Reassociate.cpp revision 081c34b725980f995be9080eaec24cd3dfaaf065
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, etc. 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> 40using namespace llvm; 41 42STATISTIC(NumLinear , "Number of insts linearized"); 43STATISTIC(NumChanged, "Number of insts reassociated"); 44STATISTIC(NumAnnihil, "Number of expr tree annihilated"); 45STATISTIC(NumFactor , "Number of multiplies factored"); 46 47namespace { 48 struct ValueEntry { 49 unsigned Rank; 50 Value *Op; 51 ValueEntry(unsigned R, Value *O) : Rank(R), Op(O) {} 52 }; 53 inline bool operator<(const ValueEntry &LHS, const ValueEntry &RHS) { 54 return LHS.Rank > RHS.Rank; // Sort so that highest rank goes to start. 55 } 56} 57 58#ifndef NDEBUG 59/// PrintOps - Print out the expression identified in the Ops list. 60/// 61static void PrintOps(Instruction *I, const SmallVectorImpl<ValueEntry> &Ops) { 62 Module *M = I->getParent()->getParent()->getParent(); 63 dbgs() << Instruction::getOpcodeName(I->getOpcode()) << " " 64 << *Ops[0].Op->getType() << '\t'; 65 for (unsigned i = 0, e = Ops.size(); i != e; ++i) { 66 dbgs() << "[ "; 67 WriteAsOperand(dbgs(), Ops[i].Op, false, M); 68 dbgs() << ", #" << Ops[i].Rank << "] "; 69 } 70} 71#endif 72 73namespace { 74 class Reassociate : public FunctionPass { 75 DenseMap<BasicBlock*, unsigned> RankMap; 76 DenseMap<AssertingVH<>, unsigned> ValueRankMap; 77 bool MadeChange; 78 public: 79 static char ID; // Pass identification, replacement for typeid 80 Reassociate() : FunctionPass(ID) { 81 initializeReassociatePass(*PassRegistry::getPassRegistry()); 82 } 83 84 bool runOnFunction(Function &F); 85 86 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 87 AU.setPreservesCFG(); 88 } 89 private: 90 void BuildRankMap(Function &F); 91 unsigned getRank(Value *V); 92 Value *ReassociateExpression(BinaryOperator *I); 93 void RewriteExprTree(BinaryOperator *I, SmallVectorImpl<ValueEntry> &Ops, 94 unsigned Idx = 0); 95 Value *OptimizeExpression(BinaryOperator *I, 96 SmallVectorImpl<ValueEntry> &Ops); 97 Value *OptimizeAdd(Instruction *I, SmallVectorImpl<ValueEntry> &Ops); 98 void LinearizeExprTree(BinaryOperator *I, SmallVectorImpl<ValueEntry> &Ops); 99 void LinearizeExpr(BinaryOperator *I); 100 Value *RemoveFactorFromExpression(Value *V, Value *Factor); 101 void ReassociateBB(BasicBlock *BB); 102 103 void RemoveDeadBinaryOp(Value *V); 104 }; 105} 106 107char Reassociate::ID = 0; 108INITIALIZE_PASS(Reassociate, "reassociate", 109 "Reassociate expressions", false, false) 110 111// Public interface to the Reassociate pass 112FunctionPass *llvm::createReassociatePass() { return new Reassociate(); } 113 114void Reassociate::RemoveDeadBinaryOp(Value *V) { 115 Instruction *Op = dyn_cast<Instruction>(V); 116 if (!Op || !isa<BinaryOperator>(Op) || !Op->use_empty()) 117 return; 118 119 Value *LHS = Op->getOperand(0), *RHS = Op->getOperand(1); 120 121 ValueRankMap.erase(Op); 122 Op->eraseFromParent(); 123 RemoveDeadBinaryOp(LHS); 124 RemoveDeadBinaryOp(RHS); 125} 126 127 128static bool isUnmovableInstruction(Instruction *I) { 129 if (I->getOpcode() == Instruction::PHI || 130 I->getOpcode() == Instruction::Alloca || 131 I->getOpcode() == Instruction::Load || 132 I->getOpcode() == Instruction::Invoke || 133 (I->getOpcode() == Instruction::Call && 134 !isa<DbgInfoIntrinsic>(I)) || 135 I->getOpcode() == Instruction::UDiv || 136 I->getOpcode() == Instruction::SDiv || 137 I->getOpcode() == Instruction::FDiv || 138 I->getOpcode() == Instruction::URem || 139 I->getOpcode() == Instruction::SRem || 140 I->getOpcode() == Instruction::FRem) 141 return true; 142 return false; 143} 144 145void Reassociate::BuildRankMap(Function &F) { 146 unsigned i = 2; 147 148 // Assign distinct ranks to function arguments 149 for (Function::arg_iterator I = F.arg_begin(), E = F.arg_end(); I != E; ++I) 150 ValueRankMap[&*I] = ++i; 151 152 ReversePostOrderTraversal<Function*> RPOT(&F); 153 for (ReversePostOrderTraversal<Function*>::rpo_iterator I = RPOT.begin(), 154 E = RPOT.end(); I != E; ++I) { 155 BasicBlock *BB = *I; 156 unsigned BBRank = RankMap[BB] = ++i << 16; 157 158 // Walk the basic block, adding precomputed ranks for any instructions that 159 // we cannot move. This ensures that the ranks for these instructions are 160 // all different in the block. 161 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) 162 if (isUnmovableInstruction(I)) 163 ValueRankMap[&*I] = ++BBRank; 164 } 165} 166 167unsigned Reassociate::getRank(Value *V) { 168 Instruction *I = dyn_cast<Instruction>(V); 169 if (I == 0) { 170 if (isa<Argument>(V)) return ValueRankMap[V]; // Function argument. 171 return 0; // Otherwise it's a global or constant, rank 0. 172 } 173 174 if (unsigned Rank = ValueRankMap[I]) 175 return Rank; // Rank already known? 