InlineCost.cpp revision dce4a407a24b04eebc6a376f8e62b41aaa7b071f
1//===- InlineCost.cpp - Cost analysis for inliner -------------------------===// 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 file implements inline cost analysis. 11// 12//===----------------------------------------------------------------------===// 13 14#include "llvm/Analysis/InlineCost.h" 15#include "llvm/ADT/STLExtras.h" 16#include "llvm/ADT/SetVector.h" 17#include "llvm/ADT/SmallPtrSet.h" 18#include "llvm/ADT/SmallVector.h" 19#include "llvm/ADT/Statistic.h" 20#include "llvm/Analysis/ConstantFolding.h" 21#include "llvm/Analysis/InstructionSimplify.h" 22#include "llvm/Analysis/TargetTransformInfo.h" 23#include "llvm/IR/CallSite.h" 24#include "llvm/IR/CallingConv.h" 25#include "llvm/IR/DataLayout.h" 26#include "llvm/IR/GetElementPtrTypeIterator.h" 27#include "llvm/IR/GlobalAlias.h" 28#include "llvm/IR/InstVisitor.h" 29#include "llvm/IR/IntrinsicInst.h" 30#include "llvm/IR/Operator.h" 31#include "llvm/Support/Debug.h" 32#include "llvm/Support/raw_ostream.h" 33 34using namespace llvm; 35 36#define DEBUG_TYPE "inline-cost" 37 38STATISTIC(NumCallsAnalyzed, "Number of call sites analyzed"); 39 40namespace { 41 42class CallAnalyzer : public InstVisitor<CallAnalyzer, bool> { 43 typedef InstVisitor<CallAnalyzer, bool> Base; 44 friend class InstVisitor<CallAnalyzer, bool>; 45 46 // DataLayout if available, or null. 47 const DataLayout *const DL; 48 49 /// The TargetTransformInfo available for this compilation. 50 const TargetTransformInfo &TTI; 51 52 // The called function. 53 Function &F; 54 55 int Threshold; 56 int Cost; 57 58 bool IsCallerRecursive; 59 bool IsRecursiveCall; 60 bool ExposesReturnsTwice; 61 bool HasDynamicAlloca; 62 bool ContainsNoDuplicateCall; 63 bool HasReturn; 64 bool HasIndirectBr; 65 66 /// Number of bytes allocated statically by the callee. 67 uint64_t AllocatedSize; 68 unsigned NumInstructions, NumVectorInstructions; 69 int FiftyPercentVectorBonus, TenPercentVectorBonus; 70 int VectorBonus; 71 72 // While we walk the potentially-inlined instructions, we build up and 73 // maintain a mapping of simplified values specific to this callsite. The 74 // idea is to propagate any special information we have about arguments to 75 // this call through the inlinable section of the function, and account for 76 // likely simplifications post-inlining. The most important aspect we track 77 // is CFG altering simplifications -- when we prove a basic block dead, that 78 // can cause dramatic shifts in the cost of inlining a function. 79 DenseMap<Value *, Constant *> SimplifiedValues; 80 81 // Keep track of the values which map back (through function arguments) to 82 // allocas on the caller stack which could be simplified through SROA. 83 DenseMap<Value *, Value *> SROAArgValues; 84 85 // The mapping of caller Alloca values to their accumulated cost savings. If 86 // we have to disable SROA for one of the allocas, this tells us how much 87 // cost must be added. 88 DenseMap<Value *, int> SROAArgCosts; 89 90 // Keep track of values which map to a pointer base and constant offset. 91 DenseMap<Value *, std::pair<Value *, APInt> > ConstantOffsetPtrs; 92 93 // Custom simplification helper routines. 94 bool isAllocaDerivedArg(Value *V); 95 bool lookupSROAArgAndCost(Value *V, Value *&Arg, 96 DenseMap<Value *, int>::iterator &CostIt); 97 void disableSROA(DenseMap<Value *, int>::iterator CostIt); 98 void disableSROA(Value *V); 99 void accumulateSROACost(DenseMap<Value *, int>::iterator CostIt, 100 int InstructionCost); 101 bool isGEPOffsetConstant(GetElementPtrInst &GEP); 102 bool accumulateGEPOffset(GEPOperator &GEP, APInt &Offset); 103 bool simplifyCallSite(Function *F, CallSite CS); 104 ConstantInt *stripAndComputeInBoundsConstantOffsets(Value *&V); 105 106 // Custom analysis routines. 107 bool analyzeBlock(BasicBlock *BB); 108 109 // Disable several entry points to the visitor so we don't accidentally use 110 // them by declaring but not defining them here. 111 void visit(Module *); void visit(Module &); 112 void visit(Function *); void visit(Function &); 113 void visit(BasicBlock *); void visit(BasicBlock &); 114 115 // Provide base case for our instruction visit. 116 bool visitInstruction(Instruction &I); 117 118 // Our visit overrides. 119 bool visitAlloca(AllocaInst &I); 120 bool visitPHI(PHINode &I); 121 bool visitGetElementPtr(GetElementPtrInst &I); 122 bool visitBitCast(BitCastInst &I); 123 bool visitPtrToInt(PtrToIntInst &I); 124 bool visitIntToPtr(IntToPtrInst &I); 125 bool visitCastInst(CastInst &I); 126 bool visitUnaryInstruction(UnaryInstruction &I); 127 bool visitCmpInst(CmpInst &I); 128 bool visitSub(BinaryOperator &I); 129 bool visitBinaryOperator(BinaryOperator &I); 130 bool visitLoad(LoadInst &I); 131 bool visitStore(StoreInst &I); 132 bool visitExtractValue(ExtractValueInst &I); 133 bool visitInsertValue(InsertValueInst &I); 134 bool visitCallSite(CallSite CS); 135 bool visitReturnInst(ReturnInst &RI); 136 bool visitBranchInst(BranchInst &BI); 137 bool visitSwitchInst(SwitchInst &SI); 138 bool visitIndirectBrInst(IndirectBrInst &IBI); 139 bool visitResumeInst(ResumeInst &RI); 140 bool visitUnreachableInst(UnreachableInst &I); 141 142public: 143 CallAnalyzer(const DataLayout *DL, const TargetTransformInfo &TTI, 144 Function &Callee, int Threshold) 145 : DL(DL), TTI(TTI), F(Callee), Threshold(Threshold), Cost(0), 146 IsCallerRecursive(false), IsRecursiveCall(false), 147 ExposesReturnsTwice(false), HasDynamicAlloca(false), 148 ContainsNoDuplicateCall(false), HasReturn(false), HasIndirectBr(false), 149 AllocatedSize(0), NumInstructions(0), NumVectorInstructions(0), 150 FiftyPercentVectorBonus(0), TenPercentVectorBonus(0), VectorBonus(0), 151 NumConstantArgs(0), NumConstantOffsetPtrArgs(0), NumAllocaArgs(0), 152 NumConstantPtrCmps(0), NumConstantPtrDiffs(0), 153 NumInstructionsSimplified(0), SROACostSavings(0), 154 SROACostSavingsLost(0) {} 155 156 bool analyzeCall(CallSite CS); 157 158 int getThreshold() { return Threshold; } 159 int getCost() { return Cost; } 160 161 // Keep a bunch of stats about the cost savings found so we can print them 162 // out when debugging. 