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, CodeMetrics &Metrics) {
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) || isa<SwitchInst>(*UI)) {
29 // We will be able to eliminate all but one of the successors.
30 const TerminatorInst &TI = cast<TerminatorInst>(**UI);
31 const unsigned NumSucc = TI.getNumSuccessors();
33 for (unsigned I = 0; I != NumSucc; ++I)
34 Instrs += Metrics.NumBBInsts[TI.getSuccessor(I)];
35 // We don't know which blocks will be eliminated, so use the average size.
36 Reduction += InlineConstants::InstrCost*Instrs*(NumSucc-1)/NumSucc;
37 } else if (CallInst *CI = dyn_cast<CallInst>(*UI)) {
38 // Turning an indirect call into a direct call is a BIG win
39 if (CI->getCalledValue() == V)
40 Reduction += InlineConstants::IndirectCallBonus;
41 } else if (InvokeInst *II = dyn_cast<InvokeInst>(*UI)) {
42 // Turning an indirect call into a direct call is a BIG win
43 if (II->getCalledValue() == V)
44 Reduction += InlineConstants::IndirectCallBonus;
46 // Figure out if this instruction will be removed due to simple constant
48 Instruction &Inst = cast<Instruction>(**UI);
50 // We can't constant propagate instructions which have effects or
53 // FIXME: It would be nice to capture the fact that a load from a
54 // pointer-to-constant-global is actually a *really* good thing to zap.
55 // Unfortunately, we don't know the pointer that may get propagated here,
56 // so we can't make this decision.
57 if (Inst.mayReadFromMemory() || Inst.mayHaveSideEffects() ||
58 isa<AllocaInst>(Inst))
61 bool AllOperandsConstant = true;
62 for (unsigned i = 0, e = Inst.getNumOperands(); i != e; ++i)
63 if (!isa<Constant>(Inst.getOperand(i)) && Inst.getOperand(i) != V) {
64 AllOperandsConstant = false;
68 if (AllOperandsConstant) {
69 // We will get to remove this instruction...
70 Reduction += InlineConstants::InstrCost;
72 // And any other instructions that use it which become constants
74 Reduction += CountCodeReductionForConstant(&Inst, Metrics);
81 // CountCodeReductionForAlloca - Figure out an approximation of how much smaller
82 // the function will be if it is inlined into a context where an argument
85 unsigned InlineCostAnalyzer::FunctionInfo::
86 CountCodeReductionForAlloca(Value *V) {
87 if (!V->getType()->isPointerTy()) return 0; // Not a pointer
88 unsigned Reduction = 0;
89 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E;++UI){
90 Instruction *I = cast<Instruction>(*UI);
91 if (isa<LoadInst>(I) || isa<StoreInst>(I))
92 Reduction += InlineConstants::InstrCost;
93 else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(I)) {
94 // If the GEP has variable indices, we won't be able to do much with it.
95 if (GEP->hasAllConstantIndices())
96 Reduction += CountCodeReductionForAlloca(GEP);
97 } else if (BitCastInst *BCI = dyn_cast<BitCastInst>(I)) {
98 // Track pointer through bitcasts.
99 Reduction += CountCodeReductionForAlloca(BCI);
101 // If there is some other strange instruction, we're not going to be able
102 // to do much if we inline this.
110 // callIsSmall - If a call is likely to lower to a single target instruction, or
111 // is otherwise deemed small return true.
112 // TODO: Perhaps calls like memcpy, strcpy, etc?
113 static bool callIsSmall(const Function *F) {
114 if (!F) return false;
116 if (F->hasLocalLinkage()) return false;
118 if (!F->hasName()) return false;
120 StringRef Name = F->getName();
122 // These will all likely lower to a single selection DAG node.
123 if (Name == "copysign" || Name == "copysignf" ||
124 Name == "fabs" || Name == "fabsf" || Name == "fabsl" ||
125 Name == "sin" || Name == "sinf" || Name == "sinl" ||
126 Name == "cos" || Name == "cosf" || Name == "cosl" ||
127 Name == "sqrt" || Name == "sqrtf" || Name == "sqrtl" )
130 // These are all likely to be optimized into something smaller.
131 if (Name == "pow" || Name == "powf" || Name == "powl" ||
132 Name == "exp2" || Name == "exp2l" || Name == "exp2f" ||
133 Name == "floor" || Name == "floorf" || Name == "ceil" ||
134 Name == "round" || Name == "ffs" || Name == "ffsl" ||
135 Name == "abs" || Name == "labs" || Name == "llabs")
141 /// analyzeBasicBlock - Fill in the current structure with information gleaned
142 /// from the specified block.
143 void CodeMetrics::analyzeBasicBlock(const BasicBlock *BB) {
145 unsigned NumInstsInThisBB = 0;
146 for (BasicBlock::const_iterator II = BB->begin(), E = BB->end();
148 if (isa<PHINode>(II)) continue; // PHI nodes don't count.
150 // Special handling for calls.
