Reassociate.cpp revision 35239934517c6fcd52e3e965f40e03f74aa4d11d
1//===- Reassociate.cpp - Reassociate binary expressions -------------------===// 2// 3// The LLVM Compiler Infrastructure 4// 5// This file is distributed under the University of Illinois Open Source 6// License. See LICENSE.TXT for details. 7// 8//===----------------------------------------------------------------------===// 9// 10// This pass reassociates commutative expressions in an order that is designed 11// to promote better constant propagation, GCSE, LICM, PRE... 12// 13// For example: 4 + (x + 5) -> x + (4 + 5) 14// 15// In the implementation of this algorithm, constants are assigned rank = 0, 16// function arguments are rank = 1, and other values are assigned ranks 17// corresponding to the reverse post order traversal of current function 18// (starting at 2), which effectively gives values in deep loops higher rank 19// than values not in loops. 20// 21//===----------------------------------------------------------------------===// 22 23#define DEBUG_TYPE "reassociate" 24#include "llvm/Transforms/Scalar.h" 25#include "llvm/Constants.h" 26#include "llvm/DerivedTypes.h" 27#include "llvm/Function.h" 28#include "llvm/Instructions.h" 29#include "llvm/IntrinsicInst.h" 30#include "llvm/Pass.h" 31#include "llvm/Assembly/Writer.h" 32#include "llvm/Support/CFG.h" 33#include "llvm/Support/Debug.h" 34#include "llvm/Support/ValueHandle.h" 35#include "llvm/Support/raw_ostream.h" 36#include "llvm/ADT/PostOrderIterator.h" 37#include "llvm/ADT/Statistic.h" 38#include "llvm/ADT/DenseMap.h" 39#include <algorithm> 40#include <map> 41using namespace llvm; 42 43STATISTIC(NumLinear , "Number of insts linearized"); 44STATISTIC(NumChanged, "Number of insts reassociated"); 45STATISTIC(NumAnnihil, "Number of expr tree annihilated"); 46STATISTIC(NumFactor , "Number of multiplies factored"); 47 48namespace { 49 struct ValueEntry { 50 unsigned Rank; 51 Value *Op; 52 ValueEntry(unsigned R, Value *O) : Rank(R), Op(O) {} 53 }; 54 inline bool operator<(const ValueEntry &LHS, const ValueEntry &RHS) { 55 return LHS.Rank > RHS.Rank; // Sort so that highest rank goes to start. 56 } 57} 58 59#ifndef NDEBUG 60/// PrintOps - Print out the expression identified in the Ops list. 61/// 62static void PrintOps(Instruction *I, const SmallVectorImpl<ValueEntry> &Ops) { 63 Module *M = I->getParent()->getParent()->getParent(); 64 errs() << Instruction::getOpcodeName(I->getOpcode()) << " " 65 << *Ops[0].Op->getType() << '\t'; 66 for (unsigned i = 0, e = Ops.size(); i != e; ++i) { 67 errs() << "[ "; 68 WriteAsOperand(errs(), Ops[i].Op, false, M); 69 errs() << ", #" << Ops[i].Rank << "] "; 70 } 71} 72#endif 73 74namespace { 75 class Reassociate : public FunctionPass { 76 std::map<BasicBlock*, unsigned> RankMap; 77 std::map<AssertingVH<>, unsigned> ValueRankMap; 78 bool MadeChange; 79 public: 80 static char ID; // Pass identification, replacement for typeid 81 Reassociate() : FunctionPass(&ID) {} 82 83 bool runOnFunction(Function &F); 84 85 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 86 AU.setPreservesCFG(); 87 } 88 private: 89 void BuildRankMap(Function &F); 90 unsigned getRank(Value *V); 91 Value *ReassociateExpression(BinaryOperator *I); 92 void RewriteExprTree(BinaryOperator *I, SmallVectorImpl<ValueEntry> &Ops, 93 unsigned Idx = 0); 94 Value *OptimizeExpression(BinaryOperator *I, 95 SmallVectorImpl<ValueEntry> &Ops); 96 Value *OptimizeAdd(Instruction *I, SmallVectorImpl<ValueEntry> &Ops); 97 void LinearizeExprTree(BinaryOperator *I, SmallVectorImpl<ValueEntry> &Ops); 98 void LinearizeExpr(BinaryOperator *I); 99 Value *RemoveFactorFromExpression(Value *V, Value *Factor); 100 void ReassociateBB(BasicBlock *BB); 101 102 void RemoveDeadBinaryOp(Value *V); 103 }; 104} 105 106char Reassociate::ID = 0; 107static RegisterPass<Reassociate> X("reassociate", "Reassociate expressions"); 108 109// Public interface to the Reassociate pass 110FunctionPass *llvm::createReassociatePass() { return new Reassociate(); } 111 112void Reassociate::RemoveDeadBinaryOp(Value *V) { 113 Instruction *Op = dyn_cast<Instruction>(V); 114 if (!Op || !isa<BinaryOperator>(Op) || !Op->use_empty()) 115 return; 116 117 Value *LHS = Op->getOperand(0), *RHS = Op->getOperand(1); 118 119 ValueRankMap.erase(Op); 120 Op->eraseFromParent(); 121 RemoveDeadBinaryOp(LHS); 122 RemoveDeadBinaryOp(RHS); 123} 124 125 126static bool isUnmovableInstruction(Instruction *I) { 127 if (I->getOpcode() == Instruction::PHI || 128 I->getOpcode() == Instruction::Alloca || 129 I->getOpcode() == Instruction::Load || 130 I->getOpcode() == Instruction::Invoke || 131 (I->getOpcode() == Instruction::Call && 132 !isa<DbgInfoIntrinsic>(I)) || 133 I->getOpcode() == Instruction::UDiv || 134 I->getOpcode() == Instruction::SDiv || 135 I->getOpcode() == Instruction::FDiv || 136 I->getOpcode() == Instruction::URem || 137 I->getOpcode() == Instruction::SRem || 138 I->getOpcode() == Instruction::FRem) 139 return true; 140 return false; 141} 142 143void Reassociate::BuildRankMap(Function &F) { 144 unsigned i = 2; 145 146 // Assign distinct ranks to function arguments 147 for (Function::arg_iterator I = F.arg_begin(), E = F.arg_end(); I != E; ++I) 148 ValueRankMap[&*I] = ++i; 149 150 ReversePostOrderTraversal<Function*> RPOT(&F); 151 for (ReversePostOrderTraversal<Function*>::rpo_iterator I = RPOT.begin(), 152 E = RPOT.end(); I != E; ++I) { 153 BasicBlock *BB = *I; 154 unsigned BBRank = RankMap[BB] = ++i << 16; 155 156 // Walk the basic block, adding precomputed ranks for any instructions that 157 // we cannot move. This ensures that the ranks for these instructions are 158 // all different in the block. 159 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) 160 if (isUnmovableInstruction(I)) 161 ValueRankMap[&*I] = ++BBRank; 162 } 163} 164 165unsigned Reassociate::getRank(Value *V) { 166 if (isa<Argument>(V)) return ValueRankMap[V]; // Function argument... 167 168 Instruction *I = dyn_cast<Instruction>(V); 169 if (I == 0) return 0; // Otherwise it's a global or constant, rank 0. 170 171 unsigned &CachedRank = ValueRankMap[I]; 172 if (CachedRank) return CachedRank; // Rank already known? 173 174 // If this is an expression, return the 1+MAX(rank(LHS), rank(RHS)) so that 175 // we can reassociate expressions for code motion! Since we do not recurse 176 // for PHI nodes, we cannot have infinite recursion here, because there 177 // cannot be loops in the value graph that do not go through PHI nodes. 178 unsigned Rank = 0, MaxRank = RankMap[I->getParent()]; 179 for (unsigned i = 0, e = I->getNumOperands(); 180 i != e && Rank != MaxRank; ++i) 181 Rank = std::max(Rank, getRank(I->getOperand(i))); 182 183 // If this is a not or neg instruction, do not count it for rank. This 184 // assures us that X and ~X will have the same rank. 185 if (!I->getType()->isInteger() || 186 (!BinaryOperator::isNot(I) && !BinaryOperator::isNeg(I))) 187 ++Rank; 188 189 //DEBUG(errs() << "Calculated Rank[" << V->getName() << "] = " 190 // << Rank << "\n"); 191 192 return CachedRank = Rank; 193} 194 195/// isReassociableOp - Return true if V is an instruction of the specified 196/// opcode and if it only has one use. 197static BinaryOperator *isReassociableOp(Value *V, unsigned Opcode) { 198 if ((V->hasOneUse() || V->use_empty()) && isa<Instruction>(V) && 199 cast<Instruction>(V)->getOpcode() == Opcode) 200 return cast<BinaryOperator>(V); 201 return 0; 202} 203 204/// LowerNegateToMultiply - Replace 0-X with X*-1. 205/// 206static Instruction *LowerNegateToMultiply(Instruction *Neg, 207 std::map<AssertingVH<>, unsigned> &ValueRankMap) { 208 Constant *Cst = Constant::getAllOnesValue(Neg->getType()); 209 210 Instruction *Res = BinaryOperator::CreateMul(Neg->getOperand(1), Cst, "",Neg); 211 ValueRankMap.erase(Neg); 212 Res->takeName(Neg); 213 Neg->replaceAllUsesWith(Res); 214 Neg->eraseFromParent(); 215 return Res; 216} 217 218// Given an expression of the form '(A+B)+(D+C)', turn it into '(((A+B)+C)+D)'. 219// Note that if D is also part of the expression tree that we recurse to 220// linearize it as well. Besides that case, this does not recurse into A,B, or 221// C. 222void Reassociate::LinearizeExpr(BinaryOperator *I) { 223 BinaryOperator *LHS = cast<BinaryOperator>(I->getOperand(0)); 224 BinaryOperator *RHS = cast<BinaryOperator>(I->getOperand(1)); 225 assert(isReassociableOp(LHS, I->getOpcode()) && 226 isReassociableOp(RHS, I->getOpcode()) && 227 "Not an expression that needs linearization?"); 228 229 DEBUG(errs() << "Linear" << *LHS << '\n' << *RHS << '\n' << *I << '\n'); 230 231 // Move the RHS instruction to live immediately before I, avoiding breaking 232 // dominator properties. 233 RHS->moveBefore(I); 234 235 // Move operands around to do the linearization. 236 I->setOperand(1, RHS->getOperand(0)); 237 RHS->setOperand(0, LHS); 238 I->setOperand(0, RHS); 239 240 ++NumLinear; 241 MadeChange = true; 242 DEBUG(errs() << "Linearized: " << *I << '\n'); 243 244 // If D is part of this expression tree, tail recurse. 245 if (isReassociableOp(I->getOperand(1), I->getOpcode())) 246 LinearizeExpr(I); 247} 248 249 250/// LinearizeExprTree - Given an associative binary expression tree, traverse 251/// all of the uses putting it into canonical form. This forces a left-linear 252/// form of the the expression (((a+b)+c)+d), and collects information about the 253/// rank of the non-tree operands. 