1 //===- Inliner.cpp - Code common to all inliners --------------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
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.
14 //===----------------------------------------------------------------------===//
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"
35 STATISTIC(NumInlined, "Number of functions inlined");
36 STATISTIC(NumCallsDeleted, "Number of call sites deleted, not inlined");
37 STATISTIC(NumDeleted, "Number of functions deleted because all callers found");
38 STATISTIC(NumMergedAllocas, "Number of allocas merged together");
41 InlineLimit("inline-threshold", cl::Hidden, cl::init(225), cl::ZeroOrMore,
42 cl::desc("Control the amount of inlining to perform (default = 225)"));
45 HintThreshold("inlinehint-threshold", cl::Hidden, cl::init(325),
46 cl::desc("Threshold for inlining functions with inline hint"));
48 // Threshold to use when optsize is specified (and there is no -inline-limit).
49 const int OptSizeThreshold = 75;
51 Inliner::Inliner(void *ID)
52 : CallGraphSCCPass(ID), InlineThreshold(InlineLimit) {}
54 Inliner::Inliner(void *ID, int Threshold)
55 : CallGraphSCCPass(ID), InlineThreshold(Threshold) {}
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);
65 typedef DenseMap<const ArrayType*, std::vector<AllocaInst*> >
66 InlinedArrayAllocasTy;
68 /// InlineCallIfPossible - If it is possible to inline the specified call site,
69 /// do so and update the CallGraph for this operation.
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, CallGraph &CG,
78 InlinedArrayAllocasTy &InlinedArrayAllocas) {
79 Function *Callee = CS.getCalledFunction();
80 Function *Caller = CS.getCaller();
82 // Try to inline the function. Get the list of static allocas that were
84 SmallVector<AllocaInst*, 16> StaticAllocas;
85 if (!InlineFunction(CS, &CG, TD, &StaticAllocas))
88 // If the inlined function had a higher stack protection level than the
89 // calling function, then bump up the caller's stack protection level.
90 if (Callee->hasFnAttr(Attribute::StackProtectReq))
91 Caller->addFnAttr(Attribute::StackProtectReq);
92 else if (Callee->hasFnAttr(Attribute::StackProtect) &&
93 !Caller->hasFnAttr(Attribute::StackProtectReq))
94 Caller->addFnAttr(Attribute::StackProtect);
97 // Look at all of the allocas that we inlined through this call site. If we
98 // have already inlined other allocas through other calls into this function,
99 // then we know that they have disjoint lifetimes and that we can merge them.
101 // There are many heuristics possible for merging these allocas, and the
102 // different options have different tradeoffs. One thing that we *really*
103 // don't want to hurt is SRoA: once inlining happens, often allocas are no
104 // longer address taken and so they can be promoted.
106 // Our "solution" for that is to only merge allocas whose outermost type is an
107 // array type. These are usually not promoted because someone is using a
108 // variable index into them. These are also often the most important ones to
111 // A better solution would be to have real memory lifetime markers in the IR
112 // and not have the inliner do any merging of allocas at all. This would
113 // allow the backend to do proper stack slot coloring of all allocas that
114 // *actually make it to the backend*, which is really what we want.
116 // Because we don't have this information, we do this simple and useful hack.
118 SmallPtrSet<AllocaInst*, 16> UsedAllocas;
120 // Loop over all the allocas we have so far and see if they can be merged with
121 // a previously inlined alloca. If not, remember that we had it.
122 for (unsigned AllocaNo = 0, e = StaticAllocas.size();
123 AllocaNo != e; ++AllocaNo) {
124 AllocaInst *AI = StaticAllocas[AllocaNo];
126 // Don't bother trying to merge array allocations (they will usually be
127 // canonicalized to be an allocation *of* an array), or allocations whose
128 // type is not itself an array (because we're afraid of pessimizing SRoA).
129 const ArrayType *ATy = dyn_cast<ArrayType>(AI->getAllocatedType());
130 if (ATy == 0 || AI->isArrayAllocation())
133 // Get the list of all available allocas for this array type.
134 std::vector<AllocaInst*> &AllocasForType = InlinedArrayAllocas[ATy];
136 // Loop over the allocas in AllocasForType to see if we can reuse one. Note
137 // that we have to be careful not to reuse the same "available" alloca for
138 // multiple different allocas that we just inlined, we use the 'UsedAllocas'
139 // set to keep track of which "available" allocas are being used by this
140 // function. Also, AllocasForType can be empty of course!
