Reassociate.cpp revision baf3c404409d5e47b13984a7f95bfbd6d1f2e79e
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/Compiler.h" 35#include "llvm/Support/Debug.h" 36#include "llvm/Support/ValueHandle.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 VISIBILITY_HIDDEN 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 cerr << Instruction::getOpcodeName(I->getOpcode()) << " " 65 << *Ops[0].Op->getType(); 66 for (unsigned i = 0, e = Ops.size(); i != e; ++i) { 67 WriteAsOperand(*cerr.stream() << " ", Ops[i].Op, false, M); 68 cerr << "," << Ops[i].Rank; 69 } 70} 71#endif 72 73namespace { 74 class VISIBILITY_HIDDEN 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::Malloc || 125 I->getOpcode() == Instruction::Invoke || 126 (I->getOpcode() == Instruction::Call && 127 !isa<DbgInfoIntrinsic>(I)) || 128 I->getOpcode() == Instruction::UDiv || 129 I->getOpcode() == Instruction::SDiv || 130 I->getOpcode() == Instruction::FDiv || 131 I->getOpcode() == Instruction::URem || 132 I->getOpcode() == Instruction::SRem || 133 I->getOpcode() == Instruction::FRem) 134 return true; 135 return false; 136} 137 138void Reassociate::BuildRankMap(Function &F) { 139 unsigned i = 2; 140 141 // Assign distinct ranks to function arguments 142 for (Function::arg_iterator I = F.arg_begin(), E = F.arg_end(); I != E; ++I) 143 ValueRankMap[&*I] = ++i; 144 145 ReversePostOrderTraversal<Function*> RPOT(&F); 146 for (ReversePostOrderTraversal<Function*>::rpo_iterator I = RPOT.begin(), 147 E = RPOT.end(); I != E; ++I) { 148 BasicBlock *BB = *I; 149 unsigned BBRank = RankMap[BB] = ++i << 16; 150 151 // Walk the basic block, adding precomputed ranks for any instructions that 152 // we cannot move. This ensures that the ranks for these instructions are 153 // all different in the block. 154 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) 155 if (isUnmovableInstruction(I)) 156 ValueRankMap[&*I] = ++BBRank; 157 } 158} 159 160unsigned Reassociate::getRank(Value *V) { 161 if (isa<Argument>(V)) return ValueRankMap[V]; // Function argument... 162 163 Instruction *I = dyn_cast<Instruction>(V); 164 if (I == 0) return 0; // Otherwise it's a global or constant, rank 0. 165 166 unsigned &CachedRank = ValueRankMap[I]; 167 if (CachedRank) return CachedRank; // Rank already known? 168 169 // If this is an expression, return the 1+MAX(rank(LHS), rank(RHS)) so that 170 // we can reassociate expressions for code motion! Since we do not recurse 171 // for PHI nodes, we cannot have infinite recursion here, because there 172 // cannot be loops in the value graph that do not go through PHI nodes. 173 unsigned Rank = 0, MaxRank = RankMap[I->getParent()]; 174 for (unsigned i = 0, e = I->getNumOperands(); 175 i != e && Rank != MaxRank; ++i) 176 Rank = std::max(Rank, getRank(I->getOperand(i))); 177 178 // If this is a not or neg instruction, do not count it for rank. This 179 // assures us that X and ~X will have the same rank. 180 if (!I->getType()->isInteger() || 181 (!BinaryOperator::isNot(I) && !BinaryOperator::isNeg(I))) 182 ++Rank; 183 184 //DOUT << "Calculated Rank[" << V->getName() << "] = " 185 // << Rank << "\n"; 186 187 return CachedRank = Rank; 188} 189 190/// isReassociableOp - Return true if V is an instruction of the specified 191/// opcode and if it only has one use. 192static BinaryOperator *isReassociableOp(Value *V, unsigned Opcode) { 193 if ((V->hasOneUse() || V->use_empty()) && isa<Instruction>(V) && 194 cast<Instruction>(V)->getOpcode() == Opcode) 195 return cast<BinaryOperator>(V); 196 return 0; 197} 198 199/// LowerNegateToMultiply - Replace 0-X with X*-1. 