151 if (isa<CallInst>(II) || isa<InvokeInst>(II)) {
152 if (isa<DbgInfoIntrinsic>(II))
153 continue; // Debug intrinsics don't count as size.
155 CallSite CS = CallSite::get(const_cast<Instruction*>(&*II));
157 // If this function contains a call to setjmp or _setjmp, never inline
158 // it. This is a hack because we depend on the user marking their local
159 // variables as volatile if they are live across a setjmp call, and they
160 // probably won't do this in callers.
161 if (Function *F = CS.getCalledFunction())
162 if (F->isDeclaration() &&
163 (F->getName() == "setjmp" || F->getName() == "_setjmp"))
166 if (!isa<IntrinsicInst>(II) && !callIsSmall(CS.getCalledFunction())) {
167 // Each argument to a call takes on average one instruction to set up.
168 NumInsts += CS.arg_size();
173 if (const AllocaInst *AI = dyn_cast<AllocaInst>(II)) {
174 if (!AI->isStaticAlloca())
175 this->usesDynamicAlloca = true;
178 if (isa<ExtractElementInst>(II) || II->getType()->isVectorTy())
181 if (const CastInst *CI = dyn_cast<CastInst>(II)) {
182 // Noop casts, including ptr <-> int, don't count.
183 if (CI->isLosslessCast() || isa<IntToPtrInst>(CI) ||
184 isa<PtrToIntInst>(CI))
186 // Result of a cmp instruction is often extended (to be used by other
187 // cmp instructions, logical or return instructions). These are usually
188 // nop on most sane targets.
189 if (isa<CmpInst>(CI->getOperand(0)))
191 } else if (const GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(II)){
192 // If a GEP has all constant indices, it will probably be folded with
194 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] = NumInstsInThisBB;
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)
246 push_back(ArgInfo(CountCodeReductionForConstant(I, Metrics),
247 CountCodeReductionForAlloca(I)));
250 // getInlineCost - The heuristic used to determine if we should inline the
251 // function call or not.
253 InlineCost InlineCostAnalyzer::getInlineCost(CallSite CS,
254 SmallPtrSet<const Function *, 16> &NeverInline) {
255 Instruction *TheCall = CS.getInstruction();
256 Function *Callee = CS.getCalledFunction();
257 Function *Caller = TheCall->getParent()->getParent();
259 // Don't inline functions which can be redefined at link-time to mean
260 // something else. Don't inline functions marked noinline.
261 if (Callee->mayBeOverridden() ||
262 Callee->hasFnAttr(Attribute::NoInline) || NeverInline.count(Callee))
263 return llvm::InlineCost::getNever();
265 // InlineCost - This value measures how good of an inline candidate this call
266 // site is to inline. A lower inline cost make is more likely for the call to
267 // be inlined. This value may go negative.
271 // If there is only one call of the function, and it has internal linkage,
272 // make it almost guaranteed to be inlined.
274 if (Callee->hasLocalLinkage() && Callee->hasOneUse())
275 InlineCost += InlineConstants::LastCallToStaticBonus;
277 // If this function uses the coldcc calling convention, prefer not to inline
279 if (Callee->getCallingConv() == CallingConv::Cold)
280 InlineCost += InlineConstants::ColdccPenalty;
282 // If the instruction after the call, or if the normal destination of the
283 // invoke is an unreachable instruction, the function is noreturn. As such,
284 // there is little point in inlining this.
285 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) {
286 if (isa<UnreachableInst>(II->getNormalDest()->begin()))
287 InlineCost += InlineConstants::NoreturnPenalty;
288 } else if (isa<UnreachableInst>(++BasicBlock::iterator(TheCall)))
289 InlineCost += InlineConstants::NoreturnPenalty;
291 // Get information about the callee...
292 FunctionInfo &CalleeFI = CachedFunctionInfo[Callee];
294 // If we haven't calculated this information yet, do so now.
295 if (CalleeFI.Metrics.NumBlocks == 0)
296 CalleeFI.analyzeFunction(Callee);
298 // If we should never inline this, return a huge cost.
299 if (CalleeFI.Metrics.NeverInline)
300 return InlineCost::getNever();
302 // FIXME: It would be nice to kill off CalleeFI.NeverInline. Then we
303 // could move this up and avoid computing the FunctionInfo for
304 // things we are going to just return always inline for. This
305 // requires handling setjmp somewhere else, however.
306 if (!Callee->isDeclaration() && Callee->hasFnAttr(Attribute::AlwaysInline))
307 return InlineCost::getAlways();
309 if (CalleeFI.Metrics.usesDynamicAlloca) {
310 // Get infomation about the caller...
311 FunctionInfo &CallerFI = CachedFunctionInfo[Caller];
313 // If we haven't calculated this information yet, do so now.
314 if (CallerFI.Metrics.NumBlocks == 0)
315 CallerFI.analyzeFunction(Caller);
317 // Don't inline a callee with dynamic alloca into a caller without them.
