFACTORING [" << MaxOcc << "]: " << *MaxOccVal << '\n'); 1640 ++NumFactor; 1641 1642 // Create a new instruction that uses the MaxOccVal twice. If we don't do 1643 // this, we could otherwise run into situations where removing a factor 1644 // from an expression will drop a use of maxocc, and this can cause 1645 // RemoveFactorFromExpression on successive values to behave differently. 1646 Instruction *DummyInst = 1647 I->getType()->isIntOrIntVectorTy() 1648 ? BinaryOperator::CreateAdd(MaxOccVal, MaxOccVal) 1649 : BinaryOperator::CreateFAdd(MaxOccVal, MaxOccVal); 1650 1651 SmallVector<WeakVH, 4> NewMulOps; 1652 for (unsigned i = 0; i != Ops.size(); ++i) { 1653 // Only try to remove factors from expressions we're allowed to. 1654 BinaryOperator *BOp = 1655 isReassociableOp(Ops[i].Op, Instruction::Mul, Instruction::FMul); 1656 if (!BOp) 1657 continue; 1658 1659 if (Value *V = RemoveFactorFromExpression(Ops[i].Op, MaxOccVal)) { 1660 // The factorized operand may occur several times. Convert them all in 1661 // one fell swoop. 1662 for (unsigned j = Ops.size(); j != i;) { 1663 --j; 1664 if (Ops[j].Op == Ops[i].Op) { 1665 NewMulOps.push_back(V); 1666 Ops.erase(Ops.begin()+j); 1667 } 1668 } 1669 --i; 1670 } 1671 } 1672 1673 // No need for extra uses anymore. 1674 delete DummyInst; 1675 1676 unsigned NumAddedValues = NewMulOps.size(); 1677 Value *V = EmitAddTreeOfValues(I, NewMulOps); 1678 1679 // Now that we have inserted the add tree, optimize it. This allows us to 1680 // handle cases that require multiple factoring steps, such as this: 1681 // A*A*B + A*A*C --> A*(A*B+A*C) --> A*(A*(B+C)) 1682 assert(NumAddedValues > 1 && "Each occurrence should contribute a value"); 1683 (void)NumAddedValues; 1684 if (Instruction *VI = dyn_cast<Instruction>(V)) 1685 RedoInsts.insert(VI); 1686 1687 // Create the multiply. 1688 Instruction *V2 = CreateMul(V, MaxOccVal, "tmp", I, I); 1689 1690 // Rerun associate on the multiply in case the inner expression turned into 1691 // a multiply. We want to make sure that we keep things in canonical form. 1692 RedoInsts.insert(V2); 1693 1694 // If every add operand included the factor (e.g. "A*B + A*C"), then the 1695 // entire result expression is just the multiply "A*(B+C)". 1696 if (Ops.empty()) 1697 return V2; 1698 1699 // Otherwise, we had some input that didn't have the factor, such as 1700 // "A*B + A*C + D" -> "A*(B+C) + D". Add the new multiply to the list of 1701 // things being added by this operation. 1702 Ops.insert(Ops.begin(), ValueEntry(getRank(V2), V2)); 1703 } 1704 1705 return nullptr; 1706} 1707 1708/// \brief Build up a vector of value/power pairs factoring a product. 1709/// 1710/// Given a series of multiplication operands, build a vector of factors and 1711/// the powers each is raised to when forming the final product. Sort them in 1712/// the order of descending power. 1713/// 1714/// (x*x) -> [(x, 2)] 1715/// ((x*x)*x) -> [(x, 3)] 1716/// ((((x*y)*x)*y)*x) -> [(x, 3), (y, 2)] 1717/// 1718/// \returns Whether any factors have a power greater than one. 1719bool Reassociate::collectMultiplyFactors(SmallVectorImpl<ValueEntry> &Ops, 1720 SmallVectorImpl<Factor> &Factors) { 1721 // FIXME: Have Ops be (ValueEntry, Multiplicity) pairs, simplifying this. 1722 // Compute the sum of powers of simplifiable factors. 1723 unsigned FactorPowerSum = 0; 1724 for (unsigned Idx = 1, Size = Ops.size(); Idx < Size; ++Idx) { 1725 Value *Op = Ops[Idx-1].Op; 1726 1727 // Count the number of occurrences of this value. 1728 unsigned Count = 1; 1729 for (; Idx < Size && Ops[Idx].Op == Op; ++Idx) 1730 ++Count; 1731 // Track for simplification all factors which occur 2 or more times. 