Reassociate.cpp revision 1e7558b65689999089f53ce40ff07564cf498c68
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 SmallVectorImpl<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 Value *ReassociateExpression(BinaryOperator *I); 92 void RewriteExprTree(BinaryOperator *I, SmallVectorImpl<ValueEntry> &Ops, 93 unsigned Idx = 0); 94 Value *OptimizeExpression(BinaryOperator *I, 95 SmallVectorImpl<ValueEntry> &Ops); 96 Value *OptimizeAdd(Instruction *I, SmallVectorImpl<ValueEntry> &Ops); 97 void LinearizeExprTree(BinaryOperator *I, SmallVectorImpl<ValueEntry> &Ops); 98 void LinearizeExpr(BinaryOperator *I); 99 Value *RemoveFactorFromExpression(Value *V, Value *Factor); 100 void ReassociateBB(BasicBlock *BB); 101 102 void RemoveDeadBinaryOp(Value *V); 103 }; 104} 105 106char Reassociate::ID = 0; 107static RegisterPass<Reassociate> X("reassociate", "Reassociate expressions"); 108 109// Public interface to the Reassociate pass 110FunctionPass *llvm::createReassociatePass() { return new Reassociate(); } 111 112void Reassociate::RemoveDeadBinaryOp(Value *V) { 113 Instruction *Op = dyn_cast<Instruction>(V); 114 if (!Op || !isa<BinaryOperator>(Op) || !Op->use_empty()) 115 return; 116 117 Value *LHS = Op->getOperand(0), *RHS = Op->getOperand(1); 118 119 ValueRankMap.erase(Op); 120 Op->eraseFromParent(); 121 RemoveDeadBinaryOp(LHS); 122 RemoveDeadBinaryOp(RHS); 123} 124 125 126static bool isUnmovableInstruction(Instruction *I) { 127 if (I->getOpcode() == Instruction::PHI || 128 I->getOpcode() == Instruction::Alloca || 129 I->getOpcode() == Instruction::Load || 130 I->getOpcode() == Instruction::Invoke || 131 (I->getOpcode() == Instruction::Call && 132 !isa<DbgInfoIntrinsic>(I)) || 133 I->getOpcode() == Instruction::UDiv || 134 I->getOpcode() == Instruction::SDiv || 135 I->getOpcode() == Instruction::FDiv || 136 I->getOpcode() == Instruction::URem || 137 I->getOpcode() == Instruction::SRem || 138 I->getOpcode() == Instruction::FRem) 139 return true; 140 return false; 141} 142 143void Reassociate::BuildRankMap(Function &F) { 144 unsigned i = 2; 145 146 // Assign distinct ranks to function arguments 147 for (Function::arg_iterator I = F.arg_begin(), E = F.arg_end(); I != E; ++I) 148 ValueRankMap[&*I] = ++i; 149 150 ReversePostOrderTraversal<Function*> RPOT(&F); 151 for (ReversePostOrderTraversal<Function*>::rpo_iterator I = RPOT.begin(), 152 E = RPOT.end(); I != E; ++I) { 153 BasicBlock *BB = *I; 154 unsigned BBRank = RankMap[BB] = ++i << 16; 155 156 // Walk the basic block, adding precomputed ranks for any instructions that 157 // we cannot move. This ensures that the ranks for these instructions are 158 // all different in the block. 159 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) 160 if (isUnmovableInstruction(I)) 161 ValueRankMap[&*I] = ++BBRank; 162 } 163} 164 165unsigned Reassociate::getRank(Value *V) { 166 if (isa<Argument>(V)) return ValueRankMap[V]; // Function argument... 167 168 Instruction *I = dyn_cast<Instruction>(V); 169 if (I == 0) return 0; // Otherwise it's a global or constant, rank 0. 170 171 unsigned &CachedRank = ValueRankMap[I]; 172 if (CachedRank) return CachedRank; // Rank already known? 173 174 // If this is an expression, return the 1+MAX(rank(LHS), rank(RHS)) so that 175 // we can reassociate expressions for code motion! Since we do not recurse 176 // for PHI nodes, we cannot have infinite recursion here, because there 177 // cannot be loops in the value graph that do not go through PHI nodes. 178 unsigned Rank = 0, MaxRank = RankMap[I->getParent()]; 179 for (unsigned i = 0, e = I->getNumOperands(); 180 i != e && Rank != MaxRank; ++i) 181 Rank = std::max(Rank, getRank(I->getOperand(i))); 182 183 // If this is a not or neg instruction, do not count it for rank. This 184 // assures us that X and ~X will have the same rank. 185 if (!I->getType()->isInteger() || 186 (!BinaryOperator::isNot(I) && !BinaryOperator::isNeg(I))) 187 ++Rank; 188 189 //DEBUG(errs() << "Calculated Rank[" << V->getName() << "] = " 190 // << Rank << "\n"); 191 192 return CachedRank = Rank; 193} 194 195/// isReassociableOp - Return true if V is an instruction of the specified 196/// opcode and if it only has one use. 197static BinaryOperator *isReassociableOp(Value *V, unsigned Opcode) { 198 if ((V->hasOneUse() || V->use_empty()) && isa<Instruction>(V) && 199 cast<Instruction>(V)->getOpcode() == Opcode) 200 return cast<BinaryOperator>(V); 201 return 0; 202} 203 204/// LowerNegateToMultiply - Replace 0-X with X*-1. 