Reassociate.cpp revision 212904
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 82 bool runOnFunction(Function &F); 83 84 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 85 AU.setPreservesCFG(); 86 } 87 private: 88 void BuildRankMap(Function &F); 89 unsigned getRank(Value *V); 90 Value *ReassociateExpression(BinaryOperator *I); 91 void RewriteExprTree(BinaryOperator *I, SmallVectorImpl<ValueEntry> &Ops, 92 unsigned Idx = 0); 93 Value *OptimizeExpression(BinaryOperator *I, 94 SmallVectorImpl<ValueEntry> &Ops); 95 Value *OptimizeAdd(Instruction *I, SmallVectorImpl<ValueEntry> &Ops); 96 void LinearizeExprTree(BinaryOperator *I, SmallVectorImpl<ValueEntry> &Ops); 97 void LinearizeExpr(BinaryOperator *I); 98 Value *RemoveFactorFromExpression(Value *V, Value *Factor); 99 void ReassociateBB(BasicBlock *BB); 100 101 void RemoveDeadBinaryOp(Value *V); 102 }; 103} 104 105char Reassociate::ID = 0; 106INITIALIZE_PASS(Reassociate, "reassociate", 107 "Reassociate expressions", false, false); 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 Instruction *I = dyn_cast<Instruction>(V); 167 if (I == 0) { 168 if (isa<Argument>(V)) return ValueRankMap[V]; // Function argument. 169 return 0; // Otherwise it's a global or constant, rank 0. 170 } 171 172 if (unsigned Rank = ValueRankMap[I]) 173 return Rank; // Rank already known? 174 175 // If this is an expression, return the 1+MAX(rank(LHS), rank(RHS)) so that 176 // we can reassociate expressions for code motion! Since we do not recurse 177 // for PHI nodes, we cannot have infinite recursion here, because there 178 // cannot be loops in the value graph that do not go through PHI nodes. 179 unsigned Rank = 0, MaxRank = RankMap[I->getParent()]; 180 for (unsigned i = 0, e = I->getNumOperands(); 181 i != e && Rank != MaxRank; ++i) 182 Rank = std::max(Rank, getRank(I->getOperand(i))); 183 184 // If this is a not or neg instruction, do not count it for rank. This 185 // assures us that X and ~X will have the same rank. 186 if (!I->getType()->isIntegerTy() || 187 (!BinaryOperator::isNot(I) && !BinaryOperator::isNeg(I))) 188 ++Rank; 189 190 //DEBUG(dbgs() << "Calculated Rank[" << V->getName() << "] = " 191 // << Rank << "\n"); 192 193 return ValueRankMap[I] = Rank; 194} 195 196/// isReassociableOp - Return true if V is an instruction of the specified 197/// opcode and if it only has one use. 198static BinaryOperator *isReassociableOp(Value *V, unsigned Opcode) { 199 if ((V->hasOneUse() || V->use_empty()) && isa<Instruction>(V) && 200 cast<Instruction>(V)->getOpcode() == Opcode) 201 return cast<BinaryOperator>(V); 202 return 0; 203} 204 205/// LowerNegateToMultiply - Replace 0-X with X*-1. 206/// 207static Instruction *LowerNegateToMultiply(Instruction *Neg, 208 DenseMap<AssertingVH<>, unsigned> &ValueRankMap) { 209 Constant *Cst = Constant::getAllOnesValue(Neg->getType()); 210 211 Instruction *Res = BinaryOperator::CreateMul(Neg->getOperand(1), Cst, "",Neg); 212 ValueRankMap.erase(Neg); 213 Res->takeName(Neg); 214 Neg->replaceAllUsesWith(Res); 215 Neg->eraseFromParent(); 216 return Res; 217} 218 219// Given an expression of the form '(A+B)+(D+C)', turn it into '(((A+B)+C)+D)'. 220// Note that if D is also part of the expression tree that we recurse to 221// linearize it as well. Besides that case, this does not recurse into A,B, or 222// C. 223void Reassociate::LinearizeExpr(BinaryOperator *I) { 224 BinaryOperator *LHS = cast<BinaryOperator>(I->getOperand(0)); 225 BinaryOperator *RHS = cast<BinaryOperator>(I->getOperand(1)); 226 assert(isReassociableOp(LHS, I->getOpcode()) && 227 isReassociableOp(RHS, I->getOpcode()) && 228 "Not an expression that needs linearization?"); 229 230 DEBUG(dbgs() << "Linear" << *LHS << '\n' << *RHS << '\n' << *I << '\n'); 231 232 // Move the RHS instruction to live immediately before I, avoiding breaking 233 // dominator properties. 234 RHS->moveBefore(I); 235 236 // Move operands around to do the linearization. 237 I->setOperand(1, RHS->getOperand(0)); 238 RHS->setOperand(0, LHS); 239 I->setOperand(0, RHS); 240 241 ++NumLinear; 242 MadeChange = true; 243 DEBUG(dbgs() << "Linearized: " << *I << '\n'); 244 245 // If D is part of this expression tree, tail recurse. 246 if (isReassociableOp(I->getOperand(1), I->getOpcode())) 247 LinearizeExpr(I); 248} 249 250 251/// LinearizeExprTree - Given an associative binary expression tree, traverse 252/// all of the uses putting it into canonical form. This forces a left-linear 253/// form of the expression (((a+b)+c)+d), and collects information about the 254/// rank of the non-tree operands. 