Reassociate.cpp revision 218893
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 // Conservatively clear all the optional flags, which may not hold 244 // after the reassociation. 245 I->clearSubclassOptionalData(); 246 LHS->clearSubclassOptionalData(); 247 RHS->clearSubclassOptionalData(); 248 249 ++NumLinear; 250 MadeChange = true; 251 DEBUG(dbgs() << "Linearized: " << *I << '\n'); 252 253 // If D is part of this expression tree, tail recurse. 254 if (isReassociableOp(I->getOperand(1), I->getOpcode())) 255 LinearizeExpr(I); 256} 257 258 259/// LinearizeExprTree - Given an associative binary expression tree, traverse 260/// all of the uses putting it into canonical form. This forces a left-linear 261/// form of the expression (((a+b)+c)+d), and collects information about the 262/// rank of the non-tree operands. 263/// 264/// NOTE: These intentionally destroys the expression tree operands (turning 265/// them into undef values) to reduce #uses of the values. This means that the 266/// caller MUST use something like RewriteExprTree to put the values back in. 267/// 268void Reassociate::LinearizeExprTree(BinaryOperator *I, 269 SmallVectorImpl<ValueEntry> &Ops) { 270 Value *LHS = I->getOperand(0), *RHS = I->getOperand(1); 271 unsigned Opcode = I->getOpcode(); 272 273 // First step, linearize the expression if it is in ((A+B)+(C+D)) form. 274 BinaryOperator *LHSBO = isReassociableOp(LHS, Opcode); 275 BinaryOperator *RHSBO = isReassociableOp(RHS, Opcode); 276 277 // If this is a multiply expression tree and it contains internal negations, 278 // transform them into multiplies by -1 so they can be reassociated. 279 if (I->getOpcode() == Instruction::Mul) { 280 if (!LHSBO && LHS->hasOneUse() && BinaryOperator::isNeg(LHS)) { 281 LHS = LowerNegateToMultiply(cast<Instruction>(LHS), ValueRankMap); 282 LHSBO = isReassociableOp(LHS, Opcode); 283 } 284 if (!RHSBO && RHS->hasOneUse() && BinaryOperator::isNeg(RHS)) { 285 RHS = LowerNegateToMultiply(cast<Instruction>(RHS), ValueRankMap); 286 RHSBO = isReassociableOp(RHS, Opcode); 287 } 288 } 289 290 if (!LHSBO) { 291 if (!RHSBO) { 292 // Neither the LHS or RHS as part of the tree, thus this is a leaf. As 293 // such, just remember these operands and their rank. 294 Ops.push_back(ValueEntry(getRank(LHS), LHS)); 295 Ops.push_back(ValueEntry(getRank(RHS), RHS)); 296 297 // Clear the leaves out. 298 I->setOperand(0, UndefValue::get(I->getType())); 299 I->setOperand(1, UndefValue::get(I->getType())); 300 return; 301 } 302 303 // Turn X+(Y+Z) -> (Y+Z)+X 304 std::swap(LHSBO, RHSBO); 305 std::swap(LHS, RHS); 306 bool Success = !I->swapOperands(); 307 assert(Success && "swapOperands failed"); 308 Success = false; 309 MadeChange = true; 310 } else if (RHSBO) { 311 // Turn (A+B)+(C+D) -> (((A+B)+C)+D). This guarantees the RHS is not 312 // part of the expression tree. 313 LinearizeExpr(I); 314 LHS = LHSBO = cast<BinaryOperator>(I->getOperand(0)); 315 RHS = I->getOperand(1); 316 RHSBO = 0; 317 } 318 319 // Okay, now we know that the LHS is a nested expression and that the RHS is 320 // not. Perform reassociation. 321 assert(!isReassociableOp(RHS, Opcode) && "LinearizeExpr failed!"); 322 323 // Move LHS right before I to make sure that the tree expression dominates all 324 // values. 325 LHSBO->moveBefore(I); 326 327 // Linearize the expression tree on the LHS. 328 LinearizeExprTree(LHSBO, Ops); 329 330 // Remember the RHS operand and its rank. 331 Ops.push_back(ValueEntry(getRank(RHS), RHS)); 332 333 // Clear the RHS leaf out. 334 I->setOperand(1, UndefValue::get(I->getType())); 335} 336 337// RewriteExprTree - Now that the operands for this expression tree are 338// linearized and optimized, emit them in-order. This function is written to be 339// tail recursive. 