163 unsigned NumConstantArgs; 164 unsigned NumConstantOffsetPtrArgs; 165 unsigned NumAllocaArgs; 166 unsigned NumConstantPtrCmps; 167 unsigned NumConstantPtrDiffs; 168 unsigned NumInstructionsSimplified; 169 unsigned SROACostSavings; 170 unsigned SROACostSavingsLost; 171 172 void dump(); 173}; 174 175} // namespace 176 177/// \brief Test whether the given value is an Alloca-derived function argument. 178bool CallAnalyzer::isAllocaDerivedArg(Value *V) { 179 return SROAArgValues.count(V); 180} 181 182/// \brief Lookup the SROA-candidate argument and cost iterator which V maps to. 183/// Returns false if V does not map to a SROA-candidate. 184bool CallAnalyzer::lookupSROAArgAndCost( 185 Value *V, Value *&Arg, DenseMap<Value *, int>::iterator &CostIt) { 186 if (SROAArgValues.empty() || SROAArgCosts.empty()) 187 return false; 188 189 DenseMap<Value *, Value *>::iterator ArgIt = SROAArgValues.find(V); 190 if (ArgIt == SROAArgValues.end()) 191 return false; 192 193 Arg = ArgIt->second; 194 CostIt = SROAArgCosts.find(Arg); 195 return CostIt != SROAArgCosts.end(); 196} 197 198/// \brief Disable SROA for the candidate marked by this cost iterator. 199/// 200/// This marks the candidate as no longer viable for SROA, and adds the cost 201/// savings associated with it back into the inline cost measurement. 202void CallAnalyzer::disableSROA(DenseMap<Value *, int>::iterator CostIt) { 203 // If we're no longer able to perform SROA we need to undo its cost savings 204 // and prevent subsequent analysis. 205 Cost += CostIt->second; 206 SROACostSavings -= CostIt->second; 207 SROACostSavingsLost += CostIt->second; 208 SROAArgCosts.erase(CostIt); 209} 210 211/// \brief If 'V' maps to a SROA candidate, disable SROA for it. 212void CallAnalyzer::disableSROA(Value *V) { 213 Value *SROAArg; 214 DenseMap<Value *, int>::iterator CostIt; 215 if (lookupSROAArgAndCost(V, SROAArg, CostIt)) 216 disableSROA(CostIt); 217} 218 219/// \brief Accumulate the given cost for a particular SROA candidate. 220void CallAnalyzer::accumulateSROACost(DenseMap<Value *, int>::iterator CostIt, 221 int InstructionCost) { 222 CostIt->second += InstructionCost; 223 SROACostSavings += InstructionCost; 224} 225 226/// \brief Check whether a GEP's indices are all constant. 227/// 228/// Respects any simplified values known during the analysis of this callsite. 229bool CallAnalyzer::isGEPOffsetConstant(GetElementPtrInst &GEP) { 230 for (User::op_iterator I = GEP.idx_begin(), E = GEP.idx_end(); I != E; ++I) 231 if (!isa<Constant>(*I) && !SimplifiedValues.lookup(*I)) 232 return false; 233 234 return true; 235} 236 237/// \brief Accumulate a constant GEP offset into an APInt if possible. 238/// 239/// Returns false if unable to compute the offset for any reason. Respects any 240/// simplified values known during the analysis of this callsite. 241bool CallAnalyzer::accumulateGEPOffset(GEPOperator &GEP, APInt &Offset) { 242 if (!DL) 243 return false; 244 245 unsigned IntPtrWidth = DL->getPointerSizeInBits(); 246 assert(IntPtrWidth == Offset.getBitWidth()); 247 248 for (gep_type_iterator GTI = gep_type_begin(GEP), GTE = gep_type_end(GEP); 249 GTI != GTE; ++GTI) { 250 ConstantInt *OpC = dyn_cast<ConstantInt>(GTI.getOperand()); 251 if (!OpC) 252 if (Constant *SimpleOp = SimplifiedValues.lookup(GTI.getOperand())) 253 OpC = dyn_cast<ConstantInt>(SimpleOp); 254 if (!OpC) 255 return false; 256 if (OpC->isZero()) continue; 257 258 // Handle a struct index, which adds its field offset to the pointer. 259 if (StructType *STy = dyn_cast<StructType>(*GTI)) { 260 unsigned ElementIdx = OpC->getZExtValue(); 261 const StructLayout *SL = DL->getStructLayout(STy); 262 Offset += APInt(IntPtrWidth, SL->getElementOffset(ElementIdx)); 263 continue; 264 } 265 266 APInt TypeSize(IntPtrWidth, DL->getTypeAllocSize(GTI.getIndexedType())); 267 Offset += OpC->getValue().sextOrTrunc(IntPtrWidth) * TypeSize; 268 } 269 return true; 270} 271 272bool CallAnalyzer::visitAlloca(AllocaInst &I) { 273 // Check whether inlining will turn a dynamic alloca into a static 274 // alloca, and handle that case. 275 if (I.isArrayAllocation()) { 276 if (Constant *Size = SimplifiedValues.lookup(I.getArraySize())) { 277 ConstantInt *AllocSize = dyn_cast<ConstantInt>(Size); 278 assert(AllocSize && "Allocation size not a constant int?"); 279 Type *Ty = I.getAllocatedType(); 280 AllocatedSize += Ty->getPrimitiveSizeInBits() * AllocSize->getZExtValue(); 281 return Base::visitAlloca(I); 282 } 283 } 284 285 // Accumulate the allocated size. 286 if (I.isStaticAlloca()) { 287 Type *Ty = I.getAllocatedType(); 288 AllocatedSize += (DL ? DL->getTypeAllocSize(Ty) : 289 Ty->getPrimitiveSizeInBits()); 290 } 291 292 // We will happily inline static alloca instructions. 293 if (I.isStaticAlloca()) 294 return Base::visitAlloca(I); 295 296 // FIXME: This is overly conservative. Dynamic allocas are inefficient for 297 // a variety of reasons, and so we would like to not inline them into 298 // functions which don't currently have a dynamic alloca. This simply 299 // disables inlining altogether in the presence of a dynamic alloca. 300 HasDynamicAlloca = true; 301 return false; 302} 303 304bool CallAnalyzer::visitPHI(PHINode &I) { 305 // FIXME: We should potentially be tracking values through phi nodes, 306 // especially when they collapse to a single value due to deleted CFG edges 307 // during inlining. 308 309 // FIXME: We need to propagate SROA *disabling* through phi nodes, even 310 // though we don't want to propagate it's bonuses. The idea is to disable 311 // SROA if it *might* be used in an inappropriate manner. 312 313 // Phi nodes are always zero-cost. 314 return true; 315} 316 317bool CallAnalyzer::visitGetElementPtr(GetElementPtrInst &I) { 318 Value *SROAArg; 319 DenseMap<Value *, int>::iterator CostIt; 320 bool SROACandidate = lookupSROAArgAndCost(I.getPointerOperand(), 321 SROAArg, CostIt); 322 323 // Try to fold GEPs of constant-offset call site argument pointers. This 324 // requires target data and inbounds GEPs. 325 if (DL && I.isInBounds()) { 326 // Check if we have a base + offset for the pointer. 327 Value *Ptr = I.getPointerOperand(); 328 std::pair<Value *, APInt> BaseAndOffset = ConstantOffsetPtrs.lookup(Ptr); 329 if (BaseAndOffset.first) { 330 // Check if the offset of this GEP is constant, and if so accumulate it 331 // into Offset. 