176 177 // If this is an expression, return the 1+MAX(rank(LHS), rank(RHS)) so that 178 // we can reassociate expressions for code motion! Since we do not recurse 179 // for PHI nodes, we cannot have infinite recursion here, because there 180 // cannot be loops in the value graph that do not go through PHI nodes. 181 unsigned Rank = 0, MaxRank = RankMap[I->getParent()]; 182 for (unsigned i = 0, e = I->getNumOperands(); 183 i != e && Rank != MaxRank; ++i) 184 Rank = std::max(Rank, getRank(I->getOperand(i))); 185 186 // If this is a not or neg instruction, do not count it for rank. This 187 // assures us that X and ~X will have the same rank. 188 if (!I->getType()->isIntegerTy() || 189 (!BinaryOperator::isNot(I) && !BinaryOperator::isNeg(I))) 190 ++Rank; 191 192 //DEBUG(dbgs() << "Calculated Rank[" << V->getName() << "] = " 193 // << Rank << "\n"); 194 195 return ValueRankMap[I] = Rank; 196} 197 198/// isReassociableOp - Return true if V is an instruction of the specified 199/// opcode and if it only has one use. 200static BinaryOperator *isReassociableOp(Value *V, unsigned Opcode) { 201 if ((V->hasOneUse() || V->use_empty()) && isa<Instruction>(V) && 202 cast<Instruction>(V)->getOpcode() == Opcode) 203 return cast<BinaryOperator>(V); 204 return 0; 205} 206 207/// LowerNegateToMultiply - Replace 0-X with X*-1. 208/// 209static Instruction *LowerNegateToMultiply(Instruction *Neg, 210 DenseMap<AssertingVH<>, unsigned> &ValueRankMap) { 211 Constant *Cst = Constant::getAllOnesValue(Neg->getType()); 212 213 Instruction *Res = BinaryOperator::CreateMul(Neg->getOperand(1), Cst, "",Neg); 214 ValueRankMap.erase(Neg); 215 Res->takeName(Neg); 216 Neg->replaceAllUsesWith(Res); 217 Neg->eraseFromParent(); 218 return Res; 219} 220 221// Given an expression of the form '(A+B)+(D+C)', turn it into '(((A+B)+C)+D)'. 222// Note that if D is also part of the expression tree that we recurse to 223// linearize it as well. Besides that case, this does not recurse into A,B, or 224// C. 225void Reassociate::LinearizeExpr(BinaryOperator *I) { 226 BinaryOperator *LHS = cast<BinaryOperator>(I->getOperand(0)); 227 BinaryOperator *RHS = cast<BinaryOperator>(I->getOperand(1)); 228 assert(isReassociableOp(LHS, I->getOpcode()) && 229 isReassociableOp(RHS, I->getOpcode()) && 230 "Not an expression that needs linearization?"); 231 232 DEBUG(dbgs() << "Linear" << *LHS << '\n' << *RHS << '\n' << *I << '\n'); 233 234 // Move the RHS instruction to live immediately before I, avoiding breaking 235 // dominator properties. 236 RHS->moveBefore(I); 237 238 // Move operands around to do the linearization. 239 I->setOperand(1, RHS->getOperand(0)); 240 RHS->setOperand(0, LHS); 241 I->setOperand(0, RHS); 242 243 ++NumLinear; 244 MadeChange = true; 245 DEBUG(dbgs() << "Linearized: " << *I << '\n'); 246 247 // If D is part of this expression tree, tail recurse. 248 if (isReassociableOp(I->getOperand(1), I->getOpcode())) 249 LinearizeExpr(I); 250} 251 252 253/// LinearizeExprTree - Given an associative binary expression tree, traverse 254/// all of the uses putting it into canonical form. This forces a left-linear 255/// form of the expression (((a+b)+c)+d), and collects information about the 256/// rank of the non-tree operands. 257/// 258/// NOTE: These intentionally destroys the expression tree operands (turning 259/// them into undef values) to reduce #uses of the values. This means that the 260/// caller MUST use something like RewriteExprTree to put the values back in. 261/// 262void Reassociate::LinearizeExprTree(BinaryOperator *I, 263 SmallVectorImpl<ValueEntry> &Ops) { 264 Value *LHS = I->getOperand(0), *RHS = I->getOperand(1); 265 unsigned Opcode = I->getOpcode(); 266 267 // First step, linearize the expression if it is in ((A+B)+(C+D)) form. 268 BinaryOperator *LHSBO = isReassociableOp(LHS, Opcode); 269 BinaryOperator *RHSBO = isReassociableOp(RHS, Opcode); 270 271 // If this is a multiply expression tree and it contains internal negations, 272 // transform them into multiplies by -1 so they can be reassociated. 273 if (I->getOpcode() == Instruction::Mul) { 274 if (!LHSBO && LHS->hasOneUse() && BinaryOperator::isNeg(LHS)) { 275 LHS = LowerNegateToMultiply(cast<Instruction>(LHS), ValueRankMap); 276 LHSBO = isReassociableOp(LHS, Opcode); 277 } 278 if (!RHSBO && RHS->hasOneUse() && BinaryOperator::isNeg(RHS)) { 279 RHS = LowerNegateToMultiply(cast<Instruction>(RHS), ValueRankMap); 280 RHSBO = isReassociableOp(RHS, Opcode); 281 } 282 } 283 284 if (!LHSBO) { 285 if (!RHSBO) { 286 // Neither the LHS or RHS as part of the tree, thus this is a leaf. As 287 // such, just remember these operands and their rank. 288 Ops.push_back(ValueEntry(getRank(LHS), LHS)); 289 Ops.push_back(ValueEntry(getRank(RHS), RHS)); 290 291 // Clear the leaves out. 292 I->setOperand(0, UndefValue::get(I->getType())); 293 I->setOperand(1, UndefValue::get(I->getType())); 294 return; 295 } 296 297 // Turn X+(Y+Z) -> (Y+Z)+X 298 std::swap(LHSBO, RHSBO); 299 std::swap(LHS, RHS); 300 bool Success = !I->swapOperands(); 301 assert(Success && "swapOperands failed"); 302 Success = false; 303 MadeChange = true; 304 } else if (RHSBO) { 305 // Turn (A+B)+(C+D) -> (((A+B)+C)+D). This guarantees the RHS is not 306 // part of the expression tree. 307 LinearizeExpr(I); 308 LHS = LHSBO = cast<BinaryOperator>(I->getOperand(0)); 309 RHS = I->getOperand(1); 310 RHSBO = 0; 311 } 312 313 // Okay, now we know that the LHS is a nested expression and that the RHS is 314 // not. Perform reassociation. 