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