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){
29 if (isa<BranchInst>(U) || isa<SwitchInst>(U)) {
30 // We will be able to eliminate all but one of the successors.
31 const TerminatorInst &TI = cast<TerminatorInst>(*U);
32 const unsigned NumSucc = TI.getNumSuccessors();
34 for (unsigned I = 0; I != NumSucc; ++I)
35 Instrs += Metrics.NumBBInsts[TI.getSuccessor(I)];
36 // We don't know which blocks will be eliminated, so use the average size.
37 Reduction += InlineConstants::InstrCost*Instrs*(NumSucc-1)/NumSucc;
38 } else if (CallInst *CI = dyn_cast<CallInst>(U)) {
39 // Turning an indirect call into a direct call is a BIG win
40 if (CI->getCalledValue() == V)
41 Reduction += InlineConstants::IndirectCallBonus;
42 } else if (InvokeInst *II = dyn_cast<InvokeInst>(U)) {
43 // Turning an indirect call into a direct call is a BIG win
44 if (II->getCalledValue() == V)
45 Reduction += InlineConstants::IndirectCallBonus;
47 // Figure out if this instruction will be removed due to simple constant
49 Instruction &Inst = cast<Instruction>(*U);
51 // We can't constant propagate instructions which have effects or
54 // FIXME: It would be nice to capture the fact that a load from a
55 // pointer-to-constant-global is actually a *really* good thing to zap.
56 // Unfortunately, we don't know the pointer that may get propagated here,
57 // so we can't make this decision.
58 if (Inst.mayReadFromMemory() || Inst.mayHaveSideEffects() ||
59 isa<AllocaInst>(Inst))
62 bool AllOperandsConstant = true;
63 for (unsigned i = 0, e = Inst.getNumOperands(); i != e; ++i)
64 if (!isa<Constant>(Inst.getOperand(i)) && Inst.getOperand(i) != V) {
65 AllOperandsConstant = false;
69 if (AllOperandsConstant) {
70 // We will get to remove this instruction...
71 Reduction += InlineConstants::InstrCost;
73 // And any other instructions that use it which become constants
75 Reduction += CountCodeReductionForConstant(&Inst);
82 // CountCodeReductionForAlloca - Figure out an approximation of how much smaller
83 // the function will be if it is inlined into a context where an argument
86 unsigned InlineCostAnalyzer::FunctionInfo::
87 CountCodeReductionForAlloca(Value *V) {
88 if (!V->getType()->isPointerTy()) return 0; // Not a pointer
89 unsigned Reduction = 0;
90 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E;++UI){
91 Instruction *I = cast<Instruction>(*UI);
92 if (isa<LoadInst>(I) || isa<StoreInst>(I))
93 Reduction += InlineConstants::InstrCost;
94 else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(I)) {
95 // If the GEP has variable indices, we won't be able to do much with it.
96 if (GEP->hasAllConstantIndices())
97 Reduction += CountCodeReductionForAlloca(GEP);
98 } else if (BitCastInst *BCI = dyn_cast<BitCastInst>(I)) {
99 // Track pointer through bitcasts.
100 Reduction += CountCodeReductionForAlloca(BCI);
102 // If there is some other strange instruction, we're not going to be able
103 // to do much if we inline this.
111 // callIsSmall - If a call is likely to lower to a single target instruction, or
112 // is otherwise deemed small return true.
113 // TODO: Perhaps calls like memcpy, strcpy, etc?
114 static bool callIsSmall(const Function *F) {
115 if (!F) return false;
117 if (F->hasLocalLinkage()) return false;
119 if (!F->hasName()) return false;
121 StringRef Name = F->getName();
123 // These will all likely lower to a single selection DAG node.
124 if (Name == "copysign" || Name == "copysignf" || Name == "copysignl" ||
125 Name == "fabs" || Name == "fabsf" || Name == "fabsl" ||
126 Name == "sin" || Name == "sinf" || Name == "sinl" ||
127 Name == "cos" || Name == "cosf" || Name == "cosl" ||
128 Name == "sqrt" || Name == "sqrtf" || Name == "sqrtl" )
131 // These are all likely to be optimized into something smaller.