318 // Functions containing dynamic alloca's are inefficient in various ways;
319 // don't create more inefficiency.
320 if (!CallerFI.Metrics.usesDynamicAlloca)
321 return InlineCost::getNever();
324 // Add to the inline quality for properties that make the call valuable to
325 // inline. This includes factors that indicate that the result of inlining
326 // the function will be optimizable. Currently this just looks at arguments
327 // passed into the function.
330 for (CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end();
331 I != E; ++I, ++ArgNo) {
332 // Each argument passed in has a cost at both the caller and the callee
333 // sides. Measurements show that each argument costs about the same as an
335 InlineCost -= InlineConstants::InstrCost;
337 // If an alloca is passed in, inlining this function is likely to allow
338 // significant future optimization possibilities (like scalar promotion, and
339 // scalarization), so encourage the inlining of the function.
341 if (isa<AllocaInst>(I)) {
342 if (ArgNo < CalleeFI.ArgumentWeights.size())
343 InlineCost -= CalleeFI.ArgumentWeights[ArgNo].AllocaWeight;
345 // If this is a constant being passed into the function, use the argument
346 // weights calculated for the callee to determine how much will be folded
347 // away with this information.
348 } else if (isa<Constant>(I)) {
349 if (ArgNo < CalleeFI.ArgumentWeights.size())
350 InlineCost -= CalleeFI.ArgumentWeights[ArgNo].ConstantWeight;
354 // Now that we have considered all of the factors that make the call site more
355 // likely to be inlined, look at factors that make us not want to inline it.
357 // Calls usually take a long time, so they make the inlining gain smaller.
358 InlineCost += CalleeFI.Metrics.NumCalls * InlineConstants::CallPenalty;
360 // Look at the size of the callee. Each instruction counts as 5.
361 InlineCost += CalleeFI.Metrics.NumInsts*InlineConstants::InstrCost;
363 return llvm::InlineCost::get(InlineCost);
366 // getInlineFudgeFactor - Return a > 1.0 factor if the inliner should use a
367 // higher threshold to determine if the function call should be inlined.
368 float InlineCostAnalyzer::getInlineFudgeFactor(CallSite CS) {
369 Function *Callee = CS.getCalledFunction();
371 // Get information about the callee...
372 FunctionInfo &CalleeFI = CachedFunctionInfo[Callee];
374 // If we haven't calculated this information yet, do so now.
375 if (CalleeFI.Metrics.NumBlocks == 0)
376 CalleeFI.analyzeFunction(Callee);
379 // Single BB functions are often written to be inlined.
380 if (CalleeFI.Metrics.NumBlocks == 1)
383 // Be more aggressive if the function contains a good chunk (if it mades up
384 // at least 10% of the instructions) of vector instructions.
385 if (CalleeFI.Metrics.NumVectorInsts > CalleeFI.Metrics.NumInsts/2)
387 else if (CalleeFI.Metrics.NumVectorInsts > CalleeFI.Metrics.NumInsts/10)
392 /// growCachedCostInfo - update the cached cost info for Caller after Callee has
395 InlineCostAnalyzer::growCachedCostInfo(Function* Caller, Function* Callee) {
396 FunctionInfo &CallerFI = CachedFunctionInfo[Caller];
398 // For small functions we prefer to recalculate the cost for better accuracy.
399 if (CallerFI.Metrics.NumBlocks < 10 || CallerFI.Metrics.NumInsts < 1000) {
400 resetCachedCostInfo(Caller);
404 // For large functions, we can save a lot of computation time by skipping
406 if (CallerFI.Metrics.NumCalls > 0)
407 --CallerFI.Metrics.NumCalls;
410 FunctionInfo &CalleeFI = CachedFunctionInfo[Callee];
411 if (!CalleeFI.Metrics.NumBlocks) {
412 resetCachedCostInfo(Caller);
415 CallerFI.Metrics.NeverInline |= CalleeFI.Metrics.NeverInline;
416 CallerFI.Metrics.usesDynamicAlloca |= CalleeFI.Metrics.usesDynamicAlloca;
418 CallerFI.Metrics.NumInsts += CalleeFI.Metrics.NumInsts;
419 CallerFI.Metrics.NumBlocks += CalleeFI.Metrics.NumBlocks;
420 CallerFI.Metrics.NumCalls += CalleeFI.Metrics.NumCalls;
421 CallerFI.Metrics.NumVectorInsts += CalleeFI.Metrics.NumVectorInsts;
422 CallerFI.Metrics.NumRets += CalleeFI.Metrics.NumRets;
424 // analyzeBasicBlock counts each function argument as an inst.
425 if (CallerFI.Metrics.NumInsts >= Callee->arg_size())
426 CallerFI.Metrics.NumInsts -= Callee->arg_size();
428 CallerFI.Metrics.NumInsts = 0;
430 // We are not updating the argumentweights. We have already determined that
431 // Caller is a fairly large function, so we accept the loss of precision.