Reassociate.cpp revision 0e5f499638c8d277b9dc4a4385712177c53b5681
1//===- Reassociate.cpp - Reassociate binary expressions -------------------===// 2// 3// The LLVM Compiler Infrastructure 4// 5// This file was developed by the LLVM research group and is distributed under 6// the University of Illinois Open Source 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... 12// 13// For example: 4 + (x + 5) -> x + (4 + 5) 14// 15// In the implementation of this algorithm, constants are assigned rank = 0, 16// function arguments are rank = 1, and other values are assigned ranks 17// corresponding to the reverse post order traversal of current function 18// (starting at 2), which effectively gives values in deep loops higher rank 19// than values not in loops. 20// 21//===----------------------------------------------------------------------===// 22 23#define DEBUG_TYPE "reassociate" 24#include "llvm/Transforms/Scalar.h" 25#include "llvm/Constants.h" 26#include "llvm/DerivedTypes.h" 27#include "llvm/Function.h" 28#include "llvm/Instructions.h" 29#include "llvm/Pass.h" 30#include "llvm/Assembly/Writer.h" 31#include "llvm/Support/CFG.h" 32#include "llvm/Support/Debug.h" 33#include "llvm/ADT/PostOrderIterator.h" 34#include "llvm/ADT/Statistic.h" 35#include <algorithm> 36using namespace llvm; 37 38STATISTIC(NumLinear , "Number of insts linearized"); 39STATISTIC(NumChanged, "Number of insts reassociated"); 40STATISTIC(NumAnnihil, "Number of expr tree annihilated"); 41STATISTIC(NumFactor , "Number of multiplies factored"); 42 43namespace { 44 struct ValueEntry { 45 unsigned Rank; 46 Value *Op; 47 ValueEntry(unsigned R, Value *O) : Rank(R), Op(O) {} 48 }; 49 inline bool operator<(const ValueEntry &LHS, const ValueEntry &RHS) { 50 return LHS.Rank > RHS.Rank; // Sort so that highest rank goes to start. 51 } 52} 53 54/// PrintOps - Print out the expression identified in the Ops list. 55/// 56static void PrintOps(Instruction *I, const std::vector<ValueEntry> &Ops) { 57 Module *M = I->getParent()->getParent()->getParent(); 58 cerr << Instruction::getOpcodeName(I->getOpcode()) << " " 59 << *Ops[0].Op->getType(); 60 for (unsigned i = 0, e = Ops.size(); i != e; ++i) 61 WriteAsOperand(*cerr.stream() << " ", Ops[i].Op, false, M) 62 << "," << Ops[i].Rank; 63} 64 65namespace { 66 class Reassociate : public FunctionPass { 67 std::map<BasicBlock*, unsigned> RankMap; 68 std::map<Value*, unsigned> ValueRankMap; 69 bool MadeChange; 70 public: 71 bool runOnFunction(Function &F); 72 73 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 74 AU.setPreservesCFG(); 75 } 76 private: 77 void BuildRankMap(Function &F); 78 unsigned getRank(Value *V); 79 void ReassociateExpression(BinaryOperator *I); 80 void RewriteExprTree(BinaryOperator *I, std::vector<ValueEntry> &Ops, 81 unsigned Idx = 0); 82 Value *OptimizeExpression(BinaryOperator *I, std::vector<ValueEntry> &Ops); 83 void LinearizeExprTree(BinaryOperator *I, std::vector<ValueEntry> &Ops); 84 void LinearizeExpr(BinaryOperator *I); 85 Value *RemoveFactorFromExpression(Value *V, Value *Factor); 86 void ReassociateBB(BasicBlock *BB); 87 88 void RemoveDeadBinaryOp(Value *V); 89 }; 90 91 RegisterPass<Reassociate> X("reassociate", "Reassociate expressions"); 92} 93 94// Public interface to the Reassociate pass 95FunctionPass *llvm::createReassociatePass() { return new Reassociate(); } 96 97void Reassociate::RemoveDeadBinaryOp(Value *V) { 98 BinaryOperator *BOp = dyn_cast<BinaryOperator>(V); 99 if (!BOp || !BOp->use_empty()) return; 100 101 Value *LHS = BOp->getOperand(0), *RHS = BOp->getOperand(1); 102 RemoveDeadBinaryOp(LHS); 103 RemoveDeadBinaryOp(RHS); 104} 105 106 107static bool isUnmovableInstruction(Instruction *I) { 108 if (I->getOpcode() == Instruction::PHI || 109 I->getOpcode() == Instruction::Alloca || 110 I->getOpcode() == Instruction::Load || 111 I->getOpcode() == Instruction::Malloc || 112 I->getOpcode() == Instruction::Invoke || 113 