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 //===----------------------------------------------------------------------===//
15 #include "llvm/Transforms/Utils/InlineCost.h"
16 #include "llvm/Analysis/MallocHelper.h"
17 #include "llvm/Support/CallSite.h"
18 #include "llvm/CallingConv.h"
19 #include "llvm/IntrinsicInst.h"
20 #include "llvm/ADT/SmallPtrSet.h"
23 // CountCodeReductionForConstant - Figure out an approximation for how many
24 // instructions will be constant folded if the specified value is constant.
26 unsigned InlineCostAnalyzer::FunctionInfo::
27 CountCodeReductionForConstant(Value *V) {
28 unsigned Reduction = 0;
29 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E; ++UI)
30 if (isa<BranchInst>(*UI))
31 Reduction += 40; // Eliminating a conditional branch is a big win
32 else if (SwitchInst *SI = dyn_cast<SwitchInst>(*UI))
33 // Eliminating a switch is a big win, proportional to the number of edges
35 Reduction += (SI->getNumSuccessors()-1) * 40;
36 else if (CallInst *CI = dyn_cast<CallInst>(*UI)) {
37 // Turning an indirect call into a direct call is a BIG win
38 Reduction += CI->getCalledValue() == V ? 500 : 0;
39 } else if (InvokeInst *II = dyn_cast<InvokeInst>(*UI)) {
40 // Turning an indirect call into a direct call is a BIG win
41 Reduction += II->getCalledValue() == V ? 500 : 0;
43 // Figure out if this instruction will be removed due to simple constant
45 Instruction &Inst = cast<Instruction>(**UI);
47 // We can't constant propagate instructions which have effects or
50 // FIXME: It would be nice to capture the fact that a load from a
51 // pointer-to-constant-global is actually a *really* good thing to zap.
52 // Unfortunately, we don't know the pointer that may get propagated here,
53 // so we can't make this decision.
54 if (Inst.mayReadFromMemory() || Inst.mayHaveSideEffects() ||
55 isa<AllocationInst>(Inst) || isMalloc(&Inst))
58 bool AllOperandsConstant = true;
59 for (unsigned i = 0, e = Inst.getNumOperands(); i != e; ++i)
60 if (!isa<Constant>(Inst.getOperand(i)) && Inst.getOperand(i) != V) {
61 AllOperandsConstant = false;
65 if (AllOperandsConstant) {
66 // We will get to remove this instruction...
69 // And any other instructions that use it which become constants
71 Reduction += CountCodeReductionForConstant(&Inst);
78 // CountCodeReductionForAlloca - Figure out an approximation of how much smaller
79 // the function will be if it is inlined into a context where an argument
82 unsigned InlineCostAnalyzer::FunctionInfo::
83 CountCodeReductionForAlloca(Value *V) {
84 if (!isa<PointerType>(V->getType())) return 0; // Not a pointer
85 unsigned Reduction = 0;
86 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E;++UI){
87 Instruction *I = cast<Instruction>(*UI);
88 if (isa<LoadInst>(I) || isa<StoreInst>(I))
90 else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(I)) {
91 // If the GEP has variable indices, we won't be able to do much with it.
92 if (!GEP->hasAllConstantIndices())
93 Reduction += CountCodeReductionForAlloca(GEP)+15;
95 // If there is some other strange instruction, we're not going to be able
96 // to do much if we inline this.
104 /// analyzeFunction - Fill in the current structure with information gleaned
105 /// from the specified function.
106 void InlineCostAnalyzer::FunctionInfo::analyzeFunction(Function *F) {
107 unsigned NumInsts = 0, NumBlocks = 0, NumVectorInsts = 0;
109 // Look at the size of the callee. Each basic block counts as 20 units, and
110 // each instruction counts as 5.
111 for (Function::const_iterator BB = F->begin(), E = F->end(); BB != E; ++BB) {
112 for (BasicBlock::const_iterator II = BB->begin(), E = BB->end();
114 if (isa<PHINode>(II)) continue; // PHI nodes don't count.
