Reassociate.cpp revision 94285e620b845e09b18939e8d6448e01e692f3ce
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 Value *OptimizeAdd(Instruction *I, std::vector<ValueEntry> &Ops); 96 void LinearizeExprTree(BinaryOperator *I, std::vector<ValueEntry> &Ops); 97 void LinearizeExpr(BinaryOperator *I); 98 Value *RemoveFactorFromExpression(Value *V, Value *Factor); 99 void ReassociateBB(BasicBlock *BB); 100 101 void RemoveDeadBinaryOp(Value *V); 102 }; 103} 104 105char Reassociate::ID = 0; 106static RegisterPass<Reassociate> X("reassociate", "Reassociate expressions"); 107 108// Public interface to the Reassociate pass 109FunctionPass *llvm::createReassociatePass() { return new Reassociate(); } 110 111void Reassociate::RemoveDeadBinaryOp(Value *V) { 112 Instruction *Op = dyn_cast<Instruction>(V); 113 if (!Op || !isa<BinaryOperator>(Op) || !isa<CmpInst>(Op) || !Op->use_empty()) 114 return; 115 116 Value *LHS = Op->getOperand(0), *RHS = Op->getOperand(1); 117 RemoveDeadBinaryOp(LHS); 118 RemoveDeadBinaryOp(RHS); 119} 120 121 122static bool isUnmovableInstruction(Instruction *I) { 123 if (I->getOpcode() == Instruction::PHI || 124 I->getOpcode() == Instruction::Alloca || 125 I->getOpcode() == Instruction::Load || 126 I->getOpcode() == Instruction::Invoke || 127 (I->getOpcode() == Instruction::Call && 128 !isa<DbgInfoIntrinsic>(I)) || 129 I->getOpcode() == Instruction::UDiv || 130 I->getOpcode() == Instruction::SDiv || 131 I->getOpcode() == Instruction::FDiv || 132 I->getOpcode() == Instruction::URem || 133 I->getOpcode() == Instruction::SRem || 134 I->getOpcode() == Instruction::FRem) 135 return true; 136 return false; 137} 138 139void Reassociate::BuildRankMap(Function &F) { 140 unsigned i = 2; 141 142 // Assign distinct ranks to function arguments 143 for (Function::arg_iterator I = F.arg_begin(), E = F.arg_end(); I != E; ++I) 144 ValueRankMap[&*I] = ++i; 145 146 ReversePostOrderTraversal<Function*> RPOT(&F); 147 for (ReversePostOrderTraversal<Function*>::rpo_iterator I = RPOT.begin(), 148 E = RPOT.end(); I != E; ++I) { 149 BasicBlock *BB = *I; 150 unsigned BBRank = RankMap[BB] = ++i << 16; 151 152 // Walk the basic block, adding precomputed ranks for any instructions that 153 // we cannot move. This ensures that the ranks for these instructions are 154 // all different in the block. 155 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) 156 if (isUnmovableInstruction(I)) 157 ValueRankMap[&*I] = ++BBRank; 158 } 159} 160 161unsigned Reassociate::getRank(Value *V) { 162 if (isa<Argument>(V)) return ValueRankMap[V]; // Function argument... 163 164 Instruction *I = dyn_cast<Instruction>(V); 165 if (I == 0) return 0; // Otherwise it's a global or constant, rank 0. 166 167 unsigned &CachedRank = ValueRankMap[I]; 168 if (CachedRank) return CachedRank; // Rank already known? 169 170 // If this is an expression, return the 1+MAX(rank(LHS), rank(RHS)) so that 171 // we can reassociate expressions for code motion! Since we do not recurse 172 // for PHI nodes, we cannot have infinite recursion here, because there 173 // cannot be loops in the value graph that do not go through PHI nodes. 