1 //===- ConstantHoisting.cpp - Prepare code for expensive constants --------===// 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 identifies expensive constants to hoist and coalesces them to 11 // better prepare it for SelectionDAG-based code generation. This works around 12 // the limitations of the basic-block-at-a-time approach. 13 // 14 // First it scans all instructions for integer constants and calculates its 15 // cost. If the constant can be folded into the instruction (the cost is 16 // TCC_Free) or the cost is just a simple operation (TCC_BASIC), then we don't 17 // consider it expensive and leave it alone. This is the default behavior and 18 // the default implementation of getIntImmCost will always return TCC_Free. 19 // 20 // If the cost is more than TCC_BASIC, then the integer constant can't be folded 21 // into the instruction and it might be beneficial to hoist the constant. 22 // Similar constants are coalesced to reduce register pressure and 23 // materialization code. 24 // 25 // When a constant is hoisted, it is also hidden behind a bitcast to force it to 26 // be live-out of the basic block. Otherwise the constant would be just 27 // duplicated and each basic block would have its own copy in the SelectionDAG. 28 // The SelectionDAG recognizes such constants as opaque and doesn't perform 29 // certain transformations on them, which would create a new expensive constant. 30 // 31 // This optimization is only applied to integer constants in instructions and 32 // simple (this means not nested) constant cast expressions. For example: 33 // %0 = load i64* inttoptr (i64 big_constant to i64*) 34 //===----------------------------------------------------------------------===// 35 36 #include "llvm/Transforms/Scalar/ConstantHoisting.h" 37 #include "llvm/ADT/APInt.h" 38 #include "llvm/ADT/DenseMap.h" 39 #include "llvm/ADT/None.h" 40 #include "llvm/ADT/Optional.h" 41 #include "llvm/ADT/SmallPtrSet.h" 42 #include "llvm/ADT/SmallVector.h" 43 #include "llvm/ADT/Statistic.h" 44 #include "llvm/Analysis/BlockFrequencyInfo.h" 45 #include "llvm/Analysis/TargetTransformInfo.h" 46 #include "llvm/IR/BasicBlock.h" 47 #include "llvm/IR/Constants.h" 48 #include "llvm/IR/Dominators.h" 49 #include "llvm/IR/Function.h" 50 #include "llvm/IR/InstrTypes.h" 51 #include "llvm/IR/Instruction.h" 52 #include "llvm/IR/Instructions.h" 53 #include "llvm/IR/IntrinsicInst.h" 54 #include "llvm/IR/Value.h" 55 #include "llvm/Pass.h" 56 #include "llvm/Support/BlockFrequency.h" 57 #include "llvm/Support/Casting.h" 58 #include "llvm/Support/CommandLine.h" 59 #include "llvm/Support/Debug.h" 60 #include "llvm/Support/raw_ostream.h" 61 #include "llvm/Transforms/Scalar.h" 62 #include "llvm/Transforms/Utils/Local.h" 63 #include <algorithm> 64 #include <cassert> 65 #include <cstdint> 66 #include <iterator> 67 #include <tuple> 68 #include <utility> 69 70 using namespace llvm; 71 using namespace consthoist; 72 73 #define DEBUG_TYPE "consthoist" 74 75 STATISTIC(NumConstantsHoisted, "Number of constants hoisted"); 76 STATISTIC(NumConstantsRebased, "Number of constants rebased"); 77 78 static cl::opt<bool> ConstHoistWithBlockFrequency( 79 "consthoist-with-block-frequency", cl::init(true), cl::Hidden, 80 cl::desc("Enable the use of the block frequency analysis to reduce the " 81 "chance to execute const materialization more frequently than " 82 "without hoisting.")); 83 84 namespace { 85 86 /// \brief The constant hoisting pass. 87 class ConstantHoistingLegacyPass : public FunctionPass { 88 public: 89 static char ID; // Pass identification, replacement for typeid 90 91 ConstantHoistingLegacyPass() : FunctionPass(ID) { 92 initializeConstantHoistingLegacyPassPass(*PassRegistry::getPassRegistry()); 93 } 94 95 bool runOnFunction(Function &Fn) override; 96 97 StringRef getPassName() const override { return "Constant Hoisting"; } 98 99 void getAnalysisUsage(AnalysisUsage &AU) const override { 100 AU.setPreservesCFG(); 101 if (ConstHoistWithBlockFrequency) 102 AU.addRequired<BlockFrequencyInfoWrapperPass>(); 103 AU.addRequired<DominatorTreeWrapperPass>(); 104 AU.addRequired<TargetTransformInfoWrapperPass>(); 105 } 106 107 void releaseMemory() override { Impl.releaseMemory(); } 108 109 private: 110 ConstantHoistingPass Impl; 111 }; 112 113 } // end anonymous namespace 114 115 char ConstantHoistingLegacyPass::ID = 0; 116 117 INITIALIZE_PASS_BEGIN(ConstantHoistingLegacyPass, "consthoist", 118 "Constant Hoisting", false, false) 119 INITIALIZE_PASS_DEPENDENCY(BlockFrequencyInfoWrapperPass) 120 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass) 121 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass) 122 INITIALIZE_PASS_END(ConstantHoistingLegacyPass, "consthoist", 123 "Constant Hoisting", false, false) 124 125 FunctionPass *llvm::createConstantHoistingPass() { 126 return new ConstantHoistingLegacyPass(); 127 } 128 129 /// \brief Perform the constant hoisting optimization for the given function. 130 bool ConstantHoistingLegacyPass::runOnFunction(Function &Fn) { 131 if (skipFunction(Fn)) 132 return false; 133 134 DEBUG(dbgs() << "********** Begin Constant Hoisting **********\n"); 135 DEBUG(dbgs() << "********** Function: " << Fn.getName() << '\n'); 136 137 bool MadeChange = 138 Impl.runImpl(Fn, getAnalysis<TargetTransformInfoWrapperPass>().getTTI(Fn), 139 getAnalysis<DominatorTreeWrapperPass>().getDomTree(), 140 ConstHoistWithBlockFrequency 141 ? &getAnalysis<BlockFrequencyInfoWrapperPass>().getBFI() 142 : nullptr, 143 Fn.getEntryBlock()); 144 145 if (MadeChange) { 146 DEBUG(dbgs() << "********** Function after Constant Hoisting: " 147 << Fn.getName() << '\n'); 148 DEBUG(dbgs() << Fn); 149 } 150 DEBUG(dbgs() << "********** End Constant Hoisting **********\n"); 151 152 return MadeChange; 153 } 154 155 /// \brief Find the constant materialization insertion point. 156 Instruction *ConstantHoistingPass::findMatInsertPt(Instruction *Inst, 157 unsigned Idx) const { 158 // If the operand is a cast instruction, then we have to materialize the 159 // constant before the cast instruction. 160 if (Idx != ~0U) { 161 Value *Opnd = Inst->getOperand(Idx); 162 if (auto CastInst = dyn_cast<Instruction>(Opnd)) 163 if (CastInst->isCast()) 164 return CastInst; 165 } 166 167 // The simple and common case. This also includes constant expressions. 168 if (!isa<PHINode>(Inst) && !Inst->isEHPad()) 169 return Inst; 170 171 // We can't insert directly before a phi node or an eh pad. Insert before 172 // the terminator of the incoming or dominating block. 173 assert(Entry != Inst->getParent() && "PHI or landing pad in entry block!"); 174 if (Idx != ~0U && isa<PHINode>(Inst)) 175 return cast<PHINode>(Inst)->getIncomingBlock(Idx)->getTerminator(); 176 177 // This must be an EH pad. Iterate over immediate dominators until we find a 178 // non-EH pad. We need to skip over catchswitch blocks, which are both EH pads 179 // and terminators. 180 auto IDom = DT->getNode(Inst->getParent())->getIDom(); 181 while (IDom->getBlock()->isEHPad()) { 182 assert(Entry != IDom->getBlock() && "eh pad in entry block"); 183 IDom = IDom->getIDom(); 184 } 185 186 return IDom->getBlock()->getTerminator(); 187 } 188 189 /// \brief Given \p BBs as input, find another set of BBs which collectively 190 /// dominates \p BBs and have the minimal sum of frequencies. Return the BB 191 /// set found in \p BBs. 192 static void findBestInsertionSet(DominatorTree &DT, BlockFrequencyInfo &BFI, 193 BasicBlock *Entry, 194 SmallPtrSet<BasicBlock *, 8> &BBs) { 195 assert(!BBs.count(Entry) && "Assume Entry is not in BBs"); 196 // Nodes on the current path to the root. 197 SmallPtrSet<BasicBlock *, 8> Path; 198 // Candidates includes any block 'BB' in set 'BBs' that is not strictly 199 // dominated by any other blocks in set 'BBs', and all nodes in the path 200 // in the dominator tree from Entry to 'BB'. 