I->getOpcode() == Instruction::Call || 114 I->getOpcode() == Instruction::UDiv || 115 I->getOpcode() == Instruction::SDiv || 116 I->getOpcode() == Instruction::FDiv || 117 I->getOpcode() == Instruction::URem || 118 I->getOpcode() == Instruction::SRem || 119 I->getOpcode() == Instruction::FRem) 120 return true; 121 return false; 122} 123 124void Reassociate::BuildRankMap(Function &F) { 125 unsigned i = 2; 126 127 // Assign distinct ranks to function arguments 128 for (Function::arg_iterator I = F.arg_begin(), E = F.arg_end(); I != E; ++I) 129 ValueRankMap[I] = ++i; 130 131 ReversePostOrderTraversal<Function*> RPOT(&F); 132 for (ReversePostOrderTraversal<Function*>::rpo_iterator I = RPOT.begin(), 133 E = RPOT.end(); I != E; ++I) { 134 BasicBlock *BB = *I; 135 unsigned BBRank = RankMap[BB] = ++i << 16; 136 137 // Walk the basic block, adding precomputed ranks for any instructions that 138 // we cannot move. This ensures that the ranks for these instructions are 139 // all different in the block. 140 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) 141 if (isUnmovableInstruction(I)) 142 ValueRankMap[I] = ++BBRank; 143 } 144} 145 146unsigned Reassociate::getRank(Value *V) { 147 if (isa<Argument>(V)) return ValueRankMap[V]; // Function argument... 148 149 Instruction *I = dyn_cast<Instruction>(V); 150 if (I == 0) return 0; // Otherwise it's a global or constant, rank 0. 151 152 unsigned &CachedRank = ValueRankMap[I]; 153 if (CachedRank) return CachedRank; // Rank already known? 154 155 // If this is an expression, return the 1+MAX(rank(LHS), rank(RHS)) so that 156 // we can reassociate expressions for code motion! Since we do not recurse 157 // for PHI nodes, we cannot have infinite recursion here, because there 158 // cannot be loops in the value graph that do not go through PHI nodes. 159 unsigned Rank = 0, MaxRank = RankMap[I->getParent()]; 160 for (unsigned i = 0, e = I->getNumOperands(); 161 i != e && Rank != MaxRank; ++i) 162 Rank = std::max(Rank, getRank(I->getOperand(i))); 163 164 // If this is a not or neg instruction, do not count it for rank. This 165 // assures us that X and ~X will have the same rank. 166 if (!I->getType()->isIntegral() || 167 (!BinaryOperator::isNot(I) && !BinaryOperator::isNeg(I))) 168 ++Rank; 169 170 //DOUT << "Calculated Rank[" << V->getName() << "] = " 171 // << Rank << "\n"; 172 173 return CachedRank = Rank; 174} 175 176/// isReassociableOp - Return true if V is an instruction of the specified 177/// opcode and if it only has one use. 178static BinaryOperator *isReassociableOp(Value *V, unsigned Opcode) { 179 if ((V->hasOneUse() || V->use_empty()) && isa<Instruction>(V) && 180 cast<Instruction>(V)->getOpcode() == Opcode) 181 return cast<BinaryOperator>(V); 182 return 0; 183} 184 185/// LowerNegateToMultiply - Replace 0-X with X*-1. 186/// 187static Instruction *LowerNegateToMultiply(Instruction *Neg) { 188 Constant *Cst; 189 if (Neg->getType()->isFloatingPoint()) 190 Cst = ConstantFP::get(Neg->getType(), -1); 191 else 192 Cst = ConstantInt::getAllOnesValue(Neg->getType()); 193 194 std::string NegName = Neg->getName(); Neg->setName(""); 195 Instruction *Res = BinaryOperator::createMul(Neg->getOperand(1), Cst, NegName, 196 Neg); 197 Neg->replaceAllUsesWith(Res); 198 Neg->eraseFromParent(); 199 return Res; 200} 201 202// Given an expression of the form '(A+B)+(D+C)', turn it into '(((A+B)+C)+D)'. 203// Note that if D is also part of the expression tree that we recurse to 204// linearize it as well. Besides that case, this does not recurse into A,B, or 205// C. 206void Reassociate::LinearizeExpr(BinaryOperator *I) { 207 BinaryOperator *LHS = cast<BinaryOperator>(I->getOperand(0)); 208 BinaryOperator *RHS = cast<BinaryOperator>(I->getOperand(1)); 209 assert(isReassociableOp(LHS, I->getOpcode()) && 210 isReassociableOp(RHS, I->getOpcode()) && 211 "Not an expression that needs linearization?"); 212 213 DOUT << "Linear" << *LHS << *RHS << *I; 214 215 // Move the RHS instruction to live immediately before I, avoiding breaking 216 // dominator properties. 