Reassociate.cpp revision ae73dc1448d25b02cabc7c64c86c64371453dda8
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/Pass.h" 30#include "llvm/Assembly/Writer.h" 31#include "llvm/Support/CFG.h" 32#include "llvm/Support/Compiler.h" 33#include "llvm/Support/Debug.h" 34#include "llvm/ADT/PostOrderIterator.h" 35#include "llvm/ADT/Statistic.h" 36#include <algorithm> 37#include <map> 38using namespace llvm; 39 40STATISTIC(NumLinear , "Number of insts linearized"); 41STATISTIC(NumChanged, "Number of insts reassociated"); 42STATISTIC(NumAnnihil, "Number of expr tree annihilated"); 43STATISTIC(NumFactor , "Number of multiplies factored"); 44 45namespace { 46 struct VISIBILITY_HIDDEN 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 cerr << Instruction::getOpcodeName(I->getOpcode()) << " " 61 << *Ops[0].Op->getType(); 62 for (unsigned i = 0, e = Ops.size(); i != e; ++i) { 63 WriteAsOperand(*cerr.stream() << " ", Ops[i].Op, false, M); 64 cerr << "," << Ops[i].Rank; 65 } 66} 67 68namespace { 69 class VISIBILITY_HIDDEN Reassociate : public FunctionPass { 70 std::map<BasicBlock*, unsigned> RankMap; 71 std::map<Value*, unsigned> ValueRankMap; 72 bool MadeChange; 73 public: 74 static char ID; // Pass identification, replacement for typeid 75 Reassociate() : FunctionPass(&ID) {} 76 77 bool runOnFunction(Function &F); 78 79 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 80 AU.setPreservesCFG(); 81 } 82 private: 83 void BuildRankMap(Function &F); 84 unsigned getRank(Value *V); 85 void ReassociateExpression(BinaryOperator *I); 86 void RewriteExprTree(BinaryOperator *I, std::vector<ValueEntry> &Ops, 87 unsigned Idx = 0); 88 Value *OptimizeExpression(BinaryOperator *I, std::vector<ValueEntry> &Ops); 89 void LinearizeExprTree(BinaryOperator *I, std::vector<ValueEntry> &Ops); 90 void LinearizeExpr(BinaryOperator *I); 91 Value *RemoveFactorFromExpression(Value *V, Value *Factor); 92 void ReassociateBB(BasicBlock *BB); 93 94 void RemoveDeadBinaryOp(Value *V); 95 }; 96} 97 98char Reassociate::ID = 0; 99static RegisterPass<Reassociate> X("reassociate", "Reassociate expressions"); 100 101// Public interface to the Reassociate pass 102FunctionPass *llvm::createReassociatePass() { return new Reassociate(); } 103 104void Reassociate::RemoveDeadBinaryOp(Value *V) { 105 Instruction *Op = dyn_cast<Instruction>(V); 106 if (!Op || !isa<BinaryOperator>(Op) || !isa<CmpInst>(Op) || !Op->use_empty()) 107 return; 108 109 Value *LHS = Op->getOperand(0), *RHS = Op->getOperand(1); 110 RemoveDeadBinaryOp(LHS); 111 RemoveDeadBinaryOp(RHS); 112} 113 114 115static bool isUnmovableInstruction(Instruction *I) { 116 if (I->getOpcode() == Instruction::PHI || 117 I->getOpcode() == Instruction::Alloca || 118 I->getOpcode() == Instruction::Load || 119 I->getOpcode() == Instruction::Malloc || 120 I->getOpcode() == Instruction::Invoke || 121 I->getOpcode() == Instruction::Call || 122 I->getOpcode() == Instruction::UDiv || 123 I->getOpcode() == Instruction::SDiv || 124 I->getOpcode() == Instruction::FDiv || 125 I->getOpcode() == Instruction::URem || 126 I->getOpcode() == Instruction::SRem || 127 I->getOpcode() == Instruction::FRem) 128 return true; 129 return false; 130} 131 132void Reassociate::BuildRankMap(Function &F) { 133 unsigned i = 2; 134 135 // Assign distinct ranks to function arguments 136 for (Function::arg_iterator I = F.arg_begin(), E = F.arg_end(); I != E; ++I) 137 ValueRankMap[I] = ++i; 138 139 ReversePostOrderTraversal<Function*> RPOT(&F); 140 for (ReversePostOrderTraversal<Function*>::rpo_iterator I = RPOT.begin(), 141 E = RPOT.end(); I != E; ++I) { 142 BasicBlock *BB = *I; 143 unsigned BBRank = RankMap[BB] = ++i << 16; 144 145 // Walk the basic block, adding precomputed ranks for any instructions that 146 // we cannot move. This ensures that the ranks for these instructions are 147 // all different in the block. 