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/Analysis/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<AllocaInst>(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 // callIsSmall - If a call is likely to lower to a single target instruction, or
103 // is otherwise deemed small return true.
104 // TODO: Perhaps calls like memcpy, strcpy, etc?
105 static bool callIsSmall(const Function *F) {
106 if (!F) return false;
108 if (F->hasLocalLinkage()) return false;
110 if (!F->hasName()) return false;
112 StringRef Name = F->getName();
114 // These will all likely lower to a single selection DAG node.
115 if (Name == "copysign" || Name == "copysignf" ||
116 Name == "fabs" || Name == "fabsf" || Name == "fabsl" ||
117 Name == "sin" || Name == "sinf" || Name == "sinl" ||
118 Name == "cos" || Name == "cosf" || Name == "cosl" ||
119 Name == "sqrt" || Name == "sqrtf" || Name == "sqrtl" )
122 // These are all likely to be optimized into something smaller.
123 if (Name == "pow" || Name == "powf" || Name == "powl" ||
124 Name == "exp2" || Name == "exp2l" || Name == "exp2f" ||
125 Name == "floor" || Name == "floorf" || Name == "ceil" ||
126 Name == "round" || Name == "ffs" || Name == "ffsl" ||
127 Name == "abs" || Name == "labs" || Name == "llabs")
133 /// analyzeBasicBlock - Fill in the current structure with information gleaned
134 /// from the specified block.
135 void CodeMetrics::analyzeBasicBlock(const BasicBlock *BB) {
138 for (BasicBlock::const_iterator II = BB->begin(), E = BB->end();
140 if (isa<PHINode>(II)) continue; // PHI nodes don't count.
142 // Special handling for calls.
143 if (isa<CallInst>(II) || isa<InvokeInst>(II)) {
144 if (isa<DbgInfoIntrinsic>(II))
145 continue; // Debug intrinsics don't count as size.
147 CallSite CS = CallSite::get(const_cast<Instruction*>(&*II));
149 // If this function contains a call to setjmp or _setjmp, never inline
150 // it. This is a hack because we depend on the user marking their local
151 // variables as volatile if they are live across a setjmp call, and they
152 // probably won't do this in callers.
153 if (Function *F = CS.getCalledFunction())
154 if (F->isDeclaration() &&
155 (F->getName() == "setjmp" || F->getName() == "_setjmp"))
158 // Calls often compile into many machine instructions. Bump up their
159 // cost to reflect this.
160 if (!isa<IntrinsicInst>(II) && !callIsSmall(CS.getCalledFunction()))
161 NumInsts += InlineConstants::CallPenalty;
164 if (const AllocaInst *AI = dyn_cast<AllocaInst>(II)) {
165 if (!AI->isStaticAlloca())
166 this->usesDynamicAlloca = true;
169 if (isa<ExtractElementInst>(II) || isa<VectorType>(II->getType()))
172 if (const CastInst *CI = dyn_cast<CastInst>(II)) {
173 // Noop casts, including ptr <-> int, don't count.
174 if (CI->isLosslessCast() || isa<IntToPtrInst>(CI) ||
175 isa<PtrToIntInst>(CI))
177 // Result of a cmp instruction is often extended (to be used by other
178 // cmp instructions, logical or return instructions). These are usually
179 // nop on most sane targets.
180 if (isa<CmpInst>(CI->getOperand(0)))
182 } else if (const GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(II)){
183 // If a GEP has all constant indices, it will probably be folded with
185 if (GEPI->hasAllConstantIndices())
192 if (isa<ReturnInst>(BB->getTerminator()))
195 // We never want to inline functions that contain an indirectbr. This is
196 // incorrect because all the blockaddress's (in static global initializers
197 // for example) would be referring to the original function, and this indirect
198 // jump would jump from the inlined copy of the function into the original
199 // function which is extremely undefined behavior.
200 if (isa<IndirectBrInst>(BB->getTerminator()))
204 /// analyzeFunction - Fill in the current structure with information gleaned
205 /// from the specified function.
206 void CodeMetrics::analyzeFunction(Function *F) {
207 // Look at the size of the callee.
208 for (Function::const_iterator BB = F->begin(), E = F->end(); BB != E; ++BB)
209 analyzeBasicBlock(&*BB);
212 /// analyzeFunction - Fill in the current structure with information gleaned
213 /// from the specified function.
214 void InlineCostAnalyzer::FunctionInfo::analyzeFunction(Function *F) {
215 Metrics.analyzeFunction(F);
217 // A function with exactly one return has it removed during the inlining
218 // process (see InlineFunction), so don't count it.
219 // FIXME: This knowledge should really be encoded outside of FunctionInfo.
220 if (Metrics.NumRets==1)
223 // Don't bother calculating argument weights if we are never going to inline
224 // the function anyway.
225 if (Metrics.NeverInline)
228 // Check out all of the arguments to the function, figuring out how much
229 // code can be eliminated if one of the arguments is a constant.
230 ArgumentWeights.reserve(F->arg_size());
231 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(); I != E; ++I)
232 ArgumentWeights.push_back(ArgInfo(CountCodeReductionForConstant(I),
233 CountCodeReductionForAlloca(I)));
236 // getInlineCost - The heuristic used to determine if we should inline the
237 // function call or not.
239 InlineCost InlineCostAnalyzer::getInlineCost(CallSite CS,
240 SmallPtrSet<const Function *, 16> &NeverInline) {
241 Instruction *TheCall = CS.getInstruction();
242 Function *Callee = CS.getCalledFunction();
243 Function *Caller = TheCall->getParent()->getParent();
245 // Don't inline functions which can be redefined at link-time to mean
246 // something else. Don't inline functions marked noinline.
