Reassociate.cpp revision a276c603b82a11b0bf0b59f0517a69e4b63adeab
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/Analysis/MallocHelper.h" 33#include "llvm/Assembly/Writer.h" 34#include "llvm/Support/CFG.h" 35#include "llvm/Support/Debug.h" 36#include "llvm/Support/ValueHandle.h" 37#include "llvm/Support/raw_ostream.h" 38#include "llvm/ADT/PostOrderIterator.h" 39#include "llvm/ADT/Statistic.h" 40#include <algorithm> 41#include <map> 42using namespace llvm; 43 44STATISTIC(NumLinear , "Number of insts linearized"); 45STATISTIC(NumChanged, "Number of insts reassociated"); 46STATISTIC(NumAnnihil, "Number of expr tree annihilated"); 47STATISTIC(NumFactor , "Number of multiplies factored"); 48 49namespace { 50 struct ValueEntry { 51 unsigned Rank; 52 Value *Op; 53 ValueEntry(unsigned R, Value *O) : Rank(R), Op(O) {} 54 }; 55 inline bool operator<(const ValueEntry &LHS, const ValueEntry &RHS) { 56 return LHS.Rank > RHS.Rank; // Sort so that highest rank goes to start. 57 } 58} 59 60#ifndef NDEBUG 61/// PrintOps - Print out the expression identified in the Ops list. 62/// 63static void PrintOps(Instruction *I, const std::vector<ValueEntry> &Ops) { 64 Module *M = I->getParent()->getParent()->getParent(); 65 errs() << Instruction::getOpcodeName(I->getOpcode()) << " " 66 << *Ops[0].Op->getType(); 67 for (unsigned i = 0, e = Ops.size(); i != e; ++i) { 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 void ReassociateExpression(BinaryOperator *I); 92 void RewriteExprTree(BinaryOperator *I, std::vector<ValueEntry> &Ops, 93 unsigned Idx = 0); 94 Value *OptimizeExpression(BinaryOperator *I, std::vector<ValueEntry> &Ops); 95 void LinearizeExprTree(BinaryOperator *I, std::vector<ValueEntry> &Ops); 96 void LinearizeExpr(BinaryOperator *I); 97 Value *RemoveFactorFromExpression(Value *V, Value *Factor); 98 void ReassociateBB(BasicBlock *BB); 99 100 void RemoveDeadBinaryOp(Value *V); 101 }; 102} 103 104char Reassociate::ID = 0; 105static RegisterPass<Reassociate> X("reassociate", "Reassociate expressions"); 106 107// Public interface to the Reassociate pass 108FunctionPass *llvm::createReassociatePass() { return new Reassociate(); } 109 110void Reassociate::RemoveDeadBinaryOp(Value *V) { 111 Instruction *Op = dyn_cast<Instruction>(V); 112 if (!Op || !isa<BinaryOperator>(Op) || !isa<CmpInst>(Op) || !Op->use_empty()) 113 return; 114 115 Value *LHS = Op->getOperand(0), *RHS = Op->getOperand(1); 116 RemoveDeadBinaryOp(LHS); 117 RemoveDeadBinaryOp(RHS); 118} 119 120 121static bool isUnmovableInstruction(Instruction *I) { 122 if (I->getOpcode() == Instruction::PHI || 123 I->getOpcode() == Instruction::Alloca || 124 I->getOpcode() == Instruction::Load || 125 isMalloc(I) || 126 I->getOpcode() == Instruction::Invoke || 127 (I->getOpcode() == Instruction::Call && 128 !isa<DbgInfoIntrinsic>(I)) || 129 I->getOpcode() == Instruction::UDiv || 130 I->getOpcode() == Instruction::SDiv || 131 I->getOpcode() == Instruction::FDiv || 132 I->getOpcode() == Instruction::URem || 133 I->getOpcode() == Instruction::SRem || 134 I->getOpcode() == Instruction::FRem) 135 return true; 136 return false; 137} 138 139void Reassociate::BuildRankMap(Function &F) { 140 unsigned i = 2; 141 142 // Assign distinct ranks to function arguments 143 for (Function::arg_iterator I = F.arg_begin(), E = F.arg_end(); I != E; ++I) 144 ValueRankMap[&*I] = ++i; 145 146 ReversePostOrderTraversal<Function*> RPOT(&F); 147 for (ReversePostOrderTraversal<Function*>::rpo_iterator I = RPOT.begin(), 148 E = RPOT.end(); I != E; ++I) { 149 BasicBlock *BB = *I; 150 unsigned BBRank = RankMap[BB] = ++i << 16; 151 152 // Walk the basic block, adding precomputed ranks for any instructions that 153 // we cannot move. This ensures that the ranks for these instructions are 154 // all different in the block. 