Reassociate.cpp revision ec531233a16605756a84d175178e1ee0fac4791c
1//===- Reassociate.cpp - Reassociate binary expressions -------------------===// 2// 3// The LLVM Compiler Infrastructure 4// 5// This file is distributed under the University of Illinois Open Source 6// License. See LICENSE.TXT for details. 7// 8//===----------------------------------------------------------------------===// 9// 10// This pass reassociates commutative expressions in an order that is designed 11// to promote better constant propagation, GCSE, LICM, PRE... 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/IntrinsicInst.h" 30#include "llvm/Pass.h" 31#include "llvm/Assembly/Writer.h" 32#include "llvm/Support/CFG.h" 33#include "llvm/Support/Debug.h" 34#include "llvm/Support/ValueHandle.h" 35#include "llvm/Support/raw_ostream.h" 36#include "llvm/ADT/PostOrderIterator.h" 37#include "llvm/ADT/Statistic.h" 38#include "llvm/ADT/DenseMap.h" 39#include <algorithm> 40#include <map> 41using namespace llvm; 42 43STATISTIC(NumLinear , "Number of insts linearized"); 44STATISTIC(NumChanged, "Number of insts reassociated"); 45STATISTIC(NumAnnihil, "Number of expr tree annihilated"); 46STATISTIC(NumFactor , "Number of multiplies factored"); 47 48namespace { 49 struct ValueEntry { 50 unsigned Rank; 51 Value *Op; 52 ValueEntry(unsigned R, Value *O) : Rank(R), Op(O) {} 53 }; 54 inline bool operator<(const ValueEntry &LHS, const ValueEntry &RHS) { 55 return LHS.Rank > RHS.Rank; // Sort so that highest rank goes to start. 56 } 57} 58 59#ifndef NDEBUG 60/// PrintOps - Print out the expression identified in the Ops list. 61/// 62static void PrintOps(Instruction *I, const std::vector<ValueEntry> &Ops) { 63 Module *M = I->getParent()->getParent()->getParent(); 64 errs() << Instruction::getOpcodeName(I->getOpcode()) << " " 65 << *Ops[0].Op->getType() << '\t'; 66 for (unsigned i = 0, e = Ops.size(); i != e; ++i) { 67 errs() << "[ "; 68 WriteAsOperand(errs(), Ops[i].Op, false, M); 69 errs() << ", #" << Ops[i].Rank << "] "; 70 } 71} 72#endif 73 74namespace { 75 class Reassociate : public FunctionPass { 76 std::map<BasicBlock*, unsigned> RankMap; 77 std::map<AssertingVH<>, unsigned> ValueRankMap; 78 bool MadeChange; 79 public: 80 static char ID; // Pass identification, replacement for typeid 81 Reassociate() : FunctionPass(&ID) {} 82 83 bool runOnFunction(Function &F); 84 85 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 86 AU.setPreservesCFG(); 87 } 88 private: 89 void BuildRankMap(Function &F); 90 unsigned getRank(Value *V); 91 void ReassociateExpression(BinaryOperator *I); 92 void RewriteExprTree(BinaryOperator *I, std::vector<ValueEntry> &Ops, 93 unsigned Idx = 0); 94 Value *OptimizeExpression(BinaryOperator *I, std::vector<ValueEntry> &Ops); 95 void LinearizeExprTree(BinaryOperator *I, std::vector<ValueEntry> &Ops); 96 void LinearizeExpr(BinaryOperator *I); 97 Value *RemoveFactorFromExpression(Value *V, Value *Factor); 98 void ReassociateBB(BasicBlock *BB); 99 100 void RemoveDeadBinaryOp(Value *V); 101 }; 102} 103 104char Reassociate::ID = 0; 105static RegisterPass<Reassociate> X("reassociate", "Reassociate expressions"); 106 107// Public interface to the Reassociate pass 108FunctionPass *llvm::createReassociatePass() { return new Reassociate(); } 109 110void Reassociate::RemoveDeadBinaryOp(Value *V) { 111 Instruction *Op = dyn_cast<Instruction>(V); 112 if (!Op || !isa<BinaryOperator>(Op) || !isa<CmpInst>(Op) || !Op->use_empty()) 113 return; 114 115 Value *LHS = Op->getOperand(0), *RHS = Op->getOperand(1); 116 RemoveDeadBinaryOp(LHS); 117 RemoveDeadBinaryOp(RHS); 118} 119 120 121static bool isUnmovableInstruction(Instruction *I) { 122 if (I->getOpcode() == Instruction::PHI || 123 I->getOpcode() == Instruction::Alloca || 124 I->getOpcode() == Instruction::Load || 125 I->getOpcode() == Instruction::Invoke || 126 (I->getOpcode() == Instruction::Call && 127 !isa<DbgInfoIntrinsic>(I)) || 128 I->getOpcode() == Instruction::UDiv || 129 I->getOpcode() == Instruction::SDiv || 130 I->getOpcode() == Instruction::FDiv || 131 I->getOpcode() == Instruction::URem || 132 I->getOpcode() == Instruction::SRem || 133 I->getOpcode() == Instruction::FRem) 134 return true; 135 return false; 136} 137 138void Reassociate::BuildRankMap(Function &F) { 139 unsigned i = 2; 140 141 // Assign distinct ranks to function arguments 142 for (Function::arg_iterator I = F.arg_begin(), E = F.arg_end(); I != E; ++I) 143 ValueRankMap[&*I] = ++i; 144 145 ReversePostOrderTraversal<Function*> RPOT(&F); 146 for (ReversePostOrderTraversal<Function*>::rpo_iterator I = RPOT.begin(), 147 E = RPOT.end(); I != E; ++I) { 148 BasicBlock *BB = *I; 149 unsigned BBRank = RankMap[BB] = ++i << 16; 150 151 // Walk the basic block, adding precomputed ranks for any instructions that 152 // we cannot move. This ensures that the ranks for these instructions are 153 // all different in the block. 154 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) 155 if (isUnmovableInstruction(I)) 156 ValueRankMap[&*I] = ++BBRank; 157 } 158} 159 160unsigned Reassociate::getRank(Value *V) { 161 if (isa<Argument>(V)) return ValueRankMap[V]; // Function argument... 162 163 Instruction *I = dyn_cast<Instruction>(V); 164 if (I == 0) return 0; // Otherwise it's a global or constant, rank 0. 165 166 unsigned &CachedRank = ValueRankMap[I]; 167 if (CachedRank) return CachedRank; // Rank already known? 168 169 // If this is an expression, return the 1+MAX(rank(LHS), rank(RHS)) so that 170 // we can reassociate expressions for code motion! Since we do not recurse 171 // for PHI nodes, we cannot have infinite recursion here, because there 172 // cannot be loops in the value graph that do not go through PHI nodes. 173 unsigned Rank = 0, MaxRank = RankMap[I->getParent()]; 174 for (unsigned i = 0, e = I->getNumOperands(); 175 i != e && Rank != MaxRank; ++i) 176 Rank = std::max(Rank, getRank(I->getOperand(i))); 177 178 // If this is a not or neg instruction, do not count it for rank. This 179 // assures us that X and ~X will have the same rank. 180 if (!I->getType()->isInteger() || 181 (!BinaryOperator::isNot(I) && !BinaryOperator::isNeg(I))) 182 ++Rank; 183 184 //DEBUG(errs() << "Calculated Rank[" << V->getName() << "] = " 185 // << Rank << "\n"); 186 187 return CachedRank = Rank; 188} 189 190/// isReassociableOp - Return true if V is an instruction of the specified 191/// opcode and if it only has one use. 192static BinaryOperator *isReassociableOp(Value *V, unsigned Opcode) { 193 if ((V->hasOneUse() || V->use_empty()) && isa<Instruction>(V) && 194 cast<Instruction>(V)->getOpcode() == Opcode) 195 return cast<BinaryOperator>(V); 196 return 0; 197} 198 199/// LowerNegateToMultiply - Replace 0-X with X*-1. 200/// 201static Instruction *LowerNegateToMultiply(Instruction *Neg, 202 std::map<AssertingVH<>, unsigned> &ValueRankMap) { 203 Constant *Cst = Constant::getAllOnesValue(Neg->getType()); 204 205 Instruction *Res = BinaryOperator::CreateMul(Neg->getOperand(1), Cst, "",Neg); 206 ValueRankMap.erase(Neg); 207 Res->takeName(Neg); 208 Neg->replaceAllUsesWith(Res); 209 Neg->eraseFromParent(); 210 return Res; 211} 212 213// Given an expression of the form '(A+B)+(D+C)', turn it into '(((A+B)+C)+D)'. 214// Note that if D is also part of the expression tree that we recurse to 215// linearize it as well. Besides that case, this does not recurse into A,B, or 216// C. 217void Reassociate::LinearizeExpr(BinaryOperator *I) { 218 BinaryOperator *LHS = cast<BinaryOperator>(I->getOperand(0)); 219 BinaryOperator *RHS = cast<BinaryOperator>(I->getOperand(1)); 220 assert(isReassociableOp(LHS, I->getOpcode()) && 221 isReassociableOp(RHS, I->getOpcode()) && 222 "Not an expression that needs linearization?"); 223 224 DEBUG(errs() << "Linear" << *LHS << '\n' << *RHS << '\n' << *I << '\n'); 225 226 // Move the RHS instruction to live immediately before I, avoiding breaking 227 // dominator properties. 228 RHS->moveBefore(I); 229 230 // Move operands around to do the linearization. 