Reassociate.cpp revision e5022fe4cd83eef91f5c3a21c943ca9b65507ab8
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/Function.h" 27#include "llvm/Instructions.h" 28#include "llvm/Pass.h" 29#include "llvm/Type.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> 36#include <iostream> 37using namespace llvm; 38 39namespace { 40 Statistic<> NumLinear ("reassociate","Number of insts linearized"); 41 Statistic<> NumChanged("reassociate","Number of insts reassociated"); 42 Statistic<> NumSwapped("reassociate","Number of insts with operands swapped"); 43 Statistic<> NumAnnihil("reassociate","Number of expr tree annihilated"); 44 Statistic<> NumFactor ("reassociate","Number of multiplies factored"); 45 46 struct ValueEntry { 47 unsigned Rank; 48 Value *Op; 49 ValueEntry(unsigned R, Value *O) : Rank(R), Op(O) {} 50 }; 51 inline bool operator<(const ValueEntry &LHS, const ValueEntry &RHS) { 52 return LHS.Rank > RHS.Rank; // Sort so that highest rank goes to start. 53 } 54} 55 56/// PrintOps - Print out the expression identified in the Ops list. 57/// 58static void PrintOps(Instruction *I, const std::vector<ValueEntry> &Ops) { 59 Module *M = I->getParent()->getParent()->getParent(); 60 std::cerr << Instruction::getOpcodeName(I->getOpcode()) << " " 61 << *Ops[0].Op->getType(); 62 for (unsigned i = 0, e = Ops.size(); i != e; ++i) 63 WriteAsOperand(std::cerr << " ", Ops[i].Op, false, true, M) 64 << "," << Ops[i].Rank; 65} 66 67namespace { 68 class Reassociate : public FunctionPass { 69 std::map<BasicBlock*, unsigned> RankMap; 70 std::map<Value*, unsigned> ValueRankMap; 71 bool MadeChange; 72 public: 73 bool runOnFunction(Function &F); 74 75 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 76 AU.setPreservesCFG(); 77 } 78 private: 79 void BuildRankMap(Function &F); 80 unsigned getRank(Value *V); 81 void RewriteExprTree(BinaryOperator *I, unsigned Idx, 82 std::vector<ValueEntry> &Ops); 83 Value *OptimizeExpression(BinaryOperator *I, std::vector<ValueEntry> &Ops); 84 void LinearizeExprTree(BinaryOperator *I, std::vector<ValueEntry> &Ops); 85 void LinearizeExpr(BinaryOperator *I); 86 Value *RemoveFactorFromExpression(Value *V, Value *Factor); 87 void ReassociateBB(BasicBlock *BB); 88 89 void RemoveDeadBinaryOp(Value *V); 90 }; 91 92 RegisterOpt<Reassociate> X("reassociate", "Reassociate expressions"); 93} 94 95// Public interface to the Reassociate pass 96FunctionPass *llvm::createReassociatePass() { return new Reassociate(); } 97 98void Reassociate::RemoveDeadBinaryOp(Value *V) { 99 BinaryOperator *BOp = dyn_cast<BinaryOperator>(V); 100 if (!BOp || !BOp->use_empty()) return; 101 102 Value *LHS = BOp->getOperand(0), *RHS = BOp->getOperand(1); 103 RemoveDeadBinaryOp(LHS); 104 RemoveDeadBinaryOp(RHS); 105} 106 107 108static bool isUnmovableInstruction(Instruction *I) { 109 if (I->getOpcode() == Instruction::PHI || 110 I->getOpcode() == Instruction::Alloca || 111 I->getOpcode() == Instruction::Load || 112 I->getOpcode() == Instruction::Malloc || 113 I->getOpcode() == Instruction::Invoke || 114 I->getOpcode() == Instruction::Call || 115 I->getOpcode() == Instruction::Div || 116 I->getOpcode() == Instruction::Rem) 117 return true; 118 return false; 119} 120 121void Reassociate::BuildRankMap(Function &F) { 122 unsigned i = 2; 123 124 // Assign distinct ranks to function arguments 125 for (Function::arg_iterator I = F.arg_begin(), E = F.arg_end(); I != E; ++I) 126 ValueRankMap[I] = ++i; 127 128 ReversePostOrderTraversal<Function*> RPOT(&F); 129 for (ReversePostOrderTraversal<Function*>::rpo_iterator I = RPOT.begin(), 130 E = RPOT.end(); I != E; ++I) { 131 BasicBlock *BB = *I; 132 unsigned BBRank = RankMap[BB] = ++i << 16; 133 134 // Walk the basic block, adding precomputed ranks for any instructions that 135 // we cannot move. This ensures that the ranks for these instructions are 136 // all different in the block. 