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/Support/CallSite.h"
22 #include "llvm/Target/TargetData.h"
23 #include "llvm/Transforms/IPO/InlinerPass.h"
24 #include "llvm/Transforms/Utils/InlineCost.h"
25 #include "llvm/Transforms/Utils/Cloning.h"
26 #include "llvm/Support/CommandLine.h"
27 #include "llvm/Support/Debug.h"
28 #include "llvm/Support/raw_ostream.h"
29 #include "llvm/ADT/SmallPtrSet.h"
30 #include "llvm/ADT/Statistic.h"
34 STATISTIC(NumInlined, "Number of functions inlined");
35 STATISTIC(NumDeleted, "Number of functions deleted because all callers found");
36 STATISTIC(NumMergedAllocas, "Number of allocas merged together");
39 InlineLimit("inline-threshold", cl::Hidden, cl::init(200), cl::ZeroOrMore,
40 cl::desc("Control the amount of inlining to perform (default = 200)"));
42 Inliner::Inliner(void *ID)
43 : CallGraphSCCPass(ID), InlineThreshold(InlineLimit) {}
45 Inliner::Inliner(void *ID, int Threshold)
46 : CallGraphSCCPass(ID), InlineThreshold(Threshold) {}
48 /// getAnalysisUsage - For this class, we declare that we require and preserve
49 /// the call graph. If the derived class implements this method, it should
50 /// always explicitly call the implementation here.
51 void Inliner::getAnalysisUsage(AnalysisUsage &Info) const {
52 CallGraphSCCPass::getAnalysisUsage(Info);
56 typedef DenseMap<const ArrayType*, std::vector<AllocaInst*> >
57 InlinedArrayAllocasTy;
59 /// InlineCallIfPossible - If it is possible to inline the specified call site,
60 /// do so and update the CallGraph for this operation.
62 /// This function also does some basic book-keeping to update the IR. The
63 /// InlinedArrayAllocas map keeps track of any allocas that are already
64 /// available from other functions inlined into the caller. If we are able to
65 /// inline this call site we attempt to reuse already available allocas or add
66 /// any new allocas to the set if not possible.
67 static bool InlineCallIfPossible(CallSite CS, CallGraph &CG,
69 InlinedArrayAllocasTy &InlinedArrayAllocas) {
70 Function *Callee = CS.getCalledFunction();
71 Function *Caller = CS.getCaller();
73 // Try to inline the function. Get the list of static allocas that were
75 SmallVector<AllocaInst*, 16> StaticAllocas;
76 if (!InlineFunction(CS, &CG, TD, &StaticAllocas))
79 // If the inlined function had a higher stack protection level than the
80 // calling function, then bump up the caller's stack protection level.
81 if (Callee->hasFnAttr(Attribute::StackProtectReq))
82 Caller->addFnAttr(Attribute::StackProtectReq);
83 else if (Callee->hasFnAttr(Attribute::StackProtect) &&
84 !Caller->hasFnAttr(Attribute::StackProtectReq))
85 Caller->addFnAttr(Attribute::StackProtect);
88 // Look at all of the allocas that we inlined through this call site. If we
89 // have already inlined other allocas through other calls into this function,
90 // then we know that they have disjoint lifetimes and that we can merge them.
92 // There are many heuristics possible for merging these allocas, and the
93 // different options have different tradeoffs. One thing that we *really*
94 // don't want to hurt is SRoA: once inlining happens, often allocas are no
95 // longer address taken and so they can be promoted.
97 // Our "solution" for that is to only merge allocas whose outermost type is an
98 // array type. These are usually not promoted because someone is using a
99 // variable index into them. These are also often the most important ones to
102 // A better solution would be to have real memory lifetime markers in the IR
103 // and not have the inliner do any merging of allocas at all. This would
104 // allow the backend to do proper stack slot coloring of all allocas that
105 // *actually make it to the backend*, which is really what we want.
107 // Because we don't have this information, we do this simple and useful hack.
109 SmallPtrSet<AllocaInst*, 16> UsedAllocas;
111 // Loop over all the allocas we have so far and see if they can be merged with
112 // a previously inlined alloca. If not, remember that we had it.
