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/Transforms/Utils/Local.h" 47 #include "llvm/IR/BasicBlock.h" 48 #include "llvm/IR/Constants.h" 49 #include "llvm/IR/DebugInfoMetadata.h" 50 #include "llvm/IR/Dominators.h" 51 #include "llvm/IR/Function.h" 52 #include "llvm/IR/InstrTypes.h" 53 #include "llvm/IR/Instruction.h" 54 #include "llvm/IR/Instructions.h" 55 #include "llvm/IR/IntrinsicInst.h" 56 #include "llvm/IR/Value.h" 57 #include "llvm/Pass.h" 58 #include "llvm/Support/BlockFrequency.h" 59 #include "llvm/Support/Casting.h" 60 #include "llvm/Support/CommandLine.h" 61 #include "llvm/Support/Debug.h" 62 #include "llvm/Support/raw_ostream.h" 63 #include "llvm/Transforms/Scalar.h" 64 #include <algorithm> 65 #include <cassert> 66 #include <cstdint> 67 #include <iterator> 68 #include <tuple> 69 #include <utility> 70 71 using namespace llvm; 72 using namespace consthoist; 73 74 #define DEBUG_TYPE "consthoist" 75 76 STATISTIC(NumConstantsHoisted, "Number of constants hoisted"); 77 STATISTIC(NumConstantsRebased, "Number of constants rebased"); 78 79 static cl::opt<bool> ConstHoistWithBlockFrequency( 80 "consthoist-with-block-frequency", cl::init(true), cl::Hidden, 81 cl::desc("Enable the use of the block frequency analysis to reduce the " 82 "chance to execute const materialization more frequently than " 83 "without hoisting.")); 84 85 namespace { 86 87 /// The constant hoisting pass. 88 class ConstantHoistingLegacyPass : public FunctionPass { 89 public: 90 static char ID; // Pass identification, replacement for typeid 91 92 ConstantHoistingLegacyPass() : FunctionPass(ID) { 93 initializeConstantHoistingLegacyPassPass(*PassRegistry::getPassRegistry()); 94 } 95 96 bool runOnFunction(Function &Fn) override; 97 98 StringRef getPassName() const override { return "Constant Hoisting"; } 99 100 void getAnalysisUsage(AnalysisUsage &AU) const override { 101 AU.setPreservesCFG(); 102 if (ConstHoistWithBlockFrequency) 103 AU.addRequired<BlockFrequencyInfoWrapperPass>(); 104 AU.addRequired<DominatorTreeWrapperPass>(); 105 AU.addRequired<TargetTransformInfoWrapperPass>(); 106 } 107 108 void releaseMemory() override { Impl.releaseMemory(); } 109 110 private: 111 ConstantHoistingPass Impl; 112 }; 113 114 } // end anonymous namespace 115 116 char ConstantHoistingLegacyPass::ID = 0; 117 118 INITIALIZE_PASS_BEGIN(ConstantHoistingLegacyPass, "consthoist", 119 "Constant Hoisting", false, false) 120 INITIALIZE_PASS_DEPENDENCY(BlockFrequencyInfoWrapperPass) 121 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass) 122 INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass) 123 INITIALIZE_PASS_END(ConstantHoistingLegacyPass, "consthoist", 124 "Constant Hoisting", false, false) 125 126 FunctionPass *llvm::createConstantHoistingPass() { 127 return new ConstantHoistingLegacyPass(); 128 } 129 130 /// Perform the constant hoisting optimization for the given function. 131 bool ConstantHoistingLegacyPass::runOnFunction(Function &Fn) { 132 if (skipFunction(Fn)) 133 return false; 134 135 LLVM_DEBUG(dbgs() << "********** Begin Constant Hoisting **********\n"); 136 LLVM_DEBUG(dbgs() << "********** Function: " << Fn.getName() << '\n'); 137 138 bool MadeChange = 139 Impl.runImpl(Fn, getAnalysis<TargetTransformInfoWrapperPass>().getTTI(Fn), 140 getAnalysis<DominatorTreeWrapperPass>().getDomTree(), 141 ConstHoistWithBlockFrequency 142 ? &getAnalysis<BlockFrequencyInfoWrapperPass>().getBFI() 143 : nullptr, 144 Fn.getEntryBlock()); 145 146 if (MadeChange) { 147 LLVM_DEBUG(dbgs() << "********** Function after Constant Hoisting: " 148 << Fn.getName() << '\n'); 149 LLVM_DEBUG(dbgs() << Fn); 150 } 151 LLVM_DEBUG(dbgs() << "********** End Constant Hoisting **********\n"); 152 153 return MadeChange; 154 } 155 156 /// Find the constant materialization insertion point. 