200/// 201static Instruction *LowerNegateToMultiply(Instruction *Neg, 202 std::map<AssertingVH<>, unsigned> &ValueRankMap, 203 LLVMContext &Context) { 204 Constant *Cst = Neg->getContext().getAllOnesValue(Neg->getType()); 205 206 Instruction *Res = BinaryOperator::CreateMul(Neg->getOperand(1), Cst, "",Neg); 207 ValueRankMap.erase(Neg); 208 Res->takeName(Neg); 209 Neg->replaceAllUsesWith(Res); 210 Neg->eraseFromParent(); 211 return Res; 212} 213 214// Given an expression of the form '(A+B)+(D+C)', turn it into '(((A+B)+C)+D)'. 215// Note that if D is also part of the expression tree that we recurse to 216// linearize it as well. Besides that case, this does not recurse into A,B, or 217// C. 218void Reassociate::LinearizeExpr(BinaryOperator *I) { 219 BinaryOperator *LHS = cast<BinaryOperator>(I->getOperand(0)); 220 BinaryOperator *RHS = cast<BinaryOperator>(I->getOperand(1)); 221 assert(isReassociableOp(LHS, I->getOpcode()) && 222 isReassociableOp(RHS, I->getOpcode()) && 223 "Not an expression that needs linearization?"); 224 225 DOUT << "Linear" << *LHS << *RHS << *I; 226 227 // Move the RHS instruction to live immediately before I, avoiding breaking 228 // dominator properties. 229 RHS->moveBefore(I); 230 231 // Move operands around to do the linearization. 232 I->setOperand(1, RHS->getOperand(0)); 233 RHS->setOperand(0, LHS); 234 I->setOperand(0, RHS); 235 236 ++NumLinear; 237 MadeChange = true; 238 DOUT << "Linearized: " << *I; 239 240 // If D is part of this expression tree, tail recurse. 241 if (isReassociableOp(I->getOperand(1), I->getOpcode())) 242 LinearizeExpr(I); 243} 244 245 246/// LinearizeExprTree - Given an associative binary expression tree, traverse 247/// all of the uses putting it into canonical form. This forces a left-linear 248/// form of the the expression (((a+b)+c)+d), and collects information about the 249/// rank of the non-tree operands. 250/// 251/// NOTE: These intentionally destroys the expression tree operands (turning 252/// them into undef values) to reduce #uses of the values. This means that the 253/// caller MUST use something like RewriteExprTree to put the values back in. 254/// 255void Reassociate::LinearizeExprTree(BinaryOperator *I, 256 std::vector<ValueEntry> &Ops) { 257 Value *LHS = I->getOperand(0), *RHS = I->getOperand(1); 258 unsigned Opcode = I->getOpcode(); 259 LLVMContext &Context = I->getContext(); 260 261 // First step, linearize the expression if it is in ((A+B)+(C+D)) form. 262 BinaryOperator *LHSBO = isReassociableOp(LHS, Opcode); 263 BinaryOperator *RHSBO = isReassociableOp(RHS, Opcode); 264 265 // If this is a multiply expression tree and it contains internal negations, 266 // transform them into multiplies by -1 so they can be reassociated. 267 if (I->getOpcode() == Instruction::Mul) { 268 if (!LHSBO && LHS->hasOneUse() && BinaryOperator::isNeg(LHS)) { 269 LHS = LowerNegateToMultiply(cast<Instruction>(LHS), 270 ValueRankMap, Context); 271 LHSBO = isReassociableOp(LHS, Opcode); 272 } 273 if (!RHSBO && RHS->hasOneUse() && BinaryOperator::isNeg(RHS)) { 274 RHS = LowerNegateToMultiply(cast<Instruction>(RHS), 275 ValueRankMap, Context); 276 RHSBO = isReassociableOp(RHS, Opcode); 277 } 278 } 279 280 if (!LHSBO) { 281 if (!RHSBO) { 282 // Neither the LHS or RHS as part of the tree, thus this is a leaf. As 283 // such, just remember these operands and their rank. 284 Ops.push_back(ValueEntry(getRank(LHS), LHS)); 285 Ops.push_back(ValueEntry(getRank(RHS), RHS)); 286 287 // Clear the leaves out. 288 I->setOperand(0, Context.getUndef(I->getType())); 289 I->setOperand(1, Context.getUndef(I->getType())); 290 return; 291 } else { 292 // Turn X+(Y+Z) -> (Y+Z)+X 293 std::swap(LHSBO, RHSBO); 294 std::swap(LHS, RHS); 295 bool Success = !I->swapOperands(); 296 assert(Success && "swapOperands failed"); 297 Success = false; 298 MadeChange = true; 299 } 300 } else if (RHSBO) { 301 // Turn (A+B)+(C+D) -> (((A+B)+C)+D). This guarantees the the RHS is not 302 // part of the expression tree. 303 LinearizeExpr(I); 304 LHS = LHSBO = cast<BinaryOperator>(I->getOperand(0)); 305 RHS = I->getOperand(1); 306 RHSBO = 0; 307 } 308 309 // Okay, now we know that the LHS is a nested expression and that the RHS is 310 // not. Perform reassociation. 