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 RespectHint("respect-inlinehint", cl::Hidden,
46 cl::desc("Respect the inlinehint attribute"));
48 // Threshold to use when inlinehint is given.
49 const int HintThreshold = 300;
51 // Threshold to use when optsize is specified (and there is no -inline-limit).
52 const int OptSizeThreshold = 75;
54 Inliner::Inliner(void *ID)
55 : CallGraphSCCPass(ID), InlineThreshold(InlineLimit) {}
57 Inliner::Inliner(void *ID, int Threshold)
58 : CallGraphSCCPass(ID), InlineThreshold(Threshold) {}
60 /// getAnalysisUsage - For this class, we declare that we require and preserve
61 /// the call graph. If the derived class implements this method, it should
62 /// always explicitly call the implementation here.
63 void Inliner::getAnalysisUsage(AnalysisUsage &Info) const {
64 CallGraphSCCPass::getAnalysisUsage(Info);
68 typedef DenseMap<const ArrayType*, std::vector<AllocaInst*> >
69 InlinedArrayAllocasTy;
71 /// InlineCallIfPossible - If it is possible to inline the specified call site,
72 /// do so and update the CallGraph for this operation.
74 /// This function also does some basic book-keeping to update the IR. The
75 /// InlinedArrayAllocas map keeps track of any allocas that are already
76 /// available from other functions inlined into the caller. If we are able to
77 /// inline this call site we attempt to reuse already available allocas or add
78 /// any new allocas to the set if not possible.
79 static bool InlineCallIfPossible(CallSite CS, CallGraph &CG,
81 InlinedArrayAllocasTy &InlinedArrayAllocas) {
82 Function *Callee = CS.getCalledFunction();
83 Function *Caller = CS.getCaller();
85 // Try to inline the function. Get the list of static allocas that were
87 SmallVector<AllocaInst*, 16> StaticAllocas;
88 if (!InlineFunction(CS, &CG, TD, &StaticAllocas))
91 // If the inlined function had a higher stack protection level than the
92 // calling function, then bump up the caller's stack protection level.
93 if (Callee->hasFnAttr(Attribute::StackProtectReq))
94 Caller->addFnAttr(Attribute::StackProtectReq);
95 else if (Callee->hasFnAttr(Attribute::StackProtect) &&
96 !Caller->hasFnAttr(Attribute::StackProtectReq))
97 Caller->addFnAttr(Attribute::StackProtect);
100 // Look at all of the allocas that we inlined through this call site. If we
101 // have already inlined other allocas through other calls into this function,
102 // then we know that they have disjoint lifetimes and that we can merge them.
104 // There are many heuristics possible for merging these allocas, and the
105 // different options have different tradeoffs. One thing that we *really*
106 // don't want to hurt is SRoA: once inlining happens, often allocas are no
107 // longer address taken and so they can be promoted.
109 // Our "solution" for that is to only merge allocas whose outermost type is an
110 // array type. These are usually not promoted because someone is using a
111 // variable index into them. These are also often the most important ones to
114 // A better solution would be to have real memory lifetime markers in the IR
115 // and not have the inliner do any merging of allocas at all. This would
116 // allow the backend to do proper stack slot coloring of all allocas that
117 // *actually make it to the backend*, which is really what we want.
119 // Because we don't have this information, we do this simple and useful hack.
121 SmallPtrSet<AllocaInst*, 16> UsedAllocas;
123 // Loop over all the allocas we have so far and see if they can be merged with
124 // a previously inlined alloca. If not, remember that we had it.
125 for (unsigned AllocaNo = 0, e = StaticAllocas.size();
126 AllocaNo != e; ++AllocaNo) {
127 AllocaInst *AI = StaticAllocas[AllocaNo];
129 // Don't bother trying to merge array allocations (they will usually be
130 // canonicalized to be an allocation *of* an array), or allocations whose
131 // type is not itself an array (because we're afraid of pessimizing SRoA).
132 const ArrayType *ATy = dyn_cast<ArrayType>(AI->getAllocatedType());
133 if (ATy == 0 || AI->isArrayAllocation())
136 // Get the list of all available allocas for this array type.
137 std::vector<AllocaInst*> &AllocasForType = InlinedArrayAllocas[ATy];
139 // Loop over the allocas in AllocasForType to see if we can reuse one. Note
140 // that we have to be careful not to reuse the same "available" alloca for
141 // multiple different allocas that we just inlined, we use the 'UsedAllocas'
142 // set to keep track of which "available" allocas are being used by this
143 // function. Also, AllocasForType can be empty of course!