174 unsigned Rank = 0, MaxRank = RankMap[I->getParent()]; 175 for (unsigned i = 0, e = I->getNumOperands(); 176 i != e && Rank != MaxRank; ++i) 177 Rank = std::max(Rank, getRank(I->getOperand(i))); 178 179 // If this is a not or neg instruction, do not count it for rank. This 180 // assures us that X and ~X will have the same rank. 181 if (!I->getType()->isInteger() || 182 (!BinaryOperator::isNot(I) && !BinaryOperator::isNeg(I))) 183 ++Rank; 184 185 //DEBUG(errs() << "Calculated Rank[" << V->getName() << "] = " 186 // << Rank << "\n"); 187 188 return CachedRank = Rank; 189} 190 191/// isReassociableOp - Return true if V is an instruction of the specified 192/// opcode and if it only has one use. 193static BinaryOperator *isReassociableOp(Value *V, unsigned Opcode) { 194 if ((V->hasOneUse() || V->use_empty()) && isa<Instruction>(V) && 195 cast<Instruction>(V)->getOpcode() == Opcode) 196 return cast<BinaryOperator>(V); 197 return 0; 198} 199 200/// LowerNegateToMultiply - Replace 0-X with X*-1. 201/// 202static Instruction *LowerNegateToMultiply(Instruction *Neg, 203 std::map<AssertingVH<>, unsigned> &ValueRankMap) { 204 Constant *Cst = Constant::getAllOnesValue(Neg->getType()); 205 206 Instruction *Res = BinaryOperator::CreateMul(Neg->getOperand(1), Cst, "",Neg); 207 ValueRankMap.erase(Neg); 208 Res->takeName(Neg); 209 Neg->replaceAllUsesWith(Res); 210 Neg->eraseFromParent(); 211 return Res; 212} 213 214// Given an expression of the form '(A+B)+(D+C)', turn it into '(((A+B)+C)+D)'. 215// Note that if D is also part of the expression tree that we recurse to 216// linearize it as well. Besides that case, this does not recurse into A,B, or 217// C. 218void Reassociate::LinearizeExpr(BinaryOperator *I) { 219 BinaryOperator *LHS = cast<BinaryOperator>(I->getOperand(0)); 220 BinaryOperator *RHS = cast<BinaryOperator>(I->getOperand(1)); 221 assert(isReassociableOp(LHS, I->getOpcode()) && 222 isReassociableOp(RHS, I->getOpcode()) && 223 "Not an expression that needs linearization?"); 224 225 DEBUG(errs() << "Linear" << *LHS << '\n' << *RHS << '\n' << *I << '\n'); 226 227 // Move the RHS instruction to live immediately before I, avoiding breaking 228 // dominator properties. 229 RHS->moveBefore(I); 230 231 // Move operands around to do the linearization. 232 I->setOperand(1, RHS->getOperand(0)); 233 RHS->setOperand(0, LHS); 234 I->setOperand(0, RHS); 235 236 ++NumLinear; 237 MadeChange = true; 238 DEBUG(errs() << "Linearized: " << *I << '\n'); 239 240 // If D is part of this expression tree, tail recurse. 241 if (isReassociableOp(I->getOperand(1), I->getOpcode())) 242 LinearizeExpr(I); 243} 244 245 246/// LinearizeExprTree - Given an associative binary expression tree, traverse 247/// all of the uses putting it into canonical form. This forces a left-linear 248/// form of the the expression (((a+b)+c)+d), and collects information about the 249/// rank of the non-tree operands. 250/// 251/// NOTE: These intentionally destroys the expression tree operands (turning 252/// them into undef values) to reduce #uses of the values. This means that the 253/// caller MUST use something like RewriteExprTree to put the values back in. 254/// 255void Reassociate::LinearizeExprTree(BinaryOperator *I, 256 std::vector<ValueEntry> &Ops) { 257 Value *LHS = I->getOperand(0), *RHS = I->getOperand(1); 258 unsigned Opcode = I->getOpcode(); 259 260 // First step, linearize the expression if it is in ((A+B)+(C+D)) form. 261 BinaryOperator *LHSBO = isReassociableOp(LHS, Opcode); 262 BinaryOperator *RHSBO = isReassociableOp(RHS, Opcode); 263 264 // If this is a multiply expression tree and it contains internal negations, 265 // transform them into multiplies by -1 so they can be reassociated. 