340void Reassociate::RewriteExprTree(BinaryOperator *I, 341 SmallVectorImpl<ValueEntry> &Ops, 342 unsigned i) { 343 if (i+2 == Ops.size()) { 344 if (I->getOperand(0) != Ops[i].Op || 345 I->getOperand(1) != Ops[i+1].Op) { 346 Value *OldLHS = I->getOperand(0); 347 DEBUG(dbgs() << "RA: " << *I << '\n'); 348 I->setOperand(0, Ops[i].Op); 349 I->setOperand(1, Ops[i+1].Op); 350 351 // Clear all the optional flags, which may not hold after the 352 // reassociation if the expression involved more than just this operation. 353 if (Ops.size() != 2) 354 I->clearSubclassOptionalData(); 355 356 DEBUG(dbgs() << "TO: " << *I << '\n'); 357 MadeChange = true; 358 ++NumChanged; 359 360 // If we reassociated a tree to fewer operands (e.g. (1+a+2) -> (a+3) 361 // delete the extra, now dead, nodes. 362 RemoveDeadBinaryOp(OldLHS); 363 } 364 return; 365 } 366 assert(i+2 < Ops.size() && "Ops index out of range!"); 367 368 if (I->getOperand(1) != Ops[i].Op) { 369 DEBUG(dbgs() << "RA: " << *I << '\n'); 370 I->setOperand(1, Ops[i].Op); 371 372 // Conservatively clear all the optional flags, which may not hold 373 // after the reassociation. 374 I->clearSubclassOptionalData(); 375 376 DEBUG(dbgs() << "TO: " << *I << '\n'); 377 MadeChange = true; 378 ++NumChanged; 379 } 380 381 BinaryOperator *LHS = cast<BinaryOperator>(I->getOperand(0)); 382 assert(LHS->getOpcode() == I->getOpcode() && 383 "Improper expression tree!"); 384 385 // Compactify the tree instructions together with each other to guarantee 386 // that the expression tree is dominated by all of Ops. 387 LHS->moveBefore(I); 388 RewriteExprTree(LHS, Ops, i+1); 389} 390 391 392 393// NegateValue - Insert instructions before the instruction pointed to by BI, 394// that computes the negative version of the value specified. The negative 395// version of the value is returned, and BI is left pointing at the instruction 396// that should be processed next by the reassociation pass. 397// 398static Value *NegateValue(Value *V, Instruction *BI) { 399 if (Constant *C = dyn_cast<Constant>(V)) 400 return ConstantExpr::getNeg(C); 401 402 // We are trying to expose opportunity for reassociation. One of the things 403 // that we want to do to achieve this is to push a negation as deep into an 404 // expression chain as possible, to expose the add instructions. In practice, 405 // this means that we turn this: 406 // X = -(A+12+C+D) into X = -A + -12 + -C + -D = -12 + -A + -C + -D 407 // so that later, a: Y = 12+X could get reassociated with the -12 to eliminate 408 // the constants. We assume that instcombine will clean up the mess later if 409 // we introduce tons of unnecessary negation instructions. 410 // 411 if (Instruction *I = dyn_cast<Instruction>(V)) 412 if (I->getOpcode() == Instruction::Add && I->hasOneUse()) { 413 // Push the negates through the add. 414 I->setOperand(0, NegateValue(I->getOperand(0), BI)); 415 I->setOperand(1, NegateValue(I->getOperand(1), BI)); 416 417 // We must move the add instruction here, because the neg instructions do 418 // not dominate the old add instruction in general. By moving it, we are 419 // assured that the neg instructions we just inserted dominate the 420 // instruction we are about to insert after them. 421 // 422 I->moveBefore(BI); 423 I->setName(I->getName()+".neg"); 424 return I; 425 } 426 427 // Okay, we need to materialize a negated version of V with an instruction. 428 // Scan the use lists of V to see if we have one already. 429 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E;++UI){ 430 User *U = *UI; 431 if (!BinaryOperator::isNeg(U)) continue; 432 433 // We found one! Now we have to make sure that the definition dominates 434 // this use. We do this by moving it to the entry block (if it is a 435 // non-instruction value) or right after the definition. These negates will 436 // be zapped by reassociate later, so we don't need much finesse here. 437 BinaryOperator *TheNeg = cast<BinaryOperator>(U); 438 439 // Verify that the negate is in this function, V might be a constant expr. 440 if (TheNeg->getParent()->getParent() != BI->getParent()->getParent()) 441 continue; 442 443 BasicBlock::iterator InsertPt; 444 if (Instruction *InstInput = dyn_cast<Instruction>(V)) { 445 if (InvokeInst *II = dyn_cast<InvokeInst>(InstInput)) { 446 InsertPt = II->getNormalDest()->begin(); 447 } else { 448 InsertPt = InstInput; 449 ++InsertPt; 450 } 451 while (isa<PHINode>(InsertPt)) ++InsertPt; 452 } else { 453 InsertPt = TheNeg->getParent()->getParent()->getEntryBlock().begin(); 454 } 455 TheNeg->moveBefore(InsertPt); 456 return TheNeg; 457 } 458 459 // Insert a 'neg' instruction that subtracts the value from zero to get the 460 // negation. 