157 Instruction *ConstantHoistingPass::findMatInsertPt(Instruction *Inst, 158 unsigned Idx) const { 159 // If the operand is a cast instruction, then we have to materialize the 160 // constant before the cast instruction. 161 if (Idx != ~0U) { 162 Value *Opnd = Inst->getOperand(Idx); 163 if (auto CastInst = dyn_cast<Instruction>(Opnd)) 164 if (CastInst->isCast()) 165 return CastInst; 166 } 167 168 // The simple and common case. This also includes constant expressions. 169 if (!isa<PHINode>(Inst) && !Inst->isEHPad()) 170 return Inst; 171 172 // We can't insert directly before a phi node or an eh pad. Insert before 173 // the terminator of the incoming or dominating block. 174 assert(Entry != Inst->getParent() && "PHI or landing pad in entry block!"); 175 if (Idx != ~0U && isa<PHINode>(Inst)) 176 return cast<PHINode>(Inst)->getIncomingBlock(Idx)->getTerminator(); 177 178 // This must be an EH pad. Iterate over immediate dominators until we find a 179 // non-EH pad. We need to skip over catchswitch blocks, which are both EH pads 180 // and terminators. 181 auto IDom = DT->getNode(Inst->getParent())->getIDom(); 182 while (IDom->getBlock()->isEHPad()) { 183 assert(Entry != IDom->getBlock() && "eh pad in entry block"); 184 IDom = IDom->getIDom(); 185 } 186 187 return IDom->getBlock()->getTerminator(); 188 } 189 190 /// Given \p BBs as input, find another set of BBs which collectively 191 /// dominates \p BBs and have the minimal sum of frequencies. Return the BB 192 /// set found in \p BBs. 193 static void findBestInsertionSet(DominatorTree &DT, BlockFrequencyInfo &BFI, 194 BasicBlock *Entry, 195 SmallPtrSet<BasicBlock *, 8> &BBs) { 196 assert(!BBs.count(Entry) && "Assume Entry is not in BBs"); 197 // Nodes on the current path to the root. 198 SmallPtrSet<BasicBlock *, 8> Path; 199 // Candidates includes any block 'BB' in set 'BBs' that is not strictly 200 // dominated by any other blocks in set 'BBs', and all nodes in the path 201 // in the dominator tree from Entry to 'BB'. 202 SmallPtrSet<BasicBlock *, 16> Candidates; 203 for (auto BB : BBs) { 204 Path.clear(); 205 // Walk up the dominator tree until Entry or another BB in BBs 206 // is reached. Insert the nodes on the way to the Path. 207 BasicBlock *Node = BB; 208 // The "Path" is a candidate path to be added into Candidates set. 209 bool isCandidate = false; 210 do { 211 Path.insert(Node); 212 if (Node == Entry || Candidates.count(Node)) { 213 isCandidate = true; 214 break; 215 } 216 assert(DT.getNode(Node)->getIDom() && 217 "Entry doens't dominate current Node"); 218 Node = DT.getNode(Node)->getIDom()->getBlock(); 219 } while (!BBs.count(Node)); 220 221 // If isCandidate is false, Node is another Block in BBs dominating 222 // current 'BB'. Drop the nodes on the Path. 223 if (!isCandidate) 224 continue; 225 226 // Add nodes on the Path into Candidates. 227 Candidates.insert(Path.begin(), Path.end()); 228 } 229 230 // Sort the nodes in Candidates in top-down order and save the nodes 231 // in Orders. 