266 if (I->getOpcode() == Instruction::Mul) { 267 if (!LHSBO && LHS->hasOneUse() && BinaryOperator::isNeg(LHS)) { 268 LHS = LowerNegateToMultiply(cast<Instruction>(LHS), ValueRankMap); 269 LHSBO = isReassociableOp(LHS, Opcode); 270 } 271 if (!RHSBO && RHS->hasOneUse() && BinaryOperator::isNeg(RHS)) { 272 RHS = LowerNegateToMultiply(cast<Instruction>(RHS), ValueRankMap); 273 RHSBO = isReassociableOp(RHS, Opcode); 274 } 275 } 276 277 if (!LHSBO) { 278 if (!RHSBO) { 279 // Neither the LHS or RHS as part of the tree, thus this is a leaf. As 280 // such, just remember these operands and their rank. 281 Ops.push_back(ValueEntry(getRank(LHS), LHS)); 282 Ops.push_back(ValueEntry(getRank(RHS), RHS)); 283 284 // Clear the leaves out. 285 I->setOperand(0, UndefValue::get(I->getType())); 286 I->setOperand(1, UndefValue::get(I->getType())); 287 return; 288 } 289 290 // Turn X+(Y+Z) -> (Y+Z)+X 291 std::swap(LHSBO, RHSBO); 292 std::swap(LHS, RHS); 293 bool Success = !I->swapOperands(); 294 assert(Success && "swapOperands failed"); 295 Success = false; 296 MadeChange = true; 297 } else if (RHSBO) { 298 // Turn (A+B)+(C+D) -> (((A+B)+C)+D). This guarantees the the RHS is not 299 // part of the expression tree. 300 LinearizeExpr(I); 301 LHS = LHSBO = cast<BinaryOperator>(I->getOperand(0)); 302 RHS = I->getOperand(1); 303 RHSBO = 0; 304 } 305 306 // Okay, now we know that the LHS is a nested expression and that the RHS is 307 // not. Perform reassociation. 308 assert(!isReassociableOp(RHS, Opcode) && "LinearizeExpr failed!"); 309 310 // Move LHS right before I to make sure that the tree expression dominates all 311 // values. 312 LHSBO->moveBefore(I); 313 314 // Linearize the expression tree on the LHS. 315 LinearizeExprTree(LHSBO, Ops); 316 317 // Remember the RHS operand and its rank. 318 Ops.push_back(ValueEntry(getRank(RHS), RHS)); 319 320 // Clear the RHS leaf out. 321 I->setOperand(1, UndefValue::get(I->getType())); 322} 323 324// RewriteExprTree - Now that the operands for this expression tree are 325// linearized and optimized, emit them in-order. This function is written to be 326// tail recursive. 327void Reassociate::RewriteExprTree(BinaryOperator *I, 328 std::vector<ValueEntry> &Ops, 329 unsigned i) { 330 if (i+2 == Ops.size()) { 331 if (I->getOperand(0) != Ops[i].Op || 332 I->getOperand(1) != Ops[i+1].Op) { 333 Value *OldLHS = I->getOperand(0); 334 DEBUG(errs() << "RA: " << *I << '\n'); 335 I->setOperand(0, Ops[i].Op); 336 I->setOperand(1, Ops[i+1].Op); 337 DEBUG(errs() << "TO: " << *I << '\n'); 338 MadeChange = true; 339 ++NumChanged; 340 341 // If we reassociated a tree to fewer operands (e.g. (1+a+2) -> (a+3) 342 // delete the extra, now dead, nodes. 343 RemoveDeadBinaryOp(OldLHS); 344 } 345 return; 346 } 347 assert(i+2 < Ops.size() && "Ops index out of range!"); 348 349 if (I->getOperand(1) != Ops[i].Op) { 350 DEBUG(errs() << "RA: " << *I << '\n'); 351 I->setOperand(1, Ops[i].Op); 352 DEBUG(errs() << "TO: " << *I << '\n'); 353 MadeChange = true; 354 ++NumChanged; 355 } 356 357 BinaryOperator *LHS = cast<BinaryOperator>(I->getOperand(0)); 358 assert(LHS->getOpcode() == I->getOpcode() && 359 "Improper expression tree!"); 360 361 // Compactify the tree instructions together with each other to guarantee 362 // that the expression tree is dominated by all of Ops. 363 LHS->moveBefore(I); 364 RewriteExprTree(LHS, Ops, i+1); 365} 366 367 368 369// NegateValue - Insert instructions before the instruction pointed to by BI, 370// that computes the negative version of the value specified. The negative 371// version of the value is returned, and BI is left pointing at the instruction 372// that should be processed next by the reassociation pass. 373// 374static Value *NegateValue(Value *V, Instruction *BI) { 375 // We are trying to expose