Inliner.cpp revision d6fc26217e194372cabe4ef9e2514beac511a943
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" 32using namespace llvm; 33 34STATISTIC(NumInlined, "Number of functions inlined"); 35STATISTIC(NumCallsDeleted, "Number of call sites deleted, not inlined"); 36STATISTIC(NumDeleted, "Number of functions deleted because all callers found"); 37STATISTIC(NumMergedAllocas, "Number of allocas merged together"); 38 39// This weirdly named statistic tracks the number of times that, when attemting 40// to inline a function A into B, we analyze the callers of B in order to see 41// if those would be more profitable and blocked inline steps. 42STATISTIC(NumCallerCallersAnalyzed, "Number of caller-callers analyzed"); 43 44static cl::opt<int> 45InlineLimit("inline-threshold", cl::Hidden, cl::init(225), cl::ZeroOrMore, 46 cl::desc("Control the amount of inlining to perform (default = 225)")); 47 48static cl::opt<int> 49HintThreshold("inlinehint-threshold", cl::Hidden, cl::init(325), 50 cl::desc("Threshold for inlining functions with inline hint")); 51 52// Threshold to use when optsize is specified (and there is no -inline-limit). 53const int OptSizeThreshold = 75; 54 55Inliner::Inliner(char &ID) 56 : CallGraphSCCPass(ID), InlineThreshold(InlineLimit), InsertLifetime(true) {} 57 58Inliner::Inliner(char &ID, int Threshold, bool InsertLifetime) 59 : CallGraphSCCPass(ID), InlineThreshold(InlineLimit.getNumOccurrences() > 0 ? 60 InlineLimit : Threshold), 61 InsertLifetime(InsertLifetime) {} 62 63/// getAnalysisUsage - For this class, we declare that we require and preserve 64/// the call graph. If the derived class implements this method, it should 65/// always explicitly call the implementation here. 66void Inliner::getAnalysisUsage(AnalysisUsage &Info) const { 67 CallGraphSCCPass::getAnalysisUsage(Info); 68} 69 70 71typedef DenseMap<ArrayType*, std::vector<AllocaInst*> > 72InlinedArrayAllocasTy; 73 74/// InlineCallIfPossible - If it is possible to inline the specified call site, 75/// do so and update the CallGraph for this operation. 76/// 77/// This function also does some basic book-keeping to update the IR. The 78/// InlinedArrayAllocas map keeps track of any allocas that are already 79/// available from other functions inlined into the caller. If we are able to 80/// inline this call site we attempt to reuse already available allocas or add 81/// any new allocas to the set if not possible. 82static bool InlineCallIfPossible(CallSite CS, InlineFunctionInfo &IFI, 83 InlinedArrayAllocasTy &InlinedArrayAllocas, 84 int InlineHistory, bool InsertLifetime) { 85 Function *Callee = CS.getCalledFunction(); 86 Function *Caller = CS.getCaller(); 87 88 // Try to inline the function. Get the list of static allocas that were 89 // inlined. 90 if (!InlineFunction(CS, IFI, InsertLifetime)) 91 return false; 92 93 // If the inlined function had a higher stack protection level than the 94 // calling function, then bump up the caller's stack protection level. 95 if (Callee->hasFnAttr(Attribute::StackProtectReq)) 96 Caller->addFnAttr(Attribute::StackProtectReq); 97 else if (Callee->hasFnAttr(Attribute::StackProtect) && 98 !Caller->hasFnAttr(Attribute::StackProtectReq)) 99 Caller->addFnAttr(Attribute::StackProtect); 100 101 // Look at all of the allocas that we inlined through this call site. If we 102 // have already inlined other allocas through other calls into this function, 103 // then we know that they have disjoint lifetimes and that we can merge them. 104 // 105 // There are many heuristics possible for merging these allocas, and the 106 // different options have different tradeoffs. One thing that we *really* 107 // don't want to hurt is SRoA: once inlining happens, often allocas are no 108 // longer address taken and so they can be promoted. 