1732 if (Count > 1) 1733 FactorPowerSum += Count; 1734 } 1735 1736 // We can only simplify factors if the sum of the powers of our simplifiable 1737 // factors is 4 or higher. When that is the case, we will *always* have 1738 // a simplification. This is an important invariant to prevent cyclicly 1739 // trying to simplify already minimal formations. 1740 if (FactorPowerSum < 4) 1741 return false; 1742 1743 // Now gather the simplifiable factors, removing them from Ops. 1744 FactorPowerSum = 0; 1745 for (unsigned Idx = 1; Idx < Ops.size(); ++Idx) { 1746 Value *Op = Ops[Idx-1].Op; 1747 1748 // Count the number of occurrences of this value. 1749 unsigned Count = 1; 1750 for (; Idx < Ops.size() && Ops[Idx].Op == Op; ++Idx) 1751 ++Count; 1752 if (Count == 1) 1753 continue; 1754 // Move an even number of occurrences to Factors. 1755 Count &= ~1U; 1756 Idx -= Count; 1757 FactorPowerSum += Count; 1758 Factors.push_back(Factor(Op, Count)); 1759 Ops.erase(Ops.begin()+Idx, Ops.begin()+Idx+Count); 1760 } 1761 1762 // None of the adjustments above should have reduced the sum of factor powers 1763 // below our mininum of '4'. 1764 assert(FactorPowerSum >= 4); 1765 1766 std::stable_sort(Factors.begin(), Factors.end(), Factor::PowerDescendingSorter()); 1767 return true; 1768} 1769 1770/// \brief Build a tree of multiplies, computing the product of Ops. 1771static Value *buildMultiplyTree(IRBuilder<> &Builder, 1772 SmallVectorImpl<Value*> &Ops) { 1773 if (Ops.size() == 1) 1774 return Ops.back(); 1775 1776 Value *LHS = Ops.pop_back_val(); 1777 do { 1778 if (LHS->getType()->isIntOrIntVectorTy()) 1779 LHS = Builder.CreateMul(LHS, Ops.pop_back_val()); 1780 else 1781 LHS = Builder.CreateFMul(LHS, Ops.pop_back_val()); 1782 } while (!Ops.empty()); 1783 1784 return LHS; 1785} 1786 1787/// \brief Build a minimal multiplication DAG for (a^x)*(b^y)*(c^z)*... 1788/// 1789/// Given a vector of values raised to various powers, where no two values are 1790/// equal and the powers are sorted in decreasing order, compute the minimal 1791/// DAG of multiplies to compute the final product, and return that product 1792/// value. 1793Value *Reassociate::buildMinimalMultiplyDAG(IRBuilder<> &Builder, 1794 SmallVectorImpl<Factor> &Factors) { 1795 assert(Factors[0].Power); 1796 SmallVector<Value *, 4> OuterProduct; 1797 for (unsigned LastIdx = 0, Idx = 1, Size = Factors.size(); 1798 Idx < Size && Factors[Idx].Power > 0; ++Idx) { 1799 if (Factors[Idx].Power != Factors[LastIdx].Power) { 1800 LastIdx = Idx; 1801 continue; 1802 } 1803 1804 // We want to multiply across all the factors with the same power so that 1805 // we can raise them to that power as a single entity. Build a mini tree 1806 // for that. 1807 SmallVector<Value *, 4> InnerProduct; 1808 InnerProduct.push_back(Factors[LastIdx].Base); 1809 do { 1810 InnerProduct.push_back(Factors[Idx].Base); 1811 ++Idx; 1812 } while (Idx < Size && Factors[Idx].Power == Factors[LastIdx].Power); 1813 1814 // Reset the base value of the first factor to the new expression tree. 1815 // We'll remove all the factors with the same power in a second pass. 1816 Value *M = Factors[LastIdx].Base = buildMultiplyTree(Builder, InnerProduct); 1817 if (Instruction *MI = dyn_cast<Instruction>(M)) 1818 RedoInsts.insert(MI); 1819 1820 LastIdx = Idx; 1821 } 1822 // Unique factors with equal powers -- we've folded them into the first one's 1823 // base. 1824 Factors.erase(std::unique(Factors.begin(), Factors.end(), 1825 Factor::PowerEqual()), 1826 Factors.end()); 1827 1828 // Iteratively collect the base of each factor with an add power into the 1829 // outer product, and halve each power in preparation for squaring the 1830 // expression. 