Reassociate.cpp revision f4978e2094a08760d4fa8a07052091cb7276eec7
15c87bf8b86a7c82ef50fb7a89697d8e02e2553beTorne (Richard Coles)//===- Reassociate.cpp - Reassociate binary expressions -------------------===// 25267f701546148b83dfbe1d151cb184385bb5c22Torne (Richard Coles)// 35267f701546148b83dfbe1d151cb184385bb5c22Torne (Richard Coles)// The LLVM Compiler Infrastructure 45c87bf8b86a7c82ef50fb7a89697d8e02e2553beTorne (Richard Coles)// 55c87bf8b86a7c82ef50fb7a89697d8e02e2553beTorne (Richard Coles)// This file is distributed under the University of Illinois Open Source 65c87bf8b86a7c82ef50fb7a89697d8e02e2553beTorne (Richard Coles)// License. See LICENSE.TXT for details. 75267f701546148b83dfbe1d151cb184385bb5c22Torne (Richard Coles)// 85c87bf8b86a7c82ef50fb7a89697d8e02e2553beTorne (Richard Coles)//===----------------------------------------------------------------------===// 95c87bf8b86a7c82ef50fb7a89697d8e02e2553beTorne (Richard Coles)// 105c87bf8b86a7c82ef50fb7a89697d8e02e2553beTorne (Richard Coles)// This pass reassociates commutative expressions in an order that is designed 115c87bf8b86a7c82ef50fb7a89697d8e02e2553beTorne (Richard Coles)// to promote better constant propagation, GCSE, LICM, PRE... 125c87bf8b86a7c82ef50fb7a89697d8e02e2553beTorne (Richard Coles)// 135c87bf8b86a7c82ef50fb7a89697d8e02e2553beTorne (Richard Coles)// For example: 4 + (x + 5) -> x + (4 + 5) 145c87bf8b86a7c82ef50fb7a89697d8e02e2553beTorne (Richard Coles)// 155c87bf8b86a7c82ef50fb7a89697d8e02e2553beTorne (Richard Coles)// In the implementation of this algorithm, constants are assigned rank = 0, 165c87bf8b86a7c82ef50fb7a89697d8e02e2553beTorne (Richard Coles)// function arguments are rank = 1, and other values are assigned ranks 175267f701546148b83dfbe1d151cb184385bb5c22Torne (Richard Coles)// corresponding to the reverse post order traversal of current function 185c87bf8b86a7c82ef50fb7a89697d8e02e2553beTorne (Richard Coles)// (starting at 2), which effectively gives values in deep loops higher rank 195c87bf8b86a7c82ef50fb7a89697d8e02e2553beTorne (Richard Coles)// than values not in loops. 205c87bf8b86a7c82ef50fb7a89697d8e02e2553beTorne (Richard Coles)// 215c87bf8b86a7c82ef50fb7a89697d8e02e2553beTorne (Richard Coles)//===----------------------------------------------------------------------===// 225c87bf8b86a7c82ef50fb7a89697d8e02e2553beTorne (Richard Coles) 235c87bf8b86a7c82ef50fb7a89697d8e02e2553beTorne (Richard Coles)#define DEBUG_TYPE "reassociate" 245c87bf8b86a7c82ef50fb7a89697d8e02e2553beTorne (Richard Coles)#include "llvm/Transforms/Scalar.h" 255c87bf8b86a7c82ef50fb7a89697d8e02e2553beTorne (Richard Coles)#include "llvm/Constants.h" 265c87bf8b86a7c82ef50fb7a89697d8e02e2553beTorne (Richard Coles)#include "llvm/DerivedTypes.h" 275c87bf8b86a7c82ef50fb7a89697d8e02e2553beTorne (Richard Coles)#include "llvm/Function.h" 285c87bf8b86a7c82ef50fb7a89697d8e02e2553beTorne (Richard Coles)#include "llvm/Instructions.h" 295c87bf8b86a7c82ef50fb7a89697d8e02e2553beTorne (Richard Coles)#include "llvm/IntrinsicInst.h" 305c87bf8b86a7c82ef50fb7a89697d8e02e2553beTorne (Richard Coles)#include "llvm/Pass.h" 31591b958dee2cf159d33a0b931e6231072eaf38d5Ben Murdoch#include "llvm/Assembly/Writer.h" 32591b958dee2cf159d33a0b931e6231072eaf38d5Ben Murdoch#include "llvm/Support/CFG.h" 335c87bf8b86a7c82ef50fb7a89697d8e02e2553beTorne (Richard Coles)#include "llvm/Support/Compiler.h" 34591b958dee2cf159d33a0b931e6231072eaf38d5Ben Murdoch#include "llvm/Support/Debug.h" 355c87bf8b86a7c82ef50fb7a89697d8e02e2553beTorne (Richard Coles)#include "llvm/ADT/PostOrderIterator.h" 36c1847b1379d12d0e05df27436bf19a9b1bf12deaTorne (Richard Coles)#include "llvm/ADT/Statistic.h" 375c87bf8b86a7c82ef50fb7a89697d8e02e2553beTorne (Richard Coles)#include <algorithm> 38591b958dee2cf159d33a0b931e6231072eaf38d5Ben Murdoch#include <map> 395c87bf8b86a7c82ef50fb7a89697d8e02e2553beTorne (Richard Coles)using namespace llvm; 40591b958dee2cf159d33a0b931e6231072eaf38d5Ben Murdoch 41591b958dee2cf159d33a0b931e6231072eaf38d5Ben MurdochSTATISTIC(NumLinear , "Number of insts linearized"); 42521d96ec04ace82590870fb04353ec4f82bb150fTorne (Richard Coles)STATISTIC(NumChanged, "Number of insts reassociated"); 43323480423219ecd77329f8326dc5e0e3b50926d4Torne (Richard Coles)STATISTIC(NumAnnihil, "Number of expr tree annihilated"); 44591b958dee2cf159d33a0b931e6231072eaf38d5Ben MurdochSTATISTIC(NumFactor , "Number of multiplies factored"); 455c87bf8b86a7c82ef50fb7a89697d8e02e2553beTorne (Richard Coles) 465c87bf8b86a7c82ef50fb7a89697d8e02e2553beTorne (Richard Coles)namespace { 47c1847b1379d12d0e05df27436bf19a9b1bf12deaTorne (Richard Coles) struct VISIBILITY_HIDDEN ValueEntry { 485c87bf8b86a7c82ef50fb7a89697d8e02e2553beTorne (Richard Coles) unsigned Rank; 49591b958dee2cf159d33a0b931e6231072eaf38d5Ben Murdoch Value *Op; 50 ValueEntry(unsigned R, Value *O) : Rank(R), Op(O) {} 51 }; 52 inline bool operator<(const ValueEntry &LHS, const ValueEntry &RHS) { 53 return LHS.Rank > RHS.Rank; // Sort so that highest rank goes to start. 