109 // 110 // Our "solution" for that is to only merge allocas whose outermost type is an 111 // array type. These are usually not promoted because someone is using a 112 // variable index into them. These are also often the most important ones to 113 // merge. 114 // 115 // A better solution would be to have real memory lifetime markers in the IR 116 // and not have the inliner do any merging of allocas at all. This would 117 // allow the backend to do proper stack slot coloring of all allocas that 118 // *actually make it to the backend*, which is really what we want. 119 // 120 // Because we don't have this information, we do this simple and useful hack. 121 // 122 SmallPtrSet<AllocaInst*, 16> UsedAllocas; 123 124 // When processing our SCC, check to see if CS was inlined from some other 125 // call site. For example, if we're processing "A" in this code: 126 // A() { B() } 127 // B() { x = alloca ... C() } 128 // C() { y = alloca ... } 129 // Assume that C was not inlined into B initially, and so we're processing A 130 // and decide to inline B into A. Doing this makes an alloca available for 131 // reuse and makes a callsite (C) available for inlining. When we process 132 // the C call site we don't want to do any alloca merging between X and Y 133 // because their scopes are not disjoint. We could make this smarter by 134 // keeping track of the inline history for each alloca in the 135 // InlinedArrayAllocas but this isn't likely to be a significant win. 136 if (InlineHistory != -1) // Only do merging for top-level call sites in SCC. 137 return true; 138 139 // Loop over all the allocas we have so far and see if they can be merged with 140 // a previously inlined alloca. If not, remember that we had it. 141 for (unsigned AllocaNo = 0, e = IFI.StaticAllocas.size(); 142 AllocaNo != e; ++AllocaNo) { 143 AllocaInst *AI = IFI.StaticAllocas[AllocaNo]; 144 145 // Don't bother trying to merge array allocations (they will usually be 146 // canonicalized to be an allocation *of* an array), or allocations whose 147 // type is not itself an array (because we're afraid of pessimizing SRoA). 148 ArrayType *ATy = dyn_cast<ArrayType>(AI->getAllocatedType()); 149 if (ATy == 0 || AI->isArrayAllocation()) 150 continue; 151 152 // Get the list of all available allocas for this array type. 153 std::vector<AllocaInst*> &AllocasForType = InlinedArrayAllocas[ATy]; 154 155 // Loop over the allocas in AllocasForType to see if we can reuse one. Note 156 // that we have to be careful not to reuse the same "available" alloca for 157 // multiple different allocas that we just inlined, we use the 'UsedAllocas' 158 // set to keep track of which "available" allocas are being used by this 159 // function. Also, AllocasForType can be empty of course! 160 bool MergedAwayAlloca = false; 161 for (unsigned i = 0, e = AllocasForType.size(); i != e; ++i) { 162 AllocaInst *AvailableAlloca = AllocasForType[i]; 163 164 // The available alloca has to be in the right function, not in some other 165 // function in this SCC. 166 if (AvailableAlloca->getParent() != AI->getParent()) 167 continue; 168 169 // If the inlined function already uses this alloca then we can't reuse 170 // it. 171 if (!UsedAllocas.insert(AvailableAlloca)) 172 continue; 173 174 // Otherwise, we *can* reuse it, RAUW AI into AvailableAlloca and declare 175 // success! 