InlineCost.cpp revision 263508
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#define DEBUG_TYPE "inline-cost" 15#include "llvm/Analysis/InlineCost.h" 16#include "llvm/ADT/STLExtras.h" 17#include "llvm/ADT/SetVector.h" 18#include "llvm/ADT/SmallPtrSet.h" 19#include "llvm/ADT/SmallVector.h" 20#include "llvm/ADT/Statistic.h" 21#include "llvm/Analysis/ConstantFolding.h" 22#include "llvm/Analysis/InstructionSimplify.h" 23#include "llvm/Analysis/TargetTransformInfo.h" 24#include "llvm/IR/CallingConv.h" 25#include "llvm/IR/DataLayout.h" 26#include "llvm/IR/GlobalAlias.h" 27#include "llvm/IR/IntrinsicInst.h" 28#include "llvm/IR/Operator.h" 29#include "llvm/InstVisitor.h" 30#include "llvm/Support/CallSite.h" 31#include "llvm/Support/Debug.h" 32#include "llvm/Support/GetElementPtrTypeIterator.h" 33#include "llvm/Support/raw_ostream.h" 34 35using namespace llvm; 36 37STATISTIC(NumCallsAnalyzed, "Number of call sites analyzed"); 38 39namespace { 40 41class CallAnalyzer : public InstVisitor<CallAnalyzer, bool> { 42 typedef InstVisitor<CallAnalyzer, bool> Base; 43 friend class InstVisitor<CallAnalyzer, bool>; 44 45 // DataLayout if available, or null. 46 const DataLayout *const TD; 47 48 /// The TargetTransformInfo available for this compilation. 49 const TargetTransformInfo &TTI; 50 51 // The called function. 52 Function &F; 53 54 int Threshold; 55 int Cost; 56 57 bool IsCallerRecursive; 58 bool IsRecursiveCall; 59 bool ExposesReturnsTwice; 60 bool HasDynamicAlloca; 61 bool ContainsNoDuplicateCall; 62 bool HasReturn; 63 bool HasIndirectBr; 64 65 /// Number of bytes allocated statically by the callee. 66 uint64_t AllocatedSize; 67 unsigned NumInstructions, NumVectorInstructions; 68 int FiftyPercentVectorBonus, TenPercentVectorBonus; 69 int VectorBonus; 70 71 // While we walk the potentially-inlined instructions, we build up and 72 // maintain a mapping of simplified values specific to this callsite. The 73 // idea is to propagate any special information we have about arguments to 74 // this call through the inlinable section of the function, and account for 75 // likely simplifications post-inlining. The most important aspect we track 76 // is CFG altering simplifications -- when we prove a basic block dead, that 77 // can cause dramatic shifts in the cost of inlining a function. 78 DenseMap<Value *, Constant *> SimplifiedValues; 79 80 // Keep track of the values which map back (through function arguments) to 81 // allocas on the caller stack which could be simplified through SROA. 82 DenseMap<Value *, Value *> SROAArgValues; 83 84 // The mapping of caller Alloca values to their accumulated cost savings. If 85 // we have to disable SROA for one of the allocas, this tells us how much 86 // cost must be added. 87 DenseMap<Value *, int> SROAArgCosts; 88 89 // Keep track of values which map to a pointer base and constant offset. 90 DenseMap<Value *, std::pair<Value *, APInt> > ConstantOffsetPtrs; 91 92 // Custom simplification helper routines. 93 bool isAllocaDerivedArg(Value *V); 94 bool lookupSROAArgAndCost(Value *V, Value *&Arg, 95 DenseMap<Value *, int>::iterator &CostIt); 96 void disableSROA(DenseMap<Value *, int>::iterator CostIt); 97 void disableSROA(Value *V); 98 void accumulateSROACost(DenseMap<Value *, int>::iterator CostIt, 99 int InstructionCost); 100 bool handleSROACandidate(bool IsSROAValid, 101 DenseMap<Value *, int>::iterator CostIt, 102 int InstructionCost); 103 bool isGEPOffsetConstant(GetElementPtrInst &GEP); 104 bool accumulateGEPOffset(GEPOperator &GEP, APInt &Offset); 105 bool simplifyCallSite(Function *F, CallSite CS); 106 ConstantInt *stripAndComputeInBoundsConstantOffsets(Value *&V); 107 108 // Custom analysis routines. 109 bool analyzeBlock(BasicBlock *BB); 110 111 // Disable several entry points to the visitor so we don't accidentally use 112 // them by declaring but not defining them here. 113 void visit(Module *); void visit(Module &); 114 void visit(Function *); void visit(Function &); 115 void visit(BasicBlock *); void visit(BasicBlock &); 116 117 // Provide base case for our instruction visit. 118 bool visitInstruction(Instruction &I); 119 120 // Our visit overrides. 121 bool visitAlloca(AllocaInst &I); 122 bool visitPHI(PHINode &I); 123 bool visitGetElementPtr(GetElementPtrInst &I); 124 bool visitBitCast(BitCastInst &I); 125 bool visitPtrToInt(PtrToIntInst &I); 126 bool visitIntToPtr(IntToPtrInst &I); 127 bool visitCastInst(CastInst &I); 128 bool visitUnaryInstruction(UnaryInstruction &I); 129 bool visitCmpInst(CmpInst &I); 130 bool visitSub(BinaryOperator &I); 131 bool visitBinaryOperator(BinaryOperator &I); 132 bool visitLoad(LoadInst &I); 133 bool visitStore(StoreInst &I); 134 bool visitExtractValue(ExtractValueInst &I); 135 bool visitInsertValue(InsertValueInst &I); 136 bool visitCallSite(CallSite CS); 137 bool visitReturnInst(ReturnInst &RI); 138 bool visitBranchInst(BranchInst &BI); 139 bool visitSwitchInst(SwitchInst &SI); 140 bool visitIndirectBrInst(IndirectBrInst &IBI); 141 bool visitResumeInst(ResumeInst &RI); 142 bool visitUnreachableInst(UnreachableInst &I); 143 144public: 145 CallAnalyzer(const DataLayout *TD, const TargetTransformInfo &TTI, 146 Function &Callee, int Threshold) 147 : TD(TD), TTI(TTI), F(Callee), Threshold(Threshold), Cost(0), 148 IsCallerRecursive(false), IsRecursiveCall(false), 149 ExposesReturnsTwice(false), HasDynamicAlloca(false), 150 ContainsNoDuplicateCall(false), HasReturn(false), HasIndirectBr(false), 151 AllocatedSize(0), NumInstructions(0), NumVectorInstructions(0), 152 FiftyPercentVectorBonus(0), TenPercentVectorBonus(0), VectorBonus(0), 153 NumConstantArgs(0), NumConstantOffsetPtrArgs(0), NumAllocaArgs(0), 154 NumConstantPtrCmps(0), NumConstantPtrDiffs(0), 155 NumInstructionsSimplified(0), SROACostSavings(0), 156 SROACostSavingsLost(0) {} 157 158 bool analyzeCall(CallSite CS); 159 160 int getThreshold() { return Threshold; } 161 int getCost() { return Cost; } 162 163 // Keep a bunch of stats about the cost savings found so we can print them 164 // out when debugging. 