255/// 256/// NOTE: These intentionally destroys the expression tree operands (turning 257/// them into undef values) to reduce #uses of the values. This means that the 258/// caller MUST use something like RewriteExprTree to put the values back in. 259/// 260void Reassociate::LinearizeExprTree(BinaryOperator *I, 261 SmallVectorImpl<ValueEntry> &Ops) { 262 Value *LHS = I->getOperand(0), *RHS = I->getOperand(1); 263 unsigned Opcode = I->getOpcode(); 264 265 // First step, linearize the expression if it is in ((A+B)+(C+D)) form. 266 BinaryOperator *LHSBO = isReassociableOp(LHS, Opcode); 267 BinaryOperator *RHSBO = isReassociableOp(RHS, Opcode); 268 269 // If this is a multiply expression tree and it contains internal negations, 270 // transform them into multiplies by -1 so they can be reassociated. 271 if (I->getOpcode() == Instruction::Mul) { 272 if (!LHSBO && LHS->hasOneUse() && BinaryOperator::isNeg(LHS)) { 273 LHS = LowerNegateToMultiply(cast<Instruction>(LHS), ValueRankMap); 274 LHSBO = isReassociableOp(LHS, Opcode); 275 } 276 if (!RHSBO && RHS->hasOneUse() && BinaryOperator::isNeg(RHS)) { 277 RHS = LowerNegateToMultiply(cast<Instruction>(RHS), ValueRankMap); 278 RHSBO = isReassociableOp(RHS, Opcode); 279 } 280 } 281 282 if (!LHSBO) { 283 if (!RHSBO) { 284 // Neither the LHS or RHS as part of the tree, thus this is a leaf. As 285 // such, just remember these operands and their rank. 286 Ops.push_back(ValueEntry(getRank(LHS), LHS)); 287 Ops.push_back(ValueEntry(getRank(RHS), RHS)); 288 289 // Clear the leaves out. 290 I->setOperand(0, UndefValue::get(I->getType())); 291 I->setOperand(1, UndefValue::get(I->getType())); 292 return; 293 } 294 295 // Turn X+(Y+Z) -> (Y+Z)+X 296 std::swap(LHSBO, RHSBO); 297 std::swap(LHS, RHS); 298 bool Success = !I->swapOperands(); 299 assert(Success && "swapOperands failed"); 300 Success = false; 301 MadeChange = true; 302 } else if (RHSBO) { 303 // Turn (A+B)+(C+D) -> (((A+B)+C)+D). This guarantees the RHS is not 304 // part of the expression tree. 305 LinearizeExpr(I); 306 LHS = LHSBO = cast<BinaryOperator>(I->getOperand(0)); 307 RHS = I->getOperand(1); 308 RHSBO = 0; 309 } 310 311 // Okay, now we know that the LHS is a nested expression and that the RHS is 312 // not. Perform reassociation. 313 assert(!isReassociableOp(RHS, Opcode) && "LinearizeExpr failed!"); 314 315 // Move LHS right before I to make sure that the tree expression dominates all 316 // values. 317 LHSBO->moveBefore(I); 318 319 // Linearize the expression tree on the LHS. 320 LinearizeExprTree(LHSBO, Ops); 321 322 // Remember the RHS operand and its rank. 323 Ops.push_back(ValueEntry(getRank(RHS), RHS)); 324 325 // Clear the RHS leaf out. 326 I->setOperand(1, UndefValue::get(I->getType())); 327} 328 329// RewriteExprTree - Now that the operands for this expression tree are 330// linearized and optimized, emit them in-order. This function is written to be 331// tail recursive. 332void Reassociate::RewriteExprTree(BinaryOperator *I, 333 SmallVectorImpl<ValueEntry> &Ops, 334 unsigned i) { 335 if (i+2 == Ops.size()) { 336 if (I->getOperand(0) != Ops[i].Op || 337 I->getOperand(1) != Ops[i+1].Op) { 338 Value *OldLHS = I->getOperand(0); 339 DEBUG(dbgs() << "RA: " << *I << '\n'); 340 I->setOperand(0, Ops[i].Op); 341 I->setOperand(1, Ops[i+1].Op); 342 DEBUG(dbgs() << "TO: " << *I << '\n'); 343 MadeChange = true; 344 ++NumChanged; 345 346 // If we reassociated a tree to fewer operands (e.g. (1+a+2) -> (a+3) 347 // delete the extra, now dead, nodes. 348 RemoveDeadBinaryOp(OldLHS); 349 } 350 return; 351 } 352 assert(i+2 < Ops.size() && "Ops index out of range!"); 353 354 if (I->getOperand(1) != Ops[i].Op) { 355 DEBUG(dbgs() << "RA: " << *I << '\n'); 356 I->setOperand(1, Ops[i].Op); 357 DEBUG(dbgs() << "TO: " << *I << '\n'); 358 MadeChange = true; 359 ++NumChanged; 360 } 361 362 BinaryOperator *LHS = cast<BinaryOperator>(I->getOperand(0)); 363 assert(LHS->getOpcode() == I->getOpcode() && 364 "Improper expression tree!"); 365 366 // Compactify the tree instructions together with each other to guarantee 367 // that the expression tree is dominated by all of Ops. 368 LHS->moveBefore(I); 369 RewriteExprTree(LHS, Ops, i+1); 370} 371 372 373 374// NegateValue - Insert instructions