176 DEBUG(dbgs() << " ***MERGED ALLOCA: " << *AI << "\n\t\tINTO: " 177 << *AvailableAlloca << '\n'); 178 179 AI->replaceAllUsesWith(AvailableAlloca); 180 AI->eraseFromParent(); 181 MergedAwayAlloca = true; 182 ++NumMergedAllocas; 183 IFI.StaticAllocas[AllocaNo] = 0; 184 break; 185 } 186 187 // If we already nuked the alloca, we're done with it. 188 if (MergedAwayAlloca) 189 continue; 190 191 // If we were unable to merge away the alloca either because there are no 192 // allocas of the right type available or because we reused them all 193 // already, remember that this alloca came from an inlined function and mark 194 // it used so we don't reuse it for other allocas from this inline 195 // operation. 196 AllocasForType.push_back(AI); 197 UsedAllocas.insert(AI); 198 } 199 200 return true; 201} 202 203unsigned Inliner::getInlineThreshold(CallSite CS) const { 204 int thres = InlineThreshold; 205 206 // Listen to optsize when -inline-limit is not given. 207 Function *Caller = CS.getCaller(); 208 if (Caller && !Caller->isDeclaration() && 209 Caller->hasFnAttr(Attribute::OptimizeForSize) && 210 InlineLimit.getNumOccurrences() == 0) 211 thres = OptSizeThreshold; 212 213 // Listen to inlinehint when it would increase the threshold. 214 Function *Callee = CS.getCalledFunction(); 215 if (HintThreshold > thres && Callee && !Callee->isDeclaration() && 216 Callee->hasFnAttr(Attribute::InlineHint)) 217 thres = HintThreshold; 218 219 return thres; 220} 221 222/// shouldInline - Return true if the inliner should attempt to inline 223/// at the given CallSite. 224bool Inliner::shouldInline(CallSite CS) { 225 InlineCost IC = getInlineCost(CS); 226 227 if (IC.isAlways()) { 228 DEBUG(dbgs() << " Inlining: cost=always" 229 << ", Call: " << *CS.getInstruction() << "\n"); 230 return true; 231 } 232 233 if (IC.isNever()) { 234 DEBUG(dbgs() << " NOT Inlining: cost=never" 235 << ", Call: " << *CS.getInstruction() << "\n"); 236 return false; 237 } 238 239 Function *Caller = CS.getCaller(); 240 if (!IC) { 241 DEBUG(dbgs() << " NOT Inlining: cost=" << IC.getCost() 242 << ", thres=" << (IC.getCostDelta() + IC.getCost()) 243 << ", Call: " << *CS.getInstruction() << "\n"); 244 return false; 245 } 246 247 // Try to detect the case where the current inlining candidate caller (call 248 // it B) is a static or linkonce-ODR function and is an inlining candidate 249 // elsewhere, and the current candidate callee (call it C) is large enough 250 // that inlining it into B would make B too big to inline later. In these 251 // circumstances it may be best not to inline C into B, but to inline B into 252 // its callers. 253 // 254 // This only applies to static and linkonce-ODR functions because those are 255 // expected to be available for inlining in the translation units where they 256 // are used. Thus we will always have the opportunity to make local inlining 257 // decisions. Importantly the linkonce-ODR linkage covers inline functions 258 // and templates in C++. 259 // 260 // FIXME: All of this logic should be sunk into getInlineCost. It relies on 261 // the internal implementation of the inline cost metrics rather than 262 // treating them as truly abstract units etc. 263 if (Caller->hasLocalLinkage() || 264 Caller->getLinkage() == GlobalValue::LinkOnceODRLinkage) { 265 int TotalSecondaryCost = 0; 266 // The candidate cost to be imposed upon the current function. 267 int CandidateCost = IC.getCost() - (InlineConstants::CallPenalty + 1); 268 // This bool tracks what happens if we do NOT inline C into B. 269 bool callerWillBeRemoved = Caller->hasLocalLinkage(); 270 // This bool tracks what happens if we DO inline C into B. 271 bool inliningPreventsSomeOuterInline = false; 272 for (Value::use_iterator I = Caller->use_begin(), E =Caller->use_end(); 273 I != E; ++I) { 274 CallSite CS2(*I); 275 276 // If this isn't a call to Caller (it could be some other sort 277 // of reference) skip it. Such references will prevent the caller 278 // from being removed. 