165 unsigned NumConstantArgs; 166 unsigned NumConstantOffsetPtrArgs; 167 unsigned NumAllocaArgs; 168 unsigned NumConstantPtrCmps; 169 unsigned NumConstantPtrDiffs; 170 unsigned NumInstructionsSimplified; 171 unsigned SROACostSavings; 172 unsigned SROACostSavingsLost; 173 174 void dump(); 175}; 176 177} // namespace 178 179/// \brief Test whether the given value is an Alloca-derived function argument. 180bool CallAnalyzer::isAllocaDerivedArg(Value *V) { 181 return SROAArgValues.count(V); 182} 183 184/// \brief Lookup the SROA-candidate argument and cost iterator which V maps to. 185/// Returns false if V does not map to a SROA-candidate. 186bool CallAnalyzer::lookupSROAArgAndCost( 187 Value *V, Value *&Arg, DenseMap<Value *, int>::iterator &CostIt) { 188 if (SROAArgValues.empty() || SROAArgCosts.empty()) 189 return false; 190 191 DenseMap<Value *, Value *>::iterator ArgIt = SROAArgValues.find(V); 192 if (ArgIt == SROAArgValues.end()) 193 return false; 194 195 Arg = ArgIt->second; 196 CostIt = SROAArgCosts.find(Arg); 197 return CostIt != SROAArgCosts.end(); 198} 199 200/// \brief Disable SROA for the candidate marked by this cost iterator. 201/// 202/// This marks the candidate as no longer viable for SROA, and adds the cost 203/// savings associated with it back into the inline cost measurement. 204void CallAnalyzer::disableSROA(DenseMap<Value *, int>::iterator CostIt) { 205 // If we're no longer able to perform SROA we need to undo its cost savings 206 // and prevent subsequent analysis. 207 Cost += CostIt->second; 208 SROACostSavings -= CostIt->second; 209 SROACostSavingsLost += CostIt->second; 210 SROAArgCosts.erase(CostIt); 211} 212 213/// \brief If 'V' maps to a SROA candidate, disable SROA for it. 214void CallAnalyzer::disableSROA(Value *V) { 215 Value *SROAArg; 216 DenseMap<Value *, int>::iterator CostIt; 217 if (lookupSROAArgAndCost(V, SROAArg, CostIt)) 218 disableSROA(CostIt); 219} 220 221/// \brief Accumulate the given cost for a particular SROA candidate. 222void CallAnalyzer::accumulateSROACost(DenseMap<Value *, int>::iterator CostIt, 223 int InstructionCost) { 224 CostIt->second += InstructionCost; 225 SROACostSavings += InstructionCost; 226} 227 228/// \brief Helper for the common pattern of handling a SROA candidate. 229/// Either accumulates the cost savings if the SROA remains valid, or disables 230/// SROA for the candidate. 231bool CallAnalyzer::handleSROACandidate(bool IsSROAValid, 232 DenseMap<Value *, int>::iterator CostIt, 233 int InstructionCost) { 234 if (IsSROAValid) { 235 accumulateSROACost(CostIt, InstructionCost); 236 return true; 237 } 238 239 disableSROA(CostIt); 240 return false; 241} 242 243/// \brief Check whether a GEP's indices are all constant. 244/// 245/// Respects any simplified values known during the analysis of this callsite. 246bool CallAnalyzer::isGEPOffsetConstant(GetElementPtrInst &GEP) { 247 for (User::op_iterator I = GEP.idx_begin(), E = GEP.idx_end(); I != E; ++I) 248 if (!isa<Constant>(*I) && !SimplifiedValues.lookup(*I)) 249 return false; 250 251 return true; 252} 253 254/// \brief Accumulate a constant GEP offset into an APInt if possible. 255/// 256/// Returns false if unable to compute the offset for any reason. Respects any 257/// simplified values known during the analysis of this callsite. 258bool CallAnalyzer::accumulateGEPOffset(GEPOperator &GEP, APInt &Offset) { 259 if (!TD) 260 return false; 261 262 unsigned IntPtrWidth = TD->getPointerSizeInBits(); 263 assert(IntPtrWidth == Offset.getBitWidth()); 264 265 for (gep_type_iterator GTI = gep_type_begin(GEP), GTE = gep_type_end(GEP); 266 GTI != GTE; ++GTI) { 267 ConstantInt *OpC = dyn_cast<ConstantInt>(GTI.getOperand()); 268 if (!OpC) 269 if (Constant *SimpleOp = SimplifiedValues.lookup(GTI.getOperand())) 270 OpC = dyn_cast<ConstantInt>(SimpleOp); 271 if (!OpC) 272 return false; 273 if (OpC->isZero()) continue; 274 275 // Handle a struct index, which adds its field offset to the pointer. 276 if (StructType *STy = dyn_cast<StructType>(*GTI)) { 277 unsigned ElementIdx = OpC->getZExtValue(); 278 const StructLayout *SL = TD->getStructLayout(STy); 279 Offset += APInt(IntPtrWidth, SL->getElementOffset(ElementIdx)); 280 continue; 281 } 282 283 APInt TypeSize(IntPtrWidth, TD->getTypeAllocSize(GTI.getIndexedType())); 284 Offset += OpC->getValue().sextOrTrunc(IntPtrWidth) * TypeSize; 285 } 286 return true; 287} 288 289bool CallAnalyzer::visitAlloca(AllocaInst &I) { 290 // FIXME: Check whether inlining will turn a dynamic alloca into a static 291 // alloca, and handle that case. 292 293 // Accumulate the allocated size. 294 if (I.isStaticAlloca()) { 295 Type *Ty = I.getAllocatedType(); 296 AllocatedSize += (TD ? TD->getTypeAllocSize(Ty) : 297 Ty->getPrimitiveSizeInBits()); 298 } 299 300 // We will happily inline static alloca instructions. 301 if (I.isStaticAlloca()) 302 return Base::visitAlloca(I); 303 304 // FIXME: This is overly conservative. Dynamic allocas are inefficient for 305 // a variety of reasons, and so we would like to not inline them into 306 // functions which don't currently have a dynamic alloca. This simply 307 // disables inlining altogether in the presence of a dynamic alloca. 308 HasDynamicAlloca = true; 309 return false; 310} 311 312bool CallAnalyzer::visitPHI(PHINode &I) { 313 // FIXME: We should potentially be tracking values through phi nodes, 314 // especially when they collapse to a single value due to deleted CFG edges 315 // during inlining. 316 317 // FIXME: We need to propagate SROA *disabling* through phi nodes, even 318 // though we don't want to propagate it's bonuses. The idea is to disable 319 // SROA if it *might* be used in an inappropriate manner. 320 321 // Phi nodes are always zero-cost. 322 return true; 323} 324 325bool CallAnalyzer::visitGetElementPtr(GetElementPtrInst &I) { 326 Value *SROAArg; 327 DenseMap<Value *, int>::iterator CostIt; 328 bool SROACandidate = lookupSROAArgAndCost(I.getPointerOperand(), 329 SROAArg, CostIt); 330 331 // Try to fold GEPs of constant-offset call site argument pointers. This 332 // requires target data and inbounds GEPs. 333 if (TD && I.isInBounds()) { 334 // Check if we have a base + offset for the pointer. 