113 for (unsigned AllocaNo = 0, e = StaticAllocas.size();
114 AllocaNo != e; ++AllocaNo) {
115 AllocaInst *AI = StaticAllocas[AllocaNo];
117 // Don't bother trying to merge array allocations (they will usually be
118 // canonicalized to be an allocation *of* an array), or allocations whose
119 // type is not itself an array (because we're afraid of pessimizing SRoA).
120 const ArrayType *ATy = dyn_cast<ArrayType>(AI->getAllocatedType());
121 if (ATy == 0 || AI->isArrayAllocation())
124 // Get the list of all available allocas for this array type.
125 std::vector<AllocaInst*> &AllocasForType = InlinedArrayAllocas[ATy];
127 // Loop over the allocas in AllocasForType to see if we can reuse one. Note
128 // that we have to be careful not to reuse the same "available" alloca for
129 // multiple different allocas that we just inlined, we use the 'UsedAllocas'
130 // set to keep track of which "available" allocas are being used by this
131 // function. Also, AllocasForType can be empty of course!
132 bool MergedAwayAlloca = false;
133 for (unsigned i = 0, e = AllocasForType.size(); i != e; ++i) {
134 AllocaInst *AvailableAlloca = AllocasForType[i];
136 // The available alloca has to be in the right function, not in some other
137 // function in this SCC.
138 if (AvailableAlloca->getParent() != AI->getParent())
141 // If the inlined function already uses this alloca then we can't reuse
143 if (!UsedAllocas.insert(AvailableAlloca))
146 // Otherwise, we *can* reuse it, RAUW AI into AvailableAlloca and declare
148 DEBUG(errs() << " ***MERGED ALLOCA: " << *AI);
150 AI->replaceAllUsesWith(AvailableAlloca);
151 AI->eraseFromParent();
152 MergedAwayAlloca = true;
157 // If we already nuked the alloca, we're done with it.
158 if (MergedAwayAlloca)
161 // If we were unable to merge away the alloca either because there are no
162 // allocas of the right type available or because we reused them all
163 // already, remember that this alloca came from an inlined function and mark
164 // it used so we don't reuse it for other allocas from this inline
166 AllocasForType.push_back(AI);
167 UsedAllocas.insert(AI);
173 /// shouldInline - Return true if the inliner should attempt to inline
174 /// at the given CallSite.
175 bool Inliner::shouldInline(CallSite CS) {
176 InlineCost IC = getInlineCost(CS);
179 DEBUG(errs() << " Inlining: cost=always"
180 << ", Call: " << *CS.getInstruction() << "\n");
185 DEBUG(errs() << " NOT Inlining: cost=never"
186 << ", Call: " << *CS.getInstruction() << "\n");
190 int Cost = IC.getValue();
191 int CurrentThreshold = InlineThreshold;
192 Function *Caller = CS.getCaller();
193 if (Caller && !Caller->isDeclaration() &&
194 Caller->hasFnAttr(Attribute::OptimizeForSize) &&
195 InlineLimit.getNumOccurrences() == 0 &&
196 InlineThreshold != 50)
197 CurrentThreshold = 50;
199 float FudgeFactor = getInlineFudgeFactor(CS);
200 if (Cost >= (int)(CurrentThreshold * FudgeFactor)) {
201 DEBUG(errs() << " NOT Inlining: cost=" << Cost
202 << ", Call: " << *CS.getInstruction() << "\n");
206 // Try to detect the case where the current inlining candidate caller
207 // (call it B) is a static function and is an inlining candidate elsewhere,
208 // and the current candidate callee (call it C) is large enough that
209 // inlining it into B would make B too big to inline later. In these
210 // circumstances it may be best not to inline C into B, but to inline B
212 if (Caller->hasLocalLinkage()) {
213 int TotalSecondaryCost = 0;
214 bool outerCallsFound = false;
215 bool allOuterCallsWillBeInlined = true;
216 bool someOuterCallWouldNotBeInlined = false;
217 for (Value::use_iterator I = Caller->use_begin(), E =Caller->use_end();
219 CallSite CS2 = CallSite::get(*I);
221 // If this isn't a call to Caller (it could be some other sort
222 // of reference) skip it.
