1 //===- Inliner.cpp - Code common to all inliners --------------------------===// 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 the mechanics required to implement inlining without 11 // missing any calls and updating the call graph. The decisions of which calls 12 // are profitable to inline are implemented elsewhere. 13 // 14 //===----------------------------------------------------------------------===// 15 16 #define DEBUG_TYPE "inline" 17 #include "llvm/Module.h" 18 #include "llvm/Instructions.h" 19 #include "llvm/IntrinsicInst.h" 20 #include "llvm/Analysis/CallGraph.h" 21 #include "llvm/Analysis/InlineCost.h" 22 #include "llvm/Target/TargetData.h" 23 #include "llvm/Transforms/IPO/InlinerPass.h" 24 #include "llvm/Transforms/Utils/Cloning.h" 25 #include "llvm/Transforms/Utils/Local.h" 26 #include "llvm/Support/CallSite.h" 27 #include "llvm/Support/CommandLine.h" 28 #include "llvm/Support/Debug.h" 29 #include "llvm/Support/raw_ostream.h" 30 #include "llvm/ADT/SmallPtrSet.h" 31 #include "llvm/ADT/Statistic.h" 32 using namespace llvm; 33 34 STATISTIC(NumInlined, "Number of functions inlined"); 35 STATISTIC(NumCallsDeleted, "Number of call sites deleted, not inlined"); 36 STATISTIC(NumDeleted, "Number of functions deleted because all callers found"); 37 STATISTIC(NumMergedAllocas, "Number of allocas merged together"); 38 39 static cl::opt<int> 40 InlineLimit("inline-threshold", cl::Hidden, cl::init(225), cl::ZeroOrMore, 41 cl::desc("Control the amount of inlining to perform (default = 225)")); 42 43 static cl::opt<int> 44 HintThreshold("inlinehint-threshold", cl::Hidden, cl::init(325), 45 cl::desc("Threshold for inlining functions with inline hint")); 46 47 // Threshold to use when optsize is specified (and there is no -inline-limit). 48 const int OptSizeThreshold = 75; 49 50 Inliner::Inliner(char &ID) 51 : CallGraphSCCPass(ID), InlineThreshold(InlineLimit) {} 52 53 Inliner::Inliner(char &ID, int Threshold) 54 : CallGraphSCCPass(ID), InlineThreshold(InlineLimit.getNumOccurrences() > 0 ? 55 InlineLimit : Threshold) {} 56 57 /// getAnalysisUsage - For this class, we declare that we require and preserve 58 /// the call graph. If the derived class implements this method, it should 59 /// always explicitly call the implementation here. 60 void Inliner::getAnalysisUsage(AnalysisUsage &Info) const { 61 CallGraphSCCPass::getAnalysisUsage(Info); 62 } 63 64 65 typedef DenseMap<ArrayType*, std::vector<AllocaInst*> > 66 InlinedArrayAllocasTy; 67 68 /// InlineCallIfPossible - If it is possible to inline the specified call site, 69 /// do so and update the CallGraph for this operation. 70 /// 71 /// This function also does some basic book-keeping to update the IR. The 72 /// InlinedArrayAllocas map keeps track of any allocas that are already 73 /// available from other functions inlined into the caller. If we are able to 74 /// inline this call site we attempt to reuse already available allocas or add 75 /// any new allocas to the set if not possible. 76 static bool InlineCallIfPossible(CallSite CS, InlineFunctionInfo &IFI, 77 InlinedArrayAllocasTy &InlinedArrayAllocas, 78 int InlineHistory) { 79 Function *Callee = CS.getCalledFunction(); 80 Function *Caller = CS.getCaller(); 81 82 // Try to inline the function. Get the list of static allocas that were 83 // inlined. 84 if (!InlineFunction(CS, IFI)) 85 return false; 86 87 // If the inlined function had a higher stack protection level than the 88 // calling function, then bump up the caller's stack protection level. 