254/// 255/// NOTE: These intentionally destroys the expression tree operands (turning 256/// them into undef values) to reduce #uses of the values. This means that the 257/// caller MUST use something like RewriteExprTree to put the values back in. 258/// 259void Reassociate::LinearizeExprTree(BinaryOperator *I, 260 SmallVectorImpl<ValueEntry> &Ops) { 261 Value *LHS = I->getOperand(0), *RHS = I->getOperand(1); 262 unsigned Opcode = I->getOpcode(); 263 264 // First step, linearize the expression if it is in ((A+B)+(C+D)) form. 265 BinaryOperator *LHSBO = isReassociableOp(LHS, Opcode); 266 BinaryOperator *RHSBO = isReassociableOp(RHS, Opcode); 267 268 // If this is a multiply expression tree and it contains internal negations, 269 // transform them into multiplies by -1 so they can be reassociated. 270 if (I->getOpcode() == Instruction::Mul) { 271 if (!LHSBO && LHS->hasOneUse() && BinaryOperator::isNeg(LHS)) { 272 LHS = LowerNegateToMultiply(cast<Instruction>(LHS), ValueRankMap); 273 LHSBO = isReassociableOp(LHS, Opcode); 274 } 275 if (!RHSBO && RHS->hasOneUse() && BinaryOperator::isNeg(RHS)) { 276 RHS = LowerNegateToMultiply(cast<Instruction>(RHS), ValueRankMap); 277 RHSBO = isReassociableOp(RHS, Opcode); 278 } 279 } 280 281 if (!LHSBO) { 282 if (!RHSBO) { 283 // Neither the LHS or RHS as part of the tree, thus this is a leaf. As 284 // such, just remember these operands and their rank. 285 Ops.push_back(ValueEntry(getRank(LHS), LHS)); 286 Ops.push_back(ValueEntry(getRank(RHS), RHS)); 287 288 // Clear the leaves out. 289 I->setOperand(0, UndefValue::get(I->getType())); 290 I->setOperand(1, UndefValue::get(I->getType())); 291 return; 292 } 293 294 // Turn X+(Y+Z) -> (Y+Z)+X 295 std::swap(LHSBO, RHSBO); 296 std::swap(LHS, RHS); 297 bool Success = !I->swapOperands(); 298 assert(Success && "swapOperands failed"); 299 Success = false; 300 MadeChange = true; 301 } else if (RHSBO) { 302 // Turn (A+B)+(C+D) -> (((A+B)+C)+D). This guarantees the the RHS is not 303 // part of the expression tree. 304 LinearizeExpr(I); 305 LHS = LHSBO = cast<BinaryOperator>(I->getOperand(0)); 306 RHS = I->getOperand(1); 307 RHSBO = 0; 308 } 309 310 // Okay, now we know that the LHS is a nested expression and that the RHS is 311 // not. Perform reassociation. 312 assert(!isReassociableOp(RHS, Opcode) && "LinearizeExpr failed!"); 313 314 // Move LHS right before I to make sure that the tree expression dominates all 315 // values. 316 LHSBO->moveBefore(I); 317 318 // Linearize the expression tree on the LHS. 319 LinearizeExprTree(LHSBO, Ops); 320 321 // Remember the RHS operand and its rank. 322 Ops.push_back(ValueEntry(getRank(RHS), RHS)); 323 324 // Clear the RHS leaf out. 325 I->setOperand(1, UndefValue::get(I->getType())); 326} 327 328// RewriteExprTree - Now that the operands for this expression tree are 329// linearized and optimized, emit them in-order. This function is written to be 330// tail recursive. 331void Reassociate::RewriteExprTree(BinaryOperator *I, 332 SmallVectorImpl<ValueEntry> &Ops, 333 unsigned i) { 334 if (i+2 == Ops.size()) { 335 if (I->getOperand(0) != Ops[i].Op || 336 I->getOperand(1) != Ops[i+1].Op) { 337 Value *OldLHS = I->getOperand(0); 338 DEBUG(errs() << "RA: " << *I << '\n'); 339 I->setOperand(0, Ops[i].Op); 340 I->setOperand(1, Ops[i+1].Op); 341 DEBUG(errs() << "TO: " << *I << '\n'); 342 MadeChange = true; 343 ++NumChanged; 344 345 // If we reassociated a tree to fewer operands (e.g. (1+a+2) -> (a+3) 346 // delete the extra, now dead, nodes. 347 RemoveDeadBinaryOp(OldLHS); 348 } 349 return; 350 } 351 assert(i+2 < Ops.size() && "Ops index out of range!"); 352 353 if (I->getOperand(1) != Ops[i].Op) { 354 DEBUG(errs() << "RA: " << *I << '\n'); 355 I->setOperand(1, Ops[i].Op); 356 DEBUG(errs() << "TO: " << *I << '\n'); 357 MadeChange = true; 358 ++NumChanged; 359 } 360 361 BinaryOperator *LHS = cast<BinaryOperator>(I->getOperand(0)); 362 assert(LHS->getOpcode() == I->getOpcode() && 363 "Improper expression tree!"); 364 365 // Compactify the tree instructions together with each other to guarantee 366 // that the expression tree is dominated by all of Ops. 367 LHS->moveBefore(I); 368 RewriteExprTree(LHS, Ops, i+1); 369} 370 371 372 373// NegateValue - Insert instructions before the instruction pointed to by BI, 374// that computes the negative version of the value specified. The negative 375// version of the value is returned, and BI is left pointing at the instruction 376// that should be processed next by the reassociation pass. 377// 378static Value *NegateValue(Value *V, Instruction *BI) { 379 if (Constant *C = dyn_cast<Constant>(V)) 380 return ConstantExpr::getNeg(C); 381 382 // We are trying to expose opportunity for reassociation. One of the things 383 // that we want to do to achieve this is to push a negation as deep into an 384 // expression chain as possible, to expose the add instructions. In practice, 385 // this means that we turn this: 386 // X = -(A+12+C+D) into X = -A + -12 + -C + -D = -12 + -A + -C + -D 387 // so that later, a: Y = 12+X could get reassociated with the -12 to eliminate 388 // the constants. We assume that instcombine will clean up the mess later if 389 // we introduce tons of unnecessary negation instructions... 