232 unsigned Idx = 0; 233 SmallVector<BasicBlock *, 16> Orders; 234 Orders.push_back(Entry); 235 while (Idx != Orders.size()) { 236 BasicBlock *Node = Orders[Idx++]; 237 for (auto ChildDomNode : DT.getNode(Node)->getChildren()) { 238 if (Candidates.count(ChildDomNode->getBlock())) 239 Orders.push_back(ChildDomNode->getBlock()); 240 } 241 } 242 243 // Visit Orders in bottom-up order. 244 using InsertPtsCostPair = 245 std::pair<SmallPtrSet<BasicBlock *, 16>, BlockFrequency>; 246 247 // InsertPtsMap is a map from a BB to the best insertion points for the 248 // subtree of BB (subtree not including the BB itself). 249 DenseMap<BasicBlock *, InsertPtsCostPair> InsertPtsMap; 250 InsertPtsMap.reserve(Orders.size() + 1); 251 for (auto RIt = Orders.rbegin(); RIt != Orders.rend(); RIt++) { 252 BasicBlock *Node = *RIt; 253 bool NodeInBBs = BBs.count(Node); 254 SmallPtrSet<BasicBlock *, 16> &InsertPts = InsertPtsMap[Node].first; 255 BlockFrequency &InsertPtsFreq = InsertPtsMap[Node].second; 256 257 // Return the optimal insert points in BBs. 258 if (Node == Entry) { 259 BBs.clear(); 260 if (InsertPtsFreq > BFI.getBlockFreq(Node) || 261 (InsertPtsFreq == BFI.getBlockFreq(Node) && InsertPts.size() > 1)) 262 BBs.insert(Entry); 263 else 264 BBs.insert(InsertPts.begin(), InsertPts.end()); 265 break; 266 } 267 268 BasicBlock *Parent = DT.getNode(Node)->getIDom()->getBlock(); 269 // Initially, ParentInsertPts is empty and ParentPtsFreq is 0. Every child 270 // will update its parent's ParentInsertPts and ParentPtsFreq. 271 SmallPtrSet<BasicBlock *, 16> &ParentInsertPts = InsertPtsMap[Parent].first; 272 BlockFrequency &ParentPtsFreq = InsertPtsMap[Parent].second; 273 // Choose to insert in Node or in subtree of Node. 274 // Don't hoist to EHPad because we may not find a proper place to insert 275 // in EHPad. 276 // If the total frequency of InsertPts is the same as the frequency of the 277 // target Node, and InsertPts contains more than one nodes, choose hoisting 278 // to reduce code size. 279 if (NodeInBBs || 280 (!Node->isEHPad() && 281 (InsertPtsFreq > BFI.getBlockFreq(Node) || 282 (InsertPtsFreq == BFI.getBlockFreq(Node) && InsertPts.size() > 1)))) { 283 ParentInsertPts.insert(Node); 284 ParentPtsFreq += BFI.getBlockFreq(Node); 285 } else { 286 ParentInsertPts.insert(InsertPts.begin(), InsertPts.end()); 287 ParentPtsFreq += InsertPtsFreq; 288 } 289 } 290 } 291 292 /// Find an insertion point that dominates all uses. 293 SmallPtrSet<Instruction *, 8> ConstantHoistingPass::findConstantInsertionPoint( 294 const ConstantInfo &ConstInfo) const { 295 assert(!ConstInfo.RebasedConstants.empty() && "Invalid constant info entry."); 296 // Collect all basic blocks. 297 SmallPtrSet<BasicBlock *, 8> BBs; 298 SmallPtrSet<Instruction *, 8> InsertPts; 299 for (auto const &RCI : ConstInfo.RebasedConstants) 300 for (auto const &U : RCI.Uses) 301 BBs.insert(findMatInsertPt(U.Inst, U.OpndIdx)->getParent()); 302 303 if (BBs.count(Entry)) { 304 InsertPts.insert(&Entry->front()); 305 return InsertPts; 306 } 307 308 if (BFI) { 309 findBestInsertionSet(*DT, *BFI, Entry, BBs); 310 for (auto BB : BBs) { 311 BasicBlock::iterator InsertPt = BB->begin(); 312 for (; isa<PHINode>(InsertPt) || InsertPt->isEHPad(); ++InsertPt) 313 ; 314 InsertPts.insert(&*InsertPt); 315 } 316 return InsertPts; 317 } 318 319 while (BBs.size() >= 2) { 320 BasicBlock *BB, *BB1, *BB2; 321 BB1 = *BBs.begin(); 322 BB2 = *std::next(BBs.begin()); 323 BB = DT->findNearestCommonDominator(BB1, BB2); 324 if (BB == Entry) { 325 InsertPts.insert(&Entry->front()); 326 return InsertPts; 327 } 328 