201 SmallPtrSet<BasicBlock *, 16> Candidates; 202 for (auto BB : BBs) { 203 Path.clear(); 204 // Walk up the dominator tree until Entry or another BB in BBs 205 // is reached. Insert the nodes on the way to the Path. 206 BasicBlock *Node = BB; 207 // The "Path" is a candidate path to be added into Candidates set. 208 bool isCandidate = false; 209 do { 210 Path.insert(Node); 211 if (Node == Entry || Candidates.count(Node)) { 212 isCandidate = true; 213 break; 214 } 215 assert(DT.getNode(Node)->getIDom() && 216 "Entry doens't dominate current Node"); 217 Node = DT.getNode(Node)->getIDom()->getBlock(); 218 } while (!BBs.count(Node)); 219 220 // If isCandidate is false, Node is another Block in BBs dominating 221 // current 'BB'. Drop the nodes on the Path. 222 if (!isCandidate) 223 continue; 224 225 // Add nodes on the Path into Candidates. 226 Candidates.insert(Path.begin(), Path.end()); 227 } 228 229 // Sort the nodes in Candidates in top-down order and save the nodes 230 // in Orders. 231 unsigned Idx = 0; 232 SmallVector<BasicBlock *, 16> Orders; 233 Orders.push_back(Entry); 234 while (Idx != Orders.size()) { 235 BasicBlock *Node = Orders[Idx++]; 236 for (auto ChildDomNode : DT.getNode(Node)->getChildren()) { 237 if (Candidates.count(ChildDomNode->getBlock())) 238 Orders.push_back(ChildDomNode->getBlock()); 239 } 240 } 241 242 // Visit Orders in bottom-up order. 243 using InsertPtsCostPair = 244 std::pair<SmallPtrSet<BasicBlock *, 16>, BlockFrequency>; 245 246 // InsertPtsMap is a map from a BB to the best insertion points for the 247 // subtree of BB (subtree not including the BB itself). 248 DenseMap<BasicBlock *, InsertPtsCostPair> InsertPtsMap; 249 InsertPtsMap.reserve(Orders.size() + 1); 250 for (auto RIt = Orders.rbegin(); RIt != Orders.rend(); RIt++) { 251 BasicBlock *Node = *RIt; 252 bool NodeInBBs = BBs.count(Node); 253 SmallPtrSet<BasicBlock *, 16> &InsertPts = InsertPtsMap[Node].first; 254 BlockFrequency &InsertPtsFreq = InsertPtsMap[Node].second; 255 256 // Return the optimal insert points in BBs. 257 if (Node == Entry) { 258 BBs.clear(); 259 if (InsertPtsFreq > BFI.getBlockFreq(Node) || 260 (InsertPtsFreq == BFI.getBlockFreq(Node) && InsertPts.size() > 1)) 261 BBs.insert(Entry); 262 else 263 BBs.insert(InsertPts.begin(), InsertPts.end()); 264 break; 265 } 266 267 BasicBlock *Parent = DT.getNode(Node)->getIDom()->getBlock(); 268 // Initially, ParentInsertPts is empty and ParentPtsFreq is 0. Every child 269 // will update its parent's ParentInsertPts and ParentPtsFreq. 270 SmallPtrSet<BasicBlock *, 16> &ParentInsertPts = InsertPtsMap[Parent].first; 271 BlockFrequency &ParentPtsFreq = InsertPtsMap[Parent].second; 272 // Choose to insert in Node or in subtree of Node. 273 // Don't hoist to EHPad because we may not find a proper place to insert 274 // in EHPad. 275 // If the total frequency of InsertPts is the same as the frequency of the 276 // target Node, and InsertPts contains more than one nodes, choose hoisting 277 // to reduce code size. 278 if (NodeInBBs || 279 (!Node->isEHPad() && 280 (InsertPtsFreq > BFI.getBlockFreq(Node) || 281 (InsertPtsFreq == BFI.getBlockFreq(Node) && InsertPts.size() > 1)))) { 282 ParentInsertPts.insert(Node); 283 ParentPtsFreq += BFI.getBlockFreq(Node); 284 } else { 285 ParentInsertPts.insert(InsertPts.begin(), InsertPts.end()); 286 ParentPtsFreq += InsertPtsFreq; 287 } 288 } 289 } 290 291 /// \brief Find an insertion point that dominates all uses. 292 SmallPtrSet<Instruction *, 8> ConstantHoistingPass::findConstantInsertionPoint( 293 const ConstantInfo &ConstInfo) const { 294 assert(!ConstInfo.RebasedConstants.empty() && "Invalid constant info entry."); 295 // Collect all basic blocks. 