Reassociate.cpp revision 198396
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/LLVMContext.h" 31#include "llvm/Pass.h" 32#include "llvm/Assembly/Writer.h" 33#include "llvm/Support/CFG.h" 34#include "llvm/Support/Debug.h" 35#include "llvm/Support/ValueHandle.h" 36#include "llvm/Support/raw_ostream.h" 37#include "llvm/ADT/PostOrderIterator.h" 38#include "llvm/ADT/Statistic.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 std::vector<ValueEntry> &Ops) { 63 Module *M = I->getParent()->getParent()->getParent(); 64 errs() << Instruction::getOpcodeName(I->getOpcode()) << " " 65 << *Ops[0].Op->getType(); 66 for (unsigned i = 0, e = Ops.size(); i != e; ++i) { 67 WriteAsOperand(errs() << " ", Ops[i].Op, false, M); 68 errs() << "," << Ops[i].Rank; 69 } 70} 71#endif 72 73namespace { 74 class Reassociate : public FunctionPass { 75 std::map<BasicBlock*, unsigned> RankMap; 76 std::map<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 void ReassociateExpression(BinaryOperator *I); 91 void RewriteExprTree(BinaryOperator *I, std::vector<ValueEntry> &Ops, 92 unsigned Idx = 0); 93 Value *OptimizeExpression(BinaryOperator *I, std::vector<ValueEntry> &Ops); 94 void LinearizeExprTree(BinaryOperator *I, std::vector<ValueEntry> &Ops); 95 void LinearizeExpr(BinaryOperator *I); 96 Value *RemoveFactorFromExpression(Value *V, Value *Factor); 97 void ReassociateBB(BasicBlock *BB); 98 99 void RemoveDeadBinaryOp(Value *V); 100 }; 101} 102 103char Reassociate::ID = 0; 104static RegisterPass<Reassociate> X("reassociate", "Reassociate expressions"); 105 106// Public interface to the Reassociate pass 107FunctionPass *llvm::createReassociatePass() { return new Reassociate(); } 108 109void Reassociate::RemoveDeadBinaryOp(Value *V) { 110 Instruction *Op = dyn_cast<Instruction>(V); 111 if (!Op || !isa<BinaryOperator>(Op) || !isa<CmpInst>(Op) || !Op->use_empty()) 112 return; 113 114 Value *LHS = Op->getOperand(0), *RHS = Op->getOperand(1); 115 RemoveDeadBinaryOp(LHS); 116 RemoveDeadBinaryOp(RHS); 117} 118 119 120static bool isUnmovableInstruction(Instruction *I) { 121 if (I->getOpcode() == Instruction::PHI || 122 I->getOpcode() == Instruction::Alloca || 123 I->getOpcode() == Instruction::Load || 124 I->getOpcode() == Instruction::Invoke || 125 (I->getOpcode() == Instruction::Call && 126 !isa<DbgInfoIntrinsic>(I)) || 127 I->getOpcode() == Instruction::UDiv || 128 I->getOpcode() == Instruction::SDiv || 129 I->getOpcode() == Instruction::FDiv || 130 I->getOpcode() == Instruction::URem || 131 I->getOpcode() == Instruction::SRem || 132 I->getOpcode() == Instruction::FRem) 133 return true; 134 return false; 135} 136 137void Reassociate::BuildRankMap(Function &F) { 138 unsigned i = 2; 139 140 // Assign distinct ranks to function arguments 141 for (Function::arg_iterator I = F.arg_begin(), E = F.arg_end(); I != E; ++I) 142 ValueRankMap[&*I] = ++i; 143 144 ReversePostOrderTraversal<Function*> RPOT(&F); 145 for (ReversePostOrderTraversal<Function*>::rpo_iterator I = RPOT.begin(), 146 E = RPOT.end(); I != E; ++I) { 147 BasicBlock *BB = *I; 148 unsigned BBRank = RankMap[BB] = ++i << 16; 149 150 // Walk the basic block, adding precomputed ranks for any instructions that 151 // we cannot move. This ensures that the ranks for these instructions are 152 // all different in the block. 153 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) 154 if (isUnmovableInstruction(I)) 155 ValueRankMap[&*I] = ++BBRank; 156 } 157} 158 159unsigned Reassociate::getRank(Value *V) { 160 if (isa<Argument>(V)) return ValueRankMap[V]; // Function argument... 161 162 Instruction *I = dyn_cast<Instruction>(V); 163 if (I == 0) return 0; // Otherwise it's a global or constant, rank 0. 164 165 unsigned &CachedRank = ValueRankMap[I]; 166 if (CachedRank) return CachedRank; // Rank already known? 