1 //===- InlineCost.cpp - Cost analysis for inliner -------------------------===//
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 inline cost analysis.
12 //===----------------------------------------------------------------------===//
14 #include "llvm/Transforms/Utils/InlineCost.h"
15 #include "llvm/Support/CallSite.h"
16 #include "llvm/CallingConv.h"
17 #include "llvm/IntrinsicInst.h"
18 #include "llvm/ADT/SmallPtrSet.h"
21 // CountCodeReductionForConstant - Figure out an approximation for how many
22 // instructions will be constant folded if the specified value is constant.
24 unsigned InlineCostAnalyzer::FunctionInfo::
25 CountCodeReductionForConstant(Value *V) {
26 unsigned Reduction = 0;
27 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; ++UI)
28 if (isa<BranchInst>(*UI))
29 Reduction += 40; // Eliminating a conditional branch is a big win
30 else if (SwitchInst *SI = dyn_cast<SwitchInst>(*UI))
31 // Eliminating a switch is a big win, proportional to the number of edges
33 Reduction += (SI->getNumSuccessors()-1) * 40;
34 else if (CallInst *CI = dyn_cast<CallInst>(*UI)) {
35 // Turning an indirect call into a direct call is a BIG win
36 Reduction += CI->getCalledValue() == V ? 500 : 0;
37 } else if (InvokeInst *II = dyn_cast<InvokeInst>(*UI)) {
38 // Turning an indirect call into a direct call is a BIG win
39 Reduction += II->getCalledValue() == V ? 500 : 0;
41 // Figure out if this instruction will be removed due to simple constant
43 Instruction &Inst = cast<Instruction>(**UI);
45 // We can't constant propagate instructions which have effects or
48 // FIXME: It would be nice to capture the fact that a load from a
49 // pointer-to-constant-global is actually a *really* good thing to zap.
50 // Unfortunately, we don't know the pointer that may get propagated here,
51 // so we can't make this decision.
52 if (Inst.mayReadFromMemory() || Inst.mayHaveSideEffects() ||
53 isa<AllocationInst>(Inst))
56 bool AllOperandsConstant = true;
57 for (unsigned i = 0, e = Inst.getNumOperands(); i != e; ++i)
58 if (!isa<Constant>(Inst.getOperand(i)) && Inst.getOperand(i) != V) {
59 AllOperandsConstant = false;
63 if (AllOperandsConstant) {
64 // We will get to remove this instruction...
67 // And any other instructions that use it which become constants
69 Reduction += CountCodeReductionForConstant(&Inst);
76 // CountCodeReductionForAlloca - Figure out an approximation of how much smaller
77 // the function will be if it is inlined into a context where an argument
80 unsigned InlineCostAnalyzer::FunctionInfo::
81 CountCodeReductionForAlloca(Value *V) {
82 if (!isa<PointerType>(V->getType())) return 0; // Not a pointer
83 unsigned Reduction = 0;
84 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E;++UI){
85 Instruction *I = cast<Instruction>(*UI);
86 if (isa<LoadInst>(I) || isa<StoreInst>(I))
88 else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(I)) {
89 // If the GEP has variable indices, we won't be able to do much with it.
90 if (!GEP->hasAllConstantIndices())
91 Reduction += CountCodeReductionForAlloca(GEP)+15;
93 // If there is some other strange instruction, we're not going to be able
94 // to do much if we inline this.
102 /// analyzeFunction - Fill in the current structure with information gleaned
103 /// from the specified function.
104 void InlineCostAnalyzer::FunctionInfo::analyzeFunction(Function *F) {
105 unsigned NumInsts = 0, NumBlocks = 0, NumVectorInsts = 0, NumRets = 0;
107 // Look at the size of the callee. Each basic block counts as 20 units, and
108 // each instruction counts as 5.
109 for (Function::const_iterator BB = F->begin(), E = F->end(); BB != E; ++BB) {
110 for (BasicBlock::const_iterator II = BB->begin(), E = BB->end();
112 if (isa<PHINode>(II)) continue; // PHI nodes don't count.
114 // Special handling for calls.