335 Value *Ptr = I.getPointerOperand(); 336 std::pair<Value *, APInt> BaseAndOffset = ConstantOffsetPtrs.lookup(Ptr); 337 if (BaseAndOffset.first) { 338 // Check if the offset of this GEP is constant, and if so accumulate it 339 // into Offset. 340 if (!accumulateGEPOffset(cast<GEPOperator>(I), BaseAndOffset.second)) { 341 // Non-constant GEPs aren't folded, and disable SROA. 342 if (SROACandidate) 343 disableSROA(CostIt); 344 return false; 345 } 346 347 // Add the result as a new mapping to Base + Offset. 348 ConstantOffsetPtrs[&I] = BaseAndOffset; 349 350 // Also handle SROA candidates here, we already know that the GEP is 351 // all-constant indexed. 352 if (SROACandidate) 353 SROAArgValues[&I] = SROAArg; 354 355 return true; 356 } 357 } 358 359 if (isGEPOffsetConstant(I)) { 360 if (SROACandidate) 361 SROAArgValues[&I] = SROAArg; 362 363 // Constant GEPs are modeled as free. 364 return true; 365 } 366 367 // Variable GEPs will require math and will disable SROA. 368 if (SROACandidate) 369 disableSROA(CostIt); 370 return false; 371} 372 373bool CallAnalyzer::visitBitCast(BitCastInst &I) { 374 // Propagate constants through bitcasts. 375 Constant *COp = dyn_cast<Constant>(I.getOperand(0)); 376 if (!COp) 377 COp = SimplifiedValues.lookup(I.getOperand(0)); 378 if (COp) 379 if (Constant *C = ConstantExpr::getBitCast(COp, I.getType())) { 380 SimplifiedValues[&I] = C; 381 return true; 382 } 383 384 // Track base/offsets through casts 385 std::pair<Value *, APInt> BaseAndOffset 386 = ConstantOffsetPtrs.lookup(I.getOperand(0)); 387 // Casts don't change the offset, just wrap it up. 388 if (BaseAndOffset.first) 389 ConstantOffsetPtrs[&I] = BaseAndOffset; 390 391 // Also look for SROA candidates here. 392 Value *SROAArg; 393 DenseMap<Value *, int>::iterator CostIt; 394 if (lookupSROAArgAndCost(I.getOperand(0), SROAArg, CostIt)) 395 SROAArgValues[&I] = SROAArg; 396 397 // Bitcasts are always zero cost. 398 return true; 399} 400 401bool CallAnalyzer::visitPtrToInt(PtrToIntInst &I) { 402 // Propagate constants through ptrtoint. 403 Constant *COp = dyn_cast<Constant>(I.getOperand(0)); 404 if (!COp) 405 COp = SimplifiedValues.lookup(I.getOperand(0)); 406 if (COp) 407 if (Constant *C = ConstantExpr::getPtrToInt(COp, I.getType())) { 408 SimplifiedValues[&I] = C; 409 return true; 410 } 411 412 // Track base/offset pairs when converted to a plain integer provided the 413 // integer is large enough to represent the pointer. 414 unsigned IntegerSize = I.getType()->getScalarSizeInBits(); 415 if (TD && IntegerSize >= TD->getPointerSizeInBits()) { 416 std::pair<Value *, APInt> BaseAndOffset 417 = ConstantOffsetPtrs.lookup(I.getOperand(0)); 418 if (BaseAndOffset.first) 419 ConstantOffsetPtrs[&I] = BaseAndOffset; 420 } 421 422 // This is really weird. Technically, ptrtoint will disable SROA. However, 423 // unless that ptrtoint is *used* somewhere in the live basic blocks after 424 // inlining, it will be nuked, and SROA should proceed. All of the uses which 425 // would block SROA would also block SROA if applied directly to a pointer, 426 // and so we can just add the integer in here. The only places where SROA is 427 // preserved either cannot fire on an integer, or won't in-and-of themselves 428 // disable SROA (ext) w/o some later use that we would see and disable. 429 Value *SROAArg; 430 DenseMap<Value *, int>::iterator CostIt; 431 if (lookupSROAArgAndCost(I.getOperand(0), SROAArg, CostIt)) 432 SROAArgValues[&I] = SROAArg; 433 434 return TargetTransformInfo::TCC_Free == TTI.getUserCost(&I); 435} 436 437bool CallAnalyzer::visitIntToPtr(IntToPtrInst &I) { 438 // Propagate constants through ptrtoint. 439 Constant *COp = dyn_cast<Constant>(I.getOperand(0)); 440 if (!COp) 441 COp = SimplifiedValues.lookup(I.getOperand(0)); 442 if (COp) 443 if (Constant *C = ConstantExpr::getIntToPtr(COp, I.getType())) { 444 SimplifiedValues[&I] = C; 445 return true; 446 } 447 448 // Track base/offset pairs when round-tripped through a pointer without 449 // modifications provided the integer is not too large. 450 Value *Op = I.getOperand(0); 451 unsigned IntegerSize = Op->getType()->getScalarSizeInBits(); 452 if (TD && IntegerSize <= TD->getPointerSizeInBits()) { 453 std::pair<Value *, APInt> BaseAndOffset = ConstantOffsetPtrs.lookup(Op); 454 if (BaseAndOffset.first) 455 ConstantOffsetPtrs[&I] = BaseAndOffset; 456 } 457 458 // "Propagate" SROA here in the same manner as we do for ptrtoint above. 459 Value *SROAArg; 460 DenseMap<Value *, int>::iterator CostIt; 461 if (lookupSROAArgAndCost(Op, SROAArg, CostIt)) 462 SROAArgValues[&I] = SROAArg; 463 464 return TargetTransformInfo::TCC_Free == TTI.getUserCost(&I); 465} 466 467bool CallAnalyzer::visitCastInst(CastInst &I) { 468 // Propagate constants through ptrtoint. 469 Constant *COp = dyn_cast<Constant>(I.getOperand(0)); 470 if (!COp) 471 COp = SimplifiedValues.lookup(I.getOperand(0)); 472 if (COp) 473 if (Constant *C = ConstantExpr::getCast(I.getOpcode(), COp, I.getType())) { 474 SimplifiedValues[&I] = C; 475 return true; 476 } 477 478 // Disable SROA in the face of arbitrary casts we don't whitelist elsewhere. 479 disableSROA(I.getOperand(0)); 480 481 return TargetTransformInfo::TCC_Free == TTI.getUserCost(&I); 482} 483 484bool CallAnalyzer::visitUnaryInstruction(UnaryInstruction &I) { 485 Value *Operand = I.getOperand(0); 486 Constant *COp = dyn_cast<Constant>(Operand); 487 if (!COp) 488 COp = SimplifiedValues.lookup(Operand); 489 if (COp) 490 if (Constant *C = ConstantFoldInstOperands(I.getOpcode(), I.getType(), 491 COp, TD)) { 492 SimplifiedValues[&I] = C; 493 return true; 494 } 495 496 // Disable any SROA on the argument to arbitrary unary operators. 497 disableSROA(Operand); 498 499 return false; 500} 501 502bool CallAnalyzer::visitCmpInst(CmpInst &I) { 503 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1); 504 // First try to handle simplified comparisons. 