167 168 // If this is an expression, return the 1+MAX(rank(LHS), rank(RHS)) so that 169 // we can reassociate expressions for code motion! Since we do not recurse 170 // for PHI nodes, we cannot have infinite recursion here, because there 171 // cannot be loops in the value graph that do not go through PHI nodes. 172 unsigned Rank = 0, MaxRank = RankMap[I->getParent()]; 173 for (unsigned i = 0, e = I->getNumOperands(); 174 i != e && Rank != MaxRank; ++i) 175 Rank = std::max(Rank, getRank(I->getOperand(i))); 176 177 // If this is a not or neg instruction, do not count it for rank. This 178 // assures us that X and ~X will have the same rank. 179 if (!I->getType()->isInteger() || 180 (!BinaryOperator::isNot(I) && !BinaryOperator::isNeg(I))) 181 ++Rank; 182 183 //DEBUG(errs() << "Calculated Rank[" << V->getName() << "] = " 184 // << Rank << "\n"); 185 186 return CachedRank = Rank; 187} 188 189/// isReassociableOp - Return true if V is an instruction of the specified 190/// opcode and if it only has one use. 191static BinaryOperator *isReassociableOp(Value *V, unsigned Opcode) { 192 if ((V->hasOneUse() || V->use_empty()) && isa<Instruction>(V) && 193 cast<Instruction>(V)->getOpcode() == Opcode) 194 return cast<BinaryOperator>(V); 195 return 0; 196} 197 198/// LowerNegateToMultiply - Replace 0-X with X*-1. 199/// 200static Instruction *LowerNegateToMultiply(Instruction *Neg, 201 std::map<AssertingVH<>, unsigned> &ValueRankMap, 202 LLVMContext &Context) { 203 Constant *Cst = Constant::getAllOnesValue(Neg->getType()); 204 205 Instruction *Res = BinaryOperator::CreateMul(Neg->getOperand(1), Cst, "",Neg); 206 ValueRankMap.erase(Neg); 207 Res->takeName(Neg); 208 Neg->replaceAllUsesWith(Res); 209 Neg->eraseFromParent(); 210 return Res; 211} 212 213// Given an expression of the form '(A+B)+(D+C)', turn it into '(((A+B)+C)+D)'. 214// Note that if D is also part of the expression tree that we recurse to 215// linearize it as well. Besides that case, this does not recurse into A,B, or 216// C. 217void Reassociate::LinearizeExpr(BinaryOperator *I) { 218 BinaryOperator *LHS = cast<BinaryOperator>(I->getOperand(0)); 219 BinaryOperator *RHS = cast<BinaryOperator>(I->getOperand(1)); 220 assert(isReassociableOp(LHS, I->getOpcode()) && 221 isReassociableOp(RHS, I->getOpcode()) && 222 "Not an expression that needs linearization?"); 223 224 DEBUG(errs() << "Linear" << *LHS << '\n' << *RHS << '\n' << *I << '\n'); 225 226 // Move the RHS instruction to live immediately before I, avoiding breaking 227 // dominator properties. 228 RHS->moveBefore(I); 229 230 // Move operands around to do the linearization. 231 I->setOperand(1, RHS->getOperand(0)); 232 RHS->setOperand(0, LHS); 233 I->setOperand(0, RHS); 234 235 ++NumLinear; 236 MadeChange = true; 237 DEBUG(errs() << "Linearized: " << *I << '\n'); 238 239 // If D is part of this expression tree, tail recurse. 240 if (isReassociableOp(I->getOperand(1), I->getOpcode())) 241 LinearizeExpr(I); 242} 243 244 245/// LinearizeExprTree - Given an associative binary expression tree, traverse 246/// all of the uses putting it into canonical form. This forces a left-linear 247/// form of the the expression (((a+b)+c)+d), and collects information about the 248/// rank of the non-tree operands. 249/// 250/// NOTE: These intentionally destroys the expression tree operands (turning 251/// them into undef values) to reduce #uses of the values. This means that the 252/// caller MUST use something like RewriteExprTree to put the values back in. 253/// 254void Reassociate::LinearizeExprTree(BinaryOperator *I, 255 std::vector<ValueEntry> &Ops) { 256 Value *LHS = I->getOperand(0), *RHS = I->getOperand(1); 257 unsigned Opcode = I->getOpcode(); 258 LLVMContext &Context = I->getContext(); 259 260 // First step, linearize the expression if it is in ((A+B)+(C+D)) form. 261 BinaryOperator *LHSBO = isReassociableOp(LHS, Opcode); 262 BinaryOperator *RHSBO = isReassociableOp(RHS, Opcode); 263 264 // If this is a multiply expression tree and it contains internal negations, 265 // transform them into multiplies by -1 so they can be reassociated. 