390 // 391 if (Instruction *I = dyn_cast<Instruction>(V)) 392 if (I->getOpcode() == Instruction::Add && I->hasOneUse()) { 393 // Push the negates through the add. 394 I->setOperand(0, NegateValue(I->getOperand(0), BI)); 395 I->setOperand(1, NegateValue(I->getOperand(1), BI)); 396 397 // We must move the add instruction here, because the neg instructions do 398 // not dominate the old add instruction in general. By moving it, we are 399 // assured that the neg instructions we just inserted dominate the 400 // instruction we are about to insert after them. 401 // 402 I->moveBefore(BI); 403 I->setName(I->getName()+".neg"); 404 return I; 405 } 406 407 // Okay, we need to materialize a negated version of V with an instruction. 408 // Scan the use lists of V to see if we have one already. 409 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E;++UI){ 410 if (!BinaryOperator::isNeg(*UI)) continue; 411 412 // We found one! Now we have to make sure that the definition dominates 413 // this use. We do this by moving it to the entry block (if it is a 414 // non-instruction value) or right after the definition. These negates will 415 // be zapped by reassociate later, so we don't need much finesse here. 416 BinaryOperator *TheNeg = cast<BinaryOperator>(*UI); 417 418 BasicBlock::iterator InsertPt; 419 if (Instruction *InstInput = dyn_cast<Instruction>(V)) { 420 if (InvokeInst *II = dyn_cast<InvokeInst>(InstInput)) { 421 InsertPt = II->getNormalDest()->begin(); 422 } else { 423 InsertPt = InstInput; 424 ++InsertPt; 425 } 426 while (isa<PHINode>(InsertPt)) ++InsertPt; 427 } else { 428 InsertPt = TheNeg->getParent()->getParent()->getEntryBlock().begin(); 429 } 430 TheNeg->moveBefore(InsertPt); 431 return TheNeg; 432 } 433 434 // Insert a 'neg' instruction that subtracts the value from zero to get the 435 // negation. 436 return BinaryOperator::CreateNeg(V, V->getName() + ".neg", BI); 437} 438 439/// ShouldBreakUpSubtract - Return true if we should break up this subtract of 440/// X-Y into (X + -Y). 441static bool ShouldBreakUpSubtract(Instruction *Sub) { 442 // If this is a negation, we can't split it up! 443 if (BinaryOperator::isNeg(Sub)) 444 return false; 445 446 // Don't bother to break this up unless either the LHS is an associable add or 447 // subtract or if this is only used by one. 448 if (isReassociableOp(Sub->getOperand(0), Instruction::Add) || 449 isReassociableOp(Sub->getOperand(0), Instruction::Sub)) 450 return true; 451 if (isReassociableOp(Sub->getOperand(1), Instruction::Add) || 452 isReassociableOp(Sub->getOperand(1), Instruction::Sub)) 453 return true; 454 if (Sub->hasOneUse() && 455 (isReassociableOp(Sub->use_back(), Instruction::Add) || 456 isReassociableOp(Sub->use_back(), Instruction::Sub))) 457 return true; 458 459 return false; 460} 461 462/// BreakUpSubtract - If we have (X-Y), and if either X is an add, or if this is 463/// only used by an add, transform this into (X+(0-Y)) to promote better 464/// reassociation. 465static Instruction *BreakUpSubtract(Instruction *Sub, 466 std::map<AssertingVH<>, unsigned> &ValueRankMap) { 467 // Convert a subtract into an add and a neg instruction... so that sub 468 // instructions can be commuted with other add instructions... 469 // 470 // Calculate the negative value of Operand 1 of the sub instruction... 471 // and set it as the RHS of the add instruction we just made... 472 // 473 Value *NegVal = NegateValue(Sub->getOperand(1), Sub); 474 Instruction *New = 475 BinaryOperator::CreateAdd(Sub->getOperand(0), NegVal, "", Sub); 476 New->takeName(Sub); 477 478 // Everyone now refers to the add instruction. 479 ValueRankMap.erase(Sub); 480 Sub->replaceAllUsesWith(New); 481 Sub->eraseFromParent(); 482 483 DEBUG(errs() << "Negated: " << *New << '\n'); 484 return New; 485} 486 487/// ConvertShiftToMul - If this is a shift of a reassociable multiply or is used 488/// by one, change this into a multiply by a constant to assist with further 489/// reassociation. 