148 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) 149 if (isUnmovableInstruction(I)) 150 ValueRankMap[I] = ++BBRank; 151 } 152} 153 154unsigned Reassociate::getRank(Value *V) { 155 if (isa<Argument>(V)) return ValueRankMap[V]; // Function argument... 156 157 Instruction *I = dyn_cast<Instruction>(V); 158 if (I == 0) return 0; // Otherwise it's a global or constant, rank 0. 159 160 unsigned &CachedRank = ValueRankMap[I]; 161 if (CachedRank) return CachedRank; // Rank already known? 162 163 // If this is an expression, return the 1+MAX(rank(LHS), rank(RHS)) so that 164 // we can reassociate expressions for code motion! Since we do not recurse 165 // for PHI nodes, we cannot have infinite recursion here, because there 166 // cannot be loops in the value graph that do not go through PHI nodes. 167 unsigned Rank = 0, MaxRank = RankMap[I->getParent()]; 168 for (unsigned i = 0, e = I->getNumOperands(); 169 i != e && Rank != MaxRank; ++i) 170 Rank = std::max(Rank, getRank(I->getOperand(i))); 171 172 // If this is a not or neg instruction, do not count it for rank. This 173 // assures us that X and ~X will have the same rank. 174 if (!I->getType()->isInteger() || 175 (!BinaryOperator::isNot(I) && !BinaryOperator::isNeg(I))) 176 ++Rank; 177 178 //DOUT << "Calculated Rank[" << V->getName() << "] = " 179 // << Rank << "\n"; 180 181 return CachedRank = Rank; 182} 183 184/// isReassociableOp - Return true if V is an instruction of the specified 185/// opcode and if it only has one use. 186static BinaryOperator *isReassociableOp(Value *V, unsigned Opcode) { 187 if ((V->hasOneUse() || V->use_empty()) && isa<Instruction>(V) && 188 cast<Instruction>(V)->getOpcode() == Opcode) 189 return cast<BinaryOperator>(V); 190 return 0; 191} 192 193/// LowerNegateToMultiply - Replace 0-X with X*-1. 194/// 195static Instruction *LowerNegateToMultiply(Instruction *Neg) { 196 Constant *Cst = ConstantInt::getAllOnesValue(Neg->getType()); 197 198 Instruction *Res = BinaryOperator::CreateMul(Neg->getOperand(1), Cst, "",Neg); 199 Res->takeName(Neg); 200 Neg->replaceAllUsesWith(Res); 201 Neg->eraseFromParent(); 202 return Res; 203} 204 205// Given an expression of the form '(A+B)+(D+C)', turn it into '(((A+B)+C)+D)'. 206// Note that if D is also part of the expression tree that we recurse to 207// linearize it as well. Besides that case, this does not recurse into A,B, or 208// C. 209void Reassociate::LinearizeExpr(BinaryOperator *I) { 210 BinaryOperator *LHS = cast<BinaryOperator>(I->getOperand(0)); 211 BinaryOperator *RHS = cast<BinaryOperator>(I->getOperand(1)); 212 assert(isReassociableOp(LHS, I->getOpcode()) && 213 isReassociableOp(RHS, I->getOpcode()) && 214 "Not an expression that needs linearization?"); 215 216 DOUT << "Linear" << *LHS << *RHS << *I; 217 218 // Move the RHS instruction to live immediately before I, avoiding breaking 219 // dominator properties. 220 RHS->moveBefore(I); 221 222 // Move operands around to do the linearization. 223 I->setOperand(1, RHS->getOperand(0)); 224 RHS->setOperand(0, LHS); 225 I->setOperand(0, RHS); 226 227 ++NumLinear; 228 MadeChange = true; 229 DOUT << "Linearized: " << *I; 230 231 // If D is part of this expression tree, tail recurse. 232 if (isReassociableOp(I->getOperand(1), I->getOpcode())) 233 LinearizeExpr(I); 234} 235 236 237/// LinearizeExprTree - Given an associative binary expression tree, traverse 238/// all of the uses putting it into canonical form. This forces a left-linear 239/// form of the the expression (((a+b)+c)+d), and collects information about the 240/// rank of the non-tree operands. 