231 I->setOperand(1, RHS->getOperand(0)); 232 RHS->setOperand(0, LHS); 233 I->setOperand(0, RHS); 234 235 ++NumLinear; 236 MadeChange = true; 237 DEBUG(errs() << "Linearized: " << *I << '\n'); 238 239 // If D is part of this expression tree, tail recurse. 240 if (isReassociableOp(I->getOperand(1), I->getOpcode())) 241 LinearizeExpr(I); 242} 243 244 245/// LinearizeExprTree - Given an associative binary expression tree, traverse 246/// all of the uses putting it into canonical form. This forces a left-linear 247/// form of the the expression (((a+b)+c)+d), and collects information about the 248/// rank of the non-tree operands. 249/// 250/// NOTE: These intentionally destroys the expression tree operands (turning 251/// them into undef values) to reduce #uses of the values. This means that the 252/// caller MUST use something like RewriteExprTree to put the values back in. 253/// 254void Reassociate::LinearizeExprTree(BinaryOperator *I, 255 std::vector<ValueEntry> &Ops) { 256 Value *LHS = I->getOperand(0), *RHS = I->getOperand(1); 257 unsigned Opcode = I->getOpcode(); 258 259 // First step, linearize the expression if it is in ((A+B)+(C+D)) form. 260 BinaryOperator *LHSBO = isReassociableOp(LHS, Opcode); 261 BinaryOperator *RHSBO = isReassociableOp(RHS, Opcode); 262 263 // If this is a multiply expression tree and it contains internal negations, 264 // transform them into multiplies by -1 so they can be reassociated. 265 if (I->getOpcode() == Instruction::Mul) { 266 if (!LHSBO && LHS->hasOneUse() && BinaryOperator::isNeg(LHS)) { 267 LHS = LowerNegateToMultiply(cast<Instruction>(LHS), ValueRankMap); 268 LHSBO = isReassociableOp(LHS, Opcode); 269 } 270 if (!RHSBO && RHS->hasOneUse() && BinaryOperator::isNeg(RHS)) { 271 RHS = LowerNegateToMultiply(cast<Instruction>(RHS), ValueRankMap); 272 RHSBO = isReassociableOp(RHS, Opcode); 273 } 274 } 275 276 if (!LHSBO) { 277 if (!RHSBO) { 278 // Neither the LHS or RHS as part of the tree, thus this is a leaf. As 279 // such, just remember these operands and their rank. 280 Ops.push_back(ValueEntry(getRank(LHS), LHS)); 281 Ops.push_back(ValueEntry(getRank(RHS), RHS)); 282 283 // Clear the leaves out. 284 I->setOperand(0, UndefValue::get(I->getType())); 285 I->setOperand(1, UndefValue::get(I->getType())); 286 return; 287 } else { 288 // Turn X+(Y+Z) -> (Y+Z)+X 289 std::swap(LHSBO, RHSBO); 290 std::swap(LHS, RHS); 291 bool Success = !I->swapOperands(); 292 assert(Success && "swapOperands failed"); 293 Success = false; 294 MadeChange = true; 295 } 296 } else if (RHSBO) { 297 // Turn (A+B)+(C+D) -> (((A+B)+C)+D). This guarantees the the RHS is not 298 // part of the expression tree. 299 LinearizeExpr(I); 300 LHS = LHSBO = cast<BinaryOperator>(I->getOperand(0)); 301 RHS = I->getOperand(1); 302 RHSBO = 0; 303 } 304 305 // Okay, now we know that the LHS is a nested expression and that the RHS is 306 // not. Perform reassociation. 307 assert(!isReassociableOp(RHS, Opcode) && "LinearizeExpr failed!"); 308 309 // Move LHS right before I to make sure that the tree expression dominates all 310 // values. 311 LHSBO->moveBefore(I); 312 313 // Linearize the expression tree on the LHS. 314 LinearizeExprTree(LHSBO, Ops); 315 316 // Remember the RHS operand and its rank. 317 Ops.push_back(ValueEntry(getRank(RHS), RHS)); 318 319 // Clear the RHS leaf out. 320 I->setOperand(1, UndefValue::get(I->getType())); 321} 322 323// RewriteExprTree - Now that the operands for this expression tree are 324// linearized and optimized, emit them in-order. This function is written to be 325// tail recursive. 326void Reassociate::RewriteExprTree(BinaryOperator *I, 327 std::vector<ValueEntry> &Ops, 328 unsigned i) { 329 if (i+2 == Ops.size()) { 330 if (I->getOperand(0) != Ops[i].Op || 331 I->getOperand(1) != Ops[i+1].Op) { 332 Value *OldLHS = I->getOperand(0); 333 DEBUG(errs() << "RA: " << *I << '\n'); 334 I->setOperand(0, Ops[i].Op); 335 I->setOperand(1, Ops[i+1].Op); 336 DEBUG(errs() << "TO: " << *I << '\n'); 337 MadeChange = true; 338 ++NumChanged; 339 340 // If we reassociated a tree to fewer operands (e.g. (1+a+2) -> (a+3) 341 // delete the extra, now dead, nodes. 