223 if (CS2.getInstruction() == 0 || CS2.getCalledFunction() != Caller)
226 InlineCost IC2 = getInlineCost(CS2);
228 allOuterCallsWillBeInlined = false;
229 if (IC2.isAlways() || IC2.isNever())
232 outerCallsFound = true;
233 int Cost2 = IC2.getValue();
234 int CurrentThreshold2 = InlineThreshold;
235 Function *Caller2 = CS2.getCaller();
236 if (Caller2 && !Caller2->isDeclaration() &&
237 Caller2->hasFnAttr(Attribute::OptimizeForSize) &&
238 InlineThreshold != 50)
239 CurrentThreshold2 = 50;
241 float FudgeFactor2 = getInlineFudgeFactor(CS2);
243 if (Cost2 >= (int)(CurrentThreshold2 * FudgeFactor2))
244 allOuterCallsWillBeInlined = false;
246 // See if we have this case. The magic 6 is what InlineCost assigns
247 // for the call instruction, which we would be deleting.
248 if (Cost2 < (int)(CurrentThreshold2 * FudgeFactor2) &&
249 Cost2 + Cost - 6 >= (int)(CurrentThreshold2 * FudgeFactor2)) {
250 someOuterCallWouldNotBeInlined = true;
251 TotalSecondaryCost += Cost2;
254 // If all outer calls to Caller would get inlined, the cost for the last
255 // one is set very low by getInlineCost, in anticipation that Caller will
256 // be removed entirely. We did not account for this above unless there
257 // is only one caller of Caller.
258 if (allOuterCallsWillBeInlined && Caller->use_begin() != Caller->use_end())
259 TotalSecondaryCost -= 15000;
261 if (outerCallsFound && someOuterCallWouldNotBeInlined &&
262 TotalSecondaryCost < Cost) {
263 DEBUG(errs() << " NOT Inlining: " << *CS.getInstruction() <<
264 " Cost = " << Cost <<
265 ", outer Cost = " << TotalSecondaryCost << '\n');
270 DEBUG(errs() << " Inlining: cost=" << Cost
271 << ", Call: " << *CS.getInstruction() << '\n');
275 bool Inliner::runOnSCC(std::vector<CallGraphNode*> &SCC) {
276 CallGraph &CG = getAnalysis<CallGraph>();
277 const TargetData *TD = getAnalysisIfAvailable<TargetData>();
279 SmallPtrSet<Function*, 8> SCCFunctions;
280 DEBUG(errs() << "Inliner visiting SCC:");
281 for (unsigned i = 0, e = SCC.size(); i != e; ++i) {
282 Function *F = SCC[i]->getFunction();
283 if (F) SCCFunctions.insert(F);
284 DEBUG(errs() << " " << (F ? F->getName() : "INDIRECTNODE"));
287 // Scan through and identify all call sites ahead of time so that we only
288 // inline call sites in the original functions, not call sites that result
289 // from inlining other functions.
290 SmallVector<CallSite, 16> CallSites;
292 for (unsigned i = 0, e = SCC.size(); i != e; ++i) {
293 Function *F = SCC[i]->getFunction();
296 for (Function::iterator BB = F->begin(), E = F->end(); BB != E; ++BB)
297 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
298 CallSite CS = CallSite::get(I);
299 // If this isn't a call, or it is a call to an intrinsic, it can
301 if (CS.getInstruction() == 0 || isa<IntrinsicInst>(I))
304 // If this is a direct call to an external function, we can never inline
305 // it. If it is an indirect call, inlining may resolve it to be a
306 // direct call, so we keep it.
307 if (CS.getCalledFunction() && CS.getCalledFunction()->isDeclaration())
310 CallSites.push_back(CS);
314 DEBUG(errs() << ": " << CallSites.size() << " call sites.\n");
316 // Now that we have all of the call sites, move the ones to functions in the
317 // current SCC to the end of the list.
318 unsigned FirstCallInSCC = CallSites.size();
319 for (unsigned i = 0; i < FirstCallInSCC; ++i)
320 if (Function *F = CallSites[i].getCalledFunction())
321 if (SCCFunctions.count(F))
322 std::swap(CallSites[i--], CallSites[--FirstCallInSCC]);
325 InlinedArrayAllocasTy InlinedArrayAllocas;
327 // Now that we have all of the call sites, loop over them and inline them if
328 // it looks profitable to do so.
329 bool Changed = false;
333 // Iterate over the outer loop because inlining functions can cause indirect
334 // calls to become direct calls.