266 if (I->getOpcode() == Instruction::Mul) { 267 if (!LHSBO && LHS->hasOneUse() && BinaryOperator::isNeg(LHS)) { 268 LHS = LowerNegateToMultiply(cast<Instruction>(LHS), 269 ValueRankMap, Context); 270 LHSBO = isReassociableOp(LHS, Opcode); 271 } 272 if (!RHSBO && RHS->hasOneUse() && BinaryOperator::isNeg(RHS)) { 273 RHS = LowerNegateToMultiply(cast<Instruction>(RHS), 274 ValueRankMap, Context); 275 RHSBO = isReassociableOp(RHS, Opcode); 276 } 277 } 278 279 if (!LHSBO) { 280 if (!RHSBO) { 281 // Neither the LHS or RHS as part of the tree, thus this is a leaf. As 282 // such, just remember these operands and their rank. 283 Ops.push_back(ValueEntry(getRank(LHS), LHS)); 284 Ops.push_back(ValueEntry(getRank(RHS), RHS)); 285 286 // Clear the leaves out. 287 I->setOperand(0, UndefValue::get(I->getType())); 288 I->setOperand(1, UndefValue::get(I->getType())); 289 return; 290 } else { 291 // Turn X+(Y+Z) -> (Y+Z)+X 292 std::swap(LHSBO, RHSBO); 293 std::swap(LHS, RHS); 294 bool Success = !I->swapOperands(); 295 assert(Success && "swapOperands failed"); 296 Success = false; 297 MadeChange = true; 298 } 299 } else if (RHSBO) { 300 // Turn (A+B)+(C+D) -> (((A+B)+C)+D). This guarantees the the RHS is not 301 // part of the expression tree. 302 LinearizeExpr(I); 303 LHS = LHSBO = cast<BinaryOperator>(I->getOperand(0)); 304 RHS = I->getOperand(1); 305 RHSBO = 0; 306 } 307 308 // Okay, now we know that the LHS is a nested expression and that the RHS is 309 // not. Perform reassociation. 310 assert(!isReassociableOp(RHS, Opcode) && "LinearizeExpr failed!"); 311 312 // Move LHS right before I to make sure that the tree expression dominates all 313 // values. 314 LHSBO->moveBefore(I); 315 316 // Linearize the expression tree on the LHS. 317 LinearizeExprTree(LHSBO, Ops); 318 319 // Remember the RHS operand and its rank. 320 Ops.push_back(ValueEntry(getRank(RHS), RHS)); 321 322 // Clear the RHS leaf out. 323 I->setOperand(1, UndefValue::get(I->getType())); 324} 325 326// RewriteExprTree - Now that the operands for this expression tree are 327// linearized and optimized, emit them in-order. This function is written to be 328// tail recursive. 329void Reassociate::RewriteExprTree(BinaryOperator *I, 330 std::vector<ValueEntry> &Ops, 331 unsigned i) { 332 if (i+2 == Ops.size()) { 333 if (I->getOperand(0) != Ops[i].Op || 334 I->getOperand(1) != Ops[i+1].Op) { 335 Value *OldLHS = I->getOperand(0); 336 DEBUG(errs() << "RA: " << *I << '\n'); 337 I->setOperand(0, Ops[i].Op); 338 I->setOperand(1, Ops[i+1].Op); 339 DEBUG(errs() << "TO: " << *I << '\n'); 340 MadeChange = true; 341 ++NumChanged; 342 343 // If we reassociated a tree to fewer operands (e.g. (1+a+2) -> (a+3) 344 // delete the extra, now dead, nodes. 345 RemoveDeadBinaryOp(OldLHS); 346 } 347 return; 348 } 349 assert(i+2 < Ops.size() && "Ops index out of range!"); 350 351 if (I->getOperand(1) != Ops[i].Op) { 352 DEBUG(errs() << "RA: " << *I << '\n'); 353 I->setOperand(1, Ops[i].Op); 354 DEBUG(errs() << "TO: " << *I << '\n'); 355 MadeChange = true; 356 ++NumChanged; 357 } 358 359 BinaryOperator *LHS = cast<BinaryOperator>(I->getOperand(0)); 360 assert(LHS->getOpcode() == I->getOpcode() && 361 "Improper expression tree!"); 362 363 // Compactify the tree instructions together with each other to guarantee 364 // that the expression tree is dominated by all of Ops. 365 LHS->moveBefore(I); 366 RewriteExprTree(LHS, Ops, i+1); 367} 368 369 370 371// NegateValue - Insert instructions before the instruction pointed to by BI, 372// that computes the negative version of the value specified. The negative 373// version of the value is returned, and BI is left pointing at the instruction 374// that should be processed next by