155 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) 156 if (isUnmovableInstruction(I)) 157 ValueRankMap[&*I] = ++BBRank; 158 } 159} 160 161unsigned Reassociate::getRank(Value *V) { 162 if (isa<Argument>(V)) return ValueRankMap[V]; // Function argument... 163 164 Instruction *I = dyn_cast<Instruction>(V); 165 if (I == 0) return 0; // Otherwise it's a global or constant, rank 0. 166 167 unsigned &CachedRank = ValueRankMap[I]; 168 if (CachedRank) return CachedRank; // Rank already known? 169 170 // If this is an expression, return the 1+MAX(rank(LHS), rank(RHS)) so that 171 // we can reassociate expressions for code motion! Since we do not recurse 172 // for PHI nodes, we cannot have infinite recursion here, because there 173 // cannot be loops in the value graph that do not go through PHI nodes. 174 unsigned Rank = 0, MaxRank = RankMap[I->getParent()]; 175 for (unsigned i = 0, e = I->getNumOperands(); 176 i != e && Rank != MaxRank; ++i) 177 Rank = std::max(Rank, getRank(I->getOperand(i))); 178 179 // If this is a not or neg instruction, do not count it for rank. This 180 // assures us that X and ~X will have the same rank. 181 if (!I->getType()->isInteger() || 182 (!BinaryOperator::isNot(I) && !BinaryOperator::isNeg(I))) 183 ++Rank; 184 185 //DEBUG(errs() << "Calculated Rank[" << V->getName() << "] = " 186 // << Rank << "\n"); 187 188 return CachedRank = Rank; 189} 190 191/// isReassociableOp - Return true if V is an instruction of the specified 192/// opcode and if it only has one use. 193static BinaryOperator *isReassociableOp(Value *V, unsigned Opcode) { 194 if ((V->hasOneUse() || V->use_empty()) && isa<Instruction>(V) && 195 cast<Instruction>(V)->getOpcode() == Opcode) 196 return cast<BinaryOperator>(V); 197 return 0; 198} 199 200/// LowerNegateToMultiply - Replace 0-X with X*-1. 201/// 202static Instruction *LowerNegateToMultiply(Instruction *Neg, 203 std::map<AssertingVH<>, unsigned> &ValueRankMap, 204 LLVMContext &Context) { 205 Constant *Cst = Constant::getAllOnesValue(Neg->getType()); 206 207 Instruction *Res = BinaryOperator::CreateMul(Neg->getOperand(1), Cst, "",Neg); 208 ValueRankMap.erase(Neg); 209 Res->takeName(Neg); 210 Neg->replaceAllUsesWith(Res); 211 Neg->eraseFromParent(); 212 return Res; 213} 214 215// Given an expression of the form '(A+B)+(D+C)', turn it into '(((A+B)+C)+D)'. 216// Note that if D is also part of the expression tree that we recurse to 217// linearize it as well. Besides that case, this does not recurse into A,B, or 218// C. 219void Reassociate::LinearizeExpr(BinaryOperator *I) { 220 BinaryOperator *LHS = cast<BinaryOperator>(I->getOperand(0)); 221 BinaryOperator *RHS = cast<BinaryOperator>(I->getOperand(1)); 222 assert(isReassociableOp(LHS, I->getOpcode()) && 223 isReassociableOp(RHS, I->getOpcode()) && 224 "Not an expression that needs linearization?"); 225 226 DEBUG(errs() << "Linear" << *LHS << '\n' << *RHS << '\n' << *I << '\n'); 227 228 // Move the RHS instruction to live immediately before I, avoiding breaking 229 // dominator properties. 230 RHS->moveBefore(I); 231 232 // Move operands around to do the linearization. 233 I->setOperand(1, RHS->getOperand(0)); 234 RHS->setOperand(0, LHS); 235 I->setOperand(0, RHS); 236 237 ++NumLinear; 238 MadeChange = true; 239 DEBUG(errs() << "Linearized: " << *I << '\n'); 240 241 // If D is part of this expression tree, tail recurse. 