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 (isa<IndirectBrInst>(*UI))
35 // Eliminating an indirect branch is a big win.
37 else if (CallInst *CI = dyn_cast<CallInst>(*UI)) {
38 // Turning an indirect call into a direct call is a BIG win
39 Reduction += CI->getCalledValue() == V ? 500 : 0;
40 } else if (InvokeInst *II = dyn_cast<InvokeInst>(*UI)) {
41 // Turning an indirect call into a direct call is a BIG win
42 Reduction += II->getCalledValue() == V ? 500 : 0;
44 // Figure out if this instruction will be removed due to simple constant
46 Instruction &Inst = cast<Instruction>(**UI);
48 // We can't constant propagate instructions which have effects or
51 // FIXME: It would be nice to capture the fact that a load from a
52 // pointer-to-constant-global is actually a *really* good thing to zap.
53 // Unfortunately, we don't know the pointer that may get propagated here,
54 // so we can't make this decision.
55 if (Inst.mayReadFromMemory() || Inst.mayHaveSideEffects() ||
56 isa<AllocaInst>(Inst))
59 bool AllOperandsConstant = true;
60 for (unsigned i = 0, e = Inst.getNumOperands(); i != e; ++i)
61 if (!isa<Constant>(Inst.getOperand(i)) && Inst.getOperand(i) != V) {
62 AllOperandsConstant = false;
66 if (AllOperandsConstant) {
67 // We will get to remove this instruction...
70 // And any other instructions that use it which become constants
72 Reduction += CountCodeReductionForConstant(&Inst);
79 // CountCodeReductionForAlloca - Figure out an approximation of how much smaller
80 // the function will be if it is inlined into a context where an argument
83 unsigned InlineCostAnalyzer::FunctionInfo::
84 CountCodeReductionForAlloca(Value *V) {
85 if (!isa<PointerType>(V->getType())) return 0; // Not a pointer
86 unsigned Reduction = 0;
87 for (Value::use_iterator UI = V->use_begin(), E = V->use_end(); UI != E;++UI){
88 Instruction *I = cast<Instruction>(*UI);
89 if (isa<LoadInst>(I) || isa<StoreInst>(I))
91 else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(I)) {
92 // If the GEP has variable indices, we won't be able to do much with it.
93 if (!GEP->hasAllConstantIndices())
94 Reduction += CountCodeReductionForAlloca(GEP)+15;
96 // If there is some other strange instruction, we're not going to be able
97 // to do much if we inline this.
105 // callIsSmall - If a call is likely to lower to a single target instruction, or
106 // is otherwise deemed small return true.
107 // TODO: Perhaps calls like memcpy, strcpy, etc?
108 static bool callIsSmall(const Function *F) {
109 if (!F) return false;
111 if (F->hasLocalLinkage()) return false;
113 if (!F->hasName()) return false;
115 StringRef Name = F->getName();
117 // These will all likely lower to a single selection DAG node.
118 if (Name == "copysign" || Name == "copysignf" ||
119 Name == "fabs" || Name == "fabsf" || Name == "fabsl" ||
120 Name == "sin" || Name == "sinf" || Name == "sinl" ||
121 Name == "cos" || Name == "cosf" || Name == "cosl" ||
122 Name == "sqrt" || Name == "sqrtf" || Name == "sqrtl" )
125 // These are all likely to be optimized into something smaller.
126 if (Name == "pow" || Name == "powf" || Name == "powl" ||
127 Name == "exp2" || Name == "exp2l" || Name == "exp2f" ||
128 Name == "floor" || Name == "floorf" || Name == "ceil" ||
129 Name == "round" || Name == "ffs" || Name == "ffsl" ||
130 Name == "abs" || Name == "labs" || Name == "llabs")
136 /// analyzeBasicBlock - Fill in the current structure with information gleaned
137 /// from the specified block.
138 void CodeMetrics::analyzeBasicBlock(const BasicBlock *BB) {
141 for (BasicBlock::const_iterator II = BB->begin(), E = BB->end();
143 if (isa<PHINode>(II)) continue; // PHI nodes don't count.
145 // Special handling for calls.
146 if (isa<CallInst>(II) || isa<InvokeInst>(II)) {
147 if (isa<DbgInfoIntrinsic>(II))
148 continue; // Debug intrinsics don't count as size.
150 CallSite CS = CallSite::get(const_cast<Instruction*>(&*II));
152 // If this function contains a call to setjmp or _setjmp, never inline
153 // it. This is a hack because we depend on the user marking their local
154 // variables as volatile if they are live across a setjmp call, and they
155 // probably won't do this in callers.