the reassociation pass. 375// 376static Value *NegateValue(LLVMContext &Context, Value *V, Instruction *BI) { 377 // We are trying to expose opportunity for reassociation. One of the things 378 // that we want to do to achieve this is to push a negation as deep into an 379 // expression chain as possible, to expose the add instructions. In practice, 380 // this means that we turn this: 381 // X = -(A+12+C+D) into X = -A + -12 + -C + -D = -12 + -A + -C + -D 382 // so that later, a: Y = 12+X could get reassociated with the -12 to eliminate 383 // the constants. We assume that instcombine will clean up the mess later if 384 // we introduce tons of unnecessary negation instructions... 385 // 386 if (Instruction *I = dyn_cast<Instruction>(V)) 387 if (I->getOpcode() == Instruction::Add && I->hasOneUse()) { 388 // Push the negates through the add. 389 I->setOperand(0, NegateValue(Context, I->getOperand(0), BI)); 390 I->setOperand(1, NegateValue(Context, I->getOperand(1), BI)); 391 392 // We must move the add instruction here, because the neg instructions do 393 // not dominate the old add instruction in general. By moving it, we are 394 // assured that the neg instructions we just inserted dominate the 395 // instruction we are about to insert after them. 396 // 397 I->moveBefore(BI); 398 I->setName(I->getName()+".neg"); 399 return I; 400 } 401 402 // Insert a 'neg' instruction that subtracts the value from zero to get the 403 // negation. 404 // 405 return BinaryOperator::CreateNeg(V, V->getName() + ".neg", BI); 406} 407 408/// ShouldBreakUpSubtract - Return true if we should break up this subtract of 409/// X-Y into (X + -Y). 410static bool ShouldBreakUpSubtract(LLVMContext &Context, Instruction *Sub) { 411 // If this is a negation, we can't split it up! 412 if (BinaryOperator::isNeg(Sub)) 413 return false; 414 415 // Don't bother to break this up unless either the LHS is an associable add or 416 // subtract or if this is only used by one. 417 if (isReassociableOp(Sub->getOperand(0), Instruction::Add) || 418 isReassociableOp(Sub->getOperand(0), Instruction::Sub)) 419 return true; 420 if (isReassociableOp(Sub->getOperand(1), Instruction::Add) || 421 isReassociableOp(Sub->getOperand(1), Instruction::Sub)) 422 return true; 423 if (Sub->hasOneUse() && 424 (isReassociableOp(Sub->use_back(), Instruction::Add) || 425 isReassociableOp(Sub->use_back(), Instruction::Sub))) 426 return true; 427 428 return false; 429} 430 431/// BreakUpSubtract - If we have (X-Y), and if either X is an add, or if this is 432/// only used by an add, transform this into (X+(0-Y)) to promote better 433/// reassociation. 434static Instruction *BreakUpSubtract(LLVMContext &Context, Instruction *Sub, 435 std::map<AssertingVH<>, unsigned> &ValueRankMap) { 436 // Convert a subtract into an add and a neg instruction... so that sub 437 // instructions can be commuted with other add instructions... 438 // 439 // Calculate the negative value of Operand 1 of the sub instruction... 440 // and set it as the RHS of the add instruction we just made... 441 // 442 Value *NegVal = NegateValue(Context, Sub->getOperand(1), Sub); 443 Instruction *New = 444 BinaryOperator::CreateAdd(Sub->getOperand(0), NegVal, "", Sub); 445 New->takeName(Sub); 446 447 // Everyone now refers to the add instruction. 448 ValueRankMap.erase(Sub); 449 Sub->replaceAllUsesWith(New); 450 Sub->eraseFromParent(); 451 452 DEBUG(errs() << "Negated: " << *New << '\n'); 453 return New; 454} 455 456/// ConvertShiftToMul - If this is a shift of a reassociable multiply or is used 457/// by one, change this into a multiply by a constant to assist with further 458/// reassociation. 459static Instruction *ConvertShiftToMul(Instruction *Shl, 460 std::map<AssertingVH<>, unsigned> &ValueRankMap, 461 LLVMContext &Context) { 462 // If an operand of this shift is a reassociable multiply, or if the shift 463 // is used by a reassociable multiply or add, turn into a multiply. 