242 if (isReassociableOp(I->getOperand(1), I->getOpcode())) 243 LinearizeExpr(I); 244} 245 246 247/// LinearizeExprTree - Given an associative binary expression tree, traverse 248/// all of the uses putting it into canonical form. This forces a left-linear 249/// form of the the expression (((a+b)+c)+d), and collects information about the 250/// rank of the non-tree operands. 251/// 252/// NOTE: These intentionally destroys the expression tree operands (turning 253/// them into undef values) to reduce #uses of the values. This means that the 254/// caller MUST use something like RewriteExprTree to put the values back in. 255/// 256void Reassociate::LinearizeExprTree(BinaryOperator *I, 257 std::vector<ValueEntry> &Ops) { 258 Value *LHS = I->getOperand(0), *RHS = I->getOperand(1); 259 unsigned Opcode = I->getOpcode(); 260 LLVMContext &Context = I->getContext(); 261 262 // First step, linearize the expression if it is in ((A+B)+(C+D)) form. 263 BinaryOperator *LHSBO = isReassociableOp(LHS, Opcode); 264 BinaryOperator *RHSBO = isReassociableOp(RHS, Opcode); 265 266 // If this is a multiply expression tree and it contains internal negations, 267 // transform them into multiplies by -1 so they can be reassociated. 268 if (I->getOpcode() == Instruction::Mul) { 269 if (!LHSBO && LHS->hasOneUse() && BinaryOperator::isNeg(LHS)) { 270 LHS = LowerNegateToMultiply(cast<Instruction>(LHS), 271 ValueRankMap, Context); 272 LHSBO = isReassociableOp(LHS, Opcode); 273 } 274 if (!RHSBO && RHS->hasOneUse() && BinaryOperator::isNeg(RHS)) { 275 RHS = LowerNegateToMultiply(cast<Instruction>(RHS), 276 ValueRankMap, Context); 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 } else { 293 // Turn X+(Y+Z) -> (Y+Z)+X 294 std::swap(LHSBO, RHSBO); 295 std::swap(LHS, RHS); 296 bool Success = !I->swapOperands(); 297 assert(Success && "swapOperands failed"); 298 Success = false; 299 MadeChange = true; 300 } 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 std::vector<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(LLVMContext &Context, Value *V, Instruction *BI) { 379 // We are trying to expose opportunity for reassociation. One of the things 380 // that we want to do to achieve this is to push a negation as deep into an 381 // expression chain as possible, to expose the add instructions. In practice, 382 // this means that we turn this: 383 // X = -(A+12+C+D) into X = -A + -12 + -C + -D = -12 + -A + -C + -D 384 // so that later, a: Y = 12+X could get reassociated with the -12 to eliminate 385 // the constants. We assume that instcombine will clean up the mess later if 386 // we introduce tons of unnecessary negation instructions... 387 // 388 if (Instruction *I = dyn_cast<Instruction>(V)) 389 if (I->getOpcode() == Instruction::Add && I->hasOneUse()) { 390 // Push the negates through the add. 391 I->setOperand(0, NegateValue(Context, I->getOperand(0), BI)); 392 I->setOperand(1, NegateValue(Context, I->getOperand(1), BI)); 393 394 // We must move the add instruction here, because the neg instructions do 395 // not dominate the old add instruction in general. By moving it, we are 396 // assured that the neg instructions we just inserted dominate the 397 // instruction we are about to insert after them. 398 // 399 I->moveBefore(BI); 400 I->setName(I->getName()+".neg"); 401 return I; 402 } 403 404 // Insert a 'neg' instruction that subtracts the value from zero to get the 405 // negation. 