311 assert(!isReassociableOp(RHS, Opcode) && "LinearizeExpr failed!"); 312 313 // Move LHS right before I to make sure that the tree expression dominates all 314 // values. 315 LHSBO->moveBefore(I); 316 317 // Linearize the expression tree on the LHS. 318 LinearizeExprTree(LHSBO, Ops); 319 320 // Remember the RHS operand and its rank. 321 Ops.push_back(ValueEntry(getRank(RHS), RHS)); 322 323 // Clear the RHS leaf out. 324 I->setOperand(1, Context.getUndef(I->getType())); 325} 326 327// RewriteExprTree - Now that the operands for this expression tree are 328// linearized and optimized, emit them in-order. This function is written to be 329// tail recursive. 330void Reassociate::RewriteExprTree(BinaryOperator *I, 331 std::vector<ValueEntry> &Ops, 332 unsigned i) { 333 if (i+2 == Ops.size()) { 334 if (I->getOperand(0) != Ops[i].Op || 335 I->getOperand(1) != Ops[i+1].Op) { 336 Value *OldLHS = I->getOperand(0); 337 DOUT << "RA: " << *I; 338 I->setOperand(0, Ops[i].Op); 339 I->setOperand(1, Ops[i+1].Op); 340 DOUT << "TO: " << *I; 341 MadeChange = true; 342 ++NumChanged; 343 344 // If we reassociated a tree to fewer operands (e.g. (1+a+2) -> (a+3) 345 // delete the extra, now dead, nodes. 346 RemoveDeadBinaryOp(OldLHS); 347 } 348 return; 349 } 350 assert(i+2 < Ops.size() && "Ops index out of range!"); 351 352 if (I->getOperand(1) != Ops[i].Op) { 353 DOUT << "RA: " << *I; 354 I->setOperand(1, Ops[i].Op); 355 DOUT << "TO: " << *I; 356 MadeChange = true; 357 ++NumChanged; 358 } 359 360 BinaryOperator *LHS = cast<BinaryOperator>(I->getOperand(0)); 361 assert(LHS->getOpcode() == I->getOpcode() && 362 "Improper expression tree!"); 363 364 // Compactify the tree instructions together with each other to guarantee 365 // that the expression tree is dominated by all of Ops. 366 LHS->moveBefore(I); 367 RewriteExprTree(LHS, Ops, i+1); 368} 369 370 371 372// NegateValue - Insert instructions before the instruction pointed to by BI, 373// that computes the negative version of the value specified. The negative 374// version of the value is returned, and BI is left pointing at the instruction 375// that should be processed next by the reassociation pass. 376// 377static Value *NegateValue(LLVMContext &Context, Value *V, Instruction *BI) { 378 // We are trying to expose opportunity for reassociation. One of the things 379 // that we want to do to achieve this is to push a negation as deep into an 380 // expression chain as possible, to expose the add instructions. In practice, 381 // this means that we turn this: 382 // X = -(A+12+C+D) into X = -A + -12 + -C + -D = -12 + -A + -C + -D 383 // so that later, a: Y = 12+X could get reassociated with the -12 to eliminate 384 // the constants. We assume that instcombine will clean up the mess later if 385 // we introduce tons of unnecessary negation instructions... 