Reassociate.cpp revision 202375
1//===- Reassociate.cpp - Reassociate binary expressions -------------------===// 2// 3// The LLVM Compiler Infrastructure 4// 5// This file is distributed under the University of Illinois Open Source 6// License. See LICENSE.TXT for details. 7// 8//===----------------------------------------------------------------------===// 9// 10// This pass reassociates commutative expressions in an order that is designed 11// to promote better constant propagation, GCSE, LICM, PRE, etc. 12// 13// For example: 4 + (x + 5) -> x + (4 + 5) 14// 15// In the implementation of this algorithm, constants are assigned rank = 0, 16// function arguments are rank = 1, and other values are assigned ranks 17// corresponding to the reverse post order traversal of current function 18// (starting at 2), which effectively gives values in deep loops higher rank 19// than values not in loops. 20// 21//===----------------------------------------------------------------------===// 22 23#define DEBUG_TYPE "reassociate" 24#include "llvm/Transforms/Scalar.h" 25#include "llvm/Constants.h" 26#include "llvm/DerivedTypes.h" 27#include "llvm/Function.h" 28#include "llvm/Instructions.h" 29#include "llvm/IntrinsicInst.h" 30#include "llvm/Pass.h" 31#include "llvm/Assembly/Writer.h" 32#include "llvm/Support/CFG.h" 33#include "llvm/Support/Debug.h" 34#include "llvm/Support/ValueHandle.h" 35#include "llvm/Support/raw_ostream.h" 36#include "llvm/ADT/PostOrderIterator.h" 37#include "llvm/ADT/Statistic.h" 38#include "llvm/ADT/DenseMap.h" 39#include <algorithm> 40using namespace llvm; 41 42STATISTIC(NumLinear , "Number of insts linearized"); 43STATISTIC(NumChanged, "Number of insts reassociated"); 44STATISTIC(NumAnnihil, "Number of expr tree annihilated"); 45STATISTIC(NumFactor , "Number of multiplies factored"); 46 47namespace { 48 struct ValueEntry { 49 unsigned Rank; 50 Value *Op; 51 ValueEntry(unsigned R, Value *O) : Rank(R), Op(O) {} 52 }; 53 inline bool operator<(const ValueEntry &LHS, const ValueEntry &RHS) { 54 return LHS.Rank > RHS.Rank; // Sort so that highest rank goes to start. 55 } 56} 57 58#ifndef NDEBUG 59/// PrintOps - Print out the expression identified in the Ops list. 60/// 61static void PrintOps(Instruction *I, const SmallVectorImpl<ValueEntry> &Ops) { 62 Module *M = I->getParent()->getParent()->getParent(); 63 dbgs() << Instruction::getOpcodeName(I->getOpcode()) << " " 64 << *Ops[0].Op->getType() << '\t'; 65 for (unsigned i = 0, e = Ops.size(); i != e; ++i) { 66 dbgs() << "[ "; 67 WriteAsOperand(dbgs(), Ops[i].Op, false, M); 68 dbgs() << ", #" << Ops[i].Rank << "] "; 69 } 70} 71#endif 72 73namespace { 74 class Reassociate : public FunctionPass { 75 DenseMap<BasicBlock*, unsigned> RankMap; 76 DenseMap<AssertingVH<>, unsigned> ValueRankMap; 77 bool MadeChange; 78 public: 79 static char ID; // Pass identification, replacement for typeid 80 Reassociate() : FunctionPass(&ID) {} 81 82 bool runOnFunction(Function &F); 83 84 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 85 AU.setPreservesCFG(); 86 } 87 private: 88 void BuildRankMap(Function &F); 89 unsigned getRank(Value *V); 90 Value *ReassociateExpression(BinaryOperator *I); 91 void RewriteExprTree(BinaryOperator *I, SmallVectorImpl<ValueEntry> &Ops, 92 unsigned Idx = 0); 93 Value *OptimizeExpression(BinaryOperator *I, 94 SmallVectorImpl<ValueEntry> &Ops); 95 Value *OptimizeAdd(Instruction *I, SmallVectorImpl<ValueEntry> &Ops); 96 void LinearizeExprTree(BinaryOperator *I, SmallVectorImpl<ValueEntry> &Ops); 97 void LinearizeExpr(BinaryOperator *I); 98 Value *RemoveFactorFromExpression(Value *V, Value *Factor); 99 void ReassociateBB(BasicBlock *BB); 100 101 void RemoveDeadBinaryOp(Value *V); 102 }; 103} 104 105char Reassociate::ID = 0; 106static RegisterPass<Reassociate> X("reassociate", "Reassociate expressions"); 107 108// Public interface to the Reassociate pass 109FunctionPass *llvm::createReassociatePass() { return new Reassociate(); } 110 111void Reassociate::RemoveDeadBinaryOp(Value *V) { 112 Instruction *Op = dyn_cast<Instruction>(V); 113 if (!Op || !isa<BinaryOperator>(Op) || !Op->use_empty()) 114 return; 115 116 Value *LHS = Op->getOperand(0), *RHS = Op->getOperand(1); 117 118 ValueRankMap.erase(Op); 119 Op->eraseFromParent(); 120 RemoveDeadBinaryOp(LHS); 121 RemoveDeadBinaryOp(RHS); 122} 123 124 125static bool isUnmovableInstruction(Instruction *I) { 126 if (I->getOpcode() == Instruction::PHI || 127 I->getOpcode() == Instruction::Alloca || 128 I->getOpcode() == Instruction::Load || 129 I->getOpcode() == Instruction::Invoke || 130 (I->getOpcode() == Instruction::Call && 131 !isa<DbgInfoIntrinsic>(I)) || 132 I->getOpcode() == Instruction::UDiv || 133 I->getOpcode() == Instruction::SDiv || 134 I->getOpcode() == Instruction::FDiv || 135 I->getOpcode() == Instruction::URem || 136 