406 // 407 return BinaryOperator::CreateNeg(V, V->getName() + ".neg", BI); 408} 409 410/// ShouldBreakUpSubtract - Return true if we should break up this subtract of 411/// X-Y into (X + -Y). 412static bool ShouldBreakUpSubtract(LLVMContext &Context, Instruction *Sub) { 413 // If this is a negation, we can't split it up! 414 if (BinaryOperator::isNeg(Sub)) 415 return false; 416 417 // Don't bother to break this up unless either the LHS is an associable add or 418 // subtract or if this is only used by one. 419 if (isReassociableOp(Sub->getOperand(0), Instruction::Add) || 420 isReassociableOp(Sub->getOperand(0), Instruction::Sub)) 421 return true; 422 if (isReassociableOp(Sub->getOperand(1), Instruction::Add) || 423 isReassociableOp(Sub->getOperand(1), Instruction::Sub)) 424 return true; 425 if (Sub->hasOneUse() && 426 (isReassociableOp(Sub->use_back(), Instruction::Add) || 427 isReassociableOp(Sub->use_back(), Instruction::Sub))) 428 return true; 429 430 return false; 431} 432 433/// BreakUpSubtract - If we have (X-Y), and if either X is an add, or if this is 434/// only used by an add, transform this into (X+(0-Y)) to promote better 435/// reassociation. 436static Instruction *BreakUpSubtract(LLVMContext &Context, Instruction *Sub, 437 std::map<AssertingVH<>, unsigned> &ValueRankMap) { 438 // Convert a subtract into an add and a neg instruction... so that sub 439 // instructions can be commuted with other add instructions... 440 // 441 // Calculate the negative value of Operand 1 of the sub instruction... 442 // and set it as the RHS of the add instruction we just made... 443 // 444 Value *NegVal = NegateValue(Context, Sub->getOperand(1), Sub); 445 Instruction *New = 446 BinaryOperator::CreateAdd(Sub->getOperand(0), NegVal, "", Sub); 447 New->takeName(Sub); 448 449 // Everyone now refers to the add instruction. 450 ValueRankMap.erase(Sub); 451 Sub->replaceAllUsesWith(New); 452 Sub->eraseFromParent(); 453 454 DEBUG(errs() << "Negated: " << *New << '\n'); 455 return New; 456} 457 458/// ConvertShiftToMul - If this is a shift of a reassociable multiply or is used 459/// by one, change this into a multiply by a constant to assist with further 460/// reassociation. 461static Instruction *ConvertShiftToMul(Instruction *Shl, 462 std::map<AssertingVH<>, unsigned> &ValueRankMap, 463 LLVMContext &Context) { 464 // If an operand of this shift is a reassociable multiply, or if the shift 465 // is used by a reassociable multiply or add, turn into a multiply. 466 if (isReassociableOp(Shl->getOperand(0), Instruction::Mul) || 467 (Shl->hasOneUse() && 468 (isReassociableOp(Shl->use_back(), Instruction::Mul) || 469 isReassociableOp(Shl->use_back(), Instruction::Add)))) { 470 Constant *MulCst = ConstantInt::get(Shl->getType(), 1); 471 MulCst = 472 ConstantExpr::getShl(MulCst, cast<Constant>(Shl->getOperand(1))); 473 474 Instruction *Mul = BinaryOperator::CreateMul(Shl->getOperand(0), MulCst, 475 "", Shl); 476 ValueRankMap.erase(Shl); 477 Mul->takeName(Shl); 478 Shl->replaceAllUsesWith(Mul); 479 Shl->eraseFromParent(); 480 return Mul; 481 } 482 return 0; 483} 484 485// Scan backwards and forwards among values with the same rank as element i to 486// see if X exists. If X does not exist, return i. 487static unsigned FindInOperandList(std::vector<ValueEntry> &Ops, unsigned i, 488 Value *X) { 489 unsigned XRank = Ops[i].Rank; 490 unsigned e = Ops.size(); 491 for (unsigned j = i+1; j != e && Ops[j].Rank == XRank; ++j) 492 if (Ops[j].Op == X) 493 return j; 494 // Scan backwards 495 for (unsigned j = i-1; j != ~0U && Ops[j].Rank == XRank; --j) 496 if (Ops[j].Op == X) 497 return j; 498 return i; 499} 500 