332 if (!accumulateGEPOffset(cast<GEPOperator>(I), BaseAndOffset.second)) { 333 // Non-constant GEPs aren't folded, and disable SROA. 334 if (SROACandidate) 335 disableSROA(CostIt); 336 return false; 337 } 338 339 // Add the result as a new mapping to Base + Offset. 340 ConstantOffsetPtrs[&I] = BaseAndOffset; 341 342 // Also handle SROA candidates here, we already know that the GEP is 343 // all-constant indexed. 344 if (SROACandidate) 345 SROAArgValues[&I] = SROAArg; 346 347 return true; 348 } 349 } 350 351 if (isGEPOffsetConstant(I)) { 352 if (SROACandidate) 353 SROAArgValues[&I] = SROAArg; 354 355 // Constant GEPs are modeled as free. 356 return true; 357 } 358 359 // Variable GEPs will require math and will disable SROA. 360 if (SROACandidate) 361 disableSROA(CostIt); 362 return false; 363} 364 365bool CallAnalyzer::visitBitCast(BitCastInst &I) { 366 // Propagate constants through bitcasts. 367 Constant *COp = dyn_cast<Constant>(I.getOperand(0)); 368 if (!COp) 369 COp = SimplifiedValues.lookup(I.getOperand(0)); 370 if (COp) 371 if (Constant *C = ConstantExpr::getBitCast(COp, I.getType())) { 372 SimplifiedValues[&I] = C; 373 return true; 374 } 375 376 // Track base/offsets through casts 377 std::pair<Value *, APInt> BaseAndOffset 378 = ConstantOffsetPtrs.lookup(I.getOperand(0)); 379 // Casts don't change the offset, just wrap it up. 380 if (BaseAndOffset.first) 381 ConstantOffsetPtrs[&I] = BaseAndOffset; 382 383 // Also look for SROA candidates here. 384 Value *SROAArg; 385 DenseMap<Value *, int>::iterator CostIt; 386 if (lookupSROAArgAndCost(I.getOperand(0), SROAArg, CostIt)) 387 SROAArgValues[&I] = SROAArg; 388 389 // Bitcasts are always zero cost. 390 return true; 391} 392 393bool CallAnalyzer::visitPtrToInt(PtrToIntInst &I) { 394 const DataLayout *DL = I.getDataLayout(); 395 // Propagate constants through ptrtoint. 396 Constant *COp = dyn_cast<Constant>(I.getOperand(0)); 397 if (!COp) 398 COp = SimplifiedValues.lookup(I.getOperand(0)); 399 if (COp) 400 if (Constant *C = ConstantExpr::getPtrToInt(COp, I.getType())) { 401 SimplifiedValues[&I] = C; 402 return true; 403 } 404 405 // Track base/offset pairs when converted to a plain integer provided the 406 // integer is large enough to represent the pointer. 407 unsigned IntegerSize = I.getType()->getScalarSizeInBits(); 408 if (DL && IntegerSize >= DL->getPointerSizeInBits()) { 409 std::pair<Value *, APInt> BaseAndOffset 410 = ConstantOffsetPtrs.lookup(I.getOperand(0)); 411 if (BaseAndOffset.first) 412 ConstantOffsetPtrs[&I] = BaseAndOffset; 413 } 414 415 // This is really weird. Technically, ptrtoint will disable SROA. However, 416 // unless that ptrtoint is *used* somewhere in the live basic blocks after 417 // inlining, it will be nuked, and SROA should proceed. All of the uses which 418 // would block SROA would also block SROA if applied directly to a pointer, 419 // and so we can just add the integer in here. The only places where SROA is 420 // preserved either cannot fire on an integer, or won't in-and-of themselves 421 // disable SROA (ext) w/o some later use that we would see and disable. 422 Value *SROAArg; 423 DenseMap<Value *, int>::iterator CostIt; 424 if (lookupSROAArgAndCost(I.getOperand(0), SROAArg, CostIt)) 425 SROAArgValues[&I] = SROAArg; 426 427 return TargetTransformInfo::TCC_Free == TTI.getUserCost(&I); 428} 429 430bool CallAnalyzer::visitIntToPtr(IntToPtrInst &I) { 431 const DataLayout *DL = I.getDataLayout(); 432 // Propagate constants through ptrtoint. 433 Constant *COp = dyn_cast<Constant>(I.getOperand(0)); 434 if (!COp) 435 COp = SimplifiedValues.lookup(I.getOperand(0)); 436 if (COp) 437 if (Constant *C = ConstantExpr::getIntToPtr(COp, I.getType())) { 438 SimplifiedValues[&I] = C; 439 return true; 440 } 441 442 // Track base/offset pairs when round-tripped through a pointer without 443 // modifications provided the integer is not too large. 444 Value *Op = I.getOperand(0); 445 unsigned IntegerSize = Op->getType()->getScalarSizeInBits(); 446 if (DL && IntegerSize <= DL->getPointerSizeInBits()) { 447 std::pair<Value *, APInt> BaseAndOffset = ConstantOffsetPtrs.lookup(Op); 448 if (BaseAndOffset.first) 449 ConstantOffsetPtrs[&I] = BaseAndOffset; 450 } 451 452 // "Propagate" SROA here in the same manner as we do for ptrtoint above. 453 Value *SROAArg; 454 DenseMap<Value *, int>::iterator CostIt; 455 if (lookupSROAArgAndCost(Op, SROAArg, CostIt)) 456 SROAArgValues[&I] = SROAArg; 457 458 return TargetTransformInfo::TCC_Free == TTI.getUserCost(&I); 459} 460 461bool CallAnalyzer::visitCastInst(CastInst &I) { 462 // Propagate constants through ptrtoint. 463 Constant *COp = dyn_cast<Constant>(I.getOperand(0)); 464 if (!COp) 465 COp = SimplifiedValues.lookup(I.getOperand(0)); 466 if (COp) 467 if (Constant *C = ConstantExpr::getCast(I.getOpcode(), COp, I.getType())) { 468 SimplifiedValues[&I] = C; 469 return true; 470 } 471 472 // Disable SROA in the face of arbitrary casts we don't whitelist elsewhere. 473 disableSROA(I.getOperand(0)); 474 475 return TargetTransformInfo::TCC_Free == TTI.getUserCost(&I); 476} 477 478bool CallAnalyzer::visitUnaryInstruction(UnaryInstruction &I) { 479 Value *Operand = I.getOperand(0); 480 Constant *COp = dyn_cast<Constant>(Operand); 481 if (!COp) 482 COp = SimplifiedValues.lookup(Operand); 483 if (COp) 484 if (Constant *C = ConstantFoldInstOperands(I.getOpcode(), I.getType(), 485 COp, DL)) { 486 SimplifiedValues[&I] = C; 487 return true; 488 } 489 490 // Disable any SROA on the argument to arbitrary unary operators. 491 disableSROA(Operand); 492 493 return false; 494} 495 496bool CallAnalyzer::visitCmpInst(CmpInst &I) { 497 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1); 498 // First try to handle simplified comparisons. 499 if (!isa<Constant>(LHS)) 500 if (Constant *SimpleLHS = SimplifiedValues.lookup(LHS)) 501 LHS = SimpleLHS; 502 if (!isa<Constant>(RHS)) 503 if (Constant *SimpleRHS = SimplifiedValues.lookup(RHS)) 504 RHS = SimpleRHS; 505 if (Constant *CLHS = dyn_cast<Constant>(LHS)) { 506 if (Constant *CRHS = dyn_cast<Constant>(RHS)) 507 if (Constant *C = ConstantExpr::getCompare(I.getPredicate(), CLHS, CRHS)) { 508 SimplifiedValues[&I] = C; 509 return true; 510 } 511 } 512 513 if (I.getOpcode() == Instruction::FCmp) 514 return false; 515 516 // Otherwise look for a comparison between constant offset pointers with 517 // a common base. 518 Value *LHSBase, *RHSBase; 519 APInt LHSOffset, RHSOffset; 520 std::tie(LHSBase, LHSOffset) = ConstantOffsetPtrs.lookup(LHS); 521 if (LHSBase) { 522 std::tie(RHSBase, RHSOffset) = ConstantOffsetPtrs.lookup(RHS); 523 if (RHSBase && LHSBase == RHSBase) { 524 // We have common bases, fold the icmp to a constant based on the 525 // offsets. 