before the instruction pointed to by BI, 375// that computes the negative version of the value specified. The negative 376// version of the value is returned, and BI is left pointing at the instruction 377// that should be processed next by the reassociation pass. 378// 379static Value *NegateValue(Value *V, Instruction *BI) { 380 if (Constant *C = dyn_cast<Constant>(V)) 381 return ConstantExpr::getNeg(C); 382 383 // We are trying to expose opportunity for reassociation. One of the things 384 // that we want to do to achieve this is to push a negation as deep into an 385 // expression chain as possible, to expose the add instructions. In practice, 386 // this means that we turn this: 387 // X = -(A+12+C+D) into X = -A + -12 + -C + -D = -12 + -A + -C + -D 388 // so that later, a: Y = 12+X could get reassociated with the -12 to eliminate 389 // the constants. We assume that instcombine will clean up the mess later if 390 // we introduce tons of unnecessary negation instructions. 391 // 392 if (Instruction *I = dyn_cast<Instruction>(V)) 393 if (I->getOpcode() == Instruction::Add && I->hasOneUse()) { 394 // Push the negates through the add. 395 I->setOperand(0, NegateValue(I->getOperand(0), BI)); 396 I->setOperand(1, NegateValue(I->getOperand(1), BI)); 397 398 // We must move the add instruction here, because the neg instructions do 399 // not dominate the old add instruction in general. By moving it, we are 400 // assured that the neg instructions we just inserted dominate the 401 // instruction we are about to insert after them. 402 // 403 I->moveBefore(BI); 404 I->setName(I->getName()+".neg"); 405 return I; 406 } 407 408 // Okay, we need to materialize a negated version of V with an instruction. 409 // Scan the use lists of V to see if we have one already. 410 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E;++UI){ 411 User *U = *UI; 412 if (!BinaryOperator::isNeg(U)) continue; 413 414 // We found one! Now we have to make sure that the definition dominates 415 // this use. We do this by moving it to the entry block (if it is a 416 // non-instruction value) or right after the definition. These negates will 417 // be zapped by reassociate later, so we don't need much finesse here. 418 BinaryOperator *TheNeg = cast<BinaryOperator>(U); 419 420 // Verify that the negate is in this function, V might be a constant expr. 421 if (TheNeg->getParent()->getParent() != BI->getParent()->getParent()) 422 continue; 423 424 BasicBlock::iterator InsertPt; 425 if (Instruction *InstInput = dyn_cast<Instruction>(V)) { 426 if (InvokeInst *II = dyn_cast<InvokeInst>(InstInput)) { 427 InsertPt = II->getNormalDest()->begin(); 428 } else { 429 InsertPt = InstInput; 430 ++InsertPt; 431 } 432 while (isa<PHINode>(InsertPt)) ++InsertPt; 433 } else { 434 InsertPt = TheNeg->getParent()->getParent()->getEntryBlock().begin(); 435 } 436 TheNeg->moveBefore(InsertPt); 437 return TheNeg; 438 } 439 440 // Insert a 'neg' instruction that subtracts the value from zero to get the 441 // negation. 442 return BinaryOperator::CreateNeg(V, V->getName() + ".neg", BI); 443} 444 445/// ShouldBreakUpSubtract - Return true if we should break up this subtract of 446/// X-Y into (X + -Y). 447static bool ShouldBreakUpSubtract(Instruction *Sub) { 448 // If this is a negation, we can't split it up! 449 if (BinaryOperator::isNeg(Sub)) 450 return false; 451 452 // Don't bother to break this up unless either the LHS is an associable add or 453 // subtract or if this is only used by one. 454 if (isReassociableOp(Sub->getOperand(0), Instruction::Add) || 455 isReassociableOp(Sub->getOperand(0), Instruction::Sub)) 456 return true; 457 if (isReassociableOp(Sub->getOperand(1), Instruction::Add) || 458 isReassociableOp(Sub->getOperand(1), Instruction::Sub)) 459 return true; 460 if (Sub->hasOneUse() && 461 (isReassociableOp(Sub->use_back(), Instruction::Add) || 462 isReassociableOp(Sub->use_back(), Instruction::Sub))) 463 return true; 464 465 return false; 466} 467 468/// BreakUpSubtract - If we have (X-Y), and if either X is an add, or if this is 469/// only used by an add, transform this into (X+(0-Y)) to promote better 470/// reassociation. 471static Instruction *BreakUpSubtract(Instruction *Sub, 472 DenseMap<AssertingVH<>, unsigned> &ValueRankMap) { 473 // Convert a subtract into an add and a neg instruction. This allows sub 474 // instructions to be commuted with other add instructions. 