opportunity for reassociation. One of the things 376 // that we want to do to achieve this is to push a negation as deep into an 377 // expression chain as possible, to expose the add instructions. In practice, 378 // this means that we turn this: 379 // X = -(A+12+C+D) into X = -A + -12 + -C + -D = -12 + -A + -C + -D 380 // so that later, a: Y = 12+X could get reassociated with the -12 to eliminate 381 // the constants. We assume that instcombine will clean up the mess later if 382 // we introduce tons of unnecessary negation instructions... 383 // 384 if (Instruction *I = dyn_cast<Instruction>(V)) 385 if (I->getOpcode() == Instruction::Add && I->hasOneUse()) { 386 // Push the negates through the add. 387 I->setOperand(0, NegateValue(I->getOperand(0), BI)); 388 I->setOperand(1, NegateValue(I->getOperand(1), BI)); 389 390 // We must move the add instruction here, because the neg instructions do 391 // not dominate the old add instruction in general. By moving it, we are 392 // assured that the neg instructions we just inserted dominate the 393 // instruction we are about to insert after them. 394 // 395 I->moveBefore(BI); 396 I->setName(I->getName()+".neg"); 397 return I; 398 } 399 400 // Insert a 'neg' instruction that subtracts the value from zero to get the 401 // negation. 402 // 403 return BinaryOperator::CreateNeg(V, V->getName() + ".neg", BI); 404} 405 406/// ShouldBreakUpSubtract - Return true if we should break up this subtract of 407/// X-Y into (X + -Y). 408static bool ShouldBreakUpSubtract(Instruction *Sub) { 409 // If this is a negation, we can't split it up! 410 if (BinaryOperator::isNeg(Sub)) 411 return false; 412 413 // Don't bother to break this up unless either the LHS is an associable add or 414 // subtract or if this is only used by one. 415 if (isReassociableOp(Sub->getOperand(0), Instruction::Add) || 416 isReassociableOp(Sub->getOperand(0), Instruction::Sub)) 417 return true; 418 if (isReassociableOp(Sub->getOperand(1), Instruction::Add) || 419 isReassociableOp(Sub->getOperand(1), Instruction::Sub)) 420 return true; 421 if (Sub->hasOneUse() && 422 (isReassociableOp(Sub->use_back(), Instruction::Add) || 423 isReassociableOp(Sub->use_back(), Instruction::Sub))) 424 return true; 425 426 return false; 427} 428 429/// BreakUpSubtract - If we have (X-Y), and if either X is an add, or if this is 430/// only used by an add, transform this into (X+(0-Y)) to promote better 431/// reassociation. 432static Instruction *BreakUpSubtract(Instruction *Sub, 433 std::map<AssertingVH<>, unsigned> &ValueRankMap) { 434 // Convert a subtract into an add and a neg instruction... so that sub 435 // instructions can be commuted with other add instructions... 436 // 437 // Calculate the negative value of Operand 1 of the sub instruction... 438 // and set it as the RHS of the add instruction we just made... 439 // 440 Value *NegVal = NegateValue(Sub->getOperand(1), Sub); 441 Instruction *New = 442 BinaryOperator::CreateAdd(Sub->getOperand(0), NegVal, "", Sub); 443 New->takeName(Sub); 444 445 // Everyone now refers to the add instruction. 446 ValueRankMap.erase(Sub); 447 Sub->replaceAllUsesWith(New); 448 Sub->eraseFromParent(); 449 450 DEBUG(errs() << "Negated: " << *New << '\n'); 451 return New; 452} 453 454/// ConvertShiftToMul - If this is a shift of a reassociable multiply or is used 455/// by one, change this into a multiply by a constant to assist with further 456/// reassociation. 