461 return BinaryOperator::CreateNeg(V, V->getName() + ".neg", BI); 462} 463 464/// ShouldBreakUpSubtract - Return true if we should break up this subtract of 465/// X-Y into (X + -Y). 466static bool ShouldBreakUpSubtract(Instruction *Sub) { 467 // If this is a negation, we can't split it up! 468 if (BinaryOperator::isNeg(Sub)) 469 return false; 470 471 // Don't bother to break this up unless either the LHS is an associable add or 472 // subtract or if this is only used by one. 473 if (isReassociableOp(Sub->getOperand(0), Instruction::Add) || 474 isReassociableOp(Sub->getOperand(0), Instruction::Sub)) 475 return true; 476 if (isReassociableOp(Sub->getOperand(1), Instruction::Add) || 477 isReassociableOp(Sub->getOperand(1), Instruction::Sub)) 478 return true; 479 if (Sub->hasOneUse() && 480 (isReassociableOp(Sub->use_back(), Instruction::Add) || 481 isReassociableOp(Sub->use_back(), Instruction::Sub))) 482 return true; 483 484 return false; 485} 486 487/// BreakUpSubtract - If we have (X-Y), and if either X is an add, or if this is 488/// only used by an add, transform this into (X+(0-Y)) to promote better 489/// reassociation. 490static Instruction *BreakUpSubtract(Instruction *Sub, 491 DenseMap<AssertingVH<>, unsigned> &ValueRankMap) { 492 // Convert a subtract into an add and a neg instruction. This allows sub 493 // instructions to be commuted with other add instructions. 494 // 495 // Calculate the negative value of Operand 1 of the sub instruction, 496 // and set it as the RHS of the add instruction we just made. 497 // 498 Value *NegVal = NegateValue(Sub->getOperand(1), Sub); 499 Instruction *New = 500 BinaryOperator::CreateAdd(Sub->getOperand(0), NegVal, "", Sub); 501 New->takeName(Sub); 502 503 // Everyone now refers to the add instruction. 504 ValueRankMap.erase(Sub); 505 Sub->replaceAllUsesWith(New); 506 Sub->eraseFromParent(); 507 508 DEBUG(dbgs() << "Negated: " << *New << '\n'); 509 return New; 510} 511 512/// ConvertShiftToMul - If this is a shift of a reassociable multiply or is used 513/// by one, change this into a multiply by a constant to assist with further 514/// reassociation. 515static Instruction *ConvertShiftToMul(Instruction *Shl, 516 DenseMap<AssertingVH<>, unsigned> &ValueRankMap) { 517 // If an operand of this shift is a reassociable multiply, or if the shift 518 // is used by a reassociable multiply or add, turn into a multiply. 519 if (isReassociableOp(Shl->getOperand(0), Instruction::Mul) || 520 (Shl->hasOneUse() && 521 (isReassociableOp(Shl->use_back(), Instruction::Mul) || 522 isReassociableOp(Shl->use_back(), Instruction::Add)))) { 523 Constant *MulCst = ConstantInt::get(Shl->getType(), 1); 524 MulCst = ConstantExpr::getShl(MulCst, cast<Constant>(Shl->getOperand(1))); 525 526 Instruction *Mul = 527 BinaryOperator::CreateMul(Shl->getOperand(0), MulCst, "", Shl); 528 ValueRankMap.erase(Shl); 529 Mul->takeName(Shl); 530 Shl->replaceAllUsesWith(Mul); 531 Shl->eraseFromParent(); 532 return Mul; 533 } 534 return 0; 535} 536 537// Scan backwards and forwards among values with the same rank as element i to 538// see if X exists. If X does not exist, return i. This is useful when 539// scanning for 'x' when we see '-x' because they both get the same rank. 540static unsigned FindInOperandList(SmallVectorImpl<ValueEntry> &Ops, unsigned i, 541 Value *X) { 542 unsigned XRank = Ops[i].Rank; 543 unsigned e = Ops.size(); 544 for (unsigned j = i+1; j != e && Ops[j].Rank == XRank; ++j) 545 if (Ops[j].Op == X) 546 return j; 547 // Scan backwards. 