Reassociate.cpp revision 193323
1251876Speter//===- Reassociate.cpp - Reassociate binary expressions -------------------===// 2251876Speter// 3251876Speter// The LLVM Compiler Infrastructure 4251876Speter// 5251876Speter// This file is distributed under the University of Illinois Open Source 6251876Speter// License. See LICENSE.TXT for details. 7251876Speter// 8251876Speter//===----------------------------------------------------------------------===// 9251876Speter// 10251876Speter// This pass reassociates commutative expressions in an order that is designed 11251876Speter// to promote better constant propagation, GCSE, LICM, PRE... 12251876Speter// 13251876Speter// For example: 4 + (x + 5) -> x + (4 + 5) 14251876Speter// 15251876Speter// In the implementation of this algorithm, constants are assigned rank = 0, 16251876Speter// function arguments are rank = 1, and other values are assigned ranks 17251876Speter// corresponding to the reverse post order traversal of current function 18251876Speter// (starting at 2), which effectively gives values in deep loops higher rank 19251876Speter// than values not in loops. 20251876Speter// 21251876Speter//===----------------------------------------------------------------------===// 22251876Speter 23251876Speter#define DEBUG_TYPE "reassociate" 24251876Speter#include "llvm/Transforms/Scalar.h" 25251876Speter#include "llvm/Constants.h" 26251876Speter#include "llvm/DerivedTypes.h" 27251876Speter#include "llvm/Function.h" 28251876Speter#include "llvm/Instructions.h" 29251876Speter#include "llvm/IntrinsicInst.h" 30251876Speter#include "llvm/Pass.h" 31251876Speter#include "llvm/Assembly/Writer.h" 32251876Speter#include "llvm/Support/CFG.h" 33251876Speter#include "llvm/Support/Compiler.h" 34251876Speter#include "llvm/Support/Debug.h" 35251876Speter#include "llvm/Support/ValueHandle.h" 36251876Speter#include "llvm/ADT/PostOrderIterator.h" 37251876Speter#include "llvm/ADT/Statistic.h" 38251876Speter#include <algorithm> 39251876Speter#include <map> 40251876Speterusing namespace llvm; 41251876Speter 42251876SpeterSTATISTIC(NumLinear , "Number of insts linearized"); 43251876SpeterSTATISTIC(NumChanged, "Number of insts reassociated"); 44251876SpeterSTATISTIC(NumAnnihil, "Number of expr tree annihilated"); 45251876SpeterSTATISTIC(NumFactor , "Number of multiplies factored"); 46251876Speter 47251876Speternamespace { 48251876Speter struct VISIBILITY_HIDDEN ValueEntry { 49251876Speter unsigned Rank; 50251876Speter Value *Op; 51251876Speter ValueEntry(unsigned R, Value *O) : Rank(R), Op(O) {} 52251876Speter }; 53251876Speter inline bool operator<(const ValueEntry &LHS, const ValueEntry &RHS) { 54253734Speter return LHS.Rank > RHS.Rank; // Sort so that highest rank goes to start. 55253734Speter } 56253734Speter} 57253734Speter 58253734Speter#ifndef NDEBUG 59251876Speter/// PrintOps - Print out the expression identified in the Ops list. 60251876Speter/// 61251876Speterstatic void PrintOps(Instruction *I, const std::vector<ValueEntry> &Ops) { 62251876Speter Module *M = I->getParent()->getParent()->getParent(); 63251876Speter cerr << Instruction::getOpcodeName(I->getOpcode()) << " " 64251876Speter << *Ops[0].Op->getType(); 65251876Speter for (unsigned i = 0, e = Ops.size(); i != e; ++i) { 66251876Speter WriteAsOperand(*cerr.stream() << " ", Ops[i].Op, false, M); 67251876Speter cerr << "," << Ops[i].Rank; 68251876Speter } 69251876Speter} 70251876Speter#endif 71251876Speter 72251876Speternamespace { 73251876Speter class VISIBILITY_HIDDEN Reassociate : public FunctionPass { 74251876Speter std::map<BasicBlock*, unsigned> RankMap; 75251876Speter std::map<AssertingVH<>, unsigned> ValueRankMap; 76251876Speter bool MadeChange; 77251876Speter public: 78251876Speter static char ID; // Pass identification, replacement for typeid 79251876Speter Reassociate() : FunctionPass(&ID) {} 80251876Speter 81251876Speter bool runOnFunction(Function &F); 82251876Speter 83251876Speter virtual void getAnalysisUsage(AnalysisUsage &AU) const { 84251876Speter AU.setPreservesCFG(); 85251876Speter } 86251876Speter private: 87251876Speter void BuildRankMap(Function &F); 88251876Speter unsigned getRank(Value *V); 89251876Speter void ReassociateExpression(BinaryOperator *I); 90251876Speter void