141 bool MergedAwayAlloca = false;
142 for (unsigned i = 0, e = AllocasForType.size(); i != e; ++i) {
143 AllocaInst *AvailableAlloca = AllocasForType[i];
145 // The available alloca has to be in the right function, not in some other
146 // function in this SCC.
147 if (AvailableAlloca->getParent() != AI->getParent())
150 // If the inlined function already uses this alloca then we can't reuse
152 if (!UsedAllocas.insert(AvailableAlloca))
155 // Otherwise, we *can* reuse it, RAUW AI into AvailableAlloca and declare
157 DEBUG(dbgs() << " ***MERGED ALLOCA: " << *AI);
159 AI->replaceAllUsesWith(AvailableAlloca);
160 AI->eraseFromParent();
161 MergedAwayAlloca = true;
166 // If we already nuked the alloca, we're done with it.
167 if (MergedAwayAlloca)
170 // If we were unable to merge away the alloca either because there are no
171 // allocas of the right type available or because we reused them all
172 // already, remember that this alloca came from an inlined function and mark
173 // it used so we don't reuse it for other allocas from this inline
175 AllocasForType.push_back(AI);
176 UsedAllocas.insert(AI);
182 unsigned Inliner::getInlineThreshold(CallSite CS) const {
183 int thres = InlineThreshold;
185 // Listen to optsize when -inline-limit is not given.
186 Function *Caller = CS.getCaller();
187 if (Caller && !Caller->isDeclaration() &&
188 Caller->hasFnAttr(Attribute::OptimizeForSize) &&
189 InlineLimit.getNumOccurrences() == 0)
190 thres = OptSizeThreshold;
192 // Listen to inlinehint when it would increase the threshold.
193 Function *Callee = CS.getCalledFunction();
194 if (HintThreshold > thres && Callee && !Callee->isDeclaration() &&
195 Callee->hasFnAttr(Attribute::InlineHint))
196 thres = HintThreshold;
201 /// shouldInline - Return true if the inliner should attempt to inline
202 /// at the given CallSite.
203 bool Inliner::shouldInline(CallSite CS) {
204 InlineCost IC = getInlineCost(CS);
207 DEBUG(dbgs() << " Inlining: cost=always"
208 << ", Call: " << *CS.getInstruction() << "\n");
213 DEBUG(dbgs() << " NOT Inlining: cost=never"
214 << ", Call: " << *CS.getInstruction() << "\n");
218 int Cost = IC.getValue();
219 Function *Caller = CS.getCaller();
220 int CurrentThreshold = getInlineThreshold(CS);
221 float FudgeFactor = getInlineFudgeFactor(CS);
222 if (Cost >= (int)(CurrentThreshold * FudgeFactor)) {
223 DEBUG(dbgs() << " NOT Inlining: cost=" << Cost
224 << ", Call: " << *CS.getInstruction() << "\n");
228 // Try to detect the case where the current inlining candidate caller
229 // (call it B) is a static function and is an inlining candidate elsewhere,
230 // and the current candidate callee (call it C) is large enough that
231 // inlining it into B would make B too big to inline later. In these
232 // circumstances it may be best not to inline C into B, but to inline B
234 if (Caller->hasLocalLinkage()) {
235 int TotalSecondaryCost = 0;
236 bool outerCallsFound = false;
237 bool allOuterCallsWillBeInlined = true;
238 bool someOuterCallWouldNotBeInlined = false;
239 for (Value::use_iterator I = Caller->use_begin(), E =Caller->use_end();
241 CallSite CS2 = CallSite::get(*I);
243 // If this isn't a call to Caller (it could be some other sort
244 // of reference) skip it.
245 if (CS2.getInstruction() == 0 || CS2.getCalledFunction() != Caller)
248 InlineCost IC2 = getInlineCost(CS2);
250 allOuterCallsWillBeInlined = false;
251 if (IC2.isAlways() || IC2.isNever())
254 outerCallsFound = true;
255 int Cost2 = IC2.getValue();
256 int CurrentThreshold2 = getInlineThreshold(CS2);
257 float FudgeFactor2 = getInlineFudgeFactor(CS2);
259 if (Cost2 >= (int)(CurrentThreshold2 * FudgeFactor2))
260 allOuterCallsWillBeInlined = false;
262 // See if we have this case. We subtract off the penalty
263 // for the call instruction, which we would be deleting.
