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(200), cl::ZeroOrMore,
42 cl::desc("Control the amount of inlining to perform (default = 200)"));
44 Inliner::Inliner(void *ID)
45 : CallGraphSCCPass(ID), InlineThreshold(InlineLimit) {}
47 Inliner::Inliner(void *ID, int Threshold)
48 : CallGraphSCCPass(ID), InlineThreshold(Threshold) {}
50 /// getAnalysisUsage - For this class, we declare that we require and preserve
51 /// the call graph. If the derived class implements this method, it should
52 /// always explicitly call the implementation here.
53 void Inliner::getAnalysisUsage(AnalysisUsage &Info) const {
54 CallGraphSCCPass::getAnalysisUsage(Info);
58 typedef DenseMap<const ArrayType*, std::vector<AllocaInst*> >
59 InlinedArrayAllocasTy;
61 /// InlineCallIfPossible - If it is possible to inline the specified call site,
62 /// do so and update the CallGraph for this operation.
64 /// This function also does some basic book-keeping to update the IR. The
65 /// InlinedArrayAllocas map keeps track of any allocas that are already
66 /// available from other functions inlined into the caller. If we are able to
67 /// inline this call site we attempt to reuse already available allocas or add
68 /// any new allocas to the set if not possible.
69 static bool InlineCallIfPossible(CallSite CS, CallGraph &CG,
71 InlinedArrayAllocasTy &InlinedArrayAllocas) {
72 Function *Callee = CS.getCalledFunction();
73 Function *Caller = CS.getCaller();
75 // Try to inline the function. Get the list of static allocas that were
77 SmallVector<AllocaInst*, 16> StaticAllocas;
78 if (!InlineFunction(CS, &CG, TD, &StaticAllocas))
81 // If the inlined function had a higher stack protection level than the
82 // calling function, then bump up the caller's stack protection level.
83 if (Callee->hasFnAttr(Attribute::StackProtectReq))
84 Caller->addFnAttr(Attribute::StackProtectReq);
85 else if (Callee->hasFnAttr(Attribute::StackProtect) &&
86 !Caller->hasFnAttr(Attribute::StackProtectReq))
87 Caller->addFnAttr(Attribute::StackProtect);
90 // Look at all of the allocas that we inlined through this call site. If we
91 // have already inlined other allocas through other calls into this function,
92 // then we know that they have disjoint lifetimes and that we can merge them.
94 // There are many heuristics possible for merging these allocas, and the
95 // different options have different tradeoffs. One thing that we *really*
96 // don't want to hurt is SRoA: once inlining happens, often allocas are no
97 // longer address taken and so they can be promoted.
99 // Our "solution" for that is to only merge allocas whose outermost type is an
100 // array type. These are usually not promoted because someone is using a
101 // variable index into them. These are also often the most important ones to
104 // A better solution would be to have real memory lifetime markers in the IR
105 // and not have the inliner do any merging of allocas at all. This would
106 // allow the backend to do proper stack slot coloring of all allocas that
107 // *actually make it to the backend*, which is really what we want.
109 // Because we don't have this information, we do this simple and useful hack.
111 SmallPtrSet<AllocaInst*, 16> UsedAllocas;
113 // Loop over all the allocas we have so far and see if they can be merged with
114 // a previously inlined alloca. If not, remember that we had it.
115 for (unsigned AllocaNo = 0, e = StaticAllocas.size();
116 AllocaNo != e; ++AllocaNo) {
117 AllocaInst *AI = StaticAllocas[AllocaNo];
119 // Don't bother trying to merge array allocations (they will usually be
120 // canonicalized to be an allocation *of* an array), or allocations whose
121 // type is not itself an array (because we're afraid of pessimizing SRoA).
122 const ArrayType *ATy = dyn_cast<ArrayType>(AI->getAllocatedType());
123 if (ATy == 0 || AI->isArrayAllocation())
126 // Get the list of all available allocas for this array type.
127 std::vector<AllocaInst*> &AllocasForType = InlinedArrayAllocas[ATy];
129 // Loop over the allocas in AllocasForType to see if we can reuse one. Note
130 // that we have to be careful not to reuse the same "available" alloca for
131 // multiple different allocas that we just inlined, we use the 'UsedAllocas'
132 // set to keep track of which "available" allocas are being used by this
133 // function. Also, AllocasForType can be empty of course!