315 assert(!isReassociableOp(RHS, Opcode) && "LinearizeExpr failed!"); 316 317 // Move LHS right before I to make sure that the tree expression dominates all 318 // values. 319 LHSBO->moveBefore(I); 320 321 // Linearize the expression tree on the LHS. 322 LinearizeExprTree(LHSBO, Ops); 323 324 // Remember the RHS operand and its rank. 325 Ops.push_back(ValueEntry(getRank(RHS), RHS)); 326 327 // Clear the RHS leaf out. 328 I->setOperand(1, UndefValue::get(I->getType())); 329} 330 331// RewriteExprTree - Now that the operands for this expression tree are 332// linearized and optimized, emit them in-order. This function is written to be 333// tail recursive. 334void Reassociate::RewriteExprTree(BinaryOperator *I, 335 SmallVectorImpl<ValueEntry> &Ops, 336 unsigned i) { 337 if (i+2 == Ops.size()) { 338 if (I->getOperand(0) != Ops[i].Op || 339 I->getOperand(1) != Ops[i+1].Op) { 340 Value *OldLHS = I->getOperand(0); 341 DEBUG(dbgs() << "RA: " << *I << '\n'); 342 I->setOperand(0, Ops[i].Op); 343 I->setOperand(1, Ops[i+1].Op); 344 DEBUG(dbgs() << "TO: " << *I << '\n'); 345 MadeChange = true; 346 ++NumChanged; 347 348 // If we reassociated a tree to fewer operands (e.g. (1+a+2) -> (a+3) 349 // delete the extra, now dead, nodes. 350 RemoveDeadBinaryOp(OldLHS); 351 } 352 return; 353 } 354 assert(i+2 < Ops.size() && "Ops index out of range!"); 355 356 if (I->getOperand(1) != Ops[i].Op) { 357 DEBUG(dbgs() << "RA: " << *I << '\n'); 358 I->setOperand(1, Ops[i].Op); 359 DEBUG(dbgs() << "TO: " << *I << '\n'); 360 MadeChange = true; 361 ++NumChanged; 362 } 363 364 BinaryOperator *LHS = cast<BinaryOperator>(I->getOperand(0)); 365 assert(LHS->getOpcode() == I->getOpcode() && 366 "Improper expression tree!"); 367 368 // Compactify the tree instructions together with each other to guarantee 369 // that the expression tree is dominated by all of Ops. 370 LHS->moveBefore(I); 371 RewriteExprTree(LHS, Ops, i+1); 372} 373 374 375 376// NegateValue - Insert instructions before the instruction pointed to by BI, 377// that computes the negative version of the value specified. The negative 378// version of the value is returned, and BI is left pointing at the instruction 379// that should be processed next by the reassociation pass. 380// 381static Value *NegateValue(Value *V, Instruction *BI) { 382 if (Constant *C = dyn_cast<Constant>(V)) 383 return ConstantExpr::getNeg(C); 384 385 // We are trying to expose opportunity for reassociation. One of the things 386 // that we want to do to achieve this is to push a negation as deep into an 387 // expression chain as possible, to expose the add instructions. In practice, 388 // this means that we turn this: 389 // X = -(A+12+C+D) into X = -A + -12 + -C + -D = -12 + -A + -C + -D 390 // so that later, a: Y = 12+X could get reassociated with the -12 to eliminate 391 // the constants. We assume that instcombine will clean up the mess later if 392 // we introduce tons of unnecessary negation instructions. 393 // 394 if (Instruction *I = dyn_cast<Instruction>(V)) 395 if (I->getOpcode() == Instruction::Add && I->hasOneUse()) { 396 // Push the negates through the add. 397 I->setOperand(0, NegateValue(I->getOperand(0), BI)); 398 I->setOperand(1, NegateValue(I->getOperand(1), BI)); 399 400 // We must move the add instruction here, because the neg instructions do 401 // not dominate the old add instruction in general. By moving it, we are 402 // assured that the neg instructions we just inserted dominate the 403 // instruction we are about to insert after them. 404 // 405 I->moveBefore(BI); 406 I->setName(I->getName()+".neg"); 407 return I; 408 } 409 410 // Okay, we need to materialize a negated version of V with an instruction. 411 // Scan the use lists of V to see if we have one already. 