89 if (Callee->hasFnAttr(Attribute::StackProtectReq)) 90 Caller->addFnAttr(Attribute::StackProtectReq); 91 else if (Callee->hasFnAttr(Attribute::StackProtect) && 92 !Caller->hasFnAttr(Attribute::StackProtectReq)) 93 Caller->addFnAttr(Attribute::StackProtect); 94 95 // Look at all of the allocas that we inlined through this call site. If we 96 // have already inlined other allocas through other calls into this function, 97 // then we know that they have disjoint lifetimes and that we can merge them. 98 // 99 // There are many heuristics possible for merging these allocas, and the 100 // different options have different tradeoffs. One thing that we *really* 101 // don't want to hurt is SRoA: once inlining happens, often allocas are no 102 // longer address taken and so they can be promoted. 103 // 104 // Our "solution" for that is to only merge allocas whose outermost type is an 105 // array type. These are usually not promoted because someone is using a 106 // variable index into them. These are also often the most important ones to 107 // merge. 108 // 109 // A better solution would be to have real memory lifetime markers in the IR 110 // and not have the inliner do any merging of allocas at all. This would 111 // allow the backend to do proper stack slot coloring of all allocas that 112 // *actually make it to the backend*, which is really what we want. 113 // 114 // Because we don't have this information, we do this simple and useful hack. 115 // 116 SmallPtrSet<AllocaInst*, 16> UsedAllocas; 117 118 // When processing our SCC, check to see if CS was inlined from some other 119 // call site. For example, if we're processing "A" in this code: 120 // A() { B() } 121 // B() { x = alloca ... C() } 122 // C() { y = alloca ... } 123 // Assume that C was not inlined into B initially, and so we're processing A 124 // and decide to inline B into A. Doing this makes an alloca available for 125 // reuse and makes a callsite (C) available for inlining. When we process 126 // the C call site we don't want to do any alloca merging between X and Y 127 // because their scopes are not disjoint. We could make this smarter by 128 // keeping track of the inline history for each alloca in the 129 // InlinedArrayAllocas but this isn't likely to be a significant win. 130 if (InlineHistory != -1) // Only do merging for top-level call sites in SCC. 131 return true; 132 133 // Loop over all the allocas we have so far and see if they can be merged with 134 // a previously inlined alloca. If not, remember that we had it. 135 for (unsigned AllocaNo = 0, e = IFI.StaticAllocas.size(); 136 AllocaNo != e; ++AllocaNo) { 137 AllocaInst *AI = IFI.StaticAllocas[AllocaNo]; 138 139 // Don't bother trying to merge array allocations (they will usually be 140 // canonicalized to be an allocation *of* an array), or allocations whose 141 // type is not itself an array (because we're afraid of pessimizing SRoA). 142 ArrayType *ATy = dyn_cast<ArrayType>(AI->getAllocatedType()); 143 if (ATy == 0 || AI->isArrayAllocation()) 144 continue; 145 146 // Get the list of all available allocas for this array type. 147 std::vector<AllocaInst*> &AllocasForType = InlinedArrayAllocas[ATy]; 148 149 // Loop over the allocas in AllocasForType to see if we can reuse one. Note 150 // that we have to be careful not to reuse the same "available" alloca for 151 // multiple different allocas that we just inlined, we use the 'UsedAllocas' 152 // set to keep track of which "available" allocas are being used by this 153 // function. Also, AllocasForType can be empty of course! 154 bool MergedAwayAlloca = false; 155 for (unsigned i = 0, e = AllocasForType.size(); i != e; ++i) { 156 AllocaInst *AvailableAlloca = AllocasForType[i]; 157 158 // The available alloca has to be in the right function, not in some other 159 // function in this SCC. 160 if (AvailableAlloca->getParent() != AI->getParent()) 161 continue; 162 163 // If the inlined function already uses this alloca then we can't reuse 164 // it. 165 if (!UsedAllocas.insert(AvailableAlloca)) 166 continue; 167 168 // Otherwise, we *can* reuse it, RAUW AI into AvailableAlloca and declare 169 // success! 