134 bool MergedAwayAlloca = false;
135 for (unsigned i = 0, e = AllocasForType.size(); i != e; ++i) {
136 AllocaInst *AvailableAlloca = AllocasForType[i];
138 // The available alloca has to be in the right function, not in some other
139 // function in this SCC.
140 if (AvailableAlloca->getParent() != AI->getParent())
143 // If the inlined function already uses this alloca then we can't reuse
145 if (!UsedAllocas.insert(AvailableAlloca))
148 // Otherwise, we *can* reuse it, RAUW AI into AvailableAlloca and declare
150 DEBUG(dbgs() << " ***MERGED ALLOCA: " << *AI);
152 AI->replaceAllUsesWith(AvailableAlloca);
153 AI->eraseFromParent();
154 MergedAwayAlloca = true;
159 // If we already nuked the alloca, we're done with it.
160 if (MergedAwayAlloca)
163 // If we were unable to merge away the alloca either because there are no
164 // allocas of the right type available or because we reused them all
165 // already, remember that this alloca came from an inlined function and mark
166 // it used so we don't reuse it for other allocas from this inline
168 AllocasForType.push_back(AI);
169 UsedAllocas.insert(AI);
175 /// shouldInline - Return true if the inliner should attempt to inline
176 /// at the given CallSite.
177 bool Inliner::shouldInline(CallSite CS) {
178 InlineCost IC = getInlineCost(CS);
181 DEBUG(dbgs() << " Inlining: cost=always"
182 << ", Call: " << *CS.getInstruction() << "\n");
187 DEBUG(dbgs() << " NOT Inlining: cost=never"
188 << ", Call: " << *CS.getInstruction() << "\n");
192 int Cost = IC.getValue();
193 int CurrentThreshold = InlineThreshold;
194 Function *Caller = CS.getCaller();
195 if (Caller && !Caller->isDeclaration() &&
196 Caller->hasFnAttr(Attribute::OptimizeForSize) &&
197 InlineLimit.getNumOccurrences() == 0 &&
198 InlineThreshold != 50)
199 CurrentThreshold = 50;
201 float FudgeFactor = getInlineFudgeFactor(CS);
202 if (Cost >= (int)(CurrentThreshold * FudgeFactor)) {
203 DEBUG(dbgs() << " NOT Inlining: cost=" << Cost
204 << ", Call: " << *CS.getInstruction() << "\n");
208 // Try to detect the case where the current inlining candidate caller
209 // (call it B) is a static function and is an inlining candidate elsewhere,
210 // and the current candidate callee (call it C) is large enough that
211 // inlining it into B would make B too big to inline later. In these
212 // circumstances it may be best not to inline C into B, but to inline B
214 if (Caller->hasLocalLinkage()) {
215 int TotalSecondaryCost = 0;
216 bool outerCallsFound = false;
217 bool allOuterCallsWillBeInlined = true;
218 bool someOuterCallWouldNotBeInlined = false;
219 for (Value::use_iterator I = Caller->use_begin(), E =Caller->use_end();
221 CallSite CS2 = CallSite::get(*I);
223 // If this isn't a call to Caller (it could be some other sort
224 // of reference) skip it.
225 if (CS2.getInstruction() == 0 || CS2.getCalledFunction() != Caller)
228 InlineCost IC2 = getInlineCost(CS2);
230 allOuterCallsWillBeInlined = false;
231 if (IC2.isAlways() || IC2.isNever())
234 outerCallsFound = true;
235 int Cost2 = IC2.getValue();
236 int CurrentThreshold2 = InlineThreshold;
237 Function *Caller2 = CS2.getCaller();
238 if (Caller2 && !Caller2->isDeclaration() &&
239 Caller2->hasFnAttr(Attribute::OptimizeForSize) &&
240 InlineThreshold != 50)
241 CurrentThreshold2 = 50;
243 float FudgeFactor2 = getInlineFudgeFactor(CS2);
245 if (Cost2 >= (int)(CurrentThreshold2 * FudgeFactor2))
246 allOuterCallsWillBeInlined = false;
248 // See if we have this case. We subtract off the penalty
249 // for the call instruction, which we would be deleting.