457static Instruction *ConvertShiftToMul(Instruction *Shl, 458 std::map<AssertingVH<>, unsigned> &ValueRankMap) { 459 // If an operand of this shift is a reassociable multiply, or if the shift 460 // is used by a reassociable multiply or add, turn into a multiply. 461 if (isReassociableOp(Shl->getOperand(0), Instruction::Mul) || 462 (Shl->hasOneUse() && 463 (isReassociableOp(Shl->use_back(), Instruction::Mul) || 464 isReassociableOp(Shl->use_back(), Instruction::Add)))) { 465 Constant *MulCst = ConstantInt::get(Shl->getType(), 1); 466 MulCst = ConstantExpr::getShl(MulCst, cast<Constant>(Shl->getOperand(1))); 467 468 Instruction *Mul = 469 BinaryOperator::CreateMul(Shl->getOperand(0), MulCst, "", 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, SmallVectorImpl<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 SmallVectorImpl<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/// OptimizeAndOrXor - Optimize a series of operands to an 'and', 'or', or 'xor' 553/// instruction. This optimizes based on identities. If it can be reduced to 554/// a single Value, it is returned, otherwise the Ops list is mutated as 555/// necessary. 556static Value *OptimizeAndOrXor(unsigned Opcode, std::vector<ValueEntry> &Ops) { 557 // Scan the operand lists looking for X and ~X pairs, along with X,X pairs. 558 // If we find any, we can simplify the expression. X&~X == 0, X|~X == -1. 559 for (unsigned i = 0, e = Ops.size(); i != e; ++i) { 560 // First, check for X and ~X in the operand list. 561 assert(i < Ops.size()); 562 if (BinaryOperator::isNot(Ops[i].Op)) { // Cannot occur for ^. 563 Value *X = BinaryOperator::getNotArgument(Ops[i].Op); 564 unsigned FoundX = FindInOperandList(Ops, i, X); 565 if (FoundX != i) { 566 if (Opcode == Instruction::And) // ...&X&~X = 0 567 return Constant::getNullValue(X->getType()); 568 569 if (Opcode == Instruction::Or) // ...|X|~X = -1 570 return Constant::getAllOnesValue(X->getType()); 571 } 572 } 573 574 // Next, check for duplicate pairs of values, which we assume are next to 575 // each other, due to our sorting criteria. 576 assert(i < Ops.size()); 577 if (i+1 != Ops.size() && Ops[i+1].Op == Ops[i].Op) { 578 if (Opcode == Instruction::And || Opcode == Instruction::Or) { 579 // Drop duplicate values. 580 Ops.erase(Ops.begin()+i); 581 --i; --e; 582 ++NumAnnihil; 583 } else { 584 assert(Opcode == Instruction::Xor); 585 if (e == 2) 586 return Constant::getNullValue(Ops[0].Op->getType()); 587 588 // ... X^X -> ... 589 Ops.erase(Ops.begin()+i, Ops.begin()+i+2); 590 i -= 1; e -= 2; 591 ++NumAnnihil; 592 } 593 } 594 } 595 return 0; 596} 597 598/// OptimizeAdd - Optimize a series of operands to an 'add' instruction. This 599/// optimizes based on identities. If it can be reduced to a single Value, it 600/// is returned, otherwise the Ops list is mutated as necessary. 601Value *Reassociate::OptimizeAdd(Instruction *I, std::vector<ValueEntry> &Ops) { 602 // Scan the operand lists looking for X and -X pairs. If we find any, we 603 // can simplify the expression. X+-X == 0. 604 for (unsigned i = 0, e = Ops.size(); i != e; ++i) { 605 assert(i < Ops.size()); 606 // Check for X and -X in the operand list. 607 if (!BinaryOperator::isNeg(Ops[i].Op)) 608 continue; 609 610 Value *X = BinaryOperator::getNegArgument(Ops[i].Op); 611 unsigned FoundX = FindInOperandList(Ops, i, X); 612 if (FoundX == i) 613 continue; 614 615 // Remove X and -X from the operand list. 616 if (Ops.size() == 2) 617 return Constant::getNullValue(X->getType()); 618 619 Ops.erase(Ops.begin()+i); 620 if (i < FoundX) 621 --FoundX; 622 else 623 --i; // Need to back up an extra one. 624 Ops.erase(Ops.begin()+FoundX); 625 ++NumAnnihil; 626 --i; // Revisit element. 