475 // 476 // Calculate the negative value of Operand 1 of the sub instruction, 477 // and set it as the RHS of the add instruction we just made. 478 // 479 Value *NegVal = NegateValue(Sub->getOperand(1), Sub); 480 Instruction *New = 481 BinaryOperator::CreateAdd(Sub->getOperand(0), NegVal, "", Sub); 482 New->takeName(Sub); 483 484 // Everyone now refers to the add instruction. 485 ValueRankMap.erase(Sub); 486 Sub->replaceAllUsesWith(New); 487 Sub->eraseFromParent(); 488 489 DEBUG(dbgs() << "Negated: " << *New << '\n'); 490 return New; 491} 492 493/// ConvertShiftToMul - If this is a shift of a reassociable multiply or is used 494/// by one, change this into a multiply by a constant to assist with further 495/// reassociation. 496static Instruction *ConvertShiftToMul(Instruction *Shl, 497 DenseMap<AssertingVH<>, unsigned> &ValueRankMap) { 498 // If an operand of this shift is a reassociable multiply, or if the shift 499 // is used by a reassociable multiply or add, turn into a multiply. 500 if (isReassociableOp(Shl->getOperand(0), Instruction::Mul) || 501 (Shl->hasOneUse() && 502 (isReassociableOp(Shl->use_back(), Instruction::Mul) || 503 isReassociableOp(Shl->use_back(), Instruction::Add)))) { 504 Constant *MulCst = ConstantInt::get(Shl->getType(), 1); 505 MulCst = ConstantExpr::getShl(MulCst, cast<Constant>(Shl->getOperand(1))); 506 507 Instruction *Mul = 508 BinaryOperator::CreateMul(Shl->getOperand(0), MulCst, "", Shl); 509 ValueRankMap.erase(Shl); 510 Mul->takeName(Shl); 511 Shl->replaceAllUsesWith(Mul); 512 Shl->eraseFromParent(); 513 return Mul; 514 } 515 return 0; 516} 517 518// Scan backwards and forwards among values with the same rank as element i to 519// see if X exists. If X does not exist, return i. This is useful when 520// scanning for 'x' when we see '-x' because they both get the same rank. 521static unsigned FindInOperandList(SmallVectorImpl<ValueEntry> &Ops, unsigned i, 522 Value *X) { 523 unsigned XRank = Ops[i].Rank; 524 unsigned e = Ops.size(); 525 for (unsigned j = i+1; j != e && Ops[j].Rank == XRank; ++j) 526 if (Ops[j].Op == X) 527 return j; 528 // Scan backwards. 529 for (unsigned j = i-1; j != ~0U && Ops[j].Rank == XRank; --j) 530 if (Ops[j].Op == X) 531 return j; 532 return i; 533} 534 535/// EmitAddTreeOfValues - Emit a tree of add instructions, summing Ops together 536/// and returning the result. Insert the tree before I. 537static Value *EmitAddTreeOfValues(Instruction *I, SmallVectorImpl<Value*> &Ops){ 538 if (Ops.size() == 1) return Ops.back(); 539 540 Value *V1 = Ops.back(); 541 Ops.pop_back(); 542 Value *V2 = EmitAddTreeOfValues(I, Ops); 543 return BinaryOperator::CreateAdd(V2, V1, "tmp", I); 544} 545 546/// RemoveFactorFromExpression - If V is an expression tree that is a 547/// multiplication sequence, and if this sequence contains a multiply by Factor, 548/// remove Factor from the tree and return the new tree. 549Value *Reassociate::RemoveFactorFromExpression(Value *V, Value *Factor) { 550 BinaryOperator *BO = isReassociableOp(V, Instruction::Mul); 551 if (!BO) return 0; 552 553 SmallVector<ValueEntry, 8> Factors; 554 LinearizeExprTree(BO, Factors); 555 556 bool FoundFactor = false; 557 bool NeedsNegate = false; 558 for (unsigned i = 0, e = Factors.size(); i != e; ++i) { 559 if (Factors[i].Op == Factor) { 560 FoundFactor = true; 561 Factors.erase(Factors.begin()+i); 562 break; 563 } 564 565 // If this is a negative version of this factor, remove it. 566 if (ConstantInt *FC1 = dyn_cast<ConstantInt>(Factor)) 567 if (ConstantInt *FC2 = dyn_cast<ConstantInt>(Factors[i].Op)) 568 if (FC1->getValue() == -FC2->getValue()) { 569 FoundFactor = NeedsNegate = true; 570 Factors.erase(Factors.begin()+i); 571 break; 572 } 573 } 574 575 if (!FoundFactor) { 576 // Make sure to restore the operands to the expression tree. 577 RewriteExprTree(BO, Factors); 578 return 0; 579 } 580 581 BasicBlock::iterator InsertPt = BO; ++InsertPt; 582 583 // If this was just a single multiply, remove the multiply and return the only 584 // remaining operand. 585 if (Factors.size() == 1) { 586 ValueRankMap.erase(BO); 587 BO->eraseFromParent(); 588 V = Factors[0].Op; 589 } else { 590 RewriteExprTree(BO, Factors); 591 V = BO; 592 } 593 594 if (NeedsNegate) 595 V = BinaryOperator::CreateNeg(V, "neg", InsertPt); 596 597 return V; 598} 599 600/// FindSingleUseMultiplyFactors - If V is a single-use multiply, recursively 601/// add its operands as factors, otherwise add V to the list of factors. 