241/// 242/// NOTE: These intentionally destroys the expression tree operands (turning 243/// them into undef values) to reduce #uses of the values. This means that the 244/// caller MUST use something like RewriteExprTree to put the values back in. 245/// 246void Reassociate::LinearizeExprTree(BinaryOperator *I, 247 std::vector<ValueEntry> &Ops) { 248 Value *LHS = I->getOperand(0), *RHS = I->getOperand(1); 249 unsigned Opcode = I->getOpcode(); 250 251 // First step, linearize the expression if it is in ((A+B)+(C+D)) form. 252 BinaryOperator *LHSBO = isReassociableOp(LHS, Opcode); 253 BinaryOperator *RHSBO = isReassociableOp(RHS, Opcode); 254 255 // If this is a multiply expression tree and it contains internal negations, 256 // transform them into multiplies by -1 so they can be reassociated. 257 if (I->getOpcode() == Instruction::Mul) { 258 if (!LHSBO && LHS->hasOneUse() && BinaryOperator::isNeg(LHS)) { 259 LHS = LowerNegateToMultiply(cast<Instruction>(LHS)); 260 LHSBO = isReassociableOp(LHS, Opcode); 261 } 262 if (!RHSBO && RHS->hasOneUse() && BinaryOperator::isNeg(RHS)) { 263 RHS = LowerNegateToMultiply(cast<Instruction>(RHS)); 264 RHSBO = isReassociableOp(RHS, Opcode); 265 } 266 } 267 268 if (!LHSBO) { 269 if (!RHSBO) { 270 // Neither the LHS or RHS as part of the tree, thus this is a leaf. As 271 // such, just remember these operands and their rank. 272 Ops.push_back(ValueEntry(getRank(LHS), LHS)); 273 Ops.push_back(ValueEntry(getRank(RHS), RHS)); 274 275 // Clear the leaves out. 276 I->setOperand(0, UndefValue::get(I->getType())); 277 I->setOperand(1, UndefValue::get(I->getType())); 278 return; 279 } else { 280 // Turn X+(Y+Z) -> (Y+Z)+X 281 std::swap(LHSBO, RHSBO); 282 std::swap(LHS, RHS); 283 bool Success = !I->swapOperands(); 284 assert(Success && "swapOperands failed"); 285 MadeChange = true; 286 } 287 } else if (RHSBO) { 288 // Turn (A+B)+(C+D) -> (((A+B)+C)+D). This guarantees the the RHS is not 289 // part of the expression tree. 290 LinearizeExpr(I); 291 LHS = LHSBO = cast<BinaryOperator>(I->getOperand(0)); 292 RHS = I->getOperand(1); 293 RHSBO = 0; 294 } 295 296 // Okay, now we know that the LHS is a nested expression and that the RHS is 297 // not. Perform reassociation. 298 assert(!isReassociableOp(RHS, Opcode) && "LinearizeExpr failed!"); 299 300 // Move LHS right before I to make sure that the tree expression dominates all 301 // values. 302 LHSBO->moveBefore(I); 303 304 // Linearize the expression tree on the LHS. 305 LinearizeExprTree(LHSBO, Ops); 306 307 // Remember the RHS operand and its rank. 308 Ops.push_back(ValueEntry(getRank(RHS), RHS)); 309 310 // Clear the RHS leaf out. 311 I->setOperand(1, UndefValue::get(I->getType())); 312} 313 314// RewriteExprTree - Now that the operands for this expression tree are 315// linearized and optimized, emit them in-order. This function is written to be 316// tail recursive. 317void Reassociate::RewriteExprTree(BinaryOperator *I, 318 std::vector<ValueEntry> &Ops, 319 unsigned i) { 320 if (i+2 == Ops.size()) { 321 if (I->getOperand(0) != Ops[i].Op || 322 I->getOperand(1) != Ops[i+1].Op) { 323 Value *OldLHS = I->getOperand(0); 324 DOUT << "RA: " << *I; 325 I->setOperand(0, Ops[i].Op); 326 I->setOperand(1, Ops[i+1].Op); 327 DOUT << "TO: " << *I; 328 MadeChange = true; 329 ++NumChanged; 330 331 // If we reassociated a tree to fewer operands (e.g. (1+a+2) -> (a+3) 332 // delete the extra, now dead, nodes. 333 RemoveDeadBinaryOp(OldLHS); 334 } 335 return; 336 } 337 assert(i+2 < Ops.size() && "Ops index out of range!"); 338 339 if (I->getOperand(1) != Ops[i].Op) { 340 DOUT << "RA: " << *I; 341 I->setOperand(1, Ops[i].Op); 342 DOUT << "TO: " << *I; 343 MadeChange = true; 344 ++NumChanged; 345 } 346 347 BinaryOperator *LHS = cast<BinaryOperator>(I->getOperand(0)); 348 assert(LHS->getOpcode() == I->getOpcode() && 349 "Improper expression tree!"); 350 351 // Compactify the tree instructions together with each other to guarantee 352 // that the expression tree is dominated by all of Ops. 353 LHS->moveBefore(I); 354 RewriteExprTree(LHS, Ops, i+1); 355} 