217 RHS->moveBefore(I); 218 219 // Move operands around to do the linearization. 220 I->setOperand(1, RHS->getOperand(0)); 221 RHS->setOperand(0, LHS); 222 I->setOperand(0, RHS); 223 224 ++NumLinear; 225 MadeChange = true; 226 DOUT << "Linearized: " << *I; 227 228 // If D is part of this expression tree, tail recurse. 229 if (isReassociableOp(I->getOperand(1), I->getOpcode())) 230 LinearizeExpr(I); 231} 232 233 234/// LinearizeExprTree - Given an associative binary expression tree, traverse 235/// all of the uses putting it into canonical form. This forces a left-linear 236/// form of the the expression (((a+b)+c)+d), and collects information about the 237/// rank of the non-tree operands. 238/// 239/// NOTE: These intentionally destroys the expression tree operands (turning 240/// them into undef values) to reduce #uses of the values. This means that the 241/// caller MUST use something like RewriteExprTree to put the values back in. 242/// 243void Reassociate::LinearizeExprTree(BinaryOperator *I, 244 std::vector<ValueEntry> &Ops) { 245 Value *LHS = I->getOperand(0), *RHS = I->getOperand(1); 246 unsigned Opcode = I->getOpcode(); 247 248 // First step, linearize the expression if it is in ((A+B)+(C+D)) form. 249 BinaryOperator *LHSBO = isReassociableOp(LHS, Opcode); 250 BinaryOperator *RHSBO = isReassociableOp(RHS, Opcode); 251 252 // If this is a multiply expression tree and it contains internal negations, 253 // transform them into multiplies by -1 so they can be reassociated. 254 if (I->getOpcode() == Instruction::Mul) { 255 if (!LHSBO && LHS->hasOneUse() && BinaryOperator::isNeg(LHS)) { 256 LHS = LowerNegateToMultiply(cast<Instruction>(LHS)); 257 LHSBO = isReassociableOp(LHS, Opcode); 258 } 259 if (!RHSBO && RHS->hasOneUse() && BinaryOperator::isNeg(RHS)) { 260 RHS = LowerNegateToMultiply(cast<Instruction>(RHS)); 261 RHSBO = isReassociableOp(RHS, Opcode); 262 } 263 } 264 265 if (!LHSBO) { 266 if (!RHSBO) { 267 // Neither the LHS or RHS as part of the tree, thus this is a leaf. As 268 // such, just remember these operands and their rank. 269 Ops.push_back(ValueEntry(getRank(LHS), LHS)); 270 Ops.push_back(ValueEntry(getRank(RHS), RHS)); 271 272 // Clear the leaves out. 273 I->setOperand(0, UndefValue::get(I->getType())); 274 I->setOperand(1, UndefValue::get(I->getType())); 275 return; 276 } else { 277 // Turn X+(Y+Z) -> (Y+Z)+X 278 std::swap(LHSBO, RHSBO); 279 std::swap(LHS, RHS); 280 bool Success = !I->swapOperands(); 281 assert(Success && "swapOperands failed"); 282 MadeChange = true; 283 } 284 } else if (RHSBO) { 285 // Turn (A+B)+(C+D) -> (((A+B)+C)+D). This guarantees the the RHS is not 286 // part of the expression tree. 287 LinearizeExpr(I); 288 LHS = LHSBO = cast<BinaryOperator>(I->getOperand(0)); 289 RHS = I->getOperand(1); 290 RHSBO = 0; 291 } 292 293 // Okay, now we know that the LHS is a nested expression and that the RHS is 294 // not. Perform reassociation. 295 assert(!isReassociableOp(RHS, Opcode) && "LinearizeExpr failed!"); 296 297 // Move LHS right before I to make sure that the tree expression dominates all 298 // values. 299 LHSBO->moveBefore(I); 300 301 // Linearize the expression tree on the LHS. 302 LinearizeExprTree(LHSBO, Ops); 303 304 // Remember the RHS operand and its rank. 305 Ops.push_back(ValueEntry(getRank(RHS), RHS)); 306 307 // Clear the RHS leaf out. 308 I->setOperand(1, UndefValue::get(I->getType())); 309} 310 311// RewriteExprTree - Now that the operands for this expression tree are 312// linearized and optimized, emit them in-order. This function is written to be 313// tail recursive. 314void Reassociate::RewriteExprTree(BinaryOperator *I, 315 std::vector<ValueEntry> &Ops, 316 unsigned i) { 317 if (i+2 == Ops.size()) { 318 if (I->getOperand(0) != Ops[i].Op || 319 I->getOperand(1) != Ops[i+1].Op) { 320 Value *OldLHS = I->getOperand(0); 321 DOUT << "RA: " << *I; 322 I->setOperand(0, Ops[i].Op); 323 I->setOperand(1, Ops[i+1].Op); 324 DOUT << "TO: " << *I; 325 MadeChange = true; 326 ++NumChanged; 327 328 // If we reassociated a tree to fewer operands (e.g. (1+a+2) -> (a+3) 329 // delete the extra, now dead, nodes. 