296 SmallPtrSet<BasicBlock *, 8> BBs; 297 SmallPtrSet<Instruction *, 8> InsertPts; 298 for (auto const &RCI : ConstInfo.RebasedConstants) 299 for (auto const &U : RCI.Uses) 300 BBs.insert(findMatInsertPt(U.Inst, U.OpndIdx)->getParent()); 301 302 if (BBs.count(Entry)) { 303 InsertPts.insert(&Entry->front()); 304 return InsertPts; 305 } 306 307 if (BFI) { 308 findBestInsertionSet(*DT, *BFI, Entry, BBs); 309 for (auto BB : BBs) { 310 BasicBlock::iterator InsertPt = BB->begin(); 311 for (; isa<PHINode>(InsertPt) || InsertPt->isEHPad(); ++InsertPt) 312 ; 313 InsertPts.insert(&*InsertPt); 314 } 315 return InsertPts; 316 } 317 318 while (BBs.size() >= 2) { 319 BasicBlock *BB, *BB1, *BB2; 320 BB1 = *BBs.begin(); 321 BB2 = *std::next(BBs.begin()); 322 BB = DT->findNearestCommonDominator(BB1, BB2); 323 if (BB == Entry) { 324 InsertPts.insert(&Entry->front()); 325 return InsertPts; 326 } 327 BBs.erase(BB1); 328 BBs.erase(BB2); 329 BBs.insert(BB); 330 } 331 assert((BBs.size() == 1) && "Expected only one element."); 332 Instruction &FirstInst = (*BBs.begin())->front(); 333 InsertPts.insert(findMatInsertPt(&FirstInst)); 334 return InsertPts; 335 } 336 337 /// \brief Record constant integer ConstInt for instruction Inst at operand 338 /// index Idx. 339 /// 340 /// The operand at index Idx is not necessarily the constant integer itself. It 341 /// could also be a cast instruction or a constant expression that uses the 342 // constant integer. 343 void ConstantHoistingPass::collectConstantCandidates( 344 ConstCandMapType &ConstCandMap, Instruction *Inst, unsigned Idx, 345 ConstantInt *ConstInt) { 346 unsigned Cost; 347 // Ask the target about the cost of materializing the constant for the given 348 // instruction and operand index. 349 if (auto IntrInst = dyn_cast<IntrinsicInst>(Inst)) 350 Cost = TTI->getIntImmCost(IntrInst->getIntrinsicID(), Idx, 351 ConstInt->getValue(), ConstInt->getType()); 352 else 353 Cost = TTI->getIntImmCost(Inst->getOpcode(), Idx, ConstInt->getValue(), 354 ConstInt->getType()); 355 356 // Ignore cheap integer constants. 357 if (Cost > TargetTransformInfo::TCC_Basic) { 358 ConstCandMapType::iterator Itr; 359 bool Inserted; 360 std::tie(Itr, Inserted) = ConstCandMap.insert(std::make_pair(ConstInt, 0)); 361 if (Inserted) { 362 ConstCandVec.push_back(ConstantCandidate(ConstInt)); 363 Itr->second = ConstCandVec.size() - 1; 364 } 365 ConstCandVec[Itr->second].addUser(Inst, Idx, Cost); 366 DEBUG(if (isa<ConstantInt>(Inst->getOperand(Idx))) 367 dbgs() << "Collect constant " << *ConstInt << " from " << *Inst 368 << " with cost " << Cost << '\n'; 369 else 370 dbgs() << "Collect constant " << *ConstInt << " indirectly from " 371 << *Inst << " via " << *Inst->getOperand(Idx) << " with cost " 372 << Cost << '\n'; 373 ); 374 } 375 } 376 377 /// \brief Check the operand for instruction Inst at index Idx. 378 void ConstantHoistingPass::collectConstantCandidates( 379 ConstCandMapType &ConstCandMap, Instruction *Inst, unsigned Idx) { 380 Value *Opnd = Inst->getOperand(Idx); 381 382 // Visit constant integers. 383 if (auto ConstInt = dyn_cast<ConstantInt>(Opnd)) { 384 collectConstantCandidates(ConstCandMap, Inst, Idx, ConstInt); 385 return; 386 } 387 388 // Visit cast instructions that have constant integers. 389 if (auto CastInst = dyn_cast<Instruction>(Opnd)) { 390 // Only visit cast instructions, which have been skipped. All other 391 // instructions should have already been visited. 392 if (!CastInst->isCast()) 393 return; 394 395 if (auto *ConstInt = dyn_cast<ConstantInt>(CastInst->getOperand(0))) { 396 // Pretend the constant is directly used by the instruction and ignore 397 // the cast instruction. 398 collectConstantCandidates(ConstCandMap, Inst, Idx, ConstInt); 399 return; 400 } 401 } 402 403 // Visit constant expressions that have constant integers. 404 if (auto ConstExpr = dyn_cast<ConstantExpr>(Opnd)) { 405 // Only visit constant cast expressions. 406 if (!ConstExpr->isCast()) 407 return; 408 409 if (auto ConstInt = dyn_cast<ConstantInt>(ConstExpr->getOperand(0))) { 410 // Pretend the constant is directly used by the instruction and ignore 411 // the constant expression. 412 collectConstantCandidates(ConstCandMap, Inst, Idx, ConstInt); 413 return; 414 } 415 } 416 } 417 418 /// \brief Scan the instruction for expensive integer constants and record them 419 /// in the constant candidate vector. 