464 if (isReassociableOp(Shl->getOperand(0), Instruction::Mul) || 465 (Shl->hasOneUse() && 466 (isReassociableOp(Shl->use_back(), Instruction::Mul) || 467 isReassociableOp(Shl->use_back(), Instruction::Add)))) { 468 Constant *MulCst = ConstantInt::get(Shl->getType(), 1); 469 MulCst = 470 ConstantExpr::getShl(MulCst, cast<Constant>(Shl->getOperand(1))); 471 472 Instruction *Mul = BinaryOperator::CreateMul(Shl->getOperand(0), MulCst, 473 "", Shl); 474 ValueRankMap.erase(Shl); 475 Mul->takeName(Shl); 476 Shl->replaceAllUsesWith(Mul); 477 Shl->eraseFromParent(); 478 return Mul; 479 } 480 return 0; 481} 482 483// Scan backwards and forwards among values with the same rank as element i to 484// see if X exists. If X does not exist, return i. 485static unsigned FindInOperandList(std::vector<ValueEntry> &Ops, unsigned i, 486 Value *X) { 487 unsigned XRank = Ops[i].Rank; 488 unsigned e = Ops.size(); 489 for (unsigned j = i+1; j != e && Ops[j].Rank == XRank; ++j) 490 if (Ops[j].Op == X) 491 return j; 492 // Scan backwards 493 for (unsigned j = i-1; j != ~0U && Ops[j].Rank == XRank; --j) 494 if (Ops[j].Op == X) 495 return j; 496 return i; 497} 498 499/// EmitAddTreeOfValues - Emit a tree of add instructions, summing Ops together 500/// and returning the result. Insert the tree before I. 501static Value *EmitAddTreeOfValues(Instruction *I, std::vector<Value*> &Ops) { 502 if (Ops.size() == 1) return Ops.back(); 503 504 Value *V1 = Ops.back(); 505 Ops.pop_back(); 506 Value *V2 = EmitAddTreeOfValues(I, Ops); 507 return BinaryOperator::CreateAdd(V2, V1, "tmp", I); 508} 509 510/// RemoveFactorFromExpression - If V is an expression tree that is a 511/// multiplication sequence, and if this sequence contains a multiply by Factor, 512/// remove Factor from the tree and return the new tree. 513Value *Reassociate::RemoveFactorFromExpression(Value *V, Value *Factor) { 514 BinaryOperator *BO = isReassociableOp(V, Instruction::Mul); 515 if (!BO) return 0; 516 517 std::vector<ValueEntry> Factors; 518 LinearizeExprTree(BO, Factors); 519 520 bool FoundFactor = false; 521 for (unsigned i = 0, e = Factors.size(); i != e; ++i) 522 if (Factors[i].Op == Factor) { 523 FoundFactor = true; 524 Factors.erase(Factors.begin()+i); 525 break; 526 } 527 if (!FoundFactor) { 528 // Make sure to restore the operands to the expression tree. 529 RewriteExprTree(BO, Factors); 530 return 0; 531 } 532 533 if (Factors.size() == 1) return Factors[0].Op; 534 535 RewriteExprTree(BO, Factors); 536 return BO; 537} 538 539/// FindSingleUseMultiplyFactors - If V is a single-use multiply, recursively 540/// add its operands as factors, otherwise add V to the list of factors. 541static void FindSingleUseMultiplyFactors(Value *V, 542 std::vector<Value*> &Factors) { 543 BinaryOperator *BO; 544 if ((!V->hasOneUse() && !V->use_empty()) || 545 !(BO = dyn_cast<BinaryOperator>(V)) || 546 BO->getOpcode() != Instruction::Mul) { 547 Factors.push_back(V); 548 return; 549 } 550 551 // Otherwise, add the LHS and RHS to the list of factors. 552 FindSingleUseMultiplyFactors(BO->getOperand(1), Factors); 553 FindSingleUseMultiplyFactors(BO->getOperand(0), Factors); 554} 555 556 557 558Value *Reassociate::OptimizeExpression(BinaryOperator *I, 559 std::vector<ValueEntry> &Ops) { 560 // Now that we have the linearized expression tree, try to optimize it. 561 // Start by folding any constants that we found. 562 bool IterateOptimization = false; 563 if (Ops.size() == 1) return Ops[0].Op; 564 565 unsigned Opcode = I->getOpcode(); 566 567 if (Constant *V1 = dyn_cast<Constant>(Ops[Ops.size()-2].Op)) 568 if (Constant *V2 = dyn_cast<Constant>(Ops.back().Op)) { 569 Ops.pop_back(); 570 Ops.back().Op = ConstantExpr::get(Opcode, V1, V2); 571 return OptimizeExpression(I, Ops); 572 } 573 574 // Check for destructive annihilation due to a constant being used. 575 if (ConstantInt *CstVal = dyn_cast<ConstantInt>(Ops.back().Op)) 576 switch (Opcode) { 577 default: break; 578 case Instruction::And: 579 if (CstVal->isZero()) { // ... & 0 -> 0 580 ++NumAnnihil; 581 return CstVal; 582 } else if (CstVal->isAllOnesValue()) { // ... & -1 -> ... 