330 RemoveDeadBinaryOp(OldLHS); 331 } 332 return; 333 } 334 assert(i+2 < Ops.size() && "Ops index out of range!"); 335 336 if (I->getOperand(1) != Ops[i].Op) { 337 DOUT << "RA: " << *I; 338 I->setOperand(1, Ops[i].Op); 339 DOUT << "TO: " << *I; 340 MadeChange = true; 341 ++NumChanged; 342 } 343 344 BinaryOperator *LHS = cast<BinaryOperator>(I->getOperand(0)); 345 assert(LHS->getOpcode() == I->getOpcode() && 346 "Improper expression tree!"); 347 348 // Compactify the tree instructions together with each other to guarantee 349 // that the expression tree is dominated by all of Ops. 350 LHS->moveBefore(I); 351 RewriteExprTree(LHS, Ops, i+1); 352} 353 354 355 356// NegateValue - Insert instructions before the instruction pointed to by BI, 357// that computes the negative version of the value specified. The negative 358// version of the value is returned, and BI is left pointing at the instruction 359// that should be processed next by the reassociation pass. 360// 361static Value *NegateValue(Value *V, Instruction *BI) { 362 // We are trying to expose opportunity for reassociation. One of the things 363 // that we want to do to achieve this is to push a negation as deep into an 364 // expression chain as possible, to expose the add instructions. In practice, 365 // this means that we turn this: 366 // X = -(A+12+C+D) into X = -A + -12 + -C + -D = -12 + -A + -C + -D 367 // so that later, a: Y = 12+X could get reassociated with the -12 to eliminate 368 // the constants. We assume that instcombine will clean up the mess later if 369 // we introduce tons of unnecessary negation instructions... 370 // 371 if (Instruction *I = dyn_cast<Instruction>(V)) 372 if (I->getOpcode() == Instruction::Add && I->hasOneUse()) { 373 // Push the negates through the add. 374 I->setOperand(0, NegateValue(I->getOperand(0), BI)); 375 I->setOperand(1, NegateValue(I->getOperand(1), BI)); 376 377 // We must move the add instruction here, because the neg instructions do 378 // not dominate the old add instruction in general. By moving it, we are 379 // assured that the neg instructions we just inserted dominate the 380 // instruction we are about to insert after them. 381 // 382 I->moveBefore(BI); 383 I->setName(I->getName()+".neg"); 384 return I; 385 } 386 387 // Insert a 'neg' instruction that subtracts the value from zero to get the 388 // negation. 389 // 390 return BinaryOperator::createNeg(V, V->getName() + ".neg", BI); 391} 392 393/// BreakUpSubtract - If we have (X-Y), and if either X is an add, or if this is 394/// only used by an add, transform this into (X+(0-Y)) to promote better 395/// reassociation. 396static Instruction *BreakUpSubtract(Instruction *Sub) { 397 // Don't bother to break this up unless either the LHS is an associable add or 398 // if this is only used by one. 399 if (!isReassociableOp(Sub->getOperand(0), Instruction::Add) && 400 !isReassociableOp(Sub->getOperand(1), Instruction::Add) && 401 !(Sub->hasOneUse() &&isReassociableOp(Sub->use_back(), Instruction::Add))) 402 return 0; 403 404 // Convert a subtract into an add and a neg instruction... so that sub 405 // instructions can be commuted with other add instructions... 406 // 407 // Calculate the negative value of Operand 1 of the sub instruction... 408 // and set it as the RHS of the add instruction we just made... 409 // 410 std::string Name = Sub->getName(); 411 Sub->setName(""); 412 Value *NegVal = NegateValue(Sub->getOperand(1), Sub); 413 Instruction *New = 414 BinaryOperator::createAdd(Sub->getOperand(0), NegVal, Name, Sub); 415 416 // Everyone now refers to the add instruction. 