137 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) 138 if (isUnmovableInstruction(I)) 139 ValueRankMap[I] = ++BBRank; 140 } 141} 142 143unsigned Reassociate::getRank(Value *V) { 144 if (isa<Argument>(V)) return ValueRankMap[V]; // Function argument... 145 146 Instruction *I = dyn_cast<Instruction>(V); 147 if (I == 0) return 0; // Otherwise it's a global or constant, rank 0. 148 149 unsigned &CachedRank = ValueRankMap[I]; 150 if (CachedRank) return CachedRank; // Rank already known? 151 152 // If this is an expression, return the 1+MAX(rank(LHS), rank(RHS)) so that 153 // we can reassociate expressions for code motion! Since we do not recurse 154 // for PHI nodes, we cannot have infinite recursion here, because there 155 // cannot be loops in the value graph that do not go through PHI nodes. 156 unsigned Rank = 0, MaxRank = RankMap[I->getParent()]; 157 for (unsigned i = 0, e = I->getNumOperands(); 158 i != e && Rank != MaxRank; ++i) 159 Rank = std::max(Rank, getRank(I->getOperand(i))); 160 161 // If this is a not or neg instruction, do not count it for rank. This 162 // assures us that X and ~X will have the same rank. 163 if (!I->getType()->isIntegral() || 164 (!BinaryOperator::isNot(I) && !BinaryOperator::isNeg(I))) 165 ++Rank; 166 167 //DEBUG(std::cerr << "Calculated Rank[" << V->getName() << "] = " 168 //<< Rank << "\n"); 169 170 return CachedRank = Rank; 171} 172 173/// isReassociableOp - Return true if V is an instruction of the specified 174/// opcode and if it only has one use. 175static BinaryOperator *isReassociableOp(Value *V, unsigned Opcode) { 176 if (V->hasOneUse() && isa<Instruction>(V) && 177 cast<Instruction>(V)->getOpcode() == Opcode) 178 return cast<BinaryOperator>(V); 179 return 0; 180} 181 182/// LowerNegateToMultiply - Replace 0-X with X*-1. 183/// 184static Instruction *LowerNegateToMultiply(Instruction *Neg) { 185 Constant *Cst; 186 if (Neg->getType()->isFloatingPoint()) 187 Cst = ConstantFP::get(Neg->getType(), -1); 188 else 189 Cst = ConstantInt::getAllOnesValue(Neg->getType()); 190 191 std::string NegName = Neg->getName(); Neg->setName(""); 192 Instruction *Res = BinaryOperator::createMul(Neg->getOperand(1), Cst, NegName, 193 Neg); 194 Neg->replaceAllUsesWith(Res); 195 Neg->eraseFromParent(); 196 return Res; 197} 198 199// Given an expression of the form '(A+B)+(D+C)', turn it into '(((A+B)+C)+D)'. 200// Note that if D is also part of the expression tree that we recurse to 201// linearize it as well. Besides that case, this does not recurse into A,B, or 202// C. 203void Reassociate::LinearizeExpr(BinaryOperator *I) { 204 BinaryOperator *LHS = cast<BinaryOperator>(I->getOperand(0)); 205 BinaryOperator *RHS = cast<BinaryOperator>(I->getOperand(1)); 206 assert(isReassociableOp(LHS, I->getOpcode()) && 207 isReassociableOp(RHS, I->getOpcode()) && 208 "Not an expression that needs linearization?"); 209 210 DEBUG(std::cerr << "Linear" << *LHS << *RHS << *I); 211 212 // Move the RHS instruction to live immediately before I, avoiding breaking 213 // dominator properties. 214 RHS->moveBefore(I); 215 216 // Move operands around to do the linearization. 217 I->setOperand(1, RHS->getOperand(0)); 218 RHS->setOperand(0, LHS); 219 I->setOperand(0, RHS); 220 221 ++NumLinear; 222 MadeChange = true; 223 DEBUG(std::cerr << "Linearized: " << *I); 224 225 // If D is part of this expression tree, tail recurse. 