BBs.erase(BB1); 329 BBs.erase(BB2); 330 BBs.insert(BB); 331 } 332 assert((BBs.size() == 1) && "Expected only one element."); 333 Instruction &FirstInst = (*BBs.begin())->front(); 334 InsertPts.insert(findMatInsertPt(&FirstInst)); 335 return InsertPts; 336 } 337 338 /// Record constant integer ConstInt for instruction Inst at operand 339 /// index Idx. 340 /// 341 /// The operand at index Idx is not necessarily the constant integer itself. It 342 /// could also be a cast instruction or a constant expression that uses the 343 // constant integer. 344 void ConstantHoistingPass::collectConstantCandidates( 345 ConstCandMapType &ConstCandMap, Instruction *Inst, unsigned Idx, 346 ConstantInt *ConstInt) { 347 unsigned Cost; 348 // Ask the target about the cost of materializing the constant for the given 349 // instruction and operand index. 350 if (auto IntrInst = dyn_cast<IntrinsicInst>(Inst)) 351 Cost = TTI->getIntImmCost(IntrInst->getIntrinsicID(), Idx, 352 ConstInt->getValue(), ConstInt->getType()); 353 else 354 Cost = TTI->getIntImmCost(Inst->getOpcode(), Idx, ConstInt->getValue(), 355 ConstInt->getType()); 356 357 // Ignore cheap integer constants. 358 if (Cost > TargetTransformInfo::TCC_Basic) { 359 ConstCandMapType::iterator Itr; 360 bool Inserted; 361 std::tie(Itr, Inserted) = ConstCandMap.insert(std::make_pair(ConstInt, 0)); 362 if (Inserted) { 363 ConstCandVec.push_back(ConstantCandidate(ConstInt)); 364 Itr->second = ConstCandVec.size() - 1; 365 } 366 ConstCandVec[Itr->second].addUser(Inst, Idx, Cost); 367 LLVM_DEBUG(if (isa<ConstantInt>(Inst->getOperand(Idx))) dbgs() 368 << "Collect constant " << *ConstInt << " from " << *Inst 369 << " with cost " << Cost << '\n'; 370 else dbgs() << "Collect constant " << *ConstInt 371 << " indirectly from " << *Inst << " via " 372 << *Inst->getOperand(Idx) << " with cost " << Cost 373 << '\n';); 374 } 375 } 376 377 /// 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 /// 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 /// 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 LLVM_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 LLVM_DEBUG(dbgs() << "= Constant: " << ConstCand->ConstInt->getValue() 511 << "\n"); 512 513 for (auto User : ConstCand->Uses) { 514 unsigned Opcode = User.Inst->getOpcode(); 515 unsigned OpndIdx = User.OpndIdx; 516 Cost += TTI->getIntImmCost(Opcode, OpndIdx, Value, Ty); 517 LLVM_DEBUG(dbgs() << "Cost: " << Cost << "\n"); 518 519 for (auto C2 = S; C2 != E; ++C2) { 520 Optional<APInt> Diff = calculateOffsetDiff( 521 C2->ConstInt->getValue(), 522 ConstCand->ConstInt->getValue()); 523 if (Diff) { 524 const int ImmCosts = 525 TTI->getIntImmCodeSizeCost(Opcode, OpndIdx, Diff.getValue(), Ty); 526 Cost -= ImmCosts; 527 LLVM_DEBUG(dbgs() << "Offset " << Diff.getValue() << " " 528 << "has penalty: " << ImmCosts << "\n" 529 << "Adjusted cost: " << Cost << "\n"); 530 } 531 } 532 } 533 LLVM_DEBUG(dbgs() << "Cumulative cost: " << Cost << "\n"); 534 if (Cost > MaxCost) { 535 MaxCost = Cost; 536 MaxCostItr = ConstCand; 537 LLVM_DEBUG(dbgs() << "New candidate: " << MaxCostItr->ConstInt->getValue() 538 << "\n"); 539 } 540 } 541 return NumUses; 542 } 543 544 /// Find the base constant within the given range and rebase all other 545 /// constants with respect to the base constant. 