54 } 55} 56 57#ifndef NDEBUG 58/// PrintOps - Print out the expression identified in the Ops list. 59/// 60static void PrintOps(Instruction *I, const std::vector<ValueEntry> &Ops) { 61 Module *M = I->getParent()->getParent()->getParent(); 62 cerr << Instruction::getOpcodeName(I->getOpcode()) << " " 63 << *Ops[0].Op->getType(); 64 for (unsigned i = 0, e = Ops.size(); i != e; ++i) { 65 WriteAsOperand(*cerr.stream() << " ", Ops[i].Op, false, M); 66 cerr << "," << Ops[i].Rank; 67 } 68} 69#endif 70 71namespace { 72 class VISIBILITY_HIDDEN Reassociate : public FunctionPass { 73 std::map<BasicBlock*, unsigned> RankMap; 74 std::map<Value*, unsigned> ValueRankMap; 75 bool MadeChange; 76 public: 77 static char ID; // Pass identification, replacement for typeid 78 Reassociate() : FunctionPass(&ID) {} 79 80 bool runOnFunction(Function &F); 81 82 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 83 AU.setPreservesCFG(); 84 } 85 private: 86 void BuildRankMap(Function &F); 87 unsigned getRank(Value *V); 88 void ReassociateExpression(BinaryOperator *I); 89 void RewriteExprTree(BinaryOperator *I, std::vector<ValueEntry> &Ops, 90 unsigned Idx = 0); 91 Value *OptimizeExpression(BinaryOperator *I, std::vector<ValueEntry> &Ops); 92 void LinearizeExprTree(BinaryOperator *I, std::vector<ValueEntry> &Ops); 93 void LinearizeExpr(BinaryOperator *I); 94 Value *RemoveFactorFromExpression(Value *V, Value *Factor); 95 void ReassociateBB(BasicBlock *BB); 96 97 void RemoveDeadBinaryOp(Value *V); 98 }; 99} 100 101char Reassociate::ID = 0; 102static RegisterPass<Reassociate> X("reassociate", "Reassociate expressions"); 103 104// Public interface to the Reassociate pass 105FunctionPass *llvm::createReassociatePass() { return new Reassociate(); } 106 107void Reassociate::RemoveDeadBinaryOp(Value *V) { 108 Instruction *Op = dyn_cast<Instruction>(V); 109 if (!Op || !isa<BinaryOperator>(Op) || !isa<CmpInst>(Op) || !Op->use_empty()) 110 return; 111 112 Value *LHS = Op->getOperand(0), *RHS = Op->getOperand(1); 113 RemoveDeadBinaryOp(LHS); 114 RemoveDeadBinaryOp(RHS); 115} 116 117 118static bool isUnmovableInstruction(Instruction *I) { 119 if (I->getOpcode() == Instruction::PHI || 120 I->getOpcode() == Instruction::Alloca || 121 I->getOpcode() == Instruction::Load || 122 I->getOpcode() == Instruction::Malloc || 123 I->getOpcode() == Instruction::Invoke || 124 (I->getOpcode() == Instruction::Call && 125 !isa<DbgInfoIntrinsic>(I)) || 126 I->getOpcode() == Instruction::UDiv || 127 I->getOpcode() == Instruction::SDiv || 128 I->getOpcode() == Instruction::FDiv || 129 I->getOpcode() == Instruction::URem || 130 I->getOpcode() == Instruction::SRem || 131 I->getOpcode() == Instruction::FRem) 132 return true; 133 return false; 134} 135 136void Reassociate::BuildRankMap(Function &F) { 137 unsigned i = 2; 138 139 // Assign distinct ranks to function arguments 140 for (Function::arg_iterator I = F.arg_begin(), E = F.arg_end(); I != E; ++I) 141 ValueRankMap[I] = ++i; 142 143 ReversePostOrderTraversal<Function*> RPOT(&F); 144 for (ReversePostOrderTraversal<Function*>::rpo_iterator I = RPOT.begin(), 145 E = RPOT.end(); I != E; ++I) { 146 BasicBlock *BB = *I; 147 unsigned BBRank = RankMap[BB] = ++i << 16; 148 149 // Walk the basic block, adding precomputed ranks for any instructions that 150 // we cannot move. This ensures that the ranks for these instructions are 151 // all different in the block. 152 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) 153 if (isUnmovableInstruction(I)) 154 ValueRankMap[I] = ++BBRank; 155 } 156} 157 158unsigned Reassociate::getRank(Value *V) { 159 if (isa<Argument>(V)) return ValueRankMap[V]; // Function argument... 160 161 Instruction *I = dyn_cast<Instruction>(V); 162 if (I == 0) return 0; // Otherwise it's a global or constant, rank 0. 163 164 unsigned &CachedRank = ValueRankMap[I]; 165 if (CachedRank) return CachedRank; // Rank already known? 