RewriteExprTree(BinaryOperator *I, std::vector<ValueEntry> &Ops, 91251876Speter unsigned Idx = 0); 92251876Speter Value *OptimizeExpression(BinaryOperator *I, std::vector<ValueEntry> &Ops); 93251876Speter void LinearizeExprTree(BinaryOperator *I, std::vector<ValueEntry> &Ops); 94251876Speter void LinearizeExpr(BinaryOperator *I); 95251876Speter Value *RemoveFactorFromExpression(Value *V, Value *Factor); 96251876Speter void ReassociateBB(BasicBlock *BB); 97251876Speter 98251876Speter void RemoveDeadBinaryOp(Value *V); 99251876Speter }; 100251876Speter} 101251876Speter 102251876Speterchar Reassociate::ID = 0; 103251876Speterstatic RegisterPass<Reassociate> X("reassociate", "Reassociate expressions"); 104251876Speter 105251876Speter// Public interface to the Reassociate pass 106251876SpeterFunctionPass *llvm::createReassociatePass() { return new Reassociate(); } 107251876Speter 108251876Spetervoid Reassociate::RemoveDeadBinaryOp(Value *V) { 109251876Speter Instruction *Op = dyn_cast<Instruction>(V); 110251876Speter if (!Op || !isa<BinaryOperator>(Op) || !isa<CmpInst>(Op) || !Op->use_empty()) 111251876Speter return; 112251876Speter 113251876Speter Value *LHS = Op->getOperand(0), *RHS = Op->getOperand(1); 114 RemoveDeadBinaryOp(LHS); 115 RemoveDeadBinaryOp(RHS); 116} 117 118 119static bool isUnmovableInstruction(Instruction *I) { 120 if (I->getOpcode() == Instruction::PHI || 121 I->getOpcode() == Instruction::Alloca || 122 I->getOpcode() == Instruction::Load || 123 I->getOpcode() == Instruction::Malloc || 124 I->getOpcode() == Instruction::Invoke || 125 (I->getOpcode() == Instruction::Call && 126 !isa<DbgInfoIntrinsic>(I)) || 127 I->getOpcode() == Instruction::UDiv || 128 I->getOpcode() == Instruction::SDiv || 129 I->getOpcode() == Instruction::FDiv || 130 I->getOpcode() == Instruction::URem || 131 I->getOpcode() == Instruction::SRem || 132 I->getOpcode() == Instruction::FRem) 133 return true; 134 return false; 135} 136 137void Reassociate::BuildRankMap(Function &F) { 138 unsigned i = 2; 139 140 // Assign distinct ranks to function arguments 141 for (Function::arg_iterator I = F.arg_begin(), E = F.arg_end(); I != E; ++I) 142 ValueRankMap[&*I] = ++i; 143 144 ReversePostOrderTraversal<Function*> RPOT(&F); 145 for (ReversePostOrderTraversal<Function*>::rpo_iterator I = RPOT.begin(), 146 E = RPOT.end(); I != E; ++I) { 147 BasicBlock *BB = *I; 148 unsigned BBRank = RankMap[BB] = ++i << 16; 149 150 // Walk the basic block, adding precomputed ranks for any instructions that 151 // we cannot move. This ensures that the ranks for these instructions are 152 // all different in the block. 153 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) 154 if (isUnmovableInstruction(I)) 155 ValueRankMap[&*I] = ++BBRank; 156 } 157} 158 159unsigned Reassociate::getRank(Value *V) { 160 if (isa<Argument>(V)) return ValueRankMap[V]; // Function argument... 161 162 Instruction *I = dyn_cast<Instruction>(V); 163 if (I == 0) return 0; // Otherwise it's a global or constant, rank 0. 164 165 unsigned &CachedRank = ValueRankMap[I]; 166 if (CachedRank) return CachedRank; // Rank already known? 167 168 // If this is an expression, return the 1+MAX(rank(LHS), rank(RHS)) so that 169 // we can reassociate expressions for code motion! Since we do not recurse 170 // for PHI nodes, we cannot have infinite recursion here, because there 171 // cannot be loops in the value graph that do not go through PHI nodes. 172 unsigned Rank = 0, MaxRank = RankMap[I->getParent()]; 173 for (unsigned i = 0, e = I->getNumOperands(); 174 i != e && Rank != MaxRank; ++i) 175 Rank = std::max(Rank, getRank(I->getOperand(i))); 176 177 // If this is a not or neg instruction, do not count it for rank. This 178 // assures us that X and ~X will have the same rank. 179 if (!I->getType()->isInteger() || 180 (!BinaryOperator::isNot(I) && !BinaryOperator::isNeg(I))) 181 ++Rank; 182 183 //DOUT << "Calculated Rank[" << V->getName() << "] = " 184 // << Rank << "\n"; 185 186 return CachedRank = Rank; 187} 188 189/// isReassociableOp - Return true if V is an instruction of the specified 190/// opcode and if it only has one use. 191static BinaryOperator *isReassociableOp(Value *V, unsigned Opcode) { 192 if ((V->hasOneUse() || V->use_empty()) && isa<Instruction>(V) && 193 cast<Instruction>(V)->getOpcode() == Opcode) 194 return cast<BinaryOperator>(V); 195 return 0; 196} 197 198/// LowerNegateToMultiply - Replace 0-X with X*-1. 