356 357 358 359// NegateValue - Insert instructions before the instruction pointed to by BI, 360// that computes the negative version of the value specified. The negative 361// version of the value is returned, and BI is left pointing at the instruction 362// that should be processed next by the reassociation pass. 363// 364static Value *NegateValue(Value *V, Instruction *BI) { 365 // We are trying to expose opportunity for reassociation. One of the things 366 // that we want to do to achieve this is to push a negation as deep into an 367 // expression chain as possible, to expose the add instructions. In practice, 368 // this means that we turn this: 369 // X = -(A+12+C+D) into X = -A + -12 + -C + -D = -12 + -A + -C + -D 370 // so that later, a: Y = 12+X could get reassociated with the -12 to eliminate 371 // the constants. We assume that instcombine will clean up the mess later if 372 // we introduce tons of unnecessary negation instructions... 373 // 374 if (Instruction *I = dyn_cast<Instruction>(V)) 375 if (I->getOpcode() == Instruction::Add && I->hasOneUse()) { 376 // Push the negates through the add. 377 I->setOperand(0, NegateValue(I->getOperand(0), BI)); 378 I->setOperand(1, NegateValue(I->getOperand(1), BI)); 379 380 // We must move the add instruction here, because the neg instructions do 381 // not dominate the old add instruction in general. By moving it, we are 382 // assured that the neg instructions we just inserted dominate the 383 // instruction we are about to insert after them. 384 // 385 I->moveBefore(BI); 386 I->setName(I->getName()+".neg"); 387 return I; 388 } 389 390 // Insert a 'neg' instruction that subtracts the value from zero to get the 391 // negation. 392 // 393 return BinaryOperator::CreateNeg(V, V->getName() + ".neg", BI); 394} 395 396/// ShouldBreakUpSubtract - Return true if we should break up this subtract of 397/// X-Y into (X + -Y). 398static bool ShouldBreakUpSubtract(Instruction *Sub) { 399 // If this is a negation, we can't split it up! 400 if (BinaryOperator::isNeg(Sub)) 401 return false; 402 403 // Don't bother to break this up unless either the LHS is an associable add or 404 // subtract or if this is only used by one. 405 if (isReassociableOp(Sub->getOperand(0), Instruction::Add) || 406 isReassociableOp(Sub->getOperand(0), Instruction::Sub)) 407 return true; 408 if (isReassociableOp(Sub->getOperand(1), Instruction::Add) || 409 isReassociableOp(Sub->getOperand(1), Instruction::Sub)) 410 return true; 411 if (Sub->hasOneUse() && 412 (isReassociableOp(Sub->use_back(), Instruction::Add) || 413 isReassociableOp(Sub->use_back(), Instruction::Sub))) 414 return true; 415 416 return false; 417} 418 419/// BreakUpSubtract - If we have (X-Y), and if either X is an add, or if this is 420/// only used by an add, transform this into (X+(0-Y)) to promote better 421/// reassociation. 422static Instruction *BreakUpSubtract(Instruction *Sub) { 423 // Convert a subtract into an add and a neg instruction... so that sub 424 // instructions can be commuted with other add instructions... 425 // 426 // Calculate the negative value of Operand 1 of the sub instruction... 427 // and set it as the RHS of the add instruction we just made... 428 // 429 Value *NegVal = NegateValue(Sub->getOperand(1), Sub); 430 Instruction *New = 431 BinaryOperator::CreateAdd(Sub->getOperand(0), NegVal, "", Sub); 432 New->takeName(Sub); 433 434 // Everyone now refers to the add instruction. 435 Sub->replaceAllUsesWith(New); 436 Sub->eraseFromParent(); 437 438 DOUT << "Negated: " << *New; 439 return New; 440} 441 442/// ConvertShiftToMul - If this is a shift of a reassociable multiply or is used 443/// by one, change this into a multiply by a constant to assist with further 444/// reassociation. 