583 Ops.pop_back(); 584 } 585 break; 586 case Instruction::Mul: 587 if (CstVal->isZero()) { // ... * 0 -> 0 588 ++NumAnnihil; 589 return CstVal; 590 } else if (cast<ConstantInt>(CstVal)->isOne()) { 591 Ops.pop_back(); // ... * 1 -> ... 592 } 593 break; 594 case Instruction::Or: 595 if (CstVal->isAllOnesValue()) { // ... | -1 -> -1 596 ++NumAnnihil; 597 return CstVal; 598 } 599 // FALLTHROUGH! 600 case Instruction::Add: 601 case Instruction::Xor: 602 if (CstVal->isZero()) // ... [|^+] 0 -> ... 603 Ops.pop_back(); 604 break; 605 } 606 if (Ops.size() == 1) return Ops[0].Op; 607 608 // Handle destructive annihilation do to identities between elements in the 609 // argument list here. 610 switch (Opcode) { 611 default: break; 612 case Instruction::And: 613 case Instruction::Or: 614 case Instruction::Xor: 615 // Scan the operand lists looking for X and ~X pairs, along with X,X pairs. 616 // If we find any, we can simplify the expression. X&~X == 0, X|~X == -1. 617 for (unsigned i = 0, e = Ops.size(); i != e; ++i) { 618 // First, check for X and ~X in the operand list. 619 assert(i < Ops.size()); 620 if (BinaryOperator::isNot(Ops[i].Op)) { // Cannot occur for ^. 621 Value *X = BinaryOperator::getNotArgument(Ops[i].Op); 622 unsigned FoundX = FindInOperandList(Ops, i, X); 623 if (FoundX != i) { 624 if (Opcode == Instruction::And) { // ...&X&~X = 0 625 ++NumAnnihil; 626 return Constant::getNullValue(X->getType()); 627 } else if (Opcode == Instruction::Or) { // ...|X|~X = -1 628 ++NumAnnihil; 629 return Constant::getAllOnesValue(X->getType()); 630 } 631 } 632 } 633 634 // Next, check for duplicate pairs of values, which we assume are next to 635 // each other, due to our sorting criteria. 636 assert(i < Ops.size()); 637 if (i+1 != Ops.size() && Ops[i+1].Op == Ops[i].Op) { 638 if (Opcode == Instruction::And || Opcode == Instruction::Or) { 639 // Drop duplicate values. 640 Ops.erase(Ops.begin()+i); 641 --i; --e; 642 IterateOptimization = true; 643 ++NumAnnihil; 644 } else { 645 assert(Opcode == Instruction::Xor); 646 if (e == 2) { 647 ++NumAnnihil; 648 return Constant::getNullValue(Ops[0].Op->getType()); 649 } 650 // ... X^X -> ... 651 Ops.erase(Ops.begin()+i, Ops.begin()+i+2); 652 i -= 1; e -= 2; 653 IterateOptimization = true; 654 ++NumAnnihil; 655 } 656 } 657 } 658 break; 659 660 case Instruction::Add: 661 // Scan the operand lists looking for X and -X pairs. If we find any, we 662 // can simplify the expression. X+-X == 0. 663 for (unsigned i = 0, e = Ops.size(); i != e; ++i) { 664 assert(i < Ops.size()); 665 // Check for X and -X in the operand list. 666 if (BinaryOperator::isNeg(Ops[i].Op)) { 667 Value *X = BinaryOperator::getNegArgument(Ops[i].Op); 668 unsigned FoundX = FindInOperandList(Ops, i, X); 669 if (FoundX != i) { 670 // Remove X and -X from the operand list. 671 if (Ops.size() == 2) { 672 ++NumAnnihil; 673 return Constant::getNullValue(X->getType()); 674 } else { 675 Ops.erase(Ops.begin()+i); 676 if (i < FoundX) 677 --FoundX; 678 else 679 --i; // Need to back up an extra one. 680 Ops.erase(Ops.begin()+FoundX); 681 IterateOptimization = true; 682 ++NumAnnihil; 683 --i; // Revisit element. 684 e -= 2; // Removed two elements. 685 } 686 } 687 } 688 } 689 690 691 // Scan the operand list, checking to see if there are any common factors 692 // between operands. Consider something like A*A+A*B*C+D. We would like to 693 // reassociate this to A*(A+B*C)+D, which reduces the number of multiplies. 