501/// EmitAddTreeOfValues - Emit a tree of add instructions, summing Ops together 502/// and returning the result. Insert the tree before I. 503static Value *EmitAddTreeOfValues(Instruction *I, std::vector<Value*> &Ops) { 504 if (Ops.size() == 1) return Ops.back(); 505 506 Value *V1 = Ops.back(); 507 Ops.pop_back(); 508 Value *V2 = EmitAddTreeOfValues(I, Ops); 509 return BinaryOperator::CreateAdd(V2, V1, "tmp", I); 510} 511 512/// RemoveFactorFromExpression - If V is an expression tree that is a 513/// multiplication sequence, and if this sequence contains a multiply by Factor, 514/// remove Factor from the tree and return the new tree. 515Value *Reassociate::RemoveFactorFromExpression(Value *V, Value *Factor) { 516 BinaryOperator *BO = isReassociableOp(V, Instruction::Mul); 517 if (!BO) return 0; 518 519 std::vector<ValueEntry> Factors; 520 LinearizeExprTree(BO, Factors); 521 522 bool FoundFactor = false; 523 for (unsigned i = 0, e = Factors.size(); i != e; ++i) 524 if (Factors[i].Op == Factor) { 525 FoundFactor = true; 526 Factors.erase(Factors.begin()+i); 527 break; 528 } 529 if (!FoundFactor) { 530 // Make sure to restore the operands to the expression tree. 531 RewriteExprTree(BO, Factors); 532 return 0; 533 } 534 535 if (Factors.size() == 1) return Factors[0].Op; 536 537 RewriteExprTree(BO, Factors); 538 return BO; 539} 540 541/// FindSingleUseMultiplyFactors - If V is a single-use multiply, recursively 542/// add its operands as factors, otherwise add V to the list of factors. 543static void FindSingleUseMultiplyFactors(Value *V, 544 std::vector<Value*> &Factors) { 545 BinaryOperator *BO; 546 if ((!V->hasOneUse() && !V->use_empty()) || 547 !(BO = dyn_cast<BinaryOperator>(V)) || 548 BO->getOpcode() != Instruction::Mul) { 549 Factors.push_back(V); 550 return; 551 } 552 553 // Otherwise, add the LHS and RHS to the list of factors. 554 FindSingleUseMultiplyFactors(BO->getOperand(1), Factors); 555 FindSingleUseMultiplyFactors(BO->getOperand(0), Factors); 556} 557 558 559 560Value *Reassociate::OptimizeExpression(BinaryOperator *I, 561 std::vector<ValueEntry> &Ops) { 562 // Now that we have the linearized expression tree, try to optimize it. 563 // Start by folding any constants that we found. 564 bool IterateOptimization = false; 565 if (Ops.size() == 1) return Ops[0].Op; 566 567 unsigned Opcode = I->getOpcode(); 568 569 if (Constant *V1 = dyn_cast<Constant>(Ops[Ops.size()-2].Op)) 570 if (Constant *V2 = dyn_cast<Constant>(Ops.back().Op)) { 571 Ops.pop_back(); 572 Ops.back().Op = ConstantExpr::get(Opcode, V1, V2); 573 return OptimizeExpression(I, Ops); 574 } 575 576 // Check for destructive annihilation due to a constant being used. 577 if (ConstantInt *CstVal = dyn_cast<ConstantInt>(Ops.back().Op)) 578 switch (Opcode) { 579 default: break; 580 case Instruction::And: 581 if (CstVal->isZero()) { // ... & 0 -> 0 582 ++NumAnnihil; 583 return CstVal; 584 } else if (CstVal->isAllOnesValue()) { // ... & -1 -> ... 585 Ops.pop_back(); 586 } 587 break; 588 case Instruction::Mul: 589 if (CstVal->isZero()) { // ... * 0 -> 0 590 ++NumAnnihil; 591 return CstVal; 592 } else if (cast<ConstantInt>(CstVal)->isOne()) { 593 Ops.pop_back(); // ... * 1 -> ... 594 } 595 break; 596 case Instruction::Or: 597 if (CstVal->isAllOnesValue()) { // ... | -1 -> -1 598 ++NumAnnihil; 599 return CstVal; 600 } 601 // FALLTHROUGH! 602 case Instruction::Add: 603 case Instruction::Xor: 604 if (CstVal->isZero()) // ... [|^+] 0 -> ... 