386 // 387 if (Instruction *I = dyn_cast<Instruction>(V)) 388 if (I->getOpcode() == Instruction::Add && I->hasOneUse()) { 389 // Push the negates through the add. 390 I->setOperand(0, NegateValue(Context, I->getOperand(0), BI)); 391 I->setOperand(1, NegateValue(Context, I->getOperand(1), BI)); 392 393 // We must move the add instruction here, because the neg instructions do 394 // not dominate the old add instruction in general. By moving it, we are 395 // assured that the neg instructions we just inserted dominate the 396 // instruction we are about to insert after them. 397 // 398 I->moveBefore(BI); 399 I->setName(I->getName()+".neg"); 400 return I; 401 } 402 403 // Insert a 'neg' instruction that subtracts the value from zero to get the 404 // negation. 405 // 406 return BinaryOperator::CreateNeg(Context, V, V->getName() + ".neg", BI); 407} 408 409/// ShouldBreakUpSubtract - Return true if we should break up this subtract of 410/// X-Y into (X + -Y). 411static bool ShouldBreakUpSubtract(LLVMContext &Context, Instruction *Sub) { 412 // If this is a negation, we can't split it up! 413 if (BinaryOperator::isNeg(Sub)) 414 return false; 415 416 // Don't bother to break this up unless either the LHS is an associable add or 417 // subtract or if this is only used by one. 418 if (isReassociableOp(Sub->getOperand(0), Instruction::Add) || 419 isReassociableOp(Sub->getOperand(0), Instruction::Sub)) 420 return true; 421 if (isReassociableOp(Sub->getOperand(1), Instruction::Add) || 422 isReassociableOp(Sub->getOperand(1), Instruction::Sub)) 423 return true; 424 if (Sub->hasOneUse() && 425 (isReassociableOp(Sub->use_back(), Instruction::Add) || 426 isReassociableOp(Sub->use_back(), Instruction::Sub))) 427 return true; 428 429 return false; 430} 431 432/// BreakUpSubtract - If we have (X-Y), and if either X is an add, or if this is 433/// only used by an add, transform this into (X+(0-Y)) to promote better 434/// reassociation. 435static Instruction *BreakUpSubtract(LLVMContext &Context, Instruction *Sub, 436 std::map<AssertingVH<>, unsigned> &ValueRankMap) { 437 // Convert a subtract into an add and a neg instruction... so that sub 438 // instructions can be commuted with other add instructions... 439 // 440 // Calculate the negative value of Operand 1 of the sub instruction... 441 // and set it as the RHS of the add instruction we just made... 442 // 443 Value *NegVal = NegateValue(Context, Sub->getOperand(1), Sub); 444 Instruction *New = 445 BinaryOperator::CreateAdd(Sub->getOperand(0), NegVal, "", Sub); 446 New->takeName(Sub); 447 448 // Everyone now refers to the add instruction. 449 ValueRankMap.erase(Sub); 450 Sub->replaceAllUsesWith(New); 451 Sub->eraseFromParent(); 452 453 DOUT << "Negated: " << *New; 454 return New; 455} 456 457/// ConvertShiftToMul - If this is a shift of a reassociable multiply or is used 458/// by one, change this into a multiply by a constant to assist with further 459/// reassociation. 460static Instruction *ConvertShiftToMul(Instruction *Shl, 461 std::map<AssertingVH<>, unsigned> &ValueRankMap, 462 LLVMContext &Context) { 463 // If an operand of this shift is a reassociable multiply, or if the shift 464 // is used by a reassociable multiply or add, turn into a multiply. 465 if (isReassociableOp(Shl->getOperand(0), Instruction::Mul) || 466 (Shl->hasOneUse() && 467 (isReassociableOp(Shl->use_back(), Instruction::Mul) || 468 isReassociableOp(Shl->use_back(), Instruction::Add)))) { 469 Constant *MulCst = ConstantInt::get(Shl->getType(), 1); 470 MulCst = 471 ConstantExpr::getShl(MulCst, cast<Constant>(Shl->getOperand(1))); 472 473 Instruction *Mul = BinaryOperator::CreateMul(Shl->getOperand(0), MulCst, 474 "", Shl); 475 ValueRankMap.erase(Shl); 476 Mul->takeName(Shl); 477 Shl->replaceAllUsesWith(Mul); 478 Shl->eraseFromParent(); 479 return Mul; 480 } 481 return 0; 482} 483 484// Scan backwards and forwards among values with the same rank as element i to 485// see if X exists. If X does not exist, return i. 486static