170 DEBUG(dbgs() << " ***MERGED ALLOCA: " << *AI << "\n\t\tINTO: " 171 << *AvailableAlloca << '\n'); 172 173 AI->replaceAllUsesWith(AvailableAlloca); 174 AI->eraseFromParent(); 175 MergedAwayAlloca = true; 176 ++NumMergedAllocas; 177 IFI.StaticAllocas[AllocaNo] = 0; 178 break; 179 } 180 181 // If we already nuked the alloca, we're done with it. 182 if (MergedAwayAlloca) 183 continue; 184 185 // If we were unable to merge away the alloca either because there are no 186 // allocas of the right type available or because we reused them all 187 // already, remember that this alloca came from an inlined function and mark 188 // it used so we don't reuse it for other allocas from this inline 189 // operation. 190 AllocasForType.push_back(AI); 191 UsedAllocas.insert(AI); 192 } 193 194 return true; 195 } 196 197 unsigned Inliner::getInlineThreshold(CallSite CS) const { 198 int thres = InlineThreshold; 199 200 // Listen to optsize when -inline-limit is not given. 201 Function *Caller = CS.getCaller(); 202 if (Caller && !Caller->isDeclaration() && 203 Caller->hasFnAttr(Attribute::OptimizeForSize) && 204 InlineLimit.getNumOccurrences() == 0) 205 thres = OptSizeThreshold; 206 207 // Listen to inlinehint when it would increase the threshold. 208 Function *Callee = CS.getCalledFunction(); 209 if (HintThreshold > thres && Callee && !Callee->isDeclaration() && 210 Callee->hasFnAttr(Attribute::InlineHint)) 211 thres = HintThreshold; 212 213 return thres; 214 } 215 216 /// shouldInline - Return true if the inliner should attempt to inline 217 /// at the given CallSite. 218 bool Inliner::shouldInline(CallSite CS) { 219 InlineCost IC = getInlineCost(CS); 220 221 if (IC.isAlways()) { 222 DEBUG(dbgs() << " Inlining: cost=always" 223 << ", Call: " << *CS.getInstruction() << "\n"); 224 return true; 225 } 226 227 if (IC.isNever()) { 228 DEBUG(dbgs() << " NOT Inlining: cost=never" 229 << ", Call: " << *CS.getInstruction() << "\n"); 230 return false; 231 } 232 233 int Cost = IC.getValue(); 234 Function *Caller = CS.getCaller(); 235 int CurrentThreshold = getInlineThreshold(CS); 236 float FudgeFactor = getInlineFudgeFactor(CS); 237 int AdjThreshold = (int)(CurrentThreshold * FudgeFactor); 238 if (Cost >= AdjThreshold) { 239 DEBUG(dbgs() << " NOT Inlining: cost=" << Cost 240 << ", thres=" << AdjThreshold 241 << ", Call: " << *CS.getInstruction() << "\n"); 242 return false; 243 } 244 245 // Try to detect the case where the current inlining candidate caller 246 // (call it B) is a static function and is an inlining candidate elsewhere, 247 // and the current candidate callee (call it C) is large enough that 248 // inlining it into B would make B too big to inline later. In these 249 // circumstances it may be best not to inline C into B, but to inline B 250 // into its callers. 251 if (Caller->hasLocalLinkage()) { 252 int TotalSecondaryCost = 0; 253 bool outerCallsFound = false; 254 // This bool tracks what happens if we do NOT inline C into B. 255 bool callerWillBeRemoved = true; 256 // This bool tracks what happens if we DO inline C into B. 257 bool inliningPreventsSomeOuterInline = false; 258 for (Value::use_iterator I = Caller->use_begin(), E =Caller->use_end(); 259 I != E; ++I) { 260 CallSite CS2(*I); 261 262 // If this isn't a call to Caller (it could be some other sort 263 // of reference) skip it. Such references will prevent the caller 264 // from being removed. 