144 bool MergedAwayAlloca = false;
145 for (unsigned i = 0, e = AllocasForType.size(); i != e; ++i) {
146 AllocaInst *AvailableAlloca = AllocasForType[i];
148 // The available alloca has to be in the right function, not in some other
149 // function in this SCC.
150 if (AvailableAlloca->getParent() != AI->getParent())
153 // If the inlined function already uses this alloca then we can't reuse
155 if (!UsedAllocas.insert(AvailableAlloca))
158 // Otherwise, we *can* reuse it, RAUW AI into AvailableAlloca and declare
160 DEBUG(dbgs() << " ***MERGED ALLOCA: " << *AI);
162 AI->replaceAllUsesWith(AvailableAlloca);
163 AI->eraseFromParent();
164 MergedAwayAlloca = true;
169 // If we already nuked the alloca, we're done with it.
170 if (MergedAwayAlloca)
173 // If we were unable to merge away the alloca either because there are no
174 // allocas of the right type available or because we reused them all
175 // already, remember that this alloca came from an inlined function and mark
176 // it used so we don't reuse it for other allocas from this inline
178 AllocasForType.push_back(AI);
179 UsedAllocas.insert(AI);
185 unsigned Inliner::getInlineThreshold(CallSite CS) const {
186 // Listen to inlinehint when -respect-inlinehint is given.
187 Function *Callee = CS.getCalledFunction();
188 if (RespectHint && Callee && !Callee->isDeclaration() &&
189 Callee->hasFnAttr(Attribute::InlineHint))
190 return HintThreshold;
192 // Listen to optsize when -inline-limit is not given.
193 Function *Caller = CS.getCaller();
194 if (Caller && !Caller->isDeclaration() &&
195 Caller->hasFnAttr(Attribute::OptimizeForSize) &&
196 InlineLimit.getNumOccurrences() == 0)
197 return OptSizeThreshold;
199 return InlineThreshold;
202 /// shouldInline - Return true if the inliner should attempt to inline
203 /// at the given CallSite.
204 bool Inliner::shouldInline(CallSite CS) {
205 InlineCost IC = getInlineCost(CS);
208 DEBUG(dbgs() << " Inlining: cost=always"
209 << ", Call: " << *CS.getInstruction() << "\n");
214 DEBUG(dbgs() << " NOT Inlining: cost=never"
215 << ", Call: " << *CS.getInstruction() << "\n");
219 int Cost = IC.getValue();
220 Function *Caller = CS.getCaller();
221 int CurrentThreshold = getInlineThreshold(CS);
222 float FudgeFactor = getInlineFudgeFactor(CS);
223 if (Cost >= (int)(CurrentThreshold * FudgeFactor)) {
224 DEBUG(dbgs() << " NOT Inlining: cost=" << Cost
225 << ", Call: " << *CS.getInstruction() << "\n");
229 // Try to detect the case where the current inlining candidate caller
230 // (call it B) is a static function and is an inlining candidate elsewhere,
231 // and the current candidate callee (call it C) is large enough that
232 // inlining it into B would make B too big to inline later. In these
233 // circumstances it may be best not to inline C into B, but to inline B
235 if (Caller->hasLocalLinkage()) {
236 int TotalSecondaryCost = 0;
237 bool outerCallsFound = false;
238 bool allOuterCallsWillBeInlined = true;
239 bool someOuterCallWouldNotBeInlined = false;
240 for (Value::use_iterator I = Caller->use_begin(), E =Caller->use_end();
242 CallSite CS2 = CallSite::get(*I);
244 // If this isn't a call to Caller (it could be some other sort
245 // of reference) skip it.
246 if (CS2.getInstruction() == 0 || CS2.getCalledFunction() != Caller)
249 InlineCost IC2 = getInlineCost(CS2);
251 allOuterCallsWillBeInlined = false;
252 if (IC2.isAlways() || IC2.isNever())
255 outerCallsFound = true;
256 int Cost2 = IC2.getValue();
257 int CurrentThreshold2 = getInlineThreshold(CS2);
258 float FudgeFactor2 = getInlineFudgeFactor(CS2);
260 if (Cost2 >= (int)(CurrentThreshold2 * FudgeFactor2))
261 allOuterCallsWillBeInlined = false;
263 // See if we have this case. We subtract off the penalty
264 // for the call instruction, which we would be deleting.