226 if (isReassociableOp(I->getOperand(1), I->getOpcode())) 227 LinearizeExpr(I); 228} 229 230 231/// LinearizeExprTree - Given an associative binary expression tree, traverse 232/// all of the uses putting it into canonical form. This forces a left-linear 233/// form of the the expression (((a+b)+c)+d), and collects information about the 234/// rank of the non-tree operands. 235/// 236void Reassociate::LinearizeExprTree(BinaryOperator *I, 237 std::vector<ValueEntry> &Ops) { 238 Value *LHS = I->getOperand(0), *RHS = I->getOperand(1); 239 unsigned Opcode = I->getOpcode(); 240 241 // First step, linearize the expression if it is in ((A+B)+(C+D)) form. 242 BinaryOperator *LHSBO = isReassociableOp(LHS, Opcode); 243 BinaryOperator *RHSBO = isReassociableOp(RHS, Opcode); 244 245 // If this is a multiply expression tree and it contains internal negations, 246 // transform them into multiplies by -1 so they can be reassociated. 247 if (I->getOpcode() == Instruction::Mul) { 248 if (!LHSBO && LHS->hasOneUse() && BinaryOperator::isNeg(LHS)) { 249 LHS = LowerNegateToMultiply(cast<Instruction>(LHS)); 250 LHSBO = isReassociableOp(LHS, Opcode); 251 } 252 if (!RHSBO && RHS->hasOneUse() && BinaryOperator::isNeg(RHS)) { 253 RHS = LowerNegateToMultiply(cast<Instruction>(RHS)); 254 RHSBO = isReassociableOp(RHS, Opcode); 255 } 256 } 257 258 if (!LHSBO) { 259 if (!RHSBO) { 260 // Neither the LHS or RHS as part of the tree, thus this is a leaf. As 261 // such, just remember these operands and their rank. 262 Ops.push_back(ValueEntry(getRank(LHS), LHS)); 263 Ops.push_back(ValueEntry(getRank(RHS), RHS)); 264 return; 265 } else { 266 // Turn X+(Y+Z) -> (Y+Z)+X 267 std::swap(LHSBO, RHSBO); 268 std::swap(LHS, RHS); 269 bool Success = !I->swapOperands(); 270 assert(Success && "swapOperands failed"); 271 MadeChange = true; 272 } 273 } else if (RHSBO) { 274 // Turn (A+B)+(C+D) -> (((A+B)+C)+D). This guarantees the the RHS is not 275 // part of the expression tree. 276 LinearizeExpr(I); 277 LHS = LHSBO = cast<BinaryOperator>(I->getOperand(0)); 278 RHS = I->getOperand(1); 279 RHSBO = 0; 280 } 281 282 // Okay, now we know that the LHS is a nested expression and that the RHS is 283 // not. Perform reassociation. 284 assert(!isReassociableOp(RHS, Opcode) && "LinearizeExpr failed!"); 285 286 // Move LHS right before I to make sure that the tree expression dominates all 287 // values. 288 LHSBO->moveBefore(I); 289 290 // Linearize the expression tree on the LHS. 291 LinearizeExprTree(LHSBO, Ops); 292 293 // Remember the RHS operand and its rank. 294 Ops.push_back(ValueEntry(getRank(RHS), RHS)); 295} 296 297// RewriteExprTree - Now that the operands for this expression tree are 298// linearized and optimized, emit them in-order. This function is written to be 299// tail recursive. 300void Reassociate::RewriteExprTree(BinaryOperator *I, unsigned i, 301 std::vector<ValueEntry> &Ops) { 302 if (i+2 == Ops.size()) { 303 if (I->getOperand(0) != Ops[i].Op || 304 I->getOperand(1) != Ops[i+1].Op) { 305 Value *OldLHS = I->getOperand(0); 306 DEBUG(std::cerr << "RA: " << *I); 307 I->setOperand(0, Ops[i].Op); 308 I->setOperand(1, Ops[i+1].Op); 309 DEBUG(std::cerr << "TO: " << *I); 310 MadeChange = true; 311 ++NumChanged; 312 313 // If we reassociated a tree to fewer operands (e.g. (1+a+2) -> (a+3) 314 // delete the extra, now dead, nodes. 