unsigned FindInOperandList(std::vector<ValueEntry> &Ops, unsigned i, 487 Value *X) { 488 unsigned XRank = Ops[i].Rank; 489 unsigned e = Ops.size(); 490 for (unsigned j = i+1; j != e && Ops[j].Rank == XRank; ++j) 491 if (Ops[j].Op == X) 492 return j; 493 // Scan backwards 494 for (unsigned j = i-1; j != ~0U && Ops[j].Rank == XRank; --j) 495 if (Ops[j].Op == X) 496 return j; 497 return i; 498} 499 500/// EmitAddTreeOfValues - Emit a tree of add instructions, summing Ops together 501/// and returning the result. Insert the tree before I. 502static Value *EmitAddTreeOfValues(Instruction *I, std::vector<Value*> &Ops) { 503 if (Ops.size() == 1) return Ops.back(); 504 505 Value *V1 = Ops.back(); 506 Ops.pop_back(); 507 Value *V2 = EmitAddTreeOfValues(I, Ops); 508 return BinaryOperator::CreateAdd(V2, V1, "tmp", I); 509} 510 511/// RemoveFactorFromExpression - If V is an expression tree that is a 512/// multiplication sequence, and if this sequence contains a multiply by Factor, 513/// remove Factor from the tree and return the new tree. 514Value *Reassociate::RemoveFactorFromExpression(Value *V, Value *Factor) { 515 BinaryOperator *BO = isReassociableOp(V, Instruction::Mul); 516 if (!BO) return 0; 517 518 std::vector<ValueEntry> Factors; 519 LinearizeExprTree(BO, Factors); 520 521 bool FoundFactor = false; 522 for (unsigned i = 0, e = Factors.size(); i != e; ++i) 523 if (Factors[i].Op == Factor) { 524 FoundFactor = true; 525 Factors.erase(Factors.begin()+i); 526 break; 527 } 528 if (!FoundFactor) { 529 // Make sure to restore the operands to the expression tree. 530 RewriteExprTree(BO, Factors); 531 return 0; 532 } 533 534 if (Factors.size() == 1) return Factors[0].Op; 535 536 RewriteExprTree(BO, Factors); 537 return BO; 538} 539 540/// FindSingleUseMultiplyFactors - If V is a single-use multiply, recursively 541/// add its operands as factors, otherwise add V to the list of factors. 542static void FindSingleUseMultiplyFactors(Value *V, 543 std::vector<Value*> &Factors) { 544 BinaryOperator *BO; 545 if ((!V->hasOneUse() && !V->use_empty()) || 546 !(BO = dyn_cast<BinaryOperator>(V)) || 547 BO->getOpcode() != Instruction::Mul) { 548 Factors.push_back(V); 549 return; 550 } 551 552 // Otherwise, add the LHS and RHS to the list of factors. 553 FindSingleUseMultiplyFactors(BO->getOperand(1), Factors); 554 FindSingleUseMultiplyFactors(BO->getOperand(0), Factors); 555} 556 557 558 559Value *Reassociate::OptimizeExpression(BinaryOperator *I, 560 std::vector<ValueEntry> &Ops) { 561 // Now that we have the linearized expression tree, try to optimize it. 562 // Start by folding any constants that we found. 563 bool IterateOptimization = false; 564 if (Ops.size() == 1) return Ops[0].Op; 565 566 LLVMContext &Context = I->getContext(); 567 568 unsigned Opcode = I->getOpcode(); 569 570 if (Constant *V1 = dyn_cast<Constant>(Ops[Ops.size()-2].Op)) 571 if (Constant *V2 = dyn_cast<Constant>(Ops.back().Op)) { 572 Ops.pop_back(); 573 Ops.back().Op = ConstantExpr::get(Opcode, V1, V2); 574 return OptimizeExpression(I, Ops); 575 } 576 577 // Check for destructive annihilation due to a constant being used. 578 if (ConstantInt *CstVal = dyn_cast<ConstantInt>(Ops.back().Op)) 579 switch (Opcode) { 580 default: break; 581 case Instruction::And: 582 if (CstVal->isZero()) { // ... & 0 -> 0 583 ++NumAnnihil; 584 return CstVal; 585 } else if (CstVal->isAllOnesValue()) { // ... & -1 -> ... 586 Ops.pop_back(); 587 } 588 break; 589 case Instruction::Mul: 590 if (CstVal->isZero()) { // ... * 0 -> 0 591 ++NumAnnihil; 592 return CstVal; 593 } else if (cast<ConstantInt>(CstVal)->isOne()) { 594 Ops.pop_back(); // ... * 1 -> ... 