1831 for (unsigned Idx = 0, Size = Factors.size(); Idx != Size; ++Idx) { 1832 if (Factors[Idx].Power & 1) 1833 OuterProduct.push_back(Factors[Idx].Base); 1834 Factors[Idx].Power >>= 1; 1835 } 1836 if (Factors[0].Power) { 1837 Value *SquareRoot = buildMinimalMultiplyDAG(Builder, Factors); 1838 OuterProduct.push_back(SquareRoot); 1839 OuterProduct.push_back(SquareRoot); 1840 } 1841 if (OuterProduct.size() == 1) 1842 return OuterProduct.front(); 1843 1844 Value *V = buildMultiplyTree(Builder, OuterProduct); 1845 return V; 1846} 1847 1848Value *Reassociate::OptimizeMul(BinaryOperator *I, 1849 SmallVectorImpl<ValueEntry> &Ops) { 1850 // We can only optimize the multiplies when there is a chain of more than 1851 // three, such that a balanced tree might require fewer total multiplies. 1852 if (Ops.size() < 4) 1853 return nullptr; 1854 1855 // Try to turn linear trees of multiplies without other uses of the 1856 // intermediate stages into minimal multiply DAGs with perfect sub-expression 1857 // re-use. 1858 SmallVector<Factor, 4> Factors; 1859 if (!collectMultiplyFactors(Ops, Factors)) 1860 return nullptr; // All distinct factors, so nothing left for us to do. 1861 1862 IRBuilder<> Builder(I); 1863 Value *V = buildMinimalMultiplyDAG(Builder, Factors); 1864 if (Ops.empty()) 1865 return V; 1866 1867 ValueEntry NewEntry = ValueEntry(getRank(V), V); 1868 Ops.insert(std::lower_bound(Ops.begin(), Ops.end(), NewEntry), NewEntry); 1869 return nullptr; 1870} 1871 1872Value *Reassociate::OptimizeExpression(BinaryOperator *I, 1873 SmallVectorImpl<ValueEntry> &Ops) { 1874 // Now that we have the linearized expression tree, try to optimize it. 1875 // Start by folding any constants that we found. 1876 Constant *Cst = nullptr; 1877 unsigned Opcode = I->getOpcode(); 1878 while (!Ops.empty() && isa<Constant>(Ops.back().Op)) { 1879 Constant *C = cast<Constant>(Ops.pop_back_val().Op); 1880 Cst = Cst ? ConstantExpr::get(Opcode, C, Cst) : C; 1881 } 1882 // If there was nothing but constants then we are done. 1883 if (Ops.empty()) 1884 return Cst; 1885 1886 // Put the combined constant back at the end of the operand list, except if 1887 // there is no point. For example, an add of 0 gets dropped here, while a 1888 // multiplication by zero turns the whole expression into zero. 1889 if (Cst && Cst != ConstantExpr::getBinOpIdentity(Opcode, I->getType())) { 1890 if (Cst == ConstantExpr::getBinOpAbsorber(Opcode, I->getType())) 1891 return Cst; 1892 Ops.push_back(ValueEntry(0, Cst)); 1893 } 1894 1895 if (Ops.size() == 1) return Ops[0].Op; 1896 1897 // Handle destructive annihilation due to identities between elements in the 1898 // argument list here. 1899 unsigned NumOps = Ops.size(); 1900 switch (Opcode) { 1901 default: break; 1902 case Instruction::And: 1903 case Instruction::Or: 1904 if (Value *Result = OptimizeAndOrXor(Opcode, Ops)) 1905 return Result; 1906 break; 1907 1908 case Instruction::Xor: 1909 if (Value *Result = OptimizeXor(I, Ops)) 1910 return Result; 1911 break; 1912 1913 case Instruction::Add: 1914 case Instruction::FAdd: 1915 if (Value *Result = OptimizeAdd(I, Ops)) 1916 return Result; 1917 break; 1918 1919 case Instruction::Mul: 1920 case Instruction::FMul: 1921 if (Value *Result = OptimizeMul(I, Ops)) 1922 return Result; 1923 break; 1924 } 1925 1926 if (Ops.size() != NumOps) 1927 return OptimizeExpression(I, Ops); 1928 return nullptr; 1929} 1930 1931// Remove dead instructions and if any operands are