I->getOpcode() == Instruction::SRem || 137 I->getOpcode() == Instruction::FRem) 138 return true; 139 return false; 140} 141 142void Reassociate::BuildRankMap(Function &F) { 143 unsigned i = 2; 144 145 // Assign distinct ranks to function arguments 146 for (Function::arg_iterator I = F.arg_begin(), E = F.arg_end(); I != E; ++I) 147 ValueRankMap[&*I] = ++i; 148 149 ReversePostOrderTraversal<Function*> RPOT(&F); 150 for (ReversePostOrderTraversal<Function*>::rpo_iterator I = RPOT.begin(), 151 E = RPOT.end(); I != E; ++I) { 152 BasicBlock *BB = *I; 153 unsigned BBRank = RankMap[BB] = ++i << 16; 154 155 // Walk the basic block, adding precomputed ranks for any instructions that 156 // we cannot move. This ensures that the ranks for these instructions are 157 // all different in the block. 158 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) 159 if (isUnmovableInstruction(I)) 160 ValueRankMap[&*I] = ++BBRank; 161 } 162} 163 164unsigned Reassociate::getRank(Value *V) { 165 Instruction *I = dyn_cast<Instruction>(V); 166 if (I == 0) { 167 if (isa<Argument>(V)) return ValueRankMap[V]; // Function argument. 168 return 0; // Otherwise it's a global or constant, rank 0. 169 } 170 171 if (unsigned Rank = ValueRankMap[I]) 172 return Rank; // Rank already known? 173 174 // If this is an expression, return the 1+MAX(rank(LHS), rank(RHS)) so that 175 // we can reassociate expressions for code motion! Since we do not recurse 176 // for PHI nodes, we cannot have infinite recursion here, because there 177 // cannot be loops in the value graph that do not go through PHI nodes. 178 unsigned Rank = 0, MaxRank = RankMap[I->getParent()]; 179 for (unsigned i = 0, e = I->getNumOperands(); 180 i != e && Rank != MaxRank; ++i) 181 Rank = std::max(Rank, getRank(I->getOperand(i))); 182 183 // If this is a not or neg instruction, do not count it for rank. This 184 // assures us that X and ~X will have the same rank. 185 if (!I->getType()->isInteger() || 186 (!BinaryOperator::isNot(I) && !BinaryOperator::isNeg(I))) 187 ++Rank; 188 189 //DEBUG(dbgs() << "Calculated Rank[" << V->getName() << "] = " 190 // << Rank << "\n"); 191 192 return ValueRankMap[I] = Rank; 193} 194 195/// isReassociableOp - Return true if V is an instruction of the specified 196/// opcode and if it only has one use. 197static BinaryOperator *isReassociableOp(Value *V, unsigned Opcode) { 198 if ((V->hasOneUse() || V->use_empty()) && isa<Instruction>(V) && 199 cast<Instruction>(V)->getOpcode() == Opcode) 200 return cast<BinaryOperator>(V); 201 return 0; 202} 203 204/// LowerNegateToMultiply - Replace 0-X with X*-1. 205/// 206static Instruction *LowerNegateToMultiply(Instruction *Neg, 207 DenseMap<AssertingVH<>, unsigned> &ValueRankMap) { 208 Constant *Cst = Constant::getAllOnesValue(Neg->getType()); 209 210 Instruction *Res = BinaryOperator::CreateMul(Neg->getOperand(1), Cst, "",Neg); 211 ValueRankMap.erase(Neg); 212 Res->takeName(Neg); 213 Neg->replaceAllUsesWith(Res); 214 Neg->eraseFromParent(); 215 return Res; 216} 217 218// Given an expression of the form '(A+B)+(D+C)', turn it into '(((A+B)+C)+D)'. 219// Note that if D is also part of the expression tree that we recurse to 220// linearize it as well. Besides that case, this does not recurse into A,B, or 221// C. 222void Reassociate::LinearizeExpr(BinaryOperator *I) { 223 BinaryOperator *LHS = cast<BinaryOperator>(I->getOperand(0)); 224 BinaryOperator *RHS = cast<BinaryOperator>(I->getOperand(1)); 225 assert(isReassociableOp(LHS, I->getOpcode()) && 226 isReassociableOp(RHS, I->getOpcode()) && 227 "Not an expression that needs linearization?"); 228 229 DEBUG(dbgs() << "Linear" << *LHS << '\n' << *RHS << '\n' << *I << '\n'); 230 231 // Move the RHS instruction to live immediately before I, avoiding breaking 232 // dominator properties. 233 RHS->moveBefore(I); 234 235 // Move operands around to do the linearization. 236 I->setOperand(1, RHS->getOperand(0)); 237 RHS->setOperand(0, LHS); 238 I->setOperand(0, RHS); 239 240 ++NumLinear; 241 MadeChange = true; 242 DEBUG(dbgs() << "Linearized: " << *I << '\n'); 243 244 // If D is part of this expression tree, tail recurse. 245 if (isReassociableOp(I->getOperand(1), I->getOpcode())) 246 LinearizeExpr(I); 247} 248 249 250/// LinearizeExprTree - Given an associative binary expression tree, traverse 251/// all of the uses putting it into canonical form. This forces a left-linear 252/// form of the the expression (((a+b)+c)+d), and collects information about the 253/// rank of the non-tree operands. 