205/// 206static Instruction *LowerNegateToMultiply(Instruction *Neg, 207 std::map<AssertingVH<>, unsigned> &ValueRankMap) { 208 Constant *Cst = Constant::getAllOnesValue(Neg->getType()); 209 210 Instruction *Res = BinaryOperator::CreateMul(Neg->getOperand(1), Cst, "",Neg); 211 ValueRankMap.erase(Neg); 212 Res->takeName(Neg); 213 Neg->replaceAllUsesWith(Res); 214 Neg->eraseFromParent(); 215 return Res; 216} 217 218// Given an expression of the form '(A+B)+(D+C)', turn it into '(((A+B)+C)+D)'. 219// Note that if D is also part of the expression tree that we recurse to 220// linearize it as well. Besides that case, this does not recurse into A,B, or 221// C. 222void Reassociate::LinearizeExpr(BinaryOperator *I) { 223 BinaryOperator *LHS = cast<BinaryOperator>(I->getOperand(0)); 224 BinaryOperator *RHS = cast<BinaryOperator>(I->getOperand(1)); 225 assert(isReassociableOp(LHS, I->getOpcode()) && 226 isReassociableOp(RHS, I->getOpcode()) && 227 "Not an expression that needs linearization?"); 228 229 DEBUG(errs() << "Linear" << *LHS << '\n' << *RHS << '\n' << *I << '\n'); 230 231 // Move the RHS instruction to live immediately before I, avoiding breaking 232 // dominator properties. 233 RHS->moveBefore(I); 234 235 // Move operands around to do the linearization. 236 I->setOperand(1, RHS->getOperand(0)); 237 RHS->setOperand(0, LHS); 238 I->setOperand(0, RHS); 239 240 ++NumLinear; 241 MadeChange = true; 242 DEBUG(errs() << "Linearized: " << *I << '\n'); 243 244 // If D is part of this expression tree, tail recurse. 245 if (isReassociableOp(I->getOperand(1), I->getOpcode())) 246 LinearizeExpr(I); 247} 248 249 250/// LinearizeExprTree - Given an associative binary expression tree, traverse 251/// all of the uses putting it into canonical form. This forces a left-linear 252/// form of the the expression (((a+b)+c)+d), and collects information about the 253/// rank of the non-tree operands. 254/// 255/// NOTE: These intentionally destroys the expression tree operands (turning 256/// them into undef values) to reduce #uses of the values. This means that the 257/// caller MUST use something like RewriteExprTree to put the values back in. 258/// 259void Reassociate::LinearizeExprTree(BinaryOperator *I, 260 SmallVectorImpl<ValueEntry> &Ops) { 261 Value *LHS = I->getOperand(0), *RHS = I->getOperand(1); 262 unsigned Opcode = I->getOpcode(); 263 264 // First step, linearize the expression if it is in ((A+B)+(C+D)) form. 265 BinaryOperator *LHSBO = isReassociableOp(LHS, Opcode); 266 BinaryOperator *RHSBO = isReassociableOp(RHS, Opcode); 267 268 // If this is a multiply expression tree and it contains internal negations, 269 // transform them into multiplies by -1 so they can be reassociated. 270 if (I->getOpcode() == Instruction::Mul) { 271 if (!LHSBO && LHS->hasOneUse() && BinaryOperator::isNeg(LHS)) { 272 LHS = LowerNegateToMultiply(cast<Instruction>(LHS), ValueRankMap); 273 LHSBO = isReassociableOp(LHS, Opcode); 274 } 275 if (!RHSBO && RHS->hasOneUse() && BinaryOperator::isNeg(RHS)) { 276 RHS = LowerNegateToMultiply(cast<Instruction>(RHS), ValueRankMap); 277 RHSBO = isReassociableOp(RHS, Opcode); 278 } 279 } 280 281 if (!LHSBO) { 282 if (!RHSBO) { 283 // Neither the LHS or RHS as part of the tree, thus this is a leaf. As 284 // such, just remember these operands and their rank. 285 Ops.push_back(ValueEntry(getRank(LHS), LHS)); 286 Ops.push_back(ValueEntry(getRank(RHS), RHS)); 287 288 // Clear the leaves out. 289 I->setOperand(0, UndefValue::get(I->getType())); 290 I->setOperand(1, UndefValue::get(I->getType())); 291 return; 292 } 293 294 // Turn X+(Y+Z) -> (Y+Z)+X 295 std::swap(LHSBO, RHSBO); 296 std::swap(LHS, RHS); 297 bool Success = !I->swapOperands(); 298 assert(Success && "swapOperands failed"); 299 Success = false; 300 MadeChange = true; 301 } else if (RHSBO) { 302 // Turn (A+B)+(C+D) -> (((A+B)+C)+D). This guarantees the the RHS is not 303 // part of the expression tree. 304 LinearizeExpr(I); 305 LHS = LHSBO = cast<BinaryOperator>(I->getOperand(0)); 306 RHS = I->getOperand(1); 307 RHSBO = 0; 308 } 309 310 // Okay, now we know that the LHS is a nested expression and that the RHS is 311 // not. Perform reassociation. 