420 void ConstantHoistingPass::collectConstantCandidates( 421 ConstCandMapType &ConstCandMap, Instruction *Inst) { 422 // Skip all cast instructions. They are visited indirectly later on. 423 if (Inst->isCast()) 424 return; 425 426 // Scan all operands. 427 for (unsigned Idx = 0, E = Inst->getNumOperands(); Idx != E; ++Idx) { 428 // The cost of materializing the constants (defined in 429 // `TargetTransformInfo::getIntImmCost`) for instructions which only take 430 // constant variables is lower than `TargetTransformInfo::TCC_Basic`. So 431 // it's safe for us to collect constant candidates from all IntrinsicInsts. 432 if (canReplaceOperandWithVariable(Inst, Idx) || isa<IntrinsicInst>(Inst)) { 433 collectConstantCandidates(ConstCandMap, Inst, Idx); 434 } 435 } // end of for all operands 436 } 437 438 /// \brief Collect all integer constants in the function that cannot be folded 439 /// into an instruction itself. 440 void ConstantHoistingPass::collectConstantCandidates(Function &Fn) { 441 ConstCandMapType ConstCandMap; 442 for (BasicBlock &BB : Fn) 443 for (Instruction &Inst : BB) 444 collectConstantCandidates(ConstCandMap, &Inst); 445 } 446 447 // This helper function is necessary to deal with values that have different 448 // bit widths (APInt Operator- does not like that). If the value cannot be 449 // represented in uint64 we return an "empty" APInt. This is then interpreted 450 // as the value is not in range. 451 static Optional<APInt> calculateOffsetDiff(const APInt &V1, const APInt &V2) { 452 Optional<APInt> Res = None; 453 unsigned BW = V1.getBitWidth() > V2.getBitWidth() ? 454 V1.getBitWidth() : V2.getBitWidth(); 455 uint64_t LimVal1 = V1.getLimitedValue(); 456 uint64_t LimVal2 = V2.getLimitedValue(); 457 458 if (LimVal1 == ~0ULL || LimVal2 == ~0ULL) 459 return Res; 460 461 uint64_t Diff = LimVal1 - LimVal2; 462 return APInt(BW, Diff, true); 463 } 464 465 // From a list of constants, one needs to picked as the base and the other 466 // constants will be transformed into an offset from that base constant. The 467 // question is which we can pick best? For example, consider these constants 468 // and their number of uses: 469 // 470 // Constants| 2 | 4 | 12 | 42 | 471 // NumUses | 3 | 2 | 8 | 7 | 472 // 473 // Selecting constant 12 because it has the most uses will generate negative 474 // offsets for constants 2 and 4 (i.e. -10 and -8 respectively). If negative 475 // offsets lead to less optimal code generation, then there might be better 476 // solutions. Suppose immediates in the range of 0..35 are most optimally 477 // supported by the architecture, then selecting constant 2 is most optimal 478 // because this will generate offsets: 0, 2, 10, 40. Offsets 0, 2 and 10 are in 479 // range 0..35, and thus 3 + 2 + 8 = 13 uses are in range. Selecting 12 would 480 // have only 8 uses in range, so choosing 2 as a base is more optimal. Thus, in 481 // selecting the base constant the range of the offsets is a very important 482 // factor too that we take into account here. This algorithm calculates a total 483 // costs for selecting a constant as the base and substract the costs if 484 // immediates are out of range. It has quadratic complexity, so we call this 485 // function only when we're optimising for size and there are less than 100 486 // constants, we fall back to the straightforward algorithm otherwise 487 // which does not do all the offset calculations. 