trivially dead add them to 1932// Insts so they will be removed as well. 1933void Reassociate::RecursivelyEraseDeadInsts( 1934 Instruction *I, SetVector<AssertingVH<Instruction>> &Insts) { 1935 assert(isInstructionTriviallyDead(I) && "Trivially dead instructions only!"); 1936 SmallVector<Value *, 4> Ops(I->op_begin(), I->op_end()); 1937 ValueRankMap.erase(I); 1938 Insts.remove(I); 1939 RedoInsts.remove(I); 1940 I->eraseFromParent(); 1941 for (auto Op : Ops) 1942 if (Instruction *OpInst = dyn_cast<Instruction>(Op)) 1943 if (OpInst->use_empty()) 1944 Insts.insert(OpInst); 1945} 1946 1947/// Zap the given instruction, adding interesting operands to the work list. 1948void Reassociate::EraseInst(Instruction *I) { 1949 assert(isInstructionTriviallyDead(I) && "Trivially dead instructions only!"); 1950 SmallVector<Value*, 8> Ops(I->op_begin(), I->op_end()); 1951 // Erase the dead instruction. 1952 ValueRankMap.erase(I); 1953 RedoInsts.remove(I); 1954 I->eraseFromParent(); 1955 // Optimize its operands. 1956 SmallPtrSet<Instruction *, 8> Visited; // Detect self-referential nodes. 1957 for (unsigned i = 0, e = Ops.size(); i != e; ++i) 1958 if (Instruction *Op = dyn_cast<Instruction>(Ops[i])) { 1959 // If this is a node in an expression tree, climb to the expression root 1960 // and add that since that's where optimization actually happens. 1961 unsigned Opcode = Op->getOpcode(); 1962 while (Op->hasOneUse() && Op->user_back()->getOpcode() == Opcode && 1963 Visited.insert(Op).second) 1964 Op = Op->user_back(); 1965 RedoInsts.insert(Op); 1966 } 1967} 1968 1969// Canonicalize expressions of the following form: 1970// x + (-Constant * y) -> x - (Constant * y) 1971// x - (-Constant * y) -> x + (Constant * y) 1972Instruction *Reassociate::canonicalizeNegConstExpr(Instruction *I) { 1973 if (!I->hasOneUse() || I->getType()->isVectorTy()) 1974 return nullptr; 1975 1976 // Must be a fmul or fdiv instruction. 1977 unsigned Opcode = I->getOpcode(); 1978 if (Opcode != Instruction::FMul && Opcode != Instruction::FDiv) 1979 return nullptr; 1980 1981 auto *C0 = dyn_cast<ConstantFP>(I->getOperand(0)); 1982 auto *C1 = dyn_cast<ConstantFP>(I->getOperand(1)); 1983 1984 // Both operands are constant, let it get constant folded away. 1985 if (C0 && C1) 1986 return nullptr; 1987 1988 ConstantFP *CF = C0 ? C0 : C1; 1989 1990 // Must have one constant operand. 1991 if (!CF) 1992 return nullptr; 1993 1994 // Must be a negative ConstantFP. 1995 if (!CF->isNegative()) 1996 return nullptr; 1997 1998 // User must be a binary operator with one or more uses. 1999 Instruction *User = I->user_back(); 2000 if (!isa<BinaryOperator>(User) || !User->hasNUsesOrMore(1)) 2001 return nullptr; 2002 2003 unsigned UserOpcode = User->getOpcode(); 2004 if (UserOpcode != Instruction::FAdd && UserOpcode != Instruction::FSub) 2005 return nullptr; 2006 2007 // Subtraction is not commutative. Explicitly, the following transform is 2008 // not valid: (-Constant * y) - x -> x + (Constant * y) 2009 if (!User->isCommutative() && User->getOperand(1) != I) 2010 return nullptr; 2011 2012 // Change the sign of the constant. 2013 APFloat Val = CF->getValueAPF(); 2014 Val.changeSign(); 2015 I->setOperand(C0 ? 0 : 1, ConstantFP::get(CF->getContext(), Val)); 2016 2017 // Canonicalize I to RHS to simplify the next bit of logic. E.g., 2018 // ((-Const*y) + x) -> (x + (-Const*y)). 