342 RemoveDeadBinaryOp(OldLHS); 343 } 344 return; 345 } 346 assert(i+2 < Ops.size() && "Ops index out of range!"); 347 348 if (I->getOperand(1) != Ops[i].Op) { 349 DEBUG(errs() << "RA: " << *I << '\n'); 350 I->setOperand(1, Ops[i].Op); 351 DEBUG(errs() << "TO: " << *I << '\n'); 352 MadeChange = true; 353 ++NumChanged; 354 } 355 356 BinaryOperator *LHS = cast<BinaryOperator>(I->getOperand(0)); 357 assert(LHS->getOpcode() == I->getOpcode() && 358 "Improper expression tree!"); 359 360 // Compactify the tree instructions together with each other to guarantee 361 // that the expression tree is dominated by all of Ops. 362 LHS->moveBefore(I); 363 RewriteExprTree(LHS, Ops, i+1); 364} 365 366 367 368// NegateValue - Insert instructions before the instruction pointed to by BI, 369// that computes the negative version of the value specified. The negative 370// version of the value is returned, and BI is left pointing at the instruction 371// that should be processed next by the reassociation pass. 372// 373static Value *NegateValue(Value *V, Instruction *BI) { 374 // We are trying to expose opportunity for reassociation. One of the things 375 // that we want to do to achieve this is to push a negation as deep into an 376 // expression chain as possible, to expose the add instructions. In practice, 377 // this means that we turn this: 378 // X = -(A+12+C+D) into X = -A + -12 + -C + -D = -12 + -A + -C + -D 379 // so that later, a: Y = 12+X could get reassociated with the -12 to eliminate 380 // the constants. We assume that instcombine will clean up the mess later if 381 // we introduce tons of unnecessary negation instructions... 382 // 383 if (Instruction *I = dyn_cast<Instruction>(V)) 384 if (I->getOpcode() == Instruction::Add && I->hasOneUse()) { 385 // Push the negates through the add. 386 I->setOperand(0, NegateValue(I->getOperand(0), BI)); 387 I->setOperand(1, NegateValue(I->getOperand(1), BI)); 388 389 // We must move the add instruction here, because the neg instructions do 390 // not dominate the old add instruction in general. By moving it, we are 391 // assured that the neg instructions we just inserted dominate the 392 // instruction we are about to insert after them. 393 // 394 I->moveBefore(BI); 395 I->setName(I->getName()+".neg"); 396 return I; 397 } 398 399 // Insert a 'neg' instruction that subtracts the value from zero to get the 400 // negation. 401 // 402 return BinaryOperator::CreateNeg(V, V->getName() + ".neg", BI); 403} 404 405/// ShouldBreakUpSubtract - Return true if we should break up this subtract of 406/// X-Y into (X + -Y). 407static bool ShouldBreakUpSubtract(Instruction *Sub) { 408 // If this is a negation, we can't split it up! 409 if (BinaryOperator::isNeg(Sub)) 410 return false; 411 412 // Don't bother to break this up unless either the LHS is an associable add or 413 // subtract or if this is only used by one. 414 if (isReassociableOp(Sub->getOperand(0), Instruction::Add) || 415 isReassociableOp(Sub->getOperand(0), Instruction::Sub)) 416 return true; 417 if (isReassociableOp(Sub->getOperand(1), Instruction::Add) || 418 isReassociableOp(Sub->getOperand(1), Instruction::Sub)) 419 return true; 420 if (Sub->hasOneUse() && 421 (isReassociableOp(Sub->use_back(), Instruction::Add) || 422 isReassociableOp(Sub->use_back(), Instruction::Sub))) 423 return true; 424 425 return false; 426} 427 428/// BreakUpSubtract - If we have (X-Y), and if either X is an add, or if this is 429/// only used by an add, transform this into (X+(0-Y)) to promote better 430/// reassociation. 431static Instruction *BreakUpSubtract(Instruction *Sub, 432 std::map<AssertingVH<>, unsigned> &ValueRankMap) { 433 // Convert a subtract into an add and a neg instruction... so that sub 434 // instructions can be commuted with other add instructions... 435 // 436 // Calculate the negative value of Operand 1 of the sub instruction... 