505 if (!isa<Constant>(LHS)) 506 if (Constant *SimpleLHS = SimplifiedValues.lookup(LHS)) 507 LHS = SimpleLHS; 508 if (!isa<Constant>(RHS)) 509 if (Constant *SimpleRHS = SimplifiedValues.lookup(RHS)) 510 RHS = SimpleRHS; 511 if (Constant *CLHS = dyn_cast<Constant>(LHS)) { 512 if (Constant *CRHS = dyn_cast<Constant>(RHS)) 513 if (Constant *C = ConstantExpr::getCompare(I.getPredicate(), CLHS, CRHS)) { 514 SimplifiedValues[&I] = C; 515 return true; 516 } 517 } 518 519 if (I.getOpcode() == Instruction::FCmp) 520 return false; 521 522 // Otherwise look for a comparison between constant offset pointers with 523 // a common base. 524 Value *LHSBase, *RHSBase; 525 APInt LHSOffset, RHSOffset; 526 llvm::tie(LHSBase, LHSOffset) = ConstantOffsetPtrs.lookup(LHS); 527 if (LHSBase) { 528 llvm::tie(RHSBase, RHSOffset) = ConstantOffsetPtrs.lookup(RHS); 529 if (RHSBase && LHSBase == RHSBase) { 530 // We have common bases, fold the icmp to a constant based on the 531 // offsets. 532 Constant *CLHS = ConstantInt::get(LHS->getContext(), LHSOffset); 533 Constant *CRHS = ConstantInt::get(RHS->getContext(), RHSOffset); 534 if (Constant *C = ConstantExpr::getICmp(I.getPredicate(), CLHS, CRHS)) { 535 SimplifiedValues[&I] = C; 536 ++NumConstantPtrCmps; 537 return true; 538 } 539 } 540 } 541 542 // If the comparison is an equality comparison with null, we can simplify it 543 // for any alloca-derived argument. 544 if (I.isEquality() && isa<ConstantPointerNull>(I.getOperand(1))) 545 if (isAllocaDerivedArg(I.getOperand(0))) { 546 // We can actually predict the result of comparisons between an 547 // alloca-derived value and null. Note that this fires regardless of 548 // SROA firing. 549 bool IsNotEqual = I.getPredicate() == CmpInst::ICMP_NE; 550 SimplifiedValues[&I] = IsNotEqual ? ConstantInt::getTrue(I.getType()) 551 : ConstantInt::getFalse(I.getType()); 552 return true; 553 } 554 555 // Finally check for SROA candidates in comparisons. 556 Value *SROAArg; 557 DenseMap<Value *, int>::iterator CostIt; 558 if (lookupSROAArgAndCost(I.getOperand(0), SROAArg, CostIt)) { 559 if (isa<ConstantPointerNull>(I.getOperand(1))) { 560 accumulateSROACost(CostIt, InlineConstants::InstrCost); 561 return true; 562 } 563 564 disableSROA(CostIt); 565 } 566 567 return false; 568} 569 570bool CallAnalyzer::visitSub(BinaryOperator &I) { 571 // Try to handle a special case: we can fold computing the difference of two 572 // constant-related pointers. 573 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1); 574 Value *LHSBase, *RHSBase; 575 APInt LHSOffset, RHSOffset; 576 llvm::tie(LHSBase, LHSOffset) = ConstantOffsetPtrs.lookup(LHS); 577 if (LHSBase) { 578 llvm::tie(RHSBase, RHSOffset) = ConstantOffsetPtrs.lookup(RHS); 579 if (RHSBase && LHSBase == RHSBase) { 580 // We have common bases, fold the subtract to a constant based on the 581 // offsets. 582 Constant *CLHS = ConstantInt::get(LHS->getContext(), LHSOffset); 583 Constant *CRHS = ConstantInt::get(RHS->getContext(), RHSOffset); 584 if (Constant *C = ConstantExpr::getSub(CLHS, CRHS)) { 585 SimplifiedValues[&I] = C; 586 ++NumConstantPtrDiffs; 587 return true; 588 } 589 } 590 } 591 592 // Otherwise, fall back to the generic logic for simplifying and handling 593 // instructions. 594 return Base::visitSub(I); 595} 596 597bool CallAnalyzer::visitBinaryOperator(BinaryOperator &I) { 598 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1); 599 if (!isa<Constant>(LHS)) 600 if (Constant *SimpleLHS = SimplifiedValues.lookup(LHS)) 601 LHS = SimpleLHS; 602 if (!isa<Constant>(RHS)) 603 if (Constant *SimpleRHS = SimplifiedValues.lookup(RHS)) 604 RHS = SimpleRHS; 605 Value *SimpleV = SimplifyBinOp(I.getOpcode(), LHS, RHS, TD); 606 if (Constant *C = dyn_cast_or_null<Constant>(SimpleV)) { 607 SimplifiedValues[&I] = C; 608 return true; 609 } 610 611 // Disable any SROA on arguments to arbitrary, unsimplified binary operators. 612 disableSROA(LHS); 613 disableSROA(RHS); 614 615 return false; 616} 617 618bool CallAnalyzer::visitLoad(LoadInst &I) { 619 Value *SROAArg; 620 DenseMap<Value *, int>::iterator CostIt; 621 if (lookupSROAArgAndCost(I.getOperand(0), SROAArg, CostIt)) { 622 if (I.isSimple()) { 623 accumulateSROACost(CostIt, InlineConstants::InstrCost); 624 return true; 625 } 626 627 disableSROA(CostIt); 628 } 629 630 return false; 631} 632 633bool CallAnalyzer::visitStore(StoreInst &I) { 634 Value *SROAArg; 635 DenseMap<Value *, int>::iterator CostIt; 636 if (lookupSROAArgAndCost(I.getOperand(0), SROAArg, CostIt)) { 637 if (I.isSimple()) { 638 accumulateSROACost(CostIt, InlineConstants::InstrCost); 639 return true; 640 } 641 642 disableSROA(CostIt); 643 } 644 645 return false; 646} 647 648bool CallAnalyzer::visitExtractValue(ExtractValueInst &I) { 649 // Constant folding for extract value is trivial. 650 Constant *C = dyn_cast<Constant>(I.getAggregateOperand()); 651 if (!C) 652 C = SimplifiedValues.lookup(I.getAggregateOperand()); 653 if (C) { 654 SimplifiedValues[&I] = ConstantExpr::getExtractValue(C, I.getIndices()); 655 return true; 656 } 657 658 // SROA can look through these but give them a cost. 659 return false; 660} 661 662bool CallAnalyzer::visitInsertValue(InsertValueInst &I) { 663 // Constant folding for insert value is trivial. 664 Constant *AggC = dyn_cast<Constant>(I.getAggregateOperand()); 665 if (!AggC) 666 AggC = SimplifiedValues.lookup(I.getAggregateOperand()); 667 Constant *InsertedC = dyn_cast<Constant>(I.getInsertedValueOperand()); 668 if (!InsertedC) 669 InsertedC = SimplifiedValues.lookup(I.getInsertedValueOperand()); 670 if (AggC && InsertedC) { 671 SimplifiedValues[&I] = ConstantExpr::getInsertValue(AggC, InsertedC, 672 I.getIndices()); 673 return true; 674 } 675 676 // SROA can look through these but give them a cost. 677 return false; 678} 679 680/// \brief Try to simplify a call site. 681/// 682/// Takes a concrete function and callsite and tries to actually simplify it by 683/// analyzing the arguments and call itself with instsimplify. Returns true if 684/// it has simplified the callsite to some other entity (a constant), making it 685/// free. 686bool CallAnalyzer::simplifyCallSite(Function *F, CallSite CS) { 687 // FIXME: Using the instsimplify logic directly for this is inefficient 688 // because we have to continually rebuild the argument list even when no 689 // simplifications can be performed. Until that is fixed with remapping 690 // inside of instsimplify, directly constant fold calls here. 691 if (!canConstantFoldCallTo(F)) 692 return false; 693 694 // Try to re-map the arguments to constants. 