254/// 255/// NOTE: These intentionally destroys the expression tree operands (turning 256/// them into undef values) to reduce #uses of the values. This means that the 257/// caller MUST use something like RewriteExprTree to put the values back in. 258/// 259void Reassociate::LinearizeExprTree(BinaryOperator *I, 260 SmallVectorImpl<ValueEntry> &Ops) { 261 Value *LHS = I->getOperand(0), *RHS = I->getOperand(1); 262 unsigned Opcode = I->getOpcode(); 263 264 // First step, linearize the expression if it is in ((A+B)+(C+D)) form. 265 BinaryOperator *LHSBO = isReassociableOp(LHS, Opcode); 266 BinaryOperator *RHSBO = isReassociableOp(RHS, Opcode); 267 268 // If this is a multiply expression tree and it contains internal negations, 269 // transform them into multiplies by -1 so they can be reassociated. 270 if (I->getOpcode() == Instruction::Mul) { 271 if (!LHSBO && LHS->hasOneUse() && BinaryOperator::isNeg(LHS)) { 272 LHS = LowerNegateToMultiply(cast<Instruction>(LHS), ValueRankMap); 273 LHSBO = isReassociableOp(LHS, Opcode); 274 } 275 if (!RHSBO && RHS->hasOneUse() && BinaryOperator::isNeg(RHS)) { 276 RHS = LowerNegateToMultiply(cast<Instruction>(RHS), ValueRankMap); 277 RHSBO = isReassociableOp(RHS, Opcode); 278 } 279 } 280 281 if (!LHSBO) { 282 if (!RHSBO) { 283 // Neither the LHS or RHS as part of the tree, thus this is a leaf. As 284 // such, just remember these operands and their rank. 285 Ops.push_back(ValueEntry(getRank(LHS), LHS)); 286 Ops.push_back(ValueEntry(getRank(RHS), RHS)); 287 288 // Clear the leaves out. 289 I->setOperand(0, UndefValue::get(I->getType())); 290 I->setOperand(1, UndefValue::get(I->getType())); 291 return; 292 } 293 294 // Turn X+(Y+Z) -> (Y+Z)+X 295 std::swap(LHSBO, RHSBO); 296 std::swap(LHS, RHS); 297 bool Success = !I->swapOperands(); 298 assert(Success && "swapOperands failed"); 299 Success = false; 300 MadeChange = true; 301 } else if (RHSBO) { 302 // Turn (A+B)+(C+D) -> (((A+B)+C)+D). This guarantees the the RHS is not 303 // part of the expression tree. 304 LinearizeExpr(I); 305 LHS = LHSBO = cast<BinaryOperator>(I->getOperand(0)); 306 RHS = I->getOperand(1); 307 RHSBO = 0; 308 } 309 310 // Okay, now we know that the LHS is a nested expression and that the RHS is 311 // not. Perform reassociation. 312 assert(!isReassociableOp(RHS, Opcode) && "LinearizeExpr failed!"); 313 314 // Move LHS right before I to make sure that the tree expression dominates all 315 // values. 316 LHSBO->moveBefore(I); 317 318 // Linearize the expression tree on the LHS. 319 LinearizeExprTree(LHSBO, Ops); 320 321 // Remember the RHS operand and its rank. 322 Ops.push_back(ValueEntry(getRank(RHS), RHS)); 323 324 // Clear the RHS leaf out. 325 I->setOperand(1, UndefValue::get(I->getType())); 326} 327 328// RewriteExprTree - Now that the operands for this expression tree are 329// linearized and optimized, emit them in-order. This function is written to be 330// tail recursive. 331void Reassociate::RewriteExprTree(BinaryOperator *I, 332 SmallVectorImpl<ValueEntry> &Ops, 333 unsigned i) { 334 if (i+2 == Ops.size()) { 335 if (I->getOperand(0) != Ops[i].Op || 336 I->getOperand(1) != Ops[i+1].Op) { 337 Value *OldLHS = I->getOperand(0); 338 DEBUG(dbgs() << "RA: " << *I << '\n'); 339 I->setOperand(0, Ops[i].Op); 340 I->setOperand(1, Ops[i+1].Op); 341 DEBUG(dbgs() << "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(dbgs() << "RA: " << *I << '\n'); 355 I->setOperand(1, Ops[i].Op); 356 DEBUG(dbgs() << "TO: " << *I << '\n'); 357 MadeChange = true; 358 ++NumChanged; 359 } 360 361 BinaryOperator *LHS = cast<BinaryOperator>(I->getOperand(0)); 362 assert(LHS->getOpcode() == I->getOpcode() && 363 "Improper expression tree!"); 364 365 // Compactify the tree instructions together with each other to guarantee 366 // that the expression tree is dominated by all of Ops. 367 LHS->moveBefore(I); 368 RewriteExprTree(LHS, Ops, i+1); 369} 370 371 372 373// NegateValue - Insert instructions before the instruction pointed to by BI, 374// that computes the negative version of the value specified. The negative 375// version of the value is returned, and BI is left pointing at the instruction 376// that should be processed next by the reassociation pass. 377// 378static Value *NegateValue(Value *V, Instruction *BI) { 379 if (Constant *C = dyn_cast<Constant>(V)) 380 return ConstantExpr::getNeg(C); 381 382 // We are trying to expose opportunity for reassociation. One of the things 383 // that we want to do to achieve this is to push a negation as deep into an 384 // expression chain as possible, to expose the add instructions. In practice, 385 // this means that we turn this: 386 // X = -(A+12+C+D) into X = -A + -12 + -C + -D = -12 + -A + -C + -D 387 // so that later, a: Y = 12+X could get reassociated with the -12 to eliminate 388 // the constants. We assume that instcombine will clean up the mess later if 389 // we introduce tons of unnecessary negation instructions. 