412 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E;++UI){ 413 User *U = *UI; 414 if (!BinaryOperator::isNeg(U)) continue; 415 416 // We found one! Now we have to make sure that the definition dominates 417 // this use. We do this by moving it to the entry block (if it is a 418 // non-instruction value) or right after the definition. These negates will 419 // be zapped by reassociate later, so we don't need much finesse here. 420 BinaryOperator *TheNeg = cast<BinaryOperator>(U); 421 422 // Verify that the negate is in this function, V might be a constant expr. 423 if (TheNeg->getParent()->getParent() != BI->getParent()->getParent()) 424 continue; 425 426 BasicBlock::iterator InsertPt; 427 if (Instruction *InstInput = dyn_cast<Instruction>(V)) { 428 if (InvokeInst *II = dyn_cast<InvokeInst>(InstInput)) { 429 InsertPt = II->getNormalDest()->begin(); 430 } else { 431 InsertPt = InstInput; 432 ++InsertPt; 433 } 434 while (isa<PHINode>(InsertPt)) ++InsertPt; 435 } else { 436 InsertPt = TheNeg->getParent()->getParent()->getEntryBlock().begin(); 437 } 438 TheNeg->moveBefore(InsertPt); 439 return TheNeg; 440 } 441 442 // Insert a 'neg' instruction that subtracts the value from zero to get the 443 // negation. 444 return BinaryOperator::CreateNeg(V, V->getName() + ".neg", BI); 445} 446 447/// ShouldBreakUpSubtract - Return true if we should break up this subtract of 448/// X-Y into (X + -Y). 449static bool ShouldBreakUpSubtract(Instruction *Sub) { 450 // If this is a negation, we can't split it up! 451 if (BinaryOperator::isNeg(Sub)) 452 return false; 453 454 // Don't bother to break this up unless either the LHS is an associable add or 455 // subtract or if this is only used by one. 456 if (isReassociableOp(Sub->getOperand(0), Instruction::Add) || 457 isReassociableOp(Sub->getOperand(0), Instruction::Sub)) 458 return true; 459 if (isReassociableOp(Sub->getOperand(1), Instruction::Add) || 460 isReassociableOp(Sub->getOperand(1), Instruction::Sub)) 461 return true; 462 if (Sub->hasOneUse() && 463 (isReassociableOp(Sub->use_back(), Instruction::Add) || 464 isReassociableOp(Sub->use_back(), Instruction::Sub))) 465 return true; 466 467 return false; 468} 469 470/// BreakUpSubtract - If we have (X-Y), and if either X is an add, or if this is 471/// only used by an add, transform this into (X+(0-Y)) to promote better 472/// reassociation. 473static Instruction *BreakUpSubtract(Instruction *Sub, 474 DenseMap<AssertingVH<>, unsigned> &ValueRankMap) { 475 // Convert a subtract into an add and a neg instruction. This allows sub 476 // instructions to be commuted with other add instructions. 477 // 478 // Calculate the negative value of Operand 1 of the sub instruction, 479 // and set it as the RHS of the add instruction we just made. 480 // 481 Value *NegVal = NegateValue(Sub->getOperand(1), Sub); 482 Instruction *New = 483 BinaryOperator::CreateAdd(Sub->getOperand(0), NegVal, "", Sub); 484 New->takeName(Sub); 485 486 // Everyone now refers to the add instruction. 487 ValueRankMap.erase(Sub); 488 Sub->replaceAllUsesWith(New); 489 Sub->eraseFromParent(); 490 491 DEBUG(dbgs() << "Negated: " << *New << '\n'); 492 return New; 493} 494 495/// ConvertShiftToMul - If this is a shift of a reassociable multiply or is used 496/// by one, change this into a multiply by a constant to assist with further 497/// reassociation. 498static Instruction *ConvertShiftToMul(Instruction *Shl, 499 DenseMap<AssertingVH<>, unsigned> &ValueRankMap) { 500 // If an operand of this shift is a reassociable multiply, or if the shift 501 // is used by a reassociable multiply or add, turn into a multiply. 502 if (isReassociableOp(Shl->getOperand(0), Instruction::Mul) || 503 (Shl->hasOneUse() && 504 (isReassociableOp(Shl->use_back(), Instruction::Mul) || 505 isReassociableOp(Shl->use_back(), Instruction::Add)))) { 506 Constant *MulCst = ConstantInt::get(Shl->getType(), 1); 507 MulCst = ConstantExpr::getShl(MulCst, cast<Constant>(Shl->getOperand(1))); 508 509 Instruction *Mul = 510 BinaryOperator::CreateMul(Shl->getOperand(0), MulCst, "", Shl); 511 ValueRankMap.erase(Shl); 512 Mul->takeName(Shl); 513 Shl->replaceAllUsesWith(Mul); 514 Shl->eraseFromParent(); 515 return Mul; 516 } 517 return 0; 518} 519 520// Scan backwards and forwards among values with the same rank as element i to 521// see if X exists. If X does not exist, return i. This is useful when 522// scanning for 'x' when we see '-x' because they both get the same rank. 523static unsigned FindInOperandList(SmallVectorImpl<ValueEntry> &Ops, unsigned i, 524 Value *X) { 525 unsigned XRank = Ops[i].Rank; 526 unsigned e = Ops.size(); 527 for (unsigned j = i+1; j != e && Ops[j].Rank == XRank; ++j) 528 if (Ops[j].Op == X) 529 return j; 530 // Scan backwards. 