694 // To efficiently find this, we count the number of times a factor occurs 695 // for any ADD operands that are MULs. 696 std::map<Value*, unsigned> FactorOccurrences; 697 unsigned MaxOcc = 0; 698 Value *MaxOccVal = 0; 699 for (unsigned i = 0, e = Ops.size(); i != e; ++i) { 700 if (BinaryOperator *BOp = dyn_cast<BinaryOperator>(Ops[i].Op)) { 701 if (BOp->getOpcode() == Instruction::Mul && BOp->use_empty()) { 702 // Compute all of the factors of this added value. 703 std::vector<Value*> Factors; 704 FindSingleUseMultiplyFactors(BOp, Factors); 705 assert(Factors.size() > 1 && "Bad linearize!"); 706 707 // Add one to FactorOccurrences for each unique factor in this op. 708 if (Factors.size() == 2) { 709 unsigned Occ = ++FactorOccurrences[Factors[0]]; 710 if (Occ > MaxOcc) { MaxOcc = Occ; MaxOccVal = Factors[0]; } 711 if (Factors[0] != Factors[1]) { // Don't double count A*A. 712 Occ = ++FactorOccurrences[Factors[1]]; 713 if (Occ > MaxOcc) { MaxOcc = Occ; MaxOccVal = Factors[1]; } 714 } 715 } else { 716 std::set<Value*> Duplicates; 717 for (unsigned i = 0, e = Factors.size(); i != e; ++i) { 718 if (Duplicates.insert(Factors[i]).second) { 719 unsigned Occ = ++FactorOccurrences[Factors[i]]; 720 if (Occ > MaxOcc) { MaxOcc = Occ; MaxOccVal = Factors[i]; } 721 } 722 } 723 } 724 } 725 } 726 } 727 728 // If any factor occurred more than one time, we can pull it out. 729 if (MaxOcc > 1) { 730 DEBUG(errs() << "\nFACTORING [" << MaxOcc << "]: " << *MaxOccVal << "\n"); 731 732 // Create a new instruction that uses the MaxOccVal twice. If we don't do 733 // this, we could otherwise run into situations where removing a factor 734 // from an expression will drop a use of maxocc, and this can cause 735 // RemoveFactorFromExpression on successive values to behave differently. 736 Instruction *DummyInst = BinaryOperator::CreateAdd(MaxOccVal, MaxOccVal); 737 std::vector<Value*> NewMulOps; 738 for (unsigned i = 0, e = Ops.size(); i != e; ++i) { 739 if (Value *V = RemoveFactorFromExpression(Ops[i].Op, MaxOccVal)) { 740 NewMulOps.push_back(V); 741 Ops.erase(Ops.begin()+i); 742 --i; --e; 743 } 744 } 745 746 // No need for extra uses anymore. 747 delete DummyInst; 748 749 unsigned NumAddedValues = NewMulOps.size(); 750 Value *V = EmitAddTreeOfValues(I, NewMulOps); 751 Value *V2 = BinaryOperator::CreateMul(V, MaxOccVal, "tmp", I); 752 753 // Now that we have inserted V and its sole use, optimize it. This allows 754 // us to handle cases that require multiple factoring steps, such as this: 755 // A*A*B + A*A*C --> A*(A*B+A*C) --> A*(A*(B+C)) 756 if (NumAddedValues > 1) 757 ReassociateExpression(cast<BinaryOperator>(V)); 758 759 ++NumFactor; 760 761 if (Ops.empty()) 762 return V2; 763 764 // Add the new value to the list of things being added. 765 Ops.insert(Ops.begin(), ValueEntry(getRank(V2), V2)); 766 767 // Rewrite the tree so that there is now a use of V. 768 RewriteExprTree(I, Ops); 769 return OptimizeExpression(I, Ops); 770 } 771 break; 772 //case Instruction::Mul: 773 } 774 775 if (IterateOptimization) 776 return OptimizeExpression(I, Ops); 777 return 0; 778} 779 780 781/// ReassociateBB - Inspect all of the instructions in this basic block, 782/// reassociating them as we go. 783void Reassociate::ReassociateBB(BasicBlock *BB) { 784 LLVMContext &Context = BB->getContext(); 785 786 for (BasicBlock::iterator BBI = BB->begin(); BBI != BB->end(); ) { 787 Instruction *BI = BBI++; 788 if (BI->getOpcode() == Instruction::Shl && 789 isa<ConstantInt>(BI->getOperand(1))) 790 if (Instruction *NI = ConvertShiftToMul(BI, ValueRankMap, Context)) { 791 MadeChange = true; 792 BI = NI; 793 } 794 795 // Reject cases where it is pointless to do this. 796 if (!isa<BinaryOperator>(BI) || BI->getType()->isFloatingPoint() || 797 isa<VectorType>(BI->getType())) 798 continue; // Floating point ops are not associative. 