264 if (Cost2 < (int)(CurrentThreshold2 * FudgeFactor2) &&
265 Cost2 + Cost - (InlineConstants::CallPenalty + 1) >=
266 (int)(CurrentThreshold2 * FudgeFactor2)) {
267 someOuterCallWouldNotBeInlined = true;
268 TotalSecondaryCost += Cost2;
271 // If all outer calls to Caller would get inlined, the cost for the last
272 // one is set very low by getInlineCost, in anticipation that Caller will
273 // be removed entirely. We did not account for this above unless there
274 // is only one caller of Caller.
275 if (allOuterCallsWillBeInlined && Caller->use_begin() != Caller->use_end())
276 TotalSecondaryCost += InlineConstants::LastCallToStaticBonus;
278 if (outerCallsFound && someOuterCallWouldNotBeInlined &&
279 TotalSecondaryCost < Cost) {
280 DEBUG(dbgs() << " NOT Inlining: " << *CS.getInstruction() <<
281 " Cost = " << Cost <<
282 ", outer Cost = " << TotalSecondaryCost << '\n');
287 DEBUG(dbgs() << " Inlining: cost=" << Cost
288 << ", Call: " << *CS.getInstruction() << '\n');
292 bool Inliner::runOnSCC(std::vector<CallGraphNode*> &SCC) {
293 CallGraph &CG = getAnalysis<CallGraph>();
294 const TargetData *TD = getAnalysisIfAvailable<TargetData>();
296 SmallPtrSet<Function*, 8> SCCFunctions;
297 DEBUG(dbgs() << "Inliner visiting SCC:");
298 for (unsigned i = 0, e = SCC.size(); i != e; ++i) {
299 Function *F = SCC[i]->getFunction();
300 if (F) SCCFunctions.insert(F);
301 DEBUG(dbgs() << " " << (F ? F->getName() : "INDIRECTNODE"));
304 // Scan through and identify all call sites ahead of time so that we only
305 // inline call sites in the original functions, not call sites that result
306 // from inlining other functions.
307 SmallVector<CallSite, 16> CallSites;
309 for (unsigned i = 0, e = SCC.size(); i != e; ++i) {
310 Function *F = SCC[i]->getFunction();
313 for (Function::iterator BB = F->begin(), E = F->end(); BB != E; ++BB)
314 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
315 CallSite CS = CallSite::get(I);
316 // If this isn't a call, or it is a call to an intrinsic, it can
318 if (CS.getInstruction() == 0 || isa<IntrinsicInst>(I))
321 // If this is a direct call to an external function, we can never inline
322 // it. If it is an indirect call, inlining may resolve it to be a
323 // direct call, so we keep it.
324 if (CS.getCalledFunction() && CS.getCalledFunction()->isDeclaration())
327 CallSites.push_back(CS);
331 DEBUG(dbgs() << ": " << CallSites.size() << " call sites.\n");
333 // Now that we have all of the call sites, move the ones to functions in the
334 // current SCC to the end of the list.
335 unsigned FirstCallInSCC = CallSites.size();
336 for (unsigned i = 0; i < FirstCallInSCC; ++i)
337 if (Function *F = CallSites[i].getCalledFunction())
338 if (SCCFunctions.count(F))
339 std::swap(CallSites[i--], CallSites[--FirstCallInSCC]);
342 InlinedArrayAllocasTy InlinedArrayAllocas;
344 // Now that we have all of the call sites, loop over them and inline them if
345 // it looks profitable to do so.
346 bool Changed = false;
350 // Iterate over the outer loop because inlining functions can cause indirect
351 // calls to become direct calls.
352 for (unsigned CSi = 0; CSi != CallSites.size(); ++CSi) {
353 CallSite CS = CallSites[CSi];
355 Function *Caller = CS.getCaller();
356 Function *Callee = CS.getCalledFunction();
358 // If this call site is dead and it is to a readonly function, we should
359 // just delete the call instead of trying to inline it, regardless of
360 // size. This happens because IPSCCP propagates the result out of the
361 // call and then we're left with the dead call.
362 if (isInstructionTriviallyDead(CS.getInstruction())) {
363 DEBUG(dbgs() << " -> Deleting dead call: "
364 << *CS.getInstruction() << "\n");
365 // Update the call graph by deleting the edge from Callee to Caller.