315 RemoveDeadBinaryOp(OldLHS); 316 } 317 return; 318 } 319 assert(i+2 < Ops.size() && "Ops index out of range!"); 320 321 if (I->getOperand(1) != Ops[i].Op) { 322 DEBUG(std::cerr << "RA: " << *I); 323 I->setOperand(1, Ops[i].Op); 324 DEBUG(std::cerr << "TO: " << *I); 325 MadeChange = true; 326 ++NumChanged; 327 } 328 329 BinaryOperator *LHS = cast<BinaryOperator>(I->getOperand(0)); 330 assert(LHS->getOpcode() == I->getOpcode() && 331 "Improper expression tree!"); 332 333 // Compactify the tree instructions together with each other to guarantee 334 // that the expression tree is dominated by all of Ops. 335 LHS->moveBefore(I); 336 RewriteExprTree(LHS, i+1, Ops); 337} 338 339 340 341// NegateValue - Insert instructions before the instruction pointed to by BI, 342// that computes the negative version of the value specified. The negative 343// version of the value is returned, and BI is left pointing at the instruction 344// that should be processed next by the reassociation pass. 345// 346static Value *NegateValue(Value *V, Instruction *BI) { 347 // We are trying to expose opportunity for reassociation. One of the things 348 // that we want to do to achieve this is to push a negation as deep into an 349 // expression chain as possible, to expose the add instructions. In practice, 350 // this means that we turn this: 351 // X = -(A+12+C+D) into X = -A + -12 + -C + -D = -12 + -A + -C + -D 352 // so that later, a: Y = 12+X could get reassociated with the -12 to eliminate 353 // the constants. We assume that instcombine will clean up the mess later if 354 // we introduce tons of unnecessary negation instructions... 355 // 356 if (Instruction *I = dyn_cast<Instruction>(V)) 357 if (I->getOpcode() == Instruction::Add && I->hasOneUse()) { 358 // Push the negates through the add. 359 I->setOperand(0, NegateValue(I->getOperand(0), BI)); 360 I->setOperand(1, NegateValue(I->getOperand(1), BI)); 361 362 // We must move the add instruction here, because the neg instructions do 363 // not dominate the old add instruction in general. By moving it, we are 364 // assured that the neg instructions we just inserted dominate the 365 // instruction we are about to insert after them. 366 // 367 I->moveBefore(BI); 368 I->setName(I->getName()+".neg"); 369 return I; 370 } 371 372 // Insert a 'neg' instruction that subtracts the value from zero to get the 373 // negation. 374 // 375 return BinaryOperator::createNeg(V, V->getName() + ".neg", BI); 376} 377 378/// BreakUpSubtract - If we have (X-Y), and if either X is an add, or if this is 379/// only used by an add, transform this into (X+(0-Y)) to promote better 380/// reassociation. 381static Instruction *BreakUpSubtract(Instruction *Sub) { 382 // Don't bother to break this up unless either the LHS is an associable add or 383 // if this is only used by one. 384 if (!isReassociableOp(Sub->getOperand(0), Instruction::Add) && 385 !isReassociableOp(Sub->getOperand(1), Instruction::Add) && 386 !(Sub->hasOneUse() &&isReassociableOp(Sub->use_back(), Instruction::Add))) 387 return 0; 388 389 // Convert a subtract into an add and a neg instruction... so that sub 390 // instructions can be commuted with other add instructions... 391 // 392 // Calculate the negative value of Operand 1 of the sub instruction... 393 // and set it as the RHS of the add instruction we just made... 394 // 395 std::string Name = Sub->getName(); 396 Sub->setName(""); 397 Value *NegVal = NegateValue(Sub->getOperand(1), Sub); 398 Instruction *New = 399 BinaryOperator::createAdd(Sub->getOperand(0), NegVal, Name, Sub); 400 401 // Everyone now refers to the add instruction. 402 Sub->replaceAllUsesWith(New); 403 Sub->eraseFromParent(); 404 405 DEBUG(std::cerr << "Negated: " << *New); 406 return New; 407} 408 409/// ConvertShiftToMul - If this is a shift of a reassociable multiply or is used 410/// by one, change this into a multiply by a constant to assist with further 411/// reassociation. 