437 // and set it as the RHS of the add instruction we just made... 438 // 439 Value *NegVal = NegateValue(Sub->getOperand(1), Sub); 440 Instruction *New = 441 BinaryOperator::CreateAdd(Sub->getOperand(0), NegVal, "", Sub); 442 New->takeName(Sub); 443 444 // Everyone now refers to the add instruction. 445 ValueRankMap.erase(Sub); 446 Sub->replaceAllUsesWith(New); 447 Sub->eraseFromParent(); 448 449 DEBUG(errs() << "Negated: " << *New << '\n'); 450 return New; 451} 452 453/// ConvertShiftToMul - If this is a shift of a reassociable multiply or is used 454/// by one, change this into a multiply by a constant to assist with further 455/// reassociation. 456static Instruction *ConvertShiftToMul(Instruction *Shl, 457 std::map<AssertingVH<>, unsigned> &ValueRankMap) { 458 // If an operand of this shift is a reassociable multiply, or if the shift 459 // is used by a reassociable multiply or add, turn into a multiply. 460 if (isReassociableOp(Shl->getOperand(0), Instruction::Mul) || 461 (Shl->hasOneUse() && 462 (isReassociableOp(Shl->use_back(), Instruction::Mul) || 463 isReassociableOp(Shl->use_back(), Instruction::Add)))) { 464 Constant *MulCst = ConstantInt::get(Shl->getType(), 1); 465 MulCst = 466 ConstantExpr::getShl(MulCst, cast<Constant>(Shl->getOperand(1))); 467 468 Instruction *Mul = BinaryOperator::CreateMul(Shl->getOperand(0), MulCst, 469 "", Shl); 470 ValueRankMap.erase(Shl); 471 Mul->takeName(Shl); 472 Shl->replaceAllUsesWith(Mul); 473 Shl->eraseFromParent(); 474 return Mul; 475 } 476 return 0; 477} 478 479// Scan backwards and forwards among values with the same rank as element i to 480// see if X exists. If X does not exist, return i. 481static unsigned FindInOperandList(std::vector<ValueEntry> &Ops, unsigned i, 482 Value *X) { 483 unsigned XRank = Ops[i].Rank; 484 unsigned e = Ops.size(); 485 for (unsigned j = i+1; j != e && Ops[j].Rank == XRank; ++j) 486 if (Ops[j].Op == X) 487 return j; 488 // Scan backwards 489 for (unsigned j = i-1; j != ~0U && Ops[j].Rank == XRank; --j) 490 if (Ops[j].Op == X) 491 return j; 492 return i; 493} 494 495/// EmitAddTreeOfValues - Emit a tree of add instructions, summing Ops together 496/// and returning the result. Insert the tree before I. 497static Value *EmitAddTreeOfValues(Instruction *I, std::vector<Value*> &Ops) { 498 if (Ops.size() == 1) return Ops.back(); 499 500 Value *V1 = Ops.back(); 501 Ops.pop_back(); 502 Value *V2 = EmitAddTreeOfValues(I, Ops); 503 return BinaryOperator::CreateAdd(V2, V1, "tmp", I); 504} 505 506/// RemoveFactorFromExpression - If V is an expression tree that is a 507/// multiplication sequence, and if this sequence contains a multiply by Factor, 508/// remove Factor from the tree and return the new tree. 509Value *Reassociate::RemoveFactorFromExpression(Value *V, Value *Factor) { 510 BinaryOperator *BO = isReassociableOp(V, Instruction::Mul); 511 if (!BO) return 0; 512 513 std::vector<ValueEntry> Factors; 514 LinearizeExprTree(BO, Factors); 515 516 bool FoundFactor = false; 517 for (unsigned i = 0, e = Factors.size(); i != e; ++i) 518 if (Factors[i].Op == Factor) { 519 FoundFactor = true; 520 Factors.erase(Factors.begin()+i); 521 break; 522 } 523 if (!FoundFactor) { 524 // Make sure to restore the operands to the expression tree. 525 RewriteExprTree(BO, Factors); 526 return 0; 527 } 528 529 if (Factors.size() == 1) return Factors[0].Op; 530 531 RewriteExprTree(BO, Factors); 532 return BO; 533} 534 535/// FindSingleUseMultiplyFactors - If V is a single-use multiply, recursively 536/// add its operands as factors, otherwise add V to the list of factors. 537static void FindSingleUseMultiplyFactors(Value *V, 538 std::vector<Value*> &Factors) { 539 BinaryOperator *BO; 540 if ((!V->hasOneUse() && !V->use_empty()) || 541 !(BO = dyn_cast<BinaryOperator>(V)) || 542 BO->getOpcode() != Instruction::Mul) { 543 Factors.push_back(V); 544 return; 545 } 546 547 // Otherwise, add the LHS and RHS to the list of factors. 548 FindSingleUseMultiplyFactors(BO->getOperand(1), Factors); 549 FindSingleUseMultiplyFactors(BO->getOperand(0), Factors); 550} 551 552 553 554Value *Reassociate::OptimizeExpression(BinaryOperator *I, 555 std::vector<ValueEntry> &Ops) { 556 // Now that we have the linearized expression tree, try to optimize it. 557 // Start by folding any constants that we found. 