199/// 200static Instruction *LowerNegateToMultiply(Instruction *Neg, 201 std::map<AssertingVH<>, unsigned> &ValueRankMap) { 202 Constant *Cst = ConstantInt::getAllOnesValue(Neg->getType()); 203 204 Instruction *Res = BinaryOperator::CreateMul(Neg->getOperand(1), Cst, "",Neg); 205 ValueRankMap.erase(Neg); 206 Res->takeName(Neg); 207 Neg->replaceAllUsesWith(Res); 208 Neg->eraseFromParent(); 209 return Res; 210} 211 212// Given an expression of the form '(A+B)+(D+C)', turn it into '(((A+B)+C)+D)'. 213// Note that if D is also part of the expression tree that we recurse to 214// linearize it as well. Besides that case, this does not recurse into A,B, or 215// C. 216void Reassociate::LinearizeExpr(BinaryOperator *I) { 217 BinaryOperator *LHS = cast<BinaryOperator>(I->getOperand(0)); 218 BinaryOperator *RHS = cast<BinaryOperator>(I->getOperand(1)); 219 assert(isReassociableOp(LHS, I->getOpcode()) && 220 isReassociableOp(RHS, I->getOpcode()) && 221 "Not an expression that needs linearization?"); 222 223 DOUT << "Linear" << *LHS << *RHS << *I; 224 225 // Move the RHS instruction to live immediately before I, avoiding breaking 226 // dominator properties. 227 RHS->moveBefore(I); 228 229 // Move operands around to do the linearization. 230 I->setOperand(1, RHS->getOperand(0)); 231 RHS->setOperand(0, LHS); 232 I->setOperand(0, RHS); 233 234 ++NumLinear; 235 MadeChange = true; 236 DOUT << "Linearized: " << *I; 237 238 // If D is part of this expression tree, tail recurse. 239 if (isReassociableOp(I->getOperand(1), I->getOpcode())) 240 LinearizeExpr(I); 241} 242 243 244/// LinearizeExprTree - Given an associative binary expression tree, traverse 245/// all of the uses putting it into canonical form. This forces a left-linear 246/// form of the the expression (((a+b)+c)+d), and collects information about the 247/// rank of the non-tree operands. 248/// 249/// NOTE: These intentionally destroys the expression tree operands (turning 250/// them into undef values) to reduce #uses of the values. This means that the 251/// caller MUST use something like RewriteExprTree to put the values back in. 252/// 253void Reassociate::LinearizeExprTree(BinaryOperator *I, 254 std::vector<ValueEntry> &Ops) { 255 Value *LHS = I->getOperand(0), *RHS = I->getOperand(1); 256 unsigned Opcode = I->getOpcode(); 257 258 // First step, linearize the expression if it is in ((A+B)+(C+D)) form. 259 BinaryOperator *LHSBO = isReassociableOp(LHS, Opcode); 260 BinaryOperator *RHSBO = isReassociableOp(RHS, Opcode); 261 262 // If this is a multiply expression tree and it contains internal negations, 263 // transform them into multiplies by -1 so they can be reassociated. 264 if (I->getOpcode() == Instruction::Mul) { 265 if (!LHSBO && LHS->hasOneUse() && BinaryOperator::isNeg(LHS)) { 266 LHS = LowerNegateToMultiply(cast<Instruction>(LHS), ValueRankMap); 267 LHSBO = isReassociableOp(LHS, Opcode); 268 } 269 if (!RHSBO && RHS->hasOneUse() && BinaryOperator::isNeg(RHS)) { 270 RHS = LowerNegateToMultiply(cast<Instruction>(RHS), ValueRankMap); 271 RHSBO = isReassociableOp(RHS, Opcode); 272 } 273 } 274 275 if (!LHSBO) { 276 if (!RHSBO) { 277 // Neither the LHS or RHS as part of the tree, thus this is a leaf. As 278 // such, just remember these operands and their rank. 279 Ops.push_back(ValueEntry(getRank(LHS), LHS)); 280 Ops.push_back(ValueEntry(getRank(RHS), RHS)); 281 282 // Clear the leaves out. 283 I->setOperand(0, UndefValue::get(I->getType())); 284 I->setOperand(1, UndefValue::get(I->getType())); 285 return; 286 } else { 287 // Turn X+(Y+Z) -> (Y+Z)+X 288 std::swap(LHSBO, RHSBO); 289 std::swap(LHS, RHS); 290 bool Success = !I->swapOperands(); 291 assert(Success && "swapOperands failed"); 292 Success = false; 293 MadeChange = true; 294 } 295 } else if (RHSBO) { 296 // Turn (A+B)+(C+D) -> (((A+B)+C)+D). This guarantees the the RHS is not 297 // part of the expression tree. 298 LinearizeExpr(I); 299 LHS = LHSBO = cast<BinaryOperator>(I->getOperand(0)); 300 RHS = I->getOperand(1); 301 RHSBO = 0; 302 } 303 304 // Okay, now we know that the LHS is a nested expression and that the RHS is 305 // not. Perform reassociation. 