531 for (unsigned j = i-1; j != ~0U && Ops[j].Rank == XRank; --j) 532 if (Ops[j].Op == X) 533 return j; 534 return i; 535} 536 537/// EmitAddTreeOfValues - Emit a tree of add instructions, summing Ops together 538/// and returning the result. Insert the tree before I. 539static Value *EmitAddTreeOfValues(Instruction *I, SmallVectorImpl<Value*> &Ops){ 540 if (Ops.size() == 1) return Ops.back(); 541 542 Value *V1 = Ops.back(); 543 Ops.pop_back(); 544 Value *V2 = EmitAddTreeOfValues(I, Ops); 545 return BinaryOperator::CreateAdd(V2, V1, "tmp", I); 546} 547 548/// RemoveFactorFromExpression - If V is an expression tree that is a 549/// multiplication sequence, and if this sequence contains a multiply by Factor, 550/// remove Factor from the tree and return the new tree. 551Value *Reassociate::RemoveFactorFromExpression(Value *V, Value *Factor) { 552 BinaryOperator *BO = isReassociableOp(V, Instruction::Mul); 553 if (!BO) return 0; 554 555 SmallVector<ValueEntry, 8> Factors; 556 LinearizeExprTree(BO, Factors); 557 558 bool FoundFactor = false; 559 bool NeedsNegate = false; 560 for (unsigned i = 0, e = Factors.size(); i != e; ++i) { 561 if (Factors[i].Op == Factor) { 562 FoundFactor = true; 563 Factors.erase(Factors.begin()+i); 564 break; 565 } 566 567 // If this is a negative version of this factor, remove it. 568 if (ConstantInt *FC1 = dyn_cast<ConstantInt>(Factor)) 569 if (ConstantInt *FC2 = dyn_cast<ConstantInt>(Factors[i].Op)) 570 if (FC1->getValue() == -FC2->getValue()) { 571 FoundFactor = NeedsNegate = true; 572 Factors.erase(Factors.begin()+i); 573 break; 574 } 575 } 576 577 if (!FoundFactor) { 578 // Make sure to restore the operands to the expression tree. 579 RewriteExprTree(BO, Factors); 580 return 0; 581 } 582 583 BasicBlock::iterator InsertPt = BO; ++InsertPt; 584 585 // If this was just a single multiply, remove the multiply and return the only 586 // remaining operand. 587 if (Factors.size() == 1) { 588 ValueRankMap.erase(BO); 589 BO->eraseFromParent(); 590 V = Factors[0].Op; 591 } else { 592 RewriteExprTree(BO, Factors); 593 V = BO; 594 } 595 596 if (NeedsNegate) 597 V = BinaryOperator::CreateNeg(V, "neg", InsertPt); 598 599 return V; 600} 601 602/// FindSingleUseMultiplyFactors - If V is a single-use multiply, recursively 603/// add its operands as factors, otherwise add V to the list of factors. 604/// 605/// Ops is the top-level list of add operands we're trying to factor. 606static void FindSingleUseMultiplyFactors(Value *V, 607 SmallVectorImpl<Value*> &Factors, 608 const SmallVectorImpl<ValueEntry> &Ops, 609 bool IsRoot) { 610 BinaryOperator *BO; 611 if (!(V->hasOneUse() || V->use_empty()) || // More than one use. 612 !(BO = dyn_cast<BinaryOperator>(V)) || 613 BO->getOpcode() != Instruction::Mul) { 614 Factors.push_back(V); 615 return; 616 } 617 618 // If this value has a single use because it is another input to the add 619 // tree we're reassociating and we dropped its use, it actually has two 620 // uses and we can't factor it. 621 if (!IsRoot) { 622 for (unsigned i = 0, e = Ops.size(); i != e; ++i) 623 if (Ops[i].Op == V) { 624 Factors.push_back(V); 625 return; 626 } 627 } 628 629 630 // Otherwise, add the LHS and RHS to the list of factors. 631 FindSingleUseMultiplyFactors(BO->getOperand(1), Factors, Ops, false); 632 FindSingleUseMultiplyFactors(BO->getOperand(0), Factors, Ops, false); 633} 634 635/// OptimizeAndOrXor - Optimize a series of operands to an 'and', 'or', or 'xor' 636/// instruction. This optimizes based on identities. If it can be reduced to 637/// a single Value, it is returned, otherwise the Ops list is mutated as 638/// necessary. 639static Value *OptimizeAndOrXor(unsigned Opcode, 640 SmallVectorImpl<ValueEntry> &Ops) { 641 // Scan the operand lists looking for X and ~X pairs, along with X,X pairs. 642 // If we find any, we can simplify the expression. X&~X == 0, X|~X == -1. 643 for (unsigned i = 0, e = Ops.size(); i != e; ++i) { 644 // First, check for X and ~X in the operand list. 645 assert(i < Ops.size()); 646 if (BinaryOperator::isNot(Ops[i].Op)) { // Cannot occur for ^. 647 Value *X = BinaryOperator::getNotArgument(Ops[i].Op); 648 unsigned FoundX = FindInOperandList(Ops, i, X); 649 if (FoundX != i) { 650 if (Opcode == Instruction::And) // ...&X&~X = 0 651 return Constant::getNullValue(X->getType()); 652 653 if (Opcode == Instruction::Or) // ...|X|~X = -1 654 return Constant::getAllOnesValue(X->getType()); 655 } 656 } 657 658 // Next, check for duplicate pairs of values, which we assume are next to 659 // each other, due to our sorting criteria. 660 assert(i < Ops.size()); 661 if (i+1 != Ops.size() && Ops[i+1].Op == Ops[i].Op) { 662 if (Opcode == Instruction::And || Opcode == Instruction::Or) { 663 // Drop duplicate values for And and Or. 664 Ops.erase(Ops.begin()+i); 665 --i; --e; 666 ++NumAnnihil; 667 continue; 668 } 669 670 // Drop pairs of values for Xor. 671 assert(Opcode == Instruction::Xor); 672 if (e == 2) 673 return Constant::getNullValue(Ops[0].Op->getType()); 674 675 // Y ^ X^X -> Y 676 Ops.erase(Ops.begin()+i, Ops.begin()+i+2); 677 i -= 1; e -= 2; 678 ++NumAnnihil; 679 } 680 } 681 return 0; 682} 683 684/// OptimizeAdd - Optimize a series of operands to an 'add' instruction. This 685/// optimizes based on identities. If it can be reduced to a single Value, it 686/// is returned, otherwise the Ops list is mutated as necessary. 