602/// 603/// Ops is the top-level list of add operands we're trying to factor. 604static void FindSingleUseMultiplyFactors(Value *V, 605 SmallVectorImpl<Value*> &Factors, 606 const SmallVectorImpl<ValueEntry> &Ops, 607 bool IsRoot) { 608 BinaryOperator *BO; 609 if (!(V->hasOneUse() || V->use_empty()) || // More than one use. 610 !(BO = dyn_cast<BinaryOperator>(V)) || 611 BO->getOpcode() != Instruction::Mul) { 612 Factors.push_back(V); 613 return; 614 } 615 616 // If this value has a single use because it is another input to the add 617 // tree we're reassociating and we dropped its use, it actually has two 618 // uses and we can't factor it. 619 if (!IsRoot) { 620 for (unsigned i = 0, e = Ops.size(); i != e; ++i) 621 if (Ops[i].Op == V) { 622 Factors.push_back(V); 623 return; 624 } 625 } 626 627 628 // Otherwise, add the LHS and RHS to the list of factors. 629 FindSingleUseMultiplyFactors(BO->getOperand(1), Factors, Ops, false); 630 FindSingleUseMultiplyFactors(BO->getOperand(0), Factors, Ops, false); 631} 632 633/// OptimizeAndOrXor - Optimize a series of operands to an 'and', 'or', or 'xor' 634/// instruction. This optimizes based on identities. If it can be reduced to 635/// a single Value, it is returned, otherwise the Ops list is mutated as 636/// necessary. 637static Value *OptimizeAndOrXor(unsigned Opcode, 638 SmallVectorImpl<ValueEntry> &Ops) { 639 // Scan the operand lists looking for X and ~X pairs, along with X,X pairs. 640 // If we find any, we can simplify the expression. X&~X == 0, X|~X == -1. 641 for (unsigned i = 0, e = Ops.size(); i != e; ++i) { 642 // First, check for X and ~X in the operand list. 643 assert(i < Ops.size()); 644 if (BinaryOperator::isNot(Ops[i].Op)) { // Cannot occur for ^. 645 Value *X = BinaryOperator::getNotArgument(Ops[i].Op); 646 unsigned FoundX = FindInOperandList(Ops, i, X); 647 if (FoundX != i) { 648 if (Opcode == Instruction::And) // ...&X&~X = 0 649 return Constant::getNullValue(X->getType()); 650 651 if (Opcode == Instruction::Or) // ...|X|~X = -1 652 return Constant::getAllOnesValue(X->getType()); 653 } 654 } 655 656 // Next, check for duplicate pairs of values, which we assume are next to 657 // each other, due to our sorting criteria. 658 assert(i < Ops.size()); 659 if (i+1 != Ops.size() && Ops[i+1].Op == Ops[i].Op) { 660 if (Opcode == Instruction::And || Opcode == Instruction::Or) { 661 // Drop duplicate values for And and Or. 662 Ops.erase(Ops.begin()+i); 663 --i; --e; 664 ++NumAnnihil; 665 continue; 666 } 667 668 // Drop pairs of values for Xor. 669 assert(Opcode == Instruction::Xor); 670 if (e == 2) 671 return Constant::getNullValue(Ops[0].Op->getType()); 672 673 // Y ^ X^X -> Y 674 Ops.erase(Ops.begin()+i, Ops.begin()+i+2); 675 i -= 1; e -= 2; 676 ++NumAnnihil; 677 } 678 } 679 return 0; 680} 681 682/// OptimizeAdd - Optimize a series of operands to an 'add' instruction. This 683/// optimizes based on identities. If it can be reduced to a single Value, it 684/// is returned, otherwise the Ops list is mutated as necessary. 685Value *Reassociate::OptimizeAdd(Instruction *I, 686 SmallVectorImpl<ValueEntry> &Ops) { 687 // Scan the operand lists looking for X and -X pairs. If we find any, we 688 // can simplify the expression. X+-X == 0. While we're at it, scan for any 689 // duplicates. We want to canonicalize Y+Y+Y+Z -> 3*Y+Z. 690 // 691 // TODO: We could handle "X + ~X" -> "-1" if we wanted, since "-X = ~X+1". 692 // 693 for (unsigned i = 0, e = Ops.size(); i != e; ++i) { 694 Value *TheOp = Ops[i].Op; 695 // Check to see if we've seen this operand before. If so, we factor all 696 // instances of the operand together. Due to our sorting criteria, we know 697 // that these need to be next to each other in the vector. 698 if (i+1 != Ops.size() && Ops[i+1].Op == TheOp) { 699 // Rescan the list, remove all instances of this operand from the expr. 700 unsigned NumFound = 0; 701 do { 702 Ops.erase(Ops.begin()+i); 703 ++NumFound; 704 } while (i != Ops.size() && Ops[i].Op == TheOp); 705 706 DEBUG(errs() << "\nFACTORING [" << NumFound << "]: " << *TheOp << '\n'); 707 ++NumFactor; 708 709 // Insert a new multiply. 