306 assert(!isReassociableOp(RHS, Opcode) && "LinearizeExpr failed!"); 307 308 // Move LHS right before I to make sure that the tree expression dominates all 309 // values. 310 LHSBO->moveBefore(I); 311 312 // Linearize the expression tree on the LHS. 313 LinearizeExprTree(LHSBO, Ops); 314 315 // Remember the RHS operand and its rank. 316 Ops.push_back(ValueEntry(getRank(RHS), RHS)); 317 318 // Clear the RHS leaf out. 319 I->setOperand(1, UndefValue::get(I->getType())); 320} 321 322// RewriteExprTree - Now that the operands for this expression tree are 323// linearized and optimized, emit them in-order. This function is written to be 324// tail recursive. 325void Reassociate::RewriteExprTree(BinaryOperator *I, 326 std::vector<ValueEntry> &Ops, 327 unsigned i) { 328 if (i+2 == Ops.size()) { 329 if (I->getOperand(0) != Ops[i].Op || 330 I->getOperand(1) != Ops[i+1].Op) { 331 Value *OldLHS = I->getOperand(0); 332 DOUT << "RA: " << *I; 333 I->setOperand(0, Ops[i].Op); 334 I->setOperand(1, Ops[i+1].Op); 335 DOUT << "TO: " << *I; 336 MadeChange = true; 337 ++NumChanged; 338 339 // If we reassociated a tree to fewer operands (e.g. (1+a+2) -> (a+3) 340 // delete the extra, now dead, nodes. 341 RemoveDeadBinaryOp(OldLHS); 342 } 343 return; 344 } 345 assert(i+2 < Ops.size() && "Ops index out of range!"); 346 347 if (I->getOperand(1) != Ops[i].Op) { 348 DOUT << "RA: " << *I; 349 I->setOperand(1, Ops[i].Op); 350 DOUT << "TO: " << *I; 351 MadeChange = true; 352 ++NumChanged; 353 } 354 355 BinaryOperator *LHS = cast<BinaryOperator>(I->getOperand(0)); 356 assert(LHS->getOpcode() == I->getOpcode() && 357 "Improper expression tree!"); 358 359 // Compactify the tree instructions together with each other to guarantee 360 // that the expression tree is dominated by all of Ops. 361 LHS->moveBefore(I); 362 RewriteExprTree(LHS, Ops, i+1); 363} 364 365 366 367// NegateValue - Insert instructions before the instruction pointed to by BI, 368// that computes the negative version of the value specified. The negative 369// version of the value is returned, and BI is left pointing at the instruction 370// that should be processed next by the reassociation pass. 371// 372static Value *NegateValue(Value *V, Instruction *BI) { 373 // We are trying to expose opportunity for reassociation. One of the things 374 // that we want to do to achieve this is to push a negation as deep into an 375 // expression chain as possible, to expose the add instructions. In practice, 376 // this means that we turn this: 377 // X = -(A+12+C+D) into X = -A + -12 + -C + -D = -12 + -A + -C + -D 378 // so that later, a: Y = 12+X could get reassociated with the -12 to eliminate 379 // the constants. We assume that instcombine will clean up the mess later if 380 // we introduce tons of unnecessary negation instructions... 381 // 382 if (Instruction *I = dyn_cast<Instruction>(V)) 383 if (I->getOpcode() == Instruction::Add && I->hasOneUse()) { 384 // Push the negates through the add. 385 I->setOperand(0, NegateValue(I->getOperand(0), BI)); 386 I->setOperand(1, NegateValue(I->getOperand(1), BI)); 387 388 // We must move the add instruction here, because the neg instructions do 389 // not dominate the old add instruction in general. By moving it, we are 390 // assured that the neg instructions we just inserted dominate the 391 // instruction we are about to insert after them. 392 // 393 I->moveBefore(BI); 394 I->setName(I->getName()+".neg"); 395 return I; 396 } 397 398 // Insert a 'neg' instruction that subtracts the value from zero to get the 399 // negation. 400 // 401 return BinaryOperator::CreateNeg(V, V->getName() + ".neg", BI); 402} 403 404/// ShouldBreakUpSubtract - Return true if we should break up this subtract of 405/// X-Y into (X + -Y). 