132 if (Name == "pow" || Name == "powf" || Name == "powl" ||
133 Name == "exp2" || Name == "exp2l" || Name == "exp2f" ||
134 Name == "floor" || Name == "floorf" || Name == "ceil" ||
135 Name == "round" || Name == "ffs" || Name == "ffsl" ||
136 Name == "abs" || Name == "labs" || Name == "llabs")
142 /// analyzeBasicBlock - Fill in the current structure with information gleaned
143 /// from the specified block.
144 void CodeMetrics::analyzeBasicBlock(const BasicBlock *BB) {
146 unsigned NumInstsBeforeThisBB = NumInsts;
147 for (BasicBlock::const_iterator II = BB->begin(), E = BB->end();
149 if (isa<PHINode>(II)) continue; // PHI nodes don't count.
151 // Special handling for calls.
152 if (isa<CallInst>(II) || isa<InvokeInst>(II)) {
153 if (isa<DbgInfoIntrinsic>(II))
154 continue; // Debug intrinsics don't count as size.
156 CallSite CS = CallSite::get(const_cast<Instruction*>(&*II));
158 // If this function contains a call to setjmp or _setjmp, never inline
159 // it. This is a hack because we depend on the user marking their local
160 // variables as volatile if they are live across a setjmp call, and they
161 // probably won't do this in callers.
162 if (Function *F = CS.getCalledFunction())
163 if (F->isDeclaration() &&
164 (F->getName() == "setjmp" || F->getName() == "_setjmp"))
167 if (!isa<IntrinsicInst>(II) && !callIsSmall(CS.getCalledFunction())) {
168 // Each argument to a call takes on average one instruction to set up.
169 NumInsts += CS.arg_size();
174 if (const AllocaInst *AI = dyn_cast<AllocaInst>(II)) {
175 if (!AI->isStaticAlloca())
176 this->usesDynamicAlloca = true;
179 if (isa<ExtractElementInst>(II) || II->getType()->isVectorTy())
182 if (const CastInst *CI = dyn_cast<CastInst>(II)) {
183 // Noop casts, including ptr <-> int, don't count.
184 if (CI->isLosslessCast() || isa<IntToPtrInst>(CI) ||
185 isa<PtrToIntInst>(CI))
187 // Result of a cmp instruction is often extended (to be used by other
188 // cmp instructions, logical or return instructions). These are usually
189 // nop on most sane targets.
190 if (isa<CmpInst>(CI->getOperand(0)))
192 } else if (const GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(II)){
193 // If a GEP has all constant indices, it will probably be folded with
195 if (GEPI->hasAllConstantIndices())
202 if (isa<ReturnInst>(BB->getTerminator()))
205 // We never want to inline functions that contain an indirectbr. This is
206 // incorrect because all the blockaddress's (in static global initializers
207 // for example) would be referring to the original function, and this indirect
208 // jump would jump from the inlined copy of the function into the original
209 // function which is extremely undefined behavior.
210 if (isa<IndirectBrInst>(BB->getTerminator()))
213 // Remember NumInsts for this BB.
214 NumBBInsts[BB] = NumInsts - NumInstsBeforeThisBB;
217 /// analyzeFunction - Fill in the current structure with information gleaned
218 /// from the specified function.
219 void CodeMetrics::analyzeFunction(Function *F) {
220 // Look at the size of the callee.
221 for (Function::const_iterator BB = F->begin(), E = F->end(); BB != E; ++BB)
222 analyzeBasicBlock(&*BB);
225 /// analyzeFunction - Fill in the current structure with information gleaned
226 /// from the specified function.
227 void InlineCostAnalyzer::FunctionInfo::analyzeFunction(Function *F) {
228 Metrics.analyzeFunction(F);
230 // A function with exactly one return has it removed during the inlining
231 // process (see InlineFunction), so don't count it.
232 // FIXME: This knowledge should really be encoded outside of FunctionInfo.
233 if (Metrics.NumRets==1)
236 // Don't bother calculating argument weights if we are never going to inline
237 // the function anyway.
238 if (Metrics.NeverInline)
241 // Check out all of the arguments to the function, figuring out how much
242 // code can be eliminated if one of the arguments is a constant.