166 167 // If this is an expression, return the 1+MAX(rank(LHS), rank(RHS)) so that 168 // we can reassociate expressions for code motion! Since we do not recurse 169 // for PHI nodes, we cannot have infinite recursion here, because there 170 // cannot be loops in the value graph that do not go through PHI nodes. 171 unsigned Rank = 0, MaxRank = RankMap[I->getParent()]; 172 for (unsigned i = 0, e = I->getNumOperands(); 173 i != e && Rank != MaxRank; ++i) 174 Rank = std::max(Rank, getRank(I->getOperand(i))); 175 176 // If this is a not or neg instruction, do not count it for rank. This 177 // assures us that X and ~X will have the same rank. 178 if (!I->getType()->isInteger() || 179 (!BinaryOperator::isNot(I) && !BinaryOperator::isNeg(I))) 180 ++Rank; 181 182 //DOUT << "Calculated Rank[" << V->getName() << "] = " 183 // << Rank << "\n"; 184 185 return CachedRank = Rank; 186} 187 188/// isReassociableOp - Return true if V is an instruction of the specified 189/// opcode and if it only has one use. 190static BinaryOperator *isReassociableOp(Value *V, unsigned Opcode) { 191 if ((V->hasOneUse() || V->use_empty()) && isa<Instruction>(V) && 192 cast<Instruction>(V)->getOpcode() == Opcode) 193 return cast<BinaryOperator>(V); 194 return 0; 195} 196 197/// LowerNegateToMultiply - Replace 0-X with X*-1. 198/// 199static Instruction *LowerNegateToMultiply(Instruction *Neg, 200 std::map<Value*, unsigned> &ValueRankMap) { 201 Constant *Cst = ConstantInt::getAllOnesValue(Neg->getType()); 202 203 Instruction *Res = BinaryOperator::CreateMul(Neg->getOperand(1), Cst, "",Neg); 204 ValueRankMap.erase(Neg); 205 Res->takeName(Neg); 206 Neg->replaceAllUsesWith(Res); 207 Neg->eraseFromParent(); 208 return Res; 209} 210 211// Given an expression of the form '(A+B)+(D+C)', turn it into '(((A+B)+C)+D)'. 212// Note that if D is also part of the expression tree that we recurse to 213// linearize it as well. Besides that case, this does not recurse into A,B, or 214// C. 215void Reassociate::LinearizeExpr(BinaryOperator *I) { 216 BinaryOperator *LHS = cast<BinaryOperator>(I->getOperand(0)); 217 BinaryOperator *RHS = cast<BinaryOperator>(I->getOperand(1)); 218 assert(isReassociableOp(LHS, I->getOpcode()) && 219 isReassociableOp(RHS, I->getOpcode()) && 220 "Not an expression that needs linearization?"); 221 222 DOUT << "Linear" << *LHS << *RHS << *I; 223 224 // Move the RHS instruction to live immediately before I, avoiding breaking 225 // dominator properties. 226 RHS->moveBefore(I); 227 228 // Move operands around to do the linearization. 229 I->setOperand(1, RHS->getOperand(0)); 230 RHS->setOperand(0, LHS); 231 I->setOperand(0, RHS); 232 233 ++NumLinear; 234 MadeChange = true; 235 DOUT << "Linearized: " << *I; 236 237 // If D is part of this expression tree, tail recurse. 238 if (isReassociableOp(I->getOperand(1), I->getOpcode())) 239 LinearizeExpr(I); 240} 241 242 243/// LinearizeExprTree - Given an associative binary expression tree, traverse 244/// all of the uses putting it into canonical form. This forces a left-linear 245/// form of the the expression (((a+b)+c)+d), and collects information about the 246/// rank of the non-tree operands. 247/// 248/// NOTE: These intentionally destroys the expression tree operands (turning 249/// them into undef values) to reduce #uses of the values. This means that the 250/// caller MUST use something like RewriteExprTree to put the values back in. 251/// 252void Reassociate::LinearizeExprTree(BinaryOperator *I, 253 std::vector<ValueEntry> &Ops) { 254 Value *LHS = I->getOperand(0), *RHS = I->getOperand(1); 255 unsigned Opcode = I->getOpcode(); 256 257 // First step, linearize the expression if it is in ((A+B)+(C+D)) form. 258 BinaryOperator *LHSBO = isReassociableOp(LHS, Opcode); 259 BinaryOperator *RHSBO = isReassociableOp(RHS, Opcode); 260 261 // If this is a multiply expression tree and it contains internal negations, 262 // transform them into multiplies by -1 so they can be reassociated. 