687Value *Reassociate::OptimizeAdd(Instruction *I, 688 SmallVectorImpl<ValueEntry> &Ops) { 689 // Scan the operand lists looking for X and -X pairs. If we find any, we 690 // can simplify the expression. X+-X == 0. While we're at it, scan for any 691 // duplicates. We want to canonicalize Y+Y+Y+Z -> 3*Y+Z. 692 // 693 // TODO: We could handle "X + ~X" -> "-1" if we wanted, since "-X = ~X+1". 694 // 695 for (unsigned i = 0, e = Ops.size(); i != e; ++i) { 696 Value *TheOp = Ops[i].Op; 697 // Check to see if we've seen this operand before. If so, we factor all 698 // instances of the operand together. Due to our sorting criteria, we know 699 // that these need to be next to each other in the vector. 700 if (i+1 != Ops.size() && Ops[i+1].Op == TheOp) { 701 // Rescan the list, remove all instances of this operand from the expr. 702 unsigned NumFound = 0; 703 do { 704 Ops.erase(Ops.begin()+i); 705 ++NumFound; 706 } while (i != Ops.size() && Ops[i].Op == TheOp); 707 708 DEBUG(errs() << "\nFACTORING [" << NumFound << "]: " << *TheOp << '\n'); 709 ++NumFactor; 710 711 // Insert a new multiply. 712 Value *Mul = ConstantInt::get(cast<IntegerType>(I->getType()), NumFound); 713 Mul = BinaryOperator::CreateMul(TheOp, Mul, "factor", I); 714 715 // Now that we have inserted a multiply, optimize it. This allows us to 716 // handle cases that require multiple factoring steps, such as this: 717 // (X*2) + (X*2) + (X*2) -> (X*2)*3 -> X*6 718 Mul = ReassociateExpression(cast<BinaryOperator>(Mul)); 719 720 // If every add operand was a duplicate, return the multiply. 721 if (Ops.empty()) 722 return Mul; 723 724 // Otherwise, we had some input that didn't have the dupe, such as 725 // "A + A + B" -> "A*2 + B". Add the new multiply to the list of 726 // things being added by this operation. 727 Ops.insert(Ops.begin(), ValueEntry(getRank(Mul), Mul)); 728 729 --i; 730 e = Ops.size(); 731 continue; 732 } 733 734 // Check for X and -X in the operand list. 735 if (!BinaryOperator::isNeg(TheOp)) 736 continue; 737 738 Value *X = BinaryOperator::getNegArgument(TheOp); 739 unsigned FoundX = FindInOperandList(Ops, i, X); 740 if (FoundX == i) 741 continue; 742 743 // Remove X and -X from the operand list. 744 if (Ops.size() == 2) 745 return Constant::getNullValue(X->getType()); 746 747 Ops.erase(Ops.begin()+i); 748 if (i < FoundX) 749 --FoundX; 750 else 751 --i; // Need to back up an extra one. 752 Ops.erase(Ops.begin()+FoundX); 753 ++NumAnnihil; 754 --i; // Revisit element. 755 e -= 2; // Removed two elements. 756 } 757 758 // Scan the operand list, checking to see if there are any common factors 759 // between operands. Consider something like A*A+A*B*C+D. We would like to 760 // reassociate this to A*(A+B*C)+D, which reduces the number of multiplies. 761 // To efficiently find this, we count the number of times a factor occurs 762 // for any ADD operands that are MULs. 763 DenseMap<Value*, unsigned> FactorOccurrences; 764 765 // Keep track of each multiply we see, to avoid triggering on (X*4)+(X*4) 766 // where they are actually the same multiply. 767 unsigned MaxOcc = 0; 768 Value *MaxOccVal = 0; 769 for (unsigned i = 0, e = Ops.size(); i != e; ++i) { 770 BinaryOperator *BOp = dyn_cast<BinaryOperator>(Ops[i].Op); 771 if (BOp == 0 || BOp->getOpcode() != Instruction::Mul || !BOp->use_empty()) 772 continue; 773 774 // Compute all of the factors of this added value. 775 SmallVector<Value*, 8> Factors; 776 FindSingleUseMultiplyFactors(BOp, Factors, Ops, true); 777 assert(Factors.size() > 1 && "Bad linearize!"); 778 779 // Add one to FactorOccurrences for each unique factor in this op. 780 SmallPtrSet<Value*, 8> Duplicates; 781 for (unsigned i = 0, e = Factors.size(); i != e; ++i) { 782 Value *Factor = Factors[i]; 783 if (!Duplicates.insert(Factor)) continue; 784 785 unsigned Occ = ++FactorOccurrences[Factor]; 786 if (Occ > MaxOcc) { MaxOcc = Occ; MaxOccVal = Factor; } 787 788 // If Factor is a negative constant, add the negated value as a factor 789 // because we can percolate the negate out. Watch for minint, which 790 // cannot be positivified. 791 if (ConstantInt *CI = dyn_cast<ConstantInt>(Factor)) 792 if (CI->getValue().isNegative() && !CI->getValue().isMinSignedValue()) { 793 Factor = ConstantInt::get(CI->getContext(), -CI->getValue()); 794 assert(!Duplicates.count(Factor) && 795 "Shouldn't have two constant factors, missed a canonicalize"); 796 797 unsigned Occ = ++FactorOccurrences[Factor]; 798 if (Occ > MaxOcc) { MaxOcc = Occ; MaxOccVal = Factor; } 799 } 800 } 801 } 802 803 // If any factor occurred more than one time, we can pull it out. 804 if (MaxOcc > 1) { 805 DEBUG(errs() << "\nFACTORING [" << MaxOcc << "]: " << *MaxOccVal << '\n'); 806 ++NumFactor; 807 808 // Create a new instruction that uses the MaxOccVal twice. If we don't do 809 // this, we could otherwise run into situations where removing a factor 810 // from an expression will drop a use of maxocc, and this can cause 811 // RemoveFactorFromExpression on successive values to behave differently. 