595 } 596 break; 597 case Instruction::Or: 598 if (CstVal->isAllOnesValue()) { // ... | -1 -> -1 599 ++NumAnnihil; 600 return CstVal; 601 } 602 // FALLTHROUGH! 603 case Instruction::Add: 604 case Instruction::Xor: 605 if (CstVal->isZero()) // ... [|^+] 0 -> ... 606 Ops.pop_back(); 607 break; 608 } 609 if (Ops.size() == 1) return Ops[0].Op; 610 611 // Handle destructive annihilation do to identities between elements in the 612 // argument list here. 613 switch (Opcode) { 614 default: break; 615 case Instruction::And: 616 case Instruction::Or: 617 case Instruction::Xor: 618 // Scan the operand lists looking for X and ~X pairs, along with X,X pairs. 619 // If we find any, we can simplify the expression. X&~X == 0, X|~X == -1. 620 for (unsigned i = 0, e = Ops.size(); i != e; ++i) { 621 // First, check for X and ~X in the operand list. 622 assert(i < Ops.size()); 623 if (BinaryOperator::isNot(Ops[i].Op)) { // Cannot occur for ^. 624 Value *X = BinaryOperator::getNotArgument(Ops[i].Op); 625 unsigned FoundX = FindInOperandList(Ops, i, X); 626 if (FoundX != i) { 627 if (Opcode == Instruction::And) { // ...&X&~X = 0 628 ++NumAnnihil; 629 return Context.getNullValue(X->getType()); 630 } else if (Opcode == Instruction::Or) { // ...|X|~X = -1 631 ++NumAnnihil; 632 return Context.getAllOnesValue(X->getType()); 633 } 634 } 635 } 636 637 // Next, check for duplicate pairs of values, which we assume are next to 638 // each other, due to our sorting criteria. 639 assert(i < Ops.size()); 640 if (i+1 != Ops.size() && Ops[i+1].Op == Ops[i].Op) { 641 if (Opcode == Instruction::And || Opcode == Instruction::Or) { 642 // Drop duplicate values. 643 Ops.erase(Ops.begin()+i); 644 --i; --e; 645 IterateOptimization = true; 646 ++NumAnnihil; 647 } else { 648 assert(Opcode == Instruction::Xor); 649 if (e == 2) { 650 ++NumAnnihil; 651 return Context.getNullValue(Ops[0].Op->getType()); 652 } 653 // ... X^X -> ... 654 Ops.erase(Ops.begin()+i, Ops.begin()+i+2); 655 i -= 1; e -= 2; 656 IterateOptimization = true; 657 ++NumAnnihil; 658 } 659 } 660 } 661 break; 662 663 case Instruction::Add: 664 // Scan the operand lists looking for X and -X pairs. If we find any, we 665 // can simplify the expression. X+-X == 0. 666 for (unsigned i = 0, e = Ops.size(); i != e; ++i) { 667 assert(i < Ops.size()); 668 // Check for X and -X in the operand list. 669 if (BinaryOperator::isNeg(Ops[i].Op)) { 670 Value *X = BinaryOperator::getNegArgument(Ops[i].Op); 671 unsigned FoundX = FindInOperandList(Ops, i, X); 672 if (FoundX != i) { 673 // Remove X and -X from the operand list. 674 if (Ops.size() == 2) { 675 ++NumAnnihil; 676 return Context.getNullValue(X->getType()); 677 } else { 678 Ops.erase(Ops.begin()+i); 679 if (i < FoundX) 680 --FoundX; 681 else 682 --i; // Need to back up an extra one. 683 Ops.erase(Ops.begin()+FoundX); 684 IterateOptimization = true; 685 ++NumAnnihil; 686 --i; // Revisit element. 687 e -= 2; // Removed two elements. 688 } 689 } 690 } 691 } 692 693 694 // Scan the operand list, checking to see if there are any common factors 695 // between operands. Consider something like A*A+A*B*C+D. We would like to 696 // reassociate this to A*(A+B*C)+D, which reduces the number of multiplies. 697 // To efficiently find this, we count the number of times a factor occurs 698 // for any ADD operands that are MULs. 699 std::map<Value*, unsigned> FactorOccurrences; 700 unsigned MaxOcc = 0; 701 Value *MaxOccVal = 0; 702 for (unsigned i = 0, e = Ops.size(); i != e; ++i) { 703 if (BinaryOperator *BOp = dyn_cast<BinaryOperator>(Ops[i].Op)) { 704 if (BOp->getOpcode() == Instruction::Mul && BOp->use_empty()) { 705 // Compute all of the factors of this added value. 