366 CG[Caller]->removeCallEdgeFor(CS);
367 CS.getInstruction()->eraseFromParent();
370 // We can only inline direct calls to non-declarations.
371 if (Callee == 0 || Callee->isDeclaration()) continue;
373 // If the policy determines that we should inline this function,
375 if (!shouldInline(CS))
378 // Attempt to inline the function...
379 if (!InlineCallIfPossible(CS, CG, TD, InlinedArrayAllocas))
384 // If we inlined or deleted the last possible call site to the function,
385 // delete the function body now.
386 if (Callee && Callee->use_empty() && Callee->hasLocalLinkage() &&
387 // TODO: Can remove if in SCC now.
388 !SCCFunctions.count(Callee) &&
390 // The function may be apparently dead, but if there are indirect
391 // callgraph references to the node, we cannot delete it yet, this
392 // could invalidate the CGSCC iterator.
393 CG[Callee]->getNumReferences() == 0) {
394 DEBUG(dbgs() << " -> Deleting dead function: "
395 << Callee->getName() << "\n");
396 CallGraphNode *CalleeNode = CG[Callee];
398 // Remove any call graph edges from the callee to its callees.
399 CalleeNode->removeAllCalledFunctions();
401 resetCachedCostInfo(Callee);
403 // Removing the node for callee from the call graph and delete it.
404 delete CG.removeFunctionFromModule(CalleeNode);
408 // Remove any cached cost info for this caller, as inlining the
409 // callee has increased the size of the caller (which may be the
410 // same as the callee).
411 resetCachedCostInfo(Caller);
413 // Remove this call site from the list. If possible, use
414 // swap/pop_back for efficiency, but do not use it if doing so would
415 // move a call site to a function in this SCC before the
416 // 'FirstCallInSCC' barrier.
417 if (SCC.size() == 1) {
418 std::swap(CallSites[CSi], CallSites.back());
419 CallSites.pop_back();
421 CallSites.erase(CallSites.begin()+CSi);
428 } while (LocalChange);
433 // doFinalization - Remove now-dead linkonce functions at the end of
434 // processing to avoid breaking the SCC traversal.
435 bool Inliner::doFinalization(CallGraph &CG) {
436 return removeDeadFunctions(CG);
439 /// removeDeadFunctions - Remove dead functions that are not included in
440 /// DNR (Do Not Remove) list.
441 bool Inliner::removeDeadFunctions(CallGraph &CG,
442 SmallPtrSet<const Function *, 16> *DNR) {
443 SmallPtrSet<CallGraphNode*, 16> FunctionsToRemove;
445 // Scan for all of the functions, looking for ones that should now be removed
446 // from the program. Insert the dead ones in the FunctionsToRemove set.
447 for (CallGraph::iterator I = CG.begin(), E = CG.end(); I != E; ++I) {
448 CallGraphNode *CGN = I->second;
449 if (CGN->getFunction() == 0)
452 Function *F = CGN->getFunction();
454 // If the only remaining users of the function are dead constants, remove
456 F->removeDeadConstantUsers();
458 if (DNR && DNR->count(F))
460 if (!F->hasLinkOnceLinkage() && !F->hasLocalLinkage() &&
461 !F->hasAvailableExternallyLinkage())
466 // Remove any call graph edges from the function to its callees.
467 CGN->removeAllCalledFunctions();
469 // Remove any edges from the external node to the function's call graph
470 // node. These edges might have been made irrelegant due to
471 // optimization of the program.
472 CG.getExternalCallingNode()->removeAnyCallEdgeTo(CGN);
474 // Removing the node for callee from the call graph and delete it.
475 FunctionsToRemove.insert(CGN);
478 // Now that we know which functions to delete, do so. We didn't want to do
479 // this inline, because that would invalidate our CallGraph::iterator
482 // Note that it doesn't matter that we are iterating over a non-stable set
483 // here to do this, it doesn't matter which order the functions are deleted
485 bool Changed = false;
486 for (SmallPtrSet<CallGraphNode*, 16>::iterator I = FunctionsToRemove.begin(),
487 E = FunctionsToRemove.end(); I != E; ++I) {
488 resetCachedCostInfo((*I)->getFunction());
489 delete CG.removeFunctionFromModule(*I);