412static Instruction *ConvertShiftToMul(Instruction *Shl) { 413 if (!isReassociableOp(Shl->getOperand(0), Instruction::Mul) && 414 !(Shl->hasOneUse() && isReassociableOp(Shl->use_back(),Instruction::Mul))) 415 return 0; 416 417 Constant *MulCst = ConstantInt::get(Shl->getType(), 1); 418 MulCst = ConstantExpr::getShl(MulCst, cast<Constant>(Shl->getOperand(1))); 419 420 std::string Name = Shl->getName(); Shl->setName(""); 421 Instruction *Mul = BinaryOperator::createMul(Shl->getOperand(0), MulCst, 422 Name, Shl); 423 Shl->replaceAllUsesWith(Mul); 424 Shl->eraseFromParent(); 425 return Mul; 426} 427 428// Scan backwards and forwards among values with the same rank as element i to 429// see if X exists. If X does not exist, return i. 430static unsigned FindInOperandList(std::vector<ValueEntry> &Ops, unsigned i, 431 Value *X) { 432 unsigned XRank = Ops[i].Rank; 433 unsigned e = Ops.size(); 434 for (unsigned j = i+1; j != e && Ops[j].Rank == XRank; ++j) 435 if (Ops[j].Op == X) 436 return j; 437 // Scan backwards 438 for (unsigned j = i-1; j != ~0U && Ops[j].Rank == XRank; --j) 439 if (Ops[j].Op == X) 440 return j; 441 return i; 442} 443 444/// EmitAddTreeOfValues - Emit a tree of add instructions, summing Ops together 445/// and returning the result. Insert the tree before I. 446static Value *EmitAddTreeOfValues(Instruction *I, std::vector<Value*> &Ops) { 447 if (Ops.size() == 1) return Ops.back(); 448 449 Value *V1 = Ops.back(); 450 Ops.pop_back(); 451 Value *V2 = EmitAddTreeOfValues(I, Ops); 452 return BinaryOperator::createAdd(V2, V1, "tmp", I); 453} 454 455/// RemoveFactorFromExpression - If V is an expression tree that is a 456/// multiplication sequence, and if this sequence contains a multiply by Factor, 457/// remove Factor from the tree and return the new tree. 458Value *Reassociate::RemoveFactorFromExpression(Value *V, Value *Factor) { 459 BinaryOperator *BO = isReassociableOp(V, Instruction::Mul); 460 if (!BO) return 0; 461 462 std::vector<ValueEntry> Factors; 463 LinearizeExprTree(BO, Factors); 464 465 bool FoundFactor = false; 466 for (unsigned i = 0, e = Factors.size(); i != e; ++i) 467 if (Factors[i].Op == Factor) { 468 FoundFactor = true; 469 Factors.erase(Factors.begin()+i); 470 break; 471 } 472 if (!FoundFactor) return 0; 473 474 if (Factors.size() == 1) return Factors[0].Op; 475 476 RewriteExprTree(BO, 0, Factors); 477 return BO; 478} 479 480 481Value *Reassociate::OptimizeExpression(BinaryOperator *I, 482 std::vector<ValueEntry> &Ops) { 483 // Now that we have the linearized expression tree, try to optimize it. 484 // Start by folding any constants that we found. 485 bool IterateOptimization = false; 486 if (Ops.size() == 1) return Ops[0].Op; 487 488 unsigned Opcode = I->getOpcode(); 489 490 if (Constant *V1 = dyn_cast<Constant>(Ops[Ops.size()-2].Op)) 491 if (Constant *V2 = dyn_cast<Constant>(Ops.back().Op)) { 492 Ops.pop_back(); 493 Ops.back().Op = ConstantExpr::get(Opcode, V1, V2); 494 return OptimizeExpression(I, Ops); 495 } 496 497 // Check for destructive annihilation due to a constant being used. 498 if (ConstantIntegral *CstVal = dyn_cast<ConstantIntegral>(Ops.back().Op)) 499 switch (Opcode) { 500 default: break; 501 case Instruction::And: 502 if (CstVal->isNullValue()) { // ... & 0 -> 0 503 ++NumAnnihil; 504 return CstVal; 505 } else if (CstVal->isAllOnesValue()) { // ... & -1 -> ... 