695 SmallVector<Constant *, 4> ConstantArgs; 696 ConstantArgs.reserve(CS.arg_size()); 697 for (CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end(); 698 I != E; ++I) { 699 Constant *C = dyn_cast<Constant>(*I); 700 if (!C) 701 C = dyn_cast_or_null<Constant>(SimplifiedValues.lookup(*I)); 702 if (!C) 703 return false; // This argument doesn't map to a constant. 704 705 ConstantArgs.push_back(C); 706 } 707 if (Constant *C = ConstantFoldCall(F, ConstantArgs)) { 708 SimplifiedValues[CS.getInstruction()] = C; 709 return true; 710 } 711 712 return false; 713} 714 715bool CallAnalyzer::visitCallSite(CallSite CS) { 716 if (CS.hasFnAttr(Attribute::ReturnsTwice) && 717 !F.getAttributes().hasAttribute(AttributeSet::FunctionIndex, 718 Attribute::ReturnsTwice)) { 719 // This aborts the entire analysis. 720 ExposesReturnsTwice = true; 721 return false; 722 } 723 if (CS.isCall() && 724 cast<CallInst>(CS.getInstruction())->hasFnAttr(Attribute::NoDuplicate)) 725 ContainsNoDuplicateCall = true; 726 727 if (Function *F = CS.getCalledFunction()) { 728 // When we have a concrete function, first try to simplify it directly. 729 if (simplifyCallSite(F, CS)) 730 return true; 731 732 // Next check if it is an intrinsic we know about. 733 // FIXME: Lift this into part of the InstVisitor. 734 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(CS.getInstruction())) { 735 switch (II->getIntrinsicID()) { 736 default: 737 return Base::visitCallSite(CS); 738 739 case Intrinsic::memset: 740 case Intrinsic::memcpy: 741 case Intrinsic::memmove: 742 // SROA can usually chew through these intrinsics, but they aren't free. 743 return false; 744 } 745 } 746 747 if (F == CS.getInstruction()->getParent()->getParent()) { 748 // This flag will fully abort the analysis, so don't bother with anything 749 // else. 750 IsRecursiveCall = true; 751 return false; 752 } 753 754 if (TTI.isLoweredToCall(F)) { 755 // We account for the average 1 instruction per call argument setup 756 // here. 757 Cost += CS.arg_size() * InlineConstants::InstrCost; 758 759 // Everything other than inline ASM will also have a significant cost 760 // merely from making the call. 761 if (!isa<InlineAsm>(CS.getCalledValue())) 762 Cost += InlineConstants::CallPenalty; 763 } 764 765 return Base::visitCallSite(CS); 766 } 767 768 // Otherwise we're in a very special case -- an indirect function call. See 769 // if we can be particularly clever about this. 770 Value *Callee = CS.getCalledValue(); 771 772 // First, pay the price of the argument setup. We account for the average 773 // 1 instruction per call argument setup here. 774 Cost += CS.arg_size() * InlineConstants::InstrCost; 775 776 // Next, check if this happens to be an indirect function call to a known 777 // function in this inline context. If not, we've done all we can. 778 Function *F = dyn_cast_or_null<Function>(SimplifiedValues.lookup(Callee)); 779 if (!F) 780 return Base::visitCallSite(CS); 781 782 // If we have a constant that we are calling as a function, we can peer 783 // through it and see the function target. This happens not infrequently 784 // during devirtualization and so we want to give it a hefty bonus for 785 // inlining, but cap that bonus in the event that inlining wouldn't pan 786 // out. Pretend to inline the function, with a custom threshold. 787 CallAnalyzer CA(TD, TTI, *F, InlineConstants::IndirectCallThreshold); 788 if (CA.analyzeCall(CS)) { 789 // We were able to inline the indirect call! Subtract the cost from the 790 // bonus we want to apply, but don't go below zero. 791 Cost -= std::max(0, InlineConstants::IndirectCallThreshold - CA.getCost()); 792 } 793 794 return Base::visitCallSite(CS); 795} 796 797bool CallAnalyzer::visitReturnInst(ReturnInst &RI) { 798 // At least one return instruction will be free after inlining. 799 bool Free = !HasReturn; 800 HasReturn = true; 801 return Free; 802} 803 804bool CallAnalyzer::visitBranchInst(BranchInst &BI) { 805 // We model unconditional branches as essentially free -- they really 806 // shouldn't exist at all, but handling them makes the behavior of the 807 // inliner more regular and predictable. Interestingly, conditional branches 808 // which will fold away are also free. 809 return BI.isUnconditional() || isa<ConstantInt>(BI.getCondition()) || 810 dyn_cast_or_null<ConstantInt>( 811 SimplifiedValues.lookup(BI.getCondition())); 812} 813 814bool CallAnalyzer::visitSwitchInst(SwitchInst &SI) { 815 // We model unconditional switches as free, see the comments on handling 816 // branches. 817 return isa<ConstantInt>(SI.getCondition()) || 818 dyn_cast_or_null<ConstantInt>( 819 SimplifiedValues.lookup(SI.getCondition())); 820} 821 822bool CallAnalyzer::visitIndirectBrInst(IndirectBrInst &IBI) { 823 // We never want to inline functions that contain an indirectbr. This is 824 // incorrect because all the blockaddress's (in static global initializers 825 // for example) would be referring to the original function, and this 826 // indirect jump would jump from the inlined copy of the function into the 827 // original function which is extremely undefined behavior. 828 // FIXME: This logic isn't really right; we can safely inline functions with 829 // indirectbr's as long as no other function or global references the 830 // blockaddress of a block within the current function. And as a QOI issue, 831 // if someone is using a blockaddress without an indirectbr, and that 832 // reference somehow ends up in another function or global, we probably don't 833 // want to inline this function. 834 HasIndirectBr = true; 835 return false; 836} 837 838bool CallAnalyzer::visitResumeInst(ResumeInst &RI) { 839 // FIXME: It's not clear that a single instruction is an accurate model for 840 // the inline cost of a resume instruction. 