279 if (!CS2 || CS2.getCalledFunction() != Caller) { 280 callerWillBeRemoved = false; 281 continue; 282 } 283 284 InlineCost IC2 = getInlineCost(CS2); 285 ++NumCallerCallersAnalyzed; 286 if (!IC2) { 287 callerWillBeRemoved = false; 288 continue; 289 } 290 if (IC2.isAlways()) 291 continue; 292 293 // See if inlining or original callsite would erase the cost delta of 294 // this callsite. We subtract off the penalty for the call instruction, 295 // which we would be deleting. 296 if (IC2.getCostDelta() <= CandidateCost) { 297 inliningPreventsSomeOuterInline = true; 298 TotalSecondaryCost += IC2.getCost(); 299 } 300 } 301 // If all outer calls to Caller would get inlined, the cost for the last 302 // one is set very low by getInlineCost, in anticipation that Caller will 303 // be removed entirely. We did not account for this above unless there 304 // is only one caller of Caller. 305 if (callerWillBeRemoved && Caller->use_begin() != Caller->use_end()) 306 TotalSecondaryCost += InlineConstants::LastCallToStaticBonus; 307 308 if (inliningPreventsSomeOuterInline && TotalSecondaryCost < IC.getCost()) { 309 DEBUG(dbgs() << " NOT Inlining: " << *CS.getInstruction() << 310 " Cost = " << IC.getCost() << 311 ", outer Cost = " << TotalSecondaryCost << '\n'); 312 return false; 313 } 314 } 315 316 DEBUG(dbgs() << " Inlining: cost=" << IC.getCost() 317 << ", thres=" << (IC.getCostDelta() + IC.getCost()) 318 << ", Call: " << *CS.getInstruction() << '\n'); 319 return true; 320} 321 322/// InlineHistoryIncludes - Return true if the specified inline history ID 323/// indicates an inline history that includes the specified function. 324static bool InlineHistoryIncludes(Function *F, int InlineHistoryID, 325 const SmallVectorImpl<std::pair<Function*, int> > &InlineHistory) { 326 while (InlineHistoryID != -1) { 327 assert(unsigned(InlineHistoryID) < InlineHistory.size() && 328 "Invalid inline history ID"); 329 if (InlineHistory[InlineHistoryID].first == F) 330 return true; 331 InlineHistoryID = InlineHistory[InlineHistoryID].second; 332 } 333 return false; 334} 335 336bool Inliner::runOnSCC(CallGraphSCC &SCC) { 337 CallGraph &CG = getAnalysis<CallGraph>(); 338 const TargetData *TD = getAnalysisIfAvailable<TargetData>(); 339 340 SmallPtrSet<Function*, 8> SCCFunctions; 341 DEBUG(dbgs() << "Inliner visiting SCC:"); 342 for (CallGraphSCC::iterator I = SCC.begin(), E = SCC.end(); I != E; ++I) { 343 Function *F = (*I)->getFunction(); 344 if (F) SCCFunctions.insert(F); 345 DEBUG(dbgs() << " " << (F ? F->getName() : "INDIRECTNODE")); 346 } 347 348 // Scan through and identify all call sites ahead of time so that we only 349 // inline call sites in the original functions, not call sites that result 350 // from inlining other functions. 351 SmallVector<std::pair<CallSite, int>, 16> CallSites; 352 353 // When inlining a callee produces new call sites, we want to keep track of 354 // the fact that they were inlined from the callee. This allows us to avoid 355 // infinite inlining in some obscure cases. To represent this, we use an 356 // index into the InlineHistory vector. 357 SmallVector<std::pair<Function*, int>, 8> InlineHistory; 358 359 for (CallGraphSCC::iterator I = SCC.begin(), E = SCC.end(); I != E; ++I) { 360 Function *F = (*I)->getFunction(); 361 if (!F) continue; 362 363 for (Function::iterator BB = F->begin(), E = F->end(); BB != E; ++BB) 364 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) { 365 CallSite CS(cast<Value>(I)); 366 // If this isn't a call, or it is a call to an intrinsic, it can 367 // never be inlined. 