2019 if (User->getOperand(0) == I && User->isCommutative()) 2020 cast<BinaryOperator>(User)->swapOperands(); 2021 2022 Value *Op0 = User->getOperand(0); 2023 Value *Op1 = User->getOperand(1); 2024 BinaryOperator *NI; 2025 switch (UserOpcode) { 2026 default: 2027 llvm_unreachable("Unexpected Opcode!"); 2028 case Instruction::FAdd: 2029 NI = BinaryOperator::CreateFSub(Op0, Op1); 2030 NI->setFastMathFlags(cast<FPMathOperator>(User)->getFastMathFlags()); 2031 break; 2032 case Instruction::FSub: 2033 NI = BinaryOperator::CreateFAdd(Op0, Op1); 2034 NI->setFastMathFlags(cast<FPMathOperator>(User)->getFastMathFlags()); 2035 break; 2036 } 2037 2038 NI->insertBefore(User); 2039 NI->setName(User->getName()); 2040 User->replaceAllUsesWith(NI); 2041 NI->setDebugLoc(I->getDebugLoc()); 2042 RedoInsts.insert(I); 2043 MadeChange = true; 2044 return NI; 2045} 2046 2047/// Inspect and optimize the given instruction. Note that erasing 2048/// instructions is not allowed. 2049void Reassociate::OptimizeInst(Instruction *I) { 2050 // Only consider operations that we understand. 2051 if (!isa<BinaryOperator>(I)) 2052 return; 2053 2054 if (I->getOpcode() == Instruction::Shl && isa<ConstantInt>(I->getOperand(1))) 2055 // If an operand of this shift is a reassociable multiply, or if the shift 2056 // is used by a reassociable multiply or add, turn into a multiply. 2057 if (isReassociableOp(I->getOperand(0), Instruction::Mul) || 2058 (I->hasOneUse() && 2059 (isReassociableOp(I->user_back(), Instruction::Mul) || 2060 isReassociableOp(I->user_back(), Instruction::Add)))) { 2061 Instruction *NI = ConvertShiftToMul(I); 2062 RedoInsts.insert(I); 2063 MadeChange = true; 2064 I = NI; 2065 } 2066 2067 // Canonicalize negative constants out of expressions. 2068 if (Instruction *Res = canonicalizeNegConstExpr(I)) 2069 I = Res; 2070 2071 // Commute binary operators, to canonicalize the order of their operands. 2072 // This can potentially expose more CSE opportunities, and makes writing other 2073 // transformations simpler. 2074 if (I->isCommutative()) 2075 canonicalizeOperands(I); 2076 2077 // TODO: We should optimize vector Xor instructions, but they are 2078 // currently unsupported. 2079 if (I->getType()->isVectorTy() && I->getOpcode() == Instruction::Xor) 2080 return; 2081 2082 // Don't optimize floating point instructions that don't have unsafe algebra. 2083 if (I->getType()->isFPOrFPVectorTy() && !I->hasUnsafeAlgebra()) 2084 return; 2085 2086 // Do not reassociate boolean (i1) expressions. We want to preserve the 2087 // original order of evaluation for short-circuited comparisons that 2088 // SimplifyCFG has folded to AND/OR expressions. If the expression 2089 // is not further optimized, it is likely to be transformed back to a 2090 // short-circuited form for code gen, and the source order may have been 2091 // optimized for the most likely conditions. 2092 if (I->getType()->isIntegerTy(1)) 2093 return; 2094 2095 // If this is a subtract instruction which is not already in negate form, 2096 // see if we can convert it to X+-Y. 2097 if (I->getOpcode() == Instruction::Sub) { 2098 if (ShouldBreakUpSubtract(I)) { 2099 Instruction *NI = BreakUpSubtract(I, RedoInsts); 2100 RedoInsts.insert(I); 2101 MadeChange = true; 2102 I = NI; 2103 } else if (BinaryOperator::isNeg(I)) { 2104 // Otherwise, this is a negation. See if the operand is a multiply tree 2105 // and if this is not an inner node of a multiply tree. 2106 if (isReassociableOp(I->getOperand(1), Instruction::Mul) && 2107 (!I->hasOneUse() || 2108 !isReassociableOp(I->user_back(), Instruction::Mul))) { 2109 Instruction *NI = LowerNegateToMultiply(I); 2110 // If the negate was simplified, revisit the users to see if we can 2111 // reassociate further. 