488 unsigned 489 ConstantHoistingPass::maximizeConstantsInRange(ConstCandVecType::iterator S, 490 ConstCandVecType::iterator E, 491 ConstCandVecType::iterator &MaxCostItr) { 492 unsigned NumUses = 0; 493 494 if(!Entry->getParent()->optForSize() || std::distance(S,E) > 100) { 495 for (auto ConstCand = S; ConstCand != E; ++ConstCand) { 496 NumUses += ConstCand->Uses.size(); 497 if (ConstCand->CumulativeCost > MaxCostItr->CumulativeCost) 498 MaxCostItr = ConstCand; 499 } 500 return NumUses; 501 } 502 503 DEBUG(dbgs() << "== Maximize constants in range ==\n"); 504 int MaxCost = -1; 505 for (auto ConstCand = S; ConstCand != E; ++ConstCand) { 506 auto Value = ConstCand->ConstInt->getValue(); 507 Type *Ty = ConstCand->ConstInt->getType(); 508 int Cost = 0; 509 NumUses += ConstCand->Uses.size(); 510 DEBUG(dbgs() << "= Constant: " << ConstCand->ConstInt->getValue() << "\n"); 511 512 for (auto User : ConstCand->Uses) { 513 unsigned Opcode = User.Inst->getOpcode(); 514 unsigned OpndIdx = User.OpndIdx; 515 Cost += TTI->getIntImmCost(Opcode, OpndIdx, Value, Ty); 516 DEBUG(dbgs() << "Cost: " << Cost << "\n"); 517 518 for (auto C2 = S; C2 != E; ++C2) { 519 Optional<APInt> Diff = calculateOffsetDiff( 520 C2->ConstInt->getValue(), 521 ConstCand->ConstInt->getValue()); 522 if (Diff) { 523 const int ImmCosts = 524 TTI->getIntImmCodeSizeCost(Opcode, OpndIdx, Diff.getValue(), Ty); 525 Cost -= ImmCosts; 526 DEBUG(dbgs() << "Offset " << Diff.getValue() << " " 527 << "has penalty: " << ImmCosts << "\n" 528 << "Adjusted cost: " << Cost << "\n"); 529 } 530 } 531 } 532 DEBUG(dbgs() << "Cumulative cost: " << Cost << "\n"); 533 if (Cost > MaxCost) { 534 MaxCost = Cost; 535 MaxCostItr = ConstCand; 536 DEBUG(dbgs() << "New candidate: " << MaxCostItr->ConstInt->getValue() 537 << "\n"); 538 } 539 } 540 return NumUses; 541 } 542 543 /// \brief Find the base constant within the given range and rebase all other 544 /// constants with respect to the base constant. 545 void ConstantHoistingPass::findAndMakeBaseConstant( 546 ConstCandVecType::iterator S, ConstCandVecType::iterator E) { 547 auto MaxCostItr = S; 548 unsigned NumUses = maximizeConstantsInRange(S, E, MaxCostItr); 549 550 // Don't hoist constants that have only one use. 551 if (NumUses <= 1) 552 return; 553 554 ConstantInfo ConstInfo; 555 ConstInfo.BaseConstant = MaxCostItr->ConstInt; 556 Type *Ty = ConstInfo.BaseConstant->getType(); 557 558 // Rebase the constants with respect to the base constant. 559 for (auto ConstCand = S; ConstCand != E; ++ConstCand) { 560 APInt Diff = ConstCand->ConstInt->getValue() - 561 ConstInfo.BaseConstant->getValue(); 562 Constant *Offset = Diff == 0 ? nullptr : ConstantInt::get(Ty, Diff); 563 ConstInfo.RebasedConstants.push_back( 564 RebasedConstantInfo(std::move(ConstCand->Uses), Offset)); 565 } 566 ConstantVec.push_back(std::move(ConstInfo)); 567 } 568 569 /// \brief Finds and combines constant candidates that can be easily 570 /// rematerialized with an add from a common base constant. 571 void ConstantHoistingPass::findBaseConstants() { 572 // Sort the constants by value and type. This invalidates the mapping! 573 std::sort(ConstCandVec.begin(), ConstCandVec.end(), 574 [](const ConstantCandidate &LHS, const ConstantCandidate &RHS) { 575 if (LHS.ConstInt->getType() != RHS.ConstInt->getType()) 576 return LHS.ConstInt->getType()->getBitWidth() < 577 RHS.ConstInt->getType()->getBitWidth(); 578 return LHS.ConstInt->getValue().ult(RHS.ConstInt->getValue()); 579 }); 580 581 // Simple linear scan through the sorted constant candidate vector for viable 582 // merge candidates. 583 auto MinValItr = ConstCandVec.begin(); 584 for (auto CC = std::next(ConstCandVec.begin()), E = ConstCandVec.end(); 585 CC != E; ++CC) { 586 if (MinValItr->ConstInt->getType() == CC->ConstInt->getType()) { 587 // Check if the constant is in range of an add with immediate. 588 APInt Diff = CC->ConstInt->getValue() - MinValItr->ConstInt->getValue(); 589 if ((Diff.getBitWidth() <= 64) && 590 TTI->isLegalAddImmediate(Diff.getSExtValue())) 591 continue; 592 } 593 // We either have now a different constant type or the constant is not in 594 // range of an add with immediate anymore. 595 findAndMakeBaseConstant(MinValItr, CC); 596 // Start a new base constant search. 597 MinValItr = CC; 598 } 599 // Finalize the last base constant search. 