706 std::vector<Value*> Factors; 707 FindSingleUseMultiplyFactors(BOp, Factors); 708 assert(Factors.size() > 1 && "Bad linearize!"); 709 710 // Add one to FactorOccurrences for each unique factor in this op. 711 if (Factors.size() == 2) { 712 unsigned Occ = ++FactorOccurrences[Factors[0]]; 713 if (Occ > MaxOcc) { MaxOcc = Occ; MaxOccVal = Factors[0]; } 714 if (Factors[0] != Factors[1]) { // Don't double count A*A. 715 Occ = ++FactorOccurrences[Factors[1]]; 716 if (Occ > MaxOcc) { MaxOcc = Occ; MaxOccVal = Factors[1]; } 717 } 718 } else { 719 std::set<Value*> Duplicates; 720 for (unsigned i = 0, e = Factors.size(); i != e; ++i) { 721 if (Duplicates.insert(Factors[i]).second) { 722 unsigned Occ = ++FactorOccurrences[Factors[i]]; 723 if (Occ > MaxOcc) { MaxOcc = Occ; MaxOccVal = Factors[i]; } 724 } 725 } 726 } 727 } 728 } 729 } 730 731 // If any factor occurred more than one time, we can pull it out. 732 if (MaxOcc > 1) { 733 DOUT << "\nFACTORING [" << MaxOcc << "]: " << *MaxOccVal << "\n"; 734 735 // Create a new instruction that uses the MaxOccVal twice. If we don't do 736 // this, we could otherwise run into situations where removing a factor 737 // from an expression will drop a use of maxocc, and this can cause 738 // RemoveFactorFromExpression on successive values to behave differently. 739 Instruction *DummyInst = BinaryOperator::CreateAdd(MaxOccVal, MaxOccVal); 740 std::vector<Value*> NewMulOps; 741 for (unsigned i = 0, e = Ops.size(); i != e; ++i) { 742 if (Value *V = RemoveFactorFromExpression(Ops[i].Op, MaxOccVal)) { 743 NewMulOps.push_back(V); 744 Ops.erase(Ops.begin()+i); 745 --i; --e; 746 } 747 } 748 749 // No need for extra uses anymore. 750 delete DummyInst; 751 752 unsigned NumAddedValues = NewMulOps.size(); 753 Value *V = EmitAddTreeOfValues(I, NewMulOps); 754 Value *V2 = BinaryOperator::CreateMul(V, MaxOccVal, "tmp", I); 755 756 // Now that we have inserted V and its sole use, optimize it. This allows 757 // us to handle cases that require multiple factoring steps, such as this: 758 // A*A*B + A*A*C --> A*(A*B+A*C) --> A*(A*(B+C)) 759 if (NumAddedValues > 1) 760 ReassociateExpression(cast<BinaryOperator>(V)); 761 762 ++NumFactor; 763 764 if (Ops.empty()) 765 return V2; 766 767 // Add the new value to the list of things being added. 768 Ops.insert(Ops.begin(), ValueEntry(getRank(V2), V2)); 769 770 // Rewrite the tree so that there is now a use of V. 771 RewriteExprTree(I, Ops); 772 return OptimizeExpression(I, Ops); 773 } 774 break; 775 //case Instruction::Mul: 776 } 777 778 if (IterateOptimization) 779 return OptimizeExpression(I, Ops); 780 return 0; 781} 782 783 784/// ReassociateBB - Inspect all of the instructions in this basic block, 785/// reassociating them as we go. 786void Reassociate::ReassociateBB(BasicBlock *BB) { 787 LLVMContext &Context = BB->getContext(); 788 789 for (BasicBlock::iterator BBI = BB->begin(); BBI != BB->end(); ) { 790 Instruction *BI = BBI++; 791 if (BI->getOpcode() == Instruction::Shl && 792 isa<ConstantInt>(BI->getOperand(1))) 793 if (Instruction *NI = ConvertShiftToMul(BI, ValueRankMap, Context)) { 794 MadeChange = true; 795 BI = NI; 796 } 797 798 // Reject cases where it is pointless to do this. 799 if (!isa<BinaryOperator>(BI) || BI->getType()->isFloatingPoint() || 800 isa<VectorType>(BI->getType())) 801 continue; // Floating point ops are not associative. 