710 Value *Mul = ConstantInt::get(cast<IntegerType>(I->getType()), NumFound); 711 Mul = BinaryOperator::CreateMul(TheOp, Mul, "factor", I); 712 713 // Now that we have inserted a multiply, optimize it. This allows us to 714 // handle cases that require multiple factoring steps, such as this: 715 // (X*2) + (X*2) + (X*2) -> (X*2)*3 -> X*6 716 Mul = ReassociateExpression(cast<BinaryOperator>(Mul)); 717 718 // If every add operand was a duplicate, return the multiply. 719 if (Ops.empty()) 720 return Mul; 721 722 // Otherwise, we had some input that didn't have the dupe, such as 723 // "A + A + B" -> "A*2 + B". Add the new multiply to the list of 724 // things being added by this operation. 725 Ops.insert(Ops.begin(), ValueEntry(getRank(Mul), Mul)); 726 727 --i; 728 e = Ops.size(); 729 continue; 730 } 731 732 // Check for X and -X in the operand list. 733 if (!BinaryOperator::isNeg(TheOp)) 734 continue; 735 736 Value *X = BinaryOperator::getNegArgument(TheOp); 737 unsigned FoundX = FindInOperandList(Ops, i, X); 738 if (FoundX == i) 739 continue; 740 741 // Remove X and -X from the operand list. 742 if (Ops.size() == 2) 743 return Constant::getNullValue(X->getType()); 744 745 Ops.erase(Ops.begin()+i); 746 if (i < FoundX) 747 --FoundX; 748 else 749 --i; // Need to back up an extra one. 750 Ops.erase(Ops.begin()+FoundX); 751 ++NumAnnihil; 752 --i; // Revisit element. 753 e -= 2; // Removed two elements. 754 } 755 756 // Scan the operand list, checking to see if there are any common factors 757 // between operands. Consider something like A*A+A*B*C+D. We would like to 758 // reassociate this to A*(A+B*C)+D, which reduces the number of multiplies. 759 // To efficiently find this, we count the number of times a factor occurs 760 // for any ADD operands that are MULs. 761 DenseMap<Value*, unsigned> FactorOccurrences; 762 763 // Keep track of each multiply we see, to avoid triggering on (X*4)+(X*4) 764 // where they are actually the same multiply. 765 unsigned MaxOcc = 0; 766 Value *MaxOccVal = 0; 767 for (unsigned i = 0, e = Ops.size(); i != e; ++i) { 768 BinaryOperator *BOp = dyn_cast<BinaryOperator>(Ops[i].Op); 769 if (BOp == 0 || BOp->getOpcode() != Instruction::Mul || !BOp->use_empty()) 770 continue; 771 772 // Compute all of the factors of this added value. 773 SmallVector<Value*, 8> Factors; 774 FindSingleUseMultiplyFactors(BOp, Factors, Ops, true); 775 assert(Factors.size() > 1 && "Bad linearize!"); 776 777 // Add one to FactorOccurrences for each unique factor in this op. 778 SmallPtrSet<Value*, 8> Duplicates; 779 for (unsigned i = 0, e = Factors.size(); i != e; ++i) { 780 Value *Factor = Factors[i]; 781 if (!Duplicates.insert(Factor)) continue; 782 783 unsigned Occ = ++FactorOccurrences[Factor]; 784 if (Occ > MaxOcc) { MaxOcc = Occ; MaxOccVal = Factor; } 785 786 // If Factor is a negative constant, add the negated value as a factor 787 // because we can percolate the negate out. Watch for minint, which 788 // cannot be positivified. 789 if (ConstantInt *CI = dyn_cast<ConstantInt>(Factor)) 790 if (CI->getValue().isNegative() && !CI->getValue().isMinSignedValue()) { 791 Factor = ConstantInt::get(CI->getContext(), -CI->getValue()); 792 assert(!Duplicates.count(Factor) && 793 "Shouldn't have two constant factors, missed a canonicalize"); 794 795 unsigned Occ = ++FactorOccurrences[Factor]; 796 if (Occ > MaxOcc) { MaxOcc = Occ; MaxOccVal = Factor; } 797 } 798 } 799 } 800 801 // If any factor occurred more than one time, we can pull it out. 802 if (MaxOcc > 1) { 803 DEBUG(errs() << "\nFACTORING [" << MaxOcc << "]: " << *MaxOccVal << '\n'); 804 ++NumFactor; 805 806 // Create a new instruction that uses the MaxOccVal twice. If we don't do 807 // this, we could otherwise run into situations where removing a factor 808 // from an expression will drop a use of maxocc, and this can cause 809 // RemoveFactorFromExpression on successive values to behave differently. 810 Instruction *DummyInst = BinaryOperator::CreateAdd(MaxOccVal, MaxOccVal); 811 SmallVector<Value*, 4> NewMulOps; 812 for (unsigned i = 0, e = Ops.size(); i != e; ++i) { 813 // Only try to remove factors from expressions we're allowed to. 814 BinaryOperator *BOp = dyn_cast<BinaryOperator>(Ops[i].Op); 815 if (BOp == 0 || BOp->getOpcode() != Instruction::Mul || !BOp->use_empty()) 816 continue; 817 818 if (Value *V = RemoveFactorFromExpression(Ops[i].Op, MaxOccVal)) { 819 NewMulOps.push_back(V); 820 Ops.erase(Ops.begin()+i); 821 --i; --e; 822 } 823 } 824 825 // No need for extra uses anymore. 