506 Ops.pop_back(); 507 } 508 break; 509 case Instruction::Mul: 510 if (CstVal->isNullValue()) { // ... * 0 -> 0 511 ++NumAnnihil; 512 return CstVal; 513 } else if (cast<ConstantInt>(CstVal)->getRawValue() == 1) { 514 Ops.pop_back(); // ... * 1 -> ... 515 } 516 break; 517 case Instruction::Or: 518 if (CstVal->isAllOnesValue()) { // ... | -1 -> -1 519 ++NumAnnihil; 520 return CstVal; 521 } 522 // FALLTHROUGH! 523 case Instruction::Add: 524 case Instruction::Xor: 525 if (CstVal->isNullValue()) // ... [|^+] 0 -> ... 526 Ops.pop_back(); 527 break; 528 } 529 if (Ops.size() == 1) return Ops[0].Op; 530 531 // Handle destructive annihilation do to identities between elements in the 532 // argument list here. 533 switch (Opcode) { 534 default: break; 535 case Instruction::And: 536 case Instruction::Or: 537 case Instruction::Xor: 538 // Scan the operand lists looking for X and ~X pairs, along with X,X pairs. 539 // If we find any, we can simplify the expression. X&~X == 0, X|~X == -1. 540 for (unsigned i = 0, e = Ops.size(); i != e; ++i) { 541 // First, check for X and ~X in the operand list. 542 assert(i < Ops.size()); 543 if (BinaryOperator::isNot(Ops[i].Op)) { // Cannot occur for ^. 544 Value *X = BinaryOperator::getNotArgument(Ops[i].Op); 545 unsigned FoundX = FindInOperandList(Ops, i, X); 546 if (FoundX != i) { 547 if (Opcode == Instruction::And) { // ...&X&~X = 0 548 ++NumAnnihil; 549 return Constant::getNullValue(X->getType()); 550 } else if (Opcode == Instruction::Or) { // ...|X|~X = -1 551 ++NumAnnihil; 552 return ConstantIntegral::getAllOnesValue(X->getType()); 553 } 554 } 555 } 556 557 // Next, check for duplicate pairs of values, which we assume are next to 558 // each other, due to our sorting criteria. 559 assert(i < Ops.size()); 560 if (i+1 != Ops.size() && Ops[i+1].Op == Ops[i].Op) { 561 if (Opcode == Instruction::And || Opcode == Instruction::Or) { 562 // Drop duplicate values. 563 Ops.erase(Ops.begin()+i); 564 --i; --e; 565 IterateOptimization = true; 566 ++NumAnnihil; 567 } else { 568 assert(Opcode == Instruction::Xor); 569 if (e == 2) { 570 ++NumAnnihil; 571 return Constant::getNullValue(Ops[0].Op->getType()); 572 } 573 // ... X^X -> ... 574 Ops.erase(Ops.begin()+i, Ops.begin()+i+2); 575 i -= 1; e -= 2; 576 IterateOptimization = true; 577 ++NumAnnihil; 578 } 579 } 580 } 581 break; 582 583 case Instruction::Add: 584 // Scan the operand lists looking for X and -X pairs. If we find any, we 585 // can simplify the expression. X+-X == 0. 586 for (unsigned i = 0, e = Ops.size(); i != e; ++i) { 587 assert(i < Ops.size()); 588 // Check for X and -X in the operand list. 589 if (BinaryOperator::isNeg(Ops[i].Op)) { 590 Value *X = BinaryOperator::getNegArgument(Ops[i].Op); 591 unsigned FoundX = FindInOperandList(Ops, i, X); 592 if (FoundX != i) { 593 // Remove X and -X from the operand list. 594 if (Ops.size() == 2) { 595 ++NumAnnihil; 596 return Constant::getNullValue(X->getType()); 597 } else { 598 Ops.erase(Ops.begin()+i); 599 if (i < FoundX) 600 --FoundX; 601 else 602 --i; // Need to back up an extra one. 603 Ops.erase(Ops.begin()+FoundX); 604 IterateOptimization = true; 605 ++NumAnnihil; 606 --i; // Revisit element. 607 e -= 2; // Removed two elements. 608 } 609 } 610 } 611 } 612 613 614 // Scan the operand list, checking to see if there are any common factors 615 // between operands. Consider something like A*A+A*B*C+D. We would like to 616 // reassociate this to A*(A+B*C)+D, which reduces the number of multiplies. 