548 for (unsigned j = i-1; j != ~0U && Ops[j].Rank == XRank; --j) 549 if (Ops[j].Op == X) 550 return j; 551 return i; 552} 553 554/// EmitAddTreeOfValues - Emit a tree of add instructions, summing Ops together 555/// and returning the result. Insert the tree before I. 556static Value *EmitAddTreeOfValues(Instruction *I, SmallVectorImpl<Value*> &Ops){ 557 if (Ops.size() == 1) return Ops.back(); 558 559 Value *V1 = Ops.back(); 560 Ops.pop_back(); 561 Value *V2 = EmitAddTreeOfValues(I, Ops); 562 return BinaryOperator::CreateAdd(V2, V1, "tmp", I); 563} 564 565/// RemoveFactorFromExpression - If V is an expression tree that is a 566/// multiplication sequence, and if this sequence contains a multiply by Factor, 567/// remove Factor from the tree and return the new tree. 568Value *Reassociate::RemoveFactorFromExpression(Value *V, Value *Factor) { 569 BinaryOperator *BO = isReassociableOp(V, Instruction::Mul); 570 if (!BO) return 0; 571 572 SmallVector<ValueEntry, 8> Factors; 573 LinearizeExprTree(BO, Factors); 574 575 bool FoundFactor = false; 576 bool NeedsNegate = false; 577 for (unsigned i = 0, e = Factors.size(); i != e; ++i) { 578 if (Factors[i].Op == Factor) { 579 FoundFactor = true; 580 Factors.erase(Factors.begin()+i); 581 break; 582 } 583 584 // If this is a negative version of this factor, remove it. 585 if (ConstantInt *FC1 = dyn_cast<ConstantInt>(Factor)) 586 if (ConstantInt *FC2 = dyn_cast<ConstantInt>(Factors[i].Op)) 587 if (FC1->getValue() == -FC2->getValue()) { 588 FoundFactor = NeedsNegate = true; 589 Factors.erase(Factors.begin()+i); 590 break; 591 } 592 } 593 594 if (!FoundFactor) { 595 // Make sure to restore the operands to the expression tree. 596 RewriteExprTree(BO, Factors); 597 return 0; 598 } 599 600 BasicBlock::iterator InsertPt = BO; ++InsertPt; 601 602 // If this was just a single multiply, remove the multiply and return the only 603 // remaining operand. 604 if (Factors.size() == 1) { 605 ValueRankMap.erase(BO); 606 BO->eraseFromParent(); 607 V = Factors[0].Op; 608 } else { 609 RewriteExprTree(BO, Factors); 610 V = BO; 611 } 612 613 if (NeedsNegate) 614 V = BinaryOperator::CreateNeg(V, "neg", InsertPt); 615 616 return V; 617} 618 619/// FindSingleUseMultiplyFactors - If V is a single-use multiply, recursively 620/// add its operands as factors, otherwise add V to the list of factors. 621/// 622/// Ops is the top-level list of add operands we're trying to factor. 623static void FindSingleUseMultiplyFactors(Value *V, 624 SmallVectorImpl<Value*> &Factors, 625 const SmallVectorImpl<ValueEntry> &Ops, 626 bool IsRoot) { 627 BinaryOperator *BO; 628 if (!(V->hasOneUse() || V->use_empty()) || // More than one use. 629 !(BO = dyn_cast<BinaryOperator>(V)) || 630 BO->getOpcode() != Instruction::Mul) { 631 Factors.push_back(V); 632 return; 633 } 634 635 // If this value has a single use because it is another input to the add 636 // tree we're reassociating and we dropped its use, it actually has two 637 // uses and we can't factor it. 638 if (!IsRoot) { 639 for (unsigned i = 0, e = Ops.size(); i != e; ++i) 640 if (Ops[i].Op == V) { 641 Factors.push_back(V); 642 return; 643 } 644 } 645 646 647 // Otherwise, add the LHS and RHS to the list of factors. 648 FindSingleUseMultiplyFactors(BO->getOperand(1), Factors, Ops, false); 649 FindSingleUseMultiplyFactors(BO->getOperand(0), Factors, Ops, false); 650} 651 652/// OptimizeAndOrXor - Optimize a series of operands to an 'and', 'or', or 'xor' 653/// instruction. This optimizes based on identities. If it can be reduced to 654/// a single Value, it is returned, otherwise the Ops list is mutated as 655/// necessary. 656static Value *OptimizeAndOrXor(unsigned Opcode, 657 SmallVectorImpl<ValueEntry> &Ops) { 658 // Scan the operand lists looking for X and ~X pairs, along with X,X pairs. 659 // If we find any, we can simplify the expression. X&~X == 0, X|~X == -1. 660 for (unsigned i = 0, e = Ops.size(); i != e; ++i) { 661 // First, check for X and ~X in the operand list. 662 assert(i < Ops.size()); 663 if (BinaryOperator::isNot(Ops[i].Op)) { // Cannot occur for ^. 