312 assert(!isReassociableOp(RHS, Opcode) && "LinearizeExpr failed!"); 313 314 // Move LHS right before I to make sure that the tree expression dominates all 315 // values. 316 LHSBO->moveBefore(I); 317 318 // Linearize the expression tree on the LHS. 319 LinearizeExprTree(LHSBO, Ops); 320 321 // Remember the RHS operand and its rank. 322 Ops.push_back(ValueEntry(getRank(RHS), RHS)); 323 324 // Clear the RHS leaf out. 325 I->setOperand(1, UndefValue::get(I->getType())); 326} 327 328// RewriteExprTree - Now that the operands for this expression tree are 329// linearized and optimized, emit them in-order. This function is written to be 330// tail recursive. 331void Reassociate::RewriteExprTree(BinaryOperator *I, 332 SmallVectorImpl<ValueEntry> &Ops, 333 unsigned i) { 334 if (i+2 == Ops.size()) { 335 if (I->getOperand(0) != Ops[i].Op || 336 I->getOperand(1) != Ops[i+1].Op) { 337 Value *OldLHS = I->getOperand(0); 338 DEBUG(errs() << "RA: " << *I << '\n'); 339 I->setOperand(0, Ops[i].Op); 340 I->setOperand(1, Ops[i+1].Op); 341 DEBUG(errs() << "TO: " << *I << '\n'); 342 MadeChange = true; 343 ++NumChanged; 344 345 // If we reassociated a tree to fewer operands (e.g. (1+a+2) -> (a+3) 346 // delete the extra, now dead, nodes. 347 RemoveDeadBinaryOp(OldLHS); 348 } 349 return; 350 } 351 assert(i+2 < Ops.size() && "Ops index out of range!"); 352 353 if (I->getOperand(1) != Ops[i].Op) { 354 DEBUG(errs() << "RA: " << *I << '\n'); 355 I->setOperand(1, Ops[i].Op); 356 DEBUG(errs() << "TO: " << *I << '\n'); 357 MadeChange = true; 358 ++NumChanged; 359 } 360 361 BinaryOperator *LHS = cast<BinaryOperator>(I->getOperand(0)); 362 assert(LHS->getOpcode() == I->getOpcode() && 363 "Improper expression tree!"); 364 365 // Compactify the tree instructions together with each other to guarantee 366 // that the expression tree is dominated by all of Ops. 367 LHS->moveBefore(I); 368 RewriteExprTree(LHS, Ops, i+1); 369} 370 371 372 373// NegateValue - Insert instructions before the instruction pointed to by BI, 374// that computes the negative version of the value specified. The negative 375// version of the value is returned, and BI is left pointing at the instruction 376// that should be processed next by the reassociation pass. 377// 378static Value *NegateValue(Value *V, Instruction *BI) { 379 // We are trying to expose opportunity for reassociation. One of the things 380 // that we want to do to achieve this is to push a negation as deep into an 381 // expression chain as possible, to expose the add instructions. In practice, 382 // this means that we turn this: 383 // X = -(A+12+C+D) into X = -A + -12 + -C + -D = -12 + -A + -C + -D 384 // so that later, a: Y = 12+X could get reassociated with the -12 to eliminate 385 // the constants. We assume that instcombine will clean up the mess later if 386 // we introduce tons of unnecessary negation instructions... 387 // 388 if (Instruction *I = dyn_cast<Instruction>(V)) 389 if (I->getOpcode() == Instruction::Add && I->hasOneUse()) { 390 // Push the negates through the add. 391 I->setOperand(0, NegateValue(I->getOperand(0), BI)); 392 I->setOperand(1, NegateValue(I->getOperand(1), BI)); 393 394 // We must move the add instruction here, because the neg instructions do 395 // not dominate the old add instruction in general. By moving it, we are 396 // assured that the neg instructions we just inserted dominate the 397 // instruction we are about to insert after them. 398 // 399 I->moveBefore(BI); 400 I->setName(I->getName()+".neg"); 401 return I; 402 } 403 404 // Insert a 'neg' instruction that subtracts the value from zero to get the 405 // negation. 