802 803 // If this is a subtract instruction which is not already in negate form, 804 // see if we can convert it to X+-Y. 805 if (BI->getOpcode() == Instruction::Sub) { 806 if (ShouldBreakUpSubtract(Context, BI)) { 807 BI = BreakUpSubtract(Context, BI, ValueRankMap); 808 MadeChange = true; 809 } else if (BinaryOperator::isNeg(BI)) { 810 // Otherwise, this is a negation. See if the operand is a multiply tree 811 // and if this is not an inner node of a multiply tree. 812 if (isReassociableOp(BI->getOperand(1), Instruction::Mul) && 813 (!BI->hasOneUse() || 814 !isReassociableOp(BI->use_back(), Instruction::Mul))) { 815 BI = LowerNegateToMultiply(BI, ValueRankMap, Context); 816 MadeChange = true; 817 } 818 } 819 } 820 821 // If this instruction is a commutative binary operator, process it. 822 if (!BI->isAssociative()) continue; 823 BinaryOperator *I = cast<BinaryOperator>(BI); 824 825 // If this is an interior node of a reassociable tree, ignore it until we 826 // get to the root of the tree, to avoid N^2 analysis. 827 if (I->hasOneUse() && isReassociableOp(I->use_back(), I->getOpcode())) 828 continue; 829 830 // If this is an add tree that is used by a sub instruction, ignore it 831 // until we process the subtract. 832 if (I->hasOneUse() && I->getOpcode() == Instruction::Add && 833 cast<Instruction>(I->use_back())->getOpcode() == Instruction::Sub) 834 continue; 835 836 ReassociateExpression(I); 837 } 838} 839 840void Reassociate::ReassociateExpression(BinaryOperator *I) { 841 842 // First, walk the expression tree, linearizing the tree, collecting 843 std::vector<ValueEntry> Ops; 844 LinearizeExprTree(I, Ops); 845 846 DOUT << "RAIn:\t"; DEBUG(PrintOps(I, Ops)); DOUT << "\n"; 847 848 // Now that we have linearized the tree to a list and have gathered all of 849 // the operands and their ranks, sort the operands by their rank. Use a 850 // stable_sort so that values with equal ranks will have their relative 851 // positions maintained (and so the compiler is deterministic). Note that 852 // this sorts so that the highest ranking values end up at the beginning of 853 // the vector. 854 std::stable_sort(Ops.begin(), Ops.end()); 855 856 // OptimizeExpression - Now that we have the expression tree in a convenient 857 // sorted form, optimize it globally if possible. 858 if (Value *V = OptimizeExpression(I, Ops)) { 859 // This expression tree simplified to something that isn't a tree, 860 // eliminate it. 861 DOUT << "Reassoc to scalar: " << *V << "\n"; 862 I->replaceAllUsesWith(V); 863 RemoveDeadBinaryOp(I); 864 return; 865 } 866 867 // We want to sink immediates as deeply as possible except in the case where 868 // this is a multiply tree used only by an add, and the immediate is a -1. 869 // In this case we reassociate to put the negation on the outside so that we 870 // can fold the negation into the add: (-X)*Y + Z -> Z-X*Y 871 if (I->getOpcode() == Instruction::Mul && I->hasOneUse() && 872 cast<Instruction>(I->use_back())->getOpcode() == Instruction::Add && 873 isa<ConstantInt>(Ops.back().Op) && 874 cast<ConstantInt>(Ops.back().Op)->isAllOnesValue()) { 875 Ops.insert(Ops.begin(), Ops.back()); 876 Ops.pop_back(); 877 } 878 879 DOUT << "RAOut:\t"; DEBUG(PrintOps(I, Ops)); DOUT << "\n"; 880 881 if (Ops.size() == 1) { 882 // This expression tree simplified to something that isn't a tree, 883 // eliminate it. 884 I->replaceAllUsesWith(Ops[0].Op); 885 RemoveDeadBinaryOp(I); 886 } else { 887 // Now that we ordered and optimized the expressions, splat them back into 888 // the expression tree, removing any unneeded nodes. 889 RewriteExprTree(I, Ops); 890 } 891} 892 893 894bool Reassociate::runOnFunction(Function &F) { 895 // Recalculate the rank map for F 896 BuildRankMap(F); 897 898 MadeChange = false; 899 for (Function::iterator FI = F.begin(), FE = F.end(); FI != FE; ++FI) 900 ReassociateBB(FI); 901 902 // We are done with the rank map... 903 RankMap.clear(); 904 ValueRankMap.clear(); 905 return MadeChange; 906} 907 908