116 // Special handling for calls.
117 if (isa<CallInst>(II) || isa<InvokeInst>(II)) {
118 if (isa<DbgInfoIntrinsic>(II))
119 continue; // Debug intrinsics don't count as size.
121 CallSite CS = CallSite::get(const_cast<Instruction*>(&*II));
123 // If this function contains a call to setjmp or _setjmp, never inline
124 // it. This is a hack because we depend on the user marking their local
125 // variables as volatile if they are live across a setjmp call, and they
126 // probably won't do this in callers.
127 if (Function *F = CS.getCalledFunction())
128 if (F->isDeclaration() &&
129 (F->getName() == "setjmp" || F->getName() == "_setjmp")) {
134 // Calls often compile into many machine instructions. Bump up their
135 // cost to reflect this.
136 if (!isa<IntrinsicInst>(II))
140 if (const AllocaInst *AI = dyn_cast<AllocaInst>(II)) {
141 if (!AI->isStaticAlloca())
142 this->usesDynamicAlloca = true;
145 if (isa<ExtractElementInst>(II) || isa<VectorType>(II->getType()))
148 // Noop casts, including ptr <-> int, don't count.
149 if (const CastInst *CI = dyn_cast<CastInst>(II)) {
150 if (CI->isLosslessCast() || isa<IntToPtrInst>(CI) ||
151 isa<PtrToIntInst>(CI))
153 } else if (const GetElementPtrInst *GEPI =
154 dyn_cast<GetElementPtrInst>(II)) {
155 // If a GEP has all constant indices, it will probably be folded with
157 if (GEPI->hasAllConstantIndices())
167 this->NumBlocks = NumBlocks;
168 this->NumInsts = NumInsts;
169 this->NumVectorInsts = NumVectorInsts;
171 // Check out all of the arguments to the function, figuring out how much
172 // code can be eliminated if one of the arguments is a constant.
173 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(); I != E; ++I)
174 ArgumentWeights.push_back(ArgInfo(CountCodeReductionForConstant(I),
175 CountCodeReductionForAlloca(I)));
180 // getInlineCost - The heuristic used to determine if we should inline the
181 // function call or not.
183 InlineCost InlineCostAnalyzer::getInlineCost(CallSite CS,
184 SmallPtrSet<const Function *, 16> &NeverInline) {
185 Instruction *TheCall = CS.getInstruction();
186 Function *Callee = CS.getCalledFunction();
187 Function *Caller = TheCall->getParent()->getParent();
189 // Don't inline functions which can be redefined at link-time to mean
191 if (Callee->mayBeOverridden() ||
192 // Don't inline functions marked noinline.
193 Callee->hasFnAttr(Attribute::NoInline) || NeverInline.count(Callee))
194 return llvm::InlineCost::getNever();
196 // InlineCost - This value measures how good of an inline candidate this call
197 // site is to inline. A lower inline cost make is more likely for the call to
198 // be inlined. This value may go negative.
202 // If there is only one call of the function, and it has internal linkage,
203 // make it almost guaranteed to be inlined.
205 if (Callee->hasLocalLinkage() && Callee->hasOneUse())
208 // If this function uses the coldcc calling convention, prefer not to inline
210 if (Callee->getCallingConv() == CallingConv::Cold)
213 // If the instruction after the call, or if the normal destination of the
214 // invoke is an unreachable instruction, the function is noreturn. As such,
215 // there is little point in inlining this.
216 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) {
217 if (isa<UnreachableInst>(II->getNormalDest()->begin()))
219 } else if (isa<UnreachableInst>(++BasicBlock::iterator(TheCall)))
222 // Get information about the callee...
223 FunctionInfo &CalleeFI = CachedFunctionInfo[Callee];
225 // If we haven't calculated this information yet, do so now.