546 void ConstantHoistingPass::findAndMakeBaseConstant( 547 ConstCandVecType::iterator S, ConstCandVecType::iterator E) { 548 auto MaxCostItr = S; 549 unsigned NumUses = maximizeConstantsInRange(S, E, MaxCostItr); 550 551 // Don't hoist constants that have only one use. 552 if (NumUses <= 1) 553 return; 554 555 ConstantInfo ConstInfo; 556 ConstInfo.BaseConstant = MaxCostItr->ConstInt; 557 Type *Ty = ConstInfo.BaseConstant->getType(); 558 559 // Rebase the constants with respect to the base constant. 560 for (auto ConstCand = S; ConstCand != E; ++ConstCand) { 561 APInt Diff = ConstCand->ConstInt->getValue() - 562 ConstInfo.BaseConstant->getValue(); 563 Constant *Offset = Diff == 0 ? nullptr : ConstantInt::get(Ty, Diff); 564 ConstInfo.RebasedConstants.push_back( 565 RebasedConstantInfo(std::move(ConstCand->Uses), Offset)); 566 } 567 ConstantVec.push_back(std::move(ConstInfo)); 568 } 569 570 /// Finds and combines constant candidates that can be easily 571 /// rematerialized with an add from a common base constant. 572 void ConstantHoistingPass::findBaseConstants() { 573 // Sort the constants by value and type. This invalidates the mapping! 574 llvm::sort(ConstCandVec.begin(), ConstCandVec.end(), 575 [](const ConstantCandidate &LHS, const ConstantCandidate &RHS) { 576 if (LHS.ConstInt->getType() != RHS.ConstInt->getType()) 577 return LHS.ConstInt->getType()->getBitWidth() < 578 RHS.ConstInt->getType()->getBitWidth(); 579 return LHS.ConstInt->getValue().ult(RHS.ConstInt->getValue()); 580 }); 581 582 // Simple linear scan through the sorted constant candidate vector for viable 583 // merge candidates. 584 auto MinValItr = ConstCandVec.begin(); 585 for (auto CC = std::next(ConstCandVec.begin()), E = ConstCandVec.end(); 586 CC != E; ++CC) { 587 if (MinValItr->ConstInt->getType() == CC->ConstInt->getType()) { 588 // Check if the constant is in range of an add with immediate. 589 APInt Diff = CC->ConstInt->getValue() - MinValItr->ConstInt->getValue(); 590 if ((Diff.getBitWidth() <= 64) && 591 TTI->isLegalAddImmediate(Diff.getSExtValue())) 592 continue; 593 } 594 // We either have now a different constant type or the constant is not in 595 // range of an add with immediate anymore. 596 findAndMakeBaseConstant(MinValItr, CC); 597 // Start a new base constant search. 598 MinValItr = CC; 599 } 600 // Finalize the last base constant search. 601 findAndMakeBaseConstant(MinValItr, ConstCandVec.end()); 602 } 603 604 /// Updates the operand at Idx in instruction Inst with the result of 605 /// instruction Mat. If the instruction is a PHI node then special 606 /// handling for duplicate values form the same incoming basic block is 607 /// required. 608 /// \return The update will always succeed, but the return value indicated if 609 /// Mat was used for the update or not. 610 static bool updateOperand(Instruction *Inst, unsigned Idx, Instruction *Mat) { 611 if (auto PHI = dyn_cast<PHINode>(Inst)) { 612 // Check if any previous operand of the PHI node has the same incoming basic 613 // block. This is a very odd case that happens when the incoming basic block 614 // has a switch statement. In this case use the same value as the previous 615 // operand(s), otherwise we will fail verification due to different values. 616 // The values are actually the same, but the variable names are different 617 // and the verifier doesn't like that. 618 BasicBlock *IncomingBB = PHI->getIncomingBlock(Idx); 619 for (unsigned i = 0; i < Idx; ++i) { 620 if (PHI->getIncomingBlock(i) == IncomingBB) { 621 Value *IncomingVal = PHI->getIncomingValue(i); 622 Inst->setOperand(Idx, IncomingVal); 623 return false; 624 } 625 } 626 } 627 628 Inst->setOperand(Idx, Mat); 629 return true; 630 } 631 632 /// Emit materialization code for all rebased constants and update their 633 /// users. 