368 if (!CS || isa<IntrinsicInst>(I)) 369 continue; 370 371 // If this is a direct call to an external function, we can never inline 372 // it. If it is an indirect call, inlining may resolve it to be a 373 // direct call, so we keep it. 374 if (CS.getCalledFunction() && CS.getCalledFunction()->isDeclaration()) 375 continue; 376 377 CallSites.push_back(std::make_pair(CS, -1)); 378 } 379 } 380 381 DEBUG(dbgs() << ": " << CallSites.size() << " call sites.\n"); 382 383 // If there are no calls in this function, exit early. 384 if (CallSites.empty()) 385 return false; 386 387 // Now that we have all of the call sites, move the ones to functions in the 388 // current SCC to the end of the list. 389 unsigned FirstCallInSCC = CallSites.size(); 390 for (unsigned i = 0; i < FirstCallInSCC; ++i) 391 if (Function *F = CallSites[i].first.getCalledFunction()) 392 if (SCCFunctions.count(F)) 393 std::swap(CallSites[i--], CallSites[--FirstCallInSCC]); 394 395 396 InlinedArrayAllocasTy InlinedArrayAllocas; 397 InlineFunctionInfo InlineInfo(&CG, TD); 398 399 // Now that we have all of the call sites, loop over them and inline them if 400 // it looks profitable to do so. 401 bool Changed = false; 402 bool LocalChange; 403 do { 404 LocalChange = false; 405 // Iterate over the outer loop because inlining functions can cause indirect 406 // calls to become direct calls. 407 for (unsigned CSi = 0; CSi != CallSites.size(); ++CSi) { 408 CallSite CS = CallSites[CSi].first; 409 410 Function *Caller = CS.getCaller(); 411 Function *Callee = CS.getCalledFunction(); 412 413 // If this call site is dead and it is to a readonly function, we should 414 // just delete the call instead of trying to inline it, regardless of 415 // size. This happens because IPSCCP propagates the result out of the 416 // call and then we're left with the dead call. 417 if (isInstructionTriviallyDead(CS.getInstruction())) { 418 DEBUG(dbgs() << " -> Deleting dead call: " 419 << *CS.getInstruction() << "\n"); 420 // Update the call graph by deleting the edge from Callee to Caller. 421 CG[Caller]->removeCallEdgeFor(CS); 422 CS.getInstruction()->eraseFromParent(); 423 ++NumCallsDeleted; 424 } else { 425 // We can only inline direct calls to non-declarations. 426 if (Callee == 0 || Callee->isDeclaration()) continue; 427 428 // If this call site was obtained by inlining another function, verify 429 // that the include path for the function did not include the callee 430 // itself. If so, we'd be recursively inlining the same function, 431 // which would provide the same callsites, which would cause us to 432 // infinitely inline. 433 int InlineHistoryID = CallSites[CSi].second; 434 if (InlineHistoryID != -1 && 435 InlineHistoryIncludes(Callee, InlineHistoryID, InlineHistory)) 436 continue; 437 438 439 // If the policy determines that we should inline this function, 440 // try to do so. 441 if (!shouldInline(CS)) 442 continue; 443 444 // Attempt to inline the function. 445 if (!InlineCallIfPossible(CS, InlineInfo, InlinedArrayAllocas, 446 InlineHistoryID, InsertLifetime)) 447 continue; 448 ++NumInlined; 449 450 // If inlining this function gave us any new call sites, throw them 451 // onto our worklist to process. They are useful inline candidates. 452 if (!InlineInfo.InlinedCalls.empty()) { 453 // Create a new inline history entry for this, so that we remember 454 // that these new callsites came about due to inlining Callee. 455 int NewHistoryID = InlineHistory.size(); 456 InlineHistory.push_back(std::make_pair(Callee, InlineHistoryID)); 457 458 for (unsigned i = 0, e = InlineInfo.InlinedCalls.size(); 459 i != e; ++i) { 460 Value *Ptr = InlineInfo.InlinedCalls[i]; 461 CallSites.push_back(std::make_pair(CallSite(Ptr), NewHistoryID)); 462 } 463 } 464 } 465 466 // If we inlined or deleted the last possible call site to the function, 467 // delete the function body now. 468 if (Callee && Callee->use_empty() && Callee->hasLocalLinkage() && 469 // TODO: Can remove if in SCC now. 470 !SCCFunctions.count(Callee) && 471 472 // The function may be apparently dead, but if there are indirect 473 // callgraph references to the node, we cannot delete it yet, this 474 // could invalidate the CGSCC iterator. 