243 ArgumentWeights.reserve(F->arg_size());
244 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(); I != E; ++I)
245 ArgumentWeights.push_back(ArgInfo(CountCodeReductionForConstant(I),
246 CountCodeReductionForAlloca(I)));
249 // getInlineCost - The heuristic used to determine if we should inline the
250 // function call or not.
252 InlineCost InlineCostAnalyzer::getInlineCost(CallSite CS,
253 SmallPtrSet<const Function *, 16> &NeverInline) {
254 Instruction *TheCall = CS.getInstruction();
255 Function *Callee = CS.getCalledFunction();
256 Function *Caller = TheCall->getParent()->getParent();
258 // Don't inline functions which can be redefined at link-time to mean
259 // something else. Don't inline functions marked noinline or call sites
261 if (Callee->mayBeOverridden() ||
262 Callee->hasFnAttr(Attribute::NoInline) || NeverInline.count(Callee) ||
264 return llvm::InlineCost::getNever();
266 // InlineCost - This value measures how good of an inline candidate this call
267 // site is to inline. A lower inline cost make is more likely for the call to
268 // be inlined. This value may go negative.
272 // If there is only one call of the function, and it has internal linkage,
273 // make it almost guaranteed to be inlined.
275 if (Callee->hasLocalLinkage() && Callee->hasOneUse())
276 InlineCost += InlineConstants::LastCallToStaticBonus;
278 // If this function uses the coldcc calling convention, prefer not to inline
280 if (Callee->getCallingConv() == CallingConv::Cold)
281 InlineCost += InlineConstants::ColdccPenalty;
283 // If the instruction after the call, or if the normal destination of the
284 // invoke is an unreachable instruction, the function is noreturn. As such,
285 // there is little point in inlining this.
286 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) {
287 if (isa<UnreachableInst>(II->getNormalDest()->begin()))
288 InlineCost += InlineConstants::NoreturnPenalty;
289 } else if (isa<UnreachableInst>(++BasicBlock::iterator(TheCall)))
290 InlineCost += InlineConstants::NoreturnPenalty;
292 // Get information about the callee...
293 FunctionInfo &CalleeFI = CachedFunctionInfo[Callee];
295 // If we haven't calculated this information yet, do so now.
296 if (CalleeFI.Metrics.NumBlocks == 0)
297 CalleeFI.analyzeFunction(Callee);
299 // If we should never inline this, return a huge cost.
300 if (CalleeFI.Metrics.NeverInline)
301 return InlineCost::getNever();
303 // FIXME: It would be nice to kill off CalleeFI.NeverInline. Then we
304 // could move this up and avoid computing the FunctionInfo for
305 // things we are going to just return always inline for. This
306 // requires handling setjmp somewhere else, however.
307 if (!Callee->isDeclaration() && Callee->hasFnAttr(Attribute::AlwaysInline))
308 return InlineCost::getAlways();
310 if (CalleeFI.Metrics.usesDynamicAlloca) {
311 // Get infomation about the caller...
312 FunctionInfo &CallerFI = CachedFunctionInfo[Caller];
314 // If we haven't calculated this information yet, do so now.
315 if (CallerFI.Metrics.NumBlocks == 0)
316 CallerFI.analyzeFunction(Caller);
318 // Don't inline a callee with dynamic alloca into a caller without them.
319 // Functions containing dynamic alloca's are inefficient in various ways;
320 // don't create more inefficiency.
321 if (!CallerFI.Metrics.usesDynamicAlloca)
322 return InlineCost::getNever();
325 // Add to the inline quality for properties that make the call valuable to
326 // inline. This includes factors that indicate that the result of inlining
327 // the function will be optimizable. Currently this just looks at arguments
328 // passed into the function.
331 for (CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end();
332 I != E; ++I, ++ArgNo) {
333 // Each argument passed in has a cost at both the caller and the callee
334 // sides. Measurements show that each argument costs about the same as an
336 InlineCost -= InlineConstants::InstrCost;
338 // If an alloca is passed in, inlining this function is likely to allow
339 // significant future optimization possibilities (like scalar promotion, and
340 // scalarization), so encourage the inlining of the function.