390 // 391 if (Instruction *I = dyn_cast<Instruction>(V)) 392 if (I->getOpcode() == Instruction::Add && I->hasOneUse()) { 393 // Push the negates through the add. 394 I->setOperand(0, NegateValue(I->getOperand(0), BI)); 395 I->setOperand(1, NegateValue(I->getOperand(1), BI)); 396 397 // We must move the add instruction here, because the neg instructions do 398 // not dominate the old add instruction in general. By moving it, we are 399 // assured that the neg instructions we just inserted dominate the 400 // instruction we are about to insert after them. 401 // 402 I->moveBefore(BI); 403 I->setName(I->getName()+".neg"); 404 return I; 405 } 406 407 // Okay, we need to materialize a negated version of V with an instruction. 408 // Scan the use lists of V to see if we have one already. 409 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E;++UI){ 410 if (!BinaryOperator::isNeg(*UI)) continue; 411 412 // We found one! Now we have to make sure that the definition dominates 413 // this use. We do this by moving it to the entry block (if it is a 414 // non-instruction value) or right after the definition. These negates will 415 // be zapped by reassociate later, so we don't need much finesse here. 416 BinaryOperator *TheNeg = cast<BinaryOperator>(*UI); 417 418 // Verify that the negate is in this function, V might be a constant expr. 419 if (TheNeg->getParent()->getParent() != BI->getParent()->getParent()) 420 continue; 421 422 BasicBlock::iterator InsertPt; 423 if (Instruction *InstInput = dyn_cast<Instruction>(V)) { 424 if (InvokeInst *II = dyn_cast<InvokeInst>(InstInput)) { 425 InsertPt = II->getNormalDest()->begin(); 426 } else { 427 InsertPt = InstInput; 428 ++InsertPt; 429 } 430 while (isa<PHINode>(InsertPt)) ++InsertPt; 431 } else { 432 InsertPt = TheNeg->getParent()->getParent()->getEntryBlock().begin(); 433 } 434 TheNeg->moveBefore(InsertPt); 435 return TheNeg; 436 } 437 438 // Insert a 'neg' instruction that subtracts the value from zero to get the 439 // negation. 440 return BinaryOperator::CreateNeg(V, V->getName() + ".neg", BI); 441} 442 443/// ShouldBreakUpSubtract - Return true if we should break up this subtract of 444/// X-Y into (X + -Y). 445static bool ShouldBreakUpSubtract(Instruction *Sub) { 446 // If this is a negation, we can't split it up! 447 if (BinaryOperator::isNeg(Sub)) 448 return false; 449 450 // Don't bother to break this up unless either the LHS is an associable add or 451 // subtract or if this is only used by one. 452 if (isReassociableOp(Sub->getOperand(0), Instruction::Add) || 453 isReassociableOp(Sub->getOperand(0), Instruction::Sub)) 454 return true; 455 if (isReassociableOp(Sub->getOperand(1), Instruction::Add) || 456 isReassociableOp(Sub->getOperand(1), Instruction::Sub)) 457 return true; 458 if (Sub->hasOneUse() && 459 (isReassociableOp(Sub->use_back(), Instruction::Add) || 460 isReassociableOp(Sub->use_back(), Instruction::Sub))) 461 return true; 462 463 return false; 464} 465 466/// BreakUpSubtract - If we have (X-Y), and if either X is an add, or if this is 467/// only used by an add, transform this into (X+(0-Y)) to promote better 468/// reassociation. 469static Instruction *BreakUpSubtract(Instruction *Sub, 470 DenseMap<AssertingVH<>, unsigned> &ValueRankMap) { 471 // Convert a subtract into an add and a neg instruction. This allows sub 472 // instructions to be commuted with other add instructions. 473 // 474 // Calculate the negative value of Operand 1 of the sub instruction, 475 // and set it as the RHS of the add instruction we just made. 476 // 477 Value *NegVal = NegateValue(Sub->getOperand(1), Sub); 478 Instruction *New = 479 BinaryOperator::CreateAdd(Sub->getOperand(0), NegVal, "", Sub); 480 New->takeName(Sub); 481 482 // Everyone now refers to the add instruction. 483 ValueRankMap.erase(Sub); 484 Sub->replaceAllUsesWith(New); 485 Sub->eraseFromParent(); 486 487 DEBUG(dbgs() << "Negated: " << *New << '\n'); 488 return New; 489} 490 491/// ConvertShiftToMul - If this is a shift of a reassociable multiply or is used 492/// by one, change this into a multiply by a constant to assist with further 493/// reassociation. 494static Instruction *ConvertShiftToMul(Instruction *Shl, 495 DenseMap<AssertingVH<>, unsigned> &ValueRankMap) { 496 // If an operand of this shift is a reassociable multiply, or if the shift 497 // is used by a reassociable multiply or add, turn into a multiply. 498 if (isReassociableOp(Shl->getOperand(0), Instruction::Mul) || 499 (Shl->hasOneUse() && 500 (isReassociableOp(Shl->use_back(), Instruction::Mul) || 501 isReassociableOp(Shl->use_back(), Instruction::Add)))) { 502 Constant *MulCst = ConstantInt::get(Shl->getType(), 1); 503 MulCst = ConstantExpr::getShl(MulCst, cast<Constant>(Shl->getOperand(1))); 504 505 Instruction *Mul = 506 BinaryOperator::CreateMul(Shl->getOperand(0), MulCst, "", Shl); 507 ValueRankMap.erase(Shl); 508 Mul->takeName(Shl); 509 Shl->replaceAllUsesWith(Mul); 510 Shl->eraseFromParent(); 511 return Mul; 512 } 513 return 0; 514} 515 516// Scan backwards and forwards among values with the same rank as element i to 517// see if X exists. If X does not exist, return i. This is useful when 518// scanning for 'x' when we see '-x' because they both get the same rank. 519static unsigned FindInOperandList(SmallVectorImpl<ValueEntry> &Ops, unsigned i, 520 Value *X) { 521 unsigned XRank = Ops[i].Rank; 522 unsigned e = Ops.size(); 523 for (unsigned j = i+1; j != e && Ops[j].Rank == XRank; ++j) 524 if (Ops[j].Op == X) 525 return j; 526 // Scan backwards. 