634 void ConstantHoistingPass::emitBaseConstants(Instruction *Base, 635 Constant *Offset, 636 const ConstantUser &ConstUser) { 637 Instruction *Mat = Base; 638 if (Offset) { 639 Instruction *InsertionPt = findMatInsertPt(ConstUser.Inst, 640 ConstUser.OpndIdx); 641 Mat = BinaryOperator::Create(Instruction::Add, Base, Offset, 642 "const_mat", InsertionPt); 643 644 LLVM_DEBUG(dbgs() << "Materialize constant (" << *Base->getOperand(0) 645 << " + " << *Offset << ") in BB " 646 << Mat->getParent()->getName() << '\n' 647 << *Mat << '\n'); 648 Mat->setDebugLoc(ConstUser.Inst->getDebugLoc()); 649 } 650 Value *Opnd = ConstUser.Inst->getOperand(ConstUser.OpndIdx); 651 652 // Visit constant integer. 653 if (isa<ConstantInt>(Opnd)) { 654 LLVM_DEBUG(dbgs() << "Update: " << *ConstUser.Inst << '\n'); 655 if (!updateOperand(ConstUser.Inst, ConstUser.OpndIdx, Mat) && Offset) 656 Mat->eraseFromParent(); 657 LLVM_DEBUG(dbgs() << "To : " << *ConstUser.Inst << '\n'); 658 return; 659 } 660 661 // Visit cast instruction. 662 if (auto CastInst = dyn_cast<Instruction>(Opnd)) { 663 assert(CastInst->isCast() && "Expected an cast instruction!"); 664 // Check if we already have visited this cast instruction before to avoid 665 // unnecessary cloning. 666 Instruction *&ClonedCastInst = ClonedCastMap[CastInst]; 667 if (!ClonedCastInst) { 668 ClonedCastInst = CastInst->clone(); 669 ClonedCastInst->setOperand(0, Mat); 670 ClonedCastInst->insertAfter(CastInst); 671 // Use the same debug location as the original cast instruction. 672 ClonedCastInst->setDebugLoc(CastInst->getDebugLoc()); 673 LLVM_DEBUG(dbgs() << "Clone instruction: " << *CastInst << '\n' 674 << "To : " << *ClonedCastInst << '\n'); 675 } 676 677 LLVM_DEBUG(dbgs() << "Update: " << *ConstUser.Inst << '\n'); 678 updateOperand(ConstUser.Inst, ConstUser.OpndIdx, ClonedCastInst); 679 LLVM_DEBUG(dbgs() << "To : " << *ConstUser.Inst << '\n'); 680 return; 681 } 682 683 // Visit constant expression. 684 if (auto ConstExpr = dyn_cast<ConstantExpr>(Opnd)) { 685 Instruction *ConstExprInst = ConstExpr->getAsInstruction(); 686 ConstExprInst->setOperand(0, Mat); 687 ConstExprInst->insertBefore(findMatInsertPt(ConstUser.Inst, 688 ConstUser.OpndIdx)); 689 690 // Use the same debug location as the instruction we are about to update. 691 ConstExprInst->setDebugLoc(ConstUser.Inst->getDebugLoc()); 692 693 LLVM_DEBUG(dbgs() << "Create instruction: " << *ConstExprInst << '\n' 694 << "From : " << *ConstExpr << '\n'); 695 LLVM_DEBUG(dbgs() << "Update: " << *ConstUser.Inst << '\n'); 696 if (!updateOperand(ConstUser.Inst, ConstUser.OpndIdx, ConstExprInst)) { 697 ConstExprInst->eraseFromParent(); 698 if (Offset) 699 Mat->eraseFromParent(); 700 } 701 LLVM_DEBUG(dbgs() << "To : " << *ConstUser.Inst << '\n'); 702 return; 703 } 704 } 705 706 /// Hoist and hide the base constant behind a bitcast and emit 707 /// materialization code for derived constants. 708 bool ConstantHoistingPass::emitBaseConstants() { 709 bool MadeChange = false; 710 for (auto const &ConstInfo : ConstantVec) { 711 // Hoist and hide the base constant behind a bitcast. 