627 e -= 2; // Removed two elements. 628 } 629 630 // Scan the operand list, checking to see if there are any common factors 631 // between operands. Consider something like A*A+A*B*C+D. We would like to 632 // reassociate this to A*(A+B*C)+D, which reduces the number of multiplies. 633 // To efficiently find this, we count the number of times a factor occurs 634 // for any ADD operands that are MULs. 635 DenseMap<Value*, unsigned> FactorOccurrences; 636 637 // Keep track of each multiply we see, to avoid triggering on (X*4)+(X*4) 638 // where they are actually the same multiply. 639 SmallPtrSet<BinaryOperator*, 4> Multiplies; 640 unsigned MaxOcc = 0; 641 Value *MaxOccVal = 0; 642 for (unsigned i = 0, e = Ops.size(); i != e; ++i) { 643 BinaryOperator *BOp = dyn_cast<BinaryOperator>(Ops[i].Op); 644 if (BOp == 0 || BOp->getOpcode() != Instruction::Mul || !BOp->use_empty()) 645 continue; 646 647 // If we've already seen this multiply, don't revisit it. 648 if (!Multiplies.insert(BOp)) continue; 649 650 // Compute all of the factors of this added value. 651 SmallVector<Value*, 8> Factors; 652 FindSingleUseMultiplyFactors(BOp, Factors); 653 assert(Factors.size() > 1 && "Bad linearize!"); 654 655 // Add one to FactorOccurrences for each unique factor in this op. 656 if (Factors.size() == 2) { 657 unsigned Occ = ++FactorOccurrences[Factors[0]]; 658 if (Occ > MaxOcc) { MaxOcc = Occ; MaxOccVal = Factors[0]; } 659 if (Factors[0] != Factors[1]) { // Don't double count A*A. 660 Occ = ++FactorOccurrences[Factors[1]]; 661 if (Occ > MaxOcc) { MaxOcc = Occ; MaxOccVal = Factors[1]; } 662 } 663 } else { 664 SmallPtrSet<Value*, 4> Duplicates; 665 for (unsigned i = 0, e = Factors.size(); i != e; ++i) { 666 if (!Duplicates.insert(Factors[i])) continue; 667 668 unsigned Occ = ++FactorOccurrences[Factors[i]]; 669 if (Occ > MaxOcc) { MaxOcc = Occ; MaxOccVal = Factors[i]; } 670 } 671 } 672 } 673 674 // If any factor occurred more than one time, we can pull it out. 675 if (MaxOcc > 1) { 676 DEBUG(errs() << "\nFACTORING [" << MaxOcc << "]: " << *MaxOccVal << "\n"); 677 ++NumFactor; 678 679 // Create a new instruction that uses the MaxOccVal twice. If we don't do 680 // this, we could otherwise run into situations where removing a factor 681 // from an expression will drop a use of maxocc, and this can cause 682 // RemoveFactorFromExpression on successive values to behave differently. 683 Instruction *DummyInst = BinaryOperator::CreateAdd(MaxOccVal, MaxOccVal); 684 SmallVector<Value*, 4> NewMulOps; 685 for (unsigned i = 0, e = Ops.size(); i != e; ++i) { 686 if (Value *V = RemoveFactorFromExpression(Ops[i].Op, MaxOccVal)) { 687 NewMulOps.push_back(V); 688 Ops.erase(Ops.begin()+i); 689 --i; --e; 690 } 691 } 692 693 // No need for extra uses anymore. 694 delete DummyInst; 695 696 unsigned NumAddedValues = NewMulOps.size(); 697 Value *V = EmitAddTreeOfValues(I, NewMulOps); 698 Value *V2 = BinaryOperator::CreateMul(V, MaxOccVal, "tmp", I); 699 700 // Now that we have inserted V and its sole use, optimize it. This allows 701 // us to handle cases that require multiple factoring steps, such as this: 702 // A*A*B + A*A*C --> A*(A*B+A*C) --> A*(A*(B+C)) 703 if (NumAddedValues > 1) 704 ReassociateExpression(cast<BinaryOperator>(V)); 705 706 // If every add operand included the factor (e.g. "A*B + A*C"), then the 707 // entire result expression is just the multiply "A*(B+C)". 