527 for (unsigned j = i-1; j != ~0U && Ops[j].Rank == XRank; --j) 528 if (Ops[j].Op == X) 529 return j; 530 return i; 531} 532 533/// EmitAddTreeOfValues - Emit a tree of add instructions, summing Ops together 534/// and returning the result. Insert the tree before I. 535static Value *EmitAddTreeOfValues(Instruction *I, SmallVectorImpl<Value*> &Ops){ 536 if (Ops.size() == 1) return Ops.back(); 537 538 Value *V1 = Ops.back(); 539 Ops.pop_back(); 540 Value *V2 = EmitAddTreeOfValues(I, Ops); 541 return BinaryOperator::CreateAdd(V2, V1, "tmp", I); 542} 543 544/// RemoveFactorFromExpression - If V is an expression tree that is a 545/// multiplication sequence, and if this sequence contains a multiply by Factor, 546/// remove Factor from the tree and return the new tree. 547Value *Reassociate::RemoveFactorFromExpression(Value *V, Value *Factor) { 548 BinaryOperator *BO = isReassociableOp(V, Instruction::Mul); 549 if (!BO) return 0; 550 551 SmallVector<ValueEntry, 8> Factors; 552 LinearizeExprTree(BO, Factors); 553 554 bool FoundFactor = false; 555 bool NeedsNegate = false; 556 for (unsigned i = 0, e = Factors.size(); i != e; ++i) { 557 if (Factors[i].Op == Factor) { 558 FoundFactor = true; 559 Factors.erase(Factors.begin()+i); 560 break; 561 } 562 563 // If this is a negative version of this factor, remove it. 564 if (ConstantInt *FC1 = dyn_cast<ConstantInt>(Factor)) 565 if (ConstantInt *FC2 = dyn_cast<ConstantInt>(Factors[i].Op)) 566 if (FC1->getValue() == -FC2->getValue()) { 567 FoundFactor = NeedsNegate = true; 568 Factors.erase(Factors.begin()+i); 569 break; 570 } 571 } 572 573 if (!FoundFactor) { 574 // Make sure to restore the operands to the expression tree. 575 RewriteExprTree(BO, Factors); 576 return 0; 577 } 578 579 BasicBlock::iterator InsertPt = BO; ++InsertPt; 580 581 // If this was just a single multiply, remove the multiply and return the only 582 // remaining operand. 583 if (Factors.size() == 1) { 584 ValueRankMap.erase(BO); 585 BO->eraseFromParent(); 586 V = Factors[0].Op; 587 } else { 588 RewriteExprTree(BO, Factors); 589 V = BO; 590 } 591 592 if (NeedsNegate) 593 V = BinaryOperator::CreateNeg(V, "neg", InsertPt); 594 595 return V; 596} 597 598/// FindSingleUseMultiplyFactors - If V is a single-use multiply, recursively 599/// add its operands as factors, otherwise add V to the list of factors. 600static void FindSingleUseMultiplyFactors(Value *V, 601 SmallVectorImpl<Value*> &Factors) { 602 BinaryOperator *BO; 603 if ((!V->hasOneUse() && !V->use_empty()) || 604 !(BO = dyn_cast<BinaryOperator>(V)) || 605 BO->getOpcode() != Instruction::Mul) { 606 Factors.push_back(V); 607 return; 608 } 609 610 // Otherwise, add the LHS and RHS to the list of factors. 611 FindSingleUseMultiplyFactors(BO->getOperand(1), Factors); 612 FindSingleUseMultiplyFactors(BO->getOperand(0), Factors); 613} 614 615/// OptimizeAndOrXor - Optimize a series of operands to an 'and', 'or', or 'xor' 616/// instruction. This optimizes based on identities. If it can be reduced to 617/// a single Value, it is returned, otherwise the Ops list is mutated as 618/// necessary. 619static Value *OptimizeAndOrXor(unsigned Opcode, 620 SmallVectorImpl<ValueEntry> &Ops) { 621 // Scan the operand lists looking for X and ~X pairs, along with X,X pairs. 622 // If we find any, we can simplify the expression. X&~X == 0, X|~X == -1. 623 for (unsigned i = 0, e = Ops.size(); i != e; ++i) { 624 // First, check for X and ~X in the operand list. 625 assert(i < Ops.size()); 626 if (BinaryOperator::isNot(Ops[i].Op)) { // Cannot occur for ^. 627 Value *X = BinaryOperator::getNotArgument(Ops[i].Op); 628 unsigned FoundX = FindInOperandList(Ops, i, X); 629 if (FoundX != i) { 630 if (Opcode == Instruction::And) // ...&X&~X = 0 631 return Constant::getNullValue(X->getType()); 632 633 if (Opcode == Instruction::Or) // ...|X|~X = -1 634 return Constant::getAllOnesValue(X->getType()); 635 } 636 } 637 638 // Next, check for duplicate pairs of values, which we assume are next to 639 // each other, due to our sorting criteria. 