664 Value *X = BinaryOperator::getNotArgument(Ops[i].Op); 665 unsigned FoundX = FindInOperandList(Ops, i, X); 666 if (FoundX != i) { 667 if (Opcode == Instruction::And) // ...&X&~X = 0 668 return Constant::getNullValue(X->getType()); 669 670 if (Opcode == Instruction::Or) // ...|X|~X = -1 671 return Constant::getAllOnesValue(X->getType()); 672 } 673 } 674 675 // Next, check for duplicate pairs of values, which we assume are next to 676 // each other, due to our sorting criteria. 677 assert(i < Ops.size()); 678 if (i+1 != Ops.size() && Ops[i+1].Op == Ops[i].Op) { 679 if (Opcode == Instruction::And || Opcode == Instruction::Or) { 680 // Drop duplicate values for And and Or. 681 Ops.erase(Ops.begin()+i); 682 --i; --e; 683 ++NumAnnihil; 684 continue; 685 } 686 687 // Drop pairs of values for Xor. 688 assert(Opcode == Instruction::Xor); 689 if (e == 2) 690 return Constant::getNullValue(Ops[0].Op->getType()); 691 692 // Y ^ X^X -> Y 693 Ops.erase(Ops.begin()+i, Ops.begin()+i+2); 694 i -= 1; e -= 2; 695 ++NumAnnihil; 696 } 697 } 698 return 0; 699} 700 701/// OptimizeAdd - Optimize a series of operands to an 'add' instruction. This 702/// optimizes based on identities. If it can be reduced to a single Value, it 703/// is returned, otherwise the Ops list is mutated as necessary. 704Value *Reassociate::OptimizeAdd(Instruction *I, 705 SmallVectorImpl<ValueEntry> &Ops) { 706 // Scan the operand lists looking for X and -X pairs. If we find any, we 707 // can simplify the expression. X+-X == 0. While we're at it, scan for any 708 // duplicates. We want to canonicalize Y+Y+Y+Z -> 3*Y+Z. 709 // 710 // TODO: We could handle "X + ~X" -> "-1" if we wanted, since "-X = ~X+1". 711 // 712 for (unsigned i = 0, e = Ops.size(); i != e; ++i) { 713 Value *TheOp = Ops[i].Op; 714 // Check to see if we've seen this operand before. If so, we factor all 715 // instances of the operand together. Due to our sorting criteria, we know 716 // that these need to be next to each other in the vector. 717 if (i+1 != Ops.size() && Ops[i+1].Op == TheOp) { 718 // Rescan the list, remove all instances of this operand from the expr. 719 unsigned NumFound = 0; 720 do { 721 Ops.erase(Ops.begin()+i); 722 ++NumFound; 723 } while (i != Ops.size() && Ops[i].Op == TheOp); 724 725 DEBUG(errs() << "\nFACTORING [" << NumFound << "]: " << *TheOp << '\n'); 726 ++NumFactor; 727 728 // Insert a new multiply. 729 Value *Mul = ConstantInt::get(cast<IntegerType>(I->getType()), NumFound); 730 Mul = BinaryOperator::CreateMul(TheOp, Mul, "factor", I); 731 732 // Now that we have inserted a multiply, optimize it. This allows us to 733 // handle cases that require multiple factoring steps, such as this: 734 // (X*2) + (X*2) + (X*2) -> (X*2)*3 -> X*6 735 Mul = ReassociateExpression(cast<BinaryOperator>(Mul)); 736 737 // If every add operand was a duplicate, return the multiply. 738 if (Ops.empty()) 739 return Mul; 740 741 // Otherwise, we had some input that didn't have the dupe, such as 742 // "A + A + B" -> "A*2 + B". Add the new multiply to the list of 743 // things being added by this operation. 744 Ops.insert(Ops.begin(), ValueEntry(getRank(Mul), Mul)); 745 746 --i; 747 e = Ops.size(); 748 continue; 749 } 750 751 // Check for X and -X in the operand list. 752 if (!BinaryOperator::isNeg(TheOp)) 753 continue; 754 755 Value *X = BinaryOperator::getNegArgument(TheOp); 756 unsigned FoundX = FindInOperandList(Ops, i, X); 757 if (FoundX == i) 758 continue; 759 760 // Remove X and -X from the operand list. 761 if (Ops.size() == 2) 762 return Constant::getNullValue(X->getType()); 763 764 Ops.erase(Ops.begin()+i); 765 if (i < FoundX) 766 --FoundX; 767 else 768 --i; // Need to back up an extra one. 769 Ops.erase(Ops.begin()+FoundX); 770 ++NumAnnihil; 771 --i; // Revisit element. 772 e -= 2; // Removed two elements. 773 } 774 775 // Scan the operand list, checking to see if there are any common factors 776 // between operands. Consider something like A*A+A*B*C+D. We would like to 777 // reassociate this to A*(A+B*C)+D, which reduces the number of multiplies. 