708 if (Ops.empty()) 709 return V2; 710 711 // Otherwise, we had some input that didn't have the fact, such as 712 // "A*B + A*C + D" -> "A*(B+C) + D". Add the new multiply to the list of 713 // things being added. 714 Ops.insert(Ops.begin(), ValueEntry(getRank(V2), V2)); 715 } 716 717 return 0; 718} 719 720Value *Reassociate::OptimizeExpression(BinaryOperator *I, 721 std::vector<ValueEntry> &Ops) { 722 // Now that we have the linearized expression tree, try to optimize it. 723 // Start by folding any constants that we found. 724 bool IterateOptimization = false; 725 if (Ops.size() == 1) return Ops[0].Op; 726 727 unsigned Opcode = I->getOpcode(); 728 729 if (Constant *V1 = dyn_cast<Constant>(Ops[Ops.size()-2].Op)) 730 if (Constant *V2 = dyn_cast<Constant>(Ops.back().Op)) { 731 Ops.pop_back(); 732 Ops.back().Op = ConstantExpr::get(Opcode, V1, V2); 733 return OptimizeExpression(I, Ops); 734 } 735 736 // Check for destructive annihilation due to a constant being used. 737 if (ConstantInt *CstVal = dyn_cast<ConstantInt>(Ops.back().Op)) 738 switch (Opcode) { 739 default: break; 740 case Instruction::And: 741 if (CstVal->isZero()) // ... & 0 -> 0 742 return CstVal; 743 if (CstVal->isAllOnesValue()) // ... & -1 -> ... 744 Ops.pop_back(); 745 break; 746 case Instruction::Mul: 747 if (CstVal->isZero()) { // ... * 0 -> 0 748 ++NumAnnihil; 749 return CstVal; 750 } 751 752 if (cast<ConstantInt>(CstVal)->isOne()) 753 Ops.pop_back(); // ... * 1 -> ... 754 break; 755 case Instruction::Or: 756 if (CstVal->isAllOnesValue()) // ... | -1 -> -1 757 return CstVal; 758 // FALLTHROUGH! 759 case Instruction::Add: 760 case Instruction::Xor: 761 if (CstVal->isZero()) // ... [|^+] 0 -> ... 762 Ops.pop_back(); 763 break; 764 } 765 if (Ops.size() == 1) return Ops[0].Op; 766 767 // Handle destructive annihilation due to identities between elements in the 768 // argument list here. 769 switch (Opcode) { 770 default: break; 771 case Instruction::And: 772 case Instruction::Or: 773 case Instruction::Xor: { 774 unsigned NumOps = Ops.size(); 775 if (Value *Result = OptimizeAndOrXor(Opcode, Ops)) 776 return Result; 777 IterateOptimization |= Ops.size() != NumOps; 778 break; 779 } 780 781 case Instruction::Add: { 782 unsigned NumOps = Ops.size(); 783 if (Value *Result = OptimizeAdd(I, Ops)) 784 return Result; 785 IterateOptimization |= Ops.size() != NumOps; 786 } 787 788 break; 789 //case Instruction::Mul: 790 } 791 792 if (IterateOptimization) 793 return OptimizeExpression(I, Ops); 794 return 0; 795} 796 797 798/// ReassociateBB - Inspect all of the instructions in this basic block, 799/// reassociating them as we go. 800void Reassociate::ReassociateBB(BasicBlock *BB) { 801 for (BasicBlock::iterator BBI = BB->begin(); BBI != BB->end(); ) { 802 Instruction *BI = BBI++; 803 if (BI->getOpcode() == Instruction::Shl && 804 isa<ConstantInt>(BI->getOperand(1))) 805 if (Instruction *NI = ConvertShiftToMul(BI, ValueRankMap)) { 806 MadeChange = true; 807 BI = NI; 808 } 809 810 // Reject cases where it is pointless to do this. 811 if (!isa<BinaryOperator>(BI) || BI->getType()->isFloatingPoint() || 812 isa<VectorType>(BI->getType())) 813 continue; // Floating point ops are not associative. 