640 assert(i < Ops.size()); 641 if (i+1 != Ops.size() && Ops[i+1].Op == Ops[i].Op) { 642 if (Opcode == Instruction::And || Opcode == Instruction::Or) { 643 // Drop duplicate values for And and Or. 644 Ops.erase(Ops.begin()+i); 645 --i; --e; 646 ++NumAnnihil; 647 continue; 648 } 649 650 // Drop pairs of values for Xor. 651 assert(Opcode == Instruction::Xor); 652 if (e == 2) 653 return Constant::getNullValue(Ops[0].Op->getType()); 654 655 // Y ^ X^X -> Y 656 Ops.erase(Ops.begin()+i, Ops.begin()+i+2); 657 i -= 1; e -= 2; 658 ++NumAnnihil; 659 } 660 } 661 return 0; 662} 663 664/// OptimizeAdd - Optimize a series of operands to an 'add' instruction. This 665/// optimizes based on identities. If it can be reduced to a single Value, it 666/// is returned, otherwise the Ops list is mutated as necessary. 667Value *Reassociate::OptimizeAdd(Instruction *I, 668 SmallVectorImpl<ValueEntry> &Ops) { 669 // Scan the operand lists looking for X and -X pairs. If we find any, we 670 // can simplify the expression. X+-X == 0. While we're at it, scan for any 671 // duplicates. We want to canonicalize Y+Y+Y+Z -> 3*Y+Z. 672 // 673 // TODO: We could handle "X + ~X" -> "-1" if we wanted, since "-X = ~X+1". 674 // 675 for (unsigned i = 0, e = Ops.size(); i != e; ++i) { 676 Value *TheOp = Ops[i].Op; 677 // Check to see if we've seen this operand before. If so, we factor all 678 // instances of the operand together. Due to our sorting criteria, we know 679 // that these need to be next to each other in the vector. 680 if (i+1 != Ops.size() && Ops[i+1].Op == TheOp) { 681 // Rescan the list, remove all instances of this operand from the expr. 682 unsigned NumFound = 0; 683 do { 684 Ops.erase(Ops.begin()+i); 685 ++NumFound; 686 } while (i != Ops.size() && Ops[i].Op == TheOp); 687 688 DEBUG(errs() << "\nFACTORING [" << NumFound << "]: " << *TheOp << '\n'); 689 ++NumFactor; 690 691 // Insert a new multiply. 692 Value *Mul = ConstantInt::get(cast<IntegerType>(I->getType()), NumFound); 693 Mul = BinaryOperator::CreateMul(TheOp, Mul, "factor", I); 694 695 // Now that we have inserted a multiply, optimize it. This allows us to 696 // handle cases that require multiple factoring steps, such as this: 697 // (X*2) + (X*2) + (X*2) -> (X*2)*3 -> X*6 698 Mul = ReassociateExpression(cast<BinaryOperator>(Mul)); 699 700 // If every add operand was a duplicate, return the multiply. 701 if (Ops.empty()) 702 return Mul; 703 704 // Otherwise, we had some input that didn't have the dupe, such as 705 // "A + A + B" -> "A*2 + B". Add the new multiply to the list of 706 // things being added by this operation. 707 Ops.insert(Ops.begin(), ValueEntry(getRank(Mul), Mul)); 708 709 --i; 710 e = Ops.size(); 711 continue; 712 } 713 714 // Check for X and -X in the operand list. 715 if (!BinaryOperator::isNeg(TheOp)) 716 continue; 717 718 Value *X = BinaryOperator::getNegArgument(TheOp); 719 unsigned FoundX = FindInOperandList(Ops, i, X); 720 if (FoundX == i) 721 continue; 722 723 // Remove X and -X from the operand list. 724 if (Ops.size() == 2) 725 return Constant::getNullValue(X->getType()); 726 727 Ops.erase(Ops.begin()+i); 728 if (i < FoundX) 729 --FoundX; 730 else 731 --i; // Need to back up an extra one. 732 Ops.erase(Ops.begin()+FoundX); 733 ++NumAnnihil; 734 --i; // Revisit element. 735 e -= 2; // Removed two elements. 736 } 737 738 // Scan the operand list, checking to see if there are any common factors 739 // between operands. Consider something like A*A+A*B*C+D. We would like to 740 // reassociate this to A*(A+B*C)+D, which reduces the number of multiplies. 741 // To efficiently find this, we count the number of times a factor occurs 742 // for any ADD operands that are MULs. 743 DenseMap<Value*, unsigned> FactorOccurrences; 744 745 // Keep track of each multiply we see, to avoid triggering on (X*4)+(X*4) 746 // where they are actually the same multiply. 747 unsigned MaxOcc = 0; 748 Value *MaxOccVal = 0; 749 for (unsigned i = 0, e = Ops.size(); i != e; ++i) { 750 BinaryOperator *BOp = dyn_cast<BinaryOperator>(Ops[i].Op); 751 if (BOp == 0 || BOp->getOpcode() != Instruction::Mul || !BOp->use_empty()) 752 continue; 753 754 // Compute all of the factors of this added value. 755 SmallVector<Value*, 8> Factors; 756 FindSingleUseMultiplyFactors(BOp, Factors); 757 assert(Factors.size() > 1 && "Bad linearize!"); 758 759 // Add one to FactorOccurrences for each unique factor in this op. 760 SmallPtrSet<Value*, 8> Duplicates; 761 for (unsigned i = 0, e = Factors.size(); i != e; ++i) { 762 Value *Factor = Factors[i]; 763 if (!Duplicates.insert(Factor)) continue; 764 765 unsigned Occ = ++FactorOccurrences[Factor]; 766 if (Occ > MaxOcc) { MaxOcc = Occ; MaxOccVal = Factor; } 767 768 // If Factor is a negative constant, add the negated value as a factor 769 // because we can percolate the negate out. Watch for minint, which 770 // cannot be positivified. 