406static bool ShouldBreakUpSubtract(Instruction *Sub) { 407 // If this is a negation, we can't split it up! 408 if (BinaryOperator::isNeg(Sub)) 409 return false; 410 411 // Don't bother to break this up unless either the LHS is an associable add or 412 // subtract or if this is only used by one. 413 if (isReassociableOp(Sub->getOperand(0), Instruction::Add) || 414 isReassociableOp(Sub->getOperand(0), Instruction::Sub)) 415 return true; 416 if (isReassociableOp(Sub->getOperand(1), Instruction::Add) || 417 isReassociableOp(Sub->getOperand(1), Instruction::Sub)) 418 return true; 419 if (Sub->hasOneUse() && 420 (isReassociableOp(Sub->use_back(), Instruction::Add) || 421 isReassociableOp(Sub->use_back(), Instruction::Sub))) 422 return true; 423 424 return false; 425} 426 427/// BreakUpSubtract - If we have (X-Y), and if either X is an add, or if this is 428/// only used by an add, transform this into (X+(0-Y)) to promote better 429/// reassociation. 430static Instruction *BreakUpSubtract(Instruction *Sub, 431 std::map<AssertingVH<>, unsigned> &ValueRankMap) { 432 // Convert a subtract into an add and a neg instruction... so that sub 433 // instructions can be commuted with other add instructions... 434 // 435 // Calculate the negative value of Operand 1 of the sub instruction... 436 // and set it as the RHS of the add instruction we just made... 437 // 438 Value *NegVal = NegateValue(Sub->getOperand(1), Sub); 439 Instruction *New = 440 BinaryOperator::CreateAdd(Sub->getOperand(0), NegVal, "", Sub); 441 New->takeName(Sub); 442 443 // Everyone now refers to the add instruction. 444 ValueRankMap.erase(Sub); 445 Sub->replaceAllUsesWith(New); 446 Sub->eraseFromParent(); 447 448 DOUT << "Negated: " << *New; 449 return New; 450} 451 452/// ConvertShiftToMul - If this is a shift of a reassociable multiply or is used 453/// by one, change this into a multiply by a constant to assist with further 454/// reassociation. 455static Instruction *ConvertShiftToMul(Instruction *Shl, 456 std::map<AssertingVH<>, unsigned> &ValueRankMap) { 457 // If an operand of this shift is a reassociable multiply, or if the shift 458 // is used by a reassociable multiply or add, turn into a multiply. 459 if (isReassociableOp(Shl->getOperand(0), Instruction::Mul) || 460 (Shl->hasOneUse() && 461 (isReassociableOp(Shl->use_back(), Instruction::Mul) || 462 isReassociableOp(Shl->use_back(), Instruction::Add)))) { 463 Constant *MulCst = ConstantInt::get(Shl->getType(), 1); 464 MulCst = ConstantExpr::getShl(MulCst, cast<Constant>(Shl->getOperand(1))); 465 466 Instruction *Mul = BinaryOperator::CreateMul(Shl->getOperand(0), MulCst, 467 "", Shl); 468 ValueRankMap.erase(Shl); 469 Mul->takeName(Shl); 470 Shl->replaceAllUsesWith(Mul); 471 Shl->eraseFromParent(); 472 return Mul; 473 } 474 return 0; 475} 476 477// Scan backwards and forwards among values with the same rank as element i to 478// see if X exists. If X does not exist, return i. 479static unsigned FindInOperandList(std::vector<ValueEntry> &Ops, unsigned i, 480 Value *X) { 481 unsigned XRank = Ops[i].Rank; 482 unsigned e = Ops.size(); 483 for (unsigned j = i+1; j != e && Ops[j].Rank == XRank; ++j) 484 if (Ops[j].Op == X) 485 return j; 486 // Scan backwards 487 for (unsigned j = i-1; j != ~0U && Ops[j].Rank == XRank; --j) 488 if (Ops[j].Op == X) 489 return j; 490 return i; 491} 492 493/// EmitAddTreeOfValues - Emit a tree of add instructions, summing Ops together 494/// and returning the result. Insert the tree before I. 495static Value *EmitAddTreeOfValues(Instruction *I, std::vector<Value*> &Ops) { 496 if (Ops.size() == 1) return Ops.back(); 497 498 Value *V1 = Ops.back(); 499 Ops.pop_back(); 500 Value *V2 = EmitAddTreeOfValues(I, Ops); 501 return BinaryOperator::CreateAdd(V2, V1, "tmp", I); 502} 503 504/// RemoveFactorFromExpression - If V is an expression tree that is a 505/// multiplication sequence, and if this sequence contains a multiply by Factor, 506/// remove Factor from the tree and return the new tree. 507Value *Reassociate::RemoveFactorFromExpression(Value *V, Value *Factor) { 508 BinaryOperator *BO = isReassociableOp(V, Instruction::Mul); 509 if (!BO) return 0; 510 511 std::vector<ValueEntry> Factors; 512 LinearizeExprTree(BO, Factors); 513 514 bool FoundFactor = false; 515 for (unsigned i = 0, e = Factors.size(); i != e; ++i) 516 if (Factors[i].Op == Factor) { 517 FoundFactor = true; 518 Factors.erase(Factors.begin()+i); 519 break; 520 } 521 if (!FoundFactor) { 522 // Make sure to restore the operands to the expression tree. 523 RewriteExprTree(BO, Factors); 524 return 0; 525 } 526 527 if (Factors.size() == 1) return Factors[0].Op; 528 529 RewriteExprTree(BO, Factors); 530 return BO; 531} 532 533/// FindSingleUseMultiplyFactors - If V is a single-use multiply, recursively 534/// add its operands as factors, otherwise add V to the list of factors. 