600 findAndMakeBaseConstant(MinValItr, ConstCandVec.end()); 601 } 602 603 /// \brief Updates the operand at Idx in instruction Inst with the result of 604 /// instruction Mat. If the instruction is a PHI node then special 605 /// handling for duplicate values form the same incoming basic block is 606 /// required. 607 /// \return The update will always succeed, but the return value indicated if 608 /// Mat was used for the update or not. 609 static bool updateOperand(Instruction *Inst, unsigned Idx, Instruction *Mat) { 610 if (auto PHI = dyn_cast<PHINode>(Inst)) { 611 // Check if any previous operand of the PHI node has the same incoming basic 612 // block. This is a very odd case that happens when the incoming basic block 613 // has a switch statement. In this case use the same value as the previous 614 // operand(s), otherwise we will fail verification due to different values. 615 // The values are actually the same, but the variable names are different 616 // and the verifier doesn't like that. 617 BasicBlock *IncomingBB = PHI->getIncomingBlock(Idx); 618 for (unsigned i = 0; i < Idx; ++i) { 619 if (PHI->getIncomingBlock(i) == IncomingBB) { 620 Value *IncomingVal = PHI->getIncomingValue(i); 621 Inst->setOperand(Idx, IncomingVal); 622 return false; 623 } 624 } 625 } 626 627 Inst->setOperand(Idx, Mat); 628 return true; 629 } 630 631 /// \brief Emit materialization code for all rebased constants and update their 632 /// users. 633 void ConstantHoistingPass::emitBaseConstants(Instruction *Base, 634 Constant *Offset, 635 const ConstantUser &ConstUser) { 636 Instruction *Mat = Base; 637 if (Offset) { 638 Instruction *InsertionPt = findMatInsertPt(ConstUser.Inst, 639 ConstUser.OpndIdx); 640 Mat = BinaryOperator::Create(Instruction::Add, Base, Offset, 641 "const_mat", InsertionPt); 642 643 DEBUG(dbgs() << "Materialize constant (" << *Base->getOperand(0) 644 << " + " << *Offset << ") in BB " 645 << Mat->getParent()->getName() << '\n' << *Mat << '\n'); 646 Mat->setDebugLoc(ConstUser.Inst->getDebugLoc()); 647 } 648 Value *Opnd = ConstUser.Inst->getOperand(ConstUser.OpndIdx); 649 650 // Visit constant integer. 651 if (isa<ConstantInt>(Opnd)) { 652 DEBUG(dbgs() << "Update: " << *ConstUser.Inst << '\n'); 653 if (!updateOperand(ConstUser.Inst, ConstUser.OpndIdx, Mat) && Offset) 654 Mat->eraseFromParent(); 655 DEBUG(dbgs() << "To : " << *ConstUser.Inst << '\n'); 656 return; 657 } 658 659 // Visit cast instruction. 660 if (auto CastInst = dyn_cast<Instruction>(Opnd)) { 661 assert(CastInst->isCast() && "Expected an cast instruction!"); 662 // Check if we already have visited this cast instruction before to avoid 663 // unnecessary cloning. 664 Instruction *&ClonedCastInst = ClonedCastMap[CastInst]; 665 if (!ClonedCastInst) { 666 ClonedCastInst = CastInst->clone(); 667 ClonedCastInst->setOperand(0, Mat); 668 ClonedCastInst->insertAfter(CastInst); 669 // Use the same debug location as the original cast instruction. 670 ClonedCastInst->setDebugLoc(CastInst->getDebugLoc()); 671 DEBUG(dbgs() << "Clone instruction: " << *CastInst << '\n' 672 << "To : " << *ClonedCastInst << '\n'); 673 } 674 675 DEBUG(dbgs() << "Update: " << *ConstUser.Inst << '\n'); 676 updateOperand(ConstUser.Inst, ConstUser.OpndIdx, ClonedCastInst); 677 DEBUG(dbgs() << "To : " << *ConstUser.Inst << '\n'); 678 return; 679 } 680 681 // Visit constant expression. 682 if (auto ConstExpr = dyn_cast<ConstantExpr>(Opnd)) { 683 Instruction *ConstExprInst = ConstExpr->getAsInstruction(); 684 ConstExprInst->setOperand(0, Mat); 685 ConstExprInst->insertBefore(findMatInsertPt(ConstUser.Inst, 686 ConstUser.OpndIdx)); 687 688 // Use the same debug location as the instruction we are about to update. 689 ConstExprInst->setDebugLoc(ConstUser.Inst->getDebugLoc()); 690 691 DEBUG(dbgs() << "Create instruction: " << *ConstExprInst << '\n' 692 << "From : " << *ConstExpr << '\n'); 693 DEBUG(dbgs() << "Update: " << *ConstUser.Inst << '\n'); 694 if (!updateOperand(ConstUser.Inst, ConstUser.OpndIdx, ConstExprInst)) { 695 ConstExprInst->eraseFromParent(); 696 if (Offset) 697 Mat->eraseFromParent(); 698 } 699 DEBUG(dbgs() << "To : " << *ConstUser.Inst << '\n'); 700 return; 701 } 702 } 703 704 /// \brief Hoist and hide the base constant behind a bitcast and emit 705 /// materialization code for derived constants. 