445static Instruction *ConvertShiftToMul(Instruction *Shl) { 446 // If an operand of this shift is a reassociable multiply, or if the shift 447 // is used by a reassociable multiply or add, turn into a multiply. 448 if (isReassociableOp(Shl->getOperand(0), Instruction::Mul) || 449 (Shl->hasOneUse() && 450 (isReassociableOp(Shl->use_back(), Instruction::Mul) || 451 isReassociableOp(Shl->use_back(), Instruction::Add)))) { 452 Constant *MulCst = ConstantInt::get(Shl->getType(), 1); 453 MulCst = ConstantExpr::getShl(MulCst, cast<Constant>(Shl->getOperand(1))); 454 455 Instruction *Mul = BinaryOperator::CreateMul(Shl->getOperand(0), MulCst, 456 "", Shl); 457 Mul->takeName(Shl); 458 Shl->replaceAllUsesWith(Mul); 459 Shl->eraseFromParent(); 460 return Mul; 461 } 462 return 0; 463} 464 465// Scan backwards and forwards among values with the same rank as element i to 466// see if X exists. If X does not exist, return i. 467static unsigned FindInOperandList(std::vector<ValueEntry> &Ops, unsigned i, 468 Value *X) { 469 unsigned XRank = Ops[i].Rank; 470 unsigned e = Ops.size(); 471 for (unsigned j = i+1; j != e && Ops[j].Rank == XRank; ++j) 472 if (Ops[j].Op == X) 473 return j; 474 // Scan backwards 475 for (unsigned j = i-1; j != ~0U && Ops[j].Rank == XRank; --j) 476 if (Ops[j].Op == X) 477 return j; 478 return i; 479} 480 481/// EmitAddTreeOfValues - Emit a tree of add instructions, summing Ops together 482/// and returning the result. Insert the tree before I. 483static Value *EmitAddTreeOfValues(Instruction *I, std::vector<Value*> &Ops) { 484 if (Ops.size() == 1) return Ops.back(); 485 486 Value *V1 = Ops.back(); 487 Ops.pop_back(); 488 Value *V2 = EmitAddTreeOfValues(I, Ops); 489 return BinaryOperator::CreateAdd(V2, V1, "tmp", I); 490} 491 492/// RemoveFactorFromExpression - If V is an expression tree that is a 493/// multiplication sequence, and if this sequence contains a multiply by Factor, 494/// remove Factor from the tree and return the new tree. 495Value *Reassociate::RemoveFactorFromExpression(Value *V, Value *Factor) { 496 BinaryOperator *BO = isReassociableOp(V, Instruction::Mul); 497 if (!BO) return 0; 498 499 std::vector<ValueEntry> Factors; 500 LinearizeExprTree(BO, Factors); 501 502 bool FoundFactor = false; 503 for (unsigned i = 0, e = Factors.size(); i != e; ++i) 504 if (Factors[i].Op == Factor) { 505 FoundFactor = true; 506 Factors.erase(Factors.begin()+i); 507 break; 508 } 509 if (!FoundFactor) { 510 // Make sure to restore the operands to the expression tree. 511 RewriteExprTree(BO, Factors); 512 return 0; 513 } 514 515 if (Factors.size() == 1) return Factors[0].Op; 516 517 RewriteExprTree(BO, Factors); 518 return BO; 519} 520 521/// FindSingleUseMultiplyFactors - If V is a single-use multiply, recursively 522/// add its operands as factors, otherwise add V to the list of factors. 523static void FindSingleUseMultiplyFactors(Value *V, 524 std::vector<Value*> &Factors) { 525 BinaryOperator *BO; 526 if ((!V->hasOneUse() && !V->use_empty()) || 527 !(BO = dyn_cast<BinaryOperator>(V)) || 528 BO->getOpcode() != Instruction::Mul) { 529 Factors.push_back(V); 530 return; 531 } 532 533 // Otherwise, add the LHS and RHS to the list of factors. 534 FindSingleUseMultiplyFactors(BO->getOperand(1), Factors); 535 FindSingleUseMultiplyFactors(BO->getOperand(0), Factors); 536} 537 538 539 540Value *Reassociate::OptimizeExpression(BinaryOperator *I, 541 std::vector<ValueEntry> &Ops) { 542 // Now that we have the linearized expression tree, try to optimize it. 543 // Start by folding any constants that we found. 