778 // To efficiently find this, we count the number of times a factor occurs 779 // for any ADD operands that are MULs. 780 DenseMap<Value*, unsigned> FactorOccurrences; 781 782 // Keep track of each multiply we see, to avoid triggering on (X*4)+(X*4) 783 // where they are actually the same multiply. 784 unsigned MaxOcc = 0; 785 Value *MaxOccVal = 0; 786 for (unsigned i = 0, e = Ops.size(); i != e; ++i) { 787 BinaryOperator *BOp = dyn_cast<BinaryOperator>(Ops[i].Op); 788 if (BOp == 0 || BOp->getOpcode() != Instruction::Mul || !BOp->use_empty()) 789 continue; 790 791 // Compute all of the factors of this added value. 792 SmallVector<Value*, 8> Factors; 793 FindSingleUseMultiplyFactors(BOp, Factors, Ops, true); 794 assert(Factors.size() > 1 && "Bad linearize!"); 795 796 // Add one to FactorOccurrences for each unique factor in this op. 797 SmallPtrSet<Value*, 8> Duplicates; 798 for (unsigned i = 0, e = Factors.size(); i != e; ++i) { 799 Value *Factor = Factors[i]; 800 if (!Duplicates.insert(Factor)) continue; 801 802 unsigned Occ = ++FactorOccurrences[Factor]; 803 if (Occ > MaxOcc) { MaxOcc = Occ; MaxOccVal = Factor; } 804 805 // If Factor is a negative constant, add the negated value as a factor 806 // because we can percolate the negate out. Watch for minint, which 807 // cannot be positivified. 808 if (ConstantInt *CI = dyn_cast<ConstantInt>(Factor)) 809 if (CI->getValue().isNegative() && !CI->getValue().isMinSignedValue()) { 810 Factor = ConstantInt::get(CI->getContext(), -CI->getValue()); 811 assert(!Duplicates.count(Factor) && 812 "Shouldn't have two constant factors, missed a canonicalize"); 813 814 unsigned Occ = ++FactorOccurrences[Factor]; 815 if (Occ > MaxOcc) { MaxOcc = Occ; MaxOccVal = Factor; } 816 } 817 } 818 } 819 820 // If any factor occurred more than one time, we can pull it out. 821 if (MaxOcc > 1) { 822 DEBUG(errs() << "\nFACTORING [" << MaxOcc << "]: " << *MaxOccVal << '\n'); 823 ++NumFactor; 824 825 // Create a new instruction that uses the MaxOccVal twice. If we don't do 826 // this, we could otherwise run into situations where removing a factor 827 // from an expression will drop a use of maxocc, and this can cause 828 // RemoveFactorFromExpression on successive values to behave differently. 829 Instruction *DummyInst = BinaryOperator::CreateAdd(MaxOccVal, MaxOccVal); 830 SmallVector<Value*, 4> NewMulOps; 831 for (unsigned i = 0; i != Ops.size(); ++i) { 832 // Only try to remove factors from expressions we're allowed to. 833 BinaryOperator *BOp = dyn_cast<BinaryOperator>(Ops[i].Op); 834 if (BOp == 0 || BOp->getOpcode() != Instruction::Mul || !BOp->use_empty()) 835 continue; 836 837 if (Value *V = RemoveFactorFromExpression(Ops[i].Op, MaxOccVal)) { 838 // The factorized operand may occur several times. Convert them all in 839 // one fell swoop. 840 for (unsigned j = Ops.size(); j != i;) { 841 --j; 842 if (Ops[j].Op == Ops[i].Op) { 843 NewMulOps.push_back(V); 844 Ops.erase(Ops.begin()+j); 845 } 846 } 847 --i; 848 } 849 } 850 851 // No need for extra uses anymore. 852 delete DummyInst; 853 854 unsigned NumAddedValues = NewMulOps.size(); 855 Value *V = EmitAddTreeOfValues(I, NewMulOps); 856 857 // Now that we have inserted the add tree, optimize it. This allows us to 858 // handle cases that require multiple factoring steps, such as this: 859 // A*A*B + A*A*C --> A*(A*B+A*C) --> A*(A*(B+C)) 860 assert(NumAddedValues > 1 && "Each occurrence should contribute a value"); 861 (void)NumAddedValues; 862 V = ReassociateExpression(cast<BinaryOperator>(V)); 863 864 // Create the multiply. 865 Value *V2 = BinaryOperator::CreateMul(V, MaxOccVal, "tmp", I); 866 867 // Rerun associate on the multiply in case the inner expression turned into 868 // a multiply. We want to make sure that we keep things in canonical form. 869 V2 = ReassociateExpression(cast<BinaryOperator>(V2)); 870 871 // If every add operand included the factor (e.g. "A*B + A*C"), then the 872 // entire result expression is just the multiply "A*(B+C)". 