406 // 407 return BinaryOperator::CreateNeg(V, V->getName() + ".neg", BI); 408} 409 410/// ShouldBreakUpSubtract - Return true if we should break up this subtract of 411/// X-Y into (X + -Y). 412static bool ShouldBreakUpSubtract(Instruction *Sub) { 413 // If this is a negation, we can't split it up! 414 if (BinaryOperator::isNeg(Sub)) 415 return false; 416 417 // Don't bother to break this up unless either the LHS is an associable add or 418 // subtract or if this is only used by one. 419 if (isReassociableOp(Sub->getOperand(0), Instruction::Add) || 420 isReassociableOp(Sub->getOperand(0), Instruction::Sub)) 421 return true; 422 if (isReassociableOp(Sub->getOperand(1), Instruction::Add) || 423 isReassociableOp(Sub->getOperand(1), Instruction::Sub)) 424 return true; 425 if (Sub->hasOneUse() && 426 (isReassociableOp(Sub->use_back(), Instruction::Add) || 427 isReassociableOp(Sub->use_back(), Instruction::Sub))) 428 return true; 429 430 return false; 431} 432 433/// BreakUpSubtract - If we have (X-Y), and if either X is an add, or if this is 434/// only used by an add, transform this into (X+(0-Y)) to promote better 435/// reassociation. 436static Instruction *BreakUpSubtract(Instruction *Sub, 437 std::map<AssertingVH<>, unsigned> &ValueRankMap) { 438 // Convert a subtract into an add and a neg instruction... so that sub 439 // instructions can be commuted with other add instructions... 440 // 441 // Calculate the negative value of Operand 1 of the sub instruction... 442 // and set it as the RHS of the add instruction we just made... 443 // 444 Value *NegVal = NegateValue(Sub->getOperand(1), Sub); 445 Instruction *New = 446 BinaryOperator::CreateAdd(Sub->getOperand(0), NegVal, "", Sub); 447 New->takeName(Sub); 448 449 // Everyone now refers to the add instruction. 450 ValueRankMap.erase(Sub); 451 Sub->replaceAllUsesWith(New); 452 Sub->eraseFromParent(); 453 454 DEBUG(errs() << "Negated: " << *New << '\n'); 455 return New; 456} 457 458/// ConvertShiftToMul - If this is a shift of a reassociable multiply or is used 459/// by one, change this into a multiply by a constant to assist with further 460/// reassociation. 461static Instruction *ConvertShiftToMul(Instruction *Shl, 462 std::map<AssertingVH<>, unsigned> &ValueRankMap) { 463 // If an operand of this shift is a reassociable multiply, or if the shift 464 // is used by a reassociable multiply or add, turn into a multiply. 465 if (isReassociableOp(Shl->getOperand(0), Instruction::Mul) || 466 (Shl->hasOneUse() && 467 (isReassociableOp(Shl->use_back(), Instruction::Mul) || 468 isReassociableOp(Shl->use_back(), Instruction::Add)))) { 469 Constant *MulCst = ConstantInt::get(Shl->getType(), 1); 470 MulCst = ConstantExpr::getShl(MulCst, cast<Constant>(Shl->getOperand(1))); 471 472 Instruction *Mul = 473 BinaryOperator::CreateMul(Shl->getOperand(0), MulCst, "", Shl); 474 ValueRankMap.erase(Shl); 475 Mul->takeName(Shl); 476 Shl->replaceAllUsesWith(Mul); 477 Shl->eraseFromParent(); 478 return Mul; 479 } 480 return 0; 481} 482 483// Scan backwards and forwards among values with the same rank as element i to 484// see if X exists. If X does not exist, return i. 485static unsigned FindInOperandList(SmallVectorImpl<ValueEntry> &Ops, unsigned i, 486 Value *X) { 487 unsigned XRank = Ops[i].Rank; 488 unsigned e = Ops.size(); 489 for (unsigned j = i+1; j != e && Ops[j].Rank == XRank; ++j) 490 if (Ops[j].Op == X) 491 return j; 492 // Scan backwards 493 for (unsigned j = i-1; j != ~0U && Ops[j].Rank == XRank; --j) 494 if (Ops[j].Op == X) 495 return j; 496 return i; 497} 498 499/// EmitAddTreeOfValues - Emit a tree of add instructions, summing Ops together 500/// and returning the result. Insert the tree before I. 501static Value *EmitAddTreeOfValues(Instruction *I, SmallVectorImpl<Value*> &Ops){ 502 if (Ops.size() == 1) return Ops.back(); 503 504 Value *V1 = Ops.back(); 505 Ops.pop_back(); 506 Value *V2 = EmitAddTreeOfValues(I, Ops); 507 return BinaryOperator::CreateAdd(V2, V1, "tmp", I); 508} 509 510/// RemoveFactorFromExpression - If V is an expression tree that is a 511/// multiplication sequence, and if this sequence contains