417 Sub->replaceAllUsesWith(New); 418 Sub->eraseFromParent(); 419 420 DOUT << "Negated: " << *New; 421 return New; 422} 423 424/// ConvertShiftToMul - If this is a shift of a reassociable multiply or is used 425/// by one, change this into a multiply by a constant to assist with further 426/// reassociation. 427static Instruction *ConvertShiftToMul(Instruction *Shl) { 428 // If an operand of this shift is a reassociable multiply, or if the shift 429 // is used by a reassociable multiply or add, turn into a multiply. 430 if (isReassociableOp(Shl->getOperand(0), Instruction::Mul) || 431 (Shl->hasOneUse() && 432 (isReassociableOp(Shl->use_back(), Instruction::Mul) || 433 isReassociableOp(Shl->use_back(), Instruction::Add)))) { 434 Constant *MulCst = ConstantInt::get(Shl->getType(), 1); 435 MulCst = ConstantExpr::getShl(MulCst, cast<Constant>(Shl->getOperand(1))); 436 437 std::string Name = Shl->getName(); Shl->setName(""); 438 Instruction *Mul = BinaryOperator::createMul(Shl->getOperand(0), MulCst, 439 Name, Shl); 440 Shl->replaceAllUsesWith(Mul); 441 Shl->eraseFromParent(); 442 return Mul; 443 } 444 return 0; 445} 446 447// Scan backwards and forwards among values with the same rank as element i to 448// see if X exists. If X does not exist, return i. 449static unsigned FindInOperandList(std::vector<ValueEntry> &Ops, unsigned i, 450 Value *X) { 451 unsigned XRank = Ops[i].Rank; 452 unsigned e = Ops.size(); 453 for (unsigned j = i+1; j != e && Ops[j].Rank == XRank; ++j) 454 if (Ops[j].Op == X) 455 return j; 456 // Scan backwards 457 for (unsigned j = i-1; j != ~0U && Ops[j].Rank == XRank; --j) 458 if (Ops[j].Op == X) 459 return j; 460 return i; 461} 462 463/// EmitAddTreeOfValues - Emit a tree of add instructions, summing Ops together 464/// and returning the result. Insert the tree before I. 465static Value *EmitAddTreeOfValues(Instruction *I, std::vector<Value*> &Ops) { 466 if (Ops.size() == 1) return Ops.back(); 467 468 Value *V1 = Ops.back(); 469 Ops.pop_back(); 470 Value *V2 = EmitAddTreeOfValues(I, Ops); 471 return BinaryOperator::createAdd(V2, V1, "tmp", I); 472} 473 474/// RemoveFactorFromExpression - If V is an expression tree that is a 475/// multiplication sequence, and if this sequence contains a multiply by Factor, 476/// remove Factor from the tree and return the new tree. 477Value *Reassociate::RemoveFactorFromExpression(Value *V, Value *Factor) { 478 BinaryOperator *BO = isReassociableOp(V, Instruction::Mul); 479 if (!BO) return 0; 480 481 std::vector<ValueEntry> Factors; 482 LinearizeExprTree(BO, Factors); 483 484 bool FoundFactor = false; 485 for (unsigned i = 0, e = Factors.size(); i != e; ++i) 486 if (Factors[i].Op == Factor) { 487 FoundFactor = true; 488 Factors.erase(Factors.begin()+i); 489 break; 490 } 491 if (!FoundFactor) { 492 // Make sure to restore the operands to the expression tree. 493 RewriteExprTree(BO, Factors); 494 return 0; 495 } 496 497 if (Factors.size() == 1) return Factors[0].Op; 498 499 RewriteExprTree(BO, Factors); 500 return BO; 501} 502 503/// FindSingleUseMultiplyFactors - If V is a single-use multiply, recursively 504/// add its operands as factors, otherwise add V to the list of factors. 505static void FindSingleUseMultiplyFactors(Value *V, 506 std::vector<Value*> &Factors) { 507 BinaryOperator *BO; 508 if ((!V->hasOneUse() && !V->use_empty()) || 509 !(BO = dyn_cast<BinaryOperator>(V)) || 510 BO->getOpcode() != Instruction::Mul) { 511 Factors.push_back(V); 512 return; 513 } 514 515 // Otherwise, add the LHS and RHS to the list of factors. 516 FindSingleUseMultiplyFactors(BO->getOperand(1), Factors); 517 FindSingleUseMultiplyFactors(BO->getOperand(0), Factors); 518} 519 520 521 522Value *Reassociate::OptimizeExpression(BinaryOperator *I, 523 std::vector<ValueEntry> &Ops) { 524 // Now that we have the linearized expression tree, try to optimize it. 525 // Start by folding any constants that we found. 