617 // To efficiently find this, we count the number of times a factor occurs 618 // for any ADD operands that are MULs. 619 std::map<Value*, unsigned> FactorOccurrences; 620 unsigned MaxOcc = 0; 621 Value *MaxOccVal = 0; 622 if (!I->getType()->isFloatingPoint()) { 623 for (unsigned i = 0, e = Ops.size(); i != e; ++i) { 624 if (BinaryOperator *BOp = dyn_cast<BinaryOperator>(Ops[i].Op)) 625 if (BOp->getOpcode() == Instruction::Mul && BOp->hasOneUse()) { 626 // Compute all of the factors of this added value. 627 std::vector<ValueEntry> Factors; 628 LinearizeExprTree(BOp, Factors); 629 assert(Factors.size() > 1 && "Bad linearize!"); 630 631 // Add one to FactorOccurrences for each unique factor in this op. 632 if (Factors.size() == 2) { 633 unsigned Occ = ++FactorOccurrences[Factors[0].Op]; 634 if (Occ > MaxOcc) { MaxOcc = Occ; MaxOccVal = Factors[0].Op; } 635 if (Factors[0].Op != Factors[1].Op) { // Don't double count A*A. 636 Occ = ++FactorOccurrences[Factors[1].Op]; 637 if (Occ > MaxOcc) { MaxOcc = Occ; MaxOccVal = Factors[1].Op; } 638 } 639 } else { 640 std::set<Value*> Duplicates; 641 for (unsigned i = 0, e = Factors.size(); i != e; ++i) 642 if (Duplicates.insert(Factors[i].Op).second) { 643 unsigned Occ = ++FactorOccurrences[Factors[i].Op]; 644 if (Occ > MaxOcc) { MaxOcc = Occ; MaxOccVal = Factors[i].Op; } 645 } 646 } 647 } 648 } 649 } 650 651 // If any factor occurred more than one time, we can pull it out. 652 if (MaxOcc > 1) { 653 DEBUG(std::cerr << "\nFACTORING [" << MaxOcc << "]: " 654 << *MaxOccVal << "\n"); 655 656 // Create a new instruction that uses the MaxOccVal twice. If we don't do 657 // this, we could otherwise run into situations where removing a factor 658 // from an expression will drop a use of maxocc, and this can cause 659 // RemoveFactorFromExpression on successive values to behave differently. 660 Instruction *DummyInst = BinaryOperator::createAdd(MaxOccVal, MaxOccVal); 661 std::vector<Value*> NewMulOps; 662 for (unsigned i = 0, e = Ops.size(); i != e; ++i) { 663 if (Value *V = RemoveFactorFromExpression(Ops[i].Op, MaxOccVal)) { 664 NewMulOps.push_back(V); 665 Ops.erase(Ops.begin()+i); 666 --i; --e; 667 } 668 } 669 670 // No need for extra uses anymore. 671 delete DummyInst; 672 673 Value *V = EmitAddTreeOfValues(I, NewMulOps); 674 // FIXME: Must optimize V now, to handle this case: 675 // A*A*B + A*A*C -> A*(A*B+A*C) -> A*(A*(B+C)) 676 V = BinaryOperator::createMul(V, MaxOccVal, "tmp", I); 677 678 ++NumFactor; 679 680 if (Ops.size() == 0) 681 return V; 682 683 // Add the new value to the list of things being added. 684 Ops.insert(Ops.begin(), ValueEntry(getRank(V), V)); 685 686 // Rewrite the tree so that there is now a use of V. 687 RewriteExprTree(I, 0, Ops); 688 return OptimizeExpression(I, Ops); 689 } 690 break; 691 //case Instruction::Mul: 692 } 693 694 if (IterateOptimization) 695 return OptimizeExpression(I, Ops); 696 return 0; 697} 698 699 700/// ReassociateBB - Inspect all of the instructions in this basic block, 701/// reassociating them as we go. 702void Reassociate::ReassociateBB(BasicBlock *BB) { 703 for (BasicBlock::iterator BBI = BB->begin(); BBI != BB->end(); ) { 704 Instruction *BI = BBI++; 705 if (BI->getOpcode() == Instruction::Shl && 706 isa<ConstantInt>(BI->getOperand(1))) 707 if (Instruction *NI = ConvertShiftToMul(BI)) { 708 MadeChange = true; 709 BI = NI; 710 } 711 712 // Reject cases where it is pointless to do this. 713 if (!isa<BinaryOperator>(BI) || BI->getType()->isFloatingPoint()) 714 continue; // Floating point ops are not associative. 