771 if (ConstantInt *CI = dyn_cast<ConstantInt>(Factor)) 772 if (CI->getValue().isNegative() && !CI->getValue().isMinSignedValue()) { 773 Factor = ConstantInt::get(CI->getContext(), -CI->getValue()); 774 assert(!Duplicates.count(Factor) && 775 "Shouldn't have two constant factors, missed a canonicalize"); 776 777 unsigned Occ = ++FactorOccurrences[Factor]; 778 if (Occ > MaxOcc) { MaxOcc = Occ; MaxOccVal = Factor; } 779 } 780 } 781 } 782 783 // If any factor occurred more than one time, we can pull it out. 784 if (MaxOcc > 1) { 785 DEBUG(errs() << "\nFACTORING [" << MaxOcc << "]: " << *MaxOccVal << '\n'); 786 ++NumFactor; 787 788 // Create a new instruction that uses the MaxOccVal twice. If we don't do 789 // this, we could otherwise run into situations where removing a factor 790 // from an expression will drop a use of maxocc, and this can cause 791 // RemoveFactorFromExpression on successive values to behave differently. 792 Instruction *DummyInst = BinaryOperator::CreateAdd(MaxOccVal, MaxOccVal); 793 SmallVector<Value*, 4> NewMulOps; 794 for (unsigned i = 0, e = Ops.size(); i != e; ++i) { 795 // Only try to remove factors from expressions we're allowed to. 796 BinaryOperator *BOp = dyn_cast<BinaryOperator>(Ops[i].Op); 797 if (BOp == 0 || BOp->getOpcode() != Instruction::Mul || !BOp->use_empty()) 798 continue; 799 800 if (Value *V = RemoveFactorFromExpression(Ops[i].Op, MaxOccVal)) { 801 NewMulOps.push_back(V); 802 Ops.erase(Ops.begin()+i); 803 --i; --e; 804 } 805 } 806 807 // No need for extra uses anymore. 808 delete DummyInst; 809 810 unsigned NumAddedValues = NewMulOps.size(); 811 Value *V = EmitAddTreeOfValues(I, NewMulOps); 812 813 // Now that we have inserted the add tree, optimize it. This allows us to 814 // handle cases that require multiple factoring steps, such as this: 815 // A*A*B + A*A*C --> A*(A*B+A*C) --> A*(A*(B+C)) 816 assert(NumAddedValues > 1 && "Each occurrence should contribute a value"); 817 (void)NumAddedValues; 818 V = ReassociateExpression(cast<BinaryOperator>(V)); 819 820 // Create the multiply. 821 Value *V2 = BinaryOperator::CreateMul(V, MaxOccVal, "tmp", I); 822 823 // Rerun associate on the multiply in case the inner expression turned into 824 // a multiply. We want to make sure that we keep things in canonical form. 825 V2 = ReassociateExpression(cast<BinaryOperator>(V2)); 826 827 // If every add operand included the factor (e.g. "A*B + A*C"), then the 828 // entire result expression is just the multiply "A*(B+C)". 829 if (Ops.empty()) 830 return V2; 831 832 // Otherwise, we had some input that didn't have the factor, such as 833 // "A*B + A*C + D" -> "A*(B+C) + D". Add the new multiply to the list of 834 // things being added by this operation. 835 Ops.insert(Ops.begin(), ValueEntry(getRank(V2), V2)); 836 } 837 838 return 0; 839} 840 841Value *Reassociate::OptimizeExpression(BinaryOperator *I, 842 SmallVectorImpl<ValueEntry> &Ops) { 843 // Now that we have the linearized expression tree, try to optimize it. 844 // Start by folding any constants that we found. 845 bool IterateOptimization = false; 846 if (Ops.size() == 1) return Ops[0].Op; 847 848 unsigned Opcode = I->getOpcode(); 849 850 if (Constant *V1 = dyn_cast<Constant>(Ops[Ops.size()-2].Op)) 851 if (Constant *V2 = dyn_cast<Constant>(Ops.back().Op)) { 852 Ops.pop_back(); 853 Ops.back().Op = ConstantExpr::get(Opcode, V1, V2); 854 return OptimizeExpression(I, Ops); 855 } 856 857 // Check for destructive annihilation due to a constant being used. 858 if (ConstantInt *CstVal = dyn_cast<ConstantInt>(Ops.back().Op)) 859 switch (Opcode) { 860 default: break; 861 case Instruction::And: 862 if (CstVal->isZero()) // X & 0 -> 0 863 return CstVal; 864 if (CstVal->isAllOnesValue()) // X & -1 -> X 865 Ops.pop_back(); 866 break; 867 case Instruction::Mul: 868 if (CstVal->isZero()) { // X * 0 -> 0 869 ++NumAnnihil; 870 return CstVal; 871 } 872 873 if (cast<ConstantInt>(CstVal)->isOne()) 874 Ops.pop_back(); // X * 1 -> X 875 break; 876 case Instruction::Or: 877 if (CstVal->isAllOnesValue()) // X | -1 -> -1 878 return CstVal; 879 // FALLTHROUGH! 880 case Instruction::Add: 881 case Instruction::Xor: 882 if (CstVal->isZero()) // X [|^+] 0 -> X 883 Ops.pop_back(); 884 break; 885 } 886 if (Ops.size() == 1) return Ops[0].Op; 887 888 // Handle destructive annihilation due to identities between elements in the 889 // argument list here. 890 switch (Opcode) { 891 default: break; 892 case Instruction::And: 893 case Instruction::Or: 894 case Instruction::Xor: { 895 unsigned NumOps = Ops.size(); 896 if (Value *Result = OptimizeAndOrXor(Opcode, Ops)) 897 return Result; 898 IterateOptimization |= Ops.size() != NumOps; 899 break; 900 } 901 902 case Instruction::Add: { 903 unsigned NumOps = Ops.size(); 904 if (Value *Result = OptimizeAdd(I, Ops)) 905 return Result; 906 IterateOptimization |= Ops.size() != NumOps; 907 } 908 909 break; 910 //case Instruction::Mul: 911 } 912 913 if (IterateOptimization) 914 return OptimizeExpression(I, Ops); 915 return 0; 916} 917 918 919/// ReassociateBB - Inspect all of the instructions in this basic block, 920/// reassociating them as we go. 921void Reassociate::ReassociateBB(BasicBlock *BB) { 922 for (BasicBlock::iterator BBI = BB->begin(); BBI != BB->end(); ) { 923 Instruction *BI = BBI++; 924 if (BI->getOpcode() == Instruction::Shl && 925 isa<ConstantInt>(BI->getOperand(1))) 926 if (Instruction *NI = ConvertShiftToMul(BI, ValueRankMap)) { 927 MadeChange = true; 928 BI = NI; 929 } 930 931 // Reject cases where it is pointless to do this. 932 if (!isa<BinaryOperator>(BI) || BI->getType()->isFloatingPoint() || 933 isa<VectorType>(BI->getType())) 934 continue; // Floating point ops are not associative. 