526 bool IterateOptimization = false; 527 if (Ops.size() == 1) return Ops[0].Op; 528 529 unsigned Opcode = I->getOpcode(); 530 531 if (Constant *V1 = dyn_cast<Constant>(Ops[Ops.size()-2].Op)) 532 if (Constant *V2 = dyn_cast<Constant>(Ops.back().Op)) { 533 Ops.pop_back(); 534 Ops.back().Op = ConstantExpr::get(Opcode, V1, V2); 535 return OptimizeExpression(I, Ops); 536 } 537 538 // Check for destructive annihilation due to a constant being used. 539 if (ConstantIntegral *CstVal = dyn_cast<ConstantIntegral>(Ops.back().Op)) 540 switch (Opcode) { 541 default: break; 542 case Instruction::And: 543 if (CstVal->isNullValue()) { // ... & 0 -> 0 544 ++NumAnnihil; 545 return CstVal; 546 } else if (CstVal->isAllOnesValue()) { // ... & -1 -> ... 547 Ops.pop_back(); 548 } 549 break; 550 case Instruction::Mul: 551 if (CstVal->isNullValue()) { // ... * 0 -> 0 552 ++NumAnnihil; 553 return CstVal; 554 } else if (cast<ConstantInt>(CstVal)->getZExtValue() == 1) { 555 Ops.pop_back(); // ... * 1 -> ... 556 } 557 break; 558 case Instruction::Or: 559 if (CstVal->isAllOnesValue()) { // ... | -1 -> -1 560 ++NumAnnihil; 561 return CstVal; 562 } 563 // FALLTHROUGH! 564 case Instruction::Add: 565 case Instruction::Xor: 566 if (CstVal->isNullValue()) // ... [|^+] 0 -> ... 567 Ops.pop_back(); 568 break; 569 } 570 if (Ops.size() == 1) return Ops[0].Op; 571 572 // Handle destructive annihilation do to identities between elements in the 573 // argument list here. 574 switch (Opcode) { 575 default: break; 576 case Instruction::And: 577 case Instruction::Or: 578 case Instruction::Xor: 579 // Scan the operand lists looking for X and ~X pairs, along with X,X pairs. 580 // If we find any, we can simplify the expression. X&~X == 0, X|~X == -1. 581 for (unsigned i = 0, e = Ops.size(); i != e; ++i) { 582 // First, check for X and ~X in the operand list. 583 assert(i < Ops.size()); 584 if (BinaryOperator::isNot(Ops[i].Op)) { // Cannot occur for ^. 585 Value *X = BinaryOperator::getNotArgument(Ops[i].Op); 586 unsigned FoundX = FindInOperandList(Ops, i, X); 587 if (FoundX != i) { 588 if (Opcode == Instruction::And) { // ...&X&~X = 0 589 ++NumAnnihil; 590 return Constant::getNullValue(X->getType()); 591 } else if (Opcode == Instruction::Or) { // ...|X|~X = -1 592 ++NumAnnihil; 593 return ConstantIntegral::getAllOnesValue(X->getType()); 594 } 595 } 596 } 597 598 // Next, check for duplicate pairs of values, which we assume are next to 599 // each other, due to our sorting criteria. 600 assert(i < Ops.size()); 601 if (i+1 != Ops.size() && Ops[i+1].Op == Ops[i].Op) { 602 if (Opcode == Instruction::And || Opcode == Instruction::Or) { 603 // Drop duplicate values. 604 Ops.erase(Ops.begin()+i); 605 --i; --e; 606 IterateOptimization = true; 607 ++NumAnnihil; 608 } else { 609 assert(Opcode == Instruction::Xor); 610 if (e == 2) { 611 ++NumAnnihil; 612 return Constant::getNullValue(Ops[0].Op->getType()); 613 } 614 // ... X^X -> ... 615 Ops.erase(Ops.begin()+i, Ops.begin()+i+2); 616 i -= 1; e -= 2; 617 IterateOptimization = true; 618 ++NumAnnihil; 619 } 620 } 621 } 622 break; 623 624 case Instruction::Add: 625 // Scan the operand lists looking for X and -X pairs. If we find any, we 626 // can simplify the expression. X+-X == 0. 627 for (unsigned i = 0, e = Ops.size(); i != e; ++i) { 628 assert(i < Ops.size()); 629 // Check for X and -X in the operand list. 630 if (BinaryOperator::isNeg(Ops[i].Op)) { 631 Value *X = BinaryOperator::getNegArgument(Ops[i].Op); 632 unsigned FoundX = FindInOperandList(Ops, i, X); 633 if (FoundX != i) { 634 // Remove X and -X from the operand list. 635 if (Ops.size() == 2) { 636 ++NumAnnihil; 637 return Constant::getNullValue(X->getType()); 638 } else { 639 Ops.erase(Ops.begin()+i); 640 if (i < FoundX) 641 --FoundX; 642 else 643 --i; // Need to back up an extra one. 644 Ops.erase(Ops.begin()+FoundX); 645 IterateOptimization = true; 646 ++NumAnnihil; 647 --i; // Revisit element. 648 e -= 2; // Removed two elements. 