558 bool IterateOptimization = false; 559 if (Ops.size() == 1) return Ops[0].Op; 560 561 unsigned Opcode = I->getOpcode(); 562 563 if (Constant *V1 = dyn_cast<Constant>(Ops[Ops.size()-2].Op)) 564 if (Constant *V2 = dyn_cast<Constant>(Ops.back().Op)) { 565 Ops.pop_back(); 566 Ops.back().Op = ConstantExpr::get(Opcode, V1, V2); 567 return OptimizeExpression(I, Ops); 568 } 569 570 // Check for destructive annihilation due to a constant being used. 571 if (ConstantInt *CstVal = dyn_cast<ConstantInt>(Ops.back().Op)) 572 switch (Opcode) { 573 default: break; 574 case Instruction::And: 575 if (CstVal->isZero()) { // ... & 0 -> 0 576 ++NumAnnihil; 577 return CstVal; 578 } else if (CstVal->isAllOnesValue()) { // ... & -1 -> ... 579 Ops.pop_back(); 580 } 581 break; 582 case Instruction::Mul: 583 if (CstVal->isZero()) { // ... * 0 -> 0 584 ++NumAnnihil; 585 return CstVal; 586 } else if (cast<ConstantInt>(CstVal)->isOne()) { 587 Ops.pop_back(); // ... * 1 -> ... 588 } 589 break; 590 case Instruction::Or: 591 if (CstVal->isAllOnesValue()) { // ... | -1 -> -1 592 ++NumAnnihil; 593 return CstVal; 594 } 595 // FALLTHROUGH! 596 case Instruction::Add: 597 case Instruction::Xor: 598 if (CstVal->isZero()) // ... [|^+] 0 -> ... 599 Ops.pop_back(); 600 break; 601 } 602 if (Ops.size() == 1) return Ops[0].Op; 603 604 // Handle destructive annihilation due to identities between elements in the 605 // argument list here. 606 switch (Opcode) { 607 default: break; 608 case Instruction::And: 609 case Instruction::Or: 610 case Instruction::Xor: 611 // Scan the operand lists looking for X and ~X pairs, along with X,X pairs. 612 // If we find any, we can simplify the expression. X&~X == 0, X|~X == -1. 613 for (unsigned i = 0, e = Ops.size(); i != e; ++i) { 614 // First, check for X and ~X in the operand list. 615 assert(i < Ops.size()); 616 if (BinaryOperator::isNot(Ops[i].Op)) { // Cannot occur for ^. 617 Value *X = BinaryOperator::getNotArgument(Ops[i].Op); 618 unsigned FoundX = FindInOperandList(Ops, i, X); 619 if (FoundX != i) { 620 if (Opcode == Instruction::And) { // ...&X&~X = 0 621 ++NumAnnihil; 622 return Constant::getNullValue(X->getType()); 623 } else if (Opcode == Instruction::Or) { // ...|X|~X = -1 624 ++NumAnnihil; 625 return Constant::getAllOnesValue(X->getType()); 626 } 627 } 628 } 629 630 // Next, check for duplicate pairs of values, which we assume are next to 631 // each other, due to our sorting criteria. 632 assert(i < Ops.size()); 633 if (i+1 != Ops.size() && Ops[i+1].Op == Ops[i].Op) { 634 if (Opcode == Instruction::And || Opcode == Instruction::Or) { 635 // Drop duplicate values. 636 Ops.erase(Ops.begin()+i); 637 --i; --e; 638 IterateOptimization = true; 639 ++NumAnnihil; 640 } else { 641 assert(Opcode == Instruction::Xor); 642 if (e == 2) { 643 ++NumAnnihil; 644 return Constant::getNullValue(Ops[0].Op->getType()); 645 } 646 // ... X^X -> ... 647 Ops.erase(Ops.begin()+i, Ops.begin()+i+2); 648 i -= 1; e -= 2; 649 IterateOptimization = true; 650 ++NumAnnihil; 651 } 652 } 653 } 654 break; 655 656 case Instruction::Add: 657 // Scan the operand lists looking for X and -X pairs. If we find any, we 658 // can simplify the expression. X+-X == 0. 659 for (unsigned i = 0, e = Ops.size(); i != e; ++i) { 660 assert(i < Ops.size()); 661 // Check for X and -X in the operand list. 662 if (BinaryOperator::isNeg(Ops[i].Op)) { 663 Value *X = BinaryOperator::getNegArgument(Ops[i].Op); 664 unsigned FoundX = FindInOperandList(Ops, i, X); 665 if (FoundX != i) { 666 // Remove X and -X from the operand list. 667 if (Ops.size() == 2) { 668 ++NumAnnihil; 669 return Constant::getNullValue(X->getType()); 670 } else { 671 Ops.erase(Ops.begin()+i); 672 if (i < FoundX) 673 --FoundX; 674 else 675 --i; // Need to back up an extra one. 676 Ops.erase(Ops.begin()+FoundX); 677 IterateOptimization = true; 678 ++NumAnnihil; 679 --i; // Revisit element. 