2112 for (User *U : NI->users()) { 2113 if (BinaryOperator *Tmp = dyn_cast<BinaryOperator>(U)) 2114 RedoInsts.insert(Tmp); 2115 } 2116 RedoInsts.insert(I); 2117 MadeChange = true; 2118 I = NI; 2119 } 2120 } 2121 } else if (I->getOpcode() == Instruction::FSub) { 2122 if (ShouldBreakUpSubtract(I)) { 2123 Instruction *NI = BreakUpSubtract(I, RedoInsts); 2124 RedoInsts.insert(I); 2125 MadeChange = true; 2126 I = NI; 2127 } else if (BinaryOperator::isFNeg(I)) { 2128 // Otherwise, this is a negation. See if the operand is a multiply tree 2129 // and if this is not an inner node of a multiply tree. 2130 if (isReassociableOp(I->getOperand(1), Instruction::FMul) && 2131 (!I->hasOneUse() || 2132 !isReassociableOp(I->user_back(), Instruction::FMul))) { 2133 // If the negate was simplified, revisit the users to see if we can 2134 // reassociate further. 2135 Instruction *NI = LowerNegateToMultiply(I); 2136 for (User *U : NI->users()) { 2137 if (BinaryOperator *Tmp = dyn_cast<BinaryOperator>(U)) 2138 RedoInsts.insert(Tmp); 2139 } 2140 RedoInsts.insert(I); 2141 MadeChange = true; 2142 I = NI; 2143 } 2144 } 2145 } 2146 2147 // If this instruction is an associative binary operator, process it. 2148 if (!I->isAssociative()) return; 2149 BinaryOperator *BO = cast<BinaryOperator>(I); 2150 2151 // If this is an interior node of a reassociable tree, ignore it until we 2152 // get to the root of the tree, to avoid N^2 analysis. 2153 unsigned Opcode = BO->getOpcode(); 2154 if (BO->hasOneUse() && BO->user_back()->getOpcode() == Opcode) { 2155 // During the initial run we will get to the root of the tree. 2156 // But if we get here while we are redoing instructions, there is no 2157 // guarantee that the root will be visited. So Redo later 2158 if (BO->user_back() != BO && 2159 BO->getParent() == BO->user_back()->getParent()) 2160 RedoInsts.insert(BO->user_back()); 2161 return; 2162 } 2163 2164 // If this is an add tree that is used by a sub instruction, ignore it 2165 // until we process the subtract. 2166 if (BO->hasOneUse() && BO->getOpcode() == Instruction::Add && 2167 cast<Instruction>(BO->user_back())->getOpcode() == Instruction::Sub) 2168 return; 2169 if (BO->hasOneUse() && BO->getOpcode() == Instruction::FAdd && 2170 cast<Instruction>(BO->user_back())->getOpcode() == Instruction::FSub) 2171 return; 2172 2173 ReassociateExpression(BO); 2174} 2175 2176void Reassociate::ReassociateExpression(BinaryOperator *I) { 2177 // First, walk the expression tree, linearizing the tree, collecting the 2178 // operand information. 2179 SmallVector<RepeatedValue, 8> Tree; 2180 MadeChange |= LinearizeExprTree(I, Tree); 2181 SmallVector<ValueEntry, 8> Ops; 2182 Ops.reserve(Tree.size()); 2183 for (unsigned i = 0, e = Tree.size(); i != e; ++i) { 2184 RepeatedValue E = Tree[i]; 2185 Ops.append(E.second.getZExtValue(), 2186 ValueEntry(getRank(E.first), E.first)); 2187 } 2188 2189 DEBUG(dbgs() << "RAIn:\t"; PrintOps(I, Ops); dbgs() << '\n'); 2190 2191 // Now that we have linearized the tree to a list and have gathered all of 2192 // the operands and their ranks, sort the operands by their rank. Use a 2193 // stable_sort so that values with equal ranks will have their relative 2194 // positions maintained (and so the compiler is deterministic). Note that 2195 // this sorts so that the highest ranking values end up at the beginning of 2196 // the vector. 2197 std::stable_sort(Ops.begin(), Ops.end()); 2198 2199 // Now that we have the expression tree in a convenient 2200 // sorted form, optimize it globally if possible. 