605 Ops.pop_back(); 606 break; 607 } 608 if (Ops.size() == 1) return Ops[0].Op; 609 610 // Handle destructive annihilation do to identities between elements in the 611 // argument list here. 612 switch (Opcode) { 613 default: break; 614 case Instruction::And: 615 case Instruction::Or: 616 case Instruction::Xor: 617 // Scan the operand lists looking for X and ~X pairs, along with X,X pairs. 618 // If we find any, we can simplify the expression. X&~X == 0, X|~X == -1. 619 for (unsigned i = 0, e = Ops.size(); i != e; ++i) { 620 // First, check for X and ~X in the operand list. 621 assert(i < Ops.size()); 622 if (BinaryOperator::isNot(Ops[i].Op)) { // Cannot occur for ^. 623 Value *X = BinaryOperator::getNotArgument(Ops[i].Op); 624 unsigned FoundX = FindInOperandList(Ops, i, X); 625 if (FoundX != i) { 626 if (Opcode == Instruction::And) { // ...&X&~X = 0 627 ++NumAnnihil; 628 return Constant::getNullValue(X->getType()); 629 } else if (Opcode == Instruction::Or) { // ...|X|~X = -1 630 ++NumAnnihil; 631 return Constant::getAllOnesValue(X->getType()); 632 } 633 } 634 } 635 636 // Next, check for duplicate pairs of values, which we assume are next to 637 // each other, due to our sorting criteria. 638 assert(i < Ops.size()); 639 if (i+1 != Ops.size() && Ops[i+1].Op == Ops[i].Op) { 640 if (Opcode == Instruction::And || Opcode == Instruction::Or) { 641 // Drop duplicate values. 642 Ops.erase(Ops.begin()+i); 643 --i; --e; 644 IterateOptimization = true; 645 ++NumAnnihil; 646 } else { 647 assert(Opcode == Instruction::Xor); 648 if (e == 2) { 649 ++NumAnnihil; 650 return Constant::getNullValue(Ops[0].Op->getType()); 651 } 652 // ... X^X -> ... 653 Ops.erase(Ops.begin()+i, Ops.begin()+i+2); 654 i -= 1; e -= 2; 655 IterateOptimization = true; 656 ++NumAnnihil; 657 } 658 } 659 } 660 break; 661 662 case Instruction::Add: 663 // Scan the operand lists looking for X and -X pairs. If we find any, we 664 // can simplify the expression. X+-X == 0. 665 for (unsigned i = 0, e = Ops.size(); i != e; ++i) { 666 assert(i < Ops.size()); 667 // Check for X and -X in the operand list. 668 if (BinaryOperator::isNeg(Ops[i].Op)) { 669 Value *X = BinaryOperator::getNegArgument(Ops[i].Op); 670 unsigned FoundX = FindInOperandList(Ops, i, X); 671 if (FoundX != i) { 672 // Remove X and -X from the operand list. 673 if (Ops.size() == 2) { 674 ++NumAnnihil; 675 return Constant::getNullValue(X->getType()); 676 } else { 677 Ops.erase(Ops.begin()+i); 678 if (i < FoundX) 679 --FoundX; 680 else 681 --i; // Need to back up an extra one. 682 Ops.erase(Ops.begin()+FoundX); 683 IterateOptimization = true; 684 ++NumAnnihil; 685 --i; // Revisit element. 686 e -= 2; // Removed two elements. 687 } 688 } 689 } 690 } 691 692 693 // Scan the operand list, checking to see if there are any common factors 694 // between operands. Consider something like A*A+A*B*C+D. We would like to 695 // reassociate this to A*(A+B*C)+D, which reduces the number of multiplies. 696 // To efficiently find this, we count the number of times a factor occurs 697 // for any ADD operands that are MULs. 698 std::map<Value*, unsigned> FactorOccurrences; 699 unsigned MaxOcc = 0; 700 Value *MaxOccVal = 0; 701 for (unsigned i = 0, e = Ops.size(); i != e; ++i) { 702 if (BinaryOperator *BOp = dyn_cast<BinaryOperator>(Ops[i].Op)) { 703 if (BOp->getOpcode() == Instruction::Mul && BOp->use_empty()) { 704 // Compute all of the factors of this added value. 