263 if (I->getOpcode() == Instruction::Mul) { 264 if (!LHSBO && LHS->hasOneUse() && BinaryOperator::isNeg(LHS)) { 265 LHS = LowerNegateToMultiply(cast<Instruction>(LHS), ValueRankMap); 266 LHSBO = isReassociableOp(LHS, Opcode); 267 } 268 if (!RHSBO && RHS->hasOneUse() && BinaryOperator::isNeg(RHS)) { 269 RHS = LowerNegateToMultiply(cast<Instruction>(RHS), ValueRankMap); 270 RHSBO = isReassociableOp(RHS, Opcode); 271 } 272 } 273 274 if (!LHSBO) { 275 if (!RHSBO) { 276 // Neither the LHS or RHS as part of the tree, thus this is a leaf. As 277 // such, just remember these operands and their rank. 278 Ops.push_back(ValueEntry(getRank(LHS), LHS)); 279 Ops.push_back(ValueEntry(getRank(RHS), RHS)); 280 281 // Clear the leaves out. 282 I->setOperand(0, UndefValue::get(I->getType())); 283 I->setOperand(1, UndefValue::get(I->getType())); 284 return; 285 } else { 286 // Turn X+(Y+Z) -> (Y+Z)+X 287 std::swap(LHSBO, RHSBO); 288 std::swap(LHS, RHS); 289 bool Success = !I->swapOperands(); 290 assert(Success && "swapOperands failed"); 291 Success = false; 292 MadeChange = true; 293 } 294 } else if (RHSBO) { 295 // Turn (A+B)+(C+D) -> (((A+B)+C)+D). This guarantees the the RHS is not 296 // part of the expression tree. 297 LinearizeExpr(I); 298 LHS = LHSBO = cast<BinaryOperator>(I->getOperand(0)); 299 RHS = I->getOperand(1); 300 RHSBO = 0; 301 } 302 303 // Okay, now we know that the LHS is a nested expression and that the RHS is 304 // not. Perform reassociation. 305 assert(!isReassociableOp(RHS, Opcode) && "LinearizeExpr failed!"); 306 307 // Move LHS right before I to make sure that the tree expression dominates all 308 // values. 309 LHSBO->moveBefore(I); 310 311 // Linearize the expression tree on the LHS. 312 LinearizeExprTree(LHSBO, Ops); 313 314 // Remember the RHS operand and its rank. 315 Ops.push_back(ValueEntry(getRank(RHS), RHS)); 316 317 // Clear the RHS leaf out. 318 I->setOperand(1, UndefValue::get(I->getType())); 319} 320 321// RewriteExprTree - Now that the operands for this expression tree are 322// linearized and optimized, emit them in-order. This function is written to be 323// tail recursive. 324void Reassociate::RewriteExprTree(BinaryOperator *I, 325 std::vector<ValueEntry> &Ops, 326 unsigned i) { 327 if (i+2 == Ops.size()) { 328 if (I->getOperand(0) != Ops[i].Op || 329 I->getOperand(1) != Ops[i+1].Op) { 330 Value *OldLHS = I->getOperand(0); 331 DOUT << "RA: " << *I; 332 I->setOperand(0, Ops[i].Op); 333 I->setOperand(1, Ops[i+1].Op); 334 DOUT << "TO: " << *I; 335 MadeChange = true; 336 ++NumChanged; 337 338 // If we reassociated a tree to fewer operands (e.g. (1+a+2) -> (a+3) 339 // delete the extra, now dead, nodes. 340 RemoveDeadBinaryOp(OldLHS); 341 } 342 return; 343 } 344 assert(i+2 < Ops.size() && "Ops index out of range!"); 345 346 if (I->getOperand(1) != Ops[i].Op) { 347 DOUT << "RA: " << *I; 348 I->setOperand(1, Ops[i].Op); 349 DOUT << "TO: " << *I; 350 MadeChange = true; 351 ++NumChanged; 352 } 353 354 BinaryOperator *LHS = cast<BinaryOperator>(I->getOperand(0)); 355 assert(LHS->getOpcode() == I->getOpcode() && 356 "Improper expression tree!"); 357 358 // Compactify the tree instructions together with each other to guarantee 359 // that the expression tree is dominated by all of Ops. 360 LHS->moveBefore(I); 361 RewriteExprTree(LHS, Ops, i+1); 362} 363 364 365 366// NegateValue - Insert instructions before the instruction pointed to by BI, 367// that computes the negative version of the value specified. The negative 368// version of the value is returned, and BI is left pointing at the instruction 369// that should be processed next by the reassociation pass. 370// 371static Value *NegateValue(Value *V, Instruction *BI) { 372 // We are trying to expose opportunity for reassociation. One of the things 373 // that we want to do to achieve this is to push a negation as deep into an 374 // expression chain as possible, to expose the add instructions. In practice, 375 // this means that we turn this: 376 // X = -(A+12+C+D) into X = -A + -12 + -C + -D = -12 + -A + -C + -D 377 // so that later, a: Y = 12+X could get reassociated with the -12 to eliminate 378 // the constants. We assume that instcombine will clean up the mess later if 379 // we introduce tons of unnecessary negation instructions... 380 // 381 if (Instruction *I = dyn_cast<Instruction>(V)) 382 if (I->getOpcode() == Instruction::Add && I->hasOneUse()) { 383 // Push the negates through the add. 384 I->setOperand(0, NegateValue(I->getOperand(0), BI)); 385 I->setOperand(1, NegateValue(I->getOperand(1), BI)); 386 387 // We must move the add instruction here, because the neg instructions do 388 // not dominate the old add instruction in general. By moving it, we are 389 // assured that the neg instructions we just inserted dominate the 390 // instruction we are about to insert after them. 391 // 392 I->moveBefore(BI); 393 I->setName(I->getName()+".neg"); 394 return I; 395 } 396 397 // Insert a 'neg' instruction that subtracts the value from zero to get the 398 // negation. 