342 if (isa<AllocaInst>(I)) {
343 if (ArgNo < CalleeFI.ArgumentWeights.size())
344 InlineCost -= CalleeFI.ArgumentWeights[ArgNo].AllocaWeight;
346 // If this is a constant being passed into the function, use the argument
347 // weights calculated for the callee to determine how much will be folded
348 // away with this information.
349 } else if (isa<Constant>(I)) {
350 if (ArgNo < CalleeFI.ArgumentWeights.size())
351 InlineCost -= CalleeFI.ArgumentWeights[ArgNo].ConstantWeight;
355 // Now that we have considered all of the factors that make the call site more
356 // likely to be inlined, look at factors that make us not want to inline it.
358 // Calls usually take a long time, so they make the inlining gain smaller.
359 InlineCost += CalleeFI.Metrics.NumCalls * InlineConstants::CallPenalty;
361 // Look at the size of the callee. Each instruction counts as 5.
362 InlineCost += CalleeFI.Metrics.NumInsts*InlineConstants::InstrCost;
364 return llvm::InlineCost::get(InlineCost);
367 // getInlineFudgeFactor - Return a > 1.0 factor if the inliner should use a
368 // higher threshold to determine if the function call should be inlined.
369 float InlineCostAnalyzer::getInlineFudgeFactor(CallSite CS) {
370 Function *Callee = CS.getCalledFunction();
372 // Get information about the callee...
373 FunctionInfo &CalleeFI = CachedFunctionInfo[Callee];
375 // If we haven't calculated this information yet, do so now.
376 if (CalleeFI.Metrics.NumBlocks == 0)
377 CalleeFI.analyzeFunction(Callee);
380 // Single BB functions are often written to be inlined.
381 if (CalleeFI.Metrics.NumBlocks == 1)
384 // Be more aggressive if the function contains a good chunk (if it mades up
385 // at least 10% of the instructions) of vector instructions.
386 if (CalleeFI.Metrics.NumVectorInsts > CalleeFI.Metrics.NumInsts/2)
388 else if (CalleeFI.Metrics.NumVectorInsts > CalleeFI.Metrics.NumInsts/10)
393 /// growCachedCostInfo - update the cached cost info for Caller after Callee has
396 InlineCostAnalyzer::growCachedCostInfo(Function* Caller, Function* Callee) {
397 FunctionInfo &CallerFI = CachedFunctionInfo[Caller];
399 // For small functions we prefer to recalculate the cost for better accuracy.
400 if (CallerFI.Metrics.NumBlocks < 10 || CallerFI.Metrics.NumInsts < 1000) {
401 resetCachedCostInfo(Caller);
405 // For large functions, we can save a lot of computation time by skipping
407 if (CallerFI.Metrics.NumCalls > 0)
408 --CallerFI.Metrics.NumCalls;
411 FunctionInfo &CalleeFI = CachedFunctionInfo[Callee];
412 if (!CalleeFI.Metrics.NumBlocks) {
413 resetCachedCostInfo(Caller);
416 CallerFI.Metrics.NeverInline |= CalleeFI.Metrics.NeverInline;
417 CallerFI.Metrics.usesDynamicAlloca |= CalleeFI.Metrics.usesDynamicAlloca;
419 CallerFI.Metrics.NumInsts += CalleeFI.Metrics.NumInsts;
420 CallerFI.Metrics.NumBlocks += CalleeFI.Metrics.NumBlocks;
421 CallerFI.Metrics.NumCalls += CalleeFI.Metrics.NumCalls;
422 CallerFI.Metrics.NumVectorInsts += CalleeFI.Metrics.NumVectorInsts;
423 CallerFI.Metrics.NumRets += CalleeFI.Metrics.NumRets;
425 // analyzeBasicBlock counts each function argument as an inst.
426 if (CallerFI.Metrics.NumInsts >= Callee->arg_size())
427 CallerFI.Metrics.NumInsts -= Callee->arg_size();
429 CallerFI.Metrics.NumInsts = 0;
431 // We are not updating the argumentweights. We have already determined that
432 // Caller is a fairly large function, so we accept the loss of precision.