705 std::vector<Value*> Factors; 706 FindSingleUseMultiplyFactors(BOp, Factors); 707 assert(Factors.size() > 1 && "Bad linearize!"); 708 709 // Add one to FactorOccurrences for each unique factor in this op. 710 if (Factors.size() == 2) { 711 unsigned Occ = ++FactorOccurrences[Factors[0]]; 712 if (Occ > MaxOcc) { MaxOcc = Occ; MaxOccVal = Factors[0]; } 713 if (Factors[0] != Factors[1]) { // Don't double count A*A. 714 Occ = ++FactorOccurrences[Factors[1]]; 715 if (Occ > MaxOcc) { MaxOcc = Occ; MaxOccVal = Factors[1]; } 716 } 717 } else { 718 std::set<Value*> Duplicates; 719 for (unsigned i = 0, e = Factors.size(); i != e; ++i) { 720 if (Duplicates.insert(Factors[i]).second) { 721 unsigned Occ = ++FactorOccurrences[Factors[i]]; 722 if (Occ > MaxOcc) { MaxOcc = Occ; MaxOccVal = Factors[i]; } 723 } 724 } 725 } 726 } 727 } 728 } 729 730 // If any factor occurred more than one time, we can pull it out. 731 if (MaxOcc > 1) { 732 DEBUG(errs() << "\nFACTORING [" << MaxOcc << "]: " << *MaxOccVal << "\n"); 733 734 // Create a new instruction that uses the MaxOccVal twice. If we don't do 735 // this, we could otherwise run into situations where removing a factor 736 // from an expression will drop a use of maxocc, and this can cause 737 // RemoveFactorFromExpression on successive values to behave differently. 738 Instruction *DummyInst = BinaryOperator::CreateAdd(MaxOccVal, MaxOccVal); 739 std::vector<Value*> NewMulOps; 740 for (unsigned i = 0, e = Ops.size(); i != e; ++i) { 741 if (Value *V = RemoveFactorFromExpression(Ops[i].Op, MaxOccVal)) { 742 NewMulOps.push_back(V); 743 Ops.erase(Ops.begin()+i); 744 --i; --e; 745 } 746 } 747 748 // No need for extra uses anymore. 749 delete DummyInst; 750 751 unsigned NumAddedValues = NewMulOps.size(); 752 Value *V = EmitAddTreeOfValues(I, NewMulOps); 753 Value *V2 = BinaryOperator::CreateMul(V, MaxOccVal, "tmp", I); 754 755 // Now that we have inserted V and its sole use, optimize it. This allows 756 // us to handle cases that require multiple factoring steps, such as this: 757 // A*A*B + A*A*C --> A*(A*B+A*C) --> A*(A*(B+C)) 758 if (NumAddedValues > 1) 759 ReassociateExpression(cast<BinaryOperator>(V)); 760 761 ++NumFactor; 762 763 if (Ops.empty()) 764 return V2; 765 766 // Add the new value to the list of things being added. 767 Ops.insert(Ops.begin(), ValueEntry(getRank(V2), V2)); 768 769 // Rewrite the tree so that there is now a use of V. 770 RewriteExprTree(I, Ops); 771 return OptimizeExpression(I, Ops); 772 } 773 break; 774 //case Instruction::Mul: 775 } 776 777 if (IterateOptimization) 778 return OptimizeExpression(I, Ops); 779 return 0; 780} 781 782 783/// ReassociateBB - Inspect all of the instructions in this basic block, 784/// reassociating them as we go. 785void Reassociate::ReassociateBB(BasicBlock *BB) { 786 LLVMContext &Context = BB->getContext(); 787 788 for (BasicBlock::iterator BBI = BB->begin(); BBI != BB->end(); ) { 789 Instruction *BI = BBI++; 790 if (BI->getOpcode() == Instruction::Shl && 791 isa<ConstantInt>(BI->getOperand(1))) 792 if (Instruction *NI = ConvertShiftToMul(BI, ValueRankMap, Context)) { 793 MadeChange = true; 794 BI = NI; 795 } 796 797 // Reject cases where it is pointless to do this. 798 if (!isa<BinaryOperator>(BI) || BI->getType()->isFloatingPoint() || 799 isa<VectorType>(BI->getType())) 800 continue; // Floating point ops are not associative. 