115 if (isa<CallInst>(II) || isa<InvokeInst>(II)) {
116 if (isa<DbgInfoIntrinsic>(II))
117 continue; // Debug intrinsics don't count as size.
119 CallSite CS = CallSite::get(const_cast<Instruction*>(&*II));
121 // If this function contains a call to setjmp or _setjmp, never inline
122 // it. This is a hack because we depend on the user marking their local
123 // variables as volatile if they are live across a setjmp call, and they
124 // probably won't do this in callers.
125 if (Function *F = CS.getCalledFunction())
126 if (F->isDeclaration() &&
127 (F->getName() == "setjmp" || F->getName() == "_setjmp")) {
132 // Calls often compile into many machine instructions. Bump up their
133 // cost to reflect this.
134 if (!isa<IntrinsicInst>(II))
135 NumInsts += InlineConstants::CallPenalty;
138 // These, too, are calls.
139 if (isa<MallocInst>(II) || isa<FreeInst>(II))
140 NumInsts += InlineConstants::CallPenalty;
142 if (const AllocaInst *AI = dyn_cast<AllocaInst>(II)) {
143 if (!AI->isStaticAlloca())
144 this->usesDynamicAlloca = true;
147 if (isa<ExtractElementInst>(II) || isa<VectorType>(II->getType()))
150 // Noop casts, including ptr <-> int, don't count.
151 if (const CastInst *CI = dyn_cast<CastInst>(II)) {
152 if (CI->isLosslessCast() || isa<IntToPtrInst>(CI) ||
153 isa<PtrToIntInst>(CI))
155 } else if (const GetElementPtrInst *GEPI =
156 dyn_cast<GetElementPtrInst>(II)) {
157 // If a GEP has all constant indices, it will probably be folded with
159 if (GEPI->hasAllConstantIndices())
163 if (isa<ReturnInst>(II))
172 // A function with exactly one return has it removed during the inlining
173 // process (see InlineFunction), so don't count it.
177 this->NumBlocks = NumBlocks;
178 this->NumInsts = NumInsts;
179 this->NumVectorInsts = NumVectorInsts;
181 // Check out all of the arguments to the function, figuring out how much
182 // code can be eliminated if one of the arguments is a constant.
183 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(); I != E; ++I)
184 ArgumentWeights.push_back(ArgInfo(CountCodeReductionForConstant(I),
185 CountCodeReductionForAlloca(I)));
190 // getInlineCost - The heuristic used to determine if we should inline the
191 // function call or not.
193 InlineCost InlineCostAnalyzer::getInlineCost(CallSite CS,
194 SmallPtrSet<const Function *, 16> &NeverInline) {
195 Instruction *TheCall = CS.getInstruction();
196 Function *Callee = CS.getCalledFunction();
197 Function *Caller = TheCall->getParent()->getParent();
199 // Don't inline functions which can be redefined at link-time to mean
200 // something else. Don't inline functions marked noinline.
201 if (Callee->mayBeOverridden() ||
202 Callee->hasFnAttr(Attribute::NoInline) || NeverInline.count(Callee))
203 return llvm::InlineCost::getNever();
205 // InlineCost - This value measures how good of an inline candidate this call
206 // site is to inline. A lower inline cost make is more likely for the call to
207 // be inlined. This value may go negative.
211 // If there is only one call of the function, and it has internal linkage,
212 // make it almost guaranteed to be inlined.
214 if (Callee->hasLocalLinkage() && Callee->hasOneUse())
215 InlineCost += InlineConstants::LastCallToStaticBonus;
217 // If this function uses the coldcc calling convention, prefer not to inline
219 if (Callee->getCallingConv() == CallingConv::Cold)
220 InlineCost += InlineConstants::ColdccPenalty;
222 // If the instruction after the call, or if the normal destination of the
223 // invoke is an unreachable instruction, the function is noreturn. As such,
224 // there is little point in inlining this.
225 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) {
226 if (isa<UnreachableInst>(II->getNormalDest()->begin()))
227 InlineCost += InlineConstants::NoreturnPenalty;
228 } else if (isa<UnreachableInst>(++BasicBlock::iterator(TheCall)))
229 InlineCost += InlineConstants::NoreturnPenalty;
231 // Get information about the callee...