826 delete DummyInst; 827 828 unsigned NumAddedValues = NewMulOps.size(); 829 Value *V = EmitAddTreeOfValues(I, NewMulOps); 830 831 // Now that we have inserted the add tree, optimize it. This allows us to 832 // handle cases that require multiple factoring steps, such as this: 833 // A*A*B + A*A*C --> A*(A*B+A*C) --> A*(A*(B+C)) 834 assert(NumAddedValues > 1 && "Each occurrence should contribute a value"); 835 (void)NumAddedValues; 836 V = ReassociateExpression(cast<BinaryOperator>(V)); 837 838 // Create the multiply. 839 Value *V2 = BinaryOperator::CreateMul(V, MaxOccVal, "tmp", I); 840 841 // Rerun associate on the multiply in case the inner expression turned into 842 // a multiply. We want to make sure that we keep things in canonical form. 843 V2 = ReassociateExpression(cast<BinaryOperator>(V2)); 844 845 // If every add operand included the factor (e.g. "A*B + A*C"), then the 846 // entire result expression is just the multiply "A*(B+C)". 847 if (Ops.empty()) 848 return V2; 849 850 // Otherwise, we had some input that didn't have the factor, such as 851 // "A*B + A*C + D" -> "A*(B+C) + D". Add the new multiply to the list of 852 // things being added by this operation. 853 Ops.insert(Ops.begin(), ValueEntry(getRank(V2), V2)); 854 } 855 856 return 0; 857} 858 859Value *Reassociate::OptimizeExpression(BinaryOperator *I, 860 SmallVectorImpl<ValueEntry> &Ops) { 861 // Now that we have the linearized expression tree, try to optimize it. 862 // Start by folding any constants that we found. 863 bool IterateOptimization = false; 864 if (Ops.size() == 1) return Ops[0].Op; 865 866 unsigned Opcode = I->getOpcode(); 867 868 if (Constant *V1 = dyn_cast<Constant>(Ops[Ops.size()-2].Op)) 869 if (Constant *V2 = dyn_cast<Constant>(Ops.back().Op)) { 870 Ops.pop_back(); 871 Ops.back().Op = ConstantExpr::get(Opcode, V1, V2); 872 return OptimizeExpression(I, Ops); 873 } 874 875 // Check for destructive annihilation due to a constant being used. 876 if (ConstantInt *CstVal = dyn_cast<ConstantInt>(Ops.back().Op)) 877 switch (Opcode) { 878 default: break; 879 case Instruction::And: 880 if (CstVal->isZero()) // X & 0 -> 0 881 return CstVal; 882 if (CstVal->isAllOnesValue()) // X & -1 -> X 883 Ops.pop_back(); 884 break; 885 case Instruction::Mul: 886 if (CstVal->isZero()) { // X * 0 -> 0 887 ++NumAnnihil; 888 return CstVal; 889 } 890 891 if (cast<ConstantInt>(CstVal)->isOne()) 892 Ops.pop_back(); // X * 1 -> X 893 break; 894 case Instruction::Or: 895 if (CstVal->isAllOnesValue()) // X | -1 -> -1 896 return CstVal; 897 // FALLTHROUGH! 898 case Instruction::Add: 899 case Instruction::Xor: 900 if (CstVal->isZero()) // X [|^+] 0 -> X 901 Ops.pop_back(); 902 break; 903 } 904 if (Ops.size() == 1) return Ops[0].Op; 905 906 // Handle destructive annihilation due to identities between elements in the 907 // argument list here. 908 switch (Opcode) { 909 default: break; 910 case Instruction::And: 911 case Instruction::Or: 912 case Instruction::Xor: { 913 unsigned NumOps = Ops.size(); 914 if (Value *Result = OptimizeAndOrXor(Opcode, Ops)) 915 return Result; 916 IterateOptimization |= Ops.size() != NumOps; 917 break; 918 } 919 920 case Instruction::Add: { 921 unsigned NumOps = Ops.size(); 922 if (Value *Result = OptimizeAdd(I, Ops)) 923 return Result; 924 IterateOptimization |= Ops.size() != NumOps; 925 } 926 927 break; 928 //case Instruction::Mul: 929 } 930 931 if (IterateOptimization) 932 return OptimizeExpression(I, Ops); 933 return 0; 934} 935 936 937/// ReassociateBB - Inspect all of the instructions in this basic block, 938/// reassociating them as we go. 939void Reassociate::ReassociateBB(BasicBlock *BB) { 940 for (BasicBlock::iterator BBI = BB->begin(); BBI != BB->end(); ) { 941 Instruction *BI = BBI++; 942 if (BI->getOpcode() == Instruction::Shl && 943 isa<ConstantInt>(BI->getOperand(1))) 944 if (Instruction *NI = ConvertShiftToMul(BI, ValueRankMap)) { 945 MadeChange = true; 946 BI = NI; 947 } 948 949 // Reject cases where it is pointless to do this. 950 if (!isa<BinaryOperator>(BI) || BI->getType()->isFloatingPointTy() || 951 BI->getType()->isVectorTy()) 952 continue; // Floating point ops are not associative. 953 954 // Do not reassociate boolean (i1) expressions. We want to preserve the 955 // original order of evaluation for short-circuited comparisons that 956 // SimplifyCFG has folded to AND/OR expressions. If the expression 957 // is not further optimized, it is likely to be transformed back to a 958 // short-circuited form for code gen, and the source order may have been 959 // optimized for the most likely conditions. 