399 // 400 return BinaryOperator::CreateNeg(V, V->getName() + ".neg", BI); 401} 402 403/// ShouldBreakUpSubtract - Return true if we should break up this subtract of 404/// X-Y into (X + -Y). 405static bool ShouldBreakUpSubtract(Instruction *Sub) { 406 // If this is a negation, we can't split it up! 407 if (BinaryOperator::isNeg(Sub)) 408 return false; 409 410 // Don't bother to break this up unless either the LHS is an associable add or 411 // subtract or if this is only used by one. 412 if (isReassociableOp(Sub->getOperand(0), Instruction::Add) || 413 isReassociableOp(Sub->getOperand(0), Instruction::Sub)) 414 return true; 415 if (isReassociableOp(Sub->getOperand(1), Instruction::Add) || 416 isReassociableOp(Sub->getOperand(1), Instruction::Sub)) 417 return true; 418 if (Sub->hasOneUse() && 419 (isReassociableOp(Sub->use_back(), Instruction::Add) || 420 isReassociableOp(Sub->use_back(), Instruction::Sub))) 421 return true; 422 423 return false; 424} 425 426/// BreakUpSubtract - If we have (X-Y), and if either X is an add, or if this is 427/// only used by an add, transform this into (X+(0-Y)) to promote better 428/// reassociation. 429static Instruction *BreakUpSubtract(Instruction *Sub, 430 std::map<Value*, unsigned> &ValueRankMap) { 431 // Convert a subtract into an add and a neg instruction... so that sub 432 // instructions can be commuted with other add instructions... 433 // 434 // Calculate the negative value of Operand 1 of the sub instruction... 435 // and set it as the RHS of the add instruction we just made... 436 // 437 Value *NegVal = NegateValue(Sub->getOperand(1), Sub); 438 Instruction *New = 439 BinaryOperator::CreateAdd(Sub->getOperand(0), NegVal, "", Sub); 440 New->takeName(Sub); 441 442 // Everyone now refers to the add instruction. 443 ValueRankMap.erase(Sub); 444 Sub->replaceAllUsesWith(New); 445 Sub->eraseFromParent(); 446 447 DOUT << "Negated: " << *New; 448 return New; 449} 450 451/// ConvertShiftToMul - If this is a shift of a reassociable multiply or is used 452/// by one, change this into a multiply by a constant to assist with further 453/// reassociation. 454static Instruction *ConvertShiftToMul(Instruction *Shl, 455 std::map<Value*, unsigned> &ValueRankMap){ 456 // If an operand of this shift is a reassociable multiply, or if the shift 457 // is used by a reassociable multiply or add, turn into a multiply. 458 if (isReassociableOp(Shl->getOperand(0), Instruction::Mul) || 459 (Shl->hasOneUse() && 460 (isReassociableOp(Shl->use_back(), Instruction::Mul) || 461 isReassociableOp(Shl->use_back(), Instruction::Add)))) { 462 Constant *MulCst = ConstantInt::get(Shl->getType(), 1); 463 MulCst = ConstantExpr::getShl(MulCst, cast<Constant>(Shl->getOperand(1))); 464 465 Instruction *Mul = BinaryOperator::CreateMul(Shl->getOperand(0), MulCst, 466 "", Shl); 467 ValueRankMap.erase(Shl); 468 Mul->takeName(Shl); 469 Shl->replaceAllUsesWith(Mul); 470 Shl->eraseFromParent(); 471 return Mul; 472 } 473 return 0; 474} 475 476// Scan backwards and forwards among values with the same rank as element i to 477// see if X exists. If X does not exist, return i. 478static unsigned FindInOperandList(std::vector<ValueEntry> &Ops, unsigned i, 479 Value *X) { 480 unsigned XRank = Ops[i].Rank; 481 unsigned e = Ops.size(); 482 for (unsigned j = i+1; j != e && Ops[j].Rank == XRank; ++j) 483 if (Ops[j].Op == X) 484 return j; 485 // Scan backwards 486 for (unsigned j = i-1; j != ~0U && Ops[j].Rank == XRank; --j) 487 if (Ops[j].Op == X) 488 return j; 489 return i; 490} 491 492/// EmitAddTreeOfValues - Emit a tree of add instructions, summing Ops together 493/// and returning the result. Insert the tree before I. 494static Value *EmitAddTreeOfValues(Instruction *I, std::vector<Value*> &Ops) { 495 if (Ops.size() == 1) return Ops.back(); 496 497 Value *V1 = Ops.back(); 498 Ops.pop_back(); 499 Value *V2 = EmitAddTreeOfValues(I, Ops); 500 return BinaryOperator::CreateAdd(V2, V1, "tmp", I); 501} 502 503/// RemoveFactorFromExpression - If V is an expression tree that is a 504/// multiplication sequence, and if this sequence contains a multiply by Factor, 505/// remove Factor from the tree and return the new tree. 506Value *Reassociate::RemoveFactorFromExpression(Value *V, Value *Factor) { 507 BinaryOperator *BO = isReassociableOp(V, Instruction::Mul); 508 if (!BO) return 