a multiply by Factor, 512/// remove Factor from the tree and return the new tree. 513Value *Reassociate::RemoveFactorFromExpression(Value *V, Value *Factor) { 514 BinaryOperator *BO = isReassociableOp(V, Instruction::Mul); 515 if (!BO) return 0; 516 517 SmallVector<ValueEntry, 8> Factors; 518 LinearizeExprTree(BO, Factors); 519 520 bool FoundFactor = false; 521 for (unsigned i = 0, e = Factors.size(); i != e; ++i) 522 if (Factors[i].Op == Factor) { 523 FoundFactor = true; 524 Factors.erase(Factors.begin()+i); 525 break; 526 } 527 if (!FoundFactor) { 528 // Make sure to restore the operands to the expression tree. 529 RewriteExprTree(BO, Factors); 530 return 0; 531 } 532 533 // If this was just a single multiply, remove the multiply and return the only 534 // remaining operand. 535 if (Factors.size() == 1) { 536 ValueRankMap.erase(BO); 537 BO->eraseFromParent(); 538 return Factors[0].Op; 539 } 540 541 RewriteExprTree(BO, Factors); 542 return BO; 543} 544 545/// FindSingleUseMultiplyFactors - If V is a single-use multiply, recursively 546/// add its operands as factors, otherwise add V to the list of factors. 547static void FindSingleUseMultiplyFactors(Value *V, 548 SmallVectorImpl<Value*> &Factors) { 549 BinaryOperator *BO; 550 if ((!V->hasOneUse() && !V->use_empty()) || 551 !(BO = dyn_cast<BinaryOperator>(V)) || 552 BO->getOpcode() != Instruction::Mul) { 553 Factors.push_back(V); 554 return; 555 } 556 557 // Otherwise, add the LHS and RHS to the list of factors. 558 FindSingleUseMultiplyFactors(BO->getOperand(1), Factors); 559 FindSingleUseMultiplyFactors(BO->getOperand(0), Factors); 560} 561 562/// OptimizeAndOrXor - Optimize a series of operands to an 'and', 'or', or 'xor' 563/// instruction. This optimizes based on identities. If it can be reduced to 564/// a single Value, it is returned, otherwise the Ops list is mutated as 565/// necessary. 566static Value *OptimizeAndOrXor(unsigned Opcode, 567 SmallVectorImpl<ValueEntry> &Ops) { 568 // Scan the operand lists looking for X and ~X pairs, along with X,X pairs. 569 // If we find any, we can simplify the expression. X&~X == 0, X|~X == -1. 570 for (unsigned i = 0, e = Ops.size(); i != e; ++i) { 571 // First, check for X and ~X in the operand list. 572 assert(i < Ops.size()); 573 if (BinaryOperator::isNot(Ops[i].Op)) { // Cannot occur for ^. 574 Value *X = BinaryOperator::getNotArgument(Ops[i].Op); 575 unsigned FoundX = FindInOperandList(Ops, i, X); 576 if (FoundX != i) { 577 if (Opcode == Instruction::And) // ...&X&~X = 0 578 return Constant::getNullValue(X->getType()); 579 580 if (Opcode == Instruction::Or) // ...|X|~X = -1 581 return Constant::getAllOnesValue(X->getType()); 582 } 583 } 584 585 // Next, check for duplicate pairs of values, which we assume are next to 586 // each other, due to our sorting criteria. 587 assert(i < Ops.size()); 588 if (i+1 != Ops.size() && Ops[i+1].Op == Ops[i].Op) { 589 if (Opcode == Instruction::And || Opcode == Instruction::Or) { 590 // Drop duplicate values. 591 Ops.erase(Ops.begin()+i); 592 --i; --e; 593 ++NumAnnihil; 594 } else { 595 assert(Opcode == Instruction::Xor); 596 if (e == 2) 597 return Constant::getNullValue(Ops[0].Op->getType()); 598 599 // ... X^X -> ... 600 Ops.erase(Ops.begin()+i, Ops.begin()+i+2); 601 i -= 1; e -= 2; 602 ++NumAnnihil; 603 } 604 } 605 } 606 return 0; 607} 608 609/// OptimizeAdd - Optimize a series of operands to an 'add' instruction. This 610/// optimizes based on identities. If it can be reduced to a single Value, it 611/// is returned, otherwise the Ops list is mutated as necessary. 612Value *Reassociate::OptimizeAdd(Instruction *I, 613 SmallVectorImpl<ValueEntry> &Ops) { 614 SmallPtrSet<Value*, 8> OperandsSeen; 615 616Restart: 617 OperandsSeen.clear(); 618 619 // Scan the operand lists looking for X and -X pairs. If we find any, we 620 // can simplify the expression. X+-X == 0. While we're at it, scan for any 621 // duplicates. We want to canonicalize Y+Y+Y+Z -> 3*Y+Z. 622 for (unsigned i = 0, e = Ops.size(); i != e; ++i) { 623 Value *TheOp = Ops[i].Op; 624 // Check to see if we've seen this operand before. If so, we factor all 625 // instances of the operand together. 