250 if (Cost2 < (int)(CurrentThreshold2 * FudgeFactor2) &&
251 Cost2 + Cost - (InlineConstants::CallPenalty + 1) >=
252 (int)(CurrentThreshold2 * FudgeFactor2)) {
253 someOuterCallWouldNotBeInlined = true;
254 TotalSecondaryCost += Cost2;
257 // If all outer calls to Caller would get inlined, the cost for the last
258 // one is set very low by getInlineCost, in anticipation that Caller will
259 // be removed entirely. We did not account for this above unless there
260 // is only one caller of Caller.
261 if (allOuterCallsWillBeInlined && Caller->use_begin() != Caller->use_end())
262 TotalSecondaryCost += InlineConstants::LastCallToStaticBonus;
264 if (outerCallsFound && someOuterCallWouldNotBeInlined &&
265 TotalSecondaryCost < Cost) {
266 DEBUG(dbgs() << " NOT Inlining: " << *CS.getInstruction() <<
267 " Cost = " << Cost <<
268 ", outer Cost = " << TotalSecondaryCost << '\n');
273 DEBUG(dbgs() << " Inlining: cost=" << Cost
274 << ", Call: " << *CS.getInstruction() << '\n');
278 bool Inliner::runOnSCC(std::vector<CallGraphNode*> &SCC) {
279 CallGraph &CG = getAnalysis<CallGraph>();
280 const TargetData *TD = getAnalysisIfAvailable<TargetData>();
282 SmallPtrSet<Function*, 8> SCCFunctions;
283 DEBUG(dbgs() << "Inliner visiting SCC:");
284 for (unsigned i = 0, e = SCC.size(); i != e; ++i) {
285 Function *F = SCC[i]->getFunction();
286 if (F) SCCFunctions.insert(F);
287 DEBUG(dbgs() << " " << (F ? F->getName() : "INDIRECTNODE"));
290 // Scan through and identify all call sites ahead of time so that we only
291 // inline call sites in the original functions, not call sites that result
292 // from inlining other functions.
293 SmallVector<CallSite, 16> CallSites;
295 for (unsigned i = 0, e = SCC.size(); i != e; ++i) {
296 Function *F = SCC[i]->getFunction();
299 for (Function::iterator BB = F->begin(), E = F->end(); BB != E; ++BB)
300 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
301 CallSite CS = CallSite::get(I);
302 // If this isn't a call, or it is a call to an intrinsic, it can
304 if (CS.getInstruction() == 0 || isa<IntrinsicInst>(I))
307 // If this is a direct call to an external function, we can never inline
308 // it. If it is an indirect call, inlining may resolve it to be a
309 // direct call, so we keep it.
310 if (CS.getCalledFunction() && CS.getCalledFunction()->isDeclaration())
313 CallSites.push_back(CS);
317 DEBUG(dbgs() << ": " << CallSites.size() << " call sites.\n");
319 // Now that we have all of the call sites, move the ones to functions in the
320 // current SCC to the end of the list.
321 unsigned FirstCallInSCC = CallSites.size();
322 for (unsigned i = 0; i < FirstCallInSCC; ++i)
323 if (Function *F = CallSites[i].getCalledFunction())
324 if (SCCFunctions.count(F))
325 std::swap(CallSites[i--], CallSites[--FirstCallInSCC]);
328 InlinedArrayAllocasTy InlinedArrayAllocas;
330 // Now that we have all of the call sites, loop over them and inline them if
331 // it looks profitable to do so.
332 bool Changed = false;
336 // Iterate over the outer loop because inlining functions can cause indirect
337 // calls to become direct calls.
338 for (unsigned CSi = 0; CSi != CallSites.size(); ++CSi) {
339 CallSite CS = CallSites[CSi];
341 Function *Caller = CS.getCaller();
342 Function *Callee = CS.getCalledFunction();
344 // If this call site is dead and it is to a readonly function, we should
345 // just delete the call instead of trying to inline it, regardless of
346 // size. This happens because IPSCCP propagates the result out of the
347 // call and then we're left with the dead call.
348 if (isInstructionTriviallyDead(CS.getInstruction())) {
349 DEBUG(dbgs() << " -> Deleting dead call: "
350 << *CS.getInstruction() << "\n");
351 // Update the call graph by deleting the edge from Callee to Caller.