873 if (Ops.empty()) 874 return V2; 875 876 // Otherwise, we had some input that didn't have the factor, such as 877 // "A*B + A*C + D" -> "A*(B+C) + D". Add the new multiply to the list of 878 // things being added by this operation. 879 Ops.insert(Ops.begin(), ValueEntry(getRank(V2), V2)); 880 } 881 882 return 0; 883} 884 885Value *Reassociate::OptimizeExpression(BinaryOperator *I, 886 SmallVectorImpl<ValueEntry> &Ops) { 887 // Now that we have the linearized expression tree, try to optimize it. 888 // Start by folding any constants that we found. 889 bool IterateOptimization = false; 890 if (Ops.size() == 1) return Ops[0].Op; 891 892 unsigned Opcode = I->getOpcode(); 893 894 if (Constant *V1 = dyn_cast<Constant>(Ops[Ops.size()-2].Op)) 895 if (Constant *V2 = dyn_cast<Constant>(Ops.back().Op)) { 896 Ops.pop_back(); 897 Ops.back().Op = ConstantExpr::get(Opcode, V1, V2); 898 return OptimizeExpression(I, Ops); 899 } 900 901 // Check for destructive annihilation due to a constant being used. 902 if (ConstantInt *CstVal = dyn_cast<ConstantInt>(Ops.back().Op)) 903 switch (Opcode) { 904 default: break; 905 case Instruction::And: 906 if (CstVal->isZero()) // X & 0 -> 0 907 return CstVal; 908 if (CstVal->isAllOnesValue()) // X & -1 -> X 909 Ops.pop_back(); 910 break; 911 case Instruction::Mul: 912 if (CstVal->isZero()) { // X * 0 -> 0 913 ++NumAnnihil; 914 return CstVal; 915 } 916 917 if (cast<ConstantInt>(CstVal)->isOne()) 918 Ops.pop_back(); // X * 1 -> X 919 break; 920 case Instruction::Or: 921 if (CstVal->isAllOnesValue()) // X | -1 -> -1 922 return CstVal; 923 // FALLTHROUGH! 924 case Instruction::Add: 925 case Instruction::Xor: 926 if (CstVal->isZero()) // X [|^+] 0 -> X 927 Ops.pop_back(); 928 break; 929 } 930 if (Ops.size() == 1) return Ops[0].Op; 931 932 // Handle destructive annihilation due to identities between elements in the 933 // argument list here. 934 switch (Opcode) { 935 default: break; 936 case Instruction::And: 937 case Instruction::Or: 938 case Instruction::Xor: { 939 unsigned NumOps = Ops.size(); 940 if (Value *Result = OptimizeAndOrXor(Opcode, Ops)) 941 return Result; 942 IterateOptimization |= Ops.size() != NumOps; 943 break; 944 } 945 946 case Instruction::Add: { 947 unsigned NumOps = Ops.size(); 948 if (Value *Result = OptimizeAdd(I, Ops)) 949 return Result; 950 IterateOptimization |= Ops.size() != NumOps; 951 } 952 953 break; 954 //case Instruction::Mul: 955 } 956 957 if (IterateOptimization) 958 return OptimizeExpression(I, Ops); 959 return 0; 960} 961 962 963/// ReassociateBB - Inspect all of the instructions in this basic block, 964/// reassociating them as we go. 965void Reassociate::ReassociateBB(BasicBlock *BB) { 966 for (BasicBlock::iterator BBI = BB->begin(); BBI != BB->end(); ) { 967 Instruction *BI = BBI++; 968 if (BI->getOpcode() == Instruction::Shl && 969 isa<ConstantInt>(BI->getOperand(1))) 970 if (Instruction *NI = ConvertShiftToMul(BI, ValueRankMap)) { 971 MadeChange = true; 972 BI = NI; 973 } 974 975 // Reject cases where it is pointless to do this. 976 if (!isa<BinaryOperator>(BI) || BI->getType()->isFloatingPointTy() || 977 BI->getType()->isVectorTy()) 978 continue; // Floating point ops are not associative. 979 980 // Do not reassociate boolean (i1) expressions. We want to preserve the 981 // original order of evaluation for short-circuited comparisons that 982 // SimplifyCFG has folded to AND/OR expressions. If the expression 983 // is not further optimized, it is likely to be transformed back to a 984 // short-circuited form for code gen, and the source order may have been 985 // optimized for the most likely conditions. 