712 SmallPtrSet<Instruction *, 8> IPSet = findConstantInsertionPoint(ConstInfo); 713 assert(!IPSet.empty() && "IPSet is empty"); 714 715 unsigned UsesNum = 0; 716 unsigned ReBasesNum = 0; 717 for (Instruction *IP : IPSet) { 718 IntegerType *Ty = ConstInfo.BaseConstant->getType(); 719 Instruction *Base = 720 new BitCastInst(ConstInfo.BaseConstant, Ty, "const", IP); 721 722 Base->setDebugLoc(IP->getDebugLoc()); 723 724 LLVM_DEBUG(dbgs() << "Hoist constant (" << *ConstInfo.BaseConstant 725 << ") to BB " << IP->getParent()->getName() << '\n' 726 << *Base << '\n'); 727 728 // Emit materialization code for all rebased constants. 729 unsigned Uses = 0; 730 for (auto const &RCI : ConstInfo.RebasedConstants) { 731 for (auto const &U : RCI.Uses) { 732 Uses++; 733 BasicBlock *OrigMatInsertBB = 734 findMatInsertPt(U.Inst, U.OpndIdx)->getParent(); 735 // If Base constant is to be inserted in multiple places, 736 // generate rebase for U using the Base dominating U. 737 if (IPSet.size() == 1 || 738 DT->dominates(Base->getParent(), OrigMatInsertBB)) { 739 emitBaseConstants(Base, RCI.Offset, U); 740 ReBasesNum++; 741 } 742 743 Base->setDebugLoc(DILocation::getMergedLocation(Base->getDebugLoc(), U.Inst->getDebugLoc())); 744 } 745 } 746 UsesNum = Uses; 747 748 // Use the same debug location as the last user of the constant. 749 assert(!Base->use_empty() && "The use list is empty!?"); 750 assert(isa<Instruction>(Base->user_back()) && 751 "All uses should be instructions."); 752 } 753 (void)UsesNum; 754 (void)ReBasesNum; 755 // Expect all uses are rebased after rebase is done. 756 assert(UsesNum == ReBasesNum && "Not all uses are rebased"); 757 758 NumConstantsHoisted++; 759 760 // Base constant is also included in ConstInfo.RebasedConstants, so 761 // deduct 1 from ConstInfo.RebasedConstants.size(). 762 NumConstantsRebased = ConstInfo.RebasedConstants.size() - 1; 763 764 MadeChange = true; 765 } 766 return MadeChange; 767 } 768 769 /// Check all cast instructions we made a copy of and remove them if they 770 /// have no more users. 771 void ConstantHoistingPass::deleteDeadCastInst() const { 772 for (auto const &I : ClonedCastMap) 773 if (I.first->use_empty()) 774 I.first->eraseFromParent(); 775 } 776 777 /// Optimize expensive integer constants in the given function. 778 bool ConstantHoistingPass::runImpl(Function &Fn, TargetTransformInfo &TTI, 779 DominatorTree &DT, BlockFrequencyInfo *BFI, 780 BasicBlock &Entry) { 781 this->TTI = &TTI; 782 this->DT = &DT; 783 this->BFI = BFI; 784 this->Entry = &Entry; 785 // Collect all constant candidates. 786 collectConstantCandidates(Fn); 787 788 // There are no constant candidates to worry about. 789 if (ConstCandVec.empty()) 790 return false; 791 792 // Combine constants that can be easily materialized with an add from a common 793 // base constant. 794 findBaseConstants(); 795 796 // There are no constants to emit. 797 if (ConstantVec.empty()) 798 return false; 799 800 // Finally hoist the base constant and emit materialization code for dependent 801 // constants. 802 bool MadeChange = emitBaseConstants(); 803 804 // Cleanup dead instructions. 805 deleteDeadCastInst(); 806 807 return MadeChange; 808 } 809 810 PreservedAnalyses ConstantHoistingPass::run(Function &F, 811 FunctionAnalysisManager &AM) { 812 auto &DT = AM.getResult<DominatorTreeAnalysis>(F); 813 auto &TTI = AM.getResult<TargetIRAnalysis>(F); 814 auto BFI = ConstHoistWithBlockFrequency 815 ? &AM.getResult<BlockFrequencyAnalysis>(F) 816 : nullptr; 817 if (!runImpl(F, TTI, DT, BFI, F.getEntryBlock())) 818 return PreservedAnalyses::all(); 819 820 PreservedAnalyses PA; 821 PA.preserveSet<CFGAnalyses>(); 822 return PA; 823 } 824