801 802 // If this is a subtract instruction which is not already in negate form, 803 // see if we can convert it to X+-Y. 804 if (BI->getOpcode() == Instruction::Sub) { 805 if (ShouldBreakUpSubtract(Context, BI)) { 806 BI = BreakUpSubtract(Context, BI, ValueRankMap); 807 MadeChange = true; 808 } else if (BinaryOperator::isNeg(BI)) { 809 // Otherwise, this is a negation. See if the operand is a multiply tree 810 // and if this is not an inner node of a multiply tree. 811 if (isReassociableOp(BI->getOperand(1), Instruction::Mul) && 812 (!BI->hasOneUse() || 813 !isReassociableOp(BI->use_back(), Instruction::Mul))) { 814 BI = LowerNegateToMultiply(BI, ValueRankMap, Context); 815 MadeChange = true; 816 } 817 } 818 } 819 820 // If this instruction is a commutative binary operator, process it. 821 if (!BI->isAssociative()) continue; 822 BinaryOperator *I = cast<BinaryOperator>(BI); 823 824 // If this is an interior node of a reassociable tree, ignore it until we 825 // get to the root of the tree, to avoid N^2 analysis. 826 if (I->hasOneUse() && isReassociableOp(I->use_back(), I->getOpcode())) 827 continue; 828 829 // If this is an add tree that is used by a sub instruction, ignore it 830 // until we process the subtract. 831 if (I->hasOneUse() && I->getOpcode() == Instruction::Add && 832 cast<Instruction>(I->use_back())->getOpcode() == Instruction::Sub) 833 continue; 834 835 ReassociateExpression(I); 836 } 837} 838 839void Reassociate::ReassociateExpression(BinaryOperator *I) { 840 841 // First, walk the expression tree, linearizing the tree, collecting 842 std::vector<ValueEntry> Ops; 843 LinearizeExprTree(I, Ops); 844 845 DEBUG(errs() << "RAIn:\t"; PrintOps(I, Ops); errs() << "\n"); 846 847 // Now that we have linearized the tree to a list and have gathered all of 848 // the operands and their ranks, sort the operands by their rank. Use a 849 // stable_sort so that values with equal ranks will have their relative 850 // positions maintained (and so the compiler is deterministic). Note that 851 // this sorts so that the highest ranking values end up at the beginning of 852 // the vector. 853 std::stable_sort(Ops.begin(), Ops.end()); 854 855 // OptimizeExpression - Now that we have the expression tree in a convenient 856 // sorted form, optimize it globally if possible. 857 if (Value *V = OptimizeExpression(I, Ops)) { 858 // This expression tree simplified to something that isn't a tree, 859 // eliminate it. 860 DEBUG(errs() << "Reassoc to scalar: " << *V << "\n"); 861 I->replaceAllUsesWith(V); 862 RemoveDeadBinaryOp(I); 863 return; 864 } 865 866 // We want to sink immediates as deeply as possible except in the case where 867 // this is a multiply tree used only by an add, and the immediate is a -1. 868 // In this case we reassociate to put the negation on the outside so that we 869 // can fold the negation into the add: (-X)*Y + Z -> Z-X*Y 870 if (I->getOpcode() == Instruction::Mul && I->hasOneUse() && 871 cast<Instruction>(I->use_back())->getOpcode() == Instruction::Add && 872 isa<ConstantInt>(Ops.back().Op) && 873 cast<ConstantInt>(Ops.back().Op)->isAllOnesValue()) { 874 Ops.insert(Ops.begin(), Ops.back()); 875 Ops.pop_back(); 876 } 877 878 DEBUG(errs() << "RAOut:\t"; PrintOps(I, Ops); errs() << "\n"); 879 880 if (Ops.size() == 1) { 881 // This expression tree simplified to something that isn't a tree, 882 // eliminate it. 883 I->replaceAllUsesWith(Ops[0].Op); 884 RemoveDeadBinaryOp(I); 885 } else { 886 // Now that we ordered and optimized the expressions, splat them back into 887 // the expression tree, removing any unneeded nodes. 888 RewriteExprTree(I, Ops); 889 } 890} 891 892 893bool Reassociate::runOnFunction(Function &F) { 894 // Recalculate the rank map for F 895 BuildRankMap(F); 896 897 MadeChange = false; 898 for (Function::iterator FI = F.begin(), FE = F.end(); FI != FE; ++FI) 899 ReassociateBB(FI); 900 901 // We are done with the rank map... 902 RankMap.clear(); 903 ValueRankMap.clear(); 904 return MadeChange; 905} 906 907