535static void FindSingleUseMultiplyFactors(Value *V, 536 std::vector<Value*> &Factors) { 537 BinaryOperator *BO; 538 if ((!V->hasOneUse() && !V->use_empty()) || 539 !(BO = dyn_cast<BinaryOperator>(V)) || 540 BO->getOpcode() != Instruction::Mul) { 541 Factors.push_back(V); 542 return; 543 } 544 545 // Otherwise, add the LHS and RHS to the list of factors. 546 FindSingleUseMultiplyFactors(BO->getOperand(1), Factors); 547 FindSingleUseMultiplyFactors(BO->getOperand(0), Factors); 548} 549 550 551 552Value *Reassociate::OptimizeExpression(BinaryOperator *I, 553 std::vector<ValueEntry> &Ops) { 554 // Now that we have the linearized expression tree, try to optimize it. 555 // Start by folding any constants that we found. 556 bool IterateOptimization = false; 557 if (Ops.size() == 1) return Ops[0].Op; 558 559 unsigned Opcode = I->getOpcode(); 560 561 if (Constant *V1 = dyn_cast<Constant>(Ops[Ops.size()-2].Op)) 562 if (Constant *V2 = dyn_cast<Constant>(Ops.back().Op)) { 563 Ops.pop_back(); 564 Ops.back().Op = ConstantExpr::get(Opcode, V1, V2); 565 return OptimizeExpression(I, Ops); 566 } 567 568 // Check for destructive annihilation due to a constant being used. 569 if (ConstantInt *CstVal = dyn_cast<ConstantInt>(Ops.back().Op)) 570 switch (Opcode) { 571 default: break; 572 case Instruction::And: 573 if (CstVal->isZero()) { // ... & 0 -> 0 574 ++NumAnnihil; 575 return CstVal; 576 } else if (CstVal->isAllOnesValue()) { // ... & -1 -> ... 577 Ops.pop_back(); 578 } 579 break; 580 case Instruction::Mul: 581 if (CstVal->isZero()) { // ... * 0 -> 0 582 ++NumAnnihil; 583 return CstVal; 584 } else if (cast<ConstantInt>(CstVal)->isOne()) { 585 Ops.pop_back(); // ... * 1 -> ... 586 } 587 break; 588 case Instruction::Or: 589 if (CstVal->isAllOnesValue()) { // ... | -1 -> -1 590 ++NumAnnihil; 591 return CstVal; 592 } 593 // FALLTHROUGH! 594 case Instruction::Add: 595 case Instruction::Xor: 596 if (CstVal->isZero()) // ... [|^+] 0 -> ... 597 Ops.pop_back(); 598 break; 599 } 600 if (Ops.size() == 1) return Ops[0].Op; 601 602 // Handle destructive annihilation do to identities between elements in the 603 // argument list here. 604 switch (Opcode) { 605 default: break; 606 case Instruction::And: 607 case Instruction::Or: 608 case Instruction::Xor: 609 // Scan the operand lists looking for X and ~X pairs, along with X,X pairs. 610 // If we find any, we can simplify the expression. X&~X == 0, X|~X == -1. 611 for (unsigned i = 0, e = Ops.size(); i != e; ++i) { 612 // First, check for X and ~X in the operand list. 613 assert(i < Ops.size()); 614 if (BinaryOperator::isNot(Ops[i].Op)) { // Cannot occur for ^. 615 Value *X = BinaryOperator::getNotArgument(Ops[i].Op); 616 unsigned FoundX = FindInOperandList(Ops, i, X); 617 if (FoundX != i) { 618 if (Opcode == Instruction::And) { // ...&X&~X = 0 619 ++NumAnnihil; 620 return Constant::getNullValue(X->getType()); 621 } else if (Opcode == Instruction::Or) { // ...|X|~X = -1 622 ++NumAnnihil; 623 return ConstantInt::getAllOnesValue(X->getType()); 624 } 625 } 626 } 627 628 // Next, check for duplicate pairs of values, which we assume are next to 629 // each other, due to our sorting criteria. 630 assert(i < Ops.size()); 631 if (i+1 != Ops.size() && Ops[i+1].Op == Ops[i].Op) { 632 if (Opcode == Instruction::And || Opcode == Instruction::Or) { 633 // Drop duplicate values. 634 Ops.erase(Ops.begin()+i); 635 --i; --e; 636 IterateOptimization = true; 637 ++NumAnnihil; 638 } else { 639 assert(Opcode == Instruction::Xor); 640 if (e == 2) { 641 ++NumAnnihil; 642 return Constant::getNullValue(Ops[0].Op->getType()); 643 } 644 // ... X^X -> ... 645 Ops.erase(Ops.begin()+i, Ops.begin()+i+2); 646 i -= 1; e -= 2; 647 IterateOptimization = true; 648 ++NumAnnihil; 649 } 650 } 651 } 652 break; 653 654 case Instruction::Add: 655 // Scan the operand lists looking for X and -X pairs. If we find any, we 656 // can simplify the expression. X+-X == 0. 