680 e -= 2; // Removed two elements. 681 } 682 } 683 } 684 } 685 686 687 // Scan the operand list, checking to see if there are any common factors 688 // between operands. Consider something like A*A+A*B*C+D. We would like to 689 // reassociate this to A*(A+B*C)+D, which reduces the number of multiplies. 690 // To efficiently find this, we count the number of times a factor occurs 691 // for any ADD operands that are MULs. 692 DenseMap<Value*, unsigned> FactorOccurrences; 693 unsigned MaxOcc = 0; 694 Value *MaxOccVal = 0; 695 for (unsigned i = 0, e = Ops.size(); i != e; ++i) { 696 if (BinaryOperator *BOp = dyn_cast<BinaryOperator>(Ops[i].Op)) { 697 if (BOp->getOpcode() == Instruction::Mul && BOp->use_empty()) { 698 // Compute all of the factors of this added value. 699 std::vector<Value*> Factors; 700 FindSingleUseMultiplyFactors(BOp, Factors); 701 assert(Factors.size() > 1 && "Bad linearize!"); 702 703 // Add one to FactorOccurrences for each unique factor in this op. 704 if (Factors.size() == 2) { 705 unsigned Occ = ++FactorOccurrences[Factors[0]]; 706 if (Occ > MaxOcc) { MaxOcc = Occ; MaxOccVal = Factors[0]; } 707 if (Factors[0] != Factors[1]) { // Don't double count A*A. 708 Occ = ++FactorOccurrences[Factors[1]]; 709 if (Occ > MaxOcc) { MaxOcc = Occ; MaxOccVal = Factors[1]; } 710 } 711 } else { 712 SmallPtrSet<Value*, 4> Duplicates; 713 for (unsigned i = 0, e = Factors.size(); i != e; ++i) { 714 if (Duplicates.insert(Factors[i])) { 715 unsigned Occ = ++FactorOccurrences[Factors[i]]; 716 if (Occ > MaxOcc) { MaxOcc = Occ; MaxOccVal = Factors[i]; } 717 } 718 } 719 } 720 } 721 } 722 } 723 724 // If any factor occurred more than one time, we can pull it out. 725 if (MaxOcc > 1) { 726 DEBUG(errs() << "\nFACTORING [" << MaxOcc << "]: " << *MaxOccVal << "\n"); 727 728 // Create a new instruction that uses the MaxOccVal twice. If we don't do 729 // this, we could otherwise run into situations where removing a factor 730 // from an expression will drop a use of maxocc, and this can cause 731 // RemoveFactorFromExpression on successive values to behave differently. 732 Instruction *DummyInst = BinaryOperator::CreateAdd(MaxOccVal, MaxOccVal); 733 std::vector<Value*> NewMulOps; 734 for (unsigned i = 0, e = Ops.size(); i != e; ++i) { 735 if (Value *V = RemoveFactorFromExpression(Ops[i].Op, MaxOccVal)) { 736 NewMulOps.push_back(V); 737 Ops.erase(Ops.begin()+i); 738 --i; --e; 739 } 740 } 741 742 // No need for extra uses anymore. 743 delete DummyInst; 744 745 unsigned NumAddedValues = NewMulOps.size(); 746 Value *V = EmitAddTreeOfValues(I, NewMulOps); 747 Value *V2 = BinaryOperator::CreateMul(V, MaxOccVal, "tmp", I); 748 749 // Now that we have inserted V and its sole use, optimize it. This allows 750 // us to handle cases that require multiple factoring steps, such as this: 751 // A*A*B + A*A*C --> A*(A*B+A*C) --> A*(A*(B+C)) 752 if (NumAddedValues > 1) 753 ReassociateExpression(cast<BinaryOperator>(V)); 754 755 ++NumFactor; 756 757 if (Ops.empty()) 758 return V2; 759 760 // Add the new value to the list of things being added. 761 Ops.insert(Ops.begin(), ValueEntry(getRank(V2), V2)); 762 763 // Rewrite the tree so that there is now a use of V. 764 RewriteExprTree(I, Ops); 765 return OptimizeExpression(I, Ops); 766 } 767 break; 768 //case Instruction::Mul: 769 } 770 771 if (IterateOptimization) 772 return OptimizeExpression(I, Ops); 773 return 0; 774} 775 776 777/// ReassociateBB - Inspect all of the instructions in this basic block, 778/// reassociating them as we go. 779void Reassociate::ReassociateBB(BasicBlock *BB) { 780 for (BasicBlock::iterator BBI = BB->begin(); BBI != BB->end(); ) { 781 Instruction *BI = BBI++; 782 if (BI->getOpcode() == Instruction::Shl && 783 isa<ConstantInt>(BI->getOperand(1))) 784 if (Instruction *NI = ConvertShiftToMul(BI, ValueRankMap)) { 785 MadeChange = true; 786 BI = NI; 787 } 788 789 // Reject cases where it is pointless to do this. 790 if (!isa<BinaryOperator>(BI) || BI->getType()->isFloatingPoint() || 791 isa<VectorType>(BI->getType())) 792 continue; // Floating point ops are not associative. 