265 if (Cost2 < (int)(CurrentThreshold2 * FudgeFactor2) &&
266 Cost2 + Cost - (InlineConstants::CallPenalty + 1) >=
267 (int)(CurrentThreshold2 * FudgeFactor2)) {
268 someOuterCallWouldNotBeInlined = true;
269 TotalSecondaryCost += Cost2;
272 // If all outer calls to Caller would get inlined, the cost for the last
273 // one is set very low by getInlineCost, in anticipation that Caller will
274 // be removed entirely. We did not account for this above unless there
275 // is only one caller of Caller.
276 if (allOuterCallsWillBeInlined && Caller->use_begin() != Caller->use_end())
277 TotalSecondaryCost += InlineConstants::LastCallToStaticBonus;
279 if (outerCallsFound && someOuterCallWouldNotBeInlined &&
280 TotalSecondaryCost < Cost) {
281 DEBUG(dbgs() << " NOT Inlining: " << *CS.getInstruction() <<
282 " Cost = " << Cost <<
283 ", outer Cost = " << TotalSecondaryCost << '\n');
288 DEBUG(dbgs() << " Inlining: cost=" << Cost
289 << ", Call: " << *CS.getInstruction() << '\n');
293 bool Inliner::runOnSCC(std::vector<CallGraphNode*> &SCC) {
294 CallGraph &CG = getAnalysis<CallGraph>();
295 const TargetData *TD = getAnalysisIfAvailable<TargetData>();
297 SmallPtrSet<Function*, 8> SCCFunctions;
298 DEBUG(dbgs() << "Inliner visiting SCC:");
299 for (unsigned i = 0, e = SCC.size(); i != e; ++i) {
300 Function *F = SCC[i]->getFunction();
301 if (F) SCCFunctions.insert(F);
302 DEBUG(dbgs() << " " << (F ? F->getName() : "INDIRECTNODE"));
305 // Scan through and identify all call sites ahead of time so that we only
306 // inline call sites in the original functions, not call sites that result
307 // from inlining other functions.
308 SmallVector<CallSite, 16> CallSites;
310 for (unsigned i = 0, e = SCC.size(); i != e; ++i) {
311 Function *F = SCC[i]->getFunction();
314 for (Function::iterator BB = F->begin(), E = F->end(); BB != E; ++BB)
315 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
316 CallSite CS = CallSite::get(I);
317 // If this isn't a call, or it is a call to an intrinsic, it can
319 if (CS.getInstruction() == 0 || isa<IntrinsicInst>(I))
322 // If this is a direct call to an external function, we can never inline
323 // it. If it is an indirect call, inlining may resolve it to be a
324 // direct call, so we keep it.
325 if (CS.getCalledFunction() && CS.getCalledFunction()->isDeclaration())
328 CallSites.push_back(CS);
332 DEBUG(dbgs() << ": " << CallSites.size() << " call sites.\n");
334 // Now that we have all of the call sites, move the ones to functions in the
335 // current SCC to the end of the list.
336 unsigned FirstCallInSCC = CallSites.size();
337 for (unsigned i = 0; i < FirstCallInSCC; ++i)
338 if (Function *F = CallSites[i].getCalledFunction())
339 if (SCCFunctions.count(F))
340 std::swap(CallSites[i--], CallSites[--FirstCallInSCC]);
343 InlinedArrayAllocasTy InlinedArrayAllocas;
345 // Now that we have all of the call sites, loop over them and inline them if
346 // it looks profitable to do so.
347 bool Changed = false;
351 // Iterate over the outer loop because inlining functions can cause indirect
352 // calls to become direct calls.
353 for (unsigned CSi = 0; CSi != CallSites.size(); ++CSi) {
354 CallSite CS = CallSites[CSi];
356 Function *Caller = CS.getCaller();
357 Function *Callee = CS.getCalledFunction();
359 // If this call site is dead and it is to a readonly function, we should
360 // just delete the call instead of trying to inline it, regardless of
361 // size. This happens because IPSCCP propagates the result out of the
362 // call and then we're left with the dead call.
363 if (isInstructionTriviallyDead(CS.getInstruction())) {
364 DEBUG(dbgs() << " -> Deleting dead call: "
365 << *CS.getInstruction() << "\n");
366 // Update the call graph by deleting the edge from Callee to Caller.