526 Constant *CLHS = ConstantInt::get(LHS->getContext(), LHSOffset); 527 Constant *CRHS = ConstantInt::get(RHS->getContext(), RHSOffset); 528 if (Constant *C = ConstantExpr::getICmp(I.getPredicate(), CLHS, CRHS)) { 529 SimplifiedValues[&I] = C; 530 ++NumConstantPtrCmps; 531 return true; 532 } 533 } 534 } 535 536 // If the comparison is an equality comparison with null, we can simplify it 537 // for any alloca-derived argument. 538 if (I.isEquality() && isa<ConstantPointerNull>(I.getOperand(1))) 539 if (isAllocaDerivedArg(I.getOperand(0))) { 540 // We can actually predict the result of comparisons between an 541 // alloca-derived value and null. Note that this fires regardless of 542 // SROA firing. 543 bool IsNotEqual = I.getPredicate() == CmpInst::ICMP_NE; 544 SimplifiedValues[&I] = IsNotEqual ? ConstantInt::getTrue(I.getType()) 545 : ConstantInt::getFalse(I.getType()); 546 return true; 547 } 548 549 // Finally check for SROA candidates in comparisons. 550 Value *SROAArg; 551 DenseMap<Value *, int>::iterator CostIt; 552 if (lookupSROAArgAndCost(I.getOperand(0), SROAArg, CostIt)) { 553 if (isa<ConstantPointerNull>(I.getOperand(1))) { 554 accumulateSROACost(CostIt, InlineConstants::InstrCost); 555 return true; 556 } 557 558 disableSROA(CostIt); 559 } 560 561 return false; 562} 563 564bool CallAnalyzer::visitSub(BinaryOperator &I) { 565 // Try to handle a special case: we can fold computing the difference of two 566 // constant-related pointers. 567 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1); 568 Value *LHSBase, *RHSBase; 569 APInt LHSOffset, RHSOffset; 570 std::tie(LHSBase, LHSOffset) = ConstantOffsetPtrs.lookup(LHS); 571 if (LHSBase) { 572 std::tie(RHSBase, RHSOffset) = ConstantOffsetPtrs.lookup(RHS); 573 if (RHSBase && LHSBase == RHSBase) { 574 // We have common bases, fold the subtract to a constant based on the 575 // offsets. 576 Constant *CLHS = ConstantInt::get(LHS->getContext(), LHSOffset); 577 Constant *CRHS = ConstantInt::get(RHS->getContext(), RHSOffset); 578 if (Constant *C = ConstantExpr::getSub(CLHS, CRHS)) { 579 SimplifiedValues[&I] = C; 580 ++NumConstantPtrDiffs; 581 return true; 582 } 583 } 584 } 585 586 // Otherwise, fall back to the generic logic for simplifying and handling 587 // instructions. 588 return Base::visitSub(I); 589} 590 591bool CallAnalyzer::visitBinaryOperator(BinaryOperator &I) { 592 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1); 593 if (!isa<Constant>(LHS)) 594 if (Constant *SimpleLHS = SimplifiedValues.lookup(LHS)) 595 LHS = SimpleLHS; 596 if (!isa<Constant>(RHS)) 597 if (Constant *SimpleRHS = SimplifiedValues.lookup(RHS)) 598 RHS = SimpleRHS; 599 Value *SimpleV = SimplifyBinOp(I.getOpcode(), LHS, RHS, DL); 600 if (Constant *C = dyn_cast_or_null<Constant>(SimpleV)) { 601 SimplifiedValues[&I] = C; 602 return true; 603 } 604 605 // Disable any SROA on arguments to arbitrary, unsimplified binary operators. 606 disableSROA(LHS); 607 disableSROA(RHS); 608 609 return false; 610} 611 612bool CallAnalyzer::visitLoad(LoadInst &I) { 613 Value *SROAArg; 614 DenseMap<Value *, int>::iterator CostIt; 615 if (lookupSROAArgAndCost(I.getOperand(0), SROAArg, CostIt)) { 616 if (I.isSimple()) { 617 accumulateSROACost(CostIt, InlineConstants::InstrCost); 618 return true; 619 } 620 621 disableSROA(CostIt); 622 } 623 624 return false; 625} 626 627bool CallAnalyzer::visitStore(StoreInst &I) { 628 Value *SROAArg; 629 DenseMap<Value *, int>::iterator CostIt; 630 if (lookupSROAArgAndCost(I.getOperand(0), SROAArg, CostIt)) { 631 if (I.isSimple()) { 632 accumulateSROACost(CostIt, InlineConstants::InstrCost); 633 return true; 634 } 635 636 disableSROA(CostIt); 637 } 638 639 return false; 640} 641 642bool CallAnalyzer::visitExtractValue(ExtractValueInst &I) { 643 // Constant folding for extract value is trivial. 644 Constant *C = dyn_cast<Constant>(I.getAggregateOperand()); 645 if (!C) 646 C = SimplifiedValues.lookup(I.getAggregateOperand()); 647 if (C) { 648 SimplifiedValues[&I] = ConstantExpr::getExtractValue(C, I.getIndices()); 649 return true; 650 } 651 652 // SROA can look through these but give them a cost. 653 return false; 654} 655 656bool CallAnalyzer::visitInsertValue(InsertValueInst &I) { 657 // Constant folding for insert value is trivial. 658 Constant *AggC = dyn_cast<Constant>(I.getAggregateOperand()); 659 if (!AggC) 660 AggC = SimplifiedValues.lookup(I.getAggregateOperand()); 661 Constant *InsertedC = dyn_cast<Constant>(I.getInsertedValueOperand()); 662 if (!InsertedC) 663 InsertedC = SimplifiedValues.lookup(I.getInsertedValueOperand()); 664 if (AggC && InsertedC) { 665 SimplifiedValues[&I] = ConstantExpr::getInsertValue(AggC, InsertedC, 666 I.getIndices()); 667 return true; 668 } 669 670 // SROA can look through these but give them a cost. 671 return false; 672} 673 674/// \brief Try to simplify a call site. 675/// 676/// Takes a concrete function and callsite and tries to actually simplify it by 677/// analyzing the arguments and call itself with instsimplify. Returns true if 678/// it has simplified the callsite to some other entity (a constant), making it 679/// free. 680bool CallAnalyzer::simplifyCallSite(Function *F, CallSite CS) { 681 // FIXME: Using the instsimplify logic directly for this is inefficient 682 // because we have to continually rebuild the argument list even when no 683 // simplifications can be performed. Until that is fixed with remapping 684 // inside of instsimplify, directly constant fold calls here. 685 if (!canConstantFoldCallTo(F)) 686 return false; 687 688 // Try to re-map the arguments to constants. 689 SmallVector<Constant *, 4> ConstantArgs; 690 ConstantArgs.reserve(CS.arg_size()); 691 for (CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end(); 692 I != E; ++I) { 693 Constant *C = dyn_cast<Constant>(*I); 694 if (!C) 695 C = dyn_cast_or_null<Constant>(SimplifiedValues.lookup(*I)); 696 if (!C) 697 return false; // This argument doesn't map to a constant. 