156 if (Function *F = CS.getCalledFunction())
157 if (F->isDeclaration() &&
158 (F->getName() == "setjmp" || F->getName() == "_setjmp"))
161 // Calls often compile into many machine instructions. Bump up their
162 // cost to reflect this.
163 if (!isa<IntrinsicInst>(II) && !callIsSmall(CS.getCalledFunction()))
164 NumInsts += InlineConstants::CallPenalty;
167 if (const AllocaInst *AI = dyn_cast<AllocaInst>(II)) {
168 if (!AI->isStaticAlloca())
169 this->usesDynamicAlloca = true;
172 if (isa<ExtractElementInst>(II) || isa<VectorType>(II->getType()))
175 if (const CastInst *CI = dyn_cast<CastInst>(II)) {
176 // Noop casts, including ptr <-> int, don't count.
177 if (CI->isLosslessCast() || isa<IntToPtrInst>(CI) ||
178 isa<PtrToIntInst>(CI))
180 // Result of a cmp instruction is often extended (to be used by other
181 // cmp instructions, logical or return instructions). These are usually
182 // nop on most sane targets.
183 if (isa<CmpInst>(CI->getOperand(0)))
185 } else if (const GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(II)){
186 // If a GEP has all constant indices, it will probably be folded with
188 if (GEPI->hasAllConstantIndices())
195 if (isa<ReturnInst>(BB->getTerminator()))
198 // We never want to inline functions that contain an indirectbr. This is
199 // incorrect because all the blockaddress's (in static global initializers
200 // for example) would be referring to the original function, and this indirect
201 // jump would jump from the inlined copy of the function into the original
202 // function which is extremely undefined behavior.
203 if (isa<IndirectBrInst>(BB->getTerminator()))
207 /// analyzeFunction - Fill in the current structure with information gleaned
208 /// from the specified function.
209 void CodeMetrics::analyzeFunction(Function *F) {
210 // Look at the size of the callee.
211 for (Function::const_iterator BB = F->begin(), E = F->end(); BB != E; ++BB)
212 analyzeBasicBlock(&*BB);
215 /// analyzeFunction - Fill in the current structure with information gleaned
216 /// from the specified function.
217 void InlineCostAnalyzer::FunctionInfo::analyzeFunction(Function *F) {
218 Metrics.analyzeFunction(F);
220 // A function with exactly one return has it removed during the inlining
221 // process (see InlineFunction), so don't count it.
222 // FIXME: This knowledge should really be encoded outside of FunctionInfo.
223 if (Metrics.NumRets==1)
226 // Don't bother calculating argument weights if we are never going to inline
227 // the function anyway.
228 if (Metrics.NeverInline)
231 // Check out all of the arguments to the function, figuring out how much
232 // code can be eliminated if one of the arguments is a constant.
233 ArgumentWeights.reserve(F->arg_size());
234 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(); I != E; ++I)
235 ArgumentWeights.push_back(ArgInfo(CountCodeReductionForConstant(I),
236 CountCodeReductionForAlloca(I)));
239 // getInlineCost - The heuristic used to determine if we should inline the
240 // function call or not.
242 InlineCost InlineCostAnalyzer::getInlineCost(CallSite CS,
243 SmallPtrSet<const Function *, 16> &NeverInline) {
244 Instruction *TheCall = CS.getInstruction();
245 Function *Callee = CS.getCalledFunction();
246 Function *Caller = TheCall->getParent()->getParent();
248 // Don't inline functions which can be redefined at link-time to mean
249 // something else. Don't inline functions marked noinline.
250 if (Callee->mayBeOverridden() ||
251 Callee->hasFnAttr(Attribute::NoInline) || NeverInline.count(Callee))
252 return llvm::InlineCost::getNever();
254 // InlineCost - This value measures how good of an inline candidate this call
255 // site is to inline. A lower inline cost make is more likely for the call to
256 // be inlined. This value may go negative.
260 // If there is only one call of the function, and it has internal linkage,
261 // make it almost guaranteed to be inlined.
263 if (Callee->hasLocalLinkage() && Callee->hasOneUse())
264 InlineCost += InlineConstants::LastCallToStaticBonus;
266 // If this function uses the coldcc calling convention, prefer not to inline
268 if (Callee->getCallingConv() == CallingConv::Cold)
269 InlineCost += InlineConstants::ColdccPenalty;
271 // If the instruction after the call, or if the normal destination of the
272 // invoke is an unreachable instruction, the function is noreturn. As such,
273 // there is little point in inlining this.