0; 509 510 std::vector<ValueEntry> Factors; 511 LinearizeExprTree(BO, Factors); 512 513 bool FoundFactor = false; 514 for (unsigned i = 0, e = Factors.size(); i != e; ++i) 515 if (Factors[i].Op == Factor) { 516 FoundFactor = true; 517 Factors.erase(Factors.begin()+i); 518 break; 519 } 520 if (!FoundFactor) { 521 // Make sure to restore the operands to the expression tree. 522 RewriteExprTree(BO, Factors); 523 return 0; 524 } 525 526 if (Factors.size() == 1) return Factors[0].Op; 527 528 RewriteExprTree(BO, Factors); 529 return BO; 530} 531 532/// FindSingleUseMultiplyFactors - If V is a single-use multiply, recursively 533/// add its operands as factors, otherwise add V to the list of factors. 534static void FindSingleUseMultiplyFactors(Value *V, 535 std::vector<Value*> &Factors) { 536 BinaryOperator *BO; 537 if ((!V->hasOneUse() && !V->use_empty()) || 538 !(BO = dyn_cast<BinaryOperator>(V)) || 539 BO->getOpcode() != Instruction::Mul) { 540 Factors.push_back(V); 541 return; 542 } 543 544 // Otherwise, add the LHS and RHS to the list of factors. 545 FindSingleUseMultiplyFactors(BO->getOperand(1), Factors); 546 FindSingleUseMultiplyFactors(BO->getOperand(0), Factors); 547} 548 549 550 551Value *Reassociate::OptimizeExpression(BinaryOperator *I, 552 std::vector<ValueEntry> &Ops) { 553 // Now that we have the linearized expression tree, try to optimize it. 554 // Start by folding any constants that we found. 555 bool IterateOptimization = false; 556 if (Ops.size() == 1) return Ops[0].Op; 557 558 unsigned Opcode = I->getOpcode(); 559 560 if (Constant *V1 = dyn_cast<Constant>(Ops[Ops.size()-2].Op)) 561 if (Constant *V2 = dyn_cast<Constant>(Ops.back().Op)) { 562 Ops.pop_back(); 563 Ops.back().Op = ConstantExpr::get(Opcode, V1, V2); 564 return OptimizeExpression(I, Ops); 565 } 566 567 // Check for destructive annihilation due to a constant being used. 568 if (ConstantInt *CstVal = dyn_cast<ConstantInt>(Ops.back().Op)) 569 switch (Opcode) { 570 default: break; 571 case Instruction::And: 572 if (CstVal->isZero()) { // ... & 0 -> 0 573 ++NumAnnihil; 574 return CstVal; 575 } else if (CstVal->isAllOnesValue()) { // ... & -1 -> ... 576 Ops.pop_back(); 577 } 578 break; 579 case Instruction::Mul: 580 if (CstVal->isZero()) { // ... * 0 -> 0 581 ++NumAnnihil; 582 return CstVal; 583 } else if (cast<ConstantInt>(CstVal)->isOne()) { 584 Ops.pop_back(); // ... * 1 -> ... 585 } 586 break; 587 case Instruction::Or: 588 if (CstVal->isAllOnesValue()) { // ... | -1 -> -1 589 ++NumAnnihil; 590 return CstVal; 591 } 592 // FALLTHROUGH! 593 case Instruction::Add: 594 case Instruction::Xor: 595 if (CstVal->isZero()) // ... [|^+] 0 -> ... 596 Ops.pop_back(); 597 break; 598 } 599 if (Ops.size() == 1) return Ops[0].Op; 600 601 // Handle destructive annihilation do to identities between elements in the 602 // argument list here. 603 switch (Opcode) { 604 default: break; 605 case Instruction::And: 606 case Instruction::Or: 607 case Instruction::Xor: 608 // Scan the operand lists looking for X and ~X pairs, along with X,X pairs. 609 // If we find any, we can simplify the expression. X&~X == 0, X|~X == -1. 610 for (unsigned i = 0, e = Ops.size(); i != e; ++i) { 611 // First, check for X and ~X in the operand list. 612 assert(i < Ops.size()); 613 if (BinaryOperator::isNot(Ops[i].Op)) { // Cannot occur for ^. 614 Value *X = BinaryOperator::getNotArgument(Ops[i].Op); 615 unsigned FoundX = FindInOperandList(Ops, i, X); 616 if (FoundX != i) { 617 if (Opcode == Instruction::And) { // ...&X&~X = 0 618 ++NumAnnihil; 619 return Constant::getNullValue(X->getType()); 620 } else if (Opcode == Instruction::Or) { // ...|X|~X = -1 621 ++NumAnnihil; 622 return ConstantInt::getAllOnesValue(X->getType()); 623 } 624 } 625 } 626 627 // Next, check for duplicate pairs of values, which we assume are next to 628 // each other, due to our sorting criteria. 629 assert(i < Ops.size()); 630 if (i+1 != Ops.size() && Ops[i+1].Op == Ops[i].Op) { 631 if (Opcode == Instruction::And || Opcode == Instruction::Or) { 632 // Drop duplicate values. 633 Ops.erase(Ops.begin()+i); 634 --i; --e; 635 IterateOptimization = true; 636 ++NumAnnihil; 637 } else { 638 assert(Opcode == Instruction::Xor); 639 if (e == 2) { 640 ++NumAnnihil; 641 return Constant::getNullValue(Ops[0].Op->getType()); 642 } 643 // ... X^X -> ... 