490static Instruction *ConvertShiftToMul(Instruction *Shl, 491 std::map<AssertingVH<>, unsigned> &ValueRankMap) { 492 // If an operand of this shift is a reassociable multiply, or if the shift 493 // is used by a reassociable multiply or add, turn into a multiply. 494 if (isReassociableOp(Shl->getOperand(0), Instruction::Mul) || 495 (Shl->hasOneUse() && 496 (isReassociableOp(Shl->use_back(), Instruction::Mul) || 497 isReassociableOp(Shl->use_back(), Instruction::Add)))) { 498 Constant *MulCst = ConstantInt::get(Shl->getType(), 1); 499 MulCst = ConstantExpr::getShl(MulCst, cast<Constant>(Shl->getOperand(1))); 500 501 Instruction *Mul = 502 BinaryOperator::CreateMul(Shl->getOperand(0), MulCst, "", Shl); 503 ValueRankMap.erase(Shl); 504 Mul->takeName(Shl); 505 Shl->replaceAllUsesWith(Mul); 506 Shl->eraseFromParent(); 507 return Mul; 508 } 509 return 0; 510} 511 512// Scan backwards and forwards among values with the same rank as element i to 513// see if X exists. If X does not exist, return i. 514static unsigned FindInOperandList(SmallVectorImpl<ValueEntry> &Ops, unsigned i, 515 Value *X) { 516 unsigned XRank = Ops[i].Rank; 517 unsigned e = Ops.size(); 518 for (unsigned j = i+1; j != e && Ops[j].Rank == XRank; ++j) 519 if (Ops[j].Op == X) 520 return j; 521 // Scan backwards 522 for (unsigned j = i-1; j != ~0U && Ops[j].Rank == XRank; --j) 523 if (Ops[j].Op == X) 524 return j; 525 return i; 526} 527 528/// EmitAddTreeOfValues - Emit a tree of add instructions, summing Ops together 529/// and returning the result. Insert the tree before I. 530static Value *EmitAddTreeOfValues(Instruction *I, SmallVectorImpl<Value*> &Ops){ 531 if (Ops.size() == 1) return Ops.back(); 532 533 Value *V1 = Ops.back(); 534 Ops.pop_back(); 535 Value *V2 = EmitAddTreeOfValues(I, Ops); 536 return BinaryOperator::CreateAdd(V2, V1, "tmp", I); 537} 538 539/// RemoveFactorFromExpression - If V is an expression tree that is a 540/// multiplication sequence, and if this sequence contains a multiply by Factor, 541/// remove Factor from the tree and return the new tree. 542Value *Reassociate::RemoveFactorFromExpression(Value *V, Value *Factor) { 543 BinaryOperator *BO = isReassociableOp(V, Instruction::Mul); 544 if (!BO) return 0; 545 546 SmallVector<ValueEntry, 8> Factors; 547 LinearizeExprTree(BO, Factors); 548 549 bool FoundFactor = false; 550 for (unsigned i = 0, e = Factors.size(); i != e; ++i) 551 if (Factors[i].Op == Factor) { 552 FoundFactor = true; 553 Factors.erase(Factors.begin()+i); 554 break; 555 } 556 if (!FoundFactor) { 557 // Make sure to restore the operands to the expression tree. 558 RewriteExprTree(BO, Factors); 559 return 0; 560 } 561 562 // If this was just a single multiply, remove the multiply and return the only 563 // remaining operand. 564 if (Factors.size() == 1) { 565 ValueRankMap.erase(BO); 566 BO->eraseFromParent(); 567 return Factors[0].Op; 568 } 569 570 RewriteExprTree(BO, Factors); 571 return BO; 572} 573 574/// FindSingleUseMultiplyFactors - If V is a single-use multiply, recursively 575/// add its operands as factors, otherwise add V to the list of factors. 576static void FindSingleUseMultiplyFactors(Value *V, 577 SmallVectorImpl<Value*> &Factors) { 578 BinaryOperator *BO; 579 if ((!V->hasOneUse() && !V->use_empty()) || 580 !(BO = dyn_cast<BinaryOperator>(V)) || 581 BO->getOpcode() != Instruction::Mul) { 582 Factors.push_back(V); 583 return; 584 } 585 586 // Otherwise, add the LHS and RHS to the list of factors. 587 FindSingleUseMultiplyFactors(BO->getOperand(1), Factors); 588 FindSingleUseMultiplyFactors(BO->getOperand(0), Factors); 589} 590 591/// OptimizeAndOrXor - Optimize a series of operands to an 'and', 'or', or 'xor' 592/// instruction. This optimizes based on identities. If it can be reduced to 593/// a single Value, it is returned, otherwise the Ops list is mutated as 594/// necessary. 595static Value *OptimizeAndOrXor(unsigned Opcode, 596 SmallVectorImpl<ValueEntry> &Ops) { 597 // Scan the operand lists looking for X and ~X pairs, along with X,X pairs. 598 // If we find any, we can simplify the expression. X&~X == 0, X|~X == -1. 599 for (unsigned i = 0, e = Ops.size(); i != e; ++i) { 600 // First, check for X and ~X in the operand list. 601 assert(i < Ops.size()); 602 if (BinaryOperator::isNot(Ops[i].Op)) { // Cannot occur for ^. 603 Value *X = BinaryOperator::getNotArgument(Ops[i].Op); 604 unsigned FoundX = FindInOperandList(Ops, i, X); 605 if (FoundX != i) { 606 if (Opcode == Instruction::And) // ...&X&~X = 0 607 return Constant::getNullValue(X->getType()); 608 609 if (Opcode == Instruction::Or) // ...|X|~X = -1 610 return Constant::getAllOnesValue(X->getType()); 611 } 612 } 613 614 // Next, check for duplicate pairs of values, which we assume are next to 615 // each other, due to our sorting criteria. 