799 800 // If this is a subtract instruction which is not already in negate form, 801 // see if we can convert it to X+-Y. 802 if (BI->getOpcode() == Instruction::Sub) { 803 if (ShouldBreakUpSubtract(Context, BI)) { 804 BI = BreakUpSubtract(Context, BI, ValueRankMap); 805 MadeChange = true; 806 } else if (BinaryOperator::isNeg(BI)) { 807 // Otherwise, this is a negation. See if the operand is a multiply tree 808 // and if this is not an inner node of a multiply tree. 809 if (isReassociableOp(BI->getOperand(1), Instruction::Mul) && 810 (!BI->hasOneUse() || 811 !isReassociableOp(BI->use_back(), Instruction::Mul))) { 812 BI = LowerNegateToMultiply(BI, ValueRankMap, Context); 813 MadeChange = true; 814 } 815 } 816 } 817 818 // If this instruction is a commutative binary operator, process it. 819 if (!BI->isAssociative()) continue; 820 BinaryOperator *I = cast<BinaryOperator>(BI); 821 822 // If this is an interior node of a reassociable tree, ignore it until we 823 // get to the root of the tree, to avoid N^2 analysis. 824 if (I->hasOneUse() && isReassociableOp(I->use_back(), I->getOpcode())) 825 continue; 826 827 // If this is an add tree that is used by a sub instruction, ignore it 828 // until we process the subtract. 829 if (I->hasOneUse() && I->getOpcode() == Instruction::Add && 830 cast<Instruction>(I->use_back())->getOpcode() == Instruction::Sub) 831 continue; 832 833 ReassociateExpression(I); 834 } 835} 836 837void Reassociate::ReassociateExpression(BinaryOperator *I) { 838 839 // First, walk the expression tree, linearizing the tree, collecting 840 std::vector<ValueEntry> Ops; 841 LinearizeExprTree(I, Ops); 842 843 DEBUG(errs() << "RAIn:\t"; PrintOps(I, Ops); errs() << "\n"); 844 845 // Now that we have linearized the tree to a list and have gathered all of 846 // the operands and their ranks, sort the operands by their rank. Use a 847 // stable_sort so that values with equal ranks will have their relative 848 // positions maintained (and so the compiler is deterministic). Note that 849 // this sorts so that the highest ranking values end up at the beginning of 850 // the vector. 851 std::stable_sort(Ops.begin(), Ops.end()); 852 853 // OptimizeExpression - Now that we have the expression tree in a convenient 854 // sorted form, optimize it globally if possible. 855 if (Value *V = OptimizeExpression(I, Ops)) { 856 // This expression tree simplified to something that isn't a tree, 857 // eliminate it. 858 DEBUG(errs() << "Reassoc to scalar: " << *V << "\n"); 859 I->replaceAllUsesWith(V); 860 RemoveDeadBinaryOp(I); 861 return; 862 } 863 864 // We want to sink immediates as deeply as possible except in the case where 865 // this is a multiply tree used only by an add, and the immediate is a -1. 866 // In this case we reassociate to put the negation on the outside so that we 867 // can fold the negation into the add: (-X)*Y + Z -> Z-X*Y 868 if (I->getOpcode() == Instruction::Mul && I->hasOneUse() && 869 cast<Instruction>(I->use_back())->getOpcode() == Instruction::Add && 870 isa<ConstantInt>(Ops.back().Op) && 871 cast<ConstantInt>(Ops.back().Op)->isAllOnesValue()) { 872 Ops.insert(Ops.begin(), Ops.back()); 873 Ops.pop_back(); 874 } 875 876 DEBUG(errs() << "RAOut:\t"; PrintOps(I, Ops); errs() << "\n"); 877 878 if (Ops.size() == 1) { 879 // This expression tree simplified to something that isn't a tree, 880 // eliminate it. 881 I->replaceAllUsesWith(Ops[0].Op); 882 RemoveDeadBinaryOp(I); 883 } else { 884 // Now that we ordered and optimized the expressions, splat them back into 885 // the expression tree, removing any unneeded nodes. 886 RewriteExprTree(I, Ops); 887 } 888} 889 890 891bool Reassociate::runOnFunction(Function &F) { 892 // Recalculate the rank map for F 893 BuildRankMap(F); 894 895 MadeChange = false; 896 for (Function::iterator FI = F.begin(), FE = F.end(); FI != FE; ++FI) 897 ReassociateBB(FI); 898 899 // We are done with the rank map... 900 RankMap.clear(); 901 ValueRankMap.clear(); 902 return MadeChange; 903} 904 905