475 CG[Callee]->getNumReferences() == 0) { 476 DEBUG(dbgs() << " -> Deleting dead function: " 477 << Callee->getName() << "\n"); 478 CallGraphNode *CalleeNode = CG[Callee]; 479 480 // Remove any call graph edges from the callee to its callees. 481 CalleeNode->removeAllCalledFunctions(); 482 483 // Removing the node for callee from the call graph and delete it. 484 delete CG.removeFunctionFromModule(CalleeNode); 485 ++NumDeleted; 486 } 487 488 // Remove this call site from the list. If possible, use 489 // swap/pop_back for efficiency, but do not use it if doing so would 490 // move a call site to a function in this SCC before the 491 // 'FirstCallInSCC' barrier. 492 if (SCC.isSingular()) { 493 CallSites[CSi] = CallSites.back(); 494 CallSites.pop_back(); 495 } else { 496 CallSites.erase(CallSites.begin()+CSi); 497 } 498 --CSi; 499 500 Changed = true; 501 LocalChange = true; 502 } 503 } while (LocalChange); 504 505 return Changed; 506} 507 508// doFinalization - Remove now-dead linkonce functions at the end of 509// processing to avoid breaking the SCC traversal. 510bool Inliner::doFinalization(CallGraph &CG) { 511 return removeDeadFunctions(CG); 512} 513 514/// removeDeadFunctions - Remove dead functions that are not included in 515/// DNR (Do Not Remove) list. 516bool Inliner::removeDeadFunctions(CallGraph &CG, bool AlwaysInlineOnly) { 517 SmallVector<CallGraphNode*, 16> FunctionsToRemove; 518 519 // Scan for all of the functions, looking for ones that should now be removed 520 // from the program. Insert the dead ones in the FunctionsToRemove set. 521 for (CallGraph::iterator I = CG.begin(), E = CG.end(); I != E; ++I) { 522 CallGraphNode *CGN = I->second; 523 Function *F = CGN->getFunction(); 524 if (!F || F->isDeclaration()) 525 continue; 526 527 // Handle the case when this function is called and we only want to care 528 // about always-inline functions. This is a bit of a hack to share code 529 // between here and the InlineAlways pass. 530 if (AlwaysInlineOnly && !F->hasFnAttr(Attribute::AlwaysInline)) 531 continue; 532 533 // If the only remaining users of the function are dead constants, remove 534 // them. 535 F->removeDeadConstantUsers(); 536 537 if (!F->isDefTriviallyDead()) 538 continue; 539 540 // Remove any call graph edges from the function to its callees. 541 CGN->removeAllCalledFunctions(); 542 543 // Remove any edges from the external node to the function's call graph 544 // node. These edges might have been made irrelegant due to 545 // optimization of the program. 546 CG.getExternalCallingNode()->removeAnyCallEdgeTo(CGN); 547 548 // Removing the node for callee from the call graph and delete it. 549 FunctionsToRemove.push_back(CGN); 550 } 551 if (FunctionsToRemove.empty()) 552 return false; 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 order 559 // here to do this, it doesn't matter which order the functions are deleted 560 // in. 561 array_pod_sort(FunctionsToRemove.begin(), FunctionsToRemove.end()); 562 FunctionsToRemove.erase(std::unique(FunctionsToRemove.begin(), 563 FunctionsToRemove.end()), 564 FunctionsToRemove.end()); 565 for (SmallVectorImpl<CallGraphNode *>::iterator I = FunctionsToRemove.begin(), 566 E = FunctionsToRemove.end(); 567 I != E; ++I) { 568 delete CG.removeFunctionFromModule(*I); 569 ++NumDeleted; 570 } 571 return true; 572} 573