644 Ops.erase(Ops.begin()+i, Ops.begin()+i+2); 645 i -= 1; e -= 2; 646 IterateOptimization = true; 647 ++NumAnnihil; 648 } 649 } 650 } 651 break; 652 653 case Instruction::Add: 654 // Scan the operand lists looking for X and -X pairs. If we find any, we 655 // can simplify the expression. X+-X == 0. 656 for (unsigned i = 0, e = Ops.size(); i != e; ++i) { 657 assert(i < Ops.size()); 658 // Check for X and -X in the operand list. 659 if (BinaryOperator::isNeg(Ops[i].Op)) { 660 Value *X = BinaryOperator::getNegArgument(Ops[i].Op); 661 unsigned FoundX = FindInOperandList(Ops, i, X); 662 if (FoundX != i) { 663 // Remove X and -X from the operand list. 664 if (Ops.size() == 2) { 665 ++NumAnnihil; 666 return Constant::getNullValue(X->getType()); 667 } else { 668 Ops.erase(Ops.begin()+i); 669 if (i < FoundX) 670 --FoundX; 671 else 672 --i; // Need to back up an extra one. 673 Ops.erase(Ops.begin()+FoundX); 674 IterateOptimization = true; 675 ++NumAnnihil; 676 --i; // Revisit element. 677 e -= 2; // Removed two elements. 678 } 679 } 680 } 681 } 682 683 684 // Scan the operand list, checking to see if there are any common factors 685 // between operands. Consider something like A*A+A*B*C+D. We would like to 686 // reassociate this to A*(A+B*C)+D, which reduces the number of multiplies. 687 // To efficiently find this, we count the number of times a factor occurs 688 // for any ADD operands that are MULs. 689 std::map<Value*, unsigned> FactorOccurrences; 690 unsigned MaxOcc = 0; 691 Value *MaxOccVal = 0; 692 for (unsigned i = 0, e = Ops.size(); i != e; ++i) { 693 if (BinaryOperator *BOp = dyn_cast<BinaryOperator>(Ops[i].Op)) { 694 if (BOp->getOpcode() == Instruction::Mul && BOp->use_empty()) { 695 // Compute all of the factors of this added value. 696 std::vector<Value*> Factors; 697 FindSingleUseMultiplyFactors(BOp, Factors); 698 assert(Factors.size() > 1 && "Bad linearize!"); 699 700 // Add one to FactorOccurrences for each unique factor in this op. 701 if (Factors.size() == 2) { 702 unsigned Occ = ++FactorOccurrences[Factors[0]]; 703 if (Occ > MaxOcc) { MaxOcc = Occ; MaxOccVal = Factors[0]; } 704 if (Factors[0] != Factors[1]) { // Don't double count A*A. 705 Occ = ++FactorOccurrences[Factors[1]]; 706 if (Occ > MaxOcc) { MaxOcc = Occ; MaxOccVal = Factors[1]; } 707 } 708 } else { 709 std::set<Value*> Duplicates; 710 for (unsigned i = 0, e = Factors.size(); i != e; ++i) { 711 if (Duplicates.insert(Factors[i]).second) { 712 unsigned Occ = ++FactorOccurrences[Factors[i]]; 713 if (Occ > MaxOcc) { MaxOcc = Occ; MaxOccVal = Factors[i]; } 714 } 715 } 716 } 717 } 718 } 719 } 720 721 // If any factor occurred more than one time, we can pull it out. 722 if (MaxOcc > 1) { 723 DOUT << "\nFACTORING [" << MaxOcc << "]: " << *MaxOccVal << "\n"; 724 725 // Create a new instruction that uses the MaxOccVal twice. If we don't do 726 // this, we could otherwise run into situations where removing a factor 727 // from an expression will drop a use of maxocc, and this can cause 728 // RemoveFactorFromExpression on successive values to behave differently. 729 Instruction *DummyInst = BinaryOperator::CreateAdd(MaxOccVal, MaxOccVal); 730 std::vector<Value*> NewMulOps; 731 for (unsigned i = 0, e = Ops.size(); i != e; ++i) { 732 if (Value *V = RemoveFactorFromExpression(Ops[i].Op, MaxOccVal)) { 733 NewMulOps.push_back(V); 734 Ops.erase(Ops.begin()+i); 735 --i; --e; 736 } 737 } 738 739 // No need for extra uses anymore. 740 delete DummyInst; 741 742 unsigned NumAddedValues = NewMulOps.size(); 743 Value *V = EmitAddTreeOfValues(I, NewMulOps); 744 Value *V2 = BinaryOperator::CreateMul(V, MaxOccVal, "tmp", I); 745 746 // Now that we have inserted V and its sole use, optimize it. This allows 747 // us to handle cases that require multiple factoring steps, such as this: 748 // A*A*B + A*A*C --> A*(A*B+A*C) --> A*(A*(B+C)) 749 if (NumAddedValues > 1) 750 ReassociateExpression(cast<BinaryOperator>(V)); 751 752 ++NumFactor; 753 754 if (Ops.empty()) 755 return V2; 756 757 // Add the new value to the list of things being added. 758 Ops.insert(Ops.begin(), ValueEntry(getRank(V2), V2)); 759 760 // Rewrite the tree so that there is now a use of V. 761 RewriteExprTree(I, Ops); 762 return OptimizeExpression(I, Ops); 763 } 764 break; 765 //case Instruction::Mul: 766 } 767 768 if (IterateOptimization) 769 return OptimizeExpression(I, Ops); 770 return 0; 771} 772 773 774/// ReassociateBB - Inspect all of the instructions in this basic block, 775/// reassociating them as we go. 776void Reassociate::ReassociateBB(BasicBlock *BB) { 777 for (BasicBlock::iterator BBI = BB->begin(); BBI != BB->end(); ) { 778 Instruction *BI = BBI++; 779 if (BI->getOpcode() == Instruction::Shl && 780 isa<ConstantInt>(BI->getOperand(1))) 781 if (Instruction *NI = ConvertShiftToMul(BI, ValueRankMap)) { 782 MadeChange = true; 783 BI = NI; 784 } 785 786 // Reject cases where it is pointless to do this. 787 if (!isa<BinaryOperator>(BI) || BI->getType()->isFloatingPoint() || 788 isa<VectorType>(BI->getType())) 789 continue; // Floating point ops are not associative. 