960 if (BI->getType()->isIntegerTy(1)) 961 continue; 962 963 // If this is a subtract instruction which is not already in negate form, 964 // see if we can convert it to X+-Y. 965 if (BI->getOpcode() == Instruction::Sub) { 966 if (ShouldBreakUpSubtract(BI)) { 967 BI = BreakUpSubtract(BI, ValueRankMap); 968 // Reset the BBI iterator in case BreakUpSubtract changed the 969 // instruction it points to. 970 BBI = BI; 971 ++BBI; 972 MadeChange = true; 973 } else if (BinaryOperator::isNeg(BI)) { 974 // Otherwise, this is a negation. See if the operand is a multiply tree 975 // and if this is not an inner node of a multiply tree. 976 if (isReassociableOp(BI->getOperand(1), Instruction::Mul) && 977 (!BI->hasOneUse() || 978 !isReassociableOp(BI->use_back(), Instruction::Mul))) { 979 BI = LowerNegateToMultiply(BI, ValueRankMap); 980 MadeChange = true; 981 } 982 } 983 } 984 985 // If this instruction is a commutative binary operator, process it. 986 if (!BI->isAssociative()) continue; 987 BinaryOperator *I = cast<BinaryOperator>(BI); 988 989 // If this is an interior node of a reassociable tree, ignore it until we 990 // get to the root of the tree, to avoid N^2 analysis. 991 if (I->hasOneUse() && isReassociableOp(I->use_back(), I->getOpcode())) 992 continue; 993 994 // If this is an add tree that is used by a sub instruction, ignore it 995 // until we process the subtract. 996 if (I->hasOneUse() && I->getOpcode() == Instruction::Add && 997 cast<Instruction>(I->use_back())->getOpcode() == Instruction::Sub) 998 continue; 999 1000 ReassociateExpression(I); 1001 } 1002} 1003 1004Value *Reassociate::ReassociateExpression(BinaryOperator *I) { 1005 1006 // First, walk the expression tree, linearizing the tree, collecting the 1007 // operand information. 1008 SmallVector<ValueEntry, 8> Ops; 1009 LinearizeExprTree(I, Ops); 1010 1011 DEBUG(dbgs() << "RAIn:\t"; PrintOps(I, Ops); dbgs() << '\n'); 1012 1013 // Now that we have linearized the tree to a list and have gathered all of 1014 // the operands and their ranks, sort the operands by their rank. Use a 1015 // stable_sort so that values with equal ranks will have their relative 1016 // positions maintained (and so the compiler is deterministic). Note that 1017 // this sorts so that the highest ranking values end up at the beginning of 1018 // the vector. 1019 std::stable_sort(Ops.begin(), Ops.end()); 1020 1021 // OptimizeExpression - Now that we have the expression tree in a convenient 1022 // sorted form, optimize it globally if possible. 1023 if (Value *V = OptimizeExpression(I, Ops)) { 1024 // This expression tree simplified to something that isn't a tree, 1025 // eliminate it. 1026 DEBUG(dbgs() << "Reassoc to scalar: " << *V << '\n'); 1027 I->replaceAllUsesWith(V); 1028 RemoveDeadBinaryOp(I); 1029 ++NumAnnihil; 1030 return V; 1031 } 1032 1033 // We want to sink immediates as deeply as possible except in the case where 1034 // this is a multiply tree used only by an add, and the immediate is a -1. 1035 // In this case we reassociate to put the negation on the outside so that we 1036 // can fold the negation into the add: (-X)*Y + Z -> Z-X*Y 1037 if (I->getOpcode() == Instruction::Mul && I->hasOneUse() && 1038 cast<Instruction>(I->use_back())->getOpcode() == Instruction::Add && 1039 isa<ConstantInt>(Ops.back().Op) && 1040 cast<ConstantInt>(Ops.back().Op)->isAllOnesValue()) { 1041 ValueEntry Tmp = Ops.pop_back_val(); 1042 Ops.insert(Ops.begin(), Tmp); 1043 } 1044 1045 DEBUG(dbgs() << "RAOut:\t"; PrintOps(I, Ops); dbgs() << '\n'); 1046 1047 if (Ops.size() == 1) { 1048 // This expression tree simplified to something that isn't a tree, 1049 // eliminate it. 1050 I->replaceAllUsesWith(Ops[0].Op); 1051 RemoveDeadBinaryOp(I); 1052 return Ops[0].Op; 1053 } 1054 1055 // Now that we ordered and optimized the expressions, splat them back into 1056 // the expression tree, removing any unneeded nodes. 1057 RewriteExprTree(I, Ops); 1058 return I; 1059} 1060 1061 1062bool Reassociate::runOnFunction(Function &F) { 1063 // Recalculate the rank map for F 1064 BuildRankMap(F); 1065 1066 MadeChange = false; 1067 for (Function::iterator FI = F.begin(), FE = F.end(); FI != FE; ++FI) 1068 ReassociateBB(FI); 1069 1070 // We are done with the rank map. 1071 RankMap.clear(); 1072 ValueRankMap.clear(); 1073 return MadeChange; 1074} 1075 1076