232 FunctionInfo &CalleeFI = CachedFunctionInfo[Callee];
234 // If we haven't calculated this information yet, do so now.
235 if (CalleeFI.NumBlocks == 0)
236 CalleeFI.analyzeFunction(Callee);
238 // If we should never inline this, return a huge cost.
239 if (CalleeFI.NeverInline)
240 return InlineCost::getNever();
242 // FIXME: It would be nice to kill off CalleeFI.NeverInline. Then we
243 // could move this up and avoid computing the FunctionInfo for
244 // things we are going to just return always inline for. This
245 // requires handling setjmp somewhere else, however.
246 if (!Callee->isDeclaration() && Callee->hasFnAttr(Attribute::AlwaysInline))
247 return InlineCost::getAlways();
249 if (CalleeFI.usesDynamicAlloca) {
250 // Get infomation about the caller...
251 FunctionInfo &CallerFI = CachedFunctionInfo[Caller];
253 // If we haven't calculated this information yet, do so now.
254 if (CallerFI.NumBlocks == 0)
255 CallerFI.analyzeFunction(Caller);
257 // Don't inline a callee with dynamic alloca into a caller without them.
258 // Functions containing dynamic alloca's are inefficient in various ways;
259 // don't create more inefficiency.
260 if (!CallerFI.usesDynamicAlloca)
261 return InlineCost::getNever();
264 // Add to the inline quality for properties that make the call valuable to
265 // inline. This includes factors that indicate that the result of inlining
266 // the function will be optimizable. Currently this just looks at arguments
267 // passed into the function.
270 for (CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end();
271 I != E; ++I, ++ArgNo) {
272 // Each argument passed in has a cost at both the caller and the callee
273 // sides. This favors functions that take many arguments over functions
274 // that take few arguments.
277 // If this is a function being passed in, it is very likely that we will be
278 // able to turn an indirect function call into a direct function call.
279 if (isa<Function>(I))
282 // If an alloca is passed in, inlining this function is likely to allow
283 // significant future optimization possibilities (like scalar promotion, and
284 // scalarization), so encourage the inlining of the function.
286 else if (isa<AllocaInst>(I)) {
287 if (ArgNo < CalleeFI.ArgumentWeights.size())
288 InlineCost -= CalleeFI.ArgumentWeights[ArgNo].AllocaWeight;
290 // If this is a constant being passed into the function, use the argument
291 // weights calculated for the callee to determine how much will be folded
292 // away with this information.
293 } else if (isa<Constant>(I)) {
294 if (ArgNo < CalleeFI.ArgumentWeights.size())
295 InlineCost -= CalleeFI.ArgumentWeights[ArgNo].ConstantWeight;
299 // Now that we have considered all of the factors that make the call site more
300 // likely to be inlined, look at factors that make us not want to inline it.
302 // Don't inline into something too big, which would make it bigger.
303 // "size" here is the number of basic blocks, not instructions.
305 InlineCost += Caller->size()/15;
307 // Look at the size of the callee. Each instruction counts as 5.
308 InlineCost += CalleeFI.NumInsts*5;
310 return llvm::InlineCost::get(InlineCost);
313 // getInlineFudgeFactor - Return a > 1.0 factor if the inliner should use a
314 // higher threshold to determine if the function call should be inlined.
315 float InlineCostAnalyzer::getInlineFudgeFactor(CallSite CS) {
316 Function *Callee = CS.getCalledFunction();
318 // Get information about the callee...
319 FunctionInfo &CalleeFI = CachedFunctionInfo[Callee];
321 // If we haven't calculated this information yet, do so now.
322 if (CalleeFI.NumBlocks == 0)
323 CalleeFI.analyzeFunction(Callee);
326 // Single BB functions are often written to be inlined.
327 if (CalleeFI.NumBlocks == 1)
330 // Be more aggressive if the function contains a good chunk (if it mades up
331 // at least 10% of the instructions) of vector instructions.
332 if (CalleeFI.NumVectorInsts > CalleeFI.NumInsts/2)
334 else if (CalleeFI.NumVectorInsts > CalleeFI.NumInsts/10)