790 791 // If this is a subtract instruction which is not already in negate form, 792 // see if we can convert it to X+-Y. 793 if (BI->getOpcode() == Instruction::Sub) { 794 if (ShouldBreakUpSubtract(BI)) { 795 BI = BreakUpSubtract(BI, ValueRankMap); 796 MadeChange = true; 797 } else if (BinaryOperator::isNeg(BI)) { 798 // Otherwise, this is a negation. See if the operand is a multiply tree 799 // and if this is not an inner node of a multiply tree. 800 if (isReassociableOp(BI->getOperand(1), Instruction::Mul) && 801 (!BI->hasOneUse() || 802 !isReassociableOp(BI->use_back(), Instruction::Mul))) { 803 BI = LowerNegateToMultiply(BI, ValueRankMap); 804 MadeChange = true; 805 } 806 } 807 } 808 809 // If this instruction is a commutative binary operator, process it. 810 if (!BI->isAssociative()) continue; 811 BinaryOperator *I = cast<BinaryOperator>(BI); 812 813 // If this is an interior node of a reassociable tree, ignore it until we 814 // get to the root of the tree, to avoid N^2 analysis. 815 if (I->hasOneUse() && isReassociableOp(I->use_back(), I->getOpcode())) 816 continue; 817 818 // If this is an add tree that is used by a sub instruction, ignore it 819 // until we process the subtract. 820 if (I->hasOneUse() && I->getOpcode() == Instruction::Add && 821 cast<Instruction>(I->use_back())->getOpcode() == Instruction::Sub) 822 continue; 823 824 ReassociateExpression(I); 825 } 826} 827 828void Reassociate::ReassociateExpression(BinaryOperator *I) { 829 830 // First, walk the expression tree, linearizing the tree, collecting 831 std::vector<ValueEntry> Ops; 832 LinearizeExprTree(I, Ops); 833 834 DOUT << "RAIn:\t"; DEBUG(PrintOps(I, Ops)); DOUT << "\n"; 835 836 // Now that we have linearized the tree to a list and have gathered all of 837 // the operands and their ranks, sort the operands by their rank. Use a 838 // stable_sort so that values with equal ranks will have their relative 839 // positions maintained (and so the compiler is deterministic). Note that 840 // this sorts so that the highest ranking values end up at the beginning of 841 // the vector. 842 std::stable_sort(Ops.begin(), Ops.end()); 843 844 // OptimizeExpression - Now that we have the expression tree in a convenient 845 // sorted form, optimize it globally if possible. 846 if (Value *V = OptimizeExpression(I, Ops)) { 847 // This expression tree simplified to something that isn't a tree, 848 // eliminate it. 849 DOUT << "Reassoc to scalar: " << *V << "\n"; 850 I->replaceAllUsesWith(V); 851 RemoveDeadBinaryOp(I); 852 return; 853 } 854 855 // We want to sink immediates as deeply as possible except in the case where 856 // this is a multiply tree used only by an add, and the immediate is a -1. 857 // In this case we reassociate to put the negation on the outside so that we 858 // can fold the negation into the add: (-X)*Y + Z -> Z-X*Y 859 if (I->getOpcode() == Instruction::Mul && I->hasOneUse() && 860 cast<Instruction>(I->use_back())->getOpcode() == Instruction::Add && 861 isa<ConstantInt>(Ops.back().Op) && 862 cast<ConstantInt>(Ops.back().Op)->isAllOnesValue()) { 863 Ops.insert(Ops.begin(), Ops.back()); 864 Ops.pop_back(); 865 } 866 867 DOUT << "RAOut:\t"; DEBUG(PrintOps(I, Ops)); DOUT << "\n"; 868 869 if (Ops.size() == 1) { 870 // This expression tree simplified to something that isn't a tree, 871 // eliminate it. 872 I->replaceAllUsesWith(Ops[0].Op); 873 RemoveDeadBinaryOp(I); 874 } else { 875 // Now that we ordered and optimized the expressions, splat them back into 876 // the expression tree, removing any unneeded nodes. 877 RewriteExprTree(I, Ops); 878 } 879} 880 881 882bool Reassociate::runOnFunction(Function &F) { 883 // Recalculate the rank map for F 884 BuildRankMap(F); 885 886 MadeChange = false; 887 for (Function::iterator FI = F.begin(), FE = F.end(); FI != FE; ++FI) 888 ReassociateBB(FI); 889 890 // We are done with the rank map... 891 RankMap.clear(); 892 ValueRankMap.clear(); 893 return MadeChange; 894} 895 896