265 if (!CS2 || CS2.getCalledFunction() != Caller) { 266 callerWillBeRemoved = false; 267 continue; 268 } 269 270 InlineCost IC2 = getInlineCost(CS2); 271 if (IC2.isNever()) 272 callerWillBeRemoved = false; 273 if (IC2.isAlways() || IC2.isNever()) 274 continue; 275 276 outerCallsFound = true; 277 int Cost2 = IC2.getValue(); 278 int CurrentThreshold2 = getInlineThreshold(CS2); 279 float FudgeFactor2 = getInlineFudgeFactor(CS2); 280 281 if (Cost2 >= (int)(CurrentThreshold2 * FudgeFactor2)) 282 callerWillBeRemoved = false; 283 284 // See if we have this case. We subtract off the penalty 285 // for the call instruction, which we would be deleting. 286 if (Cost2 < (int)(CurrentThreshold2 * FudgeFactor2) && 287 Cost2 + Cost - (InlineConstants::CallPenalty + 1) >= 288 (int)(CurrentThreshold2 * FudgeFactor2)) { 289 inliningPreventsSomeOuterInline = true; 290 TotalSecondaryCost += Cost2; 291 } 292 } 293 // If all outer calls to Caller would get inlined, the cost for the last 294 // one is set very low by getInlineCost, in anticipation that Caller will 295 // be removed entirely. We did not account for this above unless there 296 // is only one caller of Caller. 297 if (callerWillBeRemoved && Caller->use_begin() != Caller->use_end()) 298 TotalSecondaryCost += InlineConstants::LastCallToStaticBonus; 299 300 if (outerCallsFound && inliningPreventsSomeOuterInline && 301 TotalSecondaryCost < Cost) { 302 DEBUG(dbgs() << " NOT Inlining: " << *CS.getInstruction() << 303 " Cost = " << Cost << 304 ", outer Cost = " << TotalSecondaryCost << '\n'); 305 return false; 306 } 307 } 308 309 DEBUG(dbgs() << " Inlining: cost=" << Cost 310 << ", thres=" << AdjThreshold 311 << ", Call: " << *CS.getInstruction() << '\n'); 312 return true; 313 } 314 315 /// InlineHistoryIncludes - Return true if the specified inline history ID 316 /// indicates an inline history that includes the specified function. 317 static bool InlineHistoryIncludes(Function *F, int InlineHistoryID, 318 const SmallVectorImpl<std::pair<Function*, int> > &InlineHistory) { 319 while (InlineHistoryID != -1) { 320 assert(unsigned(InlineHistoryID) < InlineHistory.size() && 321 "Invalid inline history ID"); 322 if (InlineHistory[InlineHistoryID].first == F) 323 return true; 324 InlineHistoryID = InlineHistory[InlineHistoryID].second; 325 } 326 return false; 327 } 328 329 330 bool Inliner::runOnSCC(CallGraphSCC &SCC) { 331 CallGraph &CG = getAnalysis<CallGraph>(); 332 const TargetData *TD = getAnalysisIfAvailable<TargetData>(); 333 334 SmallPtrSet<Function*, 8> SCCFunctions; 335 DEBUG(dbgs() << "Inliner visiting SCC:"); 336 for (CallGraphSCC::iterator I = SCC.begin(), E = SCC.end(); I != E; ++I) { 337 Function *F = (*I)->getFunction(); 338 if (F) SCCFunctions.insert(F); 339 DEBUG(dbgs() << " " << (F ? F->getName() : "INDIRECTNODE")); 340 } 341 342 // Scan through and identify all call sites ahead of time so that we only 343 // inline call sites in the original functions, not call sites that result 344 // from inlining other functions. 345 SmallVector<std::pair<CallSite, int>, 16> CallSites; 346 347 // When inlining a callee produces new call sites, we want to keep track of 348 // the fact that they were inlined from the callee. This allows us to avoid 349 // infinite inlining in some obscure cases. To represent this, we use an 350 // index into the InlineHistory vector. 351 SmallVector<std::pair<Function*, int>, 8> InlineHistory; 352 353 for (CallGraphSCC::iterator I = SCC.begin(), E = SCC.end(); I != E; ++I) { 354 Function *F = (*I)->getFunction(); 355 if (!F) continue; 356 357 for (Function::iterator BB = F->begin(), E = F->end(); BB != E; ++BB) 358 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) { 359 CallSite CS(cast<Value>(I)); 360 // If this isn't a call, or it is a call to an intrinsic, it can 361 // never be inlined. 