841 return false; 842} 843 844bool CallAnalyzer::visitUnreachableInst(UnreachableInst &I) { 845 // FIXME: It might be reasonably to discount the cost of instructions leading 846 // to unreachable as they have the lowest possible impact on both runtime and 847 // code size. 848 return true; // No actual code is needed for unreachable. 849} 850 851bool CallAnalyzer::visitInstruction(Instruction &I) { 852 // Some instructions are free. All of the free intrinsics can also be 853 // handled by SROA, etc. 854 if (TargetTransformInfo::TCC_Free == TTI.getUserCost(&I)) 855 return true; 856 857 // We found something we don't understand or can't handle. Mark any SROA-able 858 // values in the operand list as no longer viable. 859 for (User::op_iterator OI = I.op_begin(), OE = I.op_end(); OI != OE; ++OI) 860 disableSROA(*OI); 861 862 return false; 863} 864 865 866/// \brief Analyze a basic block for its contribution to the inline cost. 867/// 868/// This method walks the analyzer over every instruction in the given basic 869/// block and accounts for their cost during inlining at this callsite. It 870/// aborts early if the threshold has been exceeded or an impossible to inline 871/// construct has been detected. It returns false if inlining is no longer 872/// viable, and true if inlining remains viable. 873bool CallAnalyzer::analyzeBlock(BasicBlock *BB) { 874 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) { 875 ++NumInstructions; 876 if (isa<ExtractElementInst>(I) || I->getType()->isVectorTy()) 877 ++NumVectorInstructions; 878 879 // If the instruction simplified to a constant, there is no cost to this 880 // instruction. Visit the instructions using our InstVisitor to account for 881 // all of the per-instruction logic. The visit tree returns true if we 882 // consumed the instruction in any way, and false if the instruction's base 883 // cost should count against inlining. 884 if (Base::visit(I)) 885 ++NumInstructionsSimplified; 886 else 887 Cost += InlineConstants::InstrCost; 888 889 // If the visit this instruction detected an uninlinable pattern, abort. 890 if (IsRecursiveCall || ExposesReturnsTwice || HasDynamicAlloca || 891 HasIndirectBr) 892 return false; 893 894 // If the caller is a recursive function then we don't want to inline 895 // functions which allocate a lot of stack space because it would increase 896 // the caller stack usage dramatically. 897 if (IsCallerRecursive && 898 AllocatedSize > InlineConstants::TotalAllocaSizeRecursiveCaller) 899 return false; 900 901 if (NumVectorInstructions > NumInstructions/2) 902 VectorBonus = FiftyPercentVectorBonus; 903 else if (NumVectorInstructions > NumInstructions/10) 904 VectorBonus = TenPercentVectorBonus; 905 else 906 VectorBonus = 0; 907 908 // Check if we've past the threshold so we don't spin in huge basic 909 // blocks that will never inline. 910 if (Cost > (Threshold + VectorBonus)) 911 return false; 912 } 913 914 return true; 915} 916 917/// \brief Compute the base pointer and cumulative constant offsets for V. 918/// 919/// This strips all constant offsets off of V, leaving it the base pointer, and 920/// accumulates the total constant offset applied in the returned constant. It 921/// returns 0 if V is not a pointer, and returns the constant '0' if there are 922/// no constant offsets applied. 923ConstantInt *CallAnalyzer::stripAndComputeInBoundsConstantOffsets(Value *&V) { 924 if (!TD || !V->getType()->isPointerTy()) 925 return 0; 926 927 unsigned IntPtrWidth = TD->getPointerSizeInBits(); 928 APInt Offset = APInt::getNullValue(IntPtrWidth); 929 930 // Even though we don't look through PHI nodes, we could be called on an 931 // instruction in an unreachable block, which may be on a cycle. 932 SmallPtrSet<Value *, 4> Visited; 933 Visited.insert(V); 934 do { 935 if (GEPOperator *GEP = dyn_cast<GEPOperator>(V)) { 936 if (!GEP->isInBounds() || !accumulateGEPOffset(*GEP, Offset)) 937 return 0; 938 V = GEP->getPointerOperand(); 939 } else if (Operator::getOpcode(V) == Instruction::BitCast) { 940 V = cast<Operator>(V)->getOperand(0); 941 } else if (GlobalAlias *GA = dyn_cast<GlobalAlias>(V)) { 942 if (GA->mayBeOverridden()) 943 break; 944 V = GA->getAliasee(); 945 } else { 946 break; 947 } 948 assert(V->getType()->isPointerTy() && "Unexpected operand type!"); 949 } while (Visited.insert(V)); 950 951 Type *IntPtrTy = TD->getIntPtrType(V->getContext()); 952 return cast<ConstantInt>(ConstantInt::get(IntPtrTy, Offset)); 953} 954 955/// \brief Analyze a call site for potential inlining. 956/// 957/// Returns true if inlining this call is viable, and false if it is not 958/// viable. It computes the cost and adjusts the threshold based on numerous 959/// factors and heuristics. If this method returns false but the computed cost 960/// is below the computed threshold, then inlining was forcibly disabled by 961/// some artifact of the routine. 962bool CallAnalyzer::analyzeCall(CallSite CS) { 963 ++NumCallsAnalyzed; 964 965 // Track whether the post-inlining function would have more than one basic 966 // block. A single basic block is often intended for inlining. Balloon the 967 // threshold by 50% until we pass the single-BB phase. 968 bool SingleBB = true; 969 int SingleBBBonus = Threshold / 2; 970 Threshold += SingleBBBonus; 971 972 // Perform some tweaks to the cost and threshold based on the direct 973 // callsite information. 974 975 // We want to more aggressively inline vector-dense kernels, so up the 976 // threshold, and we'll lower it if the % of vector instructions gets too 977 // low. 978 assert(NumInstructions == 0); 979 assert(NumVectorInstructions == 0); 980 FiftyPercentVectorBonus = Threshold; 981 TenPercentVectorBonus = Threshold / 2; 982 983 // Give out bonuses per argument, as the instructions setting them up will 984 // be gone after inlining. 985 for (unsigned I = 0, E = CS.arg_size(); I != E; ++I) { 986 if (TD && CS.isByValArgument(I)) { 987 // We approximate the number of loads and stores needed by dividing the 988 // size of the byval type by the target's pointer size. 989 PointerType *PTy = cast<PointerType>(CS.getArgument(I)->getType()); 990 unsigned TypeSize = TD->getTypeSizeInBits(PTy->getElementType()); 991 unsigned PointerSize = TD->getPointerSizeInBits(); 992 // Ceiling division. 