935 936 // If this is a subtract instruction which is not already in negate form, 937 // see if we can convert it to X+-Y. 938 if (BI->getOpcode() == Instruction::Sub) { 939 if (ShouldBreakUpSubtract(BI)) { 940 BI = BreakUpSubtract(BI, ValueRankMap); 941 // Reset the BBI iterator in case BreakUpSubtract changed the 942 // instruction it points to. 943 BBI = BI; 944 ++BBI; 945 MadeChange = true; 946 } else if (BinaryOperator::isNeg(BI)) { 947 // Otherwise, this is a negation. See if the operand is a multiply tree 948 // and if this is not an inner node of a multiply tree. 949 if (isReassociableOp(BI->getOperand(1), Instruction::Mul) && 950 (!BI->hasOneUse() || 951 !isReassociableOp(BI->use_back(), Instruction::Mul))) { 952 BI = LowerNegateToMultiply(BI, ValueRankMap); 953 MadeChange = true; 954 } 955 } 956 } 957 958 // If this instruction is a commutative binary operator, process it. 959 if (!BI->isAssociative()) continue; 960 BinaryOperator *I = cast<BinaryOperator>(BI); 961 962 // If this is an interior node of a reassociable tree, ignore it until we 963 // get to the root of the tree, to avoid N^2 analysis. 964 if (I->hasOneUse() && isReassociableOp(I->use_back(), I->getOpcode())) 965 continue; 966 967 // If this is an add tree that is used by a sub instruction, ignore it 968 // until we process the subtract. 969 if (I->hasOneUse() && I->getOpcode() == Instruction::Add && 970 cast<Instruction>(I->use_back())->getOpcode() == Instruction::Sub) 971 continue; 972 973 ReassociateExpression(I); 974 } 975} 976 977Value *Reassociate::ReassociateExpression(BinaryOperator *I) { 978 979 // First, walk the expression tree, linearizing the tree, collecting the 980 // operand information. 981 SmallVector<ValueEntry, 8> Ops; 982 LinearizeExprTree(I, Ops); 983 984 DEBUG(dbgs() << "RAIn:\t"; PrintOps(I, Ops); dbgs() << '\n'); 985 986 // Now that we have linearized the tree to a list and have gathered all of 987 // the operands and their ranks, sort the operands by their rank. Use a 988 // stable_sort so that values with equal ranks will have their relative 989 // positions maintained (and so the compiler is deterministic). Note that 990 // this sorts so that the highest ranking values end up at the beginning of 991 // the vector. 992 std::stable_sort(Ops.begin(), Ops.end()); 993 994 // OptimizeExpression - Now that we have the expression tree in a convenient 995 // sorted form, optimize it globally if possible. 996 if (Value *V = OptimizeExpression(I, Ops)) { 997 // This expression tree simplified to something that isn't a tree, 998 // eliminate it. 999 DEBUG(dbgs() << "Reassoc to scalar: " << *V << '\n'); 1000 I->replaceAllUsesWith(V); 1001 RemoveDeadBinaryOp(I); 1002 ++NumAnnihil; 1003 return V; 1004 } 1005 1006 // We want to sink immediates as deeply as possible except in the case where 1007 // this is a multiply tree used only by an add, and the immediate is a -1. 1008 // In this case we reassociate to put the negation on the outside so that we 1009 // can fold the negation into the add: (-X)*Y + Z -> Z-X*Y 1010 if (I->getOpcode() == Instruction::Mul && I->hasOneUse() && 1011 cast<Instruction>(I->use_back())->getOpcode() == Instruction::Add && 1012 isa<ConstantInt>(Ops.back().Op) && 1013 cast<ConstantInt>(Ops.back().Op)->isAllOnesValue()) { 1014 ValueEntry Tmp = Ops.pop_back_val(); 1015 Ops.insert(Ops.begin(), Tmp); 1016 } 1017 1018 DEBUG(dbgs() << "RAOut:\t"; PrintOps(I, Ops); dbgs() << '\n'); 1019 1020 if (Ops.size() == 1) { 1021 // This expression tree simplified to something that isn't a tree, 1022 // eliminate it. 1023 I->replaceAllUsesWith(Ops[0].Op); 1024 RemoveDeadBinaryOp(I); 1025 return Ops[0].Op; 1026 } 1027 1028 // Now that we ordered and optimized the expressions, splat them back into 1029 // the expression tree, removing any unneeded nodes. 1030 RewriteExprTree(I, Ops); 1031 return I; 1032} 1033 1034 1035bool Reassociate::runOnFunction(Function &F) { 1036 // Recalculate the rank map for F 1037 BuildRankMap(F); 1038 1039 MadeChange = false; 1040 for (Function::iterator FI = F.begin(), FE = F.end(); FI != FE; ++FI) 1041 ReassociateBB(FI); 1042 1043 // We are done with the rank map. 1044 RankMap.clear(); 1045 ValueRankMap.clear(); 1046 return MadeChange; 1047} 1048 1049