544 bool IterateOptimization = false; 545 if (Ops.size() == 1) return Ops[0].Op; 546 547 unsigned Opcode = I->getOpcode(); 548 549 if (Constant *V1 = dyn_cast<Constant>(Ops[Ops.size()-2].Op)) 550 if (Constant *V2 = dyn_cast<Constant>(Ops.back().Op)) { 551 Ops.pop_back(); 552 Ops.back().Op = ConstantExpr::get(Opcode, V1, V2); 553 return OptimizeExpression(I, Ops); 554 } 555 556 // Check for destructive annihilation due to a constant being used. 557 if (ConstantInt *CstVal = dyn_cast<ConstantInt>(Ops.back().Op)) 558 switch (Opcode) { 559 default: break; 560 case Instruction::And: 561 if (CstVal->isZero()) { // ... & 0 -> 0 562 ++NumAnnihil; 563 return CstVal; 564 } else if (CstVal->isAllOnesValue()) { // ... & -1 -> ... 565 Ops.pop_back(); 566 } 567 break; 568 case Instruction::Mul: 569 if (CstVal->isZero()) { // ... * 0 -> 0 570 ++NumAnnihil; 571 return CstVal; 572 } else if (cast<ConstantInt>(CstVal)->isOne()) { 573 Ops.pop_back(); // ... * 1 -> ... 574 } 575 break; 576 case Instruction::Or: 577 if (CstVal->isAllOnesValue()) { // ... | -1 -> -1 578 ++NumAnnihil; 579 return CstVal; 580 } 581 // FALLTHROUGH! 582 case Instruction::Add: 583 case Instruction::Xor: 584 if (CstVal->isZero()) // ... [|^+] 0 -> ... 585 Ops.pop_back(); 586 break; 587 } 588 if (Ops.size() == 1) return Ops[0].Op; 589 590 // Handle destructive annihilation do to identities between elements in the 591 // argument list here. 592 switch (Opcode) { 593 default: break; 594 case Instruction::And: 595 case Instruction::Or: 596 case Instruction::Xor: 597 // Scan the operand lists looking for X and ~X pairs, along with X,X pairs. 598 // If we find any, we can simplify the expression. X&~X == 0, X|~X == -1. 599 for (unsigned i = 0, e = Ops.size(); i != e; ++i) { 600 // First, check for X and ~X in the operand list. 601 assert(i < Ops.size()); 602 if (BinaryOperator::isNot(Ops[i].Op)) { // Cannot occur for ^. 603 Value *X = BinaryOperator::getNotArgument(Ops[i].Op); 604 unsigned FoundX = FindInOperandList(Ops, i, X); 605 if (FoundX != i) { 606 if (Opcode == Instruction::And) { // ...&X&~X = 0 607 ++NumAnnihil; 608 return Constant::getNullValue(X->getType()); 609 } else if (Opcode == Instruction::Or) { // ...|X|~X = -1 610 ++NumAnnihil; 611 return ConstantInt::getAllOnesValue(X->getType()); 612 } 613 } 614 } 615 616 // Next, check for duplicate pairs of values, which we assume are next to 617 // each other, due to our sorting criteria. 618 assert(i < Ops.size()); 619 if (i+1 != Ops.size() && Ops[i+1].Op == Ops[i].Op) { 620 if (Opcode == Instruction::And || Opcode == Instruction::Or) { 621 // Drop duplicate values. 622 Ops.erase(Ops.begin()+i); 623 --i; --e; 624 IterateOptimization = true; 625 ++NumAnnihil; 626 } else { 627 assert(Opcode == Instruction::Xor); 628 if (e == 2) { 629 ++NumAnnihil; 630 return Constant::getNullValue(Ops[0].Op->getType()); 631 } 632 // ... X^X -> ... 633 Ops.erase(Ops.begin()+i, Ops.begin()+i+2); 634 i -= 1; e -= 2; 635 IterateOptimization = true; 636 ++NumAnnihil; 637 } 638 } 639 } 640 break; 641 642 case Instruction::Add: 643 // Scan the operand lists looking for X and -X pairs. If we find any, we 644 // can simplify the expression. X+-X == 0. 645 for (unsigned i = 0, e = Ops.size(); i != e; ++i) { 646 assert(i < Ops.size()); 647 // Check for X and -X in the operand list. 648 if (BinaryOperator::isNeg(Ops[i].Op)) { 649 Value *X = BinaryOperator::getNegArgument(Ops[i].Op); 650 unsigned FoundX = FindInOperandList(Ops, i, X); 651 if (FoundX != i) { 652 // Remove X and -X from the operand list. 653 if (Ops.size() == 2) { 654 ++NumAnnihil; 655 return Constant::getNullValue(X->getType()); 656 } else { 657 Ops.erase(Ops.begin()+i); 658 if (i < FoundX) 659 --FoundX; 660 else 661 --i; // Need to back up an extra one. 662 Ops.erase(Ops.begin()+FoundX); 663 IterateOptimization = true; 664 ++NumAnnihil; 665 --i; // Revisit element. 666 e -= 2; // Removed two elements. 667 } 668 } 669 } 670 } 671 672 673 // Scan the operand list, checking to see if there are any common factors 674 // between operands. Consider something like A*A+A*B*C+D. We would like to 675 // reassociate this to A*(A+B*C)+D, which reduces the number of multiplies. 