616 assert(i < Ops.size()); 617 if (i+1 != Ops.size() && Ops[i+1].Op == Ops[i].Op) { 618 if (Opcode == Instruction::And || Opcode == Instruction::Or) { 619 // Drop duplicate values for And and Or. 620 Ops.erase(Ops.begin()+i); 621 --i; --e; 622 ++NumAnnihil; 623 continue; 624 } 625 626 // Drop pairs of values for Xor. 627 assert(Opcode == Instruction::Xor); 628 if (e == 2) 629 return Constant::getNullValue(Ops[0].Op->getType()); 630 631 // ... X^X -> ... 632 Ops.erase(Ops.begin()+i, Ops.begin()+i+2); 633 i -= 1; e -= 2; 634 ++NumAnnihil; 635 } 636 } 637 return 0; 638} 639 640/// OptimizeAdd - Optimize a series of operands to an 'add' instruction. This 641/// optimizes based on identities. If it can be reduced to a single Value, it 642/// is returned, otherwise the Ops list is mutated as necessary. 643Value *Reassociate::OptimizeAdd(Instruction *I, 644 SmallVectorImpl<ValueEntry> &Ops) { 645 // Scan the operand lists looking for X and -X pairs. If we find any, we 646 // can simplify the expression. X+-X == 0. While we're at it, scan for any 647 // duplicates. We want to canonicalize Y+Y+Y+Z -> 3*Y+Z. 648 for (unsigned i = 0, e = Ops.size(); i != e; ++i) { 649 Value *TheOp = Ops[i].Op; 650 // Check to see if we've seen this operand before. If so, we factor all 651 // instances of the operand together. Due to our sorting criteria, we know 652 // that these need to be next to each other in the vector. 653 if (i+1 != Ops.size() && Ops[i+1].Op == TheOp) { 654 // Rescan the list, remove all instances of this operand from the expr. 655 unsigned NumFound = 0; 656 do { 657 Ops.erase(Ops.begin()+i); 658 ++NumFound; 659 } while (i != Ops.size() && Ops[i].Op == TheOp); 660 661 DEBUG(errs() << "\nFACTORING [" << NumFound << "]: " << *TheOp << '\n'); 662 ++NumFactor; 663 664 // Insert a new multiply. 665 Value *Mul = ConstantInt::get(cast<IntegerType>(I->getType()), NumFound); 666 Mul = BinaryOperator::CreateMul(TheOp, Mul, "factor", I); 667 668 // Now that we have inserted a multiply, optimize it. This allows us to 669 // handle cases that require multiple factoring steps, such as this: 670 // (X*2) + (X*2) + (X*2) -> (X*2)*3 -> X*6 671 Mul = ReassociateExpression(cast<BinaryOperator>(Mul)); 672 673 // If every add operand was a duplicate, return the multiply. 674 if (Ops.empty()) 675 return Mul; 676 677 // Otherwise, we had some input that didn't have the dupe, such as 678 // "A + A + B" -> "A*2 + B". Add the new multiply to the list of 679 // things being added by this operation. 680 Ops.insert(Ops.begin(), ValueEntry(getRank(Mul), Mul)); 681 682 --i; 683 e = Ops.size(); 684 continue; 685 } 686 687 // Check for X and -X in the operand list. 688 if (!BinaryOperator::isNeg(TheOp)) 689 continue; 690 691 Value *X = BinaryOperator::getNegArgument(TheOp); 692 unsigned FoundX = FindInOperandList(Ops, i, X); 693 if (FoundX == i) 694 continue; 695 696 // Remove X and -X from the operand list. 697 if (Ops.size() == 2) 698 return Constant::getNullValue(X->getType()); 699 700 Ops.erase(Ops.begin()+i); 701 if (i < FoundX) 702 --FoundX; 703 else 704 --i; // Need to back up an extra one. 705 Ops.erase(Ops.begin()+FoundX); 706 ++NumAnnihil; 707 --i; // Revisit element. 708 e -= 2; // Removed two elements. 709 } 710 711 // Scan the operand list, checking to see if there are any common factors 712 // between operands. Consider something like A*A+A*B*C+D. We would like to 713 // reassociate this to A*(A+B*C)+D, which reduces the number of multiplies. 714 // To efficiently find this, we count the number of times a factor occurs 715 // for any ADD operands that are MULs. 716 DenseMap<Value*, unsigned> FactorOccurrences; 717 718 // Keep track of each multiply we see, to avoid triggering on (X*4)+(X*4) 719 // where they are actually the same multiply. 720 unsigned MaxOcc = 0; 721 Value *MaxOccVal = 0; 722 for (unsigned i = 0, e = Ops.size(); i != e; ++i) { 723 BinaryOperator *BOp = dyn_cast<BinaryOperator>(Ops[i].Op); 724 if (BOp == 0 || BOp->getOpcode() != Instruction::Mul || !BOp->use_empty()) 725 continue; 726 727 // Compute all of the factors of this added value. 