986 if (BI->getType()->isIntegerTy(1)) 987 continue; 988 989 // If this is a subtract instruction which is not already in negate form, 990 // see if we can convert it to X+-Y. 991 if (BI->getOpcode() == Instruction::Sub) { 992 if (ShouldBreakUpSubtract(BI)) { 993 BI = BreakUpSubtract(BI, ValueRankMap); 994 // Reset the BBI iterator in case BreakUpSubtract changed the 995 // instruction it points to. 996 BBI = BI; 997 ++BBI; 998 MadeChange = true; 999 } else if (BinaryOperator::isNeg(BI)) { 1000 // Otherwise, this is a negation. See if the operand is a multiply tree 1001 // and if this is not an inner node of a multiply tree. 1002 if (isReassociableOp(BI->getOperand(1), Instruction::Mul) && 1003 (!BI->hasOneUse() || 1004 !isReassociableOp(BI->use_back(), Instruction::Mul))) { 1005 BI = LowerNegateToMultiply(BI, ValueRankMap); 1006 MadeChange = true; 1007 } 1008 } 1009 } 1010 1011 // If this instruction is a commutative binary operator, process it. 1012 if (!BI->isAssociative()) continue; 1013 BinaryOperator *I = cast<BinaryOperator>(BI); 1014 1015 // If this is an interior node of a reassociable tree, ignore it until we 1016 // get to the root of the tree, to avoid N^2 analysis. 1017 if (I->hasOneUse() && isReassociableOp(I->use_back(), I->getOpcode())) 1018 continue; 1019 1020 // If this is an add tree that is used by a sub instruction, ignore it 1021 // until we process the subtract. 1022 if (I->hasOneUse() && I->getOpcode() == Instruction::Add && 1023 cast<Instruction>(I->use_back())->getOpcode() == Instruction::Sub) 1024 continue; 1025 1026 ReassociateExpression(I); 1027 } 1028} 1029 1030Value *Reassociate::ReassociateExpression(BinaryOperator *I) { 1031 1032 // First, walk the expression tree, linearizing the tree, collecting the 1033 // operand information. 1034 SmallVector<ValueEntry, 8> Ops; 1035 LinearizeExprTree(I, Ops); 1036 1037 DEBUG(dbgs() << "RAIn:\t"; PrintOps(I, Ops); dbgs() << '\n'); 1038 1039 // Now that we have linearized the tree to a list and have gathered all of 1040 // the operands and their ranks, sort the operands by their rank. Use a 1041 // stable_sort so that values with equal ranks will have their relative 1042 // positions maintained (and so the compiler is deterministic). Note that 1043 // this sorts so that the highest ranking values end up at the beginning of 1044 // the vector. 1045 std::stable_sort(Ops.begin(), Ops.end()); 1046 1047 // OptimizeExpression - Now that we have the expression tree in a convenient 1048 // sorted form, optimize it globally if possible. 1049 if (Value *V = OptimizeExpression(I, Ops)) { 1050 // This expression tree simplified to something that isn't a tree, 1051 // eliminate it. 1052 DEBUG(dbgs() << "Reassoc to scalar: " << *V << '\n'); 1053 I->replaceAllUsesWith(V); 1054 RemoveDeadBinaryOp(I); 1055 ++NumAnnihil; 1056 return V; 1057 } 1058 1059 // We want to sink immediates as deeply as possible except in the case where 1060 // this is a multiply tree used only by an add, and the immediate is a -1. 1061 // In this case we reassociate to put the negation on the outside so that we 1062 // can fold the negation into the add: (-X)*Y + Z -> Z-X*Y 1063 if (I->getOpcode() == Instruction::Mul && I->hasOneUse() && 1064 cast<Instruction>(I->use_back())->getOpcode() == Instruction::Add && 1065 isa<ConstantInt>(Ops.back().Op) && 1066 cast<ConstantInt>(Ops.back().Op)->isAllOnesValue()) { 1067 ValueEntry Tmp = Ops.pop_back_val(); 1068 Ops.insert(Ops.begin(), Tmp); 1069 } 1070 1071 DEBUG(dbgs() << "RAOut:\t"; PrintOps(I, Ops); dbgs() << '\n'); 1072 1073 if (Ops.size() == 1) { 1074 // This expression tree simplified to something that isn't a tree, 1075 // eliminate it. 1076 I->replaceAllUsesWith(Ops[0].Op); 1077 RemoveDeadBinaryOp(I); 1078 return Ops[0].Op; 1079 } 1080 1081 // Now that we ordered and optimized the expressions, splat them back into 1082 // the expression tree, removing any unneeded nodes. 1083 RewriteExprTree(I, Ops); 1084 return I; 1085} 1086 1087 1088bool Reassociate::runOnFunction(Function &F) { 1089 // Recalculate the rank map for F 1090 BuildRankMap(F); 1091 1092 MadeChange = false; 1093 for (Function::iterator FI = F.begin(), FE = F.end(); FI != FE; ++FI) 1094 ReassociateBB(FI); 1095 1096 // We are done with the rank map. 1097 RankMap.clear(); 1098 ValueRankMap.clear(); 1099 return MadeChange; 1100} 1101 1102