657 for (unsigned i = 0, e = Ops.size(); i != e; ++i) { 658 assert(i < Ops.size()); 659 // Check for X and -X in the operand list. 660 if (BinaryOperator::isNeg(Ops[i].Op)) { 661 Value *X = BinaryOperator::getNegArgument(Ops[i].Op); 662 unsigned FoundX = FindInOperandList(Ops, i, X); 663 if (FoundX != i) { 664 // Remove X and -X from the operand list. 665 if (Ops.size() == 2) { 666 ++NumAnnihil; 667 return Constant::getNullValue(X->getType()); 668 } else { 669 Ops.erase(Ops.begin()+i); 670 if (i < FoundX) 671 --FoundX; 672 else 673 --i; // Need to back up an extra one. 674 Ops.erase(Ops.begin()+FoundX); 675 IterateOptimization = true; 676 ++NumAnnihil; 677 --i; // Revisit element. 678 e -= 2; // Removed two elements. 679 } 680 } 681 } 682 } 683 684 685 // Scan the operand list, checking to see if there are any common factors 686 // between operands. Consider something like A*A+A*B*C+D. We would like to 687 // reassociate this to A*(A+B*C)+D, which reduces the number of multiplies. 688 // To efficiently find this, we count the number of times a factor occurs 689 // for any ADD operands that are MULs. 690 std::map<Value*, unsigned> FactorOccurrences; 691 unsigned MaxOcc = 0; 692 Value *MaxOccVal = 0; 693 for (unsigned i = 0, e = Ops.size(); i != e; ++i) { 694 if (BinaryOperator *BOp = dyn_cast<BinaryOperator>(Ops[i].Op)) { 695 if (BOp->getOpcode() == Instruction::Mul && BOp->use_empty()) { 696 // Compute all of the factors of this added value. 697 std::vector<Value*> Factors; 698 FindSingleUseMultiplyFactors(BOp, Factors); 699 assert(Factors.size() > 1 && "Bad linearize!"); 700 701 // Add one to FactorOccurrences for each unique factor in this op. 702 if (Factors.size() == 2) { 703 unsigned Occ = ++FactorOccurrences[Factors[0]]; 704 if (Occ > MaxOcc) { MaxOcc = Occ; MaxOccVal = Factors[0]; } 705 if (Factors[0] != Factors[1]) { // Don't double count A*A. 706 Occ = ++FactorOccurrences[Factors[1]]; 707 if (Occ > MaxOcc) { MaxOcc = Occ; MaxOccVal = Factors[1]; } 708 } 709 } else { 710 std::set<Value*> Duplicates; 711 for (unsigned i = 0, e = Factors.size(); i != e; ++i) { 712 if (Duplicates.insert(Factors[i]).second) { 713 unsigned Occ = ++FactorOccurrences[Factors[i]]; 714 if (Occ > MaxOcc) { MaxOcc = Occ; MaxOccVal = Factors[i]; } 715 } 716 } 717 } 718 } 719 } 720 } 721 722 // If any factor occurred more than one time, we can pull it out. 723 if (MaxOcc > 1) { 724 DOUT << "\nFACTORING [" << MaxOcc << "]: " << *MaxOccVal << "\n"; 725 726 // Create a new instruction that uses the MaxOccVal twice. If we don't do 727 // this, we could otherwise run into situations where removing a factor 728 // from an expression will drop a use of maxocc, and this can cause 729 // RemoveFactorFromExpression on successive values to behave differently. 730 Instruction *DummyInst = BinaryOperator::CreateAdd(MaxOccVal, MaxOccVal); 731 std::vector<Value*> NewMulOps; 732 for (unsigned i = 0, e = Ops.size(); i != e; ++i) { 733 if (Value *V = RemoveFactorFromExpression(Ops[i].Op, MaxOccVal)) { 734 NewMulOps.push_back(V); 735 Ops.erase(Ops.begin()+i); 736 --i; --e; 737 } 738 } 739 740 // No need for extra uses anymore. 741 delete DummyInst; 742 743 unsigned NumAddedValues = NewMulOps.size(); 744 Value *V = EmitAddTreeOfValues(I, NewMulOps); 745 Value *V2 = BinaryOperator::CreateMul(V, MaxOccVal, "tmp", I); 746 747 // Now that we have inserted V and its sole use, optimize it. This allows 748 // us to handle cases that require multiple factoring steps, such as this: 749 // A*A*B + A*A*C --> A*(A*B+A*C) --> A*(A*(B+C)) 750 if (NumAddedValues > 1) 751 ReassociateExpression(cast<BinaryOperator>(V)); 752 753 ++NumFactor; 754 755 if (Ops.empty()) 756 return V2; 757 758 // Add the new value to the list of things being added. 759 Ops.insert(Ops.begin(), ValueEntry(getRank(V2), V2)); 760 761 // Rewrite the tree so that there is now a use of V. 762 RewriteExprTree(I, Ops); 763 return OptimizeExpression(I, Ops); 764 } 765 break; 766 //case Instruction::Mul: 767 } 768 769 if (IterateOptimization) 770 return OptimizeExpression(I, Ops); 771 return 0; 772} 773 774 775/// ReassociateBB - Inspect all of the instructions in this basic block, 776/// reassociating them as we go. 777void Reassociate::ReassociateBB(BasicBlock *BB) { 778 for (BasicBlock::iterator BBI = BB->begin(); BBI != BB->end(); ) { 779 Instruction *BI = BBI++; 780 if (BI->getOpcode() == Instruction::Shl && 781 isa<ConstantInt>(BI->getOperand(1))) 782 if (Instruction *NI = ConvertShiftToMul(BI, ValueRankMap)) { 783 MadeChange = true; 784 BI = NI; 785 } 786 787 // Reject cases where it is pointless to do this. 788 if (!isa<BinaryOperator>(BI) || BI->getType()->isFloatingPoint() || 789 isa<VectorType>(BI->getType())) 790 continue; // Floating point ops are not associative. 