2201 if (Value *V = OptimizeExpression(I, Ops)) { 2202 if (V == I) 2203 // Self-referential expression in unreachable code. 2204 return; 2205 // This expression tree simplified to something that isn't a tree, 2206 // eliminate it. 2207 DEBUG(dbgs() << "Reassoc to scalar: " << *V << '\n'); 2208 I->replaceAllUsesWith(V); 2209 if (Instruction *VI = dyn_cast<Instruction>(V)) 2210 VI->setDebugLoc(I->getDebugLoc()); 2211 RedoInsts.insert(I); 2212 ++NumAnnihil; 2213 return; 2214 } 2215 2216 // We want to sink immediates as deeply as possible except in the case where 2217 // this is a multiply tree used only by an add, and the immediate is a -1. 2218 // In this case we reassociate to put the negation on the outside so that we 2219 // can fold the negation into the add: (-X)*Y + Z -> Z-X*Y 2220 if (I->hasOneUse()) { 2221 if (I->getOpcode() == Instruction::Mul && 2222 cast<Instruction>(I->user_back())->getOpcode() == Instruction::Add && 2223 isa<ConstantInt>(Ops.back().Op) && 2224 cast<ConstantInt>(Ops.back().Op)->isAllOnesValue()) { 2225 ValueEntry Tmp = Ops.pop_back_val(); 2226 Ops.insert(Ops.begin(), Tmp); 2227 } else if (I->getOpcode() == Instruction::FMul && 2228 cast<Instruction>(I->user_back())->getOpcode() == 2229 Instruction::FAdd && 2230 isa<ConstantFP>(Ops.back().Op) && 2231 cast<ConstantFP>(Ops.back().Op)->isExactlyValue(-1.0)) { 2232 ValueEntry Tmp = Ops.pop_back_val(); 2233 Ops.insert(Ops.begin(), Tmp); 2234 } 2235 } 2236 2237 DEBUG(dbgs() << "RAOut:\t"; PrintOps(I, Ops); dbgs() << '\n'); 2238 2239 if (Ops.size() == 1) { 2240 if (Ops[0].Op == I) 2241 // Self-referential expression in unreachable code. 2242 return; 2243 2244 // This expression tree simplified to something that isn't a tree, 2245 // eliminate it. 2246 I->replaceAllUsesWith(Ops[0].Op); 2247 if (Instruction *OI = dyn_cast<Instruction>(Ops[0].Op)) 2248 OI->setDebugLoc(I->getDebugLoc()); 2249 RedoInsts.insert(I); 2250 return; 2251 } 2252 2253 // Now that we ordered and optimized the expressions, splat them back into 2254 // the expression tree, removing any unneeded nodes. 2255 RewriteExprTree(I, Ops); 2256} 2257 2258bool Reassociate::runOnFunction(Function &F) { 2259 if (skipOptnoneFunction(F)) 2260 return false; 2261 2262 // Calculate the rank map for F 2263 BuildRankMap(F); 2264 2265 MadeChange = false; 2266 for (Function::iterator BI = F.begin(), BE = F.end(); BI != BE; ++BI) { 2267 // Optimize every instruction in the basic block. 2268 for (BasicBlock::iterator II = BI->begin(), IE = BI->end(); II != IE; ) 2269 if (isInstructionTriviallyDead(&*II)) { 2270 EraseInst(&*II++); 2271 } else { 2272 OptimizeInst(&*II); 2273 assert(II->getParent() == BI && "Moved to a different block!"); 2274 ++II; 2275 } 2276 2277 // Make a copy of all the instructions to be redone so we can remove dead 2278 // instructions. 2279 SetVector<AssertingVH<Instruction>> ToRedo(RedoInsts); 2280 // Iterate over all instructions to be reevaluated and remove trivially dead 2281 // instructions. If any operand of the trivially dead instruction becomes 2282 // dead mark it for deletion as well. Continue this process until all 2283 // trivially dead instructions have been removed. 2284 while (!ToRedo.empty()) { 2285 Instruction *I = ToRedo.pop_back_val(); 2286 if (isInstructionTriviallyDead(I)) 2287 RecursivelyEraseDeadInsts(I, ToRedo); 2288 } 2289 2290 // Now that we have removed dead instructions, we can reoptimize the 2291 // remaining instructions. 2292 while (!RedoInsts.empty()) { 2293 Instruction *I = RedoInsts.pop_back_val(); 2294 if (isInstructionTriviallyDead(I)) 2295 EraseInst(I); 2296 else 2297 OptimizeInst(I); 2298 } 2299 } 2300 2301 // We are done with the rank map. 2302 RankMap.clear(); 2303 ValueRankMap.clear(); 2304 2305 return MadeChange; 2306} 2307