676 // To efficiently find this, we count the number of times a factor occurs 677 // for any ADD operands that are MULs. 678 std::map<Value*, unsigned> FactorOccurrences; 679 unsigned MaxOcc = 0; 680 Value *MaxOccVal = 0; 681 for (unsigned i = 0, e = Ops.size(); i != e; ++i) { 682 if (BinaryOperator *BOp = dyn_cast<BinaryOperator>(Ops[i].Op)) { 683 if (BOp->getOpcode() == Instruction::Mul && BOp->use_empty()) { 684 // Compute all of the factors of this added value. 685 std::vector<Value*> Factors; 686 FindSingleUseMultiplyFactors(BOp, Factors); 687 assert(Factors.size() > 1 && "Bad linearize!"); 688 689 // Add one to FactorOccurrences for each unique factor in this op. 690 if (Factors.size() == 2) { 691 unsigned Occ = ++FactorOccurrences[Factors[0]]; 692 if (Occ > MaxOcc) { MaxOcc = Occ; MaxOccVal = Factors[0]; } 693 if (Factors[0] != Factors[1]) { // Don't double count A*A. 694 Occ = ++FactorOccurrences[Factors[1]]; 695 if (Occ > MaxOcc) { MaxOcc = Occ; MaxOccVal = Factors[1]; } 696 } 697 } else { 698 std::set<Value*> Duplicates; 699 for (unsigned i = 0, e = Factors.size(); i != e; ++i) { 700 if (Duplicates.insert(Factors[i]).second) { 701 unsigned Occ = ++FactorOccurrences[Factors[i]]; 702 if (Occ > MaxOcc) { MaxOcc = Occ; MaxOccVal = Factors[i]; } 703 } 704 } 705 } 706 } 707 } 708 } 709 710 // If any factor occurred more than one time, we can pull it out. 711 if (MaxOcc > 1) { 712 DOUT << "\nFACTORING [" << MaxOcc << "]: " << *MaxOccVal << "\n"; 713 714 // Create a new instruction that uses the MaxOccVal twice. If we don't do 715 // this, we could otherwise run into situations where removing a factor 716 // from an expression will drop a use of maxocc, and this can cause 717 // RemoveFactorFromExpression on successive values to behave differently. 718 Instruction *DummyInst = BinaryOperator::CreateAdd(MaxOccVal, MaxOccVal); 719 std::vector<Value*> NewMulOps; 720 for (unsigned i = 0, e = Ops.size(); i != e; ++i) { 721 if (Value *V = RemoveFactorFromExpression(Ops[i].Op, MaxOccVal)) { 722 NewMulOps.push_back(V); 723 Ops.erase(Ops.begin()+i); 724 --i; --e; 725 } 726 } 727 728 // No need for extra uses anymore. 729 delete DummyInst; 730 731 unsigned NumAddedValues = NewMulOps.size(); 732 Value *V = EmitAddTreeOfValues(I, NewMulOps); 733 Value *V2 = BinaryOperator::CreateMul(V, MaxOccVal, "tmp", I); 734 735 // Now that we have inserted V and its sole use, optimize it. This allows 736 // us to handle cases that require multiple factoring steps, such as this: 737 // A*A*B + A*A*C --> A*(A*B+A*C) --> A*(A*(B+C)) 738 if (NumAddedValues > 1) 739 ReassociateExpression(cast<BinaryOperator>(V)); 740 741 ++NumFactor; 742 743 if (Ops.empty()) 744 return V2; 745 746 // Add the new value to the list of things being added. 747 Ops.insert(Ops.begin(), ValueEntry(getRank(V2), V2)); 748 749 // Rewrite the tree so that there is now a use of V. 750 RewriteExprTree(I, Ops); 751 return OptimizeExpression(I, Ops); 752 } 753 break; 754 //case Instruction::Mul: 755 } 756 757 if (IterateOptimization) 758 return OptimizeExpression(I, Ops); 759 return 0; 760} 761 762 763/// ReassociateBB - Inspect all of the instructions in this basic block, 764/// reassociating them as we go. 765void Reassociate::ReassociateBB(BasicBlock *BB) { 766 for (BasicBlock::iterator BBI = BB->begin(); BBI != BB->end(); ) { 767 Instruction *BI = BBI++; 768 if (BI->getOpcode() == Instruction::Shl && 769 isa<ConstantInt>(BI->getOperand(1))) 770 if (Instruction *NI = ConvertShiftToMul(BI)) { 771 MadeChange = true; 772 BI = NI; 773 } 774 775 // Reject cases where it is pointless to do this. 776 if (!isa<BinaryOperator>(BI) || BI->getType()->isFloatingPoint() || 777 isa<VectorType>(BI->getType())) 778 continue; // Floating point ops are not associative. 