698 699 ConstantArgs.push_back(C); 700 } 701 if (Constant *C = ConstantFoldCall(F, ConstantArgs)) { 702 SimplifiedValues[CS.getInstruction()] = C; 703 return true; 704 } 705 706 return false; 707} 708 709bool CallAnalyzer::visitCallSite(CallSite CS) { 710 if (CS.hasFnAttr(Attribute::ReturnsTwice) && 711 !F.getAttributes().hasAttribute(AttributeSet::FunctionIndex, 712 Attribute::ReturnsTwice)) { 713 // This aborts the entire analysis. 714 ExposesReturnsTwice = true; 715 return false; 716 } 717 if (CS.isCall() && 718 cast<CallInst>(CS.getInstruction())->cannotDuplicate()) 719 ContainsNoDuplicateCall = true; 720 721 if (Function *F = CS.getCalledFunction()) { 722 // When we have a concrete function, first try to simplify it directly. 723 if (simplifyCallSite(F, CS)) 724 return true; 725 726 // Next check if it is an intrinsic we know about. 727 // FIXME: Lift this into part of the InstVisitor. 728 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(CS.getInstruction())) { 729 switch (II->getIntrinsicID()) { 730 default: 731 return Base::visitCallSite(CS); 732 733 case Intrinsic::memset: 734 case Intrinsic::memcpy: 735 case Intrinsic::memmove: 736 // SROA can usually chew through these intrinsics, but they aren't free. 737 return false; 738 } 739 } 740 741 if (F == CS.getInstruction()->getParent()->getParent()) { 742 // This flag will fully abort the analysis, so don't bother with anything 743 // else. 744 IsRecursiveCall = true; 745 return false; 746 } 747 748 if (TTI.isLoweredToCall(F)) { 749 // We account for the average 1 instruction per call argument setup 750 // here. 751 Cost += CS.arg_size() * InlineConstants::InstrCost; 752 753 // Everything other than inline ASM will also have a significant cost 754 // merely from making the call. 755 if (!isa<InlineAsm>(CS.getCalledValue())) 756 Cost += InlineConstants::CallPenalty; 757 } 758 759 return Base::visitCallSite(CS); 760 } 761 762 // Otherwise we're in a very special case -- an indirect function call. See 763 // if we can be particularly clever about this. 764 Value *Callee = CS.getCalledValue(); 765 766 // First, pay the price of the argument setup. We account for the average 767 // 1 instruction per call argument setup here. 768 Cost += CS.arg_size() * InlineConstants::InstrCost; 769 770 // Next, check if this happens to be an indirect function call to a known 771 // function in this inline context. If not, we've done all we can. 772 Function *F = dyn_cast_or_null<Function>(SimplifiedValues.lookup(Callee)); 773 if (!F) 774 return Base::visitCallSite(CS); 775 776 // If we have a constant that we are calling as a function, we can peer 777 // through it and see the function target. This happens not infrequently 778 // during devirtualization and so we want to give it a hefty bonus for 779 // inlining, but cap that bonus in the event that inlining wouldn't pan 780 // out. Pretend to inline the function, with a custom threshold. 781 CallAnalyzer CA(DL, TTI, *F, InlineConstants::IndirectCallThreshold); 782 if (CA.analyzeCall(CS)) { 783 // We were able to inline the indirect call! Subtract the cost from the 784 // bonus we want to apply, but don't go below zero. 785 Cost -= std::max(0, InlineConstants::IndirectCallThreshold - CA.getCost()); 786 } 787 788 return Base::visitCallSite(CS); 789} 790 791bool CallAnalyzer::visitReturnInst(ReturnInst &RI) { 792 // At least one return instruction will be free after inlining. 793 bool Free = !HasReturn; 794 HasReturn = true; 795 return Free; 796} 797 798bool CallAnalyzer::visitBranchInst(BranchInst &BI) { 799 // We model unconditional branches as essentially free -- they really 800 // shouldn't exist at all, but handling them makes the behavior of the 801 // inliner more regular and predictable. Interestingly, conditional branches 802 // which will fold away are also free. 803 return BI.isUnconditional() || isa<ConstantInt>(BI.getCondition()) || 804 dyn_cast_or_null<ConstantInt>( 805 SimplifiedValues.lookup(BI.getCondition())); 806} 807 808bool CallAnalyzer::visitSwitchInst(SwitchInst &SI) { 809 // We model unconditional switches as free, see the comments on handling 810 // branches. 811 if (isa<ConstantInt>(SI.getCondition())) 812 return true; 813 if (Value *V = SimplifiedValues.lookup(SI.getCondition())) 814 if (isa<ConstantInt>(V)) 815 return true; 816 817 // Otherwise, we need to accumulate a cost proportional to the number of 818 // distinct successor blocks. This fan-out in the CFG cannot be represented 819 // for free even if we can represent the core switch as a jumptable that 820 // takes a single instruction. 821 // 822 // NB: We convert large switches which are just used to initialize large phi 823 // nodes to lookup tables instead in simplify-cfg, so this shouldn't prevent 824 // inlining those. It will prevent inlining in cases where the optimization 825 // does not (yet) fire. 826 SmallPtrSet<BasicBlock *, 8> SuccessorBlocks; 827 SuccessorBlocks.insert(SI.getDefaultDest()); 828 for (auto I = SI.case_begin(), E = SI.case_end(); I != E; ++I) 829 SuccessorBlocks.insert(I.getCaseSuccessor()); 830 // Add cost corresponding to the number of distinct destinations. The first 831 // we model as free because of fallthrough. 832 Cost += (SuccessorBlocks.size() - 1) * InlineConstants::InstrCost; 833 return false; 834} 835 836bool CallAnalyzer::visitIndirectBrInst(IndirectBrInst &IBI) { 837 // We never want to inline functions that contain an indirectbr. This is 838 // incorrect because all the blockaddress's (in static global initializers 839 // for example) would be referring to the original function, and this 840 // indirect jump would jump from the inlined copy of the function into the 841 // original function which is extremely undefined behavior. 842 // FIXME: This logic isn't really right; we can safely inline functions with 843 // indirectbr's as long as no other function or global references the 844 // blockaddress of a block within the current function. And as a QOI issue, 845 // if someone is using a blockaddress without an indirectbr, and that 846 // reference somehow ends up in another function or global, we probably don't 847 // want to inline this function. 848 HasIndirectBr = true; 849 return false; 850} 851 852bool CallAnalyzer::visitResumeInst(ResumeInst &RI) { 853 // FIXME: It's not clear that a single instruction is an accurate model for 854 // the inline cost of a resume instruction. 