362 if (!CS || isa<IntrinsicInst>(I)) 363 continue; 364 365 // If this is a direct call to an external function, we can never inline 366 // it. If it is an indirect call, inlining may resolve it to be a 367 // direct call, so we keep it. 368 if (CS.getCalledFunction() && CS.getCalledFunction()->isDeclaration()) 369 continue; 370 371 CallSites.push_back(std::make_pair(CS, -1)); 372 } 373 } 374 375 DEBUG(dbgs() << ": " << CallSites.size() << " call sites.\n"); 376 377 // If there are no calls in this function, exit early. 378 if (CallSites.empty()) 379 return false; 380 381 // Now that we have all of the call sites, move the ones to functions in the 382 // current SCC to the end of the list. 383 unsigned FirstCallInSCC = CallSites.size(); 384 for (unsigned i = 0; i < FirstCallInSCC; ++i) 385 if (Function *F = CallSites[i].first.getCalledFunction()) 386 if (SCCFunctions.count(F)) 387 std::swap(CallSites[i--], CallSites[--FirstCallInSCC]); 388 389 390 InlinedArrayAllocasTy InlinedArrayAllocas; 391 InlineFunctionInfo InlineInfo(&CG, TD); 392 393 // Now that we have all of the call sites, loop over them and inline them if 394 // it looks profitable to do so. 395 bool Changed = false; 396 bool LocalChange; 397 do { 398 LocalChange = false; 399 // Iterate over the outer loop because inlining functions can cause indirect 400 // calls to become direct calls. 401 for (unsigned CSi = 0; CSi != CallSites.size(); ++CSi) { 402 CallSite CS = CallSites[CSi].first; 403 404 Function *Caller = CS.getCaller(); 405 Function *Callee = CS.getCalledFunction(); 406 407 // If this call site is dead and it is to a readonly function, we should 408 // just delete the call instead of trying to inline it, regardless of 409 // size. This happens because IPSCCP propagates the result out of the 410 // call and then we're left with the dead call. 411 if (isInstructionTriviallyDead(CS.getInstruction())) { 412 DEBUG(dbgs() << " -> Deleting dead call: " 413 << *CS.getInstruction() << "\n"); 414 // Update the call graph by deleting the edge from Callee to Caller. 415 CG[Caller]->removeCallEdgeFor(CS); 416 CS.getInstruction()->eraseFromParent(); 417 ++NumCallsDeleted; 418 // Update the cached cost info with the missing call 419 growCachedCostInfo(Caller, NULL); 420 } else { 421 // We can only inline direct calls to non-declarations. 422 if (Callee == 0 || Callee->isDeclaration()) continue; 423 424 // If this call site was obtained by inlining another function, verify 425 // that the include path for the function did not include the callee 426 // itself. If so, we'd be recursively inlining the same function, 427 // which would provide the same callsites, which would cause us to 428 // infinitely inline. 429 int InlineHistoryID = CallSites[CSi].second; 430 if (InlineHistoryID != -1 && 431 InlineHistoryIncludes(Callee, InlineHistoryID, InlineHistory)) 432 continue; 433 434 435 // If the policy determines that we should inline this function, 436 // try to do so. 437 if (!shouldInline(CS)) 438 continue; 439 440 // Attempt to inline the function. 441 if (!InlineCallIfPossible(CS, InlineInfo, InlinedArrayAllocas, 442 InlineHistoryID)) 443 continue; 444 ++NumInlined; 445 446 // If inlining this function gave us any new call sites, throw them 447 // onto our worklist to process. They are useful inline candidates. 448 if (!InlineInfo.InlinedCalls.empty()) { 449 // Create a new inline history entry for this, so that we remember 450 // that these new callsites came about due to inlining Callee. 