649 } 650 } 651 } 652 } 653 654 655 // Scan the operand list, checking to see if there are any common factors 656 // between operands. Consider something like A*A+A*B*C+D. We would like to 657 // reassociate this to A*(A+B*C)+D, which reduces the number of multiplies. 658 // To efficiently find this, we count the number of times a factor occurs 659 // for any ADD operands that are MULs. 660 std::map<Value*, unsigned> FactorOccurrences; 661 unsigned MaxOcc = 0; 662 Value *MaxOccVal = 0; 663 if (!I->getType()->isFloatingPoint()) { 664 for (unsigned i = 0, e = Ops.size(); i != e; ++i) { 665 if (BinaryOperator *BOp = dyn_cast<BinaryOperator>(Ops[i].Op)) 666 if (BOp->getOpcode() == Instruction::Mul && BOp->use_empty()) { 667 // Compute all of the factors of this added value. 668 std::vector<Value*> Factors; 669 FindSingleUseMultiplyFactors(BOp, Factors); 670 assert(Factors.size() > 1 && "Bad linearize!"); 671 672 // Add one to FactorOccurrences for each unique factor in this op. 673 if (Factors.size() == 2) { 674 unsigned Occ = ++FactorOccurrences[Factors[0]]; 675 if (Occ > MaxOcc) { MaxOcc = Occ; MaxOccVal = Factors[0]; } 676 if (Factors[0] != Factors[1]) { // Don't double count A*A. 677 Occ = ++FactorOccurrences[Factors[1]]; 678 if (Occ > MaxOcc) { MaxOcc = Occ; MaxOccVal = Factors[1]; } 679 } 680 } else { 681 std::set<Value*> Duplicates; 682 for (unsigned i = 0, e = Factors.size(); i != e; ++i) 683 if (Duplicates.insert(Factors[i]).second) { 684 unsigned Occ = ++FactorOccurrences[Factors[i]]; 685 if (Occ > MaxOcc) { MaxOcc = Occ; MaxOccVal = Factors[i]; } 686 } 687 } 688 } 689 } 690 } 691 692 // If any factor occurred more than one time, we can pull it out. 693 if (MaxOcc > 1) { 694 DOUT << "\nFACTORING [" << MaxOcc << "]: " << *MaxOccVal << "\n"; 695 696 // Create a new instruction that uses the MaxOccVal twice. If we don't do 697 // this, we could otherwise run into situations where removing a factor 698 // from an expression will drop a use of maxocc, and this can cause 699 // RemoveFactorFromExpression on successive values to behave differently. 700 Instruction *DummyInst = BinaryOperator::createAdd(MaxOccVal, MaxOccVal); 701 std::vector<Value*> NewMulOps; 702 for (unsigned i = 0, e = Ops.size(); i != e; ++i) { 703 if (Value *V = RemoveFactorFromExpression(Ops[i].Op, MaxOccVal)) { 704 NewMulOps.push_back(V); 705 Ops.erase(Ops.begin()+i); 706 --i; --e; 707 } 708 } 709 710 // No need for extra uses anymore. 711 delete DummyInst; 712 713 unsigned NumAddedValues = NewMulOps.size(); 714 Value *V = EmitAddTreeOfValues(I, NewMulOps); 715 Value *V2 = BinaryOperator::createMul(V, MaxOccVal, "tmp", I); 716 717 // Now that we have inserted V and its sole use, optimize it. This allows 718 // us to handle cases that require multiple factoring steps, such as this: 719 // A*A*B + A*A*C --> A*(A*B+A*C) --> A*(A*(B+C)) 720 if (NumAddedValues > 1) 721 ReassociateExpression(cast<BinaryOperator>(V)); 722 723 ++NumFactor; 724 725 if (Ops.size() == 0) 726 return V2; 727 728 // Add the new value to the list of things being added. 729 Ops.insert(Ops.begin(), ValueEntry(getRank(V2), V2)); 730 731 // Rewrite the tree so that there is now a use of V. 732 RewriteExprTree(I, Ops); 733 return OptimizeExpression(I, Ops); 734 } 735 break; 736 //case Instruction::Mul: 737 } 738 739 if (IterateOptimization) 740 return OptimizeExpression(I, Ops); 741 return 0; 742} 743 744 745/// ReassociateBB - Inspect all of the instructions in this basic block, 746/// reassociating them as we go. 747void Reassociate::ReassociateBB(BasicBlock *BB) { 748 for (BasicBlock::iterator BBI = BB->begin(); BBI != BB->end(); ) { 749 Instruction *BI = BBI++; 750 if (BI->getOpcode() == Instruction::Shl && 751 isa<ConstantInt>(BI->getOperand(1))) 752 if (Instruction *NI = ConvertShiftToMul(BI)) { 753 MadeChange = true; 754 BI = NI; 755 } 756 757 // Reject cases where it is pointless to do this. 758 if (!isa<BinaryOperator>(BI) || BI->getType()->isFloatingPoint() || 759 isa<PackedType>(BI->getType())) 760 continue; // Floating point ops are not associative. 