855 return false; 856} 857 858bool CallAnalyzer::visitUnreachableInst(UnreachableInst &I) { 859 // FIXME: It might be reasonably to discount the cost of instructions leading 860 // to unreachable as they have the lowest possible impact on both runtime and 861 // code size. 862 return true; // No actual code is needed for unreachable. 863} 864 865bool CallAnalyzer::visitInstruction(Instruction &I) { 866 // Some instructions are free. All of the free intrinsics can also be 867 // handled by SROA, etc. 868 if (TargetTransformInfo::TCC_Free == TTI.getUserCost(&I)) 869 return true; 870 871 // We found something we don't understand or can't handle. Mark any SROA-able 872 // values in the operand list as no longer viable. 873 for (User::op_iterator OI = I.op_begin(), OE = I.op_end(); OI != OE; ++OI) 874 disableSROA(*OI); 875 876 return false; 877} 878 879 880/// \brief Analyze a basic block for its contribution to the inline cost. 881/// 882/// This method walks the analyzer over every instruction in the given basic 883/// block and accounts for their cost during inlining at this callsite. It 884/// aborts early if the threshold has been exceeded or an impossible to inline 885/// construct has been detected. It returns false if inlining is no longer 886/// viable, and true if inlining remains viable. 887bool CallAnalyzer::analyzeBlock(BasicBlock *BB) { 888 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) { 889 // FIXME: Currently, the number of instructions in a function regardless of 890 // our ability to simplify them during inline to constants or dead code, 891 // are actually used by the vector bonus heuristic. As long as that's true, 892 // we have to special case debug intrinsics here to prevent differences in 893 // inlining due to debug symbols. Eventually, the number of unsimplified 894 // instructions shouldn't factor into the cost computation, but until then, 895 // hack around it here. 896 if (isa<DbgInfoIntrinsic>(I)) 897 continue; 898 899 ++NumInstructions; 900 if (isa<ExtractElementInst>(I) || I->getType()->isVectorTy()) 901 ++NumVectorInstructions; 902 903 // If the instruction simplified to a constant, there is no cost to this 904 // instruction. Visit the instructions using our InstVisitor to account for 905 // all of the per-instruction logic. The visit tree returns true if we 906 // consumed the instruction in any way, and false if the instruction's base 907 // cost should count against inlining. 908 if (Base::visit(I)) 909 ++NumInstructionsSimplified; 910 else 911 Cost += InlineConstants::InstrCost; 912 913 // If the visit this instruction detected an uninlinable pattern, abort. 914 if (IsRecursiveCall || ExposesReturnsTwice || HasDynamicAlloca || 915 HasIndirectBr) 916 return false; 917 918 // If the caller is a recursive function then we don't want to inline 919 // functions which allocate a lot of stack space because it would increase 920 // the caller stack usage dramatically. 921 if (IsCallerRecursive && 922 AllocatedSize > InlineConstants::TotalAllocaSizeRecursiveCaller) 923 return false; 924 925 if (NumVectorInstructions > NumInstructions/2) 926 VectorBonus = FiftyPercentVectorBonus; 927 else if (NumVectorInstructions > NumInstructions/10) 928 VectorBonus = TenPercentVectorBonus; 929 else 930 VectorBonus = 0; 931 932 // Check if we've past the threshold so we don't spin in huge basic 933 // blocks that will never inline. 934 if (Cost > (Threshold + VectorBonus)) 935 return false; 936 } 937 938 return true; 939} 940 941/// \brief Compute the base pointer and cumulative constant offsets for V. 942/// 943/// This strips all constant offsets off of V, leaving it the base pointer, and 944/// accumulates the total constant offset applied in the returned constant. It 945/// returns 0 if V is not a pointer, and returns the constant '0' if there are 946/// no constant offsets applied. 947ConstantInt *CallAnalyzer::stripAndComputeInBoundsConstantOffsets(Value *&V) { 948 if (!DL || !V->getType()->isPointerTy()) 949 return nullptr; 950 951 unsigned IntPtrWidth = DL->getPointerSizeInBits(); 952 APInt Offset = APInt::getNullValue(IntPtrWidth); 953 954 // Even though we don't look through PHI nodes, we could be called on an 955 // instruction in an unreachable block, which may be on a cycle. 956 SmallPtrSet<Value *, 4> Visited; 957 Visited.insert(V); 958 do { 959 if (GEPOperator *GEP = dyn_cast<GEPOperator>(V)) { 960 if (!GEP->isInBounds() || !accumulateGEPOffset(*GEP, Offset)) 961 return nullptr; 962 V = GEP->getPointerOperand(); 963 } else if (Operator::getOpcode(V) == Instruction::BitCast) { 964 V = cast<Operator>(V)->getOperand(0); 965 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) { 966 if (GA->mayBeOverridden()) 967 break; 968 V = GA->getAliasee(); 969 } else { 970 break; 971 } 972 assert(V->getType()->isPointerTy() && "Unexpected operand type!"); 973 } while (Visited.insert(V)); 974 975 Type *IntPtrTy = DL->getIntPtrType(V->getContext()); 976 return cast<ConstantInt>(ConstantInt::get(IntPtrTy, Offset)); 977} 978 979/// \brief Analyze a call site for potential inlining. 980/// 981/// Returns true if inlining this call is viable, and false if it is not 982/// viable. It computes the cost and adjusts the threshold based on numerous 983/// factors and heuristics. If this method returns false but the computed cost 984/// is below the computed threshold, then inlining was forcibly disabled by 985/// some artifact of the routine. 986bool CallAnalyzer::analyzeCall(CallSite CS) { 987 ++NumCallsAnalyzed; 988 989 // Track whether the post-inlining function would have more than one basic 990 // block. A single basic block is often intended for inlining. Balloon the 991 // threshold by 50% until we pass the single-BB phase. 992 bool SingleBB = true; 993 int SingleBBBonus = Threshold / 2; 994 Threshold += SingleBBBonus; 995 996 // Perform some tweaks to the cost and threshold based on the direct 997 // callsite information. 998 999 // We want to more aggressively inline vector-dense kernels, so up the 1000 // threshold, and we'll lower it if the % of vector instructions gets too 1001 // low. 1002 assert(NumInstructions == 0); 1003 assert(NumVectorInstructions == 0); 1004 FiftyPercentVectorBonus = Threshold; 1005 TenPercentVectorBonus = Threshold / 2; 1006 1007 // Give out bonuses per argument, as the instructions setting them up will 1008 // be gone after inlining. 