274 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) {
275 if (isa<UnreachableInst>(II->getNormalDest()->begin()))
276 InlineCost += InlineConstants::NoreturnPenalty;
277 } else if (isa<UnreachableInst>(++BasicBlock::iterator(TheCall)))
278 InlineCost += InlineConstants::NoreturnPenalty;
280 // Get information about the callee...
281 FunctionInfo &CalleeFI = CachedFunctionInfo[Callee];
283 // If we haven't calculated this information yet, do so now.
284 if (CalleeFI.Metrics.NumBlocks == 0)
285 CalleeFI.analyzeFunction(Callee);
287 // If we should never inline this, return a huge cost.
288 if (CalleeFI.Metrics.NeverInline)
289 return InlineCost::getNever();
291 // FIXME: It would be nice to kill off CalleeFI.NeverInline. Then we
292 // could move this up and avoid computing the FunctionInfo for
293 // things we are going to just return always inline for. This
294 // requires handling setjmp somewhere else, however.
295 if (!Callee->isDeclaration() && Callee->hasFnAttr(Attribute::AlwaysInline))
296 return InlineCost::getAlways();
298 if (CalleeFI.Metrics.usesDynamicAlloca) {
299 // Get infomation about the caller...
300 FunctionInfo &CallerFI = CachedFunctionInfo[Caller];
302 // If we haven't calculated this information yet, do so now.
303 if (CallerFI.Metrics.NumBlocks == 0)
304 CallerFI.analyzeFunction(Caller);
306 // Don't inline a callee with dynamic alloca into a caller without them.
307 // Functions containing dynamic alloca's are inefficient in various ways;
308 // don't create more inefficiency.
309 if (!CallerFI.Metrics.usesDynamicAlloca)
310 return InlineCost::getNever();
313 // Add to the inline quality for properties that make the call valuable to
314 // inline. This includes factors that indicate that the result of inlining
315 // the function will be optimizable. Currently this just looks at arguments
316 // passed into the function.
319 for (CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end();
320 I != E; ++I, ++ArgNo) {
321 // Each argument passed in has a cost at both the caller and the callee
322 // sides. This favors functions that take many arguments over functions
323 // that take few arguments.
326 // If this is a function being passed in, it is very likely that we will be
327 // able to turn an indirect function call into a direct function call.
328 if (isa<Function>(I))
331 // If an alloca is passed in, inlining this function is likely to allow
332 // significant future optimization possibilities (like scalar promotion, and
333 // scalarization), so encourage the inlining of the function.
335 else if (isa<AllocaInst>(I)) {
336 if (ArgNo < CalleeFI.ArgumentWeights.size())
337 InlineCost -= CalleeFI.ArgumentWeights[ArgNo].AllocaWeight;
339 // If this is a constant being passed into the function, use the argument
340 // weights calculated for the callee to determine how much will be folded
341 // away with this information.
342 } else if (isa<Constant>(I)) {
343 if (ArgNo < CalleeFI.ArgumentWeights.size())
344 InlineCost -= CalleeFI.ArgumentWeights[ArgNo].ConstantWeight;
348 // Now that we have considered all of the factors that make the call site more
349 // likely to be inlined, look at factors that make us not want to inline it.
351 // Don't inline into something too big, which would make it bigger.
352 // "size" here is the number of basic blocks, not instructions.
354 InlineCost += Caller->size()/15;
356 // Look at the size of the callee. Each instruction counts as 5.
357 InlineCost += CalleeFI.Metrics.NumInsts*5;
359 return llvm::InlineCost::get(InlineCost);
362 // getInlineFudgeFactor - Return a > 1.0 factor if the inliner should use a
363 // higher threshold to determine if the function call should be inlined.
364 float InlineCostAnalyzer::getInlineFudgeFactor(CallSite CS) {
365 Function *Callee = CS.getCalledFunction();
367 // Get information about the callee...
368 FunctionInfo &CalleeFI = CachedFunctionInfo[Callee];
370 // If we haven't calculated this information yet, do so now.
371 if (CalleeFI.Metrics.NumBlocks == 0)
372 CalleeFI.analyzeFunction(Callee);
375 // Single BB functions are often written to be inlined.
376 if (CalleeFI.Metrics.NumBlocks == 1)
379 // Be more aggressive if the function contains a good chunk (if it mades up
380 // at least 10% of the instructions) of vector instructions.
381 if (CalleeFI.Metrics.NumVectorInsts > CalleeFI.Metrics.NumInsts/2)
383 else if (CalleeFI.Metrics.NumVectorInsts > CalleeFI.Metrics.NumInsts/10)