706 bool ConstantHoistingPass::emitBaseConstants() { 707 bool MadeChange = false; 708 for (auto const &ConstInfo : ConstantVec) { 709 // Hoist and hide the base constant behind a bitcast. 710 SmallPtrSet<Instruction *, 8> IPSet = findConstantInsertionPoint(ConstInfo); 711 assert(!IPSet.empty() && "IPSet is empty"); 712 713 unsigned UsesNum = 0; 714 unsigned ReBasesNum = 0; 715 for (Instruction *IP : IPSet) { 716 IntegerType *Ty = ConstInfo.BaseConstant->getType(); 717 Instruction *Base = 718 new BitCastInst(ConstInfo.BaseConstant, Ty, "const", IP); 719 DEBUG(dbgs() << "Hoist constant (" << *ConstInfo.BaseConstant 720 << ") to BB " << IP->getParent()->getName() << '\n' 721 << *Base << '\n'); 722 723 // Emit materialization code for all rebased constants. 724 unsigned Uses = 0; 725 for (auto const &RCI : ConstInfo.RebasedConstants) { 726 for (auto const &U : RCI.Uses) { 727 Uses++; 728 BasicBlock *OrigMatInsertBB = 729 findMatInsertPt(U.Inst, U.OpndIdx)->getParent(); 730 // If Base constant is to be inserted in multiple places, 731 // generate rebase for U using the Base dominating U. 732 if (IPSet.size() == 1 || 733 DT->dominates(Base->getParent(), OrigMatInsertBB)) { 734 emitBaseConstants(Base, RCI.Offset, U); 735 ReBasesNum++; 736 } 737 } 738 } 739 UsesNum = Uses; 740 741 // Use the same debug location as the last user of the constant. 742 assert(!Base->use_empty() && "The use list is empty!?"); 743 assert(isa<Instruction>(Base->user_back()) && 744 "All uses should be instructions."); 745 Base->setDebugLoc(cast<Instruction>(Base->user_back())->getDebugLoc()); 746 } 747 (void)UsesNum; 748 (void)ReBasesNum; 749 // Expect all uses are rebased after rebase is done. 750 assert(UsesNum == ReBasesNum && "Not all uses are rebased"); 751 752 NumConstantsHoisted++; 753 754 // Base constant is also included in ConstInfo.RebasedConstants, so 755 // deduct 1 from ConstInfo.RebasedConstants.size(). 756 NumConstantsRebased = ConstInfo.RebasedConstants.size() - 1; 757 758 MadeChange = true; 759 } 760 return MadeChange; 761 } 762 763 /// \brief Check all cast instructions we made a copy of and remove them if they 764 /// have no more users. 765 void ConstantHoistingPass::deleteDeadCastInst() const { 766 for (auto const &I : ClonedCastMap) 767 if (I.first->use_empty()) 768 I.first->eraseFromParent(); 769 } 770 771 /// \brief Optimize expensive integer constants in the given function. 772 bool ConstantHoistingPass::runImpl(Function &Fn, TargetTransformInfo &TTI, 773 DominatorTree &DT, BlockFrequencyInfo *BFI, 774 BasicBlock &Entry) { 775 this->TTI = &TTI; 776 this->DT = &DT; 777 this->BFI = BFI; 778 this->Entry = &Entry; 779 // Collect all constant candidates. 780 collectConstantCandidates(Fn); 781 782 // There are no constant candidates to worry about. 783 if (ConstCandVec.empty()) 784 return false; 785 786 // Combine constants that can be easily materialized with an add from a common 787 // base constant. 788 findBaseConstants(); 789 790 // There are no constants to emit. 791 if (ConstantVec.empty()) 792 return false; 793 794 // Finally hoist the base constant and emit materialization code for dependent 795 // constants. 796 bool MadeChange = emitBaseConstants(); 797 798 // Cleanup dead instructions. 799 deleteDeadCastInst(); 800 801 return MadeChange; 802 } 803 804 PreservedAnalyses ConstantHoistingPass::run(Function &F, 805 FunctionAnalysisManager &AM) { 806 auto &DT = AM.getResult<DominatorTreeAnalysis>(F); 807 auto &TTI = AM.getResult<TargetIRAnalysis>(F); 808 auto BFI = ConstHoistWithBlockFrequency 809 ? &AM.getResult<BlockFrequencyAnalysis>(F) 810 : nullptr; 811 if (!runImpl(F, TTI, DT, BFI, F.getEntryBlock())) 812 return PreservedAnalyses::all(); 813 814 PreservedAnalyses PA; 815 PA.preserveSet<CFGAnalyses>(); 816 return PA; 817 } 818