728 SmallVector<Value*, 8> Factors; 729 FindSingleUseMultiplyFactors(BOp, Factors); 730 assert(Factors.size() > 1 && "Bad linearize!"); 731 732 // Add one to FactorOccurrences for each unique factor in this op. 733 if (Factors.size() == 2) { 734 unsigned Occ = ++FactorOccurrences[Factors[0]]; 735 if (Occ > MaxOcc) { MaxOcc = Occ; MaxOccVal = Factors[0]; } 736 if (Factors[0] != Factors[1]) { // Don't double count A*A. 737 Occ = ++FactorOccurrences[Factors[1]]; 738 if (Occ > MaxOcc) { MaxOcc = Occ; MaxOccVal = Factors[1]; } 739 } 740 } else { 741 SmallPtrSet<Value*, 4> Duplicates; 742 for (unsigned i = 0, e = Factors.size(); i != e; ++i) { 743 if (!Duplicates.insert(Factors[i])) continue; 744 745 unsigned Occ = ++FactorOccurrences[Factors[i]]; 746 if (Occ > MaxOcc) { MaxOcc = Occ; MaxOccVal = Factors[i]; } 747 } 748 } 749 } 750 751 // If any factor occurred more than one time, we can pull it out. 752 if (MaxOcc > 1) { 753 DEBUG(errs() << "\nFACTORING [" << MaxOcc << "]: " << *MaxOccVal << '\n'); 754 ++NumFactor; 755 756 // Create a new instruction that uses the MaxOccVal twice. If we don't do 757 // this, we could otherwise run into situations where removing a factor 758 // from an expression will drop a use of maxocc, and this can cause 759 // RemoveFactorFromExpression on successive values to behave differently. 760 Instruction *DummyInst = BinaryOperator::CreateAdd(MaxOccVal, MaxOccVal); 761 SmallVector<Value*, 4> NewMulOps; 762 for (unsigned i = 0, e = Ops.size(); i != e; ++i) { 763 if (Value *V = RemoveFactorFromExpression(Ops[i].Op, MaxOccVal)) { 764 NewMulOps.push_back(V); 765 Ops.erase(Ops.begin()+i); 766 --i; --e; 767 } 768 } 769 770 // No need for extra uses anymore. 771 delete DummyInst; 772 773 unsigned NumAddedValues = NewMulOps.size(); 774 Value *V = EmitAddTreeOfValues(I, NewMulOps); 775 776 // Now that we have inserted the add tree, optimize it. This allows us to 777 // handle cases that require multiple factoring steps, such as this: 778 // A*A*B + A*A*C --> A*(A*B+A*C) --> A*(A*(B+C)) 779 assert(NumAddedValues > 1 && "Each occurrence should contribute a value"); 780 V = ReassociateExpression(cast<BinaryOperator>(V)); 781 782 // Create the multiply. 783 Value *V2 = BinaryOperator::CreateMul(V, MaxOccVal, "tmp", I); 784 785 // Rerun associate on the multiply in case the inner expression turned into 786 // a multiply. We want to make sure that we keep things in canonical form. 787 V2 = ReassociateExpression(cast<BinaryOperator>(V2)); 788 789 // If every add operand included the factor (e.g. "A*B + A*C"), then the 790 // entire result expression is just the multiply "A*(B+C)". 791 if (Ops.empty()) 792 return V2; 793 794 // Otherwise, we had some input that didn't have the factor, such as 795 // "A*B + A*C + D" -> "A*(B+C) + D". Add the new multiply to the list of 796 // things being added by this operation. 797 Ops.insert(Ops.begin(), ValueEntry(getRank(V2), V2)); 798 } 799 800 return 0; 801} 802 803Value *Reassociate::OptimizeExpression(BinaryOperator *I, 804 SmallVectorImpl<ValueEntry> &Ops) { 805 // Now that we have the linearized expression tree, try to optimize it. 806 // Start by folding any constants that we found. 807 bool IterateOptimization = false; 808 if (Ops.size() == 1) return Ops[0].Op; 809 810 unsigned Opcode = I->getOpcode(); 811 812 if (Constant *V1 = dyn_cast<Constant>(Ops[Ops.size()-2].Op)) 813 if (Constant *V2 = dyn_cast<Constant>(Ops.back().Op)) { 814 Ops.pop_back(); 815 Ops.back().Op = ConstantExpr::get(Opcode, V1, V2); 816 return OptimizeExpression(I, Ops); 817 } 818 819 // Check for destructive annihilation due to a constant being used. 820 if (ConstantInt *CstVal = dyn_cast<ConstantInt>(Ops.back().Op)) 821 switch (Opcode) { 822 default: break; 823 case Instruction::And: 824 if (CstVal->isZero()) // ... & 0 -> 0 825 return CstVal; 826 if (CstVal->isAllOnesValue()) // ... & -1 -> ... 827 Ops.pop_back(); 828 break; 829 case Instruction::Mul: 830 if (CstVal->isZero()) { // ... * 0 -> 0 831 ++NumAnnihil; 832 return CstVal; 833 } 834 835 if (cast<ConstantInt>(CstVal)->isOne()) 836 Ops.pop_back(); // ... * 1 -> ... 837 break; 838 case Instruction::Or: 839 if (CstVal->isAllOnesValue()) // ... | -1 -> -1 840 return CstVal; 841 // FALLTHROUGH! 842 case Instruction::Add: 843 case Instruction::Xor: 844 if (CstVal->isZero()) // ... [|^+] 0 -> ... 845 Ops.pop_back(); 846 break; 847 } 848 if (Ops.size() == 1) return Ops[0].Op; 849 850 // Handle destructive annihilation due to identities between elements in the 851 // argument list here. 852 switch (Opcode) { 853 default: break; 854 case Instruction::And: 855 case Instruction::Or: 856 case Instruction::Xor: { 857 unsigned NumOps = Ops.size(); 858 if (Value *Result = OptimizeAndOrXor(Opcode, Ops)) 859 return Result; 860 IterateOptimization |= Ops.size() != NumOps; 861 break; 862 } 863 864 case Instruction::Add: { 865 unsigned NumOps = Ops.size(); 866 if (Value *Result = OptimizeAdd(I, Ops)) 867 return Result; 868 IterateOptimization |= Ops.size() != NumOps; 869 } 870 871 break; 872 //case Instruction::Mul: 873 } 874 875 if (IterateOptimization) 876 return OptimizeExpression(I, Ops); 877 return 0; 878} 879 880 881/// ReassociateBB - Inspect all of the instructions in this basic block, 882/// reassociating them as we go. 883void Reassociate::ReassociateBB(BasicBlock *BB) { 884 for (BasicBlock::iterator BBI = BB->begin(); BBI != BB->end(); ) { 885 Instruction *BI = BBI++; 886 if (BI->getOpcode() == Instruction::Shl && 887 isa<ConstantInt>(BI->getOperand(1))) 888 if (Instruction *NI = ConvertShiftToMul(BI, ValueRankMap)) { 889 MadeChange = true; 890 BI = NI; 891 } 892 893 // Reject cases where it is pointless to do this. 894 if (!isa<BinaryOperator>(BI) || BI->getType()->isFloatingPoint() || 895 isa<VectorType>(BI->getType())) 896 continue; // Floating point ops are not associative. 