451 int NewHistoryID = InlineHistory.size(); 452 InlineHistory.push_back(std::make_pair(Callee, InlineHistoryID)); 453 454 for (unsigned i = 0, e = InlineInfo.InlinedCalls.size(); 455 i != e; ++i) { 456 Value *Ptr = InlineInfo.InlinedCalls[i]; 457 CallSites.push_back(std::make_pair(CallSite(Ptr), NewHistoryID)); 458 } 459 } 460 461 // Update the cached cost info with the inlined call. 462 growCachedCostInfo(Caller, Callee); 463 } 464 465 // If we inlined or deleted the last possible call site to the function, 466 // delete the function body now. 467 if (Callee && Callee->use_empty() && Callee->hasLocalLinkage() && 468 // TODO: Can remove if in SCC now. 469 !SCCFunctions.count(Callee) && 470 471 // The function may be apparently dead, but if there are indirect 472 // callgraph references to the node, we cannot delete it yet, this 473 // could invalidate the CGSCC iterator. 474 CG[Callee]->getNumReferences() == 0) { 475 DEBUG(dbgs() << " -> Deleting dead function: " 476 << Callee->getName() << "\n"); 477 CallGraphNode *CalleeNode = CG[Callee]; 478 479 // Remove any call graph edges from the callee to its callees. 480 CalleeNode->removeAllCalledFunctions(); 481 482 resetCachedCostInfo(Callee); 483 484 // Removing the node for callee from the call graph and delete it. 485 delete CG.removeFunctionFromModule(CalleeNode); 486 ++NumDeleted; 487 } 488 489 // Remove this call site from the list. If possible, use 490 // swap/pop_back for efficiency, but do not use it if doing so would 491 // move a call site to a function in this SCC before the 492 // 'FirstCallInSCC' barrier. 493 if (SCC.isSingular()) { 494 CallSites[CSi] = CallSites.back(); 495 CallSites.pop_back(); 496 } else { 497 CallSites.erase(CallSites.begin()+CSi); 498 } 499 --CSi; 500 501 Changed = true; 502 LocalChange = true; 503 } 504 } while (LocalChange); 505 506 return Changed; 507 } 508 509 // doFinalization - Remove now-dead linkonce functions at the end of 510 // processing to avoid breaking the SCC traversal. 511 bool Inliner::doFinalization(CallGraph &CG) { 512 return removeDeadFunctions(CG); 513 } 514 515 /// removeDeadFunctions - Remove dead functions that are not included in 516 /// DNR (Do Not Remove) list. 517 bool Inliner::removeDeadFunctions(CallGraph &CG, 518 SmallPtrSet<const Function *, 16> *DNR) { 519 SmallPtrSet<CallGraphNode*, 16> FunctionsToRemove; 520 521 // Scan for all of the functions, looking for ones that should now be removed 522 // from the program. Insert the dead ones in the FunctionsToRemove set. 523 for (CallGraph::iterator I = CG.begin(), E = CG.end(); I != E; ++I) { 524 CallGraphNode *CGN = I->second; 525 if (CGN->getFunction() == 0) 526 continue; 527 528 Function *F = CGN->getFunction(); 529 530 // If the only remaining users of the function are dead constants, remove 531 // them. 532 F->removeDeadConstantUsers(); 533 534 if (DNR && DNR->count(F)) 535 continue; 536 if (!F->hasLinkOnceLinkage() && !F->hasLocalLinkage() && 537 !F->hasAvailableExternallyLinkage()) 538 continue; 539 if (!F->use_empty()) 540 continue; 541 542 // Remove any call graph edges from the function to its callees. 543 CGN->removeAllCalledFunctions(); 544 545 // Remove any edges from the external node to the function's call graph 546 // node. These edges might have been made irrelegant due to 547 // optimization of the program. 548 CG.getExternalCallingNode()->removeAnyCallEdgeTo(CGN); 549 550 // Removing the node for callee from the call graph and delete it. 551 FunctionsToRemove.insert(CGN); 552 } 553 554 // Now that we know which functions to delete, do so. We didn't want to do 555 // this inline, because that would invalidate our CallGraph::iterator 556 // objects. :( 557 // 558 // Note that it doesn't matter that we are iterating over a non-stable set 559 // here to do this, it doesn't matter which order the functions are deleted 560 // in. 561 bool Changed = false; 562 for (SmallPtrSet<CallGraphNode*, 16>::iterator I = FunctionsToRemove.begin(), 563 E = FunctionsToRemove.end(); I != E; ++I) { 564 resetCachedCostInfo((*I)->getFunction()); 565 delete CG.removeFunctionFromModule(*I); 566 ++NumDeleted; 567 Changed = true; 568 } 569 570 return Changed; 571 } 572