247 if (Callee->mayBeOverridden() ||
248 Callee->hasFnAttr(Attribute::NoInline) || NeverInline.count(Callee))
249 return llvm::InlineCost::getNever();
251 // InlineCost - This value measures how good of an inline candidate this call
252 // site is to inline. A lower inline cost make is more likely for the call to
253 // be inlined. This value may go negative.
257 // If there is only one call of the function, and it has internal linkage,
258 // make it almost guaranteed to be inlined.
260 if (Callee->hasLocalLinkage() && Callee->hasOneUse())
261 InlineCost += InlineConstants::LastCallToStaticBonus;
263 // If this function uses the coldcc calling convention, prefer not to inline
265 if (Callee->getCallingConv() == CallingConv::Cold)
266 InlineCost += InlineConstants::ColdccPenalty;
268 // If the instruction after the call, or if the normal destination of the
269 // invoke is an unreachable instruction, the function is noreturn. As such,
270 // there is little point in inlining this.
271 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) {
272 if (isa<UnreachableInst>(II->getNormalDest()->begin()))
273 InlineCost += InlineConstants::NoreturnPenalty;
274 } else if (isa<UnreachableInst>(++BasicBlock::iterator(TheCall)))
275 InlineCost += InlineConstants::NoreturnPenalty;
277 // Get information about the callee...
278 FunctionInfo &CalleeFI = CachedFunctionInfo[Callee];
280 // If we haven't calculated this information yet, do so now.
281 if (CalleeFI.Metrics.NumBlocks == 0)
282 CalleeFI.analyzeFunction(Callee);
284 // If we should never inline this, return a huge cost.
285 if (CalleeFI.Metrics.NeverInline)
286 return InlineCost::getNever();
288 // FIXME: It would be nice to kill off CalleeFI.NeverInline. Then we
289 // could move this up and avoid computing the FunctionInfo for
290 // things we are going to just return always inline for. This
291 // requires handling setjmp somewhere else, however.
292 if (!Callee->isDeclaration() && Callee->hasFnAttr(Attribute::AlwaysInline))
293 return InlineCost::getAlways();
295 if (CalleeFI.Metrics.usesDynamicAlloca) {
296 // Get infomation about the caller...
297 FunctionInfo &CallerFI = CachedFunctionInfo[Caller];
299 // If we haven't calculated this information yet, do so now.
300 if (CallerFI.Metrics.NumBlocks == 0)
301 CallerFI.analyzeFunction(Caller);
303 // Don't inline a callee with dynamic alloca into a caller without them.
304 // Functions containing dynamic alloca's are inefficient in various ways;
305 // don't create more inefficiency.
306 if (!CallerFI.Metrics.usesDynamicAlloca)
307 return InlineCost::getNever();
310 // Add to the inline quality for properties that make the call valuable to
311 // inline. This includes factors that indicate that the result of inlining
312 // the function will be optimizable. Currently this just looks at arguments
313 // passed into the function.
316 for (CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end();
317 I != E; ++I, ++ArgNo) {
318 // Each argument passed in has a cost at both the caller and the callee
319 // sides. This favors functions that take many arguments over functions
320 // that take few arguments.
323 // If this is a function being passed in, it is very likely that we will be
324 // able to turn an indirect function call into a direct function call.
325 if (isa<Function>(I))
328 // If an alloca is passed in, inlining this function is likely to allow
329 // significant future optimization possibilities (like scalar promotion, and
330 // scalarization), so encourage the inlining of the function.
332 else if (isa<AllocaInst>(I)) {
333 if (ArgNo < CalleeFI.ArgumentWeights.size())
334 InlineCost -= CalleeFI.ArgumentWeights[ArgNo].AllocaWeight;
336 // If this is a constant being passed into the function, use the argument
337 // weights calculated for the callee to determine how much will be folded
338 // away with this information.
339 } else if (isa<Constant>(I)) {
340 if (ArgNo < CalleeFI.ArgumentWeights.size())
341 InlineCost -= CalleeFI.ArgumentWeights[ArgNo].ConstantWeight;
345 // Now that we have considered all of the factors that make the call site more
346 // likely to be inlined, look at factors that make us not want to inline it.
348 // Don't inline into something too big, which would make it bigger.
349 // "size" here is the number of basic blocks, not instructions.
351 InlineCost += Caller->size()/15;
353 // Look at the size of the callee. Each instruction counts as 5.
354 InlineCost += CalleeFI.Metrics.NumInsts*5;
356 return llvm::InlineCost::get(InlineCost);
359 // getInlineFudgeFactor - Return a > 1.0 factor if the inliner should use a
360 // higher threshold to determine if the function call should be inlined.
361 float InlineCostAnalyzer::getInlineFudgeFactor(CallSite CS) {
362 Function *Callee = CS.getCalledFunction();
364 // Get information about the callee...
365 FunctionInfo &CalleeFI = CachedFunctionInfo[Callee];
367 // If we haven't calculated this information yet, do so now.
368 if (CalleeFI.Metrics.NumBlocks == 0)
369 CalleeFI.analyzeFunction(Callee);
372 // Single BB functions are often written to be inlined.
373 if (CalleeFI.Metrics.NumBlocks == 1)
376 // Be more aggressive if the function contains a good chunk (if it mades up
377 // at least 10% of the instructions) of vector instructions.
378 if (CalleeFI.Metrics.NumVectorInsts > CalleeFI.Metrics.NumInsts/2)
380 else if (CalleeFI.Metrics.NumVectorInsts > CalleeFI.Metrics.NumInsts/10)