791 792 // If this is a subtract instruction which is not already in negate form, 793 // see if we can convert it to X+-Y. 794 if (BI->getOpcode() == Instruction::Sub) { 795 if (ShouldBreakUpSubtract(BI)) { 796 BI = BreakUpSubtract(BI, ValueRankMap); 797 MadeChange = true; 798 } else if (BinaryOperator::isNeg(BI)) { 799 // Otherwise, this is a negation. See if the operand is a multiply tree 800 // and if this is not an inner node of a multiply tree. 801 if (isReassociableOp(BI->getOperand(1), Instruction::Mul) && 802 (!BI->hasOneUse() || 803 !isReassociableOp(BI->use_back(), Instruction::Mul))) { 804 BI = LowerNegateToMultiply(BI, ValueRankMap); 805 MadeChange = true; 806 } 807 } 808 } 809 810 // If this instruction is a commutative binary operator, process it. 811 if (!BI->isAssociative()) continue; 812 BinaryOperator *I = cast<BinaryOperator>(BI); 813 814 // If this is an interior node of a reassociable tree, ignore it until we 815 // get to the root of the tree, to avoid N^2 analysis. 816 if (I->hasOneUse() && isReassociableOp(I->use_back(), I->getOpcode())) 817 continue; 818 819 // If this is an add tree that is used by a sub instruction, ignore it 820 // until we process the subtract. 821 if (I->hasOneUse() && I->getOpcode() == Instruction::Add && 822 cast<Instruction>(I->use_back())->getOpcode() == Instruction::Sub) 823 continue; 824 825 ReassociateExpression(I); 826 } 827} 828 829void Reassociate::ReassociateExpression(BinaryOperator *I) { 830 831 // First, walk the expression tree, linearizing the tree, collecting 832 std::vector<ValueEntry> Ops; 833 LinearizeExprTree(I, Ops); 834 835 DOUT << "RAIn:\t"; DEBUG(PrintOps(I, Ops)); DOUT << "\n"; 836 837 // Now that we have linearized the tree to a list and have gathered all of 838 // the operands and their ranks, sort the operands by their rank. Use a 839 // stable_sort so that values with equal ranks will have their relative 840 // positions maintained (and so the compiler is deterministic). Note that 841 // this sorts so that the highest ranking values end up at the beginning of 842 // the vector. 843 std::stable_sort(Ops.begin(), Ops.end()); 844 845 // OptimizeExpression - Now that we have the expression tree in a convenient 846 // sorted form, optimize it globally if possible. 847 if (Value *V = OptimizeExpression(I, Ops)) { 848 // This expression tree simplified to something that isn't a tree, 849 // eliminate it. 850 DOUT << "Reassoc to scalar: " << *V << "\n"; 851 I->replaceAllUsesWith(V); 852 RemoveDeadBinaryOp(I); 853 return; 854 } 855 856 // We want to sink immediates as deeply as possible except in the case where 857 // this is a multiply tree used only by an add, and the immediate is a -1. 858 // In this case we reassociate to put the negation on the outside so that we 859 // can fold the negation into the add: (-X)*Y + Z -> Z-X*Y 860 if (I->getOpcode() == Instruction::Mul && I->hasOneUse() && 861 cast<Instruction>(I->use_back())->getOpcode() == Instruction::Add && 862 isa<ConstantInt>(Ops.back().Op) && 863 cast<ConstantInt>(Ops.back().Op)->isAllOnesValue()) { 864 Ops.insert(Ops.begin(), Ops.back()); 865 Ops.pop_back(); 866 } 867 868 DOUT << "RAOut:\t"; DEBUG(PrintOps(I, Ops)); DOUT << "\n"; 869 870 if (Ops.size() == 1) { 871 // This expression tree simplified to something that isn't a tree, 872 // eliminate it. 873 I->replaceAllUsesWith(Ops[0].Op); 874 RemoveDeadBinaryOp(I); 875 } else { 876 // Now that we ordered and optimized the expressions, splat them back into 877 // the expression tree, removing any unneeded nodes. 878 RewriteExprTree(I, Ops); 879 } 880} 881 882 883bool Reassociate::runOnFunction(Function &F) { 884 // Recalculate the rank map for F 885 BuildRankMap(F); 886 887 MadeChange = false; 888 for (Function::iterator FI = F.begin(), FE = F.end(); FI != FE; ++FI) 889 ReassociateBB(FI); 890 891 // We are done with the rank map... 892 RankMap.clear(); 893 ValueRankMap.clear(); 894 return MadeChange; 895} 896 897