993 unsigned NumStores = (TypeSize + PointerSize - 1) / PointerSize; 994 995 // If it generates more than 8 stores it is likely to be expanded as an 996 // inline memcpy so we take that as an upper bound. Otherwise we assume 997 // one load and one store per word copied. 998 // FIXME: The maxStoresPerMemcpy setting from the target should be used 999 // here instead of a magic number of 8, but it's not available via 1000 // DataLayout. 1001 NumStores = std::min(NumStores, 8U); 1002 1003 Cost -= 2 * NumStores * InlineConstants::InstrCost; 1004 } else { 1005 // For non-byval arguments subtract off one instruction per call 1006 // argument. 1007 Cost -= InlineConstants::InstrCost; 1008 } 1009 } 1010 1011 // If there is only one call of the function, and it has internal linkage, 1012 // the cost of inlining it drops dramatically. 1013 bool OnlyOneCallAndLocalLinkage = F.hasLocalLinkage() && F.hasOneUse() && 1014 &F == CS.getCalledFunction(); 1015 if (OnlyOneCallAndLocalLinkage) 1016 Cost += InlineConstants::LastCallToStaticBonus; 1017 1018 // If the instruction after the call, or if the normal destination of the 1019 // invoke is an unreachable instruction, the function is noreturn. As such, 1020 // there is little point in inlining this unless there is literally zero 1021 // cost. 1022 Instruction *Instr = CS.getInstruction(); 1023 if (InvokeInst *II = dyn_cast<InvokeInst>(Instr)) { 1024 if (isa<UnreachableInst>(II->getNormalDest()->begin())) 1025 Threshold = 1; 1026 } else if (isa<UnreachableInst>(++BasicBlock::iterator(Instr))) 1027 Threshold = 1; 1028 1029 // If this function uses the coldcc calling convention, prefer not to inline 1030 // it. 1031 if (F.getCallingConv() == CallingConv::Cold) 1032 Cost += InlineConstants::ColdccPenalty; 1033 1034 // Check if we're done. This can happen due to bonuses and penalties. 1035 if (Cost > Threshold) 1036 return false; 1037 1038 if (F.empty()) 1039 return true; 1040 1041 Function *Caller = CS.getInstruction()->getParent()->getParent(); 1042 // Check if the caller function is recursive itself. 1043 for (Value::use_iterator U = Caller->use_begin(), E = Caller->use_end(); 1044 U != E; ++U) { 1045 CallSite Site(cast<Value>(*U)); 1046 if (!Site) 1047 continue; 1048 Instruction *I = Site.getInstruction(); 1049 if (I->getParent()->getParent() == Caller) { 1050 IsCallerRecursive = true; 1051 break; 1052 } 1053 } 1054 1055 // Populate our simplified values by mapping from function arguments to call 1056 // arguments with known important simplifications. 1057 CallSite::arg_iterator CAI = CS.arg_begin(); 1058 for (Function::arg_iterator FAI = F.arg_begin(), FAE = F.arg_end(); 1059 FAI != FAE; ++FAI, ++CAI) { 1060 assert(CAI != CS.arg_end()); 1061 if (Constant *C = dyn_cast<Constant>(CAI)) 1062 SimplifiedValues[FAI] = C; 1063 1064 Value *PtrArg = *CAI; 1065 if (ConstantInt *C = stripAndComputeInBoundsConstantOffsets(PtrArg)) { 1066 ConstantOffsetPtrs[FAI] = std::make_pair(PtrArg, C->getValue()); 1067 1068 // We can SROA any pointer arguments derived from alloca instructions. 1069 if (isa<AllocaInst>(PtrArg)) { 1070 SROAArgValues[FAI] = PtrArg; 1071 SROAArgCosts[PtrArg] = 0; 1072 } 1073 } 1074 } 1075 NumConstantArgs = SimplifiedValues.size(); 1076 NumConstantOffsetPtrArgs = ConstantOffsetPtrs.size(); 1077 NumAllocaArgs = SROAArgValues.size(); 1078 1079 // The worklist of live basic blocks in the callee *after* inlining. We avoid 1080 // adding basic blocks of the callee which can be proven to be dead for this 1081 // particular call site in order to get more accurate cost estimates. This 1082 // requires a somewhat heavyweight iteration pattern: we need to walk the 1083 // basic blocks in a breadth-first order as we insert live successors. To 1084 // accomplish this, prioritizing for small iterations because we exit after 1085 // crossing our threshold, we use a small-size optimized SetVector. 1086 typedef SetVector<BasicBlock *, SmallVector<BasicBlock *, 16>, 1087 SmallPtrSet<BasicBlock *, 16> > BBSetVector; 1088 BBSetVector BBWorklist; 1089 BBWorklist.insert(&F.getEntryBlock()); 1090 // Note that we *must not* cache the size, this loop grows the worklist. 1091 for (unsigned Idx = 0; Idx != BBWorklist.size(); ++Idx) { 1092 // Bail out the moment we cross the threshold. This means we'll under-count 1093 // the cost, but only when undercounting doesn't matter. 1094 if (Cost > (Threshold + VectorBonus)) 1095 break; 1096 1097 BasicBlock *BB = BBWorklist[Idx]; 1098 if (BB->empty()) 1099 continue; 1100 1101 // Analyze the cost of this block. If we blow through the threshold, this 1102 // returns false, and we can bail on out. 1103 if (!analyzeBlock(BB)) { 1104 if (IsRecursiveCall || ExposesReturnsTwice || HasDynamicAlloca || 1105 HasIndirectBr) 1106 return false; 1107 1108 // If the caller is a recursive function then we don't want to inline 1109 // functions which allocate a lot of stack space because it would increase 1110 // the caller stack usage dramatically. 1111 if (IsCallerRecursive && 1112 AllocatedSize > InlineConstants::TotalAllocaSizeRecursiveCaller) 1113 return false; 1114 1115 break; 1116 } 1117 1118 TerminatorInst *TI = BB->getTerminator(); 1119 1120 // Add in the live successors by first checking whether we have terminator 1121 // that may be simplified based on the values simplified by this call. 1122 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) { 1123 if (BI->isConditional()) { 1124 Value *Cond = BI->getCondition(); 1125 if (ConstantInt *SimpleCond 1126 = dyn_cast_or_null<ConstantInt>(SimplifiedValues.lookup(Cond))) { 1127 BBWorklist.insert(BI->getSuccessor(SimpleCond->isZero() ? 