1009 for (unsigned I = 0, E = CS.arg_size(); I != E; ++I) { 1010 if (DL && CS.isByValArgument(I)) { 1011 // We approximate the number of loads and stores needed by dividing the 1012 // size of the byval type by the target's pointer size. 1013 PointerType *PTy = cast<PointerType>(CS.getArgument(I)->getType()); 1014 unsigned TypeSize = DL->getTypeSizeInBits(PTy->getElementType()); 1015 unsigned PointerSize = DL->getPointerSizeInBits(); 1016 // Ceiling division. 1017 unsigned NumStores = (TypeSize + PointerSize - 1) / PointerSize; 1018 1019 // If it generates more than 8 stores it is likely to be expanded as an 1020 // inline memcpy so we take that as an upper bound. Otherwise we assume 1021 // one load and one store per word copied. 1022 // FIXME: The maxStoresPerMemcpy setting from the target should be used 1023 // here instead of a magic number of 8, but it's not available via 1024 // DataLayout. 1025 NumStores = std::min(NumStores, 8U); 1026 1027 Cost -= 2 * NumStores * InlineConstants::InstrCost; 1028 } else { 1029 // For non-byval arguments subtract off one instruction per call 1030 // argument. 1031 Cost -= InlineConstants::InstrCost; 1032 } 1033 } 1034 1035 // If there is only one call of the function, and it has internal linkage, 1036 // the cost of inlining it drops dramatically. 1037 bool OnlyOneCallAndLocalLinkage = F.hasLocalLinkage() && F.hasOneUse() && 1038 &F == CS.getCalledFunction(); 1039 if (OnlyOneCallAndLocalLinkage) 1040 Cost += InlineConstants::LastCallToStaticBonus; 1041 1042 // If the instruction after the call, or if the normal destination of the 1043 // invoke is an unreachable instruction, the function is noreturn. As such, 1044 // there is little point in inlining this unless there is literally zero 1045 // cost. 1046 Instruction *Instr = CS.getInstruction(); 1047 if (InvokeInst *II = dyn_cast<InvokeInst>(Instr)) { 1048 if (isa<UnreachableInst>(II->getNormalDest()->begin())) 1049 Threshold = 1; 1050 } else if (isa<UnreachableInst>(++BasicBlock::iterator(Instr))) 1051 Threshold = 1; 1052 1053 // If this function uses the coldcc calling convention, prefer not to inline 1054 // it. 1055 if (F.getCallingConv() == CallingConv::Cold) 1056 Cost += InlineConstants::ColdccPenalty; 1057 1058 // Check if we're done. This can happen due to bonuses and penalties. 1059 if (Cost > Threshold) 1060 return false; 1061 1062 if (F.empty()) 1063 return true; 1064 1065 Function *Caller = CS.getInstruction()->getParent()->getParent(); 1066 // Check if the caller function is recursive itself. 1067 for (User *U : Caller->users()) { 1068 CallSite Site(U); 1069 if (!Site) 1070 continue; 1071 Instruction *I = Site.getInstruction(); 1072 if (I->getParent()->getParent() == Caller) { 1073 IsCallerRecursive = true; 1074 break; 1075 } 1076 } 1077 1078 // Populate our simplified values by mapping from function arguments to call 1079 // arguments with known important simplifications. 1080 CallSite::arg_iterator CAI = CS.arg_begin(); 1081 for (Function::arg_iterator FAI = F.arg_begin(), FAE = F.arg_end(); 1082 FAI != FAE; ++FAI, ++CAI) { 1083 assert(CAI != CS.arg_end()); 1084 if (Constant *C = dyn_cast<Constant>(CAI)) 1085 SimplifiedValues[FAI] = C; 1086 1087 Value *PtrArg = *CAI; 1088 if (ConstantInt *C = stripAndComputeInBoundsConstantOffsets(PtrArg)) { 1089 ConstantOffsetPtrs[FAI] = std::make_pair(PtrArg, C->getValue()); 1090 1091 // We can SROA any pointer arguments derived from alloca instructions. 1092 if (isa<AllocaInst>(PtrArg)) { 1093 SROAArgValues[FAI] = PtrArg; 1094 SROAArgCosts[PtrArg] = 0; 1095 } 1096 } 1097 } 1098 NumConstantArgs = SimplifiedValues.size(); 1099 NumConstantOffsetPtrArgs = ConstantOffsetPtrs.size(); 1100 NumAllocaArgs = SROAArgValues.size(); 1101 1102 // The worklist of live basic blocks in the callee *after* inlining. We avoid 1103 // adding basic blocks of the callee which can be proven to be dead for this 1104 // particular call site in order to get more accurate cost estimates. This 1105 // requires a somewhat heavyweight iteration pattern: we need to walk the 1106 // basic blocks in a breadth-first order as we insert live successors. To 1107 // accomplish this, prioritizing for small iterations because we exit after 1108 // crossing our threshold, we use a small-size optimized SetVector. 1109 typedef SetVector<BasicBlock *, SmallVector<BasicBlock *, 16>, 1110 SmallPtrSet<BasicBlock *, 16> > BBSetVector; 1111 BBSetVector BBWorklist; 1112 BBWorklist.insert(&F.getEntryBlock()); 1113 // Note that we *must not* cache the size, this loop grows the worklist. 1114 for (unsigned Idx = 0; Idx != BBWorklist.size(); ++Idx) { 1115 // Bail out the moment we cross the threshold. This means we'll under-count 1116 // the cost, but only when undercounting doesn't matter. 1117 if (Cost > (Threshold + VectorBonus)) 1118 break; 1119 1120 BasicBlock *BB = BBWorklist[Idx]; 1121 if (BB->empty()) 1122 continue; 1123 1124 // Analyze the cost of this block. If we blow through the threshold, this 1125 // returns false, and we can bail on out. 1126 if (!analyzeBlock(BB)) { 1127 if (IsRecursiveCall || ExposesReturnsTwice || HasDynamicAlloca || 1128 HasIndirectBr) 1129 return false; 1130 1131 // If the caller is a recursive function then we don't want to inline 1132 // functions which allocate a lot of stack space because it would increase 1133 // the caller stack usage dramatically. 1134 if (IsCallerRecursive && 1135 AllocatedSize > InlineConstants::TotalAllocaSizeRecursiveCaller) 1136 return false; 1137 1138 break; 1139 } 1140 1141 TerminatorInst *TI = BB->getTerminator(); 1142 1143 // Add in the live successors by first checking whether we have terminator 1144 // that may be simplified based on the values simplified by this call. 1145 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) { 1146 if (BI->isConditional()) { 1147 Value *Cond = BI->getCondition(); 1148 if (ConstantInt *SimpleCond 1149 = dyn_cast_or_null<ConstantInt>(SimplifiedValues.lookup(Cond))) { 1150 BBWorklist.insert(BI->getSuccessor(SimpleCond->isZero() ? 