793 794 // If this is a subtract instruction which is not already in negate form, 795 // see if we can convert it to X+-Y. 796 if (BI->getOpcode() == Instruction::Sub) { 797 if (ShouldBreakUpSubtract(BI)) { 798 BI = BreakUpSubtract(BI, ValueRankMap); 799 MadeChange = true; 800 } else if (BinaryOperator::isNeg(BI)) { 801 // Otherwise, this is a negation. See if the operand is a multiply tree 802 // and if this is not an inner node of a multiply tree. 803 if (isReassociableOp(BI->getOperand(1), Instruction::Mul) && 804 (!BI->hasOneUse() || 805 !isReassociableOp(BI->use_back(), Instruction::Mul))) { 806 BI = LowerNegateToMultiply(BI, ValueRankMap); 807 MadeChange = true; 808 } 809 } 810 } 811 812 // If this instruction is a commutative binary operator, process it. 813 if (!BI->isAssociative()) continue; 814 BinaryOperator *I = cast<BinaryOperator>(BI); 815 816 // If this is an interior node of a reassociable tree, ignore it until we 817 // get to the root of the tree, to avoid N^2 analysis. 818 if (I->hasOneUse() && isReassociableOp(I->use_back(), I->getOpcode())) 819 continue; 820 821 // If this is an add tree that is used by a sub instruction, ignore it 822 // until we process the subtract. 823 if (I->hasOneUse() && I->getOpcode() == Instruction::Add && 824 cast<Instruction>(I->use_back())->getOpcode() == Instruction::Sub) 825 continue; 826 827 ReassociateExpression(I); 828 } 829} 830 831void Reassociate::ReassociateExpression(BinaryOperator *I) { 832 833 // First, walk the expression tree, linearizing the tree, collecting 834 std::vector<ValueEntry> Ops; 835 LinearizeExprTree(I, Ops); 836 837 DEBUG(errs() << "RAIn:\t"; PrintOps(I, Ops); errs() << "\n"); 838 839 // Now that we have linearized the tree to a list and have gathered all of 840 // the operands and their ranks, sort the operands by their rank. Use a 841 // stable_sort so that values with equal ranks will have their relative 842 // positions maintained (and so the compiler is deterministic). Note that 843 // this sorts so that the highest ranking values end up at the beginning of 844 // the vector. 845 std::stable_sort(Ops.begin(), Ops.end()); 846 847 // OptimizeExpression - Now that we have the expression tree in a convenient 848 // sorted form, optimize it globally if possible. 849 if (Value *V = OptimizeExpression(I, Ops)) { 850 // This expression tree simplified to something that isn't a tree, 851 // eliminate it. 852 DEBUG(errs() << "Reassoc to scalar: " << *V << "\n"); 853 I->replaceAllUsesWith(V); 854 RemoveDeadBinaryOp(I); 855 return; 856 } 857 858 // We want to sink immediates as deeply as possible except in the case where 859 // this is a multiply tree used only by an add, and the immediate is a -1. 860 // In this case we reassociate to put the negation on the outside so that we 861 // can fold the negation into the add: (-X)*Y + Z -> Z-X*Y 862 if (I->getOpcode() == Instruction::Mul && I->hasOneUse() && 863 cast<Instruction>(I->use_back())->getOpcode() == Instruction::Add && 864 isa<ConstantInt>(Ops.back().Op) && 865 cast<ConstantInt>(Ops.back().Op)->isAllOnesValue()) { 866 Ops.insert(Ops.begin(), Ops.back()); 867 Ops.pop_back(); 868 } 869 870 DEBUG(errs() << "RAOut:\t"; PrintOps(I, Ops); errs() << "\n"); 871 872 if (Ops.size() == 1) { 873 // This expression tree simplified to something that isn't a tree, 874 // eliminate it. 875 I->replaceAllUsesWith(Ops[0].Op); 876 RemoveDeadBinaryOp(I); 877 } else { 878 // Now that we ordered and optimized the expressions, splat them back into 879 // the expression tree, removing any unneeded nodes. 880 RewriteExprTree(I, Ops); 881 } 882} 883 884 885bool Reassociate::runOnFunction(Function &F) { 886 // Recalculate the rank map for F 887 BuildRankMap(F); 888 889 MadeChange = false; 890 for (Function::iterator FI = F.begin(), FE = F.end(); FI != FE; ++FI) 891 ReassociateBB(FI); 892 893 // We are done with the rank map... 894 RankMap.clear(); 895 ValueRankMap.clear(); 896 return MadeChange; 897} 898 899