1 : 0)); 1128 continue; 1129 } 1130 } 1131 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) { 1132 Value *Cond = SI->getCondition(); 1133 if (ConstantInt *SimpleCond 1134 = dyn_cast_or_null<ConstantInt>(SimplifiedValues.lookup(Cond))) { 1135 BBWorklist.insert(SI->findCaseValue(SimpleCond).getCaseSuccessor()); 1136 continue; 1137 } 1138 } 1139 1140 // If we're unable to select a particular successor, just count all of 1141 // them. 1142 for (unsigned TIdx = 0, TSize = TI->getNumSuccessors(); TIdx != TSize; 1143 ++TIdx) 1144 BBWorklist.insert(TI->getSuccessor(TIdx)); 1145 1146 // If we had any successors at this point, than post-inlining is likely to 1147 // have them as well. Note that we assume any basic blocks which existed 1148 // due to branches or switches which folded above will also fold after 1149 // inlining. 1150 if (SingleBB && TI->getNumSuccessors() > 1) { 1151 // Take off the bonus we applied to the threshold. 1152 Threshold -= SingleBBBonus; 1153 SingleBB = false; 1154 } 1155 } 1156 1157 // If this is a noduplicate call, we can still inline as long as 1158 // inlining this would cause the removal of the caller (so the instruction 1159 // is not actually duplicated, just moved). 1160 if (!OnlyOneCallAndLocalLinkage && ContainsNoDuplicateCall) 1161 return false; 1162 1163 Threshold += VectorBonus; 1164 1165 return Cost < Threshold; 1166} 1167 1168#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) 1169/// \brief Dump stats about this call's analysis. 1170void CallAnalyzer::dump() { 1171#define DEBUG_PRINT_STAT(x) llvm::dbgs() << " " #x ": " << x << "\n" 1172 DEBUG_PRINT_STAT(NumConstantArgs); 1173 DEBUG_PRINT_STAT(NumConstantOffsetPtrArgs); 1174 DEBUG_PRINT_STAT(NumAllocaArgs); 1175 DEBUG_PRINT_STAT(NumConstantPtrCmps); 1176 DEBUG_PRINT_STAT(NumConstantPtrDiffs); 1177 DEBUG_PRINT_STAT(NumInstructionsSimplified); 1178 DEBUG_PRINT_STAT(SROACostSavings); 1179 DEBUG_PRINT_STAT(SROACostSavingsLost); 1180 DEBUG_PRINT_STAT(ContainsNoDuplicateCall); 1181#undef DEBUG_PRINT_STAT 1182} 1183#endif 1184 1185INITIALIZE_PASS_BEGIN(InlineCostAnalysis, "inline-cost", "Inline Cost Analysis", 1186 true, true) 1187INITIALIZE_AG_DEPENDENCY(TargetTransformInfo) 1188INITIALIZE_PASS_END(InlineCostAnalysis, "inline-cost", "Inline Cost Analysis", 1189 true, true) 1190 1191char InlineCostAnalysis::ID = 0; 1192 1193InlineCostAnalysis::InlineCostAnalysis() : CallGraphSCCPass(ID), TD(0) {} 1194 1195InlineCostAnalysis::~InlineCostAnalysis() {} 1196 1197void InlineCostAnalysis::getAnalysisUsage(AnalysisUsage &AU) const { 1198 AU.setPreservesAll(); 1199 AU.addRequired<TargetTransformInfo>(); 1200 CallGraphSCCPass::getAnalysisUsage(AU); 1201} 1202 1203bool InlineCostAnalysis::runOnSCC(CallGraphSCC &SCC) { 1204 TD = getAnalysisIfAvailable<DataLayout>(); 1205 TTI = &getAnalysis<TargetTransformInfo>(); 1206 return false; 1207} 1208 1209InlineCost InlineCostAnalysis::getInlineCost(CallSite CS, int Threshold) { 1210 return getInlineCost(CS, CS.getCalledFunction(), Threshold); 1211} 1212 1213/// \brief Test that two functions either have or have not the given attribute 1214/// at the same time. 1215static bool attributeMatches(Function *F1, Function *F2, 1216 Attribute::AttrKind Attr) { 1217 return F1->hasFnAttribute(Attr) == F2->hasFnAttribute(Attr); 1218} 1219 1220/// \brief Test that there are no attribute conflicts between Caller and Callee 1221/// that prevent inlining. 1222static bool functionsHaveCompatibleAttributes(Function *Caller, 1223 Function *Callee) { 1224 return attributeMatches(Caller, Callee, Attribute::SanitizeAddress) && 1225 attributeMatches(Caller, Callee, Attribute::SanitizeMemory) && 1226 attributeMatches(Caller, Callee, Attribute::SanitizeThread); 1227} 1228 1229InlineCost InlineCostAnalysis::getInlineCost(CallSite CS, Function *Callee, 1230 int Threshold) { 1231 // Cannot inline indirect calls. 1232 if (!Callee) 1233 return llvm::InlineCost::getNever(); 1234 1235 // Calls to functions with always-inline attributes should be inlined 1236 // whenever possible. 1237 if (Callee->hasFnAttribute(Attribute::AlwaysInline)) { 1238 if (isInlineViable(*Callee)) 1239 return llvm::InlineCost::getAlways(); 1240 return llvm::InlineCost::getNever(); 1241 } 1242 1243 // Never inline functions with conflicting attributes (unless callee has 1244 // always-inline attribute). 1245 if (!functionsHaveCompatibleAttributes(CS.getCaller(), Callee)) 1246 return llvm::InlineCost::getNever(); 1247 1248 // Don't inline this call if the caller has the optnone attribute. 1249 if (CS.getCaller()->hasFnAttribute(Attribute::OptimizeNone)) 1250 return llvm::InlineCost::getNever(); 1251 1252 // Don't inline functions which can be redefined at link-time to mean 1253 // something else. Don't inline functions marked noinline or call sites 1254 // marked noinline. 1255 if (Callee->mayBeOverridden() || 1256 Callee->hasFnAttribute(Attribute::NoInline) || CS.isNoInline()) 1257 return llvm::InlineCost::getNever(); 1258 1259 DEBUG(llvm::dbgs() << " Analyzing call of " << Callee->getName() 1260 << "...\n"); 1261 1262 CallAnalyzer CA(TD, *TTI, *Callee, Threshold); 1263 bool ShouldInline = CA.analyzeCall(CS); 1264 1265 DEBUG(CA.dump()); 1266 1267 // Check if there was a reason to force inlining or no inlining. 1268 if (!ShouldInline && CA.getCost() < CA.getThreshold()) 1269 return InlineCost::getNever(); 1270 if (ShouldInline && CA.getCost() >= CA.getThreshold()) 1271 return InlineCost::getAlways(); 1272 1273 return llvm::InlineCost::get(CA.getCost(), CA.getThreshold()); 1274} 1275 1276bool InlineCostAnalysis::isInlineViable(Function &F) { 1277 bool ReturnsTwice = 1278 F.getAttributes().hasAttribute(AttributeSet::FunctionIndex, 1279 Attribute::ReturnsTwice); 1280 for (Function::iterator BI = F.begin(), BE = F.end(); BI != BE; ++BI) { 1281 // Disallow inlining of functions which contain an indirect branch. 1282 if (isa<IndirectBrInst>(BI->getTerminator())) 1283 return false; 1284 1285 for (BasicBlock::iterator II = BI->begin(), IE = BI->end(); II != IE; 1286 ++II) { 1287 CallSite CS(II); 1288 if (!CS) 1289 continue; 1290 1291 // Disallow recursive calls. 1292 if (&F == CS.getCalledFunction()) 1293 return false; 1294 1295 // Disallow calls which expose returns-twice to a function not previously 1296 // attributed as such. 1297 if (!ReturnsTwice && CS.isCall() && 1298 cast<CallInst>(CS.getInstruction())->canReturnTwice()) 1299 return false; 1300 } 1301 } 1302 1303 return true; 1304} 1305