226 if (CalleeFI.NumBlocks == 0)
227 CalleeFI.analyzeFunction(Callee);
229 // If we should never inline this, return a huge cost.
230 if (CalleeFI.NeverInline)
231 return InlineCost::getNever();
233 // FIXME: It would be nice to kill off CalleeFI.NeverInline. Then we
234 // could move this up and avoid computing the FunctionInfo for
235 // things we are going to just return always inline for. This
236 // requires handling setjmp somewhere else, however.
237 if (!Callee->isDeclaration() && Callee->hasFnAttr(Attribute::AlwaysInline))
238 return InlineCost::getAlways();
240 if (CalleeFI.usesDynamicAlloca) {
241 // Get infomation about the caller...
242 FunctionInfo &CallerFI = CachedFunctionInfo[Caller];
244 // If we haven't calculated this information yet, do so now.
245 if (CallerFI.NumBlocks == 0)
246 CallerFI.analyzeFunction(Caller);
248 // Don't inline a callee with dynamic alloca into a caller without them.
249 // Functions containing dynamic alloca's are inefficient in various ways;
250 // don't create more inefficiency.
251 if (!CallerFI.usesDynamicAlloca)
252 return InlineCost::getNever();
255 // Add to the inline quality for properties that make the call valuable to
256 // inline. This includes factors that indicate that the result of inlining
257 // the function will be optimizable. Currently this just looks at arguments
258 // passed into the function.
261 for (CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end();
262 I != E; ++I, ++ArgNo) {
263 // Each argument passed in has a cost at both the caller and the callee
264 // sides. This favors functions that take many arguments over functions
265 // that take few arguments.
268 // If this is a function being passed in, it is very likely that we will be
269 // able to turn an indirect function call into a direct function call.
270 if (isa<Function>(I))
273 // If an alloca is passed in, inlining this function is likely to allow
274 // significant future optimization possibilities (like scalar promotion, and
275 // scalarization), so encourage the inlining of the function.
277 else if (isa<AllocaInst>(I)) {
278 if (ArgNo < CalleeFI.ArgumentWeights.size())
279 InlineCost -= CalleeFI.ArgumentWeights[ArgNo].AllocaWeight;
281 // If this is a constant being passed into the function, use the argument
282 // weights calculated for the callee to determine how much will be folded
283 // away with this information.
284 } else if (isa<Constant>(I)) {
285 if (ArgNo < CalleeFI.ArgumentWeights.size())
286 InlineCost -= CalleeFI.ArgumentWeights[ArgNo].ConstantWeight;
290 // Now that we have considered all of the factors that make the call site more
291 // likely to be inlined, look at factors that make us not want to inline it.
293 // Don't inline into something too big, which would make it bigger.
295 InlineCost += Caller->size()/15;
297 // Look at the size of the callee. Each instruction counts as 5.
298 InlineCost += CalleeFI.NumInsts*5;
300 return llvm::InlineCost::get(InlineCost);
303 // getInlineFudgeFactor - Return a > 1.0 factor if the inliner should use a
304 // higher threshold to determine if the function call should be inlined.
305 float InlineCostAnalyzer::getInlineFudgeFactor(CallSite CS) {
306 Function *Callee = CS.getCalledFunction();
308 // Get information about the callee...
309 FunctionInfo &CalleeFI = CachedFunctionInfo[Callee];
311 // If we haven't calculated this information yet, do so now.
312 if (CalleeFI.NumBlocks == 0)
313 CalleeFI.analyzeFunction(Callee);
316 // Single BB functions are often written to be inlined.
317 if (CalleeFI.NumBlocks == 1)
320 // Be more aggressive if the function contains a good chunk (if it mades up
321 // at least 10% of the instructions) of vector instructions.
322 if (CalleeFI.NumVectorInsts > CalleeFI.NumInsts/2)
324 else if (CalleeFI.NumVectorInsts > CalleeFI.NumInsts/10)