626 if (!OperandsSeen.insert(TheOp)) { 627 // Rescan the list, removing all instances of this operand from the expr. 628 unsigned NumFound = 0; 629 for (unsigned j = 0, je = Ops.size(); j != je; ++j) { 630 if (Ops[j].Op != TheOp) continue; 631 ++NumFound; 632 Ops.erase(Ops.begin()+j); 633 --j; --je; 634 } 635 636 DEBUG(errs() << "\nFACTORING [" << NumFound << "]: " << *TheOp << '\n'); 637 ++NumFactor; 638 639 640 // Insert a new multiply. 641 Value *Mul = ConstantInt::get(cast<IntegerType>(I->getType()), NumFound); 642 Mul = BinaryOperator::CreateMul(TheOp, Mul, "factor", I); 643 644 // Now that we have inserted a multiply, optimize it. This allows us to 645 // handle cases that require multiple factoring steps, such as this: 646 // (X*2) + (X*2) + (X*2) -> (X*2)*3 -> X*6 647 Mul = ReassociateExpression(cast<BinaryOperator>(Mul)); 648 649 // If every add operand was a duplicate, return the multiply. 650 if (Ops.empty()) 651 return Mul; 652 653 // Otherwise, we had some input that didn't have the dupe, such as 654 // "A + A + B" -> "A*2 + B". Add the new multiply to the list of 655 // things being added by this operation. 656 Ops.insert(Ops.begin(), ValueEntry(getRank(Mul), Mul)); 657 goto Restart; 658 } 659 660 // Check for X and -X in the operand list. 661 if (!BinaryOperator::isNeg(TheOp)) 662 continue; 663 664 Value *X = BinaryOperator::getNegArgument(TheOp); 665 unsigned FoundX = FindInOperandList(Ops, i, X); 666 if (FoundX == i) 667 continue; 668 669 // Remove X and -X from the operand list. 670 if (Ops.size() == 2) 671 return Constant::getNullValue(X->getType()); 672 673 Ops.erase(Ops.begin()+i); 674 if (i < FoundX) 675 --FoundX; 676 else 677 --i; // Need to back up an extra one. 678 Ops.erase(Ops.begin()+FoundX); 679 ++NumAnnihil; 680 --i; // Revisit element. 681 e -= 2; // Removed two elements. 682 } 683 684 // Scan the operand list, checking to see if there are any common factors 685 // between operands. Consider something like A*A+A*B*C+D. We would like to 686 // reassociate this to A*(A+B*C)+D, which reduces the number of multiplies. 687 // To efficiently find this, we count the number of times a factor occurs 688 // for any ADD operands that are MULs. 689 DenseMap<Value*, unsigned> FactorOccurrences; 690 691 // Keep track of each multiply we see, to avoid triggering on (X*4)+(X*4) 692 // where they are actually the same multiply. 693 unsigned MaxOcc = 0; 694 Value *MaxOccVal = 0; 695 for (unsigned i = 0, e = Ops.size(); i != e; ++i) { 696 BinaryOperator *BOp = dyn_cast<BinaryOperator>(Ops[i].Op); 697 if (BOp == 0 || BOp->getOpcode() != Instruction::Mul || !BOp->use_empty()) 698 continue; 699 700 // Compute all of the factors of this added value. 701 SmallVector<Value*, 8> Factors; 702 FindSingleUseMultiplyFactors(BOp, Factors); 703 assert(Factors.size() > 1 && "Bad linearize!"); 704 705 // Add one to FactorOccurrences for each unique factor in this op. 706 if (Factors.size() == 2) { 707 unsigned Occ = ++FactorOccurrences[Factors[0]]; 708 if (Occ > MaxOcc) { MaxOcc = Occ; MaxOccVal = Factors[0]; } 709 if (Factors[0] != Factors[1]) { // Don't double count A*A. 710 Occ = ++FactorOccurrences[Factors[1]]; 711 if (Occ > MaxOcc) { MaxOcc = Occ; MaxOccVal = Factors[1]; } 712 } 713 } else { 714 SmallPtrSet<Value*, 4> Duplicates; 715 for (unsigned i = 0, e = Factors.size(); i != e; ++i) { 716 if (!Duplicates.insert(Factors[i])) continue; 717 718 unsigned Occ = ++FactorOccurrences[Factors[i]]; 719 if (Occ > MaxOcc) { MaxOcc = Occ; MaxOccVal = Factors[i]; } 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 ++NumFactor; 728 729 // Create a new instruction that uses the MaxOccVal twice. If we don't do 730 // this, we could otherwise run into situations where removing a factor 731 // from an expression will drop a use of maxocc, and this can cause 732 // RemoveFactorFromExpression on successive values to behave differently. 