715 716 // If this is a subtract instruction which is not already in negate form, 717 // see if we can convert it to X+-Y. 718 if (BI->getOpcode() == Instruction::Sub) { 719 if (!BinaryOperator::isNeg(BI)) { 720 if (Instruction *NI = BreakUpSubtract(BI)) { 721 MadeChange = true; 722 BI = NI; 723 } 724 } else { 725 // Otherwise, this is a negation. See if the operand is a multiply tree 726 // and if this is not an inner node of a multiply tree. 727 if (isReassociableOp(BI->getOperand(1), Instruction::Mul) && 728 (!BI->hasOneUse() || 729 !isReassociableOp(BI->use_back(), Instruction::Mul))) { 730 BI = LowerNegateToMultiply(BI); 731 MadeChange = true; 732 } 733 } 734 } 735 736 // If this instruction is a commutative binary operator, process it. 737 if (!BI->isAssociative()) continue; 738 BinaryOperator *I = cast<BinaryOperator>(BI); 739 740 // If this is an interior node of a reassociable tree, ignore it until we 741 // get to the root of the tree, to avoid N^2 analysis. 742 if (I->hasOneUse() && isReassociableOp(I->use_back(), I->getOpcode())) 743 continue; 744 745 // If this is an add tree that is used by a sub instruction, ignore it 746 // until we process the subtract. 747 if (I->hasOneUse() && I->getOpcode() == Instruction::Add && 748 cast<Instruction>(I->use_back())->getOpcode() == Instruction::Sub) 749 continue; 750 751 // First, walk the expression tree, linearizing the tree, collecting 752 std::vector<ValueEntry> Ops; 753 LinearizeExprTree(I, Ops); 754 755 DEBUG(std::cerr << "RAIn:\t"; PrintOps(I, Ops); 756 std::cerr << "\n"); 757 758 // Now that we have linearized the tree to a list and have gathered all of 759 // the operands and their ranks, sort the operands by their rank. Use a 760 // stable_sort so that values with equal ranks will have their relative 761 // positions maintained (and so the compiler is deterministic). Note that 762 // this sorts so that the highest ranking values end up at the beginning of 763 // the vector. 764 std::stable_sort(Ops.begin(), Ops.end()); 765 766 // OptimizeExpression - Now that we have the expression tree in a convenient 767 // sorted form, optimize it globally if possible. 768 if (Value *V = OptimizeExpression(I, Ops)) { 769 // This expression tree simplified to something that isn't a tree, 770 // eliminate it. 771 DEBUG(std::cerr << "Reassoc to scalar: " << *V << "\n"); 772 I->replaceAllUsesWith(V); 773 RemoveDeadBinaryOp(I); 774 continue; 775 } 776 777 // We want to sink immediates as deeply as possible except in the case where 778 // this is a multiply tree used only by an add, and the immediate is a -1. 779 // In this case we reassociate to put the negation on the outside so that we 780 // can fold the negation into the add: (-X)*Y + Z -> Z-X*Y 781 if (I->getOpcode() == Instruction::Mul && I->hasOneUse() && 782 cast<Instruction>(I->use_back())->getOpcode() == Instruction::Add && 783 isa<ConstantInt>(Ops.back().Op) && 784 cast<ConstantInt>(Ops.back().Op)->isAllOnesValue()) { 785 Ops.insert(Ops.begin(), Ops.back()); 786 Ops.pop_back(); 787 } 788 789 DEBUG(std::cerr << "RAOut:\t"; PrintOps(I, Ops); 790 std::cerr << "\n"); 791 792 if (Ops.size() == 1) { 793 // This expression tree simplified to something that isn't a tree, 794 // eliminate it. 795 I->replaceAllUsesWith(Ops[0].Op); 796 RemoveDeadBinaryOp(I); 797 } else { 798 // Now that we ordered and optimized the expressions, splat them back into 799 // the expression tree, removing any unneeded nodes. 800 RewriteExprTree(I, 0, Ops); 801 } 802 } 803} 804 805 806bool Reassociate::runOnFunction(Function &F) { 807 // Recalculate the rank map for F 808 BuildRankMap(F); 809 810 MadeChange = false; 811 for (Function::iterator FI = F.begin(), FE = F.end(); FI != FE; ++FI) 812 ReassociateBB(FI); 813 814 // We are done with the rank map... 815 RankMap.clear(); 816 ValueRankMap.clear(); 817 return MadeChange; 818} 819 820