367 CG[Caller]->removeCallEdgeFor(CS);
368 CS.getInstruction()->eraseFromParent();
371 // We can only inline direct calls to non-declarations.
372 if (Callee == 0 || Callee->isDeclaration()) continue;
374 // If the policy determines that we should inline this function,
376 if (!shouldInline(CS))
379 // Attempt to inline the function...
380 if (!InlineCallIfPossible(CS, CG, TD, InlinedArrayAllocas))
385 // If we inlined or deleted the last possible call site to the function,
386 // delete the function body now.
387 if (Callee && Callee->use_empty() && Callee->hasLocalLinkage() &&
388 // TODO: Can remove if in SCC now.
389 !SCCFunctions.count(Callee) &&
391 // The function may be apparently dead, but if there are indirect
392 // callgraph references to the node, we cannot delete it yet, this
393 // could invalidate the CGSCC iterator.
394 CG[Callee]->getNumReferences() == 0) {
395 DEBUG(dbgs() << " -> Deleting dead function: "
396 << Callee->getName() << "\n");
397 CallGraphNode *CalleeNode = CG[Callee];
399 // Remove any call graph edges from the callee to its callees.
400 CalleeNode->removeAllCalledFunctions();
402 resetCachedCostInfo(Callee);
404 // Removing the node for callee from the call graph and delete it.
405 delete CG.removeFunctionFromModule(CalleeNode);
409 // Remove any cached cost info for this caller, as inlining the
410 // callee has increased the size of the caller (which may be the
411 // same as the callee).
412 resetCachedCostInfo(Caller);
414 // Remove this call site from the list. If possible, use
415 // swap/pop_back for efficiency, but do not use it if doing so would
416 // move a call site to a function in this SCC before the
417 // 'FirstCallInSCC' barrier.
418 if (SCC.size() == 1) {
419 std::swap(CallSites[CSi], CallSites.back());
420 CallSites.pop_back();
422 CallSites.erase(CallSites.begin()+CSi);
429 } while (LocalChange);
434 // doFinalization - Remove now-dead linkonce functions at the end of
435 // processing to avoid breaking the SCC traversal.
436 bool Inliner::doFinalization(CallGraph &CG) {
437 return removeDeadFunctions(CG);
440 /// removeDeadFunctions - Remove dead functions that are not included in
441 /// DNR (Do Not Remove) list.
442 bool Inliner::removeDeadFunctions(CallGraph &CG,
443 SmallPtrSet<const Function *, 16> *DNR) {
444 SmallPtrSet<CallGraphNode*, 16> FunctionsToRemove;
446 // Scan for all of the functions, looking for ones that should now be removed
447 // from the program. Insert the dead ones in the FunctionsToRemove set.
448 for (CallGraph::iterator I = CG.begin(), E = CG.end(); I != E; ++I) {
449 CallGraphNode *CGN = I->second;
450 if (CGN->getFunction() == 0)
453 Function *F = CGN->getFunction();
455 // If the only remaining users of the function are dead constants, remove
457 F->removeDeadConstantUsers();
459 if (DNR && DNR->count(F))
461 if (!F->hasLinkOnceLinkage() && !F->hasLocalLinkage() &&
462 !F->hasAvailableExternallyLinkage())
467 // Remove any call graph edges from the function to its callees.
468 CGN->removeAllCalledFunctions();
470 // Remove any edges from the external node to the function's call graph
471 // node. These edges might have been made irrelegant due to
472 // optimization of the program.
473 CG.getExternalCallingNode()->removeAnyCallEdgeTo(CGN);
475 // Removing the node for callee from the call graph and delete it.
476 FunctionsToRemove.insert(CGN);
479 // Now that we know which functions to delete, do so. We didn't want to do
480 // this inline, because that would invalidate our CallGraph::iterator
483 // Note that it doesn't matter that we are iterating over a non-stable set
484 // here to do this, it doesn't matter which order the functions are deleted
486 bool Changed = false;
487 for (SmallPtrSet<CallGraphNode*, 16>::iterator I = FunctionsToRemove.begin(),
488 E = FunctionsToRemove.end(); I != E; ++I) {
489 resetCachedCostInfo((*I)->getFunction());
490 delete CG.removeFunctionFromModule(*I);