761 762 // If this is a subtract instruction which is not already in negate form, 763 // see if we can convert it to X+-Y. 764 if (BI->getOpcode() == Instruction::Sub) { 765 if (!BinaryOperator::isNeg(BI)) { 766 if (Instruction *NI = BreakUpSubtract(BI)) { 767 MadeChange = true; 768 BI = NI; 769 } 770 } else { 771 // Otherwise, this is a negation. See if the operand is a multiply tree 772 // and if this is not an inner node of a multiply tree. 773 if (isReassociableOp(BI->getOperand(1), Instruction::Mul) && 774 (!BI->hasOneUse() || 775 !isReassociableOp(BI->use_back(), Instruction::Mul))) { 776 BI = LowerNegateToMultiply(BI); 777 MadeChange = true; 778 } 779 } 780 } 781 782 // If this instruction is a commutative binary operator, process it. 783 if (!BI->isAssociative()) continue; 784 BinaryOperator *I = cast<BinaryOperator>(BI); 785 786 // If this is an interior node of a reassociable tree, ignore it until we 787 // get to the root of the tree, to avoid N^2 analysis. 788 if (I->hasOneUse() && isReassociableOp(I->use_back(), I->getOpcode())) 789 continue; 790 791 // If this is an add tree that is used by a sub instruction, ignore it 792 // until we process the subtract. 793 if (I->hasOneUse() && I->getOpcode() == Instruction::Add && 794 cast<Instruction>(I->use_back())->getOpcode() == Instruction::Sub) 795 continue; 796 797 ReassociateExpression(I); 798 } 799} 800 801void Reassociate::ReassociateExpression(BinaryOperator *I) { 802 803 // First, walk the expression tree, linearizing the tree, collecting 804 std::vector<ValueEntry> Ops; 805 LinearizeExprTree(I, Ops); 806 807 DOUT << "RAIn:\t"; DEBUG(PrintOps(I, Ops)); DOUT << "\n"; 808 809 // Now that we have linearized the tree to a list and have gathered all of 810 // the operands and their ranks, sort the operands by their rank. Use a 811 // stable_sort so that values with equal ranks will have their relative 812 // positions maintained (and so the compiler is deterministic). Note that 813 // this sorts so that the highest ranking values end up at the beginning of 814 // the vector. 815 std::stable_sort(Ops.begin(), Ops.end()); 816 817 // OptimizeExpression - Now that we have the expression tree in a convenient 818 // sorted form, optimize it globally if possible. 819 if (Value *V = OptimizeExpression(I, Ops)) { 820 // This expression tree simplified to something that isn't a tree, 821 // eliminate it. 822 DOUT << "Reassoc to scalar: " << *V << "\n"; 823 I->replaceAllUsesWith(V); 824 RemoveDeadBinaryOp(I); 825 return; 826 } 827 828 // We want to sink immediates as deeply as possible except in the case where 829 // this is a multiply tree used only by an add, and the immediate is a -1. 830 // In this case we reassociate to put the negation on the outside so that we 831 // can fold the negation into the add: (-X)*Y + Z -> Z-X*Y 832 if (I->getOpcode() == Instruction::Mul && I->hasOneUse() && 833 cast<Instruction>(I->use_back())->getOpcode() == Instruction::Add && 834 isa<ConstantInt>(Ops.back().Op) && 835 cast<ConstantInt>(Ops.back().Op)->isAllOnesValue()) { 836 Ops.insert(Ops.begin(), Ops.back()); 837 Ops.pop_back(); 838 } 839 840 DOUT << "RAOut:\t"; DEBUG(PrintOps(I, Ops)); DOUT << "\n"; 841 842 if (Ops.size() == 1) { 843 // This expression tree simplified to something that isn't a tree, 844 // eliminate it. 845 I->replaceAllUsesWith(Ops[0].Op); 846 RemoveDeadBinaryOp(I); 847 } else { 848 // Now that we ordered and optimized the expressions, splat them back into 849 // the expression tree, removing any unneeded nodes. 850 RewriteExprTree(I, Ops); 851 } 852} 853 854 855bool Reassociate::runOnFunction(Function &F) { 856 // Recalculate the rank map for F 857 BuildRankMap(F); 858 859 MadeChange = false; 860 for (Function::iterator FI = F.begin(), FE = F.end(); FI != FE; ++FI) 861 ReassociateBB(FI); 862 863 // We are done with the rank map... 864 RankMap.clear(); 865 ValueRankMap.clear(); 866 return MadeChange; 867} 868 869