352 CG[Caller]->removeCallEdgeFor(CS);
353 CS.getInstruction()->eraseFromParent();
356 // We can only inline direct calls to non-declarations.
357 if (Callee == 0 || Callee->isDeclaration()) continue;
359 // If the policy determines that we should inline this function,
361 if (!shouldInline(CS))
364 // Attempt to inline the function...
365 if (!InlineCallIfPossible(CS, CG, TD, InlinedArrayAllocas))
370 // If we inlined or deleted the last possible call site to the function,
371 // delete the function body now.
372 if (Callee && Callee->use_empty() && Callee->hasLocalLinkage() &&
373 // TODO: Can remove if in SCC now.
374 !SCCFunctions.count(Callee) &&
376 // The function may be apparently dead, but if there are indirect
377 // callgraph references to the node, we cannot delete it yet, this
378 // could invalidate the CGSCC iterator.
379 CG[Callee]->getNumReferences() == 0) {
380 DEBUG(dbgs() << " -> Deleting dead function: "
381 << Callee->getName() << "\n");
382 CallGraphNode *CalleeNode = CG[Callee];
384 // Remove any call graph edges from the callee to its callees.
385 CalleeNode->removeAllCalledFunctions();
387 resetCachedCostInfo(Callee);
389 // Removing the node for callee from the call graph and delete it.
390 delete CG.removeFunctionFromModule(CalleeNode);
394 // Remove any cached cost info for this caller, as inlining the
395 // callee has increased the size of the caller (which may be the
396 // same as the callee).
397 resetCachedCostInfo(Caller);
399 // Remove this call site from the list. If possible, use
400 // swap/pop_back for efficiency, but do not use it if doing so would
401 // move a call site to a function in this SCC before the
402 // 'FirstCallInSCC' barrier.
403 if (SCC.size() == 1) {
404 std::swap(CallSites[CSi], CallSites.back());
405 CallSites.pop_back();
407 CallSites.erase(CallSites.begin()+CSi);
414 } while (LocalChange);
419 // doFinalization - Remove now-dead linkonce functions at the end of
420 // processing to avoid breaking the SCC traversal.
421 bool Inliner::doFinalization(CallGraph &CG) {
422 return removeDeadFunctions(CG);
425 /// removeDeadFunctions - Remove dead functions that are not included in
426 /// DNR (Do Not Remove) list.
427 bool Inliner::removeDeadFunctions(CallGraph &CG,
428 SmallPtrSet<const Function *, 16> *DNR) {
429 SmallPtrSet<CallGraphNode*, 16> FunctionsToRemove;
431 // Scan for all of the functions, looking for ones that should now be removed
432 // from the program. Insert the dead ones in the FunctionsToRemove set.
433 for (CallGraph::iterator I = CG.begin(), E = CG.end(); I != E; ++I) {
434 CallGraphNode *CGN = I->second;
435 if (CGN->getFunction() == 0)
438 Function *F = CGN->getFunction();
440 // If the only remaining users of the function are dead constants, remove
442 F->removeDeadConstantUsers();
444 if (DNR && DNR->count(F))
446 if (!F->hasLinkOnceLinkage() && !F->hasLocalLinkage() &&
447 !F->hasAvailableExternallyLinkage())
452 // Remove any call graph edges from the function to its callees.
453 CGN->removeAllCalledFunctions();
455 // Remove any edges from the external node to the function's call graph
456 // node. These edges might have been made irrelegant due to
457 // optimization of the program.
458 CG.getExternalCallingNode()->removeAnyCallEdgeTo(CGN);
460 // Removing the node for callee from the call graph and delete it.
461 FunctionsToRemove.insert(CGN);
464 // Now that we know which functions to delete, do so. We didn't want to do
465 // this inline, because that would invalidate our CallGraph::iterator
468 // Note that it doesn't matter that we are iterating over a non-stable set
469 // here to do this, it doesn't matter which order the functions are deleted
471 bool Changed = false;
472 for (SmallPtrSet<CallGraphNode*, 16>::iterator I = FunctionsToRemove.begin(),
473 E = FunctionsToRemove.end(); I != E; ++I) {
474 resetCachedCostInfo((*I)->getFunction());
475 delete CG.removeFunctionFromModule(*I);