812 Instruction *DummyInst = BinaryOperator::CreateAdd(MaxOccVal, MaxOccVal); 813 SmallVector<Value*, 4> NewMulOps; 814 for (unsigned i = 0, e = Ops.size(); i != e; ++i) { 815 // Only try to remove factors from expressions we're allowed to. 816 BinaryOperator *BOp = dyn_cast<BinaryOperator>(Ops[i].Op); 817 if (BOp == 0 || BOp->getOpcode() != Instruction::Mul || !BOp->use_empty()) 818 continue; 819 820 if (Value *V = RemoveFactorFromExpression(Ops[i].Op, MaxOccVal)) { 821 NewMulOps.push_back(V); 822 Ops.erase(Ops.begin()+i); 823 --i; --e; 824 } 825 } 826 827 // No need for extra uses anymore. 828 delete DummyInst; 829 830 unsigned NumAddedValues = NewMulOps.size(); 831 Value *V = EmitAddTreeOfValues(I, NewMulOps); 832 833 // Now that we have inserted the add tree, optimize it. This allows us to 834 // handle cases that require multiple factoring steps, such as this: 835 // A*A*B + A*A*C --> A*(A*B+A*C) --> A*(A*(B+C)) 836 assert(NumAddedValues > 1 && "Each occurrence should contribute a value"); 837 (void)NumAddedValues; 838 V = ReassociateExpression(cast<BinaryOperator>(V)); 839 840 // Create the multiply. 841 Value *V2 = BinaryOperator::CreateMul(V, MaxOccVal, "tmp", I); 842 843 // Rerun associate on the multiply in case the inner expression turned into 844 // a multiply. We want to make sure that we keep things in canonical form. 845 V2 = ReassociateExpression(cast<BinaryOperator>(V2)); 846 847 // If every add operand included the factor (e.g. "A*B + A*C"), then the 848 // entire result expression is just the multiply "A*(B+C)". 849 if (Ops.empty()) 850 return V2; 851 852 // Otherwise, we had some input that didn't have the factor, such as 853 // "A*B + A*C + D" -> "A*(B+C) + D". Add the new multiply to the list of 854 // things being added by this operation. 855 Ops.insert(Ops.begin(), ValueEntry(getRank(V2), V2)); 856 } 857 858 return 0; 859} 860 861Value *Reassociate::OptimizeExpression(BinaryOperator *I, 862 SmallVectorImpl<ValueEntry> &Ops) { 863 // Now that we have the linearized expression tree, try to optimize it. 864 // Start by folding any constants that we found. 865 bool IterateOptimization = false; 866 if (Ops.size() == 1) return Ops[0].Op; 867 868 unsigned Opcode = I->getOpcode(); 869 870 if (Constant *V1 = dyn_cast<Constant>(Ops[Ops.size()-2].Op)) 871 if (Constant *V2 = dyn_cast<Constant>(Ops.back().Op)) { 872 Ops.pop_back(); 873 Ops.back().Op = ConstantExpr::get(Opcode, V1, V2); 874 return OptimizeExpression(I, Ops); 875 } 876 877 // Check for destructive annihilation due to a constant being used. 878 if (ConstantInt *CstVal = dyn_cast<ConstantInt>(Ops.back().Op)) 879 switch (Opcode) { 880 default: break; 881 case Instruction::And: 882 if (CstVal->isZero()) // X & 0 -> 0 883 return CstVal; 884 if (CstVal->isAllOnesValue()) // X & -1 -> X 885 Ops.pop_back(); 886 break; 887 case Instruction::Mul: 888 if (CstVal->isZero()) { // X * 0 -> 0 889 ++NumAnnihil; 890 return CstVal; 891 } 892 893 if (cast<ConstantInt>(CstVal)->isOne()) 894 Ops.pop_back(); // X * 1 -> X 895 break; 896 case Instruction::Or: 897 if (CstVal->isAllOnesValue()) // X | -1 -> -1 898 return CstVal; 899 // FALLTHROUGH! 900 case Instruction::Add: 901 case Instruction::Xor: 902 if (CstVal->isZero()) // X [|^+] 0 -> X 903 Ops.pop_back(); 904 break; 905 } 906 if (Ops.size() == 1) return Ops[0].Op; 907 908 // Handle destructive annihilation due to identities between elements in the 909 // argument list here. 910 switch (Opcode) { 911 default: break; 912 case Instruction::And: 913 case Instruction::Or: 914 case Instruction::Xor: { 915 unsigned NumOps = Ops.size(); 916 if (Value *Result = OptimizeAndOrXor(Opcode, Ops)) 917 return Result; 918 IterateOptimization |= Ops.size() != NumOps; 919 break; 920 } 921 922 case Instruction::Add: { 923 unsigned NumOps = Ops.size(); 924 if (Value *Result = OptimizeAdd(I, Ops)) 925 return Result; 926 IterateOptimization |= Ops.size() != NumOps; 927 } 928 929 break; 930 //case Instruction::Mul: 931 } 932 933 if (IterateOptimization) 934 return OptimizeExpression(I, Ops); 935 return 0; 936} 937 938 939/// ReassociateBB - Inspect all of the instructions in this basic block, 940/// reassociating them as we go. 941void Reassociate::ReassociateBB(BasicBlock *BB) { 942 for (BasicBlock::iterator BBI = BB->begin(); BBI != BB->end(); ) { 943 Instruction *BI = BBI++; 944 if (BI->getOpcode() == Instruction::Shl && 945 isa<ConstantInt>(BI->getOperand(1))) 946 if (Instruction *NI = ConvertShiftToMul(BI, ValueRankMap)) { 947 MadeChange = true; 948 BI = NI; 949 } 950 951 // Reject cases where it is pointless to do this. 952 if (!isa<BinaryOperator>(BI) || BI->getType()->isFloatingPointTy() || 953 BI->getType()->isVectorTy()) 954 continue; // Floating point ops are not associative. 