733 Instruction *DummyInst = BinaryOperator::CreateAdd(MaxOccVal, MaxOccVal); 734 SmallVector<Value*, 4> NewMulOps; 735 for (unsigned i = 0, e = Ops.size(); i != e; ++i) { 736 if (Value *V = RemoveFactorFromExpression(Ops[i].Op, MaxOccVal)) { 737 NewMulOps.push_back(V); 738 Ops.erase(Ops.begin()+i); 739 --i; --e; 740 } 741 } 742 743 // No need for extra uses anymore. 744 delete DummyInst; 745 746 unsigned NumAddedValues = NewMulOps.size(); 747 Value *V = EmitAddTreeOfValues(I, NewMulOps); 748 749 // Now that we have inserted the add tree, optimize it. This allows us to 750 // 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 assert(NumAddedValues > 1 && "Each occurrence should contribute a value"); 753 V = ReassociateExpression(cast<BinaryOperator>(V)); 754 755 // Create the multiply. 756 Value *V2 = BinaryOperator::CreateMul(V, MaxOccVal, "tmp", I); 757 758 // FIXME: Should rerun 'ReassociateExpression' on the mul too?? 759 760 // If every add operand included the factor (e.g. "A*B + A*C"), then the 761 // entire result expression is just the multiply "A*(B+C)". 762 if (Ops.empty()) 763 return V2; 764 765 // Otherwise, we had some input that didn't have the factor, such as 766 // "A*B + A*C + D" -> "A*(B+C) + D". Add the new multiply to the list of 767 // things being added by this operation. 768 Ops.insert(Ops.begin(), ValueEntry(getRank(V2), V2)); 769 } 770 771 return 0; 772} 773 774Value *Reassociate::OptimizeExpression(BinaryOperator *I, 775 SmallVectorImpl<ValueEntry> &Ops) { 776 // Now that we have the linearized expression tree, try to optimize it. 777 // Start by folding any constants that we found. 778 bool IterateOptimization = false; 779 if (Ops.size() == 1) return Ops[0].Op; 780 781 unsigned Opcode = I->getOpcode(); 782 783 if (Constant *V1 = dyn_cast<Constant>(Ops[Ops.size()-2].Op)) 784 if (Constant *V2 = dyn_cast<Constant>(Ops.back().Op)) { 785 Ops.pop_back(); 786 Ops.back().Op = ConstantExpr::get(Opcode, V1, V2); 787 return OptimizeExpression(I, Ops); 788 } 789 790 // Check for destructive annihilation due to a constant being used. 791 if (ConstantInt *CstVal = dyn_cast<ConstantInt>(Ops.back().Op)) 792 switch (Opcode) { 793 default: break; 794 case Instruction::And: 795 if (CstVal->isZero()) // ... & 0 -> 0 796 return CstVal; 797 if (CstVal->isAllOnesValue()) // ... & -1 -> ... 798 Ops.pop_back(); 799 break; 800 case Instruction::Mul: 801 if (CstVal->isZero()) { // ... * 0 -> 0 802 ++NumAnnihil; 803 return CstVal; 804 } 805 806 if (cast<ConstantInt>(CstVal)->isOne()) 807 Ops.pop_back(); // ... * 1 -> ... 808 break; 809 case Instruction::Or: 810 if (CstVal->isAllOnesValue()) // ... | -1 -> -1 811 return CstVal; 812 // FALLTHROUGH! 813 case Instruction::Add: 814 case Instruction::Xor: 815 if (CstVal->isZero()) // ... [|^+] 0 -> ... 816 Ops.pop_back(); 817 break; 818 } 819 if (Ops.size() == 1) return Ops[0].Op; 820 821 // Handle destructive annihilation due to identities between elements in the 822 // argument list here. 823 switch (Opcode) { 824 default: break; 825 case Instruction::And: 826 case Instruction::Or: 827 case Instruction::Xor: { 828 unsigned NumOps = Ops.size(); 829 if (Value *Result = OptimizeAndOrXor(Opcode, Ops)) 830 return Result; 831 IterateOptimization |= Ops.size() != NumOps; 832 break; 833 } 834 835 case Instruction::Add: { 836 unsigned NumOps = Ops.size(); 837 if (Value *Result = OptimizeAdd(I, Ops)) 838 return Result; 839 IterateOptimization |= Ops.size() != NumOps; 840 } 841 842 break; 843 //case Instruction::Mul: 844 } 845 846 if (IterateOptimization) 847 return OptimizeExpression(I, Ops); 848 return 0; 849} 850 851 852/// ReassociateBB - Inspect all of the instructions in this basic block, 853/// reassociating them as we go. 854void Reassociate::ReassociateBB(BasicBlock *BB) { 855 for (BasicBlock::iterator BBI = BB->begin(); BBI != BB->end(); ) { 856 Instruction *BI = BBI++; 857 if (BI->getOpcode() == Instruction::Shl && 858 isa<ConstantInt>(BI->getOperand(1))) 859 if (Instruction *NI = ConvertShiftToMul(BI, ValueRankMap)) { 860 MadeChange = true; 861 BI = NI; 862 } 863 864 // Reject cases where it is pointless to do this. 865 if (!isa<BinaryOperator>(BI) || BI->getType()->isFloatingPoint() || 866 isa<VectorType>(BI->getType())) 867 continue; // Floating point ops are not associative. 