1 : 0)); 1151 continue; 1152 } 1153 } 1154 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) { 1155 Value *Cond = SI->getCondition(); 1156 if (ConstantInt *SimpleCond 1157 = dyn_cast_or_null<ConstantInt>(SimplifiedValues.lookup(Cond))) { 1158 BBWorklist.insert(SI->findCaseValue(SimpleCond).getCaseSuccessor()); 1159 continue; 1160 } 1161 } 1162 1163 // If we're unable to select a particular successor, just count all of 1164 // them. 1165 for (unsigned TIdx = 0, TSize = TI->getNumSuccessors(); TIdx != TSize; 1166 ++TIdx) 1167 BBWorklist.insert(TI->getSuccessor(TIdx)); 1168 1169 // If we had any successors at this point, than post-inlining is likely to 1170 // have them as well. Note that we assume any basic blocks which existed 1171 // due to branches or switches which folded above will also fold after 1172 // inlining. 1173 if (SingleBB && TI->getNumSuccessors() > 1) { 1174 // Take off the bonus we applied to the threshold. 1175 Threshold -= SingleBBBonus; 1176 SingleBB = false; 1177 } 1178 } 1179 1180 // If this is a noduplicate call, we can still inline as long as 1181 // inlining this would cause the removal of the caller (so the instruction 1182 // is not actually duplicated, just moved). 1183 if (!OnlyOneCallAndLocalLinkage && ContainsNoDuplicateCall) 1184 return false; 1185 1186 Threshold += VectorBonus; 1187 1188 return Cost < Threshold; 1189} 1190 1191#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) 1192/// \brief Dump stats about this call's analysis. 1193void CallAnalyzer::dump() { 1194#define DEBUG_PRINT_STAT(x) dbgs() << " " #x ": " << x << "\n" 1195 DEBUG_PRINT_STAT(NumConstantArgs); 1196 DEBUG_PRINT_STAT(NumConstantOffsetPtrArgs); 1197 DEBUG_PRINT_STAT(NumAllocaArgs); 1198 DEBUG_PRINT_STAT(NumConstantPtrCmps); 1199 DEBUG_PRINT_STAT(NumConstantPtrDiffs); 1200 DEBUG_PRINT_STAT(NumInstructionsSimplified); 1201 DEBUG_PRINT_STAT(SROACostSavings); 1202 DEBUG_PRINT_STAT(SROACostSavingsLost); 1203 DEBUG_PRINT_STAT(ContainsNoDuplicateCall); 1204 DEBUG_PRINT_STAT(Cost); 1205 DEBUG_PRINT_STAT(Threshold); 1206 DEBUG_PRINT_STAT(VectorBonus); 1207#undef DEBUG_PRINT_STAT 1208} 1209#endif 1210 1211INITIALIZE_PASS_BEGIN(InlineCostAnalysis, "inline-cost", "Inline Cost Analysis", 1212 true, true) 1213INITIALIZE_AG_DEPENDENCY(TargetTransformInfo) 1214INITIALIZE_PASS_END(InlineCostAnalysis, "inline-cost", "Inline Cost Analysis", 1215 true, true) 1216 1217char InlineCostAnalysis::ID = 0; 1218 1219InlineCostAnalysis::InlineCostAnalysis() : CallGraphSCCPass(ID) {} 1220 1221InlineCostAnalysis::~InlineCostAnalysis() {} 1222 1223void InlineCostAnalysis::getAnalysisUsage(AnalysisUsage &AU) const { 1224 AU.setPreservesAll(); 1225 AU.addRequired<TargetTransformInfo>(); 1226 CallGraphSCCPass::getAnalysisUsage(AU); 1227} 1228 1229bool InlineCostAnalysis::runOnSCC(CallGraphSCC &SCC) { 1230 TTI = &getAnalysis<TargetTransformInfo>(); 1231 return false; 1232} 1233 1234InlineCost InlineCostAnalysis::getInlineCost(CallSite CS, int Threshold) { 1235 return getInlineCost(CS, CS.getCalledFunction(), Threshold); 1236} 1237 1238/// \brief Test that two functions either have or have not the given attribute 1239/// at the same time. 1240static bool attributeMatches(Function *F1, Function *F2, 1241 Attribute::AttrKind Attr) { 1242 return F1->hasFnAttribute(Attr) == F2->hasFnAttribute(Attr); 1243} 1244 1245/// \brief Test that there are no attribute conflicts between Caller and Callee 1246/// that prevent inlining. 1247static bool functionsHaveCompatibleAttributes(Function *Caller, 1248 Function *Callee) { 1249 return attributeMatches(Caller, Callee, Attribute::SanitizeAddress) && 1250 attributeMatches(Caller, Callee, Attribute::SanitizeMemory) && 1251 attributeMatches(Caller, Callee, Attribute::SanitizeThread); 1252} 1253 1254InlineCost InlineCostAnalysis::getInlineCost(CallSite CS, Function *Callee, 1255 int Threshold) { 1256 // Cannot inline indirect calls. 1257 if (!Callee) 1258 return llvm::InlineCost::getNever(); 1259 1260 // Calls to functions with always-inline attributes should be inlined 1261 // whenever possible. 1262 if (CS.hasFnAttr(Attribute::AlwaysInline)) { 1263 if (isInlineViable(*Callee)) 1264 return llvm::InlineCost::getAlways(); 1265 return llvm::InlineCost::getNever(); 1266 } 1267 1268 // Never inline functions with conflicting attributes (unless callee has 1269 // always-inline attribute). 1270 if (!functionsHaveCompatibleAttributes(CS.getCaller(), Callee)) 1271 return llvm::InlineCost::getNever(); 1272 1273 // Don't inline this call if the caller has the optnone attribute. 1274 if (CS.getCaller()->hasFnAttribute(Attribute::OptimizeNone)) 1275 return llvm::InlineCost::getNever(); 1276 1277 // Don't inline functions which can be redefined at link-time to mean 1278 // something else. Don't inline functions marked noinline or call sites 1279 // marked noinline. 1280 if (Callee->mayBeOverridden() || 1281 Callee->hasFnAttribute(Attribute::NoInline) || CS.isNoInline()) 1282 return llvm::InlineCost::getNever(); 1283 1284 DEBUG(llvm::dbgs() << " Analyzing call of " << Callee->getName() 1285 << "...\n"); 1286 1287 CallAnalyzer CA(Callee->getDataLayout(), *TTI, *Callee, Threshold); 1288 bool ShouldInline = CA.analyzeCall(CS); 1289 1290 DEBUG(CA.dump()); 1291 1292 // Check if there was a reason to force inlining or no inlining. 1293 if (!ShouldInline && CA.getCost() < CA.getThreshold()) 1294 return InlineCost::getNever(); 1295 if (ShouldInline && CA.getCost() >= CA.getThreshold()) 1296 return InlineCost::getAlways(); 1297 1298 return llvm::InlineCost::get(CA.getCost(), CA.getThreshold()); 1299} 1300 1301bool InlineCostAnalysis::isInlineViable(Function &F) { 1302 bool ReturnsTwice = 1303 F.getAttributes().hasAttribute(AttributeSet::FunctionIndex, 1304 Attribute::ReturnsTwice); 1305 for (Function::iterator BI = F.begin(), BE = F.end(); BI != BE; ++BI) { 1306 // Disallow inlining of functions which contain an indirect branch. 1307 if (isa<IndirectBrInst>(BI->getTerminator())) 1308 return false; 1309 1310 for (BasicBlock::iterator II = BI->begin(), IE = BI->end(); II != IE; 1311 ++II) { 1312 CallSite CS(II); 1313 if (!CS) 1314 continue; 1315 1316 // Disallow recursive calls. 1317 if (&F == CS.getCalledFunction()) 1318 return false; 1319 1320 // Disallow calls which expose returns-twice to a function not previously 1321 // attributed as such. 1322 if (!ReturnsTwice && CS.isCall() && 1323 cast<CallInst>(CS.getInstruction())->canReturnTwice()) 1324 return false; 1325 } 1326 } 1327 1328 return true; 1329} 1330