955 956 // Do not reassociate boolean (i1) expressions. We want to preserve the 957 // original order of evaluation for short-circuited comparisons that 958 // SimplifyCFG has folded to AND/OR expressions. If the expression 959 // is not further optimized, it is likely to be transformed back to a 960 // short-circuited form for code gen, and the source order may have been 961 // optimized for the most likely conditions. 962 if (BI->getType()->isIntegerTy(1)) 963 continue; 964 965 // If this is a subtract instruction which is not already in negate form, 966 // see if we can convert it to X+-Y. 967 if (BI->getOpcode() == Instruction::Sub) { 968 if (ShouldBreakUpSubtract(BI)) { 969 BI = BreakUpSubtract(BI, ValueRankMap); 970 // Reset the BBI iterator in case BreakUpSubtract changed the 971 // instruction it points to. 972 BBI = BI; 973 ++BBI; 974 MadeChange = true; 975 } else if (BinaryOperator::isNeg(BI)) { 976 // Otherwise, this is a negation. See if the operand is a multiply tree 977 // and if this is not an inner node of a multiply tree. 978 if (isReassociableOp(BI->getOperand(1), Instruction::Mul) && 979 (!BI->hasOneUse() || 980 !isReassociableOp(BI->use_back(), Instruction::Mul))) { 981 BI = LowerNegateToMultiply(BI, ValueRankMap); 982 MadeChange = true; 983 } 984 } 985 } 986 987 // If this instruction is a commutative binary operator, process it. 988 if (!BI->isAssociative()) continue; 989 BinaryOperator *I = cast<BinaryOperator>(BI); 990 991 // If this is an interior node of a reassociable tree, ignore it until we 992 // get to the root of the tree, to avoid N^2 analysis. 993 if (I->hasOneUse() && isReassociableOp(I->use_back(), I->getOpcode())) 994 continue; 995 996 // If this is an add tree that is used by a sub instruction, ignore it 997 // until we process the subtract. 998 if (I->hasOneUse() && I->getOpcode() == Instruction::Add && 999 cast<Instruction>(I->use_back())->getOpcode() == Instruction::Sub) 1000 continue; 1001 1002 ReassociateExpression(I); 1003 } 1004} 1005 1006Value *Reassociate::ReassociateExpression(BinaryOperator *I) { 1007 1008 // First, walk the expression tree, linearizing the tree, collecting the 1009 // operand information. 1010 SmallVector<ValueEntry, 8> Ops; 1011 LinearizeExprTree(I, Ops); 1012 1013 DEBUG(dbgs() << "RAIn:\t"; PrintOps(I, Ops); dbgs() << '\n'); 1014 1015 // Now that we have linearized the tree to a list and have gathered all of 1016 // the operands and their ranks, sort the operands by their rank. Use a 1017 // stable_sort so that values with equal ranks will have their relative 1018 // positions maintained (and so the compiler is deterministic). Note that 1019 // this sorts so that the highest ranking values end up at the beginning of 1020 // the vector. 1021 std::stable_sort(Ops.begin(), Ops.end()); 1022 1023 // OptimizeExpression - Now that we have the expression tree in a convenient 1024 // sorted form, optimize it globally if possible. 1025 if (Value *V = OptimizeExpression(I, Ops)) { 1026 // This expression tree simplified to something that isn't a tree, 1027 // eliminate it. 1028 DEBUG(dbgs() << "Reassoc to scalar: " << *V << '\n'); 1029 I->replaceAllUsesWith(V); 1030 RemoveDeadBinaryOp(I); 1031 ++NumAnnihil; 1032 return V; 1033 } 1034 1035 // We want to sink immediates as deeply as possible except in the case where 1036 // this is a multiply tree used only by an add, and the immediate is a -1. 1037 // In this case we reassociate to put the negation on the outside so that we 1038 // can fold the negation into the add: (-X)*Y + Z -> Z-X*Y 1039 if (I->getOpcode() == Instruction::Mul && I->hasOneUse() && 1040 cast<Instruction>(I->use_back())->getOpcode() == Instruction::Add && 1041 isa<ConstantInt>(Ops.back().Op) && 1042 cast<ConstantInt>(Ops.back().Op)->isAllOnesValue()) { 1043 ValueEntry Tmp = Ops.pop_back_val(); 1044 Ops.insert(Ops.begin(), Tmp); 1045 } 1046 1047 DEBUG(dbgs() << "RAOut:\t"; PrintOps(I, Ops); dbgs() << '\n'); 1048 1049 if (Ops.size() == 1) { 1050 // This expression tree simplified to something that isn't a tree, 1051 // eliminate it. 1052 I->replaceAllUsesWith(Ops[0].Op); 1053 RemoveDeadBinaryOp(I); 1054 return Ops[0].Op; 1055 } 1056 1057 // Now that we ordered and optimized the expressions, splat them back into 1058 // the expression tree, removing any unneeded nodes. 1059 RewriteExprTree(I, Ops); 1060 return I; 1061} 1062 1063 1064bool Reassociate::runOnFunction(Function &F) { 1065 // Recalculate the rank map for F 1066 BuildRankMap(F); 1067 1068 MadeChange = false; 1069 for (Function::iterator FI = F.begin(), FE = F.end(); FI != FE; ++FI) 1070 ReassociateBB(FI); 1071 1072 // We are done with the rank map. 1073 RankMap.clear(); 1074 ValueRankMap.clear(); 1075 return MadeChange; 1076} 1077 1078