897 898 // If this is a subtract instruction which is not already in negate form, 899 // see if we can convert it to X+-Y. 900 if (BI->getOpcode() == Instruction::Sub) { 901 if (ShouldBreakUpSubtract(BI)) { 902 BI = BreakUpSubtract(BI, ValueRankMap); 903 MadeChange = true; 904 } else if (BinaryOperator::isNeg(BI)) { 905 // Otherwise, this is a negation. See if the operand is a multiply tree 906 // and if this is not an inner node of a multiply tree. 907 if (isReassociableOp(BI->getOperand(1), Instruction::Mul) && 908 (!BI->hasOneUse() || 909 !isReassociableOp(BI->use_back(), Instruction::Mul))) { 910 BI = LowerNegateToMultiply(BI, ValueRankMap); 911 MadeChange = true; 912 } 913 } 914 } 915 916 // If this instruction is a commutative binary operator, process it. 917 if (!BI->isAssociative()) continue; 918 BinaryOperator *I = cast<BinaryOperator>(BI); 919 920 // If this is an interior node of a reassociable tree, ignore it until we 921 // get to the root of the tree, to avoid N^2 analysis. 922 if (I->hasOneUse() && isReassociableOp(I->use_back(), I->getOpcode())) 923 continue; 924 925 // If this is an add tree that is used by a sub instruction, ignore it 926 // until we process the subtract. 927 if (I->hasOneUse() && I->getOpcode() == Instruction::Add && 928 cast<Instruction>(I->use_back())->getOpcode() == Instruction::Sub) 929 continue; 930 931 ReassociateExpression(I); 932 } 933} 934 935Value *Reassociate::ReassociateExpression(BinaryOperator *I) { 936 937 // First, walk the expression tree, linearizing the tree, collecting the 938 // operand information. 939 SmallVector<ValueEntry, 8> Ops; 940 LinearizeExprTree(I, Ops); 941 942 DEBUG(errs() << "RAIn:\t"; PrintOps(I, Ops); errs() << '\n'); 943 944 // Now that we have linearized the tree to a list and have gathered all of 945 // the operands and their ranks, sort the operands by their rank. Use a 946 // stable_sort so that values with equal ranks will have their relative 947 // positions maintained (and so the compiler is deterministic). Note that 948 // this sorts so that the highest ranking values end up at the beginning of 949 // the vector. 950 std::stable_sort(Ops.begin(), Ops.end()); 951 952 // OptimizeExpression - Now that we have the expression tree in a convenient 953 // sorted form, optimize it globally if possible. 954 if (Value *V = OptimizeExpression(I, Ops)) { 955 // This expression tree simplified to something that isn't a tree, 956 // eliminate it. 957 DEBUG(errs() << "Reassoc to scalar: " << *V << '\n'); 958 I->replaceAllUsesWith(V); 959 RemoveDeadBinaryOp(I); 960 ++NumAnnihil; 961 return V; 962 } 963 964 // We want to sink immediates as deeply as possible except in the case where 965 // this is a multiply tree used only by an add, and the immediate is a -1. 966 // In this case we reassociate to put the negation on the outside so that we 967 // can fold the negation into the add: (-X)*Y + Z -> Z-X*Y 968 if (I->getOpcode() == Instruction::Mul && I->hasOneUse() && 969 cast<Instruction>(I->use_back())->getOpcode() == Instruction::Add && 970 isa<ConstantInt>(Ops.back().Op) && 971 cast<ConstantInt>(Ops.back().Op)->isAllOnesValue()) { 972 ValueEntry Tmp = Ops.pop_back_val(); 973 Ops.insert(Ops.begin(), Tmp); 974 } 975 976 DEBUG(errs() << "RAOut:\t"; PrintOps(I, Ops); errs() << '\n'); 977 978 if (Ops.size() == 1) { 979 // This expression tree simplified to something that isn't a tree, 980 // eliminate it. 981 I->replaceAllUsesWith(Ops[0].Op); 982 RemoveDeadBinaryOp(I); 983 return Ops[0].Op; 984 } 985 986 // Now that we ordered and optimized the expressions, splat them back into 987 // the expression tree, removing any unneeded nodes. 988 RewriteExprTree(I, Ops); 989 return I; 990} 991 992 993bool Reassociate::runOnFunction(Function &F) { 994 // Recalculate the rank map for F 995 BuildRankMap(F); 996 997 MadeChange = false; 998 for (Function::iterator FI = F.begin(), FE = F.end(); FI != FE; ++FI) 999 ReassociateBB(FI); 1000 1001 // We are done with the rank map... 1002 RankMap.clear(); 1003 ValueRankMap.clear(); 1004 return MadeChange; 1005} 1006 1007