335 for (unsigned CSi = 0; CSi != CallSites.size(); ++CSi) {
336 CallSite CS = CallSites[CSi];
338 Function *Callee = CS.getCalledFunction();
339 // We can only inline direct calls to non-declarations.
340 if (Callee == 0 || Callee->isDeclaration()) continue;
342 // If the policy determines that we should inline this function,
344 if (!shouldInline(CS))
347 Function *Caller = CS.getCaller();
348 // Attempt to inline the function...
349 if (!InlineCallIfPossible(CS, CG, TD, InlinedArrayAllocas))
352 // If we inlined the last possible call site to the function, delete the
353 // function body now.
354 if (Callee->use_empty() && Callee->hasLocalLinkage() &&
355 // TODO: Can remove if in SCC now.
356 !SCCFunctions.count(Callee) &&
358 // The function may be apparently dead, but if there are indirect
359 // callgraph references to the node, we cannot delete it yet, this
360 // could invalidate the CGSCC iterator.
361 CG[Callee]->getNumReferences() == 0) {
362 DEBUG(errs() << " -> Deleting dead function: "
363 << Callee->getName() << "\n");
364 CallGraphNode *CalleeNode = CG[Callee];
366 // Remove any call graph edges from the callee to its callees.
367 CalleeNode->removeAllCalledFunctions();
369 resetCachedCostInfo(Callee);
371 // Removing the node for callee from the call graph and delete it.
372 delete CG.removeFunctionFromModule(CalleeNode);
376 // Remove any cached cost info for this caller, as inlining the
377 // callee has increased the size of the caller (which may be the
378 // same as the callee).
379 resetCachedCostInfo(Caller);
381 // Remove this call site from the list. If possible, use
382 // swap/pop_back for efficiency, but do not use it if doing so would
383 // move a call site to a function in this SCC before the
384 // 'FirstCallInSCC' barrier.
385 if (SCC.size() == 1) {
386 std::swap(CallSites[CSi], CallSites.back());
387 CallSites.pop_back();
389 CallSites.erase(CallSites.begin()+CSi);
397 } while (LocalChange);
402 // doFinalization - Remove now-dead linkonce functions at the end of
403 // processing to avoid breaking the SCC traversal.
404 bool Inliner::doFinalization(CallGraph &CG) {
405 return removeDeadFunctions(CG);
408 /// removeDeadFunctions - Remove dead functions that are not included in
409 /// DNR (Do Not Remove) list.
410 bool Inliner::removeDeadFunctions(CallGraph &CG,
411 SmallPtrSet<const Function *, 16> *DNR) {
412 SmallPtrSet<CallGraphNode*, 16> FunctionsToRemove;
414 // Scan for all of the functions, looking for ones that should now be removed
415 // from the program. Insert the dead ones in the FunctionsToRemove set.
416 for (CallGraph::iterator I = CG.begin(), E = CG.end(); I != E; ++I) {
417 CallGraphNode *CGN = I->second;
418 if (CGN->getFunction() == 0)
421 Function *F = CGN->getFunction();
423 // If the only remaining users of the function are dead constants, remove
425 F->removeDeadConstantUsers();
427 if (DNR && DNR->count(F))
429 if (!F->hasLinkOnceLinkage() && !F->hasLocalLinkage() &&
430 !F->hasAvailableExternallyLinkage())
435 // Remove any call graph edges from the function to its callees.
436 CGN->removeAllCalledFunctions();
438 // Remove any edges from the external node to the function's call graph
439 // node. These edges might have been made irrelegant due to
440 // optimization of the program.
441 CG.getExternalCallingNode()->removeAnyCallEdgeTo(CGN);
443 // Removing the node for callee from the call graph and delete it.
444 FunctionsToRemove.insert(CGN);
447 // Now that we know which functions to delete, do so. We didn't want to do
448 // this inline, because that would invalidate our CallGraph::iterator
451 // Note that it doesn't matter that we are iterating over a non-stable set
452 // here to do this, it doesn't matter which order the functions are deleted
454 bool Changed = false;
455 for (SmallPtrSet<CallGraphNode*, 16>::iterator I = FunctionsToRemove.begin(),
456 E = FunctionsToRemove.end(); I != E; ++I) {
457 resetCachedCostInfo((*I)->getFunction());
458 delete CG.removeFunctionFromModule(*I);