779 780 // If this is a subtract instruction which is not already in negate form, 781 // see if we can convert it to X+-Y. 782 if (BI->getOpcode() == Instruction::Sub) { 783 if (ShouldBreakUpSubtract(BI)) { 784 BI = BreakUpSubtract(BI); 785 MadeChange = true; 786 } else if (BinaryOperator::isNeg(BI)) { 787 // Otherwise, this is a negation. See if the operand is a multiply tree 788 // and if this is not an inner node of a multiply tree. 789 if (isReassociableOp(BI->getOperand(1), Instruction::Mul) && 790 (!BI->hasOneUse() || 791 !isReassociableOp(BI->use_back(), Instruction::Mul))) { 792 BI = LowerNegateToMultiply(BI); 793 MadeChange = true; 794 } 795 } 796 } 797 798 // If this instruction is a commutative binary operator, process it. 799 if (!BI->isAssociative()) continue; 800 BinaryOperator *I = cast<BinaryOperator>(BI); 801 802 // If this is an interior node of a reassociable tree, ignore it until we 803 // get to the root of the tree, to avoid N^2 analysis. 804 if (I->hasOneUse() && isReassociableOp(I->use_back(), I->getOpcode())) 805 continue; 806 807 // If this is an add tree that is used by a sub instruction, ignore it 808 // until we process the subtract. 809 if (I->hasOneUse() && I->getOpcode() == Instruction::Add && 810 cast<Instruction>(I->use_back())->getOpcode() == Instruction::Sub) 811 continue; 812 813 ReassociateExpression(I); 814 } 815} 816 817void Reassociate::ReassociateExpression(BinaryOperator *I) { 818 819 // First, walk the expression tree, linearizing the tree, collecting 820 std::vector<ValueEntry> Ops; 821 LinearizeExprTree(I, Ops); 822 823 DOUT << "RAIn:\t"; DEBUG(PrintOps(I, Ops)); DOUT << "\n"; 824 825 // Now that we have linearized the tree to a list and have gathered all of 826 // the operands and their ranks, sort the operands by their rank. Use a 827 // stable_sort so that values with equal ranks will have their relative 828 // positions maintained (and so the compiler is deterministic). Note that 829 // this sorts so that the highest ranking values end up at the beginning of 830 // the vector. 831 std::stable_sort(Ops.begin(), Ops.end()); 832 833 // OptimizeExpression - Now that we have the expression tree in a convenient 834 // sorted form, optimize it globally if possible. 835 if (Value *V = OptimizeExpression(I, Ops)) { 836 // This expression tree simplified to something that isn't a tree, 837 // eliminate it. 838 DOUT << "Reassoc to scalar: " << *V << "\n"; 839 I->replaceAllUsesWith(V); 840 RemoveDeadBinaryOp(I); 841 return; 842 } 843 844 // We want to sink immediates as deeply as possible except in the case where 845 // this is a multiply tree used only by an add, and the immediate is a -1. 846 // In this case we reassociate to put the negation on the outside so that we 847 // can fold the negation into the add: (-X)*Y + Z -> Z-X*Y 848 if (I->getOpcode() == Instruction::Mul && I->hasOneUse() && 849 cast<Instruction>(I->use_back())->getOpcode() == Instruction::Add && 850 isa<ConstantInt>(Ops.back().Op) && 851 cast<ConstantInt>(Ops.back().Op)->isAllOnesValue()) { 852 Ops.insert(Ops.begin(), Ops.back()); 853 Ops.pop_back(); 854 } 855 856 DOUT << "RAOut:\t"; DEBUG(PrintOps(I, Ops)); DOUT << "\n"; 857 858 if (Ops.size() == 1) { 859 // This expression tree simplified to something that isn't a tree, 860 // eliminate it. 861 I->replaceAllUsesWith(Ops[0].Op); 862 RemoveDeadBinaryOp(I); 863 } else { 864 // Now that we ordered and optimized the expressions, splat them back into 865 // the expression tree, removing any unneeded nodes. 866 RewriteExprTree(I, Ops); 867 } 868} 869 870 871bool Reassociate::runOnFunction(Function &F) { 872 // Recalculate the rank map for F 873 BuildRankMap(F); 874 875 MadeChange = false; 876 for (Function::iterator FI = F.begin(), FE = F.end(); FI != FE; ++FI) 877 ReassociateBB(FI); 878 879 // We are done with the rank map... 880 RankMap.clear(); 881 ValueRankMap.clear(); 882 return MadeChange; 883} 884 885