814 815 // If this is a subtract instruction which is not already in negate form, 816 // see if we can convert it to X+-Y. 817 if (BI->getOpcode() == Instruction::Sub) { 818 if (ShouldBreakUpSubtract(BI)) { 819 BI = BreakUpSubtract(BI, ValueRankMap); 820 MadeChange = true; 821 } else if (BinaryOperator::isNeg(BI)) { 822 // Otherwise, this is a negation. See if the operand is a multiply tree 823 // and if this is not an inner node of a multiply tree. 824 if (isReassociableOp(BI->getOperand(1), Instruction::Mul) && 825 (!BI->hasOneUse() || 826 !isReassociableOp(BI->use_back(), Instruction::Mul))) { 827 BI = LowerNegateToMultiply(BI, ValueRankMap); 828 MadeChange = true; 829 } 830 } 831 } 832 833 // If this instruction is a commutative binary operator, process it. 834 if (!BI->isAssociative()) continue; 835 BinaryOperator *I = cast<BinaryOperator>(BI); 836 837 // If this is an interior node of a reassociable tree, ignore it until we 838 // get to the root of the tree, to avoid N^2 analysis. 839 if (I->hasOneUse() && isReassociableOp(I->use_back(), I->getOpcode())) 840 continue; 841 842 // If this is an add tree that is used by a sub instruction, ignore it 843 // until we process the subtract. 844 if (I->hasOneUse() && I->getOpcode() == Instruction::Add && 845 cast<Instruction>(I->use_back())->getOpcode() == Instruction::Sub) 846 continue; 847 848 ReassociateExpression(I); 849 } 850} 851 852void Reassociate::ReassociateExpression(BinaryOperator *I) { 853 854 // First, walk the expression tree, linearizing the tree, collecting 855 std::vector<ValueEntry> Ops; 856 LinearizeExprTree(I, Ops); 857 858 DEBUG(errs() << "RAIn:\t"; PrintOps(I, Ops); errs() << '\n'); 859 860 // Now that we have linearized the tree to a list and have gathered all of 861 // the operands and their ranks, sort the operands by their rank. Use a 862 // stable_sort so that values with equal ranks will have their relative 863 // positions maintained (and so the compiler is deterministic). Note that 864 // this sorts so that the highest ranking values end up at the beginning of 865 // the vector. 866 std::stable_sort(Ops.begin(), Ops.end()); 867 868 // OptimizeExpression - Now that we have the expression tree in a convenient 869 // sorted form, optimize it globally if possible. 870 if (Value *V = OptimizeExpression(I, Ops)) { 871 // This expression tree simplified to something that isn't a tree, 872 // eliminate it. 873 DEBUG(errs() << "Reassoc to scalar: " << *V << '\n'); 874 I->replaceAllUsesWith(V); 875 RemoveDeadBinaryOp(I); 876 ++NumAnnihil; 877 return; 878 } 879 880 // We want to sink immediates as deeply as possible except in the case where 881 // this is a multiply tree used only by an add, and the immediate is a -1. 882 // In this case we reassociate to put the negation on the outside so that we 883 // can fold the negation into the add: (-X)*Y + Z -> Z-X*Y 884 if (I->getOpcode() == Instruction::Mul && I->hasOneUse() && 885 cast<Instruction>(I->use_back())->getOpcode() == Instruction::Add && 886 isa<ConstantInt>(Ops.back().Op) && 887 cast<ConstantInt>(Ops.back().Op)->isAllOnesValue()) { 888 Ops.insert(Ops.begin(), Ops.back()); 889 Ops.pop_back(); 890 } 891 892 DEBUG(errs() << "RAOut:\t"; PrintOps(I, Ops); errs() << '\n'); 893 894 if (Ops.size() == 1) { 895 // This expression tree simplified to something that isn't a tree, 896 // eliminate it. 897 I->replaceAllUsesWith(Ops[0].Op); 898 RemoveDeadBinaryOp(I); 899 } else { 900 // Now that we ordered and optimized the expressions, splat them back into 901 // the expression tree, removing any unneeded nodes. 902 RewriteExprTree(I, Ops); 903 } 904} 905 906 907bool Reassociate::runOnFunction(Function &F) { 908 // Recalculate the rank map for F 909 BuildRankMap(F); 910 911 MadeChange = false; 912 for (Function::iterator FI = F.begin(), FE = F.end(); FI != FE; ++FI) 913 ReassociateBB(FI); 914 915 // We are done with the rank map... 916 RankMap.clear(); 917 ValueRankMap.clear(); 918 return MadeChange; 919} 920 921