868 869 // If this is a subtract instruction which is not already in negate form, 870 // see if we can convert it to X+-Y. 871 if (BI->getOpcode() == Instruction::Sub) { 872 if (ShouldBreakUpSubtract(BI)) { 873 BI = BreakUpSubtract(BI, ValueRankMap); 874 MadeChange = true; 875 } else if (BinaryOperator::isNeg(BI)) { 876 // Otherwise, this is a negation. See if the operand is a multiply tree 877 // and if this is not an inner node of a multiply tree. 878 if (isReassociableOp(BI->getOperand(1), Instruction::Mul) && 879 (!BI->hasOneUse() || 880 !isReassociableOp(BI->use_back(), Instruction::Mul))) { 881 BI = LowerNegateToMultiply(BI, ValueRankMap); 882 MadeChange = true; 883 } 884 } 885 } 886 887 // If this instruction is a commutative binary operator, process it. 888 if (!BI->isAssociative()) continue; 889 BinaryOperator *I = cast<BinaryOperator>(BI); 890 891 // If this is an interior node of a reassociable tree, ignore it until we 892 // get to the root of the tree, to avoid N^2 analysis. 893 if (I->hasOneUse() && isReassociableOp(I->use_back(), I->getOpcode())) 894 continue; 895 896 // If this is an add tree that is used by a sub instruction, ignore it 897 // until we process the subtract. 898 if (I->hasOneUse() && I->getOpcode() == Instruction::Add && 899 cast<Instruction>(I->use_back())->getOpcode() == Instruction::Sub) 900 continue; 901 902 ReassociateExpression(I); 903 } 904} 905 906Value *Reassociate::ReassociateExpression(BinaryOperator *I) { 907 908 // First, walk the expression tree, linearizing the tree, collecting the 909 // operand information. 910 SmallVector<ValueEntry, 8> Ops; 911 LinearizeExprTree(I, Ops); 912 913 DEBUG(errs() << "RAIn:\t"; PrintOps(I, Ops); errs() << '\n'); 914 915 // Now that we have linearized the tree to a list and have gathered all of 916 // the operands and their ranks, sort the operands by their rank. Use a 917 // stable_sort so that values with equal ranks will have their relative 918 // positions maintained (and so the compiler is deterministic). Note that 919 // this sorts so that the highest ranking values end up at the beginning of 920 // the vector. 921 std::stable_sort(Ops.begin(), Ops.end()); 922 923 // OptimizeExpression - Now that we have the expression tree in a convenient 924 // sorted form, optimize it globally if possible. 925 if (Value *V = OptimizeExpression(I, Ops)) { 926 // This expression tree simplified to something that isn't a tree, 927 // eliminate it. 928 DEBUG(errs() << "Reassoc to scalar: " << *V << '\n'); 929 I->replaceAllUsesWith(V); 930 RemoveDeadBinaryOp(I); 931 ++NumAnnihil; 932 return V; 933 } 934 935 // We want to sink immediates as deeply as possible except in the case where 936 // this is a multiply tree used only by an add, and the immediate is a -1. 937 // In this case we reassociate to put the negation on the outside so that we 938 // can fold the negation into the add: (-X)*Y + Z -> Z-X*Y 939 if (I->getOpcode() == Instruction::Mul && I->hasOneUse() && 940 cast<Instruction>(I->use_back())->getOpcode() == Instruction::Add && 941 isa<ConstantInt>(Ops.back().Op) && 942 cast<ConstantInt>(Ops.back().Op)->isAllOnesValue()) { 943 ValueEntry Tmp = Ops.pop_back_val(); 944 Ops.insert(Ops.begin(), Tmp); 945 } 946 947 DEBUG(errs() << "RAOut:\t"; PrintOps(I, Ops); errs() << '\n'); 948 949 if (Ops.size() == 1) { 950 // This expression tree simplified to something that isn't a tree, 951 // eliminate it. 952 I->replaceAllUsesWith(Ops[0].Op); 953 RemoveDeadBinaryOp(I); 954 return Ops[0].Op; 955 } 956 957 // Now that we ordered and optimized the expressions, splat them back into 958 // the expression tree, removing any unneeded nodes. 959 RewriteExprTree(I, Ops); 960 return I; 961} 962 963 964bool Reassociate::runOnFunction(Function &F) { 965 // Recalculate the rank map for F 966 BuildRankMap(F); 967 968 MadeChange = false; 969 for (Function::iterator FI = F.begin(), FE = F.end(); FI != FE; ++FI) 970 ReassociateBB(FI); 971 972 // We are done with the rank map... 973 RankMap.clear(); 974 ValueRankMap.clear(); 975 return MadeChange; 976} 977 978